1 /* Global common subexpression elimination/Partial redundancy elimination
2 and global constant/copy propagation for GNU compiler.
3 Copyright (C) 1997, 1998, 1999, 2000, 2001 Free Software Foundation, Inc.
5 This file is part of GNU CC.
7 GNU CC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 2, or (at your option)
12 GNU CC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GNU CC; see the file COPYING. If not, write to
19 the Free Software Foundation, 59 Temple Place - Suite 330,
20 Boston, MA 02111-1307, USA. */
23 - reordering of memory allocation and freeing to be more space efficient
24 - do rough calc of how many regs are needed in each block, and a rough
25 calc of how many regs are available in each class and use that to
26 throttle back the code in cases where RTX_COST is minimal.
27 - dead store elimination
28 - a store to the same address as a load does not kill the load if the
29 source of the store is also the destination of the load. Handling this
30 allows more load motion, particularly out of loops.
31 - ability to realloc sbitmap vectors would allow one initial computation
32 of reg_set_in_block with only subsequent additions, rather than
33 recomputing it for each pass
37 /* References searched while implementing this.
39 Compilers Principles, Techniques and Tools
43 Global Optimization by Suppression of Partial Redundancies
45 communications of the acm, Vol. 22, Num. 2, Feb. 1979
47 A Portable Machine-Independent Global Optimizer - Design and Measurements
49 Stanford Ph.D. thesis, Dec. 1983
51 A Fast Algorithm for Code Movement Optimization
53 SIGPLAN Notices, Vol. 23, Num. 10, Oct. 1988
55 A Solution to a Problem with Morel and Renvoise's
56 Global Optimization by Suppression of Partial Redundancies
57 K-H Drechsler, M.P. Stadel
58 ACM TOPLAS, Vol. 10, Num. 4, Oct. 1988
60 Practical Adaptation of the Global Optimization
61 Algorithm of Morel and Renvoise
63 ACM TOPLAS, Vol. 13, Num. 2. Apr. 1991
65 Efficiently Computing Static Single Assignment Form and the Control
67 R. Cytron, J. Ferrante, B.K. Rosen, M.N. Wegman, and F.K. Zadeck
68 ACM TOPLAS, Vol. 13, Num. 4, Oct. 1991
71 J. Knoop, O. Ruthing, B. Steffen
72 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
74 What's In a Region? Or Computing Control Dependence Regions in Near-Linear
75 Time for Reducible Flow Control
77 ACM Letters on Programming Languages and Systems,
78 Vol. 2, Num. 1-4, Mar-Dec 1993
80 An Efficient Representation for Sparse Sets
81 Preston Briggs, Linda Torczon
82 ACM Letters on Programming Languages and Systems,
83 Vol. 2, Num. 1-4, Mar-Dec 1993
85 A Variation of Knoop, Ruthing, and Steffen's Lazy Code Motion
86 K-H Drechsler, M.P. Stadel
87 ACM SIGPLAN Notices, Vol. 28, Num. 5, May 1993
89 Partial Dead Code Elimination
90 J. Knoop, O. Ruthing, B. Steffen
91 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
93 Effective Partial Redundancy Elimination
94 P. Briggs, K.D. Cooper
95 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
97 The Program Structure Tree: Computing Control Regions in Linear Time
98 R. Johnson, D. Pearson, K. Pingali
99 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
101 Optimal Code Motion: Theory and Practice
102 J. Knoop, O. Ruthing, B. Steffen
103 ACM TOPLAS, Vol. 16, Num. 4, Jul. 1994
105 The power of assignment motion
106 J. Knoop, O. Ruthing, B. Steffen
107 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
109 Global code motion / global value numbering
111 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
113 Value Driven Redundancy Elimination
115 Rice University Ph.D. thesis, Apr. 1996
119 Massively Scalar Compiler Project, Rice University, Sep. 1996
121 High Performance Compilers for Parallel Computing
125 Advanced Compiler Design and Implementation
127 Morgan Kaufmann, 1997
129 Building an Optimizing Compiler
133 People wishing to speed up the code here should read:
134 Elimination Algorithms for Data Flow Analysis
135 B.G. Ryder, M.C. Paull
136 ACM Computing Surveys, Vol. 18, Num. 3, Sep. 1986
138 How to Analyze Large Programs Efficiently and Informatively
139 D.M. Dhamdhere, B.K. Rosen, F.K. Zadeck
140 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
142 People wishing to do something different can find various possibilities
143 in the above papers and elsewhere.
153 #include "hard-reg-set.h"
156 #include "insn-config.h"
158 #include "basic-block.h"
160 #include "function.h"
165 #define obstack_chunk_alloc gmalloc
166 #define obstack_chunk_free free
168 /* Maximum number of passes to perform. */
171 /* Propagate flow information through back edges and thus enable PRE's
172 moving loop invariant calculations out of loops.
174 Originally this tended to create worse overall code, but several
175 improvements during the development of PRE seem to have made following
176 back edges generally a win.
178 Note much of the loop invariant code motion done here would normally
179 be done by loop.c, which has more heuristics for when to move invariants
180 out of loops. At some point we might need to move some of those
181 heuristics into gcse.c. */
182 #define FOLLOW_BACK_EDGES 1
184 /* We support GCSE via Partial Redundancy Elimination. PRE optimizations
185 are a superset of those done by GCSE.
187 We perform the following steps:
189 1) Compute basic block information.
191 2) Compute table of places where registers are set.
193 3) Perform copy/constant propagation.
195 4) Perform global cse.
197 5) Perform another pass of copy/constant propagation.
199 Two passes of copy/constant propagation are done because the first one
200 enables more GCSE and the second one helps to clean up the copies that
201 GCSE creates. This is needed more for PRE than for Classic because Classic
202 GCSE will try to use an existing register containing the common
203 subexpression rather than create a new one. This is harder to do for PRE
204 because of the code motion (which Classic GCSE doesn't do).
206 Expressions we are interested in GCSE-ing are of the form
207 (set (pseudo-reg) (expression)).
208 Function want_to_gcse_p says what these are.
210 PRE handles moving invariant expressions out of loops (by treating them as
211 partially redundant).
213 Eventually it would be nice to replace cse.c/gcse.c with SSA (static single
214 assignment) based GVN (global value numbering). L. T. Simpson's paper
215 (Rice University) on value numbering is a useful reference for this.
217 **********************
219 We used to support multiple passes but there are diminishing returns in
220 doing so. The first pass usually makes 90% of the changes that are doable.
221 A second pass can make a few more changes made possible by the first pass.
222 Experiments show any further passes don't make enough changes to justify
225 A study of spec92 using an unlimited number of passes:
226 [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
227 [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
228 [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
230 It was found doing copy propagation between each pass enables further
233 PRE is quite expensive in complicated functions because the DFA can take
234 awhile to converge. Hence we only perform one pass. Macro MAX_PASSES can
235 be modified if one wants to experiment.
237 **********************
239 The steps for PRE are:
241 1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
243 2) Perform the data flow analysis for PRE.
245 3) Delete the redundant instructions
247 4) Insert the required copies [if any] that make the partially
248 redundant instructions fully redundant.
250 5) For other reaching expressions, insert an instruction to copy the value
251 to a newly created pseudo that will reach the redundant instruction.
253 The deletion is done first so that when we do insertions we
254 know which pseudo reg to use.
256 Various papers have argued that PRE DFA is expensive (O(n^2)) and others
257 argue it is not. The number of iterations for the algorithm to converge
258 is typically 2-4 so I don't view it as that expensive (relatively speaking).
260 PRE GCSE depends heavily on the second CSE pass to clean up the copies
261 we create. To make an expression reach the place where it's redundant,
262 the result of the expression is copied to a new register, and the redundant
263 expression is deleted by replacing it with this new register. Classic GCSE
264 doesn't have this problem as much as it computes the reaching defs of
265 each register in each block and thus can try to use an existing register.
267 **********************
269 A fair bit of simplicity is created by creating small functions for simple
270 tasks, even when the function is only called in one place. This may
271 measurably slow things down [or may not] by creating more function call
272 overhead than is necessary. The source is laid out so that it's trivial
273 to make the affected functions inline so that one can measure what speed
274 up, if any, can be achieved, and maybe later when things settle things can
277 Help stamp out big monolithic functions! */
279 /* GCSE global vars. */
282 static FILE *gcse_file;
284 /* Note whether or not we should run jump optimization after gcse. We
285 want to do this for two cases.
287 * If we changed any jumps via cprop.
289 * If we added any labels via edge splitting. */
291 static int run_jump_opt_after_gcse;
293 /* Bitmaps are normally not included in debugging dumps.
294 However it's useful to be able to print them from GDB.
295 We could create special functions for this, but it's simpler to
296 just allow passing stderr to the dump_foo fns. Since stderr can
297 be a macro, we store a copy here. */
298 static FILE *debug_stderr;
300 /* An obstack for our working variables. */
301 static struct obstack gcse_obstack;
303 /* Non-zero for each mode that supports (set (reg) (reg)).
304 This is trivially true for integer and floating point values.
305 It may or may not be true for condition codes. */
306 static char can_copy_p[(int) NUM_MACHINE_MODES];
308 /* Non-zero if can_copy_p has been initialized. */
309 static int can_copy_init_p;
311 struct reg_use {rtx reg_rtx; };
313 /* Hash table of expressions. */
317 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
319 /* Index in the available expression bitmaps. */
321 /* Next entry with the same hash. */
322 struct expr *next_same_hash;
323 /* List of anticipatable occurrences in basic blocks in the function.
324 An "anticipatable occurrence" is one that is the first occurrence in the
325 basic block, the operands are not modified in the basic block prior
326 to the occurrence and the output is not used between the start of
327 the block and the occurrence. */
328 struct occr *antic_occr;
329 /* List of available occurrence in basic blocks in the function.
330 An "available occurrence" is one that is the last occurrence in the
331 basic block and the operands are not modified by following statements in
332 the basic block [including this insn]. */
333 struct occr *avail_occr;
334 /* Non-null if the computation is PRE redundant.
335 The value is the newly created pseudo-reg to record a copy of the
336 expression in all the places that reach the redundant copy. */
340 /* Occurrence of an expression.
341 There is one per basic block. If a pattern appears more than once the
342 last appearance is used [or first for anticipatable expressions]. */
346 /* Next occurrence of this expression. */
348 /* The insn that computes the expression. */
350 /* Non-zero if this [anticipatable] occurrence has been deleted. */
352 /* Non-zero if this [available] occurrence has been copied to
354 /* ??? This is mutually exclusive with deleted_p, so they could share
359 /* Expression and copy propagation hash tables.
360 Each hash table is an array of buckets.
361 ??? It is known that if it were an array of entries, structure elements
362 `next_same_hash' and `bitmap_index' wouldn't be necessary. However, it is
363 not clear whether in the final analysis a sufficient amount of memory would
364 be saved as the size of the available expression bitmaps would be larger
365 [one could build a mapping table without holes afterwards though].
366 Someday I'll perform the computation and figure it out. */
368 /* Total size of the expression hash table, in elements. */
369 static unsigned int expr_hash_table_size;
372 This is an array of `expr_hash_table_size' elements. */
373 static struct expr **expr_hash_table;
375 /* Total size of the copy propagation hash table, in elements. */
376 static unsigned int set_hash_table_size;
379 This is an array of `set_hash_table_size' elements. */
380 static struct expr **set_hash_table;
382 /* Mapping of uids to cuids.
383 Only real insns get cuids. */
384 static int *uid_cuid;
386 /* Highest UID in UID_CUID. */
389 /* Get the cuid of an insn. */
390 #ifdef ENABLE_CHECKING
391 #define INSN_CUID(INSN) (INSN_UID (INSN) > max_uid ? (abort (), 0) : uid_cuid[INSN_UID (INSN)])
393 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
396 /* Number of cuids. */
399 /* Mapping of cuids to insns. */
400 static rtx *cuid_insn;
402 /* Get insn from cuid. */
403 #define CUID_INSN(CUID) (cuid_insn[CUID])
405 /* Maximum register number in function prior to doing gcse + 1.
406 Registers created during this pass have regno >= max_gcse_regno.
407 This is named with "gcse" to not collide with global of same name. */
408 static unsigned int max_gcse_regno;
410 /* Maximum number of cse-able expressions found. */
413 /* Maximum number of assignments for copy propagation found. */
416 /* Table of registers that are modified.
418 For each register, each element is a list of places where the pseudo-reg
421 For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
422 requires knowledge of which blocks kill which regs [and thus could use
423 a bitmap instead of the lists `reg_set_table' uses].
425 `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
426 num-regs) [however perhaps it may be useful to keep the data as is]. One
427 advantage of recording things this way is that `reg_set_table' is fairly
428 sparse with respect to pseudo regs but for hard regs could be fairly dense
429 [relatively speaking]. And recording sets of pseudo-regs in lists speeds
430 up functions like compute_transp since in the case of pseudo-regs we only
431 need to iterate over the number of times a pseudo-reg is set, not over the
432 number of basic blocks [clearly there is a bit of a slow down in the cases
433 where a pseudo is set more than once in a block, however it is believed
434 that the net effect is to speed things up]. This isn't done for hard-regs
435 because recording call-clobbered hard-regs in `reg_set_table' at each
436 function call can consume a fair bit of memory, and iterating over
437 hard-regs stored this way in compute_transp will be more expensive. */
439 typedef struct reg_set
441 /* The next setting of this register. */
442 struct reg_set *next;
443 /* The insn where it was set. */
447 static reg_set **reg_set_table;
449 /* Size of `reg_set_table'.
450 The table starts out at max_gcse_regno + slop, and is enlarged as
452 static int reg_set_table_size;
454 /* Amount to grow `reg_set_table' by when it's full. */
455 #define REG_SET_TABLE_SLOP 100
457 /* Bitmap containing one bit for each register in the program.
458 Used when performing GCSE to track which registers have been set since
459 the start of the basic block. */
460 static sbitmap reg_set_bitmap;
462 /* For each block, a bitmap of registers set in the block.
463 This is used by expr_killed_p and compute_transp.
464 It is computed during hash table computation and not by compute_sets
465 as it includes registers added since the last pass (or between cprop and
466 gcse) and it's currently not easy to realloc sbitmap vectors. */
467 static sbitmap *reg_set_in_block;
469 /* For each block, non-zero if memory is set in that block.
470 This is computed during hash table computation and is used by
471 expr_killed_p and compute_transp.
472 ??? Handling of memory is very simple, we don't make any attempt
473 to optimize things (later).
474 ??? This can be computed by compute_sets since the information
476 static char *mem_set_in_block;
478 /* Various variables for statistics gathering. */
480 /* Memory used in a pass.
481 This isn't intended to be absolutely precise. Its intent is only
482 to keep an eye on memory usage. */
483 static int bytes_used;
485 /* GCSE substitutions made. */
486 static int gcse_subst_count;
487 /* Number of copy instructions created. */
488 static int gcse_create_count;
489 /* Number of constants propagated. */
490 static int const_prop_count;
491 /* Number of copys propagated. */
492 static int copy_prop_count;
494 /* These variables are used by classic GCSE.
495 Normally they'd be defined a bit later, but `rd_gen' needs to
496 be declared sooner. */
498 /* Each block has a bitmap of each type.
