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 int expr_hash_table_size;
370 This is an array of `expr_hash_table_size' elements. */
371 static struct expr **expr_hash_table;
373 /* Total size of the copy propagation hash table, in elements. */
374 static int set_hash_table_size;
377 This is an array of `set_hash_table_size' elements. */
378 static struct expr **set_hash_table;
380 /* Mapping of uids to cuids.
381 Only real insns get cuids. */
382 static int *uid_cuid;
384 /* Highest UID in UID_CUID. */
387 /* Get the cuid of an insn. */
388 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
390 /* Number of cuids. */
393 /* Mapping of cuids to insns. */
394 static rtx *cuid_insn;
396 /* Get insn from cuid. */
397 #define CUID_INSN(CUID) (cuid_insn[CUID])
399 /* Maximum register number in function prior to doing gcse + 1.
400 Registers created during this pass have regno >= max_gcse_regno.
401 This is named with "gcse" to not collide with global of same name. */
402 static int max_gcse_regno;
404 /* Maximum number of cse-able expressions found. */
407 /* Maximum number of assignments for copy propagation found. */
410 /* Table of registers that are modified.
412 For each register, each element is a list of places where the pseudo-reg
415 For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
416 requires knowledge of which blocks kill which regs [and thus could use
417 a bitmap instead of the lists `reg_set_table' uses].
419 `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
420 num-regs) [however perhaps it may be useful to keep the data as is]. One
421 advantage of recording things this way is that `reg_set_table' is fairly
422 sparse with respect to pseudo regs but for hard regs could be fairly dense
423 [relatively speaking]. And recording sets of pseudo-regs in lists speeds
424 up functions like compute_transp since in the case of pseudo-regs we only
425 need to iterate over the number of times a pseudo-reg is set, not over the
426 number of basic blocks [clearly there is a bit of a slow down in the cases
427 where a pseudo is set more than once in a block, however it is believed
428 that the net effect is to speed things up]. This isn't done for hard-regs
429 because recording call-clobbered hard-regs in `reg_set_table' at each
430 function call can consume a fair bit of memory, and iterating over
431 hard-regs stored this way in compute_transp will be more expensive. */
433 typedef struct reg_set
435 /* The next setting of this register. */
436 struct reg_set *next;
437 /* The insn where it was set. */
441 static reg_set **reg_set_table;
443 /* Size of `reg_set_table'.
444 The table starts out at max_gcse_regno + slop, and is enlarged as
446 static int reg_set_table_size;
448 /* Amount to grow `reg_set_table' by when it's full. */
449 #define REG_SET_TABLE_SLOP 100
451 /* Bitmap containing one bit for each register in the program.
452 Used when performing GCSE to track which registers have been set since
453 the start of the basic block. */
454 static sbitmap reg_set_bitmap;
456 /* For each block, a bitmap of registers set in the block.
457 This is used by expr_killed_p and compute_transp.
458 It is computed during hash table computation and not by compute_sets
459 as it includes registers added since the last pass (or between cprop and
460 gcse) and it's currently not easy to realloc sbitmap vectors. */
461 static sbitmap *reg_set_in_block;
463 /* For each block, non-zero if memory is set in that block.
464 This is computed during hash table computation and is used by
465 expr_killed_p and compute_transp.
466 ??? Handling of memory is very simple, we don't make any attempt
467 to optimize things (later).
468 ??? This can be computed by compute_sets since the information
470 static char *mem_set_in_block;
472 /* Various variables for statistics gathering. */
474 /* Memory used in a pass.
475 This isn't intended to be absolutely precise. Its intent is only
476 to keep an eye on memory usage. */
477 static int bytes_used;
479 /* GCSE substitutions made. */
480 static int gcse_subst_count;
481 /* Number of copy instructions created. */
482 static int gcse_create_count;
483 /* Number of constants propagated. */
484 static int const_prop_count;
485 /* Number of copys propagated. */
486 static int copy_prop_count;
488 /* These variables are used by classic GCSE.
489 Normally they'd be defined a bit later, but `rd_gen' needs to
490 be declared sooner. */
492 /* A bitmap of all ones for implementing the algorithm for available
493 expressions and reaching definitions. */
494 /* ??? Available expression bitmaps have a different size than reaching
495 definition bitmaps. This should be the larger of the two, however, it
496 is not currently used for reaching definitions. */
497 static sbitmap u_bitmap;
499 /* Each block has a bitmap of each type.
500 The length of each blocks bitmap is:
502 max_cuid - for reaching definitions
503 n_exprs - for available expressions
505 Thus we view the bitmaps as 2 dimensional arrays. i.e.
506 rd_kill[block_num][cuid_num]
507 ae_kill[block_num][expr_num] */
509 /* For reaching defs */
510 static sbitmap *rd_kill, *rd_gen, *reaching_defs, *rd_out;
512 /* for available exprs */
513 static sbitmap *ae_kill, *ae_gen, *ae_in, *ae_out;
515 /* Objects of this type are passed around by the null-pointer check
517 struct null_pointer_info
519 /* The basic block being processed. */
521 /* The first register to be handled in this pass. */
523 /* One greater than the last register to be handled in this pass. */
525 sbitmap *nonnull_local;
526 sbitmap *nonnull_killed;
529 static void compute_can_copy PARAMS ((void));
530 static char *gmalloc PARAMS ((unsigned int));
531 static char *grealloc PARAMS ((char *, unsigned int));
532 static char *gcse_alloc PARAMS ((unsigned long));
533 static void alloc_gcse_mem PARAMS ((rtx));
534 static void free_gcse_mem PARAMS ((void));
535 static void alloc_reg_set_mem PARAMS ((int));
536 static void free_reg_set_mem PARAMS ((void));
537 static int get_bitmap_width PARAMS ((int, int, int));
538 static void record_one_set PARAMS ((int, rtx));
539 static void record_set_info PARAMS ((rtx, rtx, void *));
540 static void compute_sets PARAMS ((rtx));
541 static void hash_scan_insn PARAMS ((rtx, int, int));
542 static void hash_scan_set PARAMS ((rtx, rtx, int));
543 static void hash_scan_clobber PARAMS ((rtx, rtx));
544 static void hash_scan_call PARAMS ((rtx, rtx));
545 static int want_to_gcse_p PARAMS ((rtx));
546 static int oprs_unchanged_p PARAMS ((rtx, rtx, int));
547 static int oprs_anticipatable_p PARAMS ((rtx, rtx));
548 static int oprs_available_p PARAMS ((rtx, rtx));
549 static void insert_expr_in_table PARAMS ((rtx, enum machine_mode, rtx,
551 static void insert_set_in_table PARAMS ((rtx, rtx));
552 static unsigned int hash_expr PARAMS ((rtx, enum machine_mode, int *, int));
553 static unsigned int hash_expr_1 PARAMS ((rtx, enum machine_mode, int *));
554 static unsigned int hash_set PARAMS ((int, int));
555 static int expr_equiv_p PARAMS ((rtx, rtx));
556 static void record_last_reg_set_info PARAMS ((rtx, int));
557 static void record_last_mem_set_info PARAMS ((rtx));
558 static void record_last_set_info PARAMS ((rtx, rtx, void *));
559 static void compute_hash_table PARAMS ((int));
560 static void alloc_set_hash_table PARAMS ((int));
561 static void free_set_hash_table PARAMS ((void));
562 static void compute_set_hash_table PARAMS ((void));
563 static void alloc_expr_hash_table PARAMS ((int));
564 static void free_expr_hash_table PARAMS ((void));
565 static void compute_expr_hash_table PARAMS ((void));
566 static void dump_hash_table PARAMS ((FILE *, const char *, struct expr **,
568 static struct expr *lookup_expr PARAMS ((rtx));
569 static struct expr *lookup_set PARAMS ((int, rtx));
570 static struct expr *next_set PARAMS ((int, struct expr *));
571 static void reset_opr_set_tables PARAMS ((void));
572 static int oprs_not_set_p PARAMS ((rtx, rtx));
573 static void mark_call PARAMS ((rtx));
574 static void mark_set PARAMS ((rtx, rtx));
575 static void mark_clobber PARAMS ((rtx, rtx));
576 static void mark_oprs_set PARAMS ((rtx));
577 static void alloc_cprop_mem PARAMS ((int, int));
578 static void free_cprop_mem PARAMS ((void));
579 static void compute_transp PARAMS ((rtx, int, sbitmap *, int));
580 static void compute_transpout PARAMS ((void));
581 static void compute_local_properties PARAMS ((sbitmap *, sbitmap *, sbitmap *,
583 static void compute_cprop_data PARAMS ((void));
584 static void find_used_regs PARAMS ((rtx));
585 static int try_replace_reg PARAMS ((rtx, rtx, rtx));
586 static struct expr *find_avail_set PARAMS ((int, rtx));
587 static int cprop_jump PARAMS ((rtx, rtx, struct reg_use *, rtx));
589 static int cprop_cc0_jump PARAMS ((rtx, struct reg_use *, rtx));
591 static int cprop_insn PARAMS ((rtx, int));
592 static int cprop PARAMS ((int));
593 static int one_cprop_pass PARAMS ((int, int));
594 static void alloc_pre_mem PARAMS ((int, int));
595 static void free_pre_mem PARAMS ((void));
596 static void compute_pre_data PARAMS ((void));
597 static int pre_expr_reaches_here_p PARAMS ((int, struct expr *, int));
598 static void insert_insn_end_bb PARAMS ((struct expr *, int, int));
599 static void pre_insert_copy_insn PARAMS ((struct expr *, rtx));
600 static void pre_insert_copies PARAMS ((void));
601 static int pre_delete PARAMS ((void));
602 static int pre_gcse PARAMS ((void));
603 static int one_pre_gcse_pass PARAMS ((int));
604 static void add_label_notes PARAMS ((rtx, rtx));
605 static void alloc_code_hoist_mem PARAMS ((int, int));
606 static void free_code_hoist_mem PARAMS ((void));
607 static void compute_code_hoist_vbeinout PARAMS ((void));
608 static void compute_code_hoist_data PARAMS ((void));
609 static int hoist_expr_reaches_here_p PARAMS ((int, int, int, char *));
610 static void hoist_code PARAMS ((void));
611 static int one_code_hoisting_pass PARAMS ((void));
612 static void alloc_rd_mem PARAMS ((int, int));
613 static void free_rd_mem PARAMS ((void));
614 static void handle_rd_kill_set PARAMS ((rtx, int, int));
615 static void compute_kill_rd PARAMS ((void));
616 static void compute_rd PARAMS ((void));
617 static void alloc_avail_expr_mem PARAMS ((int, int));
618 static void free_avail_expr_mem PARAMS ((void));
619 static void compute_ae_gen PARAMS ((void));
620 static int expr_killed_p PARAMS ((rtx, int));
621 static void compute_ae_kill PARAMS ((sbitmap *, sbitmap *));
622 static int expr_reaches_here_p PARAMS ((struct occr *, struct expr *,
624 static rtx computing_insn PARAMS ((struct expr *, rtx));
625 static int def_reaches_here_p PARAMS ((rtx, rtx));
626 static int can_disregard_other_sets PARAMS ((struct reg_set **, rtx, int));
627 static int handle_avail_expr PARAMS ((rtx, struct expr *));
628 static int classic_gcse PARAMS ((void));
629 static int one_classic_gcse_pass PARAMS ((int));
630 static void invalidate_nonnull_info PARAMS ((rtx, rtx, void *));
631 static void delete_null_pointer_checks_1 PARAMS ((int *, sbitmap *, sbitmap *,
632 struct null_pointer_info *));
633 static rtx process_insert_insn PARAMS ((struct expr *));
634 static int pre_edge_insert PARAMS ((struct edge_list *, struct expr **));
635 static int expr_reaches_here_p_work PARAMS ((struct occr *, struct expr *,
637 static int pre_expr_reaches_here_p_work PARAMS ((int, struct expr *,
640 /* Entry point for global common subexpression elimination.
641 F is the first instruction in the function. */
649 /* Bytes used at start of pass. */
650 int initial_bytes_used;
651 /* Maximum number of bytes used by a pass. */
653 /* Point to release obstack data from for each pass. */
654 char *gcse_obstack_bottom;
656 /* We do not construct an accurate cfg in functions which call
657 setjmp, so just punt to be safe. */
658 if (current_function_calls_setjmp)
661 /* Assume that we do not need to run jump optimizations after gcse. */
662 run_jump_opt_after_gcse = 0;
664 /* For calling dump_foo fns from gdb. */
665 debug_stderr = stderr;
668 /* Identify the basic block information for this function, including
669 successors and predecessors. */
670 max_gcse_regno = max_reg_num ();
671 find_basic_blocks (f, max_gcse_regno, file);
675 dump_flow_info (file);
677 /* Return if there's nothing to do. */
678 if (n_basic_blocks <= 1)
680 free_basic_block_vars (0);
684 /* Trying to perform global optimizations on flow graphs which have
685 a high connectivity will take a long time and is unlikely to be
688 In normal circumstances a cfg should have about twice has many edges
689 as blocks. But we do not want to punish small functions which have
690 a couple switch statements. So we require a relatively large number
691 of basic blocks and the ratio of edges to blocks to be high. */
692 if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
694 free_basic_block_vars (0);
698 /* See what modes support reg/reg copy operations. */
699 if (! can_copy_init_p)
705 gcc_obstack_init (&gcse_obstack);
708 /* Record where pseudo-registers are set. This data is kept accurate
709 during each pass. ??? We could also record hard-reg information here
710 [since it's unchanging], however it is currently done during hash table
713 It may be tempting to compute MEM set information here too, but MEM sets
714 will be subject to code motion one day and thus we need to compute
715 information about memory sets when we build the hash tables. */
717 alloc_reg_set_mem (max_gcse_regno);
721 initial_bytes_used = bytes_used;
723 gcse_obstack_bottom = gcse_alloc (1);
725 while (changed && pass < MAX_PASSES)
729 fprintf (file, "GCSE pass %d\n\n", pass + 1);
731 /* Initialize bytes_used to the space for the pred/succ lists,
732 and the reg_set_table data. */
733 bytes_used = initial_bytes_used;
735 /* Each pass may create new registers, so recalculate each time. */
736 max_gcse_regno = max_reg_num ();
740 /* Don't allow constant propagation to modify jumps
742 changed = one_cprop_pass (pass + 1, 0);
745 changed |= one_classic_gcse_pass (pass + 1);
748 changed |= one_pre_gcse_pass (pass + 1);
750 alloc_reg_set_mem (max_reg_num ());
752 run_jump_opt_after_gcse = 1;
755 if (max_pass_bytes < bytes_used)
756 max_pass_bytes = bytes_used;
758 /* Free up memory, then reallocate for code hoisting. We can
759 not re-use the existing allocated memory because the tables
760 will not have info for the insns or registers created by
761 partial redundancy elimination. */
764 /* It does not make sense to run code hoisting unless we optimizing
765 for code size -- it rarely makes programs faster, and can make
766 them bigger if we did partial redundancy elimination (when optimizing
767 for space, we use a classic gcse algorithm instead of partial
768 redundancy algorithms). */
771 max_gcse_regno = max_reg_num ();
773 changed |= one_code_hoisting_pass ();
776 if (max_pass_bytes < bytes_used)
777 max_pass_bytes = bytes_used;
782 fprintf (file, "\n");
786 obstack_free (&gcse_obstack, gcse_obstack_bottom);
790 /* Do one last pass of copy propagation, including cprop into
791 conditional jumps. */
793 max_gcse_regno = max_reg_num ();
795 /* This time, go ahead and allow cprop to alter jumps. */
796 one_cprop_pass (pass + 1, 1);
801 fprintf (file, "GCSE of %s: %d basic blocks, ",
802 current_function_name, n_basic_blocks);
803 fprintf (file, "%d pass%s, %d bytes\n\n",
804 pass, pass > 1 ? "es" : "", max_pass_bytes);
807 obstack_free (&gcse_obstack, NULL_PTR);
809 free_basic_block_vars (0);
810 return run_jump_opt_after_gcse;
813 /* Misc. utilities. */
815 /* Compute which modes support reg/reg copy operations. */
821 #ifndef AVOID_CCMODE_COPIES
824 char *free_point = (char *) oballoc (1);
826 bzero (can_copy_p, NUM_MACHINE_MODES);
829 for (i = 0; i < NUM_MACHINE_MODES; i++)
830 if (GET_MODE_CLASS (i) == MODE_CC)
832 #ifdef AVOID_CCMODE_COPIES
835 reg = gen_rtx_REG ((enum machine_mode) i, LAST_VIRTUAL_REGISTER + 1);
836 insn = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
837 if (recog (PATTERN (insn), insn, NULL_PTR) >= 0)
846 /* Free the objects we just allocated. */
850 /* Cover function to xmalloc to record bytes allocated. */
857 return xmalloc (size);
860 /* Cover function to xrealloc.
861 We don't record the additional size since we don't know it.
862 It won't affect memory usage stats much anyway. */
869 return xrealloc (ptr, size);
872 /* Cover function to obstack_alloc.
