1 /* Global common subexpression elimination/Partial redundancy elimination
2 and global constant/copy propagation for GNU compiler.
3 Copyright (C) 1997, 1998, 1999, 2000 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"
164 #define obstack_chunk_alloc gmalloc
165 #define obstack_chunk_free free
167 /* Maximum number of passes to perform. */
170 /* Propagate flow information through back edges and thus enable PRE's
171 moving loop invariant calculations out of loops.
173 Originally this tended to create worse overall code, but several
174 improvements during the development of PRE seem to have made following
175 back edges generally a win.
177 Note much of the loop invariant code motion done here would normally
178 be done by loop.c, which has more heuristics for when to move invariants
179 out of loops. At some point we might need to move some of those
180 heuristics into gcse.c. */
181 #define FOLLOW_BACK_EDGES 1
183 /* We support GCSE via Partial Redundancy Elimination. PRE optimizations
184 are a superset of those done by GCSE.
186 We perform the following steps:
188 1) Compute basic block information.
190 2) Compute table of places where registers are set.
192 3) Perform copy/constant propagation.
194 4) Perform global cse.
196 5) Perform another pass of copy/constant propagation.
198 Two passes of copy/constant propagation are done because the first one
199 enables more GCSE and the second one helps to clean up the copies that
200 GCSE creates. This is needed more for PRE than for Classic because Classic
201 GCSE will try to use an existing register containing the common
202 subexpression rather than create a new one. This is harder to do for PRE
203 because of the code motion (which Classic GCSE doesn't do).
205 Expressions we are interested in GCSE-ing are of the form
206 (set (pseudo-reg) (expression)).
207 Function want_to_gcse_p says what these are.
209 PRE handles moving invariant expressions out of loops (by treating them as
210 partially redundant).
212 Eventually it would be nice to replace cse.c/gcse.c with SSA (static single
213 assignment) based GVN (global value numbering). L. T. Simpson's paper
214 (Rice University) on value numbering is a useful reference for this.
216 **********************
218 We used to support multiple passes but there are diminishing returns in
219 doing so. The first pass usually makes 90% of the changes that are doable.
220 A second pass can make a few more changes made possible by the first pass.
221 Experiments show any further passes don't make enough changes to justify
224 A study of spec92 using an unlimited number of passes:
225 [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
226 [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
227 [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
229 It was found doing copy propagation between each pass enables further
232 PRE is quite expensive in complicated functions because the DFA can take
233 awhile to converge. Hence we only perform one pass. Macro MAX_PASSES can
234 be modified if one wants to experiment.
236 **********************
238 The steps for PRE are:
240 1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
242 2) Perform the data flow analysis for PRE.
244 3) Delete the redundant instructions
246 4) Insert the required copies [if any] that make the partially
247 redundant instructions fully redundant.
249 5) For other reaching expressions, insert an instruction to copy the value
250 to a newly created pseudo that will reach the redundant instruction.
252 The deletion is done first so that when we do insertions we
253 know which pseudo reg to use.
255 Various papers have argued that PRE DFA is expensive (O(n^2)) and others
256 argue it is not. The number of iterations for the algorithm to converge
257 is typically 2-4 so I don't view it as that expensive (relatively speaking).
259 PRE GCSE depends heavily on the second CSE pass to clean up the copies
260 we create. To make an expression reach the place where it's redundant,
261 the result of the expression is copied to a new register, and the redundant
262 expression is deleted by replacing it with this new register. Classic GCSE
263 doesn't have this problem as much as it computes the reaching defs of
264 each register in each block and thus can try to use an existing register.
266 **********************
268 A fair bit of simplicity is created by creating small functions for simple
269 tasks, even when the function is only called in one place. This may
270 measurably slow things down [or may not] by creating more function call
271 overhead than is necessary. The source is laid out so that it's trivial
272 to make the affected functions inline so that one can measure what speed
273 up, if any, can be achieved, and maybe later when things settle things can
276 Help stamp out big monolithic functions! */
278 /* GCSE global vars. */
281 static FILE *gcse_file;
283 /* Note whether or not we should run jump optimization after gcse. We
284 want to do this for two cases.
286 * If we changed any jumps via cprop.
288 * If we added any labels via edge splitting. */
290 static int run_jump_opt_after_gcse;
292 /* Bitmaps are normally not included in debugging dumps.
293 However it's useful to be able to print them from GDB.
294 We could create special functions for this, but it's simpler to
295 just allow passing stderr to the dump_foo fns. Since stderr can
296 be a macro, we store a copy here. */
297 static FILE *debug_stderr;
299 /* An obstack for our working variables. */
300 static struct obstack gcse_obstack;
302 /* Non-zero for each mode that supports (set (reg) (reg)).
303 This is trivially true for integer and floating point values.
304 It may or may not be true for condition codes. */
305 static char can_copy_p[(int) NUM_MACHINE_MODES];
307 /* Non-zero if can_copy_p has been initialized. */
308 static int can_copy_init_p;
310 struct reg_use {rtx reg_rtx; };
312 /* Hash table of expressions. */
316 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
318 /* Index in the available expression bitmaps. */
320 /* Next entry with the same hash. */
321 struct expr *next_same_hash;
322 /* List of anticipatable occurrences in basic blocks in the function.
323 An "anticipatable occurrence" is one that is the first occurrence in the
324 basic block, the operands are not modified in the basic block prior
325 to the occurrence and the output is not used between the start of
326 the block and the occurrence. */
327 struct occr *antic_occr;
328 /* List of available occurrence in basic blocks in the function.
329 An "available occurrence" is one that is the last occurrence in the
330 basic block and the operands are not modified by following statements in
331 the basic block [including this insn]. */
332 struct occr *avail_occr;
333 /* Non-null if the computation is PRE redundant.
334 The value is the newly created pseudo-reg to record a copy of the
335 expression in all the places that reach the redundant copy. */
339 /* Occurrence of an expression.
340 There is one per basic block. If a pattern appears more than once the
341 last appearance is used [or first for anticipatable expressions]. */
345 /* Next occurrence of this expression. */
347 /* The insn that computes the expression. */
349 /* Non-zero if this [anticipatable] occurrence has been deleted. */
351 /* Non-zero if this [available] occurrence has been copied to
353 /* ??? This is mutually exclusive with deleted_p, so they could share
358 /* Expression and copy propagation hash tables.
359 Each hash table is an array of buckets.
360 ??? It is known that if it were an array of entries, structure elements
361 `next_same_hash' and `bitmap_index' wouldn't be necessary. However, it is
362 not clear whether in the final analysis a sufficient amount of memory would
363 be saved as the size of the available expression bitmaps would be larger
364 [one could build a mapping table without holes afterwards though].
365 Someday I'll perform the computation and figure it out. */
367 /* Total size of the expression hash table, in elements. */
368 static unsigned int expr_hash_table_size;
371 This is an array of `expr_hash_table_size' elements. */
372 static struct expr **expr_hash_table;
374 /* Total size of the copy propagation hash table, in elements. */
375 static int set_hash_table_size;
378 This is an array of `set_hash_table_size' elements. */
379 static struct expr **set_hash_table;
381 /* Mapping of uids to cuids.
382 Only real insns get cuids. */
383 static int *uid_cuid;
385 /* Highest UID in UID_CUID. */
388 /* Get the cuid of an insn. */
389 #ifdef ENABLE_CHECKING
390 #define INSN_CUID(INSN) (INSN_UID (INSN) > max_uid ? (abort (), 0) : uid_cuid[INSN_UID (INSN)])
392 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
395 /* Number of cuids. */
398 /* Mapping of cuids to insns. */
399 static rtx *cuid_insn;
401 /* Get insn from cuid. */
402 #define CUID_INSN(CUID) (cuid_insn[CUID])
404 /* Maximum register number in function prior to doing gcse + 1.
405 Registers created during this pass have regno >= max_gcse_regno.
406 This is named with "gcse" to not collide with global of same name. */
407 static unsigned int max_gcse_regno;
409 /* Maximum number of cse-able expressions found. */
412 /* Maximum number of assignments for copy propagation found. */
415 /* Table of registers that are modified.
417 For each register, each element is a list of places where the pseudo-reg
420 For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
421 requires knowledge of which blocks kill which regs [and thus could use
422 a bitmap instead of the lists `reg_set_table' uses].
424 `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
425 num-regs) [however perhaps it may be useful to keep the data as is]. One
426 advantage of recording things this way is that `reg_set_table' is fairly
427 sparse with respect to pseudo regs but for hard regs could be fairly dense
428 [relatively speaking]. And recording sets of pseudo-regs in lists speeds
429 up functions like compute_transp since in the case of pseudo-regs we only
430 need to iterate over the number of times a pseudo-reg is set, not over the
431 number of basic blocks [clearly there is a bit of a slow down in the cases
432 where a pseudo is set more than once in a block, however it is believed
433 that the net effect is to speed things up]. This isn't done for hard-regs
434 because recording call-clobbered hard-regs in `reg_set_table' at each
435 function call can consume a fair bit of memory, and iterating over
436 hard-regs stored this way in compute_transp will be more expensive. */
438 typedef struct reg_set
440 /* The next setting of this register. */
441 struct reg_set *next;
442 /* The insn where it was set. */
446 static reg_set **reg_set_table;
448 /* Size of `reg_set_table'.
449 The table starts out at max_gcse_regno + slop, and is enlarged as
451 static int reg_set_table_size;
453 /* Amount to grow `reg_set_table' by when it's full. */
454 #define REG_SET_TABLE_SLOP 100
456 /* Bitmap containing one bit for each register in the program.
457 Used when performing GCSE to track which registers have been set since
458 the start of the basic block. */
459 static sbitmap reg_set_bitmap;
461 /* For each block, a bitmap of registers set in the block.
462 This is used by expr_killed_p and compute_transp.
463 It is computed during hash table computation and not by compute_sets
464 as it includes registers added since the last pass (or between cprop and
465 gcse) and it's currently not easy to realloc sbitmap vectors. */
466 static sbitmap *reg_set_in_block;
468 /* For each block, non-zero if memory is set in that block.
469 This is computed during hash table computation and is used by
470 expr_killed_p and compute_transp.
471 ??? Handling of memory is very simple, we don't make any attempt
472 to optimize things (later).
473 ??? This can be computed by compute_sets since the information
475 static char *mem_set_in_block;
477 /* Various variables for statistics gathering. */
479 /* Memory used in a pass.
480 This isn't intended to be absolutely precise. Its intent is only
481 to keep an eye on memory usage. */
482 static int bytes_used;
484 /* GCSE substitutions made. */
485 static int gcse_subst_count;
486 /* Number of copy instructions created. */
487 static int gcse_create_count;
488 /* Number of constants propagated. */
489 static int const_prop_count;
490 /* Number of copys propagated. */
491 static int copy_prop_count;
493 /* These variables are used by classic GCSE.
494 Normally they'd be defined a bit later, but `rd_gen' needs to
495 be declared sooner. */
497 /* Each block has a bitmap of each type.
498 The length of each blocks bitmap is:
500 max_cuid - for reaching definitions
501 n_exprs - for available expressions
503 Thus we view the bitmaps as 2 dimensional arrays. i.e.
504 rd_kill[block_num][cuid_num]
505 ae_kill[block_num][expr_num] */
507 /* For reaching defs */
508 static sbitmap *rd_kill, *rd_gen, *reaching_defs, *rd_out;
510 /* for available exprs */
511 static sbitmap *ae_kill, *ae_gen, *ae_in, *ae_out;
513 /* Objects of this type are passed around by the null-pointer check
515 struct null_pointer_info
517 /* The basic block being processed. */
519 /* The first register to be handled in this pass. */
520 unsigned int min_reg;
521 /* One greater than the last register to be handled in this pass. */
522 unsigned int max_reg;
523 sbitmap *nonnull_local;
524 sbitmap *nonnull_killed;
527 static void compute_can_copy PARAMS ((void));
528 static char *gmalloc PARAMS ((unsigned int));
529 static char *grealloc PARAMS ((char *, unsigned int));
530 static char *gcse_alloc PARAMS ((unsigned long));
531 static void alloc_gcse_mem PARAMS ((rtx));
532 static void free_gcse_mem PARAMS ((void));
533 static void alloc_reg_set_mem PARAMS ((int));
534 static void free_reg_set_mem PARAMS ((void));
535 static int get_bitmap_width PARAMS ((int, int, int));
536 static void record_one_set PARAMS ((int, rtx));
537 static void record_set_info PARAMS ((rtx, rtx, void *));
538 static void compute_sets PARAMS ((rtx));
539 static void hash_scan_insn PARAMS ((rtx, int, int));
540 static void hash_scan_set PARAMS ((rtx, rtx, int));
541 static void hash_scan_clobber PARAMS ((rtx, rtx));
542 static void hash_scan_call PARAMS ((rtx, rtx));
543 static int want_to_gcse_p PARAMS ((rtx));
544 static int oprs_unchanged_p PARAMS ((rtx, rtx, int));
545 static int oprs_anticipatable_p PARAMS ((rtx, rtx));
546 static int oprs_available_p PARAMS ((rtx, rtx));
547 static void insert_expr_in_table PARAMS ((rtx, enum machine_mode, rtx,
549 static void insert_set_in_table PARAMS ((rtx, rtx));
550 static unsigned int hash_expr PARAMS ((rtx, enum machine_mode, int *, int));
551 static unsigned int hash_expr_1 PARAMS ((rtx, enum machine_mode, int *));
552 static unsigned int hash_set PARAMS ((int, int));
553 static int expr_equiv_p PARAMS ((rtx, rtx));
554 static void record_last_reg_set_info PARAMS ((rtx, int));
555 static void record_last_mem_set_info PARAMS ((rtx));
556 static void record_last_set_info PARAMS ((rtx, rtx, void *));
557 static void compute_hash_table PARAMS ((int));
558 static void alloc_set_hash_table PARAMS ((int));
559 static void free_set_hash_table PARAMS ((void));
560 static void compute_set_hash_table PARAMS ((void));
561 static void alloc_expr_hash_table PARAMS ((unsigned int));
562 static void free_expr_hash_table PARAMS ((void));
563 static void compute_expr_hash_table PARAMS ((void));
564 static void dump_hash_table PARAMS ((FILE *, const char *, struct expr **,
566 static struct expr *lookup_expr PARAMS ((rtx));
567 static struct expr *lookup_set PARAMS ((unsigned int, rtx));
568 static struct expr *next_set PARAMS ((unsigned int, struct expr *));
569 static void reset_opr_set_tables PARAMS ((void));
570 static int oprs_not_set_p PARAMS ((rtx, rtx));
571 static void mark_call PARAMS ((rtx));
572 static void mark_set PARAMS ((rtx, rtx));
573 static void mark_clobber PARAMS ((rtx, rtx));
574 static void mark_oprs_set PARAMS ((rtx));
575 static void alloc_cprop_mem PARAMS ((int, int));
576 static void free_cprop_mem PARAMS ((void));
577 static void compute_transp PARAMS ((rtx, int, sbitmap *, int));
578 static void compute_transpout PARAMS ((void));
579 static void compute_local_properties PARAMS ((sbitmap *, sbitmap *, sbitmap *,
581 static void compute_cprop_data PARAMS ((void));
582 static void find_used_regs PARAMS ((rtx));
583 static int try_replace_reg PARAMS ((rtx, rtx, rtx));
584 static struct expr *find_avail_set PARAMS ((int, rtx));
585 static int cprop_jump PARAMS ((rtx, rtx, struct reg_use *, rtx));
587 static int cprop_cc0_jump PARAMS ((rtx, struct reg_use *, rtx));
589 static int cprop_insn PARAMS ((rtx, int));
590 static int cprop PARAMS ((int));
591 static int one_cprop_pass PARAMS ((int, int));
592 static void alloc_pre_mem PARAMS ((int, int));
593 static void free_pre_mem PARAMS ((void));
594 static void compute_pre_data PARAMS ((void));
595 static int pre_expr_reaches_here_p PARAMS ((int, struct expr *, int));
596 static void insert_insn_end_bb PARAMS ((struct expr *, int, int));
597 static void pre_insert_copy_insn PARAMS ((struct expr *, rtx));
598 static void pre_insert_copies PARAMS ((void));
599 static int pre_delete PARAMS ((void));
600 static int pre_gcse PARAMS ((void));
601 static int one_pre_gcse_pass PARAMS ((int));
602 static void add_label_notes PARAMS ((rtx, rtx));
603 static void alloc_code_hoist_mem PARAMS ((int, int));
604 static void free_code_hoist_mem PARAMS ((void));
605 static void compute_code_hoist_vbeinout PARAMS ((void));
606 static void compute_code_hoist_data PARAMS ((void));
607 static int hoist_expr_reaches_here_p PARAMS ((int, int, int, char *));
608 static void hoist_code PARAMS ((void));
609 static int one_code_hoisting_pass PARAMS ((void));
610 static void alloc_rd_mem PARAMS ((int, int));
611 static void free_rd_mem PARAMS ((void));
612 static void handle_rd_kill_set PARAMS ((rtx, int, int));
613 static void compute_kill_rd PARAMS ((void));
614 static void compute_rd PARAMS ((void));
615 static void alloc_avail_expr_mem PARAMS ((int, int));
616 static void free_avail_expr_mem PARAMS ((void));
617 static void compute_ae_gen PARAMS ((void));
618 static int expr_killed_p PARAMS ((rtx, int));
619 static void compute_ae_kill PARAMS ((sbitmap *, sbitmap *));
620 static int expr_reaches_here_p PARAMS ((struct occr *, struct expr *,
622 static rtx computing_insn PARAMS ((struct expr *, rtx));
623 static int def_reaches_here_p PARAMS ((rtx, rtx));
624 static int can_disregard_other_sets PARAMS ((struct reg_set **, rtx, int));
625 static int handle_avail_expr PARAMS ((rtx, struct expr *));
626 static int classic_gcse PARAMS ((void));
627 static int one_classic_gcse_pass PARAMS ((int));
628 static void invalidate_nonnull_info PARAMS ((rtx, rtx, void *));
629 static void delete_null_pointer_checks_1 PARAMS ((unsigned int *, sbitmap *,
631 struct null_pointer_info *));
632 static rtx process_insert_insn PARAMS ((struct expr *));
633 static int pre_edge_insert PARAMS ((struct edge_list *, struct expr **));
634 static int expr_reaches_here_p_work PARAMS ((struct occr *, struct expr *,
636 static int pre_expr_reaches_here_p_work PARAMS ((int, struct expr *,
639 /* Entry point for global common subexpression elimination.
