1 /* Global common subexpression elimination/Partial redundancy elimination
2 and global constant/copy propagation for GNU compiler.
3 Copyright (C) 1997, 1998, 1999, 2000, 2001 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 2, 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 COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
23 - reordering of memory allocation and freeing to be more space efficient
24 - do rough calc of how many regs are needed in each block, and a rough
25 calc of how many regs are available in each class and use that to
26 throttle back the code in cases where RTX_COST is minimal.
27 - a store to the same address as a load does not kill the load if the
28 source of the store is also the destination of the load. Handling this
29 allows more load motion, particularly out of loops.
30 - ability to realloc sbitmap vectors would allow one initial computation
31 of reg_set_in_block with only subsequent additions, rather than
32 recomputing it for each pass
36 /* References searched while implementing this.
38 Compilers Principles, Techniques and Tools
42 Global Optimization by Suppression of Partial Redundancies
44 communications of the acm, Vol. 22, Num. 2, Feb. 1979
46 A Portable Machine-Independent Global Optimizer - Design and Measurements
48 Stanford Ph.D. thesis, Dec. 1983
50 A Fast Algorithm for Code Movement Optimization
52 SIGPLAN Notices, Vol. 23, Num. 10, Oct. 1988
54 A Solution to a Problem with Morel and Renvoise's
55 Global Optimization by Suppression of Partial Redundancies
56 K-H Drechsler, M.P. Stadel
57 ACM TOPLAS, Vol. 10, Num. 4, Oct. 1988
59 Practical Adaptation of the Global Optimization
60 Algorithm of Morel and Renvoise
62 ACM TOPLAS, Vol. 13, Num. 2. Apr. 1991
64 Efficiently Computing Static Single Assignment Form and the Control
66 R. Cytron, J. Ferrante, B.K. Rosen, M.N. Wegman, and F.K. Zadeck
67 ACM TOPLAS, Vol. 13, Num. 4, Oct. 1991
70 J. Knoop, O. Ruthing, B. Steffen
71 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
73 What's In a Region? Or Computing Control Dependence Regions in Near-Linear
74 Time for Reducible Flow Control
76 ACM Letters on Programming Languages and Systems,
77 Vol. 2, Num. 1-4, Mar-Dec 1993
79 An Efficient Representation for Sparse Sets
80 Preston Briggs, Linda Torczon
81 ACM Letters on Programming Languages and Systems,
82 Vol. 2, Num. 1-4, Mar-Dec 1993
84 A Variation of Knoop, Ruthing, and Steffen's Lazy Code Motion
85 K-H Drechsler, M.P. Stadel
86 ACM SIGPLAN Notices, Vol. 28, Num. 5, May 1993
88 Partial Dead Code Elimination
89 J. Knoop, O. Ruthing, B. Steffen
90 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
92 Effective Partial Redundancy Elimination
93 P. Briggs, K.D. Cooper
94 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
96 The Program Structure Tree: Computing Control Regions in Linear Time
97 R. Johnson, D. Pearson, K. Pingali
98 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
100 Optimal Code Motion: Theory and Practice
101 J. Knoop, O. Ruthing, B. Steffen
102 ACM TOPLAS, Vol. 16, Num. 4, Jul. 1994
104 The power of assignment motion
105 J. Knoop, O. Ruthing, B. Steffen
106 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
108 Global code motion / global value numbering
110 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
112 Value Driven Redundancy Elimination
114 Rice University Ph.D. thesis, Apr. 1996
118 Massively Scalar Compiler Project, Rice University, Sep. 1996
120 High Performance Compilers for Parallel Computing
124 Advanced Compiler Design and Implementation
126 Morgan Kaufmann, 1997
128 Building an Optimizing Compiler
132 People wishing to speed up the code here should read:
133 Elimination Algorithms for Data Flow Analysis
134 B.G. Ryder, M.C. Paull
135 ACM Computing Surveys, Vol. 18, Num. 3, Sep. 1986
137 How to Analyze Large Programs Efficiently and Informatively
138 D.M. Dhamdhere, B.K. Rosen, F.K. Zadeck
139 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
141 People wishing to do something different can find various possibilities
142 in the above papers and elsewhere.
152 #include "hard-reg-set.h"
155 #include "insn-config.h"
157 #include "basic-block.h"
159 #include "function.h"
165 #define obstack_chunk_alloc gmalloc
166 #define obstack_chunk_free free
168 /* Propagate flow information through back edges and thus enable PRE's
169 moving loop invariant calculations out of loops.
171 Originally this tended to create worse overall code, but several
172 improvements during the development of PRE seem to have made following
173 back edges generally a win.
175 Note much of the loop invariant code motion done here would normally
176 be done by loop.c, which has more heuristics for when to move invariants
177 out of loops. At some point we might need to move some of those
178 heuristics into gcse.c. */
179 #define FOLLOW_BACK_EDGES 1
181 /* We support GCSE via Partial Redundancy Elimination. PRE optimizations
182 are a superset of those done by GCSE.
184 We perform the following steps:
186 1) Compute basic block information.
188 2) Compute table of places where registers are set.
190 3) Perform copy/constant propagation.
192 4) Perform global cse.
194 5) Perform another pass of copy/constant propagation.
196 Two passes of copy/constant propagation are done because the first one
197 enables more GCSE and the second one helps to clean up the copies that
198 GCSE creates. This is needed more for PRE than for Classic because Classic
199 GCSE will try to use an existing register containing the common
200 subexpression rather than create a new one. This is harder to do for PRE
201 because of the code motion (which Classic GCSE doesn't do).
203 Expressions we are interested in GCSE-ing are of the form
204 (set (pseudo-reg) (expression)).
205 Function want_to_gcse_p says what these are.
207 PRE handles moving invariant expressions out of loops (by treating them as
208 partially redundant).
210 Eventually it would be nice to replace cse.c/gcse.c with SSA (static single
211 assignment) based GVN (global value numbering). L. T. Simpson's paper
212 (Rice University) on value numbering is a useful reference for this.
214 **********************
216 We used to support multiple passes but there are diminishing returns in
217 doing so. The first pass usually makes 90% of the changes that are doable.
218 A second pass can make a few more changes made possible by the first pass.
219 Experiments show any further passes don't make enough changes to justify
222 A study of spec92 using an unlimited number of passes:
223 [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
224 [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
225 [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
227 It was found doing copy propagation between each pass enables further
230 PRE is quite expensive in complicated functions because the DFA can take
231 awhile to converge. Hence we only perform one pass. The parameter max-gcse-passes can
232 be modified if one wants to experiment.
234 **********************
236 The steps for PRE are:
238 1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
240 2) Perform the data flow analysis for PRE.
242 3) Delete the redundant instructions
244 4) Insert the required copies [if any] that make the partially
245 redundant instructions fully redundant.
247 5) For other reaching expressions, insert an instruction to copy the value
248 to a newly created pseudo that will reach the redundant instruction.
250 The deletion is done first so that when we do insertions we
251 know which pseudo reg to use.
253 Various papers have argued that PRE DFA is expensive (O(n^2)) and others
254 argue it is not. The number of iterations for the algorithm to converge
255 is typically 2-4 so I don't view it as that expensive (relatively speaking).
257 PRE GCSE depends heavily on the second CSE pass to clean up the copies
258 we create. To make an expression reach the place where it's redundant,
259 the result of the expression is copied to a new register, and the redundant
260 expression is deleted by replacing it with this new register. Classic GCSE
261 doesn't have this problem as much as it computes the reaching defs of
262 each register in each block and thus can try to use an existing register.
264 **********************
266 A fair bit of simplicity is created by creating small functions for simple
267 tasks, even when the function is only called in one place. This may
268 measurably slow things down [or may not] by creating more function call
269 overhead than is necessary. The source is laid out so that it's trivial
270 to make the affected functions inline so that one can measure what speed
271 up, if any, can be achieved, and maybe later when things settle things can
274 Help stamp out big monolithic functions! */
276 /* GCSE global vars. */
279 static FILE *gcse_file;
281 /* Note whether or not we should run jump optimization after gcse. We
282 want to do this for two cases.
284 * If we changed any jumps via cprop.
286 * If we added any labels via edge splitting. */
288 static int run_jump_opt_after_gcse;
290 /* Bitmaps are normally not included in debugging dumps.
291 However it's useful to be able to print them from GDB.
292 We could create special functions for this, but it's simpler to
293 just allow passing stderr to the dump_foo fns. Since stderr can
294 be a macro, we store a copy here. */
295 static FILE *debug_stderr;
297 /* An obstack for our working variables. */
298 static struct obstack gcse_obstack;
300 /* Non-zero for each mode that supports (set (reg) (reg)).
301 This is trivially true for integer and floating point values.
302 It may or may not be true for condition codes. */
303 static char can_copy_p[(int) NUM_MACHINE_MODES];
305 /* Non-zero if can_copy_p has been initialized. */
306 static int can_copy_init_p;
308 struct reg_use {rtx reg_rtx; };
310 /* Hash table of expressions. */
314 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
316 /* Index in the available expression bitmaps. */
318 /* Next entry with the same hash. */
319 struct expr *next_same_hash;
320 /* List of anticipatable occurrences in basic blocks in the function.
321 An "anticipatable occurrence" is one that is the first occurrence in the
322 basic block, the operands are not modified in the basic block prior
323 to the occurrence and the output is not used between the start of
324 the block and the occurrence. */
325 struct occr *antic_occr;
326 /* List of available occurrence in basic blocks in the function.
327 An "available occurrence" is one that is the last occurrence in the
328 basic block and the operands are not modified by following statements in
329 the basic block [including this insn]. */
330 struct occr *avail_occr;
331 /* Non-null if the computation is PRE redundant.
332 The value is the newly created pseudo-reg to record a copy of the
333 expression in all the places that reach the redundant copy. */
337 /* Occurrence of an expression.
338 There is one per basic block. If a pattern appears more than once the
339 last appearance is used [or first for anticipatable expressions]. */
343 /* Next occurrence of this expression. */
345 /* The insn that computes the expression. */
347 /* Non-zero if this [anticipatable] occurrence has been deleted. */
349 /* Non-zero if this [available] occurrence has been copied to
351 /* ??? This is mutually exclusive with deleted_p, so they could share
356 /* Expression and copy propagation hash tables.
357 Each hash table is an array of buckets.
358 ??? It is known that if it were an array of entries, structure elements
359 `next_same_hash' and `bitmap_index' wouldn't be necessary. However, it is
360 not clear whether in the final analysis a sufficient amount of memory would
361 be saved as the size of the available expression bitmaps would be larger
362 [one could build a mapping table without holes afterwards though].
363 Someday I'll perform the computation and figure it out. */
365 /* Total size of the expression hash table, in elements. */
366 static unsigned int expr_hash_table_size;
369 This is an array of `expr_hash_table_size' elements. */
370 static struct expr **expr_hash_table;
372 /* Total size of the copy propagation hash table, in elements. */
373 static unsigned int set_hash_table_size;
376 This is an array of `set_hash_table_size' elements. */
377 static struct expr **set_hash_table;
379 /* Mapping of uids to cuids.
380 Only real insns get cuids. */
381 static int *uid_cuid;
383 /* Highest UID in UID_CUID. */
386 /* Get the cuid of an insn. */
387 #ifdef ENABLE_CHECKING
388 #define INSN_CUID(INSN) (INSN_UID (INSN) > max_uid ? (abort (), 0) : uid_cuid[INSN_UID (INSN)])
390 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
393 /* Number of cuids. */
396 /* Mapping of cuids to insns. */
397 static rtx *cuid_insn;
399 /* Get insn from cuid. */
400 #define CUID_INSN(CUID) (cuid_insn[CUID])
402 /* Maximum register number in function prior to doing gcse + 1.
403 Registers created during this pass have regno >= max_gcse_regno.
404 This is named with "gcse" to not collide with global of same name. */
405 static unsigned int max_gcse_regno;
407 /* Maximum number of cse-able expressions found. */
410 /* Maximum number of assignments for copy propagation found. */
413 /* Table of registers that are modified.
415 For each register, each element is a list of places where the pseudo-reg
418 For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
419 requires knowledge of which blocks kill which regs [and thus could use
420 a bitmap instead of the lists `reg_set_table' uses].
422 `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
423 num-regs) [however perhaps it may be useful to keep the data as is]. One
424 advantage of recording things this way is that `reg_set_table' is fairly
425 sparse with respect to pseudo regs but for hard regs could be fairly dense
426 [relatively speaking]. And recording sets of pseudo-regs in lists speeds
427 up functions like compute_transp since in the case of pseudo-regs we only
428 need to iterate over the number of times a pseudo-reg is set, not over the
429 number of basic blocks [clearly there is a bit of a slow down in the cases
430 where a pseudo is set more than once in a block, however it is believed
431 that the net effect is to speed things up]. This isn't done for hard-regs
432 because recording call-clobbered hard-regs in `reg_set_table' at each
433 function call can consume a fair bit of memory, and iterating over
434 hard-regs stored this way in compute_transp will be more expensive. */
436 typedef struct reg_set
438 /* The next setting of this register. */
439 struct reg_set *next;
440 /* The insn where it was set. */
444 static reg_set **reg_set_table;
446 /* Size of `reg_set_table'.
447 The table starts out at max_gcse_regno + slop, and is enlarged as
449 static int reg_set_table_size;
451 /* Amount to grow `reg_set_table' by when it's full. */
452 #define REG_SET_TABLE_SLOP 100
454 /* This is a list of expressions which are MEMs and will be used by load
456 Load motion tracks MEMs which aren't killed by
457 anything except itself. (ie, loads and stores to a single location).
458 We can then allow movement of these MEM refs with a little special
459 allowance. (all stores copy the same value to the reaching reg used
460 for the loads). This means all values used to store into memory must have
461 no side effects so we can re-issue the setter value.
462 Store Motion uses this structure as an expression table to track stores
463 which look interesting, and might be moveable towards the exit block. */
467 struct expr * expr; /* Gcse expression reference for LM. */
468 rtx pattern; /* Pattern of this mem. */
469 rtx loads; /* INSN list of loads seen. */
470 rtx stores; /* INSN list of stores seen. */
471 struct ls_expr * next; /* Next in the list. */
472 int invalid; /* Invalid for some reason. */
473 int index; /* If it maps to a bitmap index. */
474 int hash_index; /* Index when in a hash table. */
475 rtx reaching_reg; /* Register to use when re-writing. */
478 /* Head of the list of load/store memory refs. */
479 static struct ls_expr * pre_ldst_mems = NULL;
481 /* Bitmap containing one bit for each register in the program.
482 Used when performing GCSE to track which registers have been set since
483 the start of the basic block. */
484 static sbitmap reg_set_bitmap;
486 /* For each block, a bitmap of registers set in the block.
487 This is used by expr_killed_p and compute_transp.
488 It is computed during hash table computation and not by compute_sets
489 as it includes registers added since the last pass (or between cprop and
490 gcse) and it's currently not easy to realloc sbitmap vectors. */
491 static sbitmap *reg_set_in_block;
493 /* Array, indexed by basic block number for a list of insns which modify
494 memory within that block. */
495 static rtx * modify_mem_list;
497 /* This array parallels modify_mem_list, but is kept canonicalized. */
498 static rtx * canon_modify_mem_list;
499 /* Various variables for statistics gathering. */
501 /* Memory used in a pass.
502 This isn't intended to be absolutely precise. Its intent is only
503 to keep an eye on memory usage. */
504 static int bytes_used;
506 /* GCSE substitutions made. */
507 static int gcse_subst_count;
508 /* Number of copy instructions created. */
509 static int gcse_create_count;
510 /* Number of constants propagated. */
511 static int const_prop_count;
512 /* Number of copys propagated. */
513 static int copy_prop_count;
515 /* These variables are used by classic GCSE.
516 Normally they'd be defined a bit later, but `rd_gen' needs to
517 be declared sooner. */
519 /* Each block has a bitmap of each type.
520 The length of each blocks bitmap is:
522 max_cuid - for reaching definitions
523 n_exprs - for available expressions
525 Thus we view the bitmaps as 2 dimensional arrays. i.e.
526 rd_kill[block_num][cuid_num]
527 ae_kill[block_num][expr_num] */
529 /* For reaching defs */
530 static sbitmap *rd_kill, *rd_gen, *reaching_defs, *rd_out;
532 /* for available exprs */
533 static sbitmap *ae_kill, *ae_gen, *ae_in, *ae_out;
535 /* Objects of this type are passed around by the null-pointer check
537 struct null_pointer_info
539 /* The basic block being processed. */
541 /* The first register to be handled in this pass. */
542 unsigned int min_reg;
543 /* One greater than the last register to be handled in this pass. */
544 unsigned int max_reg;
545 sbitmap *nonnull_local;
546 sbitmap *nonnull_killed;
549 static void compute_can_copy PARAMS ((void));
550 static char *gmalloc PARAMS ((unsigned int));
551 static char *grealloc PARAMS ((char *, unsigned int));
552 static char *gcse_alloc PARAMS ((unsigned long));
553 static void alloc_gcse_mem PARAMS ((rtx));
554 static void free_gcse_mem PARAMS ((void));
555 static void alloc_reg_set_mem PARAMS ((int));
556 static void free_reg_set_mem PARAMS ((void));
557 static int get_bitmap_width PARAMS ((int, int, int));
558 static void record_one_set PARAMS ((int, rtx));
559 static void record_set_info PARAMS ((rtx, rtx, void *));
560 static void compute_sets PARAMS ((rtx));
561 static void hash_scan_insn PARAMS ((rtx, int, int));
562 static void hash_scan_set PARAMS ((rtx, rtx, int));
563 static void hash_scan_clobber PARAMS ((rtx, rtx));
564 static void hash_scan_call PARAMS ((rtx, rtx));
565 static int want_to_gcse_p PARAMS ((rtx));
566 static int oprs_unchanged_p PARAMS ((rtx, rtx, int));
567 static int oprs_anticipatable_p PARAMS ((rtx, rtx));
568 static int oprs_available_p PARAMS ((rtx, rtx));
569 static void insert_expr_in_table PARAMS ((rtx, enum machine_mode, rtx,
571 static void insert_set_in_table PARAMS ((rtx, rtx));
572 static unsigned int hash_expr PARAMS ((rtx, enum machine_mode, int *, int));
573 static unsigned int hash_expr_1 PARAMS ((rtx, enum machine_mode, int *));
574 static unsigned int hash_string_1 PARAMS ((const char *));
575 static unsigned int hash_set PARAMS ((int, int));
576 static int expr_equiv_p PARAMS ((rtx, rtx));
577 static void record_last_reg_set_info PARAMS ((rtx, int));
578 static void record_last_mem_set_info PARAMS ((rtx));
579 static void record_last_set_info PARAMS ((rtx, rtx, void *));
580 static void compute_hash_table PARAMS ((int));
581 static void alloc_set_hash_table PARAMS ((int));
582 static void free_set_hash_table PARAMS ((void));
583 static void compute_set_hash_table PARAMS ((void));
584 static void alloc_expr_hash_table PARAMS ((unsigned int));
585 static void free_expr_hash_table PARAMS ((void));
586 static void compute_expr_hash_table PARAMS ((void));
587 static void dump_hash_table PARAMS ((FILE *, const char *, struct expr **,
589 static struct expr *lookup_expr PARAMS ((rtx));
590 static struct expr *lookup_set PARAMS ((unsigned int, rtx));
591 static struct expr *next_set PARAMS ((unsigned int, struct expr *));
592 static void reset_opr_set_tables PARAMS ((void));
593 static int oprs_not_set_p PARAMS ((rtx, rtx));
594 static void mark_call PARAMS ((rtx));
595 static void mark_set PARAMS ((rtx, rtx));
596 static void mark_clobber PARAMS ((rtx, rtx));
597 static void mark_oprs_set PARAMS ((rtx));
598 static void alloc_cprop_mem PARAMS ((int, int));
599 static void free_cprop_mem PARAMS ((void));
600 static void compute_transp PARAMS ((rtx, int, sbitmap *, int));
601 static void compute_transpout PARAMS ((void));
602 static void compute_local_properties PARAMS ((sbitmap *, sbitmap *, sbitmap *,
604 static void compute_cprop_data PARAMS ((void));
605 static void find_used_regs PARAMS ((rtx *, void *));
606 static int try_replace_reg PARAMS ((rtx, rtx, rtx));
607 static struct expr *find_avail_set PARAMS ((int, rtx));
608 static int cprop_jump PARAMS ((basic_block, rtx, rtx, rtx));
610 static int cprop_cc0_jump PARAMS ((basic_block, rtx, struct reg_use *, rtx));
612 static void mems_conflict_for_gcse_p PARAMS ((rtx, rtx, void *));
613 static int load_killed_in_block_p PARAMS ((basic_block, int, rtx, int));
614 static void canon_list_insert PARAMS ((rtx, rtx, void *));
615 static int cprop_insn PARAMS ((basic_block, rtx, int));
616 static int cprop PARAMS ((int));
617 static int one_cprop_pass PARAMS ((int, int));
618 static void alloc_pre_mem PARAMS ((int, int));
619 static void free_pre_mem PARAMS ((void));
620 static void compute_pre_data PARAMS ((void));
621 static int pre_expr_reaches_here_p PARAMS ((basic_block, struct expr *,
623 static void insert_insn_end_bb PARAMS ((struct expr *, basic_block, int));
624 static void pre_insert_copy_insn PARAMS ((struct expr *, rtx));
625 static void pre_insert_copies PARAMS ((void));
626 static int pre_delete PARAMS ((void));
627 static int pre_gcse PARAMS ((void));
628 static int one_pre_gcse_pass PARAMS ((int));
629 static void add_label_notes PARAMS ((rtx, rtx));
630 static void alloc_code_hoist_mem PARAMS ((int, int));
631 static void free_code_hoist_mem PARAMS ((void));
632 static void compute_code_hoist_vbeinout PARAMS ((void));
633 static void compute_code_hoist_data PARAMS ((void));
634 static int hoist_expr_reaches_here_p PARAMS ((basic_block, int, basic_block,
636 static void hoist_code PARAMS ((void));
637 static int one_code_hoisting_pass PARAMS ((void));
638 static void alloc_rd_mem PARAMS ((int, int));
639 static void free_rd_mem PARAMS ((void));
640 static void handle_rd_kill_set PARAMS ((rtx, int, basic_block));
641 static void compute_kill_rd PARAMS ((void));
642 static void compute_rd PARAMS ((void));
643 static void alloc_avail_expr_mem PARAMS ((int, int));
644 static void free_avail_expr_mem PARAMS ((void));
645 static void compute_ae_gen PARAMS ((void));
646 static int expr_killed_p PARAMS ((rtx, basic_block));
647 static void compute_ae_kill PARAMS ((sbitmap *, sbitmap *));
648 static int expr_reaches_here_p PARAMS ((struct occr *, struct expr *,
650 static rtx computing_insn PARAMS ((struct expr *, rtx));
651 static int def_reaches_here_p PARAMS ((rtx, rtx));
652 static int can_disregard_other_sets PARAMS ((struct reg_set **, rtx, int));
653 static int handle_avail_expr PARAMS ((rtx, struct expr *));
654 static int classic_gcse PARAMS ((void));
655 static int one_classic_gcse_pass PARAMS ((int));
656 static void invalidate_nonnull_info PARAMS ((rtx, rtx, void *));
657 static void delete_null_pointer_checks_1 PARAMS ((varray_type *, unsigned int *,
658 sbitmap *, sbitmap *,
659 struct null_pointer_info *));
660 static rtx process_insert_insn PARAMS ((struct expr *));
661 static int pre_edge_insert PARAMS ((struct edge_list *, struct expr **));
662 static int expr_reaches_here_p_work PARAMS ((struct occr *, struct expr *,
663 basic_block, int, char *));
664 static int pre_expr_reaches_here_p_work PARAMS ((basic_block, struct expr *,
665 basic_block, char *));
666 static struct ls_expr * ldst_entry PARAMS ((rtx));
667 static void free_ldst_entry PARAMS ((struct ls_expr *));
668 static void free_ldst_mems PARAMS ((void));
669 static void print_ldst_list PARAMS ((FILE *));
670 static struct ls_expr * find_rtx_in_ldst PARAMS ((rtx));
671 static int enumerate_ldsts PARAMS ((void));
672 static inline struct ls_expr * first_ls_expr PARAMS ((void));
673 static inline struct ls_expr * next_ls_expr PARAMS ((struct ls_expr *));
674 static int simple_mem PARAMS ((rtx));
675 static void invalidate_any_buried_refs PARAMS ((rtx));
676 static void compute_ld_motion_mems PARAMS ((void));
677 static void trim_ld_motion_mems PARAMS ((void));
678 static void update_ld_motion_stores PARAMS ((struct expr *));
679 static void reg_set_info PARAMS ((rtx, rtx, void *));
680 static int store_ops_ok PARAMS ((rtx, basic_block));
681 static void find_moveable_store PARAMS ((rtx));
682 static int compute_store_table PARAMS ((void));
683 static int load_kills_store PARAMS ((rtx, rtx));
684 static int find_loads PARAMS ((rtx, rtx));
685 static int store_killed_in_insn PARAMS ((rtx, rtx));
686 static int store_killed_after PARAMS ((rtx, rtx, basic_block));
687 static int store_killed_before PARAMS ((rtx, rtx, basic_block));
688 static void build_store_vectors PARAMS ((void));
689 static void insert_insn_start_bb PARAMS ((rtx, basic_block));
690 static int insert_store PARAMS ((struct ls_expr *, edge));
691 static void replace_store_insn PARAMS ((rtx, rtx, basic_block));
692 static void delete_store PARAMS ((struct ls_expr *,
694 static void free_store_memory PARAMS ((void));
695 static void store_motion PARAMS ((void));
697 /* Entry point for global common subexpression elimination.