873 We don't need to record the bytes allocated here since
874 obstack_chunk_alloc is set to gmalloc. */
880 return (char *) obstack_alloc (&gcse_obstack, size);
883 /* Allocate memory for the cuid mapping array,
884 and reg/memory set tracking tables.
886 This is called at the start of each pass. */
895 /* Find the largest UID and create a mapping from UIDs to CUIDs.
896 CUIDs are like UIDs except they increase monotonically, have no gaps,
897 and only apply to real insns. */
899 max_uid = get_max_uid ();
900 n = (max_uid + 1) * sizeof (int);
901 uid_cuid = (int *) gmalloc (n);
902 bzero ((char *) uid_cuid, n);
903 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
905 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
906 INSN_CUID (insn) = i++;
908 INSN_CUID (insn) = i;
911 /* Create a table mapping cuids to insns. */
914 n = (max_cuid + 1) * sizeof (rtx);
915 cuid_insn = (rtx *) gmalloc (n);
916 bzero ((char *) cuid_insn, n);
917 for (insn = f, i = 0; insn; insn = NEXT_INSN (insn))
918 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
919 CUID_INSN (i++) = insn;
921 /* Allocate vars to track sets of regs. */
922 reg_set_bitmap = (sbitmap) sbitmap_alloc (max_gcse_regno);
924 /* Allocate vars to track sets of regs, memory per block. */
925 reg_set_in_block = (sbitmap *) sbitmap_vector_alloc (n_basic_blocks,
927 mem_set_in_block = (char *) gmalloc (n_basic_blocks);
930 /* Free memory allocated by alloc_gcse_mem. */
938 free (reg_set_bitmap);
940 free (reg_set_in_block);
941 free (mem_set_in_block);
944 /* Many of the global optimization algorithms work by solving dataflow
945 equations for various expressions. Initially, some local value is
946 computed for each expression in each block. Then, the values across the
947 various blocks are combined (by following flow graph edges) to arrive at
948 global values. Conceptually, each set of equations is independent. We
949 may therefore solve all the equations in parallel, solve them one at a
950 time, or pick any intermediate approach.
952 When you're going to need N two-dimensional bitmaps, each X (say, the
953 number of blocks) by Y (say, the number of expressions), call this
954 function. It's not important what X and Y represent; only that Y
955 correspond to the things that can be done in parallel. This function will
956 return an appropriate chunking factor C; you should solve C sets of
957 equations in parallel. By going through this function, we can easily
958 trade space against time; by solving fewer equations in parallel we use
962 get_bitmap_width (n, x, y)
967 /* It's not really worth figuring out *exactly* how much memory will
968 be used by a particular choice. The important thing is to get
969 something approximately right. */
970 size_t max_bitmap_memory = 10 * 1024 * 1024;
972 /* The number of bytes we'd use for a single column of minimum
974 size_t column_size = n * x * sizeof (SBITMAP_ELT_TYPE);
976 /* Often, it's reasonable just to solve all the equations in
978 if (column_size * SBITMAP_SET_SIZE (y) <= max_bitmap_memory)
981 /* Otherwise, pick the largest width we can, without going over the
983 return SBITMAP_ELT_BITS * ((max_bitmap_memory + column_size - 1)
987 /* Compute the local properties of each recorded expression.
989 Local properties are those that are defined by the block, irrespective of
992 An expression is transparent in a block if its operands are not modified
995 An expression is computed (locally available) in a block if it is computed
996 at least once and expression would contain the same value if the
997 computation was moved to the end of the block.
999 An expression is locally anticipatable in a block if it is computed at
1000 least once and expression would contain the same value if the computation
1001 was moved to the beginning of the block.
1003 We call this routine for cprop, pre and code hoisting. They all compute
1004 basically the same information and thus can easily share this code.
1006 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
1007 properties. If NULL, then it is not necessary to compute or record that
1008 particular property.
1010 SETP controls which hash table to look at. If zero, this routine looks at
1011 the expr hash table; if nonzero this routine looks at the set hash table.
1012 Additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
1016 compute_local_properties (transp, comp, antloc, setp)
1022 int i, hash_table_size;
1023 struct expr **hash_table;
1025 /* Initialize any bitmaps that were passed in. */
1029 sbitmap_vector_zero (transp, n_basic_blocks);
1031 sbitmap_vector_ones (transp, n_basic_blocks);
1035 sbitmap_vector_zero (comp, n_basic_blocks);
1037 sbitmap_vector_zero (antloc, n_basic_blocks);
1039 /* We use the same code for cprop, pre and hoisting. For cprop
1040 we care about the set hash table, for pre and hoisting we
1041 care about the expr hash table. */
1042 hash_table_size = setp ? set_hash_table_size : expr_hash_table_size;
1043 hash_table = setp ? set_hash_table : expr_hash_table;
1045 for (i = 0; i < hash_table_size; i++)
1049 for (expr = hash_table[i]; expr != NULL; expr = expr->next_same_hash)
1051 int indx = expr->bitmap_index;
1054 /* The expression is transparent in this block if it is not killed.
1055 We start by assuming all are transparent [none are killed], and
1056 then reset the bits for those that are. */
1058 compute_transp (expr->expr, indx, transp, setp);
1060 /* The occurrences recorded in antic_occr are exactly those that
1061 we want to set to non-zero in ANTLOC. */
1063 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
1065 SET_BIT (antloc[BLOCK_NUM (occr->insn)], indx);
1067 /* While we're scanning the table, this is a good place to
1069 occr->deleted_p = 0;
1072 /* The occurrences recorded in avail_occr are exactly those that
1073 we want to set to non-zero in COMP. */
1075 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
1077 SET_BIT (comp[BLOCK_NUM (occr->insn)], indx);
1079 /* While we're scanning the table, this is a good place to
1084 /* While we're scanning the table, this is a good place to
1086 expr->reaching_reg = 0;
1091 /* Register set information.
1093 `reg_set_table' records where each register is set or otherwise
1096 static struct obstack reg_set_obstack;
1099 alloc_reg_set_mem (n_regs)
1104 reg_set_table_size = n_regs + REG_SET_TABLE_SLOP;
1105 n = reg_set_table_size * sizeof (struct reg_set *);
1106 reg_set_table = (struct reg_set **) gmalloc (n);
1107 bzero ((char *) reg_set_table, n);
1109 gcc_obstack_init (®_set_obstack);
1115 free (reg_set_table);
1116 obstack_free (®_set_obstack, NULL_PTR);
1119 /* Record REGNO in the reg_set table. */
1122 record_one_set (regno, insn)
1126 /* allocate a new reg_set element and link it onto the list */
1127 struct reg_set *new_reg_info, *reg_info_ptr1, *reg_info_ptr2;
1129 /* If the table isn't big enough, enlarge it. */
1130 if (regno >= reg_set_table_size)
1132 int new_size = regno + REG_SET_TABLE_SLOP;
1135 = (struct reg_set **) grealloc ((char *) reg_set_table,
1136 new_size * sizeof (struct reg_set *));
1137 bzero ((char *) (reg_set_table + reg_set_table_size),
1138 (new_size - reg_set_table_size) * sizeof (struct reg_set *));
1139 reg_set_table_size = new_size;
1142 new_reg_info = (struct reg_set *) obstack_alloc (®_set_obstack,
1143 sizeof (struct reg_set));
1144 bytes_used += sizeof (struct reg_set);
1145 new_reg_info->insn = insn;
1146 new_reg_info->next = NULL;
1147 if (reg_set_table[regno] == NULL)
1148 reg_set_table[regno] = new_reg_info;
1151 reg_info_ptr1 = reg_info_ptr2 = reg_set_table[regno];
1152 /* ??? One could keep a "last" pointer to speed this up. */
1153 while (reg_info_ptr1 != NULL)
1155 reg_info_ptr2 = reg_info_ptr1;
1156 reg_info_ptr1 = reg_info_ptr1->next;
1159 reg_info_ptr2->next = new_reg_info;
1163 /* Called from compute_sets via note_stores to handle one SET or CLOBBER in
1164 an insn. The DATA is really the instruction in which the SET is
1168 record_set_info (dest, setter, data)
1169 rtx dest, setter ATTRIBUTE_UNUSED;
1172 rtx record_set_insn = (rtx) data;
1174 if (GET_CODE (dest) == REG && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
1175 record_one_set (REGNO (dest), record_set_insn);
1178 /* Scan the function and record each set of each pseudo-register.
1180 This is called once, at the start of the gcse pass. See the comments for
1181 `reg_set_table' for further documenation. */
1189 for (insn = f; insn != 0; insn = NEXT_INSN (insn))
1190 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
1191 note_stores (PATTERN (insn), record_set_info, insn);
1194 /* Hash table support. */
1196 /* For each register, the cuid of the first/last insn in the block to set it,
1197 or -1 if not set. */
1198 #define NEVER_SET -1
1199 static int *reg_first_set;
1200 static int *reg_last_set;
1202 /* While computing "first/last set" info, this is the CUID of first/last insn
1203 to set memory or -1 if not set. `mem_last_set' is also used when
1204 performing GCSE to record whether memory has been set since the beginning
1207 Note that handling of memory is very simple, we don't make any attempt
1208 to optimize things (later). */
1209 static int mem_first_set;
1210 static int mem_last_set;
1212 /* Perform a quick check whether X, the source of a set, is something
1213 we want to consider for GCSE. */
1219 switch (GET_CODE (x))
1235 /* Return non-zero if the operands of expression X are unchanged from the
1236 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1237 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1240 oprs_unchanged_p (x, insn, avail_p)
1251 code = GET_CODE (x);
1256 return (reg_last_set[REGNO (x)] == NEVER_SET
1257 || reg_last_set[REGNO (x)] < INSN_CUID (insn));
1259 return (reg_first_set[REGNO (x)] == NEVER_SET
1260 || reg_first_set[REGNO (x)] >= INSN_CUID (insn));
1263 if (avail_p && mem_last_set != NEVER_SET
1264 && mem_last_set >= INSN_CUID (insn))
1266 else if (! avail_p && mem_first_set != NEVER_SET
1267 && mem_first_set < INSN_CUID (insn))
1270 return oprs_unchanged_p (XEXP (x, 0), insn, avail_p);
1293 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1297 /* If we are about to do the last recursive call needed at this
1298 level, change it into iteration. This function is called enough
1301 return oprs_unchanged_p (XEXP (x, i), insn, avail_p);
1303 else if (! oprs_unchanged_p (XEXP (x, i), insn, avail_p))
1306 else if (fmt[i] == 'E')
1307 for (j = 0; j < XVECLEN (x, i); j++)
1308 if (! oprs_unchanged_p (XVECEXP (x, i, j), insn, avail_p))
1315 /* Return non-zero if the operands of expression X are unchanged from
1316 the start of INSN's basic block up to but not including INSN. */
1319 oprs_anticipatable_p (x, insn)
1322 return oprs_unchanged_p (x, insn, 0);
1325 /* Return non-zero if the operands of expression X are unchanged from
1326 INSN to the end of INSN's basic block. */
1329 oprs_available_p (x, insn)
1332 return oprs_unchanged_p (x, insn, 1);
1335 /* Hash expression X.
1337 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1338 indicating if a volatile operand is found or if the expression contains
1339 something we don't want to insert in the table.
1341 ??? One might want to merge this with canon_hash. Later. */
1344 hash_expr (x, mode, do_not_record_p, hash_table_size)
1346 enum machine_mode mode;
1347 int *do_not_record_p;
1348 int hash_table_size;
1352 *do_not_record_p = 0;
1354 hash = hash_expr_1 (x, mode, do_not_record_p);
1355 return hash % hash_table_size;
1358 /* Subroutine of hash_expr to do the actual work. */
1361 hash_expr_1 (x, mode, do_not_record_p)
1363 enum machine_mode mode;
1364 int *do_not_record_p;
1371 /* Used to turn recursion into iteration. We can't rely on GCC's
1372 tail-recursion eliminatio since we need to keep accumulating values
1379 code = GET_CODE (x);
1383 hash += ((unsigned int) REG << 7) + REGNO (x);
1387 hash += (((unsigned int) CONST_INT << 7) + (unsigned int) mode
1388 + (unsigned int) INTVAL (x));
1392 /* This is like the general case, except that it only counts
1393 the integers representing the constant. */
1394 hash += (unsigned int) code + (unsigned int) GET_MODE (x);
1395 if (GET_MODE (x) != VOIDmode)
1396 for (i = 2; i < GET_RTX_LENGTH (CONST_DOUBLE); i++)
1397 hash += (unsigned int) XWINT (x, i);
1399 hash += ((unsigned int) CONST_DOUBLE_LOW (x)
1400 + (unsigned int) CONST_DOUBLE_HIGH (x));
1403 /* Assume there is only one rtx object for any given label. */
1405 /* We don't hash on the address of the CODE_LABEL to avoid bootstrap
1406 differences and differences between each stage's debugging dumps. */
1407 hash += (((unsigned int) LABEL_REF << 7)
1408 + CODE_LABEL_NUMBER (XEXP (x, 0)));
1413 /* Don't hash on the symbol's address to avoid bootstrap differences.
1414 Different hash values may cause expressions to be recorded in
1415 different orders and thus different registers to be used in the
1416 final assembler. This also avoids differences in the dump files
1417 between various stages. */
1419 const unsigned char *p = (const unsigned char *) XSTR (x, 0);
1422 h += (h << 7) + *p++; /* ??? revisit */
1424 hash += ((unsigned int) SYMBOL_REF << 7) + h;
1429 if (MEM_VOLATILE_P (x))
1431 *do_not_record_p = 1;
1435 hash += (unsigned int) MEM;
1436 hash += MEM_ALIAS_SET (x);
1447 case UNSPEC_VOLATILE:
1448 *do_not_record_p = 1;
1452 if (MEM_VOLATILE_P (x))
1454 *do_not_record_p = 1;
1462 hash += (unsigned) code + (unsigned) GET_MODE (x);
1463 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1467 /* If we are about to do the last recursive call
1468 needed at this level, change it into iteration.
1469 This function is called enough to be worth it. */
1476 hash += hash_expr_1 (XEXP (x, i), 0, do_not_record_p);
1477 if (*do_not_record_p)
1481 else if (fmt[i] == 'E')
1482 for (j = 0; j < XVECLEN (x, i); j++)
1484 hash += hash_expr_1 (XVECEXP (x, i, j), 0, do_not_record_p);
1485 if (*do_not_record_p)
1489 else if (fmt[i] == 's')
1491 register const unsigned char *p =
1492 (const unsigned char *) XSTR (x, i);
1498 else if (fmt[i] == 'i')
1499 hash += (unsigned int) XINT (x, i);
1507 /* Hash a set of register REGNO.
1509 Sets are hashed on the register that is set. This simplifies the PRE copy
1512 ??? May need to make things more elaborate. Later, as necessary. */
1515 hash_set (regno, hash_table_size)
1517 int hash_table_size;
1522 return hash % hash_table_size;
1525 /* Return non-zero if exp1 is equivalent to exp2.
1526 ??? Borrowed from cse.c. Might want to remerge with cse.c. Later. */
1533 register enum rtx_code code;
1534 register const char *fmt;
1539 if (x == 0 || y == 0)
1542 code = GET_CODE (x);
1543 if (code != GET_CODE (y))
1546 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
1547 if (GET_MODE (x) != GET_MODE (y))
1557 return INTVAL (x) == INTVAL (y);
1560 return XEXP (x, 0) == XEXP (y, 0);
1563 return XSTR (x, 0) == XSTR (y, 0);
1566 return REGNO (x) == REGNO (y);
1569 /* Can't merge two expressions in different alias sets, since we can
1570 decide that the expression is transparent in a block when it isn't,
1571 due to it being set with the different alias set. */
1572 if (MEM_ALIAS_SET (x) != MEM_ALIAS_SET (y))
1576 /* For commutative operations, check both orders. */
1584 return ((expr_equiv_p (XEXP (x, 0), XEXP (y, 0))
1585 && expr_equiv_p (XEXP (x, 1), XEXP (y, 1)))
1586 || (expr_equiv_p (XEXP (x, 0), XEXP (y, 1))
1587 && expr_equiv_p (XEXP (x, 1), XEXP (y, 0))));
1593 /* Compare the elements. If any pair of corresponding elements
1594 fail to match, return 0 for the whole thing. */
1596 fmt = GET_RTX_FORMAT (code);
1597 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1602 if (! expr_equiv_p (XEXP (x, i), XEXP (y, i)))
1607 if (XVECLEN (x, i) != XVECLEN (y, i))
1609 for (j = 0; j < XVECLEN (x, i); j++)
1610 if (! expr_equiv_p (XVECEXP (x, i, j), XVECEXP (y, i, j)))
1615 if (strcmp (XSTR (x, i), XSTR (y, i)))
1620 if (XINT (x, i) != XINT (y, i))
1625 if (XWINT (x, i) != XWINT (y, i))
1640 /* Insert expression X in INSN in the hash table.
