1 /* RTL dead store elimination.
2 Copyright (C) 2005-2019 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"
37 #include "gimple-ssa.h"
43 #include "stor-layout.h"
46 #include "tree-pass.h"
51 #include "cfgcleanup.h"
54 /* This file contains three techniques for performing Dead Store
57 * The first technique performs dse locally on any base address. It
58 is based on the cselib which is a local value numbering technique.
59 This technique is local to a basic block but deals with a fairly
62 * The second technique performs dse globally but is restricted to
63 base addresses that are either constant or are relative to the
66 * The third technique, (which is only done after register allocation)
67 processes the spill slots. This differs from the second
68 technique because it takes advantage of the fact that spilling is
69 completely free from the effects of aliasing.
71 Logically, dse is a backwards dataflow problem. A store can be
72 deleted if it if cannot be reached in the backward direction by any
73 use of the value being stored. However, the local technique uses a
74 forwards scan of the basic block because cselib requires that the
75 block be processed in that order.
77 The pass is logically broken into 7 steps:
81 1) The local algorithm, as well as scanning the insns for the two
84 2) Analysis to see if the global algs are necessary. In the case
85 of stores base on a constant address, there must be at least two
86 stores to that address, to make it possible to delete some of the
87 stores. In the case of stores off of the frame or spill related
88 stores, only one store to an address is necessary because those
89 stores die at the end of the function.
91 3) Set up the global dataflow equations based on processing the
92 info parsed in the first step.
94 4) Solve the dataflow equations.
96 5) Delete the insns that the global analysis has indicated are
99 6) Delete insns that store the same value as preceding store
100 where the earlier store couldn't be eliminated.
104 This step uses cselib and canon_rtx to build the largest expression
105 possible for each address. This pass is a forwards pass through
106 each basic block. From the point of view of the global technique,
107 the first pass could examine a block in either direction. The
108 forwards ordering is to accommodate cselib.
110 We make a simplifying assumption: addresses fall into four broad
113 1) base has rtx_varies_p == false, offset is constant.
114 2) base has rtx_varies_p == false, offset variable.
115 3) base has rtx_varies_p == true, offset constant.
116 4) base has rtx_varies_p == true, offset variable.
118 The local passes are able to process all 4 kinds of addresses. The
119 global pass only handles 1).
121 The global problem is formulated as follows:
123 A store, S1, to address A, where A is not relative to the stack
124 frame, can be eliminated if all paths from S1 to the end of the
125 function contain another store to A before a read to A.
127 If the address A is relative to the stack frame, a store S2 to A
128 can be eliminated if there are no paths from S2 that reach the
129 end of the function that read A before another store to A. In
130 this case S2 can be deleted if there are paths from S2 to the
131 end of the function that have no reads or writes to A. This
132 second case allows stores to the stack frame to be deleted that
133 would otherwise die when the function returns. This cannot be
134 done if stores_off_frame_dead_at_return is not true. See the doc
135 for that variable for when this variable is false.
137 The global problem is formulated as a backwards set union
138 dataflow problem where the stores are the gens and reads are the
139 kills. Set union problems are rare and require some special
140 handling given our representation of bitmaps. A straightforward
141 implementation requires a lot of bitmaps filled with 1s.
142 These are expensive and cumbersome in our bitmap formulation so
143 care has been taken to avoid large vectors filled with 1s. See
144 the comments in bb_info and in the dataflow confluence functions
147 There are two places for further enhancements to this algorithm:
149 1) The original dse which was embedded in a pass called flow also
150 did local address forwarding. For example in
155 flow would replace the right hand side of the second insn with a
156 reference to r100. Most of the information is available to add this
157 to this pass. It has not done it because it is a lot of work in
158 the case that either r100 is assigned to between the first and
159 second insn and/or the second insn is a load of part of the value
160 stored by the first insn.
162 insn 5 in gcc.c-torture/compile/990203-1.c simple case.
163 insn 15 in gcc.c-torture/execute/20001017-2.c simple case.
164 insn 25 in gcc.c-torture/execute/20001026-1.c simple case.
165 insn 44 in gcc.c-torture/execute/20010910-1.c simple case.
167 2) The cleaning up of spill code is quite profitable. It currently
168 depends on reading tea leaves and chicken entrails left by reload.
169 This pass depends on reload creating a singleton alias set for each
170 spill slot and telling the next dse pass which of these alias sets
171 are the singletons. Rather than analyze the addresses of the
172 spills, dse's spill processing just does analysis of the loads and
173 stores that use those alias sets. There are three cases where this
176 a) Reload sometimes creates the slot for one mode of access, and
177 then inserts loads and/or stores for a smaller mode. In this
178 case, the current code just punts on the slot. The proper thing
179 to do is to back out and use one bit vector position for each
180 byte of the entity associated with the slot. This depends on
181 KNOWING that reload always generates the accesses for each of the
182 bytes in some canonical (read that easy to understand several
183 passes after reload happens) way.
185 b) Reload sometimes decides that spill slot it allocated was not
186 large enough for the mode and goes back and allocates more slots
187 with the same mode and alias set. The backout in this case is a
188 little more graceful than (a). In this case the slot is unmarked
189 as being a spill slot and if final address comes out to be based
190 off the frame pointer, the global algorithm handles this slot.
192 c) For any pass that may prespill, there is currently no
193 mechanism to tell the dse pass that the slot being used has the
194 special properties that reload uses. It may be that all that is
195 required is to have those passes make the same calls that reload
196 does, assuming that the alias sets can be manipulated in the same
199 /* There are limits to the size of constant offsets we model for the
200 global problem. There are certainly test cases, that exceed this
201 limit, however, it is unlikely that there are important programs
202 that really have constant offsets this size. */
203 #define MAX_OFFSET (64 * 1024)
205 /* Obstack for the DSE dataflow bitmaps. We don't want to put these
206 on the default obstack because these bitmaps can grow quite large
207 (~2GB for the small (!) test case of PR54146) and we'll hold on to
208 all that memory until the end of the compiler run.
209 As a bonus, delete_tree_live_info can destroy all the bitmaps by just
210 releasing the whole obstack. */
211 static bitmap_obstack dse_bitmap_obstack
;
213 /* Obstack for other data. As for above: Kinda nice to be able to
214 throw it all away at the end in one big sweep. */
215 static struct obstack dse_obstack
;
217 /* Scratch bitmap for cselib's cselib_expand_value_rtx. */
218 static bitmap scratch
= NULL
;
220 struct insn_info_type
;
222 /* This structure holds information about a candidate store. */
227 /* False means this is a clobber. */
230 /* False if a single HOST_WIDE_INT bitmap is used for positions_needed. */
233 /* The id of the mem group of the base address. If rtx_varies_p is
234 true, this is -1. Otherwise, it is the index into the group
238 /* This is the cselib value. */
239 cselib_val
*cse_base
;
241 /* This canonized mem. */
244 /* Canonized MEM address for use by canon_true_dependence. */
247 /* The offset of the first byte associated with the operation. */
250 /* The number of bytes covered by the operation. This is always exact
251 and known (rather than -1). */
256 /* A bitmask as wide as the number of bytes in the word that
257 contains a 1 if the byte may be needed. The store is unused if
258 all of the bits are 0. This is used if IS_LARGE is false. */
259 unsigned HOST_WIDE_INT small_bitmask
;
263 /* A bitmap with one bit per byte, or null if the number of
264 bytes isn't known at compile time. A cleared bit means
265 the position is needed. Used if IS_LARGE is true. */
268 /* When BITMAP is nonnull, this counts the number of set bits
269 (i.e. unneeded bytes) in the bitmap. If it is equal to
270 WIDTH, the whole store is unused.
273 - the store is definitely not needed when COUNT == 1
274 - all the store is needed when COUNT == 0 and RHS is nonnull
275 - otherwise we don't know which parts of the store are needed. */
280 /* The next store info for this insn. */
281 class store_info
*next
;
283 /* The right hand side of the store. This is used if there is a
284 subsequent reload of the mems address somewhere later in the
288 /* If rhs is or holds a constant, this contains that constant,
292 /* Set if this store stores the same constant value as REDUNDANT_REASON
293 insn stored. These aren't eliminated early, because doing that
294 might prevent the earlier larger store to be eliminated. */
295 struct insn_info_type
*redundant_reason
;
298 /* Return a bitmask with the first N low bits set. */
300 static unsigned HOST_WIDE_INT
301 lowpart_bitmask (int n
)
303 unsigned HOST_WIDE_INT mask
= HOST_WIDE_INT_M1U
;
304 return mask
>> (HOST_BITS_PER_WIDE_INT
- n
);
307 static object_allocator
<store_info
> cse_store_info_pool ("cse_store_info_pool");
309 static object_allocator
<store_info
> rtx_store_info_pool ("rtx_store_info_pool");
311 /* This structure holds information about a load. These are only
312 built for rtx bases. */
316 /* The id of the mem group of the base address. */
319 /* The offset of the first byte associated with the operation. */
322 /* The number of bytes covered by the operation, or -1 if not known. */
325 /* The mem being read. */
328 /* The next read_info for this insn. */
329 class read_info_type
*next
;
331 typedef class read_info_type
*read_info_t
;
333 static object_allocator
<read_info_type
> read_info_type_pool ("read_info_pool");
335 /* One of these records is created for each insn. */
337 struct insn_info_type
339 /* Set true if the insn contains a store but the insn itself cannot
340 be deleted. This is set if the insn is a parallel and there is
341 more than one non dead output or if the insn is in some way
345 /* This field is only used by the global algorithm. It is set true
346 if the insn contains any read of mem except for a (1). This is
347 also set if the insn is a call or has a clobber mem. If the insn
348 contains a wild read, the use_rec will be null. */
351 /* This is true only for CALL instructions which could potentially read
352 any non-frame memory location. This field is used by the global
354 bool non_frame_wild_read
;
356 /* This field is only used for the processing of const functions.
357 These functions cannot read memory, but they can read the stack
358 because that is where they may get their parms. We need to be
359 this conservative because, like the store motion pass, we don't
360 consider CALL_INSN_FUNCTION_USAGE when processing call insns.
361 Moreover, we need to distinguish two cases:
362 1. Before reload (register elimination), the stores related to
363 outgoing arguments are stack pointer based and thus deemed
364 of non-constant base in this pass. This requires special
365 handling but also means that the frame pointer based stores
366 need not be killed upon encountering a const function call.
367 2. After reload, the stores related to outgoing arguments can be
368 either stack pointer or hard frame pointer based. This means
369 that we have no other choice than also killing all the frame
370 pointer based stores upon encountering a const function call.
371 This field is set after reload for const function calls and before
372 reload for const tail function calls on targets where arg pointer
373 is the frame pointer. Having this set is less severe than a wild
374 read, it just means that all the frame related stores are killed
375 rather than all the stores. */
378 /* This field is only used for the processing of const functions.
379 It is set if the insn may contain a stack pointer based store. */
380 bool stack_pointer_based
;
382 /* This is true if any of the sets within the store contains a
383 cselib base. Such stores can only be deleted by the local
385 bool contains_cselib_groups
;
390 /* The list of mem sets or mem clobbers that are contained in this
391 insn. If the insn is deletable, it contains only one mem set.
392 But it could also contain clobbers. Insns that contain more than
393 one mem set are not deletable, but each of those mems are here in
394 order to provide info to delete other insns. */
395 store_info
*store_rec
;
397 /* The linked list of mem uses in this insn. Only the reads from
398 rtx bases are listed here. The reads to cselib bases are
399 completely processed during the first scan and so are never
401 read_info_t read_rec
;
403 /* The live fixed registers. We assume only fixed registers can
404 cause trouble by being clobbered from an expanded pattern;
405 storing only the live fixed registers (rather than all registers)
406 means less memory needs to be allocated / copied for the individual
408 regset fixed_regs_live
;
410 /* The prev insn in the basic block. */
411 struct insn_info_type
* prev_insn
;
413 /* The linked list of insns that are in consideration for removal in
414 the forwards pass through the basic block. This pointer may be
415 trash as it is not cleared when a wild read occurs. The only
416 time it is guaranteed to be correct is when the traversal starts
417 at active_local_stores. */
418 struct insn_info_type
* next_local_store
;
420 typedef struct insn_info_type
*insn_info_t
;
422 static object_allocator
<insn_info_type
> insn_info_type_pool ("insn_info_pool");
424 /* The linked list of stores that are under consideration in this
426 static insn_info_t active_local_stores
;
427 static int active_local_stores_len
;
429 struct dse_bb_info_type
431 /* Pointer to the insn info for the last insn in the block. These
432 are linked so this is how all of the insns are reached. During
433 scanning this is the current insn being scanned. */
434 insn_info_t last_insn
;
436 /* The info for the global dataflow problem. */
439 /* This is set if the transfer function should and in the wild_read
440 bitmap before applying the kill and gen sets. That vector knocks
441 out most of the bits in the bitmap and thus speeds up the
443 bool apply_wild_read
;
445 /* The following 4 bitvectors hold information about which positions
446 of which stores are live or dead. They are indexed by
449 /* The set of store positions that exist in this block before a wild read. */
452 /* The set of load positions that exist in this block above the
453 same position of a store. */
456 /* The set of stores that reach the top of the block without being
459 Do not represent the in if it is all ones. Note that this is
460 what the bitvector should logically be initialized to for a set
461 intersection problem. However, like the kill set, this is too
462 expensive. So initially, the in set will only be created for the
463 exit block and any block that contains a wild read. */
466 /* The set of stores that reach the bottom of the block from it's
469 Do not represent the in if it is all ones. Note that this is
470 what the bitvector should logically be initialized to for a set
471 intersection problem. However, like the kill and in set, this is
472 too expensive. So what is done is that the confluence operator
473 just initializes the vector from one of the out sets of the
474 successors of the block. */
477 /* The following bitvector is indexed by the reg number. It
478 contains the set of regs that are live at the current instruction
479 being processed. While it contains info for all of the
480 registers, only the hard registers are actually examined. It is used
481 to assure that shift and/or add sequences that are inserted do not
482 accidentally clobber live hard regs. */
486 typedef struct dse_bb_info_type
*bb_info_t
;
488 static object_allocator
<dse_bb_info_type
> dse_bb_info_type_pool
491 /* Table to hold all bb_infos. */
492 static bb_info_t
*bb_table
;
494 /* There is a group_info for each rtx base that is used to reference
495 memory. There are also not many of the rtx bases because they are
496 very limited in scope. */
500 /* The actual base of the address. */
503 /* The sequential id of the base. This allows us to have a
504 canonical ordering of these that is not based on addresses. */
507 /* True if there are any positions that are to be processed
509 bool process_globally
;
511 /* True if the base of this group is either the frame_pointer or
512 hard_frame_pointer. */
515 /* A mem wrapped around the base pointer for the group in order to do
516 read dependency. It must be given BLKmode in order to encompass all
517 the possible offsets from the base. */
520 /* Canonized version of base_mem's address. */
523 /* These two sets of two bitmaps are used to keep track of how many
524 stores are actually referencing that position from this base. We
525 only do this for rtx bases as this will be used to assign
526 positions in the bitmaps for the global problem. Bit N is set in
527 store1 on the first store for offset N. Bit N is set in store2
528 for the second store to offset N. This is all we need since we
529 only care about offsets that have two or more stores for them.
