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"
34 #include "hard-reg-set.h"
39 #include "tree-pass.h"
40 #include "alloc-pool.h"
42 #include "insn-config.h"
49 /* This file contains three techniques for performing Dead Store
52 * The first technique performs dse locally on any base address. It
53 is based on the cselib which is a local value numbering technique.
54 This technique is local to a basic block but deals with a fairly
57 * The second technique performs dse globally but is restricted to
58 base addresses that are either constant or are relative to the
61 * The third technique, (which is only done after register allocation)
62 processes the spill spill slots. This differs from the second
63 technique because it takes advantage of the fact that spilling is
64 completely free from the effects of aliasing.
66 Logically, dse is a backwards dataflow problem. A store can be
67 deleted if it if cannot be reached in the backward direction by any
68 use of the value being stored. However, the local technique uses a
69 forwards scan of the basic block because cselib requires that the
70 block be processed in that order.
72 The pass is logically broken into 7 steps:
76 1) The local algorithm, as well as scanning the insns for the two
79 2) Analysis to see if the global algs are necessary. In the case
80 of stores base on a constant address, there must be at least two
81 stores to that address, to make it possible to delete some of the
82 stores. In the case of stores off of the frame or spill related
83 stores, only one store to an address is necessary because those
84 stores die at the end of the function.
86 3) Set up the global dataflow equations based on processing the
87 info parsed in the first step.
89 4) Solve the dataflow equations.
91 5) Delete the insns that the global analysis has indicated are
96 This step uses cselib and canon_rtx to build the largest expression
97 possible for each address. This pass is a forwards pass through
98 each basic block. From the point of view of the global technique,
99 the first pass could examine a block in either direction. The
100 forwards ordering is to accommodate cselib.
102 We a simplifying assumption: addresses fall into four broad
105 1) base has rtx_varies_p == false, offset is constant.
106 2) base has rtx_varies_p == false, offset variable.
107 3) base has rtx_varies_p == true, offset constant.
108 4) base has rtx_varies_p == true, offset variable.
110 The local passes are able to process all 4 kinds of addresses. The
111 global pass only handles (1).
113 The global problem is formulated as follows:
115 A store, S1, to address A, where A is not relative to the stack
116 frame, can be eliminated if all paths from S1 to the end of the
117 of the function contain another store to A before a read to A.
119 If the address A is relative to the stack frame, a store S2 to A
120 can be eliminated if there are no paths from S1 that reach the
121 end of the function that read A before another store to A. In
122 this case S2 can be deleted if there are paths to from S2 to the
123 end of the function that have no reads or writes to A. This
124 second case allows stores to the stack frame to be deleted that
125 would otherwise die when the function returns. This cannot be
126 done if stores_off_frame_dead_at_return is not true. See the doc
127 for that variable for when this variable is false.
129 The global problem is formulated as a backwards set union
130 dataflow problem where the stores are the gens and reads are the
131 kills. Set union problems are rare and require some special
132 handling given our representation of bitmaps. A straightforward
133 implementation of requires a lot of bitmaps filled with 1s.
134 These are expensive and cumbersome in our bitmap formulation so
135 care has been taken to avoid large vectors filled with 1s. See
136 the comments in bb_info and in the dataflow confluence functions
139 There are two places for further enhancements to this algorithm:
141 1) The original dse which was embedded in a pass called flow also
142 did local address forwarding. For example in
147 flow would replace the right hand side of the second insn with a
148 reference to r100. Most of the information is available to add this
149 to this pass. It has not done it because it is a lot of work in
150 the case that either r100 is assigned to between the first and
151 second insn and/or the second insn is a load of part of the value
152 stored by the first insn.
154 insn 5 in gcc.c-torture/compile/990203-1.c simple case.
155 insn 15 in gcc.c-torture/execute/20001017-2.c simple case.
156 insn 25 in gcc.c-torture/execute/20001026-1.c simple case.
157 insn 44 in gcc.c-torture/execute/20010910-1.c simple case.
159 2) The cleaning up of spill code is quite profitable. It currently
160 depends on reading tea leaves and chicken entrails left by reload.
161 This pass depends on reload creating a singleton alias set for each
162 spill slot and telling the next dse pass which of these alias sets
163 are the singletons. Rather than analyze the addresses of the
164 spills, dse's spill processing just does analysis of the loads and
165 stores that use those alias sets. There are three cases where this
168 a) Reload sometimes creates the slot for one mode of access, and
169 then inserts loads and/or stores for a smaller mode. In this
170 case, the current code just punts on the slot. The proper thing
171 to do is to back out and use one bit vector position for each
172 byte of the entity associated with the slot. This depends on
173 KNOWING that reload always generates the accesses for each of the
174 bytes in some canonical (read that easy to understand several
175 passes after reload happens) way.
177 b) Reload sometimes decides that spill slot it allocated was not
178 large enough for the mode and goes back and allocates more slots
179 with the same mode and alias set. The backout in this case is a
180 little more graceful than (a). In this case the slot is unmarked
181 as being a spill slot and if final address comes out to be based
182 off the frame pointer, the global algorithm handles this slot.
184 c) For any pass that may prespill, there is currently no
185 mechanism to tell the dse pass that the slot being used has the
186 special properties that reload uses. It may be that all that is
187 required is to have those passes make the same calls that reload
188 does, assuming that the alias sets can be manipulated in the same
191 /* There are limits to the size of constant offsets we model for the
192 global problem. There are certainly test cases, that exceed this
193 limit, however, it is unlikely that there are important programs
194 that really have constant offsets this size. */
195 #define MAX_OFFSET (64 * 1024)
198 static bitmap scratch
= NULL
;
201 /* This structure holds information about a candidate store. */
205 /* False means this is a clobber. */
208 /* The id of the mem group of the base address. If rtx_varies_p is
209 true, this is -1. Otherwise, it is the index into the group
213 /* This is the cselib value. */
214 cselib_val
*cse_base
;
216 /* This canonized mem. */
219 /* The result of get_addr on mem. */
222 /* If this is non-zero, it is the alias set of a spill location. */
223 alias_set_type alias_set
;
225 /* The offset of the first and byte before the last byte associated
226 with the operation. */
229 /* An bitmask as wide as the number of bytes in the word that
230 contains a 1 if the byte may be needed. The store is unused if
231 all of the bits are 0. */
232 unsigned HOST_WIDE_INT positions_needed
;
234 /* The next store info for this insn. */
235 struct store_info
*next
;
237 /* The right hand side of the store. This is used if there is a
238 subsequent reload of the mems address somewhere later in the
243 /* Return a bitmask with the first N low bits set. */
245 static unsigned HOST_WIDE_INT
246 lowpart_bitmask (int n
)
248 unsigned HOST_WIDE_INT mask
= ~(unsigned HOST_WIDE_INT
) 0;
249 return mask
>> (HOST_BITS_PER_WIDE_INT
- n
);
252 typedef struct store_info
*store_info_t
;
253 static alloc_pool cse_store_info_pool
;
254 static alloc_pool rtx_store_info_pool
;
256 /* This structure holds information about a load. These are only
257 built for rtx bases. */
260 /* The id of the mem group of the base address. */
263 /* If this is non-zero, it is the alias set of a spill location. */
264 alias_set_type alias_set
;
266 /* The offset of the first and byte after the last byte associated
267 with the operation. If begin == end == 0, the read did not have
268 a constant offset. */
271 /* The mem being read. */
274 /* The next read_info for this insn. */
275 struct read_info
*next
;
277 typedef struct read_info
*read_info_t
;
278 static alloc_pool read_info_pool
;
281 /* One of these records is created for each insn. */
285 /* Set true if the insn contains a store but the insn itself cannot
286 be deleted. This is set if the insn is a parallel and there is
287 more than one non dead output or if the insn is in some way
291 /* This field is only used by the global algorithm. It is set true
292 if the insn contains any read of mem except for a (1). This is
293 also set if the insn is a call or has a clobber mem. If the insn
294 contains a wild read, the use_rec will be null. */
297 /* This field is only used for the processing of const functions.
298 These functions cannot read memory, but they can read the stack
299 because that is where they may get their parms. We need to be
300 this conservative because, like the store motion pass, we don't
301 consider CALL_INSN_FUNCTION_USAGE when processing call insns.
302 Moreover, we need to distinguish two cases:
303 1. Before reload (register elimination), the stores related to
304 outgoing arguments are stack pointer based and thus deemed
305 of non-constant base in this pass. This requires special
306 handling but also means that the frame pointer based stores
307 need not be killed upon encountering a const function call.
308 2. After reload, the stores related to outgoing arguments can be
309 either stack pointer or hard frame pointer based. This means
310 that we have no other choice than also killing all the frame
311 pointer based stores upon encountering a const function call.
312 This field is set after reload for const function calls. Having
313 this set is less severe than a wild read, it just means that all
314 the frame related stores are killed rather than all the stores. */
317 /* This field is only used for the processing of const functions.
318 It is set if the insn may contain a stack pointer based store. */
319 bool stack_pointer_based
;
321 /* This is true if any of the sets within the store contains a
322 cselib base. Such stores can only be deleted by the local
324 bool contains_cselib_groups
;
329 /* The list of mem sets or mem clobbers that are contained in this
330 insn. If the insn is deletable, it contains only one mem set.
331 But it could also contain clobbers. Insns that contain more than
332 one mem set are not deletable, but each of those mems are here in
333 order to provide info to delete other insns. */
334 store_info_t store_rec
;
336 /* The linked list of mem uses in this insn. Only the reads from
337 rtx bases are listed here. The reads to cselib bases are
338 completely processed during the first scan and so are never
340 read_info_t read_rec
;
342 /* The prev insn in the basic block. */
343 struct insn_info
* prev_insn
;
345 /* The linked list of insns that are in consideration for removal in
346 the forwards pass thru the basic block. This pointer may be
347 trash as it is not cleared when a wild read occurs. The only
348 time it is guaranteed to be correct is when the traveral starts
349 at active_local_stores. */
350 struct insn_info
* next_local_store
;
353 typedef struct insn_info
*insn_info_t
;
354 static alloc_pool insn_info_pool
;
356 /* The linked list of stores that are under consideration in this
358 static insn_info_t active_local_stores
;
363 /* Pointer to the insn info for the last insn in the block. These
364 are linked so this is how all of the insns are reached. During
365 scanning this is the current insn being scanned. */
366 insn_info_t last_insn
;
368 /* The info for the global dataflow problem. */
371 /* This is set if the transfer function should and in the wild_read
372 bitmap before applying the kill and gen sets. That vector knocks
373 out most of the bits in the bitmap and thus speeds up the
375 bool apply_wild_read
;
377 /* The set of store positions that exist in this block before a wild read. */
380 /* The set of load positions that exist in this block above the
381 same position of a store. */
384 /* The set of stores that reach the top of the block without being
387 Do not represent the in if it is all ones. Note that this is
388 what the bitvector should logically be initialized to for a set
389 intersection problem. However, like the kill set, this is too
390 expensive. So initially, the in set will only be created for the
391 exit block and any block that contains a wild read. */
394 /* The set of stores that reach the bottom of the block from it's
397 Do not represent the in if it is all ones. Note that this is
398 what the bitvector should logically be initialized to for a set
399 intersection problem. However, like the kill and in set, this is
400 too expensive. So what is done is that the confluence operator
401 just initializes the vector from one of the out sets of the
402 successors of the block. */
406 typedef struct bb_info
*bb_info_t
;
407 static alloc_pool bb_info_pool
;
409 /* Table to hold all bb_infos. */
410 static bb_info_t
*bb_table
;
412 /* There is a group_info for each rtx base that is used to reference
413 memory. There are also not many of the rtx bases because they are
414 very limited in scope. */
418 /* The actual base of the address. */
421 /* The sequential id of the base. This allows us to have a
422 canonical ordering of these that is not based on addresses. */
425 /* A mem wrapped around the base pointer for the group in order to
426 do read dependency. */
429 /* Canonized version of base_mem, most likely the same thing. */
432 /* These two sets of two bitmaps are used to keep track of how many
433 stores are actually referencing that position from this base. We
434 only do this for rtx bases as this will be used to assign
435 positions in the bitmaps for the global problem. Bit N is set in
436 store1 on the first store for offset N. Bit N is set in store2
437 for the second store to offset N. This is all we need since we
438 only care about offsets that have two or more stores for them.