1641 If it is already present, record it as the last occurrence in INSN's
1644 MODE is the mode of the value X is being stored into.
1645 It is only used if X is a CONST_INT.
1647 ANTIC_P is non-zero if X is an anticipatable expression.
1648 AVAIL_P is non-zero if X is an available expression. */
1651 insert_expr_in_table (x, mode, insn, antic_p, avail_p)
1653 enum machine_mode mode;
1655 int antic_p, avail_p;
1657 int found, do_not_record_p;
1659 struct expr *cur_expr, *last_expr = NULL;
1660 struct occr *antic_occr, *avail_occr;
1661 struct occr *last_occr = NULL;
1663 hash = hash_expr (x, mode, &do_not_record_p, expr_hash_table_size);
1665 /* Do not insert expression in table if it contains volatile operands,
1666 or if hash_expr determines the expression is something we don't want
1667 to or can't handle. */
1668 if (do_not_record_p)
1671 cur_expr = expr_hash_table[hash];
1674 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1676 /* If the expression isn't found, save a pointer to the end of
1678 last_expr = cur_expr;
1679 cur_expr = cur_expr->next_same_hash;
1684 cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
1685 bytes_used += sizeof (struct expr);
1686 if (expr_hash_table[hash] == NULL)
1687 /* This is the first pattern that hashed to this index. */
1688 expr_hash_table[hash] = cur_expr;
1690 /* Add EXPR to end of this hash chain. */
1691 last_expr->next_same_hash = cur_expr;
1693 /* Set the fields of the expr element. */
1695 cur_expr->bitmap_index = n_exprs++;
1696 cur_expr->next_same_hash = NULL;
1697 cur_expr->antic_occr = NULL;
1698 cur_expr->avail_occr = NULL;
1701 /* Now record the occurrence(s). */
1704 antic_occr = cur_expr->antic_occr;
1706 /* Search for another occurrence in the same basic block. */
1707 while (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
1709 /* If an occurrence isn't found, save a pointer to the end of
1711 last_occr = antic_occr;
1712 antic_occr = antic_occr->next;
1716 /* Found another instance of the expression in the same basic block.
1717 Prefer the currently recorded one. We want the first one in the
1718 block and the block is scanned from start to end. */
1719 ; /* nothing to do */
1722 /* First occurrence of this expression in this basic block. */
1723 antic_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1724 bytes_used += sizeof (struct occr);
1725 /* First occurrence of this expression in any block? */
1726 if (cur_expr->antic_occr == NULL)
1727 cur_expr->antic_occr = antic_occr;
1729 last_occr->next = antic_occr;
1731 antic_occr->insn = insn;
1732 antic_occr->next = NULL;
1738 avail_occr = cur_expr->avail_occr;
1740 /* Search for another occurrence in the same basic block. */
1741 while (avail_occr && BLOCK_NUM (avail_occr->insn) != BLOCK_NUM (insn))
1743 /* If an occurrence isn't found, save a pointer to the end of
1745 last_occr = avail_occr;
1746 avail_occr = avail_occr->next;
1750 /* Found another instance of the expression in the same basic block.
1751 Prefer this occurrence to the currently recorded one. We want
1752 the last one in the block and the block is scanned from start
1754 avail_occr->insn = insn;
1757 /* First occurrence of this expression in this basic block. */
1758 avail_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1759 bytes_used += sizeof (struct occr);
1761 /* First occurrence of this expression in any block? */
1762 if (cur_expr->avail_occr == NULL)
1763 cur_expr->avail_occr = avail_occr;
1765 last_occr->next = avail_occr;
1767 avail_occr->insn = insn;
1768 avail_occr->next = NULL;
1773 /* Insert pattern X in INSN in the hash table.
1774 X is a SET of a reg to either another reg or a constant.
1775 If it is already present, record it as the last occurrence in INSN's
1779 insert_set_in_table (x, insn)
1785 struct expr *cur_expr, *last_expr = NULL;
1786 struct occr *cur_occr, *last_occr = NULL;
1788 if (GET_CODE (x) != SET
1789 || GET_CODE (SET_DEST (x)) != REG)
1792 hash = hash_set (REGNO (SET_DEST (x)), set_hash_table_size);
1794 cur_expr = set_hash_table[hash];
1797 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1799 /* If the expression isn't found, save a pointer to the end of
1801 last_expr = cur_expr;
1802 cur_expr = cur_expr->next_same_hash;
1807 cur_expr = (struct expr *) gcse_alloc (sizeof (struct expr));
1808 bytes_used += sizeof (struct expr);
1809 if (set_hash_table[hash] == NULL)
1810 /* This is the first pattern that hashed to this index. */
1811 set_hash_table[hash] = cur_expr;
1813 /* Add EXPR to end of this hash chain. */
1814 last_expr->next_same_hash = cur_expr;
1816 /* Set the fields of the expr element.
1817 We must copy X because it can be modified when copy propagation is
1818 performed on its operands. */
1819 /* ??? Should this go in a different obstack? */
1820 cur_expr->expr = copy_rtx (x);
1821 cur_expr->bitmap_index = n_sets++;
1822 cur_expr->next_same_hash = NULL;
1823 cur_expr->antic_occr = NULL;
1824 cur_expr->avail_occr = NULL;
1827 /* Now record the occurrence. */
1828 cur_occr = cur_expr->avail_occr;
1830 /* Search for another occurrence in the same basic block. */
1831 while (cur_occr && BLOCK_NUM (cur_occr->insn) != BLOCK_NUM (insn))
1833 /* If an occurrence isn't found, save a pointer to the end of
1835 last_occr = cur_occr;
1836 cur_occr = cur_occr->next;
1840 /* Found another instance of the expression in the same basic block.
1841 Prefer this occurrence to the currently recorded one. We want the
1842 last one in the block and the block is scanned from start to end. */
1843 cur_occr->insn = insn;
1846 /* First occurrence of this expression in this basic block. */
1847 cur_occr = (struct occr *) gcse_alloc (sizeof (struct occr));
1848 bytes_used += sizeof (struct occr);
1850 /* First occurrence of this expression in any block? */
1851 if (cur_expr->avail_occr == NULL)
1852 cur_expr->avail_occr = cur_occr;
1854 last_occr->next = cur_occr;
1856 cur_occr->insn = insn;
1857 cur_occr->next = NULL;
1861 /* Scan pattern PAT of INSN and add an entry to the hash table. If SET_P is
1862 non-zero, this is for the assignment hash table, otherwise it is for the
1863 expression hash table. */
1866 hash_scan_set (pat, insn, set_p)
1870 rtx src = SET_SRC (pat);
1871 rtx dest = SET_DEST (pat);
1873 if (GET_CODE (src) == CALL)
1874 hash_scan_call (src, insn);
1876 if (GET_CODE (dest) == REG)
1878 int regno = REGNO (dest);
1881 /* Only record sets of pseudo-regs in the hash table. */
1883 && regno >= FIRST_PSEUDO_REGISTER
1884 /* Don't GCSE something if we can't do a reg/reg copy. */
1885 && can_copy_p [GET_MODE (dest)]
1886 /* Is SET_SRC something we want to gcse? */
1887 && want_to_gcse_p (src))
1889 /* An expression is not anticipatable if its operands are
1890 modified before this insn. */
1891 int antic_p = oprs_anticipatable_p (src, insn);
1892 /* An expression is not available if its operands are
1893 subsequently modified, including this insn. */
1894 int avail_p = oprs_available_p (src, insn);
1896 insert_expr_in_table (src, GET_MODE (dest), insn, antic_p, avail_p);
1899 /* Record sets for constant/copy propagation. */
1901 && regno >= FIRST_PSEUDO_REGISTER
1902 && ((GET_CODE (src) == REG
1903 && REGNO (src) >= FIRST_PSEUDO_REGISTER
1904 && can_copy_p [GET_MODE (dest)])
1905 || GET_CODE (src) == CONST_INT
1906 || GET_CODE (src) == SYMBOL_REF
1907 || GET_CODE (src) == CONST_DOUBLE)
1908 /* A copy is not available if its src or dest is subsequently
1909 modified. Here we want to search from INSN+1 on, but
1910 oprs_available_p searches from INSN on. */
1911 && (insn == BLOCK_END (BLOCK_NUM (insn))
1912 || ((tmp = next_nonnote_insn (insn)) != NULL_RTX
1913 && oprs_available_p (pat, tmp))))
1914 insert_set_in_table (pat, insn);
1919 hash_scan_clobber (x, insn)
1920 rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
1922 /* Currently nothing to do. */
1926 hash_scan_call (x, insn)
1927 rtx x ATTRIBUTE_UNUSED, insn ATTRIBUTE_UNUSED;
1929 /* Currently nothing to do. */
1932 /* Process INSN and add hash table entries as appropriate.
1934 Only available expressions that set a single pseudo-reg are recorded.
1936 Single sets in a PARALLEL could be handled, but it's an extra complication
1937 that isn't dealt with right now. The trick is handling the CLOBBERs that
1938 are also in the PARALLEL. Later.
1940 If SET_P is non-zero, this is for the assignment hash table,
1941 otherwise it is for the expression hash table.
1942 If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
1943 not record any expressions. */
1946 hash_scan_insn (insn, set_p, in_libcall_block)
1949 int in_libcall_block;
1951 rtx pat = PATTERN (insn);
1954 /* Pick out the sets of INSN and for other forms of instructions record
1955 what's been modified. */
1957 if (GET_CODE (pat) == SET && ! in_libcall_block)
1959 /* Ignore obvious no-ops. */
1960 if (SET_SRC (pat) != SET_DEST (pat))
1961 hash_scan_set (pat, insn, set_p);
1963 else if (GET_CODE (pat) == PARALLEL)
1964 for (i = 0; i < XVECLEN (pat, 0); i++)
1966 rtx x = XVECEXP (pat, 0, i);
1968 if (GET_CODE (x) == SET)
1970 if (GET_CODE (SET_SRC (x)) == CALL)
1971 hash_scan_call (SET_SRC (x), insn);
1973 else if (GET_CODE (x) == CLOBBER)
1974 hash_scan_clobber (x, insn);
1975 else if (GET_CODE (x) == CALL)
1976 hash_scan_call (x, insn);
1979 else if (GET_CODE (pat) == CLOBBER)
1980 hash_scan_clobber (pat, insn);
1981 else if (GET_CODE (pat) == CALL)
1982 hash_scan_call (pat, insn);
1986 dump_hash_table (file, name, table, table_size, total_size)
1989 struct expr **table;
1990 int table_size, total_size;
1993 /* Flattened out table, so it's printed in proper order. */
1994 struct expr **flat_table;
1995 unsigned int *hash_val;
1999 = (struct expr **) xcalloc (total_size, sizeof (struct expr *));
2000 hash_val = (unsigned int *) xmalloc (total_size * sizeof (unsigned int));
2002 for (i = 0; i < table_size; i++)
2003 for (expr = table[i]; expr != NULL; expr = expr->next_same_hash)
2005 flat_table[expr->bitmap_index] = expr;
2006 hash_val[expr->bitmap_index] = i;
2009 fprintf (file, "%s hash table (%d buckets, %d entries)\n",
2010 name, table_size, total_size);
2012 for (i = 0; i < total_size; i++)
2013 if (flat_table[i] != 0)
2015 expr = flat_table[i];
2016 fprintf (file, "Index %d (hash value %d)\n ",
2017 expr->bitmap_index, hash_val[i]);
2018 print_rtl (file, expr->expr);
2019 fprintf (file, "\n");
2022 fprintf (file, "\n");
2028 /* Record register first/last/block set information for REGNO in INSN.
2030 reg_first_set records the first place in the block where the register
2031 is set and is used to compute "anticipatability".
2033 reg_last_set records the last place in the block where the register
2034 is set and is used to compute "availability".
2036 reg_set_in_block records whether the register is set in the block
2037 and is used to compute "transparency". */
2040 record_last_reg_set_info (insn, regno)
2044 if (reg_first_set[regno] == NEVER_SET)
2045 reg_first_set[regno] = INSN_CUID (insn);
2047 reg_last_set[regno] = INSN_CUID (insn);
2048 SET_BIT (reg_set_in_block[BLOCK_NUM (insn)], regno);
2051 /* Record memory first/last/block set information for INSN. */
2054 record_last_mem_set_info (insn)
2057 if (mem_first_set == NEVER_SET)
2058 mem_first_set = INSN_CUID (insn);
2060 mem_last_set = INSN_CUID (insn);
2061 mem_set_in_block[BLOCK_NUM (insn)] = 1;
2064 /* Called from compute_hash_table via note_stores to handle one
2065 SET or CLOBBER in an insn. DATA is really the instruction in which
2066 the SET is taking place. */
2069 record_last_set_info (dest, setter, data)
2070 rtx dest, setter ATTRIBUTE_UNUSED;
2073 rtx last_set_insn = (rtx) data;
2075 if (GET_CODE (dest) == SUBREG)
2076 dest = SUBREG_REG (dest);
2078 if (GET_CODE (dest) == REG)
2079 record_last_reg_set_info (last_set_insn, REGNO (dest));
2080 else if (GET_CODE (dest) == MEM
2081 /* Ignore pushes, they clobber nothing. */
2082 && ! push_operand (dest, GET_MODE (dest)))
2083 record_last_mem_set_info (last_set_insn);
2086 /* Top level function to create an expression or assignment hash table.
2088 Expression entries are placed in the hash table if
2089 - they are of the form (set (pseudo-reg) src),
2090 - src is something we want to perform GCSE on,
2091 - none of the operands are subsequently modified in the block
2093 Assignment entries are placed in the hash table if
2094 - they are of the form (set (pseudo-reg) src),
2095 - src is something we want to perform const/copy propagation on,
2096 - none of the operands or target are subsequently modified in the block
2098 Currently src must be a pseudo-reg or a const_int.
2100 F is the first insn.
2101 SET_P is non-zero for computing the assignment hash table. */
2104 compute_hash_table (set_p)
2109 /* While we compute the hash table we also compute a bit array of which
2110 registers are set in which blocks.
2111 We also compute which blocks set memory, in the absence of aliasing
2112 support [which is TODO].
2113 ??? This isn't needed during const/copy propagation, but it's cheap to
2115 sbitmap_vector_zero (reg_set_in_block, n_basic_blocks);
2116 bzero ((char *) mem_set_in_block, n_basic_blocks);
2118 /* Some working arrays used to track first and last set in each block. */
2119 /* ??? One could use alloca here, but at some size a threshold is crossed
2120 beyond which one should use malloc. Are we at that threshold here? */
2121 reg_first_set = (int *) gmalloc (max_gcse_regno * sizeof (int));
2122 reg_last_set = (int *) gmalloc (max_gcse_regno * sizeof (int));
2124 for (bb = 0; bb < n_basic_blocks; bb++)
2128 int in_libcall_block;
2131 /* First pass over the instructions records information used to
2132 determine when registers and memory are first and last set.
2133 ??? The mem_set_in_block and hard-reg reg_set_in_block computation
2134 could be moved to compute_sets since they currently don't change. */
2136 for (i = 0; i < max_gcse_regno; i++)
2137 reg_first_set[i] = reg_last_set[i] = NEVER_SET;
2138 mem_first_set = NEVER_SET;
2139 mem_last_set = NEVER_SET;
2141 for (insn = BLOCK_HEAD (bb);
2142 insn && insn != NEXT_INSN (BLOCK_END (bb));
2143 insn = NEXT_INSN (insn))
2145 #ifdef NON_SAVING_SETJMP
2146 if (NON_SAVING_SETJMP && GET_CODE (insn) == NOTE
2147 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_SETJMP)
2149 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2150 record_last_reg_set_info (insn, regno);
2155 if (GET_RTX_CLASS (GET_CODE (insn)) != 'i')
2158 if (GET_CODE (insn) == CALL_INSN)
2160 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2161 if ((call_used_regs[regno]
2162 && regno != STACK_POINTER_REGNUM
2163 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2164 && regno != HARD_FRAME_POINTER_REGNUM
2166 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
2167 && ! (regno == ARG_POINTER_REGNUM && fixed_regs[regno])
2169 #if defined (PIC_OFFSET_TABLE_REGNUM) && !defined (PIC_OFFSET_TABLE_REG_CALL_CLOBBERED)
2170 && ! (regno == PIC_OFFSET_TABLE_REGNUM && flag_pic)
2173 && regno != FRAME_POINTER_REGNUM)
2174 || global_regs[regno])
2175 record_last_reg_set_info (insn, regno);
2177 if (! CONST_CALL_P (insn))
2178 record_last_mem_set_info (insn);
2181 note_stores (PATTERN (insn), record_last_set_info, insn);
2184 /* The next pass builds the hash table. */
2186 for (insn = BLOCK_HEAD (bb), in_libcall_block = 0;
2187 insn && insn != NEXT_INSN (BLOCK_END (bb));
2188 insn = NEXT_INSN (insn))
2189 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
2191 if (find_reg_note (insn, REG_LIBCALL, NULL_RTX))
2192 in_libcall_block = 1;
2193 else if (find_reg_note (insn, REG_RETVAL, NULL_RTX))
2194 in_libcall_block = 0;
2195 hash_scan_insn (insn, set_p, in_libcall_block);
2199 free (reg_first_set);
2200 free (reg_last_set);
2202 /* Catch bugs early. */
2203 reg_first_set = reg_last_set = 0;
2206 /* Allocate space for the set hash table.
