1 /* RTL dead store elimination.
2 Copyright (C) 2005-2015 Free Software Foundation, Inc.
4 Contributed by Richard Sandiford <rsandifor@codesourcery.com>
5 and Kenneth Zadeck <zadeck@naturalbridge.com>
7 This file is part of GCC.
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
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
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/>. */
27 #include "coretypes.h"
28 #include "hash-table.h"
34 #include "double-int.h"
42 #include "fold-const.h"
43 #include "stor-layout.h"
46 #include "hard-reg-set.h"
49 #include "dominance.h"
53 #include "basic-block.h"
56 #include "tree-pass.h"
57 #include "alloc-pool.h"
58 #include "insn-config.h"
61 #include "statistics.h"
62 #include "fixed-value.h"
72 #include "insn-codes.h"
77 #include "tree-ssa-alias.h"
78 #include "internal-fn.h"
79 #include "gimple-expr.h"
82 #include "gimple-ssa.h"
84 #include "cfgcleanup.h"
86 /* This file contains three techniques for performing Dead Store
89 * The first technique performs dse locally on any base address. It
90 is based on the cselib which is a local value numbering technique.
91 This technique is local to a basic block but deals with a fairly
94 * The second technique performs dse globally but is restricted to
95 base addresses that are either constant or are relative to the
98 * The third technique, (which is only done after register allocation)
99 processes the spill spill slots. This differs from the second
100 technique because it takes advantage of the fact that spilling is
101 completely free from the effects of aliasing.
103 Logically, dse is a backwards dataflow problem. A store can be
104 deleted if it if cannot be reached in the backward direction by any
105 use of the value being stored. However, the local technique uses a
106 forwards scan of the basic block because cselib requires that the
107 block be processed in that order.
109 The pass is logically broken into 7 steps:
113 1) The local algorithm, as well as scanning the insns for the two
116 2) Analysis to see if the global algs are necessary. In the case
117 of stores base on a constant address, there must be at least two
118 stores to that address, to make it possible to delete some of the
119 stores. In the case of stores off of the frame or spill related
120 stores, only one store to an address is necessary because those
121 stores die at the end of the function.
123 3) Set up the global dataflow equations based on processing the
124 info parsed in the first step.
126 4) Solve the dataflow equations.
128 5) Delete the insns that the global analysis has indicated are
131 6) Delete insns that store the same value as preceding store
132 where the earlier store couldn't be eliminated.
136 This step uses cselib and canon_rtx to build the largest expression
137 possible for each address. This pass is a forwards pass through
138 each basic block. From the point of view of the global technique,
139 the first pass could examine a block in either direction. The
140 forwards ordering is to accommodate cselib.
142 We make a simplifying assumption: addresses fall into four broad
145 1) base has rtx_varies_p == false, offset is constant.
146 2) base has rtx_varies_p == false, offset variable.
147 3) base has rtx_varies_p == true, offset constant.
148 4) base has rtx_varies_p == true, offset variable.
150 The local passes are able to process all 4 kinds of addresses. The
151 global pass only handles 1).
153 The global problem is formulated as follows:
155 A store, S1, to address A, where A is not relative to the stack
156 frame, can be eliminated if all paths from S1 to the end of the
157 function contain another store to A before a read to A.
159 If the address A is relative to the stack frame, a store S2 to A
160 can be eliminated if there are no paths from S2 that reach the
161 end of the function that read A before another store to A. In
162 this case S2 can be deleted if there are paths from S2 to the
163 end of the function that have no reads or writes to A. This
164 second case allows stores to the stack frame to be deleted that
165 would otherwise die when the function returns. This cannot be
166 done if stores_off_frame_dead_at_return is not true. See the doc
167 for that variable for when this variable is false.
169 The global problem is formulated as a backwards set union
170 dataflow problem where the stores are the gens and reads are the
171 kills. Set union problems are rare and require some special
172 handling given our representation of bitmaps. A straightforward
173 implementation requires a lot of bitmaps filled with 1s.
174 These are expensive and cumbersome in our bitmap formulation so
175 care has been taken to avoid large vectors filled with 1s. See
176 the comments in bb_info and in the dataflow confluence functions
179 There are two places for further enhancements to this algorithm:
181 1) The original dse which was embedded in a pass called flow also
182 did local address forwarding. For example in
187 flow would replace the right hand side of the second insn with a
188 reference to r100. Most of the information is available to add this
189 to this pass. It has not done it because it is a lot of work in
190 the case that either r100 is assigned to between the first and
191 second insn and/or the second insn is a load of part of the value
192 stored by the first insn.
194 insn 5 in gcc.c-torture/compile/990203-1.c simple case.
195 insn 15 in gcc.c-torture/execute/20001017-2.c simple case.
196 insn 25 in gcc.c-torture/execute/20001026-1.c simple case.
197 insn 44 in gcc.c-torture/execute/20010910-1.c simple case.
199 2) The cleaning up of spill code is quite profitable. It currently
200 depends on reading tea leaves and chicken entrails left by reload.
201 This pass depends on reload creating a singleton alias set for each
202 spill slot and telling the next dse pass which of these alias sets
203 are the singletons. Rather than analyze the addresses of the
204 spills, dse's spill processing just does analysis of the loads and
205 stores that use those alias sets. There are three cases where this
208 a) Reload sometimes creates the slot for one mode of access, and
209 then inserts loads and/or stores for a smaller mode. In this
210 case, the current code just punts on the slot. The proper thing
211 to do is to back out and use one bit vector position for each
212 byte of the entity associated with the slot. This depends on
213 KNOWING that reload always generates the accesses for each of the
214 bytes in some canonical (read that easy to understand several
215 passes after reload happens) way.
217 b) Reload sometimes decides that spill slot it allocated was not
218 large enough for the mode and goes back and allocates more slots
219 with the same mode and alias set. The backout in this case is a
220 little more graceful than (a). In this case the slot is unmarked
221 as being a spill slot and if final address comes out to be based
222 off the frame pointer, the global algorithm handles this slot.
224 c) For any pass that may prespill, there is currently no
225 mechanism to tell the dse pass that the slot being used has the
226 special properties that reload uses. It may be that all that is
227 required is to have those passes make the same calls that reload
228 does, assuming that the alias sets can be manipulated in the same
231 /* There are limits to the size of constant offsets we model for the
232 global problem. There are certainly test cases, that exceed this
233 limit, however, it is unlikely that there are important programs
234 that really have constant offsets this size. */
235 #define MAX_OFFSET (64 * 1024)
237 /* Obstack for the DSE dataflow bitmaps. We don't want to put these
238 on the default obstack because these bitmaps can grow quite large
239 (~2GB for the small (!) test case of PR54146) and we'll hold on to
240 all that memory until the end of the compiler run.
241 As a bonus, delete_tree_live_info can destroy all the bitmaps by just
242 releasing the whole obstack. */
243 static bitmap_obstack dse_bitmap_obstack
;
245 /* Obstack for other data. As for above: Kinda nice to be able to
246 throw it all away at the end in one big sweep. */
247 static struct obstack dse_obstack
;
249 /* Scratch bitmap for cselib's cselib_expand_value_rtx. */
250 static bitmap scratch
= NULL
;
254 /* This structure holds information about a candidate store. */
258 /* False means this is a clobber. */
261 /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */
264 /* The id of the mem group of the base address. If rtx_varies_p is
265 true, this is -1. Otherwise, it is the index into the group
269 /* This is the cselib value. */
270 cselib_val
*cse_base
;
272 /* This canonized mem. */
275 /* Canonized MEM address for use by canon_true_dependence. */
278 /* If this is non-zero, it is the alias set of a spill location. */
279 alias_set_type alias_set
;
281 /* The offset of the first and byte before the last byte associated
282 with the operation. */
283 HOST_WIDE_INT begin
, end
;
287 /* A bitmask as wide as the number of bytes in the word that
288 contains a 1 if the byte may be needed. The store is unused if
289 all of the bits are 0. This is used if IS_LARGE is false. */
290 unsigned HOST_WIDE_INT small_bitmask
;
294 /* A bitmap with one bit per byte. Cleared bit means the position
295 is needed. Used if IS_LARGE is false. */
298 /* Number of set bits (i.e. unneeded bytes) in BITMAP. If it is
299 equal to END - BEGIN, the whole store is unused. */
304 /* The next store info for this insn. */
305 struct store_info
*next
;
307 /* The right hand side of the store. This is used if there is a
308 subsequent reload of the mems address somewhere later in the
312 /* If rhs is or holds a constant, this contains that constant,
316 /* Set if this store stores the same constant value as REDUNDANT_REASON
317 insn stored. These aren't eliminated early, because doing that
318 might prevent the earlier larger store to be eliminated. */
319 struct insn_info
*redundant_reason
;
322 /* Return a bitmask with the first N low bits set. */
324 static unsigned HOST_WIDE_INT
325 lowpart_bitmask (int n
)
327 unsigned HOST_WIDE_INT mask
= ~(unsigned HOST_WIDE_INT
) 0;
328 return mask
>> (HOST_BITS_PER_WIDE_INT
- n
);
331 typedef struct store_info
*store_info_t
;
332 static alloc_pool cse_store_info_pool
;
333 static alloc_pool rtx_store_info_pool
;
335 /* This structure holds information about a load. These are only
336 built for rtx bases. */
339 /* The id of the mem group of the base address. */
342 /* If this is non-zero, it is the alias set of a spill location. */
343 alias_set_type alias_set
;
345 /* The offset of the first and byte after the last byte associated
346 with the operation. If begin == end == 0, the read did not have
347 a constant offset. */
350 /* The mem being read. */
353 /* The next read_info for this insn. */
354 struct read_info
*next
;
356 typedef struct read_info
*read_info_t
;
357 static alloc_pool read_info_pool
;
360 /* One of these records is created for each insn. */
364 /* Set true if the insn contains a store but the insn itself cannot
365 be deleted. This is set if the insn is a parallel and there is
366 more than one non dead output or if the insn is in some way
370 /* This field is only used by the global algorithm. It is set true
371 if the insn contains any read of mem except for a (1). This is
372 also set if the insn is a call or has a clobber mem. If the insn
373 contains a wild read, the use_rec will be null. */
376 /* This is true only for CALL instructions which could potentially read
377 any non-frame memory location. This field is used by the global
379 bool non_frame_wild_read
;
381 /* This field is only used for the processing of const functions.
382 These functions cannot read memory, but they can read the stack
383 because that is where they may get their parms. We need to be
384 this conservative because, like the store motion pass, we don't
385 consider CALL_INSN_FUNCTION_USAGE when processing call insns.
386 Moreover, we need to distinguish two cases:
387 1. Before reload (register elimination), the stores related to
388 outgoing arguments are stack pointer based and thus deemed
389 of non-constant base in this pass. This requires special
390 handling but also means that the frame pointer based stores
391 need not be killed upon encountering a const function call.
392 2. After reload, the stores related to outgoing arguments can be
393 either stack pointer or hard frame pointer based. This means
394 that we have no other choice than also killing all the frame
395 pointer based stores upon encountering a const function call.
396 This field is set after reload for const function calls and before
397 reload for const tail function calls on targets where arg pointer
398 is the frame pointer. Having this set is less severe than a wild
399 read, it just means that all the frame related stores are killed
400 rather than all the stores. */
403 /* This field is only used for the processing of const functions.
404 It is set if the insn may contain a stack pointer based store. */
405 bool stack_pointer_based
;
407 /* This is true if any of the sets within the store contains a
408 cselib base. Such stores can only be deleted by the local
410 bool contains_cselib_groups
;
415 /* The list of mem sets or mem clobbers that are contained in this
416 insn. If the insn is deletable, it contains only one mem set.
417 But it could also contain clobbers. Insns that contain more than
418 one mem set are not deletable, but each of those mems are here in
419 order to provide info to delete other insns. */
420 store_info_t store_rec
;
422 /* The linked list of mem uses in this insn. Only the reads from
423 rtx bases are listed here. The reads to cselib bases are
424 completely processed during the first scan and so are never
426 read_info_t read_rec
;
428 /* The live fixed registers. We assume only fixed registers can
429 cause trouble by being clobbered from an expanded pattern;
430 storing only the live fixed registers (rather than all registers)
431 means less memory needs to be allocated / copied for the individual
433 regset fixed_regs_live
;
435 /* The prev insn in the basic block. */
436 struct insn_info
* prev_insn
;
438 /* The linked list of insns that are in consideration for removal in
439 the forwards pass through the basic block. This pointer may be
440 trash as it is not cleared when a wild read occurs. The only
441 time it is guaranteed to be correct is when the traversal starts
442 at active_local_stores. */
443 struct insn_info
* next_local_store
;
446 typedef struct insn_info
*insn_info_t
;
447 static alloc_pool insn_info_pool
;
449 /* The linked list of stores that are under consideration in this
451 static insn_info_t active_local_stores
;
452 static int active_local_stores_len
;
457 /* Pointer to the insn info for the last insn in the block. These
458 are linked so this is how all of the insns are reached. During
459 scanning this is the current insn being scanned. */
460 insn_info_t last_insn
;
462 /* The info for the global dataflow problem. */
465 /* This is set if the transfer function should and in the wild_read
466 bitmap before applying the kill and gen sets. That vector knocks
467 out most of the bits in the bitmap and thus speeds up the
469 bool apply_wild_read
;
471 /* The following 4 bitvectors hold information about which positions
472 of which stores are live or dead. They are indexed by
475 /* The set of store positions that exist in this block before a wild read. */
478 /* The set of load positions that exist in this block above the
479 same position of a store. */
482 /* The set of stores that reach the top of the block without being
485 Do not represent the in if it is all ones. Note that this is
486 what the bitvector should logically be initialized to for a set
487 intersection problem. However, like the kill set, this is too
488 expensive. So initially, the in set will only be created for the
489 exit block and any block that contains a wild read. */
492 /* The set of stores that reach the bottom of the block from it's
495 Do not represent the in if it is all ones. Note that this is
496 what the bitvector should logically be initialized to for a set
497 intersection problem. However, like the kill and in set, this is
498 too expensive. So what is done is that the confluence operator
499 just initializes the vector from one of the out sets of the
500 successors of the block. */
503 /* The following bitvector is indexed by the reg number. It
504 contains the set of regs that are live at the current instruction
505 being processed. While it contains info for all of the
506 registers, only the hard registers are actually examined. It is used
507 to assure that shift and/or add sequences that are inserted do not
508 accidentally clobber live hard regs. */
512 typedef struct dse_bb_info
*bb_info_t
;
513 static alloc_pool bb_info_pool
;
515 /* Table to hold all bb_infos. */
516 static bb_info_t
*bb_table
;
518 /* There is a group_info for each rtx base that is used to reference
519 memory. There are also not many of the rtx bases because they are
520 very limited in scope. */
524 /* The actual base of the address. */
527 /* The sequential id of the base. This allows us to have a
528 canonical ordering of these that is not based on addresses. */
531 /* True if there are any positions that are to be processed
533 bool process_globally
;
535 /* True if the base of this group is either the frame_pointer or
536 hard_frame_pointer. */
539 /* A mem wrapped around the base pointer for the group in order to do
540 read dependency. It must be given BLKmode in order to encompass all
541 the possible offsets from the base. */
544 /* Canonized version of base_mem's address. */
547 /* These two sets of two bitmaps are used to keep track of how many
548 stores are actually referencing that position from this base. We
549 only do this for rtx bases as this will be used to assign
550 positions in the bitmaps for the global problem. Bit N is set in
551 store1 on the first store for offset N. Bit N is set in store2
552 for the second store to offset N. This is all we need since we
553 only care about offsets that have two or more stores for them.
