1 | /* Optimize by combining instructions for GNU compiler. |
2 | Copyright (C) 1987-2023 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it under |
7 | the terms of the GNU General Public License as published by the Free |
8 | Software Foundation; either version 3, or (at your option) any later |
9 | version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
14 | for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | /* This module is essentially the "combiner" phase of the U. of Arizona |
21 | Portable Optimizer, but redone to work on our list-structured |
22 | representation for RTL instead of their string representation. |
23 | |
24 | The LOG_LINKS of each insn identify the most recent assignment |
25 | to each REG used in the insn. It is a list of previous insns, |
26 | each of which contains a SET for a REG that is used in this insn |
27 | and not used or set in between. LOG_LINKs never cross basic blocks. |
28 | They were set up by the preceding pass (lifetime analysis). |
29 | |
30 | We try to combine each pair of insns joined by a logical link. |
31 | We also try to combine triplets of insns A, B and C when C has |
32 | a link back to B and B has a link back to A. Likewise for a |
33 | small number of quadruplets of insns A, B, C and D for which |
34 | there's high likelihood of success. |
35 | |
36 | We check (with modified_between_p) to avoid combining in such a way |
37 | as to move a computation to a place where its value would be different. |
38 | |
39 | Combination is done by mathematically substituting the previous |
40 | insn(s) values for the regs they set into the expressions in |
41 | the later insns that refer to these regs. If the result is a valid insn |
42 | for our target machine, according to the machine description, |
43 | we install it, delete the earlier insns, and update the data flow |
44 | information (LOG_LINKS and REG_NOTES) for what we did. |
45 | |
46 | There are a few exceptions where the dataflow information isn't |
47 | completely updated (however this is only a local issue since it is |
48 | regenerated before the next pass that uses it): |
49 | |
50 | - reg_live_length is not updated |
51 | - reg_n_refs is not adjusted in the rare case when a register is |
52 | no longer required in a computation |
53 | - there are extremely rare cases (see distribute_notes) when a |
54 | REG_DEAD note is lost |
55 | - a LOG_LINKS entry that refers to an insn with multiple SETs may be |
56 | removed because there is no way to know which register it was |
57 | linking |
58 | |
59 | To simplify substitution, we combine only when the earlier insn(s) |
60 | consist of only a single assignment. To simplify updating afterward, |
61 | we never combine when a subroutine call appears in the middle. */ |
62 | |
63 | #include "config.h" |
64 | #include "system.h" |
65 | #include "coretypes.h" |
66 | #include "backend.h" |
67 | #include "target.h" |
68 | #include "rtl.h" |
69 | #include "tree.h" |
70 | #include "cfghooks.h" |
71 | #include "predict.h" |
72 | #include "df.h" |
73 | #include "memmodel.h" |
74 | #include "tm_p.h" |
75 | #include "optabs.h" |
76 | #include "regs.h" |
77 | #include "emit-rtl.h" |
78 | #include "recog.h" |
79 | #include "cgraph.h" |
80 | #include "stor-layout.h" |
81 | #include "cfgrtl.h" |
82 | #include "cfgcleanup.h" |
83 | /* Include expr.h after insn-config.h so we get HAVE_conditional_move. */ |
84 | #include "explow.h" |
85 | #include "insn-attr.h" |
86 | #include "rtlhooks-def.h" |
87 | #include "expr.h" |
88 | #include "tree-pass.h" |
89 | #include "valtrack.h" |
90 | #include "rtl-iter.h" |
91 | #include "print-rtl.h" |
92 | #include "function-abi.h" |
93 | #include "rtlanal.h" |
94 | |
95 | /* Number of attempts to combine instructions in this function. */ |
96 | |
97 | static int combine_attempts; |
98 | |
99 | /* Number of attempts that got as far as substitution in this function. */ |
100 | |
101 | static int combine_merges; |
102 | |
103 | /* Number of instructions combined with added SETs in this function. */ |
104 | |
105 | static int ; |
106 | |
107 | /* Number of instructions combined in this function. */ |
108 | |
109 | static int combine_successes; |
110 | |
111 | /* combine_instructions may try to replace the right hand side of the |
112 | second instruction with the value of an associated REG_EQUAL note |
113 | before throwing it at try_combine. That is problematic when there |
114 | is a REG_DEAD note for a register used in the old right hand side |
115 | and can cause distribute_notes to do wrong things. This is the |
116 | second instruction if it has been so modified, null otherwise. */ |
117 | |
118 | static rtx_insn *i2mod; |
119 | |
120 | /* When I2MOD is nonnull, this is a copy of the old right hand side. */ |
121 | |
122 | static rtx i2mod_old_rhs; |
123 | |
124 | /* When I2MOD is nonnull, this is a copy of the new right hand side. */ |
125 | |
126 | static rtx i2mod_new_rhs; |
127 | |
128 | struct reg_stat_type { |
129 | /* Record last point of death of (hard or pseudo) register n. */ |
130 | rtx_insn *last_death; |
131 | |
132 | /* Record last point of modification of (hard or pseudo) register n. */ |
133 | rtx_insn *last_set; |
134 | |
135 | /* The next group of fields allows the recording of the last value assigned |
136 | to (hard or pseudo) register n. We use this information to see if an |
137 | operation being processed is redundant given a prior operation performed |
138 | on the register. For example, an `and' with a constant is redundant if |
139 | all the zero bits are already known to be turned off. |
140 | |
141 | We use an approach similar to that used by cse, but change it in the |
142 | following ways: |
143 | |
144 | (1) We do not want to reinitialize at each label. |
145 | (2) It is useful, but not critical, to know the actual value assigned |
146 | to a register. Often just its form is helpful. |
147 | |
148 | Therefore, we maintain the following fields: |
149 | |
150 | last_set_value the last value assigned |
151 | last_set_label records the value of label_tick when the |
152 | register was assigned |
153 | last_set_table_tick records the value of label_tick when a |
154 | value using the register is assigned |
155 | last_set_invalid set to true when it is not valid |
156 | to use the value of this register in some |
157 | register's value |
158 | |
159 | To understand the usage of these tables, it is important to understand |
160 | the distinction between the value in last_set_value being valid and |
161 | the register being validly contained in some other expression in the |
162 | table. |
163 | |
164 | (The next two parameters are out of date). |
165 | |
166 | reg_stat[i].last_set_value is valid if it is nonzero, and either |
167 | reg_n_sets[i] is 1 or reg_stat[i].last_set_label == label_tick. |
168 | |
169 | Register I may validly appear in any expression returned for the value |
170 | of another register if reg_n_sets[i] is 1. It may also appear in the |
171 | value for register J if reg_stat[j].last_set_invalid is zero, or |
172 | reg_stat[i].last_set_label < reg_stat[j].last_set_label. |
173 | |
174 | If an expression is found in the table containing a register which may |
175 | not validly appear in an expression, the register is replaced by |
176 | something that won't match, (clobber (const_int 0)). */ |
177 | |
178 | /* Record last value assigned to (hard or pseudo) register n. */ |
179 | |
180 | rtx last_set_value; |
181 | |
182 | /* Record the value of label_tick when an expression involving register n |
183 | is placed in last_set_value. */ |
184 | |
185 | int last_set_table_tick; |
186 | |
187 | /* Record the value of label_tick when the value for register n is placed in |
188 | last_set_value. */ |
189 | |
190 | int last_set_label; |
191 | |
192 | /* These fields are maintained in parallel with last_set_value and are |
193 | used to store the mode in which the register was last set, the bits |
194 | that were known to be zero when it was last set, and the number of |
195 | sign bits copies it was known to have when it was last set. */ |
196 | |
197 | unsigned HOST_WIDE_INT last_set_nonzero_bits; |
198 | char last_set_sign_bit_copies; |
199 | ENUM_BITFIELD(machine_mode) last_set_mode : MACHINE_MODE_BITSIZE; |
200 | |
201 | /* Set to true if references to register n in expressions should not be |
202 | used. last_set_invalid is set nonzero when this register is being |
203 | assigned to and last_set_table_tick == label_tick. */ |
204 | |
205 | bool last_set_invalid; |
206 | |
207 | /* Some registers that are set more than once and used in more than one |
208 | basic block are nevertheless always set in similar ways. For example, |
209 | a QImode register may be loaded from memory in two places on a machine |
210 | where byte loads zero extend. |
211 | |
212 | We record in the following fields if a register has some leading bits |
213 | that are always equal to the sign bit, and what we know about the |
214 | nonzero bits of a register, specifically which bits are known to be |
215 | zero. |
216 | |
217 | If an entry is zero, it means that we don't know anything special. */ |
218 | |
219 | unsigned char sign_bit_copies; |
220 | |
221 | unsigned HOST_WIDE_INT nonzero_bits; |
222 | |
223 | /* Record the value of the label_tick when the last truncation |
224 | happened. The field truncated_to_mode is only valid if |
225 | truncation_label == label_tick. */ |
226 | |
227 | int truncation_label; |
228 | |
229 | /* Record the last truncation seen for this register. If truncation |
230 | is not a nop to this mode we might be able to save an explicit |
231 | truncation if we know that value already contains a truncated |
232 | value. */ |
233 | |
234 | ENUM_BITFIELD(machine_mode) truncated_to_mode : MACHINE_MODE_BITSIZE; |
235 | }; |
236 | |
237 | |
238 | static vec<reg_stat_type> reg_stat; |
239 | |
240 | /* One plus the highest pseudo for which we track REG_N_SETS. |
241 | regstat_init_n_sets_and_refs allocates the array for REG_N_SETS just once, |
242 | but during combine_split_insns new pseudos can be created. As we don't have |
243 | updated DF information in that case, it is hard to initialize the array |
244 | after growing. The combiner only cares about REG_N_SETS (regno) == 1, |
245 | so instead of growing the arrays, just assume all newly created pseudos |
246 | during combine might be set multiple times. */ |
247 | |
248 | static unsigned int reg_n_sets_max; |
249 | |
250 | /* Record the luid of the last insn that invalidated memory |
251 | (anything that writes memory, and subroutine calls, but not pushes). */ |
252 | |
253 | static int mem_last_set; |
254 | |
255 | /* Record the luid of the last CALL_INSN |
256 | so we can tell whether a potential combination crosses any calls. */ |
257 | |
258 | static int last_call_luid; |
259 | |
260 | /* When `subst' is called, this is the insn that is being modified |
261 | (by combining in a previous insn). The PATTERN of this insn |
262 | is still the old pattern partially modified and it should not be |
263 | looked at, but this may be used to examine the successors of the insn |
264 | to judge whether a simplification is valid. */ |
265 | |
266 | static rtx_insn *subst_insn; |
267 | |
268 | /* This is the lowest LUID that `subst' is currently dealing with. |
269 | get_last_value will not return a value if the register was set at or |
270 | after this LUID. If not for this mechanism, we could get confused if |
271 | I2 or I1 in try_combine were an insn that used the old value of a register |
272 | to obtain a new value. In that case, we might erroneously get the |
273 | new value of the register when we wanted the old one. */ |
274 | |
275 | static int subst_low_luid; |
276 | |
277 | /* This contains any hard registers that are used in newpat; reg_dead_at_p |
278 | must consider all these registers to be always live. */ |
279 | |
280 | static HARD_REG_SET newpat_used_regs; |
281 | |
282 | /* This is an insn to which a LOG_LINKS entry has been added. If this |
283 | insn is the earlier than I2 or I3, combine should rescan starting at |
284 | that location. */ |
285 | |
286 | static rtx_insn *added_links_insn; |
287 | |
288 | /* And similarly, for notes. */ |
289 | |
290 | static rtx_insn *added_notes_insn; |
291 | |
292 | /* Basic block in which we are performing combines. */ |
293 | static basic_block this_basic_block; |
294 | static bool optimize_this_for_speed_p; |
295 | |
296 | |
297 | /* Length of the currently allocated uid_insn_cost array. */ |
298 | |
299 | static int max_uid_known; |
300 | |
301 | /* The following array records the insn_cost for every insn |
302 | in the instruction stream. */ |
303 | |
304 | static int *uid_insn_cost; |
305 | |
306 | /* The following array records the LOG_LINKS for every insn in the |
307 | instruction stream as struct insn_link pointers. */ |
308 | |
309 | struct insn_link { |
310 | rtx_insn *insn; |
311 | unsigned int regno; |
312 | struct insn_link *next; |
313 | }; |
314 | |
315 | static struct insn_link **uid_log_links; |
316 | |
317 | static inline int |
318 | insn_uid_check (const_rtx insn) |
319 | { |
320 | int uid = INSN_UID (insn); |
321 | gcc_checking_assert (uid <= max_uid_known); |
322 | return uid; |
323 | } |
324 | |
325 | #define INSN_COST(INSN) (uid_insn_cost[insn_uid_check (INSN)]) |
326 | #define LOG_LINKS(INSN) (uid_log_links[insn_uid_check (INSN)]) |
327 | |
328 | #define FOR_EACH_LOG_LINK(L, INSN) \ |
329 | for ((L) = LOG_LINKS (INSN); (L); (L) = (L)->next) |
330 | |
331 | /* Links for LOG_LINKS are allocated from this obstack. */ |
332 | |
333 | static struct obstack insn_link_obstack; |
334 | |
335 | /* Allocate a link. */ |
336 | |
337 | static inline struct insn_link * |
338 | alloc_insn_link (rtx_insn *insn, unsigned int regno, struct insn_link *next) |
339 | { |
340 | struct insn_link *l |
341 | = (struct insn_link *) obstack_alloc (&insn_link_obstack, |
342 | sizeof (struct insn_link)); |
343 | l->insn = insn; |
344 | l->regno = regno; |
345 | l->next = next; |
346 | return l; |
347 | } |
348 | |
349 | /* Incremented for each basic block. */ |
350 | |
351 | static int label_tick; |
352 | |
353 | /* Reset to label_tick for each extended basic block in scanning order. */ |
354 | |
355 | static int label_tick_ebb_start; |
356 | |
357 | /* Mode used to compute significance in reg_stat[].nonzero_bits. It is the |
358 | largest integer mode that can fit in HOST_BITS_PER_WIDE_INT. */ |
359 | |
360 | static scalar_int_mode nonzero_bits_mode; |
361 | |
362 | /* Nonzero when reg_stat[].nonzero_bits and reg_stat[].sign_bit_copies can |
363 | be safely used. It is zero while computing them and after combine has |
364 | completed. This former test prevents propagating values based on |
365 | previously set values, which can be incorrect if a variable is modified |
366 | in a loop. */ |
367 | |
368 | static int nonzero_sign_valid; |
369 | |
370 | |
371 | /* Record one modification to rtl structure |
372 | to be undone by storing old_contents into *where. */ |
373 | |
374 | enum undo_kind { UNDO_RTX, UNDO_INT, UNDO_MODE, UNDO_LINKS }; |
375 | |
376 | struct undo |
377 | { |
378 | struct undo *next; |
379 | enum undo_kind kind; |
380 | union { rtx r; int i; machine_mode m; struct insn_link *l; } old_contents; |
381 | union { rtx *r; int *i; int regno; struct insn_link **l; } where; |
382 | }; |
383 | |
384 | /* Record a bunch of changes to be undone, up to MAX_UNDO of them. |
385 | num_undo says how many are currently recorded. |
386 | |
387 | other_insn is nonzero if we have modified some other insn in the process |
388 | of working on subst_insn. It must be verified too. */ |
389 | |
390 | struct undobuf |
391 | { |
392 | struct undo *undos; |
393 | struct undo *frees; |
394 | rtx_insn *other_insn; |
395 | }; |
396 | |
397 | static struct undobuf undobuf; |
398 | |
399 | /* Number of times the pseudo being substituted for |
400 | was found and replaced. */ |
401 | |
402 | static int n_occurrences; |
403 | |
404 | static rtx reg_nonzero_bits_for_combine (const_rtx, scalar_int_mode, |
405 | scalar_int_mode, |
406 | unsigned HOST_WIDE_INT *); |
407 | static rtx reg_num_sign_bit_copies_for_combine (const_rtx, scalar_int_mode, |
408 | scalar_int_mode, |
409 | unsigned int *); |
410 | static void do_SUBST (rtx *, rtx); |
411 | static void do_SUBST_INT (int *, int); |
412 | static void init_reg_last (void); |
413 | static void setup_incoming_promotions (rtx_insn *); |
414 | static void set_nonzero_bits_and_sign_copies (rtx, const_rtx, void *); |
415 | static bool cant_combine_insn_p (rtx_insn *); |
416 | static bool can_combine_p (rtx_insn *, rtx_insn *, rtx_insn *, rtx_insn *, |
417 | rtx_insn *, rtx_insn *, rtx *, rtx *); |
418 | static bool combinable_i3pat (rtx_insn *, rtx *, rtx, rtx, rtx, |
419 | bool, bool, rtx *); |
420 | static bool contains_muldiv (rtx); |
421 | static rtx_insn *try_combine (rtx_insn *, rtx_insn *, rtx_insn *, rtx_insn *, |
422 | bool *, rtx_insn *); |
423 | static void undo_all (void); |
424 | static void undo_commit (void); |
425 | static rtx *find_split_point (rtx *, rtx_insn *, bool); |
426 | static rtx subst (rtx, rtx, rtx, bool, bool, bool); |
427 | static rtx combine_simplify_rtx (rtx, machine_mode, bool, bool); |
428 | static rtx simplify_if_then_else (rtx); |
429 | static rtx simplify_set (rtx); |
430 | static rtx simplify_logical (rtx); |
431 | static rtx expand_compound_operation (rtx); |
432 | static const_rtx expand_field_assignment (const_rtx); |
433 | static rtx make_extraction (machine_mode, rtx, HOST_WIDE_INT, rtx, |
434 | unsigned HOST_WIDE_INT, bool, bool, bool); |
435 | static int get_pos_from_mask (unsigned HOST_WIDE_INT, |
436 | unsigned HOST_WIDE_INT *); |
437 | static rtx canon_reg_for_combine (rtx, rtx); |
438 | static rtx force_int_to_mode (rtx, scalar_int_mode, scalar_int_mode, |
439 | scalar_int_mode, unsigned HOST_WIDE_INT, bool); |
440 | static rtx force_to_mode (rtx, machine_mode, |
441 | unsigned HOST_WIDE_INT, bool); |
442 | static rtx if_then_else_cond (rtx, rtx *, rtx *); |
443 | static rtx known_cond (rtx, enum rtx_code, rtx, rtx); |
444 | static bool rtx_equal_for_field_assignment_p (rtx, rtx, bool = false); |
445 | static rtx make_field_assignment (rtx); |
446 | static rtx apply_distributive_law (rtx); |
447 | static rtx distribute_and_simplify_rtx (rtx, int); |
448 | static rtx simplify_and_const_int_1 (scalar_int_mode, rtx, |
449 | unsigned HOST_WIDE_INT); |
450 | static rtx simplify_and_const_int (rtx, scalar_int_mode, rtx, |
451 | unsigned HOST_WIDE_INT); |
452 | static bool merge_outer_ops (enum rtx_code *, HOST_WIDE_INT *, enum rtx_code, |
453 | HOST_WIDE_INT, machine_mode, bool *); |
454 | static rtx simplify_shift_const_1 (enum rtx_code, machine_mode, rtx, int); |
455 | static rtx simplify_shift_const (rtx, enum rtx_code, machine_mode, rtx, |
456 | int); |
457 | static int recog_for_combine (rtx *, rtx_insn *, rtx *); |
458 | static rtx gen_lowpart_for_combine (machine_mode, rtx); |
459 | static enum rtx_code simplify_compare_const (enum rtx_code, machine_mode, |
460 | rtx *, rtx *); |
461 | static enum rtx_code simplify_comparison (enum rtx_code, rtx *, rtx *); |
462 | static void update_table_tick (rtx); |
463 | static void record_value_for_reg (rtx, rtx_insn *, rtx); |
464 | static void check_promoted_subreg (rtx_insn *, rtx); |
465 | static void record_dead_and_set_regs_1 (rtx, const_rtx, void *); |
466 | static void record_dead_and_set_regs (rtx_insn *); |
467 | static bool get_last_value_validate (rtx *, rtx_insn *, int, bool); |
468 | static rtx get_last_value (const_rtx); |
469 | static void reg_dead_at_p_1 (rtx, const_rtx, void *); |
470 | static bool reg_dead_at_p (rtx, rtx_insn *); |
471 | static void move_deaths (rtx, rtx, int, rtx_insn *, rtx *); |
472 | static bool reg_bitfield_target_p (rtx, rtx); |
473 | static void distribute_notes (rtx, rtx_insn *, rtx_insn *, rtx_insn *, |
474 | rtx, rtx, rtx); |
475 | static void distribute_links (struct insn_link *); |
476 | static void mark_used_regs_combine (rtx); |
477 | static void record_promoted_value (rtx_insn *, rtx); |
478 | static bool unmentioned_reg_p (rtx, rtx); |
479 | static void record_truncated_values (rtx *, void *); |
480 | static bool reg_truncated_to_mode (machine_mode, const_rtx); |
481 | static rtx gen_lowpart_or_truncate (machine_mode, rtx); |
482 | |
483 | |
484 | /* It is not safe to use ordinary gen_lowpart in combine. |
485 | See comments in gen_lowpart_for_combine. */ |
486 | #undef RTL_HOOKS_GEN_LOWPART |
487 | #define RTL_HOOKS_GEN_LOWPART gen_lowpart_for_combine |
488 | |
489 | /* Our implementation of gen_lowpart never emits a new pseudo. */ |
490 | #undef RTL_HOOKS_GEN_LOWPART_NO_EMIT |
491 | #define RTL_HOOKS_GEN_LOWPART_NO_EMIT gen_lowpart_for_combine |
492 | |
493 | #undef RTL_HOOKS_REG_NONZERO_REG_BITS |
494 | #define RTL_HOOKS_REG_NONZERO_REG_BITS reg_nonzero_bits_for_combine |
495 | |
496 | #undef RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES |
497 | #define RTL_HOOKS_REG_NUM_SIGN_BIT_COPIES reg_num_sign_bit_copies_for_combine |
498 | |
499 | #undef RTL_HOOKS_REG_TRUNCATED_TO_MODE |
500 | #define RTL_HOOKS_REG_TRUNCATED_TO_MODE reg_truncated_to_mode |
501 | |
502 | static const struct rtl_hooks combine_rtl_hooks = RTL_HOOKS_INITIALIZER; |
503 | |
504 | |
505 | /* Convenience wrapper for the canonicalize_comparison target hook. |
506 | Target hooks cannot use enum rtx_code. */ |
507 | static inline void |
508 | target_canonicalize_comparison (enum rtx_code *code, rtx *op0, rtx *op1, |
509 | bool op0_preserve_value) |
510 | { |
511 | int code_int = (int)*code; |
512 | targetm.canonicalize_comparison (&code_int, op0, op1, op0_preserve_value); |
513 | *code = (enum rtx_code)code_int; |
514 | } |
515 | |
516 | /* Try to split PATTERN found in INSN. This returns NULL_RTX if |
517 | PATTERN cannot be split. Otherwise, it returns an insn sequence. |
518 | This is a wrapper around split_insns which ensures that the |
519 | reg_stat vector is made larger if the splitter creates a new |
520 | register. */ |
521 | |
522 | static rtx_insn * |
523 | combine_split_insns (rtx pattern, rtx_insn *insn) |
524 | { |
525 | rtx_insn *ret; |
526 | unsigned int nregs; |
527 | |
528 | ret = split_insns (pattern, insn); |
529 | nregs = max_reg_num (); |
530 | if (nregs > reg_stat.length ()) |
531 | reg_stat.safe_grow_cleared (len: nregs, exact: true); |
532 | return ret; |
533 | } |
534 | |
535 | /* This is used by find_single_use to locate an rtx in LOC that |
536 | contains exactly one use of DEST, which is typically a REG. |
537 | It returns a pointer to the innermost rtx expression |
538 | containing DEST. Appearances of DEST that are being used to |
539 | totally replace it are not counted. */ |
540 | |
541 | static rtx * |
542 | find_single_use_1 (rtx dest, rtx *loc) |
543 | { |
544 | rtx x = *loc; |
545 | enum rtx_code code = GET_CODE (x); |
546 | rtx *result = NULL; |
547 | rtx *this_result; |
548 | int i; |
549 | const char *fmt; |
550 | |
551 | switch (code) |
552 | { |
553 | case CONST: |
554 | case LABEL_REF: |
555 | case SYMBOL_REF: |
556 | CASE_CONST_ANY: |
557 | case CLOBBER: |
558 | return 0; |
559 | |
560 | case SET: |
561 | /* If the destination is anything other than PC, a REG or a SUBREG |
562 | of a REG that occupies all of the REG, the insn uses DEST if |
563 | it is mentioned in the destination or the source. Otherwise, we |
564 | need just check the source. */ |
565 | if (GET_CODE (SET_DEST (x)) != PC |
566 | && !REG_P (SET_DEST (x)) |
567 | && ! (GET_CODE (SET_DEST (x)) == SUBREG |
568 | && REG_P (SUBREG_REG (SET_DEST (x))) |
569 | && !read_modify_subreg_p (SET_DEST (x)))) |
570 | break; |
571 | |
572 | return find_single_use_1 (dest, loc: &SET_SRC (x)); |
573 | |
574 | case MEM: |
575 | case SUBREG: |
576 | return find_single_use_1 (dest, loc: &XEXP (x, 0)); |
577 | |
578 | default: |
579 | break; |
580 | } |
581 | |
582 | /* If it wasn't one of the common cases above, check each expression and |
583 | vector of this code. Look for a unique usage of DEST. */ |
584 | |
585 | fmt = GET_RTX_FORMAT (code); |
586 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
587 | { |
588 | if (fmt[i] == 'e') |
589 | { |
590 | if (dest == XEXP (x, i) |
591 | || (REG_P (dest) && REG_P (XEXP (x, i)) |
592 | && REGNO (dest) == REGNO (XEXP (x, i)))) |
593 | this_result = loc; |
594 | else |
595 | this_result = find_single_use_1 (dest, loc: &XEXP (x, i)); |
596 | |
597 | if (result == NULL) |
598 | result = this_result; |
599 | else if (this_result) |
600 | /* Duplicate usage. */ |
601 | return NULL; |
602 | } |
603 | else if (fmt[i] == 'E') |
604 | { |
605 | int j; |
606 | |
607 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
608 | { |
609 | if (XVECEXP (x, i, j) == dest |
610 | || (REG_P (dest) |
611 | && REG_P (XVECEXP (x, i, j)) |
612 | && REGNO (XVECEXP (x, i, j)) == REGNO (dest))) |
613 | this_result = loc; |
614 | else |
615 | this_result = find_single_use_1 (dest, loc: &XVECEXP (x, i, j)); |
616 | |
617 | if (result == NULL) |
618 | result = this_result; |
619 | else if (this_result) |
620 | return NULL; |
621 | } |
622 | } |
623 | } |
624 | |
625 | return result; |
626 | } |
627 | |
628 | |
629 | /* See if DEST, produced in INSN, is used only a single time in the |
630 | sequel. If so, return a pointer to the innermost rtx expression in which |
631 | it is used. |
632 | |
633 | If PLOC is nonzero, *PLOC is set to the insn containing the single use. |
634 | |
635 | Otherwise, we find the single use by finding an insn that has a |
636 | LOG_LINKS pointing at INSN and has a REG_DEAD note for DEST. If DEST is |
637 | only referenced once in that insn, we know that it must be the first |
638 | and last insn referencing DEST. */ |
639 | |
640 | static rtx * |
641 | find_single_use (rtx dest, rtx_insn *insn, rtx_insn **ploc) |
642 | { |
643 | basic_block bb; |
644 | rtx_insn *next; |
645 | rtx *result; |
646 | struct insn_link *link; |
647 | |
648 | if (!REG_P (dest)) |
649 | return 0; |
650 | |
651 | bb = BLOCK_FOR_INSN (insn); |
652 | for (next = NEXT_INSN (insn); |
653 | next && BLOCK_FOR_INSN (insn: next) == bb; |
654 | next = NEXT_INSN (insn: next)) |
655 | if (NONDEBUG_INSN_P (next) && dead_or_set_p (next, dest)) |
656 | { |
657 | FOR_EACH_LOG_LINK (link, next) |
658 | if (link->insn == insn && link->regno == REGNO (dest)) |
659 | break; |
660 | |
661 | if (link) |
662 | { |
663 | result = find_single_use_1 (dest, loc: &PATTERN (insn: next)); |
664 | if (ploc) |
665 | *ploc = next; |
666 | return result; |
667 | } |
668 | } |
669 | |
670 | return 0; |
671 | } |
672 | |
673 | /* Substitute NEWVAL, an rtx expression, into INTO, a place in some |
674 | insn. The substitution can be undone by undo_all. If INTO is already |
675 | set to NEWVAL, do not record this change. Because computing NEWVAL might |
676 | also call SUBST, we have to compute it before we put anything into |
677 | the undo table. */ |
678 | |
679 | static void |
680 | do_SUBST (rtx *into, rtx newval) |
681 | { |
682 | struct undo *buf; |
683 | rtx oldval = *into; |
684 | |
685 | if (oldval == newval) |
686 | return; |
687 | |
688 | /* We'd like to catch as many invalid transformations here as |
689 | possible. Unfortunately, there are way too many mode changes |
690 | that are perfectly valid, so we'd waste too much effort for |
691 | little gain doing the checks here. Focus on catching invalid |
692 | transformations involving integer constants. */ |
693 | if (GET_MODE_CLASS (GET_MODE (oldval)) == MODE_INT |
694 | && CONST_INT_P (newval)) |
695 | { |
696 | /* Sanity check that we're replacing oldval with a CONST_INT |
697 | that is a valid sign-extension for the original mode. */ |
698 | gcc_assert (INTVAL (newval) |
699 | == trunc_int_for_mode (INTVAL (newval), GET_MODE (oldval))); |
700 | |
701 | /* Replacing the operand of a SUBREG or a ZERO_EXTEND with a |
702 | CONST_INT is not valid, because after the replacement, the |
703 | original mode would be gone. Unfortunately, we can't tell |
704 | when do_SUBST is called to replace the operand thereof, so we |
705 | perform this test on oldval instead, checking whether an |
706 | invalid replacement took place before we got here. */ |
707 | gcc_assert (!(GET_CODE (oldval) == SUBREG |
708 | && CONST_INT_P (SUBREG_REG (oldval)))); |
709 | gcc_assert (!(GET_CODE (oldval) == ZERO_EXTEND |
710 | && CONST_INT_P (XEXP (oldval, 0)))); |
711 | } |
712 | |
713 | if (undobuf.frees) |
714 | buf = undobuf.frees, undobuf.frees = buf->next; |
715 | else |
716 | buf = XNEW (struct undo); |
717 | |
718 | buf->kind = UNDO_RTX; |
719 | buf->where.r = into; |
720 | buf->old_contents.r = oldval; |
721 | *into = newval; |
722 | |
723 | buf->next = undobuf.undos, undobuf.undos = buf; |
724 | } |
725 | |
726 | #define SUBST(INTO, NEWVAL) do_SUBST (&(INTO), (NEWVAL)) |
727 | |
728 | /* Similar to SUBST, but NEWVAL is an int expression. Note that substitution |
729 | for the value of a HOST_WIDE_INT value (including CONST_INT) is |
730 | not safe. */ |
731 | |
732 | static void |
733 | do_SUBST_INT (int *into, int newval) |
734 | { |
735 | struct undo *buf; |
736 | int oldval = *into; |
737 | |
738 | if (oldval == newval) |
739 | return; |
740 | |
741 | if (undobuf.frees) |
742 | buf = undobuf.frees, undobuf.frees = buf->next; |
743 | else |
744 | buf = XNEW (struct undo); |
745 | |
746 | buf->kind = UNDO_INT; |
747 | buf->where.i = into; |
748 | buf->old_contents.i = oldval; |
749 | *into = newval; |
750 | |
751 | buf->next = undobuf.undos, undobuf.undos = buf; |
752 | } |
753 | |
754 | #define SUBST_INT(INTO, NEWVAL) do_SUBST_INT (&(INTO), (NEWVAL)) |
755 | |
756 | /* Similar to SUBST, but just substitute the mode. This is used when |
757 | changing the mode of a pseudo-register, so that any other |
758 | references to the entry in the regno_reg_rtx array will change as |
759 | well. */ |
760 | |
761 | static void |
762 | subst_mode (int regno, machine_mode newval) |
763 | { |
764 | struct undo *buf; |
765 | rtx reg = regno_reg_rtx[regno]; |
766 | machine_mode oldval = GET_MODE (reg); |
767 | |
768 | if (oldval == newval) |
769 | return; |
770 | |
771 | if (undobuf.frees) |
772 | buf = undobuf.frees, undobuf.frees = buf->next; |
773 | else |
774 | buf = XNEW (struct undo); |
775 | |
776 | buf->kind = UNDO_MODE; |
777 | buf->where.regno = regno; |
778 | buf->old_contents.m = oldval; |
779 | adjust_reg_mode (reg, newval); |
780 | |
781 | buf->next = undobuf.undos, undobuf.undos = buf; |
782 | } |
783 | |
784 | /* Similar to SUBST, but NEWVAL is a LOG_LINKS expression. */ |
785 | |
786 | static void |
787 | do_SUBST_LINK (struct insn_link **into, struct insn_link *newval) |
788 | { |
789 | struct undo *buf; |
790 | struct insn_link * oldval = *into; |
791 | |
792 | if (oldval == newval) |
793 | return; |
794 | |
795 | if (undobuf.frees) |
796 | buf = undobuf.frees, undobuf.frees = buf->next; |
797 | else |
798 | buf = XNEW (struct undo); |
799 | |
800 | buf->kind = UNDO_LINKS; |
801 | buf->where.l = into; |
802 | buf->old_contents.l = oldval; |
803 | *into = newval; |
804 | |
805 | buf->next = undobuf.undos, undobuf.undos = buf; |
806 | } |
807 | |
808 | #define SUBST_LINK(oldval, newval) do_SUBST_LINK (&oldval, newval) |
809 | |
810 | /* Subroutine of try_combine. Determine whether the replacement patterns |
811 | NEWPAT, NEWI2PAT and NEWOTHERPAT are cheaper according to insn_cost |
812 | than the original sequence I0, I1, I2, I3 and undobuf.other_insn. Note |
813 | that I0, I1 and/or NEWI2PAT may be NULL_RTX. Similarly, NEWOTHERPAT and |
814 | undobuf.other_insn may also both be NULL_RTX. Return false if the cost |
815 | of all the instructions can be estimated and the replacements are more |
816 | expensive than the original sequence. */ |
817 | |
818 | static bool |
819 | combine_validate_cost (rtx_insn *i0, rtx_insn *i1, rtx_insn *i2, rtx_insn *i3, |
820 | rtx newpat, rtx newi2pat, rtx newotherpat) |
821 | { |
822 | int i0_cost, i1_cost, i2_cost, i3_cost; |
823 | int new_i2_cost, new_i3_cost; |
824 | int old_cost, new_cost; |
825 | |
826 | /* Lookup the original insn_costs. */ |
827 | i2_cost = INSN_COST (i2); |
828 | i3_cost = INSN_COST (i3); |
829 | |
830 | if (i1) |
831 | { |
832 | i1_cost = INSN_COST (i1); |
833 | if (i0) |
834 | { |
835 | i0_cost = INSN_COST (i0); |
836 | old_cost = (i0_cost > 0 && i1_cost > 0 && i2_cost > 0 && i3_cost > 0 |
837 | ? i0_cost + i1_cost + i2_cost + i3_cost : 0); |
838 | } |
839 | else |
840 | { |
841 | old_cost = (i1_cost > 0 && i2_cost > 0 && i3_cost > 0 |
842 | ? i1_cost + i2_cost + i3_cost : 0); |
843 | i0_cost = 0; |
844 | } |
845 | } |
846 | else |
847 | { |
848 | old_cost = (i2_cost > 0 && i3_cost > 0) ? i2_cost + i3_cost : 0; |
849 | i1_cost = i0_cost = 0; |
850 | } |
851 | |
852 | /* If we have split a PARALLEL I2 to I1,I2, we have counted its cost twice; |
853 | correct that. */ |
854 | if (old_cost && i1 && INSN_UID (insn: i1) == INSN_UID (insn: i2)) |
855 | old_cost -= i1_cost; |
856 | |
857 | |
858 | /* Calculate the replacement insn_costs. */ |
859 | rtx tmp = PATTERN (insn: i3); |
860 | PATTERN (insn: i3) = newpat; |
861 | int tmpi = INSN_CODE (i3); |
862 | INSN_CODE (i3) = -1; |
863 | new_i3_cost = insn_cost (i3, optimize_this_for_speed_p); |
864 | PATTERN (insn: i3) = tmp; |
865 | INSN_CODE (i3) = tmpi; |
866 | if (newi2pat) |
867 | { |
868 | tmp = PATTERN (insn: i2); |
869 | PATTERN (insn: i2) = newi2pat; |
870 | tmpi = INSN_CODE (i2); |
871 | INSN_CODE (i2) = -1; |
872 | new_i2_cost = insn_cost (i2, optimize_this_for_speed_p); |
873 | PATTERN (insn: i2) = tmp; |
874 | INSN_CODE (i2) = tmpi; |
875 | new_cost = (new_i2_cost > 0 && new_i3_cost > 0) |
876 | ? new_i2_cost + new_i3_cost : 0; |
877 | } |
878 | else |
879 | { |
880 | new_cost = new_i3_cost; |
881 | new_i2_cost = 0; |
882 | } |
883 | |
884 | if (undobuf.other_insn) |
885 | { |
886 | int old_other_cost, new_other_cost; |
887 | |
888 | old_other_cost = INSN_COST (undobuf.other_insn); |
889 | tmp = PATTERN (insn: undobuf.other_insn); |
890 | PATTERN (insn: undobuf.other_insn) = newotherpat; |
891 | tmpi = INSN_CODE (undobuf.other_insn); |
892 | INSN_CODE (undobuf.other_insn) = -1; |
893 | new_other_cost = insn_cost (undobuf.other_insn, |
894 | optimize_this_for_speed_p); |
895 | PATTERN (insn: undobuf.other_insn) = tmp; |
896 | INSN_CODE (undobuf.other_insn) = tmpi; |
897 | if (old_other_cost > 0 && new_other_cost > 0) |
898 | { |
899 | old_cost += old_other_cost; |
900 | new_cost += new_other_cost; |
901 | } |
902 | else |
903 | old_cost = 0; |
904 | } |
905 | |
906 | /* Disallow this combination if both new_cost and old_cost are greater than |
907 | zero, and new_cost is greater than old cost. */ |
908 | bool reject = old_cost > 0 && new_cost > old_cost; |
909 | |
910 | if (dump_file) |
911 | { |
912 | fprintf (stream: dump_file, format: "%s combination of insns " , |
913 | reject ? "rejecting" : "allowing" ); |
914 | if (i0) |
915 | fprintf (stream: dump_file, format: "%d, " , INSN_UID (insn: i0)); |
916 | if (i1 && INSN_UID (insn: i1) != INSN_UID (insn: i2)) |
917 | fprintf (stream: dump_file, format: "%d, " , INSN_UID (insn: i1)); |
918 | fprintf (stream: dump_file, format: "%d and %d\n" , INSN_UID (insn: i2), INSN_UID (insn: i3)); |
919 | |
920 | fprintf (stream: dump_file, format: "original costs " ); |
921 | if (i0) |
922 | fprintf (stream: dump_file, format: "%d + " , i0_cost); |
923 | if (i1 && INSN_UID (insn: i1) != INSN_UID (insn: i2)) |
924 | fprintf (stream: dump_file, format: "%d + " , i1_cost); |
925 | fprintf (stream: dump_file, format: "%d + %d = %d\n" , i2_cost, i3_cost, old_cost); |
926 | |
927 | if (newi2pat) |
928 | fprintf (stream: dump_file, format: "replacement costs %d + %d = %d\n" , |
929 | new_i2_cost, new_i3_cost, new_cost); |
930 | else |
931 | fprintf (stream: dump_file, format: "replacement cost %d\n" , new_cost); |
932 | } |
933 | |
934 | if (reject) |
935 | return false; |
936 | |
937 | /* Update the uid_insn_cost array with the replacement costs. */ |
938 | INSN_COST (i2) = new_i2_cost; |
939 | INSN_COST (i3) = new_i3_cost; |
940 | if (i1) |
941 | { |
942 | INSN_COST (i1) = 0; |
943 | if (i0) |
944 | INSN_COST (i0) = 0; |
945 | } |
946 | |
947 | return true; |
948 | } |
949 | |
950 | |
951 | /* Delete any insns that copy a register to itself. |
952 | Return true if the CFG was changed. */ |
953 | |
954 | static bool |
955 | delete_noop_moves (void) |
956 | { |
957 | rtx_insn *insn, *next; |
958 | basic_block bb; |
959 | |
960 | bool edges_deleted = false; |
961 | |
962 | FOR_EACH_BB_FN (bb, cfun) |
963 | { |
964 | for (insn = BB_HEAD (bb); insn != NEXT_INSN (BB_END (bb)); insn = next) |
965 | { |
966 | next = NEXT_INSN (insn); |
967 | if (INSN_P (insn) && noop_move_p (insn)) |
968 | { |
969 | if (dump_file) |
970 | fprintf (stream: dump_file, format: "deleting noop move %d\n" , INSN_UID (insn)); |
971 | |
972 | edges_deleted |= delete_insn_and_edges (insn); |
973 | } |
974 | } |
975 | } |
976 | |
977 | return edges_deleted; |
978 | } |
979 | |
980 | |
981 | /* Return false if we do not want to (or cannot) combine DEF. */ |
982 | static bool |
983 | can_combine_def_p (df_ref def) |
984 | { |
985 | /* Do not consider if it is pre/post modification in MEM. */ |
986 | if (DF_REF_FLAGS (def) & DF_REF_PRE_POST_MODIFY) |
987 | return false; |
988 | |
989 | unsigned int regno = DF_REF_REGNO (def); |
990 | |
991 | /* Do not combine frame pointer adjustments. */ |
992 | if ((regno == FRAME_POINTER_REGNUM |
993 | && (!reload_completed || frame_pointer_needed)) |
994 | || (!HARD_FRAME_POINTER_IS_FRAME_POINTER |
995 | && regno == HARD_FRAME_POINTER_REGNUM |
996 | && (!reload_completed || frame_pointer_needed)) |
997 | || (FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM |
998 | && regno == ARG_POINTER_REGNUM && fixed_regs[regno])) |
999 | return false; |
1000 | |
1001 | return true; |
1002 | } |
1003 | |
1004 | /* Return false if we do not want to (or cannot) combine USE. */ |
1005 | static bool |
1006 | can_combine_use_p (df_ref use) |
1007 | { |
1008 | /* Do not consider the usage of the stack pointer by function call. */ |
1009 | if (DF_REF_FLAGS (use) & DF_REF_CALL_STACK_USAGE) |
1010 | return false; |
1011 | |
1012 | return true; |
1013 | } |
1014 | |
1015 | /* Fill in log links field for all insns. */ |
1016 | |
1017 | static void |
1018 | create_log_links (void) |
1019 | { |
1020 | basic_block bb; |
1021 | rtx_insn **next_use; |
1022 | rtx_insn *insn; |
1023 | df_ref def, use; |
1024 | |
1025 | next_use = XCNEWVEC (rtx_insn *, max_reg_num ()); |
1026 | |
1027 | /* Pass through each block from the end, recording the uses of each |
1028 | register and establishing log links when def is encountered. |
1029 | Note that we do not clear next_use array in order to save time, |
1030 | so we have to test whether the use is in the same basic block as def. |
1031 | |
1032 | There are a few cases below when we do not consider the definition or |
1033 | usage -- these are taken from original flow.c did. Don't ask me why it is |
1034 | done this way; I don't know and if it works, I don't want to know. */ |
1035 | |
1036 | FOR_EACH_BB_FN (bb, cfun) |
1037 | { |
1038 | FOR_BB_INSNS_REVERSE (bb, insn) |
1039 | { |
1040 | if (!NONDEBUG_INSN_P (insn)) |
1041 | continue; |
1042 | |
1043 | /* Log links are created only once. */ |
1044 | gcc_assert (!LOG_LINKS (insn)); |
1045 | |
1046 | FOR_EACH_INSN_DEF (def, insn) |
1047 | { |
1048 | unsigned int regno = DF_REF_REGNO (def); |
1049 | rtx_insn *use_insn; |
1050 | |
1051 | if (!next_use[regno]) |
1052 | continue; |
1053 | |
1054 | if (!can_combine_def_p (def)) |
1055 | continue; |
1056 | |
1057 | use_insn = next_use[regno]; |
1058 | next_use[regno] = NULL; |
1059 | |
1060 | if (BLOCK_FOR_INSN (insn: use_insn) != bb) |
1061 | continue; |
1062 | |
1063 | /* flow.c claimed: |
1064 | |
1065 | We don't build a LOG_LINK for hard registers contained |
1066 | in ASM_OPERANDs. If these registers get replaced, |
1067 | we might wind up changing the semantics of the insn, |
1068 | even if reload can make what appear to be valid |
1069 | assignments later. */ |
1070 | if (regno < FIRST_PSEUDO_REGISTER |
1071 | && asm_noperands (PATTERN (insn: use_insn)) >= 0) |
1072 | continue; |
1073 | |
1074 | /* Don't add duplicate links between instructions. */ |
1075 | struct insn_link *links; |
1076 | FOR_EACH_LOG_LINK (links, use_insn) |
1077 | if (insn == links->insn && regno == links->regno) |
1078 | break; |
1079 | |
1080 | if (!links) |
1081 | LOG_LINKS (use_insn) |
1082 | = alloc_insn_link (insn, regno, LOG_LINKS (use_insn)); |
1083 | } |
1084 | |
1085 | FOR_EACH_INSN_USE (use, insn) |
1086 | if (can_combine_use_p (use)) |
1087 | next_use[DF_REF_REGNO (use)] = insn; |
1088 | } |
1089 | } |
1090 | |
1091 | free (ptr: next_use); |
1092 | } |
1093 | |
1094 | /* Walk the LOG_LINKS of insn B to see if we find a reference to A. Return |
1095 | true if we found a LOG_LINK that proves that A feeds B. This only works |
1096 | if there are no instructions between A and B which could have a link |
1097 | depending on A, since in that case we would not record a link for B. */ |
1098 | |
1099 | static bool |
1100 | insn_a_feeds_b (rtx_insn *a, rtx_insn *b) |
1101 | { |
1102 | struct insn_link *links; |
1103 | FOR_EACH_LOG_LINK (links, b) |
1104 | if (links->insn == a) |
1105 | return true; |
1106 | return false; |
1107 | } |
1108 | |
1109 | /* Main entry point for combiner. F is the first insn of the function. |
1110 | NREGS is the first unused pseudo-reg number. |
1111 | |
1112 | Return nonzero if the CFG was changed (e.g. if the combiner has |
1113 | turned an indirect jump instruction into a direct jump). */ |
1114 | static bool |
1115 | combine_instructions (rtx_insn *f, unsigned int nregs) |
1116 | { |
1117 | rtx_insn *insn, *next; |
1118 | struct insn_link *links, *nextlinks; |
1119 | rtx_insn *first; |
1120 | basic_block last_bb; |
1121 | |
1122 | bool new_direct_jump_p = false; |
1123 | |
1124 | for (first = f; first && !NONDEBUG_INSN_P (first); ) |
1125 | first = NEXT_INSN (insn: first); |
1126 | if (!first) |
1127 | return false; |
1128 | |
1129 | combine_attempts = 0; |
1130 | combine_merges = 0; |
1131 | combine_extras = 0; |
1132 | combine_successes = 0; |
1133 | |
1134 | rtl_hooks = combine_rtl_hooks; |
1135 | |
1136 | reg_stat.safe_grow_cleared (len: nregs, exact: true); |
1137 | |
1138 | init_recog_no_volatile (); |
1139 | |
1140 | /* Allocate array for insn info. */ |
1141 | max_uid_known = get_max_uid (); |
1142 | uid_log_links = XCNEWVEC (struct insn_link *, max_uid_known + 1); |
1143 | uid_insn_cost = XCNEWVEC (int, max_uid_known + 1); |
1144 | gcc_obstack_init (&insn_link_obstack); |
1145 | |
1146 | nonzero_bits_mode = int_mode_for_size (HOST_BITS_PER_WIDE_INT, limit: 0).require (); |
1147 | |
1148 | /* Don't use reg_stat[].nonzero_bits when computing it. This can cause |
1149 | problems when, for example, we have j <<= 1 in a loop. */ |
1150 | |
1151 | nonzero_sign_valid = 0; |
1152 | label_tick = label_tick_ebb_start = 1; |
1153 | |
1154 | /* Scan all SETs and see if we can deduce anything about what |
1155 | bits are known to be zero for some registers and how many copies |
1156 | of the sign bit are known to exist for those registers. |
1157 | |
1158 | Also set any known values so that we can use it while searching |
1159 | for what bits are known to be set. */ |
1160 | |
1161 | setup_incoming_promotions (first); |
1162 | /* Allow the entry block and the first block to fall into the same EBB. |
1163 | Conceptually the incoming promotions are assigned to the entry block. */ |
1164 | last_bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
1165 | |
1166 | create_log_links (); |
1167 | FOR_EACH_BB_FN (this_basic_block, cfun) |
1168 | { |
1169 | optimize_this_for_speed_p = optimize_bb_for_speed_p (this_basic_block); |
1170 | last_call_luid = 0; |
1171 | mem_last_set = -1; |
1172 | |
1173 | label_tick++; |
1174 | if (!single_pred_p (bb: this_basic_block) |
1175 | || single_pred (bb: this_basic_block) != last_bb) |
1176 | label_tick_ebb_start = label_tick; |
1177 | last_bb = this_basic_block; |
1178 | |
1179 | FOR_BB_INSNS (this_basic_block, insn) |
1180 | if (INSN_P (insn) && BLOCK_FOR_INSN (insn)) |
1181 | { |
1182 | rtx links; |
1183 | |
1184 | subst_low_luid = DF_INSN_LUID (insn); |
1185 | subst_insn = insn; |
1186 | |
1187 | note_stores (insn, set_nonzero_bits_and_sign_copies, insn); |
1188 | record_dead_and_set_regs (insn); |
1189 | |
1190 | if (AUTO_INC_DEC) |
1191 | for (links = REG_NOTES (insn); links; links = XEXP (links, 1)) |
1192 | if (REG_NOTE_KIND (links) == REG_INC) |
1193 | set_nonzero_bits_and_sign_copies (XEXP (links, 0), NULL_RTX, |
1194 | insn); |
1195 | |
1196 | /* Record the current insn_cost of this instruction. */ |
1197 | INSN_COST (insn) = insn_cost (insn, optimize_this_for_speed_p); |
1198 | if (dump_file) |
1199 | { |
1200 | fprintf (stream: dump_file, format: "insn_cost %d for " , INSN_COST (insn)); |
1201 | dump_insn_slim (dump_file, insn); |
1202 | } |
1203 | } |
1204 | } |
1205 | |
1206 | nonzero_sign_valid = 1; |
1207 | |
1208 | /* Now scan all the insns in forward order. */ |
1209 | label_tick = label_tick_ebb_start = 1; |
1210 | init_reg_last (); |
1211 | setup_incoming_promotions (first); |
1212 | last_bb = ENTRY_BLOCK_PTR_FOR_FN (cfun); |
1213 | int max_combine = param_max_combine_insns; |
1214 | |
1215 | FOR_EACH_BB_FN (this_basic_block, cfun) |
1216 | { |
1217 | rtx_insn *last_combined_insn = NULL; |
1218 | |
1219 | /* Ignore instruction combination in basic blocks that are going to |
1220 | be removed as unreachable anyway. See PR82386. */ |
1221 | if (EDGE_COUNT (this_basic_block->preds) == 0) |
1222 | continue; |
1223 | |
1224 | optimize_this_for_speed_p = optimize_bb_for_speed_p (this_basic_block); |
1225 | last_call_luid = 0; |
1226 | mem_last_set = -1; |
1227 | |
1228 | label_tick++; |
1229 | if (!single_pred_p (bb: this_basic_block) |
1230 | || single_pred (bb: this_basic_block) != last_bb) |
1231 | label_tick_ebb_start = label_tick; |
1232 | last_bb = this_basic_block; |
1233 | |
1234 | rtl_profile_for_bb (this_basic_block); |
1235 | for (insn = BB_HEAD (this_basic_block); |
1236 | insn != NEXT_INSN (BB_END (this_basic_block)); |
1237 | insn = next ? next : NEXT_INSN (insn)) |
1238 | { |
1239 | next = 0; |
1240 | if (!NONDEBUG_INSN_P (insn)) |
1241 | continue; |
1242 | |
1243 | while (last_combined_insn |
1244 | && (!NONDEBUG_INSN_P (last_combined_insn) |
1245 | || last_combined_insn->deleted ())) |
1246 | last_combined_insn = PREV_INSN (insn: last_combined_insn); |
1247 | if (last_combined_insn == NULL_RTX |
1248 | || BLOCK_FOR_INSN (insn: last_combined_insn) != this_basic_block |
1249 | || DF_INSN_LUID (last_combined_insn) <= DF_INSN_LUID (insn)) |
1250 | last_combined_insn = insn; |
1251 | |
1252 | /* See if we know about function return values before this |
1253 | insn based upon SUBREG flags. */ |
1254 | check_promoted_subreg (insn, PATTERN (insn)); |
1255 | |
1256 | /* See if we can find hardregs and subreg of pseudos in |
1257 | narrower modes. This could help turning TRUNCATEs |
1258 | into SUBREGs. */ |
1259 | note_uses (&PATTERN (insn), record_truncated_values, NULL); |
1260 | |
1261 | /* Try this insn with each insn it links back to. */ |
1262 | |
1263 | FOR_EACH_LOG_LINK (links, insn) |
1264 | if ((next = try_combine (insn, links->insn, NULL, |
1265 | NULL, &new_direct_jump_p, |
1266 | last_combined_insn)) != 0) |
1267 | { |
1268 | statistics_counter_event (cfun, "two-insn combine" , 1); |
1269 | goto retry; |
1270 | } |
1271 | |
1272 | /* Try each sequence of three linked insns ending with this one. */ |
1273 | |
1274 | if (max_combine >= 3) |
1275 | FOR_EACH_LOG_LINK (links, insn) |
1276 | { |
1277 | rtx_insn *link = links->insn; |
1278 | |
1279 | /* If the linked insn has been replaced by a note, then there |
1280 | is no point in pursuing this chain any further. */ |
1281 | if (NOTE_P (link)) |
1282 | continue; |
1283 | |
1284 | FOR_EACH_LOG_LINK (nextlinks, link) |
1285 | if ((next = try_combine (insn, link, nextlinks->insn, |
1286 | NULL, &new_direct_jump_p, |
1287 | last_combined_insn)) != 0) |
1288 | { |
1289 | statistics_counter_event (cfun, "three-insn combine" , 1); |
1290 | goto retry; |
1291 | } |
1292 | } |
1293 | |
1294 | /* Try combining an insn with two different insns whose results it |
1295 | uses. */ |
1296 | if (max_combine >= 3) |
1297 | FOR_EACH_LOG_LINK (links, insn) |
1298 | for (nextlinks = links->next; nextlinks; |
1299 | nextlinks = nextlinks->next) |
1300 | if ((next = try_combine (insn, links->insn, |
1301 | nextlinks->insn, NULL, |
1302 | &new_direct_jump_p, |
1303 | last_combined_insn)) != 0) |
1304 | |
1305 | { |
1306 | statistics_counter_event (cfun, "three-insn combine" , 1); |
1307 | goto retry; |
1308 | } |
1309 | |
1310 | /* Try four-instruction combinations. */ |
1311 | if (max_combine >= 4) |
1312 | FOR_EACH_LOG_LINK (links, insn) |
1313 | { |
1314 | struct insn_link *next1; |
1315 | rtx_insn *link = links->insn; |
1316 | |
1317 | /* If the linked insn has been replaced by a note, then there |
1318 | is no point in pursuing this chain any further. */ |
1319 | if (NOTE_P (link)) |
1320 | continue; |
1321 | |
1322 | FOR_EACH_LOG_LINK (next1, link) |
1323 | { |
1324 | rtx_insn *link1 = next1->insn; |
1325 | if (NOTE_P (link1)) |
1326 | continue; |
1327 | /* I0 -> I1 -> I2 -> I3. */ |
1328 | FOR_EACH_LOG_LINK (nextlinks, link1) |
1329 | if ((next = try_combine (insn, link, link1, |
1330 | nextlinks->insn, |
1331 | &new_direct_jump_p, |
1332 | last_combined_insn)) != 0) |
1333 | { |
1334 | statistics_counter_event (cfun, "four-insn combine" , 1); |
1335 | goto retry; |
1336 | } |
1337 | /* I0, I1 -> I2, I2 -> I3. */ |
1338 | for (nextlinks = next1->next; nextlinks; |
1339 | nextlinks = nextlinks->next) |
1340 | if ((next = try_combine (insn, link, link1, |
1341 | nextlinks->insn, |
1342 | &new_direct_jump_p, |
1343 | last_combined_insn)) != 0) |
1344 | { |
1345 | statistics_counter_event (cfun, "four-insn combine" , 1); |
1346 | goto retry; |
1347 | } |
1348 | } |
1349 | |
1350 | for (next1 = links->next; next1; next1 = next1->next) |
1351 | { |
1352 | rtx_insn *link1 = next1->insn; |
1353 | if (NOTE_P (link1)) |
1354 | continue; |
1355 | /* I0 -> I2; I1, I2 -> I3. */ |
1356 | FOR_EACH_LOG_LINK (nextlinks, link) |
1357 | if ((next = try_combine (insn, link, link1, |
1358 | nextlinks->insn, |
1359 | &new_direct_jump_p, |
1360 | last_combined_insn)) != 0) |
1361 | { |
1362 | statistics_counter_event (cfun, "four-insn combine" , 1); |
1363 | goto retry; |
1364 | } |
1365 | /* I0 -> I1; I1, I2 -> I3. */ |
1366 | FOR_EACH_LOG_LINK (nextlinks, link1) |
1367 | if ((next = try_combine (insn, link, link1, |
1368 | nextlinks->insn, |
1369 | &new_direct_jump_p, |
1370 | last_combined_insn)) != 0) |
1371 | { |
1372 | statistics_counter_event (cfun, "four-insn combine" , 1); |
1373 | goto retry; |
1374 | } |
1375 | } |
1376 | } |
1377 | |
1378 | /* Try this insn with each REG_EQUAL note it links back to. */ |
1379 | FOR_EACH_LOG_LINK (links, insn) |
1380 | { |
1381 | rtx set, note; |
1382 | rtx_insn *temp = links->insn; |
1383 | if ((set = single_set (insn: temp)) != 0 |
1384 | && (note = find_reg_equal_equiv_note (temp)) != 0 |
1385 | && (note = XEXP (note, 0), GET_CODE (note)) != EXPR_LIST |
1386 | && ! side_effects_p (SET_SRC (set)) |
1387 | /* Avoid using a register that may already been marked |
1388 | dead by an earlier instruction. */ |
1389 | && ! unmentioned_reg_p (note, SET_SRC (set)) |
1390 | && (GET_MODE (note) == VOIDmode |
1391 | ? SCALAR_INT_MODE_P (GET_MODE (SET_DEST (set))) |
1392 | : (GET_MODE (SET_DEST (set)) == GET_MODE (note) |
1393 | && (GET_CODE (SET_DEST (set)) != ZERO_EXTRACT |
1394 | || (GET_MODE (XEXP (SET_DEST (set), 0)) |
1395 | == GET_MODE (note)))))) |
1396 | { |
1397 | /* Temporarily replace the set's source with the |
1398 | contents of the REG_EQUAL note. The insn will |
1399 | be deleted or recognized by try_combine. */ |
1400 | rtx orig_src = SET_SRC (set); |
1401 | rtx orig_dest = SET_DEST (set); |
1402 | if (GET_CODE (SET_DEST (set)) == ZERO_EXTRACT) |
1403 | SET_DEST (set) = XEXP (SET_DEST (set), 0); |
1404 | SET_SRC (set) = note; |
1405 | i2mod = temp; |
1406 | i2mod_old_rhs = copy_rtx (orig_src); |
1407 | i2mod_new_rhs = copy_rtx (note); |
1408 | next = try_combine (insn, i2mod, NULL, NULL, |
1409 | &new_direct_jump_p, |
1410 | last_combined_insn); |
1411 | i2mod = NULL; |
1412 | if (next) |
1413 | { |
1414 | statistics_counter_event (cfun, "insn-with-note combine" , 1); |
1415 | goto retry; |
1416 | } |
1417 | INSN_CODE (temp) = -1; |
1418 | SET_SRC (set) = orig_src; |
1419 | SET_DEST (set) = orig_dest; |
1420 | } |
1421 | } |
1422 | |
1423 | if (!NOTE_P (insn)) |
1424 | record_dead_and_set_regs (insn); |
1425 | |
1426 | retry: |
1427 | ; |
1428 | } |
1429 | } |
1430 | |
1431 | default_rtl_profile (); |
1432 | clear_bb_flags (); |
1433 | |
1434 | if (purge_all_dead_edges ()) |
1435 | new_direct_jump_p = true; |
1436 | if (delete_noop_moves ()) |
1437 | new_direct_jump_p = true; |
1438 | |
1439 | /* Clean up. */ |
1440 | obstack_free (&insn_link_obstack, NULL); |
1441 | free (ptr: uid_log_links); |
1442 | free (ptr: uid_insn_cost); |
1443 | reg_stat.release (); |
1444 | |
1445 | { |
1446 | struct undo *undo, *next; |
1447 | for (undo = undobuf.frees; undo; undo = next) |
1448 | { |
1449 | next = undo->next; |
1450 | free (ptr: undo); |
1451 | } |
1452 | undobuf.frees = 0; |
1453 | } |
1454 | |
1455 | statistics_counter_event (cfun, "attempts" , combine_attempts); |
1456 | statistics_counter_event (cfun, "merges" , combine_merges); |
1457 | statistics_counter_event (cfun, "extras" , combine_extras); |
1458 | statistics_counter_event (cfun, "successes" , combine_successes); |
1459 | |
1460 | nonzero_sign_valid = 0; |
1461 | rtl_hooks = general_rtl_hooks; |
1462 | |
1463 | /* Make recognizer allow volatile MEMs again. */ |
1464 | init_recog (); |
1465 | |
1466 | return new_direct_jump_p; |
1467 | } |
1468 | |
1469 | /* Wipe the last_xxx fields of reg_stat in preparation for another pass. */ |
1470 | |
1471 | static void |
1472 | init_reg_last (void) |
1473 | { |
1474 | unsigned int i; |
1475 | reg_stat_type *p; |
1476 | |
1477 | FOR_EACH_VEC_ELT (reg_stat, i, p) |
1478 | memset (s: p, c: 0, offsetof (reg_stat_type, sign_bit_copies)); |
1479 | } |
1480 | |
1481 | /* Set up any promoted values for incoming argument registers. */ |
1482 | |
1483 | static void |
1484 | setup_incoming_promotions (rtx_insn *first) |
1485 | { |
1486 | tree arg; |
1487 | bool strictly_local = false; |
1488 | |
1489 | for (arg = DECL_ARGUMENTS (current_function_decl); arg; |
1490 | arg = DECL_CHAIN (arg)) |
1491 | { |
1492 | rtx x, reg = DECL_INCOMING_RTL (arg); |
1493 | int uns1, uns3; |
1494 | machine_mode mode1, mode2, mode3, mode4; |
1495 | |
1496 | /* Only continue if the incoming argument is in a register. */ |
1497 | if (!REG_P (reg)) |
1498 | continue; |
1499 | |
1500 | /* Determine, if possible, whether all call sites of the current |
1501 | function lie within the current compilation unit. (This does |
1502 | take into account the exporting of a function via taking its |
1503 | address, and so forth.) */ |
1504 | strictly_local |
1505 | = cgraph_node::local_info_node (decl: current_function_decl)->local; |
1506 | |
1507 | /* The mode and signedness of the argument before any promotions happen |
1508 | (equal to the mode of the pseudo holding it at that stage). */ |
1509 | mode1 = TYPE_MODE (TREE_TYPE (arg)); |
1510 | uns1 = TYPE_UNSIGNED (TREE_TYPE (arg)); |
1511 | |
1512 | /* The mode and signedness of the argument after any source language and |
1513 | TARGET_PROMOTE_PROTOTYPES-driven promotions. */ |
1514 | mode2 = TYPE_MODE (DECL_ARG_TYPE (arg)); |
1515 | uns3 = TYPE_UNSIGNED (DECL_ARG_TYPE (arg)); |
1516 | |
1517 | /* The mode and signedness of the argument as it is actually passed, |
1518 | see assign_parm_setup_reg in function.cc. */ |
1519 | mode3 = promote_function_mode (TREE_TYPE (arg), mode1, &uns3, |
1520 | TREE_TYPE (cfun->decl), 0); |
1521 | |
1522 | /* The mode of the register in which the argument is being passed. */ |
1523 | mode4 = GET_MODE (reg); |
1524 | |
1525 | /* Eliminate sign extensions in the callee when: |
1526 | (a) A mode promotion has occurred; */ |
1527 | if (mode1 == mode3) |
1528 | continue; |
1529 | /* (b) The mode of the register is the same as the mode of |
1530 | the argument as it is passed; */ |
1531 | if (mode3 != mode4) |
1532 | continue; |
1533 | /* (c) There's no language level extension; */ |
1534 | if (mode1 == mode2) |
1535 | ; |
1536 | /* (c.1) All callers are from the current compilation unit. If that's |
1537 | the case we don't have to rely on an ABI, we only have to know |
1538 | what we're generating right now, and we know that we will do the |
1539 | mode1 to mode2 promotion with the given sign. */ |
1540 | else if (!strictly_local) |
1541 | continue; |
1542 | /* (c.2) The combination of the two promotions is useful. This is |
1543 | true when the signs match, or if the first promotion is unsigned. |
1544 | In the later case, (sign_extend (zero_extend x)) is the same as |
1545 | (zero_extend (zero_extend x)), so make sure to force UNS3 true. */ |
1546 | else if (uns1) |
1547 | uns3 = true; |
1548 | else if (uns3) |
1549 | continue; |
1550 | |
1551 | /* Record that the value was promoted from mode1 to mode3, |
1552 | so that any sign extension at the head of the current |
1553 | function may be eliminated. */ |
1554 | x = gen_rtx_CLOBBER (mode1, const0_rtx); |
1555 | x = gen_rtx_fmt_e ((uns3 ? ZERO_EXTEND : SIGN_EXTEND), mode3, x); |
1556 | record_value_for_reg (reg, first, x); |
1557 | } |
1558 | } |
1559 | |
1560 | /* If MODE has a precision lower than PREC and SRC is a non-negative constant |
1561 | that would appear negative in MODE, sign-extend SRC for use in nonzero_bits |
1562 | because some machines (maybe most) will actually do the sign-extension and |
1563 | this is the conservative approach. |
1564 | |
1565 | ??? For 2.5, try to tighten up the MD files in this regard instead of this |
1566 | kludge. */ |
1567 | |
1568 | static rtx |
1569 | sign_extend_short_imm (rtx src, machine_mode mode, unsigned int prec) |
1570 | { |
1571 | scalar_int_mode int_mode; |
1572 | if (CONST_INT_P (src) |
1573 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
1574 | && GET_MODE_PRECISION (mode: int_mode) < prec |
1575 | && INTVAL (src) > 0 |
1576 | && val_signbit_known_set_p (int_mode, INTVAL (src))) |
1577 | src = GEN_INT (INTVAL (src) | ~GET_MODE_MASK (int_mode)); |
1578 | |
1579 | return src; |
1580 | } |
1581 | |
1582 | /* Update RSP for pseudo-register X from INSN's REG_EQUAL note (if one exists) |
1583 | and SET. */ |
1584 | |
1585 | static void |
1586 | update_rsp_from_reg_equal (reg_stat_type *rsp, rtx_insn *insn, const_rtx set, |
1587 | rtx x) |
1588 | { |
1589 | rtx reg_equal_note = insn ? find_reg_equal_equiv_note (insn) : NULL_RTX; |
1590 | unsigned HOST_WIDE_INT bits = 0; |
1591 | rtx reg_equal = NULL, src = SET_SRC (set); |
1592 | unsigned int num = 0; |
1593 | |
1594 | if (reg_equal_note) |
1595 | reg_equal = XEXP (reg_equal_note, 0); |
1596 | |
1597 | if (SHORT_IMMEDIATES_SIGN_EXTEND) |
1598 | { |
1599 | src = sign_extend_short_imm (src, GET_MODE (x), BITS_PER_WORD); |
1600 | if (reg_equal) |
1601 | reg_equal = sign_extend_short_imm (src: reg_equal, GET_MODE (x), BITS_PER_WORD); |
1602 | } |
1603 | |
1604 | /* Don't call nonzero_bits if it cannot change anything. */ |
1605 | if (rsp->nonzero_bits != HOST_WIDE_INT_M1U) |
1606 | { |
1607 | machine_mode mode = GET_MODE (x); |
1608 | if (GET_MODE_CLASS (mode) == MODE_INT |
1609 | && HWI_COMPUTABLE_MODE_P (mode)) |
1610 | mode = nonzero_bits_mode; |
1611 | bits = nonzero_bits (src, mode); |
1612 | if (reg_equal && bits) |
1613 | bits &= nonzero_bits (reg_equal, mode); |
1614 | rsp->nonzero_bits |= bits; |
1615 | } |
1616 | |
1617 | /* Don't call num_sign_bit_copies if it cannot change anything. */ |
1618 | if (rsp->sign_bit_copies != 1) |
1619 | { |
1620 | num = num_sign_bit_copies (SET_SRC (set), GET_MODE (x)); |
1621 | if (reg_equal && maybe_ne (a: num, b: GET_MODE_PRECISION (GET_MODE (x)))) |
1622 | { |
1623 | unsigned int numeq = num_sign_bit_copies (reg_equal, GET_MODE (x)); |
1624 | if (num == 0 || numeq > num) |
1625 | num = numeq; |
1626 | } |
1627 | if (rsp->sign_bit_copies == 0 || num < rsp->sign_bit_copies) |
1628 | rsp->sign_bit_copies = num; |
1629 | } |
1630 | } |
1631 | |
1632 | /* Called via note_stores. If X is a pseudo that is narrower than |
1633 | HOST_BITS_PER_WIDE_INT and is being set, record what bits are known zero. |
1634 | |
1635 | If we are setting only a portion of X and we can't figure out what |
1636 | portion, assume all bits will be used since we don't know what will |
1637 | be happening. |
1638 | |
1639 | Similarly, set how many bits of X are known to be copies of the sign bit |
1640 | at all locations in the function. This is the smallest number implied |
1641 | by any set of X. */ |
1642 | |
1643 | static void |
1644 | set_nonzero_bits_and_sign_copies (rtx x, const_rtx set, void *data) |
1645 | { |
1646 | rtx_insn *insn = (rtx_insn *) data; |
1647 | scalar_int_mode mode; |
1648 | |
1649 | if (REG_P (x) |
1650 | && REGNO (x) >= FIRST_PSEUDO_REGISTER |
1651 | /* If this register is undefined at the start of the file, we can't |
1652 | say what its contents were. */ |
1653 | && ! REGNO_REG_SET_P |
1654 | (DF_LR_IN (ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb), REGNO (x)) |
1655 | && is_a <scalar_int_mode> (GET_MODE (x), result: &mode) |
1656 | && HWI_COMPUTABLE_MODE_P (mode)) |
1657 | { |
1658 | reg_stat_type *rsp = ®_stat[REGNO (x)]; |
1659 | |
1660 | if (set == 0 || GET_CODE (set) == CLOBBER) |
1661 | { |
1662 | rsp->nonzero_bits = GET_MODE_MASK (mode); |
1663 | rsp->sign_bit_copies = 1; |
1664 | return; |
1665 | } |
1666 | |
1667 | /* If this register is being initialized using itself, and the |
1668 | register is uninitialized in this basic block, and there are |
1669 | no LOG_LINKS which set the register, then part of the |
1670 | register is uninitialized. In that case we can't assume |
1671 | anything about the number of nonzero bits. |
1672 | |
1673 | ??? We could do better if we checked this in |
1674 | reg_{nonzero_bits,num_sign_bit_copies}_for_combine. Then we |
1675 | could avoid making assumptions about the insn which initially |
1676 | sets the register, while still using the information in other |
1677 | insns. We would have to be careful to check every insn |
1678 | involved in the combination. */ |
1679 | |
1680 | if (insn |
1681 | && reg_referenced_p (x, PATTERN (insn)) |
1682 | && !REGNO_REG_SET_P (DF_LR_IN (BLOCK_FOR_INSN (insn)), |
1683 | REGNO (x))) |
1684 | { |
1685 | struct insn_link *link; |
1686 | |
1687 | FOR_EACH_LOG_LINK (link, insn) |
1688 | if (dead_or_set_p (link->insn, x)) |
1689 | break; |
1690 | if (!link) |
1691 | { |
1692 | rsp->nonzero_bits = GET_MODE_MASK (mode); |
1693 | rsp->sign_bit_copies = 1; |
1694 | return; |
1695 | } |
1696 | } |
1697 | |
1698 | /* If this is a complex assignment, see if we can convert it into a |
1699 | simple assignment. */ |
1700 | set = expand_field_assignment (set); |
1701 | |
1702 | /* If this is a simple assignment, or we have a paradoxical SUBREG, |
1703 | set what we know about X. */ |
1704 | |
1705 | if (SET_DEST (set) == x |
1706 | || (paradoxical_subreg_p (SET_DEST (set)) |
1707 | && SUBREG_REG (SET_DEST (set)) == x)) |
1708 | update_rsp_from_reg_equal (rsp, insn, set, x); |
1709 | else |
1710 | { |
1711 | rsp->nonzero_bits = GET_MODE_MASK (mode); |
1712 | rsp->sign_bit_copies = 1; |
1713 | } |
1714 | } |
1715 | } |
1716 | |
1717 | /* See if INSN can be combined into I3. PRED, PRED2, SUCC and SUCC2 are |
1718 | optionally insns that were previously combined into I3 or that will be |
1719 | combined into the merger of INSN and I3. The order is PRED, PRED2, |
1720 | INSN, SUCC, SUCC2, I3. |
1721 | |
1722 | Return false if the combination is not allowed for any reason. |
1723 | |
1724 | If the combination is allowed, *PDEST will be set to the single |
1725 | destination of INSN and *PSRC to the single source, and this function |
1726 | will return true. */ |
1727 | |
1728 | static bool |
1729 | can_combine_p (rtx_insn *insn, rtx_insn *i3, rtx_insn *pred ATTRIBUTE_UNUSED, |
1730 | rtx_insn *pred2 ATTRIBUTE_UNUSED, rtx_insn *succ, rtx_insn *succ2, |
1731 | rtx *pdest, rtx *psrc) |
1732 | { |
1733 | int i; |
1734 | const_rtx set = 0; |
1735 | rtx src, dest; |
1736 | rtx_insn *p; |
1737 | rtx link; |
1738 | bool all_adjacent = true; |
1739 | bool (*is_volatile_p) (const_rtx); |
1740 | |
1741 | if (succ) |
1742 | { |
1743 | if (succ2) |
1744 | { |
1745 | if (next_active_insn (succ2) != i3) |
1746 | all_adjacent = false; |
1747 | if (next_active_insn (succ) != succ2) |
1748 | all_adjacent = false; |
1749 | } |
1750 | else if (next_active_insn (succ) != i3) |
1751 | all_adjacent = false; |
1752 | if (next_active_insn (insn) != succ) |
1753 | all_adjacent = false; |
1754 | } |
1755 | else if (next_active_insn (insn) != i3) |
1756 | all_adjacent = false; |
1757 | |
1758 | /* Can combine only if previous insn is a SET of a REG or a SUBREG, |
1759 | or a PARALLEL consisting of such a SET and CLOBBERs. |
1760 | |
1761 | If INSN has CLOBBER parallel parts, ignore them for our processing. |
1762 | By definition, these happen during the execution of the insn. When it |
1763 | is merged with another insn, all bets are off. If they are, in fact, |
1764 | needed and aren't also supplied in I3, they may be added by |
1765 | recog_for_combine. Otherwise, it won't match. |
1766 | |
1767 | We can also ignore a SET whose SET_DEST is mentioned in a REG_UNUSED |
1768 | note. |
1769 | |
1770 | Get the source and destination of INSN. If more than one, can't |
1771 | combine. */ |
1772 | |
1773 | if (GET_CODE (PATTERN (insn)) == SET) |
1774 | set = PATTERN (insn); |
1775 | else if (GET_CODE (PATTERN (insn)) == PARALLEL |
1776 | && GET_CODE (XVECEXP (PATTERN (insn), 0, 0)) == SET) |
1777 | { |
1778 | for (i = 0; i < XVECLEN (PATTERN (insn), 0); i++) |
1779 | { |
1780 | rtx elt = XVECEXP (PATTERN (insn), 0, i); |
1781 | |
1782 | switch (GET_CODE (elt)) |
1783 | { |
1784 | /* This is important to combine floating point insns |
1785 | for the SH4 port. */ |
1786 | case USE: |
1787 | /* Combining an isolated USE doesn't make sense. |
1788 | We depend here on combinable_i3pat to reject them. */ |
1789 | /* The code below this loop only verifies that the inputs of |
1790 | the SET in INSN do not change. We call reg_set_between_p |
1791 | to verify that the REG in the USE does not change between |
1792 | I3 and INSN. |
1793 | If the USE in INSN was for a pseudo register, the matching |
1794 | insn pattern will likely match any register; combining this |
1795 | with any other USE would only be safe if we knew that the |
1796 | used registers have identical values, or if there was |
1797 | something to tell them apart, e.g. different modes. For |
1798 | now, we forgo such complicated tests and simply disallow |
1799 | combining of USES of pseudo registers with any other USE. */ |
1800 | if (REG_P (XEXP (elt, 0)) |
1801 | && GET_CODE (PATTERN (i3)) == PARALLEL) |
1802 | { |
1803 | rtx i3pat = PATTERN (insn: i3); |
1804 | int i = XVECLEN (i3pat, 0) - 1; |
1805 | unsigned int regno = REGNO (XEXP (elt, 0)); |
1806 | |
1807 | do |
1808 | { |
1809 | rtx i3elt = XVECEXP (i3pat, 0, i); |
1810 | |
1811 | if (GET_CODE (i3elt) == USE |
1812 | && REG_P (XEXP (i3elt, 0)) |
1813 | && (REGNO (XEXP (i3elt, 0)) == regno |
1814 | ? reg_set_between_p (XEXP (elt, 0), |
1815 | PREV_INSN (insn), i3) |
1816 | : regno >= FIRST_PSEUDO_REGISTER)) |
1817 | return false; |
1818 | } |
1819 | while (--i >= 0); |
1820 | } |
1821 | break; |
1822 | |
1823 | /* We can ignore CLOBBERs. */ |
1824 | case CLOBBER: |
1825 | break; |
1826 | |
1827 | case SET: |
1828 | /* Ignore SETs whose result isn't used but not those that |
1829 | have side-effects. */ |
1830 | if (find_reg_note (insn, REG_UNUSED, SET_DEST (elt)) |
1831 | && insn_nothrow_p (insn) |
1832 | && !side_effects_p (elt)) |
1833 | break; |
1834 | |
1835 | /* If we have already found a SET, this is a second one and |
1836 | so we cannot combine with this insn. */ |
1837 | if (set) |
1838 | return false; |
1839 | |
1840 | set = elt; |
1841 | break; |
1842 | |
1843 | default: |
1844 | /* Anything else means we can't combine. */ |
1845 | return false; |
1846 | } |
1847 | } |
1848 | |
1849 | if (set == 0 |
1850 | /* If SET_SRC is an ASM_OPERANDS we can't throw away these CLOBBERs, |
1851 | so don't do anything with it. */ |
1852 | || GET_CODE (SET_SRC (set)) == ASM_OPERANDS) |
1853 | return false; |
1854 | } |
1855 | else |
1856 | return false; |
1857 | |
1858 | if (set == 0) |
1859 | return false; |
1860 | |
1861 | /* The simplification in expand_field_assignment may call back to |
1862 | get_last_value, so set safe guard here. */ |
1863 | subst_low_luid = DF_INSN_LUID (insn); |
1864 | |
1865 | set = expand_field_assignment (set); |
1866 | src = SET_SRC (set), dest = SET_DEST (set); |
1867 | |
1868 | /* Do not eliminate user-specified register if it is in an |
1869 | asm input because we may break the register asm usage defined |
1870 | in GCC manual if allow to do so. |
1871 | Be aware that this may cover more cases than we expect but this |
1872 | should be harmless. */ |
1873 | if (REG_P (dest) && REG_USERVAR_P (dest) && HARD_REGISTER_P (dest) |
1874 | && extract_asm_operands (PATTERN (insn: i3))) |
1875 | return false; |
1876 | |
1877 | /* Don't eliminate a store in the stack pointer. */ |
1878 | if (dest == stack_pointer_rtx |
1879 | /* Don't combine with an insn that sets a register to itself if it has |
1880 | a REG_EQUAL note. This may be part of a LIBCALL sequence. */ |
1881 | || (rtx_equal_p (src, dest) && find_reg_note (insn, REG_EQUAL, NULL_RTX)) |
1882 | /* Can't merge an ASM_OPERANDS. */ |
1883 | || GET_CODE (src) == ASM_OPERANDS |
1884 | /* Can't merge a function call. */ |
1885 | || GET_CODE (src) == CALL |
1886 | /* Don't eliminate a function call argument. */ |
1887 | || (CALL_P (i3) |
1888 | && (find_reg_fusage (i3, USE, dest) |
1889 | || (REG_P (dest) |
1890 | && REGNO (dest) < FIRST_PSEUDO_REGISTER |
1891 | && global_regs[REGNO (dest)]))) |
1892 | /* Don't substitute into an incremented register. */ |
1893 | || FIND_REG_INC_NOTE (i3, dest) |
1894 | || (succ && FIND_REG_INC_NOTE (succ, dest)) |
1895 | || (succ2 && FIND_REG_INC_NOTE (succ2, dest)) |
1896 | /* Don't substitute into a non-local goto, this confuses CFG. */ |
1897 | || (JUMP_P (i3) && find_reg_note (i3, REG_NON_LOCAL_GOTO, NULL_RTX)) |
1898 | /* Make sure that DEST is not used after INSN but before SUCC, or |
1899 | after SUCC and before SUCC2, or after SUCC2 but before I3. */ |
1900 | || (!all_adjacent |
1901 | && ((succ2 |
1902 | && (reg_used_between_p (dest, succ2, i3) |
1903 | || reg_used_between_p (dest, succ, succ2))) |
1904 | || (!succ2 && succ && reg_used_between_p (dest, succ, i3)) |
1905 | || (!succ2 && !succ && reg_used_between_p (dest, insn, i3)) |
1906 | || (succ |
1907 | /* SUCC and SUCC2 can be split halves from a PARALLEL; in |
1908 | that case SUCC is not in the insn stream, so use SUCC2 |
1909 | instead for this test. */ |
1910 | && reg_used_between_p (dest, insn, |
1911 | succ2 |
1912 | && INSN_UID (insn: succ) == INSN_UID (insn: succ2) |
1913 | ? succ2 : succ)))) |
1914 | /* Make sure that the value that is to be substituted for the register |
1915 | does not use any registers whose values alter in between. However, |
1916 | If the insns are adjacent, a use can't cross a set even though we |
1917 | think it might (this can happen for a sequence of insns each setting |
1918 | the same destination; last_set of that register might point to |
1919 | a NOTE). If INSN has a REG_EQUIV note, the register is always |
1920 | equivalent to the memory so the substitution is valid even if there |
1921 | are intervening stores. Also, don't move a volatile asm or |
1922 | UNSPEC_VOLATILE across any other insns. */ |
1923 | || (! all_adjacent |
1924 | && (((!MEM_P (src) |
1925 | || ! find_reg_note (insn, REG_EQUIV, src)) |
1926 | && modified_between_p (src, insn, i3)) |
1927 | || (GET_CODE (src) == ASM_OPERANDS && MEM_VOLATILE_P (src)) |
1928 | || GET_CODE (src) == UNSPEC_VOLATILE)) |
1929 | /* Don't combine across a CALL_INSN, because that would possibly |
1930 | change whether the life span of some REGs crosses calls or not, |
1931 | and it is a pain to update that information. |
1932 | Exception: if source is a constant, moving it later can't hurt. |
1933 | Accept that as a special case. */ |
1934 | || (DF_INSN_LUID (insn) < last_call_luid && ! CONSTANT_P (src))) |
1935 | return false; |
1936 | |
1937 | /* DEST must be a REG. */ |
1938 | if (REG_P (dest)) |
1939 | { |
1940 | /* If register alignment is being enforced for multi-word items in all |
1941 | cases except for parameters, it is possible to have a register copy |
1942 | insn referencing a hard register that is not allowed to contain the |
1943 | mode being copied and which would not be valid as an operand of most |
1944 | insns. Eliminate this problem by not combining with such an insn. |
1945 | |
1946 | Also, on some machines we don't want to extend the life of a hard |
1947 | register. */ |
1948 | |
1949 | if (REG_P (src) |
1950 | && ((REGNO (dest) < FIRST_PSEUDO_REGISTER |
1951 | && !targetm.hard_regno_mode_ok (REGNO (dest), GET_MODE (dest))) |
1952 | /* Don't extend the life of a hard register unless it is |
1953 | user variable (if we have few registers) or it can't |
1954 | fit into the desired register (meaning something special |
1955 | is going on). |