698 F is the first instruction in the function. */
706 /* Bytes used at start of pass. */
707 int initial_bytes_used;
708 /* Maximum number of bytes used by a pass. */
710 /* Point to release obstack data from for each pass. */
711 char *gcse_obstack_bottom;
713 /* Insertion of instructions on edges can create new basic blocks; we
714 need the original basic block count so that we can properly deallocate
715 arrays sized on the number of basic blocks originally in the cfg. */
717 /* We do not construct an accurate cfg in functions which call
718 setjmp, so just punt to be safe. */
719 if (current_function_calls_setjmp)
722 /* Assume that we do not need to run jump optimizations after gcse. */
723 run_jump_opt_after_gcse = 0;
725 /* For calling dump_foo fns from gdb. */
726 debug_stderr = stderr;
729 /* Identify the basic block information for this function, including
730 successors and predecessors. */
731 max_gcse_regno = max_reg_num ();
734 dump_flow_info (file);
736 orig_bb_count = n_basic_blocks;
737 /* Return if there's nothing to do. */
738 if (n_basic_blocks <= 1)
741 /* Trying to perform global optimizations on flow graphs which have
742 a high connectivity will take a long time and is unlikely to be
745 In normal circumstances a cfg should have about twice as many edges
746 as blocks. But we do not want to punish small functions which have
747 a couple switch statements. So we require a relatively large number
748 of basic blocks and the ratio of edges to blocks to be high. */
749 if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
751 if (warn_disabled_optimization)
752 warning ("GCSE disabled: %d > 1000 basic blocks and %d >= 20 edges/basic block",
753 n_basic_blocks, n_edges / n_basic_blocks);
757 /* If allocating memory for the cprop bitmap would take up too much
758 storage it's better just to disable the optimization. */
760 * SBITMAP_SET_SIZE (max_gcse_regno)
761 * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
763 if (warn_disabled_optimization)
764 warning ("GCSE disabled: %d basic blocks and %d registers",
765 n_basic_blocks, max_gcse_regno);
770 /* See what modes support reg/reg copy operations. */
771 if (! can_copy_init_p)
777 gcc_obstack_init (&gcse_obstack);
781 init_alias_analysis ();
782 /* Record where pseudo-registers are set. This data is kept accurate
783 during each pass. ??? We could also record hard-reg information here
784 [since it's unchanging], however it is currently done during hash table
787 It may be tempting to compute MEM set information here too, but MEM sets
788 will be subject to code motion one day and thus we need to compute
789 information about memory sets when we build the hash tables. */
791 alloc_reg_set_mem (max_gcse_regno);
795 initial_bytes_used = bytes_used;
797 gcse_obstack_bottom = gcse_alloc (1);
799 while (changed && pass < MAX_GCSE_PASSES)
803 fprintf (file, "GCSE pass %d\n\n", pass + 1);
805 /* Initialize bytes_used to the space for the pred/succ lists,
806 and the reg_set_table data. */
807 bytes_used = initial_bytes_used;
809 /* Each pass may create new registers, so recalculate each time. */
810 max_gcse_regno = max_reg_num ();
814 /* Don't allow constant propagation to modify jumps
816 changed = one_cprop_pass (pass + 1, 0);
819 changed |= one_classic_gcse_pass (pass + 1);
822 changed |= one_pre_gcse_pass (pass + 1);
823 /* We may have just created new basic blocks. Release and
824 recompute various things which are sized on the number of
830 for (i = 0; i < orig_bb_count; i++)
832 if (modify_mem_list[i])
833 free_INSN_LIST_list (modify_mem_list + i);
834 if (canon_modify_mem_list[i])
835 free_INSN_LIST_list (canon_modify_mem_list + i);
838 = (rtx *) gmalloc (n_basic_blocks * sizeof (rtx *));
839 canon_modify_mem_list
840 = (rtx *) gmalloc (n_basic_blocks * sizeof (rtx *));
841 memset ((char *) modify_mem_list, 0, n_basic_blocks * sizeof (rtx *));
842 memset ((char *) canon_modify_mem_list, 0, n_basic_blocks * sizeof (rtx *));
843 orig_bb_count = n_basic_blocks;
846 alloc_reg_set_mem (max_reg_num ());
848 run_jump_opt_after_gcse = 1;
851 if (max_pass_bytes < bytes_used)
852 max_pass_bytes = bytes_used;
854 /* Free up memory, then reallocate for code hoisting. We can
855 not re-use the existing allocated memory because the tables
856 will not have info for the insns or registers created by
857 partial redundancy elimination. */
860 /* It does not make sense to run code hoisting unless we optimizing
861 for code size -- it rarely makes programs faster, and can make
862 them bigger if we did partial redundancy elimination (when optimizing
863 for space, we use a classic gcse algorithm instead of partial
864 redundancy algorithms). */
867 max_gcse_regno = max_reg_num ();
869 changed |= one_code_hoisting_pass ();
872 if (max_pass_bytes < bytes_used)
873 max_pass_bytes = bytes_used;
878 fprintf (file, "\n");
882 obstack_free (&gcse_obstack, gcse_obstack_bottom);
886 /* Do one last pass of copy propagation, including cprop into
887 conditional jumps. */
889 max_gcse_regno = max_reg_num ();
891 /* This time, go ahead and allow cprop to alter jumps. */
892 one_cprop_pass (pass + 1, 1);
897 fprintf (file, "GCSE of %s: %d basic blocks, ",
898 current_function_name, n_basic_blocks);
899 fprintf (file, "%d pass%s, %d bytes\n\n",
900 pass, pass > 1 ? "es" : "", max_pass_bytes);
903 obstack_free (&gcse_obstack, NULL);
905 /* We are finished with alias. */
906 end_alias_analysis ();
907 allocate_reg_info (max_reg_num (), FALSE, FALSE);
909 if (!optimize_size && flag_gcse_sm)
911 /* Record where pseudo-registers are set. */
912 return run_jump_opt_after_gcse;
915 /* Misc. utilities. */
917 /* Compute which modes support reg/reg copy operations. */
923 #ifndef AVOID_CCMODE_COPIES
926 memset (can_copy_p, 0, NUM_MACHINE_MODES);
929 for (i = 0; i < NUM_MACHINE_MODES; i++)
930 if (GET_MODE_CLASS (i) == MODE_CC)
932 #ifdef AVOID_CCMODE_COPIES
935 reg = gen_rtx_REG ((enum machine_mode) i, LAST_VIRTUAL_REGISTER + 1);
936 insn = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
937 if (recog (PATTERN (insn), insn, NULL) >= 0)
947 /* Cover function to xmalloc to record bytes allocated. */
954 return xmalloc (size);
957 /* Cover function to xrealloc.
958 We don't record the additional size since we don't know it.
959 It won't affect memory usage stats much anyway. */
966 return xrealloc (ptr, size);
969 /* Cover function to obstack_alloc.
970 We don't need to record the bytes allocated here since
971 obstack_chunk_alloc is set to gmalloc. */
977 return (char *) obstack_alloc (&gcse_obstack, size);
980 /* Allocate memory for the cuid mapping array,
981 and reg/memory set tracking tables.
983 This is called at the start of each pass. */
992 /* Find the largest UID and create a mapping from UIDs to CUIDs.
993 CUIDs are like UIDs except they increase monotonically, have no gaps,
994 and only apply to real insns. */
996 max_uid = get_max_uid ();
997 n = (max_uid + 1) * sizeof (int);
998 uid_cuid = (int *) gmalloc (n);
999 memset ((char *) uid_cuid, 0, n);
1000 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
1003 uid_cuid[INSN_UID (insn)] = i++;
1005 uid_cuid[INSN_UID (insn)] = i;
1008 /* Create a table mapping cuids to insns. */
1011 n = (max_cuid + 1) * sizeof (rtx);
1012 cuid_insn = (rtx *) gmalloc (n);
1013 memset ((char *) cuid_insn, 0, n);
1014 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
1016 CUID_INSN (i++) = insn;
1018 /* Allocate vars to track sets of regs. */
1019 reg_set_bitmap = (sbitmap) sbitmap_alloc (max_gcse_regno);
1021 /* Allocate vars to track sets of regs, memory per block. */
1022 reg_set_in_block = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks,
1024 /* Allocate array to keep a list of insns which modify memory in each
1026 modify_mem_list = (rtx *) gmalloc (n_basic_blocks * sizeof (rtx *));
1027 canon_modify_mem_list = (rtx *) gmalloc (n_basic_blocks * sizeof (rtx *));
1028 memset ((char *) modify_mem_list, 0, n_basic_blocks * sizeof (rtx *));
1029 memset ((char *) canon_modify_mem_list, 0, n_basic_blocks * sizeof (rtx *));
1032 /* Free memory allocated by alloc_gcse_mem. */
1040 free (reg_set_bitmap);
1042 sbitmap_vector_free (reg_set_in_block);
1043 /* re-Cache any INSN_LIST nodes we have allocated. */
1047 for (i = 0; i < n_basic_blocks; i++)
1049 if (modify_mem_list[i])
1050 free_INSN_LIST_list (modify_mem_list + i);
1051 if (canon_modify_mem_list[i])
1052 free_INSN_LIST_list (canon_modify_mem_list + i);
1055 free (modify_mem_list);
1056 free (canon_modify_mem_list);
1057 modify_mem_list = 0;
1058 canon_modify_mem_list = 0;
1062 /* Many of the global optimization algorithms work by solving dataflow
1063 equations for various expressions. Initially, some local value is
1064 computed for each expression in each block. Then, the values across the
1065 various blocks are combined (by following flow graph edges) to arrive at
1066 global values. Conceptually, each set of equations is independent. We
1067 may therefore solve all the equations in parallel, solve them one at a
1068 time, or pick any intermediate approach.
1070 When you're going to need N two-dimensional bitmaps, each X (say, the
1071 number of blocks) by Y (say, the number of expressions), call this
1072 function. It's not important what X and Y represent; only that Y
1073 correspond to the things that can be done in parallel. This function will
1074 return an appropriate chunking factor C; you should solve C sets of
1075 equations in parallel. By going through this function, we can easily
1076 trade space against time; by solving fewer equations in parallel we use
1080 get_bitmap_width (n, x, y)
1085 /* It's not really worth figuring out *exactly* how much memory will
1086 be used by a particular choice. The important thing is to get
1087 something approximately right. */
1088 size_t max_bitmap_memory = 10 * 1024 * 1024;
1090 /* The number of bytes we'd use for a single column of minimum
1092 size_t column_size = n * x * sizeof (SBITMAP_ELT_TYPE);
1094 /* Often, it's reasonable just to solve all the equations in
1096 if (column_size * SBITMAP_SET_SIZE (y) <= max_bitmap_memory)
1099 /* Otherwise, pick the largest width we can, without going over the
1101 return SBITMAP_ELT_BITS * ((max_bitmap_memory + column_size - 1)
1105 /* Compute the local properties of each recorded expression.
1107 Local properties are those that are defined by the block, irrespective of
1110 An expression is transparent in a block if its operands are not modified
1113 An expression is computed (locally available) in a block if it is computed
1114 at least once and expression would contain the same value if the
1115 computation was moved to the end of the block.
1117 An expression is locally anticipatable in a block if it is computed at
1118 least once and expression would contain the same value if the computation
1119 was moved to the beginning of the block.
1121 We call this routine for cprop, pre and code hoisting. They all compute
1122 basically the same information and thus can easily share this code.
1124 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
1125 properties. If NULL, then it is not necessary to compute or record that
1126 particular property.
1128 SETP controls which hash table to look at. If zero, this routine looks at
1129 the expr hash table; if nonzero this routine looks at the set hash table.
1130 Additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
1134 compute_local_properties (transp, comp, antloc, setp)
1140 unsigned int i, hash_table_size;
1141 struct expr **hash_table;
1143 /* Initialize any bitmaps that were passed in. */
1147 sbitmap_vector_zero (transp, n_basic_blocks);
1149 sbitmap_vector_ones (transp, n_basic_blocks);
1153 sbitmap_vector_zero (comp, n_basic_blocks);
1155 sbitmap_vector_zero (antloc, n_basic_blocks);
1157 /* We use the same code for cprop, pre and hoisting. For cprop
1158 we care about the set hash table, for pre and hoisting we
1159 care about the expr hash table. */
1160 hash_table_size = setp ? set_hash_table_size : expr_hash_table_size;
1161 hash_table = setp ? set_hash_table : expr_hash_table;
1163 for (i = 0; i < hash_table_size; i++)
1167 for (expr = hash_table[i]; expr != NULL; expr = expr->next_same_hash)
1169 int indx = expr->bitmap_index;
1172 /* The expression is transparent in this block if it is not killed.
1173 We start by assuming all are transparent [none are killed], and
1174 then reset the bits for those that are. */
1176 compute_transp (expr->expr, indx, transp, setp);
1178 /* The occurrences recorded in antic_occr are exactly those that
1179 we want to set to non-zero in ANTLOC. */
1181 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
1183 SET_BIT (antloc[BLOCK_NUM (occr->insn)], indx);
1185 /* While we're scanning the table, this is a good place to
1187 occr->deleted_p = 0;
1190 /* The occurrences recorded in avail_occr are exactly those that
1191 we want to set to non-zero in COMP. */
1193 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
1195 SET_BIT (comp[BLOCK_NUM (occr->insn)], indx);
1197 /* While we're scanning the table, this is a good place to
1202 /* While we're scanning the table, this is a good place to
1204 expr->reaching_reg = 0;
1209 /* Register set information.
1211 `reg_set_table' records where each register is set or otherwise
1214 static struct obstack reg_set_obstack;
1217 alloc_reg_set_mem (n_regs)
1222 reg_set_table_size = n_regs + REG_SET_TABLE_SLOP;
1223 n = reg_set_table_size * sizeof (struct reg_set *);
1224 reg_set_table = (struct reg_set **) gmalloc (n);
1225 memset ((char *) reg_set_table, 0, n);
1227 gcc_obstack_init (®_set_obstack);
1233 free (reg_set_table);
1234 obstack_free (®_set_obstack, NULL);
1237 /* Record REGNO in the reg_set table. */
1240 record_one_set (regno, insn)
1244 /* Allocate a new reg_set element and link it onto the list. */
1245 struct reg_set *new_reg_info;
1247 /* If the table isn't big enough, enlarge it. */
1248 if (regno >= reg_set_table_size)
1250 int new_size = regno + REG_SET_TABLE_SLOP;
1253 = (struct reg_set **) grealloc ((char *) reg_set_table,
1254 new_size * sizeof (struct reg_set *));
1255 memset ((char *) (reg_set_table + reg_set_table_size), 0,
1256 (new_size - reg_set_table_size) * sizeof (struct reg_set *));
1257 reg_set_table_size = new_size;
1260 new_reg_info = (struct reg_set *) obstack_alloc (®_set_obstack,
1261 sizeof (struct reg_set));
1262 bytes_used += sizeof (struct reg_set);
1263 new_reg_info->insn = insn;
1264 new_reg_info->next = reg_set_table[regno];
1265 reg_set_table[regno] = new_reg_info;
1268 /* Called from compute_sets via note_stores to handle one SET or CLOBBER in
1269 an insn. The DATA is really the instruction in which the SET is
1273 record_set_info (dest, setter, data)
1274 rtx dest, setter ATTRIBUTE_UNUSED;
1277 rtx record_set_insn = (rtx) data;
1279 if (GET_CODE (dest) == REG && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
1280 record_one_set (REGNO (dest), record_set_insn);
1283 /* Scan the function and record each set of each pseudo-register.
1285 This is called once, at the start of the gcse pass. See the comments for
1286 `reg_set_table' for further documenation. */
1294 for (insn = f; insn != 0; insn = NEXT_INSN (insn))
1296 note_stores (PATTERN (insn), record_set_info, insn);
1299 /* Hash table support. */
1301 /* For each register, the cuid of the first/last insn in the block
1302 that set it, or -1 if not set. */
1303 #define NEVER_SET -1
1305 struct reg_avail_info
1312 static struct reg_avail_info *reg_avail_info;
1313 static int current_bb;
1316 /* See whether X, the source of a set, is something we want to consider for
1323 static rtx test_insn = 0;
1324 int num_clobbers = 0;
1327 switch (GET_CODE (x))
1340 /* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
1341 if (general_operand (x, GET_MODE (x)))
1343 else if (GET_MODE (x) == VOIDmode)
1346 /* Otherwise, check if we can make a valid insn from it. First initialize
1347 our test insn if we haven't already. */
1351 = make_insn_raw (gen_rtx_SET (VOIDmode,
1352 gen_rtx_REG (word_mode,
1353 FIRST_PSEUDO_REGISTER * 2),
1355 NEXT_INSN (test_insn) = PREV_INSN (test_insn) = 0;
1356 ggc_add_rtx_root (&test_insn, 1);
1359 /* Now make an insn like the one we would make when GCSE'ing and see if
1361 PUT_MODE (SET_DEST (PATTERN (test_insn)), GET_MODE (x));
1362 SET_SRC (PATTERN (test_insn)) = x;
1363 return ((icode = recog (PATTERN (test_insn), test_insn, &num_clobbers)) >= 0
1364 && (num_clobbers == 0 || ! added_clobbers_hard_reg_p (icode)));
1367 /* Return non-zero if the operands of expression X are unchanged from the
1368 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1369 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1372 oprs_unchanged_p (x, insn, avail_p)
1383 code = GET_CODE (x);
1388 struct reg_avail_info *info = ®_avail_info[REGNO (x)];
1390 if (info->last_bb != current_bb)
1393 return info->last_set < INSN_CUID (insn);
1395 return info->first_set >= INSN_CUID (insn);
1399 if (load_killed_in_block_p (BASIC_BLOCK (current_bb), INSN_CUID (insn),
1403 return oprs_unchanged_p (XEXP (x, 0), insn, avail_p);
1428 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1432 /* If we are about to do the last recursive call needed at this
1433 level, change it into iteration. This function is called enough
1436 return oprs_unchanged_p (XEXP (x, i), insn, avail_p);
1438 else if (! oprs_unchanged_p (XEXP (x, i), insn, avail_p))
1441 else if (fmt[i] == 'E')
1442 for (j = 0; j < XVECLEN (x, i); j++)
1443 if (! oprs_unchanged_p (XVECEXP (x, i, j), insn, avail_p))
1450 /* Used for communication between mems_conflict_for_gcse_p and
1451 load_killed_in_block_p. Nonzero if mems_conflict_for_gcse_p finds a
1452 conflict between two memory references. */
1453 static int gcse_mems_conflict_p;
1455 /* Used for communication between mems_conflict_for_gcse_p and
1456 load_killed_in_block_p. A memory reference for a load instruction,
1457 mems_conflict_for_gcse_p will see if a memory store conflicts with
1458 this memory load. */
1459 static rtx gcse_mem_operand;
1461 /* DEST is the output of an instruction. If it is a memory reference, and
1462 possibly conflicts with the load found in gcse_mem_operand, then set
1463 gcse_mems_conflict_p to a nonzero value. */
1466 mems_conflict_for_gcse_p (dest, setter, data)
1467 rtx dest, setter ATTRIBUTE_UNUSED;
1468 void *data ATTRIBUTE_UNUSED;
1470 while (GET_CODE (dest) == SUBREG
1471 || GET_CODE (dest) == ZERO_EXTRACT
1472 || GET_CODE (dest) == SIGN_EXTRACT
1473 || GET_CODE (dest) == STRICT_LOW_PART)
1474 dest = XEXP (dest, 0);
1476 /* If DEST is not a MEM, then it will not conflict with the load. Note
1477 that function calls are assumed to clobber memory, but are handled
1479 if (GET_CODE (dest) != MEM)
1482 /* If we are setting a MEM in our list of specially recognized MEMs,
1483 don't mark as killed this time. */
1485 if (dest == gcse_mem_operand && pre_ldst_mems != NULL)
1487 if (!find_rtx_in_ldst (dest))
1488 gcse_mems_conflict_p = 1;
1492 if (true_dependence (dest, GET_MODE (dest), gcse_mem_operand,
1494 gcse_mems_conflict_p = 1;
1497 /* Return nonzero if the expression in X (a memory reference) is killed
1498 in block BB before or after the insn with the CUID in UID_LIMIT.
