1 /* Post reload partially redundant load elimination
2 Copyright (C) 2004, 2005, 2006, 2007, 2008, 2010, 2011
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
23 #include "coretypes.h"
25 #include "diagnostic-core.h"
31 #include "hard-reg-set.h"
33 #include "insn-config.h"
35 #include "basic-block.h"
44 #include "tree-pass.h"
47 /* The following code implements gcse after reload, the purpose of this
48 pass is to cleanup redundant loads generated by reload and other
49 optimizations that come after gcse. It searches for simple inter-block
50 redundancies and tries to eliminate them by adding moves and loads
53 Perform partially redundant load elimination, try to eliminate redundant
54 loads created by the reload pass. We try to look for full or partial
55 redundant loads fed by one or more loads/stores in predecessor BBs,
56 and try adding loads to make them fully redundant. We also check if
57 it's worth adding loads to be able to delete the redundant load.
60 1. Build available expressions hash table:
61 For each load/store instruction, if the loaded/stored memory didn't
62 change until the end of the basic block add this memory expression to
64 2. Perform Redundancy elimination:
65 For each load instruction do the following:
66 perform partial redundancy elimination, check if it's worth adding
67 loads to make the load fully redundant. If so add loads and
68 register copies and delete the load.
69 3. Delete instructions made redundant in step 2.
72 If the loaded register is used/defined between load and some store,
73 look for some other free register between load and all its stores,
74 and replace the load with a copy from this register to the loaded
79 /* Keep statistics of this pass. */
87 /* We need to keep a hash table of expressions. The table entries are of
88 type 'struct expr', and for each expression there is a single linked
89 list of occurrences. */
91 /* The table itself. */
92 static htab_t expr_table
;
94 /* Expression elements in the hash table. */
97 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
100 /* The same hash for this entry. */
103 /* List of available occurrence in basic blocks in the function. */
104 struct occr
*avail_occr
;
107 static struct obstack expr_obstack
;
109 /* Occurrence of an expression.
110 There is at most one occurrence per basic block. If a pattern appears
111 more than once, the last appearance is used. */
115 /* Next occurrence of this expression. */
117 /* The insn that computes the expression. */
119 /* Nonzero if this [anticipatable] occurrence has been deleted. */
123 static struct obstack occr_obstack
;
125 /* The following structure holds the information about the occurrences of
126 the redundant instructions. */
134 static struct obstack unoccr_obstack
;
136 /* Array where each element is the CUID if the insn that last set the hard
137 register with the number of the element, since the start of the current
140 This array is used during the building of the hash table (step 1) to
141 determine if a reg is killed before the end of a basic block.
143 It is also used when eliminating partial redundancies (step 2) to see
144 if a reg was modified since the start of a basic block. */
145 static int *reg_avail_info
;
147 /* A list of insns that may modify memory within the current basic block. */
151 struct modifies_mem
*next
;
153 static struct modifies_mem
*modifies_mem_list
;
155 /* The modifies_mem structs also go on an obstack, only this obstack is
156 freed each time after completing the analysis or transformations on
157 a basic block. So we allocate a dummy modifies_mem_obstack_bottom
158 object on the obstack to keep track of the bottom of the obstack. */
159 static struct obstack modifies_mem_obstack
;
160 static struct modifies_mem
*modifies_mem_obstack_bottom
;
162 /* Mapping of insn UIDs to CUIDs.
