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 GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
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 to set it,
1302 or -1 if not set. */
1303 #define NEVER_SET -1
1304 static int *reg_first_set
;
1305 static int *reg_last_set
;
1308 /* See whether X, the source of a set, is something we want to consider for
1315 static rtx test_insn
= 0;
1316 int num_clobbers
= 0;
1319 switch (GET_CODE (x
))
1332 /* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
1333 if (general_operand (x
, GET_MODE (x
)))
1335 else if (GET_MODE (x
) == VOIDmode
)
1338 /* Otherwise, check if we can make a valid insn from it. First initialize
1339 our test insn if we haven't already. */
1343 = make_insn_raw (gen_rtx_SET (VOIDmode
,
1344 gen_rtx_REG (word_mode
,
1345 FIRST_PSEUDO_REGISTER
* 2),
1347 NEXT_INSN (test_insn
) = PREV_INSN (test_insn
) = 0;
1348 ggc_add_rtx_root (&test_insn
, 1);
1351 /* Now make an insn like the one we would make when GCSE'ing and see if
1353 PUT_MODE (SET_DEST (PATTERN (test_insn
)), GET_MODE (x
));
1354 SET_SRC (PATTERN (test_insn
)) = x
;
1355 return ((icode
= recog (PATTERN (test_insn
), test_insn
, &num_clobbers
)) >= 0
1356 && (num_clobbers
== 0 || ! added_clobbers_hard_reg_p (icode
)));
1359 /* Return non-zero if the operands of expression X are unchanged from the
1360 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1361 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1364 oprs_unchanged_p (x
, insn
, avail_p
)
1375 code
= GET_CODE (x
);
1380 return (reg_last_set
[REGNO (x
)] == NEVER_SET
1381 || reg_last_set
[REGNO (x
)] < INSN_CUID (insn
));
1383 return (reg_first_set
[REGNO (x
)] == NEVER_SET
1384 || reg_first_set
[REGNO (x
)] >= INSN_CUID (insn
));
1387 if (load_killed_in_block_p (BLOCK_FOR_INSN (insn
), INSN_CUID (insn
),
1391 return oprs_unchanged_p (XEXP (x
, 0), insn
, avail_p
);
1416 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
1420 /* If we are about to do the last recursive call needed at this
1421 level, change it into iteration. This function is called enough
1424 return oprs_unchanged_p (XEXP (x
, i
), insn
, avail_p
);
1426 else if (! oprs_unchanged_p (XEXP (x
, i
), insn
, avail_p
))
1429 else if (fmt
[i
] == 'E')
1430 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1431 if (! oprs_unchanged_p (XVECEXP (x
, i
, j
), insn
, avail_p
))
1438 /* Used for communication between mems_conflict_for_gcse_p and
1439 load_killed_in_block_p. Nonzero if mems_conflict_for_gcse_p finds a
1440 conflict between two memory references. */
1441 static int gcse_mems_conflict_p
;
1443 /* Used for communication between mems_conflict_for_gcse_p and
1444 load_killed_in_block_p. A memory reference for a load instruction,
1445 mems_conflict_for_gcse_p will see if a memory store conflicts with
1446 this memory load. */
1447 static rtx gcse_mem_operand
;
1449 /* DEST is the output of an instruction. If it is a memory reference, and
1450 possibly conflicts with the load found in gcse_mem_operand, then set
1451 gcse_mems_conflict_p to a nonzero value. */
1454 mems_conflict_for_gcse_p (dest
, setter
, data
)
1455 rtx dest
, setter ATTRIBUTE_UNUSED
;
1456 void *data ATTRIBUTE_UNUSED
;
1458 while (GET_CODE (dest
) == SUBREG
1459 || GET_CODE (dest
) == ZERO_EXTRACT
1460 || GET_CODE (dest
) == SIGN_EXTRACT
1461 || GET_CODE (dest
) == STRICT_LOW_PART
)
1462 dest
= XEXP (dest
, 0);
1464 /* If DEST is not a MEM, then it will not conflict with the load. Note
1465 that function calls are assumed to clobber memory, but are handled
1467 if (GET_CODE (dest
) != MEM
)
1470 /* If we are setting a MEM in our list of specially recognized MEMs,
1471 don't mark as killed this time. */
1473 if (dest
== gcse_mem_operand
&& pre_ldst_mems
!= NULL
)
1475 if (!find_rtx_in_ldst (dest
))
1476 gcse_mems_conflict_p
= 1;
1480 if (true_dependence (dest
, GET_MODE (dest
), gcse_mem_operand
,
1482 gcse_mems_conflict_p
= 1;
1485 /* Return nonzero if the expression in X (a memory reference) is killed
1486 in block BB before or after the insn with the CUID in UID_LIMIT.
1487 AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills
1490 To check the entire block, set UID_LIMIT to max_uid + 1 and
1494 load_killed_in_block_p (bb
, uid_limit
, x
, avail_p
)
1500 rtx list_entry
= modify_mem_list
[bb
->index
];
1504 /* Ignore entries in the list that do not apply. */
1506 && INSN_CUID (XEXP (list_entry
, 0)) < uid_limit
)
1508 && INSN_CUID (XEXP (list_entry
, 0)) > uid_limit
))
1510 list_entry
= XEXP (list_entry
, 1);
1514 setter
= XEXP (list_entry
, 0);
1516 /* If SETTER is a call everything is clobbered. Note that calls
1517 to pure functions are never put on the list, so we need not
1518 worry about them. */
1519 if (GET_CODE (setter
) == CALL_INSN
)
1522 /* SETTER must be an INSN of some kind that sets memory. Call
1523 note_stores to examine each hunk of memory that is modified.
1525 The note_stores interface is pretty limited, so we have to
1526 communicate via global variables. Yuk. */
1527 gcse_mem_operand
= x
;
1528 gcse_mems_conflict_p
= 0;
1529 note_stores (PATTERN (setter
), mems_conflict_for_gcse_p
, NULL
);
1530 if (gcse_mems_conflict_p
)
1532 list_entry
= XEXP (list_entry
, 1);
1537 /* Return non-zero if the operands of expression X are unchanged from
1538 the start of INSN's basic block up to but not including INSN. */
1541 oprs_anticipatable_p (x
, insn
)
1544 return oprs_unchanged_p (x
, insn
, 0);
1547 /* Return non-zero if the operands of expression X are unchanged from
1548 INSN to the end of INSN's basic block. */
1551 oprs_available_p (x
, insn
)
1554 return oprs_unchanged_p (x
, insn
, 1);
1557 /* Hash expression X.
1559 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1560 indicating if a volatile operand is found or if the expression contains
1561 something we don't want to insert in the table.
1563 ??? One might want to merge this with canon_hash. Later. */
1566 hash_expr (x
, mode
, do_not_record_p
, hash_table_size
)
1568 enum machine_mode mode
;
1569 int *do_not_record_p
;
1570 int hash_table_size
;
1574 *do_not_record_p
= 0;
1576 hash
= hash_expr_1 (x
, mode
, do_not_record_p
);
1577 return hash
% hash_table_size
;
1580 /* Hash a string. Just add its bytes up. */
1582 static inline unsigned
1587 const unsigned char *p
= (const unsigned char *)ps
;
1596 /* Subroutine of hash_expr to do the actual work. */
1599 hash_expr_1 (x
, mode
, do_not_record_p
)
1601 enum machine_mode mode
;
1602 int *do_not_record_p
;
1609 /* Used to turn recursion into iteration. We can't rely on GCC's
1610 tail-recursion eliminatio since we need to keep accumulating values
1617 code
= GET_CODE (x
);
1621 hash
+= ((unsigned int) REG
<< 7) + REGNO (x
);
1625 hash
+= (((unsigned int) CONST_INT
<< 7) + (unsigned int) mode
1626 + (unsigned int) INTVAL (x
));
1630 /* This is like the general case, except that it only counts
1631 the integers representing the constant. */
1632 hash
+= (unsigned int) code
+ (unsigned int) GET_MODE (x
);
1633 if (GET_MODE (x
) != VOIDmode
)
1634 for (i
= 2; i
< GET_RTX_LENGTH (CONST_DOUBLE
); i
++)
1635 hash
+= (unsigned int) XWINT (x
, i
);
1637 hash
+= ((unsigned int) CONST_DOUBLE_LOW (x
)
1638 + (unsigned int) CONST_DOUBLE_HIGH (x
));
1641 /* Assume there is only one rtx object for any given label. */
1643 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
1644 differences and differences between each stage's debugging dumps. */
1645 hash
+= (((unsigned int) LABEL_REF
<< 7)
1646 + CODE_LABEL_NUMBER (XEXP (x
, 0)));
1651 /* Don't hash on the symbol's address to avoid bootstrap differences.
1652 Different hash values may cause expressions to be recorded in
1653 different orders and thus different registers to be used in the
1654 final assembler. This also avoids differences in the dump files
1655 between various stages. */
1657 const unsigned char *p
= (const unsigned char *) XSTR (x
, 0);
1660 h
+= (h
<< 7) + *p
++; /* ??? revisit */
1662 hash
+= ((unsigned int) SYMBOL_REF
<< 7) + h
;
1667 if (MEM_VOLATILE_P (x
))
1669 *do_not_record_p
= 1;
1673 hash
+= (unsigned int) MEM
;
1674 hash
+= MEM_ALIAS_SET (x
);
1685 case UNSPEC_VOLATILE
:
1686 *do_not_record_p
= 1;
1690 if (MEM_VOLATILE_P (x
))
1692 *do_not_record_p
= 1;
1697 /* We don't want to take the filename and line into account. */
1698 hash
+= (unsigned) code
+ (unsigned) GET_MODE (x
)
1699 + hash_string_1 (ASM_OPERANDS_TEMPLATE (x
))
1700 + hash_string_1 (ASM_OPERANDS_OUTPUT_CONSTRAINT (x
))
1701 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x
);
1703 if (ASM_OPERANDS_INPUT_LENGTH (x
))
1705 for (i
= 1; i
< ASM_OPERANDS_INPUT_LENGTH (x
); i
++)
1707 hash
+= (hash_expr_1 (ASM_OPERANDS_INPUT (x
, i
),
1708 GET_MODE (ASM_OPERANDS_INPUT (x
, i
)),
1710 + hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT
1714 hash
+= hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT (x
, 0));
1715 x
= ASM_OPERANDS_INPUT (x
, 0);
1716 mode
= GET_MODE (x
);
1726 hash
+= (unsigned) code
+ (unsigned) GET_MODE (x
);
1727 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
1731 /* If we are about to do the last recursive call
1732 needed at this level, change it into iteration.
1733 This function is called enough to be worth it. */
1740 hash
+= hash_expr_1 (XEXP (x
, i
), 0, do_not_record_p
);
1741 if (*do_not_record_p
)
1745 else if (fmt
[i
] == 'E')
1746 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1748 hash
+= hash_expr_1 (XVECEXP (x
, i
, j
), 0, do_not_record_p
);
1749 if (*do_not_record_p
)
1753 else if (fmt
[i
] == 's')
1754 hash
+= hash_string_1 (XSTR (x
, i
));
1755 else if (fmt
[i
] == 'i')
1756 hash
+= (unsigned int) XINT (x
, i
);
1764 /* Hash a set of register REGNO.
1766 Sets are hashed on the register that is set. This simplifies the PRE copy
1769 ??? May need to make things more elaborate. Later, as necessary. */
1772 hash_set (regno
, hash_table_size
)
1774 int hash_table_size
;
1779 return hash
% hash_table_size
;
1782 /* Return non-zero if exp1 is equivalent to exp2.
1783 ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
1790 register enum rtx_code code
;
1791 register const char *fmt
;
1796 if (x
== 0 || y
== 0)
1799 code
= GET_CODE (x
);
1800 if (code
!= GET_CODE (y
))
1803 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
1804 if (GET_MODE (x
) != GET_MODE (y
))
1814 return INTVAL (x
) == INTVAL (y
);
1817 return XEXP (x
, 0) == XEXP (y
, 0);
1820 return XSTR (x
, 0) == XSTR (y
, 0);
1823 return REGNO (x
) == REGNO (y
);
1826 /* Can't merge two expressions in different alias sets, since we can
1827 decide that the expression is transparent in a block when it isn't,
1828 due to it being set with the different alias set. */
1829 if (MEM_ALIAS_SET (x
) != MEM_ALIAS_SET (y
))
1833 /* For commutative operations, check both orders. */
1841 return ((expr_equiv_p (XEXP (x
, 0), XEXP (y
, 0))
1842 && expr_equiv_p (XEXP (x
, 1), XEXP (y
, 1)))
1843 || (expr_equiv_p (XEXP (x
, 0), XEXP (y
, 1))
1844 && expr_equiv_p (XEXP (x
, 1), XEXP (y
, 0))));
1847 /* We don't use the generic code below because we want to
1848 disregard filename and line numbers. */
1850 /* A volatile asm isn't equivalent to any other. */
1851 if (MEM_VOLATILE_P (x
) || MEM_VOLATILE_P (y
))
1854 if (GET_MODE (x
) != GET_MODE (y
)
1855 || strcmp (ASM_OPERANDS_TEMPLATE (x
), ASM_OPERANDS_TEMPLATE (y
))
1856 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x
),
1857 ASM_OPERANDS_OUTPUT_CONSTRAINT (y
))
1858 || ASM_OPERANDS_OUTPUT_IDX (x
) != ASM_OPERANDS_OUTPUT_IDX (y
)
1859 || ASM_OPERANDS_INPUT_LENGTH (x
) != ASM_OPERANDS_INPUT_LENGTH (y
))
1862 if (ASM_OPERANDS_INPUT_LENGTH (x
))
1864 for (i
= ASM_OPERANDS_INPUT_LENGTH (x
) - 1; i
>= 0; i
--)
1865 if (! expr_equiv_p (ASM_OPERANDS_INPUT (x
, i
),
1866 ASM_OPERANDS_INPUT (y
, i
))
1867 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x
, i
),
1868 ASM_OPERANDS_INPUT_CONSTRAINT (y
, i
)))
1878 /* Compare the elements. If any pair of corresponding elements
1879 fail to match, return 0 for the whole thing. */
1881 fmt
= GET_RTX_FORMAT (code
);
1882 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
1887 if (! expr_equiv_p (XEXP (x
, i
), XEXP (y
, i
)))
1892 if (XVECLEN (x
, i
) != XVECLEN (y
, i
))
1894 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1895 if (! expr_equiv_p (XVECEXP (x
, i
, j
), XVECEXP (y
, i
, j
)))
1900 if (strcmp (XSTR (x
, i
), XSTR (y
, i
)))
1905 if (XINT (x
, i
) != XINT (y
, i
))
1910 if (XWINT (x
, i
) != XWINT (y
, i
))
1925 /* Insert expression X in INSN in the hash table.
1926 If it is already present, record it as the last occurrence in INSN's
1929 MODE is the mode of the value X is being stored into.
1930 It is only used if X is a CONST_INT.
1932 ANTIC_P is non-zero if X is an anticipatable expression.
1933 AVAIL_P is non-zero if X is an available expression. */
1936 insert_expr_in_table (x
, mode
, insn
, antic_p
, avail_p
)
1938 enum machine_mode mode
;
1940 int antic_p
, avail_p
;
1942 int found
, do_not_record_p
;
1944 struct expr
*cur_expr
, *last_expr
= NULL
;
1945 struct occr
*antic_occr
, *avail_occr
;
1946 struct occr
*last_occr
= NULL
;
1948 hash
= hash_expr (x
, mode
, &do_not_record_p
, expr_hash_table_size
);
1950 /* Do not insert expression in table if it contains volatile operands,
1951 or if hash_expr determines the expression is something we don't want
1952 to or can't handle. */
1953 if (do_not_record_p
)
1956 cur_expr
= expr_hash_table
[hash
];
1959 while (cur_expr
&& 0 == (found
= expr_equiv_p (cur_expr
->expr
, x
)))
1961 /* If the expression isn't found, save a pointer to the end of
1963 last_expr
= cur_expr
;
1964 cur_expr
= cur_expr
->next_same_hash
;
1969 cur_expr
= (struct expr
*) gcse_alloc (sizeof (struct expr
));
1970 bytes_used
+= sizeof (struct expr
);
1971 if (expr_hash_table
[hash
] == NULL
)
1972 /* This is the first pattern that hashed to this index. */
1973 expr_hash_table
[hash
] = cur_expr
;
1975 /* Add EXPR to end of this hash chain. */
1976 last_expr
->next_same_hash
= cur_expr
;
1978 /* Set the fields of the expr element. */
1980 cur_expr
->bitmap_index
= n_exprs
++;
1981 cur_expr
->next_same_hash
= NULL
;
1982 cur_expr
->antic_occr
= NULL
;
1983 cur_expr
->avail_occr
= NULL
;
1986 /* Now record the occurrence(s). */
1989 antic_occr
= cur_expr
->antic_occr
;
1991 /* Search for another occurrence in the same basic block. */
1992 while (antic_occr
&& BLOCK_NUM (antic_occr
->insn
) != BLOCK_NUM (insn
))
1994 /* If an occurrence isn't found, save a pointer to the end of
1996 last_occr
= antic_occr
;
1997 antic_occr
= antic_occr
->next
;
2001 /* Found another instance of the expression in the same basic block.
2002 Prefer the currently recorded one. We want the first one in the
2003 block and the block is scanned from start to end. */
2004 ; /* nothing to do */
2007 /* First occurrence of this expression in this basic block. */
2008 antic_occr
= (struct occr
*) gcse_alloc (sizeof (struct occr
));
2009 bytes_used
+= sizeof (struct occr
);
2010 /* First occurrence of this expression in any block? */
2011 if (cur_expr
->antic_occr
== NULL
)
2012 cur_expr
->antic_occr
= antic_occr
;
2014 last_occr
->next
= antic_occr
;
2016 antic_occr
->insn
= insn
;
2017 antic_occr
->next
= NULL
;
2023 avail_occr
= cur_expr
->avail_occr
;
2025 /* Search for another occurrence in the same basic block. */
2026 while (avail_occr
&& BLOCK_NUM (avail_occr
->insn
) != BLOCK_NUM (insn
))
2028 /* If an occurrence isn't found, save a pointer to the end of
2030 last_occr
= avail_occr
;
2031 avail_occr
= avail_occr
->next
;
2035 /* Found another instance of the expression in the same basic block.
2036 Prefer this occurrence to the currently recorded one. We want
2037 the last one in the block and the block is scanned from start
2039 avail_occr
->insn
= insn
;
2042 /* First occurrence of this expression in this basic block. */
2043 avail_occr
= (struct occr
*) gcse_alloc (sizeof (struct occr
));
2044 bytes_used
+= sizeof (struct occr
);
2046 /* First occurrence of this expression in any block? */
2047 if (cur_expr
->avail_occr
== NULL
)
2048 cur_expr
->avail_occr
= avail_occr
;
2050 last_occr
->next
= avail_occr
;
2052 avail_occr
->insn
= insn
;
2053 avail_occr
->next
= NULL
;
2058 /* Insert pattern X in INSN in the hash table.