440 The "_n" suffix is for offsets less than 0 and the "_p" suffix is
441 for 0 and greater offsets.
443 There is one special case here, for stores into the stack frame,
444 we will or store1 into store2 before deciding which stores look
445 at globally. This is because stores to the stack frame that have
446 no other reads before the end of the function can also be
448 bitmap store1_n
, store1_p
, store2_n
, store2_p
;
450 /* The positions in this bitmap have the same assignments as the in,
451 out, gen and kill bitmaps. This bitmap is all zeros except for
452 the positions that are occupied by stores for this group. */
455 /* True if there are any positions that are to be processed
457 bool process_globally
;
459 /* True if the base of this group is either the frame_pointer or
460 hard_frame_pointer. */
463 /* The offset_map is used to map the offsets from this base into
464 positions in the global bitmaps. It is only created after all of
465 the all of stores have been scanned and we know which ones we
467 int *offset_map_n
, *offset_map_p
;
468 int offset_map_size_n
, offset_map_size_p
;
470 typedef struct group_info
*group_info_t
;
471 typedef const struct group_info
*const_group_info_t
;
472 static alloc_pool rtx_group_info_pool
;
474 /* Tables of group_info structures, hashed by base value. */
475 static htab_t rtx_group_table
;
477 /* Index into the rtx_group_vec. */
478 static int rtx_group_next_id
;
480 DEF_VEC_P(group_info_t
);
481 DEF_VEC_ALLOC_P(group_info_t
,heap
);
483 static VEC(group_info_t
,heap
) *rtx_group_vec
;
486 /* This structure holds the set of changes that are being deferred
487 when removing read operation. See replace_read. */
488 struct deferred_change
491 /* The mem that is being replaced. */
494 /* The reg it is being replaced with. */
497 struct deferred_change
*next
;
500 typedef struct deferred_change
*deferred_change_t
;
501 static alloc_pool deferred_change_pool
;
503 static deferred_change_t deferred_change_list
= NULL
;
505 /* This are used to hold the alias sets of spill variables. Since
506 these are never aliased and there may be a lot of them, it makes
507 sense to treat them specially. This bitvector is only allocated in
508 calls from dse_record_singleton_alias_set which currently is only
509 made during reload1. So when dse is called before reload this
510 mechanism does nothing. */
512 static bitmap clear_alias_sets
= NULL
;
514 /* The set of clear_alias_sets that have been disqualified because
515 there are loads or stores using a different mode than the alias set
516 was registered with. */
517 static bitmap disqualified_clear_alias_sets
= NULL
;
519 /* The group that holds all of the clear_alias_sets. */
520 static group_info_t clear_alias_group
;
522 /* The modes of the clear_alias_sets. */
523 static htab_t clear_alias_mode_table
;
525 /* Hash table element to look up the mode for an alias set. */
526 struct clear_alias_mode_holder
528 alias_set_type alias_set
;
529 enum machine_mode mode
;
532 static alloc_pool clear_alias_mode_pool
;
534 /* This is true except if current_function_stdarg -- i.e. we cannot do
535 this for vararg functions because they play games with the frame. */
536 static bool stores_off_frame_dead_at_return
;
538 /* Counter for stats. */
539 static int globally_deleted
;
540 static int locally_deleted
;
541 static int spill_deleted
;
543 static bitmap all_blocks
;
545 /* The number of bits used in the global bitmaps. */
546 static unsigned int current_position
;
549 static bool gate_dse (void);
550 static bool gate_dse1 (void);
551 static bool gate_dse2 (void);
554 /*----------------------------------------------------------------------------
558 ----------------------------------------------------------------------------*/
560 /* Hashtable callbacks for maintaining the "bases" field of
561 store_group_info, given that the addresses are function invariants. */
564 clear_alias_mode_eq (const void *p1
, const void *p2
)
566 const struct clear_alias_mode_holder
* h1
567 = (const struct clear_alias_mode_holder
*) p1
;
568 const struct clear_alias_mode_holder
* h2
569 = (const struct clear_alias_mode_holder
*) p2
;
570 return h1
->alias_set
== h2
->alias_set
;
575 clear_alias_mode_hash (const void *p
)
577 const struct clear_alias_mode_holder
*holder
578 = (const struct clear_alias_mode_holder
*) p
;
579 return holder
->alias_set
;
583 /* Find the entry associated with ALIAS_SET. */
585 static struct clear_alias_mode_holder
*
586 clear_alias_set_lookup (alias_set_type alias_set
)
588 struct clear_alias_mode_holder tmp_holder
;
591 tmp_holder
.alias_set
= alias_set
;
592 slot
= htab_find_slot (clear_alias_mode_table
, &tmp_holder
, NO_INSERT
);
599 /* Hashtable callbacks for maintaining the "bases" field of
600 store_group_info, given that the addresses are function invariants. */
603 invariant_group_base_eq (const void *p1
, const void *p2
)
605 const_group_info_t gi1
= (const_group_info_t
) p1
;
606 const_group_info_t gi2
= (const_group_info_t
) p2
;
607 return rtx_equal_p (gi1
->rtx_base
, gi2
->rtx_base
);
612 invariant_group_base_hash (const void *p
)
614 const_group_info_t gi
= (const_group_info_t
) p
;
616 return hash_rtx (gi
->rtx_base
, Pmode
, &do_not_record
, NULL
, false);
620 /* Get the GROUP for BASE. Add a new group if it is not there. */
623 get_group_info (rtx base
)
625 struct group_info tmp_gi
;
631 /* Find the store_base_info structure for BASE, creating a new one
633 tmp_gi
.rtx_base
= base
;
634 slot
= htab_find_slot (rtx_group_table
, &tmp_gi
, INSERT
);
635 gi
= (group_info_t
) *slot
;
639 if (!clear_alias_group
)
641 clear_alias_group
= gi
= pool_alloc (rtx_group_info_pool
);
642 memset (gi
, 0, sizeof (struct group_info
));
643 gi
->id
= rtx_group_next_id
++;
644 gi
->store1_n
= BITMAP_ALLOC (NULL
);
645 gi
->store1_p
= BITMAP_ALLOC (NULL
);
646 gi
->store2_n
= BITMAP_ALLOC (NULL
);
647 gi
->store2_p
= BITMAP_ALLOC (NULL
);
648 gi
->group_kill
= BITMAP_ALLOC (NULL
);
649 gi
->process_globally
= false;
650 gi
->offset_map_size_n
= 0;
651 gi
->offset_map_size_p
= 0;
652 gi
->offset_map_n
= NULL
;
653 gi
->offset_map_p
= NULL
;
654 VEC_safe_push (group_info_t
, heap
, rtx_group_vec
, gi
);
656 return clear_alias_group
;
661 *slot
= gi
= pool_alloc (rtx_group_info_pool
);
663 gi
->id
= rtx_group_next_id
++;
664 gi
->base_mem
= gen_rtx_MEM (QImode
, base
);
665 gi
->canon_base_mem
= canon_rtx (gi
->base_mem
);
666 gi
->store1_n
= BITMAP_ALLOC (NULL
);
667 gi
->store1_p
= BITMAP_ALLOC (NULL
);
668 gi
->store2_n
= BITMAP_ALLOC (NULL
);
669 gi
->store2_p
= BITMAP_ALLOC (NULL
);
670 gi
->group_kill
= BITMAP_ALLOC (NULL
);
671 gi
->process_globally
= false;
673 (base
== frame_pointer_rtx
) || (base
== hard_frame_pointer_rtx
);
674 gi
->offset_map_size_n
= 0;
675 gi
->offset_map_size_p
= 0;
676 gi
->offset_map_n
= NULL
;
677 gi
->offset_map_p
= NULL
;
678 VEC_safe_push (group_info_t
, heap
, rtx_group_vec
, gi
);
685 /* Initialization of data structures. */
691 globally_deleted
= 0;
694 scratch
= BITMAP_ALLOC (NULL
);
697 = create_alloc_pool ("rtx_store_info_pool",
698 sizeof (struct store_info
), 100);
700 = create_alloc_pool ("read_info_pool",
701 sizeof (struct read_info
), 100);
703 = create_alloc_pool ("insn_info_pool",
704 sizeof (struct insn_info
), 100);
706 = create_alloc_pool ("bb_info_pool",
707 sizeof (struct bb_info
), 100);
709 = create_alloc_pool ("rtx_group_info_pool",
710 sizeof (struct group_info
), 100);
712 = create_alloc_pool ("deferred_change_pool",
713 sizeof (struct deferred_change
), 10);
715 rtx_group_table
= htab_create (11, invariant_group_base_hash
,
716 invariant_group_base_eq
, NULL
);
718 bb_table
= XCNEWVEC (bb_info_t
, last_basic_block
);
719 rtx_group_next_id
= 0;
721 stores_off_frame_dead_at_return
= !current_function_stdarg
;
723 init_alias_analysis ();
725 if (clear_alias_sets
)
726 clear_alias_group
= get_group_info (NULL
);
728 clear_alias_group
= NULL
;
733 /*----------------------------------------------------------------------------
736 Scan all of the insns. Any random ordering of the blocks is fine.
737 Each block is scanned in forward order to accommodate cselib which
738 is used to remove stores with non-constant bases.