1956 | Also avoid substituting a return register into I3, because |
1957 | reload can't handle a conflict with constraints of other |
1958 | inputs. */ |
1959 | || (REGNO (src) < FIRST_PSEUDO_REGISTER |
1960 | && !targetm.hard_regno_mode_ok (REGNO (src), |
1961 | GET_MODE (src))))) |
1962 | return false; |
1963 | } |
1964 | else |
1965 | return false; |
1966 | |
1967 | |
1968 | if (GET_CODE (PATTERN (i3)) == PARALLEL) |
1969 | for (i = XVECLEN (PATTERN (i3), 0) - 1; i >= 0; i--) |
1970 | if (GET_CODE (XVECEXP (PATTERN (i3), 0, i)) == CLOBBER) |
1971 | { |
1972 | rtx reg = XEXP (XVECEXP (PATTERN (i3), 0, i), 0); |
1973 | |
1974 | /* If the clobber represents an earlyclobber operand, we must not |
1975 | substitute an expression containing the clobbered register. |
1976 | As we do not analyze the constraint strings here, we have to |
1977 | make the conservative assumption. However, if the register is |
1978 | a fixed hard reg, the clobber cannot represent any operand; |
1979 | we leave it up to the machine description to either accept or |
1980 | reject use-and-clobber patterns. */ |
1981 | if (!REG_P (reg) |
1982 | || REGNO (reg) >= FIRST_PSEUDO_REGISTER |
1983 | || !fixed_regs[REGNO (reg)]) |
1984 | if (reg_overlap_mentioned_p (reg, src)) |
1985 | return false; |
1986 | } |
1987 | |
1988 | /* If INSN contains anything volatile, or is an `asm' (whether volatile |
1989 | or not), reject, unless nothing volatile comes between it and I3 */ |
1990 | |
1991 | if (GET_CODE (src) == ASM_OPERANDS || volatile_refs_p (src)) |
1992 | { |
1993 | /* Make sure neither succ nor succ2 contains a volatile reference. */ |
1994 | if (succ2 != 0 && volatile_refs_p (PATTERN (insn: succ2))) |
1995 | return false; |
1996 | if (succ != 0 && volatile_refs_p (PATTERN (insn: succ))) |
1997 | return false; |
1998 | /* We'll check insns between INSN and I3 below. */ |
1999 | } |
2000 | |
2001 | /* If INSN is an asm, and DEST is a hard register, reject, since it has |
2002 | to be an explicit register variable, and was chosen for a reason. */ |
2003 | |
2004 | if (GET_CODE (src) == ASM_OPERANDS |
2005 | && REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER) |
2006 | return false; |
2007 | |
2008 | /* If INSN contains volatile references (specifically volatile MEMs), |
2009 | we cannot combine across any other volatile references. |
2010 | Even if INSN doesn't contain volatile references, any intervening |
2011 | volatile insn might affect machine state. */ |
2012 | |
2013 | is_volatile_p = volatile_refs_p (PATTERN (insn)) |
2014 | ? volatile_refs_p |
2015 | : volatile_insn_p; |
2016 | |
2017 | for (p = NEXT_INSN (insn); p != i3; p = NEXT_INSN (insn: p)) |
2018 | if (INSN_P (p) && p != succ && p != succ2 && is_volatile_p (PATTERN (insn: p))) |
2019 | return false; |
2020 | |
2021 | /* If INSN contains an autoincrement or autodecrement, make sure that |
2022 | register is not used between there and I3, and not already used in |
2023 | I3 either. Neither must it be used in PRED or SUCC, if they exist. |
2024 | Also insist that I3 not be a jump if using LRA; if it were one |
2025 | and the incremented register were spilled, we would lose. |
2026 | Reload handles this correctly. */ |
2027 | |
2028 | if (AUTO_INC_DEC) |
2029 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) |
2030 | if (REG_NOTE_KIND (link) == REG_INC |
2031 | && ((JUMP_P (i3) && targetm.lra_p ()) |
2032 | || reg_used_between_p (XEXP (link, 0), insn, i3) |
2033 | || (pred != NULL_RTX |
2034 | && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (insn: pred))) |
2035 | || (pred2 != NULL_RTX |
2036 | && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (insn: pred2))) |
2037 | || (succ != NULL_RTX |
2038 | && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (insn: succ))) |
2039 | || (succ2 != NULL_RTX |
2040 | && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (insn: succ2))) |
2041 | || reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (insn: i3)))) |
2042 | return false; |
2043 | |
2044 | /* If we get here, we have passed all the tests and the combination is |
2045 | to be allowed. */ |
2046 | |
2047 | *pdest = dest; |
2048 | *psrc = src; |
2049 | |
2050 | return true; |
2051 | } |
2052 | |
2053 | /* LOC is the location within I3 that contains its pattern or the component |
2054 | of a PARALLEL of the pattern. We validate that it is valid for combining. |
2055 | |
2056 | One problem is if I3 modifies its output, as opposed to replacing it |
2057 | entirely, we can't allow the output to contain I2DEST, I1DEST or I0DEST as |
2058 | doing so would produce an insn that is not equivalent to the original insns. |
2059 | |
2060 | Consider: |
2061 | |
2062 | (set (reg:DI 101) (reg:DI 100)) |
2063 | (set (subreg:SI (reg:DI 101) 0) <foo>) |
2064 | |
2065 | This is NOT equivalent to: |
2066 | |
2067 | (parallel [(set (subreg:SI (reg:DI 100) 0) <foo>) |
2068 | (set (reg:DI 101) (reg:DI 100))]) |
2069 | |
2070 | Not only does this modify 100 (in which case it might still be valid |
2071 | if 100 were dead in I2), it sets 101 to the ORIGINAL value of 100. |
2072 | |
2073 | We can also run into a problem if I2 sets a register that I1 |
2074 | uses and I1 gets directly substituted into I3 (not via I2). In that |
2075 | case, we would be getting the wrong value of I2DEST into I3, so we |
2076 | must reject the combination. This case occurs when I2 and I1 both |
2077 | feed into I3, rather than when I1 feeds into I2, which feeds into I3. |
2078 | If I1_NOT_IN_SRC is nonzero, it means that finding I1 in the source |
2079 | of a SET must prevent combination from occurring. The same situation |
2080 | can occur for I0, in which case I0_NOT_IN_SRC is set. |
2081 | |
2082 | Before doing the above check, we first try to expand a field assignment |
2083 | into a set of logical operations. |
2084 | |
2085 | If PI3_DEST_KILLED is nonzero, it is a pointer to a location in which |
2086 | we place a register that is both set and used within I3. If more than one |
2087 | such register is detected, we fail. |
2088 | |
2089 | Return true if the combination is valid, false otherwise. */ |
2090 | |
2091 | static bool |
2092 | combinable_i3pat (rtx_insn *i3, rtx *loc, rtx i2dest, rtx i1dest, rtx i0dest, |
2093 | bool i1_not_in_src, bool i0_not_in_src, rtx *pi3dest_killed) |
2094 | { |
2095 | rtx x = *loc; |
2096 | |
2097 | if (GET_CODE (x) == SET) |
2098 | { |
2099 | rtx set = x ; |
2100 | rtx dest = SET_DEST (set); |
2101 | rtx src = SET_SRC (set); |
2102 | rtx inner_dest = dest; |
2103 | rtx subdest; |
2104 | |
2105 | while (GET_CODE (inner_dest) == STRICT_LOW_PART |
2106 | || GET_CODE (inner_dest) == SUBREG |
2107 | || GET_CODE (inner_dest) == ZERO_EXTRACT) |
2108 | inner_dest = XEXP (inner_dest, 0); |
2109 | |
2110 | /* Check for the case where I3 modifies its output, as discussed |
2111 | above. We don't want to prevent pseudos from being combined |
2112 | into the address of a MEM, so only prevent the combination if |
2113 | i1 or i2 set the same MEM. */ |
2114 | if ((inner_dest != dest && |
2115 | (!MEM_P (inner_dest) |
2116 | || rtx_equal_p (i2dest, inner_dest) |
2117 | || (i1dest && rtx_equal_p (i1dest, inner_dest)) |
2118 | || (i0dest && rtx_equal_p (i0dest, inner_dest))) |
2119 | && (reg_overlap_mentioned_p (i2dest, inner_dest) |
2120 | || (i1dest && reg_overlap_mentioned_p (i1dest, inner_dest)) |
2121 | || (i0dest && reg_overlap_mentioned_p (i0dest, inner_dest)))) |
2122 | |
2123 | /* This is the same test done in can_combine_p except we can't test |
2124 | all_adjacent; we don't have to, since this instruction will stay |
2125 | in place, thus we are not considering increasing the lifetime of |
2126 | INNER_DEST. |
2127 | |
2128 | Also, if this insn sets a function argument, combining it with |
2129 | something that might need a spill could clobber a previous |
2130 | function argument; the all_adjacent test in can_combine_p also |
2131 | checks this; here, we do a more specific test for this case. */ |
2132 | |
2133 | || (REG_P (inner_dest) |
2134 | && REGNO (inner_dest) < FIRST_PSEUDO_REGISTER |
2135 | && !targetm.hard_regno_mode_ok (REGNO (inner_dest), |
2136 | GET_MODE (inner_dest))) |
2137 | || (i1_not_in_src && reg_overlap_mentioned_p (i1dest, src)) |
2138 | || (i0_not_in_src && reg_overlap_mentioned_p (i0dest, src))) |
2139 | return false; |
2140 | |
2141 | /* If DEST is used in I3, it is being killed in this insn, so |
2142 | record that for later. We have to consider paradoxical |
2143 | subregs here, since they kill the whole register, but we |
2144 | ignore partial subregs, STRICT_LOW_PART, etc. |
2145 | Never add REG_DEAD notes for the FRAME_POINTER_REGNUM or the |
2146 | STACK_POINTER_REGNUM, since these are always considered to be |
2147 | live. Similarly for ARG_POINTER_REGNUM if it is fixed. */ |
2148 | subdest = dest; |
2149 | if (GET_CODE (subdest) == SUBREG && !partial_subreg_p (x: subdest)) |
2150 | subdest = SUBREG_REG (subdest); |
2151 | if (pi3dest_killed |
2152 | && REG_P (subdest) |
2153 | && reg_referenced_p (subdest, PATTERN (insn: i3)) |
2154 | && REGNO (subdest) != FRAME_POINTER_REGNUM |
2155 | && (HARD_FRAME_POINTER_IS_FRAME_POINTER |
2156 | || REGNO (subdest) != HARD_FRAME_POINTER_REGNUM) |
2157 | && (FRAME_POINTER_REGNUM == ARG_POINTER_REGNUM |
2158 | || (REGNO (subdest) != ARG_POINTER_REGNUM |
2159 | || ! fixed_regs [REGNO (subdest)])) |
2160 | && REGNO (subdest) != STACK_POINTER_REGNUM) |
2161 | { |
2162 | if (*pi3dest_killed) |
2163 | return false; |
2164 | |
2165 | *pi3dest_killed = subdest; |
2166 | } |
2167 | } |
2168 | |
2169 | else if (GET_CODE (x) == PARALLEL) |
2170 | { |
2171 | int i; |
2172 | |
2173 | for (i = 0; i < XVECLEN (x, 0); i++) |
2174 | if (! combinable_i3pat (i3, loc: &XVECEXP (x, 0, i), i2dest, i1dest, i0dest, |
2175 | i1_not_in_src, i0_not_in_src, pi3dest_killed)) |
2176 | return false; |
2177 | } |
2178 | |
2179 | return true; |
2180 | } |
2181 | |
2182 | /* Return true if X is an arithmetic expression that contains a multiplication |
2183 | and division. We don't count multiplications by powers of two here. */ |
2184 | |
2185 | static bool |
2186 | contains_muldiv (rtx x) |
2187 | { |
2188 | switch (GET_CODE (x)) |
2189 | { |
2190 | case MOD: case DIV: case UMOD: case UDIV: |
2191 | return true; |
2192 | |
2193 | case MULT: |
2194 | return ! (CONST_INT_P (XEXP (x, 1)) |
2195 | && pow2p_hwi (UINTVAL (XEXP (x, 1)))); |
2196 | default: |
2197 | if (BINARY_P (x)) |
2198 | return contains_muldiv (XEXP (x, 0)) |
2199 | || contains_muldiv (XEXP (x, 1)); |
2200 | |
2201 | if (UNARY_P (x)) |
2202 | return contains_muldiv (XEXP (x, 0)); |
2203 | |
2204 | return false; |
2205 | } |
2206 | } |
2207 | |
2208 | /* Determine whether INSN can be used in a combination. Return true if |
2209 | not. This is used in try_combine to detect early some cases where we |
2210 | can't perform combinations. */ |
2211 | |
2212 | static bool |
2213 | cant_combine_insn_p (rtx_insn *insn) |
2214 | { |
2215 | rtx set; |
2216 | rtx src, dest; |
2217 | |
2218 | /* If this isn't really an insn, we can't do anything. |
2219 | This can occur when flow deletes an insn that it has merged into an |
2220 | auto-increment address. */ |
2221 | if (!NONDEBUG_INSN_P (insn)) |
2222 | return true; |
2223 | |
2224 | /* Never combine loads and stores involving hard regs that are likely |
2225 | to be spilled. The register allocator can usually handle such |
2226 | reg-reg moves by tying. If we allow the combiner to make |
2227 | substitutions of likely-spilled regs, reload might die. |
2228 | As an exception, we allow combinations involving fixed regs; these are |
2229 | not available to the register allocator so there's no risk involved. */ |
2230 | |
2231 | set = single_set (insn); |
2232 | if (! set) |
2233 | return false; |
2234 | src = SET_SRC (set); |
2235 | dest = SET_DEST (set); |
2236 | if (GET_CODE (src) == SUBREG) |
2237 | src = SUBREG_REG (src); |
2238 | if (GET_CODE (dest) == SUBREG) |
2239 | dest = SUBREG_REG (dest); |
2240 | if (REG_P (src) && REG_P (dest) |
2241 | && ((HARD_REGISTER_P (src) |
2242 | && ! TEST_HARD_REG_BIT (fixed_reg_set, REGNO (src)) |
2243 | #ifdef LEAF_REGISTERS |
2244 | && ! LEAF_REGISTERS [REGNO (src)]) |
2245 | #else |
2246 | ) |
2247 | #endif |
2248 | || (HARD_REGISTER_P (dest) |
2249 | && ! TEST_HARD_REG_BIT (fixed_reg_set, REGNO (dest)) |
2250 | && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (dest)))))) |
2251 | return true; |
2252 | |
2253 | return false; |
2254 | } |
2255 | |
2256 | struct likely_spilled_retval_info |
2257 | { |
2258 | unsigned regno, nregs; |
2259 | unsigned mask; |
2260 | }; |
2261 | |
2262 | /* Called via note_stores by likely_spilled_retval_p. Remove from info->mask |
2263 | hard registers that are known to be written to / clobbered in full. */ |
2264 | static void |
2265 | likely_spilled_retval_1 (rtx x, const_rtx set, void *data) |
2266 | { |
2267 | struct likely_spilled_retval_info *const info = |
2268 | (struct likely_spilled_retval_info *) data; |
2269 | unsigned regno, nregs; |
2270 | unsigned new_mask; |
2271 | |
2272 | if (!REG_P (XEXP (set, 0))) |
2273 | return; |
2274 | regno = REGNO (x); |
2275 | if (regno >= info->regno + info->nregs) |
2276 | return; |
2277 | nregs = REG_NREGS (x); |
2278 | if (regno + nregs <= info->regno) |
2279 | return; |
2280 | new_mask = (2U << (nregs - 1)) - 1; |
2281 | if (regno < info->regno) |
2282 | new_mask >>= info->regno - regno; |
2283 | else |
2284 | new_mask <<= regno - info->regno; |
2285 | info->mask &= ~new_mask; |
2286 | } |
2287 | |
2288 | /* Return true iff part of the return value is live during INSN, and |
2289 | it is likely spilled. This can happen when more than one insn is needed |
2290 | to copy the return value, e.g. when we consider to combine into the |
2291 | second copy insn for a complex value. */ |
2292 | |
2293 | static bool |
2294 | likely_spilled_retval_p (rtx_insn *insn) |
2295 | { |
2296 | rtx_insn *use = BB_END (this_basic_block); |
2297 | rtx reg; |
2298 | rtx_insn *p; |
2299 | unsigned regno, nregs; |
2300 | /* We assume here that no machine mode needs more than |
2301 | 32 hard registers when the value overlaps with a register |
2302 | for which TARGET_FUNCTION_VALUE_REGNO_P is true. */ |
2303 | unsigned mask; |
2304 | struct likely_spilled_retval_info info; |
2305 | |
2306 | if (!NONJUMP_INSN_P (use) || GET_CODE (PATTERN (use)) != USE || insn == use) |
2307 | return false; |
2308 | reg = XEXP (PATTERN (use), 0); |
2309 | if (!REG_P (reg) || !targetm.calls.function_value_regno_p (REGNO (reg))) |
2310 | return false; |
2311 | regno = REGNO (reg); |
2312 | nregs = REG_NREGS (reg); |
2313 | if (nregs == 1) |
2314 | return false; |
2315 | mask = (2U << (nregs - 1)) - 1; |
2316 | |
2317 | /* Disregard parts of the return value that are set later. */ |
2318 | info.regno = regno; |
2319 | info.nregs = nregs; |
2320 | info.mask = mask; |
2321 | for (p = PREV_INSN (insn: use); info.mask && p != insn; p = PREV_INSN (insn: p)) |
2322 | if (INSN_P (p)) |
2323 | note_stores (p, likely_spilled_retval_1, &info); |
2324 | mask = info.mask; |
2325 | |
2326 | /* Check if any of the (probably) live return value registers is |
2327 | likely spilled. */ |
2328 | nregs --; |
2329 | do |
2330 | { |
2331 | if ((mask & 1 << nregs) |
2332 | && targetm.class_likely_spilled_p (REGNO_REG_CLASS (regno + nregs))) |
2333 | return true; |
2334 | } while (nregs--); |
2335 | return false; |
2336 | } |
2337 | |
2338 | /* Adjust INSN after we made a change to its destination. |
2339 | |
2340 | Changing the destination can invalidate notes that say something about |
2341 | the results of the insn and a LOG_LINK pointing to the insn. */ |
2342 | |
2343 | static void |
2344 | adjust_for_new_dest (rtx_insn *insn) |
2345 | { |
2346 | /* For notes, be conservative and simply remove them. */ |
2347 | remove_reg_equal_equiv_notes (insn, true); |
2348 | |
2349 | /* The new insn will have a destination that was previously the destination |
2350 | of an insn just above it. Call distribute_links to make a LOG_LINK from |
2351 | the next use of that destination. */ |
2352 | |
2353 | rtx set = single_set (insn); |
2354 | gcc_assert (set); |
2355 | |
2356 | rtx reg = SET_DEST (set); |
2357 | |
2358 | while (GET_CODE (reg) == ZERO_EXTRACT |
2359 | || GET_CODE (reg) == STRICT_LOW_PART |
2360 | || GET_CODE (reg) == SUBREG) |
2361 | reg = XEXP (reg, 0); |
2362 | gcc_assert (REG_P (reg)); |
2363 | |
2364 | distribute_links (alloc_insn_link (insn, REGNO (reg), NULL)); |
2365 | |
2366 | df_insn_rescan (insn); |
2367 | } |
2368 | |
2369 | /* Return TRUE if combine can reuse reg X in mode MODE. |
2370 | ADDED_SETS is trueif the original set is still required. */ |
2371 | static bool |
2372 | can_change_dest_mode (rtx x, bool added_sets, machine_mode mode) |
2373 | { |
2374 | unsigned int regno; |
2375 | |
2376 | if (!REG_P (x)) |
2377 | return false; |
2378 | |
2379 | /* Don't change between modes with different underlying register sizes, |
2380 | since this could lead to invalid subregs. */ |
2381 | if (maybe_ne (REGMODE_NATURAL_SIZE (mode), |
2382 | REGMODE_NATURAL_SIZE (GET_MODE (x)))) |
2383 | return false; |
2384 | |
2385 | regno = REGNO (x); |
2386 | /* Allow hard registers if the new mode is legal, and occupies no more |
2387 | registers than the old mode. */ |
2388 | if (regno < FIRST_PSEUDO_REGISTER) |
2389 | return (targetm.hard_regno_mode_ok (regno, mode) |
2390 | && REG_NREGS (x) >= hard_regno_nregs (regno, mode)); |
2391 | |
2392 | /* Or a pseudo that is only used once. */ |
2393 | return (regno < reg_n_sets_max |
2394 | && REG_N_SETS (regno) == 1 |
2395 | && !added_sets |
2396 | && !REG_USERVAR_P (x)); |
2397 | } |
2398 | |
2399 | |
2400 | /* Check whether X, the destination of a set, refers to part of |
2401 | the register specified by REG. */ |
2402 | |
2403 | static bool |
2404 | reg_subword_p (rtx x, rtx reg) |
2405 | { |
2406 | /* Check that reg is an integer mode register. */ |
2407 | if (!REG_P (reg) || GET_MODE_CLASS (GET_MODE (reg)) != MODE_INT) |
2408 | return false; |
2409 | |
2410 | if (GET_CODE (x) == STRICT_LOW_PART |
2411 | || GET_CODE (x) == ZERO_EXTRACT) |
2412 | x = XEXP (x, 0); |
2413 | |
2414 | return GET_CODE (x) == SUBREG |
2415 | && !paradoxical_subreg_p (x) |
2416 | && SUBREG_REG (x) == reg |
2417 | && GET_MODE_CLASS (GET_MODE (x)) == MODE_INT; |
2418 | } |
2419 | |
2420 | /* Return whether PAT is a PARALLEL of exactly N register SETs followed |
2421 | by an arbitrary number of CLOBBERs. */ |
2422 | static bool |
2423 | is_parallel_of_n_reg_sets (rtx pat, int n) |
2424 | { |
2425 | if (GET_CODE (pat) != PARALLEL) |
2426 | return false; |
2427 | |
2428 | int len = XVECLEN (pat, 0); |
2429 | if (len < n) |
2430 | return false; |
2431 | |
2432 | int i; |
2433 | for (i = 0; i < n; i++) |
2434 | if (GET_CODE (XVECEXP (pat, 0, i)) != SET |
2435 | || !REG_P (SET_DEST (XVECEXP (pat, 0, i)))) |
2436 | return false; |
2437 | for ( ; i < len; i++) |
2438 | switch (GET_CODE (XVECEXP (pat, 0, i))) |
2439 | { |
2440 | case CLOBBER: |
2441 | if (XEXP (XVECEXP (pat, 0, i), 0) == const0_rtx) |
2442 | return false; |
2443 | break; |
2444 | default: |
2445 | return false; |
2446 | } |
2447 | return true; |
2448 | } |
2449 | |
2450 | /* Return whether INSN, a PARALLEL of N register SETs (and maybe some |
2451 | CLOBBERs), can be split into individual SETs in that order, without |
2452 | changing semantics. */ |
2453 | static bool |
2454 | can_split_parallel_of_n_reg_sets (rtx_insn *insn, int n) |
2455 | { |
2456 | if (!insn_nothrow_p (insn)) |
2457 | return false; |
2458 | |
2459 | rtx pat = PATTERN (insn); |
2460 | |
2461 | int i, j; |
2462 | for (i = 0; i < n; i++) |
2463 | { |
2464 | if (side_effects_p (SET_SRC (XVECEXP (pat, 0, i)))) |
2465 | return false; |
2466 | |
2467 | rtx reg = SET_DEST (XVECEXP (pat, 0, i)); |
2468 | |
2469 | for (j = i + 1; j < n; j++) |
2470 | if (reg_referenced_p (reg, XVECEXP (pat, 0, j))) |
2471 | return false; |
2472 | } |
2473 | |
2474 | return true; |
2475 | } |
2476 | |
2477 | /* Return whether X is just a single_set, with the source |
2478 | a general_operand. */ |
2479 | static bool |
2480 | is_just_move (rtx_insn *x) |
2481 | { |
2482 | rtx set = single_set (insn: x); |
2483 | if (!set) |
2484 | return false; |
2485 | |
2486 | return general_operand (SET_SRC (set), VOIDmode); |
2487 | } |
2488 | |
2489 | /* Callback function to count autoincs. */ |
2490 | |
2491 | static int |
2492 | count_auto_inc (rtx, rtx, rtx, rtx, rtx, void *arg) |
2493 | { |
2494 | (*((int *) arg))++; |
2495 | |
2496 | return 0; |
2497 | } |
2498 | |
2499 | /* Try to combine the insns I0, I1 and I2 into I3. |
2500 | Here I0, I1 and I2 appear earlier than I3. |
2501 | I0 and I1 can be zero; then we combine just I2 into I3, or I1 and I2 into |
2502 | I3. |
2503 | |
2504 | If we are combining more than two insns and the resulting insn is not |
2505 | recognized, try splitting it into two insns. If that happens, I2 and I3 |
2506 | are retained and I1/I0 are pseudo-deleted by turning them into a NOTE. |
2507 | Otherwise, I0, I1 and I2 are pseudo-deleted. |
2508 | |
2509 | Return 0 if the combination does not work. Then nothing is changed. |
2510 | If we did the combination, return the insn at which combine should |
2511 | resume scanning. |
2512 | |
2513 | Set NEW_DIRECT_JUMP_P to true if try_combine creates a |
2514 | new direct jump instruction. |
2515 | |
2516 | LAST_COMBINED_INSN is either I3, or some insn after I3 that has |
2517 | been I3 passed to an earlier try_combine within the same basic |
2518 | block. */ |
2519 | |
2520 | static rtx_insn * |
2521 | try_combine (rtx_insn *i3, rtx_insn *i2, rtx_insn *i1, rtx_insn *i0, |
2522 | bool *new_direct_jump_p, rtx_insn *last_combined_insn) |
2523 | { |
2524 | /* New patterns for I3 and I2, respectively. */ |
2525 | rtx newpat, newi2pat = 0; |
2526 | rtvec newpat_vec_with_clobbers = 0; |
2527 | bool substed_i2 = false, substed_i1 = false, substed_i0 = false; |
2528 | /* Indicates need to preserve SET in I0, I1 or I2 in I3 if it is not |
2529 | dead. */ |
2530 | bool added_sets_0, added_sets_1, added_sets_2; |
2531 | /* Total number of SETs to put into I3. */ |
2532 | int total_sets; |
2533 | /* Nonzero if I2's or I1's body now appears in I3. */ |
2534 | int i2_is_used = 0, i1_is_used = 0; |
2535 | /* INSN_CODEs for new I3, new I2, and user of condition code. */ |
2536 | int insn_code_number, i2_code_number = 0, other_code_number = 0; |
2537 | /* Contains I3 if the destination of I3 is used in its source, which means |
2538 | that the old life of I3 is being killed. If that usage is placed into |
2539 | I2 and not in I3, a REG_DEAD note must be made. */ |
2540 | rtx i3dest_killed = 0; |
2541 | /* SET_DEST and SET_SRC of I2, I1 and I0. */ |
2542 | rtx i2dest = 0, i2src = 0, i1dest = 0, i1src = 0, i0dest = 0, i0src = 0; |
2543 | /* Copy of SET_SRC of I1 and I0, if needed. */ |
2544 | rtx i1src_copy = 0, i0src_copy = 0, i0src_copy2 = 0; |
2545 | /* Set if I2DEST was reused as a scratch register. */ |
2546 | bool i2scratch = false; |
2547 | /* The PATTERNs of I0, I1, and I2, or a copy of them in certain cases. */ |
2548 | rtx i0pat = 0, i1pat = 0, i2pat = 0; |
2549 | /* Indicates if I2DEST or I1DEST is in I2SRC or I1_SRC. */ |
2550 | bool i2dest_in_i2src = false, i1dest_in_i1src = false; |
2551 | bool i2dest_in_i1src = false, i0dest_in_i0src = false; |
2552 | bool i1dest_in_i0src = false, i2dest_in_i0src = false;; |
2553 | bool i2dest_killed = false, i1dest_killed = false, i0dest_killed = false; |
2554 | bool i1_feeds_i2_n = false, i0_feeds_i2_n = false, i0_feeds_i1_n = false; |
2555 | /* Notes that must be added to REG_NOTES in I3 and I2. */ |
2556 | rtx new_i3_notes, new_i2_notes; |
2557 | /* Notes that we substituted I3 into I2 instead of the normal case. */ |
2558 | bool i3_subst_into_i2 = false; |
2559 | /* Notes that I1, I2 or I3 is a MULT operation. */ |
2560 | bool have_mult = false; |
2561 | bool swap_i2i3 = false; |
2562 | bool split_i2i3 = false; |
2563 | bool changed_i3_dest = false; |
2564 | bool i2_was_move = false, i3_was_move = false; |
2565 | int n_auto_inc = 0; |
2566 | |
2567 | int maxreg; |
2568 | rtx_insn *temp_insn; |
2569 | rtx temp_expr; |
2570 | struct insn_link *link; |
2571 | rtx other_pat = 0; |
2572 | rtx new_other_notes; |
2573 | int i; |
2574 | scalar_int_mode dest_mode, temp_mode; |
2575 | bool has_non_call_exception = false; |
2576 | |
2577 | /* Immediately return if any of I0,I1,I2 are the same insn (I3 can |
2578 | never be). */ |
2579 | if (i1 == i2 || i0 == i2 || (i0 && i0 == i1)) |
2580 | return 0; |
2581 | |
2582 | /* Only try four-insn combinations when there's high likelihood of |
2583 | success. Look for simple insns, such as loads of constants or |
2584 | binary operations involving a constant. */ |
2585 | if (i0) |
2586 | { |
2587 | int i; |
2588 | int ngood = 0; |
2589 | int nshift = 0; |
2590 | rtx set0, set3; |
2591 | |
2592 | if (!flag_expensive_optimizations) |
2593 | return 0; |
2594 | |
2595 | for (i = 0; i < 4; i++) |
2596 | { |
2597 | rtx_insn *insn = i == 0 ? i0 : i == 1 ? i1 : i == 2 ? i2 : i3; |
2598 | rtx set = single_set (insn); |
2599 | rtx src; |
2600 | if (!set) |
2601 | continue; |
2602 | src = SET_SRC (set); |
2603 | if (CONSTANT_P (src)) |
2604 | { |
2605 | ngood += 2; |
2606 | break; |
2607 | } |
2608 | else if (BINARY_P (src) && CONSTANT_P (XEXP (src, 1))) |
2609 | ngood++; |
2610 | else if (GET_CODE (src) == ASHIFT || GET_CODE (src) == ASHIFTRT |
2611 | || GET_CODE (src) == LSHIFTRT) |
2612 | nshift++; |
2613 | } |
2614 | |
2615 | /* If I0 loads a memory and I3 sets the same memory, then I1 and I2 |
2616 | are likely manipulating its value. Ideally we'll be able to combine |
2617 | all four insns into a bitfield insertion of some kind. |
2618 | |
2619 | Note the source in I0 might be inside a sign/zero extension and the |
2620 | memory modes in I0 and I3 might be different. So extract the address |
2621 | from the destination of I3 and search for it in the source of I0. |
2622 | |
2623 | In the event that there's a match but the source/dest do not actually |
2624 | refer to the same memory, the worst that happens is we try some |
2625 | combinations that we wouldn't have otherwise. */ |
2626 | if ((set0 = single_set (insn: i0)) |
2627 | /* Ensure the source of SET0 is a MEM, possibly buried inside |
2628 | an extension. */ |
2629 | && (GET_CODE (SET_SRC (set0)) == MEM |
2630 | || ((GET_CODE (SET_SRC (set0)) == ZERO_EXTEND |
2631 | || GET_CODE (SET_SRC (set0)) == SIGN_EXTEND) |
2632 | && GET_CODE (XEXP (SET_SRC (set0), 0)) == MEM)) |
2633 | && (set3 = single_set (insn: i3)) |
2634 | /* Ensure the destination of SET3 is a MEM. */ |
2635 | && GET_CODE (SET_DEST (set3)) == MEM |
2636 | /* Would it be better to extract the base address for the MEM |
2637 | in SET3 and look for that? I don't have cases where it matters |
2638 | but I could envision such cases. */ |
2639 | && rtx_referenced_p (XEXP (SET_DEST (set3), 0), SET_SRC (set0))) |
2640 | ngood += 2; |
2641 | |
2642 | if (ngood < 2 && nshift < 2) |
2643 | return 0; |
2644 | } |
2645 | |
2646 | /* Exit early if one of the insns involved can't be used for |
2647 | combinations. */ |
2648 | if (CALL_P (i2) |
2649 | || (i1 && CALL_P (i1)) |
2650 | || (i0 && CALL_P (i0)) |
2651 | || cant_combine_insn_p (insn: i3) |
2652 | || cant_combine_insn_p (insn: i2) |
2653 | || (i1 && cant_combine_insn_p (insn: i1)) |
2654 | || (i0 && cant_combine_insn_p (insn: i0)) |
2655 | || likely_spilled_retval_p (insn: i3)) |
2656 | return 0; |
2657 | |
2658 | combine_attempts++; |
2659 | undobuf.other_insn = 0; |
2660 | |
2661 | /* Reset the hard register usage information. */ |
2662 | CLEAR_HARD_REG_SET (set&: newpat_used_regs); |
2663 | |
2664 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2665 | { |
2666 | if (i0) |
2667 | fprintf (stream: dump_file, format: "\nTrying %d, %d, %d -> %d:\n" , |
2668 | INSN_UID (insn: i0), INSN_UID (insn: i1), INSN_UID (insn: i2), INSN_UID (insn: i3)); |
2669 | else if (i1) |
2670 | fprintf (stream: dump_file, format: "\nTrying %d, %d -> %d:\n" , |
2671 | INSN_UID (insn: i1), INSN_UID (insn: i2), INSN_UID (insn: i3)); |
2672 | else |
2673 | fprintf (stream: dump_file, format: "\nTrying %d -> %d:\n" , |
2674 | INSN_UID (insn: i2), INSN_UID (insn: i3)); |
2675 | |
2676 | if (i0) |
2677 | dump_insn_slim (dump_file, i0); |
2678 | if (i1) |
2679 | dump_insn_slim (dump_file, i1); |
2680 | dump_insn_slim (dump_file, i2); |
2681 | dump_insn_slim (dump_file, i3); |
2682 | } |
2683 | |
2684 | /* If multiple insns feed into one of I2 or I3, they can be in any |
2685 | order. To simplify the code below, reorder them in sequence. */ |
2686 | if (i0 && DF_INSN_LUID (i0) > DF_INSN_LUID (i2)) |
2687 | std::swap (a&: i0, b&: i2); |
2688 | if (i0 && DF_INSN_LUID (i0) > DF_INSN_LUID (i1)) |
2689 | std::swap (a&: i0, b&: i1); |
2690 | if (i1 && DF_INSN_LUID (i1) > DF_INSN_LUID (i2)) |
2691 | std::swap (a&: i1, b&: i2); |
2692 | |
2693 | added_links_insn = 0; |
2694 | added_notes_insn = 0; |
2695 | |
2696 | /* First check for one important special case that the code below will |
2697 | not handle. Namely, the case where I1 is zero, I2 is a PARALLEL |
2698 | and I3 is a SET whose SET_SRC is a SET_DEST in I2. In that case, |
2699 | we may be able to replace that destination with the destination of I3. |
2700 | This occurs in the common code where we compute both a quotient and |
2701 | remainder into a structure, in which case we want to do the computation |
2702 | directly into the structure to avoid register-register copies. |
2703 | |
2704 | Note that this case handles both multiple sets in I2 and also cases |
2705 | where I2 has a number of CLOBBERs inside the PARALLEL. |
2706 | |
2707 | We make very conservative checks below and only try to handle the |
2708 | most common cases of this. For example, we only handle the case |
2709 | where I2 and I3 are adjacent to avoid making difficult register |
2710 | usage tests. */ |
2711 | |
2712 | if (i1 == 0 && NONJUMP_INSN_P (i3) && GET_CODE (PATTERN (i3)) == SET |
2713 | && REG_P (SET_SRC (PATTERN (i3))) |
2714 | && REGNO (SET_SRC (PATTERN (i3))) >= FIRST_PSEUDO_REGISTER |
2715 | && find_reg_note (i3, REG_DEAD, SET_SRC (PATTERN (i3))) |
2716 | && GET_CODE (PATTERN (i2)) == PARALLEL |
2717 | && ! side_effects_p (SET_DEST (PATTERN (i3))) |
2718 | /* If the dest of I3 is a ZERO_EXTRACT or STRICT_LOW_PART, the code |
2719 | below would need to check what is inside (and reg_overlap_mentioned_p |
2720 | doesn't support those codes anyway). Don't allow those destinations; |
2721 | the resulting insn isn't likely to be recognized anyway. */ |
2722 | && GET_CODE (SET_DEST (PATTERN (i3))) != ZERO_EXTRACT |
2723 | && GET_CODE (SET_DEST (PATTERN (i3))) != STRICT_LOW_PART |
2724 | && ! reg_overlap_mentioned_p (SET_SRC (PATTERN (i3)), |
2725 | SET_DEST (PATTERN (i3))) |
2726 | && next_active_insn (i2) == i3) |
2727 | { |
2728 | rtx p2 = PATTERN (insn: i2); |
2729 | |
2730 | /* Make sure that the destination of I3, |
2731 | which we are going to substitute into one output of I2, |
2732 | is not used within another output of I2. We must avoid making this: |
2733 | (parallel [(set (mem (reg 69)) ...) |
2734 | (set (reg 69) ...)]) |
2735 | which is not well-defined as to order of actions. |
2736 | (Besides, reload can't handle output reloads for this.) |
2737 | |
2738 | The problem can also happen if the dest of I3 is a memory ref, |
2739 | if another dest in I2 is an indirect memory ref. |
2740 | |
2741 | Neither can this PARALLEL be an asm. We do not allow combining |
2742 | that usually (see can_combine_p), so do not here either. */ |
2743 | bool ok = true; |
2744 | for (i = 0; ok && i < XVECLEN (p2, 0); i++) |
2745 | { |
2746 | if ((GET_CODE (XVECEXP (p2, 0, i)) == SET |
2747 | || GET_CODE (XVECEXP (p2, 0, i)) == CLOBBER) |
2748 | && reg_overlap_mentioned_p (SET_DEST (PATTERN (i3)), |
2749 | SET_DEST (XVECEXP (p2, 0, i)))) |
2750 | ok = false; |
2751 | else if (GET_CODE (XVECEXP (p2, 0, i)) == SET |
2752 | && GET_CODE (SET_SRC (XVECEXP (p2, 0, i))) == ASM_OPERANDS) |
2753 | ok = false; |
2754 | } |
2755 | |
2756 | if (ok) |
2757 | for (i = 0; i < XVECLEN (p2, 0); i++) |
2758 | if (GET_CODE (XVECEXP (p2, 0, i)) == SET |
2759 | && SET_DEST (XVECEXP (p2, 0, i)) == SET_SRC (PATTERN (i3))) |
2760 | { |
2761 | combine_merges++; |
2762 | |
2763 | subst_insn = i3; |
2764 | subst_low_luid = DF_INSN_LUID (i2); |
2765 | |
2766 | added_sets_2 = added_sets_1 = added_sets_0 = false; |
2767 | i2src = SET_SRC (XVECEXP (p2, 0, i)); |
2768 | i2dest = SET_DEST (XVECEXP (p2, 0, i)); |
2769 | i2dest_killed = dead_or_set_p (i2, i2dest); |
2770 | |
2771 | /* Replace the dest in I2 with our dest and make the resulting |
2772 | insn the new pattern for I3. Then skip to where we validate |
2773 | the pattern. Everything was set up above. */ |
2774 | SUBST (SET_DEST (XVECEXP (p2, 0, i)), SET_DEST (PATTERN (i3))); |
2775 | newpat = p2; |
2776 | i3_subst_into_i2 = true; |
2777 | goto validate_replacement; |
2778 | } |
2779 | } |
2780 | |
2781 | /* If I2 is setting a pseudo to a constant and I3 is setting some |
2782 | sub-part of it to another constant, merge them by making a new |
2783 | constant. */ |
2784 | if (i1 == 0 |
2785 | && (temp_expr = single_set (insn: i2)) != 0 |
2786 | && is_a <scalar_int_mode> (GET_MODE (SET_DEST (temp_expr)), result: &temp_mode) |
2787 | && CONST_SCALAR_INT_P (SET_SRC (temp_expr)) |
2788 | && GET_CODE (PATTERN (i3)) == SET |
2789 | && CONST_SCALAR_INT_P (SET_SRC (PATTERN (i3))) |
2790 | && reg_subword_p (SET_DEST (PATTERN (i3)), SET_DEST (temp_expr))) |
2791 | { |
2792 | rtx dest = SET_DEST (PATTERN (i3)); |
2793 | rtx temp_dest = SET_DEST (temp_expr); |
2794 | int offset = -1; |
2795 | int width = 0; |
2796 | |
2797 | if (GET_CODE (dest) == ZERO_EXTRACT) |
2798 | { |
2799 | if (CONST_INT_P (XEXP (dest, 1)) |
2800 | && CONST_INT_P (XEXP (dest, 2)) |
2801 | && is_a <scalar_int_mode> (GET_MODE (XEXP (dest, 0)), |
2802 | result: &dest_mode)) |
2803 | { |
2804 | width = INTVAL (XEXP (dest, 1)); |
2805 | offset = INTVAL (XEXP (dest, 2)); |
2806 | dest = XEXP (dest, 0); |
2807 | if (BITS_BIG_ENDIAN) |
2808 | offset = GET_MODE_PRECISION (mode: dest_mode) - width - offset; |
2809 | } |
2810 | } |
2811 | else |
2812 | { |
2813 | if (GET_CODE (dest) == STRICT_LOW_PART) |
2814 | dest = XEXP (dest, 0); |
2815 | if (is_a <scalar_int_mode> (GET_MODE (dest), result: &dest_mode)) |
2816 | { |
2817 | width = GET_MODE_PRECISION (mode: dest_mode); |
2818 | offset = 0; |
2819 | } |
2820 | } |
2821 | |
2822 | if (offset >= 0) |
2823 | { |
2824 | /* If this is the low part, we're done. */ |
2825 | if (subreg_lowpart_p (dest)) |
2826 | ; |
2827 | /* Handle the case where inner is twice the size of outer. */ |
2828 | else if (GET_MODE_PRECISION (mode: temp_mode) |
2829 | == 2 * GET_MODE_PRECISION (mode: dest_mode)) |
2830 | offset += GET_MODE_PRECISION (mode: dest_mode); |
2831 | /* Otherwise give up for now. */ |
2832 | else |
2833 | offset = -1; |
2834 | } |
2835 | |
2836 | if (offset >= 0) |
2837 | { |
2838 | rtx inner = SET_SRC (PATTERN (i3)); |
2839 | rtx outer = SET_SRC (temp_expr); |
2840 | |
2841 | wide_int o = wi::insert (x: rtx_mode_t (outer, temp_mode), |
2842 | y: rtx_mode_t (inner, dest_mode), |
2843 | offset, width); |
2844 | |
2845 | combine_merges++; |
2846 | subst_insn = i3; |
2847 | subst_low_luid = DF_INSN_LUID (i2); |
2848 | added_sets_2 = added_sets_1 = added_sets_0 = false; |
2849 | i2dest = temp_dest; |
2850 | i2dest_killed = dead_or_set_p (i2, i2dest); |
2851 | |
2852 | /* Replace the source in I2 with the new constant and make the |
2853 | resulting insn the new pattern for I3. Then skip to where we |
2854 | validate the pattern. Everything was set up above. */ |
2855 | SUBST (SET_SRC (temp_expr), |
2856 | immed_wide_int_const (o, temp_mode)); |
2857 | |
2858 | newpat = PATTERN (insn: i2); |
2859 | |
2860 | /* The dest of I3 has been replaced with the dest of I2. */ |
2861 | changed_i3_dest = true; |
2862 | goto validate_replacement; |
2863 | } |
2864 | } |
2865 | |
2866 | /* If we have no I1 and I2 looks like: |
2867 | (parallel [(set (reg:CC X) (compare:CC OP (const_int 0))) |
2868 | (set Y OP)]) |
2869 | make up a dummy I1 that is |
2870 | (set Y OP) |
2871 | and change I2 to be |
2872 | (set (reg:CC X) (compare:CC Y (const_int 0))) |
2873 | |
2874 | (We can ignore any trailing CLOBBERs.) |
2875 | |
2876 | This undoes a previous combination and allows us to match a branch-and- |
2877 | decrement insn. */ |
2878 | |
2879 | if (i1 == 0 |
2880 | && is_parallel_of_n_reg_sets (pat: PATTERN (insn: i2), n: 2) |
2881 | && (GET_MODE_CLASS (GET_MODE (SET_DEST (XVECEXP (PATTERN (i2), 0, 0)))) |
2882 | == MODE_CC) |
2883 | && GET_CODE (SET_SRC (XVECEXP (PATTERN (i2), 0, 0))) == COMPARE |
2884 | && XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 1) == const0_rtx |
2885 | && rtx_equal_p (XEXP (SET_SRC (XVECEXP (PATTERN (i2), 0, 0)), 0), |
2886 | SET_SRC (XVECEXP (PATTERN (i2), 0, 1))) |
2887 | && !reg_used_between_p (SET_DEST (XVECEXP (PATTERN (i2), 0, 0)), i2, i3) |
2888 | && !reg_used_between_p (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)), i2, i3)) |
2889 | { |
2890 | /* We make I1 with the same INSN_UID as I2. This gives it |
2891 | the same DF_INSN_LUID for value tracking. Our fake I1 will |
2892 | never appear in the insn stream so giving it the same INSN_UID |
2893 | as I2 will not cause a problem. */ |
2894 | |
2895 | i1 = gen_rtx_INSN (VOIDmode, NULL, next_insn: i2, bb: BLOCK_FOR_INSN (insn: i2), |
2896 | XVECEXP (PATTERN (i2), 0, 1), location: INSN_LOCATION (insn: i2), |
2897 | code: -1, NULL_RTX); |
2898 | INSN_UID (insn: i1) = INSN_UID (insn: i2); |
2899 | |
2900 | SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 0)); |
2901 | SUBST (XEXP (SET_SRC (PATTERN (i2)), 0), |
2902 | SET_DEST (PATTERN (i1))); |
2903 | unsigned int regno = REGNO (SET_DEST (PATTERN (i1))); |
2904 | SUBST_LINK (LOG_LINKS (i2), |
2905 | alloc_insn_link (i1, regno, LOG_LINKS (i2))); |
2906 | } |
2907 | |
2908 | /* If I2 is a PARALLEL of two SETs of REGs (and perhaps some CLOBBERs), |
2909 | make those two SETs separate I1 and I2 insns, and make an I0 that is |
2910 | the original I1. */ |
2911 | if (i0 == 0 |
2912 | && is_parallel_of_n_reg_sets (pat: PATTERN (insn: i2), n: 2) |
2913 | && can_split_parallel_of_n_reg_sets (insn: i2, n: 2) |
2914 | && !reg_used_between_p (SET_DEST (XVECEXP (PATTERN (i2), 0, 0)), i2, i3) |
2915 | && !reg_used_between_p (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)), i2, i3) |
2916 | && !reg_set_between_p (SET_DEST (XVECEXP (PATTERN (i2), 0, 0)), i2, i3) |
2917 | && !reg_set_between_p (SET_DEST (XVECEXP (PATTERN (i2), 0, 1)), i2, i3)) |
2918 | { |
2919 | /* If there is no I1, there is no I0 either. */ |
2920 | i0 = i1; |
2921 | |
2922 | /* We make I1 with the same INSN_UID as I2. This gives it |
2923 | the same DF_INSN_LUID for value tracking. Our fake I1 will |
2924 | never appear in the insn stream so giving it the same INSN_UID |
2925 | as I2 will not cause a problem. */ |
2926 | |
2927 | i1 = gen_rtx_INSN (VOIDmode, NULL, next_insn: i2, bb: BLOCK_FOR_INSN (insn: i2), |
2928 | XVECEXP (PATTERN (i2), 0, 0), location: INSN_LOCATION (insn: i2), |
2929 | code: -1, NULL_RTX); |
2930 | INSN_UID (insn: i1) = INSN_UID (insn: i2); |
2931 | |
2932 | SUBST (PATTERN (i2), XVECEXP (PATTERN (i2), 0, 1)); |
2933 | } |
2934 | |
2935 | /* Verify that I2 and maybe I1 and I0 can be combined into I3. */ |
2936 | if (!can_combine_p (insn: i2, i3, pred: i0, pred2: i1, NULL, NULL, pdest: &i2dest, psrc: &i2src)) |
2937 | { |
2938 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2939 | fprintf (stream: dump_file, format: "Can't combine i2 into i3\n" ); |
2940 | undo_all (); |
2941 | return 0; |
2942 | } |
2943 | if (i1 && !can_combine_p (insn: i1, i3, pred: i0, NULL, succ: i2, NULL, pdest: &i1dest, psrc: &i1src)) |
2944 | { |
2945 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2946 | fprintf (stream: dump_file, format: "Can't combine i1 into i3\n" ); |
2947 | undo_all (); |
2948 | return 0; |
2949 | } |
2950 | if (i0 && !can_combine_p (insn: i0, i3, NULL, NULL, succ: i1, succ2: i2, pdest: &i0dest, psrc: &i0src)) |
2951 | { |
2952 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2953 | fprintf (stream: dump_file, format: "Can't combine i0 into i3\n" ); |
2954 | undo_all (); |
2955 | return 0; |
2956 | } |
2957 | |
2958 | /* With non-call exceptions we can end up trying to combine multiple |
2959 | insns with possible EH side effects. Make sure we can combine |
2960 | that to a single insn which means there must be at most one insn |
2961 | in the combination with an EH side effect. */ |
2962 | if (cfun->can_throw_non_call_exceptions) |
2963 | { |
2964 | if (find_reg_note (i3, REG_EH_REGION, NULL_RTX) |
2965 | || find_reg_note (i2, REG_EH_REGION, NULL_RTX) |
2966 | || (i1 && find_reg_note (i1, REG_EH_REGION, NULL_RTX)) |
2967 | || (i0 && find_reg_note (i0, REG_EH_REGION, NULL_RTX))) |
2968 | { |
2969 | has_non_call_exception = true; |
2970 | if (insn_could_throw_p (i3) |
2971 | + insn_could_throw_p (i2) |
2972 | + (i1 ? insn_could_throw_p (i1) : 0) |
2973 | + (i0 ? insn_could_throw_p (i0) : 0) > 1) |
2974 | { |
2975 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2976 | fprintf (stream: dump_file, format: "Can't combine multiple insns with EH " |
2977 | "side-effects\n" ); |
2978 | undo_all (); |
2979 | return 0; |
2980 | } |
2981 | } |
2982 | } |
2983 | |
2984 | /* Record whether i2 and i3 are trivial moves. */ |
2985 | i2_was_move = is_just_move (x: i2); |
2986 | i3_was_move = is_just_move (x: i3); |
2987 | |
2988 | /* Record whether I2DEST is used in I2SRC and similarly for the other |
2989 | cases. Knowing this will help in register status updating below. */ |
2990 | i2dest_in_i2src = reg_overlap_mentioned_p (i2dest, i2src); |
2991 | i1dest_in_i1src = i1 && reg_overlap_mentioned_p (i1dest, i1src); |
2992 | i2dest_in_i1src = i1 && reg_overlap_mentioned_p (i2dest, i1src); |
2993 | i0dest_in_i0src = i0 && reg_overlap_mentioned_p (i0dest, i0src); |
2994 | i1dest_in_i0src = i0 && reg_overlap_mentioned_p (i1dest, i0src); |
2995 | i2dest_in_i0src = i0 && reg_overlap_mentioned_p (i2dest, i0src); |
2996 | i2dest_killed = dead_or_set_p (i2, i2dest); |
2997 | i1dest_killed = i1 && dead_or_set_p (i1, i1dest); |
2998 | i0dest_killed = i0 && dead_or_set_p (i0, i0dest); |
2999 | |
3000 | /* For the earlier insns, determine which of the subsequent ones they |
3001 | feed. */ |
3002 | i1_feeds_i2_n = i1 && insn_a_feeds_b (a: i1, b: i2); |
3003 | i0_feeds_i1_n = i0 && insn_a_feeds_b (a: i0, b: i1); |
3004 | i0_feeds_i2_n = (i0 && (!i0_feeds_i1_n ? insn_a_feeds_b (a: i0, b: i2) |
3005 | : (!reg_overlap_mentioned_p (i1dest, i0dest) |
3006 | && reg_overlap_mentioned_p (i0dest, i2src)))); |
3007 | |
3008 | /* Ensure that I3's pattern can be the destination of combines. */ |
3009 | if (! combinable_i3pat (i3, loc: &PATTERN (insn: i3), i2dest, i1dest, i0dest, |
3010 | i1_not_in_src: i1 && i2dest_in_i1src && !i1_feeds_i2_n, |
3011 | i0_not_in_src: i0 && ((i2dest_in_i0src && !i0_feeds_i2_n) |
3012 | || (i1dest_in_i0src && !i0_feeds_i1_n)), |
3013 | pi3dest_killed: &i3dest_killed)) |
3014 | { |
3015 | undo_all (); |
3016 | return 0; |
3017 | } |
3018 | |
3019 | /* See if any of the insns is a MULT operation. Unless one is, we will |
3020 | reject a combination that is, since it must be slower. Be conservative |
3021 | here. */ |
3022 | if (GET_CODE (i2src) == MULT |
3023 | || (i1 != 0 && GET_CODE (i1src) == MULT) |
3024 | || (i0 != 0 && GET_CODE (i0src) == MULT) |
3025 | || (GET_CODE (PATTERN (i3)) == SET |
3026 | && GET_CODE (SET_SRC (PATTERN (i3))) == MULT)) |
3027 | have_mult = true; |
3028 | |
3029 | /* If I3 has an inc, then give up if I1 or I2 uses the reg that is inc'd. |
3030 | We used to do this EXCEPT in one case: I3 has a post-inc in an |
3031 | output operand. However, that exception can give rise to insns like |
3032 | mov r3,(r3)+ |
3033 | which is a famous insn on the PDP-11 where the value of r3 used as the |
3034 | source was model-dependent. Avoid this sort of thing. */ |
3035 | |
3036 | #if 0 |
3037 | if (!(GET_CODE (PATTERN (i3)) == SET |
3038 | && REG_P (SET_SRC (PATTERN (i3))) |
3039 | && MEM_P (SET_DEST (PATTERN (i3))) |
3040 | && (GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_INC |
3041 | || GET_CODE (XEXP (SET_DEST (PATTERN (i3)), 0)) == POST_DEC))) |
3042 | /* It's not the exception. */ |
3043 | #endif |
3044 | if (AUTO_INC_DEC) |
3045 | { |
3046 | rtx link; |
3047 | for (link = REG_NOTES (i3); link; link = XEXP (link, 1)) |
3048 | if (REG_NOTE_KIND (link) == REG_INC |
3049 | && (reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (insn: i2)) |
3050 | || (i1 != 0 |
3051 | && reg_overlap_mentioned_p (XEXP (link, 0), PATTERN (insn: i1))))) |
3052 | { |
3053 | undo_all (); |
3054 | return 0; |
3055 | } |
3056 | } |
3057 | |
3058 | /* See if the SETs in I1 or I2 need to be kept around in the merged |
3059 | instruction: whenever the value set there is still needed past I3. |
3060 | For the SET in I2, this is easy: we see if I2DEST dies or is set in I3. |
3061 | |
3062 | For the SET in I1, we have two cases: if I1 and I2 independently feed |
3063 | into I3, the set in I1 needs to be kept around unless I1DEST dies |
3064 | or is set in I3. Otherwise (if I1 feeds I2 which feeds I3), the set |
3065 | in I1 needs to be kept around unless I1DEST dies or is set in either |
3066 | I2 or I3. The same considerations apply to I0. */ |
3067 | |
3068 | added_sets_2 = !dead_or_set_p (i3, i2dest); |
3069 | |
3070 | if (i1) |
3071 | added_sets_1 = !(dead_or_set_p (i3, i1dest) |
3072 | || (i1_feeds_i2_n && dead_or_set_p (i2, i1dest))); |
3073 | else |
3074 | added_sets_1 = false; |
3075 | |
3076 | if (i0) |
3077 | added_sets_0 = !(dead_or_set_p (i3, i0dest) |
3078 | || (i0_feeds_i1_n && dead_or_set_p (i1, i0dest)) |
3079 | || ((i0_feeds_i2_n || (i0_feeds_i1_n && i1_feeds_i2_n)) |
3080 | && dead_or_set_p (i2, i0dest))); |
3081 | else |
3082 | added_sets_0 = false; |
3083 | |
3084 | /* We are about to copy insns for the case where they need to be kept |
3085 | around. Check that they can be copied in the merged instruction. */ |
3086 | |
3087 | if (targetm.cannot_copy_insn_p |
3088 | && ((added_sets_2 && targetm.cannot_copy_insn_p (i2)) |
3089 | || (i1 && added_sets_1 && targetm.cannot_copy_insn_p (i1)) |
3090 | || (i0 && added_sets_0 && targetm.cannot_copy_insn_p (i0)))) |
3091 | { |
3092 | undo_all (); |
3093 | return 0; |
3094 | } |
3095 | |
3096 | /* We cannot safely duplicate volatile references in any case. */ |
3097 | |
3098 | if ((added_sets_2 && volatile_refs_p (PATTERN (insn: i2))) |
3099 | || (added_sets_1 && volatile_refs_p (PATTERN (insn: i1))) |
3100 | || (added_sets_0 && volatile_refs_p (PATTERN (insn: i0)))) |
3101 | { |
3102 | undo_all (); |
3103 | return 0; |
3104 | } |
3105 | |
3106 | /* Count how many auto_inc expressions there were in the original insns; |
3107 | we need to have the same number in the resulting patterns. */ |
3108 | |
3109 | if (i0) |
3110 | for_each_inc_dec (PATTERN (insn: i0), count_auto_inc, arg: &n_auto_inc); |
3111 | if (i1) |
3112 | for_each_inc_dec (PATTERN (insn: i1), count_auto_inc, arg: &n_auto_inc); |
3113 | for_each_inc_dec (PATTERN (insn: i2), count_auto_inc, arg: &n_auto_inc); |
3114 | for_each_inc_dec (PATTERN (insn: i3), count_auto_inc, arg: &n_auto_inc); |
3115 | |
3116 | /* If the set in I2 needs to be kept around, we must make a copy of |
3117 | PATTERN (I2), so that when we substitute I1SRC for I1DEST in |
3118 | PATTERN (I2), we are only substituting for the original I1DEST, not into |
3119 | an already-substituted copy. This also prevents making self-referential |
3120 | rtx. If I2 is a PARALLEL, we just need the piece that assigns I2SRC to |
3121 | I2DEST. */ |
3122 | |
3123 | if (added_sets_2) |
3124 | { |
3125 | if (GET_CODE (PATTERN (i2)) == PARALLEL) |
3126 | i2pat = gen_rtx_SET (i2dest, copy_rtx (i2src)); |
3127 | else |
3128 | i2pat = copy_rtx (PATTERN (insn: i2)); |
3129 | } |
3130 | |
3131 | if (added_sets_1) |
3132 | { |
3133 | if (GET_CODE (PATTERN (i1)) == PARALLEL) |
3134 | i1pat = gen_rtx_SET (i1dest, copy_rtx (i1src)); |
3135 | else |
3136 | i1pat = copy_rtx (PATTERN (insn: i1)); |
3137 | } |
3138 | |
3139 | if (added_sets_0) |
3140 | { |
3141 | if (GET_CODE (PATTERN (i0)) == PARALLEL) |
3142 | i0pat = gen_rtx_SET (i0dest, copy_rtx (i0src)); |
3143 | else |
3144 | i0pat = copy_rtx (PATTERN (insn: i0)); |
3145 | } |
3146 | |
3147 | combine_merges++; |
3148 | |
3149 | /* Substitute in the latest insn for the regs set by the earlier ones. */ |
3150 | |
3151 | maxreg = max_reg_num (); |
3152 | |
3153 | subst_insn = i3; |
3154 | |
3155 | /* Many machines have insns that can both perform an |
3156 | arithmetic operation and set the condition code. These operations will |
3157 | be represented as a PARALLEL with the first element of the vector |
3158 | being a COMPARE of an arithmetic operation with the constant zero. |
3159 | The second element of the vector will set some pseudo to the result |
3160 | of the same arithmetic operation. If we simplify the COMPARE, we won't |
3161 | match such a pattern and so will generate an extra insn. Here we test |
3162 | for this case, where both the comparison and the operation result are |
3163 | needed, and make the PARALLEL by just replacing I2DEST in I3SRC with |
3164 | I2SRC. Later we will make the PARALLEL that contains I2. */ |
3165 | |
3166 | if (i1 == 0 && added_sets_2 && GET_CODE (PATTERN (i3)) == SET |
3167 | && GET_CODE (SET_SRC (PATTERN (i3))) == COMPARE |
3168 | && CONST_INT_P (XEXP (SET_SRC (PATTERN (i3)), 1)) |
3169 | && rtx_equal_p (XEXP (SET_SRC (PATTERN (i3)), 0), i2dest)) |
3170 | { |
3171 | rtx newpat_dest; |
3172 | rtx *cc_use_loc = NULL; |
3173 | rtx_insn *cc_use_insn = NULL; |
3174 | rtx op0 = i2src, op1 = XEXP (SET_SRC (PATTERN (i3)), 1); |
3175 | machine_mode compare_mode, orig_compare_mode; |
3176 | enum rtx_code compare_code = UNKNOWN, orig_compare_code = UNKNOWN; |
3177 | scalar_int_mode mode; |
3178 | |
3179 | newpat = PATTERN (insn: i3); |
3180 | newpat_dest = SET_DEST (newpat); |
3181 | compare_mode = orig_compare_mode = GET_MODE (newpat_dest); |
3182 | |
3183 | if (undobuf.other_insn == 0 |
3184 | && (cc_use_loc = find_single_use (SET_DEST (newpat), insn: i3, |
3185 | ploc: &cc_use_insn))) |
3186 | { |
3187 | compare_code = orig_compare_code = GET_CODE (*cc_use_loc); |
3188 | if (is_a <scalar_int_mode> (GET_MODE (i2dest), result: &mode)) |
3189 | compare_code = simplify_compare_const (compare_code, mode, |
3190 | &op0, &op1); |
3191 | target_canonicalize_comparison (code: &compare_code, op0: &op0, op1: &op1, op0_preserve_value: 1); |
3192 | } |
3193 | |
3194 | /* Do the rest only if op1 is const0_rtx, which may be the |
3195 | result of simplification. */ |
3196 | if (op1 == const0_rtx) |
3197 | { |
3198 | /* If a single use of the CC is found, prepare to modify it |
3199 | when SELECT_CC_MODE returns a new CC-class mode, or when |
3200 | the above simplify_compare_const() returned a new comparison |
3201 | operator. undobuf.other_insn is assigned the CC use insn |
3202 | when modifying it. */ |
3203 | if (cc_use_loc) |
3204 | { |
3205 | #ifdef SELECT_CC_MODE |
3206 | machine_mode new_mode |
3207 | = SELECT_CC_MODE (compare_code, op0, op1); |
3208 | if (new_mode != orig_compare_mode |
3209 | && can_change_dest_mode (SET_DEST (newpat), |
3210 | added_sets: added_sets_2, mode: new_mode)) |
3211 | { |
3212 | unsigned int regno = REGNO (newpat_dest); |
3213 | compare_mode = new_mode; |
3214 | if (regno < FIRST_PSEUDO_REGISTER) |
3215 | newpat_dest = gen_rtx_REG (compare_mode, regno); |
3216 | else |
3217 | { |
3218 | subst_mode (regno, newval: compare_mode); |
3219 | newpat_dest = regno_reg_rtx[regno]; |
3220 | } |
3221 | } |
3222 | #endif |
3223 | /* Cases for modifying the CC-using comparison. */ |
3224 | if (compare_code != orig_compare_code |
3225 | /* ??? Do we need to verify the zero rtx? */ |
3226 | && XEXP (*cc_use_loc, 1) == const0_rtx) |
3227 | { |
3228 | /* Replace cc_use_loc with entire new RTX. */ |
3229 | SUBST (*cc_use_loc, |
3230 | gen_rtx_fmt_ee (compare_code, GET_MODE (*cc_use_loc), |
3231 | newpat_dest, const0_rtx)); |
3232 | undobuf.other_insn = cc_use_insn; |
3233 | } |
3234 | else if (compare_mode != orig_compare_mode) |
3235 | { |
3236 | /* Just replace the CC reg with a new mode. */ |
3237 | SUBST (XEXP (*cc_use_loc, 0), newpat_dest); |
3238 | undobuf.other_insn = cc_use_insn; |
3239 | } |
3240 | } |
3241 | |
3242 | /* Now we modify the current newpat: |
3243 | First, SET_DEST(newpat) is updated if the CC mode has been |
3244 | altered. For targets without SELECT_CC_MODE, this should be |
3245 | optimized away. */ |
3246 | if (compare_mode != orig_compare_mode) |
3247 | SUBST (SET_DEST (newpat), newpat_dest); |
3248 | /* This is always done to propagate i2src into newpat. */ |
3249 | SUBST (SET_SRC (newpat), |
3250 | gen_rtx_COMPARE (compare_mode, op0, op1)); |
3251 | /* Create new version of i2pat if needed; the below PARALLEL |
3252 | creation needs this to work correctly. */ |
3253 | if (! rtx_equal_p (i2src, op0)) |
3254 | i2pat = gen_rtx_SET (i2dest, op0); |
3255 | i2_is_used = 1; |
3256 | } |
3257 | } |
3258 | |
3259 | if (i2_is_used == 0) |
3260 | { |
3261 | /* It is possible that the source of I2 or I1 may be performing |
3262 | an unneeded operation, such as a ZERO_EXTEND of something |
3263 | that is known to have the high part zero. Handle that case |
3264 | by letting subst look at the inner insns. |
3265 | |
3266 | Another way to do this would be to have a function that tries |
3267 | to simplify a single insn instead of merging two or more |
3268 | insns. We don't do this because of the potential of infinite |
3269 | loops and because of the potential extra memory required. |
3270 | However, doing it the way we are is a bit of a kludge and |
3271 | doesn't catch all cases. |
3272 | |
3273 | But only do this if -fexpensive-optimizations since it slows |
3274 | things down and doesn't usually win. |
3275 | |
3276 | This is not done in the COMPARE case above because the |
3277 | unmodified I2PAT is used in the PARALLEL and so a pattern |
3278 | with a modified I2SRC would not match. */ |
3279 | |
3280 | if (flag_expensive_optimizations) |
3281 | { |
3282 | /* Pass pc_rtx so no substitutions are done, just |
3283 | simplifications. */ |
3284 | if (i1) |
3285 | { |
3286 | subst_low_luid = DF_INSN_LUID (i1); |
3287 | i1src = subst (i1src, pc_rtx, pc_rtx, false, false, false); |
3288 | } |
3289 | |
3290 | subst_low_luid = DF_INSN_LUID (i2); |
3291 | i2src = subst (i2src, pc_rtx, pc_rtx, false, false, false); |
3292 | } |
3293 | |
3294 | n_occurrences = 0; /* `subst' counts here */ |
3295 | subst_low_luid = DF_INSN_LUID (i2); |
3296 | |
3297 | /* If I1 feeds into I2 and I1DEST is in I1SRC, we need to make a unique |
3298 | copy of I2SRC each time we substitute it, in order to avoid creating |
3299 | self-referential RTL when we will be substituting I1SRC for I1DEST |
3300 | later. Likewise if I0 feeds into I2, either directly or indirectly |
3301 | through I1, and I0DEST is in I0SRC. */ |
3302 | newpat = subst (PATTERN (insn: i3), i2dest, i2src, false, false, |
3303 | (i1_feeds_i2_n && i1dest_in_i1src) |
3304 | || ((i0_feeds_i2_n || (i0_feeds_i1_n && i1_feeds_i2_n)) |
3305 | && i0dest_in_i0src)); |
3306 | substed_i2 = true; |
3307 | |
3308 | /* Record whether I2's body now appears within I3's body. */ |
3309 | i2_is_used = n_occurrences; |
3310 | } |
3311 | |
3312 | /* If we already got a failure, don't try to do more. Otherwise, try to |
3313 | substitute I1 if we have it. */ |
3314 | |
3315 | if (i1 && GET_CODE (newpat) != CLOBBER) |
3316 | { |
3317 | /* Before we can do this substitution, we must redo the test done |
3318 | above (see detailed comments there) that ensures I1DEST isn't |
3319 | mentioned in any SETs in NEWPAT that are field assignments. */ |
3320 | if (!combinable_i3pat (NULL, loc: &newpat, i2dest: i1dest, NULL_RTX, NULL_RTX, |
3321 | i1_not_in_src: false, i0_not_in_src: false, pi3dest_killed: 0)) |
3322 | { |
3323 | undo_all (); |
3324 | return 0; |
3325 | } |
3326 | |
3327 | n_occurrences = 0; |
3328 | subst_low_luid = DF_INSN_LUID (i1); |
3329 | |
3330 | /* If the following substitution will modify I1SRC, make a copy of it |
3331 | for the case where it is substituted for I1DEST in I2PAT later. */ |
3332 | if (added_sets_2 && i1_feeds_i2_n) |
3333 | i1src_copy = copy_rtx (i1src); |
3334 | |
3335 | /* If I0 feeds into I1 and I0DEST is in I0SRC, we need to make a unique |
3336 | copy of I1SRC each time we substitute it, in order to avoid creating |
3337 | self-referential RTL when we will be substituting I0SRC for I0DEST |
3338 | later. */ |
3339 | newpat = subst (newpat, i1dest, i1src, false, false, |
3340 | i0_feeds_i1_n && i0dest_in_i0src); |
3341 | substed_i1 = true; |
3342 | |
3343 | /* Record whether I1's body now appears within I3's body. */ |
3344 | i1_is_used = n_occurrences; |
3345 | } |
3346 | |
3347 | /* Likewise for I0 if we have it. */ |
3348 | |
3349 | if (i0 && GET_CODE (newpat) != CLOBBER) |
3350 | { |
3351 | if (!combinable_i3pat (NULL, loc: &newpat, i2dest: i0dest, NULL_RTX, NULL_RTX, |
3352 | i1_not_in_src: false, i0_not_in_src: false, pi3dest_killed: 0)) |
3353 | { |
3354 | undo_all (); |
3355 | return 0; |
3356 | } |
3357 | |
3358 | /* If the following substitution will modify I0SRC, make a copy of it |
3359 | for the case where it is substituted for I0DEST in I1PAT later. */ |
3360 | if (added_sets_1 && i0_feeds_i1_n) |
3361 | i0src_copy = copy_rtx (i0src); |
3362 | /* And a copy for I0DEST in I2PAT substitution. */ |
3363 | if (added_sets_2 && ((i0_feeds_i1_n && i1_feeds_i2_n) |
3364 | || (i0_feeds_i2_n))) |
3365 | i0src_copy2 = copy_rtx (i0src); |
3366 | |
3367 | n_occurrences = 0; |
3368 | subst_low_luid = DF_INSN_LUID (i0); |
3369 | newpat = subst (newpat, i0dest, i0src, false, false, false); |
3370 | substed_i0 = true; |
3371 | } |
3372 | |
3373 | if (n_auto_inc) |
3374 | { |
3375 | int new_n_auto_inc = 0; |
3376 | for_each_inc_dec (newpat, count_auto_inc, arg: &new_n_auto_inc); |
3377 | |
3378 | if (n_auto_inc != new_n_auto_inc) |
3379 | { |
3380 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3381 | fprintf (stream: dump_file, format: "Number of auto_inc expressions changed\n" ); |
3382 | undo_all (); |
3383 | return 0; |
3384 | } |
3385 | } |
3386 | |
3387 | /* Fail if an autoincrement side-effect has been duplicated. Be careful |
3388 | to count all the ways that I2SRC and I1SRC can be used. */ |
3389 | if ((FIND_REG_INC_NOTE (i2, NULL_RTX) != 0 |
3390 | && i2_is_used + added_sets_2 > 1) |
3391 | || (i1 != 0 && FIND_REG_INC_NOTE (i1, NULL_RTX) != 0 |
3392 | && (i1_is_used + added_sets_1 + (added_sets_2 && i1_feeds_i2_n) > 1)) |
3393 | || (i0 != 0 && FIND_REG_INC_NOTE (i0, NULL_RTX) != 0 |
3394 | && (n_occurrences + added_sets_0 |
3395 | + (added_sets_1 && i0_feeds_i1_n) |
3396 | + (added_sets_2 && i0_feeds_i2_n) > 1)) |
3397 | /* Fail if we tried to make a new register. */ |
3398 | || max_reg_num () != maxreg |
3399 | /* Fail if we couldn't do something and have a CLOBBER. */ |
3400 | || GET_CODE (newpat) == CLOBBER |
3401 | /* Fail if this new pattern is a MULT and we didn't have one before |
3402 | at the outer level. */ |
3403 | || (GET_CODE (newpat) == SET && GET_CODE (SET_SRC (newpat)) == MULT |
3404 | && ! have_mult)) |
3405 | { |
3406 | undo_all (); |
3407 | return 0; |
3408 | } |
3409 | |
3410 | /* If the actions of the earlier insns must be kept |
3411 | in addition to substituting them into the latest one, |
3412 | we must make a new PARALLEL for the latest insn |
3413 | to hold additional the SETs. */ |
3414 | |
3415 | if (added_sets_0 || added_sets_1 || added_sets_2) |
3416 | { |
3417 | int = added_sets_0 + added_sets_1 + added_sets_2; |
3418 | combine_extras++; |
3419 | |
3420 | if (GET_CODE (newpat) == PARALLEL) |
3421 | { |
3422 | rtvec old = XVEC (newpat, 0); |
3423 | total_sets = XVECLEN (newpat, 0) + extra_sets; |
3424 | newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets)); |
3425 | memcpy (XVEC (newpat, 0)->elem, src: &old->elem[0], |
3426 | n: sizeof (old->elem[0]) * old->num_elem); |
3427 | } |
3428 | else |
3429 | { |
3430 | rtx old = newpat; |
3431 | total_sets = 1 + extra_sets; |
3432 | newpat = gen_rtx_PARALLEL (VOIDmode, rtvec_alloc (total_sets)); |
3433 | XVECEXP (newpat, 0, 0) = old; |
3434 | } |
3435 | |
3436 | if (added_sets_0) |
3437 | XVECEXP (newpat, 0, --total_sets) = i0pat; |
3438 | |
3439 | if (added_sets_1) |
3440 | { |
3441 | rtx t = i1pat; |
3442 | if (i0_feeds_i1_n) |
3443 | t = subst (t, i0dest, i0src_copy ? i0src_copy : i0src, |
3444 | false, false, false); |
3445 | |
3446 | XVECEXP (newpat, 0, --total_sets) = t; |
3447 | } |
3448 | if (added_sets_2) |
3449 | { |
3450 | rtx t = i2pat; |
3451 | if (i1_feeds_i2_n) |
3452 | t = subst (t, i1dest, i1src_copy ? i1src_copy : i1src, false, false, |
3453 | i0_feeds_i1_n && i0dest_in_i0src); |
3454 | if ((i0_feeds_i1_n && i1_feeds_i2_n) || i0_feeds_i2_n) |
3455 | t = subst (t, i0dest, i0src_copy2 ? i0src_copy2 : i0src, |
3456 | false, false, false); |
3457 | |
3458 | XVECEXP (newpat, 0, --total_sets) = t; |
3459 | } |
3460 | } |
3461 | |
3462 | validate_replacement: |
3463 | |
3464 | /* Note which hard regs this insn has as inputs. */ |
3465 | mark_used_regs_combine (newpat); |
3466 | |
3467 | /* If recog_for_combine fails, it strips existing clobbers. If we'll |
3468 | consider splitting this pattern, we might need these clobbers. */ |
3469 | if (i1 && GET_CODE (newpat) == PARALLEL |
3470 | && GET_CODE (XVECEXP (newpat, 0, XVECLEN (newpat, 0) - 1)) == CLOBBER) |
3471 | { |
3472 | int len = XVECLEN (newpat, 0); |
3473 | |
3474 | newpat_vec_with_clobbers = rtvec_alloc (len); |
3475 | for (i = 0; i < len; i++) |
3476 | RTVEC_ELT (newpat_vec_with_clobbers, i) = XVECEXP (newpat, 0, i); |
3477 | } |
3478 | |
3479 | /* We have recognized nothing yet. */ |
3480 | insn_code_number = -1; |
3481 | |
3482 | /* See if this is a PARALLEL of two SETs where one SET's destination is |
3483 | a register that is unused and this isn't marked as an instruction that |
3484 | might trap in an EH region. In that case, we just need the other SET. |
3485 | We prefer this over the PARALLEL. |
3486 | |
3487 | This can occur when simplifying a divmod insn. We *must* test for this |
3488 | case here because the code below that splits two independent SETs doesn't |
3489 | handle this case correctly when it updates the register status. |
3490 | |
3491 | It's pointless doing this if we originally had two sets, one from |
3492 | i3, and one from i2. Combining then splitting the parallel results |
3493 | in the original i2 again plus an invalid insn (which we delete). |
3494 | The net effect is only to move instructions around, which makes |
3495 | debug info less accurate. |
3496 | |
3497 | If the remaining SET came from I2 its destination should not be used |
3498 | between I2 and I3. See PR82024. */ |
3499 | |
3500 | if (!(added_sets_2 && i1 == 0) |
3501 | && is_parallel_of_n_reg_sets (pat: newpat, n: 2) |
3502 | && asm_noperands (newpat) < 0) |
3503 | { |
3504 | rtx set0 = XVECEXP (newpat, 0, 0); |
3505 | rtx set1 = XVECEXP (newpat, 0, 1); |
3506 | rtx oldpat = newpat; |
3507 | |
3508 | if (((REG_P (SET_DEST (set1)) |
3509 | && find_reg_note (i3, REG_UNUSED, SET_DEST (set1))) |
3510 | || (GET_CODE (SET_DEST (set1)) == SUBREG |
3511 | && find_reg_note (i3, REG_UNUSED, SUBREG_REG (SET_DEST (set1))))) |
3512 | && insn_nothrow_p (i3) |
3513 | && !side_effects_p (SET_SRC (set1))) |
3514 | { |
3515 | newpat = set0; |
3516 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); |
3517 | } |
3518 | |
3519 | else if (((REG_P (SET_DEST (set0)) |
3520 | && find_reg_note (i3, REG_UNUSED, SET_DEST (set0))) |
3521 | || (GET_CODE (SET_DEST (set0)) == SUBREG |
3522 | && find_reg_note (i3, REG_UNUSED, |
3523 | SUBREG_REG (SET_DEST (set0))))) |
3524 | && insn_nothrow_p (i3) |
3525 | && !side_effects_p (SET_SRC (set0))) |
3526 | { |
3527 | rtx dest = SET_DEST (set1); |
3528 | if (GET_CODE (dest) == SUBREG) |
3529 | dest = SUBREG_REG (dest); |
3530 | if (!reg_used_between_p (dest, i2, i3)) |
3531 | { |
3532 | newpat = set1; |
3533 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); |
3534 | |
3535 | if (insn_code_number >= 0) |
3536 | changed_i3_dest = true; |
3537 | } |
3538 | } |
3539 | |
3540 | if (insn_code_number < 0) |
3541 | newpat = oldpat; |
3542 | } |
3543 | |
3544 | /* Is the result of combination a valid instruction? */ |
3545 | if (insn_code_number < 0) |
3546 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); |
3547 | |
3548 | /* If we were combining three insns and the result is a simple SET |
3549 | with no ASM_OPERANDS that wasn't recognized, try to split it into two |
3550 | insns. There are two ways to do this. It can be split using a |
3551 | machine-specific method (like when you have an addition of a large |
3552 | constant) or by combine in the function find_split_point. */ |
3553 | |
3554 | if (i1 && insn_code_number < 0 && GET_CODE (newpat) == SET |
3555 | && asm_noperands (newpat) < 0) |
3556 | { |
3557 | rtx parallel, *split; |
3558 | rtx_insn *m_split_insn; |
3559 | |
3560 | /* See if the MD file can split NEWPAT. If it can't, see if letting it |
3561 | use I2DEST as a scratch register will help. In the latter case, |
3562 | convert I2DEST to the mode of the source of NEWPAT if we can. */ |
3563 | |
3564 | m_split_insn = combine_split_insns (pattern: newpat, insn: i3); |
3565 | |
3566 | /* We can only use I2DEST as a scratch reg if it doesn't overlap any |
3567 | inputs of NEWPAT. */ |
3568 | |
3569 | /* ??? If I2DEST is not safe, and I1DEST exists, then it would be |
3570 | possible to try that as a scratch reg. This would require adding |
3571 | more code to make it work though. */ |
3572 | |
3573 | if (m_split_insn == 0 && ! reg_overlap_mentioned_p (i2dest, newpat)) |
3574 | { |
3575 | machine_mode new_mode = GET_MODE (SET_DEST (newpat)); |
3576 | |
3577 | /* ??? Reusing i2dest without resetting the reg_stat entry for it |
3578 | (temporarily, until we are committed to this instruction |
3579 | combination) does not work: for example, any call to nonzero_bits |
3580 | on the register (from a splitter in the MD file, for example) |
3581 | will get the old information, which is invalid. |
3582 | |
3583 | Since nowadays we can create registers during combine just fine, |
3584 | we should just create a new one here, not reuse i2dest. */ |
3585 | |
3586 | /* First try to split using the original register as a |
3587 | scratch register. */ |
3588 | parallel = gen_rtx_PARALLEL (VOIDmode, |
3589 | gen_rtvec (2, newpat, |
3590 | gen_rtx_CLOBBER (VOIDmode, |
3591 | i2dest))); |
3592 | m_split_insn = combine_split_insns (pattern: parallel, insn: i3); |
3593 | |
3594 | /* If that didn't work, try changing the mode of I2DEST if |
3595 | we can. */ |
3596 | if (m_split_insn == 0 |
3597 | && new_mode != GET_MODE (i2dest) |
3598 | && new_mode != VOIDmode |
3599 | && can_change_dest_mode (x: i2dest, added_sets: added_sets_2, mode: new_mode)) |
3600 | { |
3601 | machine_mode old_mode = GET_MODE (i2dest); |
3602 | rtx ni2dest; |
3603 | |
3604 | if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER) |
3605 | ni2dest = gen_rtx_REG (new_mode, REGNO (i2dest)); |
3606 | else |
3607 | { |
3608 | subst_mode (REGNO (i2dest), newval: new_mode); |
3609 | ni2dest = regno_reg_rtx[REGNO (i2dest)]; |
3610 | } |
3611 | |
3612 | parallel = (gen_rtx_PARALLEL |
3613 | (VOIDmode, |
3614 | gen_rtvec (2, newpat, |
3615 | gen_rtx_CLOBBER (VOIDmode, |
3616 | ni2dest)))); |
3617 | m_split_insn = combine_split_insns (pattern: parallel, insn: i3); |
3618 | |
3619 | if (m_split_insn == 0 |
3620 | && REGNO (i2dest) >= FIRST_PSEUDO_REGISTER) |
3621 | { |
3622 | struct undo *buf; |
3623 | |
3624 | adjust_reg_mode (regno_reg_rtx[REGNO (i2dest)], old_mode); |
3625 | buf = undobuf.undos; |
3626 | undobuf.undos = buf->next; |
3627 | buf->next = undobuf.frees; |
3628 | undobuf.frees = buf; |
3629 | } |
3630 | } |
3631 | |
3632 | i2scratch = m_split_insn != 0; |
3633 | } |
3634 | |
3635 | /* If recog_for_combine has discarded clobbers, try to use them |
3636 | again for the split. */ |
3637 | if (m_split_insn == 0 && newpat_vec_with_clobbers) |
3638 | { |
3639 | parallel = gen_rtx_PARALLEL (VOIDmode, newpat_vec_with_clobbers); |
3640 | m_split_insn = combine_split_insns (pattern: parallel, insn: i3); |
3641 | } |
3642 | |
3643 | if (m_split_insn && NEXT_INSN (insn: m_split_insn) == NULL_RTX) |
3644 | { |
3645 | rtx m_split_pat = PATTERN (insn: m_split_insn); |
3646 | insn_code_number = recog_for_combine (&m_split_pat, i3, &new_i3_notes); |
3647 | if (insn_code_number >= 0) |
3648 | newpat = m_split_pat; |
3649 | } |
3650 | else if (m_split_insn && NEXT_INSN (insn: NEXT_INSN (insn: m_split_insn)) == NULL_RTX |
3651 | && (next_nonnote_nondebug_insn (i2) == i3 |
3652 | || !modified_between_p (PATTERN (insn: m_split_insn), i2, i3))) |
3653 | { |
3654 | rtx i2set, i3set; |
3655 | rtx newi3pat = PATTERN (insn: NEXT_INSN (insn: m_split_insn)); |
3656 | newi2pat = PATTERN (insn: m_split_insn); |
3657 | |
3658 | i3set = single_set (insn: NEXT_INSN (insn: m_split_insn)); |
3659 | i2set = single_set (insn: m_split_insn); |
3660 | |
3661 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes); |
3662 | |
3663 | /* If I2 or I3 has multiple SETs, we won't know how to track |
3664 | register status, so don't use these insns. If I2's destination |
3665 | is used between I2 and I3, we also can't use these insns. */ |
3666 | |
3667 | if (i2_code_number >= 0 && i2set && i3set |
3668 | && (next_nonnote_nondebug_insn (i2) == i3 |
3669 | || ! reg_used_between_p (SET_DEST (i2set), i2, i3))) |
3670 | insn_code_number = recog_for_combine (&newi3pat, i3, |
3671 | &new_i3_notes); |
3672 | if (insn_code_number >= 0) |
3673 | newpat = newi3pat; |
3674 | |
3675 | /* It is possible that both insns now set the destination of I3. |
3676 | If so, we must show an extra use of it. */ |
3677 | |
3678 | if (insn_code_number >= 0) |
3679 | { |
3680 | rtx new_i3_dest = SET_DEST (i3set); |
3681 | rtx new_i2_dest = SET_DEST (i2set); |
3682 | |
3683 | while (GET_CODE (new_i3_dest) == ZERO_EXTRACT |
3684 | || GET_CODE (new_i3_dest) == STRICT_LOW_PART |
3685 | || GET_CODE (new_i3_dest) == SUBREG) |
3686 | new_i3_dest = XEXP (new_i3_dest, 0); |
3687 | |
3688 | while (GET_CODE (new_i2_dest) == ZERO_EXTRACT |
3689 | || GET_CODE (new_i2_dest) == STRICT_LOW_PART |
3690 | || GET_CODE (new_i2_dest) == SUBREG) |
3691 | new_i2_dest = XEXP (new_i2_dest, 0); |
3692 | |
3693 | if (REG_P (new_i3_dest) |
3694 | && REG_P (new_i2_dest) |
3695 | && REGNO (new_i3_dest) == REGNO (new_i2_dest) |
3696 | && REGNO (new_i2_dest) < reg_n_sets_max) |
3697 | INC_REG_N_SETS (REGNO (new_i2_dest), 1); |
3698 | } |
3699 | } |
3700 | |
3701 | /* If we can split it and use I2DEST, go ahead and see if that |
3702 | helps things be recognized. Verify that none of the registers |
3703 | are set between I2 and I3. */ |
3704 | if (insn_code_number < 0 |
3705 | && (split = find_split_point (&newpat, i3, false)) != 0 |
3706 | /* We need I2DEST in the proper mode. If it is a hard register |
3707 | or the only use of a pseudo, we can change its mode. |
3708 | Make sure we don't change a hard register to have a mode that |
3709 | isn't valid for it, or change the number of registers. */ |
3710 | && (GET_MODE (*split) == GET_MODE (i2dest) |
3711 | || GET_MODE (*split) == VOIDmode |
3712 | || can_change_dest_mode (x: i2dest, added_sets: added_sets_2, |
3713 | GET_MODE (*split))) |
3714 | && (next_nonnote_nondebug_insn (i2) == i3 |
3715 | || !modified_between_p (*split, i2, i3)) |
3716 | /* We can't overwrite I2DEST if its value is still used by |
3717 | NEWPAT. */ |
3718 | && ! reg_referenced_p (i2dest, newpat) |
3719 | /* We should not split a possibly trapping part when we |
3720 | care about non-call EH and have REG_EH_REGION notes |
3721 | to distribute. */ |
3722 | && ! (cfun->can_throw_non_call_exceptions |
3723 | && has_non_call_exception |
3724 | && may_trap_p (*split))) |
3725 | { |
3726 | rtx newdest = i2dest; |
3727 | enum rtx_code split_code = GET_CODE (*split); |
3728 | machine_mode split_mode = GET_MODE (*split); |
3729 | bool subst_done = false; |
3730 | newi2pat = NULL_RTX; |
3731 | |
3732 | i2scratch = true; |
3733 | |
3734 | /* *SPLIT may be part of I2SRC, so make sure we have the |
3735 | original expression around for later debug processing. |
3736 | We should not need I2SRC any more in other cases. */ |
3737 | if (MAY_HAVE_DEBUG_BIND_INSNS) |
3738 | i2src = copy_rtx (i2src); |
3739 | else |
3740 | i2src = NULL; |
3741 | |
3742 | /* Get NEWDEST as a register in the proper mode. We have already |
3743 | validated that we can do this. */ |
3744 | if (GET_MODE (i2dest) != split_mode && split_mode != VOIDmode) |
3745 | { |
3746 | if (REGNO (i2dest) < FIRST_PSEUDO_REGISTER) |
3747 | newdest = gen_rtx_REG (split_mode, REGNO (i2dest)); |
3748 | else |
3749 | { |
3750 | subst_mode (REGNO (i2dest), newval: split_mode); |
3751 | newdest = regno_reg_rtx[REGNO (i2dest)]; |
3752 | } |
3753 | } |
3754 | |
3755 | /* If *SPLIT is a (mult FOO (const_int pow2)), convert it to |
3756 | an ASHIFT. This can occur if it was inside a PLUS and hence |
3757 | appeared to be a memory address. This is a kludge. */ |
3758 | if (split_code == MULT |
3759 | && CONST_INT_P (XEXP (*split, 1)) |
3760 | && INTVAL (XEXP (*split, 1)) > 0 |
3761 | && (i = exact_log2 (UINTVAL (XEXP (*split, 1)))) >= 0) |
3762 | { |
3763 | rtx i_rtx = gen_int_shift_amount (split_mode, i); |
3764 | SUBST (*split, gen_rtx_ASHIFT (split_mode, |
