]>
Commit | Line | Data |
---|---|---|
5cce8171 RS |
1 | /* Early (pre-RA) rematerialization |
2 | Copyright (C) 2017 Free Software Foundation, Inc. | |
3 | ||
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
8 | Software Foundation; either version 3, or (at your option) any later | |
9 | version. | |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
17 | along with GCC; see the file COPYING3. If not see | |
18 | <http://www.gnu.org/licenses/>. */ | |
19 | ||
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
23 | #include "backend.h" | |
24 | #include "rtl.h" | |
25 | #include "df.h" | |
26 | #include "tree-pass.h" | |
27 | #include "memmodel.h" | |
28 | #include "emit-rtl.h" | |
29 | #include "insn-config.h" | |
30 | #include "recog.h" | |
31 | /* FIXME: The next two are only needed for gen_move_insn. */ | |
32 | #include "tree.h" | |
33 | #include "expr.h" | |
34 | #include "target.h" | |
35 | #include "inchash.h" | |
36 | #include "rtlhash.h" | |
37 | #include "print-rtl.h" | |
38 | #include "rtl-iter.h" | |
39 | ||
40 | /* This pass runs before register allocation and implements an aggressive | |
41 | form of rematerialization. It looks for pseudo registers R of mode M | |
42 | for which: | |
43 | ||
44 | (a) there are no call-preserved registers of mode M; and | |
45 | (b) spilling R to the stack is expensive. | |
46 | ||
47 | The assumption is that it's better to recompute R after each call instead | |
48 | of spilling it, even if this extends the live ranges of other registers. | |
49 | ||
50 | The motivating example for which these conditions hold are AArch64 SVE | |
51 | vectors and predicates. Spilling them to the stack makes the frame | |
52 | variable-sized, which we'd like to avoid if possible. It's also very | |
53 | rare for SVE values to be "naturally" live across a call: usually this | |
54 | happens as a result of CSE or other code motion. | |
55 | ||
56 | The pass is split into the following phases: | |
57 | ||
58 | Collection phase | |
59 | ================ | |
60 | ||
61 | First we go through all pseudo registers looking for any that meet | |
62 | the conditions above. For each such register R, we go through each | |
63 | instruction that defines R to see whether any of them are suitable | |
64 | rematerialization candidates. If at least one is, we treat all the | |
65 | instructions that define R as candidates, but record which ones are | |
66 | not in fact suitable. These unsuitable candidates exist only for the | |
67 | sake of calculating reaching definitions (see below). | |
68 | ||
69 | A "candidate" is a single instruction that we want to rematerialize | |
70 | and a "candidate register" is a register that is set by at least one | |
71 | candidate. | |
72 | ||
73 | Candidate sorting | |
74 | ================= | |
75 | ||
76 | Next we sort the candidates based on the cfg postorder, so that if | |
77 | candidate C1 uses candidate C2, C1 has a lower index than C2. | |
78 | This is useful when iterating through candidate bitmaps. | |
79 | ||
80 | Reaching definition calculation | |
81 | =============================== | |
82 | ||
83 | We then compute standard reaching-definition sets for each candidate. | |
84 | Each set specifies which candidates might provide the current definition | |
85 | of a live candidate register. | |
86 | ||
87 | From here on, a candidate C is "live" at a point P if the candidate | |
88 | register defined by C is live at P and if C's definition reaches P. | |
89 | An instruction I "uses" a candidate C if I takes the register defined by | |
90 | C as input and if C is one of the reaching definitions of that register. | |
91 | ||
92 | Candidate validation and value numbering | |
93 | ======================================== | |
94 | ||
95 | Next we simultaneously decide which candidates are valid and look | |
96 | for candidates that are equivalent to each other, assigning numbers | |
97 | to each unique candidate value. A candidate C is invalid if: | |
98 | ||
99 | (a) C uses an invalid candidate; | |
100 | ||
101 | (b) there is a cycle of candidate uses involving C; or | |
102 | ||
103 | (c) C takes a candidate register R as input and the reaching | |
104 | definitions of R do not have the same value number. | |
105 | ||
106 | We assign a "representative" candidate C to each value number and from | |
107 | here on replace references to other candidates with that value number | |
108 | with references to C. It is then only possible to rematerialize a | |
109 | register R at point P if (after this replacement) there is a single | |
110 | reaching definition of R at P. | |
111 | ||
112 | Local phase | |
113 | =========== | |
114 | ||
115 | During this phase we go through each block and look for cases in which: | |
116 | ||
117 | (a) an instruction I comes between two call instructions CI1 and CI2; | |
118 | ||
119 | (b) I uses a candidate register R; | |
120 | ||
121 | (c) a candidate C provides the only reaching definition of R; and | |
122 | ||
123 | (d) C does not come between CI1 and I. | |
124 | ||
125 | We then emit a copy of C after CI1, as well as the transitive closure | |
126 | TC of the candidates used by C. The copies of TC might use the original | |
127 | candidate registers or new temporary registers, depending on circumstances. | |
128 | ||
129 | For example, if elsewhere we have: | |
130 | ||
131 | C3: R3 <- f3 (...) | |
132 | ... | |
133 | C2: R2 <- f2 (...) | |
134 | ... | |
135 | C1: R1 <- f1 (R2, R3, ...) // uses C2 and C3 | |
136 | ||
137 | then for a block containing: | |
138 | ||
139 | CI1: call | |
140 | ... | |
141 | I: use R1 // uses C1 | |
142 | ... | |
143 | CI2: call | |
144 | ||
145 | we would emit: | |
146 | ||
147 | CI1: call | |
148 | C3': R3' <- f3 (...) | |
149 | C2': R2' <- f2 (...) | |
150 | C1': R1 <- f1 (R2', R3', ...) | |
151 | ... | |
152 | I: use R1 | |
153 | ... | |
154 | CI2: call | |
155 | ||
156 | where R2' and R3' might be fresh registers. If instead we had: | |
157 | ||
158 | CI1: call | |
159 | ... | |
160 | I1: use R1 // uses C1 | |
161 | ... | |
162 | I2: use R3 // uses C3 | |
163 | ... | |
164 | CI2: call | |
165 | ||
166 | we would keep the original R3: | |
167 | ||
168 | CI1: call | |
169 | C3': R3 <- f3 (...) | |
170 | C2': R2' <- f2 (...) | |
171 | C1': R1 <- f1 (R2', R3, ...) | |
172 | ... | |
173 | I1: use R1 // uses C1 | |
174 | ... | |
175 | I2: use R3 // uses C3 | |
176 | ... | |
177 | CI2: call | |
178 | ||
179 | We also record the last call in each block (if any) and compute: | |
180 | ||
181 | rd_after_call: | |
182 | The set of candidates that either (a) are defined outside the block | |
183 | and are live after the last call or (b) are defined within the block | |
184 | and reach the end of the last call. (We don't track whether the | |
185 | latter values are live or not.) | |
186 | ||
187 | required_after_call: | |
188 | The set of candidates that need to be rematerialized after the | |
189 | last call in order to satisfy uses in the block itself. | |
190 | ||
191 | required_in: | |
192 | The set of candidates that are live on entry to the block and are | |
193 | used without an intervening call. | |
194 | ||
195 | In addition, we compute the initial values of the sets required by | |
196 | the global phase below. | |
197 | ||
198 | Global phase | |
199 | ============ | |
200 | ||
201 | We next compute a maximal solution to the following availability | |
202 | problem: | |
203 | ||
204 | available_in: | |
205 | The set of candidates that are live on entry to a block and can | |
206 | be used at that point without rematerialization. | |
207 | ||
208 | available_out: | |
209 | The set of candidates that are live on exit from a block and can | |
210 | be used at that point without rematerialization. | |
211 | ||
212 | available_locally: | |
213 | The subset of available_out that is due to code in the block itself. | |
214 | It contains candidates that are defined or used in the block and | |
215 | not invalidated by a later call. | |
216 | ||
217 | We then go through each block B and look for an appropriate place | |
218 | to insert copies of required_in - available_in. Conceptually we | |
219 | start by placing the copies at the head of B, but then move the | |
220 | copy of a candidate C to predecessors if: | |
221 | ||
222 | (a) that seems cheaper; | |
223 | ||
224 | (b) there is more than one reaching definition of C's register at | |
225 | the head of B; or | |
226 | ||
227 | (c) copying C would clobber a hard register that is live on entry to B. | |
228 | ||
229 | Moving a copy of C to a predecessor block PB involves: | |
230 | ||
231 | (1) adding C to PB's required_after_call, if PB contains a call; or | |
232 | ||
233 | (2) adding C PB's required_in otherwise. | |
234 | ||
235 | C is then available on output from each PB and on input to B. | |
236 | ||
237 | Once all this is done, we emit instructions for the final required_in | |
238 | and required_after_call sets. */ | |
239 | ||
240 | namespace { | |
241 | ||
242 | /* An invalid candidate index, used to indicate that there is more than | |
243 | one reaching definition. */ | |
244 | const unsigned int MULTIPLE_CANDIDATES = -1U; | |
245 | ||
246 | /* Pass-specific information about one basic block. */ | |
247 | struct remat_block_info { | |
248 | /* The last call instruction in the block. */ | |
249 | rtx_insn *last_call; | |
250 | ||
251 | /* The set of candidates that are live on entry to the block. NULL is | |
252 | equivalent to an empty set. */ | |
253 | bitmap rd_in; | |
254 | ||
255 | /* The set of candidates that are live on exit from the block. This might | |
256 | reuse rd_in. NULL is equivalent to an empty set. */ | |
257 | bitmap rd_out; | |
258 | ||
259 | /* The subset of RD_OUT that comes from local definitions. NULL is | |
260 | equivalent to an empty set. */ | |
261 | bitmap rd_gen; | |
262 | ||
263 | /* The set of candidates that the block invalidates (because it defines | |
264 | the register to something else, or because the register's value is | |
265 | no longer important). NULL is equivalent to an empty set. */ | |
266 | bitmap rd_kill; | |
267 | ||
268 | /* The set of candidates that either (a) are defined outside the block | |
269 | and are live after LAST_CALL or (b) are defined within the block | |
270 | and reach the instruction after LAST_CALL. (We don't track whether | |
271 | the latter values are live or not.) | |
272 | ||
273 | Only used if LAST_CALL is nonnull. NULL is equivalent to an | |
274 | empty set. */ | |
275 | bitmap rd_after_call; | |
276 | ||
277 | /* Candidates that are live and available without rematerialization | |
278 | on entry to the block. NULL is equivalent to an empty set. */ | |
279 | bitmap available_in; | |
280 | ||
281 | /* Candidates that become available without rematerialization within the | |
282 | block, and remain so on exit. NULL is equivalent to an empty set. */ | |
283 | bitmap available_locally; | |
284 | ||
285 | /* Candidates that are available without rematerialization on exit from | |
286 | the block. This might reuse available_in or available_locally. */ | |
287 | bitmap available_out; | |
288 | ||
289 | /* Candidates that need to be rematerialized either at the start of the | |
290 | block or before entering the block. */ | |
291 | bitmap required_in; | |
292 | ||
293 | /* Candidates that need to be rematerialized after LAST_CALL. | |
294 | Only used if LAST_CALL is nonnull. */ | |
295 | bitmap required_after_call; | |
296 | ||
297 | /* The number of candidates in the block. */ | |
298 | unsigned int num_candidates; | |
299 | ||
300 | /* The earliest candidate in the block (i.e. the one with the | |
301 | highest index). Only valid if NUM_CANDIDATES is nonzero. */ | |
302 | unsigned int first_candidate; | |
303 | ||
304 | /* The best (lowest) execution frequency for rematerializing REQUIRED_IN. | |
305 | This is the execution frequency of the block if LOCAL_REMAT_CHEAPER_P, | |
306 | otherwise it is the sum of the execution frequencies of whichever | |
307 | predecessor blocks would do the rematerialization. */ | |
308 | int remat_frequency; | |
309 | ||
310 | /* True if the block ends with an abnormal call. */ | |
311 | unsigned int abnormal_call_p : 1; | |
312 | ||
313 | /* Used to record whether a graph traversal has visited this block. */ | |
314 | unsigned int visited_p : 1; | |
315 | ||
316 | /* True if we have calculated REMAT_FREQUENCY. */ | |
317 | unsigned int remat_frequency_valid_p : 1; | |
318 | ||