2207 N_INSNS is the number of instructions in the function.
2208 It is used to determine the number of buckets to use. */
2211 alloc_set_hash_table (n_insns)
2216 set_hash_table_size = n_insns / 4;
2217 if (set_hash_table_size < 11)
2218 set_hash_table_size = 11;
2220 /* Attempt to maintain efficient use of hash table.
2221 Making it an odd number is simplest for now.
2222 ??? Later take some measurements. */
2223 set_hash_table_size |= 1;
2224 n = set_hash_table_size * sizeof (struct expr *);
2225 set_hash_table = (struct expr **) gmalloc (n);
2228 /* Free things allocated by alloc_set_hash_table. */
2231 free_set_hash_table ()
2233 free (set_hash_table);
2236 /* Compute the hash table for doing copy/const propagation. */
2239 compute_set_hash_table ()
2241 /* Initialize count of number of entries in hash table. */
2243 bzero ((char *) set_hash_table,
2244 set_hash_table_size * sizeof (struct expr *));
2246 compute_hash_table (1);
2249 /* Allocate space for the expression hash table.
2250 N_INSNS is the number of instructions in the function.
2251 It is used to determine the number of buckets to use. */
2254 alloc_expr_hash_table (n_insns)
2259 expr_hash_table_size = n_insns / 2;
2260 /* Make sure the amount is usable. */
2261 if (expr_hash_table_size < 11)
2262 expr_hash_table_size = 11;
2264 /* Attempt to maintain efficient use of hash table.
2265 Making it an odd number is simplest for now.
2266 ??? Later take some measurements. */
2267 expr_hash_table_size |= 1;
2268 n = expr_hash_table_size * sizeof (struct expr *);
2269 expr_hash_table = (struct expr **) gmalloc (n);
2272 /* Free things allocated by alloc_expr_hash_table. */
2275 free_expr_hash_table ()
2277 free (expr_hash_table);
2280 /* Compute the hash table for doing GCSE. */
2283 compute_expr_hash_table ()
2285 /* Initialize count of number of entries in hash table. */
2287 bzero ((char *) expr_hash_table,
2288 expr_hash_table_size * sizeof (struct expr *));
2290 compute_hash_table (0);
2293 /* Expression tracking support. */
2295 /* Lookup pattern PAT in the expression table.
2296 The result is a pointer to the table entry, or NULL if not found. */
2298 static struct expr *
2302 int do_not_record_p;
2303 unsigned int hash = hash_expr (pat, GET_MODE (pat), &do_not_record_p,
2304 expr_hash_table_size);
2307 if (do_not_record_p)
2310 expr = expr_hash_table[hash];
2312 while (expr && ! expr_equiv_p (expr->expr, pat))
2313 expr = expr->next_same_hash;
2318 /* Lookup REGNO in the set table. If PAT is non-NULL look for the entry that
2319 matches it, otherwise return the first entry for REGNO. The result is a
2320 pointer to the table entry, or NULL if not found. */
2322 static struct expr *
2323 lookup_set (regno, pat)
2327 unsigned int hash = hash_set (regno, set_hash_table_size);
2330 expr = set_hash_table[hash];
2334 while (expr && ! expr_equiv_p (expr->expr, pat))
2335 expr = expr->next_same_hash;
2339 while (expr && REGNO (SET_DEST (expr->expr)) != regno)
2340 expr = expr->next_same_hash;
2346 /* Return the next entry for REGNO in list EXPR. */
2348 static struct expr *
2349 next_set (regno, expr)
2354 expr = expr->next_same_hash;
2355 while (expr && REGNO (SET_DEST (expr->expr)) != regno);
2360 /* Reset tables used to keep track of what's still available [since the
2361 start of the block]. */
2364 reset_opr_set_tables ()
2366 /* Maintain a bitmap of which regs have been set since beginning of
2368 sbitmap_zero (reg_set_bitmap);
2370 /* Also keep a record of the last instruction to modify memory.
2371 For now this is very trivial, we only record whether any memory
2372 location has been modified. */
2376 /* Return non-zero if the operands of X are not set before INSN in
2377 INSN's basic block. */
2380 oprs_not_set_p (x, insn)
2390 code = GET_CODE (x);
2405 if (mem_last_set != 0)
2408 return oprs_not_set_p (XEXP (x, 0), insn);
2411 return ! TEST_BIT (reg_set_bitmap, REGNO (x));
2417 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2421 /* If we are about to do the last recursive call
2422 needed at this level, change it into iteration.
2423 This function is called enough to be worth it. */
2425 return oprs_not_set_p (XEXP (x, i), insn);
2427 if (! oprs_not_set_p (XEXP (x, i), insn))
2430 else if (fmt[i] == 'E')
2431 for (j = 0; j < XVECLEN (x, i); j++)
2432 if (! oprs_not_set_p (XVECEXP (x, i, j), insn))
2439 /* Mark things set by a CALL. */
2445 mem_last_set = INSN_CUID (insn);
2448 /* Mark things set by a SET. */
2451 mark_set (pat, insn)
2454 rtx dest = SET_DEST (pat);
2456 while (GET_CODE (dest) == SUBREG
2457 || GET_CODE (dest) == ZERO_EXTRACT
2458 || GET_CODE (dest) == SIGN_EXTRACT
2459 || GET_CODE (dest) == STRICT_LOW_PART)
2460 dest = XEXP (dest, 0);
2462 if (GET_CODE (dest) == REG)
2463 SET_BIT (reg_set_bitmap, REGNO (dest));
2464 else if (GET_CODE (dest) == MEM)
2465 mem_last_set = INSN_CUID (insn);
2467 if (GET_CODE (SET_SRC (pat)) == CALL)
2471 /* Record things set by a CLOBBER. */
2474 mark_clobber (pat, insn)
2477 rtx clob = XEXP (pat, 0);
2479 while (GET_CODE (clob) == SUBREG || GET_CODE (clob) == STRICT_LOW_PART)
2480 clob = XEXP (clob, 0);
2482 if (GET_CODE (clob) == REG)
2483 SET_BIT (reg_set_bitmap, REGNO (clob));
2485 mem_last_set = INSN_CUID (insn);
2488 /* Record things set by INSN.
2489 This data is used by oprs_not_set_p. */
2492 mark_oprs_set (insn)
2495 rtx pat = PATTERN (insn);
2498 if (GET_CODE (pat) == SET)
2499 mark_set (pat, insn);
2500 else if (GET_CODE (pat) == PARALLEL)
2501 for (i = 0; i < XVECLEN (pat, 0); i++)
2503 rtx x = XVECEXP (pat, 0, i);
2505 if (GET_CODE (x) == SET)
2507 else if (GET_CODE (x) == CLOBBER)
2508 mark_clobber (x, insn);
2509 else if (GET_CODE (x) == CALL)
2513 else if (GET_CODE (pat) == CLOBBER)
2514 mark_clobber (pat, insn);
2515 else if (GET_CODE (pat) == CALL)
2520 /* Classic GCSE reaching definition support. */
2522 /* Allocate reaching def variables. */
2525 alloc_rd_mem (n_blocks, n_insns)
2526 int n_blocks, n_insns;
2528 rd_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2529 sbitmap_vector_zero (rd_kill, n_basic_blocks);
2531 rd_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2532 sbitmap_vector_zero (rd_gen, n_basic_blocks);
2534 reaching_defs = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2535 sbitmap_vector_zero (reaching_defs, n_basic_blocks);
2537 rd_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_insns);
2538 sbitmap_vector_zero (rd_out, n_basic_blocks);
2541 /* Free reaching def variables. */
2548 free (reaching_defs);
2552 /* Add INSN to the kills of BB. REGNO, set in BB, is killed by INSN. */
2555 handle_rd_kill_set (insn, regno, bb)
2559 struct reg_set *this_reg;
2561 for (this_reg = reg_set_table[regno]; this_reg; this_reg = this_reg ->next)
2562 if (BLOCK_NUM (this_reg->insn) != BLOCK_NUM (insn))
2563 SET_BIT (rd_kill[bb], INSN_CUID (this_reg->insn));
2566 /* Compute the set of kill's for reaching definitions. */
2575 For each set bit in `gen' of the block (i.e each insn which
2576 generates a definition in the block)
2577 Call the reg set by the insn corresponding to that bit regx
2578 Look at the linked list starting at reg_set_table[regx]
2579 For each setting of regx in the linked list, which is not in
2581 Set the bit in `kill' corresponding to that insn. */
2582 for (bb = 0; bb < n_basic_blocks; bb++)
2583 for (cuid = 0; cuid < max_cuid; cuid++)
2584 if (TEST_BIT (rd_gen[bb], cuid))
2586 rtx insn = CUID_INSN (cuid);
2587 rtx pat = PATTERN (insn);
2589 if (GET_CODE (insn) == CALL_INSN)
2591 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2593 if ((call_used_regs[regno]
2594 && regno != STACK_POINTER_REGNUM
2595 #if HARD_FRAME_POINTER_REGNUM != FRAME_POINTER_REGNUM
2596 && regno != HARD_FRAME_POINTER_REGNUM
2598 #if ARG_POINTER_REGNUM != FRAME_POINTER_REGNUM
2599 && ! (regno == ARG_POINTER_REGNUM
2600 && fixed_regs[regno])
2602 #if defined (PIC_OFFSET_TABLE_REGNUM) && !defined (PIC_OFFSET_TABLE_REG_CALL_CLOBBERED)
2603 && ! (regno == PIC_OFFSET_TABLE_REGNUM && flag_pic)
2605 && regno != FRAME_POINTER_REGNUM)
2606 || global_regs[regno])
2607 handle_rd_kill_set (insn, regno, bb);
2611 if (GET_CODE (pat) == PARALLEL)
2613 for (i = XVECLEN (pat, 0) - 1; i >= 0; i--)
2615 enum rtx_code code = GET_CODE (XVECEXP (pat, 0, i));
2617 if ((code == SET || code == CLOBBER)
2618 && GET_CODE (XEXP (XVECEXP (pat, 0, i), 0)) == REG)
2619 handle_rd_kill_set (insn,
2620 REGNO (XEXP (XVECEXP (pat, 0, i), 0)),
2624 else if (GET_CODE (pat) == SET && GET_CODE (SET_DEST (pat)) == REG)
2625 /* Each setting of this register outside of this block
2626 must be marked in the set of kills in this block. */
2627 handle_rd_kill_set (insn, REGNO (SET_DEST (pat)), bb);
2631 /* Compute the reaching definitions as in
2632 Compilers Principles, Techniques, and Tools. Aho, Sethi, Ullman,
2633 Chapter 10. It is the same algorithm as used for computing available
2634 expressions but applied to the gens and kills of reaching definitions. */
2639 int bb, changed, passes;
2641 for (bb = 0; bb < n_basic_blocks; bb++)
2642 sbitmap_copy (rd_out[bb] /*dst*/, rd_gen[bb] /*src*/);
2649 for (bb = 0; bb < n_basic_blocks; bb++)
2651 sbitmap_union_of_preds (reaching_defs[bb], rd_out, bb);
2652 changed |= sbitmap_union_of_diff (rd_out[bb], rd_gen[bb],
2653 reaching_defs[bb], rd_kill[bb]);
2659 fprintf (gcse_file, "reaching def computation: %d passes\n", passes);
2662 /* Classic GCSE available expression support. */
2664 /* Allocate memory for available expression computation. */
2667 alloc_avail_expr_mem (n_blocks, n_exprs)
2668 int n_blocks, n_exprs;
2670 ae_kill = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2671 sbitmap_vector_zero (ae_kill, n_basic_blocks);
2673 ae_gen = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2674 sbitmap_vector_zero (ae_gen, n_basic_blocks);
2676 ae_in = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2677 sbitmap_vector_zero (ae_in, n_basic_blocks);
2679 ae_out = (sbitmap *) sbitmap_vector_alloc (n_blocks, n_exprs);
2680 sbitmap_vector_zero (ae_out, n_basic_blocks);
2682 u_bitmap = (sbitmap) sbitmap_alloc (n_exprs);
2683 sbitmap_ones (u_bitmap);
2687 free_avail_expr_mem ()
2696 /* Compute the set of available expressions generated in each basic block. */
2705 /* For each recorded occurrence of each expression, set ae_gen[bb][expr].
2706 This is all we have to do because an expression is not recorded if it
2707 is not available, and the only expressions we want to work with are the
2708 ones that are recorded. */
2709 for (i = 0; i < expr_hash_table_size; i++)
2710 for (expr = expr_hash_table[i]; expr != 0; expr = expr->next_same_hash)
2711 for (occr = expr->avail_occr; occr != 0; occr = occr->next)
2712 SET_BIT (ae_gen[BLOCK_NUM (occr->insn)], expr->bitmap_index);
2715 /* Return non-zero if expression X is killed in BB. */
2718 expr_killed_p (x, bb)
2729 code = GET_CODE (x);
2733 return TEST_BIT (reg_set_in_block[bb], REGNO (x));
2736 if (mem_set_in_block[bb])
2739 return expr_killed_p (XEXP (x, 0), bb);
2756 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2760 /* If we are about to do the last recursive call
2761 needed at this level, change it into iteration.
2762 This function is called enough to be worth it. */
2764 return expr_killed_p (XEXP (x, i), bb);
2765 else if (expr_killed_p (XEXP (x, i), bb))
2768 else if (fmt[i] == 'E')
2769 for (j = 0; j < XVECLEN (x, i); j++)
2770 if (expr_killed_p (XVECEXP (x, i, j), bb))
2777 /* Compute the set of available expressions killed in each basic block. */
2780 compute_ae_kill (ae_gen, ae_kill)
2781 sbitmap *ae_gen, *ae_kill;
2786 for (bb = 0; bb < n_basic_blocks; bb++)
2787 for (i = 0; i < expr_hash_table_size; i++)
2788 for (expr = expr_hash_table[i]; expr; expr = expr->next_same_hash)
2790 /* Skip EXPR if generated in this block. */
2791 if (TEST_BIT (ae_gen[bb], expr->bitmap_index))
2794 if (expr_killed_p (expr->expr, bb))
2795 SET_BIT (ae_kill[bb], expr->bitmap_index);
2799 /* Actually perform the Classic GCSE optimizations. */
2801 /* Return non-zero if occurrence OCCR of expression EXPR reaches block BB.
2803 CHECK_SELF_LOOP is non-zero if we should consider a block reaching itself
2804 as a positive reach. We want to do this when there are two computations
2805 of the expression in the block.
2807 VISITED is a pointer to a working buffer for tracking which BB's have
2808 been visited. It is NULL for the top-level call.
2810 We treat reaching expressions that go through blocks containing the same
2811 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
2812 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
2813 2 as not reaching. The intent is to improve the probability of finding
2814 only one reaching expression and to reduce register lifetimes by picking
2815 the closest such expression. */
2818 expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited)
2822 int check_self_loop;
2827 for (pred = BASIC_BLOCK(bb)->pred; pred != NULL; pred = pred->pred_next)
2829 int pred_bb = pred->src->index;
2831 if (visited[pred_bb])
2832 /* This predecessor has already been visited. Nothing to do. */
2834 else if (pred_bb == bb)
2836 /* BB loops on itself. */
2838 && TEST_BIT (ae_gen[pred_bb], expr->bitmap_index)
2839 && BLOCK_NUM (occr->insn) == pred_bb)
2842 visited[pred_bb] = 1;
2845 /* Ignore this predecessor if it kills the expression. */
2846 else if (TEST_BIT (ae_kill[pred_bb], expr->bitmap_index))
2847 visited[pred_bb] = 1;
2849 /* Does this predecessor generate this expression? */
2850 else if (TEST_BIT (ae_gen[pred_bb], expr->bitmap_index))
2852 /* Is this the occurrence we're looking for?
2853 Note that there's only one generating occurrence per block
2854 so we just need to check the block number. */
2855 if (BLOCK_NUM (occr->insn) == pred_bb)
2858 visited[pred_bb] = 1;
2861 /* Neither gen nor kill. */
2864 visited[pred_bb] = 1;
2865 if (expr_reaches_here_p_work (occr, expr, pred_bb, check_self_loop,
2872 /* All paths have been checked. */
2876 /* This wrapper for expr_reaches_here_p_work() is to ensure that any
2877 memory allocated for that function is returned. */
2880 expr_reaches_here_p (occr, expr, bb, check_self_loop)
2884 int check_self_loop;
2887 char *visited = (char *) xcalloc (n_basic_blocks, 1);
2889 rval = expr_reaches_here_p_work (occr, expr, bb, check_self_loop, visited);
2895 /* Return the instruction that computes EXPR that reaches INSN's basic block.
2896 If there is more than one such instruction, return NULL.
2898 Called only by handle_avail_expr. */
2901 computing_insn (expr, insn)
2905 int bb = BLOCK_NUM (insn);
2907 if (expr->avail_occr->next == NULL)
2909 if (BLOCK_NUM (expr->avail_occr->insn) == bb)
2910 /* The available expression is actually itself
2911 (i.e. a loop in the flow graph) so do nothing. */
2914 /* (FIXME) Case that we found a pattern that was created by
2915 a substitution that took place. */
2916 return expr->avail_occr->insn;
2920 /* Pattern is computed more than once.
