1 | /* Generate code from machine description to recognize rtl as insns. |
2 | Copyright (C) 1987-2023 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it |
7 | under the terms of the GNU General Public License as published by |
8 | the Free Software Foundation; either version 3, or (at your option) |
9 | any later version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY |
13 | or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public |
14 | License for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | |
21 | /* This program is used to produce insn-recog.cc, which contains a |
22 | function called `recog' plus its subroutines. These functions |
23 | contain a decision tree that recognizes whether an rtx, the |
24 | argument given to recog, is a valid instruction. |
25 | |
26 | recog returns -1 if the rtx is not valid. If the rtx is valid, |
27 | recog returns a nonnegative number which is the insn code number |
28 | for the pattern that matched. This is the same as the order in the |
29 | machine description of the entry that matched. This number can be |
30 | used as an index into various insn_* tables, such as insn_template, |
31 | insn_outfun, and insn_n_operands (found in insn-output.cc). |
32 | |
33 | The third argument to recog is an optional pointer to an int. If |
34 | present, recog will accept a pattern if it matches except for |
35 | missing CLOBBER expressions at the end. In that case, the value |
36 | pointed to by the optional pointer will be set to the number of |
37 | CLOBBERs that need to be added (it should be initialized to zero by |
38 | the caller). If it is set nonzero, the caller should allocate a |
39 | PARALLEL of the appropriate size, copy the initial entries, and |
40 | call add_clobbers (found in insn-emit.cc) to fill in the CLOBBERs. |
41 | |
42 | This program also generates the function `split_insns', which |
43 | returns 0 if the rtl could not be split, or it returns the split |
44 | rtl as an INSN list. |
45 | |
46 | This program also generates the function `peephole2_insns', which |
47 | returns 0 if the rtl could not be matched. If there was a match, |
48 | the new rtl is returned in an INSN list, and LAST_INSN will point |
49 | to the last recognized insn in the old sequence. |
50 | |
51 | |
52 | At a high level, the algorithm used in this file is as follows: |
53 | |
54 | 1. Build up a decision tree for each routine, using the following |
55 | approach to matching an rtx: |
56 | |
57 | - First determine the "shape" of the rtx, based on GET_CODE, |
58 | XVECLEN and XINT. This phase examines SET_SRCs before SET_DESTs |
59 | since SET_SRCs tend to be more distinctive. It examines other |
60 | operands in numerical order, since the canonicalization rules |
61 | prefer putting complex operands of commutative operators first. |
62 | |
63 | - Next check modes and predicates. This phase examines all |
64 | operands in numerical order, even for SETs, since the mode of a |
65 | SET_DEST is exact while the mode of a SET_SRC can be VOIDmode |
66 | for constant integers. |
67 | |
68 | - Next check match_dups. |
69 | |
70 | - Finally check the C condition and (where appropriate) pnum_clobbers. |
71 | |
72 | 2. Try to optimize the tree by removing redundant tests, CSEing tests, |
73 | folding tests together, etc. |
74 | |
75 | 3. Look for common subtrees and split them out into "pattern" routines. |
76 | These common subtrees can be identical or they can differ in mode, |
77 | code, or integer (usually an UNSPEC or UNSPEC_VOLATILE code). |
78 | In the latter case the users of the pattern routine pass the |
79 | appropriate mode, etc., as argument. For example, if two patterns |
80 | contain: |
81 | |
82 | (plus:SI (match_operand:SI 1 "register_operand") |
83 | (match_operand:SI 2 "register_operand")) |
84 | |
85 | we can split the associated matching code out into a subroutine. |
86 | If a pattern contains: |
87 | |
88 | (minus:DI (match_operand:DI 1 "register_operand") |
89 | (match_operand:DI 2 "register_operand")) |
90 | |
91 | then we can consider using the same matching routine for both |
92 | the plus and minus expressions, passing PLUS and SImode in the |
93 | former case and MINUS and DImode in the latter case. |
94 | |
95 | The main aim of this phase is to reduce the compile time of the |
96 | insn-recog.cc code and to reduce the amount of object code in |
97 | insn-recog.o. |
98 | |
99 | 4. Split the matching trees into functions, trying to limit the |
100 | size of each function to a sensible amount. |
101 | |
102 | Again, the main aim of this phase is to reduce the compile time |
103 | of insn-recog.cc. (It doesn't help with the size of insn-recog.o.) |
104 | |
105 | 5. Write out C++ code for each function. */ |
106 | |
107 | #include "bconfig.h" |
108 | #define INCLUDE_ALGORITHM |
109 | #include "system.h" |
110 | #include "coretypes.h" |
111 | #include "tm.h" |
112 | #include "rtl.h" |
113 | #include "errors.h" |
114 | #include "read-md.h" |
115 | #include "gensupport.h" |
116 | |
117 | #undef GENERATOR_FILE |
118 | enum true_rtx_doe { |
119 | #define DEF_RTL_EXPR(ENUM, NAME, FORMAT, CLASS) TRUE_##ENUM, |
120 | #include "rtl.def" |
121 | #undef DEF_RTL_EXPR |
122 | FIRST_GENERATOR_RTX_CODE |
123 | }; |
124 | #define NUM_TRUE_RTX_CODE ((int) FIRST_GENERATOR_RTX_CODE) |
125 | #define GENERATOR_FILE 1 |
126 | |
127 | /* Debugging variables to control which optimizations are performed. |
128 | Note that disabling merge_states_p leads to very large output. */ |
129 | static const bool merge_states_p = true; |
130 | static const bool collapse_optional_decisions_p = true; |
131 | static const bool cse_tests_p = true; |
132 | static const bool simplify_tests_p = true; |
133 | static const bool use_operand_variables_p = true; |
134 | static const bool use_subroutines_p = true; |
135 | static const bool use_pattern_routines_p = true; |
136 | |
137 | /* Whether to add comments for optional tests that we decided to keep. |
138 | Can be useful when debugging the generator itself but is noise when |
139 | debugging the generated code. */ |
140 | static const bool mark_optional_transitions_p = false; |
141 | |
142 | /* Whether pattern routines should calculate positions relative to their |
143 | rtx parameter rather than use absolute positions. This e.g. allows |
144 | a pattern routine to be shared between a plain SET and a PARALLEL |
145 | that includes a SET. |
146 | |
147 | In principle it sounds like this should be useful, especially for |
148 | recog_for_combine, where the plain SET form is generated automatically |
149 | from a PARALLEL of a single SET and some CLOBBERs. In practice it doesn't |
150 | seem to help much and leads to slightly bigger object files. */ |
151 | static const bool relative_patterns_p = false; |
152 | |
153 | /* Whether pattern routines should be allowed to test whether pnum_clobbers |
154 | is null. This requires passing pnum_clobbers around as a parameter. */ |
155 | static const bool pattern_have_num_clobbers_p = true; |
156 | |
157 | /* Whether pattern routines should be allowed to test .md file C conditions. |
158 | This requires passing insn around as a parameter, in case the C |
159 | condition refers to it. In practice this tends to lead to bigger |
160 | object files. */ |
161 | static const bool pattern_c_test_p = false; |
162 | |
163 | /* Whether to require each parameter passed to a pattern routine to be |
164 | unique. Disabling this check for example allows unary operators with |
165 | matching modes (like NEG) and unary operators with mismatched modes |
166 | (like ZERO_EXTEND) to be matched by a single pattern. However, we then |
167 | often have cases where the same value is passed too many times. */ |
168 | static const bool force_unique_params_p = true; |
169 | |
170 | /* The maximum (approximate) depth of block nesting that an individual |
171 | routine or subroutine should have. This limit is about keeping the |
172 | output readable rather than reducing compile time. */ |
173 | static const unsigned int MAX_DEPTH = 6; |
174 | |
175 | /* The minimum number of pseudo-statements that a state must have before |
176 | we split it out into a subroutine. */ |
177 | static const unsigned int MIN_NUM_STATEMENTS = 5; |
178 | |
179 | /* The number of pseudo-statements a state can have before we consider |
180 | splitting out substates into subroutines. This limit is about avoiding |
181 | compile-time problems with very big functions (and also about keeping |
182 | functions within --param optimization limits, etc.). */ |
183 | static const unsigned int MAX_NUM_STATEMENTS = 200; |
184 | |
185 | /* The minimum number of pseudo-statements that can be used in a pattern |
186 | routine. */ |
187 | static const unsigned int MIN_COMBINE_COST = 4; |
188 | |
189 | /* The maximum number of arguments that a pattern routine can have. |
190 | The idea is to prevent one pattern getting a ridiculous number of |
191 | arguments when it would be more beneficial to have a separate pattern |
192 | routine instead. */ |
193 | static const unsigned int MAX_PATTERN_PARAMS = 5; |
194 | |
195 | /* The maximum operand number plus one. */ |
196 | int num_operands; |
197 | |
198 | /* Ways of obtaining an rtx to be tested. */ |
199 | enum position_type { |
200 | /* PATTERN (peep2_next_insn (ARG)). */ |
201 | POS_PEEP2_INSN, |
202 | |
203 | /* XEXP (BASE, ARG). */ |
204 | POS_XEXP, |
205 | |
206 | /* XVECEXP (BASE, 0, ARG). */ |
207 | POS_XVECEXP0 |
208 | }; |
209 | |
210 | /* The position of an rtx relative to X0. Each useful position is |
211 | represented by exactly one instance of this structure. */ |
212 | struct position |
213 | { |
214 | /* The parent rtx. This is the root position for POS_PEEP2_INSNs. */ |
215 | struct position *base; |
216 | |
217 | /* A position with the same BASE and TYPE, but with the next value |
218 | of ARG. */ |
219 | struct position *next; |
220 | |
221 | /* A list of all POS_XEXP positions that use this one as their base, |
222 | chained by NEXT fields. The first entry represents XEXP (this, 0), |
223 | the second represents XEXP (this, 1), and so on. */ |
224 | struct position *xexps; |
225 | |
226 | /* A list of POS_XVECEXP0 positions that use this one as their base, |
227 | chained by NEXT fields. The first entry represents XVECEXP (this, 0, 0), |
228 | the second represents XVECEXP (this, 0, 1), and so on. */ |
229 | struct position *xvecexp0s; |
230 | |
231 | /* The type of position. */ |
232 | enum position_type type; |
233 | |
234 | /* The argument to TYPE (shown as ARG in the position_type comments). */ |
235 | int arg; |
236 | |
237 | /* The instruction to which the position belongs. */ |
238 | unsigned int insn_id; |
239 | |
240 | /* The depth of this position relative to the instruction pattern. |
241 | E.g. if the instruction pattern is a SET, the SET itself has a |
242 | depth of 0 while the SET_DEST and SET_SRC have depths of 1. */ |
243 | unsigned int depth; |
244 | |
245 | /* A unique identifier for this position. */ |
246 | unsigned int id; |
247 | }; |
248 | |
249 | enum routine_type { |
250 | SUBPATTERN, RECOG, SPLIT, PEEPHOLE2 |
251 | }; |
252 | |
253 | /* The root position (x0). */ |
254 | static struct position root_pos; |
255 | |
256 | /* The number of positions created. Also one higher than the maximum |
257 | position id. */ |
258 | static unsigned int num_positions = 1; |
259 | |
260 | /* A list of all POS_PEEP2_INSNs. The entry for insn 0 is the root position, |
261 | since we are given that instruction's pattern as x0. */ |
262 | static struct position *peep2_insn_pos_list = &root_pos; |
263 | |
264 | /* Return a position with the given BASE, TYPE and ARG. NEXT_PTR |
265 | points to where the unique object that represents the position |
266 | should be stored. Create the object if it doesn't already exist, |
267 | otherwise reuse the object that is already there. */ |
268 | |
269 | static struct position * |
270 | next_position (struct position **next_ptr, struct position *base, |
271 | enum position_type type, int arg) |
272 | { |
273 | struct position *pos; |
274 | |
275 | pos = *next_ptr; |
276 | if (!pos) |
277 | { |
278 | pos = XCNEW (struct position); |
279 | pos->type = type; |
280 | pos->arg = arg; |
281 | if (type == POS_PEEP2_INSN) |
282 | { |
283 | pos->base = 0; |
284 | pos->insn_id = arg; |
285 | pos->depth = base->depth; |
286 | } |
287 | else |
288 | { |
289 | pos->base = base; |
290 | pos->insn_id = base->insn_id; |
291 | pos->depth = base->depth + 1; |
292 | } |
293 | pos->id = num_positions++; |
294 | *next_ptr = pos; |
295 | } |
296 | return pos; |
297 | } |
298 | |
299 | /* Compare positions POS1 and POS2 lexicographically. */ |
300 | |
301 | static int |
302 | compare_positions (struct position *pos1, struct position *pos2) |
303 | { |
304 | int diff; |
305 | |
306 | diff = pos1->depth - pos2->depth; |
307 | if (diff < 0) |
308 | do |
309 | pos2 = pos2->base; |
310 | while (pos1->depth != pos2->depth); |
311 | else if (diff > 0) |
312 | do |
313 | pos1 = pos1->base; |
314 | while (pos1->depth != pos2->depth); |
315 | while (pos1 != pos2) |
316 | { |
317 | diff = (int) pos1->type - (int) pos2->type; |
318 | if (diff == 0) |
319 | diff = pos1->arg - pos2->arg; |
320 | pos1 = pos1->base; |
321 | pos2 = pos2->base; |
322 | } |
323 | return diff; |
324 | } |
325 | |
326 | /* Return the most deeply-nested position that is common to both |
327 | POS1 and POS2. If the positions are from different instructions, |
328 | return the one with the lowest insn_id. */ |
329 | |
330 | static struct position * |
331 | common_position (struct position *pos1, struct position *pos2) |
332 | { |
333 | if (pos1->insn_id != pos2->insn_id) |
334 | return pos1->insn_id < pos2->insn_id ? pos1 : pos2; |
335 | if (pos1->depth > pos2->depth) |
336 | std::swap (a&: pos1, b&: pos2); |
337 | while (pos1->depth != pos2->depth) |
338 | pos2 = pos2->base; |
339 | while (pos1 != pos2) |
340 | { |
341 | pos1 = pos1->base; |
342 | pos2 = pos2->base; |
343 | } |
344 | return pos1; |
345 | } |
346 | |
347 | /* Search for and return operand N, stop when reaching node STOP. */ |
348 | |
349 | static rtx |
350 | find_operand (rtx pattern, int n, rtx stop) |
351 | { |
352 | const char *fmt; |
353 | RTX_CODE code; |
354 | int i, j, len; |
355 | rtx r; |
356 | |
357 | if (pattern == stop) |
358 | return stop; |
359 | |
360 | code = GET_CODE (pattern); |
361 | if ((code == MATCH_SCRATCH |
362 | || code == MATCH_OPERAND |
363 | || code == MATCH_OPERATOR |
364 | || code == MATCH_PARALLEL) |
365 | && XINT (pattern, 0) == n) |
366 | return pattern; |
367 | |
368 | fmt = GET_RTX_FORMAT (code); |
369 | len = GET_RTX_LENGTH (code); |
370 | for (i = 0; i < len; i++) |
371 | { |
372 | switch (fmt[i]) |
373 | { |
374 | case 'e': case 'u': |
375 | if ((r = find_operand (XEXP (pattern, i), n, stop)) != NULL_RTX) |
376 | return r; |
377 | break; |
378 | |
379 | case 'V': |
380 | if (! XVEC (pattern, i)) |
381 | break; |
382 | /* Fall through. */ |
383 | |
384 | case 'E': |
385 | for (j = 0; j < XVECLEN (pattern, i); j++) |
386 | if ((r = find_operand (XVECEXP (pattern, i, j), n, stop)) |
387 | != NULL_RTX) |
388 | return r; |
389 | break; |
390 | |
391 | case 'r': case 'p': case 'i': case 'w': case '0': case 's': |
392 | break; |
393 | |
394 | default: |
395 | gcc_unreachable (); |
396 | } |
397 | } |
398 | |
399 | return NULL; |
400 | } |
401 | |
402 | /* Search for and return operand M, such that it has a matching |
403 | constraint for operand N. */ |
404 | |
405 | static rtx |
406 | find_matching_operand (rtx pattern, int n) |
407 | { |
408 | const char *fmt; |
409 | RTX_CODE code; |
410 | int i, j, len; |
411 | rtx r; |
412 | |
413 | code = GET_CODE (pattern); |
414 | if (code == MATCH_OPERAND |
415 | && (XSTR (pattern, 2)[0] == '0' + n |
416 | || (XSTR (pattern, 2)[0] == '%' |
417 | && XSTR (pattern, 2)[1] == '0' + n))) |
418 | return pattern; |
419 | |
420 | fmt = GET_RTX_FORMAT (code); |
421 | len = GET_RTX_LENGTH (code); |
422 | for (i = 0; i < len; i++) |
423 | { |
424 | switch (fmt[i]) |
425 | { |
426 | case 'e': case 'u': |
427 | if ((r = find_matching_operand (XEXP (pattern, i), n))) |
428 | return r; |
429 | break; |
430 | |
431 | case 'V': |
432 | if (! XVEC (pattern, i)) |
433 | break; |
434 | /* Fall through. */ |
435 | |
436 | case 'E': |
437 | for (j = 0; j < XVECLEN (pattern, i); j++) |
438 | if ((r = find_matching_operand (XVECEXP (pattern, i, j), n))) |
439 | return r; |
440 | break; |
441 | |
442 | case 'r': case 'p': case 'i': case 'w': case '0': case 's': |
443 | break; |
444 | |
445 | default: |
446 | gcc_unreachable (); |
447 | } |
448 | } |
449 | |
450 | return NULL; |
451 | } |
452 | |
453 | /* In DEFINE_EXPAND, DEFINE_SPLIT, and DEFINE_PEEPHOLE2, we |
454 | don't use the MATCH_OPERAND constraint, only the predicate. |
455 | This is confusing to folks doing new ports, so help them |
456 | not make the mistake. */ |
457 | |
458 | static bool |
459 | constraints_supported_in_insn_p (rtx insn) |
460 | { |
461 | return !(GET_CODE (insn) == DEFINE_EXPAND |
462 | || GET_CODE (insn) == DEFINE_SPLIT |
463 | || GET_CODE (insn) == DEFINE_PEEPHOLE2); |
464 | } |
465 | |
466 | /* Return the name of the predicate matched by MATCH_RTX. */ |
467 | |
468 | static const char * |
469 | predicate_name (rtx match_rtx) |
470 | { |
471 | if (GET_CODE (match_rtx) == MATCH_SCRATCH) |
472 | return "scratch_operand" ; |
473 | else |
474 | return XSTR (match_rtx, 1); |
475 | } |
476 | |
477 | /* Return true if OPERAND is a MATCH_OPERAND using a special predicate |
478 | function. */ |
479 | |
480 | static bool |
481 | special_predicate_operand_p (rtx operand) |
482 | { |
483 | if (GET_CODE (operand) == MATCH_OPERAND) |
484 | { |
485 | const char *pred_name = predicate_name (match_rtx: operand); |
486 | if (pred_name[0] != 0) |
487 | { |
488 | const struct pred_data *pred; |
489 | |
490 | pred = lookup_predicate (pred_name); |
491 | return pred != NULL && pred->special; |
492 | } |
493 | } |
494 | |
495 | return false; |
496 | } |
497 | |
498 | /* Check for various errors in PATTERN, which is part of INFO. |
499 | SET is nonnull for a destination, and is the complete set pattern. |
500 | SET_CODE is '=' for normal sets, and '+' within a context that |
501 | requires in-out constraints. */ |
502 | |
503 | static void |
504 | validate_pattern (rtx pattern, md_rtx_info *info, rtx set, int set_code) |
505 | { |
506 | const char *fmt; |
507 | RTX_CODE code; |
508 | size_t i, len; |
509 | int j; |
510 | |
511 | code = GET_CODE (pattern); |
512 | switch (code) |
513 | { |
514 | case MATCH_SCRATCH: |
515 | { |
516 | const char constraints0 = XSTR (pattern, 1)[0]; |
517 | |
518 | if (!constraints_supported_in_insn_p (insn: info->def)) |
519 | { |
520 | if (constraints0) |
521 | { |
522 | error_at (info->loc, "constraints not supported in %s" , |
523 | GET_RTX_NAME (GET_CODE (info->def))); |
524 | } |
525 | return; |
526 | } |
527 | |
528 | /* If a MATCH_SCRATCH is used in a context requiring an write-only |
529 | or read/write register, validate that. */ |
530 | if (set_code == '=' |
531 | && constraints0 |
532 | && constraints0 != '=' |
533 | && constraints0 != '+') |
534 | { |
535 | error_at (info->loc, "operand %d missing output reload" , |
536 | XINT (pattern, 0)); |
537 | } |
538 | return; |
539 | } |
540 | case MATCH_DUP: |
541 | case MATCH_OP_DUP: |
542 | case MATCH_PAR_DUP: |
543 | if (find_operand (pattern: info->def, XINT (pattern, 0), stop: pattern) == pattern) |
544 | error_at (info->loc, "operand %i duplicated before defined" , |
545 | XINT (pattern, 0)); |
546 | break; |
547 | case MATCH_OPERAND: |
548 | case MATCH_OPERATOR: |
549 | { |
550 | const char *pred_name = XSTR (pattern, 1); |
551 | const struct pred_data *pred; |
552 | const char *c_test; |
553 | |
554 | c_test = get_c_test (info->def); |
555 | |
556 | if (pred_name[0] != 0) |
557 | { |
558 | pred = lookup_predicate (pred_name); |
559 | if (!pred) |
560 | error_at (info->loc, "unknown predicate '%s'" , pred_name); |
561 | } |
562 | else |
563 | pred = 0; |
564 | |
565 | if (code == MATCH_OPERAND) |
566 | { |
567 | const char *constraints = XSTR (pattern, 2); |
568 | const char constraints0 = constraints[0]; |
569 | |
570 | if (!constraints_supported_in_insn_p (insn: info->def)) |
571 | { |
572 | if (constraints0) |
573 | { |
574 | error_at (info->loc, "constraints not supported in %s" , |
575 | GET_RTX_NAME (GET_CODE (info->def))); |
576 | } |
577 | } |
578 | |
579 | /* A MATCH_OPERAND that is a SET should have an output reload. */ |
580 | else if (set && constraints0) |
581 | { |
582 | if (set_code == '+') |
583 | { |
584 | if (constraints0 == '+') |
585 | ; |
586 | /* If we've only got an output reload for this operand, |
587 | we'd better have a matching input operand. */ |
588 | else if (constraints0 == '=' |
589 | && find_matching_operand (pattern: info->def, |
590 | XINT (pattern, 0))) |
591 | ; |
592 | else |
593 | error_at (info->loc, "operand %d missing in-out reload" , |
594 | XINT (pattern, 0)); |
595 | } |
596 | else if (constraints0 != '=' && constraints0 != '+') |
597 | error_at (info->loc, "operand %d missing output reload" , |
598 | XINT (pattern, 0)); |
599 | } |
600 | |
601 | /* For matching constraint in MATCH_OPERAND, the digit must be a |
602 | smaller number than the number of the operand that uses it in the |
603 | constraint. */ |
604 | while (1) |
605 | { |
606 | while (constraints[0] |
607 | && (constraints[0] == ' ' || constraints[0] == ',')) |
608 | constraints++; |
609 | if (!constraints[0]) |
610 | break; |
611 | |
612 | if (constraints[0] >= '0' && constraints[0] <= '9') |
613 | { |
614 | int val; |
615 | |
616 | sscanf (s: constraints, format: "%d" , &val); |
617 | if (val >= XINT (pattern, 0)) |
618 | error_at (info->loc, "constraint digit %d is not" |
619 | " smaller than operand %d" , |
620 | val, XINT (pattern, 0)); |
621 | } |
622 | |
623 | while (constraints[0] && constraints[0] != ',') |
624 | constraints++; |
625 | } |
626 | } |
627 | |
628 | /* Allowing non-lvalues in destinations -- particularly CONST_INT -- |
629 | while not likely to occur at runtime, results in less efficient |
630 | code from insn-recog.cc. */ |
631 | if (set && pred && pred->allows_non_lvalue) |
632 | error_at (info->loc, "destination operand %d allows non-lvalue" , |
633 | XINT (pattern, 0)); |
634 | |
635 | /* A modeless MATCH_OPERAND can be handy when we can check for |
636 | multiple modes in the c_test. In most other cases, it is a |
637 | mistake. Only DEFINE_INSN is eligible, since SPLIT and |
638 | PEEP2 can FAIL within the output pattern. Exclude special |
639 | predicates, which check the mode themselves. Also exclude |
640 | predicates that allow only constants. Exclude the SET_DEST |
641 | of a call instruction, as that is a common idiom. */ |
642 | |
643 | if (GET_MODE (pattern) == VOIDmode |
644 | && code == MATCH_OPERAND |
645 | && GET_CODE (info->def) == DEFINE_INSN |
646 | && pred |
647 | && !pred->special |
648 | && pred->allows_non_const |
649 | && strstr (haystack: c_test, needle: "operands" ) == NULL |
650 | && ! (set |
651 | && GET_CODE (set) == SET |
652 | && GET_CODE (SET_SRC (set)) == CALL)) |
653 | message_at (info->loc, "warning: operand %d missing mode?" , |
654 | XINT (pattern, 0)); |
655 | return; |
656 | } |
657 | |
658 | case SET: |
659 | { |
660 | machine_mode dmode, smode; |
661 | rtx dest, src; |
662 | |
663 | dest = SET_DEST (pattern); |
664 | src = SET_SRC (pattern); |
665 | |
666 | /* STRICT_LOW_PART is a wrapper. Its argument is the real |
667 | destination, and it's mode should match the source. */ |
668 | if (GET_CODE (dest) == STRICT_LOW_PART) |
669 | dest = XEXP (dest, 0); |
670 | |
671 | /* Find the referent for a DUP. */ |
672 | |
673 | if (GET_CODE (dest) == MATCH_DUP |
674 | || GET_CODE (dest) == MATCH_OP_DUP |
675 | || GET_CODE (dest) == MATCH_PAR_DUP) |
676 | dest = find_operand (pattern: info->def, XINT (dest, 0), NULL); |
677 | |
678 | if (GET_CODE (src) == MATCH_DUP |
679 | || GET_CODE (src) == MATCH_OP_DUP |
680 | || GET_CODE (src) == MATCH_PAR_DUP) |
681 | src = find_operand (pattern: info->def, XINT (src, 0), NULL); |
682 | |
683 | dmode = GET_MODE (dest); |
684 | smode = GET_MODE (src); |
685 | |
686 | /* Mode checking is not performed for special predicates. */ |
687 | if (special_predicate_operand_p (operand: src) |
688 | || special_predicate_operand_p (operand: dest)) |
689 | ; |
690 | |
691 | /* The operands of a SET must have the same mode unless one |
692 | is VOIDmode. */ |
693 | else if (dmode != VOIDmode && smode != VOIDmode && dmode != smode) |
694 | error_at (info->loc, "mode mismatch in set: %smode vs %smode" , |
695 | GET_MODE_NAME (dmode), GET_MODE_NAME (smode)); |
696 | |
697 | /* If only one of the operands is VOIDmode, and PC is not involved, |
698 | it's probably a mistake. */ |
699 | else if (dmode != smode |
700 | && GET_CODE (dest) != PC |
701 | && GET_CODE (src) != PC |
702 | && !CONST_INT_P (src) |
703 | && !CONST_WIDE_INT_P (src) |
704 | && GET_CODE (src) != CALL) |
705 | { |
706 | const char *which; |
707 | which = (dmode == VOIDmode ? "destination" : "source" ); |