1499 AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills
1502 To check the entire block, set UID_LIMIT to max_uid + 1 and
1506 load_killed_in_block_p (bb, uid_limit, x, avail_p)
1512 rtx list_entry = modify_mem_list[bb->index];
1516 /* Ignore entries in the list that do not apply. */
1518 && INSN_CUID (XEXP (list_entry, 0)) < uid_limit)
1520 && INSN_CUID (XEXP (list_entry, 0)) > uid_limit))
1522 list_entry = XEXP (list_entry, 1);
1526 setter = XEXP (list_entry, 0);
1528 /* If SETTER is a call everything is clobbered. Note that calls
1529 to pure functions are never put on the list, so we need not
1530 worry about them. */
1531 if (GET_CODE (setter) == CALL_INSN)
1534 /* SETTER must be an INSN of some kind that sets memory. Call
1535 note_stores to examine each hunk of memory that is modified.
1537 The note_stores interface is pretty limited, so we have to
1538 communicate via global variables. Yuk. */
1539 gcse_mem_operand = x;
1540 gcse_mems_conflict_p = 0;
1541 note_stores (PATTERN (setter), mems_conflict_for_gcse_p, NULL);
1542 if (gcse_mems_conflict_p)
1544 list_entry = XEXP (list_entry, 1);
1549 /* Return non-zero if the operands of expression X are unchanged from
1550 the start of INSN's basic block up to but not including INSN. */
1553 oprs_anticipatable_p (x, insn)
1556 return oprs_unchanged_p (x, insn, 0);
1559 /* Return non-zero if the operands of expression X are unchanged from
1560 INSN to the end of INSN's basic block. */
1563 oprs_available_p (x, insn)
1566 return oprs_unchanged_p (x, insn, 1);
1569 /* Hash expression X.
1571 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1572 indicating if a volatile operand is found or if the expression contains
1573 something we don't want to insert in the table.
1575 ??? One might want to merge this with canon_hash. Later. */
1578 hash_expr (x, mode, do_not_record_p, hash_table_size)
1580 enum machine_mode mode;
1581 int *do_not_record_p;
1582 int hash_table_size;
1586 *do_not_record_p = 0;
1588 hash = hash_expr_1 (x, mode, do_not_record_p);
1589 return hash % hash_table_size;
1592 /* Hash a string. Just add its bytes up. */
1594 static inline unsigned
1599 const unsigned char *p = (const unsigned char *)ps;
1608 /* Subroutine of hash_expr to do the actual work. */
1611 hash_expr_1 (x, mode, do_not_record_p)
1613 enum machine_mode mode;
1614 int *do_not_record_p;
1621 /* Used to turn recursion into iteration. We can't rely on GCC's
1622 tail-recursion eliminatio since we need to keep accumulating values
1629 code = GET_CODE (x);
1633 hash += ((unsigned int) REG << 7) + REGNO (x);
1637 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
1638 + (unsigned int) INTVAL (x));
1642 /* This is like the general case, except that it only counts
1643 the integers representing the constant. */
1644 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
1645 if (GET_MODE (x) != VOIDmode)
1646 for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++)
1647 hash += (unsigned int) XWINT (x, i);
1649 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
1650 + (unsigned int) CONST_DOUBLE_HIGH (x));
1653 /* Assume there is only one rtx object for any given label. */
1655 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
1656 differences and differences between each stage's debugging dumps. */
1657 hash += (((unsigned int) LABEL_REF << 7)
1658 + CODE_LABEL_NUMBER (XEXP (x, 0)));
1663 /* Don't hash on the symbol's address to avoid bootstrap differences.
1664 Different hash values may cause expressions to be recorded in
1665 different orders and thus different registers to be used in the
1666 final assembler. This also avoids differences in the dump files
1667 between various stages. */
1669 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
1672 h += (h << 7) + *p++; /* ??? revisit */
1674 hash += ((unsigned int) SYMBOL_REF << 7) + h;
1679 if (MEM_VOLATILE_P (x))
1681 *do_not_record_p = 1;
1685 hash += (unsigned int) MEM;
1686 hash += MEM_ALIAS_SET (x);
1697 case UNSPEC_VOLATILE:
1698 *do_not_record_p = 1;
1702 if (MEM_VOLATILE_P (x))
1704 *do_not_record_p = 1;
1709 /* We don't want to take the filename and line into account. */
1710 hash += (unsigned) code + (unsigned) GET_MODE (x)
1711 + hash_string_1 (ASM_OPERANDS_TEMPLATE (x))
1712 + hash_string_1 (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
1713 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
1715 if (ASM_OPERANDS_INPUT_LENGTH (x))
1717 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
1719 hash += (hash_expr_1 (ASM_OPERANDS_INPUT (x, i),
1720 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
1722 + hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT
1726 hash += hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
1727 x = ASM_OPERANDS_INPUT (x, 0);
1728 mode = GET_MODE (x);
1738 hash += (unsigned) code + (unsigned) GET_MODE (x);
1739 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1743 /* If we are about to do the last recursive call
1744 needed at this level, change it into iteration.
1745 This function is called enough to be worth it. */
1752 hash += hash_expr_1 (XEXP (x, i), 0, do_not_record_p);
1753 if (*do_not_record_p)
1757 else if (fmt[i] == 'E')
1758 for (j = 0; j < XVECLEN (x, i); j++)
1760 hash += hash_expr_1 (XVECEXP (x, i, j), 0, do_not_record_p);
1761 if (*do_not_record_p)
1765 else if (fmt[i] == 's')
1766 hash += hash_string_1 (XSTR (x, i));
1767 else if (fmt[i] == 'i')
1768 hash += (unsigned int) XINT (x, i);
1776 /* Hash a set of register REGNO.
1778 Sets are hashed on the register that is set. This simplifies the PRE copy
1781 ??? May need to make things more elaborate. Later, as necessary. */
1784 hash_set (regno, hash_table_size)
1786 int hash_table_size;
1791 return hash % hash_table_size;
1794 /* Return non-zero if exp1 is equivalent to exp2.
1795 ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
1802 register enum rtx_code code;
1803 register const char *fmt;
1808 if (x == 0 || y == 0)
1811 code = GET_CODE (x);
1812 if (code != GET_CODE (y))
1815 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
1816 if (GET_MODE (x) != GET_MODE (y))
1826 return INTVAL (x) == INTVAL (y);
1829 return XEXP (x, 0) == XEXP (y, 0);
1832 return XSTR (x, 0) == XSTR (y, 0);
1835 return REGNO (x) == REGNO (y);
1838 /* Can't merge two expressions in different alias sets, since we can
1839 decide that the expression is transparent in a block when it isn't,
1840 due to it being set with the different alias set. */
1841 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
1845 /* For commutative operations, check both orders. */
1853 return ((expr_equiv_p (XEXP (x, 0), XEXP (y, 0))
1854 && expr_equiv_p (XEXP (x, 1), XEXP (y, 1)))
1855 || (expr_equiv_p (XEXP (x, 0), XEXP (y, 1))
1856 && expr_equiv_p (XEXP (x, 1), XEXP (y, 0))));
1859 /* We don't use the generic code below because we want to
1860 disregard filename and line numbers. */
1862 /* A volatile asm isn't equivalent to any other. */
1863 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
1866 if (GET_MODE (x) != GET_MODE (y)
1867 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
1868 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
1869 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
1870 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
1871 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
1874 if (ASM_OPERANDS_INPUT_LENGTH (x))
1876 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
1877 if (! expr_equiv_p (ASM_OPERANDS_INPUT (x, i),
1878 ASM_OPERANDS_INPUT (y, i))
1879 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
1880 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
1890 /* Compare the elements. If any pair of corresponding elements
1891 fail to match, return 0 for the whole thing. */
1893 fmt = GET_RTX_FORMAT (code);
1894 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1899 if (! expr_equiv_p (XEXP (x, i), XEXP (y, i)))
1904 if (XVECLEN (x, i) != XVECLEN (y, i))
1906 for (j = 0; j < XVECLEN (x, i); j++)
1907 if (! expr_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j)))
1912 if (strcmp (XSTR (x, i), XSTR (y, i)))
1917 if (XINT (x, i) != XINT (y, i))
1922 if (XWINT (x, i) != XWINT (y, i))
1937 /* Insert expression X in INSN in the hash table.
1938 If it is already present, record it as the last occurrence in INSN's
1941 MODE is the mode of the value X is being stored into.
1942 It is only used if X is a CONST_INT.
1944 ANTIC_P is non-zero if X is an anticipatable expression.
1945 AVAIL_P is non-zero if X is an available expression. */
1948 insert_expr_in_table (x, mode, insn, antic_p, avail_p)
1950 enum machine_mode mode;
1952 int antic_p, avail_p;
1954 int found, do_not_record_p;
1956 struct expr *cur_expr, *last_expr = NULL;
1957 struct occr *antic_occr, *avail_occr;
1958 struct occr *last_occr = NULL;
1960 hash = hash_expr (x, mode, &do_not_record_p, expr_hash_table_size);
1962 /* Do not insert expression in table if it contains volatile operands,
1963 or if hash_expr determines the expression is something we don't want
1964 to or can't handle. */
1965 if (do_not_record_p)
1968 cur_expr = expr_hash_table[hash];
1971 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1973 /* If the expression isn't found, save a pointer to the end of
1975 last_expr = cur_expr;
1976 cur_expr = cur_expr->next_same_hash;
1981 cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
1982 bytes_used += sizeof (struct expr);
1983 if (expr_hash_table[hash] == NULL)
1984 /* This is the first pattern that hashed to this index. */
1985 expr_hash_table[hash] = cur_expr;
1987 /* Add EXPR to end of this hash chain. */
1988 last_expr->next_same_hash = cur_expr;
1990 /* Set the fields of the expr element. */
1992 cur_expr->bitmap_index = n_exprs++;
1993 cur_expr->next_same_hash = NULL;
1994 cur_expr->antic_occr = NULL;
1995 cur_expr->avail_occr = NULL;
1998 /* Now record the occurrence(s). */
2001 antic_occr = cur_expr->antic_occr;
2003 /* Search for another occurrence in the same basic block. */
2004 while (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
2006 /* If an occurrence isn't found, save a pointer to the end of
2008 last_occr = antic_occr;
2009 antic_occr = antic_occr->next;
2013 /* Found another instance of the expression in the same basic block.
2014 Prefer the currently recorded one. We want the first one in the
2015 block and the block is scanned from start to end. */
2016 ; /* nothing to do */
2019 /* First occurrence of this expression in this basic block. */
2020 antic_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
2021 bytes_used += sizeof (struct occr);
2022 /* First occurrence of this expression in any block? */
2023 if (cur_expr->antic_occr == NULL)
2024 cur_expr->antic_occr = antic_occr;
2026 last_occr->next = antic_occr;
2028 antic_occr->insn = insn;
2029 antic_occr->next = NULL;
2035 avail_occr = cur_expr->avail_occr;
2037 /* Search for another occurrence in the same basic block. */
2038 while (avail_occr && BLOCK_NUM (avail_occr->insn) != BLOCK_NUM (insn))
2040 /* If an occurrence isn't found, save a pointer to the end of
2042 last_occr = avail_occr;
2043 avail_occr = avail_occr->next;
2047 /* Found another instance of the expression in the same basic block.
2048 Prefer this occurrence to the currently recorded one. We want
2049 the last one in the block and the block is scanned from start
2051 avail_occr->insn = insn;
2054 /* First occurrence of this expression in this basic block. */
2055 avail_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
2056 bytes_used += sizeof (struct occr);
2058 /* First occurrence of this expression in any block? */
2059 if (cur_expr->avail_occr == NULL)
2060 cur_expr->avail_occr = avail_occr;
2062 last_occr->next = avail_occr;
2064 avail_occr->insn = insn;
2065 avail_occr->next = NULL;
2070 /* Insert pattern X in INSN in the hash table.
2071 X is a SET of a reg to either another reg or a constant.
2072 If it is already present, record it as the last occurrence in INSN's
2076 insert_set_in_table (x, insn)
2082 struct expr *cur_expr, *last_expr = NULL;
2083 struct occr *cur_occr, *last_occr = NULL;
2085 if (GET_CODE (x) != SET
2086 || GET_CODE (SET_DEST (x)) != REG)
2089 hash = hash_set (REGNO (SET_DEST (x)), set_hash_table_size);
2091 cur_expr = set_hash_table[hash];
2094 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
2096 /* If the expression isn't found, save a pointer to the end of
2098 last_expr = cur_expr;
2099 cur_expr = cur_expr->next_same_hash;
2104 cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
2105 bytes_used += sizeof (struct expr);
2106 if (set_hash_table[hash] == NULL)
2107 /* This is the first pattern that hashed to this index. */
2108 set_hash_table[hash] = cur_expr;
2110 /* Add EXPR to end of this hash chain. */
2111 last_expr->next_same_hash = cur_expr;
2113 /* Set the fields of the expr element.
2114 We must copy X because it can be modified when copy propagation is
2115 performed on its operands. */
2116 cur_expr->expr = copy_rtx (x);
2117 cur_expr->bitmap_index = n_sets++;
2118 cur_expr->next_same_hash = NULL;
2119 cur_expr->antic_occr = NULL;
2120 cur_expr->avail_occr = NULL;
2123 /* Now record the occurrence. */
2124 cur_occr = cur_expr->avail_occr;
2126 /* Search for another occurrence in the same basic block. */
2127 while (cur_occr && BLOCK_NUM (cur_occr->insn) != BLOCK_NUM (insn))
2129 /* If an occurrence isn't found, save a pointer to the end of
2131 last_occr = cur_occr;
2132 cur_occr = cur_occr->next;
2136 /* Found another instance of the expression in the same basic block.
2137 Prefer this occurrence to the currently recorded one. We want the
2138 last one in the block and the block is scanned from start to end. */
2139 cur_occr->insn = insn;
2142 /* First occurrence of this expression in this basic block. */
2143 cur_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
2144 bytes_used += sizeof (struct occr);
2146 /* First occurrence of this expression in any block? */
2147 if (cur_expr->avail_occr == NULL)
2148 cur_expr->avail_occr = cur_occr;
2150 last_occr->next = cur_occr;
2152 cur_occr->insn = insn;
2153 cur_occr->next = NULL;
2157 /* Scan pattern PAT of INSN and add an entry to the hash table. If SET_P is
2158 non-zero, this is for the assignment hash table, otherwise it is for the
2159 expression hash table. */
2162 hash_scan_set (pat, insn, set_p)
2166 rtx src = SET_SRC (pat);
2167 rtx dest = SET_DEST (pat);
2170 if (GET_CODE (src) == CALL)
2171 hash_scan_call (src, insn);
2173 else if (GET_CODE (dest) == REG)
2175 unsigned int regno = REGNO (dest);
2178 /* If this is a single set and we are doing constant propagation,
2179 see if a REG_NOTE shows this equivalent to a constant. */
2180 if (set_p && (note = find_reg_equal_equiv_note (insn)) != 0
2181 && CONSTANT_P (XEXP (note, 0)))
2182 src = XEXP (note, 0), pat = gen_rtx_SET (VOIDmode, dest, src);
2184 /* Only record sets of pseudo-regs in the hash table. */
2186 && regno >= FIRST_PSEUDO_REGISTER
2187 /* Don't GCSE something if we can't do a reg/reg copy. */
2188 && can_copy_p [GET_MODE (dest)]
2189 /* Is SET_SRC something we want to gcse? */
2190 && want_to_gcse_p (src)
2191 /* Don't CSE a nop. */
2192 && ! set_noop_p (pat)
2193 /* Don't GCSE if it has attached REG_EQUIV note.
2194 At this point this only function parameters should have
2195 REG_EQUIV notes and if the argument slot is used somewhere
2196 explicitely, it means address of parameter has been taken,
2197 so we should not extend the lifetime of the pseudo. */
2198 && ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
2199 || GET_CODE (XEXP (note, 0)) != MEM))
2201 /* An expression is not anticipatable if its operands are
2202 modified before this insn or if this is not the only SET in
2204 int antic_p = oprs_anticipatable_p (src, insn) && single_set (insn);
2205 /* An expression is not available if its operands are
2206 subsequently modified, including this insn. It's also not
2207 available if this is a branch, because we can't insert
2208 a set after the branch. */
2209 int avail_p = (oprs_available_p (src, insn)
2210 && ! JUMP_P (insn));
2212 insert_expr_in_table (src, GET_MODE (dest), insn, antic_p, avail_p);
2215 /* Record sets for constant/copy propagation. */
2217 && regno >= FIRST_PSEUDO_REGISTER
2218 && ((GET_CODE (src) == REG
2219 && REGNO (src) >= FIRST_PSEUDO_REGISTER
2220 && can_copy_p [GET_MODE (dest)]
2221 && REGNO (src) != regno)
2222 || GET_CODE (src) == CONST_INT
2223 || GET_CODE (src) == SYMBOL_REF
2224 || GET_CODE (src) == CONST_DOUBLE)
2225 /* A copy is not available if its src or dest is subsequently
2226 modified. Here we want to search from INSN+1 on, but
2227 oprs_available_p searches from INSN on. */
2228 && (insn == BLOCK_END (BLOCK_NUM (insn))
2229 || ((tmp = next_nonnote_insn (insn)) != NULL_RTX
2230 && oprs_available_p (pat, tmp))))
2231 insert_set_in_table (pat, insn);
2236 hash_scan_clobber (x, insn)
2237 rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
2239 /* Currently nothing to do. */
2243 hash_scan_call (x, insn)
2244 rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
2246 /* Currently nothing to do. */
2249 /* Process INSN and add hash table entries as appropriate.
2251 Only available expressions that set a single pseudo-reg are recorded.
2253 Single sets in a PARALLEL could be handled, but it's an extra complication
2254 that isn't dealt with right now. The trick is handling the CLOBBERs that
2255 are also in the PARALLEL. Later.
2257 If SET_P is non-zero, this is for the assignment hash table,
2258 otherwise it is for the expression hash table.
2259 If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
2260 not record any expressions. */
2263 hash_scan_insn (insn, set_p, in_libcall_block)
2266 int in_libcall_block;
2268 rtx pat = PATTERN (insn);
2271 if (in_libcall_block)
2274 /* Pick out the sets of INSN and for other forms of instructions record
2275 what's been modified. */
2277 if (GET_CODE (pat) == SET)
2278 hash_scan_set (pat, insn, set_p);
2279 else if (GET_CODE (pat) == PARALLEL)
2280 for (i = 0; i < XVECLEN (pat, 0); i++)
2282 rtx x = XVECEXP (pat, 0, i);
2284 if (GET_CODE (x) == SET)
2285 hash_scan_set (x, insn, set_p);
2286 else if (GET_CODE (x) == CLOBBER)
2287 hash_scan_clobber (x, insn);
2288 else if (GET_CODE (x) == CALL)
2289 hash_scan_call (x, insn);
2292 else if (GET_CODE (pat) == CLOBBER)
2293 hash_scan_clobber (pat, insn);
2294 else if (GET_CODE (pat) == CALL)
2295 hash_scan_call (pat, insn);
2299 dump_hash_table (file, name, table, table_size, total_size)
2302 struct expr **table;
2303 int table_size, total_size;
2306 /* Flattened out table, so it's printed in proper order. */
2307 struct expr **flat_table;
2308 unsigned int *hash_val;
2312 = (struct expr **) xcalloc (total_size, sizeof (struct expr *));
2313 hash_val = (unsigned int *) xmalloc (total_size * sizeof (unsigned int));
2315 for (i = 0; i < table_size; i++)
2316 for (expr = table[i]; expr != NULL; expr = expr->next_same_hash)
2318 flat_table[expr->bitmap_index] = expr;
2319 hash_val[expr->bitmap_index] = i;
2322 fprintf (file, "%s hash table (%d buckets, %d entries)\n",
2323 name, table_size, total_size);
2325 for (i = 0; i < total_size; i++)
2326 if (flat_table[i] != 0)
2328 expr = flat_table[i];
2329 fprintf (file, "Index %d (hash value %d)\n ",
2330 expr->bitmap_index, hash_val[i]);
2331 print_rtl (file, expr->expr);
2332 fprintf (file, "\n");
2335 fprintf (file, "\n");
2341 /* Record register first/last/block set information for REGNO in INSN.
2343 first_set records the first place in the block where the register
2344 is set and is used to compute "anticipatability".
2346 last_set records the last place in the block where the register
2347 is set and is used to compute "availability".
2349 last_bb records the block for which first_set and last_set are
2350 valid, as a quick test to invalidate them.
2352 reg_set_in_block records whether the register is set in the block
2353 and is used to compute "transparency". */
2356 record_last_reg_set_info (insn, regno)
2360 struct reg_avail_info *info = ®_avail_info[regno];
2361 int cuid = INSN_CUID (insn);
2363 info->last_set = cuid;
2364 if (info->last_bb != current_bb)
2366 info->last_bb = current_bb;
2367 info->first_set = cuid;
2368 SET_BIT (reg_set_in_block[current_bb], regno);
2373 /* Record all of the canonicalized MEMs of record_last_mem_set_info's insn.
2374 Note we store a pair of elements in the list, so they have to be
2375 taken off pairwise. */
2378 canon_list_insert (dest, unused1, v_insn)
2379 rtx dest ATTRIBUTE_UNUSED;
2380 rtx unused1 ATTRIBUTE_UNUSED;
2383 rtx dest_addr, insn;
2385 while (GET_CODE (dest) == SUBREG
2386 || GET_CODE (dest) == ZERO_EXTRACT
2387 || GET_CODE (dest) == SIGN_EXTRACT
2388 || GET_CODE (dest) == STRICT_LOW_PART)
2389 dest = XEXP (dest, 0);
2391 /* If DEST is not a MEM, then it will not conflict with a load. Note
2392 that function calls are assumed to clobber memory, but are handled
2395 if (GET_CODE (dest) != MEM)
2398 dest_addr = get_addr (XEXP (dest, 0));
2399 dest_addr = canon_rtx (dest_addr);
2400 insn = (rtx) v_insn;
2402 canon_modify_mem_list[BLOCK_NUM (insn)] =
2403 alloc_INSN_LIST (dest_addr, canon_modify_mem_list[BLOCK_NUM (insn)]);
2404 canon_modify_mem_list[BLOCK_NUM (insn)] =
2405 alloc_INSN_LIST (dest, canon_modify_mem_list[BLOCK_NUM (insn)]);
2408 /* Record memory modification information for INSN. We do not actually care
2409 about the memory location(s) that are set, or even how they are set (consider
2410 a CALL_INSN). We merely need to record which insns modify memory. */
2413 record_last_mem_set_info (insn)
2416 /* load_killed_in_block_p will handle the case of calls clobbering
2418 modify_mem_list[BLOCK_NUM (insn)] =
2419 alloc_INSN_LIST (insn, modify_mem_list[BLOCK_NUM (insn)]);
2421 if (GET_CODE (insn) == CALL_INSN)
2423 /* Note that traversals of this loop (other than for free-ing)
2424 will break after encountering a CALL_INSN. So, there's no
2425 need to insert a pair of items, as canon_list_insert does. */
2426 canon_modify_mem_list[BLOCK_NUM (insn)] =
2427 alloc_INSN_LIST (insn, canon_modify_mem_list[BLOCK_NUM (insn)]);
2430 note_stores (PATTERN (insn), canon_list_insert, (void*)insn );
2433 /* Called from compute_hash_table via note_stores to handle one
2434 SET or CLOBBER in an insn. DATA is really the instruction in which
2435 the SET is taking place. */
2438 record_last_set_info (dest, setter, data)
2439 rtx dest, setter ATTRIBUTE_UNUSED;
2442 rtx last_set_insn = (rtx) data;
2444 if (GET_CODE (dest) == SUBREG)
2445 dest = SUBREG_REG (dest);
2447 if (GET_CODE (dest) == REG)
2448 record_last_reg_set_info (last_set_insn, REGNO (dest));
2449 else if (GET_CODE (dest) == MEM
2450 /* Ignore pushes, they clobber nothing. */
2451 && ! push_operand (dest, GET_MODE (dest)))
2452 record_last_mem_set_info (last_set_insn);
2455 /* Top level function to create an expression or assignment hash table.
