1 | /* Perform instruction reorganizations for delay slot filling. |
2 | Copyright (C) 1992-2023 Free Software Foundation, Inc. |
3 | Contributed by Richard Kenner (kenner@vlsi1.ultra.nyu.edu). |
4 | Hacked by Michael Tiemann (tiemann@cygnus.com). |
5 | |
6 | This file is part of GCC. |
7 | |
8 | GCC is free software; you can redistribute it and/or modify it under |
9 | the terms of the GNU General Public License as published by the Free |
10 | Software Foundation; either version 3, or (at your option) any later |
11 | version. |
12 | |
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
16 | for more details. |
17 | |
18 | You should have received a copy of the GNU General Public License |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ |
21 | |
22 | /* Instruction reorganization pass. |
23 | |
24 | This pass runs after register allocation and final jump |
25 | optimization. It should be the last pass to run before peephole. |
26 | It serves primarily to fill delay slots of insns, typically branch |
27 | and call insns. Other insns typically involve more complicated |
28 | interactions of data dependencies and resource constraints, and |
29 | are better handled by scheduling before register allocation (by the |
30 | function `schedule_insns'). |
31 | |
32 | The Branch Penalty is the number of extra cycles that are needed to |
33 | execute a branch insn. On an ideal machine, branches take a single |
34 | cycle, and the Branch Penalty is 0. Several RISC machines approach |
35 | branch delays differently: |
36 | |
37 | The MIPS has a single branch delay slot. Most insns |
38 | (except other branches) can be used to fill this slot. When the |
39 | slot is filled, two insns execute in two cycles, reducing the |
40 | branch penalty to zero. |
41 | |
42 | The SPARC always has a branch delay slot, but its effects can be |
43 | annulled when the branch is not taken. This means that failing to |
44 | find other sources of insns, we can hoist an insn from the branch |
45 | target that would only be safe to execute knowing that the branch |
46 | is taken. |
47 | |
48 | The HP-PA always has a branch delay slot. For unconditional branches |
49 | its effects can be annulled when the branch is taken. The effects |
50 | of the delay slot in a conditional branch can be nullified for forward |
51 | taken branches, or for untaken backward branches. This means |
52 | we can hoist insns from the fall-through path for forward branches or |
53 | steal insns from the target of backward branches. |
54 | |
55 | The TMS320C3x and C4x have three branch delay slots. When the three |
56 | slots are filled, the branch penalty is zero. Most insns can fill the |
57 | delay slots except jump insns. |
58 | |
59 | Three techniques for filling delay slots have been implemented so far: |
60 | |
61 | (1) `fill_simple_delay_slots' is the simplest, most efficient way |
62 | to fill delay slots. This pass first looks for insns which come |
63 | from before the branch and which are safe to execute after the |
64 | branch. Then it searches after the insn requiring delay slots or, |
65 | in the case of a branch, for insns that are after the point at |
66 | which the branch merges into the fallthrough code, if such a point |
67 | exists. When such insns are found, the branch penalty decreases |
68 | and no code expansion takes place. |
69 | |
70 | (2) `fill_eager_delay_slots' is more complicated: it is used for |
71 | scheduling conditional jumps, or for scheduling jumps which cannot |
72 | be filled using (1). A machine need not have annulled jumps to use |
73 | this strategy, but it helps (by keeping more options open). |
74 | `fill_eager_delay_slots' tries to guess the direction the branch |
75 | will go; if it guesses right 100% of the time, it can reduce the |
76 | branch penalty as much as `fill_simple_delay_slots' does. If it |
77 | guesses wrong 100% of the time, it might as well schedule nops. When |
78 | `fill_eager_delay_slots' takes insns from the fall-through path of |
79 | the jump, usually there is no code expansion; when it takes insns |
80 | from the branch target, there is code expansion if it is not the |
81 | only way to reach that target. |
82 | |
83 | (3) `relax_delay_slots' uses a set of rules to simplify code that |
84 | has been reorganized by (1) and (2). It finds cases where |
85 | conditional test can be eliminated, jumps can be threaded, extra |
86 | insns can be eliminated, etc. It is the job of (1) and (2) to do a |
87 | good job of scheduling locally; `relax_delay_slots' takes care of |
88 | making the various individual schedules work well together. It is |
89 | especially tuned to handle the control flow interactions of branch |
90 | insns. It does nothing for insns with delay slots that do not |
91 | branch. */ |
92 | |
93 | #include "config.h" |
94 | #include "system.h" |
95 | #include "coretypes.h" |
96 | #include "backend.h" |
97 | #include "target.h" |
98 | #include "rtl.h" |
99 | #include "tree.h" |
100 | #include "predict.h" |
101 | #include "memmodel.h" |
102 | #include "tm_p.h" |
103 | #include "expmed.h" |
104 | #include "insn-config.h" |
105 | #include "emit-rtl.h" |
106 | #include "recog.h" |
107 | #include "insn-attr.h" |
108 | #include "resource.h" |
109 | #include "tree-pass.h" |
110 | |
111 | |
112 | /* First, some functions that were used before GCC got a control flow graph. |
113 | These functions are now only used here in reorg.cc, and have therefore |
114 | been moved here to avoid inadvertent misuse elsewhere in the compiler. */ |
115 | |
116 | /* Return the last label to mark the same position as LABEL. Return LABEL |
117 | itself if it is null or any return rtx. */ |
118 | |
119 | static rtx |
120 | skip_consecutive_labels (rtx label_or_return) |
121 | { |
122 | rtx_insn *insn; |
123 | |
124 | if (label_or_return && ANY_RETURN_P (label_or_return)) |
125 | return label_or_return; |
126 | |
127 | rtx_insn *label = as_a <rtx_insn *> (p: label_or_return); |
128 | |
129 | /* __builtin_unreachable can create a CODE_LABEL followed by a BARRIER. |
130 | |
131 | Since reaching the CODE_LABEL is undefined behavior, we can return |
132 | any code label and we're OK at run time. |
133 | |
134 | However, if we return a CODE_LABEL which leads to a shrink-wrapped |
135 | epilogue, but the path does not have a prologue, then we will trip |
136 | a sanity check in the dwarf2 cfi code which wants to verify that |
137 | the CFIs are all the same on the traces leading to the epilogue. |
138 | |
139 | So we explicitly disallow looking through BARRIERS here. */ |
140 | for (insn = label; |
141 | insn != 0 && !INSN_P (insn) && !BARRIER_P (insn); |
142 | insn = NEXT_INSN (insn)) |
143 | if (LABEL_P (insn)) |
144 | label = insn; |
145 | |
146 | return label; |
147 | } |
148 | |
149 | /* Insns which have delay slots that have not yet been filled. */ |
150 | |
151 | static struct obstack unfilled_slots_obstack; |
152 | static rtx *unfilled_firstobj; |
153 | |
154 | /* Define macros to refer to the first and last slot containing unfilled |
155 | insns. These are used because the list may move and its address |
156 | should be recomputed at each use. */ |
157 | |
158 | #define unfilled_slots_base \ |
159 | ((rtx_insn **) obstack_base (&unfilled_slots_obstack)) |
160 | |
161 | #define unfilled_slots_next \ |
162 | ((rtx_insn **) obstack_next_free (&unfilled_slots_obstack)) |
163 | |
164 | /* Points to the label before the end of the function, or before a |
165 | return insn. */ |
166 | static rtx_code_label *function_return_label; |
167 | /* Likewise for a simple_return. */ |
168 | static rtx_code_label *function_simple_return_label; |
169 | |
170 | /* Mapping between INSN_UID's and position in the code since INSN_UID's do |
171 | not always monotonically increase. */ |
172 | static int *uid_to_ruid; |
173 | |
174 | /* Highest valid index in `uid_to_ruid'. */ |
175 | static int max_uid; |
176 | |
177 | static bool stop_search_p (rtx_insn *, bool); |
178 | static bool resource_conflicts_p (struct resources *, struct resources *); |
179 | static bool insn_references_resource_p (rtx, struct resources *, bool); |
180 | static bool insn_sets_resource_p (rtx, struct resources *, bool); |
181 | static rtx_code_label *find_end_label (rtx); |
182 | static rtx_insn *emit_delay_sequence (rtx_insn *, const vec<rtx_insn *> &, |
183 | int); |
184 | static void add_to_delay_list (rtx_insn *, vec<rtx_insn *> *); |
185 | static rtx_insn *delete_from_delay_slot (rtx_insn *); |
186 | static void delete_scheduled_jump (rtx_insn *); |
187 | static void note_delay_statistics (int, int); |
188 | static int get_jump_flags (const rtx_insn *, rtx); |
189 | static int mostly_true_jump (rtx); |
190 | static rtx get_branch_condition (const rtx_insn *, rtx); |
191 | static bool condition_dominates_p (rtx, const rtx_insn *); |
192 | static bool redirect_with_delay_slots_safe_p (rtx_insn *, rtx, rtx); |
193 | static bool redirect_with_delay_list_safe_p (rtx_insn *, rtx, |
194 | const vec<rtx_insn *> &); |
195 | static bool check_annul_list_true_false (bool, const vec<rtx_insn *> &); |
196 | static void steal_delay_list_from_target (rtx_insn *, rtx, rtx_sequence *, |
197 | vec<rtx_insn *> *, |
198 | struct resources *, |
199 | struct resources *, |
200 | struct resources *, |
201 | int, int *, bool *, |
202 | rtx *); |
203 | static void steal_delay_list_from_fallthrough (rtx_insn *, rtx, rtx_sequence *, |
204 | vec<rtx_insn *> *, |
205 | struct resources *, |
206 | struct resources *, |
207 | struct resources *, |
208 | int, int *, bool *); |
209 | static void try_merge_delay_insns (rtx_insn *, rtx_insn *); |
210 | static rtx_insn *redundant_insn (rtx, rtx_insn *, const vec<rtx_insn *> &); |
211 | static bool own_thread_p (rtx, rtx, bool); |
212 | static void update_block (rtx_insn *, rtx_insn *); |
213 | static bool reorg_redirect_jump (rtx_jump_insn *, rtx); |
214 | static void update_reg_dead_notes (rtx_insn *, rtx_insn *); |
215 | static void fix_reg_dead_note (rtx_insn *, rtx); |
216 | static void update_reg_unused_notes (rtx_insn *, rtx); |
217 | static void fill_simple_delay_slots (bool); |
218 | static void fill_slots_from_thread (rtx_jump_insn *, rtx, rtx, rtx, |
219 | bool, bool, bool, int, |
220 | int *, vec<rtx_insn *> *); |
221 | static void fill_eager_delay_slots (void); |
222 | static void relax_delay_slots (rtx_insn *); |
223 | static void make_return_insns (rtx_insn *); |
224 | |
225 | /* A wrapper around next_active_insn which takes care to return ret_rtx |
226 | unchanged. */ |
227 | |
228 | static rtx |
229 | first_active_target_insn (rtx insn) |
230 | { |
231 | if (ANY_RETURN_P (insn)) |
232 | return insn; |
233 | return next_active_insn (as_a <rtx_insn *> (p: insn)); |
234 | } |
235 | |
236 | /* Return true iff INSN is a simplejump, or any kind of return insn. */ |
237 | |
238 | static bool |
239 | simplejump_or_return_p (rtx insn) |
240 | { |
241 | return (JUMP_P (insn) |
242 | && (simplejump_p (as_a <rtx_insn *> (p: insn)) |
243 | || ANY_RETURN_P (PATTERN (insn)))); |
244 | } |
245 | |
246 | /* Return TRUE if this insn should stop the search for insn to fill delay |
247 | slots. LABELS_P indicates that labels should terminate the search. |
248 | In all cases, jumps terminate the search. */ |
249 | |
250 | static bool |
251 | stop_search_p (rtx_insn *insn, bool labels_p) |
252 | { |
253 | if (insn == 0) |
254 | return true; |
255 | |
256 | /* If the insn can throw an exception that is caught within the function, |
257 | it may effectively perform a jump from the viewpoint of the function. |
258 | Therefore act like for a jump. */ |
259 | if (can_throw_internal (insn)) |
260 | return true; |
261 | |
262 | switch (GET_CODE (insn)) |
263 | { |
264 | case NOTE: |
265 | case CALL_INSN: |
266 | case DEBUG_INSN: |
267 | return false; |
268 | |
269 | case CODE_LABEL: |
270 | return labels_p; |
271 | |
272 | case JUMP_INSN: |
273 | case BARRIER: |
274 | return true; |
275 | |
276 | case INSN: |
277 | /* OK unless it contains a delay slot or is an `asm' insn of some type. |
278 | We don't know anything about these. */ |
279 | return (GET_CODE (PATTERN (insn)) == SEQUENCE |
280 | || GET_CODE (PATTERN (insn)) == ASM_INPUT |
281 | || asm_noperands (PATTERN (insn)) >= 0); |
282 | |
283 | default: |
284 | gcc_unreachable (); |
285 | } |
286 | } |
287 | |
288 | /* Return TRUE if any resources are marked in both RES1 and RES2 or if either |
289 | resource set contains a volatile memory reference. Otherwise, return FALSE. */ |
290 | |
291 | static bool |
292 | resource_conflicts_p (struct resources *res1, struct resources *res2) |
293 | { |
294 | if ((res1->cc && res2->cc) || (res1->memory && res2->memory) |
295 | || res1->volatil || res2->volatil) |
296 | return true; |
297 | |
298 | return hard_reg_set_intersect_p (x: res1->regs, y: res2->regs); |
299 | } |
300 | |
301 | /* Return TRUE if any resource marked in RES, a `struct resources', is |
302 | referenced by INSN. If INCLUDE_DELAYED_EFFECTS is set, return if the called |
303 | routine is using those resources. |
304 | |
305 | We compute this by computing all the resources referenced by INSN and |
306 | seeing if this conflicts with RES. It might be faster to directly check |
307 | ourselves, and this is the way it used to work, but it means duplicating |
308 | a large block of complex code. */ |
309 | |
310 | static bool |
311 | insn_references_resource_p (rtx insn, struct resources *res, |
312 | bool include_delayed_effects) |
313 | { |
314 | struct resources insn_res; |
315 | |
316 | CLEAR_RESOURCE (&insn_res); |
317 | mark_referenced_resources (insn, &insn_res, include_delayed_effects); |
318 | return resource_conflicts_p (res1: &insn_res, res2: res); |
319 | } |
320 | |
321 | /* Return TRUE if INSN modifies resources that are marked in RES. |
322 | INCLUDE_DELAYED_EFFECTS is set if the actions of that routine should be |
323 | included. */ |
324 | |
325 | static bool |
326 | insn_sets_resource_p (rtx insn, struct resources *res, |
327 | bool include_delayed_effects) |
328 | { |
329 | struct resources insn_sets; |
330 | |
331 | CLEAR_RESOURCE (&insn_sets); |
332 | mark_set_resources (insn, &insn_sets, 0, |
333 | (include_delayed_effects |
334 | ? MARK_SRC_DEST_CALL |
335 | : MARK_SRC_DEST)); |
336 | return resource_conflicts_p (res1: &insn_sets, res2: res); |
337 | } |
338 | |
339 | /* Find a label at the end of the function or before a RETURN. If there |
340 | is none, try to make one. If that fails, returns 0. |
341 | |
342 | The property of such a label is that it is placed just before the |
343 | epilogue or a bare RETURN insn, so that another bare RETURN can be |
344 | turned into a jump to the label unconditionally. In particular, the |
345 | label cannot be placed before a RETURN insn with a filled delay slot. |
346 | |
347 | ??? There may be a problem with the current implementation. Suppose |
348 | we start with a bare RETURN insn and call find_end_label. It may set |
349 | function_return_label just before the RETURN. Suppose the machinery |
350 | is able to fill the delay slot of the RETURN insn afterwards. Then |
351 | function_return_label is no longer valid according to the property |
352 | described above and find_end_label will still return it unmodified. |
353 | Note that this is probably mitigated by the following observation: |
354 | once function_return_label is made, it is very likely the target of |
355 | a jump, so filling the delay slot of the RETURN will be much more |
356 | difficult. |
357 | KIND is either simple_return_rtx or ret_rtx, indicating which type of |
358 | return we're looking for. */ |
359 | |
360 | static rtx_code_label * |
361 | find_end_label (rtx kind) |
362 | { |
363 | rtx_insn *insn; |
364 | rtx_code_label **plabel; |
365 | |
366 | if (kind == ret_rtx) |
367 | plabel = &function_return_label; |
368 | else |
369 | { |
370 | gcc_assert (kind == simple_return_rtx); |
371 | plabel = &function_simple_return_label; |
372 | } |
373 | |
374 | /* If we found one previously, return it. */ |
375 | if (*plabel) |
376 | return *plabel; |
377 | |
378 | /* Otherwise, see if there is a label at the end of the function. If there |
379 | is, it must be that RETURN insns aren't needed, so that is our return |
380 | label and we don't have to do anything else. */ |
381 | |
382 | insn = get_last_insn (); |
383 | while (NOTE_P (insn) |
384 | || (NONJUMP_INSN_P (insn) |
385 | && (GET_CODE (PATTERN (insn)) == USE |
386 | || GET_CODE (PATTERN (insn)) == CLOBBER))) |
387 | insn = PREV_INSN (insn); |
388 | |
389 | /* When a target threads its epilogue we might already have a |
390 | suitable return insn. If so put a label before it for the |
391 | function_return_label. */ |
392 | if (BARRIER_P (insn) |
393 | && JUMP_P (PREV_INSN (insn)) |
394 | && PATTERN (insn: PREV_INSN (insn)) == kind) |
395 | { |
396 | rtx_insn *temp = PREV_INSN (insn: PREV_INSN (insn)); |
397 | rtx_code_label *label = gen_label_rtx (); |
398 | LABEL_NUSES (label) = 0; |
399 | |
400 | /* Put the label before any USE insns that may precede the RETURN |
401 | insn. */ |
402 | while (GET_CODE (temp) == USE) |
403 | temp = PREV_INSN (insn: temp); |
404 | |
405 | emit_label_after (label, temp); |
406 | *plabel = label; |
407 | } |
408 | |
409 | else if (LABEL_P (insn)) |
410 | *plabel = as_a <rtx_code_label *> (p: insn); |
411 | else |
412 | { |
413 | rtx_code_label *label = gen_label_rtx (); |
414 | LABEL_NUSES (label) = 0; |
415 | /* If the basic block reorder pass moves the return insn to |
416 | some other place try to locate it again and put our |
417 | function_return_label there. */ |
418 | while (insn && ! (JUMP_P (insn) && (PATTERN (insn) == kind))) |
419 | insn = PREV_INSN (insn); |
420 | if (insn) |
421 | { |
422 | insn = PREV_INSN (insn); |
423 | |
424 | /* Put the label before any USE insns that may precede the |
425 | RETURN insn. */ |
426 | while (GET_CODE (insn) == USE) |
427 | insn = PREV_INSN (insn); |
428 | |
429 | emit_label_after (label, insn); |
430 | } |
431 | else |
432 | { |
433 | if (targetm.have_epilogue () && ! targetm.have_return ()) |
434 | /* The RETURN insn has its delay slot filled so we cannot |
435 | emit the label just before it. Since we already have |
436 | an epilogue and cannot emit a new RETURN, we cannot |
437 | emit the label at all. */ |
438 | return NULL; |
439 | |
440 | /* Otherwise, make a new label and emit a RETURN and BARRIER, |
441 | if needed. */ |
442 | emit_label (label); |
443 | if (targetm.have_return ()) |
444 | { |
445 | /* The return we make may have delay slots too. */ |
446 | rtx_insn *pat = targetm.gen_return (); |
447 | rtx_insn *insn = emit_jump_insn (pat); |
448 | set_return_jump_label (insn); |
449 | emit_barrier (); |
450 | if (num_delay_slots (insn) > 0) |
451 | obstack_ptr_grow (&unfilled_slots_obstack, insn); |
452 | } |
453 | } |
454 | *plabel = label; |
455 | } |
456 | |
457 | /* Show one additional use for this label so it won't go away until |
458 | we are done. */ |
459 | ++LABEL_NUSES (*plabel); |
460 | |
461 | return *plabel; |
462 | } |
463 | |
464 | /* Put INSN and LIST together in a SEQUENCE rtx of LENGTH, and replace |
465 | the pattern of INSN with the SEQUENCE. |
466 | |
467 | Returns the insn containing the SEQUENCE that replaces INSN. */ |
468 | |
469 | static rtx_insn * |
470 | emit_delay_sequence (rtx_insn *insn, const vec<rtx_insn *> &list, int length) |
471 | { |
472 | /* Allocate the rtvec to hold the insns and the SEQUENCE. */ |
