1 /* Post reload partially redundant load elimination
2 Copyright (C) 2004-2016 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 3, or (at your option) any later
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
22 #include "coretypes.h"
30 #include "insn-config.h"
38 #include "tree-pass.h"
40 #include "gcse-common.h"
42 /* The following code implements gcse after reload, the purpose of this
43 pass is to cleanup redundant loads generated by reload and other
44 optimizations that come after gcse. It searches for simple inter-block
45 redundancies and tries to eliminate them by adding moves and loads
48 Perform partially redundant load elimination, try to eliminate redundant
49 loads created by the reload pass. We try to look for full or partial
50 redundant loads fed by one or more loads/stores in predecessor BBs,
51 and try adding loads to make them fully redundant. We also check if
52 it's worth adding loads to be able to delete the redundant load.
55 1. Build available expressions hash table:
56 For each load/store instruction, if the loaded/stored memory didn't
57 change until the end of the basic block add this memory expression to
59 2. Perform Redundancy elimination:
60 For each load instruction do the following:
61 perform partial redundancy elimination, check if it's worth adding
62 loads to make the load fully redundant. If so add loads and
63 register copies and delete the load.
64 3. Delete instructions made redundant in step 2.
67 If the loaded register is used/defined between load and some store,
68 look for some other free register between load and all its stores,
69 and replace the load with a copy from this register to the loaded
74 /* Keep statistics of this pass. */
82 /* We need to keep a hash table of expressions. The table entries are of
83 type 'struct expr', and for each expression there is a single linked
84 list of occurrences. */
86 /* Expression elements in the hash table. */
89 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
92 /* The same hash for this entry. */
95 /* Index in the transparent bitmaps. */
96 unsigned int bitmap_index
;
98 /* List of available occurrence in basic blocks in the function. */
99 struct occr
*avail_occr
;
102 /* Hashtable helpers. */
104 struct expr_hasher
: nofree_ptr_hash
<expr
>
106 static inline hashval_t
hash (const expr
*);
107 static inline bool equal (const expr
*, const expr
*);
111 /* Hash expression X.
112 DO_NOT_RECORD_P is a boolean indicating if a volatile operand is found
113 or if the expression contains something we don't want to insert in the
117 hash_expr (rtx x
, int *do_not_record_p
)
119 *do_not_record_p
= 0;
120 return hash_rtx (x
, GET_MODE (x
), do_not_record_p
,
121 NULL
, /*have_reg_qty=*/false);
124 /* Callback for hashtab.
125 Return the hash value for expression EXP. We don't actually hash
126 here, we just return the cached hash value. */
129 expr_hasher::hash (const expr
*exp
)
134 /* Callback for hashtab.
135 Return nonzero if exp1 is equivalent to exp2. */
138 expr_hasher::equal (const expr
*exp1
, const expr
*exp2
)
140 int equiv_p
= exp_equiv_p (exp1
->expr
, exp2
->expr
, 0, true);
142 gcc_assert (!equiv_p
|| exp1
->hash
== exp2
->hash
);
146 /* The table itself. */
147 static hash_table
<expr_hasher
> *expr_table
;
150 static struct obstack expr_obstack
;
152 /* Occurrence of an expression.
153 There is at most one occurrence per basic block. If a pattern appears
154 more than once, the last appearance is used. */
158 /* Next occurrence of this expression. */
160 /* The insn that computes the expression. */
162 /* Nonzero if this [anticipatable] occurrence has been deleted. */
166 static struct obstack occr_obstack
;
168 /* The following structure holds the information about the occurrences of
169 the redundant instructions. */
177 static struct obstack unoccr_obstack
;
179 /* Array where each element is the CUID if the insn that last set the hard
180 register with the number of the element, since the start of the current
183 This array is used during the building of the hash table (step 1) to
184 determine if a reg is killed before the end of a basic block.
186 It is also used when eliminating partial redundancies (step 2) to see
187 if a reg was modified since the start of a basic block. */
188 static int *reg_avail_info
;
190 /* A list of insns that may modify memory within the current basic block. */
194 struct modifies_mem
*next
;
196 static struct modifies_mem
*modifies_mem_list
;
198 /* The modifies_mem structs also go on an obstack, only this obstack is
199 freed each time after completing the analysis or transformations on
200 a basic block. So we allocate a dummy modifies_mem_obstack_bottom
201 object on the obstack to keep track of the bottom of the obstack. */
202 static struct obstack modifies_mem_obstack
;
203 static struct modifies_mem
*modifies_mem_obstack_bottom
;
205 /* Mapping of insn UIDs to CUIDs.