499 The length of each blocks bitmap is:
501 max_cuid - for reaching definitions
502 n_exprs - for available expressions
504 Thus we view the bitmaps as 2 dimensional arrays. i.e.
505 rd_kill[block_num][cuid_num]
506 ae_kill[block_num][expr_num] */
508 /* For reaching defs */
509 static sbitmap *rd_kill, *rd_gen, *reaching_defs, *rd_out;
511 /* for available exprs */
512 static sbitmap *ae_kill, *ae_gen, *ae_in, *ae_out;
514 /* Objects of this type are passed around by the null-pointer check
516 struct null_pointer_info
518 /* The basic block being processed. */
520 /* The first register to be handled in this pass. */
521 unsigned int min_reg;
522 /* One greater than the last register to be handled in this pass. */
523 unsigned int max_reg;
524 sbitmap *nonnull_local;
525 sbitmap *nonnull_killed;
528 static void compute_can_copy PARAMS ((void));
529 static char *gmalloc PARAMS ((unsigned int));
530 static char *grealloc PARAMS ((char *, unsigned int));
531 static char *gcse_alloc PARAMS ((unsigned long));
532 static void alloc_gcse_mem PARAMS ((rtx));
533 static void free_gcse_mem PARAMS ((void));
534 static void alloc_reg_set_mem PARAMS ((int));
535 static void free_reg_set_mem PARAMS ((void));
536 static int get_bitmap_width PARAMS ((int, int, int));
537 static void record_one_set PARAMS ((int, rtx));
538 static void record_set_info PARAMS ((rtx, rtx, void *));
539 static void compute_sets PARAMS ((rtx));
540 static void hash_scan_insn PARAMS ((rtx, int, int));
541 static void hash_scan_set PARAMS ((rtx, rtx, int));
542 static void hash_scan_clobber PARAMS ((rtx, rtx));
543 static void hash_scan_call PARAMS ((rtx, rtx));
544 static int want_to_gcse_p PARAMS ((rtx));
545 static int oprs_unchanged_p PARAMS ((rtx, rtx, int));
546 static int oprs_anticipatable_p PARAMS ((rtx, rtx));
547 static int oprs_available_p PARAMS ((rtx, rtx));
548 static void insert_expr_in_table PARAMS ((rtx, enum machine_mode, rtx,
550 static void insert_set_in_table PARAMS ((rtx, rtx));
551 static unsigned int hash_expr PARAMS ((rtx, enum machine_mode, int *, int));
552 static unsigned int hash_expr_1 PARAMS ((rtx, enum machine_mode, int *));
553 static unsigned int hash_string_1 PARAMS ((const char *));
554 static unsigned int hash_set PARAMS ((int, int));
555 static int expr_equiv_p PARAMS ((rtx, rtx));
556 static void record_last_reg_set_info PARAMS ((rtx, int));
557 static void record_last_mem_set_info PARAMS ((rtx));
558 static void record_last_set_info PARAMS ((rtx, rtx, void *));
559 static void compute_hash_table PARAMS ((int));
560 static void alloc_set_hash_table PARAMS ((int));
561 static void free_set_hash_table PARAMS ((void));
562 static void compute_set_hash_table PARAMS ((void));
563 static void alloc_expr_hash_table PARAMS ((unsigned int));
564 static void free_expr_hash_table PARAMS ((void));
565 static void compute_expr_hash_table PARAMS ((void));
566 static void dump_hash_table PARAMS ((FILE *, const char *, struct expr **,
568 static struct expr *lookup_expr PARAMS ((rtx));
569 static struct expr *lookup_set PARAMS ((unsigned int, rtx));
570 static struct expr *next_set PARAMS ((unsigned int, struct expr *));
571 static void reset_opr_set_tables PARAMS ((void));
572 static int oprs_not_set_p PARAMS ((rtx, rtx));
573 static void mark_call PARAMS ((rtx));
574 static void mark_set PARAMS ((rtx, rtx));
575 static void mark_clobber PARAMS ((rtx, rtx));
576 static void mark_oprs_set PARAMS ((rtx));
577 static void alloc_cprop_mem PARAMS ((int, int));
578 static void free_cprop_mem PARAMS ((void));
579 static void compute_transp PARAMS ((rtx, int, sbitmap *, int));
580 static void compute_transpout PARAMS ((void));
581 static void compute_local_properties PARAMS ((sbitmap *, sbitmap *, sbitmap *,
583 static void compute_cprop_data PARAMS ((void));
584 static void find_used_regs PARAMS ((rtx));
585 static int try_replace_reg PARAMS ((rtx, rtx, rtx));
586 static struct expr *find_avail_set PARAMS ((int, rtx));
587 static int cprop_jump PARAMS ((rtx, rtx, rtx));
589 static int cprop_cc0_jump PARAMS ((rtx, struct reg_use *, rtx));
591 static int cprop_insn PARAMS ((rtx, int));
592 static int cprop PARAMS ((int));
593 static int one_cprop_pass PARAMS ((int, int));
594 static void alloc_pre_mem PARAMS ((int, int));
595 static void free_pre_mem PARAMS ((void));
596 static void compute_pre_data PARAMS ((void));
597 static int pre_expr_reaches_here_p PARAMS ((int, struct expr *, int));
598 static void insert_insn_end_bb PARAMS ((struct expr *, int, int));
599 static void pre_insert_copy_insn PARAMS ((struct expr *, rtx));
600 static void pre_insert_copies PARAMS ((void));
601 static int pre_delete PARAMS ((void));
602 static int pre_gcse PARAMS ((void));
603 static int one_pre_gcse_pass PARAMS ((int));
604 static void add_label_notes PARAMS ((rtx, rtx));
605 static void alloc_code_hoist_mem PARAMS ((int, int));
606 static void free_code_hoist_mem PARAMS ((void));
607 static void compute_code_hoist_vbeinout PARAMS ((void));
608 static void compute_code_hoist_data PARAMS ((void));
609 static int hoist_expr_reaches_here_p PARAMS ((int, int, int, char *));
610 static void hoist_code PARAMS ((void));
611 static int one_code_hoisting_pass PARAMS ((void));
612 static void alloc_rd_mem PARAMS ((int, int));
613 static void free_rd_mem PARAMS ((void));
614 static void handle_rd_kill_set PARAMS ((rtx, int, int));
615 static void compute_kill_rd PARAMS ((void));
616 static void compute_rd PARAMS ((void));
617 static void alloc_avail_expr_mem PARAMS ((int, int));
618 static void free_avail_expr_mem PARAMS ((void));
619 static void compute_ae_gen PARAMS ((void));
620 static int expr_killed_p PARAMS ((rtx, int));
621 static void compute_ae_kill PARAMS ((sbitmap *, sbitmap *));
622 static int expr_reaches_here_p PARAMS ((struct occr *, struct expr *,
624 static rtx computing_insn PARAMS ((struct expr *, rtx));
625 static int def_reaches_here_p PARAMS ((rtx, rtx));
626 static int can_disregard_other_sets PARAMS ((struct reg_set **, rtx, int));
627 static int handle_avail_expr PARAMS ((rtx, struct expr *));
628 static int classic_gcse PARAMS ((void));
629 static int one_classic_gcse_pass PARAMS ((int));
630 static void invalidate_nonnull_info PARAMS ((rtx, rtx, void *));
631 static void delete_null_pointer_checks_1 PARAMS ((unsigned int *, sbitmap *,
633 struct null_pointer_info *));
634 static rtx process_insert_insn PARAMS ((struct expr *));
635 static int pre_edge_insert PARAMS ((struct edge_list *, struct expr **));
636 static int expr_reaches_here_p_work PARAMS ((struct occr *, struct expr *,
638 static int pre_expr_reaches_here_p_work PARAMS ((int, struct expr *,
641 /* Entry point for global common subexpression elimination.
642 F is the first instruction in the function. */
650 /* Bytes used at start of pass. */
651 int initial_bytes_used;
652 /* Maximum number of bytes used by a pass. */
654 /* Point to release obstack data from for each pass. */
655 char *gcse_obstack_bottom;
657 /* We do not construct an accurate cfg in functions which call
658 setjmp, so just punt to be safe. */
659 if (current_function_calls_setjmp)
662 /* Assume that we do not need to run jump optimizations after gcse. */
663 run_jump_opt_after_gcse = 0;
665 /* For calling dump_foo fns from gdb. */
666 debug_stderr = stderr;
669 /* Identify the basic block information for this function, including
670 successors and predecessors. */
671 max_gcse_regno = max_reg_num ();
674 dump_flow_info (file);
676 /* Return if there's nothing to do. */
677 if (n_basic_blocks <= 1)
680 /* Trying to perform global optimizations on flow graphs which have
681 a high connectivity will take a long time and is unlikely to be
684 In normal circumstances a cfg should have about twice has many edges
685 as blocks. But we do not want to punish small functions which have
686 a couple switch statements. So we require a relatively large number
687 of basic blocks and the ratio of edges to blocks to be high. */
688 if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
690 if (warn_disabled_optimization)
691 warning ("GCSE disabled: %d > 1000 basic blocks and %d >= 20 edges/basic block",
692 n_basic_blocks, n_edges / n_basic_blocks);
696 /* See what modes support reg/reg copy operations. */
697 if (! can_copy_init_p)
703 gcc_obstack_init (&gcse_obstack);
706 /* Record where pseudo-registers are set. This data is kept accurate
707 during each pass. ??? We could also record hard-reg information here
708 [since it's unchanging], however it is currently done during hash table
711 It may be tempting to compute MEM set information here too, but MEM sets
712 will be subject to code motion one day and thus we need to compute
713 information about memory sets when we build the hash tables. */
715 alloc_reg_set_mem (max_gcse_regno);
719 initial_bytes_used = bytes_used;
721 gcse_obstack_bottom = gcse_alloc (1);
723 while (changed && pass < MAX_PASSES)
727 fprintf (file, "GCSE pass %d\n\n", pass + 1);
729 /* Initialize bytes_used to the space for the pred/succ lists,
730 and the reg_set_table data. */
731 bytes_used = initial_bytes_used;
733 /* Each pass may create new registers, so recalculate each time. */
734 max_gcse_regno = max_reg_num ();
738 /* Don't allow constant propagation to modify jumps
740 changed = one_cprop_pass (pass + 1, 0);
743 changed |= one_classic_gcse_pass (pass + 1);
746 changed |= one_pre_gcse_pass (pass + 1);
748 alloc_reg_set_mem (max_reg_num ());
750 run_jump_opt_after_gcse = 1;
753 if (max_pass_bytes < bytes_used)
754 max_pass_bytes = bytes_used;
756 /* Free up memory, then reallocate for code hoisting. We can
757 not re-use the existing allocated memory because the tables
758 will not have info for the insns or registers created by
759 partial redundancy elimination. */
762 /* It does not make sense to run code hoisting unless we optimizing
763 for code size -- it rarely makes programs faster, and can make
764 them bigger if we did partial redundancy elimination (when optimizing
765 for space, we use a classic gcse algorithm instead of partial
766 redundancy algorithms). */
769 max_gcse_regno = max_reg_num ();
771 changed |= one_code_hoisting_pass ();
774 if (max_pass_bytes < bytes_used)
775 max_pass_bytes = bytes_used;
780 fprintf (file, "\n");
784 obstack_free (&gcse_obstack, gcse_obstack_bottom);
788 /* Do one last pass of copy propagation, including cprop into
789 conditional jumps. */
791 max_gcse_regno = max_reg_num ();
793 /* This time, go ahead and allow cprop to alter jumps. */
794 one_cprop_pass (pass + 1, 1);
799 fprintf (file, "GCSE of %s: %d basic blocks, ",
800 current_function_name, n_basic_blocks);
801 fprintf (file, "%d pass%s, %d bytes\n\n",
802 pass, pass > 1 ? "es" : "", max_pass_bytes);
805 obstack_free (&gcse_obstack, NULL_PTR);
807 return run_jump_opt_after_gcse;
810 /* Misc. utilities. */
812 /* Compute which modes support reg/reg copy operations. */
818 #ifndef AVOID_CCMODE_COPIES
821 memset (can_copy_p, 0, NUM_MACHINE_MODES);
824 for (i = 0; i < NUM_MACHINE_MODES; i++)
825 if (GET_MODE_CLASS (i) == MODE_CC)
827 #ifdef AVOID_CCMODE_COPIES
830 reg = gen_rtx_REG ((enum machine_mode) i, LAST_VIRTUAL_REGISTER + 1);
831 insn = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
832 if (recog (PATTERN (insn), insn, NULL_PTR) >= 0)
842 /* Cover function to xmalloc to record bytes allocated. */
849 return xmalloc (size);
852 /* Cover function to xrealloc.
853 We don't record the additional size since we don't know it.
854 It won't affect memory usage stats much anyway. */
861 return xrealloc (ptr, size);
864 /* Cover function to obstack_alloc.
865 We don't need to record the bytes allocated here since
866 obstack_chunk_alloc is set to gmalloc. */
872 return (char *) obstack_alloc (&gcse_obstack, size);
875 /* Allocate memory for the cuid mapping array,
876 and reg/memory set tracking tables.
878 This is called at the start of each pass. */
887 /* Find the largest UID and create a mapping from UIDs to CUIDs.
888 CUIDs are like UIDs except they increase monotonically, have no gaps,
889 and only apply to real insns. */
891 max_uid = get_max_uid ();
892 n = (max_uid + 1) * sizeof (int);
893 uid_cuid = (int *) gmalloc (n);
894 memset ((char *) uid_cuid, 0, n);
895 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
898 uid_cuid[INSN_UID (insn)] = i++;
900 uid_cuid[INSN_UID (insn)] = i;
903 /* Create a table mapping cuids to insns. */
906 n = (max_cuid + 1) * sizeof (rtx);
907 cuid_insn = (rtx *) gmalloc (n);
908 memset ((char *) cuid_insn, 0, n);
909 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
911 CUID_INSN (i++) = insn;
913 /* Allocate vars to track sets of regs. */
914 reg_set_bitmap = (sbitmap) sbitmap_alloc (max_gcse_regno);
916 /* Allocate vars to track sets of regs, memory per block. */
917 reg_set_in_block = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks,
919 mem_set_in_block = (char *) gmalloc (n_basic_blocks);
922 /* Free memory allocated by alloc_gcse_mem. */
930 free (reg_set_bitmap);
932 free (reg_set_in_block);
933 free (mem_set_in_block);
936 /* Many of the global optimization algorithms work by solving dataflow
937 equations for various expressions. Initially, some local value is
938 computed for each expression in each block. Then, the values across the
939 various blocks are combined (by following flow graph edges) to arrive at
940 global values. Conceptually, each set of equations is independent. We
941 may therefore solve all the equations in parallel, solve them one at a
942 time, or pick any intermediate approach.
944 When you're going to need N two-dimensional bitmaps, each X (say, the
945 number of blocks) by Y (say, the number of expressions), call this
946 function. It's not important what X and Y represent; only that Y
947 correspond to the things that can be done in parallel. This function will
948 return an appropriate chunking factor C; you should solve C sets of
949 equations in parallel. By going through this function, we can easily
950 trade space against time; by solving fewer equations in parallel we use
954 get_bitmap_width (n, x, y)
959 /* It's not really worth figuring out *exactly* how much memory will
960 be used by a particular choice. The important thing is to get
961 something approximately right. */
962 size_t max_bitmap_memory = 10 * 1024 * 1024;
964 /* The number of bytes we'd use for a single column of minimum
966 size_t column_size = n * x * sizeof (SBITMAP_ELT_TYPE);
968 /* Often, it's reasonable just to solve all the equations in
970 if (column_size * SBITMAP_SET_SIZE (y) <= max_bitmap_memory)
973 /* Otherwise, pick the largest width we can, without going over the
975 return SBITMAP_ELT_BITS * ((max_bitmap_memory + column_size - 1)
979 /* Compute the local properties of each recorded expression.
981 Local properties are those that are defined by the block, irrespective of
984 An expression is transparent in a block if its operands are not modified
987 An expression is computed (locally available) in a block if it is computed
988 at least once and expression would contain the same value if the
989 computation was moved to the end of the block.
991 An expression is locally anticipatable in a block if it is computed at
992 least once and expression would contain the same value if the computation
993 was moved to the beginning of the block.
995 We call this routine for cprop, pre and code hoisting. They all compute
996 basically the same information and thus can easily share this code.
998 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
999 properties. If NULL, then it is not necessary to compute or record that
1000 particular property.
1002 SETP controls which hash table to look at. If zero, this routine looks at
1003 the expr hash table; if nonzero this routine looks at the set hash table.
1004 Additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
1008 compute_local_properties (transp, comp, antloc, setp)
1014 unsigned int i, hash_table_size;
1015 struct expr **hash_table;
1017 /* Initialize any bitmaps that were passed in. */
1021 sbitmap_vector_zero (transp, n_basic_blocks);
1023 sbitmap_vector_ones (transp, n_basic_blocks);
1027 sbitmap_vector_zero (comp, n_basic_blocks);
1029 sbitmap_vector_zero (antloc, n_basic_blocks);
1031 /* We use the same code for cprop, pre and hoisting. For cprop
1032 we care about the set hash table, for pre and hoisting we
1033 care about the expr hash table. */
1034 hash_table_size = setp ? set_hash_table_size : expr_hash_table_size;
1035 hash_table = setp ? set_hash_table : expr_hash_table;
1037 for (i = 0; i < hash_table_size; i++)
1041 for (expr = hash_table[i]; expr != NULL; expr = expr->next_same_hash)
1043 int indx = expr->bitmap_index;
1046 /* The expression is transparent in this block if it is not killed.
1047 We start by assuming all are transparent [none are killed], and
1048 then reset the bits for those that are. */
1050 compute_transp (expr->expr, indx, transp, setp);
1052 /* The occurrences recorded in antic_occr are exactly those that
1053 we want to set to non-zero in ANTLOC. */
1055 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
1057 SET_BIT (antloc[BLOCK_NUM (occr->insn)], indx);
1059 /* While we're scanning the table, this is a good place to
1061 occr->deleted_p = 0;
1064 /* The occurrences recorded in avail_occr are exactly those that
1065 we want to set to non-zero in COMP. */
1067 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
1069 SET_BIT (comp[BLOCK_NUM (occr->insn)], indx);
1071 /* While we're scanning the table, this is a good place to
1076 /* While we're scanning the table, this is a good place to
1078 expr->reaching_reg = 0;
1083 /* Register set information.
1085 `reg_set_table' records where each register is set or otherwise
1088 static struct obstack reg_set_obstack;
1091 alloc_reg_set_mem (n_regs)
1096 reg_set_table_size = n_regs + REG_SET_TABLE_SLOP;
1097 n = reg_set_table_size * sizeof (struct reg_set *);
1098 reg_set_table = (struct reg_set **) gmalloc (n);
1099 memset ((char *) reg_set_table, 0, n);
1101 gcc_obstack_init (®_set_obstack);
1107 free (reg_set_table);
1108 obstack_free (®_set_obstack, NULL_PTR);
1111 /* Record REGNO in the reg_set table. */
1114 record_one_set (regno, insn)
1118 /* Allocate a new reg_set element and link it onto the list. */
1119 struct reg_set *new_reg_info;
1121 /* If the table isn't big enough, enlarge it. */
1122 if (regno >= reg_set_table_size)
1124 int new_size = regno + REG_SET_TABLE_SLOP;
1127 = (struct reg_set **) grealloc ((char *) reg_set_table,
1128 new_size * sizeof (struct reg_set *));
1129 memset ((char *) (reg_set_table + reg_set_table_size), 0,
1130 (new_size - reg_set_table_size) * sizeof (struct reg_set *));
1131 reg_set_table_size = new_size;
1134 new_reg_info = (struct reg_set *) obstack_alloc (®_set_obstack,
1135 sizeof (struct reg_set));
1136 bytes_used += sizeof (struct reg_set);
1137 new_reg_info->insn = insn;
1138 new_reg_info->next = reg_set_table[regno];
1139 reg_set_table[regno] = new_reg_info;
1142 /* Called from compute_sets via note_stores to handle one SET or CLOBBER in
1143 an insn. The DATA is really the instruction in which the SET is
1147 record_set_info (dest, setter, data)
1148 rtx dest, setter ATTRIBUTE_UNUSED;
1151 rtx record_set_insn = (rtx) data;
1153 if (GET_CODE (dest) == REG && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
1154 record_one_set (REGNO (dest), record_set_insn);
1157 /* Scan the function and record each set of each pseudo-register.