319 | /* True if it is cheaper to rematerialize candidates at the start of | |
320 | the block, rather than moving them to predecessor blocks. */ | |
321 | unsigned int local_remat_cheaper_p : 1; | |
322 | }; | |
323 | ||
324 | /* Information about a group of candidates with the same value number. */ | |
325 | struct remat_equiv_class { | |
326 | /* The candidates that have the same value number. */ | |
327 | bitmap members; | |
328 | ||
329 | /* The candidate that was first added to MEMBERS. */ | |
330 | unsigned int earliest; | |
331 | ||
332 | /* The candidate that represents the others. This is always the one | |
333 | with the highest index. */ | |
334 | unsigned int representative; | |
335 | }; | |
336 | ||
337 | /* Information about an instruction that we might want to rematerialize. */ | |
338 | struct remat_candidate { | |
339 | /* The pseudo register that the instruction sets. */ | |
340 | unsigned int regno; | |
341 | ||
342 | /* A temporary register used when rematerializing uses of this candidate, | |
343 | if REGNO doesn't have the right value or isn't worth using. */ | |
344 | unsigned int copy_regno; | |
345 | ||
346 | /* True if we intend to rematerialize this instruction by emitting | |
347 | a move of a constant into REGNO, false if we intend to emit a | |
348 | copy of the original instruction. */ | |
349 | unsigned int constant_p : 1; | |
350 | ||
351 | /* True if we still think it's possible to rematerialize INSN. */ | |
352 | unsigned int can_copy_p : 1; | |
353 | ||
354 | /* Used to record whether a graph traversal has visited this candidate. */ | |
355 | unsigned int visited_p : 1; | |
356 | ||
357 | /* True if we have verified that it's possible to rematerialize INSN. | |
358 | Once this is true, both it and CAN_COPY_P remain true. */ | |
359 | unsigned int validated_p : 1; | |
360 | ||
361 | /* True if we have "stabilized" INSN, i.e. ensured that all non-candidate | |
362 | registers read by INSN will have the same value when rematerializing INSN. | |
363 | Only ever true if CAN_COPY_P. */ | |
364 | unsigned int stabilized_p : 1; | |
365 | ||
366 | /* Hash value used for value numbering. */ | |
367 | hashval_t hash; | |
368 | ||
369 | /* The instruction that sets REGNO. */ | |
370 | rtx_insn *insn; | |
371 | ||
372 | /* If CONSTANT_P, the value that should be moved into REGNO when | |
373 | rematerializing, otherwise the pattern of the instruction that | |
374 | should be used. */ | |
375 | rtx remat_rtx; | |
376 | ||
377 | /* The set of candidates that INSN takes as input. NULL is equivalent | |
378 | to the empty set. All candidates in this set have a higher index | |
379 | than the current candidate. */ | |
380 | bitmap uses; | |
381 | ||
382 | /* The set of hard registers that would be clobbered by rematerializing | |
383 | the candidate, including (transitively) all those that would be | |
384 | clobbered by rematerializing USES. */ | |
385 | bitmap clobbers; | |
386 | ||
387 | /* The equivalence class to which the candidate belongs, or null if none. */ | |
388 | remat_equiv_class *equiv_class; | |
389 | }; | |
390 | ||
391 | /* Hash functions used for value numbering. */ | |
392 | struct remat_candidate_hasher : nofree_ptr_hash <remat_candidate> | |
393 | { | |
394 | typedef value_type compare_type; | |
395 | static hashval_t hash (const remat_candidate *); | |
396 | static bool equal (const remat_candidate *, const remat_candidate *); | |
397 | }; | |
398 | ||
399 | /* Main class for this pass. */ | |
400 | class early_remat { | |
401 | public: | |
402 | early_remat (function *, sbitmap); | |
403 | ~early_remat (); | |
404 | ||
405 | void run (void); | |
406 | ||
407 | private: | |
408 | bitmap alloc_bitmap (void); | |
409 | bitmap get_bitmap (bitmap *); | |
410 | void init_temp_bitmap (bitmap *); | |
411 | void copy_temp_bitmap (bitmap *, bitmap *); | |
412 | ||
413 | void dump_insn_id (rtx_insn *); | |
414 | void dump_candidate_bitmap (bitmap); | |
415 | void dump_all_candidates (void); | |
416 | void dump_edge_list (basic_block, bool); | |
417 | void dump_block_info (basic_block); | |
418 | void dump_all_blocks (void); | |
419 | ||
420 | bool interesting_regno_p (unsigned int); | |
421 | remat_candidate *add_candidate (rtx_insn *, unsigned int, bool); | |
422 | bool maybe_add_candidate (rtx_insn *, unsigned int); | |
423 | bool collect_candidates (void); | |
424 | void init_block_info (void); | |
425 | void sort_candidates (void); | |
426 | void finalize_candidate_indices (void); | |
427 | void record_equiv_candidates (unsigned int, unsigned int); | |
428 | static bool rd_confluence_n (edge); | |
429 | static bool rd_transfer (int); | |
430 | void compute_rd (void); | |
431 | unsigned int canon_candidate (unsigned int); | |
432 | void canon_bitmap (bitmap *); | |
433 | unsigned int resolve_reaching_def (bitmap); | |
434 | bool check_candidate_uses (unsigned int); | |
435 | void compute_clobbers (unsigned int); | |
436 | void assign_value_number (unsigned int); | |
437 | void decide_candidate_validity (void); | |
438 | bool stable_use_p (unsigned int); | |
439 | void emit_copy_before (unsigned int, rtx, rtx); | |
440 | void stabilize_pattern (unsigned int); | |
441 | void replace_dest_with_copy (unsigned int); | |
442 | void stabilize_candidate_uses (unsigned int, bitmap, bitmap, bitmap, | |
443 | bitmap); | |
444 | void emit_remat_insns (bitmap, bitmap, bitmap, rtx_insn *); | |
445 | bool set_available_out (remat_block_info *); | |
446 | void process_block (basic_block); | |
447 | void local_phase (void); | |
448 | static bool avail_confluence_n (edge); | |
449 | static bool avail_transfer (int); | |
450 | void compute_availability (void); | |
451 | void unshare_available_sets (remat_block_info *); | |
452 | bool can_move_across_edge_p (edge); | |
453 | bool local_remat_cheaper_p (unsigned int); | |
454 | bool need_to_move_candidate_p (unsigned int, unsigned int); | |
455 | void compute_minimum_move_set (unsigned int, bitmap); | |
456 | void move_to_predecessors (unsigned int, bitmap, bitmap); | |
457 | void choose_rematerialization_points (void); | |
458 | void emit_remat_insns_for_block (basic_block); | |
459 | void global_phase (void); | |
460 | ||
461 | /* The function that we're optimizing. */ | |
462 | function *m_fn; | |
463 | ||
464 | /* The modes that we want to rematerialize. */ | |
465 | sbitmap m_selected_modes; | |
466 | ||
467 | /* All rematerialization candidates, identified by their index into the | |
468 | vector. */ | |
469 | auto_vec<remat_candidate> m_candidates; | |
470 | ||
471 | /* The set of candidate registers. */ | |
472 | bitmap_head m_candidate_regnos; | |
473 | ||
474 | /* Temporary sets. */ | |
475 | bitmap_head m_tmp_bitmap; | |
476 | bitmap m_available; | |
477 | bitmap m_required; | |
478 | ||
479 | /* Information about each basic block. */ | |
480 | auto_vec<remat_block_info> m_block_info; | |
481 | ||
482 | /* A mapping from register numbers to the set of associated candidates. | |
483 | Only valid for registers in M_CANDIDATE_REGNOS. */ | |
484 | auto_vec<bitmap> m_regno_to_candidates; | |
485 | ||
486 | /* An obstack used for allocating bitmaps, so that we can free them all | |
487 | in one go. */ | |
488 | bitmap_obstack m_obstack; | |
489 | ||
490 | /* A hash table of candidates used for value numbering. If a candidate | |
491 | in the table is in an equivalence class, the candidate is marked as | |
492 | the earliest member of the class. */ | |
493 | hash_table<remat_candidate_hasher> m_value_table; | |
494 | ||
495 | /* Used temporarily by callback functions. */ | |
496 | static early_remat *er; | |
497 | }; | |
498 | ||
499 | } | |
500 | ||
501 | early_remat *early_remat::er; | |
502 | ||
503 | /* rtx_equal_p_cb callback that treats any two SCRATCHes as equal. | |
504 | This allows us to compare two copies of a pattern, even though their | |
505 | SCRATCHes are always distinct. */ | |
506 | ||
507 | static int | |
508 | scratch_equal (const_rtx *x, const_rtx *y, rtx *nx, rtx *ny) | |
509 | { | |
510 | if (GET_CODE (*x) == SCRATCH && GET_CODE (*y) == SCRATCH) | |
511 | { | |
512 | *nx = const0_rtx; | |
513 | *ny = const0_rtx; | |
514 | return 1; | |
515 | } | |
516 | return 0; | |
517 | } | |
518 | ||
519 | /* Hash callback functions for remat_candidate. */ | |
520 | ||
521 | hashval_t | |
522 | remat_candidate_hasher::hash (const remat_candidate *cand) | |
523 | { | |
524 | return cand->hash; | |
525 | } | |
526 | ||
527 | bool | |
528 | remat_candidate_hasher::equal (const remat_candidate *cand1, | |
529 | const remat_candidate *cand2) | |
530 | { | |
531 | return (cand1->regno == cand2->regno | |
532 | && cand1->constant_p == cand2->constant_p | |
533 | && (cand1->constant_p | |
534 | ? rtx_equal_p (cand1->remat_rtx, cand2->remat_rtx) | |
535 | : rtx_equal_p_cb (cand1->remat_rtx, cand2->remat_rtx, | |
536 | scratch_equal)) | |
537 | && (!cand1->uses || bitmap_equal_p (cand1->uses, cand2->uses))); | |
538 | } | |
539 | ||
540 | /* Return true if B is null or empty. */ | |
541 | ||
542 | inline bool | |
543 | empty_p (bitmap b) | |
544 | { | |
545 | return !b || bitmap_empty_p (b); | |
546 | } | |
547 | ||
548 | /* Allocate a new bitmap. It will be automatically freed at the end of | |
549 | the pass. */ | |
550 | ||
551 | inline bitmap | |
552 | early_remat::alloc_bitmap (void) | |
553 | { | |
554 | return bitmap_alloc (&m_obstack); | |
555 | } | |
556 | ||
557 | /* Initialize *PTR to an empty bitmap if it is currently null. */ | |
558 | ||
559 | inline bitmap | |
560 | early_remat::get_bitmap (bitmap *ptr) | |
561 | { | |
562 | if (!*ptr) | |
563 | *ptr = alloc_bitmap (); | |
564 | return *ptr; | |
565 | } | |
566 | ||
567 | /* *PTR is either null or empty. If it is null, initialize it to an | |
568 | empty bitmap. */ | |
569 | ||
570 | inline void | |
571 | early_remat::init_temp_bitmap (bitmap *ptr) | |
572 | { | |
573 | if (!*ptr) | |
574 | *ptr = alloc_bitmap (); | |
575 | else | |
576 | gcc_checking_assert (bitmap_empty_p (*ptr)); | |
577 | } | |
578 | ||
579 | /* Move *SRC to *DEST and leave *SRC empty. */ | |
580 | ||
581 | inline void | |
582 | early_remat::copy_temp_bitmap (bitmap *dest, bitmap *src) | |
583 | { | |
584 | if (!empty_p (*src)) | |
585 | { | |
586 | *dest = *src; | |
587 | *src = NULL; | |
588 | } | |
589 | else | |
590 | *dest = NULL; | |
591 | } | |
592 | ||
593 | /* Print INSN's identifier to the dump file. */ | |
594 | ||
595 | void | |
596 | early_remat::dump_insn_id (rtx_insn *insn) | |
597 | { | |
598 | fprintf (dump_file, "%d[bb:%d]", INSN_UID (insn), | |
599 | BLOCK_FOR_INSN (insn)->index); | |
600 | } | |
601 | ||
602 | /* Print candidate set CANDIDATES to the dump file, with a leading space. */ | |
603 | ||
604 | void | |
605 | early_remat::dump_candidate_bitmap (bitmap candidates) | |
606 | { | |
607 | if (empty_p (candidates)) | |
608 | { | |
609 | fprintf (dump_file, " none"); | |
610 | return; | |
611 | } | |
612 | ||
613 | unsigned int cand_index; | |
614 | bitmap_iterator bi; | |
615 | EXECUTE_IF_SET_IN_BITMAP (candidates, 0, cand_index, bi) | |
616 | fprintf (dump_file, " %d", cand_index); | |
617 | } | |
618 | ||
619 | /* Print information about all candidates to the dump file. */ | |
620 | ||
621 | void | |
622 | early_remat::dump_all_candidates (void) | |
623 | { | |
624 | fprintf (dump_file, "\n;; Candidates:\n;;\n"); | |
625 | fprintf (dump_file, ";; %5s %5s %8s %s\n", "#", "reg", "mode", "insn"); | |
626 | fprintf (dump_file, ";; %5s %5s %8s %s\n", "=", "===", "====", "===="); | |
627 | unsigned int cand_index; | |
628 | remat_candidate *cand; | |
629 | FOR_EACH_VEC_ELT (m_candidates, cand_index, cand) | |
630 | { | |
631 | fprintf (dump_file, ";; %5d %5d %8s ", cand_index, cand->regno, | |
632 | GET_MODE_NAME (GET_MODE (regno_reg_rtx[cand->regno]))); | |
633 | dump_insn_id (cand->insn); | |
634 | if (!cand->can_copy_p) | |
635 | fprintf (dump_file, " -- can't copy"); | |
636 | fprintf (dump_file, "\n"); | |
637 | } | |
638 | ||
639 | fprintf (dump_file, "\n;; Register-to-candidate mapping:\n;;\n"); | |
640 | unsigned int regno; | |
641 | bitmap_iterator bi; | |
642 | EXECUTE_IF_SET_IN_BITMAP (&m_candidate_regnos, 0, regno, bi) | |
643 | { | |
644 | fprintf (dump_file, ";; %5d:", regno); | |
645 | dump_candidate_bitmap (m_regno_to_candidates[regno]); | |
646 | fprintf (dump_file, "\n"); | |
647 | } | |
648 | } | |
649 | ||
650 | /* Print the predecessors or successors of BB to the dump file, with a | |