708 | message_at (info->loc, "warning: %s missing a mode?" , which); |
709 | } |
710 | |
711 | if (dest != SET_DEST (pattern)) |
712 | validate_pattern (pattern: dest, info, set: pattern, set_code: '='); |
713 | validate_pattern (SET_DEST (pattern), info, set: pattern, set_code: '='); |
714 | validate_pattern (SET_SRC (pattern), info, NULL_RTX, set_code: 0); |
715 | return; |
716 | } |
717 | |
718 | case CLOBBER: |
719 | validate_pattern (SET_DEST (pattern), info, set: pattern, set_code: '='); |
720 | return; |
721 | |
722 | case ZERO_EXTRACT: |
723 | validate_pattern (XEXP (pattern, 0), info, set, set_code: set ? '+' : 0); |
724 | validate_pattern (XEXP (pattern, 1), info, NULL_RTX, set_code: 0); |
725 | validate_pattern (XEXP (pattern, 2), info, NULL_RTX, set_code: 0); |
726 | return; |
727 | |
728 | case STRICT_LOW_PART: |
729 | validate_pattern (XEXP (pattern, 0), info, set, set_code: set ? '+' : 0); |
730 | return; |
731 | |
732 | case LABEL_REF: |
733 | if (GET_MODE (XEXP (pattern, 0)) != VOIDmode) |
734 | error_at (info->loc, "operand to label_ref %smode not VOIDmode" , |
735 | GET_MODE_NAME (GET_MODE (XEXP (pattern, 0)))); |
736 | break; |
737 | |
738 | case VEC_SELECT: |
739 | if (GET_MODE (pattern) != VOIDmode) |
740 | { |
741 | machine_mode mode = GET_MODE (pattern); |
742 | machine_mode imode = GET_MODE (XEXP (pattern, 0)); |
743 | machine_mode emode |
744 | = VECTOR_MODE_P (mode) ? GET_MODE_INNER (mode) : mode; |
745 | if (GET_CODE (XEXP (pattern, 1)) == PARALLEL) |
746 | { |
747 | int expected = 1; |
748 | unsigned int nelems; |
749 | if (VECTOR_MODE_P (mode) |
750 | && !GET_MODE_NUNITS (mode).is_constant (const_value: &expected)) |
751 | error_at (info->loc, |
752 | "vec_select with variable-sized mode %s" , |
753 | GET_MODE_NAME (mode)); |
754 | else if (XVECLEN (XEXP (pattern, 1), 0) != expected) |
755 | error_at (info->loc, |
756 | "vec_select parallel with %d elements, expected %d" , |
757 | XVECLEN (XEXP (pattern, 1), 0), expected); |
758 | else if (VECTOR_MODE_P (imode) |
759 | && GET_MODE_NUNITS (mode: imode).is_constant (const_value: &nelems)) |
760 | { |
761 | int i; |
762 | for (i = 0; i < expected; ++i) |
763 | if (CONST_INT_P (XVECEXP (XEXP (pattern, 1), 0, i)) |
764 | && (UINTVAL (XVECEXP (XEXP (pattern, 1), 0, i)) |
765 | >= nelems)) |
766 | error_at (info->loc, |
767 | "out of bounds selector %u in vec_select, " |
768 | "expected at most %u" , |
769 | (unsigned) |
770 | UINTVAL (XVECEXP (XEXP (pattern, 1), 0, i)), |
771 | nelems - 1); |
772 | } |
773 | } |
774 | if (imode != VOIDmode && !VECTOR_MODE_P (imode)) |
775 | error_at (info->loc, "%smode of first vec_select operand is not a " |
776 | "vector mode" , GET_MODE_NAME (imode)); |
777 | else if (imode != VOIDmode && GET_MODE_INNER (imode) != emode) |
778 | error_at (info->loc, "element mode mismatch between vec_select " |
779 | "%smode and its operand %smode" , |
780 | GET_MODE_NAME (emode), |
781 | GET_MODE_NAME (GET_MODE_INNER (imode))); |
782 | } |
783 | break; |
784 | |
785 | default: |
786 | break; |
787 | } |
788 | |
789 | fmt = GET_RTX_FORMAT (code); |
790 | len = GET_RTX_LENGTH (code); |
791 | for (i = 0; i < len; i++) |
792 | { |
793 | switch (fmt[i]) |
794 | { |
795 | case 'e': case 'u': |
796 | validate_pattern (XEXP (pattern, i), info, NULL_RTX, set_code: 0); |
797 | break; |
798 | |
799 | case 'E': |
800 | for (j = 0; j < XVECLEN (pattern, i); j++) |
801 | validate_pattern (XVECEXP (pattern, i, j), info, NULL_RTX, set_code: 0); |
802 | break; |
803 | |
804 | case 'r': case 'p': case 'i': case 'w': case '0': case 's': |
805 | break; |
806 | |
807 | default: |
808 | gcc_unreachable (); |
809 | } |
810 | } |
811 | } |
812 | |
813 | /* Simple list structure for items of type T, for use when being part |
814 | of a list is an inherent property of T. T must have members equivalent |
815 | to "T *prev, *next;" and a function "void set_parent (list_head <T> *)" |
816 | to set the parent list. */ |
817 | template <typename T> |
818 | class list_head |
819 | { |
820 | public: |
821 | /* A range of linked items. */ |
822 | class range |
823 | { |
824 | public: |
825 | range (T *); |
826 | range (T *, T *); |
827 | |
828 | T *start, *end; |
829 | void set_parent (list_head *); |
830 | }; |
831 | |
832 | list_head (); |
833 | range release (); |
834 | void push_back (range); |
835 | range remove (range); |
836 | void replace (range, range); |
837 | T *singleton () const; |
838 | |
839 | T *first, *last; |
840 | }; |
841 | |
842 | /* Create a range [START_IN, START_IN]. */ |
843 | |
844 | template <typename T> |
845 | list_head <T>::range::range (T *start_in) : start (start_in), end (start_in) {} |
846 | |
847 | /* Create a range [START_IN, END_IN], linked by next and prev fields. */ |
848 | |
849 | template <typename T> |
850 | list_head <T>::range::range (T *start_in, T *end_in) |
851 | : start (start_in), end (end_in) {} |
852 | |
853 | template <typename T> |
854 | void |
855 | list_head <T>::range::set_parent (list_head <T> *owner) |
856 | { |
857 | for (T *item = start; item != end; item = item->next) |
858 | item->set_parent (owner); |
859 | end->set_parent (owner); |
860 | } |
861 | |
862 | template <typename T> |
863 | list_head <T>::list_head () : first (0), last (0) {} |
864 | |
865 | /* Add R to the end of the list. */ |
866 | |
867 | template <typename T> |
868 | void |
869 | list_head <T>::push_back (range r) |
870 | { |
871 | if (last) |
872 | last->next = r.start; |
873 | else |
874 | first = r.start; |
875 | r.start->prev = last; |
876 | last = r.end; |
877 | r.set_parent (this); |
878 | } |
879 | |
880 | /* Remove R from the list. R remains valid and can be inserted into |
881 | other lists. */ |
882 | |
883 | template <typename T> |
884 | typename list_head <T>::range |
885 | list_head <T>::remove (range r) |
886 | { |
887 | if (r.start->prev) |
888 | r.start->prev->next = r.end->next; |
889 | else |
890 | first = r.end->next; |
891 | if (r.end->next) |
892 | r.end->next->prev = r.start->prev; |
893 | else |
894 | last = r.start->prev; |
895 | r.start->prev = 0; |
896 | r.end->next = 0; |
897 | r.set_parent (0); |
898 | return r; |
899 | } |
900 | |
901 | /* Replace OLDR with NEWR. OLDR remains valid and can be inserted into |
902 | other lists. */ |
903 | |
904 | template <typename T> |
905 | void |
906 | list_head <T>::replace (range oldr, range newr) |
907 | { |
908 | newr.start->prev = oldr.start->prev; |
909 | newr.end->next = oldr.end->next; |
910 | |
911 | oldr.start->prev = 0; |
912 | oldr.end->next = 0; |
913 | oldr.set_parent (0); |
914 | |
915 | if (newr.start->prev) |
916 | newr.start->prev->next = newr.start; |
917 | else |
918 | first = newr.start; |
919 | if (newr.end->next) |
920 | newr.end->next->prev = newr.end; |
921 | else |
922 | last = newr.end; |
923 | newr.set_parent (this); |
924 | } |
925 | |
926 | /* Empty the list and return the previous contents as a range that can |
927 | be inserted into other lists. */ |
928 | |
929 | template <typename T> |
930 | typename list_head <T>::range |
931 | list_head <T>::release () |
932 | { |
933 | range r (first, last); |
934 | first = 0; |
935 | last = 0; |
936 | r.set_parent (0); |
937 | return r; |
938 | } |
939 | |
940 | /* If the list contains a single item, return that item, otherwise return |
941 | null. */ |
942 | |
943 | template <typename T> |
944 | T * |
945 | list_head <T>::singleton () const |
946 | { |
947 | return first == last ? first : 0; |
948 | } |
949 | |
950 | class state; |
951 | |
952 | /* Describes a possible successful return from a routine. */ |
953 | struct acceptance_type |
954 | { |
955 | /* The type of routine we're returning from. */ |
956 | routine_type type : 16; |
957 | |
958 | /* True if this structure only really represents a partial match, |
959 | and if we must call a subroutine of type TYPE to complete the match. |
960 | In this case we'll call the subroutine and, if it succeeds, return |
961 | whatever the subroutine returned. |
962 | |
963 | False if this structure presents a full match. */ |
964 | unsigned int partial_p : 1; |
965 | |
966 | union |
967 | { |
968 | /* If PARTIAL_P, this is the number of the subroutine to call. */ |
969 | int subroutine_id; |
970 | |
971 | /* Valid if !PARTIAL_P. */ |
972 | struct |
973 | { |
974 | /* The identifier of the matching pattern. For SUBPATTERNs this |
975 | value belongs to an ad-hoc routine-specific enum. For the |
976 | others it's the number of an .md file pattern. */ |
977 | int code; |
978 | union |
979 | { |
980 | /* For RECOG, the number of clobbers that must be added to the |
981 | pattern in order for it to match CODE. */ |
982 | int num_clobbers; |
983 | |
984 | /* For PEEPHOLE2, the number of additional instructions that were |
985 | included in the optimization. */ |
986 | int match_len; |
987 | } u; |
988 | } full; |
989 | } u; |
990 | }; |
991 | |
992 | bool |
993 | operator == (const acceptance_type &a, const acceptance_type &b) |
994 | { |
995 | if (a.partial_p != b.partial_p) |
996 | return false; |
997 | if (a.partial_p) |
998 | return a.u.subroutine_id == b.u.subroutine_id; |
999 | else |
1000 | return a.u.full.code == b.u.full.code; |
1001 | } |
1002 | |
1003 | bool |
1004 | operator != (const acceptance_type &a, const acceptance_type &b) |
1005 | { |
1006 | return !operator == (a, b); |
1007 | } |
1008 | |
1009 | /* Represents a parameter to a pattern routine. */ |
1010 | class parameter |
1011 | { |
1012 | public: |
1013 | /* The C type of parameter. */ |
1014 | enum type_enum { |
1015 | /* Represents an invalid parameter. */ |
1016 | UNSET, |
1017 | |
1018 | /* A machine_mode parameter. */ |
1019 | MODE, |
1020 | |
1021 | /* An rtx_code parameter. */ |
1022 | CODE, |
1023 | |
1024 | /* An int parameter. */ |
1025 | INT, |
1026 | |
1027 | /* An unsigned int parameter. */ |
1028 | UINT, |
1029 | |
1030 | /* A HOST_WIDE_INT parameter. */ |
1031 | WIDE_INT |
1032 | }; |
1033 | |
1034 | parameter (); |
1035 | parameter (type_enum, bool, uint64_t); |
1036 | |
1037 | /* The type of the parameter. */ |
1038 | type_enum type; |
1039 | |
1040 | /* True if the value passed is variable, false if it is constant. */ |
1041 | bool is_param; |
1042 | |
1043 | /* If IS_PARAM, this is the number of the variable passed, for an "i%d" |
1044 | format string. If !IS_PARAM, this is the constant value passed. */ |
1045 | uint64_t value; |
1046 | }; |
1047 | |
1048 | parameter::parameter () |
1049 | : type (UNSET), is_param (false), value (0) {} |
1050 | |
1051 | parameter::parameter (type_enum type_in, bool is_param_in, uint64_t value_in) |
1052 | : type (type_in), is_param (is_param_in), value (value_in) {} |
1053 | |
1054 | bool |
1055 | operator == (const parameter ¶m1, const parameter ¶m2) |
1056 | { |
1057 | return (param1.type == param2.type |
1058 | && param1.is_param == param2.is_param |
1059 | && param1.value == param2.value); |
1060 | } |
1061 | |
1062 | bool |
1063 | operator != (const parameter ¶m1, const parameter ¶m2) |
1064 | { |
1065 | return !operator == (param1, param2); |
1066 | } |
1067 | |
1068 | /* Represents a routine that matches a partial rtx pattern, returning |
1069 | an ad-hoc enum value on success and -1 on failure. The routine can |
1070 | be used by any subroutine type. The match can be parameterized by |
1071 | things like mode, code and UNSPEC number. */ |
1072 | class pattern_routine |
1073 | { |
1074 | public: |
1075 | /* The state that implements the pattern. */ |
1076 | state *s; |
1077 | |
1078 | /* The deepest root position from which S can access all the rtxes it needs. |
1079 | This is NULL if the pattern doesn't need an rtx input, usually because |
1080 | all matching is done on operands[] instead. */ |
1081 | position *pos; |
1082 | |
1083 | /* A unique identifier for the routine. */ |
1084 | unsigned int pattern_id; |
1085 | |
1086 | /* True if the routine takes pnum_clobbers as argument. */ |
1087 | bool pnum_clobbers_p; |
1088 | |
1089 | /* True if the routine takes the enclosing instruction as argument. */ |
1090 | bool insn_p; |
1091 | |
1092 | /* The types of the other parameters to the routine, if any. */ |
1093 | auto_vec <parameter::type_enum, MAX_PATTERN_PARAMS> param_types; |
1094 | }; |
1095 | |
1096 | /* All defined patterns. */ |
1097 | static vec <pattern_routine *> patterns; |
1098 | |
1099 | /* Represents one use of a pattern routine. */ |
1100 | class pattern_use |
1101 | { |
1102 | public: |
1103 | /* The pattern routine to use. */ |
1104 | pattern_routine *routine; |
1105 | |
1106 | /* The values to pass as parameters. This vector has the same length |
1107 | as ROUTINE->PARAM_TYPES. */ |
1108 | auto_vec <parameter, MAX_PATTERN_PARAMS> params; |
1109 | }; |
1110 | |
1111 | /* Represents a test performed by a decision. */ |
1112 | class rtx_test |
1113 | { |
1114 | public: |
1115 | rtx_test (); |
1116 | |
1117 | /* The types of test that can be performed. Most of them take as input |
1118 | an rtx X. Some also take as input a transition label LABEL; the others |
1119 | are booleans for which the transition label is always "true". |
1120 | |
1121 | The order of the enum isn't important. */ |
1122 | enum kind_enum { |
1123 | /* Check GET_CODE (X) == LABEL. */ |
1124 | CODE, |
1125 | |
1126 | /* Check GET_MODE (X) == LABEL. */ |
1127 | MODE, |
1128 | |
1129 | /* Check REGNO (X) == LABEL. */ |
1130 | REGNO_FIELD, |
1131 | |
1132 | /* Check known_eq (SUBREG_BYTE (X), LABEL). */ |
1133 | SUBREG_FIELD, |
1134 | |
1135 | /* Check XINT (X, u.opno) == LABEL. */ |
1136 | INT_FIELD, |
1137 | |
1138 | /* Check XWINT (X, u.opno) == LABEL. */ |
1139 | WIDE_INT_FIELD, |
1140 | |
1141 | /* Check XVECLEN (X, 0) == LABEL. */ |
1142 | VECLEN, |
1143 | |
1144 | /* Check peep2_current_count >= u.min_len. */ |
1145 | PEEP2_COUNT, |
1146 | |
1147 | /* Check XVECLEN (X, 0) >= u.min_len. */ |
1148 | VECLEN_GE, |
1149 | |
1150 | /* Check whether X is a cached const_int with value u.integer. */ |
1151 | SAVED_CONST_INT, |
1152 | |
1153 | /* Check u.predicate.data (X, u.predicate.mode). */ |
1154 | PREDICATE, |
1155 | |
1156 | /* Check rtx_equal_p (X, operands[u.opno]). */ |
1157 | DUPLICATE, |
1158 | |
1159 | /* Check whether X matches pattern u.pattern. */ |
1160 | PATTERN, |
1161 | |
1162 | /* Check whether pnum_clobbers is nonnull (RECOG only). */ |
1163 | HAVE_NUM_CLOBBERS, |
1164 | |
1165 | /* Check whether general C test u.string holds. In general the condition |
1166 | needs access to "insn" and the full operand list. */ |
1167 | C_TEST, |
1168 | |
1169 | /* Execute operands[u.opno] = X. (Always succeeds.) */ |
1170 | SET_OP, |
1171 | |
1172 | /* Accept u.acceptance. Always succeeds for SUBPATTERN, RECOG and SPLIT. |
1173 | May fail for PEEPHOLE2 if the define_peephole2 C code executes FAIL. */ |
1174 | ACCEPT |
1175 | }; |
1176 | |
1177 | /* The position of rtx X in the above description, relative to the |
1178 | incoming instruction "insn". The position is null if the test |
1179 | doesn't take an X as input. */ |
1180 | position *pos; |
1181 | |
1182 | /* Which element of operands[] already contains POS, or -1 if no element |
1183 | is known to hold POS. */ |
1184 | int pos_operand; |
1185 | |
1186 | /* The type of test and its parameters, as described above. */ |
1187 | kind_enum kind; |
1188 | union |
1189 | { |
1190 | int opno; |
1191 | int min_len; |
1192 | struct |
1193 | { |
1194 | bool is_param; |
1195 | int value; |
1196 | } integer; |
1197 | struct |
1198 | { |
1199 | const struct pred_data *data; |
1200 | /* True if the mode is taken from a machine_mode parameter |
1201 | to the routine rather than a constant machine_mode. If true, |
1202 | MODE is the number of the parameter (for an "i%d" format string), |
1203 | otherwise it is the mode itself. */ |
1204 | bool mode_is_param; |
1205 | unsigned int mode; |
1206 | } predicate; |
1207 | pattern_use *pattern; |
1208 | const char *string; |
1209 | acceptance_type acceptance; |
1210 | } u; |
1211 | |
1212 | static rtx_test code (position *); |
1213 | static rtx_test mode (position *); |
1214 | static rtx_test regno_field (position *); |
1215 | static rtx_test subreg_field (position *); |
1216 | static rtx_test int_field (position *, int); |
1217 | static rtx_test wide_int_field (position *, int); |
1218 | static rtx_test veclen (position *); |
1219 | static rtx_test peep2_count (int); |
1220 | static rtx_test veclen_ge (position *, int); |
1221 | static rtx_test predicate (position *, const pred_data *, machine_mode); |
1222 | static rtx_test duplicate (position *, int); |
1223 | static rtx_test pattern (position *, pattern_use *); |
1224 | static rtx_test have_num_clobbers (); |
1225 | static rtx_test c_test (const char *); |
1226 | static rtx_test set_op (position *, int); |
1227 | static rtx_test accept (const acceptance_type &); |
1228 | |
1229 | bool terminal_p () const; |
1230 | bool single_outcome_p () const; |
1231 | |
1232 | private: |
1233 | rtx_test (position *, kind_enum); |
1234 | }; |
1235 | |
1236 | rtx_test::rtx_test () {} |
1237 | |
1238 | rtx_test::rtx_test (position *pos_in, kind_enum kind_in) |
1239 | : pos (pos_in), pos_operand (-1), kind (kind_in) {} |
1240 | |
1241 | rtx_test |
1242 | rtx_test::code (position *pos) |
1243 | { |
1244 | return rtx_test (pos, rtx_test::CODE); |
1245 | } |
1246 | |
1247 | rtx_test |
1248 | rtx_test::mode (position *pos) |
1249 | { |
1250 | return rtx_test (pos, rtx_test::MODE); |
1251 | } |
1252 | |
1253 | rtx_test |
1254 | rtx_test::regno_field (position *pos) |
1255 | { |
1256 | rtx_test res (pos, rtx_test::REGNO_FIELD); |
1257 | return res; |
1258 | } |
1259 | |
1260 | rtx_test |
1261 | rtx_test::subreg_field (position *pos) |
1262 | { |
1263 | rtx_test res (pos, rtx_test::SUBREG_FIELD); |
1264 | return res; |
1265 | } |
1266 | |
1267 | rtx_test |
1268 | rtx_test::int_field (position *pos, int opno) |
1269 | { |
1270 | rtx_test res (pos, rtx_test::INT_FIELD); |
1271 | res.u.opno = opno; |
1272 | return res; |
1273 | } |
1274 | |
1275 | rtx_test |
1276 | rtx_test::wide_int_field (position *pos, int opno) |
1277 | { |
1278 | rtx_test res (pos, rtx_test::WIDE_INT_FIELD); |
1279 | res.u.opno = opno; |
1280 | return res; |
1281 | } |
1282 | |
1283 | rtx_test |
1284 | rtx_test::veclen (position *pos) |
1285 | { |
1286 | return rtx_test (pos, rtx_test::VECLEN); |
1287 | } |
1288 | |
1289 | rtx_test |
1290 | rtx_test::peep2_count (int min_len) |
1291 | { |
1292 | rtx_test res (0, rtx_test::PEEP2_COUNT); |
1293 | res.u.min_len = min_len; |
1294 | return res; |
1295 | } |
1296 | |
1297 | rtx_test |
1298 | rtx_test::veclen_ge (position *pos, int min_len) |
1299 | { |
1300 | rtx_test res (pos, rtx_test::VECLEN_GE); |
1301 | res.u.min_len = min_len; |
1302 | return res; |
1303 | } |
1304 | |
1305 | rtx_test |
1306 | rtx_test::predicate (position *pos, const struct pred_data *data, |
1307 | machine_mode mode) |
1308 | { |
1309 | rtx_test res (pos, rtx_test::PREDICATE); |
1310 | res.u.predicate.data = data; |
1311 | res.u.predicate.mode_is_param = false; |
1312 | res.u.predicate.mode = mode; |
1313 | return res; |
1314 | } |
1315 | |
1316 | rtx_test |
1317 | rtx_test::duplicate (position *pos, int opno) |
1318 | { |
1319 | rtx_test res (pos, rtx_test::DUPLICATE); |
1320 | res.u.opno = opno; |
1321 | return res; |
1322 | } |
1323 | |
1324 | rtx_test |
1325 | rtx_test::pattern (position *pos, pattern_use *pattern) |
1326 | { |
1327 | rtx_test res (pos, rtx_test::PATTERN); |
1328 | res.u.pattern = pattern; |
1329 | return res; |
1330 | } |
1331 | |
1332 | rtx_test |
1333 | rtx_test::have_num_clobbers () |
1334 | { |
1335 | return rtx_test (0, rtx_test::HAVE_NUM_CLOBBERS); |
1336 | } |
1337 | |
1338 | rtx_test |
1339 | rtx_test::c_test (const char *string) |
1340 | { |
1341 | rtx_test res (0, rtx_test::C_TEST); |
1342 | res.u.string = string; |
1343 | return res; |
1344 | } |
1345 | |
1346 | rtx_test |
1347 | rtx_test::set_op (position *pos, int opno) |
1348 | { |
1349 | rtx_test res (pos, rtx_test::SET_OP); |
1350 | res.u.opno = opno; |
1351 | return res; |
1352 | } |
1353 | |
1354 | rtx_test |
1355 | rtx_test::accept (const acceptance_type &acceptance) |
1356 | { |
1357 | rtx_test res (0, rtx_test::ACCEPT); |
1358 | res.u.acceptance = acceptance; |
1359 | return res; |
1360 | } |
1361 | |
1362 | /* Return true if the test represents an unconditionally successful match. */ |
1363 | |
1364 | bool |
1365 | rtx_test::terminal_p () const |
1366 | { |
1367 | return kind == rtx_test::ACCEPT && u.acceptance.type != PEEPHOLE2; |
1368 | } |
1369 | |
1370 | /* Return true if the test is a boolean that is always true. */ |
1371 | |
1372 | bool |
1373 | rtx_test::single_outcome_p () const |
1374 | { |
1375 | return terminal_p () || kind == rtx_test::SET_OP; |
1376 | } |
1377 | |
1378 | bool |
1379 | operator == (const rtx_test &a, const rtx_test &b) |
1380 | { |
1381 | if (a.pos != b.pos || a.kind != b.kind) |
1382 | return false; |
1383 | switch (a.kind) |
1384 | { |
1385 | case rtx_test::CODE: |
1386 | case rtx_test::MODE: |
1387 | case rtx_test::REGNO_FIELD: |
1388 | case rtx_test::SUBREG_FIELD: |
1389 | case rtx_test::VECLEN: |
1390 | case rtx_test::HAVE_NUM_CLOBBERS: |
1391 | return true; |
1392 | |
1393 | case rtx_test::PEEP2_COUNT: |
1394 | case rtx_test::VECLEN_GE: |
1395 | return a.u.min_len == b.u.min_len; |
1396 | |
1397 | case rtx_test::INT_FIELD: |
1398 | case rtx_test::WIDE_INT_FIELD: |
1399 | case rtx_test::DUPLICATE: |
1400 | case rtx_test::SET_OP: |
1401 | return a.u.opno == b.u.opno; |
1402 | |
1403 | case rtx_test::SAVED_CONST_INT: |
1404 | return (a.u.integer.is_param == b.u.integer.is_param |
1405 | && a.u.integer.value == b.u.integer.value); |
1406 | |
1407 | case rtx_test::PREDICATE: |
1408 | return (a.u.predicate.data == b.u.predicate.data |
1409 | && a.u.predicate.mode_is_param == b.u.predicate.mode_is_param |
1410 | && a.u.predicate.mode == b.u.predicate.mode); |
1411 | |
1412 | case rtx_test::PATTERN: |
1413 | return (a.u.pattern->routine == b.u.pattern->routine |
1414 | && a.u.pattern->params == b.u.pattern->params); |
1415 | |
1416 | case rtx_test::C_TEST: |
1417 | return strcmp (s1: a.u.string, s2: b.u.string) == 0; |
1418 | |
1419 | case rtx_test::ACCEPT: |
1420 | return a.u.acceptance == b.u.acceptance; |
1421 | } |
1422 | gcc_unreachable (); |
1423 | } |
1424 | |
1425 | bool |
1426 | operator != (const rtx_test &a, const rtx_test &b) |
1427 | { |
1428 | return !operator == (a, b); |
1429 | } |
1430 | |
1431 | /* A simple set of transition labels. Most transitions have a singleton |
1432 | label, so try to make that case as efficient as possible. */ |
1433 | class int_set : public auto_vec <uint64_t, 1> |
1434 | { |
1435 | public: |
1436 | typedef uint64_t *iterator; |
1437 | |
1438 | int_set (); |
1439 | int_set (uint64_t); |
1440 | int_set (const int_set &); |
1441 | |
1442 | int_set &operator = (const int_set &); |
1443 | |
1444 | iterator begin (); |
1445 | iterator end (); |
1446 | }; |
1447 | |
1448 | int_set::int_set () : auto_vec<uint64_t, 1> () {} |
1449 | |
1450 | int_set::int_set (uint64_t label) : |
1451 | auto_vec<uint64_t, 1> () |
1452 | { |
1453 | safe_push (obj: label); |
1454 | } |
1455 | |
1456 | int_set::int_set (const int_set &other) : |
1457 | auto_vec<uint64_t, 1> () |
1458 | { |
1459 | safe_splice (src: other); |
1460 | } |
1461 | |
1462 | int_set & |
1463 | int_set::operator = (const int_set &other) |
1464 | { |
1465 | truncate (size: 0); |
1466 | safe_splice (src: other); |
1467 | return *this; |
1468 | } |
1469 | |
1470 | int_set::iterator |
1471 | int_set::begin () |
1472 | { |
1473 | return address (); |
1474 | } |
1475 | |
1476 | int_set::iterator |
1477 | int_set::end () |
1478 | { |
1479 | return address () + length (); |
1480 | } |
1481 | |
1482 | bool |
1483 | operator == (const int_set &a, const int_set &b) |
1484 | { |
1485 | if (a.length () != b.length ()) |
1486 | return false; |
1487 | for (unsigned int i = 0; i < a.length (); ++i) |
1488 | if (a[i] != b[i]) |
1489 | return false; |
1490 | return true; |
1491 | } |
1492 | |
1493 | bool |
1494 | operator != (const int_set &a, const int_set &b) |
1495 | { |
1496 | return !operator == (a, b); |
1497 | } |
1498 | |
1499 | class decision; |
1500 | |
1501 | /* Represents a transition between states, dependent on the result of |
1502 | a test T. */ |
1503 | class transition |
1504 | { |
1505 | public: |
1506 | transition (const int_set &, state *, bool); |
1507 | |
1508 | void set_parent (list_head <transition> *); |
1509 | |
1510 | /* Links to other transitions for T. Always null for boolean tests. */ |
1511 | transition *prev, *next; |
1512 | |
1513 | /* The transition should be taken when T has one of these values. |
1514 | E.g. for rtx_test::CODE this is a set of codes, while for booleans like |
1515 | rtx_test::PREDICATE it is always a singleton "true". The labels are |
1516 | sorted in ascending order. */ |
1517 | int_set labels; |
1518 | |
1519 | /* The source decision. */ |
1520 | decision *from; |
1521 | |
1522 | /* The target state. */ |
1523 | state *to; |
1524 | |
1525 | /* True if TO would function correctly even if TEST wasn't performed. |
1526 | E.g. it isn't necessary to check whether GET_MODE (x1) is SImode |
1527 | before calling register_operand (x1, SImode), since register_operand |
