1 /* Global common subexpression elimination/Partial redundancy elimination
2 and global constant/copy propagation for GNU compiler.
3 Copyright (C) 1997, 1998, 1999, 2000, 2001 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
23 - reordering of memory allocation and freeing to be more space efficient
24 - do rough calc of how many regs are needed in each block, and a rough
25 calc of how many regs are available in each class and use that to
26 throttle back the code in cases where RTX_COST is minimal.
27 - a store to the same address as a load does not kill the load if the
28 source of the store is also the destination of the load. Handling this
29 allows more load motion, particularly out of loops.
30 - ability to realloc sbitmap vectors would allow one initial computation
31 of reg_set_in_block with only subsequent additions, rather than
32 recomputing it for each pass
36 /* References searched while implementing this.
38 Compilers Principles, Techniques and Tools
42 Global Optimization by Suppression of Partial Redundancies
44 communications of the acm, Vol. 22, Num. 2, Feb. 1979
46 A Portable Machine-Independent Global Optimizer - Design and Measurements
48 Stanford Ph.D. thesis, Dec. 1983
50 A Fast Algorithm for Code Movement Optimization
52 SIGPLAN Notices, Vol. 23, Num. 10, Oct. 1988
54 A Solution to a Problem with Morel and Renvoise's
55 Global Optimization by Suppression of Partial Redundancies
56 K-H Drechsler, M.P. Stadel
57 ACM TOPLAS, Vol. 10, Num. 4, Oct. 1988
59 Practical Adaptation of the Global Optimization
60 Algorithm of Morel and Renvoise
62 ACM TOPLAS, Vol. 13, Num. 2. Apr. 1991
64 Efficiently Computing Static Single Assignment Form and the Control
66 R. Cytron, J. Ferrante, B.K. Rosen, M.N. Wegman, and F.K. Zadeck
67 ACM TOPLAS, Vol. 13, Num. 4, Oct. 1991
70 J. Knoop, O. Ruthing, B. Steffen
71 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
73 What's In a Region? Or Computing Control Dependence Regions in Near-Linear
74 Time for Reducible Flow Control
76 ACM Letters on Programming Languages and Systems,
77 Vol. 2, Num. 1-4, Mar-Dec 1993
79 An Efficient Representation for Sparse Sets
80 Preston Briggs, Linda Torczon
81 ACM Letters on Programming Languages and Systems,
82 Vol. 2, Num. 1-4, Mar-Dec 1993
84 A Variation of Knoop, Ruthing, and Steffen's Lazy Code Motion
85 K-H Drechsler, M.P. Stadel
86 ACM SIGPLAN Notices, Vol. 28, Num. 5, May 1993
88 Partial Dead Code Elimination
89 J. Knoop, O. Ruthing, B. Steffen
90 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
92 Effective Partial Redundancy Elimination
93 P. Briggs, K.D. Cooper
94 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
96 The Program Structure Tree: Computing Control Regions in Linear Time
97 R. Johnson, D. Pearson, K. Pingali
98 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
100 Optimal Code Motion: Theory and Practice
101 J. Knoop, O. Ruthing, B. Steffen
102 ACM TOPLAS, Vol. 16, Num. 4, Jul. 1994
104 The power of assignment motion
105 J. Knoop, O. Ruthing, B. Steffen
106 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
108 Global code motion / global value numbering
110 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
112 Value Driven Redundancy Elimination
114 Rice University Ph.D. thesis, Apr. 1996
118 Massively Scalar Compiler Project, Rice University, Sep. 1996
120 High Performance Compilers for Parallel Computing
124 Advanced Compiler Design and Implementation
126 Morgan Kaufmann, 1997
128 Building an Optimizing Compiler
132 People wishing to speed up the code here should read:
133 Elimination Algorithms for Data Flow Analysis
134 B.G. Ryder, M.C. Paull
135 ACM Computing Surveys, Vol. 18, Num. 3, Sep. 1986
137 How to Analyze Large Programs Efficiently and Informatively
138 D.M. Dhamdhere, B.K. Rosen, F.K. Zadeck
139 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
141 People wishing to do something different can find various possibilities
142 in the above papers and elsewhere.
152 #include "hard-reg-set.h"
155 #include "insn-config.h"
157 #include "basic-block.h"
159 #include "function.h"
165 #define obstack_chunk_alloc gmalloc
166 #define obstack_chunk_free free
168 /* Propagate flow information through back edges and thus enable PRE's
169 moving loop invariant calculations out of loops.
171 Originally this tended to create worse overall code, but several
172 improvements during the development of PRE seem to have made following
173 back edges generally a win.
175 Note much of the loop invariant code motion done here would normally
176 be done by loop.c, which has more heuristics for when to move invariants
177 out of loops. At some point we might need to move some of those
178 heuristics into gcse.c. */
179 #define FOLLOW_BACK_EDGES 1
181 /* We support GCSE via Partial Redundancy Elimination. PRE optimizations
182 are a superset of those done by GCSE.
184 We perform the following steps:
186 1) Compute basic block information.
188 2) Compute table of places where registers are set.
190 3) Perform copy/constant propagation.
192 4) Perform global cse.
194 5) Perform another pass of copy/constant propagation.
196 Two passes of copy/constant propagation are done because the first one
197 enables more GCSE and the second one helps to clean up the copies that
198 GCSE creates. This is needed more for PRE than for Classic because Classic
199 GCSE will try to use an existing register containing the common
200 subexpression rather than create a new one. This is harder to do for PRE
201 because of the code motion (which Classic GCSE doesn't do).
203 Expressions we are interested in GCSE-ing are of the form
204 (set (pseudo-reg) (expression)).
205 Function want_to_gcse_p says what these are.
207 PRE handles moving invariant expressions out of loops (by treating them as
208 partially redundant).
210 Eventually it would be nice to replace cse.c/gcse.c with SSA (static single
211 assignment) based GVN (global value numbering). L. T. Simpson's paper
212 (Rice University) on value numbering is a useful reference for this.
214 **********************
216 We used to support multiple passes but there are diminishing returns in
217 doing so. The first pass usually makes 90% of the changes that are doable.
218 A second pass can make a few more changes made possible by the first pass.
219 Experiments show any further passes don't make enough changes to justify
222 A study of spec92 using an unlimited number of passes:
223 [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
224 [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
225 [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
227 It was found doing copy propagation between each pass enables further
230 PRE is quite expensive in complicated functions because the DFA can take
231 awhile to converge. Hence we only perform one pass. The parameter max-gcse-passes can
232 be modified if one wants to experiment.
234 **********************
236 The steps for PRE are:
238 1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
240 2) Perform the data flow analysis for PRE.
242 3) Delete the redundant instructions
244 4) Insert the required copies [if any] that make the partially
245 redundant instructions fully redundant.
247 5) For other reaching expressions, insert an instruction to copy the value
248 to a newly created pseudo that will reach the redundant instruction.
250 The deletion is done first so that when we do insertions we
251 know which pseudo reg to use.
253 Various papers have argued that PRE DFA is expensive (O(n^2)) and others
254 argue it is not. The number of iterations for the algorithm to converge
255 is typically 2-4 so I don't view it as that expensive (relatively speaking).
257 PRE GCSE depends heavily on the second CSE pass to clean up the copies
258 we create. To make an expression reach the place where it's redundant,
259 the result of the expression is copied to a new register, and the redundant
260 expression is deleted by replacing it with this new register. Classic GCSE
261 doesn't have this problem as much as it computes the reaching defs of
262 each register in each block and thus can try to use an existing register.
264 **********************
266 A fair bit of simplicity is created by creating small functions for simple
267 tasks, even when the function is only called in one place. This may
268 measurably slow things down [or may not] by creating more function call
269 overhead than is necessary. The source is laid out so that it's trivial
270 to make the affected functions inline so that one can measure what speed
271 up, if any, can be achieved, and maybe later when things settle things can
274 Help stamp out big monolithic functions! */
276 /* GCSE global vars. */
279 static FILE *gcse_file
;
281 /* Note whether or not we should run jump optimization after gcse. We
282 want to do this for two cases.
284 * If we changed any jumps via cprop.
286 * If we added any labels via edge splitting. */
288 static int run_jump_opt_after_gcse
;
290 /* Bitmaps are normally not included in debugging dumps.
291 However it's useful to be able to print them from GDB.
292 We could create special functions for this, but it's simpler to
293 just allow passing stderr to the dump_foo fns. Since stderr can
294 be a macro, we store a copy here. */
295 static FILE *debug_stderr
;
297 /* An obstack for our working variables. */
298 static struct obstack gcse_obstack
;
300 /* Non-zero for each mode that supports (set (reg) (reg)).
301 This is trivially true for integer and floating point values.
302 It may or may not be true for condition codes. */
303 static char can_copy_p
[(int) NUM_MACHINE_MODES
];
305 /* Non-zero if can_copy_p has been initialized. */
306 static int can_copy_init_p
;
308 struct reg_use
{rtx reg_rtx
; };
310 /* Hash table of expressions. */
314 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
316 /* Index in the available expression bitmaps. */
318 /* Next entry with the same hash. */
319 struct expr
*next_same_hash
;
320 /* List of anticipatable occurrences in basic blocks in the function.
321 An "anticipatable occurrence" is one that is the first occurrence in the
322 basic block, the operands are not modified in the basic block prior
323 to the occurrence and the output is not used between the start of
324 the block and the occurrence. */
325 struct occr
*antic_occr
;
326 /* List of available occurrence in basic blocks in the function.
327 An "available occurrence" is one that is the last occurrence in the
328 basic block and the operands are not modified by following statements in
329 the basic block [including this insn]. */
330 struct occr
*avail_occr
;
331 /* Non-null if the computation is PRE redundant.
332 The value is the newly created pseudo-reg to record a copy of the
333 expression in all the places that reach the redundant copy. */
337 /* Occurrence of an expression.
338 There is one per basic block. If a pattern appears more than once the
339 last appearance is used [or first for anticipatable expressions]. */
343 /* Next occurrence of this expression. */
345 /* The insn that computes the expression. */
347 /* Non-zero if this [anticipatable] occurrence has been deleted. */
349 /* Non-zero if this [available] occurrence has been copied to
351 /* ??? This is mutually exclusive with deleted_p, so they could share
356 /* Expression and copy propagation hash tables.
357 Each hash table is an array of buckets.
358 ??? It is known that if it were an array of entries, structure elements
359 `next_same_hash' and `bitmap_index' wouldn't be necessary. However, it is
360 not clear whether in the final analysis a sufficient amount of memory would
361 be saved as the size of the available expression bitmaps would be larger
362 [one could build a mapping table without holes afterwards though].
363 Someday I'll perform the computation and figure it out. */
365 /* Total size of the expression hash table, in elements. */
366 static unsigned int expr_hash_table_size
;
369 This is an array of `expr_hash_table_size' elements. */
370 static struct expr
**expr_hash_table
;
372 /* Total size of the copy propagation hash table, in elements. */
373 static unsigned int set_hash_table_size
;
376 This is an array of `set_hash_table_size' elements. */
377 static struct expr
**set_hash_table
;
379 /* Mapping of uids to cuids.
380 Only real insns get cuids. */
381 static int *uid_cuid
;
383 /* Highest UID in UID_CUID. */
386 /* Get the cuid of an insn. */
387 #ifdef ENABLE_CHECKING
388 #define INSN_CUID(INSN) (INSN_UID (INSN) > max_uid ? (abort (), 0) : uid_cuid[INSN_UID (INSN)])
390 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
393 /* Number of cuids. */
396 /* Mapping of cuids to insns. */
397 static rtx
*cuid_insn
;
399 /* Get insn from cuid. */
400 #define CUID_INSN(CUID) (cuid_insn[CUID])
402 /* Maximum register number in function prior to doing gcse + 1.
403 Registers created during this pass have regno >= max_gcse_regno.
404 This is named with "gcse" to not collide with global of same name. */
405 static unsigned int max_gcse_regno
;
407 /* Maximum number of cse-able expressions found. */
410 /* Maximum number of assignments for copy propagation found. */
413 /* Table of registers that are modified.
415 For each register, each element is a list of places where the pseudo-reg
418 For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
419 requires knowledge of which blocks kill which regs [and thus could use
420 a bitmap instead of the lists `reg_set_table' uses].
422 `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
423 num-regs) [however perhaps it may be useful to keep the data as is]. One
424 advantage of recording things this way is that `reg_set_table' is fairly
425 sparse with respect to pseudo regs but for hard regs could be fairly dense
426 [relatively speaking]. And recording sets of pseudo-regs in lists speeds
427 up functions like compute_transp since in the case of pseudo-regs we only
428 need to iterate over the number of times a pseudo-reg is set, not over the
429 number of basic blocks [clearly there is a bit of a slow down in the cases
430 where a pseudo is set more than once in a block, however it is believed
431 that the net effect is to speed things up]. This isn't done for hard-regs
432 because recording call-clobbered hard-regs in `reg_set_table' at each
433 function call can consume a fair bit of memory, and iterating over
434 hard-regs stored this way in compute_transp will be more expensive. */
436 typedef struct reg_set
438 /* The next setting of this register. */
439 struct reg_set
*next
;
440 /* The insn where it was set. */
444 static reg_set
**reg_set_table
;
446 /* Size of `reg_set_table'.
447 The table starts out at max_gcse_regno + slop, and is enlarged as
449 static int reg_set_table_size
;
451 /* Amount to grow `reg_set_table' by when it's full. */
452 #define REG_SET_TABLE_SLOP 100
454 /* This is a list of expressions which are MEMs and will be used by load
456 Load motion tracks MEMs which aren't killed by
457 anything except itself. (ie, loads and stores to a single location).
458 We can then allow movement of these MEM refs with a little special
459 allowance. (all stores copy the same value to the reaching reg used
460 for the loads). This means all values used to store into memory must have
461 no side effects so we can re-issue the setter value.
462 Store Motion uses this structure as an expression table to track stores
463 which look interesting, and might be moveable towards the exit block. */
467 struct expr
* expr
; /* Gcse expression reference for LM. */
468 rtx pattern
; /* Pattern of this mem. */
469 rtx loads
; /* INSN list of loads seen. */
470 rtx stores
; /* INSN list of stores seen. */
471 struct ls_expr
* next
; /* Next in the list. */
472 int invalid
; /* Invalid for some reason. */
473 int index
; /* If it maps to a bitmap index. */
474 int hash_index
; /* Index when in a hash table. */
475 rtx reaching_reg
; /* Register to use when re-writing. */
478 /* Head of the list of load/store memory refs. */
479 static struct ls_expr
* pre_ldst_mems
= NULL
;
481 /* Bitmap containing one bit for each register in the program.
482 Used when performing GCSE to track which registers have been set since
483 the start of the basic block. */
484 static regset 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
;
496 bitmap modify_mem_list_set
;
498 /* This array parallels modify_mem_list, but is kept canonicalized. */
499 static rtx
* canon_modify_mem_list
;
500 bitmap canon_modify_mem_list_set
;
501 /* Various variables for statistics gathering. */
503 /* Memory used in a pass.
504 This isn't intended to be absolutely precise. Its intent is only
505 to keep an eye on memory usage. */
506 static int bytes_used
;
508 /* GCSE substitutions made. */
509 static int gcse_subst_count
;
510 /* Number of copy instructions created. */
511 static int gcse_create_count
;
512 /* Number of constants propagated. */
513 static int const_prop_count
;
514 /* Number of copys propagated. */
515 static int copy_prop_count
;
517 /* These variables are used by classic GCSE.
518 Normally they'd be defined a bit later, but `rd_gen' needs to
519 be declared sooner. */
521 /* Each block has a bitmap of each type.
522 The length of each blocks bitmap is:
524 max_cuid - for reaching definitions
525 n_exprs - for available expressions
527 Thus we view the bitmaps as 2 dimensional arrays. i.e.
528 rd_kill[block_num][cuid_num]
529 ae_kill[block_num][expr_num] */
531 /* For reaching defs */
532 static sbitmap
*rd_kill
, *rd_gen
, *reaching_defs
, *rd_out
;
534 /* for available exprs */
535 static sbitmap
*ae_kill
, *ae_gen
, *ae_in
, *ae_out
;
537 /* Objects of this type are passed around by the null-pointer check
539 struct null_pointer_info
541 /* The basic block being processed. */
543 /* The first register to be handled in this pass. */
544 unsigned int min_reg
;
545 /* One greater than the last register to be handled in this pass. */
546 unsigned int max_reg
;
547 sbitmap
*nonnull_local
;
548 sbitmap
*nonnull_killed
;
551 static void compute_can_copy
PARAMS ((void));
552 static char *gmalloc
PARAMS ((unsigned int));
553 static char *grealloc
PARAMS ((char *, unsigned int));
554 static char *gcse_alloc
PARAMS ((unsigned long));
555 static void alloc_gcse_mem
PARAMS ((rtx
));
556 static void free_gcse_mem
PARAMS ((void));
557 static void alloc_reg_set_mem
PARAMS ((int));
558 static void free_reg_set_mem
PARAMS ((void));
559 static int get_bitmap_width
PARAMS ((int, int, int));
560 static void record_one_set
PARAMS ((int, rtx
));
561 static void record_set_info
PARAMS ((rtx
, rtx
, void *));
562 static void compute_sets
PARAMS ((rtx
));
563 static void hash_scan_insn
PARAMS ((rtx
, int, int));
564 static void hash_scan_set
PARAMS ((rtx
, rtx
, int));
565 static void hash_scan_clobber
PARAMS ((rtx
, rtx
));
566 static void hash_scan_call
PARAMS ((rtx
, rtx
));
567 static int want_to_gcse_p
PARAMS ((rtx
));
568 static int oprs_unchanged_p
PARAMS ((rtx
, rtx
, int));
569 static int oprs_anticipatable_p
PARAMS ((rtx
, rtx
));
570 static int oprs_available_p
PARAMS ((rtx
, rtx
));
571 static void insert_expr_in_table
PARAMS ((rtx
, enum machine_mode
, rtx
,
573 static void insert_set_in_table
PARAMS ((rtx
, rtx
));
574 static unsigned int hash_expr
PARAMS ((rtx
, enum machine_mode
, int *, int));
575 static unsigned int hash_expr_1
PARAMS ((rtx
, enum machine_mode
, int *));
576 static unsigned int hash_string_1
PARAMS ((const char *));
577 static unsigned int hash_set
PARAMS ((int, int));
578 static int expr_equiv_p
PARAMS ((rtx
, rtx
));
579 static void record_last_reg_set_info
PARAMS ((rtx
, int));
580 static void record_last_mem_set_info
PARAMS ((rtx
));
581 static void record_last_set_info
PARAMS ((rtx
, rtx
, void *));
582 static void compute_hash_table
PARAMS ((int));
583 static void alloc_set_hash_table
PARAMS ((int));
584 static void free_set_hash_table
PARAMS ((void));
585 static void compute_set_hash_table
PARAMS ((void));
586 static void alloc_expr_hash_table
PARAMS ((unsigned int));
587 static void free_expr_hash_table
PARAMS ((void));
588 static void compute_expr_hash_table
PARAMS ((void));
589 static void dump_hash_table
PARAMS ((FILE *, const char *, struct expr
**,
591 static struct expr
*lookup_expr
PARAMS ((rtx
));
592 static struct expr
*lookup_set
PARAMS ((unsigned int, rtx
));
593 static struct expr
*next_set
PARAMS ((unsigned int, struct expr
*));
594 static void reset_opr_set_tables
PARAMS ((void));
595 static int oprs_not_set_p
PARAMS ((rtx
, rtx
));
596 static void mark_call
PARAMS ((rtx
));
597 static void mark_set
PARAMS ((rtx
, rtx
));
598 static void mark_clobber
PARAMS ((rtx
, rtx
));
599 static void mark_oprs_set
PARAMS ((rtx
));
600 static void alloc_cprop_mem
PARAMS ((int, int));
601 static void free_cprop_mem
PARAMS ((void));
602 static void compute_transp
PARAMS ((rtx
, int, sbitmap
*, int));
603 static void compute_transpout
PARAMS ((void));
604 static void compute_local_properties
PARAMS ((sbitmap
*, sbitmap
*, sbitmap
*,
606 static void compute_cprop_data
PARAMS ((void));
607 static void find_used_regs
PARAMS ((rtx
*, void *));
608 static int try_replace_reg
PARAMS ((rtx
, rtx
, rtx
));
609 static struct expr
*find_avail_set
PARAMS ((int, rtx
));
610 static int cprop_jump
PARAMS ((basic_block
, rtx
, rtx
, rtx
));
612 static int cprop_cc0_jump
PARAMS ((basic_block
, rtx
, struct reg_use
*, rtx
));
614 static void mems_conflict_for_gcse_p
PARAMS ((rtx
, rtx
, void *));
615 static int load_killed_in_block_p
PARAMS ((basic_block
, int, rtx
, int));
616 static void canon_list_insert
PARAMS ((rtx
, rtx
, void *));
617 static int cprop_insn
PARAMS ((basic_block
, rtx
, int));
618 static int cprop
PARAMS ((int));
619 static int one_cprop_pass
PARAMS ((int, int));
620 static void alloc_pre_mem
PARAMS ((int, int));
621 static void free_pre_mem
PARAMS ((void));
622 static void compute_pre_data
PARAMS ((void));
623 static int pre_expr_reaches_here_p
PARAMS ((basic_block
, struct expr
*,
625 static void insert_insn_end_bb
PARAMS ((struct expr
*, basic_block
, int));
626 static void pre_insert_copy_insn
PARAMS ((struct expr
*, rtx
));
627 static void pre_insert_copies
PARAMS ((void));
628 static int pre_delete
PARAMS ((void));
629 static int pre_gcse
PARAMS ((void));
630 static int one_pre_gcse_pass
PARAMS ((int));
631 static void add_label_notes
PARAMS ((rtx
, rtx
));
632 static void alloc_code_hoist_mem
PARAMS ((int, int));
633 static void free_code_hoist_mem
PARAMS ((void));
634 static void compute_code_hoist_vbeinout
PARAMS ((void));
635 static void compute_code_hoist_data
PARAMS ((void));
636 static int hoist_expr_reaches_here_p
PARAMS ((basic_block
, int, basic_block
,
638 static void hoist_code
PARAMS ((void));
639 static int one_code_hoisting_pass
PARAMS ((void));
640 static void alloc_rd_mem
PARAMS ((int, int));
641 static void free_rd_mem
PARAMS ((void));
642 static void handle_rd_kill_set
PARAMS ((rtx
, int, basic_block
));
643 static void compute_kill_rd
PARAMS ((void));
644 static void compute_rd
PARAMS ((void));
645 static void alloc_avail_expr_mem
PARAMS ((int, int));
646 static void free_avail_expr_mem
PARAMS ((void));
647 static void compute_ae_gen
PARAMS ((void));
648 static int expr_killed_p
PARAMS ((rtx
, basic_block
));
649 static void compute_ae_kill
PARAMS ((sbitmap
*, sbitmap
*));
650 static int expr_reaches_here_p
PARAMS ((struct occr
*, struct expr
*,
652 static rtx computing_insn
PARAMS ((struct expr
*, rtx
));
653 static int def_reaches_here_p
PARAMS ((rtx
, rtx
));
654 static int can_disregard_other_sets
PARAMS ((struct reg_set
**, rtx
, int));
655 static int handle_avail_expr
PARAMS ((rtx
, struct expr
*));
656 static int classic_gcse
PARAMS ((void));
657 static int one_classic_gcse_pass
PARAMS ((int));
658 static void invalidate_nonnull_info
PARAMS ((rtx
, rtx
, void *));
659 static void delete_null_pointer_checks_1
PARAMS ((varray_type
*, unsigned int *,
660 sbitmap
*, sbitmap
*,
661 struct null_pointer_info
*));
662 static rtx process_insert_insn
PARAMS ((struct expr
*));
663 static int pre_edge_insert
PARAMS ((struct edge_list
*, struct expr
**));
664 static int expr_reaches_here_p_work
PARAMS ((struct occr
*, struct expr
*,
665 basic_block
, int, char *));
666 static int pre_expr_reaches_here_p_work
PARAMS ((basic_block
, struct expr
*,
667 basic_block
, char *));
668 static struct ls_expr
* ldst_entry
PARAMS ((rtx
));
669 static void free_ldst_entry
PARAMS ((struct ls_expr
*));
670 static void free_ldst_mems
PARAMS ((void));
671 static void print_ldst_list
PARAMS ((FILE *));
672 static struct ls_expr
* find_rtx_in_ldst
PARAMS ((rtx
));
673 static int enumerate_ldsts
PARAMS ((void));
674 static inline struct ls_expr
* first_ls_expr
PARAMS ((void));
675 static inline struct ls_expr
* next_ls_expr
PARAMS ((struct ls_expr
*));
676 static int simple_mem
PARAMS ((rtx
));
677 static void invalidate_any_buried_refs
PARAMS ((rtx
));
678 static void compute_ld_motion_mems
PARAMS ((void));
679 static void trim_ld_motion_mems
PARAMS ((void));
680 static void update_ld_motion_stores
PARAMS ((struct expr
*));
681 static void reg_set_info
PARAMS ((rtx
, rtx
, void *));
682 static int store_ops_ok
PARAMS ((rtx
, basic_block
));
683 static void find_moveable_store
PARAMS ((rtx
));
684 static int compute_store_table
PARAMS ((void));
685 static int load_kills_store
PARAMS ((rtx
, rtx
));
686 static int find_loads
PARAMS ((rtx
, rtx
));
687 static int store_killed_in_insn
PARAMS ((rtx
, rtx
));
688 static int store_killed_after
PARAMS ((rtx
, rtx
, basic_block
));
689 static int store_killed_before
PARAMS ((rtx
, rtx
, basic_block
));
690 static void build_store_vectors
PARAMS ((void));
691 static void insert_insn_start_bb
PARAMS ((rtx
, basic_block
));
692 static int insert_store
PARAMS ((struct ls_expr
*, edge
));
693 static void replace_store_insn
PARAMS ((rtx
, rtx
, basic_block
));
694 static void delete_store
PARAMS ((struct ls_expr
*,
696 static void free_store_memory
PARAMS ((void));
697 static void store_motion
PARAMS ((void));
698 static void clear_modify_mem_tables
PARAMS ((void));
699 static void free_modify_mem_tables
PARAMS ((void));
701 /* Entry point for global common subexpression elimination.