2059 X is a SET of a reg to either another reg or a constant.
2060 If it is already present, record it as the last occurrence in INSN's
2064 insert_set_in_table (x
, insn
)
2070 struct expr
*cur_expr
, *last_expr
= NULL
;
2071 struct occr
*cur_occr
, *last_occr
= NULL
;
2073 if (GET_CODE (x
) != SET
2074 || GET_CODE (SET_DEST (x
)) != REG
)
2077 hash
= hash_set (REGNO (SET_DEST (x
)), set_hash_table_size
);
2079 cur_expr
= set_hash_table
[hash
];
2082 while (cur_expr
&& 0 == (found
= expr_equiv_p (cur_expr
->expr
, x
)))
2084 /* If the expression isn't found, save a pointer to the end of
2086 last_expr
= cur_expr
;
2087 cur_expr
= cur_expr
->next_same_hash
;
2092 cur_expr
= (struct expr
*) gcse_alloc (sizeof (struct expr
));
2093 bytes_used
+= sizeof (struct expr
);
2094 if (set_hash_table
[hash
] == NULL
)
2095 /* This is the first pattern that hashed to this index. */
2096 set_hash_table
[hash
] = cur_expr
;
2098 /* Add EXPR to end of this hash chain. */
2099 last_expr
->next_same_hash
= cur_expr
;
2101 /* Set the fields of the expr element.
2102 We must copy X because it can be modified when copy propagation is
2103 performed on its operands. */
2104 cur_expr
->expr
= copy_rtx (x
);
2105 cur_expr
->bitmap_index
= n_sets
++;
2106 cur_expr
->next_same_hash
= NULL
;
2107 cur_expr
->antic_occr
= NULL
;
2108 cur_expr
->avail_occr
= NULL
;
2111 /* Now record the occurrence. */
2112 cur_occr
= cur_expr
->avail_occr
;
2114 /* Search for another occurrence in the same basic block. */
2115 while (cur_occr
&& BLOCK_NUM (cur_occr
->insn
) != BLOCK_NUM (insn
))
2117 /* If an occurrence isn't found, save a pointer to the end of
2119 last_occr
= cur_occr
;
2120 cur_occr
= cur_occr
->next
;
2124 /* Found another instance of the expression in the same basic block.
2125 Prefer this occurrence to the currently recorded one. We want the
2126 last one in the block and the block is scanned from start to end. */
2127 cur_occr
->insn
= insn
;
2130 /* First occurrence of this expression in this basic block. */
2131 cur_occr
= (struct occr
*) gcse_alloc (sizeof (struct occr
));
2132 bytes_used
+= sizeof (struct occr
);
2134 /* First occurrence of this expression in any block? */
2135 if (cur_expr
->avail_occr
== NULL
)
2136 cur_expr
->avail_occr
= cur_occr
;
2138 last_occr
->next
= cur_occr
;
2140 cur_occr
->insn
= insn
;
2141 cur_occr
->next
= NULL
;
2145 /* Scan pattern PAT of INSN and add an entry to the hash table. If SET_P is
2146 non-zero, this is for the assignment hash table, otherwise it is for the
2147 expression hash table. */
2150 hash_scan_set (pat
, insn
, set_p
)
2154 rtx src
= SET_SRC (pat
);
2155 rtx dest
= SET_DEST (pat
);
2158 if (GET_CODE (src
) == CALL
)
2159 hash_scan_call (src
, insn
);
2161 else if (GET_CODE (dest
) == REG
)
2163 unsigned int regno
= REGNO (dest
);
2166 /* If this is a single set and we are doing constant propagation,
2167 see if a REG_NOTE shows this equivalent to a constant. */
2168 if (set_p
&& (note
= find_reg_equal_equiv_note (insn
)) != 0
2169 && CONSTANT_P (XEXP (note
, 0)))
2170 src
= XEXP (note
, 0), pat
= gen_rtx_SET (VOIDmode
, dest
, src
);
2172 /* Only record sets of pseudo-regs in the hash table. */
2174 && regno
>= FIRST_PSEUDO_REGISTER
2175 /* Don't GCSE something if we can't do a reg/reg copy. */
2176 && can_copy_p
[GET_MODE (dest
)]
2177 /* Is SET_SRC something we want to gcse? */
2178 && want_to_gcse_p (src
)
2179 /* Don't CSE a nop. */
2180 && ! set_noop_p (pat
)
2181 /* Don't GCSE if it has attached REG_EQUIV note.
2182 At this point this only function parameters should have
2183 REG_EQUIV notes and if the argument slot is used somewhere
2184 explicitely, it means address of parameter has been taken,
2185 so we should not extend the lifetime of the pseudo. */
2186 && ((note
= find_reg_note (insn
, REG_EQUIV
, NULL_RTX
)) == 0
2187 || GET_CODE (XEXP (note
, 0)) != MEM
))
2189 /* An expression is not anticipatable if its operands are
2190 modified before this insn or if this is not the only SET in
2192 int antic_p
= oprs_anticipatable_p (src
, insn
) && single_set (insn
);
2193 /* An expression is not available if its operands are
2194 subsequently modified, including this insn. */
2195 int avail_p
= oprs_available_p (src
, insn
);
2197 insert_expr_in_table (src
, GET_MODE (dest
), insn
, antic_p
, avail_p
);
2200 /* Record sets for constant/copy propagation. */
2202 && regno
>= FIRST_PSEUDO_REGISTER
2203 && ((GET_CODE (src
) == REG
2204 && REGNO (src
) >= FIRST_PSEUDO_REGISTER
2205 && can_copy_p
[GET_MODE (dest
)]
2206 && REGNO (src
) != regno
)
2207 || GET_CODE (src
) == CONST_INT
2208 || GET_CODE (src
) == SYMBOL_REF
2209 || GET_CODE (src
) == CONST_DOUBLE
)
2210 /* A copy is not available if its src or dest is subsequently
2211 modified. Here we want to search from INSN+1 on, but
2212 oprs_available_p searches from INSN on. */
2213 && (insn
== BLOCK_END (BLOCK_NUM (insn
))
2214 || ((tmp
= next_nonnote_insn (insn
)) != NULL_RTX
2215 && oprs_available_p (pat
, tmp
))))
2216 insert_set_in_table (pat
, insn
);
2221 hash_scan_clobber (x
, insn
)
2222 rtx x ATTRIBUTE_UNUSED
, insn ATTRIBUTE_UNUSED
;
2224 /* Currently nothing to do. */
2228 hash_scan_call (x
, insn
)
2229 rtx x ATTRIBUTE_UNUSED
, insn ATTRIBUTE_UNUSED
;
2231 /* Currently nothing to do. */
2234 /* Process INSN and add hash table entries as appropriate.
2236 Only available expressions that set a single pseudo-reg are recorded.
2238 Single sets in a PARALLEL could be handled, but it's an extra complication
2239 that isn't dealt with right now. The trick is handling the CLOBBERs that
2240 are also in the PARALLEL. Later.
2242 If SET_P is non-zero, this is for the assignment hash table,
2243 otherwise it is for the expression hash table.
2244 If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
2245 not record any expressions. */
2248 hash_scan_insn (insn
, set_p
, in_libcall_block
)
2251 int in_libcall_block
;
2253 rtx pat
= PATTERN (insn
);
2256 if (in_libcall_block
)
2259 /* Pick out the sets of INSN and for other forms of instructions record
2260 what's been modified. */
2262 if (GET_CODE (pat
) == SET
)
2263 hash_scan_set (pat
, insn
, set_p
);
2264 else if (GET_CODE (pat
) == PARALLEL
)
2265 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
2267 rtx x
= XVECEXP (pat
, 0, i
);
2269 if (GET_CODE (x
) == SET
)
2270 hash_scan_set (x
, insn
, set_p
);
2271 else if (GET_CODE (x
) == CLOBBER
)
2272 hash_scan_clobber (x
, insn
);
2273 else if (GET_CODE (x
) == CALL
)
2274 hash_scan_call (x
, insn
);
2277 else if (GET_CODE (pat
) == CLOBBER
)
2278 hash_scan_clobber (pat
, insn
);
2279 else if (GET_CODE (pat
) == CALL
)
2280 hash_scan_call (pat
, insn
);
2284 dump_hash_table (file
, name
, table
, table_size
, total_size
)
2287 struct expr
**table
;
2288 int table_size
, total_size
;
2291 /* Flattened out table, so it's printed in proper order. */
2292 struct expr
**flat_table
;
2293 unsigned int *hash_val
;
2297 = (struct expr
**) xcalloc (total_size
, sizeof (struct expr
*));
2298 hash_val
= (unsigned int *) xmalloc (total_size
* sizeof (unsigned int));
2300 for (i
= 0; i
< table_size
; i
++)
2301 for (expr
= table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
2303 flat_table
[expr
->bitmap_index
] = expr
;
2304 hash_val
[expr
->bitmap_index
] = i
;
2307 fprintf (file
, "%s hash table (%d buckets, %d entries)\n",
2308 name
, table_size
, total_size
);
2310 for (i
= 0; i
< total_size
; i
++)
2311 if (flat_table
[i
] != 0)
2313 expr
= flat_table
[i
];
2314 fprintf (file
, "Index %d (hash value %d)\n ",
2315 expr
->bitmap_index
, hash_val
[i
]);
2316 print_rtl (file
, expr
->expr
);
2317 fprintf (file
, "\n");
2320 fprintf (file
, "\n");
2326 /* Record register first/last/block set information for REGNO in INSN.
2328 reg_first_set records the first place in the block where the register
2329 is set and is used to compute "anticipatability".
2331 reg_last_set records the last place in the block where the register
2332 is set and is used to compute "availability".
2334 reg_set_in_block records whether the register is set in the block
2335 and is used to compute "transparency". */
2338 record_last_reg_set_info (insn
, regno
)
2342 if (reg_first_set
[regno
] == NEVER_SET
)
2343 reg_first_set
[regno
] = INSN_CUID (insn
);
2345 reg_last_set
[regno
] = INSN_CUID (insn
);
2346 SET_BIT (reg_set_in_block
[BLOCK_NUM (insn
)], regno
);
2350 /* Record all of the canonicalized MEMs of record_last_mem_set_info's insn.
2351 Note we store a pair of elements in the list, so they have to be
2352 taken off pairwise. */
2355 canon_list_insert (dest
, unused1
, v_insn
)
2356 rtx dest ATTRIBUTE_UNUSED
;
2357 rtx unused1 ATTRIBUTE_UNUSED
;
2360 rtx dest_addr
, insn
;
2362 while (GET_CODE (dest
) == SUBREG
2363 || GET_CODE (dest
) == ZERO_EXTRACT
2364 || GET_CODE (dest
) == SIGN_EXTRACT
2365 || GET_CODE (dest
) == STRICT_LOW_PART
)
2366 dest
= XEXP (dest
, 0);
2368 /* If DEST is not a MEM, then it will not conflict with a load. Note
2369 that function calls are assumed to clobber memory, but are handled
2372 if (GET_CODE (dest
) != MEM
)
2375 dest_addr
= get_addr (XEXP (dest
, 0));
2376 dest_addr
= canon_rtx (dest_addr
);
2377 insn
= (rtx
) v_insn
;
2379 canon_modify_mem_list
[BLOCK_NUM (insn
)] =
2380 alloc_INSN_LIST (dest_addr
, canon_modify_mem_list
[BLOCK_NUM (insn
)]);
2381 canon_modify_mem_list
[BLOCK_NUM (insn
)] =
2382 alloc_INSN_LIST (dest
, canon_modify_mem_list
[BLOCK_NUM (insn
)]);
2385 /* Record memory modification information for INSN. We do not actually care
2386 about the memory location(s) that are set, or even how they are set (consider
2387 a CALL_INSN). We merely need to record which insns modify memory. */
2390 record_last_mem_set_info (insn
)
2393 /* load_killed_in_block_p will handle the case of calls clobbering
2395 modify_mem_list
[BLOCK_NUM (insn
)] =
2396 alloc_INSN_LIST (insn
, modify_mem_list
[BLOCK_NUM (insn
)]);
2398 if (GET_CODE (insn
) == CALL_INSN
)
2400 /* Note that traversals of this loop (other than for free-ing)
2401 will break after encountering a CALL_INSN. So, there's no
2402 need to insert a pair of items, as canon_list_insert does. */
2403 canon_modify_mem_list
[BLOCK_NUM (insn
)] =
2404 alloc_INSN_LIST (insn
, canon_modify_mem_list
[BLOCK_NUM (insn
)]);
2407 note_stores (PATTERN (insn
), canon_list_insert
, (void*)insn
);
2410 /* Called from compute_hash_table via note_stores to handle one
2411 SET or CLOBBER in an insn. DATA is really the instruction in which
2412 the SET is taking place. */
2415 record_last_set_info (dest
, setter
, data
)
2416 rtx dest
, setter ATTRIBUTE_UNUSED
;
2419 rtx last_set_insn
= (rtx
) data
;
2421 if (GET_CODE (dest
) == SUBREG
)
2422 dest
= SUBREG_REG (dest
);
2424 if (GET_CODE (dest
) == REG
)
2425 record_last_reg_set_info (last_set_insn
, REGNO (dest
));
2426 else if (GET_CODE (dest
) == MEM
2427 /* Ignore pushes, they clobber nothing. */
2428 && ! push_operand (dest
, GET_MODE (dest
)))
2429 record_last_mem_set_info (last_set_insn
);
2432 /* Top level function to create an expression or assignment hash table.
2434 Expression entries are placed in the hash table if
2435 - they are of the form (set (pseudo-reg) src),
2436 - src is something we want to perform GCSE on,
2437 - none of the operands are subsequently modified in the block
2439 Assignment entries are placed in the hash table if
2440 - they are of the form (set (pseudo-reg) src),
2441 - src is something we want to perform const/copy propagation on,
2442 - none of the operands or target are subsequently modified in the block
2444 Currently src must be a pseudo-reg or a const_int.
2446 F is the first insn.
2447 SET_P is non-zero for computing the assignment hash table. */
2450 compute_hash_table (set_p
)
2455 /* While we compute the hash table we also compute a bit array of which
2456 registers are set in which blocks.
2457 ??? This isn't needed during const/copy propagation, but it's cheap to
2459 sbitmap_vector_zero (reg_set_in_block
, n_basic_blocks
);
2461 /* re-Cache any INSN_LIST nodes we have allocated. */
2464 for (i
= 0; i
< n_basic_blocks
; i
++)
2466 if (modify_mem_list
[i
])
2467 free_INSN_LIST_list (modify_mem_list
+ i
);
2468 if (canon_modify_mem_list
[i
])
2469 free_INSN_LIST_list (canon_modify_mem_list
+ i
);
2472 /* Some working arrays used to track first and last set in each block. */
2473 /* ??? One could use alloca here, but at some size a threshold is crossed
2474 beyond which one should use malloc. Are we at that threshold here? */
2475 reg_first_set
= (int *) gmalloc (max_gcse_regno
* sizeof (int));
2476 reg_last_set
= (int *) gmalloc (max_gcse_regno
* sizeof (int));
2478 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
2482 int in_libcall_block
;
2485 /* First pass over the instructions records information used to
2486 determine when registers and memory are first and last set.
2487 ??? hard-reg reg_set_in_block computation
2488 could be moved to compute_sets since they currently don't change. */
2490 for (i
= 0; i
< max_gcse_regno
; i
++)
2491 reg_first_set
[i
] = reg_last_set
[i
] = NEVER_SET
;
2494 for (insn
= BLOCK_HEAD (bb
);
2495 insn
&& insn
!= NEXT_INSN (BLOCK_END (bb
));
2496 insn
= NEXT_INSN (insn
))
2498 if (! INSN_P (insn
))
2501 if (GET_CODE (insn
) == CALL_INSN
)
2503 bool clobbers_all
= false;
2504 #ifdef NON_SAVING_SETJMP
2505 if (NON_SAVING_SETJMP
2506 && find_reg_note (insn
, REG_SETJMP
, NULL_RTX
))
2507 clobbers_all
= true;
2510 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
2512 || TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
2513 record_last_reg_set_info (insn
, regno
);
2518 note_stores (PATTERN (insn
), record_last_set_info
, insn
);
2521 /* The next pass builds the hash table. */
2523 for (insn
= BLOCK_HEAD (bb
), in_libcall_block
= 0;
2524 insn
&& insn
!= NEXT_INSN (BLOCK_END (bb
));
2525 insn
= NEXT_INSN (insn
))
2528 if (find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
))
2529 in_libcall_block
= 1;
2530 else if (find_reg_note (insn
, REG_RETVAL
, NULL_RTX
))
2531 in_libcall_block
= 0;
2532 hash_scan_insn (insn
, set_p
, in_libcall_block
);
2536 free (reg_first_set
);
2537 free (reg_last_set
);
2539 /* Catch bugs early. */
2540 reg_first_set
= reg_last_set
= 0;
2543 /* Allocate space for the set hash table.
2544 N_INSNS is the number of instructions in the function.
2545 It is used to determine the number of buckets to use. */
2548 alloc_set_hash_table (n_insns
)
2553 set_hash_table_size
= n_insns
/ 4;
2554 if (set_hash_table_size
< 11)
2555 set_hash_table_size
= 11;
2557 /* Attempt to maintain efficient use of hash table.
2558 Making it an odd number is simplest for now.
2559 ??? Later take some measurements. */
2560 set_hash_table_size
|= 1;
2561 n
= set_hash_table_size
* sizeof (struct expr
*);
2562 set_hash_table
= (struct expr
**) gmalloc (n
);
2565 /* Free things allocated by alloc_set_hash_table. */
2568 free_set_hash_table ()
2570 free (set_hash_table
);
2573 /* Compute the hash table for doing copy/const propagation. */
2576 compute_set_hash_table ()
2578 /* Initialize count of number of entries in hash table. */
2580 memset ((char *) set_hash_table
, 0,
2581 set_hash_table_size
* sizeof (struct expr
*));
2583 compute_hash_table (1);
2586 /* Allocate space for the expression hash table.
2587 N_INSNS is the number of instructions in the function.
2588 It is used to determine the number of buckets to use. */
2591 alloc_expr_hash_table (n_insns
)
2592 unsigned int n_insns
;
2596 expr_hash_table_size
= n_insns
/ 2;
2597 /* Make sure the amount is usable. */
2598 if (expr_hash_table_size
< 11)
2599 expr_hash_table_size
= 11;
2601 /* Attempt to maintain efficient use of hash table.
2602 Making it an odd number is simplest for now.
2603 ??? Later take some measurements. */
2604 expr_hash_table_size
|= 1;
2605 n
= expr_hash_table_size
* sizeof (struct expr
*);
2606 expr_hash_table
= (struct expr
**) gmalloc (n
);
2609 /* Free things allocated by alloc_expr_hash_table. */
2612 free_expr_hash_table ()
2614 free (expr_hash_table
);
2617 /* Compute the hash table for doing GCSE. */
2620 compute_expr_hash_table ()
2622 /* Initialize count of number of entries in hash table. */
2624 memset ((char *) expr_hash_table
, 0,
2625 expr_hash_table_size
* sizeof (struct expr
*));
2627 compute_hash_table (0);
2630 /* Expression tracking support. */
2632 /* Lookup pattern PAT in the expression table.
2633 The result is a pointer to the table entry, or NULL if not found. */
2635 static struct expr
*
2639 int do_not_record_p
;
2640 unsigned int hash
= hash_expr (pat
, GET_MODE (pat
), &do_not_record_p
,
2641 expr_hash_table_size
);
2644 if (do_not_record_p
)
2647 expr
= expr_hash_table
[hash
];
2649 while (expr
&& ! expr_equiv_p (expr
->expr
, pat
))
2650 expr
= expr
->next_same_hash
;
2655 /* Lookup REGNO in the set table. If PAT is non-NULL look for the entry that
2656 matches it, otherwise return the first entry for REGNO. The result is a
2657 pointer to the table entry, or NULL if not found. */
2659 static struct expr
*
2660 lookup_set (regno
, pat
)
2664 unsigned int hash
= hash_set (regno
, set_hash_table_size
);
2667 expr
= set_hash_table
[hash
];
2671 while (expr
&& ! expr_equiv_p (expr
->expr
, pat
))
2672 expr
= expr
->next_same_hash
;
2676 while (expr
&& REGNO (SET_DEST (expr
->expr
)) != regno
)
2677 expr
= expr
->next_same_hash
;
2683 /* Return the next entry for REGNO in list EXPR. */
2685 static struct expr
*
2686 next_set (regno
, expr
)
2691 expr
= expr
->next_same_hash
;
2692 while (expr
&& REGNO (SET_DEST (expr
->expr
)) != regno
);
2697 /* Reset tables used to keep track of what's still available [since the
2698 start of the block]. */
2701 reset_opr_set_tables ()
2703 /* Maintain a bitmap of which regs have been set since beginning of
2705 sbitmap_zero (reg_set_bitmap
);
2707 /* Also keep a record of the last instruction to modify memory.