739 ----------------------------------------------------------------------------*/
741 /* Delete all of the store_info recs from INSN_INFO. */
744 free_store_info (insn_info_t insn_info
)
746 store_info_t store_info
= insn_info
->store_rec
;
749 store_info_t next
= store_info
->next
;
750 if (store_info
->cse_base
)
751 pool_free (cse_store_info_pool
, store_info
);
753 pool_free (rtx_store_info_pool
, store_info
);
757 insn_info
->cannot_delete
= true;
758 insn_info
->contains_cselib_groups
= false;
759 insn_info
->store_rec
= NULL
;
769 /* Add an insn to do the add inside a x if it is a
770 PRE/POST-INC/DEC/MODIFY. D is an structure containing the insn and
771 the size of the mode of the MEM that this is inside of. */
774 replace_inc_dec (rtx
*r
, void *d
)
777 struct insn_size
*data
= (struct insn_size
*)d
;
778 switch (GET_CODE (x
))
783 rtx r1
= XEXP (x
, 0);
784 rtx c
= gen_int_mode (Pmode
, data
->size
);
785 add_insn_before (data
->insn
,
786 gen_rtx_SET (Pmode
, r1
,
787 gen_rtx_PLUS (Pmode
, r1
, c
)),
795 rtx r1
= XEXP (x
, 0);
796 rtx c
= gen_int_mode (Pmode
, -data
->size
);
797 add_insn_before (data
->insn
,
798 gen_rtx_SET (Pmode
, r1
,
799 gen_rtx_PLUS (Pmode
, r1
, c
)),
807 /* We can reuse the add because we are about to delete the
808 insn that contained it. */
809 rtx add
= XEXP (x
, 0);
810 rtx r1
= XEXP (add
, 0);
811 add_insn_before (data
->insn
,
812 gen_rtx_SET (Pmode
, r1
, add
), NULL
);
822 /* If X is a MEM, check the address to see if it is PRE/POST-INC/DEC/MODIFY
823 and generate an add to replace that. */
826 replace_inc_dec_mem (rtx
*r
, void *d
)
829 if (GET_CODE (x
) == MEM
)
831 struct insn_size data
;
833 data
.size
= GET_MODE_SIZE (GET_MODE (x
));
836 for_each_rtx (&XEXP (x
, 0), replace_inc_dec
, &data
);
843 /* Before we delete INSN, make sure that the auto inc/dec, if it is
844 there, is split into a separate insn. */
847 check_for_inc_dec (rtx insn
)
849 rtx note
= find_reg_note (insn
, REG_INC
, NULL_RTX
);
851 for_each_rtx (&insn
, replace_inc_dec_mem
, insn
);
855 /* Delete the insn and free all of the fields inside INSN_INFO. */
858 delete_dead_store_insn (insn_info_t insn_info
)
860 read_info_t read_info
;
865 check_for_inc_dec (insn_info
->insn
);
868 fprintf (dump_file
, "Locally deleting insn %d ",
869 INSN_UID (insn_info
->insn
));
870 if (insn_info
->store_rec
->alias_set
)
871 fprintf (dump_file
, "alias set %d\n",
872 (int) insn_info
->store_rec
->alias_set
);
874 fprintf (dump_file
, "\n");
877 free_store_info (insn_info
);
878 read_info
= insn_info
->read_rec
;
882 read_info_t next
= read_info
->next
;
883 pool_free (read_info_pool
, read_info
);
886 insn_info
->read_rec
= NULL
;
888 delete_insn (insn_info
->insn
);
890 insn_info
->insn
= NULL
;
892 insn_info
->wild_read
= false;
896 /* Set the store* bitmaps offset_map_size* fields in GROUP based on
900 set_usage_bits (group_info_t group
, HOST_WIDE_INT offset
, HOST_WIDE_INT width
)
904 if ((offset
> -MAX_OFFSET
) && (offset
< MAX_OFFSET
))
905 for (i
=offset
; i
<offset
+width
; i
++)
912 store1
= group
->store1_n
;
913 store2
= group
->store2_n
;
918 store1
= group
->store1_p
;
919 store2
= group
->store2_p
;
923 if (bitmap_bit_p (store1
, ai
))
924 bitmap_set_bit (store2
, ai
);
927 bitmap_set_bit (store1
, ai
);
930 if (group
->offset_map_size_n
< ai
)
931 group
->offset_map_size_n
= ai
;
935 if (group
->offset_map_size_p
< ai
)
936 group
->offset_map_size_p
= ai
;
943 /* Set the BB_INFO so that the last insn is marked as a wild read. */
946 add_wild_read (bb_info_t bb_info
)
948 insn_info_t insn_info
= bb_info
->last_insn
;
949 read_info_t
*ptr
= &insn_info
->read_rec
;
953 read_info_t next
= (*ptr
)->next
;
954 if ((*ptr
)->alias_set
== 0)
956 pool_free (read_info_pool
, *ptr
);
962 insn_info
->wild_read
= true;
963 active_local_stores
= NULL
;
967 /* Return true if X is a constant or one of the registers that behave
968 as a constant over the life of a function. This is equivalent to
969 !rtx_varies_p for memory addresses. */
972 const_or_frame_p (rtx x
)
974 switch (GET_CODE (x
))
977 return MEM_READONLY_P (x
);
988 /* Note that we have to test for the actual rtx used for the frame
989 and arg pointers and not just the register number in case we have
990 eliminated the frame and/or arg pointer and are using it
992 if (x
== frame_pointer_rtx
|| x
== hard_frame_pointer_rtx
993 /* The arg pointer varies if it is not a fixed register. */
994 || (x
== arg_pointer_rtx
&& fixed_regs
[ARG_POINTER_REGNUM
])
995 || x
== pic_offset_table_rtx
)
1004 /* Take all reasonable action to put the address of MEM into the form
1005 that we can do analysis on.
1007 The gold standard is to get the address into the form: address +
1008 OFFSET where address is something that rtx_varies_p considers a
1009 constant. When we can get the address in this form, we can do
1010 global analysis on it. Note that for constant bases, address is
1011 not actually returned, only the group_id. The address can be
1014 If that fails, we try cselib to get a value we can at least use
1015 locally. If that fails we return false.
1017 The GROUP_ID is set to -1 for cselib bases and the index of the
1018 group for non_varying bases.
1020 FOR_READ is true if this is a mem read and false if not. */
1023 canon_address (rtx mem
,
1024 alias_set_type
*alias_set_out
,
1026 HOST_WIDE_INT
*offset
,
1029 rtx mem_address
= XEXP (mem
, 0);
1030 rtx expanded_address
, address
;
1031 /* Make sure that cselib is has initialized all of the operands of
1032 the address before asking it to do the subst. */
1034 if (clear_alias_sets
)
1036 /* If this is a spill, do not do any further processing. */
1037 alias_set_type alias_set
= MEM_ALIAS_SET (mem
);
1039 fprintf (dump_file
, "found alias set %d\n", (int) alias_set
);
1040 if (bitmap_bit_p (clear_alias_sets
, alias_set
))
1042 struct clear_alias_mode_holder
*entry
1043 = clear_alias_set_lookup (alias_set
);
1045 /* If the modes do not match, we cannot process this set. */
1046 if (entry
->mode
!= GET_MODE (mem
))
1050 "disqualifying alias set %d, (%s) != (%s)\n",
1051 (int) alias_set
, GET_MODE_NAME (entry
->mode
),
1052 GET_MODE_NAME (GET_MODE (mem
)));
1054 bitmap_set_bit (disqualified_clear_alias_sets
, alias_set
);
1058 *alias_set_out
= alias_set
;
1059 *group_id
= clear_alias_group
->id
;
1066 cselib_lookup (mem_address
, Pmode
, 1);
1070 fprintf (dump_file
, " mem: ");
1071 print_inline_rtx (dump_file
, mem_address
, 0);
1072 fprintf (dump_file
, "\n");
1075 /* Use cselib to replace all of the reg references with the full
1076 expression. This will take care of the case where we have
1078 r_x = base + offset;
1083 val = *(base + offset);
1086 expanded_address
= cselib_expand_value_rtx (mem_address
, scratch
, 5);
1088 /* If this fails, just go with the mem_address. */
1089 if (!expanded_address
)
1090 expanded_address
= mem_address
;
1092 /* Split the address into canonical BASE + OFFSET terms. */
1093 address
= canon_rtx (expanded_address
);
1099 fprintf (dump_file
, "\n after cselib_expand address: ");
1100 print_inline_rtx (dump_file
, expanded_address
, 0);
1101 fprintf (dump_file
, "\n");
1103 fprintf (dump_file
, "\n after canon_rtx address: ");
1104 print_inline_rtx (dump_file
, address
, 0);
1105 fprintf (dump_file
, "\n");
1108 if (GET_CODE (address
) == CONST
)
1109 address
= XEXP (address
, 0);
1111 if (GET_CODE (address
) == PLUS
&& GET_CODE (XEXP (address
, 1)) == CONST_INT
)
1113 *offset
= INTVAL (XEXP (address
, 1));
1114 address
= XEXP (address
, 0);
1117 if (const_or_frame_p (address
))
1119 group_info_t group
= get_group_info (address
);
1122 fprintf (dump_file
, " gid=%d offset=%d \n", group
->id
, (int)*offset
);
1124 *group_id
= group
->id
;
1128 *base
= cselib_lookup (address
, Pmode
, true);
1134 fprintf (dump_file
, " no cselib val - should be a wild read.\n");
1138 fprintf (dump_file
, " varying cselib base=%d offset = %d\n",
1139 (*base
)->value
, (int)*offset
);
1145 /* Clear the rhs field from the active_local_stores array. */
1148 clear_rhs_from_active_local_stores (void)
1150 insn_info_t ptr
= active_local_stores
;
1154 store_info_t store_info
= ptr
->store_rec
;
1155 /* Skip the clobbers. */
1156 while (!store_info
->is_set
)
1157 store_info
= store_info
->next
;
1159 store_info
->rhs
= NULL
;
1161 ptr
= ptr
->next_local_store
;
1166 /* BODY is an instruction pattern that belongs to INSN. Return 1 if
1167 there is a candidate store, after adding it to the appropriate
1168 local store group if so. */
1171 record_store (rtx body
, bb_info_t bb_info
)
1174 HOST_WIDE_INT offset
= 0;
1175 HOST_WIDE_INT width
= 0;
1176 alias_set_type spill_alias_set
;
1177 insn_info_t insn_info
= bb_info
->last_insn
;
1178 store_info_t store_info
= NULL
;
1180 cselib_val
*base
= NULL
;
1181 insn_info_t ptr
, last
;
1182 bool store_is_unused
;
1184 if (GET_CODE (body
) != SET
&& GET_CODE (body
) != CLOBBER
)
1187 /* If this is not used, then this cannot be used to keep the insn
1188 from being deleted. On the other hand, it does provide something
1189 that can be used to prove that another store is dead. */
1191 = (find_reg_note (insn_info
->insn
, REG_UNUSED
, body
) != NULL
);
1193 /* Check whether that value is a suitable memory location. */
1194 mem
= SET_DEST (body
);
1197 /* If the set or clobber is unused, then it does not effect our
1198 ability to get rid of the entire insn. */
1199 if (!store_is_unused
)
1200 insn_info
->cannot_delete
= true;
1204 /* At this point we know mem is a mem. */
1205 if (GET_MODE (mem
) == BLKmode
)
1207 if (GET_CODE (XEXP (mem
, 0)) == SCRATCH
)
1210 fprintf (dump_file
, " adding wild read for (clobber (mem:BLK (scratch))\n");
1211 add_wild_read (bb_info
);
1212 insn_info
->cannot_delete
= true;
1214 else if (!store_is_unused
)
1216 /* If the set or clobber is unused, then it does not effect our
1217 ability to get rid of the entire insn. */