555 The "_n" suffix is for offsets less than 0 and the "_p" suffix is
556 for 0 and greater offsets.
558 There is one special case here, for stores into the stack frame,
559 we will or store1 into store2 before deciding which stores look
560 at globally. This is because stores to the stack frame that have
561 no other reads before the end of the function can also be
563 bitmap store1_n
, store1_p
, store2_n
, store2_p
;
565 /* These bitmaps keep track of offsets in this group escape this function.
566 An offset escapes if it corresponds to a named variable whose
567 addressable flag is set. */
568 bitmap escaped_n
, escaped_p
;
570 /* The positions in this bitmap have the same assignments as the in,
571 out, gen and kill bitmaps. This bitmap is all zeros except for
572 the positions that are occupied by stores for this group. */
575 /* The offset_map is used to map the offsets from this base into
576 positions in the global bitmaps. It is only created after all of
577 the all of stores have been scanned and we know which ones we
579 int *offset_map_n
, *offset_map_p
;
580 int offset_map_size_n
, offset_map_size_p
;
582 typedef struct group_info
*group_info_t
;
583 typedef const struct group_info
*const_group_info_t
;
584 static alloc_pool rtx_group_info_pool
;
586 /* Index into the rtx_group_vec. */
587 static int rtx_group_next_id
;
590 static vec
<group_info_t
> rtx_group_vec
;
593 /* This structure holds the set of changes that are being deferred
594 when removing read operation. See replace_read. */
595 struct deferred_change
598 /* The mem that is being replaced. */
601 /* The reg it is being replaced with. */
604 struct deferred_change
*next
;
607 typedef struct deferred_change
*deferred_change_t
;
608 static alloc_pool deferred_change_pool
;
610 static deferred_change_t deferred_change_list
= NULL
;
612 /* The group that holds all of the clear_alias_sets. */
613 static group_info_t clear_alias_group
;
615 /* The modes of the clear_alias_sets. */
616 static htab_t clear_alias_mode_table
;
618 /* Hash table element to look up the mode for an alias set. */
619 struct clear_alias_mode_holder
621 alias_set_type alias_set
;
625 /* This is true except if cfun->stdarg -- i.e. we cannot do
626 this for vararg functions because they play games with the frame. */
627 static bool stores_off_frame_dead_at_return
;
629 /* Counter for stats. */
630 static int globally_deleted
;
631 static int locally_deleted
;
632 static int spill_deleted
;
634 static bitmap all_blocks
;
636 /* Locations that are killed by calls in the global phase. */
637 static bitmap kill_on_calls
;
639 /* The number of bits used in the global bitmaps. */
640 static unsigned int current_position
;
642 /*----------------------------------------------------------------------------
646 ----------------------------------------------------------------------------*/
649 /* Find the entry associated with ALIAS_SET. */
651 static struct clear_alias_mode_holder
*
652 clear_alias_set_lookup (alias_set_type alias_set
)
654 struct clear_alias_mode_holder tmp_holder
;
657 tmp_holder
.alias_set
= alias_set
;
658 slot
= htab_find_slot (clear_alias_mode_table
, &tmp_holder
, NO_INSERT
);
661 return (struct clear_alias_mode_holder
*) *slot
;
665 /* Hashtable callbacks for maintaining the "bases" field of
666 store_group_info, given that the addresses are function invariants. */
668 struct invariant_group_base_hasher
: typed_noop_remove
<group_info
>
670 typedef group_info value_type
;
671 typedef group_info compare_type
;
672 static inline hashval_t
hash (const value_type
*);
673 static inline bool equal (const value_type
*, const compare_type
*);
677 invariant_group_base_hasher::equal (const value_type
*gi1
,
678 const compare_type
*gi2
)
680 return rtx_equal_p (gi1
->rtx_base
, gi2
->rtx_base
);
684 invariant_group_base_hasher::hash (const value_type
*gi
)
687 return hash_rtx (gi
->rtx_base
, Pmode
, &do_not_record
, NULL
, false);
690 /* Tables of group_info structures, hashed by base value. */
691 static hash_table
<invariant_group_base_hasher
> *rtx_group_table
;
694 /* Get the GROUP for BASE. Add a new group if it is not there. */
697 get_group_info (rtx base
)
699 struct group_info tmp_gi
;
705 /* Find the store_base_info structure for BASE, creating a new one
707 tmp_gi
.rtx_base
= base
;
708 slot
= rtx_group_table
->find_slot (&tmp_gi
, INSERT
);
709 gi
= (group_info_t
) *slot
;
713 if (!clear_alias_group
)
715 clear_alias_group
= gi
=
716 (group_info_t
) pool_alloc (rtx_group_info_pool
);
717 memset (gi
, 0, sizeof (struct group_info
));
718 gi
->id
= rtx_group_next_id
++;
719 gi
->store1_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
720 gi
->store1_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
721 gi
->store2_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
722 gi
->store2_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
723 gi
->escaped_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
724 gi
->escaped_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
725 gi
->group_kill
= BITMAP_ALLOC (&dse_bitmap_obstack
);
726 gi
->process_globally
= false;
727 gi
->offset_map_size_n
= 0;
728 gi
->offset_map_size_p
= 0;
729 gi
->offset_map_n
= NULL
;
730 gi
->offset_map_p
= NULL
;
731 rtx_group_vec
.safe_push (gi
);
733 return clear_alias_group
;
738 *slot
= gi
= (group_info_t
) pool_alloc (rtx_group_info_pool
);
740 gi
->id
= rtx_group_next_id
++;
741 gi
->base_mem
= gen_rtx_MEM (BLKmode
, base
);
742 gi
->canon_base_addr
= canon_rtx (base
);
743 gi
->store1_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
744 gi
->store1_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
745 gi
->store2_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
746 gi
->store2_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
747 gi
->escaped_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
748 gi
->escaped_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
749 gi
->group_kill
= BITMAP_ALLOC (&dse_bitmap_obstack
);
750 gi
->process_globally
= false;
752 (base
== frame_pointer_rtx
) || (base
== hard_frame_pointer_rtx
);
753 gi
->offset_map_size_n
= 0;
754 gi
->offset_map_size_p
= 0;
755 gi
->offset_map_n
= NULL
;
756 gi
->offset_map_p
= NULL
;
757 rtx_group_vec
.safe_push (gi
);
764 /* Initialization of data structures. */
770 globally_deleted
= 0;
773 bitmap_obstack_initialize (&dse_bitmap_obstack
);
774 gcc_obstack_init (&dse_obstack
);
776 scratch
= BITMAP_ALLOC (®_obstack
);
777 kill_on_calls
= BITMAP_ALLOC (&dse_bitmap_obstack
);
780 = create_alloc_pool ("rtx_store_info_pool",
781 sizeof (struct store_info
), 100);
783 = create_alloc_pool ("read_info_pool",
784 sizeof (struct read_info
), 100);
786 = create_alloc_pool ("insn_info_pool",
787 sizeof (struct insn_info
), 100);
789 = create_alloc_pool ("bb_info_pool",
790 sizeof (struct dse_bb_info
), 100);
792 = create_alloc_pool ("rtx_group_info_pool",
793 sizeof (struct group_info
), 100);
795 = create_alloc_pool ("deferred_change_pool",
796 sizeof (struct deferred_change
), 10);
798 rtx_group_table
= new hash_table
<invariant_group_base_hasher
> (11);
800 bb_table
= XNEWVEC (bb_info_t
, last_basic_block_for_fn (cfun
));
801 rtx_group_next_id
= 0;
803 stores_off_frame_dead_at_return
= !cfun
->stdarg
;
805 init_alias_analysis ();
807 clear_alias_group
= NULL
;
812 /*----------------------------------------------------------------------------
815 Scan all of the insns. Any random ordering of the blocks is fine.
816 Each block is scanned in forward order to accommodate cselib which
817 is used to remove stores with non-constant bases.
818 ----------------------------------------------------------------------------*/
820 /* Delete all of the store_info recs from INSN_INFO. */
823 free_store_info (insn_info_t insn_info
)
825 store_info_t store_info
= insn_info
->store_rec
;
828 store_info_t next
= store_info
->next
;
829 if (store_info
->is_large
)
830 BITMAP_FREE (store_info
->positions_needed
.large
.bmap
);
831 if (store_info
->cse_base
)
832 pool_free (cse_store_info_pool
, store_info
);
834 pool_free (rtx_store_info_pool
, store_info
);
838 insn_info
->cannot_delete
= true;
839 insn_info
->contains_cselib_groups
= false;
840 insn_info
->store_rec
= NULL
;
845 rtx_insn
*first
, *current
;
846 regset fixed_regs_live
;
848 } note_add_store_info
;
850 /* Callback for emit_inc_dec_insn_before via note_stores.
851 Check if a register is clobbered which is live afterwards. */
854 note_add_store (rtx loc
, const_rtx expr ATTRIBUTE_UNUSED
, void *data
)
857 note_add_store_info
*info
= (note_add_store_info
*) data
;
863 /* If this register is referenced by the current or an earlier insn,
864 that's OK. E.g. this applies to the register that is being incremented
865 with this addition. */
866 for (insn
= info
->first
;
867 insn
!= NEXT_INSN (info
->current
);
868 insn
= NEXT_INSN (insn
))
869 if (reg_referenced_p (loc
, PATTERN (insn
)))
872 /* If we come here, we have a clobber of a register that's only OK
873 if that register is not live. If we don't have liveness information
874 available, fail now. */
875 if (!info
->fixed_regs_live
)
877 info
->failure
= true;
880 /* Now check if this is a live fixed register. */
882 n
= hard_regno_nregs
[r
][GET_MODE (loc
)];
884 if (REGNO_REG_SET_P (info
->fixed_regs_live
, r
+n
))
885 info
->failure
= true;
888 /* Callback for for_each_inc_dec that emits an INSN that sets DEST to
889 SRC + SRCOFF before insn ARG. */
892 emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED
,
893 rtx op ATTRIBUTE_UNUSED
,
894 rtx dest
, rtx src
, rtx srcoff
, void *arg
)
896 insn_info_t insn_info
= (insn_info_t
) arg
;
897 rtx_insn
*insn
= insn_info
->insn
, *new_insn
, *cur
;
898 note_add_store_info info
;
900 /* We can reuse all operands without copying, because we are about
901 to delete the insn that contained it. */
905 emit_insn (gen_add3_insn (dest
, src
, srcoff
));
906 new_insn
= get_insns ();
910 new_insn
= as_a
<rtx_insn
*> (gen_move_insn (dest
, src
));
911 info
.first
= new_insn
;
912 info
.fixed_regs_live
= insn_info
->fixed_regs_live
;
913 info
.failure
= false;
914 for (cur
= new_insn
; cur
; cur
= NEXT_INSN (cur
))
917 note_stores (PATTERN (cur
), note_add_store
, &info
);
920 /* If a failure was flagged above, return 1 so that for_each_inc_dec will
921 return it immediately, communicating the failure to its caller. */
925 emit_insn_before (new_insn
, insn
);
930 /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it
931 is there, is split into a separate insn.