2708 For now this is very trivial, we only record whether any memory
2709 location has been modified. */
2713 /* re-Cache any INSN_LIST nodes we have allocated. */
2714 for (i
= 0; i
< n_basic_blocks
; i
++)
2716 if (modify_mem_list
[i
])
2717 free_INSN_LIST_list (modify_mem_list
+ i
);
2718 if (canon_modify_mem_list
[i
])
2719 free_INSN_LIST_list (canon_modify_mem_list
+ i
);
2724 /* Return non-zero if the operands of X are not set before INSN in
2725 INSN's basic block. */
2728 oprs_not_set_p (x
, insn
)
2738 code
= GET_CODE (x
);
2753 if (load_killed_in_block_p (BLOCK_FOR_INSN (insn
),
2754 INSN_CUID (insn
), x
, 0))
2757 return oprs_not_set_p (XEXP (x
, 0), insn
);
2760 return ! TEST_BIT (reg_set_bitmap
, REGNO (x
));
2766 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
2770 /* If we are about to do the last recursive call
2771 needed at this level, change it into iteration.
2772 This function is called enough to be worth it. */
2774 return oprs_not_set_p (XEXP (x
, i
), insn
);
2776 if (! oprs_not_set_p (XEXP (x
, i
), insn
))
2779 else if (fmt
[i
] == 'E')
2780 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
2781 if (! oprs_not_set_p (XVECEXP (x
, i
, j
), insn
))
2788 /* Mark things set by a CALL. */
2794 if (! CONST_OR_PURE_CALL_P (insn
))
2795 record_last_mem_set_info (insn
);
2798 /* Mark things set by a SET. */
2801 mark_set (pat
, insn
)
2804 rtx dest
= SET_DEST (pat
);
2806 while (GET_CODE (dest
) == SUBREG
2807 || GET_CODE (dest
) == ZERO_EXTRACT
2808 || GET_CODE (dest
) == SIGN_EXTRACT
2809 || GET_CODE (dest
) == STRICT_LOW_PART
)
2810 dest
= XEXP (dest
, 0);
2812 if (GET_CODE (dest
) == REG
)
2813 SET_BIT (reg_set_bitmap
, REGNO (dest
));
2814 else if (GET_CODE (dest
) == MEM
)
2815 record_last_mem_set_info (insn
);
2817 if (GET_CODE (SET_SRC (pat
)) == CALL
)
2821 /* Record things set by a CLOBBER. */
2824 mark_clobber (pat
, insn
)
2827 rtx clob
= XEXP (pat
, 0);
2829 while (GET_CODE (clob
) == SUBREG
|| GET_CODE (clob
) == STRICT_LOW_PART
)
2830 clob
= XEXP (clob
, 0);
2832 if (GET_CODE (clob
) == REG
)
2833 SET_BIT (reg_set_bitmap
, REGNO (clob
));
2835 record_last_mem_set_info (insn
);
2838 /* Record things set by INSN.
2839 This data is used by oprs_not_set_p. */
2842 mark_oprs_set (insn
)
2845 rtx pat
= PATTERN (insn
);
2848 if (GET_CODE (pat
) == SET
)
2849 mark_set (pat
, insn
);
2850 else if (GET_CODE (pat
) == PARALLEL
)
2851 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
2853 rtx x
= XVECEXP (pat
, 0, i
);
2855 if (GET_CODE (x
) == SET
)
2857 else if (GET_CODE (x
) == CLOBBER
)
2858 mark_clobber (x
, insn
);
2859 else if (GET_CODE (x
) == CALL
)
2863 else if (GET_CODE (pat
) == CLOBBER
)
2864 mark_clobber (pat
, insn
);
2865 else if (GET_CODE (pat
) == CALL
)
2870 /* Classic GCSE reaching definition support. */
2872 /* Allocate reaching def variables. */
2875 alloc_rd_mem (n_blocks
, n_insns
)
2876 int n_blocks
, n_insns
;
2878 rd_kill
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_insns
);
2879 sbitmap_vector_zero (rd_kill
, n_basic_blocks
);
2881 rd_gen
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_insns
);
2882 sbitmap_vector_zero (rd_gen
, n_basic_blocks
);
2884 reaching_defs
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_insns
);
2885 sbitmap_vector_zero (reaching_defs
, n_basic_blocks
);
2887 rd_out
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_insns
);
2888 sbitmap_vector_zero (rd_out
, n_basic_blocks
);
2891 /* Free reaching def variables. */
2896 sbitmap_vector_free (rd_kill
);
2897 sbitmap_vector_free (rd_gen
);
2898 sbitmap_vector_free (reaching_defs
);
2899 sbitmap_vector_free (rd_out
);
2902 /* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
2905 handle_rd_kill_set (insn
, regno
, bb
)
2910 struct reg_set
*this_reg
;
2912 for (this_reg
= reg_set_table
[regno
]; this_reg
; this_reg
= this_reg
->next
)
2913 if (BLOCK_NUM (this_reg
->insn
) != BLOCK_NUM (insn
))
2914 SET_BIT (rd_kill
[bb
->index
], INSN_CUID (this_reg
->insn
));
2917 /* Compute the set of kill's for reaching definitions. */
2927 For each set bit in `gen' of the block (i.e each insn which
2928 generates a definition in the block)
2929 Call the reg set by the insn corresponding to that bit regx
2930 Look at the linked list starting at reg_set_table[regx]
2931 For each setting of regx in the linked list, which is not in
2933 Set the bit in `kill' corresponding to that insn. */
2934 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
2935 for (cuid
= 0; cuid
< max_cuid
; cuid
++)
2936 if (TEST_BIT (rd_gen
[bb
], cuid
))
2938 rtx insn
= CUID_INSN (cuid
);
2939 rtx pat
= PATTERN (insn
);
2941 if (GET_CODE (insn
) == CALL_INSN
)
2943 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
2944 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
2945 handle_rd_kill_set (insn
, regno
, BASIC_BLOCK (bb
));
2948 if (GET_CODE (pat
) == PARALLEL
)
2950 for (i
= XVECLEN (pat
, 0) - 1; i
>= 0; i
--)
2952 enum rtx_code code
= GET_CODE (XVECEXP (pat
, 0, i
));
2954 if ((code
== SET
|| code
== CLOBBER
)
2955 && GET_CODE (XEXP (XVECEXP (pat
, 0, i
), 0)) == REG
)
2956 handle_rd_kill_set (insn
,
2957 REGNO (XEXP (XVECEXP (pat
, 0, i
), 0)),
2961 else if (GET_CODE (pat
) == SET
&& GET_CODE (SET_DEST (pat
)) == REG
)
2962 /* Each setting of this register outside of this block
2963 must be marked in the set of kills in this block. */
2964 handle_rd_kill_set (insn
, REGNO (SET_DEST (pat
)), BASIC_BLOCK (bb
));
2968 /* Compute the reaching definitions as in
2969 Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
2970 Chapter 10. It is the same algorithm as used for computing available
2971 expressions but applied to the gens and kills of reaching definitions. */
2976 int bb
, changed
, passes
;
2978 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
2979 sbitmap_copy (rd_out
[bb
] /*dst*/, rd_gen
[bb
] /*src*/);
2986 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
2988 sbitmap_union_of_preds (reaching_defs
[bb
], rd_out
, bb
);
2989 changed
|= sbitmap_union_of_diff (rd_out
[bb
], rd_gen
[bb
],
2990 reaching_defs
[bb
], rd_kill
[bb
]);
2996 fprintf (gcse_file
, "reaching def computation: %d passes\n", passes
);
2999 /* Classic GCSE available expression support. */
3001 /* Allocate memory for available expression computation. */
3004 alloc_avail_expr_mem (n_blocks
, n_exprs
)
3005 int n_blocks
, n_exprs
;
3007 ae_kill
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_exprs
);
3008 sbitmap_vector_zero (ae_kill
, n_basic_blocks
);
3010 ae_gen
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_exprs
);
3011 sbitmap_vector_zero (ae_gen
, n_basic_blocks
);
3013 ae_in
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_exprs
);
3014 sbitmap_vector_zero (ae_in
, n_basic_blocks
);
3016 ae_out
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_exprs
);
3017 sbitmap_vector_zero (ae_out
, n_basic_blocks
);
3021 free_avail_expr_mem ()
3023 sbitmap_vector_free (ae_kill
);
3024 sbitmap_vector_free (ae_gen
);
3025 sbitmap_vector_free (ae_in
);
3026 sbitmap_vector_free (ae_out
);
3029 /* Compute the set of available expressions generated in each basic block. */
3038 /* For each recorded occurrence of each expression, set ae_gen[bb][expr].
3039 This is all we have to do because an expression is not recorded if it
3040 is not available, and the only expressions we want to work with are the
3041 ones that are recorded. */
3042 for (i
= 0; i
< expr_hash_table_size
; i
++)
3043 for (expr
= expr_hash_table
[i
]; expr
!= 0; expr
= expr
->next_same_hash
)
3044 for (occr
= expr
->avail_occr
; occr
!= 0; occr
= occr
->next
)
3045 SET_BIT (ae_gen
[BLOCK_NUM (occr
->insn
)], expr
->bitmap_index
);
3048 /* Return non-zero if expression X is killed in BB. */
3051 expr_killed_p (x
, bb
)
3062 code
= GET_CODE (x
);
3066 return TEST_BIT (reg_set_in_block
[bb
->index
], REGNO (x
));
3069 if (load_killed_in_block_p (bb
, get_max_uid () + 1, x
, 0))
3072 return expr_killed_p (XEXP (x
, 0), bb
);
3089 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
3093 /* If we are about to do the last recursive call
3094 needed at this level, change it into iteration.
3095 This function is called enough to be worth it. */
3097 return expr_killed_p (XEXP (x
, i
), bb
);
3098 else if (expr_killed_p (XEXP (x
, i
), bb
))
3101 else if (fmt
[i
] == 'E')
3102 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3103 if (expr_killed_p (XVECEXP (x
, i
, j
), bb
))
3110 /* Compute the set of available expressions killed in each basic block. */
3113 compute_ae_kill (ae_gen
, ae_kill
)
3114 sbitmap
*ae_gen
, *ae_kill
;
3120 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3121 for (i
= 0; i
< expr_hash_table_size
; i
++)
3122 for (expr
= expr_hash_table
[i
]; expr
; expr
= expr
->next_same_hash
)
3124 /* Skip EXPR if generated in this block. */
3125 if (TEST_BIT (ae_gen
[bb
], expr
->bitmap_index
))
3128 if (expr_killed_p (expr
->expr
, BASIC_BLOCK (bb
)))
3129 SET_BIT (ae_kill
[bb
], expr
->bitmap_index
);
3133 /* Actually perform the Classic GCSE optimizations. */
3135 /* Return non-zero if occurrence OCCR of expression EXPR reaches block BB.
3137 CHECK_SELF_LOOP is non-zero if we should consider a block reaching itself
3138 as a positive reach. We want to do this when there are two computations
3139 of the expression in the block.
3141 VISITED is a pointer to a working buffer for tracking which BB's have
3142 been visited. It is NULL for the top-level call.
3144 We treat reaching expressions that go through blocks containing the same
3145 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
3146 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
3147 2 as not reaching. The intent is to improve the probability of finding
3148 only one reaching expression and to reduce register lifetimes by picking
3149 the closest such expression. */
3152 expr_reaches_here_p_work (occr
, expr
, bb
, check_self_loop
, visited
)
3156 int check_self_loop
;
3161 for (pred
= bb
->pred
; pred
!= NULL
; pred
= pred
->pred_next
)
3163 basic_block pred_bb
= pred
->src
;
3165 if (visited
[pred_bb
->index
])
3166 /* This predecessor has already been visited. Nothing to do. */
3168 else if (pred_bb
== bb
)
3170 /* BB loops on itself. */
3172 && TEST_BIT (ae_gen
[pred_bb
->index
], expr
->bitmap_index
)
3173 && BLOCK_NUM (occr
->insn
) == pred_bb
->index
)
3176 visited
[pred_bb
->index
] = 1;
3179 /* Ignore this predecessor if it kills the expression. */
3180 else if (TEST_BIT (ae_kill
[pred_bb
->index
], expr
->bitmap_index
))
3181 visited
[pred_bb
->index
] = 1;
3183 /* Does this predecessor generate this expression? */
3184 else if (TEST_BIT (ae_gen
[pred_bb
->index
], expr
->bitmap_index
))
3186 /* Is this the occurrence we're looking for?
3187 Note that there's only one generating occurrence per block
3188 so we just need to check the block number. */
3189 if (BLOCK_NUM (occr
->insn
) == pred_bb
->index
)
3192 visited
[pred_bb
->index
] = 1;
3195 /* Neither gen nor kill. */
3198 visited
[pred_bb
->index
] = 1;
3199 if (expr_reaches_here_p_work (occr
, expr
, pred_bb
, check_self_loop
,
3206 /* All paths have been checked. */
3210 /* This wrapper for expr_reaches_here_p_work() is to ensure that any
3211 memory allocated for that function is returned. */
3214 expr_reaches_here_p (occr
, expr
, bb
, check_self_loop
)
3218 int check_self_loop
;
3221 char *visited
= (char *) xcalloc (n_basic_blocks
, 1);
3223 rval
= expr_reaches_here_p_work (occr
, expr
, bb
, check_self_loop
, visited
);
3229 /* Return the instruction that computes EXPR that reaches INSN's basic block.
3230 If there is more than one such instruction, return NULL.
3232 Called only by handle_avail_expr. */
3235 computing_insn (expr
, insn
)
3239 basic_block bb
= BLOCK_FOR_INSN (insn
);
3241 if (expr
->avail_occr
->next
== NULL
)
3243 if (BLOCK_FOR_INSN (expr
->avail_occr
->insn
) == bb
)
3244 /* The available expression is actually itself
3245 (i.e. a loop in the flow graph) so do nothing. */
3248 /* (FIXME) Case that we found a pattern that was created by
3249 a substitution that took place. */
3250 return expr
->avail_occr
->insn
;
3254 /* Pattern is computed more than once.
3255 Search backwards from this insn to see how many of these
3256 computations actually reach this insn. */
3258 rtx insn_computes_expr
= NULL
;
3261 for (occr
= expr
->avail_occr
; occr
!= NULL
; occr
= occr
->next
)
3263 if (BLOCK_FOR_INSN (occr
->insn
) == bb
)
3265 /* The expression is generated in this block.
3266 The only time we care about this is when the expression
3267 is generated later in the block [and thus there's a loop].
3268 We let the normal cse pass handle the other cases. */
3269 if (INSN_CUID (insn
) < INSN_CUID (occr
->insn
)
3270 && expr_reaches_here_p (occr
, expr
, bb
, 1))
3276 insn_computes_expr
= occr
->insn
;
3279 else if (expr_reaches_here_p (occr
, expr
, bb
, 0))
3285 insn_computes_expr
= occr
->insn
;
3289 if (insn_computes_expr
== NULL
)
3292 return insn_computes_expr
;
3296 /* Return non-zero if the definition in DEF_INSN can reach INSN.
3297 Only called by can_disregard_other_sets. */
3300 def_reaches_here_p (insn
, def_insn
)
3305 if (TEST_BIT (reaching_defs
[BLOCK_NUM (insn
)], INSN_CUID (def_insn
)))
3308 if (BLOCK_NUM (insn
) == BLOCK_NUM (def_insn
))
3310 if (INSN_CUID (def_insn
) < INSN_CUID (insn
))
3312 if (GET_CODE (PATTERN (def_insn
)) == PARALLEL
)
3314 else if (GET_CODE (PATTERN (def_insn
)) == CLOBBER
)
3315 reg
= XEXP (PATTERN (def_insn
), 0);
3316 else if (GET_CODE (PATTERN (def_insn
)) == SET
)
3317 reg
= SET_DEST (PATTERN (def_insn
));
3321 return ! reg_set_between_p (reg
, NEXT_INSN (def_insn
), insn
);
3330 /* Return non-zero if *ADDR_THIS_REG can only have one value at INSN. The
3331 value returned is the number of definitions that reach INSN. Returning a
3332 value of zero means that [maybe] more than one definition reaches INSN and
3333 the caller can't perform whatever optimization it is trying. i.e. it is
3334 always safe to return zero. */
3337 can_disregard_other_sets (addr_this_reg
, insn
, for_combine
)
3338 struct reg_set
**addr_this_reg
;
3342 int number_of_reaching_defs
= 0;
3343 struct reg_set
*this_reg
;
3345 for (this_reg
= *addr_this_reg
; this_reg
!= 0; this_reg
= this_reg
->next
)
3346 if (def_reaches_here_p (insn
, this_reg
->insn
))
3348 number_of_reaching_defs
++;
3349 /* Ignore parallels for now. */
3350 if (GET_CODE (PATTERN (this_reg
->insn
)) == PARALLEL
)
3354 && (GET_CODE (PATTERN (this_reg
->insn
)) == CLOBBER
3355 || ! rtx_equal_p (SET_SRC (PATTERN (this_reg
->insn
)),
3356 SET_SRC (PATTERN (insn
)))))
3357 /* A setting of the reg to a different value reaches INSN. */
3360 if (number_of_reaching_defs
> 1)
3362 /* If in this setting the value the register is being set to is
3363 equal to the previous value the register was set to and this
3364 setting reaches the insn we are trying to do the substitution
3365 on then we are ok. */
3366 if (GET_CODE (PATTERN (this_reg
->insn
)) == CLOBBER
)
3368 else if (! rtx_equal_p (SET_SRC (PATTERN (this_reg
->insn
)),
3369 SET_SRC (PATTERN (insn
))))
3373 *addr_this_reg
= this_reg
;
3376 return number_of_reaching_defs
;
3379 /* Expression computed by insn is available and the substitution is legal,
3380 so try to perform the substitution.