1528 | performs its own mode check. However, checking GET_MODE can be a cheap |
1529 | way of disambiguating SImode and DImode register operands. */ |
1530 | bool optional; |
1531 | |
1532 | /* True if LABELS contains parameter numbers rather than constants. |
1533 | E.g. if this is true for a rtx_test::CODE, the label is the number |
1534 | of an rtx_code parameter rather than an rtx_code itself. |
1535 | LABELS is always a singleton when this variable is true. */ |
1536 | bool is_param; |
1537 | }; |
1538 | |
1539 | /* Represents a test and the action that should be taken on the result. |
1540 | If a transition exists for the test outcome, the machine switches |
1541 | to the transition's target state. If no suitable transition exists, |
1542 | the machine either falls through to the next decision or, if there are no |
1543 | more decisions to try, fails the match. */ |
1544 | class decision : public list_head <transition> |
1545 | { |
1546 | public: |
1547 | decision (const rtx_test &); |
1548 | |
1549 | void set_parent (list_head <decision> *s); |
1550 | bool if_statement_p (uint64_t * = 0) const; |
1551 | |
1552 | /* The state to which this decision belongs. */ |
1553 | state *s; |
1554 | |
1555 | /* Links to other decisions in the same state. */ |
1556 | decision *prev, *next; |
1557 | |
1558 | /* The test to perform. */ |
1559 | rtx_test test; |
1560 | }; |
1561 | |
1562 | /* Represents one machine state. For each state the machine tries a list |
1563 | of decisions, in order, and acts on the first match. It fails without |
1564 | further backtracking if no decisions match. */ |
1565 | class state : public list_head <decision> |
1566 | { |
1567 | public: |
1568 | void set_parent (list_head <state> *) {} |
1569 | }; |
1570 | |
1571 | transition::transition (const int_set &labels_in, state *to_in, |
1572 | bool optional_in) |
1573 | : prev (0), next (0), labels (labels_in), from (0), to (to_in), |
1574 | optional (optional_in), is_param (false) {} |
1575 | |
1576 | /* Set the source decision of the transition. */ |
1577 | |
1578 | void |
1579 | transition::set_parent (list_head <transition> *from_in) |
1580 | { |
1581 | from = static_cast <decision *> (from_in); |
1582 | } |
1583 | |
1584 | decision::decision (const rtx_test &test_in) |
1585 | : prev (0), next (0), test (test_in) {} |
1586 | |
1587 | /* Set the state to which this decision belongs. */ |
1588 | |
1589 | void |
1590 | decision::set_parent (list_head <decision> *s_in) |
1591 | { |
1592 | s = static_cast <state *> (s_in); |
1593 | } |
1594 | |
1595 | /* Return true if the decision has a single transition with a single label. |
1596 | If so, return the label in *LABEL if nonnull. */ |
1597 | |
1598 | inline bool |
1599 | decision::if_statement_p (uint64_t *label) const |
1600 | { |
1601 | if (singleton () && first->labels.length () == 1) |
1602 | { |
1603 | if (label) |
1604 | *label = first->labels[0]; |
1605 | return true; |
1606 | } |
1607 | return false; |
1608 | } |
1609 | |
1610 | /* Add to FROM a decision that performs TEST and has a single transition |
1611 | TRANS. */ |
1612 | |
1613 | static void |
1614 | add_decision (state *from, const rtx_test &test, transition *trans) |
1615 | { |
1616 | decision *d = new decision (test); |
1617 | from->push_back (r: d); |
1618 | d->push_back (r: trans); |
1619 | } |
1620 | |
1621 | /* Add a transition from FROM to a new, empty state that is taken |
1622 | when TEST == LABELS. OPTIONAL says whether the new transition |
1623 | should be optional. Return the new state. */ |
1624 | |
1625 | static state * |
1626 | add_decision (state *from, const rtx_test &test, int_set labels, bool optional) |
1627 | { |
1628 | state *to = new state; |
1629 | add_decision (from, test, trans: new transition (labels, to, optional)); |
1630 | return to; |
1631 | } |
1632 | |
1633 | /* Insert a decision before decisions R to make them dependent on |
1634 | TEST == LABELS. OPTIONAL says whether the new transition should be |
1635 | optional. */ |
1636 | |
1637 | static decision * |
1638 | insert_decision_before (state::range r, const rtx_test &test, |
1639 | const int_set &labels, bool optional) |
1640 | { |
1641 | decision *newd = new decision (test); |
1642 | state *news = new state; |
1643 | newd->push_back (r: new transition (labels, news, optional)); |
1644 | r.start->s->replace (oldr: r, newr: newd); |
1645 | news->push_back (r); |
1646 | return newd; |
1647 | } |
1648 | |
1649 | /* Remove any optional transitions from S that turned out not to be useful. */ |
1650 | |
1651 | static void |
1652 | collapse_optional_decisions (state *s) |
1653 | { |
1654 | decision *d = s->first; |
1655 | while (d) |
1656 | { |
1657 | decision *next = d->next; |
1658 | for (transition *trans = d->first; trans; trans = trans->next) |
1659 | collapse_optional_decisions (s: trans->to); |
1660 | /* A decision with a single optional transition doesn't help |
1661 | partition the potential matches and so is unlikely to be |
1662 | worthwhile. In particular, if the decision that performs the |
1663 | test is the last in the state, the best it could do is reject |
1664 | an invalid pattern slightly earlier. If instead the decision |
1665 | is not the last in the state, the condition it tests could hold |
1666 | even for the later decisions in the state. The best it can do |
1667 | is save work in some cases where only the later decisions can |
1668 | succeed. |
1669 | |
1670 | In both cases the optional transition would add extra work to |
1671 | successful matches when the tested condition holds. */ |
1672 | if (transition *trans = d->singleton ()) |
1673 | if (trans->optional) |
1674 | s->replace (oldr: d, newr: trans->to->release ()); |
1675 | d = next; |
1676 | } |
1677 | } |
1678 | |
1679 | /* Try to squash several separate tests into simpler ones. */ |
1680 | |
1681 | static void |
1682 | simplify_tests (state *s) |
1683 | { |
1684 | for (decision *d = s->first; d; d = d->next) |
1685 | { |
1686 | uint64_t label; |
1687 | /* Convert checks for GET_CODE (x) == CONST_INT and XWINT (x, 0) == N |
1688 | into checks for const_int_rtx[N'], if N is suitably small. */ |
1689 | if (d->test.kind == rtx_test::CODE |
1690 | && d->if_statement_p (label: &label) |
1691 | && label == CONST_INT) |
1692 | if (decision *second = d->first->to->singleton ()) |
1693 | if (d->test.pos == second->test.pos |
1694 | && second->test.kind == rtx_test::WIDE_INT_FIELD |
1695 | && second->test.u.opno == 0 |
1696 | && second->if_statement_p (label: &label) |
1697 | && IN_RANGE (int64_t (label), |
1698 | -MAX_SAVED_CONST_INT, MAX_SAVED_CONST_INT)) |
1699 | { |
1700 | d->test.kind = rtx_test::SAVED_CONST_INT; |
1701 | d->test.u.integer.is_param = false; |
1702 | d->test.u.integer.value = label; |
1703 | d->replace (oldr: d->first, newr: second->release ()); |
1704 | d->first->labels[0] = true; |
1705 | } |
1706 | /* If we have a CODE test followed by a PREDICATE test, rely on |
1707 | the predicate to test the code. |
1708 | |
1709 | This case exists for match_operators. We initially treat the |
1710 | CODE test for a match_operator as non-optional so that we can |
1711 | safely move down to its operands. It may turn out that all |
1712 | paths that reach that code test require the same predicate |
1713 | to be true. cse_tests will then put the predicate test in |
1714 | series with the code test. */ |
1715 | if (d->test.kind == rtx_test::CODE) |
1716 | if (transition *trans = d->singleton ()) |
1717 | { |
1718 | state *s = trans->to; |
1719 | while (decision *d2 = s->singleton ()) |
1720 | { |
1721 | if (d->test.pos != d2->test.pos) |
1722 | break; |
1723 | transition *trans2 = d2->singleton (); |
1724 | if (!trans2) |
1725 | break; |
1726 | if (d2->test.kind == rtx_test::PREDICATE) |
1727 | { |
1728 | d->test = d2->test; |
1729 | trans->labels = int_set (true); |
1730 | s->replace (oldr: d2, newr: trans2->to->release ()); |
1731 | break; |
1732 | } |
1733 | s = trans2->to; |
1734 | } |
1735 | } |
1736 | for (transition *trans = d->first; trans; trans = trans->next) |
1737 | simplify_tests (s: trans->to); |
1738 | } |
1739 | } |
1740 | |
1741 | /* Return true if all successful returns passing through D require the |
1742 | condition tested by COMMON to be true. |
1743 | |
1744 | When returning true, add all transitions like COMMON in D to WHERE. |
1745 | WHERE may contain a partial result on failure. */ |
1746 | |
1747 | static bool |
1748 | common_test_p (decision *d, transition *common, vec <transition *> *where) |
1749 | { |
1750 | if (d->test.kind == rtx_test::ACCEPT) |
1751 | /* We found a successful return that didn't require COMMON. */ |
1752 | return false; |
1753 | if (d->test == common->from->test) |
1754 | { |
1755 | transition *trans = d->singleton (); |
1756 | if (!trans |
1757 | || trans->optional != common->optional |
1758 | || trans->labels != common->labels) |
1759 | return false; |
1760 | where->safe_push (obj: trans); |
1761 | return true; |
1762 | } |
1763 | for (transition *trans = d->first; trans; trans = trans->next) |
1764 | for (decision *subd = trans->to->first; subd; subd = subd->next) |
1765 | if (!common_test_p (d: subd, common, where)) |
1766 | return false; |
1767 | return true; |
1768 | } |
1769 | |
1770 | /* Indicates that we have tested GET_CODE (X) for a particular rtx X. */ |
1771 | const unsigned char TESTED_CODE = 1; |
1772 | |
1773 | /* Indicates that we have tested XVECLEN (X, 0) for a particular rtx X. */ |
1774 | const unsigned char TESTED_VECLEN = 2; |
1775 | |
1776 | /* Represents a set of conditions that are known to hold. */ |
1777 | class known_conditions |
1778 | { |
1779 | public: |
1780 | /* A mask of TESTED_ values for each position, indexed by the position's |
1781 | id field. */ |
1782 | auto_vec <unsigned char> position_tests; |
1783 | |
1784 | /* Index N says whether operands[N] has been set. */ |
1785 | auto_vec <bool> set_operands; |
1786 | |
1787 | /* A guranteed lower bound on the value of peep2_current_count. */ |
1788 | int peep2_count; |
1789 | }; |
1790 | |
1791 | /* Return true if TEST can safely be performed at D, where |
1792 | the conditions in KC hold. TEST is known to occur along the |
1793 | first path from D (i.e. always following the first transition |
1794 | of the first decision). Any intervening tests can be used as |
1795 | negative proof that hoisting isn't safe, but only KC can be used |
1796 | as positive proof. */ |
1797 | |
1798 | static bool |
1799 | safe_to_hoist_p (decision *d, const rtx_test &test, known_conditions *kc) |
1800 | { |
1801 | switch (test.kind) |
1802 | { |
1803 | case rtx_test::C_TEST: |
1804 | /* In general, C tests require everything else to have been |
1805 | verified and all operands to have been set up. */ |
1806 | return false; |
1807 | |
1808 | case rtx_test::ACCEPT: |
1809 | /* Don't accept something before all conditions have been tested. */ |
1810 | return false; |
1811 | |
1812 | case rtx_test::PREDICATE: |
1813 | /* Don't move a predicate over a test for VECLEN_GE, since the |
1814 | predicate used in a match_parallel can legitimately expect the |
1815 | length to be checked first. */ |
1816 | for (decision *subd = d; |
1817 | subd->test != test; |
1818 | subd = subd->first->to->first) |
1819 | if (subd->test.pos == test.pos |
1820 | && subd->test.kind == rtx_test::VECLEN_GE) |
1821 | return false; |
1822 | goto any_rtx; |
1823 | |
1824 | case rtx_test::DUPLICATE: |
1825 | /* Don't test for a match_dup until the associated operand has |
1826 | been set. */ |
1827 | if (!kc->set_operands[test.u.opno]) |
1828 | return false; |
1829 | goto any_rtx; |
1830 | |
1831 | case rtx_test::CODE: |
1832 | case rtx_test::MODE: |
1833 | case rtx_test::SAVED_CONST_INT: |
1834 | case rtx_test::SET_OP: |
1835 | any_rtx: |
1836 | /* Check whether it is safe to access the rtx under test. */ |
1837 | switch (test.pos->type) |
1838 | { |
1839 | case POS_PEEP2_INSN: |
1840 | return test.pos->arg < kc->peep2_count; |
1841 | |
1842 | case POS_XEXP: |
1843 | return kc->position_tests[test.pos->base->id] & TESTED_CODE; |
1844 | |
1845 | case POS_XVECEXP0: |
1846 | return kc->position_tests[test.pos->base->id] & TESTED_VECLEN; |
1847 | } |
1848 | gcc_unreachable (); |
1849 | |
1850 | case rtx_test::REGNO_FIELD: |
1851 | case rtx_test::SUBREG_FIELD: |
1852 | case rtx_test::INT_FIELD: |
1853 | case rtx_test::WIDE_INT_FIELD: |
1854 | case rtx_test::VECLEN: |
1855 | case rtx_test::VECLEN_GE: |
1856 | /* These tests access a specific part of an rtx, so are only safe |
1857 | once we know what the rtx is. */ |
1858 | return kc->position_tests[test.pos->id] & TESTED_CODE; |
1859 | |
1860 | case rtx_test::PEEP2_COUNT: |
1861 | case rtx_test::HAVE_NUM_CLOBBERS: |
1862 | /* These tests can be performed anywhere. */ |
1863 | return true; |
1864 | |
1865 | case rtx_test::PATTERN: |
1866 | gcc_unreachable (); |
1867 | } |
1868 | gcc_unreachable (); |
1869 | } |
1870 | |
1871 | /* Look for a transition that is taken by all successful returns from a range |
1872 | of decisions starting at OUTER and that would be better performed by |
1873 | OUTER's state instead. On success, store all instances of that transition |
1874 | in WHERE and return the last decision in the range. The range could |
1875 | just be OUTER, or it could include later decisions as well. |
1876 | |
1877 | WITH_POSITION_P is true if only tests with position POS should be tried, |
1878 | false if any test should be tried. WORTHWHILE_SINGLE_P is true if the |
1879 | result is useful even when the range contains just a single decision |
1880 | with a single transition. KC are the conditions that are known to |
1881 | hold at OUTER. */ |
1882 | |
1883 | static decision * |
1884 | find_common_test (decision *outer, bool with_position_p, |
1885 | position *pos, bool worthwhile_single_p, |
1886 | known_conditions *kc, vec <transition *> *where) |
1887 | { |
1888 | /* After this, WORTHWHILE_SINGLE_P indicates whether a range that contains |
1889 | just a single decision is useful, regardless of the number of |
1890 | transitions it has. */ |
1891 | if (!outer->singleton ()) |
1892 | worthwhile_single_p = true; |
1893 | /* Quick exit if we don't have enough decisions to form a worthwhile |
1894 | range. */ |
1895 | if (!worthwhile_single_p && !outer->next) |
1896 | return 0; |
1897 | /* Follow the first chain down, as one example of a path that needs |
1898 | to contain the common test. */ |
1899 | for (decision *d = outer; d; d = d->first->to->first) |
1900 | { |
1901 | transition *trans = d->singleton (); |
1902 | if (trans |
1903 | && (!with_position_p || d->test.pos == pos) |
1904 | && safe_to_hoist_p (d: outer, test: d->test, kc)) |
1905 | { |
1906 | if (common_test_p (d: outer, common: trans, where)) |
1907 | { |
1908 | if (!outer->next) |
1909 | /* We checked above whether the move is worthwhile. */ |
1910 | return outer; |
1911 | /* See how many decisions in OUTER's chain could reuse |
1912 | the same test. */ |
1913 | decision *outer_end = outer; |
1914 | do |
1915 | { |
1916 | unsigned int length = where->length (); |
1917 | if (!common_test_p (d: outer_end->next, common: trans, where)) |
1918 | { |
1919 | where->truncate (size: length); |
1920 | break; |
1921 | } |
1922 | outer_end = outer_end->next; |
1923 | } |
1924 | while (outer_end->next); |
1925 | /* It is worth moving TRANS if it can be shared by more than |
1926 | one decision. */ |
1927 | if (outer_end != outer || worthwhile_single_p) |
1928 | return outer_end; |
1929 | } |
1930 | where->truncate (size: 0); |
1931 | } |
1932 | } |
1933 | return 0; |
1934 | } |
1935 | |
1936 | /* Try to promote common subtests in S to a single, shared decision. |
1937 | Also try to bunch tests for the same position together. POS is the |
1938 | position of the rtx tested before reaching S. KC are the conditions |
1939 | that are known to hold on entry to S. */ |
1940 | |
1941 | static void |
1942 | cse_tests (position *pos, state *s, known_conditions *kc) |
1943 | { |
1944 | for (decision *d = s->first; d; d = d->next) |
1945 | { |
1946 | auto_vec <transition *, 16> where; |
1947 | if (d->test.pos) |
1948 | { |
1949 | /* Try to find conditions that don't depend on a particular rtx, |
1950 | such as pnum_clobbers != NULL or peep2_current_count >= X. |
1951 | It's usually better to check these conditions as soon as |
1952 | possible, so the change is worthwhile even if there is |
1953 | only one copy of the test. */ |
1954 | decision *endd = find_common_test (outer: d, with_position_p: true, pos: 0, worthwhile_single_p: true, kc, where: &where); |
1955 | if (!endd && d->test.pos != pos) |
1956 | /* Try to find other conditions related to position POS |
1957 | before moving to the new position. Again, this is |
1958 | worthwhile even if there is only one copy of the test, |
1959 | since it means that fewer position variables are live |
1960 | at a given time. */ |
1961 | endd = find_common_test (outer: d, with_position_p: true, pos, worthwhile_single_p: true, kc, where: &where); |
1962 | if (!endd) |
1963 | /* Try to find any condition that is used more than once. */ |
1964 | endd = find_common_test (outer: d, with_position_p: false, pos: 0, worthwhile_single_p: false, kc, where: &where); |
1965 | if (endd) |
1966 | { |
1967 | transition *common = where[0]; |
1968 | /* Replace [D, ENDD] with a test like COMMON. We'll recurse |
1969 | on the common test and see the original D again next time. */ |
1970 | d = insert_decision_before (r: state::range (d, endd), |
1971 | test: common->from->test, |
1972 | labels: common->labels, |
1973 | optional: common->optional); |
1974 | /* Remove the old tests. */ |
1975 | while (!where.is_empty ()) |
1976 | { |
1977 | transition *trans = where.pop (); |
1978 | trans->from->s->replace (oldr: trans->from, newr: trans->to->release ()); |
1979 | } |
1980 | } |
1981 | } |
1982 | |
1983 | /* Make sure that safe_to_hoist_p isn't being overly conservative. |
1984 | It should realize that D's test is safe in the current |
1985 | environment. */ |
1986 | gcc_assert (d->test.kind == rtx_test::C_TEST |
1987 | || d->test.kind == rtx_test::ACCEPT |
1988 | || safe_to_hoist_p (d, d->test, kc)); |
1989 | |
1990 | /* D won't be changed any further by the current optimization. |
1991 | Recurse with the state temporarily updated to include D. */ |
1992 | int prev = 0; |
1993 | switch (d->test.kind) |
1994 | { |
1995 | case rtx_test::CODE: |
1996 | prev = kc->position_tests[d->test.pos->id]; |
1997 | kc->position_tests[d->test.pos->id] |= TESTED_CODE; |
1998 | break; |
1999 | |
2000 | case rtx_test::VECLEN: |
2001 | case rtx_test::VECLEN_GE: |
2002 | prev = kc->position_tests[d->test.pos->id]; |
2003 | kc->position_tests[d->test.pos->id] |= TESTED_VECLEN; |
2004 | break; |
2005 | |
2006 | case rtx_test::SET_OP: |
2007 | prev = kc->set_operands[d->test.u.opno]; |
2008 | gcc_assert (!prev); |
2009 | kc->set_operands[d->test.u.opno] = true; |
2010 | break; |
2011 | |
2012 | case rtx_test::PEEP2_COUNT: |
2013 | prev = kc->peep2_count; |
2014 | kc->peep2_count = MAX (prev, d->test.u.min_len); |
2015 | break; |
2016 | |
2017 | default: |
2018 | break; |
2019 | } |
2020 | for (transition *trans = d->first; trans; trans = trans->next) |
2021 | cse_tests (pos: d->test.pos ? d->test.pos : pos, s: trans->to, kc); |
2022 | switch (d->test.kind) |
2023 | { |
2024 | case rtx_test::CODE: |
2025 | case rtx_test::VECLEN: |
2026 | case rtx_test::VECLEN_GE: |
2027 | kc->position_tests[d->test.pos->id] = prev; |
2028 | break; |
2029 | |
2030 | case rtx_test::SET_OP: |
2031 | kc->set_operands[d->test.u.opno] = prev; |
2032 | break; |
2033 | |
2034 | case rtx_test::PEEP2_COUNT: |
2035 | kc->peep2_count = prev; |
2036 | break; |
2037 | |
2038 | default: |
2039 | break; |
2040 | } |
2041 | } |
2042 | } |
2043 | |
2044 | /* Return the type of value that can be used to parameterize test KIND, |
2045 | or parameter::UNSET if none. */ |
2046 | |
2047 | parameter::type_enum |
2048 | transition_parameter_type (rtx_test::kind_enum kind) |
2049 | { |
2050 | switch (kind) |
2051 | { |
2052 | case rtx_test::CODE: |
2053 | return parameter::CODE; |
2054 | |
2055 | case rtx_test::MODE: |
2056 | return parameter::MODE; |
2057 | |
2058 | case rtx_test::REGNO_FIELD: |
2059 | case rtx_test::SUBREG_FIELD: |
2060 | return parameter::UINT; |
2061 | |
2062 | case rtx_test::INT_FIELD: |
2063 | case rtx_test::VECLEN: |
2064 | case rtx_test::PATTERN: |
2065 | return parameter::INT; |
2066 | |
2067 | case rtx_test::WIDE_INT_FIELD: |
2068 | return parameter::WIDE_INT; |
2069 | |
2070 | case rtx_test::PEEP2_COUNT: |
2071 | case rtx_test::VECLEN_GE: |
2072 | case rtx_test::SAVED_CONST_INT: |
2073 | case rtx_test::PREDICATE: |
2074 | case rtx_test::DUPLICATE: |
2075 | case rtx_test::HAVE_NUM_CLOBBERS: |
2076 | case rtx_test::C_TEST: |
2077 | case rtx_test::SET_OP: |
2078 | case rtx_test::ACCEPT: |
2079 | return parameter::UNSET; |
2080 | } |
2081 | gcc_unreachable (); |
2082 | } |
2083 | |
2084 | /* Initialize the pos_operand fields of each state reachable from S. |
2085 | If OPERAND_POS[ID] >= 0, the position with id ID is stored in |
2086 | operands[OPERAND_POS[ID]] on entry to S. */ |
2087 | |
2088 | static void |
2089 | find_operand_positions (state *s, vec <int> &operand_pos) |
2090 | { |
2091 | for (decision *d = s->first; d; d = d->next) |
2092 | { |
2093 | int this_operand = (d->test.pos ? operand_pos[d->test.pos->id] : -1); |
2094 | if (this_operand >= 0) |
2095 | d->test.pos_operand = this_operand; |
2096 | if (d->test.kind == rtx_test::SET_OP) |
2097 | operand_pos[d->test.pos->id] = d->test.u.opno; |
2098 | for (transition *trans = d->first; trans; trans = trans->next) |
2099 | find_operand_positions (s: trans->to, operand_pos); |
2100 | if (d->test.kind == rtx_test::SET_OP) |
2101 | operand_pos[d->test.pos->id] = this_operand; |
2102 | } |
2103 | } |
2104 | |
2105 | /* Statistics about a matching routine. */ |
2106 | class stats |
2107 | { |
2108 | public: |
2109 | stats (); |
2110 | |
2111 | /* The total number of decisions in the routine, excluding trivial |
2112 | ones that never fail. */ |
2113 | unsigned int num_decisions; |
2114 | |
2115 | /* The number of non-trivial decisions on the longest path through |
2116 | the routine, and the return value that contributes most to that |
2117 | long path. */ |
2118 | unsigned int longest_path; |
2119 | int longest_path_code; |
2120 | |
2121 | /* The maximum number of times that a single call to the routine |
2122 | can backtrack, and the value returned at the end of that path. |
2123 | "Backtracking" here means failing one decision in state and |
2124 | going onto to the next. */ |
2125 | unsigned int longest_backtrack; |
2126 | int longest_backtrack_code; |
2127 | }; |
2128 | |
2129 | stats::stats () |
2130 | : num_decisions (0), longest_path (0), longest_path_code (-1), |
2131 | longest_backtrack (0), longest_backtrack_code (-1) {} |
2132 | |
2133 | /* Return statistics about S. */ |
2134 | |
2135 | static stats |
2136 | get_stats (state *s) |
2137 | { |
2138 | stats for_s; |
2139 | unsigned int longest_path = 0; |
2140 | for (decision *d = s->first; d; d = d->next) |
2141 | { |
2142 | /* Work out the statistics for D. */ |
2143 | stats for_d; |
2144 | for (transition *trans = d->first; trans; trans = trans->next) |
2145 | { |
2146 | stats for_trans = get_stats (s: trans->to); |
2147 | for_d.num_decisions += for_trans.num_decisions; |
2148 | /* Each transition is mutually-exclusive, so just pick the |
2149 | longest of the individual paths. */ |
2150 | if (for_d.longest_path <= for_trans.longest_path) |
2151 | { |
2152 | for_d.longest_path = for_trans.longest_path; |
2153 | for_d.longest_path_code = for_trans.longest_path_code; |
2154 | } |
2155 | /* Likewise for backtracking. */ |
2156 | if (for_d.longest_backtrack <= for_trans.longest_backtrack) |
2157 | { |
2158 | for_d.longest_backtrack = for_trans.longest_backtrack; |
2159 | for_d.longest_backtrack_code = for_trans.longest_backtrack_code; |
2160 | } |
2161 | } |
2162 | |
2163 | /* Account for D's test in its statistics. */ |
2164 | if (!d->test.single_outcome_p ()) |
2165 | { |
2166 | for_d.num_decisions += 1; |
2167 | for_d.longest_path += 1; |
2168 | } |
2169 | if (d->test.kind == rtx_test::ACCEPT) |
2170 | { |
2171 | for_d.longest_path_code = d->test.u.acceptance.u.full.code; |
2172 | for_d.longest_backtrack_code = d->test.u.acceptance.u.full.code; |
2173 | } |
2174 | |
2175 | /* Keep a running count of the number of backtracks. */ |
2176 | if (d->prev) |
2177 | for_s.longest_backtrack += 1; |
2178 | |
2179 | /* Accumulate D's statistics into S's. */ |
2180 | for_s.num_decisions += for_d.num_decisions; |
2181 | for_s.longest_path += for_d.longest_path; |
2182 | for_s.longest_backtrack += for_d.longest_backtrack; |
2183 | |
2184 | /* Use the code from the decision with the longest individual path, |
2185 | since that's more likely to be useful if trying to make the |
2186 | path shorter. In the event of a tie, pick the later decision, |
2187 | since that's closer to the end of the path. */ |
2188 | if (longest_path <= for_d.longest_path) |
2189 | { |
2190 | longest_path = for_d.longest_path; |
