1 /* RTL dead store elimination.
2 Copyright (C) 2005-2020 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
,
850 /* Punt on stack pushes, those don't have REG_INC notes and we are
851 unprepared to deal with distribution of REG_ARGS_SIZE notes etc. */
852 subrtx_iterator::array_type array
;
853 FOR_EACH_SUBRTX (iter
, array
, PATTERN (insn
), NONCONST
)
856 if (GET_RTX_CLASS (GET_CODE (x
)) == RTX_AUTOINC
)
864 /* Entry point for postreload. If you work on reload_cse, or you need this
865 anywhere else, consider if you can provide register liveness information
866 and add a parameter to this function so that it can be passed down in
867 insn_info.fixed_regs_live. */
869 check_for_inc_dec (rtx_insn
*insn
)
871 insn_info_type insn_info
;
874 insn_info
.insn
= insn
;
875 insn_info
.fixed_regs_live
= NULL
;
876 note
= find_reg_note (insn
, REG_INC
, NULL_RTX
);
878 return for_each_inc_dec (PATTERN (insn
), emit_inc_dec_insn_before
,
881 /* Punt on stack pushes, those don't have REG_INC notes and we are
882 unprepared to deal with distribution of REG_ARGS_SIZE notes etc. */
883 subrtx_iterator::array_type array
;
884 FOR_EACH_SUBRTX (iter
, array
, PATTERN (insn
), NONCONST
)
887 if (GET_RTX_CLASS (GET_CODE (x
)) == RTX_AUTOINC
)
894 /* Delete the insn and free all of the fields inside INSN_INFO. */
897 delete_dead_store_insn (insn_info_t insn_info
)
899 read_info_t read_info
;
904 if (!check_for_inc_dec_1 (insn_info
))
906 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
907 fprintf (dump_file
, "Locally deleting insn %d\n",
908 INSN_UID (insn_info
->insn
));
910 free_store_info (insn_info
);
911 read_info
= insn_info
->read_rec
;
915 read_info_t next
= read_info
->next
;
916 read_info_type_pool
.remove (read_info
);
919 insn_info
->read_rec
= NULL
;
921 delete_insn (insn_info
->insn
);
923 insn_info
->insn
= NULL
;
925 insn_info
->wild_read
= false;
928 /* Return whether DECL, a local variable, can possibly escape the current
932 local_variable_can_escape (tree decl
)
934 if (TREE_ADDRESSABLE (decl
))
937 /* If this is a partitioned variable, we need to consider all the variables
938 in the partition. This is necessary because a store into one of them can
939 be replaced with a store into another and this may not change the outcome
940 of the escape analysis. */
941 if (cfun
->gimple_df
->decls_to_pointers
!= NULL
)
943 tree
*namep
= cfun
->gimple_df
->decls_to_pointers
->get (decl
);
945 return TREE_ADDRESSABLE (*namep
);
951 /* Return whether EXPR can possibly escape the current function scope. */
954 can_escape (tree expr
)
959 base
= get_base_address (expr
);
961 && !may_be_aliased (base
)
963 && !DECL_EXTERNAL (base
)
964 && !TREE_STATIC (base
)
965 && local_variable_can_escape (base
)))
970 /* Set the store* bitmaps offset_map_size* fields in GROUP based on
974 set_usage_bits (group_info
*group
, poly_int64 offset
, poly_int64 width
,
977 /* Non-constant offsets and widths act as global kills, so there's no point
978 trying to use them to derive global DSE candidates. */
979 HOST_WIDE_INT i
, const_offset
, const_width
;
980 bool expr_escapes
= can_escape (expr
);
981 if (offset
.is_constant (&const_offset
)
982 && width
.is_constant (&const_width
)
983 && const_offset
> -MAX_OFFSET
984 && const_offset
+ const_width
< MAX_OFFSET
)
985 for (i
= const_offset
; i
< const_offset
+ const_width
; ++i
)
993 store1
= group
->store1_n
;
994 store2
= group
->store2_n
;
995 escaped
= group
->escaped_n
;
1000 store1
= group
->store1_p
;
1001 store2
= group
->store2_p
;
1002 escaped
= group
->escaped_p
;
1006 if (!bitmap_set_bit (store1
, ai
))
1007 bitmap_set_bit (store2
, ai
);
1012 if (group
->offset_map_size_n
< ai
)
1013 group
->offset_map_size_n
= ai
;
1017 if (group
->offset_map_size_p
< ai
)
1018 group
->offset_map_size_p
= ai
;
1022 bitmap_set_bit (escaped
, ai
);
1027 reset_active_stores (void)
1029 active_local_stores
= NULL
;
1030 active_local_stores_len
= 0;
1033 /* Free all READ_REC of the LAST_INSN of BB_INFO. */
1036 free_read_records (bb_info_t bb_info
)
1038 insn_info_t insn_info
= bb_info
->last_insn
;
1039 read_info_t
*ptr
= &insn_info
->read_rec
;
1042 read_info_t next
= (*ptr
)->next
;
1043 read_info_type_pool
.remove (*ptr
);
1048 /* Set the BB_INFO so that the last insn is marked as a wild read. */
1051 add_wild_read (bb_info_t bb_info
)
1053 insn_info_t insn_info
= bb_info
->last_insn
;
1054 insn_info
->wild_read
= true;
1055 free_read_records (bb_info
);
1056 reset_active_stores ();
1059 /* Set the BB_INFO so that the last insn is marked as a wild read of
1060 non-frame locations. */
1063 add_non_frame_wild_read (bb_info_t bb_info
)
1065 insn_info_t insn_info
= bb_info
->last_insn
;
1066 insn_info
->non_frame_wild_read
= true;
1067 free_read_records (bb_info
);
1068 reset_active_stores ();
1071 /* Return true if X is a constant or one of the registers that behave
1072 as a constant over the life of a function. This is equivalent to
1073 !rtx_varies_p for memory addresses. */
1076 const_or_frame_p (rtx x
)
1081 if (GET_CODE (x
) == REG
)
1083 /* Note that we have to test for the actual rtx used for the frame
1084 and arg pointers and not just the register number in case we have
1085 eliminated the frame and/or arg pointer and are using it
1087 if (x
== frame_pointer_rtx
|| x
== hard_frame_pointer_rtx
1088 /* The arg pointer varies if it is not a fixed register. */
1089 || (x
== arg_pointer_rtx
&& fixed_regs
[ARG_POINTER_REGNUM
])
1090 || x
== pic_offset_table_rtx
)
1098 /* Take all reasonable action to put the address of MEM into the form
1099 that we can do analysis on.
1101 The gold standard is to get the address into the form: address +
1102 OFFSET where address is something that rtx_varies_p considers a
1103 constant. When we can get the address in this form, we can do
1104 global analysis on it. Note that for constant bases, address is
1105 not actually returned, only the group_id. The address can be
1108 If that fails, we try cselib to get a value we can at least use
1109 locally. If that fails we return false.
1111 The GROUP_ID is set to -1 for cselib bases and the index of the
1112 group for non_varying bases.
1114 FOR_READ is true if this is a mem read and false if not. */
1117 canon_address (rtx mem
,
1122 machine_mode address_mode
= get_address_mode (mem
);
1123 rtx mem_address
= XEXP (mem
, 0);
1124 rtx expanded_address
, address
;
1127 cselib_lookup (mem_address
, address_mode
, 1, GET_MODE (mem
));
1129 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1131 fprintf (dump_file
, " mem: ");
1132 print_inline_rtx (dump_file
, mem_address
, 0);
1133 fprintf (dump_file
, "\n");
1136 /* First see if just canon_rtx (mem_address) is const or frame,
1137 if not, try cselib_expand_value_rtx and call canon_rtx on that. */
1139 for (expanded
= 0; expanded
< 2; expanded
++)
1143 /* Use cselib to replace all of the reg references with the full
1144 expression. This will take care of the case where we have
1146 r_x = base + offset;
1151 val = *(base + offset); */
1153 expanded_address
= cselib_expand_value_rtx (mem_address
,
1156 /* If this fails, just go with the address from first
1158 if (!expanded_address
)
1162 expanded_address
= mem_address
;
1164 /* Split the address into canonical BASE + OFFSET terms. */
1165 address
= canon_rtx (expanded_address
);
1169 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1173 fprintf (dump_file
, "\n after cselib_expand address: ");
1174 print_inline_rtx (dump_file
, expanded_address
, 0);
1175 fprintf (dump_file
, "\n");
1178 fprintf (dump_file
, "\n after canon_rtx address: ");
1179 print_inline_rtx (dump_file
, address
, 0);
1180 fprintf (dump_file
, "\n");
1183 if (GET_CODE (address
) == CONST
)
1184 address
= XEXP (address
, 0);
1186 address
= strip_offset_and_add (address
, offset
);
1188 if (ADDR_SPACE_GENERIC_P (MEM_ADDR_SPACE (mem
))
1189 && const_or_frame_p (address
))
1191 group_info
*group
= get_group_info (address
);
1193 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1195 fprintf (dump_file
, " gid=%d offset=", group
->id
);
1196 print_dec (*offset
, dump_file
);
1197 fprintf (dump_file
, "\n");
1200 *group_id
= group
->id
;
1205 *base
= cselib_lookup (address
, address_mode
, true, GET_MODE (mem
));
1210 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1211 fprintf (dump_file
, " no cselib val - should be a wild read.\n");
1214 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1216 fprintf (dump_file
, " varying cselib base=%u:%u offset = ",
1217 (*base
)->uid
, (*base
)->hash
);
1218 print_dec (*offset
, dump_file
);
1219 fprintf (dump_file
, "\n");
1225 /* Clear the rhs field from the active_local_stores array. */
1228 clear_rhs_from_active_local_stores (void)
1230 insn_info_t ptr
= active_local_stores
;
1234 store_info
*store_info
= ptr
->store_rec
;
1235 /* Skip the clobbers. */