3382 The result is non-zero if any changes were made. */
3385 handle_avail_expr (insn
, expr
)
3389 rtx pat
, insn_computes_expr
, expr_set
;
3391 struct reg_set
*this_reg
;
3392 int found_setting
, use_src
;
3395 /* We only handle the case where one computation of the expression
3396 reaches this instruction. */
3397 insn_computes_expr
= computing_insn (expr
, insn
);
3398 if (insn_computes_expr
== NULL
)
3400 expr_set
= single_set (insn_computes_expr
);
3407 /* At this point we know only one computation of EXPR outside of this
3408 block reaches this insn. Now try to find a register that the
3409 expression is computed into. */
3410 if (GET_CODE (SET_SRC (expr_set
)) == REG
)
3412 /* This is the case when the available expression that reaches
3413 here has already been handled as an available expression. */
3414 unsigned int regnum_for_replacing
3415 = REGNO (SET_SRC (expr_set
));
3417 /* If the register was created by GCSE we can't use `reg_set_table',
3418 however we know it's set only once. */
3419 if (regnum_for_replacing
>= max_gcse_regno
3420 /* If the register the expression is computed into is set only once,
3421 or only one set reaches this insn, we can use it. */
3422 || (((this_reg
= reg_set_table
[regnum_for_replacing
]),
3423 this_reg
->next
== NULL
)
3424 || can_disregard_other_sets (&this_reg
, insn
, 0)))
3433 unsigned int regnum_for_replacing
3434 = REGNO (SET_DEST (expr_set
));
3436 /* This shouldn't happen. */
3437 if (regnum_for_replacing
>= max_gcse_regno
)
3440 this_reg
= reg_set_table
[regnum_for_replacing
];
3442 /* If the register the expression is computed into is set only once,
3443 or only one set reaches this insn, use it. */
3444 if (this_reg
->next
== NULL
3445 || can_disregard_other_sets (&this_reg
, insn
, 0))
3451 pat
= PATTERN (insn
);
3453 to
= SET_SRC (expr_set
);
3455 to
= SET_DEST (expr_set
);
3456 changed
= validate_change (insn
, &SET_SRC (pat
), to
, 0);
3458 /* We should be able to ignore the return code from validate_change but
3459 to play it safe we check. */
3463 if (gcse_file
!= NULL
)
3465 fprintf (gcse_file
, "GCSE: Replacing the source in insn %d with",
3467 fprintf (gcse_file
, " reg %d %s insn %d\n",
3468 REGNO (to
), use_src
? "from" : "set in",
3469 INSN_UID (insn_computes_expr
));
3474 /* The register that the expr is computed into is set more than once. */
3475 else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
3477 /* Insert an insn after insnx that copies the reg set in insnx
3478 into a new pseudo register call this new register REGN.
3479 From insnb until end of basic block or until REGB is set
3480 replace all uses of REGB with REGN. */
3483 to
= gen_reg_rtx (GET_MODE (SET_DEST (expr_set
)));
3485 /* Generate the new insn. */
3486 /* ??? If the change fails, we return 0, even though we created
3487 an insn. I think this is ok. */
3489 = emit_insn_after (gen_rtx_SET (VOIDmode
, to
,
3490 SET_DEST (expr_set
)),
3491 insn_computes_expr
);
3493 /* Keep block number table up to date. */
3494 set_block_for_new_insns (new_insn
, BLOCK_FOR_INSN (insn_computes_expr
));
3496 /* Keep register set table up to date. */
3497 record_one_set (REGNO (to
), new_insn
);
3499 gcse_create_count
++;
3500 if (gcse_file
!= NULL
)
3502 fprintf (gcse_file
, "GCSE: Creating insn %d to copy value of reg %d",
3503 INSN_UID (NEXT_INSN (insn_computes_expr
)),
3504 REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr
)))));
3505 fprintf (gcse_file
, ", computed in insn %d,\n",
3506 INSN_UID (insn_computes_expr
));
3507 fprintf (gcse_file
, " into newly allocated reg %d\n",
3511 pat
= PATTERN (insn
);
3513 /* Do register replacement for INSN. */
3514 changed
= validate_change (insn
, &SET_SRC (pat
),
3516 (NEXT_INSN (insn_computes_expr
))),
3519 /* We should be able to ignore the return code from validate_change but
3520 to play it safe we check. */
3524 if (gcse_file
!= NULL
)
3527 "GCSE: Replacing the source in insn %d with reg %d ",
3529 REGNO (SET_DEST (PATTERN (NEXT_INSN
3530 (insn_computes_expr
)))));
3531 fprintf (gcse_file
, "set in insn %d\n",
3532 INSN_UID (insn_computes_expr
));
3540 /* Perform classic GCSE. This is called by one_classic_gcse_pass after all
3541 the dataflow analysis has been done.
3543 The result is non-zero if a change was made. */
3551 /* Note we start at block 1. */
3554 for (bb
= 1; bb
< n_basic_blocks
; bb
++)
3556 /* Reset tables used to keep track of what's still valid [since the
3557 start of the block]. */
3558 reset_opr_set_tables ();
3560 for (insn
= BLOCK_HEAD (bb
);
3561 insn
!= NULL
&& insn
!= NEXT_INSN (BLOCK_END (bb
));
3562 insn
= NEXT_INSN (insn
))
3564 /* Is insn of form (set (pseudo-reg) ...)? */
3565 if (GET_CODE (insn
) == INSN
3566 && GET_CODE (PATTERN (insn
)) == SET
3567 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
3568 && REGNO (SET_DEST (PATTERN (insn
))) >= FIRST_PSEUDO_REGISTER
)
3570 rtx pat
= PATTERN (insn
);
3571 rtx src
= SET_SRC (pat
);
3574 if (want_to_gcse_p (src
)
3575 /* Is the expression recorded? */
3576 && ((expr
= lookup_expr (src
)) != NULL
)
3577 /* Is the expression available [at the start of the
3579 && TEST_BIT (ae_in
[bb
], expr
->bitmap_index
)
3580 /* Are the operands unchanged since the start of the
3582 && oprs_not_set_p (src
, insn
))
3583 changed
|= handle_avail_expr (insn
, expr
);
3586 /* Keep track of everything modified by this insn. */
3587 /* ??? Need to be careful w.r.t. mods done to INSN. */
3589 mark_oprs_set (insn
);
3596 /* Top level routine to perform one classic GCSE pass.
3598 Return non-zero if a change was made. */
3601 one_classic_gcse_pass (pass
)
3606 gcse_subst_count
= 0;
3607 gcse_create_count
= 0;
3609 alloc_expr_hash_table (max_cuid
);
3610 alloc_rd_mem (n_basic_blocks
, max_cuid
);
3611 compute_expr_hash_table ();
3613 dump_hash_table (gcse_file
, "Expression", expr_hash_table
,
3614 expr_hash_table_size
, n_exprs
);
3620 alloc_avail_expr_mem (n_basic_blocks
, n_exprs
);
3622 compute_ae_kill (ae_gen
, ae_kill
);
3623 compute_available (ae_gen
, ae_kill
, ae_out
, ae_in
);
3624 changed
= classic_gcse ();
3625 free_avail_expr_mem ();
3629 free_expr_hash_table ();
3633 fprintf (gcse_file
, "\n");
3634 fprintf (gcse_file
, "GCSE of %s, pass %d: %d bytes needed, %d substs,",
3635 current_function_name
, pass
, bytes_used
, gcse_subst_count
);
3636 fprintf (gcse_file
, "%d insns created\n", gcse_create_count
);
3642 /* Compute copy/constant propagation working variables. */
3644 /* Local properties of assignments. */
3645 static sbitmap
*cprop_pavloc
;
3646 static sbitmap
*cprop_absaltered
;
3648 /* Global properties of assignments (computed from the local properties). */
3649 static sbitmap
*cprop_avin
;
3650 static sbitmap
*cprop_avout
;
3652 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
3653 basic blocks. N_SETS is the number of sets. */
3656 alloc_cprop_mem (n_blocks
, n_sets
)
3657 int n_blocks
, n_sets
;
3659 cprop_pavloc
= sbitmap_vector_alloc (n_blocks
, n_sets
);
3660 cprop_absaltered
= sbitmap_vector_alloc (n_blocks
, n_sets
);
3662 cprop_avin
= sbitmap_vector_alloc (n_blocks
, n_sets
);
3663 cprop_avout
= sbitmap_vector_alloc (n_blocks
, n_sets
);
3666 /* Free vars used by copy/const propagation. */
3671 sbitmap_vector_free (cprop_pavloc
);
3672 sbitmap_vector_free (cprop_absaltered
);
3673 sbitmap_vector_free (cprop_avin
);
3674 sbitmap_vector_free (cprop_avout
);
3677 /* For each block, compute whether X is transparent. X is either an
3678 expression or an assignment [though we don't care which, for this context
3679 an assignment is treated as an expression]. For each block where an
3680 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
3684 compute_transp (x
, indx
, bmap
, set_p
)
3695 /* repeat is used to turn tail-recursion into iteration since GCC
3696 can't do it when there's no return value. */
3702 code
= GET_CODE (x
);
3708 if (REGNO (x
) < FIRST_PSEUDO_REGISTER
)
3710 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3711 if (TEST_BIT (reg_set_in_block
[bb
], REGNO (x
)))
3712 SET_BIT (bmap
[bb
], indx
);
3716 for (r
= reg_set_table
[REGNO (x
)]; r
!= NULL
; r
= r
->next
)
3717 SET_BIT (bmap
[BLOCK_NUM (r
->insn
)], indx
);
3722 if (REGNO (x
) < FIRST_PSEUDO_REGISTER
)
3724 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3725 if (TEST_BIT (reg_set_in_block
[bb
], REGNO (x
)))
3726 RESET_BIT (bmap
[bb
], indx
);
3730 for (r
= reg_set_table
[REGNO (x
)]; r
!= NULL
; r
= r
->next
)
3731 RESET_BIT (bmap
[BLOCK_NUM (r
->insn
)], indx
);
3738 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3740 rtx list_entry
= canon_modify_mem_list
[bb
];
3744 rtx dest
, dest_addr
;
3746 if (GET_CODE (XEXP (list_entry
, 0)) == CALL_INSN
)
3749 SET_BIT (bmap
[bb
], indx
);
3751 RESET_BIT (bmap
[bb
], indx
);
3754 /* LIST_ENTRY must be an INSN of some kind that sets memory.
3755 Examine each hunk of memory that is modified. */
3757 dest
= XEXP (list_entry
, 0);
3758 list_entry
= XEXP (list_entry
, 1);
3759 dest_addr
= XEXP (list_entry
, 0);
3761 if (canon_true_dependence (dest
, GET_MODE (dest
), dest_addr
,
3762 x
, rtx_addr_varies_p
))
3765 SET_BIT (bmap
[bb
], indx
);
3767 RESET_BIT (bmap
[bb
], indx
);
3770 list_entry
= XEXP (list_entry
, 1);
3792 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
3796 /* If we are about to do the last recursive call
3797 needed at this level, change it into iteration.
3798 This function is called enough to be worth it. */
3805 compute_transp (XEXP (x
, i
), indx
, bmap
, set_p
);
3807 else if (fmt
[i
] == 'E')
3808 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3809 compute_transp (XVECEXP (x
, i
, j
), indx
, bmap
, set_p
);
3813 /* Top level routine to do the dataflow analysis needed by copy/const
3817 compute_cprop_data ()
3819 compute_local_properties (cprop_absaltered
, cprop_pavloc
, NULL
, 1);
3820 compute_available (cprop_pavloc
, cprop_absaltered
,
3821 cprop_avout
, cprop_avin
);
3824 /* Copy/constant propagation. */
3826 /* Maximum number of register uses in an insn that we handle. */
3829 /* Table of uses found in an insn.
3830 Allocated statically to avoid alloc/free complexity and overhead. */
3831 static struct reg_use reg_use_table
[MAX_USES
];
3833 /* Index into `reg_use_table' while building it. */
3834 static int reg_use_count
;
3836 /* Set up a list of register numbers used in INSN. The found uses are stored
3837 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
3838 and contains the number of uses in the table upon exit.
3840 ??? If a register appears multiple times we will record it multiple times.
3841 This doesn't hurt anything but it will slow things down. */
3844 find_used_regs (xptr
, data
)
3846 void *data ATTRIBUTE_UNUSED
;
3853 /* repeat is used to turn tail-recursion into iteration since GCC
3854 can't do it when there's no return value. */
3859 code
= GET_CODE (x
);
3862 if (reg_use_count
== MAX_USES
)
3865 reg_use_table
[reg_use_count
].reg_rtx
= x
;
3869 /* Recursively scan the operands of this expression. */
3871 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
3875 /* If we are about to do the last recursive call
3876 needed at this level, change it into iteration.
3877 This function is called enough to be worth it. */
3884 find_used_regs (&XEXP (x
, i
), data
);
3886 else if (fmt
[i
] == 'E')
3887 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3888 find_used_regs (&XVECEXP (x
, i
, j
), data
);
3892 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
3893 Returns non-zero is successful. */
3896 try_replace_reg (from
, to
, insn
)
3899 rtx note
= find_reg_equal_equiv_note (insn
);
3902 rtx set
= single_set (insn
);
3904 success
= validate_replace_src (from
, to
, insn
);
3906 /* If above failed and this is a single set, try to simplify the source of
3907 the set given our substitution. We could perhaps try this for multiple
3908 SETs, but it probably won't buy us anything. */
3909 if (!success
&& set
!= 0)
3911 src
= simplify_replace_rtx (SET_SRC (set
), from
, to
);
3913 if (!rtx_equal_p (src
, SET_SRC (set
))
3914 && validate_change (insn
, &SET_SRC (set
), src
, 0))
3918 /* If we've failed to do replacement, have a single SET, and don't already
3919 have a note, add a REG_EQUAL note to not lose information. */
3920 if (!success
&& note
== 0 && set
!= 0)
3921 note
= REG_NOTES (insn
)
3922 = gen_rtx_EXPR_LIST (REG_EQUAL
, src
, REG_NOTES (insn
));
3924 /* If there is already a NOTE, update the expression in it with our
3927 XEXP (note
, 0) = simplify_replace_rtx (XEXP (note
, 0), from
, to
);
3929 /* REG_EQUAL may get simplified into register.
3930 We don't allow that. Remove that note. This code ought
3931 not to hapen, because previous code ought to syntetize
3932 reg-reg move, but be on the safe side. */
3933 if (note
&& REG_P (XEXP (note
, 0)))
3934 remove_note (insn
, note
);
3939 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
3940 NULL no such set is found. */
3942 static struct expr
*
3943 find_avail_set (regno
, insn
)
3947 /* SET1 contains the last set found that can be returned to the caller for
3948 use in a substitution. */
3949 struct expr
*set1
= 0;
3951 /* Loops are not possible here. To get a loop we would need two sets
3952 available at the start of the block containing INSN. ie we would
3953 need two sets like this available at the start of the block:
3955 (set (reg X) (reg Y))
3956 (set (reg Y) (reg X))
3958 This can not happen since the set of (reg Y) would have killed the
3959 set of (reg X) making it unavailable at the start of this block. */
3963 struct expr
*set
= lookup_set (regno
, NULL_RTX
);
3965 /* Find a set that is available at the start of the block
3966 which contains INSN. */
3969 if (TEST_BIT (cprop_avin
[BLOCK_NUM (insn
)], set
->bitmap_index
))
3971 set
= next_set (regno
, set
);
3974 /* If no available set was found we've reached the end of the
3975 (possibly empty) copy chain. */
3979 if (GET_CODE (set
->expr
) != SET
)
3982 src
= SET_SRC (set
->expr
);
3984 /* We know the set is available.
3985 Now check that SRC is ANTLOC (i.e. none of the source operands
3986 have changed since the start of the block).
3988 If the source operand changed, we may still use it for the next
3989 iteration of this loop, but we may not use it for substitutions. */
3991 if (CONSTANT_P (src
) || oprs_not_set_p (src
, insn
))
3994 /* If the source of the set is anything except a register, then
3995 we have reached the end of the copy chain. */
3996 if (GET_CODE (src
) != REG
)
3999 /* Follow the copy chain, ie start another iteration of the loop
4000 and see if we have an available copy into SRC. */
4001 regno
= REGNO (src
);
4004 /* SET1 holds the last set that was available and anticipatable at
4009 /* Subroutine of cprop_insn that tries to propagate constants into
4010 JUMP_INSNS. INSN must be a conditional jump. FROM is what we will try to
4011 replace, SRC is the constant we will try to substitute for it. Returns
4012 nonzero if a change was made. We know INSN has just a SET. */
4015 cprop_jump (bb
, insn
, from
, src
)
4021 rtx set
= PATTERN (insn
);
4022 rtx
new = simplify_replace_rtx (SET_SRC (set
), from
, src
);
4024 /* If no simplification can be made, then try the next
4026 if (rtx_equal_p (new, SET_SRC (set
)))
4029 /* If this is now a no-op leave it that way, but update LABEL_NUSED if
4033 SET_SRC (set
) = new;
4035 if (JUMP_LABEL (insn
) != 0)
4036 --LABEL_NUSES (JUMP_LABEL (insn
));
4039 /* Otherwise, this must be a valid instruction. */
4040 else if (! validate_change (insn
, &SET_SRC (set
), new, 0))
4043 /* If this has turned into an unconditional jump,
4044 then put a barrier after it so that the unreachable
4045 code will be deleted. */
4046 if (GET_CODE (SET_SRC (set
)) == LABEL_REF
)
4047 emit_barrier_after (insn
);
4049 run_jump_opt_after_gcse
= 1;
4052 if (gcse_file
!= NULL
)
4055 "CONST-PROP: Replacing reg %d in insn %d with constant ",
4056 REGNO (from
), INSN_UID (insn
));
4057 print_rtl (gcse_file
, src
);
4058 fprintf (gcse_file
, "\n");
4060 purge_dead_edges (bb
);
4067 /* Subroutine of cprop_insn that tries to propagate constants into JUMP_INSNS
4068 for machines that have CC0. INSN is a single set that stores into CC0;
4069 the insn following it is a conditional jump. REG_USED is the use we will
4070 try to replace, SRC is the constant we will try to substitute for it.
4071 Returns nonzero if a change was made. */
4074 cprop_cc0_jump (bb
, insn
, reg_used
, src
)
4077 struct reg_use
*reg_used
;
4080 /* First substitute in the SET_SRC of INSN, then substitute that for
4082 rtx jump
= NEXT_INSN (insn
);
4083 rtx new_src
= simplify_replace_rtx (SET_SRC (PATTERN (insn
)),
4084 reg_used
->reg_rtx
, src
);
4086 if (! cprop_jump (bb
, jump
, cc0_rtx
, new_src
))
4089 /* If we succeeded, delete the cc0 setter. */
4090 PUT_CODE (insn
, NOTE
);
4091 NOTE_LINE_NUMBER (insn
) = NOTE_INSN_DELETED
;
4092 NOTE_SOURCE_FILE (insn
) = 0;
4098 /* Perform constant and copy propagation on INSN.
4099 The result is non-zero if a change was made. */
4102 cprop_insn (bb
, insn
, alter_jumps
)
4107 struct reg_use
*reg_used
;
4115 note_uses (&PATTERN (insn
), find_used_regs
, NULL
);
4117 note
= find_reg_equal_equiv_note (insn
);
4119 /* We may win even when propagating constants into notes. */
4121 find_used_regs (&XEXP (note
, 0), NULL
);
4123 for (reg_used
= ®_use_table
[0]; reg_use_count
> 0;
4124 reg_used
++, reg_use_count
--)
4126 unsigned int regno
= REGNO (reg_used
->reg_rtx
);
4130 /* Ignore registers created by GCSE.
4131 We do this because ... */
4132 if (regno
>= max_gcse_regno
)
4135 /* If the register has already been set in this block, there's
4136 nothing we can do. */
4137 if (! oprs_not_set_p (reg_used
->reg_rtx
, insn
))
4140 /* Find an assignment that sets reg_used and is available
4141 at the start of the block. */
4142 set
= find_avail_set (regno
, insn
);
4147 /* ??? We might be able to handle PARALLELs. Later. */
4148 if (GET_CODE (pat
) != SET
)
4151 src
= SET_SRC (pat
);
4153 /* Constant propagation. */
4154 if (GET_CODE (src
) == CONST_INT
|| GET_CODE (src
) == CONST_DOUBLE
4155 || GET_CODE (src
) == SYMBOL_REF
)
4157 /* Handle normal insns first. */
4158 if (GET_CODE (insn
) == INSN
4159 && try_replace_reg (reg_used
->reg_rtx
, src
, insn
))
4163 if (gcse_file
!= NULL
)
4165 fprintf (gcse_file
, "CONST-PROP: Replacing reg %d in ",
4167 fprintf (gcse_file
, "insn %d with constant ",
4169 print_rtl (gcse_file
, src
);
4170 fprintf (gcse_file
, "\n");
4173 /* The original insn setting reg_used may or may not now be
4174 deletable. We leave the deletion to flow. */
4177 /* Try to propagate a CONST_INT into a conditional jump.