3765 | XEXP (*split, 0), i_rtx)); |
3766 | /* Update split_code because we may not have a multiply |
3767 | anymore. */ |
3768 | split_code = GET_CODE (*split); |
3769 | } |
3770 | |
3771 | /* Similarly for (plus (mult FOO (const_int pow2))). */ |
3772 | if (split_code == PLUS |
3773 | && GET_CODE (XEXP (*split, 0)) == MULT |
3774 | && CONST_INT_P (XEXP (XEXP (*split, 0), 1)) |
3775 | && INTVAL (XEXP (XEXP (*split, 0), 1)) > 0 |
3776 | && (i = exact_log2 (UINTVAL (XEXP (XEXP (*split, 0), 1)))) >= 0) |
3777 | { |
3778 | rtx nsplit = XEXP (*split, 0); |
3779 | rtx i_rtx = gen_int_shift_amount (GET_MODE (nsplit), i); |
3780 | SUBST (XEXP (*split, 0), gen_rtx_ASHIFT (GET_MODE (nsplit), |
3781 | XEXP (nsplit, 0), |
3782 | i_rtx)); |
3783 | /* Update split_code because we may not have a multiply |
3784 | anymore. */ |
3785 | split_code = GET_CODE (*split); |
3786 | } |
3787 | |
3788 | #ifdef INSN_SCHEDULING |
3789 | /* If *SPLIT is a paradoxical SUBREG, when we split it, it should |
3790 | be written as a ZERO_EXTEND. */ |
3791 | if (split_code == SUBREG && MEM_P (SUBREG_REG (*split))) |
3792 | { |
3793 | /* Or as a SIGN_EXTEND if LOAD_EXTEND_OP says that that's |
3794 | what it really is. */ |
3795 | if (load_extend_op (GET_MODE (SUBREG_REG (*split))) |
3796 | == SIGN_EXTEND) |
3797 | SUBST (*split, gen_rtx_SIGN_EXTEND (split_mode, |
3798 | SUBREG_REG (*split))); |
3799 | else |
3800 | SUBST (*split, gen_rtx_ZERO_EXTEND (split_mode, |
3801 | SUBREG_REG (*split))); |
3802 | } |
3803 | #endif |
3804 | |
3805 | /* Attempt to split binary operators using arithmetic identities. */ |
3806 | if (BINARY_P (SET_SRC (newpat)) |
3807 | && split_mode == GET_MODE (SET_SRC (newpat)) |
3808 | && ! side_effects_p (SET_SRC (newpat))) |
3809 | { |
3810 | rtx setsrc = SET_SRC (newpat); |
3811 | machine_mode mode = GET_MODE (setsrc); |
3812 | enum rtx_code code = GET_CODE (setsrc); |
3813 | rtx src_op0 = XEXP (setsrc, 0); |
3814 | rtx src_op1 = XEXP (setsrc, 1); |
3815 | |
3816 | /* Split "X = Y op Y" as "Z = Y; X = Z op Z". */ |
3817 | if (rtx_equal_p (src_op0, src_op1)) |
3818 | { |
3819 | newi2pat = gen_rtx_SET (newdest, src_op0); |
3820 | SUBST (XEXP (setsrc, 0), newdest); |
3821 | SUBST (XEXP (setsrc, 1), newdest); |
3822 | subst_done = true; |
3823 | } |
3824 | /* Split "((P op Q) op R) op S" where op is PLUS or MULT. */ |
3825 | else if ((code == PLUS || code == MULT) |
3826 | && GET_CODE (src_op0) == code |
3827 | && GET_CODE (XEXP (src_op0, 0)) == code |
3828 | && (INTEGRAL_MODE_P (mode) |
3829 | || (FLOAT_MODE_P (mode) |
3830 | && flag_unsafe_math_optimizations))) |
3831 | { |
3832 | rtx p = XEXP (XEXP (src_op0, 0), 0); |
3833 | rtx q = XEXP (XEXP (src_op0, 0), 1); |
3834 | rtx r = XEXP (src_op0, 1); |
3835 | rtx s = src_op1; |
3836 | |
3837 | /* Split both "((X op Y) op X) op Y" and |
3838 | "((X op Y) op Y) op X" as "T op T" where T is |
3839 | "X op Y". */ |
3840 | if ((rtx_equal_p (p,r) && rtx_equal_p (q,s)) |
3841 | || (rtx_equal_p (p,s) && rtx_equal_p (q,r))) |
3842 | { |
3843 | newi2pat = gen_rtx_SET (newdest, XEXP (src_op0, 0)); |
3844 | SUBST (XEXP (setsrc, 0), newdest); |
3845 | SUBST (XEXP (setsrc, 1), newdest); |
3846 | subst_done = true; |
3847 | } |
3848 | /* Split "((X op X) op Y) op Y)" as "T op T" where |
3849 | T is "X op Y". */ |
3850 | else if (rtx_equal_p (p,q) && rtx_equal_p (r,s)) |
3851 | { |
3852 | rtx tmp = simplify_gen_binary (code, mode, op0: p, op1: r); |
3853 | newi2pat = gen_rtx_SET (newdest, tmp); |
3854 | SUBST (XEXP (setsrc, 0), newdest); |
3855 | SUBST (XEXP (setsrc, 1), newdest); |
3856 | subst_done = true; |
3857 | } |
3858 | } |
3859 | } |
3860 | |
3861 | if (!subst_done) |
3862 | { |
3863 | newi2pat = gen_rtx_SET (newdest, *split); |
3864 | SUBST (*split, newdest); |
3865 | } |
3866 | |
3867 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes); |
3868 | |
3869 | /* recog_for_combine might have added CLOBBERs to newi2pat. |
3870 | Make sure NEWPAT does not depend on the clobbered regs. */ |
3871 | if (GET_CODE (newi2pat) == PARALLEL) |
3872 | for (i = XVECLEN (newi2pat, 0) - 1; i >= 0; i--) |
3873 | if (GET_CODE (XVECEXP (newi2pat, 0, i)) == CLOBBER) |
3874 | { |
3875 | rtx reg = XEXP (XVECEXP (newi2pat, 0, i), 0); |
3876 | if (reg_overlap_mentioned_p (reg, newpat)) |
3877 | { |
3878 | undo_all (); |
3879 | return 0; |
3880 | } |
3881 | } |
3882 | |
3883 | /* If the split point was a MULT and we didn't have one before, |
3884 | don't use one now. */ |
3885 | if (i2_code_number >= 0 && ! (split_code == MULT && ! have_mult)) |
3886 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); |
3887 | } |
3888 | } |
3889 | |
3890 | /* Check for a case where we loaded from memory in a narrow mode and |
3891 | then sign extended it, but we need both registers. In that case, |
3892 | we have a PARALLEL with both loads from the same memory location. |
3893 | We can split this into a load from memory followed by a register-register |
3894 | copy. This saves at least one insn, more if register allocation can |
3895 | eliminate the copy. |
3896 | |
3897 | We cannot do this if the destination of the first assignment is a |
3898 | condition code register. We eliminate this case by making sure |
3899 | the SET_DEST and SET_SRC have the same mode. |
3900 | |
3901 | We cannot do this if the destination of the second assignment is |
3902 | a register that we have already assumed is zero-extended. Similarly |
3903 | for a SUBREG of such a register. */ |
3904 | |
3905 | else if (i1 && insn_code_number < 0 && asm_noperands (newpat) < 0 |
3906 | && GET_CODE (newpat) == PARALLEL |
3907 | && XVECLEN (newpat, 0) == 2 |
3908 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET |
3909 | && GET_CODE (SET_SRC (XVECEXP (newpat, 0, 0))) == SIGN_EXTEND |
3910 | && (GET_MODE (SET_DEST (XVECEXP (newpat, 0, 0))) |
3911 | == GET_MODE (SET_SRC (XVECEXP (newpat, 0, 0)))) |
3912 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET |
3913 | && rtx_equal_p (SET_SRC (XVECEXP (newpat, 0, 1)), |
3914 | XEXP (SET_SRC (XVECEXP (newpat, 0, 0)), 0)) |
3915 | && !modified_between_p (SET_SRC (XVECEXP (newpat, 0, 1)), i2, i3) |
3916 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT |
3917 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART |
3918 | && ! (temp_expr = SET_DEST (XVECEXP (newpat, 0, 1)), |
3919 | (REG_P (temp_expr) |
3920 | && reg_stat[REGNO (temp_expr)].nonzero_bits != 0 |
3921 | && known_lt (GET_MODE_PRECISION (GET_MODE (temp_expr)), |
3922 | BITS_PER_WORD) |
3923 | && known_lt (GET_MODE_PRECISION (GET_MODE (temp_expr)), |
3924 | HOST_BITS_PER_INT) |
3925 | && (reg_stat[REGNO (temp_expr)].nonzero_bits |
3926 | != GET_MODE_MASK (word_mode)))) |
3927 | && ! (GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) == SUBREG |
3928 | && (temp_expr = SUBREG_REG (SET_DEST (XVECEXP (newpat, 0, 1))), |
3929 | (REG_P (temp_expr) |
3930 | && reg_stat[REGNO (temp_expr)].nonzero_bits != 0 |
3931 | && known_lt (GET_MODE_PRECISION (GET_MODE (temp_expr)), |
3932 | BITS_PER_WORD) |
3933 | && known_lt (GET_MODE_PRECISION (GET_MODE (temp_expr)), |
3934 | HOST_BITS_PER_INT) |
3935 | && (reg_stat[REGNO (temp_expr)].nonzero_bits |
3936 | != GET_MODE_MASK (word_mode))))) |
3937 | && ! reg_overlap_mentioned_p (SET_DEST (XVECEXP (newpat, 0, 1)), |
3938 | SET_SRC (XVECEXP (newpat, 0, 1))) |
3939 | && ! find_reg_note (i3, REG_UNUSED, |
3940 | SET_DEST (XVECEXP (newpat, 0, 0)))) |
3941 | { |
3942 | rtx ni2dest; |
3943 | |
3944 | newi2pat = XVECEXP (newpat, 0, 0); |
3945 | ni2dest = SET_DEST (XVECEXP (newpat, 0, 0)); |
3946 | newpat = XVECEXP (newpat, 0, 1); |
3947 | SUBST (SET_SRC (newpat), |
3948 | gen_lowpart (GET_MODE (SET_SRC (newpat)), ni2dest)); |
3949 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes); |
3950 | |
3951 | if (i2_code_number >= 0) |
3952 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); |
3953 | |
3954 | if (insn_code_number >= 0) |
3955 | swap_i2i3 = 1; |
3956 | } |
3957 | |
3958 | /* Similarly, check for a case where we have a PARALLEL of two independent |
3959 | SETs but we started with three insns. In this case, we can do the sets |
3960 | as two separate insns. This case occurs when some SET allows two |
3961 | other insns to combine, but the destination of that SET is still live. |
3962 | |
3963 | Also do this if we started with two insns and (at least) one of the |
3964 | resulting sets is a noop; this noop will be deleted later. |
3965 | |
3966 | Also do this if we started with two insns neither of which was a simple |
3967 | move. */ |
3968 | |
3969 | else if (insn_code_number < 0 && asm_noperands (newpat) < 0 |
3970 | && GET_CODE (newpat) == PARALLEL |
3971 | && XVECLEN (newpat, 0) == 2 |
3972 | && GET_CODE (XVECEXP (newpat, 0, 0)) == SET |
3973 | && GET_CODE (XVECEXP (newpat, 0, 1)) == SET |
3974 | && (i1 |
3975 | || set_noop_p (XVECEXP (newpat, 0, 0)) |
3976 | || set_noop_p (XVECEXP (newpat, 0, 1)) |
3977 | || (!i2_was_move && !i3_was_move)) |
3978 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != ZERO_EXTRACT |
3979 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 0))) != STRICT_LOW_PART |
3980 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != ZERO_EXTRACT |
3981 | && GET_CODE (SET_DEST (XVECEXP (newpat, 0, 1))) != STRICT_LOW_PART |
3982 | && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 1)), |
3983 | XVECEXP (newpat, 0, 0)) |
3984 | && ! reg_referenced_p (SET_DEST (XVECEXP (newpat, 0, 0)), |
3985 | XVECEXP (newpat, 0, 1)) |
3986 | && ! (contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 0))) |
3987 | && contains_muldiv (SET_SRC (XVECEXP (newpat, 0, 1))))) |
3988 | { |
3989 | rtx set0 = XVECEXP (newpat, 0, 0); |
3990 | rtx set1 = XVECEXP (newpat, 0, 1); |
3991 | |
3992 | /* Normally, it doesn't matter which of the two is done first, but |
3993 | one which uses any regs/memory set in between i2 and i3 can't |
3994 | be first. The PARALLEL might also have been pre-existing in i3, |
3995 | so we need to make sure that we won't wrongly hoist a SET to i2 |
3996 | that would conflict with a death note present in there, or would |
3997 | have its dest modified between i2 and i3. */ |
3998 | if (!modified_between_p (SET_SRC (set1), i2, i3) |
3999 | && !(REG_P (SET_DEST (set1)) |
4000 | && find_reg_note (i2, REG_DEAD, SET_DEST (set1))) |
4001 | && !(GET_CODE (SET_DEST (set1)) == SUBREG |
4002 | && find_reg_note (i2, REG_DEAD, |
4003 | SUBREG_REG (SET_DEST (set1)))) |
4004 | && !modified_between_p (SET_DEST (set1), i2, i3) |
4005 | /* If I3 is a jump, ensure that set0 is a jump so that |
4006 | we do not create invalid RTL. */ |
4007 | && (!JUMP_P (i3) || SET_DEST (set0) == pc_rtx) |
4008 | ) |
4009 | { |
4010 | newi2pat = set1; |
4011 | newpat = set0; |
4012 | } |
4013 | else if (!modified_between_p (SET_SRC (set0), i2, i3) |
4014 | && !(REG_P (SET_DEST (set0)) |
4015 | && find_reg_note (i2, REG_DEAD, SET_DEST (set0))) |
4016 | && !(GET_CODE (SET_DEST (set0)) == SUBREG |
4017 | && find_reg_note (i2, REG_DEAD, |
4018 | SUBREG_REG (SET_DEST (set0)))) |
4019 | && !modified_between_p (SET_DEST (set0), i2, i3) |
4020 | /* If I3 is a jump, ensure that set1 is a jump so that |
4021 | we do not create invalid RTL. */ |
4022 | && (!JUMP_P (i3) || SET_DEST (set1) == pc_rtx) |
4023 | ) |
4024 | { |
4025 | newi2pat = set0; |
4026 | newpat = set1; |
4027 | } |
4028 | else |
4029 | { |
4030 | undo_all (); |
4031 | return 0; |
4032 | } |
4033 | |
4034 | i2_code_number = recog_for_combine (&newi2pat, i2, &new_i2_notes); |
4035 | |
4036 | if (i2_code_number >= 0) |
4037 | { |
4038 | /* recog_for_combine might have added CLOBBERs to newi2pat. |
4039 | Make sure NEWPAT does not depend on the clobbered regs. */ |
4040 | if (GET_CODE (newi2pat) == PARALLEL) |
4041 | { |
4042 | for (i = XVECLEN (newi2pat, 0) - 1; i >= 0; i--) |
4043 | if (GET_CODE (XVECEXP (newi2pat, 0, i)) == CLOBBER) |
4044 | { |
4045 | rtx reg = XEXP (XVECEXP (newi2pat, 0, i), 0); |
4046 | if (reg_overlap_mentioned_p (reg, newpat)) |
4047 | { |
4048 | undo_all (); |
4049 | return 0; |
4050 | } |
4051 | } |
4052 | } |
4053 | |
4054 | insn_code_number = recog_for_combine (&newpat, i3, &new_i3_notes); |
4055 | |
4056 | /* Likewise, recog_for_combine might have added clobbers to NEWPAT. |
4057 | Checking that the SET0's SET_DEST and SET1's SET_DEST aren't |
4058 | mentioned/clobbered, ensures NEWI2PAT's SET_DEST is live. */ |
4059 | if (insn_code_number >= 0 && GET_CODE (newpat) == PARALLEL) |
4060 | { |
4061 | for (i = XVECLEN (newpat, 0) - 1; i >= 0; i--) |
4062 | if (GET_CODE (XVECEXP (newpat, 0, i)) == CLOBBER) |
4063 | { |
4064 | rtx reg = XEXP (XVECEXP (newpat, 0, i), 0); |
4065 | if (reg_overlap_mentioned_p (reg, SET_DEST (set0)) |
4066 | || reg_overlap_mentioned_p (reg, SET_DEST (set1))) |
4067 | { |
4068 | undo_all (); |
4069 | return 0; |
4070 | } |
4071 | } |
4072 | } |
4073 | |
4074 | if (insn_code_number >= 0) |
4075 | split_i2i3 = true; |
4076 | } |
4077 | } |
4078 | |
4079 | /* If it still isn't recognized, fail and change things back the way they |
4080 | were. */ |
4081 | if ((insn_code_number < 0 |
4082 | /* Is the result a reasonable ASM_OPERANDS? */ |
4083 | && (! check_asm_operands (newpat) || added_sets_1 || added_sets_2))) |
4084 | { |
4085 | undo_all (); |
4086 | return 0; |
4087 | } |
4088 | |
4089 | /* If we had to change another insn, make sure it is valid also. */ |
4090 | if (undobuf.other_insn) |
4091 | { |
4092 | CLEAR_HARD_REG_SET (set&: newpat_used_regs); |
4093 | |
4094 | other_pat = PATTERN (insn: undobuf.other_insn); |
4095 | other_code_number = recog_for_combine (&other_pat, undobuf.other_insn, |
4096 | &new_other_notes); |
4097 | |
4098 | if (other_code_number < 0 && ! check_asm_operands (other_pat)) |
4099 | { |
4100 | undo_all (); |
4101 | return 0; |
4102 | } |
4103 | } |
4104 | |
4105 | /* Only allow this combination if insn_cost reports that the |
4106 | replacement instructions are cheaper than the originals. */ |
4107 | if (!combine_validate_cost (i0, i1, i2, i3, newpat, newi2pat, newotherpat: other_pat)) |
4108 | { |
4109 | undo_all (); |
4110 | return 0; |
4111 | } |
4112 | |
4113 | if (MAY_HAVE_DEBUG_BIND_INSNS) |
4114 | { |
4115 | struct undo *undo; |
4116 | |
4117 | for (undo = undobuf.undos; undo; undo = undo->next) |
4118 | if (undo->kind == UNDO_MODE) |
4119 | { |
4120 | rtx reg = regno_reg_rtx[undo->where.regno]; |
4121 | machine_mode new_mode = GET_MODE (reg); |
4122 | machine_mode old_mode = undo->old_contents.m; |
4123 | |
4124 | /* Temporarily revert mode back. */ |
4125 | adjust_reg_mode (reg, old_mode); |
4126 | |
4127 | if (reg == i2dest && i2scratch) |
4128 | { |
4129 | /* If we used i2dest as a scratch register with a |
4130 | different mode, substitute it for the original |
4131 | i2src while its original mode is temporarily |
4132 | restored, and then clear i2scratch so that we don't |
4133 | do it again later. */ |
4134 | propagate_for_debug (i2, last_combined_insn, reg, i2src, |
4135 | this_basic_block); |
4136 | i2scratch = false; |
4137 | /* Put back the new mode. */ |
4138 | adjust_reg_mode (reg, new_mode); |
4139 | } |
4140 | else |
4141 | { |
4142 | rtx tempreg = gen_raw_REG (old_mode, REGNO (reg)); |
4143 | rtx_insn *first, *last; |
4144 | |
4145 | if (reg == i2dest) |
4146 | { |
4147 | first = i2; |
4148 | last = last_combined_insn; |
4149 | } |
4150 | else |
4151 | { |
4152 | first = i3; |
4153 | last = undobuf.other_insn; |
4154 | gcc_assert (last); |
4155 | if (DF_INSN_LUID (last) |
4156 | < DF_INSN_LUID (last_combined_insn)) |
4157 | last = last_combined_insn; |
4158 | } |
4159 | |
4160 | /* We're dealing with a reg that changed mode but not |
4161 | meaning, so we want to turn it into a subreg for |
4162 | the new mode. However, because of REG sharing and |
4163 | because its mode had already changed, we have to do |
4164 | it in two steps. First, replace any debug uses of |
4165 | reg, with its original mode temporarily restored, |
4166 | with this copy we have created; then, replace the |
4167 | copy with the SUBREG of the original shared reg, |
4168 | once again changed to the new mode. */ |
4169 | propagate_for_debug (first, last, reg, tempreg, |
4170 | this_basic_block); |
4171 | adjust_reg_mode (reg, new_mode); |
4172 | propagate_for_debug (first, last, tempreg, |
4173 | lowpart_subreg (outermode: old_mode, op: reg, innermode: new_mode), |
4174 | this_basic_block); |
4175 | } |
4176 | } |
4177 | } |
4178 | |
4179 | /* If we will be able to accept this, we have made a |
4180 | change to the destination of I3. This requires us to |
4181 | do a few adjustments. */ |
4182 | |
4183 | if (changed_i3_dest) |
4184 | { |
4185 | PATTERN (insn: i3) = newpat; |
4186 | adjust_for_new_dest (insn: i3); |
4187 | } |
4188 | |
4189 | /* We now know that we can do this combination. Merge the insns and |
4190 | update the status of registers and LOG_LINKS. */ |
4191 | |
4192 | if (undobuf.other_insn) |
4193 | { |
4194 | rtx note, next; |
4195 | |
4196 | PATTERN (insn: undobuf.other_insn) = other_pat; |
4197 | |
4198 | /* If any of the notes in OTHER_INSN were REG_DEAD or REG_UNUSED, |
4199 | ensure that they are still valid. Then add any non-duplicate |
4200 | notes added by recog_for_combine. */ |
4201 | for (note = REG_NOTES (undobuf.other_insn); note; note = next) |
4202 | { |
4203 | next = XEXP (note, 1); |
4204 | |
4205 | if ((REG_NOTE_KIND (note) == REG_DEAD |
4206 | && !reg_referenced_p (XEXP (note, 0), |
4207 | PATTERN (insn: undobuf.other_insn))) |
4208 | ||(REG_NOTE_KIND (note) == REG_UNUSED |
4209 | && !reg_set_p (XEXP (note, 0), |
4210 | PATTERN (insn: undobuf.other_insn))) |
4211 | /* Simply drop equal note since it may be no longer valid |
4212 | for other_insn. It may be possible to record that CC |
4213 | register is changed and only discard those notes, but |
4214 | in practice it's unnecessary complication and doesn't |
4215 | give any meaningful improvement. |
4216 | |
4217 | See PR78559. */ |
4218 | || REG_NOTE_KIND (note) == REG_EQUAL |
4219 | || REG_NOTE_KIND (note) == REG_EQUIV) |
4220 | remove_note (undobuf.other_insn, note); |
4221 | } |
4222 | |
4223 | distribute_notes (new_other_notes, undobuf.other_insn, |
4224 | undobuf.other_insn, NULL, NULL_RTX, NULL_RTX, |
4225 | NULL_RTX); |
4226 | } |
4227 | |
4228 | if (swap_i2i3) |
4229 | { |
4230 | /* I3 now uses what used to be its destination and which is now |
4231 | I2's destination. This requires us to do a few adjustments. */ |
4232 | PATTERN (insn: i3) = newpat; |
4233 | adjust_for_new_dest (insn: i3); |
4234 | } |
4235 | |
4236 | if (swap_i2i3 || split_i2i3) |
4237 | { |
4238 | /* We might need a LOG_LINK from I3 to I2. But then we used to |
4239 | have one, so we still will. |
4240 | |
4241 | However, some later insn might be using I2's dest and have |
4242 | a LOG_LINK pointing at I3. We should change it to point at |
4243 | I2 instead. */ |
4244 | |
4245 | /* newi2pat is usually a SET here; however, recog_for_combine might |
4246 | have added some clobbers. */ |
4247 | rtx x = newi2pat; |
4248 | if (GET_CODE (x) == PARALLEL) |
4249 | x = XVECEXP (newi2pat, 0, 0); |
4250 | |
4251 | if (REG_P (SET_DEST (x)) |
4252 | || (GET_CODE (SET_DEST (x)) == SUBREG |
4253 | && REG_P (SUBREG_REG (SET_DEST (x))))) |
4254 | { |
4255 | unsigned int regno = reg_or_subregno (SET_DEST (x)); |
4256 | |
4257 | bool done = false; |
4258 | for (rtx_insn *insn = NEXT_INSN (insn: i3); |
4259 | !done |
4260 | && insn |
4261 | && INSN_P (insn) |
4262 | && BLOCK_FOR_INSN (insn) == this_basic_block; |
4263 | insn = NEXT_INSN (insn)) |
4264 | { |
4265 | if (DEBUG_INSN_P (insn)) |
4266 | continue; |
4267 | struct insn_link *link; |
4268 | FOR_EACH_LOG_LINK (link, insn) |
4269 | if (link->insn == i3 && link->regno == regno) |
4270 | { |
4271 | link->insn = i2; |
4272 | done = true; |
4273 | break; |
4274 | } |
4275 | } |
4276 | } |
4277 | } |
4278 | |
4279 | { |
4280 | rtx i3notes, i2notes, i1notes = 0, i0notes = 0; |
4281 | struct insn_link *i3links, *i2links, *i1links = 0, *i0links = 0; |
4282 | rtx midnotes = 0; |
4283 | int from_luid; |
4284 | /* Compute which registers we expect to eliminate. newi2pat may be setting |
4285 | either i3dest or i2dest, so we must check it. */ |
4286 | rtx elim_i2 = ((newi2pat && reg_set_p (i2dest, newi2pat)) |
4287 | || i2dest_in_i2src || i2dest_in_i1src || i2dest_in_i0src |
4288 | || !i2dest_killed |
4289 | ? 0 : i2dest); |
4290 | /* For i1, we need to compute both local elimination and global |
4291 | elimination information with respect to newi2pat because i1dest |
4292 | may be the same as i3dest, in which case newi2pat may be setting |
4293 | i1dest. Global information is used when distributing REG_DEAD |
4294 | note for i2 and i3, in which case it does matter if newi2pat sets |
4295 | i1dest or not. |
4296 | |
4297 | Local information is used when distributing REG_DEAD note for i1, |
4298 | in which case it doesn't matter if newi2pat sets i1dest or not. |
4299 | See PR62151, if we have four insns combination: |
4300 | i0: r0 <- i0src |
4301 | i1: r1 <- i1src (using r0) |
4302 | REG_DEAD (r0) |
4303 | i2: r0 <- i2src (using r1) |
4304 | i3: r3 <- i3src (using r0) |
4305 | ix: using r0 |
4306 | From i1's point of view, r0 is eliminated, no matter if it is set |
4307 | by newi2pat or not. In other words, REG_DEAD info for r0 in i1 |
4308 | should be discarded. |
4309 | |
4310 | Note local information only affects cases in forms like "I1->I2->I3", |
4311 | "I0->I1->I2->I3" or "I0&I1->I2, I2->I3". For other cases like |
4312 | "I0->I1, I1&I2->I3" or "I1&I2->I3", newi2pat won't set i1dest or |
4313 | i0dest anyway. */ |
4314 | rtx local_elim_i1 = (i1 == 0 || i1dest_in_i1src || i1dest_in_i0src |
4315 | || !i1dest_killed |
4316 | ? 0 : i1dest); |
4317 | rtx elim_i1 = (local_elim_i1 == 0 |
4318 | || (newi2pat && reg_set_p (i1dest, newi2pat)) |
4319 | ? 0 : i1dest); |
4320 | /* Same case as i1. */ |
4321 | rtx local_elim_i0 = (i0 == 0 || i0dest_in_i0src || !i0dest_killed |
4322 | ? 0 : i0dest); |
4323 | rtx elim_i0 = (local_elim_i0 == 0 |
4324 | || (newi2pat && reg_set_p (i0dest, newi2pat)) |
4325 | ? 0 : i0dest); |
4326 | |
4327 | /* Get the old REG_NOTES and LOG_LINKS from all our insns and |
4328 | clear them. */ |
4329 | i3notes = REG_NOTES (i3), i3links = LOG_LINKS (i3); |
4330 | i2notes = REG_NOTES (i2), i2links = LOG_LINKS (i2); |
4331 | if (i1) |
4332 | i1notes = REG_NOTES (i1), i1links = LOG_LINKS (i1); |
4333 | if (i0) |
4334 | i0notes = REG_NOTES (i0), i0links = LOG_LINKS (i0); |
4335 | |
4336 | /* Ensure that we do not have something that should not be shared but |
4337 | occurs multiple times in the new insns. Check this by first |
4338 | resetting all the `used' flags and then copying anything is shared. */ |
4339 | |
4340 | reset_used_flags (i3notes); |
4341 | reset_used_flags (i2notes); |
4342 | reset_used_flags (i1notes); |
4343 | reset_used_flags (i0notes); |
4344 | reset_used_flags (newpat); |
4345 | reset_used_flags (newi2pat); |
4346 | if (undobuf.other_insn) |
4347 | reset_used_flags (PATTERN (insn: undobuf.other_insn)); |
4348 | |
4349 | i3notes = copy_rtx_if_shared (i3notes); |
4350 | i2notes = copy_rtx_if_shared (i2notes); |
4351 | i1notes = copy_rtx_if_shared (i1notes); |
4352 | i0notes = copy_rtx_if_shared (i0notes); |
4353 | newpat = copy_rtx_if_shared (newpat); |
4354 | newi2pat = copy_rtx_if_shared (newi2pat); |
4355 | if (undobuf.other_insn) |
4356 | reset_used_flags (PATTERN (insn: undobuf.other_insn)); |
4357 | |
4358 | INSN_CODE (i3) = insn_code_number; |
4359 | PATTERN (insn: i3) = newpat; |
4360 | |
4361 | if (CALL_P (i3) && CALL_INSN_FUNCTION_USAGE (i3)) |
4362 | { |
4363 | for (rtx link = CALL_INSN_FUNCTION_USAGE (i3); link; |
4364 | link = XEXP (link, 1)) |
4365 | { |
4366 | if (substed_i2) |
4367 | { |
4368 | /* I2SRC must still be meaningful at this point. Some |
4369 | splitting operations can invalidate I2SRC, but those |
4370 | operations do not apply to calls. */ |
4371 | gcc_assert (i2src); |
4372 | XEXP (link, 0) = simplify_replace_rtx (XEXP (link, 0), |
4373 | i2dest, i2src); |
4374 | } |
4375 | if (substed_i1) |
4376 | XEXP (link, 0) = simplify_replace_rtx (XEXP (link, 0), |
4377 | i1dest, i1src); |
4378 | if (substed_i0) |
4379 | XEXP (link, 0) = simplify_replace_rtx (XEXP (link, 0), |
4380 | i0dest, i0src); |
4381 | } |
4382 | } |
4383 | |
4384 | if (undobuf.other_insn) |
4385 | INSN_CODE (undobuf.other_insn) = other_code_number; |
4386 | |
4387 | /* We had one special case above where I2 had more than one set and |
4388 | we replaced a destination of one of those sets with the destination |
4389 | of I3. In that case, we have to update LOG_LINKS of insns later |
4390 | in this basic block. Note that this (expensive) case is rare. |
4391 | |
4392 | Also, in this case, we must pretend that all REG_NOTEs for I2 |
4393 | actually came from I3, so that REG_UNUSED notes from I2 will be |
4394 | properly handled. */ |
4395 | |
4396 | if (i3_subst_into_i2) |
4397 | { |
4398 | for (i = 0; i < XVECLEN (PATTERN (i2), 0); i++) |
4399 | if ((GET_CODE (XVECEXP (PATTERN (i2), 0, i)) == SET |
4400 | || GET_CODE (XVECEXP (PATTERN (i2), 0, i)) == CLOBBER) |
4401 | && REG_P (SET_DEST (XVECEXP (PATTERN (i2), 0, i))) |
4402 | && SET_DEST (XVECEXP (PATTERN (i2), 0, i)) != i2dest |
4403 | && ! find_reg_note (i2, REG_UNUSED, |
4404 | SET_DEST (XVECEXP (PATTERN (i2), 0, i)))) |
4405 | for (temp_insn = NEXT_INSN (insn: i2); |
4406 | temp_insn |
4407 | && (this_basic_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) |
4408 | || BB_HEAD (this_basic_block) != temp_insn); |
4409 | temp_insn = NEXT_INSN (insn: temp_insn)) |
4410 | if (temp_insn != i3 && NONDEBUG_INSN_P (temp_insn)) |
4411 | FOR_EACH_LOG_LINK (link, temp_insn) |
4412 | if (link->insn == i2) |
4413 | link->insn = i3; |
4414 | |
4415 | if (i3notes) |
4416 | { |
4417 | rtx link = i3notes; |
4418 | while (XEXP (link, 1)) |
4419 | link = XEXP (link, 1); |
4420 | XEXP (link, 1) = i2notes; |
4421 | } |
4422 | else |
4423 | i3notes = i2notes; |
4424 | i2notes = 0; |
4425 | } |
4426 | |
4427 | LOG_LINKS (i3) = NULL; |
4428 | REG_NOTES (i3) = 0; |
4429 | LOG_LINKS (i2) = NULL; |
4430 | REG_NOTES (i2) = 0; |
4431 | |
4432 | if (newi2pat) |
4433 | { |
4434 | if (MAY_HAVE_DEBUG_BIND_INSNS && i2scratch) |
4435 | propagate_for_debug (i2, last_combined_insn, i2dest, i2src, |
4436 | this_basic_block); |
4437 | INSN_CODE (i2) = i2_code_number; |
4438 | PATTERN (insn: i2) = newi2pat; |
4439 | } |
4440 | else |
4441 | { |
4442 | if (MAY_HAVE_DEBUG_BIND_INSNS && i2src) |
4443 | propagate_for_debug (i2, last_combined_insn, i2dest, i2src, |
4444 | this_basic_block); |
4445 | SET_INSN_DELETED (i2); |
4446 | } |
4447 | |
4448 | if (i1) |
4449 | { |
4450 | LOG_LINKS (i1) = NULL; |
4451 | REG_NOTES (i1) = 0; |
4452 | if (MAY_HAVE_DEBUG_BIND_INSNS) |
4453 | propagate_for_debug (i1, last_combined_insn, i1dest, i1src, |
4454 | this_basic_block); |
4455 | SET_INSN_DELETED (i1); |
4456 | } |
4457 | |
4458 | if (i0) |
4459 | { |
4460 | LOG_LINKS (i0) = NULL; |
4461 | REG_NOTES (i0) = 0; |
4462 | if (MAY_HAVE_DEBUG_BIND_INSNS) |
4463 | propagate_for_debug (i0, last_combined_insn, i0dest, i0src, |
4464 | this_basic_block); |
4465 | SET_INSN_DELETED (i0); |
4466 | } |
4467 | |
4468 | /* Get death notes for everything that is now used in either I3 or |
4469 | I2 and used to die in a previous insn. If we built two new |
4470 | patterns, move from I1 to I2 then I2 to I3 so that we get the |
4471 | proper movement on registers that I2 modifies. */ |
4472 | |
4473 | if (i0) |
4474 | from_luid = DF_INSN_LUID (i0); |
4475 | else if (i1) |
4476 | from_luid = DF_INSN_LUID (i1); |
4477 | else |
4478 | from_luid = DF_INSN_LUID (i2); |
4479 | if (newi2pat) |
4480 | move_deaths (newi2pat, NULL_RTX, from_luid, i2, &midnotes); |
4481 | move_deaths (newpat, newi2pat, from_luid, i3, &midnotes); |
4482 | |
4483 | /* Distribute all the LOG_LINKS and REG_NOTES from I1, I2, and I3. */ |
4484 | if (i3notes) |
4485 | distribute_notes (i3notes, i3, i3, newi2pat ? i2 : NULL, |
4486 | elim_i2, elim_i1, elim_i0); |
4487 | if (i2notes) |
4488 | distribute_notes (i2notes, i2, i3, newi2pat ? i2 : NULL, |
4489 | elim_i2, elim_i1, elim_i0); |
4490 | if (i1notes) |
4491 | distribute_notes (i1notes, i1, i3, newi2pat ? i2 : NULL, |
4492 | elim_i2, local_elim_i1, local_elim_i0); |
4493 | if (i0notes) |
4494 | distribute_notes (i0notes, i0, i3, newi2pat ? i2 : NULL, |
4495 | elim_i2, elim_i1, local_elim_i0); |
4496 | if (midnotes) |
4497 | distribute_notes (midnotes, NULL, i3, newi2pat ? i2 : NULL, |
4498 | elim_i2, elim_i1, elim_i0); |
4499 | |
4500 | /* Distribute any notes added to I2 or I3 by recog_for_combine. We |
4501 | know these are REG_UNUSED and want them to go to the desired insn, |
4502 | so we always pass it as i3. */ |
4503 | |
4504 | if (newi2pat && new_i2_notes) |
4505 | distribute_notes (new_i2_notes, i2, i2, NULL, NULL_RTX, NULL_RTX, |
4506 | NULL_RTX); |
4507 | |
4508 | if (new_i3_notes) |
4509 | distribute_notes (new_i3_notes, i3, i3, NULL, NULL_RTX, NULL_RTX, |
4510 | NULL_RTX); |
4511 | |
4512 | /* If I3DEST was used in I3SRC, it really died in I3. We may need to |
4513 | put a REG_DEAD note for it somewhere. If NEWI2PAT exists and sets |
4514 | I3DEST, the death must be somewhere before I2, not I3. If we passed I3 |
4515 | in that case, it might delete I2. Similarly for I2 and I1. |
4516 | Show an additional death due to the REG_DEAD note we make here. If |
4517 | we discard it in distribute_notes, we will decrement it again. */ |
4518 | |
4519 | if (i3dest_killed) |
4520 | { |
4521 | rtx new_note = alloc_reg_note (REG_DEAD, i3dest_killed, NULL_RTX); |
4522 | if (newi2pat && reg_set_p (i3dest_killed, newi2pat)) |
4523 | distribute_notes (new_note, NULL, i2, NULL, elim_i2, |
4524 | elim_i1, elim_i0); |
4525 | else |
4526 | distribute_notes (new_note, NULL, i3, newi2pat ? i2 : NULL, |
4527 | elim_i2, elim_i1, elim_i0); |
4528 | } |
4529 | |
4530 | if (i2dest_in_i2src) |
4531 | { |
4532 | rtx new_note = alloc_reg_note (REG_DEAD, i2dest, NULL_RTX); |
4533 | if (newi2pat && reg_set_p (i2dest, newi2pat)) |
4534 | distribute_notes (new_note, NULL, i2, NULL, NULL_RTX, |
4535 | NULL_RTX, NULL_RTX); |
4536 | else |
4537 | distribute_notes (new_note, NULL, i3, newi2pat ? i2 : NULL, |
4538 | NULL_RTX, NULL_RTX, NULL_RTX); |
4539 | } |
4540 | |
4541 | if (i1dest_in_i1src) |
4542 | { |
4543 | rtx new_note = alloc_reg_note (REG_DEAD, i1dest, NULL_RTX); |
4544 | if (newi2pat && reg_set_p (i1dest, newi2pat)) |
4545 | distribute_notes (new_note, NULL, i2, NULL, NULL_RTX, |
4546 | NULL_RTX, NULL_RTX); |
4547 | else |
4548 | distribute_notes (new_note, NULL, i3, newi2pat ? i2 : NULL, |
4549 | NULL_RTX, NULL_RTX, NULL_RTX); |
4550 | } |
4551 | |
4552 | if (i0dest_in_i0src) |
4553 | { |
4554 | rtx new_note = alloc_reg_note (REG_DEAD, i0dest, NULL_RTX); |
4555 | if (newi2pat && reg_set_p (i0dest, newi2pat)) |
4556 | distribute_notes (new_note, NULL, i2, NULL, NULL_RTX, |
4557 | NULL_RTX, NULL_RTX); |
4558 | else |
4559 | distribute_notes (new_note, NULL, i3, newi2pat ? i2 : NULL, |
4560 | NULL_RTX, NULL_RTX, NULL_RTX); |
4561 | } |
4562 | |
4563 | distribute_links (i3links); |
4564 | distribute_links (i2links); |
4565 | distribute_links (i1links); |
4566 | distribute_links (i0links); |
4567 | |
4568 | if (REG_P (i2dest)) |
4569 | { |
4570 | struct insn_link *link; |
4571 | rtx_insn *i2_insn = 0; |
4572 | rtx i2_val = 0, set; |
4573 | |
4574 | /* The insn that used to set this register doesn't exist, and |
4575 | this life of the register may not exist either. See if one of |
4576 | I3's links points to an insn that sets I2DEST. If it does, |
4577 | that is now the last known value for I2DEST. If we don't update |
4578 | this and I2 set the register to a value that depended on its old |
4579 | contents, we will get confused. If this insn is used, thing |
4580 | will be set correctly in combine_instructions. */ |
4581 | FOR_EACH_LOG_LINK (link, i3) |
4582 | if ((set = single_set (insn: link->insn)) != 0 |
4583 | && rtx_equal_p (i2dest, SET_DEST (set))) |
4584 | i2_insn = link->insn, i2_val = SET_SRC (set); |
4585 | |
4586 | record_value_for_reg (i2dest, i2_insn, i2_val); |
4587 | |
4588 | /* If the reg formerly set in I2 died only once and that was in I3, |
4589 | zero its use count so it won't make `reload' do any work. */ |
4590 | if (! added_sets_2 |
4591 | && (newi2pat == 0 || ! reg_mentioned_p (i2dest, newi2pat)) |
4592 | && ! i2dest_in_i2src |
4593 | && REGNO (i2dest) < reg_n_sets_max) |
4594 | INC_REG_N_SETS (REGNO (i2dest), -1); |
4595 | } |
4596 | |
4597 | if (i1 && REG_P (i1dest)) |
4598 | { |
4599 | struct insn_link *link; |
4600 | rtx_insn *i1_insn = 0; |
4601 | rtx i1_val = 0, set; |
4602 | |
4603 | FOR_EACH_LOG_LINK (link, i3) |
4604 | if ((set = single_set (insn: link->insn)) != 0 |
4605 | && rtx_equal_p (i1dest, SET_DEST (set))) |
4606 | i1_insn = link->insn, i1_val = SET_SRC (set); |
4607 | |
4608 | record_value_for_reg (i1dest, i1_insn, i1_val); |
4609 | |
4610 | if (! added_sets_1 |
4611 | && ! i1dest_in_i1src |
4612 | && REGNO (i1dest) < reg_n_sets_max) |
4613 | INC_REG_N_SETS (REGNO (i1dest), -1); |
4614 | } |
4615 | |
4616 | if (i0 && REG_P (i0dest)) |
4617 | { |
4618 | struct insn_link *link; |
4619 | rtx_insn *i0_insn = 0; |
4620 | rtx i0_val = 0, set; |
4621 | |
4622 | FOR_EACH_LOG_LINK (link, i3) |
4623 | if ((set = single_set (insn: link->insn)) != 0 |
4624 | && rtx_equal_p (i0dest, SET_DEST (set))) |
4625 | i0_insn = link->insn, i0_val = SET_SRC (set); |
4626 | |
4627 | record_value_for_reg (i0dest, i0_insn, i0_val); |
4628 | |
4629 | if (! added_sets_0 |
4630 | && ! i0dest_in_i0src |
4631 | && REGNO (i0dest) < reg_n_sets_max) |
4632 | INC_REG_N_SETS (REGNO (i0dest), -1); |
4633 | } |
4634 | |
4635 | /* Update reg_stat[].nonzero_bits et al for any changes that may have |
4636 | been made to this insn. The order is important, because newi2pat |
4637 | can affect nonzero_bits of newpat. */ |
4638 | if (newi2pat) |
4639 | note_pattern_stores (newi2pat, set_nonzero_bits_and_sign_copies, NULL); |
4640 | note_pattern_stores (newpat, set_nonzero_bits_and_sign_copies, NULL); |
4641 | } |
4642 | |
4643 | if (undobuf.other_insn != NULL_RTX) |
4644 | { |
4645 | if (dump_file) |
4646 | { |
4647 | fprintf (stream: dump_file, format: "modifying other_insn " ); |
4648 | dump_insn_slim (dump_file, undobuf.other_insn); |
4649 | } |
4650 | df_insn_rescan (undobuf.other_insn); |
4651 | } |
4652 | |
4653 | if (i0 && !(NOTE_P (i0) && (NOTE_KIND (i0) == NOTE_INSN_DELETED))) |
4654 | { |
4655 | if (dump_file) |
4656 | { |
4657 | fprintf (stream: dump_file, format: "modifying insn i0 " ); |
4658 | dump_insn_slim (dump_file, i0); |
4659 | } |
4660 | df_insn_rescan (i0); |
4661 | } |
4662 | |
4663 | if (i1 && !(NOTE_P (i1) && (NOTE_KIND (i1) == NOTE_INSN_DELETED))) |
4664 | { |
4665 | if (dump_file) |
4666 | { |
4667 | fprintf (stream: dump_file, format: "modifying insn i1 " ); |
4668 | dump_insn_slim (dump_file, i1); |
4669 | } |
4670 | df_insn_rescan (i1); |
4671 | } |
4672 | |
4673 | if (i2 && !(NOTE_P (i2) && (NOTE_KIND (i2) == NOTE_INSN_DELETED))) |
4674 | { |
4675 | if (dump_file) |
4676 | { |
4677 | fprintf (stream: dump_file, format: "modifying insn i2 " ); |
4678 | dump_insn_slim (dump_file, i2); |
4679 | } |
4680 | df_insn_rescan (i2); |
4681 | } |
4682 | |
4683 | if (i3 && !(NOTE_P (i3) && (NOTE_KIND (i3) == NOTE_INSN_DELETED))) |
4684 | { |
4685 | if (dump_file) |
4686 | { |
4687 | fprintf (stream: dump_file, format: "modifying insn i3 " ); |
4688 | dump_insn_slim (dump_file, i3); |
4689 | } |
4690 | df_insn_rescan (i3); |
4691 | } |
4692 | |
4693 | /* Set new_direct_jump_p if a new return or simple jump instruction |
4694 | has been created. Adjust the CFG accordingly. */ |
4695 | if (returnjump_p (i3) || any_uncondjump_p (i3)) |
4696 | { |
4697 | *new_direct_jump_p = 1; |
4698 | mark_jump_label (PATTERN (insn: i3), i3, 0); |
4699 | update_cfg_for_uncondjump (i3); |
4700 | } |
4701 | |
4702 | if (undobuf.other_insn != NULL_RTX |
4703 | && (returnjump_p (undobuf.other_insn) |
4704 | || any_uncondjump_p (undobuf.other_insn))) |
4705 | { |
4706 | *new_direct_jump_p = 1; |
4707 | update_cfg_for_uncondjump (undobuf.other_insn); |
4708 | } |
4709 | |
4710 | if (GET_CODE (PATTERN (i3)) == TRAP_IF |
4711 | && XEXP (PATTERN (i3), 0) == const1_rtx) |
4712 | { |
4713 | basic_block bb = BLOCK_FOR_INSN (insn: i3); |
4714 | gcc_assert (bb); |
4715 | remove_edge (split_block (bb, i3)); |
4716 | emit_barrier_after_bb (bb); |
4717 | *new_direct_jump_p = 1; |
4718 | } |
4719 | |
4720 | if (undobuf.other_insn |
4721 | && GET_CODE (PATTERN (undobuf.other_insn)) == TRAP_IF |
4722 | && XEXP (PATTERN (undobuf.other_insn), 0) == const1_rtx) |
4723 | { |
4724 | basic_block bb = BLOCK_FOR_INSN (insn: undobuf.other_insn); |
4725 | gcc_assert (bb); |
4726 | remove_edge (split_block (bb, undobuf.other_insn)); |
4727 | emit_barrier_after_bb (bb); |
4728 | *new_direct_jump_p = 1; |
4729 | } |
4730 | |
4731 | /* A noop might also need cleaning up of CFG, if it comes from the |
4732 | simplification of a jump. */ |
4733 | if (JUMP_P (i3) |
4734 | && GET_CODE (newpat) == SET |
4735 | && SET_SRC (newpat) == pc_rtx |
4736 | && SET_DEST (newpat) == pc_rtx) |
4737 | { |
4738 | *new_direct_jump_p = 1; |
4739 | update_cfg_for_uncondjump (i3); |
4740 | } |
4741 | |
4742 | if (undobuf.other_insn != NULL_RTX |
4743 | && JUMP_P (undobuf.other_insn) |
4744 | && GET_CODE (PATTERN (undobuf.other_insn)) == SET |
4745 | && SET_SRC (PATTERN (undobuf.other_insn)) == pc_rtx |
4746 | && SET_DEST (PATTERN (undobuf.other_insn)) == pc_rtx) |
4747 | { |
4748 | *new_direct_jump_p = 1; |
4749 | update_cfg_for_uncondjump (undobuf.other_insn); |
4750 | } |
4751 | |
4752 | combine_successes++; |
4753 | undo_commit (); |
4754 | |
4755 | rtx_insn *ret = newi2pat ? i2 : i3; |
4756 | if (added_links_insn && DF_INSN_LUID (added_links_insn) < DF_INSN_LUID (ret)) |
4757 | ret = added_links_insn; |
4758 | if (added_notes_insn && DF_INSN_LUID (added_notes_insn) < DF_INSN_LUID (ret)) |
4759 | ret = added_notes_insn; |
4760 | |
4761 | return ret; |
4762 | } |
4763 | |
4764 | /* Get a marker for undoing to the current state. */ |
4765 | |
4766 | static void * |
4767 | get_undo_marker (void) |
4768 | { |
4769 | return undobuf.undos; |
4770 | } |
4771 | |
4772 | /* Undo the modifications up to the marker. */ |
4773 | |
4774 | static void |
4775 | undo_to_marker (void *marker) |
4776 | { |
4777 | struct undo *undo, *next; |
4778 | |
4779 | for (undo = undobuf.undos; undo != marker; undo = next) |
4780 | { |
4781 | gcc_assert (undo); |
4782 | |
4783 | next = undo->next; |
4784 | switch (undo->kind) |
4785 | { |
4786 | case UNDO_RTX: |
4787 | *undo->where.r = undo->old_contents.r; |
4788 | break; |
4789 | case UNDO_INT: |
4790 | *undo->where.i = undo->old_contents.i; |
4791 | break; |
4792 | case UNDO_MODE: |
4793 | adjust_reg_mode (regno_reg_rtx[undo->where.regno], |
4794 | undo->old_contents.m); |
4795 | break; |
4796 | case UNDO_LINKS: |
4797 | *undo->where.l = undo->old_contents.l; |
4798 | break; |
4799 | default: |
4800 | gcc_unreachable (); |
4801 | } |
4802 | |
4803 | undo->next = undobuf.frees; |
4804 | undobuf.frees = undo; |
4805 | } |
4806 | |
4807 | undobuf.undos = (struct undo *) marker; |
4808 | } |
4809 | |
4810 | /* Undo all the modifications recorded in undobuf. */ |
4811 | |
4812 | static void |
4813 | undo_all (void) |
4814 | { |
4815 | undo_to_marker (marker: 0); |
4816 | } |
4817 | |
4818 | /* We've committed to accepting the changes we made. Move all |
4819 | of the undos to the free list. */ |
4820 | |
4821 | static void |
4822 | undo_commit (void) |
4823 | { |
4824 | struct undo *undo, *next; |
4825 | |
4826 | for (undo = undobuf.undos; undo; undo = next) |
4827 | { |
4828 | next = undo->next; |
4829 | undo->next = undobuf.frees; |
4830 | undobuf.frees = undo; |
4831 | } |
4832 | undobuf.undos = 0; |
4833 | } |
4834 | |
4835 | /* Find the innermost point within the rtx at LOC, possibly LOC itself, |
4836 | where we have an arithmetic expression and return that point. LOC will |
4837 | be inside INSN. |
4838 | |
4839 | try_combine will call this function to see if an insn can be split into |
4840 | two insns. */ |
4841 | |
4842 | static rtx * |
4843 | find_split_point (rtx *loc, rtx_insn *insn, bool set_src) |
4844 | { |
4845 | rtx x = *loc; |
4846 | enum rtx_code code = GET_CODE (x); |
4847 | rtx *split; |
4848 | unsigned HOST_WIDE_INT len = 0; |
4849 | HOST_WIDE_INT pos = 0; |
4850 | bool unsignedp = false; |
4851 | rtx inner = NULL_RTX; |
4852 | scalar_int_mode mode, inner_mode; |
4853 | |
4854 | /* First special-case some codes. */ |
4855 | switch (code) |
4856 | { |
4857 | case SUBREG: |
4858 | #ifdef INSN_SCHEDULING |
4859 | /* If we are making a paradoxical SUBREG invalid, it becomes a split |
4860 | point. */ |
4861 | if (MEM_P (SUBREG_REG (x))) |
4862 | return loc; |
4863 | #endif |
4864 | return find_split_point (loc: &SUBREG_REG (x), insn, set_src: false); |
4865 | |
4866 | case MEM: |
4867 | /* If we have (mem (const ..)) or (mem (symbol_ref ...)), split it |
4868 | using LO_SUM and HIGH. */ |
4869 | if (HAVE_lo_sum && (GET_CODE (XEXP (x, 0)) == CONST |
4870 | || GET_CODE (XEXP (x, 0)) == SYMBOL_REF)) |
4871 | { |
4872 | machine_mode address_mode = get_address_mode (mem: x); |
4873 | |
4874 | SUBST (XEXP (x, 0), |
4875 | gen_rtx_LO_SUM (address_mode, |
4876 | gen_rtx_HIGH (address_mode, XEXP (x, 0)), |
4877 | XEXP (x, 0))); |
4878 | return &XEXP (XEXP (x, 0), 0); |
4879 | } |
4880 | |
4881 | /* If we have a PLUS whose second operand is a constant and the |
4882 | address is not valid, perhaps we can split it up using |
4883 | the machine-specific way to split large constants. We use |
4884 | the first pseudo-reg (one of the virtual regs) as a placeholder; |
4885 | it will not remain in the result. */ |
4886 | if (GET_CODE (XEXP (x, 0)) == PLUS |
4887 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) |
4888 | && ! memory_address_addr_space_p (GET_MODE (x), XEXP (x, 0), |
4889 | MEM_ADDR_SPACE (x))) |
4890 | { |
4891 | rtx reg = regno_reg_rtx[FIRST_PSEUDO_REGISTER]; |
4892 | rtx_insn *seq = combine_split_insns (gen_rtx_SET (reg, XEXP (x, 0)), |
4893 | insn: subst_insn); |
4894 | |
4895 | /* This should have produced two insns, each of which sets our |
4896 | placeholder. If the source of the second is a valid address, |
4897 | we can put both sources together and make a split point |
4898 | in the middle. */ |
4899 | |
4900 | if (seq |
4901 | && NEXT_INSN (insn: seq) != NULL_RTX |
4902 | && NEXT_INSN (insn: NEXT_INSN (insn: seq)) == NULL_RTX |
4903 | && NONJUMP_INSN_P (seq) |
4904 | && GET_CODE (PATTERN (seq)) == SET |
4905 | && SET_DEST (PATTERN (seq)) == reg |
4906 | && ! reg_mentioned_p (reg, |
4907 | SET_SRC (PATTERN (seq))) |
4908 | && NONJUMP_INSN_P (NEXT_INSN (seq)) |
4909 | && GET_CODE (PATTERN (NEXT_INSN (seq))) == SET |
4910 | && SET_DEST (PATTERN (NEXT_INSN (seq))) == reg |
4911 | && memory_address_addr_space_p |
4912 | (GET_MODE (x), SET_SRC (PATTERN (NEXT_INSN (seq))), |
4913 | MEM_ADDR_SPACE (x))) |
4914 | { |
4915 | rtx src1 = SET_SRC (PATTERN (seq)); |
4916 | rtx src2 = SET_SRC (PATTERN (NEXT_INSN (seq))); |
4917 | |
4918 | /* Replace the placeholder in SRC2 with SRC1. If we can |
4919 | find where in SRC2 it was placed, that can become our |
4920 | split point and we can replace this address with SRC2. |
4921 | Just try two obvious places. */ |
4922 | |
4923 | src2 = replace_rtx (src2, reg, src1); |
4924 | split = 0; |
4925 | if (XEXP (src2, 0) == src1) |
4926 | split = &XEXP (src2, 0); |
4927 | else if (GET_RTX_FORMAT (GET_CODE (XEXP (src2, 0)))[0] == 'e' |
4928 | && XEXP (XEXP (src2, 0), 0) == src1) |
4929 | split = &XEXP (XEXP (src2, 0), 0); |
4930 | |
4931 | if (split) |
4932 | { |
4933 | SUBST (XEXP (x, 0), src2); |
4934 | return split; |
4935 | } |
4936 | } |
4937 | |
4938 | /* If that didn't work and we have a nested plus, like: |
4939 | ((REG1 * CONST1) + REG2) + CONST2 and (REG1 + REG2) + CONST2 |
4940 | is valid address, try to split (REG1 * CONST1). */ |
4941 | if (GET_CODE (XEXP (XEXP (x, 0), 0)) == PLUS |
4942 | && !OBJECT_P (XEXP (XEXP (XEXP (x, 0), 0), 0)) |
4943 | && OBJECT_P (XEXP (XEXP (XEXP (x, 0), 0), 1)) |
4944 | && ! (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 0)) == SUBREG |
4945 | && OBJECT_P (SUBREG_REG (XEXP (XEXP (XEXP (x, 0), |
4946 | 0), 0))))) |
4947 | { |
4948 | rtx tem = XEXP (XEXP (XEXP (x, 0), 0), 0); |
4949 | XEXP (XEXP (XEXP (x, 0), 0), 0) = reg; |
4950 | if (memory_address_addr_space_p (GET_MODE (x), XEXP (x, 0), |
4951 | MEM_ADDR_SPACE (x))) |
4952 | { |
4953 | XEXP (XEXP (XEXP (x, 0), 0), 0) = tem; |
4954 | return &XEXP (XEXP (XEXP (x, 0), 0), 0); |
4955 | } |
4956 | XEXP (XEXP (XEXP (x, 0), 0), 0) = tem; |
4957 | } |
4958 | else if (GET_CODE (XEXP (XEXP (x, 0), 0)) == PLUS |
4959 | && OBJECT_P (XEXP (XEXP (XEXP (x, 0), 0), 0)) |
4960 | && !OBJECT_P (XEXP (XEXP (XEXP (x, 0), 0), 1)) |
4961 | && ! (GET_CODE (XEXP (XEXP (XEXP (x, 0), 0), 1)) == SUBREG |
4962 | && OBJECT_P (SUBREG_REG (XEXP (XEXP (XEXP (x, 0), |
4963 | 0), 1))))) |
4964 | { |
4965 | rtx tem = XEXP (XEXP (XEXP (x, 0), 0), 1); |
4966 | XEXP (XEXP (XEXP (x, 0), 0), 1) = reg; |
4967 | if (memory_address_addr_space_p (GET_MODE (x), XEXP (x, 0), |
4968 | MEM_ADDR_SPACE (x))) |
4969 | { |
4970 | XEXP (XEXP (XEXP (x, 0), 0), 1) = tem; |
4971 | return &XEXP (XEXP (XEXP (x, 0), 0), 1); |
4972 | } |
4973 | XEXP (XEXP (XEXP (x, 0), 0), 1) = tem; |
4974 | } |
4975 | |
4976 | /* If that didn't work, perhaps the first operand is complex and |
4977 | needs to be computed separately, so make a split point there. |
4978 | This will occur on machines that just support REG + CONST |
4979 | and have a constant moved through some previous computation. */ |
4980 | if (!OBJECT_P (XEXP (XEXP (x, 0), 0)) |
4981 | && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG |
4982 | && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0))))) |
4983 | return &XEXP (XEXP (x, 0), 0); |
4984 | } |
4985 | |
4986 | /* If we have a PLUS whose first operand is complex, try computing it |
4987 | separately by making a split there. */ |
4988 | if (GET_CODE (XEXP (x, 0)) == PLUS |
4989 | && ! memory_address_addr_space_p (GET_MODE (x), XEXP (x, 0), |
4990 | MEM_ADDR_SPACE (x)) |
4991 | && ! OBJECT_P (XEXP (XEXP (x, 0), 0)) |
4992 | && ! (GET_CODE (XEXP (XEXP (x, 0), 0)) == SUBREG |
4993 | && OBJECT_P (SUBREG_REG (XEXP (XEXP (x, 0), 0))))) |
4994 | return &XEXP (XEXP (x, 0), 0); |
4995 | break; |
4996 | |
4997 | case SET: |
4998 | /* See if we can split SET_SRC as it stands. */ |
4999 | split = find_split_point (loc: &SET_SRC (x), insn, set_src: true); |
5000 | if (split && split != &SET_SRC (x)) |
5001 | return split; |
5002 | |
5003 | /* See if we can split SET_DEST as it stands. */ |
5004 | split = find_split_point (loc: &SET_DEST (x), insn, set_src: false); |
5005 | if (split && split != &SET_DEST (x)) |
5006 | return split; |
5007 | |
5008 | /* See if this is a bitfield assignment with everything constant. If |
5009 | so, this is an IOR of an AND, so split it into that. */ |
5010 | if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT |
5011 | && is_a <scalar_int_mode> (GET_MODE (XEXP (SET_DEST (x), 0)), |
5012 | result: &inner_mode) |
5013 | && HWI_COMPUTABLE_MODE_P (mode: inner_mode) |
5014 | && CONST_INT_P (XEXP (SET_DEST (x), 1)) |
5015 | && CONST_INT_P (XEXP (SET_DEST (x), 2)) |
5016 | && CONST_INT_P (SET_SRC (x)) |
5017 | && ((INTVAL (XEXP (SET_DEST (x), 1)) |
5018 | + INTVAL (XEXP (SET_DEST (x), 2))) |
5019 | <= GET_MODE_PRECISION (mode: inner_mode)) |
5020 | && ! side_effects_p (XEXP (SET_DEST (x), 0))) |
5021 | { |
5022 | HOST_WIDE_INT pos = INTVAL (XEXP (SET_DEST (x), 2)); |
5023 | unsigned HOST_WIDE_INT len = INTVAL (XEXP (SET_DEST (x), 1)); |
5024 | rtx dest = XEXP (SET_DEST (x), 0); |
5025 | unsigned HOST_WIDE_INT mask = (HOST_WIDE_INT_1U << len) - 1; |
5026 | unsigned HOST_WIDE_INT src = INTVAL (SET_SRC (x)) & mask; |
5027 | rtx or_mask; |
5028 | |
5029 | if (BITS_BIG_ENDIAN) |
5030 | pos = GET_MODE_PRECISION (mode: inner_mode) - len - pos; |
5031 | |
5032 | or_mask = gen_int_mode (src << pos, inner_mode); |
5033 | if (src == mask) |
5034 | SUBST (SET_SRC (x), |
5035 | simplify_gen_binary (IOR, inner_mode, dest, or_mask)); |
5036 | else |
5037 | { |
5038 | rtx negmask = gen_int_mode (~(mask << pos), inner_mode); |
5039 | SUBST (SET_SRC (x), |
5040 | simplify_gen_binary (IOR, inner_mode, |
5041 | simplify_gen_binary (AND, inner_mode, |
5042 | dest, negmask), |
5043 | or_mask)); |
5044 | } |
5045 | |
5046 | SUBST (SET_DEST (x), dest); |
5047 | |
5048 | split = find_split_point (loc: &SET_SRC (x), insn, set_src: true); |
5049 | if (split && split != &SET_SRC (x)) |
5050 | return split; |
5051 | } |
5052 | |
5053 | /* Otherwise, see if this is an operation that we can split into two. |
5054 | If so, try to split that. */ |
5055 | code = GET_CODE (SET_SRC (x)); |
5056 | |
5057 | switch (code) |
5058 | { |
5059 | case AND: |
5060 | /* If we are AND'ing with a large constant that is only a single |
5061 | bit and the result is only being used in a context where we |
5062 | need to know if it is zero or nonzero, replace it with a bit |
5063 | extraction. This will avoid the large constant, which might |
5064 | have taken more than one insn to make. If the constant were |
5065 | not a valid argument to the AND but took only one insn to make, |
5066 | this is no worse, but if it took more than one insn, it will |
5067 | be better. */ |
5068 | |
5069 | if (CONST_INT_P (XEXP (SET_SRC (x), 1)) |
5070 | && REG_P (XEXP (SET_SRC (x), 0)) |
5071 | && (pos = exact_log2 (UINTVAL (XEXP (SET_SRC (x), 1)))) >= 7 |
5072 | && REG_P (SET_DEST (x)) |
5073 | && (split = find_single_use (SET_DEST (x), insn, NULL)) != 0 |
5074 | && (GET_CODE (*split) == EQ || GET_CODE (*split) == NE) |
5075 | && XEXP (*split, 0) == SET_DEST (x) |
5076 | && XEXP (*split, 1) == const0_rtx) |
5077 | { |
5078 | rtx = make_extraction (GET_MODE (SET_DEST (x)), |
5079 | XEXP (SET_SRC (x), 0), |
5080 | pos, NULL_RTX, 1, |
5081 | true, false, false); |
5082 | if (extraction != 0) |
5083 | { |
5084 | SUBST (SET_SRC (x), extraction); |
5085 | return find_split_point (loc, insn, set_src: false); |
5086 | } |
5087 | } |
5088 | break; |
5089 | |
5090 | case NE: |
5091 | /* If STORE_FLAG_VALUE is -1, this is (NE X 0) and only one bit of X |
5092 | is known to be on, this can be converted into a NEG of a shift. */ |
5093 | if (STORE_FLAG_VALUE == -1 && XEXP (SET_SRC (x), 1) == const0_rtx |
5094 | && GET_MODE (SET_SRC (x)) == GET_MODE (XEXP (SET_SRC (x), 0)) |
5095 | && ((pos = exact_log2 (x: nonzero_bits (XEXP (SET_SRC (x), 0), |
5096 | GET_MODE (XEXP (SET_SRC (x), |
5097 | 0))))) >= 1)) |
5098 | { |
5099 | machine_mode mode = GET_MODE (XEXP (SET_SRC (x), 0)); |
5100 | rtx pos_rtx = gen_int_shift_amount (mode, pos); |
5101 | SUBST (SET_SRC (x), |
5102 | gen_rtx_NEG (mode, |
5103 | gen_rtx_LSHIFTRT (mode, |
5104 | XEXP (SET_SRC (x), 0), |
5105 | pos_rtx))); |
5106 | |
5107 | split = find_split_point (loc: &SET_SRC (x), insn, set_src: true); |
5108 | if (split && split != &SET_SRC (x)) |
5109 | return split; |
5110 | } |
5111 | break; |
5112 | |
5113 | case SIGN_EXTEND: |
5114 | inner = XEXP (SET_SRC (x), 0); |
5115 | |
5116 | /* We can't optimize if either mode is a partial integer |
5117 | mode as we don't know how many bits are significant |
5118 | in those modes. */ |
5119 | if (!is_int_mode (GET_MODE (inner), int_mode: &inner_mode) |
5120 | || GET_MODE_CLASS (GET_MODE (SET_SRC (x))) == MODE_PARTIAL_INT) |
5121 | break; |
5122 | |
5123 | pos = 0; |
5124 | len = GET_MODE_PRECISION (mode: inner_mode); |
5125 | unsignedp = false; |
5126 | break; |
5127 | |
5128 | case SIGN_EXTRACT: |
5129 | case ZERO_EXTRACT: |
5130 | if (is_a <scalar_int_mode> (GET_MODE (XEXP (SET_SRC (x), 0)), |
5131 | result: &inner_mode) |
5132 | && CONST_INT_P (XEXP (SET_SRC (x), 1)) |
5133 | && CONST_INT_P (XEXP (SET_SRC (x), 2))) |
5134 | { |
5135 | inner = XEXP (SET_SRC (x), 0); |
5136 | len = INTVAL (XEXP (SET_SRC (x), 1)); |
5137 | pos = INTVAL (XEXP (SET_SRC (x), 2)); |
5138 | |
5139 | if (BITS_BIG_ENDIAN) |
5140 | pos = GET_MODE_PRECISION (mode: inner_mode) - len - pos; |
5141 | unsignedp = (code == ZERO_EXTRACT); |
5142 | } |
5143 | break; |
5144 | |
5145 | default: |
5146 | break; |
5147 | } |
5148 | |
5149 | if (len |
5150 | && known_subrange_p (pos1: pos, size1: len, |
5151 | pos2: 0, size2: GET_MODE_PRECISION (GET_MODE (inner))) |
5152 | && is_a <scalar_int_mode> (GET_MODE (SET_SRC (x)), result: &mode)) |
5153 | { |
5154 | /* For unsigned, we have a choice of a shift followed by an |
5155 | AND or two shifts. Use two shifts for field sizes where the |
5156 | constant might be too large. We assume here that we can |
5157 | always at least get 8-bit constants in an AND insn, which is |
5158 | true for every current RISC. */ |
5159 | |
5160 | if (unsignedp && len <= 8) |
5161 | { |
5162 | unsigned HOST_WIDE_INT mask |
5163 | = (HOST_WIDE_INT_1U << len) - 1; |
5164 | rtx pos_rtx = gen_int_shift_amount (mode, pos); |
5165 | SUBST (SET_SRC (x), |
5166 | gen_rtx_AND (mode, |
5167 | gen_rtx_LSHIFTRT |
5168 | (mode, gen_lowpart (mode, inner), pos_rtx), |
5169 | gen_int_mode (mask, mode))); |
5170 | |
5171 | split = find_split_point (loc: &SET_SRC (x), insn, set_src: true); |
5172 | if (split && split != &SET_SRC (x)) |
5173 | return split; |
5174 | } |
5175 | else |
5176 | { |
5177 | int left_bits = GET_MODE_PRECISION (mode) - len - pos; |
5178 | int right_bits = GET_MODE_PRECISION (mode) - len; |
5179 | SUBST (SET_SRC (x), |
5180 | gen_rtx_fmt_ee |
5181 | (unsignedp ? LSHIFTRT : ASHIFTRT, mode, |
5182 | gen_rtx_ASHIFT (mode, |
5183 | gen_lowpart (mode, inner), |
5184 | gen_int_shift_amount (mode, left_bits)), |
5185 | gen_int_shift_amount (mode, right_bits))); |
5186 | |
5187 | split = find_split_point (loc: &SET_SRC (x), insn, set_src: true); |
5188 | if (split && split != &SET_SRC (x)) |
5189 | return split; |
5190 | } |
5191 | } |
5192 | |
5193 | /* See if this is a simple operation with a constant as the second |
5194 | operand. It might be that this constant is out of range and hence |
5195 | could be used as a split point. */ |
5196 | if (BINARY_P (SET_SRC (x)) |
5197 | && CONSTANT_P (XEXP (SET_SRC (x), 1)) |
5198 | && (OBJECT_P (XEXP (SET_SRC (x), 0)) |
5199 | || (GET_CODE (XEXP (SET_SRC (x), 0)) == SUBREG |
5200 | && OBJECT_P (SUBREG_REG (XEXP (SET_SRC (x), 0)))))) |
5201 | return &XEXP (SET_SRC (x), 1); |
5202 | |
5203 | /* Finally, see if this is a simple operation with its first operand |
5204 | not in a register. The operation might require this operand in a |
5205 | register, so return it as a split point. We can always do this |
5206 | because if the first operand were another operation, we would have |
5207 | already found it as a split point. */ |
5208 | if ((BINARY_P (SET_SRC (x)) || UNARY_P (SET_SRC (x))) |
5209 | && ! register_operand (XEXP (SET_SRC (x), 0), VOIDmode)) |
5210 | return &XEXP (SET_SRC (x), 0); |
5211 | |
5212 | return 0; |
5213 | |
5214 | case AND: |
5215 | case IOR: |
5216 | /* We write NOR as (and (not A) (not B)), but if we don't have a NOR, |
5217 | it is better to write this as (not (ior A B)) so we can split it. |
5218 | Similarly for IOR. */ |
5219 | if (GET_CODE (XEXP (x, 0)) == NOT && GET_CODE (XEXP (x, 1)) == NOT) |
5220 | { |
5221 | SUBST (*loc, |
5222 | gen_rtx_NOT (GET_MODE (x), |
5223 | gen_rtx_fmt_ee (code == IOR ? AND : IOR, |
5224 | GET_MODE (x), |
5225 | XEXP (XEXP (x, 0), 0), |
5226 | XEXP (XEXP (x, 1), 0)))); |
5227 | return find_split_point (loc, insn, set_src); |
5228 | } |
5229 | |
5230 | /* Many RISC machines have a large set of logical insns. If the |
5231 | second operand is a NOT, put it first so we will try to split the |
5232 | other operand first. */ |
5233 | if (GET_CODE (XEXP (x, 1)) == NOT) |
5234 | { |
5235 | rtx tem = XEXP (x, 0); |
5236 | SUBST (XEXP (x, 0), XEXP (x, 1)); |
5237 | SUBST (XEXP (x, 1), tem); |
5238 | } |
5239 | break; |
5240 | |
5241 | case PLUS: |
5242 | case MINUS: |
5243 | /* Canonicalization can produce (minus A (mult B C)), where C is a |
5244 | constant. It may be better to try splitting (plus (mult B -C) A) |
5245 | instead if this isn't a multiply by a power of two. */ |
5246 | if (set_src && code == MINUS && GET_CODE (XEXP (x, 1)) == MULT |
5247 | && GET_CODE (XEXP (XEXP (x, 1), 1)) == CONST_INT |
5248 | && !pow2p_hwi (INTVAL (XEXP (XEXP (x, 1), 1)))) |
5249 | { |
5250 | machine_mode mode = GET_MODE (x); |
5251 | unsigned HOST_WIDE_INT this_int = INTVAL (XEXP (XEXP (x, 1), 1)); |
5252 | HOST_WIDE_INT other_int = trunc_int_for_mode (-this_int, mode); |
5253 | SUBST (*loc, gen_rtx_PLUS (mode, |
5254 | gen_rtx_MULT (mode, |
5255 | XEXP (XEXP (x, 1), 0), |
5256 | gen_int_mode (other_int, |
5257 | mode)), |
5258 | XEXP (x, 0))); |
5259 | return find_split_point (loc, insn, set_src); |
5260 | } |
5261 | |
5262 | /* Split at a multiply-accumulate instruction. However if this is |
5263 | the SET_SRC, we likely do not have such an instruction and it's |
5264 | worthless to try this split. */ |
5265 | if (!set_src |
5266 | && (GET_CODE (XEXP (x, 0)) == MULT |
5267 | || (GET_CODE (XEXP (x, 0)) == ASHIFT |
5268 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT))) |
5269 | return loc; |
5270 | |
5271 | default: |
5272 | break; |
5273 | } |
5274 | |
5275 | /* Otherwise, select our actions depending on our rtx class. */ |
5276 | switch (GET_RTX_CLASS (code)) |
5277 | { |
5278 | case RTX_BITFIELD_OPS: /* This is ZERO_EXTRACT and SIGN_EXTRACT. */ |
5279 | case RTX_TERNARY: |
5280 | split = find_split_point (loc: &XEXP (x, 2), insn, set_src: false); |
5281 | if (split) |
5282 | return split; |
5283 | /* fall through */ |
5284 | case RTX_BIN_ARITH: |
5285 | case RTX_COMM_ARITH: |
5286 | case RTX_COMPARE: |
5287 | case RTX_COMM_COMPARE: |
5288 | split = find_split_point (loc: &XEXP (x, 1), insn, set_src: false); |
5289 | if (split) |
5290 | return split; |
5291 | /* fall through */ |
5292 | case RTX_UNARY: |
5293 | /* Some machines have (and (shift ...) ...) insns. If X is not |
5294 | an AND, but XEXP (X, 0) is, use it as our split point. */ |
5295 | if (GET_CODE (x) != AND && GET_CODE (XEXP (x, 0)) == AND) |
5296 | return &XEXP (x, 0); |
5297 | |
5298 | split = find_split_point (loc: &XEXP (x, 0), insn, set_src: false); |
5299 | if (split) |
5300 | return split; |
5301 | return loc; |
5302 | |
5303 | default: |
5304 | /* Otherwise, we don't have a split point. */ |
5305 | return 0; |
5306 | } |
5307 | } |
5308 | |
5309 | /* Throughout X, replace FROM with TO, and return the result. |
5310 | The result is TO if X is FROM; |
5311 | otherwise the result is X, but its contents may have been modified. |
5312 | If they were modified, a record was made in undobuf so that |
5313 | undo_all will (among other things) return X to its original state. |
5314 | |
5315 | If the number of changes necessary is too much to record to undo, |
5316 | the excess changes are not made, so the result is invalid. |
5317 | The changes already made can still be undone. |
5318 | undobuf.num_undo is incremented for such changes, so by testing that |
5319 | the caller can tell whether the result is valid. |
5320 | |
5321 | `n_occurrences' is incremented each time FROM is replaced. |
5322 | |
5323 | IN_DEST is true if we are processing the SET_DEST of a SET. |
5324 | |
5325 | IN_COND is true if we are at the top level of a condition. |
5326 | |
5327 | UNIQUE_COPY is true if each substitution must be unique. We do this |
5328 | by copying if `n_occurrences' is nonzero. */ |
5329 | |
5330 | static rtx |
5331 | subst (rtx x, rtx from, rtx to, bool in_dest, bool in_cond, bool unique_copy) |
5332 | { |
5333 | enum rtx_code code = GET_CODE (x); |
5334 | machine_mode op0_mode = VOIDmode; |
5335 | const char *fmt; |
5336 | int len, i; |
5337 | rtx new_rtx; |
5338 | |
5339 | /* Two expressions are equal if they are identical copies of a shared |
5340 | RTX or if they are both registers with the same register number |
5341 | and mode. */ |
5342 | |
5343 | #define COMBINE_RTX_EQUAL_P(X,Y) \ |
5344 | ((X) == (Y) \ |
5345 | || (REG_P (X) && REG_P (Y) \ |
5346 | && REGNO (X) == REGNO (Y) && GET_MODE (X) == GET_MODE (Y))) |
5347 | |
5348 | /* Do not substitute into clobbers of regs -- this will never result in |
5349 | valid RTL. */ |
5350 | if (GET_CODE (x) == CLOBBER && REG_P (XEXP (x, 0))) |
5351 | return x; |
5352 | |
5353 | if (! in_dest && COMBINE_RTX_EQUAL_P (x, from)) |
5354 | { |
5355 | n_occurrences++; |
5356 | return (unique_copy && n_occurrences > 1 ? copy_rtx (to) : to); |
5357 | } |
5358 | |
5359 | /* If X and FROM are the same register but different modes, they |
5360 | will not have been seen as equal above. However, the log links code |
5361 | will make a LOG_LINKS entry for that case. If we do nothing, we |
5362 | will try to rerecognize our original insn and, when it succeeds, |
5363 | we will delete the feeding insn, which is incorrect. |
5364 | |
5365 | So force this insn not to match in this (rare) case. */ |
5366 | if (! in_dest && code == REG && REG_P (from) |
5367 | && reg_overlap_mentioned_p (x, from)) |
5368 | return gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
5369 | |
5370 | /* If this is an object, we are done unless it is a MEM or LO_SUM, both |
5371 | of which may contain things that can be combined. */ |
5372 | if (code != MEM && code != LO_SUM && OBJECT_P (x)) |
5373 | return x; |
5374 | |
5375 | /* It is possible to have a subexpression appear twice in the insn. |
5376 | Suppose that FROM is a register that appears within TO. |
5377 | Then, after that subexpression has been scanned once by `subst', |
5378 | the second time it is scanned, TO may be found. If we were |
5379 | to scan TO here, we would find FROM within it and create a |
5380 | self-referent rtl structure which is completely wrong. */ |
5381 | if (COMBINE_RTX_EQUAL_P (x, to)) |
5382 | return to; |
5383 | |
5384 | /* Parallel asm_operands need special attention because all of the |
5385 | inputs are shared across the arms. Furthermore, unsharing the |
5386 | rtl results in recognition failures. Failure to handle this case |
5387 | specially can result in circular rtl. |
5388 | |
5389 | Solve this by doing a normal pass across the first entry of the |
5390 | parallel, and only processing the SET_DESTs of the subsequent |
5391 | entries. Ug. */ |
5392 | |
5393 | if (code == PARALLEL |
5394 | && GET_CODE (XVECEXP (x, 0, 0)) == SET |
5395 | && GET_CODE (SET_SRC (XVECEXP (x, 0, 0))) == ASM_OPERANDS) |
5396 | { |
5397 | new_rtx = subst (XVECEXP (x, 0, 0), from, to, in_dest: false, in_cond: false, unique_copy); |
5398 | |
5399 | /* If this substitution failed, this whole thing fails. */ |
5400 | if (GET_CODE (new_rtx) == CLOBBER |
5401 | && XEXP (new_rtx, 0) == const0_rtx) |
5402 | return new_rtx; |
5403 | |
5404 | SUBST (XVECEXP (x, 0, 0), new_rtx); |
5405 | |
5406 | for (i = XVECLEN (x, 0) - 1; i >= 1; i--) |
5407 | { |
5408 | rtx dest = SET_DEST (XVECEXP (x, 0, i)); |
5409 | |
5410 | if (!REG_P (dest) && GET_CODE (dest) != PC) |
5411 | { |
5412 | new_rtx = subst (x: dest, from, to, in_dest: false, in_cond: false, unique_copy); |
5413 | |
5414 | /* If this substitution failed, this whole thing fails. */ |
5415 | if (GET_CODE (new_rtx) == CLOBBER |
5416 | && XEXP (new_rtx, 0) == const0_rtx) |
5417 | return new_rtx; |
5418 | |
5419 | SUBST (SET_DEST (XVECEXP (x, 0, i)), new_rtx); |
5420 | } |
5421 | } |
5422 | } |
5423 | else |
5424 | { |
5425 | len = GET_RTX_LENGTH (code); |
5426 | fmt = GET_RTX_FORMAT (code); |
5427 | |
5428 | /* We don't need to process a SET_DEST that is a register or PC, so |
5429 | set up to skip this common case. All other cases where we want |
5430 | to suppress replacing something inside a SET_SRC are handled via |
5431 | the IN_DEST operand. */ |
5432 | if (code == SET |
5433 | && (REG_P (SET_DEST (x)) |
5434 | || GET_CODE (SET_DEST (x)) == PC)) |
5435 | fmt = "ie" ; |
5436 | |
5437 | /* Trying to simplify the operands of a widening MULT is not likely |
5438 | to create RTL matching a machine insn. */ |
5439 | if (code == MULT |
5440 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND |
5441 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) |
5442 | && (GET_CODE (XEXP (x, 1)) == ZERO_EXTEND |
5443 | || GET_CODE (XEXP (x, 1)) == SIGN_EXTEND) |
5444 | && REG_P (XEXP (XEXP (x, 0), 0)) |
5445 | && REG_P (XEXP (XEXP (x, 1), 0)) |
5446 | && from == to) |
5447 | return x; |
5448 | |
5449 | |
5450 | /* Get the mode of operand 0 in case X is now a SIGN_EXTEND of a |
5451 | constant. */ |
5452 | if (fmt[0] == 'e') |
5453 | op0_mode = GET_MODE (XEXP (x, 0)); |
5454 | |
5455 | for (i = 0; i < len; i++) |
5456 | { |
5457 | if (fmt[i] == 'E') |
5458 | { |
5459 | int j; |
5460 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
5461 | { |
5462 | if (COMBINE_RTX_EQUAL_P (XVECEXP (x, i, j), from)) |
5463 | { |
5464 | new_rtx = (unique_copy && n_occurrences |
5465 | ? copy_rtx (to) : to); |
5466 | n_occurrences++; |
5467 | } |
5468 | else |
5469 | { |
5470 | new_rtx = subst (XVECEXP (x, i, j), from, to, |
5471 | in_dest: false, in_cond: false, unique_copy); |
5472 | |
5473 | /* If this substitution failed, this whole thing |
5474 | fails. */ |
5475 | if (GET_CODE (new_rtx) == CLOBBER |
5476 | && XEXP (new_rtx, 0) == const0_rtx) |
5477 | return new_rtx; |
5478 | } |
5479 | |
5480 | SUBST (XVECEXP (x, i, j), new_rtx); |
5481 | } |
5482 | } |
5483 | else if (fmt[i] == 'e') |
5484 | { |
5485 | /* If this is a register being set, ignore it. */ |
5486 | new_rtx = XEXP (x, i); |
5487 | if (in_dest |
5488 | && i == 0 |
5489 | && (((code == SUBREG || code == ZERO_EXTRACT) |
5490 | && REG_P (new_rtx)) |
5491 | || code == STRICT_LOW_PART)) |
5492 | ; |
5493 | |
5494 | else if (COMBINE_RTX_EQUAL_P (XEXP (x, i), from)) |
5495 | { |
5496 | /* In general, don't install a subreg involving two |
5497 | modes not tieable. It can worsen register |
5498 | allocation, and can even make invalid reload |
5499 | insns, since the reg inside may need to be copied |
5500 | from in the outside mode, and that may be invalid |
5501 | if it is an fp reg copied in integer mode. |
5502 | |
5503 | We allow an exception to this: It is valid if |
5504 | it is inside another SUBREG and the mode of that |
5505 | SUBREG and the mode of the inside of TO is |
5506 | tieable. */ |
5507 | |
5508 | if (GET_CODE (to) == SUBREG |
5509 | && !targetm.modes_tieable_p (GET_MODE (to), |
5510 | GET_MODE (SUBREG_REG (to))) |
5511 | && ! (code == SUBREG |
5512 | && (targetm.modes_tieable_p |
5513 | (GET_MODE (x), GET_MODE (SUBREG_REG (to)))))) |
5514 | return gen_rtx_CLOBBER (VOIDmode, const0_rtx); |
5515 | |
5516 | if (code == SUBREG |
5517 | && REG_P (to) |
5518 | && REGNO (to) < FIRST_PSEUDO_REGISTER |
5519 | && simplify_subreg_regno (REGNO (to), GET_MODE (to), |
5520 | SUBREG_BYTE (x), |
5521 | GET_MODE (x)) < 0) |
5522 | return gen_rtx_CLOBBER (VOIDmode, const0_rtx); |
5523 | |
5524 | new_rtx = (unique_copy && n_occurrences ? copy_rtx (to) : to); |
5525 | n_occurrences++; |
5526 | } |
5527 | else |
5528 | /* If we are in a SET_DEST, suppress most cases unless we |
5529 | have gone inside a MEM, in which case we want to |
5530 | simplify the address. We assume here that things that |
5531 | are actually part of the destination have their inner |
5532 | parts in the first expression. This is true for SUBREG, |
5533 | STRICT_LOW_PART, and ZERO_EXTRACT, which are the only |
5534 | things aside from REG and MEM that should appear in a |
5535 | SET_DEST. */ |
5536 | new_rtx = subst (XEXP (x, i), from, to, |
5537 | in_dest: (((in_dest |
5538 | && (code == SUBREG || code == STRICT_LOW_PART |
5539 | || code == ZERO_EXTRACT)) |
5540 | || code == SET) |
5541 | && i == 0), |
5542 | in_cond: code == IF_THEN_ELSE && i == 0, |
5543 | unique_copy); |
5544 | |
5545 | /* If we found that we will have to reject this combination, |
5546 | indicate that by returning the CLOBBER ourselves, rather than |
5547 | an expression containing it. This will speed things up as |
5548 | well as prevent accidents where two CLOBBERs are considered |
5549 | to be equal, thus producing an incorrect simplification. */ |
5550 | |
5551 | if (GET_CODE (new_rtx) == CLOBBER && XEXP (new_rtx, 0) == const0_rtx) |
5552 | return new_rtx; |
5553 | |
5554 | if (GET_CODE (x) == SUBREG && CONST_SCALAR_INT_P (new_rtx)) |
5555 | { |
5556 | machine_mode mode = GET_MODE (x); |
5557 | |
5558 | x = simplify_subreg (GET_MODE (x), op: new_rtx, |
5559 | GET_MODE (SUBREG_REG (x)), |
5560 | SUBREG_BYTE (x)); |
5561 | if (! x) |
5562 | x = gen_rtx_CLOBBER (mode, const0_rtx); |
5563 | } |
5564 | else if (CONST_SCALAR_INT_P (new_rtx) |
5565 | && (GET_CODE (x) == ZERO_EXTEND |
5566 | || GET_CODE (x) == SIGN_EXTEND |
5567 | || GET_CODE (x) == FLOAT |
5568 | || GET_CODE (x) == UNSIGNED_FLOAT)) |
5569 | { |
5570 | x = simplify_unary_operation (GET_CODE (x), GET_MODE (x), |
5571 | op: new_rtx, |
5572 | GET_MODE (XEXP (x, 0))); |
5573 | if (!x) |
5574 | return gen_rtx_CLOBBER (VOIDmode, const0_rtx); |
5575 | } |
5576 | /* CONST_INTs shouldn't be substituted into PRE_DEC, PRE_MODIFY |
5577 | etc. arguments, otherwise we can ICE before trying to recog |
5578 | it. See PR104446. */ |
5579 | else if (CONST_SCALAR_INT_P (new_rtx) |
5580 | && GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) |
5581 | return gen_rtx_CLOBBER (VOIDmode, const0_rtx); |
5582 | else |
5583 | SUBST (XEXP (x, i), new_rtx); |
5584 | } |
5585 | } |
5586 | } |
5587 | |
5588 | /* Check if we are loading something from the constant pool via float |
5589 | extension; in this case we would undo compress_float_constant |
5590 | optimization and degenerate constant load to an immediate value. */ |
5591 | if (GET_CODE (x) == FLOAT_EXTEND |
5592 | && MEM_P (XEXP (x, 0)) |
5593 | && MEM_READONLY_P (XEXP (x, 0))) |
5594 | { |
5595 | rtx tmp = avoid_constant_pool_reference (x); |
5596 | if (x != tmp) |
5597 | return x; |
5598 | } |
5599 | |
5600 | /* Try to simplify X. If the simplification changed the code, it is likely |
5601 | that further simplification will help, so loop, but limit the number |
5602 | of repetitions that will be performed. */ |
5603 | |
5604 | for (i = 0; i < 4; i++) |
5605 | { |
5606 | /* If X is sufficiently simple, don't bother trying to do anything |
5607 | with it. */ |
5608 | if (code != CONST_INT && code != REG && code != CLOBBER) |
5609 | x = combine_simplify_rtx (x, op0_mode, in_dest, in_cond); |
5610 | |
5611 | if (GET_CODE (x) == code) |
5612 | break; |
5613 | |
5614 | code = GET_CODE (x); |
5615 | |
5616 | /* We no longer know the original mode of operand 0 since we |
5617 | have changed the form of X) */ |
5618 | op0_mode = VOIDmode; |
5619 | } |
5620 | |
5621 | return x; |
5622 | } |
5623 | |
5624 | /* If X is a commutative operation whose operands are not in the canonical |
5625 | order, use substitutions to swap them. */ |
5626 | |
5627 | static void |
5628 | maybe_swap_commutative_operands (rtx x) |
5629 | { |
5630 | if (COMMUTATIVE_ARITH_P (x) |
5631 | && swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1))) |
5632 | { |
5633 | rtx temp = XEXP (x, 0); |
5634 | SUBST (XEXP (x, 0), XEXP (x, 1)); |
5635 | SUBST (XEXP (x, 1), temp); |
5636 | } |
5637 | |
5638 | unsigned n_elts = 0; |
5639 | if (GET_CODE (x) == VEC_MERGE |
5640 | && CONST_INT_P (XEXP (x, 2)) |
5641 | && GET_MODE_NUNITS (GET_MODE (x)).is_constant (const_value: &n_elts) |
5642 | && (swap_commutative_operands_p (XEXP (x, 0), XEXP (x, 1)) |
5643 | /* Two operands have same precedence, then |
5644 | first bit of mask select first operand. */ |
5645 | || (!swap_commutative_operands_p (XEXP (x, 1), XEXP (x, 0)) |
5646 | && !(UINTVAL (XEXP (x, 2)) & 1)))) |
5647 | { |
5648 | rtx temp = XEXP (x, 0); |
5649 | unsigned HOST_WIDE_INT sel = UINTVAL (XEXP (x, 2)); |
5650 | unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_1U; |
5651 | if (n_elts == HOST_BITS_PER_WIDE_INT) |
5652 | mask = -1; |
5653 | else |
5654 | mask = (HOST_WIDE_INT_1U << n_elts) - 1; |
5655 | SUBST (XEXP (x, 0), XEXP (x, 1)); |
5656 | SUBST (XEXP (x, 1), temp); |
5657 | SUBST (XEXP (x, 2), GEN_INT (~sel & mask)); |
5658 | } |
5659 | } |
5660 | |
5661 | /* Simplify X, a piece of RTL. We just operate on the expression at the |
5662 | outer level; call `subst' to simplify recursively. Return the new |
5663 | expression. |
5664 | |
5665 | OP0_MODE is the original mode of XEXP (x, 0). IN_DEST is true |
5666 | if we are inside a SET_DEST. IN_COND is true if we are at the top level |
5667 | of a condition. */ |
5668 | |
5669 | static rtx |
5670 | combine_simplify_rtx (rtx x, machine_mode op0_mode, bool in_dest, bool in_cond) |
5671 | { |
5672 | enum rtx_code code = GET_CODE (x); |
5673 | machine_mode mode = GET_MODE (x); |
5674 | scalar_int_mode int_mode; |
5675 | rtx temp; |
5676 | int i; |
5677 | |
5678 | /* If this is a commutative operation, put a constant last and a complex |
5679 | expression first. We don't need to do this for comparisons here. */ |
5680 | maybe_swap_commutative_operands (x); |
5681 | |
5682 | /* Try to fold this expression in case we have constants that weren't |
5683 | present before. */ |
5684 | temp = 0; |
5685 | switch (GET_RTX_CLASS (code)) |
5686 | { |
5687 | case RTX_UNARY: |
5688 | if (op0_mode == VOIDmode) |
5689 | op0_mode = GET_MODE (XEXP (x, 0)); |
5690 | temp = simplify_unary_operation (code, mode, XEXP (x, 0), op_mode: op0_mode); |
5691 | break; |
5692 | case RTX_COMPARE: |
5693 | case RTX_COMM_COMPARE: |
5694 | { |
5695 | machine_mode cmp_mode = GET_MODE (XEXP (x, 0)); |
5696 | if (cmp_mode == VOIDmode) |
5697 | { |
5698 | cmp_mode = GET_MODE (XEXP (x, 1)); |
5699 | if (cmp_mode == VOIDmode) |
5700 | cmp_mode = op0_mode; |
5701 | } |
5702 | temp = simplify_relational_operation (code, mode, op_mode: cmp_mode, |