1236 while (!store_info
->is_set
)
1237 store_info
= store_info
->next
;
1239 store_info
->rhs
= NULL
;
1240 store_info
->const_rhs
= NULL
;
1242 ptr
= ptr
->next_local_store
;
1247 /* Mark byte POS bytes from the beginning of store S_INFO as unneeded. */
1250 set_position_unneeded (store_info
*s_info
, int pos
)
1252 if (__builtin_expect (s_info
->is_large
, false))
1254 if (bitmap_set_bit (s_info
->positions_needed
.large
.bmap
, pos
))
1255 s_info
->positions_needed
.large
.count
++;
1258 s_info
->positions_needed
.small_bitmask
1259 &= ~(HOST_WIDE_INT_1U
<< pos
);
1262 /* Mark the whole store S_INFO as unneeded. */
1265 set_all_positions_unneeded (store_info
*s_info
)
1267 if (__builtin_expect (s_info
->is_large
, false))
1269 HOST_WIDE_INT width
;
1270 if (s_info
->width
.is_constant (&width
))
1272 bitmap_set_range (s_info
->positions_needed
.large
.bmap
, 0, width
);
1273 s_info
->positions_needed
.large
.count
= width
;
1277 gcc_checking_assert (!s_info
->positions_needed
.large
.bmap
);
1278 s_info
->positions_needed
.large
.count
= 1;
1282 s_info
->positions_needed
.small_bitmask
= HOST_WIDE_INT_0U
;
1285 /* Return TRUE if any bytes from S_INFO store are needed. */
1288 any_positions_needed_p (store_info
*s_info
)
1290 if (__builtin_expect (s_info
->is_large
, false))
1292 HOST_WIDE_INT width
;
1293 if (s_info
->width
.is_constant (&width
))
1295 gcc_checking_assert (s_info
->positions_needed
.large
.bmap
);
1296 return s_info
->positions_needed
.large
.count
< width
;
1300 gcc_checking_assert (!s_info
->positions_needed
.large
.bmap
);
1301 return s_info
->positions_needed
.large
.count
== 0;
1305 return (s_info
->positions_needed
.small_bitmask
!= HOST_WIDE_INT_0U
);
1308 /* Return TRUE if all bytes START through START+WIDTH-1 from S_INFO
1309 store are known to be needed. */
1312 all_positions_needed_p (store_info
*s_info
, poly_int64 start
,
1315 gcc_assert (s_info
->rhs
);
1316 if (!s_info
->width
.is_constant ())
1318 gcc_assert (s_info
->is_large
1319 && !s_info
->positions_needed
.large
.bmap
);
1320 return s_info
->positions_needed
.large
.count
== 0;
1323 /* Otherwise, if START and WIDTH are non-constant, we're asking about
1324 a non-constant region of a constant-sized store. We can't say for
1325 sure that all positions are needed. */
1326 HOST_WIDE_INT const_start
, const_width
;
1327 if (!start
.is_constant (&const_start
)
1328 || !width
.is_constant (&const_width
))
1331 if (__builtin_expect (s_info
->is_large
, false))
1333 for (HOST_WIDE_INT i
= const_start
; i
< const_start
+ const_width
; ++i
)
1334 if (bitmap_bit_p (s_info
->positions_needed
.large
.bmap
, i
))
1340 unsigned HOST_WIDE_INT mask
1341 = lowpart_bitmask (const_width
) << const_start
;
1342 return (s_info
->positions_needed
.small_bitmask
& mask
) == mask
;
1347 static rtx
get_stored_val (store_info
*, machine_mode
, poly_int64
,
1348 poly_int64
, basic_block
, bool);
1351 /* BODY is an instruction pattern that belongs to INSN. Return 1 if
1352 there is a candidate store, after adding it to the appropriate
1353 local store group if so. */
1356 record_store (rtx body
, bb_info_t bb_info
)
1358 rtx mem
, rhs
, const_rhs
, mem_addr
;
1359 poly_int64 offset
= 0;
1360 poly_int64 width
= 0;
1361 insn_info_t insn_info
= bb_info
->last_insn
;
1362 store_info
*store_info
= NULL
;
1364 cselib_val
*base
= NULL
;
1365 insn_info_t ptr
, last
, redundant_reason
;
1366 bool store_is_unused
;
1368 if (GET_CODE (body
) != SET
&& GET_CODE (body
) != CLOBBER
)
1371 mem
= SET_DEST (body
);
1373 /* If this is not used, then this cannot be used to keep the insn
1374 from being deleted. On the other hand, it does provide something
1375 that can be used to prove that another store is dead. */
1377 = (find_reg_note (insn_info
->insn
, REG_UNUSED
, mem
) != NULL
);
1379 /* Check whether that value is a suitable memory location. */
1382 /* If the set or clobber is unused, then it does not effect our
1383 ability to get rid of the entire insn. */
1384 if (!store_is_unused
)
1385 insn_info
->cannot_delete
= true;
1389 /* At this point we know mem is a mem. */
1390 if (GET_MODE (mem
) == BLKmode
)
1392 HOST_WIDE_INT const_size
;
1393 if (GET_CODE (XEXP (mem
, 0)) == SCRATCH
)
1395 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1396 fprintf (dump_file
, " adding wild read for (clobber (mem:BLK (scratch))\n");
1397 add_wild_read (bb_info
);
1398 insn_info
->cannot_delete
= true;
1401 /* Handle (set (mem:BLK (addr) [... S36 ...]) (const_int 0))
1402 as memset (addr, 0, 36); */
1403 else if (!MEM_SIZE_KNOWN_P (mem
)
1404 || maybe_le (MEM_SIZE (mem
), 0)
1405 /* This is a limit on the bitmap size, which is only relevant
1406 for constant-sized MEMs. */
1407 || (MEM_SIZE (mem
).is_constant (&const_size
)
1408 && const_size
> MAX_OFFSET
)
1409 || GET_CODE (body
) != SET
1410 || !CONST_INT_P (SET_SRC (body
)))
1412 if (!store_is_unused
)
1414 /* If the set or clobber is unused, then it does not effect our
1415 ability to get rid of the entire insn. */
1416 insn_info
->cannot_delete
= true;
1417 clear_rhs_from_active_local_stores ();
1423 /* We can still process a volatile mem, we just cannot delete it. */
1424 if (MEM_VOLATILE_P (mem
))
1425 insn_info
->cannot_delete
= true;
1427 if (!canon_address (mem
, &group_id
, &offset
, &base
))
1429 clear_rhs_from_active_local_stores ();
1433 if (GET_MODE (mem
) == BLKmode
)
1434 width
= MEM_SIZE (mem
);
1436 width
= GET_MODE_SIZE (GET_MODE (mem
));
1438 if (!endpoint_representable_p (offset
, width
))
1440 clear_rhs_from_active_local_stores ();
1444 if (known_eq (width
, 0))
1449 /* In the restrictive case where the base is a constant or the
1450 frame pointer we can do global analysis. */
1453 = rtx_group_vec
[group_id
];
1454 tree expr
= MEM_EXPR (mem
);
1456 store_info
= rtx_store_info_pool
.allocate ();
1457 set_usage_bits (group
, offset
, width
, expr
);
1459 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1461 fprintf (dump_file
, " processing const base store gid=%d",
1463 print_range (dump_file
, offset
, width
);
1464 fprintf (dump_file
, "\n");
1469 if (may_be_sp_based_p (XEXP (mem
, 0)))
1470 insn_info
->stack_pointer_based
= true;
1471 insn_info
->contains_cselib_groups
= true;
1473 store_info
= cse_store_info_pool
.allocate ();
1476 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1478 fprintf (dump_file
, " processing cselib store ");
1479 print_range (dump_file
, offset
, width
);
1480 fprintf (dump_file
, "\n");
1484 const_rhs
= rhs
= NULL_RTX
;
1485 if (GET_CODE (body
) == SET
1486 /* No place to keep the value after ra. */
1487 && !reload_completed
1488 && (REG_P (SET_SRC (body
))
1489 || GET_CODE (SET_SRC (body
)) == SUBREG
1490 || CONSTANT_P (SET_SRC (body
)))
1491 && !MEM_VOLATILE_P (mem
)
1492 /* Sometimes the store and reload is used for truncation and
1494 && !(FLOAT_MODE_P (GET_MODE (mem
)) && (flag_float_store
)))
1496 rhs
= SET_SRC (body
);
1497 if (CONSTANT_P (rhs
))
1499 else if (body
== PATTERN (insn_info
->insn
))
1501 rtx tem
= find_reg_note (insn_info
->insn
, REG_EQUAL
, NULL_RTX
);
1502 if (tem
&& CONSTANT_P (XEXP (tem
, 0)))
1503 const_rhs
= XEXP (tem
, 0);
1505 if (const_rhs
== NULL_RTX
&& REG_P (rhs
))
1507 rtx tem
= cselib_expand_value_rtx (rhs
, scratch
, 5);
1509 if (tem
&& CONSTANT_P (tem
))
1514 /* Check to see if this stores causes some other stores to be
1516 ptr
= active_local_stores
;
1518 redundant_reason
= NULL
;
1519 mem
= canon_rtx (mem
);
1522 mem_addr
= base
->val_rtx
;
1525 group_info
*group
= rtx_group_vec
[group_id
];
1526 mem_addr
= group
->canon_base_addr
;
1528 if (maybe_ne (offset
, 0))
1529 mem_addr
= plus_constant (get_address_mode (mem
), mem_addr
, offset
);
1533 insn_info_t next
= ptr
->next_local_store
;
1534 class store_info
*s_info
= ptr
->store_rec
;
1537 /* Skip the clobbers. We delete the active insn if this insn
1538 shadows the set. To have been put on the active list, it
1539 has exactly on set. */
1540 while (!s_info
->is_set
)
1541 s_info
= s_info
->next
;
1543 if (s_info
->group_id
== group_id
&& s_info
->cse_base
== base
)
1546 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1548 fprintf (dump_file
, " trying store in insn=%d gid=%d",
1549 INSN_UID (ptr
->insn
), s_info
->group_id
);
1550 print_range (dump_file
, s_info
->offset
, s_info
->width
);
1551 fprintf (dump_file
, "\n");
1554 /* Even if PTR won't be eliminated as unneeded, if both
1555 PTR and this insn store the same constant value, we might
1556 eliminate this insn instead. */
1557 if (s_info
->const_rhs
1559 && known_subrange_p (offset
, width
,
1560 s_info
->offset
, s_info
->width
)
1561 && all_positions_needed_p (s_info
, offset
- s_info
->offset
,
1563 /* We can only remove the later store if the earlier aliases
1564 at least all accesses the later one. */
1565 && ((MEM_ALIAS_SET (mem