531 The "_n" suffix is for offsets less than 0 and the "_p" suffix is
532 for 0 and greater offsets.
534 There is one special case here, for stores into the stack frame,
535 we will or store1 into store2 before deciding which stores look
536 at globally. This is because stores to the stack frame that have
537 no other reads before the end of the function can also be
539 bitmap store1_n
, store1_p
, store2_n
, store2_p
;
541 /* These bitmaps keep track of offsets in this group escape this function.
542 An offset escapes if it corresponds to a named variable whose
543 addressable flag is set. */
544 bitmap escaped_n
, escaped_p
;
546 /* The positions in this bitmap have the same assignments as the in,
547 out, gen and kill bitmaps. This bitmap is all zeros except for
548 the positions that are occupied by stores for this group. */
551 /* The offset_map is used to map the offsets from this base into
552 positions in the global bitmaps. It is only created after all of
553 the all of stores have been scanned and we know which ones we
555 int *offset_map_n
, *offset_map_p
;
556 int offset_map_size_n
, offset_map_size_p
;
559 static object_allocator
<group_info
> group_info_pool ("rtx_group_info_pool");
561 /* Index into the rtx_group_vec. */
562 static int rtx_group_next_id
;
565 static vec
<group_info
*> rtx_group_vec
;
568 /* This structure holds the set of changes that are being deferred
569 when removing read operation. See replace_read. */
570 struct deferred_change
573 /* The mem that is being replaced. */
576 /* The reg it is being replaced with. */
579 struct deferred_change
*next
;
582 static object_allocator
<deferred_change
> deferred_change_pool
583 ("deferred_change_pool");
585 static deferred_change
*deferred_change_list
= NULL
;
587 /* This is true except if cfun->stdarg -- i.e. we cannot do
588 this for vararg functions because they play games with the frame. */
589 static bool stores_off_frame_dead_at_return
;
591 /* Counter for stats. */
592 static int globally_deleted
;
593 static int locally_deleted
;
595 static bitmap all_blocks
;
597 /* Locations that are killed by calls in the global phase. */
598 static bitmap kill_on_calls
;
600 /* The number of bits used in the global bitmaps. */
601 static unsigned int current_position
;
603 /* Print offset range [OFFSET, OFFSET + WIDTH) to FILE. */
606 print_range (FILE *file
, poly_int64 offset
, poly_int64 width
)
609 print_dec (offset
, file
, SIGNED
);
610 fprintf (file
, "..");
611 print_dec (offset
+ width
, file
, SIGNED
);
615 /*----------------------------------------------------------------------------
619 ----------------------------------------------------------------------------*/
622 /* Hashtable callbacks for maintaining the "bases" field of
623 store_group_info, given that the addresses are function invariants. */
625 struct invariant_group_base_hasher
: nofree_ptr_hash
<group_info
>
627 static inline hashval_t
hash (const group_info
*);
628 static inline bool equal (const group_info
*, const group_info
*);
632 invariant_group_base_hasher::equal (const group_info
*gi1
,
633 const group_info
*gi2
)
635 return rtx_equal_p (gi1
->rtx_base
, gi2
->rtx_base
);
639 invariant_group_base_hasher::hash (const group_info
*gi
)
642 return hash_rtx (gi
->rtx_base
, Pmode
, &do_not_record
, NULL
, false);
645 /* Tables of group_info structures, hashed by base value. */
646 static hash_table
<invariant_group_base_hasher
> *rtx_group_table
;
649 /* Get the GROUP for BASE. Add a new group if it is not there. */
652 get_group_info (rtx base
)
654 struct group_info tmp_gi
;
658 gcc_assert (base
!= NULL_RTX
);
660 /* Find the store_base_info structure for BASE, creating a new one
662 tmp_gi
.rtx_base
= base
;
663 slot
= rtx_group_table
->find_slot (&tmp_gi
, INSERT
);
668 *slot
= gi
= group_info_pool
.allocate ();
670 gi
->id
= rtx_group_next_id
++;
671 gi
->base_mem
= gen_rtx_MEM (BLKmode
, base
);
672 gi
->canon_base_addr
= canon_rtx (base
);
673 gi
->store1_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
674 gi
->store1_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
675 gi
->store2_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
676 gi
->store2_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
677 gi
->escaped_p
= BITMAP_ALLOC (&dse_bitmap_obstack
);
678 gi
->escaped_n
= BITMAP_ALLOC (&dse_bitmap_obstack
);
679 gi
->group_kill
= BITMAP_ALLOC (&dse_bitmap_obstack
);
680 gi
->process_globally
= false;
682 (base
== frame_pointer_rtx
) || (base
== hard_frame_pointer_rtx
);
683 gi
->offset_map_size_n
= 0;
684 gi
->offset_map_size_p
= 0;
685 gi
->offset_map_n
= NULL
;
686 gi
->offset_map_p
= NULL
;
687 rtx_group_vec
.safe_push (gi
);
694 /* Initialization of data structures. */
700 globally_deleted
= 0;
702 bitmap_obstack_initialize (&dse_bitmap_obstack
);
703 gcc_obstack_init (&dse_obstack
);
705 scratch
= BITMAP_ALLOC (®_obstack
);
706 kill_on_calls
= BITMAP_ALLOC (&dse_bitmap_obstack
);
709 rtx_group_table
= new hash_table
<invariant_group_base_hasher
> (11);
711 bb_table
= XNEWVEC (bb_info_t
, last_basic_block_for_fn (cfun
));
712 rtx_group_next_id
= 0;
714 stores_off_frame_dead_at_return
= !cfun
->stdarg
;
716 init_alias_analysis ();
721 /*----------------------------------------------------------------------------
724 Scan all of the insns. Any random ordering of the blocks is fine.
725 Each block is scanned in forward order to accommodate cselib which
726 is used to remove stores with non-constant bases.
727 ----------------------------------------------------------------------------*/
729 /* Delete all of the store_info recs from INSN_INFO. */
732 free_store_info (insn_info_t insn_info
)
734 store_info
*cur
= insn_info
->store_rec
;
737 store_info
*next
= cur
->next
;
739 BITMAP_FREE (cur
->positions_needed
.large
.bmap
);
741 cse_store_info_pool
.remove (cur
);
743 rtx_store_info_pool
.remove (cur
);
747 insn_info
->cannot_delete
= true;
748 insn_info
->contains_cselib_groups
= false;
749 insn_info
->store_rec
= NULL
;
752 struct note_add_store_info
754 rtx_insn
*first
, *current
;
755 regset fixed_regs_live
;
759 /* Callback for emit_inc_dec_insn_before via note_stores.
760 Check if a register is clobbered which is live afterwards. */
763 note_add_store (rtx loc
, const_rtx expr ATTRIBUTE_UNUSED
, void *data
)
766 note_add_store_info
*info
= (note_add_store_info
*) data
;
771 /* If this register is referenced by the current or an earlier insn,
772 that's OK. E.g. this applies to the register that is being incremented
773 with this addition. */
774 for (insn
= info
->first
;
775 insn
!= NEXT_INSN (info
->current
);
776 insn
= NEXT_INSN (insn
))
777 if (reg_referenced_p (loc
, PATTERN (insn
)))
780 /* If we come here, we have a clobber of a register that's only OK
781 if that register is not live. If we don't have liveness information
782 available, fail now. */
783 if (!info
->fixed_regs_live
)
785 info
->failure
= true;
788 /* Now check if this is a live fixed register. */
789 unsigned int end_regno
= END_REGNO (loc
);
790 for (unsigned int regno
= REGNO (loc
); regno
< end_regno
; ++regno
)
791 if (REGNO_REG_SET_P (info
->fixed_regs_live
, regno
))
792 info
->failure
= true;
795 /* Callback for for_each_inc_dec that emits an INSN that sets DEST to
796 SRC + SRCOFF before insn ARG. */
799 emit_inc_dec_insn_before (rtx mem ATTRIBUTE_UNUSED
,
800 rtx op ATTRIBUTE_UNUSED
,
801 rtx dest
, rtx src
, rtx srcoff
, void *arg
)
803 insn_info_t insn_info
= (insn_info_t
) arg
;
804 rtx_insn
*insn
= insn_info
->insn
, *new_insn
, *cur
;
805 note_add_store_info info
;
807 /* We can reuse all operands without copying, because we are about
808 to delete the insn that contained it. */
812 emit_insn (gen_add3_insn (dest
, src
, srcoff
));
813 new_insn
= get_insns ();
817 new_insn
= gen_move_insn (dest
, src
);
818 info
.first
= new_insn
;
819 info
.fixed_regs_live
= insn_info
->fixed_regs_live
;
820 info
.failure
= false;
821 for (cur
= new_insn
; cur
; cur
= NEXT_INSN (cur
))
824 note_stores (cur
, note_add_store
, &info
);
827 /* If a failure was flagged above, return 1 so that for_each_inc_dec will
828 return it immediately, communicating the failure to its caller. */
832 emit_insn_before (new_insn
, insn
);
837 /* Before we delete INSN_INFO->INSN, make sure that the auto inc/dec, if it
838 is there, is split into a separate insn.