5703 | XEXP (x, 0), XEXP (x, 1)); |
5704 | } |
5705 | break; |
5706 | case RTX_COMM_ARITH: |
5707 | case RTX_BIN_ARITH: |
5708 | temp = simplify_binary_operation (code, mode, XEXP (x, 0), XEXP (x, 1)); |
5709 | break; |
5710 | case RTX_BITFIELD_OPS: |
5711 | case RTX_TERNARY: |
5712 | temp = simplify_ternary_operation (code, mode, op0_mode, XEXP (x, 0), |
5713 | XEXP (x, 1), XEXP (x, 2)); |
5714 | break; |
5715 | default: |
5716 | break; |
5717 | } |
5718 | |
5719 | if (temp) |
5720 | { |
5721 | x = temp; |
5722 | code = GET_CODE (temp); |
5723 | op0_mode = VOIDmode; |
5724 | mode = GET_MODE (temp); |
5725 | } |
5726 | |
5727 | /* If this is a simple operation applied to an IF_THEN_ELSE, try |
5728 | applying it to the arms of the IF_THEN_ELSE. This often simplifies |
5729 | things. Check for cases where both arms are testing the same |
5730 | condition. |
5731 | |
5732 | Don't do anything if all operands are very simple. */ |
5733 | |
5734 | if ((BINARY_P (x) |
5735 | && ((!OBJECT_P (XEXP (x, 0)) |
5736 | && ! (GET_CODE (XEXP (x, 0)) == SUBREG |
5737 | && OBJECT_P (SUBREG_REG (XEXP (x, 0))))) |
5738 | || (!OBJECT_P (XEXP (x, 1)) |
5739 | && ! (GET_CODE (XEXP (x, 1)) == SUBREG |
5740 | && OBJECT_P (SUBREG_REG (XEXP (x, 1))))))) |
5741 | || (UNARY_P (x) |
5742 | && (!OBJECT_P (XEXP (x, 0)) |
5743 | && ! (GET_CODE (XEXP (x, 0)) == SUBREG |
5744 | && OBJECT_P (SUBREG_REG (XEXP (x, 0))))))) |
5745 | { |
5746 | rtx cond, true_rtx, false_rtx; |
5747 | |
5748 | cond = if_then_else_cond (x, &true_rtx, &false_rtx); |
5749 | if (cond != 0 |
5750 | /* If everything is a comparison, what we have is highly unlikely |
5751 | to be simpler, so don't use it. */ |
5752 | && ! (COMPARISON_P (x) |
5753 | && (COMPARISON_P (true_rtx) || COMPARISON_P (false_rtx))) |
5754 | /* Similarly, if we end up with one of the expressions the same |
5755 | as the original, it is certainly not simpler. */ |
5756 | && ! rtx_equal_p (x, true_rtx) |
5757 | && ! rtx_equal_p (x, false_rtx)) |
5758 | { |
5759 | rtx cop1 = const0_rtx; |
5760 | enum rtx_code cond_code = simplify_comparison (NE, &cond, &cop1); |
5761 | |
5762 | if (cond_code == NE && COMPARISON_P (cond)) |
5763 | return x; |
5764 | |
5765 | /* Simplify the alternative arms; this may collapse the true and |
5766 | false arms to store-flag values. Be careful to use copy_rtx |
5767 | here since true_rtx or false_rtx might share RTL with x as a |
5768 | result of the if_then_else_cond call above. */ |
5769 | true_rtx = subst (x: copy_rtx (true_rtx), from: pc_rtx, to: pc_rtx, |
5770 | in_dest: false, in_cond: false, unique_copy: false); |
5771 | false_rtx = subst (x: copy_rtx (false_rtx), from: pc_rtx, to: pc_rtx, |
5772 | in_dest: false, in_cond: false, unique_copy: false); |
5773 | |
5774 | /* If true_rtx and false_rtx are not general_operands, an if_then_else |
5775 | is unlikely to be simpler. */ |
5776 | if (general_operand (true_rtx, VOIDmode) |
5777 | && general_operand (false_rtx, VOIDmode)) |
5778 | { |
5779 | enum rtx_code reversed; |
5780 | |
5781 | /* Restarting if we generate a store-flag expression will cause |
5782 | us to loop. Just drop through in this case. */ |
5783 | |
5784 | /* If the result values are STORE_FLAG_VALUE and zero, we can |
5785 | just make the comparison operation. */ |
5786 | if (true_rtx == const_true_rtx && false_rtx == const0_rtx) |
5787 | x = simplify_gen_relational (code: cond_code, mode, VOIDmode, |
5788 | op0: cond, op1: cop1); |
5789 | else if (true_rtx == const0_rtx && false_rtx == const_true_rtx |
5790 | && ((reversed = reversed_comparison_code_parts |
5791 | (cond_code, cond, cop1, NULL)) |
5792 | != UNKNOWN)) |
5793 | x = simplify_gen_relational (code: reversed, mode, VOIDmode, |
5794 | op0: cond, op1: cop1); |
5795 | |
5796 | /* Likewise, we can make the negate of a comparison operation |
5797 | if the result values are - STORE_FLAG_VALUE and zero. */ |
5798 | else if (CONST_INT_P (true_rtx) |
5799 | && INTVAL (true_rtx) == - STORE_FLAG_VALUE |
5800 | && false_rtx == const0_rtx) |
5801 | x = simplify_gen_unary (code: NEG, mode, |
5802 | op: simplify_gen_relational (code: cond_code, |
5803 | mode, VOIDmode, |
5804 | op0: cond, op1: cop1), |
5805 | op_mode: mode); |
5806 | else if (CONST_INT_P (false_rtx) |
5807 | && INTVAL (false_rtx) == - STORE_FLAG_VALUE |
5808 | && true_rtx == const0_rtx |
5809 | && ((reversed = reversed_comparison_code_parts |
5810 | (cond_code, cond, cop1, NULL)) |
5811 | != UNKNOWN)) |
5812 | x = simplify_gen_unary (code: NEG, mode, |
5813 | op: simplify_gen_relational (code: reversed, |
5814 | mode, VOIDmode, |
5815 | op0: cond, op1: cop1), |
5816 | op_mode: mode); |
5817 | |
5818 | code = GET_CODE (x); |
5819 | op0_mode = VOIDmode; |
5820 | } |
5821 | } |
5822 | } |
5823 | |
5824 | /* First see if we can apply the inverse distributive law. */ |
5825 | if (code == PLUS || code == MINUS |
5826 | || code == AND || code == IOR || code == XOR) |
5827 | { |
5828 | x = apply_distributive_law (x); |
5829 | code = GET_CODE (x); |
5830 | op0_mode = VOIDmode; |
5831 | } |
5832 | |
5833 | /* If CODE is an associative operation not otherwise handled, see if we |
5834 | can associate some operands. This can win if they are constants or |
5835 | if they are logically related (i.e. (a & b) & a). */ |
5836 | if ((code == PLUS || code == MINUS || code == MULT || code == DIV |
5837 | || code == AND || code == IOR || code == XOR |
5838 | || code == SMAX || code == SMIN || code == UMAX || code == UMIN) |
5839 | && ((INTEGRAL_MODE_P (mode) && code != DIV) |
5840 | || (flag_associative_math && FLOAT_MODE_P (mode)))) |
5841 | { |
5842 | if (GET_CODE (XEXP (x, 0)) == code) |
5843 | { |
5844 | rtx other = XEXP (XEXP (x, 0), 0); |
5845 | rtx inner_op0 = XEXP (XEXP (x, 0), 1); |
5846 | rtx inner_op1 = XEXP (x, 1); |
5847 | rtx inner; |
5848 | |
5849 | /* Make sure we pass the constant operand if any as the second |
5850 | one if this is a commutative operation. */ |
5851 | if (CONSTANT_P (inner_op0) && COMMUTATIVE_ARITH_P (x)) |
5852 | std::swap (a&: inner_op0, b&: inner_op1); |
5853 | inner = simplify_binary_operation (code: code == MINUS ? PLUS |
5854 | : code == DIV ? MULT |
5855 | : code, |
5856 | mode, op0: inner_op0, op1: inner_op1); |
5857 | |
5858 | /* For commutative operations, try the other pair if that one |
5859 | didn't simplify. */ |
5860 | if (inner == 0 && COMMUTATIVE_ARITH_P (x)) |
5861 | { |
5862 | other = XEXP (XEXP (x, 0), 1); |
5863 | inner = simplify_binary_operation (code, mode, |
5864 | XEXP (XEXP (x, 0), 0), |
5865 | XEXP (x, 1)); |
5866 | } |
5867 | |
5868 | if (inner) |
5869 | return simplify_gen_binary (code, mode, op0: other, op1: inner); |
5870 | } |
5871 | } |
5872 | |
5873 | /* A little bit of algebraic simplification here. */ |
5874 | switch (code) |
5875 | { |
5876 | case MEM: |
5877 | /* Ensure that our address has any ASHIFTs converted to MULT in case |
5878 | address-recognizing predicates are called later. */ |
5879 | temp = make_compound_operation (XEXP (x, 0), MEM); |
5880 | SUBST (XEXP (x, 0), temp); |
5881 | break; |
5882 | |
5883 | case SUBREG: |
5884 | if (op0_mode == VOIDmode) |
5885 | op0_mode = GET_MODE (SUBREG_REG (x)); |
5886 | |
5887 | /* See if this can be moved to simplify_subreg. */ |
5888 | if (CONSTANT_P (SUBREG_REG (x)) |
5889 | && known_eq (subreg_lowpart_offset (mode, op0_mode), SUBREG_BYTE (x)) |
5890 | /* Don't call gen_lowpart if the inner mode |
5891 | is VOIDmode and we cannot simplify it, as SUBREG without |
5892 | inner mode is invalid. */ |
5893 | && (GET_MODE (SUBREG_REG (x)) != VOIDmode |
5894 | || gen_lowpart_common (mode, SUBREG_REG (x)))) |
5895 | return gen_lowpart (mode, SUBREG_REG (x)); |
5896 | |
5897 | if (GET_MODE_CLASS (GET_MODE (SUBREG_REG (x))) == MODE_CC) |
5898 | break; |
5899 | { |
5900 | rtx temp; |
5901 | temp = simplify_subreg (outermode: mode, SUBREG_REG (x), innermode: op0_mode, |
5902 | SUBREG_BYTE (x)); |
5903 | if (temp) |
5904 | return temp; |
5905 | |
5906 | /* If op is known to have all lower bits zero, the result is zero. */ |
5907 | scalar_int_mode int_mode, int_op0_mode; |
5908 | if (!in_dest |
5909 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
5910 | && is_a <scalar_int_mode> (m: op0_mode, result: &int_op0_mode) |
5911 | && (GET_MODE_PRECISION (mode: int_mode) |
5912 | < GET_MODE_PRECISION (mode: int_op0_mode)) |
5913 | && known_eq (subreg_lowpart_offset (int_mode, int_op0_mode), |
5914 | SUBREG_BYTE (x)) |
5915 | && HWI_COMPUTABLE_MODE_P (mode: int_op0_mode) |
5916 | && ((nonzero_bits (SUBREG_REG (x), int_op0_mode) |
5917 | & GET_MODE_MASK (int_mode)) == 0) |
5918 | && !side_effects_p (SUBREG_REG (x))) |
5919 | return CONST0_RTX (int_mode); |
5920 | } |
5921 | |
5922 | /* Don't change the mode of the MEM if that would change the meaning |
5923 | of the address. */ |
5924 | if (MEM_P (SUBREG_REG (x)) |
5925 | && (MEM_VOLATILE_P (SUBREG_REG (x)) |
5926 | || mode_dependent_address_p (XEXP (SUBREG_REG (x), 0), |
5927 | MEM_ADDR_SPACE (SUBREG_REG (x))))) |
5928 | return gen_rtx_CLOBBER (mode, const0_rtx); |
5929 | |
5930 | /* Note that we cannot do any narrowing for non-constants since |
5931 | we might have been counting on using the fact that some bits were |
5932 | zero. We now do this in the SET. */ |
5933 | |
5934 | break; |
5935 | |
5936 | case NEG: |
5937 | temp = expand_compound_operation (XEXP (x, 0)); |
5938 | |
5939 | /* For C equal to the width of MODE minus 1, (neg (ashiftrt X C)) can be |
5940 | replaced by (lshiftrt X C). This will convert |
5941 | (neg (sign_extract X 1 Y)) to (zero_extract X 1 Y). */ |
5942 | |
5943 | if (GET_CODE (temp) == ASHIFTRT |
5944 | && CONST_INT_P (XEXP (temp, 1)) |
5945 | && INTVAL (XEXP (temp, 1)) == GET_MODE_UNIT_PRECISION (mode) - 1) |
5946 | return simplify_shift_const (NULL_RTX, LSHIFTRT, mode, XEXP (temp, 0), |
5947 | INTVAL (XEXP (temp, 1))); |
5948 | |
5949 | /* If X has only a single bit that might be nonzero, say, bit I, convert |
5950 | (neg X) to (ashiftrt (ashift X C-I) C-I) where C is the bitsize of |
5951 | MODE minus 1. This will convert (neg (zero_extract X 1 Y)) to |
5952 | (sign_extract X 1 Y). But only do this if TEMP isn't a register |
5953 | or a SUBREG of one since we'd be making the expression more |
5954 | complex if it was just a register. */ |
5955 | |
5956 | if (!REG_P (temp) |
5957 | && ! (GET_CODE (temp) == SUBREG |
5958 | && REG_P (SUBREG_REG (temp))) |
5959 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
5960 | && (i = exact_log2 (x: nonzero_bits (temp, int_mode))) >= 0) |
5961 | { |
5962 | rtx temp1 = simplify_shift_const |
5963 | (NULL_RTX, ASHIFTRT, int_mode, |
5964 | simplify_shift_const (NULL_RTX, ASHIFT, int_mode, temp, |
5965 | GET_MODE_PRECISION (mode: int_mode) - 1 - i), |
5966 | GET_MODE_PRECISION (mode: int_mode) - 1 - i); |
5967 | |
5968 | /* If all we did was surround TEMP with the two shifts, we |
5969 | haven't improved anything, so don't use it. Otherwise, |
5970 | we are better off with TEMP1. */ |
5971 | if (GET_CODE (temp1) != ASHIFTRT |
5972 | || GET_CODE (XEXP (temp1, 0)) != ASHIFT |
5973 | || XEXP (XEXP (temp1, 0), 0) != temp) |
5974 | return temp1; |
5975 | } |
5976 | break; |
5977 | |
5978 | case TRUNCATE: |
5979 | /* We can't handle truncation to a partial integer mode here |
5980 | because we don't know the real bitsize of the partial |
5981 | integer mode. */ |
5982 | if (GET_MODE_CLASS (mode) == MODE_PARTIAL_INT) |
5983 | break; |
5984 | |
5985 | if (HWI_COMPUTABLE_MODE_P (mode)) |
5986 | SUBST (XEXP (x, 0), |
5987 | force_to_mode (XEXP (x, 0), GET_MODE (XEXP (x, 0)), |
5988 | GET_MODE_MASK (mode), false)); |
5989 | |
5990 | /* We can truncate a constant value and return it. */ |
5991 | { |
5992 | poly_int64 c; |
5993 | if (poly_int_rtx_p (XEXP (x, 0), res: &c)) |
5994 | return gen_int_mode (c, mode); |
5995 | } |
5996 | |
5997 | /* Similarly to what we do in simplify-rtx.cc, a truncate of a register |
5998 | whose value is a comparison can be replaced with a subreg if |
5999 | STORE_FLAG_VALUE permits. */ |
6000 | if (HWI_COMPUTABLE_MODE_P (mode) |
6001 | && (STORE_FLAG_VALUE & ~GET_MODE_MASK (mode)) == 0 |
6002 | && (temp = get_last_value (XEXP (x, 0))) |
6003 | && COMPARISON_P (temp) |
6004 | && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (XEXP (x, 0)))) |
6005 | return gen_lowpart (mode, XEXP (x, 0)); |
6006 | break; |
6007 | |
6008 | case CONST: |
6009 | /* (const (const X)) can become (const X). Do it this way rather than |
6010 | returning the inner CONST since CONST can be shared with a |
6011 | REG_EQUAL note. */ |
6012 | if (GET_CODE (XEXP (x, 0)) == CONST) |
6013 | SUBST (XEXP (x, 0), XEXP (XEXP (x, 0), 0)); |
6014 | break; |
6015 | |
6016 | case LO_SUM: |
6017 | /* Convert (lo_sum (high FOO) FOO) to FOO. This is necessary so we |
6018 | can add in an offset. find_split_point will split this address up |
6019 | again if it doesn't match. */ |
6020 | if (HAVE_lo_sum && GET_CODE (XEXP (x, 0)) == HIGH |
6021 | && rtx_equal_p (XEXP (XEXP (x, 0), 0), XEXP (x, 1))) |
6022 | return XEXP (x, 1); |
6023 | break; |
6024 | |
6025 | case PLUS: |
6026 | /* (plus (xor (and <foo> (const_int pow2 - 1)) <c>) <-c>) |
6027 | when c is (const_int (pow2 + 1) / 2) is a sign extension of a |
6028 | bit-field and can be replaced by either a sign_extend or a |
6029 | sign_extract. The `and' may be a zero_extend and the two |
6030 | <c>, -<c> constants may be reversed. */ |
6031 | if (GET_CODE (XEXP (x, 0)) == XOR |
6032 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
6033 | && CONST_INT_P (XEXP (x, 1)) |
6034 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) |
6035 | && INTVAL (XEXP (x, 1)) == -INTVAL (XEXP (XEXP (x, 0), 1)) |
6036 | && ((i = exact_log2 (UINTVAL (XEXP (XEXP (x, 0), 1)))) >= 0 |
6037 | || (i = exact_log2 (UINTVAL (XEXP (x, 1)))) >= 0) |
6038 | && HWI_COMPUTABLE_MODE_P (mode: int_mode) |
6039 | && ((GET_CODE (XEXP (XEXP (x, 0), 0)) == AND |
6040 | && CONST_INT_P (XEXP (XEXP (XEXP (x, 0), 0), 1)) |
6041 | && (UINTVAL (XEXP (XEXP (XEXP (x, 0), 0), 1)) |
6042 | == (HOST_WIDE_INT_1U << (i + 1)) - 1)) |
6043 | || (GET_CODE (XEXP (XEXP (x, 0), 0)) == ZERO_EXTEND |
6044 | && known_eq ((GET_MODE_PRECISION |
6045 | (GET_MODE (XEXP (XEXP (XEXP (x, 0), 0), 0)))), |
6046 | (unsigned int) i + 1)))) |
6047 | return simplify_shift_const |
6048 | (NULL_RTX, ASHIFTRT, int_mode, |
6049 | simplify_shift_const (NULL_RTX, ASHIFT, int_mode, |
6050 | XEXP (XEXP (XEXP (x, 0), 0), 0), |
6051 | GET_MODE_PRECISION (mode: int_mode) - (i + 1)), |
6052 | GET_MODE_PRECISION (mode: int_mode) - (i + 1)); |
6053 | |
6054 | /* If only the low-order bit of X is possibly nonzero, (plus x -1) |
6055 | can become (ashiftrt (ashift (xor x 1) C) C) where C is |
6056 | the bitsize of the mode - 1. This allows simplification of |
6057 | "a = (b & 8) == 0;" */ |
6058 | if (XEXP (x, 1) == constm1_rtx |
6059 | && !REG_P (XEXP (x, 0)) |
6060 | && ! (GET_CODE (XEXP (x, 0)) == SUBREG |
6061 | && REG_P (SUBREG_REG (XEXP (x, 0)))) |
6062 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
6063 | && nonzero_bits (XEXP (x, 0), int_mode) == 1) |
6064 | return simplify_shift_const |
6065 | (NULL_RTX, ASHIFTRT, int_mode, |
6066 | simplify_shift_const (NULL_RTX, ASHIFT, int_mode, |
6067 | gen_rtx_XOR (int_mode, XEXP (x, 0), |
6068 | const1_rtx), |
6069 | GET_MODE_PRECISION (mode: int_mode) - 1), |
6070 | GET_MODE_PRECISION (mode: int_mode) - 1); |
6071 | |
6072 | /* If we are adding two things that have no bits in common, convert |
6073 | the addition into an IOR. This will often be further simplified, |
6074 | for example in cases like ((a & 1) + (a & 2)), which can |
6075 | become a & 3. */ |
6076 | |
6077 | if (HWI_COMPUTABLE_MODE_P (mode) |
6078 | && (nonzero_bits (XEXP (x, 0), mode) |
6079 | & nonzero_bits (XEXP (x, 1), mode)) == 0) |
6080 | { |
6081 | /* Try to simplify the expression further. */ |
6082 | rtx tor = simplify_gen_binary (code: IOR, mode, XEXP (x, 0), XEXP (x, 1)); |
6083 | temp = combine_simplify_rtx (x: tor, VOIDmode, in_dest, in_cond: false); |
6084 | |
6085 | /* If we could, great. If not, do not go ahead with the IOR |
6086 | replacement, since PLUS appears in many special purpose |
6087 | address arithmetic instructions. */ |
6088 | if (GET_CODE (temp) != CLOBBER |
6089 | && (GET_CODE (temp) != IOR |
6090 | || ((XEXP (temp, 0) != XEXP (x, 0) |
6091 | || XEXP (temp, 1) != XEXP (x, 1)) |
6092 | && (XEXP (temp, 0) != XEXP (x, 1) |
6093 | || XEXP (temp, 1) != XEXP (x, 0))))) |
6094 | return temp; |
6095 | } |
6096 | |
6097 | /* Canonicalize x + x into x << 1. */ |
6098 | if (GET_MODE_CLASS (mode) == MODE_INT |
6099 | && rtx_equal_p (XEXP (x, 0), XEXP (x, 1)) |
6100 | && !side_effects_p (XEXP (x, 0))) |
6101 | return simplify_gen_binary (code: ASHIFT, mode, XEXP (x, 0), const1_rtx); |
6102 | |
6103 | break; |
6104 | |
6105 | case MINUS: |
6106 | /* (minus <foo> (and <foo> (const_int -pow2))) becomes |
6107 | (and <foo> (const_int pow2-1)) */ |
6108 | if (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
6109 | && GET_CODE (XEXP (x, 1)) == AND |
6110 | && CONST_INT_P (XEXP (XEXP (x, 1), 1)) |
6111 | && pow2p_hwi (x: -UINTVAL (XEXP (XEXP (x, 1), 1))) |
6112 | && rtx_equal_p (XEXP (XEXP (x, 1), 0), XEXP (x, 0))) |
6113 | return simplify_and_const_int (NULL_RTX, int_mode, XEXP (x, 0), |
6114 | -INTVAL (XEXP (XEXP (x, 1), 1)) - 1); |
6115 | break; |
6116 | |
6117 | case MULT: |
6118 | /* If we have (mult (plus A B) C), apply the distributive law and then |
6119 | the inverse distributive law to see if things simplify. This |
6120 | occurs mostly in addresses, often when unrolling loops. */ |
6121 | |
6122 | if (GET_CODE (XEXP (x, 0)) == PLUS) |
6123 | { |
6124 | rtx result = distribute_and_simplify_rtx (x, 0); |
6125 | if (result) |
6126 | return result; |
6127 | } |
6128 | |
6129 | /* Try simplify a*(b/c) as (a*b)/c. */ |
6130 | if (FLOAT_MODE_P (mode) && flag_associative_math |
6131 | && GET_CODE (XEXP (x, 0)) == DIV) |
6132 | { |
6133 | rtx tem = simplify_binary_operation (code: MULT, mode, |
6134 | XEXP (XEXP (x, 0), 0), |
6135 | XEXP (x, 1)); |
6136 | if (tem) |
6137 | return simplify_gen_binary (code: DIV, mode, op0: tem, XEXP (XEXP (x, 0), 1)); |
6138 | } |
6139 | break; |
6140 | |
6141 | case UDIV: |
6142 | /* If this is a divide by a power of two, treat it as a shift if |
6143 | its first operand is a shift. */ |
6144 | if (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
6145 | && CONST_INT_P (XEXP (x, 1)) |
6146 | && (i = exact_log2 (UINTVAL (XEXP (x, 1)))) >= 0 |
6147 | && (GET_CODE (XEXP (x, 0)) == ASHIFT |
6148 | || GET_CODE (XEXP (x, 0)) == LSHIFTRT |
6149 | || GET_CODE (XEXP (x, 0)) == ASHIFTRT |
6150 | || GET_CODE (XEXP (x, 0)) == ROTATE |
6151 | || GET_CODE (XEXP (x, 0)) == ROTATERT)) |
6152 | return simplify_shift_const (NULL_RTX, LSHIFTRT, int_mode, |
6153 | XEXP (x, 0), i); |
6154 | break; |
6155 | |
6156 | case EQ: case NE: |
6157 | case GT: case GTU: case GE: case GEU: |
6158 | case LT: case LTU: case LE: case LEU: |
6159 | case UNEQ: case LTGT: |
6160 | case UNGT: case UNGE: |
6161 | case UNLT: case UNLE: |
6162 | case UNORDERED: case ORDERED: |
6163 | /* If the first operand is a condition code, we can't do anything |
6164 | with it. */ |
6165 | if (GET_CODE (XEXP (x, 0)) == COMPARE |
6166 | || GET_MODE_CLASS (GET_MODE (XEXP (x, 0))) != MODE_CC) |
6167 | { |
6168 | rtx op0 = XEXP (x, 0); |
6169 | rtx op1 = XEXP (x, 1); |
6170 | enum rtx_code new_code; |
6171 | |
6172 | if (GET_CODE (op0) == COMPARE) |
6173 | op1 = XEXP (op0, 1), op0 = XEXP (op0, 0); |
6174 | |
6175 | /* Simplify our comparison, if possible. */ |
6176 | new_code = simplify_comparison (code, &op0, &op1); |
6177 | |
6178 | /* If STORE_FLAG_VALUE is 1, we can convert (ne x 0) to simply X |
6179 | if only the low-order bit is possibly nonzero in X (such as when |
6180 | X is a ZERO_EXTRACT of one bit). Similarly, we can convert EQ to |
6181 | (xor X 1) or (minus 1 X); we use the former. Finally, if X is |
6182 | known to be either 0 or -1, NE becomes a NEG and EQ becomes |
6183 | (plus X 1). |
6184 | |
6185 | Remove any ZERO_EXTRACT we made when thinking this was a |
6186 | comparison. It may now be simpler to use, e.g., an AND. If a |
6187 | ZERO_EXTRACT is indeed appropriate, it will be placed back by |
6188 | the call to make_compound_operation in the SET case. |
6189 | |
6190 | Don't apply these optimizations if the caller would |
6191 | prefer a comparison rather than a value. |
6192 | E.g., for the condition in an IF_THEN_ELSE most targets need |
6193 | an explicit comparison. */ |
6194 | |
6195 | if (in_cond) |
6196 | ; |
6197 | |
6198 | else if (STORE_FLAG_VALUE == 1 |
6199 | && new_code == NE |
6200 | && is_int_mode (mode, int_mode: &int_mode) |
6201 | && op1 == const0_rtx |
6202 | && int_mode == GET_MODE (op0) |
6203 | && nonzero_bits (op0, int_mode) == 1) |
6204 | return gen_lowpart (int_mode, |
6205 | expand_compound_operation (op0)); |
6206 | |
6207 | else if (STORE_FLAG_VALUE == 1 |
6208 | && new_code == NE |
6209 | && is_int_mode (mode, int_mode: &int_mode) |
6210 | && op1 == const0_rtx |
6211 | && int_mode == GET_MODE (op0) |
6212 | && (num_sign_bit_copies (op0, int_mode) |
6213 | == GET_MODE_PRECISION (mode: int_mode))) |
6214 | { |
6215 | op0 = expand_compound_operation (op0); |
6216 | return simplify_gen_unary (code: NEG, mode: int_mode, |
6217 | gen_lowpart (int_mode, op0), |
6218 | op_mode: int_mode); |
6219 | } |
6220 | |
6221 | else if (STORE_FLAG_VALUE == 1 |
6222 | && new_code == EQ |
6223 | && is_int_mode (mode, int_mode: &int_mode) |
6224 | && op1 == const0_rtx |
6225 | && int_mode == GET_MODE (op0) |
6226 | && nonzero_bits (op0, int_mode) == 1) |
6227 | { |
6228 | op0 = expand_compound_operation (op0); |
6229 | return simplify_gen_binary (code: XOR, mode: int_mode, |
6230 | gen_lowpart (int_mode, op0), |
6231 | const1_rtx); |
6232 | } |
6233 | |
6234 | else if (STORE_FLAG_VALUE == 1 |
6235 | && new_code == EQ |
6236 | && is_int_mode (mode, int_mode: &int_mode) |
6237 | && op1 == const0_rtx |
6238 | && int_mode == GET_MODE (op0) |
6239 | && (num_sign_bit_copies (op0, int_mode) |
6240 | == GET_MODE_PRECISION (mode: int_mode))) |
6241 | { |
6242 | op0 = expand_compound_operation (op0); |
6243 | return plus_constant (int_mode, gen_lowpart (int_mode, op0), 1); |
6244 | } |
6245 | |
6246 | /* If STORE_FLAG_VALUE is -1, we have cases similar to |
6247 | those above. */ |
6248 | if (in_cond) |
6249 | ; |
6250 | |
6251 | else if (STORE_FLAG_VALUE == -1 |
6252 | && new_code == NE |
6253 | && is_int_mode (mode, int_mode: &int_mode) |
6254 | && op1 == const0_rtx |
6255 | && int_mode == GET_MODE (op0) |
6256 | && (num_sign_bit_copies (op0, int_mode) |
6257 | == GET_MODE_PRECISION (mode: int_mode))) |
6258 | return gen_lowpart (int_mode, expand_compound_operation (op0)); |
6259 | |
6260 | else if (STORE_FLAG_VALUE == -1 |
6261 | && new_code == NE |
6262 | && is_int_mode (mode, int_mode: &int_mode) |
6263 | && op1 == const0_rtx |
6264 | && int_mode == GET_MODE (op0) |
6265 | && nonzero_bits (op0, int_mode) == 1) |
6266 | { |
6267 | op0 = expand_compound_operation (op0); |
6268 | return simplify_gen_unary (code: NEG, mode: int_mode, |
6269 | gen_lowpart (int_mode, op0), |
6270 | op_mode: int_mode); |
6271 | } |
6272 | |
6273 | else if (STORE_FLAG_VALUE == -1 |
6274 | && new_code == EQ |
6275 | && is_int_mode (mode, int_mode: &int_mode) |
6276 | && op1 == const0_rtx |
6277 | && int_mode == GET_MODE (op0) |
6278 | && (num_sign_bit_copies (op0, int_mode) |
6279 | == GET_MODE_PRECISION (mode: int_mode))) |
6280 | { |
6281 | op0 = expand_compound_operation (op0); |
6282 | return simplify_gen_unary (code: NOT, mode: int_mode, |
6283 | gen_lowpart (int_mode, op0), |
6284 | op_mode: int_mode); |
6285 | } |
6286 | |
6287 | /* If X is 0/1, (eq X 0) is X-1. */ |
6288 | else if (STORE_FLAG_VALUE == -1 |
6289 | && new_code == EQ |
6290 | && is_int_mode (mode, int_mode: &int_mode) |
6291 | && op1 == const0_rtx |
6292 | && int_mode == GET_MODE (op0) |
6293 | && nonzero_bits (op0, int_mode) == 1) |
6294 | { |
6295 | op0 = expand_compound_operation (op0); |
6296 | return plus_constant (int_mode, gen_lowpart (int_mode, op0), -1); |
6297 | } |
6298 | |
6299 | /* If STORE_FLAG_VALUE says to just test the sign bit and X has just |
6300 | one bit that might be nonzero, we can convert (ne x 0) to |
6301 | (ashift x c) where C puts the bit in the sign bit. Remove any |
6302 | AND with STORE_FLAG_VALUE when we are done, since we are only |
6303 | going to test the sign bit. */ |
6304 | if (new_code == NE |
6305 | && is_int_mode (mode, int_mode: &int_mode) |
6306 | && HWI_COMPUTABLE_MODE_P (mode: int_mode) |
6307 | && val_signbit_p (int_mode, STORE_FLAG_VALUE) |
6308 | && op1 == const0_rtx |
6309 | && int_mode == GET_MODE (op0) |
6310 | && (i = exact_log2 (x: nonzero_bits (op0, int_mode))) >= 0) |
6311 | { |
6312 | x = simplify_shift_const (NULL_RTX, ASHIFT, int_mode, |
6313 | expand_compound_operation (op0), |
6314 | GET_MODE_PRECISION (mode: int_mode) - 1 - i); |
6315 | if (GET_CODE (x) == AND && XEXP (x, 1) == const_true_rtx) |
6316 | return XEXP (x, 0); |
6317 | else |
6318 | return x; |
6319 | } |
6320 | |
6321 | /* If the code changed, return a whole new comparison. |
6322 | We also need to avoid using SUBST in cases where |
6323 | simplify_comparison has widened a comparison with a CONST_INT, |
6324 | since in that case the wider CONST_INT may fail the sanity |
6325 | checks in do_SUBST. */ |
6326 | if (new_code != code |
6327 | || (CONST_INT_P (op1) |
6328 | && GET_MODE (op0) != GET_MODE (XEXP (x, 0)) |
6329 | && GET_MODE (op0) != GET_MODE (XEXP (x, 1)))) |
6330 | return gen_rtx_fmt_ee (new_code, mode, op0, op1); |
6331 | |
6332 | /* Otherwise, keep this operation, but maybe change its operands. |
6333 | This also converts (ne (compare FOO BAR) 0) to (ne FOO BAR). */ |
6334 | SUBST (XEXP (x, 0), op0); |
6335 | SUBST (XEXP (x, 1), op1); |
6336 | } |
6337 | break; |
6338 | |
6339 | case IF_THEN_ELSE: |
6340 | return simplify_if_then_else (x); |
6341 | |
6342 | case ZERO_EXTRACT: |
6343 | case SIGN_EXTRACT: |
6344 | case ZERO_EXTEND: |
6345 | case SIGN_EXTEND: |
6346 | /* If we are processing SET_DEST, we are done. */ |
6347 | if (in_dest) |
6348 | return x; |
6349 | |
6350 | return expand_compound_operation (x); |
6351 | |
6352 | case SET: |
6353 | return simplify_set (x); |
6354 | |
6355 | case AND: |
6356 | case IOR: |
6357 | return simplify_logical (x); |
6358 | |
6359 | case ASHIFT: |
6360 | case LSHIFTRT: |
6361 | case ASHIFTRT: |
6362 | case ROTATE: |
6363 | case ROTATERT: |
6364 | /* If this is a shift by a constant amount, simplify it. */ |
6365 | if (CONST_INT_P (XEXP (x, 1))) |
6366 | return simplify_shift_const (x, code, mode, XEXP (x, 0), |
6367 | INTVAL (XEXP (x, 1))); |
6368 | |
6369 | else if (SHIFT_COUNT_TRUNCATED && !REG_P (XEXP (x, 1))) |
6370 | SUBST (XEXP (x, 1), |
6371 | force_to_mode (XEXP (x, 1), GET_MODE (XEXP (x, 1)), |
6372 | (HOST_WIDE_INT_1U |
6373 | << exact_log2 (GET_MODE_UNIT_BITSIZE |
6374 | (GET_MODE (x)))) - 1, false)); |
6375 | break; |
6376 | case VEC_SELECT: |
6377 | { |
6378 | rtx trueop0 = XEXP (x, 0); |
6379 | mode = GET_MODE (trueop0); |
6380 | rtx trueop1 = XEXP (x, 1); |
6381 | /* If we select a low-part subreg, return that. */ |
6382 | if (vec_series_lowpart_p (GET_MODE (x), op_mode: mode, sel: trueop1)) |
6383 | { |
6384 | rtx new_rtx = lowpart_subreg (GET_MODE (x), op: trueop0, innermode: mode); |
6385 | if (new_rtx != NULL_RTX) |
6386 | return new_rtx; |
6387 | } |
6388 | } |
6389 | |
6390 | default: |
6391 | break; |
6392 | } |
6393 | |
6394 | return x; |
6395 | } |
6396 | |
6397 | /* Simplify X, an IF_THEN_ELSE expression. Return the new expression. */ |
6398 | |
6399 | static rtx |
6400 | simplify_if_then_else (rtx x) |
6401 | { |
6402 | machine_mode mode = GET_MODE (x); |
6403 | rtx cond = XEXP (x, 0); |
6404 | rtx true_rtx = XEXP (x, 1); |
6405 | rtx false_rtx = XEXP (x, 2); |
6406 | enum rtx_code true_code = GET_CODE (cond); |
6407 | bool comparison_p = COMPARISON_P (cond); |
6408 | rtx temp; |
6409 | int i; |
6410 | enum rtx_code false_code; |
6411 | rtx reversed; |
6412 | scalar_int_mode int_mode, inner_mode; |
6413 | |
6414 | /* Simplify storing of the truth value. */ |
6415 | if (comparison_p && true_rtx == const_true_rtx && false_rtx == const0_rtx) |
6416 | return simplify_gen_relational (code: true_code, mode, VOIDmode, |
6417 | XEXP (cond, 0), XEXP (cond, 1)); |
6418 | |
6419 | /* Also when the truth value has to be reversed. */ |
6420 | if (comparison_p |
6421 | && true_rtx == const0_rtx && false_rtx == const_true_rtx |
6422 | && (reversed = reversed_comparison (cond, mode))) |
6423 | return reversed; |
6424 | |
6425 | /* Sometimes we can simplify the arm of an IF_THEN_ELSE if a register used |
6426 | in it is being compared against certain values. Get the true and false |
6427 | comparisons and see if that says anything about the value of each arm. */ |
6428 | |
6429 | if (comparison_p |
6430 | && ((false_code = reversed_comparison_code (cond, NULL)) |
6431 | != UNKNOWN) |
6432 | && REG_P (XEXP (cond, 0))) |
6433 | { |
6434 | HOST_WIDE_INT nzb; |
6435 | rtx from = XEXP (cond, 0); |
6436 | rtx true_val = XEXP (cond, 1); |
6437 | rtx false_val = true_val; |
6438 | bool swapped = false; |
6439 | |
6440 | /* If FALSE_CODE is EQ, swap the codes and arms. */ |
6441 | |
6442 | if (false_code == EQ) |
6443 | { |
6444 | swapped = true, true_code = EQ, false_code = NE; |
6445 | std::swap (a&: true_rtx, b&: false_rtx); |
6446 | } |
6447 | |
6448 | scalar_int_mode from_mode; |
6449 | if (is_a <scalar_int_mode> (GET_MODE (from), result: &from_mode)) |
6450 | { |
6451 | /* If we are comparing against zero and the expression being |
6452 | tested has only a single bit that might be nonzero, that is |
6453 | its value when it is not equal to zero. Similarly if it is |
6454 | known to be -1 or 0. */ |
6455 | if (true_code == EQ |
6456 | && true_val == const0_rtx |
6457 | && pow2p_hwi (x: nzb = nonzero_bits (from, from_mode))) |
6458 | { |
6459 | false_code = EQ; |
6460 | false_val = gen_int_mode (nzb, from_mode); |
6461 | } |
6462 | else if (true_code == EQ |
6463 | && true_val == const0_rtx |
6464 | && (num_sign_bit_copies (from, from_mode) |
6465 | == GET_MODE_PRECISION (mode: from_mode))) |
6466 | { |
6467 | false_code = EQ; |
6468 | false_val = constm1_rtx; |
6469 | } |
6470 | } |
6471 | |
6472 | /* Now simplify an arm if we know the value of the register in the |
6473 | branch and it is used in the arm. Be careful due to the potential |
6474 | of locally-shared RTL. */ |
6475 | |
6476 | if (reg_mentioned_p (from, true_rtx)) |
6477 | true_rtx = subst (x: known_cond (copy_rtx (true_rtx), true_code, |
6478 | from, true_val), |
6479 | from: pc_rtx, to: pc_rtx, in_dest: false, in_cond: false, unique_copy: false); |
6480 | if (reg_mentioned_p (from, false_rtx)) |
6481 | false_rtx = subst (x: known_cond (copy_rtx (false_rtx), false_code, |
6482 | from, false_val), |
6483 | from: pc_rtx, to: pc_rtx, in_dest: false, in_cond: false, unique_copy: false); |
6484 | |
6485 | SUBST (XEXP (x, 1), swapped ? false_rtx : true_rtx); |
6486 | SUBST (XEXP (x, 2), swapped ? true_rtx : false_rtx); |
6487 | |
6488 | true_rtx = XEXP (x, 1); |
6489 | false_rtx = XEXP (x, 2); |
6490 | true_code = GET_CODE (cond); |
6491 | } |
6492 | |
6493 | /* If we have (if_then_else FOO (pc) (label_ref BAR)) and FOO can be |
6494 | reversed, do so to avoid needing two sets of patterns for |
6495 | subtract-and-branch insns. Similarly if we have a constant in the true |
6496 | arm, the false arm is the same as the first operand of the comparison, or |
6497 | the false arm is more complicated than the true arm. */ |
6498 | |
6499 | if (comparison_p |
6500 | && reversed_comparison_code (cond, NULL) != UNKNOWN |
6501 | && (true_rtx == pc_rtx |
6502 | || (CONSTANT_P (true_rtx) |
6503 | && !CONST_INT_P (false_rtx) && false_rtx != pc_rtx) |
6504 | || true_rtx == const0_rtx |
6505 | || (OBJECT_P (true_rtx) && !OBJECT_P (false_rtx)) |
6506 | || (GET_CODE (true_rtx) == SUBREG && OBJECT_P (SUBREG_REG (true_rtx)) |
6507 | && !OBJECT_P (false_rtx)) |
6508 | || reg_mentioned_p (true_rtx, false_rtx) |
6509 | || rtx_equal_p (false_rtx, XEXP (cond, 0)))) |
6510 | { |
6511 | SUBST (XEXP (x, 0), reversed_comparison (cond, GET_MODE (cond))); |
6512 | SUBST (XEXP (x, 1), false_rtx); |
6513 | SUBST (XEXP (x, 2), true_rtx); |
6514 | |
6515 | std::swap (a&: true_rtx, b&: false_rtx); |
6516 | cond = XEXP (x, 0); |
6517 | |
6518 | /* It is possible that the conditional has been simplified out. */ |
6519 | true_code = GET_CODE (cond); |
6520 | comparison_p = COMPARISON_P (cond); |
6521 | } |
6522 | |
6523 | /* If the two arms are identical, we don't need the comparison. */ |
6524 | |
6525 | if (rtx_equal_p (true_rtx, false_rtx) && ! side_effects_p (cond)) |
6526 | return true_rtx; |
6527 | |
6528 | /* Convert a == b ? b : a to "a". */ |
6529 | if (true_code == EQ && ! side_effects_p (cond) |
6530 | && !HONOR_NANS (mode) |
6531 | && rtx_equal_p (XEXP (cond, 0), false_rtx) |
6532 | && rtx_equal_p (XEXP (cond, 1), true_rtx)) |
6533 | return false_rtx; |
6534 | else if (true_code == NE && ! side_effects_p (cond) |
6535 | && !HONOR_NANS (mode) |
6536 | && rtx_equal_p (XEXP (cond, 0), true_rtx) |
6537 | && rtx_equal_p (XEXP (cond, 1), false_rtx)) |
6538 | return true_rtx; |
6539 | |
6540 | /* Look for cases where we have (abs x) or (neg (abs X)). */ |
6541 | |
6542 | if (GET_MODE_CLASS (mode) == MODE_INT |
6543 | && comparison_p |
6544 | && XEXP (cond, 1) == const0_rtx |
6545 | && GET_CODE (false_rtx) == NEG |
6546 | && rtx_equal_p (true_rtx, XEXP (false_rtx, 0)) |
6547 | && rtx_equal_p (true_rtx, XEXP (cond, 0)) |
6548 | && ! side_effects_p (true_rtx)) |
6549 | switch (true_code) |
6550 | { |
6551 | case GT: |
6552 | case GE: |
6553 | return simplify_gen_unary (code: ABS, mode, op: true_rtx, op_mode: mode); |
6554 | case LT: |
6555 | case LE: |
6556 | return |
6557 | simplify_gen_unary (code: NEG, mode, |
6558 | op: simplify_gen_unary (code: ABS, mode, op: true_rtx, op_mode: mode), |
6559 | op_mode: mode); |
6560 | default: |
6561 | break; |
6562 | } |
6563 | |
6564 | /* Look for MIN or MAX. */ |
6565 | |
6566 | if ((! FLOAT_MODE_P (mode) |
6567 | || (flag_unsafe_math_optimizations |
6568 | && !HONOR_NANS (mode) |
6569 | && !HONOR_SIGNED_ZEROS (mode))) |
6570 | && comparison_p |
6571 | && rtx_equal_p (XEXP (cond, 0), true_rtx) |
6572 | && rtx_equal_p (XEXP (cond, 1), false_rtx) |
6573 | && ! side_effects_p (cond)) |
6574 | switch (true_code) |
6575 | { |
6576 | case GE: |
6577 | case GT: |
6578 | return simplify_gen_binary (code: SMAX, mode, op0: true_rtx, op1: false_rtx); |
6579 | case LE: |
6580 | case LT: |
6581 | return simplify_gen_binary (code: SMIN, mode, op0: true_rtx, op1: false_rtx); |
6582 | case GEU: |
6583 | case GTU: |
6584 | return simplify_gen_binary (code: UMAX, mode, op0: true_rtx, op1: false_rtx); |
6585 | case LEU: |
6586 | case LTU: |
6587 | return simplify_gen_binary (code: UMIN, mode, op0: true_rtx, op1: false_rtx); |
6588 | default: |
6589 | break; |
6590 | } |
6591 | |
6592 | /* If we have (if_then_else COND (OP Z C1) Z) and OP is an identity when its |
6593 | second operand is zero, this can be done as (OP Z (mult COND C2)) where |
6594 | C2 = C1 * STORE_FLAG_VALUE. Similarly if OP has an outer ZERO_EXTEND or |
6595 | SIGN_EXTEND as long as Z is already extended (so we don't destroy it). |
6596 | We can do this kind of thing in some cases when STORE_FLAG_VALUE is |
6597 | neither 1 or -1, but it isn't worth checking for. */ |
6598 | |
6599 | if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
6600 | && comparison_p |
6601 | && is_int_mode (mode, int_mode: &int_mode) |
6602 | && ! side_effects_p (x)) |
6603 | { |
6604 | rtx t = make_compound_operation (true_rtx, SET); |
6605 | rtx f = make_compound_operation (false_rtx, SET); |
6606 | rtx cond_op0 = XEXP (cond, 0); |
6607 | rtx cond_op1 = XEXP (cond, 1); |
6608 | enum rtx_code op = UNKNOWN, extend_op = UNKNOWN; |
6609 | scalar_int_mode m = int_mode; |
6610 | rtx z = 0, c1 = NULL_RTX; |
6611 | |
6612 | if ((GET_CODE (t) == PLUS || GET_CODE (t) == MINUS |
6613 | || GET_CODE (t) == IOR || GET_CODE (t) == XOR |
6614 | || GET_CODE (t) == ASHIFT |
6615 | || GET_CODE (t) == LSHIFTRT || GET_CODE (t) == ASHIFTRT) |
6616 | && rtx_equal_p (XEXP (t, 0), f)) |
6617 | c1 = XEXP (t, 1), op = GET_CODE (t), z = f; |
6618 | |
6619 | /* If an identity-zero op is commutative, check whether there |
6620 | would be a match if we swapped the operands. */ |
6621 | else if ((GET_CODE (t) == PLUS || GET_CODE (t) == IOR |
6622 | || GET_CODE (t) == XOR) |
6623 | && rtx_equal_p (XEXP (t, 1), f)) |
6624 | c1 = XEXP (t, 0), op = GET_CODE (t), z = f; |
6625 | else if (GET_CODE (t) == SIGN_EXTEND |
6626 | && is_a <scalar_int_mode> (GET_MODE (XEXP (t, 0)), result: &inner_mode) |
6627 | && (GET_CODE (XEXP (t, 0)) == PLUS |
6628 | || GET_CODE (XEXP (t, 0)) == MINUS |
6629 | || GET_CODE (XEXP (t, 0)) == IOR |
6630 | || GET_CODE (XEXP (t, 0)) == XOR |
6631 | || GET_CODE (XEXP (t, 0)) == ASHIFT |
6632 | || GET_CODE (XEXP (t, 0)) == LSHIFTRT |
6633 | || GET_CODE (XEXP (t, 0)) == ASHIFTRT) |
6634 | && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG |
6635 | && subreg_lowpart_p (XEXP (XEXP (t, 0), 0)) |
6636 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f) |
6637 | && (num_sign_bit_copies (f, GET_MODE (f)) |
6638 | > (unsigned int) |
6639 | (GET_MODE_PRECISION (mode: int_mode) |
6640 | - GET_MODE_PRECISION (mode: inner_mode)))) |
6641 | { |
6642 | c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0)); |
6643 | extend_op = SIGN_EXTEND; |
6644 | m = inner_mode; |
6645 | } |
6646 | else if (GET_CODE (t) == SIGN_EXTEND |
6647 | && is_a <scalar_int_mode> (GET_MODE (XEXP (t, 0)), result: &inner_mode) |
6648 | && (GET_CODE (XEXP (t, 0)) == PLUS |
6649 | || GET_CODE (XEXP (t, 0)) == IOR |
6650 | || GET_CODE (XEXP (t, 0)) == XOR) |
6651 | && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG |
6652 | && subreg_lowpart_p (XEXP (XEXP (t, 0), 1)) |
6653 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f) |
6654 | && (num_sign_bit_copies (f, GET_MODE (f)) |
6655 | > (unsigned int) |
6656 | (GET_MODE_PRECISION (mode: int_mode) |
6657 | - GET_MODE_PRECISION (mode: inner_mode)))) |
6658 | { |
6659 | c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0)); |
6660 | extend_op = SIGN_EXTEND; |
6661 | m = inner_mode; |
6662 | } |
6663 | else if (GET_CODE (t) == ZERO_EXTEND |
6664 | && is_a <scalar_int_mode> (GET_MODE (XEXP (t, 0)), result: &inner_mode) |
6665 | && (GET_CODE (XEXP (t, 0)) == PLUS |
6666 | || GET_CODE (XEXP (t, 0)) == MINUS |
6667 | || GET_CODE (XEXP (t, 0)) == IOR |
6668 | || GET_CODE (XEXP (t, 0)) == XOR |
6669 | || GET_CODE (XEXP (t, 0)) == ASHIFT |
6670 | || GET_CODE (XEXP (t, 0)) == LSHIFTRT |
6671 | || GET_CODE (XEXP (t, 0)) == ASHIFTRT) |
6672 | && GET_CODE (XEXP (XEXP (t, 0), 0)) == SUBREG |
6673 | && HWI_COMPUTABLE_MODE_P (mode: int_mode) |
6674 | && subreg_lowpart_p (XEXP (XEXP (t, 0), 0)) |
6675 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 0)), f) |
6676 | && ((nonzero_bits (f, GET_MODE (f)) |
6677 | & ~GET_MODE_MASK (inner_mode)) |
6678 | == 0)) |
6679 | { |
6680 | c1 = XEXP (XEXP (t, 0), 1); z = f; op = GET_CODE (XEXP (t, 0)); |
6681 | extend_op = ZERO_EXTEND; |
6682 | m = inner_mode; |
6683 | } |
6684 | else if (GET_CODE (t) == ZERO_EXTEND |
6685 | && is_a <scalar_int_mode> (GET_MODE (XEXP (t, 0)), result: &inner_mode) |
6686 | && (GET_CODE (XEXP (t, 0)) == PLUS |
6687 | || GET_CODE (XEXP (t, 0)) == IOR |
6688 | || GET_CODE (XEXP (t, 0)) == XOR) |
6689 | && GET_CODE (XEXP (XEXP (t, 0), 1)) == SUBREG |
6690 | && HWI_COMPUTABLE_MODE_P (mode: int_mode) |
6691 | && subreg_lowpart_p (XEXP (XEXP (t, 0), 1)) |
6692 | && rtx_equal_p (SUBREG_REG (XEXP (XEXP (t, 0), 1)), f) |
6693 | && ((nonzero_bits (f, GET_MODE (f)) |
6694 | & ~GET_MODE_MASK (inner_mode)) |
6695 | == 0)) |
6696 | { |
6697 | c1 = XEXP (XEXP (t, 0), 0); z = f; op = GET_CODE (XEXP (t, 0)); |
6698 | extend_op = ZERO_EXTEND; |
6699 | m = inner_mode; |
6700 | } |
6701 | |
6702 | if (z) |
6703 | { |
6704 | machine_mode cm = m; |
6705 | if ((op == ASHIFT || op == LSHIFTRT || op == ASHIFTRT) |
6706 | && GET_MODE (c1) != VOIDmode) |
6707 | cm = GET_MODE (c1); |
6708 | temp = subst (x: simplify_gen_relational (code: true_code, mode: cm, VOIDmode, |
6709 | op0: cond_op0, op1: cond_op1), |
6710 | from: pc_rtx, to: pc_rtx, in_dest: false, in_cond: false, unique_copy: false); |
6711 | temp = simplify_gen_binary (code: MULT, mode: cm, op0: temp, |
6712 | op1: simplify_gen_binary (code: MULT, mode: cm, op0: c1, |
6713 | op1: const_true_rtx)); |
6714 | temp = subst (x: temp, from: pc_rtx, to: pc_rtx, in_dest: false, in_cond: false, unique_copy: false); |
6715 | temp = simplify_gen_binary (code: op, mode: m, gen_lowpart (m, z), op1: temp); |
6716 | |
6717 | if (extend_op != UNKNOWN) |
6718 | temp = simplify_gen_unary (code: extend_op, mode: int_mode, op: temp, op_mode: m); |
6719 | |
6720 | return temp; |
6721 | } |
6722 | } |
6723 | |
6724 | /* If we have (if_then_else (ne A 0) C1 0) and either A is known to be 0 or |
6725 | 1 and C1 is a single bit or A is known to be 0 or -1 and C1 is the |
6726 | negation of a single bit, we can convert this operation to a shift. We |
6727 | can actually do this more generally, but it doesn't seem worth it. */ |
6728 | |
6729 | if (true_code == NE |
6730 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
6731 | && XEXP (cond, 1) == const0_rtx |
6732 | && false_rtx == const0_rtx |
6733 | && CONST_INT_P (true_rtx) |
6734 | && ((nonzero_bits (XEXP (cond, 0), int_mode) == 1 |
6735 | && (i = exact_log2 (UINTVAL (true_rtx))) >= 0) |
6736 | || ((num_sign_bit_copies (XEXP (cond, 0), int_mode) |
6737 | == GET_MODE_PRECISION (mode: int_mode)) |
6738 | && (i = exact_log2 (x: -UINTVAL (true_rtx))) >= 0))) |
6739 | return |
6740 | simplify_shift_const (NULL_RTX, ASHIFT, int_mode, |
6741 | gen_lowpart (int_mode, XEXP (cond, 0)), i); |
6742 | |
6743 | /* (IF_THEN_ELSE (NE A 0) C1 0) is A or a zero-extend of A if the only |
6744 | non-zero bit in A is C1. */ |
6745 | if (true_code == NE && XEXP (cond, 1) == const0_rtx |
6746 | && false_rtx == const0_rtx && CONST_INT_P (true_rtx) |
6747 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
6748 | && is_a <scalar_int_mode> (GET_MODE (XEXP (cond, 0)), result: &inner_mode) |
6749 | && (UINTVAL (true_rtx) & GET_MODE_MASK (int_mode)) |
6750 | == nonzero_bits (XEXP (cond, 0), inner_mode) |
6751 | && (i = exact_log2 (UINTVAL (true_rtx) & GET_MODE_MASK (int_mode))) >= 0) |
6752 | { |
6753 | rtx val = XEXP (cond, 0); |
6754 | if (inner_mode == int_mode) |
6755 | return val; |
6756 | else if (GET_MODE_PRECISION (mode: inner_mode) < GET_MODE_PRECISION (mode: int_mode)) |
6757 | return simplify_gen_unary (code: ZERO_EXTEND, mode: int_mode, op: val, op_mode: inner_mode); |
6758 | } |
6759 | |
6760 | return x; |
6761 | } |
6762 | |
6763 | /* Simplify X, a SET expression. Return the new expression. */ |
6764 | |
6765 | static rtx |
6766 | simplify_set (rtx x) |
6767 | { |
6768 | rtx src = SET_SRC (x); |
6769 | rtx dest = SET_DEST (x); |
6770 | machine_mode mode |
6771 | = GET_MODE (src) != VOIDmode ? GET_MODE (src) : GET_MODE (dest); |
6772 | rtx_insn *other_insn; |
6773 | rtx *cc_use; |
6774 | scalar_int_mode int_mode; |
6775 | |
6776 | /* (set (pc) (return)) gets written as (return). */ |
6777 | if (GET_CODE (dest) == PC && ANY_RETURN_P (src)) |
6778 | return src; |
6779 | |
6780 | /* Now that we know for sure which bits of SRC we are using, see if we can |
6781 | simplify the expression for the object knowing that we only need the |
6782 | low-order bits. */ |
6783 | |
6784 | if (GET_MODE_CLASS (mode) == MODE_INT && HWI_COMPUTABLE_MODE_P (mode)) |
6785 | { |
6786 | src = force_to_mode (src, mode, HOST_WIDE_INT_M1U, false); |
6787 | SUBST (SET_SRC (x), src); |
6788 | } |
6789 | |
6790 | /* If the source is a COMPARE, look for the use of the comparison result |
6791 | and try to simplify it unless we already have used undobuf.other_insn. */ |
6792 | if ((GET_MODE_CLASS (mode) == MODE_CC || GET_CODE (src) == COMPARE) |
6793 | && (cc_use = find_single_use (dest, insn: subst_insn, ploc: &other_insn)) != 0 |
6794 | && (undobuf.other_insn == 0 || other_insn == undobuf.other_insn) |
6795 | && COMPARISON_P (*cc_use) |
6796 | && rtx_equal_p (XEXP (*cc_use, 0), dest)) |
6797 | { |
6798 | enum rtx_code old_code = GET_CODE (*cc_use); |
6799 | enum rtx_code new_code; |
6800 | rtx op0, op1, tmp; |
6801 | bool other_changed = false; |
6802 | rtx inner_compare = NULL_RTX; |
6803 | machine_mode compare_mode = GET_MODE (dest); |
6804 | |
6805 | if (GET_CODE (src) == COMPARE) |
6806 | { |
6807 | op0 = XEXP (src, 0), op1 = XEXP (src, 1); |
6808 | if (GET_CODE (op0) == COMPARE && op1 == const0_rtx) |
6809 | { |
6810 | inner_compare = op0; |
6811 | op0 = XEXP (inner_compare, 0), op1 = XEXP (inner_compare, 1); |
6812 | } |
6813 | } |
6814 | else |
6815 | op0 = src, op1 = CONST0_RTX (GET_MODE (src)); |
6816 | |
6817 | tmp = simplify_relational_operation (code: old_code, mode: compare_mode, VOIDmode, |
6818 | op0, op1); |
6819 | if (!tmp) |
6820 | new_code = old_code; |
6821 | else if (!CONSTANT_P (tmp)) |
6822 | { |
6823 | new_code = GET_CODE (tmp); |
6824 | op0 = XEXP (tmp, 0); |
6825 | op1 = XEXP (tmp, 1); |
6826 | } |
6827 | else |
6828 | { |
6829 | rtx pat = PATTERN (insn: other_insn); |
6830 | undobuf.other_insn = other_insn; |
6831 | SUBST (*cc_use, tmp); |
6832 | |
6833 | /* Attempt to simplify CC user. */ |
6834 | if (GET_CODE (pat) == SET) |
6835 | { |
6836 | rtx new_rtx = simplify_rtx (SET_SRC (pat)); |
6837 | if (new_rtx != NULL_RTX) |
6838 | SUBST (SET_SRC (pat), new_rtx); |
6839 | } |
6840 | |
6841 | /* Convert X into a no-op move. */ |
6842 | SUBST (SET_DEST (x), pc_rtx); |
6843 | SUBST (SET_SRC (x), pc_rtx); |
6844 | return x; |
6845 | } |
6846 | |
6847 | /* Simplify our comparison, if possible. */ |
6848 | new_code = simplify_comparison (new_code, &op0, &op1); |
6849 | |
6850 | #ifdef SELECT_CC_MODE |
6851 | /* If this machine has CC modes other than CCmode, check to see if we |
6852 | need to use a different CC mode here. */ |
6853 | if (GET_MODE_CLASS (GET_MODE (op0)) == MODE_CC) |
6854 | compare_mode = GET_MODE (op0); |
6855 | else if (inner_compare |
6856 | && GET_MODE_CLASS (GET_MODE (inner_compare)) == MODE_CC |
6857 | && new_code == old_code |
6858 | && op0 == XEXP (inner_compare, 0) |
6859 | && op1 == XEXP (inner_compare, 1)) |
6860 | compare_mode = GET_MODE (inner_compare); |
6861 | else |
6862 | compare_mode = SELECT_CC_MODE (new_code, op0, op1); |
6863 | |
6864 | /* If the mode changed, we have to change SET_DEST, the mode in the |
6865 | compare, and the mode in the place SET_DEST is used. If SET_DEST is |
6866 | a hard register, just build new versions with the proper mode. If it |
6867 | is a pseudo, we lose unless it is only time we set the pseudo, in |
6868 | which case we can safely change its mode. */ |
6869 | if (compare_mode != GET_MODE (dest)) |
6870 | { |
6871 | if (can_change_dest_mode (x: dest, added_sets: 0, mode: compare_mode)) |
6872 | { |
6873 | unsigned int regno = REGNO (dest); |
6874 | rtx new_dest; |
6875 | |
6876 | if (regno < FIRST_PSEUDO_REGISTER) |
6877 | new_dest = gen_rtx_REG (compare_mode, regno); |
6878 | else |
6879 | { |
6880 | subst_mode (regno, newval: compare_mode); |
6881 | new_dest = regno_reg_rtx[regno]; |
6882 | } |
6883 | |
6884 | SUBST (SET_DEST (x), new_dest); |
6885 | SUBST (XEXP (*cc_use, 0), new_dest); |
6886 | other_changed = true; |
6887 | |
6888 | dest = new_dest; |
6889 | } |
6890 | } |
6891 | #endif /* SELECT_CC_MODE */ |
6892 | |
6893 | /* If the code changed, we have to build a new comparison in |
6894 | undobuf.other_insn. */ |
6895 | if (new_code != old_code) |
6896 | { |
6897 | bool other_changed_previously = other_changed; |
6898 | unsigned HOST_WIDE_INT mask; |
6899 | rtx old_cc_use = *cc_use; |
6900 | |
6901 | SUBST (*cc_use, gen_rtx_fmt_ee (new_code, GET_MODE (*cc_use), |
6902 | dest, const0_rtx)); |
6903 | other_changed = true; |
6904 | |
6905 | /* If the only change we made was to change an EQ into an NE or |
6906 | vice versa, OP0 has only one bit that might be nonzero, and OP1 |
6907 | is zero, check if changing the user of the condition code will |
6908 | produce a valid insn. If it won't, we can keep the original code |
6909 | in that insn by surrounding our operation with an XOR. */ |
6910 | |
6911 | if (((old_code == NE && new_code == EQ) |
6912 | || (old_code == EQ && new_code == NE)) |
6913 | && ! other_changed_previously && op1 == const0_rtx |
6914 | && HWI_COMPUTABLE_MODE_P (GET_MODE (op0)) |
6915 | && pow2p_hwi (x: mask = nonzero_bits (op0, GET_MODE (op0)))) |
6916 | { |
6917 | rtx pat = PATTERN (insn: other_insn), note = 0; |
6918 | |
6919 | if ((recog_for_combine (&pat, other_insn, ¬e) < 0 |
6920 | && ! check_asm_operands (pat))) |
6921 | { |
6922 | *cc_use = old_cc_use; |
6923 | other_changed = false; |
6924 | |
6925 | op0 = simplify_gen_binary (code: XOR, GET_MODE (op0), op0, |
6926 | op1: gen_int_mode (mask, |
6927 | GET_MODE (op0))); |
6928 | } |
6929 | } |
6930 | } |
6931 | |
6932 | if (other_changed) |
6933 | undobuf.other_insn = other_insn; |
6934 | |
6935 | /* Don't generate a compare of a CC with 0, just use that CC. */ |
6936 | if (GET_MODE (op0) == compare_mode && op1 == const0_rtx) |
6937 | { |
6938 | SUBST (SET_SRC (x), op0); |
6939 | src = SET_SRC (x); |
6940 | } |
6941 | /* Otherwise, if we didn't previously have the same COMPARE we |
6942 | want, create it from scratch. */ |
6943 | else if (GET_CODE (src) != COMPARE || GET_MODE (src) != compare_mode |
6944 | || XEXP (src, 0) != op0 || XEXP (src, 1) != op1) |
6945 | { |
6946 | SUBST (SET_SRC (x), gen_rtx_COMPARE (compare_mode, op0, op1)); |
6947 | src = SET_SRC (x); |
6948 | } |
6949 | } |
6950 | else |
6951 | { |
6952 | /* Get SET_SRC in a form where we have placed back any |
6953 | compound expressions. Then do the checks below. */ |
6954 | src = make_compound_operation (src, SET); |
6955 | SUBST (SET_SRC (x), src); |
6956 | } |
6957 | |
6958 | /* If we have (set x (subreg:m1 (op:m2 ...) 0)) with OP being some operation, |
6959 | and X being a REG or (subreg (reg)), we may be able to convert this to |
6960 | (set (subreg:m2 x) (op)). |
6961 | |
6962 | We can always do this if M1 is narrower than M2 because that means that |
6963 | we only care about the low bits of the result. |
6964 | |
6965 | However, on machines without WORD_REGISTER_OPERATIONS defined, we cannot |
6966 | perform a narrower operation than requested since the high-order bits will |
6967 | be undefined. On machine where it is defined, this transformation is safe |
6968 | as long as M1 and M2 have the same number of words. */ |
6969 | |
6970 | if (GET_CODE (src) == SUBREG && subreg_lowpart_p (src) |
6971 | && !OBJECT_P (SUBREG_REG (src)) |
6972 | && (known_equal_after_align_up |
6973 | (a: GET_MODE_SIZE (GET_MODE (src)), |
6974 | b: GET_MODE_SIZE (GET_MODE (SUBREG_REG (src))), |
6975 | UNITS_PER_WORD)) |
6976 | && (WORD_REGISTER_OPERATIONS || !paradoxical_subreg_p (x: src)) |
6977 | && ! (REG_P (dest) && REGNO (dest) < FIRST_PSEUDO_REGISTER |
6978 | && !REG_CAN_CHANGE_MODE_P (REGNO (dest), |
6979 | GET_MODE (SUBREG_REG (src)), |
6980 | GET_MODE (src))) |
6981 | && (REG_P (dest) |
6982 | || (GET_CODE (dest) == SUBREG |
6983 | && REG_P (SUBREG_REG (dest))))) |
6984 | { |
6985 | SUBST (SET_DEST (x), |
6986 | gen_lowpart (GET_MODE (SUBREG_REG (src)), |
6987 | dest)); |
6988 | SUBST (SET_SRC (x), SUBREG_REG (src)); |
6989 | |
6990 | src = SET_SRC (x), dest = SET_DEST (x); |
6991 | } |
6992 | |
6993 | /* If we have (set FOO (subreg:M (mem:N BAR) 0)) with M wider than N, this |
6994 | would require a paradoxical subreg. Replace the subreg with a |
6995 | zero_extend to avoid the reload that would otherwise be required. |
6996 | Don't do this unless we have a scalar integer mode, otherwise the |
6997 | transformation is incorrect. */ |
6998 | |
6999 | enum rtx_code extend_op; |
7000 | if (paradoxical_subreg_p (x: src) |
7001 | && MEM_P (SUBREG_REG (src)) |
7002 | && SCALAR_INT_MODE_P (GET_MODE (src)) |
7003 | && (extend_op = load_extend_op (GET_MODE (SUBREG_REG (src)))) != UNKNOWN) |
7004 | { |
7005 | SUBST (SET_SRC (x), |
7006 | gen_rtx_fmt_e (extend_op, GET_MODE (src), SUBREG_REG (src))); |
7007 | |
7008 | src = SET_SRC (x); |
7009 | } |
7010 | |
7011 | /* If we don't have a conditional move, SET_SRC is an IF_THEN_ELSE, and we |
7012 | are comparing an item known to be 0 or -1 against 0, use a logical |
7013 | operation instead. Check for one of the arms being an IOR of the other |
7014 | arm with some value. We compute three terms to be IOR'ed together. In |
7015 | practice, at most two will be nonzero. Then we do the IOR's. */ |
7016 | |
7017 | if (GET_CODE (dest) != PC |
7018 | && GET_CODE (src) == IF_THEN_ELSE |
7019 | && is_int_mode (GET_MODE (src), int_mode: &int_mode) |
7020 | && (GET_CODE (XEXP (src, 0)) == EQ || GET_CODE (XEXP (src, 0)) == NE) |
7021 | && XEXP (XEXP (src, 0), 1) == const0_rtx |
7022 | && int_mode == GET_MODE (XEXP (XEXP (src, 0), 0)) |
7023 | && (!HAVE_conditional_move |
7024 | || ! can_conditionally_move_p (mode: int_mode)) |
7025 | && (num_sign_bit_copies (XEXP (XEXP (src, 0), 0), int_mode) |
7026 | == GET_MODE_PRECISION (mode: int_mode)) |
7027 | && ! side_effects_p (src)) |
7028 | { |
7029 | rtx true_rtx = (GET_CODE (XEXP (src, 0)) == NE |
7030 | ? XEXP (src, 1) : XEXP (src, 2)); |
7031 | rtx false_rtx = (GET_CODE (XEXP (src, 0)) == NE |
7032 | ? XEXP (src, 2) : XEXP (src, 1)); |
7033 | rtx term1 = const0_rtx, term2, term3; |
7034 | |
7035 | if (GET_CODE (true_rtx) == IOR |
7036 | && rtx_equal_p (XEXP (true_rtx, 0), false_rtx)) |
7037 | term1 = false_rtx, true_rtx = XEXP (true_rtx, 1), false_rtx = const0_rtx; |
7038 | else if (GET_CODE (true_rtx) == IOR |
7039 | && rtx_equal_p (XEXP (true_rtx, 1), false_rtx)) |
7040 | term1 = false_rtx, true_rtx = XEXP (true_rtx, 0), false_rtx = const0_rtx; |
7041 | else if (GET_CODE (false_rtx) == IOR |
7042 | && rtx_equal_p (XEXP (false_rtx, 0), true_rtx)) |
7043 | term1 = true_rtx, false_rtx = XEXP (false_rtx, 1), true_rtx = const0_rtx; |
7044 | else if (GET_CODE (false_rtx) == IOR |
7045 | && rtx_equal_p (XEXP (false_rtx, 1), true_rtx)) |
7046 | term1 = true_rtx, false_rtx = XEXP (false_rtx, 0), true_rtx = const0_rtx; |
7047 | |
7048 | term2 = simplify_gen_binary (code: AND, mode: int_mode, |
7049 | XEXP (XEXP (src, 0), 0), op1: true_rtx); |
7050 | term3 = simplify_gen_binary (code: AND, mode: int_mode, |
7051 | op0: simplify_gen_unary (code: NOT, mode: int_mode, |
7052 | XEXP (XEXP (src, 0), 0), |
7053 | op_mode: int_mode), |
7054 | op1: false_rtx); |
7055 | |
7056 | SUBST (SET_SRC (x), |
7057 | simplify_gen_binary (IOR, int_mode, |
7058 | simplify_gen_binary (IOR, int_mode, |
7059 | term1, term2), |
7060 | term3)); |
7061 | |
7062 | src = SET_SRC (x); |
7063 | } |
7064 | |
7065 | /* If either SRC or DEST is a CLOBBER of (const_int 0), make this |
7066 | whole thing fail. */ |
7067 | if (GET_CODE (src) == CLOBBER && XEXP (src, 0) == const0_rtx) |
7068 | return src; |
7069 | else if (GET_CODE (dest) == CLOBBER && XEXP (dest, 0) == const0_rtx) |
7070 | return dest; |
7071 | else |
7072 | /* Convert this into a field assignment operation, if possible. */ |
7073 | return make_field_assignment (x); |
7074 | } |
7075 | |
7076 | /* Simplify, X, and AND, IOR, or XOR operation, and return the simplified |
7077 | result. */ |
7078 | |
7079 | static rtx |
7080 | simplify_logical (rtx x) |
7081 | { |
7082 | rtx op0 = XEXP (x, 0); |
7083 | rtx op1 = XEXP (x, 1); |
7084 | scalar_int_mode mode; |
7085 | |
7086 | switch (GET_CODE (x)) |
7087 | { |
7088 | case AND: |
7089 | /* We can call simplify_and_const_int only if we don't lose |
7090 | any (sign) bits when converting INTVAL (op1) to |
7091 | "unsigned HOST_WIDE_INT". */ |
7092 | if (is_a <scalar_int_mode> (GET_MODE (x), result: &mode) |
7093 | && CONST_INT_P (op1) |
7094 | && (HWI_COMPUTABLE_MODE_P (mode) |
7095 | || INTVAL (op1) > 0)) |
7096 | { |
7097 | x = simplify_and_const_int (x, mode, op0, INTVAL (op1)); |
7098 | if (GET_CODE (x) != AND) |
7099 | return x; |
7100 | |
7101 | op0 = XEXP (x, 0); |
7102 | op1 = XEXP (x, 1); |
7103 | } |
7104 | |
7105 | /* If we have any of (and (ior A B) C) or (and (xor A B) C), |
7106 | apply the distributive law and then the inverse distributive |
7107 | law to see if things simplify. */ |
7108 | if (GET_CODE (op0) == IOR || GET_CODE (op0) == XOR) |
7109 | { |
7110 | rtx result = distribute_and_simplify_rtx (x, 0); |
7111 | if (result) |
7112 | return result; |
7113 | } |
7114 | if (GET_CODE (op1) == IOR || GET_CODE (op1) == XOR) |
7115 | { |
7116 | rtx result = distribute_and_simplify_rtx (x, 1); |
7117 | if (result) |
7118 | return result; |
7119 | } |
7120 | break; |
7121 | |
7122 | case IOR: |
7123 | /* If we have (ior (and A B) C), apply the distributive law and then |
7124 | the inverse distributive law to see if things simplify. */ |
7125 | |
7126 | if (GET_CODE (op0) == AND) |
7127 | { |
7128 | rtx result = distribute_and_simplify_rtx (x, 0); |
7129 | if (result) |
7130 | return result; |
7131 | } |
7132 | |
7133 | if (GET_CODE (op1) == AND) |
7134 | { |
7135 | rtx result = distribute_and_simplify_rtx (x, 1); |
7136 | if (result) |
7137 | return result; |
7138 | } |
7139 | break; |
7140 | |
7141 | default: |
7142 | gcc_unreachable (); |
7143 | } |
7144 | |
7145 | return x; |
7146 | } |
7147 | |
7148 | /* We consider ZERO_EXTRACT, SIGN_EXTRACT, and SIGN_EXTEND as "compound |
7149 | operations" because they can be replaced with two more basic operations. |
7150 | ZERO_EXTEND is also considered "compound" because it can be replaced with |
7151 | an AND operation, which is simpler, though only one operation. |
7152 | |
7153 | The function expand_compound_operation is called with an rtx expression |
7154 | and will convert it to the appropriate shifts and AND operations, |
7155 | simplifying at each stage. |
7156 | |
7157 | The function make_compound_operation is called to convert an expression |
7158 | consisting of shifts and ANDs into the equivalent compound expression. |
7159 | It is the inverse of this function, loosely speaking. */ |
7160 | |
7161 | static rtx |
7162 | expand_compound_operation (rtx x) |
7163 | { |
7164 | unsigned HOST_WIDE_INT pos = 0, len; |
7165 | bool unsignedp = false; |
7166 | unsigned int modewidth; |
7167 | rtx tem; |
7168 | scalar_int_mode inner_mode; |
7169 | |
7170 | switch (GET_CODE (x)) |
7171 | { |
7172 | case ZERO_EXTEND: |
7173 | unsignedp = true; |
7174 | /* FALLTHRU */ |
7175 | case SIGN_EXTEND: |
7176 | /* We can't necessarily use a const_int for a multiword mode; |
7177 | it depends on implicitly extending the value. |
7178 | Since we don't know the right way to extend it, |
7179 | we can't tell whether the implicit way is right. |
7180 | |
7181 | Even for a mode that is no wider than a const_int, |
7182 | we can't win, because we need to sign extend one of its bits through |
7183 | the rest of it, and we don't know which bit. */ |
7184 | if (CONST_INT_P (XEXP (x, 0))) |
7185 | return x; |
7186 | |
7187 | /* Reject modes that aren't scalar integers because turning vector |
7188 | or complex modes into shifts causes problems. */ |
7189 | if (!is_a <scalar_int_mode> (GET_MODE (XEXP (x, 0)), result: &inner_mode)) |
7190 | return x; |
7191 | |
7192 | /* Return if (subreg:MODE FROM 0) is not a safe replacement for |
7193 | (zero_extend:MODE FROM) or (sign_extend:MODE FROM). It is for any MEM |
7194 | because (SUBREG (MEM...)) is guaranteed to cause the MEM to be |
7195 | reloaded. If not for that, MEM's would very rarely be safe. |
7196 | |
7197 | Reject modes bigger than a word, because we might not be able |
7198 | to reference a two-register group starting with an arbitrary register |
7199 | (and currently gen_lowpart might crash for a SUBREG). */ |
7200 | |
7201 | if (GET_MODE_SIZE (mode: inner_mode) > UNITS_PER_WORD) |
7202 | return x; |
7203 | |
7204 | len = GET_MODE_PRECISION (mode: inner_mode); |
7205 | /* If the inner object has VOIDmode (the only way this can happen |
7206 | is if it is an ASM_OPERANDS), we can't do anything since we don't |
7207 | know how much masking to do. */ |
7208 | if (len == 0) |
7209 | return x; |
7210 | |
7211 | break; |
7212 | |
7213 | case ZERO_EXTRACT: |
7214 | unsignedp = true; |
7215 | |
7216 | /* fall through */ |
7217 | |
7218 | case SIGN_EXTRACT: |
7219 | /* If the operand is a CLOBBER, just return it. */ |
7220 | if (GET_CODE (XEXP (x, 0)) == CLOBBER) |
7221 | return XEXP (x, 0); |
7222 | |
7223 | if (!CONST_INT_P (XEXP (x, 1)) |
7224 | || !CONST_INT_P (XEXP (x, 2))) |
7225 | return x; |
7226 | |
7227 | /* Reject modes that aren't scalar integers because turning vector |
7228 | or complex modes into shifts causes problems. */ |
7229 | if (!is_a <scalar_int_mode> (GET_MODE (XEXP (x, 0)), result: &inner_mode)) |
7230 | return x; |
7231 | |
7232 | len = INTVAL (XEXP (x, 1)); |
7233 | pos = INTVAL (XEXP (x, 2)); |
7234 | |
7235 | /* This should stay within the object being extracted, fail otherwise. */ |
7236 | if (len + pos > GET_MODE_PRECISION (mode: inner_mode)) |
7237 | return x; |
7238 | |
7239 | if (BITS_BIG_ENDIAN) |
7240 | pos = GET_MODE_PRECISION (mode: inner_mode) - len - pos; |
7241 | |
7242 | break; |
7243 | |
7244 | default: |
7245 | return x; |
7246 | } |
7247 | |
7248 | /* We've rejected non-scalar operations by now. */ |
7249 | scalar_int_mode mode = as_a <scalar_int_mode> (GET_MODE (x)); |
7250 | |
7251 | /* Convert sign extension to zero extension, if we know that the high |
7252 | bit is not set, as this is easier to optimize. It will be converted |
7253 | back to cheaper alternative in make_extraction. */ |
7254 | if (GET_CODE (x) == SIGN_EXTEND |
7255 | && HWI_COMPUTABLE_MODE_P (mode) |
7256 | && ((nonzero_bits (XEXP (x, 0), inner_mode) |
7257 | & ~(((unsigned HOST_WIDE_INT) GET_MODE_MASK (inner_mode)) >> 1)) |
7258 | == 0)) |
7259 | { |
7260 | rtx temp = gen_rtx_ZERO_EXTEND (mode, XEXP (x, 0)); |
7261 | rtx temp2 = expand_compound_operation (x: temp); |
7262 | |
7263 | /* Make sure this is a profitable operation. */ |
7264 | if (set_src_cost (x, mode, speed_p: optimize_this_for_speed_p) |
7265 | > set_src_cost (x: temp2, mode, speed_p: optimize_this_for_speed_p)) |
7266 | return temp2; |
7267 | else if (set_src_cost (x, mode, speed_p: optimize_this_for_speed_p) |
7268 | > set_src_cost (x: temp, mode, speed_p: optimize_this_for_speed_p)) |
7269 | return temp; |
7270 | else |
7271 | return x; |
7272 | } |
7273 | |
7274 | /* We can optimize some special cases of ZERO_EXTEND. */ |
7275 | if (GET_CODE (x) == ZERO_EXTEND) |
7276 | { |
7277 | /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI if we |
7278 | know that the last value didn't have any inappropriate bits |
7279 | set. */ |
7280 | if (GET_CODE (XEXP (x, 0)) == TRUNCATE |
7281 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == mode |
7282 | && HWI_COMPUTABLE_MODE_P (mode) |
7283 | && (nonzero_bits (XEXP (XEXP (x, 0), 0), mode) |
7284 | & ~GET_MODE_MASK (inner_mode)) == 0) |
7285 | return XEXP (XEXP (x, 0), 0); |
7286 | |
7287 | /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */ |
7288 | if (GET_CODE (XEXP (x, 0)) == SUBREG |
7289 | && GET_MODE (SUBREG_REG (XEXP (x, 0))) == mode |
7290 | && subreg_lowpart_p (XEXP (x, 0)) |
7291 | && HWI_COMPUTABLE_MODE_P (mode) |
7292 | && (nonzero_bits (SUBREG_REG (XEXP (x, 0)), mode) |
7293 | & ~GET_MODE_MASK (inner_mode)) == 0) |
7294 | return SUBREG_REG (XEXP (x, 0)); |
7295 | |
7296 | /* (zero_extend:DI (truncate:SI foo:DI)) is just foo:DI when foo |
7297 | is a comparison and STORE_FLAG_VALUE permits. This is like |
7298 | the first case, but it works even when MODE is larger |
7299 | than HOST_WIDE_INT. */ |
7300 | if (GET_CODE (XEXP (x, 0)) == TRUNCATE |
7301 | && GET_MODE (XEXP (XEXP (x, 0), 0)) == mode |
7302 | && COMPARISON_P (XEXP (XEXP (x, 0), 0)) |
7303 | && GET_MODE_PRECISION (mode: inner_mode) <= HOST_BITS_PER_WIDE_INT |
7304 | && (STORE_FLAG_VALUE & ~GET_MODE_MASK (inner_mode)) == 0) |
7305 | return XEXP (XEXP (x, 0), 0); |
7306 | |
7307 | /* Likewise for (zero_extend:DI (subreg:SI foo:DI 0)). */ |
7308 | if (GET_CODE (XEXP (x, 0)) == SUBREG |
7309 | && GET_MODE (SUBREG_REG (XEXP (x, 0))) == mode |
7310 | && subreg_lowpart_p (XEXP (x, 0)) |
7311 | && COMPARISON_P (SUBREG_REG (XEXP (x, 0))) |
7312 | && GET_MODE_PRECISION (mode: inner_mode) <= HOST_BITS_PER_WIDE_INT |
7313 | && (STORE_FLAG_VALUE & ~GET_MODE_MASK (inner_mode)) == 0) |
7314 | return SUBREG_REG (XEXP (x, 0)); |
7315 | |
7316 | } |
7317 | |
7318 | /* If we reach here, we want to return a pair of shifts. The inner |
7319 | shift is a left shift of BITSIZE - POS - LEN bits. The outer |
7320 | shift is a right shift of BITSIZE - LEN bits. It is arithmetic or |
7321 | logical depending on the value of UNSIGNEDP. |
7322 | |
7323 | If this was a ZERO_EXTEND or ZERO_EXTRACT, this pair of shifts will be |
7324 | converted into an AND of a shift. |
7325 | |
7326 | We must check for the case where the left shift would have a negative |
7327 | count. This can happen in a case like (x >> 31) & 255 on machines |
7328 | that can't shift by a constant. On those machines, we would first |
7329 | combine the shift with the AND to produce a variable-position |
7330 | extraction. Then the constant of 31 would be substituted in |
7331 | to produce such a position. */ |
7332 | |
7333 | modewidth = GET_MODE_PRECISION (mode); |
7334 | if (modewidth >= pos + len) |
7335 | { |
7336 | tem = gen_lowpart (mode, XEXP (x, 0)); |
7337 | if (!tem || GET_CODE (tem) == CLOBBER) |
7338 | return x; |
7339 | tem = simplify_shift_const (NULL_RTX, ASHIFT, mode, |
7340 | tem, modewidth - pos - len); |
7341 | tem = simplify_shift_const (NULL_RTX, unsignedp ? LSHIFTRT : ASHIFTRT, |
7342 | mode, tem, modewidth - len); |
7343 | } |
7344 | else if (unsignedp && len < HOST_BITS_PER_WIDE_INT) |
7345 | { |
7346 | tem = simplify_shift_const (NULL_RTX, LSHIFTRT, inner_mode, |
7347 | XEXP (x, 0), pos); |
7348 | tem = gen_lowpart (mode, tem); |
7349 | if (!tem || GET_CODE (tem) == CLOBBER) |
7350 | return x; |
7351 | tem = simplify_and_const_int (NULL_RTX, mode, tem, |
7352 | (HOST_WIDE_INT_1U << len) - 1); |
7353 | } |
7354 | else |
7355 | /* Any other cases we can't handle. */ |
7356 | return x; |
7357 | |
7358 | /* If we couldn't do this for some reason, return the original |
7359 | expression. */ |
7360 | if (GET_CODE (tem) == CLOBBER) |
7361 | return x; |
7362 | |
7363 | return tem; |
7364 | } |
7365 | |
7366 | /* X is a SET which contains an assignment of one object into |
7367 | a part of another (such as a bit-field assignment, STRICT_LOW_PART, |
7368 | or certain SUBREGS). If possible, convert it into a series of |
7369 | logical operations. |
7370 | |
7371 | We half-heartedly support variable positions, but do not at all |
7372 | support variable lengths. */ |
7373 | |
7374 | static const_rtx |
7375 | expand_field_assignment (const_rtx x) |
7376 | { |
7377 | rtx inner; |
7378 | rtx pos; /* Always counts from low bit. */ |
7379 | int len, inner_len; |
7380 | rtx mask, cleared, masked; |
7381 | scalar_int_mode compute_mode; |
7382 | |
7383 | /* Loop until we find something we can't simplify. */ |
7384 | while (1) |
7385 | { |
7386 | if (GET_CODE (SET_DEST (x)) == STRICT_LOW_PART |
7387 | && GET_CODE (XEXP (SET_DEST (x), 0)) == SUBREG) |
7388 | { |
7389 | rtx x0 = XEXP (SET_DEST (x), 0); |
7390 | if (!GET_MODE_PRECISION (GET_MODE (x0)).is_constant (const_value: &len)) |
7391 | break; |
7392 | inner = SUBREG_REG (XEXP (SET_DEST (x), 0)); |
7393 | pos = gen_int_mode (subreg_lsb (XEXP (SET_DEST (x), 0)), |
7394 | MAX_MODE_INT); |
7395 | } |
7396 | else if (GET_CODE (SET_DEST (x)) == ZERO_EXTRACT |
7397 | && CONST_INT_P (XEXP (SET_DEST (x), 1))) |
7398 | { |
7399 | inner = XEXP (SET_DEST (x), 0); |
7400 | if (!GET_MODE_PRECISION (GET_MODE (inner)).is_constant (const_value: &inner_len)) |
7401 | break; |
7402 | |
7403 | len = INTVAL (XEXP (SET_DEST (x), 1)); |
7404 | pos = XEXP (SET_DEST (x), 2); |
7405 | |
7406 | /* A constant position should stay within the width of INNER. */ |
7407 | if (CONST_INT_P (pos) && INTVAL (pos) + len > inner_len) |
7408 | break; |
7409 | |
7410 | if (BITS_BIG_ENDIAN) |
7411 | { |
7412 | if (CONST_INT_P (pos)) |
7413 | pos = GEN_INT (inner_len - len - INTVAL (pos)); |
7414 | else if (GET_CODE (pos) == MINUS |
7415 | && CONST_INT_P (XEXP (pos, 1)) |
7416 | && INTVAL (XEXP (pos, 1)) == inner_len - len) |
7417 | /* If position is ADJUST - X, new position is X. */ |
7418 | pos = XEXP (pos, 0); |
7419 | else |
7420 | pos = simplify_gen_binary (code: MINUS, GET_MODE (pos), |
7421 | op0: gen_int_mode (inner_len - len, |
7422 | GET_MODE (pos)), |
7423 | op1: pos); |
7424 | } |
7425 | } |
7426 | |
7427 | /* If the destination is a subreg that overwrites the whole of the inner |
7428 | register, we can move the subreg to the source. */ |
7429 | else if (GET_CODE (SET_DEST (x)) == SUBREG |
7430 | /* We need SUBREGs to compute nonzero_bits properly. */ |
7431 | && nonzero_sign_valid |
7432 | && !read_modify_subreg_p (SET_DEST (x))) |
7433 | { |
7434 | x = gen_rtx_SET (SUBREG_REG (SET_DEST (x)), |
7435 | gen_lowpart |
7436 | (GET_MODE (SUBREG_REG (SET_DEST (x))), |
7437 | SET_SRC (x))); |
7438 | continue; |
7439 | } |
7440 | else |
7441 | break; |
7442 | |
7443 | while (GET_CODE (inner) == SUBREG && subreg_lowpart_p (inner)) |
7444 | inner = SUBREG_REG (inner); |
7445 | |
7446 | /* Don't attempt bitwise arithmetic on non scalar integer modes. */ |
7447 | if (!is_a <scalar_int_mode> (GET_MODE (inner), result: &compute_mode)) |
7448 | { |
7449 | /* Don't do anything for vector or complex integral types. */ |
7450 | if (! FLOAT_MODE_P (GET_MODE (inner))) |
7451 | break; |
7452 | |
7453 | /* Try to find an integral mode to pun with. */ |
7454 | if (!int_mode_for_size (size: GET_MODE_BITSIZE (GET_MODE (inner)), limit: 0) |
7455 | .exists (mode: &compute_mode)) |
7456 | break; |
7457 | |
7458 | inner = gen_lowpart (compute_mode, inner); |
7459 | } |
7460 | |
7461 | /* Compute a mask of LEN bits, if we can do this on the host machine. */ |
7462 | if (len >= HOST_BITS_PER_WIDE_INT) |
7463 | break; |
7464 | |
7465 | /* Don't try to compute in too wide unsupported modes. */ |
7466 | if (!targetm.scalar_mode_supported_p (compute_mode)) |
7467 | break; |
7468 | |
7469 | /* Now compute the equivalent expression. Make a copy of INNER |
7470 | for the SET_DEST in case it is a MEM into which we will substitute; |
7471 | we don't want shared RTL in that case. */ |
7472 | mask = gen_int_mode ((HOST_WIDE_INT_1U << len) - 1, |
7473 | compute_mode); |
7474 | cleared = simplify_gen_binary (code: AND, mode: compute_mode, |
7475 | op0: simplify_gen_unary (code: NOT, mode: compute_mode, |
7476 | op: simplify_gen_binary (code: ASHIFT, |
7477 | mode: compute_mode, |
7478 | op0: mask, op1: pos), |
7479 | op_mode: compute_mode), |
7480 | op1: inner); |
7481 | masked = simplify_gen_binary (code: ASHIFT, mode: compute_mode, |
7482 | op0: simplify_gen_binary ( |
7483 | code: AND, mode: compute_mode, |
7484 | gen_lowpart (compute_mode, SET_SRC (x)), |
7485 | op1: mask), |
7486 | op1: pos); |
7487 | |
7488 | x = gen_rtx_SET (copy_rtx (inner), |
7489 | simplify_gen_binary (IOR, compute_mode, |
7490 | cleared, masked)); |
7491 | } |
7492 | |
7493 | return x; |
7494 | } |
7495 | |
7496 | /* Return an RTX for a reference to LEN bits of INNER. If POS_RTX is nonzero, |
7497 | it is an RTX that represents the (variable) starting position; otherwise, |
7498 | POS is the (constant) starting bit position. Both are counted from the LSB. |
7499 | |
7500 | UNSIGNEDP is true for an unsigned reference and zero for a signed one. |
7501 | |
7502 | IN_DEST is true if this is a reference in the destination of a SET. |
7503 | This is used when a ZERO_ or SIGN_EXTRACT isn't needed. If nonzero, |
7504 | a STRICT_LOW_PART will be used, if zero, ZERO_EXTEND or SIGN_EXTEND will |
7505 | be used. |
7506 | |
7507 | IN_COMPARE is true if we are in a COMPARE. This means that a |
7508 | ZERO_EXTRACT should be built even for bits starting at bit 0. |
7509 | |
7510 | MODE is the desired mode of the result (if IN_DEST == 0). |
7511 | |
7512 | The result is an RTX for the extraction or NULL_RTX if the target |
7513 | can't handle it. */ |
7514 | |
7515 | static rtx |
7516 | (machine_mode mode, rtx inner, HOST_WIDE_INT pos, |
7517 | rtx pos_rtx, unsigned HOST_WIDE_INT len, bool unsignedp, |
7518 | bool in_dest, bool in_compare) |
7519 | { |
7520 | /* This mode describes the size of the storage area |
7521 | to fetch the overall value from. Within that, we |
7522 | ignore the POS lowest bits, etc. */ |
7523 | machine_mode is_mode = GET_MODE (inner); |
7524 | machine_mode inner_mode; |
7525 | scalar_int_mode wanted_inner_mode; |
7526 | scalar_int_mode wanted_inner_reg_mode = word_mode; |
7527 | scalar_int_mode pos_mode = word_mode; |
7528 | machine_mode = word_mode; |
7529 | rtx new_rtx = 0; |
7530 | rtx orig_pos_rtx = pos_rtx; |
7531 | HOST_WIDE_INT orig_pos; |
7532 | |
7533 | if (pos_rtx && CONST_INT_P (pos_rtx)) |
7534 | pos = INTVAL (pos_rtx), pos_rtx = 0; |
7535 | |
7536 | if (GET_CODE (inner) == SUBREG |
7537 | && subreg_lowpart_p (inner) |
7538 | && (paradoxical_subreg_p (x: inner) |
7539 | /* If trying or potentionally trying to extract |
7540 | bits outside of is_mode, don't look through |
7541 | non-paradoxical SUBREGs. See PR82192. */ |
7542 | || (pos_rtx == NULL_RTX |
7543 | && known_le (pos + len, GET_MODE_PRECISION (is_mode))))) |
7544 | { |
7545 | /* If going from (subreg:SI (mem:QI ...)) to (mem:QI ...), |
7546 | consider just the QI as the memory to extract from. |
