1 | /* RTL dead store elimination. |
---|---|
2 | Copyright (C) 2005-2025 Free Software Foundation, Inc. |
3 | |
4 | Contributed by Richard Sandiford <rsandifor@codesourcery.com> |
5 | and Kenneth Zadeck <zadeck@naturalbridge.com> |
6 | |
7 | This file is part of GCC. |
8 | |
9 | GCC is free software; you can redistribute it and/or modify it under |
10 | the terms of the GNU General Public License as published by the Free |
11 | Software Foundation; either version 3, or (at your option) any later |
12 | version. |
13 | |
14 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
15 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
16 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
17 | for more details. |
18 | |
19 | You should have received a copy of the GNU General Public License |
20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ |
22 | |
23 | #undef BASELINE |
24 | |
25 | #include "config.h" |
26 | #include "system.h" |
27 | #include "coretypes.h" |
28 | #include "backend.h" |
29 | #include "target.h" |
30 | #include "rtl.h" |
31 | #include "tree.h" |
32 | #include "gimple.h" |
33 | #include "predict.h" |
34 | #include "df.h" |
35 | #include "memmodel.h" |
36 | #include "tm_p.h" |
37 | #include "gimple-ssa.h" |
38 | #include "expmed.h" |
39 | #include "optabs.h" |
40 | #include "emit-rtl.h" |
41 | #include "recog.h" |
42 | #include "alias.h" |
43 | #include "stor-layout.h" |
44 | #include "cfgrtl.h" |
45 | #include "cselib.h" |
46 | #include "tree-pass.h" |
47 | #include "explow.h" |
48 | #include "expr.h" |
49 | #include "dbgcnt.h" |
50 | #include "rtl-iter.h" |
51 | #include "cfgcleanup.h" |
52 | #include "calls.h" |
53 | |
54 | /* This file contains three techniques for performing Dead Store |
55 | Elimination (dse). |
56 | |
57 | * The first technique performs dse locally on any base address. It |
58 | is based on the cselib which is a local value numbering technique. |
59 | This technique is local to a basic block but deals with a fairly |
60 | general addresses. |
61 | |
62 | * The second technique performs dse globally but is restricted to |
63 | base addresses that are either constant or are relative to the |
64 | frame_pointer. |
65 | |
66 | * The third technique, (which is only done after register allocation) |
67 | processes the spill slots. This differs from the second |
68 | technique because it takes advantage of the fact that spilling is |
69 | completely free from the effects of aliasing. |
70 | |
71 | Logically, dse is a backwards dataflow problem. A store can be |
72 | deleted if it if cannot be reached in the backward direction by any |
73 | use of the value being stored. However, the local technique uses a |
74 | forwards scan of the basic block because cselib requires that the |
75 | block be processed in that order. |
76 | |
77 | The pass is logically broken into 7 steps: |
78 | |
79 | 0) Initialization. |
80 | |
81 | 1) The local algorithm, as well as scanning the insns for the two |
82 | global algorithms. |
83 | |
84 | 2) Analysis to see if the global algs are necessary. In the case |
85 | of stores base on a constant address, there must be at least two |
86 | stores to that address, to make it possible to delete some of the |
87 | stores. In the case of stores off of the frame or spill related |
88 | stores, only one store to an address is necessary because those |
89 | stores die at the end of the function. |
90 | |
91 | 3) Set up the global dataflow equations based on processing the |
92 | info parsed in the first step. |
93 | |
94 | 4) Solve the dataflow equations. |
95 | |
96 | 5) Delete the insns that the global analysis has indicated are |
97 | unnecessary. |
98 | |
99 | 6) Delete insns that store the same value as preceding store |
100 | where the earlier store couldn't be eliminated. |
101 | |
102 | 7) Cleanup. |
103 | |
104 | This step uses cselib and canon_rtx to build the largest expression |
105 | possible for each address. This pass is a forwards pass through |
106 | each basic block. From the point of view of the global technique, |
107 | the first pass could examine a block in either direction. The |
108 | forwards ordering is to accommodate cselib. |
109 | |
110 | We make a simplifying assumption: addresses fall into four broad |
111 | categories: |
112 | |
113 | 1) base has rtx_varies_p == false, offset is constant. |
114 | 2) base has rtx_varies_p == false, offset variable. |
115 | 3) base has rtx_varies_p == true, offset constant. |
116 | 4) base has rtx_varies_p == true, offset variable. |
117 | |
118 | The local passes are able to process all 4 kinds of addresses. The |
119 | global pass only handles 1). |
120 | |
121 | The global problem is formulated as follows: |
122 | |
123 | A store, S1, to address A, where A is not relative to the stack |
124 | frame, can be eliminated if all paths from S1 to the end of the |
125 | function contain another store to A before a read to A. |
126 | |
127 | If the address A is relative to the stack frame, a store S2 to A |
128 | can be eliminated if there are no paths from S2 that reach the |
129 | end of the function that read A before another store to A. In |
130 | this case S2 can be deleted if there are paths from S2 to the |
131 | end of the function that have no reads or writes to A. This |
132 | second case allows stores to the stack frame to be deleted that |
133 | would otherwise die when the function returns. This cannot be |
134 | done if stores_off_frame_dead_at_return is not true. See the doc |
135 | for that variable for when this variable is false. |
136 | |
137 | The global problem is formulated as a backwards set union |
138 | dataflow problem where the stores are the gens and reads are the |
139 | kills. Set union problems are rare and require some special |
140 | handling given our representation of bitmaps. A straightforward |
141 | implementation requires a lot of bitmaps filled with 1s. |
142 | These are expensive and cumbersome in our bitmap formulation so |
143 | care has been taken to avoid large vectors filled with 1s. See |
144 | the comments in bb_info and in the dataflow confluence functions |
145 | for details. |
146 | |
147 | There are two places for further enhancements to this algorithm: |
148 | |
149 | 1) The original dse which was embedded in a pass called flow also |
150 | did local address forwarding. For example in |
151 | |
152 | A <- r100 |
153 | ... <- A |
154 | |
155 | flow would replace the right hand side of the second insn with a |
156 | reference to r100. Most of the information is available to add this |
157 | to this pass. It has not done it because it is a lot of work in |
158 | the case that either r100 is assigned to between the first and |
159 | second insn and/or the second insn is a load of part of the value |
160 | stored by the first insn. |
161 | |
162 | insn 5 in gcc.c-torture/compile/990203-1.c simple case. |
163 | insn 15 in gcc.c-torture/execute/20001017-2.c simple case. |
164 | insn 25 in gcc.c-torture/execute/20001026-1.c simple case. |
165 | insn 44 in gcc.c-torture/execute/20010910-1.c simple case. |
166 | |
167 | 2) The cleaning up of spill code is quite profitable. It currently |
168 | depends on reading tea leaves and chicken entrails left by reload. |
169 | This pass depends on reload creating a singleton alias set for each |
170 | spill slot and telling the next dse pass which of these alias sets |
171 | are the singletons. Rather than analyze the addresses of the |
172 | spills, dse's spill processing just does analysis of the loads and |
173 | stores that use those alias sets. There are three cases where this |
174 | falls short: |
175 | |
176 | a) Reload sometimes creates the slot for one mode of access, and |
177 | then inserts loads and/or stores for a smaller mode. In this |
178 | case, the current code just punts on the slot. The proper thing |
179 | to do is to back out and use one bit vector position for each |
180 | byte of the entity associated with the slot. This depends on |
181 | KNOWING that reload always generates the accesses for each of the |
182 | bytes in some canonical (read that easy to understand several |
183 | passes after reload happens) way. |
184 | |
185 | b) Reload sometimes decides that spill slot it allocated was not |
186 | large enough for the mode and goes back and allocates more slots |
187 | with the same mode and alias set. The backout in this case is a |
188 | little more graceful than (a). In this case the slot is unmarked |
189 | as being a spill slot and if final address comes out to be based |
190 | off the frame pointer, the global algorithm handles this slot. |
191 | |
192 | c) For any pass that may prespill, there is currently no |
193 | mechanism to tell the dse pass that the slot being used has the |
194 | special properties that reload uses. It may be that all that is |
195 | required is to have those passes make the same calls that reload |
196 | does, assuming that the alias sets can be manipulated in the same |
197 | way. */ |
198 | |
199 | /* There are limits to the size of constant offsets we model for the |
200 | global problem. There are certainly test cases, that exceed this |
201 | limit, however, it is unlikely that there are important programs |
202 | that really have constant offsets this size. */ |
203 | #define MAX_OFFSET (64 * 1024) |
204 | |
205 | /* Obstack for the DSE dataflow bitmaps. We don't want to put these |
206 | on the default obstack because these bitmaps can grow quite large |
207 | (~2GB for the small (!) test case of PR54146) and we'll hold on to |
208 | all that memory until the end of the compiler run. |
209 | As a bonus, delete_tree_live_info can destroy all the bitmaps by just |
210 | releasing the whole obstack. */ |
211 | static bitmap_obstack dse_bitmap_obstack; |
212 | |
213 | /* Obstack for other data. As for above: Kinda nice to be able to |
214 | throw it all away at the end in one big sweep. */ |
215 | static struct obstack dse_obstack; |
216 | |
217 | /* Scratch bitmap for cselib's cselib_expand_value_rtx. */ |
218 | static bitmap scratch = NULL; |
219 | |
220 | struct insn_info_type; |
221 | |
222 | /* This structure holds information about a candidate store. */ |
223 | class store_info |
224 | { |
225 | public: |
226 | |
227 | /* False means this is a clobber. */ |
228 | bool is_set; |
229 | |
230 | /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */ |
231 | bool is_large; |
232 | |
233 | /* The id of the mem group of the base address. If rtx_varies_p is |
234 | true, this is -1. Otherwise, it is the index into the group |
235 | table. */ |
236 | int group_id; |
237 | |
238 | /* This is the cselib value. */ |
239 | cselib_val *cse_base; |
240 | |
241 | /* This canonized mem. */ |
242 | rtx mem; |
243 | |
244 | /* Canonized MEM address for use by canon_true_dependence. */ |
245 | rtx mem_addr; |
246 | |
247 | /* The offset of the first byte associated with the operation. */ |
248 | poly_int64 offset; |
249 | |
250 | /* The number of bytes covered by the operation. This is always exact |
251 | and known (rather than -1). */ |
252 | poly_int64 width; |
253 | |
254 | /* The address space that the memory reference uses. */ |
255 | unsigned char addrspace; |
256 | |
257 | union |
258 | { |
259 | /* A bitmask as wide as the number of bytes in the word that |
260 | contains a 1 if the byte may be needed. The store is unused if |
261 | all of the bits are 0. This is used if IS_LARGE is false. */ |
262 | unsigned HOST_WIDE_INT small_bitmask; |
263 | |
264 | struct |
265 | { |
266 | /* A bitmap with one bit per byte, or null if the number of |
267 | bytes isn't known at compile time. A cleared bit means |
268 | the position is needed. Used if IS_LARGE is true. */ |
269 | bitmap bmap; |
270 | |
271 | /* When BITMAP is nonnull, this counts the number of set bits |
272 | (i.e. unneeded bytes) in the bitmap. If it is equal to |
273 | WIDTH, the whole store is unused. |
274 | |
275 | When BITMAP is null: |
276 | - the store is definitely not needed when COUNT == 1 |
277 | - all the store is needed when COUNT == 0 and RHS is nonnull |
278 | - otherwise we don't know which parts of the store are needed. */ |
279 | int count; |
280 | } large; |
281 | } positions_needed; |
282 | |
283 | /* The next store info for this insn. */ |
284 | class store_info *next; |
285 | |
286 | /* The right hand side of the store. This is used if there is a |
287 | subsequent reload of the mems address somewhere later in the |
288 | basic block. */ |
289 | rtx rhs; |
290 | |
291 | /* If rhs is or holds a constant, this contains that constant, |
292 | otherwise NULL. */ |
293 | rtx const_rhs; |
294 | |
295 | /* Set if this store stores the same constant value as REDUNDANT_REASON |
296 | insn stored. These aren't eliminated early, because doing that |
297 | might prevent the earlier larger store to be eliminated. */ |
298 | struct insn_info_type *redundant_reason; |
299 | }; |
300 | |
301 | /* Return a bitmask with the first N low bits set. */ |
302 | |
303 | static unsigned HOST_WIDE_INT |
304 | lowpart_bitmask (int n) |
305 | { |
306 | unsigned HOST_WIDE_INT mask = HOST_WIDE_INT_M1U; |
307 | return mask >> (HOST_BITS_PER_WIDE_INT - n); |
308 | } |
309 | |
310 | static object_allocator<store_info> cse_store_info_pool ("cse_store_info_pool"); |
311 | |
312 | static object_allocator<store_info> rtx_store_info_pool ("rtx_store_info_pool"); |
313 | |
314 | /* This structure holds information about a load. These are only |
315 | built for rtx bases. */ |
316 | class read_info_type |
317 | { |
318 | public: |
319 | /* The id of the mem group of the base address. */ |
320 | int group_id; |
321 | |
322 | /* The offset of the first byte associated with the operation. */ |
323 | poly_int64 offset; |
324 | |
325 | /* The number of bytes covered by the operation, or -1 if not known. */ |
326 | poly_int64 width; |
327 | |
328 | /* The mem being read. */ |
329 | rtx mem; |
330 | |
331 | /* The next read_info for this insn. */ |
332 | class read_info_type *next; |
333 | }; |
334 | typedef class read_info_type *read_info_t; |
335 | |
336 | static object_allocator<read_info_type> read_info_type_pool ("read_info_pool"); |
337 | |
338 | /* One of these records is created for each insn. */ |
339 | |
340 | struct insn_info_type |
341 | { |
342 | /* Set true if the insn contains a store but the insn itself cannot |
343 | be deleted. This is set if the insn is a parallel and there is |
344 | more than one non dead output or if the insn is in some way |
345 | volatile. */ |
346 | bool cannot_delete; |
347 | |
348 | /* This field is only used by the global algorithm. It is set true |
349 | if the insn contains any read of mem except for a (1). This is |
350 | also set if the insn is a call or has a clobber mem. If the insn |
351 | contains a wild read, the use_rec will be null. */ |
352 | bool wild_read; |
353 | |
354 | /* This is true only for CALL instructions which could potentially read |
355 | any non-frame memory location. This field is used by the global |
356 | algorithm. */ |
357 | bool non_frame_wild_read; |
358 | |
359 | /* This field is only used for the processing of const functions. |
360 | These functions cannot read memory, but they can read the stack |
361 | because that is where they may get their parms. We need to be |
362 | this conservative because, like the store motion pass, we don't |
363 | consider CALL_INSN_FUNCTION_USAGE when processing call insns. |
364 | Moreover, we need to distinguish two cases: |
365 | 1. Before reload (register elimination), the stores related to |
366 | outgoing arguments are stack pointer based and thus deemed |
367 | of non-constant base in this pass. This requires special |
368 | handling but also means that the frame pointer based stores |
369 | need not be killed upon encountering a const function call. |
370 | 2. After reload, the stores related to outgoing arguments can be |
371 | either stack pointer or hard frame pointer based. This means |
372 | that we have no other choice than also killing all the frame |
373 | pointer based stores upon encountering a const function call. |
374 | This field is set after reload for const function calls and before |
375 | reload for const tail function calls on targets where arg pointer |
376 | is the frame pointer. Having this set is less severe than a wild |
377 | read, it just means that all the frame related stores are killed |
378 | rather than all the stores. */ |
379 | bool frame_read; |
380 | |
381 | /* This field is only used for the processing of const functions. |
382 | It is set if the insn may contain a stack pointer based store. */ |
383 | bool stack_pointer_based; |
384 | |
385 | /* This is true if any of the sets within the store contains a |
386 | cselib base. Such stores can only be deleted by the local |
387 | algorithm. */ |
388 | bool contains_cselib_groups; |
389 | |
390 | /* The insn. */ |
391 | rtx_insn *insn; |
392 | |
393 | /* The list of mem sets or mem clobbers that are contained in this |
394 | insn. If the insn is deletable, it contains only one mem set. |
395 | But it could also contain clobbers. Insns that contain more than |
396 | one mem set are not deletable, but each of those mems are here in |
397 | order to provide info to delete other insns. */ |
398 | store_info *store_rec; |
399 | |
400 | /* The linked list of mem uses in this insn. Only the reads from |
401 | rtx bases are listed here. The reads to cselib bases are |
402 | completely processed during the first scan and so are never |
403 | created. */ |
404 | read_info_t read_rec; |
405 | |
406 | /* The live fixed registers. We assume only fixed registers can |