473 | rtvec seqv = rtvec_alloc (length + 1); |
474 | rtx seq = gen_rtx_SEQUENCE (VOIDmode, seqv); |
475 | rtx_insn *seq_insn = make_insn_raw (seq); |
476 | |
477 | /* If DELAY_INSN has a location, use it for SEQ_INSN. If DELAY_INSN does |
478 | not have a location, but one of the delayed insns does, we pick up a |
479 | location from there later. */ |
480 | INSN_LOCATION (insn: seq_insn) = INSN_LOCATION (insn); |
481 | |
482 | /* Unlink INSN from the insn chain, so that we can put it into |
483 | the SEQUENCE. Remember where we want to emit SEQUENCE in AFTER. */ |
484 | rtx_insn *after = PREV_INSN (insn); |
485 | remove_insn (insn); |
486 | SET_NEXT_INSN (insn) = SET_PREV_INSN (insn) = NULL; |
487 | |
488 | /* Build our SEQUENCE and rebuild the insn chain. */ |
489 | start_sequence (); |
490 | XVECEXP (seq, 0, 0) = emit_insn (insn); |
491 | |
492 | unsigned int delay_insns = list.length (); |
493 | gcc_assert (delay_insns == (unsigned int) length); |
494 | for (unsigned int i = 0; i < delay_insns; i++) |
495 | { |
496 | rtx_insn *tem = list[i]; |
497 | rtx note, next; |
498 | |
499 | /* Show that this copy of the insn isn't deleted. */ |
500 | tem->set_undeleted (); |
501 | |
502 | /* Unlink insn from its original place, and re-emit it into |
503 | the sequence. */ |
504 | SET_NEXT_INSN (tem) = SET_PREV_INSN (tem) = NULL; |
505 | XVECEXP (seq, 0, i + 1) = emit_insn (tem); |
506 | |
507 | /* SPARC assembler, for instance, emit warning when debug info is output |
508 | into the delay slot. */ |
509 | if (INSN_LOCATION (insn: tem) && !INSN_LOCATION (insn: seq_insn)) |
510 | INSN_LOCATION (insn: seq_insn) = INSN_LOCATION (insn: tem); |
511 | INSN_LOCATION (insn: tem) = 0; |
512 | |
513 | for (note = REG_NOTES (tem); note; note = next) |
514 | { |
515 | next = XEXP (note, 1); |
516 | switch (REG_NOTE_KIND (note)) |
517 | { |
518 | case REG_DEAD: |
519 | /* Remove any REG_DEAD notes because we can't rely on them now |
520 | that the insn has been moved. */ |
521 | remove_note (tem, note); |
522 | break; |
523 | |
524 | case REG_LABEL_OPERAND: |
525 | case REG_LABEL_TARGET: |
526 | /* Keep the label reference count up to date. */ |
527 | if (LABEL_P (XEXP (note, 0))) |
528 | LABEL_NUSES (XEXP (note, 0)) ++; |
529 | break; |
530 | |
531 | default: |
532 | break; |
533 | } |
534 | } |
535 | } |
536 | end_sequence (); |
537 | |
538 | /* Splice our SEQUENCE into the insn stream where INSN used to be. */ |
539 | add_insn_after (seq_insn, after, NULL); |
540 | |
541 | return seq_insn; |
542 | } |
543 | |
544 | /* Add INSN to DELAY_LIST and return the head of the new list. The list must |
545 | be in the order in which the insns are to be executed. */ |
546 | |
547 | static void |
548 | add_to_delay_list (rtx_insn *insn, vec<rtx_insn *> *delay_list) |
549 | { |
550 | /* If INSN has its block number recorded, clear it since we may |
551 | be moving the insn to a new block. */ |
552 | clear_hashed_info_for_insn (insn); |
553 | |
554 | delay_list->safe_push (obj: insn); |
555 | } |
556 | |
557 | /* Delete INSN from the delay slot of the insn that it is in, which may |
558 | produce an insn with no delay slots. Return the new insn. */ |
559 | |
560 | static rtx_insn * |
561 | delete_from_delay_slot (rtx_insn *insn) |
562 | { |
563 | rtx_insn *trial, *seq_insn, *prev; |
564 | rtx_sequence *seq; |
565 | bool had_barrier = false; |
566 | int i; |
567 | |
568 | /* We first must find the insn containing the SEQUENCE with INSN in its |
569 | delay slot. Do this by finding an insn, TRIAL, where |
570 | PREV_INSN (NEXT_INSN (TRIAL)) != TRIAL. */ |
571 | |
572 | for (trial = insn; |
573 | PREV_INSN (insn: NEXT_INSN (insn: trial)) == trial; |
574 | trial = NEXT_INSN (insn: trial)) |
575 | ; |
576 | |
577 | seq_insn = PREV_INSN (insn: NEXT_INSN (insn: trial)); |
578 | seq = as_a <rtx_sequence *> (p: PATTERN (insn: seq_insn)); |
579 | |
580 | if (NEXT_INSN (insn: seq_insn) && BARRIER_P (NEXT_INSN (seq_insn))) |
581 | had_barrier = true; |
582 | |
583 | /* Create a delay list consisting of all the insns other than the one |
584 | we are deleting (unless we were the only one). */ |
585 | auto_vec<rtx_insn *, 5> delay_list; |
586 | if (seq->len () > 2) |
587 | for (i = 1; i < seq->len (); i++) |
588 | if (seq->insn (index: i) != insn) |
589 | add_to_delay_list (insn: seq->insn (index: i), delay_list: &delay_list); |
590 | |
591 | /* Delete the old SEQUENCE, re-emit the insn that used to have the delay |
592 | list, and rebuild the delay list if non-empty. */ |
593 | prev = PREV_INSN (insn: seq_insn); |
594 | trial = seq->insn (index: 0); |
595 | delete_related_insns (seq_insn); |
596 | add_insn_after (trial, prev, NULL); |
597 | |
598 | /* If there was a barrier after the old SEQUENCE, remit it. */ |
599 | if (had_barrier) |
600 | emit_barrier_after (trial); |
601 | |
602 | /* If there are any delay insns, remit them. Otherwise clear the |
603 | annul flag. */ |
604 | if (!delay_list.is_empty ()) |
605 | trial = emit_delay_sequence (insn: trial, list: delay_list, XVECLEN (seq, 0) - 2); |
606 | else if (JUMP_P (trial)) |
607 | INSN_ANNULLED_BRANCH_P (trial) = 0; |
608 | |
609 | INSN_FROM_TARGET_P (insn) = 0; |
610 | |
611 | /* Show we need to fill this insn again. */ |
612 | obstack_ptr_grow (&unfilled_slots_obstack, trial); |
613 | |
614 | return trial; |
615 | } |
616 | |
617 | /* Delete INSN, a JUMP_INSN. */ |
618 | |
619 | static void |
620 | delete_scheduled_jump (rtx_insn *insn) |
621 | { |
622 | delete_related_insns (insn); |
623 | } |
624 | |
625 | /* Counters for delay-slot filling. */ |
626 | |
627 | #define NUM_REORG_FUNCTIONS 2 |
628 | #define MAX_DELAY_HISTOGRAM 3 |
629 | #define MAX_REORG_PASSES 2 |
630 | |
631 | static int num_insns_needing_delays[NUM_REORG_FUNCTIONS][MAX_REORG_PASSES]; |
632 | |
633 | static int num_filled_delays[NUM_REORG_FUNCTIONS][MAX_DELAY_HISTOGRAM+1][MAX_REORG_PASSES]; |
634 | |
635 | static int reorg_pass_number; |
636 | |
637 | static void |
638 | note_delay_statistics (int slots_filled, int index) |
639 | { |
640 | num_insns_needing_delays[index][reorg_pass_number]++; |
641 | if (slots_filled > MAX_DELAY_HISTOGRAM) |
642 | slots_filled = MAX_DELAY_HISTOGRAM; |
643 | num_filled_delays[index][slots_filled][reorg_pass_number]++; |
644 | } |
645 | |
646 | /* Optimize the following cases: |
647 | |
648 | 1. When a conditional branch skips over only one instruction, |
649 | use an annulling branch and put that insn in the delay slot. |
650 | Use either a branch that annuls when the condition if true or |
651 | invert the test with a branch that annuls when the condition is |
652 | false. This saves insns, since otherwise we must copy an insn |
653 | from the L1 target. |
654 | |
655 | (orig) (skip) (otherwise) |
656 | Bcc.n L1 Bcc',a L1 Bcc,a L1' |
657 | insn insn insn2 |
658 | L1: L1: L1: |
659 | insn2 insn2 insn2 |
660 | insn3 insn3 L1': |
661 | insn3 |
662 | |
663 | 2. When a conditional branch skips over only one instruction, |
664 | and after that, it unconditionally branches somewhere else, |
665 | perform the similar optimization. This saves executing the |
666 | second branch in the case where the inverted condition is true. |
667 | |
668 | Bcc.n L1 Bcc',a L2 |
669 | insn insn |
670 | L1: L1: |
671 | Bra L2 Bra L2 |
672 | |
673 | INSN is a JUMP_INSN. |
674 | |
675 | This should be expanded to skip over N insns, where N is the number |
676 | of delay slots required. */ |
677 | |
678 | static void |
679 | optimize_skip (rtx_jump_insn *insn, vec<rtx_insn *> *delay_list) |
680 | { |
681 | rtx_insn *trial = next_nonnote_insn (insn); |
682 | rtx_insn *next_trial = next_active_insn (trial); |
683 | int flags; |
684 | |
685 | flags = get_jump_flags (insn, JUMP_LABEL (insn)); |
686 | |
687 | if (trial == 0 |
688 | || !NONJUMP_INSN_P (trial) |
689 | || GET_CODE (PATTERN (trial)) == SEQUENCE |
690 | || recog_memoized (insn: trial) < 0 |
691 | || (! eligible_for_annul_false (insn, 0, trial, flags) |
692 | && ! eligible_for_annul_true (insn, 0, trial, flags)) |
693 | || RTX_FRAME_RELATED_P (trial) |
694 | || can_throw_internal (trial)) |
695 | return; |
696 | |
697 | /* There are two cases where we are just executing one insn (we assume |
698 | here that a branch requires only one insn; this should be generalized |
699 | at some point): Where the branch goes around a single insn or where |
700 | we have one insn followed by a branch to the same label we branch to. |
701 | In both of these cases, inverting the jump and annulling the delay |
702 | slot give the same effect in fewer insns. */ |
703 | if (next_trial == next_active_insn (JUMP_LABEL_AS_INSN (insn)) |
704 | || (next_trial != 0 |
705 | && simplejump_or_return_p (insn: next_trial) |
706 | && JUMP_LABEL (insn) == JUMP_LABEL (next_trial))) |
707 | { |
708 | if (eligible_for_annul_false (insn, 0, trial, flags)) |
709 | { |
710 | if (invert_jump (insn, JUMP_LABEL (insn), 1)) |
711 | INSN_FROM_TARGET_P (trial) = 1; |
712 | else if (! eligible_for_annul_true (insn, 0, trial, flags)) |
713 | return; |
714 | } |
715 | |
716 | add_to_delay_list (insn: trial, delay_list); |
717 | next_trial = next_active_insn (trial); |
718 | update_block (trial, trial); |
719 | delete_related_insns (trial); |
720 | |
721 | /* Also, if we are targeting an unconditional |
722 | branch, thread our jump to the target of that branch. Don't |
723 | change this into a RETURN here, because it may not accept what |
724 | we have in the delay slot. We'll fix this up later. */ |
725 | if (next_trial && simplejump_or_return_p (insn: next_trial)) |
726 | { |
727 | rtx target_label = JUMP_LABEL (next_trial); |
728 | if (ANY_RETURN_P (target_label)) |
729 | target_label = find_end_label (kind: target_label); |
730 | |
731 | if (target_label) |
732 | { |
733 | /* Recompute the flags based on TARGET_LABEL since threading |
734 | the jump to TARGET_LABEL may change the direction of the |
735 | jump (which may change the circumstances in which the |
736 | delay slot is nullified). */ |
737 | flags = get_jump_flags (insn, target_label); |
738 | if (eligible_for_annul_true (insn, 0, trial, flags)) |
739 | reorg_redirect_jump (insn, target_label); |
740 | } |
741 | } |
742 | |
743 | INSN_ANNULLED_BRANCH_P (insn) = 1; |
744 | } |
745 | } |
746 | |
747 | /* Encode and return branch direction and prediction information for |
748 | INSN assuming it will jump to LABEL. |
749 | |
750 | Non conditional branches return no direction information and |
751 | are predicted as very likely taken. */ |
752 | |
753 | static int |
754 | get_jump_flags (const rtx_insn *insn, rtx label) |
755 | { |
756 | int flags; |
757 | |
758 | /* get_jump_flags can be passed any insn with delay slots, these may |
759 | be INSNs, CALL_INSNs, or JUMP_INSNs. Only JUMP_INSNs have branch |
760 | direction information, and only if they are conditional jumps. |
761 | |
762 | If LABEL is a return, then there is no way to determine the branch |
763 | direction. */ |
764 | if (JUMP_P (insn) |
765 | && (condjump_p (insn) || condjump_in_parallel_p (insn)) |
766 | && !ANY_RETURN_P (label) |
767 | && INSN_UID (insn) <= max_uid |
768 | && INSN_UID (insn: label) <= max_uid) |
769 | flags |
770 | = (uid_to_ruid[INSN_UID (insn: label)] > uid_to_ruid[INSN_UID (insn)]) |
771 | ? ATTR_FLAG_forward : ATTR_FLAG_backward; |
772 | /* No valid direction information. */ |
773 | else |
774 | flags = 0; |
775 | |
776 | return flags; |
777 | } |
778 | |
779 | /* Return truth value of the statement that this branch |
780 | is mostly taken. If we think that the branch is extremely likely |
781 | to be taken, we return 2. If the branch is slightly more likely to be |
782 | taken, return 1. If the branch is slightly less likely to be taken, |
783 | return 0 and if the branch is highly unlikely to be taken, return -1. */ |
784 | |
785 | static int |
786 | mostly_true_jump (rtx jump_insn) |
787 | { |
788 | /* If branch probabilities are available, then use that number since it |
789 | always gives a correct answer. */ |
790 | rtx note = find_reg_note (jump_insn, REG_BR_PROB, 0); |
791 | if (note) |
792 | { |
793 | int prob = profile_probability::from_reg_br_prob_note (XINT (note, 0)) |
794 | .to_reg_br_prob_base (); |
795 | |
796 | if (prob >= REG_BR_PROB_BASE * 9 / 10) |
797 | return 2; |
798 | else if (prob >= REG_BR_PROB_BASE / 2) |
799 | return 1; |
800 | else if (prob >= REG_BR_PROB_BASE / 10) |
801 | return 0; |
802 | else |
803 | return -1; |
804 | } |
805 | |
806 | /* If there is no note, assume branches are not taken. |
807 | This should be rare. */ |
808 | return 0; |
809 | } |
810 | |
811 | /* Return the condition under which INSN will branch to TARGET. If TARGET |
812 | is zero, return the condition under which INSN will return. If INSN is |
813 | an unconditional branch, return const_true_rtx. If INSN isn't a simple |
814 | type of jump, or it doesn't go to TARGET, return 0. */ |
815 | |
816 | static rtx |
817 | get_branch_condition (const rtx_insn *insn, rtx target) |
818 | { |
819 | rtx pat = PATTERN (insn); |
820 | rtx src; |
821 | |
822 | if (condjump_in_parallel_p (insn)) |
823 | pat = XVECEXP (pat, 0, 0); |
824 | |
825 | if (ANY_RETURN_P (pat) && pat == target) |
826 | return const_true_rtx; |
827 | |
828 | if (GET_CODE (pat) != SET || SET_DEST (pat) != pc_rtx) |
829 | return 0; |
830 | |
831 | src = SET_SRC (pat); |
832 | if (GET_CODE (src) == LABEL_REF && label_ref_label (ref: src) == target) |
833 | return const_true_rtx; |
834 | |
835 | else if (GET_CODE (src) == IF_THEN_ELSE |
836 | && XEXP (src, 2) == pc_rtx |
837 | && ((GET_CODE (XEXP (src, 1)) == LABEL_REF |
838 | && label_ref_label (XEXP (src, 1)) == target) |
839 | || (ANY_RETURN_P (XEXP (src, 1)) && XEXP (src, 1) == target))) |
840 | return XEXP (src, 0); |
841 | |
842 | else if (GET_CODE (src) == IF_THEN_ELSE |
843 | && XEXP (src, 1) == pc_rtx |
844 | && ((GET_CODE (XEXP (src, 2)) == LABEL_REF |
845 | && label_ref_label (XEXP (src, 2)) == target) |
846 | || (ANY_RETURN_P (XEXP (src, 2)) && XEXP (src, 2) == target))) |
847 | { |
848 | enum rtx_code rev; |
849 | rev = reversed_comparison_code (XEXP (src, 0), insn); |
850 | if (rev != UNKNOWN) |
851 | return gen_rtx_fmt_ee (rev, GET_MODE (XEXP (src, 0)), |
852 | XEXP (XEXP (src, 0), 0), |
853 | XEXP (XEXP (src, 0), 1)); |
854 | } |
855 | |
856 | return 0; |
857 | } |
858 | |
859 | /* Return true if CONDITION is more strict than the condition of |
860 | INSN, i.e., if INSN will always branch if CONDITION is true. */ |
861 | |
862 | static bool |
863 | condition_dominates_p (rtx condition, const rtx_insn *insn) |
864 | { |
865 | rtx other_condition = get_branch_condition (insn, JUMP_LABEL (insn)); |
866 | enum rtx_code code = GET_CODE (condition); |
867 | enum rtx_code other_code; |
868 | |
869 | if (rtx_equal_p (condition, other_condition) |
870 | || other_condition == const_true_rtx) |
871 | return true; |
872 | |
873 | else if (condition == const_true_rtx || other_condition == 0) |
874 | return false; |
875 | |
876 | other_code = GET_CODE (other_condition); |
877 | if (GET_RTX_LENGTH (code) != 2 || GET_RTX_LENGTH (other_code) != 2 |
878 | || ! rtx_equal_p (XEXP (condition, 0), XEXP (other_condition, 0)) |
879 | || ! rtx_equal_p (XEXP (condition, 1), XEXP (other_condition, 1))) |
880 | return false; |
881 | |
882 | return comparison_dominates_p (code, other_code); |
883 | } |
884 | |
885 | /* Return true if redirecting JUMP to NEWLABEL does not invalidate |
886 | any insns already in the delay slot of JUMP. */ |
887 | |
888 | static bool |
889 | redirect_with_delay_slots_safe_p (rtx_insn *jump, rtx newlabel, rtx seq) |
890 | { |
891 | int flags, i; |
892 | rtx_sequence *pat = as_a <rtx_sequence *> (p: PATTERN (insn: seq)); |
893 | |
894 | /* Make sure all the delay slots of this jump would still |
895 | be valid after threading the jump. If they are still |
896 | valid, then return nonzero. */ |
897 | |
898 | flags = get_jump_flags (insn: jump, label: newlabel); |
899 | for (i = 1; i < pat->len (); i++) |
900 | if (! ( |
901 | #if ANNUL_IFFALSE_SLOTS |
902 | (INSN_ANNULLED_BRANCH_P (jump) |
903 | && INSN_FROM_TARGET_P (pat->insn (i))) |
904 | ? eligible_for_annul_false (jump, i - 1, pat->insn (i), flags) : |
905 | #endif |
906 | #if ANNUL_IFTRUE_SLOTS |
907 | (INSN_ANNULLED_BRANCH_P (jump) |
908 | && ! INSN_FROM_TARGET_P (XVECEXP (pat, 0, i))) |
909 | ? eligible_for_annul_true (jump, i - 1, pat->insn (i), flags) : |
910 | #endif |
911 | eligible_for_delay (jump, i - 1, pat->insn (index: i), flags))) |
912 | break; |
913 | |
914 | return (i == pat->len ()); |
915 | } |
916 | |
917 | /* Return true if redirecting JUMP to NEWLABEL does not invalidate |
918 | any insns we wish to place in the delay slot of JUMP. */ |
919 | |
920 | static bool |
921 | redirect_with_delay_list_safe_p (rtx_insn *jump, rtx newlabel, |
922 | const vec<rtx_insn *> &delay_list) |
923 | { |
924 | /* Make sure all the insns in DELAY_LIST would still be |
925 | valid after threading the jump. If they are still |
926 | valid, then return true. */ |
927 | |
928 | int flags = get_jump_flags (insn: jump, label: newlabel); |
929 | unsigned int delay_insns = delay_list.length (); |
930 | unsigned int i = 0; |
931 | for (; i < delay_insns; i++) |
932 | if (! ( |
933 | #if ANNUL_IFFALSE_SLOTS |
934 | (INSN_ANNULLED_BRANCH_P (jump) |
935 | && INSN_FROM_TARGET_P (delay_list[i])) |
936 | ? eligible_for_annul_false (jump, i, delay_list[i], flags) : |
937 | #endif |
938 | #if ANNUL_IFTRUE_SLOTS |
939 | (INSN_ANNULLED_BRANCH_P (jump) |
940 | && ! INSN_FROM_TARGET_P (delay_list[i])) |
941 | ? eligible_for_annul_true (jump, i, delay_list[i], flags) : |
942 | #endif |
943 | eligible_for_delay (jump, i, delay_list[i], flags))) |
944 | break; |
945 | |
946 | return i == delay_insns; |
947 | } |
948 | |
949 | /* DELAY_LIST is a list of insns that have already been placed into delay |
950 | slots. See if all of them have the same annulling status as ANNUL_TRUE_P. |
951 | If not, return false; otherwise return true. */ |
952 | |
953 | static bool |
954 | check_annul_list_true_false (bool annul_true_p, |
955 | const vec<rtx_insn *> &delay_list) |
956 | { |
957 | rtx_insn *trial; |
958 | unsigned int i; |
959 | FOR_EACH_VEC_ELT (delay_list, i, trial) |
960 | if ((annul_true_p && INSN_FROM_TARGET_P (trial)) |
961 | || (!annul_true_p && !INSN_FROM_TARGET_P (trial))) |
962 | return false; |
963 | |
964 | return true; |
965 | } |
966 | |
967 | /* INSN branches to an insn whose pattern SEQ is a SEQUENCE. Given that |
968 | the condition tested by INSN is CONDITION and the resources shown in |
969 | OTHER_NEEDED are needed after INSN, see whether INSN can take all the insns |
970 | from SEQ's delay list, in addition to whatever insns it may execute |
971 | (in DELAY_LIST). SETS and NEEDED are denote resources already set and |
972 | needed while searching for delay slot insns. Return the concatenated |
973 | delay list if possible, otherwise, return 0. |
974 | |
975 | SLOTS_TO_FILL is the total number of slots required by INSN, and |
976 | PSLOTS_FILLED points to the number filled so far (also the number of |
977 | insns in DELAY_LIST). It is updated with the number that have been |
978 | filled from the SEQUENCE, if any. |
979 | |
980 | PANNUL_P points to a nonzero value if we already know that we need |
981 | to annul INSN. If this routine determines that annulling is needed, |
982 | it may set that value to true. |
983 | |
984 | PNEW_THREAD points to a location that is to receive the place at which |