7547 | The subreg adds or removes high bits; its mode is |
7548 | irrelevant to the meaning of this extraction, |
7549 | since POS and LEN count from the lsb. */ |
7550 | if (MEM_P (SUBREG_REG (inner))) |
7551 | is_mode = GET_MODE (SUBREG_REG (inner)); |
7552 | inner = SUBREG_REG (inner); |
7553 | } |
7554 | else if (GET_CODE (inner) == ASHIFT |
7555 | && CONST_INT_P (XEXP (inner, 1)) |
7556 | && pos_rtx == 0 && pos == 0 |
7557 | && len > UINTVAL (XEXP (inner, 1))) |
7558 | { |
7559 | /* We're extracting the least significant bits of an rtx |
7560 | (ashift X (const_int C)), where LEN > C. Extract the |
7561 | least significant (LEN - C) bits of X, giving an rtx |
7562 | whose mode is MODE, then shift it left C times. */ |
7563 | new_rtx = make_extraction (mode, XEXP (inner, 0), |
7564 | pos: 0, pos_rtx: 0, len: len - INTVAL (XEXP (inner, 1)), |
7565 | unsignedp, in_dest, in_compare); |
7566 | if (new_rtx != 0) |
7567 | return gen_rtx_ASHIFT (mode, new_rtx, XEXP (inner, 1)); |
7568 | } |
7569 | else if (GET_CODE (inner) == MULT |
7570 | && CONST_INT_P (XEXP (inner, 1)) |
7571 | && pos_rtx == 0 && pos == 0) |
7572 | { |
7573 | /* We're extracting the least significant bits of an rtx |
7574 | (mult X (const_int 2^C)), where LEN > C. Extract the |
7575 | least significant (LEN - C) bits of X, giving an rtx |
7576 | whose mode is MODE, then multiply it by 2^C. */ |
7577 | const HOST_WIDE_INT shift_amt = exact_log2 (INTVAL (XEXP (inner, 1))); |
7578 | if (IN_RANGE (shift_amt, 1, len - 1)) |
7579 | { |
7580 | new_rtx = make_extraction (mode, XEXP (inner, 0), |
7581 | pos: 0, pos_rtx: 0, len: len - shift_amt, |
7582 | unsignedp, in_dest, in_compare); |
7583 | if (new_rtx) |
7584 | return gen_rtx_MULT (mode, new_rtx, XEXP (inner, 1)); |
7585 | } |
7586 | } |
7587 | else if (GET_CODE (inner) == TRUNCATE |
7588 | /* If trying or potentionally trying to extract |
7589 | bits outside of is_mode, don't look through |
7590 | TRUNCATE. See PR82192. */ |
7591 | && pos_rtx == NULL_RTX |
7592 | && known_le (pos + len, GET_MODE_PRECISION (is_mode))) |
7593 | inner = XEXP (inner, 0); |
7594 | |
7595 | inner_mode = GET_MODE (inner); |
7596 | |
7597 | /* See if this can be done without an extraction. We never can if the |
7598 | width of the field is not the same as that of some integer mode. For |
7599 | registers, we can only avoid the extraction if the position is at the |
7600 | low-order bit and this is either not in the destination or we have the |
7601 | appropriate STRICT_LOW_PART operation available. |
7602 | |
7603 | For MEM, we can avoid an extract if the field starts on an appropriate |
7604 | boundary and we can change the mode of the memory reference. */ |
7605 | |
7606 | scalar_int_mode tmode; |
7607 | if (int_mode_for_size (size: len, limit: 1).exists (mode: &tmode) |
7608 | && ((pos_rtx == 0 && (pos % BITS_PER_WORD) == 0 |
7609 | && !MEM_P (inner) |
7610 | && (pos == 0 || REG_P (inner)) |
7611 | && (inner_mode == tmode |
7612 | || !REG_P (inner) |
7613 | || TRULY_NOOP_TRUNCATION_MODES_P (tmode, inner_mode) |
7614 | || reg_truncated_to_mode (tmode, inner)) |
7615 | && (! in_dest |
7616 | || (REG_P (inner) |
7617 | && have_insn_for (STRICT_LOW_PART, tmode)))) |
7618 | || (MEM_P (inner) && pos_rtx == 0 |
7619 | && (pos |
7620 | % (STRICT_ALIGNMENT ? GET_MODE_ALIGNMENT (tmode) |
7621 | : BITS_PER_UNIT)) == 0 |
7622 | /* We can't do this if we are widening INNER_MODE (it |
7623 | may not be aligned, for one thing). */ |
7624 | && !paradoxical_subreg_p (outermode: tmode, innermode: inner_mode) |
7625 | && known_le (pos + len, GET_MODE_PRECISION (is_mode)) |
7626 | && (inner_mode == tmode |
7627 | || (! mode_dependent_address_p (XEXP (inner, 0), |
7628 | MEM_ADDR_SPACE (inner)) |
7629 | && ! MEM_VOLATILE_P (inner)))))) |
7630 | { |
7631 | /* If INNER is a MEM, make a new MEM that encompasses just the desired |
7632 | field. If the original and current mode are the same, we need not |
7633 | adjust the offset. Otherwise, we do if bytes big endian. |
7634 | |
7635 | If INNER is not a MEM, get a piece consisting of just the field |
7636 | of interest (in this case POS % BITS_PER_WORD must be 0). */ |
7637 | |
7638 | if (MEM_P (inner)) |
7639 | { |
7640 | poly_int64 offset; |
7641 | |
7642 | /* POS counts from lsb, but make OFFSET count in memory order. */ |
7643 | if (BYTES_BIG_ENDIAN) |
7644 | offset = bits_to_bytes_round_down (GET_MODE_PRECISION (is_mode) |
7645 | - len - pos); |
7646 | else |
7647 | offset = pos / BITS_PER_UNIT; |
7648 | |
7649 | new_rtx = adjust_address_nv (inner, tmode, offset); |
7650 | } |
7651 | else if (REG_P (inner)) |
7652 | { |
7653 | if (tmode != inner_mode) |
7654 | { |
7655 | /* We can't call gen_lowpart in a DEST since we |
7656 | always want a SUBREG (see below) and it would sometimes |
7657 | return a new hard register. */ |
7658 | if (pos || in_dest) |
7659 | { |
7660 | poly_uint64 offset |
7661 | = subreg_offset_from_lsb (outer_mode: tmode, inner_mode, lsb_shift: pos); |
7662 | |
7663 | /* Avoid creating invalid subregs, for example when |
7664 | simplifying (x>>32)&255. */ |
7665 | if (!validate_subreg (tmode, inner_mode, inner, offset)) |
7666 | return NULL_RTX; |
7667 | |
7668 | new_rtx = gen_rtx_SUBREG (tmode, inner, offset); |
7669 | } |
7670 | else |
7671 | new_rtx = gen_lowpart (tmode, inner); |
7672 | } |
7673 | else |
7674 | new_rtx = inner; |
7675 | } |
7676 | else |
7677 | new_rtx = force_to_mode (inner, tmode, |
7678 | len >= HOST_BITS_PER_WIDE_INT |
7679 | ? HOST_WIDE_INT_M1U |
7680 | : (HOST_WIDE_INT_1U << len) - 1, false); |
7681 | |
7682 | /* If this extraction is going into the destination of a SET, |
7683 | make a STRICT_LOW_PART unless we made a MEM. */ |
7684 | |
7685 | if (in_dest) |
7686 | return (MEM_P (new_rtx) ? new_rtx |
7687 | : (GET_CODE (new_rtx) != SUBREG |
7688 | ? gen_rtx_CLOBBER (tmode, const0_rtx) |
7689 | : gen_rtx_STRICT_LOW_PART (VOIDmode, new_rtx))); |
7690 | |
7691 | if (mode == tmode) |
7692 | return new_rtx; |
7693 | |
7694 | if (CONST_SCALAR_INT_P (new_rtx)) |
7695 | return simplify_unary_operation (code: unsignedp ? ZERO_EXTEND : SIGN_EXTEND, |
7696 | mode, op: new_rtx, op_mode: tmode); |
7697 | |
7698 | /* If we know that no extraneous bits are set, and that the high |
7699 | bit is not set, convert the extraction to the cheaper of |
7700 | sign and zero extension, that are equivalent in these cases. */ |
7701 | if (flag_expensive_optimizations |
7702 | && (HWI_COMPUTABLE_MODE_P (mode: tmode) |
7703 | && ((nonzero_bits (new_rtx, tmode) |
7704 | & ~(((unsigned HOST_WIDE_INT)GET_MODE_MASK (tmode)) >> 1)) |
7705 | == 0))) |
7706 | { |
7707 | rtx temp = gen_rtx_ZERO_EXTEND (mode, new_rtx); |
7708 | rtx temp1 = gen_rtx_SIGN_EXTEND (mode, new_rtx); |
7709 | |
7710 | /* Prefer ZERO_EXTENSION, since it gives more information to |
7711 | backends. */ |
7712 | if (set_src_cost (x: temp, mode, speed_p: optimize_this_for_speed_p) |
7713 | <= set_src_cost (x: temp1, mode, speed_p: optimize_this_for_speed_p)) |
7714 | return temp; |
7715 | return temp1; |
7716 | } |
7717 | |
7718 | /* Otherwise, sign- or zero-extend unless we already are in the |
7719 | proper mode. */ |
7720 | |
7721 | return (gen_rtx_fmt_e (unsignedp ? ZERO_EXTEND : SIGN_EXTEND, |
7722 | mode, new_rtx)); |
7723 | } |
7724 | |
7725 | /* Unless this is a COMPARE or we have a funny memory reference, |
7726 | don't do anything with zero-extending field extracts starting at |
7727 | the low-order bit since they are simple AND operations. */ |
7728 | if (pos_rtx == 0 && pos == 0 && ! in_dest |
7729 | && ! in_compare && unsignedp) |
7730 | return 0; |
7731 | |
7732 | /* Unless INNER is not MEM, reject this if we would be spanning bytes or |
7733 | if the position is not a constant and the length is not 1. In all |
7734 | other cases, we would only be going outside our object in cases when |
7735 | an original shift would have been undefined. */ |
7736 | if (MEM_P (inner) |
7737 | && ((pos_rtx == 0 && maybe_gt (pos + len, GET_MODE_PRECISION (is_mode))) |
7738 | || (pos_rtx != 0 && len != 1))) |
7739 | return 0; |
7740 | |
7741 | enum extraction_pattern pattern = (in_dest ? EP_insv |
7742 | : unsignedp ? EP_extzv : EP_extv); |
7743 | |
7744 | /* If INNER is not from memory, we want it to have the mode of a register |
7745 | extraction pattern's structure operand, or word_mode if there is no |
7746 | such pattern. The same applies to extraction_mode and pos_mode |
7747 | and their respective operands. |
7748 | |
7749 | For memory, assume that the desired extraction_mode and pos_mode |
7750 | are the same as for a register operation, since at present we don't |
7751 | have named patterns for aligned memory structures. */ |
7752 | class extraction_insn insn; |
7753 | unsigned int inner_size; |
7754 | if (GET_MODE_BITSIZE (mode: inner_mode).is_constant (const_value: &inner_size) |
7755 | && get_best_reg_extraction_insn (&insn, pattern, inner_size, mode)) |
7756 | { |
7757 | wanted_inner_reg_mode = insn.struct_mode.require (); |
7758 | pos_mode = insn.pos_mode; |
7759 | extraction_mode = insn.field_mode; |
7760 | } |
7761 | |
7762 | /* Never narrow an object, since that might not be safe. */ |
7763 | |
7764 | if (mode != VOIDmode |
7765 | && partial_subreg_p (outermode: extraction_mode, innermode: mode)) |
7766 | extraction_mode = mode; |
7767 | |
7768 | /* Punt if len is too large for extraction_mode. */ |
7769 | if (maybe_gt (len, GET_MODE_PRECISION (extraction_mode))) |
7770 | return NULL_RTX; |
7771 | |
7772 | if (!MEM_P (inner)) |
7773 | wanted_inner_mode = wanted_inner_reg_mode; |
7774 | else |
7775 | { |
7776 | /* Be careful not to go beyond the extracted object and maintain the |
7777 | natural alignment of the memory. */ |
7778 | wanted_inner_mode = smallest_int_mode_for_size (size: len); |
7779 | while (pos % GET_MODE_BITSIZE (mode: wanted_inner_mode) + len |
7780 | > GET_MODE_BITSIZE (mode: wanted_inner_mode)) |
7781 | wanted_inner_mode = GET_MODE_WIDER_MODE (m: wanted_inner_mode).require (); |
7782 | } |
7783 | |
7784 | orig_pos = pos; |
7785 | |
7786 | if (BITS_BIG_ENDIAN) |
7787 | { |
7788 | /* POS is passed as if BITS_BIG_ENDIAN == 0, so we need to convert it to |
7789 | BITS_BIG_ENDIAN style. If position is constant, compute new |
7790 | position. Otherwise, build subtraction. |
7791 | Note that POS is relative to the mode of the original argument. |
7792 | If it's a MEM we need to recompute POS relative to that. |
7793 | However, if we're extracting from (or inserting into) a register, |
7794 | we want to recompute POS relative to wanted_inner_mode. */ |
7795 | int width; |
7796 | if (!MEM_P (inner)) |
7797 | width = GET_MODE_BITSIZE (mode: wanted_inner_mode); |
7798 | else if (!GET_MODE_BITSIZE (mode: is_mode).is_constant (const_value: &width)) |
7799 | return NULL_RTX; |
7800 | |
7801 | if (pos_rtx == 0) |
7802 | pos = width - len - pos; |
7803 | else |
7804 | pos_rtx |
7805 | = gen_rtx_MINUS (GET_MODE (pos_rtx), |
7806 | gen_int_mode (width - len, GET_MODE (pos_rtx)), |
7807 | pos_rtx); |
7808 | /* POS may be less than 0 now, but we check for that below. |
7809 | Note that it can only be less than 0 if !MEM_P (inner). */ |
7810 | } |
7811 | |
7812 | /* If INNER has a wider mode, and this is a constant extraction, try to |
7813 | make it smaller and adjust the byte to point to the byte containing |
7814 | the value. */ |
7815 | if (wanted_inner_mode != VOIDmode |
7816 | && inner_mode != wanted_inner_mode |
7817 | && ! pos_rtx |
7818 | && partial_subreg_p (outermode: wanted_inner_mode, innermode: is_mode) |
7819 | && MEM_P (inner) |
7820 | && ! mode_dependent_address_p (XEXP (inner, 0), MEM_ADDR_SPACE (inner)) |
7821 | && ! MEM_VOLATILE_P (inner)) |
7822 | { |
7823 | poly_int64 offset = 0; |
7824 | |
7825 | /* The computations below will be correct if the machine is big |
7826 | endian in both bits and bytes or little endian in bits and bytes. |
7827 | If it is mixed, we must adjust. */ |
7828 | |
7829 | /* If bytes are big endian and we had a paradoxical SUBREG, we must |
7830 | adjust OFFSET to compensate. */ |
7831 | if (BYTES_BIG_ENDIAN |
7832 | && paradoxical_subreg_p (outermode: is_mode, innermode: inner_mode)) |
7833 | offset -= GET_MODE_SIZE (mode: is_mode) - GET_MODE_SIZE (mode: inner_mode); |
7834 | |
7835 | /* We can now move to the desired byte. */ |
7836 | offset += (pos / GET_MODE_BITSIZE (mode: wanted_inner_mode)) |
7837 | * GET_MODE_SIZE (mode: wanted_inner_mode); |
7838 | pos %= GET_MODE_BITSIZE (mode: wanted_inner_mode); |
7839 | |
7840 | if (BYTES_BIG_ENDIAN != BITS_BIG_ENDIAN |
7841 | && is_mode != wanted_inner_mode) |
7842 | offset = (GET_MODE_SIZE (mode: is_mode) |
7843 | - GET_MODE_SIZE (mode: wanted_inner_mode) - offset); |
7844 | |
7845 | inner = adjust_address_nv (inner, wanted_inner_mode, offset); |
7846 | } |
7847 | |
7848 | /* If INNER is not memory, get it into the proper mode. If we are changing |
7849 | its mode, POS must be a constant and smaller than the size of the new |
7850 | mode. */ |
7851 | else if (!MEM_P (inner)) |
7852 | { |
7853 | /* On the LHS, don't create paradoxical subregs implicitely truncating |
7854 | the register unless TARGET_TRULY_NOOP_TRUNCATION. */ |
7855 | if (in_dest |
7856 | && !TRULY_NOOP_TRUNCATION_MODES_P (GET_MODE (inner), |
7857 | wanted_inner_mode)) |
7858 | return NULL_RTX; |
7859 | |
7860 | if (GET_MODE (inner) != wanted_inner_mode |
7861 | && (pos_rtx != 0 |
7862 | || orig_pos + len > GET_MODE_BITSIZE (mode: wanted_inner_mode))) |
7863 | return NULL_RTX; |
7864 | |
7865 | if (orig_pos < 0) |
7866 | return NULL_RTX; |
7867 | |
7868 | inner = force_to_mode (inner, wanted_inner_mode, |
7869 | pos_rtx |
7870 | || len + orig_pos >= HOST_BITS_PER_WIDE_INT |
7871 | ? HOST_WIDE_INT_M1U |
7872 | : (((HOST_WIDE_INT_1U << len) - 1) |
7873 | << orig_pos), false); |
7874 | } |
7875 | |
7876 | /* Adjust mode of POS_RTX, if needed. If we want a wider mode, we |
7877 | have to zero extend. Otherwise, we can just use a SUBREG. |
7878 | |
7879 | We dealt with constant rtxes earlier, so pos_rtx cannot |
7880 | have VOIDmode at this point. */ |
7881 | if (pos_rtx != 0 |
7882 | && (GET_MODE_SIZE (mode: pos_mode) |
7883 | > GET_MODE_SIZE (mode: as_a <scalar_int_mode> (GET_MODE (pos_rtx))))) |
7884 | { |
7885 | rtx temp = simplify_gen_unary (code: ZERO_EXTEND, mode: pos_mode, op: pos_rtx, |
7886 | GET_MODE (pos_rtx)); |
7887 | |
7888 | /* If we know that no extraneous bits are set, and that the high |
7889 | bit is not set, convert extraction to cheaper one - either |
7890 | SIGN_EXTENSION or ZERO_EXTENSION, that are equivalent in these |
7891 | cases. */ |
7892 | if (flag_expensive_optimizations |
7893 | && (HWI_COMPUTABLE_MODE_P (GET_MODE (pos_rtx)) |
7894 | && ((nonzero_bits (pos_rtx, GET_MODE (pos_rtx)) |
7895 | & ~(((unsigned HOST_WIDE_INT) |
7896 | GET_MODE_MASK (GET_MODE (pos_rtx))) |
7897 | >> 1)) |
7898 | == 0))) |
7899 | { |
7900 | rtx temp1 = simplify_gen_unary (code: SIGN_EXTEND, mode: pos_mode, op: pos_rtx, |
7901 | GET_MODE (pos_rtx)); |
7902 | |
7903 | /* Prefer ZERO_EXTENSION, since it gives more information to |
7904 | backends. */ |
7905 | if (set_src_cost (x: temp1, mode: pos_mode, speed_p: optimize_this_for_speed_p) |
7906 | < set_src_cost (x: temp, mode: pos_mode, speed_p: optimize_this_for_speed_p)) |
7907 | temp = temp1; |
7908 | } |
7909 | pos_rtx = temp; |
7910 | } |
7911 | |
7912 | /* Make POS_RTX unless we already have it and it is correct. If we don't |
7913 | have a POS_RTX but we do have an ORIG_POS_RTX, the latter must |
7914 | be a CONST_INT. */ |
7915 | if (pos_rtx == 0 && orig_pos_rtx != 0 && INTVAL (orig_pos_rtx) == pos) |
7916 | pos_rtx = orig_pos_rtx; |
7917 | |
7918 | else if (pos_rtx == 0) |
7919 | pos_rtx = GEN_INT (pos); |
7920 | |
7921 | /* Make the required operation. See if we can use existing rtx. */ |
7922 | new_rtx = gen_rtx_fmt_eee (unsignedp ? ZERO_EXTRACT : SIGN_EXTRACT, |
7923 | extraction_mode, inner, GEN_INT (len), pos_rtx); |
7924 | if (! in_dest) |
7925 | new_rtx = gen_lowpart (mode, new_rtx); |
7926 | |
7927 | return new_rtx; |
7928 | } |
7929 | |
7930 | /* See if X (of mode MODE) contains an ASHIFT of COUNT or more bits that |
7931 | can be commuted with any other operations in X. Return X without |
7932 | that shift if so. */ |
7933 | |
7934 | static rtx |
7935 | (scalar_int_mode mode, rtx x, int count) |
7936 | { |
7937 | enum rtx_code code = GET_CODE (x); |
7938 | rtx tem; |
7939 | |
7940 | switch (code) |
7941 | { |
7942 | case ASHIFT: |
7943 | /* This is the shift itself. If it is wide enough, we will return |
7944 | either the value being shifted if the shift count is equal to |
7945 | COUNT or a shift for the difference. */ |
7946 | if (CONST_INT_P (XEXP (x, 1)) |
7947 | && INTVAL (XEXP (x, 1)) >= count) |
7948 | return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (x, 0), |
7949 | INTVAL (XEXP (x, 1)) - count); |
7950 | break; |
7951 | |
7952 | case NEG: case NOT: |
7953 | if ((tem = extract_left_shift (mode, XEXP (x, 0), count)) != 0) |
7954 | return simplify_gen_unary (code, mode, op: tem, op_mode: mode); |
7955 | |
7956 | break; |
7957 | |
7958 | case PLUS: case IOR: case XOR: case AND: |
7959 | /* If we can safely shift this constant and we find the inner shift, |
7960 | make a new operation. */ |
7961 | if (CONST_INT_P (XEXP (x, 1)) |
7962 | && (UINTVAL (XEXP (x, 1)) |
7963 | & (((HOST_WIDE_INT_1U << count)) - 1)) == 0 |
7964 | && (tem = extract_left_shift (mode, XEXP (x, 0), count)) != 0) |
7965 | { |
7966 | HOST_WIDE_INT val = INTVAL (XEXP (x, 1)) >> count; |
7967 | return simplify_gen_binary (code, mode, op0: tem, |
7968 | op1: gen_int_mode (val, mode)); |
7969 | } |
7970 | break; |
7971 | |
7972 | default: |
7973 | break; |
7974 | } |
7975 | |
7976 | return 0; |
7977 | } |
7978 | |
7979 | /* Subroutine of make_compound_operation. *X_PTR is the rtx at the current |
7980 | level of the expression and MODE is its mode. IN_CODE is as for |
7981 | make_compound_operation. *NEXT_CODE_PTR is the value of IN_CODE |
7982 | that should be used when recursing on operands of *X_PTR. |
7983 | |
7984 | There are two possible actions: |
7985 | |
7986 | - Return null. This tells the caller to recurse on *X_PTR with IN_CODE |
7987 | equal to *NEXT_CODE_PTR, after which *X_PTR holds the final value. |
7988 | |
7989 | - Return a new rtx, which the caller returns directly. */ |
7990 | |
7991 | static rtx |
7992 | make_compound_operation_int (scalar_int_mode mode, rtx *x_ptr, |
7993 | enum rtx_code in_code, |
7994 | enum rtx_code *next_code_ptr) |
7995 | { |
7996 | rtx x = *x_ptr; |
7997 | enum rtx_code next_code = *next_code_ptr; |
7998 | enum rtx_code code = GET_CODE (x); |
7999 | int mode_width = GET_MODE_PRECISION (mode); |
8000 | rtx rhs, lhs; |
8001 | rtx new_rtx = 0; |
8002 | int i; |
8003 | rtx tem; |
8004 | scalar_int_mode inner_mode; |
8005 | bool equality_comparison = false; |
8006 | |
8007 | if (in_code == EQ) |
8008 | { |
8009 | equality_comparison = true; |
8010 | in_code = COMPARE; |
8011 | } |
8012 | |
8013 | /* Process depending on the code of this operation. If NEW is set |
8014 | nonzero, it will be returned. */ |
8015 | |
8016 | switch (code) |
8017 | { |
8018 | case ASHIFT: |
8019 | /* Convert shifts by constants into multiplications if inside |
8020 | an address. */ |
8021 | if (in_code == MEM && CONST_INT_P (XEXP (x, 1)) |
8022 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT |
8023 | && INTVAL (XEXP (x, 1)) >= 0) |
8024 | { |
8025 | HOST_WIDE_INT count = INTVAL (XEXP (x, 1)); |
8026 | HOST_WIDE_INT multval = HOST_WIDE_INT_1 << count; |
8027 | |
8028 | new_rtx = make_compound_operation (XEXP (x, 0), next_code); |
8029 | if (GET_CODE (new_rtx) == NEG) |
8030 | { |
8031 | new_rtx = XEXP (new_rtx, 0); |
8032 | multval = -multval; |
8033 | } |
8034 | multval = trunc_int_for_mode (multval, mode); |
8035 | new_rtx = gen_rtx_MULT (mode, new_rtx, gen_int_mode (multval, mode)); |
8036 | } |
8037 | break; |
8038 | |
8039 | case PLUS: |
8040 | lhs = XEXP (x, 0); |
8041 | rhs = XEXP (x, 1); |
8042 | lhs = make_compound_operation (lhs, next_code); |
8043 | rhs = make_compound_operation (rhs, next_code); |
8044 | if (GET_CODE (lhs) == MULT && GET_CODE (XEXP (lhs, 0)) == NEG) |
8045 | { |
8046 | tem = simplify_gen_binary (code: MULT, mode, XEXP (XEXP (lhs, 0), 0), |
8047 | XEXP (lhs, 1)); |
8048 | new_rtx = simplify_gen_binary (code: MINUS, mode, op0: rhs, op1: tem); |
8049 | } |
8050 | else if (GET_CODE (lhs) == MULT |
8051 | && (CONST_INT_P (XEXP (lhs, 1)) && INTVAL (XEXP (lhs, 1)) < 0)) |
8052 | { |
8053 | tem = simplify_gen_binary (code: MULT, mode, XEXP (lhs, 0), |
8054 | op1: simplify_gen_unary (code: NEG, mode, |
8055 | XEXP (lhs, 1), |
8056 | op_mode: mode)); |
8057 | new_rtx = simplify_gen_binary (code: MINUS, mode, op0: rhs, op1: tem); |
8058 | } |
8059 | else |
8060 | { |
8061 | SUBST (XEXP (x, 0), lhs); |
8062 | SUBST (XEXP (x, 1), rhs); |
8063 | } |
8064 | maybe_swap_commutative_operands (x); |
8065 | return x; |
8066 | |
8067 | case MINUS: |
8068 | lhs = XEXP (x, 0); |
8069 | rhs = XEXP (x, 1); |
8070 | lhs = make_compound_operation (lhs, next_code); |
8071 | rhs = make_compound_operation (rhs, next_code); |
8072 | if (GET_CODE (rhs) == MULT && GET_CODE (XEXP (rhs, 0)) == NEG) |
8073 | { |
8074 | tem = simplify_gen_binary (code: MULT, mode, XEXP (XEXP (rhs, 0), 0), |
8075 | XEXP (rhs, 1)); |
8076 | return simplify_gen_binary (code: PLUS, mode, op0: tem, op1: lhs); |
8077 | } |
8078 | else if (GET_CODE (rhs) == MULT |
8079 | && (CONST_INT_P (XEXP (rhs, 1)) && INTVAL (XEXP (rhs, 1)) < 0)) |
8080 | { |
8081 | tem = simplify_gen_binary (code: MULT, mode, XEXP (rhs, 0), |
8082 | op1: simplify_gen_unary (code: NEG, mode, |
8083 | XEXP (rhs, 1), |
8084 | op_mode: mode)); |
8085 | return simplify_gen_binary (code: PLUS, mode, op0: tem, op1: lhs); |
8086 | } |
8087 | else |
8088 | { |
8089 | SUBST (XEXP (x, 0), lhs); |
8090 | SUBST (XEXP (x, 1), rhs); |
8091 | return x; |
8092 | } |
8093 | |
8094 | case AND: |
8095 | /* If the second operand is not a constant, we can't do anything |
8096 | with it. */ |
8097 | if (!CONST_INT_P (XEXP (x, 1))) |
8098 | break; |
8099 | |
8100 | /* If the constant is a power of two minus one and the first operand |
8101 | is a logical right shift, make an extraction. */ |
8102 | if (GET_CODE (XEXP (x, 0)) == LSHIFTRT |
8103 | && (i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0) |
8104 | { |
8105 | new_rtx = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code); |
8106 | new_rtx = make_extraction (mode, inner: new_rtx, pos: 0, XEXP (XEXP (x, 0), 1), |
8107 | len: i, unsignedp: true, in_dest: false, in_compare: in_code == COMPARE); |
8108 | } |
8109 | |
8110 | /* Same as previous, but for (subreg (lshiftrt ...)) in first op. */ |
8111 | else if (GET_CODE (XEXP (x, 0)) == SUBREG |
8112 | && subreg_lowpart_p (XEXP (x, 0)) |
8113 | && is_a <scalar_int_mode> (GET_MODE (SUBREG_REG (XEXP (x, 0))), |
8114 | result: &inner_mode) |
8115 | && GET_CODE (SUBREG_REG (XEXP (x, 0))) == LSHIFTRT |
8116 | && (i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0) |
8117 | { |
8118 | rtx inner_x0 = SUBREG_REG (XEXP (x, 0)); |
8119 | new_rtx = make_compound_operation (XEXP (inner_x0, 0), next_code); |
8120 | new_rtx = make_extraction (mode: inner_mode, inner: new_rtx, pos: 0, |
8121 | XEXP (inner_x0, 1), |
8122 | len: i, unsignedp: true, in_dest: false, in_compare: in_code == COMPARE); |
8123 | |
8124 | /* If we narrowed the mode when dropping the subreg, then we lose. */ |
8125 | if (GET_MODE_SIZE (mode: inner_mode) < GET_MODE_SIZE (mode)) |
8126 | new_rtx = NULL; |
8127 | |
8128 | /* If that didn't give anything, see if the AND simplifies on |
8129 | its own. */ |
8130 | if (!new_rtx && i >= 0) |
8131 | { |
8132 | new_rtx = make_compound_operation (XEXP (x, 0), next_code); |
8133 | new_rtx = make_extraction (mode, inner: new_rtx, pos: 0, NULL_RTX, len: i, |
8134 | unsignedp: true, in_dest: false, in_compare: in_code == COMPARE); |
8135 | } |
8136 | } |
8137 | /* Same as previous, but for (xor/ior (lshiftrt...) (lshiftrt...)). */ |
8138 | else if ((GET_CODE (XEXP (x, 0)) == XOR |
8139 | || GET_CODE (XEXP (x, 0)) == IOR) |
8140 | && GET_CODE (XEXP (XEXP (x, 0), 0)) == LSHIFTRT |
8141 | && GET_CODE (XEXP (XEXP (x, 0), 1)) == LSHIFTRT |
8142 | && (i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0) |
8143 | { |
8144 | /* Apply the distributive law, and then try to make extractions. */ |
8145 | new_rtx = gen_rtx_fmt_ee (GET_CODE (XEXP (x, 0)), mode, |
8146 | gen_rtx_AND (mode, XEXP (XEXP (x, 0), 0), |
8147 | XEXP (x, 1)), |
8148 | gen_rtx_AND (mode, XEXP (XEXP (x, 0), 1), |
8149 | XEXP (x, 1))); |
8150 | new_rtx = make_compound_operation (new_rtx, in_code); |
8151 | } |
8152 | |
8153 | /* If we are have (and (rotate X C) M) and C is larger than the number |
8154 | of bits in M, this is an extraction. */ |
8155 | |
8156 | else if (GET_CODE (XEXP (x, 0)) == ROTATE |
8157 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) |
8158 | && (i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0 |
8159 | && i <= INTVAL (XEXP (XEXP (x, 0), 1))) |
8160 | { |
8161 | new_rtx = make_compound_operation (XEXP (XEXP (x, 0), 0), next_code); |
8162 | new_rtx = make_extraction (mode, inner: new_rtx, |
8163 | pos: (GET_MODE_PRECISION (mode) |
8164 | - INTVAL (XEXP (XEXP (x, 0), 1))), |
8165 | NULL_RTX, len: i, unsignedp: true, in_dest: false, |
8166 | in_compare: in_code == COMPARE); |
8167 | } |
8168 | |
8169 | /* On machines without logical shifts, if the operand of the AND is |
8170 | a logical shift and our mask turns off all the propagated sign |
8171 | bits, we can replace the logical shift with an arithmetic shift. */ |
8172 | else if (GET_CODE (XEXP (x, 0)) == LSHIFTRT |
8173 | && !have_insn_for (LSHIFTRT, mode) |
8174 | && have_insn_for (ASHIFTRT, mode) |
8175 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) |
8176 | && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0 |
8177 | && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT |
8178 | && mode_width <= HOST_BITS_PER_WIDE_INT) |
8179 | { |
8180 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); |
8181 | |
8182 | mask >>= INTVAL (XEXP (XEXP (x, 0), 1)); |
8183 | if ((INTVAL (XEXP (x, 1)) & ~mask) == 0) |
8184 | SUBST (XEXP (x, 0), |
8185 | gen_rtx_ASHIFTRT (mode, |
8186 | make_compound_operation (XEXP (XEXP (x, |
8187 | 0), |
8188 | 0), |
8189 | next_code), |
8190 | XEXP (XEXP (x, 0), 1))); |
8191 | } |
8192 | |
8193 | /* If the constant is one less than a power of two, this might be |
8194 | representable by an extraction even if no shift is present. |
8195 | If it doesn't end up being a ZERO_EXTEND, we will ignore it unless |
8196 | we are in a COMPARE. */ |
8197 | else if ((i = exact_log2 (UINTVAL (XEXP (x, 1)) + 1)) >= 0) |
8198 | new_rtx = make_extraction (mode, |
8199 | inner: make_compound_operation (XEXP (x, 0), |
8200 | next_code), |
8201 | pos: 0, NULL_RTX, len: i, |
8202 | unsignedp: true, in_dest: false, in_compare: in_code == COMPARE); |
8203 | |
8204 | /* If we are in a comparison and this is an AND with a power of two, |
8205 | convert this into the appropriate bit extract. */ |
8206 | else if (in_code == COMPARE |
8207 | && (i = exact_log2 (UINTVAL (XEXP (x, 1)))) >= 0 |
8208 | && (equality_comparison || i < GET_MODE_PRECISION (mode) - 1)) |
8209 | new_rtx = make_extraction (mode, |
8210 | inner: make_compound_operation (XEXP (x, 0), |
8211 | next_code), |
8212 | pos: i, NULL_RTX, len: 1, unsignedp: true, in_dest: false, in_compare: true); |
8213 | |
8214 | /* If the one operand is a paradoxical subreg of a register or memory and |
8215 | the constant (limited to the smaller mode) has only zero bits where |
8216 | the sub expression has known zero bits, this can be expressed as |
8217 | a zero_extend. */ |
8218 | else if (GET_CODE (XEXP (x, 0)) == SUBREG) |
8219 | { |
8220 | rtx sub; |
8221 | |
8222 | sub = XEXP (XEXP (x, 0), 0); |
8223 | machine_mode sub_mode = GET_MODE (sub); |
8224 | int sub_width; |
8225 | if ((REG_P (sub) || MEM_P (sub)) |
8226 | && GET_MODE_PRECISION (mode: sub_mode).is_constant (const_value: &sub_width) |
8227 | && sub_width < mode_width) |
8228 | { |
8229 | unsigned HOST_WIDE_INT mode_mask = GET_MODE_MASK (sub_mode); |
8230 | unsigned HOST_WIDE_INT mask; |
8231 | |
8232 | /* original AND constant with all the known zero bits set */ |
8233 | mask = UINTVAL (XEXP (x, 1)) | (~nonzero_bits (sub, sub_mode)); |
8234 | if ((mask & mode_mask) == mode_mask) |
8235 | { |
8236 | new_rtx = make_compound_operation (sub, next_code); |
8237 | new_rtx = make_extraction (mode, inner: new_rtx, pos: 0, pos_rtx: 0, len: sub_width, |
8238 | unsignedp: true, in_dest: false, in_compare: in_code == COMPARE); |
8239 | } |
8240 | } |
8241 | } |
8242 | |
8243 | break; |
8244 | |
8245 | case LSHIFTRT: |
8246 | /* If the sign bit is known to be zero, replace this with an |
8247 | arithmetic shift. */ |
8248 | if (have_insn_for (ASHIFTRT, mode) |
8249 | && ! have_insn_for (LSHIFTRT, mode) |
8250 | && mode_width <= HOST_BITS_PER_WIDE_INT |
8251 | && (nonzero_bits (XEXP (x, 0), mode) & (1 << (mode_width - 1))) == 0) |
8252 | { |
8253 | new_rtx = gen_rtx_ASHIFTRT (mode, |
8254 | make_compound_operation (XEXP (x, 0), |
8255 | next_code), |
8256 | XEXP (x, 1)); |
8257 | break; |
8258 | } |
8259 | |
8260 | /* fall through */ |
8261 | |
8262 | case ASHIFTRT: |
8263 | lhs = XEXP (x, 0); |
8264 | rhs = XEXP (x, 1); |
8265 | |
8266 | /* If we have (ashiftrt (ashift foo C1) C2) with C2 >= C1, |
8267 | this is a SIGN_EXTRACT. */ |
8268 | if (CONST_INT_P (rhs) |
8269 | && GET_CODE (lhs) == ASHIFT |
8270 | && CONST_INT_P (XEXP (lhs, 1)) |
8271 | && INTVAL (rhs) >= INTVAL (XEXP (lhs, 1)) |
8272 | && INTVAL (XEXP (lhs, 1)) >= 0 |
8273 | && INTVAL (rhs) < mode_width) |
8274 | { |
8275 | new_rtx = make_compound_operation (XEXP (lhs, 0), next_code); |
8276 | new_rtx = make_extraction (mode, inner: new_rtx, |
8277 | INTVAL (rhs) - INTVAL (XEXP (lhs, 1)), |
8278 | NULL_RTX, len: mode_width - INTVAL (rhs), |
8279 | unsignedp: code == LSHIFTRT, in_dest: false, |
8280 | in_compare: in_code == COMPARE); |
8281 | break; |
8282 | } |
8283 | |
8284 | /* See if we have operations between an ASHIFTRT and an ASHIFT. |
8285 | If so, try to merge the shifts into a SIGN_EXTEND. We could |
8286 | also do this for some cases of SIGN_EXTRACT, but it doesn't |
8287 | seem worth the effort; the case checked for occurs on Alpha. */ |
8288 | |
8289 | if (!OBJECT_P (lhs) |
8290 | && ! (GET_CODE (lhs) == SUBREG |
8291 | && (OBJECT_P (SUBREG_REG (lhs)))) |
8292 | && CONST_INT_P (rhs) |
8293 | && INTVAL (rhs) >= 0 |
8294 | && INTVAL (rhs) < HOST_BITS_PER_WIDE_INT |
8295 | && INTVAL (rhs) < mode_width |
8296 | && (new_rtx = extract_left_shift (mode, x: lhs, INTVAL (rhs))) != 0) |
8297 | new_rtx = make_extraction (mode, inner: make_compound_operation (new_rtx, |
8298 | next_code), |
8299 | pos: 0, NULL_RTX, len: mode_width - INTVAL (rhs), |
8300 | unsignedp: code == LSHIFTRT, in_dest: false, in_compare: in_code == COMPARE); |
8301 | |
8302 | break; |
8303 | |
8304 | case SUBREG: |
8305 | /* Call ourselves recursively on the inner expression. If we are |
8306 | narrowing the object and it has a different RTL code from |
8307 | what it originally did, do this SUBREG as a force_to_mode. */ |
8308 | { |
8309 | rtx inner = SUBREG_REG (x), simplified; |
8310 | enum rtx_code subreg_code = in_code; |
8311 | |
8312 | /* If the SUBREG is masking of a logical right shift, |
8313 | make an extraction. */ |
8314 | if (GET_CODE (inner) == LSHIFTRT |
8315 | && is_a <scalar_int_mode> (GET_MODE (inner), result: &inner_mode) |
8316 | && GET_MODE_SIZE (mode) < GET_MODE_SIZE (mode: inner_mode) |
8317 | && CONST_INT_P (XEXP (inner, 1)) |
8318 | && UINTVAL (XEXP (inner, 1)) < GET_MODE_PRECISION (mode: inner_mode) |
8319 | && subreg_lowpart_p (x)) |
8320 | { |
8321 | new_rtx = make_compound_operation (XEXP (inner, 0), next_code); |
8322 | int width = GET_MODE_PRECISION (mode: inner_mode) |
8323 | - INTVAL (XEXP (inner, 1)); |
8324 | if (width > mode_width) |
8325 | width = mode_width; |
8326 | new_rtx = make_extraction (mode, inner: new_rtx, pos: 0, XEXP (inner, 1), |
8327 | len: width, unsignedp: true, in_dest: false, in_compare: in_code == COMPARE); |
8328 | break; |
8329 | } |
8330 | |
8331 | /* If in_code is COMPARE, it isn't always safe to pass it through |
8332 | to the recursive make_compound_operation call. */ |
8333 | if (subreg_code == COMPARE |
8334 | && (!subreg_lowpart_p (x) |
8335 | || GET_CODE (inner) == SUBREG |
8336 | /* (subreg:SI (and:DI (reg:DI) (const_int 0x800000000)) 0) |
8337 | is (const_int 0), rather than |
8338 | (subreg:SI (lshiftrt:DI (reg:DI) (const_int 35)) 0). |
8339 | Similarly (subreg:QI (and:SI (reg:SI) (const_int 0x80)) 0) |
8340 | for non-equality comparisons against 0 is not equivalent |
8341 | to (subreg:QI (lshiftrt:SI (reg:SI) (const_int 7)) 0). */ |
8342 | || (GET_CODE (inner) == AND |
8343 | && CONST_INT_P (XEXP (inner, 1)) |
8344 | && partial_subreg_p (x) |
8345 | && exact_log2 (UINTVAL (XEXP (inner, 1))) |
8346 | >= GET_MODE_BITSIZE (mode) - 1))) |
8347 | subreg_code = SET; |
8348 | |
8349 | tem = make_compound_operation (inner, subreg_code); |
8350 | |
8351 | simplified |
8352 | = simplify_subreg (outermode: mode, op: tem, GET_MODE (inner), SUBREG_BYTE (x)); |
8353 | if (simplified) |
8354 | tem = simplified; |
8355 | |
8356 | if (GET_CODE (tem) != GET_CODE (inner) |
8357 | && partial_subreg_p (x) |
8358 | && subreg_lowpart_p (x)) |
8359 | { |
8360 | rtx newer |
8361 | = force_to_mode (tem, mode, HOST_WIDE_INT_M1U, false); |
8362 | |
8363 | /* If we have something other than a SUBREG, we might have |
8364 | done an expansion, so rerun ourselves. */ |
8365 | if (GET_CODE (newer) != SUBREG) |
8366 | newer = make_compound_operation (newer, in_code); |
8367 | |
8368 | /* force_to_mode can expand compounds. If it just re-expanded |
8369 | the compound, use gen_lowpart to convert to the desired |
8370 | mode. */ |
8371 | if (rtx_equal_p (newer, x) |
8372 | /* Likewise if it re-expanded the compound only partially. |
8373 | This happens for SUBREG of ZERO_EXTRACT if they extract |
8374 | the same number of bits. */ |
8375 | || (GET_CODE (newer) == SUBREG |
8376 | && (GET_CODE (SUBREG_REG (newer)) == LSHIFTRT |
8377 | || GET_CODE (SUBREG_REG (newer)) == ASHIFTRT) |
8378 | && GET_CODE (inner) == AND |
8379 | && rtx_equal_p (SUBREG_REG (newer), XEXP (inner, 0)))) |
8380 | return gen_lowpart (GET_MODE (x), tem); |
8381 | |
8382 | return newer; |
8383 | } |
8384 | |
8385 | if (simplified) |
8386 | return tem; |
8387 | } |
8388 | break; |
8389 | |
8390 | default: |
8391 | break; |
8392 | } |
8393 | |
8394 | if (new_rtx) |
8395 | *x_ptr = gen_lowpart (mode, new_rtx); |
8396 | *next_code_ptr = next_code; |
8397 | return NULL_RTX; |
8398 | } |
8399 | |
8400 | /* Look at the expression rooted at X. Look for expressions |
8401 | equivalent to ZERO_EXTRACT, SIGN_EXTRACT, ZERO_EXTEND, SIGN_EXTEND. |
8402 | Form these expressions. |
8403 | |
8404 | Return the new rtx, usually just X. |
8405 | |
8406 | Also, for machines like the VAX that don't have logical shift insns, |
8407 | try to convert logical to arithmetic shift operations in cases where |
8408 | they are equivalent. This undoes the canonicalizations to logical |
8409 | shifts done elsewhere. |
8410 | |
8411 | We try, as much as possible, to re-use rtl expressions to save memory. |
8412 | |
8413 | IN_CODE says what kind of expression we are processing. Normally, it is |
8414 | SET. In a memory address it is MEM. When processing the arguments of |
8415 | a comparison or a COMPARE against zero, it is COMPARE, or EQ if more |
8416 | precisely it is an equality comparison against zero. */ |
8417 | |
8418 | rtx |
8419 | make_compound_operation (rtx x, enum rtx_code in_code) |
8420 | { |
8421 | enum rtx_code code = GET_CODE (x); |
8422 | const char *fmt; |
8423 | int i, j; |
8424 | enum rtx_code next_code; |
8425 | rtx new_rtx, tem; |
8426 | |
8427 | /* Select the code to be used in recursive calls. Once we are inside an |
8428 | address, we stay there. If we have a comparison, set to COMPARE, |
8429 | but once inside, go back to our default of SET. */ |
8430 | |
8431 | next_code = (code == MEM ? MEM |
8432 | : ((code == COMPARE || COMPARISON_P (x)) |
8433 | && XEXP (x, 1) == const0_rtx) ? COMPARE |
8434 | : in_code == COMPARE || in_code == EQ ? SET : in_code); |
8435 | |
8436 | scalar_int_mode mode; |
8437 | if (is_a <scalar_int_mode> (GET_MODE (x), result: &mode)) |
8438 | { |
8439 | rtx new_rtx = make_compound_operation_int (mode, x_ptr: &x, in_code, |
8440 | next_code_ptr: &next_code); |
8441 | if (new_rtx) |
8442 | return new_rtx; |
8443 | code = GET_CODE (x); |
8444 | } |
8445 | |
8446 | /* Now recursively process each operand of this operation. We need to |
8447 | handle ZERO_EXTEND specially so that we don't lose track of the |
8448 | inner mode. */ |
8449 | if (code == ZERO_EXTEND) |
8450 | { |
8451 | new_rtx = make_compound_operation (XEXP (x, 0), in_code: next_code); |
8452 | tem = simplify_unary_operation (code: ZERO_EXTEND, GET_MODE (x), |
8453 | op: new_rtx, GET_MODE (XEXP (x, 0))); |
8454 | if (tem) |
8455 | return tem; |
8456 | SUBST (XEXP (x, 0), new_rtx); |
8457 | return x; |
8458 | } |
8459 | |
8460 | fmt = GET_RTX_FORMAT (code); |
8461 | for (i = 0; i < GET_RTX_LENGTH (code); i++) |
8462 | if (fmt[i] == 'e') |
8463 | { |
8464 | new_rtx = make_compound_operation (XEXP (x, i), in_code: next_code); |
8465 | SUBST (XEXP (x, i), new_rtx); |
8466 | } |
8467 | else if (fmt[i] == 'E') |
8468 | for (j = 0; j < XVECLEN (x, i); j++) |
8469 | { |
8470 | new_rtx = make_compound_operation (XVECEXP (x, i, j), in_code: next_code); |
8471 | SUBST (XVECEXP (x, i, j), new_rtx); |
8472 | } |
8473 | |
8474 | maybe_swap_commutative_operands (x); |
8475 | return x; |
8476 | } |
8477 | |
8478 | /* Given M see if it is a value that would select a field of bits |
8479 | within an item, but not the entire word. Return -1 if not. |
8480 | Otherwise, return the starting position of the field, where 0 is the |
8481 | low-order bit. |
8482 | |
8483 | *PLEN is set to the length of the field. */ |
8484 | |
8485 | static int |
8486 | get_pos_from_mask (unsigned HOST_WIDE_INT m, unsigned HOST_WIDE_INT *plen) |
8487 | { |
8488 | /* Get the bit number of the first 1 bit from the right, -1 if none. */ |
8489 | int pos = m ? ctz_hwi (x: m) : -1; |
8490 | int len = 0; |
8491 | |
8492 | if (pos >= 0) |
8493 | /* Now shift off the low-order zero bits and see if we have a |
8494 | power of two minus 1. */ |
8495 | len = exact_log2 (x: (m >> pos) + 1); |
8496 | |
8497 | if (len <= 0) |
8498 | pos = -1; |
8499 | |
8500 | *plen = len; |
8501 | return pos; |
8502 | } |
8503 | |
8504 | /* If X refers to a register that equals REG in value, replace these |
8505 | references with REG. */ |
8506 | static rtx |
8507 | canon_reg_for_combine (rtx x, rtx reg) |
8508 | { |
8509 | rtx op0, op1, op2; |
8510 | const char *fmt; |
8511 | int i; |
8512 | bool copied; |
8513 | |
8514 | enum rtx_code code = GET_CODE (x); |
8515 | switch (GET_RTX_CLASS (code)) |
8516 | { |
8517 | case RTX_UNARY: |
8518 | op0 = canon_reg_for_combine (XEXP (x, 0), reg); |
8519 | if (op0 != XEXP (x, 0)) |
8520 | return simplify_gen_unary (GET_CODE (x), GET_MODE (x), op: op0, |
8521 | GET_MODE (reg)); |
8522 | break; |
8523 | |
8524 | case RTX_BIN_ARITH: |
8525 | case RTX_COMM_ARITH: |
8526 | op0 = canon_reg_for_combine (XEXP (x, 0), reg); |
8527 | op1 = canon_reg_for_combine (XEXP (x, 1), reg); |
8528 | if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1)) |
8529 | return simplify_gen_binary (GET_CODE (x), GET_MODE (x), op0, op1); |
8530 | break; |
8531 | |
8532 | case RTX_COMPARE: |
8533 | case RTX_COMM_COMPARE: |
8534 | op0 = canon_reg_for_combine (XEXP (x, 0), reg); |
8535 | op1 = canon_reg_for_combine (XEXP (x, 1), reg); |
8536 | if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1)) |
8537 | return simplify_gen_relational (GET_CODE (x), GET_MODE (x), |
8538 | GET_MODE (op0), op0, op1); |
8539 | break; |
8540 | |
8541 | case RTX_TERNARY: |
8542 | case RTX_BITFIELD_OPS: |
8543 | op0 = canon_reg_for_combine (XEXP (x, 0), reg); |
8544 | op1 = canon_reg_for_combine (XEXP (x, 1), reg); |
8545 | op2 = canon_reg_for_combine (XEXP (x, 2), reg); |
8546 | if (op0 != XEXP (x, 0) || op1 != XEXP (x, 1) || op2 != XEXP (x, 2)) |
8547 | return simplify_gen_ternary (GET_CODE (x), GET_MODE (x), |
8548 | GET_MODE (op0), op0, op1, op2); |
8549 | /* FALLTHRU */ |
8550 | |
8551 | case RTX_OBJ: |
8552 | if (REG_P (x)) |
8553 | { |
8554 | if (rtx_equal_p (get_last_value (reg), x) |
8555 | || rtx_equal_p (reg, get_last_value (x))) |
8556 | return reg; |
8557 | else |
8558 | break; |
8559 | } |
8560 | |
8561 | /* fall through */ |
8562 | |
8563 | default: |
8564 | fmt = GET_RTX_FORMAT (code); |
8565 | copied = false; |
8566 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
8567 | if (fmt[i] == 'e') |
8568 | { |
8569 | rtx op = canon_reg_for_combine (XEXP (x, i), reg); |
8570 | if (op != XEXP (x, i)) |
8571 | { |
8572 | if (!copied) |
8573 | { |
8574 | copied = true; |
8575 | x = copy_rtx (x); |
8576 | } |
8577 | XEXP (x, i) = op; |
8578 | } |
8579 | } |
8580 | else if (fmt[i] == 'E') |
8581 | { |
8582 | int j; |
8583 | for (j = 0; j < XVECLEN (x, i); j++) |
8584 | { |
8585 | rtx op = canon_reg_for_combine (XVECEXP (x, i, j), reg); |
8586 | if (op != XVECEXP (x, i, j)) |
8587 | { |
8588 | if (!copied) |
8589 | { |
8590 | copied = true; |
8591 | x = copy_rtx (x); |
8592 | } |
8593 | XVECEXP (x, i, j) = op; |
8594 | } |
8595 | } |
8596 | } |
8597 | |
8598 | break; |
8599 | } |
8600 | |
8601 | return x; |
8602 | } |
8603 | |
8604 | /* Return X converted to MODE. If the value is already truncated to |
8605 | MODE we can just return a subreg even though in the general case we |
8606 | would need an explicit truncation. */ |
8607 | |
8608 | static rtx |
8609 | gen_lowpart_or_truncate (machine_mode mode, rtx x) |
8610 | { |
8611 | if (!CONST_INT_P (x) |
8612 | && partial_subreg_p (outermode: mode, GET_MODE (x)) |
8613 | && !TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (x)) |
8614 | && !(REG_P (x) && reg_truncated_to_mode (mode, x))) |
8615 | { |
8616 | /* Bit-cast X into an integer mode. */ |
8617 | if (!SCALAR_INT_MODE_P (GET_MODE (x))) |
8618 | x = gen_lowpart (int_mode_for_mode (GET_MODE (x)).require (), x); |
8619 | x = simplify_gen_unary (code: TRUNCATE, mode: int_mode_for_mode (mode).require (), |
8620 | op: x, GET_MODE (x)); |
8621 | } |
8622 | |
8623 | return gen_lowpart (mode, x); |
8624 | } |
8625 | |
8626 | /* See if X can be simplified knowing that we will only refer to it in |
8627 | MODE and will only refer to those bits that are nonzero in MASK. |
8628 | If other bits are being computed or if masking operations are done |
8629 | that select a superset of the bits in MASK, they can sometimes be |
8630 | ignored. |
8631 | |
8632 | Return a possibly simplified expression, but always convert X to |
8633 | MODE. If X is a CONST_INT, AND the CONST_INT with MASK. |
8634 | |
8635 | If JUST_SELECT is true, don't optimize by noticing that bits in MASK |
8636 | are all off in X. This is used when X will be complemented, by either |
8637 | NOT, NEG, or XOR. */ |
8638 | |
8639 | static rtx |
8640 | force_to_mode (rtx x, machine_mode mode, unsigned HOST_WIDE_INT mask, |
8641 | bool just_select) |
8642 | { |
8643 | enum rtx_code code = GET_CODE (x); |
8644 | bool next_select = just_select || code == XOR || code == NOT || code == NEG; |
8645 | machine_mode op_mode; |
8646 | unsigned HOST_WIDE_INT nonzero; |
8647 | |
8648 | /* If this is a CALL or ASM_OPERANDS, don't do anything. Some of the |
8649 | code below will do the wrong thing since the mode of such an |
8650 | expression is VOIDmode. |
8651 | |
8652 | Also do nothing if X is a CLOBBER; this can happen if X was |
8653 | the return value from a call to gen_lowpart. */ |
8654 | if (code == CALL || code == ASM_OPERANDS || code == CLOBBER) |
8655 | return x; |
8656 | |
8657 | /* We want to perform the operation in its present mode unless we know |
8658 | that the operation is valid in MODE, in which case we do the operation |
8659 | in MODE. */ |
8660 | op_mode = ((GET_MODE_CLASS (mode) == GET_MODE_CLASS (GET_MODE (x)) |
8661 | && have_insn_for (code, mode)) |
8662 | ? mode : GET_MODE (x)); |
8663 | |
8664 | /* It is not valid to do a right-shift in a narrower mode |
8665 | than the one it came in with. */ |
8666 | if ((code == LSHIFTRT || code == ASHIFTRT) |
8667 | && partial_subreg_p (outermode: mode, GET_MODE (x))) |
8668 | op_mode = GET_MODE (x); |
8669 | |
8670 | /* Truncate MASK to fit OP_MODE. */ |
8671 | if (op_mode) |
8672 | mask &= GET_MODE_MASK (op_mode); |
8673 | |
8674 | /* Determine what bits of X are guaranteed to be (non)zero. */ |
8675 | nonzero = nonzero_bits (x, mode); |
8676 | |
8677 | /* If none of the bits in X are needed, return a zero. */ |
8678 | if (!just_select && (nonzero & mask) == 0 && !side_effects_p (x)) |
8679 | x = const0_rtx; |
8680 | |
8681 | /* If X is a CONST_INT, return a new one. Do this here since the |
8682 | test below will fail. */ |
8683 | if (CONST_INT_P (x)) |
8684 | { |
8685 | if (SCALAR_INT_MODE_P (mode)) |
8686 | return gen_int_mode (INTVAL (x) & mask, mode); |
8687 | else |
8688 | { |
8689 | x = GEN_INT (INTVAL (x) & mask); |
8690 | return gen_lowpart_common (mode, x); |
8691 | } |
8692 | } |
8693 | |
8694 | /* If X is narrower than MODE and we want all the bits in X's mode, just |
8695 | get X in the proper mode. */ |
8696 | if (paradoxical_subreg_p (outermode: mode, GET_MODE (x)) |
8697 | && (GET_MODE_MASK (GET_MODE (x)) & ~mask) == 0) |
8698 | return gen_lowpart (mode, x); |
8699 | |
8700 | /* We can ignore the effect of a SUBREG if it narrows the mode or |
8701 | if the constant masks to zero all the bits the mode doesn't have. */ |
8702 | if (GET_CODE (x) == SUBREG |
8703 | && subreg_lowpart_p (x) |
8704 | && (partial_subreg_p (x) |
8705 | || (mask |
8706 | & GET_MODE_MASK (GET_MODE (x)) |
8707 | & ~GET_MODE_MASK (GET_MODE (SUBREG_REG (x)))) == 0)) |
8708 | return force_to_mode (SUBREG_REG (x), mode, mask, just_select: next_select); |
8709 | |
8710 | scalar_int_mode int_mode, xmode; |
8711 | if (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
8712 | && is_a <scalar_int_mode> (GET_MODE (x), result: &xmode)) |
8713 | /* OP_MODE is either MODE or XMODE, so it must be a scalar |
8714 | integer too. */ |
8715 | return force_int_to_mode (x, int_mode, xmode, |
8716 | as_a <scalar_int_mode> (m: op_mode), |
8717 | mask, just_select); |
8718 | |
8719 | return gen_lowpart_or_truncate (mode, x); |
8720 | } |
8721 | |
8722 | /* Subroutine of force_to_mode that handles cases in which both X and |
8723 | the result are scalar integers. MODE is the mode of the result, |
8724 | XMODE is the mode of X, and OP_MODE says which of MODE or XMODE |
8725 | is preferred for simplified versions of X. The other arguments |
8726 | are as for force_to_mode. */ |
8727 | |
8728 | static rtx |
8729 | force_int_to_mode (rtx x, scalar_int_mode mode, scalar_int_mode xmode, |
8730 | scalar_int_mode op_mode, unsigned HOST_WIDE_INT mask, |
8731 | bool just_select) |
8732 | { |
8733 | enum rtx_code code = GET_CODE (x); |
8734 | bool next_select = just_select || code == XOR || code == NOT || code == NEG; |
8735 | unsigned HOST_WIDE_INT fuller_mask; |
8736 | rtx op0, op1, temp; |
8737 | poly_int64 const_op0; |
8738 | |
8739 | /* When we have an arithmetic operation, or a shift whose count we |
8740 | do not know, we need to assume that all bits up to the highest-order |
8741 | bit in MASK will be needed. This is how we form such a mask. */ |
8742 | if (mask & (HOST_WIDE_INT_1U << (HOST_BITS_PER_WIDE_INT - 1))) |
8743 | fuller_mask = HOST_WIDE_INT_M1U; |
8744 | else |
8745 | fuller_mask = ((HOST_WIDE_INT_1U << (floor_log2 (x: mask) + 1)) - 1); |
8746 | |
8747 | switch (code) |
8748 | { |
8749 | case CLOBBER: |
8750 | /* If X is a (clobber (const_int)), return it since we know we are |
8751 | generating something that won't match. */ |
8752 | return x; |
8753 | |
8754 | case SIGN_EXTEND: |
8755 | case ZERO_EXTEND: |
8756 | case ZERO_EXTRACT: |
8757 | case SIGN_EXTRACT: |
8758 | x = expand_compound_operation (x); |
8759 | if (GET_CODE (x) != code) |
8760 | return force_to_mode (x, mode, mask, just_select: next_select); |
8761 | break; |
8762 | |
8763 | case TRUNCATE: |
8764 | /* Similarly for a truncate. */ |
8765 | return force_to_mode (XEXP (x, 0), mode, mask, just_select: next_select); |
8766 | |
8767 | case AND: |
8768 | /* If this is an AND with a constant, convert it into an AND |
8769 | whose constant is the AND of that constant with MASK. If it |
8770 | remains an AND of MASK, delete it since it is redundant. */ |
8771 | |
8772 | if (CONST_INT_P (XEXP (x, 1))) |
8773 | { |
8774 | x = simplify_and_const_int (x, op_mode, XEXP (x, 0), |
8775 | mask & INTVAL (XEXP (x, 1))); |
8776 | xmode = op_mode; |
8777 | |
8778 | /* If X is still an AND, see if it is an AND with a mask that |
8779 | is just some low-order bits. If so, and it is MASK, we don't |
8780 | need it. */ |
8781 | |
8782 | if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1)) |
8783 | && (INTVAL (XEXP (x, 1)) & GET_MODE_MASK (xmode)) == mask) |
8784 | x = XEXP (x, 0); |
8785 | |
8786 | /* If it remains an AND, try making another AND with the bits |
8787 | in the mode mask that aren't in MASK turned on. If the |
8788 | constant in the AND is wide enough, this might make a |
8789 | cheaper constant. */ |
8790 | |
8791 | if (GET_CODE (x) == AND && CONST_INT_P (XEXP (x, 1)) |
8792 | && GET_MODE_MASK (xmode) != mask |
8793 | && HWI_COMPUTABLE_MODE_P (mode: xmode)) |
8794 | { |
8795 | unsigned HOST_WIDE_INT cval |
8796 | = UINTVAL (XEXP (x, 1)) | (GET_MODE_MASK (xmode) & ~mask); |
8797 | rtx y; |
8798 | |
8799 | y = simplify_gen_binary (code: AND, mode: xmode, XEXP (x, 0), |
8800 | op1: gen_int_mode (cval, xmode)); |
8801 | if (set_src_cost (x: y, mode: xmode, speed_p: optimize_this_for_speed_p) |
8802 | < set_src_cost (x, mode: xmode, speed_p: optimize_this_for_speed_p)) |
8803 | x = y; |
8804 | } |
8805 | |
8806 | break; |
8807 | } |
8808 | |
8809 | goto binop; |
8810 | |
8811 | case PLUS: |
8812 | /* In (and (plus FOO C1) M), if M is a mask that just turns off |
8813 | low-order bits (as in an alignment operation) and FOO is already |
8814 | aligned to that boundary, mask C1 to that boundary as well. |
8815 | This may eliminate that PLUS and, later, the AND. */ |
8816 | |
8817 | { |
8818 | unsigned int width = GET_MODE_PRECISION (mode); |
8819 | unsigned HOST_WIDE_INT smask = mask; |
8820 | |
8821 | /* If MODE is narrower than HOST_WIDE_INT and mask is a negative |
8822 | number, sign extend it. */ |
8823 | |
8824 | if (width < HOST_BITS_PER_WIDE_INT |
8825 | && (smask & (HOST_WIDE_INT_1U << (width - 1))) != 0) |
8826 | smask |= HOST_WIDE_INT_M1U << width; |
8827 | |
8828 | if (CONST_INT_P (XEXP (x, 1)) |
8829 | && pow2p_hwi (x: - smask) |
8830 | && (nonzero_bits (XEXP (x, 0), mode) & ~smask) == 0 |
8831 | && (INTVAL (XEXP (x, 1)) & ~smask) != 0) |
8832 | return force_to_mode (x: plus_constant (xmode, XEXP (x, 0), |
8833 | (INTVAL (XEXP (x, 1)) & smask)), |
8834 | mode, mask: smask, just_select: next_select); |
8835 | } |
8836 | |
8837 | /* fall through */ |
8838 | |
8839 | case MULT: |
8840 | /* Substituting into the operands of a widening MULT is not likely to |
8841 | create RTL matching a machine insn. */ |
8842 | if (code == MULT |
8843 | && (GET_CODE (XEXP (x, 0)) == ZERO_EXTEND |
8844 | || GET_CODE (XEXP (x, 0)) == SIGN_EXTEND) |
8845 | && (GET_CODE (XEXP (x, 1)) == ZERO_EXTEND |
8846 | || GET_CODE (XEXP (x, 1)) == SIGN_EXTEND) |
8847 | && REG_P (XEXP (XEXP (x, 0), 0)) |
8848 | && REG_P (XEXP (XEXP (x, 1), 0))) |
8849 | return gen_lowpart_or_truncate (mode, x); |
8850 | |
8851 | /* For PLUS, MINUS and MULT, we need any bits less significant than the |
8852 | most significant bit in MASK since carries from those bits will |
8853 | affect the bits we are interested in. */ |
8854 | mask = fuller_mask; |
8855 | goto binop; |
8856 | |
8857 | case MINUS: |
8858 | /* If X is (minus C Y) where C's least set bit is larger than any bit |
8859 | in the mask, then we may replace with (neg Y). */ |
8860 | if (poly_int_rtx_p (XEXP (x, 0), res: &const_op0) |
8861 | && known_alignment (a: poly_uint64 (const_op0)) > mask) |
8862 | { |
8863 | x = simplify_gen_unary (code: NEG, mode: xmode, XEXP (x, 1), op_mode: xmode); |
8864 | return force_to_mode (x, mode, mask, just_select: next_select); |
8865 | } |
8866 | |
8867 | /* Similarly, if C contains every bit in the fuller_mask, then we may |
8868 | replace with (not Y). */ |
8869 | if (CONST_INT_P (XEXP (x, 0)) |
8870 | && ((UINTVAL (XEXP (x, 0)) | fuller_mask) == UINTVAL (XEXP (x, 0)))) |
8871 | { |
8872 | x = simplify_gen_unary (code: NOT, mode: xmode, XEXP (x, 1), op_mode: xmode); |
8873 | return force_to_mode (x, mode, mask, just_select: next_select); |
8874 | } |
8875 | |
8876 | mask = fuller_mask; |
8877 | goto binop; |
8878 | |
8879 | case IOR: |
8880 | case XOR: |
8881 | /* If X is (ior (lshiftrt FOO C1) C2), try to commute the IOR and |
8882 | LSHIFTRT so we end up with an (and (lshiftrt (ior ...) ...) ...) |
8883 | operation which may be a bitfield extraction. Ensure that the |
8884 | constant we form is not wider than the mode of X. */ |
8885 | |
8886 | if (GET_CODE (XEXP (x, 0)) == LSHIFTRT |
8887 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) |
8888 | && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0 |
8889 | && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT |
8890 | && CONST_INT_P (XEXP (x, 1)) |
8891 | && ((INTVAL (XEXP (XEXP (x, 0), 1)) |
8892 | + floor_log2 (INTVAL (XEXP (x, 1)))) |
8893 | < GET_MODE_PRECISION (mode: xmode)) |
8894 | && (UINTVAL (XEXP (x, 1)) |
8895 | & ~nonzero_bits (XEXP (x, 0), xmode)) == 0) |
8896 | { |
8897 | temp = gen_int_mode ((INTVAL (XEXP (x, 1)) & mask) |
8898 | << INTVAL (XEXP (XEXP (x, 0), 1)), |
8899 | xmode); |
8900 | temp = simplify_gen_binary (GET_CODE (x), mode: xmode, |
8901 | XEXP (XEXP (x, 0), 0), op1: temp); |
8902 | x = simplify_gen_binary (code: LSHIFTRT, mode: xmode, op0: temp, |
8903 | XEXP (XEXP (x, 0), 1)); |
8904 | return force_to_mode (x, mode, mask, just_select: next_select); |
8905 | } |
8906 | |
8907 | binop: |
8908 | /* For most binary operations, just propagate into the operation and |
8909 | change the mode if we have an operation of that mode. */ |
8910 | |
8911 | op0 = force_to_mode (XEXP (x, 0), mode, mask, just_select: next_select); |
8912 | op1 = force_to_mode (XEXP (x, 1), mode, mask, just_select: next_select); |
8913 | |
8914 | /* If we ended up truncating both operands, truncate the result of the |
8915 | operation instead. */ |
8916 | if (GET_CODE (op0) == TRUNCATE |
8917 | && GET_CODE (op1) == TRUNCATE) |
8918 | { |
8919 | op0 = XEXP (op0, 0); |
8920 | op1 = XEXP (op1, 0); |
8921 | } |
8922 | |
8923 | op0 = gen_lowpart_or_truncate (mode: op_mode, x: op0); |
8924 | op1 = gen_lowpart_or_truncate (mode: op_mode, x: op1); |
8925 | |
8926 | if (op_mode != xmode || op0 != XEXP (x, 0) || op1 != XEXP (x, 1)) |
8927 | { |
8928 | x = simplify_gen_binary (code, mode: op_mode, op0, op1); |
8929 | xmode = op_mode; |
8930 | } |
8931 | break; |
8932 | |
8933 | case ASHIFT: |
8934 | /* For left shifts, do the same, but just for the first operand. |
8935 | However, we cannot do anything with shifts where we cannot |
8936 | guarantee that the counts are smaller than the size of the mode |
8937 | because such a count will have a different meaning in a |
8938 | wider mode. */ |
8939 | |
8940 | if (! (CONST_INT_P (XEXP (x, 1)) |
8941 | && INTVAL (XEXP (x, 1)) >= 0 |
8942 | && INTVAL (XEXP (x, 1)) < GET_MODE_PRECISION (mode)) |
8943 | && ! (GET_MODE (XEXP (x, 1)) != VOIDmode |
8944 | && (nonzero_bits (XEXP (x, 1), GET_MODE (XEXP (x, 1))) |
8945 | < (unsigned HOST_WIDE_INT) GET_MODE_PRECISION (mode)))) |
8946 | break; |
8947 | |
8948 | /* If the shift count is a constant and we can do arithmetic in |
8949 | the mode of the shift, refine which bits we need. Otherwise, use the |
8950 | conservative form of the mask. */ |
8951 | if (CONST_INT_P (XEXP (x, 1)) |
8952 | && INTVAL (XEXP (x, 1)) >= 0 |
8953 | && INTVAL (XEXP (x, 1)) < GET_MODE_PRECISION (mode: op_mode) |
8954 | && HWI_COMPUTABLE_MODE_P (mode: op_mode)) |
8955 | mask >>= INTVAL (XEXP (x, 1)); |
8956 | else |
8957 | mask = fuller_mask; |
8958 | |
8959 | op0 = gen_lowpart_or_truncate (mode: op_mode, |
8960 | x: force_to_mode (XEXP (x, 0), mode, |
8961 | mask, just_select: next_select)); |
8962 | |
8963 | if (op_mode != xmode || op0 != XEXP (x, 0)) |
8964 | { |
8965 | x = simplify_gen_binary (code, mode: op_mode, op0, XEXP (x, 1)); |
8966 | xmode = op_mode; |
8967 | } |
8968 | break; |
8969 | |
8970 | case LSHIFTRT: |
8971 | /* Here we can only do something if the shift count is a constant, |
8972 | this shift constant is valid for the host, and we can do arithmetic |
8973 | in OP_MODE. */ |
8974 | |
8975 | if (CONST_INT_P (XEXP (x, 1)) |
8976 | && INTVAL (XEXP (x, 1)) >= 0 |
8977 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT |
8978 | && HWI_COMPUTABLE_MODE_P (mode: op_mode)) |
8979 | { |
8980 | rtx inner = XEXP (x, 0); |
8981 | unsigned HOST_WIDE_INT inner_mask; |
8982 | |
8983 | /* Select the mask of the bits we need for the shift operand. */ |
8984 | inner_mask = mask << INTVAL (XEXP (x, 1)); |
8985 | |
8986 | /* We can only change the mode of the shift if we can do arithmetic |
8987 | in the mode of the shift and INNER_MASK is no wider than the |
8988 | width of X's mode. */ |
8989 | if ((inner_mask & ~GET_MODE_MASK (xmode)) != 0) |
8990 | op_mode = xmode; |
8991 | |
8992 | inner = force_to_mode (x: inner, mode: op_mode, mask: inner_mask, just_select: next_select); |
8993 | |
8994 | if (xmode != op_mode || inner != XEXP (x, 0)) |
8995 | { |
8996 | x = simplify_gen_binary (code: LSHIFTRT, mode: op_mode, op0: inner, XEXP (x, 1)); |
8997 | xmode = op_mode; |
8998 | } |
8999 | } |
9000 | |
9001 | /* If we have (and (lshiftrt FOO C1) C2) where the combination of the |
9002 | shift and AND produces only copies of the sign bit (C2 is one less |
9003 | than a power of two), we can do this with just a shift. */ |
9004 | |
9005 | if (GET_CODE (x) == LSHIFTRT |
9006 | && CONST_INT_P (XEXP (x, 1)) |
9007 | /* The shift puts one of the sign bit copies in the least significant |
9008 | bit. */ |
9009 | && ((INTVAL (XEXP (x, 1)) |
9010 | + num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0)))) |
9011 | >= GET_MODE_PRECISION (mode: xmode)) |
9012 | && pow2p_hwi (x: mask + 1) |
9013 | /* Number of bits left after the shift must be more than the mask |
9014 | needs. */ |
9015 | && ((INTVAL (XEXP (x, 1)) + exact_log2 (x: mask + 1)) |
9016 | <= GET_MODE_PRECISION (mode: xmode)) |
9017 | /* Must be more sign bit copies than the mask needs. */ |
9018 | && ((int) num_sign_bit_copies (XEXP (x, 0), GET_MODE (XEXP (x, 0))) |
9019 | >= exact_log2 (x: mask + 1))) |
9020 | { |
9021 | int nbits = GET_MODE_PRECISION (mode: xmode) - exact_log2 (x: mask + 1); |
9022 | x = simplify_gen_binary (code: LSHIFTRT, mode: xmode, XEXP (x, 0), |
9023 | op1: gen_int_shift_amount (xmode, nbits)); |
9024 | } |
9025 | goto shiftrt; |
9026 | |
9027 | case ASHIFTRT: |
9028 | /* If we are just looking for the sign bit, we don't need this shift at |
9029 | all, even if it has a variable count. */ |
9030 | if (val_signbit_p (xmode, mask)) |
9031 | return force_to_mode (XEXP (x, 0), mode, mask, just_select: next_select); |
9032 | |
9033 | /* If this is a shift by a constant, get a mask that contains those bits |
9034 | that are not copies of the sign bit. We then have two cases: If |
9035 | MASK only includes those bits, this can be a logical shift, which may |
9036 | allow simplifications. If MASK is a single-bit field not within |
9037 | those bits, we are requesting a copy of the sign bit and hence can |
9038 | shift the sign bit to the appropriate location. */ |
9039 | |
9040 | if (CONST_INT_P (XEXP (x, 1)) && INTVAL (XEXP (x, 1)) >= 0 |
9041 | && INTVAL (XEXP (x, 1)) < HOST_BITS_PER_WIDE_INT) |
9042 | { |
9043 | unsigned HOST_WIDE_INT nonzero; |
9044 | int i; |
9045 | |
9046 | /* If the considered data is wider than HOST_WIDE_INT, we can't |
9047 | represent a mask for all its bits in a single scalar. |
9048 | But we only care about the lower bits, so calculate these. */ |
9049 | |
9050 | if (GET_MODE_PRECISION (mode: xmode) > HOST_BITS_PER_WIDE_INT) |
9051 | { |
9052 | nonzero = HOST_WIDE_INT_M1U; |
9053 | |
9054 | /* GET_MODE_PRECISION (GET_MODE (x)) - INTVAL (XEXP (x, 1)) |
9055 | is the number of bits a full-width mask would have set. |
9056 | We need only shift if these are fewer than nonzero can |
9057 | hold. If not, we must keep all bits set in nonzero. */ |
9058 | |
9059 | if (GET_MODE_PRECISION (mode: xmode) - INTVAL (XEXP (x, 1)) |
9060 | < HOST_BITS_PER_WIDE_INT) |
9061 | nonzero >>= INTVAL (XEXP (x, 1)) |
9062 | + HOST_BITS_PER_WIDE_INT |
9063 | - GET_MODE_PRECISION (mode: xmode); |
9064 | } |
9065 | else |
9066 | { |
9067 | nonzero = GET_MODE_MASK (xmode); |
9068 | nonzero >>= INTVAL (XEXP (x, 1)); |
9069 | } |
9070 | |
9071 | if ((mask & ~nonzero) == 0) |
9072 | { |
9073 | x = simplify_shift_const (NULL_RTX, LSHIFTRT, xmode, |
9074 | XEXP (x, 0), INTVAL (XEXP (x, 1))); |
9075 | if (GET_CODE (x) != ASHIFTRT) |
9076 | return force_to_mode (x, mode, mask, just_select: next_select); |
9077 | } |
9078 | |
9079 | else if ((i = exact_log2 (x: mask)) >= 0) |
9080 | { |
9081 | x = simplify_shift_const |
9082 | (NULL_RTX, LSHIFTRT, xmode, XEXP (x, 0), |
9083 | GET_MODE_PRECISION (mode: xmode) - 1 - i); |
9084 | |
9085 | if (GET_CODE (x) != ASHIFTRT) |
9086 | return force_to_mode (x, mode, mask, just_select: next_select); |
9087 | } |
9088 | } |
9089 | |
9090 | /* If MASK is 1, convert this to an LSHIFTRT. This can be done |
9091 | even if the shift count isn't a constant. */ |
9092 | if (mask == 1) |
9093 | x = simplify_gen_binary (code: LSHIFTRT, mode: xmode, XEXP (x, 0), XEXP (x, 1)); |
9094 | |
9095 | shiftrt: |
9096 | |
9097 | /* If this is a zero- or sign-extension operation that just affects bits |
9098 | we don't care about, remove it. Be sure the call above returned |
9099 | something that is still a shift. */ |
9100 | |
9101 | if ((GET_CODE (x) == LSHIFTRT || GET_CODE (x) == ASHIFTRT) |
9102 | && CONST_INT_P (XEXP (x, 1)) |
9103 | && INTVAL (XEXP (x, 1)) >= 0 |
9104 | && (INTVAL (XEXP (x, 1)) |
9105 | <= GET_MODE_PRECISION (mode: xmode) - (floor_log2 (x: mask) + 1)) |
9106 | && GET_CODE (XEXP (x, 0)) == ASHIFT |
9107 | && XEXP (XEXP (x, 0), 1) == XEXP (x, 1)) |
9108 | return force_to_mode (XEXP (XEXP (x, 0), 0), mode, mask, just_select: next_select); |
9109 | |
9110 | break; |
9111 | |
9112 | case ROTATE: |
9113 | case ROTATERT: |
9114 | /* If the shift count is constant and we can do computations |
9115 | in the mode of X, compute where the bits we care about are. |
9116 | Otherwise, we can't do anything. Don't change the mode of |
9117 | the shift or propagate MODE into the shift, though. */ |
9118 | if (CONST_INT_P (XEXP (x, 1)) |
9119 | && INTVAL (XEXP (x, 1)) >= 0) |
9120 | { |
9121 | temp = simplify_binary_operation (code: code == ROTATE ? ROTATERT : ROTATE, |
9122 | mode: xmode, op0: gen_int_mode (mask, xmode), |
9123 | XEXP (x, 1)); |
9124 | if (temp && CONST_INT_P (temp)) |
9125 | x = simplify_gen_binary (code, mode: xmode, |
9126 | op0: force_to_mode (XEXP (x, 0), mode: xmode, |
9127 | INTVAL (temp), just_select: next_select), |
9128 | XEXP (x, 1)); |
9129 | } |
9130 | break; |
9131 | |
9132 | case NEG: |
9133 | /* If we just want the low-order bit, the NEG isn't needed since it |
9134 | won't change the low-order bit. */ |
9135 | if (mask == 1) |
9136 | return force_to_mode (XEXP (x, 0), mode, mask, just_select); |
9137 | |
9138 | /* We need any bits less significant than the most significant bit in |
9139 | MASK since carries from those bits will affect the bits we are |
9140 | interested in. */ |
9141 | mask = fuller_mask; |
9142 | goto unop; |
9143 | |
9144 | case NOT: |
9145 | /* (not FOO) is (xor FOO CONST), so if FOO is an LSHIFTRT, we can do the |
9146 | same as the XOR case above. Ensure that the constant we form is not |
9147 | wider than the mode of X. */ |
9148 | |
9149 | if (GET_CODE (XEXP (x, 0)) == LSHIFTRT |
9150 | && CONST_INT_P (XEXP (XEXP (x, 0), 1)) |
9151 | && INTVAL (XEXP (XEXP (x, 0), 1)) >= 0 |
9152 | && (INTVAL (XEXP (XEXP (x, 0), 1)) + floor_log2 (x: mask) |
9153 | < GET_MODE_PRECISION (mode: xmode)) |
9154 | && INTVAL (XEXP (XEXP (x, 0), 1)) < HOST_BITS_PER_WIDE_INT) |
9155 | { |
9156 | temp = gen_int_mode (mask << INTVAL (XEXP (XEXP (x, 0), 1)), xmode); |
9157 | temp = simplify_gen_binary (code: XOR, mode: xmode, XEXP (XEXP (x, 0), 0), op1: temp); |
9158 | x = simplify_gen_binary (code: LSHIFTRT, mode: xmode, |
9159 | op0: temp, XEXP (XEXP (x, 0), 1)); |
9160 | |
9161 | return force_to_mode (x, mode, mask, just_select: next_select); |
9162 | } |
9163 | |
9164 | /* (and (not FOO) CONST) is (not (or FOO (not CONST))), so we must |
9165 | use the full mask inside the NOT. */ |
9166 | mask = fuller_mask; |
9167 | |
9168 | unop: |
9169 | op0 = gen_lowpart_or_truncate (mode: op_mode, |
9170 | x: force_to_mode (XEXP (x, 0), mode, mask, |
9171 | just_select: next_select)); |
9172 | if (op_mode != xmode || op0 != XEXP (x, 0)) |
9173 | { |
9174 | x = simplify_gen_unary (code, mode: op_mode, op: op0, op_mode); |
9175 | xmode = op_mode; |
9176 | } |
9177 | break; |
9178 | |
9179 | case NE: |
9180 | /* (and (ne FOO 0) CONST) can be (and FOO CONST) if CONST is included |
9181 | in STORE_FLAG_VALUE and FOO has a single bit that might be nonzero, |
9182 | which is equal to STORE_FLAG_VALUE. */ |
9183 | if ((mask & ~STORE_FLAG_VALUE) == 0 |
9184 | && XEXP (x, 1) == const0_rtx |
9185 | && GET_MODE (XEXP (x, 0)) == mode |
9186 | && pow2p_hwi (x: nonzero_bits (XEXP (x, 0), mode)) |
9187 | && (nonzero_bits (XEXP (x, 0), mode) |
9188 | == (unsigned HOST_WIDE_INT) STORE_FLAG_VALUE)) |
9189 | return force_to_mode (XEXP (x, 0), mode, mask, just_select: next_select); |
9190 | |
9191 | break; |
9192 | |
9193 | case IF_THEN_ELSE: |
9194 | /* We have no way of knowing if the IF_THEN_ELSE can itself be |
9195 | written in a narrower mode. We play it safe and do not do so. */ |
9196 | |
9197 | op0 = gen_lowpart_or_truncate (mode: xmode, |
9198 | x: force_to_mode (XEXP (x, 1), mode, |
9199 | mask, just_select: next_select)); |
9200 | op1 = gen_lowpart_or_truncate (mode: xmode, |
9201 | x: force_to_mode (XEXP (x, 2), mode, |
9202 | mask, just_select: next_select)); |
9203 | if (op0 != XEXP (x, 1) || op1 != XEXP (x, 2)) |
9204 | x = simplify_gen_ternary (code: IF_THEN_ELSE, mode: xmode, |
9205 | GET_MODE (XEXP (x, 0)), XEXP (x, 0), |
9206 | op1: op0, op2: op1); |
9207 | break; |
9208 | |
9209 | default: |
9210 | break; |
9211 | } |
9212 | |
9213 | /* Ensure we return a value of the proper mode. */ |
9214 | return gen_lowpart_or_truncate (mode, x); |
9215 | } |
9216 | |
9217 | /* Return nonzero if X is an expression that has one of two values depending on |
9218 | whether some other value is zero or nonzero. In that case, we return the |
9219 | value that is being tested, *PTRUE is set to the value if the rtx being |
9220 | returned has a nonzero value, and *PFALSE is set to the other alternative. |
9221 | |
9222 | If we return zero, we set *PTRUE and *PFALSE to X. */ |
9223 | |
9224 | static rtx |
9225 | if_then_else_cond (rtx x, rtx *ptrue, rtx *pfalse) |
9226 | { |
9227 | machine_mode mode = GET_MODE (x); |
9228 | enum rtx_code code = GET_CODE (x); |
9229 | rtx cond0, cond1, true0, true1, false0, false1; |
9230 | unsigned HOST_WIDE_INT nz; |
9231 | scalar_int_mode int_mode; |
9232 | |
9233 | /* If we are comparing a value against zero, we are done. */ |
9234 | if ((code == NE || code == EQ) |
9235 | && XEXP (x, 1) == const0_rtx) |
9236 | { |
9237 | *ptrue = (code == NE) ? const_true_rtx : const0_rtx; |
9238 | *pfalse = (code == NE) ? const0_rtx : const_true_rtx; |
9239 | return XEXP (x, 0); |
9240 | } |
9241 | |
9242 | /* If this is a unary operation whose operand has one of two values, apply |
9243 | our opcode to compute those values. */ |
9244 | else if (UNARY_P (x) |
9245 | && (cond0 = if_then_else_cond (XEXP (x, 0), ptrue: &true0, pfalse: &false0)) != 0) |
9246 | { |
9247 | *ptrue = simplify_gen_unary (code, mode, op: true0, GET_MODE (XEXP (x, 0))); |
9248 | *pfalse = simplify_gen_unary (code, mode, op: false0, |
9249 | GET_MODE (XEXP (x, 0))); |
9250 | return cond0; |
9251 | } |
9252 | |
9253 | /* If this is a COMPARE, do nothing, since the IF_THEN_ELSE we would |
9254 | make can't possibly match and would suppress other optimizations. */ |
9255 | else if (code == COMPARE) |
9256 | ; |
9257 | |
9258 | /* If this is a binary operation, see if either side has only one of two |
9259 | values. If either one does or if both do and they are conditional on |
9260 | the same value, compute the new true and false values. */ |
9261 | else if (BINARY_P (x)) |
9262 | { |
9263 | rtx op0 = XEXP (x, 0); |
9264 | rtx op1 = XEXP (x, 1); |
9265 | cond0 = if_then_else_cond (x: op0, ptrue: &true0, pfalse: &false0); |
9266 | cond1 = if_then_else_cond (x: op1, ptrue: &true1, pfalse: &false1); |
9267 | |
9268 | if ((cond0 != 0 && cond1 != 0 && !rtx_equal_p (cond0, cond1)) |
9269 | && (REG_P (op0) || REG_P (op1))) |
9270 | { |
9271 | /* Try to enable a simplification by undoing work done by |
9272 | if_then_else_cond if it converted a REG into something more |
9273 | complex. */ |
9274 | if (REG_P (op0)) |
9275 | { |
9276 | cond0 = 0; |
9277 | true0 = false0 = op0; |
9278 | } |
9279 | else |
9280 | { |
9281 | cond1 = 0; |
9282 | true1 = false1 = op1; |
9283 | } |
9284 | } |
9285 | |
9286 | if ((cond0 != 0 || cond1 != 0) |
9287 | && ! (cond0 != 0 && cond1 != 0 && !rtx_equal_p (cond0, cond1))) |
9288 | { |
9289 | /* If if_then_else_cond returned zero, then true/false are the |
9290 | same rtl. We must copy one of them to prevent invalid rtl |
9291 | sharing. */ |
9292 | if (cond0 == 0) |
9293 | true0 = copy_rtx (true0); |
9294 | else if (cond1 == 0) |
9295 | true1 = copy_rtx (true1); |
9296 | |
9297 | if (COMPARISON_P (x)) |
9298 | { |
9299 | *ptrue = simplify_gen_relational (code, mode, VOIDmode, |
9300 | op0: true0, op1: true1); |
9301 | *pfalse = simplify_gen_relational (code, mode, VOIDmode, |
9302 | op0: false0, op1: false1); |
9303 | } |
9304 | else |
9305 | { |
9306 | *ptrue = simplify_gen_binary (code, mode, op0: true0, op1: true1); |
9307 | *pfalse = simplify_gen_binary (code, mode, op0: false0, op1: false1); |
9308 | } |
9309 | |
9310 | return cond0 ? cond0 : cond1; |
9311 | } |
9312 | |
9313 | /* See if we have PLUS, IOR, XOR, MINUS or UMAX, where one of the |
9314 | operands is zero when the other is nonzero, and vice-versa, |
9315 | and STORE_FLAG_VALUE is 1 or -1. */ |
9316 | |
9317 | if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
9318 | && (code == PLUS || code == IOR || code == XOR || code == MINUS |
9319 | || code == UMAX) |
9320 | && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT) |
9321 | { |
9322 | rtx op0 = XEXP (XEXP (x, 0), 1); |
9323 | rtx op1 = XEXP (XEXP (x, 1), 1); |
9324 | |
9325 | cond0 = XEXP (XEXP (x, 0), 0); |
9326 | cond1 = XEXP (XEXP (x, 1), 0); |
9327 | |
9328 | if (COMPARISON_P (cond0) |
9329 | && COMPARISON_P (cond1) |
9330 | && SCALAR_INT_MODE_P (mode) |
9331 | && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL) |
9332 | && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0)) |
9333 | && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1))) |
9334 | || ((swap_condition (GET_CODE (cond0)) |
9335 | == reversed_comparison_code (cond1, NULL)) |
9336 | && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1)) |
9337 | && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0)))) |
9338 | && ! side_effects_p (x)) |
9339 | { |
9340 | *ptrue = simplify_gen_binary (code: MULT, mode, op0, op1: const_true_rtx); |
9341 | *pfalse = simplify_gen_binary (code: MULT, mode, |
9342 | op0: (code == MINUS |
9343 | ? simplify_gen_unary (code: NEG, mode, |
9344 | op: op1, op_mode: mode) |
9345 | : op1), |
9346 | op1: const_true_rtx); |
9347 | return cond0; |
9348 | } |
9349 | } |
9350 | |
9351 | /* Similarly for MULT, AND and UMIN, except that for these the result |
9352 | is always zero. */ |
9353 | if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
9354 | && (code == MULT || code == AND || code == UMIN) |
9355 | && GET_CODE (XEXP (x, 0)) == MULT && GET_CODE (XEXP (x, 1)) == MULT) |
9356 | { |
9357 | cond0 = XEXP (XEXP (x, 0), 0); |
9358 | cond1 = XEXP (XEXP (x, 1), 0); |
9359 | |
9360 | if (COMPARISON_P (cond0) |
9361 | && COMPARISON_P (cond1) |
9362 | && ((GET_CODE (cond0) == reversed_comparison_code (cond1, NULL) |
9363 | && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 0)) |
9364 | && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 1))) |
9365 | || ((swap_condition (GET_CODE (cond0)) |
9366 | == reversed_comparison_code (cond1, NULL)) |
9367 | && rtx_equal_p (XEXP (cond0, 0), XEXP (cond1, 1)) |
9368 | && rtx_equal_p (XEXP (cond0, 1), XEXP (cond1, 0)))) |
9369 | && ! side_effects_p (x)) |
9370 | { |
9371 | *ptrue = *pfalse = const0_rtx; |
9372 | return cond0; |
9373 | } |
9374 | } |
9375 | } |
9376 | |
9377 | else if (code == IF_THEN_ELSE) |
9378 | { |
9379 | /* If we have IF_THEN_ELSE already, extract the condition and |
9380 | canonicalize it if it is NE or EQ. */ |
9381 | cond0 = XEXP (x, 0); |
9382 | *ptrue = XEXP (x, 1), *pfalse = XEXP (x, 2); |
9383 | if (GET_CODE (cond0) == NE && XEXP (cond0, 1) == const0_rtx) |
9384 | return XEXP (cond0, 0); |
9385 | else if (GET_CODE (cond0) == EQ && XEXP (cond0, 1) == const0_rtx) |
9386 | { |