640 F is the first instruction in the function. */
648 /* Bytes used at start of pass. */
649 int initial_bytes_used;
650 /* Maximum number of bytes used by a pass. */
652 /* Point to release obstack data from for each pass. */
653 char *gcse_obstack_bottom;
655 /* We do not construct an accurate cfg in functions which call
656 setjmp, so just punt to be safe. */
657 if (current_function_calls_setjmp)
660 /* Assume that we do not need to run jump optimizations after gcse. */
661 run_jump_opt_after_gcse = 0;
663 /* For calling dump_foo fns from gdb. */
664 debug_stderr = stderr;
667 /* Identify the basic block information for this function, including
668 successors and predecessors. */
669 max_gcse_regno = max_reg_num ();
672 dump_flow_info (file);
674 /* Return if there's nothing to do. */
675 if (n_basic_blocks <= 1)
678 /* Trying to perform global optimizations on flow graphs which have
679 a high connectivity will take a long time and is unlikely to be
682 In normal circumstances a cfg should have about twice has many edges
683 as blocks. But we do not want to punish small functions which have
684 a couple switch statements. So we require a relatively large number
685 of basic blocks and the ratio of edges to blocks to be high. */
686 if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
689 /* See what modes support reg/reg copy operations. */
690 if (! can_copy_init_p)
696 gcc_obstack_init (&gcse_obstack);
699 /* Record where pseudo-registers are set. This data is kept accurate
700 during each pass. ??? We could also record hard-reg information here
701 [since it's unchanging], however it is currently done during hash table
704 It may be tempting to compute MEM set information here too, but MEM sets
705 will be subject to code motion one day and thus we need to compute
706 information about memory sets when we build the hash tables. */
708 alloc_reg_set_mem (max_gcse_regno);
712 initial_bytes_used = bytes_used;
714 gcse_obstack_bottom = gcse_alloc (1);
716 while (changed && pass < MAX_PASSES)
720 fprintf (file, "GCSE pass %d\n\n", pass + 1);
722 /* Initialize bytes_used to the space for the pred/succ lists,
723 and the reg_set_table data. */
724 bytes_used = initial_bytes_used;
726 /* Each pass may create new registers, so recalculate each time. */
727 max_gcse_regno = max_reg_num ();
731 /* Don't allow constant propagation to modify jumps
733 changed = one_cprop_pass (pass + 1, 0);
736 changed |= one_classic_gcse_pass (pass + 1);
739 changed |= one_pre_gcse_pass (pass + 1);
741 alloc_reg_set_mem (max_reg_num ());
743 run_jump_opt_after_gcse = 1;
746 if (max_pass_bytes < bytes_used)
747 max_pass_bytes = bytes_used;
749 /* Free up memory, then reallocate for code hoisting. We can
750 not re-use the existing allocated memory because the tables
751 will not have info for the insns or registers created by
752 partial redundancy elimination. */
755 /* It does not make sense to run code hoisting unless we optimizing
756 for code size -- it rarely makes programs faster, and can make
757 them bigger if we did partial redundancy elimination (when optimizing
758 for space, we use a classic gcse algorithm instead of partial
759 redundancy algorithms). */
762 max_gcse_regno = max_reg_num ();
764 changed |= one_code_hoisting_pass ();
767 if (max_pass_bytes < bytes_used)
768 max_pass_bytes = bytes_used;
773 fprintf (file, "\n");
777 obstack_free (&gcse_obstack, gcse_obstack_bottom);
781 /* Do one last pass of copy propagation, including cprop into
782 conditional jumps. */
784 max_gcse_regno = max_reg_num ();
786 /* This time, go ahead and allow cprop to alter jumps. */
787 one_cprop_pass (pass + 1, 1);
792 fprintf (file, "GCSE of %s: %d basic blocks, ",
793 current_function_name, n_basic_blocks);
794 fprintf (file, "%d pass%s, %d bytes\n\n",
795 pass, pass > 1 ? "es" : "", max_pass_bytes);
798 obstack_free (&gcse_obstack, NULL_PTR);
800 return run_jump_opt_after_gcse;
803 /* Misc. utilities. */
805 /* Compute which modes support reg/reg copy operations. */
811 #ifndef AVOID_CCMODE_COPIES
814 char *free_point = (char *) oballoc (1);
816 bzero (can_copy_p, NUM_MACHINE_MODES);
819 for (i = 0; i < NUM_MACHINE_MODES; i++)
820 if (GET_MODE_CLASS (i) == MODE_CC)
822 #ifdef AVOID_CCMODE_COPIES
825 reg = gen_rtx_REG ((enum machine_mode) i, LAST_VIRTUAL_REGISTER + 1);
826 insn = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
827 if (recog (PATTERN (insn), insn, NULL_PTR) >= 0)
836 /* Free the objects we just allocated. */
840 /* Cover function to xmalloc to record bytes allocated. */
847 return xmalloc (size);
850 /* Cover function to xrealloc.
851 We don't record the additional size since we don't know it.
852 It won't affect memory usage stats much anyway. */
859 return xrealloc (ptr, size);
862 /* Cover function to obstack_alloc.
863 We don't need to record the bytes allocated here since
864 obstack_chunk_alloc is set to gmalloc. */
870 return (char *) obstack_alloc (&gcse_obstack, size);
873 /* Allocate memory for the cuid mapping array,
874 and reg/memory set tracking tables.
876 This is called at the start of each pass. */
885 /* Find the largest UID and create a mapping from UIDs to CUIDs.
886 CUIDs are like UIDs except they increase monotonically, have no gaps,
887 and only apply to real insns. */
889 max_uid = get_max_uid ();
890 n = (max_uid + 1) * sizeof (int);
891 uid_cuid = (int *) gmalloc (n);
892 bzero ((char *) uid_cuid, n);
893 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
896 uid_cuid[INSN_UID (insn)] = i++;
898 uid_cuid[INSN_UID (insn)] = i;
901 /* Create a table mapping cuids to insns. */
904 n = (max_cuid + 1) * sizeof (rtx);
905 cuid_insn = (rtx *) gmalloc (n);
906 bzero ((char *) cuid_insn, n);
907 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
909 CUID_INSN (i++) = insn;
911 /* Allocate vars to track sets of regs. */
912 reg_set_bitmap = (sbitmap) sbitmap_alloc (max_gcse_regno);
914 /* Allocate vars to track sets of regs, memory per block. */
915 reg_set_in_block = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks,
917 mem_set_in_block = (char *) gmalloc (n_basic_blocks);
920 /* Free memory allocated by alloc_gcse_mem. */
928 free (reg_set_bitmap);
930 free (reg_set_in_block);
931 free (mem_set_in_block);
934 /* Many of the global optimization algorithms work by solving dataflow
935 equations for various expressions. Initially, some local value is
936 computed for each expression in each block. Then, the values across the
937 various blocks are combined (by following flow graph edges) to arrive at
938 global values. Conceptually, each set of equations is independent. We
939 may therefore solve all the equations in parallel, solve them one at a
940 time, or pick any intermediate approach.
942 When you're going to need N two-dimensional bitmaps, each X (say, the
943 number of blocks) by Y (say, the number of expressions), call this
944 function. It's not important what X and Y represent; only that Y
945 correspond to the things that can be done in parallel. This function will
946 return an appropriate chunking factor C; you should solve C sets of
947 equations in parallel. By going through this function, we can easily
948 trade space against time; by solving fewer equations in parallel we use
952 get_bitmap_width (n, x, y)
957 /* It's not really worth figuring out *exactly* how much memory will
958 be used by a particular choice. The important thing is to get
959 something approximately right. */
960 size_t max_bitmap_memory = 10 * 1024 * 1024;
962 /* The number of bytes we'd use for a single column of minimum
964 size_t column_size = n * x * sizeof (SBITMAP_ELT_TYPE);
966 /* Often, it's reasonable just to solve all the equations in
968 if (column_size * SBITMAP_SET_SIZE (y) <= max_bitmap_memory)
971 /* Otherwise, pick the largest width we can, without going over the
973 return SBITMAP_ELT_BITS * ((max_bitmap_memory + column_size - 1)
977 /* Compute the local properties of each recorded expression.
979 Local properties are those that are defined by the block, irrespective of
982 An expression is transparent in a block if its operands are not modified
985 An expression is computed (locally available) in a block if it is computed
986 at least once and expression would contain the same value if the
987 computation was moved to the end of the block.
989 An expression is locally anticipatable in a block if it is computed at
990 least once and expression would contain the same value if the computation
991 was moved to the beginning of the block.
993 We call this routine for cprop, pre and code hoisting. They all compute
994 basically the same information and thus can easily share this code.
996 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
997 properties. If NULL, then it is not necessary to compute or record that
1000 SETP controls which hash table to look at. If zero, this routine looks at
1001 the expr hash table; if nonzero this routine looks at the set hash table.
1002 Additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
1006 compute_local_properties (transp, comp, antloc, setp)
1012 unsigned int i, hash_table_size;
1013 struct expr **hash_table;
1015 /* Initialize any bitmaps that were passed in. */
1019 sbitmap_vector_zero (transp, n_basic_blocks);
1021 sbitmap_vector_ones (transp, n_basic_blocks);
1025 sbitmap_vector_zero (comp, n_basic_blocks);
1027 sbitmap_vector_zero (antloc, n_basic_blocks);
1029 /* We use the same code for cprop, pre and hoisting. For cprop
1030 we care about the set hash table, for pre and hoisting we
1031 care about the expr hash table. */
1032 hash_table_size = setp ? set_hash_table_size : expr_hash_table_size;
1033 hash_table = setp ? set_hash_table : expr_hash_table;
1035 for (i = 0; i < hash_table_size; i++)
1039 for (expr = hash_table[i]; expr != NULL; expr = expr->next_same_hash)
1041 int indx = expr->bitmap_index;
1044 /* The expression is transparent in this block if it is not killed.
1045 We start by assuming all are transparent [none are killed], and
1046 then reset the bits for those that are. */
1048 compute_transp (expr->expr, indx, transp, setp);
1050 /* The occurrences recorded in antic_occr are exactly those that
1051 we want to set to non-zero in ANTLOC. */
1053 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
1055 SET_BIT (antloc[BLOCK_NUM (occr->insn)], indx);
1057 /* While we're scanning the table, this is a good place to
1059 occr->deleted_p = 0;
1062 /* The occurrences recorded in avail_occr are exactly those that
1063 we want to set to non-zero in COMP. */
1065 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
1067 SET_BIT (comp[BLOCK_NUM (occr->insn)], indx);
1069 /* While we're scanning the table, this is a good place to
1074 /* While we're scanning the table, this is a good place to
1076 expr->reaching_reg = 0;
1081 /* Register set information.
1083 `reg_set_table' records where each register is set or otherwise
1086 static struct obstack reg_set_obstack;
1089 alloc_reg_set_mem (n_regs)
1094 reg_set_table_size = n_regs + REG_SET_TABLE_SLOP;
1095 n = reg_set_table_size * sizeof (struct reg_set *);
1096 reg_set_table = (struct reg_set **) gmalloc (n);
1097 bzero ((char *) reg_set_table, n);
1099 gcc_obstack_init (®_set_obstack);
1105 free (reg_set_table);
1106 obstack_free (®_set_obstack, NULL_PTR);
1109 /* Record REGNO in the reg_set table. */
1112 record_one_set (regno, insn)
1116 /* allocate a new reg_set element and link it onto the list */
1117 struct reg_set *new_reg_info;
1119 /* If the table isn't big enough, enlarge it. */
1120 if (regno >= reg_set_table_size)
1122 int new_size = regno + REG_SET_TABLE_SLOP;
1125 = (struct reg_set **) grealloc ((char *) reg_set_table,
1126 new_size * sizeof (struct reg_set *));
1127 bzero ((char *) (reg_set_table + reg_set_table_size),
1128 (new_size - reg_set_table_size) * sizeof (struct reg_set *));
1129 reg_set_table_size = new_size;
1132 new_reg_info = (struct reg_set *) obstack_alloc (®_set_obstack,
1133 sizeof (struct reg_set));
1134 bytes_used += sizeof (struct reg_set);
1135 new_reg_info->insn = insn;
1136 new_reg_info->next = reg_set_table[regno];
1137 reg_set_table[regno] = new_reg_info;
1140 /* Called from compute_sets via note_stores to handle one SET or CLOBBER in
1141 an insn. The DATA is really the instruction in which the SET is
1145 record_set_info (dest, setter, data)
1146 rtx dest, setter ATTRIBUTE_UNUSED;
1149 rtx record_set_insn = (rtx) data;
1151 if (GET_CODE (dest) == REG && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
1152 record_one_set (REGNO (dest), record_set_insn);
1155 /* Scan the function and record each set of each pseudo-register.
1157 This is called once, at the start of the gcse pass. See the comments for
1158 `reg_set_table' for further documenation. */
1166 for (insn = f; insn != 0; insn = NEXT_INSN (insn))
1168 note_stores (PATTERN (insn), record_set_info, insn);
1171 /* Hash table support. */
1173 /* For each register, the cuid of the first/last insn in the block to set it,
1174 or -1 if not set. */
1175 #define NEVER_SET -1
1176 static int *reg_first_set;
1177 static int *reg_last_set;
1179 /* While computing "first/last set" info, this is the CUID of first/last insn
1180 to set memory or -1 if not set. `mem_last_set' is also used when
1181 performing GCSE to record whether memory has been set since the beginning
1184 Note that handling of memory is very simple, we don't make any attempt
1185 to optimize things (later). */
1186 static int mem_first_set;
1187 static int mem_last_set;
1189 /* Perform a quick check whether X, the source of a set, is something
1190 we want to consider for GCSE. */
1196 switch (GET_CODE (x))
1212 /* Return non-zero if the operands of expression X are unchanged from the
1213 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1214 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1217 oprs_unchanged_p (x, insn, avail_p)
1228 code = GET_CODE (x);
1233 return (reg_last_set[REGNO (x)] == NEVER_SET
1234 || reg_last_set[REGNO (x)] < INSN_CUID (insn));
1236 return (reg_first_set[REGNO (x)] == NEVER_SET
1237 || reg_first_set[REGNO (x)] >= INSN_CUID (insn));
1240 if (avail_p && mem_last_set != NEVER_SET
1241 && mem_last_set >= INSN_CUID (insn))
1243 else if (! avail_p && mem_first_set != NEVER_SET
1244 && mem_first_set < INSN_CUID (insn))
1247 return oprs_unchanged_p (XEXP (x, 0), insn, avail_p);
1272 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1276 /* If we are about to do the last recursive call needed at this
1277 level, change it into iteration. This function is called enough
1280 return oprs_unchanged_p (XEXP (x, i), insn, avail_p);
1282 else if (! oprs_unchanged_p (XEXP (x, i), insn, avail_p))
1285 else if (fmt[i] == 'E')
1286 for (j = 0; j < XVECLEN (x, i); j++)
1287 if (! oprs_unchanged_p (XVECEXP (x, i, j), insn, avail_p))
1294 /* Return non-zero if the operands of expression X are unchanged from
1295 the start of INSN's basic block up to but not including INSN. */
1298 oprs_anticipatable_p (x, insn)
1301 return oprs_unchanged_p (x, insn, 0);
1304 /* Return non-zero if the operands of expression X are unchanged from
1305 INSN to the end of INSN's basic block. */
1308 oprs_available_p (x, insn)
1311 return oprs_unchanged_p (x, insn, 1);
1314 /* Hash expression X.
1316 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1317 indicating if a volatile operand is found or if the expression contains
1318 something we don't want to insert in the table.
1320 ??? One might want to merge this with canon_hash. Later. */
1323 hash_expr (x, mode, do_not_record_p, hash_table_size)
1325 enum machine_mode mode;
1326 int *do_not_record_p;
1327 int hash_table_size;
1331 *do_not_record_p = 0;
1333 hash = hash_expr_1 (x, mode, do_not_record_p);
1334 return hash % hash_table_size;
1336 /* Hash a string. Just add its bytes up. */
1337 static inline unsigned
1342 const unsigned char *p = (const unsigned char *)ps;
1351 /* Subroutine of hash_expr to do the actual work. */
1354 hash_expr_1 (x, mode, do_not_record_p)
1356 enum machine_mode mode;
1357 int *do_not_record_p;
1364 /* Used to turn recursion into iteration. We can't rely on GCC's
1365 tail-recursion eliminatio since we need to keep accumulating values
1372 code = GET_CODE (x);
1376 hash += ((unsigned int) REG << 7) + REGNO (x);
1380 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
1381 + (unsigned int) INTVAL (x));
1385 /* This is like the general case, except that it only counts
1386 the integers representing the constant. */
1387 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
1388 if (GET_MODE (x) != VOIDmode)
1389 for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++)
1390 hash += (unsigned int) XWINT (x, i);
1392 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
1393 + (unsigned int) CONST_DOUBLE_HIGH (x));
1396 /* Assume there is only one rtx object for any given label. */
1398 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
1399 differences and differences between each stage's debugging dumps. */
1400 hash += (((unsigned int) LABEL_REF << 7)
1401 + CODE_LABEL_NUMBER (XEXP (x, 0)));
1406 /* Don't hash on the symbol's address to avoid bootstrap differences.