1218 insn_info
->cannot_delete
= true;
1219 clear_rhs_from_active_local_stores ();
1224 /* We can still process a volatile mem, we just cannot delete it. */
1225 if (MEM_VOLATILE_P (mem
))
1226 insn_info
->cannot_delete
= true;
1228 if (!canon_address (mem
, &spill_alias_set
, &group_id
, &offset
, &base
))
1230 clear_rhs_from_active_local_stores ();
1234 width
= GET_MODE_SIZE (GET_MODE (mem
));
1236 if (spill_alias_set
)
1238 bitmap store1
= clear_alias_group
->store1_p
;
1239 bitmap store2
= clear_alias_group
->store2_p
;
1241 if (bitmap_bit_p (store1
, spill_alias_set
))
1242 bitmap_set_bit (store2
, spill_alias_set
);
1244 bitmap_set_bit (store1
, spill_alias_set
);
1246 if (clear_alias_group
->offset_map_size_p
< spill_alias_set
)
1247 clear_alias_group
->offset_map_size_p
= spill_alias_set
;
1249 store_info
= pool_alloc (rtx_store_info_pool
);
1252 fprintf (dump_file
, " processing spill store %d(%s)\n",
1253 (int) spill_alias_set
, GET_MODE_NAME (GET_MODE (mem
)));
1255 else if (group_id
>= 0)
1257 /* In the restrictive case where the base is a constant or the
1258 frame pointer we can do global analysis. */
1261 = VEC_index (group_info_t
, rtx_group_vec
, group_id
);
1263 store_info
= pool_alloc (rtx_store_info_pool
);
1264 set_usage_bits (group
, offset
, width
);
1267 fprintf (dump_file
, " processing const base store gid=%d[%d..%d)\n",
1268 group_id
, (int)offset
, (int)(offset
+width
));
1272 rtx base_term
= find_base_term (XEXP (mem
, 0));
1274 || (GET_CODE (base_term
) == ADDRESS
1275 && GET_MODE (base_term
) == Pmode
1276 && XEXP (base_term
, 0) == stack_pointer_rtx
))
1277 insn_info
->stack_pointer_based
= true;
1278 insn_info
->contains_cselib_groups
= true;
1280 store_info
= pool_alloc (cse_store_info_pool
);
1284 fprintf (dump_file
, " processing cselib store [%d..%d)\n",
1285 (int)offset
, (int)(offset
+width
));
1288 /* Check to see if this stores causes some other stores to be
1290 ptr
= active_local_stores
;
1295 insn_info_t next
= ptr
->next_local_store
;
1296 store_info_t s_info
= ptr
->store_rec
;
1299 /* Skip the clobbers. We delete the active insn if this insn
1300 shadows the set. To have been put on the active list, it
1301 has exactly on set. */
1302 while (!s_info
->is_set
)
1303 s_info
= s_info
->next
;
1305 if (s_info
->alias_set
!= spill_alias_set
)
1307 else if (s_info
->alias_set
)
1309 struct clear_alias_mode_holder
*entry
1310 = clear_alias_set_lookup (s_info
->alias_set
);
1311 /* Generally, spills cannot be processed if and of the
1312 references to the slot have a different mode. But if
1313 we are in the same block and mode is exactly the same
1314 between this store and one before in the same block,
1315 we can still delete it. */
1316 if ((GET_MODE (mem
) == GET_MODE (s_info
->mem
))
1317 && (GET_MODE (mem
) == entry
->mode
))
1320 s_info
->positions_needed
= (unsigned HOST_WIDE_INT
) 0;
1323 fprintf (dump_file
, " trying spill store in insn=%d alias_set=%d\n",
1324 INSN_UID (ptr
->insn
), (int) s_info
->alias_set
);
1326 else if ((s_info
->group_id
== group_id
)
1327 && (s_info
->cse_base
== base
))
1331 fprintf (dump_file
, " trying store in insn=%d gid=%d[%d..%d)\n",
1332 INSN_UID (ptr
->insn
), s_info
->group_id
,
1333 (int)s_info
->begin
, (int)s_info
->end
);
1334 for (i
= offset
; i
< offset
+width
; i
++)
1335 if (i
>= s_info
->begin
&& i
< s_info
->end
)
1336 s_info
->positions_needed
1337 &= ~(((unsigned HOST_WIDE_INT
) 1) << (i
- s_info
->begin
));
1339 else if (s_info
->rhs
)
1340 /* Need to see if it is possible for this store to overwrite
1341 the value of store_info. If it is, set the rhs to NULL to
1342 keep it from being used to remove a load. */
1344 if (canon_true_dependence (s_info
->mem
,
1345 GET_MODE (s_info
->mem
),
1351 /* An insn can be deleted if every position of every one of
1352 its s_infos is zero. */
1353 if (s_info
->positions_needed
!= (unsigned HOST_WIDE_INT
) 0)
1358 insn_info_t insn_to_delete
= ptr
;
1361 last
->next_local_store
= ptr
->next_local_store
;
1363 active_local_stores
= ptr
->next_local_store
;
1365 delete_dead_store_insn (insn_to_delete
);
1373 gcc_assert ((unsigned) width
<= HOST_BITS_PER_WIDE_INT
);
1375 /* Finish filling in the store_info. */
1376 store_info
->next
= insn_info
->store_rec
;
1377 insn_info
->store_rec
= store_info
;
1378 store_info
->mem
= canon_rtx (mem
);
1379 store_info
->alias_set
= spill_alias_set
;
1380 store_info
->mem_addr
= get_addr (XEXP (mem
, 0));
1381 store_info
->cse_base
= base
;
1382 store_info
->positions_needed
= lowpart_bitmask (width
);
1383 store_info
->group_id
= group_id
;
1384 store_info
->begin
= offset
;
1385 store_info
->end
= offset
+ width
;
1386 store_info
->is_set
= GET_CODE (body
) == SET
;
1388 if (store_info
->is_set
1389 /* No place to keep the value after ra. */
1390 && !reload_completed
1391 && (REG_P (SET_SRC (body
))
1392 || GET_CODE (SET_SRC (body
)) == SUBREG
1393 || CONSTANT_P (SET_SRC (body
)))
1394 /* Sometimes the store and reload is used for truncation and
1396 && !(FLOAT_MODE_P (GET_MODE (mem
)) && (flag_float_store
)))
1397 store_info
->rhs
= SET_SRC (body
);
1399 store_info
->rhs
= NULL
;
1401 /* If this is a clobber, we return 0. We will only be able to
1402 delete this insn if there is only one store USED store, but we
1403 can use the clobber to delete other stores earlier. */
1404 return store_info
->is_set
? 1 : 0;
1409 dump_insn_info (const char * start
, insn_info_t insn_info
)
1411 fprintf (dump_file
, "%s insn=%d %s\n", start
,
1412 INSN_UID (insn_info
->insn
),
1413 insn_info
->store_rec
? "has store" : "naked");
1417 /* If the modes are different and the value's source and target do not
1418 line up, we need to extract the value from lower part of the rhs of
1419 the store, shift it, and then put it into a form that can be shoved
1420 into the read_insn. This function generates a right SHIFT of a
1421 value that is at least ACCESS_SIZE bytes wide of READ_MODE. The
1422 shift sequence is returned or NULL if we failed to find a
1426 find_shift_sequence (int access_size
,
1427 store_info_t store_info
,
1428 read_info_t read_info
,
1431 enum machine_mode store_mode
= GET_MODE (store_info
->mem
);
1432 enum machine_mode read_mode
= GET_MODE (read_info
->mem
);
1433 enum machine_mode new_mode
;
1434 rtx read_reg
= NULL
;
1436 /* Some machines like the x86 have shift insns for each size of
1437 operand. Other machines like the ppc or the ia-64 may only have
1438 shift insns that shift values within 32 or 64 bit registers.
1439 This loop tries to find the smallest shift insn that will right
1440 justify the value we want to read but is available in one insn on
1443 for (new_mode
= smallest_mode_for_size (access_size
* BITS_PER_UNIT
,
1445 GET_MODE_BITSIZE (new_mode
) <= BITS_PER_WORD
;
1446 new_mode
= GET_MODE_WIDER_MODE (new_mode
))
1448 rtx target
, new_reg
, shift_seq
, insn
, new_lhs
;
1451 /* Try a wider mode if truncating the store mode to NEW_MODE
1452 requires a real instruction. */
1453 if (GET_MODE_BITSIZE (new_mode
) < GET_MODE_BITSIZE (store_mode
)
1454 && !TRULY_NOOP_TRUNCATION (GET_MODE_BITSIZE (new_mode
),
1455 GET_MODE_BITSIZE (store_mode
)))
1458 /* Also try a wider mode if the necessary punning is either not
1459 desirable or not possible. */
1460 if (!CONSTANT_P (store_info
->rhs
)
1461 && !MODES_TIEABLE_P (new_mode
, store_mode
))
1463 new_lhs
= simplify_gen_subreg (new_mode
, copy_rtx (store_info
->rhs
),
1465 if (new_lhs
== NULL_RTX
)
1468 new_reg
= gen_reg_rtx (new_mode
);
1472 /* In theory we could also check for an ashr. Ian Taylor knows
1473 of one dsp where the cost of these two was not the same. But
1474 this really is a rare case anyway. */
1475 target
= expand_binop (new_mode
, lshr_optab
, new_reg
,
1476 GEN_INT (shift
), new_reg
, 1, OPTAB_DIRECT
);
1478 shift_seq
= get_insns ();
1481 if (target
!= new_reg
|| shift_seq
== NULL
)
1485 for (insn
= shift_seq
; insn
!= NULL_RTX
; insn
= NEXT_INSN (insn
))
1487 cost
+= insn_rtx_cost (PATTERN (insn
));
1489 /* The computation up to here is essentially independent
1490 of the arguments and could be precomputed. It may
1491 not be worth doing so. We could precompute if
1492 worthwhile or at least cache the results. The result
1493 technically depends on both SHIFT and ACCESS_SIZE,
1494 but in practice the answer will depend only on ACCESS_SIZE. */
1496 if (cost
> COSTS_N_INSNS (1))
1499 /* We found an acceptable shift. Generate a move to
1500 take the value from the store and put it into the
1501 shift pseudo, then shift it, then generate another
1502 move to put in into the target of the read. */
1503 emit_move_insn (new_reg
, new_lhs
);
1504 emit_insn (shift_seq
);
1505 read_reg
= extract_low_bits (read_mode
, new_mode
, new_reg
);
1513 /* Take a sequence of:
1536 Depending on the alignment and the mode of the store and
1540 The STORE_INFO and STORE_INSN are for the store and READ_INFO
1541 and READ_INSN are for the read. Return true if the replacement
1545 replace_read (store_info_t store_info
, insn_info_t store_insn
,
1546 read_info_t read_info
, insn_info_t read_insn
, rtx
*loc
)
1548 enum machine_mode store_mode
= GET_MODE (store_info
->mem
);
1549 enum machine_mode read_mode
= GET_MODE (read_info
->mem
);
1551 int access_size
; /* In bytes. */
1552 rtx insns
, read_reg
;
1557 /* To get here the read is within the boundaries of the write so
1558 shift will never be negative. Start out with the shift being in
1560 if (BYTES_BIG_ENDIAN
)
1561 shift
= store_info
->end
- read_info
->end
;
1563 shift
= read_info
->begin
- store_info
->begin
;
1565 access_size
= shift
+ GET_MODE_SIZE (read_mode
);
1567 /* From now on it is bits. */
1568 shift
*= BITS_PER_UNIT
;
1570 /* Create a sequence of instructions to set up the read register.
1571 This sequence goes immediately before the store and its result
1572 is read by the load.