9387 | *ptrue = XEXP (x, 2), *pfalse = XEXP (x, 1); |
9388 | return XEXP (cond0, 0); |
9389 | } |
9390 | else |
9391 | return cond0; |
9392 | } |
9393 | |
9394 | /* If X is a SUBREG, we can narrow both the true and false values |
9395 | if the inner expression, if there is a condition. */ |
9396 | else if (code == SUBREG |
9397 | && (cond0 = if_then_else_cond (SUBREG_REG (x), ptrue: &true0, |
9398 | pfalse: &false0)) != 0) |
9399 | { |
9400 | true0 = simplify_gen_subreg (outermode: mode, op: true0, |
9401 | GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x)); |
9402 | false0 = simplify_gen_subreg (outermode: mode, op: false0, |
9403 | GET_MODE (SUBREG_REG (x)), SUBREG_BYTE (x)); |
9404 | if (true0 && false0) |
9405 | { |
9406 | *ptrue = true0; |
9407 | *pfalse = false0; |
9408 | return cond0; |
9409 | } |
9410 | } |
9411 | |
9412 | /* If X is a constant, this isn't special and will cause confusions |
9413 | if we treat it as such. Likewise if it is equivalent to a constant. */ |
9414 | else if (CONSTANT_P (x) |
9415 | || ((cond0 = get_last_value (x)) != 0 && CONSTANT_P (cond0))) |
9416 | ; |
9417 | |
9418 | /* If we're in BImode, canonicalize on 0 and STORE_FLAG_VALUE, as that |
9419 | will be least confusing to the rest of the compiler. */ |
9420 | else if (mode == BImode) |
9421 | { |
9422 | *ptrue = GEN_INT (STORE_FLAG_VALUE), *pfalse = const0_rtx; |
9423 | return x; |
9424 | } |
9425 | |
9426 | /* If X is known to be either 0 or -1, those are the true and |
9427 | false values when testing X. */ |
9428 | else if (x == constm1_rtx || x == const0_rtx |
9429 | || (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
9430 | && (num_sign_bit_copies (x, int_mode) |
9431 | == GET_MODE_PRECISION (mode: int_mode)))) |
9432 | { |
9433 | *ptrue = constm1_rtx, *pfalse = const0_rtx; |
9434 | return x; |
9435 | } |
9436 | |
9437 | /* Likewise for 0 or a single bit. */ |
9438 | else if (HWI_COMPUTABLE_MODE_P (mode) |
9439 | && pow2p_hwi (x: nz = nonzero_bits (x, mode))) |
9440 | { |
9441 | *ptrue = gen_int_mode (nz, mode), *pfalse = const0_rtx; |
9442 | return x; |
9443 | } |
9444 | |
9445 | /* Otherwise fail; show no condition with true and false values the same. */ |
9446 | *ptrue = *pfalse = x; |
9447 | return 0; |
9448 | } |
9449 | |
9450 | /* Return the value of expression X given the fact that condition COND |
9451 | is known to be true when applied to REG as its first operand and VAL |
9452 | as its second. X is known to not be shared and so can be modified in |
9453 | place. |
9454 | |
9455 | We only handle the simplest cases, and specifically those cases that |
9456 | arise with IF_THEN_ELSE expressions. */ |
9457 | |
9458 | static rtx |
9459 | known_cond (rtx x, enum rtx_code cond, rtx reg, rtx val) |
9460 | { |
9461 | enum rtx_code code = GET_CODE (x); |
9462 | const char *fmt; |
9463 | int i, j; |
9464 | |
9465 | if (side_effects_p (x)) |
9466 | return x; |
9467 | |
9468 | /* If either operand of the condition is a floating point value, |
9469 | then we have to avoid collapsing an EQ comparison. */ |
9470 | if (cond == EQ |
9471 | && rtx_equal_p (x, reg) |
9472 | && ! FLOAT_MODE_P (GET_MODE (x)) |
9473 | && ! FLOAT_MODE_P (GET_MODE (val))) |
9474 | return val; |
9475 | |
9476 | if (cond == UNEQ && rtx_equal_p (x, reg)) |
9477 | return val; |
9478 | |
9479 | /* If X is (abs REG) and we know something about REG's relationship |
9480 | with zero, we may be able to simplify this. */ |
9481 | |
9482 | if (code == ABS && rtx_equal_p (XEXP (x, 0), reg) && val == const0_rtx) |
9483 | switch (cond) |
9484 | { |
9485 | case GE: case GT: case EQ: |
9486 | return XEXP (x, 0); |
9487 | case LT: case LE: |
9488 | return simplify_gen_unary (code: NEG, GET_MODE (XEXP (x, 0)), |
9489 | XEXP (x, 0), |
9490 | GET_MODE (XEXP (x, 0))); |
9491 | default: |
9492 | break; |
9493 | } |
9494 | |
9495 | /* The only other cases we handle are MIN, MAX, and comparisons if the |
9496 | operands are the same as REG and VAL. */ |
9497 | |
9498 | else if (COMPARISON_P (x) || COMMUTATIVE_ARITH_P (x)) |
9499 | { |
9500 | if (rtx_equal_p (XEXP (x, 0), val)) |
9501 | { |
9502 | std::swap (a&: val, b&: reg); |
9503 | cond = swap_condition (cond); |
9504 | } |
9505 | |
9506 | if (rtx_equal_p (XEXP (x, 0), reg) && rtx_equal_p (XEXP (x, 1), val)) |
9507 | { |
9508 | if (COMPARISON_P (x)) |
9509 | { |
9510 | if (comparison_dominates_p (cond, code)) |
9511 | return VECTOR_MODE_P (GET_MODE (x)) ? x : const_true_rtx; |
9512 | |
9513 | code = reversed_comparison_code (x, NULL); |
9514 | if (code != UNKNOWN |
9515 | && comparison_dominates_p (cond, code)) |
9516 | return CONST0_RTX (GET_MODE (x)); |
9517 | else |
9518 | return x; |
9519 | } |
9520 | else if (code == SMAX || code == SMIN |
9521 | || code == UMIN || code == UMAX) |
9522 | { |
9523 | int unsignedp = (code == UMIN || code == UMAX); |
9524 | |
9525 | /* Do not reverse the condition when it is NE or EQ. |
9526 | This is because we cannot conclude anything about |
9527 | the value of 'SMAX (x, y)' when x is not equal to y, |
9528 | but we can when x equals y. */ |
9529 | if ((code == SMAX || code == UMAX) |
9530 | && ! (cond == EQ || cond == NE)) |
9531 | cond = reverse_condition (cond); |
9532 | |
9533 | switch (cond) |
9534 | { |
9535 | case GE: case GT: |
9536 | return unsignedp ? x : XEXP (x, 1); |
9537 | case LE: case LT: |
9538 | return unsignedp ? x : XEXP (x, 0); |
9539 | case GEU: case GTU: |
9540 | return unsignedp ? XEXP (x, 1) : x; |
9541 | case LEU: case LTU: |
9542 | return unsignedp ? XEXP (x, 0) : x; |
9543 | default: |
9544 | break; |
9545 | } |
9546 | } |
9547 | } |
9548 | } |
9549 | else if (code == SUBREG) |
9550 | { |
9551 | machine_mode inner_mode = GET_MODE (SUBREG_REG (x)); |
9552 | rtx new_rtx, r = known_cond (SUBREG_REG (x), cond, reg, val); |
9553 | |
9554 | if (SUBREG_REG (x) != r) |
9555 | { |
9556 | /* We must simplify subreg here, before we lose track of the |
9557 | original inner_mode. */ |
9558 | new_rtx = simplify_subreg (GET_MODE (x), op: r, |
9559 | innermode: inner_mode, SUBREG_BYTE (x)); |
9560 | if (new_rtx) |
9561 | return new_rtx; |
9562 | else |
9563 | SUBST (SUBREG_REG (x), r); |
9564 | } |
9565 | |
9566 | return x; |
9567 | } |
9568 | /* We don't have to handle SIGN_EXTEND here, because even in the |
9569 | case of replacing something with a modeless CONST_INT, a |
9570 | CONST_INT is already (supposed to be) a valid sign extension for |
9571 | its narrower mode, which implies it's already properly |
9572 | sign-extended for the wider mode. Now, for ZERO_EXTEND, the |
9573 | story is different. */ |
9574 | else if (code == ZERO_EXTEND) |
9575 | { |
9576 | machine_mode inner_mode = GET_MODE (XEXP (x, 0)); |
9577 | rtx new_rtx, r = known_cond (XEXP (x, 0), cond, reg, val); |
9578 | |
9579 | if (XEXP (x, 0) != r) |
9580 | { |
9581 | /* We must simplify the zero_extend here, before we lose |
9582 | track of the original inner_mode. */ |
9583 | new_rtx = simplify_unary_operation (code: ZERO_EXTEND, GET_MODE (x), |
9584 | op: r, op_mode: inner_mode); |
9585 | if (new_rtx) |
9586 | return new_rtx; |
9587 | else |
9588 | SUBST (XEXP (x, 0), r); |
9589 | } |
9590 | |
9591 | return x; |
9592 | } |
9593 | |
9594 | fmt = GET_RTX_FORMAT (code); |
9595 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
9596 | { |
9597 | if (fmt[i] == 'e') |
9598 | SUBST (XEXP (x, i), known_cond (XEXP (x, i), cond, reg, val)); |
9599 | else if (fmt[i] == 'E') |
9600 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
9601 | SUBST (XVECEXP (x, i, j), known_cond (XVECEXP (x, i, j), |
9602 | cond, reg, val)); |
9603 | } |
9604 | |
9605 | return x; |
9606 | } |
9607 | |
9608 | /* See if X and Y are equal for the purposes of seeing if we can rewrite an |
9609 | assignment as a field assignment. */ |
9610 | |
9611 | static bool |
9612 | rtx_equal_for_field_assignment_p (rtx x, rtx y, bool widen_x) |
9613 | { |
9614 | if (widen_x && GET_MODE (x) != GET_MODE (y)) |
9615 | { |
9616 | if (paradoxical_subreg_p (GET_MODE (x), GET_MODE (y))) |
9617 | return false; |
9618 | if (BYTES_BIG_ENDIAN != WORDS_BIG_ENDIAN) |
9619 | return false; |
9620 | x = adjust_address_nv (x, GET_MODE (y), |
9621 | byte_lowpart_offset (GET_MODE (y), |
9622 | GET_MODE (x))); |
9623 | } |
9624 | |
9625 | if (x == y || rtx_equal_p (x, y)) |
9626 | return true; |
9627 | |
9628 | if (x == 0 || y == 0 || GET_MODE (x) != GET_MODE (y)) |
9629 | return false; |
9630 | |
9631 | /* Check for a paradoxical SUBREG of a MEM compared with the MEM. |
9632 | Note that all SUBREGs of MEM are paradoxical; otherwise they |
9633 | would have been rewritten. */ |
9634 | if (MEM_P (x) && GET_CODE (y) == SUBREG |
9635 | && MEM_P (SUBREG_REG (y)) |
9636 | && rtx_equal_p (SUBREG_REG (y), |
9637 | gen_lowpart (GET_MODE (SUBREG_REG (y)), x))) |
9638 | return true; |
9639 | |
9640 | if (MEM_P (y) && GET_CODE (x) == SUBREG |
9641 | && MEM_P (SUBREG_REG (x)) |
9642 | && rtx_equal_p (SUBREG_REG (x), |
9643 | gen_lowpart (GET_MODE (SUBREG_REG (x)), y))) |
9644 | return true; |
9645 | |
9646 | /* We used to see if get_last_value of X and Y were the same but that's |
9647 | not correct. In one direction, we'll cause the assignment to have |
9648 | the wrong destination and in the case, we'll import a register into this |
9649 | insn that might have already have been dead. So fail if none of the |
9650 | above cases are true. */ |
9651 | return false; |
9652 | } |
9653 | |
9654 | /* See if X, a SET operation, can be rewritten as a bit-field assignment. |
9655 | Return that assignment if so. |
9656 | |
9657 | We only handle the most common cases. */ |
9658 | |
9659 | static rtx |
9660 | make_field_assignment (rtx x) |
9661 | { |
9662 | rtx dest = SET_DEST (x); |
9663 | rtx src = SET_SRC (x); |
9664 | rtx assign; |
9665 | rtx rhs, lhs; |
9666 | HOST_WIDE_INT c1; |
9667 | HOST_WIDE_INT pos; |
9668 | unsigned HOST_WIDE_INT len; |
9669 | rtx other; |
9670 | |
9671 | /* All the rules in this function are specific to scalar integers. */ |
9672 | scalar_int_mode mode; |
9673 | if (!is_a <scalar_int_mode> (GET_MODE (dest), result: &mode)) |
9674 | return x; |
9675 | |
9676 | /* If SRC was (and (not (ashift (const_int 1) POS)) DEST), this is |
9677 | a clear of a one-bit field. We will have changed it to |
9678 | (and (rotate (const_int -2) POS) DEST), so check for that. Also check |
9679 | for a SUBREG. */ |
9680 | |
9681 | if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == ROTATE |
9682 | && CONST_INT_P (XEXP (XEXP (src, 0), 0)) |
9683 | && INTVAL (XEXP (XEXP (src, 0), 0)) == -2 |
9684 | && rtx_equal_for_field_assignment_p (x: dest, XEXP (src, 1))) |
9685 | { |
9686 | assign = make_extraction (VOIDmode, inner: dest, pos: 0, XEXP (XEXP (src, 0), 1), |
9687 | len: 1, unsignedp: true, in_dest: true, in_compare: false); |
9688 | if (assign != 0) |
9689 | return gen_rtx_SET (assign, const0_rtx); |
9690 | return x; |
9691 | } |
9692 | |
9693 | if (GET_CODE (src) == AND && GET_CODE (XEXP (src, 0)) == SUBREG |
9694 | && subreg_lowpart_p (XEXP (src, 0)) |
9695 | && partial_subreg_p (XEXP (src, 0)) |
9696 | && GET_CODE (SUBREG_REG (XEXP (src, 0))) == ROTATE |
9697 | && CONST_INT_P (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) |
9698 | && INTVAL (XEXP (SUBREG_REG (XEXP (src, 0)), 0)) == -2 |
9699 | && rtx_equal_for_field_assignment_p (x: dest, XEXP (src, 1))) |
9700 | { |
9701 | assign = make_extraction (VOIDmode, inner: dest, pos: 0, |
9702 | XEXP (SUBREG_REG (XEXP (src, 0)), 1), |
9703 | len: 1, unsignedp: true, in_dest: true, in_compare: false); |
9704 | if (assign != 0) |
9705 | return gen_rtx_SET (assign, const0_rtx); |
9706 | return x; |
9707 | } |
9708 | |
9709 | /* If SRC is (ior (ashift (const_int 1) POS) DEST), this is a set of a |
9710 | one-bit field. */ |
9711 | if (GET_CODE (src) == IOR && GET_CODE (XEXP (src, 0)) == ASHIFT |
9712 | && XEXP (XEXP (src, 0), 0) == const1_rtx |
9713 | && rtx_equal_for_field_assignment_p (x: dest, XEXP (src, 1))) |
9714 | { |
9715 | assign = make_extraction (VOIDmode, inner: dest, pos: 0, XEXP (XEXP (src, 0), 1), |
9716 | len: 1, unsignedp: true, in_dest: true, in_compare: false); |
9717 | if (assign != 0) |
9718 | return gen_rtx_SET (assign, const1_rtx); |
9719 | return x; |
9720 | } |
9721 | |
9722 | /* If DEST is already a field assignment, i.e. ZERO_EXTRACT, and the |
9723 | SRC is an AND with all bits of that field set, then we can discard |
9724 | the AND. */ |
9725 | if (GET_CODE (dest) == ZERO_EXTRACT |
9726 | && CONST_INT_P (XEXP (dest, 1)) |
9727 | && GET_CODE (src) == AND |
9728 | && CONST_INT_P (XEXP (src, 1))) |
9729 | { |
9730 | HOST_WIDE_INT width = INTVAL (XEXP (dest, 1)); |
9731 | unsigned HOST_WIDE_INT and_mask = INTVAL (XEXP (src, 1)); |
9732 | unsigned HOST_WIDE_INT ze_mask; |
9733 | |
9734 | if (width >= HOST_BITS_PER_WIDE_INT) |
9735 | ze_mask = -1; |
9736 | else |
9737 | ze_mask = ((unsigned HOST_WIDE_INT)1 << width) - 1; |
9738 | |
9739 | /* Complete overlap. We can remove the source AND. */ |
9740 | if ((and_mask & ze_mask) == ze_mask) |
9741 | return gen_rtx_SET (dest, XEXP (src, 0)); |
9742 | |
9743 | /* Partial overlap. We can reduce the source AND. */ |
9744 | if ((and_mask & ze_mask) != and_mask) |
9745 | { |
9746 | src = gen_rtx_AND (mode, XEXP (src, 0), |
9747 | gen_int_mode (and_mask & ze_mask, mode)); |
9748 | return gen_rtx_SET (dest, src); |
9749 | } |
9750 | } |
9751 | |
9752 | /* The other case we handle is assignments into a constant-position |
9753 | field. They look like (ior/xor (and DEST C1) OTHER). If C1 represents |
9754 | a mask that has all one bits except for a group of zero bits and |
9755 | OTHER is known to have zeros where C1 has ones, this is such an |
9756 | assignment. Compute the position and length from C1. Shift OTHER |
9757 | to the appropriate position, force it to the required mode, and |
9758 | make the extraction. Check for the AND in both operands. */ |
9759 | |
9760 | /* One or more SUBREGs might obscure the constant-position field |
9761 | assignment. The first one we are likely to encounter is an outer |
9762 | narrowing SUBREG, which we can just strip for the purposes of |
9763 | identifying the constant-field assignment. */ |
9764 | scalar_int_mode src_mode = mode; |
9765 | if (GET_CODE (src) == SUBREG |
9766 | && subreg_lowpart_p (src) |
9767 | && is_a <scalar_int_mode> (GET_MODE (SUBREG_REG (src)), result: &src_mode)) |
9768 | src = SUBREG_REG (src); |
9769 | |
9770 | if (GET_CODE (src) != IOR && GET_CODE (src) != XOR) |
9771 | return x; |
9772 | |
9773 | rhs = expand_compound_operation (XEXP (src, 0)); |
9774 | lhs = expand_compound_operation (XEXP (src, 1)); |
9775 | |
9776 | if (GET_CODE (rhs) == AND |
9777 | && CONST_INT_P (XEXP (rhs, 1)) |
9778 | && rtx_equal_for_field_assignment_p (XEXP (rhs, 0), y: dest)) |
9779 | c1 = INTVAL (XEXP (rhs, 1)), other = lhs; |
9780 | /* The second SUBREG that might get in the way is a paradoxical |
9781 | SUBREG around the first operand of the AND. We want to |
9782 | pretend the operand is as wide as the destination here. We |
9783 | do this by adjusting the MEM to wider mode for the sole |
9784 | purpose of the call to rtx_equal_for_field_assignment_p. Also |
9785 | note this trick only works for MEMs. */ |
9786 | else if (GET_CODE (rhs) == AND |
9787 | && paradoxical_subreg_p (XEXP (rhs, 0)) |
9788 | && MEM_P (SUBREG_REG (XEXP (rhs, 0))) |
9789 | && CONST_INT_P (XEXP (rhs, 1)) |
9790 | && rtx_equal_for_field_assignment_p (SUBREG_REG (XEXP (rhs, 0)), |
9791 | y: dest, widen_x: true)) |
9792 | c1 = INTVAL (XEXP (rhs, 1)), other = lhs; |
9793 | else if (GET_CODE (lhs) == AND |
9794 | && CONST_INT_P (XEXP (lhs, 1)) |
9795 | && rtx_equal_for_field_assignment_p (XEXP (lhs, 0), y: dest)) |
9796 | c1 = INTVAL (XEXP (lhs, 1)), other = rhs; |
9797 | /* The second SUBREG that might get in the way is a paradoxical |
9798 | SUBREG around the first operand of the AND. We want to |
9799 | pretend the operand is as wide as the destination here. We |
9800 | do this by adjusting the MEM to wider mode for the sole |
9801 | purpose of the call to rtx_equal_for_field_assignment_p. Also |
9802 | note this trick only works for MEMs. */ |
9803 | else if (GET_CODE (lhs) == AND |
9804 | && paradoxical_subreg_p (XEXP (lhs, 0)) |
9805 | && MEM_P (SUBREG_REG (XEXP (lhs, 0))) |
9806 | && CONST_INT_P (XEXP (lhs, 1)) |
9807 | && rtx_equal_for_field_assignment_p (SUBREG_REG (XEXP (lhs, 0)), |
9808 | y: dest, widen_x: true)) |
9809 | c1 = INTVAL (XEXP (lhs, 1)), other = rhs; |
9810 | else |
9811 | return x; |
9812 | |
9813 | pos = get_pos_from_mask (m: (~c1) & GET_MODE_MASK (mode), plen: &len); |
9814 | if (pos < 0 |
9815 | || pos + len > GET_MODE_PRECISION (mode) |
9816 | || GET_MODE_PRECISION (mode) > HOST_BITS_PER_WIDE_INT |
9817 | || (c1 & nonzero_bits (other, mode)) != 0) |
9818 | return x; |
9819 | |
9820 | assign = make_extraction (VOIDmode, inner: dest, pos, NULL_RTX, len, |
9821 | unsignedp: true, in_dest: true, in_compare: false); |
9822 | if (assign == 0) |
9823 | return x; |
9824 | |
9825 | /* The mode to use for the source is the mode of the assignment, or of |
9826 | what is inside a possible STRICT_LOW_PART. */ |
9827 | machine_mode new_mode = (GET_CODE (assign) == STRICT_LOW_PART |
9828 | ? GET_MODE (XEXP (assign, 0)) : GET_MODE (assign)); |
9829 | |
9830 | /* Shift OTHER right POS places and make it the source, restricting it |
9831 | to the proper length and mode. */ |
9832 | |
9833 | src = canon_reg_for_combine (x: simplify_shift_const (NULL_RTX, LSHIFTRT, |
9834 | src_mode, other, pos), |
9835 | reg: dest); |
9836 | src = force_to_mode (x: src, mode: new_mode, |
9837 | mask: len >= HOST_BITS_PER_WIDE_INT |
9838 | ? HOST_WIDE_INT_M1U |
9839 | : (HOST_WIDE_INT_1U << len) - 1, just_select: false); |
9840 | |
9841 | /* If SRC is masked by an AND that does not make a difference in |
9842 | the value being stored, strip it. */ |
9843 | if (GET_CODE (assign) == ZERO_EXTRACT |
9844 | && CONST_INT_P (XEXP (assign, 1)) |
9845 | && INTVAL (XEXP (assign, 1)) < HOST_BITS_PER_WIDE_INT |
9846 | && GET_CODE (src) == AND |
9847 | && CONST_INT_P (XEXP (src, 1)) |
9848 | && UINTVAL (XEXP (src, 1)) |
9849 | == (HOST_WIDE_INT_1U << INTVAL (XEXP (assign, 1))) - 1) |
9850 | src = XEXP (src, 0); |
9851 | |
9852 | return gen_rtx_SET (assign, src); |
9853 | } |
9854 | |
9855 | /* See if X is of the form (+ (* a c) (* b c)) and convert to (* (+ a b) c) |
9856 | if so. */ |
9857 | |
9858 | static rtx |
9859 | apply_distributive_law (rtx x) |
9860 | { |
9861 | enum rtx_code code = GET_CODE (x); |
9862 | enum rtx_code inner_code; |
9863 | rtx lhs, rhs, other; |
9864 | rtx tem; |
9865 | |
9866 | /* Distributivity is not true for floating point as it can change the |
9867 | value. So we don't do it unless -funsafe-math-optimizations. */ |
9868 | if (FLOAT_MODE_P (GET_MODE (x)) |
9869 | && ! flag_unsafe_math_optimizations) |
9870 | return x; |
9871 | |
9872 | /* The outer operation can only be one of the following: */ |
9873 | if (code != IOR && code != AND && code != XOR |
9874 | && code != PLUS && code != MINUS) |
9875 | return x; |
9876 | |
9877 | lhs = XEXP (x, 0); |
9878 | rhs = XEXP (x, 1); |
9879 | |
9880 | /* If either operand is a primitive we can't do anything, so get out |
9881 | fast. */ |
9882 | if (OBJECT_P (lhs) || OBJECT_P (rhs)) |
9883 | return x; |
9884 | |
9885 | lhs = expand_compound_operation (x: lhs); |
9886 | rhs = expand_compound_operation (x: rhs); |
9887 | inner_code = GET_CODE (lhs); |
9888 | if (inner_code != GET_CODE (rhs)) |
9889 | return x; |
9890 | |
9891 | /* See if the inner and outer operations distribute. */ |
9892 | switch (inner_code) |
9893 | { |
9894 | case LSHIFTRT: |
9895 | case ASHIFTRT: |
9896 | case AND: |
9897 | case IOR: |
9898 | /* These all distribute except over PLUS. */ |
9899 | if (code == PLUS || code == MINUS) |
9900 | return x; |
9901 | break; |
9902 | |
9903 | case MULT: |
9904 | if (code != PLUS && code != MINUS) |
9905 | return x; |
9906 | break; |
9907 | |
9908 | case ASHIFT: |
9909 | /* This is also a multiply, so it distributes over everything. */ |
9910 | break; |
9911 | |
9912 | /* This used to handle SUBREG, but this turned out to be counter- |
9913 | productive, since (subreg (op ...)) usually is not handled by |
9914 | insn patterns, and this "optimization" therefore transformed |
9915 | recognizable patterns into unrecognizable ones. Therefore the |
9916 | SUBREG case was removed from here. |
9917 | |
9918 | It is possible that distributing SUBREG over arithmetic operations |
9919 | leads to an intermediate result than can then be optimized further, |
9920 | e.g. by moving the outer SUBREG to the other side of a SET as done |
9921 | in simplify_set. This seems to have been the original intent of |
9922 | handling SUBREGs here. |
9923 | |
9924 | However, with current GCC this does not appear to actually happen, |
9925 | at least on major platforms. If some case is found where removing |
9926 | the SUBREG case here prevents follow-on optimizations, distributing |
9927 | SUBREGs ought to be re-added at that place, e.g. in simplify_set. */ |
9928 | |
9929 | default: |
9930 | return x; |
9931 | } |
9932 | |
9933 | /* Set LHS and RHS to the inner operands (A and B in the example |
9934 | above) and set OTHER to the common operand (C in the example). |
9935 | There is only one way to do this unless the inner operation is |
9936 | commutative. */ |
9937 | if (COMMUTATIVE_ARITH_P (lhs) |
9938 | && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 0))) |
9939 | other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 1); |
9940 | else if (COMMUTATIVE_ARITH_P (lhs) |
9941 | && rtx_equal_p (XEXP (lhs, 0), XEXP (rhs, 1))) |
9942 | other = XEXP (lhs, 0), lhs = XEXP (lhs, 1), rhs = XEXP (rhs, 0); |
9943 | else if (COMMUTATIVE_ARITH_P (lhs) |
9944 | && rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 0))) |
9945 | other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 1); |
9946 | else if (rtx_equal_p (XEXP (lhs, 1), XEXP (rhs, 1))) |
9947 | other = XEXP (lhs, 1), lhs = XEXP (lhs, 0), rhs = XEXP (rhs, 0); |
9948 | else |
9949 | return x; |
9950 | |
9951 | /* Form the new inner operation, seeing if it simplifies first. */ |
9952 | tem = simplify_gen_binary (code, GET_MODE (x), op0: lhs, op1: rhs); |
9953 | |
9954 | /* There is one exception to the general way of distributing: |
9955 | (a | c) ^ (b | c) -> (a ^ b) & ~c */ |
9956 | if (code == XOR && inner_code == IOR) |
9957 | { |
9958 | inner_code = AND; |
9959 | other = simplify_gen_unary (code: NOT, GET_MODE (x), op: other, GET_MODE (x)); |
9960 | } |
9961 | |
9962 | /* We may be able to continuing distributing the result, so call |
9963 | ourselves recursively on the inner operation before forming the |
9964 | outer operation, which we return. */ |
9965 | return simplify_gen_binary (code: inner_code, GET_MODE (x), |
9966 | op0: apply_distributive_law (x: tem), op1: other); |
9967 | } |
9968 | |
9969 | /* See if X is of the form (* (+ A B) C), and if so convert to |
9970 | (+ (* A C) (* B C)) and try to simplify. |
9971 | |
9972 | Most of the time, this results in no change. However, if some of |
9973 | the operands are the same or inverses of each other, simplifications |
9974 | will result. |
9975 | |
9976 | For example, (and (ior A B) (not B)) can occur as the result of |
9977 | expanding a bit field assignment. When we apply the distributive |
9978 | law to this, we get (ior (and (A (not B))) (and (B (not B)))), |
9979 | which then simplifies to (and (A (not B))). |
9980 | |
9981 | Note that no checks happen on the validity of applying the inverse |
9982 | distributive law. This is pointless since we can do it in the |
9983 | few places where this routine is called. |
9984 | |
9985 | N is the index of the term that is decomposed (the arithmetic operation, |
9986 | i.e. (+ A B) in the first example above). !N is the index of the term that |
9987 | is distributed, i.e. of C in the first example above. */ |
9988 | static rtx |
9989 | distribute_and_simplify_rtx (rtx x, int n) |
9990 | { |
9991 | machine_mode mode; |
9992 | enum rtx_code outer_code, inner_code; |
9993 | rtx decomposed, distributed, inner_op0, inner_op1, new_op0, new_op1, tmp; |
9994 | |
9995 | /* Distributivity is not true for floating point as it can change the |
9996 | value. So we don't do it unless -funsafe-math-optimizations. */ |
9997 | if (FLOAT_MODE_P (GET_MODE (x)) |
9998 | && ! flag_unsafe_math_optimizations) |
9999 | return NULL_RTX; |
10000 | |
10001 | decomposed = XEXP (x, n); |
10002 | if (!ARITHMETIC_P (decomposed)) |
10003 | return NULL_RTX; |
10004 | |
10005 | mode = GET_MODE (x); |
10006 | outer_code = GET_CODE (x); |
10007 | distributed = XEXP (x, !n); |
10008 | |
10009 | inner_code = GET_CODE (decomposed); |
10010 | inner_op0 = XEXP (decomposed, 0); |
10011 | inner_op1 = XEXP (decomposed, 1); |
10012 | |
10013 | /* Special case (and (xor B C) (not A)), which is equivalent to |
10014 | (xor (ior A B) (ior A C)) */ |
10015 | if (outer_code == AND && inner_code == XOR && GET_CODE (distributed) == NOT) |
10016 | { |
10017 | distributed = XEXP (distributed, 0); |
10018 | outer_code = IOR; |
10019 | } |
10020 | |
10021 | if (n == 0) |
10022 | { |
10023 | /* Distribute the second term. */ |
10024 | new_op0 = simplify_gen_binary (code: outer_code, mode, op0: inner_op0, op1: distributed); |
10025 | new_op1 = simplify_gen_binary (code: outer_code, mode, op0: inner_op1, op1: distributed); |
10026 | } |
10027 | else |
10028 | { |
10029 | /* Distribute the first term. */ |
10030 | new_op0 = simplify_gen_binary (code: outer_code, mode, op0: distributed, op1: inner_op0); |
10031 | new_op1 = simplify_gen_binary (code: outer_code, mode, op0: distributed, op1: inner_op1); |
10032 | } |
10033 | |
10034 | tmp = apply_distributive_law (x: simplify_gen_binary (code: inner_code, mode, |
10035 | op0: new_op0, op1: new_op1)); |
10036 | if (GET_CODE (tmp) != outer_code |
10037 | && (set_src_cost (x: tmp, mode, speed_p: optimize_this_for_speed_p) |
10038 | < set_src_cost (x, mode, speed_p: optimize_this_for_speed_p))) |
10039 | return tmp; |
10040 | |
10041 | return NULL_RTX; |
10042 | } |
10043 | |
10044 | /* Simplify a logical `and' of VAROP with the constant CONSTOP, to be done |
10045 | in MODE. Return an equivalent form, if different from (and VAROP |
10046 | (const_int CONSTOP)). Otherwise, return NULL_RTX. */ |
10047 | |
10048 | static rtx |
10049 | simplify_and_const_int_1 (scalar_int_mode mode, rtx varop, |
10050 | unsigned HOST_WIDE_INT constop) |
10051 | { |
10052 | unsigned HOST_WIDE_INT nonzero; |
10053 | unsigned HOST_WIDE_INT orig_constop; |
10054 | rtx orig_varop; |
10055 | int i; |
10056 | |
10057 | orig_varop = varop; |
10058 | orig_constop = constop; |
10059 | if (GET_CODE (varop) == CLOBBER) |
10060 | return NULL_RTX; |
10061 | |
10062 | /* Simplify VAROP knowing that we will be only looking at some of the |
10063 | bits in it. |
10064 | |
10065 | Note by passing in CONSTOP, we guarantee that the bits not set in |
10066 | CONSTOP are not significant and will never be examined. We must |
10067 | ensure that is the case by explicitly masking out those bits |
10068 | before returning. */ |
10069 | varop = force_to_mode (x: varop, mode, mask: constop, just_select: false); |
10070 | |
10071 | /* If VAROP is a CLOBBER, we will fail so return it. */ |
10072 | if (GET_CODE (varop) == CLOBBER) |
10073 | return varop; |
10074 | |
10075 | /* If VAROP is a CONST_INT, then we need to apply the mask in CONSTOP |
10076 | to VAROP and return the new constant. */ |
10077 | if (CONST_INT_P (varop)) |
10078 | return gen_int_mode (INTVAL (varop) & constop, mode); |
10079 | |
10080 | /* See what bits may be nonzero in VAROP. Unlike the general case of |
10081 | a call to nonzero_bits, here we don't care about bits outside |
10082 | MODE unless WORD_REGISTER_OPERATIONS is true. */ |
10083 | |
10084 | scalar_int_mode tmode = mode; |
10085 | if (WORD_REGISTER_OPERATIONS && GET_MODE_BITSIZE (mode) < BITS_PER_WORD) |
10086 | tmode = word_mode; |
10087 | nonzero = nonzero_bits (varop, tmode) & GET_MODE_MASK (tmode); |
10088 | |
10089 | /* Turn off all bits in the constant that are known to already be zero. |
10090 | Thus, if the AND isn't needed at all, we will have CONSTOP == NONZERO_BITS |
10091 | which is tested below. */ |
10092 | |
10093 | constop &= nonzero; |
10094 | |
10095 | /* If we don't have any bits left, return zero. */ |
10096 | if (constop == 0 && !side_effects_p (varop)) |
10097 | return const0_rtx; |
10098 | |
10099 | /* If VAROP is a NEG of something known to be zero or 1 and CONSTOP is |
10100 | a power of two, we can replace this with an ASHIFT. */ |
10101 | if (GET_CODE (varop) == NEG && nonzero_bits (XEXP (varop, 0), tmode) == 1 |
10102 | && (i = exact_log2 (x: constop)) >= 0) |
10103 | return simplify_shift_const (NULL_RTX, ASHIFT, mode, XEXP (varop, 0), i); |
10104 | |
10105 | /* If VAROP is an IOR or XOR, apply the AND to both branches of the IOR |
10106 | or XOR, then try to apply the distributive law. This may eliminate |
10107 | operations if either branch can be simplified because of the AND. |
10108 | It may also make some cases more complex, but those cases probably |
10109 | won't match a pattern either with or without this. */ |
10110 | |
10111 | if (GET_CODE (varop) == IOR || GET_CODE (varop) == XOR) |
10112 | { |
10113 | scalar_int_mode varop_mode = as_a <scalar_int_mode> (GET_MODE (varop)); |
10114 | return |
10115 | gen_lowpart |
10116 | (mode, |
10117 | apply_distributive_law |
10118 | (x: simplify_gen_binary (GET_CODE (varop), mode: varop_mode, |
10119 | op0: simplify_and_const_int (NULL_RTX, varop_mode, |
10120 | XEXP (varop, 0), |
10121 | constop), |
10122 | op1: simplify_and_const_int (NULL_RTX, varop_mode, |
10123 | XEXP (varop, 1), |
10124 | constop)))); |
10125 | } |
10126 | |
10127 | /* If VAROP is PLUS, and the constant is a mask of low bits, distribute |
10128 | the AND and see if one of the operands simplifies to zero. If so, we |
10129 | may eliminate it. */ |
10130 | |
10131 | if (GET_CODE (varop) == PLUS |
10132 | && pow2p_hwi (x: constop + 1)) |
10133 | { |
10134 | rtx o0, o1; |
10135 | |
10136 | o0 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 0), constop); |
10137 | o1 = simplify_and_const_int (NULL_RTX, mode, XEXP (varop, 1), constop); |
10138 | if (o0 == const0_rtx) |
10139 | return o1; |
10140 | if (o1 == const0_rtx) |
10141 | return o0; |
10142 | } |
10143 | |
10144 | /* Make a SUBREG if necessary. If we can't make it, fail. */ |
10145 | varop = gen_lowpart (mode, varop); |
10146 | if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER) |
10147 | return NULL_RTX; |
10148 | |
10149 | /* If we are only masking insignificant bits, return VAROP. */ |
10150 | if (constop == nonzero) |
10151 | return varop; |
10152 | |
10153 | if (varop == orig_varop && constop == orig_constop) |
10154 | return NULL_RTX; |
10155 | |
10156 | /* Otherwise, return an AND. */ |
10157 | return simplify_gen_binary (code: AND, mode, op0: varop, op1: gen_int_mode (constop, mode)); |
10158 | } |
10159 | |
10160 | |
10161 | /* We have X, a logical `and' of VAROP with the constant CONSTOP, to be done |
10162 | in MODE. |
10163 | |
10164 | Return an equivalent form, if different from X. Otherwise, return X. If |
10165 | X is zero, we are to always construct the equivalent form. */ |
10166 | |
10167 | static rtx |
10168 | simplify_and_const_int (rtx x, scalar_int_mode mode, rtx varop, |
10169 | unsigned HOST_WIDE_INT constop) |
10170 | { |
10171 | rtx tem = simplify_and_const_int_1 (mode, varop, constop); |
10172 | if (tem) |
10173 | return tem; |
10174 | |
10175 | if (!x) |
10176 | x = simplify_gen_binary (code: AND, GET_MODE (varop), op0: varop, |
10177 | op1: gen_int_mode (constop, mode)); |
10178 | if (GET_MODE (x) != mode) |
10179 | x = gen_lowpart (mode, x); |
10180 | return x; |
10181 | } |
10182 | |
10183 | /* Given a REG X of mode XMODE, compute which bits in X can be nonzero. |
10184 | We don't care about bits outside of those defined in MODE. |
10185 | We DO care about all the bits in MODE, even if XMODE is smaller than MODE. |
10186 | |
10187 | For most X this is simply GET_MODE_MASK (GET_MODE (MODE)), but if X is |
10188 | a shift, AND, or zero_extract, we can do better. */ |
10189 | |
10190 | static rtx |
10191 | reg_nonzero_bits_for_combine (const_rtx x, scalar_int_mode xmode, |
10192 | scalar_int_mode mode, |
10193 | unsigned HOST_WIDE_INT *nonzero) |
10194 | { |
10195 | rtx tem; |
10196 | reg_stat_type *rsp; |
10197 | |
10198 | /* If X is a register whose nonzero bits value is current, use it. |
10199 | Otherwise, if X is a register whose value we can find, use that |
10200 | value. Otherwise, use the previously-computed global nonzero bits |
10201 | for this register. */ |
10202 | |
10203 | rsp = ®_stat[REGNO (x)]; |
10204 | if (rsp->last_set_value != 0 |
10205 | && (rsp->last_set_mode == mode |
10206 | || (REGNO (x) >= FIRST_PSEUDO_REGISTER |
10207 | && GET_MODE_CLASS (rsp->last_set_mode) == MODE_INT |
10208 | && GET_MODE_CLASS (mode) == MODE_INT)) |
10209 | && ((rsp->last_set_label >= label_tick_ebb_start |
10210 | && rsp->last_set_label < label_tick) |
10211 | || (rsp->last_set_label == label_tick |
10212 | && DF_INSN_LUID (rsp->last_set) < subst_low_luid) |
10213 | || (REGNO (x) >= FIRST_PSEUDO_REGISTER |
10214 | && REGNO (x) < reg_n_sets_max |
10215 | && REG_N_SETS (REGNO (x)) == 1 |
10216 | && !REGNO_REG_SET_P |
10217 | (DF_LR_IN (ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb), |
10218 | REGNO (x))))) |
10219 | { |
10220 | /* Note that, even if the precision of last_set_mode is lower than that |
10221 | of mode, record_value_for_reg invoked nonzero_bits on the register |
10222 | with nonzero_bits_mode (because last_set_mode is necessarily integral |
10223 | and HWI_COMPUTABLE_MODE_P in this case) so bits in nonzero_bits_mode |
10224 | are all valid, hence in mode too since nonzero_bits_mode is defined |
10225 | to the largest HWI_COMPUTABLE_MODE_P mode. */ |
10226 | *nonzero &= rsp->last_set_nonzero_bits; |
10227 | return NULL; |
10228 | } |
10229 | |
10230 | tem = get_last_value (x); |
10231 | if (tem) |
10232 | { |
10233 | if (SHORT_IMMEDIATES_SIGN_EXTEND) |
10234 | tem = sign_extend_short_imm (src: tem, mode: xmode, prec: GET_MODE_PRECISION (mode)); |
10235 | |
10236 | return tem; |
10237 | } |
10238 | |
10239 | if (nonzero_sign_valid && rsp->nonzero_bits) |
10240 | { |
10241 | unsigned HOST_WIDE_INT mask = rsp->nonzero_bits; |
10242 | |
10243 | if (GET_MODE_PRECISION (mode: xmode) < GET_MODE_PRECISION (mode)) |
10244 | /* We don't know anything about the upper bits. */ |
10245 | mask |= GET_MODE_MASK (mode) ^ GET_MODE_MASK (xmode); |
10246 | |
10247 | *nonzero &= mask; |
10248 | } |
10249 | |
10250 | return NULL; |
10251 | } |
10252 | |
10253 | /* Given a reg X of mode XMODE, return the number of bits at the high-order |
10254 | end of X that are known to be equal to the sign bit. X will be used |
10255 | in mode MODE; the returned value will always be between 1 and the |
10256 | number of bits in MODE. */ |
10257 | |
10258 | static rtx |
10259 | reg_num_sign_bit_copies_for_combine (const_rtx x, scalar_int_mode xmode, |
10260 | scalar_int_mode mode, |
10261 | unsigned int *result) |
10262 | { |
10263 | rtx tem; |
10264 | reg_stat_type *rsp; |
10265 | |
10266 | rsp = ®_stat[REGNO (x)]; |
10267 | if (rsp->last_set_value != 0 |
10268 | && rsp->last_set_mode == mode |
10269 | && ((rsp->last_set_label >= label_tick_ebb_start |
10270 | && rsp->last_set_label < label_tick) |
10271 | || (rsp->last_set_label == label_tick |
10272 | && DF_INSN_LUID (rsp->last_set) < subst_low_luid) |
10273 | || (REGNO (x) >= FIRST_PSEUDO_REGISTER |
10274 | && REGNO (x) < reg_n_sets_max |
10275 | && REG_N_SETS (REGNO (x)) == 1 |
10276 | && !REGNO_REG_SET_P |
10277 | (DF_LR_IN (ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb), |
10278 | REGNO (x))))) |
10279 | { |
10280 | *result = rsp->last_set_sign_bit_copies; |
10281 | return NULL; |
10282 | } |
10283 | |
10284 | tem = get_last_value (x); |
10285 | if (tem != 0) |
10286 | return tem; |
10287 | |
10288 | if (nonzero_sign_valid && rsp->sign_bit_copies != 0 |
10289 | && GET_MODE_PRECISION (mode: xmode) == GET_MODE_PRECISION (mode)) |
10290 | *result = rsp->sign_bit_copies; |
10291 | |
10292 | return NULL; |
10293 | } |
10294 | |
10295 | /* Return the number of "extended" bits there are in X, when interpreted |
10296 | as a quantity in MODE whose signedness is indicated by UNSIGNEDP. For |
10297 | unsigned quantities, this is the number of high-order zero bits. |
10298 | For signed quantities, this is the number of copies of the sign bit |
10299 | minus 1. In both case, this function returns the number of "spare" |
10300 | bits. For example, if two quantities for which this function returns |
10301 | at least 1 are added, the addition is known not to overflow. |
10302 | |
10303 | This function will always return 0 unless called during combine, which |
10304 | implies that it must be called from a define_split. */ |
10305 | |
10306 | unsigned int |
10307 | extended_count (const_rtx x, machine_mode mode, bool unsignedp) |
10308 | { |
10309 | if (nonzero_sign_valid == 0) |
10310 | return 0; |
10311 | |
10312 | scalar_int_mode int_mode; |
10313 | return (unsignedp |
10314 | ? (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
10315 | && HWI_COMPUTABLE_MODE_P (mode: int_mode) |
10316 | ? (unsigned int) (GET_MODE_PRECISION (mode: int_mode) - 1 |
10317 | - floor_log2 (x: nonzero_bits (x, int_mode))) |
10318 | : 0) |
10319 | : num_sign_bit_copies (x, mode) - 1); |
10320 | } |
10321 | |
10322 | /* This function is called from `simplify_shift_const' to merge two |
10323 | outer operations. Specifically, we have already found that we need |
10324 | to perform operation *POP0 with constant *PCONST0 at the outermost |
10325 | position. We would now like to also perform OP1 with constant CONST1 |
10326 | (with *POP0 being done last). |
10327 | |
10328 | Return true if we can do the operation and update *POP0 and *PCONST0 with |
10329 | the resulting operation. *PCOMP_P is set to true if we would need to |
10330 | complement the innermost operand, otherwise it is unchanged. |
10331 | |
10332 | MODE is the mode in which the operation will be done. No bits outside |
10333 | the width of this mode matter. It is assumed that the width of this mode |
10334 | is smaller than or equal to HOST_BITS_PER_WIDE_INT. |
10335 | |
10336 | If *POP0 or OP1 are UNKNOWN, it means no operation is required. Only NEG, PLUS, |
10337 | IOR, XOR, and AND are supported. We may set *POP0 to SET if the proper |
10338 | result is simply *PCONST0. |
10339 | |
10340 | If the resulting operation cannot be expressed as one operation, we |
10341 | return false and do not change *POP0, *PCONST0, and *PCOMP_P. */ |
10342 | |
10343 | static bool |
10344 | merge_outer_ops (enum rtx_code *pop0, HOST_WIDE_INT *pconst0, |
10345 | enum rtx_code op1, HOST_WIDE_INT const1, |
10346 | machine_mode mode, bool *pcomp_p) |
10347 | { |
10348 | enum rtx_code op0 = *pop0; |
10349 | HOST_WIDE_INT const0 = *pconst0; |
10350 | |
10351 | const0 &= GET_MODE_MASK (mode); |
10352 | const1 &= GET_MODE_MASK (mode); |
10353 | |
10354 | /* If OP0 is an AND, clear unimportant bits in CONST1. */ |
10355 | if (op0 == AND) |
10356 | const1 &= const0; |
10357 | |
10358 | /* If OP0 or OP1 is UNKNOWN, this is easy. Similarly if they are the same or |
10359 | if OP0 is SET. */ |
10360 | |
10361 | if (op1 == UNKNOWN || op0 == SET) |
10362 | return true; |
10363 | |
10364 | else if (op0 == UNKNOWN) |
10365 | op0 = op1, const0 = const1; |
10366 | |
10367 | else if (op0 == op1) |
10368 | { |
10369 | switch (op0) |
10370 | { |
10371 | case AND: |
10372 | const0 &= const1; |
10373 | break; |
10374 | case IOR: |
10375 | const0 |= const1; |
10376 | break; |
10377 | case XOR: |
10378 | const0 ^= const1; |
10379 | break; |
10380 | case PLUS: |
10381 | const0 += const1; |
10382 | break; |
10383 | case NEG: |
10384 | op0 = UNKNOWN; |
10385 | break; |
10386 | default: |
10387 | break; |
10388 | } |
10389 | } |
10390 | |
10391 | /* Otherwise, if either is a PLUS or NEG, we can't do anything. */ |
10392 | else if (op0 == PLUS || op1 == PLUS || op0 == NEG || op1 == NEG) |
10393 | return false; |
10394 | |
10395 | /* If the two constants aren't the same, we can't do anything. The |
10396 | remaining six cases can all be done. */ |
10397 | else if (const0 != const1) |
10398 | return false; |
10399 | |
10400 | else |
10401 | switch (op0) |
10402 | { |
10403 | case IOR: |
10404 | if (op1 == AND) |
10405 | /* (a & b) | b == b */ |
10406 | op0 = SET; |
10407 | else /* op1 == XOR */ |
10408 | /* (a ^ b) | b == a | b */ |
10409 | {;} |
10410 | break; |
10411 | |
10412 | case XOR: |
10413 | if (op1 == AND) |
10414 | /* (a & b) ^ b == (~a) & b */ |
10415 | op0 = AND, *pcomp_p = true; |
10416 | else /* op1 == IOR */ |
10417 | /* (a | b) ^ b == a & ~b */ |
10418 | op0 = AND, const0 = ~const0; |
10419 | break; |
10420 | |
10421 | case AND: |
10422 | if (op1 == IOR) |
10423 | /* (a | b) & b == b */ |
10424 | op0 = SET; |
10425 | else /* op1 == XOR */ |
10426 | /* (a ^ b) & b) == (~a) & b */ |
10427 | *pcomp_p = true; |
10428 | break; |
10429 | default: |
10430 | break; |
10431 | } |
10432 | |
10433 | /* Check for NO-OP cases. */ |
10434 | const0 &= GET_MODE_MASK (mode); |
10435 | if (const0 == 0 |
10436 | && (op0 == IOR || op0 == XOR || op0 == PLUS)) |
10437 | op0 = UNKNOWN; |
10438 | else if (const0 == 0 && op0 == AND) |
10439 | op0 = SET; |
10440 | else if ((unsigned HOST_WIDE_INT) const0 == GET_MODE_MASK (mode) |
10441 | && op0 == AND) |
10442 | op0 = UNKNOWN; |
10443 | |
10444 | *pop0 = op0; |
10445 | |
10446 | /* ??? Slightly redundant with the above mask, but not entirely. |
10447 | Moving this above means we'd have to sign-extend the mode mask |
10448 | for the final test. */ |
10449 | if (op0 != UNKNOWN && op0 != NEG) |
10450 | *pconst0 = trunc_int_for_mode (const0, mode); |
10451 | |
10452 | return true; |
10453 | } |
10454 | |
10455 | /* A helper to simplify_shift_const_1 to determine the mode we can perform |
10456 | the shift in. The original shift operation CODE is performed on OP in |
10457 | ORIG_MODE. Return the wider mode MODE if we can perform the operation |
10458 | in that mode. Return ORIG_MODE otherwise. We can also assume that the |
10459 | result of the shift is subject to operation OUTER_CODE with operand |
10460 | OUTER_CONST. */ |
10461 | |
10462 | static scalar_int_mode |
10463 | try_widen_shift_mode (enum rtx_code code, rtx op, int count, |
10464 | scalar_int_mode orig_mode, scalar_int_mode mode, |
10465 | enum rtx_code outer_code, HOST_WIDE_INT outer_const) |
10466 | { |
10467 | gcc_assert (GET_MODE_PRECISION (mode) > GET_MODE_PRECISION (orig_mode)); |
10468 | |
10469 | /* In general we can't perform in wider mode for right shift and rotate. */ |
10470 | switch (code) |
10471 | { |
10472 | case ASHIFTRT: |
10473 | /* We can still widen if the bits brought in from the left are identical |
10474 | to the sign bit of ORIG_MODE. */ |
10475 | if (num_sign_bit_copies (op, mode) |
10476 | > (unsigned) (GET_MODE_PRECISION (mode) |
10477 | - GET_MODE_PRECISION (mode: orig_mode))) |
10478 | return mode; |
10479 | return orig_mode; |
10480 | |
10481 | case LSHIFTRT: |
10482 | /* Similarly here but with zero bits. */ |
10483 | if (HWI_COMPUTABLE_MODE_P (mode) |
10484 | && (nonzero_bits (op, mode) & ~GET_MODE_MASK (orig_mode)) == 0) |
10485 | return mode; |
10486 | |
10487 | /* We can also widen if the bits brought in will be masked off. This |
10488 | operation is performed in ORIG_MODE. */ |
10489 | if (outer_code == AND) |
10490 | { |
10491 | int care_bits = low_bitmask_len (orig_mode, outer_const); |
10492 | |
10493 | if (care_bits >= 0 |
10494 | && GET_MODE_PRECISION (mode: orig_mode) - care_bits >= count) |
10495 | return mode; |
10496 | } |
10497 | /* fall through */ |
10498 | |
10499 | case ROTATE: |
10500 | return orig_mode; |
10501 | |
10502 | case ROTATERT: |
10503 | gcc_unreachable (); |
10504 | |
10505 | default: |
10506 | return mode; |
10507 | } |
10508 | } |
10509 | |
10510 | /* Simplify a shift of VAROP by ORIG_COUNT bits. CODE says what kind |
10511 | of shift. The result of the shift is RESULT_MODE. Return NULL_RTX |
10512 | if we cannot simplify it. Otherwise, return a simplified value. |
10513 | |
10514 | The shift is normally computed in the widest mode we find in VAROP, as |
10515 | long as it isn't a different number of words than RESULT_MODE. Exceptions |
10516 | are ASHIFTRT and ROTATE, which are always done in their original mode. */ |
10517 | |
10518 | static rtx |
10519 | simplify_shift_const_1 (enum rtx_code code, machine_mode result_mode, |
10520 | rtx varop, int orig_count) |
10521 | { |
10522 | enum rtx_code orig_code = code; |
10523 | rtx orig_varop = varop; |
10524 | int count, log2; |
10525 | machine_mode mode = result_mode; |
10526 | machine_mode shift_mode; |
10527 | scalar_int_mode tmode, inner_mode, int_mode, int_varop_mode, int_result_mode; |
10528 | /* We form (outer_op (code varop count) (outer_const)). */ |
10529 | enum rtx_code outer_op = UNKNOWN; |
10530 | HOST_WIDE_INT outer_const = 0; |
10531 | bool complement_p = false; |
10532 | rtx new_rtx, x; |
10533 | |
10534 | /* Make sure and truncate the "natural" shift on the way in. We don't |
10535 | want to do this inside the loop as it makes it more difficult to |
10536 | combine shifts. */ |
10537 | if (SHIFT_COUNT_TRUNCATED) |
10538 | orig_count &= GET_MODE_UNIT_BITSIZE (mode) - 1; |
10539 | |
10540 | /* If we were given an invalid count, don't do anything except exactly |
10541 | what was requested. */ |
10542 | |
10543 | if (orig_count < 0 || orig_count >= (int) GET_MODE_UNIT_PRECISION (mode)) |
10544 | return NULL_RTX; |
10545 | |
10546 | count = orig_count; |
10547 | |
10548 | /* Unless one of the branches of the `if' in this loop does a `continue', |
10549 | we will `break' the loop after the `if'. */ |
10550 | |
10551 | while (count != 0) |
10552 | { |
10553 | /* If we have an operand of (clobber (const_int 0)), fail. */ |
10554 | if (GET_CODE (varop) == CLOBBER) |
10555 | return NULL_RTX; |
10556 | |
10557 | /* Convert ROTATERT to ROTATE. */ |
10558 | if (code == ROTATERT) |
10559 | { |
10560 | unsigned int bitsize = GET_MODE_UNIT_PRECISION (result_mode); |
10561 | code = ROTATE; |
10562 | count = bitsize - count; |
10563 | } |
10564 | |
10565 | shift_mode = result_mode; |
10566 | if (shift_mode != mode) |
10567 | { |
10568 | /* We only change the modes of scalar shifts. */ |
10569 | int_mode = as_a <scalar_int_mode> (m: mode); |
10570 | int_result_mode = as_a <scalar_int_mode> (m: result_mode); |
10571 | shift_mode = try_widen_shift_mode (code, op: varop, count, |
10572 | orig_mode: int_result_mode, mode: int_mode, |
10573 | outer_code: outer_op, outer_const); |
10574 | } |
10575 | |
10576 | scalar_int_mode shift_unit_mode |
10577 | = as_a <scalar_int_mode> (GET_MODE_INNER (shift_mode)); |
10578 | |
10579 | /* Handle cases where the count is greater than the size of the mode |
10580 | minus 1. For ASHIFT, use the size minus one as the count (this can |
10581 | occur when simplifying (lshiftrt (ashiftrt ..))). For rotates, |
10582 | take the count modulo the size. For other shifts, the result is |
10583 | zero. |
10584 | |
10585 | Since these shifts are being produced by the compiler by combining |
10586 | multiple operations, each of which are defined, we know what the |
10587 | result is supposed to be. */ |
10588 | |
10589 | if (count > (GET_MODE_PRECISION (mode: shift_unit_mode) - 1)) |
10590 | { |
10591 | if (code == ASHIFTRT) |
10592 | count = GET_MODE_PRECISION (mode: shift_unit_mode) - 1; |
10593 | else if (code == ROTATE || code == ROTATERT) |
10594 | count %= GET_MODE_PRECISION (mode: shift_unit_mode); |
10595 | else |
10596 | { |
10597 | /* We can't simply return zero because there may be an |
10598 | outer op. */ |
10599 | varop = const0_rtx; |
10600 | count = 0; |
10601 | break; |
10602 | } |
10603 | } |
10604 | |
10605 | /* If we discovered we had to complement VAROP, leave. Making a NOT |
10606 | here would cause an infinite loop. */ |
10607 | if (complement_p) |
10608 | break; |
10609 | |
10610 | if (shift_mode == shift_unit_mode) |
10611 | { |
10612 | /* An arithmetic right shift of a quantity known to be -1 or 0 |
10613 | is a no-op. */ |
10614 | if (code == ASHIFTRT |
10615 | && (num_sign_bit_copies (varop, shift_unit_mode) |
10616 | == GET_MODE_PRECISION (mode: shift_unit_mode))) |
10617 | { |
10618 | count = 0; |
10619 | break; |
10620 | } |
10621 | |
10622 | /* If we are doing an arithmetic right shift and discarding all but |
10623 | the sign bit copies, this is equivalent to doing a shift by the |
10624 | bitsize minus one. Convert it into that shift because it will |
10625 | often allow other simplifications. */ |
10626 | |
10627 | if (code == ASHIFTRT |
10628 | && (count + num_sign_bit_copies (varop, shift_unit_mode) |
10629 | >= GET_MODE_PRECISION (mode: shift_unit_mode))) |
10630 | count = GET_MODE_PRECISION (mode: shift_unit_mode) - 1; |
10631 | |
10632 | /* We simplify the tests below and elsewhere by converting |
10633 | ASHIFTRT to LSHIFTRT if we know the sign bit is clear. |
10634 | `make_compound_operation' will convert it to an ASHIFTRT for |
10635 | those machines (such as VAX) that don't have an LSHIFTRT. */ |
10636 | if (code == ASHIFTRT |
10637 | && HWI_COMPUTABLE_MODE_P (mode: shift_unit_mode) |
10638 | && val_signbit_known_clear_p (shift_unit_mode, |
10639 | nonzero_bits (varop, |
10640 | shift_unit_mode))) |
10641 | code = LSHIFTRT; |
10642 | |
10643 | if (((code == LSHIFTRT |
10644 | && HWI_COMPUTABLE_MODE_P (mode: shift_unit_mode) |
10645 | && !(nonzero_bits (varop, shift_unit_mode) >> count)) |
10646 | || (code == ASHIFT |
10647 | && HWI_COMPUTABLE_MODE_P (mode: shift_unit_mode) |
10648 | && !((nonzero_bits (varop, shift_unit_mode) << count) |
10649 | & GET_MODE_MASK (shift_unit_mode)))) |
10650 | && !side_effects_p (varop)) |
10651 | varop = const0_rtx; |
10652 | } |
10653 | |
10654 | switch (GET_CODE (varop)) |
10655 | { |
10656 | case SIGN_EXTEND: |
10657 | case ZERO_EXTEND: |
10658 | case SIGN_EXTRACT: |
10659 | case ZERO_EXTRACT: |
10660 | new_rtx = expand_compound_operation (x: varop); |
10661 | if (new_rtx != varop) |
10662 | { |
10663 | varop = new_rtx; |
10664 | continue; |
10665 | } |
10666 | break; |
10667 | |
10668 | case MEM: |
10669 | /* The following rules apply only to scalars. */ |
10670 | if (shift_mode != shift_unit_mode) |
10671 | break; |
10672 | int_mode = as_a <scalar_int_mode> (m: mode); |
10673 | |
10674 | /* If we have (xshiftrt (mem ...) C) and C is MODE_WIDTH |
10675 | minus the width of a smaller mode, we can do this with a |
10676 | SIGN_EXTEND or ZERO_EXTEND from the narrower memory location. */ |
10677 | if ((code == ASHIFTRT || code == LSHIFTRT) |
10678 | && ! mode_dependent_address_p (XEXP (varop, 0), |
10679 | MEM_ADDR_SPACE (varop)) |
10680 | && ! MEM_VOLATILE_P (varop) |
10681 | && (int_mode_for_size (size: GET_MODE_BITSIZE (mode: int_mode) - count, limit: 1) |
10682 | .exists (mode: &tmode))) |
10683 | { |
10684 | new_rtx = adjust_address_nv (varop, tmode, |
10685 | BYTES_BIG_ENDIAN ? 0 |
10686 | : count / BITS_PER_UNIT); |
10687 | |
10688 | varop = gen_rtx_fmt_e (code == ASHIFTRT ? SIGN_EXTEND |
10689 | : ZERO_EXTEND, int_mode, new_rtx); |
10690 | count = 0; |
10691 | continue; |
10692 | } |
10693 | break; |
10694 | |
10695 | case SUBREG: |
10696 | /* The following rules apply only to scalars. */ |
10697 | if (shift_mode != shift_unit_mode) |
10698 | break; |
10699 | int_mode = as_a <scalar_int_mode> (m: mode); |
10700 | int_varop_mode = as_a <scalar_int_mode> (GET_MODE (varop)); |
10701 | |
10702 | /* If VAROP is a SUBREG, strip it as long as the inner operand has |
10703 | the same number of words as what we've seen so far. Then store |
10704 | the widest mode in MODE. */ |
10705 | if (subreg_lowpart_p (varop) |
10706 | && is_int_mode (GET_MODE (SUBREG_REG (varop)), int_mode: &inner_mode) |
10707 | && GET_MODE_SIZE (mode: inner_mode) > GET_MODE_SIZE (mode: int_varop_mode) |
10708 | && (CEIL (GET_MODE_SIZE (inner_mode), UNITS_PER_WORD) |
10709 | == CEIL (GET_MODE_SIZE (int_mode), UNITS_PER_WORD)) |
10710 | && GET_MODE_CLASS (int_varop_mode) == MODE_INT) |
10711 | { |
10712 | varop = SUBREG_REG (varop); |
10713 | if (GET_MODE_SIZE (mode: inner_mode) > GET_MODE_SIZE (mode: int_mode)) |
10714 | mode = inner_mode; |
10715 | continue; |
10716 | } |
10717 | break; |
10718 | |
10719 | case MULT: |
10720 | /* Some machines use MULT instead of ASHIFT because MULT |
10721 | is cheaper. But it is still better on those machines to |
10722 | merge two shifts into one. */ |
10723 | if (CONST_INT_P (XEXP (varop, 1)) |
10724 | && (log2 = exact_log2 (UINTVAL (XEXP (varop, 1)))) >= 0) |
10725 | { |
10726 | rtx log2_rtx = gen_int_shift_amount (GET_MODE (varop), log2); |
10727 | varop = simplify_gen_binary (code: ASHIFT, GET_MODE (varop), |
10728 | XEXP (varop, 0), op1: log2_rtx); |
10729 | continue; |
10730 | } |
10731 | break; |
10732 | |
10733 | case UDIV: |
10734 | /* Similar, for when divides are cheaper. */ |
10735 | if (CONST_INT_P (XEXP (varop, 1)) |
10736 | && (log2 = exact_log2 (UINTVAL (XEXP (varop, 1)))) >= 0) |
10737 | { |
10738 | rtx log2_rtx = gen_int_shift_amount (GET_MODE (varop), log2); |
10739 | varop = simplify_gen_binary (code: LSHIFTRT, GET_MODE (varop), |
10740 | XEXP (varop, 0), op1: log2_rtx); |
10741 | continue; |
10742 | } |
10743 | break; |
10744 | |
10745 | case ASHIFTRT: |
10746 | /* If we are extracting just the sign bit of an arithmetic |
10747 | right shift, that shift is not needed. However, the sign |
10748 | bit of a wider mode may be different from what would be |
10749 | interpreted as the sign bit in a narrower mode, so, if |
10750 | the result is narrower, don't discard the shift. */ |
10751 | if (code == LSHIFTRT |
10752 | && count == (GET_MODE_UNIT_BITSIZE (result_mode) - 1) |
10753 | && (GET_MODE_UNIT_BITSIZE (result_mode) |
10754 | >= GET_MODE_UNIT_BITSIZE (GET_MODE (varop)))) |
10755 | { |
10756 | varop = XEXP (varop, 0); |
10757 | continue; |
10758 | } |
10759 | |
10760 | /* fall through */ |
10761 | |
10762 | case LSHIFTRT: |
10763 | case ASHIFT: |
10764 | case ROTATE: |
10765 | /* The following rules apply only to scalars. */ |
10766 | if (shift_mode != shift_unit_mode) |
10767 | break; |
10768 | int_mode = as_a <scalar_int_mode> (m: mode); |
10769 | int_varop_mode = as_a <scalar_int_mode> (GET_MODE (varop)); |
10770 | int_result_mode = as_a <scalar_int_mode> (m: result_mode); |
10771 | |
10772 | /* Here we have two nested shifts. The result is usually the |
10773 | AND of a new shift with a mask. We compute the result below. */ |
10774 | if (CONST_INT_P (XEXP (varop, 1)) |
10775 | && INTVAL (XEXP (varop, 1)) >= 0 |
10776 | && INTVAL (XEXP (varop, 1)) < GET_MODE_PRECISION (mode: int_varop_mode) |
10777 | && HWI_COMPUTABLE_MODE_P (mode: int_result_mode) |
10778 | && HWI_COMPUTABLE_MODE_P (mode: int_mode)) |
10779 | { |
10780 | enum rtx_code first_code = GET_CODE (varop); |
10781 | unsigned int first_count = INTVAL (XEXP (varop, 1)); |
10782 | unsigned HOST_WIDE_INT mask; |
10783 | rtx mask_rtx; |
10784 | |
10785 | /* We have one common special case. We can't do any merging if |
10786 | the inner code is an ASHIFTRT of a smaller mode. However, if |
10787 | we have (ashift:M1 (subreg:M1 (ashiftrt:M2 FOO C1) 0) C2) |
10788 | with C2 == GET_MODE_BITSIZE (M1) - GET_MODE_BITSIZE (M2), |
10789 | we can convert it to |
10790 | (ashiftrt:M1 (ashift:M1 (and:M1 (subreg:M1 FOO 0) C3) C2) C1). |
10791 | This simplifies certain SIGN_EXTEND operations. */ |
10792 | if (code == ASHIFT && first_code == ASHIFTRT |
10793 | && count == (GET_MODE_PRECISION (mode: int_result_mode) |
10794 | - GET_MODE_PRECISION (mode: int_varop_mode))) |
10795 | { |
10796 | /* C3 has the low-order C1 bits zero. */ |
10797 | |
10798 | mask = GET_MODE_MASK (int_mode) |
10799 | & ~((HOST_WIDE_INT_1U << first_count) - 1); |
10800 | |
10801 | varop = simplify_and_const_int (NULL_RTX, mode: int_result_mode, |
10802 | XEXP (varop, 0), constop: mask); |
10803 | varop = simplify_shift_const (NULL_RTX, ASHIFT, |
10804 | int_result_mode, varop, count); |
10805 | count = first_count; |
10806 | code = ASHIFTRT; |
10807 | continue; |
10808 | } |
10809 | |
10810 | /* If this was (ashiftrt (ashift foo C1) C2) and FOO has more |
10811 | than C1 high-order bits equal to the sign bit, we can convert |
10812 | this to either an ASHIFT or an ASHIFTRT depending on the |
10813 | two counts. |
10814 | |
10815 | We cannot do this if VAROP's mode is not SHIFT_UNIT_MODE. */ |
10816 | |
10817 | if (code == ASHIFTRT && first_code == ASHIFT |
10818 | && int_varop_mode == shift_unit_mode |
10819 | && (num_sign_bit_copies (XEXP (varop, 0), shift_unit_mode) |
10820 | > first_count)) |
10821 | { |
10822 | varop = XEXP (varop, 0); |
10823 | count -= first_count; |
10824 | if (count < 0) |
10825 | { |
10826 | count = -count; |
10827 | code = ASHIFT; |
10828 | } |
10829 | |
10830 | continue; |
10831 | } |
10832 | |
10833 | /* There are some cases we can't do. If CODE is ASHIFTRT, |
10834 | we can only do this if FIRST_CODE is also ASHIFTRT. |
10835 | |
10836 | We can't do the case when CODE is ROTATE and FIRST_CODE is |
10837 | ASHIFTRT. |
10838 | |
10839 | If the mode of this shift is not the mode of the outer shift, |
10840 | we can't do this if either shift is a right shift or ROTATE. |
10841 | |
10842 | Finally, we can't do any of these if the mode is too wide |
10843 | unless the codes are the same. |
10844 | |
10845 | Handle the case where the shift codes are the same |
10846 | first. */ |
10847 | |
10848 | if (code == first_code) |
10849 | { |
10850 | if (int_varop_mode != int_result_mode |
10851 | && (code == ASHIFTRT || code == LSHIFTRT |
10852 | || code == ROTATE)) |
10853 | break; |
10854 | |
10855 | count += first_count; |
10856 | varop = XEXP (varop, 0); |
10857 | continue; |
10858 | } |
10859 | |
10860 | if (code == ASHIFTRT |
10861 | || (code == ROTATE && first_code == ASHIFTRT) |
10862 | || GET_MODE_PRECISION (mode: int_mode) > HOST_BITS_PER_WIDE_INT |
10863 | || (int_varop_mode != int_result_mode |
10864 | && (first_code == ASHIFTRT || first_code == LSHIFTRT |
10865 | || first_code == ROTATE |
10866 | || code == ROTATE))) |
10867 | break; |
10868 | |
10869 | /* To compute the mask to apply after the shift, shift the |
10870 | nonzero bits of the inner shift the same way the |
10871 | outer shift will. */ |
10872 | |
10873 | mask_rtx = gen_int_mode (nonzero_bits (varop, int_varop_mode), |
10874 | int_result_mode); |
10875 | rtx count_rtx = gen_int_shift_amount (int_result_mode, count); |
10876 | mask_rtx |
10877 | = simplify_const_binary_operation (code, int_result_mode, |
10878 | mask_rtx, count_rtx); |
10879 | |
10880 | /* Give up if we can't compute an outer operation to use. */ |
10881 | if (mask_rtx == 0 |
10882 | || !CONST_INT_P (mask_rtx) |
10883 | || ! merge_outer_ops (pop0: &outer_op, pconst0: &outer_const, op1: AND, |
10884 | INTVAL (mask_rtx), |
10885 | mode: int_result_mode, pcomp_p: &complement_p)) |
10886 | break; |
10887 | |
10888 | /* If the shifts are in the same direction, we add the |
10889 | counts. Otherwise, we subtract them. */ |
10890 | if ((code == ASHIFTRT || code == LSHIFTRT) |
10891 | == (first_code == ASHIFTRT || first_code == LSHIFTRT)) |
10892 | count += first_count; |
10893 | else |
10894 | count -= first_count; |
10895 | |
10896 | /* If COUNT is positive, the new shift is usually CODE, |
10897 | except for the two exceptions below, in which case it is |
10898 | FIRST_CODE. If the count is negative, FIRST_CODE should |
10899 | always be used */ |
10900 | if (count > 0 |
10901 | && ((first_code == ROTATE && code == ASHIFT) |
10902 | || (first_code == ASHIFTRT && code == LSHIFTRT))) |
10903 | code = first_code; |
10904 | else if (count < 0) |
10905 | code = first_code, count = -count; |
10906 | |
10907 | varop = XEXP (varop, 0); |
10908 | continue; |
10909 | } |
10910 | |
10911 | /* If we have (A << B << C) for any shift, we can convert this to |
10912 | (A << C << B). This wins if A is a constant. Only try this if |
10913 | B is not a constant. */ |
10914 | |
10915 | else if (GET_CODE (varop) == code |
10916 | && CONST_INT_P (XEXP (varop, 0)) |
10917 | && !CONST_INT_P (XEXP (varop, 1))) |
10918 | { |
10919 | /* For ((unsigned) (cstULL >> count)) >> cst2 we have to make |
10920 | sure the result will be masked. See PR70222. */ |
10921 | if (code == LSHIFTRT |
10922 | && int_mode != int_result_mode |
10923 | && !merge_outer_ops (pop0: &outer_op, pconst0: &outer_const, op1: AND, |
10924 | GET_MODE_MASK (int_result_mode) |
10925 | >> orig_count, mode: int_result_mode, |
10926 | pcomp_p: &complement_p)) |
10927 | break; |
10928 | /* For ((int) (cstLL >> count)) >> cst2 just give up. Queuing |
10929 | up outer sign extension (often left and right shift) is |
10930 | hardly more efficient than the original. See PR70429. |
10931 | Similarly punt for rotates with different modes. |
10932 | See PR97386. */ |
10933 | if ((code == ASHIFTRT || code == ROTATE) |
10934 | && int_mode != int_result_mode) |
10935 | break; |
10936 | |
10937 | rtx count_rtx = gen_int_shift_amount (int_result_mode, count); |
10938 | rtx new_rtx = simplify_const_binary_operation (code, int_mode, |
10939 | XEXP (varop, 0), |
10940 | count_rtx); |
10941 | varop = gen_rtx_fmt_ee (code, int_mode, new_rtx, XEXP (varop, 1)); |
10942 | count = 0; |
10943 | continue; |
10944 | } |
10945 | break; |
10946 | |
10947 | case NOT: |
10948 | /* The following rules apply only to scalars. */ |
10949 | if (shift_mode != shift_unit_mode) |
10950 | break; |
10951 | |
10952 | /* Make this fit the case below. */ |
10953 | varop = gen_rtx_XOR (mode, XEXP (varop, 0), constm1_rtx); |
10954 | continue; |
10955 | |
10956 | case IOR: |
10957 | case AND: |
10958 | case XOR: |
10959 | /* The following rules apply only to scalars. */ |
10960 | if (shift_mode != shift_unit_mode) |
10961 | break; |
10962 | int_varop_mode = as_a <scalar_int_mode> (GET_MODE (varop)); |
10963 | int_result_mode = as_a <scalar_int_mode> (m: result_mode); |
10964 | |
10965 | /* If we have (xshiftrt (ior (plus X (const_int -1)) X) C) |
10966 | with C the size of VAROP - 1 and the shift is logical if |
10967 | STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1, |
10968 | we have an (le X 0) operation. If we have an arithmetic shift |
10969 | and STORE_FLAG_VALUE is 1 or we have a logical shift with |
10970 | STORE_FLAG_VALUE of -1, we have a (neg (le X 0)) operation. */ |
10971 | |
10972 | if (GET_CODE (varop) == IOR && GET_CODE (XEXP (varop, 0)) == PLUS |
10973 | && XEXP (XEXP (varop, 0), 1) == constm1_rtx |
10974 | && (STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
10975 | && (code == LSHIFTRT || code == ASHIFTRT) |
10976 | && count == (GET_MODE_PRECISION (mode: int_varop_mode) - 1) |
10977 | && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1))) |
10978 | { |
10979 | count = 0; |
10980 | varop = gen_rtx_LE (int_varop_mode, XEXP (varop, 1), |
10981 | const0_rtx); |
10982 | |
10983 | if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT) |
10984 | varop = gen_rtx_NEG (int_varop_mode, varop); |
10985 | |
10986 | continue; |
10987 | } |
10988 | |
10989 | /* If we have (shift (logical)), move the logical to the outside |
10990 | to allow it to possibly combine with another logical and the |
10991 | shift to combine with another shift. This also canonicalizes to |
10992 | what a ZERO_EXTRACT looks like. Also, some machines have |
10993 | (and (shift)) insns. */ |
10994 | |
10995 | if (CONST_INT_P (XEXP (varop, 1)) |
10996 | /* We can't do this if we have (ashiftrt (xor)) and the |
10997 | constant has its sign bit set in shift_unit_mode with |
10998 | shift_unit_mode wider than result_mode. */ |
10999 | && !(code == ASHIFTRT && GET_CODE (varop) == XOR |
11000 | && int_result_mode != shift_unit_mode |
11001 | && trunc_int_for_mode (INTVAL (XEXP (varop, 1)), |
11002 | shift_unit_mode) < 0) |
11003 | && (new_rtx = simplify_const_binary_operation |
11004 | (code, int_result_mode, |
11005 | gen_int_mode (INTVAL (XEXP (varop, 1)), int_result_mode), |
11006 | gen_int_shift_amount (int_result_mode, count))) != 0 |
11007 | && CONST_INT_P (new_rtx) |
11008 | && merge_outer_ops (pop0: &outer_op, pconst0: &outer_const, GET_CODE (varop), |
11009 | INTVAL (new_rtx), mode: int_result_mode, |
11010 | pcomp_p: &complement_p)) |
11011 | { |
11012 | varop = XEXP (varop, 0); |
11013 | continue; |
11014 | } |
11015 | |
11016 | /* If we can't do that, try to simplify the shift in each arm of the |
11017 | logical expression, make a new logical expression, and apply |
11018 | the inverse distributive law. This also can't be done for |
11019 | (ashiftrt (xor)) where we've widened the shift and the constant |
11020 | changes the sign bit. */ |
11021 | if (CONST_INT_P (XEXP (varop, 1)) |
11022 | && !(code == ASHIFTRT && GET_CODE (varop) == XOR |
11023 | && int_result_mode != shift_unit_mode |
11024 | && trunc_int_for_mode (INTVAL (XEXP (varop, 1)), |
11025 | shift_unit_mode) < 0)) |
11026 | { |
11027 | rtx lhs = simplify_shift_const (NULL_RTX, code, shift_unit_mode, |
11028 | XEXP (varop, 0), count); |
11029 | rtx rhs = simplify_shift_const (NULL_RTX, code, shift_unit_mode, |
11030 | XEXP (varop, 1), count); |
11031 | |
11032 | varop = simplify_gen_binary (GET_CODE (varop), mode: shift_unit_mode, |
11033 | op0: lhs, op1: rhs); |
11034 | varop = apply_distributive_law (x: varop); |
11035 | |
11036 | count = 0; |
11037 | continue; |
11038 | } |
11039 | break; |
11040 | |
11041 | case EQ: |
11042 | /* The following rules apply only to scalars. */ |
11043 | if (shift_mode != shift_unit_mode) |
11044 | break; |
11045 | int_result_mode = as_a <scalar_int_mode> (m: result_mode); |
11046 | |
11047 | /* Convert (lshiftrt (eq FOO 0) C) to (xor FOO 1) if STORE_FLAG_VALUE |
11048 | says that the sign bit can be tested, FOO has mode MODE, C is |
11049 | GET_MODE_PRECISION (MODE) - 1, and FOO has only its low-order bit |
11050 | that may be nonzero. */ |
11051 | if (code == LSHIFTRT |
11052 | && XEXP (varop, 1) == const0_rtx |
11053 | && GET_MODE (XEXP (varop, 0)) == int_result_mode |
11054 | && count == (GET_MODE_PRECISION (mode: int_result_mode) - 1) |
11055 | && HWI_COMPUTABLE_MODE_P (mode: int_result_mode) |
11056 | && STORE_FLAG_VALUE == -1 |
11057 | && nonzero_bits (XEXP (varop, 0), int_result_mode) == 1 |
11058 | && merge_outer_ops (pop0: &outer_op, pconst0: &outer_const, op1: XOR, const1: 1, |
11059 | mode: int_result_mode, pcomp_p: &complement_p)) |
11060 | { |
11061 | varop = XEXP (varop, 0); |
11062 | count = 0; |
11063 | continue; |
11064 | } |
11065 | break; |
11066 | |
11067 | case NEG: |
11068 | /* The following rules apply only to scalars. */ |
11069 | if (shift_mode != shift_unit_mode) |
11070 | break; |
11071 | int_result_mode = as_a <scalar_int_mode> (m: result_mode); |
11072 | |
11073 | /* (lshiftrt (neg A) C) where A is either 0 or 1 and C is one less |
11074 | than the number of bits in the mode is equivalent to A. */ |
11075 | if (code == LSHIFTRT |
11076 | && count == (GET_MODE_PRECISION (mode: int_result_mode) - 1) |
11077 | && nonzero_bits (XEXP (varop, 0), int_result_mode) == 1) |
11078 | { |
11079 | varop = XEXP (varop, 0); |
11080 | count = 0; |
11081 | continue; |
11082 | } |
11083 | |
11084 | /* NEG commutes with ASHIFT since it is multiplication. Move the |
11085 | NEG outside to allow shifts to combine. */ |
11086 | if (code == ASHIFT |
11087 | && merge_outer_ops (pop0: &outer_op, pconst0: &outer_const, op1: NEG, const1: 0, |
11088 | mode: int_result_mode, pcomp_p: &complement_p)) |
11089 | { |
11090 | varop = XEXP (varop, 0); |
11091 | continue; |
11092 | } |
11093 | break; |
11094 | |
11095 | case PLUS: |
11096 | /* The following rules apply only to scalars. */ |
11097 | if (shift_mode != shift_unit_mode) |
11098 | break; |
11099 | int_result_mode = as_a <scalar_int_mode> (m: result_mode); |
11100 | |
11101 | /* (lshiftrt (plus A -1) C) where A is either 0 or 1 and C |
11102 | is one less than the number of bits in the mode is |
11103 | equivalent to (xor A 1). */ |
11104 | if (code == LSHIFTRT |
11105 | && count == (GET_MODE_PRECISION (mode: int_result_mode) - 1) |
11106 | && XEXP (varop, 1) == constm1_rtx |
11107 | && nonzero_bits (XEXP (varop, 0), int_result_mode) == 1 |
11108 | && merge_outer_ops (pop0: &outer_op, pconst0: &outer_const, op1: XOR, const1: 1, |
11109 | mode: int_result_mode, pcomp_p: &complement_p)) |
11110 | { |
11111 | count = 0; |
11112 | varop = XEXP (varop, 0); |
11113 | continue; |
11114 | } |
11115 | |
11116 | /* If we have (xshiftrt (plus FOO BAR) C), and the only bits |
11117 | that might be nonzero in BAR are those being shifted out and those |
11118 | bits are known zero in FOO, we can replace the PLUS with FOO. |
11119 | Similarly in the other operand order. This code occurs when |
11120 | we are computing the size of a variable-size array. */ |
11121 | |
11122 | if ((code == ASHIFTRT || code == LSHIFTRT) |
11123 | && count < HOST_BITS_PER_WIDE_INT |
11124 | && nonzero_bits (XEXP (varop, 1), int_result_mode) >> count == 0 |
11125 | && (nonzero_bits (XEXP (varop, 1), int_result_mode) |
11126 | & nonzero_bits (XEXP (varop, 0), int_result_mode)) == 0) |
11127 | { |
11128 | varop = XEXP (varop, 0); |
11129 | continue; |
11130 | } |
11131 | else if ((code == ASHIFTRT || code == LSHIFTRT) |
11132 | && count < HOST_BITS_PER_WIDE_INT |
11133 | && HWI_COMPUTABLE_MODE_P (mode: int_result_mode) |
11134 | && (nonzero_bits (XEXP (varop, 0), int_result_mode) |
11135 | >> count) == 0 |
11136 | && (nonzero_bits (XEXP (varop, 0), int_result_mode) |
11137 | & nonzero_bits (XEXP (varop, 1), int_result_mode)) == 0) |
11138 | { |
11139 | varop = XEXP (varop, 1); |
11140 | continue; |
11141 | } |
11142 | |
11143 | /* (ashift (plus foo C) N) is (plus (ashift foo N) C'). */ |
11144 | if (code == ASHIFT |
11145 | && CONST_INT_P (XEXP (varop, 1)) |
11146 | && (new_rtx = simplify_const_binary_operation |
11147 | (ASHIFT, int_result_mode, |
11148 | gen_int_mode (INTVAL (XEXP (varop, 1)), int_result_mode), |
11149 | gen_int_shift_amount (int_result_mode, count))) != 0 |
11150 | && CONST_INT_P (new_rtx) |
11151 | && merge_outer_ops (pop0: &outer_op, pconst0: &outer_const, op1: PLUS, |
11152 | INTVAL (new_rtx), mode: int_result_mode, |
11153 | pcomp_p: &complement_p)) |
11154 | { |
11155 | varop = XEXP (varop, 0); |
11156 | continue; |
11157 | } |
11158 | |
11159 | /* Check for 'PLUS signbit', which is the canonical form of 'XOR |
11160 | signbit', and attempt to change the PLUS to an XOR and move it to |
11161 | the outer operation as is done above in the AND/IOR/XOR case |
11162 | leg for shift(logical). See details in logical handling above |
11163 | for reasoning in doing so. */ |
11164 | if (code == LSHIFTRT |
11165 | && CONST_INT_P (XEXP (varop, 1)) |
11166 | && mode_signbit_p (int_result_mode, XEXP (varop, 1)) |
11167 | && (new_rtx = simplify_const_binary_operation |
11168 | (code, int_result_mode, |
11169 | gen_int_mode (INTVAL (XEXP (varop, 1)), int_result_mode), |
11170 | gen_int_shift_amount (int_result_mode, count))) != 0 |
11171 | && CONST_INT_P (new_rtx) |
11172 | && merge_outer_ops (pop0: &outer_op, pconst0: &outer_const, op1: XOR, |
11173 | INTVAL (new_rtx), mode: int_result_mode, |
11174 | pcomp_p: &complement_p)) |
11175 | { |
11176 | varop = XEXP (varop, 0); |
11177 | continue; |
11178 | } |
11179 | |
11180 | break; |
11181 | |
11182 | case MINUS: |
11183 | /* The following rules apply only to scalars. */ |
11184 | if (shift_mode != shift_unit_mode) |
11185 | break; |
11186 | int_varop_mode = as_a <scalar_int_mode> (GET_MODE (varop)); |
11187 | |
11188 | /* If we have (xshiftrt (minus (ashiftrt X C)) X) C) |
11189 | with C the size of VAROP - 1 and the shift is logical if |
11190 | STORE_FLAG_VALUE is 1 and arithmetic if STORE_FLAG_VALUE is -1, |
11191 | we have a (gt X 0) operation. If the shift is arithmetic with |
11192 | STORE_FLAG_VALUE of 1 or logical with STORE_FLAG_VALUE == -1, |
11193 | we have a (neg (gt X 0)) operation. */ |
11194 | |
11195 | if ((STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1) |
11196 | && GET_CODE (XEXP (varop, 0)) == ASHIFTRT |
11197 | && count == (GET_MODE_PRECISION (mode: int_varop_mode) - 1) |
11198 | && (code == LSHIFTRT || code == ASHIFTRT) |
11199 | && CONST_INT_P (XEXP (XEXP (varop, 0), 1)) |
11200 | && INTVAL (XEXP (XEXP (varop, 0), 1)) == count |
11201 | && rtx_equal_p (XEXP (XEXP (varop, 0), 0), XEXP (varop, 1))) |
11202 | { |
11203 | count = 0; |
11204 | varop = gen_rtx_GT (int_varop_mode, XEXP (varop, 1), |
11205 | const0_rtx); |
11206 | |
11207 | if (STORE_FLAG_VALUE == 1 ? code == ASHIFTRT : code == LSHIFTRT) |
11208 | varop = gen_rtx_NEG (int_varop_mode, varop); |
11209 | |
11210 | continue; |
11211 | } |
11212 | break; |
11213 | |
11214 | case TRUNCATE: |
11215 | /* Change (lshiftrt (truncate (lshiftrt))) to (truncate (lshiftrt)) |
11216 | if the truncate does not affect the value. */ |
11217 | if (code == LSHIFTRT |
11218 | && GET_CODE (XEXP (varop, 0)) == LSHIFTRT |
11219 | && CONST_INT_P (XEXP (XEXP (varop, 0), 1)) |
11220 | && (INTVAL (XEXP (XEXP (varop, 0), 1)) |
11221 | >= (GET_MODE_UNIT_PRECISION (GET_MODE (XEXP (varop, 0))) |
11222 | - GET_MODE_UNIT_PRECISION (GET_MODE (varop))))) |
11223 | { |
11224 | rtx varop_inner = XEXP (varop, 0); |
11225 | int new_count = count + INTVAL (XEXP (varop_inner, 1)); |
11226 | rtx new_count_rtx = gen_int_shift_amount (GET_MODE (varop_inner), |
11227 | new_count); |
11228 | varop_inner = gen_rtx_LSHIFTRT (GET_MODE (varop_inner), |
11229 | XEXP (varop_inner, 0), |
11230 | new_count_rtx); |