2457 Expression entries are placed in the hash table if
2458 - they are of the form (set (pseudo-reg) src),
2459 - src is something we want to perform GCSE on,
2460 - none of the operands are subsequently modified in the block
2462 Assignment entries are placed in the hash table if
2463 - they are of the form (set (pseudo-reg) src),
2464 - src is something we want to perform const/copy propagation on,
2465 - none of the operands or target are subsequently modified in the block
2467 Currently src must be a pseudo-reg or a const_int.
2469 F is the first insn.
2470 SET_P is non-zero for computing the assignment hash table. */
2473 compute_hash_table (set_p)
2478 /* While we compute the hash table we also compute a bit array of which
2479 registers are set in which blocks.
2480 ??? This isn't needed during const/copy propagation, but it's cheap to
2482 sbitmap_vector_zero (reg_set_in_block, n_basic_blocks);
2484 /* re-Cache any INSN_LIST nodes we have allocated. */
2487 for (i = 0; i < n_basic_blocks; i++)
2489 if (modify_mem_list[i])
2490 free_INSN_LIST_list (modify_mem_list + i);
2491 if (canon_modify_mem_list[i])
2492 free_INSN_LIST_list (canon_modify_mem_list + i);
2495 /* Some working arrays used to track first and last set in each block. */
2496 reg_avail_info = (struct reg_avail_info*)
2497 gmalloc (max_gcse_regno * sizeof (struct reg_avail_info));
2499 for (i = 0; i < max_gcse_regno; ++i)
2500 reg_avail_info[i].last_bb = NEVER_SET;
2502 for (current_bb = 0; current_bb < n_basic_blocks; current_bb++)
2506 int in_libcall_block;
2508 /* First pass over the instructions records information used to
2509 determine when registers and memory are first and last set.
2510 ??? hard-reg reg_set_in_block computation
2511 could be moved to compute_sets since they currently don't change. */
2513 for (insn = BLOCK_HEAD (current_bb);
2514 insn && insn != NEXT_INSN (BLOCK_END (current_bb));
2515 insn = NEXT_INSN (insn))
2517 if (! INSN_P (insn))
2520 if (GET_CODE (insn) == CALL_INSN)
2522 bool clobbers_all = false;
2523 #ifdef NON_SAVING_SETJMP
2524 if (NON_SAVING_SETJMP
2525 && find_reg_note (insn, REG_SETJMP, NULL_RTX))
2526 clobbers_all = true;
2529 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2531 || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2532 record_last_reg_set_info (insn, regno);
2537 note_stores (PATTERN (insn), record_last_set_info, insn);
2540 /* The next pass builds the hash table. */
2542 for (insn = BLOCK_HEAD (current_bb), in_libcall_block = 0;
2543 insn && insn != NEXT_INSN (BLOCK_END (current_bb));
2544 insn = NEXT_INSN (insn))
2547 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
2548 in_libcall_block = 1;
2549 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
2550 in_libcall_block = 0;
2551 hash_scan_insn (insn, set_p, in_libcall_block);
2555 free (reg_avail_info);
2556 reg_avail_info = NULL;
2559 /* Allocate space for the set hash table.
2560 N_INSNS is the number of instructions in the function.
2561 It is used to determine the number of buckets to use. */
2564 alloc_set_hash_table (n_insns)
2569 set_hash_table_size = n_insns / 4;
2570 if (set_hash_table_size < 11)
2571 set_hash_table_size = 11;
2573 /* Attempt to maintain efficient use of hash table.
2574 Making it an odd number is simplest for now.
2575 ??? Later take some measurements. */
2576 set_hash_table_size |= 1;
2577 n = set_hash_table_size * sizeof (struct expr *);
2578 set_hash_table = (struct expr **) gmalloc (n);
2581 /* Free things allocated by alloc_set_hash_table. */
2584 free_set_hash_table ()
2586 free (set_hash_table);
2589 /* Compute the hash table for doing copy/const propagation. */
2592 compute_set_hash_table ()
2594 /* Initialize count of number of entries in hash table. */
2596 memset ((char *) set_hash_table, 0,
2597 set_hash_table_size * sizeof (struct expr *));
2599 compute_hash_table (1);
2602 /* Allocate space for the expression hash table.
2603 N_INSNS is the number of instructions in the function.
2604 It is used to determine the number of buckets to use. */
2607 alloc_expr_hash_table (n_insns)
2608 unsigned int n_insns;
2612 expr_hash_table_size = n_insns / 2;
2613 /* Make sure the amount is usable. */
2614 if (expr_hash_table_size < 11)
2615 expr_hash_table_size = 11;
2617 /* Attempt to maintain efficient use of hash table.
2618 Making it an odd number is simplest for now.
2619 ??? Later take some measurements. */
2620 expr_hash_table_size |= 1;
2621 n = expr_hash_table_size * sizeof (struct expr *);
2622 expr_hash_table = (struct expr **) gmalloc (n);
2625 /* Free things allocated by alloc_expr_hash_table. */
2628 free_expr_hash_table ()
2630 free (expr_hash_table);
2633 /* Compute the hash table for doing GCSE. */
2636 compute_expr_hash_table ()
2638 /* Initialize count of number of entries in hash table. */
2640 memset ((char *) expr_hash_table, 0,
2641 expr_hash_table_size * sizeof (struct expr *));
2643 compute_hash_table (0);
2646 /* Expression tracking support. */
2648 /* Lookup pattern PAT in the expression table.
2649 The result is a pointer to the table entry, or NULL if not found. */
2651 static struct expr *
2655 int do_not_record_p;
2656 unsigned int hash = hash_expr (pat, GET_MODE (pat), &do_not_record_p,
2657 expr_hash_table_size);
2660 if (do_not_record_p)
2663 expr = expr_hash_table[hash];
2665 while (expr && ! expr_equiv_p (expr->expr, pat))
2666 expr = expr->next_same_hash;
2671 /* Lookup REGNO in the set table. If PAT is non-NULL look for the entry that
2672 matches it, otherwise return the first entry for REGNO. The result is a
2673 pointer to the table entry, or NULL if not found. */
2675 static struct expr *
2676 lookup_set (regno, pat)
2680 unsigned int hash = hash_set (regno, set_hash_table_size);
2683 expr = set_hash_table[hash];
2687 while (expr && ! expr_equiv_p (expr->expr, pat))
2688 expr = expr->next_same_hash;
2692 while (expr && REGNO (SET_DEST (expr->expr)) != regno)
2693 expr = expr->next_same_hash;
2699 /* Return the next entry for REGNO in list EXPR. */
2701 static struct expr *
2702 next_set (regno, expr)
2707 expr = expr->next_same_hash;
2708 while (expr && REGNO (SET_DEST (expr->expr)) != regno);
2713 /* Reset tables used to keep track of what's still available [since the
2714 start of the block]. */
2717 reset_opr_set_tables ()
2719 /* Maintain a bitmap of which regs have been set since beginning of
2721 sbitmap_zero (reg_set_bitmap);
2723 /* Also keep a record of the last instruction to modify memory.
2724 For now this is very trivial, we only record whether any memory
2725 location has been modified. */
2729 /* re-Cache any INSN_LIST nodes we have allocated. */
2730 for (i = 0; i < n_basic_blocks; i++)
2732 if (modify_mem_list[i])
2733 free_INSN_LIST_list (modify_mem_list + i);
2734 if (canon_modify_mem_list[i])
2735 free_INSN_LIST_list (canon_modify_mem_list + i);
2740 /* Return non-zero if the operands of X are not set before INSN in
2741 INSN's basic block. */
2744 oprs_not_set_p (x, insn)
2754 code = GET_CODE (x);
2769 if (load_killed_in_block_p (BLOCK_FOR_INSN (insn),
2770 INSN_CUID (insn), x, 0))
2773 return oprs_not_set_p (XEXP (x, 0), insn);
2776 return ! TEST_BIT (reg_set_bitmap, REGNO (x));
2782 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2786 /* If we are about to do the last recursive call
2787 needed at this level, change it into iteration.
2788 This function is called enough to be worth it. */
2790 return oprs_not_set_p (XEXP (x, i), insn);
2792 if (! oprs_not_set_p (XEXP (x, i), insn))
2795 else if (fmt[i] == 'E')
2796 for (j = 0; j < XVECLEN (x, i); j++)
2797 if (! oprs_not_set_p (XVECEXP (x, i, j), insn))
2804 /* Mark things set by a CALL. */
2810 if (! CONST_OR_PURE_CALL_P (insn))
2811 record_last_mem_set_info (insn);
2814 /* Mark things set by a SET. */
2817 mark_set (pat, insn)
2820 rtx dest = SET_DEST (pat);
2822 while (GET_CODE (dest) == SUBREG
2823 || GET_CODE (dest) == ZERO_EXTRACT
2824 || GET_CODE (dest) == SIGN_EXTRACT
2825 || GET_CODE (dest) == STRICT_LOW_PART)
2826 dest = XEXP (dest, 0);
2828 if (GET_CODE (dest) == REG)
2829 SET_BIT (reg_set_bitmap, REGNO (dest));
2830 else if (GET_CODE (dest) == MEM)
2831 record_last_mem_set_info (insn);
2833 if (GET_CODE (SET_SRC (pat)) == CALL)
2837 /* Record things set by a CLOBBER. */
2840 mark_clobber (pat, insn)
2843 rtx clob = XEXP (pat, 0);
2845 while (GET_CODE (clob) == SUBREG || GET_CODE (clob) == STRICT_LOW_PART)
2846 clob = XEXP (clob, 0);
2848 if (GET_CODE (clob) == REG)
2849 SET_BIT (reg_set_bitmap, REGNO (clob));
2851 record_last_mem_set_info (insn);
2854 /* Record things set by INSN.
2855 This data is used by oprs_not_set_p. */
2858 mark_oprs_set (insn)
2861 rtx pat = PATTERN (insn);
2864 if (GET_CODE (pat) == SET)
2865 mark_set (pat, insn);
2866 else if (GET_CODE (pat) == PARALLEL)
2867 for (i = 0; i < XVECLEN (pat, 0); i++)
2869 rtx x = XVECEXP (pat, 0, i);
2871 if (GET_CODE (x) == SET)
2873 else if (GET_CODE (x) == CLOBBER)
2874 mark_clobber (x, insn);
2875 else if (GET_CODE (x) == CALL)
2879 else if (GET_CODE (pat) == CLOBBER)
2880 mark_clobber (pat, insn);
2881 else if (GET_CODE (pat) == CALL)
2886 /* Classic GCSE reaching definition support. */
2888 /* Allocate reaching def variables. */
2891 alloc_rd_mem (n_blocks, n_insns)
2892 int n_blocks, n_insns;
2894 rd_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2895 sbitmap_vector_zero (rd_kill, n_basic_blocks);
2897 rd_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2898 sbitmap_vector_zero (rd_gen, n_basic_blocks);
2900 reaching_defs = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2901 sbitmap_vector_zero (reaching_defs, n_basic_blocks);
2903 rd_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2904 sbitmap_vector_zero (rd_out, n_basic_blocks);
2907 /* Free reaching def variables. */
2912 sbitmap_vector_free (rd_kill);
2913 sbitmap_vector_free (rd_gen);
2914 sbitmap_vector_free (reaching_defs);
2915 sbitmap_vector_free (rd_out);
2918 /* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
2921 handle_rd_kill_set (insn, regno, bb)
2926 struct reg_set *this_reg;
2928 for (this_reg = reg_set_table[regno]; this_reg; this_reg = this_reg ->next)
2929 if (BLOCK_NUM (this_reg->insn) != BLOCK_NUM (insn))
2930 SET_BIT (rd_kill[bb->index], INSN_CUID (this_reg->insn));
2933 /* Compute the set of kill's for reaching definitions. */
2943 For each set bit in `gen' of the block (i.e each insn which
2944 generates a definition in the block)
2945 Call the reg set by the insn corresponding to that bit regx
2946 Look at the linked list starting at reg_set_table[regx]
2947 For each setting of regx in the linked list, which is not in
2949 Set the bit in `kill' corresponding to that insn. */
2950 for (bb = 0; bb < n_basic_blocks; bb++)
2951 for (cuid = 0; cuid < max_cuid; cuid++)
2952 if (TEST_BIT (rd_gen[bb], cuid))
2954 rtx insn = CUID_INSN (cuid);
2955 rtx pat = PATTERN (insn);
2957 if (GET_CODE (insn) == CALL_INSN)
2959 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2960 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2961 handle_rd_kill_set (insn, regno, BASIC_BLOCK (bb));
2964 if (GET_CODE (pat) == PARALLEL)
2966 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
2968 enum rtx_code code = GET_CODE (XVECEXP (pat, 0, i));
2970 if ((code == SET || code == CLOBBER)
2971 && GET_CODE (XEXP (XVECEXP (pat, 0, i), 0)) == REG)
2972 handle_rd_kill_set (insn,
2973 REGNO (XEXP (XVECEXP (pat, 0, i), 0)),
2977 else if (GET_CODE (pat) == SET && GET_CODE (SET_DEST (pat)) == REG)
2978 /* Each setting of this register outside of this block
2979 must be marked in the set of kills in this block. */
2980 handle_rd_kill_set (insn, REGNO (SET_DEST (pat)), BASIC_BLOCK (bb));
2984 /* Compute the reaching definitions as in
2985 Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
2986 Chapter 10. It is the same algorithm as used for computing available
2987 expressions but applied to the gens and kills of reaching definitions. */
2992 int bb, changed, passes;
2994 for (bb = 0; bb < n_basic_blocks; bb++)
2995 sbitmap_copy (rd_out[bb] /*dst*/, rd_gen[bb] /*src*/);
3002 for (bb = 0; bb < n_basic_blocks; bb++)
3004 sbitmap_union_of_preds (reaching_defs[bb], rd_out, bb);
3005 changed |= sbitmap_union_of_diff (rd_out[bb], rd_gen[bb],
3006 reaching_defs[bb], rd_kill[bb]);
3012 fprintf (gcse_file, "reaching def computation: %d passes\n", passes);
3015 /* Classic GCSE available expression support. */
3017 /* Allocate memory for available expression computation. */
3020 alloc_avail_expr_mem (n_blocks, n_exprs)
3021 int n_blocks, n_exprs;
3023 ae_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
3024 sbitmap_vector_zero (ae_kill, n_basic_blocks);
3026 ae_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
3027 sbitmap_vector_zero (ae_gen, n_basic_blocks);
3029 ae_in = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
3030 sbitmap_vector_zero (ae_in, n_basic_blocks);
3032 ae_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
3033 sbitmap_vector_zero (ae_out, n_basic_blocks);
3037 free_avail_expr_mem ()
3039 sbitmap_vector_free (ae_kill);
3040 sbitmap_vector_free (ae_gen);
3041 sbitmap_vector_free (ae_in);
3042 sbitmap_vector_free (ae_out);
3045 /* Compute the set of available expressions generated in each basic block. */
3054 /* For each recorded occurrence of each expression, set ae_gen[bb][expr].
3055 This is all we have to do because an expression is not recorded if it
3056 is not available, and the only expressions we want to work with are the
3057 ones that are recorded. */
3058 for (i = 0; i < expr_hash_table_size; i++)
3059 for (expr = expr_hash_table[i]; expr != 0; expr = expr->next_same_hash)
3060 for (occr = expr->avail_occr; occr != 0; occr = occr->next)
3061 SET_BIT (ae_gen[BLOCK_NUM (occr->insn)], expr->bitmap_index);
3064 /* Return non-zero if expression X is killed in BB. */
3067 expr_killed_p (x, bb)
3078 code = GET_CODE (x);
3082 return TEST_BIT (reg_set_in_block[bb->index], REGNO (x));
3085 if (load_killed_in_block_p (bb, get_max_uid () + 1, x, 0))
3088 return expr_killed_p (XEXP (x, 0), bb);
3105 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3109 /* If we are about to do the last recursive call
3110 needed at this level, change it into iteration.
3111 This function is called enough to be worth it. */
3113 return expr_killed_p (XEXP (x, i), bb);
3114 else if (expr_killed_p (XEXP (x, i), bb))
3117 else if (fmt[i] == 'E')
3118 for (j = 0; j < XVECLEN (x, i); j++)
3119 if (expr_killed_p (XVECEXP (x, i, j), bb))
3126 /* Compute the set of available expressions killed in each basic block. */
3129 compute_ae_kill (ae_gen, ae_kill)
3130 sbitmap *ae_gen, *ae_kill;
3136 for (bb = 0; bb < n_basic_blocks; bb++)
3137 for (i = 0; i < expr_hash_table_size; i++)
3138 for (expr = expr_hash_table[i]; expr; expr = expr->next_same_hash)
3140 /* Skip EXPR if generated in this block. */
3141 if (TEST_BIT (ae_gen[bb], expr->bitmap_index))
3144 if (expr_killed_p (expr->expr, BASIC_BLOCK (bb)))
3145 SET_BIT (ae_kill[bb], expr->bitmap_index);
3149 /* Actually perform the Classic GCSE optimizations. */
3151 /* Return non-zero if occurrence OCCR of expression EXPR reaches block BB.
3153 CHECK_SELF_LOOP is non-zero if we should consider a block reaching itself
3154 as a positive reach. We want to do this when there are two computations
3155 of the expression in the block.
3157 VISITED is a pointer to a working buffer for tracking which BB's have
3158 been visited. It is NULL for the top-level call.
3160 We treat reaching expressions that go through blocks containing the same
3161 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
3162 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
3163 2 as not reaching. The intent is to improve the probability of finding
3164 only one reaching expression and to reduce register lifetimes by picking
3165 the closest such expression. */
3168 expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited)
3172 int check_self_loop;
3177 for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
3179 basic_block pred_bb = pred->src;
3181 if (visited[pred_bb->index])
3182 /* This predecessor has already been visited. Nothing to do. */
3184 else if (pred_bb == bb)
3186 /* BB loops on itself. */
3188 && TEST_BIT (ae_gen[pred_bb->index], expr->bitmap_index)
3189 && BLOCK_NUM (occr->insn) == pred_bb->index)
3192 visited[pred_bb->index] = 1;
3195 /* Ignore this predecessor if it kills the expression. */
3196 else if (TEST_BIT (ae_kill[pred_bb->index], expr->bitmap_index))
3197 visited[pred_bb->index] = 1;
3199 /* Does this predecessor generate this expression? */
3200 else if (TEST_BIT (ae_gen[pred_bb->index], expr->bitmap_index))
3202 /* Is this the occurrence we're looking for?
3203 Note that there's only one generating occurrence per block
3204 so we just need to check the block number. */
3205 if (BLOCK_NUM (occr->insn) == pred_bb->index)
3208 visited[pred_bb->index] = 1;
3211 /* Neither gen nor kill. */
3214 visited[pred_bb->index] = 1;
3215 if (expr_reaches_here_p_work (occr, expr, pred_bb, check_self_loop,
3222 /* All paths have been checked. */
3226 /* This wrapper for expr_reaches_here_p_work() is to ensure that any
3227 memory allocated for that function is returned. */
3230 expr_reaches_here_p (occr, expr, bb, check_self_loop)
3234 int check_self_loop;
3237 char *visited = (char *) xcalloc (n_basic_blocks, 1);
3239 rval = expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited);
3245 /* Return the instruction that computes EXPR that reaches INSN's basic block.
3246 If there is more than one such instruction, return NULL.
3248 Called only by handle_avail_expr. */
3251 computing_insn (expr, insn)
3255 basic_block bb = BLOCK_FOR_INSN (insn);
3257 if (expr->avail_occr->next == NULL)
3259 if (BLOCK_FOR_INSN (expr->avail_occr->insn) == bb)
3260 /* The available expression is actually itself
3261 (i.e. a loop in the flow graph) so do nothing. */
3264 /* (FIXME) Case that we found a pattern that was created by
3265 a substitution that took place. */
3266 return expr->avail_occr->insn;
3270 /* Pattern is computed more than once.
3271 Search backwards from this insn to see how many of these
3272 computations actually reach this insn. */
3274 rtx insn_computes_expr = NULL;
3277 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
3279 if (BLOCK_FOR_INSN (occr->insn) == bb)
3281 /* The expression is generated in this block.
3282 The only time we care about this is when the expression
3283 is generated later in the block [and thus there's a loop].
3284 We let the normal cse pass handle the other cases. */
3285 if (INSN_CUID (insn) < INSN_CUID (occr->insn)
3286 && expr_reaches_here_p (occr, expr, bb, 1))
3292 insn_computes_expr = occr->insn;
3295 else if (expr_reaches_here_p (occr, expr, bb, 0))
3301 insn_computes_expr = occr->insn;
3305 if (insn_computes_expr == NULL)
3308 return insn_computes_expr;
3312 /* Return non-zero if the definition in DEF_INSN can reach INSN.
3313 Only called by can_disregard_other_sets. */
3316 def_reaches_here_p (insn, def_insn)
3321 if (TEST_BIT (reaching_defs[BLOCK_NUM (insn)], INSN_CUID (def_insn)))
3324 if (BLOCK_NUM (insn) == BLOCK_NUM (def_insn))
3326 if (INSN_CUID (def_insn) < INSN_CUID (insn))
3328 if (GET_CODE (PATTERN (def_insn)) == PARALLEL)
3330 else if (GET_CODE (PATTERN (def_insn)) == CLOBBER)
3331 reg = XEXP (PATTERN (def_insn), 0);
3332 else if (GET_CODE (PATTERN (def_insn)) == SET)
3333 reg = SET_DEST (PATTERN (def_insn));
3337 return ! reg_set_between_p (reg, NEXT_INSN (def_insn), insn);
3346 /* Return non-zero if *ADDR_THIS_REG can only have one value at INSN. The
3347 value returned is the number of definitions that reach INSN. Returning a
3348 value of zero means that [maybe] more than one definition reaches INSN and
3349 the caller can't perform whatever optimization it is trying. i.e. it is
3350 always safe to return zero. */
3353 can_disregard_other_sets (addr_this_reg, insn, for_combine)
3354 struct reg_set **addr_this_reg;
3358 int number_of_reaching_defs = 0;
3359 struct reg_set *this_reg;
3361 for (this_reg = *addr_this_reg; this_reg != 0; this_reg = this_reg->next)
3362 if (def_reaches_here_p (insn, this_reg->insn))
3364 number_of_reaching_defs++;
3365 /* Ignore parallels for now. */
3366 if (GET_CODE (PATTERN (this_reg->insn)) == PARALLEL)
3370 && (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER
3371 || ! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
3372 SET_SRC (PATTERN (insn)))))
3373 /* A setting of the reg to a different value reaches INSN. */
3376 if (number_of_reaching_defs > 1)
3378 /* If in this setting the value the register is being set to is
3379 equal to the previous value the register was set to and this
3380 setting reaches the insn we are trying to do the substitution
3381 on then we are ok. */
3382 if (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER)
3384 else if (! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
3385 SET_SRC (PATTERN (insn))))
3389 *addr_this_reg = this_reg;
3392 return number_of_reaching_defs;
3395 /* Expression computed by insn is available and the substitution is legal,
3396 so try to perform the substitution.
