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 2006, 2007, 2008, 2009 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 3, 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 COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
23 - reordering of memory allocation and freeing to be more space efficient
24 - do rough calc of how many regs are needed in each block, and a rough
25 calc of how many regs are available in each class and use that to
26 throttle back the code in cases where RTX_COST is minimal.
27 - a store to the same address as a load does not kill the load if the
28 source of the store is also the destination of the load. Handling this
29 allows more load motion, particularly out of loops.
30 - ability to realloc sbitmap vectors would allow one initial computation
31 of reg_set_in_block with only subsequent additions, rather than
32 recomputing it for each pass
36 /* References searched while implementing this.
38 Compilers Principles, Techniques and Tools
42 Global Optimization by Suppression of Partial Redundancies
44 communications of the acm, Vol. 22, Num. 2, Feb. 1979
46 A Portable Machine-Independent Global Optimizer - Design and Measurements
48 Stanford Ph.D. thesis, Dec. 1983
50 A Fast Algorithm for Code Movement Optimization
52 SIGPLAN Notices, Vol. 23, Num. 10, Oct. 1988
54 A Solution to a Problem with Morel and Renvoise's
55 Global Optimization by Suppression of Partial Redundancies
56 K-H Drechsler, M.P. Stadel
57 ACM TOPLAS, Vol. 10, Num. 4, Oct. 1988
59 Practical Adaptation of the Global Optimization
60 Algorithm of Morel and Renvoise
62 ACM TOPLAS, Vol. 13, Num. 2. Apr. 1991
64 Efficiently Computing Static Single Assignment Form and the Control
66 R. Cytron, J. Ferrante, B.K. Rosen, M.N. Wegman, and F.K. Zadeck
67 ACM TOPLAS, Vol. 13, Num. 4, Oct. 1991
70 J. Knoop, O. Ruthing, B. Steffen
71 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
73 What's In a Region? Or Computing Control Dependence Regions in Near-Linear
74 Time for Reducible Flow Control
76 ACM Letters on Programming Languages and Systems,
77 Vol. 2, Num. 1-4, Mar-Dec 1993
79 An Efficient Representation for Sparse Sets
80 Preston Briggs, Linda Torczon
81 ACM Letters on Programming Languages and Systems,
82 Vol. 2, Num. 1-4, Mar-Dec 1993
84 A Variation of Knoop, Ruthing, and Steffen's Lazy Code Motion
85 K-H Drechsler, M.P. Stadel
86 ACM SIGPLAN Notices, Vol. 28, Num. 5, May 1993
88 Partial Dead Code Elimination
89 J. Knoop, O. Ruthing, B. Steffen
90 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
92 Effective Partial Redundancy Elimination
93 P. Briggs, K.D. Cooper
94 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
96 The Program Structure Tree: Computing Control Regions in Linear Time
97 R. Johnson, D. Pearson, K. Pingali
98 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
100 Optimal Code Motion: Theory and Practice
101 J. Knoop, O. Ruthing, B. Steffen
102 ACM TOPLAS, Vol. 16, Num. 4, Jul. 1994
104 The power of assignment motion
105 J. Knoop, O. Ruthing, B. Steffen
106 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
108 Global code motion / global value numbering
110 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
112 Value Driven Redundancy Elimination
114 Rice University Ph.D. thesis, Apr. 1996
118 Massively Scalar Compiler Project, Rice University, Sep. 1996
120 High Performance Compilers for Parallel Computing
124 Advanced Compiler Design and Implementation
126 Morgan Kaufmann, 1997
128 Building an Optimizing Compiler
132 People wishing to speed up the code here should read:
133 Elimination Algorithms for Data Flow Analysis
134 B.G. Ryder, M.C. Paull
135 ACM Computing Surveys, Vol. 18, Num. 3, Sep. 1986
137 How to Analyze Large Programs Efficiently and Informatively
138 D.M. Dhamdhere, B.K. Rosen, F.K. Zadeck
139 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
141 People wishing to do something different can find various possibilities
142 in the above papers and elsewhere.
147 #include "coretypes.h"
155 #include "hard-reg-set.h"
158 #include "insn-config.h"
160 #include "basic-block.h"
162 #include "function.h"
171 #include "tree-pass.h"
176 /* Propagate flow information through back edges and thus enable PRE's
177 moving loop invariant calculations out of loops.
179 Originally this tended to create worse overall code, but several
180 improvements during the development of PRE seem to have made following
181 back edges generally a win.
183 Note much of the loop invariant code motion done here would normally
184 be done by loop.c, which has more heuristics for when to move invariants
185 out of loops. At some point we might need to move some of those
186 heuristics into gcse.c. */
188 /* We support GCSE via Partial Redundancy Elimination. PRE optimizations
189 are a superset of those done by GCSE.
191 We perform the following steps:
193 1) Compute table of places where registers are set.
195 2) Perform copy/constant propagation.
197 3) Perform global cse using lazy code motion if not optimizing
198 for size, or code hoisting if we are.
200 4) Perform another pass of copy/constant propagation. Try to bypass
201 conditional jumps if the condition can be computed from a value of
204 5) Perform store motion.
206 Two passes of copy/constant propagation are done because the first one
207 enables more GCSE and the second one helps to clean up the copies that
208 GCSE creates. This is needed more for PRE than for Classic because Classic
209 GCSE will try to use an existing register containing the common
210 subexpression rather than create a new one. This is harder to do for PRE
211 because of the code motion (which Classic GCSE doesn't do).
213 Expressions we are interested in GCSE-ing are of the form
214 (set (pseudo-reg) (expression)).
215 Function want_to_gcse_p says what these are.
217 In addition, expressions in REG_EQUAL notes are candidates for GXSE-ing.
218 This allows PRE to hoist expressions that are expressed in multiple insns,
219 such as comprex address calculations (e.g. for PIC code, or loads with a
220 high part and as lowe part).
222 PRE handles moving invariant expressions out of loops (by treating them as
223 partially redundant).
225 Eventually it would be nice to replace cse.c/gcse.c with SSA (static single
226 assignment) based GVN (global value numbering). L. T. Simpson's paper
227 (Rice University) on value numbering is a useful reference for this.
229 **********************
231 We used to support multiple passes but there are diminishing returns in
232 doing so. The first pass usually makes 90% of the changes that are doable.
233 A second pass can make a few more changes made possible by the first pass.
234 Experiments show any further passes don't make enough changes to justify
237 A study of spec92 using an unlimited number of passes:
238 [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
239 [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
240 [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
242 It was found doing copy propagation between each pass enables further
245 This study was done before expressions in REG_EQUAL notes were added as
246 candidate expressions for optimization, and before the GIMPLE optimizers
247 were added. Probably, multiple passes is even less efficient now than
248 at the time when the study was conducted.
250 PRE is quite expensive in complicated functions because the DFA can take
251 a while to converge. Hence we only perform one pass.
253 **********************
255 The steps for PRE are:
257 1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
259 2) Perform the data flow analysis for PRE.
261 3) Delete the redundant instructions
263 4) Insert the required copies [if any] that make the partially
264 redundant instructions fully redundant.
266 5) For other reaching expressions, insert an instruction to copy the value
267 to a newly created pseudo that will reach the redundant instruction.
269 The deletion is done first so that when we do insertions we
270 know which pseudo reg to use.
272 Various papers have argued that PRE DFA is expensive (O(n^2)) and others
273 argue it is not. The number of iterations for the algorithm to converge
274 is typically 2-4 so I don't view it as that expensive (relatively speaking).
276 PRE GCSE depends heavily on the second CSE pass to clean up the copies
277 we create. To make an expression reach the place where it's redundant,
278 the result of the expression is copied to a new register, and the redundant
279 expression is deleted by replacing it with this new register. Classic GCSE
280 doesn't have this problem as much as it computes the reaching defs of
281 each register in each block and thus can try to use an existing
284 /* GCSE global vars. */
286 /* Note whether or not we should run jump optimization after gcse. We
287 want to do this for two cases.
289 * If we changed any jumps via cprop.
291 * If we added any labels via edge splitting. */
292 static int run_jump_opt_after_gcse;
294 /* An obstack for our working variables. */
295 static struct obstack gcse_obstack;
297 struct reg_use {rtx reg_rtx; };
299 /* Hash table of expressions. */
303 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
305 /* Index in the available expression bitmaps. */
307 /* Next entry with the same hash. */
308 struct expr *next_same_hash;
309 /* List of anticipatable occurrences in basic blocks in the function.
310 An "anticipatable occurrence" is one that is the first occurrence in the
311 basic block, the operands are not modified in the basic block prior
312 to the occurrence and the output is not used between the start of
313 the block and the occurrence. */
314 struct occr *antic_occr;
315 /* List of available occurrence in basic blocks in the function.
316 An "available occurrence" is one that is the last occurrence in the
317 basic block and the operands are not modified by following statements in
318 the basic block [including this insn]. */
319 struct occr *avail_occr;
320 /* Non-null if the computation is PRE redundant.
321 The value is the newly created pseudo-reg to record a copy of the
322 expression in all the places that reach the redundant copy. */
326 /* Occurrence of an expression.
327 There is one per basic block. If a pattern appears more than once the
328 last appearance is used [or first for anticipatable expressions]. */
332 /* Next occurrence of this expression. */
334 /* The insn that computes the expression. */
336 /* Nonzero if this [anticipatable] occurrence has been deleted. */
338 /* Nonzero if this [available] occurrence has been copied to
340 /* ??? This is mutually exclusive with deleted_p, so they could share
345 /* Expression and copy propagation hash tables.
346 Each hash table is an array of buckets.
347 ??? It is known that if it were an array of entries, structure elements
348 `next_same_hash' and `bitmap_index' wouldn't be necessary. However, it is
349 not clear whether in the final analysis a sufficient amount of memory would
350 be saved as the size of the available expression bitmaps would be larger
351 [one could build a mapping table without holes afterwards though].
352 Someday I'll perform the computation and figure it out. */
357 This is an array of `expr_hash_table_size' elements. */
360 /* Size of the hash table, in elements. */
363 /* Number of hash table elements. */
364 unsigned int n_elems;
366 /* Whether the table is expression of copy propagation one. */
370 /* Expression hash table. */
371 static struct hash_table expr_hash_table;
373 /* Copy propagation hash table. */
374 static struct hash_table set_hash_table;
376 /* Mapping of uids to cuids.
377 Only real insns get cuids. */
378 static int *uid_cuid;
380 /* Highest UID in UID_CUID. */
383 /* Get the cuid of an insn. */
384 #ifdef ENABLE_CHECKING
385 #define INSN_CUID(INSN) \
386 (gcc_assert (INSN_UID (INSN) <= max_uid), uid_cuid[INSN_UID (INSN)])
388 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
391 /* Number of cuids. */
394 /* Maximum register number in function prior to doing gcse + 1.
395 Registers created during this pass have regno >= max_gcse_regno.
396 This is named with "gcse" to not collide with global of same name. */
397 static unsigned int max_gcse_regno;
399 /* Table of registers that are modified.
401 For each register, each element is a list of places where the pseudo-reg
404 For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
405 requires knowledge of which blocks kill which regs [and thus could use
406 a bitmap instead of the lists `reg_set_table' uses].
408 `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
409 num-regs) [however perhaps it may be useful to keep the data as is]. One
410 advantage of recording things this way is that `reg_set_table' is fairly
411 sparse with respect to pseudo regs but for hard regs could be fairly dense
412 [relatively speaking]. And recording sets of pseudo-regs in lists speeds
413 up functions like compute_transp since in the case of pseudo-regs we only
414 need to iterate over the number of times a pseudo-reg is set, not over the
415 number of basic blocks [clearly there is a bit of a slow down in the cases
416 where a pseudo is set more than once in a block, however it is believed
417 that the net effect is to speed things up]. This isn't done for hard-regs
418 because recording call-clobbered hard-regs in `reg_set_table' at each
419 function call can consume a fair bit of memory, and iterating over
420 hard-regs stored this way in compute_transp will be more expensive. */
422 typedef struct reg_set
424 /* The next setting of this register. */
425 struct reg_set *next;
426 /* The index of the block where it was set. */
430 static reg_set **reg_set_table;
432 /* Size of `reg_set_table'.
433 The table starts out at max_gcse_regno + slop, and is enlarged as
435 static int reg_set_table_size;
437 /* Amount to grow `reg_set_table' by when it's full. */
438 #define REG_SET_TABLE_SLOP 100
440 /* This is a list of expressions which are MEMs and will be used by load
442 Load motion tracks MEMs which aren't killed by
443 anything except itself. (i.e., loads and stores to a single location).
444 We can then allow movement of these MEM refs with a little special
445 allowance. (all stores copy the same value to the reaching reg used
446 for the loads). This means all values used to store into memory must have
447 no side effects so we can re-issue the setter value.
448 Store Motion uses this structure as an expression table to track stores
449 which look interesting, and might be moveable towards the exit block. */
453 struct expr * expr; /* Gcse expression reference for LM. */
454 rtx pattern; /* Pattern of this mem. */
455 rtx pattern_regs; /* List of registers mentioned by the mem. */
456 rtx loads; /* INSN list of loads seen. */
457 rtx stores; /* INSN list of stores seen. */
458 struct ls_expr * next; /* Next in the list. */
459 int invalid; /* Invalid for some reason. */
460 int index; /* If it maps to a bitmap index. */
461 unsigned int hash_index; /* Index when in a hash table. */
462 rtx reaching_reg; /* Register to use when re-writing. */
465 /* Array of implicit set patterns indexed by basic block index. */
466 static rtx *implicit_sets;
468 /* Head of the list of load/store memory refs. */
469 static struct ls_expr * pre_ldst_mems = NULL;
471 /* Hashtable for the load/store memory refs. */
472 static htab_t pre_ldst_table = NULL;
474 /* Bitmap containing one bit for each register in the program.
475 Used when performing GCSE to track which registers have been set since
476 the start of the basic block. */
477 static regset reg_set_bitmap;
479 /* For each block, a bitmap of registers set in the block.
480 This is used by compute_transp.
481 It is computed during hash table computation and not by compute_sets
482 as it includes registers added since the last pass (or between cprop and
483 gcse) and it's currently not easy to realloc sbitmap vectors. */
484 static sbitmap *reg_set_in_block;
486 /* Array, indexed by basic block number for a list of insns which modify
487 memory within that block. */
488 static rtx * modify_mem_list;
489 static bitmap modify_mem_list_set;
491 /* This array parallels modify_mem_list, but is kept canonicalized. */
492 static rtx * canon_modify_mem_list;
494 /* Bitmap indexed by block numbers to record which blocks contain
496 static bitmap blocks_with_calls;
498 /* Various variables for statistics gathering. */
500 /* Memory used in a pass.
501 This isn't intended to be absolutely precise. Its intent is only
502 to keep an eye on memory usage. */
503 static int bytes_used;
505 /* GCSE substitutions made. */
506 static int gcse_subst_count;
507 /* Number of copy instructions created. */
508 static int gcse_create_count;
509 /* Number of local constants propagated. */
510 static int local_const_prop_count;
511 /* Number of local copies propagated. */
512 static int local_copy_prop_count;
513 /* Number of global constants propagated. */
514 static int global_const_prop_count;
515 /* Number of global copies propagated. */
516 static int global_copy_prop_count;
518 /* For available exprs */
519 static sbitmap *ae_kill, *ae_gen;
521 static void compute_can_copy (void);
522 static void *gmalloc (size_t) ATTRIBUTE_MALLOC;
523 static void *gcalloc (size_t, size_t) ATTRIBUTE_MALLOC;
524 static void *grealloc (void *, size_t);
525 static void *gcse_alloc (unsigned long);
526 static void alloc_gcse_mem (void);
527 static void free_gcse_mem (void);
528 static void alloc_reg_set_mem (int);
529 static void free_reg_set_mem (void);
530 static void record_one_set (int, rtx);
531 static void record_set_info (rtx, const_rtx, void *);
532 static void compute_sets (void);
533 static void hash_scan_insn (rtx, struct hash_table *);
534 static void hash_scan_set (rtx, rtx, struct hash_table *);
535 static void hash_scan_clobber (rtx, rtx, struct hash_table *);
536 static void hash_scan_call (rtx, rtx, struct hash_table *);
537 static int want_to_gcse_p (rtx);
538 static bool can_assign_to_reg_p (rtx);
539 static bool gcse_constant_p (const_rtx);
540 static int oprs_unchanged_p (const_rtx, const_rtx, int);
541 static int oprs_anticipatable_p (const_rtx, const_rtx);
542 static int oprs_available_p (const_rtx, const_rtx);
543 static void insert_expr_in_table (rtx, enum machine_mode, rtx, int, int,
544 struct hash_table *);
545 static void insert_set_in_table (rtx, rtx, struct hash_table *);
546 static unsigned int hash_expr (const_rtx, enum machine_mode, int *, int);
547 static unsigned int hash_set (int, int);
548 static int expr_equiv_p (const_rtx, const_rtx);
549 static void record_last_reg_set_info (rtx, int);
550 static void record_last_mem_set_info (rtx);
551 static void record_last_set_info (rtx, const_rtx, void *);
552 static void compute_hash_table (struct hash_table *);
553 static void alloc_hash_table (int, struct hash_table *, int);
554 static void free_hash_table (struct hash_table *);
555 static void compute_hash_table_work (struct hash_table *);
556 static void dump_hash_table (FILE *, const char *, struct hash_table *);
557 static struct expr *lookup_set (unsigned int, struct hash_table *);
558 static struct expr *next_set (unsigned int, struct expr *);
559 static void reset_opr_set_tables (void);
560 static int oprs_not_set_p (const_rtx, const_rtx);
561 static void mark_call (rtx);
562 static void mark_set (rtx, rtx);
563 static void mark_clobber (rtx, rtx);
564 static void mark_oprs_set (rtx);
565 static void alloc_cprop_mem (int, int);
566 static void free_cprop_mem (void);
567 static void compute_transp (const_rtx, int, sbitmap *, int);
568 static void compute_transpout (void);
569 static void compute_local_properties (sbitmap *, sbitmap *, sbitmap *,
570 struct hash_table *);
571 static void compute_cprop_data (void);
572 static void find_used_regs (rtx *, void *);
573 static int try_replace_reg (rtx, rtx, rtx);
574 static struct expr *find_avail_set (int, rtx);
575 static int cprop_jump (basic_block, rtx, rtx, rtx, rtx);
576 static void mems_conflict_for_gcse_p (rtx, const_rtx, void *);
577 static int load_killed_in_block_p (const_basic_block, int, const_rtx, int);
578 static void canon_list_insert (rtx, const_rtx, void *);
579 static int cprop_insn (rtx, int);
580 static int cprop (int);
581 static void find_implicit_sets (void);
582 static int one_cprop_pass (int, bool, bool);
583 static bool constprop_register (rtx, rtx, rtx, bool);
584 static struct expr *find_bypass_set (int, int);
585 static bool reg_killed_on_edge (const_rtx, const_edge);
586 static int bypass_block (basic_block, rtx, rtx);
587 static int bypass_conditional_jumps (void);
588 static void alloc_pre_mem (int, int);
589 static void free_pre_mem (void);
590 static void compute_pre_data (void);
591 static int pre_expr_reaches_here_p (basic_block, struct expr *,
593 static void insert_insn_end_basic_block (struct expr *, basic_block, int);
594 static void pre_insert_copy_insn (struct expr *, rtx);
595 static void pre_insert_copies (void);
596 static int pre_delete (void);
597 static int pre_gcse (void);
598 static int one_pre_gcse_pass (int);
599 static void add_label_notes (rtx, rtx);
600 static void alloc_code_hoist_mem (int, int);
601 static void free_code_hoist_mem (void);
602 static void compute_code_hoist_vbeinout (void);
603 static void compute_code_hoist_data (void);
604 static int hoist_expr_reaches_here_p (basic_block, int, basic_block, char *);
605 static void hoist_code (void);
606 static int one_code_hoisting_pass (void);
607 static rtx process_insert_insn (struct expr *);
608 static int pre_edge_insert (struct edge_list *, struct expr **);
609 static int pre_expr_reaches_here_p_work (basic_block, struct expr *,
610 basic_block, char *);
611 static struct ls_expr * ldst_entry (rtx);
612 static void free_ldst_entry (struct ls_expr *);
613 static void free_ldst_mems (void);
614 static void print_ldst_list (FILE *);
615 static struct ls_expr * find_rtx_in_ldst (rtx);
616 static int enumerate_ldsts (void);
617 static inline struct ls_expr * first_ls_expr (void);
618 static inline struct ls_expr * next_ls_expr (struct ls_expr *);
619 static int simple_mem (const_rtx);
620 static void invalidate_any_buried_refs (rtx);
621 static void compute_ld_motion_mems (void);
622 static void trim_ld_motion_mems (void);
623 static void update_ld_motion_stores (struct expr *);
624 static void reg_set_info (rtx, const_rtx, void *);
625 static void reg_clear_last_set (rtx, const_rtx, void *);
626 static bool store_ops_ok (const_rtx, int *);
627 static rtx extract_mentioned_regs (rtx);
628 static rtx extract_mentioned_regs_helper (rtx, rtx);
629 static void find_moveable_store (rtx, int *, int *);
630 static int compute_store_table (void);
631 static bool load_kills_store (const_rtx, const_rtx, int);
632 static bool find_loads (const_rtx, const_rtx, int);
633 static bool store_killed_in_insn (const_rtx, const_rtx, const_rtx, int);
634 static bool store_killed_after (const_rtx, const_rtx, const_rtx, const_basic_block, int *, rtx *);
635 static bool store_killed_before (const_rtx, const_rtx, const_rtx, const_basic_block, int *);
636 static void build_store_vectors (void);
637 static void insert_insn_start_basic_block (rtx, basic_block);
638 static int insert_store (struct ls_expr *, edge);
639 static void remove_reachable_equiv_notes (basic_block, struct ls_expr *);
640 static void replace_store_insn (rtx, rtx, basic_block, struct ls_expr *);
641 static void delete_store (struct ls_expr *, basic_block);
642 static void free_store_memory (void);
643 static void store_motion (void);
644 static void free_insn_expr_list_list (rtx *);
645 static void clear_modify_mem_tables (void);
646 static void free_modify_mem_tables (void);
647 static rtx gcse_emit_move_after (rtx, rtx, rtx);
648 static void local_cprop_find_used_regs (rtx *, void *);
649 static bool do_local_cprop (rtx, rtx, bool);
650 static void local_cprop_pass (bool);
651 static bool is_too_expensive (const char *);
653 #define GNEW(T) ((T *) gmalloc (sizeof (T)))
654 #define GCNEW(T) ((T *) gcalloc (1, sizeof (T)))
656 #define GNEWVEC(T, N) ((T *) gmalloc (sizeof (T) * (N)))
657 #define GCNEWVEC(T, N) ((T *) gcalloc ((N), sizeof (T)))
658 #define GRESIZEVEC(T, P, N) ((T *) grealloc ((void *) (P), sizeof (T) * (N)))
660 #define GNEWVAR(T, S) ((T *) gmalloc ((S)))
661 #define GCNEWVAR(T, S) ((T *) gcalloc (1, (S)))
662 #define GRESIZEVAR(T, P, S) ((T *) grealloc ((P), (S)))
664 #define GOBNEW(T) ((T *) gcse_alloc (sizeof (T)))
665 #define GOBNEWVAR(T, S) ((T *) gcse_alloc ((S)))
668 /* Entry point for global common subexpression elimination.