651 | leading space. DO_SUCC is true to print successors and false to print | |
652 | predecessors. */ | |
653 | ||
654 | void | |
655 | early_remat::dump_edge_list (basic_block bb, bool do_succ) | |
656 | { | |
657 | edge e; | |
658 | edge_iterator ei; | |
659 | FOR_EACH_EDGE (e, ei, do_succ ? bb->succs : bb->preds) | |
4af78ef8 | 660 | dump_edge_info (dump_file, e, TDF_NONE, do_succ); |
5cce8171 RS |
661 | } |
662 | ||
663 | /* Print information about basic block BB to the dump file. */ | |
664 | ||
665 | void | |
666 | early_remat::dump_block_info (basic_block bb) | |
667 | { | |
668 | remat_block_info *info = &m_block_info[bb->index]; | |
669 | fprintf (dump_file, ";;\n;; Block %d:", bb->index); | |
670 | int width = 25; | |
671 | ||
672 | fprintf (dump_file, "\n;;%*s:", width, "predecessors"); | |
673 | dump_edge_list (bb, false); | |
674 | ||
675 | fprintf (dump_file, "\n;;%*s:", width, "successors"); | |
676 | dump_edge_list (bb, true); | |
677 | ||
678 | fprintf (dump_file, "\n;;%*s: %d", width, "frequency", | |
679 | bb->count.to_frequency (m_fn)); | |
680 | ||
681 | if (info->last_call) | |
682 | fprintf (dump_file, "\n;;%*s: %d", width, "last call", | |
683 | INSN_UID (info->last_call)); | |
684 | ||
685 | if (!empty_p (info->rd_in)) | |
686 | { | |
687 | fprintf (dump_file, "\n;;%*s:", width, "RD in"); | |
688 | dump_candidate_bitmap (info->rd_in); | |
689 | } | |
690 | if (!empty_p (info->rd_kill)) | |
691 | { | |
692 | fprintf (dump_file, "\n;;%*s:", width, "RD kill"); | |
693 | dump_candidate_bitmap (info->rd_kill); | |
694 | } | |
695 | if (!empty_p (info->rd_gen)) | |
696 | { | |
697 | fprintf (dump_file, "\n;;%*s:", width, "RD gen"); | |
698 | dump_candidate_bitmap (info->rd_gen); | |
699 | } | |
700 | if (!empty_p (info->rd_after_call)) | |
701 | { | |
702 | fprintf (dump_file, "\n;;%*s:", width, "RD after call"); | |
703 | dump_candidate_bitmap (info->rd_after_call); | |
704 | } | |
705 | if (!empty_p (info->rd_out)) | |
706 | { | |
707 | fprintf (dump_file, "\n;;%*s:", width, "RD out"); | |
708 | if (info->rd_in == info->rd_out) | |
709 | fprintf (dump_file, " RD in"); | |
710 | else | |
711 | dump_candidate_bitmap (info->rd_out); | |
712 | } | |
713 | if (!empty_p (info->available_in)) | |
714 | { | |
715 | fprintf (dump_file, "\n;;%*s:", width, "available in"); | |
716 | dump_candidate_bitmap (info->available_in); | |
717 | } | |
718 | if (!empty_p (info->available_locally)) | |
719 | { | |
720 | fprintf (dump_file, "\n;;%*s:", width, "available locally"); | |
721 | dump_candidate_bitmap (info->available_locally); | |
722 | } | |
723 | if (!empty_p (info->available_out)) | |
724 | { | |
725 | fprintf (dump_file, "\n;;%*s:", width, "available out"); | |
726 | if (info->available_in == info->available_out) | |
727 | fprintf (dump_file, " available in"); | |
728 | else if (info->available_locally == info->available_out) | |
729 | fprintf (dump_file, " available locally"); | |
730 | else | |
731 | dump_candidate_bitmap (info->available_out); | |
732 | } | |
733 | if (!empty_p (info->required_in)) | |
734 | { | |
735 | fprintf (dump_file, "\n;;%*s:", width, "required in"); | |
736 | dump_candidate_bitmap (info->required_in); | |
737 | } | |
738 | if (!empty_p (info->required_after_call)) | |
739 | { | |
740 | fprintf (dump_file, "\n;;%*s:", width, "required after call"); | |
741 | dump_candidate_bitmap (info->required_after_call); | |
742 | } | |
743 | fprintf (dump_file, "\n"); | |
744 | } | |
745 | ||
746 | /* Print information about all basic blocks to the dump file. */ | |
747 | ||
748 | void | |
749 | early_remat::dump_all_blocks (void) | |
750 | { | |
751 | basic_block bb; | |
752 | FOR_EACH_BB_FN (bb, m_fn) | |
753 | dump_block_info (bb); | |
754 | } | |
755 | ||
756 | /* Return true if REGNO is worth rematerializing. */ | |
757 | ||
758 | bool | |
759 | early_remat::interesting_regno_p (unsigned int regno) | |
760 | { | |
761 | /* Ignore unused registers. */ | |
762 | rtx reg = regno_reg_rtx[regno]; | |
763 | if (!reg || DF_REG_DEF_COUNT (regno) == 0) | |
764 | return false; | |
765 | ||
766 | /* Make sure the register has a mode that we want to rematerialize. */ | |
767 | if (!bitmap_bit_p (m_selected_modes, GET_MODE (reg))) | |
768 | return false; | |
769 | ||
770 | /* Ignore values that might sometimes be used uninitialized. We could | |
771 | instead add dummy candidates for the entry block definition, and so | |
772 | handle uses that are definitely not uninitialized, but the combination | |
773 | of the two should be rare in practice. */ | |
774 | if (bitmap_bit_p (DF_LR_OUT (ENTRY_BLOCK_PTR_FOR_FN (m_fn)), regno)) | |
775 | return false; | |
776 | ||
777 | return true; | |
778 | } | |
779 | ||
780 | /* Record the set of register REGNO in instruction INSN as a | |
781 | rematerialization candidate. CAN_COPY_P is true unless we already | |
782 | know that rematerialization is impossible (in which case the candidate | |
783 | only exists for the reaching definition calculation). | |
784 | ||
785 | The candidate's index is not fixed at this stage. */ | |
786 | ||
787 | remat_candidate * | |
788 | early_remat::add_candidate (rtx_insn *insn, unsigned int regno, | |
789 | bool can_copy_p) | |
790 | { | |
791 | remat_candidate cand; | |
792 | memset (&cand, 0, sizeof (cand)); | |
793 | cand.regno = regno; | |
794 | cand.insn = insn; | |
795 | cand.remat_rtx = PATTERN (insn); | |
796 | cand.can_copy_p = can_copy_p; | |
797 | m_candidates.safe_push (cand); | |
798 | ||
799 | bitmap_set_bit (&m_candidate_regnos, regno); | |
800 | ||
801 | return &m_candidates.last (); | |
802 | } | |
803 | ||
804 | /* Return true if we can rematerialize the set of register REGNO in | |
805 | instruction INSN, and add it as a candidate if so. When returning | |
806 | false, print the reason to the dump file. */ | |
807 | ||
808 | bool | |
809 | early_remat::maybe_add_candidate (rtx_insn *insn, unsigned int regno) | |
810 | { | |
811 | #define FAILURE_FORMAT ";; Can't rematerialize set of reg %d in %d[bb:%d]: " | |
812 | #define FAILURE_ARGS regno, INSN_UID (insn), BLOCK_FOR_INSN (insn)->index | |
813 | ||
814 | /* The definition must come from an ordinary instruction. */ | |
815 | basic_block bb = BLOCK_FOR_INSN (insn); | |
816 | if (!NONJUMP_INSN_P (insn) | |
817 | || (insn == BB_END (bb) | |
818 | && has_abnormal_or_eh_outgoing_edge_p (bb))) | |
819 | { | |
820 | if (dump_file) | |
821 | fprintf (dump_file, FAILURE_FORMAT "insn alters control flow\n", | |
822 | FAILURE_ARGS); | |
823 | return false; | |
824 | } | |
825 | ||
826 | /* Prefer to rematerialize constants directly -- it's much easier. */ | |
827 | machine_mode mode = GET_MODE (regno_reg_rtx[regno]); | |
828 | if (rtx note = find_reg_equal_equiv_note (insn)) | |
829 | { | |
830 | rtx val = XEXP (note, 0); | |
831 | if (CONSTANT_P (val) | |
832 | && targetm.legitimate_constant_p (mode, val)) | |
833 | { | |
834 | remat_candidate *cand = add_candidate (insn, regno, true); | |
835 | cand->constant_p = true; | |
836 | cand->remat_rtx = val; | |
837 | return true; | |
838 | } | |
839 | } | |
840 | ||
841 | /* See whether the target has reasons to prevent a copy. */ | |
842 | if (targetm.cannot_copy_insn_p && targetm.cannot_copy_insn_p (insn)) | |
843 | { | |
844 | if (dump_file) | |
845 | fprintf (dump_file, FAILURE_FORMAT "target forbids copying\n", | |
846 | FAILURE_ARGS); | |
847 | return false; | |
848 | } | |
849 | ||
850 | /* We can't copy trapping instructions. */ | |
851 | rtx pat = PATTERN (insn); | |
852 | if (may_trap_p (pat)) | |
853 | { | |
854 | if (dump_file) | |
855 | fprintf (dump_file, FAILURE_FORMAT "insn might trap\n", FAILURE_ARGS); | |
856 | return false; | |
857 | } | |
858 | ||
859 | /* We can't copy instructions that read memory, unless we know that | |
860 | the contents never change. */ | |
861 | subrtx_iterator::array_type array; | |
862 | FOR_EACH_SUBRTX (iter, array, pat, ALL) | |
863 | if (MEM_P (*iter) && !MEM_READONLY_P (*iter)) | |
864 | { | |
865 | if (dump_file) | |
866 | fprintf (dump_file, FAILURE_FORMAT "insn references non-constant" | |
867 | " memory\n", FAILURE_ARGS); | |
868 | return false; | |
869 | } | |
870 | ||
871 | /* Check each defined register. */ | |
872 | df_ref ref; | |
873 | FOR_EACH_INSN_DEF (ref, insn) | |
874 | { | |
875 | unsigned int def_regno = DF_REF_REGNO (ref); | |
876 | if (def_regno == regno) | |
877 | { | |
878 | /* Make sure the definition is write-only. (Partial definitions, | |
879 | such as setting the low part and clobbering the high part, | |
880 | are otherwise OK.) */ | |
881 | if (DF_REF_FLAGS_IS_SET (ref, DF_REF_READ_WRITE)) | |
882 | { | |
883 | if (dump_file) | |
884 | fprintf (dump_file, FAILURE_FORMAT "destination is" | |
885 | " read-modify-write\n", FAILURE_ARGS); | |
886 | return false; | |
887 | } | |
888 | } | |
889 | else | |
890 | { | |
891 | /* The instruction can set additional registers, provided that | |
892 | they're call-clobbered hard registers. This is useful for | |
893 | instructions that alter the condition codes. */ | |
894 | if (!HARD_REGISTER_NUM_P (def_regno)) | |
895 | { | |
896 | if (dump_file) | |
897 | fprintf (dump_file, FAILURE_FORMAT "insn also sets" | |
898 | " pseudo reg %d\n", FAILURE_ARGS, def_regno); | |
899 | return false; | |
900 | } | |
901 | if (global_regs[def_regno]) | |
902 | { | |
903 | if (dump_file) | |
904 | fprintf (dump_file, FAILURE_FORMAT "insn also sets" | |
905 | " global reg %d\n", FAILURE_ARGS, def_regno); | |
906 | return false; | |
907 | } | |
908 | if (!TEST_HARD_REG_BIT (regs_invalidated_by_call, def_regno)) | |
909 | { | |
910 | if (dump_file) | |
911 | fprintf (dump_file, FAILURE_FORMAT "insn also sets" | |
912 | " call-preserved reg %d\n", FAILURE_ARGS, def_regno); | |
913 | return false; | |
914 | } | |
915 | } | |
916 | } | |
917 | ||
918 | /* If the instruction uses fixed hard registers, check that those | |
919 | registers have the same value throughout the function. If the | |
920 | instruction uses non-fixed hard registers, check that we can | |
921 | replace them with pseudos. */ | |
922 | FOR_EACH_INSN_USE (ref, insn) | |
923 | { | |
924 | unsigned int use_regno = DF_REF_REGNO (ref); | |
925 | if (HARD_REGISTER_NUM_P (use_regno) && fixed_regs[use_regno]) | |
926 | { | |
927 | if (rtx_unstable_p (DF_REF_REAL_REG (ref))) | |
928 | { | |
929 | if (dump_file) | |
930 | fprintf (dump_file, FAILURE_FORMAT "insn uses fixed hard reg" | |
931 | " %d\n", FAILURE_ARGS, use_regno); | |
932 | return false; | |
933 | } | |
934 | } | |
935 | else if (HARD_REGISTER_NUM_P (use_regno)) | |
936 | { | |
937 | /* Allocate a dummy pseudo register and temporarily install it. | |
938 | Make the register number depend on the mode, which should | |
939 | provide enough sharing for match_dup while also weeding | |
940 | out cases in which operands with different modes are | |
941 | explicitly tied. */ | |
942 | rtx *loc = DF_REF_REAL_LOC (ref); | |
943 | unsigned int size = RTX_CODE_SIZE (REG); | |
944 | rtx new_reg = (rtx) alloca (size); | |
945 | memset (new_reg, 0, size); | |
946 | PUT_CODE (new_reg, REG); | |
947 | set_mode_and_regno (new_reg, GET_MODE (*loc), | |
948 | LAST_VIRTUAL_REGISTER + 1 + GET_MODE (*loc)); | |
949 | validate_change (insn, loc, new_reg, 1); | |
950 | } | |
951 | } | |
952 | bool ok_p = verify_changes (0); | |
953 | cancel_changes (0); | |
954 | if (!ok_p) | |
955 | { | |
956 | if (dump_file) | |
957 | fprintf (dump_file, FAILURE_FORMAT "insn does not allow hard" | |
958 | " register inputs to be replaced\n", FAILURE_ARGS); | |
959 | return false; | |
960 | } | |
961 | ||
962 | #undef FAILURE_ARGS | |
963 | #undef FAILURE_FORMAT | |
964 | ||
965 | add_candidate (insn, regno, true); | |
966 | return true; | |
967 | } | |
968 | ||
969 | /* Calculate the set of rematerialization candidates. Return true if | |
970 | we find at least one. */ | |
971 | ||
972 | bool | |
973 | early_remat::collect_candidates (void) | |
974 | { | |
975 | unsigned int nregs = DF_REG_SIZE (df); | |
976 | for (unsigned int regno = FIRST_PSEUDO_REGISTER; regno < nregs; ++regno) | |
977 | if (interesting_regno_p (regno)) | |
978 | { | |
979 | /* Create candidates for all suitable definitions. */ | |
980 | bitmap_clear (&m_tmp_bitmap); | |
981 | unsigned int bad = 0; | |
982 | unsigned int id = 0; | |
983 | for (df_ref ref = DF_REG_DEF_CHAIN (regno); ref; | |
984 | ref = DF_REF_NEXT_REG (ref)) | |
985 | { | |
986 | rtx_insn *insn = DF_REF_INSN (ref); | |
987 | if (maybe_add_candidate (insn, regno)) | |