839 Return true on success (or if there was nothing to do), false on failure. */
842 check_for_inc_dec_1 (insn_info_t insn_info
)
844 rtx_insn
*insn
= insn_info
->insn
;
845 rtx note
= find_reg_note (insn
, REG_INC
, NULL_RTX
);
847 return for_each_inc_dec (PATTERN (insn
), emit_inc_dec_insn_before
,
853 /* Entry point for postreload. If you work on reload_cse, or you need this
854 anywhere else, consider if you can provide register liveness information
855 and add a parameter to this function so that it can be passed down in
856 insn_info.fixed_regs_live. */
858 check_for_inc_dec (rtx_insn
*insn
)
860 insn_info_type insn_info
;
863 insn_info
.insn
= insn
;
864 insn_info
.fixed_regs_live
= NULL
;
865 note
= find_reg_note (insn
, REG_INC
, NULL_RTX
);
867 return for_each_inc_dec (PATTERN (insn
), emit_inc_dec_insn_before
,
872 /* Delete the insn and free all of the fields inside INSN_INFO. */
875 delete_dead_store_insn (insn_info_t insn_info
)
877 read_info_t read_info
;
882 if (!check_for_inc_dec_1 (insn_info
))
884 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
885 fprintf (dump_file
, "Locally deleting insn %d\n",
886 INSN_UID (insn_info
->insn
));
888 free_store_info (insn_info
);
889 read_info
= insn_info
->read_rec
;
893 read_info_t next
= read_info
->next
;
894 read_info_type_pool
.remove (read_info
);
897 insn_info
->read_rec
= NULL
;
899 delete_insn (insn_info
->insn
);
901 insn_info
->insn
= NULL
;
903 insn_info
->wild_read
= false;
906 /* Return whether DECL, a local variable, can possibly escape the current
910 local_variable_can_escape (tree decl
)
912 if (TREE_ADDRESSABLE (decl
))
915 /* If this is a partitioned variable, we need to consider all the variables
916 in the partition. This is necessary because a store into one of them can
917 be replaced with a store into another and this may not change the outcome
918 of the escape analysis. */
919 if (cfun
->gimple_df
->decls_to_pointers
!= NULL
)
921 tree
*namep
= cfun
->gimple_df
->decls_to_pointers
->get (decl
);
923 return TREE_ADDRESSABLE (*namep
);
929 /* Return whether EXPR can possibly escape the current function scope. */
932 can_escape (tree expr
)
937 base
= get_base_address (expr
);
939 && !may_be_aliased (base
)
941 && !DECL_EXTERNAL (base
)
942 && !TREE_STATIC (base
)
943 && local_variable_can_escape (base
)))
948 /* Set the store* bitmaps offset_map_size* fields in GROUP based on
952 set_usage_bits (group_info
*group
, poly_int64 offset
, poly_int64 width
,
955 /* Non-constant offsets and widths act as global kills, so there's no point
956 trying to use them to derive global DSE candidates. */
957 HOST_WIDE_INT i
, const_offset
, const_width
;
958 bool expr_escapes
= can_escape (expr
);
959 if (offset
.is_constant (&const_offset
)
960 && width
.is_constant (&const_width
)
961 && const_offset
> -MAX_OFFSET
962 && const_offset
+ const_width
< MAX_OFFSET
)
963 for (i
= const_offset
; i
< const_offset
+ const_width
; ++i
)
971 store1
= group
->store1_n
;
972 store2
= group
->store2_n
;
973 escaped
= group
->escaped_n
;
978 store1
= group
->store1_p
;
979 store2
= group
->store2_p
;
980 escaped
= group
->escaped_p
;
984 if (!bitmap_set_bit (store1
, ai
))
985 bitmap_set_bit (store2
, ai
);
990 if (group
->offset_map_size_n
< ai
)
991 group
->offset_map_size_n
= ai
;
995 if (group
->offset_map_size_p
< ai
)
996 group
->offset_map_size_p
= ai
;
1000 bitmap_set_bit (escaped
, ai
);
1005 reset_active_stores (void)
1007 active_local_stores
= NULL
;
1008 active_local_stores_len
= 0;
1011 /* Free all READ_REC of the LAST_INSN of BB_INFO. */
1014 free_read_records (bb_info_t bb_info
)
1016 insn_info_t insn_info
= bb_info
->last_insn
;
1017 read_info_t
*ptr
= &insn_info
->read_rec
;
1020 read_info_t next
= (*ptr
)->next
;
1021 read_info_type_pool
.remove (*ptr
);
1026 /* Set the BB_INFO so that the last insn is marked as a wild read. */
1029 add_wild_read (bb_info_t bb_info
)
1031 insn_info_t insn_info
= bb_info
->last_insn
;
1032 insn_info
->wild_read
= true;
1033 free_read_records (bb_info
);
1034 reset_active_stores ();
1037 /* Set the BB_INFO so that the last insn is marked as a wild read of
1038 non-frame locations. */
1041 add_non_frame_wild_read (bb_info_t bb_info
)
1043 insn_info_t insn_info
= bb_info
->last_insn
;
1044 insn_info
->non_frame_wild_read
= true;
1045 free_read_records (bb_info
);
1046 reset_active_stores ();
1049 /* Return true if X is a constant or one of the registers that behave
1050 as a constant over the life of a function. This is equivalent to
1051 !rtx_varies_p for memory addresses. */
1054 const_or_frame_p (rtx x
)
1059 if (GET_CODE (x
) == REG
)
1061 /* Note that we have to test for the actual rtx used for the frame
1062 and arg pointers and not just the register number in case we have
1063 eliminated the frame and/or arg pointer and are using it
1065 if (x
== frame_pointer_rtx
|| x
== hard_frame_pointer_rtx
1066 /* The arg pointer varies if it is not a fixed register. */
1067 || (x
== arg_pointer_rtx
&& fixed_regs
[ARG_POINTER_REGNUM
])
1068 || x
== pic_offset_table_rtx
)
1076 /* Take all reasonable action to put the address of MEM into the form
1077 that we can do analysis on.
1079 The gold standard is to get the address into the form: address +
1080 OFFSET where address is something that rtx_varies_p considers a
1081 constant. When we can get the address in this form, we can do
1082 global analysis on it. Note that for constant bases, address is
1083 not actually returned, only the group_id. The address can be
1086 If that fails, we try cselib to get a value we can at least use
1087 locally. If that fails we return false.
1089 The GROUP_ID is set to -1 for cselib bases and the index of the
1090 group for non_varying bases.
1092 FOR_READ is true if this is a mem read and false if not. */
1095 canon_address (rtx mem
,
1100 machine_mode address_mode
= get_address_mode (mem
);
1101 rtx mem_address
= XEXP (mem
, 0);
1102 rtx expanded_address
, address
;
1105 cselib_lookup (mem_address
, address_mode
, 1, GET_MODE (mem
));
1107 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1109 fprintf (dump_file
, " mem: ");
1110 print_inline_rtx (dump_file
, mem_address
, 0);
1111 fprintf (dump_file
, "\n");
1114 /* First see if just canon_rtx (mem_address) is const or frame,
1115 if not, try cselib_expand_value_rtx and call canon_rtx on that. */
1117 for (expanded
= 0; expanded
< 2; expanded
++)
1121 /* Use cselib to replace all of the reg references with the full
1122 expression. This will take care of the case where we have
1124 r_x = base + offset;
1129 val = *(base + offset); */
1131 expanded_address
= cselib_expand_value_rtx (mem_address
,
1134 /* If this fails, just go with the address from first
1136 if (!expanded_address
)
1140 expanded_address
= mem_address
;
1142 /* Split the address into canonical BASE + OFFSET terms. */
1143 address
= canon_rtx (expanded_address
);
1147 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1151 fprintf (dump_file
, "\n after cselib_expand address: ");
1152 print_inline_rtx (dump_file
, expanded_address
, 0);
1153 fprintf (dump_file
, "\n");
1156 fprintf (dump_file
, "\n after canon_rtx address: ");
1157 print_inline_rtx (dump_file
, address
, 0);
1158 fprintf (dump_file
, "\n");
1161 if (GET_CODE (address
) == CONST
)
1162 address
= XEXP (address
, 0);
1164 address
= strip_offset_and_add (address
, offset
);
1166 if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem
))
1167 && const_or_frame_p (address
))
1169 group_info
*group
= get_group_info (address
);
1171 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1173 fprintf (dump_file
, " gid=%d offset=", group
->id
);
1174 print_dec (*offset
, dump_file
);
1175 fprintf (dump_file
, "\n");
1178 *group_id
= group
->id
;
1183 *base
= cselib_lookup (address
, address_mode
, true, GET_MODE (mem
));
1188 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1189 fprintf (dump_file
, " no cselib val - should be a wild read.\n");
1192 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1194 fprintf (dump_file
, " varying cselib base=%u:%u offset = ",
1195 (*base
)->uid
, (*base
)->hash
);
1196 print_dec (*offset
, dump_file
);
1197 fprintf (dump_file
, "\n");
1203 /* Clear the rhs field from the active_local_stores array. */
1206 clear_rhs_from_active_local_stores (void)
1208 insn_info_t ptr
= active_local_stores
;
1212 store_info
*store_info
= ptr
->store_rec
;
1213 /* Skip the clobbers. */
1214 while (!store_info
->is_set
)
1215 store_info
= store_info
->next
;
1217 store_info
->rhs
= NULL
;
1218 store_info
->const_rhs
= NULL
;
1220 ptr
= ptr
->next_local_store
;
1225 /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */
1228 set_position_unneeded (store_info
*s_info
, int pos
)
1230 if (__builtin_expect (s_info
->is_large
, false))
1232 if (bitmap_set_bit (s_info
->positions_needed
.large
.bmap
, pos
))
1233 s_info
->positions_needed
.large
.count
++;
1236 s_info
->positions_needed
.small_bitmask
1237 &= ~(HOST_WIDE_INT_1U
<< pos
);
1240 /* Mark the whole store S_INFO as unneeded. */
1243 set_all_positions_unneeded (store_info
*s_info
)
1245 if (__builtin_expect (s_info
->is_large
, false))
1247 HOST_WIDE_INT width
;
1248 if (s_info
->width
.is_constant (&width
))
1250 bitmap_set_range (s_info
->positions_needed
.large
.bmap
, 0, width
);
1251 s_info
->positions_needed
.large
.count
= width
;
1255 gcc_checking_assert (!s_info
->positions_needed
.large
.bmap
);
1256 s_info
->positions_needed
.large
.count
= 1;
1260 s_info
->positions_needed
.small_bitmask
= HOST_WIDE_INT_0U
;
1263 /* Return TRUE if any bytes from S_INFO store are needed. */
1266 any_positions_needed_p (store_info
*s_info
)
1268 if (__builtin_expect (s_info
->is_large
, false))
1270 HOST_WIDE_INT width
;
1271 if (s_info
->width
.is_constant (&width
))
1273 gcc_checking_assert (s_info
->positions_needed
.large
.bmap
);
1274 return s_info
->positions_needed
.large
.count
< width
;
1278 gcc_checking_assert (!s_info
->positions_needed
.large
.bmap
);
1279 return s_info
->positions_needed
.large
.count
== 0;
1283 return (s_info
->positions_needed
.small_bitmask
!= HOST_WIDE_INT_0U
);
1286 /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO
1287 store are known to be needed. */
1290 all_positions_needed_p (store_info
*s_info
, poly_int64 start
,
1293 gcc_assert (s_info
->rhs
);
1294 if (!s_info
->width
.is_constant ())
1296 gcc_assert (s_info
->is_large
1297 && !s_info
->positions_needed
.large
.bmap
);
1298 return s_info
->positions_needed
.large
.count
== 0;
1301 /* Otherwise, if START and WIDTH are non-constant, we're asking about
1302 a non-constant region of a constant-sized store. We can't say for
1303 sure that all positions are needed. */
1304 HOST_WIDE_INT const_start
, const_width
;
1305 if (!start
.is_constant (&const_start
)
1306 || !width
.is_constant (&const_width
))
1309 if (__builtin_expect (s_info
->is_large
, false))
1311 for (HOST_WIDE_INT i
= const_start
; i
< const_start
+ const_width
; ++i
)
1312 if (bitmap_bit_p (s_info
->positions_needed
.large
.bmap
, i
))
1318 unsigned HOST_WIDE_INT mask
1319 = lowpart_bitmask (const_width
) << const_start
;
1320 return (s_info
->positions_needed
.small_bitmask
& mask
) == mask
;
1325 static rtx
get_stored_val (store_info
*, machine_mode
, poly_int64
,
1326 poly_int64
, basic_block
, bool);
1329 /* BODY is an instruction pattern that belongs to INSN. Return 1 if
1330 there is a candidate store, after adding it to the appropriate
1331 local store group if so. */
1334 record_store (rtx body
, bb_info_t bb_info
)
1336 rtx mem
, rhs
, const_rhs
, mem_addr
;
1337 poly_int64 offset
= 0;
1338 poly_int64 width
= 0;
1339 insn_info_t insn_info
= bb_info
->last_insn
;
1340 store_info
*store_info
= NULL
;
1342 cselib_val
*base
= NULL
;
1343 insn_info_t ptr
, last
, redundant_reason
;
1344 bool store_is_unused
;
1346 if (GET_CODE (body
) != SET
&& GET_CODE (body
) != CLOBBER
)
1349 mem
= SET_DEST (body
);
1351 /* If this is not used, then this cannot be used to keep the insn
1352 from being deleted. On the other hand, it does provide something
1353 that can be used to prove that another store is dead. */
1355 = (find_reg_note (insn_info
->insn
, REG_UNUSED
, mem
) != NULL
);
1357 /* Check whether that value is a suitable memory location. */
1360 /* If the set or clobber is unused, then it does not effect our
1361 ability to get rid of the entire insn. */
1362 if (!store_is_unused
)
1363 insn_info
->cannot_delete
= true;
1367 /* At this point we know mem is a mem. */
1368 if (GET_MODE (mem
) == BLKmode
)
1370 HOST_WIDE_INT const_size
;
1371 if (GET_CODE (XEXP (mem
, 0)) == SCRATCH
)
1373 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1374 fprintf (dump_file