1574 We need to keep this in perspective. We are replacing a read
1575 with a sequence of insns, but the read will almost certainly be
1576 in cache, so it is not going to be an expensive one. Thus, we
1577 are not willing to do a multi insn shift or worse a subroutine
1578 call to get rid of the read. */
1580 fprintf (dump_file
, "trying to replace %smode load in insn %d"
1581 " from %smode store in insn %d\n",
1582 GET_MODE_NAME (read_mode
), INSN_UID (read_insn
->insn
),
1583 GET_MODE_NAME (store_mode
), INSN_UID (store_insn
->insn
));
1586 read_reg
= find_shift_sequence (access_size
, store_info
, read_info
, shift
);
1588 read_reg
= extract_low_bits (read_mode
, store_mode
,
1589 copy_rtx (store_info
->rhs
));
1590 if (read_reg
== NULL_RTX
)
1594 fprintf (dump_file
, " -- could not extract bits of stored value\n");
1597 /* Force the value into a new register so that it won't be clobbered
1598 between the store and the load. */
1599 read_reg
= copy_to_mode_reg (read_mode
, read_reg
);
1600 insns
= get_insns ();
1603 if (validate_change (read_insn
->insn
, loc
, read_reg
, 0))
1605 deferred_change_t deferred_change
= pool_alloc (deferred_change_pool
);
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
);
1646 fprintf (dump_file
, " -- replaced the loaded MEM with ");
1647 print_simple_rtl (dump_file
, read_reg
);
1648 fprintf (dump_file
, "\n");
1656 fprintf (dump_file
, " -- replacing the loaded MEM with ");
1657 print_simple_rtl (dump_file
, read_reg
);
1658 fprintf (dump_file
, " led to an invalid instruction\n");
1664 /* A for_each_rtx callback in which DATA is the bb_info. Check to see
1665 if LOC is a mem and if it is look at the address and kill any
1666 appropriate stores that may be active. */
1669 check_mem_read_rtx (rtx
*loc
, void *data
)
1673 insn_info_t insn_info
;
1674 HOST_WIDE_INT offset
= 0;
1675 HOST_WIDE_INT width
= 0;
1676 alias_set_type spill_alias_set
= 0;
1677 cselib_val
*base
= NULL
;
1679 read_info_t read_info
;
1681 if (!mem
|| !MEM_P (mem
))
1684 bb_info
= (bb_info_t
) data
;
1685 insn_info
= bb_info
->last_insn
;
1687 if ((MEM_ALIAS_SET (mem
) == ALIAS_SET_MEMORY_BARRIER
)
1688 || (MEM_VOLATILE_P (mem
)))
1691 fprintf (dump_file
, " adding wild read, volatile or barrier.\n");
1692 add_wild_read (bb_info
);
1693 insn_info
->cannot_delete
= true;
1697 /* If it is reading readonly mem, then there can be no conflict with
1699 if (MEM_READONLY_P (mem
))
1702 if (!canon_address (mem
, &spill_alias_set
, &group_id
, &offset
, &base
))
1705 fprintf (dump_file
, " adding wild read, canon_address failure.\n");
1706 add_wild_read (bb_info
);
1710 if (GET_MODE (mem
) == BLKmode
)
1713 width
= GET_MODE_SIZE (GET_MODE (mem
));
1715 read_info
= pool_alloc (read_info_pool
);
1716 read_info
->group_id
= group_id
;
1717 read_info
->mem
= mem
;
1718 read_info
->alias_set
= spill_alias_set
;
1719 read_info
->begin
= offset
;
1720 read_info
->end
= offset
+ width
;
1721 read_info
->next
= insn_info
->read_rec
;
1722 insn_info
->read_rec
= read_info
;
1724 /* We ignore the clobbers in store_info. The is mildly aggressive,
1725 but there really should not be a clobber followed by a read. */
1727 if (spill_alias_set
)
1729 insn_info_t i_ptr
= active_local_stores
;
1730 insn_info_t last
= NULL
;
1733 fprintf (dump_file
, " processing spill load %d\n",
1734 (int) spill_alias_set
);
1738 store_info_t store_info
= i_ptr
->store_rec
;
1740 /* Skip the clobbers. */
1741 while (!store_info
->is_set
)
1742 store_info
= store_info
->next
;
1744 if (store_info
->alias_set
== spill_alias_set
)
1747 dump_insn_info ("removing from active", i_ptr
);
1750 last
->next_local_store
= i_ptr
->next_local_store
;
1752 active_local_stores
= i_ptr
->next_local_store
;
1756 i_ptr
= i_ptr
->next_local_store
;
1759 else if (group_id
>= 0)
1761 /* This is the restricted case where the base is a constant or
1762 the frame pointer and offset is a constant. */
1763 insn_info_t i_ptr
= active_local_stores
;
1764 insn_info_t last
= NULL
;
1769 fprintf (dump_file
, " processing const load gid=%d[BLK]\n",
1772 fprintf (dump_file
, " processing const load gid=%d[%d..%d)\n",
1773 group_id
, (int)offset
, (int)(offset
+width
));
1778 bool remove
= false;
1779 store_info_t store_info
= i_ptr
->store_rec
;
1781 /* Skip the clobbers. */
1782 while (!store_info
->is_set
)
1783 store_info
= store_info
->next
;
1785 /* There are three cases here. */
1786 if (store_info
->group_id
< 0)
1787 /* We have a cselib store followed by a read from a
1790 = canon_true_dependence (store_info
->mem
,
1791 GET_MODE (store_info
->mem
),
1792 store_info
->mem_addr
,
1795 else if (group_id
== store_info
->group_id
)
1797 /* This is a block mode load. We may get lucky and
1798 canon_true_dependence may save the day. */
1801 = canon_true_dependence (store_info
->mem
,
1802 GET_MODE (store_info
->mem
),
1803 store_info
->mem_addr
,
1806 /* If this read is just reading back something that we just
1807 stored, rewrite the read. */
1811 && (offset
>= store_info
->begin
)
1812 && (offset
+ width
<= store_info
->end
))
1814 unsigned HOST_WIDE_INT mask
1815 = (lowpart_bitmask (width
)
1816 << (offset
- store_info
->begin
));
1818 if ((store_info
->positions_needed
& mask
) == mask
1819 && replace_read (store_info
, i_ptr
,
1820 read_info
, insn_info
, loc
))
1823 /* The bases are the same, just see if the offsets
1825 if ((offset
< store_info
->end
)
1826 && (offset
+ width
> store_info
->begin
))
1832 The else case that is missing here is that the
1833 bases are constant but different. There is nothing
1834 to do here because there is no overlap. */
1839 dump_insn_info ("removing from active", i_ptr
);
1842 last
->next_local_store
= i_ptr
->next_local_store
;
1844 active_local_stores
= i_ptr
->next_local_store
;
1848 i_ptr
= i_ptr
->next_local_store
;
1853 insn_info_t i_ptr
= active_local_stores
;
1854 insn_info_t last
= NULL
;
1857 fprintf (dump_file
, " processing cselib load mem:");
1858 print_inline_rtx (dump_file
, mem
, 0);
1859 fprintf (dump_file
, "\n");
1864 bool remove
= false;
1865 store_info_t store_info
= i_ptr
->store_rec
;
1868 fprintf (dump_file
, " processing cselib load against insn %d\n",
1869 INSN_UID (i_ptr
->insn
));
1871 /* Skip the clobbers. */
1872 while (!store_info
->is_set
)
1873 store_info
= store_info
->next
;
1875 /* If this read is just reading back something that we just
1876 stored, rewrite the read. */
1878 && store_info
->group_id
== -1
1879 && store_info
->cse_base
== base
1880 && (offset
>= store_info
->begin
)
1881 && (offset
+ width
<= store_info
->end
))
1883 unsigned HOST_WIDE_INT mask
1884 = (lowpart_bitmask (width
)
1885 << (offset
- store_info
->begin
));
1887 if ((store_info
->positions_needed
& mask
) == mask
1888 && replace_read (store_info
, i_ptr
,
1889 read_info
, insn_info
, loc
))
1893 if (!store_info
->alias_set
)
1894 remove
= canon_true_dependence (store_info
->mem
,
1895 GET_MODE (store_info
->mem
),
1896 store_info
->mem_addr
,
1902 dump_insn_info ("removing from active", i_ptr
);
1905 last
->next_local_store
= i_ptr
->next_local_store
;
1907 active_local_stores
= i_ptr
->next_local_store
;
1911 i_ptr
= i_ptr
->next_local_store
;
1917 /* A for_each_rtx callback in which DATA points the INSN_INFO for
1918 as check_mem_read_rtx. Nullify the pointer if i_m_r_m_r returns
1919 true for any part of *LOC. */
1922 check_mem_read_use (rtx
*loc
, void *data
)
1924 for_each_rtx (loc
, check_mem_read_rtx
, data
);
1927 /* Apply record_store to all candidate stores in INSN. Mark INSN
1928 if some part of it is not a candidate store and assigns to a
1929 non-register target. */
1932 scan_insn (bb_info_t bb_info
, rtx insn
)
1935 insn_info_t insn_info
= pool_alloc (insn_info_pool
);
1937 memset (insn_info
, 0, sizeof (struct insn_info
));
1940 fprintf (dump_file
, "\n**scanning insn=%d\n",
1943 insn_info
->prev_insn
= bb_info
->last_insn
;
1944 insn_info
->insn
= insn
;
1945 bb_info
->last_insn
= insn_info
;
1948 /* Cselib clears the table for this case, so we have to essentially
1950 if (NONJUMP_INSN_P (insn
)
1951 && GET_CODE (PATTERN (insn
)) == ASM_OPERANDS
1952 && MEM_VOLATILE_P (PATTERN (insn
)))
1954 add_wild_read (bb_info
);
1955 insn_info
->cannot_delete
= true;
1959 /* Look at all of the uses in the insn. */
1960 note_uses (&PATTERN (insn
), check_mem_read_use
, bb_info
);
1964 insn_info
->cannot_delete
= true;
1966 /* Const functions cannot do anything bad i.e. read memory,
1967 however, they can read their parameters which may have
1968 been pushed onto the stack. */
1969 if (CONST_OR_PURE_CALL_P (insn
) && !pure_call_p (insn
))
1971 insn_info_t i_ptr
= active_local_stores
;
1972 insn_info_t last
= NULL
;
1975 fprintf (dump_file
, "const call %d\n", INSN_UID (insn
));
1977 /* See the head comment of the frame_read field. */
1978 if (reload_completed
)
1979 insn_info
->frame_read
= true;
1981 /* Loop over the active stores and remove those which are
1982 killed by the const function call. */
1985 bool remove_store
= false;
1987 /* The stack pointer based stores are always killed. */
1988 if (i_ptr
->stack_pointer_based
)
1989 remove_store
= true;
1991 /* If the frame is read, the frame related stores are killed. */
1992 else if (insn_info
->frame_read
)
1994 store_info_t store_info
= i_ptr
->store_rec
;
1996 /* Skip the clobbers. */
1997 while (!store_info
->is_set
)
1998 store_info
= store_info
->next
;
2000 if (store_info
->group_id
>= 0
2001 && VEC_index (group_info_t
, rtx_group_vec
,
2002 store_info
->group_id
)->frame_related
)
2003 remove_store
= true;
2009 dump_insn_info ("removing from active", i_ptr
);
2012 last
->next_local_store
= i_ptr
->next_local_store
;
2014 active_local_stores
= i_ptr
->next_local_store
;
2019 i_ptr
= i_ptr
->next_local_store
;
2024 /* Every other call, including pure functions, may read memory. */