407 | cause trouble by being clobbered from an expanded pattern; |
408 | storing only the live fixed registers (rather than all registers) |
409 | means less memory needs to be allocated / copied for the individual |
410 | stores. */ |
411 | regset fixed_regs_live; |
412 | |
413 | /* The prev insn in the basic block. */ |
414 | struct insn_info_type * prev_insn; |
415 | |
416 | /* The linked list of insns that are in consideration for removal in |
417 | the forwards pass through the basic block. This pointer may be |
418 | trash as it is not cleared when a wild read occurs. The only |
419 | time it is guaranteed to be correct is when the traversal starts |
420 | at active_local_stores. */ |
421 | struct insn_info_type * next_local_store; |
422 | }; |
423 | typedef struct insn_info_type *insn_info_t; |
424 | |
425 | static object_allocator<insn_info_type> insn_info_type_pool ("insn_info_pool"); |
426 | |
427 | /* The linked list of stores that are under consideration in this |
428 | basic block. */ |
429 | static insn_info_t active_local_stores; |
430 | static int active_local_stores_len; |
431 | |
432 | struct dse_bb_info_type |
433 | { |
434 | /* Pointer to the insn info for the last insn in the block. These |
435 | are linked so this is how all of the insns are reached. During |
436 | scanning this is the current insn being scanned. */ |
437 | insn_info_t last_insn; |
438 | |
439 | /* The info for the global dataflow problem. */ |
440 | |
441 | |
442 | /* This is set if the transfer function should and in the wild_read |
443 | bitmap before applying the kill and gen sets. That vector knocks |
444 | out most of the bits in the bitmap and thus speeds up the |
445 | operations. */ |
446 | bool apply_wild_read; |
447 | |
448 | /* The following 4 bitvectors hold information about which positions |
449 | of which stores are live or dead. They are indexed by |
450 | get_bitmap_index. */ |
451 | |
452 | /* The set of store positions that exist in this block before a wild read. */ |
453 | bitmap gen; |
454 | |
455 | /* The set of load positions that exist in this block above the |
456 | same position of a store. */ |
457 | bitmap kill; |
458 | |
459 | /* The set of stores that reach the top of the block without being |
460 | killed by a read. |
461 | |
462 | Do not represent the in if it is all ones. Note that this is |
463 | what the bitvector should logically be initialized to for a set |
464 | intersection problem. However, like the kill set, this is too |
465 | expensive. So initially, the in set will only be created for the |
466 | exit block and any block that contains a wild read. */ |
467 | bitmap in; |
468 | |
469 | /* The set of stores that reach the bottom of the block from it's |
470 | successors. |
471 | |
472 | Do not represent the in if it is all ones. Note that this is |
473 | what the bitvector should logically be initialized to for a set |
474 | intersection problem. However, like the kill and in set, this is |
475 | too expensive. So what is done is that the confluence operator |
476 | just initializes the vector from one of the out sets of the |
477 | successors of the block. */ |
478 | bitmap out; |
479 | |
480 | /* The following bitvector is indexed by the reg number. It |
481 | contains the set of regs that are live at the current instruction |
482 | being processed. While it contains info for all of the |
483 | registers, only the hard registers are actually examined. It is used |
484 | to assure that shift and/or add sequences that are inserted do not |
485 | accidentally clobber live hard regs. */ |
486 | bitmap regs_live; |
487 | }; |
488 | |
489 | typedef struct dse_bb_info_type *bb_info_t; |
490 | |
491 | static object_allocator<dse_bb_info_type> dse_bb_info_type_pool |
492 | ("bb_info_pool"); |
493 | |
494 | /* Table to hold all bb_infos. */ |
495 | static bb_info_t *bb_table; |
496 | |
497 | /* There is a group_info for each rtx base that is used to reference |
498 | memory. There are also not many of the rtx bases because they are |
499 | very limited in scope. */ |
500 | |
501 | struct group_info |
502 | { |
503 | /* The actual base of the address. */ |
504 | rtx rtx_base; |
505 | |
506 | /* The sequential id of the base. This allows us to have a |
507 | canonical ordering of these that is not based on addresses. */ |
508 | int id; |
509 | |
510 | /* True if there are any positions that are to be processed |
511 | globally. */ |
512 | bool process_globally; |
513 | |
514 | /* True if the base of this group is either the frame_pointer or |
515 | hard_frame_pointer. */ |
516 | bool frame_related; |
517 | |
518 | /* A mem wrapped around the base pointer for the group in order to do |
519 | read dependency. It must be given BLKmode in order to encompass all |
520 | the possible offsets from the base. */ |
521 | rtx base_mem; |
522 | |
523 | /* Canonized version of base_mem's address. */ |
524 | rtx canon_base_addr; |
525 | |
526 | /* These two sets of two bitmaps are used to keep track of how many |
527 | stores are actually referencing that position from this base. We |
528 | only do this for rtx bases as this will be used to assign |
529 | positions in the bitmaps for the global problem. Bit N is set in |
530 | store1 on the first store for offset N. Bit N is set in store2 |
531 | for the second store to offset N. This is all we need since we |
532 | only care about offsets that have two or more stores for them. |
533 | |
534 | The "_n" suffix is for offsets less than 0 and the "_p" suffix is |
535 | for 0 and greater offsets. |
536 | |
537 | There is one special case here, for stores into the stack frame, |
538 | we will or store1 into store2 before deciding which stores look |
539 | at globally. This is because stores to the stack frame that have |
540 | no other reads before the end of the function can also be |
541 | deleted. */ |
542 | bitmap store1_n, store1_p, store2_n, store2_p; |
543 | |
544 | /* These bitmaps keep track of offsets in this group escape this function. |
545 | An offset escapes if it corresponds to a named variable whose |
546 | addressable flag is set. */ |
547 | bitmap escaped_n, escaped_p; |
548 | |
549 | /* The positions in this bitmap have the same assignments as the in, |
550 | out, gen and kill bitmaps. This bitmap is all zeros except for |
551 | the positions that are occupied by stores for this group. */ |
552 | bitmap group_kill; |
553 | |
554 | /* The offset_map is used to map the offsets from this base into |
555 | positions in the global bitmaps. It is only created after all of |
556 | the all of stores have been scanned and we know which ones we |
557 | care about. */ |
558 | int *offset_map_n, *offset_map_p; |
559 | int offset_map_size_n, offset_map_size_p; |
560 | }; |
561 | |
562 | static object_allocator<group_info> group_info_pool ("rtx_group_info_pool"); |
563 | |
564 | /* Index into the rtx_group_vec. */ |
565 | static int rtx_group_next_id; |
566 | |
567 | |
568 | static vec<group_info *> rtx_group_vec; |
569 | |
570 | |
571 | /* This structure holds the set of changes that are being deferred |
572 | when removing read operation. See replace_read. */ |
573 | struct deferred_change |
574 | { |
575 | |
576 | /* The mem that is being replaced. */ |
577 | rtx *loc; |
578 | |
579 | /* The reg it is being replaced with. */ |
580 | rtx reg; |
581 | |
582 | struct deferred_change *next; |
583 | }; |
584 | |
585 | static object_allocator<deferred_change> deferred_change_pool |
586 | ("deferred_change_pool"); |
587 | |
588 | static deferred_change *deferred_change_list = NULL; |
589 | |
590 | /* This is true except if cfun->stdarg -- i.e. we cannot do |
591 | this for vararg functions because they play games with the frame. */ |
592 | static bool stores_off_frame_dead_at_return; |
593 | |
594 | /* Counter for stats. */ |
595 | static int globally_deleted; |
596 | static int locally_deleted; |
597 | |
598 | static bitmap all_blocks; |
599 | |
600 | /* Locations that are killed by calls in the global phase. */ |
601 | static bitmap kill_on_calls; |
602 | |
603 | /* The number of bits used in the global bitmaps. */ |
604 | static unsigned int current_position; |
605 | |
606 | /* Print offset range [OFFSET, OFFSET + WIDTH) to FILE. */ |
607 | |
608 | static void |
609 | print_range (FILE *file, poly_int64 offset, poly_int64 width) |
610 | { |
611 | fprintf (stream: file, format: "["); |
612 | print_dec (value: offset, file, sgn: SIGNED); |
613 | fprintf (stream: file, format: ".."); |
614 | print_dec (value: offset + width, file, sgn: SIGNED); |
615 | fprintf (stream: file, format: ")"); |
616 | } |
617 | |
618 | /*---------------------------------------------------------------------------- |
619 | Zeroth step. |
620 | |
621 | Initialization. |
622 | ----------------------------------------------------------------------------*/ |
623 | |
624 | |
625 | /* Hashtable callbacks for maintaining the "bases" field of |
626 | store_group_info, given that the addresses are function invariants. */ |
627 | |
628 | struct invariant_group_base_hasher : nofree_ptr_hash <group_info> |
629 | { |
630 | static inline hashval_t hash (const group_info *); |
631 | static inline bool equal (const group_info *, const group_info *); |
632 | }; |
633 | |
634 | inline bool |
635 | invariant_group_base_hasher::equal (const group_info *gi1, |
636 | const group_info *gi2) |
637 | { |
638 | return rtx_equal_p (gi1->rtx_base, gi2->rtx_base); |
639 | } |
640 | |
641 | inline hashval_t |
642 | invariant_group_base_hasher::hash (const group_info *gi) |
643 | { |
644 | int do_not_record; |
645 | return hash_rtx (gi->rtx_base, Pmode, &do_not_record, NULL, false); |
646 | } |
647 | |
648 | /* Tables of group_info structures, hashed by base value. */ |
649 | static hash_table<invariant_group_base_hasher> *rtx_group_table; |
650 | |
651 | |
652 | /* Get the GROUP for BASE. Add a new group if it is not there. */ |
653 | |
654 | static group_info * |
655 | get_group_info (rtx base) |
656 | { |
657 | struct group_info tmp_gi; |
658 | group_info *gi; |
659 | group_info **slot; |
660 | |
661 | gcc_assert (base != NULL_RTX); |
662 | |
663 | /* Find the store_base_info structure for BASE, creating a new one |
664 | if necessary. */ |
665 | tmp_gi.rtx_base = base; |
666 | slot = rtx_group_table->find_slot (value: &tmp_gi, insert: INSERT); |
667 | gi = *slot; |
668 | |
669 | if (gi == NULL) |
670 | { |
671 | *slot = gi = group_info_pool.allocate (); |
672 | gi->rtx_base = base; |
673 | gi->id = rtx_group_next_id++; |
674 | gi->base_mem = gen_rtx_MEM (BLKmode, base); |
675 | gi->canon_base_addr = canon_rtx (base); |
676 | gi->store1_n = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
677 | gi->store1_p = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
678 | gi->store2_n = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
679 | gi->store2_p = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
680 | gi->escaped_p = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
681 | gi->escaped_n = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
682 | gi->group_kill = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
683 | gi->process_globally = false; |
684 | gi->frame_related = |
685 | (base == frame_pointer_rtx) || (base == hard_frame_pointer_rtx) |
686 | || (base == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]); |
687 | gi->offset_map_size_n = 0; |
688 | gi->offset_map_size_p = 0; |
689 | gi->offset_map_n = NULL; |
690 | gi->offset_map_p = NULL; |
691 | rtx_group_vec.safe_push (obj: gi); |
692 | } |
693 | |
694 | return gi; |
695 | } |
696 | |
697 | |
698 | /* Initialization of data structures. */ |
699 | |
700 | static void |
701 | dse_step0 (void) |
702 | { |
703 | locally_deleted = 0; |
704 | globally_deleted = 0; |
705 | |
706 | bitmap_obstack_initialize (&dse_bitmap_obstack); |
707 | gcc_obstack_init (&dse_obstack); |
708 | |
709 | scratch = BITMAP_ALLOC (obstack: ®_obstack); |
710 | kill_on_calls = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
711 | |
712 | |
713 | rtx_group_table = new hash_table<invariant_group_base_hasher> (11); |
714 | |
715 | bb_table = XNEWVEC (bb_info_t, last_basic_block_for_fn (cfun)); |
716 | rtx_group_next_id = 0; |
717 | |
718 | stores_off_frame_dead_at_return = !cfun->stdarg; |
719 | |
720 | init_alias_analysis (); |
721 | } |
722 | |
723 | |
724 | |
725 | /*---------------------------------------------------------------------------- |
726 | First step. |
727 | |
728 | Scan all of the insns. Any random ordering of the blocks is fine. |
729 | Each block is scanned in forward order to accommodate cselib which |
730 | is used to remove stores with non-constant bases. |
731 | ----------------------------------------------------------------------------*/ |
732 | |
733 | /* Delete all of the store_info recs from INSN_INFO. */ |
734 | |
735 | static void |
736 | free_store_info (insn_info_t insn_info) |
737 | { |
738 | store_info *cur = insn_info->store_rec; |
739 | while (cur) |
740 | { |
741 | store_info *next = cur->next; |
742 | if (cur->is_large) |
743 | BITMAP_FREE (cur->positions_needed.large.bmap); |
744 | if (cur->cse_base) |
745 | cse_store_info_pool.remove (object: cur); |
746 | else |
747 | rtx_store_info_pool.remove (object: cur); |
748 | cur = next; |
749 | } |
750 | |
751 | insn_info->cannot_delete = true; |
752 | insn_info->contains_cselib_groups = false; |
753 | insn_info->store_rec = NULL; |
754 | } |
755 | |
756 | struct note_add_store_info |
757 | { |
758 | rtx_insn *first, *current; |
759 | regset fixed_regs_live; |
760 | bool failure; |
761 | }; |
762 | |
763 | /* Callback for emit_inc_dec_insn_before via note_stores. |
764 | Check if a register is clobbered which is live afterwards. */ |
765 | |
766 | static void |
767 | note_add_store (rtx loc, const_rtx expr ATTRIBUTE_UNUSED, void *data) |
768 | { |
769 | rtx_insn *insn; |
770 | note_add_store_info *info = (note_add_store_info *) data; |
771 | |
772 | if (!REG_P (loc)) |
773 | return; |
774 | |
775 | /* If this register is referenced by the current or an earlier insn, |
776 | that's OK. E.g. this applies to the register that is being incremented |
777 | with this addition. */ |
778 | for (insn = info->first; |
779 | insn != NEXT_INSN (insn: info->current); |
780 | insn = NEXT_INSN (insn)) |
781 | if (reg_referenced_p (loc, PATTERN (insn))) |
782 | return; |
783 | |
784 | /* If we come here, we have a clobber of a register that's only OK |
785 | if that register is not live. If we don't have liveness information |
786 | available, fail now. */ |
787 | if (!info->fixed_regs_live) |
788 | { |
789 | info->failure = true; |
790 | return; |
791 | } |
792 | /* Now check if this is a live fixed register. */ |
793 | unsigned int end_regno = END_REGNO (x: loc); |
794 | for (unsigned int regno = REGNO (loc); regno < end_regno; ++regno) |
795 | if (REGNO_REG_SET_P (info->fixed_regs_live, regno)) |
796 | info->failure = true; |
797 | } |
798 | |
799 | /* Callback for for_each_inc_dec that emits an INSN that sets DEST to |
800 | SRC + SRCOFF before insn ARG. */ |
801 | |
802 | static int |
803 | emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED, |
804 | rtx op ATTRIBUTE_UNUSED, |
805 | rtx dest, rtx src, rtx srcoff, void *arg) |
806 | { |
807 | insn_info_t insn_info = (insn_info_t) arg; |
808 | rtx_insn *insn = insn_info->insn, *new_insn, *cur; |
809 | note_add_store_info info; |
810 | |
811 | /* We can reuse all operands without copying, because we are about |
812 | to delete the insn that contained it. */ |
813 | if (srcoff) |
814 | { |
815 | start_sequence (); |
816 | emit_insn (gen_add3_insn (dest, src, srcoff)); |
817 | new_insn = end_sequence (); |
818 | } |
819 | else |
820 | new_insn = gen_move_insn (dest, src); |
821 | info.first = new_insn; |
822 | info.fixed_regs_live = insn_info->fixed_regs_live; |
823 | info.failure = false; |
824 | for (cur = new_insn; cur; cur = NEXT_INSN (insn: cur)) |
825 | { |
826 | info.current = cur; |
827 | note_stores (cur, note_add_store, &info); |
828 | } |
829 | |
830 | /* If a failure was flagged above, return 1 so that for_each_inc_dec will |
831 | return it immediately, communicating the failure to its caller. */ |
832 | if (info.failure) |
833 | return 1; |
834 | |
835 | emit_insn_before (new_insn, insn); |
836 | |
837 | return 0; |
838 | } |
839 | |
840 | /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it |
841 | is there, is split into a separate insn. |
842 | Return true on success (or if there was nothing to do), false on failure. */ |
843 | |
844 | static bool |
845 | check_for_inc_dec_1 (insn_info_t insn_info) |
846 | { |
847 | rtx_insn *insn = insn_info->insn; |
848 | rtx note = find_reg_note (insn, REG_INC, NULL_RTX); |
849 | if (note) |
850 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
851 | arg: insn_info) == 0; |
852 | |
853 | /* Punt on stack pushes, those don't have REG_INC notes and we are |
854 | unprepared to deal with distribution of REG_ARGS_SIZE notes etc. */ |
855 | subrtx_iterator::array_type array; |
856 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) |
857 | { |
858 | const_rtx x = *iter; |
859 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) |
860 | return false; |
861 | } |
862 | |
863 | return true; |
864 | } |
865 | |
866 | |
867 | /* Entry point for postreload. If you work on reload_cse, or you need this |
868 | anywhere else, consider if you can provide register liveness information |
869 | and add a parameter to this function so that it can be passed down in |
870 | insn_info.fixed_regs_live. */ |
871 | bool |
872 | check_for_inc_dec (rtx_insn *insn) |