2191 | for_s.longest_path_code = for_d.longest_path_code; |
2192 | } |
2193 | |
2194 | /* Later decisions in a state are necessarily in a longer backtrack |
2195 | than earlier decisions. */ |
2196 | for_s.longest_backtrack_code = for_d.longest_backtrack_code; |
2197 | } |
2198 | return for_s; |
2199 | } |
2200 | |
2201 | /* Optimize ROOT. Use TYPE to describe ROOT in status messages. */ |
2202 | |
2203 | static void |
2204 | optimize_subroutine_group (const char *type, state *root) |
2205 | { |
2206 | /* Remove optional transitions that turned out not to be worthwhile. */ |
2207 | if (collapse_optional_decisions_p) |
2208 | collapse_optional_decisions (s: root); |
2209 | |
2210 | /* Try to remove duplicated tests and to rearrange tests into a more |
2211 | logical order. */ |
2212 | if (cse_tests_p) |
2213 | { |
2214 | known_conditions kc; |
2215 | kc.position_tests.safe_grow_cleared (len: num_positions, exact: true); |
2216 | kc.set_operands.safe_grow_cleared (len: num_operands, exact: true); |
2217 | kc.peep2_count = 1; |
2218 | cse_tests (pos: &root_pos, s: root, kc: &kc); |
2219 | } |
2220 | |
2221 | /* Try to simplify two or more tests into one. */ |
2222 | if (simplify_tests_p) |
2223 | simplify_tests (s: root); |
2224 | |
2225 | /* Try to use operands[] instead of xN variables. */ |
2226 | if (use_operand_variables_p) |
2227 | { |
2228 | auto_vec <int> operand_pos (num_positions); |
2229 | for (unsigned int i = 0; i < num_positions; ++i) |
2230 | operand_pos.quick_push (obj: -1); |
2231 | find_operand_positions (s: root, operand_pos); |
2232 | } |
2233 | |
2234 | /* Print a summary of the new state. */ |
2235 | stats st = get_stats (s: root); |
2236 | fprintf (stderr, format: "Statistics for %s:\n" , type); |
2237 | fprintf (stderr, format: " Number of decisions: %6d\n" , st.num_decisions); |
2238 | fprintf (stderr, format: " longest path: %6d (code: %6d)\n" , |
2239 | st.longest_path, st.longest_path_code); |
2240 | fprintf (stderr, format: " longest backtrack: %6d (code: %6d)\n" , |
2241 | st.longest_backtrack, st.longest_backtrack_code); |
2242 | } |
2243 | |
2244 | class merge_pattern_info; |
2245 | |
2246 | /* Represents a transition from one pattern to another. */ |
2247 | class merge_pattern_transition |
2248 | { |
2249 | public: |
2250 | merge_pattern_transition (merge_pattern_info *); |
2251 | |
2252 | /* The target pattern. */ |
2253 | merge_pattern_info *to; |
2254 | |
2255 | /* The parameters that the source pattern passes to the target pattern. |
2256 | "parameter (TYPE, true, I)" represents parameter I of the source |
2257 | pattern. */ |
2258 | auto_vec <parameter, MAX_PATTERN_PARAMS> params; |
2259 | }; |
2260 | |
2261 | merge_pattern_transition::merge_pattern_transition (merge_pattern_info *to_in) |
2262 | : to (to_in) |
2263 | { |
2264 | } |
2265 | |
2266 | /* Represents a pattern that can might match several states. The pattern |
2267 | may replace parts of the test with a parameter value. It may also |
2268 | replace transition labels with parameters. */ |
2269 | class merge_pattern_info |
2270 | { |
2271 | public: |
2272 | merge_pattern_info (unsigned int); |
2273 | |
2274 | /* If PARAM_TEST_P, the state's singleton test should be generalized |
2275 | to use the runtime value of PARAMS[PARAM_TEST]. */ |
2276 | unsigned int param_test : 8; |
2277 | |
2278 | /* If PARAM_TRANSITION_P, the state's single transition label should |
2279 | be replaced by the runtime value of PARAMS[PARAM_TRANSITION]. */ |
2280 | unsigned int param_transition : 8; |
2281 | |
2282 | /* True if we have decided to generalize the root decision's test, |
2283 | as per PARAM_TEST. */ |
2284 | unsigned int param_test_p : 1; |
2285 | |
2286 | /* Likewise for the root decision's transition, as per PARAM_TRANSITION. */ |
2287 | unsigned int param_transition_p : 1; |
2288 | |
2289 | /* True if the contents of the structure are completely filled in. */ |
2290 | unsigned int complete_p : 1; |
2291 | |
2292 | /* The number of pseudo-statements in the pattern. Used to decide |
2293 | whether it's big enough to break out into a subroutine. */ |
2294 | unsigned int num_statements; |
2295 | |
2296 | /* The number of states that use this pattern. */ |
2297 | unsigned int num_users; |
2298 | |
2299 | /* The number of distinct success values that the pattern returns. */ |
2300 | unsigned int num_results; |
2301 | |
2302 | /* This array has one element for each runtime parameter to the pattern. |
2303 | PARAMS[I] gives the default value of parameter I, which is always |
2304 | constant. |
2305 | |
2306 | These default parameters are used in cases where we match the |
2307 | pattern against some state S1, then add more parameters while |
2308 | matching against some state S2. S1 is then left passing fewer |
2309 | parameters than S2. The array gives us enough informatino to |
2310 | construct a full parameter list for S1 (see update_parameters). |
2311 | |
2312 | If we decide to create a subroutine for this pattern, |
2313 | PARAMS[I].type determines the C type of parameter I. */ |
2314 | auto_vec <parameter, MAX_PATTERN_PARAMS> params; |
2315 | |
2316 | /* All states that match this pattern must have the same number of |
2317 | transitions. TRANSITIONS[I] describes the subpattern for transition |
2318 | number I; it is null if transition I represents a successful return |
2319 | from the pattern. */ |
2320 | auto_vec <merge_pattern_transition *, 1> transitions; |
2321 | |
2322 | /* The routine associated with the pattern, or null if we haven't generated |
2323 | one yet. */ |
2324 | pattern_routine *routine; |
2325 | }; |
2326 | |
2327 | merge_pattern_info::merge_pattern_info (unsigned int num_transitions) |
2328 | : param_test (0), |
2329 | param_transition (0), |
2330 | param_test_p (false), |
2331 | param_transition_p (false), |
2332 | complete_p (false), |
2333 | num_statements (0), |
2334 | num_users (0), |
2335 | num_results (0), |
2336 | routine (0) |
2337 | { |
2338 | transitions.safe_grow_cleared (len: num_transitions, exact: true); |
2339 | } |
2340 | |
2341 | /* Describes one way of matching a particular state to a particular |
2342 | pattern. */ |
2343 | class merge_state_result |
2344 | { |
2345 | public: |
2346 | merge_state_result (merge_pattern_info *, position *, merge_state_result *); |
2347 | |
2348 | /* A pattern that matches the state. */ |
2349 | merge_pattern_info *pattern; |
2350 | |
2351 | /* If we decide to use this match and create a subroutine for PATTERN, |
2352 | the state should pass the rtx at position ROOT to the pattern's |
2353 | rtx parameter. A null root means that the pattern doesn't need |
2354 | an rtx parameter; all the rtxes it matches come from elsewhere. */ |
2355 | position *root; |
2356 | |
2357 | /* The parameters that should be passed to PATTERN for this state. |
2358 | If the array is shorter than PATTERN->params, the missing entries |
2359 | should be taken from the corresponding element of PATTERN->params. */ |
2360 | auto_vec <parameter, MAX_PATTERN_PARAMS> params; |
2361 | |
2362 | /* An earlier match for the same state, or null if none. Patterns |
2363 | matched by earlier entries are smaller than PATTERN. */ |
2364 | merge_state_result *prev; |
2365 | }; |
2366 | |
2367 | merge_state_result::merge_state_result (merge_pattern_info *pattern_in, |
2368 | position *root_in, |
2369 | merge_state_result *prev_in) |
2370 | : pattern (pattern_in), root (root_in), prev (prev_in) |
2371 | {} |
2372 | |
2373 | /* Information about a state, used while trying to match it against |
2374 | a pattern. */ |
2375 | class merge_state_info |
2376 | { |
2377 | public: |
2378 | merge_state_info (state *); |
2379 | |
2380 | /* The state itself. */ |
2381 | state *s; |
2382 | |
2383 | /* Index I gives information about the target of transition I. */ |
2384 | merge_state_info *to_states; |
2385 | |
2386 | /* The number of transitions in S. */ |
2387 | unsigned int num_transitions; |
2388 | |
2389 | /* True if the state has been deleted in favor of a call to a |
2390 | pattern routine. */ |
2391 | bool merged_p; |
2392 | |
2393 | /* The previous state that might be a merge candidate for S, or null |
2394 | if no previous states could be merged with S. */ |
2395 | merge_state_info *prev_same_test; |
2396 | |
2397 | /* A list of pattern matches for this state. */ |
2398 | merge_state_result *res; |
2399 | }; |
2400 | |
2401 | merge_state_info::merge_state_info (state *s_in) |
2402 | : s (s_in), |
2403 | to_states (0), |
2404 | num_transitions (0), |
2405 | merged_p (false), |
2406 | prev_same_test (0), |
2407 | res (0) {} |
2408 | |
2409 | /* True if PAT would be useful as a subroutine. */ |
2410 | |
2411 | static bool |
2412 | useful_pattern_p (merge_pattern_info *pat) |
2413 | { |
2414 | return pat->num_statements >= MIN_COMBINE_COST; |
2415 | } |
2416 | |
2417 | /* PAT2 is a subpattern of PAT1. Return true if PAT2 should be inlined |
2418 | into PAT1's C routine. */ |
2419 | |
2420 | static bool |
2421 | same_pattern_p (merge_pattern_info *pat1, merge_pattern_info *pat2) |
2422 | { |
2423 | return pat1->num_users == pat2->num_users || !useful_pattern_p (pat: pat2); |
2424 | } |
2425 | |
2426 | /* PAT was previously matched against SINFO based on tentative matches |
2427 | for the target states of SINFO's state. Return true if the match |
2428 | still holds; that is, if the target states of SINFO's state still |
2429 | match the corresponding transitions of PAT. */ |
2430 | |
2431 | static bool |
2432 | valid_result_p (merge_pattern_info *pat, merge_state_info *sinfo) |
2433 | { |
2434 | for (unsigned int j = 0; j < sinfo->num_transitions; ++j) |
2435 | if (merge_pattern_transition *ptrans = pat->transitions[j]) |
2436 | { |
2437 | merge_state_result *to_res = sinfo->to_states[j].res; |
2438 | if (!to_res || to_res->pattern != ptrans->to) |
2439 | return false; |
2440 | } |
2441 | return true; |
2442 | } |
2443 | |
2444 | /* Remove any matches that are no longer valid from the head of SINFO's |
2445 | list of matches. */ |
2446 | |
2447 | static void |
2448 | prune_invalid_results (merge_state_info *sinfo) |
2449 | { |
2450 | while (sinfo->res && !valid_result_p (pat: sinfo->res->pattern, sinfo)) |
2451 | { |
2452 | sinfo->res = sinfo->res->prev; |
2453 | gcc_assert (sinfo->res); |
2454 | } |
2455 | } |
2456 | |
2457 | /* Return true if PAT represents the biggest posssible match for SINFO; |
2458 | that is, if the next action of SINFO's state on return from PAT will |
2459 | be something that cannot be merged with any other state. */ |
2460 | |
2461 | static bool |
2462 | complete_result_p (merge_pattern_info *pat, merge_state_info *sinfo) |
2463 | { |
2464 | for (unsigned int j = 0; j < sinfo->num_transitions; ++j) |
2465 | if (sinfo->to_states[j].res && !pat->transitions[j]) |
2466 | return false; |
2467 | return true; |
2468 | } |
2469 | |
2470 | /* Update TO for any parameters that have been added to FROM since TO |
2471 | was last set. The extra parameters in FROM will be constants or |
2472 | instructions to duplicate earlier parameters. */ |
2473 | |
2474 | static void |
2475 | update_parameters (vec <parameter> &to, const vec <parameter> &from) |
2476 | { |
2477 | for (unsigned int i = to.length (); i < from.length (); ++i) |
2478 | to.quick_push (obj: from[i]); |
2479 | } |
2480 | |
2481 | /* Return true if A and B can be tested by a single test. If the test |
2482 | can be parameterised, store the parameter value for A in *PARAMA and |
2483 | the parameter value for B in *PARAMB, otherwise leave PARAMA and |
2484 | PARAMB alone. */ |
2485 | |
2486 | static bool |
2487 | compatible_tests_p (const rtx_test &a, const rtx_test &b, |
2488 | parameter *parama, parameter *paramb) |
2489 | { |
2490 | if (a.kind != b.kind) |
2491 | return false; |
2492 | switch (a.kind) |
2493 | { |
2494 | case rtx_test::PREDICATE: |
2495 | if (a.u.predicate.data != b.u.predicate.data) |
2496 | return false; |
2497 | *parama = parameter (parameter::MODE, false, a.u.predicate.mode); |
2498 | *paramb = parameter (parameter::MODE, false, b.u.predicate.mode); |
2499 | return true; |
2500 | |
2501 | case rtx_test::SAVED_CONST_INT: |
2502 | *parama = parameter (parameter::INT, false, a.u.integer.value); |
2503 | *paramb = parameter (parameter::INT, false, b.u.integer.value); |
2504 | return true; |
2505 | |
2506 | default: |
2507 | return a == b; |
2508 | } |
2509 | } |
2510 | |
2511 | /* PARAMS is an array of the parameters that a state is going to pass |
2512 | to a pattern routine. It is still incomplete; index I has a kind of |
2513 | parameter::UNSET if we don't yet know what the state will pass |
2514 | as parameter I. Try to make parameter ID equal VALUE, returning |
2515 | true on success. */ |
2516 | |
2517 | static bool |
2518 | set_parameter (vec <parameter> ¶ms, unsigned int id, |
2519 | const parameter &value) |
2520 | { |
2521 | if (params[id].type == parameter::UNSET) |
2522 | { |
2523 | if (force_unique_params_p) |
2524 | for (unsigned int i = 0; i < params.length (); ++i) |
2525 | if (params[i] == value) |
2526 | return false; |
2527 | params[id] = value; |
2528 | return true; |
2529 | } |
2530 | return params[id] == value; |
2531 | } |
2532 | |
2533 | /* PARAMS2 is the "params" array for a pattern and PARAMS1 is the |
2534 | set of parameters that a particular state is going to pass to |
2535 | that pattern. |
2536 | |
2537 | Try to extend PARAMS1 and PARAMS2 so that there is a parameter |
2538 | that is equal to PARAM1 for the state and has a default value of |
2539 | PARAM2. Parameters beginning at START were added as part of the |
2540 | same match and so may be reused. */ |
2541 | |
2542 | static bool |
2543 | add_parameter (vec <parameter> ¶ms1, vec <parameter> ¶ms2, |
2544 | const parameter ¶m1, const parameter ¶m2, |
2545 | unsigned int start, unsigned int *res) |
2546 | { |
2547 | gcc_assert (params1.length () == params2.length ()); |
2548 | gcc_assert (!param1.is_param && !param2.is_param); |
2549 | |
2550 | for (unsigned int i = start; i < params2.length (); ++i) |
2551 | if (params1[i] == param1 && params2[i] == param2) |
2552 | { |
2553 | *res = i; |
2554 | return true; |
2555 | } |
2556 | |
2557 | if (force_unique_params_p) |
2558 | for (unsigned int i = 0; i < params2.length (); ++i) |
2559 | if (params1[i] == param1 || params2[i] == param2) |
2560 | return false; |
2561 | |
2562 | if (params2.length () >= MAX_PATTERN_PARAMS) |
2563 | return false; |
2564 | |
2565 | *res = params2.length (); |
2566 | params1.quick_push (obj: param1); |
2567 | params2.quick_push (obj: param2); |
2568 | return true; |
2569 | } |
2570 | |
2571 | /* If *ROOTA is nonnull, return true if the same sequence of steps are |
2572 | required to reach A from *ROOTA as to reach B from ROOTB. If *ROOTA |
2573 | is null, update it if necessary in order to make the condition hold. */ |
2574 | |
2575 | static bool |
2576 | merge_relative_positions (position **roota, position *a, |
2577 | position *rootb, position *b) |
2578 | { |
2579 | if (!relative_patterns_p) |
2580 | { |
2581 | if (a != b) |
2582 | return false; |
2583 | if (!*roota) |
2584 | { |
2585 | *roota = rootb; |
2586 | return true; |
2587 | } |
2588 | return *roota == rootb; |
2589 | } |
2590 | /* If B does not belong to the same instruction as ROOTB, we don't |
2591 | start with ROOTB but instead start with a call to peep2_next_insn. |
2592 | In that case the sequences for B and A are identical iff B and A |
2593 | are themselves identical. */ |
2594 | if (rootb->insn_id != b->insn_id) |
2595 | return a == b; |
2596 | while (rootb != b) |
2597 | { |
2598 | if (!a || b->type != a->type || b->arg != a->arg) |
2599 | return false; |
2600 | b = b->base; |
2601 | a = a->base; |
2602 | } |
2603 | if (!*roota) |
2604 | *roota = a; |
2605 | return *roota == a; |
2606 | } |
2607 | |
2608 | /* A hasher of states that treats two states as "equal" if they might be |
2609 | merged (but trying to be more discriminating than "return true"). */ |
2610 | struct test_pattern_hasher : nofree_ptr_hash <merge_state_info> |
2611 | { |
2612 | static inline hashval_t hash (const value_type &); |
2613 | static inline bool equal (const value_type &, const compare_type &); |
2614 | }; |
2615 | |
2616 | hashval_t |
2617 | test_pattern_hasher::hash (merge_state_info *const &sinfo) |
2618 | { |
2619 | inchash::hash h; |
2620 | decision *d = sinfo->s->singleton (); |
2621 | h.add_int (v: d->test.pos_operand + 1); |
2622 | if (!relative_patterns_p) |
2623 | h.add_int (v: d->test.pos ? d->test.pos->id + 1 : 0); |
2624 | h.add_int (v: d->test.kind); |
2625 | h.add_int (v: sinfo->num_transitions); |
2626 | return h.end (); |
2627 | } |
2628 | |
2629 | bool |
2630 | test_pattern_hasher::equal (merge_state_info *const &sinfo1, |
2631 | merge_state_info *const &sinfo2) |
2632 | { |
2633 | decision *d1 = sinfo1->s->singleton (); |
2634 | decision *d2 = sinfo2->s->singleton (); |
2635 | gcc_assert (d1 && d2); |
2636 | |
2637 | parameter new_param1, new_param2; |
2638 | return (d1->test.pos_operand == d2->test.pos_operand |
2639 | && (relative_patterns_p || d1->test.pos == d2->test.pos) |
2640 | && compatible_tests_p (a: d1->test, b: d2->test, parama: &new_param1, paramb: &new_param2) |
2641 | && sinfo1->num_transitions == sinfo2->num_transitions); |
2642 | } |
2643 | |
2644 | /* Try to make the state described by SINFO1 use the same pattern as the |
2645 | state described by SINFO2. Return true on success. |
2646 | |
2647 | SINFO1 and SINFO2 are known to have the same hash value. */ |
2648 | |
2649 | static bool |
2650 | merge_patterns (merge_state_info *sinfo1, merge_state_info *sinfo2) |
2651 | { |
2652 | merge_state_result *res2 = sinfo2->res; |
2653 | merge_pattern_info *pat = res2->pattern; |
2654 | |
2655 | /* Write to temporary arrays while matching, in case we have to abort |
2656 | half way through. */ |
2657 | auto_vec <parameter, MAX_PATTERN_PARAMS> params1; |
2658 | auto_vec <parameter, MAX_PATTERN_PARAMS> params2; |
2659 | params1.quick_grow_cleared (len: pat->params.length ()); |
2660 | params2.splice (src: pat->params); |
2661 | unsigned int start_param = params2.length (); |
2662 | |
2663 | /* An array for recording changes to PAT->transitions[?].params. |
2664 | All changes involve replacing a constant parameter with some |
2665 | PAT->params[N], where N is the second element of the pending_param. */ |
2666 | typedef std::pair <parameter *, unsigned int> pending_param; |
2667 | auto_vec <pending_param, 32> pending_params; |
2668 | |
2669 | decision *d1 = sinfo1->s->singleton (); |
2670 | decision *d2 = sinfo2->s->singleton (); |
2671 | gcc_assert (d1 && d2); |
2672 | |
2673 | /* If D2 tests a position, SINFO1's root relative to D1 is the same |
2674 | as SINFO2's root relative to D2. */ |
2675 | position *root1 = 0; |
2676 | position *root2 = res2->root; |
2677 | if (d2->test.pos_operand < 0 |
2678 | && d1->test.pos |
2679 | && !merge_relative_positions (roota: &root1, a: d1->test.pos, |
2680 | rootb: root2, b: d2->test.pos)) |
2681 | return false; |
2682 | |
2683 | /* Check whether the patterns have the same shape. */ |
2684 | unsigned int num_transitions = sinfo1->num_transitions; |
2685 | gcc_assert (num_transitions == sinfo2->num_transitions); |
2686 | for (unsigned int i = 0; i < num_transitions; ++i) |
2687 | if (merge_pattern_transition *ptrans = pat->transitions[i]) |
2688 | { |
2689 | merge_state_result *to1_res = sinfo1->to_states[i].res; |
2690 | merge_state_result *to2_res = sinfo2->to_states[i].res; |
2691 | merge_pattern_info *to_pat = ptrans->to; |
2692 | gcc_assert (to2_res && to2_res->pattern == to_pat); |
2693 | if (!to1_res || to1_res->pattern != to_pat) |
2694 | return false; |
2695 | if (to2_res->root |
2696 | && !merge_relative_positions (roota: &root1, a: to1_res->root, |
2697 | rootb: root2, b: to2_res->root)) |
2698 | return false; |
2699 | /* Match the parameters that TO1_RES passes to TO_PAT with the |
2700 | parameters that PAT passes to TO_PAT. */ |
2701 | update_parameters (to&: to1_res->params, from: to_pat->params); |
2702 | for (unsigned int j = 0; j < to1_res->params.length (); ++j) |
2703 | { |
2704 | const parameter ¶m1 = to1_res->params[j]; |
2705 | const parameter ¶m2 = ptrans->params[j]; |
2706 | gcc_assert (!param1.is_param); |
2707 | if (param2.is_param) |
2708 | { |
2709 | if (!set_parameter (params&: params1, id: param2.value, value: param1)) |
2710 | return false; |
2711 | } |
2712 | else if (param1 != param2) |
2713 | { |
2714 | unsigned int id; |
2715 | if (!add_parameter (params1, params2, |
2716 | param1, param2, start: start_param, res: &id)) |
2717 | return false; |
2718 | /* Record that PAT should now pass parameter ID to TO_PAT, |
2719 | instead of the current contents of *PARAM2. We only |
2720 | make the change if the rest of the match succeeds. */ |
2721 | pending_params.safe_push |
2722 | (obj: pending_param (&ptrans->params[j], id)); |
2723 | } |
2724 | } |
2725 | } |
2726 | |
2727 | unsigned int param_test = pat->param_test; |
2728 | unsigned int param_transition = pat->param_transition; |
2729 | bool param_test_p = pat->param_test_p; |
2730 | bool param_transition_p = pat->param_transition_p; |
2731 | |
2732 | /* If the tests don't match exactly, try to parameterize them. */ |
2733 | parameter new_param1, new_param2; |
2734 | if (!compatible_tests_p (a: d1->test, b: d2->test, parama: &new_param1, paramb: &new_param2)) |
2735 | gcc_unreachable (); |
2736 | if (new_param1.type != parameter::UNSET) |
2737 | { |
2738 | /* If the test has not already been parameterized, all existing |
2739 | matches use constant NEW_PARAM2. */ |
2740 | if (param_test_p) |
2741 | { |
2742 | if (!set_parameter (params&: params1, id: param_test, value: new_param1)) |
2743 | return false; |
2744 | } |
2745 | else if (new_param1 != new_param2) |
2746 | { |
2747 | if (!add_parameter (params1, params2, param1: new_param1, param2: new_param2, |
2748 | start: start_param, res: ¶m_test)) |
2749 | return false; |
2750 | param_test_p = true; |
2751 | } |
2752 | } |
2753 | |
2754 | /* Match the transitions. */ |
2755 | transition *trans1 = d1->first; |
2756 | transition *trans2 = d2->first; |
2757 | for (unsigned int i = 0; i < num_transitions; ++i) |
2758 | { |
2759 | if (param_transition_p || trans1->labels != trans2->labels) |
2760 | { |
2761 | /* We can only generalize a single transition with a single |
2762 | label. */ |
2763 | if (num_transitions != 1 |
2764 | || trans1->labels.length () != 1 |
2765 | || trans2->labels.length () != 1) |
2766 | return false; |
2767 | |
2768 | /* Although we can match wide-int fields, in practice it leads |
2769 | to some odd results for const_vectors. We end up |
2770 | parameterizing the first N const_ints of the vector |
2771 | and then (once we reach the maximum number of parameters) |
2772 | we go on to match the other elements exactly. */ |
2773 | if (d1->test.kind == rtx_test::WIDE_INT_FIELD) |
2774 | return false; |
2775 | |
2776 | /* See whether the label has a generalizable type. */ |
2777 | parameter::type_enum param_type |
2778 | = transition_parameter_type (kind: d1->test.kind); |
2779 | if (param_type == parameter::UNSET) |
2780 | return false; |
2781 | |
2782 | /* Match the labels using parameters. */ |
2783 | new_param1 = parameter (param_type, false, trans1->labels[0]); |
2784 | if (param_transition_p) |
2785 | { |
2786 | if (!set_parameter (params&: params1, id: param_transition, value: new_param1)) |
2787 | return false; |
2788 | } |
2789 | else |
2790 | { |
2791 | new_param2 = parameter (param_type, false, trans2->labels[0]); |
2792 | if (!add_parameter (params1, params2, param1: new_param1, param2: new_param2, |
2793 | start: start_param, res: ¶m_transition)) |
2794 | return false; |
2795 | param_transition_p = true; |
2796 | } |
2797 | } |
2798 | trans1 = trans1->next; |
2799 | trans2 = trans2->next; |
2800 | } |
2801 | |
2802 | /* Set any unset parameters to their default values. This occurs if some |
2803 | other state needed something to be parameterized in order to match SINFO2, |
2804 | but SINFO1 on its own does not. */ |
2805 | for (unsigned int i = 0; i < params1.length (); ++i) |
2806 | if (params1[i].type == parameter::UNSET) |
2807 | params1[i] = params2[i]; |
2808 | |
2809 | /* The match was successful. Commit all pending changes to PAT. */ |
2810 | update_parameters (to&: pat->params, from: params2); |
2811 | { |
2812 | pending_param *pp; |
2813 | unsigned int i; |
2814 | FOR_EACH_VEC_ELT (pending_params, i, pp) |
2815 | *pp->first = parameter (pp->first->type, true, pp->second); |
2816 | } |
2817 | pat->param_test = param_test; |
2818 | pat->param_transition = param_transition; |
2819 | pat->param_test_p = param_test_p; |
2820 | pat->param_transition_p = param_transition_p; |
2821 | |
2822 | /* Record the match of SINFO1. */ |
2823 | merge_state_result *new_res1 = new merge_state_result (pat, root1, |
2824 | sinfo1->res); |
2825 | new_res1->params.splice (src: params1); |
2826 | sinfo1->res = new_res1; |
2827 | return true; |
2828 | } |
2829 | |
2830 | /* The number of states that were removed by calling pattern routines. */ |
2831 | static unsigned int pattern_use_states; |
2832 | |