932 Return true on success (or if there was nothing to do), false on failure. */
935 check_for_inc_dec_1 (insn_info_t insn_info
)
937 rtx_insn
*insn
= insn_info
->insn
;
938 rtx note
= find_reg_note (insn
, REG_INC
, NULL_RTX
);
940 return for_each_inc_dec (PATTERN (insn
), emit_inc_dec_insn_before
,
946 /* Entry point for postreload. If you work on reload_cse, or you need this
947 anywhere else, consider if you can provide register liveness information
948 and add a parameter to this function so that it can be passed down in
949 insn_info.fixed_regs_live. */
951 check_for_inc_dec (rtx_insn
*insn
)
953 struct insn_info insn_info
;
956 insn_info
.insn
= insn
;
957 insn_info
.fixed_regs_live
= NULL
;
958 note
= find_reg_note (insn
, REG_INC
, NULL_RTX
);
960 return for_each_inc_dec (PATTERN (insn
), emit_inc_dec_insn_before
,
965 /* Delete the insn and free all of the fields inside INSN_INFO. */
968 delete_dead_store_insn (insn_info_t insn_info
)
970 read_info_t read_info
;
975 if (!check_for_inc_dec_1 (insn_info
))
977 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
979 fprintf (dump_file
, "Locally deleting insn %d ",
980 INSN_UID (insn_info
->insn
));
981 if (insn_info
->store_rec
->alias_set
)
982 fprintf (dump_file
, "alias set %d\n",
983 (int) insn_info
->store_rec
->alias_set
);
985 fprintf (dump_file
, "\n");
988 free_store_info (insn_info
);
989 read_info
= insn_info
->read_rec
;
993 read_info_t next
= read_info
->next
;
994 pool_free (read_info_pool
, read_info
);
997 insn_info
->read_rec
= NULL
;
999 delete_insn (insn_info
->insn
);
1001 insn_info
->insn
= NULL
;
1003 insn_info
->wild_read
= false;
1006 /* Return whether DECL, a local variable, can possibly escape the current
1010 local_variable_can_escape (tree decl
)
1012 if (TREE_ADDRESSABLE (decl
))
1015 /* If this is a partitioned variable, we need to consider all the variables
1016 in the partition. This is necessary because a store into one of them can
1017 be replaced with a store into another and this may not change the outcome
1018 of the escape analysis. */
1019 if (cfun
->gimple_df
->decls_to_pointers
!= NULL
)
1021 tree
*namep
= cfun
->gimple_df
->decls_to_pointers
->get (decl
);
1023 return TREE_ADDRESSABLE (*namep
);
1029 /* Return whether EXPR can possibly escape the current function scope. */
1032 can_escape (tree expr
)
1037 base
= get_base_address (expr
);
1039 && !may_be_aliased (base
)
1040 && !(TREE_CODE (base
) == VAR_DECL
1041 && !DECL_EXTERNAL (base
)
1042 && !TREE_STATIC (base
)
1043 && local_variable_can_escape (base
)))
1048 /* Set the store* bitmaps offset_map_size* fields in GROUP based on
1049 OFFSET and WIDTH. */
1052 set_usage_bits (group_info_t group
, HOST_WIDE_INT offset
, HOST_WIDE_INT width
,
1056 bool expr_escapes
= can_escape (expr
);
1057 if (offset
> -MAX_OFFSET
&& offset
+ width
< MAX_OFFSET
)
1058 for (i
=offset
; i
<offset
+width
; i
++)
1066 store1
= group
->store1_n
;
1067 store2
= group
->store2_n
;
1068 escaped
= group
->escaped_n
;
1073 store1
= group
->store1_p
;
1074 store2
= group
->store2_p
;
1075 escaped
= group
->escaped_p
;
1079 if (!bitmap_set_bit (store1
, ai
))
1080 bitmap_set_bit (store2
, ai
);
1085 if (group
->offset_map_size_n
< ai
)
1086 group
->offset_map_size_n
= ai
;
1090 if (group
->offset_map_size_p
< ai
)
1091 group
->offset_map_size_p
= ai
;
1095 bitmap_set_bit (escaped
, ai
);
1100 reset_active_stores (void)
1102 active_local_stores
= NULL
;
1103 active_local_stores_len
= 0;
1106 /* Free all READ_REC of the LAST_INSN of BB_INFO. */
1109 free_read_records (bb_info_t bb_info
)
1111 insn_info_t insn_info
= bb_info
->last_insn
;
1112 read_info_t
*ptr
= &insn_info
->read_rec
;
1115 read_info_t next
= (*ptr
)->next
;
1116 if ((*ptr
)->alias_set
== 0)
1118 pool_free (read_info_pool
, *ptr
);
1122 ptr
= &(*ptr
)->next
;
1126 /* Set the BB_INFO so that the last insn is marked as a wild read. */
1129 add_wild_read (bb_info_t bb_info
)
1131 insn_info_t insn_info
= bb_info
->last_insn
;
1132 insn_info
->wild_read
= true;
1133 free_read_records (bb_info
);
1134 reset_active_stores ();
1137 /* Set the BB_INFO so that the last insn is marked as a wild read of
1138 non-frame locations. */
1141 add_non_frame_wild_read (bb_info_t bb_info
)
1143 insn_info_t insn_info
= bb_info
->last_insn
;
1144 insn_info
->non_frame_wild_read
= true;
1145 free_read_records (bb_info
);
1146 reset_active_stores ();
1149 /* Return true if X is a constant or one of the registers that behave
1150 as a constant over the life of a function. This is equivalent to
1151 !rtx_varies_p for memory addresses. */
1154 const_or_frame_p (rtx x
)
1159 if (GET_CODE (x
) == REG
)
1161 /* Note that we have to test for the actual rtx used for the frame
1162 and arg pointers and not just the register number in case we have
1163 eliminated the frame and/or arg pointer and are using it
1165 if (x
== frame_pointer_rtx
|| x
== hard_frame_pointer_rtx
1166 /* The arg pointer varies if it is not a fixed register. */
1167 || (x
== arg_pointer_rtx
&& fixed_regs
[ARG_POINTER_REGNUM
])
1168 || x
== pic_offset_table_rtx
)
1176 /* Take all reasonable action to put the address of MEM into the form
1177 that we can do analysis on.
1179 The gold standard is to get the address into the form: address +
1180 OFFSET where address is something that rtx_varies_p considers a
1181 constant. When we can get the address in this form, we can do
1182 global analysis on it. Note that for constant bases, address is
1183 not actually returned, only the group_id. The address can be
1186 If that fails, we try cselib to get a value we can at least use
1187 locally. If that fails we return false.
1189 The GROUP_ID is set to -1 for cselib bases and the index of the
1190 group for non_varying bases.
1192 FOR_READ is true if this is a mem read and false if not. */
1195 canon_address (rtx mem
,
1196 alias_set_type
*alias_set_out
,
1198 HOST_WIDE_INT
*offset
,
1201 machine_mode address_mode
= get_address_mode (mem
);
1202 rtx mem_address
= XEXP (mem
, 0);
1203 rtx expanded_address
, address
;
1208 cselib_lookup (mem_address
, address_mode
, 1, GET_MODE (mem
));
1210 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1212 fprintf (dump_file
, " mem: ");
1213 print_inline_rtx (dump_file
, mem_address
, 0);
1214 fprintf (dump_file
, "\n");
1217 /* First see if just canon_rtx (mem_address) is const or frame,
1218 if not, try cselib_expand_value_rtx and call canon_rtx on that. */
1220 for (expanded
= 0; expanded
< 2; expanded
++)
1224 /* Use cselib to replace all of the reg references with the full
1225 expression. This will take care of the case where we have
1227 r_x = base + offset;
1232 val = *(base + offset); */
1234 expanded_address
= cselib_expand_value_rtx (mem_address
,
1237 /* If this fails, just go with the address from first
1239 if (!expanded_address
)
1243 expanded_address
= mem_address
;
1245 /* Split the address into canonical BASE + OFFSET terms. */
1246 address
= canon_rtx (expanded_address
);
1250 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1254 fprintf (dump_file
, "\n after cselib_expand address: ");
1255 print_inline_rtx (dump_file
, expanded_address
, 0);
1256 fprintf (dump_file
, "\n");
1259 fprintf (dump_file
, "\n after canon_rtx address: ");
1260 print_inline_rtx (dump_file
, address
, 0);
1261 fprintf (dump_file
, "\n");
1264 if (GET_CODE (address
) == CONST
)
1265 address
= XEXP (address
, 0);
1267 if (GET_CODE (address
) == PLUS
1268 && CONST_INT_P (XEXP (address
, 1)))
1270 *offset
= INTVAL (XEXP (address
, 1));
1271 address
= XEXP (address
, 0);
1274 if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem
))
1275 && const_or_frame_p (address
))
1277 group_info_t group
= get_group_info (address
);
1279 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1280 fprintf (dump_file
, " gid=%d offset=%d \n",
1281 group
->id
, (int)*offset
);
1283 *group_id
= group
->id
;
1288 *base
= cselib_lookup (address
, address_mode
, true, GET_MODE (mem
));
1293 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1294 fprintf (dump_file
, " no cselib val - should be a wild read.\n");
1297 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1298 fprintf (dump_file
, " varying cselib base=%u:%u offset = %d\n",
1299 (*base
)->uid
, (*base
)->hash
, (int)*offset
);
1304 /* Clear the rhs field from the active_local_stores array. */
1307 clear_rhs_from_active_local_stores (void)
1309 insn_info_t ptr
= active_local_stores
;
1313 store_info_t store_info
= ptr
->store_rec
;
1314 /* Skip the clobbers. */
1315 while (!store_info
->is_set
)
1316 store_info
= store_info
->next
;
1318 store_info
->rhs
= NULL
;
1319 store_info
->const_rhs
= NULL
;
1321 ptr
= ptr
->next_local_store
;
1326 /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */
1329 set_position_unneeded (store_info_t s_info
, int pos
)
1331 if (__builtin_expect (s_info
->is_large
, false))
1333 if (bitmap_set_bit (s_info
->positions_needed
.large
.bmap
, pos
))
1334 s_info
->positions_needed
.large
.count
++;
1337 s_info
->positions_needed
.small_bitmask
1338 &= ~(((unsigned HOST_WIDE_INT
) 1) << pos
);
1341 /* Mark the whole store S_INFO as unneeded. */
1344 set_all_positions_unneeded (store_info_t s_info
)
1346 if (__builtin_expect (s_info
->is_large
, false))
1348 int pos
, end
= s_info
->end
- s_info
->begin
;
1349 for (pos
= 0; pos
< end
; pos
++)
1350 bitmap_set_bit (s_info
->positions_needed
.large
.bmap
, pos
);
1351 s_info
->positions_needed
.large
.count
= end
;
1354 s_info
->positions_needed
.small_bitmask
= (unsigned HOST_WIDE_INT
) 0;
1357 /* Return TRUE if any bytes from S_INFO store are needed. */
1360 any_positions_needed_p (store_info_t s_info
)
1362 if (__builtin_expect (s_info
->is_large
, false))
1363 return (s_info
->positions_needed
.large
.count
1364 < s_info
->end
- s_info
->begin
);
1366 return (s_info
->positions_needed
.small_bitmask
1367 != (unsigned HOST_WIDE_INT
) 0);
1370 /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO
1371 store are needed. */
1374 all_positions_needed_p (store_info_t s_info
, int start
, int width
)
1376 if (__builtin_expect (s_info
->is_large
, false))
1378 int end
= start
+ width
;
1380 if (bitmap_bit_p (s_info
->positions_needed
.large
.bmap
, start
++))
1386 unsigned HOST_WIDE_INT mask
= lowpart_bitmask (width
) << start
;
1387 return (s_info
->positions_needed
.small_bitmask
& mask
) == mask
;
1392 static rtx
get_stored_val (store_info_t
, machine_mode
, HOST_WIDE_INT
,
1393 HOST_WIDE_INT
, basic_block
, bool);
1396 /* BODY is an instruction pattern that belongs to INSN. Return 1 if
1397 there is a candidate store, after adding it to the appropriate
1398 local store group if so. */
1401 record_store (rtx body
, bb_info_t bb_info
)
1403 rtx mem
, rhs
, const_rhs
, mem_addr
;
1404 HOST_WIDE_INT offset
= 0;
1405 HOST_WIDE_INT width
= 0;
1406 alias_set_type spill_alias_set
;
1407 insn_info_t insn_info
= bb_info
->last_insn
;
1408 store_info_t store_info
= NULL
;
1410 cselib_val
*base
= NULL
;
1411 insn_info_t ptr
, last
, redundant_reason
;
1412 bool store_is_unused
;
1414 if (GET_CODE (body
) != SET
&& GET_CODE (body
) != CLOBBER
)
1417 mem
= SET_DEST (body
);
1419 /* If this is not used, then this cannot be used to keep the insn
1420 from being deleted. On the other hand, it does provide something
1421 that can be used to prove that another store is dead. */
1423 = (find_reg_note (insn_info
->insn
, REG_UNUSED
, mem
) != NULL
);
1425 /* Check whether that value is a suitable memory location. */
1428 /* If the set or clobber is unused, then it does not effect our
1429 ability to get rid of the entire insn. */
1430 if (!store_is_unused
)
1431 insn_info
->cannot_delete
= true;
1435 /* At this point we know mem is a mem. */
1436 if (GET_MODE (mem
) == BLKmode
)
1438 if (GET_CODE (XEXP (mem
, 0)) == SCRATCH
)
1440 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1441 fprintf (dump_file
, " adding wild read for (clobber (mem:BLK (scratch))\n");
1442 add_wild_read (bb_info
);
1443 insn_info
->cannot_delete
= true;
1446 /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0))
1447 as memset (addr, 0, 36); */
1448 else if (!MEM_SIZE_KNOWN_P (mem
)
1449 || MEM_SIZE (mem
) <= 0
1450 || MEM_SIZE (mem
) > MAX_OFFSET
1451 || GET_CODE (body
) != SET
1452 || !CONST_INT_P (SET_SRC (body
)))