206 CUIDs are like UIDs except they increase monotonically in each basic
207 block, have no gaps, and only apply to real insns. */
208 static int *uid_cuid
;
209 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
211 /* Bitmap of blocks which have memory stores. */
212 static bitmap modify_mem_list_set
;
214 /* Bitmap of blocks which have calls. */
215 static bitmap blocks_with_calls
;
217 /* Vector indexed by block # with a list of all the insns that
218 modify memory within the block. */
219 static vec
<rtx_insn
*> *modify_mem_list
;
221 /* Vector indexed by block # with a canonicalized list of insns
222 that modify memory in the block. */
223 static vec
<modify_pair
> *canon_modify_mem_list
;
225 /* Vector of simple bitmaps indexed by block number. Each component sbitmap
226 indicates which expressions are transparent through the block. */
227 static sbitmap
*transp
;
230 /* Helpers for memory allocation/freeing. */
231 static void alloc_mem (void);
232 static void free_mem (void);
234 /* Support for hash table construction and transformations. */
235 static bool oprs_unchanged_p (rtx
, rtx_insn
*, bool);
236 static void record_last_reg_set_info (rtx_insn
*, rtx
);
237 static void record_last_reg_set_info_regno (rtx_insn
*, int);
238 static void record_last_mem_set_info (rtx_insn
*);
239 static void record_last_set_info (rtx
, const_rtx
, void *);
240 static void record_opr_changes (rtx_insn
*);
242 static void find_mem_conflicts (rtx
, const_rtx
, void *);
243 static int load_killed_in_block_p (int, rtx
, bool);
244 static void reset_opr_set_tables (void);
246 /* Hash table support. */
247 static hashval_t
hash_expr (rtx
, int *);
248 static void insert_expr_in_table (rtx
, rtx_insn
*);
249 static struct expr
*lookup_expr_in_table (rtx
);
250 static void dump_hash_table (FILE *);
252 /* Helpers for eliminate_partially_redundant_load. */
253 static bool reg_killed_on_edge (rtx
, edge
);
254 static bool reg_used_on_edge (rtx
, edge
);
256 static rtx
get_avail_load_store_reg (rtx_insn
*);
258 static bool bb_has_well_behaved_predecessors (basic_block
);
259 static struct occr
* get_bb_avail_insn (basic_block
, struct occr
*, int);
260 static void hash_scan_set (rtx_insn
*);
261 static void compute_hash_table (void);
263 /* The work horses of this pass. */
264 static void eliminate_partially_redundant_load (basic_block
,
267 static void eliminate_partially_redundant_loads (void);
270 /* Allocate memory for the CUID mapping array and register/memory
280 /* Find the largest UID and create a mapping from UIDs to CUIDs. */
281 uid_cuid
= XCNEWVEC (int, get_max_uid () + 1);
283 FOR_EACH_BB_FN (bb
, cfun
)
284 FOR_BB_INSNS (bb
, insn
)
287 uid_cuid
[INSN_UID (insn
)] = i
++;
289 uid_cuid
[INSN_UID (insn
)] = i
;
292 /* Allocate the available expressions hash table. We don't want to
293 make the hash table too small, but unnecessarily making it too large
294 also doesn't help. The i/4 is a gcse.c relic, and seems like a
295 reasonable choice. */
296 expr_table
= new hash_table
<expr_hasher
> (MAX (i
/ 4, 13));
298 /* We allocate everything on obstacks because we often can roll back
299 the whole obstack to some point. Freeing obstacks is very fast. */
300 gcc_obstack_init (&expr_obstack
);
301 gcc_obstack_init (&occr_obstack
);
302 gcc_obstack_init (&unoccr_obstack
);
303 gcc_obstack_init (&modifies_mem_obstack
);
305 /* Working array used to track the last set for each register
306 in the current block. */
307 reg_avail_info
= (int *) xmalloc (FIRST_PSEUDO_REGISTER
* sizeof (int));
309 /* Put a dummy modifies_mem object on the modifies_mem_obstack, so we
310 can roll it back in reset_opr_set_tables. */
311 modifies_mem_obstack_bottom
=
312 (struct modifies_mem
*) obstack_alloc (&modifies_mem_obstack
,
313 sizeof (struct modifies_mem
));
315 blocks_with_calls
= BITMAP_ALLOC (NULL
);
316 modify_mem_list_set
= BITMAP_ALLOC (NULL
);
318 modify_mem_list
= (vec_rtx_heap
*) xcalloc (last_basic_block_for_fn (cfun
),
319 sizeof (vec_rtx_heap
));
320 canon_modify_mem_list
321 = (vec_modify_pair_heap
*) xcalloc (last_basic_block_for_fn (cfun
),
322 sizeof (vec_modify_pair_heap
));
325 /* Free memory allocated by alloc_mem. */
335 obstack_free (&expr_obstack
, NULL
);
336 obstack_free (&occr_obstack
, NULL
);
337 obstack_free (&unoccr_obstack
, NULL
);
338 obstack_free (&modifies_mem_obstack
, NULL
);
342 EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set
, 0, i
, bi
)
344 modify_mem_list
[i
].release ();
345 canon_modify_mem_list
[i
].release ();
348 BITMAP_FREE (blocks_with_calls
);
349 BITMAP_FREE (modify_mem_list_set
);
350 free (reg_avail_info
);
351 free (modify_mem_list
);
352 free (canon_modify_mem_list
);
356 /* Insert expression X in INSN in the hash TABLE.
357 If it is already present, record it as the last occurrence in INSN's
361 insert_expr_in_table (rtx x
, rtx_insn
*insn
)
365 struct expr
*cur_expr
, **slot
;
366 struct occr
*avail_occr
, *last_occr
= NULL
;
368 hash
= hash_expr (x
, &do_not_record_p
);
370 /* Do not insert expression in the table if it contains volatile operands,
371 or if hash_expr determines the expression is something we don't want
372 to or can't handle. */
376 /* We anticipate that redundant expressions are rare, so for convenience
377 allocate a new hash table element here already and set its fields.