2833 | /* The number of states used while defining pattern routines. */ |
2834 | static unsigned int pattern_def_states; |
2835 | |
2836 | /* Information used while constructing a use or definition of a pattern |
2837 | routine. */ |
2838 | struct create_pattern_info |
2839 | { |
2840 | /* The routine itself. */ |
2841 | pattern_routine *routine; |
2842 | |
2843 | /* The first unclaimed return value for this particular use or definition. |
2844 | We walk the substates of uses and definitions in the same order |
2845 | so each return value always refers to the same position within |
2846 | the pattern. */ |
2847 | unsigned int next_result; |
2848 | }; |
2849 | |
2850 | static void populate_pattern_routine (create_pattern_info *, |
2851 | merge_state_info *, state *, |
2852 | const vec <parameter> &); |
2853 | |
2854 | /* SINFO matches a pattern for which we've decided to create a C routine. |
2855 | Return a decision that performs a call to the pattern routine, |
2856 | but leave the caller to add the transitions to it. Initialize CPI |
2857 | for this purpose. Also create a definition for the pattern routine, |
2858 | if it doesn't already have one. |
2859 | |
2860 | PARAMS are the parameters that SINFO passes to its pattern. */ |
2861 | |
2862 | static decision * |
2863 | init_pattern_use (create_pattern_info *cpi, merge_state_info *sinfo, |
2864 | const vec <parameter> ¶ms) |
2865 | { |
2866 | state *s = sinfo->s; |
2867 | merge_state_result *res = sinfo->res; |
2868 | merge_pattern_info *pat = res->pattern; |
2869 | cpi->routine = pat->routine; |
2870 | if (!cpi->routine) |
2871 | { |
2872 | /* We haven't defined the pattern routine yet, so create |
2873 | a definition now. */ |
2874 | pattern_routine *routine = new pattern_routine; |
2875 | pat->routine = routine; |
2876 | cpi->routine = routine; |
2877 | routine->s = new state; |
2878 | routine->insn_p = false; |
2879 | routine->pnum_clobbers_p = false; |
2880 | |
2881 | /* Create an "idempotent" mapping of parameter I to parameter I. |
2882 | Also record the C type of each parameter to the routine. */ |
2883 | auto_vec <parameter, MAX_PATTERN_PARAMS> def_params; |
2884 | for (unsigned int i = 0; i < pat->params.length (); ++i) |
2885 | { |
2886 | def_params.quick_push (obj: parameter (pat->params[i].type, true, i)); |
2887 | routine->param_types.quick_push (obj: pat->params[i].type); |
2888 | } |
2889 | |
2890 | /* Any of the states that match the pattern could be used to |
2891 | create the routine definition. We might as well use SINFO |
2892 | since it's already to hand. This means that all positions |
2893 | in the definition will be relative to RES->root. */ |
2894 | routine->pos = res->root; |
2895 | cpi->next_result = 0; |
2896 | populate_pattern_routine (cpi, sinfo, routine->s, def_params); |
2897 | gcc_assert (cpi->next_result == pat->num_results); |
2898 | |
2899 | /* Add the routine to the global list, after the subroutines |
2900 | that it calls. */ |
2901 | routine->pattern_id = patterns.length (); |
2902 | patterns.safe_push (obj: routine); |
2903 | } |
2904 | |
2905 | /* Create a decision to call the routine, passing PARAMS to it. */ |
2906 | pattern_use *use = new pattern_use; |
2907 | use->routine = pat->routine; |
2908 | use->params.splice (src: params); |
2909 | decision *d = new decision (rtx_test::pattern (pos: res->root, pattern: use)); |
2910 | |
2911 | /* If the original decision could use an element of operands[] instead |
2912 | of an rtx variable, try to transfer it to the new decision. */ |
2913 | if (s->first->test.pos && res->root == s->first->test.pos) |
2914 | d->test.pos_operand = s->first->test.pos_operand; |
2915 | |
2916 | cpi->next_result = 0; |
2917 | return d; |
2918 | } |
2919 | |
2920 | /* Make S return the next unclaimed pattern routine result for CPI. */ |
2921 | |
2922 | static void |
2923 | add_pattern_acceptance (create_pattern_info *cpi, state *s) |
2924 | { |
2925 | acceptance_type acceptance; |
2926 | acceptance.type = SUBPATTERN; |
2927 | acceptance.partial_p = false; |
2928 | acceptance.u.full.code = cpi->next_result; |
2929 | add_decision (from: s, test: rtx_test::accept (acceptance), labels: true, optional: false); |
2930 | cpi->next_result += 1; |
2931 | } |
2932 | |
2933 | /* Initialize new empty state NEWS so that it implements SINFO's pattern |
2934 | (here referred to as "P"). P may be the top level of a pattern routine |
2935 | or a subpattern that should be inlined into its parent pattern's routine |
2936 | (as per same_pattern_p). The choice of SINFO for a top-level pattern is |
2937 | arbitrary; it could be any of the states that use P. The choice for |
2938 | subpatterns follows the choice for the parent pattern. |
2939 | |
2940 | PARAMS gives the value of each parameter to P in terms of the parameters |
2941 | to the top-level pattern. If P itself is the top level pattern, PARAMS[I] |
2942 | is always "parameter (TYPE, true, I)". */ |
2943 | |
2944 | static void |
2945 | populate_pattern_routine (create_pattern_info *cpi, merge_state_info *sinfo, |
2946 | state *news, const vec <parameter> ¶ms) |
2947 | { |
2948 | pattern_def_states += 1; |
2949 | |
2950 | decision *d = sinfo->s->singleton (); |
2951 | merge_pattern_info *pat = sinfo->res->pattern; |
2952 | pattern_routine *routine = cpi->routine; |
2953 | |
2954 | /* Create a copy of D's test for the pattern routine and generalize it |
2955 | as appropriate. */ |
2956 | decision *newd = new decision (d->test); |
2957 | gcc_assert (newd->test.pos_operand >= 0 |
2958 | || !newd->test.pos |
2959 | || common_position (newd->test.pos, |
2960 | routine->pos) == routine->pos); |
2961 | if (pat->param_test_p) |
2962 | { |
2963 | const parameter ¶m = params[pat->param_test]; |
2964 | switch (newd->test.kind) |
2965 | { |
2966 | case rtx_test::PREDICATE: |
2967 | newd->test.u.predicate.mode_is_param = param.is_param; |
2968 | newd->test.u.predicate.mode = param.value; |
2969 | break; |
2970 | |
2971 | case rtx_test::SAVED_CONST_INT: |
2972 | newd->test.u.integer.is_param = param.is_param; |
2973 | newd->test.u.integer.value = param.value; |
2974 | break; |
2975 | |
2976 | default: |
2977 | gcc_unreachable (); |
2978 | break; |
2979 | } |
2980 | } |
2981 | if (d->test.kind == rtx_test::C_TEST) |
2982 | routine->insn_p = true; |
2983 | else if (d->test.kind == rtx_test::HAVE_NUM_CLOBBERS) |
2984 | routine->pnum_clobbers_p = true; |
2985 | news->push_back (r: newd); |
2986 | |
2987 | /* Fill in the transitions of NEWD. */ |
2988 | unsigned int i = 0; |
2989 | for (transition *trans = d->first; trans; trans = trans->next) |
2990 | { |
2991 | /* Create a new state to act as the target of the new transition. */ |
2992 | state *to_news = new state; |
2993 | if (merge_pattern_transition *ptrans = pat->transitions[i]) |
2994 | { |
2995 | /* The pattern hasn't finished matching yet. Get the target |
2996 | pattern and the corresponding target state of SINFO. */ |
2997 | merge_pattern_info *to_pat = ptrans->to; |
2998 | merge_state_info *to = sinfo->to_states + i; |
2999 | gcc_assert (to->res->pattern == to_pat); |
3000 | gcc_assert (ptrans->params.length () == to_pat->params.length ()); |
3001 | |
3002 | /* Express the parameters to TO_PAT in terms of the parameters |
3003 | to the top-level pattern. */ |
3004 | auto_vec <parameter, MAX_PATTERN_PARAMS> to_params; |
3005 | for (unsigned int j = 0; j < ptrans->params.length (); ++j) |
3006 | { |
3007 | const parameter ¶m = ptrans->params[j]; |
3008 | to_params.quick_push (obj: param.is_param |
3009 | ? params[param.value] |
3010 | : param); |
3011 | } |
3012 | |
3013 | if (same_pattern_p (pat1: pat, pat2: to_pat)) |
3014 | /* TO_PAT is part of the current routine, so just recurse. */ |
3015 | populate_pattern_routine (cpi, sinfo: to, news: to_news, params: to_params); |
3016 | else |
3017 | { |
3018 | /* TO_PAT should be matched by calling a separate routine. */ |
3019 | create_pattern_info sub_cpi; |
3020 | decision *subd = init_pattern_use (cpi: &sub_cpi, sinfo: to, params: to_params); |
3021 | routine->insn_p |= sub_cpi.routine->insn_p; |
3022 | routine->pnum_clobbers_p |= sub_cpi.routine->pnum_clobbers_p; |
3023 | |
3024 | /* Add the pattern routine call to the new target state. */ |
3025 | to_news->push_back (r: subd); |
3026 | |
3027 | /* Add a transition for each successful call result. */ |
3028 | for (unsigned int j = 0; j < to_pat->num_results; ++j) |
3029 | { |
3030 | state *res = new state; |
3031 | add_pattern_acceptance (cpi, s: res); |
3032 | subd->push_back (r: new transition (j, res, false)); |
3033 | } |
3034 | } |
3035 | } |
3036 | else |
3037 | /* This transition corresponds to a successful match. */ |
3038 | add_pattern_acceptance (cpi, s: to_news); |
3039 | |
3040 | /* Create the transition itself, generalizing as necessary. */ |
3041 | transition *new_trans = new transition (trans->labels, to_news, |
3042 | trans->optional); |
3043 | if (pat->param_transition_p) |
3044 | { |
3045 | const parameter ¶m = params[pat->param_transition]; |
3046 | new_trans->is_param = param.is_param; |
3047 | new_trans->labels[0] = param.value; |
3048 | } |
3049 | newd->push_back (r: new_trans); |
3050 | i += 1; |
3051 | } |
3052 | } |
3053 | |
3054 | /* USE is a decision that calls a pattern routine and SINFO is part of the |
3055 | original state tree that the call is supposed to replace. Add the |
3056 | transitions for SINFO and its substates to USE. */ |
3057 | |
3058 | static void |
3059 | populate_pattern_use (create_pattern_info *cpi, decision *use, |
3060 | merge_state_info *sinfo) |
3061 | { |
3062 | pattern_use_states += 1; |
3063 | gcc_assert (!sinfo->merged_p); |
3064 | sinfo->merged_p = true; |
3065 | merge_state_result *res = sinfo->res; |
3066 | merge_pattern_info *pat = res->pattern; |
3067 | decision *d = sinfo->s->singleton (); |
3068 | unsigned int i = 0; |
3069 | for (transition *trans = d->first; trans; trans = trans->next) |
3070 | { |
3071 | if (pat->transitions[i]) |
3072 | /* The target state is also part of the pattern. */ |
3073 | populate_pattern_use (cpi, use, sinfo: sinfo->to_states + i); |
3074 | else |
3075 | { |
3076 | /* The transition corresponds to a successful return from the |
3077 | pattern routine. */ |
3078 | use->push_back (r: new transition (cpi->next_result, trans->to, false)); |
3079 | cpi->next_result += 1; |
3080 | } |
3081 | i += 1; |
3082 | } |
3083 | } |
3084 | |
3085 | /* We have decided to replace SINFO's state with a call to a pattern |
3086 | routine. Make the change, creating a definition of the pattern routine |
3087 | if it doesn't have one already. */ |
3088 | |
3089 | static void |
3090 | use_pattern (merge_state_info *sinfo) |
3091 | { |
3092 | merge_state_result *res = sinfo->res; |
3093 | merge_pattern_info *pat = res->pattern; |
3094 | state *s = sinfo->s; |
3095 | |
3096 | /* The pattern may have acquired new parameters after it was matched |
3097 | against SINFO. Update the parameters that SINFO passes accordingly. */ |
3098 | update_parameters (to&: res->params, from: pat->params); |
3099 | |
3100 | create_pattern_info cpi; |
3101 | decision *d = init_pattern_use (cpi: &cpi, sinfo, params: res->params); |
3102 | populate_pattern_use (cpi: &cpi, use: d, sinfo); |
3103 | s->release (); |
3104 | s->push_back (r: d); |
3105 | } |
3106 | |
3107 | /* Look through the state trees in STATES for common patterns and |
3108 | split them into subroutines. */ |
3109 | |
3110 | static void |
3111 | split_out_patterns (vec <merge_state_info> &states) |
3112 | { |
3113 | unsigned int first_transition = states.length (); |
3114 | hash_table <test_pattern_hasher> hashtab (128); |
3115 | /* Stage 1: Create an order in which parent states come before their child |
3116 | states and in which sibling states are at consecutive locations. |
3117 | Having consecutive sibling states allows merge_state_info to have |
3118 | a single to_states pointer. */ |
3119 | for (unsigned int i = 0; i < states.length (); ++i) |
3120 | for (decision *d = states[i].s->first; d; d = d->next) |
3121 | for (transition *trans = d->first; trans; trans = trans->next) |
3122 | { |
3123 | states.safe_push (obj: trans->to); |
3124 | states[i].num_transitions += 1; |
3125 | } |
3126 | /* Stage 2: Now that the addresses are stable, set up the to_states |
3127 | pointers. Look for states that might be merged and enter them |
3128 | into the hash table. */ |
3129 | for (unsigned int i = 0; i < states.length (); ++i) |
3130 | { |
3131 | merge_state_info *sinfo = &states[i]; |
3132 | if (sinfo->num_transitions) |
3133 | { |
3134 | sinfo->to_states = &states[first_transition]; |
3135 | first_transition += sinfo->num_transitions; |
3136 | } |
3137 | /* For simplicity, we only try to merge states that have a single |
3138 | decision. This is in any case the best we can do for peephole2, |
3139 | since whether a peephole2 ACCEPT succeeds or not depends on the |
3140 | specific peephole2 pattern (which is unique to each ACCEPT |
3141 | and so couldn't be shared between states). */ |
3142 | if (decision *d = sinfo->s->singleton ()) |
3143 | /* ACCEPT states are unique, so don't even try to merge them. */ |
3144 | if (d->test.kind != rtx_test::ACCEPT |
3145 | && (pattern_have_num_clobbers_p |
3146 | || d->test.kind != rtx_test::HAVE_NUM_CLOBBERS) |
3147 | && (pattern_c_test_p |
3148 | || d->test.kind != rtx_test::C_TEST)) |
3149 | { |
3150 | merge_state_info **slot = hashtab.find_slot (value: sinfo, insert: INSERT); |
3151 | sinfo->prev_same_test = *slot; |
3152 | *slot = sinfo; |
3153 | } |
3154 | } |
3155 | /* Stage 3: Walk backwards through the list of states and try to merge |
3156 | them. This is a greedy, bottom-up match; parent nodes can only start |
3157 | a new leaf pattern if they fail to match when combined with all child |
3158 | nodes that have matching patterns. |
3159 | |
3160 | For each state we keep a list of potential matches, with each |
3161 | potential match being larger (and deeper) than the next match in |
3162 | the list. The final element in the list is a leaf pattern that |
3163 | matches just a single state. |
3164 | |
3165 | Each candidate pattern created in this loop is unique -- it won't |
3166 | have been seen by an earlier iteration. We try to match each pattern |
3167 | with every state that appears earlier in STATES. |
3168 | |
3169 | Because the patterns created in the loop are unique, any state |
3170 | that already has a match must have a final potential match that |
3171 | is different from any new leaf pattern. Therefore, when matching |
3172 | leaf patterns, we need only consider states whose list of matches |
3173 | is empty. |
3174 | |
3175 | The non-leaf patterns that we try are as deep as possible |
3176 | and are an extension of the state's previous best candidate match (PB). |
3177 | We need only consider states whose current potential match is also PB; |
3178 | any states that don't match as much as PB cannnot match the new pattern, |
3179 | while any states that already match more than PB must be different from |
3180 | the new pattern. */ |
3181 | for (unsigned int i2 = states.length (); i2-- > 0; ) |
3182 | { |
3183 | merge_state_info *sinfo2 = &states[i2]; |
3184 | |
3185 | /* Enforce the bottom-upness of the match: remove matches with later |
3186 | states if SINFO2's child states ended up finding a better match. */ |
3187 | prune_invalid_results (sinfo: sinfo2); |
3188 | |
3189 | /* Do nothing if the state doesn't match a later one and if there are |
3190 | no earlier states it could match. */ |
3191 | if (!sinfo2->res && !sinfo2->prev_same_test) |
3192 | continue; |
3193 | |
3194 | merge_state_result *res2 = sinfo2->res; |
3195 | decision *d2 = sinfo2->s->singleton (); |
3196 | position *root2 = (d2->test.pos_operand < 0 ? d2->test.pos : 0); |
3197 | unsigned int num_transitions = sinfo2->num_transitions; |
3198 | |
3199 | /* If RES2 is null then SINFO2's test in isolation has not been seen |
3200 | before. First try matching that on its own. */ |
3201 | if (!res2) |
3202 | { |
3203 | merge_pattern_info *new_pat |
3204 | = new merge_pattern_info (num_transitions); |
3205 | merge_state_result *new_res2 |
3206 | = new merge_state_result (new_pat, root2, res2); |
3207 | sinfo2->res = new_res2; |
3208 | |
3209 | new_pat->num_statements = !d2->test.single_outcome_p (); |
3210 | new_pat->num_results = num_transitions; |
3211 | bool matched_p = false; |
3212 | /* Look for states that don't currently match anything but |
3213 | can be made to match SINFO2 on its own. */ |
3214 | for (merge_state_info *sinfo1 = sinfo2->prev_same_test; sinfo1; |
3215 | sinfo1 = sinfo1->prev_same_test) |
3216 | if (!sinfo1->res && merge_patterns (sinfo1, sinfo2)) |
3217 | matched_p = true; |
3218 | if (!matched_p) |
3219 | { |
3220 | /* No other states match. */ |
3221 | sinfo2->res = res2; |
3222 | delete new_pat; |
3223 | delete new_res2; |
3224 | continue; |
3225 | } |
3226 | else |
3227 | res2 = new_res2; |
3228 | } |
3229 | |
3230 | /* Keep the existing pattern if it's as good as anything we'd |
3231 | create for SINFO2. */ |
3232 | if (complete_result_p (pat: res2->pattern, sinfo: sinfo2)) |
3233 | { |
3234 | res2->pattern->num_users += 1; |
3235 | continue; |
3236 | } |
3237 | |
3238 | /* Create a new pattern for SINFO2. */ |
3239 | merge_pattern_info *new_pat = new merge_pattern_info (num_transitions); |
3240 | merge_state_result *new_res2 |
3241 | = new merge_state_result (new_pat, root2, res2); |
3242 | sinfo2->res = new_res2; |
3243 | |
3244 | /* Fill in details about the pattern. */ |
3245 | new_pat->num_statements = !d2->test.single_outcome_p (); |
3246 | new_pat->num_results = 0; |
3247 | for (unsigned int j = 0; j < num_transitions; ++j) |
3248 | if (merge_state_result *to_res = sinfo2->to_states[j].res) |
3249 | { |
3250 | /* Count the target state as part of this pattern. |
3251 | First update the root position so that it can reach |
3252 | the target state's root. */ |
3253 | if (to_res->root) |
3254 | { |
3255 | if (new_res2->root) |
3256 | new_res2->root = common_position (pos1: new_res2->root, |
3257 | pos2: to_res->root); |
3258 | else |
3259 | new_res2->root = to_res->root; |
3260 | } |
3261 | merge_pattern_info *to_pat = to_res->pattern; |
3262 | merge_pattern_transition *ptrans |
3263 | = new merge_pattern_transition (to_pat); |
3264 | |
3265 | /* TO_PAT may have acquired more parameters when matching |
3266 | states earlier in STATES than TO_RES's, but the list is |
3267 | now final. Make sure that TO_RES is up to date. */ |
3268 | update_parameters (to&: to_res->params, from: to_pat->params); |
3269 | |
3270 | /* Start out by assuming that every user of NEW_PAT will |
3271 | want to pass the same (constant) parameters as TO_RES. */ |
3272 | update_parameters (to&: ptrans->params, from: to_res->params); |
3273 | |
3274 | new_pat->transitions[j] = ptrans; |
3275 | new_pat->num_statements += to_pat->num_statements; |
3276 | new_pat->num_results += to_pat->num_results; |
3277 | } |
3278 | else |
3279 | /* The target state doesn't match anything and so is not part |
3280 | of the pattern. */ |
3281 | new_pat->num_results += 1; |
3282 | |
3283 | /* See if any earlier states that match RES2's pattern also match |
3284 | NEW_PAT. */ |
3285 | bool matched_p = false; |
3286 | for (merge_state_info *sinfo1 = sinfo2->prev_same_test; sinfo1; |
3287 | sinfo1 = sinfo1->prev_same_test) |
3288 | { |
3289 | prune_invalid_results (sinfo: sinfo1); |
3290 | if (sinfo1->res |
3291 | && sinfo1->res->pattern == res2->pattern |
3292 | && merge_patterns (sinfo1, sinfo2)) |
3293 | matched_p = true; |
3294 | } |
3295 | if (!matched_p) |
3296 | { |
3297 | /* Nothing else matches NEW_PAT, so go back to the previous |
3298 | pattern (possibly just a single-state one). */ |
3299 | sinfo2->res = res2; |
3300 | delete new_pat; |
3301 | delete new_res2; |
3302 | } |
3303 | /* Assume that SINFO2 will use RES. At this point we don't know |
3304 | whether earlier states that match the same pattern will use |
3305 | that match or a different one. */ |
3306 | sinfo2->res->pattern->num_users += 1; |
3307 | } |
3308 | /* Step 4: Finalize the choice of pattern for each state, ignoring |
3309 | patterns that were only used once. Update each pattern's size |
3310 | so that it doesn't include subpatterns that are going to be split |
3311 | out into subroutines. */ |
3312 | for (unsigned int i = 0; i < states.length (); ++i) |
3313 | { |
3314 | merge_state_info *sinfo = &states[i]; |
3315 | merge_state_result *res = sinfo->res; |
3316 | /* Wind past patterns that are only used by SINFO. */ |
3317 | while (res && res->pattern->num_users == 1) |
3318 | { |
3319 | res = res->prev; |
3320 | sinfo->res = res; |
3321 | if (res) |
3322 | res->pattern->num_users += 1; |
3323 | } |
3324 | if (!res) |
3325 | continue; |
3326 | |
3327 | /* We have a shared pattern and are now committed to the match. */ |
3328 | merge_pattern_info *pat = res->pattern; |
3329 | gcc_assert (valid_result_p (pat, sinfo)); |
3330 | |
3331 | if (!pat->complete_p) |
3332 | { |
3333 | /* Look for subpatterns that are going to be split out and remove |
3334 | them from the number of statements. */ |
3335 | for (unsigned int j = 0; j < sinfo->num_transitions; ++j) |
3336 | if (merge_pattern_transition *ptrans = pat->transitions[j]) |
3337 | { |
3338 | merge_pattern_info *to_pat = ptrans->to; |
3339 | if (!same_pattern_p (pat1: pat, pat2: to_pat)) |
3340 | pat->num_statements -= to_pat->num_statements; |
3341 | } |
3342 | pat->complete_p = true; |
3343 | } |
3344 | } |
3345 | /* Step 5: Split out the patterns. */ |
3346 | for (unsigned int i = 0; i < states.length (); ++i) |
3347 | { |
3348 | merge_state_info *sinfo = &states[i]; |
3349 | merge_state_result *res = sinfo->res; |
3350 | if (!sinfo->merged_p && res && useful_pattern_p (pat: res->pattern)) |
3351 | use_pattern (sinfo); |
3352 | } |
3353 | fprintf (stderr, format: "Shared %d out of %d states by creating %d new states," |
3354 | " saving %d\n" , |
3355 | pattern_use_states, states.length (), pattern_def_states, |
3356 | pattern_use_states - pattern_def_states); |
3357 | } |
3358 | |
3359 | /* Information about a state tree that we're considering splitting into a |
3360 | subroutine. */ |
3361 | struct state_size |
3362 | { |
3363 | /* The number of pseudo-statements in the state tree. */ |
3364 | unsigned int num_statements; |
3365 | |
3366 | /* The approximate number of nested "if" and "switch" statements that |
3367 | would be required if control could fall through to a later state. */ |
3368 | unsigned int depth; |
3369 | }; |
3370 | |
3371 | /* Pairs a transition with information about its target state. */ |
3372 | typedef std::pair <transition *, state_size> subroutine_candidate; |
3373 | |
3374 | /* Sort two subroutine_candidates so that the one with the largest |
3375 | number of statements comes last. */ |
3376 | |
3377 | static int |
3378 | subroutine_candidate_cmp (const void *a, const void *b) |
3379 | { |
3380 | return int (((const subroutine_candidate *) a)->second.num_statements |
3381 | - ((const subroutine_candidate *) b)->second.num_statements); |
3382 | } |
3383 | |
3384 | /* Turn S into a subroutine of type TYPE and add it to PROCS. Return a new |
3385 | state that performs a subroutine call to S. */ |
3386 | |
3387 | static state * |
3388 | create_subroutine (routine_type type, state *s, vec <state *> &procs) |
3389 | { |
3390 | procs.safe_push (obj: s); |
3391 | acceptance_type acceptance; |
3392 | acceptance.type = type; |
3393 | acceptance.partial_p = true; |
3394 | acceptance.u.subroutine_id = procs.length (); |
3395 | state *news = new state; |
3396 | add_decision (from: news, test: rtx_test::accept (acceptance), labels: true, optional: false); |
3397 | return news; |
3398 | } |
3399 | |
3400 | /* Walk state tree S, of type TYPE, and look for subtrees that would be |
3401 | better split into subroutines. Accumulate all such subroutines in PROCS. |
3402 | Return the size of the new state tree (excluding subroutines). */ |
3403 | |
3404 | static state_size |
3405 | find_subroutines (routine_type type, state *s, vec <state *> &procs) |
3406 | { |
3407 | auto_vec <subroutine_candidate, 16> candidates; |
3408 | state_size size; |
3409 | size.num_statements = 0; |
3410 | size.depth = 0; |
3411 | for (decision *d = s->first; d; d = d->next) |
3412 | { |
3413 | if (!d->test.single_outcome_p ()) |
3414 | size.num_statements += 1; |
3415 | for (transition *trans = d->first; trans; trans = trans->next) |
3416 | { |
3417 | /* Keep chains of simple decisions together if we know that no |
3418 | change of position is required. We'll output this chain as a |
3419 | single "if" statement, so it counts as a single nesting level. */ |
3420 | if (d->test.pos && d->if_statement_p ()) |
3421 | for (;;) |
3422 | { |
3423 | decision *newd = trans->to->singleton (); |
3424 | if (!newd |
3425 | || (newd->test.pos |
3426 | && newd->test.pos_operand < 0 |
3427 | && newd->test.pos != d->test.pos) |
3428 | || !newd->if_statement_p ()) |
3429 | break; |
3430 | if (!newd->test.single_outcome_p ()) |
3431 | size.num_statements += 1; |
3432 | trans = newd->singleton (); |
3433 | if (newd->test.kind == rtx_test::SET_OP |