985 | execution should continue. */ |
986 | |
987 | static void |
988 | steal_delay_list_from_target (rtx_insn *insn, rtx condition, rtx_sequence *seq, |
989 | vec<rtx_insn *> *delay_list, |
990 | struct resources *sets, |
991 | struct resources *needed, |
992 | struct resources *other_needed, |
993 | int slots_to_fill, int *pslots_filled, |
994 | bool *pannul_p, rtx *pnew_thread) |
995 | { |
996 | int slots_remaining = slots_to_fill - *pslots_filled; |
997 | int total_slots_filled = *pslots_filled; |
998 | auto_vec<rtx_insn *, 5> new_delay_list; |
999 | bool must_annul = *pannul_p; |
1000 | bool used_annul = false; |
1001 | int i; |
1002 | struct resources cc_set; |
1003 | rtx_insn **redundant; |
1004 | |
1005 | /* We can't do anything if there are more delay slots in SEQ than we |
1006 | can handle, or if we don't know that it will be a taken branch. |
1007 | We know that it will be a taken branch if it is either an unconditional |
1008 | branch or a conditional branch with a stricter branch condition. |
1009 | |
1010 | Also, exit if the branch has more than one set, since then it is computing |
1011 | other results that can't be ignored, e.g. the HPPA mov&branch instruction. |
1012 | ??? It may be possible to move other sets into INSN in addition to |
1013 | moving the instructions in the delay slots. |
1014 | |
1015 | We cannot steal the delay list if one of the instructions in the |
1016 | current delay_list modifies the condition codes and the jump in the |
1017 | sequence is a conditional jump. We cannot do this because we cannot |
1018 | change the direction of the jump because the condition codes |
1019 | will effect the direction of the jump in the sequence. */ |
1020 | |
1021 | CLEAR_RESOURCE (&cc_set); |
1022 | |
1023 | rtx_insn *trial; |
1024 | FOR_EACH_VEC_ELT (*delay_list, i, trial) |
1025 | { |
1026 | mark_set_resources (trial, &cc_set, 0, MARK_SRC_DEST_CALL); |
1027 | if (insn_references_resource_p (insn: seq->insn (index: 0), res: &cc_set, include_delayed_effects: false)) |
1028 | return; |
1029 | } |
1030 | |
1031 | if (XVECLEN (seq, 0) - 1 > slots_remaining |
1032 | || ! condition_dominates_p (condition, insn: seq->insn (index: 0)) |
1033 | || ! single_set (insn: seq->insn (index: 0))) |
1034 | return; |
1035 | |
1036 | /* On some targets, branches with delay slots can have a limited |
1037 | displacement. Give the back end a chance to tell us we can't do |
1038 | this. */ |
1039 | if (! targetm.can_follow_jump (insn, seq->insn (index: 0))) |
1040 | return; |
1041 | |
1042 | redundant = XALLOCAVEC (rtx_insn *, XVECLEN (seq, 0)); |
1043 | for (i = 1; i < seq->len (); i++) |
1044 | { |
1045 | rtx_insn *trial = seq->insn (index: i); |
1046 | int flags; |
1047 | |
1048 | if (insn_references_resource_p (insn: trial, res: sets, include_delayed_effects: false) |
1049 | || insn_sets_resource_p (insn: trial, res: needed, include_delayed_effects: false) |
1050 | || insn_sets_resource_p (insn: trial, res: sets, include_delayed_effects: false) |
1051 | /* If TRIAL is from the fallthrough code of an annulled branch insn |
1052 | in SEQ, we cannot use it. */ |
1053 | || (INSN_ANNULLED_BRANCH_P (seq->insn (0)) |
1054 | && ! INSN_FROM_TARGET_P (trial))) |
1055 | return; |
1056 | |
1057 | /* If this insn was already done (usually in a previous delay slot), |
1058 | pretend we put it in our delay slot. */ |
1059 | redundant[i] = redundant_insn (trial, insn, new_delay_list); |
1060 | if (redundant[i]) |
1061 | continue; |
1062 | |
1063 | /* We will end up re-vectoring this branch, so compute flags |
1064 | based on jumping to the new label. */ |
1065 | flags = get_jump_flags (insn, JUMP_LABEL (seq->insn (0))); |
1066 | |
1067 | if (! must_annul |
1068 | && ((condition == const_true_rtx |
1069 | || (! insn_sets_resource_p (insn: trial, res: other_needed, include_delayed_effects: false) |
1070 | && ! may_trap_or_fault_p (PATTERN (insn: trial))))) |
1071 | ? eligible_for_delay (insn, total_slots_filled, trial, flags) |
1072 | : (must_annul || (delay_list->is_empty () && new_delay_list.is_empty ())) |
1073 | && (must_annul = true, |
1074 | check_annul_list_true_false (annul_true_p: false, delay_list: *delay_list) |
1075 | && check_annul_list_true_false (annul_true_p: false, delay_list: new_delay_list) |
1076 | && eligible_for_annul_false (insn, total_slots_filled, |
1077 | trial, flags))) |
1078 | { |
1079 | if (must_annul) |
1080 | { |
1081 | /* Frame related instructions cannot go into annulled delay |
1082 | slots, it messes up the dwarf info. */ |
1083 | if (RTX_FRAME_RELATED_P (trial)) |
1084 | return; |
1085 | used_annul = true; |
1086 | } |
1087 | rtx_insn *temp = copy_delay_slot_insn (trial); |
1088 | INSN_FROM_TARGET_P (temp) = 1; |
1089 | add_to_delay_list (insn: temp, delay_list: &new_delay_list); |
1090 | total_slots_filled++; |
1091 | |
1092 | if (--slots_remaining == 0) |
1093 | break; |
1094 | } |
1095 | else |
1096 | return; |
1097 | } |
1098 | |
1099 | /* Record the effect of the instructions that were redundant and which |
1100 | we therefore decided not to copy. */ |
1101 | for (i = 1; i < seq->len (); i++) |
1102 | if (redundant[i]) |
1103 | { |
1104 | fix_reg_dead_note (redundant[i], insn); |
1105 | update_block (seq->insn (index: i), insn); |
1106 | } |
1107 | |
1108 | /* Show the place to which we will be branching. */ |
1109 | *pnew_thread = first_active_target_insn (JUMP_LABEL (seq->insn (0))); |
1110 | |
1111 | /* Add any new insns to the delay list and update the count of the |
1112 | number of slots filled. */ |
1113 | *pslots_filled = total_slots_filled; |
1114 | if (used_annul) |
1115 | *pannul_p = true; |
1116 | |
1117 | rtx_insn *temp; |
1118 | FOR_EACH_VEC_ELT (new_delay_list, i, temp) |
1119 | add_to_delay_list (insn: temp, delay_list); |
1120 | } |
1121 | |
1122 | /* Similar to steal_delay_list_from_target except that SEQ is on the |
1123 | fallthrough path of INSN. Here we only do something if the delay insn |
1124 | of SEQ is an unconditional branch. In that case we steal its delay slot |
1125 | for INSN since unconditional branches are much easier to fill. */ |
1126 | |
1127 | static void |
1128 | steal_delay_list_from_fallthrough (rtx_insn *insn, rtx condition, |
1129 | rtx_sequence *seq, |
1130 | vec<rtx_insn *> *delay_list, |
1131 | struct resources *sets, |
1132 | struct resources *needed, |
1133 | struct resources *other_needed, |
1134 | int slots_to_fill, int *pslots_filled, |
1135 | bool *pannul_p) |
1136 | { |
1137 | int i; |
1138 | int flags; |
1139 | bool must_annul = *pannul_p; |
1140 | bool used_annul = false; |
1141 | |
1142 | flags = get_jump_flags (insn, JUMP_LABEL (insn)); |
1143 | |
1144 | /* We can't do anything if SEQ's delay insn isn't an |
1145 | unconditional branch. */ |
1146 | |
1147 | if (! simplejump_or_return_p (insn: seq->insn (index: 0))) |
1148 | return; |
1149 | |
1150 | for (i = 1; i < seq->len (); i++) |
1151 | { |
1152 | rtx_insn *trial = seq->insn (index: i); |
1153 | rtx_insn *prior_insn; |
1154 | |
1155 | if (insn_references_resource_p (insn: trial, res: sets, include_delayed_effects: false) |
1156 | || insn_sets_resource_p (insn: trial, res: needed, include_delayed_effects: false) |
1157 | || insn_sets_resource_p (insn: trial, res: sets, include_delayed_effects: false)) |
1158 | break; |
1159 | |
1160 | /* If this insn was already done, we don't need it. */ |
1161 | if ((prior_insn = redundant_insn (trial, insn, *delay_list))) |
1162 | { |
1163 | fix_reg_dead_note (prior_insn, insn); |
1164 | update_block (trial, insn); |
1165 | delete_from_delay_slot (insn: trial); |
1166 | continue; |
1167 | } |
1168 | |
1169 | if (! must_annul |
1170 | && ((condition == const_true_rtx |
1171 | || (! insn_sets_resource_p (insn: trial, res: other_needed, include_delayed_effects: false) |
1172 | && ! may_trap_or_fault_p (PATTERN (insn: trial))))) |
1173 | ? eligible_for_delay (insn, *pslots_filled, trial, flags) |
1174 | : (must_annul || delay_list->is_empty ()) && (must_annul = true, |
1175 | check_annul_list_true_false (annul_true_p: true, delay_list: *delay_list) |
1176 | && eligible_for_annul_true (insn, *pslots_filled, trial, flags))) |
1177 | { |
1178 | if (must_annul) |
1179 | used_annul = true; |
1180 | delete_from_delay_slot (insn: trial); |
1181 | add_to_delay_list (insn: trial, delay_list); |
1182 | |
1183 | if (++(*pslots_filled) == slots_to_fill) |
1184 | break; |
1185 | } |
1186 | else |
1187 | break; |
1188 | } |
1189 | |
1190 | if (used_annul) |
1191 | *pannul_p = true; |
1192 | } |
1193 | |
1194 | /* Try merging insns starting at THREAD which match exactly the insns in |
1195 | INSN's delay list. |
1196 | |
1197 | If all insns were matched and the insn was previously annulling, the |
1198 | annul bit will be cleared. |
1199 | |
1200 | For each insn that is merged, if the branch is or will be non-annulling, |
1201 | we delete the merged insn. */ |
1202 | |
1203 | static void |
1204 | try_merge_delay_insns (rtx_insn *insn, rtx_insn *thread) |
1205 | { |
1206 | rtx_insn *trial, *next_trial; |
1207 | rtx_insn *delay_insn = as_a <rtx_insn *> (XVECEXP (PATTERN (insn), 0, 0)); |
1208 | bool annul_p = JUMP_P (delay_insn) && INSN_ANNULLED_BRANCH_P (delay_insn); |
1209 | int slot_number = 1; |
1210 | int num_slots = XVECLEN (PATTERN (insn), 0); |
1211 | rtx next_to_match = XVECEXP (PATTERN (insn), 0, slot_number); |
1212 | struct resources set, needed, modified; |
1213 | auto_vec<std::pair<rtx_insn *, bool>, 10> merged_insns; |
1214 | int flags; |
1215 | |
1216 | flags = get_jump_flags (insn: delay_insn, JUMP_LABEL (delay_insn)); |
1217 | |
1218 | CLEAR_RESOURCE (&needed); |
1219 | CLEAR_RESOURCE (&set); |
1220 | |
1221 | /* If this is not an annulling branch, take into account anything needed in |
1222 | INSN's delay slot. This prevents two increments from being incorrectly |
1223 | folded into one. If we are annulling, this would be the correct |
1224 | thing to do. (The alternative, looking at things set in NEXT_TO_MATCH |
1225 | will essentially disable this optimization. This method is somewhat of |
1226 | a kludge, but I don't see a better way.) */ |
1227 | if (! annul_p) |
1228 | for (int i = 1; i < num_slots; i++) |
1229 | if (XVECEXP (PATTERN (insn), 0, i)) |
1230 | mark_referenced_resources (XVECEXP (PATTERN (insn), 0, i), &needed, |
1231 | true); |
1232 | |
1233 | for (trial = thread; !stop_search_p (insn: trial, labels_p: true); trial = next_trial) |
1234 | { |
1235 | rtx pat = PATTERN (insn: trial); |
1236 | rtx oldtrial = trial; |
1237 | |
1238 | next_trial = next_nonnote_insn (trial); |
1239 | |
1240 | /* TRIAL must be a CALL_INSN or INSN. Skip USE and CLOBBER. */ |
1241 | if (NONJUMP_INSN_P (trial) |
1242 | && (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER)) |
1243 | continue; |
1244 | |
1245 | if (GET_CODE (next_to_match) == GET_CODE (trial) |
1246 | && ! insn_references_resource_p (insn: trial, res: &set, include_delayed_effects: true) |
1247 | && ! insn_sets_resource_p (insn: trial, res: &set, include_delayed_effects: true) |
1248 | && ! insn_sets_resource_p (insn: trial, res: &needed, include_delayed_effects: true) |
1249 | && (trial = try_split (pat, trial, 0)) != 0 |
1250 | /* Update next_trial, in case try_split succeeded. */ |
1251 | && (next_trial = next_nonnote_insn (trial)) |
1252 | /* Likewise THREAD. */ |
1253 | && (thread = oldtrial == thread ? trial : thread) |
1254 | && rtx_equal_p (PATTERN (insn: next_to_match), PATTERN (insn: trial)) |
1255 | /* Have to test this condition if annul condition is different |
1256 | from (and less restrictive than) non-annulling one. */ |
1257 | && eligible_for_delay (delay_insn, slot_number - 1, trial, flags)) |
1258 | { |
1259 | |
1260 | if (! annul_p) |
1261 | { |
1262 | update_block (trial, thread); |
1263 | if (trial == thread) |
1264 | thread = next_active_insn (thread); |
1265 | |
1266 | delete_related_insns (trial); |
1267 | INSN_FROM_TARGET_P (next_to_match) = 0; |
1268 | } |
1269 | else |
1270 | merged_insns.safe_push (obj: std::pair<rtx_insn *, bool> (trial, false)); |
1271 | |
1272 | if (++slot_number == num_slots) |
1273 | break; |
1274 | |
1275 | next_to_match = XVECEXP (PATTERN (insn), 0, slot_number); |
1276 | } |
1277 | |
1278 | mark_set_resources (trial, &set, 0, MARK_SRC_DEST_CALL); |
1279 | mark_referenced_resources (trial, &needed, true); |
1280 | } |
1281 | |
1282 | /* See if we stopped on a filled insn. If we did, try to see if its |
1283 | delay slots match. */ |
1284 | if (slot_number != num_slots |
1285 | && trial && NONJUMP_INSN_P (trial) |
1286 | && GET_CODE (PATTERN (trial)) == SEQUENCE |
1287 | && !(JUMP_P (XVECEXP (PATTERN (trial), 0, 0)) |
1288 | && INSN_ANNULLED_BRANCH_P (XVECEXP (PATTERN (trial), 0, 0)))) |
1289 | { |
1290 | rtx_sequence *pat = as_a <rtx_sequence *> (p: PATTERN (insn: trial)); |
1291 | rtx filled_insn = XVECEXP (pat, 0, 0); |
1292 | |
1293 | /* Account for resources set/needed by the filled insn. */ |
1294 | mark_set_resources (filled_insn, &set, 0, MARK_SRC_DEST_CALL); |
1295 | mark_referenced_resources (filled_insn, &needed, true); |
1296 | |
1297 | for (int i = 1; i < pat->len (); i++) |
1298 | { |
1299 | rtx_insn *dtrial = pat->insn (index: i); |
1300 | |
1301 | CLEAR_RESOURCE (&modified); |
1302 | /* Account for resources set by the insn following NEXT_TO_MATCH |
1303 | inside INSN's delay list. */ |
1304 | for (int j = 1; slot_number + j < num_slots; j++) |
1305 | mark_set_resources (XVECEXP (PATTERN (insn), 0, slot_number + j), |
1306 | &modified, 0, MARK_SRC_DEST_CALL); |
1307 | /* Account for resources set by the insn before DTRIAL and inside |
1308 | TRIAL's delay list. */ |
1309 | for (int j = 1; j < i; j++) |
1310 | mark_set_resources (XVECEXP (pat, 0, j), |
1311 | &modified, 0, MARK_SRC_DEST_CALL); |
1312 | if (! insn_references_resource_p (insn: dtrial, res: &set, include_delayed_effects: true) |
1313 | && ! insn_sets_resource_p (insn: dtrial, res: &set, include_delayed_effects: true) |
1314 | && ! insn_sets_resource_p (insn: dtrial, res: &needed, include_delayed_effects: true) |
1315 | && rtx_equal_p (PATTERN (insn: next_to_match), PATTERN (insn: dtrial)) |
1316 | /* Check that DTRIAL and NEXT_TO_MATCH does not reference a |
1317 | resource modified between them (only dtrial is checked because |
1318 | next_to_match and dtrial shall to be equal in order to hit |
1319 | this line) */ |
1320 | && ! insn_references_resource_p (insn: dtrial, res: &modified, include_delayed_effects: true) |
1321 | && eligible_for_delay (delay_insn, slot_number - 1, dtrial, flags)) |
1322 | { |
1323 | if (! annul_p) |
1324 | { |
1325 | rtx_insn *new_rtx; |
1326 | |
1327 | update_block (dtrial, thread); |
1328 | new_rtx = delete_from_delay_slot (insn: dtrial); |
1329 | if (thread->deleted ()) |
1330 | thread = new_rtx; |
1331 | INSN_FROM_TARGET_P (next_to_match) = 0; |
1332 | } |
1333 | else |
1334 | merged_insns.safe_push (obj: std::pair<rtx_insn *, bool> (dtrial, |
1335 | true)); |
1336 | |
1337 | if (++slot_number == num_slots) |
1338 | break; |
1339 | |
1340 | next_to_match = XVECEXP (PATTERN (insn), 0, slot_number); |
1341 | } |
1342 | else |
1343 | { |
1344 | /* Keep track of the set/referenced resources for the delay |
1345 | slots of any trial insns we encounter. */ |
1346 | mark_set_resources (dtrial, &set, 0, MARK_SRC_DEST_CALL); |
1347 | mark_referenced_resources (dtrial, &needed, true); |
1348 | } |
1349 | } |
1350 | } |
1351 | |
1352 | /* If all insns in the delay slot have been matched and we were previously |
1353 | annulling the branch, we need not any more. In that case delete all the |
1354 | merged insns. Also clear the INSN_FROM_TARGET_P bit of each insn in |
1355 | the delay list so that we know that it isn't only being used at the |
1356 | target. */ |
1357 | if (slot_number == num_slots && annul_p) |
1358 | { |
1359 | unsigned int len = merged_insns.length (); |
1360 | for (unsigned int i = len - 1; i < len; i--) |
1361 | if (merged_insns[i].second) |
1362 | { |
1363 | update_block (merged_insns[i].first, thread); |
1364 | rtx_insn *new_rtx = delete_from_delay_slot (insn: merged_insns[i].first); |
1365 | if (thread->deleted ()) |
1366 | thread = new_rtx; |
1367 | } |
1368 | else |
1369 | { |
1370 | update_block (merged_insns[i].first, thread); |
1371 | delete_related_insns (merged_insns[i].first); |
1372 | } |
1373 | |
1374 | INSN_ANNULLED_BRANCH_P (delay_insn) = 0; |
1375 | |
1376 | for (int i = 0; i < XVECLEN (PATTERN (insn), 0); i++) |
1377 | INSN_FROM_TARGET_P (XVECEXP (PATTERN (insn), 0, i)) = 0; |
1378 | } |
1379 | } |
1380 | |
1381 | /* See if INSN is redundant with an insn in front of TARGET. Often this |
1382 | is called when INSN is a candidate for a delay slot of TARGET. |
1383 | DELAY_LIST are insns that will be placed in delay slots of TARGET in front |
1384 | of INSN. Often INSN will be redundant with an insn in a delay slot of |
1385 | some previous insn. This happens when we have a series of branches to the |
1386 | same label; in that case the first insn at the target might want to go |
1387 | into each of the delay slots. |
1388 | |
1389 | If we are not careful, this routine can take up a significant fraction |
1390 | of the total compilation time (4%), but only wins rarely. Hence we |
1391 | speed this routine up by making two passes. The first pass goes back |
1392 | until it hits a label and sees if it finds an insn with an identical |
1393 | pattern. Only in this (relatively rare) event does it check for |
1394 | data conflicts. |
1395 | |
1396 | We do not split insns we encounter. This could cause us not to find a |
1397 | redundant insn, but the cost of splitting seems greater than the possible |
1398 | gain in rare cases. */ |
1399 | |
1400 | static rtx_insn * |
1401 | redundant_insn (rtx insn, rtx_insn *target, const vec<rtx_insn *> &delay_list) |
1402 | { |
1403 | rtx target_main = target; |
1404 | rtx ipat = PATTERN (insn); |
1405 | rtx_insn *trial; |
1406 | rtx pat; |
1407 | struct resources needed, set; |
1408 | int i; |
1409 | unsigned insns_to_search; |
1410 | |
1411 | /* If INSN has any REG_UNUSED notes, it can't match anything since we |
1412 | are allowed to not actually assign to such a register. */ |
1413 | if (find_reg_note (insn, REG_UNUSED, NULL_RTX) != 0) |
1414 | return 0; |
1415 | |
1416 | /* Scan backwards looking for a match. */ |
1417 | for (trial = PREV_INSN (insn: target), |
1418 | insns_to_search = param_max_delay_slot_insn_search; |
1419 | trial && insns_to_search > 0; |
1420 | trial = PREV_INSN (insn: trial)) |
1421 | { |
1422 | /* (use (insn))s can come immediately after a barrier if the |
1423 | label that used to precede them has been deleted as dead. |
1424 | See delete_related_insns. */ |
1425 | if (LABEL_P (trial) || BARRIER_P (trial)) |
1426 | return 0; |
1427 | |
1428 | if (!INSN_P (trial)) |
1429 | continue; |
1430 | --insns_to_search; |
1431 | |
1432 | pat = PATTERN (insn: trial); |
1433 | if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER) |
1434 | continue; |
1435 | |
1436 | if (GET_CODE (trial) == DEBUG_INSN) |
1437 | continue; |
1438 | |
1439 | if (rtx_sequence *seq = dyn_cast <rtx_sequence *> (p: pat)) |
1440 | { |
1441 | /* Stop for a CALL and its delay slots because it is difficult to |
1442 | track its resource needs correctly. */ |
1443 | if (CALL_P (seq->element (0))) |
1444 | return 0; |
1445 | |
1446 | /* Stop for an INSN or JUMP_INSN with delayed effects and its delay |
