1 /* RTL dead store elimination.
2 Copyright (C) 2005, 2006, 2007 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"
33 #include "hard-reg-set.h"
38 #include "tree-pass.h"
39 #include "alloc-pool.h"
41 #include "insn-config.h"
48 /* This file contains three techniques for performing Dead Store
51 * The first technique performs dse locally on any base address. It
52 is based on the cselib which is a local value numbering technique.
53 This technique is local to a basic block but deals with a fairly
56 * The second technique performs dse globally but is restricted to
57 base addresses that are either constant or are relative to the
60 * The third technique, (which is only done after register allocation)
61 processes the spill spill slots. This differs from the second
62 technique because it takes advantage of the fact that spilling is
63 completely free from the effects of aliasing.
65 Logically, dse is a backwards dataflow problem. A store can be
66 deleted if it if cannot be reached in the backward direction by any
67 use of the value being stored. However, the local technique uses a
68 forwards scan of the basic block because cselib requires that the
69 block be processed in that order.
71 The pass is logically broken into 7 steps:
75 1) The local algorithm, as well as scanning the insns for the two
78 2) Analysis to see if the global algs are necessary. In the case
79 of stores base on a constant address, there must be at least two
80 stores to that address, to make it possible to delete some of the
81 stores. In the case of stores off of the frame or spill related
82 stores, only one store to an address is necessary because those
83 stores die at the end of the function.
85 3) Set up the global dataflow equations based on processing the
86 info parsed in the first step.
88 4) Solve the dataflow equations.
90 5) Delete the insns that the global analysis has indicated are
95 This step uses cselib and canon_rtx to build the largest expression
96 possible for each address. This pass is a forwards pass through
97 each basic block. From the point of view of the global technique,
98 the first pass could examine a block in either direction. The
99 forwards ordering is to accommodate cselib.
101 We a simplifying assumption: addresses fall into four broad
104 1) base has rtx_varies_p == false, offset is constant.
105 2) base has rtx_varies_p == false, offset variable.
106 3) base has rtx_varies_p == true, offset constant.
107 4) base has rtx_varies_p == true, offset variable.
109 The local passes are able to process all 4 kinds of addresses. The
110 global pass only handles (1).
112 The global problem is formulated as follows:
114 A store, S1, to address A, where A is not relative to the stack
115 frame, can be eliminated if all paths from S1 to the end of the
116 of the function contain another store to A before a read to A.
118 If the address A is relative to the stack frame, a store S2 to A
119 can be eliminated if there are no paths from S1 that reach the
120 end of the function that read A before another store to A. In
121 this case S2 can be deleted if there are paths to from S2 to the
122 end of the function that have no reads or writes to A. This
123 second case allows stores to the stack frame to be deleted that
124 would otherwise die when the function returns. This cannot be
125 done if stores_off_frame_dead_at_return is not true. See the doc
126 for that variable for when this variable is false.
128 The global problem is formulated as a backwards set union
129 dataflow problem where the stores are the gens and reads are the
130 kills. Set union problems are rare and require some special
131 handling given our representation of bitmaps. A straightforward
132 implementation of requires a lot of bitmaps filled with 1s.
133 These are expensive and cumbersome in our bitmap formulation so
134 care has been taken to avoid large vectors filled with 1s. See
135 the comments in bb_info and in the dataflow confluence functions
138 There are two places for further enhancements to this algorithm:
140 1) The original dse which was embedded in a pass called flow also
141 did local address forwarding. For example in
146 flow would replace the right hand side of the second insn with a
147 reference to r100. Most of the information is available to add this
148 to this pass. It has not done it because it is a lot of work in
149 the case that either r100 is assigned to between the first and
150 second insn and/or the second insn is a load of part of the value
151 stored by the first insn.
153 insn 5 in gcc.c-torture/compile/990203-1.c simple case.
154 insn 15 in gcc.c-torture/execute/20001017-2.c simple case.
155 insn 25 in gcc.c-torture/execute/20001026-1.c simple case.
156 insn 44 in gcc.c-torture/execute/20010910-1.c simple case.
158 2) The cleaning up of spill code is quite profitable. It currently
159 depends on reading tea leaves and chicken entrails left by reload.
160 This pass depends on reload creating a singleton alias set for each
161 spill slot and telling the next dse pass which of these alias sets
162 are the singletons. Rather than analyze the addresses of the
163 spills, dse's spill processing just does analysis of the loads and
164 stores that use those alias sets. There are three cases where this
167 a) Reload sometimes creates the slot for one mode of access, and
168 then inserts loads and/or stores for a smaller mode. In this
169 case, the current code just punts on the slot. The proper thing
170 to do is to back out and use one bit vector position for each
171 byte of the entity associated with the slot. This depends on
172 KNOWING that reload always generates the accesses for each of the
173 bytes in some canonical (read that easy to understand several
174 passes after reload happens) way.
176 b) Reload sometimes decides that spill slot it allocated was not
177 large enough for the mode and goes back and allocates more slots
178 with the same mode and alias set. The backout in this case is a
179 little more graceful than (a). In this case the slot is unmarked
180 as being a spill slot and if final address comes out to be based
181 off the frame pointer, the global algorithm handles this slot.
183 c) For any pass that may prespill, there is currently no
184 mechanism to tell the dse pass that the slot being used has the
185 special properties that reload uses. It may be that all that is
186 required is to have those passes make the same calls that reload
187 does, assuming that the alias sets can be manipulated in the same
190 /* There are limits to the size of constant offsets we model for the
191 global problem. There are certainly test cases, that exceed this
192 limit, however, it is unlikely that there are important programs
193 that really have constant offsets this size. */
194 #define MAX_OFFSET (64 * 1024)
197 static bitmap scratch
= NULL
;
200 /* This structure holds information about a candidate store. */
204 /* False means this is a clobber. */
207 /* The id of the mem group of the base address. If rtx_varies_p is
208 true, this is -1. Otherwise, it is the index into the group
212 /* This is the cselib value. */
213 cselib_val
*cse_base
;
215 /* This canonized mem. */
218 /* The result of get_addr on mem. */
221 /* If this is non-zero, it is the alias set of a spill location. */
222 alias_set_type alias_set
;
224 /* The offset of the first and byte before the last byte associated
225 with the operation. */
228 /* An bitmask as wide as the number of bytes in the word that
229 contains a 1 if the byte may be needed. The store is unused if
230 all of the bits are 0. */
231 long positions_needed
;
233 /* The next store info for this insn. */
234 struct store_info
*next
;
236 /* The right hand side of the store. This is used if there is a
237 subsequent reload of the mems address somewhere later in the
242 typedef struct store_info
*store_info_t
;
243 static alloc_pool cse_store_info_pool
;
244 static alloc_pool rtx_store_info_pool
;
246 /* This structure holds information about a load. These are only
247 built for rtx bases. */
250 /* The id of the mem group of the base address. */
253 /* If this is non-zero, it is the alias set of a spill location. */
254 alias_set_type alias_set
;
256 /* The offset of the first and byte after the last byte associated
257 with the operation. If begin == end == 0, the read did not have
258 a constant offset. */
261 /* The mem being read. */
264 /* The next read_info for this insn. */
265 struct read_info
*next
;
267 typedef struct read_info
*read_info_t
;
268 static alloc_pool read_info_pool
;
271 /* One of these records is created for each insn. */
275 /* Set true if the insn contains a store but the insn itself cannot
276 be deleted. This is set if the insn is a parallel and there is
277 more than one non dead output or if the insn is in some way
281 /* This field is only used by the global algorithm. It is set true
282 if the insn contains any read of mem except for a (1). This is
283 also set if the insn is a call or has a clobber mem. If the insn
284 contains a wild read, the use_rec will be null. */
287 /* This field is set for const function calls. Const functions
288 cannot read memory, but they can read the stack because that is
289 where they may get their parms. So having this set is less
290 severe than a wild read, it just means that all of the stores to
291 the stack are killed rather than all stores. */
294 /* This is true if any of the sets within the store contains a
295 cselib base. Such stores can only be deleted by the local
297 bool contains_cselib_groups
;
302 /* The list of mem sets or mem clobbers that are contained in this
303 insn. If the insn is deletable, it contains only one mem set.
304 But it could also contain clobbers. Insns that contain more than
305 one mem set are not deletable, but each of those mems are here in
306 order to provide info to delete other insns. */
307 store_info_t store_rec
;
309 /* The linked list of mem uses in this insn. Only the reads from
310 rtx bases are listed here. The reads to cselib bases are
311 completely processed during the first scan and so are never
313 read_info_t read_rec
;
315 /* The prev insn in the basic block. */
316 struct insn_info
* prev_insn
;
318 /* The linked list of insns that are in consideration for removal in
319 the forwards pass thru the basic block. This pointer may be
320 trash as it is not cleared when a wild read occurs. The only
321 time it is guaranteed to be correct is when the traveral starts
322 at active_local_stores. */
323 struct insn_info
* next_local_store
;
326 typedef struct insn_info
*insn_info_t
;
327 static alloc_pool insn_info_pool
;
329 /* The linked list of stores that are under consideration in this
331 static insn_info_t active_local_stores
;
336 /* Pointer to the insn info for the last insn in the block. These
337 are linked so this is how all of the insns are reached. During
338 scanning this is the current insn being scanned. */
339 insn_info_t last_insn
;
341 /* The info for the global dataflow problem. */
344 /* This is set if the transfer function should and in the wild_read
345 bitmap before applying the kill and gen sets. That vector knocks
346 out most of the bits in the bitmap and thus speeds up the
348 bool apply_wild_read
;
350 /* The set of store positions that exist in this block before a wild read. */
353 /* The set of load positions that exist in this block above the
354 same position of a store. */
357 /* The set of stores that reach the top of the block without being
360 Do not represent the in if it is all ones. Note that this is
361 what the bitvector should logically be initialized to for a set
362 intersection problem. However, like the kill set, this is too
363 expensive. So initially, the in set will only be created for the
364 exit block and any block that contains a wild read. */
367 /* The set of stores that reach the bottom of the block from it's
370 Do not represent the in if it is all ones. Note that this is
371 what the bitvector should logically be initialized to for a set
372 intersection problem. However, like the kill and in set, this is
373 too expensive. So what is done is that the confluence operator
374 just initializes the vector from one of the out sets of the
375 successors of the block. */
379 typedef struct bb_info
*bb_info_t
;
380 static alloc_pool bb_info_pool
;
382 /* Table to hold all bb_infos. */
383 static bb_info_t
*bb_table
;
385 /* There is a group_info for each rtx base that is used to reference
386 memory. There are also not many of the rtx bases because they are
387 very limited in scope. */
391 /* The actual base of the address. */
394 /* The sequential id of the base. This allows us to have a
395 canonical ordering of these that is not based on addresses. */
398 /* A mem wrapped around the base pointer for the group in order to
399 do read dependency. */
402 /* Canonized version of base_mem, most likely the same thing. */
405 /* These two sets of two bitmaps are used to keep track of how many
406 stores are actually referencing that position from this base. We
407 only do this for rtx bases as this will be used to assign
408 positions in the bitmaps for the global problem. Bit N is set in
409 store1 on the first store for offset N. Bit N is set in store2
410 for the second store to offset N. This is all we need since we
411 only care about offsets that have two or more stores for them.