873 | { |
874 | insn_info_type insn_info; |
875 | rtx note; |
876 | |
877 | insn_info.insn = insn; |
878 | insn_info.fixed_regs_live = NULL; |
879 | note = find_reg_note (insn, REG_INC, NULL_RTX); |
880 | if (note) |
881 | return for_each_inc_dec (PATTERN (insn), emit_inc_dec_insn_before, |
882 | arg: &insn_info) == 0; |
883 | |
884 | /* Punt on stack pushes, those don't have REG_INC notes and we are |
885 | unprepared to deal with distribution of REG_ARGS_SIZE notes etc. */ |
886 | subrtx_iterator::array_type array; |
887 | FOR_EACH_SUBRTX (iter, array, PATTERN (insn), NONCONST) |
888 | { |
889 | const_rtx x = *iter; |
890 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) |
891 | return false; |
892 | } |
893 | |
894 | return true; |
895 | } |
896 | |
897 | /* Delete the insn and free all of the fields inside INSN_INFO. */ |
898 | |
899 | static void |
900 | delete_dead_store_insn (insn_info_t insn_info) |
901 | { |
902 | read_info_t read_info; |
903 | |
904 | if (!dbg_cnt (index: dse)) |
905 | return; |
906 | |
907 | if (!check_for_inc_dec_1 (insn_info)) |
908 | return; |
909 | if (dump_file && (dump_flags & TDF_DETAILS)) |
910 | fprintf (stream: dump_file, format: "Locally deleting insn %d\n", |
911 | INSN_UID (insn: insn_info->insn)); |
912 | |
913 | free_store_info (insn_info); |
914 | read_info = insn_info->read_rec; |
915 | |
916 | while (read_info) |
917 | { |
918 | read_info_t next = read_info->next; |
919 | read_info_type_pool.remove (object: read_info); |
920 | read_info = next; |
921 | } |
922 | insn_info->read_rec = NULL; |
923 | |
924 | delete_insn (insn_info->insn); |
925 | locally_deleted++; |
926 | insn_info->insn = NULL; |
927 | |
928 | insn_info->wild_read = false; |
929 | } |
930 | |
931 | /* Return whether DECL, a local variable, can possibly escape the current |
932 | function scope. */ |
933 | |
934 | static bool |
935 | local_variable_can_escape (tree decl) |
936 | { |
937 | if (TREE_ADDRESSABLE (decl)) |
938 | return true; |
939 | |
940 | /* If this is a partitioned variable, we need to consider all the variables |
941 | in the partition. This is necessary because a store into one of them can |
942 | be replaced with a store into another and this may not change the outcome |
943 | of the escape analysis. */ |
944 | if (cfun->gimple_df->decls_to_pointers != NULL) |
945 | { |
946 | tree *namep = cfun->gimple_df->decls_to_pointers->get (k: decl); |
947 | if (namep) |
948 | return TREE_ADDRESSABLE (*namep); |
949 | } |
950 | |
951 | return false; |
952 | } |
953 | |
954 | /* Return whether EXPR can possibly escape the current function scope. */ |
955 | |
956 | static bool |
957 | can_escape (tree expr) |
958 | { |
959 | tree base; |
960 | if (!expr) |
961 | return true; |
962 | base = get_base_address (t: expr); |
963 | if (DECL_P (base) |
964 | && !may_be_aliased (var: base) |
965 | && !(VAR_P (base) |
966 | && !DECL_EXTERNAL (base) |
967 | && !TREE_STATIC (base) |
968 | && local_variable_can_escape (decl: base))) |
969 | return false; |
970 | return true; |
971 | } |
972 | |
973 | /* Set the store* bitmaps offset_map_size* fields in GROUP based on |
974 | OFFSET and WIDTH. */ |
975 | |
976 | static void |
977 | set_usage_bits (group_info *group, poly_int64 offset, poly_int64 width, |
978 | tree expr) |
979 | { |
980 | /* Non-constant offsets and widths act as global kills, so there's no point |
981 | trying to use them to derive global DSE candidates. */ |
982 | HOST_WIDE_INT i, const_offset, const_width; |
983 | bool expr_escapes = can_escape (expr); |
984 | if (offset.is_constant (const_value: &const_offset) |
985 | && width.is_constant (const_value: &const_width) |
986 | && const_offset > -MAX_OFFSET |
987 | && const_offset + const_width < MAX_OFFSET) |
988 | for (i = const_offset; i < const_offset + const_width; ++i) |
989 | { |
990 | bitmap store1; |
991 | bitmap store2; |
992 | bitmap escaped; |
993 | int ai; |
994 | if (i < 0) |
995 | { |
996 | store1 = group->store1_n; |
997 | store2 = group->store2_n; |
998 | escaped = group->escaped_n; |
999 | ai = -i; |
1000 | } |
1001 | else |
1002 | { |
1003 | store1 = group->store1_p; |
1004 | store2 = group->store2_p; |
1005 | escaped = group->escaped_p; |
1006 | ai = i; |
1007 | } |
1008 | |
1009 | if (!bitmap_set_bit (store1, ai)) |
1010 | bitmap_set_bit (store2, ai); |
1011 | else |
1012 | { |
1013 | if (i < 0) |
1014 | { |
1015 | if (group->offset_map_size_n < ai) |
1016 | group->offset_map_size_n = ai; |
1017 | } |
1018 | else |
1019 | { |
1020 | if (group->offset_map_size_p < ai) |
1021 | group->offset_map_size_p = ai; |
1022 | } |
1023 | } |
1024 | if (expr_escapes) |
1025 | bitmap_set_bit (escaped, ai); |
1026 | } |
1027 | } |
1028 | |
1029 | static void |
1030 | reset_active_stores (void) |
1031 | { |
1032 | active_local_stores = NULL; |
1033 | active_local_stores_len = 0; |
1034 | } |
1035 | |
1036 | /* Free all READ_REC of the LAST_INSN of BB_INFO. */ |
1037 | |
1038 | static void |
1039 | free_read_records (bb_info_t bb_info) |
1040 | { |
1041 | insn_info_t insn_info = bb_info->last_insn; |
1042 | read_info_t *ptr = &insn_info->read_rec; |
1043 | while (*ptr) |
1044 | { |
1045 | read_info_t next = (*ptr)->next; |
1046 | read_info_type_pool.remove (object: *ptr); |
1047 | *ptr = next; |
1048 | } |
1049 | } |
1050 | |
1051 | /* Set the BB_INFO so that the last insn is marked as a wild read. */ |
1052 | |
1053 | static void |
1054 | add_wild_read (bb_info_t bb_info) |
1055 | { |
1056 | insn_info_t insn_info = bb_info->last_insn; |
1057 | insn_info->wild_read = true; |
1058 | free_read_records (bb_info); |
1059 | reset_active_stores (); |
1060 | } |
1061 | |
1062 | /* Set the BB_INFO so that the last insn is marked as a wild read of |
1063 | non-frame locations. */ |
1064 | |
1065 | static void |
1066 | add_non_frame_wild_read (bb_info_t bb_info) |
1067 | { |
1068 | insn_info_t insn_info = bb_info->last_insn; |
1069 | insn_info->non_frame_wild_read = true; |
1070 | free_read_records (bb_info); |
1071 | reset_active_stores (); |
1072 | } |
1073 | |
1074 | /* Return true if X is a constant or one of the registers that behave |
1075 | as a constant over the life of a function. This is equivalent to |
1076 | !rtx_varies_p for memory addresses. */ |
1077 | |
1078 | static bool |
1079 | const_or_frame_p (rtx x) |
1080 | { |
1081 | if (CONSTANT_P (x)) |
1082 | return true; |
1083 | |
1084 | if (GET_CODE (x) == REG) |
1085 | { |
1086 | /* Note that we have to test for the actual rtx used for the frame |
1087 | and arg pointers and not just the register number in case we have |
1088 | eliminated the frame and/or arg pointer and are using it |
1089 | for pseudos. */ |
1090 | if (x == frame_pointer_rtx || x == hard_frame_pointer_rtx |
1091 | /* The arg pointer varies if it is not a fixed register. */ |
1092 | || (x == arg_pointer_rtx && fixed_regs[ARG_POINTER_REGNUM]) |
1093 | || x == pic_offset_table_rtx) |
1094 | return true; |
1095 | return false; |
1096 | } |
1097 | |
1098 | return false; |
1099 | } |
1100 | |
1101 | /* Take all reasonable action to put the address of MEM into the form |
1102 | that we can do analysis on. |
1103 | |
1104 | The gold standard is to get the address into the form: address + |
1105 | OFFSET where address is something that rtx_varies_p considers a |
1106 | constant. When we can get the address in this form, we can do |
1107 | global analysis on it. Note that for constant bases, address is |
1108 | not actually returned, only the group_id. The address can be |
1109 | obtained from that. |
1110 | |
1111 | If that fails, we try cselib to get a value we can at least use |
1112 | locally. If that fails we return false. |
1113 | |
1114 | The GROUP_ID is set to -1 for cselib bases and the index of the |
1115 | group for non_varying bases. |
1116 | |
1117 | FOR_READ is true if this is a mem read and false if not. */ |
1118 | |
1119 | static bool |
1120 | canon_address (rtx mem, |
1121 | int *group_id, |
1122 | poly_int64 *offset, |
1123 | cselib_val **base) |
1124 | { |
1125 | machine_mode address_mode = get_address_mode (mem); |
1126 | rtx mem_address = XEXP (mem, 0); |
1127 | rtx expanded_address, address; |
1128 | int expanded; |
1129 | |
1130 | cselib_lookup (mem_address, address_mode, 1, GET_MODE (mem)); |
1131 | |
1132 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1133 | { |
1134 | fprintf (stream: dump_file, format: " mem: "); |
1135 | print_inline_rtx (dump_file, mem_address, 0); |
1136 | fprintf (stream: dump_file, format: "\n"); |
1137 | } |
1138 | |
1139 | /* First see if just canon_rtx (mem_address) is const or frame, |
1140 | if not, try cselib_expand_value_rtx and call canon_rtx on that. */ |
1141 | address = NULL_RTX; |
1142 | for (expanded = 0; expanded < 2; expanded++) |
1143 | { |
1144 | if (expanded) |
1145 | { |
1146 | /* Use cselib to replace all of the reg references with the full |
1147 | expression. This will take care of the case where we have |
1148 | |
1149 | r_x = base + offset; |
1150 | val = *r_x; |
1151 | |
1152 | by making it into |
1153 | |
1154 | val = *(base + offset); */ |
1155 | |
1156 | expanded_address = cselib_expand_value_rtx (mem_address, |
1157 | scratch, 5); |
1158 | |
1159 | /* If this fails, just go with the address from first |
1160 | iteration. */ |
1161 | if (!expanded_address) |
1162 | break; |
1163 | } |
1164 | else |
1165 | expanded_address = mem_address; |
1166 | |
1167 | /* Split the address into canonical BASE + OFFSET terms. */ |
1168 | address = canon_rtx (expanded_address); |
1169 | |
1170 | *offset = 0; |
1171 | |
1172 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1173 | { |
1174 | if (expanded) |
1175 | { |
1176 | fprintf (stream: dump_file, format: "\n after cselib_expand address: "); |
1177 | print_inline_rtx (dump_file, expanded_address, 0); |
1178 | fprintf (stream: dump_file, format: "\n"); |
1179 | } |
1180 | |
1181 | fprintf (stream: dump_file, format: "\n after canon_rtx address: "); |
1182 | print_inline_rtx (dump_file, address, 0); |
1183 | fprintf (stream: dump_file, format: "\n"); |
1184 | } |
1185 | |
1186 | if (GET_CODE (address) == CONST) |
1187 | address = XEXP (address, 0); |
1188 | |
1189 | address = strip_offset_and_add (x: address, offset); |
1190 | |
1191 | if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem)) |
1192 | && const_or_frame_p (x: address) |
1193 | /* Literal addresses can alias any base, avoid creating a |
1194 | group for them. */ |
1195 | && ! CONST_SCALAR_INT_P (address)) |
1196 | { |
1197 | group_info *group = get_group_info (base: address); |
1198 | |
1199 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1200 | { |
1201 | fprintf (stream: dump_file, format: " gid=%d offset=", group->id); |
1202 | print_dec (value: *offset, file: dump_file); |
1203 | fprintf (stream: dump_file, format: "\n"); |
1204 | } |
1205 | *base = NULL; |
1206 | *group_id = group->id; |
1207 | return true; |
1208 | } |
1209 | } |
1210 | |
1211 | *base = cselib_lookup (address, address_mode, true, GET_MODE (mem)); |
1212 | *group_id = -1; |
1213 | |
1214 | if (*base == NULL) |
1215 | { |
1216 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1217 | fprintf (stream: dump_file, format: " no cselib val - should be a wild read.\n"); |
1218 | return false; |
1219 | } |
1220 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1221 | { |
1222 | fprintf (stream: dump_file, format: " varying cselib base=%u:%u offset = ", |
1223 | (*base)->uid, (*base)->hash); |
1224 | print_dec (value: *offset, file: dump_file); |
1225 | fprintf (stream: dump_file, format: "\n"); |
1226 | } |
1227 | return true; |
1228 | } |
1229 | |
1230 | |
1231 | /* Clear the rhs field from the active_local_stores array. */ |
1232 | |
1233 | static void |
1234 | clear_rhs_from_active_local_stores (void) |
1235 | { |
1236 | insn_info_t ptr = active_local_stores; |
1237 | |
1238 | while (ptr) |
1239 | { |
1240 | store_info *store_info = ptr->store_rec; |
1241 | /* Skip the clobbers. */ |
1242 | while (!store_info->is_set) |
1243 | store_info = store_info->next; |
1244 | |
1245 | store_info->rhs = NULL; |
1246 | store_info->const_rhs = NULL; |
1247 | |
1248 | ptr = ptr->next_local_store; |
1249 | } |
1250 | } |
1251 | |
1252 | |
1253 | /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */ |
1254 | |
1255 | static inline void |
1256 | set_position_unneeded (store_info *s_info, int pos) |
1257 | { |
1258 | if (UNLIKELY (s_info->is_large)) |
1259 | { |
1260 | if (bitmap_set_bit (s_info->positions_needed.large.bmap, pos)) |
1261 | s_info->positions_needed.large.count++; |
1262 | } |
1263 | else |
1264 | s_info->positions_needed.small_bitmask |
1265 | &= ~(HOST_WIDE_INT_1U << pos); |
1266 | } |
1267 | |
1268 | /* Mark the whole store S_INFO as unneeded. */ |
1269 | |
1270 | static inline void |
1271 | set_all_positions_unneeded (store_info *s_info) |
1272 | { |
1273 | if (UNLIKELY (s_info->is_large)) |
1274 | { |
1275 | HOST_WIDE_INT width; |
1276 | if (s_info->width.is_constant (const_value: &width)) |
1277 | { |
1278 | bitmap_set_range (s_info->positions_needed.large.bmap, 0, width); |
1279 | s_info->positions_needed.large.count = width; |
1280 | } |
1281 | else |
1282 | { |
1283 | gcc_checking_assert (!s_info->positions_needed.large.bmap); |
1284 | s_info->positions_needed.large.count = 1; |
1285 | } |
1286 | } |
1287 | else |
1288 | s_info->positions_needed.small_bitmask = HOST_WIDE_INT_0U; |
1289 | } |
1290 | |
1291 | /* Return TRUE if any bytes from S_INFO store are needed. */ |
1292 | |
1293 | static inline bool |
1294 | any_positions_needed_p (store_info *s_info) |
1295 | { |
1296 | if (UNLIKELY (s_info->is_large)) |
1297 | { |
1298 | HOST_WIDE_INT width; |
1299 | if (s_info->width.is_constant (const_value: &width)) |
1300 | { |
1301 | gcc_checking_assert (s_info->positions_needed.large.bmap); |
1302 | return s_info->positions_needed.large.count < width; |
1303 | } |
1304 | else |
1305 | { |
1306 | gcc_checking_assert (!s_info->positions_needed.large.bmap); |
1307 | return s_info->positions_needed.large.count == 0; |
1308 | } |
1309 | } |
1310 | else |
1311 | return (s_info->positions_needed.small_bitmask != HOST_WIDE_INT_0U); |
1312 | } |
1313 | |
1314 | /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO |
1315 | store are known to be needed. */ |
1316 | |
1317 | static inline bool |
1318 | all_positions_needed_p (store_info *s_info, poly_int64 start, |
1319 | poly_int64 width) |
1320 | { |
1321 | gcc_assert (s_info->rhs); |
1322 | if (!s_info->width.is_constant ()) |
1323 | { |
1324 | gcc_assert (s_info->is_large |
1325 | && !s_info->positions_needed.large.bmap); |
1326 | return s_info->positions_needed.large.count == 0; |
1327 | } |
1328 | |
1329 | /* Otherwise, if START and WIDTH are non-constant, we're asking about |
1330 | a non-constant region of a constant-sized store. We can't say for |
1331 | sure that all positions are needed. */ |
1332 | HOST_WIDE_INT const_start, const_width; |
1333 | if (!start.is_constant (const_value: &const_start) |
1334 | || !width.is_constant (const_value: &const_width)) |
1335 | return false; |
1336 | |
1337 | if (UNLIKELY (s_info->is_large)) |
1338 | { |
1339 | for (HOST_WIDE_INT i = const_start; i < const_start + const_width; ++i) |
1340 | if (bitmap_bit_p (s_info->positions_needed.large.bmap, i)) |
1341 | return false; |
1342 | return true; |
1343 | } |
1344 | else |
1345 | { |
1346 | unsigned HOST_WIDE_INT mask |
1347 | = lowpart_bitmask (n: const_width) << const_start; |
1348 | return (s_info->positions_needed.small_bitmask & mask) == mask; |
1349 | } |
1350 | } |
1351 | |
1352 | |
1353 | static rtx get_stored_val (store_info *, machine_mode, poly_int64, |
1354 | poly_int64, basic_block, bool); |
1355 | |
1356 | |
1357 | /* BODY is an instruction pattern that belongs to INSN. Return 1 if |
1358 | there is a candidate store, after adding it to the appropriate |
1359 | local store group if so. */ |
1360 | |
1361 | static int |
1362 | record_store (rtx body, bb_info_t bb_info) |
1363 | { |
1364 | rtx mem, rhs, const_rhs, mem_addr; |
1365 | poly_int64 offset = 0; |
1366 | poly_int64 width = 0; |
1367 | insn_info_t insn_info = bb_info->last_insn; |
1368 | store_info *store_info = NULL; |
1369 | int group_id; |
1370 | cselib_val *base = NULL; |
1371 | insn_info_t ptr, last, redundant_reason; |
1372 | bool store_is_unused; |
1373 | |
1374 | if (GET_CODE (body) != SET && GET_CODE (body) != CLOBBER) |
1375 | return 0; |
1376 | |
1377 | mem = SET_DEST (body); |
1378 | |
1379 | /* If this is not used, then this cannot be used to keep the insn |
1380 | from being deleted. On the other hand, it does provide something |
1381 | that can be used to prove that another store is dead. */ |
1382 | store_is_unused |
1383 | = (find_reg_note (insn_info->insn, REG_UNUSED, mem) != NULL); |
1384 | |
1385 | /* Check whether that value is a suitable memory location. */ |
1386 | if (!MEM_P (mem)) |
1387 | { |
1388 | /* If the set or clobber is unused, then it does not effect our |
1389 | ability to get rid of the entire insn. */ |
1390 | if (!store_is_unused) |
1391 | insn_info->cannot_delete = true; |
1392 | return 0; |
1393 | } |
1394 | |
1395 | /* At this point we know mem is a mem. */ |
1396 | if (GET_MODE (mem) == BLKmode) |
1397 | { |
1398 | HOST_WIDE_INT const_size; |
1399 | if (GET_CODE (XEXP (mem, 0)) == SCRATCH) |
1400 | { |
1401 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1402 | fprintf (stream: dump_file, format: " adding wild read for (clobber (mem:BLK (scratch))\n"); |
1403 | add_wild_read (bb_info); |
1404 | insn_info->cannot_delete = true; |
1405 | return 0; |
1406 | } |
1407 | /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0)) |