1159 This is called once, at the start of the gcse pass. See the comments for
1160 `reg_set_table' for further documenation. */
1168 for (insn = f; insn != 0; insn = NEXT_INSN (insn))
1170 note_stores (PATTERN (insn), record_set_info, insn);
1173 /* Hash table support. */
1175 /* For each register, the cuid of the first/last insn in the block to set it,
1176 or -1 if not set. */
1177 #define NEVER_SET -1
1178 static int *reg_first_set;
1179 static int *reg_last_set;
1181 /* While computing "first/last set" info, this is the CUID of first/last insn
1182 to set memory or -1 if not set. `mem_last_set' is also used when
1183 performing GCSE to record whether memory has been set since the beginning
1186 Note that handling of memory is very simple, we don't make any attempt
1187 to optimize things (later). */
1188 static int mem_first_set;
1189 static int mem_last_set;
1191 /* See whether X, the source of a set, is something we want to consider for
1198 static rtx test_insn = 0;
1199 int num_clobbers = 0;
1202 switch (GET_CODE (x))
1215 /* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
1216 if (general_operand (x, GET_MODE (x)))
1218 else if (GET_MODE (x) == VOIDmode)
1221 /* Otherwise, check if we can make a valid insn from it. First initialize
1222 our test insn if we haven't already. */
1226 = make_insn_raw (gen_rtx_SET (VOIDmode,
1227 gen_rtx_REG (word_mode,
1228 FIRST_PSEUDO_REGISTER * 2),
1230 NEXT_INSN (test_insn) = PREV_INSN (test_insn) = 0;
1231 ggc_add_rtx_root (&test_insn, 1);
1234 /* Now make an insn like the one we would make when GCSE'ing and see if
1236 PUT_MODE (SET_DEST (PATTERN (test_insn)), GET_MODE (x));
1237 SET_SRC (PATTERN (test_insn)) = x;
1238 return ((icode = recog (PATTERN (test_insn), test_insn, &num_clobbers)) >= 0
1239 && (num_clobbers == 0 || ! added_clobbers_hard_reg_p (icode)));
1242 /* Return non-zero if the operands of expression X are unchanged from the
1243 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1244 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1247 oprs_unchanged_p (x, insn, avail_p)
1258 code = GET_CODE (x);
1263 return (reg_last_set[REGNO (x)] == NEVER_SET
1264 || reg_last_set[REGNO (x)] < INSN_CUID (insn));
1266 return (reg_first_set[REGNO (x)] == NEVER_SET
1267 || reg_first_set[REGNO (x)] >= INSN_CUID (insn));
1270 if (avail_p && mem_last_set != NEVER_SET
1271 && mem_last_set >= INSN_CUID (insn))
1273 else if (! avail_p && mem_first_set != NEVER_SET
1274 && mem_first_set < INSN_CUID (insn))
1277 return oprs_unchanged_p (XEXP (x, 0), insn, avail_p);
1302 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1306 /* If we are about to do the last recursive call needed at this
1307 level, change it into iteration. This function is called enough
1310 return oprs_unchanged_p (XEXP (x, i), insn, avail_p);
1312 else if (! oprs_unchanged_p (XEXP (x, i), insn, avail_p))
1315 else if (fmt[i] == 'E')
1316 for (j = 0; j < XVECLEN (x, i); j++)
1317 if (! oprs_unchanged_p (XVECEXP (x, i, j), insn, avail_p))
1324 /* Return non-zero if the operands of expression X are unchanged from
1325 the start of INSN's basic block up to but not including INSN. */
1328 oprs_anticipatable_p (x, insn)
1331 return oprs_unchanged_p (x, insn, 0);
1334 /* Return non-zero if the operands of expression X are unchanged from
1335 INSN to the end of INSN's basic block. */
1338 oprs_available_p (x, insn)
1341 return oprs_unchanged_p (x, insn, 1);
1344 /* Hash expression X.
1346 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1347 indicating if a volatile operand is found or if the expression contains
1348 something we don't want to insert in the table.
1350 ??? One might want to merge this with canon_hash. Later. */
1353 hash_expr (x, mode, do_not_record_p, hash_table_size)
1355 enum machine_mode mode;
1356 int *do_not_record_p;
1357 int hash_table_size;
1361 *do_not_record_p = 0;
1363 hash = hash_expr_1 (x, mode, do_not_record_p);
1364 return hash % hash_table_size;
1367 /* Hash a string. Just add its bytes up. */
1369 static inline unsigned
1374 const unsigned char *p = (const unsigned char *)ps;
1383 /* Subroutine of hash_expr to do the actual work. */
1386 hash_expr_1 (x, mode, do_not_record_p)
1388 enum machine_mode mode;
1389 int *do_not_record_p;
1396 /* Used to turn recursion into iteration. We can't rely on GCC's
1397 tail-recursion eliminatio since we need to keep accumulating values
1404 code = GET_CODE (x);
1408 hash += ((unsigned int) REG << 7) + REGNO (x);
1412 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
1413 + (unsigned int) INTVAL (x));
1417 /* This is like the general case, except that it only counts
1418 the integers representing the constant. */
1419 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
1420 if (GET_MODE (x) != VOIDmode)
1421 for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++)
1422 hash += (unsigned int) XWINT (x, i);
1424 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
1425 + (unsigned int) CONST_DOUBLE_HIGH (x));
1428 /* Assume there is only one rtx object for any given label. */
1430 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
1431 differences and differences between each stage's debugging dumps. */
1432 hash += (((unsigned int) LABEL_REF << 7)
1433 + CODE_LABEL_NUMBER (XEXP (x, 0)));
1438 /* Don't hash on the symbol's address to avoid bootstrap differences.
1439 Different hash values may cause expressions to be recorded in
1440 different orders and thus different registers to be used in the
1441 final assembler. This also avoids differences in the dump files
1442 between various stages. */
1444 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
1447 h += (h << 7) + *p++; /* ??? revisit */
1449 hash += ((unsigned int) SYMBOL_REF << 7) + h;
1454 if (MEM_VOLATILE_P (x))
1456 *do_not_record_p = 1;
1460 hash += (unsigned int) MEM;
1461 hash += MEM_ALIAS_SET (x);
1472 case UNSPEC_VOLATILE:
1473 *do_not_record_p = 1;
1477 if (MEM_VOLATILE_P (x))
1479 *do_not_record_p = 1;
1484 /* We don't want to take the filename and line into account. */
1485 hash += (unsigned) code + (unsigned) GET_MODE (x)
1486 + hash_string_1 (ASM_OPERANDS_TEMPLATE (x))
1487 + hash_string_1 (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
1488 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
1490 if (ASM_OPERANDS_INPUT_LENGTH (x))
1492 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
1494 hash += (hash_expr_1 (ASM_OPERANDS_INPUT (x, i),
1495 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
1497 + hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT
1501 hash += hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
1502 x = ASM_OPERANDS_INPUT (x, 0);
1503 mode = GET_MODE (x);
1513 hash += (unsigned) code + (unsigned) GET_MODE (x);
1514 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1518 /* If we are about to do the last recursive call
1519 needed at this level, change it into iteration.
1520 This function is called enough to be worth it. */
1527 hash += hash_expr_1 (XEXP (x, i), 0, do_not_record_p);
1528 if (*do_not_record_p)
1532 else if (fmt[i] == 'E')
1533 for (j = 0; j < XVECLEN (x, i); j++)
1535 hash += hash_expr_1 (XVECEXP (x, i, j), 0, do_not_record_p);
1536 if (*do_not_record_p)
1540 else if (fmt[i] == 's')
1541 hash += hash_string_1 (XSTR (x, i));
1542 else if (fmt[i] == 'i')
1543 hash += (unsigned int) XINT (x, i);
1551 /* Hash a set of register REGNO.
1553 Sets are hashed on the register that is set. This simplifies the PRE copy
1556 ??? May need to make things more elaborate. Later, as necessary. */
1559 hash_set (regno, hash_table_size)
1561 int hash_table_size;
1566 return hash % hash_table_size;
1569 /* Return non-zero if exp1 is equivalent to exp2.
1570 ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
1577 register enum rtx_code code;
1578 register const char *fmt;
1583 if (x == 0 || y == 0)
1586 code = GET_CODE (x);
1587 if (code != GET_CODE (y))
1590 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
1591 if (GET_MODE (x) != GET_MODE (y))
1601 return INTVAL (x) == INTVAL (y);
1604 return XEXP (x, 0) == XEXP (y, 0);
1607 return XSTR (x, 0) == XSTR (y, 0);
1610 return REGNO (x) == REGNO (y);
1613 /* Can't merge two expressions in different alias sets, since we can
1614 decide that the expression is transparent in a block when it isn't,
1615 due to it being set with the different alias set. */
1616 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
1620 /* For commutative operations, check both orders. */
1628 return ((expr_equiv_p (XEXP (x, 0), XEXP (y, 0))
1629 && expr_equiv_p (XEXP (x, 1), XEXP (y, 1)))
1630 || (expr_equiv_p (XEXP (x, 0), XEXP (y, 1))
1631 && expr_equiv_p (XEXP (x, 1), XEXP (y, 0))));
1634 /* We don't use the generic code below because we want to
1635 disregard filename and line numbers. */
1637 /* A volatile asm isn't equivalent to any other. */
1638 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
1641 if (GET_MODE (x) != GET_MODE (y)
1642 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
1643 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
1644 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
1645 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
1646 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
1649 if (ASM_OPERANDS_INPUT_LENGTH (x))
1651 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
1652 if (! expr_equiv_p (ASM_OPERANDS_INPUT (x, i),
1653 ASM_OPERANDS_INPUT (y, i))
1654 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
1655 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
1665 /* Compare the elements. If any pair of corresponding elements
1666 fail to match, return 0 for the whole thing. */
1668 fmt = GET_RTX_FORMAT (code);
1669 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1674 if (! expr_equiv_p (XEXP (x, i), XEXP (y, i)))
1679 if (XVECLEN (x, i) != XVECLEN (y, i))
1681 for (j = 0; j < XVECLEN (x, i); j++)
1682 if (! expr_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j)))
1687 if (strcmp (XSTR (x, i), XSTR (y, i)))
1692 if (XINT (x, i) != XINT (y, i))
1697 if (XWINT (x, i) != XWINT (y, i))
1712 /* Insert expression X in INSN in the hash table.
1713 If it is already present, record it as the last occurrence in INSN's
1716 MODE is the mode of the value X is being stored into.
1717 It is only used if X is a CONST_INT.
1719 ANTIC_P is non-zero if X is an anticipatable expression.
1720 AVAIL_P is non-zero if X is an available expression. */
1723 insert_expr_in_table (x, mode, insn, antic_p, avail_p)
1725 enum machine_mode mode;
1727 int antic_p, avail_p;
1729 int found, do_not_record_p;
1731 struct expr *cur_expr, *last_expr = NULL;
1732 struct occr *antic_occr, *avail_occr;
1733 struct occr *last_occr = NULL;
1735 hash = hash_expr (x, mode, &do_not_record_p, expr_hash_table_size);
1737 /* Do not insert expression in table if it contains volatile operands,
1738 or if hash_expr determines the expression is something we don't want
1739 to or can't handle. */
1740 if (do_not_record_p)
1743 cur_expr = expr_hash_table[hash];
1746 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1748 /* If the expression isn't found, save a pointer to the end of
1750 last_expr = cur_expr;
1751 cur_expr = cur_expr->next_same_hash;
1756 cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
1757 bytes_used += sizeof (struct expr);
1758 if (expr_hash_table[hash] == NULL)
1759 /* This is the first pattern that hashed to this index. */
1760 expr_hash_table[hash] = cur_expr;
1762 /* Add EXPR to end of this hash chain. */
1763 last_expr->next_same_hash = cur_expr;
1765 /* Set the fields of the expr element. */
1767 cur_expr->bitmap_index = n_exprs++;
1768 cur_expr->next_same_hash = NULL;
1769 cur_expr->antic_occr = NULL;
1770 cur_expr->avail_occr = NULL;
1773 /* Now record the occurrence(s). */
1776 antic_occr = cur_expr->antic_occr;
1778 /* Search for another occurrence in the same basic block. */
1779 while (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
1781 /* If an occurrence isn't found, save a pointer to the end of
1783 last_occr = antic_occr;
1784 antic_occr = antic_occr->next;
1788 /* Found another instance of the expression in the same basic block.
1789 Prefer the currently recorded one. We want the first one in the
1790 block and the block is scanned from start to end. */
1791 ; /* nothing to do */
1794 /* First occurrence of this expression in this basic block. */
1795 antic_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1796 bytes_used += sizeof (struct occr);
1797 /* First occurrence of this expression in any block? */
1798 if (cur_expr->antic_occr == NULL)
1799 cur_expr->antic_occr = antic_occr;
1801 last_occr->next = antic_occr;
1803 antic_occr->insn = insn;
1804 antic_occr->next = NULL;
1810 avail_occr = cur_expr->avail_occr;
1812 /* Search for another occurrence in the same basic block. */
1813 while (avail_occr && BLOCK_NUM (avail_occr->insn) != BLOCK_NUM (insn))
1815 /* If an occurrence isn't found, save a pointer to the end of
1817 last_occr = avail_occr;
1818 avail_occr = avail_occr->next;
1822 /* Found another instance of the expression in the same basic block.
1823 Prefer this occurrence to the currently recorded one. We want
1824 the last one in the block and the block is scanned from start
1826 avail_occr->insn = insn;
1829 /* First occurrence of this expression in this basic block. */
1830 avail_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1831 bytes_used += sizeof (struct occr);
1833 /* First occurrence of this expression in any block? */
1834 if (cur_expr->avail_occr == NULL)
1835 cur_expr->avail_occr = avail_occr;
1837 last_occr->next = avail_occr;
1839 avail_occr->insn = insn;
1840 avail_occr->next = NULL;
1845 /* Insert pattern X in INSN in the hash table.
1846 X is a SET of a reg to either another reg or a constant.
1847 If it is already present, record it as the last occurrence in INSN's
1851 insert_set_in_table (x, insn)
1857 struct expr *cur_expr, *last_expr = NULL;
1858 struct occr *cur_occr, *last_occr = NULL;
1860 if (GET_CODE (x) != SET
1861 || GET_CODE (SET_DEST (x)) != REG)
1864 hash = hash_set (REGNO (SET_DEST (x)), set_hash_table_size);
1866 cur_expr = set_hash_table[hash];
1869 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1871 /* If the expression isn't found, save a pointer to the end of
1873 last_expr = cur_expr;
1874 cur_expr = cur_expr->next_same_hash;
1879 cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
1880 bytes_used += sizeof (struct expr);
1881 if (set_hash_table[hash] == NULL)
1882 /* This is the first pattern that hashed to this index. */
1883 set_hash_table[hash] = cur_expr;
1885 /* Add EXPR to end of this hash chain. */
1886 last_expr->next_same_hash = cur_expr;
1888 /* Set the fields of the expr element.
1889 We must copy X because it can be modified when copy propagation is
1890 performed on its operands. */
1891 cur_expr->expr = copy_rtx (x);
1892 cur_expr->bitmap_index = n_sets++;
1893 cur_expr->next_same_hash = NULL;
1894 cur_expr->antic_occr = NULL;
1895 cur_expr->avail_occr = NULL;
1898 /* Now record the occurrence. */
1899 cur_occr = cur_expr->avail_occr;
1901 /* Search for another occurrence in the same basic block. */
1902 while (cur_occr && BLOCK_NUM (cur_occr->insn) != BLOCK_NUM (insn))
1904 /* If an occurrence isn't found, save a pointer to the end of
1906 last_occr = cur_occr;
1907 cur_occr = cur_occr->next;
1911 /* Found another instance of the expression in the same basic block.
1912 Prefer this occurrence to the currently recorded one. We want the
1913 last one in the block and the block is scanned from start to end. */
1914 cur_occr->insn = insn;
1917 /* First occurrence of this expression in this basic block. */
1918 cur_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1919 bytes_used += sizeof (struct occr);
1921 /* First occurrence of this expression in any block? */
1922 if (cur_expr->avail_occr == NULL)
1923 cur_expr->avail_occr = cur_occr;
1925 last_occr->next = cur_occr;
1927 cur_occr->insn = insn;
1928 cur_occr->next = NULL;
1932 /* Scan pattern PAT of INSN and add an entry to the hash table. If SET_P is
1933 non-zero, this is for the assignment hash table, otherwise it is for the
1934 expression hash table. */
1937 hash_scan_set (pat, insn, set_p)
1941 rtx src = SET_SRC (pat);
1942 rtx dest = SET_DEST (pat);
1945 if (GET_CODE (src) == CALL)
1946 hash_scan_call (src, insn);
1948 else if (GET_CODE (dest) == REG)
1950 unsigned int regno = REGNO (dest);
1953 /* If this is a single set and we are doing constant propagation,
1954 see if a REG_NOTE shows this equivalent to a constant. */
1955 if (set_p && (note = find_reg_equal_equiv_note (insn)) != 0
1956 && CONSTANT_P (XEXP (note, 0)))
1957 src = XEXP (note, 0), pat = gen_rtx_SET (VOIDmode, dest, src);
1959 /* Only record sets of pseudo-regs in the hash table. */
1961 && regno >= FIRST_PSEUDO_REGISTER
1962 /* Don't GCSE something if we can't do a reg/reg copy. */
1963 && can_copy_p [GET_MODE (dest)]
1964 /* Is SET_SRC something we want to gcse? */
1965 && want_to_gcse_p (src)
1966 /* Don't CSE a nop. */
1969 /* An expression is not anticipatable if its operands are
1970 modified before this insn. */
1971 int antic_p = oprs_anticipatable_p (src, insn);
1972 /* An expression is not available if its operands are
1973 subsequently modified, including this insn. */
1974 int avail_p = oprs_available_p (src, insn);
1976 insert_expr_in_table (src, GET_MODE (dest), insn, antic_p, avail_p);
1979 /* Record sets for constant/copy propagation. */
1981 && regno >= FIRST_PSEUDO_REGISTER
1982 && ((GET_CODE (src) == REG
1983 && REGNO (src) >= FIRST_PSEUDO_REGISTER
1984 && can_copy_p [GET_MODE (dest)]
1985 && REGNO (src) != regno)
1986 || GET_CODE (src) == CONST_INT
1987 || GET_CODE (src) == SYMBOL_REF
1988 || GET_CODE (src) == CONST_DOUBLE)
1989 /* A copy is not available if its src or dest is subsequently
1990 modified. Here we want to search from INSN+1 on, but
1991 oprs_available_p searches from INSN on. */
1992 && (insn == BLOCK_END (BLOCK_NUM (insn))
1993 || ((tmp = next_nonnote_insn (insn)) != NULL_RTX
1994 && oprs_available_p (pat, tmp))))
1995 insert_set_in_table (pat, insn);
2000 hash_scan_clobber (x, insn)
2001 rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
2003 /* Currently nothing to do. */
2007 hash_scan_call (x, insn)
2008 rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
2010 /* Currently nothing to do. */
2013 /* Process INSN and add hash table entries as appropriate.