988 | bitmap_set_bit (&m_tmp_bitmap, id); | |
989 | else | |
990 | bad += 1; | |
991 | id += 1; | |
992 | } | |
993 | ||
994 | /* If we found at least one suitable definition, add dummy | |
995 | candidates for the rest, so that we can see which definitions | |
996 | are live where. */ | |
997 | if (!bitmap_empty_p (&m_tmp_bitmap) && bad) | |
998 | { | |
999 | id = 0; | |
1000 | for (df_ref ref = DF_REG_DEF_CHAIN (regno); ref; | |
1001 | ref = DF_REF_NEXT_REG (ref)) | |
1002 | { | |
1003 | if (!bitmap_bit_p (&m_tmp_bitmap, id)) | |
1004 | add_candidate (DF_REF_INSN (ref), regno, false); | |
1005 | id += 1; | |
1006 | } | |
1007 | } | |
1008 | } | |
1009 | ||
1010 | ||
1011 | return !m_candidates.is_empty (); | |
1012 | } | |
1013 | ||
1014 | /* Initialize the m_block_info array. */ | |
1015 | ||
1016 | void | |
1017 | early_remat::init_block_info (void) | |
1018 | { | |
1019 | unsigned int n_blocks = last_basic_block_for_fn (m_fn); | |
1020 | m_block_info.safe_grow_cleared (n_blocks); | |
1021 | } | |
1022 | ||
1023 | /* Maps basic block indices to their position in the post order. */ | |
1024 | static unsigned int *postorder_index; | |
1025 | ||
1026 | /* Order remat_candidates X_IN and Y_IN according to the cfg postorder. */ | |
1027 | ||
1028 | static int | |
1029 | compare_candidates (const void *x_in, const void *y_in) | |
1030 | { | |
1031 | const remat_candidate *x = (const remat_candidate *) x_in; | |
1032 | const remat_candidate *y = (const remat_candidate *) y_in; | |
1033 | basic_block x_bb = BLOCK_FOR_INSN (x->insn); | |
1034 | basic_block y_bb = BLOCK_FOR_INSN (y->insn); | |
1035 | if (x_bb != y_bb) | |
1036 | /* Make X and Y follow block postorder. */ | |
1037 | return postorder_index[x_bb->index] - postorder_index[y_bb->index]; | |
1038 | ||
1039 | /* Make X and Y follow a backward traversal of the containing block. */ | |
1040 | return DF_INSN_LUID (y->insn) - DF_INSN_LUID (x->insn); | |
1041 | } | |
1042 | ||
1043 | /* Sort the collected rematerialization candidates so that they follow | |
1044 | cfg postorder. */ | |
1045 | ||
1046 | void | |
1047 | early_remat::sort_candidates (void) | |
1048 | { | |
1049 | /* Make sure the DF LUIDs are up-to-date for all the blocks we | |
1050 | care about. */ | |
1051 | bitmap_clear (&m_tmp_bitmap); | |
1052 | unsigned int cand_index; | |
1053 | remat_candidate *cand; | |
1054 | FOR_EACH_VEC_ELT (m_candidates, cand_index, cand) | |
1055 | { | |
1056 | basic_block bb = BLOCK_FOR_INSN (cand->insn); | |
1057 | if (bitmap_set_bit (&m_tmp_bitmap, bb->index)) | |
1058 | df_recompute_luids (bb); | |
1059 | } | |
1060 | ||
1061 | /* Create a mapping from block numbers to their position in the | |
1062 | postorder. */ | |
1063 | unsigned int n_blocks = last_basic_block_for_fn (m_fn); | |
1064 | int *postorder = df_get_postorder (DF_BACKWARD); | |
1065 | unsigned int postorder_len = df_get_n_blocks (DF_BACKWARD); | |
1066 | postorder_index = new unsigned int[n_blocks]; | |
1067 | for (unsigned int i = 0; i < postorder_len; ++i) | |
1068 | postorder_index[postorder[i]] = i; | |
1069 | ||
1070 | m_candidates.qsort (compare_candidates); | |
1071 | ||
1072 | delete postorder_index; | |
1073 | } | |
1074 | ||
1075 | /* Commit to the current candidate indices and initialize cross-references. */ | |
1076 | ||
1077 | void | |
1078 | early_remat::finalize_candidate_indices (void) | |
1079 | { | |
1080 | /* Create a bitmap for each candidate register. */ | |
1081 | m_regno_to_candidates.safe_grow (max_reg_num ()); | |
1082 | unsigned int regno; | |
1083 | bitmap_iterator bi; | |
1084 | EXECUTE_IF_SET_IN_BITMAP (&m_candidate_regnos, 0, regno, bi) | |
1085 | m_regno_to_candidates[regno] = alloc_bitmap (); | |
1086 | ||
1087 | /* Go through each candidate and record its index. */ | |
1088 | unsigned int cand_index; | |
1089 | remat_candidate *cand; | |
1090 | FOR_EACH_VEC_ELT (m_candidates, cand_index, cand) | |
1091 | { | |
1092 | basic_block bb = BLOCK_FOR_INSN (cand->insn); | |
1093 | remat_block_info *info = &m_block_info[bb->index]; | |
1094 | info->num_candidates += 1; | |
1095 | info->first_candidate = cand_index; | |
1096 | bitmap_set_bit (m_regno_to_candidates[cand->regno], cand_index); | |
1097 | } | |
1098 | } | |
1099 | ||
1100 | /* Record that candidates CAND1_INDEX and CAND2_INDEX are equivalent. | |
1101 | CAND1_INDEX might already have an equivalence class, but CAND2_INDEX | |
1102 | doesn't. */ | |
1103 | ||
1104 | void | |
1105 | early_remat::record_equiv_candidates (unsigned int cand1_index, | |
1106 | unsigned int cand2_index) | |
1107 | { | |
1108 | if (dump_file) | |
1109 | fprintf (dump_file, ";; Candidate %d is equivalent to candidate %d\n", | |
1110 | cand2_index, cand1_index); | |
1111 | ||
1112 | remat_candidate *cand1 = &m_candidates[cand1_index]; | |
1113 | remat_candidate *cand2 = &m_candidates[cand2_index]; | |
1114 | gcc_checking_assert (!cand2->equiv_class); | |
1115 | ||
1116 | remat_equiv_class *ec = cand1->equiv_class; | |
1117 | if (!ec) | |
1118 | { | |
1119 | ec = XOBNEW (&m_obstack.obstack, remat_equiv_class); | |
1120 | ec->members = alloc_bitmap (); | |
1121 | bitmap_set_bit (ec->members, cand1_index); | |
1122 | ec->earliest = cand1_index; | |
1123 | ec->representative = cand1_index; | |
1124 | cand1->equiv_class = ec; | |
1125 | } | |
1126 | cand1 = &m_candidates[ec->representative]; | |
1127 | cand2->equiv_class = ec; | |
1128 | bitmap_set_bit (ec->members, cand2_index); | |
1129 | if (cand2_index > ec->representative) | |
1130 | ec->representative = cand2_index; | |
1131 | } | |
1132 | ||
1133 | /* Propagate information from the rd_out set of E->src to the rd_in set | |
1134 | of E->dest, when computing global reaching definitions. Return true | |
1135 | if something changed. */ | |
1136 | ||
1137 | bool | |
1138 | early_remat::rd_confluence_n (edge e) | |
1139 | { | |
1140 | remat_block_info *src = &er->m_block_info[e->src->index]; | |
1141 | remat_block_info *dest = &er->m_block_info[e->dest->index]; | |
1142 | ||
1143 | /* available_in temporarily contains the set of candidates whose | |
1144 | registers are live on entry. */ | |
1145 | if (empty_p (src->rd_out) || empty_p (dest->available_in)) | |
1146 | return false; | |
1147 | ||
1148 | return bitmap_ior_and_into (er->get_bitmap (&dest->rd_in), | |
1149 | src->rd_out, dest->available_in); | |
1150 | } | |
1151 | ||
1152 | /* Propagate information from the rd_in set of block BB_INDEX to rd_out. | |
1153 | Return true if something changed. */ | |
1154 | ||
1155 | bool | |
1156 | early_remat::rd_transfer (int bb_index) | |
1157 | { | |
1158 | remat_block_info *info = &er->m_block_info[bb_index]; | |
1159 | ||
1160 | if (empty_p (info->rd_in)) | |
1161 | return false; | |
1162 | ||
1163 | if (empty_p (info->rd_kill)) | |
1164 | { | |
1165 | gcc_checking_assert (empty_p (info->rd_gen)); | |
1166 | if (!info->rd_out) | |
1167 | info->rd_out = info->rd_in; | |
1168 | else | |
1169 | gcc_checking_assert (info->rd_out == info->rd_in); | |
1170 | /* Assume that we only get called if something changed. */ | |
1171 | return true; | |
1172 | } | |
1173 | ||
1174 | if (empty_p (info->rd_gen)) | |
1175 | return bitmap_and_compl (er->get_bitmap (&info->rd_out), | |
1176 | info->rd_in, info->rd_kill); | |
1177 | ||
1178 | return bitmap_ior_and_compl (er->get_bitmap (&info->rd_out), info->rd_gen, | |
1179 | info->rd_in, info->rd_kill); | |
1180 | } | |
1181 | ||
1182 | /* Calculate the rd_* sets for each block. */ | |
1183 | ||
1184 | void | |
1185 | early_remat::compute_rd (void) | |
1186 | { | |
1187 | /* First calculate the rd_kill and rd_gen sets, using the fact | |
1188 | that m_candidates is sorted in order of decreasing LUID. */ | |
1189 | unsigned int cand_index; | |
1190 | remat_candidate *cand; | |
1191 | FOR_EACH_VEC_ELT_REVERSE (m_candidates, cand_index, cand) | |
1192 | { | |
1193 | rtx_insn *insn = cand->insn; | |
1194 | basic_block bb = BLOCK_FOR_INSN (insn); | |
1195 | remat_block_info *info = &m_block_info[bb->index]; | |
1196 | bitmap kill = m_regno_to_candidates[cand->regno]; | |
1197 | bitmap_ior_into (get_bitmap (&info->rd_kill), kill); | |
1198 | if (bitmap_bit_p (DF_LR_OUT (bb), cand->regno)) | |
1199 | { | |
1200 | bitmap_and_compl_into (get_bitmap (&info->rd_gen), kill); | |
1201 | bitmap_set_bit (info->rd_gen, cand_index); | |
1202 | } | |
1203 | } | |
1204 | ||
1205 | /* Set up the initial values of the other sets. */ | |
1206 | basic_block bb; | |
1207 | FOR_EACH_BB_FN (bb, m_fn) | |
1208 | { | |
1209 | remat_block_info *info = &m_block_info[bb->index]; | |
1210 | unsigned int regno; | |
1211 | bitmap_iterator bi; | |
1212 | EXECUTE_IF_AND_IN_BITMAP (DF_LR_IN (bb), &m_candidate_regnos, | |
1213 | 0, regno, bi) | |
1214 | { | |
1215 | /* Use available_in to record the set of candidates whose | |
1216 | registers are live on entry (i.e. a maximum bound on rd_in). */ | |
1217 | bitmap_ior_into (get_bitmap (&info->available_in), | |
1218 | m_regno_to_candidates[regno]); | |
1219 | ||
1220 | /* Add registers that die in a block to the block's kill set, | |
1221 | so that we don't needlessly propagate them through the rest | |
1222 | of the function. */ | |
1223 | if (!bitmap_bit_p (DF_LR_OUT (bb), regno)) | |
1224 | bitmap_ior_into (get_bitmap (&info->rd_kill), | |
1225 | m_regno_to_candidates[regno]); | |
1226 | } | |
1227 | ||
1228 | /* Initialize each block's rd_out to the minimal set (the set of | |
1229 | local definitions). */ | |
1230 | if (!empty_p (info->rd_gen)) | |
1231 | bitmap_copy (get_bitmap (&info->rd_out), info->rd_gen); | |
1232 | } | |
1233 | ||
1234 | /* Iterate until we reach a fixed point. */ | |
1235 | er = this; | |
1236 | bitmap_clear (&m_tmp_bitmap); | |
1237 | bitmap_set_range (&m_tmp_bitmap, 0, last_basic_block_for_fn (m_fn)); | |
1238 | df_simple_dataflow (DF_FORWARD, NULL, NULL, rd_confluence_n, rd_transfer, | |
1239 | &m_tmp_bitmap, df_get_postorder (DF_FORWARD), | |
1240 | df_get_n_blocks (DF_FORWARD)); | |
1241 | er = 0; | |
1242 | ||
1243 | /* Work out which definitions reach which candidates, again taking | |
1244 | advantage of the candidate order. */ | |
1245 | bitmap_head reaching; | |
1246 | bitmap_initialize (&reaching, &m_obstack); | |
1247 | basic_block old_bb = NULL; | |
1248 | FOR_EACH_VEC_ELT_REVERSE (m_candidates, cand_index, cand) | |
1249 | { | |
1250 | bb = BLOCK_FOR_INSN (cand->insn); | |
1251 | if (bb != old_bb) | |
1252 | { | |
1253 | /* Get the definitions that reach the start of the new block. */ | |
1254 | remat_block_info *info = &m_block_info[bb->index]; | |
1255 | if (info->rd_in) | |
1256 | bitmap_copy (&reaching, info->rd_in); | |
1257 | else | |
1258 | bitmap_clear (&reaching); | |
1259 | old_bb = bb; | |
1260 | } | |
1261 | else | |
1262 | { | |
1263 | /* Process the definitions of the previous instruction. */ | |
1264 | bitmap kill = m_regno_to_candidates[cand[1].regno]; | |
1265 | bitmap_and_compl_into (&reaching, kill); | |
1266 | bitmap_set_bit (&reaching, cand_index + 1); | |
1267 | } | |
1268 | ||
1269 | if (cand->can_copy_p && !cand->constant_p) | |
1270 | { | |
1271 | df_ref ref; | |
1272 | FOR_EACH_INSN_USE (ref, cand->insn) | |
1273 | { | |
1274 | unsigned int regno = DF_REF_REGNO (ref); | |
1275 | if (bitmap_bit_p (&m_candidate_regnos, regno)) | |
1276 | { | |
1277 | bitmap defs = m_regno_to_candidates[regno]; | |
1278 | bitmap_and (&m_tmp_bitmap, defs, &reaching); | |
1279 | bitmap_ior_into (get_bitmap (&cand->uses), &m_tmp_bitmap); | |
1280 | } | |
1281 | } | |
1282 | } | |
1283 | } | |
1284 | bitmap_clear (&reaching); | |
1285 | } | |
1286 | ||
1287 | /* If CAND_INDEX is in an equivalence class, return the representative | |
1288 | of the class, otherwise return CAND_INDEX. */ | |
1289 | ||
1290 | inline unsigned int | |
1291 | early_remat::canon_candidate (unsigned int cand_index) | |
1292 | { | |
1293 | if (remat_equiv_class *ec = m_candidates[cand_index].equiv_class) | |
1294 | return ec->representative; | |
1295 | return cand_index; | |
1296 | } | |
1297 | ||
1298 | /* Make candidate set *PTR refer to candidates using the representative | |
1299 | of each equivalence class. */ | |
1300 | ||
1301 | void | |
1302 | early_remat::canon_bitmap (bitmap *ptr) | |
1303 | { | |
1304 | bitmap old_set = *ptr; | |
1305 | if (empty_p (old_set)) | |
1306 | return; | |
1307 | ||
1308 | bitmap new_set = NULL; | |
1309 | unsigned int old_index; | |
1310 | bitmap_iterator bi; | |
1311 | EXECUTE_IF_SET_IN_BITMAP (old_set, 0, old_index, bi) | |
1312 | { | |