413 The "_n" suffix is for offsets less than 0 and the "_p" suffix is
414 for 0 and greater offsets.
416 There is one special case here, for stores into the stack frame,
417 we will or store1 into store2 before deciding which stores look
418 at globally. This is because stores to the stack frame that have
419 no other reads before the end of the function can also be
421 bitmap store1_n
, store1_p
, store2_n
, store2_p
;
423 /* The positions in this bitmap have the same assignments as the in,
424 out, gen and kill bitmaps. This bitmap is all zeros except for
425 the positions that are occupied by stores for this group. */
428 /* True if there are any positions that are to be processed
430 bool process_globally
;
432 /* True if the base of this group is either the frame_pointer or
433 hard_frame_pointer. */
436 /* The offset_map is used to map the offsets from this base into
437 positions in the global bitmaps. It is only created after all of
438 the all of stores have been scanned and we know which ones we
440 int *offset_map_n
, *offset_map_p
;
441 int offset_map_size_n
, offset_map_size_p
;
443 typedef struct group_info
*group_info_t
;
444 typedef const struct group_info
*const_group_info_t
;
445 static alloc_pool rtx_group_info_pool
;
447 /* Tables of group_info structures, hashed by base value. */
448 static htab_t rtx_group_table
;
450 /* Index into the rtx_group_vec. */
451 static int rtx_group_next_id
;
453 DEF_VEC_P(group_info_t
);
454 DEF_VEC_ALLOC_P(group_info_t
,heap
);
456 static VEC(group_info_t
,heap
) *rtx_group_vec
;
459 /* This structure holds the set of changes that are being deferred
460 when removing read operation. See replace_read. */
461 struct deferred_change
464 /* The mem that is being replaced. */
467 /* The reg it is being replaced with. */
470 struct deferred_change
*next
;
473 typedef struct deferred_change
*deferred_change_t
;
474 static alloc_pool deferred_change_pool
;
476 static deferred_change_t deferred_change_list
= NULL
;
478 /* This are used to hold the alias sets of spill variables. Since
479 these are never aliased and there may be a lot of them, it makes
480 sense to treat them specially. This bitvector is only allocated in
481 calls from dse_record_singleton_alias_set which currently is only
482 made during reload1. So when dse is called before reload this
483 mechanism does nothing. */
485 static bitmap clear_alias_sets
= NULL
;
487 /* The set of clear_alias_sets that have been disqualified because
488 there are loads or stores using a different mode than the alias set
489 was registered with. */
490 static bitmap disqualified_clear_alias_sets
= NULL
;
492 /* The group that holds all of the clear_alias_sets. */
493 static group_info_t clear_alias_group
;
495 /* The modes of the clear_alias_sets. */
496 static htab_t clear_alias_mode_table
;
498 /* Hash table element to look up the mode for an alias set. */
499 struct clear_alias_mode_holder
501 alias_set_type alias_set
;
502 enum machine_mode mode
;
505 static alloc_pool clear_alias_mode_pool
;
507 /* This is true except for two cases:
508 (1) current_function_stdarg -- i.e. we cannot do this
509 for vararg functions because they play games with the frame.
510 (2) In ada, it is sometimes not safe to do assume that any stores
511 based off the stack frame go dead at the exit to a function. */
512 static bool stores_off_frame_dead_at_return
;
514 /* Counter for stats. */
515 static int globally_deleted
;
516 static int locally_deleted
;
517 static int spill_deleted
;
519 static bitmap all_blocks
;
521 /* The number of bits used in the global bitmaps. */
522 static unsigned int current_position
;
525 static bool gate_dse (void);
528 /*----------------------------------------------------------------------------
532 ----------------------------------------------------------------------------*/
534 /* Hashtable callbacks for maintaining the "bases" field of
535 store_group_info, given that the addresses are function invariants. */
538 clear_alias_mode_eq (const void *p1
, const void *p2
)
540 const struct clear_alias_mode_holder
* h1
541 = (const struct clear_alias_mode_holder
*) p1
;
542 const struct clear_alias_mode_holder
* h2
543 = (const struct clear_alias_mode_holder
*) p2
;
544 return h1
->alias_set
== h2
->alias_set
;
549 clear_alias_mode_hash (const void *p
)
551 const struct clear_alias_mode_holder
*holder
552 = (const struct clear_alias_mode_holder
*) p
;
553 return holder
->alias_set
;
557 /* Find the entry associated with ALIAS_SET. */
559 static struct clear_alias_mode_holder
*
560 clear_alias_set_lookup (alias_set_type alias_set
)
562 struct clear_alias_mode_holder tmp_holder
;
565 tmp_holder
.alias_set
= alias_set
;
566 slot
= htab_find_slot (clear_alias_mode_table
, &tmp_holder
, NO_INSERT
);
573 /* Hashtable callbacks for maintaining the "bases" field of
574 store_group_info, given that the addresses are function invariants. */
577 invariant_group_base_eq (const void *p1
, const void *p2
)
579 const_group_info_t gi1
= (const_group_info_t
) p1
;
580 const_group_info_t gi2
= (const_group_info_t
) p2
;
581 return rtx_equal_p (gi1
->rtx_base
, gi2
->rtx_base
);
586 invariant_group_base_hash (const void *p
)
588 const_group_info_t gi
= (const_group_info_t
) p
;
590 return hash_rtx (gi
->rtx_base
, Pmode
, &do_not_record
, NULL
, false);
594 /* Get the GROUP for BASE. Add a new group if it is not there. */
597 get_group_info (rtx base
)
599 struct group_info tmp_gi
;
605 /* Find the store_base_info structure for BASE, creating a new one
607 tmp_gi
.rtx_base
= base
;
608 slot
= htab_find_slot (rtx_group_table
, &tmp_gi
, INSERT
);
609 gi
= (group_info_t
) *slot
;
613 if (!clear_alias_group
)
615 clear_alias_group
= gi
= pool_alloc (rtx_group_info_pool
);
616 memset (gi
, 0, sizeof (struct group_info
));
617 gi
->id
= rtx_group_next_id
++;
618 gi
->store1_n
= BITMAP_ALLOC (NULL
);
619 gi
->store1_p
= BITMAP_ALLOC (NULL
);
620 gi
->store2_n
= BITMAP_ALLOC (NULL
);
621 gi
->store2_p
= BITMAP_ALLOC (NULL
);
622 gi
->group_kill
= BITMAP_ALLOC (NULL
);
623 gi
->process_globally
= false;
624 gi
->offset_map_size_n
= 0;
625 gi
->offset_map_size_p
= 0;
626 gi
->offset_map_n
= NULL
;
627 gi
->offset_map_p
= NULL
;
628 VEC_safe_push (group_info_t
, heap
, rtx_group_vec
, gi
);
630 return clear_alias_group
;
635 *slot
= gi
= pool_alloc (rtx_group_info_pool
);
637 gi
->id
= rtx_group_next_id
++;
638 gi
->base_mem
= gen_rtx_MEM (QImode
, base
);
639 gi
->canon_base_mem
= canon_rtx (gi
->base_mem
);
640 gi
->store1_n
= BITMAP_ALLOC (NULL
);
641 gi
->store1_p
= BITMAP_ALLOC (NULL
);
642 gi
->store2_n
= BITMAP_ALLOC (NULL
);
643 gi
->store2_p
= BITMAP_ALLOC (NULL
);
644 gi
->group_kill
= BITMAP_ALLOC (NULL
);
645 gi
->process_globally
= false;
647 (base
== frame_pointer_rtx
) || (base
== hard_frame_pointer_rtx
);
648 gi
->offset_map_size_n
= 0;
649 gi
->offset_map_size_p
= 0;
650 gi
->offset_map_n
= NULL
;
651 gi
->offset_map_p
= NULL
;
652 VEC_safe_push (group_info_t
, heap
, rtx_group_vec
, gi
);
659 /* Initialization of data structures. */
665 globally_deleted
= 0;
668 scratch
= BITMAP_ALLOC (NULL
);
671 = create_alloc_pool ("rtx_store_info_pool",
672 sizeof (struct store_info
), 100);
674 = create_alloc_pool ("read_info_pool",
675 sizeof (struct read_info
), 100);
677 = create_alloc_pool ("insn_info_pool",
678 sizeof (struct insn_info
), 100);
680 = create_alloc_pool ("bb_info_pool",
681 sizeof (struct bb_info
), 100);
683 = create_alloc_pool ("rtx_group_info_pool",
684 sizeof (struct group_info
), 100);
686 = create_alloc_pool ("deferred_change_pool",
687 sizeof (struct deferred_change
), 10);
689 rtx_group_table
= htab_create (11, invariant_group_base_hash
,
690 invariant_group_base_eq
, NULL
);
692 bb_table
= XCNEWVEC (bb_info_t
, last_basic_block
);
693 rtx_group_next_id
= 0;
695 stores_off_frame_dead_at_return
=
696 (!(TREE_CODE (TREE_TYPE (current_function_decl
)) == FUNCTION_TYPE
697 && (TYPE_RETURNS_STACK_DEPRESSED (TREE_TYPE (current_function_decl
)))))
698 && (!current_function_stdarg
);
700 init_alias_analysis ();
702 if (clear_alias_sets
)
703 clear_alias_group
= get_group_info (NULL
);
705 clear_alias_group
= NULL
;
710 /*----------------------------------------------------------------------------
713 Scan all of the insns. Any random ordering of the blocks is fine.
714 Each block is scanned in forward order to accommodate cselib which
715 is used to remove stores with non-constant bases.