378 If we don't do this, we need a hack with a static struct expr. Anyway,
379 obstack_free is really fast and one more obstack_alloc doesn't hurt if
380 we're going to see more expressions later on. */
381 cur_expr
= (struct expr
*) obstack_alloc (&expr_obstack
,
382 sizeof (struct expr
));
384 cur_expr
->hash
= hash
;
385 cur_expr
->avail_occr
= NULL
;
387 slot
= expr_table
->find_slot_with_hash (cur_expr
, hash
, INSERT
);
391 /* The expression isn't found, so insert it. */
394 /* Anytime we add an entry to the table, record the index
395 of the new entry. The bitmap index starts counting
397 cur_expr
->bitmap_index
= expr_table
->elements () - 1;
401 /* The expression is already in the table, so roll back the
402 obstack and use the existing table entry. */
403 obstack_free (&expr_obstack
, cur_expr
);
407 /* Search for another occurrence in the same basic block. */
408 avail_occr
= cur_expr
->avail_occr
;
410 && BLOCK_FOR_INSN (avail_occr
->insn
) != BLOCK_FOR_INSN (insn
))
412 /* If an occurrence isn't found, save a pointer to the end of
414 last_occr
= avail_occr
;
415 avail_occr
= avail_occr
->next
;
419 /* Found another instance of the expression in the same basic block.
420 Prefer this occurrence to the currently recorded one. We want
421 the last one in the block and the block is scanned from start
423 avail_occr
->insn
= insn
;
426 /* First occurrence of this expression in this basic block. */
427 avail_occr
= (struct occr
*) obstack_alloc (&occr_obstack
,
428 sizeof (struct occr
));
430 /* First occurrence of this expression in any block? */
431 if (cur_expr
->avail_occr
== NULL
)
432 cur_expr
->avail_occr
= avail_occr
;
434 last_occr
->next
= avail_occr
;
436 avail_occr
->insn
= insn
;
437 avail_occr
->next
= NULL
;
438 avail_occr
->deleted_p
= 0;
443 /* Lookup pattern PAT in the expression hash table.
444 The result is a pointer to the table entry, or NULL if not found. */
447 lookup_expr_in_table (rtx pat
)
450 struct expr
**slot
, *tmp_expr
;
451 hashval_t hash
= hash_expr (pat
, &do_not_record_p
);
456 tmp_expr
= (struct expr
*) obstack_alloc (&expr_obstack
,
457 sizeof (struct expr
));
458 tmp_expr
->expr
= pat
;
459 tmp_expr
->hash
= hash
;
460 tmp_expr
->avail_occr
= NULL
;
462 slot
= expr_table
->find_slot_with_hash (tmp_expr
, hash
, INSERT
);
463 obstack_free (&expr_obstack
, tmp_expr
);
472 /* Dump all expressions and occurrences that are currently in the
473 expression hash table to FILE. */
475 /* This helper is called via htab_traverse. */
477 dump_expr_hash_table_entry (expr
**slot
, FILE *file
)
479 struct expr
*exprs
= *slot
;
482 fprintf (file
, "expr: ");
483 print_rtl (file
, exprs
->expr
);
484 fprintf (file
,"\nhashcode: %u\n", exprs
->hash
);
485 fprintf (file
,"list of occurrences:\n");
486 occr
= exprs
->avail_occr
;
489 rtx_insn
*insn
= occr
->insn
;
490 print_rtl_single (file
, insn
);
491 fprintf (file
, "\n");
494 fprintf (file
, "\n");
499 dump_hash_table (FILE *file
)
501 fprintf (file
, "\n\nexpression hash table\n");
502 fprintf (file
, "size %ld, %ld elements, %f collision/search ratio\n",
503 (long) expr_table
->size (),
504 (long) expr_table
->elements (),
505 expr_table
->collisions ());
506 if (expr_table
->elements () > 0)
508 fprintf (file
, "\n\ntable entries:\n");
509 expr_table
->traverse
<FILE *, dump_expr_hash_table_entry
> (file
);
511 fprintf (file
, "\n");
514 /* Return true if register X is recorded as being set by an instruction
515 whose CUID is greater than the one given. */
518 reg_changed_after_insn_p (rtx x
, int cuid
)
520 unsigned int regno
, end_regno
;
523 end_regno
= END_REGNO (x
);
525 if (reg_avail_info
[regno
] > cuid
)
527 while (++regno
< end_regno
);
531 /* Return nonzero if the operands of expression X are unchanged
532 1) from the start of INSN's basic block up to but not including INSN
533 if AFTER_INSN is false, or
534 2) from INSN to the end of INSN's basic block if AFTER_INSN is true. */
537 oprs_unchanged_p (rtx x
, rtx_insn
*insn
, bool after_insn
)
550 /* We are called after register allocation. */
551 gcc_assert (REGNO (x
) < FIRST_PSEUDO_REGISTER
);
553 return !reg_changed_after_insn_p (x
, INSN_CUID (insn
) - 1);
555 return !reg_changed_after_insn_p (x
, 0);
558 if (load_killed_in_block_p (INSN_CUID (insn
), x
, after_insn
))
561 return oprs_unchanged_p (XEXP (x
, 0), insn
, after_insn
);
587 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
591 if (! oprs_unchanged_p (XEXP (x
, i
), insn
, after_insn
))
594 else if (fmt
[i
] == 'E')
595 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
596 if (! oprs_unchanged_p (XVECEXP (x
, i
, j
), insn
, after_insn
))
604 /* Used for communication between find_mem_conflicts and
605 load_killed_in_block_p. Nonzero if find_mem_conflicts finds a
606 conflict between two memory references.