702 F is the first instruction in the function. */
710 /* Bytes used at start of pass. */
711 int initial_bytes_used
;
712 /* Maximum number of bytes used by a pass. */
714 /* Point to release obstack data from for each pass. */
715 char *gcse_obstack_bottom
;
717 /* Insertion of instructions on edges can create new basic blocks; we
718 need the original basic block count so that we can properly deallocate
719 arrays sized on the number of basic blocks originally in the cfg. */
721 /* We do not construct an accurate cfg in functions which call
722 setjmp, so just punt to be safe. */
723 if (current_function_calls_setjmp
)
726 /* Assume that we do not need to run jump optimizations after gcse. */
727 run_jump_opt_after_gcse
= 0;
729 /* For calling dump_foo fns from gdb. */
730 debug_stderr
= stderr
;
733 /* Identify the basic block information for this function, including
734 successors and predecessors. */
735 max_gcse_regno
= max_reg_num ();
738 dump_flow_info (file
);
740 orig_bb_count
= n_basic_blocks
;
741 /* Return if there's nothing to do. */
742 if (n_basic_blocks
<= 1)
745 /* Trying to perform global optimizations on flow graphs which have
746 a high connectivity will take a long time and is unlikely to be
749 In normal circumstances a cfg should have about twice as many edges
750 as blocks. But we do not want to punish small functions which have
751 a couple switch statements. So we require a relatively large number
752 of basic blocks and the ratio of edges to blocks to be high. */
753 if (n_basic_blocks
> 1000 && n_edges
/ n_basic_blocks
>= 20)
755 if (warn_disabled_optimization
)
756 warning ("GCSE disabled: %d > 1000 basic blocks and %d >= 20 edges/basic block",
757 n_basic_blocks
, n_edges
/ n_basic_blocks
);
761 /* If allocating memory for the cprop bitmap would take up too much
762 storage it's better just to disable the optimization. */
764 * SBITMAP_SET_SIZE (max_gcse_regno
)
765 * sizeof (SBITMAP_ELT_TYPE
)) > MAX_GCSE_MEMORY
)
767 if (warn_disabled_optimization
)
768 warning ("GCSE disabled: %d basic blocks and %d registers",
769 n_basic_blocks
, max_gcse_regno
);
774 /* See what modes support reg/reg copy operations. */
775 if (! can_copy_init_p
)
781 gcc_obstack_init (&gcse_obstack
);
785 init_alias_analysis ();
786 /* Record where pseudo-registers are set. This data is kept accurate
787 during each pass. ??? We could also record hard-reg information here
788 [since it's unchanging], however it is currently done during hash table
791 It may be tempting to compute MEM set information here too, but MEM sets
792 will be subject to code motion one day and thus we need to compute
793 information about memory sets when we build the hash tables. */
795 alloc_reg_set_mem (max_gcse_regno
);
799 initial_bytes_used
= bytes_used
;
801 gcse_obstack_bottom
= gcse_alloc (1);
803 while (changed
&& pass
< MAX_GCSE_PASSES
)
807 fprintf (file
, "GCSE pass %d\n\n", pass
+ 1);
809 /* Initialize bytes_used to the space for the pred/succ lists,
810 and the reg_set_table data. */
811 bytes_used
= initial_bytes_used
;
813 /* Each pass may create new registers, so recalculate each time. */
814 max_gcse_regno
= max_reg_num ();
818 /* Don't allow constant propagation to modify jumps
820 changed
= one_cprop_pass (pass
+ 1, 0);
823 changed
|= one_classic_gcse_pass (pass
+ 1);
826 changed
|= one_pre_gcse_pass (pass
+ 1);
827 /* We may have just created new basic blocks. Release and
828 recompute various things which are sized on the number of
832 free_modify_mem_tables ();
834 = (rtx
*) gmalloc (n_basic_blocks
* sizeof (rtx
*));
835 canon_modify_mem_list
836 = (rtx
*) gmalloc (n_basic_blocks
* sizeof (rtx
*));
837 memset ((char *) modify_mem_list
, 0, n_basic_blocks
* sizeof (rtx
*));
838 memset ((char *) canon_modify_mem_list
, 0, n_basic_blocks
* sizeof (rtx
*));
839 orig_bb_count
= n_basic_blocks
;
842 alloc_reg_set_mem (max_reg_num ());
844 run_jump_opt_after_gcse
= 1;
847 if (max_pass_bytes
< bytes_used
)
848 max_pass_bytes
= bytes_used
;
850 /* Free up memory, then reallocate for code hoisting. We can
851 not re-use the existing allocated memory because the tables
852 will not have info for the insns or registers created by
853 partial redundancy elimination. */
856 /* It does not make sense to run code hoisting unless we optimizing
857 for code size -- it rarely makes programs faster, and can make
858 them bigger if we did partial redundancy elimination (when optimizing
859 for space, we use a classic gcse algorithm instead of partial
860 redundancy algorithms). */
863 max_gcse_regno
= max_reg_num ();
865 changed
|= one_code_hoisting_pass ();
868 if (max_pass_bytes
< bytes_used
)
869 max_pass_bytes
= bytes_used
;
874 fprintf (file
, "\n");
878 obstack_free (&gcse_obstack
, gcse_obstack_bottom
);
882 /* Do one last pass of copy propagation, including cprop into
883 conditional jumps. */
885 max_gcse_regno
= max_reg_num ();
887 /* This time, go ahead and allow cprop to alter jumps. */
888 one_cprop_pass (pass
+ 1, 1);
893 fprintf (file
, "GCSE of %s: %d basic blocks, ",
894 current_function_name
, n_basic_blocks
);
895 fprintf (file
, "%d pass%s, %d bytes\n\n",
896 pass
, pass
> 1 ? "es" : "", max_pass_bytes
);
899 obstack_free (&gcse_obstack
, NULL
);
901 /* We are finished with alias. */
902 end_alias_analysis ();
903 allocate_reg_info (max_reg_num (), FALSE
, FALSE
);
905 if (!optimize_size
&& flag_gcse_sm
)
907 /* Record where pseudo-registers are set. */
908 return run_jump_opt_after_gcse
;
911 /* Misc. utilities. */
913 /* Compute which modes support reg/reg copy operations. */
919 #ifndef AVOID_CCMODE_COPIES
922 memset (can_copy_p
, 0, NUM_MACHINE_MODES
);
925 for (i
= 0; i
< NUM_MACHINE_MODES
; i
++)
926 if (GET_MODE_CLASS (i
) == MODE_CC
)
928 #ifdef AVOID_CCMODE_COPIES
931 reg
= gen_rtx_REG ((enum machine_mode
) i
, LAST_VIRTUAL_REGISTER
+ 1);
932 insn
= emit_insn (gen_rtx_SET (VOIDmode
, reg
, reg
));
933 if (recog (PATTERN (insn
), insn
, NULL
) >= 0)
943 /* Cover function to xmalloc to record bytes allocated. */
950 return xmalloc (size
);
953 /* Cover function to xrealloc.
954 We don't record the additional size since we don't know it.
955 It won't affect memory usage stats much anyway. */
962 return xrealloc (ptr
, size
);
965 /* Cover function to obstack_alloc.
966 We don't need to record the bytes allocated here since
967 obstack_chunk_alloc is set to gmalloc. */
973 return (char *) obstack_alloc (&gcse_obstack
, size
);
976 /* Allocate memory for the cuid mapping array,
977 and reg/memory set tracking tables.
979 This is called at the start of each pass. */
988 /* Find the largest UID and create a mapping from UIDs to CUIDs.
989 CUIDs are like UIDs except they increase monotonically, have no gaps,
990 and only apply to real insns. */
992 max_uid
= get_max_uid ();
993 n
= (max_uid
+ 1) * sizeof (int);
994 uid_cuid
= (int *) gmalloc (n
);
995 memset ((char *) uid_cuid
, 0, n
);
996 for (insn
= f
, i
= 0; insn
; insn
= NEXT_INSN (insn
))
999 uid_cuid
[INSN_UID (insn
)] = i
++;
1001 uid_cuid
[INSN_UID (insn
)] = i
;
1004 /* Create a table mapping cuids to insns. */
1007 n
= (max_cuid
+ 1) * sizeof (rtx
);
1008 cuid_insn
= (rtx
*) gmalloc (n
);
1009 memset ((char *) cuid_insn
, 0, n
);
1010 for (insn
= f
, i
= 0; insn
; insn
= NEXT_INSN (insn
))
1012 CUID_INSN (i
++) = insn
;
1014 /* Allocate vars to track sets of regs. */
1015 reg_set_bitmap
= BITMAP_XMALLOC ();
1017 /* Allocate vars to track sets of regs, memory per block. */
1018 reg_set_in_block
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
,
1020 /* Allocate array to keep a list of insns which modify memory in each
1022 modify_mem_list
= (rtx
*) gmalloc (n_basic_blocks
* sizeof (rtx
*));
1023 canon_modify_mem_list
= (rtx
*) gmalloc (n_basic_blocks
* sizeof (rtx
*));
1024 memset ((char *) modify_mem_list
, 0, n_basic_blocks
* sizeof (rtx
*));
1025 memset ((char *) canon_modify_mem_list
, 0, n_basic_blocks
* sizeof (rtx
*));
1026 modify_mem_list_set
= BITMAP_XMALLOC ();
1027 canon_modify_mem_list_set
= BITMAP_XMALLOC ();
1030 /* Free memory allocated by alloc_gcse_mem. */
1038 BITMAP_XFREE (reg_set_bitmap
);
1040 sbitmap_vector_free (reg_set_in_block
);
1041 free_modify_mem_tables ();
1042 BITMAP_XFREE (modify_mem_list_set
);
1043 BITMAP_XFREE (canon_modify_mem_list_set
);
1046 /* Many of the global optimization algorithms work by solving dataflow
1047 equations for various expressions. Initially, some local value is
1048 computed for each expression in each block. Then, the values across the
1049 various blocks are combined (by following flow graph edges) to arrive at
1050 global values. Conceptually, each set of equations is independent. We
1051 may therefore solve all the equations in parallel, solve them one at a
1052 time, or pick any intermediate approach.
1054 When you're going to need N two-dimensional bitmaps, each X (say, the
1055 number of blocks) by Y (say, the number of expressions), call this
1056 function. It's not important what X and Y represent; only that Y
1057 correspond to the things that can be done in parallel. This function will
1058 return an appropriate chunking factor C; you should solve C sets of
1059 equations in parallel. By going through this function, we can easily
1060 trade space against time; by solving fewer equations in parallel we use
1064 get_bitmap_width (n
, x
, y
)
1069 /* It's not really worth figuring out *exactly* how much memory will
1070 be used by a particular choice. The important thing is to get
1071 something approximately right. */
1072 size_t max_bitmap_memory
= 10 * 1024 * 1024;
1074 /* The number of bytes we'd use for a single column of minimum
1076 size_t column_size
= n
* x
* sizeof (SBITMAP_ELT_TYPE
);
1078 /* Often, it's reasonable just to solve all the equations in
1080 if (column_size
* SBITMAP_SET_SIZE (y
) <= max_bitmap_memory
)
1083 /* Otherwise, pick the largest width we can, without going over the
1085 return SBITMAP_ELT_BITS
* ((max_bitmap_memory
+ column_size
- 1)
1089 /* Compute the local properties of each recorded expression.
1091 Local properties are those that are defined by the block, irrespective of
1094 An expression is transparent in a block if its operands are not modified
1097 An expression is computed (locally available) in a block if it is computed
1098 at least once and expression would contain the same value if the
1099 computation was moved to the end of the block.
1101 An expression is locally anticipatable in a block if it is computed at
1102 least once and expression would contain the same value if the computation
1103 was moved to the beginning of the block.
1105 We call this routine for cprop, pre and code hoisting. They all compute
1106 basically the same information and thus can easily share this code.
1108 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
1109 properties. If NULL, then it is not necessary to compute or record that
1110 particular property.
1112 SETP controls which hash table to look at. If zero, this routine looks at
1113 the expr hash table; if nonzero this routine looks at the set hash table.
1114 Additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
1118 compute_local_properties (transp
, comp
, antloc
, setp
)
1124 unsigned int i
, hash_table_size
;
1125 struct expr
**hash_table
;
1127 /* Initialize any bitmaps that were passed in. */
1131 sbitmap_vector_zero (transp
, n_basic_blocks
);
1133 sbitmap_vector_ones (transp
, n_basic_blocks
);
1137 sbitmap_vector_zero (comp
, n_basic_blocks
);
1139 sbitmap_vector_zero (antloc
, n_basic_blocks
);
1141 /* We use the same code for cprop, pre and hoisting. For cprop
1142 we care about the set hash table, for pre and hoisting we
1143 care about the expr hash table. */
1144 hash_table_size
= setp
? set_hash_table_size
: expr_hash_table_size
;
1145 hash_table
= setp
? set_hash_table
: expr_hash_table
;
1147 for (i
= 0; i
< hash_table_size
; i
++)
1151 for (expr
= hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
1153 int indx
= expr
->bitmap_index
;
1156 /* The expression is transparent in this block if it is not killed.
1157 We start by assuming all are transparent [none are killed], and
1158 then reset the bits for those that are. */
1160 compute_transp (expr
->expr
, indx
, transp
, setp
);
1162 /* The occurrences recorded in antic_occr are exactly those that
1163 we want to set to non-zero in ANTLOC. */
1165 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
1167 SET_BIT (antloc
[BLOCK_NUM (occr
->insn
)], indx
);
1169 /* While we're scanning the table, this is a good place to
1171 occr
->deleted_p
= 0;
1174 /* The occurrences recorded in avail_occr are exactly those that
1175 we want to set to non-zero in COMP. */
1177 for (occr
= expr
->avail_occr
; occr
!= NULL
; occr
= occr
->next
)
1179 SET_BIT (comp
[BLOCK_NUM (occr
->insn
)], indx
);
1181 /* While we're scanning the table, this is a good place to
1186 /* While we're scanning the table, this is a good place to
1188 expr
->reaching_reg
= 0;
1193 /* Register set information.
1195 `reg_set_table' records where each register is set or otherwise
1198 static struct obstack reg_set_obstack
;
1201 alloc_reg_set_mem (n_regs
)
1206 reg_set_table_size
= n_regs
+ REG_SET_TABLE_SLOP
;
1207 n
= reg_set_table_size
* sizeof (struct reg_set
*);
1208 reg_set_table
= (struct reg_set
**) gmalloc (n
);
1209 memset ((char *) reg_set_table
, 0, n
);
1211 gcc_obstack_init (®_set_obstack
);
1217 free (reg_set_table
);
1218 obstack_free (®_set_obstack
, NULL
);
1221 /* Record REGNO in the reg_set table. */
1224 record_one_set (regno
, insn
)
1228 /* Allocate a new reg_set element and link it onto the list. */
1229 struct reg_set
*new_reg_info
;
1231 /* If the table isn't big enough, enlarge it. */
1232 if (regno
>= reg_set_table_size
)
1234 int new_size
= regno
+ REG_SET_TABLE_SLOP
;
1237 = (struct reg_set
**) grealloc ((char *) reg_set_table
,
1238 new_size
* sizeof (struct reg_set
*));
1239 memset ((char *) (reg_set_table
+ reg_set_table_size
), 0,
1240 (new_size
- reg_set_table_size
) * sizeof (struct reg_set
*));
1241 reg_set_table_size
= new_size
;
1244 new_reg_info
= (struct reg_set
*) obstack_alloc (®_set_obstack
,
1245 sizeof (struct reg_set
));
1246 bytes_used
+= sizeof (struct reg_set
);
1247 new_reg_info
->insn
= insn
;
1248 new_reg_info
->next
= reg_set_table
[regno
];
1249 reg_set_table
[regno
] = new_reg_info
;
1252 /* Called from compute_sets via note_stores to handle one SET or CLOBBER in
1253 an insn. The DATA is really the instruction in which the SET is
1257 record_set_info (dest
, setter
, data
)
1258 rtx dest
, setter ATTRIBUTE_UNUSED
;
1261 rtx record_set_insn
= (rtx
) data
;
1263 if (GET_CODE (dest
) == REG
&& REGNO (dest
) >= FIRST_PSEUDO_REGISTER
)
1264 record_one_set (REGNO (dest
), record_set_insn
);
1267 /* Scan the function and record each set of each pseudo-register.
1269 This is called once, at the start of the gcse pass. See the comments for
1270 `reg_set_table' for further documenation. */
1278 for (insn
= f
; insn
!= 0; insn
= NEXT_INSN (insn
))
1280 note_stores (PATTERN (insn
), record_set_info
, insn
);
1283 /* Hash table support. */
1285 /* For each register, the cuid of the first/last insn in the block
1286 that set it, or -1 if not set. */
1287 #define NEVER_SET -1
1289 struct reg_avail_info
1296 static struct reg_avail_info
*reg_avail_info
;
1297 static int current_bb
;
1300 /* See whether X, the source of a set, is something we want to consider for
1307 static rtx test_insn
= 0;
1308 int num_clobbers
= 0;
1311 switch (GET_CODE (x
))
1324 /* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
1325 if (general_operand (x
, GET_MODE (x
)))
1327 else if (GET_MODE (x
) == VOIDmode
)
1330 /* Otherwise, check if we can make a valid insn from it. First initialize
1331 our test insn if we haven't already. */
1335 = make_insn_raw (gen_rtx_SET (VOIDmode
,
1336 gen_rtx_REG (word_mode
,
1337 FIRST_PSEUDO_REGISTER
* 2),
1339 NEXT_INSN (test_insn
) = PREV_INSN (test_insn
) = 0;
1340 ggc_add_rtx_root (&test_insn
, 1);
1343 /* Now make an insn like the one we would make when GCSE'ing and see if
1345 PUT_MODE (SET_DEST (PATTERN (test_insn
)), GET_MODE (x
));
1346 SET_SRC (PATTERN (test_insn
)) = x
;
1347 return ((icode
= recog (PATTERN (test_insn
), test_insn
, &num_clobbers
)) >= 0
1348 && (num_clobbers
== 0 || ! added_clobbers_hard_reg_p (icode
)));
1351 /* Return non-zero if the operands of expression X are unchanged from the
1352 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1353 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1356 oprs_unchanged_p (x
, insn
, avail_p
)
1367 code
= GET_CODE (x
);
1372 struct reg_avail_info
*info
= ®_avail_info
[REGNO (x
)];
1374 if (info
->last_bb
!= current_bb
)
1377 return info
->last_set
< INSN_CUID (insn
);
1379 return info
->first_set
>= INSN_CUID (insn
);
1383 if (load_killed_in_block_p (BASIC_BLOCK (current_bb
), INSN_CUID (insn
),
1387 return oprs_unchanged_p (XEXP (x
, 0), insn
, avail_p
);
1412 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
1416 /* If we are about to do the last recursive call needed at this
1417 level, change it into iteration. This function is called enough
1420 return oprs_unchanged_p (XEXP (x
, i
), insn
, avail_p
);
1422 else if (! oprs_unchanged_p (XEXP (x
, i
), insn
, avail_p
))
1425 else if (fmt
[i
] == 'E')
1426 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1427 if (! oprs_unchanged_p (XVECEXP (x
, i
, j
), insn
, avail_p
))
1434 /* Used for communication between mems_conflict_for_gcse_p and
1435 load_killed_in_block_p. Nonzero if mems_conflict_for_gcse_p finds a
1436 conflict between two memory references. */
1437 static int gcse_mems_conflict_p
;
1439 /* Used for communication between mems_conflict_for_gcse_p and
1440 load_killed_in_block_p. A memory reference for a load instruction,
1441 mems_conflict_for_gcse_p will see if a memory store conflicts with
1442 this memory load. */
1443 static rtx gcse_mem_operand
;
1445 /* DEST is the output of an instruction. If it is a memory reference, and
1446 possibly conflicts with the load found in gcse_mem_operand, then set
1447 gcse_mems_conflict_p to a nonzero value. */
1450 mems_conflict_for_gcse_p (dest
, setter
, data
)
1451 rtx dest
, setter ATTRIBUTE_UNUSED
;
1452 void *data ATTRIBUTE_UNUSED
;
1454 while (GET_CODE (dest
) == SUBREG
1455 || GET_CODE (dest
) == ZERO_EXTRACT
1456 || GET_CODE (dest
) == SIGN_EXTRACT
1457 || GET_CODE (dest
) == STRICT_LOW_PART
)
1458 dest
= XEXP (dest
, 0);
1460 /* If DEST is not a MEM, then it will not conflict with the load. Note
1461 that function calls are assumed to clobber memory, but are handled
1463 if (GET_CODE (dest
) != MEM
)
1466 /* If we are setting a MEM in our list of specially recognized MEMs,
1467 don't mark as killed this time. */
1469 if (dest
== gcse_mem_operand
&& pre_ldst_mems
!= NULL
)
1471 if (!find_rtx_in_ldst (dest
))
1472 gcse_mems_conflict_p
= 1;
1476 if (true_dependence (dest
, GET_MODE (dest
), gcse_mem_operand
,
1478 gcse_mems_conflict_p
= 1;
1481 /* Return nonzero if the expression in X (a memory reference) is killed
1482 in block BB before or after the insn with the CUID in UID_LIMIT.
1483 AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills
1486 To check the entire block, set UID_LIMIT to max_uid + 1 and
1490 load_killed_in_block_p (bb
, uid_limit
, x
, avail_p
)
1496 rtx list_entry
= modify_mem_list
[bb
->index
];
1500 /* Ignore entries in the list that do not apply. */
1502 && INSN_CUID (XEXP (list_entry
, 0)) < uid_limit
)
1504 && INSN_CUID (XEXP (list_entry
, 0)) > uid_limit
))
1506 list_entry
= XEXP (list_entry
, 1);
1510 setter
= XEXP (list_entry
, 0);
1512 /* If SETTER is a call everything is clobbered. Note that calls
1513 to pure functions are never put on the list, so we need not
1514 worry about them. */
1515 if (GET_CODE (setter
) == CALL_INSN
)
1518 /* SETTER must be an INSN of some kind that sets memory. Call
1519 note_stores to examine each hunk of memory that is modified.
1521 The note_stores interface is pretty limited, so we have to
1522 communicate via global variables. Yuk. */
1523 gcse_mem_operand
= x
;
1524 gcse_mems_conflict_p
= 0;
1525 note_stores (PATTERN (setter
), mems_conflict_for_gcse_p
, NULL
);
1526 if (gcse_mems_conflict_p
)
1528 list_entry
= XEXP (list_entry
, 1);
1533 /* Return non-zero if the operands of expression X are unchanged from
1534 the start of INSN's basic block up to but not including INSN. */
1537 oprs_anticipatable_p (x
, insn
)
1540 return oprs_unchanged_p (x
, insn
, 0);
1543 /* Return non-zero if the operands of expression X are unchanged from
1544 INSN to the end of INSN's basic block. */
1547 oprs_available_p (x
, insn
)
1550 return oprs_unchanged_p (x
, insn
, 1);
1553 /* Hash expression X.
1555 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1556 indicating if a volatile operand is found or if the expression contains
1557 something we don't want to insert in the table.