1408 | as memset (addr, 0, 36); */ |
1409 | else if (!MEM_SIZE_KNOWN_P (mem) |
1410 | || maybe_le (MEM_SIZE (mem), b: 0) |
1411 | /* This is a limit on the bitmap size, which is only relevant |
1412 | for constant-sized MEMs. */ |
1413 | || (MEM_SIZE (mem).is_constant (const_value: &const_size) |
1414 | && const_size > MAX_OFFSET) |
1415 | || GET_CODE (body) != SET |
1416 | || !CONST_INT_P (SET_SRC (body))) |
1417 | { |
1418 | if (!store_is_unused) |
1419 | { |
1420 | /* If the set or clobber is unused, then it does not effect our |
1421 | ability to get rid of the entire insn. */ |
1422 | insn_info->cannot_delete = true; |
1423 | clear_rhs_from_active_local_stores (); |
1424 | } |
1425 | return 0; |
1426 | } |
1427 | } |
1428 | |
1429 | /* We can still process a volatile mem, we just cannot delete it. */ |
1430 | if (MEM_VOLATILE_P (mem)) |
1431 | insn_info->cannot_delete = true; |
1432 | |
1433 | if (!canon_address (mem, group_id: &group_id, offset: &offset, base: &base)) |
1434 | { |
1435 | clear_rhs_from_active_local_stores (); |
1436 | return 0; |
1437 | } |
1438 | |
1439 | if (GET_MODE (mem) == BLKmode) |
1440 | width = MEM_SIZE (mem); |
1441 | else |
1442 | width = GET_MODE_SIZE (GET_MODE (mem)); |
1443 | |
1444 | if (!endpoint_representable_p (pos: offset, size: width)) |
1445 | { |
1446 | clear_rhs_from_active_local_stores (); |
1447 | return 0; |
1448 | } |
1449 | |
1450 | if (known_eq (width, 0)) |
1451 | return 0; |
1452 | |
1453 | if (group_id >= 0) |
1454 | { |
1455 | /* In the restrictive case where the base is a constant or the |
1456 | frame pointer we can do global analysis. */ |
1457 | |
1458 | group_info *group |
1459 | = rtx_group_vec[group_id]; |
1460 | tree expr = MEM_EXPR (mem); |
1461 | |
1462 | store_info = rtx_store_info_pool.allocate (); |
1463 | set_usage_bits (group, offset, width, expr); |
1464 | |
1465 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1466 | { |
1467 | fprintf (stream: dump_file, format: " processing const base store gid=%d", |
1468 | group_id); |
1469 | print_range (file: dump_file, offset, width); |
1470 | fprintf (stream: dump_file, format: "\n"); |
1471 | } |
1472 | } |
1473 | else |
1474 | { |
1475 | if (may_be_sp_based_p (XEXP (mem, 0))) |
1476 | insn_info->stack_pointer_based = true; |
1477 | insn_info->contains_cselib_groups = true; |
1478 | |
1479 | store_info = cse_store_info_pool.allocate (); |
1480 | group_id = -1; |
1481 | |
1482 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1483 | { |
1484 | fprintf (stream: dump_file, format: " processing cselib store "); |
1485 | print_range (file: dump_file, offset, width); |
1486 | fprintf (stream: dump_file, format: "\n"); |
1487 | } |
1488 | } |
1489 | |
1490 | const_rhs = rhs = NULL_RTX; |
1491 | if (GET_CODE (body) == SET |
1492 | /* No place to keep the value after ra. */ |
1493 | && !reload_completed |
1494 | && (REG_P (SET_SRC (body)) |
1495 | || GET_CODE (SET_SRC (body)) == SUBREG |
1496 | || CONSTANT_P (SET_SRC (body))) |
1497 | && !MEM_VOLATILE_P (mem) |
1498 | /* Sometimes the store and reload is used for truncation and |
1499 | rounding. */ |
1500 | && !(FLOAT_MODE_P (GET_MODE (mem)) && (flag_float_store))) |
1501 | { |
1502 | rhs = SET_SRC (body); |
1503 | if (CONSTANT_P (rhs)) |
1504 | const_rhs = rhs; |
1505 | else if (body == PATTERN (insn: insn_info->insn)) |
1506 | { |
1507 | rtx tem = find_reg_note (insn_info->insn, REG_EQUAL, NULL_RTX); |
1508 | if (tem && CONSTANT_P (XEXP (tem, 0))) |
1509 | const_rhs = XEXP (tem, 0); |
1510 | } |
1511 | if (const_rhs == NULL_RTX && REG_P (rhs)) |
1512 | { |
1513 | rtx tem = cselib_expand_value_rtx (rhs, scratch, 5); |
1514 | |
1515 | if (tem && CONSTANT_P (tem)) |
1516 | const_rhs = tem; |
1517 | else |
1518 | { |
1519 | /* If RHS is set only once to a constant, set CONST_RHS |
1520 | to the constant. */ |
1521 | rtx def_src = df_find_single_def_src (rhs); |
1522 | if (def_src != nullptr && CONSTANT_P (def_src)) |
1523 | const_rhs = def_src; |
1524 | } |
1525 | } |
1526 | } |
1527 | |
1528 | /* Check to see if this stores causes some other stores to be |
1529 | dead. */ |
1530 | ptr = active_local_stores; |
1531 | last = NULL; |
1532 | redundant_reason = NULL; |
1533 | unsigned char addrspace = MEM_ADDR_SPACE (mem); |
1534 | mem = canon_rtx (mem); |
1535 | |
1536 | if (group_id < 0) |
1537 | mem_addr = base->val_rtx; |
1538 | else |
1539 | { |
1540 | group_info *group = rtx_group_vec[group_id]; |
1541 | mem_addr = group->canon_base_addr; |
1542 | } |
1543 | if (maybe_ne (a: offset, b: 0)) |
1544 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
1545 | |
1546 | while (ptr) |
1547 | { |
1548 | insn_info_t next = ptr->next_local_store; |
1549 | class store_info *s_info = ptr->store_rec; |
1550 | bool del = true; |
1551 | |
1552 | /* Skip the clobbers. We delete the active insn if this insn |
1553 | shadows the set. To have been put on the active list, it |
1554 | has exactly on set. */ |
1555 | while (!s_info->is_set) |
1556 | s_info = s_info->next; |
1557 | |
1558 | if (s_info->group_id == group_id |
1559 | && s_info->cse_base == base |
1560 | && s_info->addrspace == addrspace) |
1561 | { |
1562 | HOST_WIDE_INT i; |
1563 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1564 | { |
1565 | fprintf (stream: dump_file, format: " trying store in insn=%d gid=%d", |
1566 | INSN_UID (insn: ptr->insn), s_info->group_id); |
1567 | print_range (file: dump_file, offset: s_info->offset, width: s_info->width); |
1568 | fprintf (stream: dump_file, format: "\n"); |
1569 | } |
1570 | |
1571 | /* Even if PTR won't be eliminated as unneeded, if both |
1572 | PTR and this insn store the same constant value, we might |
1573 | eliminate this insn instead. */ |
1574 | if (s_info->const_rhs |
1575 | && const_rhs |
1576 | && known_subrange_p (pos1: offset, size1: width, |
1577 | pos2: s_info->offset, size2: s_info->width) |
1578 | && all_positions_needed_p (s_info, start: offset - s_info->offset, |
1579 | width) |
1580 | /* We can only remove the later store if the earlier aliases |
1581 | at least all accesses the later one. */ |
1582 | && mems_same_for_tbaa_p (s_info->mem, mem)) |
1583 | { |
1584 | if (GET_MODE (mem) == BLKmode) |
1585 | { |
1586 | if (GET_MODE (s_info->mem) == BLKmode |
1587 | && s_info->const_rhs == const_rhs) |
1588 | redundant_reason = ptr; |
1589 | } |
1590 | else if (s_info->const_rhs == const0_rtx |
1591 | && const_rhs == const0_rtx) |
1592 | redundant_reason = ptr; |
1593 | else |
1594 | { |
1595 | rtx val; |
1596 | start_sequence (); |
1597 | val = get_stored_val (s_info, GET_MODE (mem), offset, width, |
1598 | BLOCK_FOR_INSN (insn: insn_info->insn), |
1599 | true); |
1600 | if (get_insns () != NULL) |
1601 | val = NULL_RTX; |
1602 | end_sequence (); |
1603 | if (val && rtx_equal_p (val, const_rhs)) |
1604 | redundant_reason = ptr; |
1605 | } |
1606 | } |
1607 | |
1608 | HOST_WIDE_INT begin_unneeded, const_s_width, const_width; |
1609 | if (known_subrange_p (pos1: s_info->offset, size1: s_info->width, pos2: offset, size2: width)) |
1610 | /* The new store touches every byte that S_INFO does. */ |
1611 | set_all_positions_unneeded (s_info); |
1612 | else if ((offset - s_info->offset).is_constant (const_value: &begin_unneeded) |
1613 | && s_info->width.is_constant (const_value: &const_s_width) |
1614 | && width.is_constant (const_value: &const_width)) |
1615 | { |
1616 | HOST_WIDE_INT end_unneeded = begin_unneeded + const_width; |
1617 | begin_unneeded = MAX (begin_unneeded, 0); |
1618 | end_unneeded = MIN (end_unneeded, const_s_width); |
1619 | for (i = begin_unneeded; i < end_unneeded; ++i) |
1620 | set_position_unneeded (s_info, pos: i); |
1621 | } |
1622 | else |
1623 | { |
1624 | /* We don't know which parts of S_INFO are needed and |
1625 | which aren't, so invalidate the RHS. */ |
1626 | s_info->rhs = NULL; |
1627 | s_info->const_rhs = NULL; |
1628 | } |
1629 | } |
1630 | else if (s_info->rhs) |
1631 | /* Need to see if it is possible for this store to overwrite |
1632 | the value of store_info. If it is, set the rhs to NULL to |
1633 | keep it from being used to remove a load. */ |
1634 | { |
1635 | if (canon_output_dependence (s_info->mem, true, |
1636 | mem, GET_MODE (mem), |
1637 | mem_addr)) |
1638 | { |
1639 | s_info->rhs = NULL; |
1640 | s_info->const_rhs = NULL; |
1641 | } |
1642 | } |
1643 | |
1644 | /* An insn can be deleted if every position of every one of |
1645 | its s_infos is zero. */ |
1646 | if (any_positions_needed_p (s_info)) |
1647 | del = false; |
1648 | |
1649 | if (del) |
1650 | { |
1651 | insn_info_t insn_to_delete = ptr; |
1652 | |
1653 | active_local_stores_len--; |
1654 | if (last) |
1655 | last->next_local_store = ptr->next_local_store; |
1656 | else |
1657 | active_local_stores = ptr->next_local_store; |
1658 | |
1659 | if (!insn_to_delete->cannot_delete) |
1660 | delete_dead_store_insn (insn_info: insn_to_delete); |
1661 | } |
1662 | else |
1663 | last = ptr; |
1664 | |
1665 | ptr = next; |
1666 | } |
1667 | |
1668 | /* Finish filling in the store_info. */ |
1669 | store_info->next = insn_info->store_rec; |
1670 | insn_info->store_rec = store_info; |
1671 | store_info->mem = mem; |
1672 | store_info->mem_addr = mem_addr; |
1673 | store_info->cse_base = base; |
1674 | HOST_WIDE_INT const_width; |
1675 | if (!width.is_constant (const_value: &const_width)) |
1676 | { |
1677 | store_info->is_large = true; |
1678 | store_info->positions_needed.large.count = 0; |
1679 | store_info->positions_needed.large.bmap = NULL; |
1680 | } |
1681 | else if (const_width > HOST_BITS_PER_WIDE_INT) |
1682 | { |
1683 | store_info->is_large = true; |
1684 | store_info->positions_needed.large.count = 0; |
1685 | store_info->positions_needed.large.bmap = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
1686 | } |
1687 | else |
1688 | { |
1689 | store_info->is_large = false; |
1690 | store_info->positions_needed.small_bitmask |
1691 | = lowpart_bitmask (n: const_width); |
1692 | } |
1693 | store_info->group_id = group_id; |
1694 | store_info->offset = offset; |
1695 | store_info->width = width; |
1696 | store_info->is_set = GET_CODE (body) == SET; |
1697 | store_info->rhs = rhs; |
1698 | store_info->const_rhs = const_rhs; |
1699 | store_info->redundant_reason = redundant_reason; |
1700 | store_info->addrspace = addrspace; |
1701 | |
1702 | /* If this is a clobber, we return 0. We will only be able to |
1703 | delete this insn if there is only one store USED store, but we |
1704 | can use the clobber to delete other stores earlier. */ |
1705 | return store_info->is_set ? 1 : 0; |
1706 | } |
1707 | |
1708 | |
1709 | static void |
1710 | dump_insn_info (const char * start, insn_info_t insn_info) |
1711 | { |
1712 | fprintf (stream: dump_file, format: "%s insn=%d %s\n", start, |
1713 | INSN_UID (insn: insn_info->insn), |
1714 | insn_info->store_rec ? "has store": "naked"); |
1715 | } |
1716 | |
1717 | |
1718 | /* If the modes are different and the value's source and target do not |
1719 | line up, we need to extract the value from lower part of the rhs of |
1720 | the store, shift it, and then put it into a form that can be shoved |
1721 | into the read_insn. This function generates a right SHIFT of a |
1722 | value that is at least ACCESS_BYTES bytes wide of READ_MODE. The |
1723 | shift sequence is returned or NULL if we failed to find a |
1724 | shift. */ |
1725 | |
1726 | static rtx |
1727 | find_shift_sequence (poly_int64 access_bytes, |
1728 | store_info *store_info, |
1729 | machine_mode read_mode, |
1730 | poly_int64 shift, bool speed, bool require_cst) |
1731 | { |
1732 | machine_mode store_mode = GET_MODE (store_info->mem); |
1733 | scalar_int_mode new_mode; |
1734 | rtx read_reg = NULL; |
1735 | |
1736 | /* If a constant was stored into memory, try to simplify it here, |
1737 | otherwise the cost of the shift might preclude this optimization |
1738 | e.g. at -Os, even when no actual shift will be needed. */ |
1739 | auto access_bits = access_bytes * BITS_PER_UNIT; |
1740 | if (store_info->const_rhs |
1741 | && known_le (access_bytes, GET_MODE_SIZE (MAX_MODE_INT)) |
1742 | && smallest_int_mode_for_size (size: access_bits).exists (mode: &new_mode)) |
1743 | { |
1744 | auto byte = subreg_lowpart_offset (outermode: new_mode, innermode: store_mode); |
1745 | rtx ret |
1746 | = simplify_subreg (outermode: new_mode, op: store_info->const_rhs, innermode: store_mode, byte); |
1747 | if (ret && CONSTANT_P (ret)) |
1748 | { |
1749 | rtx shift_rtx = gen_int_shift_amount (new_mode, shift); |
1750 | ret = simplify_const_binary_operation (LSHIFTRT, new_mode, ret, |
1751 | shift_rtx); |
1752 | if (ret && CONSTANT_P (ret)) |
1753 | { |
1754 | byte = subreg_lowpart_offset (outermode: read_mode, innermode: new_mode); |
1755 | ret = simplify_subreg (outermode: read_mode, op: ret, innermode: new_mode, byte); |
1756 | if (ret && CONSTANT_P (ret) |
1757 | && (set_src_cost (x: ret, mode: read_mode, speed_p: speed) |
1758 | <= COSTS_N_INSNS (1))) |
1759 | return ret; |
1760 | } |
1761 | } |
1762 | } |
1763 | |
1764 | if (require_cst) |
1765 | return NULL_RTX; |
1766 | |
1767 | /* Some machines like the x86 have shift insns for each size of |
1768 | operand. Other machines like the ppc or the ia-64 may only have |
1769 | shift insns that shift values within 32 or 64 bit registers. |
1770 | This loop tries to find the smallest shift insn that will right |
1771 | justify the value we want to read but is available in one insn on |
1772 | the machine. */ |
1773 | |
1774 | opt_scalar_int_mode new_mode_iter; |
1775 | FOR_EACH_MODE_IN_CLASS (new_mode_iter, MODE_INT) |
1776 | { |
1777 | rtx target, new_reg, new_lhs; |
1778 | rtx_insn *shift_seq, *insn; |
1779 | int cost; |
1780 | |
1781 | new_mode = new_mode_iter.require (); |
1782 | if (GET_MODE_BITSIZE (mode: new_mode) > BITS_PER_WORD) |
1783 | break; |
1784 | if (maybe_lt (a: GET_MODE_SIZE (mode: new_mode), b: GET_MODE_SIZE (mode: read_mode))) |
1785 | continue; |
1786 | |
1787 | /* Try a wider mode if truncating the store mode to NEW_MODE |
1788 | requires a real instruction. */ |
1789 | if (maybe_lt (a: GET_MODE_SIZE (mode: new_mode), b: GET_MODE_SIZE (mode: store_mode)) |
1790 | && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode, store_mode)) |
1791 | continue; |
1792 | |
1793 | /* Also try a wider mode if the necessary punning is either not |
1794 | desirable or not possible. */ |
1795 | if (!CONSTANT_P (store_info->rhs) |
1796 | && !targetm.modes_tieable_p (new_mode, store_mode)) |
1797 | continue; |
1798 | |
1799 | if (multiple_p (a: shift, b: GET_MODE_BITSIZE (mode: new_mode)) |
1800 | && known_le (GET_MODE_SIZE (new_mode), GET_MODE_SIZE (store_mode))) |
1801 | { |
1802 | /* Try to implement the shift using a subreg. */ |
1803 | poly_int64 offset |
1804 | = subreg_offset_from_lsb (outer_mode: new_mode, inner_mode: store_mode, lsb_shift: shift); |
1805 | rtx rhs_subreg = simplify_gen_subreg (outermode: new_mode, op: store_info->rhs, |
1806 | innermode: store_mode, byte: offset); |
1807 | if (rhs_subreg) |
1808 | { |
1809 | read_reg |
1810 | = extract_low_bits (read_mode, new_mode, copy_rtx (rhs_subreg)); |
1811 | break; |
1812 | } |
1813 | } |
1814 | |
1815 | if (maybe_lt (a: GET_MODE_SIZE (mode: new_mode), b: access_bytes)) |
1816 | continue; |
1817 | |
1818 | new_reg = gen_reg_rtx (new_mode); |
1819 | |
1820 | start_sequence (); |
1821 | |
1822 | /* In theory we could also check for an ashr. Ian Taylor knows |
1823 | of one dsp where the cost of these two was not the same. But |
1824 | this really is a rare case anyway. */ |
1825 | target = expand_binop (new_mode, lshr_optab, new_reg, |
1826 | gen_int_shift_amount (new_mode, shift), |
1827 | new_reg, 1, OPTAB_DIRECT); |
1828 | |
1829 | shift_seq = end_sequence (); |
1830 | |
1831 | if (target != new_reg || shift_seq == NULL) |
1832 | continue; |
1833 | |
1834 | cost = 0; |
1835 | for (insn = shift_seq; insn != NULL_RTX; insn = NEXT_INSN (insn)) |
1836 | if (INSN_P (insn)) |
1837 | cost += insn_cost (insn, speed); |
1838 | |
1839 | /* The computation up to here is essentially independent |
1840 | of the arguments and could be precomputed. It may |
1841 | not be worth doing so. We could precompute if |
1842 | worthwhile or at least cache the results. The result |
1843 | technically depends on both SHIFT and ACCESS_BYTES, |
1844 | but in practice the answer will depend only on ACCESS_BYTES. */ |
1845 | |
1846 | if (cost > COSTS_N_INSNS (1)) |
1847 | continue; |
1848 | |
1849 | new_lhs = extract_low_bits (new_mode, store_mode, |
1850 | copy_rtx (store_info->rhs)); |
1851 | if (new_lhs == NULL_RTX) |
1852 | continue; |
1853 | |
1854 | /* We found an acceptable shift. Generate a move to |
1855 | take the value from the store and put it into the |
1856 | shift pseudo, then shift it, then generate another |
1857 | move to put in into the target of the read. */ |
1858 | emit_move_insn (new_reg, new_lhs); |
1859 | emit_insn (shift_seq); |
1860 | read_reg = extract_low_bits (read_mode, new_mode, new_reg); |
1861 | break; |
1862 | } |
1863 | |
1864 | return read_reg; |
1865 | } |
1866 | |
1867 | |
1868 | /* Call back for note_stores to find the hard regs set or clobbered by |
1869 | insn. Data is a bitmap of the hardregs set so far. */ |
1870 | |
1871 | static void |
1872 | look_for_hardregs (rtx x, const_rtx pat ATTRIBUTE_UNUSED, void *data) |
1873 | { |
1874 | bitmap regs_set = (bitmap) data; |
1875 | |
1876 | if (REG_P (x) |
1877 | && HARD_REGISTER_P (x)) |
1878 | bitmap_set_range (regs_set, REGNO (x), REG_NREGS (x)); |
1879 | } |
1880 | |
1881 | /* Helper function for replace_read and record_store. |