1313 | unsigned int new_index = canon_candidate (old_index); | |
1314 | if (old_index != new_index) | |
1315 | { | |
1316 | if (!new_set) | |
1317 | { | |
1318 | new_set = alloc_bitmap (); | |
1319 | bitmap_copy (new_set, old_set); | |
1320 | } | |
1321 | bitmap_clear_bit (new_set, old_index); | |
1322 | bitmap_set_bit (new_set, new_index); | |
1323 | } | |
1324 | } | |
1325 | if (new_set) | |
1326 | { | |
1327 | BITMAP_FREE (*ptr); | |
1328 | *ptr = new_set; | |
1329 | } | |
1330 | } | |
1331 | ||
1332 | /* If the candidates in REACHING all have the same value, return the | |
1333 | earliest instance of that value (i.e. the first one to be added | |
1334 | to m_value_table), otherwise return MULTIPLE_CANDIDATES. */ | |
1335 | ||
1336 | unsigned int | |
1337 | early_remat::resolve_reaching_def (bitmap reaching) | |
1338 | { | |
1339 | unsigned int cand_index = bitmap_first_set_bit (reaching); | |
1340 | if (remat_equiv_class *ec = m_candidates[cand_index].equiv_class) | |
1341 | { | |
1342 | if (!bitmap_intersect_compl_p (reaching, ec->members)) | |
1343 | return ec->earliest; | |
1344 | } | |
1345 | else if (bitmap_single_bit_set_p (reaching)) | |
1346 | return cand_index; | |
1347 | ||
1348 | return MULTIPLE_CANDIDATES; | |
1349 | } | |
1350 | ||
1351 | /* Check whether all candidate registers used by candidate CAND_INDEX have | |
1352 | unique definitions. Return true if so, replacing the candidate's uses | |
1353 | set with the appropriate form for value numbering. */ | |
1354 | ||
1355 | bool | |
1356 | early_remat::check_candidate_uses (unsigned int cand_index) | |
1357 | { | |
1358 | remat_candidate *cand = &m_candidates[cand_index]; | |
1359 | ||
1360 | /* Process the uses for each register in turn. */ | |
1361 | bitmap_head uses; | |
1362 | bitmap_initialize (&uses, &m_obstack); | |
1363 | bitmap_copy (&uses, cand->uses); | |
1364 | bitmap uses_ec = alloc_bitmap (); | |
1365 | while (!bitmap_empty_p (&uses)) | |
1366 | { | |
1367 | /* Get the register for the lowest-indexed candidate remaining, | |
1368 | and the reaching definitions of that register. */ | |
1369 | unsigned int first = bitmap_first_set_bit (&uses); | |
1370 | unsigned int regno = m_candidates[first].regno; | |
1371 | bitmap_and (&m_tmp_bitmap, &uses, m_regno_to_candidates[regno]); | |
1372 | ||
1373 | /* See whether all reaching definitions have the same value and if | |
1374 | so get the index of the first candidate we saw with that value. */ | |
1375 | unsigned int def = resolve_reaching_def (&m_tmp_bitmap); | |
1376 | if (def == MULTIPLE_CANDIDATES) | |
1377 | { | |
1378 | if (dump_file) | |
1379 | fprintf (dump_file, ";; Removing candidate %d because there is" | |
1380 | " more than one reaching definition of reg %d\n", | |
1381 | cand_index, regno); | |
1382 | cand->can_copy_p = false; | |
1383 | break; | |
1384 | } | |
1385 | bitmap_set_bit (uses_ec, def); | |
1386 | bitmap_and_compl_into (&uses, &m_tmp_bitmap); | |
1387 | } | |
1388 | BITMAP_FREE (cand->uses); | |
1389 | cand->uses = uses_ec; | |
1390 | return cand->can_copy_p; | |
1391 | } | |
1392 | ||
1393 | /* Calculate the set of hard registers that would be clobbered by | |
1394 | rematerializing candidate CAND_INDEX. At this point the candidate's | |
1395 | set of uses is final. */ | |
1396 | ||
1397 | void | |
1398 | early_remat::compute_clobbers (unsigned int cand_index) | |
1399 | { | |
1400 | remat_candidate *cand = &m_candidates[cand_index]; | |
1401 | if (cand->uses) | |
1402 | { | |
1403 | unsigned int use_index; | |
1404 | bitmap_iterator bi; | |
1405 | EXECUTE_IF_SET_IN_BITMAP (cand->uses, 0, use_index, bi) | |
1406 | if (bitmap clobbers = m_candidates[use_index].clobbers) | |
1407 | bitmap_ior_into (get_bitmap (&cand->clobbers), clobbers); | |
1408 | } | |
1409 | ||
1410 | df_ref ref; | |
1411 | FOR_EACH_INSN_DEF (ref, cand->insn) | |
1412 | { | |
1413 | unsigned int def_regno = DF_REF_REGNO (ref); | |
1414 | if (def_regno != cand->regno) | |
1415 | bitmap_set_bit (get_bitmap (&cand->clobbers), def_regno); | |
1416 | } | |
1417 | } | |
1418 | ||
1419 | /* Mark candidate CAND_INDEX as validated and add it to the value table. */ | |
1420 | ||
1421 | void | |
1422 | early_remat::assign_value_number (unsigned int cand_index) | |
1423 | { | |
1424 | remat_candidate *cand = &m_candidates[cand_index]; | |
1425 | gcc_checking_assert (cand->can_copy_p && !cand->validated_p); | |
1426 | ||
1427 | compute_clobbers (cand_index); | |
1428 | cand->validated_p = true; | |
1429 | ||
1430 | inchash::hash h; | |
1431 | h.add_int (cand->regno); | |
1432 | inchash::add_rtx (cand->remat_rtx, h); | |
1433 | cand->hash = h.end (); | |
1434 | ||
1435 | remat_candidate **slot | |
1436 | = m_value_table.find_slot_with_hash (cand, cand->hash, INSERT); | |
1437 | if (!*slot) | |
1438 | { | |
1439 | *slot = cand; | |
1440 | if (dump_file) | |
1441 | fprintf (dump_file, ";; Candidate %d is not equivalent to" | |
1442 | " others seen so far\n", cand_index); | |
1443 | } | |
1444 | else | |
1445 | record_equiv_candidates (*slot - m_candidates.address (), cand_index); | |
1446 | } | |
1447 | ||
1448 | /* Make a final decision about which candidates are valid and assign | |
1449 | value numbers to those that are. */ | |
1450 | ||
1451 | void | |
1452 | early_remat::decide_candidate_validity (void) | |
1453 | { | |
1454 | auto_vec<unsigned int, 16> stack; | |
1455 | unsigned int cand1_index; | |
1456 | remat_candidate *cand1; | |
1457 | FOR_EACH_VEC_ELT_REVERSE (m_candidates, cand1_index, cand1) | |
1458 | { | |
1459 | if (!cand1->can_copy_p || cand1->validated_p) | |
1460 | continue; | |
1461 | ||
1462 | if (empty_p (cand1->uses)) | |
1463 | { | |
1464 | assign_value_number (cand1_index); | |
1465 | continue; | |
1466 | } | |
1467 | ||
1468 | stack.safe_push (cand1_index); | |
1469 | while (!stack.is_empty ()) | |
1470 | { | |
1471 | unsigned int cand2_index = stack.last (); | |
1472 | unsigned int watermark = stack.length (); | |
1473 | remat_candidate *cand2 = &m_candidates[cand2_index]; | |
1474 | if (!cand2->can_copy_p || cand2->validated_p) | |
1475 | { | |
1476 | stack.pop (); | |
1477 | continue; | |
1478 | } | |
1479 | cand2->visited_p = true; | |
1480 | unsigned int cand3_index; | |
1481 | bitmap_iterator bi; | |
1482 | EXECUTE_IF_SET_IN_BITMAP (cand2->uses, 0, cand3_index, bi) | |
1483 | { | |
1484 | remat_candidate *cand3 = &m_candidates[cand3_index]; | |
1485 | if (!cand3->can_copy_p) | |
1486 | { | |
1487 | if (dump_file) | |
1488 | fprintf (dump_file, ";; Removing candidate %d because" | |
1489 | " it uses removed candidate %d\n", cand2_index, | |
1490 | cand3_index); | |
1491 | cand2->can_copy_p = false; | |
1492 | break; | |
1493 | } | |
1494 | if (!cand3->validated_p) | |
1495 | { | |
1496 | if (empty_p (cand3->uses)) | |
1497 | assign_value_number (cand3_index); | |
1498 | else if (cand3->visited_p) | |
1499 | { | |
1500 | if (dump_file) | |
1501 | fprintf (dump_file, ";; Removing candidate %d" | |
1502 | " because its definition is cyclic\n", | |
1503 | cand2_index); | |
1504 | cand2->can_copy_p = false; | |
1505 | break; | |
1506 | } | |
1507 | else | |
1508 | stack.safe_push (cand3_index); | |
1509 | } | |
1510 | } | |
1511 | if (!cand2->can_copy_p) | |
1512 | { | |
1513 | cand2->visited_p = false; | |
1514 | stack.truncate (watermark - 1); | |
1515 | } | |
1516 | else if (watermark == stack.length ()) | |
1517 | { | |
1518 | cand2->visited_p = false; | |
1519 | if (check_candidate_uses (cand2_index)) | |
1520 | assign_value_number (cand2_index); | |
1521 | stack.pop (); | |
1522 | } | |
1523 | } | |
1524 | } | |
1525 | ||
1526 | /* Ensure that the candidates always use the same candidate index | |
1527 | to refer to an equivalence class. */ | |
1528 | FOR_EACH_VEC_ELT_REVERSE (m_candidates, cand1_index, cand1) | |
1529 | if (cand1->can_copy_p && !empty_p (cand1->uses)) | |
1530 | { | |
1531 | canon_bitmap (&cand1->uses); | |
1532 | gcc_checking_assert (bitmap_first_set_bit (cand1->uses) > cand1_index); | |
1533 | } | |
1534 | } | |
1535 | ||
1536 | /* Assuming that every path reaching a point P contains a copy of a | |
1537 | use U of REGNO, return true if another copy of U at P would have | |
1538 | access to the same value of REGNO. */ | |
1539 | ||
1540 | bool | |
1541 | early_remat::stable_use_p (unsigned int regno) | |
1542 | { | |
1543 | /* Conservatively assume not for hard registers. */ | |
1544 | if (HARD_REGISTER_NUM_P (regno)) | |
1545 | return false; | |
1546 | ||
1547 | /* See if REGNO has a single definition and is never used uninitialized. | |
1548 | In this case the definition of REGNO dominates the common dominator | |
1549 | of the uses U, which in turn dominates P. */ | |
1550 | if (DF_REG_DEF_COUNT (regno) == 1 | |
1551 | && !bitmap_bit_p (DF_LR_OUT (ENTRY_BLOCK_PTR_FOR_FN (m_fn)), regno)) | |
1552 | return true; | |
1553 | ||
1554 | return false; | |
1555 | } | |
1556 | ||
1557 | /* Emit a copy from register DEST to register SRC before candidate | |
1558 | CAND_INDEX's instruction. */ | |
1559 | ||
1560 | void | |
1561 | early_remat::emit_copy_before (unsigned int cand_index, rtx dest, rtx src) | |
1562 | { | |
1563 | remat_candidate *cand = &m_candidates[cand_index]; | |
1564 | if (dump_file) | |
1565 | { | |
1566 | fprintf (dump_file, ";; Stabilizing insn "); | |
1567 | dump_insn_id (cand->insn); | |
1568 | fprintf (dump_file, " by copying source reg %d:%s to temporary reg %d\n", | |
1569 | REGNO (src), GET_MODE_NAME (GET_MODE (src)), REGNO (dest)); | |
1570 | } | |
1571 | emit_insn_before (gen_move_insn (dest, src), cand->insn); | |
1572 | } | |
1573 | ||
1574 | /* Check whether any inputs to candidate CAND_INDEX's instruction could | |
1575 | change at rematerialization points and replace them with new pseudo | |
1576 | registers if so. */ | |
1577 | ||
1578 | void | |
1579 | early_remat::stabilize_pattern (unsigned int cand_index) | |
1580 | { | |
1581 | remat_candidate *cand = &m_candidates[cand_index]; | |
1582 | if (cand->stabilized_p) | |
1583 | return; | |
1584 | ||
1585 | remat_equiv_class *ec = cand->equiv_class; | |
1586 | gcc_checking_assert (!ec || cand_index == ec->representative); | |
1587 | ||
1588 | /* Record the replacements we've made so far, so that we don't | |
1589 | create two separate registers for match_dups. Lookup is O(n), | |
1590 | but the n is very small. */ | |
1591 | typedef std::pair<rtx, rtx> reg_pair; | |
1592 | auto_vec<reg_pair, 16> reg_map; | |
1593 | ||
1594 | rtx_insn *insn = cand->insn; | |
1595 | df_ref ref; | |
1596 | FOR_EACH_INSN_USE (ref, insn) | |
1597 | { | |
1598 | unsigned int old_regno = DF_REF_REGNO (ref); | |
1599 | rtx *loc = DF_REF_REAL_LOC (ref); | |
1600 | ||
1601 | if (HARD_REGISTER_NUM_P (old_regno) && fixed_regs[old_regno]) | |
1602 | { | |
1603 | /* We checked when adding the candidate that the value is stable. */ | |
1604 | gcc_checking_assert (!rtx_unstable_p (*loc)); | |
1605 | continue; | |
1606 | } | |
1607 | ||
1608 | if (bitmap_bit_p (&m_candidate_regnos, old_regno)) | |
1609 | /* We already know which candidate provides the definition | |
1610 | and will handle it during copying. */ | |
1611 | continue; | |
1612 | ||
1613 | if (stable_use_p (old_regno)) | |
1614 | /* We can continue to use the existing register. */ | |
1615 | continue; | |
1616 | ||
1617 | /* We need to replace the register. See whether we've already | |
1618 | created a suitable copy. */ | |
1619 | rtx old_reg = *loc; | |
1620 | rtx new_reg = NULL_RTX; | |
1621 | machine_mode mode = GET_MODE (old_reg); | |
1622 | reg_pair *p; | |
1623 | unsigned int pi; | |
1624 | FOR_EACH_VEC_ELT (reg_map, pi, p) | |
1625 | if (REGNO (p->first) == old_regno | |
1626 | && GET_MODE (p->first) == mode) | |
1627 | { | |
1628 | new_reg = p->second; | |
1629 | break; | |
1630 | } | |
1631 | ||
1632 | if (!new_reg) | |
1633 | { | |
1634 | /* Create a new register and initialize it just before | |
1635 | the instruction. */ | |
1636 | new_reg = gen_reg_rtx (mode); | |
1637 | reg_map.safe_push (reg_pair (old_reg, new_reg)); | |
1638 | if (ec) | |
1639 | { | |
1640 | unsigned int member_index; | |
1641 | bitmap_iterator bi; | |
1642 | EXECUTE_IF_SET_IN_BITMAP (ec->members, 0, member_index, bi) | |
1643 | emit_copy_before (member_index, new_reg, old_reg); | |
1644 | } | |
1645 | else | |
1646 | emit_copy_before (cand_index, new_reg, old_reg); | |
1647 | } | |
1648 | validate_change (insn, loc, new_reg, true); | |
1649 | } | |
1650 | if (num_changes_pending ()) | |
1651 | { | |
1652 | if (!apply_change_group ()) | |