1454 if (!store_is_unused
)
1456 /* If the set or clobber is unused, then it does not effect our
1457 ability to get rid of the entire insn. */
1458 insn_info
->cannot_delete
= true;
1459 clear_rhs_from_active_local_stores ();
1465 /* We can still process a volatile mem, we just cannot delete it. */
1466 if (MEM_VOLATILE_P (mem
))
1467 insn_info
->cannot_delete
= true;
1469 if (!canon_address (mem
, &spill_alias_set
, &group_id
, &offset
, &base
))
1471 clear_rhs_from_active_local_stores ();
1475 if (GET_MODE (mem
) == BLKmode
)
1476 width
= MEM_SIZE (mem
);
1478 width
= GET_MODE_SIZE (GET_MODE (mem
));
1480 if (spill_alias_set
)
1482 bitmap store1
= clear_alias_group
->store1_p
;
1483 bitmap store2
= clear_alias_group
->store2_p
;
1485 gcc_assert (GET_MODE (mem
) != BLKmode
);
1487 if (!bitmap_set_bit (store1
, spill_alias_set
))
1488 bitmap_set_bit (store2
, spill_alias_set
);
1490 if (clear_alias_group
->offset_map_size_p
< spill_alias_set
)
1491 clear_alias_group
->offset_map_size_p
= spill_alias_set
;
1493 store_info
= (store_info_t
) pool_alloc (rtx_store_info_pool
);
1495 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1496 fprintf (dump_file
, " processing spill store %d(%s)\n",
1497 (int) spill_alias_set
, GET_MODE_NAME (GET_MODE (mem
)));
1499 else if (group_id
>= 0)
1501 /* In the restrictive case where the base is a constant or the
1502 frame pointer we can do global analysis. */
1505 = rtx_group_vec
[group_id
];
1506 tree expr
= MEM_EXPR (mem
);
1508 store_info
= (store_info_t
) pool_alloc (rtx_store_info_pool
);
1509 set_usage_bits (group
, offset
, width
, expr
);
1511 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1512 fprintf (dump_file
, " processing const base store gid=%d[%d..%d)\n",
1513 group_id
, (int)offset
, (int)(offset
+width
));
1517 if (may_be_sp_based_p (XEXP (mem
, 0)))
1518 insn_info
->stack_pointer_based
= true;
1519 insn_info
->contains_cselib_groups
= true;
1521 store_info
= (store_info_t
) pool_alloc (cse_store_info_pool
);
1524 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1525 fprintf (dump_file
, " processing cselib store [%d..%d)\n",
1526 (int)offset
, (int)(offset
+width
));
1529 const_rhs
= rhs
= NULL_RTX
;
1530 if (GET_CODE (body
) == SET
1531 /* No place to keep the value after ra. */
1532 && !reload_completed
1533 && (REG_P (SET_SRC (body
))
1534 || GET_CODE (SET_SRC (body
)) == SUBREG
1535 || CONSTANT_P (SET_SRC (body
)))
1536 && !MEM_VOLATILE_P (mem
)
1537 /* Sometimes the store and reload is used for truncation and
1539 && !(FLOAT_MODE_P (GET_MODE (mem
)) && (flag_float_store
)))
1541 rhs
= SET_SRC (body
);
1542 if (CONSTANT_P (rhs
))
1544 else if (body
== PATTERN (insn_info
->insn
))
1546 rtx tem
= find_reg_note (insn_info
->insn
, REG_EQUAL
, NULL_RTX
);
1547 if (tem
&& CONSTANT_P (XEXP (tem
, 0)))
1548 const_rhs
= XEXP (tem
, 0);
1550 if (const_rhs
== NULL_RTX
&& REG_P (rhs
))
1552 rtx tem
= cselib_expand_value_rtx (rhs
, scratch
, 5);
1554 if (tem
&& CONSTANT_P (tem
))
1559 /* Check to see if this stores causes some other stores to be
1561 ptr
= active_local_stores
;
1563 redundant_reason
= NULL
;
1564 mem
= canon_rtx (mem
);
1565 /* For alias_set != 0 canon_true_dependence should be never called. */
1566 if (spill_alias_set
)
1567 mem_addr
= NULL_RTX
;
1571 mem_addr
= base
->val_rtx
;
1575 = rtx_group_vec
[group_id
];
1576 mem_addr
= group
->canon_base_addr
;
1579 mem_addr
= plus_constant (get_address_mode (mem
), mem_addr
, offset
);
1584 insn_info_t next
= ptr
->next_local_store
;
1585 store_info_t s_info
= ptr
->store_rec
;
1588 /* Skip the clobbers. We delete the active insn if this insn
1589 shadows the set. To have been put on the active list, it
1590 has exactly on set. */
1591 while (!s_info
->is_set
)
1592 s_info
= s_info
->next
;
1594 if (s_info
->alias_set
!= spill_alias_set
)
1596 else if (s_info
->alias_set
)
1598 struct clear_alias_mode_holder
*entry
1599 = clear_alias_set_lookup (s_info
->alias_set
);
1600 /* Generally, spills cannot be processed if and of the
1601 references to the slot have a different mode. But if
1602 we are in the same block and mode is exactly the same
1603 between this store and one before in the same block,
1604 we can still delete it. */
1605 if ((GET_MODE (mem
) == GET_MODE (s_info
->mem
))
1606 && (GET_MODE (mem
) == entry
->mode
))
1609 set_all_positions_unneeded (s_info
);
1611 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1612 fprintf (dump_file
, " trying spill store in insn=%d alias_set=%d\n",
1613 INSN_UID (ptr
->insn
), (int) s_info
->alias_set
);
1615 else if ((s_info
->group_id
== group_id
)
1616 && (s_info
->cse_base
== base
))
1619 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1620 fprintf (dump_file
, " trying store in insn=%d gid=%d[%d..%d)\n",
1621 INSN_UID (ptr
->insn
), s_info
->group_id
,
1622 (int)s_info
->begin
, (int)s_info
->end
);
1624 /* Even if PTR won't be eliminated as unneeded, if both
1625 PTR and this insn store the same constant value, we might
1626 eliminate this insn instead. */
1627 if (s_info
->const_rhs
1629 && offset
>= s_info
->begin
1630 && offset
+ width
<= s_info
->end
1631 && all_positions_needed_p (s_info
, offset
- s_info
->begin
,
1634 if (GET_MODE (mem
) == BLKmode
)
1636 if (GET_MODE (s_info
->mem
) == BLKmode
1637 && s_info
->const_rhs
== const_rhs
)
1638 redundant_reason
= ptr
;
1640 else if (s_info
->const_rhs
== const0_rtx
1641 && const_rhs
== const0_rtx
)
1642 redundant_reason
= ptr
;
1647 val
= get_stored_val (s_info
, GET_MODE (mem
),
1648 offset
, offset
+ width
,
1649 BLOCK_FOR_INSN (insn_info
->insn
),
1651 if (get_insns () != NULL
)
1654 if (val
&& rtx_equal_p (val
, const_rhs
))
1655 redundant_reason
= ptr
;
1659 for (i
= MAX (offset
, s_info
->begin
);
1660 i
< offset
+ width
&& i
< s_info
->end
;
1662 set_position_unneeded (s_info
, i
- s_info
->begin
);
1664 else if (s_info
->rhs
)
1665 /* Need to see if it is possible for this store to overwrite
1666 the value of store_info. If it is, set the rhs to NULL to
1667 keep it from being used to remove a load. */
1669 if (canon_true_dependence (s_info
->mem
,
1670 GET_MODE (s_info
->mem
),
1675 s_info
->const_rhs
= NULL
;
1679 /* An insn can be deleted if every position of every one of
1680 its s_infos is zero. */
1681 if (any_positions_needed_p (s_info
))
1686 insn_info_t insn_to_delete
= ptr
;
1688 active_local_stores_len
--;
1690 last
->next_local_store
= ptr
->next_local_store
;
1692 active_local_stores
= ptr
->next_local_store
;
1694 if (!insn_to_delete
->cannot_delete
)
1695 delete_dead_store_insn (insn_to_delete
);
1703 /* Finish filling in the store_info. */
1704 store_info
->next
= insn_info
->store_rec
;
1705 insn_info
->store_rec
= store_info
;
1706 store_info
->mem
= mem
;
1707 store_info
->alias_set
= spill_alias_set
;
1708 store_info
->mem_addr
= mem_addr
;
1709 store_info
->cse_base
= base
;
1710 if (width
> HOST_BITS_PER_WIDE_INT
)
1712 store_info
->is_large
= true;
1713 store_info
->positions_needed
.large
.count
= 0;
1714 store_info
->positions_needed
.large
.bmap
= BITMAP_ALLOC (&dse_bitmap_obstack
);
1718 store_info
->is_large
= false;
1719 store_info
->positions_needed
.small_bitmask
= lowpart_bitmask (width
);
1721 store_info
->group_id
= group_id
;
1722 store_info
->begin
= offset
;
1723 store_info
->end
= offset
+ width
;
1724 store_info
->is_set
= GET_CODE (body
) == SET
;
1725 store_info
->rhs
= rhs
;
1726 store_info
->const_rhs
= const_rhs
;
1727 store_info
->redundant_reason
= redundant_reason
;
1729 /* If this is a clobber, we return 0. We will only be able to
1730 delete this insn if there is only one store USED store, but we
1731 can use the clobber to delete other stores earlier. */
1732 return store_info
->is_set
? 1 : 0;
1737 dump_insn_info (const char * start
, insn_info_t insn_info
)
1739 fprintf (dump_file
, "%s insn=%d %s\n", start
,
1740 INSN_UID (insn_info
->insn
),
1741 insn_info
->store_rec
? "has store" : "naked");
1745 /* If the modes are different and the value's source and target do not
1746 line up, we need to extract the value from lower part of the rhs of
1747 the store, shift it, and then put it into a form that can be shoved
1748 into the read_insn. This function generates a right SHIFT of a
1749 value that is at least ACCESS_SIZE bytes wide of READ_MODE. The
1750 shift sequence is returned or NULL if we failed to find a
1754 find_shift_sequence (int access_size
,
1755 store_info_t store_info
,
1756 machine_mode read_mode
,
1757 int shift
, bool speed
, bool require_cst
)
1759 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1760 machine_mode new_mode
;
1761 rtx read_reg
= NULL
;
1763 /* Some machines like the x86 have shift insns for each size of
1764 operand. Other machines like the ppc or the ia-64 may only have
1765 shift insns that shift values within 32 or 64 bit registers.
1766 This loop tries to find the smallest shift insn that will right
1767 justify the value we want to read but is available in one insn on
1770 for (new_mode
= smallest_mode_for_size (access_size
* BITS_PER_UNIT
,
1772 GET_MODE_BITSIZE (new_mode
) <= BITS_PER_WORD
;
1773 new_mode
= GET_MODE_WIDER_MODE (new_mode
))
1775 rtx target
, new_reg
, new_lhs
;
1776 rtx_insn
*shift_seq
, *insn
;
1779 /* If a constant was stored into memory, try to simplify it here,
1780 otherwise the cost of the shift might preclude this optimization
1781 e.g. at -Os, even when no actual shift will be needed. */
1782 if (store_info
->const_rhs
)
1784 unsigned int byte
= subreg_lowpart_offset (new_mode
, store_mode
);
1785 rtx ret
= simplify_subreg (new_mode
, store_info
->const_rhs
,
1787 if (ret
&& CONSTANT_P (ret
))
1789 ret
= simplify_const_binary_operation (LSHIFTRT
, new_mode
,
1790 ret
, GEN_INT (shift
));
1791 if (ret
&& CONSTANT_P (ret
))
1793 byte
= subreg_lowpart_offset (read_mode
, new_mode
);
1794 ret
= simplify_subreg (read_mode
, ret
, new_mode
, byte
);
1795 if (ret
&& CONSTANT_P (ret
)
1796 && set_src_cost (ret
, speed
) <= COSTS_N_INSNS (1))
1805 /* Try a wider mode if truncating the store mode to NEW_MODE
1806 requires a real instruction. */
1807 if (GET_MODE_BITSIZE (new_mode
) < GET_MODE_BITSIZE (store_mode
)
1808 && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode
, store_mode
))
1811 /* Also try a wider mode if the necessary punning is either not
1812 desirable or not possible. */
1813 if (!CONSTANT_P (store_info
->rhs
)
1814 && !MODES_TIEABLE_P (new_mode
, store_mode
))
1817 new_reg
= gen_reg_rtx (new_mode
);
1821 /* In theory we could also check for an ashr. Ian Taylor knows
1822 of one dsp where the cost of these two was not the same. But
1823 this really is a rare case anyway. */
1824 target
= expand_binop (new_mode
, lshr_optab
, new_reg
,
1825 GEN_INT (shift
), new_reg
, 1, OPTAB_DIRECT
);
1827 shift_seq
= get_insns ();
1830 if (target
!= new_reg
|| shift_seq
== NULL
)
1834 for (insn
= shift_seq
; insn
!= NULL_RTX
; insn
= NEXT_INSN (insn
))
1836 cost
+= insn_rtx_cost (PATTERN (insn
), speed
);
1838 /* The computation up to here is essentially independent
1839 of the arguments and could be precomputed. It may
1840 not be worth doing so. We could precompute if
1841 worthwhile or at least cache the results. The result
1842 technically depends on both SHIFT and ACCESS_SIZE,
1843 but in practice the answer will depend only on ACCESS_SIZE. */
1845 if (cost
> COSTS_N_INSNS (1))
1848 new_lhs
= extract_low_bits (new_mode
, store_mode
,
1849 copy_rtx (store_info
->rhs
));
1850 if (new_lhs
== NULL_RTX
)
1853 /* We found an acceptable shift. Generate a move to
1854 take the value from the store and put it into the
1855 shift pseudo, then shift it, then generate another
1856 move to put in into the target of the read. */
1857 emit_move_insn (new_reg
, new_lhs
);
1858 emit_insn (shift_seq
);
1859 read_reg
= extract_low_bits (read_mode
, new_mode
, new_reg
);
1867 /* Call back for note_stores to find the hard regs set or clobbered by
1868 insn. Data is a bitmap of the hardregs set so far. */
1871 look_for_hardregs (rtx x
, const_rtx pat ATTRIBUTE_UNUSED
, void *data
)
1873 bitmap regs_set
= (bitmap
) data
;
1876 && HARD_REGISTER_P (x
))
1878 unsigned int regno
= REGNO (x
);
1879 bitmap_set_range (regs_set
, regno
,
1880 hard_regno_nregs
[regno
][GET_MODE (x
)]);
1884 /* Helper function for replace_read and record_store.