669 F is the first instruction in the function. Return nonzero if a
673 gcse_main (rtx f ATTRIBUTE_UNUSED)
676 /* Point to release obstack data from for each pass. */
677 char *gcse_obstack_bottom;
679 /* We do not construct an accurate cfg in functions which call
680 setjmp, so just punt to be safe. */
681 if (cfun->calls_setjmp)
684 /* Assume that we do not need to run jump optimizations after gcse. */
685 run_jump_opt_after_gcse = 0;
687 /* Identify the basic block information for this function, including
688 successors and predecessors. */
689 max_gcse_regno = max_reg_num ();
691 df_note_add_problem ();
695 dump_flow_info (dump_file, dump_flags);
697 /* Return if there's nothing to do, or it is too expensive. */
698 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
699 || is_too_expensive (_("GCSE disabled")))
702 gcc_obstack_init (&gcse_obstack);
706 init_alias_analysis ();
708 /* Record where pseudo-registers are set. This data is kept accurate
709 during each pass. ??? We could also record hard-reg information here
710 [since it's unchanging], however it is currently done during hash table
713 It may be tempting to compute MEM set information here too, but MEM sets
714 will be subject to code motion one day and thus we need to compute
715 information about memory sets when we build the hash tables.
717 ??? Actually, we already know the information that compute_sets computes
718 because it is available from DF. FIXME. */
720 alloc_reg_set_mem (max_gcse_regno);
723 gcse_obstack_bottom = GOBNEWVAR (char, 1);
727 fprintf (dump_file, "GCSE pass\n\n");
729 max_gcse_regno = max_reg_num ();
733 /* Don't allow constant propagation to modify jumps
735 if (dbg_cnt (cprop1))
737 timevar_push (TV_CPROP1);
738 changed = one_cprop_pass (1, false, false);
739 timevar_pop (TV_CPROP1);
742 if (optimize_function_for_speed_p (cfun))
744 timevar_push (TV_PRE);
745 changed |= one_pre_gcse_pass (1);
746 /* We may have just created new basic blocks. Release and
747 recompute various things which are sized on the number of
749 ??? There would be no need for this if we used a block
750 based Lazy Code Motion variant, with all (or selected)
751 edges split before running the pass. That would also
752 help find_implicit_sets for cprop. FIXME. */
755 free_modify_mem_tables ();
756 modify_mem_list = GCNEWVEC (rtx, last_basic_block);
757 canon_modify_mem_list = GCNEWVEC (rtx, last_basic_block);
760 /* ??? When we allocate this at the start of the function,
761 the comment says that "this data is kept accurate during
762 each pass". Apparently this is not so? FIXME. */
764 alloc_reg_set_mem (max_reg_num ());
766 run_jump_opt_after_gcse = 1;
767 timevar_pop (TV_PRE);
771 /* This function is being optimized for code size.
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). */
776 timevar_push (TV_HOIST);
777 max_gcse_regno = max_reg_num ();
779 one_code_hoisting_pass ();
780 timevar_pop (TV_HOIST);
787 fprintf (dump_file, "\n");
791 obstack_free (&gcse_obstack, gcse_obstack_bottom);
793 /* Do the second const/copy propagation pass, including cprop into
794 conditional jumps. */
795 if (dbg_cnt (cprop2))
797 max_gcse_regno = max_reg_num ();
800 /* This time, go ahead and allow cprop to alter jumps. */
801 timevar_push (TV_CPROP2);
802 one_cprop_pass (2, true, true);
803 timevar_pop (TV_CPROP2);
809 fprintf (dump_file, "GCSE of %s: %d basic blocks, ",
810 current_function_name (), n_basic_blocks);
811 fprintf (dump_file, "pass 1, %d bytes\n\n", bytes_used);
814 obstack_free (&gcse_obstack, NULL);
817 /* We are finished with alias.
818 ??? Actually we recompute alias in store_motion. */
819 end_alias_analysis ();
821 /* Run store motion. */
822 if (optimize_function_for_speed_p (cfun) && flag_gcse_sm)
824 timevar_push (TV_LSM);
826 timevar_pop (TV_LSM);
829 /* Record where pseudo-registers are set. */
830 return run_jump_opt_after_gcse;
833 /* Misc. utilities. */
835 /* Nonzero for each mode that supports (set (reg) (reg)).
836 This is trivially true for integer and floating point values.
837 It may or may not be true for condition codes. */
838 static char can_copy[(int) NUM_MACHINE_MODES];
840 /* Compute which modes support reg/reg copy operations. */
843 compute_can_copy (void)
846 #ifndef AVOID_CCMODE_COPIES
849 memset (can_copy, 0, NUM_MACHINE_MODES);
852 for (i = 0; i < NUM_MACHINE_MODES; i++)
853 if (GET_MODE_CLASS (i) == MODE_CC)
855 #ifdef AVOID_CCMODE_COPIES
858 reg = gen_rtx_REG ((enum machine_mode) i, LAST_VIRTUAL_REGISTER + 1);
859 insn = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
860 if (recog (PATTERN (insn), insn, NULL) >= 0)
870 /* Returns whether the mode supports reg/reg copy operations. */
873 can_copy_p (enum machine_mode mode)
875 static bool can_copy_init_p = false;
877 if (! can_copy_init_p)
880 can_copy_init_p = true;
883 return can_copy[mode] != 0;
886 /* Cover function to xmalloc to record bytes allocated. */
889 gmalloc (size_t size)
892 return xmalloc (size);
895 /* Cover function to xcalloc to record bytes allocated. */
898 gcalloc (size_t nelem, size_t elsize)
900 bytes_used += nelem * elsize;
901 return xcalloc (nelem, elsize);
904 /* Cover function to xrealloc.
905 We don't record the additional size since we don't know it.
906 It won't affect memory usage stats much anyway. */
909 grealloc (void *ptr, size_t size)
911 return xrealloc (ptr, size);
914 /* Cover function to obstack_alloc. */
917 gcse_alloc (unsigned long size)
920 return obstack_alloc (&gcse_obstack, size);
923 /* Allocate memory for the cuid mapping array,
924 and reg/memory set tracking tables.
926 This is called at the start of each pass. */
929 alloc_gcse_mem (void)
935 /* Find the largest UID and create a mapping from UIDs to CUIDs.
936 CUIDs are like UIDs except they increase monotonically, have no gaps,
937 and only apply to real insns.
938 (Actually, there are gaps, for insn that are not inside a basic block.
939 but we should never see those anyway, so this is OK.) */
941 max_uid = get_max_uid ();
942 uid_cuid = GCNEWVEC (int, max_uid + 1);
945 FOR_BB_INSNS (bb, insn)
948 uid_cuid[INSN_UID (insn)] = i++;
950 uid_cuid[INSN_UID (insn)] = i;
955 /* Allocate vars to track sets of regs. */
956 reg_set_bitmap = BITMAP_ALLOC (NULL);
958 /* Allocate vars to track sets of regs, memory per block. */
959 reg_set_in_block = sbitmap_vector_alloc (last_basic_block, max_gcse_regno);
960 /* Allocate array to keep a list of insns which modify memory in each
962 modify_mem_list = GCNEWVEC (rtx, last_basic_block);
963 canon_modify_mem_list = GCNEWVEC (rtx, last_basic_block);
964 modify_mem_list_set = BITMAP_ALLOC (NULL);
965 blocks_with_calls = BITMAP_ALLOC (NULL);
968 /* Free memory allocated by alloc_gcse_mem. */
975 BITMAP_FREE (reg_set_bitmap);
977 sbitmap_vector_free (reg_set_in_block);
978 free_modify_mem_tables ();
979 BITMAP_FREE (modify_mem_list_set);
980 BITMAP_FREE (blocks_with_calls);
983 /* Compute the local properties of each recorded expression.
985 Local properties are those that are defined by the block, irrespective of
988 An expression is transparent in a block if its operands are not modified
991 An expression is computed (locally available) in a block if it is computed
992 at least once and expression would contain the same value if the
993 computation was moved to the end of the block.
995 An expression is locally anticipatable in a block if it is computed at
996 least once and expression would contain the same value if the computation
997 was moved to the beginning of the block.
999 We call this routine for cprop, pre and code hoisting. They all compute
1000 basically the same information and thus can easily share this code.
1002 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
1003 properties. If NULL, then it is not necessary to compute or record that
1004 particular property.
1006 TABLE controls which hash table to look at. If it is set hash table,
1007 additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
1011 compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc,
1012 struct hash_table *table)
1016 /* Initialize any bitmaps that were passed in. */
1020 sbitmap_vector_zero (transp, last_basic_block);
1022 sbitmap_vector_ones (transp, last_basic_block);
1026 sbitmap_vector_zero (comp, last_basic_block);
1028 sbitmap_vector_zero (antloc, last_basic_block);
1030 for (i = 0; i < table->size; i++)
1034 for (expr = table->table[i]; expr != NULL; expr = expr->next_same_hash)
1036 int indx = expr->bitmap_index;
1039 /* The expression is transparent in this block if it is not killed.
1040 We start by assuming all are transparent [none are killed], and
1041 then reset the bits for those that are. */
1043 compute_transp (expr->expr, indx, transp, table->set_p);
1045 /* The occurrences recorded in antic_occr are exactly those that
1046 we want to set to nonzero in ANTLOC. */
1048 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
1050 SET_BIT (antloc[BLOCK_NUM (occr->insn)], indx);
1052 /* While we're scanning the table, this is a good place to
1054 occr->deleted_p = 0;
1057 /* The occurrences recorded in avail_occr are exactly those that
1058 we want to set to nonzero in COMP. */
1060 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
1062 SET_BIT (comp[BLOCK_NUM (occr->insn)], indx);
1064 /* While we're scanning the table, this is a good place to
1069 /* While we're scanning the table, this is a good place to
1071 expr->reaching_reg = 0;
1076 /* Register set information.
1078 `reg_set_table' records where each register is set or otherwise
1081 static struct obstack reg_set_obstack;
1084 alloc_reg_set_mem (int n_regs)
1086 reg_set_table_size = n_regs + REG_SET_TABLE_SLOP;
1087 reg_set_table = GCNEWVEC (struct reg_set *, reg_set_table_size);
1089 gcc_obstack_init (®_set_obstack);
1093 free_reg_set_mem (void)
1095 free (reg_set_table);
1096 obstack_free (®_set_obstack, NULL);
1099 /* Record REGNO in the reg_set table. */
1102 record_one_set (int regno, rtx insn)
1104 /* Allocate a new reg_set element and link it onto the list. */
1105 struct reg_set *new_reg_info;
1107 /* If the table isn't big enough, enlarge it. */
1108 if (regno >= reg_set_table_size)
1110 int new_size = regno + REG_SET_TABLE_SLOP;
1112 reg_set_table = GRESIZEVEC (struct reg_set *, reg_set_table, new_size);
1113 memset (reg_set_table + reg_set_table_size, 0,
1114 (new_size - reg_set_table_size) * sizeof (struct reg_set *));
1115 reg_set_table_size = new_size;
1118 new_reg_info = XOBNEW (®_set_obstack, struct reg_set);
1119 bytes_used += sizeof (struct reg_set);
1120 new_reg_info->bb_index = BLOCK_NUM (insn);
1121 new_reg_info->next = reg_set_table[regno];
1122 reg_set_table[regno] = new_reg_info;
1125 /* Called from compute_sets via note_stores to handle one SET or CLOBBER in
1126 an insn. The DATA is really the instruction in which the SET is
1130 record_set_info (rtx dest, const_rtx setter ATTRIBUTE_UNUSED, void *data)
1132 rtx record_set_insn = (rtx) data;
1134 if (REG_P (dest) && REGNO (dest) >= FIRST_PSEUDO_REGISTER)
1135 record_one_set (REGNO (dest), record_set_insn);
1138 /* Scan the function and record each set of each pseudo-register.
1140 This is called once, at the start of the gcse pass. See the comments for
1141 `reg_set_table' for further documentation. */
1150 FOR_BB_INSNS (bb, insn)
1152 note_stores (PATTERN (insn), record_set_info, insn);
1155 /* Hash table support. */
1157 struct reg_avail_info
1159 basic_block last_bb;
1164 static struct reg_avail_info *reg_avail_info;
1165 static basic_block current_bb;
1168 /* See whether X, the source of a set, is something we want to consider for
1172 want_to_gcse_p (rtx x)
1175 /* On register stack architectures, don't GCSE constants from the
1176 constant pool, as the benefits are often swamped by the overhead
1177 of shuffling the register stack between basic blocks. */
1178 if (IS_STACK_MODE (GET_MODE (x)))
1179 x = avoid_constant_pool_reference (x);
1182 switch (GET_CODE (x))
1194 return can_assign_to_reg_p (x);
1198 /* Used internally by can_assign_to_reg_p. */
1200 static GTY(()) rtx test_insn;
1202 /* Return true if we can assign X to a pseudo register. */
1205 can_assign_to_reg_p (rtx x)
1207 int num_clobbers = 0;
1210 /* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
1211 if (general_operand (x, GET_MODE (x)))
1213 else if (GET_MODE (x) == VOIDmode)
1216 /* Otherwise, check if we can make a valid insn from it. First initialize
1217 our test insn if we haven't already. */
1221 = make_insn_raw (gen_rtx_SET (VOIDmode,
1222 gen_rtx_REG (word_mode,
1223 FIRST_PSEUDO_REGISTER * 2),
1225 NEXT_INSN (test_insn) = PREV_INSN (test_insn) = 0;
1228 /* Now make an insn like the one we would make when GCSE'ing and see if
1230 PUT_MODE (SET_DEST (PATTERN (test_insn)), GET_MODE (x));
1231 SET_SRC (PATTERN (test_insn)) = x;
1232 return ((icode = recog (PATTERN (test_insn), test_insn, &num_clobbers)) >= 0
1233 && (num_clobbers == 0 || ! added_clobbers_hard_reg_p (icode)));
1236 /* Return nonzero if the operands of expression X are unchanged from the
1237 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1238 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1241 oprs_unchanged_p (const_rtx x, const_rtx insn, int avail_p)
1250 code = GET_CODE (x);
1255 struct reg_avail_info *info = ®_avail_info[REGNO (x)];
1257 if (info->last_bb != current_bb)
1260 return info->last_set < INSN_CUID (insn);
1262 return info->first_set >= INSN_CUID (insn);
1266 if (load_killed_in_block_p (current_bb, INSN_CUID (insn),
1270 return oprs_unchanged_p (XEXP (x, 0), insn, avail_p);
1297 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1301 /* If we are about to do the last recursive call needed at this
1302 level, change it into iteration. This function is called enough
1305 return oprs_unchanged_p (XEXP (x, i), insn, avail_p);
1307 else if (! oprs_unchanged_p (XEXP (x, i), insn, avail_p))
1310 else if (fmt[i] == 'E')
1311 for (j = 0; j < XVECLEN (x, i); j++)
1312 if (! oprs_unchanged_p (XVECEXP (x, i, j), insn, avail_p))
1319 /* Used for communication between mems_conflict_for_gcse_p and
1320 load_killed_in_block_p. Nonzero if mems_conflict_for_gcse_p finds a
1321 conflict between two memory references. */
1322 static int gcse_mems_conflict_p;
1324 /* Used for communication between mems_conflict_for_gcse_p and
1325 load_killed_in_block_p. A memory reference for a load instruction,
1326 mems_conflict_for_gcse_p will see if a memory store conflicts with
1327 this memory load. */
1328 static const_rtx gcse_mem_operand;
1330 /* DEST is the output of an instruction. If it is a memory reference, and
1331 possibly conflicts with the load found in gcse_mem_operand, then set
1332 gcse_mems_conflict_p to a nonzero value. */
1335 mems_conflict_for_gcse_p (rtx dest, const_rtx setter ATTRIBUTE_UNUSED,
1336 void *data ATTRIBUTE_UNUSED)
1338 while (GET_CODE (dest) == SUBREG
1339 || GET_CODE (dest) == ZERO_EXTRACT
1340 || GET_CODE (dest) == STRICT_LOW_PART)
1341 dest = XEXP (dest, 0);
1343 /* If DEST is not a MEM, then it will not conflict with the load. Note
1344 that function calls are assumed to clobber memory, but are handled
1349 /* If we are setting a MEM in our list of specially recognized MEMs,
1350 don't mark as killed this time. */
1352 if (expr_equiv_p (dest, gcse_mem_operand) && pre_ldst_mems != NULL)
1354 if (!find_rtx_in_ldst (dest))
1355 gcse_mems_conflict_p = 1;
1359 if (true_dependence (dest, GET_MODE (dest), gcse_mem_operand,
1361 gcse_mems_conflict_p = 1;
1364 /* Return nonzero if the expression in X (a memory reference) is killed
1365 in block BB before or after the insn with the CUID in UID_LIMIT.
1366 AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills
1369 To check the entire block, set UID_LIMIT to max_uid + 1 and
1373 load_killed_in_block_p (const_basic_block bb, int uid_limit, const_rtx x, int avail_p)
1375 rtx list_entry = modify_mem_list[bb->index];
1377 /* If this is a readonly then we aren't going to be changing it. */
1378 if (MEM_READONLY_P (x))
1384 /* Ignore entries in the list that do not apply. */
1386 && INSN_CUID (XEXP (list_entry, 0)) < uid_limit)
1388 && INSN_CUID (XEXP (list_entry, 0)) > uid_limit))
1390 list_entry = XEXP (list_entry, 1);
1394 setter = XEXP (list_entry, 0);
1396 /* If SETTER is a call everything is clobbered. Note that calls
1397 to pure functions are never put on the list, so we need not
1398 worry about them. */
1399 if (CALL_P (setter))
1402 /* SETTER must be an INSN of some kind that sets memory. Call
1403 note_stores to examine each hunk of memory that is modified.
1405 The note_stores interface is pretty limited, so we have to
1406 communicate via global variables. Yuk. */
1407 gcse_mem_operand = x;
1408 gcse_mems_conflict_p = 0;
1409 note_stores (PATTERN (setter), mems_conflict_for_gcse_p, NULL);
1410 if (gcse_mems_conflict_p)
1412 list_entry = XEXP (list_entry, 1);
1417 /* Return nonzero if the operands of expression X are unchanged from
1418 the start of INSN's basic block up to but not including INSN. */
1421 oprs_anticipatable_p (const_rtx x, const_rtx insn)
1423 return oprs_unchanged_p (x, insn, 0);
1426 /* Return nonzero if the operands of expression X are unchanged from
1427 INSN to the end of INSN's basic block. */
1430 oprs_available_p (const_rtx x, const_rtx insn)
1432 return oprs_unchanged_p (x, insn, 1);
1435 /* Hash expression X.
1437 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1438 indicating if a volatile operand is found or if the expression contains
1439 something we don't want to insert in the table. HASH_TABLE_SIZE is
1440 the current size of the hash table to be probed. */
1443 hash_expr (const_rtx x, enum machine_mode mode, int *do_not_record_p,
1444 int hash_table_size)
1448 *do_not_record_p = 0;
1450 hash = hash_rtx (x, mode, do_not_record_p,
1451 NULL, /*have_reg_qty=*/false);
1452 return hash % hash_table_size;
1455 /* Hash a set of register REGNO.
1457 Sets are hashed on the register that is set. This simplifies the PRE copy
1460 ??? May need to make things more elaborate. Later, as necessary. */
1463 hash_set (int regno, int hash_table_size)
1468 return hash % hash_table_size;
1471 /* Return nonzero if exp1 is equivalent to exp2. */
1474 expr_equiv_p (const_rtx x, const_rtx y)
1476 return exp_equiv_p (x, y, 0, true);
1479 /* Insert expression X in INSN in the hash TABLE.