, " adding wild read for (clobber (mem:BLK (scratch))\n");
1375 add_wild_read (bb_info
);
1376 insn_info
->cannot_delete
= true;
1379 /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0))
1380 as memset (addr, 0, 36); */
1381 else if (!MEM_SIZE_KNOWN_P (mem
)
1382 || maybe_le (MEM_SIZE (mem
), 0)
1383 /* This is a limit on the bitmap size, which is only relevant
1384 for constant-sized MEMs. */
1385 || (MEM_SIZE (mem
).is_constant (&const_size
)
1386 && const_size
> MAX_OFFSET
)
1387 || GET_CODE (body
) != SET
1388 || !CONST_INT_P (SET_SRC (body
)))
1390 if (!store_is_unused
)
1392 /* If the set or clobber is unused, then it does not effect our
1393 ability to get rid of the entire insn. */
1394 insn_info
->cannot_delete
= true;
1395 clear_rhs_from_active_local_stores ();
1401 /* We can still process a volatile mem, we just cannot delete it. */
1402 if (MEM_VOLATILE_P (mem
))
1403 insn_info
->cannot_delete
= true;
1405 if (!canon_address (mem
, &group_id
, &offset
, &base
))
1407 clear_rhs_from_active_local_stores ();
1411 if (GET_MODE (mem
) == BLKmode
)
1412 width
= MEM_SIZE (mem
);
1414 width
= GET_MODE_SIZE (GET_MODE (mem
));
1416 if (!endpoint_representable_p (offset
, width
))
1418 clear_rhs_from_active_local_stores ();
1422 if (known_eq (width
, 0))
1427 /* In the restrictive case where the base is a constant or the
1428 frame pointer we can do global analysis. */
1431 = rtx_group_vec
[group_id
];
1432 tree expr
= MEM_EXPR (mem
);
1434 store_info
= rtx_store_info_pool
.allocate ();
1435 set_usage_bits (group
, offset
, width
, expr
);
1437 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1439 fprintf (dump_file
, " processing const base store gid=%d",
1441 print_range (dump_file
, offset
, width
);
1442 fprintf (dump_file
, "\n");
1447 if (may_be_sp_based_p (XEXP (mem
, 0)))
1448 insn_info
->stack_pointer_based
= true;
1449 insn_info
->contains_cselib_groups
= true;
1451 store_info
= cse_store_info_pool
.allocate ();
1454 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1456 fprintf (dump_file
, " processing cselib store ");
1457 print_range (dump_file
, offset
, width
);
1458 fprintf (dump_file
, "\n");
1462 const_rhs
= rhs
= NULL_RTX
;
1463 if (GET_CODE (body
) == SET
1464 /* No place to keep the value after ra. */
1465 && !reload_completed
1466 && (REG_P (SET_SRC (body
))
1467 || GET_CODE (SET_SRC (body
)) == SUBREG
1468 || CONSTANT_P (SET_SRC (body
)))
1469 && !MEM_VOLATILE_P (mem
)
1470 /* Sometimes the store and reload is used for truncation and
1472 && !(FLOAT_MODE_P (GET_MODE (mem
)) && (flag_float_store
)))
1474 rhs
= SET_SRC (body
);
1475 if (CONSTANT_P (rhs
))
1477 else if (body
== PATTERN (insn_info
->insn
))
1479 rtx tem
= find_reg_note (insn_info
->insn
, REG_EQUAL
, NULL_RTX
);
1480 if (tem
&& CONSTANT_P (XEXP (tem
, 0)))
1481 const_rhs
= XEXP (tem
, 0);
1483 if (const_rhs
== NULL_RTX
&& REG_P (rhs
))
1485 rtx tem
= cselib_expand_value_rtx (rhs
, scratch
, 5);
1487 if (tem
&& CONSTANT_P (tem
))
1492 /* Check to see if this stores causes some other stores to be
1494 ptr
= active_local_stores
;
1496 redundant_reason
= NULL
;
1497 mem
= canon_rtx (mem
);
1500 mem_addr
= base
->val_rtx
;
1503 group_info
*group
= rtx_group_vec
[group_id
];
1504 mem_addr
= group
->canon_base_addr
;
1506 if (maybe_ne (offset
, 0))
1507 mem_addr
= plus_constant (get_address_mode (mem
), mem_addr
, offset
);
1511 insn_info_t next
= ptr
->next_local_store
;
1512 class store_info
*s_info
= ptr
->store_rec
;
1515 /* Skip the clobbers. We delete the active insn if this insn
1516 shadows the set. To have been put on the active list, it
1517 has exactly on set. */
1518 while (!s_info
->is_set
)
1519 s_info
= s_info
->next
;
1521 if (s_info
->group_id
== group_id
&& s_info
->cse_base
== base
)
1524 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1526 fprintf (dump_file
, " trying store in insn=%d gid=%d",
1527 INSN_UID (ptr
->insn
), s_info
->group_id
);
1528 print_range (dump_file
, s_info
->offset
, s_info
->width
);
1529 fprintf (dump_file
, "\n");
1532 /* Even if PTR won't be eliminated as unneeded, if both
1533 PTR and this insn store the same constant value, we might
1534 eliminate this insn instead. */
1535 if (s_info
->const_rhs
1537 && known_subrange_p (offset
, width
,
1538 s_info
->offset
, s_info
->width
)
1539 && all_positions_needed_p (s_info
, offset
- s_info
->offset
,
1541 /* We can only remove the later store if the earlier aliases
1542 at least all accesses the later one. */
1543 && (MEM_ALIAS_SET (mem
) == MEM_ALIAS_SET (s_info
->mem
)
1544 || alias_set_subset_of (MEM_ALIAS_SET (mem
),
1545 MEM_ALIAS_SET (s_info
->mem
))))
1547 if (GET_MODE (mem
) == BLKmode
)
1549 if (GET_MODE (s_info
->mem
) == BLKmode
1550 && s_info
->const_rhs
== const_rhs
)
1551 redundant_reason
= ptr
;
1553 else if (s_info
->const_rhs
== const0_rtx
1554 && const_rhs
== const0_rtx
)
1555 redundant_reason
= ptr
;
1560 val
= get_stored_val (s_info
, GET_MODE (mem
), offset
, width
,
1561 BLOCK_FOR_INSN (insn_info
->insn
),
1563 if (get_insns () != NULL
)
1566 if (val
&& rtx_equal_p (val
, const_rhs
))
1567 redundant_reason
= ptr
;
1571 HOST_WIDE_INT begin_unneeded
, const_s_width
, const_width
;
1572 if (known_subrange_p (s_info
->offset
, s_info
->width
, offset
, width
))
1573 /* The new store touches every byte that S_INFO does. */
1574 set_all_positions_unneeded (s_info
);
1575 else if ((offset
- s_info
->offset
).is_constant (&begin_unneeded
)
1576 && s_info
->width
.is_constant (&const_s_width
)
1577 && width
.is_constant (&const_width
))
1579 HOST_WIDE_INT end_unneeded
= begin_unneeded
+ const_width
;
1580 begin_unneeded
= MAX (begin_unneeded
, 0);
1581 end_unneeded
= MIN (end_unneeded
, const_s_width
);
1582 for (i
= begin_unneeded
; i
< end_unneeded
; ++i
)
1583 set_position_unneeded (s_info
, i
);
1587 /* We don't know which parts of S_INFO are needed and
1588 which aren't, so invalidate the RHS. */
1590 s_info
->const_rhs
= NULL
;
1593 else if (s_info
->rhs
)
1594 /* Need to see if it is possible for this store to overwrite
1595 the value of store_info. If it is, set the rhs to NULL to
1596 keep it from being used to remove a load. */
1598 if (canon_output_dependence (s_info
->mem
, true,
1599 mem
, GET_MODE (mem
),
1603 s_info
->const_rhs
= NULL
;
1607 /* An insn can be deleted if every position of every one of
1608 its s_infos is zero. */
1609 if (any_positions_needed_p (s_info
))
1614 insn_info_t insn_to_delete
= ptr
;
1616 active_local_stores_len
--;
1618 last
->next_local_store
= ptr
->next_local_store
;
1620 active_local_stores
= ptr
->next_local_store
;
1622 if (!insn_to_delete
->cannot_delete
)
1623 delete_dead_store_insn (insn_to_delete
);
1631 /* Finish filling in the store_info. */
1632 store_info
->next
= insn_info
->store_rec
;
1633 insn_info
->store_rec
= store_info
;
1634 store_info
->mem
= mem
;
1635 store_info
->mem_addr
= mem_addr
;
1636 store_info
->cse_base
= base
;
1637 HOST_WIDE_INT const_width
;
1638 if (!width
.is_constant (&const_width
))
1640 store_info
->is_large
= true;
1641 store_info
->positions_needed
.large
.count
= 0;
1642 store_info
->positions_needed
.large
.bmap
= NULL
;
1644 else if (const_width
> HOST_BITS_PER_WIDE_INT
)
1646 store_info
->is_large
= true;
1647 store_info
->positions_needed
.large
.count
= 0;
1648 store_info
->positions_needed
.large
.bmap
= BITMAP_ALLOC (&dse_bitmap_obstack
);
1652 store_info
->is_large
= false;
1653 store_info
->positions_needed
.small_bitmask
1654 = lowpart_bitmask (const_width
);
1656 store_info
->group_id
= group_id
;
1657 store_info
->offset
= offset
;
1658 store_info
->width
= width
;
1659 store_info
->is_set
= GET_CODE (body
) == SET
;
1660 store_info
->rhs
= rhs
;
1661 store_info
->const_rhs
= const_rhs
;
1662 store_info
->redundant_reason
= redundant_reason
;
1664 /* If this is a clobber, we return 0. We will only be able to
1665 delete this insn if there is only one store USED store, but we
1666 can use the clobber to delete other stores earlier. */
1667 return store_info
->is_set
? 1 : 0;
1672 dump_insn_info (const char * start
, insn_info_t insn_info
)
1674 fprintf (dump_file
, "%s insn=%d %s\n", start
,
1675 INSN_UID (insn_info
->insn
),
1676 insn_info
->store_rec
? "has store" : "naked");
1680 /* If the modes are different and the value's source and target do not
1681 line up, we need to extract the value from lower part of the rhs of
1682 the store, shift it, and then put it into a form that can be shoved
1683 into the read_insn. This function generates a right SHIFT of a
1684 value that is at least ACCESS_SIZE bytes wide of READ_MODE. The
1685 shift sequence is returned or NULL if we failed to find a
1689 find_shift_sequence (poly_int64 access_size
,
1690 store_info
*store_info
,
1691 machine_mode read_mode
,
1692 poly_int64 shift
, bool speed
, bool require_cst
)
1694 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1695 scalar_int_mode new_mode
;
1696 rtx read_reg
= NULL
;
1698 /* Some machines like the x86 have shift insns for each size of
1699 operand. Other machines like the ppc or the ia-64 may only have
1700 shift insns that shift values within 32 or 64 bit registers.
1701 This loop tries to find the smallest shift insn that will right
1702 justify the value we want to read but is available in one insn on
1705 opt_scalar_int_mode new_mode_iter
;
1706 FOR_EACH_MODE_FROM (new_mode_iter
,
1707 smallest_int_mode_for_size (access_size
* BITS_PER_UNIT
))
1709 rtx target
, new_reg
, new_lhs
;
1710 rtx_insn
*shift_seq
, *insn
;
1713 new_mode
= new_mode_iter
.require ();
1714 if (GET_MODE_BITSIZE (new_mode
) > BITS_PER_WORD
)
1717 /* If a constant was stored into memory, try to simplify it here,
1718 otherwise the cost of the shift might preclude this optimization
1719 e.g. at -Os, even when no actual shift will be needed. */
1720 if (store_info
->const_rhs
)
1722 poly_uint64 byte
= subreg_lowpart_offset (new_mode
, store_mode
);
1723 rtx ret
= simplify_subreg (new_mode
, store_info
->const_rhs
,
1725 if (ret
&& CONSTANT_P (ret
))
1727 rtx shift_rtx
= gen_int_shift_amount (new_mode
, shift
);
1728 ret
= simplify_const_binary_operation (LSHIFTRT
, new_mode
,
1730 if (ret
&& CONSTANT_P (ret
))
1732 byte
= subreg_lowpart_offset (read_mode
, new_mode
);
1733 ret
= simplify_subreg (read_mode
, ret
, new_mode
, byte
);
1734 if (ret
&& CONSTANT_P (ret
)
1735 && (set_src_cost (ret
, read_mode
, speed
)
1736 <= COSTS_N_INSNS (1)))
1745 /* Try a wider mode if truncating the store mode to NEW_MODE
1746 requires a real instruction. */
1747 if (maybe_lt (GET_MODE_SIZE (new_mode
), GET_MODE_SIZE (store_mode
))
1748 && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode
, store_mode
))
1751 /* Also try a wider mode if the necessary punning is either not
1752 desirable or not possible. */
1753 if (!CONSTANT_P (store_info
->rhs
)
1754 && !targetm
.modes_tieable_p (new_mode
, store_mode
))
1757 new_reg
= gen_reg_rtx (new_mode
);
1761 /* In theory we could also check for an ashr. Ian Taylor knows
1762 of one dsp where the cost of these two was not the same. But
1763 this really is a rare case anyway. */
1764 target
= expand_binop (new_mode
, lshr_optab
, new_reg
,
1765 gen_int_shift_amount (new_mode
, shift
),
1766 new_reg
, 1, OPTAB_DIRECT
);
1768 shift_seq
= get_insns ();
1771 if (target
!= new_reg
|| shift_seq
== NULL
)
1775 for (insn
= shift_seq
; insn
!= NULL_RTX
; insn
= NEXT_INSN (insn
))
1777 cost
+= insn_cost (insn
, speed
);
1779 /* The computation up to here is essentially independent
1780 of the arguments and could be precomputed. It may
1781 not be worth doing so. We could precompute if
1782 worthwhile or at least cache the results. The result
1783 technically depends on both SHIFT and ACCESS_SIZE,
1784 but in practice the answer will depend only on ACCESS_SIZE. */
1786 if (cost
> COSTS_N_INSNS (1))
1789 new_lhs
= extract_low_bits (new_mode
, store_mode
,
1790 copy_rtx (store_info
->rhs
));
1791 if (new_lhs
== NULL_RTX
)
1794 /* We found an acceptable shift. Generate a move to
1795 take the value from the store and put it into the
1796 shift pseudo, then shift it, then generate another
1797 move to put in into the target of the read. */
1798 emit_move_insn (new_reg
, new_lhs
);
1799 emit_insn (shift_seq
);
1800 read_reg
= extract_low_bits (read_mode
, new_mode
, new_reg
);
1808 /* Call back for note_stores to find the hard regs set or clobbered by
1809 insn. Data is a bitmap of the hardregs set so far. */
1812 look_for_hardregs (rtx x
, const_rtx pat ATTRIBUTE_UNUSED
, void *data
)
1814 bitmap regs_set
= (bitmap
) data
;
1817 && HARD_REGISTER_P (x
))
1818 bitmap_set_range (regs_set
, REGNO (x
), REG_NREGS (x
));
1821 /* Helper function for replace_read and record_store.