1885 Attempt to return a value stored in STORE_INFO, from READ_BEGIN
1886 to one before READ_END bytes read in READ_MODE. Return NULL
1887 if not successful. If REQUIRE_CST is true, return always constant. */
1890 get_stored_val (store_info_t store_info
, machine_mode read_mode
,
1891 HOST_WIDE_INT read_begin
, HOST_WIDE_INT read_end
,
1892 basic_block bb
, bool require_cst
)
1894 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1896 int access_size
; /* In bytes. */
1899 /* To get here the read is within the boundaries of the write so
1900 shift will never be negative. Start out with the shift being in
1902 if (store_mode
== BLKmode
)
1904 else if (BYTES_BIG_ENDIAN
)
1905 shift
= store_info
->end
- read_end
;
1907 shift
= read_begin
- store_info
->begin
;
1909 access_size
= shift
+ GET_MODE_SIZE (read_mode
);
1911 /* From now on it is bits. */
1912 shift
*= BITS_PER_UNIT
;
1915 read_reg
= find_shift_sequence (access_size
, store_info
, read_mode
, shift
,
1916 optimize_bb_for_speed_p (bb
),
1918 else if (store_mode
== BLKmode
)
1920 /* The store is a memset (addr, const_val, const_size). */
1921 gcc_assert (CONST_INT_P (store_info
->rhs
));
1922 store_mode
= int_mode_for_mode (read_mode
);
1923 if (store_mode
== BLKmode
)
1924 read_reg
= NULL_RTX
;
1925 else if (store_info
->rhs
== const0_rtx
)
1926 read_reg
= extract_low_bits (read_mode
, store_mode
, const0_rtx
);
1927 else if (GET_MODE_BITSIZE (store_mode
) > HOST_BITS_PER_WIDE_INT
1928 || BITS_PER_UNIT
>= HOST_BITS_PER_WIDE_INT
)
1929 read_reg
= NULL_RTX
;
1932 unsigned HOST_WIDE_INT c
1933 = INTVAL (store_info
->rhs
)
1934 & (((HOST_WIDE_INT
) 1 << BITS_PER_UNIT
) - 1);
1935 int shift
= BITS_PER_UNIT
;
1936 while (shift
< HOST_BITS_PER_WIDE_INT
)
1941 read_reg
= gen_int_mode (c
, store_mode
);
1942 read_reg
= extract_low_bits (read_mode
, store_mode
, read_reg
);
1945 else if (store_info
->const_rhs
1947 || GET_MODE_CLASS (read_mode
) != GET_MODE_CLASS (store_mode
)))
1948 read_reg
= extract_low_bits (read_mode
, store_mode
,
1949 copy_rtx (store_info
->const_rhs
));
1951 read_reg
= extract_low_bits (read_mode
, store_mode
,
1952 copy_rtx (store_info
->rhs
));
1953 if (require_cst
&& read_reg
&& !CONSTANT_P (read_reg
))
1954 read_reg
= NULL_RTX
;
1958 /* Take a sequence of:
1981 Depending on the alignment and the mode of the store and
1985 The STORE_INFO and STORE_INSN are for the store and READ_INFO
1986 and READ_INSN are for the read. Return true if the replacement
1990 replace_read (store_info_t store_info
, insn_info_t store_insn
,
1991 read_info_t read_info
, insn_info_t read_insn
, rtx
*loc
,
1994 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1995 machine_mode read_mode
= GET_MODE (read_info
->mem
);
1996 rtx_insn
*insns
, *this_insn
;
2003 /* Create a sequence of instructions to set up the read register.
2004 This sequence goes immediately before the store and its result
2005 is read by the load.
2007 We need to keep this in perspective. We are replacing a read
2008 with a sequence of insns, but the read will almost certainly be
2009 in cache, so it is not going to be an expensive one. Thus, we
2010 are not willing to do a multi insn shift or worse a subroutine
2011 call to get rid of the read. */
2012 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2013 fprintf (dump_file
, "trying to replace %smode load in insn %d"
2014 " from %smode store in insn %d\n",
2015 GET_MODE_NAME (read_mode
), INSN_UID (read_insn
->insn
),
2016 GET_MODE_NAME (store_mode
), INSN_UID (store_insn
->insn
));
2018 bb
= BLOCK_FOR_INSN (read_insn
->insn
);
2019 read_reg
= get_stored_val (store_info
,
2020 read_mode
, read_info
->begin
, read_info
->end
,
2022 if (read_reg
== NULL_RTX
)
2025 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2026 fprintf (dump_file
, " -- could not extract bits of stored value\n");
2029 /* Force the value into a new register so that it won't be clobbered
2030 between the store and the load. */
2031 read_reg
= copy_to_mode_reg (read_mode
, read_reg
);
2032 insns
= get_insns ();
2035 if (insns
!= NULL_RTX
)
2037 /* Now we have to scan the set of new instructions to see if the
2038 sequence contains and sets of hardregs that happened to be
2039 live at this point. For instance, this can happen if one of
2040 the insns sets the CC and the CC happened to be live at that
2041 point. This does occasionally happen, see PR 37922. */
2042 bitmap regs_set
= BITMAP_ALLOC (®_obstack
);
2044 for (this_insn
= insns
; this_insn
!= NULL_RTX
; this_insn
= NEXT_INSN (this_insn
))
2045 note_stores (PATTERN (this_insn
), look_for_hardregs
, regs_set
);
2047 bitmap_and_into (regs_set
, regs_live
);
2048 if (!bitmap_empty_p (regs_set
))
2050 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2053 "abandoning replacement because sequence clobbers live hardregs:");
2054 df_print_regset (dump_file
, regs_set
);
2057 BITMAP_FREE (regs_set
);
2060 BITMAP_FREE (regs_set
);
2063 if (validate_change (read_insn
->insn
, loc
, read_reg
, 0))
2065 deferred_change_t deferred_change
=
2066 (deferred_change_t
) pool_alloc (deferred_change_pool
);
2068 /* Insert this right before the store insn where it will be safe
2069 from later insns that might change it before the read. */
2070 emit_insn_before (insns
, store_insn
->insn
);
2072 /* And now for the kludge part: cselib croaks if you just
2073 return at this point. There are two reasons for this:
2075 1) Cselib has an idea of how many pseudos there are and
2076 that does not include the new ones we just added.
2078 2) Cselib does not know about the move insn we added
2079 above the store_info, and there is no way to tell it
2080 about it, because it has "moved on".
2082 Problem (1) is fixable with a certain amount of engineering.
2083 Problem (2) is requires starting the bb from scratch. This
2086 So we are just going to have to lie. The move/extraction
2087 insns are not really an issue, cselib did not see them. But
2088 the use of the new pseudo read_insn is a real problem because
2089 cselib has not scanned this insn. The way that we solve this
2090 problem is that we are just going to put the mem back for now
2091 and when we are finished with the block, we undo this. We
2092 keep a table of mems to get rid of. At the end of the basic
2093 block we can put them back. */
2095 *loc
= read_info
->mem
;
2096 deferred_change
->next
= deferred_change_list
;
2097 deferred_change_list
= deferred_change
;
2098 deferred_change
->loc
= loc
;
2099 deferred_change
->reg
= read_reg
;
2101 /* Get rid of the read_info, from the point of view of the
2102 rest of dse, play like this read never happened. */
2103 read_insn
->read_rec
= read_info
->next
;
2104 pool_free (read_info_pool
, read_info
);
2105 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2107 fprintf (dump_file
, " -- replaced the loaded MEM with ");
2108 print_simple_rtl (dump_file
, read_reg
);
2109 fprintf (dump_file
, "\n");
2115 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2117 fprintf (dump_file
, " -- replacing the loaded MEM with ");
2118 print_simple_rtl (dump_file
, read_reg
);
2119 fprintf (dump_file
, " led to an invalid instruction\n");
2125 /* Check the address of MEM *LOC and kill any appropriate stores that may
2129 check_mem_read_rtx (rtx
*loc
, bb_info_t bb_info
)
2131 rtx mem
= *loc
, mem_addr
;
2132 insn_info_t insn_info
;
2133 HOST_WIDE_INT offset
= 0;
2134 HOST_WIDE_INT width
= 0;
2135 alias_set_type spill_alias_set
= 0;
2136 cselib_val
*base
= NULL
;
2138 read_info_t read_info
;
2140 insn_info
= bb_info
->last_insn
;
2142 if ((MEM_ALIAS_SET (mem
) == ALIAS_SET_MEMORY_BARRIER
)
2143 || (MEM_VOLATILE_P (mem
)))
2145 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2146 fprintf (dump_file
, " adding wild read, volatile or barrier.\n");
2147 add_wild_read (bb_info
);
2148 insn_info
->cannot_delete
= true;
2152 /* If it is reading readonly mem, then there can be no conflict with
2154 if (MEM_READONLY_P (mem
))
2157 if (!canon_address (mem
, &spill_alias_set
, &group_id
, &offset
, &base
))
2159 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2160 fprintf (dump_file
, " adding wild read, canon_address failure.\n");
2161 add_wild_read (bb_info
);
2165 if (GET_MODE (mem
) == BLKmode
)
2168 width
= GET_MODE_SIZE (GET_MODE (mem
));
2170 read_info
= (read_info_t
) pool_alloc (read_info_pool
);
2171 read_info
->group_id
= group_id
;
2172 read_info
->mem
= mem
;
2173 read_info
->alias_set
= spill_alias_set
;
2174 read_info
->begin
= offset
;
2175 read_info
->end
= offset
+ width
;
2176 read_info
->next
= insn_info
->read_rec
;
2177 insn_info
->read_rec
= read_info
;
2178 /* For alias_set != 0 canon_true_dependence should be never called. */
2179 if (spill_alias_set
)
2180 mem_addr
= NULL_RTX
;
2184 mem_addr
= base
->val_rtx
;
2188 = rtx_group_vec
[group_id
];
2189 mem_addr
= group
->canon_base_addr
;
2192 mem_addr
= plus_constant (get_address_mode (mem
), mem_addr
, offset
);
2195 /* We ignore the clobbers in store_info. The is mildly aggressive,
2196 but there really should not be a clobber followed by a read. */
2198 if (spill_alias_set
)
2200 insn_info_t i_ptr
= active_local_stores
;
2201 insn_info_t last
= NULL
;
2203 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2204 fprintf (dump_file
, " processing spill load %d\n",
2205 (int) spill_alias_set
);
2209 store_info_t store_info
= i_ptr
->store_rec
;
2211 /* Skip the clobbers. */
2212 while (!store_info
->is_set
)
2213 store_info
= store_info
->next
;
2215 if (store_info
->alias_set
== spill_alias_set
)
2217 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2218 dump_insn_info ("removing from active", i_ptr
);
2220 active_local_stores_len
--;
2222 last
->next_local_store
= i_ptr
->next_local_store
;
2224 active_local_stores
= i_ptr
->next_local_store
;
2228 i_ptr
= i_ptr
->next_local_store
;
2231 else if (group_id
>= 0)
2233 /* This is the restricted case where the base is a constant or
2234 the frame pointer and offset is a constant. */
2235 insn_info_t i_ptr
= active_local_stores
;
2236 insn_info_t last
= NULL
;
2238 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2241 fprintf (dump_file
, " processing const load gid=%d[BLK]\n",
2244 fprintf (dump_file
, " processing const load gid=%d[%d..%d)\n",
2245 group_id
, (int)offset
, (int)(offset
+width
));
2250 bool remove
= false;
2251 store_info_t store_info
= i_ptr
->store_rec
;
2253 /* Skip the clobbers. */
2254 while (!store_info
->is_set
)
2255 store_info
= store_info
->next
;
2257 /* There are three cases here. */
2258 if (store_info
->group_id
< 0)
2259 /* We have a cselib store followed by a read from a
2262 = canon_true_dependence (store_info
->mem
,
2263 GET_MODE (store_info
->mem
),
2264 store_info
->mem_addr
,
2267 else if (group_id
== store_info
->group_id
)
2269 /* This is a block mode load. We may get lucky and
2270 canon_true_dependence may save the day. */
2273 = canon_true_dependence (store_info
->mem
,
2274 GET_MODE (store_info
->mem
),
2275 store_info
->mem_addr
,
2278 /* If this read is just reading back something that we just
2279 stored, rewrite the read. */
2283 && offset
>= store_info
->begin
2284 && offset
+ width
<= store_info
->end
2285 && all_positions_needed_p (store_info
,
2286 offset
- store_info
->begin
,
2288 && replace_read (store_info
, i_ptr
, read_info
,
2289 insn_info
, loc
, bb_info
->regs_live
))
2292 /* The bases are the same, just see if the offsets
2294 if ((offset
< store_info
->end
)
2295 && (offset
+ width
> store_info
->begin
))
2301 The else case that is missing here is that the
2302 bases are constant but different. There is nothing
2303 to do here because there is no overlap. */
2307 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2308 dump_insn_info ("removing from active", i_ptr
);
2310 active_local_stores_len
--;
2312 last
->next_local_store
= i_ptr
->next_local_store
;
2314 active_local_stores
= i_ptr
->next_local_store
;
2318 i_ptr
= i_ptr
->next_local_store
;
2323 insn_info_t i_ptr
= active_local_stores
;
2324 insn_info_t last
= NULL
;
2325 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2327 fprintf (dump_file
, " processing cselib load mem:");
2328 print_inline_rtx (dump_file
, mem
, 0);
2329 fprintf (dump_file
, "\n");
2334 bool remove
= false;
2335 store_info_t store_info
= i_ptr
->store_rec
;
2337 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2338 fprintf (dump_file
, " processing cselib load against insn %d\n",
2339 INSN_UID (i_ptr
->insn
));
2341 /* Skip the clobbers. */
2342 while (!store_info
->is_set
)
2343 store_info
= store_info
->next
;
2345 /* If this read is just reading back something that we just
2346 stored, rewrite the read. */
2348 && store_info
->group_id
== -1
2349 && store_info
->cse_base
== base
2351 && offset
>= store_info
->begin
2352 && offset
+ width
<= store_info
->end
2353 && all_positions_needed_p (store_info
,
2354 offset
- store_info
->begin
, width
)
2355 && replace_read (store_info
, i_ptr
, read_info
, insn_info
, loc
,
2356 bb_info
->regs_live
))
2359 if (!store_info
->alias_set
)
2360 remove
= canon_true_dependence (store_info
->mem
,
2361 GET_MODE (store_info
->mem
),
2362 store_info
->mem_addr
,
2367 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2368 dump_insn_info ("removing from active", i_ptr
);
2370 active_local_stores_len
--;
2372 last
->next_local_store
= i_ptr
->next_local_store
;
2374 active_local_stores
= i_ptr
->next_local_store
;
2378 i_ptr
= i_ptr
->next_local_store
;
2383 /* A note_uses callback in which DATA points the INSN_INFO for
2384 as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns
2385 true for any part of *LOC. */
2388 check_mem_read_use (rtx
*loc
, void *data
)
2390 subrtx_ptr_iterator::array_type array
;
2391 FOR_EACH_SUBRTX_PTR (iter
, array
, loc
, NONCONST
)
2395 check_mem_read_rtx (loc
, (bb_info_t
) data
);
2400 /* Get arguments passed to CALL_INSN. Return TRUE if successful.