1447 | slots because it is difficult to track its resource needs |
1448 | correctly. */ |
1449 | |
1450 | if (INSN_SETS_ARE_DELAYED (seq->insn (0))) |
1451 | return 0; |
1452 | |
1453 | if (INSN_REFERENCES_ARE_DELAYED (seq->insn (0))) |
1454 | return 0; |
1455 | |
1456 | /* See if any of the insns in the delay slot match, updating |
1457 | resource requirements as we go. */ |
1458 | for (i = seq->len () - 1; i > 0; i--) |
1459 | if (GET_CODE (seq->element (i)) == GET_CODE (insn) |
1460 | && rtx_equal_p (PATTERN (insn: seq->element (index: i)), ipat) |
1461 | && ! find_reg_note (seq->element (index: i), REG_UNUSED, NULL_RTX)) |
1462 | break; |
1463 | |
1464 | /* If found a match, exit this loop early. */ |
1465 | if (i > 0) |
1466 | break; |
1467 | } |
1468 | |
1469 | else if (GET_CODE (trial) == GET_CODE (insn) && rtx_equal_p (pat, ipat) |
1470 | && ! find_reg_note (trial, REG_UNUSED, NULL_RTX)) |
1471 | break; |
1472 | } |
1473 | |
1474 | /* If we didn't find an insn that matches, return 0. */ |
1475 | if (trial == 0) |
1476 | return 0; |
1477 | |
1478 | /* See what resources this insn sets and needs. If they overlap, it |
1479 | can't be redundant. */ |
1480 | |
1481 | CLEAR_RESOURCE (&needed); |
1482 | CLEAR_RESOURCE (&set); |
1483 | mark_set_resources (insn, &set, 0, MARK_SRC_DEST_CALL); |
1484 | mark_referenced_resources (insn, &needed, true); |
1485 | |
1486 | /* If TARGET is a SEQUENCE, get the main insn. */ |
1487 | if (NONJUMP_INSN_P (target) && GET_CODE (PATTERN (target)) == SEQUENCE) |
1488 | target_main = XVECEXP (PATTERN (target), 0, 0); |
1489 | |
1490 | if (resource_conflicts_p (res1: &needed, res2: &set) |
1491 | /* The insn requiring the delay may not set anything needed or set by |
1492 | INSN. */ |
1493 | || insn_sets_resource_p (insn: target_main, res: &needed, include_delayed_effects: true) |
1494 | || insn_sets_resource_p (insn: target_main, res: &set, include_delayed_effects: true)) |
1495 | return 0; |
1496 | |
1497 | /* Insns we pass may not set either NEEDED or SET, so merge them for |
1498 | simpler tests. */ |
1499 | needed.memory |= set.memory; |
1500 | needed.regs |= set.regs; |
1501 | |
1502 | /* This insn isn't redundant if it conflicts with an insn that either is |
1503 | or will be in a delay slot of TARGET. */ |
1504 | |
1505 | unsigned int j; |
1506 | rtx_insn *temp; |
1507 | FOR_EACH_VEC_ELT (delay_list, j, temp) |
1508 | if (insn_sets_resource_p (insn: temp, res: &needed, include_delayed_effects: true)) |
1509 | return 0; |
1510 | |
1511 | if (NONJUMP_INSN_P (target) && GET_CODE (PATTERN (target)) == SEQUENCE) |
1512 | for (i = 1; i < XVECLEN (PATTERN (target), 0); i++) |
1513 | if (insn_sets_resource_p (XVECEXP (PATTERN (target), 0, i), res: &needed, |
1514 | include_delayed_effects: true)) |
1515 | return 0; |
1516 | |
1517 | /* Scan backwards until we reach a label or an insn that uses something |
1518 | INSN sets or sets something insn uses or sets. */ |
1519 | |
1520 | for (trial = PREV_INSN (insn: target), |
1521 | insns_to_search = param_max_delay_slot_insn_search; |
1522 | trial && !LABEL_P (trial) && insns_to_search > 0; |
1523 | trial = PREV_INSN (insn: trial)) |
1524 | { |
1525 | if (!INSN_P (trial)) |
1526 | continue; |
1527 | --insns_to_search; |
1528 | |
1529 | pat = PATTERN (insn: trial); |
1530 | if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER) |
1531 | continue; |
1532 | |
1533 | if (GET_CODE (trial) == DEBUG_INSN) |
1534 | continue; |
1535 | |
1536 | if (rtx_sequence *seq = dyn_cast <rtx_sequence *> (p: pat)) |
1537 | { |
1538 | bool annul_p = false; |
1539 | rtx_insn *control = seq->insn (index: 0); |
1540 | |
1541 | /* If this is a CALL_INSN and its delay slots, it is hard to track |
1542 | the resource needs properly, so give up. */ |
1543 | if (CALL_P (control)) |
1544 | return 0; |
1545 | |
1546 | /* If this is an INSN or JUMP_INSN with delayed effects, it |
1547 | is hard to track the resource needs properly, so give up. */ |
1548 | |
1549 | if (INSN_SETS_ARE_DELAYED (control)) |
1550 | return 0; |
1551 | |
1552 | if (INSN_REFERENCES_ARE_DELAYED (control)) |
1553 | return 0; |
1554 | |
1555 | if (JUMP_P (control)) |
1556 | annul_p = INSN_ANNULLED_BRANCH_P (control); |
1557 | |
1558 | /* See if any of the insns in the delay slot match, updating |
1559 | resource requirements as we go. */ |
1560 | for (i = seq->len () - 1; i > 0; i--) |
1561 | { |
1562 | rtx_insn *candidate = seq->insn (index: i); |
1563 | |
1564 | /* If an insn will be annulled if the branch is false, it isn't |
1565 | considered as a possible duplicate insn. */ |
1566 | if (rtx_equal_p (PATTERN (insn: candidate), ipat) |
1567 | && ! (annul_p && INSN_FROM_TARGET_P (candidate))) |
1568 | { |
1569 | /* Show that this insn will be used in the sequel. */ |
1570 | INSN_FROM_TARGET_P (candidate) = 0; |
1571 | return candidate; |
1572 | } |
1573 | |
1574 | /* Unless this is an annulled insn from the target of a branch, |
1575 | we must stop if it sets anything needed or set by INSN. */ |
1576 | if ((!annul_p || !INSN_FROM_TARGET_P (candidate)) |
1577 | && insn_sets_resource_p (insn: candidate, res: &needed, include_delayed_effects: true)) |
1578 | return 0; |
1579 | } |
1580 | |
1581 | /* If the insn requiring the delay slot conflicts with INSN, we |
1582 | must stop. */ |
1583 | if (insn_sets_resource_p (insn: control, res: &needed, include_delayed_effects: true)) |
1584 | return 0; |
1585 | } |
1586 | else |
1587 | { |
1588 | /* See if TRIAL is the same as INSN. */ |
1589 | pat = PATTERN (insn: trial); |
1590 | if (rtx_equal_p (pat, ipat)) |
1591 | return trial; |
1592 | |
1593 | /* Can't go any further if TRIAL conflicts with INSN. */ |
1594 | if (insn_sets_resource_p (insn: trial, res: &needed, include_delayed_effects: true)) |
1595 | return 0; |
1596 | } |
1597 | } |
1598 | |
1599 | return 0; |
1600 | } |
1601 | |
1602 | /* Return true if THREAD can only be executed in one way. If LABEL is nonzero, |
1603 | it is the target of the branch insn being scanned. If ALLOW_FALLTHROUGH |
1604 | is true, we are allowed to fall into this thread; otherwise, we are not. |
1605 | |
1606 | If LABEL is used more than one or we pass a label other than LABEL before |
1607 | finding an active insn, we do not own this thread. */ |
1608 | |
1609 | static bool |
1610 | own_thread_p (rtx thread, rtx label, bool allow_fallthrough) |
1611 | { |
1612 | rtx_insn *active_insn; |
1613 | rtx_insn *insn; |
1614 | |
1615 | /* We don't own the function end. */ |
1616 | if (thread == 0 || ANY_RETURN_P (thread)) |
1617 | return false; |
1618 | |
1619 | /* We have a non-NULL insn. */ |
1620 | rtx_insn *thread_insn = as_a <rtx_insn *> (p: thread); |
1621 | |
1622 | /* Get the first active insn, or THREAD_INSN, if it is an active insn. */ |
1623 | active_insn = next_active_insn (PREV_INSN (insn: thread_insn)); |
1624 | |
1625 | for (insn = thread_insn; insn != active_insn; insn = NEXT_INSN (insn)) |
1626 | if (LABEL_P (insn) |
1627 | && (insn != label || LABEL_NUSES (insn) != 1)) |
1628 | return false; |
1629 | |
1630 | if (allow_fallthrough) |
1631 | return true; |
1632 | |
1633 | /* Ensure that we reach a BARRIER before any insn or label. */ |
1634 | for (insn = prev_nonnote_insn (thread_insn); |
1635 | insn == 0 || !BARRIER_P (insn); |
1636 | insn = prev_nonnote_insn (insn)) |
1637 | if (insn == 0 |
1638 | || LABEL_P (insn) |
1639 | || (NONJUMP_INSN_P (insn) |
1640 | && GET_CODE (PATTERN (insn)) != USE |
1641 | && GET_CODE (PATTERN (insn)) != CLOBBER)) |
1642 | return false; |
1643 | |
1644 | return true; |
1645 | } |
1646 | |
1647 | /* Called when INSN is being moved from a location near the target of a jump. |
1648 | We leave a marker of the form (use (INSN)) immediately in front of WHERE |
1649 | for mark_target_live_regs. These markers will be deleted at the end. |
1650 | |
1651 | We used to try to update the live status of registers if WHERE is at |
1652 | the start of a basic block, but that can't work since we may remove a |
1653 | BARRIER in relax_delay_slots. */ |
1654 | |
1655 | static void |
1656 | update_block (rtx_insn *insn, rtx_insn *where) |
1657 | { |
1658 | emit_insn_before (gen_rtx_USE (VOIDmode, insn), where); |
1659 | |
1660 | /* INSN might be making a value live in a block where it didn't use to |
1661 | be. So recompute liveness information for this block. */ |
1662 | incr_ticks_for_insn (insn); |
1663 | } |
1664 | |
1665 | /* Similar to REDIRECT_JUMP except that we update the BB_TICKS entry for |
1666 | the basic block containing the jump. */ |
1667 | |
1668 | static bool |
1669 | reorg_redirect_jump (rtx_jump_insn *jump, rtx nlabel) |
1670 | { |
1671 | incr_ticks_for_insn (jump); |
1672 | return redirect_jump (jump, nlabel, 1); |
1673 | } |
1674 | |
1675 | /* Called when INSN is being moved forward into a delay slot of DELAYED_INSN. |
1676 | We check every instruction between INSN and DELAYED_INSN for REG_DEAD notes |
1677 | that reference values used in INSN. If we find one, then we move the |
1678 | REG_DEAD note to INSN. |
1679 | |
1680 | This is needed to handle the case where a later insn (after INSN) has a |
1681 | REG_DEAD note for a register used by INSN, and this later insn subsequently |
1682 | gets moved before a CODE_LABEL because it is a redundant insn. In this |
1683 | case, mark_target_live_regs may be confused into thinking the register |
1684 | is dead because it sees a REG_DEAD note immediately before a CODE_LABEL. */ |
1685 | |
1686 | static void |
1687 | update_reg_dead_notes (rtx_insn *insn, rtx_insn *delayed_insn) |
1688 | { |
1689 | rtx link, next; |
1690 | rtx_insn *p; |
1691 | |
1692 | for (p = next_nonnote_insn (insn); p != delayed_insn; |
1693 | p = next_nonnote_insn (p)) |
1694 | for (link = REG_NOTES (p); link; link = next) |
1695 | { |
1696 | next = XEXP (link, 1); |
1697 | |
1698 | if (REG_NOTE_KIND (link) != REG_DEAD |
1699 | || !REG_P (XEXP (link, 0))) |
1700 | continue; |
1701 | |
1702 | if (reg_referenced_p (XEXP (link, 0), PATTERN (insn))) |
1703 | { |
1704 | /* Move the REG_DEAD note from P to INSN. */ |
1705 | remove_note (p, link); |
1706 | XEXP (link, 1) = REG_NOTES (insn); |
1707 | REG_NOTES (insn) = link; |
1708 | } |
1709 | } |
1710 | } |
1711 | |
1712 | /* Called when an insn redundant with start_insn is deleted. If there |
1713 | is a REG_DEAD note for the target of start_insn between start_insn |
1714 | and stop_insn, then the REG_DEAD note needs to be deleted since the |
1715 | value no longer dies there. |
1716 | |
1717 | If the REG_DEAD note isn't deleted, then mark_target_live_regs may be |
1718 | confused into thinking the register is dead. */ |
1719 | |
1720 | static void |
1721 | fix_reg_dead_note (rtx_insn *start_insn, rtx stop_insn) |
1722 | { |
1723 | rtx link, next; |
1724 | rtx_insn *p; |
1725 | |
1726 | for (p = next_nonnote_insn (start_insn); p != stop_insn; |
1727 | p = next_nonnote_insn (p)) |
1728 | for (link = REG_NOTES (p); link; link = next) |
1729 | { |
1730 | next = XEXP (link, 1); |
1731 | |
1732 | if (REG_NOTE_KIND (link) != REG_DEAD |
1733 | || !REG_P (XEXP (link, 0))) |
1734 | continue; |
1735 | |
1736 | if (reg_set_p (XEXP (link, 0), PATTERN (insn: start_insn))) |
1737 | { |
1738 | remove_note (p, link); |
1739 | return; |
1740 | } |
1741 | } |
1742 | } |
1743 | |
1744 | /* Delete any REG_UNUSED notes that exist on INSN but not on OTHER_INSN. |
1745 | |
1746 | This handles the case of udivmodXi4 instructions which optimize their |
1747 | output depending on whether any REG_UNUSED notes are present. We must |
1748 | make sure that INSN calculates as many results as OTHER_INSN does. */ |
1749 | |
1750 | static void |
1751 | update_reg_unused_notes (rtx_insn *insn, rtx other_insn) |
1752 | { |
1753 | rtx link, next; |
1754 | |
1755 | for (link = REG_NOTES (insn); link; link = next) |
1756 | { |
1757 | next = XEXP (link, 1); |
1758 | |
1759 | if (REG_NOTE_KIND (link) != REG_UNUSED |
1760 | || !REG_P (XEXP (link, 0))) |
1761 | continue; |
1762 | |
1763 | if (!find_regno_note (other_insn, REG_UNUSED, REGNO (XEXP (link, 0)))) |
1764 | remove_note (insn, link); |
1765 | } |
1766 | } |
1767 | |
1768 | static vec <rtx> sibling_labels; |
1769 | |
1770 | /* Return the label before INSN, or put a new label there. If SIBLING is |
1771 | non-zero, it is another label associated with the new label (if any), |
1772 | typically the former target of the jump that will be redirected to |
1773 | the new label. */ |
1774 | |
1775 | static rtx_insn * |
1776 | get_label_before (rtx_insn *insn, rtx sibling) |
1777 | { |
1778 | rtx_insn *label; |
1779 | |
1780 | /* Find an existing label at this point |
1781 | or make a new one if there is none. */ |
1782 | label = prev_nonnote_insn (insn); |
1783 | |
1784 | if (label == 0 || !LABEL_P (label)) |
1785 | { |
1786 | rtx_insn *prev = PREV_INSN (insn); |
1787 | |
1788 | label = gen_label_rtx (); |
1789 | emit_label_after (label, prev); |
1790 | LABEL_NUSES (label) = 0; |
1791 | if (sibling) |
1792 | { |
1793 | sibling_labels.safe_push (obj: label); |
1794 | sibling_labels.safe_push (obj: sibling); |
1795 | } |
1796 | } |
1797 | return label; |
1798 | } |
1799 | |
1800 | /* Scan a function looking for insns that need a delay slot and find insns to |
1801 | put into the delay slot. |
1802 | |
1803 | NON_JUMPS_P is true if we are to only try to fill non-jump insns (such |
1804 | as calls). We do these first since we don't want jump insns (that are |
1805 | easier to fill) to get the only insns that could be used for non-jump insns. |
1806 | When it is zero, only try to fill JUMP_INSNs. |
1807 | |
1808 | When slots are filled in this manner, the insns (including the |
1809 | delay_insn) are put together in a SEQUENCE rtx. In this fashion, |
1810 | it is possible to tell whether a delay slot has really been filled |
1811 | or not. `final' knows how to deal with this, by communicating |
1812 | through FINAL_SEQUENCE. */ |
1813 | |
1814 | static void |
1815 | fill_simple_delay_slots (bool non_jumps_p) |
1816 | { |
1817 | rtx_insn *insn, *trial, *next_trial; |
1818 | rtx pat; |
1819 | int i; |
1820 | int num_unfilled_slots = unfilled_slots_next - unfilled_slots_base; |
1821 | struct resources needed, set; |
1822 | int slots_to_fill, slots_filled; |
1823 | auto_vec<rtx_insn *, 5> delay_list; |
1824 | |
1825 | for (i = 0; i < num_unfilled_slots; i++) |
1826 | { |
1827 | int flags; |
1828 | /* Get the next insn to fill. If it has already had any slots assigned, |
1829 | we can't do anything with it. Maybe we'll improve this later. */ |
1830 | |
1831 | insn = unfilled_slots_base[i]; |
1832 | if (insn == 0 |
1833 | || insn->deleted () |
1834 | || (NONJUMP_INSN_P (insn) |
1835 | && GET_CODE (PATTERN (insn)) == SEQUENCE) |
1836 | || (JUMP_P (insn) && non_jumps_p) |
1837 | || (!JUMP_P (insn) && ! non_jumps_p)) |
1838 | continue; |
1839 | |
1840 | /* It may have been that this insn used to need delay slots, but |
1841 | now doesn't; ignore in that case. This can happen, for example, |
1842 | on the HP PA RISC, where the number of delay slots depends on |
1843 | what insns are nearby. */ |
1844 | slots_to_fill = num_delay_slots (insn); |
1845 | |
1846 | /* Some machine description have defined instructions to have |
1847 | delay slots only in certain circumstances which may depend on |
1848 | nearby insns (which change due to reorg's actions). |
1849 | |
1850 | For example, the PA port normally has delay slots for unconditional |
1851 | jumps. |
1852 | |
1853 | However, the PA port claims such jumps do not have a delay slot |
1854 | if they are immediate successors of certain CALL_INSNs. This |
1855 | allows the port to favor filling the delay slot of the call with |
1856 | the unconditional jump. */ |
1857 | if (slots_to_fill == 0) |
1858 | continue; |
1859 | |
1860 | /* This insn needs, or can use, some delay slots. SLOTS_TO_FILL |
1861 | says how many. After initialization, first try optimizing |
1862 | |
1863 | call _foo call _foo |
1864 | nop add %o7,.-L1,%o7 |
1865 | b,a L1 |
1866 | nop |
1867 | |
1868 | If this case applies, the delay slot of the call is filled with |
1869 | the unconditional jump. This is done first to avoid having the |
1870 | delay slot of the call filled in the backward scan. Also, since |
1871 | the unconditional jump is likely to also have a delay slot, that |
1872 | insn must exist when it is subsequently scanned. |
1873 | |
1874 | This is tried on each insn with delay slots as some machines |
1875 | have insns which perform calls, but are not represented as |
1876 | CALL_INSNs. */ |
1877 | |
1878 | slots_filled = 0; |
1879 | delay_list.truncate (size: 0); |
1880 | |
1881 | if (JUMP_P (insn)) |
1882 | flags = get_jump_flags (insn, JUMP_LABEL (insn)); |
1883 | else |
1884 | flags = get_jump_flags (insn, NULL_RTX); |
1885 | |
1886 | if ((trial = next_active_insn (insn)) |
1887 | && JUMP_P (trial) |
1888 | && simplejump_p (trial) |
1889 | && eligible_for_delay (insn, slots_filled, trial, flags) |
1890 | && no_labels_between_p (insn, trial) |
1891 | && ! can_throw_internal (trial)) |
1892 | { |
1893 | rtx_insn **tmp; |
1894 | slots_filled++; |
1895 | add_to_delay_list (insn: trial, delay_list: &delay_list); |
1896 | |
1897 | /* TRIAL may have had its delay slot filled, then unfilled. When |
1898 | the delay slot is unfilled, TRIAL is placed back on the unfilled |
1899 | slots obstack. Unfortunately, it is placed on the end of the |
1900 | obstack, not in its original location. Therefore, we must search |
1901 | from entry i + 1 to the end of the unfilled slots obstack to |
1902 | try and find TRIAL. */ |
1903 | tmp = &unfilled_slots_base[i + 1]; |
1904 | while (*tmp != trial && tmp != unfilled_slots_next) |
1905 | tmp++; |
1906 | |
1907 | /* Remove the unconditional jump from consideration for delay slot |
1908 | filling and unthread it. */ |
1909 | if (*tmp == trial) |
1910 | *tmp = 0; |
1911 | { |
1912 | rtx_insn *next = NEXT_INSN (insn: trial); |
1913 | rtx_insn *prev = PREV_INSN (insn: trial); |
1914 | if (prev) |
1915 | SET_NEXT_INSN (prev) = next; |
1916 | if (next) |
1917 | SET_PREV_INSN (next) = prev; |
1918 | } |
1919 | } |
1920 | |
1921 | /* Now, scan backwards from the insn to search for a potential |
1922 | delay-slot candidate. Stop searching when a label or jump is hit. |
1923 | |
1924 | For each candidate, if it is to go into the delay slot (moved |
1925 | forward in execution sequence), it must not need or set any resources |
1926 | that were set by later insns and must not set any resources that |
1927 | are needed for those insns. |
1928 | |
1929 | The delay slot insn itself sets resources unless it is a call |
1930 | (in which case the called routine, not the insn itself, is doing |
1931 | the setting). */ |
1932 | |
1933 | if (slots_filled < slots_to_fill) |
1934 | { |
1935 | /* If the flags register is dead after the insn, then we want to be |
1936 | able to accept a candidate that clobbers it. For this purpose, |
1937 | we need to filter the flags register during life analysis, so |
1938 | that it doesn't create RAW and WAW dependencies, while still |
1939 | creating the necessary WAR dependencies. */ |
1940 | bool filter_flags |
1941 | = (slots_to_fill == 1 |
1942 | && targetm.flags_regnum != INVALID_REGNUM |
1943 | && find_regno_note (insn, REG_DEAD, targetm.flags_regnum)); |
1944 | struct resources fset; |
1945 | CLEAR_RESOURCE (&needed); |
1946 | CLEAR_RESOURCE (&set); |
1947 | mark_set_resources (insn, &set, 0, MARK_SRC_DEST); |
1948 | if (filter_flags) |
1949 | { |