) == MEM_ALIAS_SET (s_info
->mem
)
1566 || alias_set_subset_of (MEM_ALIAS_SET (mem
),
1567 MEM_ALIAS_SET (s_info
->mem
)))
1568 && (!MEM_EXPR (s_info
->mem
)
1569 || refs_same_for_tbaa_p (MEM_EXPR (s_info
->mem
),
1572 if (GET_MODE (mem
) == BLKmode
)
1574 if (GET_MODE (s_info
->mem
) == BLKmode
1575 && s_info
->const_rhs
== const_rhs
)
1576 redundant_reason
= ptr
;
1578 else if (s_info
->const_rhs
== const0_rtx
1579 && const_rhs
== const0_rtx
)
1580 redundant_reason
= ptr
;
1585 val
= get_stored_val (s_info
, GET_MODE (mem
), offset
, width
,
1586 BLOCK_FOR_INSN (insn_info
->insn
),
1588 if (get_insns () != NULL
)
1591 if (val
&& rtx_equal_p (val
, const_rhs
))
1592 redundant_reason
= ptr
;
1596 HOST_WIDE_INT begin_unneeded
, const_s_width
, const_width
;
1597 if (known_subrange_p (s_info
->offset
, s_info
->width
, offset
, width
))
1598 /* The new store touches every byte that S_INFO does. */
1599 set_all_positions_unneeded (s_info
);
1600 else if ((offset
- s_info
->offset
).is_constant (&begin_unneeded
)
1601 && s_info
->width
.is_constant (&const_s_width
)
1602 && width
.is_constant (&const_width
))
1604 HOST_WIDE_INT end_unneeded
= begin_unneeded
+ const_width
;
1605 begin_unneeded
= MAX (begin_unneeded
, 0);
1606 end_unneeded
= MIN (end_unneeded
, const_s_width
);
1607 for (i
= begin_unneeded
; i
< end_unneeded
; ++i
)
1608 set_position_unneeded (s_info
, i
);
1612 /* We don't know which parts of S_INFO are needed and
1613 which aren't, so invalidate the RHS. */
1615 s_info
->const_rhs
= NULL
;
1618 else if (s_info
->rhs
)
1619 /* Need to see if it is possible for this store to overwrite
1620 the value of store_info. If it is, set the rhs to NULL to
1621 keep it from being used to remove a load. */
1623 if (canon_output_dependence (s_info
->mem
, true,
1624 mem
, GET_MODE (mem
),
1628 s_info
->const_rhs
= NULL
;
1632 /* An insn can be deleted if every position of every one of
1633 its s_infos is zero. */
1634 if (any_positions_needed_p (s_info
))
1639 insn_info_t insn_to_delete
= ptr
;
1641 active_local_stores_len
--;
1643 last
->next_local_store
= ptr
->next_local_store
;
1645 active_local_stores
= ptr
->next_local_store
;
1647 if (!insn_to_delete
->cannot_delete
)
1648 delete_dead_store_insn (insn_to_delete
);
1656 /* Finish filling in the store_info. */
1657 store_info
->next
= insn_info
->store_rec
;
1658 insn_info
->store_rec
= store_info
;
1659 store_info
->mem
= mem
;
1660 store_info
->mem_addr
= mem_addr
;
1661 store_info
->cse_base
= base
;
1662 HOST_WIDE_INT const_width
;
1663 if (!width
.is_constant (&const_width
))
1665 store_info
->is_large
= true;
1666 store_info
->positions_needed
.large
.count
= 0;
1667 store_info
->positions_needed
.large
.bmap
= NULL
;
1669 else if (const_width
> HOST_BITS_PER_WIDE_INT
)
1671 store_info
->is_large
= true;
1672 store_info
->positions_needed
.large
.count
= 0;
1673 store_info
->positions_needed
.large
.bmap
= BITMAP_ALLOC (&dse_bitmap_obstack
);
1677 store_info
->is_large
= false;
1678 store_info
->positions_needed
.small_bitmask
1679 = lowpart_bitmask (const_width
);
1681 store_info
->group_id
= group_id
;
1682 store_info
->offset
= offset
;
1683 store_info
->width
= width
;
1684 store_info
->is_set
= GET_CODE (body
) == SET
;
1685 store_info
->rhs
= rhs
;
1686 store_info
->const_rhs
= const_rhs
;
1687 store_info
->redundant_reason
= redundant_reason
;
1689 /* If this is a clobber, we return 0. We will only be able to
1690 delete this insn if there is only one store USED store, but we
1691 can use the clobber to delete other stores earlier. */
1692 return store_info
->is_set
? 1 : 0;
1697 dump_insn_info (const char * start
, insn_info_t insn_info
)
1699 fprintf (dump_file
, "%s insn=%d %s\n", start
,
1700 INSN_UID (insn_info
->insn
),
1701 insn_info
->store_rec
? "has store" : "naked");
1705 /* If the modes are different and the value's source and target do not
1706 line up, we need to extract the value from lower part of the rhs of
1707 the store, shift it, and then put it into a form that can be shoved
1708 into the read_insn. This function generates a right SHIFT of a
1709 value that is at least ACCESS_SIZE bytes wide of READ_MODE. The
1710 shift sequence is returned or NULL if we failed to find a
1714 find_shift_sequence (poly_int64 access_size
,
1715 store_info
*store_info
,
1716 machine_mode read_mode
,
1717 poly_int64 shift
, bool speed
, bool require_cst
)
1719 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1720 scalar_int_mode new_mode
;
1721 rtx read_reg
= NULL
;
1723 /* If a constant was stored into memory, try to simplify it here,
1724 otherwise the cost of the shift might preclude this optimization
1725 e.g. at -Os, even when no actual shift will be needed. */
1726 if (store_info
->const_rhs
)
1728 auto new_mode
= smallest_int_mode_for_size (access_size
* BITS_PER_UNIT
);
1729 auto byte
= subreg_lowpart_offset (new_mode
, store_mode
);
1731 = simplify_subreg (new_mode
, store_info
->const_rhs
, store_mode
, byte
);
1732 if (ret
&& CONSTANT_P (ret
))
1734 rtx shift_rtx
= gen_int_shift_amount (new_mode
, shift
);
1735 ret
= simplify_const_binary_operation (LSHIFTRT
, new_mode
, ret
,
1737 if (ret
&& CONSTANT_P (ret
))
1739 byte
= subreg_lowpart_offset (read_mode
, new_mode
);
1740 ret
= simplify_subreg (read_mode
, ret
, new_mode
, byte
);
1741 if (ret
&& CONSTANT_P (ret
)
1742 && (set_src_cost (ret
, read_mode
, speed
)
1743 <= COSTS_N_INSNS (1)))
1752 /* Some machines like the x86 have shift insns for each size of
1753 operand. Other machines like the ppc or the ia-64 may only have
1754 shift insns that shift values within 32 or 64 bit registers.
1755 This loop tries to find the smallest shift insn that will right
1756 justify the value we want to read but is available in one insn on
1759 opt_scalar_int_mode new_mode_iter
;
1760 FOR_EACH_MODE_FROM (new_mode_iter
,
1761 smallest_int_mode_for_size (GET_MODE_BITSIZE (read_mode
)))
1763 rtx target
, new_reg
, new_lhs
;
1764 rtx_insn
*shift_seq
, *insn
;
1767 new_mode
= new_mode_iter
.require ();
1768 if (GET_MODE_BITSIZE (new_mode
) > BITS_PER_WORD
)
1771 /* Try a wider mode if truncating the store mode to NEW_MODE
1772 requires a real instruction. */
1773 if (maybe_lt (GET_MODE_SIZE (new_mode
), GET_MODE_SIZE (store_mode
))
1774 && !TRULY_NOOP_TRUNCATION_MODES_P (new_mode
, store_mode
))
1777 /* Also try a wider mode if the necessary punning is either not
1778 desirable or not possible. */
1779 if (!CONSTANT_P (store_info
->rhs
)
1780 && !targetm
.modes_tieable_p (new_mode
, store_mode
))
1783 if (multiple_p (shift
, GET_MODE_BITSIZE (new_mode
))
1784 && known_le (GET_MODE_SIZE (new_mode
), GET_MODE_SIZE (store_mode
)))
1786 /* Try to implement the shift using a subreg. */
1788 = subreg_offset_from_lsb (new_mode
, store_mode
, shift
);
1789 rtx rhs_subreg
= simplify_gen_subreg (new_mode
, store_info
->rhs
,
1790 store_mode
, offset
);
1794 = extract_low_bits (read_mode
, new_mode
, copy_rtx (rhs_subreg
));
1799 if (maybe_lt (GET_MODE_SIZE (new_mode
), access_size
))
1802 new_reg
= gen_reg_rtx (new_mode
);
1806 /* In theory we could also check for an ashr. Ian Taylor knows
1807 of one dsp where the cost of these two was not the same. But
1808 this really is a rare case anyway. */
1809 target
= expand_binop (new_mode
, lshr_optab
, new_reg
,
1810 gen_int_shift_amount (new_mode
, shift
),
1811 new_reg
, 1, OPTAB_DIRECT
);
1813 shift_seq
= get_insns ();
1816 if (target
!= new_reg
|| shift_seq
== NULL
)
1820 for (insn
= shift_seq
; insn
!= NULL_RTX
; insn
= NEXT_INSN (insn
))
1822 cost
+= insn_cost (insn
, speed
);
1824 /* The computation up to here is essentially independent
1825 of the arguments and could be precomputed. It may
1826 not be worth doing so. We could precompute if
1827 worthwhile or at least cache the results. The result
1828 technically depends on both SHIFT and ACCESS_SIZE,
1829 but in practice the answer will depend only on ACCESS_SIZE. */
1831 if (cost
> COSTS_N_INSNS (1))
1834 new_lhs
= extract_low_bits (new_mode
, store_mode
,
1835 copy_rtx (store_info
->rhs
));
1836 if (new_lhs
== NULL_RTX
)
1839 /* We found an acceptable shift. Generate a move to
1840 take the value from the store and put it into the
1841 shift pseudo, then shift it, then generate another
1842 move to put in into the target of the read. */
1843 emit_move_insn (new_reg
, new_lhs
);
1844 emit_insn (shift_seq
);
1845 read_reg
= extract_low_bits (read_mode
, new_mode
, new_reg
);
1853 /* Call back for note_stores to find the hard regs set or clobbered by
1854 insn. Data is a bitmap of the hardregs set so far. */
1857 look_for_hardregs (rtx x
, const_rtx pat ATTRIBUTE_UNUSED
, void *data
)
1859 bitmap regs_set
= (bitmap
) data
;
1862 && HARD_REGISTER_P (x
))
1863 bitmap_set_range (regs_set
, REGNO (x
), REG_NREGS (x
));
1866 /* Helper function for replace_read and record_store.