1559 ??? One might want to merge this with canon_hash. Later. */
1562 hash_expr (x
, mode
, do_not_record_p
, hash_table_size
)
1564 enum machine_mode mode
;
1565 int *do_not_record_p
;
1566 int hash_table_size
;
1570 *do_not_record_p
= 0;
1572 hash
= hash_expr_1 (x
, mode
, do_not_record_p
);
1573 return hash
% hash_table_size
;
1576 /* Hash a string. Just add its bytes up. */
1578 static inline unsigned
1583 const unsigned char *p
= (const unsigned char *)ps
;
1592 /* Subroutine of hash_expr to do the actual work. */
1595 hash_expr_1 (x
, mode
, do_not_record_p
)
1597 enum machine_mode mode
;
1598 int *do_not_record_p
;
1605 /* Used to turn recursion into iteration. We can't rely on GCC's
1606 tail-recursion eliminatio since we need to keep accumulating values
1613 code
= GET_CODE (x
);
1617 hash
+= ((unsigned int) REG
<< 7) + REGNO (x
);
1621 hash
+= (((unsigned int) CONST_INT
<< 7) + (unsigned int) mode
1622 + (unsigned int) INTVAL (x
));
1626 /* This is like the general case, except that it only counts
1627 the integers representing the constant. */
1628 hash
+= (unsigned int) code
+ (unsigned int) GET_MODE (x
);
1629 if (GET_MODE (x
) != VOIDmode
)
1630 for (i
= 2; i
< GET_RTX_LENGTH (CONST_DOUBLE
); i
++)
1631 hash
+= (unsigned int) XWINT (x
, i
);
1633 hash
+= ((unsigned int) CONST_DOUBLE_LOW (x
)
1634 + (unsigned int) CONST_DOUBLE_HIGH (x
));
1637 /* Assume there is only one rtx object for any given label. */
1639 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
1640 differences and differences between each stage's debugging dumps. */
1641 hash
+= (((unsigned int) LABEL_REF
<< 7)
1642 + CODE_LABEL_NUMBER (XEXP (x
, 0)));
1647 /* Don't hash on the symbol's address to avoid bootstrap differences.
1648 Different hash values may cause expressions to be recorded in
1649 different orders and thus different registers to be used in the
1650 final assembler. This also avoids differences in the dump files
1651 between various stages. */
1653 const unsigned char *p
= (const unsigned char *) XSTR (x
, 0);
1656 h
+= (h
<< 7) + *p
++; /* ??? revisit */
1658 hash
+= ((unsigned int) SYMBOL_REF
<< 7) + h
;
1663 if (MEM_VOLATILE_P (x
))
1665 *do_not_record_p
= 1;
1669 hash
+= (unsigned int) MEM
;
1670 hash
+= MEM_ALIAS_SET (x
);
1681 case UNSPEC_VOLATILE
:
1682 *do_not_record_p
= 1;
1686 if (MEM_VOLATILE_P (x
))
1688 *do_not_record_p
= 1;
1693 /* We don't want to take the filename and line into account. */
1694 hash
+= (unsigned) code
+ (unsigned) GET_MODE (x
)
1695 + hash_string_1 (ASM_OPERANDS_TEMPLATE (x
))
1696 + hash_string_1 (ASM_OPERANDS_OUTPUT_CONSTRAINT (x
))
1697 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x
);
1699 if (ASM_OPERANDS_INPUT_LENGTH (x
))
1701 for (i
= 1; i
< ASM_OPERANDS_INPUT_LENGTH (x
); i
++)
1703 hash
+= (hash_expr_1 (ASM_OPERANDS_INPUT (x
, i
),
1704 GET_MODE (ASM_OPERANDS_INPUT (x
, i
)),
1706 + hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT
1710 hash
+= hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT (x
, 0));
1711 x
= ASM_OPERANDS_INPUT (x
, 0);
1712 mode
= GET_MODE (x
);
1722 hash
+= (unsigned) code
+ (unsigned) GET_MODE (x
);
1723 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
1727 /* If we are about to do the last recursive call
1728 needed at this level, change it into iteration.
1729 This function is called enough to be worth it. */
1736 hash
+= hash_expr_1 (XEXP (x
, i
), 0, do_not_record_p
);
1737 if (*do_not_record_p
)
1741 else if (fmt
[i
] == 'E')
1742 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1744 hash
+= hash_expr_1 (XVECEXP (x
, i
, j
), 0, do_not_record_p
);
1745 if (*do_not_record_p
)
1749 else if (fmt
[i
] == 's')
1750 hash
+= hash_string_1 (XSTR (x
, i
));
1751 else if (fmt
[i
] == 'i')
1752 hash
+= (unsigned int) XINT (x
, i
);
1760 /* Hash a set of register REGNO.
1762 Sets are hashed on the register that is set. This simplifies the PRE copy
1765 ??? May need to make things more elaborate. Later, as necessary. */
1768 hash_set (regno
, hash_table_size
)
1770 int hash_table_size
;
1775 return hash
% hash_table_size
;
1778 /* Return non-zero if exp1 is equivalent to exp2.
1779 ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
1792 if (x
== 0 || y
== 0)
1795 code
= GET_CODE (x
);
1796 if (code
!= GET_CODE (y
))
1799 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
1800 if (GET_MODE (x
) != GET_MODE (y
))
1810 return INTVAL (x
) == INTVAL (y
);
1813 return XEXP (x
, 0) == XEXP (y
, 0);
1816 return XSTR (x
, 0) == XSTR (y
, 0);
1819 return REGNO (x
) == REGNO (y
);
1822 /* Can't merge two expressions in different alias sets, since we can
1823 decide that the expression is transparent in a block when it isn't,
1824 due to it being set with the different alias set. */
1825 if (MEM_ALIAS_SET (x
) != MEM_ALIAS_SET (y
))
1829 /* For commutative operations, check both orders. */
1837 return ((expr_equiv_p (XEXP (x
, 0), XEXP (y
, 0))
1838 && expr_equiv_p (XEXP (x
, 1), XEXP (y
, 1)))
1839 || (expr_equiv_p (XEXP (x
, 0), XEXP (y
, 1))
1840 && expr_equiv_p (XEXP (x
, 1), XEXP (y
, 0))));
1843 /* We don't use the generic code below because we want to
1844 disregard filename and line numbers. */
1846 /* A volatile asm isn't equivalent to any other. */
1847 if (MEM_VOLATILE_P (x
) || MEM_VOLATILE_P (y
))
1850 if (GET_MODE (x
) != GET_MODE (y
)
1851 || strcmp (ASM_OPERANDS_TEMPLATE (x
), ASM_OPERANDS_TEMPLATE (y
))
1852 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x
),
1853 ASM_OPERANDS_OUTPUT_CONSTRAINT (y
))
1854 || ASM_OPERANDS_OUTPUT_IDX (x
) != ASM_OPERANDS_OUTPUT_IDX (y
)
1855 || ASM_OPERANDS_INPUT_LENGTH (x
) != ASM_OPERANDS_INPUT_LENGTH (y
))
1858 if (ASM_OPERANDS_INPUT_LENGTH (x
))
1860 for (i
= ASM_OPERANDS_INPUT_LENGTH (x
) - 1; i
>= 0; i
--)
1861 if (! expr_equiv_p (ASM_OPERANDS_INPUT (x
, i
),
1862 ASM_OPERANDS_INPUT (y
, i
))
1863 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x
, i
),
1864 ASM_OPERANDS_INPUT_CONSTRAINT (y
, i
)))
1874 /* Compare the elements. If any pair of corresponding elements
1875 fail to match, return 0 for the whole thing. */
1877 fmt
= GET_RTX_FORMAT (code
);
1878 for (i
= GET_RTX_LENGTH (code
) - 1; i
>= 0; i
--)
1883 if (! expr_equiv_p (XEXP (x
, i
), XEXP (y
, i
)))
1888 if (XVECLEN (x
, i
) != XVECLEN (y
, i
))
1890 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1891 if (! expr_equiv_p (XVECEXP (x
, i
, j
), XVECEXP (y
, i
, j
)))
1896 if (strcmp (XSTR (x
, i
), XSTR (y
, i
)))
1901 if (XINT (x
, i
) != XINT (y
, i
))
1906 if (XWINT (x
, i
) != XWINT (y
, i
))
1921 /* Insert expression X in INSN in the hash table.
1922 If it is already present, record it as the last occurrence in INSN's
1925 MODE is the mode of the value X is being stored into.
1926 It is only used if X is a CONST_INT.
1928 ANTIC_P is non-zero if X is an anticipatable expression.
1929 AVAIL_P is non-zero if X is an available expression. */
1932 insert_expr_in_table (x
, mode
, insn
, antic_p
, avail_p
)
1934 enum machine_mode mode
;
1936 int antic_p
, avail_p
;
1938 int found
, do_not_record_p
;
1940 struct expr
*cur_expr
, *last_expr
= NULL
;
1941 struct occr
*antic_occr
, *avail_occr
;
1942 struct occr
*last_occr
= NULL
;
1944 hash
= hash_expr (x
, mode
, &do_not_record_p
, expr_hash_table_size
);
1946 /* Do not insert expression in table if it contains volatile operands,
1947 or if hash_expr determines the expression is something we don't want
1948 to or can't handle. */
1949 if (do_not_record_p
)
1952 cur_expr
= expr_hash_table
[hash
];
1955 while (cur_expr
&& 0 == (found
= expr_equiv_p (cur_expr
->expr
, x
)))
1957 /* If the expression isn't found, save a pointer to the end of
1959 last_expr
= cur_expr
;
1960 cur_expr
= cur_expr
->next_same_hash
;
1965 cur_expr
= (struct expr
*) gcse_alloc (sizeof (struct expr
));
1966 bytes_used
+= sizeof (struct expr
);
1967 if (expr_hash_table
[hash
] == NULL
)
1968 /* This is the first pattern that hashed to this index. */
1969 expr_hash_table
[hash
] = cur_expr
;
1971 /* Add EXPR to end of this hash chain. */
1972 last_expr
->next_same_hash
= cur_expr
;
1974 /* Set the fields of the expr element. */
1976 cur_expr
->bitmap_index
= n_exprs
++;
1977 cur_expr
->next_same_hash
= NULL
;
1978 cur_expr
->antic_occr
= NULL
;
1979 cur_expr
->avail_occr
= NULL
;
1982 /* Now record the occurrence(s). */
1985 antic_occr
= cur_expr
->antic_occr
;
1987 /* Search for another occurrence in the same basic block. */
1988 while (antic_occr
&& BLOCK_NUM (antic_occr
->insn
) != BLOCK_NUM (insn
))
1990 /* If an occurrence isn't found, save a pointer to the end of
1992 last_occr
= antic_occr
;
1993 antic_occr
= antic_occr
->next
;
1997 /* Found another instance of the expression in the same basic block.
1998 Prefer the currently recorded one. We want the first one in the
1999 block and the block is scanned from start to end. */
2000 ; /* nothing to do */
2003 /* First occurrence of this expression in this basic block. */
2004 antic_occr
= (struct occr
*) gcse_alloc (sizeof (struct occr
));
2005 bytes_used
+= sizeof (struct occr
);
2006 /* First occurrence of this expression in any block? */
2007 if (cur_expr
->antic_occr
== NULL
)
2008 cur_expr
->antic_occr
= antic_occr
;
2010 last_occr
->next
= antic_occr
;
2012 antic_occr
->insn
= insn
;
2013 antic_occr
->next
= NULL
;
2019 avail_occr
= cur_expr
->avail_occr
;
2021 /* Search for another occurrence in the same basic block. */
2022 while (avail_occr
&& BLOCK_NUM (avail_occr
->insn
) != BLOCK_NUM (insn
))
2024 /* If an occurrence isn't found, save a pointer to the end of
2026 last_occr
= avail_occr
;
2027 avail_occr
= avail_occr
->next
;
2031 /* Found another instance of the expression in the same basic block.
2032 Prefer this occurrence to the currently recorded one. We want
2033 the last one in the block and the block is scanned from start
2035 avail_occr
->insn
= insn
;
2038 /* First occurrence of this expression in this basic block. */
2039 avail_occr
= (struct occr
*) gcse_alloc (sizeof (struct occr
));
2040 bytes_used
+= sizeof (struct occr
);
2042 /* First occurrence of this expression in any block? */
2043 if (cur_expr
->avail_occr
== NULL
)
2044 cur_expr
->avail_occr
= avail_occr
;
2046 last_occr
->next
= avail_occr
;
2048 avail_occr
->insn
= insn
;
2049 avail_occr
->next
= NULL
;
2054 /* Insert pattern X in INSN in the hash table.
2055 X is a SET of a reg to either another reg or a constant.
2056 If it is already present, record it as the last occurrence in INSN's
2060 insert_set_in_table (x
, insn
)
2066 struct expr
*cur_expr
, *last_expr
= NULL
;
2067 struct occr
*cur_occr
, *last_occr
= NULL
;
2069 if (GET_CODE (x
) != SET
2070 || GET_CODE (SET_DEST (x
)) != REG
)
2073 hash
= hash_set (REGNO (SET_DEST (x
)), set_hash_table_size
);
2075 cur_expr
= set_hash_table
[hash
];
2078 while (cur_expr
&& 0 == (found
= expr_equiv_p (cur_expr
->expr
, x
)))
2080 /* If the expression isn't found, save a pointer to the end of
2082 last_expr
= cur_expr
;
2083 cur_expr
= cur_expr
->next_same_hash
;
2088 cur_expr
= (struct expr
*) gcse_alloc (sizeof (struct expr
));
2089 bytes_used
+= sizeof (struct expr
);
2090 if (set_hash_table
[hash
] == NULL
)
2091 /* This is the first pattern that hashed to this index. */
2092 set_hash_table
[hash
] = cur_expr
;
2094 /* Add EXPR to end of this hash chain. */
2095 last_expr
->next_same_hash
= cur_expr
;
2097 /* Set the fields of the expr element.
2098 We must copy X because it can be modified when copy propagation is
2099 performed on its operands. */
2100 cur_expr
->expr
= copy_rtx (x
);
2101 cur_expr
->bitmap_index
= n_sets
++;
2102 cur_expr
->next_same_hash
= NULL
;
2103 cur_expr
->antic_occr
= NULL
;
2104 cur_expr
->avail_occr
= NULL
;
2107 /* Now record the occurrence. */
2108 cur_occr
= cur_expr
->avail_occr
;
2110 /* Search for another occurrence in the same basic block. */
2111 while (cur_occr
&& BLOCK_NUM (cur_occr
->insn
) != BLOCK_NUM (insn
))
2113 /* If an occurrence isn't found, save a pointer to the end of
2115 last_occr
= cur_occr
;
2116 cur_occr
= cur_occr
->next
;
2120 /* Found another instance of the expression in the same basic block.
2121 Prefer this occurrence to the currently recorded one. We want the
2122 last one in the block and the block is scanned from start to end. */
2123 cur_occr
->insn
= insn
;
2126 /* First occurrence of this expression in this basic block. */
2127 cur_occr
= (struct occr
*) gcse_alloc (sizeof (struct occr
));
2128 bytes_used
+= sizeof (struct occr
);
2130 /* First occurrence of this expression in any block? */
2131 if (cur_expr
->avail_occr
== NULL
)
2132 cur_expr
->avail_occr
= cur_occr
;
2134 last_occr
->next
= cur_occr
;
2136 cur_occr
->insn
= insn
;
2137 cur_occr
->next
= NULL
;
2141 /* Scan pattern PAT of INSN and add an entry to the hash table. If SET_P is
2142 non-zero, this is for the assignment hash table, otherwise it is for the
2143 expression hash table. */
2146 hash_scan_set (pat
, insn
, set_p
)
2150 rtx src
= SET_SRC (pat
);
2151 rtx dest
= SET_DEST (pat
);
2154 if (GET_CODE (src
) == CALL
)
2155 hash_scan_call (src
, insn
);
2157 else if (GET_CODE (dest
) == REG
)
2159 unsigned int regno
= REGNO (dest
);
2162 /* If this is a single set and we are doing constant propagation,
2163 see if a REG_NOTE shows this equivalent to a constant. */
2164 if (set_p
&& (note
= find_reg_equal_equiv_note (insn
)) != 0
2165 && CONSTANT_P (XEXP (note
, 0)))
2166 src
= XEXP (note
, 0), pat
= gen_rtx_SET (VOIDmode
, dest
, src
);
2168 /* Only record sets of pseudo-regs in the hash table. */
2170 && regno
>= FIRST_PSEUDO_REGISTER
2171 /* Don't GCSE something if we can't do a reg/reg copy. */
2172 && can_copy_p
[GET_MODE (dest
)]
2173 /* Is SET_SRC something we want to gcse? */
2174 && want_to_gcse_p (src
)
2175 /* Don't CSE a nop. */
2176 && ! set_noop_p (pat
)
2177 /* Don't GCSE if it has attached REG_EQUIV note.
2178 At this point this only function parameters should have
2179 REG_EQUIV notes and if the argument slot is used somewhere
2180 explicitely, it means address of parameter has been taken,
2181 so we should not extend the lifetime of the pseudo. */
2182 && ((note
= find_reg_note (insn
, REG_EQUIV
, NULL_RTX
)) == 0
2183 || GET_CODE (XEXP (note
, 0)) != MEM
))
2185 /* An expression is not anticipatable if its operands are
2186 modified before this insn or if this is not the only SET in
2188 int antic_p
= oprs_anticipatable_p (src
, insn
) && single_set (insn
);
2189 /* An expression is not available if its operands are
2190 subsequently modified, including this insn. It's also not
2191 available if this is a branch, because we can't insert
2192 a set after the branch. */
2193 int avail_p
= (oprs_available_p (src
, insn
)
2194 && ! JUMP_P (insn
));
2196 insert_expr_in_table (src
, GET_MODE (dest
), insn
, antic_p
, avail_p
);
2199 /* Record sets for constant/copy propagation. */
2201 && regno
>= FIRST_PSEUDO_REGISTER
2202 && ((GET_CODE (src
) == REG
2203 && REGNO (src
) >= FIRST_PSEUDO_REGISTER
2204 && can_copy_p
[GET_MODE (dest
)]
2205 && REGNO (src
) != regno
)
2206 || GET_CODE (src
) == CONST_INT
2207 || GET_CODE (src
) == SYMBOL_REF
2208 || GET_CODE (src
) == CONST_DOUBLE
)
2209 /* A copy is not available if its src or dest is subsequently
2210 modified. Here we want to search from INSN+1 on, but
2211 oprs_available_p searches from INSN on. */
2212 && (insn
== BLOCK_END (BLOCK_NUM (insn
))
2213 || ((tmp
= next_nonnote_insn (insn
)) != NULL_RTX
2214 && oprs_available_p (pat
, tmp
))))
2215 insert_set_in_table (pat
, insn
);
2220 hash_scan_clobber (x
, insn
)
2221 rtx x ATTRIBUTE_UNUSED
, insn ATTRIBUTE_UNUSED
;
2223 /* Currently nothing to do. */
2227 hash_scan_call (x
, insn
)
2228 rtx x ATTRIBUTE_UNUSED
, insn ATTRIBUTE_UNUSED
;
2230 /* Currently nothing to do. */
2233 /* Process INSN and add hash table entries as appropriate.
2235 Only available expressions that set a single pseudo-reg are recorded.
2237 Single sets in a PARALLEL could be handled, but it's an extra complication
2238 that isn't dealt with right now. The trick is handling the CLOBBERs that
2239 are also in the PARALLEL. Later.
2241 If SET_P is non-zero, this is for the assignment hash table,
2242 otherwise it is for the expression hash table.
2243 If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
2244 not record any expressions. */
2247 hash_scan_insn (insn
, set_p
, in_libcall_block
)
2250 int in_libcall_block
;
2252 rtx pat
= PATTERN (insn
);
2255 if (in_libcall_block
)
2258 /* Pick out the sets of INSN and for other forms of instructions record
2259 what's been modified. */
2261 if (GET_CODE (pat
) == SET
)
2262 hash_scan_set (pat
, insn
, set_p
);
2263 else if (GET_CODE (pat
) == PARALLEL
)
2264 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
2266 rtx x
= XVECEXP (pat
, 0, i
);
2268 if (GET_CODE (x
) == SET
)
2269 hash_scan_set (x
, insn
, set_p
);
2270 else if (GET_CODE (x
) == CLOBBER
)
2271 hash_scan_clobber (x
, insn
);
2272 else if (GET_CODE (x
) == CALL
)
2273 hash_scan_call (x
, insn
);
2276 else if (GET_CODE (pat
) == CLOBBER
)
2277 hash_scan_clobber (pat
, insn
);
2278 else if (GET_CODE (pat
) == CALL
)
2279 hash_scan_call (pat
, insn
);
2283 dump_hash_table (file
, name
, table
, table_size
, total_size
)
2286 struct expr
**table
;
2287 int table_size
, total_size
;
2290 /* Flattened out table, so it's printed in proper order. */
2291 struct expr
**flat_table
;
2292 unsigned int *hash_val
;
2296 = (struct expr
**) xcalloc (total_size
, sizeof (struct expr
*));
2297 hash_val
= (unsigned int *) xmalloc (total_size
* sizeof (unsigned int));
2299 for (i
= 0; i
< table_size
; i
++)
2300 for (expr
= table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
2302 flat_table
[expr
->bitmap_index
] = expr
;
2303 hash_val
[expr
->bitmap_index
] = i
;
2306 fprintf (file
, "%s hash table (%d buckets, %d entries)\n",
2307 name
, table_size
, total_size
);
2309 for (i
= 0; i
< total_size
; i
++)
2310 if (flat_table
[i
] != 0)
2312 expr
= flat_table
[i
];
2313 fprintf (file
, "Index %d (hash value %d)\n ",
2314 expr
->bitmap_index
, hash_val
[i
]);
2315 print_rtl (file
, expr
->expr
);
2316 fprintf (file
, "\n");
2319 fprintf (file
, "\n");
2325 /* Record register first/last/block set information for REGNO in INSN.
2327 first_set records the first place in the block where the register
2328 is set and is used to compute "anticipatability".
2330 last_set records the last place in the block where the register
2331 is set and is used to compute "availability".
2333 last_bb records the block for which first_set and last_set are
2334 valid, as a quick test to invalidate them.
2336 reg_set_in_block records whether the register is set in the block
2337 and is used to compute "transparency". */
2340 record_last_reg_set_info (insn
, regno
)
2344 struct reg_avail_info
*info
= ®_avail_info
[regno
];
2345 int cuid
= INSN_CUID (insn
);
2347 info
->last_set
= cuid
;
2348 if (info
->last_bb
!= current_bb
)
2350 info
->last_bb
= current_bb
;
2351 info
->first_set
= cuid
;
2352 SET_BIT (reg_set_in_block
[current_bb
], regno
);
2357 /* Record all of the canonicalized MEMs of record_last_mem_set_info's insn.
2358 Note we store a pair of elements in the list, so they have to be
2359 taken off pairwise. */
2362 canon_list_insert (dest
, unused1
, v_insn
)
2363 rtx dest ATTRIBUTE_UNUSED
;
2364 rtx unused1 ATTRIBUTE_UNUSED
;
2367 rtx dest_addr
, insn
;
2369 while (GET_CODE (dest
) == SUBREG
2370 || GET_CODE (dest
) == ZERO_EXTRACT
2371 || GET_CODE (dest
) == SIGN_EXTRACT
2372 || GET_CODE (dest
) == STRICT_LOW_PART
)
2373 dest
= XEXP (dest
, 0);
2375 /* If DEST is not a MEM, then it will not conflict with a load. Note
2376 that function calls are assumed to clobber memory, but are handled
2379 if (GET_CODE (dest
) != MEM
)
2382 dest_addr
= get_addr (XEXP (dest
, 0));
2383 dest_addr
= canon_rtx (dest_addr
);
2384 insn
= (rtx
) v_insn
;
2386 canon_modify_mem_list
[BLOCK_NUM (insn
)] =
2387 alloc_INSN_LIST (dest_addr
, canon_modify_mem_list
[BLOCK_NUM (insn
)]);
2388 canon_modify_mem_list
[BLOCK_NUM (insn
)] =
2389 alloc_INSN_LIST (dest
, canon_modify_mem_list
[BLOCK_NUM (insn
)]);
2390 bitmap_set_bit (canon_modify_mem_list_set
, BLOCK_NUM (insn
));
2393 /* Record memory modification information for INSN. We do not actually care
2394 about the memory location(s) that are set, or even how they are set (consider
2395 a CALL_INSN). We merely need to record which insns modify memory. */
2398 record_last_mem_set_info (insn
)
2401 /* load_killed_in_block_p will handle the case of calls clobbering
2403 modify_mem_list
[BLOCK_NUM (insn
)] =
2404 alloc_INSN_LIST (insn
, modify_mem_list
[BLOCK_NUM (insn
)]);
2405 bitmap_set_bit (modify_mem_list_set
, BLOCK_NUM (insn
));
2407 if (GET_CODE (insn
) == CALL_INSN
)
2409 /* Note that traversals of this loop (other than for free-ing)
2410 will break after encountering a CALL_INSN. So, there's no
2411 need to insert a pair of items, as canon_list_insert does. */
2412 canon_modify_mem_list
[BLOCK_NUM (insn
)] =
2413 alloc_INSN_LIST (insn
, canon_modify_mem_list
[BLOCK_NUM (insn
)]);
2414 bitmap_set_bit (canon_modify_mem_list_set
, BLOCK_NUM (insn
));
2417 note_stores (PATTERN (insn
), canon_list_insert
, (void*)insn
);
2420 /* Called from compute_hash_table via note_stores to handle one
2421 SET or CLOBBER in an insn. DATA is really the instruction in which
2422 the SET is taking place. */
2425 record_last_set_info (dest
, setter
, data
)
2426 rtx dest
, setter ATTRIBUTE_UNUSED
;
2429 rtx last_set_insn
= (rtx
) data
;
2431 if (GET_CODE (dest
) == SUBREG
)
2432 dest
= SUBREG_REG (dest
);
2434 if (GET_CODE (dest
) == REG
)
2435 record_last_reg_set_info (last_set_insn
, REGNO (dest
));
2436 else if (GET_CODE (dest
) == MEM
2437 /* Ignore pushes, they clobber nothing. */
2438 && ! push_operand (dest
, GET_MODE (dest
)))
2439 record_last_mem_set_info (last_set_insn
);
2442 /* Top level function to create an expression or assignment hash table.