4178 We're pretty specific about what we will handle in this
4179 code, we can extend this as necessary over time.
4181 Right now the insn in question must look like
4182 (set (pc) (if_then_else ...)) */
4183 else if (alter_jumps
4184 && GET_CODE (insn
) == JUMP_INSN
4185 && condjump_p (insn
)
4186 && ! simplejump_p (insn
))
4187 changed
|= cprop_jump (bb
, insn
, reg_used
->reg_rtx
, src
);
4190 /* Similar code for machines that use a pair of CC0 setter and
4191 conditional jump insn. */
4192 else if (alter_jumps
4193 && GET_CODE (PATTERN (insn
)) == SET
4194 && SET_DEST (PATTERN (insn
)) == cc0_rtx
4195 && GET_CODE (NEXT_INSN (insn
)) == JUMP_INSN
4196 && condjump_p (NEXT_INSN (insn
))
4197 && ! simplejump_p (NEXT_INSN (insn
))
4198 && cprop_cc0_jump (bb
, insn
, reg_used
, src
))
4205 else if (GET_CODE (src
) == REG
4206 && REGNO (src
) >= FIRST_PSEUDO_REGISTER
4207 && REGNO (src
) != regno
)
4209 if (try_replace_reg (reg_used
->reg_rtx
, src
, insn
))
4213 if (gcse_file
!= NULL
)
4215 fprintf (gcse_file
, "COPY-PROP: Replacing reg %d in insn %d",
4216 regno
, INSN_UID (insn
));
4217 fprintf (gcse_file
, " with reg %d\n", REGNO (src
));
4220 /* The original insn setting reg_used may or may not now be
4221 deletable. We leave the deletion to flow. */
4222 /* FIXME: If it turns out that the insn isn't deletable,
4223 then we may have unnecessarily extended register lifetimes
4224 and made things worse. */
4232 /* Forward propagate copies. This includes copies and constants. Return
4233 non-zero if a change was made. */
4242 /* Note we start at block 1. */
4245 for (bb
= 1; bb
< n_basic_blocks
; bb
++)
4247 /* Reset tables used to keep track of what's still valid [since the
4248 start of the block]. */
4249 reset_opr_set_tables ();
4251 for (insn
= BLOCK_HEAD (bb
);
4252 insn
!= NULL
&& insn
!= NEXT_INSN (BLOCK_END (bb
));
4253 insn
= NEXT_INSN (insn
))
4256 changed
|= cprop_insn (BASIC_BLOCK (bb
), insn
, alter_jumps
);
4258 /* Keep track of everything modified by this insn. */
4259 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
4260 call mark_oprs_set if we turned the insn into a NOTE. */
4261 if (GET_CODE (insn
) != NOTE
)
4262 mark_oprs_set (insn
);
4266 if (gcse_file
!= NULL
)
4267 fprintf (gcse_file
, "\n");
4272 /* Perform one copy/constant propagation pass.
4273 F is the first insn in the function.
4274 PASS is the pass count. */
4277 one_cprop_pass (pass
, alter_jumps
)
4283 const_prop_count
= 0;
4284 copy_prop_count
= 0;
4286 alloc_set_hash_table (max_cuid
);
4287 compute_set_hash_table ();
4289 dump_hash_table (gcse_file
, "SET", set_hash_table
, set_hash_table_size
,
4293 alloc_cprop_mem (n_basic_blocks
, n_sets
);
4294 compute_cprop_data ();
4295 changed
= cprop (alter_jumps
);
4299 free_set_hash_table ();
4303 fprintf (gcse_file
, "CPROP of %s, pass %d: %d bytes needed, ",
4304 current_function_name
, pass
, bytes_used
);
4305 fprintf (gcse_file
, "%d const props, %d copy props\n\n",
4306 const_prop_count
, copy_prop_count
);
4312 /* Compute PRE+LCM working variables. */
4314 /* Local properties of expressions. */
4315 /* Nonzero for expressions that are transparent in the block. */
4316 static sbitmap
*transp
;
4318 /* Nonzero for expressions that are transparent at the end of the block.
4319 This is only zero for expressions killed by abnormal critical edge
4320 created by a calls. */
4321 static sbitmap
*transpout
;
4323 /* Nonzero for expressions that are computed (available) in the block. */
4324 static sbitmap
*comp
;
4326 /* Nonzero for expressions that are locally anticipatable in the block. */
4327 static sbitmap
*antloc
;
4329 /* Nonzero for expressions where this block is an optimal computation
4331 static sbitmap
*pre_optimal
;
4333 /* Nonzero for expressions which are redundant in a particular block. */
4334 static sbitmap
*pre_redundant
;
4336 /* Nonzero for expressions which should be inserted on a specific edge. */
4337 static sbitmap
*pre_insert_map
;
4339 /* Nonzero for expressions which should be deleted in a specific block. */
4340 static sbitmap
*pre_delete_map
;
4342 /* Contains the edge_list returned by pre_edge_lcm. */
4343 static struct edge_list
*edge_list
;
4345 /* Redundant insns. */
4346 static sbitmap pre_redundant_insns
;
4348 /* Allocate vars used for PRE analysis. */
4351 alloc_pre_mem (n_blocks
, n_exprs
)
4352 int n_blocks
, n_exprs
;
4354 transp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4355 comp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4356 antloc
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4359 pre_redundant
= NULL
;
4360 pre_insert_map
= NULL
;
4361 pre_delete_map
= NULL
;
4364 ae_kill
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4366 /* pre_insert and pre_delete are allocated later. */
4369 /* Free vars used for PRE analysis. */
4374 sbitmap_vector_free (transp
);
4375 sbitmap_vector_free (comp
);
4377 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
4380 sbitmap_vector_free (pre_optimal
);
4382 sbitmap_vector_free (pre_redundant
);
4384 sbitmap_vector_free (pre_insert_map
);
4386 sbitmap_vector_free (pre_delete_map
);
4388 sbitmap_vector_free (ae_in
);
4390 sbitmap_vector_free (ae_out
);
4392 transp
= comp
= NULL
;
4393 pre_optimal
= pre_redundant
= pre_insert_map
= pre_delete_map
= NULL
;
4394 ae_in
= ae_out
= NULL
;
4397 /* Top level routine to do the dataflow analysis needed by PRE. */
4402 sbitmap trapping_expr
;
4406 compute_local_properties (transp
, comp
, antloc
, 0);
4407 sbitmap_vector_zero (ae_kill
, n_basic_blocks
);
4409 /* Collect expressions which might trap. */
4410 trapping_expr
= sbitmap_alloc (n_exprs
);
4411 sbitmap_zero (trapping_expr
);
4412 for (ui
= 0; ui
< expr_hash_table_size
; ui
++)
4415 for (e
= expr_hash_table
[ui
]; e
!= NULL
; e
= e
->next_same_hash
)
4416 if (may_trap_p (e
->expr
))
4417 SET_BIT (trapping_expr
, e
->bitmap_index
);
4420 /* Compute ae_kill for each basic block using:
4424 This is significantly faster than compute_ae_kill. */
4426 for (i
= 0; i
< n_basic_blocks
; i
++)
4430 /* If the current block is the destination of an abnormal edge, we
4431 kill all trapping expressions because we won't be able to properly
4432 place the instruction on the edge. So make them neither
4433 anticipatable nor transparent. This is fairly conservative. */
4434 for (e
= BASIC_BLOCK (i
)->pred
; e
; e
= e
->pred_next
)
4435 if (e
->flags
& EDGE_ABNORMAL
)
4437 sbitmap_difference (antloc
[i
], antloc
[i
], trapping_expr
);
4438 sbitmap_difference (transp
[i
], transp
[i
], trapping_expr
);
4442 sbitmap_a_or_b (ae_kill
[i
], transp
[i
], comp
[i
]);
4443 sbitmap_not (ae_kill
[i
], ae_kill
[i
]);
4446 edge_list
= pre_edge_lcm (gcse_file
, n_exprs
, transp
, comp
, antloc
,
4447 ae_kill
, &pre_insert_map
, &pre_delete_map
);
4448 sbitmap_vector_free (antloc
);
4450 sbitmap_vector_free (ae_kill
);
4452 free (trapping_expr
);
4457 /* Return non-zero if an occurrence of expression EXPR in OCCR_BB would reach
4460 VISITED is a pointer to a working buffer for tracking which BB's have
4461 been visited. It is NULL for the top-level call.
4463 We treat reaching expressions that go through blocks containing the same
4464 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
4465 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
4466 2 as not reaching. The intent is to improve the probability of finding
4467 only one reaching expression and to reduce register lifetimes by picking
4468 the closest such expression. */
4471 pre_expr_reaches_here_p_work (occr_bb
, expr
, bb
, visited
)
4472 basic_block occr_bb
;
4479 for (pred
= bb
->pred
; pred
!= NULL
; pred
= pred
->pred_next
)
4481 basic_block pred_bb
= pred
->src
;
4483 if (pred
->src
== ENTRY_BLOCK_PTR
4484 /* Has predecessor has already been visited? */
4485 || visited
[pred_bb
->index
])
4486 ;/* Nothing to do. */
4488 /* Does this predecessor generate this expression? */
4489 else if (TEST_BIT (comp
[pred_bb
->index
], expr
->bitmap_index
))
4491 /* Is this the occurrence we're looking for?
4492 Note that there's only one generating occurrence per block
4493 so we just need to check the block number. */
4494 if (occr_bb
== pred_bb
)
4497 visited
[pred_bb
->index
] = 1;
4499 /* Ignore this predecessor if it kills the expression. */
4500 else if (! TEST_BIT (transp
[pred_bb
->index
], expr
->bitmap_index
))
4501 visited
[pred_bb
->index
] = 1;
4503 /* Neither gen nor kill. */
4506 visited
[pred_bb
->index
] = 1;
4507 if (pre_expr_reaches_here_p_work (occr_bb
, expr
, pred_bb
, visited
))
4512 /* All paths have been checked. */
4516 /* The wrapper for pre_expr_reaches_here_work that ensures that any
4517 memory allocated for that function is returned. */
4520 pre_expr_reaches_here_p (occr_bb
, expr
, bb
)
4521 basic_block occr_bb
;
4526 char *visited
= (char *) xcalloc (n_basic_blocks
, 1);
4528 rval
= pre_expr_reaches_here_p_work(occr_bb
, expr
, bb
, visited
);
4535 /* Given an expr, generate RTL which we can insert at the end of a BB,
4536 or on an edge. Set the block number of any insns generated to
4540 process_insert_insn (expr
)
4543 rtx reg
= expr
->reaching_reg
;
4544 rtx exp
= copy_rtx (expr
->expr
);
4549 /* If the expression is something that's an operand, like a constant,
4550 just copy it to a register. */
4551 if (general_operand (exp
, GET_MODE (reg
)))
4552 emit_move_insn (reg
, exp
);
4554 /* Otherwise, make a new insn to compute this expression and make sure the
4555 insn will be recognized (this also adds any needed CLOBBERs). Copy the
4556 expression to make sure we don't have any sharing issues. */
4557 else if (insn_invalid_p (emit_insn (gen_rtx_SET (VOIDmode
, reg
, exp
))))
4560 pat
= gen_sequence ();
4566 /* Add EXPR to the end of basic block BB.
4568 This is used by both the PRE and code hoisting.
4570 For PRE, we want to verify that the expr is either transparent
4571 or locally anticipatable in the target block. This check makes
4572 no sense for code hoisting. */
4575 insert_insn_end_bb (expr
, bb
, pre
)
4582 rtx reg
= expr
->reaching_reg
;
4583 int regno
= REGNO (reg
);
4587 pat
= process_insert_insn (expr
);
4589 /* If the last insn is a jump, insert EXPR in front [taking care to
4590 handle cc0, etc. properly]. */
4592 if (GET_CODE (insn
) == JUMP_INSN
)
4598 /* If this is a jump table, then we can't insert stuff here. Since
4599 we know the previous real insn must be the tablejump, we insert
4600 the new instruction just before the tablejump. */
4601 if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
4602 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
4603 insn
= prev_real_insn (insn
);
4606 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
4607 if cc0 isn't set. */
4608 note
= find_reg_note (insn
, REG_CC_SETTER
, NULL_RTX
);
4610 insn
= XEXP (note
, 0);
4613 rtx maybe_cc0_setter
= prev_nonnote_insn (insn
);
4614 if (maybe_cc0_setter
4615 && INSN_P (maybe_cc0_setter
)
4616 && sets_cc0_p (PATTERN (maybe_cc0_setter
)))
4617 insn
= maybe_cc0_setter
;
4620 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4621 new_insn
= emit_block_insn_before (pat
, insn
, bb
);
4624 /* Likewise if the last insn is a call, as will happen in the presence
4625 of exception handling. */
4626 else if (GET_CODE (insn
) == CALL_INSN
)
4628 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4629 we search backward and place the instructions before the first
4630 parameter is loaded. Do this for everyone for consistency and a
4631 presumtion that we'll get better code elsewhere as well.
4633 It should always be the case that we can put these instructions
4634 anywhere in the basic block with performing PRE optimizations.
4638 && !TEST_BIT (antloc
[bb
->index
], expr
->bitmap_index
)
4639 && !TEST_BIT (transp
[bb
->index
], expr
->bitmap_index
))
4642 /* Since different machines initialize their parameter registers
4643 in different orders, assume nothing. Collect the set of all
4644 parameter registers. */
4645 insn
= find_first_parameter_load (insn
, bb
->head
);
4647 /* If we found all the parameter loads, then we want to insert
4648 before the first parameter load.
4650 If we did not find all the parameter loads, then we might have
4651 stopped on the head of the block, which could be a CODE_LABEL.
4652 If we inserted before the CODE_LABEL, then we would be putting
4653 the insn in the wrong basic block. In that case, put the insn
4654 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4655 while (GET_CODE (insn
) == CODE_LABEL
4656 || NOTE_INSN_BASIC_BLOCK_P (insn
))
4657 insn
= NEXT_INSN (insn
);
4659 new_insn
= emit_block_insn_before (pat
, insn
, bb
);
4663 new_insn
= emit_insn_after (pat
, insn
);
4667 /* Keep block number table up to date.
4668 Note, PAT could be a multiple insn sequence, we have to make
4669 sure that each insn in the sequence is handled. */
4670 if (GET_CODE (pat
) == SEQUENCE
)
4672 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
4674 rtx insn
= XVECEXP (pat
, 0, i
);
4676 set_block_for_insn (insn
, bb
);
4678 add_label_notes (PATTERN (insn
), new_insn
);
4680 note_stores (PATTERN (insn
), record_set_info
, insn
);
4685 add_label_notes (SET_SRC (pat
), new_insn
);
4686 set_block_for_new_insns (new_insn
, bb
);
4688 /* Keep register set table up to date. */
4689 record_one_set (regno
, new_insn
);
4692 gcse_create_count
++;
4696 fprintf (gcse_file
, "PRE/HOIST: end of bb %d, insn %d, ",
4697 bb
->index
, INSN_UID (new_insn
));
4698 fprintf (gcse_file
, "copying expression %d to reg %d\n",
4699 expr
->bitmap_index
, regno
);
4703 /* Insert partially redundant expressions on edges in the CFG to make
4704 the expressions fully redundant. */
4707 pre_edge_insert (edge_list
, index_map
)
4708 struct edge_list
*edge_list
;
4709 struct expr
**index_map
;
4711 int e
, i
, j
, num_edges
, set_size
, did_insert
= 0;
4714 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4715 if it reaches any of the deleted expressions. */
4717 set_size
= pre_insert_map
[0]->size
;
4718 num_edges
= NUM_EDGES (edge_list
);
4719 inserted
= sbitmap_vector_alloc (num_edges
, n_exprs
);
4720 sbitmap_vector_zero (inserted
, num_edges
);
4722 for (e
= 0; e
< num_edges
; e
++)
4725 basic_block bb
= INDEX_EDGE_PRED_BB (edge_list
, e
);
4727 for (i
= indx
= 0; i
< set_size
; i
++, indx
+= SBITMAP_ELT_BITS
)
4729 SBITMAP_ELT_TYPE insert
= pre_insert_map
[e
]->elms
[i
];
4731 for (j
= indx
; insert
&& j
< n_exprs
; j
++, insert
>>= 1)
4732 if ((insert
& 1) != 0 && index_map
[j
]->reaching_reg
!= NULL_RTX
)
4734 struct expr
*expr
= index_map
[j
];
4737 /* Now look at each deleted occurence of this expression. */
4738 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4740 if (! occr
->deleted_p
)
4743 /* Insert this expression on this edge if if it would
4744 reach the deleted occurence in BB. */
4745 if (!TEST_BIT (inserted
[e
], j
))
4748 edge eg
= INDEX_EDGE (edge_list
, e
);
4750 /* We can't insert anything on an abnormal and
4751 critical edge, so we insert the insn at the end of
4752 the previous block. There are several alternatives
4753 detailed in Morgans book P277 (sec 10.5) for
4754 handling this situation. This one is easiest for
4757 if ((eg
->flags
& EDGE_ABNORMAL
) == EDGE_ABNORMAL
)
4758 insert_insn_end_bb (index_map
[j
], bb
, 0);
4761 insn
= process_insert_insn (index_map
[j
]);
4762 insert_insn_on_edge (insn
, eg
);
4767 fprintf (gcse_file
, "PRE/HOIST: edge (%d,%d), ",
4769 INDEX_EDGE_SUCC_BB (edge_list
, e
)->index
);
4770 fprintf (gcse_file
, "copy expression %d\n",
4771 expr
->bitmap_index
);
4774 update_ld_motion_stores (expr
);
4775 SET_BIT (inserted
[e
], j
);
4777 gcse_create_count
++;
4784 sbitmap_vector_free (inserted
);
4788 /* Copy the result of INSN to REG. INDX is the expression number. */
4791 pre_insert_copy_insn (expr
, insn
)
4795 rtx reg
= expr
->reaching_reg
;
4796 int regno
= REGNO (reg
);
4797 int indx
= expr
->bitmap_index
;
4798 rtx set
= single_set (insn
);
4800 basic_block bb
= BLOCK_FOR_INSN (insn
);
4805 new_insn
= emit_insn_after (gen_move_insn (reg
, SET_DEST (set
)), insn
);
4807 /* Keep block number table up to date. */
4808 set_block_for_new_insns (new_insn
, bb
);
4810 /* Keep register set table up to date. */
4811 record_one_set (regno
, new_insn
);
4812 if (insn
== bb
->end
)
4815 gcse_create_count
++;
4819 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4820 BLOCK_NUM (insn
), INSN_UID (new_insn
), indx
,
4821 INSN_UID (insn
), regno
);
4822 update_ld_motion_stores (expr
);
4825 /* Copy available expressions that reach the redundant expression
4826 to `reaching_reg'. */
4829 pre_insert_copies ()
4836 /* For each available expression in the table, copy the result to
4837 `reaching_reg' if the expression reaches a deleted one.
4839 ??? The current algorithm is rather brute force.