163 CUIDs are like UIDs except they increase monotonically in each basic
164 block, have no gaps, and only apply to real insns. */
165 static int *uid_cuid
;
166 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
169 /* Helpers for memory allocation/freeing. */
170 static void alloc_mem (void);
171 static void free_mem (void);
173 /* Support for hash table construction and transformations. */
174 static bool oprs_unchanged_p (rtx
, rtx
, bool);
175 static void record_last_reg_set_info (rtx
, rtx
);
176 static void record_last_reg_set_info_regno (rtx
, int);
177 static void record_last_mem_set_info (rtx
);
178 static void record_last_set_info (rtx
, const_rtx
, void *);
179 static void record_opr_changes (rtx
);
181 static void find_mem_conflicts (rtx
, const_rtx
, void *);
182 static int load_killed_in_block_p (int, rtx
, bool);
183 static void reset_opr_set_tables (void);
185 /* Hash table support. */
186 static hashval_t
hash_expr (rtx
, int *);
187 static hashval_t
hash_expr_for_htab (const void *);
188 static int expr_equiv_p (const void *, const void *);
189 static void insert_expr_in_table (rtx
, rtx
);
190 static struct expr
*lookup_expr_in_table (rtx
);
191 static int dump_hash_table_entry (void **, void *);
192 static void dump_hash_table (FILE *);
194 /* Helpers for eliminate_partially_redundant_load. */
195 static bool reg_killed_on_edge (rtx
, edge
);
196 static bool reg_used_on_edge (rtx
, edge
);
198 static rtx
get_avail_load_store_reg (rtx
);
200 static bool bb_has_well_behaved_predecessors (basic_block
);
201 static struct occr
* get_bb_avail_insn (basic_block
, struct occr
*);
202 static void hash_scan_set (rtx
);
203 static void compute_hash_table (void);
205 /* The work horses of this pass. */
206 static void eliminate_partially_redundant_load (basic_block
,
209 static void eliminate_partially_redundant_loads (void);
212 /* Allocate memory for the CUID mapping array and register/memory
222 /* Find the largest UID and create a mapping from UIDs to CUIDs. */
223 uid_cuid
= XCNEWVEC (int, get_max_uid () + 1);
226 FOR_BB_INSNS (bb
, insn
)
229 uid_cuid
[INSN_UID (insn
)] = i
++;
231 uid_cuid
[INSN_UID (insn
)] = i
;
234 /* Allocate the available expressions hash table. We don't want to
235 make the hash table too small, but unnecessarily making it too large
236 also doesn't help. The i/4 is a gcse.c relic, and seems like a
237 reasonable choice. */
238 expr_table
= htab_create (MAX (i
/ 4, 13),
239 hash_expr_for_htab
, expr_equiv_p
, NULL
);
241 /* We allocate everything on obstacks because we often can roll back
242 the whole obstack to some point. Freeing obstacks is very fast. */
243 gcc_obstack_init (&expr_obstack
);
244 gcc_obstack_init (&occr_obstack
);
245 gcc_obstack_init (&unoccr_obstack
);
246 gcc_obstack_init (&modifies_mem_obstack
);
248 /* Working array used to track the last set for each register
249 in the current block. */
250 reg_avail_info
= (int *) xmalloc (FIRST_PSEUDO_REGISTER
* sizeof (int));
252 /* Put a dummy modifies_mem object on the modifies_mem_obstack, so we
253 can roll it back in reset_opr_set_tables. */
254 modifies_mem_obstack_bottom
=
255 (struct modifies_mem
*) obstack_alloc (&modifies_mem_obstack
,
256 sizeof (struct modifies_mem
));
259 /* Free memory allocated by alloc_mem. */
266 htab_delete (expr_table
);
268 obstack_free (&expr_obstack
, NULL
);
269 obstack_free (&occr_obstack
, NULL
);
270 obstack_free (&unoccr_obstack
, NULL
);
271 obstack_free (&modifies_mem_obstack
, NULL
);
273 free (reg_avail_info
);
277 /* Hash expression X.
278 DO_NOT_RECORD_P is a boolean indicating if a volatile operand is found
279 or if the expression contains something we don't want to insert in the
283 hash_expr (rtx x
, int *do_not_record_p
)
285 *do_not_record_p
= 0;
286 return hash_rtx (x
, GET_MODE (x
), do_not_record_p
,
287 NULL
, /*have_reg_qty=*/false);
290 /* Callback for hashtab.
291 Return the hash value for expression EXP. We don't actually hash
292 here, we just return the cached hash value. */
295 hash_expr_for_htab (const void *expp
)
297 const struct expr
*const exp
= (const struct expr
*) expp
;
301 /* Callback for hashtab.
302 Return nonzero if exp1 is equivalent to exp2. */
305 expr_equiv_p (const void *exp1p
, const void *exp2p
)
307 const struct expr
*const exp1
= (const struct expr
*) exp1p
;
308 const struct expr
*const exp2
= (const struct expr
*) exp2p
;
309 int equiv_p
= exp_equiv_p (exp1
->expr
, exp2
->expr
, 0, true);
311 gcc_assert (!equiv_p
|| exp1
->hash
== exp2
->hash
);
316 /* Insert expression X in INSN in the hash TABLE.
317 If it is already present, record it as the last occurrence in INSN's
321 insert_expr_in_table (rtx x
, rtx insn
)
325 struct expr
*cur_expr
, **slot
;
326 struct occr
*avail_occr
, *last_occr
= NULL
;
328 hash
= hash_expr (x
, &do_not_record_p
);
330 /* Do not insert expression in the table if it contains volatile operands,
331 or if hash_expr determines the expression is something we don't want
332 to or can't handle. */
336 /* We anticipate that redundant expressions are rare, so for convenience
337 allocate a new hash table element here already and set its fields.