1882 | Attempt to return a value of mode READ_MODE stored in STORE_INFO, |
1883 | consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL |
1884 | if not successful. If REQUIRE_CST is true, return always constant. */ |
1885 | |
1886 | static rtx |
1887 | get_stored_val (store_info *store_info, machine_mode read_mode, |
1888 | poly_int64 read_offset, poly_int64 read_width, |
1889 | basic_block bb, bool require_cst) |
1890 | { |
1891 | machine_mode store_mode = GET_MODE (store_info->mem); |
1892 | poly_int64 gap; |
1893 | rtx read_reg; |
1894 | |
1895 | /* To get here the read is within the boundaries of the write so |
1896 | shift will never be negative. Start out with the shift being in |
1897 | bytes. */ |
1898 | if (store_mode == BLKmode) |
1899 | gap = 0; |
1900 | else if (BYTES_BIG_ENDIAN) |
1901 | gap = ((store_info->offset + store_info->width) |
1902 | - (read_offset + read_width)); |
1903 | else |
1904 | gap = read_offset - store_info->offset; |
1905 | |
1906 | if (maybe_ne (a: gap, b: 0)) |
1907 | { |
1908 | if (!gap.is_constant ()) |
1909 | return NULL_RTX; |
1910 | |
1911 | poly_int64 shift = gap * BITS_PER_UNIT; |
1912 | poly_int64 access_size = GET_MODE_SIZE (mode: read_mode) + gap; |
1913 | read_reg = find_shift_sequence (access_bytes: access_size, store_info, read_mode, |
1914 | shift, speed: optimize_bb_for_speed_p (bb), |
1915 | require_cst); |
1916 | } |
1917 | else if (store_mode == BLKmode) |
1918 | { |
1919 | /* The store is a memset (addr, const_val, const_size). */ |
1920 | gcc_assert (CONST_INT_P (store_info->rhs)); |
1921 | scalar_int_mode int_store_mode; |
1922 | if (!int_mode_for_mode (read_mode).exists (mode: &int_store_mode)) |
1923 | read_reg = NULL_RTX; |
1924 | else if (store_info->rhs == const0_rtx) |
1925 | read_reg = extract_low_bits (read_mode, int_store_mode, const0_rtx); |
1926 | else if (GET_MODE_BITSIZE (mode: int_store_mode) > HOST_BITS_PER_WIDE_INT |
1927 | || BITS_PER_UNIT >= HOST_BITS_PER_WIDE_INT) |
1928 | read_reg = NULL_RTX; |
1929 | else |
1930 | { |
1931 | unsigned HOST_WIDE_INT c |
1932 | = INTVAL (store_info->rhs) |
1933 | & ((HOST_WIDE_INT_1 << BITS_PER_UNIT) - 1); |
1934 | int shift = BITS_PER_UNIT; |
1935 | while (shift < HOST_BITS_PER_WIDE_INT) |
1936 | { |
1937 | c |= (c << shift); |
1938 | shift <<= 1; |
1939 | } |
1940 | read_reg = gen_int_mode (c, int_store_mode); |
1941 | read_reg = extract_low_bits (read_mode, int_store_mode, read_reg); |
1942 | } |
1943 | } |
1944 | else if (store_info->const_rhs |
1945 | && (require_cst |
1946 | || GET_MODE_CLASS (read_mode) != GET_MODE_CLASS (store_mode))) |
1947 | read_reg = extract_low_bits (read_mode, store_mode, |
1948 | copy_rtx (store_info->const_rhs)); |
1949 | else if (VECTOR_MODE_P (read_mode) && VECTOR_MODE_P (store_mode) |
1950 | && known_le (GET_MODE_BITSIZE (read_mode), GET_MODE_BITSIZE (store_mode)) |
1951 | && targetm.modes_tieable_p (read_mode, store_mode) |
1952 | && validate_subreg (read_mode, store_mode, copy_rtx (store_info->rhs), |
1953 | subreg_lowpart_offset (outermode: read_mode, innermode: store_mode))) |
1954 | read_reg = gen_lowpart (read_mode, copy_rtx (store_info->rhs)); |
1955 | else |
1956 | read_reg = extract_low_bits (read_mode, store_mode, |
1957 | copy_rtx (store_info->rhs)); |
1958 | if (require_cst && read_reg && !CONSTANT_P (read_reg)) |
1959 | read_reg = NULL_RTX; |
1960 | return read_reg; |
1961 | } |
1962 | |
1963 | /* Take a sequence of: |
1964 | A <- r1 |
1965 | ... |
1966 | ... <- A |
1967 | |
1968 | and change it into |
1969 | r2 <- r1 |
1970 | A <- r1 |
1971 | ... |
1972 | ... <- r2 |
1973 | |
1974 | or |
1975 | |
1976 | r3 <- extract (r1) |
1977 | r3 <- r3 >> shift |
1978 | r2 <- extract (r3) |
1979 | ... <- r2 |
1980 | |
1981 | or |
1982 | |
1983 | r2 <- extract (r1) |
1984 | ... <- r2 |
1985 | |
1986 | Depending on the alignment and the mode of the store and |
1987 | subsequent load. |
1988 | |
1989 | |
1990 | The STORE_INFO and STORE_INSN are for the store and READ_INFO |
1991 | and READ_INSN are for the read. Return true if the replacement |
1992 | went ok. */ |
1993 | |
1994 | static bool |
1995 | replace_read (store_info *store_info, insn_info_t store_insn, |
1996 | read_info_t read_info, insn_info_t read_insn, rtx *loc) |
1997 | { |
1998 | machine_mode store_mode = GET_MODE (store_info->mem); |
1999 | machine_mode read_mode = GET_MODE (read_info->mem); |
2000 | rtx_insn *insns, *this_insn; |
2001 | rtx read_reg; |
2002 | basic_block bb; |
2003 | |
2004 | if (!dbg_cnt (index: dse)) |
2005 | return false; |
2006 | |
2007 | /* Create a sequence of instructions to set up the read register. |
2008 | This sequence goes immediately before the store and its result |
2009 | is read by the load. |
2010 | |
2011 | We need to keep this in perspective. We are replacing a read |
2012 | with a sequence of insns, but the read will almost certainly be |
2013 | in cache, so it is not going to be an expensive one. Thus, we |
2014 | are not willing to do a multi insn shift or worse a subroutine |
2015 | call to get rid of the read. */ |
2016 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2017 | fprintf (stream: dump_file, format: "trying to replace %smode load in insn %d" |
2018 | " from %smode store in insn %d\n", |
2019 | GET_MODE_NAME (read_mode), INSN_UID (insn: read_insn->insn), |
2020 | GET_MODE_NAME (store_mode), INSN_UID (insn: store_insn->insn)); |
2021 | start_sequence (); |
2022 | bb = BLOCK_FOR_INSN (insn: read_insn->insn); |
2023 | read_reg = get_stored_val (store_info, |
2024 | read_mode, read_offset: read_info->offset, read_width: read_info->width, |
2025 | bb, require_cst: false); |
2026 | if (read_reg == NULL_RTX) |
2027 | { |
2028 | end_sequence (); |
2029 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2030 | fprintf (stream: dump_file, format: " -- could not extract bits of stored value\n"); |
2031 | return false; |
2032 | } |
2033 | /* Force the value into a new register so that it won't be clobbered |
2034 | between the store and the load. */ |
2035 | if (WORD_REGISTER_OPERATIONS |
2036 | && GET_CODE (read_reg) == SUBREG |
2037 | && REG_P (SUBREG_REG (read_reg)) |
2038 | && GET_MODE (SUBREG_REG (read_reg)) == word_mode) |
2039 | { |
2040 | /* For WORD_REGISTER_OPERATIONS with subreg of word_mode register |
2041 | force SUBREG_REG into a new register rather than the SUBREG. */ |
2042 | rtx r = copy_to_mode_reg (word_mode, SUBREG_REG (read_reg)); |
2043 | read_reg = shallow_copy_rtx (read_reg); |
2044 | SUBREG_REG (read_reg) = r; |
2045 | } |
2046 | else |
2047 | read_reg = copy_to_mode_reg (read_mode, read_reg); |
2048 | insns = end_sequence (); |
2049 | |
2050 | if (insns != NULL_RTX) |
2051 | { |
2052 | /* Now we have to scan the set of new instructions to see if the |
2053 | sequence contains and sets of hardregs that happened to be |
2054 | live at this point. For instance, this can happen if one of |
2055 | the insns sets the CC and the CC happened to be live at that |
2056 | point. This does occasionally happen, see PR 37922. */ |
2057 | bitmap regs_set = BITMAP_ALLOC (obstack: ®_obstack); |
2058 | |
2059 | for (this_insn = insns; |
2060 | this_insn != NULL_RTX; this_insn = NEXT_INSN (insn: this_insn)) |
2061 | { |
2062 | if (insn_invalid_p (this_insn, false)) |
2063 | { |
2064 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2065 | { |
2066 | fprintf (stream: dump_file, format: " -- replacing the loaded MEM with "); |
2067 | print_simple_rtl (dump_file, read_reg); |
2068 | fprintf (stream: dump_file, format: " led to an invalid instruction\n"); |
2069 | } |
2070 | BITMAP_FREE (regs_set); |
2071 | return false; |
2072 | } |
2073 | note_stores (this_insn, look_for_hardregs, regs_set); |
2074 | } |
2075 | |
2076 | if (store_insn->fixed_regs_live) |
2077 | bitmap_and_into (regs_set, store_insn->fixed_regs_live); |
2078 | if (!bitmap_empty_p (map: regs_set)) |
2079 | { |
2080 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2081 | { |
2082 | fprintf (stream: dump_file, format: "abandoning replacement because sequence " |
2083 | "clobbers live hardregs:"); |
2084 | df_print_regset (file: dump_file, r: regs_set); |
2085 | } |
2086 | |
2087 | BITMAP_FREE (regs_set); |
2088 | return false; |
2089 | } |
2090 | BITMAP_FREE (regs_set); |
2091 | } |
2092 | |
2093 | subrtx_iterator::array_type array; |
2094 | FOR_EACH_SUBRTX (iter, array, *loc, NONCONST) |
2095 | { |
2096 | const_rtx x = *iter; |
2097 | if (GET_RTX_CLASS (GET_CODE (x)) == RTX_AUTOINC) |
2098 | { |
2099 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2100 | fprintf (stream: dump_file, format: " -- replacing the MEM failed due to address " |
2101 | "side-effects\n"); |
2102 | return false; |
2103 | } |
2104 | } |
2105 | |
2106 | if (validate_change (read_insn->insn, loc, read_reg, 0)) |
2107 | { |
2108 | deferred_change *change = deferred_change_pool.allocate (); |
2109 | |
2110 | /* Insert this right before the store insn where it will be safe |
2111 | from later insns that might change it before the read. */ |
2112 | emit_insn_before (insns, store_insn->insn); |
2113 | |
2114 | /* And now for the kludge part: cselib croaks if you just |
2115 | return at this point. There are two reasons for this: |
2116 | |
2117 | 1) Cselib has an idea of how many pseudos there are and |
2118 | that does not include the new ones we just added. |
2119 | |
2120 | 2) Cselib does not know about the move insn we added |
2121 | above the store_info, and there is no way to tell it |
2122 | about it, because it has "moved on". |
2123 | |
2124 | Problem (1) is fixable with a certain amount of engineering. |
2125 | Problem (2) is requires starting the bb from scratch. This |
2126 | could be expensive. |
2127 | |
2128 | So we are just going to have to lie. The move/extraction |
2129 | insns are not really an issue, cselib did not see them. But |
2130 | the use of the new pseudo read_insn is a real problem because |
2131 | cselib has not scanned this insn. The way that we solve this |
2132 | problem is that we are just going to put the mem back for now |
2133 | and when we are finished with the block, we undo this. We |
2134 | keep a table of mems to get rid of. At the end of the basic |
2135 | block we can put them back. */ |
2136 | |
2137 | *loc = read_info->mem; |
2138 | change->next = deferred_change_list; |
2139 | deferred_change_list = change; |
2140 | change->loc = loc; |
2141 | change->reg = read_reg; |
2142 | |
2143 | /* Get rid of the read_info, from the point of view of the |
2144 | rest of dse, play like this read never happened. */ |
2145 | read_insn->read_rec = read_info->next; |
2146 | read_info_type_pool.remove (object: read_info); |
2147 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2148 | { |
2149 | fprintf (stream: dump_file, format: " -- replaced the loaded MEM with "); |
2150 | print_simple_rtl (dump_file, read_reg); |
2151 | fprintf (stream: dump_file, format: "\n"); |
2152 | } |
2153 | return true; |
2154 | } |
2155 | else |
2156 | { |
2157 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2158 | { |
2159 | fprintf (stream: dump_file, format: " -- replacing the loaded MEM with "); |
2160 | print_simple_rtl (dump_file, read_reg); |
2161 | fprintf (stream: dump_file, format: " led to an invalid instruction\n"); |
2162 | } |
2163 | return false; |
2164 | } |
2165 | } |
2166 | |
2167 | /* Check the address of MEM *LOC and kill any appropriate stores that may |
2168 | be active. */ |
2169 | |
2170 | static void |
2171 | check_mem_read_rtx (rtx *loc, bb_info_t bb_info, bool used_in_call = false) |
2172 | { |
2173 | rtx mem = *loc, mem_addr; |
2174 | insn_info_t insn_info; |
2175 | poly_int64 offset = 0; |
2176 | poly_int64 width = 0; |
2177 | cselib_val *base = NULL; |
2178 | int group_id; |
2179 | read_info_t read_info; |
2180 | |
2181 | insn_info = bb_info->last_insn; |
2182 | |
2183 | if ((MEM_ALIAS_SET (mem) == ALIAS_SET_MEMORY_BARRIER) |
2184 | || MEM_VOLATILE_P (mem)) |
2185 | { |
2186 | if (crtl->stack_protect_guard |
2187 | && (MEM_EXPR (mem) == crtl->stack_protect_guard |
2188 | || (crtl->stack_protect_guard_decl |
2189 | && MEM_EXPR (mem) == crtl->stack_protect_guard_decl)) |
2190 | && MEM_VOLATILE_P (mem)) |
2191 | { |
2192 | /* This is either the stack protector canary on the stack, |
2193 | which ought to be written by a MEM_VOLATILE_P store and |
2194 | thus shouldn't be deleted and is read at the very end of |
2195 | function, but shouldn't conflict with any other store. |
2196 | Or it is __stack_chk_guard variable or TLS or whatever else |
2197 | MEM holding the canary value, which really shouldn't be |
2198 | ever modified in -fstack-protector* protected functions, |
2199 | otherwise the prologue store wouldn't match the epilogue |
2200 | check. */ |
2201 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2202 | fprintf (stream: dump_file, format: " stack protector canary read ignored.\n"); |
2203 | insn_info->cannot_delete = true; |
2204 | return; |
2205 | } |
2206 | |
2207 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2208 | fprintf (stream: dump_file, format: " adding wild read, volatile or barrier.\n"); |
2209 | add_wild_read (bb_info); |
2210 | insn_info->cannot_delete = true; |
2211 | return; |
2212 | } |
2213 | |
2214 | /* If it is reading readonly mem, then there can be no conflict with |
2215 | another write. */ |
2216 | if (MEM_READONLY_P (mem)) |
2217 | return; |
2218 | |
2219 | if (!canon_address (mem, group_id: &group_id, offset: &offset, base: &base)) |
2220 | { |
2221 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2222 | fprintf (stream: dump_file, format: " adding wild read, canon_address failure.\n"); |
2223 | add_wild_read (bb_info); |
2224 | return; |
2225 | } |
2226 | |
2227 | if (GET_MODE (mem) == BLKmode) |
2228 | width = -1; |
2229 | else |
2230 | width = GET_MODE_SIZE (GET_MODE (mem)); |
2231 | |
2232 | if (!endpoint_representable_p (pos: offset, known_eq (width, -1) ? 1 : width)) |
2233 | { |
2234 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2235 | fprintf (stream: dump_file, format: " adding wild read, due to overflow.\n"); |
2236 | add_wild_read (bb_info); |
2237 | return; |
2238 | } |
2239 | |
2240 | read_info = read_info_type_pool.allocate (); |
2241 | read_info->group_id = group_id; |
2242 | read_info->mem = mem; |
2243 | read_info->offset = offset; |
2244 | read_info->width = width; |
2245 | read_info->next = insn_info->read_rec; |
2246 | insn_info->read_rec = read_info; |
2247 | if (group_id < 0) |
2248 | mem_addr = base->val_rtx; |
2249 | else |
2250 | { |
2251 | group_info *group = rtx_group_vec[group_id]; |
2252 | mem_addr = group->canon_base_addr; |
2253 | } |
2254 | if (maybe_ne (a: offset, b: 0)) |
2255 | mem_addr = plus_constant (get_address_mode (mem), mem_addr, offset); |
2256 | /* Avoid passing VALUE RTXen as mem_addr to canon_true_dependence |
2257 | which will over and over re-create proper RTL and re-apply the |
2258 | offset above. See PR80960 where we almost allocate 1.6GB of PLUS |
2259 | RTXen that way. */ |
2260 | mem_addr = get_addr (mem_addr); |
2261 | |
2262 | if (group_id >= 0) |
2263 | { |
2264 | /* This is the restricted case where the base is a constant or |
2265 | the frame pointer and offset is a constant. */ |
2266 | insn_info_t i_ptr = active_local_stores; |
2267 | insn_info_t last = NULL; |
2268 | |
2269 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2270 | { |
2271 | if (!known_size_p (a: width)) |
2272 | fprintf (stream: dump_file, format: " processing const load gid=%d[BLK]\n", |
2273 | group_id); |
2274 | else |
2275 | { |
2276 | fprintf (stream: dump_file, format: " processing const load gid=%d", group_id); |
2277 | print_range (file: dump_file, offset, width); |
2278 | fprintf (stream: dump_file, format: "\n"); |
2279 | } |
2280 | } |
2281 | |
2282 | while (i_ptr) |
2283 | { |
2284 | bool remove = false; |
2285 | store_info *store_info = i_ptr->store_rec; |
2286 | |
2287 | /* Skip the clobbers. */ |
2288 | while (!store_info->is_set) |
2289 | store_info = store_info->next; |
2290 | |
2291 | /* There are three cases here. */ |
2292 | if (store_info->group_id < 0) |
2293 | /* We have a cselib store followed by a read from a |
2294 | const base. */ |
2295 | remove |
2296 | = canon_true_dependence (store_info->mem, |
2297 | GET_MODE (store_info->mem), |
2298 | store_info->mem_addr, |
2299 | mem, mem_addr); |
2300 | |
2301 | else if (group_id == store_info->group_id) |
2302 | { |
2303 | /* This is a block mode load. We may get lucky and |
2304 | canon_true_dependence may save the day. */ |
2305 | if (!known_size_p (a: width)) |
2306 | remove |
2307 | = canon_true_dependence (store_info->mem, |
2308 | GET_MODE (store_info->mem), |
2309 | store_info->mem_addr, |
2310 | mem, mem_addr); |
2311 | |
2312 | /* If this read is just reading back something that we just |
2313 | stored, rewrite the read. */ |
2314 | else |
2315 | { |
2316 | if (!used_in_call |
2317 | && store_info->rhs |
2318 | && known_subrange_p (pos1: offset, size1: width, pos2: store_info->offset, |
2319 | size2: store_info->width) |
2320 | && all_positions_needed_p (s_info: store_info, |
2321 | start: offset - store_info->offset, |
2322 | width) |
2323 | && replace_read (store_info, store_insn: i_ptr, read_info, |
2324 | read_insn: insn_info, loc)) |
2325 | return; |
2326 | |
2327 | /* The bases are the same, just see if the offsets |
2328 | could overlap. */ |
2329 | if (ranges_maybe_overlap_p (pos1: offset, size1: width, |
2330 | pos2: store_info->offset, |
2331 | size2: store_info->width)) |
2332 | remove = true; |
2333 | } |
2334 | } |
2335 | |
2336 | /* else |
2337 | The else case that is missing here is that the |