2401 So far it only handles arguments passed in registers. */
2404 get_call_args (rtx call_insn
, tree fn
, rtx
*args
, int nargs
)
2406 CUMULATIVE_ARGS args_so_far_v
;
2407 cumulative_args_t args_so_far
;
2411 INIT_CUMULATIVE_ARGS (args_so_far_v
, TREE_TYPE (fn
), NULL_RTX
, 0, 3);
2412 args_so_far
= pack_cumulative_args (&args_so_far_v
);
2414 arg
= TYPE_ARG_TYPES (TREE_TYPE (fn
));
2416 arg
!= void_list_node
&& idx
< nargs
;
2417 arg
= TREE_CHAIN (arg
), idx
++)
2419 machine_mode mode
= TYPE_MODE (TREE_VALUE (arg
));
2421 reg
= targetm
.calls
.function_arg (args_so_far
, mode
, NULL_TREE
, true);
2422 if (!reg
|| !REG_P (reg
) || GET_MODE (reg
) != mode
2423 || GET_MODE_CLASS (mode
) != MODE_INT
)
2426 for (link
= CALL_INSN_FUNCTION_USAGE (call_insn
);
2428 link
= XEXP (link
, 1))
2429 if (GET_CODE (XEXP (link
, 0)) == USE
)
2431 args
[idx
] = XEXP (XEXP (link
, 0), 0);
2432 if (REG_P (args
[idx
])
2433 && REGNO (args
[idx
]) == REGNO (reg
)
2434 && (GET_MODE (args
[idx
]) == mode
2435 || (GET_MODE_CLASS (GET_MODE (args
[idx
])) == MODE_INT
2436 && (GET_MODE_SIZE (GET_MODE (args
[idx
]))
2438 && (GET_MODE_SIZE (GET_MODE (args
[idx
]))
2439 > GET_MODE_SIZE (mode
)))))
2445 tmp
= cselib_expand_value_rtx (args
[idx
], scratch
, 5);
2446 if (GET_MODE (args
[idx
]) != mode
)
2448 if (!tmp
|| !CONST_INT_P (tmp
))
2450 tmp
= gen_int_mode (INTVAL (tmp
), mode
);
2455 targetm
.calls
.function_arg_advance (args_so_far
, mode
, NULL_TREE
, true);
2457 if (arg
!= void_list_node
|| idx
!= nargs
)
2462 /* Return a bitmap of the fixed registers contained in IN. */
2465 copy_fixed_regs (const_bitmap in
)
2469 ret
= ALLOC_REG_SET (NULL
);
2470 bitmap_and (ret
, in
, fixed_reg_set_regset
);
2474 /* Apply record_store to all candidate stores in INSN. Mark INSN
2475 if some part of it is not a candidate store and assigns to a
2476 non-register target. */
2479 scan_insn (bb_info_t bb_info
, rtx_insn
*insn
)
2482 insn_info_t insn_info
= (insn_info_t
) pool_alloc (insn_info_pool
);
2484 memset (insn_info
, 0, sizeof (struct insn_info
));
2486 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2487 fprintf (dump_file
, "\n**scanning insn=%d\n",
2490 insn_info
->prev_insn
= bb_info
->last_insn
;
2491 insn_info
->insn
= insn
;
2492 bb_info
->last_insn
= insn_info
;
2494 if (DEBUG_INSN_P (insn
))
2496 insn_info
->cannot_delete
= true;
2500 /* Look at all of the uses in the insn. */
2501 note_uses (&PATTERN (insn
), check_mem_read_use
, bb_info
);
2506 tree memset_call
= NULL_TREE
;
2508 insn_info
->cannot_delete
= true;
2510 /* Const functions cannot do anything bad i.e. read memory,
2511 however, they can read their parameters which may have
2512 been pushed onto the stack.
2513 memset and bzero don't read memory either. */
2514 const_call
= RTL_CONST_CALL_P (insn
);
2517 rtx call
= get_call_rtx_from (insn
);
2518 if (call
&& GET_CODE (XEXP (XEXP (call
, 0), 0)) == SYMBOL_REF
)
2520 rtx symbol
= XEXP (XEXP (call
, 0), 0);
2521 if (SYMBOL_REF_DECL (symbol
)
2522 && TREE_CODE (SYMBOL_REF_DECL (symbol
)) == FUNCTION_DECL
)
2524 if ((DECL_BUILT_IN_CLASS (SYMBOL_REF_DECL (symbol
))
2526 && (DECL_FUNCTION_CODE (SYMBOL_REF_DECL (symbol
))
2527 == BUILT_IN_MEMSET
))
2528 || SYMBOL_REF_DECL (symbol
) == block_clear_fn
)
2529 memset_call
= SYMBOL_REF_DECL (symbol
);
2533 if (const_call
|| memset_call
)
2535 insn_info_t i_ptr
= active_local_stores
;
2536 insn_info_t last
= NULL
;
2538 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2539 fprintf (dump_file
, "%s call %d\n",
2540 const_call
? "const" : "memset", INSN_UID (insn
));
2542 /* See the head comment of the frame_read field. */
2543 if (reload_completed
2544 /* Tail calls are storing their arguments using
2545 arg pointer. If it is a frame pointer on the target,
2546 even before reload we need to kill frame pointer based
2548 || (SIBLING_CALL_P (insn
)
2549 && HARD_FRAME_POINTER_IS_ARG_POINTER
))
2550 insn_info
->frame_read
= true;
2552 /* Loop over the active stores and remove those which are
2553 killed by the const function call. */
2556 bool remove_store
= false;
2558 /* The stack pointer based stores are always killed. */
2559 if (i_ptr
->stack_pointer_based
)
2560 remove_store
= true;
2562 /* If the frame is read, the frame related stores are killed. */
2563 else if (insn_info
->frame_read
)
2565 store_info_t store_info
= i_ptr
->store_rec
;
2567 /* Skip the clobbers. */
2568 while (!store_info
->is_set
)
2569 store_info
= store_info
->next
;
2571 if (store_info
->group_id
>= 0
2572 && rtx_group_vec
[store_info
->group_id
]->frame_related
)
2573 remove_store
= true;
2578 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2579 dump_insn_info ("removing from active", i_ptr
);
2581 active_local_stores_len
--;
2583 last
->next_local_store
= i_ptr
->next_local_store
;
2585 active_local_stores
= i_ptr
->next_local_store
;
2590 i_ptr
= i_ptr
->next_local_store
;
2596 if (get_call_args (insn
, memset_call
, args
, 3)
2597 && CONST_INT_P (args
[1])
2598 && CONST_INT_P (args
[2])
2599 && INTVAL (args
[2]) > 0)
2601 rtx mem
= gen_rtx_MEM (BLKmode
, args
[0]);
2602 set_mem_size (mem
, INTVAL (args
[2]));
2603 body
= gen_rtx_SET (VOIDmode
, mem
, args
[1]);
2604 mems_found
+= record_store (body
, bb_info
);
2605 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2606 fprintf (dump_file
, "handling memset as BLKmode store\n");
2607 if (mems_found
== 1)
2609 if (active_local_stores_len
++
2610 >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES
))
2612 active_local_stores_len
= 1;
2613 active_local_stores
= NULL
;
2615 insn_info
->fixed_regs_live
2616 = copy_fixed_regs (bb_info
->regs_live
);
2617 insn_info
->next_local_store
= active_local_stores
;
2618 active_local_stores
= insn_info
;
2623 else if (SIBLING_CALL_P (insn
) && reload_completed
)
2624 /* Arguments for a sibling call that are pushed to memory are passed
2625 using the incoming argument pointer of the current function. After
2626 reload that might be (and likely is) frame pointer based. */
2627 add_wild_read (bb_info
);
2629 /* Every other call, including pure functions, may read any memory
2630 that is not relative to the frame. */
2631 add_non_frame_wild_read (bb_info
);
2636 /* Assuming that there are sets in these insns, we cannot delete
2638 if ((GET_CODE (PATTERN (insn
)) == CLOBBER
)
2639 || volatile_refs_p (PATTERN (insn
))
2640 || (!cfun
->can_delete_dead_exceptions
&& !insn_nothrow_p (insn
))
2641 || (RTX_FRAME_RELATED_P (insn
))
2642 || find_reg_note (insn
, REG_FRAME_RELATED_EXPR
, NULL_RTX
))
2643 insn_info
->cannot_delete
= true;
2645 body
= PATTERN (insn
);
2646 if (GET_CODE (body
) == PARALLEL
)
2649 for (i
= 0; i
< XVECLEN (body
, 0); i
++)
2650 mems_found
+= record_store (XVECEXP (body
, 0, i
), bb_info
);
2653 mems_found
+= record_store (body
, bb_info
);
2655 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2656 fprintf (dump_file
, "mems_found = %d, cannot_delete = %s\n",
2657 mems_found
, insn_info
->cannot_delete
? "true" : "false");
2659 /* If we found some sets of mems, add it into the active_local_stores so
2660 that it can be locally deleted if found dead or used for
2661 replace_read and redundant constant store elimination. Otherwise mark
2662 it as cannot delete. This simplifies the processing later. */
2663 if (mems_found
== 1)
2665 if (active_local_stores_len
++
2666 >= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES
))
2668 active_local_stores_len
= 1;
2669 active_local_stores
= NULL
;
2671 insn_info
->fixed_regs_live
= copy_fixed_regs (bb_info
->regs_live
);
2672 insn_info
->next_local_store
= active_local_stores
;
2673 active_local_stores
= insn_info
;
2676 insn_info
->cannot_delete
= true;
2680 /* Remove BASE from the set of active_local_stores. This is a
2681 callback from cselib that is used to get rid of the stores in
2682 active_local_stores. */
2685 remove_useless_values (cselib_val
*base
)
2687 insn_info_t insn_info
= active_local_stores
;
2688 insn_info_t last
= NULL
;
2692 store_info_t store_info
= insn_info
->store_rec
;
2695 /* If ANY of the store_infos match the cselib group that is
2696 being deleted, then the insn can not be deleted. */
2699 if ((store_info
->group_id
== -1)
2700 && (store_info
->cse_base
== base
))
2705 store_info
= store_info
->next
;
2710 active_local_stores_len
--;
2712 last
->next_local_store
= insn_info
->next_local_store
;
2714 active_local_stores
= insn_info
->next_local_store
;
2715 free_store_info (insn_info
);
2720 insn_info
= insn_info
->next_local_store
;
2725 /* Do all of step 1. */
2731 bitmap regs_live
= BITMAP_ALLOC (®_obstack
);
2734 all_blocks
= BITMAP_ALLOC (NULL
);
2735 bitmap_set_bit (all_blocks
, ENTRY_BLOCK
);
2736 bitmap_set_bit (all_blocks
, EXIT_BLOCK
);
2738 FOR_ALL_BB_FN (bb
, cfun
)
2741 bb_info_t bb_info
= (bb_info_t
) pool_alloc (bb_info_pool
);
2743 memset (bb_info
, 0, sizeof (struct dse_bb_info
));
2744 bitmap_set_bit (all_blocks
, bb
->index
);
2745 bb_info
->regs_live
= regs_live
;
2747 bitmap_copy (regs_live
, DF_LR_IN (bb
));
2748 df_simulate_initialize_forwards (bb
, regs_live
);
2750 bb_table
[bb
->index
] = bb_info
;
2751 cselib_discard_hook
= remove_useless_values
;
2753 if (bb
->index
>= NUM_FIXED_BLOCKS
)
2758 = create_alloc_pool ("cse_store_info_pool",
2759 sizeof (struct store_info
), 100);
2760 active_local_stores
= NULL
;
2761 active_local_stores_len
= 0;
2762 cselib_clear_table ();
2764 /* Scan the insns. */
2765 FOR_BB_INSNS (bb
, insn
)
2768 scan_insn (bb_info
, insn
);
2769 cselib_process_insn (insn
);
2771 df_simulate_one_insn_forwards (bb
, insn
, regs_live
);
2774 /* This is something of a hack, because the global algorithm
2775 is supposed to take care of the case where stores go dead
2776 at the end of the function. However, the global
2777 algorithm must take a more conservative view of block
2778 mode reads than the local alg does. So to get the case
2779 where you have a store to the frame followed by a non
2780 overlapping block more read, we look at the active local
2781 stores at the end of the function and delete all of the
2782 frame and spill based ones. */
2783 if (stores_off_frame_dead_at_return
2784 && (EDGE_COUNT (bb
->succs
) == 0
2785 || (single_succ_p (bb
)
2786 && single_succ (bb
) == EXIT_BLOCK_PTR_FOR_FN (cfun
)
2787 && ! crtl
->calls_eh_return
)))
2789 insn_info_t i_ptr
= active_local_stores
;
2792 store_info_t store_info
= i_ptr
->store_rec
;
2794 /* Skip the clobbers. */
2795 while (!store_info
->is_set
)
2796 store_info
= store_info
->next
;
2797 if (store_info
->alias_set
&& !i_ptr
->cannot_delete
)
2798 delete_dead_store_insn (i_ptr
);
2800 if (store_info
->group_id
>= 0)
2803 = rtx_group_vec
[store_info
->group_id
];
2804 if (group
->frame_related
&& !i_ptr
->cannot_delete
)
2805 delete_dead_store_insn (i_ptr
);
2808 i_ptr
= i_ptr
->next_local_store
;
2812 /* Get rid of the loads that were discovered in
2813 replace_read. Cselib is finished with this block. */
2814 while (deferred_change_list
)
2816 deferred_change_t next
= deferred_change_list
->next
;
2818 /* There is no reason to validate this change. That was
2820 *deferred_change_list
->loc
= deferred_change_list
->reg
;
2821 pool_free (deferred_change_pool
, deferred_change_list
);
2822 deferred_change_list
= next
;
2825 /* Get rid of all of the cselib based store_infos in this
2826 block and mark the containing insns as not being
2828 ptr
= bb_info
->last_insn
;
2831 if (ptr
->contains_cselib_groups
)
2833 store_info_t s_info
= ptr
->store_rec
;
2834 while (s_info
&& !s_info
->is_set
)
2835 s_info
= s_info
->next
;
2837 && s_info
->redundant_reason
2838 && s_info
->redundant_reason
->insn
2839 && !ptr
->cannot_delete
)
2841 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2842 fprintf (dump_file
, "Locally deleting insn %d "
2843 "because insn %d stores the "
2844 "same value and couldn't be "
2846 INSN_UID (ptr
->insn
),
2847 INSN_UID (s_info
->redundant_reason
->insn
));
2848 delete_dead_store_insn (ptr
);
2850 free_store_info (ptr
);
2854 store_info_t s_info
;
2856 /* Free at least positions_needed bitmaps. */
2857 for (s_info
= ptr
->store_rec
; s_info
; s_info
= s_info
->next
)
2858 if (s_info
->is_large
)
2860 BITMAP_FREE (s_info
->positions_needed
.large
.bmap
);
2861 s_info
->is_large
= false;
2864 ptr
= ptr
->prev_insn
;
2867 free_alloc_pool (cse_store_info_pool
);
2869 bb_info
->regs_live
= NULL
;
2872 BITMAP_FREE (regs_live
);
2874 rtx_group_table
->empty ();
2878 /*----------------------------------------------------------------------------
2881 Assign each byte position in the stores that we are going to
2882 analyze globally to a position in the bitmaps. Returns true if
2883 there are any bit positions assigned.