4840 Need to do some profiling. */
4842 for (i
= 0; i
< expr_hash_table_size
; i
++)
4843 for (expr
= expr_hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4845 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4846 we don't want to insert a copy here because the expression may not
4847 really be redundant. So only insert an insn if the expression was
4848 deleted. This test also avoids further processing if the
4849 expression wasn't deleted anywhere. */
4850 if (expr
->reaching_reg
== NULL
)
4853 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4855 if (! occr
->deleted_p
)
4858 for (avail
= expr
->avail_occr
; avail
!= NULL
; avail
= avail
->next
)
4860 rtx insn
= avail
->insn
;
4862 /* No need to handle this one if handled already. */
4863 if (avail
->copied_p
)
4866 /* Don't handle this one if it's a redundant one. */
4867 if (TEST_BIT (pre_redundant_insns
, INSN_CUID (insn
)))
4870 /* Or if the expression doesn't reach the deleted one. */
4871 if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail
->insn
),
4873 BLOCK_FOR_INSN (occr
->insn
)))
4876 /* Copy the result of avail to reaching_reg. */
4877 pre_insert_copy_insn (expr
, insn
);
4878 avail
->copied_p
= 1;
4884 /* Delete redundant computations.
4885 Deletion is done by changing the insn to copy the `reaching_reg' of
4886 the expression into the result of the SET. It is left to later passes
4887 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4889 Returns non-zero if a change is made. */
4900 for (i
= 0; i
< expr_hash_table_size
; i
++)
4901 for (expr
= expr_hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4903 int indx
= expr
->bitmap_index
;
4905 /* We only need to search antic_occr since we require
4908 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4910 rtx insn
= occr
->insn
;
4912 basic_block bb
= BLOCK_FOR_INSN (insn
);
4914 if (TEST_BIT (pre_delete_map
[bb
->index
], indx
))
4916 set
= single_set (insn
);
4920 /* Create a pseudo-reg to store the result of reaching
4921 expressions into. Get the mode for the new pseudo from
4922 the mode of the original destination pseudo. */
4923 if (expr
->reaching_reg
== NULL
)
4925 = gen_reg_rtx (GET_MODE (SET_DEST (set
)));
4927 /* In theory this should never fail since we're creating
4930 However, on the x86 some of the movXX patterns actually
4931 contain clobbers of scratch regs. This may cause the
4932 insn created by validate_change to not match any pattern
4933 and thus cause validate_change to fail. */
4934 if (validate_change (insn
, &SET_SRC (set
),
4935 expr
->reaching_reg
, 0))
4937 occr
->deleted_p
= 1;
4938 SET_BIT (pre_redundant_insns
, INSN_CUID (insn
));
4946 "PRE: redundant insn %d (expression %d) in ",
4947 INSN_UID (insn
), indx
);
4948 fprintf (gcse_file
, "bb %d, reaching reg is %d\n",
4949 bb
->index
, REGNO (expr
->reaching_reg
));
4958 /* Perform GCSE optimizations using PRE.
4959 This is called by one_pre_gcse_pass after all the dataflow analysis
4962 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4963 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4964 Compiler Design and Implementation.
4966 ??? A new pseudo reg is created to hold the reaching expression. The nice
4967 thing about the classical approach is that it would try to use an existing
4968 reg. If the register can't be adequately optimized [i.e. we introduce
4969 reload problems], one could add a pass here to propagate the new register
4972 ??? We don't handle single sets in PARALLELs because we're [currently] not
4973 able to copy the rest of the parallel when we insert copies to create full
4974 redundancies from partial redundancies. However, there's no reason why we
4975 can't handle PARALLELs in the cases where there are no partial
4982 int did_insert
, changed
;
4983 struct expr
**index_map
;
4986 /* Compute a mapping from expression number (`bitmap_index') to
4987 hash table entry. */
4989 index_map
= (struct expr
**) xcalloc (n_exprs
, sizeof (struct expr
*));
4990 for (i
= 0; i
< expr_hash_table_size
; i
++)
4991 for (expr
= expr_hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4992 index_map
[expr
->bitmap_index
] = expr
;
4994 /* Reset bitmap used to track which insns are redundant. */
4995 pre_redundant_insns
= sbitmap_alloc (max_cuid
);
4996 sbitmap_zero (pre_redundant_insns
);
4998 /* Delete the redundant insns first so that
4999 - we know what register to use for the new insns and for the other
5000 ones with reaching expressions
5001 - we know which insns are redundant when we go to create copies */
5003 changed
= pre_delete ();
5005 did_insert
= pre_edge_insert (edge_list
, index_map
);
5007 /* In other places with reaching expressions, copy the expression to the
5008 specially allocated pseudo-reg that reaches the redundant expr. */
5009 pre_insert_copies ();
5012 commit_edge_insertions ();
5017 free (pre_redundant_insns
);
5021 /* Top level routine to perform one PRE GCSE pass.
5023 Return non-zero if a change was made. */
5026 one_pre_gcse_pass (pass
)
5031 gcse_subst_count
= 0;
5032 gcse_create_count
= 0;
5034 alloc_expr_hash_table (max_cuid
);
5035 add_noreturn_fake_exit_edges ();
5037 compute_ld_motion_mems ();
5039 compute_expr_hash_table ();
5040 trim_ld_motion_mems ();
5042 dump_hash_table (gcse_file
, "Expression", expr_hash_table
,
5043 expr_hash_table_size
, n_exprs
);
5047 alloc_pre_mem (n_basic_blocks
, n_exprs
);
5048 compute_pre_data ();
5049 changed
|= pre_gcse ();
5050 free_edge_list (edge_list
);
5055 remove_fake_edges ();
5056 free_expr_hash_table ();
5060 fprintf (gcse_file
, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
5061 current_function_name
, pass
, bytes_used
);
5062 fprintf (gcse_file
, "%d substs, %d insns created\n",
5063 gcse_subst_count
, gcse_create_count
);
5069 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to INSN.
5070 If notes are added to an insn which references a CODE_LABEL, the
5071 LABEL_NUSES count is incremented. We have to add REG_LABEL notes,
5072 because the following loop optimization pass requires them. */
5074 /* ??? This is very similar to the loop.c add_label_notes function. We
5075 could probably share code here. */
5077 /* ??? If there was a jump optimization pass after gcse and before loop,
5078 then we would not need to do this here, because jump would add the
5079 necessary REG_LABEL notes. */
5082 add_label_notes (x
, insn
)
5086 enum rtx_code code
= GET_CODE (x
);
5090 if (code
== LABEL_REF
&& !LABEL_REF_NONLOCAL_P (x
))
5092 /* This code used to ignore labels that referred to dispatch tables to
5093 avoid flow generating (slighly) worse code.
5095 We no longer ignore such label references (see LABEL_REF handling in
5096 mark_jump_label for additional information). */
5098 REG_NOTES (insn
) = gen_rtx_EXPR_LIST (REG_LABEL
, XEXP (x
, 0),
5100 if (LABEL_P (XEXP (x
, 0)))
5101 LABEL_NUSES (XEXP (x
, 0))++;
5105 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
5108 add_label_notes (XEXP (x
, i
), insn
);
5109 else if (fmt
[i
] == 'E')
5110 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
5111 add_label_notes (XVECEXP (x
, i
, j
), insn
);
5115 /* Compute transparent outgoing information for each block.
5117 An expression is transparent to an edge unless it is killed by
5118 the edge itself. This can only happen with abnormal control flow,
5119 when the edge is traversed through a call. This happens with
5120 non-local labels and exceptions.
5122 This would not be necessary if we split the edge. While this is
5123 normally impossible for abnormal critical edges, with some effort
5124 it should be possible with exception handling, since we still have
5125 control over which handler should be invoked. But due to increased
5126 EH table sizes, this may not be worthwhile. */
5129 compute_transpout ()
5135 sbitmap_vector_ones (transpout
, n_basic_blocks
);
5137 for (bb
= 0; bb
< n_basic_blocks
; ++bb
)
5139 /* Note that flow inserted a nop a the end of basic blocks that
5140 end in call instructions for reasons other than abnormal
5142 if (GET_CODE (BLOCK_END (bb
)) != CALL_INSN
)
5145 for (i
= 0; i
< expr_hash_table_size
; i
++)
5146 for (expr
= expr_hash_table
[i
]; expr
; expr
= expr
->next_same_hash
)
5147 if (GET_CODE (expr
->expr
) == MEM
)
5149 if (GET_CODE (XEXP (expr
->expr
, 0)) == SYMBOL_REF
5150 && CONSTANT_POOL_ADDRESS_P (XEXP (expr
->expr
, 0)))
5153 /* ??? Optimally, we would use interprocedural alias
5154 analysis to determine if this mem is actually killed
5156 RESET_BIT (transpout
[bb
], expr
->bitmap_index
);
5161 /* Removal of useless null pointer checks */
5163 /* Called via note_stores. X is set by SETTER. If X is a register we must
5164 invalidate nonnull_local and set nonnull_killed. DATA is really a
5165 `null_pointer_info *'.
5167 We ignore hard registers. */
5170 invalidate_nonnull_info (x
, setter
, data
)
5172 rtx setter ATTRIBUTE_UNUSED
;
5176 struct null_pointer_info
*npi
= (struct null_pointer_info
*) data
;
5178 while (GET_CODE (x
) == SUBREG
)
5181 /* Ignore anything that is not a register or is a hard register. */
5182 if (GET_CODE (x
) != REG
5183 || REGNO (x
) < npi
->min_reg
5184 || REGNO (x
) >= npi
->max_reg
)
5187 regno
= REGNO (x
) - npi
->min_reg
;
5189 RESET_BIT (npi
->nonnull_local
[npi
->current_block
], regno
);
5190 SET_BIT (npi
->nonnull_killed
[npi
->current_block
], regno
);
5193 /* Do null-pointer check elimination for the registers indicated in
5194 NPI. NONNULL_AVIN and NONNULL_AVOUT are pre-allocated sbitmaps;
5195 they are not our responsibility to free. */
5198 delete_null_pointer_checks_1 (delete_list
, block_reg
, nonnull_avin
,
5200 varray_type
*delete_list
;
5201 unsigned int *block_reg
;
5202 sbitmap
*nonnull_avin
;
5203 sbitmap
*nonnull_avout
;
5204 struct null_pointer_info
*npi
;
5208 sbitmap
*nonnull_local
= npi
->nonnull_local
;
5209 sbitmap
*nonnull_killed
= npi
->nonnull_killed
;
5211 /* Compute local properties, nonnull and killed. A register will have
5212 the nonnull property if at the end of the current block its value is
5213 known to be nonnull. The killed property indicates that somewhere in
5214 the block any information we had about the register is killed.
5216 Note that a register can have both properties in a single block. That
5217 indicates that it's killed, then later in the block a new value is
5219 sbitmap_vector_zero (nonnull_local
, n_basic_blocks
);
5220 sbitmap_vector_zero (nonnull_killed
, n_basic_blocks
);
5222 for (current_block
= 0; current_block
< n_basic_blocks
; current_block
++)
5224 rtx insn
, stop_insn
;
5226 /* Set the current block for invalidate_nonnull_info. */
5227 npi
->current_block
= current_block
;
5229 /* Scan each insn in the basic block looking for memory references and
5231 stop_insn
= NEXT_INSN (BLOCK_END (current_block
));
5232 for (insn
= BLOCK_HEAD (current_block
);
5234 insn
= NEXT_INSN (insn
))
5239 /* Ignore anything that is not a normal insn. */
5240 if (! INSN_P (insn
))
5243 /* Basically ignore anything that is not a simple SET. We do have
5244 to make sure to invalidate nonnull_local and set nonnull_killed
5245 for such insns though. */
5246 set
= single_set (insn
);
5249 note_stores (PATTERN (insn
), invalidate_nonnull_info
, npi
);
5253 /* See if we've got a useable memory load. We handle it first
5254 in case it uses its address register as a dest (which kills
5255 the nonnull property). */
5256 if (GET_CODE (SET_SRC (set
)) == MEM
5257 && GET_CODE ((reg
= XEXP (SET_SRC (set
), 0))) == REG
5258 && REGNO (reg
) >= npi
->min_reg
5259 && REGNO (reg
) < npi
->max_reg
)
5260 SET_BIT (nonnull_local
[current_block
],
5261 REGNO (reg
) - npi
->min_reg
);
5263 /* Now invalidate stuff clobbered by this insn. */
5264 note_stores (PATTERN (insn
), invalidate_nonnull_info
, npi
);
5266 /* And handle stores, we do these last since any sets in INSN can
5267 not kill the nonnull property if it is derived from a MEM
5268 appearing in a SET_DEST. */
5269 if (GET_CODE (SET_DEST (set
)) == MEM
5270 && GET_CODE ((reg
= XEXP (SET_DEST (set
), 0))) == REG
5271 && REGNO (reg
) >= npi
->min_reg
5272 && REGNO (reg
) < npi
->max_reg
)
5273 SET_BIT (nonnull_local
[current_block
],
5274 REGNO (reg
) - npi
->min_reg
);
5278 /* Now compute global properties based on the local properties. This
5279 is a classic global availablity algorithm. */
5280 compute_available (nonnull_local
, nonnull_killed
,
5281 nonnull_avout
, nonnull_avin
);
5283 /* Now look at each bb and see if it ends with a compare of a value
5285 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
5287 rtx last_insn
= BLOCK_END (bb
);
5288 rtx condition
, earliest
;
5289 int compare_and_branch
;
5291 /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
5292 since BLOCK_REG[BB] is zero if this block did not end with a
5293 comparison against zero, this condition works. */
5294 if (block_reg
[bb
] < npi
->min_reg
5295 || block_reg
[bb
] >= npi
->max_reg
)
5298 /* LAST_INSN is a conditional jump. Get its condition. */
5299 condition
= get_condition (last_insn
, &earliest
);
5301 /* If we can't determine the condition then skip. */
5305 /* Is the register known to have a nonzero value? */
5306 if (!TEST_BIT (nonnull_avout
[bb
], block_reg
[bb
] - npi
->min_reg
))
5309 /* Try to compute whether the compare/branch at the loop end is one or
5310 two instructions. */
5311 if (earliest
== last_insn
)
5312 compare_and_branch
= 1;
5313 else if (earliest
== prev_nonnote_insn (last_insn
))
5314 compare_and_branch
= 2;
5318 /* We know the register in this comparison is nonnull at exit from
5319 this block. We can optimize this comparison. */
5320 if (GET_CODE (condition
) == NE
)
5324 new_jump
= emit_jump_insn_before (gen_jump (JUMP_LABEL (last_insn
)),
5326 JUMP_LABEL (new_jump
) = JUMP_LABEL (last_insn
);
5327 LABEL_NUSES (JUMP_LABEL (new_jump
))++;
5328 emit_barrier_after (new_jump
);
5331 VARRAY_RTX_INIT (*delete_list
, 10, "delete_list");
5333 VARRAY_PUSH_RTX (*delete_list
, last_insn
);
5334 if (compare_and_branch
== 2)
5335 VARRAY_PUSH_RTX (*delete_list
, earliest
);
5337 /* Don't check this block again. (Note that BLOCK_END is
5338 invalid here; we deleted the last instruction in the
5344 /* Find EQ/NE comparisons against zero which can be (indirectly) evaluated
5347 This is conceptually similar to global constant/copy propagation and
5348 classic global CSE (it even uses the same dataflow equations as cprop).
5350 If a register is used as memory address with the form (mem (reg)), then we
5351 know that REG can not be zero at that point in the program. Any instruction
5352 which sets REG "kills" this property.
5354 So, if every path leading to a conditional branch has an available memory
5355 reference of that form, then we know the register can not have the value
5356 zero at the conditional branch.
5358 So we merely need to compute the local properies and propagate that data
5359 around the cfg, then optimize where possible.
5361 We run this pass two times. Once before CSE, then again after CSE. This
5362 has proven to be the most profitable approach. It is rare for new
5363 optimization opportunities of this nature to appear after the first CSE
5366 This could probably be integrated with global cprop with a little work. */
5369 delete_null_pointer_checks (f
)
5370 rtx f ATTRIBUTE_UNUSED
;
5372 sbitmap
*nonnull_avin
, *nonnull_avout
;
5373 unsigned int *block_reg
;
5374 varray_type delete_list
= NULL
;
5380 struct null_pointer_info npi
;
5382 /* If we have only a single block, then there's nothing to do. */
5383 if (n_basic_blocks
<= 1)
5386 /* Trying to perform global optimizations on flow graphs which have
5387 a high connectivity will take a long time and is unlikely to be
5388 particularly useful.
5390 In normal circumstances a cfg should have about twice as many edges
5391 as blocks. But we do not want to punish small functions which have
5392 a couple switch statements. So we require a relatively large number
5393 of basic blocks and the ratio of edges to blocks to be high. */
5394 if (n_basic_blocks
> 1000 && n_edges
/ n_basic_blocks
>= 20)
5397 /* We need four bitmaps, each with a bit for each register in each
5399 max_reg
= max_reg_num ();
5400 regs_per_pass
= get_bitmap_width (4, n_basic_blocks
, max_reg
);
5402 /* Allocate bitmaps to hold local and global properties. */
5403 npi
.nonnull_local
= sbitmap_vector_alloc (n_basic_blocks
, regs_per_pass
);
5404 npi
.nonnull_killed
= sbitmap_vector_alloc (n_basic_blocks
, regs_per_pass
);
5405 nonnull_avin
= sbitmap_vector_alloc (n_basic_blocks
, regs_per_pass
);
5406 nonnull_avout
= sbitmap_vector_alloc (n_basic_blocks
, regs_per_pass
);
5408 /* Go through the basic blocks, seeing whether or not each block
5409 ends with a conditional branch whose condition is a comparison
5410 against zero. Record the register compared in BLOCK_REG. */
5411 block_reg
= (unsigned int *) xcalloc (n_basic_blocks
, sizeof (int));
5412 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
5414 rtx last_insn
= BLOCK_END (bb
);
5415 rtx condition
, earliest
, reg
;
5417 /* We only want conditional branches. */
5418 if (GET_CODE (last_insn
) != JUMP_INSN
5419 || !any_condjump_p (last_insn
)
5420 || !onlyjump_p (last_insn
))
5423 /* LAST_INSN is a conditional jump. Get its condition. */
5424 condition
= get_condition (last_insn
, &earliest
);
5426 /* If we were unable to get the condition, or it is not a equality
5427 comparison against zero then there's nothing we can do. */
5429 || (GET_CODE (condition
) != NE
&& GET_CODE (condition
) != EQ
)
5430 || GET_CODE (XEXP (condition
, 1)) != CONST_INT
5431 || (XEXP (condition
, 1)
5432 != CONST0_RTX (GET_MODE (XEXP (condition
, 0)))))
5435 /* We must be checking a register against zero. */
5436 reg
= XEXP (condition
, 0);
5437 if (GET_CODE (reg
) != REG
)
5440 block_reg
[bb
] = REGNO (reg
);
5443 /* Go through the algorithm for each block of registers. */
5444 for (reg
= FIRST_PSEUDO_REGISTER
; reg
< max_reg
; reg
+= regs_per_pass
)
5447 npi
.max_reg
= MIN (reg
+ regs_per_pass
, max_reg
);
5448 delete_null_pointer_checks_1 (&delete_list
, block_reg
, nonnull_avin
,
5449 nonnull_avout
, &npi
);
5452 /* Now delete the instructions all at once. This breaks the CFG. */
5455 for (i
= 0; i
< VARRAY_ACTIVE_SIZE (delete_list
); i
++)
5456 delete_insn (VARRAY_RTX (delete_list
, i
));
5457 VARRAY_FREE (delete_list
);
5460 /* Free the table of registers compared at the end of every block. */
5464 sbitmap_vector_free (npi
.nonnull_local
);
5465 sbitmap_vector_free (npi
.nonnull_killed
);
5466 sbitmap_vector_free (nonnull_avin
);
5467 sbitmap_vector_free (nonnull_avout
);
5470 /* Code Hoisting variables and subroutines. */
5472 /* Very busy expressions. */
5473 static sbitmap
*hoist_vbein
;
5474 static sbitmap
*hoist_vbeout
;
5476 /* Hoistable expressions. */
5477 static sbitmap
*hoist_exprs
;
5479 /* Dominator bitmaps. */
5480 static sbitmap
*dominators
;
5482 /* ??? We could compute post dominators and run this algorithm in
5483 reverse to to perform tail merging, doing so would probably be
5484 more effective than the tail merging code in jump.c.