2338 | bases are constant but different. There is nothing |
2339 | to do here because there is no overlap. */ |
2340 | |
2341 | if (remove) |
2342 | { |
2343 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2344 | dump_insn_info (start: "removing from active", insn_info: i_ptr); |
2345 | |
2346 | active_local_stores_len--; |
2347 | if (last) |
2348 | last->next_local_store = i_ptr->next_local_store; |
2349 | else |
2350 | active_local_stores = i_ptr->next_local_store; |
2351 | } |
2352 | else |
2353 | last = i_ptr; |
2354 | i_ptr = i_ptr->next_local_store; |
2355 | } |
2356 | } |
2357 | else |
2358 | { |
2359 | insn_info_t i_ptr = active_local_stores; |
2360 | insn_info_t last = NULL; |
2361 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2362 | { |
2363 | fprintf (stream: dump_file, format: " processing cselib load mem:"); |
2364 | print_inline_rtx (dump_file, mem, 0); |
2365 | fprintf (stream: dump_file, format: "\n"); |
2366 | } |
2367 | |
2368 | while (i_ptr) |
2369 | { |
2370 | bool remove = false; |
2371 | store_info *store_info = i_ptr->store_rec; |
2372 | |
2373 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2374 | fprintf (stream: dump_file, format: " processing cselib load against insn %d\n", |
2375 | INSN_UID (insn: i_ptr->insn)); |
2376 | |
2377 | /* Skip the clobbers. */ |
2378 | while (!store_info->is_set) |
2379 | store_info = store_info->next; |
2380 | |
2381 | /* If this read is just reading back something that we just |
2382 | stored, rewrite the read. */ |
2383 | if (!used_in_call |
2384 | && store_info->rhs |
2385 | && store_info->group_id == -1 |
2386 | && store_info->cse_base == base |
2387 | && known_subrange_p (pos1: offset, size1: width, pos2: store_info->offset, |
2388 | size2: store_info->width) |
2389 | && all_positions_needed_p (s_info: store_info, |
2390 | start: offset - store_info->offset, width) |
2391 | && replace_read (store_info, store_insn: i_ptr, read_info, read_insn: insn_info, loc)) |
2392 | return; |
2393 | |
2394 | remove = canon_true_dependence (store_info->mem, |
2395 | GET_MODE (store_info->mem), |
2396 | store_info->mem_addr, |
2397 | mem, mem_addr); |
2398 | |
2399 | if (remove) |
2400 | { |
2401 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2402 | dump_insn_info (start: "removing from active", insn_info: i_ptr); |
2403 | |
2404 | active_local_stores_len--; |
2405 | if (last) |
2406 | last->next_local_store = i_ptr->next_local_store; |
2407 | else |
2408 | active_local_stores = i_ptr->next_local_store; |
2409 | } |
2410 | else |
2411 | last = i_ptr; |
2412 | i_ptr = i_ptr->next_local_store; |
2413 | } |
2414 | } |
2415 | } |
2416 | |
2417 | /* A note_uses callback in which DATA points the INSN_INFO for |
2418 | as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns |
2419 | true for any part of *LOC. */ |
2420 | |
2421 | static void |
2422 | check_mem_read_use (rtx *loc, void *data) |
2423 | { |
2424 | subrtx_ptr_iterator::array_type array; |
2425 | FOR_EACH_SUBRTX_PTR (iter, array, loc, NONCONST) |
2426 | { |
2427 | rtx *loc = *iter; |
2428 | if (MEM_P (*loc)) |
2429 | check_mem_read_rtx (loc, bb_info: (bb_info_t) data); |
2430 | } |
2431 | } |
2432 | |
2433 | |
2434 | /* Get arguments passed to CALL_INSN. Return TRUE if successful. |
2435 | So far it only handles arguments passed in registers. */ |
2436 | |
2437 | static bool |
2438 | get_call_args (rtx call_insn, tree fn, rtx *args, int nargs) |
2439 | { |
2440 | CUMULATIVE_ARGS args_so_far_v; |
2441 | cumulative_args_t args_so_far; |
2442 | tree arg; |
2443 | int idx; |
2444 | |
2445 | INIT_CUMULATIVE_ARGS (args_so_far_v, TREE_TYPE (fn), NULL_RTX, 0, 3); |
2446 | args_so_far = pack_cumulative_args (arg: &args_so_far_v); |
2447 | |
2448 | arg = TYPE_ARG_TYPES (TREE_TYPE (fn)); |
2449 | for (idx = 0; |
2450 | arg != void_list_node && idx < nargs; |
2451 | arg = TREE_CHAIN (arg), idx++) |
2452 | { |
2453 | scalar_int_mode mode; |
2454 | rtx reg, link, tmp; |
2455 | |
2456 | if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg)), int_mode: &mode)) |
2457 | return false; |
2458 | |
2459 | function_arg_info arg (mode, /*named=*/true); |
2460 | reg = targetm.calls.function_arg (args_so_far, arg); |
2461 | if (!reg || !REG_P (reg) || GET_MODE (reg) != mode) |
2462 | return false; |
2463 | |
2464 | for (link = CALL_INSN_FUNCTION_USAGE (call_insn); |
2465 | link; |
2466 | link = XEXP (link, 1)) |
2467 | if (GET_CODE (XEXP (link, 0)) == USE) |
2468 | { |
2469 | scalar_int_mode arg_mode; |
2470 | args[idx] = XEXP (XEXP (link, 0), 0); |
2471 | if (REG_P (args[idx]) |
2472 | && REGNO (args[idx]) == REGNO (reg) |
2473 | && (GET_MODE (args[idx]) == mode |
2474 | || (is_int_mode (GET_MODE (args[idx]), int_mode: &arg_mode) |
2475 | && (GET_MODE_SIZE (mode: arg_mode) <= UNITS_PER_WORD) |
2476 | && (GET_MODE_SIZE (mode: arg_mode) > GET_MODE_SIZE (mode))))) |
2477 | break; |
2478 | } |
2479 | if (!link) |
2480 | return false; |
2481 | |
2482 | tmp = cselib_expand_value_rtx (args[idx], scratch, 5); |
2483 | if (GET_MODE (args[idx]) != mode) |
2484 | { |
2485 | if (!tmp || !CONST_INT_P (tmp)) |
2486 | return false; |
2487 | tmp = gen_int_mode (INTVAL (tmp), mode); |
2488 | } |
2489 | if (tmp) |
2490 | args[idx] = tmp; |
2491 | |
2492 | targetm.calls.function_arg_advance (args_so_far, arg); |
2493 | } |
2494 | if (arg != void_list_node || idx != nargs) |
2495 | return false; |
2496 | return true; |
2497 | } |
2498 | |
2499 | /* Return a bitmap of the fixed registers contained in IN. */ |
2500 | |
2501 | static bitmap |
2502 | copy_fixed_regs (const_bitmap in) |
2503 | { |
2504 | bitmap ret; |
2505 | |
2506 | ret = ALLOC_REG_SET (NULL); |
2507 | bitmap_and (ret, in, bitmap_view<HARD_REG_SET> (fixed_reg_set)); |
2508 | return ret; |
2509 | } |
2510 | |
2511 | /* Apply record_store to all candidate stores in INSN. Mark INSN |
2512 | if some part of it is not a candidate store and assigns to a |
2513 | non-register target. */ |
2514 | |
2515 | static void |
2516 | scan_insn (bb_info_t bb_info, rtx_insn *insn, int max_active_local_stores) |
2517 | { |
2518 | rtx body; |
2519 | insn_info_type *insn_info = insn_info_type_pool.allocate (); |
2520 | int mems_found = 0; |
2521 | memset (s: insn_info, c: 0, n: sizeof (struct insn_info_type)); |
2522 | |
2523 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2524 | fprintf (stream: dump_file, format: "\n**scanning insn=%d\n", |
2525 | INSN_UID (insn)); |
2526 | |
2527 | insn_info->prev_insn = bb_info->last_insn; |
2528 | insn_info->insn = insn; |
2529 | bb_info->last_insn = insn_info; |
2530 | |
2531 | if (DEBUG_INSN_P (insn)) |
2532 | { |
2533 | insn_info->cannot_delete = true; |
2534 | return; |
2535 | } |
2536 | |
2537 | /* Look at all of the uses in the insn. */ |
2538 | note_uses (&PATTERN (insn), check_mem_read_use, bb_info); |
2539 | |
2540 | if (CALL_P (insn)) |
2541 | { |
2542 | bool const_call; |
2543 | rtx call, sym; |
2544 | tree memset_call = NULL_TREE; |
2545 | |
2546 | insn_info->cannot_delete = true; |
2547 | |
2548 | /* Const functions cannot do anything bad i.e. read memory, |
2549 | however, they can read their parameters which may have |
2550 | been pushed onto the stack. |
2551 | memset and bzero don't read memory either. */ |
2552 | const_call = RTL_CONST_CALL_P (insn); |
2553 | if (!const_call |
2554 | && (call = get_call_rtx_from (insn)) |
2555 | && (sym = XEXP (XEXP (call, 0), 0)) |
2556 | && GET_CODE (sym) == SYMBOL_REF |
2557 | && SYMBOL_REF_DECL (sym) |
2558 | && TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL |
2559 | && fndecl_built_in_p (SYMBOL_REF_DECL (sym), name1: BUILT_IN_MEMSET)) |
2560 | memset_call = SYMBOL_REF_DECL (sym); |
2561 | |
2562 | if (const_call || memset_call) |
2563 | { |
2564 | insn_info_t i_ptr = active_local_stores; |
2565 | insn_info_t last = NULL; |
2566 | |
2567 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2568 | fprintf (stream: dump_file, format: "%s call %d\n", |
2569 | const_call ? "const": "memset", INSN_UID (insn)); |
2570 | |
2571 | /* See the head comment of the frame_read field. */ |
2572 | if (reload_completed |
2573 | /* Tail calls are storing their arguments using |
2574 | arg pointer. If it is a frame pointer on the target, |
2575 | even before reload we need to kill frame pointer based |
2576 | stores. */ |
2577 | || (SIBLING_CALL_P (insn) |
2578 | && HARD_FRAME_POINTER_IS_ARG_POINTER)) |
2579 | insn_info->frame_read = true; |
2580 | |
2581 | /* Loop over the active stores and remove those which are |
2582 | killed by the const function call. */ |
2583 | while (i_ptr) |
2584 | { |
2585 | bool remove_store = false; |
2586 | |
2587 | /* The stack pointer based stores are always killed. */ |
2588 | if (i_ptr->stack_pointer_based) |
2589 | remove_store = true; |
2590 | |
2591 | /* If the frame is read, the frame related stores are killed. */ |
2592 | else if (insn_info->frame_read) |
2593 | { |
2594 | store_info *store_info = i_ptr->store_rec; |
2595 | |
2596 | /* Skip the clobbers. */ |
2597 | while (!store_info->is_set) |
2598 | store_info = store_info->next; |
2599 | |
2600 | if (store_info->group_id >= 0 |
2601 | && rtx_group_vec[store_info->group_id]->frame_related) |
2602 | remove_store = true; |
2603 | } |
2604 | |
2605 | if (remove_store) |
2606 | { |
2607 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2608 | dump_insn_info (start: "removing from active", insn_info: i_ptr); |
2609 | |
2610 | active_local_stores_len--; |
2611 | if (last) |
2612 | last->next_local_store = i_ptr->next_local_store; |
2613 | else |
2614 | active_local_stores = i_ptr->next_local_store; |
2615 | } |
2616 | else |
2617 | last = i_ptr; |
2618 | |
2619 | i_ptr = i_ptr->next_local_store; |
2620 | } |
2621 | |
2622 | if (memset_call) |
2623 | { |
2624 | rtx args[3]; |
2625 | if (get_call_args (call_insn: insn, fn: memset_call, args, nargs: 3) |
2626 | && CONST_INT_P (args[1]) |
2627 | && CONST_INT_P (args[2]) |
2628 | && INTVAL (args[2]) > 0) |
2629 | { |
2630 | rtx mem = gen_rtx_MEM (BLKmode, args[0]); |
2631 | set_mem_size (mem, INTVAL (args[2])); |
2632 | body = gen_rtx_SET (mem, args[1]); |
2633 | mems_found += record_store (body, bb_info); |
2634 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2635 | fprintf (stream: dump_file, format: "handling memset as BLKmode store\n"); |
2636 | if (mems_found == 1) |
2637 | { |
2638 | if (active_local_stores_len++ >= max_active_local_stores) |
2639 | { |
2640 | active_local_stores_len = 1; |
2641 | active_local_stores = NULL; |
2642 | } |
2643 | insn_info->fixed_regs_live |
2644 | = copy_fixed_regs (in: bb_info->regs_live); |
2645 | insn_info->next_local_store = active_local_stores; |
2646 | active_local_stores = insn_info; |
2647 | } |
2648 | } |
2649 | else |
2650 | clear_rhs_from_active_local_stores (); |
2651 | } |
2652 | } |
2653 | else if (SIBLING_CALL_P (insn) |
2654 | && (reload_completed || HARD_FRAME_POINTER_IS_ARG_POINTER)) |
2655 | /* Arguments for a sibling call that are pushed to memory are passed |
2656 | using the incoming argument pointer of the current function. After |
2657 | reload that might be (and likely is) frame pointer based. And, if |
2658 | it is a frame pointer on the target, even before reload we need to |
2659 | kill frame pointer based stores. */ |
2660 | add_wild_read (bb_info); |
2661 | else |
2662 | /* Every other call, including pure functions, may read any memory |
2663 | that is not relative to the frame. */ |
2664 | add_non_frame_wild_read (bb_info); |
2665 | |
2666 | for (rtx link = CALL_INSN_FUNCTION_USAGE (insn); |
2667 | link != NULL_RTX; |
2668 | link = XEXP (link, 1)) |
2669 | if (GET_CODE (XEXP (link, 0)) == USE && MEM_P (XEXP (XEXP (link, 0),0))) |
2670 | check_mem_read_rtx (loc: &XEXP (XEXP (link, 0),0), bb_info, used_in_call: true); |
2671 | |
2672 | return; |
2673 | } |
2674 | |
2675 | /* Assuming that there are sets in these insns, we cannot delete |
2676 | them. */ |
2677 | if ((GET_CODE (PATTERN (insn)) == CLOBBER) |
2678 | || volatile_refs_p (PATTERN (insn)) |
2679 | || (!cfun->can_delete_dead_exceptions && !insn_nothrow_p (insn)) |
2680 | || (RTX_FRAME_RELATED_P (insn)) |
2681 | || find_reg_note (insn, REG_FRAME_RELATED_EXPR, NULL_RTX)) |
2682 | insn_info->cannot_delete = true; |
2683 | |
2684 | body = PATTERN (insn); |
2685 | if (GET_CODE (body) == PARALLEL) |
2686 | { |
2687 | int i; |
2688 | for (i = 0; i < XVECLEN (body, 0); i++) |
2689 | mems_found += record_store (XVECEXP (body, 0, i), bb_info); |
2690 | } |
2691 | else |
2692 | mems_found += record_store (body, bb_info); |
2693 | |
2694 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2695 | fprintf (stream: dump_file, format: "mems_found = %d, cannot_delete = %s\n", |
2696 | mems_found, insn_info->cannot_delete ? "true": "false"); |
2697 | |
2698 | /* If we found some sets of mems, add it into the active_local_stores so |
2699 | that it can be locally deleted if found dead or used for |
2700 | replace_read and redundant constant store elimination. Otherwise mark |
2701 | it as cannot delete. This simplifies the processing later. */ |
2702 | if (mems_found == 1) |
2703 | { |
2704 | if (active_local_stores_len++ >= max_active_local_stores) |
2705 | { |
2706 | active_local_stores_len = 1; |
2707 | active_local_stores = NULL; |
2708 | } |
2709 | insn_info->fixed_regs_live = copy_fixed_regs (in: bb_info->regs_live); |
2710 | insn_info->next_local_store = active_local_stores; |
2711 | active_local_stores = insn_info; |
2712 | } |
2713 | else |
2714 | insn_info->cannot_delete = true; |
2715 | } |
2716 | |
2717 | |
2718 | /* Remove BASE from the set of active_local_stores. This is a |
2719 | callback from cselib that is used to get rid of the stores in |
2720 | active_local_stores. */ |
2721 | |
2722 | static void |
2723 | remove_useless_values (cselib_val *base) |
2724 | { |
2725 | insn_info_t insn_info = active_local_stores; |
2726 | insn_info_t last = NULL; |
2727 | |
2728 | while (insn_info) |
2729 | { |
2730 | store_info *store_info = insn_info->store_rec; |
2731 | bool del = false; |
2732 | |
2733 | /* If ANY of the store_infos match the cselib group that is |
2734 | being deleted, then the insn cannot be deleted. */ |
2735 | while (store_info) |
2736 | { |
2737 | if ((store_info->group_id == -1) |
2738 | && (store_info->cse_base == base)) |
2739 | { |
2740 | del = true; |
2741 | break; |
2742 | } |
2743 | store_info = store_info->next; |
2744 | } |
2745 | |
2746 | if (del) |
2747 | { |
2748 | active_local_stores_len--; |
2749 | if (last) |
2750 | last->next_local_store = insn_info->next_local_store; |
2751 | else |
2752 | active_local_stores = insn_info->next_local_store; |
2753 | free_store_info (insn_info); |
2754 | } |
2755 | else |
2756 | last = insn_info; |
2757 | |
2758 | insn_info = insn_info->next_local_store; |
2759 | } |
2760 | } |
2761 | |
2762 | |
2763 | /* Do all of step 1. */ |
2764 | |
2765 | static void |
2766 | dse_step1 (void) |
2767 | { |
2768 | basic_block bb; |
2769 | bitmap regs_live = BITMAP_ALLOC (obstack: ®_obstack); |
2770 | |
2771 | cselib_init (0); |
2772 | all_blocks = BITMAP_ALLOC (NULL); |
2773 | bitmap_set_bit (all_blocks, ENTRY_BLOCK); |
2774 | bitmap_set_bit (all_blocks, EXIT_BLOCK); |
2775 | |
2776 | /* For -O1 reduce the maximum number of active local stores for RTL DSE |
2777 | since this can consume huge amounts of memory (PR89115). */ |
2778 | int max_active_local_stores = param_max_dse_active_local_stores; |
2779 | if (optimize < 2) |
2780 | max_active_local_stores /= 10; |
2781 | |
2782 | FOR_ALL_BB_FN (bb, cfun) |
2783 | { |
2784 | insn_info_t ptr; |
2785 | bb_info_t bb_info = dse_bb_info_type_pool.allocate (); |
2786 | |
2787 | memset (s: bb_info, c: 0, n: sizeof (dse_bb_info_type)); |
2788 | bitmap_set_bit (all_blocks, bb->index); |
2789 | bb_info->regs_live = regs_live; |
2790 | |
2791 | bitmap_copy (regs_live, DF_LR_IN (bb)); |
2792 | df_simulate_initialize_forwards (bb, regs_live); |
2793 | |
2794 | bb_table[bb->index] = bb_info; |
2795 | cselib_discard_hook = remove_useless_values; |
2796 | |
2797 | if (bb->index >= NUM_FIXED_BLOCKS) |
2798 | { |
2799 | rtx_insn *insn; |
2800 | |
2801 | active_local_stores = NULL; |
2802 | active_local_stores_len = 0; |
2803 | cselib_clear_table (); |
2804 | |
2805 | /* Scan the insns. */ |
2806 | FOR_BB_INSNS (bb, insn) |
2807 | { |
2808 | if (INSN_P (insn)) |
2809 | scan_insn (bb_info, insn, max_active_local_stores); |
2810 | cselib_process_insn (insn); |
2811 | if (INSN_P (insn)) |
2812 | df_simulate_one_insn_forwards (bb, insn, regs_live); |
2813 | } |
2814 | |
2815 | /* This is something of a hack, because the global algorithm |
2816 | is supposed to take care of the case where stores go dead |
2817 | at the end of the function. However, the global |
2818 | algorithm must take a more conservative view of block |
2819 | mode reads than the local alg does. So to get the case |
2820 | where you have a store to the frame followed by a non |
2821 | overlapping block more read, we look at the active local |
2822 | stores at the end of the function and delete all of the |
2823 | frame and spill based ones. */ |
2824 | if (stores_off_frame_dead_at_return |
2825 | && (EDGE_COUNT (bb->succs) == 0 |
2826 | || (single_succ_p (bb) |
2827 | && single_succ (bb) == EXIT_BLOCK_PTR_FOR_FN (cfun) |
2828 | && ! crtl->calls_eh_return))) |
2829 | { |
2830 | insn_info_t i_ptr = active_local_stores; |
2831 | while (i_ptr) |
2832 | { |
2833 | store_info *store_info = i_ptr->store_rec; |
2834 | |
2835 | /* Skip the clobbers. */ |
2836 | while (!store_info->is_set) |
2837 | store_info = store_info->next; |
2838 | if (store_info->group_id >= 0) |
2839 | { |
2840 | group_info *group = rtx_group_vec[store_info->group_id]; |