2444 Expression entries are placed in the hash table if
2445 - they are of the form (set (pseudo-reg) src),
2446 - src is something we want to perform GCSE on,
2447 - none of the operands are subsequently modified in the block
2449 Assignment entries are placed in the hash table if
2450 - they are of the form (set (pseudo-reg) src),
2451 - src is something we want to perform const/copy propagation on,
2452 - none of the operands or target are subsequently modified in the block
2454 Currently src must be a pseudo-reg or a const_int.
2456 F is the first insn.
2457 SET_P is non-zero for computing the assignment hash table. */
2460 compute_hash_table (set_p
)
2465 /* While we compute the hash table we also compute a bit array of which
2466 registers are set in which blocks.
2467 ??? This isn't needed during const/copy propagation, but it's cheap to
2469 sbitmap_vector_zero (reg_set_in_block
, n_basic_blocks
);
2471 /* re-Cache any INSN_LIST nodes we have allocated. */
2472 clear_modify_mem_tables ();
2473 /* Some working arrays used to track first and last set in each block. */
2474 reg_avail_info
= (struct reg_avail_info
*)
2475 gmalloc (max_gcse_regno
* sizeof (struct reg_avail_info
));
2477 for (i
= 0; i
< max_gcse_regno
; ++i
)
2478 reg_avail_info
[i
].last_bb
= NEVER_SET
;
2480 for (current_bb
= 0; current_bb
< n_basic_blocks
; current_bb
++)
2484 int in_libcall_block
;
2486 /* First pass over the instructions records information used to
2487 determine when registers and memory are first and last set.
2488 ??? hard-reg reg_set_in_block computation
2489 could be moved to compute_sets since they currently don't change. */
2491 for (insn
= BLOCK_HEAD (current_bb
);
2492 insn
&& insn
!= NEXT_INSN (BLOCK_END (current_bb
));
2493 insn
= NEXT_INSN (insn
))
2495 if (! INSN_P (insn
))
2498 if (GET_CODE (insn
) == CALL_INSN
)
2500 bool clobbers_all
= false;
2501 #ifdef NON_SAVING_SETJMP
2502 if (NON_SAVING_SETJMP
2503 && find_reg_note (insn
, REG_SETJMP
, NULL_RTX
))
2504 clobbers_all
= true;
2507 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
2509 || TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
2510 record_last_reg_set_info (insn
, regno
);
2515 note_stores (PATTERN (insn
), record_last_set_info
, insn
);
2518 /* The next pass builds the hash table. */
2520 for (insn
= BLOCK_HEAD (current_bb
), in_libcall_block
= 0;
2521 insn
&& insn
!= NEXT_INSN (BLOCK_END (current_bb
));
2522 insn
= NEXT_INSN (insn
))
2525 if (find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
))
2526 in_libcall_block
= 1;
2527 else if (set_p
&& find_reg_note (insn
, REG_RETVAL
, NULL_RTX
))
2528 in_libcall_block
= 0;
2529 hash_scan_insn (insn
, set_p
, in_libcall_block
);
2530 if (!set_p
&& find_reg_note (insn
, REG_RETVAL
, NULL_RTX
))
2531 in_libcall_block
= 0;
2535 free (reg_avail_info
);
2536 reg_avail_info
= NULL
;
2539 /* Allocate space for the set hash table.
2540 N_INSNS is the number of instructions in the function.
2541 It is used to determine the number of buckets to use. */
2544 alloc_set_hash_table (n_insns
)
2549 set_hash_table_size
= n_insns
/ 4;
2550 if (set_hash_table_size
< 11)
2551 set_hash_table_size
= 11;
2553 /* Attempt to maintain efficient use of hash table.
2554 Making it an odd number is simplest for now.
2555 ??? Later take some measurements. */
2556 set_hash_table_size
|= 1;
2557 n
= set_hash_table_size
* sizeof (struct expr
*);
2558 set_hash_table
= (struct expr
**) gmalloc (n
);
2561 /* Free things allocated by alloc_set_hash_table. */
2564 free_set_hash_table ()
2566 free (set_hash_table
);
2569 /* Compute the hash table for doing copy/const propagation. */
2572 compute_set_hash_table ()
2574 /* Initialize count of number of entries in hash table. */
2576 memset ((char *) set_hash_table
, 0,
2577 set_hash_table_size
* sizeof (struct expr
*));
2579 compute_hash_table (1);
2582 /* Allocate space for the expression hash table.
2583 N_INSNS is the number of instructions in the function.
2584 It is used to determine the number of buckets to use. */
2587 alloc_expr_hash_table (n_insns
)
2588 unsigned int n_insns
;
2592 expr_hash_table_size
= n_insns
/ 2;
2593 /* Make sure the amount is usable. */
2594 if (expr_hash_table_size
< 11)
2595 expr_hash_table_size
= 11;
2597 /* Attempt to maintain efficient use of hash table.
2598 Making it an odd number is simplest for now.
2599 ??? Later take some measurements. */
2600 expr_hash_table_size
|= 1;
2601 n
= expr_hash_table_size
* sizeof (struct expr
*);
2602 expr_hash_table
= (struct expr
**) gmalloc (n
);
2605 /* Free things allocated by alloc_expr_hash_table. */
2608 free_expr_hash_table ()
2610 free (expr_hash_table
);
2613 /* Compute the hash table for doing GCSE. */
2616 compute_expr_hash_table ()
2618 /* Initialize count of number of entries in hash table. */
2620 memset ((char *) expr_hash_table
, 0,
2621 expr_hash_table_size
* sizeof (struct expr
*));
2623 compute_hash_table (0);
2626 /* Expression tracking support. */
2628 /* Lookup pattern PAT in the expression table.
2629 The result is a pointer to the table entry, or NULL if not found. */
2631 static struct expr
*
2635 int do_not_record_p
;
2636 unsigned int hash
= hash_expr (pat
, GET_MODE (pat
), &do_not_record_p
,
2637 expr_hash_table_size
);
2640 if (do_not_record_p
)
2643 expr
= expr_hash_table
[hash
];
2645 while (expr
&& ! expr_equiv_p (expr
->expr
, pat
))
2646 expr
= expr
->next_same_hash
;
2651 /* Lookup REGNO in the set table. If PAT is non-NULL look for the entry that
2652 matches it, otherwise return the first entry for REGNO. The result is a
2653 pointer to the table entry, or NULL if not found. */
2655 static struct expr
*
2656 lookup_set (regno
, pat
)
2660 unsigned int hash
= hash_set (regno
, set_hash_table_size
);
2663 expr
= set_hash_table
[hash
];
2667 while (expr
&& ! expr_equiv_p (expr
->expr
, pat
))
2668 expr
= expr
->next_same_hash
;
2672 while (expr
&& REGNO (SET_DEST (expr
->expr
)) != regno
)
2673 expr
= expr
->next_same_hash
;
2679 /* Return the next entry for REGNO in list EXPR. */
2681 static struct expr
*
2682 next_set (regno
, expr
)
2687 expr
= expr
->next_same_hash
;
2688 while (expr
&& REGNO (SET_DEST (expr
->expr
)) != regno
);
2693 /* Clear canon_modify_mem_list and modify_mem_list tables. */
2695 clear_modify_mem_tables ()
2699 EXECUTE_IF_SET_IN_BITMAP
2700 (canon_modify_mem_list_set
, 0, i
,
2701 free_INSN_LIST_list (modify_mem_list
+ i
));
2702 bitmap_clear (canon_modify_mem_list_set
);
2704 EXECUTE_IF_SET_IN_BITMAP
2705 (canon_modify_mem_list_set
, 0, i
,
2706 free_INSN_LIST_list (canon_modify_mem_list
+ i
));
2707 bitmap_clear (modify_mem_list_set
);
2710 /* Release memory used by modify_mem_list_set and canon_modify_mem_list_set. */
2713 free_modify_mem_tables ()
2715 clear_modify_mem_tables ();
2716 free (modify_mem_list
);
2717 free (canon_modify_mem_list
);
2718 modify_mem_list
= 0;
2719 canon_modify_mem_list
= 0;
2722 /* Reset tables used to keep track of what's still available [since the
2723 start of the block]. */
2726 reset_opr_set_tables ()
2728 /* Maintain a bitmap of which regs have been set since beginning of
2730 CLEAR_REG_SET (reg_set_bitmap
);
2732 /* Also keep a record of the last instruction to modify memory.
2733 For now this is very trivial, we only record whether any memory
2734 location has been modified. */
2735 clear_modify_mem_tables ();
2738 /* Return non-zero if the operands of X are not set before INSN in
2739 INSN's basic block. */
2742 oprs_not_set_p (x
, insn
)
2752 code
= GET_CODE (x
);
2767 if (load_killed_in_block_p (BLOCK_FOR_INSN (insn
),
2768 INSN_CUID (insn
), x
, 0))
2771 return oprs_not_set_p (XEXP (x
, 0), insn
);
2774 return ! REGNO_REG_SET_P (reg_set_bitmap
, REGNO (x
));
2780 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
2784 /* If we are about to do the last recursive call
2785 needed at this level, change it into iteration.
2786 This function is called enough to be worth it. */
2788 return oprs_not_set_p (XEXP (x
, i
), insn
);
2790 if (! oprs_not_set_p (XEXP (x
, i
), insn
))
2793 else if (fmt
[i
] == 'E')
2794 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
2795 if (! oprs_not_set_p (XVECEXP (x
, i
, j
), insn
))
2802 /* Mark things set by a CALL. */
2808 if (! CONST_OR_PURE_CALL_P (insn
))
2809 record_last_mem_set_info (insn
);
2812 /* Mark things set by a SET. */
2815 mark_set (pat
, insn
)
2818 rtx dest
= SET_DEST (pat
);
2820 while (GET_CODE (dest
) == SUBREG
2821 || GET_CODE (dest
) == ZERO_EXTRACT
2822 || GET_CODE (dest
) == SIGN_EXTRACT
2823 || GET_CODE (dest
) == STRICT_LOW_PART
)
2824 dest
= XEXP (dest
, 0);
2826 if (GET_CODE (dest
) == REG
)
2827 SET_REGNO_REG_SET (reg_set_bitmap
, REGNO (dest
));
2828 else if (GET_CODE (dest
) == MEM
)
2829 record_last_mem_set_info (insn
);
2831 if (GET_CODE (SET_SRC (pat
)) == CALL
)
2835 /* Record things set by a CLOBBER. */
2838 mark_clobber (pat
, insn
)
2841 rtx clob
= XEXP (pat
, 0);
2843 while (GET_CODE (clob
) == SUBREG
|| GET_CODE (clob
) == STRICT_LOW_PART
)
2844 clob
= XEXP (clob
, 0);
2846 if (GET_CODE (clob
) == REG
)
2847 SET_REGNO_REG_SET (reg_set_bitmap
, REGNO (clob
));
2849 record_last_mem_set_info (insn
);
2852 /* Record things set by INSN.
2853 This data is used by oprs_not_set_p. */
2856 mark_oprs_set (insn
)
2859 rtx pat
= PATTERN (insn
);
2862 if (GET_CODE (pat
) == SET
)
2863 mark_set (pat
, insn
);
2864 else if (GET_CODE (pat
) == PARALLEL
)
2865 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
2867 rtx x
= XVECEXP (pat
, 0, i
);
2869 if (GET_CODE (x
) == SET
)
2871 else if (GET_CODE (x
) == CLOBBER
)
2872 mark_clobber (x
, insn
);
2873 else if (GET_CODE (x
) == CALL
)
2877 else if (GET_CODE (pat
) == CLOBBER
)
2878 mark_clobber (pat
, insn
);
2879 else if (GET_CODE (pat
) == CALL
)
2884 /* Classic GCSE reaching definition support. */
2886 /* Allocate reaching def variables. */
2889 alloc_rd_mem (n_blocks
, n_insns
)
2890 int n_blocks
, n_insns
;
2892 rd_kill
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_insns
);
2893 sbitmap_vector_zero (rd_kill
, n_basic_blocks
);
2895 rd_gen
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_insns
);
2896 sbitmap_vector_zero (rd_gen
, n_basic_blocks
);
2898 reaching_defs
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_insns
);
2899 sbitmap_vector_zero (reaching_defs
, n_basic_blocks
);
2901 rd_out
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_insns
);
2902 sbitmap_vector_zero (rd_out
, n_basic_blocks
);
2905 /* Free reaching def variables. */
2910 sbitmap_vector_free (rd_kill
);
2911 sbitmap_vector_free (rd_gen
);
2912 sbitmap_vector_free (reaching_defs
);
2913 sbitmap_vector_free (rd_out
);
2916 /* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
2919 handle_rd_kill_set (insn
, regno
, bb
)
2924 struct reg_set
*this_reg
;
2926 for (this_reg
= reg_set_table
[regno
]; this_reg
; this_reg
= this_reg
->next
)
2927 if (BLOCK_NUM (this_reg
->insn
) != BLOCK_NUM (insn
))
2928 SET_BIT (rd_kill
[bb
->index
], INSN_CUID (this_reg
->insn
));
2931 /* Compute the set of kill's for reaching definitions. */
2941 For each set bit in `gen' of the block (i.e each insn which
2942 generates a definition in the block)
2943 Call the reg set by the insn corresponding to that bit regx
2944 Look at the linked list starting at reg_set_table[regx]
2945 For each setting of regx in the linked list, which is not in
2947 Set the bit in `kill' corresponding to that insn. */
2948 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
2949 for (cuid
= 0; cuid
< max_cuid
; cuid
++)
2950 if (TEST_BIT (rd_gen
[bb
], cuid
))
2952 rtx insn
= CUID_INSN (cuid
);
2953 rtx pat
= PATTERN (insn
);
2955 if (GET_CODE (insn
) == CALL_INSN
)
2957 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
2958 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
2959 handle_rd_kill_set (insn
, regno
, BASIC_BLOCK (bb
));
2962 if (GET_CODE (pat
) == PARALLEL
)
2964 for (i
= XVECLEN (pat
, 0) - 1; i
>= 0; i
--)
2966 enum rtx_code code
= GET_CODE (XVECEXP (pat
, 0, i
));
2968 if ((code
== SET
|| code
== CLOBBER
)
2969 && GET_CODE (XEXP (XVECEXP (pat
, 0, i
), 0)) == REG
)
2970 handle_rd_kill_set (insn
,
2971 REGNO (XEXP (XVECEXP (pat
, 0, i
), 0)),
2975 else if (GET_CODE (pat
) == SET
&& GET_CODE (SET_DEST (pat
)) == REG
)
2976 /* Each setting of this register outside of this block
2977 must be marked in the set of kills in this block. */
2978 handle_rd_kill_set (insn
, REGNO (SET_DEST (pat
)), BASIC_BLOCK (bb
));
2982 /* Compute the reaching definitions as in
2983 Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
2984 Chapter 10. It is the same algorithm as used for computing available
2985 expressions but applied to the gens and kills of reaching definitions. */
2990 int bb
, changed
, passes
;
2992 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
2993 sbitmap_copy (rd_out
[bb
] /*dst*/, rd_gen
[bb
] /*src*/);
3000 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3002 sbitmap_union_of_preds (reaching_defs
[bb
], rd_out
, bb
);
3003 changed
|= sbitmap_union_of_diff (rd_out
[bb
], rd_gen
[bb
],
3004 reaching_defs
[bb
], rd_kill
[bb
]);
3010 fprintf (gcse_file
, "reaching def computation: %d passes\n", passes
);
3013 /* Classic GCSE available expression support. */
3015 /* Allocate memory for available expression computation. */
3018 alloc_avail_expr_mem (n_blocks
, n_exprs
)
3019 int n_blocks
, n_exprs
;
3021 ae_kill
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_exprs
);
3022 sbitmap_vector_zero (ae_kill
, n_basic_blocks
);
3024 ae_gen
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_exprs
);
3025 sbitmap_vector_zero (ae_gen
, n_basic_blocks
);
3027 ae_in
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_exprs
);
3028 sbitmap_vector_zero (ae_in
, n_basic_blocks
);
3030 ae_out
= (sbitmap
*) sbitmap_vector_alloc (n_blocks
, n_exprs
);
3031 sbitmap_vector_zero (ae_out
, n_basic_blocks
);
3035 free_avail_expr_mem ()
3037 sbitmap_vector_free (ae_kill
);
3038 sbitmap_vector_free (ae_gen
);
3039 sbitmap_vector_free (ae_in
);
3040 sbitmap_vector_free (ae_out
);
3043 /* Compute the set of available expressions generated in each basic block. */
3052 /* For each recorded occurrence of each expression, set ae_gen[bb][expr].
3053 This is all we have to do because an expression is not recorded if it
3054 is not available, and the only expressions we want to work with are the
3055 ones that are recorded. */
3056 for (i
= 0; i
< expr_hash_table_size
; i
++)
3057 for (expr
= expr_hash_table
[i
]; expr
!= 0; expr
= expr
->next_same_hash
)
3058 for (occr
= expr
->avail_occr
; occr
!= 0; occr
= occr
->next
)
3059 SET_BIT (ae_gen
[BLOCK_NUM (occr
->insn
)], expr
->bitmap_index
);
3062 /* Return non-zero if expression X is killed in BB. */
3065 expr_killed_p (x
, bb
)
3076 code
= GET_CODE (x
);
3080 return TEST_BIT (reg_set_in_block
[bb
->index
], REGNO (x
));
3083 if (load_killed_in_block_p (bb
, get_max_uid () + 1, x
, 0))
3086 return expr_killed_p (XEXP (x
, 0), bb
);
3103 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
3107 /* If we are about to do the last recursive call
3108 needed at this level, change it into iteration.
3109 This function is called enough to be worth it. */
3111 return expr_killed_p (XEXP (x
, i
), bb
);
3112 else if (expr_killed_p (XEXP (x
, i
), bb
))
3115 else if (fmt
[i
] == 'E')
3116 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3117 if (expr_killed_p (XVECEXP (x
, i
, j
), bb
))
3124 /* Compute the set of available expressions killed in each basic block. */
3127 compute_ae_kill (ae_gen
, ae_kill
)
3128 sbitmap
*ae_gen
, *ae_kill
;
3134 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3135 for (i
= 0; i
< expr_hash_table_size
; i
++)
3136 for (expr
= expr_hash_table
[i
]; expr
; expr
= expr
->next_same_hash
)
3138 /* Skip EXPR if generated in this block. */
3139 if (TEST_BIT (ae_gen
[bb
], expr
->bitmap_index
))
3142 if (expr_killed_p (expr
->expr
, BASIC_BLOCK (bb
)))
3143 SET_BIT (ae_kill
[bb
], expr
->bitmap_index
);
3147 /* Actually perform the Classic GCSE optimizations. */
3149 /* Return non-zero if occurrence OCCR of expression EXPR reaches block BB.
3151 CHECK_SELF_LOOP is non-zero if we should consider a block reaching itself
3152 as a positive reach. We want to do this when there are two computations
3153 of the expression in the block.
3155 VISITED is a pointer to a working buffer for tracking which BB's have
3156 been visited. It is NULL for the top-level call.
3158 We treat reaching expressions that go through blocks containing the same
3159 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
3160 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
3161 2 as not reaching. The intent is to improve the probability of finding
3162 only one reaching expression and to reduce register lifetimes by picking
3163 the closest such expression. */
3166 expr_reaches_here_p_work (occr
, expr
, bb
, check_self_loop
, visited
)
3170 int check_self_loop
;
3175 for (pred
= bb
->pred
; pred
!= NULL
; pred
= pred
->pred_next
)
3177 basic_block pred_bb
= pred
->src
;
3179 if (visited
[pred_bb
->index
])
3180 /* This predecessor has already been visited. Nothing to do. */
3182 else if (pred_bb
== bb
)
3184 /* BB loops on itself. */
3186 && TEST_BIT (ae_gen
[pred_bb
->index
], expr
->bitmap_index
)
3187 && BLOCK_NUM (occr
->insn
) == pred_bb
->index
)
3190 visited
[pred_bb
->index
] = 1;
3193 /* Ignore this predecessor if it kills the expression. */
3194 else if (TEST_BIT (ae_kill
[pred_bb
->index
], expr
->bitmap_index
))
3195 visited
[pred_bb
->index
] = 1;
3197 /* Does this predecessor generate this expression? */
3198 else if (TEST_BIT (ae_gen
[pred_bb
->index
], expr
->bitmap_index
))
3200 /* Is this the occurrence we're looking for?
3201 Note that there's only one generating occurrence per block
3202 so we just need to check the block number. */
3203 if (BLOCK_NUM (occr
->insn
) == pred_bb
->index
)
3206 visited
[pred_bb
->index
] = 1;
3209 /* Neither gen nor kill. */
3212 visited
[pred_bb
->index
] = 1;
3213 if (expr_reaches_here_p_work (occr
, expr
, pred_bb
, check_self_loop
,
3220 /* All paths have been checked. */
3224 /* This wrapper for expr_reaches_here_p_work() is to ensure that any
3225 memory allocated for that function is returned. */
3228 expr_reaches_here_p (occr
, expr
, bb
, check_self_loop
)
3232 int check_self_loop
;
3235 char *visited
= (char *) xcalloc (n_basic_blocks
, 1);
3237 rval
= expr_reaches_here_p_work (occr
, expr
, bb
, check_self_loop
, visited
);
3243 /* Return the instruction that computes EXPR that reaches INSN's basic block.
3244 If there is more than one such instruction, return NULL.
3246 Called only by handle_avail_expr. */
3249 computing_insn (expr
, insn
)
3253 basic_block bb
= BLOCK_FOR_INSN (insn
);
3255 if (expr
->avail_occr
->next
== NULL
)
3257 if (BLOCK_FOR_INSN (expr
->avail_occr
->insn
) == bb
)
3258 /* The available expression is actually itself
3259 (i.e. a loop in the flow graph) so do nothing. */
3262 /* (FIXME) Case that we found a pattern that was created by
3263 a substitution that took place. */
3264 return expr
->avail_occr
->insn
;
3268 /* Pattern is computed more than once.
3269 Search backwards from this insn to see how many of these
3270 computations actually reach this insn. */
3272 rtx insn_computes_expr
= NULL
;
3275 for (occr
= expr
->avail_occr
; occr
!= NULL
; occr
= occr
->next
)
3277 if (BLOCK_FOR_INSN (occr
->insn
) == bb
)
3279 /* The expression is generated in this block.
3280 The only time we care about this is when the expression
3281 is generated later in the block [and thus there's a loop].
3282 We let the normal cse pass handle the other cases. */
3283 if (INSN_CUID (insn
) < INSN_CUID (occr
->insn
)
3284 && expr_reaches_here_p (occr
, expr
, bb
, 1))
3290 insn_computes_expr
= occr
->insn
;
3293 else if (expr_reaches_here_p (occr
, expr
, bb
, 0))
3299 insn_computes_expr
= occr
->insn
;
3303 if (insn_computes_expr
== NULL
)
3306 return insn_computes_expr
;
3310 /* Return non-zero if the definition in DEF_INSN can reach INSN.