338 If we don't do this, we need a hack with a static struct expr. Anyway,
339 obstack_free is really fast and one more obstack_alloc doesn't hurt if
340 we're going to see more expressions later on. */
341 cur_expr
= (struct expr
*) obstack_alloc (&expr_obstack
,
342 sizeof (struct expr
));
344 cur_expr
->hash
= hash
;
345 cur_expr
->avail_occr
= NULL
;
347 slot
= (struct expr
**) htab_find_slot_with_hash (expr_table
, cur_expr
,
351 /* The expression isn't found, so insert it. */
355 /* The expression is already in the table, so roll back the
356 obstack and use the existing table entry. */
357 obstack_free (&expr_obstack
, cur_expr
);
361 /* Search for another occurrence in the same basic block. */
362 avail_occr
= cur_expr
->avail_occr
;
364 && BLOCK_FOR_INSN (avail_occr
->insn
) != BLOCK_FOR_INSN (insn
))
366 /* If an occurrence isn't found, save a pointer to the end of
368 last_occr
= avail_occr
;
369 avail_occr
= avail_occr
->next
;
373 /* Found another instance of the expression in the same basic block.
374 Prefer this occurrence to the currently recorded one. We want
375 the last one in the block and the block is scanned from start
377 avail_occr
->insn
= insn
;
380 /* First occurrence of this expression in this basic block. */
381 avail_occr
= (struct occr
*) obstack_alloc (&occr_obstack
,
382 sizeof (struct occr
));
384 /* First occurrence of this expression in any block? */
385 if (cur_expr
->avail_occr
== NULL
)
386 cur_expr
->avail_occr
= avail_occr
;
388 last_occr
->next
= avail_occr
;
390 avail_occr
->insn
= insn
;
391 avail_occr
->next
= NULL
;
392 avail_occr
->deleted_p
= 0;
397 /* Lookup pattern PAT in the expression hash table.
398 The result is a pointer to the table entry, or NULL if not found. */
401 lookup_expr_in_table (rtx pat
)
404 struct expr
**slot
, *tmp_expr
;
405 hashval_t hash
= hash_expr (pat
, &do_not_record_p
);
410 tmp_expr
= (struct expr
*) obstack_alloc (&expr_obstack
,
411 sizeof (struct expr
));
412 tmp_expr
->expr
= pat
;
413 tmp_expr
->hash
= hash
;
414 tmp_expr
->avail_occr
= NULL
;
416 slot
= (struct expr
**) htab_find_slot_with_hash (expr_table
, tmp_expr
,
418 obstack_free (&expr_obstack
, tmp_expr
);
427 /* Dump all expressions and occurrences that are currently in the
428 expression hash table to FILE. */
430 /* This helper is called via htab_traverse. */
432 dump_hash_table_entry (void **slot
, void *filep
)
434 struct expr
*expr
= (struct expr
*) *slot
;
435 FILE *file
= (FILE *) filep
;
438 fprintf (file
, "expr: ");
439 print_rtl (file
, expr
->expr
);
440 fprintf (file
,"\nhashcode: %u\n", expr
->hash
);
441 fprintf (file
,"list of occurrences:\n");
442 occr
= expr
->avail_occr
;
445 rtx insn
= occr
->insn
;
446 print_rtl_single (file
, insn
);
447 fprintf (file
, "\n");
450 fprintf (file
, "\n");
455 dump_hash_table (FILE *file
)
457 fprintf (file
, "\n\nexpression hash table\n");
458 fprintf (file
, "size %ld, %ld elements, %f collision/search ratio\n",
459 (long) htab_size (expr_table
),
460 (long) htab_elements (expr_table
),
461 htab_collisions (expr_table
));
462 if (htab_elements (expr_table
) > 0)
464 fprintf (file
, "\n\ntable entries:\n");
465 htab_traverse (expr_table
, dump_hash_table_entry
, file
);
467 fprintf (file
, "\n");
470 /* Return true if register X is recorded as being set by an instruction
471 whose CUID is greater than the one given. */
474 reg_changed_after_insn_p (rtx x
, int cuid
)
476 unsigned int regno
, end_regno
;
479 end_regno
= END_HARD_REGNO (x
);
481 if (reg_avail_info
[regno
] > cuid
)
483 while (++regno
< end_regno
);
487 /* Return nonzero if the operands of expression X are unchanged
488 1) from the start of INSN's basic block up to but not including INSN
489 if AFTER_INSN is false, or
490 2) from INSN to the end of INSN's basic block if AFTER_INSN is true. */
493 oprs_unchanged_p (rtx x
, rtx insn
, bool after_insn
)
506 /* We are called after register allocation. */
507 gcc_assert (REGNO (x
) < FIRST_PSEUDO_REGISTER
);
509 return !reg_changed_after_insn_p (x
, INSN_CUID (insn
) - 1);
511 return !reg_changed_after_insn_p (x
, 0);
514 if (load_killed_in_block_p (INSN_CUID (insn
), x
, after_insn
))
517 return oprs_unchanged_p (XEXP (x
, 0), insn
, after_insn
);
543 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
547 if (! oprs_unchanged_p (XEXP (x
, i
), insn
, after_insn
))
550 else if (fmt
[i
] == 'E')
551 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
552 if (! oprs_unchanged_p (XVECEXP (x
, i
, j
), insn
, after_insn
))
560 /* Used for communication between find_mem_conflicts and
561 load_killed_in_block_p. Nonzero if find_mem_conflicts finds a
562 conflict between two memory references.