1480 If it is already present, record it as the last occurrence in INSN's
1483 MODE is the mode of the value X is being stored into.
1484 It is only used if X is a CONST_INT.
1486 ANTIC_P is nonzero if X is an anticipatable expression.
1487 AVAIL_P is nonzero if X is an available expression. */
1490 insert_expr_in_table (rtx x, enum machine_mode mode, rtx insn, int antic_p,
1491 int avail_p, struct hash_table *table)
1493 int found, do_not_record_p;
1495 struct expr *cur_expr, *last_expr = NULL;
1496 struct occr *antic_occr, *avail_occr;
1498 hash = hash_expr (x, mode, &do_not_record_p, table->size);
1500 /* Do not insert expression in table if it contains volatile operands,
1501 or if hash_expr determines the expression is something we don't want
1502 to or can't handle. */
1503 if (do_not_record_p)
1506 cur_expr = table->table[hash];
1509 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1511 /* If the expression isn't found, save a pointer to the end of
1513 last_expr = cur_expr;
1514 cur_expr = cur_expr->next_same_hash;
1519 cur_expr = GOBNEW (struct expr);
1520 bytes_used += sizeof (struct expr);
1521 if (table->table[hash] == NULL)
1522 /* This is the first pattern that hashed to this index. */
1523 table->table[hash] = cur_expr;
1525 /* Add EXPR to end of this hash chain. */
1526 last_expr->next_same_hash = cur_expr;
1528 /* Set the fields of the expr element. */
1530 cur_expr->bitmap_index = table->n_elems++;
1531 cur_expr->next_same_hash = NULL;
1532 cur_expr->antic_occr = NULL;
1533 cur_expr->avail_occr = NULL;
1536 /* Now record the occurrence(s). */
1539 antic_occr = cur_expr->antic_occr;
1541 if (antic_occr && BLOCK_NUM (antic_occr->insn) != BLOCK_NUM (insn))
1545 /* Found another instance of the expression in the same basic block.
1546 Prefer the currently recorded one. We want the first one in the
1547 block and the block is scanned from start to end. */
1548 ; /* nothing to do */
1551 /* First occurrence of this expression in this basic block. */
1552 antic_occr = GOBNEW (struct occr);
1553 bytes_used += sizeof (struct occr);
1554 antic_occr->insn = insn;
1555 antic_occr->next = cur_expr->antic_occr;
1556 antic_occr->deleted_p = 0;
1557 cur_expr->antic_occr = antic_occr;
1563 avail_occr = cur_expr->avail_occr;
1565 if (avail_occr && BLOCK_NUM (avail_occr->insn) == BLOCK_NUM (insn))
1567 /* Found another instance of the expression in the same basic block.
1568 Prefer this occurrence to the currently recorded one. We want
1569 the last one in the block and the block is scanned from start
1571 avail_occr->insn = insn;
1575 /* First occurrence of this expression in this basic block. */
1576 avail_occr = GOBNEW (struct occr);
1577 bytes_used += sizeof (struct occr);
1578 avail_occr->insn = insn;
1579 avail_occr->next = cur_expr->avail_occr;
1580 avail_occr->deleted_p = 0;
1581 cur_expr->avail_occr = avail_occr;
1586 /* Insert pattern X in INSN in the hash table.
1587 X is a SET of a reg to either another reg or a constant.
1588 If it is already present, record it as the last occurrence in INSN's
1592 insert_set_in_table (rtx x, rtx insn, struct hash_table *table)
1596 struct expr *cur_expr, *last_expr = NULL;
1597 struct occr *cur_occr;
1599 gcc_assert (GET_CODE (x) == SET && REG_P (SET_DEST (x)));
1601 hash = hash_set (REGNO (SET_DEST (x)), table->size);
1603 cur_expr = table->table[hash];
1606 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1608 /* If the expression isn't found, save a pointer to the end of
1610 last_expr = cur_expr;
1611 cur_expr = cur_expr->next_same_hash;
1616 cur_expr = GOBNEW (struct expr);
1617 bytes_used += sizeof (struct expr);
1618 if (table->table[hash] == NULL)
1619 /* This is the first pattern that hashed to this index. */
1620 table->table[hash] = cur_expr;
1622 /* Add EXPR to end of this hash chain. */
1623 last_expr->next_same_hash = cur_expr;
1625 /* Set the fields of the expr element.
1626 We must copy X because it can be modified when copy propagation is
1627 performed on its operands. */
1628 cur_expr->expr = copy_rtx (x);
1629 cur_expr->bitmap_index = table->n_elems++;
1630 cur_expr->next_same_hash = NULL;
1631 cur_expr->antic_occr = NULL;
1632 cur_expr->avail_occr = NULL;
1635 /* Now record the occurrence. */
1636 cur_occr = cur_expr->avail_occr;
1638 if (cur_occr && BLOCK_NUM (cur_occr->insn) == BLOCK_NUM (insn))
1640 /* Found another instance of the expression in the same basic block.
1641 Prefer this occurrence to the currently recorded one. We want
1642 the last one in the block and the block is scanned from start
1644 cur_occr->insn = insn;
1648 /* First occurrence of this expression in this basic block. */
1649 cur_occr = GOBNEW (struct occr);
1650 bytes_used += sizeof (struct occr);
1652 cur_occr->insn = insn;
1653 cur_occr->next = cur_expr->avail_occr;
1654 cur_occr->deleted_p = 0;
1655 cur_expr->avail_occr = cur_occr;
1659 /* Determine whether the rtx X should be treated as a constant for
1660 the purposes of GCSE's constant propagation. */
1663 gcse_constant_p (const_rtx x)
1665 /* Consider a COMPARE of two integers constant. */
1666 if (GET_CODE (x) == COMPARE
1667 && GET_CODE (XEXP (x, 0)) == CONST_INT
1668 && GET_CODE (XEXP (x, 1)) == CONST_INT)
1671 /* Consider a COMPARE of the same registers is a constant
1672 if they are not floating point registers. */
1673 if (GET_CODE(x) == COMPARE
1674 && REG_P (XEXP (x, 0)) && REG_P (XEXP (x, 1))
1675 && REGNO (XEXP (x, 0)) == REGNO (XEXP (x, 1))
1676 && ! FLOAT_MODE_P (GET_MODE (XEXP (x, 0)))
1677 && ! FLOAT_MODE_P (GET_MODE (XEXP (x, 1))))
1680 return CONSTANT_P (x);
1683 /* Scan pattern PAT of INSN and add an entry to the hash TABLE (set or
1687 hash_scan_set (rtx pat, rtx insn, struct hash_table *table)
1689 rtx src = SET_SRC (pat);
1690 rtx dest = SET_DEST (pat);
1693 if (GET_CODE (src) == CALL)
1694 hash_scan_call (src, insn, table);
1696 else if (REG_P (dest))
1698 unsigned int regno = REGNO (dest);
1701 /* See if a REG_EQUAL note shows this equivalent to a simpler expression.
1703 This allows us to do a single GCSE pass and still eliminate
1704 redundant constants, addresses or other expressions that are
1705 constructed with multiple instructions.
1707 However, keep the original SRC if INSN is a simple reg-reg move. In
1708 In this case, there will almost always be a REG_EQUAL note on the
1709 insn that sets SRC. By recording the REG_EQUAL value here as SRC
1710 for INSN, we miss copy propagation opportunities and we perform the
1711 same PRE GCSE operation repeatedly on the same REG_EQUAL value if we
1712 do more than one PRE GCSE pass.
1714 Note that this does not impede profitable constant propagations. We
1715 "look through" reg-reg sets in lookup_avail_set. */
1716 note = find_reg_equal_equiv_note (insn);
1718 && REG_NOTE_KIND (note) == REG_EQUAL
1721 ? gcse_constant_p (XEXP (note, 0))
1722 : want_to_gcse_p (XEXP (note, 0))))
1723 src = XEXP (note, 0), pat = gen_rtx_SET (VOIDmode, dest, src);
1725 /* Only record sets of pseudo-regs in the hash table. */
1727 && regno >= FIRST_PSEUDO_REGISTER
1728 /* Don't GCSE something if we can't do a reg/reg copy. */
1729 && can_copy_p (GET_MODE (dest))
1730 /* GCSE commonly inserts instruction after the insn. We can't
1731 do that easily for EH_REGION notes so disable GCSE on these
1733 && !find_reg_note (insn, REG_EH_REGION, NULL_RTX)
1734 /* Is SET_SRC something we want to gcse? */
1735 && want_to_gcse_p (src)
1736 /* Don't CSE a nop. */
1737 && ! set_noop_p (pat)
1738 /* Don't GCSE if it has attached REG_EQUIV note.
1739 At this point this only function parameters should have
1740 REG_EQUIV notes and if the argument slot is used somewhere
1741 explicitly, it means address of parameter has been taken,
1742 so we should not extend the lifetime of the pseudo. */
1743 && (note == NULL_RTX || ! MEM_P (XEXP (note, 0))))
1745 /* An expression is not anticipatable if its operands are
1746 modified before this insn or if this is not the only SET in
1747 this insn. The latter condition does not have to mean that
1748 SRC itself is not anticipatable, but we just will not be
1749 able to handle code motion of insns with multiple sets. */
1750 int antic_p = oprs_anticipatable_p (src, insn)
1751 && !multiple_sets (insn);
1752 /* An expression is not available if its operands are
1753 subsequently modified, including this insn. It's also not
1754 available if this is a branch, because we can't insert
1755 a set after the branch. */
1756 int avail_p = (oprs_available_p (src, insn)
1757 && ! JUMP_P (insn));
1759 insert_expr_in_table (src, GET_MODE (dest), insn, antic_p, avail_p, table);
1762 /* Record sets for constant/copy propagation. */
1763 else if (table->set_p
1764 && regno >= FIRST_PSEUDO_REGISTER
1766 && REGNO (src) >= FIRST_PSEUDO_REGISTER
1767 && can_copy_p (GET_MODE (dest))
1768 && REGNO (src) != regno)
1769 || gcse_constant_p (src))
1770 /* A copy is not available if its src or dest is subsequently
1771 modified. Here we want to search from INSN+1 on, but
1772 oprs_available_p searches from INSN on. */
1773 && (insn == BB_END (BLOCK_FOR_INSN (insn))
1774 || (tmp = next_nonnote_insn (insn)) == NULL_RTX
1775 || BLOCK_FOR_INSN (tmp) != BLOCK_FOR_INSN (insn)
1776 || oprs_available_p (pat, tmp)))
1777 insert_set_in_table (pat, insn, table);
1779 /* In case of store we want to consider the memory value as available in
1780 the REG stored in that memory. This makes it possible to remove
1781 redundant loads from due to stores to the same location. */
1782 else if (flag_gcse_las && REG_P (src) && MEM_P (dest))
1784 unsigned int regno = REGNO (src);
1786 /* Do not do this for constant/copy propagation. */
1788 /* Only record sets of pseudo-regs in the hash table. */
1789 && regno >= FIRST_PSEUDO_REGISTER
1790 /* Don't GCSE something if we can't do a reg/reg copy. */
1791 && can_copy_p (GET_MODE (src))
1792 /* GCSE commonly inserts instruction after the insn. We can't
1793 do that easily for EH_REGION notes so disable GCSE on these
1795 && ! find_reg_note (insn, REG_EH_REGION, NULL_RTX)
1796 /* Is SET_DEST something we want to gcse? */
1797 && want_to_gcse_p (dest)
1798 /* Don't CSE a nop. */
1799 && ! set_noop_p (pat)
1800 /* Don't GCSE if it has attached REG_EQUIV note.
1801 At this point this only function parameters should have
1802 REG_EQUIV notes and if the argument slot is used somewhere
1803 explicitly, it means address of parameter has been taken,
1804 so we should not extend the lifetime of the pseudo. */
1805 && ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
1806 || ! MEM_P (XEXP (note, 0))))
1808 /* Stores are never anticipatable. */
1810 /* An expression is not available if its operands are
1811 subsequently modified, including this insn. It's also not
1812 available if this is a branch, because we can't insert
1813 a set after the branch. */
1814 int avail_p = oprs_available_p (dest, insn)
1817 /* Record the memory expression (DEST) in the hash table. */
1818 insert_expr_in_table (dest, GET_MODE (dest), insn,
1819 antic_p, avail_p, table);
1825 hash_scan_clobber (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
1826 struct hash_table *table ATTRIBUTE_UNUSED)
1828 /* Currently nothing to do. */
1832 hash_scan_call (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
1833 struct hash_table *table ATTRIBUTE_UNUSED)
1835 /* Currently nothing to do. */
1838 /* Process INSN and add hash table entries as appropriate.
1840 Only available expressions that set a single pseudo-reg are recorded.
1842 Single sets in a PARALLEL could be handled, but it's an extra complication
1843 that isn't dealt with right now. The trick is handling the CLOBBERs that
1844 are also in the PARALLEL. Later.
1846 If SET_P is nonzero, this is for the assignment hash table,
1847 otherwise it is for the expression hash table. */
1850 hash_scan_insn (rtx insn, struct hash_table *table)
1852 rtx pat = PATTERN (insn);
1855 /* Pick out the sets of INSN and for other forms of instructions record
1856 what's been modified. */
1858 if (GET_CODE (pat) == SET)
1859 hash_scan_set (pat, insn, table);
1860 else if (GET_CODE (pat) == PARALLEL)
1861 for (i = 0; i < XVECLEN (pat, 0); i++)
1863 rtx x = XVECEXP (pat, 0, i);
1865 if (GET_CODE (x) == SET)
1866 hash_scan_set (x, insn, table);
1867 else if (GET_CODE (x) == CLOBBER)
1868 hash_scan_clobber (x, insn, table);
1869 else if (GET_CODE (x) == CALL)
1870 hash_scan_call (x, insn, table);
1873 else if (GET_CODE (pat) == CLOBBER)
1874 hash_scan_clobber (pat, insn, table);
1875 else if (GET_CODE (pat) == CALL)
1876 hash_scan_call (pat, insn, table);
1880 dump_hash_table (FILE *file, const char *name, struct hash_table *table)
1883 /* Flattened out table, so it's printed in proper order. */
1884 struct expr **flat_table;
1885 unsigned int *hash_val;
1888 flat_table = XCNEWVEC (struct expr *, table->n_elems);
1889 hash_val = XNEWVEC (unsigned int, table->n_elems);
1891 for (i = 0; i < (int) table->size; i++)
1892 for (expr = table->table[i]; expr != NULL; expr = expr->next_same_hash)
1894 flat_table[expr->bitmap_index] = expr;
1895 hash_val[expr->bitmap_index] = i;
1898 fprintf (file, "%s hash table (%d buckets, %d entries)\n",
1899 name, table->size, table->n_elems);
1901 for (i = 0; i < (int) table->n_elems; i++)
1902 if (flat_table[i] != 0)
1904 expr = flat_table[i];
1905 fprintf (file, "Index %d (hash value %d)\n ",
1906 expr->bitmap_index, hash_val[i]);
1907 print_rtl (file, expr->expr);
1908 fprintf (file, "\n");
1911 fprintf (file, "\n");
1917 /* Record register first/last/block set information for REGNO in INSN.
1919 first_set records the first place in the block where the register
1920 is set and is used to compute "anticipatability".
1922 last_set records the last place in the block where the register
1923 is set and is used to compute "availability".
1925 last_bb records the block for which first_set and last_set are
1926 valid, as a quick test to invalidate them.
1928 reg_set_in_block records whether the register is set in the block
1929 and is used to compute "transparency". */
1932 record_last_reg_set_info (rtx insn, int regno)
1934 struct reg_avail_info *info = ®_avail_info[regno];
1935 int cuid = INSN_CUID (insn);
1937 info->last_set = cuid;
1938 if (info->last_bb != current_bb)
1940 info->last_bb = current_bb;
1941 info->first_set = cuid;
1942 SET_BIT (reg_set_in_block[current_bb->index], regno);
1947 /* Record all of the canonicalized MEMs of record_last_mem_set_info's insn.
1948 Note we store a pair of elements in the list, so they have to be
1949 taken off pairwise. */
1952 canon_list_insert (rtx dest ATTRIBUTE_UNUSED, const_rtx unused1 ATTRIBUTE_UNUSED,
1955 rtx dest_addr, insn;
1958 while (GET_CODE (dest) == SUBREG
1959 || GET_CODE (dest) == ZERO_EXTRACT
1960 || GET_CODE (dest) == STRICT_LOW_PART)
1961 dest = XEXP (dest, 0);
1963 /* If DEST is not a MEM, then it will not conflict with a load. Note
1964 that function calls are assumed to clobber memory, but are handled
1970 dest_addr = get_addr (XEXP (dest, 0));
1971 dest_addr = canon_rtx (dest_addr);
1972 insn = (rtx) v_insn;
1973 bb = BLOCK_NUM (insn);
1975 canon_modify_mem_list[bb] =
1976 alloc_EXPR_LIST (VOIDmode, dest_addr, canon_modify_mem_list[bb]);
1977 canon_modify_mem_list[bb] =
1978 alloc_EXPR_LIST (VOIDmode, dest, canon_modify_mem_list[bb]);
1981 /* Record memory modification information for INSN. We do not actually care
1982 about the memory location(s) that are set, or even how they are set (consider
1983 a CALL_INSN). We merely need to record which insns modify memory. */
1986 record_last_mem_set_info (rtx insn)
1988 int bb = BLOCK_NUM (insn);
1990 /* load_killed_in_block_p will handle the case of calls clobbering
1992 modify_mem_list[bb] = alloc_INSN_LIST (insn, modify_mem_list[bb]);
1993 bitmap_set_bit (modify_mem_list_set, bb);
1997 /* Note that traversals of this loop (other than for free-ing)
1998 will break after encountering a CALL_INSN. So, there's no
1999 need to insert a pair of items, as canon_list_insert does. */
2000 canon_modify_mem_list[bb] =
2001 alloc_INSN_LIST (insn, canon_modify_mem_list[bb]);
2002 bitmap_set_bit (blocks_with_calls, bb);
2005 note_stores (PATTERN (insn), canon_list_insert, (void*) insn);
2008 /* Called from compute_hash_table via note_stores to handle one
2009 SET or CLOBBER in an insn. DATA is really the instruction in which
2010 the SET is taking place. */
2013 record_last_set_info (rtx dest, const_rtx setter ATTRIBUTE_UNUSED, void *data)
2015 rtx last_set_insn = (rtx) data;
2017 if (GET_CODE (dest) == SUBREG)
2018 dest = SUBREG_REG (dest);
2021 record_last_reg_set_info (last_set_insn, REGNO (dest));
2022 else if (MEM_P (dest)
2023 /* Ignore pushes, they clobber nothing. */
2024 && ! push_operand (dest, GET_MODE (dest)))
2025 record_last_mem_set_info (last_set_insn);
2028 /* Top level function to create an expression or assignment hash table.
2030 Expression entries are placed in the hash table if
2031 - they are of the form (set (pseudo-reg) src),
2032 - src is something we want to perform GCSE on,
2033 - none of the operands are subsequently modified in the block
2035 Assignment entries are placed in the hash table if
2036 - they are of the form (set (pseudo-reg) src),
2037 - src is something we want to perform const/copy propagation on,
2038 - none of the operands or target are subsequently modified in the block
2040 Currently src must be a pseudo-reg or a const_int.
2042 TABLE is the table computed. */
2045 compute_hash_table_work (struct hash_table *table)
2049 /* While we compute the hash table we also compute a bit array of which
2050 registers are set in which blocks.
2051 ??? This isn't needed during const/copy propagation, but it's cheap to
2053 sbitmap_vector_zero (reg_set_in_block, last_basic_block);
2055 /* re-Cache any INSN_LIST nodes we have allocated. */
2056 clear_modify_mem_tables ();
2057 /* Some working arrays used to track first and last set in each block. */
2058 reg_avail_info = GNEWVEC (struct reg_avail_info, max_gcse_regno);
2060 for (i = 0; i < max_gcse_regno; ++i)
2061 reg_avail_info[i].last_bb = NULL;
2063 FOR_EACH_BB (current_bb)
2068 /* First pass over the instructions records information used to
2069 determine when registers and memory are first and last set.
2070 ??? hard-reg reg_set_in_block computation
2071 could be moved to compute_sets since they currently don't change. */
2073 FOR_BB_INSNS (current_bb, insn)
2075 if (! INSN_P (insn))
2080 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
2081 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
2082 record_last_reg_set_info (insn, regno);
2087 note_stores (PATTERN (insn), record_last_set_info, insn);
2090 /* Insert implicit sets in the hash table. */
2092 && implicit_sets[current_bb->index] != NULL_RTX)
2093 hash_scan_set (implicit_sets[current_bb->index],
2094 BB_HEAD (current_bb), table);
2096 /* The next pass builds the hash table. */
2097 FOR_BB_INSNS (current_bb, insn)
2099 hash_scan_insn (insn, table);
2102 free (reg_avail_info);
2103 reg_avail_info = NULL;
2106 /* Allocate space for the set/expr hash TABLE.
2107 N_INSNS is the number of instructions in the function.
2108 It is used to determine the number of buckets to use.
2109 SET_P determines whether set or expression table will
2113 alloc_hash_table (int n_insns, struct hash_table *table, int set_p)
2117 table->size = n_insns / 4;
2118 if (table->size < 11)
2121 /* Attempt to maintain efficient use of hash table.