1822 Attempt to return a value of mode READ_MODE stored in STORE_INFO,
1823 consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL
1824 if not successful. If REQUIRE_CST is true, return always constant. */
1827 get_stored_val (store_info
*store_info
, machine_mode read_mode
,
1828 poly_int64 read_offset
, poly_int64 read_width
,
1829 basic_block bb
, bool require_cst
)
1831 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1835 /* To get here the read is within the boundaries of the write so
1836 shift will never be negative. Start out with the shift being in
1838 if (store_mode
== BLKmode
)
1840 else if (BYTES_BIG_ENDIAN
)
1841 gap
= ((store_info
->offset
+ store_info
->width
)
1842 - (read_offset
+ read_width
));
1844 gap
= read_offset
- store_info
->offset
;
1846 if (gap
.is_constant () && maybe_ne (gap
, 0))
1848 poly_int64 shift
= gap
* BITS_PER_UNIT
;
1849 poly_int64 access_size
= GET_MODE_SIZE (read_mode
) + gap
;
1850 read_reg
= find_shift_sequence (access_size
, store_info
, read_mode
,
1851 shift
, optimize_bb_for_speed_p (bb
),
1854 else if (store_mode
== BLKmode
)
1856 /* The store is a memset (addr, const_val, const_size). */
1857 gcc_assert (CONST_INT_P (store_info
->rhs
));
1858 scalar_int_mode int_store_mode
;
1859 if (!int_mode_for_mode (read_mode
).exists (&int_store_mode
))
1860 read_reg
= NULL_RTX
;
1861 else if (store_info
->rhs
== const0_rtx
)
1862 read_reg
= extract_low_bits (read_mode
, int_store_mode
, const0_rtx
);
1863 else if (GET_MODE_BITSIZE (int_store_mode
) > HOST_BITS_PER_WIDE_INT
1864 || BITS_PER_UNIT
>= HOST_BITS_PER_WIDE_INT
)
1865 read_reg
= NULL_RTX
;
1868 unsigned HOST_WIDE_INT c
1869 = INTVAL (store_info
->rhs
)
1870 & ((HOST_WIDE_INT_1
<< BITS_PER_UNIT
) - 1);
1871 int shift
= BITS_PER_UNIT
;
1872 while (shift
< HOST_BITS_PER_WIDE_INT
)
1877 read_reg
= gen_int_mode (c
, int_store_mode
);
1878 read_reg
= extract_low_bits (read_mode
, int_store_mode
, read_reg
);
1881 else if (store_info
->const_rhs
1883 || GET_MODE_CLASS (read_mode
) != GET_MODE_CLASS (store_mode
)))
1884 read_reg
= extract_low_bits (read_mode
, store_mode
,
1885 copy_rtx (store_info
->const_rhs
));
1887 read_reg
= extract_low_bits (read_mode
, store_mode
,
1888 copy_rtx (store_info
->rhs
));
1889 if (require_cst
&& read_reg
&& !CONSTANT_P (read_reg
))
1890 read_reg
= NULL_RTX
;
1894 /* Take a sequence of:
1917 Depending on the alignment and the mode of the store and
1921 The STORE_INFO and STORE_INSN are for the store and READ_INFO
1922 and READ_INSN are for the read. Return true if the replacement
1926 replace_read (store_info
*store_info
, insn_info_t store_insn
,
1927 read_info_t read_info
, insn_info_t read_insn
, rtx
*loc
,
1930 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1931 machine_mode read_mode
= GET_MODE (read_info
->mem
);
1932 rtx_insn
*insns
, *this_insn
;
1939 /* Create a sequence of instructions to set up the read register.
1940 This sequence goes immediately before the store and its result
1941 is read by the load.
1943 We need to keep this in perspective. We are replacing a read
1944 with a sequence of insns, but the read will almost certainly be
1945 in cache, so it is not going to be an expensive one. Thus, we
1946 are not willing to do a multi insn shift or worse a subroutine
1947 call to get rid of the read. */
1948 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1949 fprintf (dump_file
, "trying to replace %smode load in insn %d"
1950 " from %smode store in insn %d\n",
1951 GET_MODE_NAME (read_mode
), INSN_UID (read_insn
->insn
),
1952 GET_MODE_NAME (store_mode
), INSN_UID (store_insn
->insn
));
1954 bb
= BLOCK_FOR_INSN (read_insn
->insn
);
1955 read_reg
= get_stored_val (store_info
,
1956 read_mode
, read_info
->offset
, read_info
->width
,
1958 if (read_reg
== NULL_RTX
)
1961 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1962 fprintf (dump_file
, " -- could not extract bits of stored value\n");
1965 /* Force the value into a new register so that it won't be clobbered
1966 between the store and the load. */
1967 read_reg
= copy_to_mode_reg (read_mode
, read_reg
);
1968 insns
= get_insns ();
1971 if (insns
!= NULL_RTX
)
1973 /* Now we have to scan the set of new instructions to see if the
1974 sequence contains and sets of hardregs that happened to be
1975 live at this point. For instance, this can happen if one of
1976 the insns sets the CC and the CC happened to be live at that
1977 point. This does occasionally happen, see PR 37922. */
1978 bitmap regs_set
= BITMAP_ALLOC (®_obstack
);
1980 for (this_insn
= insns
; this_insn
!= NULL_RTX
; this_insn
= NEXT_INSN (this_insn
))
1981 note_stores (this_insn
, look_for_hardregs
, regs_set
);
1983 bitmap_and_into (regs_set
, regs_live
);
1984 if (!bitmap_empty_p (regs_set
))
1986 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1989 "abandoning replacement because sequence clobbers live hardregs:");
1990 df_print_regset (dump_file
, regs_set
);
1993 BITMAP_FREE (regs_set
);
1996 BITMAP_FREE (regs_set
);
1999 if (validate_change (read_insn
->insn
, loc
, read_reg
, 0))
2001 deferred_change
*change
= deferred_change_pool
.allocate ();
2003 /* Insert this right before the store insn where it will be safe
2004 from later insns that might change it before the read. */
2005 emit_insn_before (insns
, store_insn
->insn
);
2007 /* And now for the kludge part: cselib croaks if you just
2008 return at this point. There are two reasons for this:
2010 1) Cselib has an idea of how many pseudos there are and
2011 that does not include the new ones we just added.
2013 2) Cselib does not know about the move insn we added
2014 above the store_info, and there is no way to tell it
2015 about it, because it has "moved on".
2017 Problem (1) is fixable with a certain amount of engineering.
2018 Problem (2) is requires starting the bb from scratch. This
2021 So we are just going to have to lie. The move/extraction
2022 insns are not really an issue, cselib did not see them. But
2023 the use of the new pseudo read_insn is a real problem because
2024 cselib has not scanned this insn. The way that we solve this
2025 problem is that we are just going to put the mem back for now
2026 and when we are finished with the block, we undo this. We
2027 keep a table of mems to get rid of. At the end of the basic
2028 block we can put them back. */
2030 *loc
= read_info
->mem
;
2031 change
->next
= deferred_change_list
;
2032 deferred_change_list
= change
;
2034 change
->reg
= read_reg
;
2036 /* Get rid of the read_info, from the point of view of the
2037 rest of dse, play like this read never happened. */
2038 read_insn
->read_rec
= read_info
->next
;
2039 read_info_type_pool
.remove (read_info
);
2040 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2042 fprintf (dump_file
, " -- replaced the loaded MEM with ");
2043 print_simple_rtl (dump_file
, read_reg
);
2044 fprintf (dump_file
, "\n");
2050 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2052 fprintf (dump_file
, " -- replacing the loaded MEM with ");
2053 print_simple_rtl (dump_file
, read_reg
);
2054 fprintf (dump_file
, " led to an invalid instruction\n");
2060 /* Check the address of MEM *LOC and kill any appropriate stores that may
2064 check_mem_read_rtx (rtx
*loc
, bb_info_t bb_info
)
2066 rtx mem
= *loc
, mem_addr
;
2067 insn_info_t insn_info
;
2068 poly_int64 offset
= 0;
2069 poly_int64 width
= 0;
2070 cselib_val
*base
= NULL
;
2072 read_info_t read_info
;
2074 insn_info
= bb_info
->last_insn
;
2076 if ((MEM_ALIAS_SET (mem
) == ALIAS_SET_MEMORY_BARRIER
)
2077 || MEM_VOLATILE_P (mem
))
2079 if (crtl
->stack_protect_guard
2080 && (MEM_EXPR (mem
) == crtl
->stack_protect_guard
2081 || (crtl
->stack_protect_guard_decl
2082 && MEM_EXPR (mem
) == crtl
->stack_protect_guard_decl
))
2083 && MEM_VOLATILE_P (mem
))
2085 /* This is either the stack protector canary on the stack,
2086 which ought to be written by a MEM_VOLATILE_P store and
2087 thus shouldn't be deleted and is read at the very end of
2088 function, but shouldn't conflict with any other store.
2089 Or it is __stack_chk_guard variable or TLS or whatever else
2090 MEM holding the canary value, which really shouldn't be
2091 ever modified in -fstack-protector* protected functions,
2092 otherwise the prologue store wouldn't match the epilogue
2094 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2095 fprintf (dump_file
, " stack protector canary read ignored.\n");
2096 insn_info
->cannot_delete
= true;
2100 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2101 fprintf (dump_file
, " adding wild read, volatile or barrier.\n");
2102 add_wild_read (bb_info
);
2103 insn_info
->cannot_delete
= true;
2107 /* If it is reading readonly mem, then there can be no conflict with
2109 if (MEM_READONLY_P (mem
))
2112 if (!canon_address (mem
, &group_id
, &offset
, &base
))
2114 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2115 fprintf (dump_file
, " adding wild read, canon_address failure.\n");
2116 add_wild_read (bb_info
);
2120 if (GET_MODE (mem
) == BLKmode
)
2123 width
= GET_MODE_SIZE (GET_MODE (mem
));
2125 if (!endpoint_representable_p (offset
, known_eq (width
, -1) ? 1 : width
))
2127 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2128 fprintf (dump_file
, " adding wild read, due to overflow.\n");
2129 add_wild_read (bb_info
);
2133 read_info
= read_info_type_pool
.allocate ();
2134 read_info
->group_id
= group_id
;
2135 read_info
->mem
= mem
;
2136 read_info
->offset
= offset
;
2137 read_info
->width
= width
;
2138 read_info
->next
= insn_info
->read_rec
;
2139 insn_info
->read_rec
= read_info
;
2141 mem_addr
= base
->val_rtx
;
2144 group_info
*group
= rtx_group_vec
[group_id
];
2145 mem_addr
= group
->canon_base_addr
;
2147 if (maybe_ne (offset
, 0))
2148 mem_addr
= plus_constant (get_address_mode (mem
), mem_addr
, offset
);
2152 /* This is the restricted case where the base is a constant or
2153 the frame pointer and offset is a constant. */
2154 insn_info_t i_ptr
= active_local_stores
;
2155 insn_info_t last
= NULL
;
2157 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2159 if (!known_size_p (width
))
2160 fprintf (dump_file
, " processing const load gid=%d[BLK]\n",
2164 fprintf (dump_file
, " processing const load gid=%d", group_id
);
2165 print_range (dump_file
, offset
, width
);
2166 fprintf (dump_file
, "\n");
2172 bool remove
= false;
2173 store_info
*store_info
= i_ptr
->store_rec
;
2175 /* Skip the clobbers. */
2176 while (!store_info
->is_set
)
2177 store_info
= store_info
->next
;
2179 /* There are three cases here. */
2180 if (store_info
->group_id
< 0)
2181 /* We have a cselib store followed by a read from a
2184 = canon_true_dependence (store_info
->mem
,
2185 GET_MODE (store_info
->mem
),
2186 store_info
->mem_addr
,
2189 else if (group_id
== store_info
->group_id
)
2191 /* This is a block mode load. We may get lucky and
2192 canon_true_dependence may save the day. */
2193 if (!known_size_p (width
))
2195 = canon_true_dependence (store_info
->mem
,
2196 GET_MODE (store_info
->mem
),
2197 store_info
->mem_addr
,
2200 /* If this read is just reading back something that we just
2201 stored, rewrite the read. */
2205 && known_subrange_p (offset
, width
, store_info
->offset
,
2207 && all_positions_needed_p (store_info
,
2208 offset
- store_info
->offset
,
2210 && replace_read (store_info
, i_ptr
, read_info
,
2211 insn_info
, loc
, bb_info
->regs_live
))
2214 /* The bases are the same, just see if the offsets
2216 if (ranges_maybe_overlap_p (offset
, width
,
2224 The else case that is missing here is that the
2225 bases are constant but different. There is nothing
2226 to do here because there is no overlap. */
2230 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2231 dump_insn_info ("removing from active", i_ptr
);
2233 active_local_stores_len
--;
2235 last
->next_local_store
= i_ptr
->next_local_store
;
2237 active_local_stores
= i_ptr
->next_local_store
;
2241 i_ptr
= i_ptr
->next_local_store
;
2246 insn_info_t i_ptr
= active_local_stores
;
2247 insn_info_t last
= NULL
;
2248 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2250 fprintf (dump_file
, " processing cselib load mem:");
2251 print_inline_rtx (dump_file
, mem
, 0);
2252 fprintf (dump_file
, "\n");
2257 bool remove
= false;
2258 store_info
*store_info
= i_ptr
->store_rec
;
2260 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2261 fprintf (dump_file
, " processing cselib load against insn %d\n",
2262 INSN_UID (i_ptr
->insn
));
2264 /* Skip the clobbers. */
2265 while (!store_info
->is_set
)
2266 store_info
= store_info
->next
;
2268 /* If this read is just reading back something that we just
2269 stored, rewrite the read. */
2271 && store_info
->group_id
== -1
2272 && store_info
->cse_base
== base
2273 && known_subrange_p (offset
, width
, store_info
->offset
,
2275 && all_positions_needed_p (store_info
,
2276 offset
- store_info
->offset
, width
)
2277 && replace_read (store_info
, i_ptr
, read_info
, insn_info
, loc
,
2278 bb_info
->regs_live
))
2281 remove
= canon_true_dependence (store_info
->mem
,
2282 GET_MODE (store_info
->mem
),
2283 store_info
->mem_addr
,
2288 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2289 dump_insn_info ("removing from active", i_ptr
);
2291 active_local_stores_len
--;
2293 last
->next_local_store
= i_ptr
->next_local_store
;
2295 active_local_stores
= i_ptr
->next_local_store
;
2299 i_ptr
= i_ptr
->next_local_store
;
2304 /* A note_uses callback in which DATA points the INSN_INFO for
2305 as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns
2306 true for any part of *LOC. */
2309 check_mem_read_use (rtx
*loc
, void *data
)
2311 subrtx_ptr_iterator::array_type array
;
2312 FOR_EACH_SUBRTX_PTR (iter
, array
, loc
, NONCONST
)
2316 check_mem_read_rtx (loc
, (bb_info_t
) data
);
2321 /* Get arguments passed to CALL_INSN. Return TRUE if successful.