1407 Different hash values may cause expressions to be recorded in
1408 different orders and thus different registers to be used in the
1409 final assembler. This also avoids differences in the dump files
1410 between various stages. */
1412 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
1415 h += (h << 7) + *p++; /* ??? revisit */
1417 hash += ((unsigned int) SYMBOL_REF << 7) + h;
1422 if (MEM_VOLATILE_P (x))
1424 *do_not_record_p = 1;
1428 hash += (unsigned int) MEM;
1429 hash += MEM_ALIAS_SET (x);
1440 case UNSPEC_VOLATILE:
1441 *do_not_record_p = 1;
1445 if (MEM_VOLATILE_P (x))
1447 *do_not_record_p = 1;
1452 /* We don't want to take the filename and line into account. */
1453 hash += (unsigned) code + (unsigned) GET_MODE (x)
1454 + hash_string_1 (ASM_OPERANDS_TEMPLATE (x))
1455 + hash_string_1 (ASM_OPERANDS_OUTPUT_CONSTRAINT (x))
1456 + (unsigned) ASM_OPERANDS_OUTPUT_IDX (x);
1458 if (ASM_OPERANDS_INPUT_LENGTH (x))
1460 for (i = 1; i < ASM_OPERANDS_INPUT_LENGTH (x); i++)
1462 hash += (hash_expr_1 (ASM_OPERANDS_INPUT (x, i),
1463 GET_MODE (ASM_OPERANDS_INPUT (x, i)),
1465 + hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT
1469 hash += hash_string_1 (ASM_OPERANDS_INPUT_CONSTRAINT (x, 0));
1470 x = ASM_OPERANDS_INPUT (x, 0);
1471 mode = GET_MODE (x);
1481 hash += (unsigned) code + (unsigned) GET_MODE (x);
1482 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1486 /* If we are about to do the last recursive call
1487 needed at this level, change it into iteration.
1488 This function is called enough to be worth it. */
1495 hash += hash_expr_1 (XEXP (x, i), 0, do_not_record_p);
1496 if (*do_not_record_p)
1500 else if (fmt[i] == 'E')
1501 for (j = 0; j < XVECLEN (x, i); j++)
1503 hash += hash_expr_1 (XVECEXP (x, i, j), 0, do_not_record_p);
1504 if (*do_not_record_p)
1508 else if (fmt[i] == 's')
1509 hash += hash_string_1 (XSTR (x, i));
1510 else if (fmt[i] == 'i')
1511 hash += (unsigned int) XINT (x, i);
1519 /* Hash a set of register REGNO.
1521 Sets are hashed on the register that is set. This simplifies the PRE copy
1524 ??? May need to make things more elaborate. Later, as necessary. */
1527 hash_set (regno, hash_table_size)
1529 int hash_table_size;
1534 return hash % hash_table_size;
1537 /* Return non-zero if exp1 is equivalent to exp2.
1538 ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
1545 register enum rtx_code code;
1546 register const char *fmt;
1551 if (x == 0 || y == 0)
1554 code = GET_CODE (x);
1555 if (code != GET_CODE (y))
1558 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
1559 if (GET_MODE (x) != GET_MODE (y))
1569 return INTVAL (x) == INTVAL (y);
1572 return XEXP (x, 0) == XEXP (y, 0);
1575 return XSTR (x, 0) == XSTR (y, 0);
1578 return REGNO (x) == REGNO (y);
1581 /* Can't merge two expressions in different alias sets, since we can
1582 decide that the expression is transparent in a block when it isn't,
1583 due to it being set with the different alias set. */
1584 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
1588 /* For commutative operations, check both orders. */
1596 return ((expr_equiv_p (XEXP (x, 0), XEXP (y, 0))
1597 && expr_equiv_p (XEXP (x, 1), XEXP (y, 1)))
1598 || (expr_equiv_p (XEXP (x, 0), XEXP (y, 1))
1599 && expr_equiv_p (XEXP (x, 1), XEXP (y, 0))));
1602 /* We don't use the generic code below because we want to
1603 disregard filename and line numbers. */
1605 /* A volatile asm isn't equivalent to any other. */
1606 if (MEM_VOLATILE_P (x) || MEM_VOLATILE_P (y))
1609 if (GET_MODE (x) != GET_MODE (y)
1610 || strcmp (ASM_OPERANDS_TEMPLATE (x), ASM_OPERANDS_TEMPLATE (y))
1611 || strcmp (ASM_OPERANDS_OUTPUT_CONSTRAINT (x),
1612 ASM_OPERANDS_OUTPUT_CONSTRAINT (y))
1613 || ASM_OPERANDS_OUTPUT_IDX (x) != ASM_OPERANDS_OUTPUT_IDX (y)
1614 || ASM_OPERANDS_INPUT_LENGTH (x) != ASM_OPERANDS_INPUT_LENGTH (y))
1617 if (ASM_OPERANDS_INPUT_LENGTH (x))
1619 for (i = ASM_OPERANDS_INPUT_LENGTH (x) - 1; i >= 0; i--)
1620 if (! expr_equiv_p (ASM_OPERANDS_INPUT (x, i),
1621 ASM_OPERANDS_INPUT (y, i))
1622 || strcmp (ASM_OPERANDS_INPUT_CONSTRAINT (x, i),
1623 ASM_OPERANDS_INPUT_CONSTRAINT (y, i)))
1633 /* Compare the elements. If any pair of corresponding elements
1634 fail to match, return 0 for the whole thing. */
1636 fmt = GET_RTX_FORMAT (code);
1637 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1642 if (! expr_equiv_p (XEXP (x, i), XEXP (y, i)))
1647 if (XVECLEN (x, i) != XVECLEN (y, i))
1649 for (j = 0; j < XVECLEN (x, i); j++)
1650 if (! expr_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j)))
1655 if (strcmp (XSTR (x, i), XSTR (y, i)))
1660 if (XINT (x, i) != XINT (y, i))
1665 if (XWINT (x, i) != XWINT (y, i))
1680 /* Insert expression X in INSN in the hash table.
1681 If it is already present, record it as the last occurrence in INSN's
1684 MODE is the mode of the value X is being stored into.
1685 It is only used if X is a CONST_INT.
1687 ANTIC_P is non-zero if X is an anticipatable expression.
1688 AVAIL_P is non-zero if X is an available expression. */
1691 insert_expr_in_table (x, mode, insn, antic_p, avail_p)
1693 enum machine_mode mode;
1695 int antic_p, avail_p;
1697 int found, do_not_record_p;
1699 struct expr *cur_expr, *last_expr = NULL;
1700 struct occr *antic_occr, *avail_occr;
1701 struct occr *last_occr = NULL;
1703 hash = hash_expr (x, mode, &do_not_record_p, expr_hash_table_size);
1705 /* Do not insert expression in table if it contains volatile operands,
1706 or if hash_expr determines the expression is something we don't want
1707 to or can't handle. */
1708 if (do_not_record_p)
1711 cur_expr = expr_hash_table[hash];
1714 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1716 /* If the expression isn't found, save a pointer to the end of
1718 last_expr = cur_expr;
1719 cur_expr = cur_expr->next_same_hash;
1724 cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
1725 bytes_used += sizeof (struct expr);
1726 if (expr_hash_table[hash] == NULL)
1727 /* This is the first pattern that hashed to this index. */
1728 expr_hash_table[hash] = cur_expr;
1730 /* Add EXPR to end of this hash chain. */
1731 last_expr->next_same_hash = cur_expr;
1733 /* Set the fields of the expr element. */
1735 cur_expr->bitmap_index = n_exprs++;
1736 cur_expr->next_same_hash = NULL;
1737 cur_expr->antic_occr = NULL;
1738 cur_expr->avail_occr = NULL;
1741 /* Now record the occurrence(s). */
1744 antic_occr = cur_expr->antic_occr;
1746 /* Search for another occurrence in the same basic block. */
1747 while (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
1749 /* If an occurrence isn't found, save a pointer to the end of
1751 last_occr = antic_occr;
1752 antic_occr = antic_occr->next;
1756 /* Found another instance of the expression in the same basic block.
1757 Prefer the currently recorded one. We want the first one in the
1758 block and the block is scanned from start to end. */
1759 ; /* nothing to do */
1762 /* First occurrence of this expression in this basic block. */
1763 antic_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1764 bytes_used += sizeof (struct occr);
1765 /* First occurrence of this expression in any block? */
1766 if (cur_expr->antic_occr == NULL)
1767 cur_expr->antic_occr = antic_occr;
1769 last_occr->next = antic_occr;
1771 antic_occr->insn = insn;
1772 antic_occr->next = NULL;
1778 avail_occr = cur_expr->avail_occr;
1780 /* Search for another occurrence in the same basic block. */
1781 while (avail_occr && BLOCK_NUM (avail_occr->insn) != BLOCK_NUM (insn))
1783 /* If an occurrence isn't found, save a pointer to the end of
1785 last_occr = avail_occr;
1786 avail_occr = avail_occr->next;
1790 /* Found another instance of the expression in the same basic block.
1791 Prefer this occurrence to the currently recorded one. We want
1792 the last one in the block and the block is scanned from start
1794 avail_occr->insn = insn;
1797 /* First occurrence of this expression in this basic block. */
1798 avail_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1799 bytes_used += sizeof (struct occr);
1801 /* First occurrence of this expression in any block? */
1802 if (cur_expr->avail_occr == NULL)
1803 cur_expr->avail_occr = avail_occr;
1805 last_occr->next = avail_occr;
1807 avail_occr->insn = insn;
1808 avail_occr->next = NULL;
1813 /* Insert pattern X in INSN in the hash table.
1814 X is a SET of a reg to either another reg or a constant.
1815 If it is already present, record it as the last occurrence in INSN's
1819 insert_set_in_table (x, insn)
1825 struct expr *cur_expr, *last_expr = NULL;
1826 struct occr *cur_occr, *last_occr = NULL;
1828 if (GET_CODE (x) != SET
1829 || GET_CODE (SET_DEST (x)) != REG)
1832 hash = hash_set (REGNO (SET_DEST (x)), set_hash_table_size);
1834 cur_expr = set_hash_table[hash];
1837 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1839 /* If the expression isn't found, save a pointer to the end of
1841 last_expr = cur_expr;
1842 cur_expr = cur_expr->next_same_hash;
1847 cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
1848 bytes_used += sizeof (struct expr);
1849 if (set_hash_table[hash] == NULL)
1850 /* This is the first pattern that hashed to this index. */
1851 set_hash_table[hash] = cur_expr;
1853 /* Add EXPR to end of this hash chain. */
1854 last_expr->next_same_hash = cur_expr;
1856 /* Set the fields of the expr element.
1857 We must copy X because it can be modified when copy propagation is
1858 performed on its operands. */
1859 /* ??? Should this go in a different obstack? */
1860 cur_expr->expr = copy_rtx (x);
1861 cur_expr->bitmap_index = n_sets++;
1862 cur_expr->next_same_hash = NULL;
1863 cur_expr->antic_occr = NULL;
1864 cur_expr->avail_occr = NULL;
1867 /* Now record the occurrence. */
1868 cur_occr = cur_expr->avail_occr;
1870 /* Search for another occurrence in the same basic block. */
1871 while (cur_occr && BLOCK_NUM (cur_occr->insn) != BLOCK_NUM (insn))
1873 /* If an occurrence isn't found, save a pointer to the end of
1875 last_occr = cur_occr;
1876 cur_occr = cur_occr->next;
1880 /* Found another instance of the expression in the same basic block.
1881 Prefer this occurrence to the currently recorded one. We want the
1882 last one in the block and the block is scanned from start to end. */
1883 cur_occr->insn = insn;
1886 /* First occurrence of this expression in this basic block. */
1887 cur_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1888 bytes_used += sizeof (struct occr);
1890 /* First occurrence of this expression in any block? */
1891 if (cur_expr->avail_occr == NULL)
1892 cur_expr->avail_occr = cur_occr;
1894 last_occr->next = cur_occr;
1896 cur_occr->insn = insn;
1897 cur_occr->next = NULL;
1901 /* Scan pattern PAT of INSN and add an entry to the hash table. If SET_P is
1902 non-zero, this is for the assignment hash table, otherwise it is for the
1903 expression hash table. */
1906 hash_scan_set (pat, insn, set_p)
1910 rtx src = SET_SRC (pat);
1911 rtx dest = SET_DEST (pat);
1913 if (GET_CODE (src) == CALL)
1914 hash_scan_call (src, insn);
1916 if (GET_CODE (dest) == REG)
1918 int regno = REGNO (dest);
1921 /* Only record sets of pseudo-regs in the hash table. */
1923 && regno >= FIRST_PSEUDO_REGISTER
1924 /* Don't GCSE something if we can't do a reg/reg copy. */
1925 && can_copy_p [GET_MODE (dest)]
1926 /* Is SET_SRC something we want to gcse? */
1927 && want_to_gcse_p (src))
1929 /* An expression is not anticipatable if its operands are
1930 modified before this insn. */
1931 int antic_p = oprs_anticipatable_p (src, insn);
1932 /* An expression is not available if its operands are
1933 subsequently modified, including this insn. */
1934 int avail_p = oprs_available_p (src, insn);
1936 insert_expr_in_table (src, GET_MODE (dest), insn, antic_p, avail_p);
1939 /* Record sets for constant/copy propagation. */
1941 && regno >= FIRST_PSEUDO_REGISTER
1942 && ((GET_CODE (src) == REG
1943 && REGNO (src) >= FIRST_PSEUDO_REGISTER
1944 && can_copy_p [GET_MODE (dest)])
1945 || GET_CODE (src) == CONST_INT
1946 || GET_CODE (src) == SYMBOL_REF
1947 || GET_CODE (src) == CONST_DOUBLE)
1948 /* A copy is not available if its src or dest is subsequently
1949 modified. Here we want to search from INSN+1 on, but
1950 oprs_available_p searches from INSN on. */
1951 && (insn == BLOCK_END (BLOCK_NUM (insn))
1952 || ((tmp = next_nonnote_insn (insn)) != NULL_RTX
1953 && oprs_available_p (pat, tmp))))
1954 insert_set_in_table (pat, insn);
1959 hash_scan_clobber (x, insn)
1960 rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
1962 /* Currently nothing to do. */
1966 hash_scan_call (x, insn)
1967 rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
1969 /* Currently nothing to do. */
1972 /* Process INSN and add hash table entries as appropriate.
1974 Only available expressions that set a single pseudo-reg are recorded.
1976 Single sets in a PARALLEL could be handled, but it's an extra complication
1977 that isn't dealt with right now. The trick is handling the CLOBBERs that
1978 are also in the PARALLEL. Later.
1980 If SET_P is non-zero, this is for the assignment hash table,
1981 otherwise it is for the expression hash table.
1982 If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
1983 not record any expressions. */
1986 hash_scan_insn (insn, set_p, in_libcall_block)
1989 int in_libcall_block;
1991 rtx pat = PATTERN (insn);
1994 /* Pick out the sets of INSN and for other forms of instructions record
1995 what's been modified. */
1997 if (GET_CODE (pat) == SET && ! in_libcall_block)
1999 /* Ignore obvious no-ops. */
2000 if (SET_SRC (pat) != SET_DEST (pat))
2001 hash_scan_set (pat, insn, set_p);
2003 else if (GET_CODE (pat) == PARALLEL)
2004 for (i = 0; i < XVECLEN (pat, 0); i++)
2006 rtx x = XVECEXP (pat, 0, i);
2008 if (GET_CODE (x) == SET)
2010 if (GET_CODE (SET_SRC (x)) == CALL)
2011 hash_scan_call (SET_SRC (x), insn);
2013 else if (GET_CODE (x) == CLOBBER)
2014 hash_scan_clobber (x, insn);
2015 else if (GET_CODE (x) == CALL)
2016 hash_scan_call (x, insn);
2019 else if (GET_CODE (pat) == CLOBBER)
2020 hash_scan_clobber (pat, insn);
2021 else if (GET_CODE (pat) == CALL)
2022 hash_scan_call (pat, insn);
2026 dump_hash_table (file, name, table, table_size, total_size)
2029 struct expr **table;
2030 int table_size, total_size;
2033 /* Flattened out table, so it's printed in proper order. */
2034 struct expr **flat_table;
2035 unsigned int *hash_val;
2039 = (struct expr **) xcalloc (total_size, sizeof (struct expr *));
2040 hash_val = (unsigned int *) xmalloc (total_size * sizeof (unsigned int));
2042 for (i = 0; i < table_size; i++)
2043 for (expr = table[i]; expr != NULL; expr = expr->next_same_hash)
2045 flat_table[expr->bitmap_index] = expr;
2046 hash_val[expr->bitmap_index] = i;
2049 fprintf (file, "%s hash table (%d buckets, %d entries)\n",
2050 name, table_size, total_size);
2052 for (i = 0; i < total_size; i++)
2053 if (flat_table[i] != 0)
2055 expr = flat_table[i];
2056 fprintf (file, "Index %d (hash value %d)\n ",
2057 expr->bitmap_index, hash_val[i]);
2058 print_rtl (file, expr->expr);
2059 fprintf (file, "\n");
2062 fprintf (file, "\n");
2068 /* Record register first/last/block set information for REGNO in INSN.