2841 | if (group->frame_related && !i_ptr->cannot_delete) |
2842 | delete_dead_store_insn (insn_info: i_ptr); |
2843 | } |
2844 | |
2845 | i_ptr = i_ptr->next_local_store; |
2846 | } |
2847 | } |
2848 | |
2849 | /* Get rid of the loads that were discovered in |
2850 | replace_read. Cselib is finished with this block. */ |
2851 | while (deferred_change_list) |
2852 | { |
2853 | deferred_change *next = deferred_change_list->next; |
2854 | |
2855 | /* There is no reason to validate this change. That was |
2856 | done earlier. */ |
2857 | *deferred_change_list->loc = deferred_change_list->reg; |
2858 | deferred_change_pool.remove (object: deferred_change_list); |
2859 | deferred_change_list = next; |
2860 | } |
2861 | |
2862 | /* Get rid of all of the cselib based store_infos in this |
2863 | block and mark the containing insns as not being |
2864 | deletable. */ |
2865 | ptr = bb_info->last_insn; |
2866 | while (ptr) |
2867 | { |
2868 | if (ptr->contains_cselib_groups) |
2869 | { |
2870 | store_info *s_info = ptr->store_rec; |
2871 | while (s_info && !s_info->is_set) |
2872 | s_info = s_info->next; |
2873 | if (s_info |
2874 | && s_info->redundant_reason |
2875 | && s_info->redundant_reason->insn |
2876 | && !ptr->cannot_delete) |
2877 | { |
2878 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2879 | fprintf (stream: dump_file, format: "Locally deleting insn %d " |
2880 | "because insn %d stores the " |
2881 | "same value and couldn't be " |
2882 | "eliminated\n", |
2883 | INSN_UID (insn: ptr->insn), |
2884 | INSN_UID (insn: s_info->redundant_reason->insn)); |
2885 | delete_dead_store_insn (insn_info: ptr); |
2886 | } |
2887 | free_store_info (insn_info: ptr); |
2888 | } |
2889 | else |
2890 | { |
2891 | store_info *s_info; |
2892 | |
2893 | /* Free at least positions_needed bitmaps. */ |
2894 | for (s_info = ptr->store_rec; s_info; s_info = s_info->next) |
2895 | if (s_info->is_large) |
2896 | { |
2897 | BITMAP_FREE (s_info->positions_needed.large.bmap); |
2898 | s_info->is_large = false; |
2899 | } |
2900 | } |
2901 | ptr = ptr->prev_insn; |
2902 | } |
2903 | |
2904 | cse_store_info_pool.release (); |
2905 | } |
2906 | bb_info->regs_live = NULL; |
2907 | } |
2908 | |
2909 | BITMAP_FREE (regs_live); |
2910 | cselib_finish (); |
2911 | rtx_group_table->empty (); |
2912 | } |
2913 | |
2914 | |
2915 | /*---------------------------------------------------------------------------- |
2916 | Second step. |
2917 | |
2918 | Assign each byte position in the stores that we are going to |
2919 | analyze globally to a position in the bitmaps. Returns true if |
2920 | there are any bit positions assigned. |
2921 | ----------------------------------------------------------------------------*/ |
2922 | |
2923 | static void |
2924 | dse_step2_init (void) |
2925 | { |
2926 | unsigned int i; |
2927 | group_info *group; |
2928 | |
2929 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
2930 | { |
2931 | /* For all non stack related bases, we only consider a store to |
2932 | be deletable if there are two or more stores for that |
2933 | position. This is because it takes one store to make the |
2934 | other store redundant. However, for the stores that are |
2935 | stack related, we consider them if there is only one store |
2936 | for the position. We do this because the stack related |
2937 | stores can be deleted if their is no read between them and |
2938 | the end of the function. |
2939 | |
2940 | To make this work in the current framework, we take the stack |
2941 | related bases add all of the bits from store1 into store2. |
2942 | This has the effect of making the eligible even if there is |
2943 | only one store. */ |
2944 | |
2945 | if (stores_off_frame_dead_at_return && group->frame_related) |
2946 | { |
2947 | bitmap_ior_into (group->store2_n, group->store1_n); |
2948 | bitmap_ior_into (group->store2_p, group->store1_p); |
2949 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2950 | fprintf (stream: dump_file, format: "group %d is frame related ", i); |
2951 | } |
2952 | |
2953 | group->offset_map_size_n++; |
2954 | group->offset_map_n = XOBNEWVEC (&dse_obstack, int, |
2955 | group->offset_map_size_n); |
2956 | group->offset_map_size_p++; |
2957 | group->offset_map_p = XOBNEWVEC (&dse_obstack, int, |
2958 | group->offset_map_size_p); |
2959 | group->process_globally = false; |
2960 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2961 | { |
2962 | fprintf (stream: dump_file, format: "group %d(%d+%d): ", i, |
2963 | (int)bitmap_count_bits (group->store2_n), |
2964 | (int)bitmap_count_bits (group->store2_p)); |
2965 | bitmap_print (dump_file, group->store2_n, "n ", " "); |
2966 | bitmap_print (dump_file, group->store2_p, "p ", "\n"); |
2967 | } |
2968 | } |
2969 | } |
2970 | |
2971 | |
2972 | /* Init the offset tables. */ |
2973 | |
2974 | static bool |
2975 | dse_step2 (void) |
2976 | { |
2977 | unsigned int i; |
2978 | group_info *group; |
2979 | /* Position 0 is unused because 0 is used in the maps to mean |
2980 | unused. */ |
2981 | current_position = 1; |
2982 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
2983 | { |
2984 | bitmap_iterator bi; |
2985 | unsigned int j; |
2986 | |
2987 | memset (s: group->offset_map_n, c: 0, n: sizeof (int) * group->offset_map_size_n); |
2988 | memset (s: group->offset_map_p, c: 0, n: sizeof (int) * group->offset_map_size_p); |
2989 | bitmap_clear (group->group_kill); |
2990 | |
2991 | EXECUTE_IF_SET_IN_BITMAP (group->store2_n, 0, j, bi) |
2992 | { |
2993 | bitmap_set_bit (group->group_kill, current_position); |
2994 | if (bitmap_bit_p (group->escaped_n, j)) |
2995 | bitmap_set_bit (kill_on_calls, current_position); |
2996 | group->offset_map_n[j] = current_position++; |
2997 | group->process_globally = true; |
2998 | } |
2999 | EXECUTE_IF_SET_IN_BITMAP (group->store2_p, 0, j, bi) |
3000 | { |
3001 | bitmap_set_bit (group->group_kill, current_position); |
3002 | if (bitmap_bit_p (group->escaped_p, j)) |
3003 | bitmap_set_bit (kill_on_calls, current_position); |
3004 | group->offset_map_p[j] = current_position++; |
3005 | group->process_globally = true; |
3006 | } |
3007 | } |
3008 | return current_position != 1; |
3009 | } |
3010 | |
3011 | |
3012 | |
3013 | /*---------------------------------------------------------------------------- |
3014 | Third step. |
3015 | |
3016 | Build the bit vectors for the transfer functions. |
3017 | ----------------------------------------------------------------------------*/ |
3018 | |
3019 | |
3020 | /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not |
3021 | there, return 0. */ |
3022 | |
3023 | static int |
3024 | get_bitmap_index (group_info *group_info, HOST_WIDE_INT offset) |
3025 | { |
3026 | if (offset < 0) |
3027 | { |
3028 | HOST_WIDE_INT offset_p = -offset; |
3029 | if (offset_p >= group_info->offset_map_size_n) |
3030 | return 0; |
3031 | return group_info->offset_map_n[offset_p]; |
3032 | } |
3033 | else |
3034 | { |
3035 | if (offset >= group_info->offset_map_size_p) |
3036 | return 0; |
3037 | return group_info->offset_map_p[offset]; |
3038 | } |
3039 | } |
3040 | |
3041 | |
3042 | /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL |
3043 | may be NULL. */ |
3044 | |
3045 | static void |
3046 | scan_stores (store_info *store_info, bitmap gen, bitmap kill) |
3047 | { |
3048 | while (store_info) |
3049 | { |
3050 | HOST_WIDE_INT i, offset, width; |
3051 | group_info *group_info |
3052 | = rtx_group_vec[store_info->group_id]; |
3053 | /* We can (conservatively) ignore stores whose bounds aren't known; |
3054 | they simply don't generate new global dse opportunities. */ |
3055 | if (group_info->process_globally |
3056 | && store_info->offset.is_constant (const_value: &offset) |
3057 | && store_info->width.is_constant (const_value: &width)) |
3058 | { |
3059 | HOST_WIDE_INT end = offset + width; |
3060 | for (i = offset; i < end; i++) |
3061 | { |
3062 | int index = get_bitmap_index (group_info, offset: i); |
3063 | if (index != 0) |
3064 | { |
3065 | bitmap_set_bit (gen, index); |
3066 | if (kill) |
3067 | bitmap_clear_bit (kill, index); |
3068 | } |
3069 | } |
3070 | } |
3071 | store_info = store_info->next; |
3072 | } |
3073 | } |
3074 | |
3075 | |
3076 | /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL |
3077 | may be NULL. */ |
3078 | |
3079 | static void |
3080 | scan_reads (insn_info_t insn_info, bitmap gen, bitmap kill) |
3081 | { |
3082 | read_info_t read_info = insn_info->read_rec; |
3083 | int i; |
3084 | group_info *group; |
3085 | |
3086 | /* If this insn reads the frame, kill all the frame related stores. */ |
3087 | if (insn_info->frame_read) |
3088 | { |
3089 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3090 | if (group->process_globally && group->frame_related) |
3091 | { |
3092 | if (kill) |
3093 | bitmap_ior_into (kill, group->group_kill); |
3094 | bitmap_and_compl_into (gen, group->group_kill); |
3095 | } |
3096 | } |
3097 | if (insn_info->non_frame_wild_read) |
3098 | { |
3099 | /* Kill all non-frame related stores. Kill all stores of variables that |
3100 | escape. */ |
3101 | if (kill) |
3102 | bitmap_ior_into (kill, kill_on_calls); |
3103 | bitmap_and_compl_into (gen, kill_on_calls); |
3104 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3105 | if (group->process_globally && !group->frame_related) |
3106 | { |
3107 | if (kill) |
3108 | bitmap_ior_into (kill, group->group_kill); |
3109 | bitmap_and_compl_into (gen, group->group_kill); |
3110 | } |
3111 | } |
3112 | while (read_info) |
3113 | { |
3114 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3115 | { |
3116 | if (group->process_globally) |
3117 | { |
3118 | if (i == read_info->group_id) |
3119 | { |
3120 | HOST_WIDE_INT offset, width; |
3121 | /* Reads with non-constant size kill all DSE opportunities |
3122 | in the group. */ |
3123 | if (!read_info->offset.is_constant (const_value: &offset) |
3124 | || !read_info->width.is_constant (const_value: &width) |
3125 | || !known_size_p (a: width)) |
3126 | { |
3127 | /* Handle block mode reads. */ |
3128 | if (kill) |
3129 | bitmap_ior_into (kill, group->group_kill); |
3130 | bitmap_and_compl_into (gen, group->group_kill); |
3131 | } |
3132 | else |
3133 | { |
3134 | /* The groups are the same, just process the |
3135 | offsets. */ |
3136 | HOST_WIDE_INT j; |
3137 | HOST_WIDE_INT end = offset + width; |
3138 | for (j = offset; j < end; j++) |
3139 | { |
3140 | int index = get_bitmap_index (group_info: group, offset: j); |
3141 | if (index != 0) |
3142 | { |
3143 | if (kill) |
3144 | bitmap_set_bit (kill, index); |
3145 | bitmap_clear_bit (gen, index); |
3146 | } |
3147 | } |
3148 | } |
3149 | } |
3150 | else |
3151 | { |
3152 | /* The groups are different, if the alias sets |
3153 | conflict, clear the entire group. We only need |
3154 | to apply this test if the read_info is a cselib |
3155 | read. Anything with a constant base cannot alias |
3156 | something else with a different constant |
3157 | base. */ |
3158 | if ((read_info->group_id < 0) |
3159 | && canon_true_dependence (group->base_mem, |
3160 | GET_MODE (group->base_mem), |
3161 | group->canon_base_addr, |
3162 | read_info->mem, NULL_RTX)) |
3163 | { |
3164 | if (kill) |
3165 | bitmap_ior_into (kill, group->group_kill); |
3166 | bitmap_and_compl_into (gen, group->group_kill); |
3167 | } |
3168 | } |
3169 | } |
3170 | } |
3171 | |
3172 | read_info = read_info->next; |
3173 | } |
3174 | } |
3175 | |
3176 | |
3177 | /* Return the insn in BB_INFO before the first wild read or if there |
3178 | are no wild reads in the block, return the last insn. */ |
3179 | |
3180 | static insn_info_t |
3181 | find_insn_before_first_wild_read (bb_info_t bb_info) |
3182 | { |
3183 | insn_info_t insn_info = bb_info->last_insn; |
3184 | insn_info_t last_wild_read = NULL; |
3185 | |
3186 | while (insn_info) |
3187 | { |
3188 | if (insn_info->wild_read) |
3189 | { |
3190 | last_wild_read = insn_info->prev_insn; |
3191 | /* Block starts with wild read. */ |
3192 | if (!last_wild_read) |
3193 | return NULL; |
3194 | } |
3195 | |
3196 | insn_info = insn_info->prev_insn; |
3197 | } |
3198 | |
3199 | if (last_wild_read) |
3200 | return last_wild_read; |
3201 | else |
3202 | return bb_info->last_insn; |
3203 | } |
3204 | |
3205 | |
3206 | /* Scan the insns in BB_INFO starting at PTR and going to the top of |
3207 | the block in order to build the gen and kill sets for the block. |
3208 | We start at ptr which may be the last insn in the block or may be |
3209 | the first insn with a wild read. In the latter case we are able to |
3210 | skip the rest of the block because it just does not matter: |
3211 | anything that happens is hidden by the wild read. */ |
3212 | |
3213 | static void |
3214 | dse_step3_scan (basic_block bb) |
3215 | { |
3216 | bb_info_t bb_info = bb_table[bb->index]; |
3217 | insn_info_t insn_info; |
3218 | |
3219 | insn_info = find_insn_before_first_wild_read (bb_info); |
3220 | |
3221 | /* In the spill case or in the no_spill case if there is no wild |
3222 | read in the block, we will need a kill set. */ |
3223 | if (insn_info == bb_info->last_insn) |
3224 | { |
3225 | if (bb_info->kill) |
3226 | bitmap_clear (bb_info->kill); |
3227 | else |
3228 | bb_info->kill = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
3229 | } |
3230 | else |
3231 | if (bb_info->kill) |
3232 | BITMAP_FREE (bb_info->kill); |
3233 | |
3234 | while (insn_info) |
3235 | { |
3236 | /* There may have been code deleted by the dce pass run before |
3237 | this phase. */ |
3238 | if (insn_info->insn && INSN_P (insn_info->insn)) |
3239 | { |
3240 | scan_stores (store_info: insn_info->store_rec, gen: bb_info->gen, kill: bb_info->kill); |
3241 | scan_reads (insn_info, gen: bb_info->gen, kill: bb_info->kill); |
3242 | } |
3243 | |
3244 | insn_info = insn_info->prev_insn; |
3245 | } |
3246 | } |
3247 | |
3248 | |
3249 | /* Set the gen set of the exit block, and also any block with no |
3250 | successors that does not have a wild read. */ |
3251 | |
3252 | static void |
3253 | dse_step3_exit_block_scan (bb_info_t bb_info) |
3254 | { |
3255 | /* The gen set is all 0's for the exit block except for the |
3256 | frame_pointer_group. */ |
3257 | |
3258 | if (stores_off_frame_dead_at_return) |
3259 | { |
3260 | unsigned int i; |
3261 | group_info *group; |
3262 | |
3263 | FOR_EACH_VEC_ELT (rtx_group_vec, i, group) |
3264 | { |
3265 | if (group->process_globally && group->frame_related) |
3266 | bitmap_ior_into (bb_info->gen, group->group_kill); |
3267 | } |
3268 | } |
3269 | } |
3270 | |
3271 | |
3272 | /* Find all of the blocks that are not backwards reachable from the |
3273 | exit block or any block with no successors (BB). These are the |
3274 | infinite loops or infinite self loops. These blocks will still |
3275 | have their bits set in UNREACHABLE_BLOCKS. */ |
3276 | |
3277 | static void |
3278 | mark_reachable_blocks (sbitmap unreachable_blocks, basic_block bb) |
3279 | { |
3280 | edge e; |
3281 | edge_iterator ei; |
3282 | |
3283 | if (bitmap_bit_p (map: unreachable_blocks, bitno: bb->index)) |
3284 | { |
3285 | bitmap_clear_bit (map: unreachable_blocks, bitno: bb->index); |
3286 | FOR_EACH_EDGE (e, ei, bb->preds) |
3287 | { |
3288 | mark_reachable_blocks (unreachable_blocks, bb: e->src); |
3289 | } |
3290 | } |
3291 | } |
3292 | |
3293 | /* Build the transfer functions for the function. */ |
3294 | |
3295 | static void |
3296 | dse_step3 () |
3297 | { |
3298 | basic_block bb; |
3299 | sbitmap_iterator sbi; |
3300 | bitmap all_ones = NULL; |
3301 | unsigned int i; |
3302 | |
3303 | auto_sbitmap unreachable_blocks (last_basic_block_for_fn (cfun)); |
3304 | bitmap_ones (unreachable_blocks); |
3305 | |
3306 | FOR_ALL_BB_FN (bb, cfun) |
3307 | { |
3308 | bb_info_t bb_info = bb_table[bb->index]; |
3309 | if (bb_info->gen) |
3310 | bitmap_clear (bb_info->gen); |
3311 | else |
3312 | bb_info->gen = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
3313 | |
3314 | if (bb->index == ENTRY_BLOCK) |
3315 | ; |
3316 | else if (bb->index == EXIT_BLOCK) |
3317 | dse_step3_exit_block_scan (bb_info); |
3318 | else |
3319 | dse_step3_scan (bb); |
3320 | if (EDGE_COUNT (bb->succs) == 0) |
3321 | mark_reachable_blocks (unreachable_blocks, bb); |
3322 | |
3323 | /* If this is the second time dataflow is run, delete the old |
3324 | sets. */ |
3325 | if (bb_info->in) |
3326 | BITMAP_FREE (bb_info->in); |
3327 | if (bb_info->out) |
3328 | BITMAP_FREE (bb_info->out); |
3329 | } |
3330 | |
3331 | /* For any block in an infinite loop, we must initialize the out set |
3332 | to all ones. This could be expensive, but almost never occurs in |
3333 | practice. However, it is common in regression tests. */ |
3334 | EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks, 0, i, sbi) |
3335 | { |
3336 | if (bitmap_bit_p (all_blocks, i)) |
3337 | { |
3338 | bb_info_t bb_info = bb_table[i]; |
3339 | if (!all_ones) |
3340 | { |
3341 | unsigned int j; |
3342 | group_info *group; |
3343 | |
3344 | all_ones = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
3345 | FOR_EACH_VEC_ELT (rtx_group_vec, j, group) |
3346 | bitmap_ior_into (all_ones, group->group_kill); |
3347 | } |
3348 | if (!bb_info->out) |
3349 | { |
3350 | bb_info->out = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
3351 | bitmap_copy (bb_info->out, all_ones); |
3352 | } |
3353 | } |
3354 | } |
3355 | |
3356 | if (all_ones) |
3357 | BITMAP_FREE (all_ones); |
3358 | } |
3359 | |
3360 | |
3361 | |
3362 | /*---------------------------------------------------------------------------- |
3363 | Fourth step. |
3364 | |
3365 | Solve the bitvector equations. |