3398 The result is non-zero if any changes were made. */
3401 handle_avail_expr (insn, expr)
3405 rtx pat, insn_computes_expr, expr_set;
3407 struct reg_set *this_reg;
3408 int found_setting, use_src;
3411 /* We only handle the case where one computation of the expression
3412 reaches this instruction. */
3413 insn_computes_expr = computing_insn (expr, insn);
3414 if (insn_computes_expr == NULL)
3416 expr_set = single_set (insn_computes_expr);
3423 /* At this point we know only one computation of EXPR outside of this
3424 block reaches this insn. Now try to find a register that the
3425 expression is computed into. */
3426 if (GET_CODE (SET_SRC (expr_set)) == REG)
3428 /* This is the case when the available expression that reaches
3429 here has already been handled as an available expression. */
3430 unsigned int regnum_for_replacing
3431 = REGNO (SET_SRC (expr_set));
3433 /* If the register was created by GCSE we can't use `reg_set_table',
3434 however we know it's set only once. */
3435 if (regnum_for_replacing >= max_gcse_regno
3436 /* If the register the expression is computed into is set only once,
3437 or only one set reaches this insn, we can use it. */
3438 || (((this_reg = reg_set_table[regnum_for_replacing]),
3439 this_reg->next == NULL)
3440 || can_disregard_other_sets (&this_reg, insn, 0)))
3449 unsigned int regnum_for_replacing
3450 = REGNO (SET_DEST (expr_set));
3452 /* This shouldn't happen. */
3453 if (regnum_for_replacing >= max_gcse_regno)
3456 this_reg = reg_set_table[regnum_for_replacing];
3458 /* If the register the expression is computed into is set only once,
3459 or only one set reaches this insn, use it. */
3460 if (this_reg->next == NULL
3461 || can_disregard_other_sets (&this_reg, insn, 0))
3467 pat = PATTERN (insn);
3469 to = SET_SRC (expr_set);
3471 to = SET_DEST (expr_set);
3472 changed = validate_change (insn, &SET_SRC (pat), to, 0);
3474 /* We should be able to ignore the return code from validate_change but
3475 to play it safe we check. */
3479 if (gcse_file != NULL)
3481 fprintf (gcse_file, "GCSE: Replacing the source in insn %d with",
3483 fprintf (gcse_file, " reg %d %s insn %d\n",
3484 REGNO (to), use_src ? "from" : "set in",
3485 INSN_UID (insn_computes_expr));
3490 /* The register that the expr is computed into is set more than once. */
3491 else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
3493 /* Insert an insn after insnx that copies the reg set in insnx
3494 into a new pseudo register call this new register REGN.
3495 From insnb until end of basic block or until REGB is set
3496 replace all uses of REGB with REGN. */
3499 to = gen_reg_rtx (GET_MODE (SET_DEST (expr_set)));
3501 /* Generate the new insn. */
3502 /* ??? If the change fails, we return 0, even though we created
3503 an insn. I think this is ok. */
3505 = emit_insn_after (gen_rtx_SET (VOIDmode, to,
3506 SET_DEST (expr_set)),
3507 insn_computes_expr);
3509 /* Keep block number table up to date. */
3510 set_block_for_new_insns (new_insn, BLOCK_FOR_INSN (insn_computes_expr));
3512 /* Keep register set table up to date. */
3513 record_one_set (REGNO (to), new_insn);
3515 gcse_create_count++;
3516 if (gcse_file != NULL)
3518 fprintf (gcse_file, "GCSE: Creating insn %d to copy value of reg %d",
3519 INSN_UID (NEXT_INSN (insn_computes_expr)),
3520 REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr)))));
3521 fprintf (gcse_file, ", computed in insn %d,\n",
3522 INSN_UID (insn_computes_expr));
3523 fprintf (gcse_file, " into newly allocated reg %d\n",
3527 pat = PATTERN (insn);
3529 /* Do register replacement for INSN. */
3530 changed = validate_change (insn, &SET_SRC (pat),
3532 (NEXT_INSN (insn_computes_expr))),
3535 /* We should be able to ignore the return code from validate_change but
3536 to play it safe we check. */
3540 if (gcse_file != NULL)
3543 "GCSE: Replacing the source in insn %d with reg %d ",
3545 REGNO (SET_DEST (PATTERN (NEXT_INSN
3546 (insn_computes_expr)))));
3547 fprintf (gcse_file, "set in insn %d\n",
3548 INSN_UID (insn_computes_expr));
3556 /* Perform classic GCSE. This is called by one_classic_gcse_pass after all
3557 the dataflow analysis has been done.
3559 The result is non-zero if a change was made. */
3567 /* Note we start at block 1. */
3570 for (bb = 1; bb < n_basic_blocks; bb++)
3572 /* Reset tables used to keep track of what's still valid [since the
3573 start of the block]. */
3574 reset_opr_set_tables ();
3576 for (insn = BLOCK_HEAD (bb);
3577 insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
3578 insn = NEXT_INSN (insn))
3580 /* Is insn of form (set (pseudo-reg) ...)? */
3581 if (GET_CODE (insn) == INSN
3582 && GET_CODE (PATTERN (insn)) == SET
3583 && GET_CODE (SET_DEST (PATTERN (insn))) == REG
3584 && REGNO (SET_DEST (PATTERN (insn))) >= FIRST_PSEUDO_REGISTER)
3586 rtx pat = PATTERN (insn);
3587 rtx src = SET_SRC (pat);
3590 if (want_to_gcse_p (src)
3591 /* Is the expression recorded? */
3592 && ((expr = lookup_expr (src)) != NULL)
3593 /* Is the expression available [at the start of the
3595 && TEST_BIT (ae_in[bb], expr->bitmap_index)
3596 /* Are the operands unchanged since the start of the
3598 && oprs_not_set_p (src, insn))
3599 changed |= handle_avail_expr (insn, expr);
3602 /* Keep track of everything modified by this insn. */
3603 /* ??? Need to be careful w.r.t. mods done to INSN. */
3605 mark_oprs_set (insn);
3612 /* Top level routine to perform one classic GCSE pass.
3614 Return non-zero if a change was made. */
3617 one_classic_gcse_pass (pass)
3622 gcse_subst_count = 0;
3623 gcse_create_count = 0;
3625 alloc_expr_hash_table (max_cuid);
3626 alloc_rd_mem (n_basic_blocks, max_cuid);
3627 compute_expr_hash_table ();
3629 dump_hash_table (gcse_file, "Expression", expr_hash_table,
3630 expr_hash_table_size, n_exprs);
3636 alloc_avail_expr_mem (n_basic_blocks, n_exprs);
3638 compute_ae_kill (ae_gen, ae_kill);
3639 compute_available (ae_gen, ae_kill, ae_out, ae_in);
3640 changed = classic_gcse ();
3641 free_avail_expr_mem ();
3645 free_expr_hash_table ();
3649 fprintf (gcse_file, "\n");
3650 fprintf (gcse_file, "GCSE of %s, pass %d: %d bytes needed, %d substs,",
3651 current_function_name, pass, bytes_used, gcse_subst_count);
3652 fprintf (gcse_file, "%d insns created\n", gcse_create_count);
3658 /* Compute copy/constant propagation working variables. */
3660 /* Local properties of assignments. */
3661 static sbitmap *cprop_pavloc;
3662 static sbitmap *cprop_absaltered;
3664 /* Global properties of assignments (computed from the local properties). */
3665 static sbitmap *cprop_avin;
3666 static sbitmap *cprop_avout;
3668 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
3669 basic blocks. N_SETS is the number of sets. */
3672 alloc_cprop_mem (n_blocks, n_sets)
3673 int n_blocks, n_sets;
3675 cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
3676 cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
3678 cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
3679 cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
3682 /* Free vars used by copy/const propagation. */
3687 sbitmap_vector_free (cprop_pavloc);
3688 sbitmap_vector_free (cprop_absaltered);
3689 sbitmap_vector_free (cprop_avin);
3690 sbitmap_vector_free (cprop_avout);
3693 /* For each block, compute whether X is transparent. X is either an
3694 expression or an assignment [though we don't care which, for this context
3695 an assignment is treated as an expression]. For each block where an
3696 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
3700 compute_transp (x, indx, bmap, set_p)
3711 /* repeat is used to turn tail-recursion into iteration since GCC
3712 can't do it when there's no return value. */
3718 code = GET_CODE (x);
3724 if (REGNO (x) < FIRST_PSEUDO_REGISTER)
3726 for (bb = 0; bb < n_basic_blocks; bb++)
3727 if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
3728 SET_BIT (bmap[bb], indx);
3732 for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
3733 SET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
3738 if (REGNO (x) < FIRST_PSEUDO_REGISTER)
3740 for (bb = 0; bb < n_basic_blocks; bb++)
3741 if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
3742 RESET_BIT (bmap[bb], indx);
3746 for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
3747 RESET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
3754 for (bb = 0; bb < n_basic_blocks; bb++)
3756 rtx list_entry = canon_modify_mem_list[bb];
3760 rtx dest, dest_addr;
3762 if (GET_CODE (XEXP (list_entry, 0)) == CALL_INSN)
3765 SET_BIT (bmap[bb], indx);
3767 RESET_BIT (bmap[bb], indx);
3770 /* LIST_ENTRY must be an INSN of some kind that sets memory.
3771 Examine each hunk of memory that is modified. */
3773 dest = XEXP (list_entry, 0);
3774 list_entry = XEXP (list_entry, 1);
3775 dest_addr = XEXP (list_entry, 0);
3777 if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
3778 x, rtx_addr_varies_p))
3781 SET_BIT (bmap[bb], indx);
3783 RESET_BIT (bmap[bb], indx);
3786 list_entry = XEXP (list_entry, 1);
3808 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3812 /* If we are about to do the last recursive call
3813 needed at this level, change it into iteration.
3814 This function is called enough to be worth it. */
3821 compute_transp (XEXP (x, i), indx, bmap, set_p);
3823 else if (fmt[i] == 'E')
3824 for (j = 0; j < XVECLEN (x, i); j++)
3825 compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
3829 /* Top level routine to do the dataflow analysis needed by copy/const
3833 compute_cprop_data ()
3835 compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, 1);
3836 compute_available (cprop_pavloc, cprop_absaltered,
3837 cprop_avout, cprop_avin);
3840 /* Copy/constant propagation. */
3842 /* Maximum number of register uses in an insn that we handle. */
3845 /* Table of uses found in an insn.
3846 Allocated statically to avoid alloc/free complexity and overhead. */
3847 static struct reg_use reg_use_table[MAX_USES];
3849 /* Index into `reg_use_table' while building it. */
3850 static int reg_use_count;
3852 /* Set up a list of register numbers used in INSN. The found uses are stored
3853 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
3854 and contains the number of uses in the table upon exit.
3856 ??? If a register appears multiple times we will record it multiple times.
3857 This doesn't hurt anything but it will slow things down. */
3860 find_used_regs (xptr, data)
3862 void *data ATTRIBUTE_UNUSED;
3869 /* repeat is used to turn tail-recursion into iteration since GCC
3870 can't do it when there's no return value. */
3875 code = GET_CODE (x);
3878 if (reg_use_count == MAX_USES)
3881 reg_use_table[reg_use_count].reg_rtx = x;
3885 /* Recursively scan the operands of this expression. */
3887 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3891 /* If we are about to do the last recursive call
3892 needed at this level, change it into iteration.
3893 This function is called enough to be worth it. */
3900 find_used_regs (&XEXP (x, i), data);
3902 else if (fmt[i] == 'E')
3903 for (j = 0; j < XVECLEN (x, i); j++)
3904 find_used_regs (&XVECEXP (x, i, j), data);
3908 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
3909 Returns non-zero is successful. */
3912 try_replace_reg (from, to, insn)
3915 rtx note = find_reg_equal_equiv_note (insn);
3918 rtx set = single_set (insn);
3920 success = validate_replace_src (from, to, insn);
3922 /* If above failed and this is a single set, try to simplify the source of
3923 the set given our substitution. We could perhaps try this for multiple
3924 SETs, but it probably won't buy us anything. */
3925 if (!success && set != 0)
3927 src = simplify_replace_rtx (SET_SRC (set), from, to);
3929 if (!rtx_equal_p (src, SET_SRC (set))
3930 && validate_change (insn, &SET_SRC (set), src, 0))
3934 /* If we've failed to do replacement, have a single SET, and don't already
3935 have a note, add a REG_EQUAL note to not lose information. */
3936 if (!success && note == 0 && set != 0)
3937 note = REG_NOTES (insn)
3938 = gen_rtx_EXPR_LIST (REG_EQUAL, src, REG_NOTES (insn));
3940 /* If there is already a NOTE, update the expression in it with our
3943 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), from, to);
3945 /* REG_EQUAL may get simplified into register.
3946 We don't allow that. Remove that note. This code ought
3947 not to hapen, because previous code ought to syntetize
3948 reg-reg move, but be on the safe side. */
3949 if (note && REG_P (XEXP (note, 0)))
3950 remove_note (insn, note);
3955 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
3956 NULL no such set is found. */
3958 static struct expr *
3959 find_avail_set (regno, insn)
3963 /* SET1 contains the last set found that can be returned to the caller for
3964 use in a substitution. */
3965 struct expr *set1 = 0;
3967 /* Loops are not possible here. To get a loop we would need two sets
3968 available at the start of the block containing INSN. ie we would
3969 need two sets like this available at the start of the block:
3971 (set (reg X) (reg Y))
3972 (set (reg Y) (reg X))
3974 This can not happen since the set of (reg Y) would have killed the
3975 set of (reg X) making it unavailable at the start of this block. */
3979 struct expr *set = lookup_set (regno, NULL_RTX);
3981 /* Find a set that is available at the start of the block
3982 which contains INSN. */
3985 if (TEST_BIT (cprop_avin[BLOCK_NUM (insn)], set->bitmap_index))
3987 set = next_set (regno, set);
3990 /* If no available set was found we've reached the end of the
3991 (possibly empty) copy chain. */
3995 if (GET_CODE (set->expr) != SET)
3998 src = SET_SRC (set->expr);
4000 /* We know the set is available.
4001 Now check that SRC is ANTLOC (i.e. none of the source operands
4002 have changed since the start of the block).
4004 If the source operand changed, we may still use it for the next
4005 iteration of this loop, but we may not use it for substitutions. */
4007 if (CONSTANT_P (src) || oprs_not_set_p (src, insn))
4010 /* If the source of the set is anything except a register, then
4011 we have reached the end of the copy chain. */
4012 if (GET_CODE (src) != REG)
4015 /* Follow the copy chain, ie start another iteration of the loop
4016 and see if we have an available copy into SRC. */
4017 regno = REGNO (src);
4020 /* SET1 holds the last set that was available and anticipatable at
4025 /* Subroutine of cprop_insn that tries to propagate constants into
4026 JUMP_INSNS. INSN must be a conditional jump. FROM is what we will try to
4027 replace, SRC is the constant we will try to substitute for it. Returns
4028 nonzero if a change was made. We know INSN has just a SET. */
4031 cprop_jump (bb, insn, from, src)
4037 rtx set = PATTERN (insn);
4038 rtx new = simplify_replace_rtx (SET_SRC (set), from, src);
4040 /* If no simplification can be made, then try the next
4042 if (rtx_equal_p (new, SET_SRC (set)))
4045 /* If this is now a no-op leave it that way, but update LABEL_NUSED if
4049 SET_SRC (set) = new;
4051 if (JUMP_LABEL (insn) != 0)
4052 --LABEL_NUSES (JUMP_LABEL (insn));
4055 /* Otherwise, this must be a valid instruction. */
4056 else if (! validate_change (insn, &SET_SRC (set), new, 0))
4059 /* If this has turned into an unconditional jump,
4060 then put a barrier after it so that the unreachable
4061 code will be deleted. */
4062 if (GET_CODE (SET_SRC (set)) == LABEL_REF)
4063 emit_barrier_after (insn);
4065 run_jump_opt_after_gcse = 1;
4068 if (gcse_file != NULL)
4071 "CONST-PROP: Replacing reg %d in insn %d with constant ",
4072 REGNO (from), INSN_UID (insn));
4073 print_rtl (gcse_file, src);
4074 fprintf (gcse_file, "\n");
4076 purge_dead_edges (bb);
4083 /* Subroutine of cprop_insn that tries to propagate constants into JUMP_INSNS
4084 for machines that have CC0. INSN is a single set that stores into CC0;
4085 the insn following it is a conditional jump. REG_USED is the use we will
4086 try to replace, SRC is the constant we will try to substitute for it.
4087 Returns nonzero if a change was made. */
4090 cprop_cc0_jump (bb, insn, reg_used, src)
4093 struct reg_use *reg_used;
4096 /* First substitute in the SET_SRC of INSN, then substitute that for
4098 rtx jump = NEXT_INSN (insn);
4099 rtx new_src = simplify_replace_rtx (SET_SRC (PATTERN (insn)),
4100 reg_used->reg_rtx, src);
4102 if (! cprop_jump (bb, jump, cc0_rtx, new_src))
4105 /* If we succeeded, delete the cc0 setter. */
4106 PUT_CODE (insn, NOTE);
4107 NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
4108 NOTE_SOURCE_FILE (insn) = 0;
4114 /* Perform constant and copy propagation on INSN.
4115 The result is non-zero if a change was made. */
4118 cprop_insn (bb, insn, alter_jumps)
4123 struct reg_use *reg_used;
4131 note_uses (&PATTERN (insn), find_used_regs, NULL);
4133 note = find_reg_equal_equiv_note (insn);
4135 /* We may win even when propagating constants into notes. */
4137 find_used_regs (&XEXP (note, 0), NULL);
4139 for (reg_used = ®_use_table[0]; reg_use_count > 0;
4140 reg_used++, reg_use_count--)
4142 unsigned int regno = REGNO (reg_used->reg_rtx);
4146 /* Ignore registers created by GCSE.
4147 We do this because ... */
4148 if (regno >= max_gcse_regno)
4151 /* If the register has already been set in this block, there's
4152 nothing we can do. */
4153 if (! oprs_not_set_p (reg_used->reg_rtx, insn))
4156 /* Find an assignment that sets reg_used and is available
4157 at the start of the block. */
4158 set = find_avail_set (regno, insn);
4163 /* ??? We might be able to handle PARALLELs. Later. */
4164 if (GET_CODE (pat) != SET)
4167 src = SET_SRC (pat);
4169 /* Constant propagation. */
4170 if (GET_CODE (src) == CONST_INT || GET_CODE (src) == CONST_DOUBLE
4171 || GET_CODE (src) == SYMBOL_REF)
4173 /* Handle normal insns first. */
4174 if (GET_CODE (insn) == INSN
4175 && try_replace_reg (reg_used->reg_rtx, src, insn))
4179 if (gcse_file != NULL)
4181 fprintf (gcse_file, "CONST-PROP: Replacing reg %d in ",
4183 fprintf (gcse_file, "insn %d with constant ",
4185 print_rtl (gcse_file, src);
4186 fprintf (gcse_file, "\n");
4189 /* The original insn setting reg_used may or may not now be
4190 deletable. We leave the deletion to flow. */
4193 /* Try to propagate a CONST_INT into a conditional jump.
4194 We're pretty specific about what we will handle in this
4195 code, we can extend this as necessary over time.
4197 Right now the insn in question must look like
4198 (set (pc) (if_then_else ...)) */
4199 else if (alter_jumps
4200 && GET_CODE (insn) == JUMP_INSN
4201 && condjump_p (insn)
4202 && ! simplejump_p (insn))
4203 changed |= cprop_jump (bb, insn, reg_used->reg_rtx, src);
4206 /* Similar code for machines that use a pair of CC0 setter and
4207 conditional jump insn. */
4208 else if (alter_jumps
4209 && GET_CODE (PATTERN (insn)) == SET
4210 && SET_DEST (PATTERN (insn)) == cc0_rtx
4211 && GET_CODE (NEXT_INSN (insn)) == JUMP_INSN
4212 && condjump_p (NEXT_INSN (insn))
4213 && ! simplejump_p (NEXT_INSN (insn))
4214 && cprop_cc0_jump (bb, insn, reg_used, src))
4221 else if (GET_CODE (src) == REG
4222 && REGNO (src) >= FIRST_PSEUDO_REGISTER
4223 && REGNO (src) != regno)
4225 if (try_replace_reg (reg_used->reg_rtx, src, insn))
4229 if (gcse_file != NULL)
4231 fprintf (gcse_file, "COPY-PROP: Replacing reg %d in insn %d",
4232 regno, INSN_UID (insn));
4233 fprintf (gcse_file, " with reg %d\n", REGNO (src));
4236 /* The original insn setting reg_used may or may not now be
4237 deletable. We leave the deletion to flow. */
4238 /* FIXME: If it turns out that the insn isn't deletable,
4239 then we may have unnecessarily extended register lifetimes
4240 and made things worse. */
4248 /* Forward propagate copies. This includes copies and constants. Return
4249 non-zero if a change was made. */
4258 /* Note we start at block 1. */
4261 for (bb = 1; bb < n_basic_blocks; bb++)
4263 /* Reset tables used to keep track of what's still valid [since the
4264 start of the block]. */
4265 reset_opr_set_tables ();
4267 for (insn = BLOCK_HEAD (bb);
4268 insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
4269 insn = NEXT_INSN (insn))
4272 changed |= cprop_insn (BASIC_BLOCK (bb), insn, alter_jumps);
4274 /* Keep track of everything modified by this insn. */
4275 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
4276 call mark_oprs_set if we turned the insn into a NOTE. */
4277 if (GET_CODE (insn) != NOTE)
4278 mark_oprs_set (insn);
4282 if (gcse_file != NULL)
4283 fprintf (gcse_file, "\n");
4288 /* Perform one copy/constant propagation pass.