1653 | /* We checked when adding the candidates that the pattern allows | |
1654 | hard registers to be replaced. Nothing else should make the | |
1655 | changes invalid. */ | |
1656 | gcc_unreachable (); | |
1657 | ||
1658 | if (ec) | |
1659 | { | |
1660 | /* Copy the new pattern to other members of the equivalence | |
1661 | class. */ | |
1662 | unsigned int member_index; | |
1663 | bitmap_iterator bi; | |
1664 | EXECUTE_IF_SET_IN_BITMAP (ec->members, 0, member_index, bi) | |
1665 | if (cand_index != member_index) | |
1666 | { | |
1667 | rtx_insn *other_insn = m_candidates[member_index].insn; | |
1668 | if (!validate_change (other_insn, &PATTERN (other_insn), | |
1669 | copy_insn (PATTERN (insn)), 0)) | |
1670 | /* If the original instruction was valid then the copy | |
1671 | should be too. */ | |
1672 | gcc_unreachable (); | |
1673 | } | |
1674 | } | |
1675 | } | |
1676 | ||
1677 | cand->stabilized_p = true; | |
1678 | } | |
1679 | ||
1680 | /* Change CAND's instruction so that it sets CAND->copy_regno instead | |
1681 | of CAND->regno. */ | |
1682 | ||
1683 | void | |
1684 | early_remat::replace_dest_with_copy (unsigned int cand_index) | |
1685 | { | |
1686 | remat_candidate *cand = &m_candidates[cand_index]; | |
1687 | df_ref def; | |
1688 | FOR_EACH_INSN_DEF (def, cand->insn) | |
1689 | if (DF_REF_REGNO (def) == cand->regno) | |
1690 | validate_change (cand->insn, DF_REF_REAL_LOC (def), | |
1691 | regno_reg_rtx[cand->copy_regno], 1); | |
1692 | } | |
1693 | ||
1694 | /* Make sure that the candidates used by candidate CAND_INDEX are available. | |
1695 | There are two ways of doing this for an input candidate I: | |
1696 | ||
1697 | (1) Using the existing register number and ensuring that I is available. | |
1698 | ||
1699 | (2) Using a new register number (recorded in copy_regno) and adding I | |
1700 | to VIA_COPY. This guarantees that making I available does not | |
1701 | conflict with other uses of the original register. | |
1702 | ||
1703 | REQUIRED is the set of candidates that are required but not available | |
1704 | before the copy of CAND_INDEX. AVAILABLE is the set of candidates | |
1705 | that are already available before the copy of CAND_INDEX. REACHING | |
1706 | is the set of candidates that reach the copy of CAND_INDEX. VIA_COPY | |
1707 | is the set of candidates that will use new register numbers recorded | |
1708 | in copy_regno instead of the original ones. */ | |
1709 | ||
1710 | void | |
1711 | early_remat::stabilize_candidate_uses (unsigned int cand_index, | |
1712 | bitmap required, bitmap available, | |
1713 | bitmap reaching, bitmap via_copy) | |
1714 | { | |
1715 | remat_candidate *cand = &m_candidates[cand_index]; | |
1716 | df_ref use; | |
1717 | FOR_EACH_INSN_USE (use, cand->insn) | |
1718 | { | |
1719 | unsigned int regno = DF_REF_REGNO (use); | |
1720 | if (!bitmap_bit_p (&m_candidate_regnos, regno)) | |
1721 | continue; | |
1722 | ||
1723 | /* Work out which candidate provides the definition. */ | |
1724 | bitmap defs = m_regno_to_candidates[regno]; | |
1725 | bitmap_and (&m_tmp_bitmap, cand->uses, defs); | |
1726 | gcc_checking_assert (bitmap_single_bit_set_p (&m_tmp_bitmap)); | |
1727 | unsigned int def_index = bitmap_first_set_bit (&m_tmp_bitmap); | |
1728 | ||
1729 | /* First see if DEF_INDEX is the only reaching definition of REGNO | |
1730 | at this point too and if it is or will become available. We can | |
1731 | continue to use REGNO if so. */ | |
1732 | bitmap_and (&m_tmp_bitmap, reaching, defs); | |
1733 | if (bitmap_single_bit_set_p (&m_tmp_bitmap) | |
1734 | && bitmap_first_set_bit (&m_tmp_bitmap) == def_index | |
1735 | && ((available && bitmap_bit_p (available, def_index)) | |
1736 | || bitmap_bit_p (required, def_index))) | |
1737 | { | |
1738 | if (dump_file) | |
1739 | fprintf (dump_file, ";; Keeping reg %d for use of candidate %d" | |
1740 | " in candidate %d\n", regno, def_index, cand_index); | |
1741 | continue; | |
1742 | } | |
1743 | ||
1744 | /* Otherwise fall back to using a copy. There are other cases | |
1745 | in which we *could* continue to use REGNO, but there's not | |
1746 | really much point. Using a separate register ought to make | |
1747 | things easier for the register allocator. */ | |
1748 | remat_candidate *def_cand = &m_candidates[def_index]; | |
1749 | rtx *loc = DF_REF_REAL_LOC (use); | |
1750 | rtx new_reg; | |
1751 | if (bitmap_set_bit (via_copy, def_index)) | |
1752 | { | |
1753 | new_reg = gen_reg_rtx (GET_MODE (*loc)); | |
1754 | def_cand->copy_regno = REGNO (new_reg); | |
1755 | if (dump_file) | |
1756 | fprintf (dump_file, ";; Creating reg %d for use of candidate %d" | |
1757 | " in candidate %d\n", REGNO (new_reg), def_index, | |
1758 | cand_index); | |
1759 | } | |
1760 | else | |
1761 | new_reg = regno_reg_rtx[def_cand->copy_regno]; | |
1762 | validate_change (cand->insn, loc, new_reg, 1); | |
1763 | } | |
1764 | } | |
1765 | ||
1766 | /* Rematerialize the candidates in REQUIRED after instruction INSN, | |
1767 | given that the candidates in AVAILABLE are already available | |
1768 | and that REACHING is the set of candidates live after INSN. | |
1769 | REQUIRED and AVAILABLE are disjoint on entry. | |
1770 | ||
1771 | Clear REQUIRED on exit. */ | |
1772 | ||
1773 | void | |
1774 | early_remat::emit_remat_insns (bitmap required, bitmap available, | |
1775 | bitmap reaching, rtx_insn *insn) | |
1776 | { | |
1777 | /* Quick exit if there's nothing to do. */ | |
1778 | if (empty_p (required)) | |
1779 | return; | |
1780 | ||
1781 | /* Only reaching definitions should be available or required. */ | |
1782 | gcc_checking_assert (!bitmap_intersect_compl_p (required, reaching)); | |
1783 | if (available) | |
1784 | gcc_checking_assert (!bitmap_intersect_compl_p (available, reaching)); | |
1785 | ||
1786 | bitmap_head via_copy; | |
1787 | bitmap_initialize (&via_copy, &m_obstack); | |
1788 | while (!bitmap_empty_p (required) || !bitmap_empty_p (&via_copy)) | |
1789 | { | |
1790 | /* Pick the lowest-indexed candidate left. */ | |
1791 | unsigned int required_index = (bitmap_empty_p (required) | |
1792 | ? ~0U : bitmap_first_set_bit (required)); | |
1793 | unsigned int via_copy_index = (bitmap_empty_p (&via_copy) | |
1794 | ? ~0U : bitmap_first_set_bit (&via_copy)); | |
1795 | unsigned int cand_index = MIN (required_index, via_copy_index); | |
1796 | remat_candidate *cand = &m_candidates[cand_index]; | |
1797 | ||
1798 | bool via_copy_p = (cand_index == via_copy_index); | |
1799 | if (via_copy_p) | |
1800 | bitmap_clear_bit (&via_copy, cand_index); | |
1801 | else | |
1802 | { | |
1803 | /* Remove all candidates for the same register from REQUIRED. */ | |
1804 | bitmap_and (&m_tmp_bitmap, reaching, | |
1805 | m_regno_to_candidates[cand->regno]); | |
1806 | bitmap_and_compl_into (required, &m_tmp_bitmap); | |
1807 | gcc_checking_assert (!bitmap_bit_p (required, cand_index)); | |
1808 | ||
1809 | /* Only rematerialize if we have a single reaching definition | |
1810 | of the register. */ | |
1811 | if (!bitmap_single_bit_set_p (&m_tmp_bitmap)) | |
1812 | { | |
1813 | if (dump_file) | |
1814 | { | |
1815 | fprintf (dump_file, ";; Can't rematerialize reg %d after ", | |
1816 | cand->regno); | |
1817 | dump_insn_id (insn); | |
1818 | fprintf (dump_file, ": more than one reaching definition\n"); | |
1819 | } | |
1820 | continue; | |
1821 | } | |
1822 | ||
1823 | /* Skip candidates that can't be rematerialized. */ | |
1824 | if (!cand->can_copy_p) | |
1825 | continue; | |
1826 | ||
1827 | /* Check the function precondition. */ | |
1828 | gcc_checking_assert (!available | |
1829 | || !bitmap_bit_p (available, cand_index)); | |
1830 | } | |
1831 | ||
1832 | /* Invalid candidates should have been weeded out by now. */ | |
1833 | gcc_assert (cand->can_copy_p); | |
1834 | ||
1835 | rtx new_pattern; | |
1836 | if (cand->constant_p) | |
1837 | { | |
1838 | /* Emit a simple move. */ | |
1839 | unsigned int regno = via_copy_p ? cand->copy_regno : cand->regno; | |
1840 | new_pattern = gen_move_insn (regno_reg_rtx[regno], cand->remat_rtx); | |
1841 | } | |
1842 | else | |
1843 | { | |
1844 | /* If this is the first time we've copied the instruction, make | |
1845 | sure that any inputs will have the same value after INSN. */ | |
1846 | stabilize_pattern (cand_index); | |
1847 | ||
1848 | /* Temporarily adjust the original instruction so that it has | |
1849 | the right form for the copy. */ | |
1850 | if (via_copy_p) | |
1851 | replace_dest_with_copy (cand_index); | |
1852 | if (cand->uses) | |
1853 | stabilize_candidate_uses (cand_index, required, available, | |
1854 | reaching, &via_copy); | |
1855 | ||
1856 | /* Get the new instruction pattern. */ | |
1857 | new_pattern = copy_insn (cand->remat_rtx); | |
1858 | ||
1859 | /* Undo the temporary changes. */ | |
1860 | cancel_changes (0); | |
1861 | } | |
1862 | ||
1863 | /* Emit the new instruction. */ | |
1864 | rtx_insn *new_insn = emit_insn_after (new_pattern, insn); | |
1865 | ||
1866 | if (dump_file) | |
1867 | { | |
1868 | fprintf (dump_file, ";; Rematerializing candidate %d after ", | |
1869 | cand_index); | |
1870 | dump_insn_id (insn); | |
1871 | if (via_copy_p) | |
1872 | fprintf (dump_file, " with new destination reg %d", | |
1873 | cand->copy_regno); | |
1874 | fprintf (dump_file, ":\n\n"); | |
1875 | print_rtl_single (dump_file, new_insn); | |
1876 | fprintf (dump_file, "\n"); | |
1877 | } | |
1878 | } | |
1879 | } | |
1880 | ||
1881 | /* Recompute INFO's available_out set, given that it's distinct from | |
1882 | available_in and available_locally. */ | |
1883 | ||
1884 | bool | |
1885 | early_remat::set_available_out (remat_block_info *info) | |
1886 | { | |
1887 | if (empty_p (info->available_locally)) | |
1888 | return bitmap_and_compl (get_bitmap (&info->available_out), | |
1889 | info->available_in, info->rd_kill); | |
1890 | ||
1891 | if (empty_p (info->rd_kill)) | |
1892 | return bitmap_ior (get_bitmap (&info->available_out), | |
1893 | info->available_locally, info->available_in); | |
1894 | ||
1895 | return bitmap_ior_and_compl (get_bitmap (&info->available_out), | |
1896 | info->available_locally, info->available_in, | |
1897 | info->rd_kill); | |
1898 | } | |
1899 | ||
1900 | /* If BB has more than one call, decide which candidates should be | |
1901 | rematerialized after the non-final calls and emit the associated | |
1902 | instructions. Record other information about the block in preparation | |
1903 | for the global phase. */ | |
1904 | ||
1905 | void | |
1906 | early_remat::process_block (basic_block bb) | |
1907 | { | |
1908 | remat_block_info *info = &m_block_info[bb->index]; | |
1909 | rtx_insn *last_call = NULL; | |
1910 | rtx_insn *insn; | |
1911 | ||
1912 | /* Ensure that we always use the same candidate index to refer to an | |
1913 | equivalence class. */ | |
1914 | if (info->rd_out == info->rd_in) | |
1915 | { | |
1916 | canon_bitmap (&info->rd_in); | |
1917 | info->rd_out = info->rd_in; | |
1918 | } | |
1919 | else | |
1920 | { | |
1921 | canon_bitmap (&info->rd_in); | |
1922 | canon_bitmap (&info->rd_out); | |
1923 | } | |
1924 | canon_bitmap (&info->rd_kill); | |
1925 | canon_bitmap (&info->rd_gen); | |
1926 | ||
1927 | /* The set of candidates that should be rematerialized on entry to the | |
1928 | block or after the previous call (whichever is more recent). */ | |
1929 | init_temp_bitmap (&m_required); | |
1930 | ||
1931 | /* The set of candidates that reach the current instruction (i.e. are | |
1932 | live just before the instruction). */ | |
1933 | bitmap_head reaching; | |
1934 | bitmap_initialize (&reaching, &m_obstack); | |
1935 | if (info->rd_in) | |
1936 | bitmap_copy (&reaching, info->rd_in); | |
1937 | ||
1938 | /* The set of candidates that are live and available without | |
1939 | rematerialization just before the current instruction. This only | |
1940 | accounts for earlier candidates in the block, or those that become | |
1941 | available by being added to M_REQUIRED. */ | |
1942 | init_temp_bitmap (&m_available); | |
1943 | ||
1944 | /* Get the range of candidates in the block. */ | |
1945 | unsigned int next_candidate = info->first_candidate; | |
1946 | unsigned int num_candidates = info->num_candidates; | |
1947 | remat_candidate *next_def = (num_candidates > 0 | |
1948 | ? &m_candidates[next_candidate] | |
1949 | : NULL); | |
1950 | ||
1951 | FOR_BB_INSNS (bb, insn) | |
1952 | { | |
1953 | if (!NONDEBUG_INSN_P (insn)) | |
1954 | continue; | |
1955 | ||
1956 | /* First process uses, since this is a forward walk. */ | |