1950 | CLEAR_RESOURCE (&fset); |
1951 | mark_set_resources (insn, &fset, 0, MARK_SRC_DEST); |
1952 | } |
1953 | mark_referenced_resources (insn, &needed, false); |
1954 | |
1955 | for (trial = prev_nonnote_insn (insn); ! stop_search_p (insn: trial, labels_p: true); |
1956 | trial = next_trial) |
1957 | { |
1958 | next_trial = prev_nonnote_insn (trial); |
1959 | |
1960 | /* This must be an INSN or CALL_INSN. */ |
1961 | pat = PATTERN (insn: trial); |
1962 | |
1963 | /* Stand-alone USE and CLOBBER are just for flow. */ |
1964 | if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER) |
1965 | continue; |
1966 | |
1967 | /* And DEBUG_INSNs never go into delay slots. */ |
1968 | if (GET_CODE (trial) == DEBUG_INSN) |
1969 | continue; |
1970 | |
1971 | /* Check for resource conflict first, to avoid unnecessary |
1972 | splitting. */ |
1973 | if (! insn_references_resource_p (insn: trial, res: &set, include_delayed_effects: true) |
1974 | && ! insn_sets_resource_p (insn: trial, |
1975 | res: filter_flags ? &fset : &set, |
1976 | include_delayed_effects: true) |
1977 | && ! insn_sets_resource_p (insn: trial, res: &needed, include_delayed_effects: true) |
1978 | && ! can_throw_internal (trial)) |
1979 | { |
1980 | trial = try_split (pat, trial, 1); |
1981 | next_trial = prev_nonnote_insn (trial); |
1982 | if (eligible_for_delay (insn, slots_filled, trial, flags)) |
1983 | { |
1984 | /* In this case, we are searching backward, so if we |
1985 | find insns to put on the delay list, we want |
1986 | to put them at the head, rather than the |
1987 | tail, of the list. */ |
1988 | |
1989 | update_reg_dead_notes (insn: trial, delayed_insn: insn); |
1990 | delay_list.safe_insert (ix: 0, obj: trial); |
1991 | update_block (insn: trial, where: trial); |
1992 | delete_related_insns (trial); |
1993 | if (slots_to_fill == ++slots_filled) |
1994 | break; |
1995 | continue; |
1996 | } |
1997 | } |
1998 | |
1999 | mark_set_resources (trial, &set, 0, MARK_SRC_DEST_CALL); |
2000 | if (filter_flags) |
2001 | { |
2002 | mark_set_resources (trial, &fset, 0, MARK_SRC_DEST_CALL); |
2003 | /* If the flags register is set, then it doesn't create RAW |
2004 | dependencies any longer and it also doesn't create WAW |
2005 | dependencies since it's dead after the original insn. */ |
2006 | if (TEST_HARD_REG_BIT (set: fset.regs, bit: targetm.flags_regnum)) |
2007 | { |
2008 | CLEAR_HARD_REG_BIT (set&: needed.regs, bit: targetm.flags_regnum); |
2009 | CLEAR_HARD_REG_BIT (set&: fset.regs, bit: targetm.flags_regnum); |
2010 | } |
2011 | } |
2012 | mark_referenced_resources (trial, &needed, true); |
2013 | } |
2014 | } |
2015 | |
2016 | /* If all needed slots haven't been filled, we come here. */ |
2017 | |
2018 | /* Try to optimize case of jumping around a single insn. */ |
2019 | if ((ANNUL_IFTRUE_SLOTS || ANNUL_IFFALSE_SLOTS) |
2020 | && slots_filled != slots_to_fill |
2021 | && delay_list.is_empty () |
2022 | && JUMP_P (insn) |
2023 | && (condjump_p (insn) || condjump_in_parallel_p (insn)) |
2024 | && !ANY_RETURN_P (JUMP_LABEL (insn))) |
2025 | { |
2026 | optimize_skip (insn: as_a <rtx_jump_insn *> (p: insn), delay_list: &delay_list); |
2027 | if (!delay_list.is_empty ()) |
2028 | slots_filled += 1; |
2029 | } |
2030 | |
2031 | /* Try to get insns from beyond the insn needing the delay slot. |
2032 | These insns can neither set or reference resources set in insns being |
2033 | skipped, cannot set resources in the insn being skipped, and, if this |
2034 | is a CALL_INSN (or a CALL_INSN is passed), cannot trap (because the |
2035 | call might not return). |
2036 | |
2037 | There used to be code which continued past the target label if |
2038 | we saw all uses of the target label. This code did not work, |
2039 | because it failed to account for some instructions which were |
2040 | both annulled and marked as from the target. This can happen as a |
2041 | result of optimize_skip. Since this code was redundant with |
2042 | fill_eager_delay_slots anyways, it was just deleted. */ |
2043 | |
2044 | if (slots_filled != slots_to_fill |
2045 | /* If this instruction could throw an exception which is |
2046 | caught in the same function, then it's not safe to fill |
2047 | the delay slot with an instruction from beyond this |
2048 | point. For example, consider: |
2049 | |
2050 | int i = 2; |
2051 | |
2052 | try { |
2053 | f(); |
2054 | i = 3; |
2055 | } catch (...) {} |
2056 | |
2057 | return i; |
2058 | |
2059 | Even though `i' is a local variable, we must be sure not |
2060 | to put `i = 3' in the delay slot if `f' might throw an |
2061 | exception. |
2062 | |
2063 | Presumably, we should also check to see if we could get |
2064 | back to this function via `setjmp'. */ |
2065 | && ! can_throw_internal (insn) |
2066 | && !JUMP_P (insn)) |
2067 | { |
2068 | bool maybe_never = false; |
2069 | rtx pat, trial_delay; |
2070 | |
2071 | CLEAR_RESOURCE (&needed); |
2072 | CLEAR_RESOURCE (&set); |
2073 | mark_set_resources (insn, &set, 0, MARK_SRC_DEST_CALL); |
2074 | mark_referenced_resources (insn, &needed, true); |
2075 | |
2076 | if (CALL_P (insn)) |
2077 | maybe_never = true; |
2078 | |
2079 | for (trial = next_nonnote_insn (insn); !stop_search_p (insn: trial, labels_p: true); |
2080 | trial = next_trial) |
2081 | { |
2082 | next_trial = next_nonnote_insn (trial); |
2083 | |
2084 | /* This must be an INSN or CALL_INSN. */ |
2085 | pat = PATTERN (insn: trial); |
2086 | |
2087 | /* Stand-alone USE and CLOBBER are just for flow. */ |
2088 | if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER) |
2089 | continue; |
2090 | |
2091 | /* And DEBUG_INSNs do not go in delay slots. */ |
2092 | if (GET_CODE (trial) == DEBUG_INSN) |
2093 | continue; |
2094 | |
2095 | /* If this already has filled delay slots, get the insn needing |
2096 | the delay slots. */ |
2097 | if (GET_CODE (pat) == SEQUENCE) |
2098 | trial_delay = XVECEXP (pat, 0, 0); |
2099 | else |
2100 | trial_delay = trial; |
2101 | |
2102 | /* Stop our search when seeing a jump. */ |
2103 | if (JUMP_P (trial_delay)) |
2104 | break; |
2105 | |
2106 | /* See if we have a resource problem before we try to split. */ |
2107 | if (GET_CODE (pat) != SEQUENCE |
2108 | && ! insn_references_resource_p (insn: trial, res: &set, include_delayed_effects: true) |
2109 | && ! insn_sets_resource_p (insn: trial, res: &set, include_delayed_effects: true) |
2110 | && ! insn_sets_resource_p (insn: trial, res: &needed, include_delayed_effects: true) |
2111 | && ! (maybe_never && may_trap_or_fault_p (pat)) |
2112 | && (trial = try_split (pat, trial, 0)) |
2113 | && eligible_for_delay (insn, slots_filled, trial, flags) |
2114 | && ! can_throw_internal (trial)) |
2115 | { |
2116 | next_trial = next_nonnote_insn (trial); |
2117 | add_to_delay_list (insn: trial, delay_list: &delay_list); |
2118 | |
2119 | delete_related_insns (trial); |
2120 | if (slots_to_fill == ++slots_filled) |
2121 | break; |
2122 | continue; |
2123 | } |
2124 | |
2125 | mark_set_resources (trial, &set, 0, MARK_SRC_DEST_CALL); |
2126 | mark_referenced_resources (trial, &needed, true); |
2127 | |
2128 | /* Ensure we don't put insns between the setting of cc and the |
2129 | comparison by moving a setting of cc into an earlier delay |
2130 | slot since these insns could clobber the condition code. */ |
2131 | set.cc = 1; |
2132 | |
2133 | /* If this is a call, we might not get here. */ |
2134 | if (CALL_P (trial_delay)) |
2135 | maybe_never = true; |
2136 | } |
2137 | |
2138 | /* If there are slots left to fill and our search was stopped by an |
2139 | unconditional branch, try the insn at the branch target. We can |
2140 | redirect the branch if it works. |
2141 | |
2142 | Don't do this if the insn at the branch target is a branch. */ |
2143 | if (slots_to_fill != slots_filled |
2144 | && trial |
2145 | && jump_to_label_p (trial) |
2146 | && simplejump_p (trial) |
2147 | && (next_trial = next_active_insn (JUMP_LABEL_AS_INSN (insn: trial))) != 0 |
2148 | && ! (NONJUMP_INSN_P (next_trial) |
2149 | && GET_CODE (PATTERN (next_trial)) == SEQUENCE) |
2150 | && !JUMP_P (next_trial) |
2151 | && ! insn_references_resource_p (insn: next_trial, res: &set, include_delayed_effects: true) |
2152 | && ! insn_sets_resource_p (insn: next_trial, res: &set, include_delayed_effects: true) |
2153 | && ! insn_sets_resource_p (insn: next_trial, res: &needed, include_delayed_effects: true) |
2154 | && ! (maybe_never && may_trap_or_fault_p (PATTERN (insn: next_trial))) |
2155 | && (next_trial = try_split (PATTERN (insn: next_trial), next_trial, 0)) |
2156 | && eligible_for_delay (insn, slots_filled, next_trial, flags) |
2157 | && ! can_throw_internal (trial)) |
2158 | { |
2159 | /* See comment in relax_delay_slots about necessity of using |
2160 | next_real_nondebug_insn here. */ |
2161 | rtx_insn *new_label = next_real_nondebug_insn (next_trial); |
2162 | |
2163 | if (new_label != 0) |
2164 | new_label = get_label_before (insn: new_label, JUMP_LABEL (trial)); |
2165 | else |
2166 | new_label = find_end_label (kind: simple_return_rtx); |
2167 | |
2168 | if (new_label) |
2169 | { |
2170 | add_to_delay_list (insn: copy_delay_slot_insn (next_trial), |
2171 | delay_list: &delay_list); |
2172 | slots_filled++; |
2173 | reorg_redirect_jump (jump: as_a <rtx_jump_insn *> (p: trial), |
2174 | nlabel: new_label); |
2175 | } |
2176 | } |
2177 | } |
2178 | |
2179 | /* If this is an unconditional jump, then try to get insns from the |
2180 | target of the jump. */ |
2181 | rtx_jump_insn *jump_insn; |
2182 | if ((jump_insn = dyn_cast <rtx_jump_insn *> (p: insn)) |
2183 | && simplejump_p (jump_insn) |
2184 | && slots_filled != slots_to_fill) |
2185 | fill_slots_from_thread (jump_insn, const_true_rtx, |
2186 | next_active_insn (JUMP_LABEL_AS_INSN (insn)), |
2187 | NULL, 1, 1, own_thread_p (JUMP_LABEL (insn), |
2188 | JUMP_LABEL (insn), allow_fallthrough: false), |
2189 | slots_to_fill, &slots_filled, &delay_list); |
2190 | |
2191 | if (!delay_list.is_empty ()) |
2192 | unfilled_slots_base[i] |
2193 | = emit_delay_sequence (insn, list: delay_list, length: slots_filled); |
2194 | |
2195 | if (slots_to_fill == slots_filled) |
2196 | unfilled_slots_base[i] = 0; |
2197 | |
2198 | note_delay_statistics (slots_filled, index: 0); |
2199 | } |
2200 | } |
2201 | |
2202 | /* Follow any unconditional jump at LABEL, for the purpose of redirecting JUMP; |
2203 | return the ultimate label reached by any such chain of jumps. |
2204 | Return a suitable return rtx if the chain ultimately leads to a |
2205 | return instruction. |
2206 | If LABEL is not followed by a jump, return LABEL. |
2207 | If the chain loops or we can't find end, return LABEL, |
2208 | since that tells caller to avoid changing the insn. |
2209 | If the returned label is obtained by following a crossing jump, |
2210 | set *CROSSING to true, otherwise set it to false. */ |
2211 | |
2212 | static rtx |
2213 | follow_jumps (rtx label, rtx_insn *jump, bool *crossing) |
2214 | { |
2215 | rtx_insn *insn; |
2216 | rtx_insn *next; |
2217 | int depth; |
2218 | |
2219 | *crossing = false; |
2220 | if (ANY_RETURN_P (label)) |
2221 | return label; |
2222 | |
2223 | rtx_insn *value = as_a <rtx_insn *> (p: label); |
2224 | |
2225 | for (depth = 0; |
2226 | (depth < 10 |
2227 | && (insn = next_active_insn (value)) != 0 |
2228 | && JUMP_P (insn) |
2229 | && JUMP_LABEL (insn) != NULL_RTX |
2230 | && ((any_uncondjump_p (insn) && onlyjump_p (insn)) |
2231 | || ANY_RETURN_P (PATTERN (insn))) |
2232 | && (next = NEXT_INSN (insn)) |
2233 | && BARRIER_P (next)); |
2234 | depth++) |
2235 | { |
2236 | rtx this_label_or_return = JUMP_LABEL (insn); |
2237 | |
2238 | /* If we have found a cycle, make the insn jump to itself. */ |
2239 | if (this_label_or_return == label) |
2240 | return label; |
2241 | |
2242 | /* Cannot follow returns and cannot look through tablejumps. */ |
2243 | if (ANY_RETURN_P (this_label_or_return)) |
2244 | return this_label_or_return; |
2245 | |
2246 | rtx_insn *this_label = as_a <rtx_insn *> (p: this_label_or_return); |
2247 | if (NEXT_INSN (insn: this_label) |
2248 | && JUMP_TABLE_DATA_P (NEXT_INSN (this_label))) |
2249 | break; |
2250 | |
2251 | if (!targetm.can_follow_jump (jump, insn)) |
2252 | break; |
2253 | if (!*crossing) |
2254 | *crossing = CROSSING_JUMP_P (jump); |
2255 | value = this_label; |
2256 | } |
2257 | if (depth == 10) |
2258 | return label; |
2259 | return value; |
2260 | } |
2261 | |
2262 | /* Try to find insns to place in delay slots. |
2263 | |
2264 | INSN is the jump needing SLOTS_TO_FILL delay slots. It tests CONDITION |
2265 | or is an unconditional branch if CONDITION is const_true_rtx. |
2266 | *PSLOTS_FILLED is updated with the number of slots that we have filled. |
2267 | |
2268 | THREAD is a flow-of-control, either the insns to be executed if the |
2269 | branch is true or if the branch is false, THREAD_IF_TRUE says which. |
2270 | |
2271 | OPPOSITE_THREAD is the thread in the opposite direction. It is used |
2272 | to see if any potential delay slot insns set things needed there. |
2273 | |
2274 | LIKELY is true if it is extremely likely that the branch will be |
2275 | taken and THREAD_IF_TRUE is set. This is used for the branch at the |
2276 | end of a loop back up to the top. |
2277 | |
2278 | OWN_THREAD is true if we are the only user of the thread, i.e. it is |
2279 | the target of the jump when we are the only jump going there. |
2280 | |
2281 | If OWN_THREAD is false, it must be the "true" thread of a jump. In that |
2282 | case, we can only take insns from the head of the thread for our delay |
2283 | slot. We then adjust the jump to point after the insns we have taken. */ |
2284 | |
2285 | static void |
2286 | fill_slots_from_thread (rtx_jump_insn *insn, rtx condition, |
2287 | rtx thread_or_return, rtx opposite_thread, bool likely, |
2288 | bool thread_if_true, bool own_thread, int slots_to_fill, |
2289 | int *pslots_filled, vec<rtx_insn *> *delay_list) |
2290 | { |
2291 | rtx new_thread; |
2292 | struct resources opposite_needed, set, needed; |
2293 | rtx_insn *trial; |
2294 | bool lose = false; |
2295 | bool must_annul = false; |
2296 | int flags; |
2297 | |
2298 | /* Validate our arguments. */ |
2299 | gcc_assert (condition != const_true_rtx || thread_if_true); |
2300 | gcc_assert (own_thread || thread_if_true); |
2301 | |
2302 | flags = get_jump_flags (insn, JUMP_LABEL (insn)); |
2303 | |
2304 | /* If our thread is the end of subroutine, we can't get any delay |
2305 | insns from that. */ |
2306 | if (thread_or_return == NULL_RTX || ANY_RETURN_P (thread_or_return)) |
2307 | return; |
2308 | |
2309 | rtx_insn *thread = as_a <rtx_insn *> (p: thread_or_return); |
2310 | |
2311 | /* If this is an unconditional branch, nothing is needed at the |
2312 | opposite thread. Otherwise, compute what is needed there. */ |
2313 | if (condition == const_true_rtx) |
2314 | CLEAR_RESOURCE (&opposite_needed); |
2315 | else |
2316 | mark_target_live_regs (get_insns (), opposite_thread, &opposite_needed); |
2317 | |
2318 | /* If the insn at THREAD can be split, do it here to avoid having to |
2319 | update THREAD and NEW_THREAD if it is done in the loop below. Also |
2320 | initialize NEW_THREAD. */ |
2321 | |
2322 | new_thread = thread = try_split (PATTERN (insn: thread), thread, 0); |
2323 | |
2324 | /* Scan insns at THREAD. We are looking for an insn that can be removed |
2325 | from THREAD (it neither sets nor references resources that were set |
2326 | ahead of it and it doesn't set anything needs by the insns ahead of |
2327 | it) and that either can be placed in an annulling insn or aren't |
2328 | needed at OPPOSITE_THREAD. */ |
2329 | |
2330 | CLEAR_RESOURCE (&needed); |
2331 | CLEAR_RESOURCE (&set); |
2332 | |
2333 | /* Handle the flags register specially, to be able to accept a |
2334 | candidate that clobbers it. See also fill_simple_delay_slots. */ |
2335 | bool filter_flags |
2336 | = (slots_to_fill == 1 |
2337 | && targetm.flags_regnum != INVALID_REGNUM |
2338 | && find_regno_note (insn, REG_DEAD, targetm.flags_regnum)); |
2339 | struct resources fset; |
2340 | struct resources flags_res; |
2341 | if (filter_flags) |
2342 | { |
2343 | CLEAR_RESOURCE (&fset); |
2344 | CLEAR_RESOURCE (&flags_res); |
2345 | SET_HARD_REG_BIT (set&: flags_res.regs, bit: targetm.flags_regnum); |
2346 | } |
2347 | |
2348 | /* If we do not own this thread, we must stop as soon as we find |
2349 | something that we can't put in a delay slot, since all we can do |
2350 | is branch into THREAD at a later point. Therefore, labels stop |
2351 | the search if this is not the `true' thread. */ |
2352 | |
2353 | for (trial = thread; |
2354 | ! stop_search_p (insn: trial, labels_p: ! thread_if_true) && (! lose || own_thread); |
2355 | trial = next_nonnote_insn (trial)) |
2356 | { |
2357 | rtx pat, old_trial; |
2358 | |
2359 | /* If we have passed a label, we no longer own this thread. */ |
2360 | if (LABEL_P (trial)) |
2361 | { |
2362 | own_thread = 0; |
2363 | continue; |
2364 | } |
2365 | |
2366 | pat = PATTERN (insn: trial); |
2367 | if (GET_CODE (pat) == USE || GET_CODE (pat) == CLOBBER) |
2368 | continue; |
2369 | |
2370 | if (GET_CODE (trial) == DEBUG_INSN) |
2371 | continue; |
2372 | |
2373 | /* If TRIAL conflicts with the insns ahead of it, we lose. */ |
2374 | if (! insn_references_resource_p (insn: trial, res: &set, include_delayed_effects: true) |
2375 | && ! insn_sets_resource_p (insn: trial, res: filter_flags ? &fset : &set, include_delayed_effects: true) |
2376 | && ! insn_sets_resource_p (insn: trial, res: &needed, include_delayed_effects: true) |
2377 | /* If we're handling sets to the flags register specially, we |
2378 | only allow an insn into a delay-slot, if it either: |
2379 | - doesn't set the flags register, |
2380 | - the "set" of the flags register isn't used (clobbered), |
2381 | - insns between the delay-slot insn and the trial-insn |
2382 | as accounted in "set", have not affected the flags register. */ |
2383 | && (! filter_flags |
2384 | || ! insn_sets_resource_p (insn: trial, res: &flags_res, include_delayed_effects: true) |
2385 | || find_regno_note (trial, REG_UNUSED, targetm.flags_regnum) |
2386 | || ! TEST_HARD_REG_BIT (set: set.regs, bit: targetm.flags_regnum)) |
2387 | && ! can_throw_internal (trial)) |
2388 | { |
2389 | rtx_insn *prior_insn; |
2390 | |
2391 | /* If TRIAL is redundant with some insn before INSN, we don't |
2392 | actually need to add it to the delay list; we can merely pretend |
2393 | we did. */ |
2394 | if ((prior_insn = redundant_insn (insn: trial, target: insn, delay_list: *delay_list))) |
2395 | { |
2396 | fix_reg_dead_note (start_insn: prior_insn, stop_insn: insn); |
2397 | if (own_thread) |
2398 | { |
2399 | update_block (insn: trial, where: thread); |
2400 | if (trial == thread) |
2401 | { |
2402 | thread = next_active_insn (thread); |
2403 | if (new_thread == trial) |
2404 | new_thread = thread; |
2405 | } |
2406 | |
2407 | delete_related_insns (trial); |
2408 | } |
2409 | else |
2410 | { |
2411 | update_reg_unused_notes (insn: prior_insn, other_insn: trial); |
2412 | new_thread = next_active_insn (trial); |
2413 | } |
2414 | |
2415 | continue; |
2416 | } |
2417 | |
2418 | /* There are two ways we can win: If TRIAL doesn't set anything |
2419 | needed at the opposite thread and can't trap, or if it can |
2420 | go into an annulled delay slot. But we want neither to copy |
2421 | nor to speculate frame-related insns. */ |
2422 | if (!must_annul |
2423 | && ((condition == const_true_rtx |
