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 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 basic block information.
195 2) Compute table of places where registers are set.
197 3) Perform copy/constant propagation.
199 4) Perform global cse using lazy code motion if not optimizing
200 for size, or code hoisting if we are.
202 5) Perform another pass of copy/constant propagation.
204 Two passes of copy/constant propagation are done because the first one
205 enables more GCSE and the second one helps to clean up the copies that
206 GCSE creates. This is needed more for PRE than for Classic because Classic
207 GCSE will try to use an existing register containing the common
208 subexpression rather than create a new one. This is harder to do for PRE
209 because of the code motion (which Classic GCSE doesn't do).
211 Expressions we are interested in GCSE-ing are of the form
212 (set (pseudo-reg) (expression)).
213 Function want_to_gcse_p says what these are.
215 PRE handles moving invariant expressions out of loops (by treating them as
216 partially redundant).
218 Eventually it would be nice to replace cse.c/gcse.c with SSA (static single
219 assignment) based GVN (global value numbering). L. T. Simpson's paper
220 (Rice University) on value numbering is a useful reference for this.
222 **********************
224 We used to support multiple passes but there are diminishing returns in
225 doing so. The first pass usually makes 90% of the changes that are doable.
226 A second pass can make a few more changes made possible by the first pass.
227 Experiments show any further passes don't make enough changes to justify
230 A study of spec92 using an unlimited number of passes:
231 [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
232 [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
233 [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
235 It was found doing copy propagation between each pass enables further
238 PRE is quite expensive in complicated functions because the DFA can take
239 a while to converge. Hence we only perform one pass. The parameter
240 max-gcse-passes can be modified if one wants to experiment.
242 **********************
244 The steps for PRE are:
246 1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
248 2) Perform the data flow analysis for PRE.
250 3) Delete the redundant instructions
252 4) Insert the required copies [if any] that make the partially
253 redundant instructions fully redundant.
255 5) For other reaching expressions, insert an instruction to copy the value
256 to a newly created pseudo that will reach the redundant instruction.
258 The deletion is done first so that when we do insertions we
259 know which pseudo reg to use.
261 Various papers have argued that PRE DFA is expensive (O(n^2)) and others
262 argue it is not. The number of iterations for the algorithm to converge
263 is typically 2-4 so I don't view it as that expensive (relatively speaking).
265 PRE GCSE depends heavily on the second CSE pass to clean up the copies
266 we create. To make an expression reach the place where it's redundant,
267 the result of the expression is copied to a new register, and the redundant
268 expression is deleted by replacing it with this new register. Classic GCSE
269 doesn't have this problem as much as it computes the reaching defs of
270 each register in each block and thus can try to use an existing
273 /* GCSE global vars. */
275 /* Note whether or not we should run jump optimization after gcse. We
276 want to do this for two cases.
278 * If we changed any jumps via cprop.
280 * If we added any labels via edge splitting. */
281 static int run_jump_opt_after_gcse
;
283 /* An obstack for our working variables. */
284 static struct obstack gcse_obstack
;
286 struct reg_use
{rtx reg_rtx
; };
288 /* Hash table of expressions. */
292 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
294 /* Index in the available expression bitmaps. */
296 /* Next entry with the same hash. */
297 struct expr
*next_same_hash
;
298 /* List of anticipatable occurrences in basic blocks in the function.
299 An "anticipatable occurrence" is one that is the first occurrence in the
300 basic block, the operands are not modified in the basic block prior
301 to the occurrence and the output is not used between the start of
302 the block and the occurrence. */
303 struct occr
*antic_occr
;
304 /* List of available occurrence in basic blocks in the function.
305 An "available occurrence" is one that is the last occurrence in the
306 basic block and the operands are not modified by following statements in
307 the basic block [including this insn]. */
308 struct occr
*avail_occr
;
309 /* Non-null if the computation is PRE redundant.
310 The value is the newly created pseudo-reg to record a copy of the
311 expression in all the places that reach the redundant copy. */
315 /* Occurrence of an expression.