563 This is a bit of a hack to work around the limitations of note_stores. */
564 static int mems_conflict_p
;
566 /* DEST is the output of an instruction. If it is a memory reference, and
567 possibly conflicts with the load found in DATA, then set mems_conflict_p
568 to a nonzero value. */
571 find_mem_conflicts (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
,
574 rtx mem_op
= (rtx
) data
;
576 while (GET_CODE (dest
) == SUBREG
577 || GET_CODE (dest
) == ZERO_EXTRACT
578 || GET_CODE (dest
) == STRICT_LOW_PART
)
579 dest
= XEXP (dest
, 0);
581 /* If DEST is not a MEM, then it will not conflict with the load. Note
582 that function calls are assumed to clobber memory, but are handled
587 if (true_dependence (dest
, GET_MODE (dest
), mem_op
))
592 /* Return nonzero if the expression in X (a memory reference) is killed
593 in the current basic block before (if AFTER_INSN is false) or after
594 (if AFTER_INSN is true) the insn with the CUID in UID_LIMIT.
596 This function assumes that the modifies_mem table is flushed when
597 the hash table construction or redundancy elimination phases start
598 processing a new basic block. */
601 load_killed_in_block_p (int uid_limit
, rtx x
, bool after_insn
)
603 struct modifies_mem
*list_entry
= modifies_mem_list
;
607 rtx setter
= list_entry
->insn
;
609 /* Ignore entries in the list that do not apply. */
611 && INSN_CUID (setter
) < uid_limit
)
613 && INSN_CUID (setter
) > uid_limit
))
615 list_entry
= list_entry
->next
;
619 /* If SETTER is a call everything is clobbered. Note that calls
620 to pure functions are never put on the list, so we need not
625 /* SETTER must be an insn of some kind that sets memory. Call
626 note_stores to examine each hunk of memory that is modified.
627 It will set mems_conflict_p to nonzero if there may be a
628 conflict between X and SETTER. */
630 note_stores (PATTERN (setter
), find_mem_conflicts
, x
);
634 list_entry
= list_entry
->next
;
640 /* Record register first/last/block set information for REGNO in INSN. */
643 record_last_reg_set_info (rtx insn
, rtx reg
)
645 unsigned int regno
, end_regno
;
648 end_regno
= END_HARD_REGNO (reg
);
650 reg_avail_info
[regno
] = INSN_CUID (insn
);
651 while (++regno
< end_regno
);
655 record_last_reg_set_info_regno (rtx insn
, int regno
)
657 reg_avail_info
[regno
] = INSN_CUID (insn
);
661 /* Record memory modification information for INSN. We do not actually care
662 about the memory location(s) that are set, or even how they are set (consider
663 a CALL_INSN). We merely need to record which insns modify memory. */
666 record_last_mem_set_info (rtx insn
)
668 struct modifies_mem
*list_entry
;
670 list_entry
= (struct modifies_mem
*) obstack_alloc (&modifies_mem_obstack
,
671 sizeof (struct modifies_mem
));
672 list_entry
->insn
= insn
;
673 list_entry
->next
= modifies_mem_list
;
674 modifies_mem_list
= list_entry
;
677 /* Called from compute_hash_table via note_stores to handle one
678 SET or CLOBBER in an insn. DATA is really the instruction in which
679 the SET is taking place. */
682 record_last_set_info (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
, void *data
)
684 rtx last_set_insn
= (rtx
) data
;
686 if (GET_CODE (dest
) == SUBREG
)
687 dest
= SUBREG_REG (dest
);
690 record_last_reg_set_info (last_set_insn
, dest
);
691 else if (MEM_P (dest
))
693 /* Ignore pushes, they don't clobber memory. They may still
694 clobber the stack pointer though. Some targets do argument
695 pushes without adding REG_INC notes. See e.g. PR25196,
696 where a pushsi2 on i386 doesn't have REG_INC notes. Note
697 such changes here too. */
698 if (! push_operand (dest
, GET_MODE (dest
)))
699 record_last_mem_set_info (last_set_insn
);
701 record_last_reg_set_info_regno (last_set_insn
, STACK_POINTER_REGNUM
);
706 /* Reset tables used to keep track of what's still available since the
707 start of the block. */
710 reset_opr_set_tables (void)
712 memset (reg_avail_info
, 0, FIRST_PSEUDO_REGISTER
* sizeof (int));
713 obstack_free (&modifies_mem_obstack
, modifies_mem_obstack_bottom
);
714 modifies_mem_list
= NULL
;
718 /* Record things set by INSN.