607 This is a bit of a hack to work around the limitations of note_stores. */
608 static int mems_conflict_p
;
610 /* DEST is the output of an instruction. If it is a memory reference, and
611 possibly conflicts with the load found in DATA, then set mems_conflict_p
612 to a nonzero value. */
615 find_mem_conflicts (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
,
618 rtx mem_op
= (rtx
) data
;
620 while (GET_CODE (dest
) == SUBREG
621 || GET_CODE (dest
) == ZERO_EXTRACT
622 || GET_CODE (dest
) == STRICT_LOW_PART
)
623 dest
= XEXP (dest
, 0);
625 /* If DEST is not a MEM, then it will not conflict with the load. Note
626 that function calls are assumed to clobber memory, but are handled
631 if (true_dependence (dest
, GET_MODE (dest
), mem_op
))
636 /* Return nonzero if the expression in X (a memory reference) is killed
637 in the current basic block before (if AFTER_INSN is false) or after
638 (if AFTER_INSN is true) the insn with the CUID in UID_LIMIT.
640 This function assumes that the modifies_mem table is flushed when
641 the hash table construction or redundancy elimination phases start
642 processing a new basic block. */
645 load_killed_in_block_p (int uid_limit
, rtx x
, bool after_insn
)
647 struct modifies_mem
*list_entry
= modifies_mem_list
;
651 rtx_insn
*setter
= list_entry
->insn
;
653 /* Ignore entries in the list that do not apply. */
655 && INSN_CUID (setter
) < uid_limit
)
657 && INSN_CUID (setter
) > uid_limit
))
659 list_entry
= list_entry
->next
;
663 /* If SETTER is a call everything is clobbered. Note that calls
664 to pure functions are never put on the list, so we need not
669 /* SETTER must be an insn of some kind that sets memory. Call
670 note_stores to examine each hunk of memory that is modified.
671 It will set mems_conflict_p to nonzero if there may be a
672 conflict between X and SETTER. */
674 note_stores (PATTERN (setter
), find_mem_conflicts
, x
);
678 list_entry
= list_entry
->next
;
684 /* Record register first/last/block set information for REGNO in INSN. */
687 record_last_reg_set_info (rtx_insn
*insn
, rtx reg
)
689 unsigned int regno
, end_regno
;
692 end_regno
= END_REGNO (reg
);
694 reg_avail_info
[regno
] = INSN_CUID (insn
);
695 while (++regno
< end_regno
);
699 record_last_reg_set_info_regno (rtx_insn
*insn
, int regno
)
701 reg_avail_info
[regno
] = INSN_CUID (insn
);
705 /* Record memory modification information for INSN. We do not actually care
706 about the memory location(s) that are set, or even how they are set (consider
707 a CALL_INSN). We merely need to record which insns modify memory. */
710 record_last_mem_set_info (rtx_insn
*insn
)
712 struct modifies_mem
*list_entry
;
714 list_entry
= (struct modifies_mem
*) obstack_alloc (&modifies_mem_obstack
,
715 sizeof (struct modifies_mem
));
716 list_entry
->insn
= insn
;
717 list_entry
->next
= modifies_mem_list
;
718 modifies_mem_list
= list_entry
;
720 record_last_mem_set_info_common (insn
, modify_mem_list
,
721 canon_modify_mem_list
,
726 /* Called from compute_hash_table via note_stores to handle one
727 SET or CLOBBER in an insn. DATA is really the instruction in which
728 the SET is taking place. */
731 record_last_set_info (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
, void *data
)
733 rtx_insn
*last_set_insn
= (rtx_insn
*) data
;
735 if (GET_CODE (dest
) == SUBREG
)
736 dest
= SUBREG_REG (dest
);
739 record_last_reg_set_info (last_set_insn
, dest
);
740 else if (MEM_P (dest
))
742 /* Ignore pushes, they don't clobber memory. They may still
743 clobber the stack pointer though. Some targets do argument
744 pushes without adding REG_INC notes. See e.g. PR25196,
745 where a pushsi2 on i386 doesn't have REG_INC notes. Note
746 such changes here too. */
747 if (! push_operand (dest
, GET_MODE (dest
)))
748 record_last_mem_set_info (last_set_insn
);
750 record_last_reg_set_info_regno (last_set_insn
, STACK_POINTER_REGNUM
);
755 /* Reset tables used to keep track of what's still available since the
756 start of the block. */
759 reset_opr_set_tables (void)
761 memset (reg_avail_info
, 0, FIRST_PSEUDO_REGISTER
* sizeof (int));
762 obstack_free (&modifies_mem_obstack
, modifies_mem_obstack_bottom
);
763 modifies_mem_list
= NULL
;
767 /* Record things set by INSN.