3434 | || newd->test.kind == rtx_test::ACCEPT) |
3435 | break; |
3436 | } |
3437 | /* The target of TRANS is a subroutine candidate. First recurse |
3438 | on it to see how big it is after subroutines have been |
3439 | split out. */ |
3440 | state_size to_size = find_subroutines (type, s: trans->to, procs); |
3441 | if (d->next && to_size.depth > MAX_DEPTH) |
3442 | /* Keeping the target state in the same routine would lead |
3443 | to an excessive nesting of "if" and "switch" statements. |
3444 | Split it out into a subroutine so that it can use |
3445 | inverted tests that return early on failure. */ |
3446 | trans->to = create_subroutine (type, s: trans->to, procs); |
3447 | else |
3448 | { |
3449 | size.num_statements += to_size.num_statements; |
3450 | if (to_size.num_statements < MIN_NUM_STATEMENTS) |
3451 | /* The target state is too small to be worth splitting. |
3452 | Keep it in the same routine as S. */ |
3453 | size.depth = MAX (size.depth, to_size.depth); |
3454 | else |
3455 | /* Assume for now that we'll keep the target state in the |
3456 | same routine as S, but record it as a subroutine candidate |
3457 | if S grows too big. */ |
3458 | candidates.safe_push (obj: subroutine_candidate (trans, to_size)); |
3459 | } |
3460 | } |
3461 | } |
3462 | if (size.num_statements > MAX_NUM_STATEMENTS) |
3463 | { |
3464 | /* S is too big. Sort the subroutine candidates so that bigger ones |
3465 | are nearer the end. */ |
3466 | candidates.qsort (subroutine_candidate_cmp); |
3467 | while (!candidates.is_empty () |
3468 | && size.num_statements > MAX_NUM_STATEMENTS) |
3469 | { |
3470 | /* Peel off a candidate and force it into a subroutine. */ |
3471 | subroutine_candidate cand = candidates.pop (); |
3472 | size.num_statements -= cand.second.num_statements; |
3473 | cand.first->to = create_subroutine (type, s: cand.first->to, procs); |
3474 | } |
3475 | } |
3476 | /* Update the depth for subroutine candidates that we decided not to |
3477 | split out. */ |
3478 | for (unsigned int i = 0; i < candidates.length (); ++i) |
3479 | size.depth = MAX (size.depth, candidates[i].second.depth); |
3480 | size.depth += 1; |
3481 | return size; |
3482 | } |
3483 | |
3484 | /* Return true if, for all X, PRED (X, MODE) implies that X has mode MODE. */ |
3485 | |
3486 | static bool |
3487 | safe_predicate_mode (const struct pred_data *pred, machine_mode mode) |
3488 | { |
3489 | /* Scalar integer constants have VOIDmode. */ |
3490 | if (GET_MODE_CLASS (mode) == MODE_INT |
3491 | && (pred->codes[CONST_INT] |
3492 | || pred->codes[CONST_DOUBLE] |
3493 | || pred->codes[CONST_WIDE_INT] |
3494 | || pred->codes[LABEL_REF])) |
3495 | return false; |
3496 | |
3497 | return !pred->special && mode != VOIDmode; |
3498 | } |
3499 | |
3500 | /* Fill CODES with the set of codes that could be matched by PRED. */ |
3501 | |
3502 | static void |
3503 | get_predicate_codes (const struct pred_data *pred, int_set *codes) |
3504 | { |
3505 | for (int i = 0; i < NUM_TRUE_RTX_CODE; ++i) |
3506 | if (!pred || pred->codes[i]) |
3507 | codes->safe_push (obj: i); |
3508 | } |
3509 | |
3510 | /* Return true if the first path through D1 tests the same thing as D2. */ |
3511 | |
3512 | static bool |
3513 | has_same_test_p (decision *d1, decision *d2) |
3514 | { |
3515 | do |
3516 | { |
3517 | if (d1->test == d2->test) |
3518 | return true; |
3519 | d1 = d1->first->to->first; |
3520 | } |
3521 | while (d1); |
3522 | return false; |
3523 | } |
3524 | |
3525 | /* Return true if D1 and D2 cannot match the same rtx. All states reachable |
3526 | from D2 have single decisions and all those decisions have single |
3527 | transitions. */ |
3528 | |
3529 | static bool |
3530 | mutually_exclusive_p (decision *d1, decision *d2) |
3531 | { |
3532 | /* If one path through D1 fails to test the same thing as D2, assume |
3533 | that D2's test could be true for D1 and look for a later, more useful, |
3534 | test. This isn't as expensive as it looks in practice. */ |
3535 | while (!has_same_test_p (d1, d2)) |
3536 | { |
3537 | d2 = d2->singleton ()->to->singleton (); |
3538 | if (!d2) |
3539 | return false; |
3540 | } |
3541 | if (d1->test == d2->test) |
3542 | { |
3543 | /* Look for any transitions from D1 that have the same labels as |
3544 | the transition from D2. */ |
3545 | transition *trans2 = d2->singleton (); |
3546 | for (transition *trans1 = d1->first; trans1; trans1 = trans1->next) |
3547 | { |
3548 | int_set::iterator i1 = trans1->labels.begin (); |
3549 | int_set::iterator end1 = trans1->labels.end (); |
3550 | int_set::iterator i2 = trans2->labels.begin (); |
3551 | int_set::iterator end2 = trans2->labels.end (); |
3552 | while (i1 != end1 && i2 != end2) |
3553 | if (*i1 < *i2) |
3554 | ++i1; |
3555 | else if (*i2 < *i1) |
3556 | ++i2; |
3557 | else |
3558 | { |
3559 | /* TRANS1 has some labels in common with TRANS2. Assume |
3560 | that D1 and D2 could match the same rtx if the target |
3561 | of TRANS1 could match the same rtx as D2. */ |
3562 | for (decision *subd1 = trans1->to->first; |
3563 | subd1; subd1 = subd1->next) |
3564 | if (!mutually_exclusive_p (d1: subd1, d2)) |
3565 | return false; |
3566 | break; |
3567 | } |
3568 | } |
3569 | return true; |
3570 | } |
3571 | for (transition *trans1 = d1->first; trans1; trans1 = trans1->next) |
3572 | for (decision *subd1 = trans1->to->first; subd1; subd1 = subd1->next) |
3573 | if (!mutually_exclusive_p (d1: subd1, d2)) |
3574 | return false; |
3575 | return true; |
3576 | } |
3577 | |
3578 | /* Try to merge S2's decision into D1, given that they have the same test. |
3579 | Fail only if EXCLUDE is nonnull and the new transition would have the |
3580 | same labels as *EXCLUDE. When returning true, set *NEXT_S1, *NEXT_S2 |
3581 | and *NEXT_EXCLUDE as for merge_into_state_1, or set *NEXT_S2 to null |
3582 | if the merge is complete. */ |
3583 | |
3584 | static bool |
3585 | merge_into_decision (decision *d1, state *s2, const int_set *exclude, |
3586 | state **next_s1, state **next_s2, |
3587 | const int_set **next_exclude) |
3588 | { |
3589 | decision *d2 = s2->singleton (); |
3590 | transition *trans2 = d2->singleton (); |
3591 | |
3592 | /* Get a list of the transitions that intersect TRANS2. */ |
3593 | auto_vec <transition *, 32> intersecting; |
3594 | for (transition *trans1 = d1->first; trans1; trans1 = trans1->next) |
3595 | { |
3596 | int_set::iterator i1 = trans1->labels.begin (); |
3597 | int_set::iterator end1 = trans1->labels.end (); |
3598 | int_set::iterator i2 = trans2->labels.begin (); |
3599 | int_set::iterator end2 = trans2->labels.end (); |
3600 | bool trans1_is_subset = true; |
3601 | bool trans2_is_subset = true; |
3602 | bool intersect_p = false; |
3603 | while (i1 != end1 && i2 != end2) |
3604 | if (*i1 < *i2) |
3605 | { |
3606 | trans1_is_subset = false; |
3607 | ++i1; |
3608 | } |
3609 | else if (*i2 < *i1) |
3610 | { |
3611 | trans2_is_subset = false; |
3612 | ++i2; |
3613 | } |
3614 | else |
3615 | { |
3616 | intersect_p = true; |
3617 | ++i1; |
3618 | ++i2; |
3619 | } |
3620 | if (i1 != end1) |
3621 | trans1_is_subset = false; |
3622 | if (i2 != end2) |
3623 | trans2_is_subset = false; |
3624 | if (trans1_is_subset && trans2_is_subset) |
3625 | { |
3626 | /* There's already a transition that matches exactly. |
3627 | Merge the target states. */ |
3628 | trans1->optional &= trans2->optional; |
3629 | *next_s1 = trans1->to; |
3630 | *next_s2 = trans2->to; |
3631 | *next_exclude = 0; |
3632 | return true; |
3633 | } |
3634 | if (trans2_is_subset) |
3635 | { |
3636 | /* TRANS1 has all the labels that TRANS2 needs. Merge S2 into |
3637 | the target of TRANS1, but (to avoid infinite recursion) |
3638 | make sure that we don't end up creating another transition |
3639 | like TRANS1. */ |
3640 | *next_s1 = trans1->to; |
3641 | *next_s2 = s2; |
3642 | *next_exclude = &trans1->labels; |
3643 | return true; |
3644 | } |
3645 | if (intersect_p) |
3646 | intersecting.safe_push (obj: trans1); |
3647 | } |
3648 | |
3649 | if (intersecting.is_empty ()) |
3650 | { |
3651 | /* No existing labels intersect the new ones. We can just add |
3652 | TRANS2 itself. */ |
3653 | d1->push_back (r: d2->release ()); |
3654 | *next_s1 = 0; |
3655 | *next_s2 = 0; |
3656 | *next_exclude = 0; |
3657 | return true; |
3658 | } |
3659 | |
3660 | /* Take the union of the labels in INTERSECTING and TRANS2. Store the |
3661 | result in COMBINED and use NEXT as a temporary. */ |
3662 | int_set tmp1 = trans2->labels, tmp2; |
3663 | int_set *combined = &tmp1, *next = &tmp2; |
3664 | for (unsigned int i = 0; i < intersecting.length (); ++i) |
3665 | { |
3666 | transition *trans1 = intersecting[i]; |
3667 | next->truncate (size: 0); |
3668 | next->safe_grow (len: trans1->labels.length () + combined->length (), exact: true); |
3669 | int_set::iterator end |
3670 | = std::set_union (first1: trans1->labels.begin (), last1: trans1->labels.end (), |
3671 | first2: combined->begin (), last2: combined->end (), |
3672 | result: next->begin ()); |
3673 | next->truncate (size: end - next->begin ()); |
3674 | std::swap (a&: next, b&: combined); |
3675 | } |
3676 | |
3677 | /* Stop now if we've been told not to create a transition with these |
3678 | labels. */ |
3679 | if (exclude && *combined == *exclude) |
3680 | return false; |
3681 | |
3682 | /* Get the transition that should carry the new labels. */ |
3683 | transition *new_trans = intersecting[0]; |
3684 | if (intersecting.length () == 1) |
3685 | { |
3686 | /* We're merging with one existing transition whose labels are a |
3687 | subset of those required. If both transitions are optional, |
3688 | we can just expand the set of labels so that it's suitable |
3689 | for both transitions. It isn't worth preserving the original |
3690 | transitions since we know that they can't be merged; we would |
3691 | need to backtrack to S2 if TRANS1->to fails. In contrast, |
3692 | we might be able to merge the targets of the transitions |
3693 | without any backtracking. |
3694 | |
3695 | If instead the existing transition is not optional, ensure that |
3696 | all target decisions are suitably protected. Some decisions |
3697 | might already have a more specific requirement than NEW_TRANS, |
3698 | in which case there's no point testing NEW_TRANS as well. E.g. this |
3699 | would have happened if a test for an (eq ...) rtx had been |
3700 | added to a decision that tested whether the code is suitable |
3701 | for comparison_operator. The original comparison_operator |
3702 | transition would have been non-optional and the (eq ...) test |
3703 | would be performed by a second decision in the target of that |
3704 | transition. |
3705 | |
3706 | The remaining case -- keeping the original optional transition |
3707 | when adding a non-optional TRANS2 -- is a wash. Preserving |
3708 | the optional transition only helps if we later merge another |
3709 | state S3 that is mutually exclusive with S2 and whose labels |
3710 | belong to *COMBINED - TRANS1->labels. We can then test the |
3711 | original NEW_TRANS and S3 in the same decision. We keep the |
3712 | optional transition around for that case, but it occurs very |
3713 | rarely. */ |
3714 | gcc_assert (new_trans->labels != *combined); |
3715 | if (!new_trans->optional || !trans2->optional) |
3716 | { |
3717 | decision *start = 0; |
3718 | for (decision *end = new_trans->to->first; end; end = end->next) |
3719 | { |
3720 | if (!start && end->test != d1->test) |
3721 | /* END belongs to a range of decisions that need to be |
3722 | protected by NEW_TRANS. */ |
3723 | start = end; |
3724 | if (start && (!end->next || end->next->test == d1->test)) |
3725 | { |
3726 | /* Protect [START, END] with NEW_TRANS. The decisions |
3727 | move to NEW_S and NEW_D becomes part of NEW_TRANS->to. */ |
3728 | state *new_s = new state; |
3729 | decision *new_d = new decision (d1->test); |
3730 | new_d->push_back (r: new transition (new_trans->labels, new_s, |
3731 | new_trans->optional)); |
3732 | state::range r (start, end); |
3733 | new_trans->to->replace (oldr: r, newr: new_d); |
3734 | new_s->push_back (r); |
3735 | |
3736 | /* Continue with an empty range. */ |
3737 | start = 0; |
3738 | |
3739 | /* Continue from the decision after NEW_D. */ |
3740 | end = new_d; |
3741 | } |
3742 | } |
3743 | } |
3744 | new_trans->optional = true; |
3745 | new_trans->labels = *combined; |
3746 | } |
3747 | else |
3748 | { |
3749 | /* We're merging more than one existing transition together. |
3750 | Those transitions are successfully dividing the matching space |
3751 | and so we want to preserve them, even if they're optional. |
3752 | |
3753 | Create a new transition with the union set of labels and make |
3754 | it go to a state that has the original transitions. */ |
3755 | decision *new_d = new decision (d1->test); |
3756 | for (unsigned int i = 0; i < intersecting.length (); ++i) |
3757 | new_d->push_back (r: d1->remove (r: intersecting[i])); |
3758 | |
3759 | state *new_s = new state; |
3760 | new_s->push_back (r: new_d); |
3761 | |
3762 | new_trans = new transition (*combined, new_s, true); |
3763 | d1->push_back (r: new_trans); |
3764 | } |
3765 | |
3766 | /* We now have an optional transition with labels *COMBINED. Decide |
3767 | whether we can use it as TRANS2 or whether we need to merge S2 |
3768 | into the target of NEW_TRANS. */ |
3769 | gcc_assert (new_trans->optional); |
3770 | if (new_trans->labels == trans2->labels) |
3771 | { |
3772 | /* NEW_TRANS matches TRANS2. Just merge the target states. */ |
3773 | new_trans->optional = trans2->optional; |
3774 | *next_s1 = new_trans->to; |
3775 | *next_s2 = trans2->to; |
3776 | *next_exclude = 0; |
3777 | } |
3778 | else |
3779 | { |
3780 | /* Try to merge TRANS2 into the target of the overlapping transition, |
3781 | but (to prevent infinite recursion or excessive redundancy) without |
3782 | creating another transition of the same type. */ |
3783 | *next_s1 = new_trans->to; |
3784 | *next_s2 = s2; |
3785 | *next_exclude = &new_trans->labels; |
3786 | } |
3787 | return true; |
3788 | } |
3789 | |
3790 | /* Make progress in merging S2 into S1, given that each state in S2 |
3791 | has a single decision. If EXCLUDE is nonnull, avoid creating a new |
3792 | transition with the same test as S2's decision and with the labels |
3793 | in *EXCLUDE. |
3794 | |
3795 | Return true if there is still work to do. When returning true, |
3796 | set *NEXT_S1, *NEXT_S2 and *NEXT_EXCLUDE to the values that |
3797 | S1, S2 and EXCLUDE should have next time round. |
3798 | |
3799 | If S1 and S2 both match a particular rtx, give priority to S1. */ |
3800 | |
3801 | static bool |
3802 | merge_into_state_1 (state *s1, state *s2, const int_set *exclude, |
3803 | state **next_s1, state **next_s2, |
3804 | const int_set **next_exclude) |
3805 | { |
3806 | decision *d2 = s2->singleton (); |
3807 | if (decision *d1 = s1->last) |
3808 | { |
3809 | if (d1->test.terminal_p ()) |
3810 | /* D1 is an unconditional return, so S2 can never match. This can |
3811 | sometimes be a bug in the .md description, but might also happen |
3812 | if genconditions forces some conditions to true for certain |
3813 | configurations. */ |
3814 | return false; |
3815 | |
3816 | /* Go backwards through the decisions in S1, stopping once we find one |
3817 | that could match the same thing as S2. */ |
3818 | while (d1->prev && mutually_exclusive_p (d1, d2)) |
3819 | d1 = d1->prev; |
3820 | |
3821 | /* Search forwards from that point, merging D2 into the first |
3822 | decision we can. */ |
3823 | for (; d1; d1 = d1->next) |
3824 | { |
3825 | /* If S2 performs some optional tests before testing the same thing |
3826 | as D1, those tests do not help to distinguish D1 and S2, so it's |
3827 | better to drop them. Search through such optional decisions |
3828 | until we find something that tests the same thing as D1. */ |
3829 | state *sub_s2 = s2; |
3830 | for (;;) |
3831 | { |
3832 | decision *sub_d2 = sub_s2->singleton (); |
3833 | if (d1->test == sub_d2->test) |
3834 | { |
3835 | /* Only apply EXCLUDE if we're testing the same thing |
3836 | as D2. */ |
3837 | const int_set *sub_exclude = (d2 == sub_d2 ? exclude : 0); |
3838 | |
3839 | /* Try to merge SUB_S2 into D1. This can only fail if |
3840 | it would involve creating a new transition with |
3841 | labels SUB_EXCLUDE. */ |
3842 | if (merge_into_decision (d1, s2: sub_s2, exclude: sub_exclude, |
3843 | next_s1, next_s2, next_exclude)) |
3844 | return *next_s2 != 0; |
3845 | |
3846 | /* Can't merge with D1; try a later decision. */ |
3847 | break; |
3848 | } |
3849 | transition *sub_trans2 = sub_d2->singleton (); |
3850 | if (!sub_trans2->optional) |
3851 | /* Can't merge with D1; try a later decision. */ |
3852 | break; |
3853 | sub_s2 = sub_trans2->to; |
3854 | } |
3855 | } |
3856 | } |
3857 | |
3858 | /* We can't merge D2 with any existing decision. Just add it to the end. */ |
3859 | s1->push_back (r: s2->release ()); |
3860 | return false; |
3861 | } |
3862 | |
3863 | /* Merge S2 into S1. If they both match a particular rtx, give |
3864 | priority to S1. Each state in S2 has a single decision. */ |
3865 | |
3866 | static void |
3867 | merge_into_state (state *s1, state *s2) |
3868 | { |
3869 | const int_set *exclude = 0; |
3870 | while (s2 && merge_into_state_1 (s1, s2, exclude, next_s1: &s1, next_s2: &s2, next_exclude: &exclude)) |
3871 | continue; |
3872 | } |
3873 | |
3874 | /* Pairs a pattern that needs to be matched with the rtx position at |
3875 | which the pattern should occur. */ |
3876 | class pattern_pos { |
3877 | public: |
3878 | pattern_pos () {} |
3879 | pattern_pos (rtx, position *); |
3880 | |
3881 | rtx pattern; |
3882 | position *pos; |
3883 | }; |
3884 | |
3885 | pattern_pos::pattern_pos (rtx pattern_in, position *pos_in) |
3886 | : pattern (pattern_in), pos (pos_in) |
3887 | {} |
3888 | |
3889 | /* Compare entries according to their depth-first order. There shouldn't |
3890 | be two entries at the same position. */ |
3891 | |
3892 | bool |
3893 | operator < (const pattern_pos &e1, const pattern_pos &e2) |
3894 | { |
3895 | int diff = compare_positions (pos1: e1.pos, pos2: e2.pos); |
3896 | gcc_assert (diff != 0 || e1.pattern == e2.pattern); |
3897 | return diff < 0; |
3898 | } |
3899 | |
3900 | /* Add new decisions to S that check whether the rtx at position POS |
3901 | matches PATTERN. Return the state that is reached in that case. |
3902 | TOP_PATTERN is the overall pattern, as passed to match_pattern_1. */ |
3903 | |
3904 | static state * |
3905 | match_pattern_2 (state *s, md_rtx_info *info, position *pos, rtx pattern) |
3906 | { |
3907 | auto_vec <pattern_pos, 32> worklist; |
3908 | auto_vec <pattern_pos, 32> pred_and_mode_tests; |
3909 | auto_vec <pattern_pos, 32> dup_tests; |
3910 | |
3911 | worklist.safe_push (obj: pattern_pos (pattern, pos)); |
3912 | while (!worklist.is_empty ()) |
3913 | { |
3914 | pattern_pos next = worklist.pop (); |
3915 | pattern = next.pattern; |
3916 | pos = next.pos; |
3917 | unsigned int reverse_s = worklist.length (); |
3918 | |
3919 | enum rtx_code code = GET_CODE (pattern); |
3920 | switch (code) |
3921 | { |
3922 | case MATCH_OP_DUP: |
3923 | case MATCH_DUP: |
3924 | case MATCH_PAR_DUP: |
3925 | /* Add a test that the rtx matches the earlier one, but only |
3926 | after the structure and predicates have been checked. */ |
3927 | dup_tests.safe_push (obj: pattern_pos (pattern, pos)); |
3928 | |
3929 | /* Use the same code check as the original operand. */ |
3930 | pattern = find_operand (pattern: info->def, XINT (pattern, 0), NULL_RTX); |
3931 | /* Fall through. */ |
3932 | |
3933 | case MATCH_PARALLEL: |
3934 | case MATCH_OPERAND: |
3935 | case MATCH_SCRATCH: |
3936 | case MATCH_OPERATOR: |
3937 | { |
3938 | const char *pred_name = predicate_name (match_rtx: pattern); |
3939 | const struct pred_data *pred = 0; |
3940 | if (pred_name[0] != 0) |
3941 | { |
3942 | pred = lookup_predicate (pred_name); |
3943 | /* Only report errors once per rtx. */ |
3944 | if (code == GET_CODE (pattern)) |
3945 | { |
3946 | if (!pred) |
3947 | error_at (info->loc, "unknown predicate '%s' used in %s" , |
3948 | pred_name, GET_RTX_NAME (code)); |
3949 | else if (code == MATCH_PARALLEL |
3950 | && pred->singleton != PARALLEL) |
3951 | error_at (info->loc, "predicate '%s' used in" |
3952 | " match_parallel does not allow only PARALLEL" , |
3953 | pred->name); |
3954 | } |
3955 | } |
3956 | |
3957 | if (code == MATCH_PARALLEL || code == MATCH_PAR_DUP) |
3958 | { |
3959 | /* Check that we have a parallel with enough elements. */ |
3960 | s = add_decision (from: s, test: rtx_test::code (pos), labels: PARALLEL, optional: false); |
3961 | int min_len = XVECLEN (pattern, 2); |
3962 | s = add_decision (from: s, test: rtx_test::veclen_ge (pos, min_len), |
3963 | labels: true, optional: false); |
3964 | } |
3965 | else |
3966 | { |
3967 | /* Check that the rtx has one of codes accepted by the |
3968 | predicate. This is necessary when matching suboperands |
3969 | of a MATCH_OPERATOR or MATCH_OP_DUP, since we can't |
3970 | call XEXP (X, N) without checking that X has at least |
3971 | N+1 operands. */ |
3972 | int_set codes; |
3973 | get_predicate_codes (pred, codes: &codes); |
3974 | bool need_codes = (pred |
3975 | && (code == MATCH_OPERATOR |
3976 | || code == MATCH_OP_DUP)); |
3977 | s = add_decision (from: s, test: rtx_test::code (pos), labels: codes, optional: !need_codes); |
3978 | } |
3979 | |
3980 | /* Postpone the predicate check until we've checked the rest |
3981 | of the rtx structure. */ |
3982 | if (code == GET_CODE (pattern)) |
3983 | pred_and_mode_tests.safe_push (obj: pattern_pos (pattern, pos)); |
3984 | |
3985 | /* If we need to match suboperands, add them to the worklist. */ |
3986 | if (code == MATCH_OPERATOR || code == MATCH_PARALLEL) |
3987 | { |
3988 | position **subpos_ptr; |
3989 | enum position_type pos_type; |
3990 | int i; |
3991 | if (code == MATCH_OPERATOR || code == MATCH_OP_DUP) |
3992 | { |
3993 | pos_type = POS_XEXP; |
3994 | subpos_ptr = &pos->xexps; |
3995 | i = (code == MATCH_OPERATOR ? 2 : 1); |
3996 | } |
3997 | else |
3998 | { |
3999 | pos_type = POS_XVECEXP0; |
4000 | subpos_ptr = &pos->xvecexp0s; |
4001 | i = 2; |
4002 | } |
4003 | for (int j = 0; j < XVECLEN (pattern, i); ++j) |
4004 | { |
4005 | position *subpos = next_position (next_ptr: subpos_ptr, base: pos, |
4006 | type: pos_type, arg: j); |
4007 | worklist.safe_push (obj: pattern_pos (XVECEXP (pattern, i, j), |
4008 | subpos)); |
4009 | subpos_ptr = &subpos->next; |
4010 | } |
4011 | } |
4012 | break; |
4013 | } |
4014 | |
4015 | default: |
4016 | { |
4017 | /* Check that the rtx has the right code. */ |
4018 | s = add_decision (from: s, test: rtx_test::code (pos), labels: code, optional: false); |
4019 | |
4020 | /* Queue a test for the mode if one is specified. */ |
4021 | if (GET_MODE (pattern) != VOIDmode) |
4022 | pred_and_mode_tests.safe_push (obj: pattern_pos (pattern, pos)); |
4023 | |
4024 | /* Push subrtxes onto the worklist. Match nonrtx operands now. */ |
4025 | const char *fmt = GET_RTX_FORMAT (code); |
4026 | position **subpos_ptr = &pos->xexps; |
4027 | for (size_t i = 0; fmt[i]; ++i) |
4028 | { |
4029 | position *subpos = next_position (next_ptr: subpos_ptr, base: pos, |
4030 | type: POS_XEXP, arg: i); |
4031 | switch (fmt[i]) |
4032 | { |
4033 | case 'e': case 'u': |
4034 | worklist.safe_push (obj: pattern_pos (XEXP (pattern, i), |
4035 | subpos)); |
4036 | break; |
4037 | |
4038 | case 'E': |
4039 | { |
4040 | /* Make sure the vector has the right number of |
4041 | elements. */ |
4042 | int length = XVECLEN (pattern, i); |
4043 | s = add_decision (from: s, test: rtx_test::veclen (pos), |
4044 | labels: length, optional: false); |
4045 | |
4046 | position **subpos2_ptr = &pos->xvecexp0s; |
4047 | for (int j = 0; j < length; j++) |
4048 | { |
4049 | position *subpos2 = next_position (next_ptr: subpos2_ptr, base: pos, |
4050 | type: POS_XVECEXP0, arg: j); |
4051 | rtx x = XVECEXP (pattern, i, j); |
4052 | worklist.safe_push (obj: pattern_pos (x, subpos2)); |
4053 | subpos2_ptr = &subpos2->next; |
4054 | } |
4055 | break; |
4056 | } |
4057 | |
4058 | case 'i': |
4059 | /* Make sure that XINT (X, I) has the right value. */ |
4060 | s = add_decision (from: s, test: rtx_test::int_field (pos, opno: i), |
4061 | XINT (pattern, i), optional: false); |
4062 | break; |
4063 | |
4064 | case 'r': |
4065 | /* Make sure that REGNO (X) has the right value. */ |
4066 | gcc_assert (i == 0); |
4067 | s = add_decision (from: s, test: rtx_test::regno_field (pos), |
4068 | REGNO (pattern), optional: false); |
4069 | break; |
4070 | |
4071 | case 'w': |
4072 | /* Make sure that XWINT (X, I) has the right value. */ |
4073 | s = add_decision (from: s, test: rtx_test::wide_int_field (pos, opno: i), |
4074 | XWINT (pattern, 0), optional: false); |
4075 | break; |
4076 | |
4077 | case 'p': |
4078 | /* We don't have a way of parsing polynomial offsets yet, |
4079 | and hopefully never will. */ |
4080 | s = add_decision (from: s, test: rtx_test::subreg_field (pos), |
4081 | SUBREG_BYTE (pattern).to_constant (), |
4082 | optional: false); |
4083 | break; |
4084 | |
4085 | case '0': |
4086 | break; |
4087 | |
4088 | default: |