1867 Attempt to return a value of mode READ_MODE stored in STORE_INFO,
1868 consisting of READ_WIDTH bytes starting from READ_OFFSET. Return NULL
1869 if not successful. If REQUIRE_CST is true, return always constant. */
1872 get_stored_val (store_info
*store_info
, machine_mode read_mode
,
1873 poly_int64 read_offset
, poly_int64 read_width
,
1874 basic_block bb
, bool require_cst
)
1876 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1880 /* To get here the read is within the boundaries of the write so
1881 shift will never be negative. Start out with the shift being in
1883 if (store_mode
== BLKmode
)
1885 else if (BYTES_BIG_ENDIAN
)
1886 gap
= ((store_info
->offset
+ store_info
->width
)
1887 - (read_offset
+ read_width
));
1889 gap
= read_offset
- store_info
->offset
;
1891 if (gap
.is_constant () && maybe_ne (gap
, 0))
1893 poly_int64 shift
= gap
* BITS_PER_UNIT
;
1894 poly_int64 access_size
= GET_MODE_SIZE (read_mode
) + gap
;
1895 read_reg
= find_shift_sequence (access_size
, store_info
, read_mode
,
1896 shift
, optimize_bb_for_speed_p (bb
),
1899 else if (store_mode
== BLKmode
)
1901 /* The store is a memset (addr, const_val, const_size). */
1902 gcc_assert (CONST_INT_P (store_info
->rhs
));
1903 scalar_int_mode int_store_mode
;
1904 if (!int_mode_for_mode (read_mode
).exists (&int_store_mode
))
1905 read_reg
= NULL_RTX
;
1906 else if (store_info
->rhs
== const0_rtx
)
1907 read_reg
= extract_low_bits (read_mode
, int_store_mode
, const0_rtx
);
1908 else if (GET_MODE_BITSIZE (int_store_mode
) > HOST_BITS_PER_WIDE_INT
1909 || BITS_PER_UNIT
>= HOST_BITS_PER_WIDE_INT
)
1910 read_reg
= NULL_RTX
;
1913 unsigned HOST_WIDE_INT c
1914 = INTVAL (store_info
->rhs
)
1915 & ((HOST_WIDE_INT_1
<< BITS_PER_UNIT
) - 1);
1916 int shift
= BITS_PER_UNIT
;
1917 while (shift
< HOST_BITS_PER_WIDE_INT
)
1922 read_reg
= gen_int_mode (c
, int_store_mode
);
1923 read_reg
= extract_low_bits (read_mode
, int_store_mode
, read_reg
);
1926 else if (store_info
->const_rhs
1928 || GET_MODE_CLASS (read_mode
) != GET_MODE_CLASS (store_mode
)))
1929 read_reg
= extract_low_bits (read_mode
, store_mode
,
1930 copy_rtx (store_info
->const_rhs
));
1932 read_reg
= extract_low_bits (read_mode
, store_mode
,
1933 copy_rtx (store_info
->rhs
));
1934 if (require_cst
&& read_reg
&& !CONSTANT_P (read_reg
))
1935 read_reg
= NULL_RTX
;
1939 /* Take a sequence of:
1962 Depending on the alignment and the mode of the store and
1966 The STORE_INFO and STORE_INSN are for the store and READ_INFO
1967 and READ_INSN are for the read. Return true if the replacement
1971 replace_read (store_info
*store_info
, insn_info_t store_insn
,
1972 read_info_t read_info
, insn_info_t read_insn
, rtx
*loc
,
1975 machine_mode store_mode
= GET_MODE (store_info
->mem
);
1976 machine_mode read_mode
= GET_MODE (read_info
->mem
);
1977 rtx_insn
*insns
, *this_insn
;
1984 /* Create a sequence of instructions to set up the read register.
1985 This sequence goes immediately before the store and its result
1986 is read by the load.
1988 We need to keep this in perspective. We are replacing a read
1989 with a sequence of insns, but the read will almost certainly be
1990 in cache, so it is not going to be an expensive one. Thus, we
1991 are not willing to do a multi insn shift or worse a subroutine
1992 call to get rid of the read. */
1993 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1994 fprintf (dump_file
, "trying to replace %smode load in insn %d"
1995 " from %smode store in insn %d\n",
1996 GET_MODE_NAME (read_mode
), INSN_UID (read_insn
->insn
),
1997 GET_MODE_NAME (store_mode
), INSN_UID (store_insn
->insn
));
1999 bb
= BLOCK_FOR_INSN (read_insn
->insn
);
2000 read_reg
= get_stored_val (store_info
,
2001 read_mode
, read_info
->offset
, read_info
->width
,
2003 if (read_reg
== NULL_RTX
)
2006 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2007 fprintf (dump_file
, " -- could not extract bits of stored value\n");
2010 /* Force the value into a new register so that it won't be clobbered
2011 between the store and the load. */
2012 read_reg
= copy_to_mode_reg (read_mode
, read_reg
);
2013 insns
= get_insns ();
2016 if (insns
!= NULL_RTX
)
2018 /* Now we have to scan the set of new instructions to see if the
2019 sequence contains and sets of hardregs that happened to be
2020 live at this point. For instance, this can happen if one of
2021 the insns sets the CC and the CC happened to be live at that
2022 point. This does occasionally happen, see PR 37922. */
2023 bitmap regs_set
= BITMAP_ALLOC (®_obstack
);
2025 for (this_insn
= insns
;
2026 this_insn
!= NULL_RTX
; this_insn
= NEXT_INSN (this_insn
))
2028 if (insn_invalid_p (this_insn
, false))
2030 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2032 fprintf (dump_file
, " -- replacing the loaded MEM with ");
2033 print_simple_rtl (dump_file
, read_reg
);
2034 fprintf (dump_file
, " led to an invalid instruction\n");
2036 BITMAP_FREE (regs_set
);
2039 note_stores (this_insn
, look_for_hardregs
, regs_set
);
2042 bitmap_and_into (regs_set
, regs_live
);
2043 if (!bitmap_empty_p (regs_set
))
2045 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2047 fprintf (dump_file
, "abandoning replacement because sequence "
2048 "clobbers live hardregs:");
2049 df_print_regset (dump_file
, regs_set
);
2052 BITMAP_FREE (regs_set
);
2055 BITMAP_FREE (regs_set
);
2058 subrtx_iterator::array_type array
;
2059 FOR_EACH_SUBRTX (iter
, array
, *loc
, NONCONST
)
2061 const_rtx x
= *iter
;
2062 if (GET_RTX_CLASS (GET_CODE (x
)) == RTX_AUTOINC
)
2064 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2065 fprintf (dump_file
, " -- replacing the MEM failed due to address "
2071 if (validate_change (read_insn
->insn
, loc
, read_reg
, 0))
2073 deferred_change
*change
= deferred_change_pool
.allocate ();
2075 /* Insert this right before the store insn where it will be safe
2076 from later insns that might change it before the read. */
2077 emit_insn_before (insns
, store_insn
->insn
);
2079 /* And now for the kludge part: cselib croaks if you just
2080 return at this point. There are two reasons for this:
2082 1) Cselib has an idea of how many pseudos there are and
2083 that does not include the new ones we just added.
2085 2) Cselib does not know about the move insn we added
2086 above the store_info, and there is no way to tell it
2087 about it, because it has "moved on".
2089 Problem (1) is fixable with a certain amount of engineering.
2090 Problem (2) is requires starting the bb from scratch. This
2093 So we are just going to have to lie. The move/extraction
2094 insns are not really an issue, cselib did not see them. But
2095 the use of the new pseudo read_insn is a real problem because
2096 cselib has not scanned this insn. The way that we solve this
2097 problem is that we are just going to put the mem back for now
2098 and when we are finished with the block, we undo this. We
2099 keep a table of mems to get rid of. At the end of the basic
2100 block we can put them back. */
2102 *loc
= read_info
->mem
;
2103 change
->next
= deferred_change_list
;
2104 deferred_change_list
= change
;
2106 change
->reg
= read_reg
;
2108 /* Get rid of the read_info, from the point of view of the
2109 rest of dse, play like this read never happened. */
2110 read_insn
->read_rec
= read_info
->next
;
2111 read_info_type_pool
.remove (read_info
);
2112 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2114 fprintf (dump_file
, " -- replaced the loaded MEM with ");
2115 print_simple_rtl (dump_file
, read_reg
);
2116 fprintf (dump_file
, "\n");
2122 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2124 fprintf (dump_file
, " -- replacing the loaded MEM with ");
2125 print_simple_rtl (dump_file
, read_reg
);
2126 fprintf (dump_file
, " led to an invalid instruction\n");
2132 /* Check the address of MEM *LOC and kill any appropriate stores that may
2136 check_mem_read_rtx (rtx
*loc
, bb_info_t bb_info
)
2138 rtx mem
= *loc
, mem_addr
;
2139 insn_info_t insn_info
;
2140 poly_int64 offset
= 0;
2141 poly_int64 width
= 0;
2142 cselib_val
*base
= NULL
;
2144 read_info_t read_info
;
2146 insn_info
= bb_info
->last_insn
;
2148 if ((MEM_ALIAS_SET (mem
) == ALIAS_SET_MEMORY_BARRIER
)
2149 || MEM_VOLATILE_P (mem
))
2151 if (crtl
->stack_protect_guard
2152 && (MEM_EXPR (mem
) == crtl
->stack_protect_guard
2153 || (crtl
->stack_protect_guard_decl
2154 && MEM_EXPR (mem
) == crtl
->stack_protect_guard_decl
))
2155 && MEM_VOLATILE_P (mem
))
2157 /* This is either the stack protector canary on the stack,
2158 which ought to be written by a MEM_VOLATILE_P store and
2159 thus shouldn't be deleted and is read at the very end of
2160 function, but shouldn't conflict with any other store.