3311 Only called by can_disregard_other_sets. */
3314 def_reaches_here_p (insn
, def_insn
)
3319 if (TEST_BIT (reaching_defs
[BLOCK_NUM (insn
)], INSN_CUID (def_insn
)))
3322 if (BLOCK_NUM (insn
) == BLOCK_NUM (def_insn
))
3324 if (INSN_CUID (def_insn
) < INSN_CUID (insn
))
3326 if (GET_CODE (PATTERN (def_insn
)) == PARALLEL
)
3328 else if (GET_CODE (PATTERN (def_insn
)) == CLOBBER
)
3329 reg
= XEXP (PATTERN (def_insn
), 0);
3330 else if (GET_CODE (PATTERN (def_insn
)) == SET
)
3331 reg
= SET_DEST (PATTERN (def_insn
));
3335 return ! reg_set_between_p (reg
, NEXT_INSN (def_insn
), insn
);
3344 /* Return non-zero if *ADDR_THIS_REG can only have one value at INSN. The
3345 value returned is the number of definitions that reach INSN. Returning a
3346 value of zero means that [maybe] more than one definition reaches INSN and
3347 the caller can't perform whatever optimization it is trying. i.e. it is
3348 always safe to return zero. */
3351 can_disregard_other_sets (addr_this_reg
, insn
, for_combine
)
3352 struct reg_set
**addr_this_reg
;
3356 int number_of_reaching_defs
= 0;
3357 struct reg_set
*this_reg
;
3359 for (this_reg
= *addr_this_reg
; this_reg
!= 0; this_reg
= this_reg
->next
)
3360 if (def_reaches_here_p (insn
, this_reg
->insn
))
3362 number_of_reaching_defs
++;
3363 /* Ignore parallels for now. */
3364 if (GET_CODE (PATTERN (this_reg
->insn
)) == PARALLEL
)
3368 && (GET_CODE (PATTERN (this_reg
->insn
)) == CLOBBER
3369 || ! rtx_equal_p (SET_SRC (PATTERN (this_reg
->insn
)),
3370 SET_SRC (PATTERN (insn
)))))
3371 /* A setting of the reg to a different value reaches INSN. */
3374 if (number_of_reaching_defs
> 1)
3376 /* If in this setting the value the register is being set to is
3377 equal to the previous value the register was set to and this
3378 setting reaches the insn we are trying to do the substitution
3379 on then we are ok. */
3380 if (GET_CODE (PATTERN (this_reg
->insn
)) == CLOBBER
)
3382 else if (! rtx_equal_p (SET_SRC (PATTERN (this_reg
->insn
)),
3383 SET_SRC (PATTERN (insn
))))
3387 *addr_this_reg
= this_reg
;
3390 return number_of_reaching_defs
;
3393 /* Expression computed by insn is available and the substitution is legal,
3394 so try to perform the substitution.
3396 The result is non-zero if any changes were made. */
3399 handle_avail_expr (insn
, expr
)
3403 rtx pat
, insn_computes_expr
, expr_set
;
3405 struct reg_set
*this_reg
;
3406 int found_setting
, use_src
;
3409 /* We only handle the case where one computation of the expression
3410 reaches this instruction. */
3411 insn_computes_expr
= computing_insn (expr
, insn
);
3412 if (insn_computes_expr
== NULL
)
3414 expr_set
= single_set (insn_computes_expr
);
3421 /* At this point we know only one computation of EXPR outside of this
3422 block reaches this insn. Now try to find a register that the
3423 expression is computed into. */
3424 if (GET_CODE (SET_SRC (expr_set
)) == REG
)
3426 /* This is the case when the available expression that reaches
3427 here has already been handled as an available expression. */
3428 unsigned int regnum_for_replacing
3429 = REGNO (SET_SRC (expr_set
));
3431 /* If the register was created by GCSE we can't use `reg_set_table',
3432 however we know it's set only once. */
3433 if (regnum_for_replacing
>= max_gcse_regno
3434 /* If the register the expression is computed into is set only once,
3435 or only one set reaches this insn, we can use it. */
3436 || (((this_reg
= reg_set_table
[regnum_for_replacing
]),
3437 this_reg
->next
== NULL
)
3438 || can_disregard_other_sets (&this_reg
, insn
, 0)))
3447 unsigned int regnum_for_replacing
3448 = REGNO (SET_DEST (expr_set
));
3450 /* This shouldn't happen. */
3451 if (regnum_for_replacing
>= max_gcse_regno
)
3454 this_reg
= reg_set_table
[regnum_for_replacing
];
3456 /* If the register the expression is computed into is set only once,
3457 or only one set reaches this insn, use it. */
3458 if (this_reg
->next
== NULL
3459 || can_disregard_other_sets (&this_reg
, insn
, 0))
3465 pat
= PATTERN (insn
);
3467 to
= SET_SRC (expr_set
);
3469 to
= SET_DEST (expr_set
);
3470 changed
= validate_change (insn
, &SET_SRC (pat
), to
, 0);
3472 /* We should be able to ignore the return code from validate_change but
3473 to play it safe we check. */
3477 if (gcse_file
!= NULL
)
3479 fprintf (gcse_file
, "GCSE: Replacing the source in insn %d with",
3481 fprintf (gcse_file
, " reg %d %s insn %d\n",
3482 REGNO (to
), use_src
? "from" : "set in",
3483 INSN_UID (insn_computes_expr
));
3488 /* The register that the expr is computed into is set more than once. */
3489 else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
3491 /* Insert an insn after insnx that copies the reg set in insnx
3492 into a new pseudo register call this new register REGN.
3493 From insnb until end of basic block or until REGB is set
3494 replace all uses of REGB with REGN. */
3497 to
= gen_reg_rtx (GET_MODE (SET_DEST (expr_set
)));
3499 /* Generate the new insn. */
3500 /* ??? If the change fails, we return 0, even though we created
3501 an insn. I think this is ok. */
3503 = emit_insn_after (gen_rtx_SET (VOIDmode
, to
,
3504 SET_DEST (expr_set
)),
3505 insn_computes_expr
);
3507 /* Keep register set table up to date. */
3508 record_one_set (REGNO (to
), new_insn
);
3510 gcse_create_count
++;
3511 if (gcse_file
!= NULL
)
3513 fprintf (gcse_file
, "GCSE: Creating insn %d to copy value of reg %d",
3514 INSN_UID (NEXT_INSN (insn_computes_expr
)),
3515 REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr
)))));
3516 fprintf (gcse_file
, ", computed in insn %d,\n",
3517 INSN_UID (insn_computes_expr
));
3518 fprintf (gcse_file
, " into newly allocated reg %d\n",
3522 pat
= PATTERN (insn
);
3524 /* Do register replacement for INSN. */
3525 changed
= validate_change (insn
, &SET_SRC (pat
),
3527 (NEXT_INSN (insn_computes_expr
))),
3530 /* We should be able to ignore the return code from validate_change but
3531 to play it safe we check. */
3535 if (gcse_file
!= NULL
)
3538 "GCSE: Replacing the source in insn %d with reg %d ",
3540 REGNO (SET_DEST (PATTERN (NEXT_INSN
3541 (insn_computes_expr
)))));
3542 fprintf (gcse_file
, "set in insn %d\n",
3543 INSN_UID (insn_computes_expr
));
3551 /* Perform classic GCSE. This is called by one_classic_gcse_pass after all
3552 the dataflow analysis has been done.
3554 The result is non-zero if a change was made. */
3562 /* Note we start at block 1. */
3565 for (bb
= 1; bb
< n_basic_blocks
; bb
++)
3567 /* Reset tables used to keep track of what's still valid [since the
3568 start of the block]. */
3569 reset_opr_set_tables ();
3571 for (insn
= BLOCK_HEAD (bb
);
3572 insn
!= NULL
&& insn
!= NEXT_INSN (BLOCK_END (bb
));
3573 insn
= NEXT_INSN (insn
))
3575 /* Is insn of form (set (pseudo-reg) ...)? */
3576 if (GET_CODE (insn
) == INSN
3577 && GET_CODE (PATTERN (insn
)) == SET
3578 && GET_CODE (SET_DEST (PATTERN (insn
))) == REG
3579 && REGNO (SET_DEST (PATTERN (insn
))) >= FIRST_PSEUDO_REGISTER
)
3581 rtx pat
= PATTERN (insn
);
3582 rtx src
= SET_SRC (pat
);
3585 if (want_to_gcse_p (src
)
3586 /* Is the expression recorded? */
3587 && ((expr
= lookup_expr (src
)) != NULL
)
3588 /* Is the expression available [at the start of the
3590 && TEST_BIT (ae_in
[bb
], expr
->bitmap_index
)
3591 /* Are the operands unchanged since the start of the
3593 && oprs_not_set_p (src
, insn
))
3594 changed
|= handle_avail_expr (insn
, expr
);
3597 /* Keep track of everything modified by this insn. */
3598 /* ??? Need to be careful w.r.t. mods done to INSN. */
3600 mark_oprs_set (insn
);
3607 /* Top level routine to perform one classic GCSE pass.
3609 Return non-zero if a change was made. */
3612 one_classic_gcse_pass (pass
)
3617 gcse_subst_count
= 0;
3618 gcse_create_count
= 0;
3620 alloc_expr_hash_table (max_cuid
);
3621 alloc_rd_mem (n_basic_blocks
, max_cuid
);
3622 compute_expr_hash_table ();
3624 dump_hash_table (gcse_file
, "Expression", expr_hash_table
,
3625 expr_hash_table_size
, n_exprs
);
3631 alloc_avail_expr_mem (n_basic_blocks
, n_exprs
);
3633 compute_ae_kill (ae_gen
, ae_kill
);
3634 compute_available (ae_gen
, ae_kill
, ae_out
, ae_in
);
3635 changed
= classic_gcse ();
3636 free_avail_expr_mem ();
3640 free_expr_hash_table ();
3644 fprintf (gcse_file
, "\n");
3645 fprintf (gcse_file
, "GCSE of %s, pass %d: %d bytes needed, %d substs,",
3646 current_function_name
, pass
, bytes_used
, gcse_subst_count
);
3647 fprintf (gcse_file
, "%d insns created\n", gcse_create_count
);
3653 /* Compute copy/constant propagation working variables. */
3655 /* Local properties of assignments. */
3656 static sbitmap
*cprop_pavloc
;
3657 static sbitmap
*cprop_absaltered
;
3659 /* Global properties of assignments (computed from the local properties). */
3660 static sbitmap
*cprop_avin
;
3661 static sbitmap
*cprop_avout
;
3663 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
3664 basic blocks. N_SETS is the number of sets. */
3667 alloc_cprop_mem (n_blocks
, n_sets
)
3668 int n_blocks
, n_sets
;
3670 cprop_pavloc
= sbitmap_vector_alloc (n_blocks
, n_sets
);
3671 cprop_absaltered
= sbitmap_vector_alloc (n_blocks
, n_sets
);
3673 cprop_avin
= sbitmap_vector_alloc (n_blocks
, n_sets
);
3674 cprop_avout
= sbitmap_vector_alloc (n_blocks
, n_sets
);
3677 /* Free vars used by copy/const propagation. */
3682 sbitmap_vector_free (cprop_pavloc
);
3683 sbitmap_vector_free (cprop_absaltered
);
3684 sbitmap_vector_free (cprop_avin
);
3685 sbitmap_vector_free (cprop_avout
);
3688 /* For each block, compute whether X is transparent. X is either an
3689 expression or an assignment [though we don't care which, for this context
3690 an assignment is treated as an expression]. For each block where an
3691 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
3695 compute_transp (x
, indx
, bmap
, set_p
)
3706 /* repeat is used to turn tail-recursion into iteration since GCC
3707 can't do it when there's no return value. */
3713 code
= GET_CODE (x
);
3719 if (REGNO (x
) < FIRST_PSEUDO_REGISTER
)
3721 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3722 if (TEST_BIT (reg_set_in_block
[bb
], REGNO (x
)))
3723 SET_BIT (bmap
[bb
], indx
);
3727 for (r
= reg_set_table
[REGNO (x
)]; r
!= NULL
; r
= r
->next
)
3728 SET_BIT (bmap
[BLOCK_NUM (r
->insn
)], indx
);
3733 if (REGNO (x
) < FIRST_PSEUDO_REGISTER
)
3735 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3736 if (TEST_BIT (reg_set_in_block
[bb
], REGNO (x
)))
3737 RESET_BIT (bmap
[bb
], indx
);
3741 for (r
= reg_set_table
[REGNO (x
)]; r
!= NULL
; r
= r
->next
)
3742 RESET_BIT (bmap
[BLOCK_NUM (r
->insn
)], indx
);
3749 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
3751 rtx list_entry
= canon_modify_mem_list
[bb
];
3755 rtx dest
, dest_addr
;
3757 if (GET_CODE (XEXP (list_entry
, 0)) == CALL_INSN
)
3760 SET_BIT (bmap
[bb
], indx
);
3762 RESET_BIT (bmap
[bb
], indx
);
3765 /* LIST_ENTRY must be an INSN of some kind that sets memory.
3766 Examine each hunk of memory that is modified. */
3768 dest
= XEXP (list_entry
, 0);
3769 list_entry
= XEXP (list_entry
, 1);
3770 dest_addr
= XEXP (list_entry
, 0);
3772 if (canon_true_dependence (dest
, GET_MODE (dest
), dest_addr
,
3773 x
, rtx_addr_varies_p
))
3776 SET_BIT (bmap
[bb
], indx
);
3778 RESET_BIT (bmap
[bb
], indx
);
3781 list_entry
= XEXP (list_entry
, 1);
3803 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
3807 /* If we are about to do the last recursive call
3808 needed at this level, change it into iteration.
3809 This function is called enough to be worth it. */
3816 compute_transp (XEXP (x
, i
), indx
, bmap
, set_p
);
3818 else if (fmt
[i
] == 'E')
3819 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3820 compute_transp (XVECEXP (x
, i
, j
), indx
, bmap
, set_p
);
3824 /* Top level routine to do the dataflow analysis needed by copy/const
3828 compute_cprop_data ()
3830 compute_local_properties (cprop_absaltered
, cprop_pavloc
, NULL
, 1);
3831 compute_available (cprop_pavloc
, cprop_absaltered
,
3832 cprop_avout
, cprop_avin
);
3835 /* Copy/constant propagation. */
3837 /* Maximum number of register uses in an insn that we handle. */
3840 /* Table of uses found in an insn.
3841 Allocated statically to avoid alloc/free complexity and overhead. */
3842 static struct reg_use reg_use_table
[MAX_USES
];
3844 /* Index into `reg_use_table' while building it. */
3845 static int reg_use_count
;
3847 /* Set up a list of register numbers used in INSN. The found uses are stored
3848 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
3849 and contains the number of uses in the table upon exit.
3851 ??? If a register appears multiple times we will record it multiple times.
3852 This doesn't hurt anything but it will slow things down. */
3855 find_used_regs (xptr
, data
)
3857 void *data ATTRIBUTE_UNUSED
;
3864 /* repeat is used to turn tail-recursion into iteration since GCC
3865 can't do it when there's no return value. */
3870 code
= GET_CODE (x
);
3873 if (reg_use_count
== MAX_USES
)
3876 reg_use_table
[reg_use_count
].reg_rtx
= x
;
3880 /* Recursively scan the operands of this expression. */
3882 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
3886 /* If we are about to do the last recursive call
3887 needed at this level, change it into iteration.
3888 This function is called enough to be worth it. */
3895 find_used_regs (&XEXP (x
, i
), data
);
3897 else if (fmt
[i
] == 'E')
3898 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
3899 find_used_regs (&XVECEXP (x
, i
, j
), data
);
3903 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
3904 Returns non-zero is successful. */
3907 try_replace_reg (from
, to
, insn
)
3910 rtx note
= find_reg_equal_equiv_note (insn
);
3913 rtx set
= single_set (insn
);
3915 success
= validate_replace_src (from
, to
, insn
);
3917 /* If above failed and this is a single set, try to simplify the source of
3918 the set given our substitution. We could perhaps try this for multiple
3919 SETs, but it probably won't buy us anything. */
3920 if (!success
&& set
!= 0)
3922 src
= simplify_replace_rtx (SET_SRC (set
), from
, to
);
3924 if (!rtx_equal_p (src
, SET_SRC (set
))
3925 && validate_change (insn
, &SET_SRC (set
), src
, 0))
3929 /* If we've failed to do replacement, have a single SET, and don't already
3930 have a note, add a REG_EQUAL note to not lose information. */
3931 if (!success
&& note
== 0 && set
!= 0)
3932 note
= set_unique_reg_note (insn
, REG_EQUAL
, src
);
3934 /* If there is already a NOTE, update the expression in it with our
3937 XEXP (note
, 0) = simplify_replace_rtx (XEXP (note
, 0), from
, to
);
3939 /* REG_EQUAL may get simplified into register.
3940 We don't allow that. Remove that note. This code ought
3941 not to hapen, because previous code ought to syntetize
3942 reg-reg move, but be on the safe side. */
3943 if (note
&& REG_P (XEXP (note
, 0)))
3944 remove_note (insn
, note
);
3949 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
3950 NULL no such set is found. */
3952 static struct expr
*
3953 find_avail_set (regno
, insn
)
3957 /* SET1 contains the last set found that can be returned to the caller for
3958 use in a substitution. */
3959 struct expr
*set1
= 0;
3961 /* Loops are not possible here. To get a loop we would need two sets
3962 available at the start of the block containing INSN. ie we would
3963 need two sets like this available at the start of the block:
3965 (set (reg X) (reg Y))
3966 (set (reg Y) (reg X))
3968 This can not happen since the set of (reg Y) would have killed the
3969 set of (reg X) making it unavailable at the start of this block. */
3973 struct expr
*set
= lookup_set (regno
, NULL_RTX
);
3975 /* Find a set that is available at the start of the block
3976 which contains INSN. */
3979 if (TEST_BIT (cprop_avin
[BLOCK_NUM (insn
)], set
->bitmap_index
))
3981 set
= next_set (regno
, set
);
3984 /* If no available set was found we've reached the end of the
3985 (possibly empty) copy chain. */
3989 if (GET_CODE (set
->expr
) != SET
)
3992 src
= SET_SRC (set
->expr
);
3994 /* We know the set is available.
3995 Now check that SRC is ANTLOC (i.e. none of the source operands
3996 have changed since the start of the block).
3998 If the source operand changed, we may still use it for the next
3999 iteration of this loop, but we may not use it for substitutions. */
4001 if (CONSTANT_P (src
) || oprs_not_set_p (src
, insn
))
4004 /* If the source of the set is anything except a register, then
4005 we have reached the end of the copy chain. */
4006 if (GET_CODE (src
) != REG
)
4009 /* Follow the copy chain, ie start another iteration of the loop
4010 and see if we have an available copy into SRC. */
4011 regno
= REGNO (src
);
4014 /* SET1 holds the last set that was available and anticipatable at
4019 /* Subroutine of cprop_insn that tries to propagate constants into
4020 JUMP_INSNS. INSN must be a conditional jump. FROM is what we will try to
4021 replace, SRC is the constant we will try to substitute for it. Returns
4022 nonzero if a change was made. We know INSN has just a SET. */
4025 cprop_jump (bb
, insn
, from
, src
)
4031 rtx set
= PATTERN (insn
);
4032 rtx
new = simplify_replace_rtx (SET_SRC (set
), from
, src
);
4034 /* If no simplification can be made, then try the next
4036 if (rtx_equal_p (new, SET_SRC (set
)))
4039 /* If this is now a no-op leave it that way, but update LABEL_NUSED if
4043 SET_SRC (set
) = new;
4045 if (JUMP_LABEL (insn
) != 0)
4046 --LABEL_NUSES (JUMP_LABEL (insn
));
4049 /* Otherwise, this must be a valid instruction. */
4050 else if (! validate_change (insn
, &SET_SRC (set
), new, 0))
4053 /* If this has turned into an unconditional jump,
4054 then put a barrier after it so that the unreachable
4055 code will be deleted. */
4056 if (GET_CODE (SET_SRC (set
)) == LABEL_REF
)
4057 emit_barrier_after (insn
);
4059 run_jump_opt_after_gcse
= 1;
4062 if (gcse_file
!= NULL
)
4065 "CONST-PROP: Replacing reg %d in insn %d with constant ",
4066 REGNO (from
), INSN_UID (insn
));
4067 print_rtl (gcse_file
, src
);
4068 fprintf (gcse_file
, "\n");
4070 purge_dead_edges (bb
);
4077 /* Subroutine of cprop_insn that tries to propagate constants into JUMP_INSNS
4078 for machines that have CC0. INSN is a single set that stores into CC0;
4079 the insn following it is a conditional jump. REG_USED is the use we will
4080 try to replace, SRC is the constant we will try to substitute for it.
4081 Returns nonzero if a change was made. */
4084 cprop_cc0_jump (bb
, insn
, reg_used
, src
)
4087 struct reg_use
*reg_used
;
4090 /* First substitute in the SET_SRC of INSN, then substitute that for
4092 rtx jump
= NEXT_INSN (insn
);
4093 rtx new_src
= simplify_replace_rtx (SET_SRC (PATTERN (insn
)),
4094 reg_used
->reg_rtx
, src
);
4096 if (! cprop_jump (bb
, jump
, cc0_rtx
, new_src
))
4099 /* If we succeeded, delete the cc0 setter. */
4106 /* Perform constant and copy propagation on INSN.
4107 The result is non-zero if a change was made. */
4110 cprop_insn (bb
, insn
, alter_jumps
)
4115 struct reg_use
*reg_used
;
4123 note_uses (&PATTERN (insn
), find_used_regs
, NULL
);
4125 note
= find_reg_equal_equiv_note (insn
);
4127 /* We may win even when propagating constants into notes. */
4129 find_used_regs (&XEXP (note
, 0), NULL
);
4131 for (reg_used
= ®_use_table
[0]; reg_use_count
> 0;
4132 reg_used
++, reg_use_count
--)
4134 unsigned int regno
= REGNO (reg_used
->reg_rtx
);
4138 /* Ignore registers created by GCSE.
4139 We do this because ... */
4140 if (regno
>= max_gcse_regno
)
4143 /* If the register has already been set in this block, there's
4144 nothing we can do. */
4145 if (! oprs_not_set_p (reg_used
->reg_rtx
, insn
))
4148 /* Find an assignment that sets reg_used and is available
4149 at the start of the block. */
4150 set
= find_avail_set (regno
, insn
);
4155 /* ??? We might be able to handle PARALLELs. Later. */
4156 if (GET_CODE (pat
) != SET
)
4159 src
= SET_SRC (pat
);
4161 /* Constant propagation. */
4162 if (GET_CODE (src
) == CONST_INT
|| GET_CODE (src
) == CONST_DOUBLE
4163 || GET_CODE (src
) == SYMBOL_REF
)
4165 /* Handle normal insns first. */
4166 if (GET_CODE (insn
) == INSN
4167 && try_replace_reg (reg_used
->reg_rtx
, src
, insn
))
4171 if (gcse_file
!= NULL
)
4173 fprintf (gcse_file
, "CONST-PROP: Replacing reg %d in ",
4175 fprintf (gcse_file
, "insn %d with constant ",
4177 print_rtl (gcse_file
, src
);
4178 fprintf (gcse_file
, "\n");
4181 /* The original insn setting reg_used may or may not now be
4182 deletable. We leave the deletion to flow. */
4185 /* Try to propagate a CONST_INT into a conditional jump.
4186 We're pretty specific about what we will handle in this
4187 code, we can extend this as necessary over time.
4189 Right now the insn in question must look like
4190 (set (pc) (if_then_else ...)) */
4191 else if (alter_jumps
4192 && GET_CODE (insn
) == JUMP_INSN
4193 && condjump_p (insn
)
4194 && ! simplejump_p (insn
))
4195 changed
|= cprop_jump (bb
, insn
, reg_used
->reg_rtx
, src
);
4198 /* Similar code for machines that use a pair of CC0 setter and
4199 conditional jump insn. */
4200 else if (alter_jumps
4201 && GET_CODE (PATTERN (insn
)) == SET
4202 && SET_DEST (PATTERN (insn
)) == cc0_rtx
4203 && GET_CODE (NEXT_INSN (insn
)) == JUMP_INSN
4204 && condjump_p (NEXT_INSN (insn
))
4205 && ! simplejump_p (NEXT_INSN (insn
))
4206 && cprop_cc0_jump (bb
, insn
, reg_used
, src
))
4213 else if (GET_CODE (src
) == REG
4214 && REGNO (src
) >= FIRST_PSEUDO_REGISTER
4215 && REGNO (src
) != regno
)
4217 if (try_replace_reg (reg_used
->reg_rtx
, src
, insn
))
4221 if (gcse_file
!= NULL
)
4223 fprintf (gcse_file
, "COPY-PROP: Replacing reg %d in insn %d",
4224 regno
, INSN_UID (insn
));
4225 fprintf (gcse_file
, " with reg %d\n", REGNO (src
));
4228 /* The original insn setting reg_used may or may not now be
4229 deletable. We leave the deletion to flow. */
4230 /* FIXME: If it turns out that the insn isn't deletable,
4231 then we may have unnecessarily extended register lifetimes
4232 and made things worse. */
4240 /* Forward propagate copies. This includes copies and constants. Return
4241 non-zero if a change was made. */
4250 /* Note we start at block 1. */
4253 for (bb
= 1; bb
< n_basic_blocks
; bb
++)
4255 /* Reset tables used to keep track of what's still valid [since the
4256 start of the block]. */
4257 reset_opr_set_tables ();
4259 for (insn
= BLOCK_HEAD (bb
);
4260 insn
!= NULL
&& insn
!= NEXT_INSN (BLOCK_END (bb
));
4261 insn
= NEXT_INSN (insn
))
4264 changed
|= cprop_insn (BASIC_BLOCK (bb
), insn
, alter_jumps
);
4266 /* Keep track of everything modified by this insn. */
4267 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
4268 call mark_oprs_set if we turned the insn into a NOTE. */
4269 if (GET_CODE (insn
) != NOTE
)
4270 mark_oprs_set (insn
);
4274 if (gcse_file
!= NULL
)
4275 fprintf (gcse_file
, "\n");
4280 /* Perform one copy/constant propagation pass.