2070 reg_first_set records the first place in the block where the register
2071 is set and is used to compute "anticipatability".
2073 reg_last_set records the last place in the block where the register
2074 is set and is used to compute "availability".
2076 reg_set_in_block records whether the register is set in the block
2077 and is used to compute "transparency". */
2080 record_last_reg_set_info (insn, regno)
2084 if (reg_first_set[regno] == NEVER_SET)
2085 reg_first_set[regno] = INSN_CUID (insn);
2087 reg_last_set[regno] = INSN_CUID (insn);
2088 SET_BIT (reg_set_in_block[BLOCK_NUM (insn)], regno);
2091 /* Record memory first/last/block set information for INSN. */
2094 record_last_mem_set_info (insn)
2097 if (mem_first_set == NEVER_SET)
2098 mem_first_set = INSN_CUID (insn);
2100 mem_last_set = INSN_CUID (insn);
2101 mem_set_in_block[BLOCK_NUM (insn)] = 1;
2104 /* Called from compute_hash_table via note_stores to handle one
2105 SET or CLOBBER in an insn. DATA is really the instruction in which
2106 the SET is taking place. */
2109 record_last_set_info (dest, setter, data)
2110 rtx dest, setter ATTRIBUTE_UNUSED;
2113 rtx last_set_insn = (rtx) data;
2115 if (GET_CODE (dest) == SUBREG)
2116 dest = SUBREG_REG (dest);
2118 if (GET_CODE (dest) == REG)
2119 record_last_reg_set_info (last_set_insn, REGNO (dest));
2120 else if (GET_CODE (dest) == MEM
2121 /* Ignore pushes, they clobber nothing. */
2122 && ! push_operand (dest, GET_MODE (dest)))
2123 record_last_mem_set_info (last_set_insn);
2126 /* Top level function to create an expression or assignment hash table.
2128 Expression entries are placed in the hash table if
2129 - they are of the form (set (pseudo-reg) src),
2130 - src is something we want to perform GCSE on,
2131 - none of the operands are subsequently modified in the block
2133 Assignment entries are placed in the hash table if
2134 - they are of the form (set (pseudo-reg) src),
2135 - src is something we want to perform const/copy propagation on,
2136 - none of the operands or target are subsequently modified in the block
2138 Currently src must be a pseudo-reg or a const_int.
2140 F is the first insn.
2141 SET_P is non-zero for computing the assignment hash table. */
2144 compute_hash_table (set_p)
2149 /* While we compute the hash table we also compute a bit array of which
2150 registers are set in which blocks.
2151 We also compute which blocks set memory, in the absence of aliasing
2152 support [which is TODO].
2153 ??? This isn't needed during const/copy propagation, but it's cheap to
2155 sbitmap_vector_zero (reg_set_in_block, n_basic_blocks);
2156 bzero ((char *) mem_set_in_block, n_basic_blocks);
2158 /* Some working arrays used to track first and last set in each block. */
2159 /* ??? One could use alloca here, but at some size a threshold is crossed
2160 beyond which one should use malloc. Are we at that threshold here? */
2161 reg_first_set = (int *) gmalloc (max_gcse_regno * sizeof (int));
2162 reg_last_set = (int *) gmalloc (max_gcse_regno * sizeof (int));
2164 for (bb = 0; bb < n_basic_blocks; bb++)
2168 int in_libcall_block;
2171 /* First pass over the instructions records information used to
2172 determine when registers and memory are first and last set.
2173 ??? The mem_set_in_block and hard-reg reg_set_in_block computation
2174 could be moved to compute_sets since they currently don't change. */
2176 for (i = 0; i < max_gcse_regno; i++)
2177 reg_first_set[i] = reg_last_set[i] = NEVER_SET;
2179 mem_first_set = NEVER_SET;
2180 mem_last_set = NEVER_SET;
2182 for (insn = BLOCK_HEAD (bb);
2183 insn && insn != NEXT_INSN (BLOCK_END (bb));
2184 insn = NEXT_INSN (insn))
2186 #ifdef NON_SAVING_SETJMP
2187 if (NON_SAVING_SETJMP && GET_CODE (insn) == NOTE
2188 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_SETJMP)
2190 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2191 record_last_reg_set_info (insn, regno);
2196 if (! INSN_P (insn))
2199 if (GET_CODE (insn) == CALL_INSN)
2201 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2202 if ((call_used_regs[regno]
2203 && regno != STACK_POINTER_REGNUM
2204 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2205 && regno != HARD_FRAME_POINTER_REGNUM
2207 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
2208 && ! (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
2210 #if defined (PIC_OFFSET_TABLE_REGNUM) && !defined (PIC_OFFSET_TABLE_REG_CALL_CLOBBERED)
2211 && ! (regno == PIC_OFFSET_TABLE_REGNUM && flag_pic)
2214 && regno != FRAME_POINTER_REGNUM)
2215 || global_regs[regno])
2216 record_last_reg_set_info (insn, regno);
2218 if (! CONST_CALL_P (insn))
2219 record_last_mem_set_info (insn);
2222 note_stores (PATTERN (insn), record_last_set_info, insn);
2225 /* The next pass builds the hash table. */
2227 for (insn = BLOCK_HEAD (bb), in_libcall_block = 0;
2228 insn && insn != NEXT_INSN (BLOCK_END (bb));
2229 insn = NEXT_INSN (insn))
2232 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
2233 in_libcall_block = 1;
2234 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
2235 in_libcall_block = 0;
2236 hash_scan_insn (insn, set_p, in_libcall_block);
2240 free (reg_first_set);
2241 free (reg_last_set);
2243 /* Catch bugs early. */
2244 reg_first_set = reg_last_set = 0;
2247 /* Allocate space for the set hash table.
2248 N_INSNS is the number of instructions in the function.
2249 It is used to determine the number of buckets to use. */
2252 alloc_set_hash_table (n_insns)
2257 set_hash_table_size = n_insns / 4;
2258 if (set_hash_table_size < 11)
2259 set_hash_table_size = 11;
2261 /* Attempt to maintain efficient use of hash table.
2262 Making it an odd number is simplest for now.
2263 ??? Later take some measurements. */
2264 set_hash_table_size |= 1;
2265 n = set_hash_table_size * sizeof (struct expr *);
2266 set_hash_table = (struct expr **) gmalloc (n);
2269 /* Free things allocated by alloc_set_hash_table. */
2272 free_set_hash_table ()
2274 free (set_hash_table);
2277 /* Compute the hash table for doing copy/const propagation. */
2280 compute_set_hash_table ()
2282 /* Initialize count of number of entries in hash table. */
2284 bzero ((char *) set_hash_table,
2285 set_hash_table_size * sizeof (struct expr *));
2287 compute_hash_table (1);
2290 /* Allocate space for the expression hash table.
2291 N_INSNS is the number of instructions in the function.
2292 It is used to determine the number of buckets to use. */
2295 alloc_expr_hash_table (n_insns)
2296 unsigned int n_insns;
2300 expr_hash_table_size = n_insns / 2;
2301 /* Make sure the amount is usable. */
2302 if (expr_hash_table_size < 11)
2303 expr_hash_table_size = 11;
2305 /* Attempt to maintain efficient use of hash table.
2306 Making it an odd number is simplest for now.
2307 ??? Later take some measurements. */
2308 expr_hash_table_size |= 1;
2309 n = expr_hash_table_size * sizeof (struct expr *);
2310 expr_hash_table = (struct expr **) gmalloc (n);
2313 /* Free things allocated by alloc_expr_hash_table. */
2316 free_expr_hash_table ()
2318 free (expr_hash_table);
2321 /* Compute the hash table for doing GCSE. */
2324 compute_expr_hash_table ()
2326 /* Initialize count of number of entries in hash table. */
2328 bzero ((char *) expr_hash_table,
2329 expr_hash_table_size * sizeof (struct expr *));
2331 compute_hash_table (0);
2334 /* Expression tracking support. */
2336 /* Lookup pattern PAT in the expression table.
2337 The result is a pointer to the table entry, or NULL if not found. */
2339 static struct expr *
2343 int do_not_record_p;
2344 unsigned int hash = hash_expr (pat, GET_MODE (pat), &do_not_record_p,
2345 expr_hash_table_size);
2348 if (do_not_record_p)
2351 expr = expr_hash_table[hash];
2353 while (expr && ! expr_equiv_p (expr->expr, pat))
2354 expr = expr->next_same_hash;
2359 /* Lookup REGNO in the set table. If PAT is non-NULL look for the entry that
2360 matches it, otherwise return the first entry for REGNO. The result is a
2361 pointer to the table entry, or NULL if not found. */
2363 static struct expr *
2364 lookup_set (regno, pat)
2368 unsigned int hash = hash_set (regno, set_hash_table_size);
2371 expr = set_hash_table[hash];
2375 while (expr && ! expr_equiv_p (expr->expr, pat))
2376 expr = expr->next_same_hash;
2380 while (expr && REGNO (SET_DEST (expr->expr)) != regno)
2381 expr = expr->next_same_hash;
2387 /* Return the next entry for REGNO in list EXPR. */
2389 static struct expr *
2390 next_set (regno, expr)
2395 expr = expr->next_same_hash;
2396 while (expr && REGNO (SET_DEST (expr->expr)) != regno);
2401 /* Reset tables used to keep track of what's still available [since the
2402 start of the block]. */
2405 reset_opr_set_tables ()
2407 /* Maintain a bitmap of which regs have been set since beginning of
2409 sbitmap_zero (reg_set_bitmap);
2411 /* Also keep a record of the last instruction to modify memory.
2412 For now this is very trivial, we only record whether any memory
2413 location has been modified. */
2417 /* Return non-zero if the operands of X are not set before INSN in
2418 INSN's basic block. */
2421 oprs_not_set_p (x, insn)
2431 code = GET_CODE (x);
2446 if (mem_last_set != 0)
2449 return oprs_not_set_p (XEXP (x, 0), insn);
2452 return ! TEST_BIT (reg_set_bitmap, REGNO (x));
2458 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2462 /* If we are about to do the last recursive call
2463 needed at this level, change it into iteration.
2464 This function is called enough to be worth it. */
2466 return oprs_not_set_p (XEXP (x, i), insn);
2468 if (! oprs_not_set_p (XEXP (x, i), insn))
2471 else if (fmt[i] == 'E')
2472 for (j = 0; j < XVECLEN (x, i); j++)
2473 if (! oprs_not_set_p (XVECEXP (x, i, j), insn))
2480 /* Mark things set by a CALL. */
2486 mem_last_set = INSN_CUID (insn);
2489 /* Mark things set by a SET. */
2492 mark_set (pat, insn)
2495 rtx dest = SET_DEST (pat);
2497 while (GET_CODE (dest) == SUBREG
2498 || GET_CODE (dest) == ZERO_EXTRACT
2499 || GET_CODE (dest) == SIGN_EXTRACT
2500 || GET_CODE (dest) == STRICT_LOW_PART)
2501 dest = XEXP (dest, 0);
2503 if (GET_CODE (dest) == REG)
2504 SET_BIT (reg_set_bitmap, REGNO (dest));
2505 else if (GET_CODE (dest) == MEM)
2506 mem_last_set = INSN_CUID (insn);
2508 if (GET_CODE (SET_SRC (pat)) == CALL)
2512 /* Record things set by a CLOBBER. */
2515 mark_clobber (pat, insn)
2518 rtx clob = XEXP (pat, 0);
2520 while (GET_CODE (clob) == SUBREG || GET_CODE (clob) == STRICT_LOW_PART)
2521 clob = XEXP (clob, 0);
2523 if (GET_CODE (clob) == REG)
2524 SET_BIT (reg_set_bitmap, REGNO (clob));
2526 mem_last_set = INSN_CUID (insn);
2529 /* Record things set by INSN.
2530 This data is used by oprs_not_set_p. */
2533 mark_oprs_set (insn)
2536 rtx pat = PATTERN (insn);
2539 if (GET_CODE (pat) == SET)
2540 mark_set (pat, insn);
2541 else if (GET_CODE (pat) == PARALLEL)
2542 for (i = 0; i < XVECLEN (pat, 0); i++)
2544 rtx x = XVECEXP (pat, 0, i);
2546 if (GET_CODE (x) == SET)
2548 else if (GET_CODE (x) == CLOBBER)
2549 mark_clobber (x, insn);
2550 else if (GET_CODE (x) == CALL)
2554 else if (GET_CODE (pat) == CLOBBER)
2555 mark_clobber (pat, insn);
2556 else if (GET_CODE (pat) == CALL)
2561 /* Classic GCSE reaching definition support. */
2563 /* Allocate reaching def variables. */
2566 alloc_rd_mem (n_blocks, n_insns)
2567 int n_blocks, n_insns;
2569 rd_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2570 sbitmap_vector_zero (rd_kill, n_basic_blocks);
2572 rd_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2573 sbitmap_vector_zero (rd_gen, n_basic_blocks);
2575 reaching_defs = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2576 sbitmap_vector_zero (reaching_defs, n_basic_blocks);
2578 rd_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2579 sbitmap_vector_zero (rd_out, n_basic_blocks);
2582 /* Free reaching def variables. */
2589 free (reaching_defs);
2593 /* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
2596 handle_rd_kill_set (insn, regno, bb)
2600 struct reg_set *this_reg;
2602 for (this_reg = reg_set_table[regno]; this_reg; this_reg = this_reg ->next)
2603 if (BLOCK_NUM (this_reg->insn) != BLOCK_NUM (insn))
2604 SET_BIT (rd_kill[bb], INSN_CUID (this_reg->insn));
2607 /* Compute the set of kill's for reaching definitions. */
2616 For each set bit in `gen' of the block (i.e each insn which
2617 generates a definition in the block)
2618 Call the reg set by the insn corresponding to that bit regx
2619 Look at the linked list starting at reg_set_table[regx]
2620 For each setting of regx in the linked list, which is not in
2622 Set the bit in `kill' corresponding to that insn. */
2623 for (bb = 0; bb < n_basic_blocks; bb++)
2624 for (cuid = 0; cuid < max_cuid; cuid++)
2625 if (TEST_BIT (rd_gen[bb], cuid))
2627 rtx insn = CUID_INSN (cuid);
2628 rtx pat = PATTERN (insn);
2630 if (GET_CODE (insn) == CALL_INSN)
2632 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2634 if ((call_used_regs[regno]
2635 && regno != STACK_POINTER_REGNUM
2636 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2637 && regno != HARD_FRAME_POINTER_REGNUM
2639 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
2640 && ! (regno == ARG_POINTER_REGNUM
2641 && fixed_regs[regno])
2643 #if defined (PIC_OFFSET_TABLE_REGNUM) && !defined (PIC_OFFSET_TABLE_REG_CALL_CLOBBERED)
2644 && ! (regno == PIC_OFFSET_TABLE_REGNUM && flag_pic)
2646 && regno != FRAME_POINTER_REGNUM)
2647 || global_regs[regno])
2648 handle_rd_kill_set (insn, regno, bb);
2652 if (GET_CODE (pat) == PARALLEL)
2654 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
2656 enum rtx_code code = GET_CODE (XVECEXP (pat, 0, i));
2658 if ((code == SET || code == CLOBBER)
2659 && GET_CODE (XEXP (XVECEXP (pat, 0, i), 0)) == REG)
2660 handle_rd_kill_set (insn,
2661 REGNO (XEXP (XVECEXP (pat, 0, i), 0)),
2665 else if (GET_CODE (pat) == SET && GET_CODE (SET_DEST (pat)) == REG)
2666 /* Each setting of this register outside of this block
2667 must be marked in the set of kills in this block. */
2668 handle_rd_kill_set (insn, REGNO (SET_DEST (pat)), bb);
2672 /* Compute the reaching definitions as in
2673 Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
2674 Chapter 10. It is the same algorithm as used for computing available
2675 expressions but applied to the gens and kills of reaching definitions. */
2680 int bb, changed, passes;
2682 for (bb = 0; bb < n_basic_blocks; bb++)
2683 sbitmap_copy (rd_out[bb] /*dst*/, rd_gen[bb] /*src*/);
2690 for (bb = 0; bb < n_basic_blocks; bb++)
2692 sbitmap_union_of_preds (reaching_defs[bb], rd_out, bb);
2693 changed |= sbitmap_union_of_diff (rd_out[bb], rd_gen[bb],
2694 reaching_defs[bb], rd_kill[bb]);
2700 fprintf (gcse_file, "reaching def computation: %d passes\n", passes);
2703 /* Classic GCSE available expression support. */
2705 /* Allocate memory for available expression computation. */
2708 alloc_avail_expr_mem (n_blocks, n_exprs)
2709 int n_blocks, n_exprs;
2711 ae_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2712 sbitmap_vector_zero (ae_kill, n_basic_blocks);
2714 ae_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2715 sbitmap_vector_zero (ae_gen, n_basic_blocks);
2717 ae_in = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2718 sbitmap_vector_zero (ae_in, n_basic_blocks);
2720 ae_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2721 sbitmap_vector_zero (ae_out, n_basic_blocks);
2725 free_avail_expr_mem ()
2733 /* Compute the set of available expressions generated in each basic block. */
2742 /* For each recorded occurrence of each expression, set ae_gen[bb][expr].
