1 /* Global common subexpression elimination/Partial redundancy elimination
2 and global constant/copy propagation for GNU compiler.
3 Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005
4 Free Software Foundation, Inc.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 2, or (at your option) any later
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING. If not, write to the Free
20 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
24 - reordering of memory allocation and freeing to be more space efficient
25 - do rough calc of how many regs are needed in each block, and a rough
26 calc of how many regs are available in each class and use that to
27 throttle back the code in cases where RTX_COST is minimal.
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.
148 #include "coretypes.h"
156 #include "hard-reg-set.h"
159 #include "insn-config.h"
161 #include "basic-block.h"
163 #include "function.h"
172 #include "tree-pass.h"
175 /* Propagate flow information through back edges and thus enable PRE's
176 moving loop invariant calculations out of loops.
178 Originally this tended to create worse overall code, but several
179 improvements during the development of PRE seem to have made following
180 back edges generally a win.
182 Note much of the loop invariant code motion done here would normally
183 be done by loop.c, which has more heuristics for when to move invariants
184 out of loops. At some point we might need to move some of those
185 heuristics into gcse.c. */
187 /* We support GCSE via Partial Redundancy Elimination. PRE optimizations
188 are a superset of those done by GCSE.
190 We perform the following steps:
192 1) Compute basic block information.
194 2) Compute table of places where registers are set.
196 3) Perform copy/constant propagation.
198 4) Perform global cse using lazy code motion if not optimizing
199 for size, or code hoisting if we are.
201 5) Perform another pass of copy/constant propagation.
203 Two passes of copy/constant propagation are done because the first one
204 enables more GCSE and the second one helps to clean up the copies that
205 GCSE creates. This is needed more for PRE than for Classic because Classic
206 GCSE will try to use an existing register containing the common
207 subexpression rather than create a new one. This is harder to do for PRE
208 because of the code motion (which Classic GCSE doesn't do).
210 Expressions we are interested in GCSE-ing are of the form
211 (set (pseudo-reg) (expression)).
212 Function want_to_gcse_p says what these are.
214 PRE handles moving invariant expressions out of loops (by treating them as
215 partially redundant).
217 Eventually it would be nice to replace cse.c/gcse.c with SSA (static single
218 assignment) based GVN (global value numbering). L. T. Simpson's paper
219 (Rice University) on value numbering is a useful reference for this.
221 **********************
223 We used to support multiple passes but there are diminishing returns in
224 doing so. The first pass usually makes 90% of the changes that are doable.
225 A second pass can make a few more changes made possible by the first pass.
226 Experiments show any further passes don't make enough changes to justify
229 A study of spec92 using an unlimited number of passes:
230 [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
231 [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
232 [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
234 It was found doing copy propagation between each pass enables further
237 PRE is quite expensive in complicated functions because the DFA can take
238 a while to converge. Hence we only perform one pass. The parameter
239 max-gcse-passes can be modified if one wants to experiment.
241 **********************
243 The steps for PRE are:
245 1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
247 2) Perform the data flow analysis for PRE.
249 3) Delete the redundant instructions
251 4) Insert the required copies [if any] that make the partially
252 redundant instructions fully redundant.
254 5) For other reaching expressions, insert an instruction to copy the value
255 to a newly created pseudo that will reach the redundant instruction.
257 The deletion is done first so that when we do insertions we
258 know which pseudo reg to use.
260 Various papers have argued that PRE DFA is expensive (O(n^2)) and others
261 argue it is not. The number of iterations for the algorithm to converge
262 is typically 2-4 so I don't view it as that expensive (relatively speaking).
264 PRE GCSE depends heavily on the second CSE pass to clean up the copies
265 we create. To make an expression reach the place where it's redundant,
266 the result of the expression is copied to a new register, and the redundant
267 expression is deleted by replacing it with this new register. Classic GCSE
268 doesn't have this problem as much as it computes the reaching defs of
269 each register in each block and thus can try to use an existing
272 /* GCSE global vars. */
275 static FILE *gcse_file
;
277 /* Note whether or not we should run jump optimization after gcse. We
278 want to do this for two cases.
280 * If we changed any jumps via cprop.
282 * If we added any labels via edge splitting. */
283 static int run_jump_opt_after_gcse
;
285 /* Bitmaps are normally not included in debugging dumps.
286 However it's useful to be able to print them from GDB.
287 We could create special functions for this, but it's simpler to
288 just allow passing stderr to the dump_foo fns. Since stderr can
289 be a macro, we store a copy here. */
290 static FILE *debug_stderr
;
292 /* An obstack for our working variables. */
293 static struct obstack gcse_obstack
;
295 struct reg_use
{rtx reg_rtx
; };
297 /* Hash table of expressions. */
301 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
303 /* Index in the available expression bitmaps. */
305 /* Next entry with the same hash. */
306 struct expr
*next_same_hash
;
307 /* List of anticipatable occurrences in basic blocks in the function.
308 An "anticipatable occurrence" is one that is the first occurrence in the
309 basic block, the operands are not modified in the basic block prior
310 to the occurrence and the output is not used between the start of
311 the block and the occurrence. */
312 struct occr
*antic_occr
;
313 /* List of available occurrence in basic blocks in the function.
314 An "available occurrence" is one that is the last occurrence in the
315 basic block and the operands are not modified by following statements in
316 the basic block [including this insn]. */
317 struct occr
*avail_occr
;
318 /* Non-null if the computation is PRE redundant.
319 The value is the newly created pseudo-reg to record a copy of the
320 expression in all the places that reach the redundant copy. */
324 /* Occurrence of an expression.
325 There is one per basic block. If a pattern appears more than once the
326 last appearance is used [or first for anticipatable expressions]. */
330 /* Next occurrence of this expression. */
332 /* The insn that computes the expression. */
334 /* Nonzero if this [anticipatable] occurrence has been deleted. */
336 /* Nonzero if this [available] occurrence has been copied to
338 /* ??? This is mutually exclusive with deleted_p, so they could share
343 /* Expression and copy propagation hash tables.
344 Each hash table is an array of buckets.
345 ??? It is known that if it were an array of entries, structure elements
346 `next_same_hash' and `bitmap_index' wouldn't be necessary. However, it is
347 not clear whether in the final analysis a sufficient amount of memory would
348 be saved as the size of the available expression bitmaps would be larger
349 [one could build a mapping table without holes afterwards though].
350 Someday I'll perform the computation and figure it out. */
355 This is an array of `expr_hash_table_size' elements. */
358 /* Size of the hash table, in elements. */
361 /* Number of hash table elements. */
362 unsigned int n_elems
;
364 /* Whether the table is expression of copy propagation one. */
368 /* Expression hash table. */
369 static struct hash_table expr_hash_table
;
371 /* Copy propagation hash table. */
372 static struct hash_table set_hash_table
;
374 /* Mapping of uids to cuids.
375 Only real insns get cuids. */
376 static int *uid_cuid
;
378 /* Highest UID in UID_CUID. */
381 /* Get the cuid of an insn. */
382 #ifdef ENABLE_CHECKING
383 #define INSN_CUID(INSN) \
384 (gcc_assert (INSN_UID (INSN) <= max_uid), uid_cuid[INSN_UID (INSN)])
386 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
389 /* Number of cuids. */
392 /* Mapping of cuids to insns. */
393 static rtx
*cuid_insn
;
395 /* Get insn from cuid. */
396 #define CUID_INSN(CUID) (cuid_insn[CUID])
398 /* Maximum register number in function prior to doing gcse + 1.
399 Registers created during this pass have regno >= max_gcse_regno.
400 This is named with "gcse" to not collide with global of same name. */
401 static unsigned int max_gcse_regno
;
403 /* Table of registers that are modified.
405 For each register, each element is a list of places where the pseudo-reg
408 For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
409 requires knowledge of which blocks kill which regs [and thus could use
410 a bitmap instead of the lists `reg_set_table' uses].
412 `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
413 num-regs) [however perhaps it may be useful to keep the data as is]. One
414 advantage of recording things this way is that `reg_set_table' is fairly
415 sparse with respect to pseudo regs but for hard regs could be fairly dense
416 [relatively speaking]. And recording sets of pseudo-regs in lists speeds
417 up functions like compute_transp since in the case of pseudo-regs we only
418 need to iterate over the number of times a pseudo-reg is set, not over the
419 number of basic blocks [clearly there is a bit of a slow down in the cases
420 where a pseudo is set more than once in a block, however it is believed
421 that the net effect is to speed things up]. This isn't done for hard-regs
422 because recording call-clobbered hard-regs in `reg_set_table' at each
423 function call can consume a fair bit of memory, and iterating over
424 hard-regs stored this way in compute_transp will be more expensive. */
426 typedef struct reg_set
428 /* The next setting of this register. */
429 struct reg_set
*next
;
430 /* The index of the block where it was set. */
434 static reg_set
**reg_set_table
;
436 /* Size of `reg_set_table'.
437 The table starts out at max_gcse_regno + slop, and is enlarged as
439 static int reg_set_table_size
;
441 /* Amount to grow `reg_set_table' by when it's full. */
442 #define REG_SET_TABLE_SLOP 100
444 /* This is a list of expressions which are MEMs and will be used by load
446 Load motion tracks MEMs which aren't killed by
447 anything except itself. (i.e., loads and stores to a single location).
448 We can then allow movement of these MEM refs with a little special
449 allowance. (all stores copy the same value to the reaching reg used
450 for the loads). This means all values used to store into memory must have
451 no side effects so we can re-issue the setter value.
452 Store Motion uses this structure as an expression table to track stores
453 which look interesting, and might be moveable towards the exit block. */
457 struct expr
* expr
; /* Gcse expression reference for LM. */
458 rtx pattern
; /* Pattern of this mem. */
459 rtx pattern_regs
; /* List of registers mentioned by the mem. */
460 rtx loads
; /* INSN list of loads seen. */
461 rtx stores
; /* INSN list of stores seen. */
462 struct ls_expr
* next
; /* Next in the list. */
463 int invalid
; /* Invalid for some reason. */
464 int index
; /* If it maps to a bitmap index. */
465 unsigned int hash_index
; /* Index when in a hash table. */
466 rtx reaching_reg
; /* Register to use when re-writing. */
469 /* Array of implicit set patterns indexed by basic block index. */
470 static rtx
*implicit_sets
;
472 /* Head of the list of load/store memory refs. */
473 static struct ls_expr
* pre_ldst_mems
= NULL
;
475 /* Hashtable for the load/store memory refs. */
476 static htab_t pre_ldst_table
= NULL
;
478 /* Bitmap containing one bit for each register in the program.
479 Used when performing GCSE to track which registers have been set since
480 the start of the basic block. */
481 static regset reg_set_bitmap
;
483 /* For each block, a bitmap of registers set in the block.
484 This is used by compute_transp.
485 It is computed during hash table computation and not by compute_sets
486 as it includes registers added since the last pass (or between cprop and
487 gcse) and it's currently not easy to realloc sbitmap vectors. */
488 static sbitmap
*reg_set_in_block
;
490 /* Array, indexed by basic block number for a list of insns which modify
491 memory within that block. */
492 static rtx
* modify_mem_list
;
493 static bitmap modify_mem_list_set
;
495 /* This array parallels modify_mem_list, but is kept canonicalized. */
496 static rtx
* canon_modify_mem_list
;
498 /* Bitmap indexed by block numbers to record which blocks contain
500 static bitmap blocks_with_calls
;
502 /* Various variables for statistics gathering. */
504 /* Memory used in a pass.
505 This isn't intended to be absolutely precise. Its intent is only
506 to keep an eye on memory usage. */
507 static int bytes_used
;
509 /* GCSE substitutions made. */
510 static int gcse_subst_count
;
511 /* Number of copy instructions created. */
512 static int gcse_create_count
;
513 /* Number of local constants propagated. */
514 static int local_const_prop_count
;
515 /* Number of local copies propagated. */
516 static int local_copy_prop_count
;
517 /* Number of global constants propagated. */
518 static int global_const_prop_count
;
519 /* Number of global copies propagated. */
520 static int global_copy_prop_count
;
522 /* For available exprs */
523 static sbitmap
*ae_kill
, *ae_gen
;
525 static void compute_can_copy (void);
526 static void *gmalloc (size_t) ATTRIBUTE_MALLOC
;
527 static void *gcalloc (size_t, size_t) ATTRIBUTE_MALLOC
;
528 static void *grealloc (void *, size_t);
529 static void *gcse_alloc (unsigned long);
530 static void alloc_gcse_mem (void);
531 static void free_gcse_mem (void);
532 static void alloc_reg_set_mem (int);
533 static void free_reg_set_mem (void);
534 static void record_one_set (int, rtx
);
535 static void record_set_info (rtx
, rtx
, void *);
536 static void compute_sets (void);
537 static void hash_scan_insn (rtx
, struct hash_table
*, int);
538 static void hash_scan_set (rtx
, rtx
, struct hash_table
*);
539 static void hash_scan_clobber (rtx
, rtx
, struct hash_table
*);
540 static void hash_scan_call (rtx
, rtx
, struct hash_table
*);
541 static int want_to_gcse_p (rtx
);
542 static bool can_assign_to_reg_p (rtx
);
543 static bool gcse_constant_p (rtx
);
544 static int oprs_unchanged_p (rtx
, rtx
, int);
545 static int oprs_anticipatable_p (rtx
, rtx
);
546 static int oprs_available_p (rtx
, rtx
);
547 static void insert_expr_in_table (rtx
, enum machine_mode
, rtx
, int, int,
548 struct hash_table
*);
549 static void insert_set_in_table (rtx
, rtx
, struct hash_table
*);
550 static unsigned int hash_expr (rtx
, enum machine_mode
, int *, int);
551 static unsigned int hash_set (int, int);
552 static int expr_equiv_p (rtx
, rtx
);
553 static void record_last_reg_set_info (rtx
, int);
554 static void record_last_mem_set_info (rtx
);
555 static void record_last_set_info (rtx
, rtx
, void *);
556 static void compute_hash_table (struct hash_table
*);
557 static void alloc_hash_table (int, struct hash_table
*, int);
558 static void free_hash_table (struct hash_table
*);
559 static void compute_hash_table_work (struct hash_table
*);
560 static void dump_hash_table (FILE *, const char *, struct hash_table
*);
561 static struct expr
*lookup_set (unsigned int, struct hash_table
*);
562 static struct expr
*next_set (unsigned int, struct expr
*);
563 static void reset_opr_set_tables (void);
564 static int oprs_not_set_p (rtx
, rtx
);
565 static void mark_call (rtx
);
566 static void mark_set (rtx
, rtx
);
567 static void mark_clobber (rtx
, rtx
);
568 static void mark_oprs_set (rtx
);
569 static void alloc_cprop_mem (int, int);
570 static void free_cprop_mem (void);
571 static void compute_transp (rtx
, int, sbitmap
*, int);
572 static void compute_transpout (void);
573 static void compute_local_properties (sbitmap
*, sbitmap
*, sbitmap
*,
574 struct hash_table
*);
575 static void compute_cprop_data (void);
576 static void find_used_regs (rtx
*, void *);
577 static int try_replace_reg (rtx
, rtx
, rtx
);
578 static struct expr
*find_avail_set (int, rtx
);
579 static int cprop_jump (basic_block
, rtx
, rtx
, rtx
, rtx
);
580 static void mems_conflict_for_gcse_p (rtx
, rtx
, void *);
581 static int load_killed_in_block_p (basic_block
, int, rtx
, int);
582 static void canon_list_insert (rtx
, rtx
, void *);
583 static int cprop_insn (rtx
, int);
584 static int cprop (int);
585 static void find_implicit_sets (void);
586 static int one_cprop_pass (int, bool, bool);
587 static bool constprop_register (rtx
, rtx
, rtx
, bool);
588 static struct expr
*find_bypass_set (int, int);
589 static bool reg_killed_on_edge (rtx
, edge
);
590 static int bypass_block (basic_block
, rtx
, rtx
);
591 static int bypass_conditional_jumps (void);
592 static void alloc_pre_mem (int, int);
593 static void free_pre_mem (void);
594 static void compute_pre_data (void);
595 static int pre_expr_reaches_here_p (basic_block
, struct expr
*,
597 static void insert_insn_end_bb (struct expr
*, basic_block
, int);
598 static void pre_insert_copy_insn (struct expr
*, rtx
);
599 static void pre_insert_copies (void);
600 static int pre_delete (void);
601 static int pre_gcse (void);
602 static int one_pre_gcse_pass (int);
603 static void add_label_notes (rtx
, rtx
);
604 static void alloc_code_hoist_mem (int, int);
605 static void free_code_hoist_mem (void);
606 static void compute_code_hoist_vbeinout (void);
607 static void compute_code_hoist_data (void);
608 static int hoist_expr_reaches_here_p (basic_block
, int, basic_block
, char *);
609 static void hoist_code (void);
610 static int one_code_hoisting_pass (void);
611 static rtx
process_insert_insn (struct expr
*);
612 static int pre_edge_insert (struct edge_list
*, struct expr
**);
613 static int pre_expr_reaches_here_p_work (basic_block
, struct expr
*,
614 basic_block
, char *);
615 static struct ls_expr
* ldst_entry (rtx
);
616 static void free_ldst_entry (struct ls_expr
*);
617 static void free_ldst_mems (void);
618 static void print_ldst_list (FILE *);
619 static struct ls_expr
* find_rtx_in_ldst (rtx
);
620 static int enumerate_ldsts (void);
621 static inline struct ls_expr
* first_ls_expr (void);
622 static inline struct ls_expr
* next_ls_expr (struct ls_expr
*);
623 static int simple_mem (rtx
);
624 static void invalidate_any_buried_refs (rtx
);
625 static void compute_ld_motion_mems (void);
626 static void trim_ld_motion_mems (void);
627 static void update_ld_motion_stores (struct expr
*);
628 static void reg_set_info (rtx
, rtx
, void *);
629 static void reg_clear_last_set (rtx
, rtx
, void *);
630 static bool store_ops_ok (rtx
, int *);
631 static rtx
extract_mentioned_regs (rtx
);
632 static rtx
extract_mentioned_regs_helper (rtx
, rtx
);
633 static void find_moveable_store (rtx
, int *, int *);
634 static int compute_store_table (void);
635 static bool load_kills_store (rtx
, rtx
, int);
636 static bool find_loads (rtx
, rtx
, int);
637 static bool store_killed_in_insn (rtx
, rtx
, rtx
, int);
638 static bool store_killed_after (rtx
, rtx
, rtx
, basic_block
, int *, rtx
*);
639 static bool store_killed_before (rtx
, rtx
, rtx
, basic_block
, int *);
640 static void build_store_vectors (void);
641 static void insert_insn_start_bb (rtx
, basic_block
);
642 static int insert_store (struct ls_expr
*, edge
);
643 static void remove_reachable_equiv_notes (basic_block
, struct ls_expr
*);
644 static void replace_store_insn (rtx
, rtx
, basic_block
, struct ls_expr
*);
645 static void delete_store (struct ls_expr
*, basic_block
);
646 static void free_store_memory (void);
647 static void store_motion (void);
648 static void free_insn_expr_list_list (rtx
*);
649 static void clear_modify_mem_tables (void);
650 static void free_modify_mem_tables (void);
651 static rtx
gcse_emit_move_after (rtx
, rtx
, rtx
);
652 static void local_cprop_find_used_regs (rtx
*, void *);
653 static bool do_local_cprop (rtx
, rtx
, bool, rtx
*);
654 static bool adjust_libcall_notes (rtx
, rtx
, rtx
, rtx
*);
655 static void local_cprop_pass (bool);
656 static bool is_too_expensive (const char *);
659 /* Entry point for global common subexpression elimination.
660 F is the first instruction in the function. Return nonzero if a
664 gcse_main (rtx f ATTRIBUTE_UNUSED
, FILE *file
)
667 /* Bytes used at start of pass. */
668 int initial_bytes_used
;
669 /* Maximum number of bytes used by a pass. */
671 /* Point to release obstack data from for each pass. */
672 char *gcse_obstack_bottom
;
674 /* We do not construct an accurate cfg in functions which call
675 setjmp, so just punt to be safe. */
676 if (current_function_calls_setjmp
)
679 /* Assume that we do not need to run jump optimizations after gcse. */
680 run_jump_opt_after_gcse
= 0;
682 /* For calling dump_foo fns from gdb. */
683 debug_stderr
= stderr
;
686 /* Identify the basic block information for this function, including
687 successors and predecessors. */
688 max_gcse_regno
= max_reg_num ();
691 dump_flow_info (file
);
693 /* Return if there's nothing to do, or it is too expensive. */
694 if (n_basic_blocks
<= NUM_FIXED_BLOCKS
+ 1
695 || is_too_expensive (_("GCSE disabled")))
698 gcc_obstack_init (&gcse_obstack
);
702 init_alias_analysis ();
703 /* Record where pseudo-registers are set. This data is kept accurate
704 during each pass. ??? We could also record hard-reg information here
705 [since it's unchanging], however it is currently done during hash table
708 It may be tempting to compute MEM set information here too, but MEM sets
709 will be subject to code motion one day and thus we need to compute
710 information about memory sets when we build the hash tables. */
712 alloc_reg_set_mem (max_gcse_regno
);
716 initial_bytes_used
= bytes_used
;
718 gcse_obstack_bottom
= gcse_alloc (1);
720 while (changed
&& pass
< MAX_GCSE_PASSES
)
724 fprintf (file
, "GCSE pass %d\n\n", pass
+ 1);
726 /* Initialize bytes_used to the space for the pred/succ lists,
727 and the reg_set_table data. */
728 bytes_used
= initial_bytes_used
;
730 /* Each pass may create new registers, so recalculate each time. */
731 max_gcse_regno
= max_reg_num ();
735 /* Don't allow constant propagation to modify jumps
737 timevar_push (TV_CPROP1
);
738 changed
= one_cprop_pass (pass
+ 1, false, false);
739 timevar_pop (TV_CPROP1
);
745 timevar_push (TV_PRE
);
746 changed
|= one_pre_gcse_pass (pass
+ 1);
747 /* We may have just created new basic blocks. Release and
748 recompute various things which are sized on the number of
752 free_modify_mem_tables ();
753 modify_mem_list
= gcalloc (last_basic_block
, sizeof (rtx
));
754 canon_modify_mem_list
= gcalloc (last_basic_block
, sizeof (rtx
));
757 alloc_reg_set_mem (max_reg_num ());
759 run_jump_opt_after_gcse
= 1;
760 timevar_pop (TV_PRE
);
763 if (max_pass_bytes
< bytes_used
)
764 max_pass_bytes
= bytes_used
;
766 /* Free up memory, then reallocate for code hoisting. We can
767 not re-use the existing allocated memory because the tables
768 will not have info for the insns or registers created by
769 partial redundancy elimination. */
772 /* It does not make sense to run code hoisting unless we are optimizing
773 for code size -- it rarely makes programs faster, and can make
774 them bigger if we did partial redundancy elimination (when optimizing
775 for space, we don't run the partial redundancy algorithms). */
778 timevar_push (TV_HOIST
);
779 max_gcse_regno
= max_reg_num ();
781 changed
|= one_code_hoisting_pass ();
784 if (max_pass_bytes
< bytes_used
)
785 max_pass_bytes
= bytes_used
;
786 timevar_pop (TV_HOIST
);
791 fprintf (file
, "\n");
795 obstack_free (&gcse_obstack
, gcse_obstack_bottom
);
799 /* Do one last pass of copy propagation, including cprop into
800 conditional jumps. */
802 max_gcse_regno
= max_reg_num ();
804 /* This time, go ahead and allow cprop to alter jumps. */
805 timevar_push (TV_CPROP2
);
806 one_cprop_pass (pass
+ 1, true, false);
807 timevar_pop (TV_CPROP2
);
812 fprintf (file
, "GCSE of %s: %d basic blocks, ",
813 current_function_name (), n_basic_blocks
);
814 fprintf (file
, "%d pass%s, %d bytes\n\n",
815 pass
, pass
> 1 ? "es" : "", max_pass_bytes
);
818 obstack_free (&gcse_obstack
, NULL
);
821 /* We are finished with alias. */
822 end_alias_analysis ();
823 allocate_reg_info (max_reg_num (), FALSE
, FALSE
);
825 if (!optimize_size
&& flag_gcse_sm
)
827 timevar_push (TV_LSM
);
829 timevar_pop (TV_LSM
);
832 /* Record where pseudo-registers are set. */
833 return run_jump_opt_after_gcse
;
836 /* Misc. utilities. */
838 /* Nonzero for each mode that supports (set (reg) (reg)).