768 This data is used by oprs_unchanged_p. */
771 record_opr_changes (rtx_insn
*insn
)
775 /* Find all stores and record them. */
776 note_stores (PATTERN (insn
), record_last_set_info
, insn
);
778 /* Also record autoincremented REGs for this insn as changed. */
779 for (note
= REG_NOTES (insn
); note
; note
= XEXP (note
, 1))
780 if (REG_NOTE_KIND (note
) == REG_INC
)
781 record_last_reg_set_info (insn
, XEXP (note
, 0));
783 /* Finally, if this is a call, record all call clobbers. */
788 hard_reg_set_iterator hrsi
;
789 EXECUTE_IF_SET_IN_HARD_REG_SET (regs_invalidated_by_call
, 0, regno
, hrsi
)
790 record_last_reg_set_info_regno (insn
, regno
);
792 for (link
= CALL_INSN_FUNCTION_USAGE (insn
); link
; link
= XEXP (link
, 1))
793 if (GET_CODE (XEXP (link
, 0)) == CLOBBER
)
795 x
= XEXP (XEXP (link
, 0), 0);
798 gcc_assert (HARD_REGISTER_P (x
));
799 record_last_reg_set_info (insn
, x
);
803 if (! RTL_CONST_OR_PURE_CALL_P (insn
))
804 record_last_mem_set_info (insn
);
809 /* Scan the pattern of INSN and add an entry to the hash TABLE.
810 After reload we are interested in loads/stores only. */
813 hash_scan_set (rtx_insn
*insn
)
815 rtx pat
= PATTERN (insn
);
816 rtx src
= SET_SRC (pat
);
817 rtx dest
= SET_DEST (pat
);
819 /* We are only interested in loads and stores. */
820 if (! MEM_P (src
) && ! MEM_P (dest
))
823 /* Don't mess with jumps and nops. */
824 if (JUMP_P (insn
) || set_noop_p (pat
))
829 if (/* Don't CSE something if we can't do a reg/reg copy. */
830 can_copy_p (GET_MODE (dest
))
831 /* Is SET_SRC something we want to gcse? */
832 && general_operand (src
, GET_MODE (src
))
834 /* Never consider insns touching the register stack. It may
835 create situations that reg-stack cannot handle (e.g. a stack
836 register live across an abnormal edge). */
837 && (REGNO (dest
) < FIRST_STACK_REG
|| REGNO (dest
) > LAST_STACK_REG
)
839 /* An expression is not available if its operands are
840 subsequently modified, including this insn. */
841 && oprs_unchanged_p (src
, insn
, true))
843 insert_expr_in_table (src
, insn
);
846 else if (REG_P (src
))
848 /* Only record sets of pseudo-regs in the hash table. */
849 if (/* Don't CSE something if we can't do a reg/reg copy. */
850 can_copy_p (GET_MODE (src
))
851 /* Is SET_DEST something we want to gcse? */
852 && general_operand (dest
, GET_MODE (dest
))
854 /* As above for STACK_REGS. */
855 && (REGNO (src
) < FIRST_STACK_REG
|| REGNO (src
) > LAST_STACK_REG
)
857 && ! (flag_float_store
&& FLOAT_MODE_P (GET_MODE (dest
)))
858 /* Check if the memory expression is killed after insn. */
859 && ! load_killed_in_block_p (INSN_CUID (insn
) + 1, dest
, true)
860 && oprs_unchanged_p (XEXP (dest
, 0), insn
, true))
862 insert_expr_in_table (dest
, insn
);
868 /* Create hash table of memory expressions available at end of basic
869 blocks. Basically you should think of this hash table as the
870 representation of AVAIL_OUT. This is the set of expressions that
871 is generated in a basic block and not killed before the end of the
872 same basic block. Notice that this is really a local computation. */
875 compute_hash_table (void)
879 FOR_EACH_BB_FN (bb
, cfun
)
883 /* First pass over the instructions records information used to
884 determine when registers and memory are last set.
885 Since we compute a "local" AVAIL_OUT, reset the tables that
886 help us keep track of what has been modified since the start
888 reset_opr_set_tables ();
889 FOR_BB_INSNS (bb
, insn
)
892 record_opr_changes (insn
);
895 /* The next pass actually builds the hash table. */
896 FOR_BB_INSNS (bb
, insn
)
897 if (INSN_P (insn
) && GET_CODE (PATTERN (insn
)) == SET
)
898 hash_scan_set (insn
);
903 /* Check if register REG is killed in any insn waiting to be inserted on
904 edge E. This function is required to check that our data flow analysis
905 is still valid prior to commit_edge_insertions. */
908 reg_killed_on_edge (rtx reg
, edge e
)
912 for (insn
= e
->insns
.r
; insn
; insn
= NEXT_INSN (insn
))
913 if (INSN_P (insn
) && reg_set_p (reg
, insn
))
919 /* Similar to above - check if register REG is used in any insn waiting
920 to be inserted on edge E.