316 There is one per basic block. If a pattern appears more than once the
317 last appearance is used [or first for anticipatable expressions]. */
321 /* Next occurrence of this expression. */
323 /* The insn that computes the expression. */
325 /* Nonzero if this [anticipatable] occurrence has been deleted. */
327 /* Nonzero if this [available] occurrence has been copied to
329 /* ??? This is mutually exclusive with deleted_p, so they could share
334 /* Expression and copy propagation hash tables.
335 Each hash table is an array of buckets.
336 ??? It is known that if it were an array of entries, structure elements
337 `next_same_hash' and `bitmap_index' wouldn't be necessary. However, it is
338 not clear whether in the final analysis a sufficient amount of memory would
339 be saved as the size of the available expression bitmaps would be larger
340 [one could build a mapping table without holes afterwards though].
341 Someday I'll perform the computation and figure it out. */
346 This is an array of `expr_hash_table_size' elements. */
349 /* Size of the hash table, in elements. */
352 /* Number of hash table elements. */
353 unsigned int n_elems
;
355 /* Whether the table is expression of copy propagation one. */
359 /* Expression hash table. */
360 static struct hash_table expr_hash_table
;
362 /* Copy propagation hash table. */
363 static struct hash_table set_hash_table
;
365 /* Mapping of uids to cuids.
366 Only real insns get cuids. */
367 static int *uid_cuid
;
369 /* Highest UID in UID_CUID. */
372 /* Get the cuid of an insn. */
373 #ifdef ENABLE_CHECKING
374 #define INSN_CUID(INSN) \
375 (gcc_assert (INSN_UID (INSN) <= max_uid), uid_cuid[INSN_UID (INSN)])
377 #define INSN_CUID(INSN) (uid_cuid[INSN_UID (INSN)])
380 /* Number of cuids. */
383 /* Mapping of cuids to insns. */
384 static rtx
*cuid_insn
;
386 /* Get insn from cuid. */
387 #define CUID_INSN(CUID) (cuid_insn[CUID])
389 /* Maximum register number in function prior to doing gcse + 1.
390 Registers created during this pass have regno >= max_gcse_regno.
391 This is named with "gcse" to not collide with global of same name. */
392 static unsigned int max_gcse_regno
;
394 /* Table of registers that are modified.
396 For each register, each element is a list of places where the pseudo-reg
399 For simplicity, GCSE is done on sets of pseudo-regs only. PRE GCSE only
400 requires knowledge of which blocks kill which regs [and thus could use
401 a bitmap instead of the lists `reg_set_table' uses].
403 `reg_set_table' and could be turned into an array of bitmaps (num-bbs x
404 num-regs) [however perhaps it may be useful to keep the data as is]. One
405 advantage of recording things this way is that `reg_set_table' is fairly
406 sparse with respect to pseudo regs but for hard regs could be fairly dense
407 [relatively speaking]. And recording sets of pseudo-regs in lists speeds
408 up functions like compute_transp since in the case of pseudo-regs we only
409 need to iterate over the number of times a pseudo-reg is set, not over the
410 number of basic blocks [clearly there is a bit of a slow down in the cases
411 where a pseudo is set more than once in a block, however it is believed
412 that the net effect is to speed things up]. This isn't done for hard-regs
413 because recording call-clobbered hard-regs in `reg_set_table' at each
414 function call can consume a fair bit of memory, and iterating over
415 hard-regs stored this way in compute_transp will be more expensive. */
417 typedef struct reg_set
419 /* The next setting of this register. */
420 struct reg_set
*next
;
421 /* The index of the block where it was set. */
425 static reg_set
**reg_set_table
;
427 /* Size of `reg_set_table'.
428 The table starts out at max_gcse_regno + slop, and is enlarged as
430 static int reg_set_table_size
;
432 /* Amount to grow `reg_set_table' by when it's full. */
433 #define REG_SET_TABLE_SLOP 100
435 /* This is a list of expressions which are MEMs and will be used by load
437 Load motion tracks MEMs which aren't killed by
438 anything except itself. (i.e., loads and stores to a single location).