2322 So far it only handles arguments passed in registers. */
2325 get_call_args (rtx call_insn
, tree fn
, rtx
*args
, int nargs
)
2327 CUMULATIVE_ARGS args_so_far_v
;
2328 cumulative_args_t args_so_far
;
2332 INIT_CUMULATIVE_ARGS (args_so_far_v
, TREE_TYPE (fn
), NULL_RTX
, 0, 3);
2333 args_so_far
= pack_cumulative_args (&args_so_far_v
);
2335 arg
= TYPE_ARG_TYPES (TREE_TYPE (fn
));
2337 arg
!= void_list_node
&& idx
< nargs
;
2338 arg
= TREE_CHAIN (arg
), idx
++)
2340 scalar_int_mode mode
;
2343 if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg
)), &mode
))
2346 function_arg_info
arg (mode
, /*named=*/true);
2347 reg
= targetm
.calls
.function_arg (args_so_far
, arg
);
2348 if (!reg
|| !REG_P (reg
) || GET_MODE (reg
) != mode
)
2351 for (link
= CALL_INSN_FUNCTION_USAGE (call_insn
);
2353 link
= XEXP (link
, 1))
2354 if (GET_CODE (XEXP (link
, 0)) == USE
)
2356 scalar_int_mode arg_mode
;
2357 args
[idx
] = XEXP (XEXP (link
, 0), 0);
2358 if (REG_P (args
[idx
])
2359 && REGNO (args
[idx
]) == REGNO (reg
)
2360 && (GET_MODE (args
[idx
]) == mode
2361 || (is_int_mode (GET_MODE (args
[idx
]), &arg_mode
)
2362 && (GET_MODE_SIZE (arg_mode
) <= UNITS_PER_WORD
)
2363 && (GET_MODE_SIZE (arg_mode
) > GET_MODE_SIZE (mode
)))))
2369 tmp
= cselib_expand_value_rtx (args
[idx
], scratch
, 5);
2370 if (GET_MODE (args
[idx
]) != mode
)
2372 if (!tmp
|| !CONST_INT_P (tmp
))
2374 tmp
= gen_int_mode (INTVAL (tmp
), mode
);
2379 targetm
.calls
.function_arg_advance (args_so_far
, arg
);
2381 if (arg
!= void_list_node
|| idx
!= nargs
)
2386 /* Return a bitmap of the fixed registers contained in IN. */
2389 copy_fixed_regs (const_bitmap in
)
2393 ret
= ALLOC_REG_SET (NULL
);
2394 bitmap_and (ret
, in
, bitmap_view
<HARD_REG_SET
> (fixed_reg_set
));
2398 /* Apply record_store to all candidate stores in INSN. Mark INSN
2399 if some part of it is not a candidate store and assigns to a
2400 non-register target. */
2403 scan_insn (bb_info_t bb_info
, rtx_insn
*insn
, int max_active_local_stores
)
2406 insn_info_type
*insn_info
= insn_info_type_pool
.allocate ();
2408 memset (insn_info
, 0, sizeof (struct insn_info_type
));
2410 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2411 fprintf (dump_file
, "\n**scanning insn=%d\n",
2414 insn_info
->prev_insn
= bb_info
->last_insn
;
2415 insn_info
->insn
= insn
;
2416 bb_info
->last_insn
= insn_info
;
2418 if (DEBUG_INSN_P (insn
))
2420 insn_info
->cannot_delete
= true;
2424 /* Look at all of the uses in the insn. */
2425 note_uses (&PATTERN (insn
), check_mem_read_use
, bb_info
);
2431 tree memset_call
= NULL_TREE
;
2433 insn_info
->cannot_delete
= true;
2435 /* Const functions cannot do anything bad i.e. read memory,
2436 however, they can read their parameters which may have
2437 been pushed onto the stack.
2438 memset and bzero don't read memory either. */
2439 const_call
= RTL_CONST_CALL_P (insn
);
2441 && (call
= get_call_rtx_from (insn
))
2442 && (sym
= XEXP (XEXP (call
, 0), 0))
2443 && GET_CODE (sym
) == SYMBOL_REF
2444 && SYMBOL_REF_DECL (sym
)
2445 && TREE_CODE (SYMBOL_REF_DECL (sym
)) == FUNCTION_DECL
2446 && fndecl_built_in_p (SYMBOL_REF_DECL (sym
), BUILT_IN_MEMSET
))
2447 memset_call
= SYMBOL_REF_DECL (sym
);
2449 if (const_call
|| memset_call
)
2451 insn_info_t i_ptr
= active_local_stores
;
2452 insn_info_t last
= NULL
;
2454 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2455 fprintf (dump_file
, "%s call %d\n",
2456 const_call
? "const" : "memset", INSN_UID (insn
));
2458 /* See the head comment of the frame_read field. */
2459 if (reload_completed
2460 /* Tail calls are storing their arguments using
2461 arg pointer. If it is a frame pointer on the target,
2462 even before reload we need to kill frame pointer based
2464 || (SIBLING_CALL_P (insn
)
2465 && HARD_FRAME_POINTER_IS_ARG_POINTER
))
2466 insn_info
->frame_read
= true;
2468 /* Loop over the active stores and remove those which are
2469 killed by the const function call. */
2472 bool remove_store
= false;
2474 /* The stack pointer based stores are always killed. */
2475 if (i_ptr
->stack_pointer_based
)
2476 remove_store
= true;
2478 /* If the frame is read, the frame related stores are killed. */
2479 else if (insn_info
->frame_read
)
2481 store_info
*store_info
= i_ptr
->store_rec
;
2483 /* Skip the clobbers. */
2484 while (!store_info
->is_set
)
2485 store_info
= store_info
->next
;
2487 if (store_info
->group_id
>= 0
2488 && rtx_group_vec
[store_info
->group_id
]->frame_related
)
2489 remove_store
= true;
2494 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2495 dump_insn_info ("removing from active", i_ptr
);
2497 active_local_stores_len
--;
2499 last
->next_local_store
= i_ptr
->next_local_store
;
2501 active_local_stores
= i_ptr
->next_local_store
;
2506 i_ptr
= i_ptr
->next_local_store
;
2512 if (get_call_args (insn
, memset_call
, args
, 3)
2513 && CONST_INT_P (args
[1])
2514 && CONST_INT_P (args
[2])
2515 && INTVAL (args
[2]) > 0)
2517 rtx mem
= gen_rtx_MEM (BLKmode
, args
[0]);
2518 set_mem_size (mem
, INTVAL (args
[2]));
2519 body
= gen_rtx_SET (mem
, args
[1]);
2520 mems_found
+= record_store (body
, bb_info
);
2521 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2522 fprintf (dump_file
, "handling memset as BLKmode store\n");
2523 if (mems_found
== 1)
2525 if (active_local_stores_len
++ >= max_active_local_stores
)
2527 active_local_stores_len
= 1;
2528 active_local_stores
= NULL
;
2530 insn_info
->fixed_regs_live
2531 = copy_fixed_regs (bb_info
->regs_live
);
2532 insn_info
->next_local_store
= active_local_stores
;
2533 active_local_stores
= insn_info
;
2537 clear_rhs_from_active_local_stores ();
2540 else if (SIBLING_CALL_P (insn
)
2541 && (reload_completed
|| HARD_FRAME_POINTER_IS_ARG_POINTER
))
2542 /* Arguments for a sibling call that are pushed to memory are passed
2543 using the incoming argument pointer of the current function. After
2544 reload that might be (and likely is) frame pointer based. And, if
2545 it is a frame pointer on the target, even before reload we need to
2546 kill frame pointer based stores. */
2547 add_wild_read (bb_info
);
2549 /* Every other call, including pure functions, may read any memory
2550 that is not relative to the frame. */
2551 add_non_frame_wild_read (bb_info
);
2556 /* Assuming that there are sets in these insns, we cannot delete
2558 if ((GET_CODE (PATTERN (insn
)) == CLOBBER
)
2559 || volatile_refs_p (PATTERN (insn
))
2560 || (!cfun
->can_delete_dead_exceptions
&& !insn_nothrow_p (insn
))
2561 || (RTX_FRAME_RELATED_P (insn
))
2562 || find_reg_note (insn
, REG_FRAME_RELATED_EXPR
, NULL_RTX
))
2563 insn_info
->cannot_delete
= true;
2565 body
= PATTERN (insn
);
2566 if (GET_CODE (body
) == PARALLEL
)
2569 for (i
= 0; i
< XVECLEN (body
, 0); i
++)
2570 mems_found
+= record_store (XVECEXP (body
, 0, i
), bb_info
);
2573 mems_found
+= record_store (body
, bb_info
);
2575 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2576 fprintf (dump_file
, "mems_found = %d, cannot_delete = %s\n",
2577 mems_found
, insn_info
->cannot_delete
? "true" : "false");
2579 /* If we found some sets of mems, add it into the active_local_stores so
2580 that it can be locally deleted if found dead or used for
2581 replace_read and redundant constant store elimination. Otherwise mark
2582 it as cannot delete. This simplifies the processing later. */
2583 if (mems_found
== 1)
2585 if (active_local_stores_len
++ >= max_active_local_stores
)
2587 active_local_stores_len
= 1;
2588 active_local_stores
= NULL
;
2590 insn_info
->fixed_regs_live
= copy_fixed_regs (bb_info
->regs_live
);
2591 insn_info
->next_local_store
= active_local_stores
;
2592 active_local_stores
= insn_info
;
2595 insn_info
->cannot_delete
= true;
2599 /* Remove BASE from the set of active_local_stores. This is a
2600 callback from cselib that is used to get rid of the stores in
2601 active_local_stores. */
2604 remove_useless_values (cselib_val
*base
)
2606 insn_info_t insn_info
= active_local_stores
;
2607 insn_info_t last
= NULL
;
2611 store_info
*store_info
= insn_info
->store_rec
;
2614 /* If ANY of the store_infos match the cselib group that is
2615 being deleted, then the insn cannot be deleted. */
2618 if ((store_info
->group_id
== -1)
2619 && (store_info
->cse_base
== base
))
2624 store_info
= store_info
->next
;
2629 active_local_stores_len
--;
2631 last
->next_local_store
= insn_info
->next_local_store
;
2633 active_local_stores
= insn_info
->next_local_store
;
2634 free_store_info (insn_info
);
2639 insn_info
= insn_info
->next_local_store
;
2644 /* Do all of step 1. */
2650 bitmap regs_live
= BITMAP_ALLOC (®_obstack
);
2653 all_blocks
= BITMAP_ALLOC (NULL
);
2654 bitmap_set_bit (all_blocks
, ENTRY_BLOCK
);
2655 bitmap_set_bit (all_blocks
, EXIT_BLOCK
);
2657 /* For -O1 reduce the maximum number of active local stores for RTL DSE
2658 since this can consume huge amounts of memory (PR89115). */
2659 int max_active_local_stores
= param_max_dse_active_local_stores
;
2661 max_active_local_stores
/= 10;
2663 FOR_ALL_BB_FN (bb
, cfun
)
2666 bb_info_t bb_info
= dse_bb_info_type_pool
.allocate ();
2668 memset (bb_info
, 0, sizeof (dse_bb_info_type
));
2669 bitmap_set_bit (all_blocks
, bb
->index
);
2670 bb_info
->regs_live
= regs_live
;
2672 bitmap_copy (regs_live
, DF_LR_IN (bb
));
2673 df_simulate_initialize_forwards (bb
, regs_live
);
2675 bb_table
[bb
->index
] = bb_info
;
2676 cselib_discard_hook
= remove_useless_values
;
2678 if (bb
->index
>= NUM_FIXED_BLOCKS
)
2682 active_local_stores
= NULL
;
2683 active_local_stores_len
= 0;
2684 cselib_clear_table ();
2686 /* Scan the insns. */
2687 FOR_BB_INSNS (bb
, insn
)
2690 scan_insn (bb_info
, insn
, max_active_local_stores
);
2691 cselib_process_insn (insn
);
2693 df_simulate_one_insn_forwards (bb
, insn
, regs_live
);
2696 /* This is something of a hack, because the global algorithm
2697 is supposed to take care of the case where stores go dead
2698 at the end of the function. However, the global
2699 algorithm must take a more conservative view of block
2700 mode reads than the local alg does. So to get the case
2701 where you have a store to the frame followed by a non
2702 overlapping block more read, we look at the active local
2703 stores at the end of the function and delete all of the
2704 frame and spill based ones. */
2705 if (stores_off_frame_dead_at_return
2706 && (EDGE_COUNT (bb
->succs
) == 0
2707 || (single_succ_p (bb
)
2708 && single_succ (bb
) == EXIT_BLOCK_PTR_FOR_FN (cfun
)
2709 && ! crtl
->calls_eh_return
)))
2711 insn_info_t i_ptr
= active_local_stores
;
2714 store_info
*store_info
= i_ptr
->store_rec
;
2716 /* Skip the clobbers. */
2717 while (!store_info
->is_set
)
2718 store_info
= store_info
->next
;
2719 if (store_info
->group_id
>= 0)
2721 group_info
*group
= rtx_group_vec
[store_info
->group_id
];
2722 if (group
->frame_related
&& !i_ptr
->cannot_delete
)
2723 delete_dead_store_insn (i_ptr
);
2726 i_ptr
= i_ptr
->next_local_store
;
2730 /* Get rid of the loads that were discovered in
2731 replace_read. Cselib is finished with this block. */
2732 while (deferred_change_list
)
2734 deferred_change
*next
= deferred_change_list
->next
;
2736 /* There is no reason to validate this change. That was
2738 *deferred_change_list
->loc
= deferred_change_list
->reg
;
2739 deferred_change_pool
.remove (deferred_change_list
);
2740 deferred_change_list
= next
;
2743 /* Get rid of all of the cselib based store_infos in this
2744 block and mark the containing insns as not being
2746 ptr
= bb_info
->last_insn
;
2749 if (ptr
->contains_cselib_groups
)
2751 store_info
*s_info
= ptr
->store_rec
;
2752 while (s_info
&& !s_info
->is_set
)
2753 s_info
= s_info
->next
;
2755 && s_info
->redundant_reason
2756 && s_info
->redundant_reason
->insn
2757 && !ptr
->cannot_delete
)
2759 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2760 fprintf (dump_file
, "Locally deleting insn %d "
2761 "because insn %d stores the "
2762 "same value and couldn't be "
2764 INSN_UID (ptr
->insn
),
2765 INSN_UID (s_info
->redundant_reason
->insn
));
2766 delete_dead_store_insn (ptr
);
2768 free_store_info (ptr
);
2774 /* Free at least positions_needed bitmaps. */
2775 for (s_info
= ptr
->store_rec
; s_info
; s_info
= s_info
->next
)
2776 if (s_info
->is_large
)
2778 BITMAP_FREE (s_info
->positions_needed
.large
.bmap
);
2779 s_info
->is_large
= false;
2782 ptr
= ptr
->prev_insn
;
2785 cse_store_info_pool
.release ();
2787 bb_info
->regs_live
= NULL
;
2790 BITMAP_FREE (regs_live
);
2792 rtx_group_table
->empty ();
2796 /*----------------------------------------------------------------------------
2799 Assign each byte position in the stores that we are going to
2800 analyze globally to a position in the bitmaps. Returns true if
2801 there are any bit positions assigned.