4281 F is the first insn in the function.
4282 PASS is the pass count. */
4285 one_cprop_pass (pass
, alter_jumps
)
4291 const_prop_count
= 0;
4292 copy_prop_count
= 0;
4294 alloc_set_hash_table (max_cuid
);
4295 compute_set_hash_table ();
4297 dump_hash_table (gcse_file
, "SET", set_hash_table
, set_hash_table_size
,
4301 alloc_cprop_mem (n_basic_blocks
, n_sets
);
4302 compute_cprop_data ();
4303 changed
= cprop (alter_jumps
);
4307 free_set_hash_table ();
4311 fprintf (gcse_file
, "CPROP of %s, pass %d: %d bytes needed, ",
4312 current_function_name
, pass
, bytes_used
);
4313 fprintf (gcse_file
, "%d const props, %d copy props\n\n",
4314 const_prop_count
, copy_prop_count
);
4320 /* Compute PRE+LCM working variables. */
4322 /* Local properties of expressions. */
4323 /* Nonzero for expressions that are transparent in the block. */
4324 static sbitmap
*transp
;
4326 /* Nonzero for expressions that are transparent at the end of the block.
4327 This is only zero for expressions killed by abnormal critical edge
4328 created by a calls. */
4329 static sbitmap
*transpout
;
4331 /* Nonzero for expressions that are computed (available) in the block. */
4332 static sbitmap
*comp
;
4334 /* Nonzero for expressions that are locally anticipatable in the block. */
4335 static sbitmap
*antloc
;
4337 /* Nonzero for expressions where this block is an optimal computation
4339 static sbitmap
*pre_optimal
;
4341 /* Nonzero for expressions which are redundant in a particular block. */
4342 static sbitmap
*pre_redundant
;
4344 /* Nonzero for expressions which should be inserted on a specific edge. */
4345 static sbitmap
*pre_insert_map
;
4347 /* Nonzero for expressions which should be deleted in a specific block. */
4348 static sbitmap
*pre_delete_map
;
4350 /* Contains the edge_list returned by pre_edge_lcm. */
4351 static struct edge_list
*edge_list
;
4353 /* Redundant insns. */
4354 static sbitmap pre_redundant_insns
;
4356 /* Allocate vars used for PRE analysis. */
4359 alloc_pre_mem (n_blocks
, n_exprs
)
4360 int n_blocks
, n_exprs
;
4362 transp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4363 comp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4364 antloc
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4367 pre_redundant
= NULL
;
4368 pre_insert_map
= NULL
;
4369 pre_delete_map
= NULL
;
4372 ae_kill
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4374 /* pre_insert and pre_delete are allocated later. */
4377 /* Free vars used for PRE analysis. */
4382 sbitmap_vector_free (transp
);
4383 sbitmap_vector_free (comp
);
4385 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
4388 sbitmap_vector_free (pre_optimal
);
4390 sbitmap_vector_free (pre_redundant
);
4392 sbitmap_vector_free (pre_insert_map
);
4394 sbitmap_vector_free (pre_delete_map
);
4396 sbitmap_vector_free (ae_in
);
4398 sbitmap_vector_free (ae_out
);
4400 transp
= comp
= NULL
;
4401 pre_optimal
= pre_redundant
= pre_insert_map
= pre_delete_map
= NULL
;
4402 ae_in
= ae_out
= NULL
;
4405 /* Top level routine to do the dataflow analysis needed by PRE. */
4410 sbitmap trapping_expr
;
4414 compute_local_properties (transp
, comp
, antloc
, 0);
4415 sbitmap_vector_zero (ae_kill
, n_basic_blocks
);
4417 /* Collect expressions which might trap. */
4418 trapping_expr
= sbitmap_alloc (n_exprs
);
4419 sbitmap_zero (trapping_expr
);
4420 for (ui
= 0; ui
< expr_hash_table_size
; ui
++)
4423 for (e
= expr_hash_table
[ui
]; e
!= NULL
; e
= e
->next_same_hash
)
4424 if (may_trap_p (e
->expr
))
4425 SET_BIT (trapping_expr
, e
->bitmap_index
);
4428 /* Compute ae_kill for each basic block using:
4432 This is significantly faster than compute_ae_kill. */
4434 for (i
= 0; i
< n_basic_blocks
; i
++)
4438 /* If the current block is the destination of an abnormal edge, we
4439 kill all trapping expressions because we won't be able to properly
4440 place the instruction on the edge. So make them neither
4441 anticipatable nor transparent. This is fairly conservative. */
4442 for (e
= BASIC_BLOCK (i
)->pred
; e
; e
= e
->pred_next
)
4443 if (e
->flags
& EDGE_ABNORMAL
)
4445 sbitmap_difference (antloc
[i
], antloc
[i
], trapping_expr
);
4446 sbitmap_difference (transp
[i
], transp
[i
], trapping_expr
);
4450 sbitmap_a_or_b (ae_kill
[i
], transp
[i
], comp
[i
]);
4451 sbitmap_not (ae_kill
[i
], ae_kill
[i
]);
4454 edge_list
= pre_edge_lcm (gcse_file
, n_exprs
, transp
, comp
, antloc
,
4455 ae_kill
, &pre_insert_map
, &pre_delete_map
);
4456 sbitmap_vector_free (antloc
);
4458 sbitmap_vector_free (ae_kill
);
4460 free (trapping_expr
);
4465 /* Return non-zero if an occurrence of expression EXPR in OCCR_BB would reach
4468 VISITED is a pointer to a working buffer for tracking which BB's have
4469 been visited. It is NULL for the top-level call.
4471 We treat reaching expressions that go through blocks containing the same
4472 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
4473 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
4474 2 as not reaching. The intent is to improve the probability of finding
4475 only one reaching expression and to reduce register lifetimes by picking
4476 the closest such expression. */
4479 pre_expr_reaches_here_p_work (occr_bb
, expr
, bb
, visited
)
4480 basic_block occr_bb
;
4487 for (pred
= bb
->pred
; pred
!= NULL
; pred
= pred
->pred_next
)
4489 basic_block pred_bb
= pred
->src
;
4491 if (pred
->src
== ENTRY_BLOCK_PTR
4492 /* Has predecessor has already been visited? */
4493 || visited
[pred_bb
->index
])
4494 ;/* Nothing to do. */
4496 /* Does this predecessor generate this expression? */
4497 else if (TEST_BIT (comp
[pred_bb
->index
], expr
->bitmap_index
))
4499 /* Is this the occurrence we're looking for?
4500 Note that there's only one generating occurrence per block
4501 so we just need to check the block number. */
4502 if (occr_bb
== pred_bb
)
4505 visited
[pred_bb
->index
] = 1;
4507 /* Ignore this predecessor if it kills the expression. */
4508 else if (! TEST_BIT (transp
[pred_bb
->index
], expr
->bitmap_index
))
4509 visited
[pred_bb
->index
] = 1;
4511 /* Neither gen nor kill. */
4514 visited
[pred_bb
->index
] = 1;
4515 if (pre_expr_reaches_here_p_work (occr_bb
, expr
, pred_bb
, visited
))
4520 /* All paths have been checked. */
4524 /* The wrapper for pre_expr_reaches_here_work that ensures that any
4525 memory allocated for that function is returned. */
4528 pre_expr_reaches_here_p (occr_bb
, expr
, bb
)
4529 basic_block occr_bb
;
4534 char *visited
= (char *) xcalloc (n_basic_blocks
, 1);
4536 rval
= pre_expr_reaches_here_p_work(occr_bb
, expr
, bb
, visited
);
4543 /* Given an expr, generate RTL which we can insert at the end of a BB,
4544 or on an edge. Set the block number of any insns generated to
4548 process_insert_insn (expr
)
4551 rtx reg
= expr
->reaching_reg
;
4552 rtx exp
= copy_rtx (expr
->expr
);
4557 /* If the expression is something that's an operand, like a constant,
4558 just copy it to a register. */
4559 if (general_operand (exp
, GET_MODE (reg
)))
4560 emit_move_insn (reg
, exp
);
4562 /* Otherwise, make a new insn to compute this expression and make sure the
4563 insn will be recognized (this also adds any needed CLOBBERs). Copy the
4564 expression to make sure we don't have any sharing issues. */
4565 else if (insn_invalid_p (emit_insn (gen_rtx_SET (VOIDmode
, reg
, exp
))))
4568 pat
= gen_sequence ();
4574 /* Add EXPR to the end of basic block BB.
4576 This is used by both the PRE and code hoisting.
4578 For PRE, we want to verify that the expr is either transparent
4579 or locally anticipatable in the target block. This check makes
4580 no sense for code hoisting. */
4583 insert_insn_end_bb (expr
, bb
, pre
)
4590 rtx reg
= expr
->reaching_reg
;
4591 int regno
= REGNO (reg
);
4595 pat
= process_insert_insn (expr
);
4597 /* If the last insn is a jump, insert EXPR in front [taking care to
4598 handle cc0, etc. properly]. */
4600 if (GET_CODE (insn
) == JUMP_INSN
)
4606 /* If this is a jump table, then we can't insert stuff here. Since
4607 we know the previous real insn must be the tablejump, we insert
4608 the new instruction just before the tablejump. */
4609 if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
4610 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
4611 insn
= prev_real_insn (insn
);
4614 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
4615 if cc0 isn't set. */
4616 note
= find_reg_note (insn
, REG_CC_SETTER
, NULL_RTX
);
4618 insn
= XEXP (note
, 0);
4621 rtx maybe_cc0_setter
= prev_nonnote_insn (insn
);
4622 if (maybe_cc0_setter
4623 && INSN_P (maybe_cc0_setter
)
4624 && sets_cc0_p (PATTERN (maybe_cc0_setter
)))
4625 insn
= maybe_cc0_setter
;
4628 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4629 new_insn
= emit_insn_before (pat
, insn
);
4632 /* Likewise if the last insn is a call, as will happen in the presence
4633 of exception handling. */
4634 else if (GET_CODE (insn
) == CALL_INSN
)
4636 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4637 we search backward and place the instructions before the first
4638 parameter is loaded. Do this for everyone for consistency and a
4639 presumtion that we'll get better code elsewhere as well.
4641 It should always be the case that we can put these instructions
4642 anywhere in the basic block with performing PRE optimizations.
4646 && !TEST_BIT (antloc
[bb
->index
], expr
->bitmap_index
)
4647 && !TEST_BIT (transp
[bb
->index
], expr
->bitmap_index
))
4650 /* Since different machines initialize their parameter registers
4651 in different orders, assume nothing. Collect the set of all
4652 parameter registers. */
4653 insn
= find_first_parameter_load (insn
, bb
->head
);
4655 /* If we found all the parameter loads, then we want to insert
4656 before the first parameter load.
4658 If we did not find all the parameter loads, then we might have
4659 stopped on the head of the block, which could be a CODE_LABEL.
4660 If we inserted before the CODE_LABEL, then we would be putting
4661 the insn in the wrong basic block. In that case, put the insn
4662 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4663 while (GET_CODE (insn
) == CODE_LABEL
4664 || NOTE_INSN_BASIC_BLOCK_P (insn
))
4665 insn
= NEXT_INSN (insn
);
4667 new_insn
= emit_insn_before (pat
, insn
);
4670 new_insn
= emit_insn_after (pat
, insn
);
4672 /* Keep block number table up to date.
4673 Note, PAT could be a multiple insn sequence, we have to make
4674 sure that each insn in the sequence is handled. */
4675 if (GET_CODE (pat
) == SEQUENCE
)
4677 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
4679 rtx insn
= XVECEXP (pat
, 0, i
);
4681 add_label_notes (PATTERN (insn
), new_insn
);
4683 note_stores (PATTERN (insn
), record_set_info
, insn
);
4688 add_label_notes (pat
, new_insn
);
4690 /* Keep register set table up to date. */
4691 record_one_set (regno
, new_insn
);
4694 gcse_create_count
++;
4698 fprintf (gcse_file
, "PRE/HOIST: end of bb %d, insn %d, ",
4699 bb
->index
, INSN_UID (new_insn
));
4700 fprintf (gcse_file
, "copying expression %d to reg %d\n",
4701 expr
->bitmap_index
, regno
);
4705 /* Insert partially redundant expressions on edges in the CFG to make
4706 the expressions fully redundant. */
4709 pre_edge_insert (edge_list
, index_map
)
4710 struct edge_list
*edge_list
;
4711 struct expr
**index_map
;
4713 int e
, i
, j
, num_edges
, set_size
, did_insert
= 0;
4716 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4717 if it reaches any of the deleted expressions. */
4719 set_size
= pre_insert_map
[0]->size
;
4720 num_edges
= NUM_EDGES (edge_list
);
4721 inserted
= sbitmap_vector_alloc (num_edges
, n_exprs
);
4722 sbitmap_vector_zero (inserted
, num_edges
);
4724 for (e
= 0; e
< num_edges
; e
++)
4727 basic_block bb
= INDEX_EDGE_PRED_BB (edge_list
, e
);
4729 for (i
= indx
= 0; i
< set_size
; i
++, indx
+= SBITMAP_ELT_BITS
)
4731 SBITMAP_ELT_TYPE insert
= pre_insert_map
[e
]->elms
[i
];
4733 for (j
= indx
; insert
&& j
< n_exprs
; j
++, insert
>>= 1)
4734 if ((insert
& 1) != 0 && index_map
[j
]->reaching_reg
!= NULL_RTX
)
4736 struct expr
*expr
= index_map
[j
];
4739 /* Now look at each deleted occurrence of this expression. */
4740 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4742 if (! occr
->deleted_p
)
4745 /* Insert this expression on this edge if if it would
4746 reach the deleted occurrence in BB. */
4747 if (!TEST_BIT (inserted
[e
], j
))
4750 edge eg
= INDEX_EDGE (edge_list
, e
);
4752 /* We can't insert anything on an abnormal and
4753 critical edge, so we insert the insn at the end of
4754 the previous block. There are several alternatives
4755 detailed in Morgans book P277 (sec 10.5) for
4756 handling this situation. This one is easiest for
4759 if ((eg
->flags
& EDGE_ABNORMAL
) == EDGE_ABNORMAL
)
4760 insert_insn_end_bb (index_map
[j
], bb
, 0);
4763 insn
= process_insert_insn (index_map
[j
]);
4764 insert_insn_on_edge (insn
, eg
);
4769 fprintf (gcse_file
, "PRE/HOIST: edge (%d,%d), ",
4771 INDEX_EDGE_SUCC_BB (edge_list
, e
)->index
);
4772 fprintf (gcse_file
, "copy expression %d\n",
4773 expr
->bitmap_index
);
4776 update_ld_motion_stores (expr
);
4777 SET_BIT (inserted
[e
], j
);
4779 gcse_create_count
++;
4786 sbitmap_vector_free (inserted
);
4790 /* Copy the result of INSN to REG. INDX is the expression number. */
4793 pre_insert_copy_insn (expr
, insn
)
4797 rtx reg
= expr
->reaching_reg
;
4798 int regno
= REGNO (reg
);
4799 int indx
= expr
->bitmap_index
;
4800 rtx set
= single_set (insn
);
4806 new_insn
= emit_insn_after (gen_move_insn (reg
, SET_DEST (set
)), insn
);
4808 /* Keep register set table up to date. */
4809 record_one_set (regno
, new_insn
);
4811 gcse_create_count
++;
4815 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4816 BLOCK_NUM (insn
), INSN_UID (new_insn
), indx
,
4817 INSN_UID (insn
), regno
);
4818 update_ld_motion_stores (expr
);
4821 /* Copy available expressions that reach the redundant expression
4822 to `reaching_reg'. */
4825 pre_insert_copies ()
4832 /* For each available expression in the table, copy the result to
4833 `reaching_reg' if the expression reaches a deleted one.
4835 ??? The current algorithm is rather brute force.
4836 Need to do some profiling. */
4838 for (i
= 0; i
< expr_hash_table_size
; i
++)
4839 for (expr
= expr_hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4841 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4842 we don't want to insert a copy here because the expression may not
4843 really be redundant. So only insert an insn if the expression was
4844 deleted. This test also avoids further processing if the
4845 expression wasn't deleted anywhere. */
4846 if (expr
->reaching_reg
== NULL
)
4849 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4851 if (! occr
->deleted_p
)
4854 for (avail
= expr
->avail_occr
; avail
!= NULL
; avail
= avail
->next
)
4856 rtx insn
= avail
->insn
;
4858 /* No need to handle this one if handled already. */
4859 if (avail
->copied_p
)
4862 /* Don't handle this one if it's a redundant one. */
4863 if (TEST_BIT (pre_redundant_insns
, INSN_CUID (insn
)))
4866 /* Or if the expression doesn't reach the deleted one. */
4867 if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail
->insn
),
4869 BLOCK_FOR_INSN (occr
->insn
)))
4872 /* Copy the result of avail to reaching_reg. */
4873 pre_insert_copy_insn (expr
, insn
);
4874 avail
->copied_p
= 1;
4880 /* Delete redundant computations.
4881 Deletion is done by changing the insn to copy the `reaching_reg' of
4882 the expression into the result of the SET. It is left to later passes
4883 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4885 Returns non-zero if a change is made. */
4896 for (i
= 0; i
< expr_hash_table_size
; i
++)
4897 for (expr
= expr_hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4899 int indx
= expr
->bitmap_index
;
4901 /* We only need to search antic_occr since we require
4904 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4906 rtx insn
= occr
->insn
;
4908 basic_block bb
= BLOCK_FOR_INSN (insn
);
4910 if (TEST_BIT (pre_delete_map
[bb
->index
], indx
))
4912 set
= single_set (insn
);
4916 /* Create a pseudo-reg to store the result of reaching
4917 expressions into. Get the mode for the new pseudo from
4918 the mode of the original destination pseudo. */
4919 if (expr
->reaching_reg
== NULL
)
4921 = gen_reg_rtx (GET_MODE (SET_DEST (set
)));
4923 /* In theory this should never fail since we're creating
4926 However, on the x86 some of the movXX patterns actually
4927 contain clobbers of scratch regs. This may cause the
4928 insn created by validate_change to not match any pattern
4929 and thus cause validate_change to fail. */
4930 if (validate_change (insn
, &SET_SRC (set
),
4931 expr
->reaching_reg
, 0))
4933 occr
->deleted_p
= 1;
4934 SET_BIT (pre_redundant_insns
, INSN_CUID (insn
));
4942 "PRE: redundant insn %d (expression %d) in ",
4943 INSN_UID (insn
), indx
);
4944 fprintf (gcse_file
, "bb %d, reaching reg is %d\n",
4945 bb
->index
, REGNO (expr
->reaching_reg
));
4954 /* Perform GCSE optimizations using PRE.
4955 This is called by one_pre_gcse_pass after all the dataflow analysis
4958 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4959 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4960 Compiler Design and Implementation.
4962 ??? A new pseudo reg is created to hold the reaching expression. The nice
4963 thing about the classical approach is that it would try to use an existing
4964 reg. If the register can't be adequately optimized [i.e. we introduce
4965 reload problems], one could add a pass here to propagate the new register
4968 ??? We don't handle single sets in PARALLELs because we're [currently] not
4969 able to copy the rest of the parallel when we insert copies to create full
4970 redundancies from partial redundancies. However, there's no reason why we
4971 can't handle PARALLELs in the cases where there are no partial
4978 int did_insert
, changed
;
4979 struct expr
**index_map
;
4982 /* Compute a mapping from expression number (`bitmap_index') to
4983 hash table entry. */
4985 index_map
= (struct expr
**) xcalloc (n_exprs
, sizeof (struct expr
*));
4986 for (i
= 0; i
< expr_hash_table_size
; i
++)
4987 for (expr
= expr_hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4988 index_map
[expr
->bitmap_index
] = expr
;
4990 /* Reset bitmap used to track which insns are redundant. */
4991 pre_redundant_insns
= sbitmap_alloc (max_cuid
);
4992 sbitmap_zero (pre_redundant_insns
);
4994 /* Delete the redundant insns first so that
4995 - we know what register to use for the new insns and for the other
4996 ones with reaching expressions
4997 - we know which insns are redundant when we go to create copies */
4999 changed
= pre_delete ();
5001 did_insert
= pre_edge_insert (edge_list
, index_map
);
5003 /* In other places with reaching expressions, copy the expression to the
5004 specially allocated pseudo-reg that reaches the redundant expr. */
5005 pre_insert_copies ();
5008 commit_edge_insertions ();
5013 free (pre_redundant_insns
);
5017 /* Top level routine to perform one PRE GCSE pass.
5019 Return non-zero if a change was made. */
5022 one_pre_gcse_pass (pass
)
5027 gcse_subst_count
= 0;
5028 gcse_create_count
= 0;
5030 alloc_expr_hash_table (max_cuid
);
5031 add_noreturn_fake_exit_edges ();
5033 compute_ld_motion_mems ();
5035 compute_expr_hash_table ();
5036 trim_ld_motion_mems ();
5038 dump_hash_table (gcse_file
, "Expression", expr_hash_table
,
5039 expr_hash_table_size
, n_exprs
);
5043 alloc_pre_mem (n_basic_blocks
, n_exprs
);
5044 compute_pre_data ();
5045 changed
|= pre_gcse ();
5046 free_edge_list (edge_list
);
5051 remove_fake_edges ();
5052 free_expr_hash_table ();
5056 fprintf (gcse_file
, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
5057 current_function_name
, pass
, bytes_used
);
5058 fprintf (gcse_file
, "%d substs, %d insns created\n",
5059 gcse_subst_count
, gcse_create_count
);
5065 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to INSN.
5066 If notes are added to an insn which references a CODE_LABEL, the
5067 LABEL_NUSES count is incremented. We have to add REG_LABEL notes,
5068 because the following loop optimization pass requires them. */
5070 /* ??? This is very similar to the loop.c add_label_notes function. We
5071 could probably share code here. */
5073 /* ??? If there was a jump optimization pass after gcse and before loop,
5074 then we would not need to do this here, because jump would add the
5075 necessary REG_LABEL notes. */
5078 add_label_notes (x
, insn
)
5082 enum rtx_code code
= GET_CODE (x
);
5086 if (code
== LABEL_REF
&& !LABEL_REF_NONLOCAL_P (x
))
5088 /* This code used to ignore labels that referred to dispatch tables to
5089 avoid flow generating (slighly) worse code.
5091 We no longer ignore such label references (see LABEL_REF handling in
5092 mark_jump_label for additional information). */
5094 REG_NOTES (insn
) = gen_rtx_INSN_LIST (REG_LABEL
, XEXP (x
, 0),
5096 if (LABEL_P (XEXP (x
, 0)))
5097 LABEL_NUSES (XEXP (x
, 0))++;
5101 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
5104 add_label_notes (XEXP (x
, i
), insn
);
5105 else if (fmt
[i
] == 'E')
5106 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
5107 add_label_notes (XVECEXP (x
, i
, j
), insn
);
5111 /* Compute transparent outgoing information for each block.
5113 An expression is transparent to an edge unless it is killed by
5114 the edge itself. This can only happen with abnormal control flow,
5115 when the edge is traversed through a call. This happens with
5116 non-local labels and exceptions.
5118 This would not be necessary if we split the edge. While this is
5119 normally impossible for abnormal critical edges, with some effort
5120 it should be possible with exception handling, since we still have
5121 control over which handler should be invoked. But due to increased
5122 EH table sizes, this may not be worthwhile. */
5125 compute_transpout ()
5131 sbitmap_vector_ones (transpout
, n_basic_blocks
);
5133 for (bb
= 0; bb
< n_basic_blocks
; ++bb
)
5135 /* Note that flow inserted a nop a the end of basic blocks that
5136 end in call instructions for reasons other than abnormal
5138 if (GET_CODE (BLOCK_END (bb
)) != CALL_INSN
)
5141 for (i
= 0; i
< expr_hash_table_size
; i
++)
5142 for (expr
= expr_hash_table
[i
]; expr
; expr
= expr
->next_same_hash
)
5143 if (GET_CODE (expr
->expr
) == MEM
)
5145 if (GET_CODE (XEXP (expr
->expr
, 0)) == SYMBOL_REF
5146 && CONSTANT_POOL_ADDRESS_P (XEXP (expr
->expr
, 0)))
5149 /* ??? Optimally, we would use interprocedural alias
5150 analysis to determine if this mem is actually killed
5152 RESET_BIT (transpout
[bb
], expr
->bitmap_index
);
5157 /* Removal of useless null pointer checks */
5159 /* Called via note_stores. X is set by SETTER. If X is a register we must
5160 invalidate nonnull_local and set nonnull_killed. DATA is really a
5161 `null_pointer_info *'.