719 This data is used by oprs_unchanged_p. */
722 record_opr_changes (rtx insn
)
726 /* Find all stores and record them. */
727 note_stores (PATTERN (insn
), record_last_set_info
, insn
);
729 /* Also record autoincremented REGs for this insn as changed. */
730 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
731 if (REG_NOTE_KIND (note
) == REG_INC
)
732 record_last_reg_set_info (insn
, XEXP (note
, 0));
734 /* Finally, if this is a call, record all call clobbers. */
740 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
741 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
742 record_last_reg_set_info_regno (insn
, regno
);
744 for (link
= CALL_INSN_FUNCTION_USAGE (insn
); link
; link
= XEXP (link
, 1))
745 if (GET_CODE (XEXP (link
, 0)) == CLOBBER
)
747 x
= XEXP (XEXP (link
, 0), 0);
750 gcc_assert (HARD_REGISTER_P (x
));
751 record_last_reg_set_info (insn
, x
);
755 if (! RTL_CONST_OR_PURE_CALL_P (insn
))
756 record_last_mem_set_info (insn
);
761 /* Scan the pattern of INSN and add an entry to the hash TABLE.
762 After reload we are interested in loads/stores only. */
765 hash_scan_set (rtx insn
)
767 rtx pat
= PATTERN (insn
);
768 rtx src
= SET_SRC (pat
);
769 rtx dest
= SET_DEST (pat
);
771 /* We are only interested in loads and stores. */
772 if (! MEM_P (src
) && ! MEM_P (dest
))
775 /* Don't mess with jumps and nops. */
776 if (JUMP_P (insn
) || set_noop_p (pat
))
781 if (/* Don't CSE something if we can't do a reg/reg copy. */
782 can_copy_p (GET_MODE (dest
))
783 /* Is SET_SRC something we want to gcse? */
784 && general_operand (src
, GET_MODE (src
))
786 /* Never consider insns touching the register stack. It may
787 create situations that reg-stack cannot handle (e.g. a stack
788 register live across an abnormal edge). */
789 && (REGNO (dest
) < FIRST_STACK_REG
|| REGNO (dest
) > LAST_STACK_REG
)
791 /* An expression is not available if its operands are
792 subsequently modified, including this insn. */
793 && oprs_unchanged_p (src
, insn
, true))
795 insert_expr_in_table (src
, insn
);
798 else if (REG_P (src
))
800 /* Only record sets of pseudo-regs in the hash table. */
801 if (/* Don't CSE something if we can't do a reg/reg copy. */
802 can_copy_p (GET_MODE (src
))
803 /* Is SET_DEST something we want to gcse? */
804 && general_operand (dest
, GET_MODE (dest
))
806 /* As above for STACK_REGS. */
807 && (REGNO (src
) < FIRST_STACK_REG
|| REGNO (src
) > LAST_STACK_REG
)
809 && ! (flag_float_store
&& FLOAT_MODE_P (GET_MODE (dest
)))
810 /* Check if the memory expression is killed after insn. */
811 && ! load_killed_in_block_p (INSN_CUID (insn
) + 1, dest
, true)
812 && oprs_unchanged_p (XEXP (dest
, 0), insn
, true))
814 insert_expr_in_table (dest
, insn
);
820 /* Create hash table of memory expressions available at end of basic
821 blocks. Basically you should think of this hash table as the
822 representation of AVAIL_OUT. This is the set of expressions that
823 is generated in a basic block and not killed before the end of the
824 same basic block. Notice that this is really a local computation. */
827 compute_hash_table (void)
835 /* First pass over the instructions records information used to
836 determine when registers and memory are last set.
837 Since we compute a "local" AVAIL_OUT, reset the tables that
838 help us keep track of what has been modified since the start
840 reset_opr_set_tables ();
841 FOR_BB_INSNS (bb
, insn
)
844 record_opr_changes (insn
);
847 /* The next pass actually builds the hash table. */
848 FOR_BB_INSNS (bb
, insn
)
849 if (INSN_P (insn
) && GET_CODE (PATTERN (insn
)) == SET
)
850 hash_scan_set (insn
);
855 /* Check if register REG is killed in any insn waiting to be inserted on
856 edge E. This function is required to check that our data flow analysis
857 is still valid prior to commit_edge_insertions. */
860 reg_killed_on_edge (rtx reg
, edge e
)
864 for (insn
= e
->insns
.r
; insn
; insn
= NEXT_INSN (insn
))
865 if (INSN_P (insn
) && reg_set_p (reg
, insn
))
871 /* Similar to above - check if register REG is used in any insn waiting
872 to be inserted on edge E.