2802 ----------------------------------------------------------------------------*/
2805 dse_step2_init (void)
2810 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2812 /* For all non stack related bases, we only consider a store to
2813 be deletable if there are two or more stores for that
2814 position. This is because it takes one store to make the
2815 other store redundant. However, for the stores that are
2816 stack related, we consider them if there is only one store
2817 for the position. We do this because the stack related
2818 stores can be deleted if their is no read between them and
2819 the end of the function.
2821 To make this work in the current framework, we take the stack
2822 related bases add all of the bits from store1 into store2.
2823 This has the effect of making the eligible even if there is
2826 if (stores_off_frame_dead_at_return
&& group
->frame_related
)
2828 bitmap_ior_into (group
->store2_n
, group
->store1_n
);
2829 bitmap_ior_into (group
->store2_p
, group
->store1_p
);
2830 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2831 fprintf (dump_file
, "group %d is frame related ", i
);
2834 group
->offset_map_size_n
++;
2835 group
->offset_map_n
= XOBNEWVEC (&dse_obstack
, int,
2836 group
->offset_map_size_n
);
2837 group
->offset_map_size_p
++;
2838 group
->offset_map_p
= XOBNEWVEC (&dse_obstack
, int,
2839 group
->offset_map_size_p
);
2840 group
->process_globally
= false;
2841 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2843 fprintf (dump_file
, "group %d(%d+%d): ", i
,
2844 (int)bitmap_count_bits (group
->store2_n
),
2845 (int)bitmap_count_bits (group
->store2_p
));
2846 bitmap_print (dump_file
, group
->store2_n
, "n ", " ");
2847 bitmap_print (dump_file
, group
->store2_p
, "p ", "\n");
2853 /* Init the offset tables. */
2860 /* Position 0 is unused because 0 is used in the maps to mean
2862 current_position
= 1;
2863 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2868 memset (group
->offset_map_n
, 0, sizeof (int) * group
->offset_map_size_n
);
2869 memset (group
->offset_map_p
, 0, sizeof (int) * group
->offset_map_size_p
);
2870 bitmap_clear (group
->group_kill
);
2872 EXECUTE_IF_SET_IN_BITMAP (group
->store2_n
, 0, j
, bi
)
2874 bitmap_set_bit (group
->group_kill
, current_position
);
2875 if (bitmap_bit_p (group
->escaped_n
, j
))
2876 bitmap_set_bit (kill_on_calls
, current_position
);
2877 group
->offset_map_n
[j
] = current_position
++;
2878 group
->process_globally
= true;
2880 EXECUTE_IF_SET_IN_BITMAP (group
->store2_p
, 0, j
, bi
)
2882 bitmap_set_bit (group
->group_kill
, current_position
);
2883 if (bitmap_bit_p (group
->escaped_p
, j
))
2884 bitmap_set_bit (kill_on_calls
, current_position
);
2885 group
->offset_map_p
[j
] = current_position
++;
2886 group
->process_globally
= true;
2889 return current_position
!= 1;
2894 /*----------------------------------------------------------------------------
2897 Build the bit vectors for the transfer functions.
2898 ----------------------------------------------------------------------------*/
2901 /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not
2905 get_bitmap_index (group_info
*group_info
, HOST_WIDE_INT offset
)
2909 HOST_WIDE_INT offset_p
= -offset
;
2910 if (offset_p
>= group_info
->offset_map_size_n
)
2912 return group_info
->offset_map_n
[offset_p
];
2916 if (offset
>= group_info
->offset_map_size_p
)
2918 return group_info
->offset_map_p
[offset
];
2923 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL
2927 scan_stores (store_info
*store_info
, bitmap gen
, bitmap kill
)
2931 HOST_WIDE_INT i
, offset
, width
;
2932 group_info
*group_info
2933 = rtx_group_vec
[store_info
->group_id
];
2934 /* We can (conservatively) ignore stores whose bounds aren't known;
2935 they simply don't generate new global dse opportunities. */
2936 if (group_info
->process_globally
2937 && store_info
->offset
.is_constant (&offset
)
2938 && store_info
->width
.is_constant (&width
))
2940 HOST_WIDE_INT end
= offset
+ width
;
2941 for (i
= offset
; i
< end
; i
++)
2943 int index
= get_bitmap_index (group_info
, i
);
2946 bitmap_set_bit (gen
, index
);
2948 bitmap_clear_bit (kill
, index
);
2952 store_info
= store_info
->next
;
2957 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL
2961 scan_reads (insn_info_t insn_info
, bitmap gen
, bitmap kill
)
2963 read_info_t read_info
= insn_info
->read_rec
;
2967 /* If this insn reads the frame, kill all the frame related stores. */
2968 if (insn_info
->frame_read
)
2970 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2971 if (group
->process_globally
&& group
->frame_related
)
2974 bitmap_ior_into (kill
, group
->group_kill
);
2975 bitmap_and_compl_into (gen
, group
->group_kill
);
2978 if (insn_info
->non_frame_wild_read
)
2980 /* Kill all non-frame related stores. Kill all stores of variables that
2983 bitmap_ior_into (kill
, kill_on_calls
);
2984 bitmap_and_compl_into (gen
, kill_on_calls
);
2985 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2986 if (group
->process_globally
&& !group
->frame_related
)
2989 bitmap_ior_into (kill
, group
->group_kill
);
2990 bitmap_and_compl_into (gen
, group
->group_kill
);
2995 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2997 if (group
->process_globally
)
2999 if (i
== read_info
->group_id
)
3001 HOST_WIDE_INT offset
, width
;
3002 /* Reads with non-constant size kill all DSE opportunities
3004 if (!read_info
->offset
.is_constant (&offset
)
3005 || !read_info
->width
.is_constant (&width
)
3006 || !known_size_p (width
))
3008 /* Handle block mode reads. */
3010 bitmap_ior_into (kill
, group
->group_kill
);
3011 bitmap_and_compl_into (gen
, group
->group_kill
);
3015 /* The groups are the same, just process the
3018 HOST_WIDE_INT end
= offset
+ width
;
3019 for (j
= offset
; j
< end
; j
++)
3021 int index
= get_bitmap_index (group
, j
);
3025 bitmap_set_bit (kill
, index
);
3026 bitmap_clear_bit (gen
, index
);
3033 /* The groups are different, if the alias sets
3034 conflict, clear the entire group. We only need
3035 to apply this test if the read_info is a cselib
3036 read. Anything with a constant base cannot alias
3037 something else with a different constant
3039 if ((read_info
->group_id
< 0)
3040 && canon_true_dependence (group
->base_mem
,
3041 GET_MODE (group
->base_mem
),
3042 group
->canon_base_addr
,
3043 read_info
->mem
, NULL_RTX
))
3046 bitmap_ior_into (kill
, group
->group_kill
);
3047 bitmap_and_compl_into (gen
, group
->group_kill
);
3053 read_info
= read_info
->next
;
3058 /* Return the insn in BB_INFO before the first wild read or if there
3059 are no wild reads in the block, return the last insn. */
3062 find_insn_before_first_wild_read (bb_info_t bb_info
)
3064 insn_info_t insn_info
= bb_info
->last_insn
;
3065 insn_info_t last_wild_read
= NULL
;
3069 if (insn_info
->wild_read
)
3071 last_wild_read
= insn_info
->prev_insn
;
3072 /* Block starts with wild read. */
3073 if (!last_wild_read
)
3077 insn_info
= insn_info
->prev_insn
;
3081 return last_wild_read
;
3083 return bb_info
->last_insn
;
3087 /* Scan the insns in BB_INFO starting at PTR and going to the top of
3088 the block in order to build the gen and kill sets for the block.