11231 | varop = gen_rtx_TRUNCATE (GET_MODE (varop), varop_inner); |
11232 | count = 0; |
11233 | continue; |
11234 | } |
11235 | break; |
11236 | |
11237 | default: |
11238 | break; |
11239 | } |
11240 | |
11241 | break; |
11242 | } |
11243 | |
11244 | shift_mode = result_mode; |
11245 | if (shift_mode != mode) |
11246 | { |
11247 | /* We only change the modes of scalar shifts. */ |
11248 | int_mode = as_a <scalar_int_mode> (m: mode); |
11249 | int_result_mode = as_a <scalar_int_mode> (m: result_mode); |
11250 | shift_mode = try_widen_shift_mode (code, op: varop, count, orig_mode: int_result_mode, |
11251 | mode: int_mode, outer_code: outer_op, outer_const); |
11252 | } |
11253 | |
11254 | /* We have now finished analyzing the shift. The result should be |
11255 | a shift of type CODE with SHIFT_MODE shifting VAROP COUNT places. If |
11256 | OUTER_OP is non-UNKNOWN, it is an operation that needs to be applied |
11257 | to the result of the shift. OUTER_CONST is the relevant constant, |
11258 | but we must turn off all bits turned off in the shift. */ |
11259 | |
11260 | if (outer_op == UNKNOWN |
11261 | && orig_code == code && orig_count == count |
11262 | && varop == orig_varop |
11263 | && shift_mode == GET_MODE (varop)) |
11264 | return NULL_RTX; |
11265 | |
11266 | /* Make a SUBREG if necessary. If we can't make it, fail. */ |
11267 | varop = gen_lowpart (shift_mode, varop); |
11268 | if (varop == NULL_RTX || GET_CODE (varop) == CLOBBER) |
11269 | return NULL_RTX; |
11270 | |
11271 | /* If we have an outer operation and we just made a shift, it is |
11272 | possible that we could have simplified the shift were it not |
11273 | for the outer operation. So try to do the simplification |
11274 | recursively. */ |
11275 | |
11276 | if (outer_op != UNKNOWN) |
11277 | x = simplify_shift_const_1 (code, result_mode: shift_mode, varop, orig_count: count); |
11278 | else |
11279 | x = NULL_RTX; |
11280 | |
11281 | if (x == NULL_RTX) |
11282 | x = simplify_gen_binary (code, mode: shift_mode, op0: varop, |
11283 | op1: gen_int_shift_amount (shift_mode, count)); |
11284 | |
11285 | /* If we were doing an LSHIFTRT in a wider mode than it was originally, |
11286 | turn off all the bits that the shift would have turned off. */ |
11287 | if (orig_code == LSHIFTRT && result_mode != shift_mode) |
11288 | /* We only change the modes of scalar shifts. */ |
11289 | x = simplify_and_const_int (NULL_RTX, mode: as_a <scalar_int_mode> (m: shift_mode), |
11290 | varop: x, GET_MODE_MASK (result_mode) >> orig_count); |
11291 | |
11292 | /* Do the remainder of the processing in RESULT_MODE. */ |
11293 | x = gen_lowpart_or_truncate (mode: result_mode, x); |
11294 | |
11295 | /* If COMPLEMENT_P is set, we have to complement X before doing the outer |
11296 | operation. */ |
11297 | if (complement_p) |
11298 | x = simplify_gen_unary (code: NOT, mode: result_mode, op: x, op_mode: result_mode); |
11299 | |
11300 | if (outer_op != UNKNOWN) |
11301 | { |
11302 | int_result_mode = as_a <scalar_int_mode> (m: result_mode); |
11303 | |
11304 | if (GET_RTX_CLASS (outer_op) != RTX_UNARY |
11305 | && GET_MODE_PRECISION (mode: int_result_mode) < HOST_BITS_PER_WIDE_INT) |
11306 | outer_const = trunc_int_for_mode (outer_const, int_result_mode); |
11307 | |
11308 | if (outer_op == AND) |
11309 | x = simplify_and_const_int (NULL_RTX, mode: int_result_mode, varop: x, constop: outer_const); |
11310 | else if (outer_op == SET) |
11311 | { |
11312 | /* This means that we have determined that the result is |
11313 | equivalent to a constant. This should be rare. */ |
11314 | if (!side_effects_p (x)) |
11315 | x = GEN_INT (outer_const); |
11316 | } |
11317 | else if (GET_RTX_CLASS (outer_op) == RTX_UNARY) |
11318 | x = simplify_gen_unary (code: outer_op, mode: int_result_mode, op: x, op_mode: int_result_mode); |
11319 | else |
11320 | x = simplify_gen_binary (code: outer_op, mode: int_result_mode, op0: x, |
11321 | GEN_INT (outer_const)); |
11322 | } |
11323 | |
11324 | return x; |
11325 | } |
11326 | |
11327 | /* Simplify a shift of VAROP by COUNT bits. CODE says what kind of shift. |
11328 | The result of the shift is RESULT_MODE. If we cannot simplify it, |
11329 | return X or, if it is NULL, synthesize the expression with |
11330 | simplify_gen_binary. Otherwise, return a simplified value. |
11331 | |
11332 | The shift is normally computed in the widest mode we find in VAROP, as |
11333 | long as it isn't a different number of words than RESULT_MODE. Exceptions |
11334 | are ASHIFTRT and ROTATE, which are always done in their original mode. */ |
11335 | |
11336 | static rtx |
11337 | simplify_shift_const (rtx x, enum rtx_code code, machine_mode result_mode, |
11338 | rtx varop, int count) |
11339 | { |
11340 | rtx tem = simplify_shift_const_1 (code, result_mode, varop, orig_count: count); |
11341 | if (tem) |
11342 | return tem; |
11343 | |
11344 | if (!x) |
11345 | x = simplify_gen_binary (code, GET_MODE (varop), op0: varop, |
11346 | op1: gen_int_shift_amount (GET_MODE (varop), count)); |
11347 | if (GET_MODE (x) != result_mode) |
11348 | x = gen_lowpart (result_mode, x); |
11349 | return x; |
11350 | } |
11351 | |
11352 | |
11353 | /* A subroutine of recog_for_combine. See there for arguments and |
11354 | return value. */ |
11355 | |
11356 | static int |
11357 | recog_for_combine_1 (rtx *pnewpat, rtx_insn *insn, rtx *pnotes) |
11358 | { |
11359 | rtx pat = *pnewpat; |
11360 | rtx pat_without_clobbers; |
11361 | int insn_code_number; |
11362 | int num_clobbers_to_add = 0; |
11363 | int i; |
11364 | rtx notes = NULL_RTX; |
11365 | rtx old_notes, old_pat; |
11366 | int old_icode; |
11367 | |
11368 | /* If PAT is a PARALLEL, check to see if it contains the CLOBBER |
11369 | we use to indicate that something didn't match. If we find such a |
11370 | thing, force rejection. */ |
11371 | if (GET_CODE (pat) == PARALLEL) |
11372 | for (i = XVECLEN (pat, 0) - 1; i >= 0; i--) |
11373 | if (GET_CODE (XVECEXP (pat, 0, i)) == CLOBBER |
11374 | && XEXP (XVECEXP (pat, 0, i), 0) == const0_rtx) |
11375 | return -1; |
11376 | |
11377 | old_pat = PATTERN (insn); |
11378 | old_notes = REG_NOTES (insn); |
11379 | PATTERN (insn) = pat; |
11380 | REG_NOTES (insn) = NULL_RTX; |
11381 | |
11382 | insn_code_number = recog (pat, insn, &num_clobbers_to_add); |
11383 | if (dump_file && (dump_flags & TDF_DETAILS)) |
11384 | { |
11385 | if (insn_code_number < 0) |
11386 | fputs (s: "Failed to match this instruction:\n" , stream: dump_file); |
11387 | else |
11388 | fputs (s: "Successfully matched this instruction:\n" , stream: dump_file); |
11389 | print_rtl_single (dump_file, pat); |
11390 | } |
11391 | |
11392 | /* If it isn't, there is the possibility that we previously had an insn |
11393 | that clobbered some register as a side effect, but the combined |
11394 | insn doesn't need to do that. So try once more without the clobbers |
11395 | unless this represents an ASM insn. */ |
11396 | |
11397 | if (insn_code_number < 0 && ! check_asm_operands (pat) |
11398 | && GET_CODE (pat) == PARALLEL) |
11399 | { |
11400 | int pos; |
11401 | |
11402 | for (pos = 0, i = 0; i < XVECLEN (pat, 0); i++) |
11403 | if (GET_CODE (XVECEXP (pat, 0, i)) != CLOBBER) |
11404 | { |
11405 | if (i != pos) |
11406 | SUBST (XVECEXP (pat, 0, pos), XVECEXP (pat, 0, i)); |
11407 | pos++; |
11408 | } |
11409 | |
11410 | SUBST_INT (XVECLEN (pat, 0), pos); |
11411 | |
11412 | if (pos == 1) |
11413 | pat = XVECEXP (pat, 0, 0); |
11414 | |
11415 | PATTERN (insn) = pat; |
11416 | insn_code_number = recog (pat, insn, &num_clobbers_to_add); |
11417 | if (dump_file && (dump_flags & TDF_DETAILS)) |
11418 | { |
11419 | if (insn_code_number < 0) |
11420 | fputs (s: "Failed to match this instruction:\n" , stream: dump_file); |
11421 | else |
11422 | fputs (s: "Successfully matched this instruction:\n" , stream: dump_file); |
11423 | print_rtl_single (dump_file, pat); |
11424 | } |
11425 | } |
11426 | |
11427 | pat_without_clobbers = pat; |
11428 | |
11429 | PATTERN (insn) = old_pat; |
11430 | REG_NOTES (insn) = old_notes; |
11431 | |
11432 | /* Recognize all noop sets, these will be killed by followup pass. */ |
11433 | if (insn_code_number < 0 && GET_CODE (pat) == SET && set_noop_p (pat)) |
11434 | insn_code_number = NOOP_MOVE_INSN_CODE, num_clobbers_to_add = 0; |
11435 | |
11436 | /* If we had any clobbers to add, make a new pattern than contains |
11437 | them. Then check to make sure that all of them are dead. */ |
11438 | if (num_clobbers_to_add) |
11439 | { |
11440 | rtx newpat = gen_rtx_PARALLEL (VOIDmode, |
11441 | rtvec_alloc (GET_CODE (pat) == PARALLEL |
11442 | ? (XVECLEN (pat, 0) |
11443 | + num_clobbers_to_add) |
11444 | : num_clobbers_to_add + 1)); |
11445 | |
11446 | if (GET_CODE (pat) == PARALLEL) |
11447 | for (i = 0; i < XVECLEN (pat, 0); i++) |
11448 | XVECEXP (newpat, 0, i) = XVECEXP (pat, 0, i); |
11449 | else |
11450 | XVECEXP (newpat, 0, 0) = pat; |
11451 | |
11452 | add_clobbers (newpat, insn_code_number); |
11453 | |
11454 | for (i = XVECLEN (newpat, 0) - num_clobbers_to_add; |
11455 | i < XVECLEN (newpat, 0); i++) |
11456 | { |
11457 | if (REG_P (XEXP (XVECEXP (newpat, 0, i), 0)) |
11458 | && ! reg_dead_at_p (XEXP (XVECEXP (newpat, 0, i), 0), insn)) |
11459 | return -1; |
11460 | if (GET_CODE (XEXP (XVECEXP (newpat, 0, i), 0)) != SCRATCH) |
11461 | { |
11462 | gcc_assert (REG_P (XEXP (XVECEXP (newpat, 0, i), 0))); |
11463 | notes = alloc_reg_note (REG_UNUSED, |
11464 | XEXP (XVECEXP (newpat, 0, i), 0), notes); |
11465 | } |
11466 | } |
11467 | pat = newpat; |
11468 | } |
11469 | |
11470 | if (insn_code_number >= 0 |
11471 | && insn_code_number != NOOP_MOVE_INSN_CODE) |
11472 | { |
11473 | old_pat = PATTERN (insn); |
11474 | old_notes = REG_NOTES (insn); |
11475 | old_icode = INSN_CODE (insn); |
11476 | PATTERN (insn) = pat; |
11477 | REG_NOTES (insn) = notes; |
11478 | INSN_CODE (insn) = insn_code_number; |
11479 | |
11480 | /* Allow targets to reject combined insn. */ |
11481 | if (!targetm.legitimate_combined_insn (insn)) |
11482 | { |
11483 | if (dump_file && (dump_flags & TDF_DETAILS)) |
11484 | fputs (s: "Instruction not appropriate for target." , |
11485 | stream: dump_file); |
11486 | |
11487 | /* Callers expect recog_for_combine to strip |
11488 | clobbers from the pattern on failure. */ |
11489 | pat = pat_without_clobbers; |
11490 | notes = NULL_RTX; |
11491 | |
11492 | insn_code_number = -1; |
11493 | } |
11494 | |
11495 | PATTERN (insn) = old_pat; |
11496 | REG_NOTES (insn) = old_notes; |
11497 | INSN_CODE (insn) = old_icode; |
11498 | } |
11499 | |
11500 | *pnewpat = pat; |
11501 | *pnotes = notes; |
11502 | |
11503 | return insn_code_number; |
11504 | } |
11505 | |
11506 | /* Change every ZERO_EXTRACT and ZERO_EXTEND of a SUBREG that can be |
11507 | expressed as an AND and maybe an LSHIFTRT, to that formulation. |
11508 | Return whether anything was so changed. */ |
11509 | |
11510 | static bool |
11511 | change_zero_ext (rtx pat) |
11512 | { |
11513 | bool changed = false; |
11514 | rtx *src = &SET_SRC (pat); |
11515 | |
11516 | subrtx_ptr_iterator::array_type array; |
11517 | FOR_EACH_SUBRTX_PTR (iter, array, src, NONCONST) |
11518 | { |
11519 | rtx x = **iter; |
11520 | scalar_int_mode mode, inner_mode; |
11521 | if (!is_a <scalar_int_mode> (GET_MODE (x), result: &mode)) |
11522 | continue; |
11523 | int size; |
11524 | |
11525 | if (GET_CODE (x) == ZERO_EXTRACT |
11526 | && CONST_INT_P (XEXP (x, 1)) |
11527 | && CONST_INT_P (XEXP (x, 2)) |
11528 | && is_a <scalar_int_mode> (GET_MODE (XEXP (x, 0)), result: &inner_mode) |
11529 | && GET_MODE_PRECISION (mode: inner_mode) <= GET_MODE_PRECISION (mode)) |
11530 | { |
11531 | size = INTVAL (XEXP (x, 1)); |
11532 | |
11533 | int start = INTVAL (XEXP (x, 2)); |
11534 | if (BITS_BIG_ENDIAN) |
11535 | start = GET_MODE_PRECISION (mode: inner_mode) - size - start; |
11536 | |
11537 | if (start != 0) |
11538 | x = gen_rtx_LSHIFTRT (inner_mode, XEXP (x, 0), |
11539 | gen_int_shift_amount (inner_mode, start)); |
11540 | else |
11541 | x = XEXP (x, 0); |
11542 | |
11543 | if (mode != inner_mode) |
11544 | { |
11545 | if (REG_P (x) && HARD_REGISTER_P (x) |
11546 | && !can_change_dest_mode (x, added_sets: 0, mode)) |
11547 | continue; |
11548 | |
11549 | x = gen_lowpart_SUBREG (mode, x); |
11550 | } |
11551 | } |
11552 | else if (GET_CODE (x) == ZERO_EXTEND |
11553 | && GET_CODE (XEXP (x, 0)) == SUBREG |
11554 | && SCALAR_INT_MODE_P (GET_MODE (SUBREG_REG (XEXP (x, 0)))) |
11555 | && !paradoxical_subreg_p (XEXP (x, 0)) |
11556 | && subreg_lowpart_p (XEXP (x, 0))) |
11557 | { |
11558 | inner_mode = as_a <scalar_int_mode> (GET_MODE (XEXP (x, 0))); |
11559 | size = GET_MODE_PRECISION (mode: inner_mode); |
11560 | x = SUBREG_REG (XEXP (x, 0)); |
11561 | if (GET_MODE (x) != mode) |
11562 | { |
11563 | if (REG_P (x) && HARD_REGISTER_P (x) |
11564 | && !can_change_dest_mode (x, added_sets: 0, mode)) |
11565 | continue; |
11566 | |
11567 | x = gen_lowpart_SUBREG (mode, x); |
11568 | } |
11569 | } |
11570 | else if (GET_CODE (x) == ZERO_EXTEND |
11571 | && REG_P (XEXP (x, 0)) |
11572 | && HARD_REGISTER_P (XEXP (x, 0)) |
11573 | && can_change_dest_mode (XEXP (x, 0), added_sets: 0, mode)) |
11574 | { |
11575 | inner_mode = as_a <scalar_int_mode> (GET_MODE (XEXP (x, 0))); |
11576 | size = GET_MODE_PRECISION (mode: inner_mode); |
11577 | x = gen_rtx_REG (mode, REGNO (XEXP (x, 0))); |
11578 | } |
11579 | else |
11580 | continue; |
11581 | |
11582 | if (!(GET_CODE (x) == LSHIFTRT |
11583 | && CONST_INT_P (XEXP (x, 1)) |
11584 | && size + INTVAL (XEXP (x, 1)) == GET_MODE_PRECISION (mode))) |
11585 | { |
11586 | wide_int mask = wi::mask (width: size, negate_p: false, precision: GET_MODE_PRECISION (mode)); |
11587 | x = gen_rtx_AND (mode, x, immed_wide_int_const (mask, mode)); |
11588 | } |
11589 | |
11590 | SUBST (**iter, x); |
11591 | changed = true; |
11592 | } |
11593 | |
11594 | if (changed) |
11595 | FOR_EACH_SUBRTX_PTR (iter, array, src, NONCONST) |
11596 | maybe_swap_commutative_operands (x: **iter); |
11597 | |
11598 | rtx *dst = &SET_DEST (pat); |
11599 | scalar_int_mode mode; |
11600 | if (GET_CODE (*dst) == ZERO_EXTRACT |
11601 | && REG_P (XEXP (*dst, 0)) |
11602 | && is_a <scalar_int_mode> (GET_MODE (XEXP (*dst, 0)), result: &mode) |
11603 | && CONST_INT_P (XEXP (*dst, 1)) |
11604 | && CONST_INT_P (XEXP (*dst, 2))) |
11605 | { |
11606 | rtx reg = XEXP (*dst, 0); |
11607 | int width = INTVAL (XEXP (*dst, 1)); |
11608 | int offset = INTVAL (XEXP (*dst, 2)); |
11609 | int reg_width = GET_MODE_PRECISION (mode); |
11610 | if (BITS_BIG_ENDIAN) |
11611 | offset = reg_width - width - offset; |
11612 | |
11613 | rtx x, y, z, w; |
11614 | wide_int mask = wi::shifted_mask (start: offset, width, negate_p: true, precision: reg_width); |
11615 | wide_int mask2 = wi::shifted_mask (start: offset, width, negate_p: false, precision: reg_width); |
11616 | x = gen_rtx_AND (mode, reg, immed_wide_int_const (mask, mode)); |
11617 | if (offset) |
11618 | y = gen_rtx_ASHIFT (mode, SET_SRC (pat), GEN_INT (offset)); |
11619 | else |
11620 | y = SET_SRC (pat); |
11621 | z = gen_rtx_AND (mode, y, immed_wide_int_const (mask2, mode)); |
11622 | w = gen_rtx_IOR (mode, x, z); |
11623 | SUBST (SET_DEST (pat), reg); |
11624 | SUBST (SET_SRC (pat), w); |
11625 | |
11626 | changed = true; |
11627 | } |
11628 | |
11629 | return changed; |
11630 | } |
11631 | |
11632 | /* Like recog, but we receive the address of a pointer to a new pattern. |
11633 | We try to match the rtx that the pointer points to. |
11634 | If that fails, we may try to modify or replace the pattern, |
11635 | storing the replacement into the same pointer object. |
11636 | |
11637 | Modifications include deletion or addition of CLOBBERs. If the |
11638 | instruction will still not match, we change ZERO_EXTEND and ZERO_EXTRACT |
11639 | to the equivalent AND and perhaps LSHIFTRT patterns, and try with that |
11640 | (and undo if that fails). |
11641 | |
11642 | PNOTES is a pointer to a location where any REG_UNUSED notes added for |
11643 | the CLOBBERs are placed. |
11644 | |
11645 | The value is the final insn code from the pattern ultimately matched, |
11646 | or -1. */ |
11647 | |
11648 | static int |
11649 | recog_for_combine (rtx *pnewpat, rtx_insn *insn, rtx *pnotes) |
11650 | { |
11651 | rtx pat = *pnewpat; |
11652 | int insn_code_number = recog_for_combine_1 (pnewpat, insn, pnotes); |
11653 | if (insn_code_number >= 0 || check_asm_operands (pat)) |
11654 | return insn_code_number; |
11655 | |
11656 | void *marker = get_undo_marker (); |
11657 | bool changed = false; |
11658 | |
11659 | if (GET_CODE (pat) == SET) |
11660 | { |
11661 | /* For an unrecognized single set of a constant, try placing it in |
11662 | the constant pool, if this function already uses one. */ |
11663 | rtx src = SET_SRC (pat); |
11664 | if (CONSTANT_P (src) |
11665 | && !CONST_INT_P (src) |
11666 | && crtl->uses_const_pool) |
11667 | { |
11668 | machine_mode mode = GET_MODE (src); |
11669 | if (mode == VOIDmode) |
11670 | mode = GET_MODE (SET_DEST (pat)); |
11671 | src = force_const_mem (mode, src); |
11672 | if (src) |
11673 | { |
11674 | SUBST (SET_SRC (pat), src); |
11675 | changed = true; |
11676 | } |
11677 | } |
11678 | else |
11679 | changed = change_zero_ext (pat); |
11680 | } |
11681 | else if (GET_CODE (pat) == PARALLEL) |
11682 | { |
11683 | int i; |
11684 | for (i = 0; i < XVECLEN (pat, 0); i++) |
11685 | { |
11686 | rtx set = XVECEXP (pat, 0, i); |
11687 | if (GET_CODE (set) == SET) |
11688 | changed |= change_zero_ext (pat: set); |
11689 | } |
11690 | } |
11691 | |
11692 | if (changed) |
11693 | { |
11694 | insn_code_number = recog_for_combine_1 (pnewpat, insn, pnotes); |
11695 | |
11696 | if (insn_code_number < 0) |
11697 | undo_to_marker (marker); |
11698 | } |
11699 | |
11700 | return insn_code_number; |
11701 | } |
11702 | |
11703 | /* Like gen_lowpart_general but for use by combine. In combine it |
11704 | is not possible to create any new pseudoregs. However, it is |
11705 | safe to create invalid memory addresses, because combine will |
11706 | try to recognize them and all they will do is make the combine |
11707 | attempt fail. |
11708 | |
11709 | If for some reason this cannot do its job, an rtx |
11710 | (clobber (const_int 0)) is returned. |
11711 | An insn containing that will not be recognized. */ |
11712 | |
11713 | static rtx |
11714 | gen_lowpart_for_combine (machine_mode omode, rtx x) |
11715 | { |
11716 | machine_mode imode = GET_MODE (x); |
11717 | rtx result; |
11718 | |
11719 | if (omode == imode) |
11720 | return x; |
11721 | |
11722 | /* We can only support MODE being wider than a word if X is a |
11723 | constant integer or has a mode the same size. */ |
11724 | if (maybe_gt (GET_MODE_SIZE (omode), UNITS_PER_WORD) |
11725 | && ! (CONST_SCALAR_INT_P (x) |
11726 | || known_eq (GET_MODE_SIZE (imode), GET_MODE_SIZE (omode)))) |
11727 | goto fail; |
11728 | |
11729 | /* X might be a paradoxical (subreg (mem)). In that case, gen_lowpart |
11730 | won't know what to do. So we will strip off the SUBREG here and |
11731 | process normally. */ |
11732 | if (GET_CODE (x) == SUBREG && MEM_P (SUBREG_REG (x))) |
11733 | { |
11734 | x = SUBREG_REG (x); |
11735 | |
11736 | /* For use in case we fall down into the address adjustments |
11737 | further below, we need to adjust the known mode and size of |
11738 | x; imode and isize, since we just adjusted x. */ |
11739 | imode = GET_MODE (x); |
11740 | |
11741 | if (imode == omode) |
11742 | return x; |
11743 | } |
11744 | |
11745 | result = gen_lowpart_common (omode, x); |
11746 | |
11747 | if (result) |
11748 | return result; |
11749 | |
11750 | if (MEM_P (x)) |
11751 | { |
11752 | /* Refuse to work on a volatile memory ref or one with a mode-dependent |
11753 | address. */ |
11754 | if (MEM_VOLATILE_P (x) |
11755 | || mode_dependent_address_p (XEXP (x, 0), MEM_ADDR_SPACE (x))) |
11756 | goto fail; |
11757 | |
11758 | /* If we want to refer to something bigger than the original memref, |
11759 | generate a paradoxical subreg instead. That will force a reload |
11760 | of the original memref X. */ |
11761 | if (paradoxical_subreg_p (outermode: omode, innermode: imode)) |
11762 | return gen_rtx_SUBREG (omode, x, 0); |
11763 | |
11764 | poly_int64 offset = byte_lowpart_offset (omode, imode); |
11765 | return adjust_address_nv (x, omode, offset); |
11766 | } |
11767 | |
11768 | /* If X is a comparison operator, rewrite it in a new mode. This |
11769 | probably won't match, but may allow further simplifications. */ |
11770 | else if (COMPARISON_P (x) |
11771 | && SCALAR_INT_MODE_P (imode) |
11772 | && SCALAR_INT_MODE_P (omode)) |
11773 | return gen_rtx_fmt_ee (GET_CODE (x), omode, XEXP (x, 0), XEXP (x, 1)); |
11774 | |
11775 | /* If we couldn't simplify X any other way, just enclose it in a |
11776 | SUBREG. Normally, this SUBREG won't match, but some patterns may |
11777 | include an explicit SUBREG or we may simplify it further in combine. */ |
11778 | else |
11779 | { |
11780 | rtx res; |
11781 | |
11782 | if (imode == VOIDmode) |
11783 | { |
11784 | imode = int_mode_for_mode (omode).require (); |
11785 | x = gen_lowpart_common (imode, x); |
11786 | if (x == NULL) |
11787 | goto fail; |
11788 | } |
11789 | res = lowpart_subreg (outermode: omode, op: x, innermode: imode); |
11790 | if (res) |
11791 | return res; |
11792 | } |
11793 | |
11794 | fail: |
11795 | return gen_rtx_CLOBBER (omode, const0_rtx); |
11796 | } |
11797 | |
11798 | /* Try to simplify a comparison between OP0 and a constant OP1, |
11799 | where CODE is the comparison code that will be tested, into a |
11800 | (CODE OP0 const0_rtx) form. |
11801 | |
11802 | The result is a possibly different comparison code to use. |
11803 | *POP0 and *POP1 may be updated. */ |
11804 | |
11805 | static enum rtx_code |
11806 | simplify_compare_const (enum rtx_code code, machine_mode mode, |
11807 | rtx *pop0, rtx *pop1) |
11808 | { |
11809 | scalar_int_mode int_mode; |
11810 | rtx op0 = *pop0; |
11811 | HOST_WIDE_INT const_op = INTVAL (*pop1); |
11812 | |
11813 | /* Get the constant we are comparing against and turn off all bits |
11814 | not on in our mode. */ |
11815 | if (mode != VOIDmode) |
11816 | const_op = trunc_int_for_mode (const_op, mode); |
11817 | |
11818 | /* If we are comparing against a constant power of two and the value |
11819 | being compared can only have that single bit nonzero (e.g., it was |
11820 | `and'ed with that bit), we can replace this with a comparison |
11821 | with zero. */ |
11822 | if (const_op |
11823 | && (code == EQ || code == NE || code == GE || code == GEU |
11824 | || code == LT || code == LTU) |
11825 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
11826 | && GET_MODE_PRECISION (mode: int_mode) - 1 < HOST_BITS_PER_WIDE_INT |
11827 | && pow2p_hwi (x: const_op & GET_MODE_MASK (int_mode)) |
11828 | && (nonzero_bits (op0, int_mode) |
11829 | == (unsigned HOST_WIDE_INT) (const_op & GET_MODE_MASK (int_mode)))) |
11830 | { |
11831 | code = (code == EQ || code == GE || code == GEU ? NE : EQ); |
11832 | const_op = 0; |
11833 | } |
11834 | |
11835 | /* Similarly, if we are comparing a value known to be either -1 or |
11836 | 0 with -1, change it to the opposite comparison against zero. */ |
11837 | if (const_op == -1 |
11838 | && (code == EQ || code == NE || code == GT || code == LE |
11839 | || code == GEU || code == LTU) |
11840 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
11841 | && num_sign_bit_copies (op0, int_mode) == GET_MODE_PRECISION (mode: int_mode)) |
11842 | { |
11843 | code = (code == EQ || code == LE || code == GEU ? NE : EQ); |
11844 | const_op = 0; |
11845 | } |
11846 | |
11847 | /* Do some canonicalizations based on the comparison code. We prefer |
11848 | comparisons against zero and then prefer equality comparisons. |
11849 | If we can reduce the size of a constant, we will do that too. */ |
11850 | switch (code) |
11851 | { |
11852 | case LT: |
11853 | /* < C is equivalent to <= (C - 1) */ |
11854 | if (const_op > 0) |
11855 | { |
11856 | const_op -= 1; |
11857 | code = LE; |
11858 | /* ... fall through to LE case below. */ |
11859 | gcc_fallthrough (); |
11860 | } |
11861 | else |
11862 | break; |
11863 | |
11864 | case LE: |
11865 | /* <= C is equivalent to < (C + 1); we do this for C < 0 */ |
11866 | if (const_op < 0) |
11867 | { |
11868 | const_op += 1; |
11869 | code = LT; |
11870 | } |
11871 | |
11872 | /* If we are doing a <= 0 comparison on a value known to have |
11873 | a zero sign bit, we can replace this with == 0. */ |
11874 | else if (const_op == 0 |
11875 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
11876 | && GET_MODE_PRECISION (mode: int_mode) - 1 < HOST_BITS_PER_WIDE_INT |
11877 | && (nonzero_bits (op0, int_mode) |
11878 | & (HOST_WIDE_INT_1U << (GET_MODE_PRECISION (mode: int_mode) - 1))) |
11879 | == 0) |
11880 | code = EQ; |
11881 | break; |
11882 | |
11883 | case GE: |
11884 | /* >= C is equivalent to > (C - 1). */ |
11885 | if (const_op > 0) |
11886 | { |
11887 | const_op -= 1; |
11888 | code = GT; |
11889 | /* ... fall through to GT below. */ |
11890 | gcc_fallthrough (); |
11891 | } |
11892 | else |
11893 | break; |
11894 | |
11895 | case GT: |
11896 | /* > C is equivalent to >= (C + 1); we do this for C < 0. */ |
11897 | if (const_op < 0) |
11898 | { |
11899 | const_op += 1; |
11900 | code = GE; |
11901 | } |
11902 | |
11903 | /* If we are doing a > 0 comparison on a value known to have |
11904 | a zero sign bit, we can replace this with != 0. */ |
11905 | else if (const_op == 0 |
11906 | && is_a <scalar_int_mode> (m: mode, result: &int_mode) |
11907 | && GET_MODE_PRECISION (mode: int_mode) - 1 < HOST_BITS_PER_WIDE_INT |
11908 | && (nonzero_bits (op0, int_mode) |
11909 | & (HOST_WIDE_INT_1U << (GET_MODE_PRECISION (mode: int_mode) - 1))) |
11910 | == 0) |
11911 | code = NE; |
11912 | break; |
11913 | |
11914 | case LTU: |
11915 | /* < C is equivalent to <= (C - 1). */ |
11916 | if (const_op > 0) |
11917 | { |
11918 | const_op -= 1; |
11919 | code = LEU; |
11920 | /* ... fall through ... */ |
11921 | gcc_fallthrough (); |
11922 | } |
11923 | /* (unsigned) < 0x80000000 is equivalent to >= 0. */ |
11924 | else if (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
11925 | && GET_MODE_PRECISION (mode: int_mode) - 1 < HOST_BITS_PER_WIDE_INT |
11926 | && (((unsigned HOST_WIDE_INT) const_op & GET_MODE_MASK (int_mode)) |
11927 | == HOST_WIDE_INT_1U << (GET_MODE_PRECISION (mode: int_mode) - 1))) |
11928 | { |
11929 | const_op = 0; |
11930 | code = GE; |
11931 | break; |
11932 | } |
11933 | else |
11934 | break; |
11935 | |
11936 | case LEU: |
11937 | /* unsigned <= 0 is equivalent to == 0 */ |
11938 | if (const_op == 0) |
11939 | code = EQ; |
11940 | /* (unsigned) <= 0x7fffffff is equivalent to >= 0. */ |
11941 | else if (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
11942 | && GET_MODE_PRECISION (mode: int_mode) - 1 < HOST_BITS_PER_WIDE_INT |
11943 | && ((unsigned HOST_WIDE_INT) const_op |
11944 | == ((HOST_WIDE_INT_1U |
11945 | << (GET_MODE_PRECISION (mode: int_mode) - 1)) - 1))) |
11946 | { |
11947 | const_op = 0; |
11948 | code = GE; |
11949 | } |
11950 | break; |
11951 | |
11952 | case GEU: |
11953 | /* >= C is equivalent to > (C - 1). */ |
11954 | if (const_op > 1) |
11955 | { |
11956 | const_op -= 1; |
11957 | code = GTU; |
11958 | /* ... fall through ... */ |
11959 | gcc_fallthrough (); |
11960 | } |
11961 | |
11962 | /* (unsigned) >= 0x80000000 is equivalent to < 0. */ |
11963 | else if (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
11964 | && GET_MODE_PRECISION (mode: int_mode) - 1 < HOST_BITS_PER_WIDE_INT |
11965 | && (((unsigned HOST_WIDE_INT) const_op & GET_MODE_MASK (int_mode)) |
11966 | == HOST_WIDE_INT_1U << (GET_MODE_PRECISION (mode: int_mode) - 1))) |
11967 | { |
11968 | const_op = 0; |
11969 | code = LT; |
11970 | break; |
11971 | } |
11972 | else |
11973 | break; |
11974 | |
11975 | case GTU: |
11976 | /* unsigned > 0 is equivalent to != 0 */ |
11977 | if (const_op == 0) |
11978 | code = NE; |
11979 | /* (unsigned) > 0x7fffffff is equivalent to < 0. */ |
11980 | else if (is_a <scalar_int_mode> (m: mode, result: &int_mode) |
11981 | && GET_MODE_PRECISION (mode: int_mode) - 1 < HOST_BITS_PER_WIDE_INT |
11982 | && ((unsigned HOST_WIDE_INT) const_op |
11983 | == (HOST_WIDE_INT_1U |
11984 | << (GET_MODE_PRECISION (mode: int_mode) - 1)) - 1)) |
11985 | { |
11986 | const_op = 0; |
11987 | code = LT; |
11988 | } |
11989 | break; |
11990 | |
11991 | default: |
11992 | break; |
11993 | } |
11994 | |
11995 | /* Narrow non-symmetric comparison of memory and constant as e.g. |
11996 | x0...x7 <= 0x3fffffffffffffff into x0 <= 0x3f where x0 is the most |
11997 | significant byte. Likewise, transform x0...x7 >= 0x4000000000000000 into |
11998 | x0 >= 0x40. */ |
11999 | if ((code == LEU || code == LTU || code == GEU || code == GTU) |
12000 | && is_a <scalar_int_mode> (GET_MODE (op0), result: &int_mode) |
12001 | && HWI_COMPUTABLE_MODE_P (mode: int_mode) |
12002 | && MEM_P (op0) |
12003 | && !MEM_VOLATILE_P (op0) |
12004 | /* The optimization makes only sense for constants which are big enough |
12005 | so that we have a chance to chop off something at all. */ |
12006 | && ((unsigned HOST_WIDE_INT) const_op & GET_MODE_MASK (int_mode)) > 0xff |
12007 | /* Ensure that we do not overflow during normalization. */ |
12008 | && (code != GTU |
12009 | || ((unsigned HOST_WIDE_INT) const_op & GET_MODE_MASK (int_mode)) |
12010 | < HOST_WIDE_INT_M1U) |
12011 | && trunc_int_for_mode (const_op, int_mode) == const_op) |
12012 | { |
12013 | unsigned HOST_WIDE_INT n |
12014 | = (unsigned HOST_WIDE_INT) const_op & GET_MODE_MASK (int_mode); |
12015 | enum rtx_code adjusted_code; |
12016 | |
12017 | /* Normalize code to either LEU or GEU. */ |
12018 | if (code == LTU) |
12019 | { |
12020 | --n; |
12021 | adjusted_code = LEU; |
12022 | } |
12023 | else if (code == GTU) |
12024 | { |
12025 | ++n; |
12026 | adjusted_code = GEU; |
12027 | } |
12028 | else |
12029 | adjusted_code = code; |
12030 | |
12031 | scalar_int_mode narrow_mode_iter; |
12032 | FOR_EACH_MODE_UNTIL (narrow_mode_iter, int_mode) |
12033 | { |
12034 | unsigned nbits = GET_MODE_PRECISION (mode: int_mode) |
12035 | - GET_MODE_PRECISION (mode: narrow_mode_iter); |
12036 | unsigned HOST_WIDE_INT mask = (HOST_WIDE_INT_1U << nbits) - 1; |
12037 | unsigned HOST_WIDE_INT lower_bits = n & mask; |
12038 | if ((adjusted_code == LEU && lower_bits == mask) |
12039 | || (adjusted_code == GEU && lower_bits == 0)) |
12040 | { |
12041 | n >>= nbits; |
12042 | break; |
12043 | } |
12044 | } |
12045 | |
12046 | if (narrow_mode_iter < int_mode) |
12047 | { |
12048 | if (dump_file && (dump_flags & TDF_DETAILS)) |
12049 | { |
12050 | fprintf ( |
12051 | stream: dump_file, format: "narrow comparison from mode %s to %s: (MEM %s " |
12052 | HOST_WIDE_INT_PRINT_HEX ") to (MEM %s " |
12053 | HOST_WIDE_INT_PRINT_HEX ").\n" , GET_MODE_NAME (int_mode), |
12054 | GET_MODE_NAME (narrow_mode_iter), GET_RTX_NAME (code), |
12055 | (unsigned HOST_WIDE_INT) const_op & GET_MODE_MASK (int_mode), |
12056 | GET_RTX_NAME (adjusted_code), n); |
12057 | } |
12058 | poly_int64 offset = (BYTES_BIG_ENDIAN |
12059 | ? 0 |
12060 | : (GET_MODE_SIZE (mode: int_mode) |
12061 | - GET_MODE_SIZE (mode: narrow_mode_iter))); |
12062 | *pop0 = adjust_address_nv (op0, narrow_mode_iter, offset); |
12063 | *pop1 = gen_int_mode (n, narrow_mode_iter); |
12064 | return adjusted_code; |
12065 | } |
12066 | } |
12067 | |
12068 | *pop1 = GEN_INT (const_op); |
12069 | return code; |
12070 | } |
12071 | |
12072 | /* Simplify a comparison between *POP0 and *POP1 where CODE is the |
12073 | comparison code that will be tested. |
12074 | |
12075 | The result is a possibly different comparison code to use. *POP0 and |
12076 | *POP1 may be updated. |
12077 | |
12078 | It is possible that we might detect that a comparison is either always |
12079 | true or always false. However, we do not perform general constant |
12080 | folding in combine, so this knowledge isn't useful. Such tautologies |
12081 | should have been detected earlier. Hence we ignore all such cases. */ |
12082 | |
12083 | static enum rtx_code |
12084 | simplify_comparison (enum rtx_code code, rtx *pop0, rtx *pop1) |
12085 | { |
12086 | rtx op0 = *pop0; |
12087 | rtx op1 = *pop1; |
12088 | rtx tem, tem1; |
12089 | int i; |
12090 | scalar_int_mode mode, inner_mode, tmode; |
12091 | opt_scalar_int_mode tmode_iter; |
12092 | |
12093 | /* Try a few ways of applying the same transformation to both operands. */ |
12094 | while (1) |
12095 | { |
12096 | /* The test below this one won't handle SIGN_EXTENDs on these machines, |
12097 | so check specially. */ |
12098 | if (!WORD_REGISTER_OPERATIONS |
12099 | && code != GTU && code != GEU && code != LTU && code != LEU |
12100 | && GET_CODE (op0) == ASHIFTRT && GET_CODE (op1) == ASHIFTRT |
12101 | && GET_CODE (XEXP (op0, 0)) == ASHIFT |
12102 | && GET_CODE (XEXP (op1, 0)) == ASHIFT |
12103 | && GET_CODE (XEXP (XEXP (op0, 0), 0)) == SUBREG |
12104 | && GET_CODE (XEXP (XEXP (op1, 0), 0)) == SUBREG |
12105 | && is_a <scalar_int_mode> (GET_MODE (op0), result: &mode) |
12106 | && (is_a <scalar_int_mode> |
12107 | (GET_MODE (SUBREG_REG (XEXP (XEXP (op0, 0), 0))), result: &inner_mode)) |
12108 | && inner_mode == GET_MODE (SUBREG_REG (XEXP (XEXP (op1, 0), 0))) |
12109 | && CONST_INT_P (XEXP (op0, 1)) |
12110 | && XEXP (op0, 1) == XEXP (op1, 1) |
12111 | && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1) |
12112 | && XEXP (op0, 1) == XEXP (XEXP (op1, 0), 1) |
12113 | && (INTVAL (XEXP (op0, 1)) |
12114 | == (GET_MODE_PRECISION (mode) |
12115 | - GET_MODE_PRECISION (mode: inner_mode)))) |
12116 | { |
12117 | op0 = SUBREG_REG (XEXP (XEXP (op0, 0), 0)); |
12118 | op1 = SUBREG_REG (XEXP (XEXP (op1, 0), 0)); |
12119 | } |
12120 | |
12121 | /* If both operands are the same constant shift, see if we can ignore the |
12122 | shift. We can if the shift is a rotate or if the bits shifted out of |
12123 | this shift are known to be zero for both inputs and if the type of |
12124 | comparison is compatible with the shift. */ |
12125 | if (GET_CODE (op0) == GET_CODE (op1) |
12126 | && HWI_COMPUTABLE_MODE_P (GET_MODE (op0)) |
12127 | && ((GET_CODE (op0) == ROTATE && (code == NE || code == EQ)) |
12128 | || ((GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFT) |
12129 | && (code != GT && code != LT && code != GE && code != LE)) |
12130 | || (GET_CODE (op0) == ASHIFTRT |
12131 | && (code != GTU && code != LTU |
12132 | && code != GEU && code != LEU))) |
12133 | && CONST_INT_P (XEXP (op0, 1)) |
12134 | && INTVAL (XEXP (op0, 1)) >= 0 |
12135 | && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT |
12136 | && XEXP (op0, 1) == XEXP (op1, 1)) |
12137 | { |
12138 | machine_mode mode = GET_MODE (op0); |
12139 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); |
12140 | int shift_count = INTVAL (XEXP (op0, 1)); |
12141 | |
12142 | if (GET_CODE (op0) == LSHIFTRT || GET_CODE (op0) == ASHIFTRT) |
12143 | mask &= (mask >> shift_count) << shift_count; |
12144 | else if (GET_CODE (op0) == ASHIFT) |
12145 | mask = (mask & (mask << shift_count)) >> shift_count; |
12146 | |
12147 | if ((nonzero_bits (XEXP (op0, 0), mode) & ~mask) == 0 |
12148 | && (nonzero_bits (XEXP (op1, 0), mode) & ~mask) == 0) |
12149 | op0 = XEXP (op0, 0), op1 = XEXP (op1, 0); |
12150 | else |
12151 | break; |
12152 | } |
12153 | |
12154 | /* If both operands are AND's of a paradoxical SUBREG by constant, the |
12155 | SUBREGs are of the same mode, and, in both cases, the AND would |
12156 | be redundant if the comparison was done in the narrower mode, |
12157 | do the comparison in the narrower mode (e.g., we are AND'ing with 1 |
12158 | and the operand's possibly nonzero bits are 0xffffff01; in that case |
12159 | if we only care about QImode, we don't need the AND). This case |
12160 | occurs if the output mode of an scc insn is not SImode and |
12161 | STORE_FLAG_VALUE == 1 (e.g., the 386). |
12162 | |
12163 | Similarly, check for a case where the AND's are ZERO_EXTEND |
12164 | operations from some narrower mode even though a SUBREG is not |
12165 | present. */ |
12166 | |
12167 | else if (GET_CODE (op0) == AND && GET_CODE (op1) == AND |
12168 | && CONST_INT_P (XEXP (op0, 1)) |
12169 | && CONST_INT_P (XEXP (op1, 1))) |
12170 | { |
12171 | rtx inner_op0 = XEXP (op0, 0); |
12172 | rtx inner_op1 = XEXP (op1, 0); |
12173 | HOST_WIDE_INT c0 = INTVAL (XEXP (op0, 1)); |
12174 | HOST_WIDE_INT c1 = INTVAL (XEXP (op1, 1)); |
12175 | bool changed = false; |
12176 | |
12177 | if (paradoxical_subreg_p (x: inner_op0) |
12178 | && GET_CODE (inner_op1) == SUBREG |
12179 | && HWI_COMPUTABLE_MODE_P (GET_MODE (SUBREG_REG (inner_op0))) |
12180 | && (GET_MODE (SUBREG_REG (inner_op0)) |
12181 | == GET_MODE (SUBREG_REG (inner_op1))) |
12182 | && ((~c0) & nonzero_bits (SUBREG_REG (inner_op0), |
12183 | GET_MODE (SUBREG_REG (inner_op0)))) == 0 |
12184 | && ((~c1) & nonzero_bits (SUBREG_REG (inner_op1), |
12185 | GET_MODE (SUBREG_REG (inner_op1)))) == 0) |
12186 | { |
12187 | op0 = SUBREG_REG (inner_op0); |
12188 | op1 = SUBREG_REG (inner_op1); |
12189 | |
12190 | /* The resulting comparison is always unsigned since we masked |
12191 | off the original sign bit. */ |
12192 | code = unsigned_condition (code); |
12193 | |
12194 | changed = true; |
12195 | } |
12196 | |
12197 | else if (c0 == c1) |
12198 | FOR_EACH_MODE_UNTIL (tmode, |
12199 | as_a <scalar_int_mode> (GET_MODE (op0))) |
12200 | if ((unsigned HOST_WIDE_INT) c0 == GET_MODE_MASK (tmode)) |
12201 | { |
12202 | op0 = gen_lowpart_or_truncate (mode: tmode, x: inner_op0); |
12203 | op1 = gen_lowpart_or_truncate (mode: tmode, x: inner_op1); |
12204 | code = unsigned_condition (code); |
12205 | changed = true; |
12206 | break; |
12207 | } |
12208 | |
12209 | if (! changed) |
12210 | break; |
12211 | } |
12212 | |
12213 | /* If both operands are NOT, we can strip off the outer operation |
12214 | and adjust the comparison code for swapped operands; similarly for |
12215 | NEG, except that this must be an equality comparison. */ |
12216 | else if ((GET_CODE (op0) == NOT && GET_CODE (op1) == NOT) |
12217 | || (GET_CODE (op0) == NEG && GET_CODE (op1) == NEG |
12218 | && (code == EQ || code == NE))) |
12219 | op0 = XEXP (op0, 0), op1 = XEXP (op1, 0), code = swap_condition (code); |
12220 | |
12221 | else |
12222 | break; |
12223 | } |
12224 | |
12225 | /* If the first operand is a constant, swap the operands and adjust the |
12226 | comparison code appropriately, but don't do this if the second operand |
12227 | is already a constant integer. */ |
12228 | if (swap_commutative_operands_p (op0, op1)) |
12229 | { |
12230 | std::swap (a&: op0, b&: op1); |
12231 | code = swap_condition (code); |
12232 | } |
12233 | |
12234 | /* We now enter a loop during which we will try to simplify the comparison. |
12235 | For the most part, we only are concerned with comparisons with zero, |
12236 | but some things may really be comparisons with zero but not start |
12237 | out looking that way. */ |
12238 | |
12239 | while (CONST_INT_P (op1)) |
12240 | { |
12241 | machine_mode raw_mode = GET_MODE (op0); |
12242 | scalar_int_mode int_mode; |
12243 | int equality_comparison_p; |
12244 | int sign_bit_comparison_p; |
12245 | int unsigned_comparison_p; |
12246 | HOST_WIDE_INT const_op; |
12247 | |
12248 | /* We only want to handle integral modes. This catches VOIDmode, |
12249 | CCmode, and the floating-point modes. An exception is that we |
12250 | can handle VOIDmode if OP0 is a COMPARE or a comparison |
12251 | operation. */ |
12252 | |
12253 | if (GET_MODE_CLASS (raw_mode) != MODE_INT |
12254 | && ! (raw_mode == VOIDmode |
12255 | && (GET_CODE (op0) == COMPARE || COMPARISON_P (op0)))) |
12256 | break; |
12257 | |
12258 | /* Try to simplify the compare to constant, possibly changing the |
12259 | comparison op, and/or changing op1 to zero. */ |
12260 | code = simplify_compare_const (code, mode: raw_mode, pop0: &op0, pop1: &op1); |
12261 | const_op = INTVAL (op1); |
12262 | |
12263 | /* Compute some predicates to simplify code below. */ |
12264 | |
12265 | equality_comparison_p = (code == EQ || code == NE); |
12266 | sign_bit_comparison_p = ((code == LT || code == GE) && const_op == 0); |
12267 | unsigned_comparison_p = (code == LTU || code == LEU || code == GTU |
12268 | || code == GEU); |
12269 | |
12270 | /* If this is a sign bit comparison and we can do arithmetic in |
12271 | MODE, say that we will only be needing the sign bit of OP0. */ |
12272 | if (sign_bit_comparison_p |
12273 | && is_a <scalar_int_mode> (m: raw_mode, result: &int_mode) |
12274 | && HWI_COMPUTABLE_MODE_P (mode: int_mode)) |
12275 | op0 = force_to_mode (x: op0, mode: int_mode, |
12276 | HOST_WIDE_INT_1U |
12277 | << (GET_MODE_PRECISION (mode: int_mode) - 1), just_select: false); |
12278 | |
12279 | if (COMPARISON_P (op0)) |
12280 | { |
12281 | /* We can't do anything if OP0 is a condition code value, rather |
12282 | than an actual data value. */ |
12283 | if (const_op != 0 |
12284 | || GET_MODE_CLASS (GET_MODE (XEXP (op0, 0))) == MODE_CC) |
12285 | break; |
12286 | |
12287 | /* Get the two operands being compared. */ |
12288 | if (GET_CODE (XEXP (op0, 0)) == COMPARE) |
12289 | tem = XEXP (XEXP (op0, 0), 0), tem1 = XEXP (XEXP (op0, 0), 1); |
12290 | else |
12291 | tem = XEXP (op0, 0), tem1 = XEXP (op0, 1); |
12292 | |
12293 | /* Check for the cases where we simply want the result of the |
12294 | earlier test or the opposite of that result. */ |
12295 | if (code == NE || code == EQ |
12296 | || (val_signbit_known_set_p (raw_mode, STORE_FLAG_VALUE) |
12297 | && (code == LT || code == GE))) |
12298 | { |
12299 | enum rtx_code new_code; |
12300 | if (code == LT || code == NE) |
12301 | new_code = GET_CODE (op0); |
12302 | else |
12303 | new_code = reversed_comparison_code (op0, NULL); |
12304 | |
12305 | if (new_code != UNKNOWN) |
12306 | { |
12307 | code = new_code; |
12308 | op0 = tem; |
12309 | op1 = tem1; |
12310 | continue; |
12311 | } |
12312 | } |
12313 | break; |
12314 | } |
12315 | |
12316 | if (raw_mode == VOIDmode) |
12317 | break; |
12318 | scalar_int_mode mode = as_a <scalar_int_mode> (m: raw_mode); |
12319 | |
12320 | /* Now try cases based on the opcode of OP0. If none of the cases |
12321 | does a "continue", we exit this loop immediately after the |
12322 | switch. */ |
12323 | |
12324 | unsigned int mode_width = GET_MODE_PRECISION (mode); |
12325 | unsigned HOST_WIDE_INT mask = GET_MODE_MASK (mode); |
12326 | switch (GET_CODE (op0)) |
12327 | { |
12328 | case ZERO_EXTRACT: |
12329 | /* If we are extracting a single bit from a variable position in |
12330 | a constant that has only a single bit set and are comparing it |
12331 | with zero, we can convert this into an equality comparison |
12332 | between the position and the location of the single bit. */ |
12333 | /* Except we can't if SHIFT_COUNT_TRUNCATED is set, since we might |
12334 | have already reduced the shift count modulo the word size. */ |
12335 | if (!SHIFT_COUNT_TRUNCATED |
12336 | && CONST_INT_P (XEXP (op0, 0)) |
12337 | && XEXP (op0, 1) == const1_rtx |
12338 | && equality_comparison_p && const_op == 0 |
12339 | && (i = exact_log2 (UINTVAL (XEXP (op0, 0)))) >= 0) |
12340 | { |
12341 | if (BITS_BIG_ENDIAN) |
12342 | i = BITS_PER_WORD - 1 - i; |
12343 | |
12344 | op0 = XEXP (op0, 2); |
12345 | op1 = GEN_INT (i); |
12346 | const_op = i; |
12347 | |
12348 | /* Result is nonzero iff shift count is equal to I. */ |
12349 | code = reverse_condition (code); |
12350 | continue; |
12351 | } |
12352 | |
12353 | /* fall through */ |
12354 | |
12355 | case SIGN_EXTRACT: |
12356 | tem = expand_compound_operation (x: op0); |
12357 | if (tem != op0) |
12358 | { |
12359 | op0 = tem; |
12360 | continue; |
12361 | } |
12362 | break; |
12363 | |
12364 | case NOT: |
12365 | /* If testing for equality, we can take the NOT of the constant. */ |
12366 | if (equality_comparison_p |
12367 | && (tem = simplify_unary_operation (code: NOT, mode, op: op1, op_mode: mode)) != 0) |
12368 | { |
12369 | op0 = XEXP (op0, 0); |
12370 | op1 = tem; |
12371 | continue; |
12372 | } |
12373 | |
12374 | /* If just looking at the sign bit, reverse the sense of the |
12375 | comparison. */ |
12376 | if (sign_bit_comparison_p) |
12377 | { |
12378 | op0 = XEXP (op0, 0); |
12379 | code = (code == GE ? LT : GE); |
12380 | continue; |
12381 | } |
12382 | break; |
12383 | |
12384 | case NEG: |
12385 | /* If testing for equality, we can take the NEG of the constant. */ |
12386 | if (equality_comparison_p |
12387 | && (tem = simplify_unary_operation (code: NEG, mode, op: op1, op_mode: mode)) != 0) |
12388 | { |
12389 | op0 = XEXP (op0, 0); |
12390 | op1 = tem; |
12391 | continue; |
12392 | } |
12393 | |
12394 | /* The remaining cases only apply to comparisons with zero. */ |
12395 | if (const_op != 0) |
12396 | break; |
12397 | |
12398 | /* When X is ABS or is known positive, |
12399 | (neg X) is < 0 if and only if X != 0. */ |
12400 | |
12401 | if (sign_bit_comparison_p |
12402 | && (GET_CODE (XEXP (op0, 0)) == ABS |
12403 | || (mode_width <= HOST_BITS_PER_WIDE_INT |
12404 | && (nonzero_bits (XEXP (op0, 0), mode) |
12405 | & (HOST_WIDE_INT_1U << (mode_width - 1))) |
12406 | == 0))) |
12407 | { |
12408 | op0 = XEXP (op0, 0); |
12409 | code = (code == LT ? NE : EQ); |
12410 | continue; |
12411 | } |
12412 | |
12413 | /* If we have NEG of something whose two high-order bits are the |
12414 | same, we know that "(-a) < 0" is equivalent to "a > 0". */ |
12415 | if (num_sign_bit_copies (op0, mode) >= 2) |
12416 | { |
12417 | op0 = XEXP (op0, 0); |
12418 | code = swap_condition (code); |
12419 | continue; |
12420 | } |
12421 | break; |
12422 | |
12423 | case ROTATE: |
12424 | /* If we are testing equality and our count is a constant, we |
12425 | can perform the inverse operation on our RHS. */ |
12426 | if (equality_comparison_p && CONST_INT_P (XEXP (op0, 1)) |
12427 | && (tem = simplify_binary_operation (code: ROTATERT, mode, |
12428 | op0: op1, XEXP (op0, 1))) != 0) |
12429 | { |
12430 | op0 = XEXP (op0, 0); |
12431 | op1 = tem; |
12432 | continue; |
12433 | } |
12434 | |
12435 | /* If we are doing a < 0 or >= 0 comparison, it means we are testing |
12436 | a particular bit. Convert it to an AND of a constant of that |
12437 | bit. This will be converted into a ZERO_EXTRACT. */ |
12438 | if (const_op == 0 && sign_bit_comparison_p |
12439 | && CONST_INT_P (XEXP (op0, 1)) |
12440 | && mode_width <= HOST_BITS_PER_WIDE_INT |
12441 | && UINTVAL (XEXP (op0, 1)) < mode_width) |
12442 | { |
12443 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), |
12444 | constop: (HOST_WIDE_INT_1U |
12445 | << (mode_width - 1 |
12446 | - INTVAL (XEXP (op0, 1))))); |
12447 | code = (code == LT ? NE : EQ); |
12448 | continue; |
12449 | } |
12450 | |
12451 | /* Fall through. */ |
12452 | |
12453 | case ABS: |
12454 | /* ABS is ignorable inside an equality comparison with zero. */ |
12455 | if (const_op == 0 && equality_comparison_p) |
12456 | { |
12457 | op0 = XEXP (op0, 0); |
12458 | continue; |
12459 | } |
12460 | break; |
12461 | |
12462 | case SIGN_EXTEND: |
12463 | /* Can simplify (compare (zero/sign_extend FOO) CONST) to |
12464 | (compare FOO CONST) if CONST fits in FOO's mode and we |
12465 | are either testing inequality or have an unsigned |
12466 | comparison with ZERO_EXTEND or a signed comparison with |
12467 | SIGN_EXTEND. But don't do it if we don't have a compare |
12468 | insn of the given mode, since we'd have to revert it |
12469 | later on, and then we wouldn't know whether to sign- or |
12470 | zero-extend. */ |
12471 | if (is_int_mode (GET_MODE (XEXP (op0, 0)), int_mode: &mode) |
12472 | && ! unsigned_comparison_p |
12473 | && HWI_COMPUTABLE_MODE_P (mode) |
12474 | && trunc_int_for_mode (const_op, mode) == const_op |
12475 | && have_insn_for (COMPARE, mode)) |
12476 | { |
12477 | op0 = XEXP (op0, 0); |
12478 | continue; |
12479 | } |
12480 | break; |
12481 | |
12482 | case SUBREG: |
12483 | /* Check for the case where we are comparing A - C1 with C2, that is |
12484 | |
12485 | (subreg:MODE (plus (A) (-C1))) op (C2) |
12486 | |
12487 | with C1 a constant, and try to lift the SUBREG, i.e. to do the |
12488 | comparison in the wider mode. One of the following two conditions |
12489 | must be true in order for this to be valid: |
12490 | |
12491 | 1. The mode extension results in the same bit pattern being added |
12492 | on both sides and the comparison is equality or unsigned. As |
12493 | C2 has been truncated to fit in MODE, the pattern can only be |
12494 | all 0s or all 1s. |
12495 | |
12496 | 2. The mode extension results in the sign bit being copied on |
12497 | each side. |
12498 | |
12499 | The difficulty here is that we have predicates for A but not for |
12500 | (A - C1) so we need to check that C1 is within proper bounds so |
12501 | as to perturbate A as little as possible. */ |
12502 | |
12503 | if (mode_width <= HOST_BITS_PER_WIDE_INT |
12504 | && subreg_lowpart_p (op0) |
12505 | && is_a <scalar_int_mode> (GET_MODE (SUBREG_REG (op0)), |
12506 | result: &inner_mode) |
12507 | && GET_MODE_PRECISION (mode: inner_mode) > mode_width |
12508 | && GET_CODE (SUBREG_REG (op0)) == PLUS |
12509 | && CONST_INT_P (XEXP (SUBREG_REG (op0), 1))) |
12510 | { |
12511 | rtx a = XEXP (SUBREG_REG (op0), 0); |
12512 | HOST_WIDE_INT c1 = -INTVAL (XEXP (SUBREG_REG (op0), 1)); |
12513 | |
12514 | if ((c1 > 0 |
12515 | && (unsigned HOST_WIDE_INT) c1 |
12516 | < HOST_WIDE_INT_1U << (mode_width - 1) |
12517 | && (equality_comparison_p || unsigned_comparison_p) |
12518 | /* (A - C1) zero-extends if it is positive and sign-extends |
12519 | if it is negative, C2 both zero- and sign-extends. */ |
12520 | && (((nonzero_bits (a, inner_mode) |
12521 | & ~GET_MODE_MASK (mode)) == 0 |
12522 | && const_op >= 0) |
12523 | /* (A - C1) sign-extends if it is positive and 1-extends |
12524 | if it is negative, C2 both sign- and 1-extends. */ |
12525 | || (num_sign_bit_copies (a, inner_mode) |
12526 | > (unsigned int) (GET_MODE_PRECISION (mode: inner_mode) |
12527 | - mode_width) |
12528 | && const_op < 0))) |
12529 | || ((unsigned HOST_WIDE_INT) c1 |
12530 | < HOST_WIDE_INT_1U << (mode_width - 2) |
12531 | /* (A - C1) always sign-extends, like C2. */ |
12532 | && num_sign_bit_copies (a, inner_mode) |
12533 | > (unsigned int) (GET_MODE_PRECISION (mode: inner_mode) |
12534 | - (mode_width - 1)))) |
12535 | { |
12536 | op0 = SUBREG_REG (op0); |
12537 | continue; |
12538 | } |
12539 | } |
12540 | |
12541 | /* If the inner mode is narrower and we are extracting the low part, |
12542 | we can treat the SUBREG as if it were a ZERO_EXTEND. */ |
12543 | if (paradoxical_subreg_p (x: op0)) |
12544 | ; |
12545 | else if (subreg_lowpart_p (op0) |
12546 | && GET_MODE_CLASS (mode) == MODE_INT |
12547 | && is_int_mode (GET_MODE (SUBREG_REG (op0)), int_mode: &inner_mode) |
12548 | && (code == NE || code == EQ) |
12549 | && GET_MODE_PRECISION (mode: inner_mode) <= HOST_BITS_PER_WIDE_INT |
12550 | && !paradoxical_subreg_p (x: op0) |
12551 | && (nonzero_bits (SUBREG_REG (op0), inner_mode) |
12552 | & ~GET_MODE_MASK (mode)) == 0) |
12553 | { |
12554 | /* Remove outer subregs that don't do anything. */ |
12555 | tem = gen_lowpart (inner_mode, op1); |
12556 | |
12557 | if ((nonzero_bits (tem, inner_mode) |
12558 | & ~GET_MODE_MASK (mode)) == 0) |
12559 | { |
12560 | op0 = SUBREG_REG (op0); |
12561 | op1 = tem; |
12562 | continue; |
12563 | } |
12564 | break; |
12565 | } |
12566 | else |
12567 | break; |
12568 | |
12569 | /* FALLTHROUGH */ |
12570 | |
12571 | case ZERO_EXTEND: |
12572 | if (is_int_mode (GET_MODE (XEXP (op0, 0)), int_mode: &mode) |
12573 | && (unsigned_comparison_p || equality_comparison_p) |
12574 | && HWI_COMPUTABLE_MODE_P (mode) |
12575 | && (unsigned HOST_WIDE_INT) const_op <= GET_MODE_MASK (mode) |
12576 | && const_op >= 0 |
12577 | && have_insn_for (COMPARE, mode)) |
12578 | { |
12579 | op0 = XEXP (op0, 0); |
12580 | continue; |
12581 | } |
12582 | break; |
12583 | |
12584 | case PLUS: |
12585 | /* (eq (plus X A) B) -> (eq X (minus B A)). We can only do |
12586 | this for equality comparisons due to pathological cases involving |
12587 | overflows. */ |
12588 | if (equality_comparison_p |
12589 | && (tem = simplify_binary_operation (code: MINUS, mode, |
12590 | op0: op1, XEXP (op0, 1))) != 0) |
12591 | { |
12592 | op0 = XEXP (op0, 0); |
12593 | op1 = tem; |
12594 | continue; |
12595 | } |
12596 | |
12597 | /* (plus (abs X) (const_int -1)) is < 0 if and only if X == 0. */ |
12598 | if (const_op == 0 && XEXP (op0, 1) == constm1_rtx |
12599 | && GET_CODE (XEXP (op0, 0)) == ABS && sign_bit_comparison_p) |
12600 | { |
12601 | op0 = XEXP (XEXP (op0, 0), 0); |
12602 | code = (code == LT ? EQ : NE); |
12603 | continue; |
12604 | } |
12605 | break; |
12606 | |
12607 | case MINUS: |
12608 | /* We used to optimize signed comparisons against zero, but that |
12609 | was incorrect. Unsigned comparisons against zero (GTU, LEU) |
12610 | arrive here as equality comparisons, or (GEU, LTU) are |
12611 | optimized away. No need to special-case them. */ |
12612 | |
12613 | /* (eq (minus A B) C) -> (eq A (plus B C)) or |
12614 | (eq B (minus A C)), whichever simplifies. We can only do |
12615 | this for equality comparisons due to pathological cases involving |
12616 | overflows. */ |
12617 | if (equality_comparison_p |
12618 | && (tem = simplify_binary_operation (code: PLUS, mode, |
12619 | XEXP (op0, 1), op1)) != 0) |
12620 | { |
12621 | op0 = XEXP (op0, 0); |
12622 | op1 = tem; |
12623 | continue; |
12624 | } |
12625 | |
12626 | if (equality_comparison_p |
12627 | && (tem = simplify_binary_operation (code: MINUS, mode, |
12628 | XEXP (op0, 0), op1)) != 0) |
12629 | { |
12630 | op0 = XEXP (op0, 1); |
12631 | op1 = tem; |
12632 | continue; |
12633 | } |
12634 | |
12635 | /* The sign bit of (minus (ashiftrt X C) X), where C is the number |
12636 | of bits in X minus 1, is one iff X > 0. */ |
12637 | if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == ASHIFTRT |
12638 | && CONST_INT_P (XEXP (XEXP (op0, 0), 1)) |
12639 | && UINTVAL (XEXP (XEXP (op0, 0), 1)) == mode_width - 1 |
12640 | && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1))) |
12641 | { |
12642 | op0 = XEXP (op0, 1); |
12643 | code = (code == GE ? LE : GT); |
12644 | continue; |
12645 | } |
12646 | break; |
12647 | |
12648 | case XOR: |
12649 | /* (eq (xor A B) C) -> (eq A (xor B C)). This is a simplification |
12650 | if C is zero or B is a constant. */ |
12651 | if (equality_comparison_p |
12652 | && (tem = simplify_binary_operation (code: XOR, mode, |
12653 | XEXP (op0, 1), op1)) != 0) |
12654 | { |
12655 | op0 = XEXP (op0, 0); |
12656 | op1 = tem; |
12657 | continue; |
12658 | } |
12659 | break; |
12660 | |
12661 | |
12662 | case IOR: |
12663 | /* The sign bit of (ior (plus X (const_int -1)) X) is nonzero |
12664 | iff X <= 0. */ |
12665 | if (sign_bit_comparison_p && GET_CODE (XEXP (op0, 0)) == PLUS |
12666 | && XEXP (XEXP (op0, 0), 1) == constm1_rtx |
12667 | && rtx_equal_p (XEXP (XEXP (op0, 0), 0), XEXP (op0, 1))) |
12668 | { |
12669 | op0 = XEXP (op0, 1); |
12670 | code = (code == GE ? GT : LE); |
12671 | continue; |
12672 | } |
12673 | break; |
12674 | |
12675 | case AND: |
12676 | /* Convert (and (xshift 1 X) Y) to (and (lshiftrt Y X) 1). This |
12677 | will be converted to a ZERO_EXTRACT later. */ |
12678 | if (const_op == 0 && equality_comparison_p |
12679 | && GET_CODE (XEXP (op0, 0)) == ASHIFT |
12680 | && XEXP (XEXP (op0, 0), 0) == const1_rtx) |
12681 | { |
12682 | op0 = gen_rtx_LSHIFTRT (mode, XEXP (op0, 1), |
12683 | XEXP (XEXP (op0, 0), 1)); |
12684 | op0 = simplify_and_const_int (NULL_RTX, mode, varop: op0, constop: 1); |
12685 | continue; |
12686 | } |
12687 | |
12688 | /* If we are comparing (and (lshiftrt X C1) C2) for equality with |
12689 | zero and X is a comparison and C1 and C2 describe only bits set |
12690 | in STORE_FLAG_VALUE, we can compare with X. */ |
12691 | if (const_op == 0 && equality_comparison_p |
12692 | && mode_width <= HOST_BITS_PER_WIDE_INT |
12693 | && CONST_INT_P (XEXP (op0, 1)) |
12694 | && GET_CODE (XEXP (op0, 0)) == LSHIFTRT |
12695 | && CONST_INT_P (XEXP (XEXP (op0, 0), 1)) |
12696 | && INTVAL (XEXP (XEXP (op0, 0), 1)) >= 0 |
12697 | && INTVAL (XEXP (XEXP (op0, 0), 1)) < HOST_BITS_PER_WIDE_INT) |
12698 | { |
12699 | mask = ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode)) |
12700 | << INTVAL (XEXP (XEXP (op0, 0), 1))); |
12701 | if ((~STORE_FLAG_VALUE & mask) == 0 |
12702 | && (COMPARISON_P (XEXP (XEXP (op0, 0), 0)) |
12703 | || ((tem = get_last_value (XEXP (XEXP (op0, 0), 0))) != 0 |
12704 | && COMPARISON_P (tem)))) |
12705 | { |
12706 | op0 = XEXP (XEXP (op0, 0), 0); |
12707 | continue; |
12708 | } |
12709 | } |
12710 | |
12711 | /* If we are doing an equality comparison of an AND of a bit equal |
12712 | to the sign bit, replace this with a LT or GE comparison of |
12713 | the underlying value. */ |
12714 | if (equality_comparison_p |
12715 | && const_op == 0 |
12716 | && CONST_INT_P (XEXP (op0, 1)) |
12717 | && mode_width <= HOST_BITS_PER_WIDE_INT |
12718 | && ((INTVAL (XEXP (op0, 1)) & GET_MODE_MASK (mode)) |
12719 | == HOST_WIDE_INT_1U << (mode_width - 1))) |
12720 | { |
12721 | op0 = XEXP (op0, 0); |
12722 | code = (code == EQ ? GE : LT); |
12723 | continue; |
12724 | } |
12725 | |
12726 | /* If this AND operation is really a ZERO_EXTEND from a narrower |
12727 | mode, the constant fits within that mode, and this is either an |
12728 | equality or unsigned comparison, try to do this comparison in |
12729 | the narrower mode. |
12730 | |
12731 | Note that in: |
12732 | |
12733 | (ne:DI (and:DI (reg:DI 4) (const_int 0xffffffff)) (const_int 0)) |
12734 | -> (ne:DI (reg:SI 4) (const_int 0)) |
12735 | |
12736 | unless TARGET_TRULY_NOOP_TRUNCATION allows it or the register is |
12737 | known to hold a value of the required mode the |
12738 | transformation is invalid. */ |
12739 | if ((equality_comparison_p || unsigned_comparison_p) |
12740 | && CONST_INT_P (XEXP (op0, 1)) |
12741 | && (i = exact_log2 (x: (UINTVAL (XEXP (op0, 1)) |
12742 | & GET_MODE_MASK (mode)) |
12743 | + 1)) >= 0 |
12744 | && const_op >> i == 0 |
12745 | && int_mode_for_size (size: i, limit: 1).exists (mode: &tmode)) |
12746 | { |
12747 | op0 = gen_lowpart_or_truncate (mode: tmode, XEXP (op0, 0)); |
12748 | continue; |
12749 | } |
12750 | |
12751 | /* If this is (and:M1 (subreg:M1 X:M2 0) (const_int C1)) where C1 |
12752 | fits in both M1 and M2 and the SUBREG is either paradoxical |
12753 | or represents the low part, permute the SUBREG and the AND |
12754 | and try again. */ |
12755 | if (GET_CODE (XEXP (op0, 0)) == SUBREG |
12756 | && CONST_INT_P (XEXP (op0, 1))) |
12757 | { |
12758 | unsigned HOST_WIDE_INT c1 = INTVAL (XEXP (op0, 1)); |
12759 | /* Require an integral mode, to avoid creating something like |
12760 | (AND:SF ...). */ |
12761 | if ((is_a <scalar_int_mode> |
12762 | (GET_MODE (SUBREG_REG (XEXP (op0, 0))), result: &tmode)) |
12763 | /* It is unsafe to commute the AND into the SUBREG if the |
12764 | SUBREG is paradoxical and WORD_REGISTER_OPERATIONS is |
12765 | not defined. As originally written the upper bits |
12766 | have a defined value due to the AND operation. |
12767 | However, if we commute the AND inside the SUBREG then |
12768 | they no longer have defined values and the meaning of |
12769 | the code has been changed. |
12770 | Also C1 should not change value in the smaller mode, |
12771 | see PR67028 (a positive C1 can become negative in the |
12772 | smaller mode, so that the AND does no longer mask the |
12773 | upper bits). */ |
12774 | && ((WORD_REGISTER_OPERATIONS |
12775 | && mode_width > GET_MODE_PRECISION (mode: tmode) |
12776 | && mode_width <= BITS_PER_WORD |
12777 | && trunc_int_for_mode (c1, tmode) == (HOST_WIDE_INT) c1) |
12778 | || (mode_width <= GET_MODE_PRECISION (mode: tmode) |
12779 | && subreg_lowpart_p (XEXP (op0, 0)))) |
12780 | && mode_width <= HOST_BITS_PER_WIDE_INT |
12781 | && HWI_COMPUTABLE_MODE_P (mode: tmode) |
12782 | && (c1 & ~mask) == 0 |
12783 | && (c1 & ~GET_MODE_MASK (tmode)) == 0 |
12784 | && c1 != mask |
12785 | && c1 != GET_MODE_MASK (tmode)) |
12786 | { |
12787 | op0 = simplify_gen_binary (code: AND, mode: tmode, |
12788 | SUBREG_REG (XEXP (op0, 0)), |
12789 | op1: gen_int_mode (c1, tmode)); |
12790 | op0 = gen_lowpart (mode, op0); |
12791 | continue; |
12792 | } |
12793 | } |
12794 | |
12795 | /* Convert (ne (and (not X) 1) 0) to (eq (and X 1) 0). */ |
12796 | if (const_op == 0 && equality_comparison_p |
12797 | && XEXP (op0, 1) == const1_rtx |
12798 | && GET_CODE (XEXP (op0, 0)) == NOT) |
12799 | { |
12800 | op0 = simplify_and_const_int (NULL_RTX, mode, |
12801 | XEXP (XEXP (op0, 0), 0), constop: 1); |
12802 | code = (code == NE ? EQ : NE); |
12803 | continue; |
12804 | } |
12805 | |
12806 | /* Convert (ne (and (lshiftrt (not X)) 1) 0) to |
12807 | (eq (and (lshiftrt X) 1) 0). |
12808 | Also handle the case where (not X) is expressed using xor. */ |
12809 | if (const_op == 0 && equality_comparison_p |
12810 | && XEXP (op0, 1) == const1_rtx |
12811 | && GET_CODE (XEXP (op0, 0)) == LSHIFTRT) |
12812 | { |
12813 | rtx shift_op = XEXP (XEXP (op0, 0), 0); |
12814 | rtx shift_count = XEXP (XEXP (op0, 0), 1); |
12815 | |
12816 | if (GET_CODE (shift_op) == NOT |
12817 | || (GET_CODE (shift_op) == XOR |
12818 | && CONST_INT_P (XEXP (shift_op, 1)) |
12819 | && CONST_INT_P (shift_count) |
12820 | && HWI_COMPUTABLE_MODE_P (mode) |
12821 | && (UINTVAL (XEXP (shift_op, 1)) |
12822 | == HOST_WIDE_INT_1U |
12823 | << INTVAL (shift_count)))) |
12824 | { |
12825 | op0 |
12826 | = gen_rtx_LSHIFTRT (mode, XEXP (shift_op, 0), shift_count); |
12827 | op0 = simplify_and_const_int (NULL_RTX, mode, varop: op0, constop: 1); |
12828 | code = (code == NE ? EQ : NE); |
12829 | continue; |
12830 | } |
12831 | } |
12832 | break; |
12833 | |
12834 | case ASHIFT: |
12835 | /* If we have (compare (ashift FOO N) (const_int C)) and |
12836 | the high order N bits of FOO (N+1 if an inequality comparison) |
12837 | are known to be zero, we can do this by comparing FOO with C |
12838 | shifted right N bits so long as the low-order N bits of C are |
12839 | zero. */ |
12840 | if (CONST_INT_P (XEXP (op0, 1)) |
12841 | && INTVAL (XEXP (op0, 1)) >= 0 |
12842 | && ((INTVAL (XEXP (op0, 1)) + ! equality_comparison_p) |
12843 | < HOST_BITS_PER_WIDE_INT) |
12844 | && (((unsigned HOST_WIDE_INT) const_op |
12845 | & ((HOST_WIDE_INT_1U << INTVAL (XEXP (op0, 1))) |
12846 | - 1)) == 0) |
12847 | && mode_width <= HOST_BITS_PER_WIDE_INT |
12848 | && (nonzero_bits (XEXP (op0, 0), mode) |
12849 | & ~(mask >> (INTVAL (XEXP (op0, 1)) |
12850 | + ! equality_comparison_p))) == 0) |
12851 | { |
12852 | /* We must perform a logical shift, not an arithmetic one, |
12853 | as we want the top N bits of C to be zero. */ |
12854 | unsigned HOST_WIDE_INT temp = const_op & GET_MODE_MASK (mode); |
12855 | |
12856 | temp >>= INTVAL (XEXP (op0, 1)); |
12857 | op1 = gen_int_mode (temp, mode); |
12858 | op0 = XEXP (op0, 0); |
12859 | continue; |
12860 | } |
12861 | |
12862 | /* If we are doing a sign bit comparison, it means we are testing |
12863 | a particular bit. Convert it to the appropriate AND. */ |
12864 | if (sign_bit_comparison_p && CONST_INT_P (XEXP (op0, 1)) |
12865 | && mode_width <= HOST_BITS_PER_WIDE_INT) |
12866 | { |
12867 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), |
12868 | constop: (HOST_WIDE_INT_1U |
12869 | << (mode_width - 1 |
12870 | - INTVAL (XEXP (op0, 1))))); |
12871 | code = (code == LT ? NE : EQ); |
12872 | continue; |
12873 | } |
12874 | |
12875 | /* If this an equality comparison with zero and we are shifting |
12876 | the low bit to the sign bit, we can convert this to an AND of the |
12877 | low-order bit. */ |
12878 | if (const_op == 0 && equality_comparison_p |
12879 | && CONST_INT_P (XEXP (op0, 1)) |
12880 | && UINTVAL (XEXP (op0, 1)) == mode_width - 1) |
12881 | { |
12882 | op0 = simplify_and_const_int (NULL_RTX, mode, XEXP (op0, 0), constop: 1); |
12883 | continue; |
12884 | } |
12885 | break; |
12886 | |
12887 | case ASHIFTRT: |
12888 | /* If this is an equality comparison with zero, we can do this |
12889 | as a logical shift, which might be much simpler. */ |
12890 | if (equality_comparison_p && const_op == 0 |
12891 | && CONST_INT_P (XEXP (op0, 1))) |
12892 | { |
12893 | op0 = simplify_shift_const (NULL_RTX, code: LSHIFTRT, result_mode: mode, |
12894 | XEXP (op0, 0), |
12895 | INTVAL (XEXP (op0, 1))); |
12896 | continue; |
12897 | } |
12898 | |
12899 | /* If OP0 is a sign extension and CODE is not an unsigned comparison, |
12900 | do the comparison in a narrower mode. */ |
12901 | if (! unsigned_comparison_p |
12902 | && CONST_INT_P (XEXP (op0, 1)) |
12903 | && GET_CODE (XEXP (op0, 0)) == ASHIFT |
12904 | && XEXP (op0, 1) == XEXP (XEXP (op0, 0), 1) |
12905 | && (int_mode_for_size (size: mode_width - INTVAL (XEXP (op0, 1)), limit: 1) |
12906 | .exists (mode: &tmode)) |
12907 | && (((unsigned HOST_WIDE_INT) const_op |
12908 | + (GET_MODE_MASK (tmode) >> 1) + 1) |
12909 | <= GET_MODE_MASK (tmode))) |
12910 | { |
12911 | op0 = gen_lowpart (tmode, XEXP (XEXP (op0, 0), 0)); |
12912 | continue; |
12913 | } |
12914 | |
12915 | /* Likewise if OP0 is a PLUS of a sign extension with a |
12916 | constant, which is usually represented with the PLUS |
12917 | between the shifts. */ |
12918 | if (! unsigned_comparison_p |
12919 | && CONST_INT_P (XEXP (op0, 1)) |
12920 | && GET_CODE (XEXP (op0, 0)) == PLUS |
12921 | && CONST_INT_P (XEXP (XEXP (op0, 0), 1)) |
12922 | && GET_CODE (XEXP (XEXP (op0, 0), 0)) == ASHIFT |
12923 | && XEXP (op0, 1) == XEXP (XEXP (XEXP (op0, 0), 0), 1) |
12924 | && (int_mode_for_size (size: mode_width - INTVAL (XEXP (op0, 1)), limit: 1) |
12925 | .exists (mode: &tmode)) |
12926 | && (((unsigned HOST_WIDE_INT) const_op |
12927 | + (GET_MODE_MASK (tmode) >> 1) + 1) |
12928 | <= GET_MODE_MASK (tmode))) |
12929 | { |
12930 | rtx inner = XEXP (XEXP (XEXP (op0, 0), 0), 0); |
12931 | rtx add_const = XEXP (XEXP (op0, 0), 1); |
12932 | rtx new_const = simplify_gen_binary (code: ASHIFTRT, mode, |
12933 | op0: add_const, XEXP (op0, 1)); |
12934 | |
12935 | op0 = simplify_gen_binary (code: PLUS, mode: tmode, |
12936 | gen_lowpart (tmode, inner), |
12937 | op1: new_const); |
12938 | continue; |
12939 | } |
12940 | |
12941 | /* FALLTHROUGH */ |
12942 | case LSHIFTRT: |
12943 | /* If we have (compare (xshiftrt FOO N) (const_int C)) and |
12944 | the low order N bits of FOO are known to be zero, we can do this |
12945 | by comparing FOO with C shifted left N bits so long as no |
12946 | overflow occurs. Even if the low order N bits of FOO aren't known |
12947 | to be zero, if the comparison is >= or < we can use the same |
12948 | optimization and for > or <= by setting all the low |
12949 | order N bits in the comparison constant. */ |
12950 | if (CONST_INT_P (XEXP (op0, 1)) |
12951 | && INTVAL (XEXP (op0, 1)) > 0 |
12952 | && INTVAL (XEXP (op0, 1)) < HOST_BITS_PER_WIDE_INT |
12953 | && mode_width <= HOST_BITS_PER_WIDE_INT |
12954 | && (((unsigned HOST_WIDE_INT) const_op |
12955 | + (GET_CODE (op0) != LSHIFTRT |
12956 | ? ((GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1)) >> 1) |
12957 | + 1) |
12958 | : 0)) |
12959 | <= GET_MODE_MASK (mode) >> INTVAL (XEXP (op0, 1)))) |
12960 | { |
12961 | unsigned HOST_WIDE_INT low_bits |
12962 | = (nonzero_bits (XEXP (op0, 0), mode) |
12963 | & ((HOST_WIDE_INT_1U |
12964 | << INTVAL (XEXP (op0, 1))) - 1)); |
12965 | if (low_bits == 0 || !equality_comparison_p) |
12966 | { |
12967 | /* If the shift was logical, then we must make the condition |
12968 | unsigned. */ |
12969 | if (GET_CODE (op0) == LSHIFTRT) |
12970 | code = unsigned_condition (code); |
12971 | |
12972 | const_op = (unsigned HOST_WIDE_INT) const_op |
12973 | << INTVAL (XEXP (op0, 1)); |
12974 | if (low_bits != 0 |
12975 | && (code == GT || code == GTU |
12976 | || code == LE || code == LEU)) |
12977 | const_op |
12978 | |= ((HOST_WIDE_INT_1 << INTVAL (XEXP (op0, 1))) - 1); |
12979 | op1 = GEN_INT (const_op); |
12980 | op0 = XEXP (op0, 0); |
12981 | continue; |
12982 | } |
12983 | } |
12984 | |
12985 | /* If we are using this shift to extract just the sign bit, we |
12986 | can replace this with an LT or GE comparison. */ |
12987 | if (const_op == 0 |
12988 | && (equality_comparison_p || sign_bit_comparison_p) |
12989 | && CONST_INT_P (XEXP (op0, 1)) |
12990 | && UINTVAL (XEXP (op0, 1)) == mode_width - 1) |
12991 | { |
12992 | op0 = XEXP (op0, 0); |
12993 | code = (code == NE || code == GT ? LT : GE); |
12994 | continue; |
12995 | } |
12996 | break; |
12997 | |
12998 | default: |
12999 | break; |
13000 | } |
13001 | |
13002 | break; |
13003 | } |
13004 | |
13005 | /* Now make any compound operations involved in this comparison. Then, |
13006 | check for an outmost SUBREG on OP0 that is not doing anything or is |
13007 | paradoxical. The latter transformation must only be performed when |
13008 | it is known that the "extra" bits will be the same in op0 and op1 or |
13009 | that they don't matter. There are three cases to consider: |
13010 | |
13011 | 1. SUBREG_REG (op0) is a register. In this case the bits are don't |
13012 | care bits and we can assume they have any convenient value. So |
13013 | making the transformation is safe. |
13014 | |
13015 | 2. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is UNKNOWN. |
13016 | In this case the upper bits of op0 are undefined. We should not make |
13017 | the simplification in that case as we do not know the contents of |
13018 | those bits. |
13019 | |
13020 | 3. SUBREG_REG (op0) is a memory and LOAD_EXTEND_OP is not UNKNOWN. |
13021 | In that case we know those bits are zeros or ones. We must also be |
13022 | sure that they are the same as the upper bits of op1. |
13023 | |
13024 | We can never remove a SUBREG for a non-equality comparison because |
13025 | the sign bit is in a different place in the underlying object. */ |
13026 | |
13027 | rtx_code op0_mco_code = SET; |
13028 | if (op1 == const0_rtx) |
13029 | op0_mco_code = code == NE || code == EQ ? EQ : COMPARE; |
13030 | |
13031 | op0 = make_compound_operation (x: op0, in_code: op0_mco_code); |
13032 | op1 = make_compound_operation (x: op1, in_code: SET); |
13033 | |
13034 | if (GET_CODE (op0) == SUBREG && subreg_lowpart_p (op0) |
13035 | && is_int_mode (GET_MODE (op0), int_mode: &mode) |
13036 | && is_int_mode (GET_MODE (SUBREG_REG (op0)), int_mode: &inner_mode) |
13037 | && (code == NE || code == EQ)) |
13038 | { |
13039 | if (paradoxical_subreg_p (x: op0)) |
13040 | { |
13041 | /* For paradoxical subregs, allow case 1 as above. Case 3 isn't |
13042 | implemented. */ |
13043 | if (REG_P (SUBREG_REG (op0))) |
13044 | { |
13045 | op0 = SUBREG_REG (op0); |
13046 | op1 = gen_lowpart (inner_mode, op1); |
13047 | } |
13048 | } |
13049 | else if (GET_MODE_PRECISION (mode: inner_mode) <= HOST_BITS_PER_WIDE_INT |
13050 | && (nonzero_bits (SUBREG_REG (op0), inner_mode) |
13051 | & ~GET_MODE_MASK (mode)) == 0) |
13052 | { |
13053 | tem = gen_lowpart (inner_mode, op1); |
13054 | |
13055 | if ((nonzero_bits (tem, inner_mode) & ~GET_MODE_MASK (mode)) == 0) |
13056 | op0 = SUBREG_REG (op0), op1 = tem; |
13057 | } |
13058 | } |
13059 | |
13060 | /* We now do the opposite procedure: Some machines don't have compare |
13061 | insns in all modes. If OP0's mode is an integer mode smaller than a |
13062 | word and we can't do a compare in that mode, see if there is a larger |
13063 | mode for which we can do the compare. There are a number of cases in |
13064 | which we can use the wider mode. */ |
13065 | |
13066 | if (is_int_mode (GET_MODE (op0), int_mode: &mode) |
13067 | && GET_MODE_SIZE (mode) < UNITS_PER_WORD |
13068 | && ! have_insn_for (COMPARE, mode)) |
13069 | FOR_EACH_WIDER_MODE (tmode_iter, mode) |
13070 | { |
13071 | tmode = tmode_iter.require (); |
13072 | if (!HWI_COMPUTABLE_MODE_P (mode: tmode)) |
13073 | break; |
13074 | if (have_insn_for (COMPARE, tmode)) |
13075 | { |
13076 | int zero_extended; |
13077 | |
13078 | /* If this is a test for negative, we can make an explicit |
13079 | test of the sign bit. Test this first so we can use |
13080 | a paradoxical subreg to extend OP0. */ |
13081 | |
13082 | if (op1 == const0_rtx && (code == LT || code == GE) |
13083 | && HWI_COMPUTABLE_MODE_P (mode)) |
13084 | { |
13085 | unsigned HOST_WIDE_INT sign |
13086 | = HOST_WIDE_INT_1U << (GET_MODE_BITSIZE (mode) - 1); |
13087 | op0 = simplify_gen_binary (code: AND, mode: tmode, |
13088 | gen_lowpart (tmode, op0), |
13089 | op1: gen_int_mode (sign, tmode)); |
13090 | code = (code == LT) ? NE : EQ; |
13091 | break; |
13092 | } |
13093 | |
13094 | /* If the only nonzero bits in OP0 and OP1 are those in the |
13095 | narrower mode and this is an equality or unsigned comparison, |
13096 | we can use the wider mode. Similarly for sign-extended |
13097 | values, in which case it is true for all comparisons. */ |
13098 | zero_extended = ((code == EQ || code == NE |
13099 | || code == GEU || code == GTU |
13100 | || code == LEU || code == LTU) |
13101 | && (nonzero_bits (op0, tmode) |
13102 | & ~GET_MODE_MASK (mode)) == 0 |
13103 | && ((CONST_INT_P (op1) |
13104 | || (nonzero_bits (op1, tmode) |
13105 | & ~GET_MODE_MASK (mode)) == 0))); |
13106 | |
13107 | if (zero_extended |
13108 | || ((num_sign_bit_copies (op0, tmode) |
13109 | > (unsigned int) (GET_MODE_PRECISION (mode: tmode) |
13110 | - GET_MODE_PRECISION (mode))) |
13111 | && (num_sign_bit_copies (op1, tmode) |
13112 | > (unsigned int) (GET_MODE_PRECISION (mode: tmode) |
13113 | - GET_MODE_PRECISION (mode))))) |
13114 | { |
13115 | /* If OP0 is an AND and we don't have an AND in MODE either, |
13116 | make a new AND in the proper mode. */ |
13117 | if (GET_CODE (op0) == AND |