839 This is trivially true for integer and floating point values.
840 It may or may not be true for condition codes. */
841 static char can_copy
[(int) NUM_MACHINE_MODES
];
843 /* Compute which modes support reg/reg copy operations. */
846 compute_can_copy (void)
849 #ifndef AVOID_CCMODE_COPIES
852 memset (can_copy
, 0, NUM_MACHINE_MODES
);
855 for (i
= 0; i
< NUM_MACHINE_MODES
; i
++)
856 if (GET_MODE_CLASS (i
) == MODE_CC
)
858 #ifdef AVOID_CCMODE_COPIES
861 reg
= gen_rtx_REG ((enum machine_mode
) i
, LAST_VIRTUAL_REGISTER
+ 1);
862 insn
= emit_insn (gen_rtx_SET (VOIDmode
, reg
, reg
));
863 if (recog (PATTERN (insn
), insn
, NULL
) >= 0)
873 /* Returns whether the mode supports reg/reg copy operations. */
876 can_copy_p (enum machine_mode mode
)
878 static bool can_copy_init_p
= false;
880 if (! can_copy_init_p
)
883 can_copy_init_p
= true;
886 return can_copy
[mode
] != 0;
889 /* Cover function to xmalloc to record bytes allocated. */
892 gmalloc (size_t size
)
895 return xmalloc (size
);
898 /* Cover function to xcalloc to record bytes allocated. */
901 gcalloc (size_t nelem
, size_t elsize
)
903 bytes_used
+= nelem
* elsize
;
904 return xcalloc (nelem
, elsize
);
907 /* Cover function to xrealloc.
908 We don't record the additional size since we don't know it.
909 It won't affect memory usage stats much anyway. */
912 grealloc (void *ptr
, size_t size
)
914 return xrealloc (ptr
, size
);
917 /* Cover function to obstack_alloc. */
920 gcse_alloc (unsigned long size
)
923 return obstack_alloc (&gcse_obstack
, size
);
926 /* Allocate memory for the cuid mapping array,
927 and reg/memory set tracking tables.
929 This is called at the start of each pass. */
932 alloc_gcse_mem (void)
938 /* Find the largest UID and create a mapping from UIDs to CUIDs.
939 CUIDs are like UIDs except they increase monotonically, have no gaps,
940 and only apply to real insns.
941 (Actually, there are gaps, for insn that are not inside a basic block.
942 but we should never see those anyway, so this is OK.) */
944 max_uid
= get_max_uid ();
945 uid_cuid
= gcalloc (max_uid
+ 1, sizeof (int));
948 FOR_BB_INSNS (bb
, insn
)
951 uid_cuid
[INSN_UID (insn
)] = i
++;
953 uid_cuid
[INSN_UID (insn
)] = i
;
956 /* Create a table mapping cuids to insns. */
959 cuid_insn
= gcalloc (max_cuid
+ 1, sizeof (rtx
));
962 FOR_BB_INSNS (bb
, insn
)
964 CUID_INSN (i
++) = insn
;
966 /* Allocate vars to track sets of regs. */
967 reg_set_bitmap
= BITMAP_ALLOC (NULL
);
969 /* Allocate vars to track sets of regs, memory per block. */
970 reg_set_in_block
= sbitmap_vector_alloc (last_basic_block
, max_gcse_regno
);
971 /* Allocate array to keep a list of insns which modify memory in each
973 modify_mem_list
= gcalloc (last_basic_block
, sizeof (rtx
));
974 canon_modify_mem_list
= gcalloc (last_basic_block
, sizeof (rtx
));
975 modify_mem_list_set
= BITMAP_ALLOC (NULL
);
976 blocks_with_calls
= BITMAP_ALLOC (NULL
);
979 /* Free memory allocated by alloc_gcse_mem. */
987 BITMAP_FREE (reg_set_bitmap
);
989 sbitmap_vector_free (reg_set_in_block
);
990 free_modify_mem_tables ();
991 BITMAP_FREE (modify_mem_list_set
);
992 BITMAP_FREE (blocks_with_calls
);
995 /* Compute the local properties of each recorded expression.
997 Local properties are those that are defined by the block, irrespective of
1000 An expression is transparent in a block if its operands are not modified
1003 An expression is computed (locally available) in a block if it is computed
1004 at least once and expression would contain the same value if the
1005 computation was moved to the end of the block.
1007 An expression is locally anticipatable in a block if it is computed at
1008 least once and expression would contain the same value if the computation
1009 was moved to the beginning of the block.
1011 We call this routine for cprop, pre and code hoisting. They all compute
1012 basically the same information and thus can easily share this code.
1014 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
1015 properties. If NULL, then it is not necessary to compute or record that
1016 particular property.
1018 TABLE controls which hash table to look at. If it is set hash table,
1019 additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
1023 compute_local_properties (sbitmap
*transp
, sbitmap
*comp
, sbitmap
*antloc
,
1024 struct hash_table
*table
)
1028 /* Initialize any bitmaps that were passed in. */
1032 sbitmap_vector_zero (transp
, last_basic_block
);
1034 sbitmap_vector_ones (transp
, last_basic_block
);
1038 sbitmap_vector_zero (comp
, last_basic_block
);
1040 sbitmap_vector_zero (antloc
, last_basic_block
);
1042 for (i
= 0; i
< table
->size
; i
++)
1046 for (expr
= table
->table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
1048 int indx
= expr
->bitmap_index
;
1051 /* The expression is transparent in this block if it is not killed.
1052 We start by assuming all are transparent [none are killed], and
1053 then reset the bits for those that are. */
1055 compute_transp (expr
->expr
, indx
, transp
, table
->set_p
);
1057 /* The occurrences recorded in antic_occr are exactly those that
1058 we want to set to nonzero in ANTLOC. */
1060 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
1062 SET_BIT (antloc
[BLOCK_NUM (occr
->insn
)], indx
);
1064 /* While we're scanning the table, this is a good place to
1066 occr
->deleted_p
= 0;
1069 /* The occurrences recorded in avail_occr are exactly those that
1070 we want to set to nonzero in COMP. */
1072 for (occr
= expr
->avail_occr
; occr
!= NULL
; occr
= occr
->next
)
1074 SET_BIT (comp
[BLOCK_NUM (occr
->insn
)], indx
);
1076 /* While we're scanning the table, this is a good place to
1081 /* While we're scanning the table, this is a good place to
1083 expr
->reaching_reg
= 0;
1088 /* Register set information.
1090 `reg_set_table' records where each register is set or otherwise
1093 static struct obstack reg_set_obstack
;
1096 alloc_reg_set_mem (int n_regs
)
1098 reg_set_table_size
= n_regs
+ REG_SET_TABLE_SLOP
;
1099 reg_set_table
= gcalloc (reg_set_table_size
, sizeof (struct reg_set
*));
1101 gcc_obstack_init (®_set_obstack
);
1105 free_reg_set_mem (void)
1107 free (reg_set_table
);
1108 obstack_free (®_set_obstack
, NULL
);
1111 /* Record REGNO in the reg_set table. */
1114 record_one_set (int regno
, rtx insn
)
1116 /* Allocate a new reg_set element and link it onto the list. */
1117 struct reg_set
*new_reg_info
;
1119 /* If the table isn't big enough, enlarge it. */
1120 if (regno
>= reg_set_table_size
)
1122 int new_size
= regno
+ REG_SET_TABLE_SLOP
;
1124 reg_set_table
= grealloc (reg_set_table
,
1125 new_size
* sizeof (struct reg_set
*));
1126 memset (reg_set_table
+ reg_set_table_size
, 0,
1127 (new_size
- reg_set_table_size
) * sizeof (struct reg_set
*));
1128 reg_set_table_size
= new_size
;
1131 new_reg_info
= obstack_alloc (®_set_obstack
, sizeof (struct reg_set
));
1132 bytes_used
+= sizeof (struct reg_set
);
1133 new_reg_info
->bb_index
= BLOCK_NUM (insn
);
1134 new_reg_info
->next
= reg_set_table
[regno
];
1135 reg_set_table
[regno
] = new_reg_info
;
1138 /* Called from compute_sets via note_stores to handle one SET or CLOBBER in
1139 an insn. The DATA is really the instruction in which the SET is
1143 record_set_info (rtx dest
, rtx setter ATTRIBUTE_UNUSED
, void *data
)
1145 rtx record_set_insn
= (rtx
) data
;
1147 if (REG_P (dest
) && REGNO (dest
) >= FIRST_PSEUDO_REGISTER
)
1148 record_one_set (REGNO (dest
), record_set_insn
);
1151 /* Scan the function and record each set of each pseudo-register.
1153 This is called once, at the start of the gcse pass. See the comments for
1154 `reg_set_table' for further documentation. */
1163 FOR_BB_INSNS (bb
, insn
)
1165 note_stores (PATTERN (insn
), record_set_info
, insn
);
1168 /* Hash table support. */
1170 struct reg_avail_info
1172 basic_block last_bb
;
1177 static struct reg_avail_info
*reg_avail_info
;
1178 static basic_block current_bb
;
1181 /* See whether X, the source of a set, is something we want to consider for
1185 want_to_gcse_p (rtx x
)
1187 switch (GET_CODE (x
))
1198 return can_assign_to_reg_p (x
);
1202 /* Used internally by can_assign_to_reg_p. */
1204 static GTY(()) rtx test_insn
;
1206 /* Return true if we can assign X to a pseudo register. */
1209 can_assign_to_reg_p (rtx x
)
1211 int num_clobbers
= 0;
1214 /* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
1215 if (general_operand (x
, GET_MODE (x
)))
1217 else if (GET_MODE (x
) == VOIDmode
)
1220 /* Otherwise, check if we can make a valid insn from it. First initialize
1221 our test insn if we haven't already. */
1225 = make_insn_raw (gen_rtx_SET (VOIDmode
,
1226 gen_rtx_REG (word_mode
,
1227 FIRST_PSEUDO_REGISTER
* 2),
1229 NEXT_INSN (test_insn
) = PREV_INSN (test_insn
) = 0;
1232 /* Now make an insn like the one we would make when GCSE'ing and see if
1234 PUT_MODE (SET_DEST (PATTERN (test_insn
)), GET_MODE (x
));
1235 SET_SRC (PATTERN (test_insn
)) = x
;
1236 return ((icode
= recog (PATTERN (test_insn
), test_insn
, &num_clobbers
)) >= 0
1237 && (num_clobbers
== 0 || ! added_clobbers_hard_reg_p (icode
)));
1240 /* Return nonzero if the operands of expression X are unchanged from the
1241 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1242 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1245 oprs_unchanged_p (rtx x
, rtx insn
, int avail_p
)
1254 code
= GET_CODE (x
);
1259 struct reg_avail_info
*info
= ®_avail_info
[REGNO (x
)];
1261 if (info
->last_bb
!= current_bb
)
1264 return info
->last_set
< INSN_CUID (insn
);
1266 return info
->first_set
>= INSN_CUID (insn
);
1270 if (load_killed_in_block_p (current_bb
, INSN_CUID (insn
),
1274 return oprs_unchanged_p (XEXP (x
, 0), insn
, avail_p
);
1300 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
1304 /* If we are about to do the last recursive call needed at this
1305 level, change it into iteration. This function is called enough
1308 return oprs_unchanged_p (XEXP (x
, i
), insn
, avail_p
);
1310 else if (! oprs_unchanged_p (XEXP (x
, i
), insn
, avail_p
))
1313 else if (fmt
[i
] == 'E')
1314 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1315 if (! oprs_unchanged_p (XVECEXP (x
, i
, j
), insn
, avail_p
))
1322 /* Used for communication between mems_conflict_for_gcse_p and
1323 load_killed_in_block_p. Nonzero if mems_conflict_for_gcse_p finds a
1324 conflict between two memory references. */
1325 static int gcse_mems_conflict_p
;
1327 /* Used for communication between mems_conflict_for_gcse_p and
1328 load_killed_in_block_p. A memory reference for a load instruction,
1329 mems_conflict_for_gcse_p will see if a memory store conflicts with
1330 this memory load. */
1331 static rtx gcse_mem_operand
;
1333 /* DEST is the output of an instruction. If it is a memory reference, and
1334 possibly conflicts with the load found in gcse_mem_operand, then set
1335 gcse_mems_conflict_p to a nonzero value. */
1338 mems_conflict_for_gcse_p (rtx dest
, rtx setter ATTRIBUTE_UNUSED
,
1339 void *data ATTRIBUTE_UNUSED
)
1341 while (GET_CODE (dest
) == SUBREG
1342 || GET_CODE (dest
) == ZERO_EXTRACT
1343 || GET_CODE (dest
) == STRICT_LOW_PART
)
1344 dest
= XEXP (dest
, 0);
1346 /* If DEST is not a MEM, then it will not conflict with the load. Note
1347 that function calls are assumed to clobber memory, but are handled
1352 /* If we are setting a MEM in our list of specially recognized MEMs,
1353 don't mark as killed this time. */
1355 if (expr_equiv_p (dest
, gcse_mem_operand
) && pre_ldst_mems
!= NULL
)
1357 if (!find_rtx_in_ldst (dest
))
1358 gcse_mems_conflict_p
= 1;
1362 if (true_dependence (dest
, GET_MODE (dest
), gcse_mem_operand
,
1364 gcse_mems_conflict_p
= 1;
1367 /* Return nonzero if the expression in X (a memory reference) is killed
1368 in block BB before or after the insn with the CUID in UID_LIMIT.
1369 AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills
1372 To check the entire block, set UID_LIMIT to max_uid + 1 and
1376 load_killed_in_block_p (basic_block bb
, int uid_limit
, rtx x
, int avail_p
)
1378 rtx list_entry
= modify_mem_list
[bb
->index
];
1380 /* If this is a readonly then we aren't going to be changing it. */
1381 if (MEM_READONLY_P (x
))
1387 /* Ignore entries in the list that do not apply. */
1389 && INSN_CUID (XEXP (list_entry
, 0)) < uid_limit
)
1391 && INSN_CUID (XEXP (list_entry
, 0)) > uid_limit
))
1393 list_entry
= XEXP (list_entry
, 1);
1397 setter
= XEXP (list_entry
, 0);
1399 /* If SETTER is a call everything is clobbered. Note that calls
1400 to pure functions are never put on the list, so we need not
1401 worry about them. */
1402 if (CALL_P (setter
))
1405 /* SETTER must be an INSN of some kind that sets memory. Call
1406 note_stores to examine each hunk of memory that is modified.
1408 The note_stores interface is pretty limited, so we have to
1409 communicate via global variables. Yuk. */
1410 gcse_mem_operand
= x
;
1411 gcse_mems_conflict_p
= 0;
1412 note_stores (PATTERN (setter
), mems_conflict_for_gcse_p
, NULL
);
1413 if (gcse_mems_conflict_p
)
1415 list_entry
= XEXP (list_entry
, 1);
1420 /* Return nonzero if the operands of expression X are unchanged from
1421 the start of INSN's basic block up to but not including INSN. */
1424 oprs_anticipatable_p (rtx x
, rtx insn
)
1426 return oprs_unchanged_p (x
, insn
, 0);
1429 /* Return nonzero if the operands of expression X are unchanged from
1430 INSN to the end of INSN's basic block. */
1433 oprs_available_p (rtx x
, rtx insn
)
1435 return oprs_unchanged_p (x
, insn
, 1);
1438 /* Hash expression X.
1440 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1441 indicating if a volatile operand is found or if the expression contains
1442 something we don't want to insert in the table. HASH_TABLE_SIZE is
1443 the current size of the hash table to be probed. */
1446 hash_expr (rtx x
, enum machine_mode mode
, int *do_not_record_p
,
1447 int hash_table_size
)
1451 *do_not_record_p
= 0;
1453 hash
= hash_rtx (x
, mode
, do_not_record_p
,
1454 NULL
, /*have_reg_qty=*/false);
1455 return hash
% hash_table_size
;
1458 /* Hash a set of register REGNO.
1460 Sets are hashed on the register that is set. This simplifies the PRE copy
1463 ??? May need to make things more elaborate. Later, as necessary. */
1466 hash_set (int regno
, int hash_table_size
)
1471 return hash
% hash_table_size
;
1474 /* Return nonzero if exp1 is equivalent to exp2. */
1477 expr_equiv_p (rtx x
, rtx y
)
1479 return exp_equiv_p (x
, y
, 0, true);
1482 /* Insert expression X in INSN in the hash TABLE.
1483 If it is already present, record it as the last occurrence in INSN's
1486 MODE is the mode of the value X is being stored into.
1487 It is only used if X is a CONST_INT.
1489 ANTIC_P is nonzero if X is an anticipatable expression.
1490 AVAIL_P is nonzero if X is an available expression. */
1493 insert_expr_in_table (rtx x
, enum machine_mode mode
, rtx insn
, int antic_p
,
1494 int avail_p
, struct hash_table
*table
)
1496 int found
, do_not_record_p
;
1498 struct expr
*cur_expr
, *last_expr
= NULL
;
1499 struct occr
*antic_occr
, *avail_occr
;
1501 hash
= hash_expr (x
, mode
, &do_not_record_p
, table
->size
);
1503 /* Do not insert expression in table if it contains volatile operands,
1504 or if hash_expr determines the expression is something we don't want
1505 to or can't handle. */
1506 if (do_not_record_p
)
1509 cur_expr
= table
->table
[hash
];
1512 while (cur_expr
&& 0 == (found
= expr_equiv_p (cur_expr
->expr
, x
)))
1514 /* If the expression isn't found, save a pointer to the end of
1516 last_expr
= cur_expr
;
1517 cur_expr
= cur_expr
->next_same_hash
;
1522 cur_expr
= gcse_alloc (sizeof (struct expr
));
1523 bytes_used
+= sizeof (struct expr
);
1524 if (table
->table
[hash
] == NULL
)
1525 /* This is the first pattern that hashed to this index. */
1526 table
->table
[hash
] = cur_expr
;
1528 /* Add EXPR to end of this hash chain. */
1529 last_expr
->next_same_hash
= cur_expr
;
1531 /* Set the fields of the expr element. */
1533 cur_expr
->bitmap_index
= table
->n_elems
++;
1534 cur_expr
->next_same_hash
= NULL
;
1535 cur_expr
->antic_occr
= NULL
;
1536 cur_expr
->avail_occr
= NULL
;
1539 /* Now record the occurrence(s). */
1542 antic_occr
= cur_expr
->antic_occr
;
1544 if (antic_occr
&& BLOCK_NUM (antic_occr
->insn
) != BLOCK_NUM (insn
))
1548 /* Found another instance of the expression in the same basic block.
1549 Prefer the currently recorded one. We want the first one in the
1550 block and the block is scanned from start to end. */
1551 ; /* nothing to do */
1554 /* First occurrence of this expression in this basic block. */
1555 antic_occr
= gcse_alloc (sizeof (struct occr
));
1556 bytes_used
+= sizeof (struct occr
);
1557 antic_occr
->insn
= insn
;
1558 antic_occr
->next
= cur_expr
->antic_occr
;
1559 antic_occr
->deleted_p
= 0;
1560 cur_expr
->antic_occr
= antic_occr
;
1566 avail_occr
= cur_expr
->avail_occr
;
1568 if (avail_occr
&& BLOCK_NUM (avail_occr
->insn
) == BLOCK_NUM (insn
))
1570 /* Found another instance of the expression in the same basic block.
1571 Prefer this occurrence to the currently recorded one. We want
1572 the last one in the block and the block is scanned from start
1574 avail_occr
->insn
= insn
;
1578 /* First occurrence of this expression in this basic block. */
1579 avail_occr
= gcse_alloc (sizeof (struct occr
));
1580 bytes_used
+= sizeof (struct occr
);
1581 avail_occr
->insn
= insn
;
1582 avail_occr
->next
= cur_expr
->avail_occr
;
1583 avail_occr
->deleted_p
= 0;
1584 cur_expr
->avail_occr
= avail_occr
;
1589 /* Insert pattern X in INSN in the hash table.
1590 X is a SET of a reg to either another reg or a constant.
1591 If it is already present, record it as the last occurrence in INSN's
1595 insert_set_in_table (rtx x
, rtx insn
, struct hash_table
*table
)
1599 struct expr
*cur_expr
, *last_expr
= NULL
;
1600 struct occr
*cur_occr
;
1602 gcc_assert (GET_CODE (x
) == SET
&& REG_P (SET_DEST (x
)));
1604 hash
= hash_set (REGNO (SET_DEST (x
)), table
->size
);
1606 cur_expr
= table
->table
[hash
];
1609 while (cur_expr
&& 0 == (found
= expr_equiv_p (cur_expr
->expr
, x
)))
1611 /* If the expression isn't found, save a pointer to the end of
1613 last_expr
= cur_expr
;
1614 cur_expr
= cur_expr
->next_same_hash
;
1619 cur_expr
= gcse_alloc (sizeof (struct expr
));
1620 bytes_used
+= sizeof (struct expr
);
1621 if (table
->table
[hash
] == NULL
)
1622 /* This is the first pattern that hashed to this index. */
1623 table
->table
[hash
] = cur_expr
;
1625 /* Add EXPR to end of this hash chain. */
1626 last_expr
->next_same_hash
= cur_expr
;
1628 /* Set the fields of the expr element.
1629 We must copy X because it can be modified when copy propagation is
1630 performed on its operands. */
1631 cur_expr
->expr
= copy_rtx (x
);
1632 cur_expr
->bitmap_index
= table
->n_elems
++;
1633 cur_expr
->next_same_hash
= NULL
;
1634 cur_expr
->antic_occr
= NULL
;
1635 cur_expr
->avail_occr
= NULL
;
1638 /* Now record the occurrence. */
1639 cur_occr
= cur_expr
->avail_occr
;
1641 if (cur_occr
&& BLOCK_NUM (cur_occr
->insn
) == BLOCK_NUM (insn
))
1643 /* Found another instance of the expression in the same basic block.
1644 Prefer this occurrence to the currently recorded one. We want
1645 the last one in the block and the block is scanned from start
1647 cur_occr
->insn
= insn
;
1651 /* First occurrence of this expression in this basic block. */
1652 cur_occr
= gcse_alloc (sizeof (struct occr
));
1653 bytes_used
+= sizeof (struct occr
);
1655 cur_occr
->insn
= insn
;
1656 cur_occr
->next
= cur_expr
->avail_occr
;
1657 cur_occr
->deleted_p
= 0;
1658 cur_expr
->avail_occr
= cur_occr
;
1662 /* Determine whether the rtx X should be treated as a constant for
1663 the purposes of GCSE's constant propagation. */
1666 gcse_constant_p (rtx x
)
1668 /* Consider a COMPARE of two integers constant. */
1669 if (GET_CODE (x
) == COMPARE
1670 && GET_CODE (XEXP (x
, 0)) == CONST_INT
1671 && GET_CODE (XEXP (x
, 1)) == CONST_INT
)
1674 /* Consider a COMPARE of the same registers is a constant
1675 if they are not floating point registers. */
1676 if (GET_CODE(x
) == COMPARE
1677 && REG_P (XEXP (x
, 0)) && REG_P (XEXP (x
, 1))
1678 && REGNO (XEXP (x
, 0)) == REGNO (XEXP (x
, 1))
1679 && ! FLOAT_MODE_P (GET_MODE (XEXP (x
, 0)))
1680 && ! FLOAT_MODE_P (GET_MODE (XEXP (x
, 1))))
1683 return CONSTANT_P (x
);
1686 /* Scan pattern PAT of INSN and add an entry to the hash TABLE (set or
1690 hash_scan_set (rtx pat
, rtx insn
, struct hash_table
*table
)
1692 rtx src
= SET_SRC (pat
);
1693 rtx dest
= SET_DEST (pat
);
1696 if (GET_CODE (src
) == CALL
)
1697 hash_scan_call (src
, insn
, table
);
1699 else if (REG_P (dest
))
1701 unsigned int regno
= REGNO (dest
);
1704 /* See if a REG_NOTE shows this equivalent to a simpler expression.