873 Assumes no such insn can be a CALL_INSN; if so call reg_used_between_p
874 with PREV(insn),NEXT(insn) instead of calling reg_overlap_mentioned_p. */
877 reg_used_on_edge (rtx reg
, edge e
)
881 for (insn
= e
->insns
.r
; insn
; insn
= NEXT_INSN (insn
))
882 if (INSN_P (insn
) && reg_overlap_mentioned_p (reg
, PATTERN (insn
)))
888 /* Return the loaded/stored register of a load/store instruction. */
891 get_avail_load_store_reg (rtx insn
)
893 if (REG_P (SET_DEST (PATTERN (insn
))))
895 return SET_DEST(PATTERN(insn
));
899 gcc_assert (REG_P (SET_SRC (PATTERN (insn
))));
900 return SET_SRC (PATTERN (insn
));
904 /* Return nonzero if the predecessors of BB are "well behaved". */
907 bb_has_well_behaved_predecessors (basic_block bb
)
912 if (EDGE_COUNT (bb
->preds
) == 0)
915 FOR_EACH_EDGE (pred
, ei
, bb
->preds
)
917 if ((pred
->flags
& EDGE_ABNORMAL
) && EDGE_CRITICAL_P (pred
))
920 if ((pred
->flags
& EDGE_ABNORMAL_CALL
) && cfun
->has_nonlocal_label
)
923 if (JUMP_TABLE_DATA_P (BB_END (pred
->src
)))
930 /* Search for the occurrences of expression in BB. */
933 get_bb_avail_insn (basic_block bb
, struct occr
*occr
)
935 for (; occr
!= NULL
; occr
= occr
->next
)
936 if (BLOCK_FOR_INSN (occr
->insn
) == bb
)
942 /* This handles the case where several stores feed a partially redundant
943 load. It checks if the redundancy elimination is possible and if it's
946 Redundancy elimination is possible if,
947 1) None of the operands of an insn have been modified since the start
948 of the current basic block.
949 2) In any predecessor of the current basic block, the same expression
952 See the function body for the heuristics that determine if eliminating
953 a redundancy is also worth doing, assuming it is possible. */
956 eliminate_partially_redundant_load (basic_block bb
, rtx insn
,
960 rtx avail_insn
= NULL_RTX
;
964 struct unoccr
*occr
, *avail_occrs
= NULL
;
965 struct unoccr
*unoccr
, *unavail_occrs
= NULL
, *rollback_unoccr
= NULL
;
967 gcov_type ok_count
= 0; /* Redundant load execution count. */
968 gcov_type critical_count
= 0; /* Execution count of critical edges. */
970 bool critical_edge_split
= false;
972 /* The execution count of the loads to be added to make the
973 load fully redundant. */
974 gcov_type not_ok_count
= 0;
977 pat
= PATTERN (insn
);
978 dest
= SET_DEST (pat
);
980 /* Check that the loaded register is not used, set, or killed from the
981 beginning of the block. */
982 if (reg_changed_after_insn_p (dest
, 0)
983 || reg_used_between_p (dest
, PREV_INSN (BB_HEAD (bb
)), insn
))
986 /* Check potential for replacing load with copy for predecessors. */
987 FOR_EACH_EDGE (pred
, ei
, bb
->preds
)
989 rtx next_pred_bb_end
;
991 avail_insn
= NULL_RTX
;
992 avail_reg
= NULL_RTX
;
994 next_pred_bb_end
= NEXT_INSN (BB_END (pred_bb
));
995 for (a_occr
= get_bb_avail_insn (pred_bb
, expr
->avail_occr
); a_occr
;
996 a_occr
= get_bb_avail_insn (pred_bb
, a_occr
->next
))
998 /* Check if the loaded register is not used. */
999 avail_insn
= a_occr
->insn
;
1000 avail_reg
= get_avail_load_store_reg (avail_insn
);
1001 gcc_assert (avail_reg
);
1003 /* Make sure we can generate a move from register avail_reg to
1005 extract_insn (gen_move_insn (copy_rtx (dest
),
1006 copy_rtx (avail_reg
)));
1007 if (! constrain_operands (1)
1008 || reg_killed_on_edge (avail_reg
, pred
)
1009 || reg_used_on_edge (dest
, pred
))
1014 if (!reg_set_between_p (avail_reg
, avail_insn
, next_pred_bb_end
))
1015 /* AVAIL_INSN remains non-null. */
1021 if (EDGE_CRITICAL_P (pred
))
1022 critical_count
+= pred
->count
;
1024 if (avail_insn
!= NULL_RTX
)
1027 ok_count
+= pred
->count
;
1028 if (! set_noop_p (PATTERN (gen_move_insn (copy_rtx (dest