921 Assumes no such insn can be a CALL_INSN; if so call reg_used_between_p
922 with PREV(insn),NEXT(insn) instead of calling reg_overlap_mentioned_p. */
925 reg_used_on_edge (rtx reg
, edge e
)
929 for (insn
= e
->insns
.r
; insn
; insn
= NEXT_INSN (insn
))
930 if (INSN_P (insn
) && reg_overlap_mentioned_p (reg
, PATTERN (insn
)))
936 /* Return the loaded/stored register of a load/store instruction. */
939 get_avail_load_store_reg (rtx_insn
*insn
)
941 if (REG_P (SET_DEST (PATTERN (insn
))))
943 return SET_DEST (PATTERN (insn
));
947 gcc_assert (REG_P (SET_SRC (PATTERN (insn
))));
948 return SET_SRC (PATTERN (insn
));
952 /* Return nonzero if the predecessors of BB are "well behaved". */
955 bb_has_well_behaved_predecessors (basic_block bb
)
960 if (EDGE_COUNT (bb
->preds
) == 0)
963 FOR_EACH_EDGE (pred
, ei
, bb
->preds
)
965 /* commit_one_edge_insertion refuses to insert on abnormal edges even if
966 the source has only one successor so EDGE_CRITICAL_P is too weak. */
967 if ((pred
->flags
& EDGE_ABNORMAL
) && !single_pred_p (pred
->dest
))
970 if ((pred
->flags
& EDGE_ABNORMAL_CALL
) && cfun
->has_nonlocal_label
)
973 if (tablejump_p (BB_END (pred
->src
), NULL
, NULL
))
980 /* Search for the occurrences of expression in BB. */
983 get_bb_avail_insn (basic_block bb
, struct occr
*orig_occr
, int bitmap_index
)
985 struct occr
*occr
= orig_occr
;
987 for (; occr
!= NULL
; occr
= occr
->next
)
988 if (BLOCK_FOR_INSN (occr
->insn
) == bb
)
991 /* If we could not find an occurrence in BB, see if BB
992 has a single predecessor with an occurrence that is
993 transparent through BB. */
994 if (single_pred_p (bb
)
995 && bitmap_bit_p (transp
[bb
->index
], bitmap_index
)
996 && (occr
= get_bb_avail_insn (single_pred (bb
), orig_occr
, bitmap_index
)))
998 rtx avail_reg
= get_avail_load_store_reg (occr
->insn
);
999 if (!reg_set_between_p (avail_reg
,
1000 PREV_INSN (BB_HEAD (bb
)),
1001 NEXT_INSN (BB_END (bb
)))
1002 && !reg_killed_on_edge (avail_reg
, single_pred_edge (bb
)))
1010 /* This helper is called via htab_traverse. */
1012 compute_expr_transp (expr
**slot
, FILE *dump_file ATTRIBUTE_UNUSED
)
1014 struct expr
*expr
= *slot
;
1016 compute_transp (expr
->expr
, expr
->bitmap_index
, transp
,
1017 blocks_with_calls
, modify_mem_list_set
,
1018 canon_modify_mem_list
);
1022 /* This handles the case where several stores feed a partially redundant
1023 load. It checks if the redundancy elimination is possible and if it's
1026 Redundancy elimination is possible if,
1027 1) None of the operands of an insn have been modified since the start
1028 of the current basic block.
1029 2) In any predecessor of the current basic block, the same expression
1032 See the function body for the heuristics that determine if eliminating
1033 a redundancy is also worth doing, assuming it is possible. */
1036 eliminate_partially_redundant_load (basic_block bb
, rtx_insn
*insn
,
1040 rtx_insn
*avail_insn
= NULL
;
1043 struct occr
*a_occr
;
1044 struct unoccr
*occr
, *avail_occrs
= NULL
;
1045 struct unoccr
*unoccr
, *unavail_occrs
= NULL
, *rollback_unoccr
= NULL
;
1047 gcov_type ok_count
= 0; /* Redundant load execution count. */
1048 gcov_type critical_count
= 0; /* Execution count of critical edges. */
1050 bool critical_edge_split
= false;
1052 /* The execution count of the loads to be added to make the
1053 load fully redundant. */
1054 gcov_type not_ok_count
= 0;
1055 basic_block pred_bb
;
1057 pat
= PATTERN (insn
);
1058 dest
= SET_DEST (pat
);
1060 /* Check that the loaded register is not used, set, or killed from the
1061 beginning of the block. */
1062 if (reg_changed_after_insn_p (dest
, 0)
1063 || reg_used_between_p (dest
, PREV_INSN (BB_HEAD (bb
)), insn
))
1066 /* Check potential for replacing load with copy for predecessors. */
1067 FOR_EACH_EDGE (pred
, ei
, bb
->preds
)
1069 rtx_insn
*next_pred_bb_end
;
1072 avail_reg
= NULL_RTX
;
1073 pred_bb
= pred
->src
;
1074 for (a_occr
= get_bb_avail_insn (pred_bb
,
1076 expr
->bitmap_index
);
1078 a_occr
= get_bb_avail_insn (pred_bb
,
1080 expr
->bitmap_index
))
1082 /* Check if the loaded register is not used. */
1083 avail_insn
= a_occr
->insn
;
1084 avail_reg
= get_avail_load_store_reg (avail_insn
);
1085 gcc_assert (avail_reg
);
1087 /* Make sure we can generate a move from register avail_reg to
1089 rtx_insn
*move
= gen_move_insn (copy_rtx (dest
),
1090 copy_rtx (avail_reg
));
1091 extract_insn (move
);
1092 if (! constrain_operands (1, get_preferred_alternatives (insn