1957 | df_ref ref; | |
1958 | FOR_EACH_INSN_USE (ref, insn) | |
1959 | { | |
1960 | unsigned int regno = DF_REF_REGNO (ref); | |
1961 | if (bitmap_bit_p (&m_candidate_regnos, regno)) | |
1962 | { | |
1963 | bitmap defs = m_regno_to_candidates[regno]; | |
1964 | bitmap_and (&m_tmp_bitmap, defs, &reaching); | |
1965 | gcc_checking_assert (!bitmap_empty_p (&m_tmp_bitmap)); | |
1966 | if (!bitmap_intersect_p (defs, m_available)) | |
1967 | { | |
1968 | /* There has been no definition of the register since | |
1969 | the last call or the start of the block (whichever | |
1970 | is most recent). Mark the reaching definitions | |
1971 | as required at that point and thus available here. */ | |
1972 | bitmap_ior_into (m_required, &m_tmp_bitmap); | |
1973 | bitmap_ior_into (m_available, &m_tmp_bitmap); | |
1974 | } | |
1975 | } | |
1976 | } | |
1977 | ||
1978 | if (CALL_P (insn)) | |
1979 | { | |
1980 | if (!last_call) | |
1981 | { | |
1982 | /* The first call in the block. Record which candidates are | |
1983 | required at the start of the block. */ | |
1984 | copy_temp_bitmap (&info->required_in, &m_required); | |
1985 | init_temp_bitmap (&m_required); | |
1986 | } | |
1987 | else | |
1988 | /* The fully-local case: candidates that need to be | |
1989 | rematerialized after a previous call in the block. */ | |
1990 | emit_remat_insns (m_required, NULL, info->rd_after_call, | |
1991 | last_call); | |
1992 | last_call = insn; | |
1993 | bitmap_clear (m_available); | |
1994 | gcc_checking_assert (empty_p (m_required)); | |
1995 | } | |
1996 | ||
1997 | /* Now process definitions. */ | |
1998 | if (next_def && insn == next_def->insn) | |
1999 | { | |
2000 | unsigned int gen = canon_candidate (next_candidate); | |
2001 | ||
2002 | /* Other candidates with the same regno are not available | |
2003 | any more. */ | |
2004 | bitmap kill = m_regno_to_candidates[next_def->regno]; | |
2005 | bitmap_and_compl_into (m_available, kill); | |
2006 | bitmap_and_compl_into (&reaching, kill); | |
2007 | ||
2008 | /* Record that this candidate is available without | |
2009 | rematerialization. */ | |
2010 | bitmap_set_bit (m_available, gen); | |
2011 | bitmap_set_bit (&reaching, gen); | |
2012 | ||
2013 | /* Find the next candidate in the block. */ | |
2014 | num_candidates -= 1; | |
2015 | next_candidate -= 1; | |
2016 | if (num_candidates > 0) | |
2017 | next_def -= 1; | |
2018 | else | |
2019 | next_def = NULL; | |
2020 | } | |
2021 | ||
2022 | if (insn == last_call) | |
2023 | bitmap_copy (get_bitmap (&info->rd_after_call), &reaching); | |
2024 | } | |
2025 | bitmap_clear (&reaching); | |
2026 | gcc_checking_assert (num_candidates == 0); | |
2027 | ||
2028 | /* Remove values from the available set if they aren't live (and so | |
2029 | aren't interesting to successor blocks). */ | |
2030 | if (info->rd_out) | |
2031 | bitmap_and_into (m_available, info->rd_out); | |
2032 | ||
2033 | /* Record the accumulated information. */ | |
2034 | info->last_call = last_call; | |
2035 | info->abnormal_call_p = (last_call | |
2036 | && last_call == BB_END (bb) | |
2037 | && has_abnormal_or_eh_outgoing_edge_p (bb)); | |
2038 | copy_temp_bitmap (&info->available_locally, &m_available); | |
2039 | if (last_call) | |
2040 | copy_temp_bitmap (&info->required_after_call, &m_required); | |
2041 | else | |
2042 | copy_temp_bitmap (&info->required_in, &m_required); | |
2043 | ||
2044 | /* Assume at first that all live-in values are available without | |
2045 | rematerialization (i.e. start with the most optimistic assumption). */ | |
2046 | if (info->available_in) | |
2047 | { | |
2048 | if (info->rd_in) | |
2049 | bitmap_copy (info->available_in, info->rd_in); | |
2050 | else | |
2051 | BITMAP_FREE (info->available_in); | |
2052 | } | |
2053 | ||
2054 | if (last_call || empty_p (info->available_in)) | |
2055 | /* The values available on exit from the block are exactly those that | |
2056 | are available locally. This set doesn't change. */ | |
2057 | info->available_out = info->available_locally; | |
2058 | else if (empty_p (info->available_locally) && empty_p (info->rd_kill)) | |
2059 | /* The values available on exit are the same as those available on entry. | |
2060 | Updating one updates the other. */ | |
2061 | info->available_out = info->available_in; | |
2062 | else | |
2063 | set_available_out (info); | |
2064 | } | |
2065 | ||
2066 | /* Process each block as for process_block, visiting dominators before | |
2067 | the blocks they dominate. */ | |
2068 | ||
2069 | void | |
2070 | early_remat::local_phase (void) | |
2071 | { | |
2072 | if (dump_file) | |
2073 | fprintf (dump_file, "\n;; Local phase:\n"); | |
2074 | ||
2075 | int *postorder = df_get_postorder (DF_BACKWARD); | |
2076 | unsigned int postorder_len = df_get_n_blocks (DF_BACKWARD); | |
2077 | for (unsigned int i = postorder_len; i-- > 0; ) | |
2078 | if (postorder[i] >= NUM_FIXED_BLOCKS) | |
2079 | process_block (BASIC_BLOCK_FOR_FN (m_fn, postorder[i])); | |
2080 | } | |
2081 | ||
2082 | /* Return true if available values survive across edge E. */ | |
2083 | ||
2084 | static inline bool | |
2085 | available_across_edge_p (edge e) | |
2086 | { | |
2087 | return (e->flags & EDGE_EH) == 0; | |
2088 | } | |
2089 | ||
2090 | /* Propagate information from the available_out set of E->src to the | |
2091 | available_in set of E->dest, when computing global availability. | |
2092 | Return true if something changed. */ | |
2093 | ||
2094 | bool | |
2095 | early_remat::avail_confluence_n (edge e) | |
2096 | { | |
2097 | remat_block_info *src = &er->m_block_info[e->src->index]; | |
2098 | remat_block_info *dest = &er->m_block_info[e->dest->index]; | |
2099 | ||
2100 | if (!available_across_edge_p (e)) | |
2101 | return false; | |
2102 | ||
2103 | if (empty_p (dest->available_in)) | |
2104 | return false; | |
2105 | ||
2106 | if (!src->available_out) | |
2107 | { | |
2108 | bitmap_clear (dest->available_in); | |
2109 | return true; | |
2110 | } | |
2111 | ||
2112 | return bitmap_and_into (dest->available_in, src->available_out); | |
2113 | } | |
2114 | ||
2115 | /* Propagate information from the available_in set of block BB_INDEX | |
2116 | to available_out. Return true if something changed. */ | |
2117 | ||
2118 | bool | |
2119 | early_remat::avail_transfer (int bb_index) | |
2120 | { | |
2121 | remat_block_info *info = &er->m_block_info[bb_index]; | |
2122 | ||
2123 | if (info->available_out == info->available_locally) | |
2124 | return false; | |
2125 | ||
2126 | if (info->available_out == info->available_in) | |
2127 | /* Assume that we are only called if the input changed. */ | |
2128 | return true; | |
2129 | ||
2130 | return er->set_available_out (info); | |
2131 | } | |
2132 | ||
2133 | /* Compute global availability for the function, starting with the local | |
2134 | information computed by local_phase. */ | |
2135 | ||
2136 | void | |
2137 | early_remat::compute_availability (void) | |
2138 | { | |
2139 | /* We use df_simple_dataflow instead of the lcm routines for three reasons: | |
2140 | ||
2141 | (1) it avoids recomputing the traversal order; | |
2142 | (2) many of the sets are likely to be sparse, so we don't necessarily | |
2143 | want to use sbitmaps; and | |
2144 | (3) it means we can avoid creating an explicit kill set for the call. */ | |
2145 | er = this; | |
2146 | bitmap_clear (&m_tmp_bitmap); | |
2147 | bitmap_set_range (&m_tmp_bitmap, 0, last_basic_block_for_fn (m_fn)); | |
2148 | df_simple_dataflow (DF_FORWARD, NULL, NULL, | |
2149 | avail_confluence_n, avail_transfer, | |
2150 | &m_tmp_bitmap, df_get_postorder (DF_FORWARD), | |
2151 | df_get_n_blocks (DF_FORWARD)); | |
2152 | er = 0; | |
2153 | ||
2154 | /* Restrict the required_in sets to values that aren't available. */ | |
2155 | basic_block bb; | |
2156 | FOR_EACH_BB_FN (bb, m_fn) | |
2157 | { | |
2158 | remat_block_info *info = &m_block_info[bb->index]; | |
2159 | if (info->required_in && info->available_in) | |
2160 | bitmap_and_compl_into (info->required_in, info->available_in); | |
2161 | } | |
2162 | } | |
2163 | ||
2164 | /* Make sure that INFO's available_out and available_in sets are unique. */ | |
2165 | ||
2166 | inline void | |
2167 | early_remat::unshare_available_sets (remat_block_info *info) | |
2168 | { | |
2169 | if (info->available_in && info->available_in == info->available_out) | |
2170 | { | |
2171 | info->available_in = alloc_bitmap (); | |
2172 | bitmap_copy (info->available_in, info->available_out); | |
2173 | } | |
2174 | } | |
2175 | ||
2176 | /* Return true if it is possible to move rematerializations from the | |
2177 | destination of E to the source of E. */ | |
2178 | ||
2179 | inline bool | |
2180 | early_remat::can_move_across_edge_p (edge e) | |
2181 | { | |
2182 | return (available_across_edge_p (e) | |
2183 | && !m_block_info[e->src->index].abnormal_call_p); | |
2184 | } | |
2185 | ||
2186 | /* Return true if it is cheaper to rematerialize values at the head of | |
2187 | block QUERY_BB_INDEX instead of rematerializing in its predecessors. */ | |
2188 | ||
2189 | bool | |
2190 | early_remat::local_remat_cheaper_p (unsigned int query_bb_index) | |
2191 | { | |
2192 | if (m_block_info[query_bb_index].remat_frequency_valid_p) | |
2193 | return m_block_info[query_bb_index].local_remat_cheaper_p; | |
2194 | ||
2195 | /* Iteratively compute the cost of rematerializing values in the | |
2196 | predecessor blocks, then compare that with the cost of | |
2197 | rematerializing at the head of the block. | |
2198 | ||
2199 | A cycle indicates that there is no call on that execution path, | |
2200 | so it isn't necessary to rematerialize on that path. */ | |
2201 | auto_vec<basic_block, 16> stack; | |
2202 | stack.quick_push (BASIC_BLOCK_FOR_FN (m_fn, query_bb_index)); | |
2203 | while (!stack.is_empty ()) | |
2204 | { | |
2205 | basic_block bb = stack.last (); | |
2206 | remat_block_info *info = &m_block_info[bb->index]; | |
2207 | if (info->remat_frequency_valid_p) | |
2208 | { | |
2209 | stack.pop (); | |
2210 | continue; | |
2211 | } | |
2212 | ||
2213 | info->visited_p = true; | |
2214 | int frequency = 0; | |
2215 | bool can_move_p = true; | |
2216 | edge e; | |
2217 | edge_iterator ei; | |
2218 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2219 | if (!can_move_across_edge_p (e)) | |
2220 | { | |
2221 | can_move_p = false; | |
2222 | break; | |
2223 | } | |
2224 | else if (m_block_info[e->src->index].last_call) | |
2225 | /* We'll rematerialize after the call. */ | |
2226 | frequency += e->src->count.to_frequency (m_fn); | |
2227 | else if (m_block_info[e->src->index].remat_frequency_valid_p) | |
2228 | /* Add the cost of rematerializing at the head of E->src | |
2229 | or in its predecessors (whichever is cheaper). */ | |
2230 | frequency += m_block_info[e->src->index].remat_frequency; | |
2231 | else if (!m_block_info[e->src->index].visited_p) | |
2232 | /* Queue E->src and then revisit this block again. */ | |
2233 | stack.safe_push (e->src); | |
2234 | ||
2235 | /* Come back to this block later if we need to process some of | |
2236 | its predecessors. */ | |
2237 | if (stack.last () != bb) | |
2238 | continue; | |
2239 | ||
2240 | /* If rematerializing in and before the block have equal cost, prefer | |
2241 | rematerializing in the block. This should shorten the live range. */ | |
2242 | int bb_frequency = bb->count.to_frequency (m_fn); | |
2243 | if (!can_move_p || frequency >= bb_frequency) | |
2244 | { | |
2245 | info->local_remat_cheaper_p = true; | |
2246 | info->remat_frequency = bb_frequency; | |
2247 | } | |
2248 | else | |
2249 | info->remat_frequency = frequency; | |
2250 | info->remat_frequency_valid_p = true; | |
2251 | info->visited_p = false; | |
2252 | if (dump_file) | |
2253 | { | |
2254 | if (!can_move_p) | |
2255 | fprintf (dump_file, ";; Need to rematerialize at the head of" | |
2256 | " block %d; cannot move to predecessors.\n", bb->index); | |
2257 | else | |
2258 | { | |
2259 | fprintf (dump_file, ";; Block %d has frequency %d," | |
2260 | " rematerializing in predecessors has frequency %d", | |
2261 | bb->index, bb_frequency, frequency); | |
2262 | if (info->local_remat_cheaper_p) | |
2263 | fprintf (dump_file, "; prefer to rematerialize" | |
2264 | " in the block\n"); | |
2265 | else | |
2266 | fprintf (dump_file, "; prefer to rematerialize" | |
2267 | " in predecessors\n"); | |
2268 | } | |
2269 | } | |
2270 | stack.pop (); | |
2271 | } | |
2272 | return m_block_info[query_bb_index].local_remat_cheaper_p; | |