716 ----------------------------------------------------------------------------*/
718 /* Delete all of the store_info recs from INSN_INFO. */
721 free_store_info (insn_info_t insn_info
)
723 store_info_t store_info
= insn_info
->store_rec
;
726 store_info_t next
= store_info
->next
;
727 if (store_info
->cse_base
)
728 pool_free (cse_store_info_pool
, store_info
);
730 pool_free (rtx_store_info_pool
, store_info
);
734 insn_info
->cannot_delete
= true;
735 insn_info
->contains_cselib_groups
= false;
736 insn_info
->store_rec
= NULL
;
746 /* Add an insn to do the add inside a x if it is a
747 PRE/POST-INC/DEC/MODIFY. D is an structure containing the insn and
748 the size of the mode of the MEM that this is inside of. */
751 replace_inc_dec (rtx
*r
, void *d
)
754 struct insn_size
*data
= (struct insn_size
*)d
;
755 switch (GET_CODE (x
))
760 rtx r1
= XEXP (x
, 0);
761 rtx c
= gen_int_mode (Pmode
, data
->size
);
762 add_insn_before (data
->insn
,
763 gen_rtx_SET (Pmode
, r1
,
764 gen_rtx_PLUS (Pmode
, r1
, c
)),
772 rtx r1
= XEXP (x
, 0);
773 rtx c
= gen_int_mode (Pmode
, -data
->size
);
774 add_insn_before (data
->insn
,
775 gen_rtx_SET (Pmode
, r1
,
776 gen_rtx_PLUS (Pmode
, r1
, c
)),
784 /* We can reuse the add because we are about to delete the
785 insn that contained it. */
786 rtx add
= XEXP (x
, 0);
787 rtx r1
= XEXP (add
, 0);
788 add_insn_before (data
->insn
,
789 gen_rtx_SET (Pmode
, r1
, add
), NULL
);
799 /* If X is a MEM, check the address to see if it is PRE/POST-INC/DEC/MODIFY
800 and generate an add to replace that. */
803 replace_inc_dec_mem (rtx
*r
, void *d
)
806 if (GET_CODE (x
) == MEM
)
808 struct insn_size data
;
810 data
.size
= GET_MODE_SIZE (GET_MODE (x
));
813 for_each_rtx (&XEXP (x
, 0), replace_inc_dec
, &data
);
820 /* Before we delete INSN, make sure that the auto inc/dec, if it is
821 there, is split into a separate insn. */
824 check_for_inc_dec (rtx insn
)
826 rtx note
= find_reg_note (insn
, REG_INC
, NULL_RTX
);
828 for_each_rtx (&insn
, replace_inc_dec_mem
, insn
);
832 /* Delete the insn and free all of the fields inside INSN_INFO. */
835 delete_dead_store_insn (insn_info_t insn_info
)
837 read_info_t read_info
;
842 check_for_inc_dec (insn_info
->insn
);
845 fprintf (dump_file
, "Locally deleting insn %d ",
846 INSN_UID (insn_info
->insn
));
847 if (insn_info
->store_rec
->alias_set
)
848 fprintf (dump_file
, "alias set %d\n",
849 (int) insn_info
->store_rec
->alias_set
);
851 fprintf (dump_file
, "\n");
854 free_store_info (insn_info
);
855 read_info
= insn_info
->read_rec
;
859 read_info_t next
= read_info
->next
;
860 pool_free (read_info_pool
, read_info
);
863 insn_info
->read_rec
= NULL
;
865 delete_insn (insn_info
->insn
);
867 insn_info
->insn
= NULL
;
869 insn_info
->wild_read
= false;
873 /* Set the store* bitmaps offset_map_size* fields in GROUP based on
877 set_usage_bits (group_info_t group
, HOST_WIDE_INT offset
, HOST_WIDE_INT width
)
881 if ((offset
> -MAX_OFFSET
) && (offset
< MAX_OFFSET
))
882 for (i
=offset
; i
<offset
+width
; i
++)
889 store1
= group
->store1_n
;
890 store2
= group
->store2_n
;
895 store1
= group
->store1_p
;
896 store2
= group
->store2_p
;
900 if (bitmap_bit_p (store1
, ai
))
901 bitmap_set_bit (store2
, ai
);
904 bitmap_set_bit (store1
, ai
);
907 if (group
->offset_map_size_n
< ai
)
908 group
->offset_map_size_n
= ai
;
912 if (group
->offset_map_size_p
< ai
)
913 group
->offset_map_size_p
= ai
;
920 /* Set the BB_INFO so that the last insn is marked as a wild read. */
923 add_wild_read (bb_info_t bb_info
)
925 insn_info_t insn_info
= bb_info
->last_insn
;
926 read_info_t
*ptr
= &insn_info
->read_rec
;
930 read_info_t next
= (*ptr
)->next
;
931 if ((*ptr
)->alias_set
== 0)
933 pool_free (read_info_pool
, *ptr
);
939 insn_info
->wild_read
= true;
940 active_local_stores
= NULL
;
944 /* Return true if X is a constant or one of the registers that behaves
945 as a constant over the life of a function. */
948 const_or_frame_p (rtx x
)
950 switch (GET_CODE (x
))
953 return MEM_READONLY_P (x
);
964 /* Note that we have to test for the actual rtx used for the frame
965 and arg pointers and not just the register number in case we have
966 eliminated the frame and/or arg pointer and are using it
968 if (x
== frame_pointer_rtx
|| x
== hard_frame_pointer_rtx
969 /* The arg pointer varies if it is not a fixed register. */
970 || (x
== arg_pointer_rtx
&& fixed_regs
[ARG_POINTER_REGNUM
])
971 || x
== pic_offset_table_rtx
)
980 /* Take all reasonable action to put the address of MEM into the form
981 that we can do analysis on.
983 The gold standard is to get the address into the form: address +
984 OFFSET where address is something that rtx_varies_p considers a
985 constant. When we can get the address in this form, we can do
986 global analysis on it. Note that for constant bases, address is
987 not actually returned, only the group_id. The address can be
990 If that fails, we try cselib to get a value we can at least use
991 locally. If that fails we return false.
993 The GROUP_ID is set to -1 for cselib bases and the index of the
994 group for non_varying bases.
996 FOR_READ is true if this is a mem read and false if not. */
999 canon_address (rtx mem
,
1000 alias_set_type
*alias_set_out
,
1002 HOST_WIDE_INT
*offset
,
1005 rtx mem_address
= XEXP (mem
, 0);
1006 rtx expanded_address
, address
;
1007 /* Make sure that cselib is has initialized all of the operands of
1008 the address before asking it to do the subst. */
1010 if (clear_alias_sets
)
1012 /* If this is a spill, do not do any further processing. */
1013 alias_set_type alias_set
= MEM_ALIAS_SET (mem
);
1015 fprintf (dump_file
, "found alias set %d\n", (int) alias_set
);
1016 if (bitmap_bit_p (clear_alias_sets
, alias_set
))
1018 struct clear_alias_mode_holder
*entry
1019 = clear_alias_set_lookup (alias_set
);
1021 /* If the modes do not match, we cannot process this set. */
1022 if (entry
->mode
!= GET_MODE (mem
))
1026 "disqualifying alias set %d, (%s) != (%s)\n",
1027 (int) alias_set
, GET_MODE_NAME (entry
->mode
),
1028 GET_MODE_NAME (GET_MODE (mem
)));
1030 bitmap_set_bit (disqualified_clear_alias_sets
, alias_set
);
1034 *alias_set_out
= alias_set
;
1035 *group_id
= clear_alias_group
->id
;
1042 cselib_lookup (mem_address
, Pmode
, 1);
1046 fprintf (dump_file
, " mem: ");
1047 print_inline_rtx (dump_file
, mem_address
, 0);
1048 fprintf (dump_file
, "\n");
1051 /* Use cselib to replace all of the reg references with the full
1052 expression. This will take care of the case where we have
1054 r_x = base + offset;
1059 val = *(base + offset);
1062 expanded_address
= cselib_expand_value_rtx (mem_address
, scratch
, 5);
1064 /* If this fails, just go with the mem_address. */
1065 if (!expanded_address
)
1066 expanded_address
= mem_address
;
1068 /* Split the address into canonical BASE + OFFSET terms. */
1069 address
= canon_rtx (expanded_address
);
1075 fprintf (dump_file
, "\n after cselib_expand address: ");
1076 print_inline_rtx (dump_file
, expanded_address
, 0);
1077 fprintf (dump_file
, "\n");
1079 fprintf (dump_file
, "\n after canon_rtx address: ");
1080 print_inline_rtx (dump_file
, address
, 0);
1081 fprintf (dump_file
, "\n");
1084 if (GET_CODE (address
) == CONST
)
1085 address
= XEXP (address
, 0);
1087 if (GET_CODE (address
) == PLUS
&& GET_CODE (XEXP (address
, 1)) == CONST_INT
)
1089 *offset
= INTVAL (XEXP (address
, 1));
1090 address
= XEXP (address
, 0);
1093 if (const_or_frame_p (address
))
1095 group_info_t group
= get_group_info (address
);
1098 fprintf (dump_file
, " gid=%d offset=%d \n", group
->id
, (int)*offset
);
1100 *group_id
= group
->id
;
1104 *base
= cselib_lookup (address
, Pmode
, true);
1110 fprintf (dump_file
, " no cselib val - should be a wild read.\n");
1114 fprintf (dump_file
, " varying cselib base=%d offset = %d\n",
1115 (*base
)->value
, (int)*offset
);
1121 /* Clear the rhs field from the active_local_stores array. */
1124 clear_rhs_from_active_local_stores (void)
1126 insn_info_t ptr
= active_local_stores
;
1130 store_info_t store_info
= ptr
->store_rec
;
1131 /* Skip the clobbers. */
1132 while (!store_info
->is_set
)
1133 store_info
= store_info
->next
;
1135 store_info
->rhs
= NULL
;
1137 ptr
= ptr
->next_local_store
;
1142 /* BODY is an instruction pattern that belongs to INSN. Return 1 if
1143 there is a candidate store, after adding it to the appropriate
1144 local store group if so. */
1147 record_store (rtx body
, bb_info_t bb_info
)
1150 HOST_WIDE_INT offset
= 0;
1151 HOST_WIDE_INT width
= 0;
1152 alias_set_type spill_alias_set
;
1153 insn_info_t insn_info
= bb_info
->last_insn
;
1154 store_info_t store_info
= NULL
;
1156 cselib_val
*base
= NULL
;
1157 insn_info_t ptr
, last
;
1158 bool store_is_unused
;
1160 if (GET_CODE (body
) != SET
&& GET_CODE (body
) != CLOBBER
)
1163 /* If this is not used, then this cannot be used to keep the insn
1164 from being deleted. On the other hand, it does provide something
1165 that can be used to prove that another store is dead. */
1167 = (find_reg_note (insn_info
->insn
, REG_UNUSED
, body
) != NULL
);
1169 /* Check whether that value is a suitable memory location. */
1170 mem
= SET_DEST (body
);
1173 /* If the set or clobber is unused, then it does not effect our
1174 ability to get rid of the entire insn. */
1175 if (!store_is_unused
)
1176 insn_info
->cannot_delete
= true;
1180 /* At this point we know mem is a mem. */
1181 if (GET_MODE (mem
) == BLKmode
)
1183 if (GET_CODE (XEXP (mem
, 0)) == SCRATCH
)
1186 fprintf (dump_file
, " adding wild read for (clobber (mem:BLK (scratch))\n");
1187 add_wild_read (bb_info
);
1188 insn_info
->cannot_delete
= true;
1190 else if (!store_is_unused
)
1192 /* If the set or clobber is unused, then it does not effect our
1193 ability to get rid of the entire insn. */
1194 insn_info
->cannot_delete
= true;
1195 clear_rhs_from_active_local_stores ();
1200 /* We can still process a volatile mem, we just cannot delete it. */
1201 if (MEM_VOLATILE_P (mem
))
1202 insn_info
->cannot_delete
= true;
1204 if (!canon_address (mem
, &spill_alias_set
, &group_id
, &offset
, &base
))
1206 clear_rhs_from_active_local_stores ();
1210 width
= GET_MODE_SIZE (GET_MODE (mem
));
1212 if (spill_alias_set
)
1214 bitmap store1
= clear_alias_group
->store1_p
;
1215 bitmap store2
= clear_alias_group
->store2_p
;
1217 if (bitmap_bit_p (store1
, spill_alias_set
))
1218 bitmap_set_bit (store2
, spill_alias_set
);
1220 bitmap_set_bit (store1
, spill_alias_set
);
1222 if (clear_alias_group
->offset_map_size_p
< spill_alias_set
)
1223 clear_alias_group
->offset_map_size_p
= spill_alias_set
;
1225 store_info
= pool_alloc (rtx_store_info_pool
);
1228 fprintf (dump_file
, " processing spill store %d(%s)\n",