4089 | gcc_unreachable (); |
4090 | } |
4091 | subpos_ptr = &subpos->next; |
4092 | } |
4093 | } |
4094 | break; |
4095 | } |
4096 | /* Operands are pushed onto the worklist so that later indices are |
4097 | nearer the top. That's what we want for SETs, since a SET_SRC |
4098 | is a better discriminator than a SET_DEST. In other cases it's |
4099 | usually better to match earlier indices first. This is especially |
4100 | true of PARALLELs, where the first element tends to be the most |
4101 | individual. It's also true for commutative operators, where the |
4102 | canonicalization rules say that the more complex operand should |
4103 | come first. */ |
4104 | if (code != SET && worklist.length () > reverse_s) |
4105 | std::reverse (first: &worklist[0] + reverse_s, |
4106 | last: &worklist[0] + worklist.length ()); |
4107 | } |
4108 | |
4109 | /* Sort the predicate and mode tests so that they're in depth-first order. |
4110 | The main goal of this is to put SET_SRC match_operands after SET_DEST |
4111 | match_operands and after mode checks for the enclosing SET_SRC operators |
4112 | (such as the mode of a PLUS in an addition instruction). The latter |
4113 | two types of test can determine the mode exactly, whereas a SET_SRC |
4114 | match_operand often has to cope with the possibility of the operand |
4115 | being a modeless constant integer. E.g. something that matches |
4116 | register_operand (x, SImode) never matches register_operand (x, DImode), |
4117 | but a const_int that matches immediate_operand (x, SImode) also matches |
4118 | immediate_operand (x, DImode). The register_operand cases can therefore |
4119 | be distinguished by a switch on the mode, but the immediate_operand |
4120 | cases can't. */ |
4121 | if (pred_and_mode_tests.length () > 1) |
4122 | std::sort (first: &pred_and_mode_tests[0], |
4123 | last: &pred_and_mode_tests[0] + pred_and_mode_tests.length ()); |
4124 | |
4125 | /* Add the mode and predicate tests. */ |
4126 | pattern_pos *e; |
4127 | unsigned int i; |
4128 | FOR_EACH_VEC_ELT (pred_and_mode_tests, i, e) |
4129 | { |
4130 | switch (GET_CODE (e->pattern)) |
4131 | { |
4132 | case MATCH_PARALLEL: |
4133 | case MATCH_OPERAND: |
4134 | case MATCH_SCRATCH: |
4135 | case MATCH_OPERATOR: |
4136 | { |
4137 | int opno = XINT (e->pattern, 0); |
4138 | num_operands = MAX (num_operands, opno + 1); |
4139 | const char *pred_name = predicate_name (match_rtx: e->pattern); |
4140 | if (pred_name[0]) |
4141 | { |
4142 | const struct pred_data *pred = lookup_predicate (pred_name); |
4143 | /* Check the mode first, to distinguish things like SImode |
4144 | and DImode register_operands, as described above. */ |
4145 | machine_mode mode = GET_MODE (e->pattern); |
4146 | if (pred && safe_predicate_mode (pred, mode)) |
4147 | s = add_decision (from: s, test: rtx_test::mode (pos: e->pos), labels: mode, optional: true); |
4148 | |
4149 | /* Assign to operands[] first, so that the rtx usually doesn't |
4150 | need to be live across the call to the predicate. |
4151 | |
4152 | This shouldn't cause a problem with dirtying the page, |
4153 | since we fully expect to assign to operands[] at some point, |
4154 | and since the caller usually writes to other parts of |
4155 | recog_data anyway. */ |
4156 | s = add_decision (from: s, test: rtx_test::set_op (pos: e->pos, opno), |
4157 | labels: true, optional: false); |
4158 | s = add_decision (from: s, test: rtx_test::predicate (pos: e->pos, data: pred, mode), |
4159 | labels: true, optional: false); |
4160 | } |
4161 | else |
4162 | /* Historically we've ignored the mode when there's no |
4163 | predicate. Just set up operands[] unconditionally. */ |
4164 | s = add_decision (from: s, test: rtx_test::set_op (pos: e->pos, opno), |
4165 | labels: true, optional: false); |
4166 | break; |
4167 | } |
4168 | |
4169 | default: |
4170 | s = add_decision (from: s, test: rtx_test::mode (pos: e->pos), |
4171 | GET_MODE (e->pattern), optional: false); |
4172 | break; |
4173 | } |
4174 | } |
4175 | |
4176 | /* Finally add rtx_equal_p checks for duplicated operands. */ |
4177 | FOR_EACH_VEC_ELT (dup_tests, i, e) |
4178 | s = add_decision (from: s, test: rtx_test::duplicate (pos: e->pos, XINT (e->pattern, 0)), |
4179 | labels: true, optional: false); |
4180 | return s; |
4181 | } |
4182 | |
4183 | /* Add new decisions to S that make it return ACCEPTANCE if: |
4184 | |
4185 | (1) the rtx doesn't match anything already matched by S |
4186 | (2) the rtx matches TOP_PATTERN and |
4187 | (3) the C test required by INFO->def is true |
4188 | |
4189 | For peephole2, TOP_PATTERN is a SEQUENCE of the instruction patterns |
4190 | to match, otherwise it is a single instruction pattern. */ |
4191 | |
4192 | static void |
4193 | match_pattern_1 (state *s, md_rtx_info *info, rtx pattern, |
4194 | acceptance_type acceptance) |
4195 | { |
4196 | if (acceptance.type == PEEPHOLE2) |
4197 | { |
4198 | /* Match each individual instruction. */ |
4199 | position **subpos_ptr = &peep2_insn_pos_list; |
4200 | int count = 0; |
4201 | for (int i = 0; i < XVECLEN (pattern, 0); ++i) |
4202 | { |
4203 | rtx x = XVECEXP (pattern, 0, i); |
4204 | position *subpos = next_position (next_ptr: subpos_ptr, base: &root_pos, |
4205 | type: POS_PEEP2_INSN, arg: count); |
4206 | if (count > 0) |
4207 | s = add_decision (from: s, test: rtx_test::peep2_count (min_len: count + 1), |
4208 | labels: true, optional: false); |
4209 | s = match_pattern_2 (s, info, pos: subpos, pattern: x); |
4210 | subpos_ptr = &subpos->next; |
4211 | count += 1; |
4212 | } |
4213 | acceptance.u.full.u.match_len = count - 1; |
4214 | } |
4215 | else |
4216 | { |
4217 | /* Make the rtx itself. */ |
4218 | s = match_pattern_2 (s, info, pos: &root_pos, pattern); |
4219 | |
4220 | /* If the match is only valid when extra clobbers are added, |
4221 | make sure we're able to pass that information to the caller. */ |
4222 | if (acceptance.type == RECOG && acceptance.u.full.u.num_clobbers) |
4223 | s = add_decision (from: s, test: rtx_test::have_num_clobbers (), labels: true, optional: false); |
4224 | } |
4225 | |
4226 | /* Make sure that the C test is true. */ |
4227 | const char *c_test = get_c_test (info->def); |
4228 | if (maybe_eval_c_test (c_test) != 1) |
4229 | s = add_decision (from: s, test: rtx_test::c_test (string: c_test), labels: true, optional: false); |
4230 | |
4231 | /* Accept the pattern. */ |
4232 | add_decision (from: s, test: rtx_test::accept (acceptance), labels: true, optional: false); |
4233 | } |
4234 | |
4235 | /* Like match_pattern_1, but (if merge_states_p) try to merge the |
4236 | decisions with what's already in S, to reduce the amount of |
4237 | backtracking. */ |
4238 | |
4239 | static void |
4240 | match_pattern (state *s, md_rtx_info *info, rtx pattern, |
4241 | acceptance_type acceptance) |
4242 | { |
4243 | if (merge_states_p) |
4244 | { |
4245 | state root; |
4246 | /* Add the decisions to a fresh state and then merge the full tree |
4247 | into the existing one. */ |
4248 | match_pattern_1 (s: &root, info, pattern, acceptance); |
4249 | merge_into_state (s1: s, s2: &root); |
4250 | } |
4251 | else |
4252 | match_pattern_1 (s, info, pattern, acceptance); |
4253 | } |
4254 | |
4255 | /* Begin the output file. */ |
4256 | |
4257 | static void |
4258 | (void) |
4259 | { |
4260 | puts (s: "\ |
4261 | /* Generated automatically by the program `genrecog' from the target\n\ |
4262 | machine description file. */\n\ |
4263 | \n\ |
4264 | #define IN_TARGET_CODE 1\n\ |
4265 | \n\ |
4266 | #include \"config.h\"\n\ |
4267 | #include \"system.h\"\n\ |
4268 | #include \"coretypes.h\"\n\ |
4269 | #include \"backend.h\"\n\ |
4270 | #include \"predict.h\"\n\ |
4271 | #include \"rtl.h\"\n\ |
4272 | #include \"memmodel.h\"\n\ |
4273 | #include \"tm_p.h\"\n\ |
4274 | #include \"emit-rtl.h\"\n\ |
4275 | #include \"insn-config.h\"\n\ |
4276 | #include \"recog.h\"\n\ |
4277 | #include \"output.h\"\n\ |
4278 | #include \"flags.h\"\n\ |
4279 | #include \"df.h\"\n\ |
4280 | #include \"resource.h\"\n\ |
4281 | #include \"diagnostic-core.h\"\n\ |
4282 | #include \"reload.h\"\n\ |
4283 | #include \"regs.h\"\n\ |
4284 | #include \"tm-constrs.h\"\n\ |
4285 | \n" ); |
4286 | |
4287 | puts (s: "\n\ |
4288 | /* `recog' contains a decision tree that recognizes whether the rtx\n\ |
4289 | X0 is a valid instruction.\n\ |
4290 | \n\ |
4291 | recog returns -1 if the rtx is not valid. If the rtx is valid, recog\n\ |
4292 | returns a nonnegative number which is the insn code number for the\n\ |
4293 | pattern that matched. This is the same as the order in the machine\n\ |
4294 | description of the entry that matched. This number can be used as an\n\ |
4295 | index into `insn_data' and other tables.\n" ); |
4296 | puts (s: "\ |
4297 | The third parameter to recog is an optional pointer to an int. If\n\ |
4298 | present, recog will accept a pattern if it matches except for missing\n\ |
4299 | CLOBBER expressions at the end. In that case, the value pointed to by\n\ |
4300 | the optional pointer will be set to the number of CLOBBERs that need\n\ |
4301 | to be added (it should be initialized to zero by the caller). If it" ); |
4302 | puts (s: "\ |
4303 | is set nonzero, the caller should allocate a PARALLEL of the\n\ |
4304 | appropriate size, copy the initial entries, and call add_clobbers\n\ |
4305 | (found in insn-emit.cc) to fill in the CLOBBERs.\n\ |
4306 | " ); |
4307 | |
4308 | puts (s: "\n\ |
4309 | The function split_insns returns 0 if the rtl could not\n\ |
4310 | be split or the split rtl as an INSN list if it can be.\n\ |
4311 | \n\ |
4312 | The function peephole2_insns returns 0 if the rtl could not\n\ |
4313 | be matched. If there was a match, the new rtl is returned in an INSN list,\n\ |
4314 | and LAST_INSN will point to the last recognized insn in the old sequence.\n\ |
4315 | */\n\n" ); |
4316 | } |
4317 | |
4318 | /* Return the C type of a parameter with type TYPE. */ |
4319 | |
4320 | static const char * |
4321 | parameter_type_string (parameter::type_enum type) |
4322 | { |
4323 | switch (type) |
4324 | { |
4325 | case parameter::UNSET: |
4326 | break; |
4327 | |
4328 | case parameter::CODE: |
4329 | return "rtx_code" ; |
4330 | |
4331 | case parameter::MODE: |
4332 | return "machine_mode" ; |
4333 | |
4334 | case parameter::INT: |
4335 | return "int" ; |
4336 | |
4337 | case parameter::UINT: |
4338 | return "unsigned int" ; |
4339 | |
4340 | case parameter::WIDE_INT: |
4341 | return "HOST_WIDE_INT" ; |
4342 | } |
4343 | gcc_unreachable (); |
4344 | } |
4345 | |
4346 | /* Return true if ACCEPTANCE requires only a single C statement even in |
4347 | a backtracking context. */ |
4348 | |
4349 | static bool |
4350 | single_statement_p (const acceptance_type &acceptance) |
4351 | { |
4352 | if (acceptance.partial_p) |
4353 | /* We need to handle failures of the subroutine. */ |
4354 | return false; |
4355 | switch (acceptance.type) |
4356 | { |
4357 | case SUBPATTERN: |
4358 | case SPLIT: |
4359 | return true; |
4360 | |
4361 | case RECOG: |
4362 | /* False if we need to assign to pnum_clobbers. */ |
4363 | return acceptance.u.full.u.num_clobbers == 0; |
4364 | |
4365 | case PEEPHOLE2: |
4366 | /* We need to assign to pmatch_len_ and handle null returns from the |
4367 | peephole2 routine. */ |
4368 | return false; |
4369 | } |
4370 | gcc_unreachable (); |
4371 | } |
4372 | |
4373 | /* Return the C failure value for a routine of type TYPE. */ |
4374 | |
4375 | static const char * |
4376 | get_failure_return (routine_type type) |
4377 | { |
4378 | switch (type) |
4379 | { |
4380 | case SUBPATTERN: |
4381 | case RECOG: |
4382 | return "-1" ; |
4383 | |
4384 | case SPLIT: |
4385 | case PEEPHOLE2: |
4386 | return "NULL" ; |
4387 | } |
4388 | gcc_unreachable (); |
4389 | } |
4390 | |
4391 | /* Indicates whether a block of code always returns or whether it can fall |
4392 | through. */ |
4393 | |
4394 | enum exit_state { |
4395 | ES_RETURNED, |
4396 | ES_FALLTHROUGH |
4397 | }; |
4398 | |
4399 | /* Information used while writing out code. */ |
4400 | |
4401 | class output_state |
4402 | { |
4403 | public: |
4404 | /* The type of routine that we're generating. */ |
4405 | routine_type type; |
4406 | |
4407 | /* Maps position ids to xN variable numbers. The entry is only valid if |
4408 | it is less than the length of VAR_TO_ID, but this holds for every position |
4409 | tested by a state when writing out that state. */ |
4410 | auto_vec <unsigned int> id_to_var; |
4411 | |
4412 | /* Maps xN variable numbers to position ids. */ |
4413 | auto_vec <unsigned int> var_to_id; |
4414 | |
4415 | /* Index N is true if variable xN has already been set. */ |
4416 | auto_vec <bool> seen_vars; |
4417 | }; |
4418 | |
4419 | /* Return true if D is a call to a pattern routine and if there is some X |
4420 | such that the transition for pattern result N goes to a successful return |
4421 | with code X+N. When returning true, set *BASE_OUT to this X and *COUNT_OUT |
4422 | to the number of return values. (We know that every PATTERN decision has |
4423 | a transition for every successful return.) */ |
4424 | |
4425 | static bool |
4426 | terminal_pattern_p (decision *d, unsigned int *base_out, |
4427 | unsigned int *count_out) |
4428 | { |
4429 | if (d->test.kind != rtx_test::PATTERN) |
4430 | return false; |
4431 | unsigned int base = 0; |
4432 | unsigned int count = 0; |
4433 | for (transition *trans = d->first; trans; trans = trans->next) |
4434 | { |
4435 | if (trans->is_param || trans->labels.length () != 1) |
4436 | return false; |
4437 | decision *subd = trans->to->singleton (); |
4438 | if (!subd || subd->test.kind != rtx_test::ACCEPT) |
4439 | return false; |
4440 | unsigned int this_base = (subd->test.u.acceptance.u.full.code |
4441 | - trans->labels[0]); |
4442 | if (trans == d->first) |
4443 | base = this_base; |
4444 | else if (base != this_base) |
4445 | return false; |
4446 | count += 1; |
4447 | } |
4448 | *base_out = base; |
4449 | *count_out = count; |
4450 | return true; |
4451 | } |
4452 | |
4453 | /* Return true if TEST doesn't test an rtx or if the rtx it tests is |
4454 | already available in state OS. */ |
4455 | |
4456 | static bool |
4457 | test_position_available_p (output_state *os, const rtx_test &test) |
4458 | { |
4459 | return (!test.pos |
4460 | || test.pos_operand >= 0 |
4461 | || os->seen_vars[os->id_to_var[test.pos->id]]); |
4462 | } |
4463 | |
4464 | /* Like printf, but print INDENT spaces at the beginning. */ |
4465 | |
4466 | static void ATTRIBUTE_PRINTF_2 |
4467 | printf_indent (unsigned int indent, const char *format, ...) |
4468 | { |
4469 | va_list ap; |
4470 | va_start (ap, format); |
4471 | printf (format: "%*s" , indent, "" ); |
4472 | vprintf (fmt: format, arg: ap); |
4473 | va_end (ap); |
4474 | } |
4475 | |
4476 | /* Emit code to initialize the variable associated with POS, if it isn't |
4477 | already valid in state OS. Indent each line by INDENT spaces. Update |
4478 | OS with the new state. */ |
4479 | |
4480 | static void |
4481 | change_state (output_state *os, position *pos, unsigned int indent) |
4482 | { |
4483 | unsigned int var = os->id_to_var[pos->id]; |
4484 | gcc_assert (var < os->var_to_id.length () && os->var_to_id[var] == pos->id); |
4485 | if (os->seen_vars[var]) |
4486 | return; |
4487 | switch (pos->type) |
4488 | { |
4489 | case POS_PEEP2_INSN: |
4490 | printf_indent (indent, format: "x%d = PATTERN (peep2_next_insn (%d));\n" , |
4491 | var, pos->arg); |
4492 | break; |
4493 | |
4494 | case POS_XEXP: |
4495 | change_state (os, pos: pos->base, indent); |
4496 | printf_indent (indent, format: "x%d = XEXP (x%d, %d);\n" , |
4497 | var, os->id_to_var[pos->base->id], pos->arg); |
4498 | break; |
4499 | |
4500 | case POS_XVECEXP0: |
4501 | change_state (os, pos: pos->base, indent); |
4502 | printf_indent (indent, format: "x%d = XVECEXP (x%d, 0, %d);\n" , |
4503 | var, os->id_to_var[pos->base->id], pos->arg); |
4504 | break; |
4505 | } |
4506 | os->seen_vars[var] = true; |
4507 | } |
4508 | |
4509 | /* Print the enumerator constant for CODE -- the upcase version of |
4510 | the name. */ |
4511 | |
4512 | static void |
4513 | print_code (enum rtx_code code) |
4514 | { |
4515 | const char *p; |
4516 | for (p = GET_RTX_NAME (code); *p; p++) |
4517 | putchar (TOUPPER (*p)); |
4518 | } |
4519 | |
4520 | /* Emit a uint64_t as an integer constant expression. We need to take |
4521 | special care to avoid "decimal constant is so large that it is unsigned" |
4522 | warnings in the resulting code. */ |
4523 | |
4524 | static void |
4525 | print_host_wide_int (uint64_t val) |
4526 | { |
4527 | uint64_t min = uint64_t (1) << (HOST_BITS_PER_WIDE_INT - 1); |
4528 | if (val == min) |
4529 | printf (format: "(" HOST_WIDE_INT_PRINT_DEC_C " - 1)" , val + 1); |
4530 | else |
4531 | printf (HOST_WIDE_INT_PRINT_DEC_C, val); |
4532 | } |
4533 | |
4534 | /* Print the C expression for actual parameter PARAM. */ |
4535 | |
4536 | static void |
4537 | print_parameter_value (const parameter ¶m) |
4538 | { |
4539 | if (param.is_param) |
4540 | printf (format: "i%d" , (int) param.value + 1); |
4541 | else |
4542 | switch (param.type) |
4543 | { |
4544 | case parameter::UNSET: |
4545 | gcc_unreachable (); |
4546 | break; |
4547 | |
4548 | case parameter::CODE: |
4549 | print_code (code: (enum rtx_code) param.value); |
4550 | break; |
4551 | |
4552 | case parameter::MODE: |
4553 | printf (format: "E_%smode" , GET_MODE_NAME ((machine_mode) param.value)); |
4554 | break; |
4555 | |
4556 | case parameter::INT: |
4557 | printf (format: "%d" , (int) param.value); |
4558 | break; |
4559 | |
4560 | case parameter::UINT: |
4561 | printf (format: "%u" , (unsigned int) param.value); |
4562 | break; |
4563 | |
4564 | case parameter::WIDE_INT: |
4565 | print_host_wide_int (val: param.value); |
4566 | break; |
4567 | } |
4568 | } |
4569 | |
4570 | /* Print the C expression for the rtx tested by TEST. */ |
4571 | |
4572 | static void |
4573 | print_test_rtx (output_state *os, const rtx_test &test) |
4574 | { |
4575 | if (test.pos_operand >= 0) |
4576 | printf (format: "operands[%d]" , test.pos_operand); |
4577 | else |
4578 | printf (format: "x%d" , os->id_to_var[test.pos->id]); |
4579 | } |
4580 | |
4581 | /* Print the C expression for non-boolean test TEST. */ |
4582 | |
4583 | static void |
4584 | print_nonbool_test (output_state *os, const rtx_test &test) |
4585 | { |
4586 | switch (test.kind) |
4587 | { |
4588 | case rtx_test::CODE: |
4589 | printf (format: "GET_CODE (" ); |
4590 | print_test_rtx (os, test); |
4591 | printf (format: ")" ); |
4592 | break; |
4593 | |
4594 | case rtx_test::MODE: |
4595 | printf (format: "GET_MODE (" ); |
4596 | print_test_rtx (os, test); |
4597 | printf (format: ")" ); |
4598 | break; |
4599 | |
4600 | case rtx_test::VECLEN: |
4601 | printf (format: "XVECLEN (" ); |
4602 | print_test_rtx (os, test); |
4603 | printf (format: ", 0)" ); |
4604 | break; |
4605 | |
4606 | case rtx_test::INT_FIELD: |
4607 | printf (format: "XINT (" ); |
4608 | print_test_rtx (os, test); |
4609 | printf (format: ", %d)" , test.u.opno); |
4610 | break; |
4611 | |
4612 | case rtx_test::REGNO_FIELD: |
4613 | printf (format: "REGNO (" ); |
4614 | print_test_rtx (os, test); |
4615 | printf (format: ")" ); |
4616 | break; |
4617 | |
4618 | case rtx_test::SUBREG_FIELD: |
4619 | printf (format: "SUBREG_BYTE (" ); |
4620 | print_test_rtx (os, test); |
4621 | printf (format: ")" ); |
4622 | break; |
4623 | |
4624 | case rtx_test::WIDE_INT_FIELD: |
4625 | printf (format: "XWINT (" ); |
4626 | print_test_rtx (os, test); |
4627 | printf (format: ", %d)" , test.u.opno); |
4628 | break; |
4629 | |
4630 | case rtx_test::PATTERN: |
4631 | { |
4632 | pattern_routine *routine = test.u.pattern->routine; |
4633 | printf (format: "pattern%d (" , routine->pattern_id); |
4634 | const char *sep = "" ; |
4635 | if (test.pos) |
4636 | { |
4637 | print_test_rtx (os, test); |
4638 | sep = ", " ; |
4639 | } |
4640 | if (routine->insn_p) |
4641 | { |
4642 | printf (format: "%sinsn" , sep); |
4643 | sep = ", " ; |
4644 | } |
4645 | if (routine->pnum_clobbers_p) |
4646 | { |
4647 | printf (format: "%spnum_clobbers" , sep); |
4648 | sep = ", " ; |
4649 | } |
4650 | for (unsigned int i = 0; i < test.u.pattern->params.length (); ++i) |
4651 | { |
4652 | fputs (sep, stdout); |
4653 | print_parameter_value (param: test.u.pattern->params[i]); |
4654 | sep = ", " ; |
4655 | } |
4656 | printf (format: ")" ); |
4657 | break; |
4658 | } |
4659 | |
4660 | case rtx_test::PEEP2_COUNT: |
4661 | case rtx_test::VECLEN_GE: |
4662 | case rtx_test::SAVED_CONST_INT: |
4663 | case rtx_test::DUPLICATE: |
4664 | case rtx_test::PREDICATE: |
4665 | case rtx_test::SET_OP: |
4666 | case rtx_test::HAVE_NUM_CLOBBERS: |
4667 | case rtx_test::C_TEST: |
4668 | case rtx_test::ACCEPT: |
4669 | gcc_unreachable (); |
4670 | } |
4671 | } |
4672 | |
4673 | /* IS_PARAM and LABEL are taken from a transition whose source |
4674 | decision performs TEST. Print the C code for the label. */ |
4675 | |
4676 | static void |
4677 | print_label_value (const rtx_test &test, bool is_param, uint64_t value) |
4678 | { |
4679 | print_parameter_value (param: parameter (transition_parameter_type (kind: test.kind), |
4680 | is_param, value)); |
4681 | } |
4682 | |
4683 | /* If IS_PARAM, print code to compare TEST with the C variable i<VALUE+1>. |
4684 | If !IS_PARAM, print code to compare TEST with the C constant VALUE. |
4685 | Test for inequality if INVERT_P, otherwise test for equality. */ |
4686 | |
4687 | static void |
4688 | print_test (output_state *os, const rtx_test &test, bool is_param, |
4689 | uint64_t value, bool invert_p) |
4690 | { |
4691 | switch (test.kind) |
4692 | { |
4693 | /* Handle the non-boolean TESTs. */ |
4694 | case rtx_test::CODE: |
4695 | case rtx_test::MODE: |
4696 | case rtx_test::VECLEN: |
4697 | case rtx_test::REGNO_FIELD: |
4698 | case rtx_test::INT_FIELD: |
4699 | case rtx_test::WIDE_INT_FIELD: |
4700 | case rtx_test::PATTERN: |
4701 | print_nonbool_test (os, test); |
4702 | printf (format: " %s " , invert_p ? "!=" : "==" ); |
4703 | print_label_value (test, is_param, value); |
4704 | break; |
4705 | |
4706 | case rtx_test::SUBREG_FIELD: |
4707 | printf (format: "%s (" , invert_p ? "maybe_ne" : "known_eq" ); |
4708 | print_nonbool_test (os, test); |
4709 | printf (format: ", " ); |
4710 | print_label_value (test, is_param, value); |
4711 | printf (format: ")" ); |
4712 | break; |
4713 | |
4714 | case rtx_test::SAVED_CONST_INT: |
4715 | gcc_assert (!is_param && value == 1); |
4716 | print_test_rtx (os, test); |
4717 | printf (format: " %s const_int_rtx[MAX_SAVED_CONST_INT + " , |
4718 | invert_p ? "!=" : "==" ); |
4719 | print_parameter_value (param: parameter (parameter::INT, |
4720 | test.u.integer.is_param, |
4721 | test.u.integer.value)); |
4722 | printf (format: "]" ); |
4723 | break; |
4724 | |
4725 | case rtx_test::PEEP2_COUNT: |
4726 | gcc_assert (!is_param && value == 1); |
4727 | printf (format: "peep2_current_count %s %d" , invert_p ? "<" : ">=" , |
4728 | test.u.min_len); |
4729 | break; |
4730 | |
4731 | case rtx_test::VECLEN_GE: |
4732 | gcc_assert (!is_param && value == 1); |
4733 | printf (format: "XVECLEN (" ); |
4734 | print_test_rtx (os, test); |
4735 | printf (format: ", 0) %s %d" , invert_p ? "<" : ">=" , test.u.min_len); |
4736 | break; |
4737 | |
4738 | case rtx_test::PREDICATE: |
4739 | gcc_assert (!is_param && value == 1); |
4740 | printf (format: "%s%s (" , invert_p ? "!" : "" , test.u.predicate.data->name); |
4741 | print_test_rtx (os, test); |
4742 | printf (format: ", " ); |
4743 | print_parameter_value (param: parameter (parameter::MODE, |
4744 | test.u.predicate.mode_is_param, |
4745 | test.u.predicate.mode)); |
4746 | printf (format: ")" ); |
4747 | break; |
4748 | |
4749 | case rtx_test::DUPLICATE: |
4750 | gcc_assert (!is_param && value == 1); |
4751 | printf (format: "%srtx_equal_p (" , invert_p ? "!" : "" ); |
4752 | print_test_rtx (os, test); |
4753 | printf (format: ", operands[%d])" , test.u.opno); |
4754 | break; |
4755 | |
4756 | case rtx_test::HAVE_NUM_CLOBBERS: |
4757 | gcc_assert (!is_param && value == 1); |
4758 | printf (format: "pnum_clobbers %s NULL" , invert_p ? "==" : "!=" ); |
4759 | break; |
4760 | |
4761 | case rtx_test::C_TEST: |
4762 | gcc_assert (!is_param && value == 1); |
4763 | if (invert_p) |
4764 | printf (format: "!" ); |
4765 | rtx_reader_ptr->print_c_condition (cond: test.u.string); |
4766 | break; |
4767 | |
4768 | case rtx_test::ACCEPT: |
4769 | case rtx_test::SET_OP: |
4770 | gcc_unreachable (); |
4771 | } |
4772 | } |
4773 | |
4774 | static exit_state print_decision (output_state *, decision *, |
4775 | unsigned int, bool); |
4776 | |
4777 | /* Print code to perform S, indent each line by INDENT spaces. |
4778 | IS_FINAL is true if there are no fallback decisions to test on failure; |
4779 | if the state fails then the entire routine fails. */ |
4780 | |
4781 | static exit_state |
4782 | print_state (output_state *os, state *s, unsigned int indent, bool is_final) |
4783 | { |
4784 | exit_state es = ES_FALLTHROUGH; |
4785 | for (decision *d = s->first; d; d = d->next) |
4786 | es = print_decision (os, d, indent, is_final && !d->next); |