2015 Only available expressions that set a single pseudo-reg are recorded.
2017 Single sets in a PARALLEL could be handled, but it's an extra complication
2018 that isn't dealt with right now. The trick is handling the CLOBBERs that
2019 are also in the PARALLEL. Later.
2021 If SET_P is non-zero, this is for the assignment hash table,
2022 otherwise it is for the expression hash table.
2023 If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
2024 not record any expressions. */
2027 hash_scan_insn (insn, set_p, in_libcall_block)
2030 int in_libcall_block;
2032 rtx pat = PATTERN (insn);
2035 if (in_libcall_block)
2038 /* Pick out the sets of INSN and for other forms of instructions record
2039 what's been modified. */
2041 if (GET_CODE (pat) == SET)
2042 hash_scan_set (pat, insn, set_p);
2043 else if (GET_CODE (pat) == PARALLEL)
2044 for (i = 0; i < XVECLEN (pat, 0); i++)
2046 rtx x = XVECEXP (pat, 0, i);
2048 if (GET_CODE (x) == SET)
2049 hash_scan_set (x, insn, set_p);
2050 else if (GET_CODE (x) == CLOBBER)
2051 hash_scan_clobber (x, insn);
2052 else if (GET_CODE (x) == CALL)
2053 hash_scan_call (x, insn);
2056 else if (GET_CODE (pat) == CLOBBER)
2057 hash_scan_clobber (pat, insn);
2058 else if (GET_CODE (pat) == CALL)
2059 hash_scan_call (pat, insn);
2063 dump_hash_table (file, name, table, table_size, total_size)
2066 struct expr **table;
2067 int table_size, total_size;
2070 /* Flattened out table, so it's printed in proper order. */
2071 struct expr **flat_table;
2072 unsigned int *hash_val;
2076 = (struct expr **) xcalloc (total_size, sizeof (struct expr *));
2077 hash_val = (unsigned int *) xmalloc (total_size * sizeof (unsigned int));
2079 for (i = 0; i < table_size; i++)
2080 for (expr = table[i]; expr != NULL; expr = expr->next_same_hash)
2082 flat_table[expr->bitmap_index] = expr;
2083 hash_val[expr->bitmap_index] = i;
2086 fprintf (file, "%s hash table (%d buckets, %d entries)\n",
2087 name, table_size, total_size);
2089 for (i = 0; i < total_size; i++)
2090 if (flat_table[i] != 0)
2092 expr = flat_table[i];
2093 fprintf (file, "Index %d (hash value %d)\n ",
2094 expr->bitmap_index, hash_val[i]);
2095 print_rtl (file, expr->expr);
2096 fprintf (file, "\n");
2099 fprintf (file, "\n");
2105 /* Record register first/last/block set information for REGNO in INSN.
2107 reg_first_set records the first place in the block where the register
2108 is set and is used to compute "anticipatability".
2110 reg_last_set records the last place in the block where the register
2111 is set and is used to compute "availability".
2113 reg_set_in_block records whether the register is set in the block
2114 and is used to compute "transparency". */
2117 record_last_reg_set_info (insn, regno)
2121 if (reg_first_set[regno] == NEVER_SET)
2122 reg_first_set[regno] = INSN_CUID (insn);
2124 reg_last_set[regno] = INSN_CUID (insn);
2125 SET_BIT (reg_set_in_block[BLOCK_NUM (insn)], regno);
2128 /* Record memory first/last/block set information for INSN. */
2131 record_last_mem_set_info (insn)
2134 if (mem_first_set == NEVER_SET)
2135 mem_first_set = INSN_CUID (insn);
2137 mem_last_set = INSN_CUID (insn);
2138 mem_set_in_block[BLOCK_NUM (insn)] = 1;
2141 /* Called from compute_hash_table via note_stores to handle one
2142 SET or CLOBBER in an insn. DATA is really the instruction in which
2143 the SET is taking place. */
2146 record_last_set_info (dest, setter, data)
2147 rtx dest, setter ATTRIBUTE_UNUSED;
2150 rtx last_set_insn = (rtx) data;
2152 if (GET_CODE (dest) == SUBREG)
2153 dest = SUBREG_REG (dest);
2155 if (GET_CODE (dest) == REG)
2156 record_last_reg_set_info (last_set_insn, REGNO (dest));
2157 else if (GET_CODE (dest) == MEM
2158 /* Ignore pushes, they clobber nothing. */
2159 && ! push_operand (dest, GET_MODE (dest)))
2160 record_last_mem_set_info (last_set_insn);
2163 /* Top level function to create an expression or assignment hash table.
2165 Expression entries are placed in the hash table if
2166 - they are of the form (set (pseudo-reg) src),
2167 - src is something we want to perform GCSE on,
2168 - none of the operands are subsequently modified in the block
2170 Assignment entries are placed in the hash table if
2171 - they are of the form (set (pseudo-reg) src),
2172 - src is something we want to perform const/copy propagation on,
2173 - none of the operands or target are subsequently modified in the block
2175 Currently src must be a pseudo-reg or a const_int.
2177 F is the first insn.
2178 SET_P is non-zero for computing the assignment hash table. */
2181 compute_hash_table (set_p)
2186 /* While we compute the hash table we also compute a bit array of which
2187 registers are set in which blocks.
2188 We also compute which blocks set memory, in the absence of aliasing
2189 support [which is TODO].
2190 ??? This isn't needed during const/copy propagation, but it's cheap to
2192 sbitmap_vector_zero (reg_set_in_block, n_basic_blocks);
2193 memset ((char *) mem_set_in_block, 0, n_basic_blocks);
2195 /* Some working arrays used to track first and last set in each block. */
2196 /* ??? One could use alloca here, but at some size a threshold is crossed
2197 beyond which one should use malloc. Are we at that threshold here? */
2198 reg_first_set = (int *) gmalloc (max_gcse_regno * sizeof (int));
2199 reg_last_set = (int *) gmalloc (max_gcse_regno * sizeof (int));
2201 for (bb = 0; bb < n_basic_blocks; bb++)
2205 int in_libcall_block;
2208 /* First pass over the instructions records information used to
2209 determine when registers and memory are first and last set.
2210 ??? The mem_set_in_block and hard-reg reg_set_in_block computation
2211 could be moved to compute_sets since they currently don't change. */
2213 for (i = 0; i < max_gcse_regno; i++)
2214 reg_first_set[i] = reg_last_set[i] = NEVER_SET;
2216 mem_first_set = NEVER_SET;
2217 mem_last_set = NEVER_SET;
2219 for (insn = BLOCK_HEAD (bb);
2220 insn && insn != NEXT_INSN (BLOCK_END (bb));
2221 insn = NEXT_INSN (insn))
2223 #ifdef NON_SAVING_SETJMP
2224 if (NON_SAVING_SETJMP && GET_CODE (insn) == NOTE
2225 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_SETJMP)
2227 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2228 record_last_reg_set_info (insn, regno);
2233 if (! INSN_P (insn))
2236 if (GET_CODE (insn) == CALL_INSN)
2238 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2239 if ((call_used_regs[regno]
2240 && regno != STACK_POINTER_REGNUM
2241 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2242 && regno != HARD_FRAME_POINTER_REGNUM
2244 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
2245 && ! (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
2247 #if !defined (PIC_OFFSET_TABLE_REG_CALL_CLOBBERED)
2248 && ! (regno == PIC_OFFSET_TABLE_REGNUM && flag_pic)
2251 && regno != FRAME_POINTER_REGNUM)
2252 || global_regs[regno])
2253 record_last_reg_set_info (insn, regno);
2255 if (! CONST_CALL_P (insn))
2256 record_last_mem_set_info (insn);
2259 note_stores (PATTERN (insn), record_last_set_info, insn);
2262 /* The next pass builds the hash table. */
2264 for (insn = BLOCK_HEAD (bb), in_libcall_block = 0;
2265 insn && insn != NEXT_INSN (BLOCK_END (bb));
2266 insn = NEXT_INSN (insn))
2269 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
2270 in_libcall_block = 1;
2271 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
2272 in_libcall_block = 0;
2273 hash_scan_insn (insn, set_p, in_libcall_block);
2277 free (reg_first_set);
2278 free (reg_last_set);
2280 /* Catch bugs early. */
2281 reg_first_set = reg_last_set = 0;
2284 /* Allocate space for the set hash table.
2285 N_INSNS is the number of instructions in the function.
2286 It is used to determine the number of buckets to use. */
2289 alloc_set_hash_table (n_insns)
2294 set_hash_table_size = n_insns / 4;
2295 if (set_hash_table_size < 11)
2296 set_hash_table_size = 11;
2298 /* Attempt to maintain efficient use of hash table.
2299 Making it an odd number is simplest for now.
2300 ??? Later take some measurements. */
2301 set_hash_table_size |= 1;
2302 n = set_hash_table_size * sizeof (struct expr *);
2303 set_hash_table = (struct expr **) gmalloc (n);
2306 /* Free things allocated by alloc_set_hash_table. */
2309 free_set_hash_table ()
2311 free (set_hash_table);
2314 /* Compute the hash table for doing copy/const propagation. */
2317 compute_set_hash_table ()
2319 /* Initialize count of number of entries in hash table. */
2321 memset ((char *) set_hash_table, 0,
2322 set_hash_table_size * sizeof (struct expr *));
2324 compute_hash_table (1);
2327 /* Allocate space for the expression hash table.
2328 N_INSNS is the number of instructions in the function.
2329 It is used to determine the number of buckets to use. */
2332 alloc_expr_hash_table (n_insns)
2333 unsigned int n_insns;
2337 expr_hash_table_size = n_insns / 2;
2338 /* Make sure the amount is usable. */
2339 if (expr_hash_table_size < 11)
2340 expr_hash_table_size = 11;
2342 /* Attempt to maintain efficient use of hash table.
2343 Making it an odd number is simplest for now.
2344 ??? Later take some measurements. */
2345 expr_hash_table_size |= 1;
2346 n = expr_hash_table_size * sizeof (struct expr *);
2347 expr_hash_table = (struct expr **) gmalloc (n);
2350 /* Free things allocated by alloc_expr_hash_table. */
2353 free_expr_hash_table ()
2355 free (expr_hash_table);
2358 /* Compute the hash table for doing GCSE. */
2361 compute_expr_hash_table ()
2363 /* Initialize count of number of entries in hash table. */
2365 memset ((char *) expr_hash_table, 0,
2366 expr_hash_table_size * sizeof (struct expr *));
2368 compute_hash_table (0);
2371 /* Expression tracking support. */
2373 /* Lookup pattern PAT in the expression table.
2374 The result is a pointer to the table entry, or NULL if not found. */
2376 static struct expr *
2380 int do_not_record_p;
2381 unsigned int hash = hash_expr (pat, GET_MODE (pat), &do_not_record_p,
2382 expr_hash_table_size);
2385 if (do_not_record_p)
2388 expr = expr_hash_table[hash];
2390 while (expr && ! expr_equiv_p (expr->expr, pat))
2391 expr = expr->next_same_hash;
2396 /* Lookup REGNO in the set table. If PAT is non-NULL look for the entry that
2397 matches it, otherwise return the first entry for REGNO. The result is a
2398 pointer to the table entry, or NULL if not found. */
2400 static struct expr *
2401 lookup_set (regno, pat)
2405 unsigned int hash = hash_set (regno, set_hash_table_size);
2408 expr = set_hash_table[hash];
2412 while (expr && ! expr_equiv_p (expr->expr, pat))
2413 expr = expr->next_same_hash;
2417 while (expr && REGNO (SET_DEST (expr->expr)) != regno)
2418 expr = expr->next_same_hash;
2424 /* Return the next entry for REGNO in list EXPR. */
2426 static struct expr *
2427 next_set (regno, expr)
2432 expr = expr->next_same_hash;
2433 while (expr && REGNO (SET_DEST (expr->expr)) != regno);
2438 /* Reset tables used to keep track of what's still available [since the
2439 start of the block]. */
2442 reset_opr_set_tables ()
2444 /* Maintain a bitmap of which regs have been set since beginning of
2446 sbitmap_zero (reg_set_bitmap);
2448 /* Also keep a record of the last instruction to modify memory.
2449 For now this is very trivial, we only record whether any memory
2450 location has been modified. */
2454 /* Return non-zero if the operands of X are not set before INSN in
2455 INSN's basic block. */
2458 oprs_not_set_p (x, insn)
2468 code = GET_CODE (x);
2483 if (mem_last_set != 0)
2486 return oprs_not_set_p (XEXP (x, 0), insn);
2489 return ! TEST_BIT (reg_set_bitmap, REGNO (x));
2495 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2499 /* If we are about to do the last recursive call
2500 needed at this level, change it into iteration.
2501 This function is called enough to be worth it. */
2503 return oprs_not_set_p (XEXP (x, i), insn);
2505 if (! oprs_not_set_p (XEXP (x, i), insn))
2508 else if (fmt[i] == 'E')
2509 for (j = 0; j < XVECLEN (x, i); j++)
2510 if (! oprs_not_set_p (XVECEXP (x, i, j), insn))
2517 /* Mark things set by a CALL. */
2523 mem_last_set = INSN_CUID (insn);
2526 /* Mark things set by a SET. */
2529 mark_set (pat, insn)
2532 rtx dest = SET_DEST (pat);
2534 while (GET_CODE (dest) == SUBREG
2535 || GET_CODE (dest) == ZERO_EXTRACT
2536 || GET_CODE (dest) == SIGN_EXTRACT
2537 || GET_CODE (dest) == STRICT_LOW_PART)
2538 dest = XEXP (dest, 0);
2540 if (GET_CODE (dest) == REG)
2541 SET_BIT (reg_set_bitmap, REGNO (dest));
2542 else if (GET_CODE (dest) == MEM)
2543 mem_last_set = INSN_CUID (insn);
2545 if (GET_CODE (SET_SRC (pat)) == CALL)
2549 /* Record things set by a CLOBBER. */
2552 mark_clobber (pat, insn)
2555 rtx clob = XEXP (pat, 0);
2557 while (GET_CODE (clob) == SUBREG || GET_CODE (clob) == STRICT_LOW_PART)
2558 clob = XEXP (clob, 0);
2560 if (GET_CODE (clob) == REG)
2561 SET_BIT (reg_set_bitmap, REGNO (clob));
2563 mem_last_set = INSN_CUID (insn);
2566 /* Record things set by INSN.
2567 This data is used by oprs_not_set_p. */
2570 mark_oprs_set (insn)
2573 rtx pat = PATTERN (insn);
2576 if (GET_CODE (pat) == SET)
2577 mark_set (pat, insn);
2578 else if (GET_CODE (pat) == PARALLEL)
2579 for (i = 0; i < XVECLEN (pat, 0); i++)
2581 rtx x = XVECEXP (pat, 0, i);
2583 if (GET_CODE (x) == SET)
2585 else if (GET_CODE (x) == CLOBBER)
2586 mark_clobber (x, insn);
2587 else if (GET_CODE (x) == CALL)
2591 else if (GET_CODE (pat) == CLOBBER)
2592 mark_clobber (pat, insn);
2593 else if (GET_CODE (pat) == CALL)
2598 /* Classic GCSE reaching definition support. */
2600 /* Allocate reaching def variables. */
2603 alloc_rd_mem (n_blocks, n_insns)
2604 int n_blocks, n_insns;
2606 rd_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2607 sbitmap_vector_zero (rd_kill, n_basic_blocks);
2609 rd_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2610 sbitmap_vector_zero (rd_gen, n_basic_blocks);
2612 reaching_defs = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2613 sbitmap_vector_zero (reaching_defs, n_basic_blocks);
2615 rd_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2616 sbitmap_vector_zero (rd_out, n_basic_blocks);
2619 /* Free reaching def variables. */
2626 free (reaching_defs);
2630 /* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
2633 handle_rd_kill_set (insn, regno, bb)
2637 struct reg_set *this_reg;
2639 for (this_reg = reg_set_table[regno]; this_reg; this_reg = this_reg ->next)
2640 if (BLOCK_NUM (this_reg->insn) != BLOCK_NUM (insn))
2641 SET_BIT (rd_kill[bb], INSN_CUID (this_reg->insn));
2644 /* Compute the set of kill's for reaching definitions. */
2654 For each set bit in `gen' of the block (i.e each insn which
2655 generates a definition in the block)
2656 Call the reg set by the insn corresponding to that bit regx
2657 Look at the linked list starting at reg_set_table[regx]
2658 For each setting of regx in the linked list, which is not in
2660 Set the bit in `kill' corresponding to that insn. */
2661 for (bb = 0; bb < n_basic_blocks; bb++)
2662 for (cuid = 0; cuid < max_cuid; cuid++)
2663 if (TEST_BIT (rd_gen[bb], cuid))
2665 rtx insn = CUID_INSN (cuid);
2666 rtx pat = PATTERN (insn);
2668 if (GET_CODE (insn) == CALL_INSN)
2670 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2672 if ((call_used_regs[regno]
2673 && regno != STACK_POINTER_REGNUM
2674 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2675 && regno != HARD_FRAME_POINTER_REGNUM
2677 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
2678 && ! (regno == ARG_POINTER_REGNUM
2679 && fixed_regs[regno])
2681 #if !defined (PIC_OFFSET_TABLE_REG_CALL_CLOBBERED)
2682 && ! (regno == PIC_OFFSET_TABLE_REGNUM && flag_pic)
2684 && regno != FRAME_POINTER_REGNUM)
2685 || global_regs[regno])
2686 handle_rd_kill_set (insn, regno, bb);
2690 if (GET_CODE (pat) == PARALLEL)
2692 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
2694 enum rtx_code code = GET_CODE (XVECEXP (pat, 0, i));
2696 if ((code == SET || code == CLOBBER)
2697 && GET_CODE (XEXP (XVECEXP (pat, 0, i), 0)) == REG)
2698 handle_rd_kill_set (insn,
2699 REGNO (XEXP (XVECEXP (pat, 0, i), 0)),
2703 else if (GET_CODE (pat) == SET && GET_CODE (SET_DEST (pat)) == REG)
2704 /* Each setting of this register outside of this block
2705 must be marked in the set of kills in this block. */
2706 handle_rd_kill_set (insn, REGNO (SET_DEST (pat)), bb);
2710 /* Compute the reaching definitions as in
2711 Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
2712 Chapter 10. It is the same algorithm as used for computing available
2713 expressions but applied to the gens and kills of reaching definitions. */
2718 int bb, changed, passes;
2720 for (bb = 0; bb < n_basic_blocks; bb++)
2721 sbitmap_copy (rd_out[bb] /*dst*/, rd_gen[bb] /*src*/);
2728 for (bb = 0; bb < n_basic_blocks; bb++)
2730 sbitmap_union_of_preds (reaching_defs[bb], rd_out, bb);
2731 changed |= sbitmap_union_of_diff (rd_out[bb], rd_gen[bb],
2732 reaching_defs[bb], rd_kill[bb]);
2738 fprintf (gcse_file, "reaching def computation: %d passes\n", passes);
2741 /* Classic GCSE available expression support. */
2743 /* Allocate memory for available expression computation. */
2746 alloc_avail_expr_mem (n_blocks, n_exprs)
2747 int n_blocks, n_exprs;
2749 ae_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2750 sbitmap_vector_zero (ae_kill, n_basic_blocks);
2752 ae_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2753 sbitmap_vector_zero (ae_gen, n_basic_blocks);
2755 ae_in = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2756 sbitmap_vector_zero (ae_in, n_basic_blocks);
2758 ae_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2759 sbitmap_vector_zero (ae_out, n_basic_blocks);
2763 free_avail_expr_mem ()
2771 /* Compute the set of available expressions generated in each basic block. */
2780 /* For each recorded occurrence of each expression, set ae_gen[bb][expr].