1705 This allows us to do a single GCSE pass and still eliminate
1706 redundant constants, addresses or other expressions that are
1707 constructed with multiple instructions. */
1708 note
= find_reg_equal_equiv_note (insn
);
1711 ? gcse_constant_p (XEXP (note
, 0))
1712 : want_to_gcse_p (XEXP (note
, 0))))
1713 src
= XEXP (note
, 0), pat
= gen_rtx_SET (VOIDmode
, dest
, src
);
1715 /* Only record sets of pseudo-regs in the hash table. */
1717 && regno
>= FIRST_PSEUDO_REGISTER
1718 /* Don't GCSE something if we can't do a reg/reg copy. */
1719 && can_copy_p (GET_MODE (dest
))
1720 /* GCSE commonly inserts instruction after the insn. We can't
1721 do that easily for EH_REGION notes so disable GCSE on these
1723 && !find_reg_note (insn
, REG_EH_REGION
, NULL_RTX
)
1724 /* Is SET_SRC something we want to gcse? */
1725 && want_to_gcse_p (src
)
1726 /* Don't CSE a nop. */
1727 && ! set_noop_p (pat
)
1728 /* Don't GCSE if it has attached REG_EQUIV note.
1729 At this point this only function parameters should have
1730 REG_EQUIV notes and if the argument slot is used somewhere
1731 explicitly, it means address of parameter has been taken,
1732 so we should not extend the lifetime of the pseudo. */
1733 && (note
== NULL_RTX
|| ! MEM_P (XEXP (note
, 0))))
1735 /* An expression is not anticipatable if its operands are
1736 modified before this insn or if this is not the only SET in
1738 int antic_p
= oprs_anticipatable_p (src
, insn
) && single_set (insn
);
1739 /* An expression is not available if its operands are
1740 subsequently modified, including this insn. It's also not
1741 available if this is a branch, because we can't insert
1742 a set after the branch. */
1743 int avail_p
= (oprs_available_p (src
, insn
)
1744 && ! JUMP_P (insn
));
1746 insert_expr_in_table (src
, GET_MODE (dest
), insn
, antic_p
, avail_p
, table
);
1749 /* Record sets for constant/copy propagation. */
1750 else if (table
->set_p
1751 && regno
>= FIRST_PSEUDO_REGISTER
1753 && REGNO (src
) >= FIRST_PSEUDO_REGISTER
1754 && can_copy_p (GET_MODE (dest
))
1755 && REGNO (src
) != regno
)
1756 || gcse_constant_p (src
))
1757 /* A copy is not available if its src or dest is subsequently
1758 modified. Here we want to search from INSN+1 on, but
1759 oprs_available_p searches from INSN on. */
1760 && (insn
== BB_END (BLOCK_FOR_INSN (insn
))
1761 || ((tmp
= next_nonnote_insn (insn
)) != NULL_RTX
1762 && oprs_available_p (pat
, tmp
))))
1763 insert_set_in_table (pat
, insn
, table
);
1765 /* In case of store we want to consider the memory value as available in
1766 the REG stored in that memory. This makes it possible to remove
1767 redundant loads from due to stores to the same location. */
1768 else if (flag_gcse_las
&& REG_P (src
) && MEM_P (dest
))
1770 unsigned int regno
= REGNO (src
);
1772 /* Do not do this for constant/copy propagation. */
1774 /* Only record sets of pseudo-regs in the hash table. */
1775 && regno
>= FIRST_PSEUDO_REGISTER
1776 /* Don't GCSE something if we can't do a reg/reg copy. */
1777 && can_copy_p (GET_MODE (src
))
1778 /* GCSE commonly inserts instruction after the insn. We can't
1779 do that easily for EH_REGION notes so disable GCSE on these
1781 && ! find_reg_note (insn
, REG_EH_REGION
, NULL_RTX
)
1782 /* Is SET_DEST something we want to gcse? */
1783 && want_to_gcse_p (dest
)
1784 /* Don't CSE a nop. */
1785 && ! set_noop_p (pat
)
1786 /* Don't GCSE if it has attached REG_EQUIV note.
1787 At this point this only function parameters should have
1788 REG_EQUIV notes and if the argument slot is used somewhere
1789 explicitly, it means address of parameter has been taken,
1790 so we should not extend the lifetime of the pseudo. */
1791 && ((note
= find_reg_note (insn
, REG_EQUIV
, NULL_RTX
)) == 0
1792 || ! MEM_P (XEXP (note
, 0))))
1794 /* Stores are never anticipatable. */
1796 /* An expression is not available if its operands are
1797 subsequently modified, including this insn. It's also not
1798 available if this is a branch, because we can't insert
1799 a set after the branch. */
1800 int avail_p
= oprs_available_p (dest
, insn
)
1803 /* Record the memory expression (DEST) in the hash table. */
1804 insert_expr_in_table (dest
, GET_MODE (dest
), insn
,
1805 antic_p
, avail_p
, table
);
1811 hash_scan_clobber (rtx x ATTRIBUTE_UNUSED
, rtx insn ATTRIBUTE_UNUSED
,
1812 struct hash_table
*table ATTRIBUTE_UNUSED
)
1814 /* Currently nothing to do. */
1818 hash_scan_call (rtx x ATTRIBUTE_UNUSED
, rtx insn ATTRIBUTE_UNUSED
,
1819 struct hash_table
*table ATTRIBUTE_UNUSED
)
1821 /* Currently nothing to do. */
1824 /* Process INSN and add hash table entries as appropriate.
1826 Only available expressions that set a single pseudo-reg are recorded.
1828 Single sets in a PARALLEL could be handled, but it's an extra complication
1829 that isn't dealt with right now. The trick is handling the CLOBBERs that
1830 are also in the PARALLEL. Later.
1832 If SET_P is nonzero, this is for the assignment hash table,
1833 otherwise it is for the expression hash table.
1834 If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
1835 not record any expressions. */
1838 hash_scan_insn (rtx insn
, struct hash_table
*table
, int in_libcall_block
)
1840 rtx pat
= PATTERN (insn
);
1843 if (in_libcall_block
)
1846 /* Pick out the sets of INSN and for other forms of instructions record
1847 what's been modified. */
1849 if (GET_CODE (pat
) == SET
)
1850 hash_scan_set (pat
, insn
, table
);
1851 else if (GET_CODE (pat
) == PARALLEL
)
1852 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
1854 rtx x
= XVECEXP (pat
, 0, i
);
1856 if (GET_CODE (x
) == SET
)
1857 hash_scan_set (x
, insn
, table
);
1858 else if (GET_CODE (x
) == CLOBBER
)
1859 hash_scan_clobber (x
, insn
, table
);
1860 else if (GET_CODE (x
) == CALL
)
1861 hash_scan_call (x
, insn
, table
);
1864 else if (GET_CODE (pat
) == CLOBBER
)
1865 hash_scan_clobber (pat
, insn
, table
);
1866 else if (GET_CODE (pat
) == CALL
)
1867 hash_scan_call (pat
, insn
, table
);
1871 dump_hash_table (FILE *file
, const char *name
, struct hash_table
*table
)
1874 /* Flattened out table, so it's printed in proper order. */
1875 struct expr
**flat_table
;
1876 unsigned int *hash_val
;
1879 flat_table
= xcalloc (table
->n_elems
, sizeof (struct expr
*));
1880 hash_val
= xmalloc (table
->n_elems
* sizeof (unsigned int));
1882 for (i
= 0; i
< (int) table
->size
; i
++)
1883 for (expr
= table
->table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
1885 flat_table
[expr
->bitmap_index
] = expr
;
1886 hash_val
[expr
->bitmap_index
] = i
;
1889 fprintf (file
, "%s hash table (%d buckets, %d entries)\n",
1890 name
, table
->size
, table
->n_elems
);
1892 for (i
= 0; i
< (int) table
->n_elems
; i
++)
1893 if (flat_table
[i
] != 0)
1895 expr
= flat_table
[i
];
1896 fprintf (file
, "Index %d (hash value %d)\n ",
1897 expr
->bitmap_index
, hash_val
[i
]);
1898 print_rtl (file
, expr
->expr
);
1899 fprintf (file
, "\n");
1902 fprintf (file
, "\n");
1908 /* Record register first/last/block set information for REGNO in INSN.
1910 first_set records the first place in the block where the register
1911 is set and is used to compute "anticipatability".
1913 last_set records the last place in the block where the register
1914 is set and is used to compute "availability".
1916 last_bb records the block for which first_set and last_set are
1917 valid, as a quick test to invalidate them.
1919 reg_set_in_block records whether the register is set in the block
1920 and is used to compute "transparency". */
1923 record_last_reg_set_info (rtx insn
, int regno
)
1925 struct reg_avail_info
*info
= ®_avail_info
[regno
];
1926 int cuid
= INSN_CUID (insn
);
1928 info
->last_set
= cuid
;
1929 if (info
->last_bb
!= current_bb
)
1931 info
->last_bb
= current_bb
;
1932 info
->first_set
= cuid
;
1933 SET_BIT (reg_set_in_block
[current_bb
->index
], regno
);
1938 /* Record all of the canonicalized MEMs of record_last_mem_set_info's insn.
1939 Note we store a pair of elements in the list, so they have to be
1940 taken off pairwise. */
1943 canon_list_insert (rtx dest ATTRIBUTE_UNUSED
, rtx unused1 ATTRIBUTE_UNUSED
,
1946 rtx dest_addr
, insn
;
1949 while (GET_CODE (dest
) == SUBREG
1950 || GET_CODE (dest
) == ZERO_EXTRACT
1951 || GET_CODE (dest
) == STRICT_LOW_PART
)
1952 dest
= XEXP (dest
, 0);
1954 /* If DEST is not a MEM, then it will not conflict with a load. Note
1955 that function calls are assumed to clobber memory, but are handled
1961 dest_addr
= get_addr (XEXP (dest
, 0));
1962 dest_addr
= canon_rtx (dest_addr
);
1963 insn
= (rtx
) v_insn
;
1964 bb
= BLOCK_NUM (insn
);
1966 canon_modify_mem_list
[bb
] =
1967 alloc_EXPR_LIST (VOIDmode
, dest_addr
, canon_modify_mem_list
[bb
]);
1968 canon_modify_mem_list
[bb
] =
1969 alloc_EXPR_LIST (VOIDmode
, dest
, canon_modify_mem_list
[bb
]);
1972 /* Record memory modification information for INSN. We do not actually care
1973 about the memory location(s) that are set, or even how they are set (consider
1974 a CALL_INSN). We merely need to record which insns modify memory. */
1977 record_last_mem_set_info (rtx insn
)
1979 int bb
= BLOCK_NUM (insn
);
1981 /* load_killed_in_block_p will handle the case of calls clobbering
1983 modify_mem_list
[bb
] = alloc_INSN_LIST (insn
, modify_mem_list
[bb
]);
1984 bitmap_set_bit (modify_mem_list_set
, bb
);
1988 /* Note that traversals of this loop (other than for free-ing)
1989 will break after encountering a CALL_INSN. So, there's no
1990 need to insert a pair of items, as canon_list_insert does. */
1991 canon_modify_mem_list
[bb
] =
1992 alloc_INSN_LIST (insn
, canon_modify_mem_list
[bb
]);
1993 bitmap_set_bit (blocks_with_calls
, bb
);
1996 note_stores (PATTERN (insn
), canon_list_insert
, (void*) insn
);
1999 /* Called from compute_hash_table via note_stores to handle one
2000 SET or CLOBBER in an insn. DATA is really the instruction in which
2001 the SET is taking place. */
2004 record_last_set_info (rtx dest
, rtx setter ATTRIBUTE_UNUSED
, void *data
)
2006 rtx last_set_insn
= (rtx
) data
;
2008 if (GET_CODE (dest
) == SUBREG
)
2009 dest
= SUBREG_REG (dest
);
2012 record_last_reg_set_info (last_set_insn
, REGNO (dest
));
2013 else if (MEM_P (dest
)
2014 /* Ignore pushes, they clobber nothing. */
2015 && ! push_operand (dest
, GET_MODE (dest
)))
2016 record_last_mem_set_info (last_set_insn
);
2019 /* Top level function to create an expression or assignment hash table.
2021 Expression entries are placed in the hash table if
2022 - they are of the form (set (pseudo-reg) src),
2023 - src is something we want to perform GCSE on,
2024 - none of the operands are subsequently modified in the block
2026 Assignment entries are placed in the hash table if
2027 - they are of the form (set (pseudo-reg) src),
2028 - src is something we want to perform const/copy propagation on,
2029 - none of the operands or target are subsequently modified in the block
2031 Currently src must be a pseudo-reg or a const_int.
2033 TABLE is the table computed. */
2036 compute_hash_table_work (struct hash_table
*table
)
2040 /* While we compute the hash table we also compute a bit array of which
2041 registers are set in which blocks.
2042 ??? This isn't needed during const/copy propagation, but it's cheap to
2044 sbitmap_vector_zero (reg_set_in_block
, last_basic_block
);
2046 /* re-Cache any INSN_LIST nodes we have allocated. */
2047 clear_modify_mem_tables ();
2048 /* Some working arrays used to track first and last set in each block. */
2049 reg_avail_info
= gmalloc (max_gcse_regno
* sizeof (struct reg_avail_info
));
2051 for (i
= 0; i
< max_gcse_regno
; ++i
)
2052 reg_avail_info
[i
].last_bb
= NULL
;
2054 FOR_EACH_BB (current_bb
)
2058 int in_libcall_block
;
2060 /* First pass over the instructions records information used to
2061 determine when registers and memory are first and last set.
2062 ??? hard-reg reg_set_in_block computation
2063 could be moved to compute_sets since they currently don't change. */
2065 FOR_BB_INSNS (current_bb
, insn
)
2067 if (! INSN_P (insn
))
2072 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
2073 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
2074 record_last_reg_set_info (insn
, regno
);
2079 note_stores (PATTERN (insn
), record_last_set_info
, insn
);
2082 /* Insert implicit sets in the hash table. */
2084 && implicit_sets
[current_bb
->index
] != NULL_RTX
)
2085 hash_scan_set (implicit_sets
[current_bb
->index
],
2086 BB_HEAD (current_bb
), table
);
2088 /* The next pass builds the hash table. */
2089 in_libcall_block
= 0;
2090 FOR_BB_INSNS (current_bb
, insn
)
2093 if (find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
))
2094 in_libcall_block
= 1;
2095 else if (table
->set_p
&& find_reg_note (insn
, REG_RETVAL
, NULL_RTX
))
2096 in_libcall_block
= 0;
2097 hash_scan_insn (insn
, table
, in_libcall_block
);
2098 if (!table
->set_p
&& find_reg_note (insn
, REG_RETVAL
, NULL_RTX
))
2099 in_libcall_block
= 0;
2103 free (reg_avail_info
);
2104 reg_avail_info
= NULL
;
2107 /* Allocate space for the set/expr hash TABLE.
2108 N_INSNS is the number of instructions in the function.
2109 It is used to determine the number of buckets to use.
2110 SET_P determines whether set or expression table will
2114 alloc_hash_table (int n_insns
, struct hash_table
*table
, int set_p
)
2118 table
->size
= n_insns
/ 4;
2119 if (table
->size
< 11)
2122 /* Attempt to maintain efficient use of hash table.
2123 Making it an odd number is simplest for now.
2124 ??? Later take some measurements. */
2126 n
= table
->size
* sizeof (struct expr
*);
2127 table
->table
= gmalloc (n
);
2128 table
->set_p
= set_p
;
2131 /* Free things allocated by alloc_hash_table. */
2134 free_hash_table (struct hash_table
*table
)
2136 free (table
->table
);
2139 /* Compute the hash TABLE for doing copy/const propagation or
2140 expression hash table. */
2143 compute_hash_table (struct hash_table
*table
)
2145 /* Initialize count of number of entries in hash table. */
2147 memset (table
->table
, 0, table
->size
* sizeof (struct expr
*));
2149 compute_hash_table_work (table
);
2152 /* Expression tracking support. */
2154 /* Lookup REGNO in the set TABLE. The result is a pointer to the
2155 table entry, or NULL if not found. */
2157 static struct expr
*
2158 lookup_set (unsigned int regno
, struct hash_table
*table
)
2160 unsigned int hash
= hash_set (regno
, table
->size
);
2163 expr
= table
->table
[hash
];
2165 while (expr
&& REGNO (SET_DEST (expr
->expr
)) != regno
)
2166 expr
= expr
->next_same_hash
;
2171 /* Return the next entry for REGNO in list EXPR. */
2173 static struct expr
*
2174 next_set (unsigned int regno
, struct expr
*expr
)
2177 expr
= expr
->next_same_hash
;
2178 while (expr
&& REGNO (SET_DEST (expr
->expr
)) != regno
);
2183 /* Like free_INSN_LIST_list or free_EXPR_LIST_list, except that the node
2184 types may be mixed. */
2187 free_insn_expr_list_list (rtx
*listp
)
2191 for (list
= *listp
; list
; list
= next
)
2193 next
= XEXP (list
, 1);
2194 if (GET_CODE (list
) == EXPR_LIST
)
2195 free_EXPR_LIST_node (list
);
2197 free_INSN_LIST_node (list
);
2203 /* Clear canon_modify_mem_list and modify_mem_list tables. */
2205 clear_modify_mem_tables (void)
2210 EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set
, 0, i
, bi
)
2212 free_INSN_LIST_list (modify_mem_list
+ i
);
2213 free_insn_expr_list_list (canon_modify_mem_list
+ i
);
2215 bitmap_clear (modify_mem_list_set
);
2216 bitmap_clear (blocks_with_calls
);
2219 /* Release memory used by modify_mem_list_set. */
2222 free_modify_mem_tables (void)
2224 clear_modify_mem_tables ();
2225 free (modify_mem_list
);
2226 free (canon_modify_mem_list
);
2227 modify_mem_list
= 0;
2228 canon_modify_mem_list
= 0;
2231 /* Reset tables used to keep track of what's still available [since the
2232 start of the block]. */
2235 reset_opr_set_tables (void)
2237 /* Maintain a bitmap of which regs have been set since beginning of
2239 CLEAR_REG_SET (reg_set_bitmap
);
2241 /* Also keep a record of the last instruction to modify memory.
2242 For now this is very trivial, we only record whether any memory
2243 location has been modified. */
2244 clear_modify_mem_tables ();
2247 /* Return nonzero if the operands of X are not set before INSN in
2248 INSN's basic block. */
2251 oprs_not_set_p (rtx x
, rtx insn
)
2260 code
= GET_CODE (x
);
2276 if (load_killed_in_block_p (BLOCK_FOR_INSN (insn
),
2277 INSN_CUID (insn
), x
, 0))
2280 return oprs_not_set_p (XEXP (x
, 0), insn
);
2283 return ! REGNO_REG_SET_P (reg_set_bitmap
, REGNO (x
));
2289 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
2293 /* If we are about to do the last recursive call
2294 needed at this level, change it into iteration.
2295 This function is called enough to be worth it. */
2297 return oprs_not_set_p (XEXP (x
, i
), insn
);
2299 if (! oprs_not_set_p (XEXP (x
, i
), insn
))
2302 else if (fmt
[i
] == 'E')
2303 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
2304 if (! oprs_not_set_p (XVECEXP (x
, i
, j
), insn
))
2311 /* Mark things set by a CALL. */
2314 mark_call (rtx insn
)
2316 if (! CONST_OR_PURE_CALL_P (insn
))
2317 record_last_mem_set_info (insn
);
2320 /* Mark things set by a SET. */
2323 mark_set (rtx pat
, rtx insn
)
2325 rtx dest
= SET_DEST (pat
);
2327 while (GET_CODE (dest
) == SUBREG
2328 || GET_CODE (dest
) == ZERO_EXTRACT
2329 || GET_CODE (dest
) == STRICT_LOW_PART
)
2330 dest
= XEXP (dest
, 0);
2333 SET_REGNO_REG_SET (reg_set_bitmap
, REGNO (dest
));
2334 else if (MEM_P (dest
))
2335 record_last_mem_set_info (insn
);
2337 if (GET_CODE (SET_SRC (pat
)) == CALL
)
2341 /* Record things set by a CLOBBER. */
2344 mark_clobber (rtx pat
, rtx insn
)
2346 rtx clob
= XEXP (pat
, 0);
2348 while (GET_CODE (clob
) == SUBREG
|| GET_CODE (clob
) == STRICT_LOW_PART
)
2349 clob
= XEXP (clob
, 0);
2352 SET_REGNO_REG_SET (reg_set_bitmap
, REGNO (clob
));
2354 record_last_mem_set_info (insn
);
2357 /* Record things set by INSN.
2358 This data is used by oprs_not_set_p. */
2361 mark_oprs_set (rtx insn
)
2363 rtx pat
= PATTERN (insn
);
2366 if (GET_CODE (pat
) == SET
)
2367 mark_set (pat
, insn
);
2368 else if (GET_CODE (pat
) == PARALLEL
)
2369 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
2371 rtx x
= XVECEXP (pat
, 0, i
);
2373 if (GET_CODE (x
) == SET
)
2375 else if (GET_CODE (x
) == CLOBBER
)
2376 mark_clobber (x
, insn
);
2377 else if (GET_CODE (x
) == CALL
)
2381 else if (GET_CODE (pat
) == CLOBBER
)
2382 mark_clobber (pat
, insn
);
2383 else if (GET_CODE (pat
) == CALL
)
2388 /* Compute copy/constant propagation working variables. */
2390 /* Local properties of assignments. */
2391 static sbitmap
*cprop_pavloc
;
2392 static sbitmap
*cprop_absaltered
;
2394 /* Global properties of assignments (computed from the local properties). */
2395 static sbitmap
*cprop_avin
;
2396 static sbitmap
*cprop_avout
;
2398 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
2399 basic blocks. N_SETS is the number of sets. */
2402 alloc_cprop_mem (int n_blocks
, int n_sets
)
2404 cprop_pavloc
= sbitmap_vector_alloc (n_blocks
, n_sets
);
2405 cprop_absaltered
= sbitmap_vector_alloc (n_blocks
, n_sets
);
2407 cprop_avin
= sbitmap_vector_alloc (n_blocks
, n_sets
);
2408 cprop_avout
= sbitmap_vector_alloc (n_blocks
, n_sets
);
2411 /* Free vars used by copy/const propagation. */
2414 free_cprop_mem (void)
2416 sbitmap_vector_free (cprop_pavloc
);
2417 sbitmap_vector_free (cprop_absaltered
);
2418 sbitmap_vector_free (cprop_avin
);
2419 sbitmap_vector_free (cprop_avout
);
2422 /* For each block, compute whether X is transparent. X is either an
2423 expression or an assignment [though we don't care which, for this context
2424 an assignment is treated as an expression]. For each block where an
2425 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
2429 compute_transp (rtx x
, int indx
, sbitmap
*bmap
, int set_p
)
2437 /* repeat is used to turn tail-recursion into iteration since GCC
2438 can't do it when there's no return value. */
2444 code
= GET_CODE (x
);
2450 if (REGNO (x
) < FIRST_PSEUDO_REGISTER
)
2453 if (TEST_BIT (reg_set_in_block
[bb
->index
], REGNO (x
)))
2454 SET_BIT (bmap
[bb
->index
], indx
);
2458 for (r
= reg_set_table
[REGNO (x
)]; r
!= NULL
; r
= r
->next
)
2459 SET_BIT (bmap
[r
->bb_index
], indx
);
2464 if (REGNO (x
) < FIRST_PSEUDO_REGISTER
)
2467 if (TEST_BIT (reg_set_in_block
[bb
->index
], REGNO (x
)))
2468 RESET_BIT (bmap
[bb
->index
], indx
);
2472 for (r
= reg_set_table
[REGNO (x
)]; r
!= NULL
; r
= r
->next
)
2473 RESET_BIT (bmap
[r
->bb_index
], indx
);
2480 if (! MEM_READONLY_P (x
))
2485 /* First handle all the blocks with calls. We don't need to
2486 do any list walking for them. */
2487 EXECUTE_IF_SET_IN_BITMAP (blocks_with_calls
, 0, bb_index
, bi
)
2490 SET_BIT (bmap
[bb_index
], indx
);
2492 RESET_BIT (bmap
[bb_index
], indx
);
2495 /* Now iterate over the blocks which have memory modifications
2496 but which do not have any calls. */
2497 EXECUTE_IF_AND_COMPL_IN_BITMAP (modify_mem_list_set
,
2501 rtx list_entry
= canon_modify_mem_list
[bb_index
];
2505 rtx dest
, dest_addr
;
2507 /* LIST_ENTRY must be an INSN of some kind that sets memory.