2884 ----------------------------------------------------------------------------*/
2887 dse_step2_init (void)
2892 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2894 /* For all non stack related bases, we only consider a store to
2895 be deletable if there are two or more stores for that
2896 position. This is because it takes one store to make the
2897 other store redundant. However, for the stores that are
2898 stack related, we consider them if there is only one store
2899 for the position. We do this because the stack related
2900 stores can be deleted if their is no read between them and
2901 the end of the function.
2903 To make this work in the current framework, we take the stack
2904 related bases add all of the bits from store1 into store2.
2905 This has the effect of making the eligible even if there is
2908 if (stores_off_frame_dead_at_return
&& group
->frame_related
)
2910 bitmap_ior_into (group
->store2_n
, group
->store1_n
);
2911 bitmap_ior_into (group
->store2_p
, group
->store1_p
);
2912 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2913 fprintf (dump_file
, "group %d is frame related ", i
);
2916 group
->offset_map_size_n
++;
2917 group
->offset_map_n
= XOBNEWVEC (&dse_obstack
, int,
2918 group
->offset_map_size_n
);
2919 group
->offset_map_size_p
++;
2920 group
->offset_map_p
= XOBNEWVEC (&dse_obstack
, int,
2921 group
->offset_map_size_p
);
2922 group
->process_globally
= false;
2923 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2925 fprintf (dump_file
, "group %d(%d+%d): ", i
,
2926 (int)bitmap_count_bits (group
->store2_n
),
2927 (int)bitmap_count_bits (group
->store2_p
));
2928 bitmap_print (dump_file
, group
->store2_n
, "n ", " ");
2929 bitmap_print (dump_file
, group
->store2_p
, "p ", "\n");
2935 /* Init the offset tables for the normal case. */
2938 dse_step2_nospill (void)
2942 /* Position 0 is unused because 0 is used in the maps to mean
2944 current_position
= 1;
2945 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2950 if (group
== clear_alias_group
)
2953 memset (group
->offset_map_n
, 0, sizeof (int) * group
->offset_map_size_n
);
2954 memset (group
->offset_map_p
, 0, sizeof (int) * group
->offset_map_size_p
);
2955 bitmap_clear (group
->group_kill
);
2957 EXECUTE_IF_SET_IN_BITMAP (group
->store2_n
, 0, j
, bi
)
2959 bitmap_set_bit (group
->group_kill
, current_position
);
2960 if (bitmap_bit_p (group
->escaped_n
, j
))
2961 bitmap_set_bit (kill_on_calls
, current_position
);
2962 group
->offset_map_n
[j
] = current_position
++;
2963 group
->process_globally
= true;
2965 EXECUTE_IF_SET_IN_BITMAP (group
->store2_p
, 0, j
, bi
)
2967 bitmap_set_bit (group
->group_kill
, current_position
);
2968 if (bitmap_bit_p (group
->escaped_p
, j
))
2969 bitmap_set_bit (kill_on_calls
, current_position
);
2970 group
->offset_map_p
[j
] = current_position
++;
2971 group
->process_globally
= true;
2974 return current_position
!= 1;
2979 /*----------------------------------------------------------------------------
2982 Build the bit vectors for the transfer functions.
2983 ----------------------------------------------------------------------------*/
2986 /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not
2990 get_bitmap_index (group_info_t group_info
, HOST_WIDE_INT offset
)
2994 HOST_WIDE_INT offset_p
= -offset
;
2995 if (offset_p
>= group_info
->offset_map_size_n
)
2997 return group_info
->offset_map_n
[offset_p
];
3001 if (offset
>= group_info
->offset_map_size_p
)
3003 return group_info
->offset_map_p
[offset
];
3008 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL
3012 scan_stores_nospill (store_info_t store_info
, bitmap gen
, bitmap kill
)
3017 group_info_t group_info
3018 = rtx_group_vec
[store_info
->group_id
];
3019 if (group_info
->process_globally
)
3020 for (i
= store_info
->begin
; i
< store_info
->end
; i
++)
3022 int index
= get_bitmap_index (group_info
, i
);
3025 bitmap_set_bit (gen
, index
);
3027 bitmap_clear_bit (kill
, index
);
3030 store_info
= store_info
->next
;
3035 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL
3039 scan_stores_spill (store_info_t store_info
, bitmap gen
, bitmap kill
)
3043 if (store_info
->alias_set
)
3045 int index
= get_bitmap_index (clear_alias_group
,
3046 store_info
->alias_set
);
3049 bitmap_set_bit (gen
, index
);
3051 bitmap_clear_bit (kill
, index
);
3054 store_info
= store_info
->next
;
3059 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL
3063 scan_reads_nospill (insn_info_t insn_info
, bitmap gen
, bitmap kill
)
3065 read_info_t read_info
= insn_info
->read_rec
;
3069 /* If this insn reads the frame, kill all the frame related stores. */
3070 if (insn_info
->frame_read
)
3072 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3073 if (group
->process_globally
&& group
->frame_related
)
3076 bitmap_ior_into (kill
, group
->group_kill
);
3077 bitmap_and_compl_into (gen
, group
->group_kill
);
3080 if (insn_info
->non_frame_wild_read
)
3082 /* Kill all non-frame related stores. Kill all stores of variables that
3085 bitmap_ior_into (kill
, kill_on_calls
);
3086 bitmap_and_compl_into (gen
, kill_on_calls
);
3087 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3088 if (group
->process_globally
&& !group
->frame_related
)
3091 bitmap_ior_into (kill
, group
->group_kill
);
3092 bitmap_and_compl_into (gen
, group
->group_kill
);
3097 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3099 if (group
->process_globally
)
3101 if (i
== read_info
->group_id
)
3103 if (read_info
->begin
> read_info
->end
)
3105 /* Begin > end for block mode reads. */
3107 bitmap_ior_into (kill
, group
->group_kill
);
3108 bitmap_and_compl_into (gen
, group
->group_kill
);
3112 /* The groups are the same, just process the
3115 for (j
= read_info
->begin
; j
< read_info
->end
; j
++)
3117 int index
= get_bitmap_index (group
, j
);
3121 bitmap_set_bit (kill
, index
);
3122 bitmap_clear_bit (gen
, index
);
3129 /* The groups are different, if the alias sets
3130 conflict, clear the entire group. We only need
3131 to apply this test if the read_info is a cselib
3132 read. Anything with a constant base cannot alias
3133 something else with a different constant
3135 if ((read_info
->group_id
< 0)
3136 && canon_true_dependence (group
->base_mem
,
3137 GET_MODE (group
->base_mem
),
3138 group
->canon_base_addr
,
3139 read_info
->mem
, NULL_RTX
))
3142 bitmap_ior_into (kill
, group
->group_kill
);
3143 bitmap_and_compl_into (gen
, group
->group_kill
);
3149 read_info
= read_info
->next
;
3153 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL
3157 scan_reads_spill (read_info_t read_info
, bitmap gen
, bitmap kill
)
3161 if (read_info
->alias_set
)
3163 int index
= get_bitmap_index (clear_alias_group
,
3164 read_info
->alias_set
);
3168 bitmap_set_bit (kill
, index
);
3169 bitmap_clear_bit (gen
, index
);
3173 read_info
= read_info
->next
;
3178 /* Return the insn in BB_INFO before the first wild read or if there
3179 are no wild reads in the block, return the last insn. */
3182 find_insn_before_first_wild_read (bb_info_t bb_info
)
3184 insn_info_t insn_info
= bb_info
->last_insn
;
3185 insn_info_t last_wild_read
= NULL
;
3189 if (insn_info
->wild_read
)
3191 last_wild_read
= insn_info
->prev_insn
;
3192 /* Block starts with wild read. */
3193 if (!last_wild_read
)
3197 insn_info
= insn_info
->prev_insn
;
3201 return last_wild_read
;
3203 return bb_info
->last_insn
;
3207 /* Scan the insns in BB_INFO starting at PTR and going to the top of
3208 the block in order to build the gen and kill sets for the block.