2781 This is all we have to do because an expression is not recorded if it
2782 is not available, and the only expressions we want to work with are the
2783 ones that are recorded. */
2784 for (i = 0; i < expr_hash_table_size; i++)
2785 for (expr = expr_hash_table[i]; expr != 0; expr = expr->next_same_hash)
2786 for (occr = expr->avail_occr; occr != 0; occr = occr->next)
2787 SET_BIT (ae_gen[BLOCK_NUM (occr->insn)], expr->bitmap_index);
2790 /* Return non-zero if expression X is killed in BB. */
2793 expr_killed_p (x, bb)
2804 code = GET_CODE (x);
2808 return TEST_BIT (reg_set_in_block[bb], REGNO (x));
2811 if (mem_set_in_block[bb])
2814 return expr_killed_p (XEXP (x, 0), bb);
2831 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2835 /* If we are about to do the last recursive call
2836 needed at this level, change it into iteration.
2837 This function is called enough to be worth it. */
2839 return expr_killed_p (XEXP (x, i), bb);
2840 else if (expr_killed_p (XEXP (x, i), bb))
2843 else if (fmt[i] == 'E')
2844 for (j = 0; j < XVECLEN (x, i); j++)
2845 if (expr_killed_p (XVECEXP (x, i, j), bb))
2852 /* Compute the set of available expressions killed in each basic block. */
2855 compute_ae_kill (ae_gen, ae_kill)
2856 sbitmap *ae_gen, *ae_kill;
2862 for (bb = 0; bb < n_basic_blocks; bb++)
2863 for (i = 0; i < expr_hash_table_size; i++)
2864 for (expr = expr_hash_table[i]; expr; expr = expr->next_same_hash)
2866 /* Skip EXPR if generated in this block. */
2867 if (TEST_BIT (ae_gen[bb], expr->bitmap_index))
2870 if (expr_killed_p (expr->expr, bb))
2871 SET_BIT (ae_kill[bb], expr->bitmap_index);
2875 /* Actually perform the Classic GCSE optimizations. */
2877 /* Return non-zero if occurrence OCCR of expression EXPR reaches block BB.
2879 CHECK_SELF_LOOP is non-zero if we should consider a block reaching itself
2880 as a positive reach. We want to do this when there are two computations
2881 of the expression in the block.
2883 VISITED is a pointer to a working buffer for tracking which BB's have
2884 been visited. It is NULL for the top-level call.
2886 We treat reaching expressions that go through blocks containing the same
2887 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
2888 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
2889 2 as not reaching. The intent is to improve the probability of finding
2890 only one reaching expression and to reduce register lifetimes by picking
2891 the closest such expression. */
2894 expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited)
2898 int check_self_loop;
2903 for (pred = BASIC_BLOCK(bb)->pred; pred != NULL; pred = pred->pred_next)
2905 int pred_bb = pred->src->index;
2907 if (visited[pred_bb])
2908 /* This predecessor has already been visited. Nothing to do. */
2910 else if (pred_bb == bb)
2912 /* BB loops on itself. */
2914 && TEST_BIT (ae_gen[pred_bb], expr->bitmap_index)
2915 && BLOCK_NUM (occr->insn) == pred_bb)
2918 visited[pred_bb] = 1;
2921 /* Ignore this predecessor if it kills the expression. */
2922 else if (TEST_BIT (ae_kill[pred_bb], expr->bitmap_index))
2923 visited[pred_bb] = 1;
2925 /* Does this predecessor generate this expression? */
2926 else if (TEST_BIT (ae_gen[pred_bb], expr->bitmap_index))
2928 /* Is this the occurrence we're looking for?
2929 Note that there's only one generating occurrence per block
2930 so we just need to check the block number. */
2931 if (BLOCK_NUM (occr->insn) == pred_bb)
2934 visited[pred_bb] = 1;
2937 /* Neither gen nor kill. */
2940 visited[pred_bb] = 1;
2941 if (expr_reaches_here_p_work (occr, expr, pred_bb, check_self_loop,
2948 /* All paths have been checked. */
2952 /* This wrapper for expr_reaches_here_p_work() is to ensure that any
2953 memory allocated for that function is returned. */
2956 expr_reaches_here_p (occr, expr, bb, check_self_loop)
2960 int check_self_loop;
2963 char *visited = (char *) xcalloc (n_basic_blocks, 1);
2965 rval = expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited);
2971 /* Return the instruction that computes EXPR that reaches INSN's basic block.
2972 If there is more than one such instruction, return NULL.
2974 Called only by handle_avail_expr. */
2977 computing_insn (expr, insn)
2981 int bb = BLOCK_NUM (insn);
2983 if (expr->avail_occr->next == NULL)
2985 if (BLOCK_NUM (expr->avail_occr->insn) == bb)
2986 /* The available expression is actually itself
2987 (i.e. a loop in the flow graph) so do nothing. */
2990 /* (FIXME) Case that we found a pattern that was created by
2991 a substitution that took place. */
2992 return expr->avail_occr->insn;
2996 /* Pattern is computed more than once.
2997 Search backwards from this insn to see how many of these
2998 computations actually reach this insn. */
3000 rtx insn_computes_expr = NULL;
3003 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
3005 if (BLOCK_NUM (occr->insn) == bb)
3007 /* The expression is generated in this block.
3008 The only time we care about this is when the expression
3009 is generated later in the block [and thus there's a loop].
3010 We let the normal cse pass handle the other cases. */
3011 if (INSN_CUID (insn) < INSN_CUID (occr->insn)
3012 && expr_reaches_here_p (occr, expr, bb, 1))
3018 insn_computes_expr = occr->insn;
3021 else if (expr_reaches_here_p (occr, expr, bb, 0))
3027 insn_computes_expr = occr->insn;
3031 if (insn_computes_expr == NULL)
3034 return insn_computes_expr;
3038 /* Return non-zero if the definition in DEF_INSN can reach INSN.
3039 Only called by can_disregard_other_sets. */
3042 def_reaches_here_p (insn, def_insn)
3047 if (TEST_BIT (reaching_defs[BLOCK_NUM (insn)], INSN_CUID (def_insn)))
3050 if (BLOCK_NUM (insn) == BLOCK_NUM (def_insn))
3052 if (INSN_CUID (def_insn) < INSN_CUID (insn))
3054 if (GET_CODE (PATTERN (def_insn)) == PARALLEL)
3056 else if (GET_CODE (PATTERN (def_insn)) == CLOBBER)
3057 reg = XEXP (PATTERN (def_insn), 0);
3058 else if (GET_CODE (PATTERN (def_insn)) == SET)
3059 reg = SET_DEST (PATTERN (def_insn));
3063 return ! reg_set_between_p (reg, NEXT_INSN (def_insn), insn);
3072 /* Return non-zero if *ADDR_THIS_REG can only have one value at INSN. The
3073 value returned is the number of definitions that reach INSN. Returning a
3074 value of zero means that [maybe] more than one definition reaches INSN and
3075 the caller can't perform whatever optimization it is trying. i.e. it is
3076 always safe to return zero. */
3079 can_disregard_other_sets (addr_this_reg, insn, for_combine)
3080 struct reg_set **addr_this_reg;
3084 int number_of_reaching_defs = 0;
3085 struct reg_set *this_reg;
3087 for (this_reg = *addr_this_reg; this_reg != 0; this_reg = this_reg->next)
3088 if (def_reaches_here_p (insn, this_reg->insn))
3090 number_of_reaching_defs++;
3091 /* Ignore parallels for now. */
3092 if (GET_CODE (PATTERN (this_reg->insn)) == PARALLEL)
3096 && (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER
3097 || ! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
3098 SET_SRC (PATTERN (insn)))))
3099 /* A setting of the reg to a different value reaches INSN. */
3102 if (number_of_reaching_defs > 1)
3104 /* If in this setting the value the register is being set to is
3105 equal to the previous value the register was set to and this
3106 setting reaches the insn we are trying to do the substitution
3107 on then we are ok. */
3108 if (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER)
3110 else if (! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
3111 SET_SRC (PATTERN (insn))))
3115 *addr_this_reg = this_reg;
3118 return number_of_reaching_defs;
3121 /* Expression computed by insn is available and the substitution is legal,
3122 so try to perform the substitution.
3124 The result is non-zero if any changes were made. */
3127 handle_avail_expr (insn, expr)
3131 rtx pat, insn_computes_expr;
3133 struct reg_set *this_reg;
3134 int found_setting, use_src;
3137 /* We only handle the case where one computation of the expression
3138 reaches this instruction. */
3139 insn_computes_expr = computing_insn (expr, insn);
3140 if (insn_computes_expr == NULL)
3146 /* At this point we know only one computation of EXPR outside of this
3147 block reaches this insn. Now try to find a register that the
3148 expression is computed into. */
3149 if (GET_CODE (SET_SRC (PATTERN (insn_computes_expr))) == REG)
3151 /* This is the case when the available expression that reaches
3152 here has already been handled as an available expression. */
3153 unsigned int regnum_for_replacing
3154 = REGNO (SET_SRC (PATTERN (insn_computes_expr)));
3156 /* If the register was created by GCSE we can't use `reg_set_table',
3157 however we know it's set only once. */
3158 if (regnum_for_replacing >= max_gcse_regno
3159 /* If the register the expression is computed into is set only once,
3160 or only one set reaches this insn, we can use it. */
3161 || (((this_reg = reg_set_table[regnum_for_replacing]),
3162 this_reg->next == NULL)
3163 || can_disregard_other_sets (&this_reg, insn, 0)))
3172 unsigned int regnum_for_replacing
3173 = REGNO (SET_DEST (PATTERN (insn_computes_expr)));
3175 /* This shouldn't happen. */
3176 if (regnum_for_replacing >= max_gcse_regno)
3179 this_reg = reg_set_table[regnum_for_replacing];
3181 /* If the register the expression is computed into is set only once,
3182 or only one set reaches this insn, use it. */
3183 if (this_reg->next == NULL
3184 || can_disregard_other_sets (&this_reg, insn, 0))
3190 pat = PATTERN (insn);
3192 to = SET_SRC (PATTERN (insn_computes_expr));
3194 to = SET_DEST (PATTERN (insn_computes_expr));
3195 changed = validate_change (insn, &SET_SRC (pat), to, 0);
3197 /* We should be able to ignore the return code from validate_change but
3198 to play it safe we check. */
3202 if (gcse_file != NULL)
3204 fprintf (gcse_file, "GCSE: Replacing the source in insn %d with",
3206 fprintf (gcse_file, " reg %d %s insn %d\n",
3207 REGNO (to), use_src ? "from" : "set in",
3208 INSN_UID (insn_computes_expr));
3213 /* The register that the expr is computed into is set more than once. */
3214 else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
3216 /* Insert an insn after insnx that copies the reg set in insnx
3217 into a new pseudo register call this new register REGN.
3218 From insnb until end of basic block or until REGB is set
3219 replace all uses of REGB with REGN. */
3222 to = gen_reg_rtx (GET_MODE (SET_DEST (PATTERN (insn_computes_expr))));
3224 /* Generate the new insn. */
3225 /* ??? If the change fails, we return 0, even though we created
3226 an insn. I think this is ok. */
3228 = emit_insn_after (gen_rtx_SET (VOIDmode, to,
3230 (insn_computes_expr))),
3231 insn_computes_expr);
3233 /* Keep block number table up to date. */
3234 set_block_num (new_insn, BLOCK_NUM (insn_computes_expr));
3236 /* Keep register set table up to date. */
3237 record_one_set (REGNO (to), new_insn);
3239 gcse_create_count++;
3240 if (gcse_file != NULL)
3242 fprintf (gcse_file, "GCSE: Creating insn %d to copy value of reg %d",
3243 INSN_UID (NEXT_INSN (insn_computes_expr)),
3244 REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr)))));
3245 fprintf (gcse_file, ", computed in insn %d,\n",
3246 INSN_UID (insn_computes_expr));
3247 fprintf (gcse_file, " into newly allocated reg %d\n",
3251 pat = PATTERN (insn);
3253 /* Do register replacement for INSN. */
3254 changed = validate_change (insn, &SET_SRC (pat),
3256 (NEXT_INSN (insn_computes_expr))),
3259 /* We should be able to ignore the return code from validate_change but
3260 to play it safe we check. */
3264 if (gcse_file != NULL)
3267 "GCSE: Replacing the source in insn %d with reg %d ",
3269 REGNO (SET_DEST (PATTERN (NEXT_INSN
3270 (insn_computes_expr)))));
3271 fprintf (gcse_file, "set in insn %d\n",
3272 INSN_UID (insn_computes_expr));
3280 /* Perform classic GCSE. This is called by one_classic_gcse_pass after all
3281 the dataflow analysis has been done.
3283 The result is non-zero if a change was made. */
3291 /* Note we start at block 1. */
3294 for (bb = 1; bb < n_basic_blocks; bb++)
3296 /* Reset tables used to keep track of what's still valid [since the
3297 start of the block]. */
3298 reset_opr_set_tables ();
3300 for (insn = BLOCK_HEAD (bb);
3301 insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
3302 insn = NEXT_INSN (insn))
3304 /* Is insn of form (set (pseudo-reg) ...)? */
3305 if (GET_CODE (insn) == INSN
3306 && GET_CODE (PATTERN (insn)) == SET
3307 && GET_CODE (SET_DEST (PATTERN (insn))) == REG
3308 && REGNO (SET_DEST (PATTERN (insn))) >= FIRST_PSEUDO_REGISTER)
3310 rtx pat = PATTERN (insn);
3311 rtx src = SET_SRC (pat);
3314 if (want_to_gcse_p (src)
3315 /* Is the expression recorded? */
3316 && ((expr = lookup_expr (src)) != NULL)
3317 /* Is the expression available [at the start of the
3319 && TEST_BIT (ae_in[bb], expr->bitmap_index)
3320 /* Are the operands unchanged since the start of the
3322 && oprs_not_set_p (src, insn))
3323 changed |= handle_avail_expr (insn, expr);
3326 /* Keep track of everything modified by this insn. */
3327 /* ??? Need to be careful w.r.t. mods done to INSN. */
3329 mark_oprs_set (insn);
3336 /* Top level routine to perform one classic GCSE pass.
3338 Return non-zero if a change was made. */
3341 one_classic_gcse_pass (pass)
3346 gcse_subst_count = 0;
3347 gcse_create_count = 0;
3349 alloc_expr_hash_table (max_cuid);
3350 alloc_rd_mem (n_basic_blocks, max_cuid);
3351 compute_expr_hash_table ();
3353 dump_hash_table (gcse_file, "Expression", expr_hash_table,
3354 expr_hash_table_size, n_exprs);
3360 alloc_avail_expr_mem (n_basic_blocks, n_exprs);
3362 compute_ae_kill (ae_gen, ae_kill);
3363 compute_available (ae_gen, ae_kill, ae_out, ae_in);
3364 changed = classic_gcse ();
3365 free_avail_expr_mem ();
3369 free_expr_hash_table ();
3373 fprintf (gcse_file, "\n");
3374 fprintf (gcse_file, "GCSE of %s, pass %d: %d bytes needed, %d substs,",
3375 current_function_name, pass, bytes_used, gcse_subst_count);
3376 fprintf (gcse_file, "%d insns created\n", gcse_create_count);
3382 /* Compute copy/constant propagation working variables. */
3384 /* Local properties of assignments. */
3385 static sbitmap *cprop_pavloc;
3386 static sbitmap *cprop_absaltered;
3388 /* Global properties of assignments (computed from the local properties). */
3389 static sbitmap *cprop_avin;
3390 static sbitmap *cprop_avout;
3392 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
3393 basic blocks. N_SETS is the number of sets. */
3396 alloc_cprop_mem (n_blocks, n_sets)
3397 int n_blocks, n_sets;
3399 cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
3400 cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
3402 cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
3403 cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
3406 /* Free vars used by copy/const propagation. */
3411 free (cprop_pavloc);
3412 free (cprop_absaltered);
3417 /* For each block, compute whether X is transparent. X is either an
3418 expression or an assignment [though we don't care which, for this context
3419 an assignment is treated as an expression]. For each block where an
3420 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
3424 compute_transp (x, indx, bmap, set_p)
3435 /* repeat is used to turn tail-recursion into iteration since GCC
3436 can't do it when there's no return value. */
3442 code = GET_CODE (x);
3448 if (REGNO (x) < FIRST_PSEUDO_REGISTER)
3450 for (bb = 0; bb < n_basic_blocks; bb++)
3451 if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
3452 SET_BIT (bmap[bb], indx);
3456 for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
3457 SET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
3462 if (REGNO (x) < FIRST_PSEUDO_REGISTER)
3464 for (bb = 0; bb < n_basic_blocks; bb++)
3465 if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
3466 RESET_BIT (bmap[bb], indx);
3470 for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
3471 RESET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
3480 for (bb = 0; bb < n_basic_blocks; bb++)
3481 if (mem_set_in_block[bb])
3482 SET_BIT (bmap[bb], indx);
3486 for (bb = 0; bb < n_basic_blocks; bb++)
3487 if (mem_set_in_block[bb])
3488 RESET_BIT (bmap[bb], indx);
3509 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3513 /* If we are about to do the last recursive call
3514 needed at this level, change it into iteration.