2273 | } | |
2274 | ||
2275 | /* Return true if we cannot rematerialize candidate CAND_INDEX at the head of | |
2276 | block BB_INDEX. */ | |
2277 | ||
2278 | bool | |
2279 | early_remat::need_to_move_candidate_p (unsigned int bb_index, | |
2280 | unsigned int cand_index) | |
2281 | { | |
2282 | remat_block_info *info = &m_block_info[bb_index]; | |
2283 | remat_candidate *cand = &m_candidates[cand_index]; | |
2284 | basic_block bb = BASIC_BLOCK_FOR_FN (m_fn, bb_index); | |
2285 | ||
2286 | /* If there is more than one reaching definition of REGNO, | |
2287 | we'll need to rematerialize in predecessors instead. */ | |
2288 | bitmap_and (&m_tmp_bitmap, info->rd_in, m_regno_to_candidates[cand->regno]); | |
2289 | if (!bitmap_single_bit_set_p (&m_tmp_bitmap)) | |
2290 | { | |
2291 | if (dump_file) | |
2292 | fprintf (dump_file, ";; Cannot rematerialize %d at the" | |
2293 | " head of block %d because there is more than one" | |
2294 | " reaching definition of reg %d\n", cand_index, | |
2295 | bb_index, cand->regno); | |
2296 | return true; | |
2297 | } | |
2298 | ||
2299 | /* Likewise if rematerializing CAND here would clobber a live register. */ | |
2300 | if (cand->clobbers | |
2301 | && bitmap_intersect_p (cand->clobbers, DF_LR_IN (bb))) | |
2302 | { | |
2303 | if (dump_file) | |
2304 | fprintf (dump_file, ";; Cannot rematerialize %d at the" | |
2305 | " head of block %d because it would clobber live" | |
2306 | " registers\n", cand_index, bb_index); | |
2307 | return true; | |
2308 | } | |
2309 | ||
2310 | return false; | |
2311 | } | |
2312 | ||
2313 | /* Set REQUIRED to the minimum set of candidates that must be rematerialized | |
2314 | in predecessors of block BB_INDEX instead of at the start of the block. */ | |
2315 | ||
2316 | void | |
2317 | early_remat::compute_minimum_move_set (unsigned int bb_index, | |
2318 | bitmap required) | |
2319 | { | |
2320 | remat_block_info *info = &m_block_info[bb_index]; | |
2321 | bitmap_head remaining; | |
2322 | ||
2323 | bitmap_clear (required); | |
2324 | bitmap_initialize (&remaining, &m_obstack); | |
2325 | bitmap_copy (&remaining, info->required_in); | |
2326 | while (!bitmap_empty_p (&remaining)) | |
2327 | { | |
2328 | unsigned int cand_index = bitmap_first_set_bit (&remaining); | |
2329 | remat_candidate *cand = &m_candidates[cand_index]; | |
2330 | bitmap_clear_bit (&remaining, cand_index); | |
2331 | ||
2332 | /* Leave invalid candidates where they are. */ | |
2333 | if (!cand->can_copy_p) | |
2334 | continue; | |
2335 | ||
2336 | /* Decide whether to move this candidate. */ | |
2337 | if (!bitmap_bit_p (required, cand_index)) | |
2338 | { | |
2339 | if (!need_to_move_candidate_p (bb_index, cand_index)) | |
2340 | continue; | |
2341 | bitmap_set_bit (required, cand_index); | |
2342 | } | |
2343 | ||
2344 | /* Also move values used by the candidate, so that we don't | |
2345 | rematerialize them twice. */ | |
2346 | if (cand->uses) | |
2347 | { | |
2348 | bitmap_ior_and_into (required, cand->uses, info->required_in); | |
2349 | bitmap_ior_and_into (&remaining, cand->uses, info->required_in); | |
2350 | } | |
2351 | } | |
2352 | } | |
2353 | ||
2354 | /* Make the predecessors of BB_INDEX rematerialize the candidates in | |
2355 | REQUIRED. Add any blocks whose required_in set changes to | |
2356 | PENDING_BLOCKS. */ | |
2357 | ||
2358 | void | |
2359 | early_remat::move_to_predecessors (unsigned int bb_index, bitmap required, | |
2360 | bitmap pending_blocks) | |
2361 | { | |
2362 | if (empty_p (required)) | |
2363 | return; | |
2364 | remat_block_info *dest_info = &m_block_info[bb_index]; | |
2365 | basic_block bb = BASIC_BLOCK_FOR_FN (m_fn, bb_index); | |
2366 | edge e; | |
2367 | edge_iterator ei; | |
2368 | FOR_EACH_EDGE (e, ei, bb->preds) | |
2369 | { | |
2370 | remat_block_info *src_info = &m_block_info[e->src->index]; | |
2371 | ||
2372 | /* Restrict the set we add to the reaching definitions. */ | |
2373 | bitmap_and (&m_tmp_bitmap, required, src_info->rd_out); | |
2374 | if (bitmap_empty_p (&m_tmp_bitmap)) | |
2375 | continue; | |
2376 | ||
2377 | if (!can_move_across_edge_p (e)) | |
2378 | { | |
2379 | /* We can't move the rematerialization and we can't do it at | |
2380 | the start of the block either. In this case we just give up | |
2381 | and rely on spilling to make the values available across E. */ | |
2382 | if (dump_file) | |
2383 | { | |
2384 | fprintf (dump_file, ";; Cannot rematerialize the following" | |
2385 | " candidates in block %d:", e->src->index); | |
2386 | dump_candidate_bitmap (required); | |
2387 | fprintf (dump_file, "\n"); | |
2388 | } | |
2389 | continue; | |
2390 | } | |
2391 | ||
2392 | /* Remove candidates that are already available. */ | |
2393 | if (src_info->available_out) | |
2394 | { | |
2395 | bitmap_and_compl_into (&m_tmp_bitmap, src_info->available_out); | |
2396 | if (bitmap_empty_p (&m_tmp_bitmap)) | |
2397 | continue; | |
2398 | } | |
2399 | ||
2400 | /* Add the remaining candidates to the appropriate required set. */ | |
2401 | if (dump_file) | |
2402 | { | |
2403 | fprintf (dump_file, ";; Moving this set from block %d" | |
2404 | " to block %d:", bb_index, e->src->index); | |
2405 | dump_candidate_bitmap (&m_tmp_bitmap); | |
2406 | fprintf (dump_file, "\n"); | |
2407 | } | |
2408 | /* If the source block contains a call, we want to rematerialize | |
2409 | after the call, otherwise we want to rematerialize at the start | |
2410 | of the block. */ | |
2411 | bitmap src_required = get_bitmap (src_info->last_call | |
2412 | ? &src_info->required_after_call | |
2413 | : &src_info->required_in); | |
2414 | if (bitmap_ior_into (src_required, &m_tmp_bitmap)) | |
2415 | { | |
2416 | if (!src_info->last_call) | |
2417 | bitmap_set_bit (pending_blocks, e->src->index); | |
2418 | unshare_available_sets (src_info); | |
2419 | bitmap_ior_into (get_bitmap (&src_info->available_out), | |
2420 | &m_tmp_bitmap); | |
2421 | } | |
2422 | } | |
2423 | ||
2424 | /* The candidates are now available on entry to the block. */ | |
2425 | bitmap_and_compl_into (dest_info->required_in, required); | |
2426 | unshare_available_sets (dest_info); | |
2427 | bitmap_ior_into (get_bitmap (&dest_info->available_in), required); | |
2428 | } | |
2429 | ||
2430 | /* Go through the candidates that are currently marked as being | |
2431 | rematerialized at the beginning of a block. Decide in each case | |
2432 | whether that's valid and profitable; if it isn't, move the | |
2433 | rematerialization to predecessor blocks instead. */ | |
2434 | ||
2435 | void | |
2436 | early_remat::choose_rematerialization_points (void) | |
2437 | { | |
2438 | bitmap_head required; | |
2439 | bitmap_head pending_blocks; | |
2440 | ||
2441 | int *postorder = df_get_postorder (DF_BACKWARD); | |
2442 | unsigned int postorder_len = df_get_n_blocks (DF_BACKWARD); | |
2443 | bitmap_initialize (&required, &m_obstack); | |
2444 | bitmap_initialize (&pending_blocks, &m_obstack); | |
2445 | do | |
2446 | /* Process the blocks in postorder, to reduce the number of iterations | |
2447 | of the outer loop. */ | |
2448 | for (unsigned int i = 0; i < postorder_len; ++i) | |
2449 | { | |
2450 | unsigned int bb_index = postorder[i]; | |
2451 | remat_block_info *info = &m_block_info[bb_index]; | |
2452 | bitmap_clear_bit (&pending_blocks, bb_index); | |
2453 | ||
2454 | if (empty_p (info->required_in)) | |
2455 | continue; | |
2456 | ||
2457 | if (info->available_in) | |
2458 | gcc_checking_assert (!bitmap_intersect_p (info->required_in, | |
2459 | info->available_in)); | |
2460 | ||
2461 | if (local_remat_cheaper_p (bb_index)) | |
2462 | { | |
2463 | /* We'd prefer to rematerialize at the head of the block. | |
2464 | Only move candidates if we need to. */ | |
2465 | compute_minimum_move_set (bb_index, &required); | |
2466 | move_to_predecessors (bb_index, &required, &pending_blocks); | |
2467 | } | |
2468 | else | |
2469 | move_to_predecessors (bb_index, info->required_in, | |
2470 | &pending_blocks); | |
2471 | } | |
2472 | while (!bitmap_empty_p (&pending_blocks)); | |
2473 | bitmap_clear (&required); | |
2474 | } | |
2475 | ||
2476 | /* Emit all rematerialization instructions queued for BB. */ | |
2477 | ||
2478 | void | |
2479 | early_remat::emit_remat_insns_for_block (basic_block bb) | |
2480 | { | |
2481 | remat_block_info *info = &m_block_info[bb->index]; | |
2482 | ||
2483 | if (info->last_call && !empty_p (info->required_after_call)) | |
2484 | emit_remat_insns (info->required_after_call, NULL, | |
2485 | info->rd_after_call, info->last_call); | |
2486 | ||
2487 | if (!empty_p (info->required_in)) | |
2488 | { | |
2489 | rtx_insn *insn = BB_HEAD (bb); | |
2490 | while (insn != BB_END (bb) | |
2491 | && !INSN_P (NEXT_INSN (insn))) | |
2492 | insn = NEXT_INSN (insn); | |
2493 | emit_remat_insns (info->required_in, info->available_in, | |
2494 | info->rd_in, insn); | |
2495 | } | |
2496 | } | |
2497 | ||
2498 | /* Decide which candidates in each block's REQUIRED_IN set need to be | |
2499 | rematerialized and decide where the rematerialization instructions | |
2500 | should go. Emit queued rematerialization instructions at the start | |
2501 | of blocks and after the last calls in blocks. */ | |
2502 | ||
2503 | void | |
2504 | early_remat::global_phase (void) | |
2505 | { | |
2506 | compute_availability (); | |
2507 | if (dump_file) | |
2508 | { | |
2509 | fprintf (dump_file, "\n;; Blocks after computing global" | |
2510 | " availability:\n"); | |
2511 | dump_all_blocks (); | |
2512 | } | |
2513 | ||
2514 | choose_rematerialization_points (); | |
2515 | if (dump_file) | |
2516 | { | |
2517 | fprintf (dump_file, "\n;; Blocks after choosing rematerialization" | |
2518 | " points:\n"); | |
2519 | dump_all_blocks (); | |
2520 | } | |
2521 | ||
2522 | basic_block bb; | |
2523 | FOR_EACH_BB_FN (bb, m_fn) | |
2524 | emit_remat_insns_for_block (bb); | |
2525 | } | |
2526 | ||
2527 | /* Main function for the pass. */ | |
2528 | ||
2529 | void | |
2530 | early_remat::run (void) | |
2531 | { | |
2532 | df_analyze (); | |
2533 | ||
2534 | if (!collect_candidates ()) | |
2535 | return; | |
2536 | ||
2537 | init_block_info (); | |
2538 | sort_candidates (); | |
2539 | finalize_candidate_indices (); | |
2540 | if (dump_file) | |
2541 | dump_all_candidates (); | |
2542 | ||
2543 | compute_rd (); | |
2544 | decide_candidate_validity (); | |
2545 | local_phase (); | |
2546 | global_phase (); | |
2547 | } | |
2548 | ||
2549 | early_remat::early_remat (function *fn, sbitmap selected_modes) | |
2550 | : m_fn (fn), | |
2551 | m_selected_modes (selected_modes), | |
2552 | m_available (0), | |
2553 | m_required (0), | |
2554 | m_value_table (63) | |
2555 | { | |
2556 | bitmap_obstack_initialize (&m_obstack); | |
2557 | bitmap_initialize (&m_candidate_regnos, &m_obstack); | |
2558 | bitmap_initialize (&m_tmp_bitmap, &m_obstack); | |
2559 | } | |
2560 | ||
2561 | early_remat::~early_remat () | |
2562 | { | |
2563 | bitmap_obstack_release (&m_obstack); | |
2564 | } | |
2565 | ||
2566 | namespace { | |
2567 | ||
2568 | const pass_data pass_data_early_remat = | |
2569 | { | |
2570 | RTL_PASS, /* type */ | |
2571 | "early_remat", /* name */ | |
2572 | OPTGROUP_NONE, /* optinfo_flags */ | |
2573 | TV_EARLY_REMAT, /* tv_id */ | |
2574 | 0, /* properties_required */ | |
2575 | 0, /* properties_provided */ | |
2576 | 0, /* properties_destroyed */ | |
2577 | 0, /* todo_flags_start */ | |
2578 | TODO_df_finish, /* todo_flags_finish */ | |
2579 | }; | |
2580 | ||
2581 | class pass_early_remat : public rtl_opt_pass | |
2582 | { | |
2583 | public: | |
2584 | pass_early_remat (gcc::context *ctxt) | |
2585 | : rtl_opt_pass (pass_data_early_remat, ctxt) | |
2586 | {} | |
2587 | ||
2588 | /* opt_pass methods: */ | |
2589 | virtual bool gate (function *) | |
2590 | { | |
2591 | return optimize > 1 && NUM_POLY_INT_COEFFS > 1; | |
2592 | } | |
2593 | ||
2594 | virtual unsigned int execute (function *f) | |
2595 | { | |
2596 | auto_sbitmap selected_modes (NUM_MACHINE_MODES); | |
2597 | bitmap_clear (selected_modes); | |
2598 | targetm.select_early_remat_modes (selected_modes); | |
2599 | if (!bitmap_empty_p (selected_modes)) | |
2600 | early_remat (f, selected_modes).run (); | |
2601 | return 0; | |
2602 | } | |
2603 | }; // class pass_early_remat | |
2604 | ||
2605 | } // anon namespace | |
2606 | ||
2607 | rtl_opt_pass * | |
2608 | make_pass_early_remat (gcc::context *ctxt) | |
2609 | { | |
2610 | return new pass_early_remat (ctxt); | |
2611 | } |