2424 | && (own_thread || !RTX_FRAME_RELATED_P (trial))) |
2425 | || (! insn_sets_resource_p (insn: trial, res: &opposite_needed, include_delayed_effects: true) |
2426 | && ! may_trap_or_fault_p (pat) |
2427 | && ! RTX_FRAME_RELATED_P (trial)))) |
2428 | { |
2429 | old_trial = trial; |
2430 | trial = try_split (pat, trial, 0); |
2431 | if (new_thread == old_trial) |
2432 | new_thread = trial; |
2433 | if (thread == old_trial) |
2434 | thread = trial; |
2435 | pat = PATTERN (insn: trial); |
2436 | if (eligible_for_delay (insn, *pslots_filled, trial, flags)) |
2437 | goto winner; |
2438 | } |
2439 | else if (!RTX_FRAME_RELATED_P (trial) |
2440 | && ((ANNUL_IFTRUE_SLOTS && ! thread_if_true) |
2441 | || (ANNUL_IFFALSE_SLOTS && thread_if_true))) |
2442 | { |
2443 | old_trial = trial; |
2444 | trial = try_split (pat, trial, 0); |
2445 | if (new_thread == old_trial) |
2446 | new_thread = trial; |
2447 | if (thread == old_trial) |
2448 | thread = trial; |
2449 | pat = PATTERN (insn: trial); |
2450 | if ((must_annul || delay_list->is_empty ()) && (thread_if_true |
2451 | ? check_annul_list_true_false (annul_true_p: false, delay_list: *delay_list) |
2452 | && eligible_for_annul_false (insn, *pslots_filled, trial, flags) |
2453 | : check_annul_list_true_false (annul_true_p: true, delay_list: *delay_list) |
2454 | && eligible_for_annul_true (insn, *pslots_filled, trial, flags))) |
2455 | { |
2456 | rtx_insn *temp; |
2457 | |
2458 | must_annul = true; |
2459 | winner: |
2460 | |
2461 | /* If we own this thread, delete the insn. If this is the |
2462 | destination of a branch, show that a basic block status |
2463 | may have been updated. In any case, mark the new |
2464 | starting point of this thread. */ |
2465 | if (own_thread) |
2466 | { |
2467 | rtx note; |
2468 | |
2469 | update_block (insn: trial, where: thread); |
2470 | if (trial == thread) |
2471 | { |
2472 | thread = next_active_insn (thread); |
2473 | if (new_thread == trial) |
2474 | new_thread = thread; |
2475 | } |
2476 | |
2477 | /* We are moving this insn, not deleting it. We must |
2478 | temporarily increment the use count on any referenced |
2479 | label lest it be deleted by delete_related_insns. */ |
2480 | for (note = REG_NOTES (trial); |
2481 | note != NULL_RTX; |
2482 | note = XEXP (note, 1)) |
2483 | if (REG_NOTE_KIND (note) == REG_LABEL_OPERAND |
2484 | || REG_NOTE_KIND (note) == REG_LABEL_TARGET) |
2485 | { |
2486 | /* REG_LABEL_OPERAND could be |
2487 | NOTE_INSN_DELETED_LABEL too. */ |
2488 | if (LABEL_P (XEXP (note, 0))) |
2489 | LABEL_NUSES (XEXP (note, 0))++; |
2490 | else |
2491 | gcc_assert (REG_NOTE_KIND (note) |
2492 | == REG_LABEL_OPERAND); |
2493 | } |
2494 | if (jump_to_label_p (trial)) |
2495 | LABEL_NUSES (JUMP_LABEL (trial))++; |
2496 | |
2497 | delete_related_insns (trial); |
2498 | |
2499 | for (note = REG_NOTES (trial); |
2500 | note != NULL_RTX; |
2501 | note = XEXP (note, 1)) |
2502 | if (REG_NOTE_KIND (note) == REG_LABEL_OPERAND |
2503 | || REG_NOTE_KIND (note) == REG_LABEL_TARGET) |
2504 | { |
2505 | /* REG_LABEL_OPERAND could be |
2506 | NOTE_INSN_DELETED_LABEL too. */ |
2507 | if (LABEL_P (XEXP (note, 0))) |
2508 | LABEL_NUSES (XEXP (note, 0))--; |
2509 | else |
2510 | gcc_assert (REG_NOTE_KIND (note) |
2511 | == REG_LABEL_OPERAND); |
2512 | } |
2513 | if (jump_to_label_p (trial)) |
2514 | LABEL_NUSES (JUMP_LABEL (trial))--; |
2515 | } |
2516 | else |
2517 | new_thread = next_active_insn (trial); |
2518 | |
2519 | temp = own_thread ? trial : copy_delay_slot_insn (trial); |
2520 | if (thread_if_true) |
2521 | INSN_FROM_TARGET_P (temp) = 1; |
2522 | |
2523 | add_to_delay_list (insn: temp, delay_list); |
2524 | |
2525 | if (slots_to_fill == ++(*pslots_filled)) |
2526 | { |
2527 | /* Even though we have filled all the slots, we |
2528 | may be branching to a location that has a |
2529 | redundant insn. Skip any if so. */ |
2530 | while (new_thread && ! own_thread |
2531 | && ! insn_sets_resource_p (insn: new_thread, res: &set, include_delayed_effects: true) |
2532 | && ! insn_sets_resource_p (insn: new_thread, res: &needed, |
2533 | include_delayed_effects: true) |
2534 | && ! insn_references_resource_p (insn: new_thread, |
2535 | res: &set, include_delayed_effects: true) |
2536 | && (prior_insn |
2537 | = redundant_insn (insn: new_thread, target: insn, |
2538 | delay_list: *delay_list))) |
2539 | { |
2540 | /* We know we do not own the thread, so no need |
2541 | to call update_block and delete_insn. */ |
2542 | fix_reg_dead_note (start_insn: prior_insn, stop_insn: insn); |
2543 | update_reg_unused_notes (insn: prior_insn, other_insn: new_thread); |
2544 | new_thread |
2545 | = next_active_insn (as_a<rtx_insn *> (p: new_thread)); |
2546 | } |
2547 | break; |
2548 | } |
2549 | |
2550 | continue; |
2551 | } |
2552 | } |
2553 | } |
2554 | |
2555 | /* This insn can't go into a delay slot. */ |
2556 | lose = true; |
2557 | mark_set_resources (trial, &set, 0, MARK_SRC_DEST_CALL); |
2558 | mark_referenced_resources (trial, &needed, true); |
2559 | if (filter_flags) |
2560 | { |
2561 | mark_set_resources (trial, &fset, 0, MARK_SRC_DEST_CALL); |
2562 | |
2563 | /* Groups of flags-register setters with users should not |
2564 | affect opportunities to move flags-register-setting insns |
2565 | (clobbers) into the delay-slot. */ |
2566 | CLEAR_HARD_REG_BIT (set&: needed.regs, bit: targetm.flags_regnum); |
2567 | CLEAR_HARD_REG_BIT (set&: fset.regs, bit: targetm.flags_regnum); |
2568 | } |
2569 | |
2570 | /* Ensure we don't put insns between the setting of cc and the comparison |
2571 | by moving a setting of cc into an earlier delay slot since these insns |
2572 | could clobber the condition code. */ |
2573 | set.cc = 1; |
2574 | |
2575 | /* If this insn is a register-register copy and the next insn has |
2576 | a use of our destination, change it to use our source. That way, |
2577 | it will become a candidate for our delay slot the next time |
2578 | through this loop. This case occurs commonly in loops that |
2579 | scan a list. |
2580 | |
2581 | We could check for more complex cases than those tested below, |
2582 | but it doesn't seem worth it. It might also be a good idea to try |
2583 | to swap the two insns. That might do better. |
2584 | |
2585 | We can't do this if the next insn modifies our destination, because |
2586 | that would make the replacement into the insn invalid. We also can't |
2587 | do this if it modifies our source, because it might be an earlyclobber |
2588 | operand. This latter test also prevents updating the contents of |
2589 | a PRE_INC. We also can't do this if there's overlap of source and |
2590 | destination. Overlap may happen for larger-than-register-size modes. */ |
2591 | |
2592 | if (NONJUMP_INSN_P (trial) && GET_CODE (pat) == SET |
2593 | && REG_P (SET_SRC (pat)) |
2594 | && REG_P (SET_DEST (pat)) |
2595 | && !reg_overlap_mentioned_p (SET_DEST (pat), SET_SRC (pat))) |
2596 | { |
2597 | rtx_insn *next = next_nonnote_insn (trial); |
2598 | |
2599 | if (next && NONJUMP_INSN_P (next) |
2600 | && GET_CODE (PATTERN (next)) != USE |
2601 | && ! reg_set_p (SET_DEST (pat), next) |
2602 | && ! reg_set_p (SET_SRC (pat), next) |
2603 | && reg_referenced_p (SET_DEST (pat), PATTERN (insn: next)) |
2604 | && ! modified_in_p (SET_DEST (pat), next)) |
2605 | validate_replace_rtx (SET_DEST (pat), SET_SRC (pat), next); |
2606 | } |
2607 | } |
2608 | |
2609 | /* If we stopped on a branch insn that has delay slots, see if we can |
2610 | steal some of the insns in those slots. */ |
2611 | if (trial && NONJUMP_INSN_P (trial) |
2612 | && GET_CODE (PATTERN (trial)) == SEQUENCE |
2613 | && JUMP_P (XVECEXP (PATTERN (trial), 0, 0))) |
2614 | { |
2615 | rtx_sequence *sequence = as_a <rtx_sequence *> (p: PATTERN (insn: trial)); |
2616 | /* If this is the `true' thread, we will want to follow the jump, |
2617 | so we can only do this if we have taken everything up to here. */ |
2618 | if (thread_if_true && trial == new_thread) |
2619 | { |
2620 | steal_delay_list_from_target (insn, condition, seq: sequence, |
2621 | delay_list, sets: &set, needed: &needed, |
2622 | other_needed: &opposite_needed, slots_to_fill, |
2623 | pslots_filled, pannul_p: &must_annul, |
2624 | pnew_thread: &new_thread); |
2625 | /* If we owned the thread and are told that it branched |
2626 | elsewhere, make sure we own the thread at the new location. */ |
2627 | if (own_thread && trial != new_thread) |
2628 | own_thread = own_thread_p (thread: new_thread, label: new_thread, allow_fallthrough: false); |
2629 | } |
2630 | else if (! thread_if_true) |
2631 | steal_delay_list_from_fallthrough (insn, condition, seq: sequence, |
2632 | delay_list, sets: &set, needed: &needed, |
2633 | other_needed: &opposite_needed, slots_to_fill, |
2634 | pslots_filled, pannul_p: &must_annul); |
2635 | } |
2636 | |
2637 | /* If we haven't found anything for this delay slot and it is very |
2638 | likely that the branch will be taken, see if the insn at our target |
2639 | increments or decrements a register with an increment that does not |
2640 | depend on the destination register. If so, try to place the opposite |
2641 | arithmetic insn after the jump insn and put the arithmetic insn in the |
2642 | delay slot. If we can't do this, return. */ |
2643 | if (delay_list->is_empty () && likely |
2644 | && new_thread && !ANY_RETURN_P (new_thread) |
2645 | && NONJUMP_INSN_P (new_thread) |
2646 | && !RTX_FRAME_RELATED_P (new_thread) |
2647 | && GET_CODE (PATTERN (new_thread)) != ASM_INPUT |
2648 | && asm_noperands (PATTERN (insn: new_thread)) < 0) |
2649 | { |
2650 | rtx dest; |
2651 | rtx src; |
2652 | |
2653 | /* We know "new_thread" is an insn due to NONJUMP_INSN_P (new_thread) |
2654 | above. */ |
2655 | trial = as_a <rtx_insn *> (p: new_thread); |
2656 | rtx pat = PATTERN (insn: trial); |
2657 | |
2658 | if (!NONJUMP_INSN_P (trial) |
2659 | || GET_CODE (pat) != SET |
2660 | || ! eligible_for_delay (insn, 0, trial, flags) |
2661 | || can_throw_internal (trial)) |
2662 | return; |
2663 | |
2664 | dest = SET_DEST (pat), src = SET_SRC (pat); |
2665 | if ((GET_CODE (src) == PLUS || GET_CODE (src) == MINUS) |
2666 | && rtx_equal_p (XEXP (src, 0), dest) |
2667 | && (!FLOAT_MODE_P (GET_MODE (src)) |
2668 | || flag_unsafe_math_optimizations) |
2669 | && ! reg_overlap_mentioned_p (dest, XEXP (src, 1)) |
2670 | && ! side_effects_p (pat)) |
2671 | { |
2672 | rtx other = XEXP (src, 1); |
2673 | rtx new_arith; |
2674 | rtx_insn *ninsn; |
2675 | |
2676 | /* If this is a constant adjustment, use the same code with |
2677 | the negated constant. Otherwise, reverse the sense of the |
2678 | arithmetic. */ |
2679 | if (CONST_INT_P (other)) |
2680 | new_arith = gen_rtx_fmt_ee (GET_CODE (src), GET_MODE (src), dest, |
2681 | negate_rtx (GET_MODE (src), other)); |
2682 | else |
2683 | new_arith = gen_rtx_fmt_ee (GET_CODE (src) == PLUS ? MINUS : PLUS, |
2684 | GET_MODE (src), dest, other); |
2685 | |
2686 | ninsn = emit_insn_after (gen_rtx_SET (dest, new_arith), insn); |
2687 | |
2688 | if (recog_memoized (insn: ninsn) < 0 |
2689 | || (extract_insn (ninsn), |
2690 | !constrain_operands (1, get_preferred_alternatives (ninsn)))) |
2691 | { |
2692 | delete_related_insns (ninsn); |
2693 | return; |
2694 | } |
2695 | |
2696 | if (own_thread) |
2697 | { |
2698 | update_block (insn: trial, where: thread); |
2699 | if (trial == thread) |
2700 | { |
2701 | thread = next_active_insn (thread); |
2702 | if (new_thread == trial) |
2703 | new_thread = thread; |
2704 | } |
2705 | delete_related_insns (trial); |
2706 | } |
2707 | else |
2708 | new_thread = next_active_insn (trial); |
2709 | |
2710 | ninsn = own_thread ? trial : copy_delay_slot_insn (trial); |
2711 | if (thread_if_true) |
2712 | INSN_FROM_TARGET_P (ninsn) = 1; |
2713 | |
2714 | add_to_delay_list (insn: ninsn, delay_list); |
2715 | (*pslots_filled)++; |
2716 | } |
2717 | } |
2718 | |
2719 | if (!delay_list->is_empty () && must_annul) |
2720 | INSN_ANNULLED_BRANCH_P (insn) = 1; |
2721 | |
2722 | /* If we are to branch into the middle of this thread, find an appropriate |
2723 | label or make a new one if none, and redirect INSN to it. If we hit the |
2724 | end of the function, use the end-of-function label. */ |
2725 | if (new_thread != thread) |
2726 | { |
2727 | rtx label; |
2728 | bool crossing = false; |
2729 | |
2730 | gcc_assert (thread_if_true); |
2731 | |
2732 | if (new_thread && simplejump_or_return_p (insn: new_thread) |
2733 | && redirect_with_delay_list_safe_p (jump: insn, |
2734 | JUMP_LABEL (new_thread), |
2735 | delay_list: *delay_list)) |
2736 | new_thread = follow_jumps (JUMP_LABEL (new_thread), jump: insn, |
2737 | crossing: &crossing); |
2738 | |
2739 | if (ANY_RETURN_P (new_thread)) |
2740 | label = find_end_label (kind: new_thread); |
2741 | else if (LABEL_P (new_thread)) |
2742 | label = new_thread; |
2743 | else |
2744 | label = get_label_before (insn: as_a <rtx_insn *> (p: new_thread), |
2745 | JUMP_LABEL (insn)); |
2746 | |
2747 | if (label) |
2748 | { |
2749 | reorg_redirect_jump (jump: insn, nlabel: label); |
2750 | if (crossing) |
2751 | CROSSING_JUMP_P (insn) = 1; |
2752 | } |
2753 | } |
2754 | } |
2755 | |
2756 | /* Make another attempt to find insns to place in delay slots. |
2757 | |
2758 | We previously looked for insns located in front of the delay insn |
2759 | and, for non-jump delay insns, located behind the delay insn. |
2760 | |
2761 | Here only try to schedule jump insns and try to move insns from either |
2762 | the target or the following insns into the delay slot. If annulling is |
2763 | supported, we will be likely to do this. Otherwise, we can do this only |
2764 | if safe. */ |
2765 | |
2766 | static void |
2767 | fill_eager_delay_slots (void) |
2768 | { |
2769 | rtx_insn *insn; |
2770 | int i; |
2771 | int num_unfilled_slots = unfilled_slots_next - unfilled_slots_base; |
2772 | |
2773 | for (i = 0; i < num_unfilled_slots; i++) |
2774 | { |
2775 | rtx condition; |
2776 | rtx target_label, insn_at_target; |
2777 | rtx_insn *fallthrough_insn; |
2778 | auto_vec<rtx_insn *, 5> delay_list; |
2779 | rtx_jump_insn *jump_insn; |
2780 | bool own_target; |
2781 | bool own_fallthrough; |
2782 | int prediction, slots_to_fill, slots_filled; |
2783 | |
2784 | insn = unfilled_slots_base[i]; |
2785 | if (insn == 0 |
2786 | || insn->deleted () |
2787 | || ! (jump_insn = dyn_cast <rtx_jump_insn *> (p: insn)) |
2788 | || ! (condjump_p (jump_insn) || condjump_in_parallel_p (jump_insn))) |
2789 | continue; |
2790 | |
2791 | slots_to_fill = num_delay_slots (jump_insn); |
2792 | /* Some machine description have defined instructions to have |
2793 | delay slots only in certain circumstances which may depend on |
2794 | nearby insns (which change due to reorg's actions). |
2795 | |
2796 | For example, the PA port normally has delay slots for unconditional |
2797 | jumps. |
2798 | |
2799 | However, the PA port claims such jumps do not have a delay slot |
2800 | if they are immediate successors of certain CALL_INSNs. This |
2801 | allows the port to favor filling the delay slot of the call with |
2802 | the unconditional jump. */ |
2803 | if (slots_to_fill == 0) |
2804 | continue; |
2805 | |
2806 | slots_filled = 0; |
2807 | target_label = JUMP_LABEL (jump_insn); |
2808 | condition = get_branch_condition (insn: jump_insn, target: target_label); |
2809 | |
2810 | if (condition == 0) |
2811 | continue; |
2812 | |
2813 | /* Get the next active fallthrough and target insns and see if we own |
2814 | them. Then see whether the branch is likely true. We don't need |
2815 | to do a lot of this for unconditional branches. */ |
2816 | |
2817 | insn_at_target = first_active_target_insn (insn: target_label); |
2818 | own_target = own_thread_p (thread: target_label, label: target_label, allow_fallthrough: false); |
2819 | |
2820 | if (condition == const_true_rtx) |
2821 | { |
2822 | own_fallthrough = false; |
2823 | fallthrough_insn = 0; |
2824 | prediction = 2; |
2825 | } |
2826 | else |
2827 | { |
2828 | fallthrough_insn = next_active_insn (jump_insn); |
2829 | own_fallthrough = own_thread_p (thread: NEXT_INSN (insn: jump_insn), |
2830 | NULL_RTX, allow_fallthrough: true); |
2831 | prediction = mostly_true_jump (jump_insn); |
2832 | } |
2833 | |
2834 | /* If this insn is expected to branch, first try to get insns from our |
2835 | target, then our fallthrough insns. If it is not expected to branch, |
2836 | try the other order. */ |
2837 | |
2838 | if (prediction > 0) |
2839 | { |
2840 | fill_slots_from_thread (insn: jump_insn, condition, thread_or_return: insn_at_target, |
2841 | opposite_thread: fallthrough_insn, likely: prediction == 2, thread_if_true: true, |
2842 | own_thread: own_target, slots_to_fill, |
2843 | pslots_filled: &slots_filled, delay_list: &delay_list); |
2844 | |
2845 | if (delay_list.is_empty () && own_fallthrough) |
2846 | { |
2847 | /* Even though we didn't find anything for delay slots, |
2848 | we might have found a redundant insn which we deleted |
2849 | from the thread that was filled. So we have to recompute |
2850 | the next insn at the target. */ |
2851 | target_label = JUMP_LABEL (jump_insn); |
2852 | insn_at_target = first_active_target_insn (insn: target_label); |
2853 | |
2854 | fill_slots_from_thread (insn: jump_insn, condition, thread_or_return: fallthrough_insn, |
2855 | opposite_thread: insn_at_target, likely: false, thread_if_true: false, |
2856 | own_thread: own_fallthrough, slots_to_fill, |
2857 | pslots_filled: &slots_filled, delay_list: &delay_list); |
2858 | } |
2859 | } |
2860 | else |
2861 | { |
2862 | if (own_fallthrough) |
2863 | fill_slots_from_thread (insn: jump_insn, condition, thread_or_return: fallthrough_insn, |
2864 | opposite_thread: insn_at_target, likely: false, thread_if_true: false, |
2865 | own_thread: own_fallthrough, slots_to_fill, |
2866 | pslots_filled: &slots_filled, delay_list: &delay_list); |
2867 | |
2868 | if (delay_list.is_empty ()) |
2869 | fill_slots_from_thread (insn: jump_insn, condition, thread_or_return: insn_at_target, |
2870 | opposite_thread: next_active_insn (insn), likely: false, thread_if_true: true, |
2871 | own_thread: own_target, slots_to_fill, |
2872 | pslots_filled: &slots_filled, delay_list: &delay_list); |
2873 | } |
2874 | |
2875 | if (!delay_list.is_empty ()) |
2876 | unfilled_slots_base[i] |
2877 | = emit_delay_sequence (insn: jump_insn, list: delay_list, length: slots_filled); |
2878 | |
2879 | if (slots_to_fill == slots_filled) |
2880 | unfilled_slots_base[i] = 0; |
2881 | |
2882 | note_delay_statistics (slots_filled, index: 1); |
2883 | } |
2884 | } |
2885 | |
2886 | static void delete_computation (rtx_insn *insn); |
2887 | |
2888 | /* Recursively delete prior insns that compute the value (used only by INSN |
2889 | which the caller is deleting) stored in the register mentioned by NOTE |
2890 | which is a REG_DEAD note associated with INSN. */ |
2891 | |
2892 | static void |
2893 | delete_prior_computation (rtx note, rtx_insn *insn) |
2894 | { |
2895 | rtx_insn *our_prev; |
2896 | rtx reg = XEXP (note, 0); |
2897 | |
2898 | for (our_prev = prev_nonnote_insn (insn); |
2899 | our_prev && (NONJUMP_INSN_P (our_prev) |
2900 | || CALL_P (our_prev)); |
2901 | our_prev = prev_nonnote_insn (our_prev)) |
2902 | { |
2903 | rtx pat = PATTERN (insn: our_prev); |
2904 | |
2905 | /* If we reach a CALL which is not calling a const function |
2906 | or the callee pops the arguments, then give up. */ |
2907 | if (CALL_P (our_prev) |