2921 Search backwards from this insn to see how many of these
2922 computations actually reach this insn. */
2924 rtx insn_computes_expr = NULL;
2927 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
2929 if (BLOCK_NUM (occr->insn) == bb)
2931 /* The expression is generated in this block.
2932 The only time we care about this is when the expression
2933 is generated later in the block [and thus there's a loop].
2934 We let the normal cse pass handle the other cases. */
2935 if (INSN_CUID (insn) < INSN_CUID (occr->insn)
2936 && expr_reaches_here_p (occr, expr, bb, 1))
2942 insn_computes_expr = occr->insn;
2945 else if (expr_reaches_here_p (occr, expr, bb, 0))
2951 insn_computes_expr = occr->insn;
2955 if (insn_computes_expr == NULL)
2958 return insn_computes_expr;
2962 /* Return non-zero if the definition in DEF_INSN can reach INSN.
2963 Only called by can_disregard_other_sets. */
2966 def_reaches_here_p (insn, def_insn)
2971 if (TEST_BIT (reaching_defs[BLOCK_NUM (insn)], INSN_CUID (def_insn)))
2974 if (BLOCK_NUM (insn) == BLOCK_NUM (def_insn))
2976 if (INSN_CUID (def_insn) < INSN_CUID (insn))
2978 if (GET_CODE (PATTERN (def_insn)) == PARALLEL)
2980 else if (GET_CODE (PATTERN (def_insn)) == CLOBBER)
2981 reg = XEXP (PATTERN (def_insn), 0);
2982 else if (GET_CODE (PATTERN (def_insn)) == SET)
2983 reg = SET_DEST (PATTERN (def_insn));
2987 return ! reg_set_between_p (reg, NEXT_INSN (def_insn), insn);
2996 /* Return non-zero if *ADDR_THIS_REG can only have one value at INSN. The
2997 value returned is the number of definitions that reach INSN. Returning a
2998 value of zero means that [maybe] more than one definition reaches INSN and
2999 the caller can't perform whatever optimization it is trying. i.e. it is
3000 always safe to return zero. */
3003 can_disregard_other_sets (addr_this_reg, insn, for_combine)
3004 struct reg_set **addr_this_reg;
3008 int number_of_reaching_defs = 0;
3009 struct reg_set *this_reg;
3011 for (this_reg = *addr_this_reg; this_reg != 0; this_reg = this_reg->next)
3012 if (def_reaches_here_p (insn, this_reg->insn))
3014 number_of_reaching_defs++;
3015 /* Ignore parallels for now. */
3016 if (GET_CODE (PATTERN (this_reg->insn)) == PARALLEL)
3020 && (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER
3021 || ! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
3022 SET_SRC (PATTERN (insn)))))
3023 /* A setting of the reg to a different value reaches INSN. */
3026 if (number_of_reaching_defs > 1)
3028 /* If in this setting the value the register is being set to is
3029 equal to the previous value the register was set to and this
3030 setting reaches the insn we are trying to do the substitution
3031 on then we are ok. */
3032 if (GET_CODE (PATTERN (this_reg->insn)) == CLOBBER)
3034 else if (! rtx_equal_p (SET_SRC (PATTERN (this_reg->insn)),
3035 SET_SRC (PATTERN (insn))))
3039 *addr_this_reg = this_reg;
3042 return number_of_reaching_defs;
3045 /* Expression computed by insn is available and the substitution is legal,
3046 so try to perform the substitution.
3048 The result is non-zero if any changes were made. */
3051 handle_avail_expr (insn, expr)
3055 rtx pat, insn_computes_expr;
3057 struct reg_set *this_reg;
3058 int found_setting, use_src;
3061 /* We only handle the case where one computation of the expression
3062 reaches this instruction. */
3063 insn_computes_expr = computing_insn (expr, insn);
3064 if (insn_computes_expr == NULL)
3070 /* At this point we know only one computation of EXPR outside of this
3071 block reaches this insn. Now try to find a register that the
3072 expression is computed into. */
3073 if (GET_CODE (SET_SRC (PATTERN (insn_computes_expr))) == REG)
3075 /* This is the case when the available expression that reaches
3076 here has already been handled as an available expression. */
3077 int regnum_for_replacing
3078 = REGNO (SET_SRC (PATTERN (insn_computes_expr)));
3080 /* If the register was created by GCSE we can't use `reg_set_table',
3081 however we know it's set only once. */
3082 if (regnum_for_replacing >= max_gcse_regno
3083 /* If the register the expression is computed into is set only once,
3084 or only one set reaches this insn, we can use it. */
3085 || (((this_reg = reg_set_table[regnum_for_replacing]),
3086 this_reg->next == NULL)
3087 || can_disregard_other_sets (&this_reg, insn, 0)))
3096 int regnum_for_replacing
3097 = REGNO (SET_DEST (PATTERN (insn_computes_expr)));
3099 /* This shouldn't happen. */
3100 if (regnum_for_replacing >= max_gcse_regno)
3103 this_reg = reg_set_table[regnum_for_replacing];
3105 /* If the register the expression is computed into is set only once,
3106 or only one set reaches this insn, use it. */
3107 if (this_reg->next == NULL
3108 || can_disregard_other_sets (&this_reg, insn, 0))
3114 pat = PATTERN (insn);
3116 to = SET_SRC (PATTERN (insn_computes_expr));
3118 to = SET_DEST (PATTERN (insn_computes_expr));
3119 changed = validate_change (insn, &SET_SRC (pat), to, 0);
3121 /* We should be able to ignore the return code from validate_change but
3122 to play it safe we check. */
3126 if (gcse_file != NULL)
3128 fprintf (gcse_file, "GCSE: Replacing the source in insn %d with",
3130 fprintf (gcse_file, " reg %d %s insn %d\n",
3131 REGNO (to), use_src ? "from" : "set in",
3132 INSN_UID (insn_computes_expr));
3137 /* The register that the expr is computed into is set more than once. */
3138 else if (1 /*expensive_op(this_pattrn->op) && do_expensive_gcse)*/)
3140 /* Insert an insn after insnx that copies the reg set in insnx
3141 into a new pseudo register call this new register REGN.
3142 From insnb until end of basic block or until REGB is set
3143 replace all uses of REGB with REGN. */
3146 to = gen_reg_rtx (GET_MODE (SET_DEST (PATTERN (insn_computes_expr))));
3148 /* Generate the new insn. */
3149 /* ??? If the change fails, we return 0, even though we created
3150 an insn. I think this is ok. */
3152 = emit_insn_after (gen_rtx_SET (VOIDmode, to,
3154 (insn_computes_expr))),
3155 insn_computes_expr);
3157 /* Keep block number table up to date. */
3158 set_block_num (new_insn, BLOCK_NUM (insn_computes_expr));
3160 /* Keep register set table up to date. */
3161 record_one_set (REGNO (to), new_insn);
3163 gcse_create_count++;
3164 if (gcse_file != NULL)
3166 fprintf (gcse_file, "GCSE: Creating insn %d to copy value of reg %d",
3167 INSN_UID (NEXT_INSN (insn_computes_expr)),
3168 REGNO (SET_SRC (PATTERN (NEXT_INSN (insn_computes_expr)))));
3169 fprintf (gcse_file, ", computed in insn %d,\n",
3170 INSN_UID (insn_computes_expr));
3171 fprintf (gcse_file, " into newly allocated reg %d\n",
3175 pat = PATTERN (insn);
3177 /* Do register replacement for INSN. */
3178 changed = validate_change (insn, &SET_SRC (pat),
3180 (NEXT_INSN (insn_computes_expr))),
3183 /* We should be able to ignore the return code from validate_change but
3184 to play it safe we check. */
3188 if (gcse_file != NULL)
3191 "GCSE: Replacing the source in insn %d with reg %d ",
3193 REGNO (SET_DEST (PATTERN (NEXT_INSN
3194 (insn_computes_expr)))));
3195 fprintf (gcse_file, "set in insn %d\n",
3196 INSN_UID (insn_computes_expr));
3204 /* Perform classic GCSE. This is called by one_classic_gcse_pass after all
3205 the dataflow analysis has been done.
3207 The result is non-zero if a change was made. */
3215 /* Note we start at block 1. */
3218 for (bb = 1; bb < n_basic_blocks; bb++)
3220 /* Reset tables used to keep track of what's still valid [since the
3221 start of the block]. */
3222 reset_opr_set_tables ();
3224 for (insn = BLOCK_HEAD (bb);
3225 insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
3226 insn = NEXT_INSN (insn))
3228 /* Is insn of form (set (pseudo-reg) ...)? */
3229 if (GET_CODE (insn) == INSN
3230 && GET_CODE (PATTERN (insn)) == SET
3231 && GET_CODE (SET_DEST (PATTERN (insn))) == REG
3232 && REGNO (SET_DEST (PATTERN (insn))) >= FIRST_PSEUDO_REGISTER)
3234 rtx pat = PATTERN (insn);
3235 rtx src = SET_SRC (pat);
3238 if (want_to_gcse_p (src)
3239 /* Is the expression recorded? */
3240 && ((expr = lookup_expr (src)) != NULL)
3241 /* Is the expression available [at the start of the
3243 && TEST_BIT (ae_in[bb], expr->bitmap_index)
3244 /* Are the operands unchanged since the start of the
3246 && oprs_not_set_p (src, insn))
3247 changed |= handle_avail_expr (insn, expr);
3250 /* Keep track of everything modified by this insn. */
3251 /* ??? Need to be careful w.r.t. mods done to INSN. */
3252 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3253 mark_oprs_set (insn);
3260 /* Top level routine to perform one classic GCSE pass.
3262 Return non-zero if a change was made. */
3265 one_classic_gcse_pass (pass)
3270 gcse_subst_count = 0;
3271 gcse_create_count = 0;
3273 alloc_expr_hash_table (max_cuid);
3274 alloc_rd_mem (n_basic_blocks, max_cuid);
3275 compute_expr_hash_table ();
3277 dump_hash_table (gcse_file, "Expression", expr_hash_table,
3278 expr_hash_table_size, n_exprs);
3284 alloc_avail_expr_mem (n_basic_blocks, n_exprs);
3286 compute_ae_kill (ae_gen, ae_kill);
3287 compute_available (ae_gen, ae_kill, ae_out, ae_in);
3288 changed = classic_gcse ();
3289 free_avail_expr_mem ();
3293 free_expr_hash_table ();
3297 fprintf (gcse_file, "\n");
3298 fprintf (gcse_file, "GCSE of %s, pass %d: %d bytes needed, %d substs,",
3299 current_function_name, pass, bytes_used, gcse_subst_count);
3300 fprintf (gcse_file, "%d insns created\n", gcse_create_count);
3306 /* Compute copy/constant propagation working variables. */
3308 /* Local properties of assignments. */
3309 static sbitmap *cprop_pavloc;
3310 static sbitmap *cprop_absaltered;
3312 /* Global properties of assignments (computed from the local properties). */
3313 static sbitmap *cprop_avin;
3314 static sbitmap *cprop_avout;
3316 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
3317 basic blocks. N_SETS is the number of sets. */
3320 alloc_cprop_mem (n_blocks, n_sets)
3321 int n_blocks, n_sets;
3323 cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
3324 cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
3326 cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
3327 cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
3330 /* Free vars used by copy/const propagation. */
3335 free (cprop_pavloc);
3336 free (cprop_absaltered);
3341 /* For each block, compute whether X is transparent. X is either an
3342 expression or an assignment [though we don't care which, for this context
3343 an assignment is treated as an expression]. For each block where an
3344 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
3348 compute_transp (x, indx, bmap, set_p)
3359 /* repeat is used to turn tail-recursion into iteration since GCC
3360 can't do it when there's no return value. */
3366 code = GET_CODE (x);
3372 if (REGNO (x) < FIRST_PSEUDO_REGISTER)
3374 for (bb = 0; bb < n_basic_blocks; bb++)
3375 if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
3376 SET_BIT (bmap[bb], indx);
3380 for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
3381 SET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
3386 if (REGNO (x) < FIRST_PSEUDO_REGISTER)
3388 for (bb = 0; bb < n_basic_blocks; bb++)
3389 if (TEST_BIT (reg_set_in_block[bb], REGNO (x)))
3390 RESET_BIT (bmap[bb], indx);
3394 for (r = reg_set_table[REGNO (x)]; r != NULL; r = r->next)
3395 RESET_BIT (bmap[BLOCK_NUM (r->insn)], indx);
3404 for (bb = 0; bb < n_basic_blocks; bb++)
3405 if (mem_set_in_block[bb])
3406 SET_BIT (bmap[bb], indx);
3410 for (bb = 0; bb < n_basic_blocks; bb++)
3411 if (mem_set_in_block[bb])
3412 RESET_BIT (bmap[bb], indx);
3433 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3437 /* If we are about to do the last recursive call
3438 needed at this level, change it into iteration.
3439 This function is called enough to be worth it. */
3446 compute_transp (XEXP (x, i), indx, bmap, set_p);
3448 else if (fmt[i] == 'E')
3449 for (j = 0; j < XVECLEN (x, i); j++)
3450 compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
3454 /* Top level routine to do the dataflow analysis needed by copy/const
3458 compute_cprop_data ()
3460 compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, 1);
3461 compute_available (cprop_pavloc, cprop_absaltered,
3462 cprop_avout, cprop_avin);
3465 /* Copy/constant propagation. */
3467 /* Maximum number of register uses in an insn that we handle. */
3470 /* Table of uses found in an insn.
3471 Allocated statically to avoid alloc/free complexity and overhead. */
3472 static struct reg_use reg_use_table[MAX_USES];
3474 /* Index into `reg_use_table' while building it. */
3475 static int reg_use_count;
3477 /* Set up a list of register numbers used in INSN. The found uses are stored
3478 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
3479 and contains the number of uses in the table upon exit.
3481 ??? If a register appears multiple times we will record it multiple times.
3482 This doesn't hurt anything but it will slow things down. */
3492 /* repeat is used to turn tail-recursion into iteration since GCC
3493 can't do it when there's no return value. */
3499 code = GET_CODE (x);
3503 if (reg_use_count == MAX_USES)
3506 reg_use_table[reg_use_count].reg_rtx = x;
3524 case ASM_INPUT: /*FIXME*/
3528 if (GET_CODE (SET_DEST (x)) == MEM)
3529 find_used_regs (SET_DEST (x));
3537 /* Recursively scan the operands of this expression. */
3539 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
3543 /* If we are about to do the last recursive call
3544 needed at this level, change it into iteration.
3545 This function is called enough to be worth it. */
3552 find_used_regs (XEXP (x, i));
3554 else if (fmt[i] == 'E')
3555 for (j = 0; j < XVECLEN (x, i); j++)
3556 find_used_regs (XVECEXP (x, i, j));
3560 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
3561 Returns non-zero is successful. */
3564 try_replace_reg (from, to, insn)
3572 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3575 note = find_reg_note (insn, REG_EQUIV, NULL_RTX);
3577 /* If this fails we could try to simplify the result of the
3578 replacement and attempt to recognize the simplified insn.
3580 But we need a general simplify_rtx that doesn't have pass
3581 specific state variables. I'm not aware of one at the moment. */
3583 success = validate_replace_src (from, to, insn);
3584 set = single_set (insn);
3586 /* We've failed to do replacement. Try to add REG_EQUAL note to not loose
3588 if (!success && !note)
3593 note = REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_EQUAL,
3594 copy_rtx (SET_SRC (set)),
3598 /* Always do the replacement in REQ_EQUAL and REG_EQUIV notes. Also
3599 try to simplify them. */
3604 src = XEXP (note, 0);
3605 replace_rtx (src, from, to);
3607 /* Try to simplify resulting note. */
3608 simplified = simplify_rtx (src);
3612 XEXP (note, 0) = src;
3615 /* REG_EQUAL may get simplified into register.
3616 We don't allow that. Remove that note. This code ought
3617 not to hapen, because previous code ought to syntetize
3618 reg-reg move, but be on the safe side. */
3619 else if (REG_P (src))
3620 remove_note (insn, note);
3625 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
3626 NULL no such set is found. */
3628 static struct expr *
3629 find_avail_set (regno, insn)
3633 /* SET1 contains the last set found that can be returned to the caller for
3634 use in a substitution. */
3635 struct expr *set1 = 0;
3637 /* Loops are not possible here. To get a loop we would need two sets
3638 available at the start of the block containing INSN. ie we would
3639 need two sets like this available at the start of the block:
3641 (set (reg X) (reg Y))
3642 (set (reg Y) (reg X))
3644 This can not happen since the set of (reg Y) would have killed the
3645 set of (reg X) making it unavailable at the start of this block. */
3649 struct expr *set = lookup_set (regno, NULL_RTX);
3651 /* Find a set that is available at the start of the block
3652 which contains INSN. */
3655 if (TEST_BIT (cprop_avin[BLOCK_NUM (insn)], set->bitmap_index))
3657 set = next_set (regno, set);
3660 /* If no available set was found we've reached the end of the
3661 (possibly empty) copy chain. */
3665 if (GET_CODE (set->expr) != SET)
3668 src = SET_SRC (set->expr);
3670 /* We know the set is available.