2743 This is all we have to do because an expression is not recorded if it
2744 is not available, and the only expressions we want to work with are the
2745 ones that are recorded. */
2746 for (i = 0; i < expr_hash_table_size; i++)
2747 for (expr = expr_hash_table[i]; expr != 0; expr = expr->next_same_hash)
2748 for (occr = expr->avail_occr; occr != 0; occr = occr->next)
2749 SET_BIT (ae_gen[BLOCK_NUM (occr->insn)], expr->bitmap_index);
2752 /* Return non-zero if expression X is killed in BB. */
2755 expr_killed_p (x, bb)
2766 code = GET_CODE (x);
2770 return TEST_BIT (reg_set_in_block[bb], REGNO (x));
2773 if (mem_set_in_block[bb])
2776 return expr_killed_p (XEXP (x, 0), bb);
2793 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2797 /* If we are about to do the last recursive call
2798 needed at this level, change it into iteration.
2799 This function is called enough to be worth it. */
2801 return expr_killed_p (XEXP (x, i), bb);
2802 else if (expr_killed_p (XEXP (x, i), bb))
2805 else if (fmt[i] == 'E')
2806 for (j = 0; j < XVECLEN (x, i); j++)
2807 if (expr_killed_p (XVECEXP (x, i, j), bb))
2814 /* Compute the set of available expressions killed in each basic block. */
2817 compute_ae_kill (ae_gen, ae_kill)
2818 sbitmap *ae_gen, *ae_kill;
2824 for (bb = 0; bb < n_basic_blocks; bb++)
2825 for (i = 0; i < expr_hash_table_size; i++)
2826 for (expr = expr_hash_table[i]; expr; expr = expr->next_same_hash)
2828 /* Skip EXPR if generated in this block. */
2829 if (TEST_BIT (ae_gen[bb], expr->bitmap_index))
2832 if (expr_killed_p (expr->expr, bb))
2833 SET_BIT (ae_kill[bb], expr->bitmap_index);
2837 /* Actually perform the Classic GCSE optimizations. */
2839 /* Return non-zero if occurrence OCCR of expression EXPR reaches block BB.
2841 CHECK_SELF_LOOP is non-zero if we should consider a block reaching itself
2842 as a positive reach. We want to do this when there are two computations
2843 of the expression in the block.
2845 VISITED is a pointer to a working buffer for tracking which BB's have
2846 been visited. It is NULL for the top-level call.
2848 We treat reaching expressions that go through blocks containing the same
2849 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
2850 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
2851 2 as not reaching. The intent is to improve the probability of finding
2852 only one reaching expression and to reduce register lifetimes by picking
2853 the closest such expression. */
2856 expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited)
2860 int check_self_loop;
2865 for (pred = BASIC_BLOCK(bb)->pred; pred != NULL; pred = pred->pred_next)
2867 int pred_bb = pred->src->index;
2869 if (visited[pred_bb])
2870 /* This predecessor has already been visited. Nothing to do. */
2872 else if (pred_bb == bb)
2874 /* BB loops on itself. */
2876 && TEST_BIT (ae_gen[pred_bb], expr->bitmap_index)
2877 && BLOCK_NUM (occr->insn) == pred_bb)
2880 visited[pred_bb] = 1;
2883 /* Ignore this predecessor if it kills the expression. */
2884 else if (TEST_BIT (ae_kill[pred_bb], expr->bitmap_index))
2885 visited[pred_bb] = 1;
2887 /* Does this predecessor generate this expression? */
2888 else if (TEST_BIT (ae_gen[pred_bb], expr->bitmap_index))
2890 /* Is this the occurrence we're looking for?
2891 Note that there's only one generating occurrence per block
2892 so we just need to check the block number. */
2893 if (BLOCK_NUM (occr->insn) == pred_bb)
2896 visited[pred_bb] = 1;
2899 /* Neither gen nor kill. */
2902 visited[pred_bb] = 1;
2903 if (expr_reaches_here_p_work (occr, expr, pred_bb, check_self_loop,
2910 /* All paths have been checked. */
2914 /* This wrapper for expr_reaches_here_p_work() is to ensure that any
2915 memory allocated for that function is returned. */
2918 expr_reaches_here_p (occr, expr, bb, check_self_loop)
2922 int check_self_loop;
2925 char *visited = (char *) xcalloc (n_basic_blocks, 1);
2927 rval = expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited);
2933 /* Return the instruction that computes EXPR that reaches INSN's basic block.
2934 If there is more than one such instruction, return NULL.
2936 Called only by handle_avail_expr. */
2939 computing_insn (expr, insn)
2943 int bb = BLOCK_NUM (insn);
2945 if (expr->avail_occr->next == NULL)
2947 if (BLOCK_NUM (expr->avail_occr->insn) == bb)
2948 /* The available expression is actually itself
2949 (i.e. a loop in the flow graph) so do nothing. */
2952 /* (FIXME) Case that we found a pattern that was created by
2953 a substitution that took place. */
2954 return expr->avail_occr->insn;
2958 /* Pattern is computed more than once.
2959 Search backwards from this insn to see how many of these
2960 computations actually reach this insn. */
2962 rtx insn_computes_expr = NULL;
2965 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
2967 if (BLOCK_NUM (occr->insn) == bb)
2969 /* The expression is generated in this block.
2970 The only time we care about this is when the expression
2971 is generated later in the block [and thus there's a loop].
2972 We let the normal cse pass handle the other cases. */
2973 if (INSN_CUID (insn) < INSN_CUID (occr->insn)
2974 && expr_reaches_here_p (occr, expr, bb, 1))
2980 insn_computes_expr = occr->insn;
2983 else if (expr_reaches_here_p (occr, expr, bb, 0))
2989 insn_computes_expr = occr->insn;
2993 if (insn_computes_expr == NULL)
2996 return insn_computes_expr;
3000 /* Return non-zero if the definition in DEF_INSN can reach INSN.
3001 Only called by can_disregard_other_sets. */
3004 def_reaches_here_p (insn, def_insn)
3009 if (TEST_BIT (reaching_defs[BLOCK_NUM (insn)], INSN_CUID (def_insn)))
3012 if (BLOCK_NUM (insn) == BLOCK_NUM (def_insn))
3014 if (INSN_CUID (def_insn) < INSN_CUID (insn))
3016 if (GET_CODE (PATTERN (def_insn)) == PARALLEL)
3018 else if (GET_CODE (PATTERN (def_insn)) == CLOBBER)
3019 reg = XEXP (PATTERN (def_insn), 0);
3020 else if (GET_CODE (PATTERN (def_insn)) == SET)
3021 reg = SET_DEST (PATTERN (def_insn));
3025 return ! reg_set_between_p (reg, NEXT_INSN (def_insn), insn);
3034 /* Return non-zero if *ADDR_THIS_REG can only have one value at INSN. The
3035 value returned is the number of definitions that reach INSN. Returning a
3036 value of zero means that [maybe] more than one definition reaches INSN and
3037 the caller can't perform whatever optimization it is trying. i.e. it is
3038 always safe to return zero. */
3041 can_disregard_other_sets (addr_this_reg, insn, for_combine)
3042 struct reg_set **addr_this_reg;
3046 int number_of_reaching_defs = 0;
3047 struct reg_set *this_reg;
3049 for (this_reg = *addr_this_reg; this_reg != 0; this_reg = this_reg->next)
3050 if (def_reaches_here_p (insn, this_reg->insn))
3052 number_of_reaching_defs++;
3053 /* Ignore parallels for now. */
3054 if (GET_CODE (PATTERN (this_reg->insn)) == PARALLEL)
3058 && (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER
3059 || ! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
3060 SET_SRC (PATTERN (insn)))))
3061 /* A setting of the reg to a different value reaches INSN. */
3064 if (number_of_reaching_defs > 1)
3066 /* If in this setting the value the register is being set to is
3067 equal to the previous value the register was set to and this
3068 setting reaches the insn we are trying to do the substitution
3069 on then we are ok. */
3070 if (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER)
3072 else if (! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
3073 SET_SRC (PATTERN (insn))))
3077 *addr_this_reg = this_reg;
3080 return number_of_reaching_defs;
3083 /* Expression computed by insn is available and the substitution is legal,
3084 so try to perform the substitution.
3086 The result is non-zero if any changes were made. */
3089 handle_avail_expr (insn, expr)
3093 rtx pat, insn_computes_expr;
3095 struct reg_set *this_reg;
3096 int found_setting, use_src;
3099 /* We only handle the case where one computation of the expression
3100 reaches this instruction. */
3101 insn_computes_expr = computing_insn (expr, insn);
3102 if (insn_computes_expr == NULL)
3108 /* At this point we know only one computation of EXPR outside of this
3109 block reaches this insn. Now try to find a register that the
3110 expression is computed into. */
3111 if (GET_CODE (SET_SRC (PATTERN (insn_computes_expr))) == REG)
3113 /* This is the case when the available expression that reaches
3114 here has already been handled as an available expression. */
3115 unsigned int regnum_for_replacing
3116 = REGNO (SET_SRC (PATTERN (insn_computes_expr)));
3118 /* If the register was created by GCSE we can't use `reg_set_table',
3119 however we know it's set only once. */
3120 if (regnum_for_replacing >= max_gcse_regno
3121 /* If the register the expression is computed into is set only once,
3122 or only one set reaches this insn, we can use it. */
3123 || (((this_reg = reg_set_table[regnum_for_replacing]),
3124 this_reg->next == NULL)
3125 || can_disregard_other_sets (&this_reg, insn, 0)))
3134 unsigned int regnum_for_replacing
3135 = REGNO (SET_DEST (PATTERN (insn_computes_expr)));
3137 /* This shouldn't happen. */
3138 if (regnum_for_replacing >= max_gcse_regno)
3141 this_reg = reg_set_table[regnum_for_replacing];
3143 /* If the register the expression is computed into is set only once,
3144 or only one set reaches this insn, use it. */
3145 if (this_reg->next == NULL
3146 || can_disregard_other_sets (&this_reg, insn, 0))
3152 pat = PATTERN (insn);
3154 to = SET_SRC (PATTERN (insn_computes_expr));
3156 to = SET_DEST (PATTERN (insn_computes_expr));
3157 changed = validate_change (insn, &SET_SRC (pat), to, 0);
3159 /* We should be able to ignore the return code from validate_change but
3160 to play it safe we check. */
3164 if (gcse_file != NULL)
3166 fprintf (gcse_file, "GCSE: Replacing the source in insn %d with",
3168 fprintf (gcse_file, " reg %d %s insn %d\n",
3169 REGNO (to), use_src ? "from" : "set in",
3170 INSN_UID (insn_computes_expr));
3175 /* The register that the expr is computed into is set more than once. */
3176 else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
3178 /* Insert an insn after insnx that copies the reg set in insnx
3179 into a new pseudo register call this new register REGN.
3180 From insnb until end of basic block or until REGB is set
3181 replace all uses of REGB with REGN. */
3184 to = gen_reg_rtx (GET_MODE (SET_DEST (PATTERN (insn_computes_expr))));
3186 /* Generate the new insn. */
3187 /* ??? If the change fails, we return 0, even though we created
3188 an insn. I think this is ok. */
3190 = emit_insn_after (gen_rtx_SET (VOIDmode, to,
3192 (insn_computes_expr))),
3193 insn_computes_expr);
3195 /* Keep block number table up to date. */
3196 set_block_num (new_insn, BLOCK_NUM (insn_computes_expr));
3198 /* Keep register set table up to date. */
3199 record_one_set (REGNO (to), new_insn);
3201 gcse_create_count++;
3202 if (gcse_file != NULL)
3204 fprintf (gcse_file, "GCSE: Creating insn %d to copy value of reg %d",
3205 INSN_UID (NEXT_INSN (insn_computes_expr)),
3206 REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr)))));
3207 fprintf (gcse_file, ", computed in insn %d,\n",
3208 INSN_UID (insn_computes_expr));
3209 fprintf (gcse_file, " into newly allocated reg %d\n",
3213 pat = PATTERN (insn);
3215 /* Do register replacement for INSN. */
3216 changed = validate_change (insn, &SET_SRC (pat),
3218 (NEXT_INSN (insn_computes_expr))),
3221 /* We should be able to ignore the return code from validate_change but
3222 to play it safe we check. */
3226 if (gcse_file != NULL)
3229 "GCSE: Replacing the source in insn %d with reg %d ",
3231 REGNO (SET_DEST (PATTERN (NEXT_INSN
3232 (insn_computes_expr)))));
3233 fprintf (gcse_file, "set in insn %d\n",
3234 INSN_UID (insn_computes_expr));
3242 /* Perform classic GCSE. This is called by one_classic_gcse_pass after all
3243 the dataflow analysis has been done.
3245 The result is non-zero if a change was made. */
3253 /* Note we start at block 1. */
3256 for (bb = 1; bb < n_basic_blocks; bb++)
3258 /* Reset tables used to keep track of what's still valid [since the
3259 start of the block]. */
3260 reset_opr_set_tables ();
3262 for (insn = BLOCK_HEAD (bb);
3263 insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
3264 insn = NEXT_INSN (insn))
3266 /* Is insn of form (set (pseudo-reg) ...)? */
3267 if (GET_CODE (insn) == INSN
3268 && GET_CODE (PATTERN (insn)) == SET
3269 && GET_CODE (SET_DEST (PATTERN (insn))) == REG
3270 && REGNO (SET_DEST (PATTERN (insn))) >= FIRST_PSEUDO_REGISTER)
3272 rtx pat = PATTERN (insn);
3273 rtx src = SET_SRC (pat);
3276 if (want_to_gcse_p (src)
3277 /* Is the expression recorded? */
3278 && ((expr = lookup_expr (src)) != NULL)
3279 /* Is the expression available [at the start of the
3281 && TEST_BIT (ae_in[bb], expr->bitmap_index)
3282 /* Are the operands unchanged since the start of the
3284 && oprs_not_set_p (src, insn))
3285 changed |= handle_avail_expr (insn, expr);
3288 /* Keep track of everything modified by this insn. */
3289 /* ??? Need to be careful w.r.t. mods done to INSN. */
3291 mark_oprs_set (insn);
3298 /* Top level routine to perform one classic GCSE pass.
3300 Return non-zero if a change was made. */
3303 one_classic_gcse_pass (pass)
3308 gcse_subst_count = 0;
3309 gcse_create_count = 0;
3311 alloc_expr_hash_table (max_cuid);
3312 alloc_rd_mem (n_basic_blocks, max_cuid);
3313 compute_expr_hash_table ();
3315 dump_hash_table (gcse_file, "Expression", expr_hash_table,
3316 expr_hash_table_size, n_exprs);
3322 alloc_avail_expr_mem (n_basic_blocks, n_exprs);
3324 compute_ae_kill (ae_gen, ae_kill);
3325 compute_available (ae_gen, ae_kill, ae_out, ae_in);
3326 changed = classic_gcse ();
3327 free_avail_expr_mem ();
3331 free_expr_hash_table ();
3335 fprintf (gcse_file, "\n");
3336 fprintf (gcse_file, "GCSE of %s, pass %d: %d bytes needed, %d substs,",
3337 current_function_name, pass, bytes_used, gcse_subst_count);
3338 fprintf (gcse_file, "%d insns created\n", gcse_create_count);
3344 /* Compute copy/constant propagation working variables. */
3346 /* Local properties of assignments. */
3347 static sbitmap *cprop_pavloc;
3348 static sbitmap *cprop_absaltered;
3350 /* Global properties of assignments (computed from the local properties). */
3351 static sbitmap *cprop_avin;
3352 static sbitmap *cprop_avout;
3354 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
3355 basic blocks. N_SETS is the number of sets. */
3358 alloc_cprop_mem (n_blocks, n_sets)
3359 int n_blocks, n_sets;
3361 cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
3362 cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
3364 cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
3365 cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
3368 /* Free vars used by copy/const propagation. */
3373 free (cprop_pavloc);
3374 free (cprop_absaltered);
3379 /* For each block, compute whether X is transparent. X is either an
3380 expression or an assignment [though we don't care which, for this context
3381 an assignment is treated as an expression]. For each block where an
3382 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
3386 compute_transp (x, indx, bmap, set_p)
3397 /* repeat is used to turn tail-recursion into iteration since GCC
3398 can't do it when there's no return value. */
3404 code = GET_CODE (x);
3410 if (REGNO (x) < FIRST_PSEUDO_REGISTER)
3412 for (bb = 0; bb < n_basic_blocks; bb++)
3413 if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
3414 SET_BIT (bmap[bb], indx);
3418 for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
3419 SET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
3424 if (REGNO (x) < FIRST_PSEUDO_REGISTER)
3426 for (bb = 0; bb < n_basic_blocks; bb++)
3427 if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
3428 RESET_BIT (bmap[bb], indx);
3432 for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
3433 RESET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
3442 for (bb = 0; bb < n_basic_blocks; bb++)
3443 if (mem_set_in_block[bb])
3444 SET_BIT (bmap[bb], indx);
3448 for (bb = 0; bb < n_basic_blocks; bb++)
3449 if (mem_set_in_block[bb])
3450 RESET_BIT (bmap[bb], indx);
3471 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3475 /* If we are about to do the last recursive call
3476 needed at this level, change it into iteration.
3477 This function is called enough to be worth it. */
3484 compute_transp (XEXP (x, i), indx, bmap, set_p);
3486 else if (fmt[i] == 'E')
3487 for (j = 0; j < XVECLEN (x, i); j++)
3488 compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
3492 /* Top level routine to do the dataflow analysis needed by copy/const
3496 compute_cprop_data ()
3498 compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, 1);
3499 compute_available (cprop_pavloc, cprop_absaltered,
3500 cprop_avout, cprop_avin);
3503 /* Copy/constant propagation. */
3505 /* Maximum number of register uses in an insn that we handle. */
3508 /* Table of uses found in an insn.