5163 We ignore hard registers. */
5166 invalidate_nonnull_info (x
, setter
, data
)
5168 rtx setter ATTRIBUTE_UNUSED
;
5172 struct null_pointer_info
*npi
= (struct null_pointer_info
*) data
;
5174 while (GET_CODE (x
) == SUBREG
)
5177 /* Ignore anything that is not a register or is a hard register. */
5178 if (GET_CODE (x
) != REG
5179 || REGNO (x
) < npi
->min_reg
5180 || REGNO (x
) >= npi
->max_reg
)
5183 regno
= REGNO (x
) - npi
->min_reg
;
5185 RESET_BIT (npi
->nonnull_local
[npi
->current_block
], regno
);
5186 SET_BIT (npi
->nonnull_killed
[npi
->current_block
], regno
);
5189 /* Do null-pointer check elimination for the registers indicated in
5190 NPI. NONNULL_AVIN and NONNULL_AVOUT are pre-allocated sbitmaps;
5191 they are not our responsibility to free. */
5194 delete_null_pointer_checks_1 (delete_list
, block_reg
, nonnull_avin
,
5196 varray_type
*delete_list
;
5197 unsigned int *block_reg
;
5198 sbitmap
*nonnull_avin
;
5199 sbitmap
*nonnull_avout
;
5200 struct null_pointer_info
*npi
;
5204 sbitmap
*nonnull_local
= npi
->nonnull_local
;
5205 sbitmap
*nonnull_killed
= npi
->nonnull_killed
;
5207 /* Compute local properties, nonnull and killed. A register will have
5208 the nonnull property if at the end of the current block its value is
5209 known to be nonnull. The killed property indicates that somewhere in
5210 the block any information we had about the register is killed.
5212 Note that a register can have both properties in a single block. That
5213 indicates that it's killed, then later in the block a new value is
5215 sbitmap_vector_zero (nonnull_local
, n_basic_blocks
);
5216 sbitmap_vector_zero (nonnull_killed
, n_basic_blocks
);
5218 for (current_block
= 0; current_block
< n_basic_blocks
; current_block
++)
5220 rtx insn
, stop_insn
;
5222 /* Set the current block for invalidate_nonnull_info. */
5223 npi
->current_block
= current_block
;
5225 /* Scan each insn in the basic block looking for memory references and
5227 stop_insn
= NEXT_INSN (BLOCK_END (current_block
));
5228 for (insn
= BLOCK_HEAD (current_block
);
5230 insn
= NEXT_INSN (insn
))
5235 /* Ignore anything that is not a normal insn. */
5236 if (! INSN_P (insn
))
5239 /* Basically ignore anything that is not a simple SET. We do have
5240 to make sure to invalidate nonnull_local and set nonnull_killed
5241 for such insns though. */
5242 set
= single_set (insn
);
5245 note_stores (PATTERN (insn
), invalidate_nonnull_info
, npi
);
5249 /* See if we've got a usable memory load. We handle it first
5250 in case it uses its address register as a dest (which kills
5251 the nonnull property). */
5252 if (GET_CODE (SET_SRC (set
)) == MEM
5253 && GET_CODE ((reg
= XEXP (SET_SRC (set
), 0))) == REG
5254 && REGNO (reg
) >= npi
->min_reg
5255 && REGNO (reg
) < npi
->max_reg
)
5256 SET_BIT (nonnull_local
[current_block
],
5257 REGNO (reg
) - npi
->min_reg
);
5259 /* Now invalidate stuff clobbered by this insn. */
5260 note_stores (PATTERN (insn
), invalidate_nonnull_info
, npi
);
5262 /* And handle stores, we do these last since any sets in INSN can
5263 not kill the nonnull property if it is derived from a MEM
5264 appearing in a SET_DEST. */
5265 if (GET_CODE (SET_DEST (set
)) == MEM
5266 && GET_CODE ((reg
= XEXP (SET_DEST (set
), 0))) == REG
5267 && REGNO (reg
) >= npi
->min_reg
5268 && REGNO (reg
) < npi
->max_reg
)
5269 SET_BIT (nonnull_local
[current_block
],
5270 REGNO (reg
) - npi
->min_reg
);
5274 /* Now compute global properties based on the local properties. This
5275 is a classic global availablity algorithm. */
5276 compute_available (nonnull_local
, nonnull_killed
,
5277 nonnull_avout
, nonnull_avin
);
5279 /* Now look at each bb and see if it ends with a compare of a value
5281 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
5283 rtx last_insn
= BLOCK_END (bb
);
5284 rtx condition
, earliest
;
5285 int compare_and_branch
;
5287 /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
5288 since BLOCK_REG[BB] is zero if this block did not end with a
5289 comparison against zero, this condition works. */
5290 if (block_reg
[bb
] < npi
->min_reg
5291 || block_reg
[bb
] >= npi
->max_reg
)
5294 /* LAST_INSN is a conditional jump. Get its condition. */
5295 condition
= get_condition (last_insn
, &earliest
);
5297 /* If we can't determine the condition then skip. */
5301 /* Is the register known to have a nonzero value? */
5302 if (!TEST_BIT (nonnull_avout
[bb
], block_reg
[bb
] - npi
->min_reg
))
5305 /* Try to compute whether the compare/branch at the loop end is one or
5306 two instructions. */
5307 if (earliest
== last_insn
)
5308 compare_and_branch
= 1;
5309 else if (earliest
== prev_nonnote_insn (last_insn
))
5310 compare_and_branch
= 2;
5314 /* We know the register in this comparison is nonnull at exit from
5315 this block. We can optimize this comparison. */
5316 if (GET_CODE (condition
) == NE
)
5320 new_jump
= emit_jump_insn_before (gen_jump (JUMP_LABEL (last_insn
)),
5322 JUMP_LABEL (new_jump
) = JUMP_LABEL (last_insn
);
5323 LABEL_NUSES (JUMP_LABEL (new_jump
))++;
5324 emit_barrier_after (new_jump
);
5327 VARRAY_RTX_INIT (*delete_list
, 10, "delete_list");
5329 VARRAY_PUSH_RTX (*delete_list
, last_insn
);
5330 if (compare_and_branch
== 2)
5331 VARRAY_PUSH_RTX (*delete_list
, earliest
);
5333 /* Don't check this block again. (Note that BLOCK_END is
5334 invalid here; we deleted the last instruction in the
5340 /* Find EQ/NE comparisons against zero which can be (indirectly) evaluated
5343 This is conceptually similar to global constant/copy propagation and
5344 classic global CSE (it even uses the same dataflow equations as cprop).
5346 If a register is used as memory address with the form (mem (reg)), then we
5347 know that REG can not be zero at that point in the program. Any instruction
5348 which sets REG "kills" this property.
5350 So, if every path leading to a conditional branch has an available memory
5351 reference of that form, then we know the register can not have the value
5352 zero at the conditional branch.
5354 So we merely need to compute the local properies and propagate that data
5355 around the cfg, then optimize where possible.
5357 We run this pass two times. Once before CSE, then again after CSE. This
5358 has proven to be the most profitable approach. It is rare for new
5359 optimization opportunities of this nature to appear after the first CSE
5362 This could probably be integrated with global cprop with a little work. */
5365 delete_null_pointer_checks (f
)
5366 rtx f ATTRIBUTE_UNUSED
;
5368 sbitmap
*nonnull_avin
, *nonnull_avout
;
5369 unsigned int *block_reg
;
5370 varray_type delete_list
= NULL
;
5376 struct null_pointer_info npi
;
5378 /* If we have only a single block, then there's nothing to do. */
5379 if (n_basic_blocks
<= 1)
5382 /* Trying to perform global optimizations on flow graphs which have
5383 a high connectivity will take a long time and is unlikely to be
5384 particularly useful.
5386 In normal circumstances a cfg should have about twice as many edges
5387 as blocks. But we do not want to punish small functions which have
5388 a couple switch statements. So we require a relatively large number
5389 of basic blocks and the ratio of edges to blocks to be high. */
5390 if (n_basic_blocks
> 1000 && n_edges
/ n_basic_blocks
>= 20)
5393 /* We need four bitmaps, each with a bit for each register in each
5395 max_reg
= max_reg_num ();
5396 regs_per_pass
= get_bitmap_width (4, n_basic_blocks
, max_reg
);
5398 /* Allocate bitmaps to hold local and global properties. */
5399 npi
.nonnull_local
= sbitmap_vector_alloc (n_basic_blocks
, regs_per_pass
);
5400 npi
.nonnull_killed
= sbitmap_vector_alloc (n_basic_blocks
, regs_per_pass
);
5401 nonnull_avin
= sbitmap_vector_alloc (n_basic_blocks
, regs_per_pass
);
5402 nonnull_avout
= sbitmap_vector_alloc (n_basic_blocks
, regs_per_pass
);
5404 /* Go through the basic blocks, seeing whether or not each block
5405 ends with a conditional branch whose condition is a comparison
5406 against zero. Record the register compared in BLOCK_REG. */
5407 block_reg
= (unsigned int *) xcalloc (n_basic_blocks
, sizeof (int));
5408 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
5410 rtx last_insn
= BLOCK_END (bb
);
5411 rtx condition
, earliest
, reg
;
5413 /* We only want conditional branches. */
5414 if (GET_CODE (last_insn
) != JUMP_INSN
5415 || !any_condjump_p (last_insn
)
5416 || !onlyjump_p (last_insn
))
5419 /* LAST_INSN is a conditional jump. Get its condition. */
5420 condition
= get_condition (last_insn
, &earliest
);
5422 /* If we were unable to get the condition, or it is not an equality
5423 comparison against zero then there's nothing we can do. */
5425 || (GET_CODE (condition
) != NE
&& GET_CODE (condition
) != EQ
)
5426 || GET_CODE (XEXP (condition
, 1)) != CONST_INT
5427 || (XEXP (condition
, 1)
5428 != CONST0_RTX (GET_MODE (XEXP (condition
, 0)))))
5431 /* We must be checking a register against zero. */
5432 reg
= XEXP (condition
, 0);
5433 if (GET_CODE (reg
) != REG
)
5436 block_reg
[bb
] = REGNO (reg
);
5439 /* Go through the algorithm for each block of registers. */
5440 for (reg
= FIRST_PSEUDO_REGISTER
; reg
< max_reg
; reg
+= regs_per_pass
)
5443 npi
.max_reg
= MIN (reg
+ regs_per_pass
, max_reg
);
5444 delete_null_pointer_checks_1 (&delete_list
, block_reg
, nonnull_avin
,
5445 nonnull_avout
, &npi
);
5448 /* Now delete the instructions all at once. This breaks the CFG. */
5451 for (i
= 0; i
< VARRAY_ACTIVE_SIZE (delete_list
); i
++)
5452 delete_related_insns (VARRAY_RTX (delete_list
, i
));
5453 VARRAY_FREE (delete_list
);
5456 /* Free the table of registers compared at the end of every block. */
5460 sbitmap_vector_free (npi
.nonnull_local
);
5461 sbitmap_vector_free (npi
.nonnull_killed
);
5462 sbitmap_vector_free (nonnull_avin
);
5463 sbitmap_vector_free (nonnull_avout
);
5466 /* Code Hoisting variables and subroutines. */
5468 /* Very busy expressions. */
5469 static sbitmap
*hoist_vbein
;
5470 static sbitmap
*hoist_vbeout
;
5472 /* Hoistable expressions. */
5473 static sbitmap
*hoist_exprs
;
5475 /* Dominator bitmaps. */
5476 static sbitmap
*dominators
;
5478 /* ??? We could compute post dominators and run this algorithm in
5479 reverse to to perform tail merging, doing so would probably be
5480 more effective than the tail merging code in jump.c.
5482 It's unclear if tail merging could be run in parallel with
5483 code hoisting. It would be nice. */
5485 /* Allocate vars used for code hoisting analysis. */
5488 alloc_code_hoist_mem (n_blocks
, n_exprs
)
5489 int n_blocks
, n_exprs
;
5491 antloc
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5492 transp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5493 comp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5495 hoist_vbein
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5496 hoist_vbeout
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5497 hoist_exprs
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5498 transpout
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
5500 dominators
= sbitmap_vector_alloc (n_blocks
, n_blocks
);
5503 /* Free vars used for code hoisting analysis. */
5506 free_code_hoist_mem ()
5508 sbitmap_vector_free (antloc
);
5509 sbitmap_vector_free (transp
);
5510 sbitmap_vector_free (comp
);
5512 sbitmap_vector_free (hoist_vbein
);
5513 sbitmap_vector_free (hoist_vbeout
);
5514 sbitmap_vector_free (hoist_exprs
);
5515 sbitmap_vector_free (transpout
);
5517 sbitmap_vector_free (dominators
);
5520 /* Compute the very busy expressions at entry/exit from each block.
5522 An expression is very busy if all paths from a given point
5523 compute the expression. */
5526 compute_code_hoist_vbeinout ()
5528 int bb
, changed
, passes
;
5530 sbitmap_vector_zero (hoist_vbeout
, n_basic_blocks
);
5531 sbitmap_vector_zero (hoist_vbein
, n_basic_blocks
);
5540 /* We scan the blocks in the reverse order to speed up
5542 for (bb
= n_basic_blocks
- 1; bb
>= 0; bb
--)
5544 changed
|= sbitmap_a_or_b_and_c (hoist_vbein
[bb
], antloc
[bb
],
5545 hoist_vbeout
[bb
], transp
[bb
]);
5546 if (bb
!= n_basic_blocks
- 1)
5547 sbitmap_intersection_of_succs (hoist_vbeout
[bb
], hoist_vbein
, bb
);
5554 fprintf (gcse_file
, "hoisting vbeinout computation: %d passes\n", passes
);
5557 /* Top level routine to do the dataflow analysis needed by code hoisting. */
5560 compute_code_hoist_data ()
5562 compute_local_properties (transp
, comp
, antloc
, 0);
5563 compute_transpout ();
5564 compute_code_hoist_vbeinout ();
5565 calculate_dominance_info (NULL
, dominators
, CDI_DOMINATORS
);
5567 fprintf (gcse_file
, "\n");
5570 /* Determine if the expression identified by EXPR_INDEX would
5571 reach BB unimpared if it was placed at the end of EXPR_BB.
5573 It's unclear exactly what Muchnick meant by "unimpared". It seems
5574 to me that the expression must either be computed or transparent in
5575 *every* block in the path(s) from EXPR_BB to BB. Any other definition
5576 would allow the expression to be hoisted out of loops, even if
5577 the expression wasn't a loop invariant.
5579 Contrast this to reachability for PRE where an expression is
5580 considered reachable if *any* path reaches instead of *all*
5584 hoist_expr_reaches_here_p (expr_bb
, expr_index
, bb
, visited
)
5585 basic_block expr_bb
;
5591 int visited_allocated_locally
= 0;
5594 if (visited
== NULL
)
5596 visited_allocated_locally
= 1;
5597 visited
= xcalloc (n_basic_blocks
, 1);
5600 for (pred
= bb
->pred
; pred
!= NULL
; pred
= pred
->pred_next
)
5602 basic_block pred_bb
= pred
->src
;
5604 if (pred
->src
== ENTRY_BLOCK_PTR
)
5606 else if (visited
[pred_bb
->index
])
5609 /* Does this predecessor generate this expression? */
5610 else if (TEST_BIT (comp
[pred_bb
->index
], expr_index
))
5612 else if (! TEST_BIT (transp
[pred_bb
->index
], expr_index
))
5618 visited
[pred_bb
->index
] = 1;
5619 if (! hoist_expr_reaches_here_p (expr_bb
, expr_index
,
5624 if (visited_allocated_locally
)
5627 return (pred
== NULL
);
5630 /* Actually perform code hoisting. */
5637 struct expr
**index_map
;
5640 sbitmap_vector_zero (hoist_exprs
, n_basic_blocks
);
5642 /* Compute a mapping from expression number (`bitmap_index') to
5643 hash table entry. */
5645 index_map
= (struct expr
**) xcalloc (n_exprs
, sizeof (struct expr
*));
5646 for (i
= 0; i
< expr_hash_table_size
; i
++)
5647 for (expr
= expr_hash_table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
5648 index_map
[expr
->bitmap_index
] = expr
;
5650 /* Walk over each basic block looking for potentially hoistable
5651 expressions, nothing gets hoisted from the entry block. */
5652 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
5655 int insn_inserted_p
;
5657 /* Examine each expression that is very busy at the exit of this
5658 block. These are the potentially hoistable expressions. */
5659 for (i
= 0; i
< hoist_vbeout
[bb
]->n_bits
; i
++)
5663 if (TEST_BIT (hoist_vbeout
[bb
], i
) && TEST_BIT (transpout
[bb
], i
))
5665 /* We've found a potentially hoistable expression, now
5666 we look at every block BB dominates to see if it
5667 computes the expression. */
5668 for (dominated
= 0; dominated
< n_basic_blocks
; dominated
++)
5670 /* Ignore self dominance. */
5672 || ! TEST_BIT (dominators
[dominated
], bb
))
5675 /* We've found a dominated block, now see if it computes
5676 the busy expression and whether or not moving that
5677 expression to the "beginning" of that block is safe. */
5678 if (!TEST_BIT (antloc
[dominated
], i
))
5681 /* Note if the expression would reach the dominated block
5682 unimpared if it was placed at the end of BB.
5684 Keep track of how many times this expression is hoistable
5685 from a dominated block into BB. */
5686 if (hoist_expr_reaches_here_p (BASIC_BLOCK (bb
), i
,
5687 BASIC_BLOCK (dominated
), NULL
))
5691 /* If we found more than one hoistable occurrence of this
5692 expression, then note it in the bitmap of expressions to
5693 hoist. It makes no sense to hoist things which are computed
5694 in only one BB, and doing so tends to pessimize register
5695 allocation. One could increase this value to try harder
5696 to avoid any possible code expansion due to register
5697 allocation issues; however experiments have shown that
5698 the vast majority of hoistable expressions are only movable
5699 from two successors, so raising this threshhold is likely
5700 to nullify any benefit we get from code hoisting. */
5703 SET_BIT (hoist_exprs
[bb
], i
);
5709 /* If we found nothing to hoist, then quit now. */
5713 /* Loop over all the hoistable expressions. */
5714 for (i
= 0; i
< hoist_exprs
[bb
]->n_bits
; i
++)
5716 /* We want to insert the expression into BB only once, so
5717 note when we've inserted it. */
5718 insn_inserted_p
= 0;
5720 /* These tests should be the same as the tests above. */
5721 if (TEST_BIT (hoist_vbeout
[bb
], i
))
5723 /* We've found a potentially hoistable expression, now
5724 we look at every block BB dominates to see if it
5725 computes the expression. */
5726 for (dominated
= 0; dominated
< n_basic_blocks
; dominated
++)
5728 /* Ignore self dominance. */
5730 || ! TEST_BIT (dominators
[dominated
], bb
))
5733 /* We've found a dominated block, now see if it computes
5734 the busy expression and whether or not moving that
5735 expression to the "beginning" of that block is safe. */
5736 if (!TEST_BIT (antloc
[dominated
], i
))
5739 /* The expression is computed in the dominated block and
5740 it would be safe to compute it at the start of the
5741 dominated block. Now we have to determine if the
5742 expression would reach the dominated block if it was
5743 placed at the end of BB. */
5744 if (hoist_expr_reaches_here_p (BASIC_BLOCK (bb
), i
,
5745 BASIC_BLOCK (dominated
), NULL
))
5747 struct expr
*expr
= index_map
[i
];
5748 struct occr
*occr
= expr
->antic_occr
;
5752 /* Find the right occurrence of this expression. */
5753 while (BLOCK_NUM (occr
->insn
) != dominated
&& occr
)
5756 /* Should never happen. */
5762 set
= single_set (insn
);
5766 /* Create a pseudo-reg to store the result of reaching
5767 expressions into. Get the mode for the new pseudo
5768 from the mode of the original destination pseudo. */
5769 if (expr
->reaching_reg
== NULL
)
5771 = gen_reg_rtx (GET_MODE (SET_DEST (set
)));
5773 /* In theory this should never fail since we're creating
5776 However, on the x86 some of the movXX patterns
5777 actually contain clobbers of scratch regs. This may
5778 cause the insn created by validate_change to not
5779 match any pattern and thus cause validate_change to
5781 if (validate_change (insn
, &SET_SRC (set
),
5782 expr
->reaching_reg
, 0))
5784 occr
->deleted_p
= 1;
5785 if (!insn_inserted_p
)
5787 insert_insn_end_bb (index_map
[i
],
5788 BASIC_BLOCK (bb
), 0);
5789 insn_inserted_p
= 1;
5801 /* Top level routine to perform one code hoisting (aka unification) pass
5803 Return non-zero if a change was made. */
5806 one_code_hoisting_pass ()
5810 alloc_expr_hash_table (max_cuid
);
5811 compute_expr_hash_table ();
5813 dump_hash_table (gcse_file
, "Code Hosting Expressions", expr_hash_table
,
5814 expr_hash_table_size
, n_exprs
);
5818 alloc_code_hoist_mem (n_basic_blocks
, n_exprs
);
5819 compute_code_hoist_data ();
5821 free_code_hoist_mem ();
5824 free_expr_hash_table ();
5829 /* Here we provide the things required to do store motion towards
5830 the exit. In order for this to be effective, gcse also needed to
5831 be taught how to move a load when it is kill only by a store to itself.
5836 void foo(float scale)
5838 for (i=0; i<10; i++)
5842 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
5843 the load out since its live around the loop, and stored at the bottom
5846 The 'Load Motion' referred to and implemented in this file is
5847 an enhancement to gcse which when using edge based lcm, recognizes
5848 this situation and allows gcse to move the load out of the loop.
5850 Once gcse has hoisted the load, store motion can then push this
5851 load towards the exit, and we end up with no loads or stores of 'i'
5854 /* This will search the ldst list for a matching expression. If it
5855 doesn't find one, we create one and initialize it. */
5857 static struct ls_expr
*
5861 struct ls_expr
* ptr
;
5863 for (ptr
= first_ls_expr(); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
5864 if (expr_equiv_p (ptr
->pattern
, x
))
5869 ptr
= (struct ls_expr
*) xmalloc (sizeof (struct ls_expr
));
5871 ptr
->next
= pre_ldst_mems
;
5874 ptr
->loads
= NULL_RTX
;
5875 ptr
->stores
= NULL_RTX
;
5876 ptr
->reaching_reg
= NULL_RTX
;
5879 ptr
->hash_index
= 0;
5880 pre_ldst_mems
= ptr
;
5886 /* Free up an individual ldst entry. */
5889 free_ldst_entry (ptr
)
5890 struct ls_expr
* ptr
;
5892 free_INSN_LIST_list (& ptr
->loads
);
5893 free_INSN_LIST_list (& ptr
->stores
);
5898 /* Free up all memory associated with the ldst list. */
5903 while (pre_ldst_mems
)
5905 struct ls_expr
* tmp
= pre_ldst_mems
;
5907 pre_ldst_mems
= pre_ldst_mems
->next
;
5909 free_ldst_entry (tmp
);
5912 pre_ldst_mems
= NULL
;
5915 /* Dump debugging info about the ldst list. */
5918 print_ldst_list (file
)
5921 struct ls_expr
* ptr
;
5923 fprintf (file
, "LDST list: \n");
5925 for (ptr
= first_ls_expr(); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
5927 fprintf (file
, " Pattern (%3d): ", ptr
->index
);
5929 print_rtl (file
, ptr
->pattern
);
5931 fprintf (file
, "\n Loads : ");
5934 print_rtl (file
, ptr
->loads
);
5936 fprintf (file
, "(nil)");
5938 fprintf (file
, "\n Stores : ");
5941 print_rtl (file
, ptr
->stores
);
5943 fprintf (file
, "(nil)");
5945 fprintf (file
, "\n\n");
5948 fprintf (file
, "\n");
5951 /* Returns 1 if X is in the list of ldst only expressions. */
5953 static struct ls_expr
*
5954 find_rtx_in_ldst (x
)
5957 struct ls_expr
* ptr
;
5959 for (ptr
= pre_ldst_mems
; ptr
!= NULL
; ptr
= ptr
->next
)
5960 if (expr_equiv_p (ptr
->pattern
, x
) && ! ptr
->invalid
)
5966 /* Assign each element of the list of mems a monotonically increasing value. */
5971 struct ls_expr
* ptr
;
5974 for (ptr
= pre_ldst_mems
; ptr
!= NULL
; ptr
= ptr
->next
)
5980 /* Return first item in the list. */
5982 static inline struct ls_expr
*
5985 return pre_ldst_mems
;
5988 /* Return the next item in ther list after the specified one. */
5990 static inline struct ls_expr
*
5992 struct ls_expr
* ptr
;
5997 /* Load Motion for loads which only kill themselves. */
5999 /* Return true if x is a simple MEM operation, with no registers or
6000 side effects. These are the types of loads we consider for the
6001 ld_motion list, otherwise we let the usual aliasing take care of it. */
6007 if (GET_CODE (x
) != MEM
)
6010 if (MEM_VOLATILE_P (x
))
6013 if (GET_MODE (x
) == BLKmode
)
6016 if (!rtx_varies_p (XEXP (x
, 0), 0))
6022 /* Make sure there isn't a buried reference in this pattern anywhere.