2908 | && (! RTL_CONST_CALL_P (our_prev) |
2909 | || GET_CODE (pat) != SET || GET_CODE (SET_SRC (pat)) != CALL)) |
2910 | break; |
2911 | |
2912 | /* If we reach a SEQUENCE, it is too complex to try to |
2913 | do anything with it, so give up. We can be run during |
2914 | and after reorg, so SEQUENCE rtl can legitimately show |
2915 | up here. */ |
2916 | if (GET_CODE (pat) == SEQUENCE) |
2917 | break; |
2918 | |
2919 | if (GET_CODE (pat) == USE |
2920 | && NONJUMP_INSN_P (XEXP (pat, 0))) |
2921 | /* reorg creates USEs that look like this. We leave them |
2922 | alone because reorg needs them for its own purposes. */ |
2923 | break; |
2924 | |
2925 | if (reg_set_p (reg, pat)) |
2926 | { |
2927 | if (side_effects_p (pat) && !CALL_P (our_prev)) |
2928 | break; |
2929 | |
2930 | if (GET_CODE (pat) == PARALLEL) |
2931 | { |
2932 | /* If we find a SET of something else, we can't |
2933 | delete the insn. */ |
2934 | |
2935 | int i; |
2936 | |
2937 | for (i = 0; i < XVECLEN (pat, 0); i++) |
2938 | { |
2939 | rtx part = XVECEXP (pat, 0, i); |
2940 | |
2941 | if (GET_CODE (part) == SET |
2942 | && SET_DEST (part) != reg) |
2943 | break; |
2944 | } |
2945 | |
2946 | if (i == XVECLEN (pat, 0)) |
2947 | delete_computation (insn: our_prev); |
2948 | } |
2949 | else if (GET_CODE (pat) == SET |
2950 | && REG_P (SET_DEST (pat))) |
2951 | { |
2952 | int dest_regno = REGNO (SET_DEST (pat)); |
2953 | int dest_endregno = END_REGNO (SET_DEST (pat)); |
2954 | int regno = REGNO (reg); |
2955 | int endregno = END_REGNO (x: reg); |
2956 | |
2957 | if (dest_regno >= regno |
2958 | && dest_endregno <= endregno) |
2959 | delete_computation (insn: our_prev); |
2960 | |
2961 | /* We may have a multi-word hard register and some, but not |
2962 | all, of the words of the register are needed in subsequent |
2963 | insns. Write REG_UNUSED notes for those parts that were not |
2964 | needed. */ |
2965 | else if (dest_regno <= regno |
2966 | && dest_endregno >= endregno) |
2967 | { |
2968 | int i; |
2969 | |
2970 | add_reg_note (our_prev, REG_UNUSED, reg); |
2971 | |
2972 | for (i = dest_regno; i < dest_endregno; i++) |
2973 | if (! find_regno_note (our_prev, REG_UNUSED, i)) |
2974 | break; |
2975 | |
2976 | if (i == dest_endregno) |
2977 | delete_computation (insn: our_prev); |
2978 | } |
2979 | } |
2980 | |
2981 | break; |
2982 | } |
2983 | |
2984 | /* If PAT references the register that dies here, it is an |
2985 | additional use. Hence any prior SET isn't dead. However, this |
2986 | insn becomes the new place for the REG_DEAD note. */ |
2987 | if (reg_overlap_mentioned_p (reg, pat)) |
2988 | { |
2989 | XEXP (note, 1) = REG_NOTES (our_prev); |
2990 | REG_NOTES (our_prev) = note; |
2991 | break; |
2992 | } |
2993 | } |
2994 | } |
2995 | |
2996 | /* Delete INSN and recursively delete insns that compute values used only |
2997 | by INSN. This uses the REG_DEAD notes computed during flow analysis. |
2998 | |
2999 | Look at all our REG_DEAD notes. If a previous insn does nothing other |
3000 | than set a register that dies in this insn, we can delete that insn |
3001 | as well. */ |
3002 | |
3003 | static void |
3004 | delete_computation (rtx_insn *insn) |
3005 | { |
3006 | rtx note, next; |
3007 | |
3008 | for (note = REG_NOTES (insn); note; note = next) |
3009 | { |
3010 | next = XEXP (note, 1); |
3011 | |
3012 | if (REG_NOTE_KIND (note) != REG_DEAD |
3013 | /* Verify that the REG_NOTE is legitimate. */ |
3014 | || !REG_P (XEXP (note, 0))) |
3015 | continue; |
3016 | |
3017 | delete_prior_computation (note, insn); |
3018 | } |
3019 | |
3020 | delete_related_insns (insn); |
3021 | } |
3022 | |
3023 | /* If all INSN does is set the pc, delete it, |
3024 | and delete the insn that set the condition codes for it |
3025 | if that's what the previous thing was. */ |
3026 | |
3027 | static void |
3028 | delete_jump (rtx_insn *insn) |
3029 | { |
3030 | rtx set = single_set (insn); |
3031 | |
3032 | if (set && GET_CODE (SET_DEST (set)) == PC) |
3033 | delete_computation (insn); |
3034 | } |
3035 | |
3036 | static rtx_insn * |
3037 | label_before_next_insn (rtx_insn *x, rtx scan_limit) |
3038 | { |
3039 | rtx_insn *insn = next_active_insn (x); |
3040 | while (insn) |
3041 | { |
3042 | insn = PREV_INSN (insn); |
3043 | if (insn == scan_limit || insn == NULL_RTX) |
3044 | return NULL; |
3045 | if (LABEL_P (insn)) |
3046 | break; |
3047 | } |
3048 | return insn; |
3049 | } |
3050 | |
3051 | /* Return TRUE if there is a NOTE_INSN_SWITCH_TEXT_SECTIONS note in between |
3052 | BEG and END. */ |
3053 | |
3054 | static bool |
3055 | switch_text_sections_between_p (const rtx_insn *beg, const rtx_insn *end) |
3056 | { |
3057 | const rtx_insn *p; |
3058 | for (p = beg; p != end; p = NEXT_INSN (insn: p)) |
3059 | if (NOTE_P (p) && NOTE_KIND (p) == NOTE_INSN_SWITCH_TEXT_SECTIONS) |
3060 | return true; |
3061 | return false; |
3062 | } |
3063 | |
3064 | |
3065 | /* Once we have tried two ways to fill a delay slot, make a pass over the |
3066 | code to try to improve the results and to do such things as more jump |
3067 | threading. */ |
3068 | |
3069 | static void |
3070 | relax_delay_slots (rtx_insn *first) |
3071 | { |
3072 | rtx_insn *insn, *next; |
3073 | rtx_sequence *pat; |
3074 | rtx_insn *delay_insn; |
3075 | rtx target_label; |
3076 | |
3077 | /* Look at every JUMP_INSN and see if we can improve it. */ |
3078 | for (insn = first; insn; insn = next) |
3079 | { |
3080 | rtx_insn *other, *prior_insn; |
3081 | bool crossing; |
3082 | |
3083 | next = next_active_insn (insn); |
3084 | |
3085 | /* If this is a jump insn, see if it now jumps to a jump, jumps to |
3086 | the next insn, or jumps to a label that is not the last of a |
3087 | group of consecutive labels. */ |
3088 | if (is_a <rtx_jump_insn *> (p: insn) |
3089 | && (condjump_p (insn) || condjump_in_parallel_p (insn)) |
3090 | && !ANY_RETURN_P (target_label = JUMP_LABEL (insn))) |
3091 | { |
3092 | rtx_jump_insn *jump_insn = as_a <rtx_jump_insn *> (p: insn); |
3093 | target_label |
3094 | = skip_consecutive_labels (label_or_return: follow_jumps (label: target_label, jump: jump_insn, |
3095 | crossing: &crossing)); |
3096 | if (ANY_RETURN_P (target_label)) |
3097 | target_label = find_end_label (kind: target_label); |
3098 | |
3099 | if (target_label |
3100 | && next_active_insn (as_a<rtx_insn *> (p: target_label)) == next |
3101 | && ! condjump_in_parallel_p (jump_insn) |
3102 | && ! (next && switch_text_sections_between_p (beg: jump_insn, end: next))) |
3103 | { |
3104 | rtx_insn *direct_label = as_a<rtx_insn *> (JUMP_LABEL (insn)); |
3105 | rtx_insn *prev = prev_nonnote_insn (direct_label); |
3106 | |
3107 | /* If the insn jumps over a BARRIER and is the only way to reach |
3108 | its target, then we need to delete the BARRIER before the jump |
3109 | because, otherwise, the target may end up being considered as |
3110 | unreachable and thus also deleted. */ |
3111 | if (BARRIER_P (prev) && LABEL_NUSES (direct_label) == 1) |
3112 | { |
3113 | delete_related_insns (prev); |
3114 | |
3115 | /* We have just removed a BARRIER, which means that the block |
3116 | number of the next insns has effectively been changed (see |
3117 | find_basic_block in resource.cc), so clear it. */ |
3118 | clear_hashed_info_until_next_barrier (direct_label); |
3119 | } |
3120 | |
3121 | delete_jump (insn: jump_insn); |
3122 | continue; |
3123 | } |
3124 | |
3125 | if (target_label && target_label != JUMP_LABEL (jump_insn)) |
3126 | { |
3127 | reorg_redirect_jump (jump: jump_insn, nlabel: target_label); |
3128 | if (crossing) |
3129 | CROSSING_JUMP_P (jump_insn) = 1; |
3130 | } |
3131 | |
3132 | /* See if this jump conditionally branches around an unconditional |
3133 | jump. If so, invert this jump and point it to the target of the |
3134 | second jump. Check if it's possible on the target. */ |
3135 | if (next && simplejump_or_return_p (insn: next) |
3136 | && any_condjump_p (jump_insn) |
3137 | && target_label |
3138 | && (next_active_insn (as_a<rtx_insn *> (p: target_label)) |
3139 | == next_active_insn (next)) |
3140 | && no_labels_between_p (jump_insn, next) |
3141 | && targetm.can_follow_jump (jump_insn, next)) |
3142 | { |
3143 | rtx label = JUMP_LABEL (next); |
3144 | |
3145 | /* Be careful how we do this to avoid deleting code or |
3146 | labels that are momentarily dead. See similar optimization |
3147 | in jump.cc. |
3148 | |
3149 | We also need to ensure we properly handle the case when |
3150 | invert_jump fails. */ |
3151 | |
3152 | ++LABEL_NUSES (target_label); |
3153 | if (!ANY_RETURN_P (label)) |
3154 | ++LABEL_NUSES (label); |
3155 | |
3156 | if (invert_jump (jump_insn, label, 1)) |
3157 | { |
3158 | rtx_insn *from = delete_related_insns (next); |
3159 | |
3160 | /* We have just removed a BARRIER, which means that the block |
3161 | number of the next insns has effectively been changed (see |
3162 | find_basic_block in resource.cc), so clear it. */ |
3163 | if (from) |
3164 | clear_hashed_info_until_next_barrier (from); |
3165 | |
3166 | next = jump_insn; |
3167 | } |
3168 | |
3169 | if (!ANY_RETURN_P (label)) |
3170 | --LABEL_NUSES (label); |
3171 | |
3172 | if (--LABEL_NUSES (target_label) == 0) |
3173 | delete_related_insns (target_label); |
3174 | |
3175 | continue; |
3176 | } |
3177 | } |
3178 | |
3179 | /* If this is an unconditional jump and the previous insn is a |
3180 | conditional jump, try reversing the condition of the previous |
3181 | insn and swapping our targets. The next pass might be able to |
3182 | fill the slots. |
3183 | |
3184 | Don't do this if we expect the conditional branch to be true, because |
3185 | we would then be making the more common case longer. */ |
3186 | |
3187 | if (simplejump_or_return_p (insn) |
3188 | && (other = prev_active_insn (insn)) != 0 |
3189 | && any_condjump_p (other) |
3190 | && no_labels_between_p (other, insn) |
3191 | && mostly_true_jump (jump_insn: other) < 0) |
3192 | { |
3193 | rtx other_target = JUMP_LABEL (other); |
3194 | target_label = JUMP_LABEL (insn); |
3195 | |
3196 | if (invert_jump (as_a <rtx_jump_insn *> (p: other), target_label, 0)) |
3197 | reorg_redirect_jump (jump: as_a <rtx_jump_insn *> (p: insn), nlabel: other_target); |
3198 | } |
3199 | |
3200 | /* Now look only at cases where we have a filled delay slot. */ |
3201 | if (!NONJUMP_INSN_P (insn) || GET_CODE (PATTERN (insn)) != SEQUENCE) |
3202 | continue; |
3203 | |
3204 | pat = as_a <rtx_sequence *> (p: PATTERN (insn)); |
3205 | delay_insn = pat->insn (index: 0); |
3206 | |
3207 | /* See if the first insn in the delay slot is redundant with some |
3208 | previous insn. Remove it from the delay slot if so; then set up |
3209 | to reprocess this insn. */ |
3210 | if ((prior_insn = redundant_insn (insn: pat->insn (index: 1), target: delay_insn, delay_list: vNULL))) |
3211 | { |
3212 | fix_reg_dead_note (start_insn: prior_insn, stop_insn: insn); |
3213 | update_block (insn: pat->insn (index: 1), where: insn); |
3214 | delete_from_delay_slot (insn: pat->insn (index: 1)); |
3215 | next = prev_active_insn (next); |
3216 | continue; |
3217 | } |
3218 | |
3219 | /* See if we have a RETURN insn with a filled delay slot followed |
3220 | by a RETURN insn with an unfilled a delay slot. If so, we can delete |
3221 | the first RETURN (but not its delay insn). This gives the same |
3222 | effect in fewer instructions. |
3223 | |
3224 | Only do so if optimizing for size since this results in slower, but |
3225 | smaller code. */ |
3226 | if (optimize_function_for_size_p (cfun) |
3227 | && ANY_RETURN_P (PATTERN (delay_insn)) |
3228 | && next |
3229 | && JUMP_P (next) |
3230 | && PATTERN (insn: next) == PATTERN (insn: delay_insn)) |
3231 | { |
3232 | rtx_insn *after; |
3233 | int i; |
3234 | |
3235 | /* Delete the RETURN and just execute the delay list insns. |
3236 | |
3237 | We do this by deleting the INSN containing the SEQUENCE, then |
3238 | re-emitting the insns separately, and then deleting the RETURN. |
3239 | This allows the count of the jump target to be properly |
3240 | decremented. |
3241 | |
3242 | Note that we need to change the INSN_UID of the re-emitted insns |
3243 | since it is used to hash the insns for mark_target_live_regs and |
3244 | the re-emitted insns will no longer be wrapped up in a SEQUENCE. |
3245 | |
3246 | Clear the from target bit, since these insns are no longer |
3247 | in delay slots. */ |
3248 | for (i = 0; i < XVECLEN (pat, 0); i++) |
3249 | INSN_FROM_TARGET_P (XVECEXP (pat, 0, i)) = 0; |
3250 | |
3251 | rtx_insn *prev = PREV_INSN (insn); |
3252 | delete_related_insns (insn); |
3253 | gcc_assert (GET_CODE (pat) == SEQUENCE); |
3254 | add_insn_after (delay_insn, prev, NULL); |
3255 | after = delay_insn; |
3256 | for (i = 1; i < pat->len (); i++) |
3257 | after = emit_copy_of_insn_after (pat->insn (index: i), after); |
3258 | delete_scheduled_jump (insn: delay_insn); |
3259 | continue; |
3260 | } |
3261 | |
3262 | /* Now look only at the cases where we have a filled JUMP_INSN. */ |
3263 | rtx_jump_insn *delay_jump_insn = |
3264 | dyn_cast <rtx_jump_insn *> (p: delay_insn); |
3265 | if (! delay_jump_insn || !(condjump_p (delay_jump_insn) |
3266 | || condjump_in_parallel_p (delay_jump_insn))) |
3267 | continue; |
3268 | |
3269 | target_label = JUMP_LABEL (delay_jump_insn); |
3270 | if (target_label && ANY_RETURN_P (target_label)) |
3271 | continue; |
3272 | |
3273 | /* If this jump goes to another unconditional jump, thread it, but |
3274 | don't convert a jump into a RETURN here. */ |
3275 | rtx trial = skip_consecutive_labels (label_or_return: follow_jumps (label: target_label, |
3276 | jump: delay_jump_insn, |
3277 | crossing: &crossing)); |
3278 | if (ANY_RETURN_P (trial)) |
3279 | trial = find_end_label (kind: trial); |
3280 | |
3281 | if (trial && trial != target_label |
3282 | && redirect_with_delay_slots_safe_p (jump: delay_jump_insn, newlabel: trial, seq: insn)) |
3283 | { |
3284 | reorg_redirect_jump (jump: delay_jump_insn, nlabel: trial); |
3285 | target_label = trial; |
3286 | if (crossing) |
3287 | CROSSING_JUMP_P (delay_jump_insn) = 1; |
3288 | } |
3289 | |
3290 | /* If the first insn at TARGET_LABEL is redundant with a previous |
3291 | insn, redirect the jump to the following insn and process again. |
3292 | We use next_real_nondebug_insn instead of next_active_insn so we |
3293 | don't skip USE-markers, or we'll end up with incorrect |
3294 | liveness info. */ |
3295 | trial = next_real_nondebug_insn (target_label); |
3296 | if (trial && GET_CODE (PATTERN (trial)) != SEQUENCE |
3297 | && redundant_insn (insn: trial, target: insn, delay_list: vNULL) |
3298 | && ! can_throw_internal (trial)) |
3299 | { |
3300 | /* Figure out where to emit the special USE insn so we don't |
3301 | later incorrectly compute register live/death info. */ |
3302 | rtx_insn *tmp = next_active_insn (as_a<rtx_insn *> (p: trial)); |
3303 | if (tmp == 0) |
3304 | tmp = find_end_label (kind: simple_return_rtx); |
3305 | |
3306 | if (tmp) |
3307 | { |
3308 | /* Insert the special USE insn and update dataflow info. |
3309 | We know "trial" is an insn here as it is the output of |
3310 | next_real_nondebug_insn () above. */ |
3311 | update_block (insn: as_a <rtx_insn *> (p: trial), where: tmp); |
3312 | |
3313 | /* Now emit a label before the special USE insn, and |
3314 | redirect our jump to the new label. */ |
3315 | target_label = get_label_before (insn: PREV_INSN (insn: tmp), sibling: target_label); |
3316 | reorg_redirect_jump (jump: delay_jump_insn, nlabel: target_label); |
3317 | next = insn; |
3318 | continue; |
3319 | } |
3320 | } |
3321 | |
3322 | /* Similarly, if it is an unconditional jump with one insn in its |
3323 | delay list and that insn is redundant, thread the jump. */ |
3324 | rtx_sequence *trial_seq = |
3325 | trial ? dyn_cast <rtx_sequence *> (p: PATTERN (insn: trial)) : NULL; |
3326 | if (trial_seq |
3327 | && trial_seq->len () == 2 |
3328 | && JUMP_P (trial_seq->insn (0)) |
3329 | && simplejump_or_return_p (insn: trial_seq->insn (index: 0)) |
3330 | && redundant_insn (insn: trial_seq->insn (index: 1), target: insn, delay_list: vNULL)) |
3331 | { |
3332 | rtx temp_label = JUMP_LABEL (trial_seq->insn (0)); |
3333 | if (ANY_RETURN_P (temp_label)) |
3334 | temp_label = find_end_label (kind: temp_label); |
3335 | |
3336 | if (temp_label |
3337 | && redirect_with_delay_slots_safe_p (jump: delay_jump_insn, |
3338 | newlabel: temp_label, seq: insn)) |
3339 | { |
3340 | update_block (insn: trial_seq->insn (index: 1), where: insn); |
3341 | reorg_redirect_jump (jump: delay_jump_insn, nlabel: temp_label); |
3342 | next = insn; |
3343 | continue; |
3344 | } |
3345 | } |
3346 | |
3347 | /* See if we have a simple (conditional) jump that is useless. */ |
3348 | if (!CROSSING_JUMP_P (delay_jump_insn) |
3349 | && !INSN_ANNULLED_BRANCH_P (delay_jump_insn) |
3350 | && !condjump_in_parallel_p (delay_jump_insn) |
3351 | && prev_active_insn (as_a<rtx_insn *> (p: target_label)) == insn |
3352 | && !BARRIER_P (prev_nonnote_insn (as_a<rtx_insn *> (target_label)))) |
3353 | { |
3354 | rtx_insn *after; |
3355 | int i; |
3356 | |
3357 | /* All this insn does is execute its delay list and jump to the |
3358 | following insn. So delete the jump and just execute the delay |
3359 | list insns. |
3360 | |
3361 | We do this by deleting the INSN containing the SEQUENCE, then |
3362 | re-emitting the insns separately, and then deleting the jump. |
3363 | This allows the count of the jump target to be properly |
3364 | decremented. |
3365 | |
3366 | Note that we need to change the INSN_UID of the re-emitted insns |
3367 | since it is used to hash the insns for mark_target_live_regs and |
3368 | the re-emitted insns will no longer be wrapped up in a SEQUENCE. |
3369 | |
3370 | Clear the from target bit, since these insns are no longer |
3371 | in delay slots. */ |
3372 | for (i = 0; i < XVECLEN (pat, 0); i++) |
3373 | INSN_FROM_TARGET_P (XVECEXP (pat, 0, i)) = 0; |
3374 | |
3375 | rtx_insn *prev = PREV_INSN (insn); |
3376 | delete_related_insns (insn); |
3377 | gcc_assert (GET_CODE (pat) == SEQUENCE); |
3378 | add_insn_after (delay_jump_insn, prev, NULL); |
3379 | after = delay_jump_insn; |
3380 | for (i = 1; i < pat->len (); i++) |
3381 | after = emit_copy_of_insn_after (pat->insn (index: i), after); |
3382 | delete_scheduled_jump (insn: delay_jump_insn); |
3383 | continue; |
3384 | } |
3385 | |
3386 | /* See if this is an unconditional jump around a single insn which is |
3387 | identical to the one in its delay slot. In this case, we can just |
3388 | delete the branch and the insn in its delay slot. */ |
3389 | if (next && NONJUMP_INSN_P (next) |
3390 | && label_before_next_insn (x: next, scan_limit: insn) == target_label |
3391 | && simplejump_p (insn) |
3392 | && XVECLEN (pat, 0) == 2 |
3393 | && rtx_equal_p (PATTERN (insn: next), PATTERN (insn: pat->insn (index: 1)))) |
3394 | { |
3395 | delete_related_insns (insn); |
3396 | continue; |
3397 | } |
3398 | |
3399 | /* See if this jump (with its delay slots) conditionally branches |
3400 | around an unconditional jump (without delay slots). If so, invert |
3401 | this jump and point it to the target of the second jump. We cannot |
3402 | do this for annulled jumps, though. Again, don't convert a jump to |
3403 | a RETURN here. */ |
3404 | if (! INSN_ANNULLED_BRANCH_P (delay_jump_insn) |
3405 | && any_condjump_p (delay_jump_insn) |
3406 | && next && simplejump_or_return_p (insn: next) |
3407 | && (next_active_insn (as_a<rtx_insn *> (p: target_label)) |