3366 | ----------------------------------------------------------------------------*/ |
3367 | |
3368 | |
3369 | /* Confluence function for blocks with no successors. Create an out |
3370 | set from the gen set of the exit block. This block logically has |
3371 | the exit block as a successor. */ |
3372 | |
3373 | |
3374 | |
3375 | static void |
3376 | dse_confluence_0 (basic_block bb) |
3377 | { |
3378 | bb_info_t bb_info = bb_table[bb->index]; |
3379 | |
3380 | if (bb->index == EXIT_BLOCK) |
3381 | return; |
3382 | |
3383 | if (!bb_info->out) |
3384 | { |
3385 | bb_info->out = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
3386 | bitmap_copy (bb_info->out, bb_table[EXIT_BLOCK]->gen); |
3387 | } |
3388 | } |
3389 | |
3390 | /* Propagate the information from the in set of the dest of E to the |
3391 | out set of the src of E. If the various in or out sets are not |
3392 | there, that means they are all ones. */ |
3393 | |
3394 | static bool |
3395 | dse_confluence_n (edge e) |
3396 | { |
3397 | bb_info_t src_info = bb_table[e->src->index]; |
3398 | bb_info_t dest_info = bb_table[e->dest->index]; |
3399 | |
3400 | if (dest_info->in) |
3401 | { |
3402 | if (src_info->out) |
3403 | bitmap_and_into (src_info->out, dest_info->in); |
3404 | else |
3405 | { |
3406 | src_info->out = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
3407 | bitmap_copy (src_info->out, dest_info->in); |
3408 | } |
3409 | } |
3410 | return true; |
3411 | } |
3412 | |
3413 | |
3414 | /* Propagate the info from the out to the in set of BB_INDEX's basic |
3415 | block. There are three cases: |
3416 | |
3417 | 1) The block has no kill set. In this case the kill set is all |
3418 | ones. It does not matter what the out set of the block is, none of |
3419 | the info can reach the top. The only thing that reaches the top is |
3420 | the gen set and we just copy the set. |
3421 | |
3422 | 2) There is a kill set but no out set and bb has successors. In |
3423 | this case we just return. Eventually an out set will be created and |
3424 | it is better to wait than to create a set of ones. |
3425 | |
3426 | 3) There is both a kill and out set. We apply the obvious transfer |
3427 | function. |
3428 | */ |
3429 | |
3430 | static bool |
3431 | dse_transfer_function (int bb_index) |
3432 | { |
3433 | bb_info_t bb_info = bb_table[bb_index]; |
3434 | |
3435 | if (bb_info->kill) |
3436 | { |
3437 | if (bb_info->out) |
3438 | { |
3439 | /* Case 3 above. */ |
3440 | if (bb_info->in) |
3441 | return bitmap_ior_and_compl (DST: bb_info->in, A: bb_info->gen, |
3442 | B: bb_info->out, C: bb_info->kill); |
3443 | else |
3444 | { |
3445 | bb_info->in = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
3446 | bitmap_ior_and_compl (DST: bb_info->in, A: bb_info->gen, |
3447 | B: bb_info->out, C: bb_info->kill); |
3448 | return true; |
3449 | } |
3450 | } |
3451 | else |
3452 | /* Case 2 above. */ |
3453 | return false; |
3454 | } |
3455 | else |
3456 | { |
3457 | /* Case 1 above. If there is already an in set, nothing |
3458 | happens. */ |
3459 | if (bb_info->in) |
3460 | return false; |
3461 | else |
3462 | { |
3463 | bb_info->in = BITMAP_ALLOC (obstack: &dse_bitmap_obstack); |
3464 | bitmap_copy (bb_info->in, bb_info->gen); |
3465 | return true; |
3466 | } |
3467 | } |
3468 | } |
3469 | |
3470 | /* Solve the dataflow equations. */ |
3471 | |
3472 | static void |
3473 | dse_step4 (void) |
3474 | { |
3475 | df_simple_dataflow (DF_BACKWARD, NULL, dse_confluence_0, |
3476 | dse_confluence_n, dse_transfer_function, |
3477 | all_blocks, df_get_postorder (DF_BACKWARD), |
3478 | df_get_n_blocks (DF_BACKWARD)); |
3479 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3480 | { |
3481 | basic_block bb; |
3482 | |
3483 | fprintf (stream: dump_file, format: "\n\n*** Global dataflow info after analysis.\n"); |
3484 | FOR_ALL_BB_FN (bb, cfun) |
3485 | { |
3486 | bb_info_t bb_info = bb_table[bb->index]; |
3487 | |
3488 | df_print_bb_index (bb, file: dump_file); |
3489 | if (bb_info->in) |
3490 | bitmap_print (dump_file, bb_info->in, " in: ", "\n"); |
3491 | else |
3492 | fprintf (stream: dump_file, format: " in: *MISSING*\n"); |
3493 | if (bb_info->gen) |
3494 | bitmap_print (dump_file, bb_info->gen, " gen: ", "\n"); |
3495 | else |
3496 | fprintf (stream: dump_file, format: " gen: *MISSING*\n"); |
3497 | if (bb_info->kill) |
3498 | bitmap_print (dump_file, bb_info->kill, " kill: ", "\n"); |
3499 | else |
3500 | fprintf (stream: dump_file, format: " kill: *MISSING*\n"); |
3501 | if (bb_info->out) |
3502 | bitmap_print (dump_file, bb_info->out, " out: ", "\n"); |
3503 | else |
3504 | fprintf (stream: dump_file, format: " out: *MISSING*\n\n"); |
3505 | } |
3506 | } |
3507 | } |
3508 | |
3509 | |
3510 | |
3511 | /*---------------------------------------------------------------------------- |
3512 | Fifth step. |
3513 | |
3514 | Delete the stores that can only be deleted using the global information. |
3515 | ----------------------------------------------------------------------------*/ |
3516 | |
3517 | |
3518 | static void |
3519 | dse_step5 (void) |
3520 | { |
3521 | basic_block bb; |
3522 | FOR_EACH_BB_FN (bb, cfun) |
3523 | { |
3524 | bb_info_t bb_info = bb_table[bb->index]; |
3525 | insn_info_t insn_info = bb_info->last_insn; |
3526 | bitmap v = bb_info->out; |
3527 | |
3528 | while (insn_info) |
3529 | { |
3530 | bool deleted = false; |
3531 | if (dump_file && insn_info->insn) |
3532 | { |
3533 | fprintf (stream: dump_file, format: "starting to process insn %d\n", |
3534 | INSN_UID (insn: insn_info->insn)); |
3535 | bitmap_print (dump_file, v, " v: ", "\n"); |
3536 | } |
3537 | |
3538 | /* There may have been code deleted by the dce pass run before |
3539 | this phase. */ |
3540 | if (insn_info->insn |
3541 | && INSN_P (insn_info->insn) |
3542 | && (!insn_info->cannot_delete) |
3543 | && (!bitmap_empty_p (map: v))) |
3544 | { |
3545 | store_info *store_info = insn_info->store_rec; |
3546 | |
3547 | /* Try to delete the current insn. */ |
3548 | deleted = true; |
3549 | |
3550 | /* Skip the clobbers. */ |
3551 | while (!store_info->is_set) |
3552 | store_info = store_info->next; |
3553 | |
3554 | HOST_WIDE_INT i, offset, width; |
3555 | group_info *group_info = rtx_group_vec[store_info->group_id]; |
3556 | |
3557 | if (!store_info->offset.is_constant (const_value: &offset) |
3558 | || !store_info->width.is_constant (const_value: &width)) |
3559 | deleted = false; |
3560 | else |
3561 | { |
3562 | HOST_WIDE_INT end = offset + width; |
3563 | for (i = offset; i < end; i++) |
3564 | { |
3565 | int index = get_bitmap_index (group_info, offset: i); |
3566 | |
3567 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3568 | fprintf (stream: dump_file, format: "i = %d, index = %d\n", |
3569 | (int) i, index); |
3570 | if (index == 0 || !bitmap_bit_p (v, index)) |
3571 | { |
3572 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3573 | fprintf (stream: dump_file, format: "failing at i = %d\n", |
3574 | (int) i); |
3575 | deleted = false; |
3576 | break; |
3577 | } |
3578 | } |
3579 | } |
3580 | if (deleted) |
3581 | { |
3582 | if (dbg_cnt (index: dse) |
3583 | && check_for_inc_dec_1 (insn_info)) |
3584 | { |
3585 | delete_insn (insn_info->insn); |
3586 | insn_info->insn = NULL; |
3587 | globally_deleted++; |
3588 | } |
3589 | } |
3590 | } |
3591 | /* We do want to process the local info if the insn was |
3592 | deleted. For instance, if the insn did a wild read, we |
3593 | no longer need to trash the info. */ |
3594 | if (insn_info->insn |
3595 | && INSN_P (insn_info->insn) |
3596 | && (!deleted)) |
3597 | { |
3598 | scan_stores (store_info: insn_info->store_rec, gen: v, NULL); |
3599 | if (insn_info->wild_read) |
3600 | { |
3601 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3602 | fprintf (stream: dump_file, format: "wild read\n"); |
3603 | bitmap_clear (v); |
3604 | } |
3605 | else if (insn_info->read_rec |
3606 | || insn_info->non_frame_wild_read |
3607 | || insn_info->frame_read) |
3608 | { |
3609 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3610 | { |
3611 | if (!insn_info->non_frame_wild_read |
3612 | && !insn_info->frame_read) |
3613 | fprintf (stream: dump_file, format: "regular read\n"); |
3614 | if (insn_info->non_frame_wild_read) |
3615 | fprintf (stream: dump_file, format: "non-frame wild read\n"); |
3616 | if (insn_info->frame_read) |
3617 | fprintf (stream: dump_file, format: "frame read\n"); |
3618 | } |
3619 | scan_reads (insn_info, gen: v, NULL); |
3620 | } |
3621 | } |
3622 | |
3623 | insn_info = insn_info->prev_insn; |
3624 | } |
3625 | } |
3626 | } |
3627 | |
3628 | |
3629 | |
3630 | /*---------------------------------------------------------------------------- |
3631 | Sixth step. |
3632 | |
3633 | Delete stores made redundant by earlier stores (which store the same |
3634 | value) that couldn't be eliminated. |
3635 | ----------------------------------------------------------------------------*/ |
3636 | |
3637 | static void |
3638 | dse_step6 (void) |
3639 | { |
3640 | basic_block bb; |
3641 | |
3642 | FOR_ALL_BB_FN (bb, cfun) |
3643 | { |
3644 | bb_info_t bb_info = bb_table[bb->index]; |
3645 | insn_info_t insn_info = bb_info->last_insn; |
3646 | |
3647 | while (insn_info) |
3648 | { |
3649 | /* There may have been code deleted by the dce pass run before |
3650 | this phase. */ |
3651 | if (insn_info->insn |
3652 | && INSN_P (insn_info->insn) |
3653 | && !insn_info->cannot_delete) |
3654 | { |
3655 | store_info *s_info = insn_info->store_rec; |
3656 | |
3657 | while (s_info && !s_info->is_set) |
3658 | s_info = s_info->next; |
3659 | if (s_info |
3660 | && s_info->redundant_reason |
3661 | && s_info->redundant_reason->insn |
3662 | && INSN_P (s_info->redundant_reason->insn)) |
3663 | { |
3664 | rtx_insn *rinsn = s_info->redundant_reason->insn; |
3665 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3666 | fprintf (stream: dump_file, format: "Locally deleting insn %d " |
3667 | "because insn %d stores the " |
3668 | "same value and couldn't be " |
3669 | "eliminated\n", |
3670 | INSN_UID (insn: insn_info->insn), |
3671 | INSN_UID (insn: rinsn)); |
3672 | delete_dead_store_insn (insn_info); |
3673 | } |
3674 | } |
3675 | insn_info = insn_info->prev_insn; |
3676 | } |
3677 | } |
3678 | } |
3679 | |
3680 | /*---------------------------------------------------------------------------- |
3681 | Seventh step. |
3682 | |
3683 | Destroy everything left standing. |
3684 | ----------------------------------------------------------------------------*/ |
3685 | |
3686 | static void |
3687 | dse_step7 (void) |
3688 | { |
3689 | bitmap_obstack_release (&dse_bitmap_obstack); |
3690 | obstack_free (&dse_obstack, NULL); |
3691 | |
3692 | end_alias_analysis (); |
3693 | free (ptr: bb_table); |
3694 | delete rtx_group_table; |
3695 | rtx_group_table = NULL; |
3696 | rtx_group_vec.release (); |
3697 | BITMAP_FREE (all_blocks); |
3698 | BITMAP_FREE (scratch); |
3699 | |
3700 | rtx_store_info_pool.release (); |
3701 | read_info_type_pool.release (); |
3702 | insn_info_type_pool.release (); |
3703 | dse_bb_info_type_pool.release (); |
3704 | group_info_pool.release (); |
3705 | deferred_change_pool.release (); |
3706 | } |
3707 | |
3708 | |
3709 | /* ------------------------------------------------------------------------- |
3710 | DSE |
3711 | ------------------------------------------------------------------------- */ |
3712 | |
3713 | /* Callback for running pass_rtl_dse. */ |
3714 | |
3715 | static unsigned int |
3716 | rest_of_handle_dse (void) |
3717 | { |
3718 | df_set_flags (DF_DEFER_INSN_RESCAN); |
3719 | |
3720 | /* Need the notes since we must track live hardregs in the forwards |
3721 | direction. */ |
3722 | df_note_add_problem (); |
3723 | df_analyze (); |
3724 | |
3725 | dse_step0 (); |
3726 | dse_step1 (); |
3727 | /* DSE can eliminate potentially-trapping MEMs. |
3728 | Remove any EH edges associated with them, since otherwise |
3729 | DF_LR_RUN_DCE will complain later. */ |
3730 | if ((locally_deleted || globally_deleted) |
3731 | && cfun->can_throw_non_call_exceptions |
3732 | && purge_all_dead_edges ()) |
3733 | { |
3734 | free_dominance_info (CDI_DOMINATORS); |
3735 | delete_unreachable_blocks (); |
3736 | } |
3737 | dse_step2_init (); |
3738 | if (dse_step2 ()) |
3739 | { |
3740 | df_set_flags (DF_LR_RUN_DCE); |
3741 | df_analyze (); |
3742 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3743 | fprintf (stream: dump_file, format: "doing global processing\n"); |
3744 | dse_step3 (); |
3745 | dse_step4 (); |
3746 | dse_step5 (); |
3747 | } |
3748 | |
3749 | dse_step6 (); |
3750 | dse_step7 (); |
3751 | |
3752 | if (dump_file) |
3753 | fprintf (stream: dump_file, format: "dse: local deletions = %d, global deletions = %d\n", |
3754 | locally_deleted, globally_deleted); |
3755 | |
3756 | /* DSE can eliminate potentially-trapping MEMs. |
3757 | Remove any EH edges associated with them. */ |
3758 | if ((locally_deleted || globally_deleted) |
3759 | && cfun->can_throw_non_call_exceptions |
3760 | && purge_all_dead_edges ()) |
3761 | { |
3762 | free_dominance_info (CDI_DOMINATORS); |
3763 | cleanup_cfg (0); |
3764 | } |
3765 | |
3766 | return 0; |
3767 | } |
3768 | |
3769 | namespace { |
3770 | |
3771 | const pass_data pass_data_rtl_dse1 = |
3772 | { |
3773 | .type: RTL_PASS, /* type */ |
3774 | .name: "dse1", /* name */ |
3775 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
3776 | .tv_id: TV_DSE1, /* tv_id */ |
3777 | .properties_required: 0, /* properties_required */ |
3778 | .properties_provided: 0, /* properties_provided */ |
3779 | .properties_destroyed: 0, /* properties_destroyed */ |
3780 | .todo_flags_start: 0, /* todo_flags_start */ |
3781 | TODO_df_finish, /* todo_flags_finish */ |
3782 | }; |
3783 | |
3784 | class pass_rtl_dse1 : public rtl_opt_pass |
3785 | { |
3786 | public: |
3787 | pass_rtl_dse1 (gcc::context *ctxt) |
3788 | : rtl_opt_pass (pass_data_rtl_dse1, ctxt) |
3789 | {} |
3790 | |
3791 | /* opt_pass methods: */ |
3792 | bool gate (function *) final override |
3793 | { |
3794 | return optimize > 0 && flag_dse && dbg_cnt (index: dse1); |
3795 | } |
3796 | |
3797 | unsigned int execute (function *) final override |
3798 | { |
3799 | return rest_of_handle_dse (); |
3800 | } |
3801 | |
3802 | }; // class pass_rtl_dse1 |
3803 | |
3804 | } // anon namespace |
3805 | |
3806 | rtl_opt_pass * |
3807 | make_pass_rtl_dse1 (gcc::context *ctxt) |
3808 | { |
3809 | return new pass_rtl_dse1 (ctxt); |
3810 | } |
3811 | |
3812 | namespace { |
3813 | |
3814 | const pass_data pass_data_rtl_dse2 = |
3815 | { |
3816 | .type: RTL_PASS, /* type */ |
3817 | .name: "dse2", /* name */ |
3818 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
3819 | .tv_id: TV_DSE2, /* tv_id */ |
3820 | .properties_required: 0, /* properties_required */ |
3821 | .properties_provided: 0, /* properties_provided */ |
3822 | .properties_destroyed: 0, /* properties_destroyed */ |
3823 | .todo_flags_start: 0, /* todo_flags_start */ |
3824 | TODO_df_finish, /* todo_flags_finish */ |
3825 | }; |
3826 | |
3827 | class pass_rtl_dse2 : public rtl_opt_pass |
3828 | { |
3829 | public: |
3830 | pass_rtl_dse2 (gcc::context *ctxt) |
3831 | : rtl_opt_pass (pass_data_rtl_dse2, ctxt) |
3832 | {} |
3833 | |
3834 | /* opt_pass methods: */ |
3835 | bool gate (function *) final override |
3836 | { |
3837 | return optimize > 0 && flag_dse && dbg_cnt (index: dse2); |
3838 | } |
3839 | |
3840 | unsigned int execute (function *) final override |
3841 | { |
3842 | return rest_of_handle_dse (); |
3843 | } |
3844 | |
3845 | }; // class pass_rtl_dse2 |
3846 | |
3847 | } // anon namespace |
3848 | |
3849 | rtl_opt_pass * |
3850 | make_pass_rtl_dse2 (gcc::context *ctxt) |
3851 | { |
3852 | return new pass_rtl_dse2 (ctxt); |
3853 | } |
3854 |
Definitions
- dse_bitmap_obstack
- dse_obstack
- scratch
- store_info
- lowpart_bitmask
- cse_store_info_pool
- rtx_store_info_pool
- read_info_type
- read_info_type_pool
- insn_info_type
- insn_info_type_pool
- active_local_stores
- active_local_stores_len
- dse_bb_info_type
- dse_bb_info_type_pool
- bb_table
- group_info
- group_info_pool
- rtx_group_next_id
- rtx_group_vec
- deferred_change
- deferred_change_pool
- deferred_change_list
- stores_off_frame_dead_at_return
- globally_deleted
- locally_deleted
- all_blocks
- kill_on_calls
- current_position
- print_range
- invariant_group_base_hasher
- equal
- hash
- rtx_group_table
- get_group_info
- dse_step0
- free_store_info
- note_add_store_info
- note_add_store
- emit_inc_dec_insn_before
- check_for_inc_dec_1
- check_for_inc_dec
- delete_dead_store_insn
- local_variable_can_escape
- can_escape
- set_usage_bits
- reset_active_stores
- free_read_records
- add_wild_read
- add_non_frame_wild_read
- const_or_frame_p
- canon_address
- clear_rhs_from_active_local_stores
- set_position_unneeded
- set_all_positions_unneeded
- any_positions_needed_p
- all_positions_needed_p
- record_store
- dump_insn_info
- find_shift_sequence
- look_for_hardregs
- get_stored_val
- replace_read
- check_mem_read_rtx
- check_mem_read_use
- get_call_args
- copy_fixed_regs
- scan_insn
- remove_useless_values
- dse_step1
- dse_step2_init
- dse_step2
- get_bitmap_index
- scan_stores
- scan_reads
- find_insn_before_first_wild_read
- dse_step3_scan
- dse_step3_exit_block_scan
- mark_reachable_blocks
- dse_step3
- dse_confluence_0
- dse_confluence_n
- dse_transfer_function
- dse_step4
- dse_step5
- dse_step6
- dse_step7
- rest_of_handle_dse
- pass_data_rtl_dse1
- pass_rtl_dse1
- pass_rtl_dse1
- gate
- execute
- make_pass_rtl_dse1
- pass_data_rtl_dse2
- pass_rtl_dse2
- pass_rtl_dse2
- gate
- execute
Learn to use CMake with our Intro Training
Find out more