6023 If there is, invalidate the entry for it since we're not capable
6024 of fixing it up just yet.. We have to be sure we know about ALL
6025 loads since the aliasing code will allow all entries in the
6026 ld_motion list to not-alias itself. If we miss a load, we will get
6027 the wrong value since gcse might common it and we won't know to
6031 invalidate_any_buried_refs (x
)
6036 struct ls_expr
* ptr
;
6038 /* Invalidate it in the list. */
6039 if (GET_CODE (x
) == MEM
&& simple_mem (x
))
6041 ptr
= ldst_entry (x
);
6045 /* Recursively process the insn. */
6046 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
6048 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
6051 invalidate_any_buried_refs (XEXP (x
, i
));
6052 else if (fmt
[i
] == 'E')
6053 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
6054 invalidate_any_buried_refs (XVECEXP (x
, i
, j
));
6058 /* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
6059 being defined as MEM loads and stores to symbols, with no
6060 side effects and no registers in the expression. If there are any
6061 uses/defs which don't match this criteria, it is invalidated and
6062 trimmed out later. */
6065 compute_ld_motion_mems ()
6067 struct ls_expr
* ptr
;
6071 pre_ldst_mems
= NULL
;
6073 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
6075 for (insn
= BLOCK_HEAD (bb
);
6076 insn
&& insn
!= NEXT_INSN (BLOCK_END (bb
));
6077 insn
= NEXT_INSN (insn
))
6079 if (GET_RTX_CLASS (GET_CODE (insn
)) == 'i')
6081 if (GET_CODE (PATTERN (insn
)) == SET
)
6083 rtx src
= SET_SRC (PATTERN (insn
));
6084 rtx dest
= SET_DEST (PATTERN (insn
));
6086 /* Check for a simple LOAD... */
6087 if (GET_CODE (src
) == MEM
&& simple_mem (src
))
6089 ptr
= ldst_entry (src
);
6090 if (GET_CODE (dest
) == REG
)
6091 ptr
->loads
= alloc_INSN_LIST (insn
, ptr
->loads
);
6097 /* Make sure there isn't a buried load somewhere. */
6098 invalidate_any_buried_refs (src
);
6101 /* Check for stores. Don't worry about aliased ones, they
6102 will block any movement we might do later. We only care
6103 about this exact pattern since those are the only
6104 circumstance that we will ignore the aliasing info. */
6105 if (GET_CODE (dest
) == MEM
&& simple_mem (dest
))
6107 ptr
= ldst_entry (dest
);
6109 if (GET_CODE (src
) != MEM
6110 && GET_CODE (src
) != ASM_OPERANDS
)
6111 ptr
->stores
= alloc_INSN_LIST (insn
, ptr
->stores
);
6117 invalidate_any_buried_refs (PATTERN (insn
));
6123 /* Remove any references that have been either invalidated or are not in the
6124 expression list for pre gcse. */
6127 trim_ld_motion_mems ()
6129 struct ls_expr
* last
= NULL
;
6130 struct ls_expr
* ptr
= first_ls_expr ();
6134 int del
= ptr
->invalid
;
6135 struct expr
* expr
= NULL
;
6137 /* Delete if entry has been made invalid. */
6143 /* Delete if we cannot find this mem in the expression list. */
6144 for (i
= 0; i
< expr_hash_table_size
&& del
; i
++)
6146 for (expr
= expr_hash_table
[i
];
6148 expr
= expr
->next_same_hash
)
6149 if (expr_equiv_p (expr
->expr
, ptr
->pattern
))
6161 last
->next
= ptr
->next
;
6162 free_ldst_entry (ptr
);
6167 pre_ldst_mems
= pre_ldst_mems
->next
;
6168 free_ldst_entry (ptr
);
6169 ptr
= pre_ldst_mems
;
6174 /* Set the expression field if we are keeping it. */
6181 /* Show the world what we've found. */
6182 if (gcse_file
&& pre_ldst_mems
!= NULL
)
6183 print_ldst_list (gcse_file
);
6186 /* This routine will take an expression which we are replacing with
6187 a reaching register, and update any stores that are needed if
6188 that expression is in the ld_motion list. Stores are updated by
6189 copying their SRC to the reaching register, and then storeing
6190 the reaching register into the store location. These keeps the
6191 correct value in the reaching register for the loads. */
6194 update_ld_motion_stores (expr
)
6197 struct ls_expr
* mem_ptr
;
6199 if ((mem_ptr
= find_rtx_in_ldst (expr
->expr
)))
6201 /* We can try to find just the REACHED stores, but is shouldn't
6202 matter to set the reaching reg everywhere... some might be
6203 dead and should be eliminated later. */
6205 /* We replace SET mem = expr with
6207 SET mem = reg , where reg is the
6208 reaching reg used in the load. */
6209 rtx list
= mem_ptr
->stores
;
6211 for ( ; list
!= NULL_RTX
; list
= XEXP (list
, 1))
6213 rtx insn
= XEXP (list
, 0);
6214 rtx pat
= PATTERN (insn
);
6215 rtx src
= SET_SRC (pat
);
6216 rtx reg
= expr
->reaching_reg
;
6219 /* If we've already copied it, continue. */
6220 if (expr
->reaching_reg
== src
)
6225 fprintf (gcse_file
, "PRE: store updated with reaching reg ");
6226 print_rtl (gcse_file
, expr
->reaching_reg
);
6227 fprintf (gcse_file
, ":\n ");
6228 print_inline_rtx (gcse_file
, insn
, 8);
6229 fprintf (gcse_file
, "\n");
6232 copy
= gen_move_insn ( reg
, SET_SRC (pat
));
6233 new = emit_insn_before (copy
, insn
);
6234 record_one_set (REGNO (reg
), new);
6235 SET_SRC (pat
) = reg
;
6237 /* un-recognize this pattern since it's probably different now. */
6238 INSN_CODE (insn
) = -1;
6239 gcse_create_count
++;
6244 /* Store motion code. */
6246 /* This is used to communicate the target bitvector we want to use in the
6247 reg_set_info routine when called via the note_stores mechanism. */
6248 static sbitmap
* regvec
;
6250 /* Used in computing the reverse edge graph bit vectors. */
6251 static sbitmap
* st_antloc
;
6253 /* Global holding the number of store expressions we are dealing with. */
6254 static int num_stores
;
6256 /* Checks to set if we need to mark a register set. Called from note_stores. */
6259 reg_set_info (dest
, setter
, data
)
6260 rtx dest
, setter ATTRIBUTE_UNUSED
;
6261 void * data ATTRIBUTE_UNUSED
;
6263 if (GET_CODE (dest
) == SUBREG
)
6264 dest
= SUBREG_REG (dest
);
6266 if (GET_CODE (dest
) == REG
)
6267 SET_BIT (*regvec
, REGNO (dest
));
6270 /* Return non-zero if the register operands of expression X are killed
6271 anywhere in basic block BB. */
6274 store_ops_ok (x
, bb
)
6282 /* Repeat is used to turn tail-recursion into iteration. */
6288 code
= GET_CODE (x
);
6292 /* If a reg has changed after us in this
6293 block, the operand has been killed. */
6294 return TEST_BIT (reg_set_in_block
[bb
->index
], REGNO (x
));
6321 i
= GET_RTX_LENGTH (code
) - 1;
6322 fmt
= GET_RTX_FORMAT (code
);
6328 rtx tem
= XEXP (x
, i
);
6330 /* If we are about to do the last recursive call
6331 needed at this level, change it into iteration.
6332 This function is called enough to be worth it. */
6339 if (! store_ops_ok (tem
, bb
))
6342 else if (fmt
[i
] == 'E')
6346 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
6348 if (! store_ops_ok (XVECEXP (x
, i
, j
), bb
))
6357 /* Determine whether insn is MEM store pattern that we will consider moving. */
6360 find_moveable_store (insn
)
6363 struct ls_expr
* ptr
;
6364 rtx dest
= PATTERN (insn
);
6366 if (GET_CODE (dest
) != SET
6367 || GET_CODE (SET_SRC (dest
)) == ASM_OPERANDS
)
6370 dest
= SET_DEST (dest
);
6372 if (GET_CODE (dest
) != MEM
|| MEM_VOLATILE_P (dest
)
6373 || GET_MODE (dest
) == BLKmode
)
6376 if (GET_CODE (XEXP (dest
, 0)) != SYMBOL_REF
)
6379 if (rtx_varies_p (XEXP (dest
, 0), 0))
6382 ptr
= ldst_entry (dest
);
6383 ptr
->stores
= alloc_INSN_LIST (insn
, ptr
->stores
);
6386 /* Perform store motion. Much like gcse, except we move expressions the
6387 other way by looking at the flowgraph in reverse. */
6390 compute_store_table ()
6396 max_gcse_regno
= max_reg_num ();
6398 reg_set_in_block
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
,
6400 sbitmap_vector_zero (reg_set_in_block
, n_basic_blocks
);
6403 /* Find all the stores we care about. */
6404 for (bb
= 0; bb
< n_basic_blocks
; bb
++)
6406 regvec
= & (reg_set_in_block
[bb
]);
6407 for (insn
= BLOCK_END (bb
);
6408 insn
&& insn
!= PREV_INSN (BLOCK_HEAD (bb
));
6409 insn
= PREV_INSN (insn
))
6411 /* Ignore anything that is not a normal insn. */
6412 if (! INSN_P (insn
))
6415 if (GET_CODE (insn
) == CALL_INSN
)
6417 bool clobbers_all
= false;
6418 #ifdef NON_SAVING_SETJMP
6419 if (NON_SAVING_SETJMP
6420 && find_reg_note (insn
, REG_SETJMP
, NULL_RTX
))
6421 clobbers_all
= true;
6424 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
6426 || TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
6427 SET_BIT (reg_set_in_block
[bb
], regno
);
6430 pat
= PATTERN (insn
);
6431 note_stores (pat
, reg_set_info
, NULL
);
6433 /* Now that we've marked regs, look for stores. */
6434 if (GET_CODE (pat
) == SET
)
6435 find_moveable_store (insn
);
6439 ret
= enumerate_ldsts ();
6443 fprintf (gcse_file
, "Store Motion Expressions.\n");
6444 print_ldst_list (gcse_file
);
6450 /* Check to see if the load X is aliased with STORE_PATTERN. */
6453 load_kills_store (x
, store_pattern
)
6454 rtx x
, store_pattern
;
6456 if (true_dependence (x
, GET_MODE (x
), store_pattern
, rtx_addr_varies_p
))
6461 /* Go through the entire insn X, looking for any loads which might alias
6462 STORE_PATTERN. Return 1 if found. */
6465 find_loads (x
, store_pattern
)
6466 rtx x
, store_pattern
;
6475 if (GET_CODE (x
) == SET
)
6478 if (GET_CODE (x
) == MEM
)
6480 if (load_kills_store (x
, store_pattern
))
6484 /* Recursively process the insn. */
6485 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
6487 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0 && !ret
; i
--)
6490 ret
|= find_loads (XEXP (x
, i
), store_pattern
);
6491 else if (fmt
[i
] == 'E')
6492 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
6493 ret
|= find_loads (XVECEXP (x
, i
, j
), store_pattern
);
6498 /* Check if INSN kills the store pattern X (is aliased with it).
6499 Return 1 if it it does. */
6502 store_killed_in_insn (x
, insn
)
6505 if (GET_RTX_CLASS (GET_CODE (insn
)) != 'i')
6508 if (GET_CODE (insn
) == CALL_INSN
)
6510 if (CONST_OR_PURE_CALL_P (insn
))
6516 if (GET_CODE (PATTERN (insn
)) == SET
)
6518 rtx pat
= PATTERN (insn
);
6519 /* Check for memory stores to aliased objects. */
6520 if (GET_CODE (SET_DEST (pat
)) == MEM
&& !expr_equiv_p (SET_DEST (pat
), x
))
6521 /* pretend its a load and check for aliasing. */
6522 if (find_loads (SET_DEST (pat
), x
))
6524 return find_loads (SET_SRC (pat
), x
);
6527 return find_loads (PATTERN (insn
), x
);
6530 /* Returns 1 if the expression X is loaded or clobbered on or after INSN
6531 within basic block BB. */
6534 store_killed_after (x
, insn
, bb
)
6543 /* Check if the register operands of the store are OK in this block.
6544 Note that if registers are changed ANYWHERE in the block, we'll
6545 decide we can't move it, regardless of whether it changed above
6546 or below the store. This could be improved by checking the register
6547 operands while lookinng for aliasing in each insn. */
6548 if (!store_ops_ok (XEXP (x
, 0), bb
))
6551 for ( ; insn
&& insn
!= NEXT_INSN (last
); insn
= NEXT_INSN (insn
))
6552 if (store_killed_in_insn (x
, insn
))
6558 /* Returns 1 if the expression X is loaded or clobbered on or before INSN
6559 within basic block BB. */
6561 store_killed_before (x
, insn
, bb
)
6565 rtx first
= bb
->head
;
6568 return store_killed_in_insn (x
, insn
);
6570 /* Check if the register operands of the store are OK in this block.
6571 Note that if registers are changed ANYWHERE in the block, we'll
6572 decide we can't move it, regardless of whether it changed above
6573 or below the store. This could be improved by checking the register
6574 operands while lookinng for aliasing in each insn. */
6575 if (!store_ops_ok (XEXP (x
, 0), bb
))
6578 for ( ; insn
&& insn
!= PREV_INSN (first
); insn
= PREV_INSN (insn
))
6579 if (store_killed_in_insn (x
, insn
))
6585 #define ANTIC_STORE_LIST(x) ((x)->loads)
6586 #define AVAIL_STORE_LIST(x) ((x)->stores)
6588 /* Given the table of available store insns at the end of blocks,
6589 determine which ones are not killed by aliasing, and generate
6590 the appropriate vectors for gen and killed. */
6592 build_store_vectors ()
6597 struct ls_expr
* ptr
;
6599 /* Build the gen_vector. This is any store in the table which is not killed
6600 by aliasing later in its block. */
6601 ae_gen
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
, num_stores
);
6602 sbitmap_vector_zero (ae_gen
, n_basic_blocks
);
6604 st_antloc
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
, num_stores
);
6605 sbitmap_vector_zero (st_antloc
, n_basic_blocks
);
6607 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6609 /* Put all the stores into either the antic list, or the avail list,
6611 rtx store_list
= ptr
->stores
;
6612 ptr
->stores
= NULL_RTX
;
6614 for (st
= store_list
; st
!= NULL
; st
= XEXP (st
, 1))
6616 insn
= XEXP (st
, 0);
6617 bb
= BLOCK_FOR_INSN (insn
);
6619 if (!store_killed_after (ptr
->pattern
, insn
, bb
))
6621 /* If we've already seen an availale expression in this block,
6622 we can delete the one we saw already (It occurs earlier in
6623 the block), and replace it with this one). We'll copy the
6624 old SRC expression to an unused register in case there
6625 are any side effects. */
6626 if (TEST_BIT (ae_gen
[bb
->index
], ptr
->index
))
6628 /* Find previous store. */
6630 for (st
= AVAIL_STORE_LIST (ptr
); st
; st
= XEXP (st
, 1))
6631 if (BLOCK_FOR_INSN (XEXP (st
, 0)) == bb
)
6635 rtx r
= gen_reg_rtx (GET_MODE (ptr
->pattern
));
6637 fprintf(gcse_file
, "Removing redundant store:\n");
6638 replace_store_insn (r
, XEXP (st
, 0), bb
);
6639 XEXP (st
, 0) = insn
;
6643 SET_BIT (ae_gen
[bb
->index
], ptr
->index
);
6644 AVAIL_STORE_LIST (ptr
) = alloc_INSN_LIST (insn
,
6645 AVAIL_STORE_LIST (ptr
));
6648 if (!store_killed_before (ptr
->pattern
, insn
, bb
))
6650 SET_BIT (st_antloc
[BLOCK_NUM (insn
)], ptr
->index
);
6651 ANTIC_STORE_LIST (ptr
) = alloc_INSN_LIST (insn
,
6652 ANTIC_STORE_LIST (ptr
));
6656 /* Free the original list of store insns. */
6657 free_INSN_LIST_list (&store_list
);
6660 ae_kill
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
, num_stores
);
6661 sbitmap_vector_zero (ae_kill
, n_basic_blocks
);
6663 transp
= (sbitmap
*) sbitmap_vector_alloc (n_basic_blocks
, num_stores
);
6664 sbitmap_vector_zero (transp
, n_basic_blocks
);
6666 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6667 for (b
= 0; b
< n_basic_blocks
; b
++)
6669 if (store_killed_after (ptr
->pattern
, BLOCK_HEAD (b
), BASIC_BLOCK (b
)))
6671 /* The anticipatable expression is not killed if it's gen'd. */
6673 We leave this check out for now. If we have a code sequence
6674 in a block which looks like:
6678 We should flag this as having an ANTIC expression, NOT
6679 transparent, NOT killed, and AVAIL.
6680 Unfortunately, since we haven't re-written all loads to
6681 use the reaching reg, we'll end up doing an incorrect
6682 Load in the middle here if we push the store down. It happens in
6683 gcc.c-torture/execute/960311-1.c with -O3
6684 If we always kill it in this case, we'll sometimes do
6685 uneccessary work, but it shouldn't actually hurt anything.
6686 if (!TEST_BIT (ae_gen[b], ptr->index)). */
6687 SET_BIT (ae_kill
[b
], ptr
->index
);
6690 SET_BIT (transp
[b
], ptr
->index
);
6693 /* Any block with no exits calls some non-returning function, so
6694 we better mark the store killed here, or we might not store to
6695 it at all. If we knew it was abort, we wouldn't have to store,
6696 but we don't know that for sure. */
6699 fprintf (gcse_file
, "ST_avail and ST_antic (shown under loads..)\n");
6700 print_ldst_list (gcse_file
);
6701 dump_sbitmap_vector (gcse_file
, "st_antloc", "", st_antloc
, n_basic_blocks
);
6702 dump_sbitmap_vector (gcse_file
, "st_kill", "", ae_kill
, n_basic_blocks
);
6703 dump_sbitmap_vector (gcse_file
, "Transpt", "", transp
, n_basic_blocks
);
6704 dump_sbitmap_vector (gcse_file
, "st_avloc", "", ae_gen
, n_basic_blocks
);
6708 /* Insert an instruction at the begining of a basic block, and update
6709 the BLOCK_HEAD if needed. */
6712 insert_insn_start_bb (insn
, bb
)
6716 /* Insert at start of successor block. */
6717 rtx prev
= PREV_INSN (bb
->head
);
6718 rtx before
= bb
->head
;
6721 if (GET_CODE (before
) != CODE_LABEL
6722 && (GET_CODE (before
) != NOTE
6723 || NOTE_LINE_NUMBER (before
) != NOTE_INSN_BASIC_BLOCK
))
6726 if (prev
== bb
->end
)
6728 before
= NEXT_INSN (before
);
6731 insn
= emit_insn_after (insn
, prev
);
6735 fprintf (gcse_file
, "STORE_MOTION insert store at start of BB %d:\n",
6737 print_inline_rtx (gcse_file
, insn
, 6);
6738 fprintf (gcse_file
, "\n");
6742 /* This routine will insert a store on an edge. EXPR is the ldst entry for
6743 the memory reference, and E is the edge to insert it on. Returns non-zero
6744 if an edge insertion was performed. */
6747 insert_store (expr
, e
)
6748 struct ls_expr
* expr
;
6755 /* We did all the deleted before this insert, so if we didn't delete a
6756 store, then we haven't set the reaching reg yet either. */
6757 if (expr
->reaching_reg
== NULL_RTX
)
6760 reg
= expr
->reaching_reg
;
6761 insn
= gen_move_insn (expr
->pattern
, reg
);
6763 /* If we are inserting this expression on ALL predecessor edges of a BB,
6764 insert it at the start of the BB, and reset the insert bits on the other
6765 edges so we don't try to insert it on the other edges. */
6767 for (tmp
= e
->dest
->pred
; tmp
; tmp
= tmp
->pred_next
)
6769 int index
= EDGE_INDEX (edge_list
, tmp
->src
, tmp
->dest
);
6770 if (index
== EDGE_INDEX_NO_EDGE
)
6772 if (! TEST_BIT (pre_insert_map
[index
], expr
->index
))
6776 /* If tmp is NULL, we found an insertion on every edge, blank the
6777 insertion vector for these edges, and insert at the start of the BB. */
6778 if (!tmp
&& bb
!= EXIT_BLOCK_PTR
)
6780 for (tmp
= e
->dest
->pred
; tmp
; tmp
= tmp
->pred_next
)
6782 int index
= EDGE_INDEX (edge_list
, tmp
->src
, tmp
->dest
);
6783 RESET_BIT (pre_insert_map
[index
], expr
->index
);
6785 insert_insn_start_bb (insn
, bb
);
6789 /* We can't insert on this edge, so we'll insert at the head of the
6790 successors block. See Morgan, sec 10.5. */
6791 if ((e
->flags
& EDGE_ABNORMAL
) == EDGE_ABNORMAL
)
6793 insert_insn_start_bb (insn
, bb
);
6797 insert_insn_on_edge (insn
, e
);
6801 fprintf (gcse_file
, "STORE_MOTION insert insn on edge (%d, %d):\n",
6802 e
->src
->index
, e
->dest
->index
);
6803 print_inline_rtx (gcse_file
, insn
, 6);
6804 fprintf (gcse_file
, "\n");
6810 /* This routine will replace a store with a SET to a specified register. */
6813 replace_store_insn (reg
, del
, bb
)
6819 insn
= gen_move_insn (reg
, SET_SRC (PATTERN (del
)));
6820 insn
= emit_insn_after (insn
, del
);
6825 "STORE_MOTION delete insn in BB %d:\n ", bb
->index
);
6826 print_inline_rtx (gcse_file
, del
, 6);
6827 fprintf(gcse_file
, "\nSTORE MOTION replaced with insn:\n ");
6828 print_inline_rtx (gcse_file
, insn
, 6);
6829 fprintf(gcse_file
, "\n");
6836 /* Delete a store, but copy the value that would have been stored into
6837 the reaching_reg for later storing. */
6840 delete_store (expr
, bb
)
6841 struct ls_expr
* expr
;
6846 if (expr
->reaching_reg
== NULL_RTX
)
6847 expr
->reaching_reg
= gen_reg_rtx (GET_MODE (expr
->pattern
));
6850 /* If there is more than 1 store, the earlier ones will be dead,
6851 but it doesn't hurt to replace them here. */
6852 reg
= expr
->reaching_reg
;
6854 for (i
= AVAIL_STORE_LIST (expr
); i
; i
= XEXP (i
, 1))
6857 if (BLOCK_FOR_INSN (del
) == bb
)
6859 /* We know there is only one since we deleted redundant
6860 ones during the available computation. */
6861 replace_store_insn (reg
, del
, bb
);
6867 /* Free memory used by store motion. */
6870 free_store_memory ()
6875 sbitmap_vector_free (ae_gen
);
6877 sbitmap_vector_free (ae_kill
);
6879 sbitmap_vector_free (transp
);
6881 sbitmap_vector_free (st_antloc
);
6883 sbitmap_vector_free (pre_insert_map
);
6885 sbitmap_vector_free (pre_delete_map
);
6886 if (reg_set_in_block
)
6887 sbitmap_vector_free (reg_set_in_block
);
6889 ae_gen
= ae_kill
= transp
= st_antloc
= NULL
;
6890 pre_insert_map
= pre_delete_map
= reg_set_in_block
= NULL
;
6893 /* Perform store motion. Much like gcse, except we move expressions the
6894 other way by looking at the flowgraph in reverse. */
6900 struct ls_expr
* ptr
;
6901 int update_flow
= 0;
6905 fprintf (gcse_file
, "before store motion\n");
6906 print_rtl (gcse_file
, get_insns ());
6910 init_alias_analysis ();
6912 /* Find all the stores that are live to the end of their block. */
6913 num_stores
= compute_store_table ();
6914 if (num_stores
== 0)
6916 sbitmap_vector_free (reg_set_in_block
);
6917 end_alias_analysis ();
6921 /* Now compute whats actually available to move. */
6922 add_noreturn_fake_exit_edges ();
6923 build_store_vectors ();
6925 edge_list
= pre_edge_rev_lcm (gcse_file
, num_stores
, transp
, ae_gen
,
6926 st_antloc
, ae_kill
, &pre_insert_map
,
6929 /* Now we want to insert the new stores which are going to be needed. */
6930 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6932 for (x
= 0; x
< n_basic_blocks
; x
++)
6933 if (TEST_BIT (pre_delete_map
[x
], ptr
->index
))
6934 delete_store (ptr
, BASIC_BLOCK (x
));
6936 for (x
= 0; x
< NUM_EDGES (edge_list
); x
++)
6937 if (TEST_BIT (pre_insert_map
[x
], ptr
->index
))
6938 update_flow
|= insert_store (ptr
, INDEX_EDGE (edge_list
, x
));
6942 commit_edge_insertions ();
6944 free_store_memory ();
6945 free_edge_list (edge_list
);
6946 remove_fake_edges ();
6947 end_alias_analysis ();