2025 add_wild_read (bb_info
);
2030 /* Assuming that there are sets in these insns, we cannot delete
2032 if ((GET_CODE (PATTERN (insn
)) == CLOBBER
)
2033 || volatile_refs_p (PATTERN (insn
))
2034 || (flag_non_call_exceptions
&& may_trap_p (PATTERN (insn
)))
2035 || (RTX_FRAME_RELATED_P (insn
))
2036 || find_reg_note (insn
, REG_FRAME_RELATED_EXPR
, NULL_RTX
))
2037 insn_info
->cannot_delete
= true;
2039 body
= PATTERN (insn
);
2040 if (GET_CODE (body
) == PARALLEL
)
2043 for (i
= 0; i
< XVECLEN (body
, 0); i
++)
2044 mems_found
+= record_store (XVECEXP (body
, 0, i
), bb_info
);
2047 mems_found
+= record_store (body
, bb_info
);
2050 fprintf (dump_file
, "mems_found = %d, cannot_delete = %s\n",
2051 mems_found
, insn_info
->cannot_delete
? "true" : "false");
2053 /* If we found some sets of mems, and the insn has not been marked
2054 cannot delete, add it into the active_local_stores so that it can
2055 be locally deleted if found dead. Otherwise mark it as cannot
2056 delete. This simplifies the processing later. */
2057 if (mems_found
== 1 && !insn_info
->cannot_delete
)
2059 insn_info
->next_local_store
= active_local_stores
;
2060 active_local_stores
= insn_info
;
2063 insn_info
->cannot_delete
= true;
2067 /* Remove BASE from the set of active_local_stores. This is a
2068 callback from cselib that is used to get rid of the stores in
2069 active_local_stores. */
2072 remove_useless_values (cselib_val
*base
)
2074 insn_info_t insn_info
= active_local_stores
;
2075 insn_info_t last
= NULL
;
2079 store_info_t store_info
= insn_info
->store_rec
;
2080 bool delete = false;
2082 /* If ANY of the store_infos match the cselib group that is
2083 being deleted, then the insn can not be deleted. */
2086 if ((store_info
->group_id
== -1)
2087 && (store_info
->cse_base
== base
))
2092 store_info
= store_info
->next
;
2098 last
->next_local_store
= insn_info
->next_local_store
;
2100 active_local_stores
= insn_info
->next_local_store
;
2101 free_store_info (insn_info
);
2106 insn_info
= insn_info
->next_local_store
;
2111 /* Do all of step 1. */
2118 cselib_init (false);
2119 all_blocks
= BITMAP_ALLOC (NULL
);
2120 bitmap_set_bit (all_blocks
, ENTRY_BLOCK
);
2121 bitmap_set_bit (all_blocks
, EXIT_BLOCK
);
2126 bb_info_t bb_info
= pool_alloc (bb_info_pool
);
2128 memset (bb_info
, 0, sizeof (struct bb_info
));
2129 bitmap_set_bit (all_blocks
, bb
->index
);
2131 bb_table
[bb
->index
] = bb_info
;
2132 cselib_discard_hook
= remove_useless_values
;
2134 if (bb
->index
>= NUM_FIXED_BLOCKS
)
2139 = create_alloc_pool ("cse_store_info_pool",
2140 sizeof (struct store_info
), 100);
2141 active_local_stores
= NULL
;
2142 cselib_clear_table ();
2144 /* Scan the insns. */
2145 FOR_BB_INSNS (bb
, insn
)
2148 scan_insn (bb_info
, insn
);
2149 cselib_process_insn (insn
);
2152 /* This is something of a hack, because the global algorithm
2153 is supposed to take care of the case where stores go dead
2154 at the end of the function. However, the global
2155 algorithm must take a more conservative view of block
2156 mode reads than the local alg does. So to get the case
2157 where you have a store to the frame followed by a non
2158 overlapping block more read, we look at the active local
2159 stores at the end of the function and delete all of the
2160 frame and spill based ones. */
2161 if (stores_off_frame_dead_at_return
2162 && (EDGE_COUNT (bb
->succs
) == 0
2163 || (single_succ_p (bb
)
2164 && single_succ (bb
) == EXIT_BLOCK_PTR
2165 && ! current_function_calls_eh_return
)))
2167 insn_info_t i_ptr
= active_local_stores
;
2170 store_info_t store_info
= i_ptr
->store_rec
;
2172 /* Skip the clobbers. */
2173 while (!store_info
->is_set
)
2174 store_info
= store_info
->next
;
2175 if (store_info
->alias_set
)
2176 delete_dead_store_insn (i_ptr
);
2178 if (store_info
->group_id
>= 0)
2181 = VEC_index (group_info_t
, rtx_group_vec
, store_info
->group_id
);
2182 if (group
->frame_related
)
2183 delete_dead_store_insn (i_ptr
);
2186 i_ptr
= i_ptr
->next_local_store
;
2190 /* Get rid of the loads that were discovered in
2191 replace_read. Cselib is finished with this block. */
2192 while (deferred_change_list
)
2194 deferred_change_t next
= deferred_change_list
->next
;
2196 /* There is no reason to validate this change. That was
2198 *deferred_change_list
->loc
= deferred_change_list
->reg
;
2199 pool_free (deferred_change_pool
, deferred_change_list
);
2200 deferred_change_list
= next
;
2203 /* Get rid of all of the cselib based store_infos in this
2204 block and mark the containing insns as not being
2206 ptr
= bb_info
->last_insn
;
2209 if (ptr
->contains_cselib_groups
)
2210 free_store_info (ptr
);
2211 ptr
= ptr
->prev_insn
;
2214 free_alloc_pool (cse_store_info_pool
);
2219 htab_empty (rtx_group_table
);
2223 /*----------------------------------------------------------------------------
2226 Assign each byte position in the stores that we are going to
2227 analyze globally to a position in the bitmaps. Returns true if
2228 there are any bit positions assigned.
2229 ----------------------------------------------------------------------------*/
2232 dse_step2_init (void)
2237 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2239 /* For all non stack related bases, we only consider a store to
2240 be deletable if there are two or more stores for that
2241 position. This is because it takes one store to make the
2242 other store redundant. However, for the stores that are
2243 stack related, we consider them if there is only one store
2244 for the position. We do this because the stack related
2245 stores can be deleted if their is no read between them and
2246 the end of the function.
2248 To make this work in the current framework, we take the stack
2249 related bases add all of the bits from store1 into store2.
2250 This has the effect of making the eligible even if there is
2253 if (stores_off_frame_dead_at_return
&& group
->frame_related
)
2255 bitmap_ior_into (group
->store2_n
, group
->store1_n
);
2256 bitmap_ior_into (group
->store2_p
, group
->store1_p
);
2258 fprintf (dump_file
, "group %d is frame related ", i
);
2261 group
->offset_map_size_n
++;
2262 group
->offset_map_n
= XNEWVEC (int, group
->offset_map_size_n
);
2263 group
->offset_map_size_p
++;
2264 group
->offset_map_p
= XNEWVEC (int, group
->offset_map_size_p
);
2265 group
->process_globally
= false;
2268 fprintf (dump_file
, "group %d(%d+%d): ", i
,
2269 (int)bitmap_count_bits (group
->store2_n
),
2270 (int)bitmap_count_bits (group
->store2_p
));
2271 bitmap_print (dump_file
, group
->store2_n
, "n ", " ");
2272 bitmap_print (dump_file
, group
->store2_p
, "p ", "\n");
2278 /* Init the offset tables for the normal case. */
2281 dse_step2_nospill (void)
2285 /* Position 0 is unused because 0 is used in the maps to mean
2287 current_position
= 1;
2289 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2294 if (group
== clear_alias_group
)
2297 memset (group
->offset_map_n
, 0, sizeof(int) * group
->offset_map_size_n
);
2298 memset (group
->offset_map_p
, 0, sizeof(int) * group
->offset_map_size_p
);
2299 bitmap_clear (group
->group_kill
);
2301 EXECUTE_IF_SET_IN_BITMAP (group
->store2_n
, 0, j
, bi
)
2303 bitmap_set_bit (group
->group_kill
, current_position
);
2304 group
->offset_map_n
[j
] = current_position
++;
2305 group
->process_globally
= true;
2307 EXECUTE_IF_SET_IN_BITMAP (group
->store2_p
, 0, j
, bi
)
2309 bitmap_set_bit (group
->group_kill
, current_position
);
2310 group
->offset_map_p
[j
] = current_position
++;
2311 group
->process_globally
= true;
2314 return current_position
!= 1;
2318 /* Init the offset tables for the spill case. */
2321 dse_step2_spill (void)
2324 group_info_t group
= clear_alias_group
;
2327 /* Position 0 is unused because 0 is used in the maps to mean
2329 current_position
= 1;
2333 bitmap_print (dump_file
, clear_alias_sets
,
2334 "clear alias sets ", "\n");
2335 bitmap_print (dump_file
, disqualified_clear_alias_sets
,
2336 "disqualified clear alias sets ", "\n");
2339 memset (group
->offset_map_n
, 0, sizeof(int) * group
->offset_map_size_n
);
2340 memset (group
->offset_map_p
, 0, sizeof(int) * group
->offset_map_size_p
);
2341 bitmap_clear (group
->group_kill
);
2343 /* Remove the disqualified positions from the store2_p set. */
2344 bitmap_and_compl_into (group
->store2_p
, disqualified_clear_alias_sets
);
2346 /* We do not need to process the store2_n set because
2347 alias_sets are always positive. */
2348 EXECUTE_IF_SET_IN_BITMAP (group
->store2_p
, 0, j
, bi
)
2350 bitmap_set_bit (group
->group_kill
, current_position
);
2351 group
->offset_map_p
[j
] = current_position
++;
2352 group
->process_globally
= true;
2355 return current_position
!= 1;
2360 /*----------------------------------------------------------------------------
2363 Build the bit vectors for the transfer functions.
2364 ----------------------------------------------------------------------------*/
2367 /* Note that this is NOT a general purpose function. Any mem that has
2368 an alias set registered here expected to be COMPLETELY unaliased:
2369 i.e it's addresses are not and need not be examined.
2371 It is known that all references to this address will have this
2372 alias set and there are NO other references to this address in the
2375 Currently the only place that is known to be clean enough to use
2376 this interface is the code that assigns the spill locations.
2378 All of the mems that have alias_sets registered are subjected to a
2379 very powerful form of dse where function calls, volatile reads and
2380 writes, and reads from random location are not taken into account.
2382 It is also assumed that these locations go dead when the function
2383 returns. This assumption could be relaxed if there were found to
2384 be places that this assumption was not correct.