1229 (int) spill_alias_set
, GET_MODE_NAME (GET_MODE (mem
)));
1231 else if (group_id
>= 0)
1233 /* In the restrictive case where the base is a constant or the
1234 frame pointer we can do global analysis. */
1237 = VEC_index (group_info_t
, rtx_group_vec
, group_id
);
1239 store_info
= pool_alloc (rtx_store_info_pool
);
1240 set_usage_bits (group
, offset
, width
);
1243 fprintf (dump_file
, " processing const base store gid=%d[%d..%d)\n",
1244 group_id
, (int)offset
, (int)(offset
+width
));
1248 store_info
= pool_alloc (cse_store_info_pool
);
1249 insn_info
->contains_cselib_groups
= true;
1253 fprintf (dump_file
, " processing cselib store [%d..%d)\n",
1254 (int)offset
, (int)(offset
+width
));
1257 /* Check to see if this stores causes some other stores to be
1259 ptr
= active_local_stores
;
1264 insn_info_t next
= ptr
->next_local_store
;
1265 store_info_t s_info
= ptr
->store_rec
;
1268 /* Skip the clobbers. We delete the active insn if this insn
1269 shadows the set. To have been put on the active list, it
1270 has exactly on set. */
1271 while (!s_info
->is_set
)
1272 s_info
= s_info
->next
;
1274 if (s_info
->alias_set
!= spill_alias_set
)
1276 else if (s_info
->alias_set
)
1278 struct clear_alias_mode_holder
*entry
1279 = clear_alias_set_lookup (s_info
->alias_set
);
1280 /* Generally, spills cannot be processed if and of the
1281 references to the slot have a different mode. But if
1282 we are in the same block and mode is exactly the same
1283 between this store and one before in the same block,
1284 we can still delete it. */
1285 if ((GET_MODE (mem
) == GET_MODE (s_info
->mem
))
1286 && (GET_MODE (mem
) == entry
->mode
))
1289 s_info
->positions_needed
= 0;
1292 fprintf (dump_file
, " trying spill store in insn=%d alias_set=%d\n",
1293 INSN_UID (ptr
->insn
), (int) s_info
->alias_set
);
1295 else if ((s_info
->group_id
== group_id
)
1296 && (s_info
->cse_base
== base
))
1300 fprintf (dump_file
, " trying store in insn=%d gid=%d[%d..%d)\n",
1301 INSN_UID (ptr
->insn
), s_info
->group_id
,
1302 (int)s_info
->begin
, (int)s_info
->end
);
1303 for (i
= offset
; i
< offset
+width
; i
++)
1304 if (i
>= s_info
->begin
&& i
< s_info
->end
)
1305 s_info
->positions_needed
&= ~(1L << (i
- s_info
->begin
));
1307 else if (s_info
->rhs
)
1308 /* Need to see if it is possible for this store to overwrite
1309 the value of store_info. If it is, set the rhs to NULL to
1310 keep it from being used to remove a load. */
1312 if (canon_true_dependence (s_info
->mem
,
1313 GET_MODE (s_info
->mem
),
1319 /* An insn can be deleted if every position of every one of
1320 its s_infos is zero. */
1321 if (s_info
->positions_needed
!= 0)
1326 insn_info_t insn_to_delete
= ptr
;
1329 last
->next_local_store
= ptr
->next_local_store
;
1331 active_local_stores
= ptr
->next_local_store
;
1333 delete_dead_store_insn (insn_to_delete
);
1341 gcc_assert ((unsigned) width
< sizeof (store_info
->positions_needed
) * CHAR_BIT
);
1343 /* Finish filling in the store_info. */
1344 store_info
->next
= insn_info
->store_rec
;
1345 insn_info
->store_rec
= store_info
;
1346 store_info
->mem
= canon_rtx (mem
);
1347 store_info
->alias_set
= spill_alias_set
;
1348 store_info
->mem_addr
= get_addr (XEXP (mem
, 0));
1349 store_info
->cse_base
= base
;
1350 store_info
->positions_needed
= (1L << width
) - 1;
1351 store_info
->group_id
= group_id
;
1352 store_info
->begin
= offset
;
1353 store_info
->end
= offset
+ width
;
1354 store_info
->is_set
= GET_CODE (body
) == SET
;
1356 if (store_info
->is_set
1357 /* No place to keep the value after ra. */
1358 && !reload_completed
1359 /* The careful reviewer may wish to comment my checking that the
1360 rhs of a store is always a reg. */
1361 && REG_P (SET_SRC (body
))
1362 /* Sometimes the store and reload is used for truncation and
1364 && !(FLOAT_MODE_P (GET_MODE (mem
)) && (flag_float_store
)))
1365 store_info
->rhs
= SET_SRC (body
);
1367 store_info
->rhs
= NULL
;
1369 /* If this is a clobber, we return 0. We will only be able to
1370 delete this insn if there is only one store USED store, but we
1371 can use the clobber to delete other stores earlier. */
1372 return store_info
->is_set
? 1 : 0;
1377 dump_insn_info (const char * start
, insn_info_t insn_info
)
1379 fprintf (dump_file
, "%s insn=%d %s\n", start
,
1380 INSN_UID (insn_info
->insn
),
1381 insn_info
->store_rec
? "has store" : "naked");
1385 /* If the modes are different and the value's source and target do not
1386 line up, we need to extract the value from lower part of the rhs of
1387 the store, shift it, and then put it into a form that can be shoved
1388 into the read_insn. This function generates a right SHIFT of a
1389 value that is at least ACCESS_SIZE bytes wide of READ_MODE. The
1390 shift sequence is returned or NULL if we failed to find a
1394 find_shift_sequence (rtx read_reg
,
1396 store_info_t store_info
,
1397 read_info_t read_info
,
1400 enum machine_mode store_mode
= GET_MODE (store_info
->mem
);
1401 enum machine_mode read_mode
= GET_MODE (read_info
->mem
);
1402 rtx chosen_seq
= NULL
;
1404 /* Some machines like the x86 have shift insns for each size of
1405 operand. Other machines like the ppc or the ia-64 may only have
1406 shift insns that shift values within 32 or 64 bit registers.
1407 This loop tries to find the smallest shift insn that will right
1408 justify the value we want to read but is available in one insn on
1411 for (; access_size
< UNITS_PER_WORD
; access_size
*= 2)
1413 rtx target
, new_reg
, shift_seq
, insn
;
1414 enum machine_mode new_mode
;
1417 /* Try a wider mode if truncating the store mode to ACCESS_SIZE
1418 bytes requires a real instruction. */
1419 if (access_size
< GET_MODE_SIZE (store_mode
)
1420 && !TRULY_NOOP_TRUNCATION (access_size
* BITS_PER_UNIT
,
1421 GET_MODE_BITSIZE (store_mode
)))
1424 new_mode
= smallest_mode_for_size (access_size
* BITS_PER_UNIT
,
1425 GET_MODE_CLASS (read_mode
));
1426 new_reg
= gen_reg_rtx (new_mode
);
1430 /* In theory we could also check for an ashr. Ian Taylor knows
1431 of one dsp where the cost of these two was not the same. But
1432 this really is a rare case anyway. */
1433 target
= expand_binop (new_mode
, lshr_optab
, new_reg
,
1434 GEN_INT (shift
), new_reg
, 1, OPTAB_DIRECT
);
1436 shift_seq
= get_insns ();
1439 if (target
!= new_reg
|| shift_seq
== NULL
)
1443 for (insn
= shift_seq
; insn
!= NULL_RTX
; insn
= NEXT_INSN (insn
))
1445 cost
+= insn_rtx_cost (PATTERN (insn
));
1447 /* The computation up to here is essentially independent
1448 of the arguments and could be precomputed. It may
1449 not be worth doing so. We could precompute if
1450 worthwhile or at least cache the results. The result
1451 technically depends on SHIFT, ACCESS_SIZE, and
1452 GET_MODE_CLASS (READ_MODE). But in practice the
1453 answer will depend only on ACCESS_SIZE. */
1455 if (cost
> COSTS_N_INSNS (1))
1458 /* We found an acceptable shift. Generate a move to
1459 take the value from the store and put it into the
1460 shift pseudo, then shift it, then generate another
1461 move to put in into the target of the read. */
1463 emit_move_insn (new_reg
, gen_lowpart (new_mode
, store_info
->rhs
));
1464 emit_insn (shift_seq
);
1465 convert_move (read_reg
, new_reg
, 1);
1469 fprintf (dump_file
, " -- adding extract insn r%d:%s = r%d:%s\n",
1470 REGNO (new_reg
), GET_MODE_NAME (new_mode
),
1471 REGNO (store_info
->rhs
), GET_MODE_NAME (store_mode
));
1473 fprintf (dump_file
, " -- with shift of r%d by %d\n",
1474 REGNO(new_reg
), shift
);
1475 fprintf (dump_file
, " -- and second extract insn r%d:%s = r%d:%s\n",
1476 REGNO (read_reg
), GET_MODE_NAME (read_mode
),
1477 REGNO (new_reg
), GET_MODE_NAME (new_mode
));
1480 /* Get the three insn sequence and return it. */
1481 chosen_seq
= get_insns ();
1490 /* Take a sequence of:
1513 Depending on the alignment and the mode of the store and
1517 The STORE_INFO and STORE_INSN are for the store and READ_INFO
1518 and READ_INSN are for the read. Return true if the replacement
1522 replace_read (store_info_t store_info
, insn_info_t store_insn
,
1523 read_info_t read_info
, insn_info_t read_insn
, rtx
*loc
)
1525 enum machine_mode store_mode
= GET_MODE (store_info
->mem
);
1526 enum machine_mode read_mode
= GET_MODE (read_info
->mem
);
1528 int access_size
; /* In bytes. */
1529 rtx read_reg
= gen_reg_rtx (read_mode
);
1530 rtx shift_seq
= NULL
;
1535 if (GET_MODE_CLASS (read_mode
) != GET_MODE_CLASS (store_mode
))
1538 /* To get here the read is within the boundaries of the write so
1539 shift will never be negative. Start out with the shift being in
1541 if (BYTES_BIG_ENDIAN
)
1542 shift
= store_info
->end
- read_info
->end
;
1544 shift
= read_info
->begin
- store_info
->begin
;
1546 access_size
= shift
+ GET_MODE_SIZE (read_mode
);
1548 /* From now on it is bits. */
1549 shift
*= BITS_PER_UNIT
;
1551 /* We need to keep this in perspective. We are replacing a read
1552 with a sequence of insns, but the read will almost certainly be
1553 in cache, so it is not going to be an expensive one. Thus, we
1554 are not willing to do a multi insn shift or worse a subroutine
1555 call to get rid of the read. */
1558 if (access_size
> UNITS_PER_WORD
|| FLOAT_MODE_P (store_mode
))
1561 shift_seq
= find_shift_sequence (read_reg
, access_size
, store_info
,
1568 fprintf (dump_file
, "replacing load at %d from store at %d\n",
1569 INSN_UID (read_insn
->insn
), INSN_UID (store_insn
->insn
));
1571 if (validate_change (read_insn
->insn
, loc
, read_reg
, 0))
1574 deferred_change_t deferred_change
= pool_alloc (deferred_change_pool
);
1576 if (read_mode
== store_mode
)
1580 /* The modes are the same and everything lines up. Just
1581 generate a simple move. */
1582 emit_move_insn (read_reg
, store_info
->rhs
);
1584 fprintf (dump_file
, " -- adding move insn r%d = r%d\n",
1585 REGNO (read_reg
), REGNO (store_info
->rhs
));
1586 insns
= get_insns ();
1593 /* The modes are different but the lsb are in the same
1594 place, we need to extract the value in the right from the
1595 rhs of the store. */
1597 convert_move (read_reg
, store_info
->rhs
, 1);
1600 fprintf (dump_file
, " -- adding extract insn r%d:%s = r%d:%s\n",
1601 REGNO (read_reg
), GET_MODE_NAME (read_mode
),
1602 REGNO (store_info
->rhs
), GET_MODE_NAME (store_mode
));
1603 insns
= get_insns ();
1607 /* Insert this right before the store insn where it will be safe
1608 from later insns that might change it before the read. */
1609 emit_insn_before (insns
, store_insn
->insn
);
1611 /* And now for the kludge part: cselib croaks if you just
1612 return at this point. There are two reasons for this:
1614 1) Cselib has an idea of how many pseudos there are and
1615 that does not include the new ones we just added.