3408 | == next_active_insn (next)) |
3409 | && no_labels_between_p (insn, next)) |
3410 | { |
3411 | rtx label = JUMP_LABEL (next); |
3412 | rtx old_label = JUMP_LABEL (delay_jump_insn); |
3413 | |
3414 | if (ANY_RETURN_P (label)) |
3415 | label = find_end_label (kind: label); |
3416 | |
3417 | /* find_end_label can generate a new label. Check this first. */ |
3418 | if (label |
3419 | && no_labels_between_p (insn, next) |
3420 | && redirect_with_delay_slots_safe_p (jump: delay_jump_insn, |
3421 | newlabel: label, seq: insn)) |
3422 | { |
3423 | /* Be careful how we do this to avoid deleting code or labels |
3424 | that are momentarily dead. See similar optimization in |
3425 | jump.cc */ |
3426 | if (old_label) |
3427 | ++LABEL_NUSES (old_label); |
3428 | |
3429 | if (invert_jump (delay_jump_insn, label, 1)) |
3430 | { |
3431 | /* Must update the INSN_FROM_TARGET_P bits now that |
3432 | the branch is reversed, so that mark_target_live_regs |
3433 | will handle the delay slot insn correctly. */ |
3434 | for (int i = 1; i < XVECLEN (PATTERN (insn), 0); i++) |
3435 | { |
3436 | rtx slot = XVECEXP (PATTERN (insn), 0, i); |
3437 | INSN_FROM_TARGET_P (slot) = ! INSN_FROM_TARGET_P (slot); |
3438 | } |
3439 | |
3440 | /* We have just removed a BARRIER, which means that the block |
3441 | number of the next insns has effectively been changed (see |
3442 | find_basic_block in resource.cc), so clear it. */ |
3443 | rtx_insn *from = delete_related_insns (next); |
3444 | if (from) |
3445 | clear_hashed_info_until_next_barrier (from); |
3446 | |
3447 | next = insn; |
3448 | } |
3449 | |
3450 | if (old_label && --LABEL_NUSES (old_label) == 0) |
3451 | delete_related_insns (old_label); |
3452 | continue; |
3453 | } |
3454 | } |
3455 | |
3456 | /* If we own the thread opposite the way this insn branches, see if we |
3457 | can merge its delay slots with following insns. */ |
3458 | if (INSN_FROM_TARGET_P (pat->insn (1)) |
3459 | && own_thread_p (thread: NEXT_INSN (insn), label: 0, allow_fallthrough: true)) |
3460 | try_merge_delay_insns (insn, thread: next); |
3461 | else if (! INSN_FROM_TARGET_P (pat->insn (1)) |
3462 | && own_thread_p (thread: target_label, label: target_label, allow_fallthrough: false)) |
3463 | try_merge_delay_insns (insn, |
3464 | thread: next_active_insn (as_a<rtx_insn *> (p: target_label))); |
3465 | |
3466 | /* If we get here, we haven't deleted INSN. But we may have deleted |
3467 | NEXT, so recompute it. */ |
3468 | next = next_active_insn (insn); |
3469 | } |
3470 | } |
3471 | |
3472 | |
3473 | /* Look for filled jumps to the end of function label. We can try to convert |
3474 | them into RETURN insns if the insns in the delay slot are valid for the |
3475 | RETURN as well. */ |
3476 | |
3477 | static void |
3478 | make_return_insns (rtx_insn *first) |
3479 | { |
3480 | rtx_insn *insn; |
3481 | rtx_jump_insn *jump_insn; |
3482 | rtx real_return_label = function_return_label; |
3483 | rtx real_simple_return_label = function_simple_return_label; |
3484 | int slots, i; |
3485 | |
3486 | /* See if there is a RETURN insn in the function other than the one we |
3487 | made for END_OF_FUNCTION_LABEL. If so, set up anything we can't change |
3488 | into a RETURN to jump to it. */ |
3489 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
3490 | if (JUMP_P (insn) && ANY_RETURN_P (PATTERN (insn))) |
3491 | { |
3492 | rtx t = get_label_before (insn, NULL_RTX); |
3493 | if (PATTERN (insn) == ret_rtx) |
3494 | real_return_label = t; |
3495 | else |
3496 | real_simple_return_label = t; |
3497 | break; |
3498 | } |
3499 | |
3500 | /* Show an extra usage of REAL_RETURN_LABEL so it won't go away if it |
3501 | was equal to END_OF_FUNCTION_LABEL. */ |
3502 | if (real_return_label) |
3503 | LABEL_NUSES (real_return_label)++; |
3504 | if (real_simple_return_label) |
3505 | LABEL_NUSES (real_simple_return_label)++; |
3506 | |
3507 | /* Clear the list of insns to fill so we can use it. */ |
3508 | obstack_free (&unfilled_slots_obstack, unfilled_firstobj); |
3509 | |
3510 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
3511 | { |
3512 | int flags; |
3513 | rtx kind, real_label; |
3514 | |
3515 | /* Only look at filled JUMP_INSNs that go to the end of function |
3516 | label. */ |
3517 | if (!NONJUMP_INSN_P (insn)) |
3518 | continue; |
3519 | |
3520 | if (GET_CODE (PATTERN (insn)) != SEQUENCE) |
3521 | continue; |
3522 | |
3523 | rtx_sequence *pat = as_a <rtx_sequence *> (p: PATTERN (insn)); |
3524 | |
3525 | if (!jump_to_label_p (pat->insn (index: 0))) |
3526 | continue; |
3527 | |
3528 | if (JUMP_LABEL (pat->insn (0)) == function_return_label) |
3529 | { |
3530 | kind = ret_rtx; |
3531 | real_label = real_return_label; |
3532 | } |
3533 | else if (JUMP_LABEL (pat->insn (0)) == function_simple_return_label) |
3534 | { |
3535 | kind = simple_return_rtx; |
3536 | real_label = real_simple_return_label; |
3537 | } |
3538 | else |
3539 | continue; |
3540 | |
3541 | jump_insn = as_a <rtx_jump_insn *> (p: pat->insn (index: 0)); |
3542 | |
3543 | /* If we can't make the jump into a RETURN, try to redirect it to the best |
3544 | RETURN and go on to the next insn. */ |
3545 | if (!reorg_redirect_jump (jump: jump_insn, nlabel: kind)) |
3546 | { |
3547 | /* Make sure redirecting the jump will not invalidate the delay |
3548 | slot insns. */ |
3549 | if (redirect_with_delay_slots_safe_p (jump: jump_insn, newlabel: real_label, seq: insn)) |
3550 | reorg_redirect_jump (jump: jump_insn, nlabel: real_label); |
3551 | continue; |
3552 | } |
3553 | |
3554 | /* See if this RETURN can accept the insns current in its delay slot. |
3555 | It can if it has more or an equal number of slots and the contents |
3556 | of each is valid. */ |
3557 | |
3558 | flags = get_jump_flags (insn: jump_insn, JUMP_LABEL (jump_insn)); |
3559 | slots = num_delay_slots (jump_insn); |
3560 | if (slots >= XVECLEN (pat, 0) - 1) |
3561 | { |
3562 | for (i = 1; i < XVECLEN (pat, 0); i++) |
3563 | if (! ( |
3564 | #if ANNUL_IFFALSE_SLOTS |
3565 | (INSN_ANNULLED_BRANCH_P (jump_insn) |
3566 | && INSN_FROM_TARGET_P (pat->insn (i))) |
3567 | ? eligible_for_annul_false (jump_insn, i - 1, |
3568 | pat->insn (i), flags) : |
3569 | #endif |
3570 | #if ANNUL_IFTRUE_SLOTS |
3571 | (INSN_ANNULLED_BRANCH_P (jump_insn) |
3572 | && ! INSN_FROM_TARGET_P (pat->insn (i))) |
3573 | ? eligible_for_annul_true (jump_insn, i - 1, |
3574 | pat->insn (i), flags) : |
3575 | #endif |
3576 | eligible_for_delay (jump_insn, i - 1, |
3577 | pat->insn (index: i), flags))) |
3578 | break; |
3579 | } |
3580 | else |
3581 | i = 0; |
3582 | |
3583 | if (i == XVECLEN (pat, 0)) |
3584 | continue; |
3585 | |
3586 | /* We have to do something with this insn. If it is an unconditional |
3587 | RETURN, delete the SEQUENCE and output the individual insns, |
3588 | followed by the RETURN. Then set things up so we try to find |
3589 | insns for its delay slots, if it needs some. */ |
3590 | if (ANY_RETURN_P (PATTERN (jump_insn))) |
3591 | { |
3592 | rtx_insn *after = PREV_INSN (insn); |
3593 | |
3594 | delete_related_insns (insn); |
3595 | insn = jump_insn; |
3596 | for (i = 1; i < pat->len (); i++) |
3597 | after = emit_copy_of_insn_after (pat->insn (index: i), after); |
3598 | add_insn_after (insn, after, NULL); |
3599 | emit_barrier_after (insn); |
3600 | |
3601 | if (slots) |
3602 | obstack_ptr_grow (&unfilled_slots_obstack, insn); |
3603 | } |
3604 | else |
3605 | /* It is probably more efficient to keep this with its current |
3606 | delay slot as a branch to a RETURN. */ |
3607 | reorg_redirect_jump (jump: jump_insn, nlabel: real_label); |
3608 | } |
3609 | |
3610 | /* Now delete REAL_RETURN_LABEL if we never used it. Then try to fill any |
3611 | new delay slots we have created. */ |
3612 | if (real_return_label != NULL_RTX && --LABEL_NUSES (real_return_label) == 0) |
3613 | delete_related_insns (real_return_label); |
3614 | if (real_simple_return_label != NULL_RTX |
3615 | && --LABEL_NUSES (real_simple_return_label) == 0) |
3616 | delete_related_insns (real_simple_return_label); |
3617 | |
3618 | fill_simple_delay_slots (non_jumps_p: true); |
3619 | fill_simple_delay_slots (non_jumps_p: false); |
3620 | } |
3621 | |
3622 | /* Try to find insns to place in delay slots. */ |
3623 | |
3624 | static void |
3625 | dbr_schedule (rtx_insn *first) |
3626 | { |
3627 | rtx_insn *insn, *next, *epilogue_insn = 0; |
3628 | bool need_return_insns; |
3629 | int i; |
3630 | |
3631 | /* If the current function has no insns other than the prologue and |
3632 | epilogue, then do not try to fill any delay slots. */ |
3633 | if (n_basic_blocks_for_fn (cfun) == NUM_FIXED_BLOCKS) |
3634 | return; |
3635 | |
3636 | /* Find the highest INSN_UID and allocate and initialize our map from |
3637 | INSN_UID's to position in code. */ |
3638 | for (max_uid = 0, insn = first; insn; insn = NEXT_INSN (insn)) |
3639 | { |
3640 | if (INSN_UID (insn) > max_uid) |
3641 | max_uid = INSN_UID (insn); |
3642 | if (NOTE_P (insn) |
3643 | && NOTE_KIND (insn) == NOTE_INSN_EPILOGUE_BEG) |
3644 | epilogue_insn = insn; |
3645 | } |
3646 | |
3647 | uid_to_ruid = XNEWVEC (int, max_uid + 1); |
3648 | for (i = 0, insn = first; insn; i++, insn = NEXT_INSN (insn)) |
3649 | uid_to_ruid[INSN_UID (insn)] = i; |
3650 | |
3651 | /* Initialize the list of insns that need filling. */ |
3652 | if (unfilled_firstobj == 0) |
3653 | { |
3654 | gcc_obstack_init (&unfilled_slots_obstack); |
3655 | unfilled_firstobj = XOBNEWVAR (&unfilled_slots_obstack, rtx, 0); |
3656 | } |
3657 | |
3658 | for (insn = next_active_insn (first); insn; insn = next_active_insn (insn)) |
3659 | { |
3660 | rtx target; |
3661 | |
3662 | /* Skip vector tables. We can't get attributes for them. */ |
3663 | if (JUMP_TABLE_DATA_P (insn)) |
3664 | continue; |
3665 | |
3666 | if (JUMP_P (insn)) |
3667 | INSN_ANNULLED_BRANCH_P (insn) = 0; |
3668 | INSN_FROM_TARGET_P (insn) = 0; |
3669 | |
3670 | if (num_delay_slots (insn) > 0) |
3671 | obstack_ptr_grow (&unfilled_slots_obstack, insn); |
3672 | |
3673 | /* Ensure all jumps go to the last of a set of consecutive labels. */ |
3674 | if (JUMP_P (insn) |
3675 | && (condjump_p (insn) || condjump_in_parallel_p (insn)) |
3676 | && !ANY_RETURN_P (JUMP_LABEL (insn)) |
3677 | && ((target = skip_consecutive_labels (JUMP_LABEL (insn))) |
3678 | != JUMP_LABEL (insn))) |
3679 | redirect_jump (as_a <rtx_jump_insn *> (p: insn), target, 1); |
3680 | } |
3681 | |
3682 | init_resource_info (epilogue_insn); |
3683 | |
3684 | /* Show we haven't computed an end-of-function label yet. */ |
3685 | function_return_label = function_simple_return_label = NULL; |
3686 | |
3687 | /* Initialize the statistics for this function. */ |
3688 | memset (s: num_insns_needing_delays, c: 0, n: sizeof num_insns_needing_delays); |
3689 | memset (s: num_filled_delays, c: 0, n: sizeof num_filled_delays); |
3690 | |
3691 | /* Now do the delay slot filling. Try everything twice in case earlier |
3692 | changes make more slots fillable. */ |
3693 | |
3694 | for (reorg_pass_number = 0; |
3695 | reorg_pass_number < MAX_REORG_PASSES; |
3696 | reorg_pass_number++) |
3697 | { |
3698 | fill_simple_delay_slots (non_jumps_p: true); |
3699 | fill_simple_delay_slots (non_jumps_p: false); |
3700 | if (!targetm.no_speculation_in_delay_slots_p ()) |
3701 | fill_eager_delay_slots (); |
3702 | relax_delay_slots (first); |
3703 | } |
3704 | |
3705 | /* If we made an end of function label, indicate that it is now |
3706 | safe to delete it by undoing our prior adjustment to LABEL_NUSES. |
3707 | If it is now unused, delete it. */ |
3708 | if (function_return_label && --LABEL_NUSES (function_return_label) == 0) |
3709 | delete_related_insns (function_return_label); |
3710 | if (function_simple_return_label |
3711 | && --LABEL_NUSES (function_simple_return_label) == 0) |
3712 | delete_related_insns (function_simple_return_label); |
3713 | |
3714 | need_return_insns = false; |
3715 | need_return_insns |= targetm.have_return () && function_return_label != 0; |
3716 | need_return_insns |= (targetm.have_simple_return () |
3717 | && function_simple_return_label != 0); |
3718 | if (need_return_insns) |
3719 | make_return_insns (first); |
3720 | |
3721 | /* Delete any USE insns made by update_block; subsequent passes don't need |
3722 | them or know how to deal with them. */ |
3723 | for (insn = first; insn; insn = next) |
3724 | { |
3725 | next = NEXT_INSN (insn); |
3726 | |
3727 | if (NONJUMP_INSN_P (insn) && GET_CODE (PATTERN (insn)) == USE |
3728 | && INSN_P (XEXP (PATTERN (insn), 0))) |
3729 | next = delete_related_insns (insn); |
3730 | } |
3731 | |
3732 | obstack_free (&unfilled_slots_obstack, unfilled_firstobj); |
3733 | |
3734 | /* It is not clear why the line below is needed, but it does seem to be. */ |
3735 | unfilled_firstobj = XOBNEWVAR (&unfilled_slots_obstack, rtx, 0); |
3736 | |
3737 | if (dump_file) |
3738 | { |
3739 | int i, j, need_comma; |
3740 | int total_delay_slots[MAX_DELAY_HISTOGRAM + 1]; |
3741 | int total_annul_slots[MAX_DELAY_HISTOGRAM + 1]; |
3742 | |
3743 | for (reorg_pass_number = 0; |
3744 | reorg_pass_number < MAX_REORG_PASSES; |
3745 | reorg_pass_number++) |
3746 | { |
3747 | fprintf (stream: dump_file, format: ";; Reorg pass #%d:\n" , reorg_pass_number + 1); |
3748 | for (i = 0; i < NUM_REORG_FUNCTIONS; i++) |
3749 | { |
3750 | need_comma = 0; |
3751 | fprintf (stream: dump_file, format: ";; Reorg function #%d\n" , i); |
3752 | |
3753 | fprintf (stream: dump_file, format: ";; %d insns needing delay slots\n;; " , |
3754 | num_insns_needing_delays[i][reorg_pass_number]); |
3755 | |
3756 | for (j = 0; j < MAX_DELAY_HISTOGRAM + 1; j++) |
3757 | if (num_filled_delays[i][j][reorg_pass_number]) |
3758 | { |
3759 | if (need_comma) |
3760 | fprintf (stream: dump_file, format: ", " ); |
3761 | need_comma = 1; |
3762 | fprintf (stream: dump_file, format: "%d got %d delays" , |
3763 | num_filled_delays[i][j][reorg_pass_number], j); |
3764 | } |
3765 | fprintf (stream: dump_file, format: "\n" ); |
3766 | } |
3767 | } |
3768 | memset (s: total_delay_slots, c: 0, n: sizeof total_delay_slots); |
3769 | memset (s: total_annul_slots, c: 0, n: sizeof total_annul_slots); |
3770 | for (insn = first; insn; insn = NEXT_INSN (insn)) |
3771 | { |
3772 | if (! insn->deleted () |
3773 | && NONJUMP_INSN_P (insn) |
3774 | && GET_CODE (PATTERN (insn)) != USE |
3775 | && GET_CODE (PATTERN (insn)) != CLOBBER) |
3776 | { |
3777 | if (GET_CODE (PATTERN (insn)) == SEQUENCE) |
3778 | { |
3779 | rtx control; |
3780 | j = XVECLEN (PATTERN (insn), 0) - 1; |
3781 | if (j > MAX_DELAY_HISTOGRAM) |
3782 | j = MAX_DELAY_HISTOGRAM; |
3783 | control = XVECEXP (PATTERN (insn), 0, 0); |
3784 | if (JUMP_P (control) && INSN_ANNULLED_BRANCH_P (control)) |
3785 | total_annul_slots[j]++; |
3786 | else |
3787 | total_delay_slots[j]++; |
3788 | } |
3789 | else if (num_delay_slots (insn) > 0) |
3790 | total_delay_slots[0]++; |
3791 | } |
3792 | } |
3793 | fprintf (stream: dump_file, format: ";; Reorg totals: " ); |
3794 | need_comma = 0; |
3795 | for (j = 0; j < MAX_DELAY_HISTOGRAM + 1; j++) |
3796 | { |
3797 | if (total_delay_slots[j]) |
3798 | { |
3799 | if (need_comma) |
3800 | fprintf (stream: dump_file, format: ", " ); |
3801 | need_comma = 1; |
3802 | fprintf (stream: dump_file, format: "%d got %d delays" , total_delay_slots[j], j); |
3803 | } |
3804 | } |
3805 | fprintf (stream: dump_file, format: "\n" ); |
3806 | |
3807 | if (ANNUL_IFTRUE_SLOTS || ANNUL_IFFALSE_SLOTS) |
3808 | { |
3809 | fprintf (stream: dump_file, format: ";; Reorg annuls: " ); |
3810 | need_comma = 0; |
3811 | for (j = 0; j < MAX_DELAY_HISTOGRAM + 1; j++) |
3812 | { |
3813 | if (total_annul_slots[j]) |
3814 | { |
3815 | if (need_comma) |
3816 | fprintf (stream: dump_file, format: ", " ); |
3817 | need_comma = 1; |
3818 | fprintf (stream: dump_file, format: "%d got %d delays" , total_annul_slots[j], j); |
3819 | } |
3820 | } |
3821 | fprintf (stream: dump_file, format: "\n" ); |
3822 | } |
3823 | |
3824 | fprintf (stream: dump_file, format: "\n" ); |
3825 | } |
3826 | |
3827 | if (!sibling_labels.is_empty ()) |
3828 | { |
3829 | update_alignments (sibling_labels); |
3830 | sibling_labels.release (); |
3831 | } |
3832 | |
3833 | free_resource_info (); |
3834 | free (ptr: uid_to_ruid); |
3835 | crtl->dbr_scheduled_p = true; |
3836 | } |
3837 | |
3838 | /* Run delay slot optimization. */ |
3839 | static void |
3840 | rest_of_handle_delay_slots (void) |
3841 | { |
3842 | if (DELAY_SLOTS) |
3843 | dbr_schedule (first: get_insns ()); |
3844 | } |
3845 | |
3846 | namespace { |
3847 | |
3848 | const pass_data pass_data_delay_slots = |
3849 | { |
3850 | .type: RTL_PASS, /* type */ |
3851 | .name: "dbr" , /* name */ |
3852 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
3853 | .tv_id: TV_DBR_SCHED, /* tv_id */ |
3854 | .properties_required: 0, /* properties_required */ |
3855 | .properties_provided: 0, /* properties_provided */ |
3856 | .properties_destroyed: 0, /* properties_destroyed */ |
3857 | .todo_flags_start: 0, /* todo_flags_start */ |
3858 | .todo_flags_finish: 0, /* todo_flags_finish */ |
3859 | }; |
3860 | |
3861 | class pass_delay_slots : public rtl_opt_pass |
3862 | { |
3863 | public: |
3864 | pass_delay_slots (gcc::context *ctxt) |
3865 | : rtl_opt_pass (pass_data_delay_slots, ctxt) |
3866 | {} |
3867 | |
3868 | /* opt_pass methods: */ |
3869 | bool gate (function *) final override; |
3870 | unsigned int execute (function *) final override |
3871 | { |
3872 | rest_of_handle_delay_slots (); |
3873 | return 0; |
3874 | } |
3875 | |
3876 | }; // class pass_delay_slots |
3877 | |
3878 | bool |
3879 | pass_delay_slots::gate (function *) |
3880 | { |
3881 | /* At -O0 dataflow info isn't updated after RA. */ |
3882 | if (DELAY_SLOTS) |
3883 | return optimize > 0 && flag_delayed_branch && !crtl->dbr_scheduled_p; |
3884 | |
3885 | return false; |
3886 | } |
3887 | |
3888 | } // anon namespace |
3889 | |
3890 | rtl_opt_pass * |
3891 | make_pass_delay_slots (gcc::context *ctxt) |
3892 | { |
3893 | return new pass_delay_slots (ctxt); |
3894 | } |
3895 | |
3896 | /* Machine dependent reorg pass. */ |
3897 | |
3898 | namespace { |
3899 | |
3900 | const pass_data pass_data_machine_reorg = |
3901 | { |
3902 | .type: RTL_PASS, /* type */ |
3903 | .name: "mach" , /* name */ |
3904 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
3905 | .tv_id: TV_MACH_DEP, /* tv_id */ |
3906 | .properties_required: 0, /* properties_required */ |
3907 | .properties_provided: 0, /* properties_provided */ |
3908 | .properties_destroyed: 0, /* properties_destroyed */ |
3909 | .todo_flags_start: 0, /* todo_flags_start */ |
3910 | .todo_flags_finish: 0, /* todo_flags_finish */ |
3911 | }; |
3912 | |
3913 | class pass_machine_reorg : public rtl_opt_pass |
3914 | { |
3915 | public: |
3916 | pass_machine_reorg (gcc::context *ctxt) |
3917 | : rtl_opt_pass (pass_data_machine_reorg, ctxt) |
3918 | {} |
3919 | |
3920 | /* opt_pass methods: */ |
3921 | bool gate (function *) final override |
3922 | { |
3923 | return targetm.machine_dependent_reorg != 0; |
3924 | } |
3925 | |
3926 | unsigned int execute (function *) final override |
3927 | { |
3928 | targetm.machine_dependent_reorg (); |
3929 | return 0; |
3930 | } |
3931 | |
3932 | }; // class pass_machine_reorg |
3933 | |
3934 | } // anon namespace |
3935 | |
3936 | rtl_opt_pass * |
3937 | make_pass_machine_reorg (gcc::context *ctxt) |
3938 | { |
3939 | return new pass_machine_reorg (ctxt); |
3940 | } |
3941 | |