439 We can then allow movement of these MEM refs with a little special
440 allowance. (all stores copy the same value to the reaching reg used
441 for the loads). This means all values used to store into memory must have
442 no side effects so we can re-issue the setter value.
443 Store Motion uses this structure as an expression table to track stores
444 which look interesting, and might be moveable towards the exit block. */
448 struct expr
* expr
; /* Gcse expression reference for LM. */
449 rtx pattern
; /* Pattern of this mem. */
450 rtx pattern_regs
; /* List of registers mentioned by the mem. */
451 rtx loads
; /* INSN list of loads seen. */
452 rtx stores
; /* INSN list of stores seen. */
453 struct ls_expr
* next
; /* Next in the list. */
454 int invalid
; /* Invalid for some reason. */
455 int index
; /* If it maps to a bitmap index. */
456 unsigned int hash_index
; /* Index when in a hash table. */
457 rtx reaching_reg
; /* Register to use when re-writing. */
460 /* Array of implicit set patterns indexed by basic block index. */
461 static rtx
*implicit_sets
;
463 /* Head of the list of load/store memory refs. */
464 static struct ls_expr
* pre_ldst_mems
= NULL
;
466 /* Hashtable for the load/store memory refs. */
467 static htab_t pre_ldst_table
= NULL
;
469 /* Bitmap containing one bit for each register in the program.
470 Used when performing GCSE to track which registers have been set since
471 the start of the basic block. */
472 static regset reg_set_bitmap
;
474 /* For each block, a bitmap of registers set in the block.
475 This is used by compute_transp.
476 It is computed during hash table computation and not by compute_sets
477 as it includes registers added since the last pass (or between cprop and
478 gcse) and it's currently not easy to realloc sbitmap vectors. */
479 static sbitmap
*reg_set_in_block
;
481 /* Array, indexed by basic block number for a list of insns which modify
482 memory within that block. */
483 static rtx
* modify_mem_list
;
484 static bitmap modify_mem_list_set
;
486 /* This array parallels modify_mem_list, but is kept canonicalized. */
487 static rtx
* canon_modify_mem_list
;
489 /* Bitmap indexed by block numbers to record which blocks contain
491 static bitmap blocks_with_calls
;
493 /* Various variables for statistics gathering. */
495 /* Memory used in a pass.
496 This isn't intended to be absolutely precise. Its intent is only
497 to keep an eye on memory usage. */
498 static int bytes_used
;
500 /* GCSE substitutions made. */
501 static int gcse_subst_count
;
502 /* Number of copy instructions created. */
503 static int gcse_create_count
;
504 /* Number of local constants propagated. */
505 static int local_const_prop_count
;
506 /* Number of local copies propagated. */
507 static int local_copy_prop_count
;
508 /* Number of global constants propagated. */
509 static int global_const_prop_count
;
510 /* Number of global copies propagated. */
511 static int global_copy_prop_count
;
513 /* For available exprs */
514 static sbitmap
*ae_kill
, *ae_gen
;
516 static void compute_can_copy (void);
517 static void *gmalloc (size_t) ATTRIBUTE_MALLOC
;
518 static void *gcalloc (size_t, size_t) ATTRIBUTE_MALLOC
;
519 static void *grealloc (void *, size_t);
520 static void *gcse_alloc (unsigned long);
521 static void alloc_gcse_mem (void);
522 static void free_gcse_mem (void);