1617 2) Cselib does not know about the move insn we added
1618 above the store_info, and there is no way to tell it
1619 about it, because it has "moved on".
1621 Problem (1) is fixable with a certain amount of engineering.
1622 Problem (2) is requires starting the bb from scratch. This
1625 So we are just going to have to lie. The move/extraction
1626 insns are not really an issue, cselib did not see them. But
1627 the use of the new pseudo read_insn is a real problem because
1628 cselib has not scanned this insn. The way that we solve this
1629 problem is that we are just going to put the mem back for now
1630 and when we are finished with the block, we undo this. We
1631 keep a table of mems to get rid of. At the end of the basic
1632 block we can put them back. */
1634 *loc
= read_info
->mem
;
1635 deferred_change
->next
= deferred_change_list
;
1636 deferred_change_list
= deferred_change
;
1637 deferred_change
->loc
= loc
;
1638 deferred_change
->reg
= read_reg
;
1640 /* Get rid of the read_info, from the point of view of the
1641 rest of dse, play like this read never happened. */
1642 read_insn
->read_rec
= read_info
->next
;
1643 pool_free (read_info_pool
, read_info
);
1649 fprintf (dump_file
, " -- validation failure\n");
1654 /* A for_each_rtx callback in which DATA is the bb_info. Check to see
1655 if LOC is a mem and if it is look at the address and kill any
1656 appropriate stores that may be active. */
1659 check_mem_read_rtx (rtx
*loc
, void *data
)
1663 insn_info_t insn_info
;
1664 HOST_WIDE_INT offset
= 0;
1665 HOST_WIDE_INT width
= 0;
1666 alias_set_type spill_alias_set
= 0;
1667 cselib_val
*base
= NULL
;
1669 read_info_t read_info
;
1671 if (!mem
|| !MEM_P (mem
))
1674 bb_info
= (bb_info_t
) data
;
1675 insn_info
= bb_info
->last_insn
;
1677 if ((MEM_ALIAS_SET (mem
) == ALIAS_SET_MEMORY_BARRIER
)
1678 || (MEM_VOLATILE_P (mem
)))
1681 fprintf (dump_file
, " adding wild read, volatile or barrier.\n");
1682 add_wild_read (bb_info
);
1683 insn_info
->cannot_delete
= true;
1687 /* If it is reading readonly mem, then there can be no conflict with
1689 if (MEM_READONLY_P (mem
))
1692 if (!canon_address (mem
, &spill_alias_set
, &group_id
, &offset
, &base
))
1695 fprintf (dump_file
, " adding wild read, canon_address failure.\n");
1696 add_wild_read (bb_info
);
1700 if (GET_MODE (mem
) == BLKmode
)
1703 width
= GET_MODE_SIZE (GET_MODE (mem
));
1705 read_info
= pool_alloc (read_info_pool
);
1706 read_info
->group_id
= group_id
;
1707 read_info
->mem
= mem
;
1708 read_info
->alias_set
= spill_alias_set
;
1709 read_info
->begin
= offset
;
1710 read_info
->end
= offset
+ width
;
1711 read_info
->next
= insn_info
->read_rec
;
1712 insn_info
->read_rec
= read_info
;
1714 /* We ignore the clobbers in store_info. The is mildly aggressive,
1715 but there really should not be a clobber followed by a read. */
1717 if (spill_alias_set
)
1719 insn_info_t i_ptr
= active_local_stores
;
1720 insn_info_t last
= NULL
;
1723 fprintf (dump_file
, " processing spill load %d\n",
1724 (int) spill_alias_set
);
1728 store_info_t store_info
= i_ptr
->store_rec
;
1730 /* Skip the clobbers. */
1731 while (!store_info
->is_set
)
1732 store_info
= store_info
->next
;
1734 if (store_info
->alias_set
== spill_alias_set
)
1737 dump_insn_info ("removing from active", i_ptr
);
1740 last
->next_local_store
= i_ptr
->next_local_store
;
1742 active_local_stores
= i_ptr
->next_local_store
;
1746 i_ptr
= i_ptr
->next_local_store
;
1749 else if (group_id
>= 0)
1751 /* This is the restricted case where the base is a constant or
1752 the frame pointer and offset is a constant. */
1753 insn_info_t i_ptr
= active_local_stores
;
1754 insn_info_t last
= NULL
;
1759 fprintf (dump_file
, " processing const load gid=%d[BLK]\n",
1762 fprintf (dump_file
, " processing const load gid=%d[%d..%d)\n",
1763 group_id
, (int)offset
, (int)(offset
+width
));
1768 bool remove
= false;
1769 store_info_t store_info
= i_ptr
->store_rec
;
1771 /* Skip the clobbers. */
1772 while (!store_info
->is_set
)
1773 store_info
= store_info
->next
;
1775 /* There are three cases here. */
1776 if (store_info
->group_id
< 0)
1777 /* We have a cselib store followed by a read from a
1780 = canon_true_dependence (store_info
->mem
,
1781 GET_MODE (store_info
->mem
),
1782 store_info
->mem_addr
,
1785 else if (group_id
== store_info
->group_id
)
1787 /* This is a block mode load. We may get lucky and
1788 canon_true_dependence may save the day. */
1791 = canon_true_dependence (store_info
->mem
,
1792 GET_MODE (store_info
->mem
),
1793 store_info
->mem_addr
,
1796 /* If this read is just reading back something that we just
1797 stored, rewrite the read. */
1801 && (offset
>= store_info
->begin
)
1802 && (offset
+ width
<= store_info
->end
))
1804 int mask
= ((1L << width
) - 1) << (offset
- store_info
->begin
);
1806 if ((store_info
->positions_needed
& mask
) == mask
1807 && replace_read (store_info
, i_ptr
,
1808 read_info
, insn_info
, loc
))
1811 /* The bases are the same, just see if the offsets
1813 if ((offset
< store_info
->end
)
1814 && (offset
+ width
> store_info
->begin
))
1820 The else case that is missing here is that the
1821 bases are constant but different. There is nothing
1822 to do here because there is no overlap. */
1827 dump_insn_info ("removing from active", i_ptr
);
1830 last
->next_local_store
= i_ptr
->next_local_store
;
1832 active_local_stores
= i_ptr
->next_local_store
;
1836 i_ptr
= i_ptr
->next_local_store
;
1841 insn_info_t i_ptr
= active_local_stores
;
1842 insn_info_t last
= NULL
;
1845 fprintf (dump_file
, " processing cselib load mem:");
1846 print_inline_rtx (dump_file
, mem
, 0);
1847 fprintf (dump_file
, "\n");
1852 bool remove
= false;
1853 store_info_t store_info
= i_ptr
->store_rec
;
1856 fprintf (dump_file
, " processing cselib load against insn %d\n",
1857 INSN_UID (i_ptr
->insn
));
1859 /* Skip the clobbers. */
1860 while (!store_info
->is_set
)
1861 store_info
= store_info
->next
;
1863 /* If this read is just reading back something that we just
1864 stored, rewrite the read. */
1866 && store_info
->group_id
== -1
1867 && store_info
->cse_base
== base
1868 && (offset
>= store_info
->begin
)
1869 && (offset
+ width
<= store_info
->end
))
1871 int mask
= ((1L << width
) - 1) << (offset
- store_info
->begin
);
1873 if ((store_info
->positions_needed
& mask
) == mask
1874 && replace_read (store_info
, i_ptr
,
1875 read_info
, insn_info
, loc
))
1879 if (!store_info
->alias_set
)
1880 remove
= canon_true_dependence (store_info
->mem
,
1881 GET_MODE (store_info
->mem
),
1882 store_info
->mem_addr
,
1888 dump_insn_info ("removing from active", i_ptr
);
1891 last
->next_local_store
= i_ptr
->next_local_store
;
1893 active_local_stores
= i_ptr
->next_local_store
;
1897 i_ptr
= i_ptr
->next_local_store
;
1903 /* A for_each_rtx callback in which DATA points the INSN_INFO for
1904 as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns
1905 true for any part of *LOC. */
1908 check_mem_read_use (rtx
*loc
, void *data
)
1910 for_each_rtx (loc
, check_mem_read_rtx
, data
);
1913 /* Apply record_store to all candidate stores in INSN. Mark INSN
1914 if some part of it is not a candidate store and assigns to a
1915 non-register target. */
1918 scan_insn (bb_info_t bb_info
, rtx insn
)
1921 insn_info_t insn_info
= pool_alloc (insn_info_pool
);
1923 memset (insn_info
, 0, sizeof (struct insn_info
));
1926 fprintf (dump_file
, "\n**scanning insn=%d\n",
1929 insn_info
->prev_insn
= bb_info
->last_insn
;
1930 insn_info
->insn
= insn
;
1931 bb_info
->last_insn
= insn_info
;
1934 /* Cselib clears the table for this case, so we have to essentially
1936 if (NONJUMP_INSN_P (insn
)
1937 && GET_CODE (PATTERN (insn
)) == ASM_OPERANDS
1938 && MEM_VOLATILE_P (PATTERN (insn
)))
1940 add_wild_read (bb_info
);
1941 insn_info
->cannot_delete
= true;
1945 /* Look at all of the uses in the insn. */
1946 note_uses (&PATTERN (insn
), check_mem_read_use
, bb_info
);
1950 insn_info
->cannot_delete
= true;
1951 /* Const functions cannot do anything bad i.e. read memory,
1952 however, they can read their parameters which may have been
1953 pushed onto the stack. */
1954 if (CONST_OR_PURE_CALL_P (insn
) && !pure_call_p (insn
))
1956 insn_info_t i_ptr
= active_local_stores
;
1957 insn_info_t last
= NULL
;
1960 fprintf (dump_file
, "const call %d\n", INSN_UID (insn
));
1964 store_info_t store_info
= i_ptr
->store_rec
;
1966 /* Skip the clobbers. */
1967 while (!store_info
->is_set
)
1968 store_info
= store_info
->next
;
1970 /* Remove the frame related stores. */
1971 if (store_info
->group_id
>= 0
1972 && VEC_index (group_info_t
, rtx_group_vec
, store_info
->group_id
)->frame_related
)
1975 dump_insn_info ("removing from active", i_ptr
);
1978 last
->next_local_store
= i_ptr
->next_local_store
;
1980 active_local_stores
= i_ptr
->next_local_store
;
1984 i_ptr
= i_ptr
->next_local_store
;
1987 insn_info
->stack_read
= true;
1992 /* Every other call, including pure functions may read memory. */
1993 add_wild_read (bb_info
);
1997 /* Assuming that there are sets in these insns, we cannot delete
1999 if ((GET_CODE (PATTERN (insn
)) == CLOBBER
)
2000 || volatile_refs_p (PATTERN (insn
))
2001 || (flag_non_call_exceptions
&& may_trap_p (PATTERN (insn
)))
2002 || (RTX_FRAME_RELATED_P (insn
))
2003 || find_reg_note (insn
, REG_FRAME_RELATED_EXPR
, NULL_RTX
))
2004 insn_info
->cannot_delete
= true;
2006 body