2161 Or it is __stack_chk_guard variable or TLS or whatever else
2162 MEM holding the canary value, which really shouldn't be
2163 ever modified in -fstack-protector* protected functions,
2164 otherwise the prologue store wouldn't match the epilogue
2166 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2167 fprintf (dump_file
, " stack protector canary read ignored.\n");
2168 insn_info
->cannot_delete
= true;
2172 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2173 fprintf (dump_file
, " adding wild read, volatile or barrier.\n");
2174 add_wild_read (bb_info
);
2175 insn_info
->cannot_delete
= true;
2179 /* If it is reading readonly mem, then there can be no conflict with
2181 if (MEM_READONLY_P (mem
))
2184 if (!canon_address (mem
, &group_id
, &offset
, &base
))
2186 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2187 fprintf (dump_file
, " adding wild read, canon_address failure.\n");
2188 add_wild_read (bb_info
);
2192 if (GET_MODE (mem
) == BLKmode
)
2195 width
= GET_MODE_SIZE (GET_MODE (mem
));
2197 if (!endpoint_representable_p (offset
, known_eq (width
, -1) ? 1 : width
))
2199 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2200 fprintf (dump_file
, " adding wild read, due to overflow.\n");
2201 add_wild_read (bb_info
);
2205 read_info
= read_info_type_pool
.allocate ();
2206 read_info
->group_id
= group_id
;
2207 read_info
->mem
= mem
;
2208 read_info
->offset
= offset
;
2209 read_info
->width
= width
;
2210 read_info
->next
= insn_info
->read_rec
;
2211 insn_info
->read_rec
= read_info
;
2213 mem_addr
= base
->val_rtx
;
2216 group_info
*group
= rtx_group_vec
[group_id
];
2217 mem_addr
= group
->canon_base_addr
;
2219 if (maybe_ne (offset
, 0))
2220 mem_addr
= plus_constant (get_address_mode (mem
), mem_addr
, offset
);
2224 /* This is the restricted case where the base is a constant or
2225 the frame pointer and offset is a constant. */
2226 insn_info_t i_ptr
= active_local_stores
;
2227 insn_info_t last
= NULL
;
2229 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2231 if (!known_size_p (width
))
2232 fprintf (dump_file
, " processing const load gid=%d[BLK]\n",
2236 fprintf (dump_file
, " processing const load gid=%d", group_id
);
2237 print_range (dump_file
, offset
, width
);
2238 fprintf (dump_file
, "\n");
2244 bool remove
= false;
2245 store_info
*store_info
= i_ptr
->store_rec
;
2247 /* Skip the clobbers. */
2248 while (!store_info
->is_set
)
2249 store_info
= store_info
->next
;
2251 /* There are three cases here. */
2252 if (store_info
->group_id
< 0)
2253 /* We have a cselib store followed by a read from a
2256 = canon_true_dependence (store_info
->mem
,
2257 GET_MODE (store_info
->mem
),
2258 store_info
->mem_addr
,
2261 else if (group_id
== store_info
->group_id
)
2263 /* This is a block mode load. We may get lucky and
2264 canon_true_dependence may save the day. */
2265 if (!known_size_p (width
))
2267 = canon_true_dependence (store_info
->mem
,
2268 GET_MODE (store_info
->mem
),
2269 store_info
->mem_addr
,
2272 /* If this read is just reading back something that we just
2273 stored, rewrite the read. */
2277 && known_subrange_p (offset
, width
, store_info
->offset
,
2279 && all_positions_needed_p (store_info
,
2280 offset
- store_info
->offset
,
2282 && replace_read (store_info
, i_ptr
, read_info
,
2283 insn_info
, loc
, bb_info
->regs_live
))
2286 /* The bases are the same, just see if the offsets
2288 if (ranges_maybe_overlap_p (offset
, width
,
2296 The else case that is missing here is that the
2297 bases are constant but different. There is nothing
2298 to do here because there is no overlap. */
2302 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2303 dump_insn_info ("removing from active", i_ptr
);
2305 active_local_stores_len
--;
2307 last
->next_local_store
= i_ptr
->next_local_store
;
2309 active_local_stores
= i_ptr
->next_local_store
;
2313 i_ptr
= i_ptr
->next_local_store
;
2318 insn_info_t i_ptr
= active_local_stores
;
2319 insn_info_t last
= NULL
;
2320 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2322 fprintf (dump_file
, " processing cselib load mem:");
2323 print_inline_rtx (dump_file
, mem
, 0);
2324 fprintf (dump_file
, "\n");
2329 bool remove
= false;
2330 store_info
*store_info
= i_ptr
->store_rec
;
2332 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2333 fprintf (dump_file
, " processing cselib load against insn %d\n",
2334 INSN_UID (i_ptr
->insn
));
2336 /* Skip the clobbers. */
2337 while (!store_info
->is_set
)
2338 store_info
= store_info
->next
;
2340 /* If this read is just reading back something that we just
2341 stored, rewrite the read. */
2343 && store_info
->group_id
== -1
2344 && store_info
->cse_base
== base
2345 && known_subrange_p (offset
, width
, store_info
->offset
,
2347 && all_positions_needed_p (store_info
,
2348 offset
- store_info
->offset
, width
)
2349 && replace_read (store_info
, i_ptr
, read_info
, insn_info
, loc
,
2350 bb_info
->regs_live
))
2353 remove
= canon_true_dependence (store_info
->mem
,
2354 GET_MODE (store_info
->mem
),
2355 store_info
->mem_addr
,
2360 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2361 dump_insn_info ("removing from active", i_ptr
);
2363 active_local_stores_len
--;
2365 last
->next_local_store
= i_ptr
->next_local_store
;
2367 active_local_stores
= i_ptr
->next_local_store
;
2371 i_ptr
= i_ptr
->next_local_store
;
2376 /* A note_uses callback in which DATA points the INSN_INFO for
2377 as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns
2378 true for any part of *LOC. */
2381 check_mem_read_use (rtx
*loc
, void *data
)
2383 subrtx_ptr_iterator::array_type array
;
2384 FOR_EACH_SUBRTX_PTR (iter
, array
, loc
, NONCONST
)
2388 check_mem_read_rtx (loc
, (bb_info_t
) data
);
2393 /* Get arguments passed to CALL_INSN. Return TRUE if successful.
2394 So far it only handles arguments passed in registers. */
2397 get_call_args (rtx call_insn
, tree fn
, rtx
*args
, int nargs
)
2399 CUMULATIVE_ARGS args_so_far_v
;
2400 cumulative_args_t args_so_far
;
2404 INIT_CUMULATIVE_ARGS (args_so_far_v
, TREE_TYPE (fn
), NULL_RTX
, 0, 3);
2405 args_so_far
= pack_cumulative_args (&args_so_far_v
);
2407 arg
= TYPE_ARG_TYPES (TREE_TYPE (fn
));
2409 arg
!= void_list_node
&& idx
< nargs
;
2410 arg
= TREE_CHAIN (arg
), idx
++)
2412 scalar_int_mode mode
;
2415 if (!is_int_mode (TYPE_MODE (TREE_VALUE (arg
)), &mode
))
2418 function_arg_info
arg (mode
, /*named=*/true);
2419 reg
= targetm
.calls
.function_arg (args_so_far
, arg
);
2420 if (!reg
|| !REG_P (reg
) || GET_MODE (reg
) != mode
)
2423 for (link
= CALL_INSN_FUNCTION_USAGE (call_insn
);
2425 link
= XEXP (link
, 1))
2426 if (GET_CODE (XEXP (link
, 0)) == USE
)
2428 scalar_int_mode arg_mode
;
2429 args
[idx
] = XEXP (XEXP (link
, 0), 0);
2430 if (REG_P (args
[idx
])
2431 && REGNO (args
[idx
]) == REGNO (reg
)
2432 && (GET_MODE (args
[idx
]) == mode
2433 || (is_int_mode (GET_MODE (args
[idx
]), &arg_mode
)
2434 && (GET_MODE_SIZE (arg_mode
) <= UNITS_PER_WORD
)
2435 && (GET_MODE_SIZE (arg_mode
) > GET_MODE_SIZE (mode
)))))
2441 tmp
= cselib_expand_value_rtx (args
[idx
], scratch
, 5);
2442 if (GET_MODE (args
[idx
]) != mode
)
2444 if (!tmp
|| !CONST_INT_P (tmp
))
2446 tmp
= gen_int_mode (INTVAL (tmp
), mode
);
2451 targetm
.calls
.function_arg_advance (args_so_far
, arg
);
2453 if (arg
!= void_list_node
|| idx
!= nargs
)
2458 /* Return a bitmap of the fixed registers contained in IN. */
2461 copy_fixed_regs (const_bitmap in
)
2465 ret
= ALLOC_REG_SET (NULL
);
2466 bitmap_and (ret
, in
, bitmap_view
<HARD_REG_SET
> (fixed_reg_set
));
2470 /* Apply record_store to all candidate stores in INSN. Mark INSN
2471 if some part of it is not a candidate store and assigns to a
2472 non-register target. */
2475 scan_insn (bb_info_t bb_info
, rtx_insn
*insn
, int max_active_local_stores
)
2478 insn_info_type
*insn_info
= insn_info_type_pool
.allocate ();
2480 memset (insn_info
, 0, sizeof (struct insn_info_type
));
2482 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2483 fprintf (dump_file
, "\n**scanning insn=%d\n",
2486 insn_info
->prev_insn
= bb_info
->last_insn
;
2487 insn_info
->insn
= insn
;
2488 bb_info
->last_insn
= insn_info
;
2490 if (DEBUG_INSN_P (insn
))
2492 insn_info
->cannot_delete
= true;
2496 /* Look at all of the uses in the insn. */
2497 note_uses (&PATTERN (insn
), check_mem_read_use
, bb_info
);
2503 tree memset_call
= NULL_TREE
;
2505 insn_info
->cannot_delete
= true;
2507 /* Const functions cannot do anything bad i.e. read memory,
2508 however, they can read their parameters which may have
2509 been pushed onto the stack.