2508 Examine each hunk of memory that is modified. */
2510 dest
= XEXP (list_entry
, 0);
2511 list_entry
= XEXP (list_entry
, 1);
2512 dest_addr
= XEXP (list_entry
, 0);
2514 if (canon_true_dependence (dest
, GET_MODE (dest
), dest_addr
,
2515 x
, rtx_addr_varies_p
))
2518 SET_BIT (bmap
[bb_index
], indx
);
2520 RESET_BIT (bmap
[bb_index
], indx
);
2523 list_entry
= XEXP (list_entry
, 1);
2547 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
2551 /* If we are about to do the last recursive call
2552 needed at this level, change it into iteration.
2553 This function is called enough to be worth it. */
2560 compute_transp (XEXP (x
, i
), indx
, bmap
, set_p
);
2562 else if (fmt
[i
] == 'E')
2563 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
2564 compute_transp (XVECEXP (x
, i
, j
), indx
, bmap
, set_p
);
2568 /* Top level routine to do the dataflow analysis needed by copy/const
2572 compute_cprop_data (void)
2574 compute_local_properties (cprop_absaltered
, cprop_pavloc
, NULL
, &set_hash_table
);
2575 compute_available (cprop_pavloc
, cprop_absaltered
,
2576 cprop_avout
, cprop_avin
);
2579 /* Copy/constant propagation. */
2581 /* Maximum number of register uses in an insn that we handle. */
2584 /* Table of uses found in an insn.
2585 Allocated statically to avoid alloc/free complexity and overhead. */
2586 static struct reg_use reg_use_table
[MAX_USES
];
2588 /* Index into `reg_use_table' while building it. */
2589 static int reg_use_count
;
2591 /* Set up a list of register numbers used in INSN. The found uses are stored
2592 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
2593 and contains the number of uses in the table upon exit.
2595 ??? If a register appears multiple times we will record it multiple times.
2596 This doesn't hurt anything but it will slow things down. */
2599 find_used_regs (rtx
*xptr
, void *data ATTRIBUTE_UNUSED
)
2606 /* repeat is used to turn tail-recursion into iteration since GCC
2607 can't do it when there's no return value. */
2612 code
= GET_CODE (x
);
2615 if (reg_use_count
== MAX_USES
)
2618 reg_use_table
[reg_use_count
].reg_rtx
= x
;
2622 /* Recursively scan the operands of this expression. */
2624 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
2628 /* If we are about to do the last recursive call
2629 needed at this level, change it into iteration.
2630 This function is called enough to be worth it. */
2637 find_used_regs (&XEXP (x
, i
), data
);
2639 else if (fmt
[i
] == 'E')
2640 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
2641 find_used_regs (&XVECEXP (x
, i
, j
), data
);
2645 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
2646 Returns nonzero is successful. */
2649 try_replace_reg (rtx from
, rtx to
, rtx insn
)
2651 rtx note
= find_reg_equal_equiv_note (insn
);
2654 rtx set
= single_set (insn
);
2656 validate_replace_src_group (from
, to
, insn
);
2657 if (num_changes_pending () && apply_change_group ())
2660 /* Try to simplify SET_SRC if we have substituted a constant. */
2661 if (success
&& set
&& CONSTANT_P (to
))
2663 src
= simplify_rtx (SET_SRC (set
));
2666 validate_change (insn
, &SET_SRC (set
), src
, 0);
2669 /* If there is already a NOTE, update the expression in it with our
2672 XEXP (note
, 0) = simplify_replace_rtx (XEXP (note
, 0), from
, to
);
2674 if (!success
&& set
&& reg_mentioned_p (from
, SET_SRC (set
)))
2676 /* If above failed and this is a single set, try to simplify the source of
2677 the set given our substitution. We could perhaps try this for multiple
2678 SETs, but it probably won't buy us anything. */
2679 src
= simplify_replace_rtx (SET_SRC (set
), from
, to
);
2681 if (!rtx_equal_p (src
, SET_SRC (set
))
2682 && validate_change (insn
, &SET_SRC (set
), src
, 0))
2685 /* If we've failed to do replacement, have a single SET, don't already
2686 have a note, and have no special SET, add a REG_EQUAL note to not
2687 lose information. */
2688 if (!success
&& note
== 0 && set
!= 0
2689 && GET_CODE (SET_DEST (set
)) != ZERO_EXTRACT
2690 && GET_CODE (SET_DEST (set
)) != STRICT_LOW_PART
)
2691 note
= set_unique_reg_note (insn
, REG_EQUAL
, copy_rtx (src
));
2694 /* REG_EQUAL may get simplified into register.
2695 We don't allow that. Remove that note. This code ought
2696 not to happen, because previous code ought to synthesize
2697 reg-reg move, but be on the safe side. */
2698 if (note
&& REG_P (XEXP (note
, 0)))
2699 remove_note (insn
, note
);
2704 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
2705 NULL no such set is found. */
2707 static struct expr
*
2708 find_avail_set (int regno
, rtx insn
)
2710 /* SET1 contains the last set found that can be returned to the caller for
2711 use in a substitution. */
2712 struct expr
*set1
= 0;
2714 /* Loops are not possible here. To get a loop we would need two sets
2715 available at the start of the block containing INSN. i.e. we would
2716 need two sets like this available at the start of the block:
2718 (set (reg X) (reg Y))
2719 (set (reg Y) (reg X))
2721 This can not happen since the set of (reg Y) would have killed the
2722 set of (reg X) making it unavailable at the start of this block. */
2726 struct expr
*set
= lookup_set (regno
, &set_hash_table
);
2728 /* Find a set that is available at the start of the block
2729 which contains INSN. */
2732 if (TEST_BIT (cprop_avin
[BLOCK_NUM (insn
)], set
->bitmap_index
))
2734 set
= next_set (regno
, set
);
2737 /* If no available set was found we've reached the end of the
2738 (possibly empty) copy chain. */
2742 gcc_assert (GET_CODE (set
->expr
) == SET
);
2744 src
= SET_SRC (set
->expr
);
2746 /* We know the set is available.
2747 Now check that SRC is ANTLOC (i.e. none of the source operands
2748 have changed since the start of the block).
2750 If the source operand changed, we may still use it for the next
2751 iteration of this loop, but we may not use it for substitutions. */
2753 if (gcse_constant_p (src
) || oprs_not_set_p (src
, insn
))
2756 /* If the source of the set is anything except a register, then
2757 we have reached the end of the copy chain. */
2761 /* Follow the copy chain, i.e. start another iteration of the loop
2762 and see if we have an available copy into SRC. */
2763 regno
= REGNO (src
);
2766 /* SET1 holds the last set that was available and anticipatable at
2771 /* Subroutine of cprop_insn that tries to propagate constants into
2772 JUMP_INSNS. JUMP must be a conditional jump. If SETCC is non-NULL
2773 it is the instruction that immediately precedes JUMP, and must be a
2774 single SET of a register. FROM is what we will try to replace,
2775 SRC is the constant we will try to substitute for it. Returns nonzero
2776 if a change was made. */
2779 cprop_jump (basic_block bb
, rtx setcc
, rtx jump
, rtx from
, rtx src
)
2781 rtx
new, set_src
, note_src
;
2782 rtx set
= pc_set (jump
);
2783 rtx note
= find_reg_equal_equiv_note (jump
);
2787 note_src
= XEXP (note
, 0);
2788 if (GET_CODE (note_src
) == EXPR_LIST
)
2789 note_src
= NULL_RTX
;
2791 else note_src
= NULL_RTX
;
2793 /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
2794 set_src
= note_src
? note_src
: SET_SRC (set
);
2796 /* First substitute the SETCC condition into the JUMP instruction,
2797 then substitute that given values into this expanded JUMP. */
2798 if (setcc
!= NULL_RTX
2799 && !modified_between_p (from
, setcc
, jump
)
2800 && !modified_between_p (src
, setcc
, jump
))
2803 rtx setcc_set
= single_set (setcc
);
2804 rtx setcc_note
= find_reg_equal_equiv_note (setcc
);
2805 setcc_src
= (setcc_note
&& GET_CODE (XEXP (setcc_note
, 0)) != EXPR_LIST
)
2806 ? XEXP (setcc_note
, 0) : SET_SRC (setcc_set
);
2807 set_src
= simplify_replace_rtx (set_src
, SET_DEST (setcc_set
),
2813 new = simplify_replace_rtx (set_src
, from
, src
);
2815 /* If no simplification can be made, then try the next register. */
2816 if (rtx_equal_p (new, SET_SRC (set
)))
2819 /* If this is now a no-op delete it, otherwise this must be a valid insn. */
2824 /* Ensure the value computed inside the jump insn to be equivalent
2825 to one computed by setcc. */
2826 if (setcc
&& modified_in_p (new, setcc
))
2828 if (! validate_change (jump
, &SET_SRC (set
), new, 0))
2830 /* When (some) constants are not valid in a comparison, and there
2831 are two registers to be replaced by constants before the entire
2832 comparison can be folded into a constant, we need to keep
2833 intermediate information in REG_EQUAL notes. For targets with
2834 separate compare insns, such notes are added by try_replace_reg.
2835 When we have a combined compare-and-branch instruction, however,
2836 we need to attach a note to the branch itself to make this
2837 optimization work. */
2839 if (!rtx_equal_p (new, note_src
))
2840 set_unique_reg_note (jump
, REG_EQUAL
, copy_rtx (new));
2844 /* Remove REG_EQUAL note after simplification. */
2846 remove_note (jump
, note
);
2848 /* If this has turned into an unconditional jump,
2849 then put a barrier after it so that the unreachable
2850 code will be deleted. */
2851 if (GET_CODE (SET_SRC (set
)) == LABEL_REF
)
2852 emit_barrier_after (jump
);
2856 /* Delete the cc0 setter. */
2857 if (setcc
!= NULL
&& CC0_P (SET_DEST (single_set (setcc
))))
2858 delete_insn (setcc
);
2861 run_jump_opt_after_gcse
= 1;
2863 global_const_prop_count
++;
2864 if (gcse_file
!= NULL
)
2867 "GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
2868 REGNO (from
), INSN_UID (jump
));
2869 print_rtl (gcse_file
, src
);
2870 fprintf (gcse_file
, "\n");
2872 purge_dead_edges (bb
);
2878 constprop_register (rtx insn
, rtx from
, rtx to
, bool alter_jumps
)
2882 /* Check for reg or cc0 setting instructions followed by
2883 conditional branch instructions first. */
2885 && (sset
= single_set (insn
)) != NULL
2887 && any_condjump_p (NEXT_INSN (insn
)) && onlyjump_p (NEXT_INSN (insn
)))
2889 rtx dest
= SET_DEST (sset
);
2890 if ((REG_P (dest
) || CC0_P (dest
))
2891 && cprop_jump (BLOCK_FOR_INSN (insn
), insn
, NEXT_INSN (insn
), from
, to
))
2895 /* Handle normal insns next. */
2896 if (NONJUMP_INSN_P (insn
)
2897 && try_replace_reg (from
, to
, insn
))
2900 /* Try to propagate a CONST_INT into a conditional jump.
2901 We're pretty specific about what we will handle in this
2902 code, we can extend this as necessary over time.
2904 Right now the insn in question must look like
2905 (set (pc) (if_then_else ...)) */
2906 else if (alter_jumps
&& any_condjump_p (insn
) && onlyjump_p (insn
))
2907 return cprop_jump (BLOCK_FOR_INSN (insn
), NULL
, insn
, from
, to
);
2911 /* Perform constant and copy propagation on INSN.
2912 The result is nonzero if a change was made. */
2915 cprop_insn (rtx insn
, int alter_jumps
)
2917 struct reg_use
*reg_used
;
2925 note_uses (&PATTERN (insn
), find_used_regs
, NULL
);
2927 note
= find_reg_equal_equiv_note (insn
);
2929 /* We may win even when propagating constants into notes. */
2931 find_used_regs (&XEXP (note
, 0), NULL
);
2933 for (reg_used
= ®_use_table
[0]; reg_use_count
> 0;
2934 reg_used
++, reg_use_count
--)
2936 unsigned int regno
= REGNO (reg_used
->reg_rtx
);
2940 /* Ignore registers created by GCSE.
2941 We do this because ... */
2942 if (regno
>= max_gcse_regno
)
2945 /* If the register has already been set in this block, there's
2946 nothing we can do. */
2947 if (! oprs_not_set_p (reg_used
->reg_rtx
, insn
))
2950 /* Find an assignment that sets reg_used and is available
2951 at the start of the block. */
2952 set
= find_avail_set (regno
, insn
);
2957 /* ??? We might be able to handle PARALLELs. Later. */
2958 gcc_assert (GET_CODE (pat
) == SET
);
2960 src
= SET_SRC (pat
);
2962 /* Constant propagation. */
2963 if (gcse_constant_p (src
))
2965 if (constprop_register (insn
, reg_used
->reg_rtx
, src
, alter_jumps
))
2968 global_const_prop_count
++;
2969 if (gcse_file
!= NULL
)
2971 fprintf (gcse_file
, "GLOBAL CONST-PROP: Replacing reg %d in ", regno
);
2972 fprintf (gcse_file
, "insn %d with constant ", INSN_UID (insn
));
2973 print_rtl (gcse_file
, src
);
2974 fprintf (gcse_file
, "\n");
2976 if (INSN_DELETED_P (insn
))
2980 else if (REG_P (src
)
2981 && REGNO (src
) >= FIRST_PSEUDO_REGISTER
2982 && REGNO (src
) != regno
)
2984 if (try_replace_reg (reg_used
->reg_rtx
, src
, insn
))
2987 global_copy_prop_count
++;
2988 if (gcse_file
!= NULL
)
2990 fprintf (gcse_file
, "GLOBAL COPY-PROP: Replacing reg %d in insn %d",
2991 regno
, INSN_UID (insn
));
2992 fprintf (gcse_file
, " with reg %d\n", REGNO (src
));
2995 /* The original insn setting reg_used may or may not now be
2996 deletable. We leave the deletion to flow. */
2997 /* FIXME: If it turns out that the insn isn't deletable,
2998 then we may have unnecessarily extended register lifetimes
2999 and made things worse. */
3007 /* Like find_used_regs, but avoid recording uses that appear in
3008 input-output contexts such as zero_extract or pre_dec. This
3009 restricts the cases we consider to those for which local cprop
3010 can legitimately make replacements. */
3013 local_cprop_find_used_regs (rtx
*xptr
, void *data
)
3020 switch (GET_CODE (x
))
3024 case STRICT_LOW_PART
:
3033 /* Can only legitimately appear this early in the context of
3034 stack pushes for function arguments, but handle all of the
3035 codes nonetheless. */
3039 /* Setting a subreg of a register larger than word_mode leaves
3040 the non-written words unchanged. */
3041 if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x
))) > BITS_PER_WORD
)
3049 find_used_regs (xptr
, data
);
3052 /* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
3053 their REG_EQUAL notes need updating. */
3056 do_local_cprop (rtx x
, rtx insn
, bool alter_jumps
, rtx
*libcall_sp
)
3058 rtx newreg
= NULL
, newcnst
= NULL
;
3060 /* Rule out USE instructions and ASM statements as we don't want to
3061 change the hard registers mentioned. */
3063 && (REGNO (x
) >= FIRST_PSEUDO_REGISTER
3064 || (GET_CODE (PATTERN (insn
)) != USE
3065 && asm_noperands (PATTERN (insn
)) < 0)))
3067 cselib_val
*val
= cselib_lookup (x
, GET_MODE (x
), 0);
3068 struct elt_loc_list
*l
;
3072 for (l
= val
->locs
; l
; l
= l
->next
)
3074 rtx this_rtx
= l
->loc
;
3077 /* Don't CSE non-constant values out of libcall blocks. */
3078 if (l
->in_libcall
&& ! CONSTANT_P (this_rtx
))
3081 if (gcse_constant_p (this_rtx
))
3083 if (REG_P (this_rtx
) && REGNO (this_rtx
) >= FIRST_PSEUDO_REGISTER
3084 /* Don't copy propagate if it has attached REG_EQUIV note.
3085 At this point this only function parameters should have
3086 REG_EQUIV notes and if the argument slot is used somewhere
3087 explicitly, it means address of parameter has been taken,
3088 so we should not extend the lifetime of the pseudo. */
3089 && (!(note
= find_reg_note (l
->setting_insn
, REG_EQUIV
, NULL_RTX
))
3090 || ! MEM_P (XEXP (note
, 0))))
3093 if (newcnst
&& constprop_register (insn
, x
, newcnst
, alter_jumps
))
3095 /* If we find a case where we can't fix the retval REG_EQUAL notes
3096 match the new register, we either have to abandon this replacement
3097 or fix delete_trivially_dead_insns to preserve the setting insn,
3098 or make it delete the REG_EUAQL note, and fix up all passes that
3099 require the REG_EQUAL note there. */
3102 adjusted
= adjust_libcall_notes (x
, newcnst
, insn
, libcall_sp
);
3103 gcc_assert (adjusted
);
3105 if (gcse_file
!= NULL
)
3107 fprintf (gcse_file
, "LOCAL CONST-PROP: Replacing reg %d in ",
3109 fprintf (gcse_file
, "insn %d with constant ",
3111 print_rtl (gcse_file
, newcnst
);
3112 fprintf (gcse_file
, "\n");
3114 local_const_prop_count
++;
3117 else if (newreg
&& newreg
!= x
&& try_replace_reg (x
, newreg
, insn
))
3119 adjust_libcall_notes (x
, newreg
, insn
, libcall_sp
);
3120 if (gcse_file
!= NULL
)
3123 "LOCAL COPY-PROP: Replacing reg %d in insn %d",
3124 REGNO (x
), INSN_UID (insn
));
3125 fprintf (gcse_file
, " with reg %d\n", REGNO (newreg
));
3127 local_copy_prop_count
++;
3134 /* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
3135 their REG_EQUAL notes need updating to reflect that OLDREG has been
3136 replaced with NEWVAL in INSN. Return true if all substitutions could
3139 adjust_libcall_notes (rtx oldreg
, rtx newval
, rtx insn
, rtx
*libcall_sp
)
3143 while ((end
= *libcall_sp
++))
3145 rtx note
= find_reg_equal_equiv_note (end
);
3152 if (reg_set_between_p (newval
, PREV_INSN (insn
), end
))
3156 note
= find_reg_equal_equiv_note (end
);
3159 if (reg_mentioned_p (newval
, XEXP (note
, 0)))
3162 while ((end
= *libcall_sp
++));
3166 XEXP (note
, 0) = simplify_replace_rtx (XEXP (note
, 0), oldreg
, newval
);
3172 #define MAX_NESTED_LIBCALLS 9
3174 /* Do local const/copy propagation (i.e. within each basic block).
3175 If ALTER_JUMPS is true, allow propagating into jump insns, which
3176 could modify the CFG. */
3179 local_cprop_pass (bool alter_jumps
)
3183 struct reg_use
*reg_used
;
3184 rtx libcall_stack
[MAX_NESTED_LIBCALLS
+ 1], *libcall_sp
;
3185 bool changed
= false;
3187 cselib_init (false);
3188 libcall_sp
= &libcall_stack
[MAX_NESTED_LIBCALLS
];
3192 FOR_BB_INSNS (bb
, insn
)
3196 rtx note
= find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
);
3200 gcc_assert (libcall_sp
!= libcall_stack
);
3201 *--libcall_sp
= XEXP (note
, 0);
3203 note
= find_reg_note (insn
, REG_RETVAL
, NULL_RTX
);
3206 note
= find_reg_equal_equiv_note (insn
);
3210 note_uses (&PATTERN (insn
), local_cprop_find_used_regs
,
3213 local_cprop_find_used_regs (&XEXP (note
, 0), NULL
);
3215 for (reg_used
= ®_use_table
[0]; reg_use_count
> 0;
3216 reg_used
++, reg_use_count
--)
3217 if (do_local_cprop (reg_used
->reg_rtx
, insn
, alter_jumps
,
3223 if (INSN_DELETED_P (insn
))
3226 while (reg_use_count
);
3228 cselib_process_insn (insn
);
3231 /* Forget everything at the end of a basic block. Make sure we are
3232 not inside a libcall, they should never cross basic blocks. */
3233 cselib_clear_table ();
3234 gcc_assert (libcall_sp
== &libcall_stack
[MAX_NESTED_LIBCALLS
]);
3239 /* Global analysis may get into infinite loops for unreachable blocks. */
3240 if (changed
&& alter_jumps
)
3242 delete_unreachable_blocks ();
3243 free_reg_set_mem ();
3244 alloc_reg_set_mem (max_reg_num ());
3249 /* Forward propagate copies. This includes copies and constants. Return
3250 nonzero if a change was made. */
3253 cprop (int alter_jumps
)
3259 /* Note we start at block 1. */
3260 if (ENTRY_BLOCK_PTR
->next_bb
== EXIT_BLOCK_PTR
)
3262 if (gcse_file
!= NULL
)
3263 fprintf (gcse_file
, "\n");
3268 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
->next_bb
->next_bb
, EXIT_BLOCK_PTR
, next_bb
)
3270 /* Reset tables used to keep track of what's still valid [since the
3271 start of the block]. */
3272 reset_opr_set_tables ();
3274 FOR_BB_INSNS (bb
, insn
)
3277 changed
|= cprop_insn (insn
, alter_jumps
);
3279 /* Keep track of everything modified by this insn. */
3280 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
3281 call mark_oprs_set if we turned the insn into a NOTE. */
3282 if (! NOTE_P (insn
))
3283 mark_oprs_set (insn
);
3287 if (gcse_file
!= NULL
)
3288 fprintf (gcse_file
, "\n");
3293 /* Similar to get_condition, only the resulting condition must be
3294 valid at JUMP, instead of at EARLIEST.
3296 This differs from noce_get_condition in ifcvt.c in that we prefer not to
3297 settle for the condition variable in the jump instruction being integral.