= PATTERN (insn
);
2007 if (GET_CODE (body
) == PARALLEL
)
2010 for (i
= 0; i
< XVECLEN (body
, 0); i
++)
2011 mems_found
+= record_store (XVECEXP (body
, 0, i
), bb_info
);
2014 mems_found
+= record_store (body
, bb_info
);
2017 fprintf (dump_file
, "mems_found = %d, cannot_delete = %s\n",
2018 mems_found
, insn_info
->cannot_delete
? "true" : "false");
2020 /* If we found some sets of mems, and the insn has not been marked
2021 cannot delete, add it into the active_local_stores so that it can
2022 be locally deleted if found dead. Otherwise mark it as cannot
2023 delete. This simplifies the processing later. */
2024 if (mems_found
== 1 && !insn_info
->cannot_delete
)
2026 insn_info
->next_local_store
= active_local_stores
;
2027 active_local_stores
= insn_info
;
2030 insn_info
->cannot_delete
= true;
2034 /* Remove BASE from the set of active_local_stores. This is a
2035 callback from cselib that is used to get rid of the stores in
2036 active_local_stores. */
2039 remove_useless_values (cselib_val
*base
)
2041 insn_info_t insn_info
= active_local_stores
;
2042 insn_info_t last
= NULL
;
2046 store_info_t store_info
= insn_info
->store_rec
;
2047 bool delete = false;
2049 /* If ANY of the store_infos match the cselib group that is
2050 being deleted, then the insn can not be deleted. */
2053 if ((store_info
->group_id
== -1)
2054 && (store_info
->cse_base
== base
))
2059 store_info
= store_info
->next
;
2065 last
->next_local_store
= insn_info
->next_local_store
;
2067 active_local_stores
= insn_info
->next_local_store
;
2068 free_store_info (insn_info
);
2073 insn_info
= insn_info
->next_local_store
;
2078 /* Do all of step 1. */
2085 cselib_init (false);
2086 all_blocks
= BITMAP_ALLOC (NULL
);
2087 bitmap_set_bit (all_blocks
, ENTRY_BLOCK
);
2088 bitmap_set_bit (all_blocks
, EXIT_BLOCK
);
2093 bb_info_t bb_info
= pool_alloc (bb_info_pool
);
2095 memset (bb_info
, 0, sizeof (struct bb_info
));
2096 bitmap_set_bit (all_blocks
, bb
->index
);
2098 bb_table
[bb
->index
] = bb_info
;
2099 cselib_discard_hook
= remove_useless_values
;
2101 if (bb
->index
>= NUM_FIXED_BLOCKS
)
2106 = create_alloc_pool ("cse_store_info_pool",
2107 sizeof (struct store_info
), 100);
2108 active_local_stores
= NULL
;
2109 cselib_clear_table ();
2111 /* Scan the insns. */
2112 FOR_BB_INSNS (bb
, insn
)
2115 scan_insn (bb_info
, insn
);
2116 cselib_process_insn (insn
);
2119 /* This is something of a hack, because the global algorithm
2120 is supposed to take care of the case where stores go dead
2121 at the end of the function. However, the global
2122 algorithm must take a more conservative view of block
2123 mode reads than the local alg does. So to get the case
2124 where you have a store to the frame followed by a non
2125 overlapping block more read, we look at the active local
2126 stores at the end of the function and delete all of the
2127 frame and spill based ones. */
2128 if (stores_off_frame_dead_at_return
2129 && (EDGE_COUNT (bb
->succs
) == 0
2130 || (single_succ_p (bb
)
2131 && single_succ (bb
) == EXIT_BLOCK_PTR
2132 && ! current_function_calls_eh_return
)))
2134 insn_info_t i_ptr
= active_local_stores
;
2137 store_info_t store_info
= i_ptr
->store_rec
;
2139 /* Skip the clobbers. */
2140 while (!store_info
->is_set
)
2141 store_info
= store_info
->next
;
2142 if (store_info
->alias_set
)
2143 delete_dead_store_insn (i_ptr
);
2145 if (store_info
->group_id
>= 0)
2148 = VEC_index (group_info_t
, rtx_group_vec
, store_info
->group_id
);
2149 if (group
->frame_related
)
2150 delete_dead_store_insn (i_ptr
);
2153 i_ptr
= i_ptr
->next_local_store
;
2157 /* Get rid of the loads that were discovered in
2158 replace_read. Cselib is finished with this block. */
2159 while (deferred_change_list
)
2161 deferred_change_t next
= deferred_change_list
->next
;
2163 /* There is no reason to validate this change. That was
2165 *deferred_change_list
->loc
= deferred_change_list
->reg
;
2166 pool_free (deferred_change_pool
, deferred_change_list
);
2167 deferred_change_list
= next
;
2170 /* Get rid of all of the cselib based store_infos in this
2171 block and mark the containing insns as not being
2173 ptr
= bb_info
->last_insn
;
2176 if (ptr
->contains_cselib_groups
)
2177 free_store_info (ptr
);
2178 ptr
= ptr
->prev_insn
;
2181 free_alloc_pool (cse_store_info_pool
);
2186 htab_empty (rtx_group_table
);
2190 /*----------------------------------------------------------------------------
2193 Assign each byte position in the stores that we are going to
2194 analyze globally to a position in the bitmaps. Returns true if
2195 there are any bit positions assigned.
2196 ----------------------------------------------------------------------------*/
2199 dse_step2_init (void)
2204 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2206 /* For all non stack related bases, we only consider a store to
2207 be deletable if there are two or more stores for that
2208 position. This is because it takes one store to make the
2209 other store redundant. However, for the stores that are
2210 stack related, we consider them if there is only one store
2211 for the position. We do this because the stack related
2212 stores can be deleted if their is no read between them and
2213 the end of the function.
2215 To make this work in the current framework, we take the stack
2216 related bases add all of the bits from store1 into store2.
2217 This has the effect of making the eligible even if there is
2220 if (stores_off_frame_dead_at_return
&& group
->frame_related
)
2222 bitmap_ior_into (group
->store2_n
, group
->store1_n
);
2223 bitmap_ior_into (group
->store2_p
, group
->store1_p
);
2225 fprintf (dump_file
, "group %d is frame related ", i
);
2228 group
->offset_map_size_n
++;
2229 group
->offset_map_n
= XNEWVEC (int, group
->offset_map_size_n
);
2230 group
->offset_map_size_p
++;
2231 group
->offset_map_p
= XNEWVEC (int, group
->offset_map_size_p
);
2232 group
->process_globally
= false;
2235 fprintf (dump_file
, "group %d(%d+%d): ", i
,
2236 (int)bitmap_count_bits (group
->store2_n
),
2237 (int)bitmap_count_bits (group
->store2_p
));
2238 bitmap_print (dump_file
, group
->store2_n
, "n ", " ");
2239 bitmap_print (dump_file
, group
->store2_p
, "p ", "\n");
2245 /* Init the offset tables for the normal case. */
2248 dse_step2_nospill (void)
2252 /* Position 0 is unused because 0 is used in the maps to mean
2254 current_position
= 1;
2256 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2261 if (group
== clear_alias_group
)
2264 memset (group
->offset_map_n
, 0, sizeof(int) * group
->offset_map_size_n
);
2265 memset (group
->offset_map_p
, 0, sizeof(int) * group
->offset_map_size_p
);
2266 bitmap_clear (group
->group_kill
);
2268 EXECUTE_IF_SET_IN_BITMAP (group
->store2_n
, 0, j
, bi
)
2270 bitmap_set_bit (group
->group_kill
, current_position
);
2271 group
->offset_map_n
[j
] = current_position
++;
2272 group
->process_globally
= true;
2274 EXECUTE_IF_SET_IN_BITMAP (group
->store2_p
, 0, j
, bi
)
2276 bitmap_set_bit (group
->group_kill
, current_position
);
2277 group
->offset_map_p
[j
] = current_position
++;
2278 group
->process_globally
= true;
2281 return current_position
!= 1;
2285 /* Init the offset tables for the spill case. */
2288 dse_step2_spill (void)
2291 group_info_t group
= clear_alias_group
;
2294 /* Position 0 is unused because 0 is used in the maps to mean
2296 current_position
= 1;
2300 bitmap_print (dump_file
, clear_alias_sets
,
2301 "clear alias sets ", "\n");
2302 bitmap_print (dump_file
, disqualified_clear_alias_sets
,
2303 "disqualified clear alias sets ", "\n");
2306 memset (group
->offset_map_n
, 0, sizeof(int) * group
->offset_map_size_n
);
2307 memset (group
->offset_map_p
, 0, sizeof(int) * group
->offset_map_size_p
);
2308 bitmap_clear (group
->group_kill
);
2310 /* Remove the disqualified positions from the store2_p set. */
2311 bitmap_and_compl_into (group
->store2_p
, disqualified_clear_alias_sets
);
2313 /* We do not need to process the store2_n set because
2314 alias_sets are always positive. */
2315 EXECUTE_IF_SET_IN_BITMAP (group
->store2_p
, 0, j
, bi
)
2317 bitmap_set_bit (group
->group_kill
, current_position
);
2318 group
->offset_map_p
[j
] = current_position
++;
2319 group
->process_globally
= true;
2322 return current_position
!= 1;
2327 /*----------------------------------------------------------------------------
2330 Build the bit vectors for the transfer functions.
2331 ----------------------------------------------------------------------------*/
2334 /* Note that this is NOT a general purpose function. Any mem that has
2335 an alias set registered here expected to be COMPLETELY unaliased:
2336 i.e it's addresses are not and need not be examined.
2338 It is known that all references to this address will have this
2339 alias set and there are NO other references to this address in the
2342 Currently the only place that is known to be clean enough to use
2343 this interface is the code that assigns the spill locations.
2345 All of the mems that have alias_sets registered are subjected to a
2346 very powerful form of dse where function calls, volatile reads and
2347 writes, and reads from random location are not taken into account.
2349 It is also assumed that these locations go dead when the function
2350 returns. This assumption could be relaxed if there were found to
2351 be places that this assumption was not correct.