4289 F is the first insn in the function.
4290 PASS is the pass count. */
4293 one_cprop_pass (pass, alter_jumps)
4299 const_prop_count = 0;
4300 copy_prop_count = 0;
4302 alloc_set_hash_table (max_cuid);
4303 compute_set_hash_table ();
4305 dump_hash_table (gcse_file, "SET", set_hash_table, set_hash_table_size,
4309 alloc_cprop_mem (n_basic_blocks, n_sets);
4310 compute_cprop_data ();
4311 changed = cprop (alter_jumps);
4315 free_set_hash_table ();
4319 fprintf (gcse_file, "CPROP of %s, pass %d: %d bytes needed, ",
4320 current_function_name, pass, bytes_used);
4321 fprintf (gcse_file, "%d const props, %d copy props\n\n",
4322 const_prop_count, copy_prop_count);
4328 /* Compute PRE+LCM working variables. */
4330 /* Local properties of expressions. */
4331 /* Nonzero for expressions that are transparent in the block. */
4332 static sbitmap *transp;
4334 /* Nonzero for expressions that are transparent at the end of the block.
4335 This is only zero for expressions killed by abnormal critical edge
4336 created by a calls. */
4337 static sbitmap *transpout;
4339 /* Nonzero for expressions that are computed (available) in the block. */
4340 static sbitmap *comp;
4342 /* Nonzero for expressions that are locally anticipatable in the block. */
4343 static sbitmap *antloc;
4345 /* Nonzero for expressions where this block is an optimal computation
4347 static sbitmap *pre_optimal;
4349 /* Nonzero for expressions which are redundant in a particular block. */
4350 static sbitmap *pre_redundant;
4352 /* Nonzero for expressions which should be inserted on a specific edge. */
4353 static sbitmap *pre_insert_map;
4355 /* Nonzero for expressions which should be deleted in a specific block. */
4356 static sbitmap *pre_delete_map;
4358 /* Contains the edge_list returned by pre_edge_lcm. */
4359 static struct edge_list *edge_list;
4361 /* Redundant insns. */
4362 static sbitmap pre_redundant_insns;
4364 /* Allocate vars used for PRE analysis. */
4367 alloc_pre_mem (n_blocks, n_exprs)
4368 int n_blocks, n_exprs;
4370 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
4371 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
4372 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
4375 pre_redundant = NULL;
4376 pre_insert_map = NULL;
4377 pre_delete_map = NULL;
4380 ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
4382 /* pre_insert and pre_delete are allocated later. */
4385 /* Free vars used for PRE analysis. */
4390 sbitmap_vector_free (transp);
4391 sbitmap_vector_free (comp);
4393 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
4396 sbitmap_vector_free (pre_optimal);
4398 sbitmap_vector_free (pre_redundant);
4400 sbitmap_vector_free (pre_insert_map);
4402 sbitmap_vector_free (pre_delete_map);
4404 sbitmap_vector_free (ae_in);
4406 sbitmap_vector_free (ae_out);
4408 transp = comp = NULL;
4409 pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
4410 ae_in = ae_out = NULL;
4413 /* Top level routine to do the dataflow analysis needed by PRE. */
4418 sbitmap trapping_expr;
4422 compute_local_properties (transp, comp, antloc, 0);
4423 sbitmap_vector_zero (ae_kill, n_basic_blocks);
4425 /* Collect expressions which might trap. */
4426 trapping_expr = sbitmap_alloc (n_exprs);
4427 sbitmap_zero (trapping_expr);
4428 for (ui = 0; ui < expr_hash_table_size; ui++)
4431 for (e = expr_hash_table[ui]; e != NULL; e = e->next_same_hash)
4432 if (may_trap_p (e->expr))
4433 SET_BIT (trapping_expr, e->bitmap_index);
4436 /* Compute ae_kill for each basic block using:
4440 This is significantly faster than compute_ae_kill. */
4442 for (i = 0; i < n_basic_blocks; i++)
4446 /* If the current block is the destination of an abnormal edge, we
4447 kill all trapping expressions because we won't be able to properly
4448 place the instruction on the edge. So make them neither
4449 anticipatable nor transparent. This is fairly conservative. */
4450 for (e = BASIC_BLOCK (i)->pred; e ; e = e->pred_next)
4451 if (e->flags & EDGE_ABNORMAL)
4453 sbitmap_difference (antloc[i], antloc[i], trapping_expr);
4454 sbitmap_difference (transp[i], transp[i], trapping_expr);
4458 sbitmap_a_or_b (ae_kill[i], transp[i], comp[i]);
4459 sbitmap_not (ae_kill[i], ae_kill[i]);
4462 edge_list = pre_edge_lcm (gcse_file, n_exprs, transp, comp, antloc,
4463 ae_kill, &pre_insert_map, &pre_delete_map);
4464 sbitmap_vector_free (antloc);
4466 sbitmap_vector_free (ae_kill);
4468 free (trapping_expr);
4473 /* Return non-zero if an occurrence of expression EXPR in OCCR_BB would reach
4476 VISITED is a pointer to a working buffer for tracking which BB's have
4477 been visited. It is NULL for the top-level call.
4479 We treat reaching expressions that go through blocks containing the same
4480 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
4481 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
4482 2 as not reaching. The intent is to improve the probability of finding
4483 only one reaching expression and to reduce register lifetimes by picking
4484 the closest such expression. */
4487 pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited)
4488 basic_block occr_bb;
4495 for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
4497 basic_block pred_bb = pred->src;
4499 if (pred->src == ENTRY_BLOCK_PTR
4500 /* Has predecessor has already been visited? */
4501 || visited[pred_bb->index])
4502 ;/* Nothing to do. */
4504 /* Does this predecessor generate this expression? */
4505 else if (TEST_BIT (comp[pred_bb->index], expr->bitmap_index))
4507 /* Is this the occurrence we're looking for?
4508 Note that there's only one generating occurrence per block
4509 so we just need to check the block number. */
4510 if (occr_bb == pred_bb)
4513 visited[pred_bb->index] = 1;
4515 /* Ignore this predecessor if it kills the expression. */
4516 else if (! TEST_BIT (transp[pred_bb->index], expr->bitmap_index))
4517 visited[pred_bb->index] = 1;
4519 /* Neither gen nor kill. */
4522 visited[pred_bb->index] = 1;
4523 if (pre_expr_reaches_here_p_work (occr_bb, expr, pred_bb, visited))
4528 /* All paths have been checked. */
4532 /* The wrapper for pre_expr_reaches_here_work that ensures that any
4533 memory allocated for that function is returned. */
4536 pre_expr_reaches_here_p (occr_bb, expr, bb)
4537 basic_block occr_bb;
4542 char *visited = (char *) xcalloc (n_basic_blocks, 1);
4544 rval = pre_expr_reaches_here_p_work(occr_bb, expr, bb, visited);
4551 /* Given an expr, generate RTL which we can insert at the end of a BB,
4552 or on an edge. Set the block number of any insns generated to
4556 process_insert_insn (expr)
4559 rtx reg = expr->reaching_reg;
4560 rtx exp = copy_rtx (expr->expr);
4565 /* If the expression is something that's an operand, like a constant,
4566 just copy it to a register. */
4567 if (general_operand (exp, GET_MODE (reg)))
4568 emit_move_insn (reg, exp);
4570 /* Otherwise, make a new insn to compute this expression and make sure the
4571 insn will be recognized (this also adds any needed CLOBBERs). Copy the
4572 expression to make sure we don't have any sharing issues. */
4573 else if (insn_invalid_p (emit_insn (gen_rtx_SET (VOIDmode, reg, exp))))
4576 pat = gen_sequence ();
4582 /* Add EXPR to the end of basic block BB.
4584 This is used by both the PRE and code hoisting.
4586 For PRE, we want to verify that the expr is either transparent
4587 or locally anticipatable in the target block. This check makes
4588 no sense for code hoisting. */
4591 insert_insn_end_bb (expr, bb, pre)
4598 rtx reg = expr->reaching_reg;
4599 int regno = REGNO (reg);
4603 pat = process_insert_insn (expr);
4605 /* If the last insn is a jump, insert EXPR in front [taking care to
4606 handle cc0, etc. properly]. */
4608 if (GET_CODE (insn) == JUMP_INSN)
4614 /* If this is a jump table, then we can't insert stuff here. Since
4615 we know the previous real insn must be the tablejump, we insert
4616 the new instruction just before the tablejump. */
4617 if (GET_CODE (PATTERN (insn)) == ADDR_VEC
4618 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
4619 insn = prev_real_insn (insn);
4622 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
4623 if cc0 isn't set. */
4624 note = find_reg_note (insn, REG_CC_SETTER, NULL_RTX);
4626 insn = XEXP (note, 0);
4629 rtx maybe_cc0_setter = prev_nonnote_insn (insn);
4630 if (maybe_cc0_setter
4631 && INSN_P (maybe_cc0_setter)
4632 && sets_cc0_p (PATTERN (maybe_cc0_setter)))
4633 insn = maybe_cc0_setter;
4636 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4637 new_insn = emit_block_insn_before (pat, insn, bb);
4640 /* Likewise if the last insn is a call, as will happen in the presence
4641 of exception handling. */
4642 else if (GET_CODE (insn) == CALL_INSN)
4644 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4645 we search backward and place the instructions before the first
4646 parameter is loaded. Do this for everyone for consistency and a
4647 presumtion that we'll get better code elsewhere as well.
4649 It should always be the case that we can put these instructions
4650 anywhere in the basic block with performing PRE optimizations.
4654 && !TEST_BIT (antloc[bb->index], expr->bitmap_index)
4655 && !TEST_BIT (transp[bb->index], expr->bitmap_index))
4658 /* Since different machines initialize their parameter registers
4659 in different orders, assume nothing. Collect the set of all
4660 parameter registers. */
4661 insn = find_first_parameter_load (insn, bb->head);
4663 /* If we found all the parameter loads, then we want to insert
4664 before the first parameter load.
4666 If we did not find all the parameter loads, then we might have
4667 stopped on the head of the block, which could be a CODE_LABEL.
4668 If we inserted before the CODE_LABEL, then we would be putting
4669 the insn in the wrong basic block. In that case, put the insn
4670 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4671 while (GET_CODE (insn) == CODE_LABEL
4672 || NOTE_INSN_BASIC_BLOCK_P (insn))
4673 insn = NEXT_INSN (insn);
4675 new_insn = emit_block_insn_before (pat, insn, bb);
4679 new_insn = emit_insn_after (pat, insn);
4683 /* Keep block number table up to date.
4684 Note, PAT could be a multiple insn sequence, we have to make
4685 sure that each insn in the sequence is handled. */
4686 if (GET_CODE (pat) == SEQUENCE)
4688 for (i = 0; i < XVECLEN (pat, 0); i++)
4690 rtx insn = XVECEXP (pat, 0, i);
4692 set_block_for_insn (insn, bb);
4694 add_label_notes (PATTERN (insn), new_insn);
4696 note_stores (PATTERN (insn), record_set_info, insn);
4701 add_label_notes (SET_SRC (pat), new_insn);
4702 set_block_for_new_insns (new_insn, bb);
4704 /* Keep register set table up to date. */
4705 record_one_set (regno, new_insn);
4708 gcse_create_count++;
4712 fprintf (gcse_file, "PRE/HOIST: end of bb %d, insn %d, ",
4713 bb->index, INSN_UID (new_insn));
4714 fprintf (gcse_file, "copying expression %d to reg %d\n",
4715 expr->bitmap_index, regno);
4719 /* Insert partially redundant expressions on edges in the CFG to make
4720 the expressions fully redundant. */
4723 pre_edge_insert (edge_list, index_map)
4724 struct edge_list *edge_list;
4725 struct expr **index_map;
4727 int e, i, j, num_edges, set_size, did_insert = 0;
4730 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4731 if it reaches any of the deleted expressions. */
4733 set_size = pre_insert_map[0]->size;
4734 num_edges = NUM_EDGES (edge_list);
4735 inserted = sbitmap_vector_alloc (num_edges, n_exprs);
4736 sbitmap_vector_zero (inserted, num_edges);
4738 for (e = 0; e < num_edges; e++)
4741 basic_block bb = INDEX_EDGE_PRED_BB (edge_list, e);
4743 for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
4745 SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
4747 for (j = indx; insert && j < n_exprs; j++, insert >>= 1)
4748 if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
4750 struct expr *expr = index_map[j];
4753 /* Now look at each deleted occurence of this expression. */
4754 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4756 if (! occr->deleted_p)
4759 /* Insert this expression on this edge if if it would
4760 reach the deleted occurence in BB. */
4761 if (!TEST_BIT (inserted[e], j))
4764 edge eg = INDEX_EDGE (edge_list, e);
4766 /* We can't insert anything on an abnormal and
4767 critical edge, so we insert the insn at the end of
4768 the previous block. There are several alternatives
4769 detailed in Morgans book P277 (sec 10.5) for
4770 handling this situation. This one is easiest for
4773 if ((eg->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
4774 insert_insn_end_bb (index_map[j], bb, 0);
4777 insn = process_insert_insn (index_map[j]);
4778 insert_insn_on_edge (insn, eg);
4783 fprintf (gcse_file, "PRE/HOIST: edge (%d,%d), ",
4785 INDEX_EDGE_SUCC_BB (edge_list, e)->index);
4786 fprintf (gcse_file, "copy expression %d\n",
4787 expr->bitmap_index);
4790 update_ld_motion_stores (expr);
4791 SET_BIT (inserted[e], j);
4793 gcse_create_count++;
4800 sbitmap_vector_free (inserted);
4804 /* Copy the result of INSN to REG. INDX is the expression number. */
4807 pre_insert_copy_insn (expr, insn)
4811 rtx reg = expr->reaching_reg;
4812 int regno = REGNO (reg);
4813 int indx = expr->bitmap_index;
4814 rtx set = single_set (insn);
4816 basic_block bb = BLOCK_FOR_INSN (insn);
4821 new_insn = emit_insn_after (gen_move_insn (reg, SET_DEST (set)), insn);
4823 /* Keep block number table up to date. */
4824 set_block_for_new_insns (new_insn, bb);
4826 /* Keep register set table up to date. */
4827 record_one_set (regno, new_insn);
4828 if (insn == bb->end)
4831 gcse_create_count++;
4835 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4836 BLOCK_NUM (insn), INSN_UID (new_insn), indx,
4837 INSN_UID (insn), regno);
4838 update_ld_motion_stores (expr);
4841 /* Copy available expressions that reach the redundant expression
4842 to `reaching_reg'. */
4845 pre_insert_copies ()
4852 /* For each available expression in the table, copy the result to
4853 `reaching_reg' if the expression reaches a deleted one.
4855 ??? The current algorithm is rather brute force.
4856 Need to do some profiling. */
4858 for (i = 0; i < expr_hash_table_size; i++)
4859 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4861 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4862 we don't want to insert a copy here because the expression may not
4863 really be redundant. So only insert an insn if the expression was
4864 deleted. This test also avoids further processing if the
4865 expression wasn't deleted anywhere. */
4866 if (expr->reaching_reg == NULL)
4869 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4871 if (! occr->deleted_p)
4874 for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
4876 rtx insn = avail->insn;
4878 /* No need to handle this one if handled already. */
4879 if (avail->copied_p)
4882 /* Don't handle this one if it's a redundant one. */
4883 if (TEST_BIT (pre_redundant_insns, INSN_CUID (insn)))
4886 /* Or if the expression doesn't reach the deleted one. */
4887 if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail->insn),
4889 BLOCK_FOR_INSN (occr->insn)))
4892 /* Copy the result of avail to reaching_reg. */
4893 pre_insert_copy_insn (expr, insn);
4894 avail->copied_p = 1;
4900 /* Delete redundant computations.
4901 Deletion is done by changing the insn to copy the `reaching_reg' of
4902 the expression into the result of the SET. It is left to later passes
4903 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4905 Returns non-zero if a change is made. */
4916 for (i = 0; i < expr_hash_table_size; i++)
4917 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4919 int indx = expr->bitmap_index;
4921 /* We only need to search antic_occr since we require
4924 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4926 rtx insn = occr->insn;
4928 basic_block bb = BLOCK_FOR_INSN (insn);
4930 if (TEST_BIT (pre_delete_map[bb->index], indx))
4932 set = single_set (insn);
4936 /* Create a pseudo-reg to store the result of reaching
4937 expressions into. Get the mode for the new pseudo from
4938 the mode of the original destination pseudo. */
4939 if (expr->reaching_reg == NULL)
4941 = gen_reg_rtx (GET_MODE (SET_DEST (set)));
4943 /* In theory this should never fail since we're creating
4946 However, on the x86 some of the movXX patterns actually
4947 contain clobbers of scratch regs. This may cause the
4948 insn created by validate_change to not match any pattern
4949 and thus cause validate_change to fail. */
4950 if (validate_change (insn, &SET_SRC (set),
4951 expr->reaching_reg, 0))
4953 occr->deleted_p = 1;
4954 SET_BIT (pre_redundant_insns, INSN_CUID (insn));
4962 "PRE: redundant insn %d (expression %d) in ",
4963 INSN_UID (insn), indx);
4964 fprintf (gcse_file, "bb %d, reaching reg is %d\n",
4965 bb->index, REGNO (expr->reaching_reg));
4974 /* Perform GCSE optimizations using PRE.
4975 This is called by one_pre_gcse_pass after all the dataflow analysis
4978 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4979 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4980 Compiler Design and Implementation.
4982 ??? A new pseudo reg is created to hold the reaching expression. The nice
4983 thing about the classical approach is that it would try to use an existing
4984 reg. If the register can't be adequately optimized [i.e. we introduce
4985 reload problems], one could add a pass here to propagate the new register
4988 ??? We don't handle single sets in PARALLELs because we're [currently] not
4989 able to copy the rest of the parallel when we insert copies to create full
4990 redundancies from partial redundancies. However, there's no reason why we
4991 can't handle PARALLELs in the cases where there are no partial
4998 int did_insert, changed;
4999 struct expr **index_map;
5002 /* Compute a mapping from expression number (`bitmap_index') to
5003 hash table entry. */
5005 index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
5006 for (i = 0; i < expr_hash_table_size; i++)
5007 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
5008 index_map[expr->bitmap_index] = expr;
5010 /* Reset bitmap used to track which insns are redundant. */
5011 pre_redundant_insns = sbitmap_alloc (max_cuid);
5012 sbitmap_zero (pre_redundant_insns);
5014 /* Delete the redundant insns first so that
5015 - we know what register to use for the new insns and for the other
5016 ones with reaching expressions
5017 - we know which insns are redundant when we go to create copies */
5019 changed = pre_delete ();
5021 did_insert = pre_edge_insert (edge_list, index_map);
5023 /* In other places with reaching expressions, copy the expression to the
5024 specially allocated pseudo-reg that reaches the redundant expr. */
5025 pre_insert_copies ();
5028 commit_edge_insertions ();
5033 free (pre_redundant_insns);
5037 /* Top level routine to perform one PRE GCSE pass.
5039 Return non-zero if a change was made. */
5042 one_pre_gcse_pass (pass)
5047 gcse_subst_count = 0;
5048 gcse_create_count = 0;
5050 alloc_expr_hash_table (max_cuid);
5051 add_noreturn_fake_exit_edges ();
5053 compute_ld_motion_mems ();
5055 compute_expr_hash_table ();
5056 trim_ld_motion_mems ();
5058 dump_hash_table (gcse_file, "Expression", expr_hash_table,
5059 expr_hash_table_size, n_exprs);
5063 alloc_pre_mem (n_basic_blocks, n_exprs);
5064 compute_pre_data ();
5065 changed |= pre_gcse ();
5066 free_edge_list (edge_list);
5071 remove_fake_edges ();
5072 free_expr_hash_table ();
5076 fprintf (gcse_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
5077 current_function_name, pass, bytes_used);
5078 fprintf (gcse_file, "%d substs, %d insns created\n",
5079 gcse_subst_count, gcse_create_count);
5085 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to INSN.
5086 If notes are added to an insn which references a CODE_LABEL, the
5087 LABEL_NUSES count is incremented. We have to add REG_LABEL notes,
5088 because the following loop optimization pass requires them. */
5090 /* ??? This is very similar to the loop.c add_label_notes function. We
5091 could probably share code here. */
5093 /* ??? If there was a jump optimization pass after gcse and before loop,
5094 then we would not need to do this here, because jump would add the
5095 necessary REG_LABEL notes. */
5098 add_label_notes (x, insn)
5102 enum rtx_code code = GET_CODE (x);
5106 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
5108 /* This code used to ignore labels that referred to dispatch tables to
5109 avoid flow generating (slighly) worse code.
5111 We no longer ignore such label references (see LABEL_REF handling in
5112 mark_jump_label for additional information). */
5114 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
5116 if (LABEL_P (XEXP (x, 0)))
5117 LABEL_NUSES (XEXP (x, 0))++;
5121 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
5124 add_label_notes (XEXP (x, i), insn);
5125 else if (fmt[i] == 'E')
5126 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5127 add_label_notes (XVECEXP (x, i, j), insn);
5131 /* Compute transparent outgoing information for each block.
5133 An expression is transparent to an edge unless it is killed by
5134 the edge itself. This can only happen with abnormal control flow,
5135 when the edge is traversed through a call. This happens with
5136 non-local labels and exceptions.
5138 This would not be necessary if we split the edge. While this is
5139 normally impossible for abnormal critical edges, with some effort
5140 it should be possible with exception handling, since we still have
5141 control over which handler should be invoked. But due to increased
5142 EH table sizes, this may not be worthwhile. */
5145 compute_transpout ()
5151 sbitmap_vector_ones (transpout, n_basic_blocks);
5153 for (bb = 0; bb < n_basic_blocks; ++bb)
5155 /* Note that flow inserted a nop a the end of basic blocks that
5156 end in call instructions for reasons other than abnormal
5158 if (GET_CODE (BLOCK_END (bb)) != CALL_INSN)
5161 for (i = 0; i < expr_hash_table_size; i++)
5162 for (expr = expr_hash_table[i]; expr ; expr = expr->next_same_hash)
5163 if (GET_CODE (expr->expr) == MEM)
5165 if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
5166 && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
5169 /* ??? Optimally, we would use interprocedural alias
5170 analysis to determine if this mem is actually killed
5172 RESET_BIT (transpout[bb], expr->bitmap_index);
5177 /* Removal of useless null pointer checks */
5179 /* Called via note_stores. X is set by SETTER. If X is a register we must
5180 invalidate nonnull_local and set nonnull_killed. DATA is really a
5181 `null_pointer_info *'.
5183 We ignore hard registers. */
5186 invalidate_nonnull_info (x, setter, data)
5188 rtx setter ATTRIBUTE_UNUSED;
5192 struct null_pointer_info *npi = (struct null_pointer_info *) data;
5194 while (GET_CODE (x) == SUBREG)
5197 /* Ignore anything that is not a register or is a hard register. */
5198 if (GET_CODE (x) != REG
5199 || REGNO (x) < npi->min_reg
5200 || REGNO (x) >= npi->max_reg)
5203 regno = REGNO (x) - npi->min_reg;
5205 RESET_BIT (npi->nonnull_local[npi->current_block], regno);
5206 SET_BIT (npi->nonnull_killed[npi->current_block], regno);
5209 /* Do null-pointer check elimination for the registers indicated in
5210 NPI. NONNULL_AVIN and NONNULL_AVOUT are pre-allocated sbitmaps;
5211 they are not our responsibility to free. */
5214 delete_null_pointer_checks_1 (delete_list, block_reg, nonnull_avin,
5216 varray_type *delete_list;
5217 unsigned int *block_reg;
5218 sbitmap *nonnull_avin;
5219 sbitmap *nonnull_avout;
5220 struct null_pointer_info *npi;
5224 sbitmap *nonnull_local = npi->nonnull_local;
5225 sbitmap *nonnull_killed = npi->nonnull_killed;
5227 /* Compute local properties, nonnull and killed. A register will have
5228 the nonnull property if at the end of the current block its value is
5229 known to be nonnull. The killed property indicates that somewhere in
5230 the block any information we had about the register is killed.