),
1029 copy_rtx (avail_reg
)))))
1031 /* Check if there is going to be a split. */
1032 if (EDGE_CRITICAL_P (pred
))
1033 critical_edge_split
= true;
1035 else /* Its a dead move no need to generate. */
1037 occr
= (struct unoccr
*) obstack_alloc (&unoccr_obstack
,
1038 sizeof (struct unoccr
));
1039 occr
->insn
= avail_insn
;
1041 occr
->next
= avail_occrs
;
1043 if (! rollback_unoccr
)
1044 rollback_unoccr
= occr
;
1048 /* Adding a load on a critical edge will cause a split. */
1049 if (EDGE_CRITICAL_P (pred
))
1050 critical_edge_split
= true;
1051 not_ok_count
+= pred
->count
;
1052 unoccr
= (struct unoccr
*) obstack_alloc (&unoccr_obstack
,
1053 sizeof (struct unoccr
));
1054 unoccr
->insn
= NULL_RTX
;
1055 unoccr
->pred
= pred
;
1056 unoccr
->next
= unavail_occrs
;
1057 unavail_occrs
= unoccr
;
1058 if (! rollback_unoccr
)
1059 rollback_unoccr
= unoccr
;
1063 if (/* No load can be replaced by copy. */
1065 /* Prevent exploding the code. */
1066 || (optimize_bb_for_size_p (bb
) && npred_ok
> 1)
1067 /* If we don't have profile information we cannot tell if splitting
1068 a critical edge is profitable or not so don't do it. */
1069 || ((! profile_info
|| ! flag_branch_probabilities
1070 || targetm
.cannot_modify_jumps_p ())
1071 && critical_edge_split
))
1074 /* Check if it's worth applying the partial redundancy elimination. */
1075 if (ok_count
< GCSE_AFTER_RELOAD_PARTIAL_FRACTION
* not_ok_count
)
1077 if (ok_count
< GCSE_AFTER_RELOAD_CRITICAL_FRACTION
* critical_count
)
1080 /* Generate moves to the loaded register from where
1081 the memory is available. */
1082 for (occr
= avail_occrs
; occr
; occr
= occr
->next
)
1084 avail_insn
= occr
->insn
;
1086 /* Set avail_reg to be the register having the value of the
1088 avail_reg
= get_avail_load_store_reg (avail_insn
);
1089 gcc_assert (avail_reg
);
1091 insert_insn_on_edge (gen_move_insn (copy_rtx (dest
),
1092 copy_rtx (avail_reg
)),
1094 stats
.moves_inserted
++;
1098 "generating move from %d to %d on edge from %d to %d\n",
1105 /* Regenerate loads where the memory is unavailable. */
1106 for (unoccr
= unavail_occrs
; unoccr
; unoccr
= unoccr
->next
)
1108 pred
= unoccr
->pred
;
1109 insert_insn_on_edge (copy_insn (PATTERN (insn
)), pred
);
1110 stats
.copies_inserted
++;
1115 "generating on edge from %d to %d a copy of load: ",
1118 print_rtl (dump_file
, PATTERN (insn
));
1119 fprintf (dump_file
, "\n");
1123 /* Delete the insn if it is not available in this block and mark it
1124 for deletion if it is available. If insn is available it may help
1125 discover additional redundancies, so mark it for later deletion. */
1126 for (a_occr
= get_bb_avail_insn (bb
, expr
->avail_occr
);
1127 a_occr
&& (a_occr
->insn
!= insn
);
1128 a_occr
= get_bb_avail_insn (bb
, a_occr
->next
))
1133 stats
.insns_deleted
++;
1137 fprintf (dump_file
, "deleting insn:\n");
1138 print_rtl_single (dump_file
, insn
);
1139 fprintf (dump_file
, "\n");
1144 a_occr
->deleted_p
= 1;
1147 if (rollback_unoccr
)
1148 obstack_free (&unoccr_obstack
, rollback_unoccr
);
1151 /* Performing the redundancy elimination as described before. */
1154 eliminate_partially_redundant_loads (void)
1159 /* Note we start at block 1. */
1161 if (ENTRY_BLOCK_PTR
->next_bb
== EXIT_BLOCK_PTR
)
1165 ENTRY_BLOCK_PTR
->next_bb
->next_bb
,
1169 /* Don't try anything on basic blocks with strange predecessors. */
1170 if (! bb_has_well_behaved_predecessors (bb
))
1173 /* Do not try anything on cold basic blocks. */
1174 if (optimize_bb_for_size_p (bb
))
1177 /* Reset the table of things changed since the start of the current
1179 reset_opr_set_tables ();
1181 /* Look at all insns in the current basic block and see if there are