2122 Making it an odd number is simplest for now.
2123 ??? Later take some measurements. */
2125 n = table->size * sizeof (struct expr *);
2126 table->table = GNEWVAR (struct expr *, n);
2127 table->set_p = set_p;
2130 /* Free things allocated by alloc_hash_table. */
2133 free_hash_table (struct hash_table *table)
2135 free (table->table);
2138 /* Compute the hash TABLE for doing copy/const propagation or
2139 expression hash table. */
2142 compute_hash_table (struct hash_table *table)
2144 /* Initialize count of number of entries in hash table. */
2146 memset (table->table, 0, table->size * sizeof (struct expr *));
2148 compute_hash_table_work (table);
2151 /* Expression tracking support. */
2153 /* Lookup REGNO in the set TABLE. The result is a pointer to the
2154 table entry, or NULL if not found. */
2156 static struct expr *
2157 lookup_set (unsigned int regno, struct hash_table *table)
2159 unsigned int hash = hash_set (regno, table->size);
2162 expr = table->table[hash];
2164 while (expr && REGNO (SET_DEST (expr->expr)) != regno)
2165 expr = expr->next_same_hash;
2170 /* Return the next entry for REGNO in list EXPR. */
2172 static struct expr *
2173 next_set (unsigned int regno, struct expr *expr)
2176 expr = expr->next_same_hash;
2177 while (expr && REGNO (SET_DEST (expr->expr)) != regno);
2182 /* Like free_INSN_LIST_list or free_EXPR_LIST_list, except that the node
2183 types may be mixed. */
2186 free_insn_expr_list_list (rtx *listp)
2190 for (list = *listp; list ; list = next)
2192 next = XEXP (list, 1);
2193 if (GET_CODE (list) == EXPR_LIST)
2194 free_EXPR_LIST_node (list);
2196 free_INSN_LIST_node (list);
2202 /* Clear canon_modify_mem_list and modify_mem_list tables. */
2204 clear_modify_mem_tables (void)
2209 EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set, 0, i, bi)
2211 free_INSN_LIST_list (modify_mem_list + i);
2212 free_insn_expr_list_list (canon_modify_mem_list + i);
2214 bitmap_clear (modify_mem_list_set);
2215 bitmap_clear (blocks_with_calls);
2218 /* Release memory used by modify_mem_list_set. */
2221 free_modify_mem_tables (void)
2223 clear_modify_mem_tables ();
2224 free (modify_mem_list);
2225 free (canon_modify_mem_list);
2226 modify_mem_list = 0;
2227 canon_modify_mem_list = 0;
2230 /* Reset tables used to keep track of what's still available [since the
2231 start of the block]. */
2234 reset_opr_set_tables (void)
2236 /* Maintain a bitmap of which regs have been set since beginning of
2238 CLEAR_REG_SET (reg_set_bitmap);
2240 /* Also keep a record of the last instruction to modify memory.
2241 For now this is very trivial, we only record whether any memory
2242 location has been modified. */
2243 clear_modify_mem_tables ();
2246 /* Return nonzero if the operands of X are not set before INSN in
2247 INSN's basic block. */
2250 oprs_not_set_p (const_rtx x, const_rtx insn)
2259 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 (! RTL_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 (const_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, NULL_RTX, 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);
2548 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2552 /* If we are about to do the last recursive call
2553 needed at this level, change it into iteration.
2554 This function is called enough to be worth it. */
2561 compute_transp (XEXP (x, i), indx, bmap, set_p);
2563 else if (fmt[i] == 'E')
2564 for (j = 0; j < XVECLEN (x, i); j++)
2565 compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
2569 /* Top level routine to do the dataflow analysis needed by copy/const
2573 compute_cprop_data (void)
2575 compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, &set_hash_table);
2576 compute_available (cprop_pavloc, cprop_absaltered,
2577 cprop_avout, cprop_avin);
2580 /* Copy/constant propagation. */
2582 /* Maximum number of register uses in an insn that we handle. */
2585 /* Table of uses found in an insn.
2586 Allocated statically to avoid alloc/free complexity and overhead. */
2587 static struct reg_use reg_use_table[MAX_USES];
2589 /* Index into `reg_use_table' while building it. */
2590 static int reg_use_count;
2592 /* Set up a list of register numbers used in INSN. The found uses are stored
2593 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
2594 and contains the number of uses in the table upon exit.
2596 ??? If a register appears multiple times we will record it multiple times.
2597 This doesn't hurt anything but it will slow things down. */
2600 find_used_regs (rtx *xptr, void *data ATTRIBUTE_UNUSED)
2607 /* repeat is used to turn tail-recursion into iteration since GCC
2608 can't do it when there's no return value. */
2613 code = GET_CODE (x);
2616 if (reg_use_count == MAX_USES)
2619 reg_use_table[reg_use_count].reg_rtx = x;
2623 /* Recursively scan the operands of this expression. */
2625 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2629 /* If we are about to do the last recursive call
2630 needed at this level, change it into iteration.
2631 This function is called enough to be worth it. */
2638 find_used_regs (&XEXP (x, i), data);
2640 else if (fmt[i] == 'E')
2641 for (j = 0; j < XVECLEN (x, i); j++)
2642 find_used_regs (&XVECEXP (x, i, j), data);
2646 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
2647 Returns nonzero is successful. */
2650 try_replace_reg (rtx from, rtx to, rtx insn)
2652 rtx note = find_reg_equal_equiv_note (insn);
2655 rtx set = single_set (insn);
2657 /* Usually we substitute easy stuff, so we won't copy everything.
2658 We however need to take care to not duplicate non-trivial CONST
2662 validate_replace_src_group (from, to, insn);
2663 if (num_changes_pending () && apply_change_group ())
2666 /* Try to simplify SET_SRC if we have substituted a constant. */
2667 if (success && set && CONSTANT_P (to))
2669 src = simplify_rtx (SET_SRC (set));
2672 validate_change (insn, &SET_SRC (set), src, 0);
2675 /* If there is already a REG_EQUAL note, update the expression in it
2676 with our replacement. */
2677 if (note != 0 && REG_NOTE_KIND (note) == REG_EQUAL)
2678 set_unique_reg_note (insn, REG_EQUAL,
2679 simplify_replace_rtx (XEXP (note, 0), from,
2681 if (!success && set && reg_mentioned_p (from, SET_SRC (set)))
2683 /* If above failed and this is a single set, try to simplify the source of
2684 the set given our substitution. We could perhaps try this for multiple
2685 SETs, but it probably won't buy us anything. */
2686 src = simplify_replace_rtx (SET_SRC (set), from, to);
2688 if (!rtx_equal_p (src, SET_SRC (set))
2689 && validate_change (insn, &SET_SRC (set), src, 0))
2692 /* If we've failed to do replacement, have a single SET, don't already
2693 have a note, and have no special SET, add a REG_EQUAL note to not
2694 lose information. */
2695 if (!success && note == 0 && set != 0
2696 && GET_CODE (SET_DEST (set)) != ZERO_EXTRACT
2697 && GET_CODE (SET_DEST (set)) != STRICT_LOW_PART)
2698 note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src));
2701 /* REG_EQUAL may get simplified into register.
2702 We don't allow that. Remove that note. This code ought
2703 not to happen, because previous code ought to synthesize
2704 reg-reg move, but be on the safe side. */
2705 if (note && REG_NOTE_KIND (note) == REG_EQUAL && REG_P (XEXP (note, 0)))
2706 remove_note (insn, note);
2711 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
2712 NULL no such set is found. */
2714 static struct expr *
2715 find_avail_set (int regno, rtx insn)
2717 /* SET1 contains the last set found that can be returned to the caller for
2718 use in a substitution. */
2719 struct expr *set1 = 0;
2721 /* Loops are not possible here. To get a loop we would need two sets
2722 available at the start of the block containing INSN. i.e. we would
2723 need two sets like this available at the start of the block:
2725 (set (reg X) (reg Y))
2726 (set (reg Y) (reg X))
2728 This can not happen since the set of (reg Y) would have killed the
2729 set of (reg X) making it unavailable at the start of this block. */
2733 struct expr *set = lookup_set (regno, &set_hash_table);
2735 /* Find a set that is available at the start of the block
2736 which contains INSN. */
2739 if (TEST_BIT (cprop_avin[BLOCK_NUM (insn)], set->bitmap_index))
2741 set = next_set (regno, set);
2744 /* If no available set was found we've reached the end of the
2745 (possibly empty) copy chain. */
2749 gcc_assert (GET_CODE (set->expr) == SET);
2751 src = SET_SRC (set->expr);
2753 /* We know the set is available.
2754 Now check that SRC is ANTLOC (i.e. none of the source operands
2755 have changed since the start of the block).
2757 If the source operand changed, we may still use it for the next
2758 iteration of this loop, but we may not use it for substitutions. */
2760 if (gcse_constant_p (src) || oprs_not_set_p (src, insn))
2763 /* If the source of the set is anything except a register, then
2764 we have reached the end of the copy chain. */
2768 /* Follow the copy chain, i.e. start another iteration of the loop
2769 and see if we have an available copy into SRC. */
2770 regno = REGNO (src);
2773 /* SET1 holds the last set that was available and anticipatable at
2778 /* Subroutine of cprop_insn that tries to propagate constants into
2779 JUMP_INSNS. JUMP must be a conditional jump. If SETCC is non-NULL
2780 it is the instruction that immediately precedes JUMP, and must be a
2781 single SET of a register. FROM is what we will try to replace,
2782 SRC is the constant we will try to substitute for it. Returns nonzero
2783 if a change was made. */
2786 cprop_jump (basic_block bb, rtx setcc, rtx jump, rtx from, rtx src)
2788 rtx new_rtx, set_src, note_src;
2789 rtx set = pc_set (jump);
2790 rtx note = find_reg_equal_equiv_note (jump);
2794 note_src = XEXP (note, 0);
2795 if (GET_CODE (note_src) == EXPR_LIST)
2796 note_src = NULL_RTX;
2798 else note_src = NULL_RTX;
2800 /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
2801 set_src = note_src ? note_src : SET_SRC (set);
2803 /* First substitute the SETCC condition into the JUMP instruction,
2804 then substitute that given values into this expanded JUMP. */
2805 if (setcc != NULL_RTX
2806 && !modified_between_p (from, setcc, jump)
2807 && !modified_between_p (src, setcc, jump))
2810 rtx setcc_set = single_set (setcc);
2811 rtx setcc_note = find_reg_equal_equiv_note (setcc);
2812 setcc_src = (setcc_note && GET_CODE (XEXP (setcc_note, 0)) != EXPR_LIST)
2813 ? XEXP (setcc_note, 0) : SET_SRC (setcc_set);
2814 set_src = simplify_replace_rtx (set_src, SET_DEST (setcc_set),
2820 new_rtx = simplify_replace_rtx (set_src, from, src);
2822 /* If no simplification can be made, then try the next register. */
2823 if (rtx_equal_p (new_rtx, SET_SRC (set)))
2826 /* If this is now a no-op delete it, otherwise this must be a valid insn. */
2827 if (new_rtx == pc_rtx)
2831 /* Ensure the value computed inside the jump insn to be equivalent
2832 to one computed by setcc. */
2833 if (setcc && modified_in_p (new_rtx, setcc))
2835 if (! validate_unshare_change (jump, &SET_SRC (set), new_rtx, 0))
2837 /* When (some) constants are not valid in a comparison, and there
2838 are two registers to be replaced by constants before the entire
2839 comparison can be folded into a constant, we need to keep
2840 intermediate information in REG_EQUAL notes. For targets with
2841 separate compare insns, such notes are added by try_replace_reg.
2842 When we have a combined compare-and-branch instruction, however,
2843 we need to attach a note to the branch itself to make this
2844 optimization work. */
2846 if (!rtx_equal_p (new_rtx, note_src))
2847 set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new_rtx));
2851 /* Remove REG_EQUAL note after simplification. */
2853 remove_note (jump, note);
2857 /* Delete the cc0 setter. */
2858 if (setcc != NULL && CC0_P (SET_DEST (single_set (setcc))))
2859 delete_insn (setcc);
2862 run_jump_opt_after_gcse = 1;
2864 global_const_prop_count++;
2865 if (dump_file != NULL)
2868 "GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
2869 REGNO (from), INSN_UID (jump));
2870 print_rtl (dump_file, src);
2871 fprintf (dump_file, "\n");
2873 purge_dead_edges (bb);
2875 /* If a conditional jump has been changed into unconditional jump, remove
2876 the jump and make the edge fallthru - this is always called in
2878 if (new_rtx != pc_rtx && simplejump_p (jump))
2883 for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); ei_next (&ei))
2884 if (e->dest != EXIT_BLOCK_PTR
2885 && BB_HEAD (e->dest) == JUMP_LABEL (jump))
2887 e->flags |= EDGE_FALLTHRU;
2897 constprop_register (rtx insn, rtx from, rtx to, bool alter_jumps)
2901 /* Check for reg or cc0 setting instructions followed by
2902 conditional branch instructions first. */
2904 && (sset = single_set (insn)) != NULL
2906 && any_condjump_p (NEXT_INSN (insn)) && onlyjump_p (NEXT_INSN (insn)))
2908 rtx dest = SET_DEST (sset);
2909 if ((REG_P (dest) || CC0_P (dest))
2910 && cprop_jump (BLOCK_FOR_INSN (insn), insn, NEXT_INSN (insn), from, to))
2914 /* Handle normal insns next. */
2915 if (NONJUMP_INSN_P (insn)
2916 && try_replace_reg (from, to, insn))
2919 /* Try to propagate a CONST_INT into a conditional jump.
2920 We're pretty specific about what we will handle in this
2921 code, we can extend this as necessary over time.
2923 Right now the insn in question must look like
2924 (set (pc) (if_then_else ...)) */
2925 else if (alter_jumps && any_condjump_p (insn) && onlyjump_p (insn))
2926 return cprop_jump (BLOCK_FOR_INSN (insn), NULL, insn, from, to);
2930 /* Perform constant and copy propagation on INSN.
2931 The result is nonzero if a change was made. */
2934 cprop_insn (rtx insn, int alter_jumps)
2936 struct reg_use *reg_used;
2944 note_uses (&PATTERN (insn), find_used_regs, NULL);
2946 note = find_reg_equal_equiv_note (insn);
2948 /* We may win even when propagating constants into notes. */
2950 find_used_regs (&XEXP (note, 0), NULL);
2952 for (reg_used = ®_use_table[0]; reg_use_count > 0;
2953 reg_used++, reg_use_count--)
2955 unsigned int regno = REGNO (reg_used->reg_rtx);
2959 /* Ignore registers created by GCSE.
2960 We do this because ... */
2961 if (regno >= max_gcse_regno)
2964 /* If the register has already been set in this block, there's
2965 nothing we can do. */
2966 if (! oprs_not_set_p (reg_used->reg_rtx, insn))
2969 /* Find an assignment that sets reg_used and is available
2970 at the start of the block. */
2971 set = find_avail_set (regno, insn);
2976 /* ??? We might be able to handle PARALLELs. Later. */
2977 gcc_assert (GET_CODE (pat) == SET);
2979 src = SET_SRC (pat);
2981 /* Constant propagation. */
2982 if (gcse_constant_p (src))
2984 if (constprop_register (insn, reg_used->reg_rtx, src, alter_jumps))
2987 global_const_prop_count++;
2988 if (dump_file != NULL)
2990 fprintf (dump_file, "GLOBAL CONST-PROP: Replacing reg %d in ", regno);
2991 fprintf (dump_file, "insn %d with constant ", INSN_UID (insn));
2992 print_rtl (dump_file, src);
2993 fprintf (dump_file, "\n");
2995 if (INSN_DELETED_P (insn))
2999 else if (REG_P (src)
3000 && REGNO (src) >= FIRST_PSEUDO_REGISTER
3001 && REGNO (src) != regno)
3003 if (try_replace_reg (reg_used->reg_rtx, src, insn))
3006 global_copy_prop_count++;
3007 if (dump_file != NULL)
3009 fprintf (dump_file, "GLOBAL COPY-PROP: Replacing reg %d in insn %d",
3010 regno, INSN_UID (insn));
3011 fprintf (dump_file, " with reg %d\n", REGNO (src));
3014 /* The original insn setting reg_used may or may not now be
3015 deletable. We leave the deletion to flow. */
3016 /* FIXME: If it turns out that the insn isn't deletable,
3017 then we may have unnecessarily extended register lifetimes
3018 and made things worse. */
3026 /* Like find_used_regs, but avoid recording uses that appear in
3027 input-output contexts such as zero_extract or pre_dec. This
3028 restricts the cases we consider to those for which local cprop
3029 can legitimately make replacements. */
3032 local_cprop_find_used_regs (rtx *xptr, void *data)
3039 switch (GET_CODE (x))
3043 case STRICT_LOW_PART:
3052 /* Can only legitimately appear this early in the context of
3053 stack pushes for function arguments, but handle all of the
3054 codes nonetheless. */
3058 /* Setting a subreg of a register larger than word_mode leaves
3059 the non-written words unchanged. */
3060 if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) > BITS_PER_WORD)
3068 find_used_regs (xptr, data);
3071 /* Try to perform local const/copy propagation on X in INSN.
3072 If ALTER_JUMPS is false, changing jump insns is not allowed. */
3075 do_local_cprop (rtx x, rtx insn, bool alter_jumps)
3077 rtx newreg = NULL, newcnst = NULL;
3079 /* Rule out USE instructions and ASM statements as we don't want to
3080 change the hard registers mentioned. */
3082 && (REGNO (x) >= FIRST_PSEUDO_REGISTER
3083 || (GET_CODE (PATTERN (insn)) != USE
3084 && asm_noperands (PATTERN (insn)) < 0)))
3086 cselib_val *val = cselib_lookup (x, GET_MODE (x), 0);
3087 struct elt_loc_list *l;
3091 for (l = val->locs; l; l = l->next)
3093 rtx this_rtx = l->loc;
3096 if (gcse_constant_p (this_rtx))
3098 if (REG_P (this_rtx) && REGNO (this_rtx) >= FIRST_PSEUDO_REGISTER
3099 /* Don't copy propagate if it has attached REG_EQUIV note.
3100 At this point this only function parameters should have
3101 REG_EQUIV notes and if the argument slot is used somewhere
3102 explicitly, it means address of parameter has been taken,
3103 so we should not extend the lifetime of the pseudo. */
3104 && (!(note = find_reg_note (l->setting_insn, REG_EQUIV, NULL_RTX))
3105 || ! MEM_P (XEXP (note, 0))))
3108 if (newcnst && constprop_register (insn, x, newcnst, alter_jumps))
3110 if (dump_file != NULL)
3112 fprintf (dump_file, "LOCAL CONST-PROP: Replacing reg %d in ",
3114 fprintf (dump_file, "insn %d with constant ",
3116 print_rtl (dump_file, newcnst);
3117 fprintf (dump_file, "\n");
3119 local_const_prop_count++;
3122 else if (newreg && newreg != x && try_replace_reg (x, newreg, insn))
3124 if (dump_file != NULL)
3127 "LOCAL COPY-PROP: Replacing reg %d in insn %d",
3128 REGNO (x), INSN_UID (insn));
3129 fprintf (dump_file, " with reg %d\n", REGNO (newreg));
3131 local_copy_prop_count++;
3138 /* Do local const/copy propagation (i.e. within each basic block).
3139 If ALTER_JUMPS is true, allow propagating into jump insns, which
3140 could modify the CFG. */
3143 local_cprop_pass (bool alter_jumps)
3147 struct reg_use *reg_used;
3148 bool changed = false;
3150 cselib_init (false);
3153 FOR_BB_INSNS (bb, insn)
3157 rtx note = find_reg_equal_equiv_note (insn);
3161 note_uses (&PATTERN (insn), local_cprop_find_used_regs,
3164 local_cprop_find_used_regs (&XEXP (note, 0), NULL);
3166 for (reg_used = ®_use_table[0]; reg_use_count > 0;
3167 reg_used++, reg_use_count--)
3169 if (do_local_cprop (reg_used->reg_rtx, insn, alter_jumps))
3175 if (INSN_DELETED_P (insn))
3178 while (reg_use_count);
3180 cselib_process_insn (insn);
3183 /* Forget everything at the end of a basic block. */
3184 cselib_clear_table ();
3189 /* Global analysis may get into infinite loops for unreachable blocks. */
3190 if (changed && alter_jumps)
3192 delete_unreachable_blocks ();
3193 free_reg_set_mem ();
3194 alloc_reg_set_mem (max_reg_num ());
3199 /* Forward propagate copies. This includes copies and constants. Return
3200 nonzero if a change was made. */
3203 cprop (int alter_jumps)
3209 /* Note we start at block 1. */
3210 if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
3212 if (dump_file != NULL)
3213 fprintf (dump_file, "\n");
3218 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb, EXIT_BLOCK_PTR, next_bb)
3220 /* Reset tables used to keep track of what's still valid [since the
3221 start of the block]. */
3222 reset_opr_set_tables ();
3224 FOR_BB_INSNS (bb, insn)
3227 changed |= cprop_insn (insn, alter_jumps);
3229 /* Keep track of everything modified by this insn. */
3230 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
3231 call mark_oprs_set if we turned the insn into a NOTE. */
3232 if (! NOTE_P (insn))
3233 mark_oprs_set (insn);
3237 if (dump_file != NULL)
3238 fprintf (dump_file, "\n");
3243 /* Similar to get_condition, only the resulting condition must be
3244 valid at JUMP, instead of at EARLIEST.