3671 Now check that SRC is ANTLOC (i.e. none of the source operands
3672 have changed since the start of the block).
3674 If the source operand changed, we may still use it for the next
3675 iteration of this loop, but we may not use it for substitutions. */
3677 if (CONSTANT_P (src) || oprs_not_set_p (src, insn))
3680 /* If the source of the set is anything except a register, then
3681 we have reached the end of the copy chain. */
3682 if (GET_CODE (src) != REG)
3685 /* Follow the copy chain, ie start another iteration of the loop
3686 and see if we have an available copy into SRC. */
3687 regno = REGNO (src);
3690 /* SET1 holds the last set that was available and anticipatable at
3695 /* Subroutine of cprop_insn that tries to propagate constants into
3696 JUMP_INSNS. INSN must be a conditional jump; COPY is a copy of it
3697 that we can use for substitutions.
3698 REG_USED is the use we will try to replace, SRC is the constant we
3699 will try to substitute for it.
3700 Returns nonzero if a change was made. */
3703 cprop_jump (insn, copy, reg_used, src)
3705 struct reg_use *reg_used;
3708 rtx set = PATTERN (copy);
3711 /* Replace the register with the appropriate constant. */
3712 replace_rtx (SET_SRC (set), reg_used->reg_rtx, src);
3714 temp = simplify_ternary_operation (GET_CODE (SET_SRC (set)),
3715 GET_MODE (SET_SRC (set)),
3716 GET_MODE (XEXP (SET_SRC (set), 0)),
3717 XEXP (SET_SRC (set), 0),
3718 XEXP (SET_SRC (set), 1),
3719 XEXP (SET_SRC (set), 2));
3721 /* If no simplification can be made, then try the next
3726 SET_SRC (set) = temp;
3728 /* That may have changed the structure of TEMP, so
3729 force it to be rerecognized if it has not turned
3730 into a nop or unconditional jump. */
3732 INSN_CODE (copy) = -1;
3733 if ((SET_DEST (set) == pc_rtx
3734 && (SET_SRC (set) == pc_rtx
3735 || GET_CODE (SET_SRC (set)) == LABEL_REF))
3736 || recog (PATTERN (copy), copy, NULL) >= 0)
3738 /* This has either become an unconditional jump
3739 or a nop-jump. We'd like to delete nop jumps
3740 here, but doing so confuses gcse. So we just
3741 make the replacement and let later passes
3743 PATTERN (insn) = set;
3744 INSN_CODE (insn) = -1;
3746 /* One less use of the label this insn used to jump to
3747 if we turned this into a NOP jump. */
3748 if (SET_SRC (set) == pc_rtx && JUMP_LABEL (insn) != 0)
3749 --LABEL_NUSES (JUMP_LABEL (insn));
3751 /* If this has turned into an unconditional jump,
3752 then put a barrier after it so that the unreachable
3753 code will be deleted. */
3754 if (GET_CODE (SET_SRC (set)) == LABEL_REF)
3755 emit_barrier_after (insn);
3757 run_jump_opt_after_gcse = 1;
3760 if (gcse_file != NULL)
3763 "CONST-PROP: Replacing reg %d in insn %d with constant ",
3764 REGNO (reg_used->reg_rtx), INSN_UID (insn));
3765 print_rtl (gcse_file, src);
3766 fprintf (gcse_file, "\n");
3776 /* Subroutine of cprop_insn that tries to propagate constants into JUMP_INSNS
3777 for machines that have CC0. INSN is a single set that stores into CC0;
3778 the insn following it is a conditional jump. REG_USED is the use we will
3779 try to replace, SRC is the constant we will try to substitute for it.
3780 Returns nonzero if a change was made. */
3783 cprop_cc0_jump (insn, reg_used, src)
3785 struct reg_use *reg_used;
3788 rtx jump = NEXT_INSN (insn);
3789 rtx copy = copy_rtx (jump);
3790 rtx set = PATTERN (copy);
3792 /* We need to copy the source of the cc0 setter, as cprop_jump is going to
3793 substitute into it. */
3794 replace_rtx (SET_SRC (set), cc0_rtx, copy_rtx (SET_SRC (PATTERN (insn))));
3795 if (! cprop_jump (jump, copy, reg_used, src))
3798 /* If we succeeded, delete the cc0 setter. */
3799 PUT_CODE (insn, NOTE);
3800 NOTE_LINE_NUMBER (insn) = NOTE_INSN_DELETED;
3801 NOTE_SOURCE_FILE (insn) = 0;
3806 /* Perform constant and copy propagation on INSN.
3807 The result is non-zero if a change was made. */
3810 cprop_insn (insn, alter_jumps)
3814 struct reg_use *reg_used;
3818 /* Only propagate into SETs. Note that a conditional jump is a
3819 SET with pc_rtx as the destination. */
3820 if ((GET_CODE (insn) != INSN
3821 && GET_CODE (insn) != JUMP_INSN)
3822 || GET_CODE (PATTERN (insn)) != SET)
3826 find_used_regs (PATTERN (insn));
3828 note = find_reg_note (insn, REG_EQUIV, NULL_RTX);
3830 note = find_reg_note (insn, REG_EQUAL, NULL_RTX);
3832 /* We may win even when propagating constants into notes. */
3834 find_used_regs (XEXP (note, 0));
3836 for (reg_used = ®_use_table[0]; reg_use_count > 0;
3837 reg_used++, reg_use_count--)
3839 int regno = REGNO (reg_used->reg_rtx);
3843 /* Ignore registers created by GCSE.
3844 We do this because ... */
3845 if (regno >= max_gcse_regno)
3848 /* If the register has already been set in this block, there's
3849 nothing we can do. */
3850 if (! oprs_not_set_p (reg_used->reg_rtx, insn))
3853 /* Find an assignment that sets reg_used and is available
3854 at the start of the block. */
3855 set = find_avail_set (regno, insn);
3860 /* ??? We might be able to handle PARALLELs. Later. */
3861 if (GET_CODE (pat) != SET)
3864 src = SET_SRC (pat);
3866 /* Constant propagation. */
3867 if (GET_CODE (src) == CONST_INT || GET_CODE (src) == CONST_DOUBLE
3868 || GET_CODE (src) == SYMBOL_REF)
3870 /* Handle normal insns first. */
3871 if (GET_CODE (insn) == INSN
3872 && try_replace_reg (reg_used->reg_rtx, src, insn))
3876 if (gcse_file != NULL)
3878 fprintf (gcse_file, "CONST-PROP: Replacing reg %d in ",
3880 fprintf (gcse_file, "insn %d with constant ",
3882 print_rtl (gcse_file, src);
3883 fprintf (gcse_file, "\n");
3886 /* The original insn setting reg_used may or may not now be
3887 deletable. We leave the deletion to flow. */
3890 /* Try to propagate a CONST_INT into a conditional jump.
3891 We're pretty specific about what we will handle in this
3892 code, we can extend this as necessary over time.
3894 Right now the insn in question must look like
3895 (set (pc) (if_then_else ...)) */
3896 else if (alter_jumps
3897 && GET_CODE (insn) == JUMP_INSN
3898 && condjump_p (insn)
3899 && ! simplejump_p (insn))
3900 changed |= cprop_jump (insn, copy_rtx (insn), reg_used, src);
3902 /* Similar code for machines that use a pair of CC0 setter and
3903 conditional jump insn. */
3904 else if (alter_jumps
3905 && GET_CODE (PATTERN (insn)) == SET
3906 && SET_DEST (PATTERN (insn)) == cc0_rtx
3907 && GET_CODE (NEXT_INSN (insn)) == JUMP_INSN
3908 && condjump_p (NEXT_INSN (insn))
3909 && ! simplejump_p (NEXT_INSN (insn)))
3910 changed |= cprop_cc0_jump (insn, reg_used, src);
3913 else if (GET_CODE (src) == REG
3914 && REGNO (src) >= FIRST_PSEUDO_REGISTER
3915 && REGNO (src) != regno)
3917 if (try_replace_reg (reg_used->reg_rtx, src, insn))
3921 if (gcse_file != NULL)
3923 fprintf (gcse_file, "COPY-PROP: Replacing reg %d in insn %d",
3924 regno, INSN_UID (insn));
3925 fprintf (gcse_file, " with reg %d\n", REGNO (src));
3928 /* The original insn setting reg_used may or may not now be
3929 deletable. We leave the deletion to flow. */
3930 /* FIXME: If it turns out that the insn isn't deletable,
3931 then we may have unnecessarily extended register lifetimes
3932 and made things worse. */
3940 /* Forward propagate copies. This includes copies and constants. Return
3941 non-zero if a change was made. */
3950 /* Note we start at block 1. */
3953 for (bb = 1; bb < n_basic_blocks; bb++)
3955 /* Reset tables used to keep track of what's still valid [since the
3956 start of the block]. */
3957 reset_opr_set_tables ();
3959 for (insn = BLOCK_HEAD (bb);
3960 insn != NULL && insn != NEXT_INSN (BLOCK_END (bb));
3961 insn = NEXT_INSN (insn))
3963 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
3965 changed |= cprop_insn (insn, alter_jumps);
3967 /* Keep track of everything modified by this insn. */
3968 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
3969 call mark_oprs_set if we turned the insn into a NOTE. */
3970 if (GET_CODE (insn) != NOTE)
3971 mark_oprs_set (insn);
3976 if (gcse_file != NULL)
3977 fprintf (gcse_file, "\n");
3982 /* Perform one copy/constant propagation pass.
3983 F is the first insn in the function.
3984 PASS is the pass count. */
3987 one_cprop_pass (pass, alter_jumps)
3993 const_prop_count = 0;
3994 copy_prop_count = 0;
3996 alloc_set_hash_table (max_cuid);
3997 compute_set_hash_table ();
3999 dump_hash_table (gcse_file, "SET", set_hash_table, set_hash_table_size,
4003 alloc_cprop_mem (n_basic_blocks, n_sets);
4004 compute_cprop_data ();
4005 changed = cprop (alter_jumps);
4009 free_set_hash_table ();
4013 fprintf (gcse_file, "CPROP of %s, pass %d: %d bytes needed, ",
4014 current_function_name, pass, bytes_used);
4015 fprintf (gcse_file, "%d const props, %d copy props\n\n",
4016 const_prop_count, copy_prop_count);
4022 /* Compute PRE+LCM working variables. */
4024 /* Local properties of expressions. */
4025 /* Nonzero for expressions that are transparent in the block. */
4026 static sbitmap *transp;
4028 /* Nonzero for expressions that are transparent at the end of the block.
4029 This is only zero for expressions killed by abnormal critical edge
4030 created by a calls. */
4031 static sbitmap *transpout;
4033 /* Nonzero for expressions that are computed (available) in the block. */
4034 static sbitmap *comp;
4036 /* Nonzero for expressions that are locally anticipatable in the block. */
4037 static sbitmap *antloc;
4039 /* Nonzero for expressions where this block is an optimal computation
4041 static sbitmap *pre_optimal;
4043 /* Nonzero for expressions which are redundant in a particular block. */
4044 static sbitmap *pre_redundant;
4046 /* Nonzero for expressions which should be inserted on a specific edge. */
4047 static sbitmap *pre_insert_map;
4049 /* Nonzero for expressions which should be deleted in a specific block. */
4050 static sbitmap *pre_delete_map;
4052 /* Contains the edge_list returned by pre_edge_lcm. */
4053 static struct edge_list *edge_list;
4055 static sbitmap *temp_bitmap;
4057 /* Redundant insns. */
4058 static sbitmap pre_redundant_insns;
4060 /* Allocate vars used for PRE analysis. */
4063 alloc_pre_mem (n_blocks, n_exprs)
4064 int n_blocks, n_exprs;
4066 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
4067 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
4068 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
4069 temp_bitmap = sbitmap_vector_alloc (n_blocks, n_exprs);
4072 pre_redundant = NULL;
4073 pre_insert_map = NULL;
4074 pre_delete_map = NULL;
4078 transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
4079 ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
4081 /* pre_insert and pre_delete are allocated later. */
4084 /* Free vars used for PRE analysis. */
4097 free (pre_redundant);
4099 free (pre_insert_map);
4101 free (pre_delete_map);
4114 transp = comp = antloc = NULL;
4115 pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
4116 transpout = ae_in = ae_out = ae_kill = NULL;
4121 /* Top level routine to do the dataflow analysis needed by PRE. */
4126 compute_local_properties (transp, comp, antloc, 0);
4127 compute_transpout ();
4128 sbitmap_vector_zero (ae_kill, n_basic_blocks);
4129 compute_ae_kill (comp, ae_kill);
4130 edge_list = pre_edge_lcm (gcse_file, n_exprs, transp, comp, antloc,
4131 ae_kill, &pre_insert_map, &pre_delete_map);
4136 /* Return non-zero if an occurrence of expression EXPR in OCCR_BB would reach
4139 VISITED is a pointer to a working buffer for tracking which BB's have
4140 been visited. It is NULL for the top-level call.
4142 We treat reaching expressions that go through blocks containing the same
4143 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
4144 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
4145 2 as not reaching. The intent is to improve the probability of finding
4146 only one reaching expression and to reduce register lifetimes by picking
4147 the closest such expression. */
4150 pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited)
4158 for (pred = BASIC_BLOCK (bb)->pred; pred != NULL; pred = pred->pred_next)
4160 int pred_bb = pred->src->index;
4162 if (pred->src == ENTRY_BLOCK_PTR
4163 /* Has predecessor has already been visited? */
4164 || visited[pred_bb])
4165 ;/* Nothing to do. */
4167 /* Does this predecessor generate this expression? */
4168 else if (TEST_BIT (comp[pred_bb], expr->bitmap_index))
4170 /* Is this the occurrence we're looking for?
4171 Note that there's only one generating occurrence per block
4172 so we just need to check the block number. */
4173 if (occr_bb == pred_bb)
4176 visited[pred_bb] = 1;
4178 /* Ignore this predecessor if it kills the expression. */
4179 else if (! TEST_BIT (transp[pred_bb], expr->bitmap_index))
4180 visited[pred_bb] = 1;
4182 /* Neither gen nor kill. */
4185 visited[pred_bb] = 1;
4186 if (pre_expr_reaches_here_p_work (occr_bb, expr, pred_bb, visited))
4191 /* All paths have been checked. */
4195 /* The wrapper for pre_expr_reaches_here_work that ensures that any
4196 memory allocated for that function is returned. */
4199 pre_expr_reaches_here_p (occr_bb, expr, bb)
4205 char *visited = (char *) xcalloc (n_basic_blocks, 1);
4207 rval = pre_expr_reaches_here_p_work(occr_bb, expr, bb, visited);
4214 /* Given an expr, generate RTL which we can insert at the end of a BB,
4215 or on an edge. Set the block number of any insns generated to
4219 process_insert_insn (expr)
4222 rtx reg = expr->reaching_reg;
4223 rtx pat, copied_expr;
4227 copied_expr = copy_rtx (expr->expr);
4228 emit_move_insn (reg, copied_expr);
4229 first_new_insn = get_insns ();
4230 pat = gen_sequence ();
4236 /* Add EXPR to the end of basic block BB.
4238 This is used by both the PRE and code hoisting.
4240 For PRE, we want to verify that the expr is either transparent
4241 or locally anticipatable in the target block. This check makes
4242 no sense for code hoisting. */
4245 insert_insn_end_bb (expr, bb, pre)
4250 rtx insn = BLOCK_END (bb);
4252 rtx reg = expr->reaching_reg;
4253 int regno = REGNO (reg);
4257 pat = process_insert_insn (expr);
4259 /* If the last insn is a jump, insert EXPR in front [taking care to
4260 handle cc0, etc. properly]. */
4262 if (GET_CODE (insn) == JUMP_INSN)
4268 /* If this is a jump table, then we can't insert stuff here. Since
4269 we know the previous real insn must be the tablejump, we insert
4270 the new instruction just before the tablejump. */
4271 if (GET_CODE (PATTERN (insn)) == ADDR_VEC
4272 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
4273 insn = prev_real_insn (insn);
4276 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
4277 if cc0 isn't set. */
4278 note = find_reg_note (insn, REG_CC_SETTER, NULL_RTX);
4280 insn = XEXP (note, 0);
4283 rtx maybe_cc0_setter = prev_nonnote_insn (insn);
4284 if (maybe_cc0_setter
4285 && GET_RTX_CLASS (GET_CODE (maybe_cc0_setter)) == 'i'
4286 && sets_cc0_p (PATTERN (maybe_cc0_setter)))
4287 insn = maybe_cc0_setter;
4290 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4291 new_insn = emit_block_insn_before (pat, insn, BASIC_BLOCK (bb));
4294 /* Likewise if the last insn is a call, as will happen in the presence
4295 of exception handling. */
4296 else if (GET_CODE (insn) == CALL_INSN)
4298 HARD_REG_SET parm_regs;
4302 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4303 we search backward and place the instructions before the first
4304 parameter is loaded. Do this for everyone for consistency and a
4305 presumtion that we'll get better code elsewhere as well.