2353 The MODE is passed in and saved. The mode of each load or store to
2354 a mem with ALIAS_SET is checked against MEM. If the size of that
2355 load or store is different from MODE, processing is halted on this
2356 alias set. For the vast majority of aliases sets, all of the loads
2357 and stores will use the same mode. But vectors are treated
2358 differently: the alias set is established for the entire vector,
2359 but reload will insert loads and stores for individual elements and
2360 we do not necessarily have the information to track those separate
2361 elements. So when we see a mode mismatch, we just bail. */
2365 dse_record_singleton_alias_set (alias_set_type alias_set
,
2366 enum machine_mode mode
)
2368 struct clear_alias_mode_holder tmp_holder
;
2369 struct clear_alias_mode_holder
*entry
;
2372 /* If we are not going to run dse, we need to return now or there
2373 will be problems with allocating the bitmaps. */
2374 if ((!gate_dse()) || !alias_set
)
2377 if (!clear_alias_sets
)
2379 clear_alias_sets
= BITMAP_ALLOC (NULL
);
2380 disqualified_clear_alias_sets
= BITMAP_ALLOC (NULL
);
2381 clear_alias_mode_table
= htab_create (11, clear_alias_mode_hash
,
2382 clear_alias_mode_eq
, NULL
);
2383 clear_alias_mode_pool
= create_alloc_pool ("clear_alias_mode_pool",
2384 sizeof (struct clear_alias_mode_holder
), 100);
2387 bitmap_set_bit (clear_alias_sets
, alias_set
);
2389 tmp_holder
.alias_set
= alias_set
;
2391 slot
= htab_find_slot (clear_alias_mode_table
, &tmp_holder
, INSERT
);
2392 gcc_assert (*slot
== NULL
);
2394 *slot
= entry
= pool_alloc (clear_alias_mode_pool
);
2395 entry
->alias_set
= alias_set
;
2400 /* Remove ALIAS_SET from the sets of stack slots being considered. */
2403 dse_invalidate_singleton_alias_set (alias_set_type alias_set
)
2405 if ((!gate_dse()) || !alias_set
)
2408 bitmap_clear_bit (clear_alias_sets
, alias_set
);
2412 /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not
2416 get_bitmap_index (group_info_t group_info
, HOST_WIDE_INT offset
)
2420 HOST_WIDE_INT offset_p
= -offset
;
2421 if (offset_p
>= group_info
->offset_map_size_n
)
2423 return group_info
->offset_map_n
[offset_p
];
2427 if (offset
>= group_info
->offset_map_size_p
)
2429 return group_info
->offset_map_p
[offset
];
2434 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL
2438 scan_stores_nospill (store_info_t store_info
, bitmap gen
, bitmap kill
)
2443 group_info_t group_info
2444 = VEC_index (group_info_t
, rtx_group_vec
, store_info
->group_id
);
2445 if (group_info
->process_globally
)
2446 for (i
= store_info
->begin
; i
< store_info
->end
; i
++)
2448 int index
= get_bitmap_index (group_info
, i
);
2451 bitmap_set_bit (gen
, index
);
2453 bitmap_clear_bit (kill
, index
);
2456 store_info
= store_info
->next
;
2461 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL
2465 scan_stores_spill (store_info_t store_info
, bitmap gen
, bitmap kill
)
2469 if (store_info
->alias_set
)
2471 int index
= get_bitmap_index (clear_alias_group
,
2472 store_info
->alias_set
);
2475 bitmap_set_bit (gen
, index
);
2477 bitmap_clear_bit (kill
, index
);
2480 store_info
= store_info
->next
;
2485 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL
2489 scan_reads_nospill (insn_info_t insn_info
, bitmap gen
, bitmap kill
)
2491 read_info_t read_info
= insn_info
->read_rec
;
2495 /* For const function calls kill the stack related stores. */
2496 if (insn_info
->stack_read
)
2498 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2499 if (group
->process_globally
&& group
->frame_related
)
2502 bitmap_ior_into (kill
, group
->group_kill
);
2503 bitmap_and_compl_into (gen
, group
->group_kill
);
2509 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2511 if (group
->process_globally
)
2513 if (i
== read_info
->group_id
)
2515 if (read_info
->begin
> read_info
->end
)
2517 /* Begin > end for block mode reads. */
2519 bitmap_ior_into (kill
, group
->group_kill
);
2520 bitmap_and_compl_into (gen
, group
->group_kill
);
2524 /* The groups are the same, just process the
2527 for (j
= read_info
->begin
; j
< read_info
->end
; j
++)
2529 int index
= get_bitmap_index (group
, j
);
2533 bitmap_set_bit (kill
, index
);
2534 bitmap_clear_bit (gen
, index
);
2541 /* The groups are different, if the alias sets
2542 conflict, clear the entire group. We only need
2543 to apply this test if the read_info is a cselib
2544 read. Anything with a constant base cannot alias
2545 something else with a different constant
2547 if ((read_info
->group_id
< 0)
2548 && canon_true_dependence (group
->base_mem
,
2550 group
->canon_base_mem
,
2551 read_info
->mem
, rtx_varies_p
))
2554 bitmap_ior_into (kill
, group
->group_kill
);
2555 bitmap_and_compl_into (gen
, group
->group_kill
);
2561 read_info
= read_info
->next
;
2565 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL
2569 scan_reads_spill (read_info_t read_info
, bitmap gen
, bitmap kill
)
2573 if (read_info
->alias_set
)
2575 int index
= get_bitmap_index (clear_alias_group
,
2576 read_info
->alias_set
);
2580 bitmap_set_bit (kill
, index
);
2581 bitmap_clear_bit (gen
, index
);
2585 read_info
= read_info
->next
;
2590 /* Return the insn in BB_INFO before the first wild read or if there
2591 are no wild reads in the block, return the last insn. */
2594 find_insn_before_first_wild_read (bb_info_t bb_info
)
2596 insn_info_t insn_info
= bb_info
->last_insn
;
2597 insn_info_t last_wild_read
= NULL
;
2601 if (insn_info
->wild_read
)
2603 last_wild_read
= insn_info
->prev_insn
;
2604 /* Block starts with wild read. */
2605 if (!last_wild_read
)
2609 insn_info
= insn_info
->prev_insn
;
2613 return last_wild_read
;
2615 return bb_info
->last_insn
;
2619 /* Scan the insns in BB_INFO starting at PTR and going to the top of
2620 the block in order to build the gen and kill sets for the block.
2621 We start at ptr which may be the last insn in the block or may be
2622 the first insn with a wild read. In the latter case we are able to
2623 skip the rest of the block because it just does not matter:
2624 anything that happens is hidden by the wild read. */
2627 dse_step3_scan (bool for_spills
, basic_block bb
)
2629 bb_info_t bb_info
= bb_table
[bb
->index
];
2630 insn_info_t insn_info
;
2633 /* There are no wild reads in the spill case. */
2634 insn_info
= bb_info
->last_insn
;
2636 insn_info
= find_insn_before_first_wild_read (bb_info
);
2638 /* In the spill case or in the no_spill case if there is no wild
2639 read in the block, we will need a kill set. */
2640 if (insn_info
== bb_info
->last_insn
)
2643 bitmap_clear (bb_info
->kill
);
2645 bb_info
->kill
= BITMAP_ALLOC (NULL
);
2649 BITMAP_FREE (bb_info
->kill
);
2653 /* There may have been code deleted by the dce pass run before
2655 if (insn_info
->insn
&& INSN_P (insn_info
->insn
))
2657 /* Process the read(s) last. */
2660 scan_stores_spill (insn_info
->store_rec
, bb_info
->gen
, bb_info
->kill
);
2661 scan_reads_spill (insn_info
->read_rec
, bb_info
->gen
, bb_info
->kill
);
2665 scan_stores_nospill (insn_info
->store_rec
, bb_info
->gen
, bb_info
->kill
);
2666 scan_reads_nospill (insn_info
, bb_info
->gen
, bb_info
->kill
);
2670 insn_info
= insn_info
->prev_insn
;
2675 /* Set the gen set of the exit block, and also any block with no
2676 successors that does not have a wild read. */
2679 dse_step3_exit_block_scan (bb_info_t bb_info
)
2681 /* The gen set is all 0's for the exit block except for the
2682 frame_pointer_group. */
2684 if (stores_off_frame_dead_at_return
)
2689 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2691 if (group
->process_globally
&& group
->frame_related
)
2692 bitmap_ior_into (bb_info
->gen
, group
->group_kill
);
2698 /* Find all of the blocks that are not backwards reachable from the
2699 exit block or any block with no successors (BB). These are the
2700 infinite loops or infinite self loops. These blocks will still
2701 have their bits set in UNREACHABLE_BLOCKS. */
2704 mark_reachable_blocks (sbitmap unreachable_blocks
, basic_block bb
)
2709 if (TEST_BIT (unreachable_blocks
, bb
->index
))
2711 RESET_BIT (unreachable_blocks
, bb
->index
);
2712 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2714 mark_reachable_blocks (unreachable_blocks
, e
->src
);
2719 /* Build the transfer functions for the function. */
2722 dse_step3 (bool for_spills
)
2725 sbitmap unreachable_blocks
= sbitmap_alloc (last_basic_block
);
2726 sbitmap_iterator sbi
;
2727 bitmap all_ones
= NULL
;
2730 sbitmap_ones (unreachable_blocks
);
2734 bb_info_t bb_info
= bb_table
[bb
->index
];
2736 bitmap_clear (bb_info
->gen
);
2738 bb_info
->gen
= BITMAP_ALLOC (NULL
);
2740 if (bb
->index
== ENTRY_BLOCK
)
2742 else if (bb
->index
== EXIT_BLOCK
)
2743 dse_step3_exit_block_scan (bb_info
);
2745 dse_step3_scan (for_spills
, bb
);
2746 if (EDGE_COUNT (bb
->succs
) == 0)
2747 mark_reachable_blocks (unreachable_blocks
, bb
);
2749 /* If this is the second time dataflow is run, delete the old
2752 BITMAP_FREE (bb_info
->in
);
2754 BITMAP_FREE (bb_info
->out
);
2757 /* For any block in an infinite loop, we must initialize the out set
2758 to all ones. This could be expensive, but almost never occurs in
2759 practice. However, it is common in regression tests. */
2760 EXECUTE_IF_SET_IN_SBITMAP (unreachable_blocks
, 0, i
, sbi
)
2762 if (bitmap_bit_p (all_blocks
, i
))
2764 bb_info_t bb_info
= bb_table
[i
];
2770 all_ones
= BITMAP_ALLOC (NULL
);
2771 for (j
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, j
, group
); j
++)
2772 bitmap_ior_into (all_ones
, group
->group_kill
);
2776 bb_info
->out
= BITMAP_ALLOC (NULL
);
2777 bitmap_copy (bb_info
->out
, all_ones
);
2783 BITMAP_FREE (all_ones
);
2784 sbitmap_free (unreachable_blocks
);
2789 /*----------------------------------------------------------------------------
2792 Solve the bitvector equations.
2793 ----------------------------------------------------------------------------*/
2796 /* Confluence function for blocks with no successors. Create an out
2797 set from the gen set of the exit block. This block logically has
2798 the exit block as a successor. */
2803 dse_confluence_0 (basic_block bb
)
2805 bb_info_t bb_info
= bb_table
[bb
->index
];
2807 if (bb
->index
== EXIT_BLOCK
)
2812 bb_info
->out
= BITMAP_ALLOC (NULL
);
2813 bitmap_copy (bb_info
->out
, bb_table
[EXIT_BLOCK
]->gen
);
2817 /* Propagate the information from the in set of the dest of E to the
2818 out set of the src of E. If the various in or out sets are not
2819 there, that means they are all ones. */
2822 dse_confluence_n (edge e
)
2824 bb_info_t src_info
= bb_table
[e
->src
->index
];
2825 bb_info_t dest_info
= bb_table
[e
->dest
->index
];
2830 bitmap_and_into (src_info
->out
, dest_info
->in
);
2833 src_info
->out
= BITMAP_ALLOC (NULL
);
2834 bitmap_copy (src_info
->out
, dest_info
->in
);
2840 /* Propagate the info from the out to the in set of BB_INDEX's basic
2841 block. There are three cases:
2843 1) The block has no kill set. In this case the kill set is all
2844 ones. It does not matter what the out set of the block is, none of
2845 the info can reach the top. The only thing that reaches the top is
2846 the gen set and we just copy the set.