3246 This differs from noce_get_condition in ifcvt.c in that we prefer not to
3247 settle for the condition variable in the jump instruction being integral.
3248 We prefer to be able to record the value of a user variable, rather than
3249 the value of a temporary used in a condition. This could be solved by
3250 recording the value of *every* register scanned by canonicalize_condition,
3251 but this would require some code reorganization. */
3254 fis_get_condition (rtx jump)
3256 return get_condition (jump, NULL, false, true);
3259 /* Check the comparison COND to see if we can safely form an implicit set from
3260 it. COND is either an EQ or NE comparison. */
3263 implicit_set_cond_p (const_rtx cond)
3265 const enum machine_mode mode = GET_MODE (XEXP (cond, 0));
3266 const_rtx cst = XEXP (cond, 1);
3268 /* We can't perform this optimization if either operand might be or might
3269 contain a signed zero. */
3270 if (HONOR_SIGNED_ZEROS (mode))
3272 /* It is sufficient to check if CST is or contains a zero. We must
3273 handle float, complex, and vector. If any subpart is a zero, then
3274 the optimization can't be performed. */
3275 /* ??? The complex and vector checks are not implemented yet. We just
3276 always return zero for them. */
3277 if (GET_CODE (cst) == CONST_DOUBLE)
3280 REAL_VALUE_FROM_CONST_DOUBLE (d, cst);
3281 if (REAL_VALUES_EQUAL (d, dconst0))
3288 return gcse_constant_p (cst);
3291 /* Find the implicit sets of a function. An "implicit set" is a constraint
3292 on the value of a variable, implied by a conditional jump. For example,
3293 following "if (x == 2)", the then branch may be optimized as though the
3294 conditional performed an "explicit set", in this example, "x = 2". This
3295 function records the set patterns that are implicit at the start of each
3299 find_implicit_sets (void)
3301 basic_block bb, dest;
3307 /* Check for more than one successor. */
3308 if (EDGE_COUNT (bb->succs) > 1)
3310 cond = fis_get_condition (BB_END (bb));
3313 && (GET_CODE (cond) == EQ || GET_CODE (cond) == NE)
3314 && REG_P (XEXP (cond, 0))
3315 && REGNO (XEXP (cond, 0)) >= FIRST_PSEUDO_REGISTER
3316 && implicit_set_cond_p (cond))
3318 dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
3319 : FALLTHRU_EDGE (bb)->dest;
3321 if (dest && single_pred_p (dest)
3322 && dest != EXIT_BLOCK_PTR)
3324 new_rtx = gen_rtx_SET (VOIDmode, XEXP (cond, 0),
3326 implicit_sets[dest->index] = new_rtx;
3329 fprintf(dump_file, "Implicit set of reg %d in ",
3330 REGNO (XEXP (cond, 0)));
3331 fprintf(dump_file, "basic block %d\n", dest->index);
3339 fprintf (dump_file, "Found %d implicit sets\n", count);
3342 /* Perform one copy/constant propagation pass.
3343 PASS is the pass count. If CPROP_JUMPS is true, perform constant
3344 propagation into conditional jumps. If BYPASS_JUMPS is true,
3345 perform conditional jump bypassing optimizations. */
3348 one_cprop_pass (int pass, bool cprop_jumps, bool bypass_jumps)
3352 global_const_prop_count = local_const_prop_count = 0;
3353 global_copy_prop_count = local_copy_prop_count = 0;
3356 local_cprop_pass (cprop_jumps);
3358 /* Determine implicit sets. */
3359 implicit_sets = XCNEWVEC (rtx, last_basic_block);
3360 find_implicit_sets ();
3362 alloc_hash_table (max_cuid, &set_hash_table, 1);
3363 compute_hash_table (&set_hash_table);
3365 /* Free implicit_sets before peak usage. */
3366 free (implicit_sets);
3367 implicit_sets = NULL;
3370 dump_hash_table (dump_file, "SET", &set_hash_table);
3371 if (set_hash_table.n_elems > 0)
3373 alloc_cprop_mem (last_basic_block, set_hash_table.n_elems);
3374 compute_cprop_data ();
3375 changed = cprop (cprop_jumps);
3377 changed |= bypass_conditional_jumps ();
3381 free_hash_table (&set_hash_table);
3385 fprintf (dump_file, "CPROP of %s, pass %d: %d bytes needed, ",
3386 current_function_name (), pass, bytes_used);
3387 fprintf (dump_file, "%d local const props, %d local copy props, ",
3388 local_const_prop_count, local_copy_prop_count);
3389 fprintf (dump_file, "%d global const props, %d global copy props\n\n",
3390 global_const_prop_count, global_copy_prop_count);
3392 /* Global analysis may get into infinite loops for unreachable blocks. */
3393 if (changed && cprop_jumps)
3394 delete_unreachable_blocks ();
3399 /* Bypass conditional jumps. */
3401 /* The value of last_basic_block at the beginning of the jump_bypass
3402 pass. The use of redirect_edge_and_branch_force may introduce new
3403 basic blocks, but the data flow analysis is only valid for basic
3404 block indices less than bypass_last_basic_block. */
3406 static int bypass_last_basic_block;
3408 /* Find a set of REGNO to a constant that is available at the end of basic
3409 block BB. Returns NULL if no such set is found. Based heavily upon
3412 static struct expr *
3413 find_bypass_set (int regno, int bb)
3415 struct expr *result = 0;
3420 struct expr *set = lookup_set (regno, &set_hash_table);
3424 if (TEST_BIT (cprop_avout[bb], set->bitmap_index))
3426 set = next_set (regno, set);
3432 gcc_assert (GET_CODE (set->expr) == SET);
3434 src = SET_SRC (set->expr);
3435 if (gcse_constant_p (src))
3441 regno = REGNO (src);
3447 /* Subroutine of bypass_block that checks whether a pseudo is killed by
3448 any of the instructions inserted on an edge. Jump bypassing places
3449 condition code setters on CFG edges using insert_insn_on_edge. This
3450 function is required to check that our data flow analysis is still
3451 valid prior to commit_edge_insertions. */
3454 reg_killed_on_edge (const_rtx reg, const_edge e)
3458 for (insn = e->insns.r; insn; insn = NEXT_INSN (insn))
3459 if (INSN_P (insn) && reg_set_p (reg, insn))
3465 /* Subroutine of bypass_conditional_jumps that attempts to bypass the given
3466 basic block BB which has more than one predecessor. If not NULL, SETCC
3467 is the first instruction of BB, which is immediately followed by JUMP_INSN
3468 JUMP. Otherwise, SETCC is NULL, and JUMP is the first insn of BB.
3469 Returns nonzero if a change was made.
3471 During the jump bypassing pass, we may place copies of SETCC instructions
3472 on CFG edges. The following routine must be careful to pay attention to
3473 these inserted insns when performing its transformations. */
3476 bypass_block (basic_block bb, rtx setcc, rtx jump)
3481 int may_be_loop_header;
3485 insn = (setcc != NULL) ? setcc : jump;
3487 /* Determine set of register uses in INSN. */
3489 note_uses (&PATTERN (insn), find_used_regs, NULL);
3490 note = find_reg_equal_equiv_note (insn);
3492 find_used_regs (&XEXP (note, 0), NULL);
3494 may_be_loop_header = false;
3495 FOR_EACH_EDGE (e, ei, bb->preds)
3496 if (e->flags & EDGE_DFS_BACK)
3498 may_be_loop_header = true;
3503 for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); )
3507 if (e->flags & EDGE_COMPLEX)
3513 /* We can't redirect edges from new basic blocks. */
3514 if (e->src->index >= bypass_last_basic_block)
3520 /* The irreducible loops created by redirecting of edges entering the
3521 loop from outside would decrease effectiveness of some of the following
3522 optimizations, so prevent this. */
3523 if (may_be_loop_header
3524 && !(e->flags & EDGE_DFS_BACK))
3530 for (i = 0; i < reg_use_count; i++)
3532 struct reg_use *reg_used = ®_use_table[i];
3533 unsigned int regno = REGNO (reg_used->reg_rtx);
3534 basic_block dest, old_dest;
3538 if (regno >= max_gcse_regno)
3541 set = find_bypass_set (regno, e->src->index);
3546 /* Check the data flow is valid after edge insertions. */
3547 if (e->insns.r && reg_killed_on_edge (reg_used->reg_rtx, e))
3550 src = SET_SRC (pc_set (jump));
3553 src = simplify_replace_rtx (src,
3554 SET_DEST (PATTERN (setcc)),
3555 SET_SRC (PATTERN (setcc)));
3557 new_rtx = simplify_replace_rtx (src, reg_used->reg_rtx,
3558 SET_SRC (set->expr));
3560 /* Jump bypassing may have already placed instructions on
3561 edges of the CFG. We can't bypass an outgoing edge that
3562 has instructions associated with it, as these insns won't
3563 get executed if the incoming edge is redirected. */
3565 if (new_rtx == pc_rtx)
3567 edest = FALLTHRU_EDGE (bb);
3568 dest = edest->insns.r ? NULL : edest->dest;
3570 else if (GET_CODE (new_rtx) == LABEL_REF)
3572 dest = BLOCK_FOR_INSN (XEXP (new_rtx, 0));
3573 /* Don't bypass edges containing instructions. */
3574 edest = find_edge (bb, dest);
3575 if (edest && edest->insns.r)
3581 /* Avoid unification of the edge with other edges from original
3582 branch. We would end up emitting the instruction on "both"
3585 if (dest && setcc && !CC0_P (SET_DEST (PATTERN (setcc)))
3586 && find_edge (e->src, dest))
3592 && dest != EXIT_BLOCK_PTR)
3594 redirect_edge_and_branch_force (e, dest);
3596 /* Copy the register setter to the redirected edge.
3597 Don't copy CC0 setters, as CC0 is dead after jump. */
3600 rtx pat = PATTERN (setcc);
3601 if (!CC0_P (SET_DEST (pat)))
3602 insert_insn_on_edge (copy_insn (pat), e);
3605 if (dump_file != NULL)
3607 fprintf (dump_file, "JUMP-BYPASS: Proved reg %d "
3608 "in jump_insn %d equals constant ",
3609 regno, INSN_UID (jump));
3610 print_rtl (dump_file, SET_SRC (set->expr));
3611 fprintf (dump_file, "\nBypass edge from %d->%d to %d\n",
3612 e->src->index, old_dest->index, dest->index);
3625 /* Find basic blocks with more than one predecessor that only contain a
3626 single conditional jump. If the result of the comparison is known at
3627 compile-time from any incoming edge, redirect that edge to the
3628 appropriate target. Returns nonzero if a change was made.
3630 This function is now mis-named, because we also handle indirect jumps. */
3633 bypass_conditional_jumps (void)
3641 /* Note we start at block 1. */
3642 if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
3645 bypass_last_basic_block = last_basic_block;
3646 mark_dfs_back_edges ();
3649 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb,
3650 EXIT_BLOCK_PTR, next_bb)
3652 /* Check for more than one predecessor. */
3653 if (!single_pred_p (bb))
3656 FOR_BB_INSNS (bb, insn)
3657 if (NONJUMP_INSN_P (insn))
3661 if (GET_CODE (PATTERN (insn)) != SET)
3664 dest = SET_DEST (PATTERN (insn));
3665 if (REG_P (dest) || CC0_P (dest))
3670 else if (JUMP_P (insn))
3672 if ((any_condjump_p (insn) || computed_jump_p (insn))
3673 && onlyjump_p (insn))
3674 changed |= bypass_block (bb, setcc, insn);
3677 else if (INSN_P (insn))
3682 /* If we bypassed any register setting insns, we inserted a
3683 copy on the redirected edge. These need to be committed. */
3685 commit_edge_insertions ();
3690 /* Compute PRE+LCM working variables. */
3692 /* Local properties of expressions. */
3693 /* Nonzero for expressions that are transparent in the block. */
3694 static sbitmap *transp;
3696 /* Nonzero for expressions that are transparent at the end of the block.
3697 This is only zero for expressions killed by abnormal critical edge
3698 created by a calls. */
3699 static sbitmap *transpout;
3701 /* Nonzero for expressions that are computed (available) in the block. */
3702 static sbitmap *comp;
3704 /* Nonzero for expressions that are locally anticipatable in the block. */
3705 static sbitmap *antloc;
3707 /* Nonzero for expressions where this block is an optimal computation
3709 static sbitmap *pre_optimal;
3711 /* Nonzero for expressions which are redundant in a particular block. */
3712 static sbitmap *pre_redundant;
3714 /* Nonzero for expressions which should be inserted on a specific edge. */
3715 static sbitmap *pre_insert_map;
3717 /* Nonzero for expressions which should be deleted in a specific block. */
3718 static sbitmap *pre_delete_map;
3720 /* Contains the edge_list returned by pre_edge_lcm. */
3721 static struct edge_list *edge_list;
3723 /* Redundant insns. */
3724 static sbitmap pre_redundant_insns;
3726 /* Allocate vars used for PRE analysis. */
3729 alloc_pre_mem (int n_blocks, int n_exprs)
3731 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
3732 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
3733 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
3736 pre_redundant = NULL;
3737 pre_insert_map = NULL;
3738 pre_delete_map = NULL;
3739 ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
3741 /* pre_insert and pre_delete are allocated later. */
3744 /* Free vars used for PRE analysis. */
3749 sbitmap_vector_free (transp);
3750 sbitmap_vector_free (comp);
3752 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
3755 sbitmap_vector_free (pre_optimal);
3757 sbitmap_vector_free (pre_redundant);
3759 sbitmap_vector_free (pre_insert_map);
3761 sbitmap_vector_free (pre_delete_map);
3763 transp = comp = NULL;
3764 pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
3767 /* Top level routine to do the dataflow analysis needed by PRE. */
3770 compute_pre_data (void)
3772 sbitmap trapping_expr;
3776 compute_local_properties (transp, comp, antloc, &expr_hash_table);
3777 sbitmap_vector_zero (ae_kill, last_basic_block);
3779 /* Collect expressions which might trap. */
3780 trapping_expr = sbitmap_alloc (expr_hash_table.n_elems);
3781 sbitmap_zero (trapping_expr);
3782 for (ui = 0; ui < expr_hash_table.size; ui++)
3785 for (e = expr_hash_table.table[ui]; e != NULL; e = e->next_same_hash)
3786 if (may_trap_p (e->expr))
3787 SET_BIT (trapping_expr, e->bitmap_index);
3790 /* Compute ae_kill for each basic block using:
3800 /* If the current block is the destination of an abnormal edge, we
3801 kill all trapping expressions because we won't be able to properly
3802 place the instruction on the edge. So make them neither
3803 anticipatable nor transparent. This is fairly conservative. */
3804 FOR_EACH_EDGE (e, ei, bb->preds)
3805 if (e->flags & EDGE_ABNORMAL)
3807 sbitmap_difference (antloc[bb->index], antloc[bb->index], trapping_expr);
3808 sbitmap_difference (transp[bb->index], transp[bb->index], trapping_expr);
3812 sbitmap_a_or_b (ae_kill[bb->index], transp[bb->index], comp[bb->index]);
3813 sbitmap_not (ae_kill[bb->index], ae_kill[bb->index]);
3816 edge_list = pre_edge_lcm (expr_hash_table.n_elems, transp, comp, antloc,
3817 ae_kill, &pre_insert_map, &pre_delete_map);
3818 sbitmap_vector_free (antloc);
3820 sbitmap_vector_free (ae_kill);
3822 sbitmap_free (trapping_expr);
3827 /* Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach
3830 VISITED is a pointer to a working buffer for tracking which BB's have
3831 been visited. It is NULL for the top-level call.
3833 We treat reaching expressions that go through blocks containing the same
3834 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
3835 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
3836 2 as not reaching. The intent is to improve the probability of finding
3837 only one reaching expression and to reduce register lifetimes by picking
3838 the closest such expression. */
3841 pre_expr_reaches_here_p_work (basic_block occr_bb, struct expr *expr, basic_block bb, char *visited)
3846 FOR_EACH_EDGE (pred, ei, bb->preds)
3848 basic_block pred_bb = pred->src;
3850 if (pred->src == ENTRY_BLOCK_PTR
3851 /* Has predecessor has already been visited? */
3852 || visited[pred_bb->index])
3853 ;/* Nothing to do. */
3855 /* Does this predecessor generate this expression? */
3856 else if (TEST_BIT (comp[pred_bb->index], expr->bitmap_index))
3858 /* Is this the occurrence we're looking for?
3859 Note that there's only one generating occurrence per block
3860 so we just need to check the block number. */
3861 if (occr_bb == pred_bb)
3864 visited[pred_bb->index] = 1;
3866 /* Ignore this predecessor if it kills the expression. */
3867 else if (! TEST_BIT (transp[pred_bb->index], expr->bitmap_index))
3868 visited[pred_bb->index] = 1;
3870 /* Neither gen nor kill. */
3873 visited[pred_bb->index] = 1;
3874 if (pre_expr_reaches_here_p_work (occr_bb, expr, pred_bb, visited))
3879 /* All paths have been checked. */
3883 /* The wrapper for pre_expr_reaches_here_work that ensures that any
3884 memory allocated for that function is returned. */
3887 pre_expr_reaches_here_p (basic_block occr_bb, struct expr *expr, basic_block bb)
3890 char *visited = XCNEWVEC (char, last_basic_block);
3892 rval = pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited);
3899 /* Given an expr, generate RTL which we can insert at the end of a BB,
3900 or on an edge. Set the block number of any insns generated to
3904 process_insert_insn (struct expr *expr)
3906 rtx reg = expr->reaching_reg;
3907 rtx exp = copy_rtx (expr->expr);
3912 /* If the expression is something that's an operand, like a constant,
3913 just copy it to a register. */
3914 if (general_operand (exp, GET_MODE (reg)))
3915 emit_move_insn (reg, exp);
3917 /* Otherwise, make a new insn to compute this expression and make sure the
3918 insn will be recognized (this also adds any needed CLOBBERs). Copy the
3919 expression to make sure we don't have any sharing issues. */
3922 rtx insn = emit_insn (gen_rtx_SET (VOIDmode, reg, exp));
3924 if (insn_invalid_p (insn))
3935 /* Add EXPR to the end of basic block BB.
3937 This is used by both the PRE and code hoisting.
3939 For PRE, we want to verify that the expr is either transparent
3940 or locally anticipatable in the target block. This check makes
3941 no sense for code hoisting. */
3944 insert_insn_end_basic_block (struct expr *expr, basic_block bb, int pre)
3946 rtx insn = BB_END (bb);
3948 rtx reg = expr->reaching_reg;
3949 int regno = REGNO (reg);
3952 pat = process_insert_insn (expr);
3953 gcc_assert (pat && INSN_P (pat));
3956 while (NEXT_INSN (pat_end) != NULL_RTX)
3957 pat_end = NEXT_INSN (pat_end);
3959 /* If the last insn is a jump, insert EXPR in front [taking care to
3960 handle cc0, etc. properly]. Similarly we need to care trapping
3961 instructions in presence of non-call exceptions. */
3964 || (NONJUMP_INSN_P (insn)
3965 && (!single_succ_p (bb)
3966 || single_succ_edge (bb)->flags & EDGE_ABNORMAL)))
3971 /* It should always be the case that we can put these instructions
3972 anywhere in the basic block with performing PRE optimizations.
3974 gcc_assert (!NONJUMP_INSN_P (insn) || !pre
3975 || TEST_BIT (antloc[bb->index], expr->bitmap_index)
3976 || TEST_BIT (transp[bb->index], expr->bitmap_index));
3978 /* If this is a jump table, then we can't insert stuff here. Since
3979 we know the previous real insn must be the tablejump, we insert
3980 the new instruction just before the tablejump. */
3981 if (GET_CODE (PATTERN (insn)) == ADDR_VEC
3982 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
3983 insn = prev_real_insn (insn);
3986 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
3987 if cc0 isn't set. */
3988 note = find_reg_note (insn, REG_CC_SETTER, NULL_RTX);
3990 insn = XEXP (note, 0);
3993 rtx maybe_cc0_setter = prev_nonnote_insn (insn);
3994 if (maybe_cc0_setter
3995 && INSN_P (maybe_cc0_setter)
3996 && sets_cc0_p (PATTERN (maybe_cc0_setter)))
3997 insn = maybe_cc0_setter;
4000 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4001 new_insn = emit_insn_before_noloc (pat, insn, bb);
4004 /* Likewise if the last insn is a call, as will happen in the presence
4005 of exception handling. */
4006 else if (CALL_P (insn)
4007 && (!single_succ_p (bb)
4008 || single_succ_edge (bb)->flags & EDGE_ABNORMAL))
4010 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4011 we search backward and place the instructions before the first
4012 parameter is loaded. Do this for everyone for consistency and a
4013 presumption that we'll get better code elsewhere as well.
4015 It should always be the case that we can put these instructions
4016 anywhere in the basic block with performing PRE optimizations.
4020 || TEST_BIT (antloc[bb->index], expr->bitmap_index)
4021 || TEST_BIT (transp[bb->index], expr->bitmap_index));
4023 /* Since different machines initialize their parameter registers
4024 in different orders, assume nothing. Collect the set of all
4025 parameter registers. */
4026 insn = find_first_parameter_load (insn, BB_HEAD (bb));
4028 /* If we found all the parameter loads, then we want to insert
4029 before the first parameter load.