3515 This function is called enough to be worth it. */
3522 compute_transp (XEXP (x, i), indx, bmap, set_p);
3524 else if (fmt[i] == 'E')
3525 for (j = 0; j < XVECLEN (x, i); j++)
3526 compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
3530 /* Top level routine to do the dataflow analysis needed by copy/const
3534 compute_cprop_data ()
3536 compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, 1);
3537 compute_available (cprop_pavloc, cprop_absaltered,
3538 cprop_avout, cprop_avin);
3541 /* Copy/constant propagation. */
3543 /* Maximum number of register uses in an insn that we handle. */
3546 /* Table of uses found in an insn.
3547 Allocated statically to avoid alloc/free complexity and overhead. */
3548 static struct reg_use reg_use_table[MAX_USES];
3550 /* Index into `reg_use_table' while building it. */
3551 static int reg_use_count;
3553 /* Set up a list of register numbers used in INSN. The found uses are stored
3554 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
3555 and contains the number of uses in the table upon exit.
3557 ??? If a register appears multiple times we will record it multiple times.
3558 This doesn't hurt anything but it will slow things down. */
3568 /* repeat is used to turn tail-recursion into iteration since GCC
3569 can't do it when there's no return value. */
3575 code = GET_CODE (x);
3579 if (reg_use_count == MAX_USES)
3582 reg_use_table[reg_use_count].reg_rtx = x;
3600 case ASM_INPUT: /*FIXME*/
3604 if (GET_CODE (SET_DEST (x)) == MEM)
3605 find_used_regs (SET_DEST (x));
3613 /* Recursively scan the operands of this expression. */
3615 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3619 /* If we are about to do the last recursive call
3620 needed at this level, change it into iteration.
3621 This function is called enough to be worth it. */
3628 find_used_regs (XEXP (x, i));
3630 else if (fmt[i] == 'E')
3631 for (j = 0; j < XVECLEN (x, i); j++)
3632 find_used_regs (XVECEXP (x, i, j));
3636 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
3637 Returns non-zero is successful. */
3640 try_replace_reg (from, to, insn)
3643 rtx note = find_reg_equal_equiv_note (insn);
3646 rtx set = single_set (insn);
3648 /* If this is a single set, try to simplify the source of the set given
3649 our substitution. We could perhaps try this for multiple SETs, but
3650 it probably won't buy us anything. */
3653 src = simplify_replace_rtx (SET_SRC (set), from, to);
3655 /* Try this two ways: first just replace SET_SRC. If that doesn't
3656 work and this is a PARALLEL, try to replace the whole pattern
3658 if (validate_change (insn, &SET_SRC (set), src, 0))
3660 else if (GET_CODE (PATTERN (insn)) == PARALLEL
3661 && validate_change (insn, &PATTERN (insn),
3662 gen_rtx_SET (VOIDmode, SET_DEST (set),
3668 /* Otherwise, try to do a global replacement within the insn. */
3670 success = validate_replace_src (from, to, insn);
3672 /* If we've failed to do replacement, have a single SET, and don't already
3673 have a note, add a REG_EQUAL note to not lose information. */
3674 if (!success && note == 0 && set != 0)
3675 note = REG_NOTES (insn)
3676 = gen_rtx_EXPR_LIST (REG_EQUAL, src, REG_NOTES (insn));
3678 /* If there is already a NOTE, update the expression in it with our
3681 XEXP (note, 0) = simplify_replace_rtx (XEXP (note, 0), from, to);
3683 /* REG_EQUAL may get simplified into register.
3684 We don't allow that. Remove that note. This code ought
3685 not to hapen, because previous code ought to syntetize
3686 reg-reg move, but be on the safe side. */
3687 if (note && REG_P (XEXP (note, 0)))
3688 remove_note (insn, note);
3693 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
3694 NULL no such set is found. */
3696 static struct expr *
3697 find_avail_set (regno, insn)
3701 /* SET1 contains the last set found that can be returned to the caller for
3702 use in a substitution. */
3703 struct expr *set1 = 0;
3705 /* Loops are not possible here. To get a loop we would need two sets
3706 available at the start of the block containing INSN. ie we would
3707 need two sets like this available at the start of the block:
3709 (set (reg X) (reg Y))
3710 (set (reg Y) (reg X))
3712 This can not happen since the set of (reg Y) would have killed the
3713 set of (reg X) making it unavailable at the start of this block. */
3717 struct expr *set = lookup_set (regno, NULL_RTX);
3719 /* Find a set that is available at the start of the block
3720 which contains INSN. */
3723 if (TEST_BIT (cprop_avin[BLOCK_NUM (insn)], set->bitmap_index))
3725 set = next_set (regno, set);
3728 /* If no available set was found we've reached the end of the
3729 (possibly empty) copy chain. */
3733 if (GET_CODE (set->expr) != SET)
3736 src = SET_SRC (set->expr);
3738 /* We know the set is available.
3739 Now check that SRC is ANTLOC (i.e. none of the source operands
3740 have changed since the start of the block).
3742 If the source operand changed, we may still use it for the next
3743 iteration of this loop, but we may not use it for substitutions. */
3745 if (CONSTANT_P (src) || oprs_not_set_p (src, insn))
3748 /* If the source of the set is anything except a register, then
3749 we have reached the end of the copy chain. */
3750 if (GET_CODE (src) != REG)
3753 /* Follow the copy chain, ie start another iteration of the loop
3754 and see if we have an available copy into SRC. */
3755 regno = REGNO (src);
3758 /* SET1 holds the last set that was available and anticipatable at
3763 /* Subroutine of cprop_insn that tries to propagate constants into
3764 JUMP_INSNS. INSN must be a conditional jump. FROM is what we will try to
3765 replace, SRC is the constant we will try to substitute for it. Returns
3766 nonzero if a change was made. We know INSN has just a SET. */
3769 cprop_jump (insn, from, src)
3774 rtx set = PATTERN (insn);
3775 rtx new = simplify_replace_rtx (SET_SRC (set), from, src);
3777 /* If no simplification can be made, then try the next
3779 if (rtx_equal_p (new, SET_SRC (set)))
3782 /* If this is now a no-op leave it that way, but update LABEL_NUSED if
3786 SET_SRC (set) = new;
3788 if (JUMP_LABEL (insn) != 0)
3789 --LABEL_NUSES (JUMP_LABEL (insn));
3792 /* Otherwise, this must be a valid instruction. */
3793 else if (! validate_change (insn, &SET_SRC (set), new, 0))
3796 /* If this has turned into an unconditional jump,
3797 then put a barrier after it so that the unreachable
3798 code will be deleted. */
3799 if (GET_CODE (SET_SRC (set)) == LABEL_REF)
3800 emit_barrier_after (insn);
3802 run_jump_opt_after_gcse = 1;
3805 if (gcse_file != NULL)
3808 "CONST-PROP: Replacing reg %d in insn %d with constant ",
3809 REGNO (from), INSN_UID (insn));
3810 print_rtl (gcse_file, src);
3811 fprintf (gcse_file, "\n");
3819 /* Subroutine of cprop_insn that tries to propagate constants into JUMP_INSNS
3820 for machines that have CC0. INSN is a single set that stores into CC0;
3821 the insn following it is a conditional jump. REG_USED is the use we will
3822 try to replace, SRC is the constant we will try to substitute for it.
3823 Returns nonzero if a change was made. */
3826 cprop_cc0_jump (insn, reg_used, src)
3828 struct reg_use *reg_used;
3831 /* First substitute in the SET_SRC of INSN, then substitute that for
3833 rtx jump = NEXT_INSN (insn);
3834 rtx new_src = simplify_replace_rtx (SET_SRC (PATTERN (insn)),
3835 reg_used->reg_rtx, src);
3837 if (! cprop_jump (jump, cc0_rtx, new_src))
3840 /* If we succeeded, delete the cc0 setter. */
3841 PUT_CODE (insn, NOTE);
3842 NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
3843 NOTE_SOURCE_FILE (insn) = 0;
3849 /* Perform constant and copy propagation on INSN.
3850 The result is non-zero if a change was made. */
3853 cprop_insn (insn, alter_jumps)
3857 struct reg_use *reg_used;
3861 /* Only propagate into SETs. Note that a conditional jump is a
3862 SET with pc_rtx as the destination. */
3863 if (GET_CODE (insn) != INSN && GET_CODE (insn) != JUMP_INSN)
3867 find_used_regs (PATTERN (insn));
3869 note = find_reg_equal_equiv_note (insn);
3871 /* We may win even when propagating constants into notes. */
3873 find_used_regs (XEXP (note, 0));
3875 for (reg_used = ®_use_table[0]; reg_use_count > 0;
3876 reg_used++, reg_use_count--)
3878 unsigned int regno = REGNO (reg_used->reg_rtx);
3882 /* Ignore registers created by GCSE.
3883 We do this because ... */
3884 if (regno >= max_gcse_regno)
3887 /* If the register has already been set in this block, there's
3888 nothing we can do. */
3889 if (! oprs_not_set_p (reg_used->reg_rtx, insn))
3892 /* Find an assignment that sets reg_used and is available
3893 at the start of the block. */
3894 set = find_avail_set (regno, insn);
3899 /* ??? We might be able to handle PARALLELs. Later. */
3900 if (GET_CODE (pat) != SET)
3903 src = SET_SRC (pat);
3905 /* Constant propagation. */
3906 if (GET_CODE (src) == CONST_INT || GET_CODE (src) == CONST_DOUBLE
3907 || GET_CODE (src) == SYMBOL_REF)
3909 /* Handle normal insns first. */
3910 if (GET_CODE (insn) == INSN
3911 && try_replace_reg (reg_used->reg_rtx, src, insn))
3915 if (gcse_file != NULL)
3917 fprintf (gcse_file, "CONST-PROP: Replacing reg %d in ",
3919 fprintf (gcse_file, "insn %d with constant ",
3921 print_rtl (gcse_file, src);
3922 fprintf (gcse_file, "\n");
3925 /* The original insn setting reg_used may or may not now be
3926 deletable. We leave the deletion to flow. */
3929 /* Try to propagate a CONST_INT into a conditional jump.
3930 We're pretty specific about what we will handle in this
3931 code, we can extend this as necessary over time.
3933 Right now the insn in question must look like
3934 (set (pc) (if_then_else ...)) */
3935 else if (alter_jumps
3936 && GET_CODE (insn) == JUMP_INSN
3937 && condjump_p (insn)
3938 && ! simplejump_p (insn))
3939 changed |= cprop_jump (insn, reg_used->reg_rtx, src);
3942 /* Similar code for machines that use a pair of CC0 setter and
3943 conditional jump insn. */
3944 else if (alter_jumps
3945 && GET_CODE (PATTERN (insn)) == SET
3946 && SET_DEST (PATTERN (insn)) == cc0_rtx
3947 && GET_CODE (NEXT_INSN (insn)) == JUMP_INSN
3948 && condjump_p (NEXT_INSN (insn))
3949 && ! simplejump_p (NEXT_INSN (insn))
3950 && cprop_cc0_jump (insn, reg_used, src))
3957 else if (GET_CODE (src) == REG
3958 && REGNO (src) >= FIRST_PSEUDO_REGISTER
3959 && REGNO (src) != regno)
3961 if (try_replace_reg (reg_used->reg_rtx, src, insn))
3965 if (gcse_file != NULL)
3967 fprintf (gcse_file, "COPY-PROP: Replacing reg %d in insn %d",
3968 regno, INSN_UID (insn));
3969 fprintf (gcse_file, " with reg %d\n", REGNO (src));
3972 /* The original insn setting reg_used may or may not now be
3973 deletable. We leave the deletion to flow. */
3974 /* FIXME: If it turns out that the insn isn't deletable,
3975 then we may have unnecessarily extended register lifetimes
3976 and made things worse. */
3984 /* Forward propagate copies. This includes copies and constants. Return
3985 non-zero if a change was made. */
3994 /* Note we start at block 1. */
3997 for (bb = 1; bb < n_basic_blocks; bb++)
3999 /* Reset tables used to keep track of what's still valid [since the
4000 start of the block]. */
4001 reset_opr_set_tables ();
4003 for (insn = BLOCK_HEAD (bb);
4004 insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
4005 insn = NEXT_INSN (insn))
4008 changed |= cprop_insn (insn, alter_jumps);
4010 /* Keep track of everything modified by this insn. */
4011 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
4012 call mark_oprs_set if we turned the insn into a NOTE. */
4013 if (GET_CODE (insn) != NOTE)
4014 mark_oprs_set (insn);
4018 if (gcse_file != NULL)
4019 fprintf (gcse_file, "\n");
4024 /* Perform one copy/constant propagation pass.
4025 F is the first insn in the function.
4026 PASS is the pass count. */
4029 one_cprop_pass (pass, alter_jumps)
4035 const_prop_count = 0;
4036 copy_prop_count = 0;
4038 alloc_set_hash_table (max_cuid);
4039 compute_set_hash_table ();
4041 dump_hash_table (gcse_file, "SET", set_hash_table, set_hash_table_size,
4045 alloc_cprop_mem (n_basic_blocks, n_sets);
4046 compute_cprop_data ();
4047 changed = cprop (alter_jumps);
4051 free_set_hash_table ();
4055 fprintf (gcse_file, "CPROP of %s, pass %d: %d bytes needed, ",
4056 current_function_name, pass, bytes_used);
4057 fprintf (gcse_file, "%d const props, %d copy props\n\n",
4058 const_prop_count, copy_prop_count);
4064 /* Compute PRE+LCM working variables. */
4066 /* Local properties of expressions. */
4067 /* Nonzero for expressions that are transparent in the block. */
4068 static sbitmap *transp;
4070 /* Nonzero for expressions that are transparent at the end of the block.
4071 This is only zero for expressions killed by abnormal critical edge
4072 created by a calls. */
4073 static sbitmap *transpout;
4075 /* Nonzero for expressions that are computed (available) in the block. */
4076 static sbitmap *comp;
4078 /* Nonzero for expressions that are locally anticipatable in the block. */
4079 static sbitmap *antloc;
4081 /* Nonzero for expressions where this block is an optimal computation
4083 static sbitmap *pre_optimal;
4085 /* Nonzero for expressions which are redundant in a particular block. */
4086 static sbitmap *pre_redundant;
4088 /* Nonzero for expressions which should be inserted on a specific edge. */
4089 static sbitmap *pre_insert_map;
4091 /* Nonzero for expressions which should be deleted in a specific block. */
4092 static sbitmap *pre_delete_map;
4094 /* Contains the edge_list returned by pre_edge_lcm. */
4095 static struct edge_list *edge_list;
4097 /* Redundant insns. */
4098 static sbitmap pre_redundant_insns;
4100 /* Allocate vars used for PRE analysis. */
4103 alloc_pre_mem (n_blocks, n_exprs)
4104 int n_blocks, n_exprs;
4106 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
4107 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
4108 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
4111 pre_redundant = NULL;
4112 pre_insert_map = NULL;
4113 pre_delete_map = NULL;
4116 ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
4118 /* pre_insert and pre_delete are allocated later. */
4121 /* Free vars used for PRE analysis. */
4129 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
4134 free (pre_redundant);
4136 free (pre_insert_map);
4138 free (pre_delete_map);
4145 transp = comp = NULL;
4146 pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
4147 ae_in = ae_out = NULL;
4150 /* Top level routine to do the dataflow analysis needed by PRE. */
4155 sbitmap trapping_expr;
4159 compute_local_properties (transp, comp, antloc, 0);
4160 sbitmap_vector_zero (ae_kill, n_basic_blocks);
4162 /* Collect expressions which might trap. */
4163 trapping_expr = sbitmap_alloc (n_exprs);
4164 sbitmap_zero (trapping_expr);
4165 for (ui = 0; ui < expr_hash_table_size; ui++)
4168 for (e = expr_hash_table[ui]; e != NULL; e = e->next_same_hash)
4169 if (may_trap_p (e->expr))
4170 SET_BIT (trapping_expr, e->bitmap_index);
4173 /* Compute ae_kill for each basic block using:
4177 This is significantly faster than compute_ae_kill. */
4179 for (i = 0; i < n_basic_blocks; i++)
4183 /* If the current block is the destination of an abnormal edge, we
4184 kill all trapping expressions because we won't be able to properly
4185 place the instruction on the edge. So make them neither
4186 anticipatable nor transparent. This is fairly conservative. */
4187 for (e = BASIC_BLOCK (i)->pred; e ; e = e->pred_next)
4188 if (e->flags & EDGE_ABNORMAL)
4190 sbitmap_difference (antloc[i], antloc[i], trapping_expr);
4191 sbitmap_difference (transp[i], transp[i], trapping_expr);
4195 sbitmap_a_or_b (ae_kill[i], transp[i], comp[i]);
4196 sbitmap_not (ae_kill[i], ae_kill[i]);
4199 edge_list = pre_edge_lcm (gcse_file, n_exprs, transp, comp, antloc,
4200 ae_kill, &pre_insert_map, &pre_delete_map);
4205 free (trapping_expr);
4210 /* Return non-zero if an occurrence of expression EXPR in OCCR_BB would reach
4213 VISITED is a pointer to a working buffer for tracking which BB's have
4214 been visited. It is NULL for the top-level call.
4216 We treat reaching expressions that go through blocks containing the same
4217 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
4218 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
4219 2 as not reaching. The intent is to improve the probability of finding
4220 only one reaching expression and to reduce register lifetimes by picking
4221 the closest such expression. */
4224 pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited)
4232 for (pred = BASIC_BLOCK (bb)->pred; pred != NULL; pred = pred->pred_next)
4234 int pred_bb = pred->src->index;
4236 if (pred->src == ENTRY_BLOCK_PTR
4237 /* Has predecessor has already been visited? */
4238 || visited[pred_bb])
4239 ;/* Nothing to do. */
4241 /* Does this predecessor generate this expression? */
4242 else if (TEST_BIT (comp[pred_bb], expr->bitmap_index))
4244 /* Is this the occurrence we're looking for?