4307 It should always be the case that we can put these instructions
4308 anywhere in the basic block with performing PRE optimizations.
4312 && !TEST_BIT (antloc[bb], expr->bitmap_index)
4313 && !TEST_BIT (transp[bb], expr->bitmap_index))
4316 /* Since different machines initialize their parameter registers
4317 in different orders, assume nothing. Collect the set of all
4318 parameter registers. */
4319 CLEAR_HARD_REG_SET (parm_regs);
4321 for (p = CALL_INSN_FUNCTION_USAGE (insn); p ; p = XEXP (p, 1))
4322 if (GET_CODE (XEXP (p, 0)) == USE
4323 && GET_CODE (XEXP (XEXP (p, 0), 0)) == REG)
4325 if (REGNO (XEXP (XEXP (p, 0), 0)) >= FIRST_PSEUDO_REGISTER)
4328 SET_HARD_REG_BIT (parm_regs, REGNO (XEXP (XEXP (p, 0), 0)));
4332 /* Search backward for the first set of a register in this set. */
4333 while (nparm_regs && BLOCK_HEAD (bb) != insn)
4335 insn = PREV_INSN (insn);
4336 p = single_set (insn);
4337 if (p && GET_CODE (SET_DEST (p)) == REG
4338 && REGNO (SET_DEST (p)) < FIRST_PSEUDO_REGISTER
4339 && TEST_HARD_REG_BIT (parm_regs, REGNO (SET_DEST (p))))
4341 CLEAR_HARD_REG_BIT (parm_regs, REGNO (SET_DEST (p)));
4346 /* If we found all the parameter loads, then we want to insert
4347 before the first parameter load.
4349 If we did not find all the parameter loads, then we might have
4350 stopped on the head of the block, which could be a CODE_LABEL.
4351 If we inserted before the CODE_LABEL, then we would be putting
4352 the insn in the wrong basic block. In that case, put the insn
4353 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4354 if (GET_CODE (insn) == CODE_LABEL)
4355 insn = NEXT_INSN (insn);
4356 else if (GET_CODE (insn) == NOTE
4357 && NOTE_LINE_NUMBER (insn) == NOTE_INSN_BASIC_BLOCK)
4358 insn = NEXT_INSN (insn);
4360 new_insn = emit_block_insn_before (pat, insn, BASIC_BLOCK (bb));
4364 new_insn = emit_insn_after (pat, insn);
4365 BLOCK_END (bb) = new_insn;
4368 /* Keep block number table up to date.
4369 Note, PAT could be a multiple insn sequence, we have to make
4370 sure that each insn in the sequence is handled. */
4371 if (GET_CODE (pat) == SEQUENCE)
4373 for (i = 0; i < XVECLEN (pat, 0); i++)
4375 rtx insn = XVECEXP (pat, 0, i);
4377 set_block_num (insn, bb);
4378 if (GET_RTX_CLASS (GET_CODE (insn)) == 'i')
4379 add_label_notes (PATTERN (insn), new_insn);
4381 note_stores (PATTERN (insn), record_set_info, insn);
4386 add_label_notes (SET_SRC (pat), new_insn);
4387 set_block_num (new_insn, bb);
4389 /* Keep register set table up to date. */
4390 record_one_set (regno, new_insn);
4393 gcse_create_count++;
4397 fprintf (gcse_file, "PRE/HOIST: end of bb %d, insn %d, ",
4398 bb, INSN_UID (new_insn));
4399 fprintf (gcse_file, "copying expression %d to reg %d\n",
4400 expr->bitmap_index, regno);
4404 /* Insert partially redundant expressions on edges in the CFG to make
4405 the expressions fully redundant. */
4408 pre_edge_insert (edge_list, index_map)
4409 struct edge_list *edge_list;
4410 struct expr **index_map;
4412 int e, i, j, num_edges, set_size, did_insert = 0;
4415 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4416 if it reaches any of the deleted expressions. */
4418 set_size = pre_insert_map[0]->size;
4419 num_edges = NUM_EDGES (edge_list);
4420 inserted = sbitmap_vector_alloc (num_edges, n_exprs);
4421 sbitmap_vector_zero (inserted, num_edges);
4423 for (e = 0; e < num_edges; e++)
4426 basic_block pred = INDEX_EDGE_PRED_BB (edge_list, e);
4427 int bb = pred->index;
4429 for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
4431 SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
4433 for (j = indx; insert && j < n_exprs; j++, insert >>= 1)
4434 if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
4436 struct expr *expr = index_map[j];
4439 /* Now look at each deleted occurence of this expression. */
4440 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4442 if (! occr->deleted_p)
4445 /* Insert this expression on this edge if if it would
4446 reach the deleted occurence in BB. */
4447 if (!TEST_BIT (inserted[e], j))
4450 edge eg = INDEX_EDGE (edge_list, e);
4452 /* We can't insert anything on an abnormal and
4453 critical edge, so we insert the insn at the end of
4454 the previous block. There are several alternatives
4455 detailed in Morgans book P277 (sec 10.5) for
4456 handling this situation. This one is easiest for
4459 if ((eg->flags & EDGE_ABNORMAL) == EDGE_ABNORMAL)
4460 insert_insn_end_bb (index_map[j], bb, 0);
4463 insn = process_insert_insn (index_map[j]);
4464 insert_insn_on_edge (insn, eg);
4469 fprintf (gcse_file, "PRE/HOIST: edge (%d,%d), ",
4471 INDEX_EDGE_SUCC_BB (edge_list, e)->index);
4472 fprintf (gcse_file, "copy expression %d\n",
4473 expr->bitmap_index);
4476 SET_BIT (inserted[e], j);
4478 gcse_create_count++;
4489 /* Copy the result of INSN to REG. INDX is the expression number. */
4492 pre_insert_copy_insn (expr, insn)
4496 rtx reg = expr->reaching_reg;
4497 int regno = REGNO (reg);
4498 int indx = expr->bitmap_index;
4499 rtx set = single_set (insn);
4501 int bb = BLOCK_NUM (insn);
4506 new_insn = emit_insn_after (gen_rtx_SET (VOIDmode, reg, SET_DEST (set)),
4509 /* Keep block number table up to date. */
4510 set_block_num (new_insn, bb);
4512 /* Keep register set table up to date. */
4513 record_one_set (regno, new_insn);
4514 if (insn == BLOCK_END (bb))
4515 BLOCK_END (bb) = new_insn;
4517 gcse_create_count++;
4521 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4522 BLOCK_NUM (insn), INSN_UID (new_insn), indx,
4523 INSN_UID (insn), regno);
4526 /* Copy available expressions that reach the redundant expression
4527 to `reaching_reg'. */
4530 pre_insert_copies ()
4537 /* For each available expression in the table, copy the result to
4538 `reaching_reg' if the expression reaches a deleted one.
4540 ??? The current algorithm is rather brute force.
4541 Need to do some profiling. */
4543 for (i = 0; i < expr_hash_table_size; i++)
4544 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4546 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4547 we don't want to insert a copy here because the expression may not
4548 really be redundant. So only insert an insn if the expression was
4549 deleted. This test also avoids further processing if the
4550 expression wasn't deleted anywhere. */
4551 if (expr->reaching_reg == NULL)
4554 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4556 if (! occr->deleted_p)
4559 for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
4561 rtx insn = avail->insn;
4563 /* No need to handle this one if handled already. */
4564 if (avail->copied_p)
4567 /* Don't handle this one if it's a redundant one. */
4568 if (TEST_BIT (pre_redundant_insns, INSN_CUID (insn)))
4571 /* Or if the expression doesn't reach the deleted one. */
4572 if (! pre_expr_reaches_here_p (BLOCK_NUM (avail->insn), expr,
4573 BLOCK_NUM (occr->insn)))
4576 /* Copy the result of avail to reaching_reg. */
4577 pre_insert_copy_insn (expr, insn);
4578 avail->copied_p = 1;
4584 /* Delete redundant computations.
4585 Deletion is done by changing the insn to copy the `reaching_reg' of
4586 the expression into the result of the SET. It is left to later passes
4587 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4589 Returns non-zero if a change is made. */
4598 /* Compute the expressions which are redundant and need to be replaced by
4599 copies from the reaching reg to the target reg. */
4600 for (bb = 0; bb < n_basic_blocks; bb++)
4601 sbitmap_copy (temp_bitmap[bb], pre_delete_map[bb]);
4604 for (i = 0; i < expr_hash_table_size; i++)
4605 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4607 int indx = expr->bitmap_index;
4609 /* We only need to search antic_occr since we require
4612 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4614 rtx insn = occr->insn;
4616 int bb = BLOCK_NUM (insn);
4618 if (TEST_BIT (temp_bitmap[bb], indx))
4620 set = single_set (insn);
4624 /* Create a pseudo-reg to store the result of reaching
4625 expressions into. Get the mode for the new pseudo from
4626 the mode of the original destination pseudo. */
4627 if (expr->reaching_reg == NULL)
4629 = gen_reg_rtx (GET_MODE (SET_DEST (set)));
4631 /* In theory this should never fail since we're creating
4634 However, on the x86 some of the movXX patterns actually
4635 contain clobbers of scratch regs. This may cause the
4636 insn created by validate_change to not match any pattern
4637 and thus cause validate_change to fail. */
4638 if (validate_change (insn, &SET_SRC (set),
4639 expr->reaching_reg, 0))
4641 occr->deleted_p = 1;
4642 SET_BIT (pre_redundant_insns, INSN_CUID (insn));
4650 "PRE: redundant insn %d (expression %d) in ",
4651 INSN_UID (insn), indx);
4652 fprintf (gcse_file, "bb %d, reaching reg is %d\n",
4653 bb, REGNO (expr->reaching_reg));
4662 /* Perform GCSE optimizations using PRE.
4663 This is called by one_pre_gcse_pass after all the dataflow analysis
4666 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4667 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4668 Compiler Design and Implementation.
4670 ??? A new pseudo reg is created to hold the reaching expression. The nice
4671 thing about the classical approach is that it would try to use an existing
4672 reg. If the register can't be adequately optimized [i.e. we introduce
4673 reload problems], one could add a pass here to propagate the new register
4676 ??? We don't handle single sets in PARALLELs because we're [currently] not
4677 able to copy the rest of the parallel when we insert copies to create full
4678 redundancies from partial redundancies. However, there's no reason why we
4679 can't handle PARALLELs in the cases where there are no partial
4687 struct expr **index_map;
4690 /* Compute a mapping from expression number (`bitmap_index') to
4691 hash table entry. */
4693 index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
4694 for (i = 0; i < expr_hash_table_size; i++)
4695 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
4696 index_map[expr->bitmap_index] = expr;
4698 /* Reset bitmap used to track which insns are redundant. */
4699 pre_redundant_insns = sbitmap_alloc (max_cuid);
4700 sbitmap_zero (pre_redundant_insns);
4702 /* Delete the redundant insns first so that
4703 - we know what register to use for the new insns and for the other
4704 ones with reaching expressions
4705 - we know which insns are redundant when we go to create copies */
4707 changed = pre_delete ();
4709 did_insert = pre_edge_insert (edge_list, index_map);
4711 /* In other places with reaching expressions, copy the expression to the
4712 specially allocated pseudo-reg that reaches the redundant expr. */
4713 pre_insert_copies ();
4716 commit_edge_insertions ();
4721 free (pre_redundant_insns);
4725 /* Top level routine to perform one PRE GCSE pass.
4727 Return non-zero if a change was made. */
4730 one_pre_gcse_pass (pass)
4735 gcse_subst_count = 0;
4736 gcse_create_count = 0;
4738 alloc_expr_hash_table (max_cuid);
4739 add_noreturn_fake_exit_edges ();
4740 compute_expr_hash_table ();
4742 dump_hash_table (gcse_file, "Expression", expr_hash_table,
4743 expr_hash_table_size, n_exprs);
4747 alloc_pre_mem (n_basic_blocks, n_exprs);
4748 compute_pre_data ();
4749 changed |= pre_gcse ();
4750 free_edge_list (edge_list);
4754 remove_fake_edges ();
4755 free_expr_hash_table ();
4759 fprintf (gcse_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
4760 current_function_name, pass, bytes_used);
4761 fprintf (gcse_file, "%d substs, %d insns created\n",
4762 gcse_subst_count, gcse_create_count);
4768 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to INSN.
4769 We have to add REG_LABEL notes, because the following loop optimization
4770 pass requires them. */
4772 /* ??? This is very similar to the loop.c add_label_notes function. We
4773 could probably share code here. */
4775 /* ??? If there was a jump optimization pass after gcse and before loop,
4776 then we would not need to do this here, because jump would add the
4777 necessary REG_LABEL notes. */
4780 add_label_notes (x, insn)
4784 enum rtx_code code = GET_CODE (x);
4788 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
4790 /* This code used to ignore labels that referred to dispatch tables to
4791 avoid flow generating (slighly) worse code.
4793 We no longer ignore such label references (see LABEL_REF handling in
4794 mark_jump_label for additional information). */
4796 REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_LABEL, XEXP (x, 0),
4801 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
4804 add_label_notes (XEXP (x, i), insn);
4805 else if (fmt[i] == 'E')
4806 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4807 add_label_notes (XVECEXP (x, i, j), insn);
4811 /* Compute transparent outgoing information for each block.
4813 An expression is transparent to an edge unless it is killed by
4814 the edge itself. This can only happen with abnormal control flow,
4815 when the edge is traversed through a call. This happens with
4816 non-local labels and exceptions.
4818 This would not be necessary if we split the edge. While this is
4819 normally impossible for abnormal critical edges, with some effort
4820 it should be possible with exception handling, since we still have
4821 control over which handler should be invoked. But due to increased
4822 EH table sizes, this may not be worthwhile. */
4825 compute_transpout ()
4831 sbitmap_vector_ones (transpout, n_basic_blocks);
4833 for (bb = 0; bb < n_basic_blocks; ++bb)
4835 /* Note that flow inserted a nop a the end of basic blocks that
4836 end in call instructions for reasons other than abnormal
4838 if (GET_CODE (BLOCK_END (bb)) != CALL_INSN)
4841 for (i = 0; i < expr_hash_table_size; i++)
4842 for (expr = expr_hash_table[i]; expr ; expr = expr->next_same_hash)
4843 if (GET_CODE (expr->expr) == MEM)
4845 if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
4846 && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
4849 /* ??? Optimally, we would use interprocedural alias
4850 analysis to determine if this mem is actually killed
4852 RESET_BIT (transpout[bb], expr->bitmap_index);
4857 /* Removal of useless null pointer checks */
4859 /* Called via note_stores. X is set by SETTER. If X is a register we must
4860 invalidate nonnull_local and set nonnull_killed. DATA is really a
4861 `null_pointer_info *'.
4863 We ignore hard registers. */
4866 invalidate_nonnull_info (x, setter, data)
4868 rtx setter ATTRIBUTE_UNUSED;
4872 struct null_pointer_info* npi = (struct null_pointer_info *) data;
4876 while (GET_CODE (x) == SUBREG)
4879 /* Ignore anything that is not a register or is a hard register. */
4880 if (GET_CODE (x) != REG
4881 || REGNO (x) < npi->min_reg
4882 || REGNO (x) >= npi->max_reg)
4885 regno = REGNO (x) - npi->min_reg;
4887 RESET_BIT (npi->nonnull_local[npi->current_block], regno);
4888 SET_BIT (npi->nonnull_killed[npi->current_block], regno);
4891 /* Do null-pointer check elimination for the registers indicated in
4892 NPI. NONNULL_AVIN and NONNULL_AVOUT are pre-allocated sbitmaps;
4893 they are not our responsibility to free. */
4896 delete_null_pointer_checks_1 (block_reg, nonnull_avin, nonnull_avout, npi)
4898 sbitmap *nonnull_avin;
4899 sbitmap *nonnull_avout;
4900 struct null_pointer_info *npi;
4904 sbitmap *nonnull_local = npi->nonnull_local;
4905 sbitmap *nonnull_killed = npi->nonnull_killed;
4907 /* Compute local properties, nonnull and killed. A register will have
4908 the nonnull property if at the end of the current block its value is
4909 known to be nonnull. The killed property indicates that somewhere in
4910 the block any information we had about the register is killed.