5486 It's unclear if tail merging could be run in parallel with
5487 code hoisting. It would be nice. */
5489 /* Allocate vars used for code hoisting analysis. */
5492 alloc_code_hoist_mem (n_blocks
, n_exprs
)
5493 int n_blocks
, n_exprs
;
5495 antloc
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5496 transp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5497 comp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5499 hoist_vbein
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5500 hoist_vbeout
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5501 hoist_exprs
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5502 transpout
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5504 dominators
= sbitmap_vector_alloc (n_blocks
, n_blocks
);
5507 /* Free vars used for code hoisting analysis. */
5510 free_code_hoist_mem ()
5512 sbitmap_vector_free (antloc
);
5513 sbitmap_vector_free (transp
);
5514 sbitmap_vector_free (comp
);
5516 sbitmap_vector_free (hoist_vbein
);
5517 sbitmap_vector_free (hoist_vbeout
);
5518 sbitmap_vector_free (hoist_exprs
);
5519 sbitmap_vector_free (transpout
);
5521 sbitmap_vector_free (dominators
);
5524 /* Compute the very busy expressions at entry/exit from each block.
5526 An expression is very busy if all paths from a given point
5527 compute the expression. */
5530 compute_code_hoist_vbeinout ()
5532 int bb
, changed
, passes
;
5534 sbitmap_vector_zero (hoist_vbeout
, n_basic_blocks
);
5535 sbitmap_vector_zero (hoist_vbein
, n_basic_blocks
);
5544 /* We scan the blocks in the reverse order to speed up
5546 for (bb
= n_basic_blocks
- 1; bb
>= 0; bb
--)
5548 changed
|= sbitmap_a_or_b_and_c (hoist_vbein
[bb
], antloc
[bb
],
5549 hoist_vbeout
[bb
], transp
[bb
]);
5550 if (bb
!= n_basic_blocks
- 1)
5551 sbitmap_intersection_of_succs (hoist_vbeout
[bb
], hoist_vbein
, bb
);
5558 fprintf (gcse_file
, "hoisting vbeinout computation: %d passes\n", passes
);
5561 /* Top level routine to do the dataflow analysis needed by code hoisting. */
5564 compute_code_hoist_data ()
5566 compute_local_properties (transp
, comp
, antloc
, 0);
5567 compute_transpout ();
5568 compute_code_hoist_vbeinout ();
5569 calculate_dominance_info (NULL
, dominators
, CDI_DOMINATORS
);
5571 fprintf (gcse_file
, "\n");
5574 /* Determine if the expression identified by EXPR_INDEX would
5575 reach BB unimpared if it was placed at the end of EXPR_BB.
5577 It's unclear exactly what Muchnick meant by "unimpared". It seems
5578 to me that the expression must either be computed or transparent in
5579 *every* block in the path(s) from EXPR_BB to BB. Any other definition
5580 would allow the expression to be hoisted out of loops, even if
5581 the expression wasn't a loop invariant.
5583 Contrast this to reachability for PRE where an expression is
5584 considered reachable if *any* path reaches instead of *all*
5588 hoist_expr_reaches_here_p (expr_bb
, expr_index
, bb
, visited
)
5589 basic_block expr_bb
;
5595 int visited_allocated_locally
= 0;
5598 if (visited
== NULL
)
5600 visited_allocated_locally
= 1;
5601 visited
= xcalloc (n_basic_blocks
, 1);
5604 for (pred
= bb
->pred
; pred
!= NULL
; pred
= pred
->pred_next
)
5606 basic_block pred_bb
= pred
->src
;
5608 if (pred
->src
== ENTRY_BLOCK_PTR
)
5610 else if (visited
[pred_bb
->index
])
5613 /* Does this predecessor generate this expression? */
5614 else if (TEST_BIT (comp
[pred_bb
->index
], expr_index
))
5616 else if (! TEST_BIT (transp
[pred_bb
->index
], expr_index
))
5622 visited
[pred_bb
->index
] = 1;
5623 if (! hoist_expr_reaches_here_p (expr_bb
, expr_index
,
5628 if (visited_allocated_locally
)
5631 return (pred
== NULL
);
5634 /* Actually perform code hoisting. */
5641 struct expr
**index_map
;
5644 sbitmap_vector_zero (hoist_exprs
, n_basic_blocks
);
5646 /* Compute a mapping from expression number (`bitmap_index') to
5647 hash table entry. */
5649 index_map
= (struct expr
**) xcalloc (n_exprs
, sizeof (struct expr
*));
5650 for (i
= 0; i
< expr_hash_table_size
; i
++)
5651 for (expr
= expr_hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
5652 index_map
[expr
->bitmap_index
] = expr
;
5654 /* Walk over each basic block looking for potentially hoistable
5655 expressions, nothing gets hoisted from the entry block. */
5656 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
5659 int insn_inserted_p
;
5661 /* Examine each expression that is very busy at the exit of this
5662 block. These are the potentially hoistable expressions. */
5663 for (i
= 0; i
< hoist_vbeout
[bb
]->n_bits
; i
++)
5667 if (TEST_BIT (hoist_vbeout
[bb
], i
) && TEST_BIT (transpout
[bb
], i
))
5669 /* We've found a potentially hoistable expression, now
5670 we look at every block BB dominates to see if it
5671 computes the expression. */
5672 for (dominated
= 0; dominated
< n_basic_blocks
; dominated
++)
5674 /* Ignore self dominance. */
5676 || ! TEST_BIT (dominators
[dominated
], bb
))
5679 /* We've found a dominated block, now see if it computes
5680 the busy expression and whether or not moving that
5681 expression to the "beginning" of that block is safe. */
5682 if (!TEST_BIT (antloc
[dominated
], i
))
5685 /* Note if the expression would reach the dominated block
5686 unimpared if it was placed at the end of BB.
5688 Keep track of how many times this expression is hoistable
5689 from a dominated block into BB. */
5690 if (hoist_expr_reaches_here_p (BASIC_BLOCK (bb
), i
,
5691 BASIC_BLOCK (dominated
), NULL
))
5695 /* If we found more than one hoistable occurence of this
5696 expression, then note it in the bitmap of expressions to
5697 hoist. It makes no sense to hoist things which are computed
5698 in only one BB, and doing so tends to pessimize register
5699 allocation. One could increase this value to try harder
5700 to avoid any possible code expansion due to register
5701 allocation issues; however experiments have shown that
5702 the vast majority of hoistable expressions are only movable
5703 from two successors, so raising this threshhold is likely
5704 to nullify any benefit we get from code hoisting. */
5707 SET_BIT (hoist_exprs
[bb
], i
);
5713 /* If we found nothing to hoist, then quit now. */
5717 /* Loop over all the hoistable expressions. */
5718 for (i
= 0; i
< hoist_exprs
[bb
]->n_bits
; i
++)
5720 /* We want to insert the expression into BB only once, so
5721 note when we've inserted it. */
5722 insn_inserted_p
= 0;
5724 /* These tests should be the same as the tests above. */
5725 if (TEST_BIT (hoist_vbeout
[bb
], i
))
5727 /* We've found a potentially hoistable expression, now
5728 we look at every block BB dominates to see if it
5729 computes the expression. */
5730 for (dominated
= 0; dominated
< n_basic_blocks
; dominated
++)
5732 /* Ignore self dominance. */
5734 || ! TEST_BIT (dominators
[dominated
], bb
))
5737 /* We've found a dominated block, now see if it computes
5738 the busy expression and whether or not moving that
5739 expression to the "beginning" of that block is safe. */
5740 if (!TEST_BIT (antloc
[dominated
], i
))
5743 /* The expression is computed in the dominated block and
5744 it would be safe to compute it at the start of the
5745 dominated block. Now we have to determine if the
5746 expresion would reach the dominated block if it was
5747 placed at the end of BB. */
5748 if (hoist_expr_reaches_here_p (BASIC_BLOCK (bb
), i
,
5749 BASIC_BLOCK (dominated
), NULL
))
5751 struct expr
*expr
= index_map
[i
];
5752 struct occr
*occr
= expr
->antic_occr
;
5756 /* Find the right occurence of this expression. */
5757 while (BLOCK_NUM (occr
->insn
) != dominated
&& occr
)
5760 /* Should never happen. */
5766 set
= single_set (insn
);
5770 /* Create a pseudo-reg to store the result of reaching
5771 expressions into. Get the mode for the new pseudo
5772 from the mode of the original destination pseudo. */
5773 if (expr
->reaching_reg
== NULL
)
5775 = gen_reg_rtx (GET_MODE (SET_DEST (set
)));
5777 /* In theory this should never fail since we're creating
5780 However, on the x86 some of the movXX patterns
5781 actually contain clobbers of scratch regs. This may
5782 cause the insn created by validate_change to not
5783 match any pattern and thus cause validate_change to
5785 if (validate_change (insn
, &SET_SRC (set
),
5786 expr
->reaching_reg
, 0))
5788 occr
->deleted_p
= 1;
5789 if (!insn_inserted_p
)
5791 insert_insn_end_bb (index_map
[i
],
5792 BASIC_BLOCK (bb
), 0);
5793 insn_inserted_p
= 1;
5805 /* Top level routine to perform one code hoisting (aka unification) pass
5807 Return non-zero if a change was made. */
5810 one_code_hoisting_pass ()
5814 alloc_expr_hash_table (max_cuid
);
5815 compute_expr_hash_table ();
5817 dump_hash_table (gcse_file
, "Code Hosting Expressions", expr_hash_table
,
5818 expr_hash_table_size
, n_exprs
);
5822 alloc_code_hoist_mem (n_basic_blocks
, n_exprs
);
5823 compute_code_hoist_data ();
5825 free_code_hoist_mem ();
5828 free_expr_hash_table ();
5833 /* Here we provide the things required to do store motion towards
5834 the exit. In order for this to be effective, gcse also needed to
5835 be taught how to move a load when it is kill only by a store to itself.
5840 void foo(float scale)
5842 for (i=0; i<10; i++)
5846 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
5847 the load out since its live around the loop, and stored at the bottom
5850 The 'Load Motion' referred to and implemented in this file is
5851 an enhancement to gcse which when using edge based lcm, recognizes
5852 this situation and allows gcse to move the load out of the loop.
5854 Once gcse has hoisted the load, store motion can then push this
5855 load towards the exit, and we end up with no loads or stores of 'i'
5858 /* This will search the ldst list for a matching expresion. If it
5859 doesn't find one, we create one and initialize it. */
5861 static struct ls_expr
*
5865 struct ls_expr
* ptr
;
5867 for (ptr
= first_ls_expr(); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
5868 if (expr_equiv_p (ptr
->pattern
, x
))
5873 ptr
= (struct ls_expr
*) xmalloc (sizeof (struct ls_expr
));
5875 ptr
->next
= pre_ldst_mems
;
5878 ptr
->loads
= NULL_RTX
;
5879 ptr
->stores
= NULL_RTX
;
5880 ptr
->reaching_reg
= NULL_RTX
;
5883 ptr
->hash_index
= 0;
5884 pre_ldst_mems
= ptr
;
5890 /* Free up an individual ldst entry. */
5893 free_ldst_entry (ptr
)
5894 struct ls_expr
* ptr
;
5896 free_INSN_LIST_list (& ptr
->loads
);
5897 free_INSN_LIST_list (& ptr
->stores
);
5902 /* Free up all memory associated with the ldst list. */
5907 while (pre_ldst_mems
)
5909 struct ls_expr
* tmp
= pre_ldst_mems
;
5911 pre_ldst_mems
= pre_ldst_mems
->next
;
5913 free_ldst_entry (tmp
);
5916 pre_ldst_mems
= NULL
;
5919 /* Dump debugging info about the ldst list. */
5922 print_ldst_list (file
)
5925 struct ls_expr
* ptr
;
5927 fprintf (file
, "LDST list: \n");
5929 for (ptr
= first_ls_expr(); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
5931 fprintf (file
, " Pattern (%3d): ", ptr
->index
);
5933 print_rtl (file
, ptr
->pattern
);
5935 fprintf (file
, "\n Loads : ");
5938 print_rtl (file
, ptr
->loads
);
5940 fprintf (file
, "(nil)");
5942 fprintf (file
, "\n Stores : ");
5945 print_rtl (file
, ptr
->stores
);
5947 fprintf (file
, "(nil)");
5949 fprintf (file
, "\n\n");
5952 fprintf (file
, "\n");
5955 /* Returns 1 if X is in the list of ldst only expressions. */
5957 static struct ls_expr
*
5958 find_rtx_in_ldst (x
)
5961 struct ls_expr
* ptr
;
5963 for (ptr
= pre_ldst_mems
; ptr
!= NULL
; ptr
= ptr
->next
)
5964 if (expr_equiv_p (ptr
->pattern
, x
) && ! ptr
->invalid
)
5970 /* Assign each element of the list of mems a monotonically increasing value. */
5975 struct ls_expr
* ptr
;
5978 for (ptr
= pre_ldst_mems
; ptr
!= NULL
; ptr
= ptr
->next
)
5984 /* Return first item in the list. */
5986 static inline struct ls_expr
*
5989 return pre_ldst_mems
;
5992 /* Return the next item in ther list after the specified one. */
5994 static inline struct ls_expr
*
5996 struct ls_expr
* ptr
;
6001 /* Load Motion for loads which only kill themselves. */
6003 /* Return true if x is a simple MEM operation, with no registers or
6004 side effects. These are the types of loads we consider for the
6005 ld_motion list, otherwise we let the usual aliasing take care of it. */
6011 if (GET_CODE (x
) != MEM
)
6014 if (MEM_VOLATILE_P (x
))
6017 if (GET_MODE (x
) == BLKmode
)
6020 if (!rtx_varies_p (XEXP (x
, 0), 0))
6026 /* Make sure there isn't a buried reference in this pattern anywhere.
6027 If there is, invalidate the entry for it since we're not capable
6028 of fixing it up just yet.. We have to be sure we know about ALL
6029 loads since the aliasing code will allow all entries in the
6030 ld_motion list to not-alias itself. If we miss a load, we will get
6031 the wrong value since gcse might common it and we won't know to
6035 invalidate_any_buried_refs (x
)
6040 struct ls_expr
* ptr
;
6042 /* Invalidate it in the list. */
6043 if (GET_CODE (x
) == MEM
&& simple_mem (x
))
6045 ptr
= ldst_entry (x
);
6049 /* Recursively process the insn. */
6050 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
6052 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
6055 invalidate_any_buried_refs (XEXP (x
, i
));
6056 else if (fmt
[i
] == 'E')
6057 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
6058 invalidate_any_buried_refs (XVECEXP (x
, i
, j
));
6062 /* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
6063 being defined as MEM loads and stores to symbols, with no
6064 side effects and no registers in the expression. If there are any
6065 uses/defs which dont match this criteria, it is invalidated and
6066 trimmed out later. */
6069 compute_ld_motion_mems ()
6071 struct ls_expr
* ptr
;
6075 pre_ldst_mems
= NULL
;
6077 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
6079 for (insn
= BLOCK_HEAD (bb
);
6080 insn
&& insn
!= NEXT_INSN (BLOCK_END (bb
));
6081 insn
= NEXT_INSN (insn
))
6083 if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i')
6085 if (GET_CODE (PATTERN (insn
)) == SET
)
6087 rtx src
= SET_SRC (PATTERN (insn
));
6088 rtx dest
= SET_DEST (PATTERN (insn
));
6090 /* Check for a simple LOAD... */
6091 if (GET_CODE (src
) == MEM
&& simple_mem (src
))
6093 ptr
= ldst_entry (src
);
6094 if (GET_CODE (dest
) == REG
)
6095 ptr
->loads
= alloc_INSN_LIST (insn
, ptr
->loads
);
6101 /* Make sure there isn't a buried load somewhere. */
6102 invalidate_any_buried_refs (src
);
6105 /* Check for stores. Don't worry about aliased ones, they
6106 will block any movement we might do later. We only care
6107 about this exact pattern since those are the only
6108 circumstance that we will ignore the aliasing info. */
6109 if (GET_CODE (dest
) == MEM
&& simple_mem (dest
))
6111 ptr
= ldst_entry (dest
);
6113 if (GET_CODE (src
) != MEM
6114 && GET_CODE (src
) != ASM_OPERANDS
)
6115 ptr
->stores
= alloc_INSN_LIST (insn
, ptr
->stores
);
6121 invalidate_any_buried_refs (PATTERN (insn
));
6127 /* Remove any references that have been either invalidated or are not in the
6128 expression list for pre gcse. */
6131 trim_ld_motion_mems ()
6133 struct ls_expr
* last
= NULL
;
6134 struct ls_expr
* ptr
= first_ls_expr ();
6138 int del
= ptr
->invalid
;
6139 struct expr
* expr
= NULL
;
6141 /* Delete if entry has been made invalid. */
6147 /* Delete if we cannot find this mem in the expression list. */
6148 for (i
= 0; i
< expr_hash_table_size
&& del
; i
++)
6150 for (expr
= expr_hash_table
[i
];
6152 expr
= expr
->next_same_hash
)
6153 if (expr_equiv_p (expr
->expr
, ptr
->pattern
))
6165 last
->next
= ptr
->next
;
6166 free_ldst_entry (ptr
);
6171 pre_ldst_mems
= pre_ldst_mems
->next
;
6172 free_ldst_entry (ptr
);
6173 ptr
= pre_ldst_mems
;
6178 /* Set the expression field if we are keeping it. */
6185 /* Show the world what we've found. */
6186 if (gcse_file
&& pre_ldst_mems
!= NULL
)
6187 print_ldst_list (gcse_file
);
6190 /* This routine will take an expression which we are replacing with
6191 a reaching register, and update any stores that are needed if
6192 that expression is in the ld_motion list. Stores are updated by
6193 copying their SRC to the reaching register, and then storeing
6194 the reaching register into the store location. These keeps the
6195 correct value in the reaching register for the loads. */
6198 update_ld_motion_stores (expr
)
6201 struct ls_expr
* mem_ptr
;
6203 if ((mem_ptr
= find_rtx_in_ldst (expr
->expr
)))
6205 /* We can try to find just the REACHED stores, but is shouldn't
6206 matter to set the reaching reg everywhere... some might be
6207 dead and should be eliminated later. */
6209 /* We replace SET mem = expr with
6211 SET mem = reg , where reg is the
6212 reaching reg used in the load. */
6213 rtx list
= mem_ptr
->stores
;
6215 for ( ; list
!= NULL_RTX
; list
= XEXP (list
, 1))
6217 rtx insn
= XEXP (list
, 0);
6218 rtx pat
= PATTERN (insn
);
6219 rtx src
= SET_SRC (pat
);
6220 rtx reg
= expr
->reaching_reg
;
6223 /* If we've already copied it, continue. */
6224 if (expr
->reaching_reg
== src
)
6229 fprintf (gcse_file
, "PRE: store updated with reaching reg ");
6230 print_rtl (gcse_file
, expr
->reaching_reg
);
6231 fprintf (gcse_file
, ":\n ");
6232 print_inline_rtx (gcse_file
, insn
, 8);
6233 fprintf (gcse_file
, "\n");
6236 copy
= gen_move_insn ( reg
, SET_SRC (pat
));
6237 new = emit_insn_before (copy
, insn
);
6238 record_one_set (REGNO (reg
), new);
6239 set_block_for_new_insns (new, BLOCK_FOR_INSN (insn
));
6240 SET_SRC (pat
) = reg
;
6242 /* un-recognize this pattern since it's probably different now. */
6243 INSN_CODE (insn
) = -1;
6244 gcse_create_count
++;
6249 /* Store motion code. */
6251 /* This is used to communicate the target bitvector we want to use in the
6252 reg_set_info routine when called via the note_stores mechanism. */
6253 static sbitmap
* regvec
;
6255 /* Used in computing the reverse edge graph bit vectors. */
6256 static sbitmap
* st_antloc
;
6258 /* Global holding the number of store expressions we are dealing with. */
6259 static int num_stores
;
6261 /* Checks to set if we need to mark a register set. Called from note_stores. */
6264 reg_set_info (dest
, setter
, data
)
6265 rtx dest
, setter ATTRIBUTE_UNUSED
;
6266 void * data ATTRIBUTE_UNUSED
;
6268 if (GET_CODE (dest
) == SUBREG
)
6269 dest
= SUBREG_REG (dest
);
6271 if (GET_CODE (dest
) == REG
)
6272 SET_BIT (*regvec
, REGNO (dest
));
6275 /* Return non-zero if the register operands of expression X are killed
6276 anywhere in basic block BB. */
6279 store_ops_ok (x
, bb
)
6287 /* Repeat is used to turn tail-recursion into iteration. */
6293 code
= GET_CODE (x
);
6297 /* If a reg has changed after us in this
6298 block, the operand has been killed. */
6299 return TEST_BIT (reg_set_in_block
[bb
->index
], REGNO (x
));
6326 i
= GET_RTX_LENGTH (code
) - 1;
6327 fmt
= GET_RTX_FORMAT (code
);
6333 rtx tem
= XEXP (x
, i
);
6335 /* If we are about to do the last recursive call
6336 needed at this level, change it into iteration.