,
1094 || reg_killed_on_edge (avail_reg
, pred
)
1095 || reg_used_on_edge (dest
, pred
))
1100 next_pred_bb_end
= NEXT_INSN (BB_END (BLOCK_FOR_INSN (avail_insn
)));
1101 if (!reg_set_between_p (avail_reg
, avail_insn
, next_pred_bb_end
))
1102 /* AVAIL_INSN remains non-null. */
1108 if (EDGE_CRITICAL_P (pred
))
1109 critical_count
+= pred
->count
;
1111 if (avail_insn
!= NULL_RTX
)
1114 ok_count
+= pred
->count
;
1115 if (! set_noop_p (PATTERN (gen_move_insn (copy_rtx (dest
),
1116 copy_rtx (avail_reg
)))))
1118 /* Check if there is going to be a split. */
1119 if (EDGE_CRITICAL_P (pred
))
1120 critical_edge_split
= true;
1122 else /* Its a dead move no need to generate. */
1124 occr
= (struct unoccr
*) obstack_alloc (&unoccr_obstack
,
1125 sizeof (struct unoccr
));
1126 occr
->insn
= avail_insn
;
1128 occr
->next
= avail_occrs
;
1130 if (! rollback_unoccr
)
1131 rollback_unoccr
= occr
;
1135 /* Adding a load on a critical edge will cause a split. */
1136 if (EDGE_CRITICAL_P (pred
))
1137 critical_edge_split
= true;
1138 not_ok_count
+= pred
->count
;
1139 unoccr
= (struct unoccr
*) obstack_alloc (&unoccr_obstack
,
1140 sizeof (struct unoccr
));
1141 unoccr
->insn
= NULL
;
1142 unoccr
->pred
= pred
;
1143 unoccr
->next
= unavail_occrs
;
1144 unavail_occrs
= unoccr
;
1145 if (! rollback_unoccr
)
1146 rollback_unoccr
= unoccr
;
1150 if (/* No load can be replaced by copy. */
1152 /* Prevent exploding the code. */
1153 || (optimize_bb_for_size_p (bb
) && npred_ok
> 1)
1154 /* If we don't have profile information we cannot tell if splitting
1155 a critical edge is profitable or not so don't do it. */
1156 || ((! profile_info
|| ! flag_branch_probabilities
1157 || targetm
.cannot_modify_jumps_p ())
1158 && critical_edge_split
))
1161 /* Check if it's worth applying the partial redundancy elimination. */
1162 if (ok_count
< GCSE_AFTER_RELOAD_PARTIAL_FRACTION
* not_ok_count
)
1164 if (ok_count
< GCSE_AFTER_RELOAD_CRITICAL_FRACTION
* critical_count
)
1167 /* Generate moves to the loaded register from where
1168 the memory is available. */
1169 for (occr
= avail_occrs
; occr
; occr
= occr
->next
)
1171 avail_insn
= occr
->insn
;
1173 /* Set avail_reg to be the register having the value of the
1175 avail_reg
= get_avail_load_store_reg (avail_insn
);
1176 gcc_assert (avail_reg
);
1178 insert_insn_on_edge (gen_move_insn (copy_rtx (dest
),
1179 copy_rtx (avail_reg
)),
1181 stats
.moves_inserted
++;
1185 "generating move from %d to %d on edge from %d to %d\n",
1192 /* Regenerate loads where the memory is unavailable. */
1193 for (unoccr
= unavail_occrs
; unoccr
; unoccr
= unoccr
->next
)
1195 pred
= unoccr
->pred
;
1196 insert_insn_on_edge (copy_insn (PATTERN (insn
)), pred
);
1197 stats
.copies_inserted
++;
1202 "generating on edge from %d to %d a copy of load: ",
1205 print_rtl (dump_file
, PATTERN (insn
));
1206 fprintf (dump_file
, "\n");
1210 /* Delete the insn if it is not available in this block and mark it
1211 for deletion if it is available. If insn is available it may help
1212 discover additional redundancies, so mark it for later deletion. */
1213 for (a_occr
= get_bb_avail_insn (bb
, expr
->avail_occr
, expr
->bitmap_index
);
1214 a_occr
&& (a_occr
->insn
!= insn
);
1215 a_occr
= get_bb_avail_insn (bb
, a_occr
->next
, expr
->bitmap_index
))
1220 stats
.insns_deleted
++;
1224 fprintf (dump_file
, "deleting insn:\n");
1225 print_rtl_single (dump_file
, insn
);
1226 fprintf (dump_file
, "\n");
1231 a_occr
->deleted_p
= 1;
1234 if (rollback_unoccr
)
1235 obstack_free (&unoccr_obstack
, rollback_unoccr
);
1238 /* Performing the redundancy elimination as described before. */
1241 eliminate_partially_redundant_loads (void)
1246 /* Note we start at block 1. */
1248 if (ENTRY_BLOCK_PTR_FOR_FN (cfun
)->next_bb
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
1252 ENTRY_BLOCK_PTR_FOR_FN (cfun
)->next_bb
->next_bb
,
1253 EXIT_BLOCK_PTR_FOR_FN (cfun
),
1256 /* Don't try anything on basic blocks with strange predecessors. */
1257 if (! bb_has_well_behaved_predecessors (bb
))
1260 /* Do not try anything on cold basic blocks. */
1261 if (optimize_bb_for_size_p (bb
))
1264 /* Reset the table of things changed since the start of the current
1266 reset_opr_set_tables ();
1268 /* Look at all insns in the current basic block and see if there are
1269 any loads in it that we can record. */
1270 FOR_BB_INSNS (bb
, insn
)
1272 /* Is it a load - of the form (set (reg) (mem))? */