4031 If we did not find all the parameter loads, then we might have
4032 stopped on the head of the block, which could be a CODE_LABEL.
4033 If we inserted before the CODE_LABEL, then we would be putting
4034 the insn in the wrong basic block. In that case, put the insn
4035 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4036 while (LABEL_P (insn)
4037 || NOTE_INSN_BASIC_BLOCK_P (insn))
4038 insn = NEXT_INSN (insn);
4040 new_insn = emit_insn_before_noloc (pat, insn, bb);
4043 new_insn = emit_insn_after_noloc (pat, insn, bb);
4049 add_label_notes (PATTERN (pat), new_insn);
4050 note_stores (PATTERN (pat), record_set_info, pat);
4054 pat = NEXT_INSN (pat);
4057 gcse_create_count++;
4061 fprintf (dump_file, "PRE/HOIST: end of bb %d, insn %d, ",
4062 bb->index, INSN_UID (new_insn));
4063 fprintf (dump_file, "copying expression %d to reg %d\n",
4064 expr->bitmap_index, regno);
4068 /* Insert partially redundant expressions on edges in the CFG to make
4069 the expressions fully redundant. */
4072 pre_edge_insert (struct edge_list *edge_list, struct expr **index_map)
4074 int e, i, j, num_edges, set_size, did_insert = 0;
4077 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4078 if it reaches any of the deleted expressions. */
4080 set_size = pre_insert_map[0]->size;
4081 num_edges = NUM_EDGES (edge_list);
4082 inserted = sbitmap_vector_alloc (num_edges, expr_hash_table.n_elems);
4083 sbitmap_vector_zero (inserted, num_edges);
4085 for (e = 0; e < num_edges; e++)
4088 basic_block bb = INDEX_EDGE_PRED_BB (edge_list, e);
4090 for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
4092 SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
4094 for (j = indx; insert && j < (int) expr_hash_table.n_elems; j++, insert >>= 1)
4095 if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
4097 struct expr *expr = index_map[j];
4100 /* Now look at each deleted occurrence of this expression. */
4101 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4103 if (! occr->deleted_p)
4106 /* Insert this expression on this edge if it would
4107 reach the deleted occurrence in BB. */
4108 if (!TEST_BIT (inserted[e], j))
4111 edge eg = INDEX_EDGE (edge_list, e);
4113 /* We can't insert anything on an abnormal and
4114 critical edge, so we insert the insn at the end of
4115 the previous block. There are several alternatives
4116 detailed in Morgans book P277 (sec 10.5) for
4117 handling this situation. This one is easiest for
4120 if (eg->flags & EDGE_ABNORMAL)
4121 insert_insn_end_basic_block (index_map[j], bb, 0);
4124 insn = process_insert_insn (index_map[j]);
4125 insert_insn_on_edge (insn, eg);
4130 fprintf (dump_file, "PRE/HOIST: edge (%d,%d), ",
4132 INDEX_EDGE_SUCC_BB (edge_list, e)->index);
4133 fprintf (dump_file, "copy expression %d\n",
4134 expr->bitmap_index);
4137 update_ld_motion_stores (expr);
4138 SET_BIT (inserted[e], j);
4140 gcse_create_count++;
4147 sbitmap_vector_free (inserted);
4151 /* Copy the result of EXPR->EXPR generated by INSN to EXPR->REACHING_REG.
4152 Given "old_reg <- expr" (INSN), instead of adding after it
4153 reaching_reg <- old_reg
4154 it's better to do the following:
4155 reaching_reg <- expr
4156 old_reg <- reaching_reg
4157 because this way copy propagation can discover additional PRE
4158 opportunities. But if this fails, we try the old way.
4159 When "expr" is a store, i.e.
4160 given "MEM <- old_reg", instead of adding after it
4161 reaching_reg <- old_reg
4162 it's better to add it before as follows:
4163 reaching_reg <- old_reg
4164 MEM <- reaching_reg. */
4167 pre_insert_copy_insn (struct expr *expr, rtx insn)
4169 rtx reg = expr->reaching_reg;
4170 int regno = REGNO (reg);
4171 int indx = expr->bitmap_index;
4172 rtx pat = PATTERN (insn);
4173 rtx set, first_set, new_insn;
4177 /* This block matches the logic in hash_scan_insn. */
4178 switch (GET_CODE (pat))
4185 /* Search through the parallel looking for the set whose
4186 source was the expression that we're interested in. */
4187 first_set = NULL_RTX;
4189 for (i = 0; i < XVECLEN (pat, 0); i++)
4191 rtx x = XVECEXP (pat, 0, i);
4192 if (GET_CODE (x) == SET)
4194 /* If the source was a REG_EQUAL or REG_EQUIV note, we
4195 may not find an equivalent expression, but in this
4196 case the PARALLEL will have a single set. */
4197 if (first_set == NULL_RTX)
4199 if (expr_equiv_p (SET_SRC (x), expr->expr))
4207 gcc_assert (first_set);
4208 if (set == NULL_RTX)
4216 if (REG_P (SET_DEST (set)))
4218 old_reg = SET_DEST (set);
4219 /* Check if we can modify the set destination in the original insn. */
4220 if (validate_change (insn, &SET_DEST (set), reg, 0))
4222 new_insn = gen_move_insn (old_reg, reg);
4223 new_insn = emit_insn_after (new_insn, insn);
4225 /* Keep register set table up to date. */
4226 record_one_set (regno, insn);
4230 new_insn = gen_move_insn (reg, old_reg);
4231 new_insn = emit_insn_after (new_insn, insn);
4233 /* Keep register set table up to date. */
4234 record_one_set (regno, new_insn);
4237 else /* This is possible only in case of a store to memory. */
4239 old_reg = SET_SRC (set);
4240 new_insn = gen_move_insn (reg, old_reg);
4242 /* Check if we can modify the set source in the original insn. */
4243 if (validate_change (insn, &SET_SRC (set), reg, 0))
4244 new_insn = emit_insn_before (new_insn, insn);
4246 new_insn = emit_insn_after (new_insn, insn);
4248 /* Keep register set table up to date. */
4249 record_one_set (regno, new_insn);
4252 gcse_create_count++;
4256 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4257 BLOCK_NUM (insn), INSN_UID (new_insn), indx,
4258 INSN_UID (insn), regno);
4261 /* Copy available expressions that reach the redundant expression
4262 to `reaching_reg'. */
4265 pre_insert_copies (void)
4267 unsigned int i, added_copy;
4272 /* For each available expression in the table, copy the result to
4273 `reaching_reg' if the expression reaches a deleted one.
4275 ??? The current algorithm is rather brute force.
4276 Need to do some profiling. */
4278 for (i = 0; i < expr_hash_table.size; i++)
4279 for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
4281 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4282 we don't want to insert a copy here because the expression may not
4283 really be redundant. So only insert an insn if the expression was
4284 deleted. This test also avoids further processing if the
4285 expression wasn't deleted anywhere. */
4286 if (expr->reaching_reg == NULL)
4289 /* Set when we add a copy for that expression. */
4292 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4294 if (! occr->deleted_p)
4297 for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
4299 rtx insn = avail->insn;
4301 /* No need to handle this one if handled already. */
4302 if (avail->copied_p)
4305 /* Don't handle this one if it's a redundant one. */
4306 if (TEST_BIT (pre_redundant_insns, INSN_CUID (insn)))
4309 /* Or if the expression doesn't reach the deleted one. */
4310 if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail->insn),
4312 BLOCK_FOR_INSN (occr->insn)))
4317 /* Copy the result of avail to reaching_reg. */
4318 pre_insert_copy_insn (expr, insn);
4319 avail->copied_p = 1;
4324 update_ld_motion_stores (expr);
4328 /* Emit move from SRC to DEST noting the equivalence with expression computed
4331 gcse_emit_move_after (rtx src, rtx dest, rtx insn)
4334 rtx set = single_set (insn), set2;
4338 /* This should never fail since we're creating a reg->reg copy
4339 we've verified to be valid. */
4341 new_rtx = emit_insn_after (gen_move_insn (dest, src), insn);
4343 /* Note the equivalence for local CSE pass. */
4344 set2 = single_set (new_rtx);
4345 if (!set2 || !rtx_equal_p (SET_DEST (set2), dest))
4347 if ((note = find_reg_equal_equiv_note (insn)))
4348 eqv = XEXP (note, 0);
4350 eqv = SET_SRC (set);
4352 set_unique_reg_note (new_rtx, REG_EQUAL, copy_insn_1 (eqv));
4357 /* Delete redundant computations.
4358 Deletion is done by changing the insn to copy the `reaching_reg' of
4359 the expression into the result of the SET. It is left to later passes
4360 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4362 Returns nonzero if a change is made. */
4373 for (i = 0; i < expr_hash_table.size; i++)
4374 for (expr = expr_hash_table.table[i];
4376 expr = expr->next_same_hash)
4378 int indx = expr->bitmap_index;
4380 /* We only need to search antic_occr since we require
4383 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
4385 rtx insn = occr->insn;
4387 basic_block bb = BLOCK_FOR_INSN (insn);
4389 /* We only delete insns that have a single_set. */
4390 if (TEST_BIT (pre_delete_map[bb->index], indx)
4391 && (set = single_set (insn)) != 0
4392 && dbg_cnt (pre_insn))
4394 /* Create a pseudo-reg to store the result of reaching
4395 expressions into. Get the mode for the new pseudo from
4396 the mode of the original destination pseudo. */
4397 if (expr->reaching_reg == NULL)
4398 expr->reaching_reg = gen_reg_rtx_and_attrs (SET_DEST (set));
4400 gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
4402 occr->deleted_p = 1;
4403 SET_BIT (pre_redundant_insns, INSN_CUID (insn));
4410 "PRE: redundant insn %d (expression %d) in ",
4411 INSN_UID (insn), indx);
4412 fprintf (dump_file, "bb %d, reaching reg is %d\n",
4413 bb->index, REGNO (expr->reaching_reg));
4422 /* Perform GCSE optimizations using PRE.
4423 This is called by one_pre_gcse_pass after all the dataflow analysis
4426 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4427 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4428 Compiler Design and Implementation.
4430 ??? A new pseudo reg is created to hold the reaching expression. The nice
4431 thing about the classical approach is that it would try to use an existing
4432 reg. If the register can't be adequately optimized [i.e. we introduce
4433 reload problems], one could add a pass here to propagate the new register
4436 ??? We don't handle single sets in PARALLELs because we're [currently] not
4437 able to copy the rest of the parallel when we insert copies to create full
4438 redundancies from partial redundancies. However, there's no reason why we
4439 can't handle PARALLELs in the cases where there are no partial
4446 int did_insert, changed;
4447 struct expr **index_map;
4450 /* Compute a mapping from expression number (`bitmap_index') to
4451 hash table entry. */
4453 index_map = XCNEWVEC (struct expr *, expr_hash_table.n_elems);
4454 for (i = 0; i < expr_hash_table.size; i++)
4455 for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
4456 index_map[expr->bitmap_index] = expr;
4458 /* Reset bitmap used to track which insns are redundant. */
4459 pre_redundant_insns = sbitmap_alloc (max_cuid);
4460 sbitmap_zero (pre_redundant_insns);
4462 /* Delete the redundant insns first so that
4463 - we know what register to use for the new insns and for the other
4464 ones with reaching expressions
4465 - we know which insns are redundant when we go to create copies */
4467 changed = pre_delete ();
4468 did_insert = pre_edge_insert (edge_list, index_map);
4470 /* In other places with reaching expressions, copy the expression to the
4471 specially allocated pseudo-reg that reaches the redundant expr. */
4472 pre_insert_copies ();
4475 commit_edge_insertions ();
4480 sbitmap_free (pre_redundant_insns);
4484 /* Top level routine to perform one PRE GCSE pass.
4486 Return nonzero if a change was made. */
4489 one_pre_gcse_pass (int pass)
4493 gcse_subst_count = 0;
4494 gcse_create_count = 0;
4496 alloc_hash_table (max_cuid, &expr_hash_table, 0);
4497 add_noreturn_fake_exit_edges ();
4499 compute_ld_motion_mems ();
4501 compute_hash_table (&expr_hash_table);
4502 trim_ld_motion_mems ();
4504 dump_hash_table (dump_file, "Expression", &expr_hash_table);
4506 if (expr_hash_table.n_elems > 0)
4508 alloc_pre_mem (last_basic_block, expr_hash_table.n_elems);
4509 compute_pre_data ();
4510 changed |= pre_gcse ();
4511 free_edge_list (edge_list);
4516 remove_fake_exit_edges ();
4517 free_hash_table (&expr_hash_table);
4521 fprintf (dump_file, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
4522 current_function_name (), pass, bytes_used);
4523 fprintf (dump_file, "%d substs, %d insns created\n",
4524 gcse_subst_count, gcse_create_count);
4530 /* If X contains any LABEL_REF's, add REG_LABEL_OPERAND notes for them
4531 to INSN. If such notes are added to an insn which references a
4532 CODE_LABEL, the LABEL_NUSES count is incremented. We have to add
4533 that note, because the following loop optimization pass requires
4536 /* ??? If there was a jump optimization pass after gcse and before loop,
4537 then we would not need to do this here, because jump would add the
4538 necessary REG_LABEL_OPERAND and REG_LABEL_TARGET notes. */
4541 add_label_notes (rtx x, rtx insn)
4543 enum rtx_code code = GET_CODE (x);
4547 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
4549 /* This code used to ignore labels that referred to dispatch tables to
4550 avoid flow generating (slightly) worse code.
4552 We no longer ignore such label references (see LABEL_REF handling in
4553 mark_jump_label for additional information). */
4555 /* There's no reason for current users to emit jump-insns with
4556 such a LABEL_REF, so we don't have to handle REG_LABEL_TARGET
4558 gcc_assert (!JUMP_P (insn));
4559 add_reg_note (insn, REG_LABEL_OPERAND, XEXP (x, 0));
4561 if (LABEL_P (XEXP (x, 0)))
4562 LABEL_NUSES (XEXP (x, 0))++;
4567 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
4570 add_label_notes (XEXP (x, i), insn);
4571 else if (fmt[i] == 'E')
4572 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4573 add_label_notes (XVECEXP (x, i, j), insn);
4577 /* Compute transparent outgoing information for each block.
4579 An expression is transparent to an edge unless it is killed by
4580 the edge itself. This can only happen with abnormal control flow,
4581 when the edge is traversed through a call. This happens with
4582 non-local labels and exceptions.
4584 This would not be necessary if we split the edge. While this is
4585 normally impossible for abnormal critical edges, with some effort
4586 it should be possible with exception handling, since we still have
4587 control over which handler should be invoked. But due to increased
4588 EH table sizes, this may not be worthwhile. */
4591 compute_transpout (void)
4597 sbitmap_vector_ones (transpout, last_basic_block);
4601 /* Note that flow inserted a nop at the end of basic blocks that
4602 end in call instructions for reasons other than abnormal
4604 if (! CALL_P (BB_END (bb)))
4607 for (i = 0; i < expr_hash_table.size; i++)
4608 for (expr = expr_hash_table.table[i]; expr ; expr = expr->next_same_hash)
4609 if (MEM_P (expr->expr))
4611 if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
4612 && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
4615 /* ??? Optimally, we would use interprocedural alias
4616 analysis to determine if this mem is actually killed
4618 RESET_BIT (transpout[bb->index], expr->bitmap_index);
4623 /* Code Hoisting variables and subroutines. */
4625 /* Very busy expressions. */
4626 static sbitmap *hoist_vbein;
4627 static sbitmap *hoist_vbeout;
4629 /* Hoistable expressions. */
4630 static sbitmap *hoist_exprs;
4632 /* ??? We could compute post dominators and run this algorithm in
4633 reverse to perform tail merging, doing so would probably be
4634 more effective than the tail merging code in jump.c.
4636 It's unclear if tail merging could be run in parallel with
4637 code hoisting. It would be nice. */
4639 /* Allocate vars used for code hoisting analysis. */
4642 alloc_code_hoist_mem (int n_blocks, int n_exprs)
4644 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
4645 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
4646 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
4648 hoist_vbein = sbitmap_vector_alloc (n_blocks, n_exprs);
4649 hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
4650 hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
4651 transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
4654 /* Free vars used for code hoisting analysis. */
4657 free_code_hoist_mem (void)
4659 sbitmap_vector_free (antloc);
4660 sbitmap_vector_free (transp);
4661 sbitmap_vector_free (comp);
4663 sbitmap_vector_free (hoist_vbein);
4664 sbitmap_vector_free (hoist_vbeout);
4665 sbitmap_vector_free (hoist_exprs);
4666 sbitmap_vector_free (transpout);
4668 free_dominance_info (CDI_DOMINATORS);
4671 /* Compute the very busy expressions at entry/exit from each block.
4673 An expression is very busy if all paths from a given point
4674 compute the expression. */
4677 compute_code_hoist_vbeinout (void)
4679 int changed, passes;
4682 sbitmap_vector_zero (hoist_vbeout, last_basic_block);
4683 sbitmap_vector_zero (hoist_vbein, last_basic_block);
4692 /* We scan the blocks in the reverse order to speed up
4694 FOR_EACH_BB_REVERSE (bb)
4696 if (bb->next_bb != EXIT_BLOCK_PTR)
4697 sbitmap_intersection_of_succs (hoist_vbeout[bb->index],
4698 hoist_vbein, bb->index);
4700 changed |= sbitmap_a_or_b_and_c_cg (hoist_vbein[bb->index],
4702 hoist_vbeout[bb->index],
4710 fprintf (dump_file, "hoisting vbeinout computation: %d passes\n", passes);
4713 /* Top level routine to do the dataflow analysis needed by code hoisting. */
4716 compute_code_hoist_data (void)
4718 compute_local_properties (transp, comp, antloc, &expr_hash_table);
4719 compute_transpout ();
4720 compute_code_hoist_vbeinout ();
4721 calculate_dominance_info (CDI_DOMINATORS);
4723 fprintf (dump_file, "\n");
4726 /* Determine if the expression identified by EXPR_INDEX would
4727 reach BB unimpared if it was placed at the end of EXPR_BB.
4729 It's unclear exactly what Muchnick meant by "unimpared". It seems
4730 to me that the expression must either be computed or transparent in
4731 *every* block in the path(s) from EXPR_BB to BB. Any other definition
4732 would allow the expression to be hoisted out of loops, even if
4733 the expression wasn't a loop invariant.
4735 Contrast this to reachability for PRE where an expression is
4736 considered reachable if *any* path reaches instead of *all*
4740 hoist_expr_reaches_here_p (basic_block expr_bb, int expr_index, basic_block bb, char *visited)
4744 int visited_allocated_locally = 0;
4747 if (visited == NULL)
4749 visited_allocated_locally = 1;
4750 visited = XCNEWVEC (char, last_basic_block);
4753 FOR_EACH_EDGE (pred, ei, bb->preds)
4755 basic_block pred_bb = pred->src;
4757 if (pred->src == ENTRY_BLOCK_PTR)
4759 else if (pred_bb == expr_bb)
4761 else if (visited[pred_bb->index])
4764 /* Does this predecessor generate this expression? */
4765 else if (TEST_BIT (comp[pred_bb->index], expr_index))
4767 else if (! TEST_BIT (transp[pred_bb->index], expr_index))
4773 visited[pred_bb->index] = 1;
4774 if (! hoist_expr_reaches_here_p (expr_bb, expr_index,
4779 if (visited_allocated_locally)
4782 return (pred == NULL);
4785 /* Actually perform code hoisting. */
4790 basic_block bb, dominated;
4791 VEC (basic_block, heap) *domby;
4793 struct expr **index_map;
4796 sbitmap_vector_zero (hoist_exprs, last_basic_block);
4798 /* Compute a mapping from expression number (`bitmap_index') to
4799 hash table entry. */
4801 index_map = XCNEWVEC (struct expr *, expr_hash_table.n_elems);
4802 for (i = 0; i < expr_hash_table.size; i++)
4803 for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
4804 index_map[expr->bitmap_index] = expr;
4806 /* Walk over each basic block looking for potentially hoistable
4807 expressions, nothing gets hoisted from the entry block. */
4811 int insn_inserted_p;
4813 domby = get_dominated_by (CDI_DOMINATORS, bb);
4814 /* Examine each expression that is very busy at the exit of this
4815 block. These are the potentially hoistable expressions. */
4816 for (i = 0; i < hoist_vbeout[bb->index]->n_bits; i++)
4820 if (TEST_BIT (hoist_vbeout[bb->index], i)
4821 && TEST_BIT (transpout[bb->index], i))
4823 /* We've found a potentially hoistable expression, now
4824 we look at every block BB dominates to see if it
4825 computes the expression. */
4826 for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
4828 /* Ignore self dominance. */
4829 if (bb == dominated)
4831 /* We've found a dominated block, now see if it computes
4832 the busy expression and whether or not moving that
4833 expression to the "beginning" of that block is safe. */
4834 if (!TEST_BIT (antloc[dominated->index], i))
4837 /* Note if the expression would reach the dominated block
4838 unimpared if it was placed at the end of BB.