4787 | if (es != ES_RETURNED && is_final) |
4788 | { |
4789 | printf_indent (indent, format: "return %s;\n" , get_failure_return (type: os->type)); |
4790 | es = ES_RETURNED; |
4791 | } |
4792 | return es; |
4793 | } |
4794 | |
4795 | /* Print the code for subroutine call ACCEPTANCE (for which partial_p |
4796 | is known to be true). Return the C condition that indicates a successful |
4797 | match. */ |
4798 | |
4799 | static const char * |
4800 | print_subroutine_call (const acceptance_type &acceptance) |
4801 | { |
4802 | switch (acceptance.type) |
4803 | { |
4804 | case SUBPATTERN: |
4805 | gcc_unreachable (); |
4806 | |
4807 | case RECOG: |
4808 | printf (format: "recog_%d (x1, insn, pnum_clobbers)" , |
4809 | acceptance.u.subroutine_id); |
4810 | return ">= 0" ; |
4811 | |
4812 | case SPLIT: |
4813 | printf (format: "split_%d (x1, insn)" , acceptance.u.subroutine_id); |
4814 | return "!= NULL_RTX" ; |
4815 | |
4816 | case PEEPHOLE2: |
4817 | printf (format: "peephole2_%d (x1, insn, pmatch_len_)" , |
4818 | acceptance.u.subroutine_id); |
4819 | return "!= NULL_RTX" ; |
4820 | } |
4821 | gcc_unreachable (); |
4822 | } |
4823 | |
4824 | /* Print code for the successful match described by ACCEPTANCE. |
4825 | INDENT and IS_FINAL are as for print_state. */ |
4826 | |
4827 | static exit_state |
4828 | print_acceptance (const acceptance_type &acceptance, unsigned int indent, |
4829 | bool is_final) |
4830 | { |
4831 | if (acceptance.partial_p) |
4832 | { |
4833 | /* Defer the rest of the match to a subroutine. */ |
4834 | if (is_final) |
4835 | { |
4836 | printf_indent (indent, format: "return " ); |
4837 | print_subroutine_call (acceptance); |
4838 | printf (format: ";\n" ); |
4839 | return ES_RETURNED; |
4840 | } |
4841 | else |
4842 | { |
4843 | printf_indent (indent, format: "res = " ); |
4844 | const char *res_test = print_subroutine_call (acceptance); |
4845 | printf (format: ";\n" ); |
4846 | printf_indent (indent, format: "if (res %s)\n" , res_test); |
4847 | printf_indent (indent: indent + 2, format: "return res;\n" ); |
4848 | return ES_FALLTHROUGH; |
4849 | } |
4850 | } |
4851 | switch (acceptance.type) |
4852 | { |
4853 | case SUBPATTERN: |
4854 | printf_indent (indent, format: "return %d;\n" , acceptance.u.full.code); |
4855 | return ES_RETURNED; |
4856 | |
4857 | case RECOG: |
4858 | if (acceptance.u.full.u.num_clobbers != 0) |
4859 | printf_indent (indent, format: "*pnum_clobbers = %d;\n" , |
4860 | acceptance.u.full.u.num_clobbers); |
4861 | printf_indent (indent, format: "return %d; /* %s */\n" , acceptance.u.full.code, |
4862 | get_insn_name (acceptance.u.full.code)); |
4863 | return ES_RETURNED; |
4864 | |
4865 | case SPLIT: |
4866 | printf_indent (indent, format: "return gen_split_%d (insn, operands);\n" , |
4867 | acceptance.u.full.code); |
4868 | return ES_RETURNED; |
4869 | |
4870 | case PEEPHOLE2: |
4871 | printf_indent (indent, format: "*pmatch_len_ = %d;\n" , |
4872 | acceptance.u.full.u.match_len); |
4873 | if (is_final) |
4874 | { |
4875 | printf_indent (indent, format: "return gen_peephole2_%d (insn, operands);\n" , |
4876 | acceptance.u.full.code); |
4877 | return ES_RETURNED; |
4878 | } |
4879 | else |
4880 | { |
4881 | printf_indent (indent, format: "res = gen_peephole2_%d (insn, operands);\n" , |
4882 | acceptance.u.full.code); |
4883 | printf_indent (indent, format: "if (res != NULL_RTX)\n" ); |
4884 | printf_indent (indent: indent + 2, format: "return res;\n" ); |
4885 | return ES_FALLTHROUGH; |
4886 | } |
4887 | } |
4888 | gcc_unreachable (); |
4889 | } |
4890 | |
4891 | /* Print code to perform D. INDENT and IS_FINAL are as for print_state. */ |
4892 | |
4893 | static exit_state |
4894 | print_decision (output_state *os, decision *d, unsigned int indent, |
4895 | bool is_final) |
4896 | { |
4897 | uint64_t label; |
4898 | unsigned int base, count; |
4899 | |
4900 | /* Make sure the rtx under test is available either in operands[] or |
4901 | in an xN variable. */ |
4902 | if (d->test.pos && d->test.pos_operand < 0) |
4903 | change_state (os, pos: d->test.pos, indent); |
4904 | |
4905 | /* Look for cases where a pattern routine P1 calls another pattern routine |
4906 | P2 and where P1 returns X + BASE whenever P2 returns X. If IS_FINAL |
4907 | is true and BASE is zero we can simply use: |
4908 | |
4909 | return patternN (...); |
4910 | |
4911 | Otherwise we can use: |
4912 | |
4913 | res = patternN (...); |
4914 | if (res >= 0) |
4915 | return res + BASE; |
4916 | |
4917 | However, if BASE is nonzero and patternN only returns 0 or -1, |
4918 | the usual "return BASE;" is better than "return res + BASE;". |
4919 | If BASE is zero, "return res;" should be better than "return 0;", |
4920 | since no assignment to the return register is required. */ |
4921 | if (os->type == SUBPATTERN |
4922 | && terminal_pattern_p (d, base_out: &base, count_out: &count) |
4923 | && (base == 0 || count > 1)) |
4924 | { |
4925 | if (is_final && base == 0) |
4926 | { |
4927 | printf_indent (indent, format: "return " ); |
4928 | print_nonbool_test (os, test: d->test); |
4929 | printf (format: "; /* [-1, %d] */\n" , count - 1); |
4930 | return ES_RETURNED; |
4931 | } |
4932 | else |
4933 | { |
4934 | printf_indent (indent, format: "res = " ); |
4935 | print_nonbool_test (os, test: d->test); |
4936 | printf (format: ";\n" ); |
4937 | printf_indent (indent, format: "if (res >= 0)\n" ); |
4938 | printf_indent (indent: indent + 2, format: "return res" ); |
4939 | if (base != 0) |
4940 | printf (format: " + %d" , base); |
4941 | printf (format: "; /* [%d, %d] */\n" , base, base + count - 1); |
4942 | return ES_FALLTHROUGH; |
4943 | } |
4944 | } |
4945 | else if (d->test.kind == rtx_test::ACCEPT) |
4946 | return print_acceptance (acceptance: d->test.u.acceptance, indent, is_final); |
4947 | else if (d->test.kind == rtx_test::SET_OP) |
4948 | { |
4949 | printf_indent (indent, format: "operands[%d] = " , d->test.u.opno); |
4950 | print_test_rtx (os, test: d->test); |
4951 | printf (format: ";\n" ); |
4952 | return print_state (os, s: d->singleton ()->to, indent, is_final); |
4953 | } |
4954 | /* Handle decisions with a single transition and a single transition |
4955 | label. */ |
4956 | else if (d->if_statement_p (label: &label)) |
4957 | { |
4958 | transition *trans = d->singleton (); |
4959 | if (mark_optional_transitions_p && trans->optional) |
4960 | printf_indent (indent, format: "/* OPTIONAL IF */\n" ); |
4961 | |
4962 | /* Print the condition associated with TRANS. Invert it if IS_FINAL, |
4963 | so that we return immediately on failure and fall through on |
4964 | success. */ |
4965 | printf_indent (indent, format: "if (" ); |
4966 | print_test (os, test: d->test, is_param: trans->is_param, value: label, invert_p: is_final); |
4967 | |
4968 | /* Look for following states that would be handled by this code |
4969 | on recursion. If they don't need any preparatory statements, |
4970 | include them in the current "if" statement rather than creating |
4971 | a new one. */ |
4972 | for (;;) |
4973 | { |
4974 | d = trans->to->singleton (); |
4975 | if (!d |
4976 | || d->test.kind == rtx_test::ACCEPT |
4977 | || d->test.kind == rtx_test::SET_OP |
4978 | || !d->if_statement_p (label: &label) |
4979 | || !test_position_available_p (os, test: d->test)) |
4980 | break; |
4981 | trans = d->first; |
4982 | printf (format: "\n" ); |
4983 | if (mark_optional_transitions_p && trans->optional) |
4984 | printf_indent (indent: indent + 4, format: "/* OPTIONAL IF */\n" ); |
4985 | printf_indent (indent: indent + 4, format: "%s " , is_final ? "||" : "&&" ); |
4986 | print_test (os, test: d->test, is_param: trans->is_param, value: label, invert_p: is_final); |
4987 | } |
4988 | printf (format: ")\n" ); |
4989 | |
4990 | /* Print the conditional code with INDENT + 2 and the fallthrough |
4991 | code with indent INDENT. */ |
4992 | state *to = trans->to; |
4993 | if (is_final) |
4994 | { |
4995 | /* We inverted the condition above, so return failure in the |
4996 | "if" body and fall through to the target of the transition. */ |
4997 | printf_indent (indent: indent + 2, format: "return %s;\n" , |
4998 | get_failure_return (type: os->type)); |
4999 | return print_state (os, s: to, indent, is_final); |
5000 | } |
5001 | else if (to->singleton () |
5002 | && to->first->test.kind == rtx_test::ACCEPT |
5003 | && single_statement_p (acceptance: to->first->test.u.acceptance)) |
5004 | { |
5005 | /* The target of the transition is a simple "return" statement. |
5006 | It doesn't need any braces and doesn't fall through. */ |
5007 | if (print_acceptance (acceptance: to->first->test.u.acceptance, |
5008 | indent: indent + 2, is_final: true) != ES_RETURNED) |
5009 | gcc_unreachable (); |
5010 | return ES_FALLTHROUGH; |
5011 | } |
5012 | else |
5013 | { |
5014 | /* The general case. Output code for the target of the transition |
5015 | in braces. This will not invalidate any of the xN variables |
5016 | that are already valid, but we mustn't rely on any that are |
5017 | set by the "if" body. */ |
5018 | auto_vec <bool, 32> old_seen; |
5019 | old_seen.safe_splice (src: os->seen_vars); |
5020 | |
5021 | printf_indent (indent: indent + 2, format: "{\n" ); |
5022 | print_state (os, s: trans->to, indent: indent + 4, is_final); |
5023 | printf_indent (indent: indent + 2, format: "}\n" ); |
5024 | |
5025 | os->seen_vars.truncate (size: 0); |
5026 | os->seen_vars.splice (src: old_seen); |
5027 | return ES_FALLTHROUGH; |
5028 | } |
5029 | } |
5030 | else |
5031 | { |
5032 | /* Output the decision as a switch statement. */ |
5033 | printf_indent (indent, format: "switch (" ); |
5034 | print_nonbool_test (os, test: d->test); |
5035 | printf (format: ")\n" ); |
5036 | |
5037 | /* Each case statement starts with the same set of valid variables. |
5038 | These are also the only variables will be valid on fallthrough. */ |
5039 | auto_vec <bool, 32> old_seen; |
5040 | old_seen.safe_splice (src: os->seen_vars); |
5041 | |
5042 | printf_indent (indent: indent + 2, format: "{\n" ); |
5043 | for (transition *trans = d->first; trans; trans = trans->next) |
5044 | { |
5045 | gcc_assert (!trans->is_param); |
5046 | if (mark_optional_transitions_p && trans->optional) |
5047 | printf_indent (indent: indent + 2, format: "/* OPTIONAL CASE */\n" ); |
5048 | for (int_set::iterator j = trans->labels.begin (); |
5049 | j != trans->labels.end (); ++j) |
5050 | { |
5051 | printf_indent (indent: indent + 2, format: "case " ); |
5052 | print_label_value (test: d->test, is_param: trans->is_param, value: *j); |
5053 | printf (format: ":\n" ); |
5054 | } |
5055 | if (print_state (os, s: trans->to, indent: indent + 4, is_final)) |
5056 | { |
5057 | /* The state can fall through. Add an explicit break. */ |
5058 | gcc_assert (!is_final); |
5059 | printf_indent (indent: indent + 4, format: "break;\n" ); |
5060 | } |
5061 | printf (format: "\n" ); |
5062 | |
5063 | /* Restore the original set of valid variables. */ |
5064 | os->seen_vars.truncate (size: 0); |
5065 | os->seen_vars.splice (src: old_seen); |
5066 | } |
5067 | /* Add a default case. */ |
5068 | printf_indent (indent: indent + 2, format: "default:\n" ); |
5069 | if (is_final) |
5070 | printf_indent (indent: indent + 4, format: "return %s;\n" , |
5071 | get_failure_return (type: os->type)); |
5072 | else |
5073 | printf_indent (indent: indent + 4, format: "break;\n" ); |
5074 | printf_indent (indent: indent + 2, format: "}\n" ); |
5075 | return is_final ? ES_RETURNED : ES_FALLTHROUGH; |
5076 | } |
5077 | } |
5078 | |
5079 | /* Make sure that OS has a position variable for POS. ROOT_P is true if |
5080 | POS is the root position for the routine. */ |
5081 | |
5082 | static void |
5083 | assign_position_var (output_state *os, position *pos, bool root_p) |
5084 | { |
5085 | unsigned int idx = os->id_to_var[pos->id]; |
5086 | if (idx < os->var_to_id.length () && os->var_to_id[idx] == pos->id) |
5087 | return; |
5088 | if (!root_p && pos->type != POS_PEEP2_INSN) |
5089 | assign_position_var (os, pos: pos->base, root_p: false); |
5090 | os->id_to_var[pos->id] = os->var_to_id.length (); |
5091 | os->var_to_id.safe_push (obj: pos->id); |
5092 | } |
5093 | |
5094 | /* Make sure that OS has the position variables required by S. */ |
5095 | |
5096 | static void |
5097 | assign_position_vars (output_state *os, state *s) |
5098 | { |
5099 | for (decision *d = s->first; d; d = d->next) |
5100 | { |
5101 | /* Positions associated with operands can be read from the |
5102 | operands[] array. */ |
5103 | if (d->test.pos && d->test.pos_operand < 0) |
5104 | assign_position_var (os, pos: d->test.pos, root_p: false); |
5105 | for (transition *trans = d->first; trans; trans = trans->next) |
5106 | assign_position_vars (os, s: trans->to); |
5107 | } |
5108 | } |
5109 | |
5110 | /* Print the open brace and variable definitions for a routine that |
5111 | implements S. ROOT is the deepest rtx from which S can access all |
5112 | relevant parts of the first instruction it matches. Initialize OS |
5113 | so that every relevant position has an rtx variable xN and so that |
5114 | only ROOT's variable has a valid value. */ |
5115 | |
5116 | static void |
5117 | print_subroutine_start (output_state *os, state *s, position *root) |
5118 | { |
5119 | printf (format: "{\n rtx * const operands ATTRIBUTE_UNUSED" |
5120 | " = &recog_data.operand[0];\n" ); |
5121 | os->var_to_id.truncate (size: 0); |
5122 | os->seen_vars.truncate (size: 0); |
5123 | if (root) |
5124 | { |
5125 | /* Create a fake entry for position 0 so that an id_to_var of 0 |
5126 | is always invalid. This also makes the xN variables naturally |
5127 | 1-based rather than 0-based. */ |
5128 | os->var_to_id.safe_push (obj: num_positions); |
5129 | |
5130 | /* Associate ROOT with x1. */ |
5131 | assign_position_var (os, pos: root, root_p: true); |
5132 | |
5133 | /* Assign xN variables to all other relevant positions. */ |
5134 | assign_position_vars (os, s); |
5135 | |
5136 | /* Output the variable declarations (except for ROOT's, which is |
5137 | passed in as a parameter). */ |
5138 | unsigned int num_vars = os->var_to_id.length (); |
5139 | if (num_vars > 2) |
5140 | { |
5141 | for (unsigned int i = 2; i < num_vars; ++i) |
5142 | /* Print 8 rtx variables to a line. */ |
5143 | printf (format: "%s x%d" , |
5144 | i == 2 ? " rtx" : (i - 2) % 8 == 0 ? ";\n rtx" : "," , i); |
5145 | printf (format: ";\n" ); |
5146 | } |
5147 | |
5148 | /* Say that x1 is valid and the rest aren't. */ |
5149 | os->seen_vars.safe_grow_cleared (len: num_vars, exact: true); |
5150 | os->seen_vars[1] = true; |
5151 | } |
5152 | if (os->type == SUBPATTERN || os->type == RECOG) |
5153 | printf (format: " int res ATTRIBUTE_UNUSED;\n" ); |
5154 | else |
5155 | printf (format: " rtx_insn *res ATTRIBUTE_UNUSED;\n" ); |
5156 | } |
5157 | |
5158 | /* Output the definition of pattern routine ROUTINE. */ |
5159 | |
5160 | static void |
5161 | print_pattern (output_state *os, pattern_routine *routine) |
5162 | { |
5163 | printf (format: "\nstatic int\npattern%d (" , routine->pattern_id); |
5164 | const char *sep = "" ; |
5165 | /* Add the top-level rtx parameter, if any. */ |
5166 | if (routine->pos) |
5167 | { |
5168 | printf (format: "%srtx x1" , sep); |
5169 | sep = ", " ; |
5170 | } |
5171 | /* Add the optional parameters. */ |
5172 | if (routine->insn_p) |
5173 | { |
5174 | /* We can't easily tell whether a C condition actually reads INSN, |
5175 | so add an ATTRIBUTE_UNUSED just in case. */ |
5176 | printf (format: "%srtx_insn *insn ATTRIBUTE_UNUSED" , sep); |
5177 | sep = ", " ; |
5178 | } |
5179 | if (routine->pnum_clobbers_p) |
5180 | { |
5181 | printf (format: "%sint *pnum_clobbers" , sep); |
5182 | sep = ", " ; |
5183 | } |
5184 | /* Add the "i" parameters. */ |
5185 | for (unsigned int i = 0; i < routine->param_types.length (); ++i) |
5186 | { |
5187 | printf (format: "%s%s i%d" , sep, |
5188 | parameter_type_string (type: routine->param_types[i]), i + 1); |
5189 | sep = ", " ; |
5190 | } |
5191 | printf (format: ")\n" ); |
5192 | os->type = SUBPATTERN; |
5193 | print_subroutine_start (os, s: routine->s, root: routine->pos); |
5194 | print_state (os, s: routine->s, indent: 2, is_final: true); |
5195 | printf (format: "}\n" ); |
5196 | } |
5197 | |
5198 | /* Output a routine of type TYPE that implements S. PROC_ID is the |
5199 | number of the subroutine associated with S, or 0 if S is the main |
5200 | routine. */ |
5201 | |
5202 | static void |
5203 | print_subroutine (output_state *os, state *s, int proc_id) |
5204 | { |
5205 | printf (format: "\n" ); |
5206 | switch (os->type) |
5207 | { |
5208 | case SUBPATTERN: |
5209 | gcc_unreachable (); |
5210 | |
5211 | case RECOG: |
5212 | if (proc_id) |
5213 | printf (format: "static int\nrecog_%d" , proc_id); |
5214 | else |
5215 | printf (format: "int\nrecog" ); |
5216 | printf (format: " (rtx x1 ATTRIBUTE_UNUSED,\n" |
5217 | "\trtx_insn *insn ATTRIBUTE_UNUSED,\n" |
5218 | "\tint *pnum_clobbers ATTRIBUTE_UNUSED)\n" ); |
5219 | break; |
5220 | |
5221 | case SPLIT: |
5222 | if (proc_id) |
5223 | printf (format: "static rtx_insn *\nsplit_%d" , proc_id); |
5224 | else |
5225 | printf (format: "rtx_insn *\nsplit_insns" ); |
5226 | printf (format: " (rtx x1 ATTRIBUTE_UNUSED, rtx_insn *insn ATTRIBUTE_UNUSED)\n" ); |
5227 | break; |
5228 | |
5229 | case PEEPHOLE2: |
5230 | if (proc_id) |
5231 | printf (format: "static rtx_insn *\npeephole2_%d" , proc_id); |
5232 | else |
5233 | printf (format: "rtx_insn *\npeephole2_insns" ); |
5234 | printf (format: " (rtx x1 ATTRIBUTE_UNUSED,\n" |
5235 | "\trtx_insn *insn ATTRIBUTE_UNUSED,\n" |
5236 | "\tint *pmatch_len_ ATTRIBUTE_UNUSED)\n" ); |
5237 | break; |
5238 | } |
5239 | print_subroutine_start (os, s, root: &root_pos); |
5240 | if (proc_id == 0) |
5241 | { |
5242 | printf (format: " recog_data.insn = NULL;\n" ); |
5243 | } |
5244 | print_state (os, s, indent: 2, is_final: true); |
5245 | printf (format: "}\n" ); |
5246 | } |
5247 | |
5248 | /* Print out a routine of type TYPE that performs ROOT. */ |
5249 | |
5250 | static void |
5251 | print_subroutine_group (output_state *os, routine_type type, state *root) |
5252 | { |
5253 | os->type = type; |
5254 | if (use_subroutines_p) |
5255 | { |
5256 | /* Split ROOT up into smaller pieces, both for readability and to |
5257 | help the compiler. */ |
5258 | auto_vec <state *> subroutines; |
5259 | find_subroutines (type, s: root, procs&: subroutines); |
5260 | |
5261 | /* Output the subroutines (but not ROOT itself). */ |
5262 | unsigned int i; |
5263 | state *s; |
5264 | FOR_EACH_VEC_ELT (subroutines, i, s) |
5265 | print_subroutine (os, s, proc_id: i + 1); |
5266 | } |
5267 | /* Output the main routine. */ |
5268 | print_subroutine (os, s: root, proc_id: 0); |
5269 | } |
5270 | |
5271 | /* Return the rtx pattern for the list of rtxes in a define_peephole2. */ |
5272 | |
5273 | static rtx |
5274 | get_peephole2_pattern (md_rtx_info *info) |
5275 | { |
5276 | int i, j; |
5277 | rtvec vec = XVEC (info->def, 0); |
5278 | rtx pattern = rtx_alloc (SEQUENCE); |
5279 | XVEC (pattern, 0) = rtvec_alloc (GET_NUM_ELEM (vec)); |
5280 | for (i = j = 0; i < GET_NUM_ELEM (vec); i++) |
5281 | { |
5282 | rtx x = RTVEC_ELT (vec, i); |
5283 | /* Ignore scratch register requirements. */ |
5284 | if (GET_CODE (x) != MATCH_SCRATCH && GET_CODE (x) != MATCH_DUP) |
5285 | { |
5286 | XVECEXP (pattern, 0, j) = x; |
5287 | j++; |
5288 | } |
5289 | } |
5290 | XVECLEN (pattern, 0) = j; |
5291 | if (j == 0) |
5292 | error_at (info->loc, "empty define_peephole2" ); |
5293 | return pattern; |
5294 | } |
5295 | |
5296 | /* Return true if *PATTERN_PTR is a PARALLEL in which at least one trailing |
5297 | rtx can be added automatically by add_clobbers. If so, update |
5298 | *ACCEPTANCE_PTR so that its num_clobbers field contains the number |
5299 | of such trailing rtxes and update *PATTERN_PTR so that it contains |
5300 | the pattern without those rtxes. */ |
5301 | |
5302 | static bool |
5303 | remove_clobbers (acceptance_type *acceptance_ptr, rtx *pattern_ptr) |
5304 | { |
5305 | int i; |
5306 | rtx new_pattern; |
5307 | |
5308 | /* Find the last non-clobber in the parallel. */ |
5309 | rtx pattern = *pattern_ptr; |
5310 | for (i = XVECLEN (pattern, 0); i > 0; i--) |
5311 | { |
5312 | rtx x = XVECEXP (pattern, 0, i - 1); |
5313 | if (GET_CODE (x) != CLOBBER |
5314 | || (!REG_P (XEXP (x, 0)) |
5315 | && GET_CODE (XEXP (x, 0)) != MATCH_SCRATCH)) |
5316 | break; |
5317 | } |
5318 | |
5319 | if (i == XVECLEN (pattern, 0)) |
5320 | return false; |
5321 | |
5322 | /* Build a similar insn without the clobbers. */ |
5323 | if (i == 1) |
5324 | new_pattern = XVECEXP (pattern, 0, 0); |
5325 | else |
5326 | { |
5327 | new_pattern = rtx_alloc (PARALLEL); |
5328 | XVEC (new_pattern, 0) = rtvec_alloc (i); |
5329 | for (int j = 0; j < i; ++j) |
5330 | XVECEXP (new_pattern, 0, j) = XVECEXP (pattern, 0, j); |
5331 | } |
5332 | |
5333 | /* Recognize it. */ |
5334 | acceptance_ptr->u.full.u.num_clobbers = XVECLEN (pattern, 0) - i; |
5335 | *pattern_ptr = new_pattern; |
5336 | return true; |
5337 | } |
5338 | |
5339 | int |
5340 | main (int argc, const char **argv) |
5341 | { |
5342 | state insn_root, split_root, peephole2_root; |
5343 | |
5344 | progname = "genrecog" ; |
5345 | |
5346 | if (!init_rtx_reader_args (argc, argv)) |
5347 | return (FATAL_EXIT_CODE); |
5348 | |
5349 | write_header (); |
5350 | |
5351 | /* Read the machine description. */ |
5352 | |
5353 | md_rtx_info info; |
5354 | while (read_md_rtx (&info)) |
5355 | { |
5356 | rtx def = info.def; |
5357 | |
5358 | acceptance_type acceptance; |
5359 | acceptance.partial_p = false; |
5360 | acceptance.u.full.code = info.index; |
5361 | |
5362 | rtx pattern; |
5363 | switch (GET_CODE (def)) |
5364 | { |
5365 | case DEFINE_INSN: |
5366 | { |
5367 | /* Match the instruction in the original .md form. */ |
5368 | acceptance.type = RECOG; |
5369 | acceptance.u.full.u.num_clobbers = 0; |
5370 | pattern = add_implicit_parallel (XVEC (def, 1)); |
5371 | validate_pattern (pattern, info: &info, NULL_RTX, set_code: 0); |
5372 | match_pattern (s: &insn_root, info: &info, pattern, acceptance); |
5373 | |
5374 | /* If the pattern is a PARALLEL with trailing CLOBBERs, |
5375 | allow recog_for_combine to match without the clobbers. */ |
5376 | if (GET_CODE (pattern) == PARALLEL |
5377 | && remove_clobbers (acceptance_ptr: &acceptance, pattern_ptr: &pattern)) |
5378 | match_pattern (s: &insn_root, info: &info, pattern, acceptance); |
5379 | break; |
5380 | } |
5381 | |
5382 | case DEFINE_SPLIT: |
5383 | acceptance.type = SPLIT; |
5384 | pattern = add_implicit_parallel (XVEC (def, 0)); |
5385 | validate_pattern (pattern, info: &info, NULL_RTX, set_code: 0); |
5386 | match_pattern (s: &split_root, info: &info, pattern, acceptance); |
5387 | |
5388 | /* Declare the gen_split routine that we'll call if the |
5389 | pattern matches. The definition comes from insn-emit.cc. */ |
5390 | printf (format: "extern rtx_insn *gen_split_%d (rtx_insn *, rtx *);\n" , |
5391 | info.index); |
5392 | break; |
5393 | |
5394 | case DEFINE_PEEPHOLE2: |
5395 | acceptance.type = PEEPHOLE2; |
5396 | pattern = get_peephole2_pattern (info: &info); |
5397 | validate_pattern (pattern, info: &info, NULL_RTX, set_code: 0); |
5398 | match_pattern (s: &peephole2_root, info: &info, pattern, acceptance); |
5399 | |
5400 | /* Declare the gen_peephole2 routine that we'll call if the |
5401 | pattern matches. The definition comes from insn-emit.cc. */ |
5402 | printf (format: "extern rtx_insn *gen_peephole2_%d (rtx_insn *, rtx *);\n" , |
5403 | info.index); |
5404 | break; |
5405 | |
5406 | default: |
5407 | /* do nothing */; |
5408 | } |
5409 | } |
5410 | |
5411 | if (have_error) |
5412 | return FATAL_EXIT_CODE; |
5413 | |
5414 | puts (s: "\n\n" ); |
5415 | |
5416 | /* Optimize each routine in turn. */ |
5417 | optimize_subroutine_group (type: "recog" , root: &insn_root); |
5418 | optimize_subroutine_group (type: "split_insns" , root: &split_root); |
5419 | optimize_subroutine_group (type: "peephole2_insns" , root: &peephole2_root); |
5420 | |
5421 | output_state os; |
5422 | os.id_to_var.safe_grow_cleared (len: num_positions, exact: true); |
5423 | |
5424 | if (use_pattern_routines_p) |
5425 | { |
5426 | /* Look for common patterns and split them out into subroutines. */ |
5427 | auto_vec <merge_state_info> states; |
5428 | states.safe_push (obj: &insn_root); |
5429 | states.safe_push (obj: &split_root); |
5430 | states.safe_push (obj: &peephole2_root); |
5431 | split_out_patterns (states); |
5432 | |
5433 | /* Print out the routines that we just created. */ |
5434 | unsigned int i; |
5435 | pattern_routine *routine; |
5436 | FOR_EACH_VEC_ELT (patterns, i, routine) |
5437 | print_pattern (os: &os, routine); |
5438 | } |
5439 | |
5440 | /* Print out the matching routines. */ |
5441 | print_subroutine_group (os: &os, type: RECOG, root: &insn_root); |
5442 | print_subroutine_group (os: &os, type: SPLIT, root: &split_root); |
5443 | print_subroutine_group (os: &os, type: PEEPHOLE2, root: &peephole2_root); |
5444 | |
5445 | fflush (stdout); |
5446 | return (ferror (stdout) != 0 ? FATAL_EXIT_CODE : SUCCESS_EXIT_CODE); |
5447 | } |
5448 | |