1273 if (NONJUMP_INSN_P (insn
)
1274 && GET_CODE (PATTERN (insn
)) == SET
1275 && REG_P (SET_DEST (PATTERN (insn
)))
1276 && MEM_P (SET_SRC (PATTERN (insn
))))
1278 rtx pat
= PATTERN (insn
);
1279 rtx src
= SET_SRC (pat
);
1282 if (!MEM_VOLATILE_P (src
)
1283 && GET_MODE (src
) != BLKmode
1284 && general_operand (src
, GET_MODE (src
))
1285 /* Are the operands unchanged since the start of the
1287 && oprs_unchanged_p (src
, insn
, false)
1288 && !(cfun
->can_throw_non_call_exceptions
&& may_trap_p (src
))
1289 && !side_effects_p (src
)
1290 /* Is the expression recorded? */
1291 && (expr
= lookup_expr_in_table (src
)) != NULL
)
1293 /* We now have a load (insn) and an available memory at
1294 its BB start (expr). Try to remove the loads if it is
1296 eliminate_partially_redundant_load (bb
, insn
, expr
);
1300 /* Keep track of everything modified by this insn, so that we
1301 know what has been modified since the start of the current
1304 record_opr_changes (insn
);
1308 commit_edge_insertions ();
1311 /* Go over the expression hash table and delete insns that were
1312 marked for later deletion. */
1314 /* This helper is called via htab_traverse. */
1316 delete_redundant_insns_1 (expr
**slot
, void *data ATTRIBUTE_UNUSED
)
1318 struct expr
*exprs
= *slot
;
1321 for (occr
= exprs
->avail_occr
; occr
!= NULL
; occr
= occr
->next
)
1323 if (occr
->deleted_p
&& dbg_cnt (gcse2_delete
))
1325 delete_insn (occr
->insn
);
1326 stats
.insns_deleted
++;
1330 fprintf (dump_file
, "deleting insn:\n");
1331 print_rtl_single (dump_file
, occr
->insn
);
1332 fprintf (dump_file
, "\n");
1341 delete_redundant_insns (void)
1343 expr_table
->traverse
<void *, delete_redundant_insns_1
> (NULL
);
1345 fprintf (dump_file
, "\n");
1348 /* Main entry point of the GCSE after reload - clean some redundant loads
1352 gcse_after_reload_main (rtx f ATTRIBUTE_UNUSED
)
1355 memset (&stats
, 0, sizeof (stats
));
1357 /* Allocate memory for this pass.
1358 Also computes and initializes the insns' CUIDs. */
1361 /* We need alias analysis. */
1362 init_alias_analysis ();
1364 compute_hash_table ();
1367 dump_hash_table (dump_file
);
1369 if (expr_table
->elements () > 0)
1371 /* Knowing which MEMs are transparent through a block can signifiantly
1372 increase the number of redundant loads found. So compute transparency
1373 information for each memory expression in the hash table. */
1375 /* This can not be part of the normal allocation routine because
1376 we have to know the number of elements in the hash table. */
1377 transp
= sbitmap_vector_alloc (last_basic_block_for_fn (cfun
),
1378 expr_table
->elements ());
1379 bitmap_vector_ones (transp
, last_basic_block_for_fn (cfun
));
1380 expr_table
->traverse
<FILE *, compute_expr_transp
> (dump_file
);
1381 eliminate_partially_redundant_loads ();
1382 delete_redundant_insns ();
1383 sbitmap_vector_free (transp
);
1387 fprintf (dump_file
, "GCSE AFTER RELOAD stats:\n");
1388 fprintf (dump_file
, "copies inserted: %d\n", stats
.copies_inserted
);
1389 fprintf (dump_file
, "moves inserted: %d\n", stats
.moves_inserted
);
1390 fprintf (dump_file
, "insns deleted: %d\n", stats
.insns_deleted
);
1391 fprintf (dump_file
, "\n\n");
1394 statistics_counter_event (cfun
, "copies inserted",
1395 stats
.copies_inserted
);
1396 statistics_counter_event (cfun
, "moves inserted",
1397 stats
.moves_inserted
);
1398 statistics_counter_event (cfun
, "insns deleted",
1399 stats
.insns_deleted
);
1402 /* We are finished with alias. */
1403 end_alias_analysis ();
1411 rest_of_handle_gcse2 (void)
1413 gcse_after_reload_main (get_insns ());
1414 rebuild_jump_labels (get_insns ());
1420 const pass_data pass_data_gcse2
=
1422 RTL_PASS
, /* type */
1424 OPTGROUP_NONE
, /* optinfo_flags */
1425 TV_GCSE_AFTER_RELOAD
, /* tv_id */
1426 0, /* properties_required */
1427 0, /* properties_provided */
1428 0, /* properties_destroyed */
1429 0, /* todo_flags_start */
1430 0, /* todo_flags_finish */
1433 class pass_gcse2
: public rtl_opt_pass
1436 pass_gcse2 (gcc::context
*ctxt
)
1437 : rtl_opt_pass (pass_data_gcse2
, ctxt
)
1440 /* opt_pass methods: */
1441 virtual bool gate (function
*fun
)
1443 return (optimize
> 0 && flag_gcse_after_reload
1444 && optimize_function_for_speed_p (fun
));
1447 virtual unsigned int execute (function
*) { return rest_of_handle_gcse2 (); }
1449 }; // class pass_gcse2
1454 make_pass_gcse2 (gcc::context
*ctxt
)
1456 return new pass_gcse2 (ctxt
);