4840 Keep track of how many times this expression is hoistable
4841 from a dominated block into BB. */
4842 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
4846 /* If we found more than one hoistable occurrence of this
4847 expression, then note it in the bitmap of expressions to
4848 hoist. It makes no sense to hoist things which are computed
4849 in only one BB, and doing so tends to pessimize register
4850 allocation. One could increase this value to try harder
4851 to avoid any possible code expansion due to register
4852 allocation issues; however experiments have shown that
4853 the vast majority of hoistable expressions are only movable
4854 from two successors, so raising this threshold is likely
4855 to nullify any benefit we get from code hoisting. */
4858 SET_BIT (hoist_exprs[bb->index], i);
4863 /* If we found nothing to hoist, then quit now. */
4866 VEC_free (basic_block, heap, domby);
4870 /* Loop over all the hoistable expressions. */
4871 for (i = 0; i < hoist_exprs[bb->index]->n_bits; i++)
4873 /* We want to insert the expression into BB only once, so
4874 note when we've inserted it. */
4875 insn_inserted_p = 0;
4877 /* These tests should be the same as the tests above. */
4878 if (TEST_BIT (hoist_exprs[bb->index], i))
4880 /* We've found a potentially hoistable expression, now
4881 we look at every block BB dominates to see if it
4882 computes the expression. */
4883 for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
4885 /* Ignore self dominance. */
4886 if (bb == dominated)
4889 /* We've found a dominated block, now see if it computes
4890 the busy expression and whether or not moving that
4891 expression to the "beginning" of that block is safe. */
4892 if (!TEST_BIT (antloc[dominated->index], i))
4895 /* The expression is computed in the dominated block and
4896 it would be safe to compute it at the start of the
4897 dominated block. Now we have to determine if the
4898 expression would reach the dominated block if it was
4899 placed at the end of BB. */
4900 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
4902 struct expr *expr = index_map[i];
4903 struct occr *occr = expr->antic_occr;
4907 /* Find the right occurrence of this expression. */
4908 while (BLOCK_FOR_INSN (occr->insn) != dominated && occr)
4913 set = single_set (insn);
4916 /* Create a pseudo-reg to store the result of reaching
4917 expressions into. Get the mode for the new pseudo
4918 from the mode of the original destination pseudo. */
4919 if (expr->reaching_reg == NULL)
4921 = gen_reg_rtx_and_attrs (SET_DEST (set));
4923 gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
4925 occr->deleted_p = 1;
4926 if (!insn_inserted_p)
4928 insert_insn_end_basic_block (index_map[i], bb, 0);
4929 insn_inserted_p = 1;
4935 VEC_free (basic_block, heap, domby);
4941 /* Top level routine to perform one code hoisting (aka unification) pass
4943 Return nonzero if a change was made. */
4946 one_code_hoisting_pass (void)
4950 alloc_hash_table (max_cuid, &expr_hash_table, 0);
4951 compute_hash_table (&expr_hash_table);
4953 dump_hash_table (dump_file, "Code Hosting Expressions", &expr_hash_table);
4955 if (expr_hash_table.n_elems > 0)
4957 alloc_code_hoist_mem (last_basic_block, expr_hash_table.n_elems);
4958 compute_code_hoist_data ();
4960 free_code_hoist_mem ();
4963 free_hash_table (&expr_hash_table);
4968 /* Here we provide the things required to do store motion towards
4969 the exit. In order for this to be effective, gcse also needed to
4970 be taught how to move a load when it is kill only by a store to itself.
4975 void foo(float scale)
4977 for (i=0; i<10; i++)
4981 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
4982 the load out since its live around the loop, and stored at the bottom
4985 The 'Load Motion' referred to and implemented in this file is
4986 an enhancement to gcse which when using edge based lcm, recognizes
4987 this situation and allows gcse to move the load out of the loop.
4989 Once gcse has hoisted the load, store motion can then push this
4990 load towards the exit, and we end up with no loads or stores of 'i'
4994 pre_ldst_expr_hash (const void *p)
4996 int do_not_record_p = 0;
4997 const struct ls_expr *const x = (const struct ls_expr *) p;
4998 return hash_rtx (x->pattern, GET_MODE (x->pattern), &do_not_record_p, NULL, false);
5002 pre_ldst_expr_eq (const void *p1, const void *p2)
5004 const struct ls_expr *const ptr1 = (const struct ls_expr *) p1,
5005 *const ptr2 = (const struct ls_expr *) p2;
5006 return expr_equiv_p (ptr1->pattern, ptr2->pattern);
5009 /* This will search the ldst list for a matching expression. If it
5010 doesn't find one, we create one and initialize it. */
5012 static struct ls_expr *
5015 int do_not_record_p = 0;
5016 struct ls_expr * ptr;
5021 hash = hash_rtx (x, GET_MODE (x), &do_not_record_p,
5022 NULL, /*have_reg_qty=*/false);
5025 slot = htab_find_slot_with_hash (pre_ldst_table, &e, hash, INSERT);
5027 return (struct ls_expr *)*slot;
5029 ptr = XNEW (struct ls_expr);
5031 ptr->next = pre_ldst_mems;
5034 ptr->pattern_regs = NULL_RTX;
5035 ptr->loads = NULL_RTX;
5036 ptr->stores = NULL_RTX;
5037 ptr->reaching_reg = NULL_RTX;
5040 ptr->hash_index = hash;
5041 pre_ldst_mems = ptr;
5047 /* Free up an individual ldst entry. */
5050 free_ldst_entry (struct ls_expr * ptr)
5052 free_INSN_LIST_list (& ptr->loads);
5053 free_INSN_LIST_list (& ptr->stores);
5058 /* Free up all memory associated with the ldst list. */
5061 free_ldst_mems (void)
5064 htab_delete (pre_ldst_table);
5065 pre_ldst_table = NULL;
5067 while (pre_ldst_mems)
5069 struct ls_expr * tmp = pre_ldst_mems;
5071 pre_ldst_mems = pre_ldst_mems->next;
5073 free_ldst_entry (tmp);
5076 pre_ldst_mems = NULL;
5079 /* Dump debugging info about the ldst list. */
5082 print_ldst_list (FILE * file)
5084 struct ls_expr * ptr;
5086 fprintf (file, "LDST list: \n");
5088 for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
5090 fprintf (file, " Pattern (%3d): ", ptr->index);
5092 print_rtl (file, ptr->pattern);
5094 fprintf (file, "\n Loads : ");
5097 print_rtl (file, ptr->loads);
5099 fprintf (file, "(nil)");
5101 fprintf (file, "\n Stores : ");
5104 print_rtl (file, ptr->stores);
5106 fprintf (file, "(nil)");
5108 fprintf (file, "\n\n");
5111 fprintf (file, "\n");
5114 /* Returns 1 if X is in the list of ldst only expressions. */
5116 static struct ls_expr *
5117 find_rtx_in_ldst (rtx x)
5121 if (!pre_ldst_table)
5124 slot = htab_find_slot (pre_ldst_table, &e, NO_INSERT);
5125 if (!slot || ((struct ls_expr *)*slot)->invalid)
5127 return (struct ls_expr *) *slot;
5130 /* Assign each element of the list of mems a monotonically increasing value. */
5133 enumerate_ldsts (void)
5135 struct ls_expr * ptr;
5138 for (ptr = pre_ldst_mems; ptr != NULL; ptr = ptr->next)
5144 /* Return first item in the list. */
5146 static inline struct ls_expr *
5147 first_ls_expr (void)
5149 return pre_ldst_mems;
5152 /* Return the next item in the list after the specified one. */
5154 static inline struct ls_expr *
5155 next_ls_expr (struct ls_expr * ptr)
5160 /* Load Motion for loads which only kill themselves. */
5162 /* Return true if x is a simple MEM operation, with no registers or
5163 side effects. These are the types of loads we consider for the
5164 ld_motion list, otherwise we let the usual aliasing take care of it. */
5167 simple_mem (const_rtx x)
5172 if (MEM_VOLATILE_P (x))
5175 if (GET_MODE (x) == BLKmode)
5178 /* If we are handling exceptions, we must be careful with memory references
5179 that may trap. If we are not, the behavior is undefined, so we may just
5181 if (flag_non_call_exceptions && may_trap_p (x))
5184 if (side_effects_p (x))
5187 /* Do not consider function arguments passed on stack. */
5188 if (reg_mentioned_p (stack_pointer_rtx, x))
5191 if (flag_float_store && FLOAT_MODE_P (GET_MODE (x)))
5197 /* Make sure there isn't a buried reference in this pattern anywhere.
5198 If there is, invalidate the entry for it since we're not capable
5199 of fixing it up just yet.. We have to be sure we know about ALL
5200 loads since the aliasing code will allow all entries in the
5201 ld_motion list to not-alias itself. If we miss a load, we will get
5202 the wrong value since gcse might common it and we won't know to
5206 invalidate_any_buried_refs (rtx x)
5210 struct ls_expr * ptr;
5212 /* Invalidate it in the list. */
5213 if (MEM_P (x) && simple_mem (x))
5215 ptr = ldst_entry (x);
5219 /* Recursively process the insn. */
5220 fmt = GET_RTX_FORMAT (GET_CODE (x));
5222 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
5225 invalidate_any_buried_refs (XEXP (x, i));
5226 else if (fmt[i] == 'E')
5227 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5228 invalidate_any_buried_refs (XVECEXP (x, i, j));
5232 /* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
5233 being defined as MEM loads and stores to symbols, with no side effects
5234 and no registers in the expression. For a MEM destination, we also
5235 check that the insn is still valid if we replace the destination with a
5236 REG, as is done in update_ld_motion_stores. If there are any uses/defs
5237 which don't match this criteria, they are invalidated and trimmed out
5241 compute_ld_motion_mems (void)
5243 struct ls_expr * ptr;
5247 pre_ldst_mems = NULL;
5248 pre_ldst_table = htab_create (13, pre_ldst_expr_hash,
5249 pre_ldst_expr_eq, NULL);
5253 FOR_BB_INSNS (bb, insn)
5257 if (GET_CODE (PATTERN (insn)) == SET)
5259 rtx src = SET_SRC (PATTERN (insn));
5260 rtx dest = SET_DEST (PATTERN (insn));
5262 /* Check for a simple LOAD... */
5263 if (MEM_P (src) && simple_mem (src))
5265 ptr = ldst_entry (src);
5267 ptr->loads = alloc_INSN_LIST (insn, ptr->loads);
5273 /* Make sure there isn't a buried load somewhere. */
5274 invalidate_any_buried_refs (src);
5277 /* Check for stores. Don't worry about aliased ones, they
5278 will block any movement we might do later. We only care
5279 about this exact pattern since those are the only
5280 circumstance that we will ignore the aliasing info. */
5281 if (MEM_P (dest) && simple_mem (dest))
5283 ptr = ldst_entry (dest);
5286 && GET_CODE (src) != ASM_OPERANDS
5287 /* Check for REG manually since want_to_gcse_p
5288 returns 0 for all REGs. */
5289 && can_assign_to_reg_p (src))
5290 ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
5296 invalidate_any_buried_refs (PATTERN (insn));
5302 /* Remove any references that have been either invalidated or are not in the
5303 expression list for pre gcse. */
5306 trim_ld_motion_mems (void)
5308 struct ls_expr * * last = & pre_ldst_mems;
5309 struct ls_expr * ptr = pre_ldst_mems;
5315 /* Delete if entry has been made invalid. */
5318 /* Delete if we cannot find this mem in the expression list. */
5319 unsigned int hash = ptr->hash_index % expr_hash_table.size;
5321 for (expr = expr_hash_table.table[hash];
5323 expr = expr->next_same_hash)
5324 if (expr_equiv_p (expr->expr, ptr->pattern))
5328 expr = (struct expr *) 0;
5332 /* Set the expression field if we are keeping it. */
5340 htab_remove_elt_with_hash (pre_ldst_table, ptr, ptr->hash_index);
5341 free_ldst_entry (ptr);
5346 /* Show the world what we've found. */
5347 if (dump_file && pre_ldst_mems != NULL)
5348 print_ldst_list (dump_file);
5351 /* This routine will take an expression which we are replacing with
5352 a reaching register, and update any stores that are needed if
5353 that expression is in the ld_motion list. Stores are updated by
5354 copying their SRC to the reaching register, and then storing
5355 the reaching register into the store location. These keeps the
5356 correct value in the reaching register for the loads. */
5359 update_ld_motion_stores (struct expr * expr)
5361 struct ls_expr * mem_ptr;
5363 if ((mem_ptr = find_rtx_in_ldst (expr->expr)))
5365 /* We can try to find just the REACHED stores, but is shouldn't
5366 matter to set the reaching reg everywhere... some might be
5367 dead and should be eliminated later. */
5369 /* We replace (set mem expr) with (set reg expr) (set mem reg)
5370 where reg is the reaching reg used in the load. We checked in
5371 compute_ld_motion_mems that we can replace (set mem expr) with
5372 (set reg expr) in that insn. */
5373 rtx list = mem_ptr->stores;
5375 for ( ; list != NULL_RTX; list = XEXP (list, 1))
5377 rtx insn = XEXP (list, 0);
5378 rtx pat = PATTERN (insn);
5379 rtx src = SET_SRC (pat);
5380 rtx reg = expr->reaching_reg;
5383 /* If we've already copied it, continue. */
5384 if (expr->reaching_reg == src)
5389 fprintf (dump_file, "PRE: store updated with reaching reg ");
5390 print_rtl (dump_file, expr->reaching_reg);
5391 fprintf (dump_file, ":\n ");
5392 print_inline_rtx (dump_file, insn, 8);
5393 fprintf (dump_file, "\n");
5396 copy = gen_move_insn ( reg, copy_rtx (SET_SRC (pat)));
5397 new_rtx = emit_insn_before (copy, insn);
5398 record_one_set (REGNO (reg), new_rtx);
5399 SET_SRC (pat) = reg;
5400 df_insn_rescan (insn);
5402 /* un-recognize this pattern since it's probably different now. */
5403 INSN_CODE (insn) = -1;
5404 gcse_create_count++;
5409 /* Store motion code. */
5411 #define ANTIC_STORE_LIST(x) ((x)->loads)
5412 #define AVAIL_STORE_LIST(x) ((x)->stores)
5413 #define LAST_AVAIL_CHECK_FAILURE(x) ((x)->reaching_reg)
5415 /* This is used to communicate the target bitvector we want to use in the
5416 reg_set_info routine when called via the note_stores mechanism. */
5417 static int * regvec;
5419 /* And current insn, for the same routine. */
5420 static rtx compute_store_table_current_insn;
5422 /* Used in computing the reverse edge graph bit vectors. */
5423 static sbitmap * st_antloc;
5425 /* Global holding the number of store expressions we are dealing with. */
5426 static int num_stores;
5428 /* Checks to set if we need to mark a register set. Called from
5432 reg_set_info (rtx dest, const_rtx setter ATTRIBUTE_UNUSED,
5435 sbitmap bb_reg = (sbitmap) data;
5437 if (GET_CODE (dest) == SUBREG)
5438 dest = SUBREG_REG (dest);
5442 regvec[REGNO (dest)] = INSN_UID (compute_store_table_current_insn);
5444 SET_BIT (bb_reg, REGNO (dest));
5448 /* Clear any mark that says that this insn sets dest. Called from
5452 reg_clear_last_set (rtx dest, const_rtx setter ATTRIBUTE_UNUSED,
5455 int *dead_vec = (int *) data;
5457 if (GET_CODE (dest) == SUBREG)
5458 dest = SUBREG_REG (dest);
5461 dead_vec[REGNO (dest)] == INSN_UID (compute_store_table_current_insn))
5462 dead_vec[REGNO (dest)] = 0;
5465 /* Return zero if some of the registers in list X are killed
5466 due to set of registers in bitmap REGS_SET. */
5469 store_ops_ok (const_rtx x, int *regs_set)
5473 for (; x; x = XEXP (x, 1))
5476 if (regs_set[REGNO(reg)])
5483 /* Returns a list of registers mentioned in X. */
5485 extract_mentioned_regs (rtx x)
5487 return extract_mentioned_regs_helper (x, NULL_RTX);
5490 /* Helper for extract_mentioned_regs; ACCUM is used to accumulate used
5493 extract_mentioned_regs_helper (rtx x, rtx accum)
5499 /* Repeat is used to turn tail-recursion into iteration. */
5505 code = GET_CODE (x);
5509 return alloc_EXPR_LIST (0, x, accum);
5521 /* We do not run this function with arguments having side effects. */
5541 i = GET_RTX_LENGTH (code) - 1;
5542 fmt = GET_RTX_FORMAT (code);
5548 rtx tem = XEXP (x, i);
5550 /* If we are about to do the last recursive call
5551 needed at this level, change it into iteration. */
5558 accum = extract_mentioned_regs_helper (tem, accum);
5560 else if (fmt[i] == 'E')
5564 for (j = 0; j < XVECLEN (x, i); j++)
5565 accum = extract_mentioned_regs_helper (XVECEXP (x, i, j), accum);
5572 /* Determine whether INSN is MEM store pattern that we will consider moving.
5573 REGS_SET_BEFORE is bitmap of registers set before (and including) the
5574 current insn, REGS_SET_AFTER is bitmap of registers set after (and
5575 including) the insn in this basic block. We must be passing through BB from
5576 head to end, as we are using this fact to speed things up.
5578 The results are stored this way:
5580 -- the first anticipatable expression is added into ANTIC_STORE_LIST
5581 -- if the processed expression is not anticipatable, NULL_RTX is added
5582 there instead, so that we can use it as indicator that no further
5583 expression of this type may be anticipatable
5584 -- if the expression is available, it is added as head of AVAIL_STORE_LIST;
5585 consequently, all of them but this head are dead and may be deleted.
5586 -- if the expression is not available, the insn due to that it fails to be
5587 available is stored in reaching_reg.
5589 The things are complicated a bit by fact that there already may be stores
5590 to the same MEM from other blocks; also caller must take care of the
5591 necessary cleanup of the temporary markers after end of the basic block.
5595 find_moveable_store (rtx insn, int *regs_set_before, int *regs_set_after)
5597 struct ls_expr * ptr;
5599 int check_anticipatable, check_available;
5600 basic_block bb = BLOCK_FOR_INSN (insn);
5602 set = single_set (insn);
5606 dest = SET_DEST (set);
5608 if (! MEM_P (dest) || MEM_VOLATILE_P (dest)
5609 || GET_MODE (dest) == BLKmode)
5612 if (side_effects_p (dest))
5615 /* If we are handling exceptions, we must be careful with memory references
5616 that may trap. If we are not, the behavior is undefined, so we may just
5618 if (flag_non_call_exceptions && may_trap_p (dest))
5621 /* Even if the destination cannot trap, the source may. In this case we'd
5622 need to handle updating the REG_EH_REGION note. */
5623 if (find_reg_note (insn, REG_EH_REGION, NULL_RTX))
5626 /* Make sure that the SET_SRC of this store insns can be assigned to
5627 a register, or we will fail later on in replace_store_insn, which
5628 assumes that we can do this. But sometimes the target machine has
5629 oddities like MEM read-modify-write instruction. See for example
5631 if (!can_assign_to_reg_p (SET_SRC (set)))
5634 ptr = ldst_entry (dest);
5635 if (!ptr->pattern_regs)
5636 ptr->pattern_regs = extract_mentioned_regs (dest);
5638 /* Do not check for anticipatability if we either found one anticipatable
5639 store already, or tested for one and found out that it was killed. */
5640 check_anticipatable = 0;
5641 if (!ANTIC_STORE_LIST (ptr))
5642 check_anticipatable = 1;
5645 tmp = XEXP (ANTIC_STORE_LIST (ptr), 0);
5647 && BLOCK_FOR_INSN (tmp) != bb)
5648 check_anticipatable = 1;
5650 if (check_anticipatable)
5652 if (store_killed_before (dest, ptr->pattern_regs, insn, bb, regs_set_before))
5656 ANTIC_STORE_LIST (ptr) = alloc_INSN_LIST (tmp,
5657 ANTIC_STORE_LIST (ptr));
5660 /* It is not necessary to check whether store is available if we did
5661 it successfully before; if we failed before, do not bother to check
5662 until we reach the insn that caused us to fail. */
5663 check_available = 0;
5664 if (!AVAIL_STORE_LIST (ptr))
5665 check_available = 1;
5668 tmp = XEXP (AVAIL_STORE_LIST (ptr), 0);
5669 if (BLOCK_FOR_INSN (tmp) != bb)
5670 check_available = 1;
5672 if (check_available)
5674 /* Check that we have already reached the insn at that the check
5675 failed last time. */
5676 if (LAST_AVAIL_CHECK_FAILURE (ptr))
5678 for (tmp = BB_END (bb);
5679 tmp != insn && tmp != LAST_AVAIL_CHECK_FAILURE (ptr);
5680 tmp = PREV_INSN (tmp))
5683 check_available = 0;
5686 check_available = store_killed_after (dest, ptr->pattern_regs, insn,
5688 &LAST_AVAIL_CHECK_FAILURE (ptr));
5690 if (!check_available)
5691 AVAIL_STORE_LIST (ptr) = alloc_INSN_LIST (insn, AVAIL_STORE_LIST (ptr));
5694 /* Find available and anticipatable stores. */
5697 compute_store_table (void)
5703 int *last_set_in, *already_set;
5704 struct ls_expr * ptr, **prev_next_ptr_ptr;
5706 max_gcse_regno = max_reg_num ();
5708 reg_set_in_block = sbitmap_vector_alloc (last_basic_block,
5710 sbitmap_vector_zero (reg_set_in_block, last_basic_block);
5712 pre_ldst_table = htab_create (13, pre_ldst_expr_hash,
5713 pre_ldst_expr_eq, NULL);
5714 last_set_in = XCNEWVEC (int, max_gcse_regno);
5715 already_set = XNEWVEC (int, max_gcse_regno);
5717 /* Find all the stores we care about. */
5720 /* First compute the registers set in this block. */
5721 regvec = last_set_in;
5723 FOR_BB_INSNS (bb, insn)
5725 if (! INSN_P (insn))
5730 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
5731 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
5733 last_set_in[regno] = INSN_UID (insn);
5734 SET_BIT (reg_set_in_block[bb->index], regno);
5738 pat = PATTERN (insn);
5739 compute_store_table_current_insn = insn;
5740 note_stores (pat, reg_set_info, reg_set_in_block[bb->index]);
5743 /* Now find the stores. */
5744 memset (already_set, 0, sizeof (int) * max_gcse_regno);
5745 regvec = already_set;
5746 FOR_BB_INSNS (bb, insn)
5748 if (! INSN_P (insn))
5753 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
5754 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
5755 already_set[regno] = 1;
5758 pat = PATTERN (insn);
5759 note_stores (pat, reg_set_info, NULL);
5761 /* Now that we've marked regs, look for stores. */
5762 find_moveable_store (insn, already_set, last_set_in);
5764 /* Unmark regs that are no longer set. */
5765 compute_store_table_current_insn = insn;
5766 note_stores (pat, reg_clear_last_set, last_set_in);
5769 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
5770 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno)
5771 && last_set_in[regno] == INSN_UID (insn))
5772 last_set_in[regno] = 0;
5776 #ifdef ENABLE_CHECKING
5777 /* last_set_in should now be all-zero. */
5778 for (regno = 0; regno < max_gcse_regno; regno++)
5779 gcc_assert (!last_set_in[regno]);
5782 /* Clear temporary marks. */
5783 for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
5785 LAST_AVAIL_CHECK_FAILURE(ptr) = NULL_RTX;
5786 if (ANTIC_STORE_LIST (ptr)
5787 && (tmp = XEXP (ANTIC_STORE_LIST (ptr), 0)) == NULL_RTX)
5788 ANTIC_STORE_LIST (ptr) = XEXP (ANTIC_STORE_LIST (ptr), 1);
5792 /* Remove the stores that are not available anywhere, as there will
5793 be no opportunity to optimize them. */
5794 for (ptr = pre_ldst_mems, prev_next_ptr_ptr = &pre_ldst_mems;
5796 ptr = *prev_next_ptr_ptr)
5798 if (!AVAIL_STORE_LIST (ptr))
5800 *prev_next_ptr_ptr = ptr->next;
5801 htab_remove_elt_with_hash (pre_ldst_table, ptr, ptr->hash_index);
5802 free_ldst_entry (ptr);
5805 prev_next_ptr_ptr = &ptr->next;
5808 ret = enumerate_ldsts ();
5812 fprintf (dump_file, "ST_avail and ST_antic (shown under loads..)\n");
5813 print_ldst_list (dump_file);
5821 /* Check to see if the load X is aliased with STORE_PATTERN.