3298 We prefer to be able to record the value of a user variable, rather than
3299 the value of a temporary used in a condition. This could be solved by
3300 recording the value of *every* register scanned by canonicalize_condition,
3301 but this would require some code reorganization. */
3304 fis_get_condition (rtx jump
)
3306 return get_condition (jump
, NULL
, false, true);
3309 /* Check the comparison COND to see if we can safely form an implicit set from
3310 it. COND is either an EQ or NE comparison. */
3313 implicit_set_cond_p (rtx cond
)
3315 enum machine_mode mode
= GET_MODE (XEXP (cond
, 0));
3316 rtx cst
= XEXP (cond
, 1);
3318 /* We can't perform this optimization if either operand might be or might
3319 contain a signed zero. */
3320 if (HONOR_SIGNED_ZEROS (mode
))
3322 /* It is sufficient to check if CST is or contains a zero. We must
3323 handle float, complex, and vector. If any subpart is a zero, then
3324 the optimization can't be performed. */
3325 /* ??? The complex and vector checks are not implemented yet. We just
3326 always return zero for them. */
3327 if (GET_CODE (cst
) == CONST_DOUBLE
)
3330 REAL_VALUE_FROM_CONST_DOUBLE (d
, cst
);
3331 if (REAL_VALUES_EQUAL (d
, dconst0
))
3338 return gcse_constant_p (cst
);
3341 /* Find the implicit sets of a function. An "implicit set" is a constraint
3342 on the value of a variable, implied by a conditional jump. For example,
3343 following "if (x == 2)", the then branch may be optimized as though the
3344 conditional performed an "explicit set", in this example, "x = 2". This
3345 function records the set patterns that are implicit at the start of each
3349 find_implicit_sets (void)
3351 basic_block bb
, dest
;
3357 /* Check for more than one successor. */
3358 if (EDGE_COUNT (bb
->succs
) > 1)
3360 cond
= fis_get_condition (BB_END (bb
));
3363 && (GET_CODE (cond
) == EQ
|| GET_CODE (cond
) == NE
)
3364 && REG_P (XEXP (cond
, 0))
3365 && REGNO (XEXP (cond
, 0)) >= FIRST_PSEUDO_REGISTER
3366 && implicit_set_cond_p (cond
))
3368 dest
= GET_CODE (cond
) == EQ
? BRANCH_EDGE (bb
)->dest
3369 : FALLTHRU_EDGE (bb
)->dest
;
3371 if (dest
&& single_pred_p (dest
)
3372 && dest
!= EXIT_BLOCK_PTR
)
3374 new = gen_rtx_SET (VOIDmode
, XEXP (cond
, 0),
3376 implicit_sets
[dest
->index
] = new;
3379 fprintf(gcse_file
, "Implicit set of reg %d in ",
3380 REGNO (XEXP (cond
, 0)));
3381 fprintf(gcse_file
, "basic block %d\n", dest
->index
);
3389 fprintf (gcse_file
, "Found %d implicit sets\n", count
);
3392 /* Perform one copy/constant propagation pass.
3393 PASS is the pass count. If CPROP_JUMPS is true, perform constant
3394 propagation into conditional jumps. If BYPASS_JUMPS is true,
3395 perform conditional jump bypassing optimizations. */
3398 one_cprop_pass (int pass
, bool cprop_jumps
, bool bypass_jumps
)
3402 global_const_prop_count
= local_const_prop_count
= 0;
3403 global_copy_prop_count
= local_copy_prop_count
= 0;
3405 local_cprop_pass (cprop_jumps
);
3407 /* Determine implicit sets. */
3408 implicit_sets
= xcalloc (last_basic_block
, sizeof (rtx
));
3409 find_implicit_sets ();
3411 alloc_hash_table (max_cuid
, &set_hash_table
, 1);
3412 compute_hash_table (&set_hash_table
);
3414 /* Free implicit_sets before peak usage. */
3415 free (implicit_sets
);
3416 implicit_sets
= NULL
;
3419 dump_hash_table (gcse_file
, "SET", &set_hash_table
);
3420 if (set_hash_table
.n_elems
> 0)
3422 alloc_cprop_mem (last_basic_block
, set_hash_table
.n_elems
);
3423 compute_cprop_data ();
3424 changed
= cprop (cprop_jumps
);
3426 changed
|= bypass_conditional_jumps ();
3430 free_hash_table (&set_hash_table
);
3434 fprintf (gcse_file
, "CPROP of %s, pass %d: %d bytes needed, ",
3435 current_function_name (), pass
, bytes_used
);
3436 fprintf (gcse_file
, "%d local const props, %d local copy props, ",
3437 local_const_prop_count
, local_copy_prop_count
);
3438 fprintf (gcse_file
, "%d global const props, %d global copy props\n\n",
3439 global_const_prop_count
, global_copy_prop_count
);
3441 /* Global analysis may get into infinite loops for unreachable blocks. */
3442 if (changed
&& cprop_jumps
)
3443 delete_unreachable_blocks ();
3448 /* Bypass conditional jumps. */
3450 /* The value of last_basic_block at the beginning of the jump_bypass
3451 pass. The use of redirect_edge_and_branch_force may introduce new
3452 basic blocks, but the data flow analysis is only valid for basic
3453 block indices less than bypass_last_basic_block. */
3455 static int bypass_last_basic_block
;
3457 /* Find a set of REGNO to a constant that is available at the end of basic
3458 block BB. Returns NULL if no such set is found. Based heavily upon
3461 static struct expr
*
3462 find_bypass_set (int regno
, int bb
)
3464 struct expr
*result
= 0;
3469 struct expr
*set
= lookup_set (regno
, &set_hash_table
);
3473 if (TEST_BIT (cprop_avout
[bb
], set
->bitmap_index
))
3475 set
= next_set (regno
, set
);
3481 gcc_assert (GET_CODE (set
->expr
) == SET
);
3483 src
= SET_SRC (set
->expr
);
3484 if (gcse_constant_p (src
))
3490 regno
= REGNO (src
);
3496 /* Subroutine of bypass_block that checks whether a pseudo is killed by
3497 any of the instructions inserted on an edge. Jump bypassing places
3498 condition code setters on CFG edges using insert_insn_on_edge. This
3499 function is required to check that our data flow analysis is still
3500 valid prior to commit_edge_insertions. */
3503 reg_killed_on_edge (rtx reg
, edge e
)
3507 for (insn
= e
->insns
.r
; insn
; insn
= NEXT_INSN (insn
))
3508 if (INSN_P (insn
) && reg_set_p (reg
, insn
))
3514 /* Subroutine of bypass_conditional_jumps that attempts to bypass the given
3515 basic block BB which has more than one predecessor. If not NULL, SETCC
3516 is the first instruction of BB, which is immediately followed by JUMP_INSN
3517 JUMP. Otherwise, SETCC is NULL, and JUMP is the first insn of BB.
3518 Returns nonzero if a change was made.
3520 During the jump bypassing pass, we may place copies of SETCC instructions
3521 on CFG edges. The following routine must be careful to pay attention to
3522 these inserted insns when performing its transformations. */
3525 bypass_block (basic_block bb
, rtx setcc
, rtx jump
)
3530 int may_be_loop_header
;
3534 insn
= (setcc
!= NULL
) ? setcc
: jump
;
3536 /* Determine set of register uses in INSN. */
3538 note_uses (&PATTERN (insn
), find_used_regs
, NULL
);
3539 note
= find_reg_equal_equiv_note (insn
);
3541 find_used_regs (&XEXP (note
, 0), NULL
);
3543 may_be_loop_header
= false;
3544 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3545 if (e
->flags
& EDGE_DFS_BACK
)
3547 may_be_loop_header
= true;
3552 for (ei
= ei_start (bb
->preds
); (e
= ei_safe_edge (ei
)); )
3556 if (e
->flags
& EDGE_COMPLEX
)
3562 /* We can't redirect edges from new basic blocks. */
3563 if (e
->src
->index
>= bypass_last_basic_block
)
3569 /* The irreducible loops created by redirecting of edges entering the
3570 loop from outside would decrease effectiveness of some of the following
3571 optimizations, so prevent this. */
3572 if (may_be_loop_header
3573 && !(e
->flags
& EDGE_DFS_BACK
))
3579 for (i
= 0; i
< reg_use_count
; i
++)
3581 struct reg_use
*reg_used
= ®_use_table
[i
];
3582 unsigned int regno
= REGNO (reg_used
->reg_rtx
);
3583 basic_block dest
, old_dest
;
3587 if (regno
>= max_gcse_regno
)
3590 set
= find_bypass_set (regno
, e
->src
->index
);
3595 /* Check the data flow is valid after edge insertions. */
3596 if (e
->insns
.r
&& reg_killed_on_edge (reg_used
->reg_rtx
, e
))
3599 src
= SET_SRC (pc_set (jump
));
3602 src
= simplify_replace_rtx (src
,
3603 SET_DEST (PATTERN (setcc
)),
3604 SET_SRC (PATTERN (setcc
)));
3606 new = simplify_replace_rtx (src
, reg_used
->reg_rtx
,
3607 SET_SRC (set
->expr
));
3609 /* Jump bypassing may have already placed instructions on
3610 edges of the CFG. We can't bypass an outgoing edge that
3611 has instructions associated with it, as these insns won't
3612 get executed if the incoming edge is redirected. */
3616 edest
= FALLTHRU_EDGE (bb
);
3617 dest
= edest
->insns
.r
? NULL
: edest
->dest
;
3619 else if (GET_CODE (new) == LABEL_REF
)
3621 dest
= BLOCK_FOR_INSN (XEXP (new, 0));
3622 /* Don't bypass edges containing instructions. */
3623 edest
= find_edge (bb
, dest
);
3624 if (edest
&& edest
->insns
.r
)
3630 /* Avoid unification of the edge with other edges from original
3631 branch. We would end up emitting the instruction on "both"
3634 if (dest
&& setcc
&& !CC0_P (SET_DEST (PATTERN (setcc
)))
3635 && find_edge (e
->src
, dest
))
3641 && dest
!= EXIT_BLOCK_PTR
)
3643 redirect_edge_and_branch_force (e
, dest
);
3645 /* Copy the register setter to the redirected edge.
3646 Don't copy CC0 setters, as CC0 is dead after jump. */
3649 rtx pat
= PATTERN (setcc
);
3650 if (!CC0_P (SET_DEST (pat
)))
3651 insert_insn_on_edge (copy_insn (pat
), e
);
3654 if (gcse_file
!= NULL
)
3656 fprintf (gcse_file
, "JUMP-BYPASS: Proved reg %d "
3657 "in jump_insn %d equals constant ",
3658 regno
, INSN_UID (jump
));
3659 print_rtl (gcse_file
, SET_SRC (set
->expr
));
3660 fprintf (gcse_file
, "\nBypass edge from %d->%d to %d\n",
3661 e
->src
->index
, old_dest
->index
, dest
->index
);
3674 /* Find basic blocks with more than one predecessor that only contain a
3675 single conditional jump. If the result of the comparison is known at
3676 compile-time from any incoming edge, redirect that edge to the
3677 appropriate target. Returns nonzero if a change was made.
3679 This function is now mis-named, because we also handle indirect jumps. */
3682 bypass_conditional_jumps (void)
3690 /* Note we start at block 1. */
3691 if (ENTRY_BLOCK_PTR
->next_bb
== EXIT_BLOCK_PTR
)
3694 bypass_last_basic_block
= last_basic_block
;
3695 mark_dfs_back_edges ();
3698 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
->next_bb
->next_bb
,
3699 EXIT_BLOCK_PTR
, next_bb
)
3701 /* Check for more than one predecessor. */
3702 if (!single_pred_p (bb
))
3705 FOR_BB_INSNS (bb
, insn
)
3706 if (NONJUMP_INSN_P (insn
))
3710 if (GET_CODE (PATTERN (insn
)) != SET
)
3713 dest
= SET_DEST (PATTERN (insn
));
3714 if (REG_P (dest
) || CC0_P (dest
))
3719 else if (JUMP_P (insn
))
3721 if ((any_condjump_p (insn
) || computed_jump_p (insn
))
3722 && onlyjump_p (insn
))
3723 changed
|= bypass_block (bb
, setcc
, insn
);
3726 else if (INSN_P (insn
))
3731 /* If we bypassed any register setting insns, we inserted a
3732 copy on the redirected edge. These need to be committed. */
3734 commit_edge_insertions();
3739 /* Compute PRE+LCM working variables. */
3741 /* Local properties of expressions. */
3742 /* Nonzero for expressions that are transparent in the block. */
3743 static sbitmap
*transp
;
3745 /* Nonzero for expressions that are transparent at the end of the block.
3746 This is only zero for expressions killed by abnormal critical edge
3747 created by a calls. */
3748 static sbitmap
*transpout
;
3750 /* Nonzero for expressions that are computed (available) in the block. */
3751 static sbitmap
*comp
;
3753 /* Nonzero for expressions that are locally anticipatable in the block. */
3754 static sbitmap
*antloc
;
3756 /* Nonzero for expressions where this block is an optimal computation
3758 static sbitmap
*pre_optimal
;
3760 /* Nonzero for expressions which are redundant in a particular block. */
3761 static sbitmap
*pre_redundant
;
3763 /* Nonzero for expressions which should be inserted on a specific edge. */
3764 static sbitmap
*pre_insert_map
;
3766 /* Nonzero for expressions which should be deleted in a specific block. */
3767 static sbitmap
*pre_delete_map
;
3769 /* Contains the edge_list returned by pre_edge_lcm. */
3770 static struct edge_list
*edge_list
;
3772 /* Redundant insns. */
3773 static sbitmap pre_redundant_insns
;
3775 /* Allocate vars used for PRE analysis. */
3778 alloc_pre_mem (int n_blocks
, int n_exprs
)
3780 transp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
3781 comp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
3782 antloc
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
3785 pre_redundant
= NULL
;
3786 pre_insert_map
= NULL
;
3787 pre_delete_map
= NULL
;
3788 ae_kill
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
3790 /* pre_insert and pre_delete are allocated later. */
3793 /* Free vars used for PRE analysis. */
3798 sbitmap_vector_free (transp
);
3799 sbitmap_vector_free (comp
);
3801 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
3804 sbitmap_vector_free (pre_optimal
);
3806 sbitmap_vector_free (pre_redundant
);
3808 sbitmap_vector_free (pre_insert_map
);
3810 sbitmap_vector_free (pre_delete_map
);
3812 transp
= comp
= NULL
;
3813 pre_optimal
= pre_redundant
= pre_insert_map
= pre_delete_map
= NULL
;
3816 /* Top level routine to do the dataflow analysis needed by PRE. */
3819 compute_pre_data (void)
3821 sbitmap trapping_expr
;
3825 compute_local_properties (transp
, comp
, antloc
, &expr_hash_table
);
3826 sbitmap_vector_zero (ae_kill
, last_basic_block
);
3828 /* Collect expressions which might trap. */
3829 trapping_expr
= sbitmap_alloc (expr_hash_table
.n_elems
);
3830 sbitmap_zero (trapping_expr
);
3831 for (ui
= 0; ui
< expr_hash_table
.size
; ui
++)
3834 for (e
= expr_hash_table
.table
[ui
]; e
!= NULL
; e
= e
->next_same_hash
)
3835 if (may_trap_p (e
->expr
))
3836 SET_BIT (trapping_expr
, e
->bitmap_index
);
3839 /* Compute ae_kill for each basic block using:
3849 /* If the current block is the destination of an abnormal edge, we
3850 kill all trapping expressions because we won't be able to properly
3851 place the instruction on the edge. So make them neither
3852 anticipatable nor transparent. This is fairly conservative. */
3853 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3854 if (e
->flags
& EDGE_ABNORMAL
)
3856 sbitmap_difference (antloc
[bb
->index
], antloc
[bb
->index
], trapping_expr
);
3857 sbitmap_difference (transp
[bb
->index
], transp
[bb
->index
], trapping_expr
);
3861 sbitmap_a_or_b (ae_kill
[bb
->index
], transp
[bb
->index
], comp
[bb
->index
]);
3862 sbitmap_not (ae_kill
[bb
->index
], ae_kill
[bb
->index
]);
3865 edge_list
= pre_edge_lcm (gcse_file
, expr_hash_table
.n_elems
, transp
, comp
, antloc
,
3866 ae_kill
, &pre_insert_map
, &pre_delete_map
);
3867 sbitmap_vector_free (antloc
);
3869 sbitmap_vector_free (ae_kill
);
3871 sbitmap_free (trapping_expr
);
3876 /* Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach
3879 VISITED is a pointer to a working buffer for tracking which BB's have
3880 been visited. It is NULL for the top-level call.
3882 We treat reaching expressions that go through blocks containing the same
3883 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
3884 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
3885 2 as not reaching. The intent is to improve the probability of finding
3886 only one reaching expression and to reduce register lifetimes by picking
3887 the closest such expression. */
3890 pre_expr_reaches_here_p_work (basic_block occr_bb
, struct expr
*expr
, basic_block bb
, char *visited
)
3895 FOR_EACH_EDGE (pred
, ei
, bb
->preds
)
3897 basic_block pred_bb
= pred
->src
;
3899 if (pred
->src
== ENTRY_BLOCK_PTR
3900 /* Has predecessor has already been visited? */
3901 || visited
[pred_bb
->index
])
3902 ;/* Nothing to do. */
3904 /* Does this predecessor generate this expression? */
3905 else if (TEST_BIT (comp
[pred_bb
->index
], expr
->bitmap_index
))
3907 /* Is this the occurrence we're looking for?
3908 Note that there's only one generating occurrence per block
3909 so we just need to check the block number. */
3910 if (occr_bb
== pred_bb
)
3913 visited
[pred_bb
->index
] = 1;
3915 /* Ignore this predecessor if it kills the expression. */
3916 else if (! TEST_BIT (transp
[pred_bb
->index
], expr
->bitmap_index
))
3917 visited
[pred_bb
->index
] = 1;
3919 /* Neither gen nor kill. */
3922 visited
[pred_bb
->index
] = 1;
3923 if (pre_expr_reaches_here_p_work (occr_bb
, expr
, pred_bb
, visited
))
3928 /* All paths have been checked. */
3932 /* The wrapper for pre_expr_reaches_here_work that ensures that any
3933 memory allocated for that function is returned. */
3936 pre_expr_reaches_here_p (basic_block occr_bb
, struct expr
*expr
, basic_block bb
)
3939 char *visited
= xcalloc (last_basic_block
, 1);
3941 rval
= pre_expr_reaches_here_p_work (occr_bb
, expr
, bb
, visited
);
3948 /* Given an expr, generate RTL which we can insert at the end of a BB,
3949 or on an edge. Set the block number of any insns generated to
3953 process_insert_insn (struct expr
*expr
)
3955 rtx reg
= expr
->reaching_reg
;
3956 rtx exp
= copy_rtx (expr
->expr
);
3961 /* If the expression is something that's an operand, like a constant,
3962 just copy it to a register. */
3963 if (general_operand (exp
, GET_MODE (reg
)))
3964 emit_move_insn (reg
, exp
);
3966 /* Otherwise, make a new insn to compute this expression and make sure the
3967 insn will be recognized (this also adds any needed CLOBBERs). Copy the
3968 expression to make sure we don't have any sharing issues. */
3971 rtx insn
= emit_insn (gen_rtx_SET (VOIDmode
, reg
, exp
));
3973 if (insn_invalid_p (insn
))
3984 /* Add EXPR to the end of basic block BB.
3986 This is used by both the PRE and code hoisting.
3988 For PRE, we want to verify that the expr is either transparent
3989 or locally anticipatable in the target block. This check makes
3990 no sense for code hoisting. */
3993 insert_insn_end_bb (struct expr
*expr
, basic_block bb
, int pre
)
3995 rtx insn
= BB_END (bb
);
3997 rtx reg
= expr
->reaching_reg
;
3998 int regno
= REGNO (reg
);
4001 pat
= process_insert_insn (expr
);
4002 gcc_assert (pat
&& INSN_P (pat
));
4005 while (NEXT_INSN (pat_end
) != NULL_RTX
)
4006 pat_end
= NEXT_INSN (pat_end
);
4008 /* If the last insn is a jump, insert EXPR in front [taking care to
4009 handle cc0, etc. properly]. Similarly we need to care trapping
4010 instructions in presence of non-call exceptions. */
4013 || (NONJUMP_INSN_P (insn
)
4014 && (!single_succ_p (bb
)
4015 || single_succ_edge (bb
)->flags
& EDGE_ABNORMAL
)))
4020 /* It should always be the case that we can put these instructions
4021 anywhere in the basic block with performing PRE optimizations.
4023 gcc_assert (!NONJUMP_INSN_P (insn
) || !pre
4024 || TEST_BIT (antloc
[bb
->index
], expr
->bitmap_index
)
4025 || TEST_BIT (transp
[bb
->index
], expr
->bitmap_index
));
4027 /* If this is a jump table, then we can't insert stuff here. Since
4028 we know the previous real insn must be the tablejump, we insert
4029 the new instruction just before the tablejump. */
4030 if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
4031 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
4032 insn
= prev_real_insn (insn
);
4035 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
4036 if cc0 isn't set. */
4037 note
= find_reg_note (insn
, REG_CC_SETTER
, NULL_RTX
);
4039 insn
= XEXP (note
, 0);
4042 rtx maybe_cc0_setter
= prev_nonnote_insn (insn
);
4043 if (maybe_cc0_setter
4044 && INSN_P (maybe_cc0_setter
)
4045 && sets_cc0_p (PATTERN (maybe_cc0_setter
)))
4046 insn
= maybe_cc0_setter
;
4049 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4050 new_insn
= emit_insn_before_noloc (pat
, insn
);
4053 /* Likewise if the last insn is a call, as will happen in the presence
4054 of exception handling. */
4055 else if (CALL_P (insn
)
4056 && (!single_succ_p (bb
)
4057 || single_succ_edge (bb
)->flags
& EDGE_ABNORMAL
))
4059 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4060 we search backward and place the instructions before the first
4061 parameter is loaded. Do this for everyone for consistency and a
4062 presumption that we'll get better code elsewhere as well.
4064 It should always be the case that we can put these instructions
4065 anywhere in the basic block with performing PRE optimizations.
4069 || TEST_BIT (antloc
[bb
->index
], expr
->bitmap_index
)
4070 || TEST_BIT (transp
[bb
->index
], expr
->bitmap_index
));
4072 /* Since different machines initialize their parameter registers
4073 in different orders, assume nothing. Collect the set of all
4074 parameter registers. */
4075 insn
= find_first_parameter_load (insn
, BB_HEAD (bb
));
4077 /* If we found all the parameter loads, then we want to insert
4078 before the first parameter load.
4080 If we did not find all the parameter loads, then we might have
4081 stopped on the head of the block, which could be a CODE_LABEL.
4082 If we inserted before the CODE_LABEL, then we would be putting
4083 the insn in the wrong basic block. In that case, put the insn
4084 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4085 while (LABEL_P (insn
)
4086 || NOTE_INSN_BASIC_BLOCK_P (insn
))
4087 insn
= NEXT_INSN (insn
);
4089 new_insn
= emit_insn_before_noloc (pat
, insn
);
4092 new_insn
= emit_insn_after_noloc (pat
, insn
);
4098 add_label_notes (PATTERN (pat
), new_insn
);
4099 note_stores (PATTERN (pat
), record_set_info
, pat
);
4103 pat
= NEXT_INSN (pat
);
4106 gcse_create_count
++;
4110 fprintf (gcse_file
, "PRE/HOIST: end of bb %d, insn %d, ",
4111 bb
->index
, INSN_UID (new_insn
));
4112 fprintf (gcse_file
, "copying expression %d to reg %d\n",
4113 expr
->bitmap_index
, regno
);
4117 /* Insert partially redundant expressions on edges in the CFG to make
4118 the expressions fully redundant. */
4121 pre_edge_insert (struct edge_list
*edge_list
, struct expr
**index_map
)
4123 int e
, i
, j
, num_edges
, set_size
, did_insert
= 0;
4126 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4127 if it reaches any of the deleted expressions. */
4129 set_size
= pre_insert_map
[0]->size
;
4130 num_edges
= NUM_EDGES (edge_list
);
4131 inserted
= sbitmap_vector_alloc (num_edges
, expr_hash_table
.n_elems
);
4132 sbitmap_vector_zero (inserted
, num_edges
);
4134 for (e
= 0; e
< num_edges
; e
++)
4137 basic_block bb
= INDEX_EDGE_PRED_BB (edge_list
, e
);
4139 for (i
= indx
= 0; i
< set_size
; i
++, indx
+= SBITMAP_ELT_BITS
)
4141 SBITMAP_ELT_TYPE insert
= pre_insert_map
[e
]->elms
[i
];
4143 for (j
= indx
; insert
&& j
< (int) expr_hash_table
.n_elems
; j
++, insert
>>= 1)
4144 if ((insert
& 1) != 0 && index_map
[j
]->reaching_reg
!= NULL_RTX
)
4146 struct expr
*expr
= index_map
[j
];
4149 /* Now look at each deleted occurrence of this expression. */
4150 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4152 if (! occr
->deleted_p
)
4155 /* Insert this expression on this edge if it would
4156 reach the deleted occurrence in BB. */
4157 if (!TEST_BIT (inserted
[e
], j
))
4160 edge eg
= INDEX_EDGE (edge_list
, e
);
4162 /* We can't insert anything on an abnormal and
4163 critical edge, so we insert the insn at the end of
4164 the previous block. There are several alternatives
4165 detailed in Morgans book P277 (sec 10.5) for
4166 handling this situation. This one is easiest for
4169 if (eg
->flags
& EDGE_ABNORMAL
)
4170 insert_insn_end_bb (index_map
[j
], bb
, 0);
4173 insn
= process_insert_insn (index_map
[j
]);
4174 insert_insn_on_edge (insn
, eg
);
4179 fprintf (gcse_file
, "PRE/HOIST: edge (%d,%d), ",
4181 INDEX_EDGE_SUCC_BB (edge_list
, e
)->index
);
4182 fprintf (gcse_file
, "copy expression %d\n",
4183 expr
->bitmap_index
);
4186 update_ld_motion_stores (expr
);
4187 SET_BIT (inserted
[e
], j
);
4189 gcse_create_count
++;
4196 sbitmap_vector_free (inserted
);
4200 /* Copy the result of EXPR->EXPR generated by INSN to EXPR->REACHING_REG.