3209 We start at ptr which may be the last insn in the block or may be
3210 the first insn with a wild read. In the latter case we are able to
3211 skip the rest of the block because it just does not matter:
3212 anything that happens is hidden by the wild read. */
3215 dse_step3_scan (bool for_spills
, basic_block bb
)
3217 bb_info_t bb_info
= bb_table
[bb
->index
];
3218 insn_info_t insn_info
;
3221 /* There are no wild reads in the spill case. */
3222 insn_info
= bb_info
->last_insn
;
3224 insn_info
= find_insn_before_first_wild_read (bb_info
);
3226 /* In the spill case or in the no_spill case if there is no wild
3227 read in the block, we will need a kill set. */
3228 if (insn_info
== bb_info
->last_insn
)
3231 bitmap_clear (bb_info
->kill
);
3233 bb_info
->kill
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3237 BITMAP_FREE (bb_info
->kill
);
3241 /* There may have been code deleted by the dce pass run before
3243 if (insn_info
->insn
&& INSN_P (insn_info
->insn
))
3245 /* Process the read(s) last. */
3248 scan_stores_spill (insn_info
->store_rec
, bb_info
->gen
, bb_info
->kill
);
3249 scan_reads_spill (insn_info
->read_rec
, bb_info
->gen
, bb_info
->kill
);
3253 scan_stores_nospill (insn_info
->store_rec
, bb_info
->gen
, bb_info
->kill
);
3254 scan_reads_nospill (insn_info
, bb_info
->gen
, bb_info
->kill
);
3258 insn_info
= insn_info
->prev_insn
;
3263 /* Set the gen set of the exit block, and also any block with no
3264 successors that does not have a wild read. */
3267 dse_step3_exit_block_scan (bb_info_t bb_info
)
3269 /* The gen set is all 0's for the exit block except for the
3270 frame_pointer_group. */
3272 if (stores_off_frame_dead_at_return
)
3277 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3279 if (group
->process_globally
&& group
->frame_related
)
3280 bitmap_ior_into (bb_info
->gen
, group
->group_kill
);
3286 /* Find all of the blocks that are not backwards reachable from the
3287 exit block or any block with no successors (BB). These are the
3288 infinite loops or infinite self loops. These blocks will still
3289 have their bits set in UNREACHABLE_BLOCKS. */
3292 mark_reachable_blocks (sbitmap unreachable_blocks
, basic_block bb
)
3297 if (bitmap_bit_p (unreachable_blocks
, bb
->index
))
3299 bitmap_clear_bit (unreachable_blocks
, bb
->index
);
3300 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3302 mark_reachable_blocks (unreachable_blocks
, e
->src
);
3307 /* Build the transfer functions for the function. */
3310 dse_step3 (bool for_spills
)
3313 sbitmap unreachable_blocks
= sbitmap_alloc (last_basic_block_for_fn (cfun
));
3314 sbitmap_iterator sbi
;
3315 bitmap all_ones
= NULL
;
3318 bitmap_ones (unreachable_blocks
);
3320 FOR_ALL_BB_FN (bb
, cfun
)
3322 bb_info_t bb_info
= bb_table
[bb
->index
];
3324 bitmap_clear (bb_info
->gen
);
3326 bb_info
->gen
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3328 if (bb
->index
== ENTRY_BLOCK
)
3330 else if (bb
->index
== EXIT_BLOCK
)
3331 dse_step3_exit_block_scan (bb_info
);
3333 dse_step3_scan (for_spills
, bb
);
3334 if (EDGE_COUNT (bb
->succs
) == 0)
3335 mark_reachable_blocks (unreachable_blocks
, bb
);
3337 /* If this is the second time dataflow is run, delete the old
3340 BITMAP_FREE (bb_info
->in
);
3342 BITMAP_FREE (bb_info
->out
);
3345 /* For any block in an infinite loop, we must initialize the out set
3346 to all ones. This could be expensive, but almost never occurs in
3347 practice. However, it is common in regression tests. */
3348 EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks
, 0, i
, sbi
)
3350 if (bitmap_bit_p (all_blocks
, i
))
3352 bb_info_t bb_info
= bb_table
[i
];
3358 all_ones
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3359 FOR_EACH_VEC_ELT (rtx_group_vec
, j
, group
)
3360 bitmap_ior_into (all_ones
, group
->group_kill
);
3364 bb_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3365 bitmap_copy (bb_info
->out
, all_ones
);
3371 BITMAP_FREE (all_ones
);
3372 sbitmap_free (unreachable_blocks
);
3377 /*----------------------------------------------------------------------------
3380 Solve the bitvector equations.
3381 ----------------------------------------------------------------------------*/
3384 /* Confluence function for blocks with no successors. Create an out
3385 set from the gen set of the exit block. This block logically has
3386 the exit block as a successor. */
3391 dse_confluence_0 (basic_block bb
)
3393 bb_info_t bb_info
= bb_table
[bb
->index
];
3395 if (bb
->index
== EXIT_BLOCK
)
3400 bb_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3401 bitmap_copy (bb_info
->out
, bb_table
[EXIT_BLOCK
]->gen
);
3405 /* Propagate the information from the in set of the dest of E to the
3406 out set of the src of E. If the various in or out sets are not
3407 there, that means they are all ones. */
3410 dse_confluence_n (edge e
)
3412 bb_info_t src_info
= bb_table
[e
->src
->index
];
3413 bb_info_t dest_info
= bb_table
[e
->dest
->index
];
3418 bitmap_and_into (src_info
->out
, dest_info
->in
);
3421 src_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3422 bitmap_copy (src_info
->out
, dest_info
->in
);
3429 /* Propagate the info from the out to the in set of BB_INDEX's basic
3430 block. There are three cases:
3432 1) The block has no kill set. In this case the kill set is all
3433 ones. It does not matter what the out set of the block is, none of
3434 the info can reach the top. The only thing that reaches the top is
3435 the gen set and we just copy the set.
3437 2) There is a kill set but no out set and bb has successors. In
3438 this case we just return. Eventually an out set will be created and
3439 it is better to wait than to create a set of ones.
3441 3) There is both a kill and out set. We apply the obvious transfer
3446 dse_transfer_function (int bb_index
)
3448 bb_info_t bb_info
= bb_table
[bb_index
];
3456 return bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
3457 bb_info
->out
, bb_info
->kill
);
3460 bb_info
->in
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3461 bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
3462 bb_info
->out
, bb_info
->kill
);
3472 /* Case 1 above. If there is already an in set, nothing
3478 bb_info
->in
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3479 bitmap_copy (bb_info
->in
, bb_info
->gen
);
3485 /* Solve the dataflow equations. */
3490 df_simple_dataflow (DF_BACKWARD
, NULL
, dse_confluence_0
,
3491 dse_confluence_n
, dse_transfer_function
,
3492 all_blocks
, df_get_postorder (DF_BACKWARD
),
3493 df_get_n_blocks (DF_BACKWARD
));
3494 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3498 fprintf (dump_file
, "\n\n*** Global dataflow info after analysis.\n");
3499 FOR_ALL_BB_FN (bb
, cfun
)
3501 bb_info_t bb_info
= bb_table
[bb
->index
];
3503 df_print_bb_index (bb
, dump_file
);
3505 bitmap_print (dump_file
, bb_info
->in
, " in: ", "\n");
3507 fprintf (dump_file
, " in: *MISSING*\n");
3509 bitmap_print (dump_file
, bb_info
->gen
, " gen: ", "\n");
3511 fprintf (dump_file
, " gen: *MISSING*\n");
3513 bitmap_print (dump_file
, bb_info
->kill
, " kill: ", "\n");
3515 fprintf (dump_file
, " kill: *MISSING*\n");
3517 bitmap_print (dump_file
, bb_info
->out
, " out: ", "\n");
3519 fprintf (dump_file
, " out: *MISSING*\n\n");
3526 /*----------------------------------------------------------------------------
3529 Delete the stores that can only be deleted using the global information.
3530 ----------------------------------------------------------------------------*/
3534 dse_step5_nospill (void)
3537 FOR_EACH_BB_FN (bb
, cfun
)
3539 bb_info_t bb_info
= bb_table
[bb
->index
];
3540 insn_info_t insn_info
= bb_info
->last_insn
;
3541 bitmap v
= bb_info
->out
;
3545 bool deleted
= false;
3546 if (dump_file
&& insn_info
->insn
)
3548 fprintf (dump_file
, "starting to process insn %d\n",
3549 INSN_UID (insn_info
->insn
));
3550 bitmap_print (dump_file
, v
, " v: ", "\n");
3553 /* There may have been code deleted by the dce pass run before
3556 && INSN_P (insn_info
->insn
)
3557 && (!insn_info
->cannot_delete
)
3558 && (!bitmap_empty_p (v
)))
3560 store_info_t store_info
= insn_info
->store_rec
;
3562 /* Try to delete the current insn. */
3565 /* Skip the clobbers. */
3566 while (!store_info
->is_set
)
3567 store_info
= store_info
->next
;
3569 if (store_info
->alias_set
)
3574 group_info_t group_info
3575 = rtx_group_vec
[store_info
->group_id
];
3577 for (i
= store_info
->begin
; i
< store_info
->end
; i
++)
3579 int index
= get_bitmap_index (group_info
, i
);
3581 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3582 fprintf (dump_file
, "i = %d, index = %d\n", (int)i
, index
);
3583 if (index
== 0 || !bitmap_bit_p (v
, index
))
3585 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3586 fprintf (dump_file
, "failing at i = %d\n", (int)i
);
3595 && check_for_inc_dec_1 (insn_info
))
3597 delete_insn (insn_info
->insn
);
3598 insn_info
->insn
= NULL
;
3603 /* We do want to process the local info if the insn was
3604 deleted. For instance, if the insn did a wild read, we
3605 no longer need to trash the info. */
3607 && INSN_P (insn_info
->insn
)
3610 scan_stores_nospill (insn_info
->store_rec
, v
, NULL
);
3611 if (insn_info
->wild_read
)
3613 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3614 fprintf (dump_file
, "wild read\n");
3617 else if (insn_info
->read_rec
3618 || insn_info
->non_frame_wild_read
)
3620 if (dump_file
&& !insn_info
->non_frame_wild_read
)
3621 fprintf (dump_file
, "regular read\n");
3622 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3623 fprintf (dump_file
, "non-frame wild read\n");
3624 scan_reads_nospill (insn_info
, v
, NULL
);
3628 insn_info
= insn_info
->prev_insn
;
3635 /*----------------------------------------------------------------------------
3638 Delete stores made redundant by earlier stores (which store the same
3639 value) that couldn't be eliminated.
3640 ----------------------------------------------------------------------------*/
3647 FOR_ALL_BB_FN (bb
, cfun
)
3649 bb_info_t bb_info
= bb_table
[bb
->index
];
3650 insn_info_t insn_info
= bb_info
->last_insn
;
3654 /* There may have been code deleted by the dce pass run before
3657 && INSN_P (insn_info
->insn
)
3658 && !insn_info
->cannot_delete
)
3660 store_info_t s_info
= insn_info
->store_rec
;
3662 while (s_info
&& !s_info
->is_set
)
3663 s_info
= s_info
->next
;
3665 && s_info
->redundant_reason
3666 && s_info
->redundant_reason
->insn
3667 && INSN_P (s_info
->redundant_reason
->insn
))
3669 rtx_insn
*rinsn
= s_info
->redundant_reason
->insn
;
3670 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3671 fprintf (dump_file
, "Locally deleting insn %d "
3672 "because insn %d stores the "
3673 "same value and couldn't be "
3675 INSN_UID (insn_info
->insn
),
3677 delete_dead_store_insn (insn_info
);
3680 insn_info
= insn_info
->prev_insn
;
3685 /*----------------------------------------------------------------------------
3688 Destroy everything left standing.
3689 ----------------------------------------------------------------------------*/
3694 bitmap_obstack_release (&dse_bitmap_obstack
);
3695 obstack_free (&dse_obstack
, NULL
);
3697 end_alias_analysis ();
3699 delete rtx_group_table
;
3700 rtx_group_table
= NULL
;
3701 rtx_group_vec
.release ();
3702 BITMAP_FREE (all_blocks
);
3703 BITMAP_FREE (scratch
);
3705 free_alloc_pool (rtx_store_info_pool
);
3706 free_alloc_pool (read_info_pool
);
3707 free_alloc_pool (insn_info_pool
);
3708 free_alloc_pool (bb_info_pool
);
3709 free_alloc_pool (rtx_group_info_pool
);
3710 free_alloc_pool (deferred_change_pool
);
3714 /* -------------------------------------------------------------------------
3716 ------------------------------------------------------------------------- */
3718 /* Callback for running pass_rtl_dse. */
3721 rest_of_handle_dse (void)
3723 df_set_flags (DF_DEFER_INSN_RESCAN
);
3725 /* Need the notes since we must track live hardregs in the forwards
3727 df_note_add_problem ();
3733 if (dse_step2_nospill ())
3735 df_set_flags (DF_LR_RUN_DCE
);
3737 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3738 fprintf (dump_file
, "doing global processing\n");
3741 dse_step5_nospill ();
3748 fprintf (dump_file
, "dse: local deletions = %d, global deletions = %d, spill deletions = %d\n",
3749 locally_deleted
, globally_deleted
, spill_deleted
);
3751 /* DSE can eliminate potentially-trapping MEMs.
3752 Remove any EH edges associated with them. */
3753 if ((locally_deleted
|| globally_deleted
)
3754 && cfun
->can_throw_non_call_exceptions
3755 && purge_all_dead_edges ())
3763 const pass_data pass_data_rtl_dse1
=
3765 RTL_PASS
, /* type */
3767 OPTGROUP_NONE
, /* optinfo_flags */
3768 TV_DSE1
, /* tv_id */
3769 0, /* properties_required */
3770 0, /* properties_provided */
3771 0, /* properties_destroyed */
3772 0, /* todo_flags_start */
3773 TODO_df_finish
, /* todo_flags_finish */
3776 class pass_rtl_dse1
: public rtl_opt_pass
3779 pass_rtl_dse1 (gcc::context
*ctxt
)
3780 : rtl_opt_pass (pass_data_rtl_dse1
, ctxt
)
3783 /* opt_pass methods: */
3784 virtual bool gate (function
*)
3786 return optimize
> 0 && flag_dse
&& dbg_cnt (dse1
);
3789 virtual unsigned int execute (function
*) { return rest_of_handle_dse (); }
3791 }; // class pass_rtl_dse1
3796 make_pass_rtl_dse1 (gcc::context
*ctxt
)
3798 return new pass_rtl_dse1 (ctxt
);
3803 const pass_data pass_data_rtl_dse2
=
3805 RTL_PASS
, /* type */
3807 OPTGROUP_NONE
, /* optinfo_flags */
3808 TV_DSE2
, /* tv_id */
3809 0, /* properties_required */
3810 0, /* properties_provided */
3811 0, /* properties_destroyed */
3812 0, /* todo_flags_start */
3813 TODO_df_finish
, /* todo_flags_finish */
3816 class pass_rtl_dse2
: public rtl_opt_pass
3819 pass_rtl_dse2 (gcc::context
*ctxt
)
3820 : rtl_opt_pass (pass_data_rtl_dse2
, ctxt
)
3823 /* opt_pass methods: */
3824 virtual bool gate (function
*)
3826 return optimize
> 0 && flag_dse
&& dbg_cnt (dse2
);
3829 virtual unsigned int execute (function
*) { return rest_of_handle_dse (); }
3831 }; // class pass_rtl_dse2
3836 make_pass_rtl_dse2 (gcc::context
*ctxt
)
3838 return new pass_rtl_dse2 (ctxt
);