2510 memset and bzero don't read memory either. */
2511 const_call
= RTL_CONST_CALL_P (insn
);
2513 && (call
= get_call_rtx_from (insn
))
2514 && (sym
= XEXP (XEXP (call
, 0), 0))
2515 && GET_CODE (sym
) == SYMBOL_REF
2516 && SYMBOL_REF_DECL (sym
)
2517 && TREE_CODE (SYMBOL_REF_DECL (sym
)) == FUNCTION_DECL
2518 && fndecl_built_in_p (SYMBOL_REF_DECL (sym
), BUILT_IN_MEMSET
))
2519 memset_call
= SYMBOL_REF_DECL (sym
);
2521 if (const_call
|| memset_call
)
2523 insn_info_t i_ptr
= active_local_stores
;
2524 insn_info_t last
= NULL
;
2526 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2527 fprintf (dump_file
, "%s call %d\n",
2528 const_call
? "const" : "memset", INSN_UID (insn
));
2530 /* See the head comment of the frame_read field. */
2531 if (reload_completed
2532 /* Tail calls are storing their arguments using
2533 arg pointer. If it is a frame pointer on the target,
2534 even before reload we need to kill frame pointer based
2536 || (SIBLING_CALL_P (insn
)
2537 && HARD_FRAME_POINTER_IS_ARG_POINTER
))
2538 insn_info
->frame_read
= true;
2540 /* Loop over the active stores and remove those which are
2541 killed by the const function call. */
2544 bool remove_store
= false;
2546 /* The stack pointer based stores are always killed. */
2547 if (i_ptr
->stack_pointer_based
)
2548 remove_store
= true;
2550 /* If the frame is read, the frame related stores are killed. */
2551 else if (insn_info
->frame_read
)
2553 store_info
*store_info
= i_ptr
->store_rec
;
2555 /* Skip the clobbers. */
2556 while (!store_info
->is_set
)
2557 store_info
= store_info
->next
;
2559 if (store_info
->group_id
>= 0
2560 && rtx_group_vec
[store_info
->group_id
]->frame_related
)
2561 remove_store
= true;
2566 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2567 dump_insn_info ("removing from active", i_ptr
);
2569 active_local_stores_len
--;
2571 last
->next_local_store
= i_ptr
->next_local_store
;
2573 active_local_stores
= i_ptr
->next_local_store
;
2578 i_ptr
= i_ptr
->next_local_store
;
2584 if (get_call_args (insn
, memset_call
, args
, 3)
2585 && CONST_INT_P (args
[1])
2586 && CONST_INT_P (args
[2])
2587 && INTVAL (args
[2]) > 0)
2589 rtx mem
= gen_rtx_MEM (BLKmode
, args
[0]);
2590 set_mem_size (mem
, INTVAL (args
[2]));
2591 body
= gen_rtx_SET (mem
, args
[1]);
2592 mems_found
+= record_store (body
, bb_info
);
2593 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2594 fprintf (dump_file
, "handling memset as BLKmode store\n");
2595 if (mems_found
== 1)
2597 if (active_local_stores_len
++ >= max_active_local_stores
)
2599 active_local_stores_len
= 1;
2600 active_local_stores
= NULL
;
2602 insn_info
->fixed_regs_live
2603 = copy_fixed_regs (bb_info
->regs_live
);
2604 insn_info
->next_local_store
= active_local_stores
;
2605 active_local_stores
= insn_info
;
2609 clear_rhs_from_active_local_stores ();
2612 else if (SIBLING_CALL_P (insn
)
2613 && (reload_completed
|| HARD_FRAME_POINTER_IS_ARG_POINTER
))
2614 /* Arguments for a sibling call that are pushed to memory are passed
2615 using the incoming argument pointer of the current function. After
2616 reload that might be (and likely is) frame pointer based. And, if
2617 it is a frame pointer on the target, even before reload we need to
2618 kill frame pointer based stores. */
2619 add_wild_read (bb_info
);
2621 /* Every other call, including pure functions, may read any memory
2622 that is not relative to the frame. */
2623 add_non_frame_wild_read (bb_info
);
2628 /* Assuming that there are sets in these insns, we cannot delete
2630 if ((GET_CODE (PATTERN (insn
)) == CLOBBER
)
2631 || volatile_refs_p (PATTERN (insn
))
2632 || (!cfun
->can_delete_dead_exceptions
&& !insn_nothrow_p (insn
))
2633 || (RTX_FRAME_RELATED_P (insn
))
2634 || find_reg_note (insn
, REG_FRAME_RELATED_EXPR
, NULL_RTX
))
2635 insn_info
->cannot_delete
= true;
2637 body
= PATTERN (insn
);
2638 if (GET_CODE (body
) == PARALLEL
)
2641 for (i
= 0; i
< XVECLEN (body
, 0); i
++)
2642 mems_found
+= record_store (XVECEXP (body
, 0, i
), bb_info
);
2645 mems_found
+= record_store (body
, bb_info
);
2647 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2648 fprintf (dump_file
, "mems_found = %d, cannot_delete = %s\n",
2649 mems_found
, insn_info
->cannot_delete
? "true" : "false");
2651 /* If we found some sets of mems, add it into the active_local_stores so
2652 that it can be locally deleted if found dead or used for
2653 replace_read and redundant constant store elimination. Otherwise mark
2654 it as cannot delete. This simplifies the processing later. */
2655 if (mems_found
== 1)
2657 if (active_local_stores_len
++ >= max_active_local_stores
)
2659 active_local_stores_len
= 1;
2660 active_local_stores
= NULL
;
2662 insn_info
->fixed_regs_live
= copy_fixed_regs (bb_info
->regs_live
);
2663 insn_info
->next_local_store
= active_local_stores
;
2664 active_local_stores
= insn_info
;
2667 insn_info
->cannot_delete
= true;
2671 /* Remove BASE from the set of active_local_stores. This is a
2672 callback from cselib that is used to get rid of the stores in
2673 active_local_stores. */
2676 remove_useless_values (cselib_val
*base
)
2678 insn_info_t insn_info
= active_local_stores
;
2679 insn_info_t last
= NULL
;
2683 store_info
*store_info
= insn_info
->store_rec
;
2686 /* If ANY of the store_infos match the cselib group that is
2687 being deleted, then the insn cannot be deleted. */
2690 if ((store_info
->group_id
== -1)
2691 && (store_info
->cse_base
== base
))
2696 store_info
= store_info
->next
;
2701 active_local_stores_len
--;
2703 last
->next_local_store
= insn_info
->next_local_store
;
2705 active_local_stores
= insn_info
->next_local_store
;
2706 free_store_info (insn_info
);
2711 insn_info
= insn_info
->next_local_store
;
2716 /* Do all of step 1. */
2722 bitmap regs_live
= BITMAP_ALLOC (®_obstack
);
2725 all_blocks
= BITMAP_ALLOC (NULL
);
2726 bitmap_set_bit (all_blocks
, ENTRY_BLOCK
);
2727 bitmap_set_bit (all_blocks
, EXIT_BLOCK
);
2729 /* For -O1 reduce the maximum number of active local stores for RTL DSE
2730 since this can consume huge amounts of memory (PR89115). */
2731 int max_active_local_stores
= param_max_dse_active_local_stores
;
2733 max_active_local_stores
/= 10;
2735 FOR_ALL_BB_FN (bb
, cfun
)
2738 bb_info_t bb_info
= dse_bb_info_type_pool
.allocate ();
2740 memset (bb_info
, 0, sizeof (dse_bb_info_type
));
2741 bitmap_set_bit (all_blocks
, bb
->index
);
2742 bb_info
->regs_live
= regs_live
;
2744 bitmap_copy (regs_live
, DF_LR_IN (bb
));
2745 df_simulate_initialize_forwards (bb
, regs_live
);
2747 bb_table
[bb
->index
] = bb_info
;
2748 cselib_discard_hook
= remove_useless_values
;
2750 if (bb
->index
>= NUM_FIXED_BLOCKS
)
2754 active_local_stores
= NULL
;
2755 active_local_stores_len
= 0;
2756 cselib_clear_table ();
2758 /* Scan the insns. */
2759 FOR_BB_INSNS (bb
, insn
)
2762 scan_insn (bb_info
, insn
, max_active_local_stores
);
2763 cselib_process_insn (insn
);
2765 df_simulate_one_insn_forwards (bb
, insn
, regs_live
);
2768 /* This is something of a hack, because the global algorithm
2769 is supposed to take care of the case where stores go dead
2770 at the end of the function. However, the global
2771 algorithm must take a more conservative view of block
2772 mode reads than the local alg does. So to get the case
2773 where you have a store to the frame followed by a non
2774 overlapping block more read, we look at the active local
2775 stores at the end of the function and delete all of the
2776 frame and spill based ones. */
2777 if (stores_off_frame_dead_at_return
2778 && (EDGE_COUNT (bb
->succs
) == 0
2779 || (single_succ_p (bb
)
2780 && single_succ (bb
) == EXIT_BLOCK_PTR_FOR_FN (cfun
)
2781 && ! crtl
->calls_eh_return
)))
2783 insn_info_t i_ptr
= active_local_stores
;
2786 store_info
*store_info
= i_ptr
->store_rec
;
2788 /* Skip the clobbers. */
2789 while (!store_info
->is_set
)
2790 store_info
= store_info
->next
;
2791 if (store_info
->group_id
>= 0)
2793 group_info
*group
= rtx_group_vec
[store_info
->group_id
];
2794 if (group
->frame_related
&& !i_ptr
->cannot_delete
)
2795 delete_dead_store_insn (i_ptr
);
2798 i_ptr
= i_ptr
->next_local_store
;
2802 /* Get rid of the loads that were discovered in
2803 replace_read. Cselib is finished with this block. */
2804 while (deferred_change_list
)
2806 deferred_change
*next
= deferred_change_list
->next
;
2808 /* There is no reason to validate this change. That was
2810 *deferred_change_list
->loc
= deferred_change_list
->reg
;
2811 deferred_change_pool
.remove (deferred_change_list
);
2812 deferred_change_list
= next
;
2815 /* Get rid of all of the cselib based store_infos in this
2816 block and mark the containing insns as not being
2818 ptr
= bb_info
->last_insn
;
2821 if (ptr
->contains_cselib_groups
)
2823 store_info
*s_info
= ptr
->store_rec
;
2824 while (s_info
&& !s_info
->is_set
)
2825 s_info
= s_info
->next
;
2827 && s_info
->redundant_reason
2828 && s_info
->redundant_reason
->insn
2829 && !ptr
->cannot_delete
)
2831 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2832 fprintf (dump_file
, "Locally deleting insn %d "
2833 "because insn %d stores the "
2834 "same value and couldn't be "
2836 INSN_UID (ptr
->insn
),
2837 INSN_UID (s_info
->redundant_reason
->insn
));
2838 delete_dead_store_insn (ptr
);
2840 free_store_info (ptr
);
2846 /* Free at least positions_needed bitmaps. */
2847 for (s_info
= ptr
->store_rec
; s_info
; s_info
= s_info
->next
)
2848 if (s_info
->is_large
)
2850 BITMAP_FREE (s_info
->positions_needed
.large
.bmap
);
2851 s_info
->is_large
= false;
2854 ptr
= ptr
->prev_insn
;
2857 cse_store_info_pool
.release ();
2859 bb_info
->regs_live
= NULL
;
2862 BITMAP_FREE (regs_live
);
2864 rtx_group_table
->empty ();
2868 /*----------------------------------------------------------------------------
2871 Assign each byte position in the stores that we are going to
2872 analyze globally to a position in the bitmaps. Returns true if
2873 there are any bit positions assigned.
2874 ----------------------------------------------------------------------------*/
2877 dse_step2_init (void)
2882 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2884 /* For all non stack related bases, we only consider a store to
2885 be deletable if there are two or more stores for that
2886 position. This is because it takes one store to make the
2887 other store redundant. However, for the stores that are
2888 stack related, we consider them if there is only one store
2889 for the position. We do this because the stack related
2890 stores can be deleted if their is no read between them and
2891 the end of the function.
2893 To make this work in the current framework, we take the stack
2894 related bases add all of the bits from store1 into store2.