4245 Note that there's only one generating occurrence per block
4246 so we just need to check the block number. */
4247 if (occr_bb == pred_bb)
4250 visited[pred_bb] = 1;
4252 /* Ignore this predecessor if it kills the expression. */
4253 else if (! TEST_BIT (transp[pred_bb], expr->bitmap_index))
4254 visited[pred_bb] = 1;
4256 /* Neither gen nor kill. */
4259 visited[pred_bb] = 1;
4260 if (pre_expr_reaches_here_p_work (occr_bb, expr, pred_bb, visited))
4265 /* All paths have been checked. */
4269 /* The wrapper for pre_expr_reaches_here_work that ensures that any
4270 memory allocated for that function is returned. */
4273 pre_expr_reaches_here_p (occr_bb, expr, bb)
4279 char *visited = (char *) xcalloc (n_basic_blocks, 1);
4281 rval = pre_expr_reaches_here_p_work(occr_bb, expr, bb, visited);
4288 /* Given an expr, generate RTL which we can insert at the end of a BB,
4289 or on an edge. Set the block number of any insns generated to
4293 process_insert_insn (expr)
4296 rtx reg = expr->reaching_reg;
4297 rtx exp = copy_rtx (expr->expr);
4302 /* If the expression is something that's an operand, like a constant,
4303 just copy it to a register. */
4304 if (general_operand (exp, GET_MODE (reg)))
4305 emit_move_insn (reg, exp);
4307 /* Otherwise, make a new insn to compute this expression and make sure the
4308 insn will be recognized (this also adds any needed CLOBBERs). Copy the
4309 expression to make sure we don't have any sharing issues. */
4310 else if (insn_invalid_p (emit_insn (gen_rtx_SET (VOIDmode, reg, exp))))
4313 pat = gen_sequence ();
4319 /* Add EXPR to the end of basic block BB.
4321 This is used by both the PRE and code hoisting.
4323 For PRE, we want to verify that the expr is either transparent
4324 or locally anticipatable in the target block. This check makes
4325 no sense for code hoisting. */
4328 insert_insn_end_bb (expr, bb, pre)
4333 rtx insn = BLOCK_END (bb);
4335 rtx reg = expr->reaching_reg;
4336 int regno = REGNO (reg);
4340 pat = process_insert_insn (expr);
4342 /* If the last insn is a jump, insert EXPR in front [taking care to
4343 handle cc0, etc. properly]. */
4345 if (GET_CODE (insn) == JUMP_INSN)
4351 /* If this is a jump table, then we can't insert stuff here. Since
4352 we know the previous real insn must be the tablejump, we insert
4353 the new instruction just before the tablejump. */
4354 if (GET_CODE (PATTERN (insn)) == ADDR_VEC
4355 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
4356 insn = prev_real_insn (insn);
4359 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
4360 if cc0 isn't set. */
4361 note = find_reg_note (insn, REG_CC_SETTER, NULL_RTX);
4363 insn = XEXP (note, 0);
4366 rtx maybe_cc0_setter = prev_nonnote_insn (insn);
4367 if (maybe_cc0_setter
4368 && INSN_P (maybe_cc0_setter)
4369 && sets_cc0_p (PATTERN (maybe_cc0_setter)))
4370 insn = maybe_cc0_setter;
4373 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4374 new_insn = emit_block_insn_before (pat, insn, BASIC_BLOCK (bb));
4377 /* Likewise if the last insn is a call, as will happen in the presence
4378 of exception handling. */
4379 else if (GET_CODE (insn) == CALL_INSN)
4381 HARD_REG_SET parm_regs;
4385 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4386 we search backward and place the instructions before the first
4387 parameter is loaded. Do this for everyone for consistency and a
4388 presumtion that we'll get better code elsewhere as well.
4390 It should always be the case that we can put these instructions
4391 anywhere in the basic block with performing PRE optimizations.
4395 && !TEST_BIT (antloc[bb], expr->bitmap_index)
4396 && !TEST_BIT (transp[bb], expr->bitmap_index))
4399 /* Since different machines initialize their parameter registers
4400 in different orders, assume nothing. Collect the set of all
4401 parameter registers. */
4402 CLEAR_HARD_REG_SET (parm_regs);
4404 for (p = CALL_INSN_FUNCTION_USAGE (insn); p ; p = XEXP (p, 1))
4405 if (GET_CODE (XEXP (p, 0)) == USE
4406 && GET_CODE (XEXP (XEXP (p, 0), 0)) == REG)
4408 if (REGNO (XEXP (XEXP (p, 0), 0)) >= FIRST_PSEUDO_REGISTER)
4411 SET_HARD_REG_BIT (parm_regs, REGNO (XEXP (XEXP (p, 0), 0)));
4415 /* Search backward for the first set of a register in this set. */
4416 while (nparm_regs && BLOCK_HEAD (bb) != insn)
4418 insn = PREV_INSN (insn);
4419 p = single_set (insn);
4420 if (p && GET_CODE (SET_DEST (p)) == REG
4421 && REGNO (SET_DEST (p)) < FIRST_PSEUDO_REGISTER
4422 && TEST_HARD_REG_BIT (parm_regs, REGNO (SET_DEST (p))))
4424 CLEAR_HARD_REG_BIT (parm_regs, REGNO (SET_DEST (p)));
4429 /* If we found all the parameter loads, then we want to insert
4430 before the first parameter load.
4432 If we did not find all the parameter loads, then we might have
4433 stopped on the head of the block, which could be a CODE_LABEL.
4434 If we inserted before the CODE_LABEL, then we would be putting
4435 the insn in the wrong basic block. In that case, put the insn
4436 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4437 while (GET_CODE (insn) == CODE_LABEL
4438 || NOTE_INSN_BASIC_BLOCK_P (insn))
4439 insn = NEXT_INSN (insn);
4441 new_insn = emit_block_insn_before (pat, insn, BASIC_BLOCK (bb));
4445 new_insn = emit_insn_after (pat, insn);
4446 BLOCK_END (bb) = new_insn;
4449 /* Keep block number table up to date.
4450 Note, PAT could be a multiple insn sequence, we have to make
4451 sure that each insn in the sequence is handled. */
4452 if (GET_CODE (pat) == SEQUENCE)
4454 for (i = 0; i < XVECLEN (pat, 0); i++)
4456 rtx insn = XVECEXP (pat, 0, i);
4458 set_block_num (insn, bb);
4460 add_label_notes (PATTERN (insn), new_insn);
4462 note_stores (PATTERN (insn), record_set_info, insn);
4467 add_label_notes (SET_SRC (pat), new_insn);
4468 set_block_num (new_insn, bb);
4470 /* Keep register set table up to date. */
4471 record_one_set (regno, new_insn);
4474 gcse_create_count++;
4478 fprintf (gcse_file, "PRE/HOIST: end of bb %d, insn %d, ",
4479 bb, INSN_UID (new_insn));
4480 fprintf (gcse_file, "copying expression %d to reg %d\n",
4481 expr->bitmap_index, regno);
4485 /* Insert partially redundant expressions on edges in the CFG to make
4486 the expressions fully redundant. */
4489 pre_edge_insert (edge_list, index_map)
4490 struct edge_list *edge_list;
4491 struct expr **index_map;
4493 int e, i, j, num_edges, set_size, did_insert = 0;
4496 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4497 if it reaches any of the deleted expressions. */
4499 set_size = pre_insert_map[0]->size;
4500 num_edges = NUM_EDGES (edge_list);
4501 inserted = sbitmap_vector_alloc (num_edges, n_exprs);
4502 sbitmap_vector_zero (inserted, num_edges);
4504 for (e = 0; e < num_edges; e++)
4507 basic_block pred = INDEX_EDGE_PRED_BB (edge_list, e);
4508 int bb = pred->index;
4510 for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
4512 SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
4514 for (j = indx; insert && j < n_exprs; j++, insert >>= 1)
4515 if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
4517 struct expr *expr = index_map[j];
4520 /* Now look at each deleted occurence of this expression. */
4521 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4523 if (! occr->deleted_p)
4526 /* Insert this expression on this edge if if it would
4527 reach the deleted occurence in BB. */
4528 if (!TEST_BIT (inserted[e], j))
4531 edge eg = INDEX_EDGE (edge_list, e);
4533 /* We can't insert anything on an abnormal and
4534 critical edge, so we insert the insn at the end of
4535 the previous block. There are several alternatives
4536 detailed in Morgans book P277 (sec 10.5) for
4537 handling this situation. This one is easiest for
4540 if ((eg->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
4541 insert_insn_end_bb (index_map[j], bb, 0);
4544 insn = process_insert_insn (index_map[j]);
4545 insert_insn_on_edge (insn, eg);
4550 fprintf (gcse_file, "PRE/HOIST: edge (%d,%d), ",
4552 INDEX_EDGE_SUCC_BB (edge_list, e)->index);
4553 fprintf (gcse_file, "copy expression %d\n",
4554 expr->bitmap_index);
4557 SET_BIT (inserted[e], j);
4559 gcse_create_count++;
4570 /* Copy the result of INSN to REG. INDX is the expression number. */
4573 pre_insert_copy_insn (expr, insn)
4577 rtx reg = expr->reaching_reg;
4578 int regno = REGNO (reg);
4579 int indx = expr->bitmap_index;
4580 rtx set = single_set (insn);
4582 int bb = BLOCK_NUM (insn);
4587 new_insn = emit_insn_after (gen_rtx_SET (VOIDmode, reg, SET_DEST (set)),
4590 /* Keep block number table up to date. */
4591 set_block_num (new_insn, bb);
4593 /* Keep register set table up to date. */
4594 record_one_set (regno, new_insn);
4595 if (insn == BLOCK_END (bb))
4596 BLOCK_END (bb) = new_insn;
4598 gcse_create_count++;
4602 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4603 BLOCK_NUM (insn), INSN_UID (new_insn), indx,
4604 INSN_UID (insn), regno);
4607 /* Copy available expressions that reach the redundant expression
4608 to `reaching_reg'. */
4611 pre_insert_copies ()
4618 /* For each available expression in the table, copy the result to
4619 `reaching_reg' if the expression reaches a deleted one.
4621 ??? The current algorithm is rather brute force.
4622 Need to do some profiling. */
4624 for (i = 0; i < expr_hash_table_size; i++)
4625 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4627 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4628 we don't want to insert a copy here because the expression may not
4629 really be redundant. So only insert an insn if the expression was
4630 deleted. This test also avoids further processing if the
4631 expression wasn't deleted anywhere. */
4632 if (expr->reaching_reg == NULL)
4635 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4637 if (! occr->deleted_p)
4640 for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
4642 rtx insn = avail->insn;
4644 /* No need to handle this one if handled already. */
4645 if (avail->copied_p)
4648 /* Don't handle this one if it's a redundant one. */
4649 if (TEST_BIT (pre_redundant_insns, INSN_CUID (insn)))
4652 /* Or if the expression doesn't reach the deleted one. */
4653 if (! pre_expr_reaches_here_p (BLOCK_NUM (avail->insn), expr,
4654 BLOCK_NUM (occr->insn)))
4657 /* Copy the result of avail to reaching_reg. */
4658 pre_insert_copy_insn (expr, insn);
4659 avail->copied_p = 1;
4665 /* Delete redundant computations.
4666 Deletion is done by changing the insn to copy the `reaching_reg' of
4667 the expression into the result of the SET. It is left to later passes
4668 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4670 Returns non-zero if a change is made. */
4681 for (i = 0; i < expr_hash_table_size; i++)
4682 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4684 int indx = expr->bitmap_index;
4686 /* We only need to search antic_occr since we require
4689 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4691 rtx insn = occr->insn;
4693 int bb = BLOCK_NUM (insn);
4695 if (TEST_BIT (pre_delete_map[bb], indx))
4697 set = single_set (insn);
4701 /* Create a pseudo-reg to store the result of reaching
4702 expressions into. Get the mode for the new pseudo from
4703 the mode of the original destination pseudo. */
4704 if (expr->reaching_reg == NULL)
4706 = gen_reg_rtx (GET_MODE (SET_DEST (set)));
4708 /* In theory this should never fail since we're creating
4711 However, on the x86 some of the movXX patterns actually
4712 contain clobbers of scratch regs. This may cause the
4713 insn created by validate_change to not match any pattern
4714 and thus cause validate_change to fail. */
4715 if (validate_change (insn, &SET_SRC (set),
4716 expr->reaching_reg, 0))
4718 occr->deleted_p = 1;
4719 SET_BIT (pre_redundant_insns, INSN_CUID (insn));
4727 "PRE: redundant insn %d (expression %d) in ",
4728 INSN_UID (insn), indx);
4729 fprintf (gcse_file, "bb %d, reaching reg is %d\n",
4730 bb, REGNO (expr->reaching_reg));
4739 /* Perform GCSE optimizations using PRE.
4740 This is called by one_pre_gcse_pass after all the dataflow analysis
4743 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4744 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4745 Compiler Design and Implementation.
4747 ??? A new pseudo reg is created to hold the reaching expression. The nice
4748 thing about the classical approach is that it would try to use an existing
4749 reg. If the register can't be adequately optimized [i.e. we introduce
4750 reload problems], one could add a pass here to propagate the new register
4753 ??? We don't handle single sets in PARALLELs because we're [currently] not
4754 able to copy the rest of the parallel when we insert copies to create full
4755 redundancies from partial redundancies. However, there's no reason why we
4756 can't handle PARALLELs in the cases where there are no partial
4763 int did_insert, changed;
4764 struct expr **index_map;
4767 /* Compute a mapping from expression number (`bitmap_index') to
4768 hash table entry. */
4770 index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
4771 for (i = 0; i < expr_hash_table_size; i++)
4772 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4773 index_map[expr->bitmap_index] = expr;
4775 /* Reset bitmap used to track which insns are redundant. */
4776 pre_redundant_insns = sbitmap_alloc (max_cuid);
4777 sbitmap_zero (pre_redundant_insns);
4779 /* Delete the redundant insns first so that
4780 - we know what register to use for the new insns and for the other
4781 ones with reaching expressions
4782 - we know which insns are redundant when we go to create copies */
4784 changed = pre_delete ();
4786 did_insert = pre_edge_insert (edge_list, index_map);
4788 /* In other places with reaching expressions, copy the expression to the
4789 specially allocated pseudo-reg that reaches the redundant expr. */
4790 pre_insert_copies ();
4793 commit_edge_insertions ();
4798 free (pre_redundant_insns);
4802 /* Top level routine to perform one PRE GCSE pass.
4804 Return non-zero if a change was made. */
4807 one_pre_gcse_pass (pass)
4812 gcse_subst_count = 0;
4813 gcse_create_count = 0;
4815 alloc_expr_hash_table (max_cuid);
4816 add_noreturn_fake_exit_edges ();
4817 compute_expr_hash_table ();
4819 dump_hash_table (gcse_file, "Expression", expr_hash_table,
4820 expr_hash_table_size, n_exprs);
4824 alloc_pre_mem (n_basic_blocks, n_exprs);
4825 compute_pre_data ();
4826 changed |= pre_gcse ();
4827 free_edge_list (edge_list);
4831 remove_fake_edges ();
4832 free_expr_hash_table ();
4836 fprintf (gcse_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
4837 current_function_name, pass, bytes_used);
4838 fprintf (gcse_file, "%d substs, %d insns created\n",
4839 gcse_subst_count, gcse_create_count);
4845 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to INSN.
4846 If notes are added to an insn which references a CODE_LABEL, the
4847 LABEL_NUSES count is incremented. We have to add REG_LABEL notes,
4848 because the following loop optimization pass requires them. */
4850 /* ??? This is very similar to the loop.c add_label_notes function. We
4851 could probably share code here. */
4853 /* ??? If there was a jump optimization pass after gcse and before loop,
4854 then we would not need to do this here, because jump would add the
4855 necessary REG_LABEL notes. */
4858 add_label_notes (x, insn)
4862 enum rtx_code code = GET_CODE (x);
4866 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
4868 /* This code used to ignore labels that referred to dispatch tables to
4869 avoid flow generating (slighly) worse code.
4871 We no longer ignore such label references (see LABEL_REF handling in
4872 mark_jump_label for additional information). */
4874 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
4876 if (LABEL_P (XEXP (x, 0)))
4877 LABEL_NUSES (XEXP (x, 0))++;
4881 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
4884 add_label_notes (XEXP (x, i), insn);
4885 else if (fmt[i] == 'E')
4886 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4887 add_label_notes (XVECEXP (x, i, j), insn);
4891 /* Compute transparent outgoing information for each block.
4893 An expression is transparent to an edge unless it is killed by
4894 the edge itself. This can only happen with abnormal control flow,
4895 when the edge is traversed through a call. This happens with
4896 non-local labels and exceptions.
4898 This would not be necessary if we split the edge. While this is
4899 normally impossible for abnormal critical edges, with some effort
4900 it should be possible with exception handling, since we still have
4901 control over which handler should be invoked. But due to increased
4902 EH table sizes, this may not be worthwhile. */
4905 compute_transpout ()
4911 sbitmap_vector_ones (transpout, n_basic_blocks);
4913 for (bb = 0; bb < n_basic_blocks; ++bb)
4915 /* Note that flow inserted a nop a the end of basic blocks that
4916 end in call instructions for reasons other than abnormal
4918 if (GET_CODE (BLOCK_END (bb)) != CALL_INSN)
4921 for (i = 0; i < expr_hash_table_size; i++)
4922 for (expr = expr_hash_table[i]; expr ; expr = expr->next_same_hash)
4923 if (GET_CODE (expr->expr) == MEM)
4925 if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
4926 && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
4929 /* ??? Optimally, we would use interprocedural alias
4930 analysis to determine if this mem is actually killed
4932 RESET_BIT (transpout[bb], expr->bitmap_index);
4937 /* Removal of useless null pointer checks */
4939 /* Called via note_stores. X is set by SETTER. If X is a register we must
4940 invalidate nonnull_local and set nonnull_killed. DATA is really a
4941 `null_pointer_info *'.