1182 any loads in it that we can record. */
1183 FOR_BB_INSNS (bb
, insn
)
1185 /* Is it a load - of the form (set (reg) (mem))? */
1186 if (NONJUMP_INSN_P (insn
)
1187 && GET_CODE (PATTERN (insn
)) == SET
1188 && REG_P (SET_DEST (PATTERN (insn
)))
1189 && MEM_P (SET_SRC (PATTERN (insn
))))
1191 rtx pat
= PATTERN (insn
);
1192 rtx src
= SET_SRC (pat
);
1195 if (!MEM_VOLATILE_P (src
)
1196 && GET_MODE (src
) != BLKmode
1197 && general_operand (src
, GET_MODE (src
))
1198 /* Are the operands unchanged since the start of the
1200 && oprs_unchanged_p (src
, insn
, false)
1201 && !(cfun
->can_throw_non_call_exceptions
&& may_trap_p (src
))
1202 && !side_effects_p (src
)
1203 /* Is the expression recorded? */
1204 && (expr
= lookup_expr_in_table (src
)) != NULL
)
1206 /* We now have a load (insn) and an available memory at
1207 its BB start (expr). Try to remove the loads if it is
1209 eliminate_partially_redundant_load (bb
, insn
, expr
);
1213 /* Keep track of everything modified by this insn, so that we
1214 know what has been modified since the start of the current
1217 record_opr_changes (insn
);
1221 commit_edge_insertions ();
1224 /* Go over the expression hash table and delete insns that were
1225 marked for later deletion. */
1227 /* This helper is called via htab_traverse. */
1229 delete_redundant_insns_1 (void **slot
, void *data ATTRIBUTE_UNUSED
)
1231 struct expr
*expr
= (struct expr
*) *slot
;
1234 for (occr
= expr
->avail_occr
; occr
!= NULL
; occr
= occr
->next
)
1236 if (occr
->deleted_p
&& dbg_cnt (gcse2_delete
))
1238 delete_insn (occr
->insn
);
1239 stats
.insns_deleted
++;
1243 fprintf (dump_file
, "deleting insn:\n");
1244 print_rtl_single (dump_file
, occr
->insn
);
1245 fprintf (dump_file
, "\n");
1254 delete_redundant_insns (void)
1256 htab_traverse (expr_table
, delete_redundant_insns_1
, NULL
);
1258 fprintf (dump_file
, "\n");
1261 /* Main entry point of the GCSE after reload - clean some redundant loads
1265 gcse_after_reload_main (rtx f ATTRIBUTE_UNUSED
)
1268 memset (&stats
, 0, sizeof (stats
));
1270 /* Allocate memory for this pass.
1271 Also computes and initializes the insns' CUIDs. */
1274 /* We need alias analysis. */
1275 init_alias_analysis ();
1277 compute_hash_table ();
1280 dump_hash_table (dump_file
);
1282 if (htab_elements (expr_table
) > 0)
1284 eliminate_partially_redundant_loads ();
1285 delete_redundant_insns ();
1289 fprintf (dump_file
, "GCSE AFTER RELOAD stats:\n");
1290 fprintf (dump_file
, "copies inserted: %d\n", stats
.copies_inserted
);
1291 fprintf (dump_file
, "moves inserted: %d\n", stats
.moves_inserted
);
1292 fprintf (dump_file
, "insns deleted: %d\n", stats
.insns_deleted
);
1293 fprintf (dump_file
, "\n\n");
1296 statistics_counter_event (cfun
, "copies inserted",
1297 stats
.copies_inserted
);
1298 statistics_counter_event (cfun
, "moves inserted",
1299 stats
.moves_inserted
);
1300 statistics_counter_event (cfun
, "insns deleted",
1301 stats
.insns_deleted
);
1304 /* We are finished with alias. */
1305 end_alias_analysis ();
1312 gate_handle_gcse2 (void)
1314 return (optimize
> 0 && flag_gcse_after_reload
1315 && optimize_function_for_speed_p (cfun
));
1320 rest_of_handle_gcse2 (void)
1322 gcse_after_reload_main (get_insns ());
1323 rebuild_jump_labels (get_insns ());
1327 struct rtl_opt_pass pass_gcse2
=
1332 gate_handle_gcse2
, /* gate */
1333 rest_of_handle_gcse2
, /* execute */
1336 0, /* static_pass_number */
1337 TV_GCSE_AFTER_RELOAD
, /* tv_id */
1338 0, /* properties_required */
1339 0, /* properties_provided */
1340 0, /* properties_destroyed */
1341 0, /* todo_flags_start */
1342 TODO_verify_rtl_sharing
1343 | TODO_verify_flow
| TODO_ggc_collect
/* todo_flags_finish */