2895 This has the effect of making the eligible even if there is
2898 if (stores_off_frame_dead_at_return
&& group
->frame_related
)
2900 bitmap_ior_into (group
->store2_n
, group
->store1_n
);
2901 bitmap_ior_into (group
->store2_p
, group
->store1_p
);
2902 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2903 fprintf (dump_file
, "group %d is frame related ", i
);
2906 group
->offset_map_size_n
++;
2907 group
->offset_map_n
= XOBNEWVEC (&dse_obstack
, int,
2908 group
->offset_map_size_n
);
2909 group
->offset_map_size_p
++;
2910 group
->offset_map_p
= XOBNEWVEC (&dse_obstack
, int,
2911 group
->offset_map_size_p
);
2912 group
->process_globally
= false;
2913 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2915 fprintf (dump_file
, "group %d(%d+%d): ", i
,
2916 (int)bitmap_count_bits (group
->store2_n
),
2917 (int)bitmap_count_bits (group
->store2_p
));
2918 bitmap_print (dump_file
, group
->store2_n
, "n ", " ");
2919 bitmap_print (dump_file
, group
->store2_p
, "p ", "\n");
2925 /* Init the offset tables. */
2932 /* Position 0 is unused because 0 is used in the maps to mean
2934 current_position
= 1;
2935 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
2940 memset (group
->offset_map_n
, 0, sizeof (int) * group
->offset_map_size_n
);
2941 memset (group
->offset_map_p
, 0, sizeof (int) * group
->offset_map_size_p
);
2942 bitmap_clear (group
->group_kill
);
2944 EXECUTE_IF_SET_IN_BITMAP (group
->store2_n
, 0, j
, bi
)
2946 bitmap_set_bit (group
->group_kill
, current_position
);
2947 if (bitmap_bit_p (group
->escaped_n
, j
))
2948 bitmap_set_bit (kill_on_calls
, current_position
);
2949 group
->offset_map_n
[j
] = current_position
++;
2950 group
->process_globally
= true;
2952 EXECUTE_IF_SET_IN_BITMAP (group
->store2_p
, 0, j
, bi
)
2954 bitmap_set_bit (group
->group_kill
, current_position
);
2955 if (bitmap_bit_p (group
->escaped_p
, j
))
2956 bitmap_set_bit (kill_on_calls
, current_position
);
2957 group
->offset_map_p
[j
] = current_position
++;
2958 group
->process_globally
= true;
2961 return current_position
!= 1;
2966 /*----------------------------------------------------------------------------
2969 Build the bit vectors for the transfer functions.
2970 ----------------------------------------------------------------------------*/
2973 /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not
2977 get_bitmap_index (group_info
*group_info
, HOST_WIDE_INT offset
)
2981 HOST_WIDE_INT offset_p
= -offset
;
2982 if (offset_p
>= group_info
->offset_map_size_n
)
2984 return group_info
->offset_map_n
[offset_p
];
2988 if (offset
>= group_info
->offset_map_size_p
)
2990 return group_info
->offset_map_p
[offset
];
2995 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL
2999 scan_stores (store_info
*store_info
, bitmap gen
, bitmap kill
)
3003 HOST_WIDE_INT i
, offset
, width
;
3004 group_info
*group_info
3005 = rtx_group_vec
[store_info
->group_id
];
3006 /* We can (conservatively) ignore stores whose bounds aren't known;
3007 they simply don't generate new global dse opportunities. */
3008 if (group_info
->process_globally
3009 && store_info
->offset
.is_constant (&offset
)
3010 && store_info
->width
.is_constant (&width
))
3012 HOST_WIDE_INT end
= offset
+ width
;
3013 for (i
= offset
; i
< end
; i
++)
3015 int index
= get_bitmap_index (group_info
, i
);
3018 bitmap_set_bit (gen
, index
);
3020 bitmap_clear_bit (kill
, index
);
3024 store_info
= store_info
->next
;
3029 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL
3033 scan_reads (insn_info_t insn_info
, bitmap gen
, bitmap kill
)
3035 read_info_t read_info
= insn_info
->read_rec
;
3039 /* If this insn reads the frame, kill all the frame related stores. */
3040 if (insn_info
->frame_read
)
3042 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3043 if (group
->process_globally
&& group
->frame_related
)
3046 bitmap_ior_into (kill
, group
->group_kill
);
3047 bitmap_and_compl_into (gen
, group
->group_kill
);
3050 if (insn_info
->non_frame_wild_read
)
3052 /* Kill all non-frame related stores. Kill all stores of variables that
3055 bitmap_ior_into (kill
, kill_on_calls
);
3056 bitmap_and_compl_into (gen
, kill_on_calls
);
3057 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3058 if (group
->process_globally
&& !group
->frame_related
)
3061 bitmap_ior_into (kill
, group
->group_kill
);
3062 bitmap_and_compl_into (gen
, group
->group_kill
);
3067 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3069 if (group
->process_globally
)
3071 if (i
== read_info
->group_id
)
3073 HOST_WIDE_INT offset
, width
;
3074 /* Reads with non-constant size kill all DSE opportunities
3076 if (!read_info
->offset
.is_constant (&offset
)
3077 || !read_info
->width
.is_constant (&width
)
3078 || !known_size_p (width
))
3080 /* Handle block mode reads. */
3082 bitmap_ior_into (kill
, group
->group_kill
);
3083 bitmap_and_compl_into (gen
, group
->group_kill
);
3087 /* The groups are the same, just process the
3090 HOST_WIDE_INT end
= offset
+ width
;
3091 for (j
= offset
; j
< end
; j
++)
3093 int index
= get_bitmap_index (group
, j
);
3097 bitmap_set_bit (kill
, index
);
3098 bitmap_clear_bit (gen
, index
);
3105 /* The groups are different, if the alias sets
3106 conflict, clear the entire group. We only need
3107 to apply this test if the read_info is a cselib
3108 read. Anything with a constant base cannot alias
3109 something else with a different constant
3111 if ((read_info
->group_id
< 0)
3112 && canon_true_dependence (group
->base_mem
,
3113 GET_MODE (group
->base_mem
),
3114 group
->canon_base_addr
,
3115 read_info
->mem
, NULL_RTX
))
3118 bitmap_ior_into (kill
, group
->group_kill
);
3119 bitmap_and_compl_into (gen
, group
->group_kill
);
3125 read_info
= read_info
->next
;
3130 /* Return the insn in BB_INFO before the first wild read or if there
3131 are no wild reads in the block, return the last insn. */
3134 find_insn_before_first_wild_read (bb_info_t bb_info
)
3136 insn_info_t insn_info
= bb_info
->last_insn
;
3137 insn_info_t last_wild_read
= NULL
;
3141 if (insn_info
->wild_read
)
3143 last_wild_read
= insn_info
->prev_insn
;
3144 /* Block starts with wild read. */
3145 if (!last_wild_read
)
3149 insn_info
= insn_info
->prev_insn
;
3153 return last_wild_read
;
3155 return bb_info
->last_insn
;
3159 /* Scan the insns in BB_INFO starting at PTR and going to the top of
3160 the block in order to build the gen and kill sets for the block.
3161 We start at ptr which may be the last insn in the block or may be
3162 the first insn with a wild read. In the latter case we are able to
3163 skip the rest of the block because it just does not matter:
3164 anything that happens is hidden by the wild read. */
3167 dse_step3_scan (basic_block bb
)
3169 bb_info_t bb_info
= bb_table
[bb
->index
];
3170 insn_info_t insn_info
;
3172 insn_info
= find_insn_before_first_wild_read (bb_info
);
3174 /* In the spill case or in the no_spill case if there is no wild
3175 read in the block, we will need a kill set. */
3176 if (insn_info
== bb_info
->last_insn
)
3179 bitmap_clear (bb_info
->kill
);
3181 bb_info
->kill
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3185 BITMAP_FREE (bb_info
->kill
);
3189 /* There may have been code deleted by the dce pass run before
3191 if (insn_info
->insn
&& INSN_P (insn_info
->insn
))
3193 scan_stores (insn_info
->store_rec
, bb_info
->gen
, bb_info
->kill
);
3194 scan_reads (insn_info
, bb_info
->gen
, bb_info
->kill
);
3197 insn_info
= insn_info
->prev_insn
;
3202 /* Set the gen set of the exit block, and also any block with no
3203 successors that does not have a wild read. */
3206 dse_step3_exit_block_scan (bb_info_t bb_info
)
3208 /* The gen set is all 0's for the exit block except for the
3209 frame_pointer_group. */
3211 if (stores_off_frame_dead_at_return
)
3216 FOR_EACH_VEC_ELT (rtx_group_vec
, i
, group
)
3218 if (group
->process_globally
&& group
->frame_related
)
3219 bitmap_ior_into (bb_info
->gen
, group
->group_kill
);
3225 /* Find all of the blocks that are not backwards reachable from the
3226 exit block or any block with no successors (BB). These are the
3227 infinite loops or infinite self loops. These blocks will still
3228 have their bits set in UNREACHABLE_BLOCKS. */
3231 mark_reachable_blocks (sbitmap unreachable_blocks
, basic_block bb
)
3236 if (bitmap_bit_p (unreachable_blocks
, bb
->index
))
3238 bitmap_clear_bit (unreachable_blocks
, bb
->index
);
3239 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3241 mark_reachable_blocks (unreachable_blocks
, e
->src
);
3246 /* Build the transfer functions for the function. */
3252 sbitmap_iterator sbi
;
3253 bitmap all_ones
= NULL
;
3256 auto_sbitmap
unreachable_blocks (last_basic_block_for_fn (cfun
));
3257 bitmap_ones (unreachable_blocks
);
3259 FOR_ALL_BB_FN (bb
, cfun
)
3261 bb_info_t bb_info
= bb_table
[bb
->index
];
3263 bitmap_clear (bb_info
->gen
);
3265 bb_info
->gen
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3267 if (bb
->index
== ENTRY_BLOCK
)
3269 else if (bb
->index
== EXIT_BLOCK
)
3270 dse_step3_exit_block_scan (bb_info
);
3272 dse_step3_scan (bb
);
3273 if (EDGE_COUNT (bb
->succs
) == 0)
3274 mark_reachable_blocks (unreachable_blocks
, bb
);
3276 /* If this is the second time dataflow is run, delete the old
3279 BITMAP_FREE (bb_info
->in
);
3281 BITMAP_FREE (bb_info
->out
);
3284 /* For any block in an infinite loop, we must initialize the out set
3285 to all ones. This could be expensive, but almost never occurs in
3286 practice. However, it is common in regression tests. */
3287 EXECUTE_IF_SET_IN_BITMAP (unreachable_blocks
, 0, i
, sbi
)
3289 if (bitmap_bit_p (all_blocks
, i
))
3291 bb_info_t bb_info
= bb_table
[i
];
3297 all_ones
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3298 FOR_EACH_VEC_ELT (rtx_group_vec
, j
, group
)
3299 bitmap_ior_into (all_ones
, group
->group_kill
);
3303 bb_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3304 bitmap_copy (bb_info
->out
, all_ones
);
3310 BITMAP_FREE (all_ones
);
3315 /*----------------------------------------------------------------------------
3318 Solve the bitvector equations.
3319 ----------------------------------------------------------------------------*/
3322 /* Confluence function for blocks with no successors. Create an out
3323 set from the gen set of the exit block. This block logically has
3324 the exit block as a successor. */
3329 dse_confluence_0 (basic_block bb
)
3331 bb_info_t bb_info
= bb_table
[bb
->index
];
3333 if (bb
->index
== EXIT_BLOCK
)
3338 bb_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3339 bitmap_copy (bb_info
->out
, bb_table
[EXIT_BLOCK
]->gen
);
3343 /* Propagate the information from the in set of the dest of E to the
3344 out set of the src of E. If the various in or out sets are not
3345 there, that means they are all ones. */
3348 dse_confluence_n (edge e
)
3350 bb_info_t src_info
= bb_table
[e
->src
->index
];
3351 bb_info_t dest_info
= bb_table
[e
->dest
->index
];
3356 bitmap_and_into (src_info
->out
, dest_info
->in
);
3359 src_info
->out
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3360 bitmap_copy (src_info
->out
, dest_info
->in
);
3367 /* Propagate the info from the out to the in set of BB_INDEX's basic
3368 block. There are three cases:
3370 1) The block has no kill set. In this case the kill set is all
3371 ones. It does not matter what the out set of the block is, none of
3372 the info can reach the top. The only thing that reaches the top is
3373 the gen set and we just copy the set.