2848 2) There is a kill set but no out set and bb has successors. In
2849 this case we just return. Eventually an out set will be created and
2850 it is better to wait than to create a set of ones.
2852 3) There is both a kill and out set. We apply the obvious transfer
2857 dse_transfer_function (int bb_index
)
2859 bb_info_t bb_info
= bb_table
[bb_index
];
2867 return bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
2868 bb_info
->out
, bb_info
->kill
);
2871 bb_info
->in
= BITMAP_ALLOC (NULL
);
2872 bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
2873 bb_info
->out
, bb_info
->kill
);
2883 /* Case 1 above. If there is already an in set, nothing
2889 bb_info
->in
= BITMAP_ALLOC (NULL
);
2890 bitmap_copy (bb_info
->in
, bb_info
->gen
);
2896 /* Solve the dataflow equations. */
2901 df_simple_dataflow (DF_BACKWARD
, NULL
, dse_confluence_0
,
2902 dse_confluence_n
, dse_transfer_function
,
2903 all_blocks
, df_get_postorder (DF_BACKWARD
),
2904 df_get_n_blocks (DF_BACKWARD
));
2909 fprintf (dump_file
, "\n\n*** Global dataflow info after analysis.\n");
2912 bb_info_t bb_info
= bb_table
[bb
->index
];
2914 df_print_bb_index (bb
, dump_file
);
2916 bitmap_print (dump_file
, bb_info
->in
, " in: ", "\n");
2918 fprintf (dump_file
, " in: *MISSING*\n");
2920 bitmap_print (dump_file
, bb_info
->gen
, " gen: ", "\n");
2922 fprintf (dump_file
, " gen: *MISSING*\n");
2924 bitmap_print (dump_file
, bb_info
->kill
, " kill: ", "\n");
2926 fprintf (dump_file
, " kill: *MISSING*\n");
2928 bitmap_print (dump_file
, bb_info
->out
, " out: ", "\n");
2930 fprintf (dump_file
, " out: *MISSING*\n\n");
2937 /*----------------------------------------------------------------------------
2940 Delete the stores that can only be deleted using the global information.
2941 ----------------------------------------------------------------------------*/
2945 dse_step5_nospill (void)
2950 bb_info_t bb_info
= bb_table
[bb
->index
];
2951 insn_info_t insn_info
= bb_info
->last_insn
;
2952 bitmap v
= bb_info
->out
;
2956 bool deleted
= false;
2957 if (dump_file
&& insn_info
->insn
)
2959 fprintf (dump_file
, "starting to process insn %d\n",
2960 INSN_UID (insn_info
->insn
));
2961 bitmap_print (dump_file
, v
, " v: ", "\n");
2964 /* There may have been code deleted by the dce pass run before
2967 && INSN_P (insn_info
->insn
)
2968 && (!insn_info
->cannot_delete
)
2969 && (!bitmap_empty_p (v
)))
2971 store_info_t store_info
= insn_info
->store_rec
;
2973 /* Try to delete the current insn. */
2976 /* Skip the clobbers. */
2977 while (!store_info
->is_set
)
2978 store_info
= store_info
->next
;
2980 if (store_info
->alias_set
)
2985 group_info_t group_info
2986 = VEC_index (group_info_t
, rtx_group_vec
, store_info
->group_id
);
2988 for (i
= store_info
->begin
; i
< store_info
->end
; i
++)
2990 int index
= get_bitmap_index (group_info
, i
);
2993 fprintf (dump_file
, "i = %d, index = %d\n", (int)i
, index
);
2994 if (index
== 0 || !bitmap_bit_p (v
, index
))
2997 fprintf (dump_file
, "failing at i = %d\n", (int)i
);
3007 check_for_inc_dec (insn_info
->insn
);
3008 delete_insn (insn_info
->insn
);
3009 insn_info
->insn
= NULL
;
3014 /* We do want to process the local info if the insn was
3015 deleted. For instance, if the insn did a wild read, we
3016 no longer need to trash the info. */
3018 && INSN_P (insn_info
->insn
)
3021 scan_stores_nospill (insn_info
->store_rec
, v
, NULL
);
3022 if (insn_info
->wild_read
)
3025 fprintf (dump_file
, "wild read\n");
3028 else if (insn_info
->read_rec
)
3031 fprintf (dump_file
, "regular read\n");
3032 scan_reads_nospill (insn_info
, v
, NULL
);
3036 insn_info
= insn_info
->prev_insn
;
3043 dse_step5_spill (void)
3048 bb_info_t bb_info
= bb_table
[bb
->index
];
3049 insn_info_t insn_info
= bb_info
->last_insn
;
3050 bitmap v
= bb_info
->out
;
3054 bool deleted
= false;
3055 /* There may have been code deleted by the dce pass run before
3058 && INSN_P (insn_info
->insn
)
3059 && (!insn_info
->cannot_delete
)
3060 && (!bitmap_empty_p (v
)))
3062 /* Try to delete the current insn. */
3063 store_info_t store_info
= insn_info
->store_rec
;
3068 if (store_info
->alias_set
)
3070 int index
= get_bitmap_index (clear_alias_group
,
3071 store_info
->alias_set
);
3072 if (index
== 0 || !bitmap_bit_p (v
, index
))
3080 store_info
= store_info
->next
;
3082 if (deleted
&& dbg_cnt (dse
))
3085 fprintf (dump_file
, "Spill deleting insn %d\n",
3086 INSN_UID (insn_info
->insn
));
3087 check_for_inc_dec (insn_info
->insn
);
3088 delete_insn (insn_info
->insn
);
3090 insn_info
->insn
= NULL
;
3095 && INSN_P (insn_info
->insn
)
3098 scan_stores_spill (insn_info
->store_rec
, v
, NULL
);
3099 scan_reads_spill (insn_info
->read_rec
, v
, NULL
);
3102 insn_info
= insn_info
->prev_insn
;
3109 /*----------------------------------------------------------------------------
3112 Destroy everything left standing.
3113 ----------------------------------------------------------------------------*/
3116 dse_step6 (bool global_done
)
3124 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
3126 free (group
->offset_map_n
);
3127 free (group
->offset_map_p
);
3128 BITMAP_FREE (group
->store1_n
);
3129 BITMAP_FREE (group
->store1_p
);
3130 BITMAP_FREE (group
->store2_n
);
3131 BITMAP_FREE (group
->store2_p
);
3132 BITMAP_FREE (group
->group_kill
);
3137 bb_info_t bb_info
= bb_table
[bb
->index
];
3138 BITMAP_FREE (bb_info
->gen
);
3140 BITMAP_FREE (bb_info
->kill
);
3142 BITMAP_FREE (bb_info
->in
);
3144 BITMAP_FREE (bb_info
->out
);
3149 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
3151 BITMAP_FREE (group
->store1_n
);
3152 BITMAP_FREE (group
->store1_p
);
3153 BITMAP_FREE (group
->store2_n
);
3154 BITMAP_FREE (group
->store2_p
);
3155 BITMAP_FREE (group
->group_kill
);
3159 if (clear_alias_sets
)
3161 BITMAP_FREE (clear_alias_sets
);
3162 BITMAP_FREE (disqualified_clear_alias_sets
);
3163 free_alloc_pool (clear_alias_mode_pool
);
3164 htab_delete (clear_alias_mode_table
);
3167 end_alias_analysis ();
3169 htab_delete (rtx_group_table
);
3170 VEC_free (group_info_t
, heap
, rtx_group_vec
);
3171 BITMAP_FREE (all_blocks
);
3172 BITMAP_FREE (scratch
);
3174 free_alloc_pool (rtx_store_info_pool
);
3175 free_alloc_pool (read_info_pool
);
3176 free_alloc_pool (insn_info_pool
);
3177 free_alloc_pool (bb_info_pool
);
3178 free_alloc_pool (rtx_group_info_pool
);
3179 free_alloc_pool (deferred_change_pool
);
3184 /* -------------------------------------------------------------------------
3186 ------------------------------------------------------------------------- */
3188 /* Callback for running pass_rtl_dse. */
3191 rest_of_handle_dse (void)
3193 bool did_global
= false;
3195 df_set_flags (DF_DEFER_INSN_RESCAN
);
3200 if (dse_step2_nospill ())
3202 df_set_flags (DF_LR_RUN_DCE
);
3206 fprintf (dump_file
, "doing global processing\n");
3209 dse_step5_nospill ();
3212 /* For the instance of dse that runs after reload, we make a special
3213 pass to process the spills. These are special in that they are
3214 totally transparent, i.e, there is no aliasing issues that need
3215 to be considered. This means that the wild reads that kill
3216 everything else do not apply here. */
3217 if (clear_alias_sets
&& dse_step2_spill ())
3221 df_set_flags (DF_LR_RUN_DCE
);
3226 fprintf (dump_file
, "doing global spill processing\n");
3232 dse_step6 (did_global
);
3235 fprintf (dump_file
, "dse: local deletions = %d, global deletions = %d, spill deletions = %d\n",
3236 locally_deleted
, globally_deleted
, spill_deleted
);
3243 return optimize
> 0 && flag_dse
;
3246 struct tree_opt_pass pass_rtl_dse1
=
3249 gate_dse
, /* gate */
3250 rest_of_handle_dse
, /* execute */
3253 0, /* static_pass_number */
3254 TV_DSE1
, /* tv_id */
3255 0, /* properties_required */
3256 0, /* properties_provided */
3257 0, /* properties_destroyed */
3258 0, /* todo_flags_start */
3260 TODO_df_finish
| TODO_verify_rtl_sharing
|
3261 TODO_ggc_collect
, /* todo_flags_finish */
3265 struct tree_opt_pass pass_rtl_dse2
=
3268 gate_dse
, /* gate */
3269 rest_of_handle_dse
, /* execute */
3272 0, /* static_pass_number */
3273 TV_DSE2
, /* tv_id */
3274 0, /* properties_required */
3275 0, /* properties_provided */
3276 0, /* properties_destroyed */
3277 0, /* todo_flags_start */
3279 TODO_df_finish
| TODO_verify_rtl_sharing
|
3280 TODO_ggc_collect
, /* todo_flags_finish */