3509 Allocated statically to avoid alloc/free complexity and overhead. */
3510 static struct reg_use reg_use_table[MAX_USES];
3512 /* Index into `reg_use_table' while building it. */
3513 static int reg_use_count;
3515 /* Set up a list of register numbers used in INSN. The found uses are stored
3516 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
3517 and contains the number of uses in the table upon exit.
3519 ??? If a register appears multiple times we will record it multiple times.
3520 This doesn't hurt anything but it will slow things down. */
3530 /* repeat is used to turn tail-recursion into iteration since GCC
3531 can't do it when there's no return value. */
3537 code = GET_CODE (x);
3541 if (reg_use_count == MAX_USES)
3544 reg_use_table[reg_use_count].reg_rtx = x;
3562 case ASM_INPUT: /*FIXME*/
3566 if (GET_CODE (SET_DEST (x)) == MEM)
3567 find_used_regs (SET_DEST (x));
3575 /* Recursively scan the operands of this expression. */
3577 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3581 /* If we are about to do the last recursive call
3582 needed at this level, change it into iteration.
3583 This function is called enough to be worth it. */
3590 find_used_regs (XEXP (x, i));
3592 else if (fmt[i] == 'E')
3593 for (j = 0; j < XVECLEN (x, i); j++)
3594 find_used_regs (XVECEXP (x, i, j));
3598 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
3599 Returns non-zero is successful. */
3602 try_replace_reg (from, to, insn)
3610 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3613 note = find_reg_note (insn, REG_EQUIV, NULL_RTX);
3615 /* If this fails we could try to simplify the result of the
3616 replacement and attempt to recognize the simplified insn.
3618 But we need a general simplify_rtx that doesn't have pass
3619 specific state variables. I'm not aware of one at the moment. */
3621 success = validate_replace_src (from, to, insn);
3622 set = single_set (insn);
3624 /* We've failed to do replacement. Try to add REG_EQUAL note to not loose
3626 if (!success && !note)
3631 note = REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL,
3632 copy_rtx (SET_SRC (set)),
3636 /* Always do the replacement in REQ_EQUAL and REG_EQUIV notes. Also
3637 try to simplify them. */
3642 if (!validate_replace_rtx_subexp (from, to, insn, &XEXP (note, 0)))
3645 src = XEXP (note, 0);
3647 /* Try to simplify resulting note. */
3648 simplified = simplify_rtx (src);
3652 XEXP (note, 0) = src;
3655 /* REG_EQUAL may get simplified into register.
3656 We don't allow that. Remove that note. This code ought
3657 not to hapen, because previous code ought to syntetize
3658 reg-reg move, but be on the safe side. */
3659 else if (REG_P (src))
3660 remove_note (insn, note);
3665 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
3666 NULL no such set is found. */
3668 static struct expr *
3669 find_avail_set (regno, insn)
3673 /* SET1 contains the last set found that can be returned to the caller for
3674 use in a substitution. */
3675 struct expr *set1 = 0;
3677 /* Loops are not possible here. To get a loop we would need two sets
3678 available at the start of the block containing INSN. ie we would
3679 need two sets like this available at the start of the block:
3681 (set (reg X) (reg Y))
3682 (set (reg Y) (reg X))
3684 This can not happen since the set of (reg Y) would have killed the
3685 set of (reg X) making it unavailable at the start of this block. */
3689 struct expr *set = lookup_set (regno, NULL_RTX);
3691 /* Find a set that is available at the start of the block
3692 which contains INSN. */
3695 if (TEST_BIT (cprop_avin[BLOCK_NUM (insn)], set->bitmap_index))
3697 set = next_set (regno, set);
3700 /* If no available set was found we've reached the end of the
3701 (possibly empty) copy chain. */
3705 if (GET_CODE (set->expr) != SET)
3708 src = SET_SRC (set->expr);
3710 /* We know the set is available.
3711 Now check that SRC is ANTLOC (i.e. none of the source operands
3712 have changed since the start of the block).
3714 If the source operand changed, we may still use it for the next
3715 iteration of this loop, but we may not use it for substitutions. */
3717 if (CONSTANT_P (src) || oprs_not_set_p (src, insn))
3720 /* If the source of the set is anything except a register, then
3721 we have reached the end of the copy chain. */
3722 if (GET_CODE (src) != REG)
3725 /* Follow the copy chain, ie start another iteration of the loop
3726 and see if we have an available copy into SRC. */
3727 regno = REGNO (src);
3730 /* SET1 holds the last set that was available and anticipatable at
3735 /* Subroutine of cprop_insn that tries to propagate constants into
3736 JUMP_INSNS. INSN must be a conditional jump; COPY is a copy of it
3737 that we can use for substitutions.
3738 REG_USED is the use we will try to replace, SRC is the constant we
3739 will try to substitute for it.
3740 Returns nonzero if a change was made. */
3743 cprop_jump (insn, copy, reg_used, src)
3745 struct reg_use *reg_used;
3748 rtx set = PATTERN (copy);
3751 /* Replace the register with the appropriate constant. */
3752 replace_rtx (SET_SRC (set), reg_used->reg_rtx, src);
3754 temp = simplify_ternary_operation (GET_CODE (SET_SRC (set)),
3755 GET_MODE (SET_SRC (set)),
3756 GET_MODE (XEXP (SET_SRC (set), 0)),
3757 XEXP (SET_SRC (set), 0),
3758 XEXP (SET_SRC (set), 1),
3759 XEXP (SET_SRC (set), 2));
3761 /* If no simplification can be made, then try the next
3766 SET_SRC (set) = temp;
3768 /* That may have changed the structure of TEMP, so
3769 force it to be rerecognized if it has not turned
3770 into a nop or unconditional jump. */
3772 INSN_CODE (copy) = -1;
3773 if ((SET_DEST (set) == pc_rtx
3774 && (SET_SRC (set) == pc_rtx
3775 || GET_CODE (SET_SRC (set)) == LABEL_REF))
3776 || recog (PATTERN (copy), copy, NULL) >= 0)
3778 /* This has either become an unconditional jump
3779 or a nop-jump. We'd like to delete nop jumps
3780 here, but doing so confuses gcse. So we just
3781 make the replacement and let later passes
3783 PATTERN (insn) = set;
3784 INSN_CODE (insn) = -1;
3786 /* One less use of the label this insn used to jump to
3787 if we turned this into a NOP jump. */
3788 if (SET_SRC (set) == pc_rtx && JUMP_LABEL (insn) != 0)
3789 --LABEL_NUSES (JUMP_LABEL (insn));
3791 /* If this has turned into an unconditional jump,
3792 then put a barrier after it so that the unreachable
3793 code will be deleted. */
3794 if (GET_CODE (SET_SRC (set)) == LABEL_REF)
3795 emit_barrier_after (insn);
3797 run_jump_opt_after_gcse = 1;
3800 if (gcse_file != NULL)
3803 "CONST-PROP: Replacing reg %d in insn %d with constant ",
3804 REGNO (reg_used->reg_rtx), INSN_UID (insn));
3805 print_rtl (gcse_file, src);
3806 fprintf (gcse_file, "\n");
3816 /* Subroutine of cprop_insn that tries to propagate constants into JUMP_INSNS
3817 for machines that have CC0. INSN is a single set that stores into CC0;
3818 the insn following it is a conditional jump. REG_USED is the use we will
3819 try to replace, SRC is the constant we will try to substitute for it.
3820 Returns nonzero if a change was made. */
3823 cprop_cc0_jump (insn, reg_used, src)
3825 struct reg_use *reg_used;
3828 rtx jump = NEXT_INSN (insn);
3829 rtx copy = copy_rtx (jump);
3830 rtx set = PATTERN (copy);
3832 /* We need to copy the source of the cc0 setter, as cprop_jump is going to
3833 substitute into it. */
3834 replace_rtx (SET_SRC (set), cc0_rtx, copy_rtx (SET_SRC (PATTERN (insn))));
3835 if (! cprop_jump (jump, copy, reg_used, src))
3838 /* If we succeeded, delete the cc0 setter. */
3839 PUT_CODE (insn, NOTE);
3840 NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
3841 NOTE_SOURCE_FILE (insn) = 0;
3846 /* Perform constant and copy propagation on INSN.
3847 The result is non-zero if a change was made. */
3850 cprop_insn (insn, alter_jumps)
3854 struct reg_use *reg_used;
3858 /* Only propagate into SETs. Note that a conditional jump is a
3859 SET with pc_rtx as the destination. */
3860 if ((GET_CODE (insn) != INSN
3861 && GET_CODE (insn) != JUMP_INSN)
3862 || GET_CODE (PATTERN (insn)) != SET)
3866 find_used_regs (PATTERN (insn));
3868 note = find_reg_note (insn, REG_EQUIV, NULL_RTX);
3870 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3872 /* We may win even when propagating constants into notes. */
3874 find_used_regs (XEXP (note, 0));
3876 for (reg_used = ®_use_table[0]; reg_use_count > 0;
3877 reg_used++, reg_use_count--)
3879 unsigned int regno = REGNO (reg_used->reg_rtx);
3883 /* Ignore registers created by GCSE.
3884 We do this because ... */
3885 if (regno >= max_gcse_regno)
3888 /* If the register has already been set in this block, there's
3889 nothing we can do. */
3890 if (! oprs_not_set_p (reg_used->reg_rtx, insn))
3893 /* Find an assignment that sets reg_used and is available
3894 at the start of the block. */
3895 set = find_avail_set (regno, insn);
3900 /* ??? We might be able to handle PARALLELs. Later. */
3901 if (GET_CODE (pat) != SET)
3904 src = SET_SRC (pat);
3906 /* Constant propagation. */
3907 if (GET_CODE (src) == CONST_INT || GET_CODE (src) == CONST_DOUBLE
3908 || GET_CODE (src) == SYMBOL_REF)
3910 /* Handle normal insns first. */
3911 if (GET_CODE (insn) == INSN
3912 && try_replace_reg (reg_used->reg_rtx, src, insn))
3916 if (gcse_file != NULL)
3918 fprintf (gcse_file, "CONST-PROP: Replacing reg %d in ",
3920 fprintf (gcse_file, "insn %d with constant ",
3922 print_rtl (gcse_file, src);
3923 fprintf (gcse_file, "\n");
3926 /* The original insn setting reg_used may or may not now be
3927 deletable. We leave the deletion to flow. */
3930 /* Try to propagate a CONST_INT into a conditional jump.
3931 We're pretty specific about what we will handle in this
3932 code, we can extend this as necessary over time.
3934 Right now the insn in question must look like
3935 (set (pc) (if_then_else ...)) */
3936 else if (alter_jumps
3937 && GET_CODE (insn) == JUMP_INSN
3938 && condjump_p (insn)
3939 && ! simplejump_p (insn))
3940 changed |= cprop_jump (insn, copy_rtx (insn), reg_used, 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 changed |= cprop_cc0_jump (insn, reg_used, src);
3953 else if (GET_CODE (src) == REG
3954 && REGNO (src) >= FIRST_PSEUDO_REGISTER
3955 && REGNO (src) != regno)
3957 if (try_replace_reg (reg_used->reg_rtx, src, insn))
3961 if (gcse_file != NULL)
3963 fprintf (gcse_file, "COPY-PROP: Replacing reg %d in insn %d",
3964 regno, INSN_UID (insn));
3965 fprintf (gcse_file, " with reg %d\n", REGNO (src));
3968 /* The original insn setting reg_used may or may not now be
3969 deletable. We leave the deletion to flow. */
3970 /* FIXME: If it turns out that the insn isn't deletable,
3971 then we may have unnecessarily extended register lifetimes
3972 and made things worse. */
3980 /* Forward propagate copies. This includes copies and constants. Return
3981 non-zero if a change was made. */
3990 /* Note we start at block 1. */
3993 for (bb = 1; bb < n_basic_blocks; bb++)
3995 /* Reset tables used to keep track of what's still valid [since the
3996 start of the block]. */
3997 reset_opr_set_tables ();
3999 for (insn = BLOCK_HEAD (bb);
4000 insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
4001 insn = NEXT_INSN (insn))
4005 changed |= cprop_insn (insn, alter_jumps);
4007 /* Keep track of everything modified by this insn. */
4008 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
4009 call mark_oprs_set if we turned the insn into a NOTE. */
4010 if (GET_CODE (insn) != NOTE)
4011 mark_oprs_set (insn);
4016 if (gcse_file != NULL)
4017 fprintf (gcse_file, "\n");
4022 /* Perform one copy/constant propagation pass.
4023 F is the first insn in the function.
4024 PASS is the pass count. */
4027 one_cprop_pass (pass, alter_jumps)
4033 const_prop_count = 0;
4034 copy_prop_count = 0;
4036 alloc_set_hash_table (max_cuid);
4037 compute_set_hash_table ();
4039 dump_hash_table (gcse_file, "SET", set_hash_table, set_hash_table_size,
4043 alloc_cprop_mem (n_basic_blocks, n_sets);
4044 compute_cprop_data ();
4045 changed = cprop (alter_jumps);
4049 free_set_hash_table ();
4053 fprintf (gcse_file, "CPROP of %s, pass %d: %d bytes needed, ",
4054 current_function_name, pass, bytes_used);
4055 fprintf (gcse_file, "%d const props, %d copy props\n\n",
4056 const_prop_count, copy_prop_count);
4062 /* Compute PRE+LCM working variables. */
4064 /* Local properties of expressions. */
4065 /* Nonzero for expressions that are transparent in the block. */
4066 static sbitmap *transp;
4068 /* Nonzero for expressions that are transparent at the end of the block.
4069 This is only zero for expressions killed by abnormal critical edge
4070 created by a calls. */
4071 static sbitmap *transpout;
4073 /* Nonzero for expressions that are computed (available) in the block. */
4074 static sbitmap *comp;
4076 /* Nonzero for expressions that are locally anticipatable in the block. */
4077 static sbitmap *antloc;
4079 /* Nonzero for expressions where this block is an optimal computation
4081 static sbitmap *pre_optimal;
4083 /* Nonzero for expressions which are redundant in a particular block. */
4084 static sbitmap *pre_redundant;
4086 /* Nonzero for expressions which should be inserted on a specific edge. */
4087 static sbitmap *pre_insert_map;
4089 /* Nonzero for expressions which should be deleted in a specific block. */
4090 static sbitmap *pre_delete_map;
4092 /* Contains the edge_list returned by pre_edge_lcm. */
4093 static struct edge_list *edge_list;
4095 /* Redundant insns. */
4096 static sbitmap pre_redundant_insns;
4098 /* Allocate vars used for PRE analysis. */
4101 alloc_pre_mem (n_blocks, n_exprs)
4102 int n_blocks, n_exprs;
4104 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
4105 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
4106 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
4109 pre_redundant = NULL;
4110 pre_insert_map = NULL;
4111 pre_delete_map = NULL;
4114 ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
4116 /* pre_insert and pre_delete are allocated later. */
4119 /* Free vars used for PRE analysis. */
4127 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
4132 free (pre_redundant);
4134 free (pre_insert_map);
4136 free (pre_delete_map);
4143 transp = comp = NULL;
4144 pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
4145 ae_in = ae_out = NULL;
4148 /* Top level routine to do the dataflow analysis needed by PRE. */
4155 compute_local_properties (transp, comp, antloc, 0);
4156 sbitmap_vector_zero (ae_kill, n_basic_blocks);
4158 /* Compute ae_kill for each basic block using:
4162 This is significantly faster than compute_ae_kill. */
4164 for (i = 0; i < n_basic_blocks; i++)
4166 sbitmap_a_or_b (ae_kill[i], transp[i], comp[i]);
4167 sbitmap_not (ae_kill[i], ae_kill[i]);
4170 edge_list = pre_edge_lcm (gcse_file, n_exprs, transp, comp, antloc,
4171 ae_kill, &pre_insert_map, &pre_delete_map);
4180 /* Return non-zero if an occurrence of expression EXPR in OCCR_BB would reach
4183 VISITED is a pointer to a working buffer for tracking which BB's have
4184 been visited. It is NULL for the top-level call.
4186 We treat reaching expressions that go through blocks containing the same
4187 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
4188 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
4189 2 as not reaching. The intent is to improve the probability of finding
4190 only one reaching expression and to reduce register lifetimes by picking
4191 the closest such expression. */
4194 pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited)
4202 for (pred = BASIC_BLOCK (bb)->pred; pred != NULL; pred = pred->pred_next)
4204 int pred_bb = pred->src->index;
4206 if (pred->src == ENTRY_BLOCK_PTR
4207 /* Has predecessor has already been visited? */
4208 || visited[pred_bb])
4209 ;/* Nothing to do. */
4211 /* Does this predecessor generate this expression? */
4212 else if (TEST_BIT (comp[pred_bb], expr->bitmap_index))
4214 /* Is this the occurrence we're looking for?