2386 The MODE is passed in and saved. The mode of each load or store to
2387 a mem with ALIAS_SET is checked against MEM. If the size of that
2388 load or store is different from MODE, processing is halted on this
2389 alias set. For the vast majority of aliases sets, all of the loads
2390 and stores will use the same mode. But vectors are treated
2391 differently: the alias set is established for the entire vector,
2392 but reload will insert loads and stores for individual elements and
2393 we do not necessarily have the information to track those separate
2394 elements. So when we see a mode mismatch, we just bail. */
2398 dse_record_singleton_alias_set (alias_set_type alias_set
,
2399 enum machine_mode mode
)
2401 struct clear_alias_mode_holder tmp_holder
;
2402 struct clear_alias_mode_holder
*entry
;
2405 /* If we are not going to run dse, we need to return now or there
2406 will be problems with allocating the bitmaps. */
2407 if ((!gate_dse()) || !alias_set
)
2410 if (!clear_alias_sets
)
2412 clear_alias_sets
= BITMAP_ALLOC (NULL
);
2413 disqualified_clear_alias_sets
= BITMAP_ALLOC (NULL
);
2414 clear_alias_mode_table
= htab_create (11, clear_alias_mode_hash
,
2415 clear_alias_mode_eq
, NULL
);
2416 clear_alias_mode_pool
= create_alloc_pool ("clear_alias_mode_pool",
2417 sizeof (struct clear_alias_mode_holder
), 100);
2420 bitmap_set_bit (clear_alias_sets
, alias_set
);
2422 tmp_holder
.alias_set
= alias_set
;
2424 slot
= htab_find_slot (clear_alias_mode_table
, &tmp_holder
, INSERT
);
2425 gcc_assert (*slot
== NULL
);
2427 *slot
= entry
= pool_alloc (clear_alias_mode_pool
);
2428 entry
->alias_set
= alias_set
;
2433 /* Remove ALIAS_SET from the sets of stack slots being considered. */
2436 dse_invalidate_singleton_alias_set (alias_set_type alias_set
)
2438 if ((!gate_dse()) || !alias_set
)
2441 bitmap_clear_bit (clear_alias_sets
, alias_set
);
2445 /* Look up the bitmap index for OFFSET in GROUP_INFO. If it is not
2449 get_bitmap_index (group_info_t group_info
, HOST_WIDE_INT offset
)
2453 HOST_WIDE_INT offset_p
= -offset
;
2454 if (offset_p
>= group_info
->offset_map_size_n
)
2456 return group_info
->offset_map_n
[offset_p
];
2460 if (offset
>= group_info
->offset_map_size_p
)
2462 return group_info
->offset_map_p
[offset
];
2467 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL
2471 scan_stores_nospill (store_info_t store_info
, bitmap gen
, bitmap kill
)
2476 group_info_t group_info
2477 = VEC_index (group_info_t
, rtx_group_vec
, store_info
->group_id
);
2478 if (group_info
->process_globally
)
2479 for (i
= store_info
->begin
; i
< store_info
->end
; i
++)
2481 int index
= get_bitmap_index (group_info
, i
);
2484 bitmap_set_bit (gen
, index
);
2486 bitmap_clear_bit (kill
, index
);
2489 store_info
= store_info
->next
;
2494 /* Process the STORE_INFOs into the bitmaps into GEN and KILL. KILL
2498 scan_stores_spill (store_info_t store_info
, bitmap gen
, bitmap kill
)
2502 if (store_info
->alias_set
)
2504 int index
= get_bitmap_index (clear_alias_group
,
2505 store_info
->alias_set
);
2508 bitmap_set_bit (gen
, index
);
2510 bitmap_clear_bit (kill
, index
);
2513 store_info
= store_info
->next
;
2518 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL
2522 scan_reads_nospill (insn_info_t insn_info
, bitmap gen
, bitmap kill
)
2524 read_info_t read_info
= insn_info
->read_rec
;
2528 /* If this insn reads the frame, kill all the frame related stores. */
2529 if (insn_info
->frame_read
)
2531 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2532 if (group
->process_globally
&& group
->frame_related
)
2535 bitmap_ior_into (kill
, group
->group_kill
);
2536 bitmap_and_compl_into (gen
, group
->group_kill
);
2542 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2544 if (group
->process_globally
)
2546 if (i
== read_info
->group_id
)
2548 if (read_info
->begin
> read_info
->end
)
2550 /* Begin > end for block mode reads. */
2552 bitmap_ior_into (kill
, group
->group_kill
);
2553 bitmap_and_compl_into (gen
, group
->group_kill
);
2557 /* The groups are the same, just process the
2560 for (j
= read_info
->begin
; j
< read_info
->end
; j
++)
2562 int index
= get_bitmap_index (group
, j
);
2566 bitmap_set_bit (kill
, index
);
2567 bitmap_clear_bit (gen
, index
);
2574 /* The groups are different, if the alias sets
2575 conflict, clear the entire group. We only need
2576 to apply this test if the read_info is a cselib
2577 read. Anything with a constant base cannot alias
2578 something else with a different constant
2580 if ((read_info
->group_id
< 0)
2581 && canon_true_dependence (group
->base_mem
,
2583 group
->canon_base_mem
,
2584 read_info
->mem
, rtx_varies_p
))
2587 bitmap_ior_into (kill
, group
->group_kill
);
2588 bitmap_and_compl_into (gen
, group
->group_kill
);
2594 read_info
= read_info
->next
;
2598 /* Process the READ_INFOs into the bitmaps into GEN and KILL. KILL
2602 scan_reads_spill (read_info_t read_info
, bitmap gen
, bitmap kill
)
2606 if (read_info
->alias_set
)
2608 int index
= get_bitmap_index (clear_alias_group
,
2609 read_info
->alias_set
);
2613 bitmap_set_bit (kill
, index
);
2614 bitmap_clear_bit (gen
, index
);
2618 read_info
= read_info
->next
;
2623 /* Return the insn in BB_INFO before the first wild read or if there
2624 are no wild reads in the block, return the last insn. */
2627 find_insn_before_first_wild_read (bb_info_t bb_info
)
2629 insn_info_t insn_info
= bb_info
->last_insn
;
2630 insn_info_t last_wild_read
= NULL
;
2634 if (insn_info
->wild_read
)
2636 last_wild_read
= insn_info
->prev_insn
;
2637 /* Block starts with wild read. */
2638 if (!last_wild_read
)
2642 insn_info
= insn_info
->prev_insn
;
2646 return last_wild_read
;
2648 return bb_info
->last_insn
;
2652 /* Scan the insns in BB_INFO starting at PTR and going to the top of
2653 the block in order to build the gen and kill sets for the block.
2654 We start at ptr which may be the last insn in the block or may be
2655 the first insn with a wild read. In the latter case we are able to
2656 skip the rest of the block because it just does not matter:
2657 anything that happens is hidden by the wild read. */
2660 dse_step3_scan (bool for_spills
, basic_block bb
)
2662 bb_info_t bb_info
= bb_table
[bb
->index
];
2663 insn_info_t insn_info
;
2666 /* There are no wild reads in the spill case. */
2667 insn_info
= bb_info
->last_insn
;
2669 insn_info
= find_insn_before_first_wild_read (bb_info
);
2671 /* In the spill case or in the no_spill case if there is no wild
2672 read in the block, we will need a kill set. */
2673 if (insn_info
== bb_info
->last_insn
)
2676 bitmap_clear (bb_info
->kill
);
2678 bb_info
->kill
= BITMAP_ALLOC (NULL
);
2682 BITMAP_FREE (bb_info
->kill
);
2686 /* There may have been code deleted by the dce pass run before
2688 if (insn_info
->insn
&& INSN_P (insn_info
->insn
))
2690 /* Process the read(s) last. */
2693 scan_stores_spill (insn_info
->store_rec
, bb_info
->gen
, bb_info
->kill
);
2694 scan_reads_spill (insn_info
->read_rec
, bb_info
->gen
, bb_info
->kill
);
2698 scan_stores_nospill (insn_info
->store_rec
, bb_info
->gen
, bb_info
->kill
);
2699 scan_reads_nospill (insn_info
, bb_info
->gen
, bb_info
->kill
);
2703 insn_info
= insn_info
->prev_insn
;
2708 /* Set the gen set of the exit block, and also any block with no
2709 successors that does not have a wild read. */
2712 dse_step3_exit_block_scan (bb_info_t bb_info
)
2714 /* The gen set is all 0's for the exit block except for the
2715 frame_pointer_group. */
2717 if (stores_off_frame_dead_at_return
)
2722 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
2724 if (group
->process_globally
&& group
->frame_related
)
2725 bitmap_ior_into (bb_info
->gen
, group
->group_kill
);
2731 /* Find all of the blocks that are not backwards reachable from the
2732 exit block or any block with no successors (BB). These are the
2733 infinite loops or infinite self loops. These blocks will still
2734 have their bits set in UNREACHABLE_BLOCKS. */
2737 mark_reachable_blocks (sbitmap unreachable_blocks
, basic_block bb
)
2742 if (TEST_BIT (unreachable_blocks
, bb
->index
))
2744 RESET_BIT (unreachable_blocks
, bb
->index
);
2745 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
2747 mark_reachable_blocks (unreachable_blocks
, e
->src
);
2752 /* Build the transfer functions for the function. */
2755 dse_step3 (bool for_spills
)
2758 sbitmap unreachable_blocks
= sbitmap_alloc (last_basic_block
);
2759 sbitmap_iterator sbi
;
2760 bitmap all_ones
= NULL
;
2763 sbitmap_ones (unreachable_blocks
);
2767 bb_info_t bb_info
= bb_table
[bb
->index
];
2769 bitmap_clear (bb_info
->gen
);
2771 bb_info
->gen
= BITMAP_ALLOC (NULL
);
2773 if (bb
->index
== ENTRY_BLOCK
)
2775 else if (bb
->index
== EXIT_BLOCK
)
2776 dse_step3_exit_block_scan (bb_info
);
2778 dse_step3_scan (for_spills
, bb
);
2779 if (EDGE_COUNT (bb
->succs
) == 0)
2780 mark_reachable_blocks (unreachable_blocks
, bb
);
2782 /* If this is the second time dataflow is run, delete the old
2785 BITMAP_FREE (bb_info
->in
);
2787 BITMAP_FREE (bb_info
->out
);
2790 /* For any block in an infinite loop, we must initialize the out set
2791 to all ones. This could be expensive, but almost never occurs in
2792 practice. However, it is common in regression tests. */
2793 EXECUTE_IF_SET_IN_SBITMAP (unreachable_blocks
, 0, i
, sbi
)
2795 if (bitmap_bit_p (all_blocks
, i
))
2797 bb_info_t bb_info
= bb_table
[i
];
2803 all_ones
= BITMAP_ALLOC (NULL
);
2804 for (j
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, j
, group
); j
++)
2805 bitmap_ior_into (all_ones
, group
->group_kill
);
2809 bb_info
->out
= BITMAP_ALLOC (NULL
);
2810 bitmap_copy (bb_info
->out
, all_ones
);
2816 BITMAP_FREE (all_ones
);
2817 sbitmap_free (unreachable_blocks
);
2822 /*----------------------------------------------------------------------------
2825 Solve the bitvector equations.
2826 ----------------------------------------------------------------------------*/
2829 /* Confluence function for blocks with no successors. Create an out
2830 set from the gen set of the exit block. This block logically has
2831 the exit block as a successor. */
2836 dse_confluence_0 (basic_block bb
)
2838 bb_info_t bb_info
= bb_table
[bb
->index
];
2840 if (bb
->index
== EXIT_BLOCK
)
2845 bb_info
->out
= BITMAP_ALLOC (NULL
);
2846 bitmap_copy (bb_info
->out
, bb_table
[EXIT_BLOCK
]->gen
);
2850 /* Propagate the information from the in set of the dest of E to the
2851 out set of the src of E. If the various in or out sets are not
2852 there, that means they are all ones. */
2855 dse_confluence_n (edge e
)
2857 bb_info_t src_info
= bb_table
[e
->src
->index
];
2858 bb_info_t dest_info
= bb_table
[e
->dest
->index
];
2863 bitmap_and_into (src_info
->out
, dest_info
->in
);
2866 src_info
->out
= BITMAP_ALLOC (NULL
);
2867 bitmap_copy (src_info
->out
, dest_info
->in
);
2873 /* Propagate the info from the out to the in set of BB_INDEX's basic
2874 block. There are three cases:
2876 1) The block has no kill set. In this case the kill set is all
2877 ones. It does not matter what the out set of the block is, none of
2878 the info can reach the top. The only thing that reaches the top is
2879 the gen set and we just copy the set.