523 static void alloc_reg_set_mem (int);
524 static void free_reg_set_mem (void);
525 static void record_one_set (int, rtx
);
526 static void record_set_info (rtx
, const_rtx
, void *);
527 static void compute_sets (void);
528 static void hash_scan_insn (rtx
, struct hash_table
*, int);
529 static void hash_scan_set (rtx
, rtx
, struct hash_table
*);
530 static void hash_scan_clobber (rtx
, rtx
, struct hash_table
*);
531 static void hash_scan_call (rtx
, rtx
, struct hash_table
*);
532 static int want_to_gcse_p (rtx
);
533 static bool can_assign_to_reg_p (rtx
);
534 static bool gcse_constant_p (const_rtx
);
535 static int oprs_unchanged_p (const_rtx
, const_rtx
, int);
536 static int oprs_anticipatable_p (const_rtx
, const_rtx
);
537 static int oprs_available_p (const_rtx
, const_rtx
);
538 static void insert_expr_in_table (rtx
, enum machine_mode
, rtx
, int, int,
539 struct hash_table
*);
540 static void insert_set_in_table (rtx
, rtx
, struct hash_table
*);
541 static unsigned int hash_expr (const_rtx
, enum machine_mode
, int *, int);
542 static unsigned int hash_set (int, int);
543 static int expr_equiv_p (const_rtx
, const_rtx
);
544 static void record_last_reg_set_info (rtx
, int);
545 static void record_last_mem_set_info (rtx
);
546 static void record_last_set_info (rtx
, const_rtx
, void *);
547 static void compute_hash_table (struct hash_table
*);
548 static void alloc_hash_table (int, struct hash_table
*, int);
549 static void free_hash_table (struct hash_table
*);
550 static void compute_hash_table_work (struct hash_table
*);
551 static void dump_hash_table (FILE *, const char *, struct hash_table
*);
552 static struct expr
*lookup_set (unsigned int, struct hash_table
*);
553 static struct expr
*next_set (unsigned int, struct expr
*);
554 static void reset_opr_set_tables (void);
555 static int oprs_not_set_p (const_rtx
, const_rtx
);
556 static void mark_call (rtx
);
557 static void mark_set (rtx
, rtx
);
558 static void mark_clobber (rtx
, rtx
);
559 static void mark_oprs_set (rtx
);
560 static void alloc_cprop_mem (int, int);
561 static void free_cprop_mem (void);
562 static void compute_transp (const_rtx
, int, sbitmap
*, int);
563 static void compute_transpout (void);
564 static void compute_local_properties (sbitmap
*, sbitmap
*, sbitmap
*,
565 struct hash_table
*);
566 static void compute_cprop_data (void);
567 static void find_used_regs (rtx
*, void *);
568 static int try_replace_reg (rtx
, rtx
, rtx
);
569 static struct expr
*find_avail_set (int, rtx
);
570 static int cprop_jump (basic_block
, rtx
, rtx
, rtx
, rtx
);
571 static void mems_conflict_for_gcse_p (rtx
, const_rtx
, void *);
572 static int load_killed_in_block_p (const_basic_block
, int, const_rtx
, int);
573 static void canon_list_insert (rtx
, const_rtx
, void *);
574 static int cprop_insn (rtx
, int);
575 static int cprop (int);
576 static void find_implicit_sets (void);
577 static int one_cprop_pass (int, bool, bool);
578 static bool constprop_register (rtx
, rtx
, rtx
, bool);
579 static struct expr
*find_bypass_set (int, int);
580 static bool reg_killed_on_edge (const_rtx
, const_edge
);
581 static int bypass_block (basic_block
, rtx
, rtx
);
582 static int bypass_conditional_jumps (void);
583 static void alloc_pre_mem (int, int);
584 static void free_pre_mem (void);
585 static void compute_pre_data (void);
586 static int pre_expr_reaches_here_p (basic_block