4215 Note that there's only one generating occurrence per block
4216 so we just need to check the block number. */
4217 if (occr_bb == pred_bb)
4220 visited[pred_bb] = 1;
4222 /* Ignore this predecessor if it kills the expression. */
4223 else if (! TEST_BIT (transp[pred_bb], expr->bitmap_index))
4224 visited[pred_bb] = 1;
4226 /* Neither gen nor kill. */
4229 visited[pred_bb] = 1;
4230 if (pre_expr_reaches_here_p_work (occr_bb, expr, pred_bb, visited))
4235 /* All paths have been checked. */
4239 /* The wrapper for pre_expr_reaches_here_work that ensures that any
4240 memory allocated for that function is returned. */
4243 pre_expr_reaches_here_p (occr_bb, expr, bb)
4249 char *visited = (char *) xcalloc (n_basic_blocks, 1);
4251 rval = pre_expr_reaches_here_p_work(occr_bb, expr, bb, visited);
4258 /* Given an expr, generate RTL which we can insert at the end of a BB,
4259 or on an edge. Set the block number of any insns generated to
4263 process_insert_insn (expr)
4266 rtx reg = expr->reaching_reg;
4267 rtx pat, copied_expr;
4271 copied_expr = copy_rtx (expr->expr);
4272 emit_move_insn (reg, copied_expr);
4273 first_new_insn = get_insns ();
4274 pat = gen_sequence ();
4280 /* Add EXPR to the end of basic block BB.
4282 This is used by both the PRE and code hoisting.
4284 For PRE, we want to verify that the expr is either transparent
4285 or locally anticipatable in the target block. This check makes
4286 no sense for code hoisting. */
4289 insert_insn_end_bb (expr, bb, pre)
4294 rtx insn = BLOCK_END (bb);
4296 rtx reg = expr->reaching_reg;
4297 int regno = REGNO (reg);
4301 pat = process_insert_insn (expr);
4303 /* If the last insn is a jump, insert EXPR in front [taking care to
4304 handle cc0, etc. properly]. */
4306 if (GET_CODE (insn) == JUMP_INSN)
4312 /* If this is a jump table, then we can't insert stuff here. Since
4313 we know the previous real insn must be the tablejump, we insert
4314 the new instruction just before the tablejump. */
4315 if (GET_CODE (PATTERN (insn)) == ADDR_VEC
4316 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
4317 insn = prev_real_insn (insn);
4320 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
4321 if cc0 isn't set. */
4322 note = find_reg_note (insn, REG_CC_SETTER, NULL_RTX);
4324 insn = XEXP (note, 0);
4327 rtx maybe_cc0_setter = prev_nonnote_insn (insn);
4328 if (maybe_cc0_setter
4329 && INSN_P (maybe_cc0_setter)
4330 && sets_cc0_p (PATTERN (maybe_cc0_setter)))
4331 insn = maybe_cc0_setter;
4334 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4335 new_insn = emit_block_insn_before (pat, insn, BASIC_BLOCK (bb));
4338 /* Likewise if the last insn is a call, as will happen in the presence
4339 of exception handling. */
4340 else if (GET_CODE (insn) == CALL_INSN)
4342 HARD_REG_SET parm_regs;
4346 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4347 we search backward and place the instructions before the first
4348 parameter is loaded. Do this for everyone for consistency and a
4349 presumtion that we'll get better code elsewhere as well.
4351 It should always be the case that we can put these instructions
4352 anywhere in the basic block with performing PRE optimizations.
4356 && !TEST_BIT (antloc[bb], expr->bitmap_index)
4357 && !TEST_BIT (transp[bb], expr->bitmap_index))
4360 /* Since different machines initialize their parameter registers
4361 in different orders, assume nothing. Collect the set of all
4362 parameter registers. */
4363 CLEAR_HARD_REG_SET (parm_regs);
4365 for (p = CALL_INSN_FUNCTION_USAGE (insn); p ; p = XEXP (p, 1))
4366 if (GET_CODE (XEXP (p, 0)) == USE
4367 && GET_CODE (XEXP (XEXP (p, 0), 0)) == REG)
4369 if (REGNO (XEXP (XEXP (p, 0), 0)) >= FIRST_PSEUDO_REGISTER)
4372 SET_HARD_REG_BIT (parm_regs, REGNO (XEXP (XEXP (p, 0), 0)));
4376 /* Search backward for the first set of a register in this set. */
4377 while (nparm_regs && BLOCK_HEAD (bb) != insn)
4379 insn = PREV_INSN (insn);
4380 p = single_set (insn);
4381 if (p && GET_CODE (SET_DEST (p)) == REG
4382 && REGNO (SET_DEST (p)) < FIRST_PSEUDO_REGISTER
4383 && TEST_HARD_REG_BIT (parm_regs, REGNO (SET_DEST (p))))
4385 CLEAR_HARD_REG_BIT (parm_regs, REGNO (SET_DEST (p)));
4390 /* If we found all the parameter loads, then we want to insert
4391 before the first parameter load.
4393 If we did not find all the parameter loads, then we might have
4394 stopped on the head of the block, which could be a CODE_LABEL.
4395 If we inserted before the CODE_LABEL, then we would be putting
4396 the insn in the wrong basic block. In that case, put the insn
4397 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4398 while (GET_CODE (insn) == CODE_LABEL
4399 || NOTE_INSN_BASIC_BLOCK_P (insn))
4400 insn = NEXT_INSN (insn);
4402 new_insn = emit_block_insn_before (pat, insn, BASIC_BLOCK (bb));
4406 new_insn = emit_insn_after (pat, insn);
4407 BLOCK_END (bb) = new_insn;
4410 /* Keep block number table up to date.
4411 Note, PAT could be a multiple insn sequence, we have to make
4412 sure that each insn in the sequence is handled. */
4413 if (GET_CODE (pat) == SEQUENCE)
4415 for (i = 0; i < XVECLEN (pat, 0); i++)
4417 rtx insn = XVECEXP (pat, 0, i);
4419 set_block_num (insn, bb);
4421 add_label_notes (PATTERN (insn), new_insn);
4423 note_stores (PATTERN (insn), record_set_info, insn);
4428 add_label_notes (SET_SRC (pat), new_insn);
4429 set_block_num (new_insn, bb);
4431 /* Keep register set table up to date. */
4432 record_one_set (regno, new_insn);
4435 gcse_create_count++;
4439 fprintf (gcse_file, "PRE/HOIST: end of bb %d, insn %d, ",
4440 bb, INSN_UID (new_insn));
4441 fprintf (gcse_file, "copying expression %d to reg %d\n",
4442 expr->bitmap_index, regno);
4446 /* Insert partially redundant expressions on edges in the CFG to make
4447 the expressions fully redundant. */
4450 pre_edge_insert (edge_list, index_map)
4451 struct edge_list *edge_list;
4452 struct expr **index_map;
4454 int e, i, j, num_edges, set_size, did_insert = 0;
4457 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4458 if it reaches any of the deleted expressions. */
4460 set_size = pre_insert_map[0]->size;
4461 num_edges = NUM_EDGES (edge_list);
4462 inserted = sbitmap_vector_alloc (num_edges, n_exprs);
4463 sbitmap_vector_zero (inserted, num_edges);
4465 for (e = 0; e < num_edges; e++)
4468 basic_block pred = INDEX_EDGE_PRED_BB (edge_list, e);
4469 int bb = pred->index;
4471 for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
4473 SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
4475 for (j = indx; insert && j < n_exprs; j++, insert >>= 1)
4476 if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
4478 struct expr *expr = index_map[j];
4481 /* Now look at each deleted occurence of this expression. */
4482 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4484 if (! occr->deleted_p)
4487 /* Insert this expression on this edge if if it would
4488 reach the deleted occurence in BB. */
4489 if (!TEST_BIT (inserted[e], j))
4492 edge eg = INDEX_EDGE (edge_list, e);
4494 /* We can't insert anything on an abnormal and
4495 critical edge, so we insert the insn at the end of
4496 the previous block. There are several alternatives
4497 detailed in Morgans book P277 (sec 10.5) for
4498 handling this situation. This one is easiest for
4501 if ((eg->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
4502 insert_insn_end_bb (index_map[j], bb, 0);
4505 insn = process_insert_insn (index_map[j]);
4506 insert_insn_on_edge (insn, eg);
4511 fprintf (gcse_file, "PRE/HOIST: edge (%d,%d), ",
4513 INDEX_EDGE_SUCC_BB (edge_list, e)->index);
4514 fprintf (gcse_file, "copy expression %d\n",
4515 expr->bitmap_index);
4518 SET_BIT (inserted[e], j);
4520 gcse_create_count++;
4531 /* Copy the result of INSN to REG. INDX is the expression number. */
4534 pre_insert_copy_insn (expr, insn)
4538 rtx reg = expr->reaching_reg;
4539 int regno = REGNO (reg);
4540 int indx = expr->bitmap_index;
4541 rtx set = single_set (insn);
4543 int bb = BLOCK_NUM (insn);
4548 new_insn = emit_insn_after (gen_rtx_SET (VOIDmode, reg, SET_DEST (set)),
4551 /* Keep block number table up to date. */
4552 set_block_num (new_insn, bb);
4554 /* Keep register set table up to date. */
4555 record_one_set (regno, new_insn);
4556 if (insn == BLOCK_END (bb))
4557 BLOCK_END (bb) = new_insn;
4559 gcse_create_count++;
4563 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4564 BLOCK_NUM (insn), INSN_UID (new_insn), indx,
4565 INSN_UID (insn), regno);
4568 /* Copy available expressions that reach the redundant expression
4569 to `reaching_reg'. */
4572 pre_insert_copies ()
4579 /* For each available expression in the table, copy the result to
4580 `reaching_reg' if the expression reaches a deleted one.
4582 ??? The current algorithm is rather brute force.
4583 Need to do some profiling. */
4585 for (i = 0; i < expr_hash_table_size; i++)
4586 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4588 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4589 we don't want to insert a copy here because the expression may not
4590 really be redundant. So only insert an insn if the expression was
4591 deleted. This test also avoids further processing if the
4592 expression wasn't deleted anywhere. */
4593 if (expr->reaching_reg == NULL)
4596 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4598 if (! occr->deleted_p)
4601 for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
4603 rtx insn = avail->insn;
4605 /* No need to handle this one if handled already. */
4606 if (avail->copied_p)
4609 /* Don't handle this one if it's a redundant one. */
4610 if (TEST_BIT (pre_redundant_insns, INSN_CUID (insn)))
4613 /* Or if the expression doesn't reach the deleted one. */
4614 if (! pre_expr_reaches_here_p (BLOCK_NUM (avail->insn), expr,
4615 BLOCK_NUM (occr->insn)))
4618 /* Copy the result of avail to reaching_reg. */
4619 pre_insert_copy_insn (expr, insn);
4620 avail->copied_p = 1;
4626 /* Delete redundant computations.
4627 Deletion is done by changing the insn to copy the `reaching_reg' of
4628 the expression into the result of the SET. It is left to later passes
4629 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4631 Returns non-zero if a change is made. */
4642 for (i = 0; i < expr_hash_table_size; i++)
4643 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4645 int indx = expr->bitmap_index;
4647 /* We only need to search antic_occr since we require
4650 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4652 rtx insn = occr->insn;
4654 int bb = BLOCK_NUM (insn);
4656 if (TEST_BIT (pre_delete_map[bb], indx))
4658 set = single_set (insn);
4662 /* Create a pseudo-reg to store the result of reaching
4663 expressions into. Get the mode for the new pseudo from
4664 the mode of the original destination pseudo. */
4665 if (expr->reaching_reg == NULL)
4667 = gen_reg_rtx (GET_MODE (SET_DEST (set)));
4669 /* In theory this should never fail since we're creating
4672 However, on the x86 some of the movXX patterns actually
4673 contain clobbers of scratch regs. This may cause the
4674 insn created by validate_change to not match any pattern
4675 and thus cause validate_change to fail. */
4676 if (validate_change (insn, &SET_SRC (set),
4677 expr->reaching_reg, 0))
4679 occr->deleted_p = 1;
4680 SET_BIT (pre_redundant_insns, INSN_CUID (insn));
4688 "PRE: redundant insn %d (expression %d) in ",
4689 INSN_UID (insn), indx);
4690 fprintf (gcse_file, "bb %d, reaching reg is %d\n",
4691 bb, REGNO (expr->reaching_reg));
4700 /* Perform GCSE optimizations using PRE.
4701 This is called by one_pre_gcse_pass after all the dataflow analysis
4704 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4705 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4706 Compiler Design and Implementation.
4708 ??? A new pseudo reg is created to hold the reaching expression. The nice
4709 thing about the classical approach is that it would try to use an existing
4710 reg. If the register can't be adequately optimized [i.e. we introduce
4711 reload problems], one could add a pass here to propagate the new register
4714 ??? We don't handle single sets in PARALLELs because we're [currently] not
4715 able to copy the rest of the parallel when we insert copies to create full
4716 redundancies from partial redundancies. However, there's no reason why we
4717 can't handle PARALLELs in the cases where there are no partial
4724 int did_insert, changed;
4725 struct expr **index_map;
4728 /* Compute a mapping from expression number (`bitmap_index') to
4729 hash table entry. */
4731 index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
4732 for (i = 0; i < expr_hash_table_size; i++)
4733 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4734 index_map[expr->bitmap_index] = expr;
4736 /* Reset bitmap used to track which insns are redundant. */
4737 pre_redundant_insns = sbitmap_alloc (max_cuid);
4738 sbitmap_zero (pre_redundant_insns);
4740 /* Delete the redundant insns first so that
4741 - we know what register to use for the new insns and for the other
4742 ones with reaching expressions
4743 - we know which insns are redundant when we go to create copies */
4745 changed = pre_delete ();
4747 did_insert = pre_edge_insert (edge_list, index_map);
4749 /* In other places with reaching expressions, copy the expression to the
4750 specially allocated pseudo-reg that reaches the redundant expr. */
4751 pre_insert_copies ();
4754 commit_edge_insertions ();
4759 free (pre_redundant_insns);
4763 /* Top level routine to perform one PRE GCSE pass.
4765 Return non-zero if a change was made. */
4768 one_pre_gcse_pass (pass)
4773 gcse_subst_count = 0;
4774 gcse_create_count = 0;
4776 alloc_expr_hash_table (max_cuid);
4777 add_noreturn_fake_exit_edges ();
4778 compute_expr_hash_table ();
4780 dump_hash_table (gcse_file, "Expression", expr_hash_table,
4781 expr_hash_table_size, n_exprs);
4785 alloc_pre_mem (n_basic_blocks, n_exprs);
4786 compute_pre_data ();
4787 changed |= pre_gcse ();
4788 free_edge_list (edge_list);
4792 remove_fake_edges ();
4793 free_expr_hash_table ();
4797 fprintf (gcse_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
4798 current_function_name, pass, bytes_used);
4799 fprintf (gcse_file, "%d substs, %d insns created\n",
4800 gcse_subst_count, gcse_create_count);
4806 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to INSN.
4807 We have to add REG_LABEL notes, because the following loop optimization
4808 pass requires them. */
4810 /* ??? This is very similar to the loop.c add_label_notes function. We
4811 could probably share code here. */
4813 /* ??? If there was a jump optimization pass after gcse and before loop,
4814 then we would not need to do this here, because jump would add the
4815 necessary REG_LABEL notes. */
4818 add_label_notes (x, insn)
4822 enum rtx_code code = GET_CODE (x);
4826 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
4828 /* This code used to ignore labels that referred to dispatch tables to
4829 avoid flow generating (slighly) worse code.
4831 We no longer ignore such label references (see LABEL_REF handling in
4832 mark_jump_label for additional information). */
4834 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
4839 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
4842 add_label_notes (XEXP (x, i), insn);
4843 else if (fmt[i] == 'E')
4844 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4845 add_label_notes (XVECEXP (x, i, j), insn);
4849 /* Compute transparent outgoing information for each block.
4851 An expression is transparent to an edge unless it is killed by
4852 the edge itself. This can only happen with abnormal control flow,
4853 when the edge is traversed through a call. This happens with
4854 non-local labels and exceptions.
4856 This would not be necessary if we split the edge. While this is
4857 normally impossible for abnormal critical edges, with some effort
4858 it should be possible with exception handling, since we still have
4859 control over which handler should be invoked. But due to increased
4860 EH table sizes, this may not be worthwhile. */
4863 compute_transpout ()
4869 sbitmap_vector_ones (transpout, n_basic_blocks);
4871 for (bb = 0; bb < n_basic_blocks; ++bb)
4873 /* Note that flow inserted a nop a the end of basic blocks that
4874 end in call instructions for reasons other than abnormal
4876 if (GET_CODE (BLOCK_END (bb)) != CALL_INSN)
4879 for (i = 0; i < expr_hash_table_size; i++)
4880 for (expr = expr_hash_table[i]; expr ; expr = expr->next_same_hash)
4881 if (GET_CODE (expr->expr) == MEM)
4883 if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
4884 && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
4887 /* ??? Optimally, we would use interprocedural alias
4888 analysis to determine if this mem is actually killed
4890 RESET_BIT (transpout[bb], expr->bitmap_index);
4895 /* Removal of useless null pointer checks */
4897 /* Called via note_stores. X is set by SETTER. If X is a register we must
4898 invalidate nonnull_local and set nonnull_killed. DATA is really a
4899 `null_pointer_info *'.
4901 We ignore hard registers. */
4904 invalidate_nonnull_info (x, setter, data)
4906 rtx setter ATTRIBUTE_UNUSED;
4910 struct null_pointer_info *npi = (struct null_pointer_info *) data;
4912 while (GET_CODE (x) == SUBREG)
4915 /* Ignore anything that is not a register or is a hard register. */
4916 if (GET_CODE (x) != REG
4917 || REGNO (x) < npi->min_reg
4918 || REGNO (x) >= npi->max_reg)
4921 regno = REGNO (x) - npi->min_reg;
4923 RESET_BIT (npi->nonnull_local[npi->current_block], regno);
4924 SET_BIT (npi->nonnull_killed[npi->current_block], regno);
4927 /* Do null-pointer check elimination for the registers indicated in
4928 NPI. NONNULL_AVIN and NONNULL_AVOUT are pre-allocated sbitmaps;
4929 they are not our responsibility to free. */
4932 delete_null_pointer_checks_1 (block_reg, nonnull_avin, nonnull_avout, npi)
4933 unsigned int *block_reg;
4934 sbitmap *nonnull_avin;
4935 sbitmap *nonnull_avout;
4936 struct null_pointer_info *npi;
4940 sbitmap *nonnull_local = npi->nonnull_local;
4941 sbitmap *nonnull_killed = npi->nonnull_killed;
4943 /* Compute local properties, nonnull and killed. A register will have
4944 the nonnull property if at the end of the current block its value is
4945 known to be nonnull. The killed property indicates that somewhere in
4946 the block any information we had about the register is killed.