3375 2) There is a kill set but no out set and bb has successors. In
3376 this case we just return. Eventually an out set will be created and
3377 it is better to wait than to create a set of ones.
3379 3) There is both a kill and out set. We apply the obvious transfer
3384 dse_transfer_function (int bb_index
)
3386 bb_info_t bb_info
= bb_table
[bb_index
];
3394 return bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
3395 bb_info
->out
, bb_info
->kill
);
3398 bb_info
->in
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3399 bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
3400 bb_info
->out
, bb_info
->kill
);
3410 /* Case 1 above. If there is already an in set, nothing
3416 bb_info
->in
= BITMAP_ALLOC (&dse_bitmap_obstack
);
3417 bitmap_copy (bb_info
->in
, bb_info
->gen
);
3423 /* Solve the dataflow equations. */
3428 df_simple_dataflow (DF_BACKWARD
, NULL
, dse_confluence_0
,
3429 dse_confluence_n
, dse_transfer_function
,
3430 all_blocks
, df_get_postorder (DF_BACKWARD
),
3431 df_get_n_blocks (DF_BACKWARD
));
3432 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3436 fprintf (dump_file
, "\n\n*** Global dataflow info after analysis.\n");
3437 FOR_ALL_BB_FN (bb
, cfun
)
3439 bb_info_t bb_info
= bb_table
[bb
->index
];
3441 df_print_bb_index (bb
, dump_file
);
3443 bitmap_print (dump_file
, bb_info
->in
, " in: ", "\n");
3445 fprintf (dump_file
, " in: *MISSING*\n");
3447 bitmap_print (dump_file
, bb_info
->gen
, " gen: ", "\n");
3449 fprintf (dump_file
, " gen: *MISSING*\n");
3451 bitmap_print (dump_file
, bb_info
->kill
, " kill: ", "\n");
3453 fprintf (dump_file
, " kill: *MISSING*\n");
3455 bitmap_print (dump_file
, bb_info
->out
, " out: ", "\n");
3457 fprintf (dump_file
, " out: *MISSING*\n\n");
3464 /*----------------------------------------------------------------------------
3467 Delete the stores that can only be deleted using the global information.
3468 ----------------------------------------------------------------------------*/
3475 FOR_EACH_BB_FN (bb
, cfun
)
3477 bb_info_t bb_info
= bb_table
[bb
->index
];
3478 insn_info_t insn_info
= bb_info
->last_insn
;
3479 bitmap v
= bb_info
->out
;
3483 bool deleted
= false;
3484 if (dump_file
&& insn_info
->insn
)
3486 fprintf (dump_file
, "starting to process insn %d\n",
3487 INSN_UID (insn_info
->insn
));
3488 bitmap_print (dump_file
, v
, " v: ", "\n");
3491 /* There may have been code deleted by the dce pass run before
3494 && INSN_P (insn_info
->insn
)
3495 && (!insn_info
->cannot_delete
)
3496 && (!bitmap_empty_p (v
)))
3498 store_info
*store_info
= insn_info
->store_rec
;
3500 /* Try to delete the current insn. */
3503 /* Skip the clobbers. */
3504 while (!store_info
->is_set
)
3505 store_info
= store_info
->next
;
3507 HOST_WIDE_INT i
, offset
, width
;
3508 group_info
*group_info
= rtx_group_vec
[store_info
->group_id
];
3510 if (!store_info
->offset
.is_constant (&offset
)
3511 || !store_info
->width
.is_constant (&width
))
3515 HOST_WIDE_INT end
= offset
+ width
;
3516 for (i
= offset
; i
< end
; i
++)
3518 int index
= get_bitmap_index (group_info
, i
);
3520 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3521 fprintf (dump_file
, "i = %d, index = %d\n",
3523 if (index
== 0 || !bitmap_bit_p (v
, index
))
3525 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3526 fprintf (dump_file
, "failing at i = %d\n",
3536 && check_for_inc_dec_1 (insn_info
))
3538 delete_insn (insn_info
->insn
);
3539 insn_info
->insn
= NULL
;
3544 /* We do want to process the local info if the insn was
3545 deleted. For instance, if the insn did a wild read, we
3546 no longer need to trash the info. */
3548 && INSN_P (insn_info
->insn
)
3551 scan_stores (insn_info
->store_rec
, v
, NULL
);
3552 if (insn_info
->wild_read
)
3554 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3555 fprintf (dump_file
, "wild read\n");
3558 else if (insn_info
->read_rec
3559 || insn_info
->non_frame_wild_read
3560 || insn_info
->frame_read
)
3562 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3564 if (!insn_info
->non_frame_wild_read
3565 && !insn_info
->frame_read
)
3566 fprintf (dump_file
, "regular read\n");
3567 if (insn_info
->non_frame_wild_read
)
3568 fprintf (dump_file
, "non-frame wild read\n");
3569 if (insn_info
->frame_read
)
3570 fprintf (dump_file
, "frame read\n");
3572 scan_reads (insn_info
, v
, NULL
);
3576 insn_info
= insn_info
->prev_insn
;
3583 /*----------------------------------------------------------------------------
3586 Delete stores made redundant by earlier stores (which store the same
3587 value) that couldn't be eliminated.
3588 ----------------------------------------------------------------------------*/
3595 FOR_ALL_BB_FN (bb
, cfun
)
3597 bb_info_t bb_info
= bb_table
[bb
->index
];
3598 insn_info_t insn_info
= bb_info
->last_insn
;
3602 /* There may have been code deleted by the dce pass run before
3605 && INSN_P (insn_info
->insn
)
3606 && !insn_info
->cannot_delete
)
3608 store_info
*s_info
= insn_info
->store_rec
;
3610 while (s_info
&& !s_info
->is_set
)
3611 s_info
= s_info
->next
;
3613 && s_info
->redundant_reason
3614 && s_info
->redundant_reason
->insn
3615 && INSN_P (s_info
->redundant_reason
->insn
))
3617 rtx_insn
*rinsn
= s_info
->redundant_reason
->insn
;
3618 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3619 fprintf (dump_file
, "Locally deleting insn %d "
3620 "because insn %d stores the "
3621 "same value and couldn't be "
3623 INSN_UID (insn_info
->insn
),
3625 delete_dead_store_insn (insn_info
);
3628 insn_info
= insn_info
->prev_insn
;
3633 /*----------------------------------------------------------------------------
3636 Destroy everything left standing.
3637 ----------------------------------------------------------------------------*/
3642 bitmap_obstack_release (&dse_bitmap_obstack
);
3643 obstack_free (&dse_obstack
, NULL
);
3645 end_alias_analysis ();
3647 delete rtx_group_table
;
3648 rtx_group_table
= NULL
;
3649 rtx_group_vec
.release ();
3650 BITMAP_FREE (all_blocks
);
3651 BITMAP_FREE (scratch
);
3653 rtx_store_info_pool
.release ();
3654 read_info_type_pool
.release ();
3655 insn_info_type_pool
.release ();
3656 dse_bb_info_type_pool
.release ();
3657 group_info_pool
.release ();
3658 deferred_change_pool
.release ();
3662 /* -------------------------------------------------------------------------
3664 ------------------------------------------------------------------------- */
3666 /* Callback for running pass_rtl_dse. */
3669 rest_of_handle_dse (void)
3671 df_set_flags (DF_DEFER_INSN_RESCAN
);
3673 /* Need the notes since we must track live hardregs in the forwards
3675 df_note_add_problem ();
3683 df_set_flags (DF_LR_RUN_DCE
);
3685 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3686 fprintf (dump_file
, "doing global processing\n");
3696 fprintf (dump_file
, "dse: local deletions = %d, global deletions = %d\n",
3697 locally_deleted
, globally_deleted
);
3699 /* DSE can eliminate potentially-trapping MEMs.
3700 Remove any EH edges associated with them. */
3701 if ((locally_deleted
|| globally_deleted
)
3702 && cfun
->can_throw_non_call_exceptions
3703 && purge_all_dead_edges ())
3705 free_dominance_info (CDI_DOMINATORS
);
3714 const pass_data pass_data_rtl_dse1
=
3716 RTL_PASS
, /* type */
3718 OPTGROUP_NONE
, /* optinfo_flags */
3719 TV_DSE1
, /* tv_id */
3720 0, /* properties_required */
3721 0, /* properties_provided */
3722 0, /* properties_destroyed */
3723 0, /* todo_flags_start */
3724 TODO_df_finish
, /* todo_flags_finish */
3727 class pass_rtl_dse1
: public rtl_opt_pass
3730 pass_rtl_dse1 (gcc::context
*ctxt
)
3731 : rtl_opt_pass (pass_data_rtl_dse1
, ctxt
)
3734 /* opt_pass methods: */
3735 virtual bool gate (function
*)
3737 return optimize
> 0 && flag_dse
&& dbg_cnt (dse1
);
3740 virtual unsigned int execute (function
*) { return rest_of_handle_dse (); }
3742 }; // class pass_rtl_dse1
3747 make_pass_rtl_dse1 (gcc::context
*ctxt
)
3749 return new pass_rtl_dse1 (ctxt
);
3754 const pass_data pass_data_rtl_dse2
=
3756 RTL_PASS
, /* type */
3758 OPTGROUP_NONE
, /* optinfo_flags */
3759 TV_DSE2
, /* tv_id */
3760 0, /* properties_required */
3761 0, /* properties_provided */
3762 0, /* properties_destroyed */
3763 0, /* todo_flags_start */
3764 TODO_df_finish
, /* todo_flags_finish */
3767 class pass_rtl_dse2
: public rtl_opt_pass
3770 pass_rtl_dse2 (gcc::context
*ctxt
)
3771 : rtl_opt_pass (pass_data_rtl_dse2
, ctxt
)
3774 /* opt_pass methods: */
3775 virtual bool gate (function
*)
3777 return optimize
> 0 && flag_dse
&& dbg_cnt (dse2
);
3780 virtual unsigned int execute (function
*) { return rest_of_handle_dse (); }
3782 }; // class pass_rtl_dse2
3787 make_pass_rtl_dse2 (gcc::context
*ctxt
)
3789 return new pass_rtl_dse2 (ctxt
);