, struct expr
*,
588 static void insert_insn_end_basic_block (struct expr
*, basic_block
, int);
589 static void pre_insert_copy_insn (struct expr
*, rtx
);
590 static void pre_insert_copies (void);
591 static int pre_delete (void);
592 static int pre_gcse (void);
593 static int one_pre_gcse_pass (int);
594 static void add_label_notes (rtx
, rtx
);
595 static void alloc_code_hoist_mem (int, int);
596 static void free_code_hoist_mem (void);
597 static void compute_code_hoist_vbeinout (void);
598 static void compute_code_hoist_data (void);
599 static int hoist_expr_reaches_here_p (basic_block
, int, basic_block
, char *);
600 static void hoist_code (void);
601 static int one_code_hoisting_pass (void);
602 static rtx
process_insert_insn (struct expr
*);
603 static int pre_edge_insert (struct edge_list
*, struct expr
**);
604 static int pre_expr_reaches_here_p_work (basic_block
, struct expr
*,
605 basic_block
, char *);
606 static struct ls_expr
* ldst_entry (rtx
);
607 static void free_ldst_entry (struct ls_expr
*);
608 static void free_ldst_mems (void);
609 static void print_ldst_list (FILE *);
610 static struct ls_expr
* find_rtx_in_ldst (rtx
);
611 static int enumerate_ldsts (void);
612 static inline struct ls_expr
* first_ls_expr (void);
613 static inline struct ls_expr
* next_ls_expr (struct ls_expr
*);
614 static int simple_mem (const_rtx
);
615 static void invalidate_any_buried_refs (rtx
);
616 static void compute_ld_motion_mems (void);
617 static void trim_ld_motion_mems (void);
618 static void update_ld_motion_stores (struct expr
*);
619 static void reg_set_info (rtx
, const_rtx
, void *);
620 static void reg_clear_last_set (rtx
, const_rtx
, void *);
621 static bool store_ops_ok (const_rtx
, int *);
622 static rtx
extract_mentioned_regs (rtx
);
623 static rtx
extract_mentioned_regs_helper (rtx
, rtx
);
624 static void find_moveable_store (rtx
, int *, int *);
625 static int compute_store_table (void);
626 static bool load_kills_store (const_rtx
, const_rtx
, int);
627 static bool find_loads (const_rtx
, const_rtx
, int);
628 static bool store_killed_in_insn (const_rtx
, const_rtx
, const_rtx
, int);
629 static bool store_killed_after (const_rtx
, const_rtx
, const_rtx
, const_basic_block
, int *, rtx
*);
630 static bool store_killed_before (const_rtx
, const_rtx
, const_rtx
, const_basic_block
, int *);
631 static void build_store_vectors (void);
632 static void insert_insn_start_basic_block (rtx
, basic_block
);
633 static int insert_store (struct ls_expr
*, edge
);
634 static void remove_reachable_equiv_notes (basic_block
, struct ls_expr
*);
635 static void replace_store_insn (rtx
, rtx
, basic_block
, struct ls_expr
*);
636 static void delete_store (struct ls_expr
*, basic_block
);
637 static void free_store_memory (void);
638 static void store_motion (void);
639 static void free_insn_expr_list_list (rtx
*);
640 static void clear_modify_mem_tables (void);
641 static void free_modify_mem_tables (void);
642 static rtx
gcse_emit_move_after (rtx
, rtx
, rtx
);
643 static void local_cprop_find_used_regs (rtx
*, void *);
644 static bool do_local_cprop (rtx
, rtx
, bool, rtx
*);
645 static bool adjust_libcall_notes (rtx
, rtx
, rtx
, rtx
*);
646 static void local_cprop_pass (bool);
647 static bool is_too_expensive (const char *);
650 /* Entry point for global common subexpression elimination.