4943 We ignore hard registers. */
4946 invalidate_nonnull_info (x, setter, data)
4948 rtx setter ATTRIBUTE_UNUSED;
4952 struct null_pointer_info *npi = (struct null_pointer_info *) data;
4954 while (GET_CODE (x) == SUBREG)
4957 /* Ignore anything that is not a register or is a hard register. */
4958 if (GET_CODE (x) != REG
4959 || REGNO (x) < npi->min_reg
4960 || REGNO (x) >= npi->max_reg)
4963 regno = REGNO (x) - npi->min_reg;
4965 RESET_BIT (npi->nonnull_local[npi->current_block], regno);
4966 SET_BIT (npi->nonnull_killed[npi->current_block], regno);
4969 /* Do null-pointer check elimination for the registers indicated in
4970 NPI. NONNULL_AVIN and NONNULL_AVOUT are pre-allocated sbitmaps;
4971 they are not our responsibility to free. */
4974 delete_null_pointer_checks_1 (block_reg, nonnull_avin, nonnull_avout, npi)
4975 unsigned int *block_reg;
4976 sbitmap *nonnull_avin;
4977 sbitmap *nonnull_avout;
4978 struct null_pointer_info *npi;
4982 sbitmap *nonnull_local = npi->nonnull_local;
4983 sbitmap *nonnull_killed = npi->nonnull_killed;
4985 /* Compute local properties, nonnull and killed. A register will have
4986 the nonnull property if at the end of the current block its value is
4987 known to be nonnull. The killed property indicates that somewhere in
4988 the block any information we had about the register is killed.
4990 Note that a register can have both properties in a single block. That
4991 indicates that it's killed, then later in the block a new value is
4993 sbitmap_vector_zero (nonnull_local, n_basic_blocks);
4994 sbitmap_vector_zero (nonnull_killed, n_basic_blocks);
4996 for (current_block = 0; current_block < n_basic_blocks; current_block++)
4998 rtx insn, stop_insn;
5000 /* Set the current block for invalidate_nonnull_info. */
5001 npi->current_block = current_block;
5003 /* Scan each insn in the basic block looking for memory references and
5005 stop_insn = NEXT_INSN (BLOCK_END (current_block));
5006 for (insn = BLOCK_HEAD (current_block);
5008 insn = NEXT_INSN (insn))
5013 /* Ignore anything that is not a normal insn. */
5014 if (! INSN_P (insn))
5017 /* Basically ignore anything that is not a simple SET. We do have
5018 to make sure to invalidate nonnull_local and set nonnull_killed
5019 for such insns though. */
5020 set = single_set (insn);
5023 note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
5027 /* See if we've got a useable memory load. We handle it first
5028 in case it uses its address register as a dest (which kills
5029 the nonnull property). */
5030 if (GET_CODE (SET_SRC (set)) == MEM
5031 && GET_CODE ((reg = XEXP (SET_SRC (set), 0))) == REG
5032 && REGNO (reg) >= npi->min_reg
5033 && REGNO (reg) < npi->max_reg)
5034 SET_BIT (nonnull_local[current_block],
5035 REGNO (reg) - npi->min_reg);
5037 /* Now invalidate stuff clobbered by this insn. */
5038 note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
5040 /* And handle stores, we do these last since any sets in INSN can
5041 not kill the nonnull property if it is derived from a MEM
5042 appearing in a SET_DEST. */
5043 if (GET_CODE (SET_DEST (set)) == MEM
5044 && GET_CODE ((reg = XEXP (SET_DEST (set), 0))) == REG
5045 && REGNO (reg) >= npi->min_reg
5046 && REGNO (reg) < npi->max_reg)
5047 SET_BIT (nonnull_local[current_block],
5048 REGNO (reg) - npi->min_reg);
5052 /* Now compute global properties based on the local properties. This
5053 is a classic global availablity algorithm. */
5054 compute_available (nonnull_local, nonnull_killed,
5055 nonnull_avout, nonnull_avin);
5057 /* Now look at each bb and see if it ends with a compare of a value
5059 for (bb = 0; bb < n_basic_blocks; bb++)
5061 rtx last_insn = BLOCK_END (bb);
5062 rtx condition, earliest;
5063 int compare_and_branch;
5065 /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
5066 since BLOCK_REG[BB] is zero if this block did not end with a
5067 comparison against zero, this condition works. */
5068 if (block_reg[bb] < npi->min_reg
5069 || block_reg[bb] >= npi->max_reg)
5072 /* LAST_INSN is a conditional jump. Get its condition. */
5073 condition = get_condition (last_insn, &earliest);
5075 /* If we can't determine the condition then skip. */
5079 /* Is the register known to have a nonzero value? */
5080 if (!TEST_BIT (nonnull_avout[bb], block_reg[bb] - npi->min_reg))
5083 /* Try to compute whether the compare/branch at the loop end is one or
5084 two instructions. */
5085 if (earliest == last_insn)
5086 compare_and_branch = 1;
5087 else if (earliest == prev_nonnote_insn (last_insn))
5088 compare_and_branch = 2;
5092 /* We know the register in this comparison is nonnull at exit from
5093 this block. We can optimize this comparison. */
5094 if (GET_CODE (condition) == NE)
5098 new_jump = emit_jump_insn_before (gen_jump (JUMP_LABEL (last_insn)),
5100 JUMP_LABEL (new_jump) = JUMP_LABEL (last_insn);
5101 LABEL_NUSES (JUMP_LABEL (new_jump))++;
5102 emit_barrier_after (new_jump);
5104 delete_insn (last_insn);
5105 if (compare_and_branch == 2)
5106 delete_insn (earliest);
5108 /* Don't check this block again. (Note that BLOCK_END is
5109 invalid here; we deleted the last instruction in the
5115 /* Find EQ/NE comparisons against zero which can be (indirectly) evaluated
5118 This is conceptually similar to global constant/copy propagation and
5119 classic global CSE (it even uses the same dataflow equations as cprop).
5121 If a register is used as memory address with the form (mem (reg)), then we
5122 know that REG can not be zero at that point in the program. Any instruction
5123 which sets REG "kills" this property.
5125 So, if every path leading to a conditional branch has an available memory
5126 reference of that form, then we know the register can not have the value
5127 zero at the conditional branch.
5129 So we merely need to compute the local properies and propagate that data
5130 around the cfg, then optimize where possible.
5132 We run this pass two times. Once before CSE, then again after CSE. This
5133 has proven to be the most profitable approach. It is rare for new
5134 optimization opportunities of this nature to appear after the first CSE
5137 This could probably be integrated with global cprop with a little work. */
5140 delete_null_pointer_checks (f)
5141 rtx f ATTRIBUTE_UNUSED;
5143 sbitmap *nonnull_avin, *nonnull_avout;
5144 unsigned int *block_reg;
5149 struct null_pointer_info npi;
5151 /* If we have only a single block, then there's nothing to do. */
5152 if (n_basic_blocks <= 1)
5155 /* Trying to perform global optimizations on flow graphs which have
5156 a high connectivity will take a long time and is unlikely to be
5157 particularly useful.
5159 In normal circumstances a cfg should have about twice has many edges
5160 as blocks. But we do not want to punish small functions which have
5161 a couple switch statements. So we require a relatively large number
5162 of basic blocks and the ratio of edges to blocks to be high. */
5163 if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
5166 /* We need four bitmaps, each with a bit for each register in each
5168 max_reg = max_reg_num ();
5169 regs_per_pass = get_bitmap_width (4, n_basic_blocks, max_reg);
5171 /* Allocate bitmaps to hold local and global properties. */
5172 npi.nonnull_local = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5173 npi.nonnull_killed = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5174 nonnull_avin = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5175 nonnull_avout = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5177 /* Go through the basic blocks, seeing whether or not each block
5178 ends with a conditional branch whose condition is a comparison
5179 against zero. Record the register compared in BLOCK_REG. */
5180 block_reg = (unsigned int *) xcalloc (n_basic_blocks, sizeof (int));
5181 for (bb = 0; bb < n_basic_blocks; bb++)
5183 rtx last_insn = BLOCK_END (bb);
5184 rtx condition, earliest, reg;
5186 /* We only want conditional branches. */
5187 if (GET_CODE (last_insn) != JUMP_INSN
5188 || !any_condjump_p (last_insn)
5189 || !onlyjump_p (last_insn))
5192 /* LAST_INSN is a conditional jump. Get its condition. */
5193 condition = get_condition (last_insn, &earliest);
5195 /* If we were unable to get the condition, or it is not a equality
5196 comparison against zero then there's nothing we can do. */
5198 || (GET_CODE (condition) != NE && GET_CODE (condition) != EQ)
5199 || GET_CODE (XEXP (condition, 1)) != CONST_INT
5200 || (XEXP (condition, 1)
5201 != CONST0_RTX (GET_MODE (XEXP (condition, 0)))))
5204 /* We must be checking a register against zero. */
5205 reg = XEXP (condition, 0);
5206 if (GET_CODE (reg) != REG)
5209 block_reg[bb] = REGNO (reg);
5212 /* Go through the algorithm for each block of registers. */
5213 for (reg = FIRST_PSEUDO_REGISTER; reg < max_reg; reg += regs_per_pass)
5216 npi.max_reg = MIN (reg + regs_per_pass, max_reg);
5217 delete_null_pointer_checks_1 (block_reg, nonnull_avin,
5218 nonnull_avout, &npi);
5221 /* Free the table of registers compared at the end of every block. */
5225 free (npi.nonnull_local);
5226 free (npi.nonnull_killed);
5227 free (nonnull_avin);
5228 free (nonnull_avout);
5231 /* Code Hoisting variables and subroutines. */
5233 /* Very busy expressions. */
5234 static sbitmap *hoist_vbein;
5235 static sbitmap *hoist_vbeout;
5237 /* Hoistable expressions. */
5238 static sbitmap *hoist_exprs;
5240 /* Dominator bitmaps. */
5241 static sbitmap *dominators;
5243 /* ??? We could compute post dominators and run this algorithm in
5244 reverse to to perform tail merging, doing so would probably be
5245 more effective than the tail merging code in jump.c.
5247 It's unclear if tail merging could be run in parallel with
5248 code hoisting. It would be nice. */
5250 /* Allocate vars used for code hoisting analysis. */
5253 alloc_code_hoist_mem (n_blocks, n_exprs)
5254 int n_blocks, n_exprs;
5256 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
5257 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
5258 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
5260 hoist_vbein = sbitmap_vector_alloc (n_blocks, n_exprs);
5261 hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
5262 hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
5263 transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
5265 dominators = sbitmap_vector_alloc (n_blocks, n_blocks);
5268 /* Free vars used for code hoisting analysis. */
5271 free_code_hoist_mem ()
5278 free (hoist_vbeout);
5285 /* Compute the very busy expressions at entry/exit from each block.
5287 An expression is very busy if all paths from a given point
5288 compute the expression. */
5291 compute_code_hoist_vbeinout ()
5293 int bb, changed, passes;
5295 sbitmap_vector_zero (hoist_vbeout, n_basic_blocks);
5296 sbitmap_vector_zero (hoist_vbein, n_basic_blocks);
5305 /* We scan the blocks in the reverse order to speed up
5307 for (bb = n_basic_blocks - 1; bb >= 0; bb--)
5309 changed |= sbitmap_a_or_b_and_c (hoist_vbein[bb], antloc[bb],
5310 hoist_vbeout[bb], transp[bb]);
5311 if (bb != n_basic_blocks - 1)
5312 sbitmap_intersection_of_succs (hoist_vbeout[bb], hoist_vbein, bb);
5319 fprintf (gcse_file, "hoisting vbeinout computation: %d passes\n", passes);
5322 /* Top level routine to do the dataflow analysis needed by code hoisting. */
5325 compute_code_hoist_data ()
5327 compute_local_properties (transp, comp, antloc, 0);
5328 compute_transpout ();
5329 compute_code_hoist_vbeinout ();
5330 calculate_dominance_info (NULL, dominators, CDI_DOMINATORS);
5332 fprintf (gcse_file, "\n");
5335 /* Determine if the expression identified by EXPR_INDEX would
5336 reach BB unimpared if it was placed at the end of EXPR_BB.
5338 It's unclear exactly what Muchnick meant by "unimpared". It seems
5339 to me that the expression must either be computed or transparent in
5340 *every* block in the path(s) from EXPR_BB to BB. Any other definition
5341 would allow the expression to be hoisted out of loops, even if
5342 the expression wasn't a loop invariant.
5344 Contrast this to reachability for PRE where an expression is
5345 considered reachable if *any* path reaches instead of *all*
5349 hoist_expr_reaches_here_p (expr_bb, expr_index, bb, visited)
5356 int visited_allocated_locally = 0;
5359 if (visited == NULL)
5361 visited_allocated_locally = 1;
5362 visited = xcalloc (n_basic_blocks, 1);
5365 for (pred = BASIC_BLOCK (bb)->pred; pred != NULL; pred = pred->pred_next)
5367 int pred_bb = pred->src->index;
5369 if (pred->src == ENTRY_BLOCK_PTR)
5371 else if (visited[pred_bb])
5374 /* Does this predecessor generate this expression? */
5375 else if (TEST_BIT (comp[pred_bb], expr_index))
5377 else if (! TEST_BIT (transp[pred_bb], expr_index))
5383 visited[pred_bb] = 1;
5384 if (! hoist_expr_reaches_here_p (expr_bb, expr_index,
5389 if (visited_allocated_locally)
5392 return (pred == NULL);
5395 /* Actually perform code hoisting. */
5402 struct expr **index_map;
5405 sbitmap_vector_zero (hoist_exprs, n_basic_blocks);
5407 /* Compute a mapping from expression number (`bitmap_index') to
5408 hash table entry. */
5410 index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
5411 for (i = 0; i < expr_hash_table_size; i++)
5412 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
5413 index_map[expr->bitmap_index] = expr;
5415 /* Walk over each basic block looking for potentially hoistable
5416 expressions, nothing gets hoisted from the entry block. */
5417 for (bb = 0; bb < n_basic_blocks; bb++)
5420 int insn_inserted_p;
5422 /* Examine each expression that is very busy at the exit of this
5423 block. These are the potentially hoistable expressions. */
5424 for (i = 0; i < hoist_vbeout[bb]->n_bits; i++)
5428 if (TEST_BIT (hoist_vbeout[bb], i) && TEST_BIT (transpout[bb], i))
5430 /* We've found a potentially hoistable expression, now
5431 we look at every block BB dominates to see if it
5432 computes the expression. */
5433 for (dominated = 0; dominated < n_basic_blocks; dominated++)
5435 /* Ignore self dominance. */
5437 || ! TEST_BIT (dominators[dominated], bb))
5440 /* We've found a dominated block, now see if it computes
5441 the busy expression and whether or not moving that
5442 expression to the "beginning" of that block is safe. */
5443 if (!TEST_BIT (antloc[dominated], i))
5446 /* Note if the expression would reach the dominated block
5447 unimpared if it was placed at the end of BB.
5449 Keep track of how many times this expression is hoistable
5450 from a dominated block into BB. */
5451 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
5455 /* If we found more than one hoistable occurence of this
5456 expression, then note it in the bitmap of expressions to
5457 hoist. It makes no sense to hoist things which are computed
5458 in only one BB, and doing so tends to pessimize register
5459 allocation. One could increase this value to try harder
5460 to avoid any possible code expansion due to register
5461 allocation issues; however experiments have shown that
5462 the vast majority of hoistable expressions are only movable
5463 from two successors, so raising this threshhold is likely
5464 to nullify any benefit we get from code hoisting. */
5467 SET_BIT (hoist_exprs[bb], i);
5473 /* If we found nothing to hoist, then quit now. */
5477 /* Loop over all the hoistable expressions. */
5478 for (i = 0; i < hoist_exprs[bb]->n_bits; i++)
5480 /* We want to insert the expression into BB only once, so
5481 note when we've inserted it. */
5482 insn_inserted_p = 0;
5484 /* These tests should be the same as the tests above. */
5485 if (TEST_BIT (hoist_vbeout[bb], i))
5487 /* We've found a potentially hoistable expression, now
5488 we look at every block BB dominates to see if it
5489 computes the expression. */
5490 for (dominated = 0; dominated < n_basic_blocks; dominated++)
5492 /* Ignore self dominance. */
5494 || ! TEST_BIT (dominators[dominated], bb))
5497 /* We've found a dominated block, now see if it computes
5498 the busy expression and whether or not moving that
5499 expression to the "beginning" of that block is safe. */
5500 if (!TEST_BIT (antloc[dominated], i))
5503 /* The expression is computed in the dominated block and
5504 it would be safe to compute it at the start of the
5505 dominated block. Now we have to determine if the
5506 expresion would reach the dominated block if it was
5507 placed at the end of BB. */
5508 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
5510 struct expr *expr = index_map[i];
5511 struct occr *occr = expr->antic_occr;
5515 /* Find the right occurence of this expression. */
5516 while (BLOCK_NUM (occr->insn) != dominated && occr)
5519 /* Should never happen. */
5525 set = single_set (insn);
5529 /* Create a pseudo-reg to store the result of reaching
5530 expressions into. Get the mode for the new pseudo
5531 from the mode of the original destination pseudo. */
5532 if (expr->reaching_reg == NULL)
5534 = gen_reg_rtx (GET_MODE (SET_DEST (set)));
5536 /* In theory this should never fail since we're creating
5539 However, on the x86 some of the movXX patterns
5540 actually contain clobbers of scratch regs. This may
5541 cause the insn created by validate_change to not
5542 match any pattern and thus cause validate_change to
5544 if (validate_change (insn, &SET_SRC (set),
5545 expr->reaching_reg, 0))
5547 occr->deleted_p = 1;
5548 if (!insn_inserted_p)
5550 insert_insn_end_bb (index_map[i], bb, 0);
5551 insn_inserted_p = 1;
5563 /* Top level routine to perform one code hoisting (aka unification) pass
5565 Return non-zero if a change was made. */
5568 one_code_hoisting_pass ()
5572 alloc_expr_hash_table (max_cuid);
5573 compute_expr_hash_table ();
5575 dump_hash_table (gcse_file, "Code Hosting Expressions", expr_hash_table,
5576 expr_hash_table_size, n_exprs);
5580 alloc_code_hoist_mem (n_basic_blocks, n_exprs);
5581 compute_code_hoist_data ();
5583 free_code_hoist_mem ();
5586 free_expr_hash_table ();