4201 Given "old_reg <- expr" (INSN), instead of adding after it
4202 reaching_reg <- old_reg
4203 it's better to do the following:
4204 reaching_reg <- expr
4205 old_reg <- reaching_reg
4206 because this way copy propagation can discover additional PRE
4207 opportunities. But if this fails, we try the old way.
4208 When "expr" is a store, i.e.
4209 given "MEM <- old_reg", instead of adding after it
4210 reaching_reg <- old_reg
4211 it's better to add it before as follows:
4212 reaching_reg <- old_reg
4213 MEM <- reaching_reg. */
4216 pre_insert_copy_insn (struct expr
*expr
, rtx insn
)
4218 rtx reg
= expr
->reaching_reg
;
4219 int regno
= REGNO (reg
);
4220 int indx
= expr
->bitmap_index
;
4221 rtx pat
= PATTERN (insn
);
4226 /* This block matches the logic in hash_scan_insn. */
4227 switch (GET_CODE (pat
))
4234 /* Search through the parallel looking for the set whose
4235 source was the expression that we're interested in. */
4237 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
4239 rtx x
= XVECEXP (pat
, 0, i
);
4240 if (GET_CODE (x
) == SET
4241 && expr_equiv_p (SET_SRC (x
), expr
->expr
))
4253 if (REG_P (SET_DEST (set
)))
4255 old_reg
= SET_DEST (set
);
4256 /* Check if we can modify the set destination in the original insn. */
4257 if (validate_change (insn
, &SET_DEST (set
), reg
, 0))
4259 new_insn
= gen_move_insn (old_reg
, reg
);
4260 new_insn
= emit_insn_after (new_insn
, insn
);
4262 /* Keep register set table up to date. */
4263 record_one_set (regno
, insn
);
4267 new_insn
= gen_move_insn (reg
, old_reg
);
4268 new_insn
= emit_insn_after (new_insn
, insn
);
4270 /* Keep register set table up to date. */
4271 record_one_set (regno
, new_insn
);
4274 else /* This is possible only in case of a store to memory. */
4276 old_reg
= SET_SRC (set
);
4277 new_insn
= gen_move_insn (reg
, old_reg
);
4279 /* Check if we can modify the set source in the original insn. */
4280 if (validate_change (insn
, &SET_SRC (set
), reg
, 0))
4281 new_insn
= emit_insn_before (new_insn
, insn
);
4283 new_insn
= emit_insn_after (new_insn
, insn
);
4285 /* Keep register set table up to date. */
4286 record_one_set (regno
, new_insn
);
4289 gcse_create_count
++;
4293 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4294 BLOCK_NUM (insn
), INSN_UID (new_insn
), indx
,
4295 INSN_UID (insn
), regno
);
4298 /* Copy available expressions that reach the redundant expression
4299 to `reaching_reg'. */
4302 pre_insert_copies (void)
4304 unsigned int i
, added_copy
;
4309 /* For each available expression in the table, copy the result to
4310 `reaching_reg' if the expression reaches a deleted one.
4312 ??? The current algorithm is rather brute force.
4313 Need to do some profiling. */
4315 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4316 for (expr
= expr_hash_table
.table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4318 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4319 we don't want to insert a copy here because the expression may not
4320 really be redundant. So only insert an insn if the expression was
4321 deleted. This test also avoids further processing if the
4322 expression wasn't deleted anywhere. */
4323 if (expr
->reaching_reg
== NULL
)
4326 /* Set when we add a copy for that expression. */
4329 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4331 if (! occr
->deleted_p
)
4334 for (avail
= expr
->avail_occr
; avail
!= NULL
; avail
= avail
->next
)
4336 rtx insn
= avail
->insn
;
4338 /* No need to handle this one if handled already. */
4339 if (avail
->copied_p
)
4342 /* Don't handle this one if it's a redundant one. */
4343 if (TEST_BIT (pre_redundant_insns
, INSN_CUID (insn
)))
4346 /* Or if the expression doesn't reach the deleted one. */
4347 if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail
->insn
),
4349 BLOCK_FOR_INSN (occr
->insn
)))
4354 /* Copy the result of avail to reaching_reg. */
4355 pre_insert_copy_insn (expr
, insn
);
4356 avail
->copied_p
= 1;
4361 update_ld_motion_stores (expr
);
4365 /* Emit move from SRC to DEST noting the equivalence with expression computed
4368 gcse_emit_move_after (rtx src
, rtx dest
, rtx insn
)
4371 rtx set
= single_set (insn
), set2
;
4375 /* This should never fail since we're creating a reg->reg copy
4376 we've verified to be valid. */
4378 new = emit_insn_after (gen_move_insn (dest
, src
), insn
);
4380 /* Note the equivalence for local CSE pass. */
4381 set2
= single_set (new);
4382 if (!set2
|| !rtx_equal_p (SET_DEST (set2
), dest
))
4384 if ((note
= find_reg_equal_equiv_note (insn
)))
4385 eqv
= XEXP (note
, 0);
4387 eqv
= SET_SRC (set
);
4389 set_unique_reg_note (new, REG_EQUAL
, copy_insn_1 (eqv
));
4394 /* Delete redundant computations.
4395 Deletion is done by changing the insn to copy the `reaching_reg' of
4396 the expression into the result of the SET. It is left to later passes
4397 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4399 Returns nonzero if a change is made. */
4410 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4411 for (expr
= expr_hash_table
.table
[i
];
4413 expr
= expr
->next_same_hash
)
4415 int indx
= expr
->bitmap_index
;
4417 /* We only need to search antic_occr since we require
4420 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4422 rtx insn
= occr
->insn
;
4424 basic_block bb
= BLOCK_FOR_INSN (insn
);
4426 /* We only delete insns that have a single_set. */
4427 if (TEST_BIT (pre_delete_map
[bb
->index
], indx
)
4428 && (set
= single_set (insn
)) != 0)
4430 /* Create a pseudo-reg to store the result of reaching
4431 expressions into. Get the mode for the new pseudo from
4432 the mode of the original destination pseudo. */
4433 if (expr
->reaching_reg
== NULL
)
4435 = gen_reg_rtx (GET_MODE (SET_DEST (set
)));
4437 gcse_emit_move_after (expr
->reaching_reg
, SET_DEST (set
), insn
);
4439 occr
->deleted_p
= 1;
4440 SET_BIT (pre_redundant_insns
, INSN_CUID (insn
));
4447 "PRE: redundant insn %d (expression %d) in ",
4448 INSN_UID (insn
), indx
);
4449 fprintf (gcse_file
, "bb %d, reaching reg is %d\n",
4450 bb
->index
, REGNO (expr
->reaching_reg
));
4459 /* Perform GCSE optimizations using PRE.
4460 This is called by one_pre_gcse_pass after all the dataflow analysis
4463 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4464 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4465 Compiler Design and Implementation.
4467 ??? A new pseudo reg is created to hold the reaching expression. The nice
4468 thing about the classical approach is that it would try to use an existing
4469 reg. If the register can't be adequately optimized [i.e. we introduce
4470 reload problems], one could add a pass here to propagate the new register
4473 ??? We don't handle single sets in PARALLELs because we're [currently] not
4474 able to copy the rest of the parallel when we insert copies to create full
4475 redundancies from partial redundancies. However, there's no reason why we
4476 can't handle PARALLELs in the cases where there are no partial
4483 int did_insert
, changed
;
4484 struct expr
**index_map
;
4487 /* Compute a mapping from expression number (`bitmap_index') to
4488 hash table entry. */
4490 index_map
= xcalloc (expr_hash_table
.n_elems
, sizeof (struct expr
*));
4491 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4492 for (expr
= expr_hash_table
.table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4493 index_map
[expr
->bitmap_index
] = expr
;
4495 /* Reset bitmap used to track which insns are redundant. */
4496 pre_redundant_insns
= sbitmap_alloc (max_cuid
);
4497 sbitmap_zero (pre_redundant_insns
);
4499 /* Delete the redundant insns first so that
4500 - we know what register to use for the new insns and for the other
4501 ones with reaching expressions
4502 - we know which insns are redundant when we go to create copies */
4504 changed
= pre_delete ();
4506 did_insert
= pre_edge_insert (edge_list
, index_map
);
4508 /* In other places with reaching expressions, copy the expression to the
4509 specially allocated pseudo-reg that reaches the redundant expr. */
4510 pre_insert_copies ();
4513 commit_edge_insertions ();
4518 sbitmap_free (pre_redundant_insns
);
4522 /* Top level routine to perform one PRE GCSE pass.
4524 Return nonzero if a change was made. */
4527 one_pre_gcse_pass (int pass
)
4531 gcse_subst_count
= 0;
4532 gcse_create_count
= 0;
4534 alloc_hash_table (max_cuid
, &expr_hash_table
, 0);
4535 add_noreturn_fake_exit_edges ();
4537 compute_ld_motion_mems ();
4539 compute_hash_table (&expr_hash_table
);
4540 trim_ld_motion_mems ();
4542 dump_hash_table (gcse_file
, "Expression", &expr_hash_table
);
4544 if (expr_hash_table
.n_elems
> 0)
4546 alloc_pre_mem (last_basic_block
, expr_hash_table
.n_elems
);
4547 compute_pre_data ();
4548 changed
|= pre_gcse ();
4549 free_edge_list (edge_list
);
4554 remove_fake_exit_edges ();
4555 free_hash_table (&expr_hash_table
);
4559 fprintf (gcse_file
, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
4560 current_function_name (), pass
, bytes_used
);
4561 fprintf (gcse_file
, "%d substs, %d insns created\n",
4562 gcse_subst_count
, gcse_create_count
);
4568 /* If X contains any LABEL_REF's, add REG_LABEL notes for them to INSN.
4569 If notes are added to an insn which references a CODE_LABEL, the
4570 LABEL_NUSES count is incremented. We have to add REG_LABEL notes,
4571 because the following loop optimization pass requires them. */
4573 /* ??? This is very similar to the loop.c add_label_notes function. We
4574 could probably share code here. */
4576 /* ??? If there was a jump optimization pass after gcse and before loop,
4577 then we would not need to do this here, because jump would add the
4578 necessary REG_LABEL notes. */
4581 add_label_notes (rtx x
, rtx insn
)
4583 enum rtx_code code
= GET_CODE (x
);
4587 if (code
== LABEL_REF
&& !LABEL_REF_NONLOCAL_P (x
))
4589 /* This code used to ignore labels that referred to dispatch tables to
4590 avoid flow generating (slightly) worse code.
4592 We no longer ignore such label references (see LABEL_REF handling in
4593 mark_jump_label for additional information). */
4595 REG_NOTES (insn
) = gen_rtx_INSN_LIST (REG_LABEL
, XEXP (x
, 0),
4597 if (LABEL_P (XEXP (x
, 0)))
4598 LABEL_NUSES (XEXP (x
, 0))++;
4602 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
4605 add_label_notes (XEXP (x
, i
), insn
);
4606 else if (fmt
[i
] == 'E')
4607 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
4608 add_label_notes (XVECEXP (x
, i
, j
), insn
);
4612 /* Compute transparent outgoing information for each block.
4614 An expression is transparent to an edge unless it is killed by
4615 the edge itself. This can only happen with abnormal control flow,
4616 when the edge is traversed through a call. This happens with
4617 non-local labels and exceptions.
4619 This would not be necessary if we split the edge. While this is
4620 normally impossible for abnormal critical edges, with some effort
4621 it should be possible with exception handling, since we still have
4622 control over which handler should be invoked. But due to increased
4623 EH table sizes, this may not be worthwhile. */
4626 compute_transpout (void)
4632 sbitmap_vector_ones (transpout
, last_basic_block
);
4636 /* Note that flow inserted a nop a the end of basic blocks that
4637 end in call instructions for reasons other than abnormal
4639 if (! CALL_P (BB_END (bb
)))
4642 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4643 for (expr
= expr_hash_table
.table
[i
]; expr
; expr
= expr
->next_same_hash
)
4644 if (MEM_P (expr
->expr
))
4646 if (GET_CODE (XEXP (expr
->expr
, 0)) == SYMBOL_REF
4647 && CONSTANT_POOL_ADDRESS_P (XEXP (expr
->expr
, 0)))
4650 /* ??? Optimally, we would use interprocedural alias
4651 analysis to determine if this mem is actually killed
4653 RESET_BIT (transpout
[bb
->index
], expr
->bitmap_index
);
4658 /* Code Hoisting variables and subroutines. */
4660 /* Very busy expressions. */
4661 static sbitmap
*hoist_vbein
;
4662 static sbitmap
*hoist_vbeout
;
4664 /* Hoistable expressions. */
4665 static sbitmap
*hoist_exprs
;
4667 /* ??? We could compute post dominators and run this algorithm in
4668 reverse to perform tail merging, doing so would probably be
4669 more effective than the tail merging code in jump.c.
4671 It's unclear if tail merging could be run in parallel with
4672 code hoisting. It would be nice. */
4674 /* Allocate vars used for code hoisting analysis. */
4677 alloc_code_hoist_mem (int n_blocks
, int n_exprs
)
4679 antloc
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4680 transp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4681 comp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4683 hoist_vbein
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4684 hoist_vbeout
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4685 hoist_exprs
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4686 transpout
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4689 /* Free vars used for code hoisting analysis. */
4692 free_code_hoist_mem (void)
4694 sbitmap_vector_free (antloc
);
4695 sbitmap_vector_free (transp
);
4696 sbitmap_vector_free (comp
);
4698 sbitmap_vector_free (hoist_vbein
);
4699 sbitmap_vector_free (hoist_vbeout
);
4700 sbitmap_vector_free (hoist_exprs
);
4701 sbitmap_vector_free (transpout
);
4703 free_dominance_info (CDI_DOMINATORS
);
4706 /* Compute the very busy expressions at entry/exit from each block.
4708 An expression is very busy if all paths from a given point
4709 compute the expression. */
4712 compute_code_hoist_vbeinout (void)
4714 int changed
, passes
;
4717 sbitmap_vector_zero (hoist_vbeout
, last_basic_block
);
4718 sbitmap_vector_zero (hoist_vbein
, last_basic_block
);
4727 /* We scan the blocks in the reverse order to speed up
4729 FOR_EACH_BB_REVERSE (bb
)
4731 changed
|= sbitmap_a_or_b_and_c_cg (hoist_vbein
[bb
->index
], antloc
[bb
->index
],
4732 hoist_vbeout
[bb
->index
], transp
[bb
->index
]);
4733 if (bb
->next_bb
!= EXIT_BLOCK_PTR
)
4734 sbitmap_intersection_of_succs (hoist_vbeout
[bb
->index
], hoist_vbein
, bb
->index
);
4741 fprintf (gcse_file
, "hoisting vbeinout computation: %d passes\n", passes
);
4744 /* Top level routine to do the dataflow analysis needed by code hoisting. */
4747 compute_code_hoist_data (void)
4749 compute_local_properties (transp
, comp
, antloc
, &expr_hash_table
);
4750 compute_transpout ();
4751 compute_code_hoist_vbeinout ();
4752 calculate_dominance_info (CDI_DOMINATORS
);
4754 fprintf (gcse_file
, "\n");
4757 /* Determine if the expression identified by EXPR_INDEX would
4758 reach BB unimpared if it was placed at the end of EXPR_BB.
4760 It's unclear exactly what Muchnick meant by "unimpared". It seems
4761 to me that the expression must either be computed or transparent in
4762 *every* block in the path(s) from EXPR_BB to BB. Any other definition
4763 would allow the expression to be hoisted out of loops, even if
4764 the expression wasn't a loop invariant.
4766 Contrast this to reachability for PRE where an expression is
4767 considered reachable if *any* path reaches instead of *all*
4771 hoist_expr_reaches_here_p (basic_block expr_bb
, int expr_index
, basic_block bb
, char *visited
)
4775 int visited_allocated_locally
= 0;
4778 if (visited
== NULL
)
4780 visited_allocated_locally
= 1;
4781 visited
= xcalloc (last_basic_block
, 1);
4784 FOR_EACH_EDGE (pred
, ei
, bb
->preds
)
4786 basic_block pred_bb
= pred
->src
;
4788 if (pred
->src
== ENTRY_BLOCK_PTR
)
4790 else if (pred_bb
== expr_bb
)
4792 else if (visited
[pred_bb
->index
])
4795 /* Does this predecessor generate this expression? */
4796 else if (TEST_BIT (comp
[pred_bb
->index
], expr_index
))
4798 else if (! TEST_BIT (transp
[pred_bb
->index
], expr_index
))
4804 visited
[pred_bb
->index
] = 1;
4805 if (! hoist_expr_reaches_here_p (expr_bb
, expr_index
,
4810 if (visited_allocated_locally
)
4813 return (pred
== NULL
);
4816 /* Actually perform code hoisting. */
4821 basic_block bb
, dominated
;
4823 unsigned int domby_len
;
4825 struct expr
**index_map
;
4828 sbitmap_vector_zero (hoist_exprs
, last_basic_block
);
4830 /* Compute a mapping from expression number (`bitmap_index') to
4831 hash table entry. */
4833 index_map
= xcalloc (expr_hash_table
.n_elems
, sizeof (struct expr
*));
4834 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4835 for (expr
= expr_hash_table
.table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4836 index_map
[expr
->bitmap_index
] = expr
;
4838 /* Walk over each basic block looking for potentially hoistable
4839 expressions, nothing gets hoisted from the entry block. */
4843 int insn_inserted_p
;
4845 domby_len
= get_dominated_by (CDI_DOMINATORS
, bb
, &domby
);
4846 /* Examine each expression that is very busy at the exit of this
4847 block. These are the potentially hoistable expressions. */
4848 for (i
= 0; i
< hoist_vbeout
[bb
->index
]->n_bits
; i
++)
4852 if (TEST_BIT (hoist_vbeout
[bb
->index
], i
)
4853 && TEST_BIT (transpout
[bb
->index
], i
))
4855 /* We've found a potentially hoistable expression, now
4856 we look at every block BB dominates to see if it
4857 computes the expression. */
4858 for (j
= 0; j
< domby_len
; j
++)
4860 dominated
= domby
[j
];
4861 /* Ignore self dominance. */
4862 if (bb
== dominated
)
4864 /* We've found a dominated block, now see if it computes
4865 the busy expression and whether or not moving that
4866 expression to the "beginning" of that block is safe. */
4867 if (!TEST_BIT (antloc
[dominated
->index
], i
))
4870 /* Note if the expression would reach the dominated block
4871 unimpared if it was placed at the end of BB.
4873 Keep track of how many times this expression is hoistable
4874 from a dominated block into BB. */
4875 if (hoist_expr_reaches_here_p (bb
, i
, dominated
, NULL
))
4879 /* If we found more than one hoistable occurrence of this
4880 expression, then note it in the bitmap of expressions to
4881 hoist. It makes no sense to hoist things which are computed
4882 in only one BB, and doing so tends to pessimize register
4883 allocation. One could increase this value to try harder
4884 to avoid any possible code expansion due to register
4885 allocation issues; however experiments have shown that
4886 the vast majority of hoistable expressions are only movable
4887 from two successors, so raising this threshold is likely
4888 to nullify any benefit we get from code hoisting. */
4891 SET_BIT (hoist_exprs
[bb
->index
], i
);
4896 /* If we found nothing to hoist, then quit now. */
4903 /* Loop over all the hoistable expressions. */
4904 for (i
= 0; i
< hoist_exprs
[bb
->index
]->n_bits
; i
++)
4906 /* We want to insert the expression into BB only once, so
4907 note when we've inserted it. */
4908 insn_inserted_p
= 0;
4910 /* These tests should be the same as the tests above. */
4911 if (TEST_BIT (hoist_exprs
[bb
->index
], i
))
4913 /* We've found a potentially hoistable expression, now
4914 we look at every block BB dominates to see if it
4915 computes the expression. */
4916 for (j
= 0; j
< domby_len
; j
++)
4918 dominated
= domby
[j
];
4919 /* Ignore self dominance. */
4920 if (bb
== dominated
)
4923 /* We've found a dominated block, now see if it computes
4924 the busy expression and whether or not moving that
4925 expression to the "beginning" of that block is safe. */
4926 if (!TEST_BIT (antloc
[dominated
->index
], i
))
4929 /* The expression is computed in the dominated block and
4930 it would be safe to compute it at the start of the
4931 dominated block. Now we have to determine if the
4932 expression would reach the dominated block if it was
4933 placed at the end of BB. */
4934 if (hoist_expr_reaches_here_p (bb
, i
, dominated
, NULL
))
4936 struct expr
*expr
= index_map
[i
];
4937 struct occr
*occr
= expr
->antic_occr
;
4941 /* Find the right occurrence of this expression. */
4942 while (BLOCK_FOR_INSN (occr
->insn
) != dominated
&& occr
)
4947 set
= single_set (insn
);
4950 /* Create a pseudo-reg to store the result of reaching
4951 expressions into. Get the mode for the new pseudo
4952 from the mode of the original destination pseudo. */
4953 if (expr
->reaching_reg
== NULL
)
4955 = gen_reg_rtx (GET_MODE (SET_DEST (set
)));
4957 gcse_emit_move_after (expr
->reaching_reg
, SET_DEST (set
), insn
);
4959 occr
->deleted_p
= 1;
4960 if (!insn_inserted_p
)
4962 insert_insn_end_bb (index_map
[i
], bb
, 0);
4963 insn_inserted_p
= 1;
4975 /* Top level routine to perform one code hoisting (aka unification) pass
4977 Return nonzero if a change was made. */
4980 one_code_hoisting_pass (void)
4984 alloc_hash_table (max_cuid
, &expr_hash_table
, 0);
4985 compute_hash_table (&expr_hash_table
);
4987 dump_hash_table (gcse_file
, "Code Hosting Expressions", &expr_hash_table
);
4989 if (expr_hash_table
.n_elems
> 0)
4991 alloc_code_hoist_mem (last_basic_block
, expr_hash_table
.n_elems
);
4992 compute_code_hoist_data ();
4994 free_code_hoist_mem ();
4997 free_hash_table (&expr_hash_table
);
5002 /* Here we provide the things required to do store motion towards
5003 the exit. In order for this to be effective, gcse also needed to
5004 be taught how to move a load when it is kill only by a store to itself.