5232 Note that a register can have both properties in a single block. That
5233 indicates that it's killed, then later in the block a new value is
5235 sbitmap_vector_zero (nonnull_local, n_basic_blocks);
5236 sbitmap_vector_zero (nonnull_killed, n_basic_blocks);
5238 for (current_block = 0; current_block < n_basic_blocks; current_block++)
5240 rtx insn, stop_insn;
5242 /* Set the current block for invalidate_nonnull_info. */
5243 npi->current_block = current_block;
5245 /* Scan each insn in the basic block looking for memory references and
5247 stop_insn = NEXT_INSN (BLOCK_END (current_block));
5248 for (insn = BLOCK_HEAD (current_block);
5250 insn = NEXT_INSN (insn))
5255 /* Ignore anything that is not a normal insn. */
5256 if (! INSN_P (insn))
5259 /* Basically ignore anything that is not a simple SET. We do have
5260 to make sure to invalidate nonnull_local and set nonnull_killed
5261 for such insns though. */
5262 set = single_set (insn);
5265 note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
5269 /* See if we've got a useable memory load. We handle it first
5270 in case it uses its address register as a dest (which kills
5271 the nonnull property). */
5272 if (GET_CODE (SET_SRC (set)) == MEM
5273 && GET_CODE ((reg = XEXP (SET_SRC (set), 0))) == REG
5274 && REGNO (reg) >= npi->min_reg
5275 && REGNO (reg) < npi->max_reg)
5276 SET_BIT (nonnull_local[current_block],
5277 REGNO (reg) - npi->min_reg);
5279 /* Now invalidate stuff clobbered by this insn. */
5280 note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
5282 /* And handle stores, we do these last since any sets in INSN can
5283 not kill the nonnull property if it is derived from a MEM
5284 appearing in a SET_DEST. */
5285 if (GET_CODE (SET_DEST (set)) == MEM
5286 && GET_CODE ((reg = XEXP (SET_DEST (set), 0))) == REG
5287 && REGNO (reg) >= npi->min_reg
5288 && REGNO (reg) < npi->max_reg)
5289 SET_BIT (nonnull_local[current_block],
5290 REGNO (reg) - npi->min_reg);
5294 /* Now compute global properties based on the local properties. This
5295 is a classic global availablity algorithm. */
5296 compute_available (nonnull_local, nonnull_killed,
5297 nonnull_avout, nonnull_avin);
5299 /* Now look at each bb and see if it ends with a compare of a value
5301 for (bb = 0; bb < n_basic_blocks; bb++)
5303 rtx last_insn = BLOCK_END (bb);
5304 rtx condition, earliest;
5305 int compare_and_branch;
5307 /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
5308 since BLOCK_REG[BB] is zero if this block did not end with a
5309 comparison against zero, this condition works. */
5310 if (block_reg[bb] < npi->min_reg
5311 || block_reg[bb] >= npi->max_reg)
5314 /* LAST_INSN is a conditional jump. Get its condition. */
5315 condition = get_condition (last_insn, &earliest);
5317 /* If we can't determine the condition then skip. */
5321 /* Is the register known to have a nonzero value? */
5322 if (!TEST_BIT (nonnull_avout[bb], block_reg[bb] - npi->min_reg))
5325 /* Try to compute whether the compare/branch at the loop end is one or
5326 two instructions. */
5327 if (earliest == last_insn)
5328 compare_and_branch = 1;
5329 else if (earliest == prev_nonnote_insn (last_insn))
5330 compare_and_branch = 2;
5334 /* We know the register in this comparison is nonnull at exit from
5335 this block. We can optimize this comparison. */
5336 if (GET_CODE (condition) == NE)
5340 new_jump = emit_jump_insn_before (gen_jump (JUMP_LABEL (last_insn)),
5342 JUMP_LABEL (new_jump) = JUMP_LABEL (last_insn);
5343 LABEL_NUSES (JUMP_LABEL (new_jump))++;
5344 emit_barrier_after (new_jump);
5347 VARRAY_RTX_INIT (*delete_list, 10, "delete_list");
5349 VARRAY_PUSH_RTX (*delete_list, last_insn);
5350 if (compare_and_branch == 2)
5351 VARRAY_PUSH_RTX (*delete_list, earliest);
5353 /* Don't check this block again. (Note that BLOCK_END is
5354 invalid here; we deleted the last instruction in the
5360 /* Find EQ/NE comparisons against zero which can be (indirectly) evaluated
5363 This is conceptually similar to global constant/copy propagation and
5364 classic global CSE (it even uses the same dataflow equations as cprop).
5366 If a register is used as memory address with the form (mem (reg)), then we
5367 know that REG can not be zero at that point in the program. Any instruction
5368 which sets REG "kills" this property.
5370 So, if every path leading to a conditional branch has an available memory
5371 reference of that form, then we know the register can not have the value
5372 zero at the conditional branch.
5374 So we merely need to compute the local properies and propagate that data
5375 around the cfg, then optimize where possible.
5377 We run this pass two times. Once before CSE, then again after CSE. This
5378 has proven to be the most profitable approach. It is rare for new
5379 optimization opportunities of this nature to appear after the first CSE
5382 This could probably be integrated with global cprop with a little work. */
5385 delete_null_pointer_checks (f)
5386 rtx f ATTRIBUTE_UNUSED;
5388 sbitmap *nonnull_avin, *nonnull_avout;
5389 unsigned int *block_reg;
5390 varray_type delete_list = NULL;
5396 struct null_pointer_info npi;
5398 /* If we have only a single block, then there's nothing to do. */
5399 if (n_basic_blocks <= 1)
5402 /* Trying to perform global optimizations on flow graphs which have
5403 a high connectivity will take a long time and is unlikely to be
5404 particularly useful.
5406 In normal circumstances a cfg should have about twice as many edges
5407 as blocks. But we do not want to punish small functions which have
5408 a couple switch statements. So we require a relatively large number
5409 of basic blocks and the ratio of edges to blocks to be high. */
5410 if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
5413 /* We need four bitmaps, each with a bit for each register in each
5415 max_reg = max_reg_num ();
5416 regs_per_pass = get_bitmap_width (4, n_basic_blocks, max_reg);
5418 /* Allocate bitmaps to hold local and global properties. */
5419 npi.nonnull_local = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5420 npi.nonnull_killed = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5421 nonnull_avin = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5422 nonnull_avout = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5424 /* Go through the basic blocks, seeing whether or not each block
5425 ends with a conditional branch whose condition is a comparison
5426 against zero. Record the register compared in BLOCK_REG. */
5427 block_reg = (unsigned int *) xcalloc (n_basic_blocks, sizeof (int));
5428 for (bb = 0; bb < n_basic_blocks; bb++)
5430 rtx last_insn = BLOCK_END (bb);
5431 rtx condition, earliest, reg;
5433 /* We only want conditional branches. */
5434 if (GET_CODE (last_insn) != JUMP_INSN
5435 || !any_condjump_p (last_insn)
5436 || !onlyjump_p (last_insn))
5439 /* LAST_INSN is a conditional jump. Get its condition. */
5440 condition = get_condition (last_insn, &earliest);
5442 /* If we were unable to get the condition, or it is not a equality
5443 comparison against zero then there's nothing we can do. */
5445 || (GET_CODE (condition) != NE && GET_CODE (condition) != EQ)
5446 || GET_CODE (XEXP (condition, 1)) != CONST_INT
5447 || (XEXP (condition, 1)
5448 != CONST0_RTX (GET_MODE (XEXP (condition, 0)))))
5451 /* We must be checking a register against zero. */
5452 reg = XEXP (condition, 0);
5453 if (GET_CODE (reg) != REG)
5456 block_reg[bb] = REGNO (reg);
5459 /* Go through the algorithm for each block of registers. */
5460 for (reg = FIRST_PSEUDO_REGISTER; reg < max_reg; reg += regs_per_pass)
5463 npi.max_reg = MIN (reg + regs_per_pass, max_reg);
5464 delete_null_pointer_checks_1 (&delete_list, block_reg, nonnull_avin,
5465 nonnull_avout, &npi);
5468 /* Now delete the instructions all at once. This breaks the CFG. */
5471 for (i = 0; i < VARRAY_ACTIVE_SIZE (delete_list); i++)
5472 delete_insn (VARRAY_RTX (delete_list, i));
5473 VARRAY_FREE (delete_list);
5476 /* Free the table of registers compared at the end of every block. */
5480 sbitmap_vector_free (npi.nonnull_local);
5481 sbitmap_vector_free (npi.nonnull_killed);
5482 sbitmap_vector_free (nonnull_avin);
5483 sbitmap_vector_free (nonnull_avout);
5486 /* Code Hoisting variables and subroutines. */
5488 /* Very busy expressions. */
5489 static sbitmap *hoist_vbein;
5490 static sbitmap *hoist_vbeout;
5492 /* Hoistable expressions. */
5493 static sbitmap *hoist_exprs;
5495 /* Dominator bitmaps. */
5496 static sbitmap *dominators;
5498 /* ??? We could compute post dominators and run this algorithm in
5499 reverse to to perform tail merging, doing so would probably be
5500 more effective than the tail merging code in jump.c.
5502 It's unclear if tail merging could be run in parallel with
5503 code hoisting. It would be nice. */
5505 /* Allocate vars used for code hoisting analysis. */
5508 alloc_code_hoist_mem (n_blocks, n_exprs)
5509 int n_blocks, n_exprs;
5511 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
5512 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
5513 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
5515 hoist_vbein = sbitmap_vector_alloc (n_blocks, n_exprs);
5516 hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
5517 hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
5518 transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
5520 dominators = sbitmap_vector_alloc (n_blocks, n_blocks);
5523 /* Free vars used for code hoisting analysis. */
5526 free_code_hoist_mem ()
5528 sbitmap_vector_free (antloc);
5529 sbitmap_vector_free (transp);
5530 sbitmap_vector_free (comp);
5532 sbitmap_vector_free (hoist_vbein);
5533 sbitmap_vector_free (hoist_vbeout);
5534 sbitmap_vector_free (hoist_exprs);
5535 sbitmap_vector_free (transpout);
5537 sbitmap_vector_free (dominators);
5540 /* Compute the very busy expressions at entry/exit from each block.
5542 An expression is very busy if all paths from a given point
5543 compute the expression. */
5546 compute_code_hoist_vbeinout ()
5548 int bb, changed, passes;
5550 sbitmap_vector_zero (hoist_vbeout, n_basic_blocks);
5551 sbitmap_vector_zero (hoist_vbein, n_basic_blocks);
5560 /* We scan the blocks in the reverse order to speed up
5562 for (bb = n_basic_blocks - 1; bb >= 0; bb--)
5564 changed |= sbitmap_a_or_b_and_c (hoist_vbein[bb], antloc[bb],
5565 hoist_vbeout[bb], transp[bb]);
5566 if (bb != n_basic_blocks - 1)
5567 sbitmap_intersection_of_succs (hoist_vbeout[bb], hoist_vbein, bb);
5574 fprintf (gcse_file, "hoisting vbeinout computation: %d passes\n", passes);
5577 /* Top level routine to do the dataflow analysis needed by code hoisting. */
5580 compute_code_hoist_data ()
5582 compute_local_properties (transp, comp, antloc, 0);
5583 compute_transpout ();
5584 compute_code_hoist_vbeinout ();
5585 calculate_dominance_info (NULL, dominators, CDI_DOMINATORS);
5587 fprintf (gcse_file, "\n");
5590 /* Determine if the expression identified by EXPR_INDEX would
5591 reach BB unimpared if it was placed at the end of EXPR_BB.
5593 It's unclear exactly what Muchnick meant by "unimpared". It seems
5594 to me that the expression must either be computed or transparent in
5595 *every* block in the path(s) from EXPR_BB to BB. Any other definition
5596 would allow the expression to be hoisted out of loops, even if
5597 the expression wasn't a loop invariant.
5599 Contrast this to reachability for PRE where an expression is
5600 considered reachable if *any* path reaches instead of *all*
5604 hoist_expr_reaches_here_p (expr_bb, expr_index, bb, visited)
5605 basic_block expr_bb;
5611 int visited_allocated_locally = 0;
5614 if (visited == NULL)
5616 visited_allocated_locally = 1;
5617 visited = xcalloc (n_basic_blocks, 1);
5620 for (pred = bb->pred; pred != NULL; pred = pred->pred_next)
5622 basic_block pred_bb = pred->src;
5624 if (pred->src == ENTRY_BLOCK_PTR)
5626 else if (visited[pred_bb->index])
5629 /* Does this predecessor generate this expression? */
5630 else if (TEST_BIT (comp[pred_bb->index], expr_index))
5632 else if (! TEST_BIT (transp[pred_bb->index], expr_index))
5638 visited[pred_bb->index] = 1;
5639 if (! hoist_expr_reaches_here_p (expr_bb, expr_index,
5644 if (visited_allocated_locally)
5647 return (pred == NULL);
5650 /* Actually perform code hoisting. */
5657 struct expr **index_map;
5660 sbitmap_vector_zero (hoist_exprs, n_basic_blocks);
5662 /* Compute a mapping from expression number (`bitmap_index') to
5663 hash table entry. */
5665 index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
5666 for (i = 0; i < expr_hash_table_size; i++)
5667 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
5668 index_map[expr->bitmap_index] = expr;
5670 /* Walk over each basic block looking for potentially hoistable
5671 expressions, nothing gets hoisted from the entry block. */
5672 for (bb = 0; bb < n_basic_blocks; bb++)
5675 int insn_inserted_p;
5677 /* Examine each expression that is very busy at the exit of this
5678 block. These are the potentially hoistable expressions. */
5679 for (i = 0; i < hoist_vbeout[bb]->n_bits; i++)
5683 if (TEST_BIT (hoist_vbeout[bb], i) && TEST_BIT (transpout[bb], i))
5685 /* We've found a potentially hoistable expression, now
5686 we look at every block BB dominates to see if it
5687 computes the expression. */
5688 for (dominated = 0; dominated < n_basic_blocks; dominated++)
5690 /* Ignore self dominance. */
5692 || ! TEST_BIT (dominators[dominated], bb))
5695 /* We've found a dominated block, now see if it computes
5696 the busy expression and whether or not moving that
5697 expression to the "beginning" of that block is safe. */
5698 if (!TEST_BIT (antloc[dominated], i))
5701 /* Note if the expression would reach the dominated block
5702 unimpared if it was placed at the end of BB.
5704 Keep track of how many times this expression is hoistable
5705 from a dominated block into BB. */
5706 if (hoist_expr_reaches_here_p (BASIC_BLOCK (bb), i,
5707 BASIC_BLOCK (dominated), NULL))
5711 /* If we found more than one hoistable occurence of this
5712 expression, then note it in the bitmap of expressions to
5713 hoist. It makes no sense to hoist things which are computed
5714 in only one BB, and doing so tends to pessimize register
5715 allocation. One could increase this value to try harder
5716 to avoid any possible code expansion due to register
5717 allocation issues; however experiments have shown that
5718 the vast majority of hoistable expressions are only movable
5719 from two successors, so raising this threshhold is likely
5720 to nullify any benefit we get from code hoisting. */
5723 SET_BIT (hoist_exprs[bb], i);
5729 /* If we found nothing to hoist, then quit now. */
5733 /* Loop over all the hoistable expressions. */
5734 for (i = 0; i < hoist_exprs[bb]->n_bits; i++)
5736 /* We want to insert the expression into BB only once, so
5737 note when we've inserted it. */
5738 insn_inserted_p = 0;
5740 /* These tests should be the same as the tests above. */
5741 if (TEST_BIT (hoist_vbeout[bb], i))
5743 /* We've found a potentially hoistable expression, now
5744 we look at every block BB dominates to see if it
5745 computes the expression. */
5746 for (dominated = 0; dominated < n_basic_blocks; dominated++)
5748 /* Ignore self dominance. */
5750 || ! TEST_BIT (dominators[dominated], bb))
5753 /* We've found a dominated block, now see if it computes
5754 the busy expression and whether or not moving that
5755 expression to the "beginning" of that block is safe. */
5756 if (!TEST_BIT (antloc[dominated], i))
5759 /* The expression is computed in the dominated block and
5760 it would be safe to compute it at the start of the
5761 dominated block. Now we have to determine if the
5762 expresion would reach the dominated block if it was
5763 placed at the end of BB. */
5764 if (hoist_expr_reaches_here_p (BASIC_BLOCK (bb), i,
5765 BASIC_BLOCK (dominated), NULL))
5767 struct expr *expr = index_map[i];
5768 struct occr *occr = expr->antic_occr;
5772 /* Find the right occurence of this expression. */
5773 while (BLOCK_NUM (occr->insn) != dominated && occr)
5776 /* Should never happen. */
5782 set = single_set (insn);
5786 /* Create a pseudo-reg to store the result of reaching
5787 expressions into. Get the mode for the new pseudo
5788 from the mode of the original destination pseudo. */
5789 if (expr->reaching_reg == NULL)
5791 = gen_reg_rtx (GET_MODE (SET_DEST (set)));
5793 /* In theory this should never fail since we're creating
5796 However, on the x86 some of the movXX patterns
5797 actually contain clobbers of scratch regs. This may
5798 cause the insn created by validate_change to not
5799 match any pattern and thus cause validate_change to
5801 if (validate_change (insn, &SET_SRC (set),
5802 expr->reaching_reg, 0))
5804 occr->deleted_p = 1;
5805 if (!insn_inserted_p)
5807 insert_insn_end_bb (index_map[i],
5808 BASIC_BLOCK (bb), 0);
5809 insn_inserted_p = 1;
5821 /* Top level routine to perform one code hoisting (aka unification) pass
5823 Return non-zero if a change was made. */
5826 one_code_hoisting_pass ()
5830 alloc_expr_hash_table (max_cuid);
5831 compute_expr_hash_table ();
5833 dump_hash_table (gcse_file, "Code Hosting Expressions", expr_hash_table,
5834 expr_hash_table_size, n_exprs);
5838 alloc_code_hoist_mem (n_basic_blocks, n_exprs);
5839 compute_code_hoist_data ();
5841 free_code_hoist_mem ();
5844 free_expr_hash_table ();
5849 /* Here we provide the things required to do store motion towards
5850 the exit. In order for this to be effective, gcse also needed to
5851 be taught how to move a load when it is kill only by a store to itself.
5856 void foo(float scale)
5858 for (i=0; i<10; i++)
5862 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
5863 the load out since its live around the loop, and stored at the bottom
5866 The 'Load Motion' referred to and implemented in this file is
5867 an enhancement to gcse which when using edge based lcm, recognizes
5868 this situation and allows gcse to move the load out of the loop.
5870 Once gcse has hoisted the load, store motion can then push this
5871 load towards the exit, and we end up with no loads or stores of 'i'
5874 /* This will search the ldst list for a matching expresion. If it
5875 doesn't find one, we create one and initialize it. */
5877 static struct ls_expr *
5881 struct ls_expr * ptr;
5883 for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
5884 if (expr_equiv_p (ptr->pattern, x))
5889 ptr = (struct ls_expr *) xmalloc (sizeof (struct ls_expr));
5891 ptr->next = pre_ldst_mems;
5894 ptr->loads = NULL_RTX;
5895 ptr->stores = NULL_RTX;
5896 ptr->reaching_reg = NULL_RTX;
5899 ptr->hash_index = 0;
5900 pre_ldst_mems = ptr;
5906 /* Free up an individual ldst entry. */
5909 free_ldst_entry (ptr)
5910 struct ls_expr * ptr;
5912 free_INSN_LIST_list (& ptr->loads);
5913 free_INSN_LIST_list (& ptr->stores);
5918 /* Free up all memory associated with the ldst list. */
5923 while (pre_ldst_mems)
5925 struct ls_expr * tmp = pre_ldst_mems;
5927 pre_ldst_mems = pre_ldst_mems->next;
5929 free_ldst_entry (tmp);
5932 pre_ldst_mems = NULL;
5935 /* Dump debugging info about the ldst list. */
5938 print_ldst_list (file)
5941 struct ls_expr * ptr;
5943 fprintf (file, "LDST list: \n");
5945 for (ptr = first_ls_expr(); ptr != NULL; ptr = next_ls_expr (ptr))
5947 fprintf (file, " Pattern (%3d): ", ptr->index);
5949 print_rtl (file, ptr->pattern);
5951 fprintf (file, "\n Loads : ");
5954 print_rtl (file, ptr->loads);
5956 fprintf (file, "(nil)");
5958 fprintf (file, "\n Stores : ");
5961 print_rtl (file, ptr->stores);
5963 fprintf (file, "(nil)");
5965 fprintf (file, "\n\n");
5968 fprintf (file, "\n");
5971 /* Returns 1 if X is in the list of ldst only expressions. */
5973 static struct ls_expr *
5974 find_rtx_in_ldst (x)
5977 struct ls_expr * ptr;
5979 for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
5980 if (expr_equiv_p (ptr->pattern, x) && ! ptr->invalid)
5986 /* Assign each element of the list of mems a monotonically increasing value. */
5991 struct ls_expr * ptr;
5994 for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
6000 /* Return first item in the list. */
6002 static inline struct ls_expr *
6005 return pre_ldst_mems;
6008 /* Return the next item in ther list after the specified one. */
6010 static inline struct ls_expr *
6012 struct ls_expr * ptr;
6017 /* Load Motion for loads which only kill themselves. */
6019 /* Return true if x is a simple MEM operation, with no registers or
6020 side effects. These are the types of loads we consider for the
6021 ld_motion list, otherwise we let the usual aliasing take care of it. */
6027 if (GET_CODE (x) != MEM)
6030 if (MEM_VOLATILE_P (x))
6033 if (GET_MODE (x) == BLKmode)
6036 if (!rtx_varies_p (XEXP (x, 0), 0))
6042 /* Make sure there isn't a buried reference in this pattern anywhere.