13118 | && !have_insn_for (AND, mode)) |
13119 | op0 = simplify_gen_binary (code: AND, mode: tmode, |
13120 | gen_lowpart (tmode, |
13121 | XEXP (op0, 0)), |
13122 | gen_lowpart (tmode, |
13123 | XEXP (op0, 1))); |
13124 | else |
13125 | { |
13126 | if (zero_extended) |
13127 | { |
13128 | op0 = simplify_gen_unary (code: ZERO_EXTEND, mode: tmode, |
13129 | op: op0, op_mode: mode); |
13130 | op1 = simplify_gen_unary (code: ZERO_EXTEND, mode: tmode, |
13131 | op: op1, op_mode: mode); |
13132 | } |
13133 | else |
13134 | { |
13135 | op0 = simplify_gen_unary (code: SIGN_EXTEND, mode: tmode, |
13136 | op: op0, op_mode: mode); |
13137 | op1 = simplify_gen_unary (code: SIGN_EXTEND, mode: tmode, |
13138 | op: op1, op_mode: mode); |
13139 | } |
13140 | break; |
13141 | } |
13142 | } |
13143 | } |
13144 | } |
13145 | |
13146 | /* We may have changed the comparison operands. Re-canonicalize. */ |
13147 | if (swap_commutative_operands_p (op0, op1)) |
13148 | { |
13149 | std::swap (a&: op0, b&: op1); |
13150 | code = swap_condition (code); |
13151 | } |
13152 | |
13153 | /* If this machine only supports a subset of valid comparisons, see if we |
13154 | can convert an unsupported one into a supported one. */ |
13155 | target_canonicalize_comparison (code: &code, op0: &op0, op1: &op1, op0_preserve_value: 0); |
13156 | |
13157 | *pop0 = op0; |
13158 | *pop1 = op1; |
13159 | |
13160 | return code; |
13161 | } |
13162 | |
13163 | /* Utility function for record_value_for_reg. Count number of |
13164 | rtxs in X. */ |
13165 | static int |
13166 | count_rtxs (rtx x) |
13167 | { |
13168 | enum rtx_code code = GET_CODE (x); |
13169 | const char *fmt; |
13170 | int i, j, ret = 1; |
13171 | |
13172 | if (GET_RTX_CLASS (code) == RTX_BIN_ARITH |
13173 | || GET_RTX_CLASS (code) == RTX_COMM_ARITH) |
13174 | { |
13175 | rtx x0 = XEXP (x, 0); |
13176 | rtx x1 = XEXP (x, 1); |
13177 | |
13178 | if (x0 == x1) |
13179 | return 1 + 2 * count_rtxs (x: x0); |
13180 | |
13181 | if ((GET_RTX_CLASS (GET_CODE (x1)) == RTX_BIN_ARITH |
13182 | || GET_RTX_CLASS (GET_CODE (x1)) == RTX_COMM_ARITH) |
13183 | && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1))) |
13184 | return 2 + 2 * count_rtxs (x: x0) |
13185 | + count_rtxs (x: x == XEXP (x1, 0) |
13186 | ? XEXP (x1, 1) : XEXP (x1, 0)); |
13187 | |
13188 | if ((GET_RTX_CLASS (GET_CODE (x0)) == RTX_BIN_ARITH |
13189 | || GET_RTX_CLASS (GET_CODE (x0)) == RTX_COMM_ARITH) |
13190 | && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1))) |
13191 | return 2 + 2 * count_rtxs (x: x1) |
13192 | + count_rtxs (x: x == XEXP (x0, 0) |
13193 | ? XEXP (x0, 1) : XEXP (x0, 0)); |
13194 | } |
13195 | |
13196 | fmt = GET_RTX_FORMAT (code); |
13197 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
13198 | if (fmt[i] == 'e') |
13199 | ret += count_rtxs (XEXP (x, i)); |
13200 | else if (fmt[i] == 'E') |
13201 | for (j = 0; j < XVECLEN (x, i); j++) |
13202 | ret += count_rtxs (XVECEXP (x, i, j)); |
13203 | |
13204 | return ret; |
13205 | } |
13206 | |
13207 | /* Utility function for following routine. Called when X is part of a value |
13208 | being stored into last_set_value. Sets last_set_table_tick |
13209 | for each register mentioned. Similar to mention_regs in cse.cc */ |
13210 | |
13211 | static void |
13212 | update_table_tick (rtx x) |
13213 | { |
13214 | enum rtx_code code = GET_CODE (x); |
13215 | const char *fmt = GET_RTX_FORMAT (code); |
13216 | int i, j; |
13217 | |
13218 | if (code == REG) |
13219 | { |
13220 | unsigned int regno = REGNO (x); |
13221 | unsigned int endregno = END_REGNO (x); |
13222 | unsigned int r; |
13223 | |
13224 | for (r = regno; r < endregno; r++) |
13225 | { |
13226 | reg_stat_type *rsp = ®_stat[r]; |
13227 | rsp->last_set_table_tick = label_tick; |
13228 | } |
13229 | |
13230 | return; |
13231 | } |
13232 | |
13233 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
13234 | if (fmt[i] == 'e') |
13235 | { |
13236 | /* Check for identical subexpressions. If x contains |
13237 | identical subexpression we only have to traverse one of |
13238 | them. */ |
13239 | if (i == 0 && ARITHMETIC_P (x)) |
13240 | { |
13241 | /* Note that at this point x1 has already been |
13242 | processed. */ |
13243 | rtx x0 = XEXP (x, 0); |
13244 | rtx x1 = XEXP (x, 1); |
13245 | |
13246 | /* If x0 and x1 are identical then there is no need to |
13247 | process x0. */ |
13248 | if (x0 == x1) |
13249 | break; |
13250 | |
13251 | /* If x0 is identical to a subexpression of x1 then while |
13252 | processing x1, x0 has already been processed. Thus we |
13253 | are done with x. */ |
13254 | if (ARITHMETIC_P (x1) |
13255 | && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1))) |
13256 | break; |
13257 | |
13258 | /* If x1 is identical to a subexpression of x0 then we |
13259 | still have to process the rest of x0. */ |
13260 | if (ARITHMETIC_P (x0) |
13261 | && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1))) |
13262 | { |
13263 | update_table_tick (XEXP (x0, x1 == XEXP (x0, 0) ? 1 : 0)); |
13264 | break; |
13265 | } |
13266 | } |
13267 | |
13268 | update_table_tick (XEXP (x, i)); |
13269 | } |
13270 | else if (fmt[i] == 'E') |
13271 | for (j = 0; j < XVECLEN (x, i); j++) |
13272 | update_table_tick (XVECEXP (x, i, j)); |
13273 | } |
13274 | |
13275 | /* Record that REG is set to VALUE in insn INSN. If VALUE is zero, we |
13276 | are saying that the register is clobbered and we no longer know its |
13277 | value. If INSN is zero, don't update reg_stat[].last_set; this is |
13278 | only permitted with VALUE also zero and is used to invalidate the |
13279 | register. */ |
13280 | |
13281 | static void |
13282 | record_value_for_reg (rtx reg, rtx_insn *insn, rtx value) |
13283 | { |
13284 | unsigned int regno = REGNO (reg); |
13285 | unsigned int endregno = END_REGNO (x: reg); |
13286 | unsigned int i; |
13287 | reg_stat_type *rsp; |
13288 | |
13289 | /* If VALUE contains REG and we have a previous value for REG, substitute |
13290 | the previous value. */ |
13291 | if (value && insn && reg_overlap_mentioned_p (reg, value)) |
13292 | { |
13293 | rtx tem; |
13294 | |
13295 | /* Set things up so get_last_value is allowed to see anything set up to |
13296 | our insn. */ |
13297 | subst_low_luid = DF_INSN_LUID (insn); |
13298 | tem = get_last_value (reg); |
13299 | |
13300 | /* If TEM is simply a binary operation with two CLOBBERs as operands, |
13301 | it isn't going to be useful and will take a lot of time to process, |
13302 | so just use the CLOBBER. */ |
13303 | |
13304 | if (tem) |
13305 | { |
13306 | if (ARITHMETIC_P (tem) |
13307 | && GET_CODE (XEXP (tem, 0)) == CLOBBER |
13308 | && GET_CODE (XEXP (tem, 1)) == CLOBBER) |
13309 | tem = XEXP (tem, 0); |
13310 | else if (count_occurrences (value, reg, 1) >= 2) |
13311 | { |
13312 | /* If there are two or more occurrences of REG in VALUE, |
13313 | prevent the value from growing too much. */ |
13314 | if (count_rtxs (x: tem) > param_max_last_value_rtl) |
13315 | tem = gen_rtx_CLOBBER (GET_MODE (tem), const0_rtx); |
13316 | } |
13317 | |
13318 | value = replace_rtx (copy_rtx (value), reg, tem); |
13319 | } |
13320 | } |
13321 | |
13322 | /* For each register modified, show we don't know its value, that |
13323 | we don't know about its bitwise content, that its value has been |
13324 | updated, and that we don't know the location of the death of the |
13325 | register. */ |
13326 | for (i = regno; i < endregno; i++) |
13327 | { |
13328 | rsp = ®_stat[i]; |
13329 | |
13330 | if (insn) |
13331 | rsp->last_set = insn; |
13332 | |
13333 | rsp->last_set_value = 0; |
13334 | rsp->last_set_mode = VOIDmode; |
13335 | rsp->last_set_nonzero_bits = 0; |
13336 | rsp->last_set_sign_bit_copies = 0; |
13337 | rsp->last_death = 0; |
13338 | rsp->truncated_to_mode = VOIDmode; |
13339 | } |
13340 | |
13341 | /* Mark registers that are being referenced in this value. */ |
13342 | if (value) |
13343 | update_table_tick (x: value); |
13344 | |
13345 | /* Now update the status of each register being set. |
13346 | If someone is using this register in this block, set this register |
13347 | to invalid since we will get confused between the two lives in this |
13348 | basic block. This makes using this register always invalid. In cse, we |
13349 | scan the table to invalidate all entries using this register, but this |
13350 | is too much work for us. */ |
13351 | |
13352 | for (i = regno; i < endregno; i++) |
13353 | { |
13354 | rsp = ®_stat[i]; |
13355 | rsp->last_set_label = label_tick; |
13356 | if (!insn |
13357 | || (value && rsp->last_set_table_tick >= label_tick_ebb_start)) |
13358 | rsp->last_set_invalid = true; |
13359 | else |
13360 | rsp->last_set_invalid = false; |
13361 | } |
13362 | |
13363 | /* The value being assigned might refer to X (like in "x++;"). In that |
13364 | case, we must replace it with (clobber (const_int 0)) to prevent |
13365 | infinite loops. */ |
13366 | rsp = ®_stat[regno]; |
13367 | if (value && !get_last_value_validate (&value, insn, label_tick, false)) |
13368 | { |
13369 | value = copy_rtx (value); |
13370 | if (!get_last_value_validate (&value, insn, label_tick, true)) |
13371 | value = 0; |
13372 | } |
13373 | |
13374 | /* For the main register being modified, update the value, the mode, the |
13375 | nonzero bits, and the number of sign bit copies. */ |
13376 | |
13377 | rsp->last_set_value = value; |
13378 | |
13379 | if (value) |
13380 | { |
13381 | machine_mode mode = GET_MODE (reg); |
13382 | subst_low_luid = DF_INSN_LUID (insn); |
13383 | rsp->last_set_mode = mode; |
13384 | if (GET_MODE_CLASS (mode) == MODE_INT |
13385 | && HWI_COMPUTABLE_MODE_P (mode)) |
13386 | mode = nonzero_bits_mode; |
13387 | rsp->last_set_nonzero_bits = nonzero_bits (value, mode); |
13388 | rsp->last_set_sign_bit_copies |
13389 | = num_sign_bit_copies (value, GET_MODE (reg)); |
13390 | } |
13391 | } |
13392 | |
13393 | /* Called via note_stores from record_dead_and_set_regs to handle one |
13394 | SET or CLOBBER in an insn. DATA is the instruction in which the |
13395 | set is occurring. */ |
13396 | |
13397 | static void |
13398 | record_dead_and_set_regs_1 (rtx dest, const_rtx setter, void *data) |
13399 | { |
13400 | rtx_insn *record_dead_insn = (rtx_insn *) data; |
13401 | |
13402 | if (GET_CODE (dest) == SUBREG) |
13403 | dest = SUBREG_REG (dest); |
13404 | |
13405 | if (!record_dead_insn) |
13406 | { |
13407 | if (REG_P (dest)) |
13408 | record_value_for_reg (reg: dest, NULL, NULL_RTX); |
13409 | return; |
13410 | } |
13411 | |
13412 | if (REG_P (dest)) |
13413 | { |
13414 | /* If we are setting the whole register, we know its value. */ |
13415 | if (GET_CODE (setter) == SET && dest == SET_DEST (setter)) |
13416 | record_value_for_reg (reg: dest, insn: record_dead_insn, SET_SRC (setter)); |
13417 | /* We can handle a SUBREG if it's the low part, but we must be |
13418 | careful with paradoxical SUBREGs on RISC architectures because |
13419 | we cannot strip e.g. an extension around a load and record the |
13420 | naked load since the RTL middle-end considers that the upper bits |
13421 | are defined according to LOAD_EXTEND_OP. */ |
13422 | else if (GET_CODE (setter) == SET |
13423 | && GET_CODE (SET_DEST (setter)) == SUBREG |
13424 | && SUBREG_REG (SET_DEST (setter)) == dest |
13425 | && known_le (GET_MODE_PRECISION (GET_MODE (dest)), |
13426 | BITS_PER_WORD) |
13427 | && subreg_lowpart_p (SET_DEST (setter))) |
13428 | { |
13429 | if (WORD_REGISTER_OPERATIONS |
13430 | && word_register_operation_p (SET_SRC (setter)) |
13431 | && paradoxical_subreg_p (SET_DEST (setter))) |
13432 | record_value_for_reg (reg: dest, insn: record_dead_insn, SET_SRC (setter)); |
13433 | else if (!partial_subreg_p (SET_DEST (setter))) |
13434 | record_value_for_reg (reg: dest, insn: record_dead_insn, |
13435 | gen_lowpart (GET_MODE (dest), |
13436 | SET_SRC (setter))); |
13437 | else |
13438 | { |
13439 | record_value_for_reg (reg: dest, insn: record_dead_insn, |
13440 | gen_lowpart (GET_MODE (dest), |
13441 | SET_SRC (setter))); |
13442 | |
13443 | unsigned HOST_WIDE_INT mask; |
13444 | reg_stat_type *rsp = ®_stat[REGNO (dest)]; |
13445 | mask = GET_MODE_MASK (GET_MODE (SET_DEST (setter))); |
13446 | rsp->last_set_nonzero_bits |= ~mask; |
13447 | rsp->last_set_sign_bit_copies = 1; |
13448 | } |
13449 | } |
13450 | /* Otherwise show that we don't know the value. */ |
13451 | else |
13452 | record_value_for_reg (reg: dest, insn: record_dead_insn, NULL_RTX); |
13453 | } |
13454 | else if (MEM_P (dest) |
13455 | /* Ignore pushes, they clobber nothing. */ |
13456 | && ! push_operand (dest, GET_MODE (dest))) |
13457 | mem_last_set = DF_INSN_LUID (record_dead_insn); |
13458 | } |
13459 | |
13460 | /* Update the records of when each REG was most recently set or killed |
13461 | for the things done by INSN. This is the last thing done in processing |
13462 | INSN in the combiner loop. |
13463 | |
13464 | We update reg_stat[], in particular fields last_set, last_set_value, |
13465 | last_set_mode, last_set_nonzero_bits, last_set_sign_bit_copies, |
13466 | last_death, and also the similar information mem_last_set (which insn |
13467 | most recently modified memory) and last_call_luid (which insn was the |
13468 | most recent subroutine call). */ |
13469 | |
13470 | static void |
13471 | record_dead_and_set_regs (rtx_insn *insn) |
13472 | { |
13473 | rtx link; |
13474 | unsigned int i; |
13475 | |
13476 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) |
13477 | { |
13478 | if (REG_NOTE_KIND (link) == REG_DEAD |
13479 | && REG_P (XEXP (link, 0))) |
13480 | { |
13481 | unsigned int regno = REGNO (XEXP (link, 0)); |
13482 | unsigned int endregno = END_REGNO (XEXP (link, 0)); |
13483 | |
13484 | for (i = regno; i < endregno; i++) |
13485 | { |
13486 | reg_stat_type *rsp; |
13487 | |
13488 | rsp = ®_stat[i]; |
13489 | rsp->last_death = insn; |
13490 | } |
13491 | } |
13492 | else if (REG_NOTE_KIND (link) == REG_INC) |
13493 | record_value_for_reg (XEXP (link, 0), insn, NULL_RTX); |
13494 | } |
13495 | |
13496 | if (CALL_P (insn)) |
13497 | { |
13498 | HARD_REG_SET callee_clobbers |
13499 | = insn_callee_abi (insn).full_and_partial_reg_clobbers (); |
13500 | hard_reg_set_iterator hrsi; |
13501 | EXECUTE_IF_SET_IN_HARD_REG_SET (callee_clobbers, 0, i, hrsi) |
13502 | { |
13503 | reg_stat_type *rsp; |
13504 | |
13505 | /* ??? We could try to preserve some information from the last |
13506 | set of register I if the call doesn't actually clobber |
13507 | (reg:last_set_mode I), which might be true for ABIs with |
13508 | partial clobbers. However, it would be difficult to |
13509 | update last_set_nonzero_bits and last_sign_bit_copies |
13510 | to account for the part of I that actually was clobbered. |
13511 | It wouldn't help much anyway, since we rarely see this |
13512 | situation before RA. */ |
13513 | rsp = ®_stat[i]; |
13514 | rsp->last_set_invalid = true; |
13515 | rsp->last_set = insn; |
13516 | rsp->last_set_value = 0; |
13517 | rsp->last_set_mode = VOIDmode; |
13518 | rsp->last_set_nonzero_bits = 0; |
13519 | rsp->last_set_sign_bit_copies = 0; |
13520 | rsp->last_death = 0; |
13521 | rsp->truncated_to_mode = VOIDmode; |
13522 | } |
13523 | |
13524 | last_call_luid = mem_last_set = DF_INSN_LUID (insn); |
13525 | |
13526 | /* We can't combine into a call pattern. Remember, though, that |
13527 | the return value register is set at this LUID. We could |
13528 | still replace a register with the return value from the |
13529 | wrong subroutine call! */ |
13530 | note_stores (insn, record_dead_and_set_regs_1, NULL_RTX); |
13531 | } |
13532 | else |
13533 | note_stores (insn, record_dead_and_set_regs_1, insn); |
13534 | } |
13535 | |
13536 | /* If a SUBREG has the promoted bit set, it is in fact a property of the |
13537 | register present in the SUBREG, so for each such SUBREG go back and |
13538 | adjust nonzero and sign bit information of the registers that are |
13539 | known to have some zero/sign bits set. |
13540 | |
13541 | This is needed because when combine blows the SUBREGs away, the |
13542 | information on zero/sign bits is lost and further combines can be |
13543 | missed because of that. */ |
13544 | |
13545 | static void |
13546 | record_promoted_value (rtx_insn *insn, rtx subreg) |
13547 | { |
13548 | struct insn_link *links; |
13549 | rtx set; |
13550 | unsigned int regno = REGNO (SUBREG_REG (subreg)); |
13551 | machine_mode mode = GET_MODE (subreg); |
13552 | |
13553 | if (!HWI_COMPUTABLE_MODE_P (mode)) |
13554 | return; |
13555 | |
13556 | for (links = LOG_LINKS (insn); links;) |
13557 | { |
13558 | reg_stat_type *rsp; |
13559 | |
13560 | insn = links->insn; |
13561 | set = single_set (insn); |
13562 | |
13563 | if (! set || !REG_P (SET_DEST (set)) |
13564 | || REGNO (SET_DEST (set)) != regno |
13565 | || GET_MODE (SET_DEST (set)) != GET_MODE (SUBREG_REG (subreg))) |
13566 | { |
13567 | links = links->next; |
13568 | continue; |
13569 | } |
13570 | |
13571 | rsp = ®_stat[regno]; |
13572 | if (rsp->last_set == insn) |
13573 | { |
13574 | if (SUBREG_PROMOTED_UNSIGNED_P (subreg)) |
13575 | rsp->last_set_nonzero_bits &= GET_MODE_MASK (mode); |
13576 | } |
13577 | |
13578 | if (REG_P (SET_SRC (set))) |
13579 | { |
13580 | regno = REGNO (SET_SRC (set)); |
13581 | links = LOG_LINKS (insn); |
13582 | } |
13583 | else |
13584 | break; |
13585 | } |
13586 | } |
13587 | |
13588 | /* Check if X, a register, is known to contain a value already |
13589 | truncated to MODE. In this case we can use a subreg to refer to |
13590 | the truncated value even though in the generic case we would need |
13591 | an explicit truncation. */ |
13592 | |
13593 | static bool |
13594 | reg_truncated_to_mode (machine_mode mode, const_rtx x) |
13595 | { |
13596 | reg_stat_type *rsp = ®_stat[REGNO (x)]; |
13597 | machine_mode truncated = rsp->truncated_to_mode; |
13598 | |
13599 | if (truncated == 0 |
13600 | || rsp->truncation_label < label_tick_ebb_start) |
13601 | return false; |
13602 | if (!partial_subreg_p (outermode: mode, innermode: truncated)) |
13603 | return true; |
13604 | if (TRULY_NOOP_TRUNCATION_MODES_P (mode, truncated)) |
13605 | return true; |
13606 | return false; |
13607 | } |
13608 | |
13609 | /* If X is a hard reg or a subreg record the mode that the register is |
13610 | accessed in. For non-TARGET_TRULY_NOOP_TRUNCATION targets we might be |
13611 | able to turn a truncate into a subreg using this information. Return true |
13612 | if traversing X is complete. */ |
13613 | |
13614 | static bool |
13615 | record_truncated_value (rtx x) |
13616 | { |
13617 | machine_mode truncated_mode; |
13618 | reg_stat_type *rsp; |
13619 | |
13620 | if (GET_CODE (x) == SUBREG && REG_P (SUBREG_REG (x))) |
13621 | { |
13622 | machine_mode original_mode = GET_MODE (SUBREG_REG (x)); |
13623 | truncated_mode = GET_MODE (x); |
13624 | |
13625 | if (!partial_subreg_p (outermode: truncated_mode, innermode: original_mode)) |
13626 | return true; |
13627 | |
13628 | truncated_mode = GET_MODE (x); |
13629 | if (TRULY_NOOP_TRUNCATION_MODES_P (truncated_mode, original_mode)) |
13630 | return true; |
13631 | |
13632 | x = SUBREG_REG (x); |
13633 | } |
13634 | /* ??? For hard-regs we now record everything. We might be able to |
13635 | optimize this using last_set_mode. */ |
13636 | else if (REG_P (x) && REGNO (x) < FIRST_PSEUDO_REGISTER) |
13637 | truncated_mode = GET_MODE (x); |
13638 | else |
13639 | return false; |
13640 | |
13641 | rsp = ®_stat[REGNO (x)]; |
13642 | if (rsp->truncated_to_mode == 0 |
13643 | || rsp->truncation_label < label_tick_ebb_start |
13644 | || partial_subreg_p (outermode: truncated_mode, innermode: rsp->truncated_to_mode)) |
13645 | { |
13646 | rsp->truncated_to_mode = truncated_mode; |
13647 | rsp->truncation_label = label_tick; |
13648 | } |
13649 | |
13650 | return true; |
13651 | } |
13652 | |
13653 | /* Callback for note_uses. Find hardregs and subregs of pseudos and |
13654 | the modes they are used in. This can help truning TRUNCATEs into |
13655 | SUBREGs. */ |
13656 | |
13657 | static void |
13658 | record_truncated_values (rtx *loc, void *data ATTRIBUTE_UNUSED) |
13659 | { |
13660 | subrtx_var_iterator::array_type array; |
13661 | FOR_EACH_SUBRTX_VAR (iter, array, *loc, NONCONST) |
13662 | if (record_truncated_value (x: *iter)) |
13663 | iter.skip_subrtxes (); |
13664 | } |
13665 | |
13666 | /* Scan X for promoted SUBREGs. For each one found, |
13667 | note what it implies to the registers used in it. */ |
13668 | |
13669 | static void |
13670 | check_promoted_subreg (rtx_insn *insn, rtx x) |
13671 | { |
13672 | if (GET_CODE (x) == SUBREG |
13673 | && SUBREG_PROMOTED_VAR_P (x) |
13674 | && REG_P (SUBREG_REG (x))) |
13675 | record_promoted_value (insn, subreg: x); |
13676 | else |
13677 | { |
13678 | const char *format = GET_RTX_FORMAT (GET_CODE (x)); |
13679 | int i, j; |
13680 | |
13681 | for (i = 0; i < GET_RTX_LENGTH (GET_CODE (x)); i++) |
13682 | switch (format[i]) |
13683 | { |
13684 | case 'e': |
13685 | check_promoted_subreg (insn, XEXP (x, i)); |
13686 | break; |
13687 | case 'V': |
13688 | case 'E': |
13689 | if (XVEC (x, i) != 0) |
13690 | for (j = 0; j < XVECLEN (x, i); j++) |
13691 | check_promoted_subreg (insn, XVECEXP (x, i, j)); |
13692 | break; |
13693 | } |
13694 | } |
13695 | } |
13696 | |
13697 | /* Verify that all the registers and memory references mentioned in *LOC are |
13698 | still valid. *LOC was part of a value set in INSN when label_tick was |
13699 | equal to TICK. Return false if some are not. If REPLACE is true, replace |
13700 | the invalid references with (clobber (const_int 0)) and return true. This |
13701 | replacement is useful because we often can get useful information about |
13702 | the form of a value (e.g., if it was produced by a shift that always |
13703 | produces -1 or 0) even though we don't know exactly what registers it |
13704 | was produced from. */ |
13705 | |
13706 | static bool |
13707 | get_last_value_validate (rtx *loc, rtx_insn *insn, int tick, bool replace) |
13708 | { |
13709 | rtx x = *loc; |
13710 | const char *fmt = GET_RTX_FORMAT (GET_CODE (x)); |
13711 | int len = GET_RTX_LENGTH (GET_CODE (x)); |
13712 | int i, j; |
13713 | |
13714 | if (REG_P (x)) |
13715 | { |
13716 | unsigned int regno = REGNO (x); |
13717 | unsigned int endregno = END_REGNO (x); |
13718 | unsigned int j; |
13719 | |
13720 | for (j = regno; j < endregno; j++) |
13721 | { |
13722 | reg_stat_type *rsp = ®_stat[j]; |
13723 | if (rsp->last_set_invalid |
13724 | /* If this is a pseudo-register that was only set once and not |
13725 | live at the beginning of the function, it is always valid. */ |
13726 | || (! (regno >= FIRST_PSEUDO_REGISTER |
13727 | && regno < reg_n_sets_max |
13728 | && REG_N_SETS (regno) == 1 |
13729 | && (!REGNO_REG_SET_P |
13730 | (DF_LR_IN (ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb), |
13731 | regno))) |
13732 | && rsp->last_set_label > tick)) |
13733 | { |
13734 | if (replace) |
13735 | *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
13736 | return replace; |
13737 | } |
13738 | } |
13739 | |
13740 | return true; |
13741 | } |
13742 | /* If this is a memory reference, make sure that there were no stores after |
13743 | it that might have clobbered the value. We don't have alias info, so we |
13744 | assume any store invalidates it. Moreover, we only have local UIDs, so |
13745 | we also assume that there were stores in the intervening basic blocks. */ |
13746 | else if (MEM_P (x) && !MEM_READONLY_P (x) |
13747 | && (tick != label_tick || DF_INSN_LUID (insn) <= mem_last_set)) |
13748 | { |
13749 | if (replace) |
13750 | *loc = gen_rtx_CLOBBER (GET_MODE (x), const0_rtx); |
13751 | return replace; |
13752 | } |
13753 | |
13754 | for (i = 0; i < len; i++) |
13755 | { |
13756 | if (fmt[i] == 'e') |
13757 | { |
13758 | /* Check for identical subexpressions. If x contains |
13759 | identical subexpression we only have to traverse one of |
13760 | them. */ |
13761 | if (i == 1 && ARITHMETIC_P (x)) |
13762 | { |
13763 | /* Note that at this point x0 has already been checked |
13764 | and found valid. */ |
13765 | rtx x0 = XEXP (x, 0); |
13766 | rtx x1 = XEXP (x, 1); |
13767 | |
13768 | /* If x0 and x1 are identical then x is also valid. */ |
13769 | if (x0 == x1) |
13770 | return true; |
13771 | |
13772 | /* If x1 is identical to a subexpression of x0 then |
13773 | while checking x0, x1 has already been checked. Thus |
13774 | it is valid and so as x. */ |
13775 | if (ARITHMETIC_P (x0) |
13776 | && (x1 == XEXP (x0, 0) || x1 == XEXP (x0, 1))) |
13777 | return true; |
13778 | |
13779 | /* If x0 is identical to a subexpression of x1 then x is |
13780 | valid iff the rest of x1 is valid. */ |
13781 | if (ARITHMETIC_P (x1) |
13782 | && (x0 == XEXP (x1, 0) || x0 == XEXP (x1, 1))) |
13783 | return |
13784 | get_last_value_validate (loc: &XEXP (x1, |
13785 | x0 == XEXP (x1, 0) ? 1 : 0), |
13786 | insn, tick, replace); |
13787 | } |
13788 | |
13789 | if (!get_last_value_validate (loc: &XEXP (x, i), insn, tick, replace)) |
13790 | return false; |
13791 | } |
13792 | else if (fmt[i] == 'E') |
13793 | for (j = 0; j < XVECLEN (x, i); j++) |
13794 | if (!get_last_value_validate (loc: &XVECEXP (x, i, j), |
13795 | insn, tick, replace)) |
13796 | return false; |
13797 | } |
13798 | |
13799 | /* If we haven't found a reason for it to be invalid, it is valid. */ |
13800 | return true; |
13801 | } |
13802 | |
13803 | /* Get the last value assigned to X, if known. Some registers |
13804 | in the value may be replaced with (clobber (const_int 0)) if their value |
13805 | is known longer known reliably. */ |
13806 | |
13807 | static rtx |
13808 | get_last_value (const_rtx x) |
13809 | { |
13810 | unsigned int regno; |
13811 | rtx value; |
13812 | reg_stat_type *rsp; |
13813 | |
13814 | /* If this is a non-paradoxical SUBREG, get the value of its operand and |
13815 | then convert it to the desired mode. If this is a paradoxical SUBREG, |
13816 | we cannot predict what values the "extra" bits might have. */ |
13817 | if (GET_CODE (x) == SUBREG |
13818 | && subreg_lowpart_p (x) |
13819 | && !paradoxical_subreg_p (x) |
13820 | && (value = get_last_value (SUBREG_REG (x))) != 0) |
13821 | return gen_lowpart (GET_MODE (x), value); |
13822 | |
13823 | if (!REG_P (x)) |
13824 | return 0; |
13825 | |
13826 | regno = REGNO (x); |
13827 | rsp = ®_stat[regno]; |
13828 | value = rsp->last_set_value; |
13829 | |
13830 | /* If we don't have a value, or if it isn't for this basic block and |
13831 | it's either a hard register, set more than once, or it's a live |
13832 | at the beginning of the function, return 0. |
13833 | |
13834 | Because if it's not live at the beginning of the function then the reg |
13835 | is always set before being used (is never used without being set). |
13836 | And, if it's set only once, and it's always set before use, then all |
13837 | uses must have the same last value, even if it's not from this basic |
13838 | block. */ |
13839 | |
13840 | if (value == 0 |
13841 | || (rsp->last_set_label < label_tick_ebb_start |
13842 | && (regno < FIRST_PSEUDO_REGISTER |
13843 | || regno >= reg_n_sets_max |
13844 | || REG_N_SETS (regno) != 1 |
13845 | || REGNO_REG_SET_P |
13846 | (DF_LR_IN (ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb), regno)))) |
13847 | return 0; |
13848 | |
13849 | /* If the value was set in a later insn than the ones we are processing, |
13850 | we can't use it even if the register was only set once. */ |
13851 | if (rsp->last_set_label == label_tick |
13852 | && DF_INSN_LUID (rsp->last_set) >= subst_low_luid) |
13853 | return 0; |
13854 | |
13855 | /* If fewer bits were set than what we are asked for now, we cannot use |
13856 | the value. */ |
13857 | if (maybe_lt (a: GET_MODE_PRECISION (mode: rsp->last_set_mode), |
13858 | b: GET_MODE_PRECISION (GET_MODE (x)))) |
13859 | return 0; |
13860 | |
13861 | /* If the value has all its registers valid, return it. */ |
13862 | if (get_last_value_validate (loc: &value, insn: rsp->last_set, |
13863 | tick: rsp->last_set_label, replace: false)) |
13864 | return value; |
13865 | |
13866 | /* Otherwise, make a copy and replace any invalid register with |
13867 | (clobber (const_int 0)). If that fails for some reason, return 0. */ |
13868 | |
13869 | value = copy_rtx (value); |
13870 | if (get_last_value_validate (loc: &value, insn: rsp->last_set, |
13871 | tick: rsp->last_set_label, replace: true)) |
13872 | return value; |
13873 | |
13874 | return 0; |
13875 | } |
13876 | |
13877 | /* Define three variables used for communication between the following |
13878 | routines. */ |
13879 | |
13880 | static unsigned int reg_dead_regno, reg_dead_endregno; |
13881 | static int reg_dead_flag; |
13882 | rtx reg_dead_reg; |
13883 | |
13884 | /* Function called via note_stores from reg_dead_at_p. |
13885 | |
13886 | If DEST is within [reg_dead_regno, reg_dead_endregno), set |
13887 | reg_dead_flag to 1 if X is a CLOBBER and to -1 it is a SET. */ |
13888 | |
13889 | static void |
13890 | reg_dead_at_p_1 (rtx dest, const_rtx x, void *data ATTRIBUTE_UNUSED) |
13891 | { |
13892 | unsigned int regno, endregno; |
13893 | |
13894 | if (!REG_P (dest)) |
13895 | return; |
13896 | |
13897 | regno = REGNO (dest); |
13898 | endregno = END_REGNO (x: dest); |
13899 | if (reg_dead_endregno > regno && reg_dead_regno < endregno) |
13900 | reg_dead_flag = (GET_CODE (x) == CLOBBER) ? 1 : -1; |
13901 | } |
13902 | |
13903 | /* Return true if REG is known to be dead at INSN. |
13904 | |
13905 | We scan backwards from INSN. If we hit a REG_DEAD note or a CLOBBER |
13906 | referencing REG, it is dead. If we hit a SET referencing REG, it is |
13907 | live. Otherwise, see if it is live or dead at the start of the basic |
13908 | block we are in. Hard regs marked as being live in NEWPAT_USED_REGS |
13909 | must be assumed to be always live. */ |
13910 | |
13911 | static bool |
13912 | reg_dead_at_p (rtx reg, rtx_insn *insn) |
13913 | { |
13914 | basic_block block; |
13915 | unsigned int i; |
13916 | |
13917 | /* Set variables for reg_dead_at_p_1. */ |
13918 | reg_dead_regno = REGNO (reg); |
13919 | reg_dead_endregno = END_REGNO (x: reg); |
13920 | reg_dead_reg = reg; |
13921 | |
13922 | reg_dead_flag = 0; |
13923 | |
13924 | /* Check that reg isn't mentioned in NEWPAT_USED_REGS. For fixed registers |
13925 | we allow the machine description to decide whether use-and-clobber |
13926 | patterns are OK. */ |
13927 | if (reg_dead_regno < FIRST_PSEUDO_REGISTER) |
13928 | { |
13929 | for (i = reg_dead_regno; i < reg_dead_endregno; i++) |
13930 | if (!fixed_regs[i] && TEST_HARD_REG_BIT (set: newpat_used_regs, bit: i)) |
13931 | return false; |
13932 | } |
13933 | |
13934 | /* Scan backwards until we find a REG_DEAD note, SET, CLOBBER, or |
13935 | beginning of basic block. */ |
13936 | block = BLOCK_FOR_INSN (insn); |
13937 | for (;;) |
13938 | { |
13939 | if (INSN_P (insn)) |
13940 | { |
13941 | if (find_regno_note (insn, REG_UNUSED, reg_dead_regno)) |
13942 | return true; |
13943 | |
13944 | note_stores (insn, reg_dead_at_p_1, NULL); |
13945 | if (reg_dead_flag) |
13946 | return reg_dead_flag == 1 ? 1 : 0; |
13947 | |
13948 | if (find_regno_note (insn, REG_DEAD, reg_dead_regno)) |
13949 | return true; |
13950 | } |
13951 | |
13952 | if (insn == BB_HEAD (block)) |
13953 | break; |
13954 | |
13955 | insn = PREV_INSN (insn); |
13956 | } |
13957 | |
13958 | /* Look at live-in sets for the basic block that we were in. */ |
13959 | for (i = reg_dead_regno; i < reg_dead_endregno; i++) |
13960 | if (REGNO_REG_SET_P (df_get_live_in (block), i)) |
13961 | return false; |
13962 | |
13963 | return true; |
13964 | } |
13965 | |
13966 | /* Note hard registers in X that are used. */ |
13967 | |
13968 | static void |
13969 | mark_used_regs_combine (rtx x) |
13970 | { |
13971 | RTX_CODE code = GET_CODE (x); |
13972 | unsigned int regno; |
13973 | int i; |
13974 | |
13975 | switch (code) |
13976 | { |
13977 | case LABEL_REF: |
13978 | case SYMBOL_REF: |
13979 | case CONST: |
13980 | CASE_CONST_ANY: |
13981 | case PC: |
13982 | case ADDR_VEC: |
13983 | case ADDR_DIFF_VEC: |
13984 | case ASM_INPUT: |
13985 | return; |
13986 | |
13987 | case CLOBBER: |
13988 | /* If we are clobbering a MEM, mark any hard registers inside the |
13989 | address as used. */ |
13990 | if (MEM_P (XEXP (x, 0))) |
13991 | mark_used_regs_combine (XEXP (XEXP (x, 0), 0)); |
13992 | return; |
13993 | |
13994 | case REG: |
13995 | regno = REGNO (x); |
13996 | /* A hard reg in a wide mode may really be multiple registers. |
13997 | If so, mark all of them just like the first. */ |
13998 | if (regno < FIRST_PSEUDO_REGISTER) |
13999 | { |
14000 | /* None of this applies to the stack, frame or arg pointers. */ |
14001 | if (regno == STACK_POINTER_REGNUM |
14002 | || (!HARD_FRAME_POINTER_IS_FRAME_POINTER |
14003 | && regno == HARD_FRAME_POINTER_REGNUM) |
14004 | || (FRAME_POINTER_REGNUM != ARG_POINTER_REGNUM |
14005 | && regno == ARG_POINTER_REGNUM && fixed_regs[regno]) |
14006 | || regno == FRAME_POINTER_REGNUM) |
14007 | return; |
14008 | |
14009 | add_to_hard_reg_set (regs: &newpat_used_regs, GET_MODE (x), regno); |
14010 | } |
14011 | return; |
14012 | |
14013 | case SET: |
14014 | { |
14015 | /* If setting a MEM, or a SUBREG of a MEM, then note any hard regs in |
14016 | the address. */ |
14017 | rtx testreg = SET_DEST (x); |
14018 | |
14019 | while (GET_CODE (testreg) == SUBREG |
14020 | || GET_CODE (testreg) == ZERO_EXTRACT |
14021 | || GET_CODE (testreg) == STRICT_LOW_PART) |
14022 | testreg = XEXP (testreg, 0); |
14023 | |
14024 | if (MEM_P (testreg)) |
14025 | mark_used_regs_combine (XEXP (testreg, 0)); |
14026 | |
14027 | mark_used_regs_combine (SET_SRC (x)); |
14028 | } |
14029 | return; |
14030 | |
14031 | default: |
14032 | break; |
14033 | } |
14034 | |
14035 | /* Recursively scan the operands of this expression. */ |
14036 | |
14037 | { |
14038 | const char *fmt = GET_RTX_FORMAT (code); |
14039 | |
14040 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) |
14041 | { |
14042 | if (fmt[i] == 'e') |
14043 | mark_used_regs_combine (XEXP (x, i)); |
14044 | else if (fmt[i] == 'E') |
14045 | { |
14046 | int j; |
14047 | |
14048 | for (j = 0; j < XVECLEN (x, i); j++) |
14049 | mark_used_regs_combine (XVECEXP (x, i, j)); |
14050 | } |
14051 | } |
14052 | } |
14053 | } |
14054 | |
14055 | /* Remove register number REGNO from the dead registers list of INSN. |
14056 | |
14057 | Return the note used to record the death, if there was one. */ |
14058 | |
14059 | rtx |
14060 | remove_death (unsigned int regno, rtx_insn *insn) |
14061 | { |
14062 | rtx note = find_regno_note (insn, REG_DEAD, regno); |
14063 | |
14064 | if (note) |
14065 | remove_note (insn, note); |
14066 | |
14067 | return note; |
14068 | } |
14069 | |
14070 | /* For each register (hardware or pseudo) used within expression X, if its |
14071 | death is in an instruction with luid between FROM_LUID (inclusive) and |
14072 | TO_INSN (exclusive), put a REG_DEAD note for that register in the |
14073 | list headed by PNOTES. |
14074 | |
14075 | That said, don't move registers killed by maybe_kill_insn. |
14076 | |
14077 | This is done when X is being merged by combination into TO_INSN. These |
14078 | notes will then be distributed as needed. */ |
14079 | |
14080 | static void |
14081 | move_deaths (rtx x, rtx maybe_kill_insn, int from_luid, rtx_insn *to_insn, |
14082 | rtx *pnotes) |
14083 | { |
14084 | const char *fmt; |
14085 | int len, i; |
14086 | enum rtx_code code = GET_CODE (x); |
14087 | |
14088 | if (code == REG) |
14089 | { |
14090 | unsigned int regno = REGNO (x); |
14091 | rtx_insn *where_dead = reg_stat[regno].last_death; |
14092 | |
14093 | /* If we do not know where the register died, it may still die between |
14094 | FROM_LUID and TO_INSN. If so, find it. This is PR83304. */ |
14095 | if (!where_dead || DF_INSN_LUID (where_dead) >= DF_INSN_LUID (to_insn)) |
14096 | { |
14097 | rtx_insn *insn = prev_real_nondebug_insn (to_insn); |
14098 | while (insn |
14099 | && BLOCK_FOR_INSN (insn) == BLOCK_FOR_INSN (insn: to_insn) |
14100 | && DF_INSN_LUID (insn) >= from_luid) |
14101 | { |
14102 | if (dead_or_set_regno_p (insn, regno)) |
14103 | { |
14104 | if (find_regno_note (insn, REG_DEAD, regno)) |
14105 | where_dead = insn; |
14106 | break; |
14107 | } |
14108 | |
14109 | insn = prev_real_nondebug_insn (insn); |
14110 | } |
14111 | } |
14112 | |
14113 | /* Don't move the register if it gets killed in between from and to. */ |
14114 | if (maybe_kill_insn && reg_set_p (x, maybe_kill_insn) |
14115 | && ! reg_referenced_p (x, maybe_kill_insn)) |
14116 | return; |
14117 | |
14118 | if (where_dead |
14119 | && BLOCK_FOR_INSN (insn: where_dead) == BLOCK_FOR_INSN (insn: to_insn) |
14120 | && DF_INSN_LUID (where_dead) >= from_luid |
14121 | && DF_INSN_LUID (where_dead) < DF_INSN_LUID (to_insn)) |
14122 | { |
14123 | rtx note = remove_death (regno, insn: where_dead); |
14124 | |
14125 | /* It is possible for the call above to return 0. This can occur |
14126 | when last_death points to I2 or I1 that we combined with. |
14127 | In that case make a new note. |
14128 | |
14129 | We must also check for the case where X is a hard register |
14130 | and NOTE is a death note for a range of hard registers |
14131 | including X. In that case, we must put REG_DEAD notes for |
14132 | the remaining registers in place of NOTE. */ |
14133 | |
14134 | if (note != 0 && regno < FIRST_PSEUDO_REGISTER |
14135 | && partial_subreg_p (GET_MODE (x), GET_MODE (XEXP (note, 0)))) |
14136 | { |
14137 | unsigned int deadregno = REGNO (XEXP (note, 0)); |
14138 | unsigned int deadend = END_REGNO (XEXP (note, 0)); |
14139 | unsigned int ourend = END_REGNO (x); |
14140 | unsigned int i; |
14141 | |
14142 | for (i = deadregno; i < deadend; i++) |
14143 | if (i < regno || i >= ourend) |
14144 | add_reg_note (where_dead, REG_DEAD, regno_reg_rtx[i]); |
14145 | } |
14146 | |
14147 | /* If we didn't find any note, or if we found a REG_DEAD note that |
14148 | covers only part of the given reg, and we have a multi-reg hard |
14149 | register, then to be safe we must check for REG_DEAD notes |
14150 | for each register other than the first. They could have |
14151 | their own REG_DEAD notes lying around. */ |
14152 | else if ((note == 0 |
14153 | || (note != 0 |
14154 | && partial_subreg_p (GET_MODE (XEXP (note, 0)), |
14155 | GET_MODE (x)))) |
14156 | && regno < FIRST_PSEUDO_REGISTER |
14157 | && REG_NREGS (x) > 1) |
14158 | { |
14159 | unsigned int ourend = END_REGNO (x); |
14160 | unsigned int i, offset; |
14161 | rtx oldnotes = 0; |
14162 | |
14163 | if (note) |
14164 | offset = hard_regno_nregs (regno, GET_MODE (XEXP (note, 0))); |
14165 | else |
14166 | offset = 1; |
14167 | |
14168 | for (i = regno + offset; i < ourend; i++) |
14169 | move_deaths (x: regno_reg_rtx[i], |
14170 | maybe_kill_insn, from_luid, to_insn, pnotes: &oldnotes); |
14171 | } |
14172 | |
14173 | if (note != 0 && GET_MODE (XEXP (note, 0)) == GET_MODE (x)) |
14174 | { |
14175 | XEXP (note, 1) = *pnotes; |
14176 | *pnotes = note; |
14177 | } |
14178 | else |
14179 | *pnotes = alloc_reg_note (REG_DEAD, x, *pnotes); |
14180 | } |
14181 | |
14182 | return; |
14183 | } |
14184 | |
14185 | else if (GET_CODE (x) == SET) |
14186 | { |
14187 | rtx dest = SET_DEST (x); |
14188 | |
14189 | move_deaths (SET_SRC (x), maybe_kill_insn, from_luid, to_insn, pnotes); |
14190 | |
14191 | /* In the case of a ZERO_EXTRACT, a STRICT_LOW_PART, or a SUBREG |
14192 | that accesses one word of a multi-word item, some |
14193 | piece of everything register in the expression is used by |
14194 | this insn, so remove any old death. */ |
14195 | /* ??? So why do we test for equality of the sizes? */ |
14196 | |
14197 | if (GET_CODE (dest) == ZERO_EXTRACT |
14198 | || GET_CODE (dest) == STRICT_LOW_PART |
14199 | || (GET_CODE (dest) == SUBREG |
14200 | && !read_modify_subreg_p (dest))) |
14201 | { |
14202 | move_deaths (x: dest, maybe_kill_insn, from_luid, to_insn, pnotes); |
14203 | return; |
14204 | } |
14205 | |
14206 | /* If this is some other SUBREG, we know it replaces the entire |
14207 | value, so use that as the destination. */ |
14208 | if (GET_CODE (dest) == SUBREG) |
14209 | dest = SUBREG_REG (dest); |
14210 | |
14211 | /* If this is a MEM, adjust deaths of anything used in the address. |
14212 | For a REG (the only other possibility), the entire value is |
14213 | being replaced so the old value is not used in this insn. */ |
14214 | |
14215 | if (MEM_P (dest)) |
14216 | move_deaths (XEXP (dest, 0), maybe_kill_insn, from_luid, |
14217 | to_insn, pnotes); |
14218 | return; |
14219 | } |
14220 | |
14221 | else if (GET_CODE (x) == CLOBBER) |
14222 | return; |
14223 | |
14224 | len = GET_RTX_LENGTH (code); |
14225 | fmt = GET_RTX_FORMAT (code); |
14226 | |
14227 | for (i = 0; i < len; i++) |
14228 | { |
14229 | if (fmt[i] == 'E') |
14230 | { |
14231 | int j; |
14232 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) |
14233 | move_deaths (XVECEXP (x, i, j), maybe_kill_insn, from_luid, |
14234 | to_insn, pnotes); |
14235 | } |
14236 | else if (fmt[i] == 'e') |
14237 | move_deaths (XEXP (x, i), maybe_kill_insn, from_luid, to_insn, pnotes); |
14238 | } |
14239 | } |
14240 | |
14241 | /* Return true if X is the target of a bit-field assignment in BODY, the |
14242 | pattern of an insn. X must be a REG. */ |
14243 | |
14244 | static bool |
14245 | reg_bitfield_target_p (rtx x, rtx body) |
14246 | { |
14247 | int i; |
14248 | |
14249 | if (GET_CODE (body) == SET) |
14250 | { |
14251 | rtx dest = SET_DEST (body); |
14252 | rtx target; |
14253 | unsigned int regno, tregno, endregno, endtregno; |
14254 | |
14255 | if (GET_CODE (dest) == ZERO_EXTRACT) |
14256 | target = XEXP (dest, 0); |
14257 | else if (GET_CODE (dest) == STRICT_LOW_PART) |
14258 | target = SUBREG_REG (XEXP (dest, 0)); |
14259 | else |
14260 | return false; |
14261 | |
14262 | if (GET_CODE (target) == SUBREG) |
14263 | target = SUBREG_REG (target); |
14264 | |
14265 | if (!REG_P (target)) |
14266 | return false; |
14267 | |
14268 | tregno = REGNO (target), regno = REGNO (x); |
14269 | if (tregno >= FIRST_PSEUDO_REGISTER || regno >= FIRST_PSEUDO_REGISTER) |
14270 | return target == x; |
14271 | |
14272 | endtregno = end_hard_regno (GET_MODE (target), regno: tregno); |
14273 | endregno = end_hard_regno (GET_MODE (x), regno); |
14274 | |
14275 | return endregno > tregno && regno < endtregno; |
14276 | } |
14277 | |
14278 | else if (GET_CODE (body) == PARALLEL) |
14279 | for (i = XVECLEN (body, 0) - 1; i >= 0; i--) |
14280 | if (reg_bitfield_target_p (x, XVECEXP (body, 0, i))) |
14281 | return true; |
14282 | |
14283 | return false; |
14284 | } |
14285 | |
14286 | /* Given a chain of REG_NOTES originally from FROM_INSN, try to place them |
14287 | as appropriate. I3 and I2 are the insns resulting from the combination |
14288 | insns including FROM (I2 may be zero). |
14289 | |
14290 | ELIM_I2 and ELIM_I1 are either zero or registers that we know will |
14291 | not need REG_DEAD notes because they are being substituted for. This |
14292 | saves searching in the most common cases. |
14293 | |
14294 | Each note in the list is either ignored or placed on some insns, depending |
14295 | on the type of note. */ |
14296 | |
14297 | static void |
14298 | distribute_notes (rtx notes, rtx_insn *from_insn, rtx_insn *i3, rtx_insn *i2, |
14299 | rtx elim_i2, rtx elim_i1, rtx elim_i0) |
14300 | { |
14301 | rtx note, next_note; |
14302 | rtx tem_note; |
14303 | rtx_insn *tem_insn; |
14304 | |
14305 | for (note = notes; note; note = next_note) |
14306 | { |
14307 | rtx_insn *place = 0, *place2 = 0; |
14308 | |
14309 | next_note = XEXP (note, 1); |
14310 | switch (REG_NOTE_KIND (note)) |
14311 | { |
14312 | case REG_BR_PROB: |
14313 | case REG_BR_PRED: |
14314 | /* Doesn't matter much where we put this, as long as it's somewhere. |
14315 | It is preferable to keep these notes on branches, which is most |
14316 | likely to be i3. */ |
14317 | place = i3; |
14318 | break; |
14319 | |
14320 | case REG_NON_LOCAL_GOTO: |
14321 | if (JUMP_P (i3)) |
14322 | place = i3; |
14323 | else |
14324 | { |
14325 | gcc_assert (i2 && JUMP_P (i2)); |
14326 | place = i2; |
14327 | } |
14328 | break; |
14329 | |
14330 | case REG_EH_REGION: |
14331 | { |
14332 | /* The landing pad handling needs to be kept in sync with the |
14333 | prerequisite checking in try_combine. */ |
14334 | int lp_nr = INTVAL (XEXP (note, 0)); |
14335 | /* A REG_EH_REGION note transfering control can only ever come |
14336 | from i3. */ |
14337 | if (lp_nr > 0) |
14338 | gcc_assert (from_insn == i3); |
14339 | /* We are making sure there is a single effective REG_EH_REGION |
14340 | note and it's valid to put it on i3. */ |
14341 | if (!insn_could_throw_p (from_insn) |
14342 | && !(lp_nr == INT_MIN && can_nonlocal_goto (from_insn))) |
14343 | /* Throw away stray notes on insns that can never throw or |
14344 | make a nonlocal goto. */ |
14345 | ; |
14346 | else |
14347 | { |
14348 | if (CALL_P (i3)) |
14349 | place = i3; |
14350 | else |
14351 | { |
14352 | gcc_assert (cfun->can_throw_non_call_exceptions); |
14353 | /* If i3 can still trap preserve the note, otherwise we've |
14354 | combined things such that we can now prove that the |
14355 | instructions can't trap. Drop the note in this case. */ |
14356 | if (may_trap_p (i3)) |
14357 | place = i3; |
14358 | } |
14359 | } |
14360 | break; |
14361 | } |
14362 | |
14363 | case REG_ARGS_SIZE: |
14364 | /* ??? How to distribute between i3-i1. Assume i3 contains the |
14365 | entire adjustment. Assert i3 contains at least some adjust. */ |
14366 | if (!noop_move_p (i3)) |
14367 | { |
14368 | poly_int64 old_size, args_size = get_args_size (note); |
14369 | /* fixup_args_size_notes looks at REG_NORETURN note, |
14370 | so ensure the note is placed there first. */ |
14371 | if (CALL_P (i3)) |
14372 | { |
14373 | rtx *np; |
14374 | for (np = &next_note; *np; np = &XEXP (*np, 1)) |
14375 | if (REG_NOTE_KIND (*np) == REG_NORETURN) |
14376 | { |
14377 | rtx n = *np; |
14378 | *np = XEXP (n, 1); |
14379 | XEXP (n, 1) = REG_NOTES (i3); |
14380 | REG_NOTES (i3) = n; |
14381 | break; |
14382 | } |
14383 | } |
14384 | old_size = fixup_args_size_notes (PREV_INSN (insn: i3), i3, args_size); |
14385 | /* emit_call_1 adds for !ACCUMULATE_OUTGOING_ARGS |
14386 | REG_ARGS_SIZE note to all noreturn calls, allow that here. */ |
14387 | gcc_assert (maybe_ne (old_size, args_size) |
14388 | || (CALL_P (i3) |
14389 | && !ACCUMULATE_OUTGOING_ARGS |
14390 | && find_reg_note (i3, REG_NORETURN, NULL_RTX))); |
14391 | } |
14392 | break; |
14393 | |
14394 | case REG_NORETURN: |
14395 | case REG_SETJMP: |
14396 | case REG_TM: |
14397 | case REG_CALL_DECL: |
14398 | case REG_UNTYPED_CALL: |
14399 | case REG_CALL_NOCF_CHECK: |
14400 | /* These notes must remain with the call. It should not be |
14401 | possible for both I2 and I3 to be a call. */ |
14402 | if (CALL_P (i3)) |
14403 | place = i3; |
14404 | else |
14405 | { |
14406 | gcc_assert (i2 && CALL_P (i2)); |
14407 | place = i2; |
14408 | } |
14409 | break; |
14410 | |
14411 | case REG_UNUSED: |
14412 | /* Any clobbers for i3 may still exist, and so we must process |
14413 | REG_UNUSED notes from that insn. |
14414 | |
14415 | Any clobbers from i2 or i1 can only exist if they were added by |
14416 | recog_for_combine. In that case, recog_for_combine created the |
14417 | necessary REG_UNUSED notes. Trying to keep any original |
14418 | REG_UNUSED notes from these insns can cause incorrect output |
14419 | if it is for the same register as the original i3 dest. |
14420 | In that case, we will notice that the register is set in i3, |
14421 | and then add a REG_UNUSED note for the destination of i3, which |
14422 | is wrong. However, it is possible to have REG_UNUSED notes from |
14423 | i2 or i1 for register which were both used and clobbered, so |
14424 | we keep notes from i2 or i1 if they will turn into REG_DEAD |
14425 | notes. */ |
14426 | |
14427 | /* If this register is set or clobbered between FROM_INSN and I3, |
14428 | we should not create a note for it. */ |
14429 | if (reg_set_between_p (XEXP (note, 0), from_insn, i3)) |
14430 | break; |
14431 | |
14432 | /* If this register is set or clobbered in I3, put the note there |
14433 | unless there is one already. */ |
14434 | if (reg_set_p (XEXP (note, 0), PATTERN (insn: i3))) |
14435 | { |
14436 | if (from_insn != i3) |
14437 | break; |
14438 | |
14439 | if (! (REG_P (XEXP (note, 0)) |
14440 | ? find_regno_note (i3, REG_UNUSED, REGNO (XEXP (note, 0))) |
14441 | : find_reg_note (i3, REG_UNUSED, XEXP (note, 0)))) |
14442 | place = i3; |
14443 | } |
14444 | /* Otherwise, if this register is used by I3, then this register |
14445 | now dies here, so we must put a REG_DEAD note here unless there |
14446 | is one already. */ |
14447 | else if (reg_referenced_p (XEXP (note, 0), PATTERN (insn: i3)) |
14448 | && ! (REG_P (XEXP (note, 0)) |
14449 | ? find_regno_note (i3, REG_DEAD, |
14450 | REGNO (XEXP (note, 0))) |
14451 | : find_reg_note (i3, REG_DEAD, XEXP (note, 0)))) |
14452 | { |
14453 | PUT_REG_NOTE_KIND (note, REG_DEAD); |
14454 | place = i3; |
14455 | } |
14456 | |
14457 | /* A SET or CLOBBER of the REG_UNUSED reg has been removed, |
14458 | but we can't tell which at this point. We must reset any |
14459 | expectations we had about the value that was previously |
14460 | stored in the reg. ??? Ideally, we'd adjust REG_N_SETS |
14461 | and, if appropriate, restore its previous value, but we |
14462 | don't have enough information for that at this point. */ |
14463 | else |
14464 | { |
14465 | record_value_for_reg (XEXP (note, 0), NULL, NULL_RTX); |
14466 | |
14467 | /* Otherwise, if this register is now referenced in i2 |
14468 | then the register used to be modified in one of the |
14469 | original insns. If it was i3 (say, in an unused |
14470 | parallel), it's now completely gone, so the note can |
14471 | be discarded. But if it was modified in i2, i1 or i0 |
14472 | and we still reference it in i2, then we're |
14473 | referencing the previous value, and since the |
14474 | register was modified and REG_UNUSED, we know that |
14475 | the previous value is now dead. So, if we only |
14476 | reference the register in i2, we change the note to |
14477 | REG_DEAD, to reflect the previous value. However, if |
14478 | we're also setting or clobbering the register as |
14479 | scratch, we know (because the register was not |
14480 | referenced in i3) that it's unused, just as it was |
14481 | unused before, and we place the note in i2. */ |
14482 | if (from_insn != i3 && i2 && INSN_P (i2) |
14483 | && reg_referenced_p (XEXP (note, 0), PATTERN (insn: i2))) |
14484 | { |
14485 | if (!reg_set_p (XEXP (note, 0), PATTERN (insn: i2))) |
14486 | PUT_REG_NOTE_KIND (note, REG_DEAD); |
14487 | if (! (REG_P (XEXP (note, 0)) |
14488 | ? find_regno_note (i2, REG_NOTE_KIND (note), |
14489 | REGNO (XEXP (note, 0))) |
14490 | : find_reg_note (i2, REG_NOTE_KIND (note), |
14491 | XEXP (note, 0)))) |
14492 | place = i2; |
14493 | } |
14494 | } |
14495 | |
14496 | break; |
14497 | |
14498 | case REG_EQUAL: |
14499 | case REG_EQUIV: |
14500 | case REG_NOALIAS: |
14501 | /* These notes say something about results of an insn. We can |
14502 | only support them if they used to be on I3 in which case they |
14503 | remain on I3. Otherwise they are ignored. |
14504 | |
14505 | If the note refers to an expression that is not a constant, we |
14506 | must also ignore the note since we cannot tell whether the |
14507 | equivalence is still true. It might be possible to do |
14508 | slightly better than this (we only have a problem if I2DEST |
14509 | or I1DEST is present in the expression), but it doesn't |
14510 | seem worth the trouble. */ |
14511 | |
14512 | if (from_insn == i3 |
14513 | && (XEXP (note, 0) == 0 || CONSTANT_P (XEXP (note, 0)))) |
14514 | place = i3; |
14515 | break; |
14516 | |
14517 | case REG_INC: |
14518 | /* These notes say something about how a register is used. They must |
14519 | be present on any use of the register in I2 or I3. */ |
14520 | if (reg_mentioned_p (XEXP (note, 0), PATTERN (insn: i3))) |
14521 | place = i3; |
14522 | |
14523 | if (i2 && reg_mentioned_p (XEXP (note, 0), PATTERN (insn: i2))) |
14524 | { |
14525 | if (place) |
14526 | place2 = i2; |
14527 | else |
14528 | place = i2; |
14529 | } |
14530 | break; |
14531 | |
14532 | case REG_LABEL_TARGET: |
14533 | case REG_LABEL_OPERAND: |
14534 | /* This can show up in several ways -- either directly in the |
14535 | pattern, or hidden off in the constant pool with (or without?) |
14536 | a REG_EQUAL note. */ |
14537 | /* ??? Ignore the without-reg_equal-note problem for now. */ |
14538 | if (reg_mentioned_p (XEXP (note, 0), PATTERN (insn: i3)) |
14539 | || ((tem_note = find_reg_note (i3, REG_EQUAL, NULL_RTX)) |
14540 | && GET_CODE (XEXP (tem_note, 0)) == LABEL_REF |
14541 | && label_ref_label (XEXP (tem_note, 0)) == XEXP (note, 0))) |
14542 | place = i3; |
14543 | |
14544 | if (i2 |
14545 | && (reg_mentioned_p (XEXP (note, 0), PATTERN (insn: i2)) |
14546 | || ((tem_note = find_reg_note (i2, REG_EQUAL, NULL_RTX)) |
14547 | && GET_CODE (XEXP (tem_note, 0)) == LABEL_REF |
14548 | && label_ref_label (XEXP (tem_note, 0)) == XEXP (note, 0)))) |
14549 | { |
14550 | if (place) |
14551 | place2 = i2; |
14552 | else |
14553 | place = i2; |
14554 | } |
14555 | |
14556 | /* For REG_LABEL_TARGET on a JUMP_P, we prefer to put the note |
14557 | as a JUMP_LABEL or decrement LABEL_NUSES if it's already |
14558 | there. */ |
14559 | if (place && JUMP_P (place) |
14560 | && REG_NOTE_KIND (note) == REG_LABEL_TARGET |
14561 | && (JUMP_LABEL (place) == NULL |
14562 | || JUMP_LABEL (place) == XEXP (note, 0))) |
14563 | { |
14564 | rtx label = JUMP_LABEL (place); |
14565 | |
14566 | if (!label) |
14567 | JUMP_LABEL (place) = XEXP (note, 0); |
14568 | else if (LABEL_P (label)) |
14569 | LABEL_NUSES (label)--; |
14570 | } |
14571 | |
14572 | if (place2 && JUMP_P (place2) |
14573 | && REG_NOTE_KIND (note) == REG_LABEL_TARGET |
14574 | && (JUMP_LABEL (place2) == NULL |
14575 | || JUMP_LABEL (place2) == XEXP (note, 0))) |
14576 | { |
14577 | rtx label = JUMP_LABEL (place2); |
14578 | |
14579 | if (!label) |
14580 | JUMP_LABEL (place2) = XEXP (note, 0); |
14581 | else if (LABEL_P (label)) |
14582 | LABEL_NUSES (label)--; |
14583 | place2 = 0; |
14584 | } |
14585 | break; |
14586 | |
14587 | case REG_NONNEG: |
14588 | /* This note says something about the value of a register prior |
14589 | to the execution of an insn. It is too much trouble to see |
14590 | if the note is still correct in all situations. It is better |
14591 | to simply delete it. */ |
14592 | break; |
14593 | |
14594 | case REG_DEAD: |
14595 | /* If we replaced the right hand side of FROM_INSN with a |
14596 | REG_EQUAL note, the original use of the dying register |
14597 | will not have been combined into I3 and I2. In such cases, |
14598 | FROM_INSN is guaranteed to be the first of the combined |
14599 | instructions, so we simply need to search back before |
14600 | FROM_INSN for the previous use or set of this register, |
14601 | then alter the notes there appropriately. |
14602 | |
14603 | If the register is used as an input in I3, it dies there. |
14604 | Similarly for I2, if it is nonzero and adjacent to I3. |
14605 | |
14606 | If the register is not used as an input in either I3 or I2 |
14607 | and it is not one of the registers we were supposed to eliminate, |
14608 | there are two possibilities. We might have a non-adjacent I2 |
14609 | or we might have somehow eliminated an additional register |
14610 | from a computation. For example, we might have had A & B where |
14611 | we discover that B will always be zero. In this case we will |
14612 | eliminate the reference to A. |
14613 | |
14614 | In both cases, we must search to see if we can find a previous |
14615 | use of A and put the death note there. */ |
14616 | |
14617 | if (from_insn |
14618 | && from_insn == i2mod |
14619 | && !reg_overlap_mentioned_p (XEXP (note, 0), i2mod_new_rhs)) |
14620 | tem_insn = from_insn; |
14621 | else |
14622 | { |
14623 | if (from_insn |
14624 | && CALL_P (from_insn) |
14625 | && find_reg_fusage (from_insn, USE, XEXP (note, 0))) |
14626 | place = from_insn; |
14627 | else if (i2 && reg_set_p (XEXP (note, 0), PATTERN (insn: i2))) |
14628 | { |
14629 | /* If the new I2 sets the same register that is marked |
14630 | dead in the note, we do not in general know where to |
14631 | put the note. One important case we _can_ handle is |
14632 | when the note comes from I3. */ |
14633 | if (from_insn == i3) |
14634 | place = i3; |
14635 | else |
14636 | break; |
14637 | } |
14638 | else if (reg_referenced_p (XEXP (note, 0), PATTERN (insn: i3))) |
14639 | place = i3; |
14640 | else if (i2 != 0 && next_nonnote_nondebug_insn (i2) == i3 |
14641 | && reg_referenced_p (XEXP (note, 0), PATTERN (insn: i2))) |
14642 | place = i2; |
14643 | else if ((rtx_equal_p (XEXP (note, 0), elim_i2) |
14644 | && !(i2mod |
14645 | && reg_overlap_mentioned_p (XEXP (note, 0), |
14646 | i2mod_old_rhs))) |
14647 | || rtx_equal_p (XEXP (note, 0), elim_i1) |
14648 | || rtx_equal_p (XEXP (note, 0), elim_i0)) |
14649 | break; |
14650 | tem_insn = i3; |
14651 | } |
14652 | |
14653 | if (place == 0) |
14654 | { |
14655 | basic_block bb = this_basic_block; |
14656 | |
14657 | for (tem_insn = PREV_INSN (insn: tem_insn); place == 0; tem_insn = PREV_INSN (insn: tem_insn)) |
14658 | { |
14659 | if (!NONDEBUG_INSN_P (tem_insn)) |
14660 | { |
14661 | if (tem_insn == BB_HEAD (bb)) |
14662 | break; |
14663 | continue; |
14664 | } |
14665 | |
14666 | /* If the register is being set at TEM_INSN, see if that is all |
14667 | TEM_INSN is doing. If so, delete TEM_INSN. Otherwise, make this |
14668 | into a REG_UNUSED note instead. Don't delete sets to |
14669 | global register vars. */ |
14670 | if ((REGNO (XEXP (note, 0)) >= FIRST_PSEUDO_REGISTER |
14671 | || !global_regs[REGNO (XEXP (note, 0))]) |
14672 | && reg_set_p (XEXP (note, 0), PATTERN (insn: tem_insn))) |
14673 | { |
14674 | rtx set = single_set (insn: tem_insn); |
14675 | rtx inner_dest = 0; |
14676 | |
14677 | if (set != 0) |
14678 | for (inner_dest = SET_DEST (set); |
14679 | (GET_CODE (inner_dest) == STRICT_LOW_PART |
14680 | || GET_CODE (inner_dest) == SUBREG |
14681 | || GET_CODE (inner_dest) == ZERO_EXTRACT); |
14682 | inner_dest = XEXP (inner_dest, 0)) |
14683 | ; |
14684 | |
14685 | /* Verify that it was the set, and not a clobber that |
14686 | modified the register. |
14687 | |
14688 | If we cannot delete the setter due to side |
14689 | effects, mark the user with an UNUSED note instead |
14690 | of deleting it. */ |
14691 | |
14692 | if (set != 0 && ! side_effects_p (SET_SRC (set)) |
14693 | && rtx_equal_p (XEXP (note, 0), inner_dest)) |
14694 | { |
14695 | /* Move the notes and links of TEM_INSN elsewhere. |
14696 | This might delete other dead insns recursively. |
14697 | First set the pattern to something that won't use |
14698 | any register. */ |
14699 | rtx old_notes = REG_NOTES (tem_insn); |
14700 | |
14701 | PATTERN (insn: tem_insn) = pc_rtx; |
14702 | REG_NOTES (tem_insn) = NULL; |
14703 | |
14704 | distribute_notes (notes: old_notes, from_insn: tem_insn, i3: tem_insn, NULL, |
14705 | NULL_RTX, NULL_RTX, NULL_RTX); |
14706 | distribute_links (LOG_LINKS (tem_insn)); |
14707 | |
14708 | unsigned int regno = REGNO (XEXP (note, 0)); |
14709 | reg_stat_type *rsp = ®_stat[regno]; |
14710 | if (rsp->last_set == tem_insn) |
14711 | record_value_for_reg (XEXP (note, 0), NULL, NULL_RTX); |
14712 | |
14713 | SET_INSN_DELETED (tem_insn); |
14714 | if (tem_insn == i2) |
14715 | i2 = NULL; |
14716 | } |
14717 | else |
14718 | { |
14719 | PUT_REG_NOTE_KIND (note, REG_UNUSED); |
14720 | |
14721 | /* If there isn't already a REG_UNUSED note, put one |
14722 | here. Do not place a REG_DEAD note, even if |
14723 | the register is also used here; that would not |
14724 | match the algorithm used in lifetime analysis |
14725 | and can cause the consistency check in the |
14726 | scheduler to fail. */ |
14727 | if (! find_regno_note (tem_insn, REG_UNUSED, |
14728 | REGNO (XEXP (note, 0)))) |
14729 | place = tem_insn; |
14730 | break; |
14731 | } |
14732 | } |
14733 | else if (reg_referenced_p (XEXP (note, 0), PATTERN (insn: tem_insn)) |
14734 | || (CALL_P (tem_insn) |
14735 | && find_reg_fusage (tem_insn, USE, XEXP (note, 0)))) |
14736 | { |
14737 | place = tem_insn; |
14738 | |
14739 | /* If we are doing a 3->2 combination, and we have a |
14740 | register which formerly died in i3 and was not used |
14741 | by i2, which now no longer dies in i3 and is used in |
14742 | i2 but does not die in i2, and place is between i2 |
14743 | and i3, then we may need to move a link from place to |
14744 | i2. */ |
14745 | if (i2 && DF_INSN_LUID (place) > DF_INSN_LUID (i2) |
14746 | && from_insn |
14747 | && DF_INSN_LUID (from_insn) > DF_INSN_LUID (i2) |
14748 | && reg_referenced_p (XEXP (note, 0), PATTERN (insn: i2))) |
14749 | { |
14750 | struct insn_link *links = LOG_LINKS (place); |
14751 | LOG_LINKS (place) = NULL; |
14752 | distribute_links (links); |
14753 | } |
14754 | break; |
14755 | } |
14756 | |
14757 | if (tem_insn == BB_HEAD (bb)) |
14758 | break; |
14759 | } |
14760 | |
14761 | } |
14762 | |
14763 | /* If the register is set or already dead at PLACE, we needn't do |
14764 | anything with this note if it is still a REG_DEAD note. |
14765 | We check here if it is set at all, not if is it totally replaced, |
14766 | which is what `dead_or_set_p' checks, so also check for it being |
14767 | set partially. */ |
14768 | |
14769 | if (place && REG_NOTE_KIND (note) == REG_DEAD) |
14770 | { |
14771 | unsigned int regno = REGNO (XEXP (note, 0)); |
14772 | reg_stat_type *rsp = ®_stat[regno]; |
14773 | |
14774 | if (dead_or_set_p (place, XEXP (note, 0)) |
14775 | || reg_bitfield_target_p (XEXP (note, 0), body: PATTERN (insn: place))) |
14776 | { |
14777 | /* Unless the register previously died in PLACE, clear |
14778 | last_death. [I no longer understand why this is |
14779 | being done.] */ |
14780 | if (rsp->last_death != place) |
14781 | rsp->last_death = 0; |
14782 | place = 0; |
14783 | } |
14784 | else |
14785 | rsp->last_death = place; |
14786 | |
14787 | /* If this is a death note for a hard reg that is occupying |
14788 | multiple registers, ensure that we are still using all |
14789 | parts of the object. If we find a piece of the object |
14790 | that is unused, we must arrange for an appropriate REG_DEAD |
14791 | note to be added for it. However, we can't just emit a USE |
14792 | and tag the note to it, since the register might actually |
14793 | be dead; so we recourse, and the recursive call then finds |
14794 | the previous insn that used this register. */ |
14795 | |
14796 | if (place && REG_NREGS (XEXP (note, 0)) > 1) |
14797 | { |
14798 | unsigned int endregno = END_REGNO (XEXP (note, 0)); |
14799 | bool all_used = true; |
14800 | unsigned int i; |
14801 | |
14802 | for (i = regno; i < endregno; i++) |
14803 | if ((! refers_to_regno_p (regnum: i, x: PATTERN (insn: place)) |
14804 | && ! find_regno_fusage (place, USE, i)) |
14805 | || dead_or_set_regno_p (place, i)) |
14806 | { |
14807 | all_used = false; |
14808 | break; |
14809 | } |
14810 | |
14811 | if (! all_used) |
14812 | { |
14813 | /* Put only REG_DEAD notes for pieces that are |
14814 | not already dead or set. */ |
14815 | |
14816 | for (i = regno; i < endregno; |
14817 | i += hard_regno_nregs (regno: i, reg_raw_mode[i])) |
14818 | { |
14819 | rtx piece = regno_reg_rtx[i]; |
14820 | basic_block bb = this_basic_block; |
14821 | |
14822 | if (! dead_or_set_p (place, piece) |
14823 | && ! reg_bitfield_target_p (x: piece, |
14824 | body: PATTERN (insn: place))) |
14825 | { |
14826 | rtx new_note = alloc_reg_note (REG_DEAD, piece, |
14827 | NULL_RTX); |
14828 | |
14829 | distribute_notes (notes: new_note, from_insn: place, i3: place, |
14830 | NULL, NULL_RTX, NULL_RTX, |
14831 | NULL_RTX); |
14832 | } |
14833 | else if (! refers_to_regno_p (regnum: i, x: PATTERN (insn: place)) |
14834 | && ! find_regno_fusage (place, USE, i)) |
14835 | for (tem_insn = PREV_INSN (insn: place); ; |
14836 | tem_insn = PREV_INSN (insn: tem_insn)) |
14837 | { |
14838 | if (!NONDEBUG_INSN_P (tem_insn)) |
14839 | { |
14840 | if (tem_insn == BB_HEAD (bb)) |
14841 | break; |
14842 | continue; |
14843 | } |
14844 | if (dead_or_set_p (tem_insn, piece) |
14845 | || reg_bitfield_target_p (x: piece, |
14846 | body: PATTERN (insn: tem_insn))) |
14847 | { |
14848 | add_reg_note (tem_insn, REG_UNUSED, piece); |
14849 | break; |
14850 | } |
14851 | } |
14852 | } |
14853 | |
14854 | place = 0; |
14855 | } |
14856 | } |
14857 | } |
14858 | break; |
14859 | |
14860 | default: |
14861 | /* Any other notes should not be present at this point in the |
14862 | compilation. */ |
14863 | gcc_unreachable (); |
14864 | } |
14865 | |
14866 | if (place) |
14867 | { |
14868 | XEXP (note, 1) = REG_NOTES (place); |
14869 | REG_NOTES (place) = note; |
14870 | |
14871 | /* Set added_notes_insn to the earliest insn we added a note to. */ |
14872 | if (added_notes_insn == 0 |
14873 | || DF_INSN_LUID (added_notes_insn) > DF_INSN_LUID (place)) |
14874 | added_notes_insn = place; |
14875 | } |
14876 | |
14877 | if (place2) |
14878 | { |
14879 | add_shallow_copy_of_reg_note (place2, note); |
14880 | |
14881 | /* Set added_notes_insn to the earliest insn we added a note to. */ |
14882 | if (added_notes_insn == 0 |
14883 | || DF_INSN_LUID (added_notes_insn) > DF_INSN_LUID (place2)) |
14884 | added_notes_insn = place2; |
14885 | } |
14886 | } |
14887 | } |
14888 | |
14889 | /* Similarly to above, distribute the LOG_LINKS that used to be present on |
14890 | I3, I2, and I1 to new locations. This is also called to add a link |
14891 | pointing at I3 when I3's destination is changed. */ |
14892 | |
14893 | static void |
14894 | distribute_links (struct insn_link *links) |
14895 | { |
14896 | struct insn_link *link, *next_link; |
14897 | |
14898 | for (link = links; link; link = next_link) |
14899 | { |
14900 | rtx_insn *place = 0; |
14901 | rtx_insn *insn; |
14902 | rtx set, reg; |
14903 | |
14904 | next_link = link->next; |
14905 | |
14906 | /* If the insn that this link points to is a NOTE, ignore it. */ |
14907 | if (NOTE_P (link->insn)) |
14908 | continue; |
14909 | |
14910 | set = 0; |
14911 | rtx pat = PATTERN (insn: link->insn); |
14912 | if (GET_CODE (pat) == SET) |
14913 | set = pat; |
14914 | else if (GET_CODE (pat) == PARALLEL) |
14915 | { |
14916 | int i; |
14917 | for (i = 0; i < XVECLEN (pat, 0); i++) |
14918 | { |
14919 | set = XVECEXP (pat, 0, i); |
14920 | if (GET_CODE (set) != SET) |
14921 | continue; |
14922 | |
14923 | reg = SET_DEST (set); |
14924 | while (GET_CODE (reg) == ZERO_EXTRACT |
14925 | || GET_CODE (reg) == STRICT_LOW_PART |
14926 | || GET_CODE (reg) == SUBREG) |
14927 | reg = XEXP (reg, 0); |
14928 | |
14929 | if (!REG_P (reg)) |
14930 | continue; |
14931 | |
14932 | if (REGNO (reg) == link->regno) |
14933 | break; |
14934 | } |
14935 | if (i == XVECLEN (pat, 0)) |
14936 | continue; |
14937 | } |
14938 | else |
14939 | continue; |
14940 | |
14941 | reg = SET_DEST (set); |
14942 | |
14943 | while (GET_CODE (reg) == ZERO_EXTRACT |
14944 | || GET_CODE (reg) == STRICT_LOW_PART |
14945 | || GET_CODE (reg) == SUBREG) |
14946 | reg = XEXP (reg, 0); |
14947 | |
14948 | if (reg == pc_rtx) |
14949 | continue; |
14950 | |
14951 | /* A LOG_LINK is defined as being placed on the first insn that uses |
14952 | a register and points to the insn that sets the register. Start |
14953 | searching at the next insn after the target of the link and stop |
14954 | when we reach a set of the register or the end of the basic block. |
14955 | |
14956 | Note that this correctly handles the link that used to point from |
14957 | I3 to I2. Also note that not much searching is typically done here |
14958 | since most links don't point very far away. */ |
14959 | |
14960 | for (insn = NEXT_INSN (insn: link->insn); |
14961 | (insn && (this_basic_block->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun) |
14962 | || BB_HEAD (this_basic_block->next_bb) != insn)); |
14963 | insn = NEXT_INSN (insn)) |
14964 | if (DEBUG_INSN_P (insn)) |
14965 | continue; |
14966 | else if (INSN_P (insn) && reg_overlap_mentioned_p (reg, PATTERN (insn))) |
14967 | { |
14968 | if (reg_referenced_p (reg, PATTERN (insn))) |
14969 | place = insn; |
14970 | break; |
14971 | } |
14972 | else if (CALL_P (insn) |
14973 | && find_reg_fusage (insn, USE, reg)) |
14974 | { |
14975 | place = insn; |
14976 | break; |
14977 | } |
14978 | else if (INSN_P (insn) && reg_set_p (reg, insn)) |
14979 | break; |
14980 | |
14981 | /* If we found a place to put the link, place it there unless there |
14982 | is already a link to the same insn as LINK at that point. */ |
14983 | |
14984 | if (place) |
14985 | { |
14986 | struct insn_link *link2; |
14987 | |
14988 | FOR_EACH_LOG_LINK (link2, place) |
14989 | if (link2->insn == link->insn && link2->regno == link->regno) |
14990 | break; |
14991 | |
14992 | if (link2 == NULL) |
14993 | { |
14994 | link->next = LOG_LINKS (place); |
14995 | LOG_LINKS (place) = link; |
14996 | |
14997 | /* Set added_links_insn to the earliest insn we added a |
14998 | link to. */ |
14999 | if (added_links_insn == 0 |
15000 | || DF_INSN_LUID (added_links_insn) > DF_INSN_LUID (place)) |
15001 | added_links_insn = place; |
15002 | } |
15003 | } |
15004 | } |
15005 | } |
15006 | |
15007 | /* Check for any register or memory mentioned in EQUIV that is not |
15008 | mentioned in EXPR. This is used to restrict EQUIV to "specializations" |
15009 | of EXPR where some registers may have been replaced by constants. */ |
15010 | |
15011 | static bool |
15012 | unmentioned_reg_p (rtx equiv, rtx expr) |
15013 | { |
15014 | subrtx_iterator::array_type array; |
15015 | FOR_EACH_SUBRTX (iter, array, equiv, NONCONST) |
15016 | { |
15017 | const_rtx x = *iter; |
15018 | if ((REG_P (x) || MEM_P (x)) |
15019 | && !reg_mentioned_p (x, expr)) |
15020 | return true; |
15021 | } |
15022 | return false; |
15023 | } |
15024 | |
15025 | /* Make pseudo-to-pseudo copies after every hard-reg-to-pseudo-copy, because |
15026 | the reg-to-reg copy can usefully combine with later instructions, but we |
15027 | do not want to combine the hard reg into later instructions, for that |
15028 | restricts register allocation. */ |
15029 | static void |
15030 | make_more_copies (void) |
15031 | { |
15032 | basic_block bb; |
15033 | |
15034 | FOR_EACH_BB_FN (bb, cfun) |
15035 | { |
15036 | rtx_insn *insn; |
15037 | |
15038 | FOR_BB_INSNS (bb, insn) |
15039 | { |
15040 | if (!NONDEBUG_INSN_P (insn)) |
15041 | continue; |
15042 | |
15043 | rtx set = single_set (insn); |
15044 | if (!set) |
15045 | continue; |
15046 | |
15047 | rtx dest = SET_DEST (set); |
15048 | if (!(REG_P (dest) && !HARD_REGISTER_P (dest))) |
15049 | continue; |
15050 | |
15051 | rtx src = SET_SRC (set); |
15052 | if (!(REG_P (src) && HARD_REGISTER_P (src))) |
15053 | continue; |
15054 | if (TEST_HARD_REG_BIT (fixed_reg_set, REGNO (src))) |
15055 | continue; |
15056 | |
15057 | rtx new_reg = gen_reg_rtx (GET_MODE (dest)); |
15058 | rtx_insn *new_insn = gen_move_insn (new_reg, src); |
15059 | SET_SRC (set) = new_reg; |
15060 | emit_insn_before (new_insn, insn); |
15061 | df_insn_rescan (insn); |
15062 | } |
15063 | } |
15064 | } |
15065 | |
15066 | /* Try combining insns through substitution. */ |
15067 | static void |
15068 | rest_of_handle_combine (void) |
15069 | { |
15070 | make_more_copies (); |
15071 | |
15072 | df_set_flags (DF_LR_RUN_DCE + DF_DEFER_INSN_RESCAN); |
15073 | df_note_add_problem (); |
15074 | df_analyze (); |
15075 | |
15076 | regstat_init_n_sets_and_refs (); |
15077 | reg_n_sets_max = max_reg_num (); |
15078 | |
15079 | bool rebuild_jump_labels_after_combine |
15080 | = combine_instructions (f: get_insns (), nregs: max_reg_num ()); |
15081 | |
15082 | /* Combining insns may have turned an indirect jump into a |
15083 | direct jump. Rebuild the JUMP_LABEL fields of jumping |
15084 | instructions. */ |
15085 | if (rebuild_jump_labels_after_combine) |
15086 | { |
15087 | if (dom_info_available_p (CDI_DOMINATORS)) |
15088 | free_dominance_info (CDI_DOMINATORS); |
15089 | timevar_push (tv: TV_JUMP); |
15090 | rebuild_jump_labels (get_insns ()); |
15091 | cleanup_cfg (0); |
15092 | timevar_pop (tv: TV_JUMP); |
15093 | } |
15094 | |
15095 | regstat_free_n_sets_and_refs (); |
15096 | } |
15097 | |
15098 | namespace { |
15099 | |
15100 | const pass_data pass_data_combine = |
15101 | { |
15102 | .type: RTL_PASS, /* type */ |
15103 | .name: "combine" , /* name */ |
15104 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
15105 | .tv_id: TV_COMBINE, /* tv_id */ |
15106 | PROP_cfglayout, /* properties_required */ |
15107 | .properties_provided: 0, /* properties_provided */ |
15108 | .properties_destroyed: 0, /* properties_destroyed */ |
15109 | .todo_flags_start: 0, /* todo_flags_start */ |
15110 | TODO_df_finish, /* todo_flags_finish */ |
15111 | }; |
15112 | |
15113 | class pass_combine : public rtl_opt_pass |
15114 | { |
15115 | public: |
15116 | pass_combine (gcc::context *ctxt) |
15117 | : rtl_opt_pass (pass_data_combine, ctxt) |
15118 | {} |
15119 | |
15120 | /* opt_pass methods: */ |
15121 | bool gate (function *) final override { return (optimize > 0); } |
15122 | unsigned int execute (function *) final override |
15123 | { |
15124 | rest_of_handle_combine (); |
15125 | return 0; |
15126 | } |
15127 | |
15128 | }; // class pass_combine |
15129 | |
15130 | } // anon namespace |
15131 | |
15132 | rtl_opt_pass * |
15133 | make_pass_combine (gcc::context *ctxt) |
15134 | { |
15135 | return new pass_combine (ctxt); |
15136 | } |
15137 | |