6337 This function is called enough to be worth it. */
6344 if (! store_ops_ok (tem
, bb
))
6347 else if (fmt
[i
] == 'E')
6351 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
6353 if (! store_ops_ok (XVECEXP (x
, i
, j
), bb
))
6362 /* Determine whether insn is MEM store pattern that we will consider moving. */
6365 find_moveable_store (insn
)
6368 struct ls_expr
* ptr
;
6369 rtx dest
= PATTERN (insn
);
6371 if (GET_CODE (dest
) != SET
6372 || GET_CODE (SET_SRC (dest
)) == ASM_OPERANDS
)
6375 dest
= SET_DEST (dest
);
6377 if (GET_CODE (dest
) != MEM
|| MEM_VOLATILE_P (dest
)
6378 || GET_MODE (dest
) == BLKmode
)
6381 if (GET_CODE (XEXP (dest
, 0)) != SYMBOL_REF
)
6384 if (rtx_varies_p (XEXP (dest
, 0), 0))
6387 ptr
= ldst_entry (dest
);
6388 ptr
->stores
= alloc_INSN_LIST (insn
, ptr
->stores
);
6391 /* Perform store motion. Much like gcse, except we move expressions the
6392 other way by looking at the flowgraph in reverse. */
6395 compute_store_table ()
6401 max_gcse_regno
= max_reg_num ();
6403 reg_set_in_block
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
,
6405 sbitmap_vector_zero (reg_set_in_block
, n_basic_blocks
);
6408 /* Find all the stores we care about. */
6409 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
6411 regvec
= & (reg_set_in_block
[bb
]);
6412 for (insn
= BLOCK_END (bb
);
6413 insn
&& insn
!= PREV_INSN (BLOCK_HEAD (bb
));
6414 insn
= PREV_INSN (insn
))
6416 /* Ignore anything that is not a normal insn. */
6417 if (! INSN_P (insn
))
6420 if (GET_CODE (insn
) == CALL_INSN
)
6422 bool clobbers_all
= false;
6423 #ifdef NON_SAVING_SETJMP
6424 if (NON_SAVING_SETJMP
6425 && find_reg_note (insn
, REG_SETJMP
, NULL_RTX
))
6426 clobbers_all
= true;
6429 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
6431 || TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
6432 SET_BIT (reg_set_in_block
[bb
], regno
);
6435 pat
= PATTERN (insn
);
6436 note_stores (pat
, reg_set_info
, NULL
);
6438 /* Now that we've marked regs, look for stores. */
6439 if (GET_CODE (pat
) == SET
)
6440 find_moveable_store (insn
);
6444 ret
= enumerate_ldsts ();
6448 fprintf (gcse_file
, "Store Motion Expressions.\n");
6449 print_ldst_list (gcse_file
);
6455 /* Check to see if the load X is aliased with STORE_PATTERN. */
6458 load_kills_store (x
, store_pattern
)
6459 rtx x
, store_pattern
;
6461 if (true_dependence (x
, GET_MODE (x
), store_pattern
, rtx_addr_varies_p
))
6466 /* Go through the entire insn X, looking for any loads which might alias
6467 STORE_PATTERN. Return 1 if found. */
6470 find_loads (x
, store_pattern
)
6471 rtx x
, store_pattern
;
6480 if (GET_CODE (x
) == SET
)
6483 if (GET_CODE (x
) == MEM
)
6485 if (load_kills_store (x
, store_pattern
))
6489 /* Recursively process the insn. */
6490 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
6492 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0 && !ret
; i
--)
6495 ret
|= find_loads (XEXP (x
, i
), store_pattern
);
6496 else if (fmt
[i
] == 'E')
6497 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
6498 ret
|= find_loads (XVECEXP (x
, i
, j
), store_pattern
);
6503 /* Check if INSN kills the store pattern X (is aliased with it).
6504 Return 1 if it it does. */
6507 store_killed_in_insn (x
, insn
)
6510 if (GET_RTX_CLASS (GET_CODE (insn
)) != 'i')
6513 if (GET_CODE (insn
) == CALL_INSN
)
6515 if (CONST_OR_PURE_CALL_P (insn
))
6521 if (GET_CODE (PATTERN (insn
)) == SET
)
6523 rtx pat
= PATTERN (insn
);
6524 /* Check for memory stores to aliased objects. */
6525 if (GET_CODE (SET_DEST (pat
)) == MEM
&& !expr_equiv_p (SET_DEST (pat
), x
))
6526 /* pretend its a load and check for aliasing. */
6527 if (find_loads (SET_DEST (pat
), x
))
6529 return find_loads (SET_SRC (pat
), x
);
6532 return find_loads (PATTERN (insn
), x
);
6535 /* Returns 1 if the expression X is loaded or clobbered on or after INSN
6536 within basic block BB. */
6539 store_killed_after (x
, insn
, bb
)
6548 /* Check if the register operands of the store are OK in this block.
6549 Note that if registers are changed ANYWHERE in the block, we'll
6550 decide we can't move it, regardless of whether it changed above
6551 or below the store. This could be improved by checking the register
6552 operands while lookinng for aliasing in each insn. */
6553 if (!store_ops_ok (XEXP (x
, 0), bb
))
6556 for ( ; insn
&& insn
!= NEXT_INSN (last
); insn
= NEXT_INSN (insn
))
6557 if (store_killed_in_insn (x
, insn
))
6563 /* Returns 1 if the expression X is loaded or clobbered on or before INSN
6564 within basic block BB. */
6566 store_killed_before (x
, insn
, bb
)
6570 rtx first
= bb
->head
;
6573 return store_killed_in_insn (x
, insn
);
6575 /* Check if the register operands of the store are OK in this block.
6576 Note that if registers are changed ANYWHERE in the block, we'll
6577 decide we can't move it, regardless of whether it changed above
6578 or below the store. This could be improved by checking the register
6579 operands while lookinng for aliasing in each insn. */
6580 if (!store_ops_ok (XEXP (x
, 0), bb
))
6583 for ( ; insn
&& insn
!= PREV_INSN (first
); insn
= PREV_INSN (insn
))
6584 if (store_killed_in_insn (x
, insn
))
6590 #define ANTIC_STORE_LIST(x) ((x)->loads)
6591 #define AVAIL_STORE_LIST(x) ((x)->stores)
6593 /* Given the table of available store insns at the end of blocks,
6594 determine which ones are not killed by aliasing, and generate
6595 the appropriate vectors for gen and killed. */
6597 build_store_vectors ()
6602 struct ls_expr
* ptr
;
6604 /* Build the gen_vector. This is any store in the table which is not killed
6605 by aliasing later in its block. */
6606 ae_gen
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
, num_stores
);
6607 sbitmap_vector_zero (ae_gen
, n_basic_blocks
);
6609 st_antloc
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
, num_stores
);
6610 sbitmap_vector_zero (st_antloc
, n_basic_blocks
);
6612 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6614 /* Put all the stores into either the antic list, or the avail list,
6616 rtx store_list
= ptr
->stores
;
6617 ptr
->stores
= NULL_RTX
;
6619 for (st
= store_list
; st
!= NULL
; st
= XEXP (st
, 1))
6621 insn
= XEXP (st
, 0);
6622 bb
= BLOCK_FOR_INSN (insn
);
6624 if (!store_killed_after (ptr
->pattern
, insn
, bb
))
6626 /* If we've already seen an availale expression in this block,
6627 we can delete the one we saw already (It occurs earlier in
6628 the block), and replace it with this one). We'll copy the
6629 old SRC expression to an unused register in case there
6630 are any side effects. */
6631 if (TEST_BIT (ae_gen
[bb
->index
], ptr
->index
))
6633 /* Find previous store. */
6635 for (st
= AVAIL_STORE_LIST (ptr
); st
; st
= XEXP (st
, 1))
6636 if (BLOCK_FOR_INSN (XEXP (st
, 0)) == bb
)
6640 rtx r
= gen_reg_rtx (GET_MODE (ptr
->pattern
));
6642 fprintf(gcse_file
, "Removing redundant store:\n");
6643 replace_store_insn (r
, XEXP (st
, 0), bb
);
6644 XEXP (st
, 0) = insn
;
6648 SET_BIT (ae_gen
[bb
->index
], ptr
->index
);
6649 AVAIL_STORE_LIST (ptr
) = alloc_INSN_LIST (insn
,
6650 AVAIL_STORE_LIST (ptr
));
6653 if (!store_killed_before (ptr
->pattern
, insn
, bb
))
6655 SET_BIT (st_antloc
[BLOCK_NUM (insn
)], ptr
->index
);
6656 ANTIC_STORE_LIST (ptr
) = alloc_INSN_LIST (insn
,
6657 ANTIC_STORE_LIST (ptr
));
6661 /* Free the original list of store insns. */
6662 free_INSN_LIST_list (&store_list
);
6665 ae_kill
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
, num_stores
);
6666 sbitmap_vector_zero (ae_kill
, n_basic_blocks
);
6668 transp
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
, num_stores
);
6669 sbitmap_vector_zero (transp
, n_basic_blocks
);
6671 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6672 for (b
= 0; b
< n_basic_blocks
; b
++)
6674 if (store_killed_after (ptr
->pattern
, BLOCK_HEAD (b
), BASIC_BLOCK (b
)))
6676 /* The anticipatable expression is not killed if it's gen'd. */
6678 We leave this check out for now. If we have a code sequence
6679 in a block which looks like:
6683 We should flag this as having an ANTIC expression, NOT
6684 transparent, NOT killed, and AVAIL.
6685 Unfortunately, since we haven't re-written all loads to
6686 use the reaching reg, we'll end up doing an incorrect
6687 Load in the middle here if we push the store down. It happens in
6688 gcc.c-torture/execute/960311-1.c with -O3
6689 If we always kill it in this case, we'll sometimes do
6690 uneccessary work, but it shouldn't actually hurt anything.
6691 if (!TEST_BIT (ae_gen[b], ptr->index)). */
6692 SET_BIT (ae_kill
[b
], ptr
->index
);
6695 SET_BIT (transp
[b
], ptr
->index
);
6698 /* Any block with no exits calls some non-returning function, so
6699 we better mark the store killed here, or we might not store to
6700 it at all. If we knew it was abort, we wouldn't have to store,
6701 but we don't know that for sure. */
6704 fprintf (gcse_file
, "ST_avail and ST_antic (shown under loads..)\n");
6705 print_ldst_list (gcse_file
);
6706 dump_sbitmap_vector (gcse_file
, "st_antloc", "", st_antloc
, n_basic_blocks
);
6707 dump_sbitmap_vector (gcse_file
, "st_kill", "", ae_kill
, n_basic_blocks
);
6708 dump_sbitmap_vector (gcse_file
, "Transpt", "", transp
, n_basic_blocks
);
6709 dump_sbitmap_vector (gcse_file
, "st_avloc", "", ae_gen
, n_basic_blocks
);
6713 /* Insert an instruction at the begining of a basic block, and update
6714 the BLOCK_HEAD if needed. */
6717 insert_insn_start_bb (insn
, bb
)
6721 /* Insert at start of successor block. */
6722 rtx prev
= PREV_INSN (bb
->head
);
6723 rtx before
= bb
->head
;
6726 if (GET_CODE (before
) != CODE_LABEL
6727 && (GET_CODE (before
) != NOTE
6728 || NOTE_LINE_NUMBER (before
) != NOTE_INSN_BASIC_BLOCK
))
6731 if (prev
== bb
->end
)
6733 before
= NEXT_INSN (before
);
6736 insn
= emit_insn_after (insn
, prev
);
6738 if (prev
== bb
->end
)
6741 set_block_for_new_insns (insn
, bb
);
6745 fprintf (gcse_file
, "STORE_MOTION insert store at start of BB %d:\n",
6747 print_inline_rtx (gcse_file
, insn
, 6);
6748 fprintf (gcse_file
, "\n");
6752 /* This routine will insert a store on an edge. EXPR is the ldst entry for
6753 the memory reference, and E is the edge to insert it on. Returns non-zero
6754 if an edge insertion was performed. */
6757 insert_store (expr
, e
)
6758 struct ls_expr
* expr
;
6765 /* We did all the deleted before this insert, so if we didn't delete a
6766 store, then we haven't set the reaching reg yet either. */
6767 if (expr
->reaching_reg
== NULL_RTX
)
6770 reg
= expr
->reaching_reg
;
6771 insn
= gen_move_insn (expr
->pattern
, reg
);
6773 /* If we are inserting this expression on ALL predecessor edges of a BB,
6774 insert it at the start of the BB, and reset the insert bits on the other
6775 edges so we don;t try to insert it on the other edges. */
6777 for (tmp
= e
->dest
->pred
; tmp
; tmp
= tmp
->pred_next
)
6779 int index
= EDGE_INDEX (edge_list
, tmp
->src
, tmp
->dest
);
6780 if (index
== EDGE_INDEX_NO_EDGE
)
6782 if (! TEST_BIT (pre_insert_map
[index
], expr
->index
))
6786 /* If tmp is NULL, we found an insertion on every edge, blank the
6787 insertion vector for these edges, and insert at the start of the BB. */
6788 if (!tmp
&& bb
!= EXIT_BLOCK_PTR
)
6790 for (tmp
= e
->dest
->pred
; tmp
; tmp
= tmp
->pred_next
)
6792 int index
= EDGE_INDEX (edge_list
, tmp
->src
, tmp
->dest
);
6793 RESET_BIT (pre_insert_map
[index
], expr
->index
);
6795 insert_insn_start_bb (insn
, bb
);
6799 /* We can't insert on this edge, so we'll insert at the head of the
6800 successors block. See Morgan, sec 10.5. */
6801 if ((e
->flags
& EDGE_ABNORMAL
) == EDGE_ABNORMAL
)
6803 insert_insn_start_bb (insn
, bb
);
6807 insert_insn_on_edge (insn
, e
);
6811 fprintf (gcse_file
, "STORE_MOTION insert insn on edge (%d, %d):\n",
6812 e
->src
->index
, e
->dest
->index
);
6813 print_inline_rtx (gcse_file
, insn
, 6);
6814 fprintf (gcse_file
, "\n");
6820 /* This routine will replace a store with a SET to a specified register. */
6823 replace_store_insn (reg
, del
, bb
)
6829 insn
= gen_move_insn (reg
, SET_SRC (PATTERN (del
)));
6830 insn
= emit_insn_after (insn
, del
);
6831 set_block_for_new_insns (insn
, bb
);
6836 "STORE_MOTION delete insn in BB %d:\n ", bb
->index
);
6837 print_inline_rtx (gcse_file
, del
, 6);
6838 fprintf(gcse_file
, "\nSTORE MOTION replaced with insn:\n ");
6839 print_inline_rtx (gcse_file
, insn
, 6);
6840 fprintf(gcse_file
, "\n");
6846 if (bb
->head
== del
)
6853 /* Delete a store, but copy the value that would have been stored into
6854 the reaching_reg for later storing. */
6857 delete_store (expr
, bb
)
6858 struct ls_expr
* expr
;
6863 if (expr
->reaching_reg
== NULL_RTX
)
6864 expr
->reaching_reg
= gen_reg_rtx (GET_MODE (expr
->pattern
));
6867 /* If there is more than 1 store, the earlier ones will be dead,
6868 but it doesn't hurt to replace them here. */
6869 reg
= expr
->reaching_reg
;
6871 for (i
= AVAIL_STORE_LIST (expr
); i
; i
= XEXP (i
, 1))
6874 if (BLOCK_FOR_INSN (del
) == bb
)
6876 /* We know there is only one since we deleted redundant
6877 ones during the available computation. */
6878 replace_store_insn (reg
, del
, bb
);
6884 /* Free memory used by store motion. */
6887 free_store_memory ()
6892 sbitmap_vector_free (ae_gen
);
6894 sbitmap_vector_free (ae_kill
);
6896 sbitmap_vector_free (transp
);
6898 sbitmap_vector_free (st_antloc
);
6900 sbitmap_vector_free (pre_insert_map
);
6902 sbitmap_vector_free (pre_delete_map
);
6903 if (reg_set_in_block
)
6904 sbitmap_vector_free (reg_set_in_block
);
6906 ae_gen
= ae_kill
= transp
= st_antloc
= NULL
;
6907 pre_insert_map
= pre_delete_map
= reg_set_in_block
= NULL
;
6910 /* Perform store motion. Much like gcse, except we move expressions the
6911 other way by looking at the flowgraph in reverse. */
6917 struct ls_expr
* ptr
;
6918 int update_flow
= 0;
6922 fprintf (gcse_file
, "before store motion\n");
6923 print_rtl (gcse_file
, get_insns ());
6927 init_alias_analysis ();
6929 /* Find all the stores that are live to the end of their block. */
6930 num_stores
= compute_store_table ();
6931 if (num_stores
== 0)
6933 sbitmap_vector_free (reg_set_in_block
);
6934 end_alias_analysis ();
6938 /* Now compute whats actually available to move. */
6939 add_noreturn_fake_exit_edges ();
6940 build_store_vectors ();
6942 edge_list
= pre_edge_rev_lcm (gcse_file
, num_stores
, transp
, ae_gen
,
6943 st_antloc
, ae_kill
, &pre_insert_map
,
6946 /* Now we want to insert the new stores which are going to be needed. */
6947 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6949 for (x
= 0; x
< n_basic_blocks
; x
++)
6950 if (TEST_BIT (pre_delete_map
[x
], ptr
->index
))
6951 delete_store (ptr
, BASIC_BLOCK (x
));
6953 for (x
= 0; x
< NUM_EDGES (edge_list
); x
++)
6954 if (TEST_BIT (pre_insert_map
[x
], ptr
->index
))
6955 update_flow
|= insert_store (ptr
, INDEX_EDGE (edge_list
, x
));
6959 commit_edge_insertions ();
6961 free_store_memory ();
6962 free_edge_list (edge_list
);
6963 remove_fake_edges ();
6964 end_alias_analysis ();