3089 We start at ptr which may be the last insn in the block or may be
3090 the first insn with a wild read. In the latter case we are able to
3091 skip the rest of the block because it just does not matter:
3092 anything that happens is hidden by the wild read. */
3095 dse_step3_scan (basic_block bb
)
3097 bb_info_t bb_info
= bb_table
[bb
->index
];
3098 insn_info_t insn_info
;
3100 insn_info
= find_insn_before_first_wild_read (bb_info
);
3102 /* In the spill case or in the no_spill case if there is no wild
3103 read in the block, we will need a kill set. */
3104 if (insn_info
== bb_info
->last_insn
)
3107 bitmap_clear (bb_info
->kill
);
3109 bb_info
->kill
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3113 BITMAP_FREE (bb_info
->kill
);
3117 /* There may have been code deleted by the dce pass run before
3119 if (insn_info
->insn
&& INSN_P (insn_info
->insn
))
3121 scan_stores (insn_info
->store_rec
, bb_info
->gen
, bb_info
->kill
);
3122 scan_reads (insn_info
, bb_info
->gen
, bb_info
->kill
);
3125 insn_info
= insn_info
->prev_insn
;
3130 /* Set the gen set of the exit block, and also any block with no
3131 successors that does not have a wild read. */
3134 dse_step3_exit_block_scan (bb_info_t bb_info
)
3136 /* The gen set is all 0's for the exit block except for the
3137 frame_pointer_group. */
3139 if (stores_off_frame_dead_at_return
)
3144 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3146 if (group
->process_globally
&& group
->frame_related
)
3147 bitmap_ior_into (bb_info
->gen
, group
->group_kill
);
3153 /* Find all of the blocks that are not backwards reachable from the
3154 exit block or any block with no successors (BB). These are the
3155 infinite loops or infinite self loops. These blocks will still
3156 have their bits set in UNREACHABLE_BLOCKS. */
3159 mark_reachable_blocks (sbitmap unreachable_blocks
, basic_block bb
)
3164 if (bitmap_bit_p (unreachable_blocks
, bb
->index
))
3166 bitmap_clear_bit (unreachable_blocks
, bb
->index
);
3167 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3169 mark_reachable_blocks (unreachable_blocks
, e
->src
);
3174 /* Build the transfer functions for the function. */
3180 sbitmap_iterator sbi
;
3181 bitmap all_ones
= NULL
;
3184 auto_sbitmap
unreachable_blocks (last_basic_block_for_fn (cfun
));
3185 bitmap_ones (unreachable_blocks
);
3187 FOR_ALL_BB_FN (bb
, cfun
)
3189 bb_info_t bb_info
= bb_table
[bb
->index
];
3191 bitmap_clear (bb_info
->gen
);
3193 bb_info
->gen
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3195 if (bb
->index
== ENTRY_BLOCK
)
3197 else if (bb
->index
== EXIT_BLOCK
)
3198 dse_step3_exit_block_scan (bb_info
);
3200 dse_step3_scan (bb
);
3201 if (EDGE_COUNT (bb
->succs
) == 0)
3202 mark_reachable_blocks (unreachable_blocks
, bb
);
3204 /* If this is the second time dataflow is run, delete the old
3207 BITMAP_FREE (bb_info
->in
);
3209 BITMAP_FREE (bb_info
->out
);
3212 /* For any block in an infinite loop, we must initialize the out set
3213 to all ones. This could be expensive, but almost never occurs in
3214 practice. However, it is common in regression tests. */
3215 EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks
, 0, i
, sbi
)
3217 if (bitmap_bit_p (all_blocks
, i
))
3219 bb_info_t bb_info
= bb_table
[i
];
3225 all_ones
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3226 FOR_EACH_VEC_ELT (rtx_group_vec
, j
, group
)
3227 bitmap_ior_into (all_ones
, group
->group_kill
);
3231 bb_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3232 bitmap_copy (bb_info
->out
, all_ones
);
3238 BITMAP_FREE (all_ones
);
3243 /*----------------------------------------------------------------------------
3246 Solve the bitvector equations.
3247 ----------------------------------------------------------------------------*/
3250 /* Confluence function for blocks with no successors. Create an out
3251 set from the gen set of the exit block. This block logically has
3252 the exit block as a successor. */
3257 dse_confluence_0 (basic_block bb
)
3259 bb_info_t bb_info
= bb_table
[bb
->index
];
3261 if (bb
->index
== EXIT_BLOCK
)
3266 bb_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3267 bitmap_copy (bb_info
->out
, bb_table
[EXIT_BLOCK
]->gen
);
3271 /* Propagate the information from the in set of the dest of E to the
3272 out set of the src of E. If the various in or out sets are not
3273 there, that means they are all ones. */
3276 dse_confluence_n (edge e
)
3278 bb_info_t src_info
= bb_table
[e
->src
->index
];
3279 bb_info_t dest_info
= bb_table
[e
->dest
->index
];
3284 bitmap_and_into (src_info
->out
, dest_info
->in
);
3287 src_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3288 bitmap_copy (src_info
->out
, dest_info
->in
);
3295 /* Propagate the info from the out to the in set of BB_INDEX's basic
3296 block. There are three cases:
3298 1) The block has no kill set. In this case the kill set is all
3299 ones. It does not matter what the out set of the block is, none of
3300 the info can reach the top. The only thing that reaches the top is
3301 the gen set and we just copy the set.
3303 2) There is a kill set but no out set and bb has successors. In
3304 this case we just return. Eventually an out set will be created and
3305 it is better to wait than to create a set of ones.
3307 3) There is both a kill and out set. We apply the obvious transfer
3312 dse_transfer_function (int bb_index
)
3314 bb_info_t bb_info
= bb_table
[bb_index
];
3322 return bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
3323 bb_info
->out
, bb_info
->kill
);
3326 bb_info
->in
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3327 bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
3328 bb_info
->out
, bb_info
->kill
);
3338 /* Case 1 above. If there is already an in set, nothing
3344 bb_info
->in
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3345 bitmap_copy (bb_info
->in
, bb_info
->gen
);
3351 /* Solve the dataflow equations. */
3356 df_simple_dataflow (DF_BACKWARD
, NULL
, dse_confluence_0
,
3357 dse_confluence_n
, dse_transfer_function
,
3358 all_blocks
, df_get_postorder (DF_BACKWARD
),
3359 df_get_n_blocks (DF_BACKWARD
));
3360 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3364 fprintf (dump_file
, "\n\n*** Global dataflow info after analysis.\n");
3365 FOR_ALL_BB_FN (bb
, cfun
)
3367 bb_info_t bb_info
= bb_table
[bb
->index
];
3369 df_print_bb_index (bb
, dump_file
);
3371 bitmap_print (dump_file
, bb_info
->in
, " in: ", "\n");
3373 fprintf (dump_file
, " in: *MISSING*\n");
3375 bitmap_print (dump_file
, bb_info
->gen
, " gen: ", "\n");
3377 fprintf (dump_file
, " gen: *MISSING*\n");
3379 bitmap_print (dump_file
, bb_info
->kill
, " kill: ", "\n");
3381 fprintf (dump_file
, " kill: *MISSING*\n");
3383 bitmap_print (dump_file
, bb_info
->out
, " out: ", "\n");
3385 fprintf (dump_file
, " out: *MISSING*\n\n");
3392 /*----------------------------------------------------------------------------
3395 Delete the stores that can only be deleted using the global information.
3396 ----------------------------------------------------------------------------*/
3403 FOR_EACH_BB_FN (bb
, cfun
)
3405 bb_info_t bb_info
= bb_table
[bb
->index
];
3406 insn_info_t insn_info
= bb_info
->last_insn
;
3407 bitmap v
= bb_info
->out
;
3411 bool deleted
= false;
3412 if (dump_file
&& insn_info
->insn
)
3414 fprintf (dump_file
, "starting to process insn %d\n",
3415 INSN_UID (insn_info
->insn
));
3416 bitmap_print (dump_file
, v
, " v: ", "\n");
3419 /* There may have been code deleted by the dce pass run before
3422 && INSN_P (insn_info
->insn
)
3423 && (!insn_info
->cannot_delete
)
3424 && (!bitmap_empty_p (v
)))
3426 store_info
*store_info
= insn_info
->store_rec
;
3428 /* Try to delete the current insn. */
3431 /* Skip the clobbers. */
3432 while (!store_info
->is_set
)
3433 store_info
= store_info
->next
;
3435 HOST_WIDE_INT i
, offset
, width
;
3436 group_info
*group_info
= rtx_group_vec
[store_info
->group_id
];
3438 if (!store_info
->offset
.is_constant (&offset
)
3439 || !store_info
->width
.is_constant (&width
))
3443 HOST_WIDE_INT end
= offset
+ width
;
3444 for (i
= offset
; i
< end
; i
++)
3446 int index
= get_bitmap_index (group_info
, i
);
3448 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3449 fprintf (dump_file
, "i = %d, index = %d\n",
3451 if (index
== 0 || !bitmap_bit_p (v
, index
))
3453 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3454 fprintf (dump_file
, "failing at i = %d\n",
3464 && check_for_inc_dec_1 (insn_info
))
3466 delete_insn (insn_info
->insn
);
3467 insn_info
->insn
= NULL
;
3472 /* We do want to process the local info if the insn was
3473 deleted. For instance, if the insn did a wild read, we
3474 no longer need to trash the info. */
3476 && INSN_P (insn_info
->insn
)
3479 scan_stores (insn_info
->store_rec
, v
, NULL
);
3480 if (insn_info
->wild_read
)
3482 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3483 fprintf (dump_file
, "wild read\n");
3486 else if (insn_info
->read_rec
3487 || insn_info
->non_frame_wild_read
3488 || insn_info
->frame_read
)
3490 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3492 if (!insn_info
->non_frame_wild_read
3493 && !insn_info
->frame_read
)
3494 fprintf (dump_file
, "regular read\n");
3495 if (insn_info
->non_frame_wild_read
)
3496 fprintf (dump_file
, "non-frame wild read\n");
3497 if (insn_info
->frame_read
)
3498 fprintf (dump_file
, "frame read\n");
3500 scan_reads (insn_info
, v
, NULL
);
3504 insn_info
= insn_info
->prev_insn
;
3511 /*----------------------------------------------------------------------------
3514 Delete stores made redundant by earlier stores (which store the same
3515 value) that couldn't be eliminated.
3516 ----------------------------------------------------------------------------*/
3523 FOR_ALL_BB_FN (bb
, cfun
)
3525 bb_info_t bb_info
= bb_table
[bb
->index
];
3526 insn_info_t insn_info
= bb_info
->last_insn
;
3530 /* There may have been code deleted by the dce pass run before
3533 && INSN_P (insn_info
->insn
)
3534 && !insn_info
->cannot_delete
)
3536 store_info
*s_info
= insn_info
->store_rec
;
3538 while (s_info
&& !s_info
->is_set
)
3539 s_info
= s_info
->next
;
3541 && s_info
->redundant_reason
3542 && s_info
->redundant_reason
->insn
3543 && INSN_P (s_info
->redundant_reason
->insn
))
3545 rtx_insn
*rinsn
= s_info
->redundant_reason
->insn
;
3546 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3547 fprintf (dump_file
, "Locally deleting insn %d "
3548 "because insn %d stores the "
3549 "same value and couldn't be "
3551 INSN_UID (insn_info
->insn
),
3553 delete_dead_store_insn (insn_info
);
3556 insn_info
= insn_info
->prev_insn
;
3561 /*----------------------------------------------------------------------------
3564 Destroy everything left standing.
3565 ----------------------------------------------------------------------------*/
3570 bitmap_obstack_release (&dse_bitmap_obstack
);
3571 obstack_free (&dse_obstack
, NULL
);
3573 end_alias_analysis ();
3575 delete rtx_group_table
;
3576 rtx_group_table
= NULL
;
3577 rtx_group_vec
.release ();
3578 BITMAP_FREE (all_blocks
);
3579 BITMAP_FREE (scratch
);
3581 rtx_store_info_pool
.release ();
3582 read_info_type_pool
.release ();
3583 insn_info_type_pool
.release ();
3584 dse_bb_info_type_pool
.release ();
3585 group_info_pool
.release ();
3586 deferred_change_pool
.release ();
3590 /* -------------------------------------------------------------------------
3592 ------------------------------------------------------------------------- */
3594 /* Callback for running pass_rtl_dse. */
3597 rest_of_handle_dse (void)
3599 df_set_flags (DF_DEFER_INSN_RESCAN
);
3601 /* Need the notes since we must track live hardregs in the forwards
3603 df_note_add_problem ();
3611 df_set_flags (DF_LR_RUN_DCE
);
3613 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3614 fprintf (dump_file
, "doing global processing\n");
3624 fprintf (dump_file
, "dse: local deletions = %d, global deletions = %d\n",
3625 locally_deleted
, globally_deleted
);
3627 /* DSE can eliminate potentially-trapping MEMs.
3628 Remove any EH edges associated with them. */
3629 if ((locally_deleted
|| globally_deleted
)
3630 && cfun
->can_throw_non_call_exceptions
3631 && purge_all_dead_edges ())
3633 free_dominance_info (CDI_DOMINATORS
);
3642 const pass_data pass_data_rtl_dse1
=
3644 RTL_PASS
, /* type */
3646 OPTGROUP_NONE
, /* optinfo_flags */
3647 TV_DSE1
, /* tv_id */
3648 0, /* properties_required */
3649 0, /* properties_provided */
3650 0, /* properties_destroyed */
3651 0, /* todo_flags_start */
3652 TODO_df_finish
, /* todo_flags_finish */
3655 class pass_rtl_dse1
: public rtl_opt_pass
3658 pass_rtl_dse1 (gcc::context
*ctxt
)
3659 : rtl_opt_pass (pass_data_rtl_dse1
, ctxt
)
3662 /* opt_pass methods: */
3663 virtual bool gate (function
*)
3665 return optimize
> 0 && flag_dse
&& dbg_cnt (dse1
);
3668 virtual unsigned int execute (function
*) { return rest_of_handle_dse (); }
3670 }; // class pass_rtl_dse1
3675 make_pass_rtl_dse1 (gcc::context
*ctxt
)
3677 return new pass_rtl_dse1 (ctxt
);
3682 const pass_data pass_data_rtl_dse2
=
3684 RTL_PASS
, /* type */
3686 OPTGROUP_NONE
, /* optinfo_flags */
3687 TV_DSE2
, /* tv_id */
3688 0, /* properties_required */
3689 0, /* properties_provided */
3690 0, /* properties_destroyed */
3691 0, /* todo_flags_start */
3692 TODO_df_finish
, /* todo_flags_finish */
3695 class pass_rtl_dse2
: public rtl_opt_pass
3698 pass_rtl_dse2 (gcc::context
*ctxt
)
3699 : rtl_opt_pass (pass_data_rtl_dse2
, ctxt
)
3702 /* opt_pass methods: */
3703 virtual bool gate (function
*)
3705 return optimize
> 0 && flag_dse
&& dbg_cnt (dse2
);
3708 virtual unsigned int execute (function
*) { return rest_of_handle_dse (); }
3710 }; // class pass_rtl_dse2
3715 make_pass_rtl_dse2 (gcc::context
*ctxt
)
3717 return new pass_rtl_dse2 (ctxt
);