6043 If there is, invalidate the entry for it since we're not capable
6044 of fixing it up just yet.. We have to be sure we know about ALL
6045 loads since the aliasing code will allow all entries in the
6046 ld_motion list to not-alias itself. If we miss a load, we will get
6047 the wrong value since gcse might common it and we won't know to
6051 invalidate_any_buried_refs (x)
6056 struct ls_expr * ptr;
6058 /* Invalidate it in the list. */
6059 if (GET_CODE (x) == MEM && simple_mem (x))
6061 ptr = ldst_entry (x);
6065 /* Recursively process the insn. */
6066 fmt = GET_RTX_FORMAT (GET_CODE (x));
6068 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
6071 invalidate_any_buried_refs (XEXP (x, i));
6072 else if (fmt[i] == 'E')
6073 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6074 invalidate_any_buried_refs (XVECEXP (x, i, j));
6078 /* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
6079 being defined as MEM loads and stores to symbols, with no
6080 side effects and no registers in the expression. If there are any
6081 uses/defs which dont match this criteria, it is invalidated and
6082 trimmed out later. */
6085 compute_ld_motion_mems ()
6087 struct ls_expr * ptr;
6091 pre_ldst_mems = NULL;
6093 for (bb = 0; bb < n_basic_blocks; bb++)
6095 for (insn = BLOCK_HEAD (bb);
6096 insn && insn != NEXT_INSN (BLOCK_END (bb));
6097 insn = NEXT_INSN (insn))
6099 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
6101 if (GET_CODE (PATTERN (insn)) == SET)
6103 rtx src = SET_SRC (PATTERN (insn));
6104 rtx dest = SET_DEST (PATTERN (insn));
6106 /* Check for a simple LOAD... */
6107 if (GET_CODE (src) == MEM && simple_mem (src))
6109 ptr = ldst_entry (src);
6110 if (GET_CODE (dest) == REG)
6111 ptr->loads = alloc_INSN_LIST (insn, ptr->loads);
6117 /* Make sure there isn't a buried load somewhere. */
6118 invalidate_any_buried_refs (src);
6121 /* Check for stores. Don't worry about aliased ones, they
6122 will block any movement we might do later. We only care
6123 about this exact pattern since those are the only
6124 circumstance that we will ignore the aliasing info. */
6125 if (GET_CODE (dest) == MEM && simple_mem (dest))
6127 ptr = ldst_entry (dest);
6129 if (GET_CODE (src) != MEM
6130 && GET_CODE (src) != ASM_OPERANDS)
6131 ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
6137 invalidate_any_buried_refs (PATTERN (insn));
6143 /* Remove any references that have been either invalidated or are not in the
6144 expression list for pre gcse. */
6147 trim_ld_motion_mems ()
6149 struct ls_expr * last = NULL;
6150 struct ls_expr * ptr = first_ls_expr ();
6154 int del = ptr->invalid;
6155 struct expr * expr = NULL;
6157 /* Delete if entry has been made invalid. */
6163 /* Delete if we cannot find this mem in the expression list. */
6164 for (i = 0; i < expr_hash_table_size && del; i++)
6166 for (expr = expr_hash_table[i];
6168 expr = expr->next_same_hash)
6169 if (expr_equiv_p (expr->expr, ptr->pattern))
6181 last->next = ptr->next;
6182 free_ldst_entry (ptr);
6187 pre_ldst_mems = pre_ldst_mems->next;
6188 free_ldst_entry (ptr);
6189 ptr = pre_ldst_mems;
6194 /* Set the expression field if we are keeping it. */
6201 /* Show the world what we've found. */
6202 if (gcse_file && pre_ldst_mems != NULL)
6203 print_ldst_list (gcse_file);
6206 /* This routine will take an expression which we are replacing with
6207 a reaching register, and update any stores that are needed if
6208 that expression is in the ld_motion list. Stores are updated by
6209 copying their SRC to the reaching register, and then storeing
6210 the reaching register into the store location. These keeps the
6211 correct value in the reaching register for the loads. */
6214 update_ld_motion_stores (expr)
6217 struct ls_expr * mem_ptr;
6219 if ((mem_ptr = find_rtx_in_ldst (expr->expr)))
6221 /* We can try to find just the REACHED stores, but is shouldn't
6222 matter to set the reaching reg everywhere... some might be
6223 dead and should be eliminated later. */
6225 /* We replace SET mem = expr with
6227 SET mem = reg , where reg is the
6228 reaching reg used in the load. */
6229 rtx list = mem_ptr->stores;
6231 for ( ; list != NULL_RTX; list = XEXP (list, 1))
6233 rtx insn = XEXP (list, 0);
6234 rtx pat = PATTERN (insn);
6235 rtx src = SET_SRC (pat);
6236 rtx reg = expr->reaching_reg;
6239 /* If we've already copied it, continue. */
6240 if (expr->reaching_reg == src)
6245 fprintf (gcse_file, "PRE: store updated with reaching reg ");
6246 print_rtl (gcse_file, expr->reaching_reg);
6247 fprintf (gcse_file, ":\n ");
6248 print_inline_rtx (gcse_file, insn, 8);
6249 fprintf (gcse_file, "\n");
6252 copy = gen_move_insn ( reg, SET_SRC (pat));
6253 new = emit_insn_before (copy, insn);
6254 record_one_set (REGNO (reg), new);
6255 set_block_for_new_insns (new, BLOCK_FOR_INSN (insn));
6256 SET_SRC (pat) = reg;
6258 /* un-recognize this pattern since it's probably different now. */
6259 INSN_CODE (insn) = -1;
6260 gcse_create_count++;
6265 /* Store motion code. */
6267 /* This is used to communicate the target bitvector we want to use in the
6268 reg_set_info routine when called via the note_stores mechanism. */
6269 static sbitmap * regvec;
6271 /* Used in computing the reverse edge graph bit vectors. */
6272 static sbitmap * st_antloc;
6274 /* Global holding the number of store expressions we are dealing with. */
6275 static int num_stores;
6277 /* Checks to set if we need to mark a register set. Called from note_stores. */
6280 reg_set_info (dest, setter, data)
6281 rtx dest, setter ATTRIBUTE_UNUSED;
6282 void * data ATTRIBUTE_UNUSED;
6284 if (GET_CODE (dest) == SUBREG)
6285 dest = SUBREG_REG (dest);
6287 if (GET_CODE (dest) == REG)
6288 SET_BIT (*regvec, REGNO (dest));
6291 /* Return non-zero if the register operands of expression X are killed
6292 anywhere in basic block BB. */
6295 store_ops_ok (x, bb)
6303 /* Repeat is used to turn tail-recursion into iteration. */
6309 code = GET_CODE (x);
6313 /* If a reg has changed after us in this
6314 block, the operand has been killed. */
6315 return TEST_BIT (reg_set_in_block[bb->index], REGNO (x));
6342 i = GET_RTX_LENGTH (code) - 1;
6343 fmt = GET_RTX_FORMAT (code);
6349 rtx tem = XEXP (x, i);
6351 /* If we are about to do the last recursive call
6352 needed at this level, change it into iteration.
6353 This function is called enough to be worth it. */
6360 if (! store_ops_ok (tem, bb))
6363 else if (fmt[i] == 'E')
6367 for (j = 0; j < XVECLEN (x, i); j++)
6369 if (! store_ops_ok (XVECEXP (x, i, j), bb))
6378 /* Determine whether insn is MEM store pattern that we will consider moving. */
6381 find_moveable_store (insn)
6384 struct ls_expr * ptr;
6385 rtx dest = PATTERN (insn);
6387 if (GET_CODE (dest) != SET
6388 || GET_CODE (SET_SRC (dest)) == ASM_OPERANDS)
6391 dest = SET_DEST (dest);
6393 if (GET_CODE (dest) != MEM || MEM_VOLATILE_P (dest)
6394 || GET_MODE (dest) == BLKmode)
6397 if (GET_CODE (XEXP (dest, 0)) != SYMBOL_REF)
6400 if (rtx_varies_p (XEXP (dest, 0), 0))
6403 ptr = ldst_entry (dest);
6404 ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
6407 /* Perform store motion. Much like gcse, except we move expressions the
6408 other way by looking at the flowgraph in reverse. */
6411 compute_store_table ()
6417 max_gcse_regno = max_reg_num ();
6419 reg_set_in_block = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks,
6421 sbitmap_vector_zero (reg_set_in_block, n_basic_blocks);
6424 /* Find all the stores we care about. */
6425 for (bb = 0; bb < n_basic_blocks; bb++)
6427 regvec = & (reg_set_in_block[bb]);
6428 for (insn = BLOCK_END (bb);
6429 insn && insn != PREV_INSN (BLOCK_HEAD (bb));
6430 insn = PREV_INSN (insn))
6432 /* Ignore anything that is not a normal insn. */
6433 if (! INSN_P (insn))
6436 if (GET_CODE (insn) == CALL_INSN)
6438 bool clobbers_all = false;
6439 #ifdef NON_SAVING_SETJMP
6440 if (NON_SAVING_SETJMP
6441 && find_reg_note (insn, REG_SETJMP, NULL_RTX))
6442 clobbers_all = true;
6445 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
6447 || TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
6448 SET_BIT (reg_set_in_block[bb], regno);
6451 pat = PATTERN (insn);
6452 note_stores (pat, reg_set_info, NULL);
6454 /* Now that we've marked regs, look for stores. */
6455 if (GET_CODE (pat) == SET)
6456 find_moveable_store (insn);
6460 ret = enumerate_ldsts ();
6464 fprintf (gcse_file, "Store Motion Expressions.\n");
6465 print_ldst_list (gcse_file);
6471 /* Check to see if the load X is aliased with STORE_PATTERN. */
6474 load_kills_store (x, store_pattern)
6475 rtx x, store_pattern;
6477 if (true_dependence (x, GET_MODE (x), store_pattern, rtx_addr_varies_p))
6482 /* Go through the entire insn X, looking for any loads which might alias
6483 STORE_PATTERN. Return 1 if found. */
6486 find_loads (x, store_pattern)
6487 rtx x, store_pattern;
6496 if (GET_CODE (x) == SET)
6499 if (GET_CODE (x) == MEM)
6501 if (load_kills_store (x, store_pattern))
6505 /* Recursively process the insn. */
6506 fmt = GET_RTX_FORMAT (GET_CODE (x));
6508 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0 && !ret; i--)
6511 ret |= find_loads (XEXP (x, i), store_pattern);
6512 else if (fmt[i] == 'E')
6513 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
6514 ret |= find_loads (XVECEXP (x, i, j), store_pattern);
6519 /* Check if INSN kills the store pattern X (is aliased with it).
6520 Return 1 if it it does. */
6523 store_killed_in_insn (x, insn)
6526 if (GET_RTX_CLASS (GET_CODE (insn)) != 'i')
6529 if (GET_CODE (insn) == CALL_INSN)
6531 if (CONST_OR_PURE_CALL_P (insn))
6537 if (GET_CODE (PATTERN (insn)) == SET)
6539 rtx pat = PATTERN (insn);
6540 /* Check for memory stores to aliased objects. */
6541 if (GET_CODE (SET_DEST (pat)) == MEM && !expr_equiv_p (SET_DEST (pat), x))
6542 /* pretend its a load and check for aliasing. */
6543 if (find_loads (SET_DEST (pat), x))
6545 return find_loads (SET_SRC (pat), x);
6548 return find_loads (PATTERN (insn), x);
6551 /* Returns 1 if the expression X is loaded or clobbered on or after INSN
6552 within basic block BB. */
6555 store_killed_after (x, insn, bb)
6564 /* Check if the register operands of the store are OK in this block.
6565 Note that if registers are changed ANYWHERE in the block, we'll
6566 decide we can't move it, regardless of whether it changed above
6567 or below the store. This could be improved by checking the register
6568 operands while lookinng for aliasing in each insn. */
6569 if (!store_ops_ok (XEXP (x, 0), bb))
6572 for ( ; insn && insn != NEXT_INSN (last); insn = NEXT_INSN (insn))
6573 if (store_killed_in_insn (x, insn))
6579 /* Returns 1 if the expression X is loaded or clobbered on or before INSN
6580 within basic block BB. */
6582 store_killed_before (x, insn, bb)
6586 rtx first = bb->head;
6589 return store_killed_in_insn (x, insn);
6591 /* Check if the register operands of the store are OK in this block.
6592 Note that if registers are changed ANYWHERE in the block, we'll
6593 decide we can't move it, regardless of whether it changed above
6594 or below the store. This could be improved by checking the register
6595 operands while lookinng for aliasing in each insn. */
6596 if (!store_ops_ok (XEXP (x, 0), bb))
6599 for ( ; insn && insn != PREV_INSN (first); insn = PREV_INSN (insn))
6600 if (store_killed_in_insn (x, insn))
6606 #define ANTIC_STORE_LIST(x) ((x)->loads)
6607 #define AVAIL_STORE_LIST(x) ((x)->stores)
6609 /* Given the table of available store insns at the end of blocks,
6610 determine which ones are not killed by aliasing, and generate
6611 the appropriate vectors for gen and killed. */
6613 build_store_vectors ()
6618 struct ls_expr * ptr;
6620 /* Build the gen_vector. This is any store in the table which is not killed
6621 by aliasing later in its block. */
6622 ae_gen = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks, num_stores);
6623 sbitmap_vector_zero (ae_gen, n_basic_blocks);
6625 st_antloc = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks, num_stores);
6626 sbitmap_vector_zero (st_antloc, n_basic_blocks);
6628 for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
6630 /* Put all the stores into either the antic list, or the avail list,
6632 rtx store_list = ptr->stores;
6633 ptr->stores = NULL_RTX;
6635 for (st = store_list; st != NULL; st = XEXP (st, 1))
6637 insn = XEXP (st, 0);
6638 bb = BLOCK_FOR_INSN (insn);
6640 if (!store_killed_after (ptr->pattern, insn, bb))
6642 /* If we've already seen an availale expression in this block,
6643 we can delete the one we saw already (It occurs earlier in
6644 the block), and replace it with this one). We'll copy the
6645 old SRC expression to an unused register in case there
6646 are any side effects. */
6647 if (TEST_BIT (ae_gen[bb->index], ptr->index))
6649 /* Find previous store. */
6651 for (st = AVAIL_STORE_LIST (ptr); st ; st = XEXP (st, 1))
6652 if (BLOCK_FOR_INSN (XEXP (st, 0)) == bb)
6656 rtx r = gen_reg_rtx (GET_MODE (ptr->pattern));
6658 fprintf(gcse_file, "Removing redundant store:\n");
6659 replace_store_insn (r, XEXP (st, 0), bb);
6660 XEXP (st, 0) = insn;
6664 SET_BIT (ae_gen[bb->index], ptr->index);
6665 AVAIL_STORE_LIST (ptr) = alloc_INSN_LIST (insn,
6666 AVAIL_STORE_LIST (ptr));
6669 if (!store_killed_before (ptr->pattern, insn, bb))
6671 SET_BIT (st_antloc[BLOCK_NUM (insn)], ptr->index);
6672 ANTIC_STORE_LIST (ptr) = alloc_INSN_LIST (insn,
6673 ANTIC_STORE_LIST (ptr));
6677 /* Free the original list of store insns. */
6678 free_INSN_LIST_list (&store_list);
6681 ae_kill = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks, num_stores);
6682 sbitmap_vector_zero (ae_kill, n_basic_blocks);
6684 transp = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks, num_stores);
6685 sbitmap_vector_zero (transp, n_basic_blocks);
6687 for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
6688 for (b = 0; b < n_basic_blocks; b++)
6690 if (store_killed_after (ptr->pattern, BLOCK_HEAD (b), BASIC_BLOCK (b)))
6692 /* The anticipatable expression is not killed if it's gen'd. */
6694 We leave this check out for now. If we have a code sequence
6695 in a block which looks like:
6699 We should flag this as having an ANTIC expression, NOT
6700 transparent, NOT killed, and AVAIL.
6701 Unfortunately, since we haven't re-written all loads to
6702 use the reaching reg, we'll end up doing an incorrect
6703 Load in the middle here if we push the store down. It happens in
6704 gcc.c-torture/execute/960311-1.c with -O3
6705 If we always kill it in this case, we'll sometimes do
6706 uneccessary work, but it shouldn't actually hurt anything.
6707 if (!TEST_BIT (ae_gen[b], ptr->index)). */
6708 SET_BIT (ae_kill[b], ptr->index);
6711 SET_BIT (transp[b], ptr->index);
6714 /* Any block with no exits calls some non-returning function, so
6715 we better mark the store killed here, or we might not store to
6716 it at all. If we knew it was abort, we wouldn't have to store,
6717 but we don't know that for sure. */
6720 fprintf (gcse_file, "ST_avail and ST_antic (shown under loads..)\n");
6721 print_ldst_list (gcse_file);
6722 dump_sbitmap_vector (gcse_file, "st_antloc", "", st_antloc, n_basic_blocks);
6723 dump_sbitmap_vector (gcse_file, "st_kill", "", ae_kill, n_basic_blocks);
6724 dump_sbitmap_vector (gcse_file, "Transpt", "", transp, n_basic_blocks);
6725 dump_sbitmap_vector (gcse_file, "st_avloc", "", ae_gen, n_basic_blocks);
6729 /* Insert an instruction at the begining of a basic block, and update
6730 the BLOCK_HEAD if needed. */
6733 insert_insn_start_bb (insn, bb)
6737 /* Insert at start of successor block. */
6738 rtx prev = PREV_INSN (bb->head);
6739 rtx before = bb->head;
6742 if (GET_CODE (before) != CODE_LABEL
6743 && (GET_CODE (before) != NOTE
6744 || NOTE_LINE_NUMBER (before) != NOTE_INSN_BASIC_BLOCK))
6747 if (prev == bb->end)
6749 before = NEXT_INSN (before);
6752 insn = emit_insn_after (insn, prev);
6754 if (prev == bb->end)
6757 set_block_for_new_insns (insn, bb);
6761 fprintf (gcse_file, "STORE_MOTION insert store at start of BB %d:\n",
6763 print_inline_rtx (gcse_file, insn, 6);
6764 fprintf (gcse_file, "\n");
6768 /* This routine will insert a store on an edge. EXPR is the ldst entry for
6769 the memory reference, and E is the edge to insert it on. Returns non-zero
6770 if an edge insertion was performed. */
6773 insert_store (expr, e)
6774 struct ls_expr * expr;
6781 /* We did all the deleted before this insert, so if we didn't delete a
6782 store, then we haven't set the reaching reg yet either. */
6783 if (expr->reaching_reg == NULL_RTX)
6786 reg = expr->reaching_reg;
6787 insn = gen_move_insn (expr->pattern, reg);
6789 /* If we are inserting this expression on ALL predecessor edges of a BB,
6790 insert it at the start of the BB, and reset the insert bits on the other
6791 edges so we don;t try to insert it on the other edges. */
6793 for (tmp = e->dest->pred; tmp ; tmp = tmp->pred_next)
6795 int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
6796 if (index == EDGE_INDEX_NO_EDGE)
6798 if (! TEST_BIT (pre_insert_map[index], expr->index))
6802 /* If tmp is NULL, we found an insertion on every edge, blank the
6803 insertion vector for these edges, and insert at the start of the BB. */
6804 if (!tmp && bb != EXIT_BLOCK_PTR)
6806 for (tmp = e->dest->pred; tmp ; tmp = tmp->pred_next)
6808 int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
6809 RESET_BIT (pre_insert_map[index], expr->index);
6811 insert_insn_start_bb (insn, bb);
6815 /* We can't insert on this edge, so we'll insert at the head of the
6816 successors block. See Morgan, sec 10.5. */
6817 if ((e->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
6819 insert_insn_start_bb (insn, bb);
6823 insert_insn_on_edge (insn, e);
6827 fprintf (gcse_file, "STORE_MOTION insert insn on edge (%d, %d):\n",
6828 e->src->index, e->dest->index);
6829 print_inline_rtx (gcse_file, insn, 6);
6830 fprintf (gcse_file, "\n");
6836 /* This routine will replace a store with a SET to a specified register. */
6839 replace_store_insn (reg, del, bb)
6845 insn = gen_move_insn (reg, SET_SRC (PATTERN (del)));
6846 insn = emit_insn_after (insn, del);
6847 set_block_for_new_insns (insn, bb);
6852 "STORE_MOTION delete insn in BB %d:\n ", bb->index);
6853 print_inline_rtx (gcse_file, del, 6);
6854 fprintf(gcse_file, "\nSTORE MOTION replaced with insn:\n ");
6855 print_inline_rtx (gcse_file, insn, 6);
6856 fprintf(gcse_file, "\n");
6862 if (bb->head == del)
6869 /* Delete a store, but copy the value that would have been stored into
6870 the reaching_reg for later storing. */
6873 delete_store (expr, bb)
6874 struct ls_expr * expr;
6879 if (expr->reaching_reg == NULL_RTX)
6880 expr->reaching_reg = gen_reg_rtx (GET_MODE (expr->pattern));
6883 /* If there is more than 1 store, the earlier ones will be dead,
6884 but it doesn't hurt to replace them here. */
6885 reg = expr->reaching_reg;
6887 for (i = AVAIL_STORE_LIST (expr); i; i = XEXP (i, 1))
6890 if (BLOCK_FOR_INSN (del) == bb)
6892 /* We know there is only one since we deleted redundant
6893 ones during the available computation. */
6894 replace_store_insn (reg, del, bb);
6900 /* Free memory used by store motion. */
6903 free_store_memory ()
6908 sbitmap_vector_free (ae_gen);
6910 sbitmap_vector_free (ae_kill);
6912 sbitmap_vector_free (transp);
6914 sbitmap_vector_free (st_antloc);
6916 sbitmap_vector_free (pre_insert_map);
6918 sbitmap_vector_free (pre_delete_map);
6919 if (reg_set_in_block)
6920 sbitmap_vector_free (reg_set_in_block);
6922 ae_gen = ae_kill = transp = st_antloc = NULL;
6923 pre_insert_map = pre_delete_map = reg_set_in_block = NULL;
6926 /* Perform store motion. Much like gcse, except we move expressions the
6927 other way by looking at the flowgraph in reverse. */
6933 struct ls_expr * ptr;
6934 int update_flow = 0;
6938 fprintf (gcse_file, "before store motion\n");
6939 print_rtl (gcse_file, get_insns ());
6943 init_alias_analysis ();
6945 /* Find all the stores that are live to the end of their block. */
6946 num_stores = compute_store_table ();
6947 if (num_stores == 0)
6949 sbitmap_vector_free (reg_set_in_block);
6950 end_alias_analysis ();
6954 /* Now compute whats actually available to move. */
6955 add_noreturn_fake_exit_edges ();
6956 build_store_vectors ();
6958 edge_list = pre_edge_rev_lcm (gcse_file, num_stores, transp, ae_gen,
6959 st_antloc, ae_kill, &pre_insert_map,
6962 /* Now we want to insert the new stores which are going to be needed. */
6963 for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
6965 for (x = 0; x < n_basic_blocks; x++)
6966 if (TEST_BIT (pre_delete_map[x], ptr->index))
6967 delete_store (ptr, BASIC_BLOCK (x));
6969 for (x = 0; x < NUM_EDGES (edge_list); x++)
6970 if (TEST_BIT (pre_insert_map[x], ptr->index))
6971 update_flow |= insert_store (ptr, INDEX_EDGE (edge_list, x));
6975 commit_edge_insertions ();
6977 free_store_memory ();
6978 free_edge_list (edge_list);
6979 remove_fake_edges ();
6980 end_alias_analysis ();