2881 2) There is a kill set but no out set and bb has successors. In
2882 this case we just return. Eventually an out set will be created and
2883 it is better to wait than to create a set of ones.
2885 3) There is both a kill and out set. We apply the obvious transfer
2890 dse_transfer_function (int bb_index
)
2892 bb_info_t bb_info
= bb_table
[bb_index
];
2900 return bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
2901 bb_info
->out
, bb_info
->kill
);
2904 bb_info
->in
= BITMAP_ALLOC (NULL
);
2905 bitmap_ior_and_compl (bb_info
->in
, bb_info
->gen
,
2906 bb_info
->out
, bb_info
->kill
);
2916 /* Case 1 above. If there is already an in set, nothing
2922 bb_info
->in
= BITMAP_ALLOC (NULL
);
2923 bitmap_copy (bb_info
->in
, bb_info
->gen
);
2929 /* Solve the dataflow equations. */
2934 df_simple_dataflow (DF_BACKWARD
, NULL
, dse_confluence_0
,
2935 dse_confluence_n
, dse_transfer_function
,
2936 all_blocks
, df_get_postorder (DF_BACKWARD
),
2937 df_get_n_blocks (DF_BACKWARD
));
2942 fprintf (dump_file
, "\n\n*** Global dataflow info after analysis.\n");
2945 bb_info_t bb_info
= bb_table
[bb
->index
];
2947 df_print_bb_index (bb
, dump_file
);
2949 bitmap_print (dump_file
, bb_info
->in
, " in: ", "\n");
2951 fprintf (dump_file
, " in: *MISSING*\n");
2953 bitmap_print (dump_file
, bb_info
->gen
, " gen: ", "\n");
2955 fprintf (dump_file
, " gen: *MISSING*\n");
2957 bitmap_print (dump_file
, bb_info
->kill
, " kill: ", "\n");
2959 fprintf (dump_file
, " kill: *MISSING*\n");
2961 bitmap_print (dump_file
, bb_info
->out
, " out: ", "\n");
2963 fprintf (dump_file
, " out: *MISSING*\n\n");
2970 /*----------------------------------------------------------------------------
2973 Delete the stores that can only be deleted using the global information.
2974 ----------------------------------------------------------------------------*/
2978 dse_step5_nospill (void)
2983 bb_info_t bb_info
= bb_table
[bb
->index
];
2984 insn_info_t insn_info
= bb_info
->last_insn
;
2985 bitmap v
= bb_info
->out
;
2989 bool deleted
= false;
2990 if (dump_file
&& insn_info
->insn
)
2992 fprintf (dump_file
, "starting to process insn %d\n",
2993 INSN_UID (insn_info
->insn
));
2994 bitmap_print (dump_file
, v
, " v: ", "\n");
2997 /* There may have been code deleted by the dce pass run before
3000 && INSN_P (insn_info
->insn
)
3001 && (!insn_info
->cannot_delete
)
3002 && (!bitmap_empty_p (v
)))
3004 store_info_t store_info
= insn_info
->store_rec
;
3006 /* Try to delete the current insn. */
3009 /* Skip the clobbers. */
3010 while (!store_info
->is_set
)
3011 store_info
= store_info
->next
;
3013 if (store_info
->alias_set
)
3018 group_info_t group_info
3019 = VEC_index (group_info_t
, rtx_group_vec
, store_info
->group_id
);
3021 for (i
= store_info
->begin
; i
< store_info
->end
; i
++)
3023 int index
= get_bitmap_index (group_info
, i
);
3026 fprintf (dump_file
, "i = %d, index = %d\n", (int)i
, index
);
3027 if (index
== 0 || !bitmap_bit_p (v
, index
))
3030 fprintf (dump_file
, "failing at i = %d\n", (int)i
);
3040 check_for_inc_dec (insn_info
->insn
);
3041 delete_insn (insn_info
->insn
);
3042 insn_info
->insn
= NULL
;
3047 /* We do want to process the local info if the insn was
3048 deleted. For instance, if the insn did a wild read, we
3049 no longer need to trash the info. */
3051 && INSN_P (insn_info
->insn
)
3054 scan_stores_nospill (insn_info
->store_rec
, v
, NULL
);
3055 if (insn_info
->wild_read
)
3058 fprintf (dump_file
, "wild read\n");
3061 else if (insn_info
->read_rec
)
3064 fprintf (dump_file
, "regular read\n");
3065 scan_reads_nospill (insn_info
, v
, NULL
);
3069 insn_info
= insn_info
->prev_insn
;
3076 dse_step5_spill (void)
3081 bb_info_t bb_info
= bb_table
[bb
->index
];
3082 insn_info_t insn_info
= bb_info
->last_insn
;
3083 bitmap v
= bb_info
->out
;
3087 bool deleted
= false;
3088 /* There may have been code deleted by the dce pass run before
3091 && INSN_P (insn_info
->insn
)
3092 && (!insn_info
->cannot_delete
)
3093 && (!bitmap_empty_p (v
)))
3095 /* Try to delete the current insn. */
3096 store_info_t store_info
= insn_info
->store_rec
;
3101 if (store_info
->alias_set
)
3103 int index
= get_bitmap_index (clear_alias_group
,
3104 store_info
->alias_set
);
3105 if (index
== 0 || !bitmap_bit_p (v
, index
))
3113 store_info
= store_info
->next
;
3115 if (deleted
&& dbg_cnt (dse
))
3118 fprintf (dump_file
, "Spill deleting insn %d\n",
3119 INSN_UID (insn_info
->insn
));
3120 check_for_inc_dec (insn_info
->insn
);
3121 delete_insn (insn_info
->insn
);
3123 insn_info
->insn
= NULL
;
3128 && INSN_P (insn_info
->insn
)
3131 scan_stores_spill (insn_info
->store_rec
, v
, NULL
);
3132 scan_reads_spill (insn_info
->read_rec
, v
, NULL
);
3135 insn_info
= insn_info
->prev_insn
;
3142 /*----------------------------------------------------------------------------
3145 Destroy everything left standing.
3146 ----------------------------------------------------------------------------*/
3149 dse_step6 (bool global_done
)
3157 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
3159 free (group
->offset_map_n
);
3160 free (group
->offset_map_p
);
3161 BITMAP_FREE (group
->store1_n
);
3162 BITMAP_FREE (group
->store1_p
);
3163 BITMAP_FREE (group
->store2_n
);
3164 BITMAP_FREE (group
->store2_p
);
3165 BITMAP_FREE (group
->group_kill
);
3170 bb_info_t bb_info
= bb_table
[bb
->index
];
3171 BITMAP_FREE (bb_info
->gen
);
3173 BITMAP_FREE (bb_info
->kill
);
3175 BITMAP_FREE (bb_info
->in
);
3177 BITMAP_FREE (bb_info
->out
);
3182 for (i
= 0; VEC_iterate (group_info_t
, rtx_group_vec
, i
, group
); i
++)
3184 BITMAP_FREE (group
->store1_n
);
3185 BITMAP_FREE (group
->store1_p
);
3186 BITMAP_FREE (group
->store2_n
);
3187 BITMAP_FREE (group
->store2_p
);
3188 BITMAP_FREE (group
->group_kill
);
3192 if (clear_alias_sets
)
3194 BITMAP_FREE (clear_alias_sets
);
3195 BITMAP_FREE (disqualified_clear_alias_sets
);
3196 free_alloc_pool (clear_alias_mode_pool
);
3197 htab_delete (clear_alias_mode_table
);
3200 end_alias_analysis ();
3202 htab_delete (rtx_group_table
);
3203 VEC_free (group_info_t
, heap
, rtx_group_vec
);
3204 BITMAP_FREE (all_blocks
);
3205 BITMAP_FREE (scratch
);
3207 free_alloc_pool (rtx_store_info_pool
);
3208 free_alloc_pool (read_info_pool
);
3209 free_alloc_pool (insn_info_pool
);
3210 free_alloc_pool (bb_info_pool
);
3211 free_alloc_pool (rtx_group_info_pool
);
3212 free_alloc_pool (deferred_change_pool
);
3217 /* -------------------------------------------------------------------------
3219 ------------------------------------------------------------------------- */
3221 /* Callback for running pass_rtl_dse. */
3224 rest_of_handle_dse (void)
3226 bool did_global
= false;
3228 df_set_flags (DF_DEFER_INSN_RESCAN
);
3233 if (dse_step2_nospill ())
3235 df_set_flags (DF_LR_RUN_DCE
);
3239 fprintf (dump_file
, "doing global processing\n");
3242 dse_step5_nospill ();
3245 /* For the instance of dse that runs after reload, we make a special
3246 pass to process the spills. These are special in that they are
3247 totally transparent, i.e, there is no aliasing issues that need
3248 to be considered. This means that the wild reads that kill
3249 everything else do not apply here. */
3250 if (clear_alias_sets
&& dse_step2_spill ())
3254 df_set_flags (DF_LR_RUN_DCE
);
3259 fprintf (dump_file
, "doing global spill processing\n");
3265 dse_step6 (did_global
);
3268 fprintf (dump_file
, "dse: local deletions = %d, global deletions = %d, spill deletions = %d\n",
3269 locally_deleted
, globally_deleted
, spill_deleted
);
3276 return gate_dse1 () || gate_dse2 ();
3282 return optimize
> 0 && flag_dse
3289 return optimize
> 0 && flag_dse
3293 struct rtl_opt_pass pass_rtl_dse1
=
3298 gate_dse1
, /* gate */
3299 rest_of_handle_dse
, /* execute */
3302 0, /* static_pass_number */
3303 TV_DSE1
, /* tv_id */
3304 0, /* properties_required */
3305 0, /* properties_provided */
3306 0, /* properties_destroyed */
3307 0, /* todo_flags_start */
3309 TODO_df_finish
| TODO_verify_rtl_sharing
|
3310 TODO_ggc_collect
/* todo_flags_finish */
3314 struct rtl_opt_pass pass_rtl_dse2
=
3319 gate_dse2
, /* gate */
3320 rest_of_handle_dse
, /* execute */
3323 0, /* static_pass_number */
3324 TV_DSE2
, /* tv_id */
3325 0, /* properties_required */
3326 0, /* properties_provided */
3327 0, /* properties_destroyed */
3328 0, /* todo_flags_start */
3330 TODO_df_finish
| TODO_verify_rtl_sharing
|
3331 TODO_ggc_collect
/* todo_flags_finish */