4912 Note that a register can have both properties in a single block. That
4913 indicates that it's killed, then later in the block a new value is
4915 sbitmap_vector_zero (nonnull_local, n_basic_blocks);
4916 sbitmap_vector_zero (nonnull_killed, n_basic_blocks);
4918 for (current_block = 0; current_block < n_basic_blocks; current_block++)
4920 rtx insn, stop_insn;
4922 /* Set the current block for invalidate_nonnull_info. */
4923 npi->current_block = current_block;
4925 /* Scan each insn in the basic block looking for memory references and
4927 stop_insn = NEXT_INSN (BLOCK_END (current_block));
4928 for (insn = BLOCK_HEAD (current_block);
4930 insn = NEXT_INSN (insn))
4935 /* Ignore anything that is not a normal insn. */
4936 if (GET_RTX_CLASS (GET_CODE (insn)) != 'i')
4939 /* Basically ignore anything that is not a simple SET. We do have
4940 to make sure to invalidate nonnull_local and set nonnull_killed
4941 for such insns though. */
4942 set = single_set (insn);
4945 note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
4949 /* See if we've got a useable memory load. We handle it first
4950 in case it uses its address register as a dest (which kills
4951 the nonnull property). */
4952 if (GET_CODE (SET_SRC (set)) == MEM
4953 && GET_CODE ((reg = XEXP (SET_SRC (set), 0))) == REG
4954 && REGNO (reg) >= npi->min_reg
4955 && REGNO (reg) < npi->max_reg)
4956 SET_BIT (nonnull_local[current_block],
4957 REGNO (reg) - npi->min_reg);
4959 /* Now invalidate stuff clobbered by this insn. */
4960 note_stores (PATTERN (insn), invalidate_nonnull_info, npi);
4962 /* And handle stores, we do these last since any sets in INSN can
4963 not kill the nonnull property if it is derived from a MEM
4964 appearing in a SET_DEST. */
4965 if (GET_CODE (SET_DEST (set)) == MEM
4966 && GET_CODE ((reg = XEXP (SET_DEST (set), 0))) == REG
4967 && REGNO (reg) >= npi->min_reg
4968 && REGNO (reg) < npi->max_reg)
4969 SET_BIT (nonnull_local[current_block],
4970 REGNO (reg) - npi->min_reg);
4974 /* Now compute global properties based on the local properties. This
4975 is a classic global availablity algorithm. */
4976 compute_available (nonnull_local, nonnull_killed,
4977 nonnull_avout, nonnull_avin);
4979 /* Now look at each bb and see if it ends with a compare of a value
4981 for (bb = 0; bb < n_basic_blocks; bb++)
4983 rtx last_insn = BLOCK_END (bb);
4984 rtx condition, earliest;
4985 int compare_and_branch;
4987 /* Since MIN_REG is always at least FIRST_PSEUDO_REGISTER, and
4988 since BLOCK_REG[BB] is zero if this block did not end with a
4989 comparison against zero, this condition works. */
4990 if (block_reg[bb] < npi->min_reg
4991 || block_reg[bb] >= npi->max_reg)
4994 /* LAST_INSN is a conditional jump. Get its condition. */
4995 condition = get_condition (last_insn, &earliest);
4997 /* If we can't determine the condition then skip. */
5001 /* Is the register known to have a nonzero value? */
5002 if (!TEST_BIT (nonnull_avout[bb], block_reg[bb] - npi->min_reg))
5005 /* Try to compute whether the compare/branch at the loop end is one or
5006 two instructions. */
5007 if (earliest == last_insn)
5008 compare_and_branch = 1;
5009 else if (earliest == prev_nonnote_insn (last_insn))
5010 compare_and_branch = 2;
5014 /* We know the register in this comparison is nonnull at exit from
5015 this block. We can optimize this comparison. */
5016 if (GET_CODE (condition) == NE)
5020 new_jump = emit_jump_insn_before (gen_jump (JUMP_LABEL (last_insn)),
5022 JUMP_LABEL (new_jump) = JUMP_LABEL (last_insn);
5023 LABEL_NUSES (JUMP_LABEL (new_jump))++;
5024 emit_barrier_after (new_jump);
5026 delete_insn (last_insn);
5027 if (compare_and_branch == 2)
5028 delete_insn (earliest);
5030 /* Don't check this block again. (Note that BLOCK_END is
5031 invalid here; we deleted the last instruction in the
5037 /* Find EQ/NE comparisons against zero which can be (indirectly) evaluated
5040 This is conceptually similar to global constant/copy propagation and
5041 classic global CSE (it even uses the same dataflow equations as cprop).
5043 If a register is used as memory address with the form (mem (reg)), then we
5044 know that REG can not be zero at that point in the program. Any instruction
5045 which sets REG "kills" this property.
5047 So, if every path leading to a conditional branch has an available memory
5048 reference of that form, then we know the register can not have the value
5049 zero at the conditional branch.
5051 So we merely need to compute the local properies and propagate that data
5052 around the cfg, then optimize where possible.
5054 We run this pass two times. Once before CSE, then again after CSE. This
5055 has proven to be the most profitable approach. It is rare for new
5056 optimization opportunities of this nature to appear after the first CSE
5059 This could probably be integrated with global cprop with a little work. */
5062 delete_null_pointer_checks (f)
5065 sbitmap *nonnull_avin, *nonnull_avout;
5071 struct null_pointer_info npi;
5073 /* First break the program into basic blocks. */
5074 find_basic_blocks (f, max_reg_num (), NULL);
5077 /* If we have only a single block, then there's nothing to do. */
5078 if (n_basic_blocks <= 1)
5080 /* Free storage allocated by find_basic_blocks. */
5081 free_basic_block_vars (0);
5085 /* Trying to perform global optimizations on flow graphs which have
5086 a high connectivity will take a long time and is unlikely to be
5087 particularly useful.
5089 In normal circumstances a cfg should have about twice has many edges
5090 as blocks. But we do not want to punish small functions which have
5091 a couple switch statements. So we require a relatively large number
5092 of basic blocks and the ratio of edges to blocks to be high. */
5093 if (n_basic_blocks > 1000 && n_edges / n_basic_blocks >= 20)
5095 /* Free storage allocated by find_basic_blocks. */
5096 free_basic_block_vars (0);
5100 /* We need four bitmaps, each with a bit for each register in each
5102 max_reg = max_reg_num ();
5103 regs_per_pass = get_bitmap_width (4, n_basic_blocks, max_reg);
5105 /* Allocate bitmaps to hold local and global properties. */
5106 npi.nonnull_local = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5107 npi.nonnull_killed = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5108 nonnull_avin = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5109 nonnull_avout = sbitmap_vector_alloc (n_basic_blocks, regs_per_pass);
5111 /* Go through the basic blocks, seeing whether or not each block
5112 ends with a conditional branch whose condition is a comparison
5113 against zero. Record the register compared in BLOCK_REG. */
5114 block_reg = (int *) xcalloc (n_basic_blocks, sizeof (int));
5115 for (bb = 0; bb < n_basic_blocks; bb++)
5117 rtx last_insn = BLOCK_END (bb);
5118 rtx condition, earliest, reg;
5120 /* We only want conditional branches. */
5121 if (GET_CODE (last_insn) != JUMP_INSN
5122 || !condjump_p (last_insn)
5123 || simplejump_p (last_insn))
5126 /* LAST_INSN is a conditional jump. Get its condition. */
5127 condition = get_condition (last_insn, &earliest);
5129 /* If we were unable to get the condition, or it is not a equality
5130 comparison against zero then there's nothing we can do. */
5132 || (GET_CODE (condition) != NE && GET_CODE (condition) != EQ)
5133 || GET_CODE (XEXP (condition, 1)) != CONST_INT
5134 || (XEXP (condition, 1)
5135 != CONST0_RTX (GET_MODE (XEXP (condition, 0)))))
5138 /* We must be checking a register against zero. */
5139 reg = XEXP (condition, 0);
5140 if (GET_CODE (reg) != REG)
5143 block_reg[bb] = REGNO (reg);
5146 /* Go through the algorithm for each block of registers. */
5147 for (reg = FIRST_PSEUDO_REGISTER; reg < max_reg; reg += regs_per_pass)
5150 npi.max_reg = MIN (reg + regs_per_pass, max_reg);
5151 delete_null_pointer_checks_1 (block_reg, nonnull_avin,
5152 nonnull_avout, &npi);
5155 /* Free storage allocated by find_basic_blocks. */
5156 free_basic_block_vars (0);
5158 /* Free the table of registers compared at the end of every block. */
5162 free (npi.nonnull_local);
5163 free (npi.nonnull_killed);
5164 free (nonnull_avin);
5165 free (nonnull_avout);
5168 /* Code Hoisting variables and subroutines. */
5170 /* Very busy expressions. */
5171 static sbitmap *hoist_vbein;
5172 static sbitmap *hoist_vbeout;
5174 /* Hoistable expressions. */
5175 static sbitmap *hoist_exprs;
5177 /* Dominator bitmaps. */
5178 static sbitmap *dominators;
5180 /* ??? We could compute post dominators and run this algorithm in
5181 reverse to to perform tail merging, doing so would probably be
5182 more effective than the tail merging code in jump.c.
5184 It's unclear if tail merging could be run in parallel with
5185 code hoisting. It would be nice. */
5187 /* Allocate vars used for code hoisting analysis. */
5190 alloc_code_hoist_mem (n_blocks, n_exprs)
5191 int n_blocks, n_exprs;
5193 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
5194 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
5195 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
5197 hoist_vbein = sbitmap_vector_alloc (n_blocks, n_exprs);
5198 hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
5199 hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
5200 transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
5202 dominators = sbitmap_vector_alloc (n_blocks, n_blocks);
5205 /* Free vars used for code hoisting analysis. */
5208 free_code_hoist_mem ()
5215 free (hoist_vbeout);
5222 /* Compute the very busy expressions at entry/exit from each block.
5224 An expression is very busy if all paths from a given point
5225 compute the expression. */
5228 compute_code_hoist_vbeinout ()
5230 int bb, changed, passes;
5232 sbitmap_vector_zero (hoist_vbeout, n_basic_blocks);
5233 sbitmap_vector_zero (hoist_vbein, n_basic_blocks);
5242 /* We scan the blocks in the reverse order to speed up
5244 for (bb = n_basic_blocks - 1; bb >= 0; bb--)
5246 changed |= sbitmap_a_or_b_and_c (hoist_vbein[bb], antloc[bb],
5247 hoist_vbeout[bb], transp[bb]);
5248 if (bb != n_basic_blocks - 1)
5249 sbitmap_intersection_of_succs (hoist_vbeout[bb], hoist_vbein, bb);
5256 fprintf (gcse_file, "hoisting vbeinout computation: %d passes\n", passes);
5259 /* Top level routine to do the dataflow analysis needed by code hoisting. */
5262 compute_code_hoist_data ()
5264 compute_local_properties (transp, comp, antloc, 0);
5265 compute_transpout ();
5266 compute_code_hoist_vbeinout ();
5267 compute_flow_dominators (dominators, NULL);
5269 fprintf (gcse_file, "\n");
5272 /* Determine if the expression identified by EXPR_INDEX would
5273 reach BB unimpared if it was placed at the end of EXPR_BB.
5275 It's unclear exactly what Muchnick meant by "unimpared". It seems
5276 to me that the expression must either be computed or transparent in
5277 *every* block in the path(s) from EXPR_BB to BB. Any other definition
5278 would allow the expression to be hoisted out of loops, even if
5279 the expression wasn't a loop invariant.
5281 Contrast this to reachability for PRE where an expression is
5282 considered reachable if *any* path reaches instead of *all*
5286 hoist_expr_reaches_here_p (expr_bb, expr_index, bb, visited)
5293 int visited_allocated_locally = 0;
5296 if (visited == NULL)
5298 visited_allocated_locally = 1;
5299 visited = xcalloc (n_basic_blocks, 1);
5302 visited[expr_bb] = 1;
5303 for (pred = BASIC_BLOCK (bb)->pred; pred != NULL; pred = pred->pred_next)
5305 int pred_bb = pred->src->index;
5307 if (pred->src == ENTRY_BLOCK_PTR)
5309 else if (visited[pred_bb])
5312 /* Does this predecessor generate this expression? */
5313 else if (TEST_BIT (comp[pred_bb], expr_index))
5315 else if (! TEST_BIT (transp[pred_bb], expr_index))
5321 visited[pred_bb] = 1;
5322 if (! hoist_expr_reaches_here_p (expr_bb, expr_index,
5327 if (visited_allocated_locally)
5330 return (pred == NULL);
5333 /* Actually perform code hoisting. */
5338 int bb, dominated, i;
5339 struct expr **index_map;
5342 sbitmap_vector_zero (hoist_exprs, n_basic_blocks);
5344 /* Compute a mapping from expression number (`bitmap_index') to
5345 hash table entry. */
5347 index_map = (struct expr **) xcalloc (n_exprs, sizeof (struct expr *));
5348 for (i = 0; i < expr_hash_table_size; i++)
5349 for (expr = expr_hash_table[i]; expr != NULL; expr = expr->next_same_hash)
5350 index_map[expr->bitmap_index] = expr;
5352 /* Walk over each basic block looking for potentially hoistable
5353 expressions, nothing gets hoisted from the entry block. */
5354 for (bb = 0; bb < n_basic_blocks; bb++)
5357 int insn_inserted_p;
5359 /* Examine each expression that is very busy at the exit of this
5360 block. These are the potentially hoistable expressions. */
5361 for (i = 0; i < hoist_vbeout[bb]->n_bits; i++)
5365 if (TEST_BIT (hoist_vbeout[bb], i) && TEST_BIT (transpout[bb], i))
5367 /* We've found a potentially hoistable expression, now
5368 we look at every block BB dominates to see if it
5369 computes the expression. */
5370 for (dominated = 0; dominated < n_basic_blocks; dominated++)
5372 /* Ignore self dominance. */
5374 || ! TEST_BIT (dominators[dominated], bb))
5377 /* We've found a dominated block, now see if it computes
5378 the busy expression and whether or not moving that
5379 expression to the "beginning" of that block is safe. */
5380 if (!TEST_BIT (antloc[dominated], i))
5383 /* Note if the expression would reach the dominated block
5384 unimpared if it was placed at the end of BB.
5386 Keep track of how many times this expression is hoistable
5387 from a dominated block into BB. */
5388 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
5392 /* If we found more than one hoistable occurence of this
5393 expression, then note it in the bitmap of expressions to
5394 hoist. It makes no sense to hoist things which are computed
5395 in only one BB, and doing so tends to pessimize register
5396 allocation. One could increase this value to try harder
5397 to avoid any possible code expansion due to register
5398 allocation issues; however experiments have shown that
5399 the vast majority of hoistable expressions are only movable
5400 from two successors, so raising this threshhold is likely
5401 to nullify any benefit we get from code hoisting. */
5404 SET_BIT (hoist_exprs[bb], i);
5410 /* If we found nothing to hoist, then quit now. */
5414 /* Loop over all the hoistable expressions. */
5415 for (i = 0; i < hoist_exprs[bb]->n_bits; i++)
5417 /* We want to insert the expression into BB only once, so
5418 note when we've inserted it. */
5419 insn_inserted_p = 0;
5421 /* These tests should be the same as the tests above. */
5422 if (TEST_BIT (hoist_vbeout[bb], i))
5424 /* We've found a potentially hoistable expression, now
5425 we look at every block BB dominates to see if it
5426 computes the expression. */
5427 for (dominated = 0; dominated < n_basic_blocks; dominated++)
5429 /* Ignore self dominance. */
5431 || ! TEST_BIT (dominators[dominated], bb))
5434 /* We've found a dominated block, now see if it computes
5435 the busy expression and whether or not moving that
5436 expression to the "beginning" of that block is safe. */
5437 if (!TEST_BIT (antloc[dominated], i))
5440 /* The expression is computed in the dominated block and
5441 it would be safe to compute it at the start of the
5442 dominated block. Now we have to determine if the
5443 expresion would reach the dominated block if it was
5444 placed at the end of BB. */
5445 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
5447 struct expr *expr = index_map[i];
5448 struct occr *occr = expr->antic_occr;
5452 /* Find the right occurence of this expression. */
5453 while (BLOCK_NUM (occr->insn) != dominated && occr)
5456 /* Should never happen. */
5462 set = single_set (insn);
5466 /* Create a pseudo-reg to store the result of reaching
5467 expressions into. Get the mode for the new pseudo
5468 from the mode of the original destination pseudo. */
5469 if (expr->reaching_reg == NULL)
5471 = gen_reg_rtx (GET_MODE (SET_DEST (set)));
5473 /* In theory this should never fail since we're creating
5476 However, on the x86 some of the movXX patterns
5477 actually contain clobbers of scratch regs. This may
5478 cause the insn created by validate_change to not
5479 match any pattern and thus cause validate_change to
5481 if (validate_change (insn, &SET_SRC (set),
5482 expr->reaching_reg, 0))
5484 occr->deleted_p = 1;
5485 if (!insn_inserted_p)
5487 insert_insn_end_bb (index_map[i], bb, 0);
5488 insn_inserted_p = 1;
5500 /* Top level routine to perform one code hoisting (aka unification) pass
5502 Return non-zero if a change was made. */
5505 one_code_hoisting_pass ()
5509 alloc_expr_hash_table (max_cuid);
5510 compute_expr_hash_table ();
5512 dump_hash_table (gcse_file, "Code Hosting Expressions", expr_hash_table,
5513 expr_hash_table_size, n_exprs);
5517 alloc_code_hoist_mem (n_basic_blocks, n_exprs);
5518 compute_code_hoist_data ();
5520 free_code_hoist_mem ();
5523 free_expr_hash_table ();