5822 AFTER is true if we are checking the case when STORE_PATTERN occurs
5826 load_kills_store (const_rtx x, const_rtx store_pattern, int after)
5829 return anti_dependence (x, store_pattern);
5831 return true_dependence (store_pattern, GET_MODE (store_pattern), x,
5835 /* Go through the entire insn X, looking for any loads which might alias
5836 STORE_PATTERN. Return true if found.
5837 AFTER is true if we are checking the case when STORE_PATTERN occurs
5838 after the insn X. */
5841 find_loads (const_rtx x, const_rtx store_pattern, int after)
5850 if (GET_CODE (x) == SET)
5855 if (load_kills_store (x, store_pattern, after))
5859 /* Recursively process the insn. */
5860 fmt = GET_RTX_FORMAT (GET_CODE (x));
5862 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0 && !ret; i--)
5865 ret |= find_loads (XEXP (x, i), store_pattern, after);
5866 else if (fmt[i] == 'E')
5867 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5868 ret |= find_loads (XVECEXP (x, i, j), store_pattern, after);
5874 store_killed_in_pat (const_rtx x, const_rtx pat, int after)
5876 if (GET_CODE (pat) == SET)
5878 rtx dest = SET_DEST (pat);
5880 if (GET_CODE (dest) == ZERO_EXTRACT)
5881 dest = XEXP (dest, 0);
5883 /* Check for memory stores to aliased objects. */
5885 && !expr_equiv_p (dest, x))
5889 if (output_dependence (dest, x))
5894 if (output_dependence (x, dest))
5900 if (find_loads (pat, x, after))
5906 /* Check if INSN kills the store pattern X (is aliased with it).
5907 AFTER is true if we are checking the case when store X occurs
5908 after the insn. Return true if it does. */
5911 store_killed_in_insn (const_rtx x, const_rtx x_regs, const_rtx insn, int after)
5913 const_rtx reg, base, note, pat;
5920 /* A normal or pure call might read from pattern,
5921 but a const call will not. */
5922 if (!RTL_CONST_CALL_P (insn))
5925 /* But even a const call reads its parameters. Check whether the
5926 base of some of registers used in mem is stack pointer. */
5927 for (reg = x_regs; reg; reg = XEXP (reg, 1))
5929 base = find_base_term (XEXP (reg, 0));
5931 || (GET_CODE (base) == ADDRESS
5932 && GET_MODE (base) == Pmode
5933 && XEXP (base, 0) == stack_pointer_rtx))
5940 pat = PATTERN (insn);
5941 if (GET_CODE (pat) == SET)
5943 if (store_killed_in_pat (x, pat, after))
5946 else if (GET_CODE (pat) == PARALLEL)
5950 for (i = 0; i < XVECLEN (pat, 0); i++)
5951 if (store_killed_in_pat (x, XVECEXP (pat, 0, i), 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 (const_rtx x, const_rtx x_regs, const_rtx insn, const_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 (const_rtx x, const_rtx x_regs, const_rtx insn, const_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_and_attrs (ptr->pattern);
6057 fprintf (dump_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 = XNEWVEC (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 (dump_file, "st_antloc", "", st_antloc, last_basic_block);
6105 dump_sbitmap_vector (dump_file, "st_kill", "", ae_kill, last_basic_block);
6106 dump_sbitmap_vector (dump_file, "Transpt", "", transp, last_basic_block);
6107 dump_sbitmap_vector (dump_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_basic_block (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_INSN_BASIC_BLOCK_P (before))
6126 if (prev == BB_END (bb))
6128 before = NEXT_INSN (before);
6131 insn = emit_insn_after_noloc (insn, prev, bb);
6135 fprintf (dump_file, "STORE_MOTION insert store at start of BB %d:\n",
6137 print_inline_rtx (dump_file, insn, 6);
6138 fprintf (dump_file, "\n");
6142 /* This routine will insert a store on an edge. EXPR is the ldst entry for
6143 the memory reference, and E is the edge to insert it on. Returns nonzero
6144 if an edge insertion was performed. */
6147 insert_store (struct ls_expr * expr, edge e)
6154 /* We did all the deleted before this insert, so if we didn't delete a
6155 store, then we haven't set the reaching reg yet either. */
6156 if (expr->reaching_reg == NULL_RTX)
6159 if (e->flags & EDGE_FAKE)
6162 reg = expr->reaching_reg;
6163 insn = gen_move_insn (copy_rtx (expr->pattern), reg);
6165 /* If we are inserting this expression on ALL predecessor edges of a BB,
6166 insert it at the start of the BB, and reset the insert bits on the other
6167 edges so we don't try to insert it on the other edges. */
6169 FOR_EACH_EDGE (tmp, ei, e->dest->preds)
6170 if (!(tmp->flags & EDGE_FAKE))
6172 int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
6174 gcc_assert (index != EDGE_INDEX_NO_EDGE);
6175 if (! TEST_BIT (pre_insert_map[index], expr->index))
6179 /* If tmp is NULL, we found an insertion on every edge, blank the
6180 insertion vector for these edges, and insert at the start of the BB. */
6181 if (!tmp && bb != EXIT_BLOCK_PTR)
6183 FOR_EACH_EDGE (tmp, ei, e->dest->preds)
6185 int index = EDGE_INDEX (edge_list, tmp->src, tmp->dest);
6186 RESET_BIT (pre_insert_map[index], expr->index);
6188 insert_insn_start_basic_block (insn, bb);
6192 /* We can't put stores in the front of blocks pointed to by abnormal
6193 edges since that may put a store where one didn't used to be. */
6194 gcc_assert (!(e->flags & EDGE_ABNORMAL));
6196 insert_insn_on_edge (insn, e);
6200 fprintf (dump_file, "STORE_MOTION insert insn on edge (%d, %d):\n",
6201 e->src->index, e->dest->index);
6202 print_inline_rtx (dump_file, insn, 6);
6203 fprintf (dump_file, "\n");
6209 /* Remove any REG_EQUAL or REG_EQUIV notes containing a reference to the
6210 memory location in SMEXPR set in basic block BB.
6212 This could be rather expensive. */
6215 remove_reachable_equiv_notes (basic_block bb, struct ls_expr *smexpr)
6217 edge_iterator *stack, ei;
6220 sbitmap visited = sbitmap_alloc (last_basic_block);
6221 rtx last, insn, note;
6222 rtx mem = smexpr->pattern;
6224 stack = XNEWVEC (edge_iterator, n_basic_blocks);
6226 ei = ei_start (bb->succs);
6228 sbitmap_zero (visited);
6230 act = (EDGE_COUNT (ei_container (ei)) > 0 ? EDGE_I (ei_container (ei), 0) : NULL);
6238 sbitmap_free (visited);
6241 act = ei_edge (stack[--sp]);
6245 if (bb == EXIT_BLOCK_PTR
6246 || TEST_BIT (visited, bb->index))
6250 act = (! ei_end_p (ei)) ? ei_edge (ei) : NULL;
6253 SET_BIT (visited, bb->index);
6255 if (TEST_BIT (st_antloc[bb->index], smexpr->index))
6257 for (last = ANTIC_STORE_LIST (smexpr);
6258 BLOCK_FOR_INSN (XEXP (last, 0)) != bb;
6259 last = XEXP (last, 1))
6261 last = XEXP (last, 0);
6264 last = NEXT_INSN (BB_END (bb));
6266 for (insn = BB_HEAD (bb); insn != last; insn = NEXT_INSN (insn))
6269 note = find_reg_equal_equiv_note (insn);
6270 if (!note || !expr_equiv_p (XEXP (note, 0), mem))
6274 fprintf (dump_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
6276 remove_note (insn, note);
6281 act = (! ei_end_p (ei)) ? ei_edge (ei) : NULL;
6283 if (EDGE_COUNT (bb->succs) > 0)
6287 ei = ei_start (bb->succs);
6288 act = (EDGE_COUNT (ei_container (ei)) > 0 ? EDGE_I (ei_container (ei), 0) : NULL);
6293 /* This routine will replace a store with a SET to a specified register. */
6296 replace_store_insn (rtx reg, rtx del, basic_block bb, struct ls_expr *smexpr)
6298 rtx insn, mem, note, set, ptr;
6300 mem = smexpr->pattern;
6301 insn = gen_move_insn (reg, SET_SRC (single_set (del)));
6303 for (ptr = ANTIC_STORE_LIST (smexpr); ptr; ptr = XEXP (ptr, 1))
6304 if (XEXP (ptr, 0) == del)
6306 XEXP (ptr, 0) = insn;
6310 /* Move the notes from the deleted insn to its replacement. */
6311 REG_NOTES (insn) = REG_NOTES (del);
6313 /* Emit the insn AFTER all the notes are transferred.
6314 This is cheaper since we avoid df rescanning for the note change. */
6315 insn = emit_insn_after (insn, del);
6320 "STORE_MOTION delete insn in BB %d:\n ", bb->index);
6321 print_inline_rtx (dump_file, del, 6);
6322 fprintf (dump_file, "\nSTORE MOTION replaced with insn:\n ");
6323 print_inline_rtx (dump_file, insn, 6);
6324 fprintf (dump_file, "\n");
6329 /* Now we must handle REG_EQUAL notes whose contents is equal to the mem;
6330 they are no longer accurate provided that they are reached by this
6331 definition, so drop them. */
6332 for (; insn != NEXT_INSN (BB_END (bb)); insn = NEXT_INSN (insn))
6335 set = single_set (insn);
6338 if (expr_equiv_p (SET_DEST (set), mem))
6340 note = find_reg_equal_equiv_note (insn);
6341 if (!note || !expr_equiv_p (XEXP (note, 0), mem))
6345 fprintf (dump_file, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
6347 remove_note (insn, note);
6349 remove_reachable_equiv_notes (bb, smexpr);
6353 /* Delete a store, but copy the value that would have been stored into
6354 the reaching_reg for later storing. */
6357 delete_store (struct ls_expr * expr, basic_block bb)
6361 if (expr->reaching_reg == NULL_RTX)
6362 expr->reaching_reg = gen_reg_rtx_and_attrs (expr->pattern);
6364 reg = expr->reaching_reg;
6366 for (i = AVAIL_STORE_LIST (expr); i; i = XEXP (i, 1))
6369 if (BLOCK_FOR_INSN (del) == bb)
6371 /* We know there is only one since we deleted redundant
6372 ones during the available computation. */
6373 replace_store_insn (reg, del, bb, expr);
6379 /* Free memory used by store motion. */
6382 free_store_memory (void)
6387 sbitmap_vector_free (ae_gen);
6389 sbitmap_vector_free (ae_kill);
6391 sbitmap_vector_free (transp);
6393 sbitmap_vector_free (st_antloc);
6395 sbitmap_vector_free (pre_insert_map);
6397 sbitmap_vector_free (pre_delete_map);
6398 if (reg_set_in_block)
6399 sbitmap_vector_free (reg_set_in_block);
6401 ae_gen = ae_kill = transp = st_antloc = NULL;
6402 pre_insert_map = pre_delete_map = reg_set_in_block = NULL;
6405 /* Perform store motion. Much like gcse, except we move expressions the
6406 other way by looking at the flowgraph in reverse. */
6413 struct ls_expr * ptr;
6414 int update_flow = 0;
6418 fprintf (dump_file, "before store motion\n");
6419 print_rtl (dump_file, get_insns ());
6422 init_alias_analysis ();
6424 /* Find all the available and anticipatable stores. */
6425 num_stores = compute_store_table ();
6426 if (num_stores == 0)
6428 htab_delete (pre_ldst_table);
6429 pre_ldst_table = NULL;
6430 sbitmap_vector_free (reg_set_in_block);
6431 end_alias_analysis ();
6435 /* Now compute kill & transp vectors. */
6436 build_store_vectors ();
6437 add_noreturn_fake_exit_edges ();
6438 connect_infinite_loops_to_exit ();
6440 edge_list = pre_edge_rev_lcm (num_stores, transp, ae_gen,
6441 st_antloc, ae_kill, &pre_insert_map,
6444 /* Now we want to insert the new stores which are going to be needed. */
6445 for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
6447 /* If any of the edges we have above are abnormal, we can't move this
6449 for (x = NUM_EDGES (edge_list) - 1; x >= 0; x--)
6450 if (TEST_BIT (pre_insert_map[x], ptr->index)
6451 && (INDEX_EDGE (edge_list, x)->flags & EDGE_ABNORMAL))
6456 if (dump_file != NULL)
6458 "Can't replace store %d: abnormal edge from %d to %d\n",
6459 ptr->index, INDEX_EDGE (edge_list, x)->src->index,
6460 INDEX_EDGE (edge_list, x)->dest->index);
6464 /* Now we want to insert the new stores which are going to be needed. */
6467 if (TEST_BIT (pre_delete_map[bb->index], ptr->index))
6468 delete_store (ptr, bb);
6470 for (x = 0; x < NUM_EDGES (edge_list); x++)
6471 if (TEST_BIT (pre_insert_map[x], ptr->index))
6472 update_flow |= insert_store (ptr, INDEX_EDGE (edge_list, x));
6476 commit_edge_insertions ();
6478 free_store_memory ();
6479 free_edge_list (edge_list);
6480 remove_fake_exit_edges ();
6481 end_alias_analysis ();
6485 /* Entry point for jump bypassing optimization pass. */
6492 /* We do not construct an accurate cfg in functions which call
6493 setjmp, so just punt to be safe. */
6494 if (cfun->calls_setjmp)
6497 /* Identify the basic block information for this function, including
6498 successors and predecessors. */
6499 max_gcse_regno = max_reg_num ();
6502 dump_flow_info (dump_file, dump_flags);
6504 /* Return if there's nothing to do, or it is too expensive. */
6505 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
6506 || is_too_expensive (_ ("jump bypassing disabled")))
6509 gcc_obstack_init (&gcse_obstack);
6512 /* We need alias. */
6513 init_alias_analysis ();
6515 /* Record where pseudo-registers are set. This data is kept accurate
6516 during each pass. ??? We could also record hard-reg information here
6517 [since it's unchanging], however it is currently done during hash table
6520 It may be tempting to compute MEM set information here too, but MEM sets
6521 will be subject to code motion one day and thus we need to compute
6522 information about memory sets when we build the hash tables. */
6524 alloc_reg_set_mem (max_gcse_regno);
6527 max_gcse_regno = max_reg_num ();
6529 changed = one_cprop_pass (3, true, true);
6534 fprintf (dump_file, "BYPASS of %s: %d basic blocks, ",
6535 current_function_name (), n_basic_blocks);
6536 fprintf (dump_file, "%d bytes\n\n", bytes_used);
6539 obstack_free (&gcse_obstack, NULL);
6540 free_reg_set_mem ();
6542 /* We are finished with alias. */
6543 end_alias_analysis ();
6548 /* Return true if the graph is too expensive to optimize. PASS is the
6549 optimization about to be performed. */
6552 is_too_expensive (const char *pass)
6554 /* Trying to perform global optimizations on flow graphs which have
6555 a high connectivity will take a long time and is unlikely to be
6556 particularly useful.
6558 In normal circumstances a cfg should have about twice as many
6559 edges as blocks. But we do not want to punish small functions
6560 which have a couple switch statements. Rather than simply
6561 threshold the number of blocks, uses something with a more
6562 graceful degradation. */
6563 if (n_edges > 20000 + n_basic_blocks * 4)
6565 warning (OPT_Wdisabled_optimization,
6566 "%s: %d basic blocks and %d edges/basic block",
6567 pass, n_basic_blocks, n_edges / n_basic_blocks);
6572 /* If allocating memory for the cprop bitmap would take up too much
6573 storage it's better just to disable the optimization. */
6575 * SBITMAP_SET_SIZE (max_reg_num ())
6576 * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
6578 warning (OPT_Wdisabled_optimization,
6579 "%s: %d basic blocks and %d registers",
6580 pass, n_basic_blocks, max_reg_num ());
6589 gate_handle_jump_bypass (void)
6591 return optimize > 0 && flag_gcse
6592 && dbg_cnt (jump_bypass);
6595 /* Perform jump bypassing and control flow optimizations. */
6597 rest_of_handle_jump_bypass (void)
6599 delete_unreachable_blocks ();
6600 if (bypass_jumps ())
6602 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6603 rebuild_jump_labels (get_insns ());
6609 struct rtl_opt_pass pass_jump_bypass =
6613 "bypass", /* name */
6614 gate_handle_jump_bypass, /* gate */
6615 rest_of_handle_jump_bypass, /* execute */
6618 0, /* static_pass_number */
6619 TV_BYPASS, /* tv_id */
6620 PROP_cfglayout, /* properties_required */
6621 0, /* properties_provided */
6622 0, /* properties_destroyed */
6623 0, /* todo_flags_start */
6625 TODO_ggc_collect | TODO_verify_flow /* todo_flags_finish */
6631 gate_handle_gcse (void)
6633 return optimize > 0 && flag_gcse
6639 rest_of_handle_gcse (void)
6641 int save_csb, save_cfj;
6643 tem = gcse_main (get_insns ());
6644 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6645 rebuild_jump_labels (get_insns ());
6646 save_csb = flag_cse_skip_blocks;
6647 save_cfj = flag_cse_follow_jumps;
6648 flag_cse_skip_blocks = flag_cse_follow_jumps = 0;
6650 /* If -fexpensive-optimizations, re-run CSE to clean up things done
6652 if (flag_expensive_optimizations)
6654 timevar_push (TV_CSE);
6655 tem2 = cse_main (get_insns (), max_reg_num ());
6656 df_finish_pass (false);
6657 purge_all_dead_edges ();
6658 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6659 timevar_pop (TV_CSE);
6660 cse_not_expected = !flag_rerun_cse_after_loop;
6663 /* If gcse or cse altered any jumps, rerun jump optimizations to clean
6665 if (tem || tem2 == 2)
6667 timevar_push (TV_JUMP);
6668 rebuild_jump_labels (get_insns ());
6670 timevar_pop (TV_JUMP);
6675 flag_cse_skip_blocks = save_csb;
6676 flag_cse_follow_jumps = save_cfj;
6680 struct rtl_opt_pass pass_gcse =
6685 gate_handle_gcse, /* gate */
6686 rest_of_handle_gcse, /* execute */
6689 0, /* static_pass_number */
6690 TV_GCSE, /* tv_id */
6691 PROP_cfglayout, /* properties_required */
6692 0, /* properties_provided */
6693 0, /* properties_destroyed */
6694 0, /* todo_flags_start */
6695 TODO_df_finish | TODO_verify_rtl_sharing |
6697 TODO_verify_flow | TODO_ggc_collect /* todo_flags_finish */
6702 #include "gt-gcse.h"