5009 void foo(float scale)
5011 for (i=0; i<10; i++)
5015 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
5016 the load out since its live around the loop, and stored at the bottom
5019 The 'Load Motion' referred to and implemented in this file is
5020 an enhancement to gcse which when using edge based lcm, recognizes
5021 this situation and allows gcse to move the load out of the loop.
5023 Once gcse has hoisted the load, store motion can then push this
5024 load towards the exit, and we end up with no loads or stores of 'i'
5028 pre_ldst_expr_hash (const void *p
)
5030 int do_not_record_p
= 0;
5031 const struct ls_expr
*x
= p
;
5032 return hash_rtx (x
->pattern
, GET_MODE (x
->pattern
), &do_not_record_p
, NULL
, false);
5036 pre_ldst_expr_eq (const void *p1
, const void *p2
)
5038 const struct ls_expr
*ptr1
= p1
, *ptr2
= p2
;
5039 return expr_equiv_p (ptr1
->pattern
, ptr2
->pattern
);
5042 /* This will search the ldst list for a matching expression. If it
5043 doesn't find one, we create one and initialize it. */
5045 static struct ls_expr
*
5048 int do_not_record_p
= 0;
5049 struct ls_expr
* ptr
;
5054 hash
= hash_rtx (x
, GET_MODE (x
), &do_not_record_p
,
5055 NULL
, /*have_reg_qty=*/false);
5058 slot
= htab_find_slot_with_hash (pre_ldst_table
, &e
, hash
, INSERT
);
5060 return (struct ls_expr
*)*slot
;
5062 ptr
= xmalloc (sizeof (struct ls_expr
));
5064 ptr
->next
= pre_ldst_mems
;
5067 ptr
->pattern_regs
= NULL_RTX
;
5068 ptr
->loads
= NULL_RTX
;
5069 ptr
->stores
= NULL_RTX
;
5070 ptr
->reaching_reg
= NULL_RTX
;
5073 ptr
->hash_index
= hash
;
5074 pre_ldst_mems
= ptr
;
5080 /* Free up an individual ldst entry. */
5083 free_ldst_entry (struct ls_expr
* ptr
)
5085 free_INSN_LIST_list (& ptr
->loads
);
5086 free_INSN_LIST_list (& ptr
->stores
);
5091 /* Free up all memory associated with the ldst list. */
5094 free_ldst_mems (void)
5097 htab_delete (pre_ldst_table
);
5098 pre_ldst_table
= NULL
;
5100 while (pre_ldst_mems
)
5102 struct ls_expr
* tmp
= pre_ldst_mems
;
5104 pre_ldst_mems
= pre_ldst_mems
->next
;
5106 free_ldst_entry (tmp
);
5109 pre_ldst_mems
= NULL
;
5112 /* Dump debugging info about the ldst list. */
5115 print_ldst_list (FILE * file
)
5117 struct ls_expr
* ptr
;
5119 fprintf (file
, "LDST list: \n");
5121 for (ptr
= first_ls_expr(); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
5123 fprintf (file
, " Pattern (%3d): ", ptr
->index
);
5125 print_rtl (file
, ptr
->pattern
);
5127 fprintf (file
, "\n Loads : ");
5130 print_rtl (file
, ptr
->loads
);
5132 fprintf (file
, "(nil)");
5134 fprintf (file
, "\n Stores : ");
5137 print_rtl (file
, ptr
->stores
);
5139 fprintf (file
, "(nil)");
5141 fprintf (file
, "\n\n");
5144 fprintf (file
, "\n");
5147 /* Returns 1 if X is in the list of ldst only expressions. */
5149 static struct ls_expr
*
5150 find_rtx_in_ldst (rtx x
)
5154 if (!pre_ldst_table
)
5157 slot
= htab_find_slot (pre_ldst_table
, &e
, NO_INSERT
);
5158 if (!slot
|| ((struct ls_expr
*)*slot
)->invalid
)
5163 /* Assign each element of the list of mems a monotonically increasing value. */
5166 enumerate_ldsts (void)
5168 struct ls_expr
* ptr
;
5171 for (ptr
= pre_ldst_mems
; ptr
!= NULL
; ptr
= ptr
->next
)
5177 /* Return first item in the list. */
5179 static inline struct ls_expr
*
5180 first_ls_expr (void)
5182 return pre_ldst_mems
;
5185 /* Return the next item in the list after the specified one. */
5187 static inline struct ls_expr
*
5188 next_ls_expr (struct ls_expr
* ptr
)
5193 /* Load Motion for loads which only kill themselves. */
5195 /* Return true if x is a simple MEM operation, with no registers or
5196 side effects. These are the types of loads we consider for the
5197 ld_motion list, otherwise we let the usual aliasing take care of it. */
5205 if (MEM_VOLATILE_P (x
))
5208 if (GET_MODE (x
) == BLKmode
)
5211 /* If we are handling exceptions, we must be careful with memory references
5212 that may trap. If we are not, the behavior is undefined, so we may just
5214 if (flag_non_call_exceptions
&& may_trap_p (x
))
5217 if (side_effects_p (x
))
5220 /* Do not consider function arguments passed on stack. */
5221 if (reg_mentioned_p (stack_pointer_rtx
, x
))
5224 if (flag_float_store
&& FLOAT_MODE_P (GET_MODE (x
)))
5230 /* Make sure there isn't a buried reference in this pattern anywhere.
5231 If there is, invalidate the entry for it since we're not capable
5232 of fixing it up just yet.. We have to be sure we know about ALL
5233 loads since the aliasing code will allow all entries in the
5234 ld_motion list to not-alias itself. If we miss a load, we will get
5235 the wrong value since gcse might common it and we won't know to
5239 invalidate_any_buried_refs (rtx x
)
5243 struct ls_expr
* ptr
;
5245 /* Invalidate it in the list. */
5246 if (MEM_P (x
) && simple_mem (x
))
5248 ptr
= ldst_entry (x
);
5252 /* Recursively process the insn. */
5253 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5255 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5258 invalidate_any_buried_refs (XEXP (x
, i
));
5259 else if (fmt
[i
] == 'E')
5260 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
5261 invalidate_any_buried_refs (XVECEXP (x
, i
, j
));
5265 /* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
5266 being defined as MEM loads and stores to symbols, with no side effects
5267 and no registers in the expression. For a MEM destination, we also
5268 check that the insn is still valid if we replace the destination with a
5269 REG, as is done in update_ld_motion_stores. If there are any uses/defs
5270 which don't match this criteria, they are invalidated and trimmed out
5274 compute_ld_motion_mems (void)
5276 struct ls_expr
* ptr
;
5280 pre_ldst_mems
= NULL
;
5281 pre_ldst_table
= htab_create (13, pre_ldst_expr_hash
,
5282 pre_ldst_expr_eq
, NULL
);
5286 FOR_BB_INSNS (bb
, insn
)
5290 if (GET_CODE (PATTERN (insn
)) == SET
)
5292 rtx src
= SET_SRC (PATTERN (insn
));
5293 rtx dest
= SET_DEST (PATTERN (insn
));
5295 /* Check for a simple LOAD... */
5296 if (MEM_P (src
) && simple_mem (src
))
5298 ptr
= ldst_entry (src
);
5300 ptr
->loads
= alloc_INSN_LIST (insn
, ptr
->loads
);
5306 /* Make sure there isn't a buried load somewhere. */
5307 invalidate_any_buried_refs (src
);
5310 /* Check for stores. Don't worry about aliased ones, they
5311 will block any movement we might do later. We only care
5312 about this exact pattern since those are the only
5313 circumstance that we will ignore the aliasing info. */
5314 if (MEM_P (dest
) && simple_mem (dest
))
5316 ptr
= ldst_entry (dest
);
5319 && GET_CODE (src
) != ASM_OPERANDS
5320 /* Check for REG manually since want_to_gcse_p
5321 returns 0 for all REGs. */
5322 && can_assign_to_reg_p (src
))
5323 ptr
->stores
= alloc_INSN_LIST (insn
, ptr
->stores
);
5329 invalidate_any_buried_refs (PATTERN (insn
));
5335 /* Remove any references that have been either invalidated or are not in the
5336 expression list for pre gcse. */
5339 trim_ld_motion_mems (void)
5341 struct ls_expr
* * last
= & pre_ldst_mems
;
5342 struct ls_expr
* ptr
= pre_ldst_mems
;
5348 /* Delete if entry has been made invalid. */
5351 /* Delete if we cannot find this mem in the expression list. */
5352 unsigned int hash
= ptr
->hash_index
% expr_hash_table
.size
;
5354 for (expr
= expr_hash_table
.table
[hash
];
5356 expr
= expr
->next_same_hash
)
5357 if (expr_equiv_p (expr
->expr
, ptr
->pattern
))
5361 expr
= (struct expr
*) 0;
5365 /* Set the expression field if we are keeping it. */
5373 htab_remove_elt_with_hash (pre_ldst_table
, ptr
, ptr
->hash_index
);
5374 free_ldst_entry (ptr
);
5379 /* Show the world what we've found. */
5380 if (gcse_file
&& pre_ldst_mems
!= NULL
)
5381 print_ldst_list (gcse_file
);
5384 /* This routine will take an expression which we are replacing with
5385 a reaching register, and update any stores that are needed if
5386 that expression is in the ld_motion list. Stores are updated by
5387 copying their SRC to the reaching register, and then storing
5388 the reaching register into the store location. These keeps the
5389 correct value in the reaching register for the loads. */
5392 update_ld_motion_stores (struct expr
* expr
)
5394 struct ls_expr
* mem_ptr
;
5396 if ((mem_ptr
= find_rtx_in_ldst (expr
->expr
)))
5398 /* We can try to find just the REACHED stores, but is shouldn't
5399 matter to set the reaching reg everywhere... some might be
5400 dead and should be eliminated later. */
5402 /* We replace (set mem expr) with (set reg expr) (set mem reg)
5403 where reg is the reaching reg used in the load. We checked in
5404 compute_ld_motion_mems that we can replace (set mem expr) with
5405 (set reg expr) in that insn. */
5406 rtx list
= mem_ptr
->stores
;
5408 for ( ; list
!= NULL_RTX
; list
= XEXP (list
, 1))
5410 rtx insn
= XEXP (list
, 0);
5411 rtx pat
= PATTERN (insn
);
5412 rtx src
= SET_SRC (pat
);
5413 rtx reg
= expr
->reaching_reg
;
5416 /* If we've already copied it, continue. */
5417 if (expr
->reaching_reg
== src
)
5422 fprintf (gcse_file
, "PRE: store updated with reaching reg ");
5423 print_rtl (gcse_file
, expr
->reaching_reg
);
5424 fprintf (gcse_file
, ":\n ");
5425 print_inline_rtx (gcse_file
, insn
, 8);
5426 fprintf (gcse_file
, "\n");
5429 copy
= gen_move_insn ( reg
, copy_rtx (SET_SRC (pat
)));
5430 new = emit_insn_before (copy
, insn
);
5431 record_one_set (REGNO (reg
), new);
5432 SET_SRC (pat
) = reg
;
5434 /* un-recognize this pattern since it's probably different now. */
5435 INSN_CODE (insn
) = -1;
5436 gcse_create_count
++;
5441 /* Store motion code. */
5443 #define ANTIC_STORE_LIST(x) ((x)->loads)
5444 #define AVAIL_STORE_LIST(x) ((x)->stores)
5445 #define LAST_AVAIL_CHECK_FAILURE(x) ((x)->reaching_reg)
5447 /* This is used to communicate the target bitvector we want to use in the
5448 reg_set_info routine when called via the note_stores mechanism. */
5449 static int * regvec
;
5451 /* And current insn, for the same routine. */
5452 static rtx compute_store_table_current_insn
;
5454 /* Used in computing the reverse edge graph bit vectors. */
5455 static sbitmap
* st_antloc
;
5457 /* Global holding the number of store expressions we are dealing with. */
5458 static int num_stores
;
5460 /* Checks to set if we need to mark a register set. Called from
5464 reg_set_info (rtx dest
, rtx setter ATTRIBUTE_UNUSED
,
5467 sbitmap bb_reg
= data
;
5469 if (GET_CODE (dest
) == SUBREG
)
5470 dest
= SUBREG_REG (dest
);
5474 regvec
[REGNO (dest
)] = INSN_UID (compute_store_table_current_insn
);
5476 SET_BIT (bb_reg
, REGNO (dest
));
5480 /* Clear any mark that says that this insn sets dest. Called from
5484 reg_clear_last_set (rtx dest
, rtx setter ATTRIBUTE_UNUSED
,
5487 int *dead_vec
= data
;
5489 if (GET_CODE (dest
) == SUBREG
)
5490 dest
= SUBREG_REG (dest
);
5493 dead_vec
[REGNO (dest
)] == INSN_UID (compute_store_table_current_insn
))
5494 dead_vec
[REGNO (dest
)] = 0;
5497 /* Return zero if some of the registers in list X are killed
5498 due to set of registers in bitmap REGS_SET. */
5501 store_ops_ok (rtx x
, int *regs_set
)
5505 for (; x
; x
= XEXP (x
, 1))
5508 if (regs_set
[REGNO(reg
)])
5515 /* Returns a list of registers mentioned in X. */
5517 extract_mentioned_regs (rtx x
)
5519 return extract_mentioned_regs_helper (x
, NULL_RTX
);
5522 /* Helper for extract_mentioned_regs; ACCUM is used to accumulate used
5525 extract_mentioned_regs_helper (rtx x
, rtx accum
)
5531 /* Repeat is used to turn tail-recursion into iteration. */
5537 code
= GET_CODE (x
);
5541 return alloc_EXPR_LIST (0, x
, accum
);
5551 /* We do not run this function with arguments having side effects. */
5570 i
= GET_RTX_LENGTH (code
) - 1;
5571 fmt
= GET_RTX_FORMAT (code
);
5577 rtx tem
= XEXP (x
, i
);
5579 /* If we are about to do the last recursive call
5580 needed at this level, change it into iteration. */
5587 accum
= extract_mentioned_regs_helper (tem
, accum
);
5589 else if (fmt
[i
] == 'E')
5593 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5594 accum
= extract_mentioned_regs_helper (XVECEXP (x
, i
, j
), accum
);
5601 /* Determine whether INSN is MEM store pattern that we will consider moving.
5602 REGS_SET_BEFORE is bitmap of registers set before (and including) the
5603 current insn, REGS_SET_AFTER is bitmap of registers set after (and
5604 including) the insn in this basic block. We must be passing through BB from
5605 head to end, as we are using this fact to speed things up.
5607 The results are stored this way:
5609 -- the first anticipatable expression is added into ANTIC_STORE_LIST
5610 -- if the processed expression is not anticipatable, NULL_RTX is added
5611 there instead, so that we can use it as indicator that no further
5612 expression of this type may be anticipatable
5613 -- if the expression is available, it is added as head of AVAIL_STORE_LIST;
5614 consequently, all of them but this head are dead and may be deleted.
5615 -- if the expression is not available, the insn due to that it fails to be
5616 available is stored in reaching_reg.
5618 The things are complicated a bit by fact that there already may be stores
5619 to the same MEM from other blocks; also caller must take care of the
5620 necessary cleanup of the temporary markers after end of the basic block.
5624 find_moveable_store (rtx insn
, int *regs_set_before
, int *regs_set_after
)
5626 struct ls_expr
* ptr
;
5628 int check_anticipatable
, check_available
;
5629 basic_block bb
= BLOCK_FOR_INSN (insn
);
5631 set
= single_set (insn
);
5635 dest
= SET_DEST (set
);
5637 if (! MEM_P (dest
) || MEM_VOLATILE_P (dest
)
5638 || GET_MODE (dest
) == BLKmode
)
5641 if (side_effects_p (dest
))
5644 /* If we are handling exceptions, we must be careful with memory references
5645 that may trap. If we are not, the behavior is undefined, so we may just
5647 if (flag_non_call_exceptions
&& may_trap_p (dest
))
5650 /* Even if the destination cannot trap, the source may. In this case we'd
5651 need to handle updating the REG_EH_REGION note. */
5652 if (find_reg_note (insn
, REG_EH_REGION
, NULL_RTX
))
5655 ptr
= ldst_entry (dest
);
5656 if (!ptr
->pattern_regs
)
5657 ptr
->pattern_regs
= extract_mentioned_regs (dest
);
5659 /* Do not check for anticipatability if we either found one anticipatable
5660 store already, or tested for one and found out that it was killed. */
5661 check_anticipatable
= 0;
5662 if (!ANTIC_STORE_LIST (ptr
))
5663 check_anticipatable
= 1;
5666 tmp
= XEXP (ANTIC_STORE_LIST (ptr
), 0);
5668 && BLOCK_FOR_INSN (tmp
) != bb
)
5669 check_anticipatable
= 1;
5671 if (check_anticipatable
)
5673 if (store_killed_before (dest
, ptr
->pattern_regs
, insn
, bb
, regs_set_before
))
5677 ANTIC_STORE_LIST (ptr
) = alloc_INSN_LIST (tmp
,
5678 ANTIC_STORE_LIST (ptr
));
5681 /* It is not necessary to check whether store is available if we did
5682 it successfully before; if we failed before, do not bother to check
5683 until we reach the insn that caused us to fail. */
5684 check_available
= 0;
5685 if (!AVAIL_STORE_LIST (ptr
))
5686 check_available
= 1;
5689 tmp
= XEXP (AVAIL_STORE_LIST (ptr
), 0);
5690 if (BLOCK_FOR_INSN (tmp
) != bb
)
5691 check_available
= 1;
5693 if (check_available
)
5695 /* Check that we have already reached the insn at that the check
5696 failed last time. */
5697 if (LAST_AVAIL_CHECK_FAILURE (ptr
))
5699 for (tmp
= BB_END (bb
);
5700 tmp
!= insn
&& tmp
!= LAST_AVAIL_CHECK_FAILURE (ptr
);
5701 tmp
= PREV_INSN (tmp
))
5704 check_available
= 0;
5707 check_available
= store_killed_after (dest
, ptr
->pattern_regs
, insn
,
5709 &LAST_AVAIL_CHECK_FAILURE (ptr
));
5711 if (!check_available
)
5712 AVAIL_STORE_LIST (ptr
) = alloc_INSN_LIST (insn
, AVAIL_STORE_LIST (ptr
));
5715 /* Find available and anticipatable stores. */
5718 compute_store_table (void)
5724 int *last_set_in
, *already_set
;
5725 struct ls_expr
* ptr
, **prev_next_ptr_ptr
;
5727 max_gcse_regno
= max_reg_num ();
5729 reg_set_in_block
= sbitmap_vector_alloc (last_basic_block
,
5731 sbitmap_vector_zero (reg_set_in_block
, last_basic_block
);
5733 pre_ldst_table
= htab_create (13, pre_ldst_expr_hash
,
5734 pre_ldst_expr_eq
, NULL
);
5735 last_set_in
= xcalloc (max_gcse_regno
, sizeof (int));
5736 already_set
= xmalloc (sizeof (int) * max_gcse_regno
);
5738 /* Find all the stores we care about. */
5741 /* First compute the registers set in this block. */
5742 regvec
= last_set_in
;
5744 FOR_BB_INSNS (bb
, insn
)
5746 if (! INSN_P (insn
))
5751 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
5752 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
5754 last_set_in
[regno
] = INSN_UID (insn
);
5755 SET_BIT (reg_set_in_block
[bb
->index
], regno
);
5759 pat
= PATTERN (insn
);
5760 compute_store_table_current_insn
= insn
;
5761 note_stores (pat
, reg_set_info
, reg_set_in_block
[bb
->index
]);
5764 /* Now find the stores. */
5765 memset (already_set
, 0, sizeof (int) * max_gcse_regno
);
5766 regvec
= already_set
;
5767 FOR_BB_INSNS (bb
, insn
)
5769 if (! INSN_P (insn
))
5774 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
5775 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
5776 already_set
[regno
] = 1;
5779 pat
= PATTERN (insn
);
5780 note_stores (pat
, reg_set_info
, NULL
);
5782 /* Now that we've marked regs, look for stores. */
5783 find_moveable_store (insn
, already_set
, last_set_in
);
5785 /* Unmark regs that are no longer set. */
5786 compute_store_table_current_insn
= insn
;
5787 note_stores (pat
, reg_clear_last_set
, last_set_in
);
5790 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
5791 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
)
5792 && last_set_in
[regno
] == INSN_UID (insn
))
5793 last_set_in
[regno
] = 0;
5797 #ifdef ENABLE_CHECKING
5798 /* last_set_in should now be all-zero. */
5799 for (regno
= 0; regno
< max_gcse_regno
; regno
++)
5800 gcc_assert (!last_set_in
[regno
]);
5803 /* Clear temporary marks. */
5804 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
5806 LAST_AVAIL_CHECK_FAILURE(ptr
) = NULL_RTX
;
5807 if (ANTIC_STORE_LIST (ptr
)
5808 && (tmp
= XEXP (ANTIC_STORE_LIST (ptr
), 0)) == NULL_RTX
)
5809 ANTIC_STORE_LIST (ptr
) = XEXP (ANTIC_STORE_LIST (ptr
), 1);
5813 /* Remove the stores that are not available anywhere, as there will
5814 be no opportunity to optimize them. */
5815 for (ptr
= pre_ldst_mems
, prev_next_ptr_ptr
= &pre_ldst_mems
;
5817 ptr
= *prev_next_ptr_ptr
)
5819 if (!AVAIL_STORE_LIST (ptr
))
5821 *prev_next_ptr_ptr
= ptr
->next
;
5822 htab_remove_elt_with_hash (pre_ldst_table
, ptr
, ptr
->hash_index
);
5823 free_ldst_entry (ptr
);
5826 prev_next_ptr_ptr
= &ptr
->next
;
5829 ret
= enumerate_ldsts ();
5833 fprintf (gcse_file
, "ST_avail and ST_antic (shown under loads..)\n");
5834 print_ldst_list (gcse_file
);
5842 /* Check to see if the load X is aliased with STORE_PATTERN.
5843 AFTER is true if we are checking the case when STORE_PATTERN occurs
5847 load_kills_store (rtx x
, rtx store_pattern
, int after
)
5850 return anti_dependence (x
, store_pattern
);
5852 return true_dependence (store_pattern
, GET_MODE (store_pattern
), x
,
5856 /* Go through the entire insn X, looking for any loads which might alias
5857 STORE_PATTERN. Return true if found.
5858 AFTER is true if we are checking the case when STORE_PATTERN occurs
5859 after the insn X. */
5862 find_loads (rtx x
, rtx store_pattern
, int after
)
5871 if (GET_CODE (x
) == SET
)
5876 if (load_kills_store (x
, store_pattern
, after
))
5880 /* Recursively process the insn. */
5881 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5883 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0 && !ret
; i
--)
5886 ret
|= find_loads (XEXP (x
, i
), store_pattern
, after
);
5887 else if (fmt
[i
] == 'E')
5888 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
5889 ret
|= find_loads (XVECEXP (x
, i
, j
), store_pattern
, after
);
5894 /* Check if INSN kills the store pattern X (is aliased with it).