4948 Note that a register can have both properties in a single block. That
4949 indicates that it's killed, then later in the block a new value is
4951 sbitmap_vector_zero (nonnull_local, n_basic_blocks);
4952 sbitmap_vector_zero (nonnull_killed, n_basic_blocks);
4954 for (current_block = 0; current_block < n_basic_blocks; current_block++)
4956 rtx insn, stop_insn;
4958 /* Set the current block for invalidate_nonnull_info. */
4959 npi->current_block = current_block;
4961 /* Scan each insn in the basic block looking for memory references and
4963 stop_insn = NEXT_INSN (BLOCK_END (current_block));
4964 for (insn = BLOCK_HEAD (current_block);
4966 insn = NEXT_INSN (insn))
4971 /* Ignore anything that is not a normal insn. */
4972 if (! INSN_P (insn))
4975 /* Basically ignore anything that is not a simple SET. We do have
4976 to make sure to invalidate nonnull_local and set nonnull_killed
4977 for such insns though. */
4978 set = single_set (insn);
4981 note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
4985 /* See if we've got a useable memory load. We handle it first
4986 in case it uses its address register as a dest (which kills
4987 the nonnull property). */
4988 if (GET_CODE (SET_SRC (set)) == MEM
4989 && GET_CODE ((reg = XEXP (SET_SRC (set), 0))) == REG
4990 && REGNO (reg) >= npi->min_reg
4991 && REGNO (reg) < npi->max_reg)
4992 SET_BIT (nonnull_local[current_block],
4993 REGNO (reg) - npi->min_reg);
4995 /* Now invalidate stuff clobbered by this insn. */
4996 note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
4998 /* And handle stores, we do these last since any sets in INSN can
4999 not kill the nonnull property if it is derived from a MEM
5000 appearing in a SET_DEST. */
5001 if (GET_CODE (SET_DEST (set)) == MEM
5002 && GET_CODE ((reg = XEXP (SET_DEST (set), 0))) == REG
5003 && REGNO (reg) >= npi->min_reg
5004 && REGNO (reg) < npi->max_reg)
5005 SET_BIT (nonnull_local[current_block],
5006 REGNO (reg) - npi->min_reg);
5010 /* Now compute global properties based on the local properties. This
5011 is a classic global availablity algorithm. */
5012 compute_available (nonnull_local, nonnull_killed,
5013 nonnull_avout, nonnull_avin);
5015 /* Now look at each bb and see if it ends with a compare of a value
5017 for (bb = 0; bb < n_basic_blocks; bb++)
5019 rtx last_insn = BLOCK_END (bb);
5020 rtx condition, earliest;
5021 int compare_and_branch;
5023 /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
5024 since BLOCK_REG[BB] is zero if this block did not end with a
5025 comparison against zero, this condition works. */
5026 if (block_reg[bb] < npi->min_reg
5027 || block_reg[bb] >= npi->max_reg)
5030 /* LAST_INSN is a conditional jump. Get its condition. */
5031 condition = get_condition (last_insn, &earliest);
5033 /* If we can't determine the condition then skip. */
5037 /* Is the register known to have a nonzero value? */
5038 if (!TEST_BIT (nonnull_avout[bb], block_reg[bb] - npi->min_reg))
5041 /* Try to compute whether the compare/branch at the loop end is one or
5042 two instructions. */
5043 if (earliest == last_insn)
5044 compare_and_branch = 1;
5045 else if (earliest == prev_nonnote_insn (last_insn))
5046 compare_and_branch = 2;
5050 /* We know the register in this comparison is nonnull at exit from
5051 this block. We can optimize this comparison. */
5052 if (GET_CODE (condition) == NE)
5056 new_jump = emit_jump_insn_before (gen_jump (JUMP_LABEL (last_insn)),
5058 JUMP_LABEL (new_jump) = JUMP_LABEL (last_insn);
5059 LABEL_NUSES (JUMP_LABEL (new_jump))++;
5060 emit_barrier_after (new_jump);
5062 delete_insn (last_insn);
5063 if (compare_and_branch == 2)
5064 delete_insn (earliest);
5066 /* Don't check this block again. (Note that BLOCK_END is
5067 invalid here; we deleted the last instruction in the
5073 /* Find EQ/NE comparisons against zero which can be (indirectly) evaluated
5076 This is conceptually similar to global constant/copy propagation and
5077 classic global CSE (it even uses the same dataflow equations as cprop).
5079 If a register is used as memory address with the form (mem (reg)), then we
5080 know that REG can not be zero at that point in the program. Any instruction
5081 which sets REG "kills" this property.
5083 So, if every path leading to a conditional branch has an available memory
5084 reference of that form, then we know the register can not have the value
5085 zero at the conditional branch.
5087 So we merely need to compute the local properies and propagate that data
5088 around the cfg, then optimize where possible.
5090 We run this pass two times. Once before CSE, then again after CSE. This
5091 has proven to be the most profitable approach. It is rare for new
5092 optimization opportunities of this nature to appear after the first CSE
5095 This could probably be integrated with global cprop with a little work. */
5098 delete_null_pointer_checks (f)
5099 rtx f ATTRIBUTE_UNUSED;
5101 sbitmap *nonnull_avin, *nonnull_avout;
5102 unsigned int *block_reg;
5107 struct null_pointer_info npi;
5109 /* If we have only a single block, then there's nothing to do. */
5110 if (n_basic_blocks <= 1)
5113 /* Trying to perform global optimizations on flow graphs which have
5114 a high connectivity will take a long time and is unlikely to be
5115 particularly useful.
5117 In normal circumstances a cfg should have about twice has many edges
5118 as blocks. But we do not want to punish small functions which have
5119 a couple switch statements. So we require a relatively large number
5120 of basic blocks and the ratio of edges to blocks to be high. */
5121 if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
5124 /* We need four bitmaps, each with a bit for each register in each
5126 max_reg = max_reg_num ();
5127 regs_per_pass = get_bitmap_width (4, n_basic_blocks, max_reg);
5129 /* Allocate bitmaps to hold local and global properties. */
5130 npi.nonnull_local = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5131 npi.nonnull_killed = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5132 nonnull_avin = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5133 nonnull_avout = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5135 /* Go through the basic blocks, seeing whether or not each block
5136 ends with a conditional branch whose condition is a comparison
5137 against zero. Record the register compared in BLOCK_REG. */
5138 block_reg = (unsigned int *) xcalloc (n_basic_blocks, sizeof (int));
5139 for (bb = 0; bb < n_basic_blocks; bb++)
5141 rtx last_insn = BLOCK_END (bb);
5142 rtx condition, earliest, reg;
5144 /* We only want conditional branches. */
5145 if (GET_CODE (last_insn) != JUMP_INSN
5146 || !any_condjump_p (last_insn)
5147 || !onlyjump_p (last_insn))
5150 /* LAST_INSN is a conditional jump. Get its condition. */
5151 condition = get_condition (last_insn, &earliest);
5153 /* If we were unable to get the condition, or it is not a equality
5154 comparison against zero then there's nothing we can do. */
5156 || (GET_CODE (condition) != NE && GET_CODE (condition) != EQ)
5157 || GET_CODE (XEXP (condition, 1)) != CONST_INT
5158 || (XEXP (condition, 1)
5159 != CONST0_RTX (GET_MODE (XEXP (condition, 0)))))
5162 /* We must be checking a register against zero. */
5163 reg = XEXP (condition, 0);
5164 if (GET_CODE (reg) != REG)
5167 block_reg[bb] = REGNO (reg);
5170 /* Go through the algorithm for each block of registers. */
5171 for (reg = FIRST_PSEUDO_REGISTER; reg < max_reg; reg += regs_per_pass)
5174 npi.max_reg = MIN (reg + regs_per_pass, max_reg);
5175 delete_null_pointer_checks_1 (block_reg, nonnull_avin,
5176 nonnull_avout, &npi);
5179 /* Free the table of registers compared at the end of every block. */
5183 free (npi.nonnull_local);
5184 free (npi.nonnull_killed);
5185 free (nonnull_avin);
5186 free (nonnull_avout);
5189 /* Code Hoisting variables and subroutines. */
5191 /* Very busy expressions. */
5192 static sbitmap *hoist_vbein;
5193 static sbitmap *hoist_vbeout;
5195 /* Hoistable expressions. */
5196 static sbitmap *hoist_exprs;
5198 /* Dominator bitmaps. */
5199 static sbitmap *dominators;
5201 /* ??? We could compute post dominators and run this algorithm in
5202 reverse to to perform tail merging, doing so would probably be
5203 more effective than the tail merging code in jump.c.
5205 It's unclear if tail merging could be run in parallel with
5206 code hoisting. It would be nice. */
5208 /* Allocate vars used for code hoisting analysis. */
5211 alloc_code_hoist_mem (n_blocks, n_exprs)
5212 int n_blocks, n_exprs;
5214 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
5215 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
5216 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
5218 hoist_vbein = sbitmap_vector_alloc (n_blocks, n_exprs);
5219 hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
5220 hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
5221 transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
5223 dominators = sbitmap_vector_alloc (n_blocks, n_blocks);
5226 /* Free vars used for code hoisting analysis. */
5229 free_code_hoist_mem ()
5236 free (hoist_vbeout);
5243 /* Compute the very busy expressions at entry/exit from each block.
5245 An expression is very busy if all paths from a given point
5246 compute the expression. */
5249 compute_code_hoist_vbeinout ()
5251 int bb, changed, passes;
5253 sbitmap_vector_zero (hoist_vbeout, n_basic_blocks);
5254 sbitmap_vector_zero (hoist_vbein, n_basic_blocks);
5263 /* We scan the blocks in the reverse order to speed up
5265 for (bb = n_basic_blocks - 1; bb >= 0; bb--)
5267 changed |= sbitmap_a_or_b_and_c (hoist_vbein[bb], antloc[bb],
5268 hoist_vbeout[bb], transp[bb]);
5269 if (bb != n_basic_blocks - 1)
5270 sbitmap_intersection_of_succs (hoist_vbeout[bb], hoist_vbein, bb);
5277 fprintf (gcse_file, "hoisting vbeinout computation: %d passes\n", passes);
5280 /* Top level routine to do the dataflow analysis needed by code hoisting. */
5283 compute_code_hoist_data ()
5285 compute_local_properties (transp, comp, antloc, 0);
5286 compute_transpout ();
5287 compute_code_hoist_vbeinout ();
5288 compute_flow_dominators (dominators, NULL);
5290 fprintf (gcse_file, "\n");
5293 /* Determine if the expression identified by EXPR_INDEX would
5294 reach BB unimpared if it was placed at the end of EXPR_BB.
5296 It's unclear exactly what Muchnick meant by "unimpared". It seems
5297 to me that the expression must either be computed or transparent in
5298 *every* block in the path(s) from EXPR_BB to BB. Any other definition
5299 would allow the expression to be hoisted out of loops, even if
5300 the expression wasn't a loop invariant.
5302 Contrast this to reachability for PRE where an expression is
5303 considered reachable if *any* path reaches instead of *all*
5307 hoist_expr_reaches_here_p (expr_bb, expr_index, bb, visited)
5314 int visited_allocated_locally = 0;
5317 if (visited == NULL)
5319 visited_allocated_locally = 1;
5320 visited = xcalloc (n_basic_blocks, 1);
5323 visited[expr_bb] = 1;
5324 for (pred = BASIC_BLOCK (bb)->pred; pred != NULL; pred = pred->pred_next)
5326 int pred_bb = pred->src->index;
5328 if (pred->src == ENTRY_BLOCK_PTR)
5330 else if (visited[pred_bb])
5333 /* Does this predecessor generate this expression? */
5334 else if (TEST_BIT (comp[pred_bb], expr_index))
5336 else if (! TEST_BIT (transp[pred_bb], expr_index))
5342 visited[pred_bb] = 1;
5343 if (! hoist_expr_reaches_here_p (expr_bb, expr_index,
5348 if (visited_allocated_locally)
5351 return (pred == NULL);
5354 /* Actually perform code hoisting. */
5361 struct expr **index_map;
5364 sbitmap_vector_zero (hoist_exprs, n_basic_blocks);
5366 /* Compute a mapping from expression number (`bitmap_index') to
5367 hash table entry. */
5369 index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
5370 for (i = 0; i < expr_hash_table_size; i++)
5371 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
5372 index_map[expr->bitmap_index] = expr;
5374 /* Walk over each basic block looking for potentially hoistable
5375 expressions, nothing gets hoisted from the entry block. */
5376 for (bb = 0; bb < n_basic_blocks; bb++)
5379 int insn_inserted_p;
5381 /* Examine each expression that is very busy at the exit of this
5382 block. These are the potentially hoistable expressions. */
5383 for (i = 0; i < hoist_vbeout[bb]->n_bits; i++)
5387 if (TEST_BIT (hoist_vbeout[bb], i) && TEST_BIT (transpout[bb], i))
5389 /* We've found a potentially hoistable expression, now
5390 we look at every block BB dominates to see if it
5391 computes the expression. */
5392 for (dominated = 0; dominated < n_basic_blocks; dominated++)
5394 /* Ignore self dominance. */
5396 || ! TEST_BIT (dominators[dominated], bb))
5399 /* We've found a dominated block, now see if it computes
5400 the busy expression and whether or not moving that
5401 expression to the "beginning" of that block is safe. */
5402 if (!TEST_BIT (antloc[dominated], i))
5405 /* Note if the expression would reach the dominated block
5406 unimpared if it was placed at the end of BB.
5408 Keep track of how many times this expression is hoistable
5409 from a dominated block into BB. */
5410 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
5414 /* If we found more than one hoistable occurence of this
5415 expression, then note it in the bitmap of expressions to
5416 hoist. It makes no sense to hoist things which are computed
5417 in only one BB, and doing so tends to pessimize register
5418 allocation. One could increase this value to try harder
5419 to avoid any possible code expansion due to register
5420 allocation issues; however experiments have shown that
5421 the vast majority of hoistable expressions are only movable
5422 from two successors, so raising this threshhold is likely
5423 to nullify any benefit we get from code hoisting. */
5426 SET_BIT (hoist_exprs[bb], i);
5432 /* If we found nothing to hoist, then quit now. */
5436 /* Loop over all the hoistable expressions. */
5437 for (i = 0; i < hoist_exprs[bb]->n_bits; i++)
5439 /* We want to insert the expression into BB only once, so
5440 note when we've inserted it. */
5441 insn_inserted_p = 0;
5443 /* These tests should be the same as the tests above. */
5444 if (TEST_BIT (hoist_vbeout[bb], i))
5446 /* We've found a potentially hoistable expression, now
5447 we look at every block BB dominates to see if it
5448 computes the expression. */
5449 for (dominated = 0; dominated < n_basic_blocks; dominated++)
5451 /* Ignore self dominance. */
5453 || ! TEST_BIT (dominators[dominated], bb))
5456 /* We've found a dominated block, now see if it computes
5457 the busy expression and whether or not moving that
5458 expression to the "beginning" of that block is safe. */
5459 if (!TEST_BIT (antloc[dominated], i))
5462 /* The expression is computed in the dominated block and
5463 it would be safe to compute it at the start of the
5464 dominated block. Now we have to determine if the
5465 expresion would reach the dominated block if it was
5466 placed at the end of BB. */
5467 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
5469 struct expr *expr = index_map[i];
5470 struct occr *occr = expr->antic_occr;
5474 /* Find the right occurence of this expression. */
5475 while (BLOCK_NUM (occr->insn) != dominated && occr)
5478 /* Should never happen. */
5484 set = single_set (insn);
5488 /* Create a pseudo-reg to store the result of reaching
5489 expressions into. Get the mode for the new pseudo
5490 from the mode of the original destination pseudo. */
5491 if (expr->reaching_reg == NULL)
5493 = gen_reg_rtx (GET_MODE (SET_DEST (set)));
5495 /* In theory this should never fail since we're creating
5498 However, on the x86 some of the movXX patterns
5499 actually contain clobbers of scratch regs. This may
5500 cause the insn created by validate_change to not
5501 match any pattern and thus cause validate_change to
5503 if (validate_change (insn, &SET_SRC (set),
5504 expr->reaching_reg, 0))
5506 occr->deleted_p = 1;
5507 if (!insn_inserted_p)
5509 insert_insn_end_bb (index_map[i], bb, 0);
5510 insn_inserted_p = 1;
5522 /* Top level routine to perform one code hoisting (aka unification) pass
5524 Return non-zero if a change was made. */
5527 one_code_hoisting_pass ()
5531 alloc_expr_hash_table (max_cuid);
5532 compute_expr_hash_table ();
5534 dump_hash_table (gcse_file, "Code Hosting Expressions", expr_hash_table,
5535 expr_hash_table_size, n_exprs);
5539 alloc_code_hoist_mem (n_basic_blocks, n_exprs);
5540 compute_code_hoist_data ();
5542 free_code_hoist_mem ();
5545 free_expr_hash_table ();