5895 AFTER is true if we are checking the case when store X occurs
5896 after the insn. Return true if it does. */
5899 store_killed_in_insn (rtx x
, rtx x_regs
, rtx insn
, int after
)
5901 rtx reg
, base
, note
;
5908 /* A normal or pure call might read from pattern,
5909 but a const call will not. */
5910 if (! CONST_OR_PURE_CALL_P (insn
) || pure_call_p (insn
))
5913 /* But even a const call reads its parameters. Check whether the
5914 base of some of registers used in mem is stack pointer. */
5915 for (reg
= x_regs
; reg
; reg
= XEXP (reg
, 1))
5917 base
= find_base_term (XEXP (reg
, 0));
5919 || (GET_CODE (base
) == ADDRESS
5920 && GET_MODE (base
) == Pmode
5921 && XEXP (base
, 0) == stack_pointer_rtx
))
5928 if (GET_CODE (PATTERN (insn
)) == SET
)
5930 rtx pat
= PATTERN (insn
);
5931 rtx dest
= SET_DEST (pat
);
5933 if (GET_CODE (dest
) == ZERO_EXTRACT
)
5934 dest
= XEXP (dest
, 0);
5936 /* Check for memory stores to aliased objects. */
5938 && !expr_equiv_p (dest
, x
))
5942 if (output_dependence (dest
, x
))
5947 if (output_dependence (x
, dest
))
5951 if (find_loads (SET_SRC (pat
), x
, after
))
5954 else if (find_loads (PATTERN (insn
), x
, after
))
5957 /* If this insn has a REG_EQUAL or REG_EQUIV note referencing a memory
5958 location aliased with X, then this insn kills X. */
5959 note
= find_reg_equal_equiv_note (insn
);
5962 note
= XEXP (note
, 0);
5964 /* However, if the note represents a must alias rather than a may
5965 alias relationship, then it does not kill X. */
5966 if (expr_equiv_p (note
, x
))
5969 /* See if there are any aliased loads in the note. */
5970 return find_loads (note
, x
, after
);
5973 /* Returns true if the expression X is loaded or clobbered on or after INSN
5974 within basic block BB. REGS_SET_AFTER is bitmap of registers set in
5975 or after the insn. X_REGS is list of registers mentioned in X. If the store
5976 is killed, return the last insn in that it occurs in FAIL_INSN. */
5979 store_killed_after (rtx x
, rtx x_regs
, rtx insn
, basic_block bb
,
5980 int *regs_set_after
, rtx
*fail_insn
)
5982 rtx last
= BB_END (bb
), act
;
5984 if (!store_ops_ok (x_regs
, regs_set_after
))
5986 /* We do not know where it will happen. */
5988 *fail_insn
= NULL_RTX
;
5992 /* Scan from the end, so that fail_insn is determined correctly. */
5993 for (act
= last
; act
!= PREV_INSN (insn
); act
= PREV_INSN (act
))
5994 if (store_killed_in_insn (x
, x_regs
, act
, false))
6004 /* Returns true if the expression X is loaded or clobbered on or before INSN
6005 within basic block BB. X_REGS is list of registers mentioned in X.
6006 REGS_SET_BEFORE is bitmap of registers set before or in this insn. */
6008 store_killed_before (rtx x
, rtx x_regs
, rtx insn
, basic_block bb
,
6009 int *regs_set_before
)
6011 rtx first
= BB_HEAD (bb
);
6013 if (!store_ops_ok (x_regs
, regs_set_before
))
6016 for ( ; insn
!= PREV_INSN (first
); insn
= PREV_INSN (insn
))
6017 if (store_killed_in_insn (x
, x_regs
, insn
, true))
6023 /* Fill in available, anticipatable, transparent and kill vectors in
6024 STORE_DATA, based on lists of available and anticipatable stores. */
6026 build_store_vectors (void)
6029 int *regs_set_in_block
;
6031 struct ls_expr
* ptr
;
6034 /* Build the gen_vector. This is any store in the table which is not killed
6035 by aliasing later in its block. */
6036 ae_gen
= sbitmap_vector_alloc (last_basic_block
, num_stores
);
6037 sbitmap_vector_zero (ae_gen
, last_basic_block
);
6039 st_antloc
= sbitmap_vector_alloc (last_basic_block
, num_stores
);
6040 sbitmap_vector_zero (st_antloc
, last_basic_block
);
6042 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6044 for (st
= AVAIL_STORE_LIST (ptr
); st
!= NULL
; st
= XEXP (st
, 1))
6046 insn
= XEXP (st
, 0);
6047 bb
= BLOCK_FOR_INSN (insn
);
6049 /* If we've already seen an available expression in this block,
6050 we can delete this one (It occurs earlier in the block). We'll
6051 copy the SRC expression to an unused register in case there
6052 are any side effects. */
6053 if (TEST_BIT (ae_gen
[bb
->index
], ptr
->index
))
6055 rtx r
= gen_reg_rtx (GET_MODE (ptr
->pattern
));
6057 fprintf (gcse_file
, "Removing redundant store:\n");
6058 replace_store_insn (r
, XEXP (st
, 0), bb
, ptr
);
6061 SET_BIT (ae_gen
[bb
->index
], ptr
->index
);
6064 for (st
= ANTIC_STORE_LIST (ptr
); st
!= NULL
; st
= XEXP (st
, 1))
6066 insn
= XEXP (st
, 0);
6067 bb
= BLOCK_FOR_INSN (insn
);
6068 SET_BIT (st_antloc
[bb
->index
], ptr
->index
);
6072 ae_kill
= sbitmap_vector_alloc (last_basic_block
, num_stores
);
6073 sbitmap_vector_zero (ae_kill
, last_basic_block
);
6075 transp
= sbitmap_vector_alloc (last_basic_block
, num_stores
);
6076 sbitmap_vector_zero (transp
, last_basic_block
);
6077 regs_set_in_block
= xmalloc (sizeof (int) * max_gcse_regno
);
6081 for (regno
= 0; regno
< max_gcse_regno
; regno
++)
6082 regs_set_in_block
[regno
] = TEST_BIT (reg_set_in_block
[bb
->index
], regno
);
6084 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6086 if (store_killed_after (ptr
->pattern
, ptr
->pattern_regs
, BB_HEAD (bb
),
6087 bb
, regs_set_in_block
, NULL
))
6089 /* It should not be necessary to consider the expression
6090 killed if it is both anticipatable and available. */
6091 if (!TEST_BIT (st_antloc
[bb
->index
], ptr
->index
)
6092 || !TEST_BIT (ae_gen
[bb
->index
], ptr
->index
))
6093 SET_BIT (ae_kill
[bb
->index
], ptr
->index
);
6096 SET_BIT (transp
[bb
->index
], ptr
->index
);
6100 free (regs_set_in_block
);
6104 dump_sbitmap_vector (gcse_file
, "st_antloc", "", st_antloc
, last_basic_block
);
6105 dump_sbitmap_vector (gcse_file
, "st_kill", "", ae_kill
, last_basic_block
);
6106 dump_sbitmap_vector (gcse_file
, "Transpt", "", transp
, last_basic_block
);
6107 dump_sbitmap_vector (gcse_file
, "st_avloc", "", ae_gen
, last_basic_block
);
6111 /* Insert an instruction at the beginning of a basic block, and update
6112 the BB_HEAD if needed. */
6115 insert_insn_start_bb (rtx insn
, basic_block bb
)
6117 /* Insert at start of successor block. */
6118 rtx prev
= PREV_INSN (BB_HEAD (bb
));
6119 rtx before
= BB_HEAD (bb
);
6122 if (! LABEL_P (before
)
6123 && (! NOTE_P (before
)
6124 || NOTE_LINE_NUMBER (before
) != NOTE_INSN_BASIC_BLOCK
))
6127 if (prev
== BB_END (bb
))
6129 before
= NEXT_INSN (before
);
6132 insn
= emit_insn_after_noloc (insn
, prev
);
6136 fprintf (gcse_file
, "STORE_MOTION insert store at start of BB %d:\n",
6138 print_inline_rtx (gcse_file
, insn
, 6);
6139 fprintf (gcse_file
, "\n");
6143 /* This routine will insert a store on an edge. EXPR is the ldst entry for
6144 the memory reference, and E is the edge to insert it on. Returns nonzero
6145 if an edge insertion was performed. */
6148 insert_store (struct ls_expr
* expr
, edge e
)
6155 /* We did all the deleted before this insert, so if we didn't delete a
6156 store, then we haven't set the reaching reg yet either. */
6157 if (expr
->reaching_reg
== NULL_RTX
)
6160 if (e
->flags
& EDGE_FAKE
)
6163 reg
= expr
->reaching_reg
;
6164 insn
= gen_move_insn (copy_rtx (expr
->pattern
), reg
);
6166 /* If we are inserting this expression on ALL predecessor edges of a BB,
6167 insert it at the start of the BB, and reset the insert bits on the other
6168 edges so we don't try to insert it on the other edges. */
6170 FOR_EACH_EDGE (tmp
, ei
, e
->dest
->preds
)
6171 if (!(tmp
->flags
& EDGE_FAKE
))
6173 int index
= EDGE_INDEX (edge_list
, tmp
->src
, tmp
->dest
);
6175 gcc_assert (index
!= EDGE_INDEX_NO_EDGE
);
6176 if (! TEST_BIT (pre_insert_map
[index
], expr
->index
))
6180 /* If tmp is NULL, we found an insertion on every edge, blank the
6181 insertion vector for these edges, and insert at the start of the BB. */
6182 if (!tmp
&& bb
!= EXIT_BLOCK_PTR
)
6184 FOR_EACH_EDGE (tmp
, ei
, e
->dest
->preds
)
6186 int index
= EDGE_INDEX (edge_list
, tmp
->src
, tmp
->dest
);
6187 RESET_BIT (pre_insert_map
[index
], expr
->index
);
6189 insert_insn_start_bb (insn
, bb
);
6193 /* We can't put stores in the front of blocks pointed to by abnormal
6194 edges since that may put a store where one didn't used to be. */
6195 gcc_assert (!(e
->flags
& EDGE_ABNORMAL
));
6197 insert_insn_on_edge (insn
, e
);
6201 fprintf (gcse_file
, "STORE_MOTION insert insn on edge (%d, %d):\n",
6202 e
->src
->index
, e
->dest
->index
);
6203 print_inline_rtx (gcse_file
, insn
, 6);
6204 fprintf (gcse_file
, "\n");
6210 /* Remove any REG_EQUAL or REG_EQUIV notes containing a reference to the
6211 memory location in SMEXPR set in basic block BB.
6213 This could be rather expensive. */
6216 remove_reachable_equiv_notes (basic_block bb
, struct ls_expr
*smexpr
)
6218 edge_iterator
*stack
, ei
;
6221 sbitmap visited
= sbitmap_alloc (last_basic_block
);
6222 rtx last
, insn
, note
;
6223 rtx mem
= smexpr
->pattern
;
6225 stack
= xmalloc (sizeof (edge_iterator
) * n_basic_blocks
);
6227 ei
= ei_start (bb
->succs
);
6229 sbitmap_zero (visited
);
6231 act
= (EDGE_COUNT (ei_container (ei
)) > 0 ? EDGE_I (ei_container (ei
), 0) : NULL
);
6239 sbitmap_free (visited
);
6242 act
= ei_edge (stack
[--sp
]);
6246 if (bb
== EXIT_BLOCK_PTR
6247 || TEST_BIT (visited
, bb
->index
))
6251 act
= (! ei_end_p (ei
)) ? ei_edge (ei
) : NULL
;
6254 SET_BIT (visited
, bb
->index
);
6256 if (TEST_BIT (st_antloc
[bb
->index
], smexpr
->index
))
6258 for (last
= ANTIC_STORE_LIST (smexpr
);
6259 BLOCK_FOR_INSN (XEXP (last
, 0)) != bb
;
6260 last
= XEXP (last
, 1))
6262 last
= XEXP (last
, 0);
6265 last
= NEXT_INSN (BB_END (bb
));
6267 for (insn
= BB_HEAD (bb
); insn
!= last
; insn
= NEXT_INSN (insn
))
6270 note
= find_reg_equal_equiv_note (insn
);
6271 if (!note
|| !expr_equiv_p (XEXP (note
, 0), mem
))
6275 fprintf (gcse_file
, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
6277 remove_note (insn
, note
);
6282 act
= (! ei_end_p (ei
)) ? ei_edge (ei
) : NULL
;
6284 if (EDGE_COUNT (bb
->succs
) > 0)
6288 ei
= ei_start (bb
->succs
);
6289 act
= (EDGE_COUNT (ei_container (ei
)) > 0 ? EDGE_I (ei_container (ei
), 0) : NULL
);
6294 /* This routine will replace a store with a SET to a specified register. */
6297 replace_store_insn (rtx reg
, rtx del
, basic_block bb
, struct ls_expr
*smexpr
)
6299 rtx insn
, mem
, note
, set
, ptr
, pair
;
6301 mem
= smexpr
->pattern
;
6302 insn
= gen_move_insn (reg
, SET_SRC (single_set (del
)));
6303 insn
= emit_insn_after (insn
, del
);
6308 "STORE_MOTION delete insn in BB %d:\n ", bb
->index
);
6309 print_inline_rtx (gcse_file
, del
, 6);
6310 fprintf (gcse_file
, "\nSTORE MOTION replaced with insn:\n ");
6311 print_inline_rtx (gcse_file
, insn
, 6);
6312 fprintf (gcse_file
, "\n");
6315 for (ptr
= ANTIC_STORE_LIST (smexpr
); ptr
; ptr
= XEXP (ptr
, 1))
6316 if (XEXP (ptr
, 0) == del
)
6318 XEXP (ptr
, 0) = insn
;
6322 /* Move the notes from the deleted insn to its replacement, and patch
6323 up the LIBCALL notes. */
6324 REG_NOTES (insn
) = REG_NOTES (del
);
6326 note
= find_reg_note (insn
, REG_RETVAL
, NULL_RTX
);
6329 pair
= XEXP (note
, 0);
6330 note
= find_reg_note (pair
, REG_LIBCALL
, NULL_RTX
);
6331 XEXP (note
, 0) = insn
;
6333 note
= find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
);
6336 pair
= XEXP (note
, 0);
6337 note
= find_reg_note (pair
, REG_RETVAL
, NULL_RTX
);
6338 XEXP (note
, 0) = insn
;
6343 /* Now we must handle REG_EQUAL notes whose contents is equal to the mem;
6344 they are no longer accurate provided that they are reached by this
6345 definition, so drop them. */
6346 for (; insn
!= NEXT_INSN (BB_END (bb
)); insn
= NEXT_INSN (insn
))
6349 set
= single_set (insn
);
6352 if (expr_equiv_p (SET_DEST (set
), mem
))
6354 note
= find_reg_equal_equiv_note (insn
);
6355 if (!note
|| !expr_equiv_p (XEXP (note
, 0), mem
))
6359 fprintf (gcse_file
, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
6361 remove_note (insn
, note
);
6363 remove_reachable_equiv_notes (bb
, smexpr
);
6367 /* Delete a store, but copy the value that would have been stored into
6368 the reaching_reg for later storing. */
6371 delete_store (struct ls_expr
* expr
, basic_block bb
)
6375 if (expr
->reaching_reg
== NULL_RTX
)
6376 expr
->reaching_reg
= gen_reg_rtx (GET_MODE (expr
->pattern
));
6378 reg
= expr
->reaching_reg
;
6380 for (i
= AVAIL_STORE_LIST (expr
); i
; i
= XEXP (i
, 1))
6383 if (BLOCK_FOR_INSN (del
) == bb
)
6385 /* We know there is only one since we deleted redundant
6386 ones during the available computation. */
6387 replace_store_insn (reg
, del
, bb
, expr
);
6393 /* Free memory used by store motion. */
6396 free_store_memory (void)
6401 sbitmap_vector_free (ae_gen
);
6403 sbitmap_vector_free (ae_kill
);
6405 sbitmap_vector_free (transp
);
6407 sbitmap_vector_free (st_antloc
);
6409 sbitmap_vector_free (pre_insert_map
);
6411 sbitmap_vector_free (pre_delete_map
);
6412 if (reg_set_in_block
)
6413 sbitmap_vector_free (reg_set_in_block
);
6415 ae_gen
= ae_kill
= transp
= st_antloc
= NULL
;
6416 pre_insert_map
= pre_delete_map
= reg_set_in_block
= NULL
;
6419 /* Perform store motion. Much like gcse, except we move expressions the
6420 other way by looking at the flowgraph in reverse. */
6427 struct ls_expr
* ptr
;
6428 int update_flow
= 0;
6432 fprintf (gcse_file
, "before store motion\n");
6433 print_rtl (gcse_file
, get_insns ());
6436 init_alias_analysis ();
6438 /* Find all the available and anticipatable stores. */
6439 num_stores
= compute_store_table ();
6440 if (num_stores
== 0)
6442 htab_delete (pre_ldst_table
);
6443 pre_ldst_table
= NULL
;
6444 sbitmap_vector_free (reg_set_in_block
);
6445 end_alias_analysis ();
6449 /* Now compute kill & transp vectors. */
6450 build_store_vectors ();
6451 add_noreturn_fake_exit_edges ();
6452 connect_infinite_loops_to_exit ();
6454 edge_list
= pre_edge_rev_lcm (gcse_file
, num_stores
, transp
, ae_gen
,
6455 st_antloc
, ae_kill
, &pre_insert_map
,
6458 /* Now we want to insert the new stores which are going to be needed. */
6459 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6461 /* If any of the edges we have above are abnormal, we can't move this
6463 for (x
= NUM_EDGES (edge_list
) - 1; x
>= 0; x
--)
6464 if (TEST_BIT (pre_insert_map
[x
], ptr
->index
)
6465 && (INDEX_EDGE (edge_list
, x
)->flags
& EDGE_ABNORMAL
))
6470 if (gcse_file
!= NULL
)
6472 "Can't replace store %d: abnormal edge from %d to %d\n",
6473 ptr
->index
, INDEX_EDGE (edge_list
, x
)->src
->index
,
6474 INDEX_EDGE (edge_list
, x
)->dest
->index
);
6478 /* Now we want to insert the new stores which are going to be needed. */
6481 if (TEST_BIT (pre_delete_map
[bb
->index
], ptr
->index
))
6482 delete_store (ptr
, bb
);
6484 for (x
= 0; x
< NUM_EDGES (edge_list
); x
++)
6485 if (TEST_BIT (pre_insert_map
[x
], ptr
->index
))
6486 update_flow
|= insert_store (ptr
, INDEX_EDGE (edge_list
, x
));
6490 commit_edge_insertions ();
6492 free_store_memory ();
6493 free_edge_list (edge_list
);
6494 remove_fake_exit_edges ();
6495 end_alias_analysis ();
6499 /* Entry point for jump bypassing optimization pass. */
6502 bypass_jumps (FILE *file
)
6506 /* We do not construct an accurate cfg in functions which call
6507 setjmp, so just punt to be safe. */
6508 if (current_function_calls_setjmp
)
6511 /* For calling dump_foo fns from gdb. */
6512 debug_stderr
= stderr
;
6515 /* Identify the basic block information for this function, including
6516 successors and predecessors. */
6517 max_gcse_regno
= max_reg_num ();
6520 dump_flow_info (file
);
6522 /* Return if there's nothing to do, or it is too expensive. */
6523 if (n_basic_blocks
<= NUM_FIXED_BLOCKS
+ 1
6524 || is_too_expensive (_ ("jump bypassing disabled")))
6527 gcc_obstack_init (&gcse_obstack
);
6530 /* We need alias. */
6531 init_alias_analysis ();
6533 /* Record where pseudo-registers are set. This data is kept accurate
6534 during each pass. ??? We could also record hard-reg information here
6535 [since it's unchanging], however it is currently done during hash table
6538 It may be tempting to compute MEM set information here too, but MEM sets
6539 will be subject to code motion one day and thus we need to compute
6540 information about memory sets when we build the hash tables. */
6542 alloc_reg_set_mem (max_gcse_regno
);
6545 max_gcse_regno
= max_reg_num ();
6547 changed
= one_cprop_pass (MAX_GCSE_PASSES
+ 2, true, true);
6552 fprintf (file
, "BYPASS of %s: %d basic blocks, ",
6553 current_function_name (), n_basic_blocks
);
6554 fprintf (file
, "%d bytes\n\n", bytes_used
);
6557 obstack_free (&gcse_obstack
, NULL
);
6558 free_reg_set_mem ();
6560 /* We are finished with alias. */
6561 end_alias_analysis ();
6562 allocate_reg_info (max_reg_num (), FALSE
, FALSE
);
6567 /* Return true if the graph is too expensive to optimize. PASS is the
6568 optimization about to be performed. */
6571 is_too_expensive (const char *pass
)
6573 /* Trying to perform global optimizations on flow graphs which have
6574 a high connectivity will take a long time and is unlikely to be
6575 particularly useful.
6577 In normal circumstances a cfg should have about twice as many
6578 edges as blocks. But we do not want to punish small functions
6579 which have a couple switch statements. Rather than simply
6580 threshold the number of blocks, uses something with a more
6581 graceful degradation. */
6582 if (n_edges
> 20000 + n_basic_blocks
* 4)
6584 warning (OPT_Wdisabled_optimization
,
6585 "%s: %d basic blocks and %d edges/basic block",
6586 pass
, n_basic_blocks
, n_edges
/ n_basic_blocks
);
6591 /* If allocating memory for the cprop bitmap would take up too much
6592 storage it's better just to disable the optimization. */
6594 * SBITMAP_SET_SIZE (max_reg_num ())
6595 * sizeof (SBITMAP_ELT_TYPE
)) > MAX_GCSE_MEMORY
)
6597 warning (OPT_Wdisabled_optimization
,
6598 "%s: %d basic blocks and %d registers",
6599 pass
, n_basic_blocks
, max_reg_num ());
6608 gate_handle_jump_bypass (void)
6610 return optimize
> 0 && flag_gcse
;
6613 /* Perform jump bypassing and control flow optimizations. */
6615 rest_of_handle_jump_bypass (void)
6617 cleanup_cfg (CLEANUP_EXPENSIVE
);
6618 reg_scan (get_insns (), max_reg_num ());
6620 if (bypass_jumps (dump_file
))
6622 rebuild_jump_labels (get_insns ());
6623 cleanup_cfg (CLEANUP_EXPENSIVE
);
6624 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6628 struct tree_opt_pass pass_jump_bypass
=
6630 "bypass", /* name */
6631 gate_handle_jump_bypass
, /* gate */
6632 rest_of_handle_jump_bypass
, /* execute */
6635 0, /* static_pass_number */
6636 TV_BYPASS
, /* tv_id */
6637 0, /* properties_required */
6638 0, /* properties_provided */
6639 0, /* properties_destroyed */
6640 0, /* todo_flags_start */
6642 TODO_ggc_collect
| TODO_verify_flow
, /* todo_flags_finish */
6648 gate_handle_gcse (void)
6650 return optimize
> 0 && flag_gcse
;
6655 rest_of_handle_gcse (void)
6657 int save_csb
, save_cfj
;
6660 tem
= gcse_main (get_insns (), dump_file
);
6661 rebuild_jump_labels (get_insns ());
6662 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6664 save_csb
= flag_cse_skip_blocks
;
6665 save_cfj
= flag_cse_follow_jumps
;
6666 flag_cse_skip_blocks
= flag_cse_follow_jumps
= 0;
6668 /* If -fexpensive-optimizations, re-run CSE to clean up things done
6670 if (flag_expensive_optimizations
)
6672 timevar_push (TV_CSE
);
6673 reg_scan (get_insns (), max_reg_num ());
6674 tem2
= cse_main (get_insns (), max_reg_num (), dump_file
);
6675 purge_all_dead_edges ();
6676 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6677 timevar_pop (TV_CSE
);
6678 cse_not_expected
= !flag_rerun_cse_after_loop
;
6681 /* If gcse or cse altered any jumps, rerun jump optimizations to clean
6685 timevar_push (TV_JUMP
);
6686 rebuild_jump_labels (get_insns ());
6687 delete_dead_jumptables ();
6688 cleanup_cfg (CLEANUP_EXPENSIVE
| CLEANUP_PRE_LOOP
);
6689 timevar_pop (TV_JUMP
);
6692 flag_cse_skip_blocks
= save_csb
;
6693 flag_cse_follow_jumps
= save_cfj
;
6696 struct tree_opt_pass pass_gcse
=
6699 gate_handle_gcse
, /* gate */
6700 rest_of_handle_gcse
, /* execute */
6703 0, /* static_pass_number */
6704 TV_GCSE
, /* tv_id */
6705 0, /* properties_required */
6706 0, /* properties_provided */
6707 0, /* properties_destroyed */
6708 0, /* todo_flags_start */
6710 TODO_verify_flow
| TODO_ggc_collect
, /* todo_flags_finish */
6715 #include "gt-gcse.h"