651 F is the first instruction in the function. Return nonzero if a
655 gcse_main (rtx f ATTRIBUTE_UNUSED
)
658 /* Bytes used at start of pass. */
659 int initial_bytes_used
;
660 /* Maximum number of bytes used by a pass. */
662 /* Point to release obstack data from for each pass. */
663 char *gcse_obstack_bottom
;
665 /* We do not construct an accurate cfg in functions which call
666 setjmp, so just punt to be safe. */
667 if (current_function_calls_setjmp
)
670 /* Assume that we do not need to run jump optimizations after gcse. */
671 run_jump_opt_after_gcse
= 0;
673 /* Identify the basic block information for this function, including
674 successors and predecessors. */
675 max_gcse_regno
= max_reg_num ();
677 df_note_add_problem ();
681 dump_flow_info (dump_file
, dump_flags
);
683 /* Return if there's nothing to do, or it is too expensive. */
684 if (n_basic_blocks
<= NUM_FIXED_BLOCKS
+ 1
685 || is_too_expensive (_("GCSE disabled")))
688 gcc_obstack_init (&gcse_obstack
);
692 init_alias_analysis ();
693 /* Record where pseudo-registers are set. This data is kept accurate
694 during each pass. ??? We could also record hard-reg information here
695 [since it's unchanging], however it is currently done during hash table
698 It may be tempting to compute MEM set information here too, but MEM sets
699 will be subject to code motion one day and thus we need to compute
700 information about memory sets when we build the hash tables. */
702 alloc_reg_set_mem (max_gcse_regno
);
706 initial_bytes_used
= bytes_used
;
708 gcse_obstack_bottom
= gcse_alloc (1);
710 while (changed
&& pass
< MAX_GCSE_PASSES
)
714 fprintf (dump_file
, "GCSE pass %d\n\n", pass
+ 1);
716 /* Initialize bytes_used to the space for the pred/succ lists,
717 and the reg_set_table data. */
718 bytes_used
= initial_bytes_used
;
720 /* Each pass may create new registers, so recalculate each time. */
721 max_gcse_regno
= max_reg_num ();
725 /* Don't allow constant propagation to modify jumps
727 timevar_push (TV_CPROP1
);
728 changed
= one_cprop_pass (pass
+ 1, false, false);
729 timevar_pop (TV_CPROP1
);
735 timevar_push (TV_PRE
);
736 changed
|= one_pre_gcse_pass (pass
+ 1);
737 /* We may have just created new basic blocks. Release and
738 recompute various things which are sized on the number of
742 free_modify_mem_tables ();
743 modify_mem_list
= gcalloc (last_basic_block
, sizeof (rtx
));
744 canon_modify_mem_list
= gcalloc (last_basic_block
, sizeof (rtx
));
747 alloc_reg_set_mem (max_reg_num ());
749 run_jump_opt_after_gcse
= 1;
750 timevar_pop (TV_PRE
);
753 if (max_pass_bytes
< bytes_used
)
754 max_pass_bytes
= bytes_used
;
756 /* Free up memory, then reallocate for code hoisting. We can
757 not re-use the existing allocated memory because the tables
758 will not have info for the insns or registers created by
759 partial redundancy elimination. */
762 /* It does not make sense to run code hoisting unless we are optimizing
763 for code size -- it rarely makes programs faster, and can make
764 them bigger if we did partial redundancy elimination (when optimizing
765 for space, we don't run the partial redundancy algorithms). */
768 timevar_push (TV_HOIST
);
769 max_gcse_regno
= max_reg_num ();
771 changed
|= one_code_hoisting_pass ();
774 if (max_pass_bytes
< bytes_used
)
775 max_pass_bytes
= bytes_used
;
776 timevar_pop (TV_HOIST
);
781 fprintf (dump_file
, "\n");
785 obstack_free (&gcse_obstack
, gcse_obstack_bottom
);
789 /* Do one last pass of copy propagation, including cprop into
790 conditional jumps. */
792 max_gcse_regno
= max_reg_num ();
794 /* This time, go ahead and allow cprop to alter jumps. */
795 timevar_push (TV_CPROP2
);
796 one_cprop_pass (pass
+ 1, true, true);
797 timevar_pop (TV_CPROP2
);
802 fprintf (dump_file
, "GCSE of %s: %d basic blocks, ",
803 current_function_name (), n_basic_blocks
);
804 fprintf (dump_file
, "%d pass%s, %d bytes\n\n",
805 pass
, pass
> 1 ? "es" : "", max_pass_bytes
);
808 obstack_free (&gcse_obstack
, NULL
);
811 /* We are finished with alias. */
812 end_alias_analysis ();
814 if (!optimize_size
&& flag_gcse_sm
)
816 timevar_push (TV_LSM
);
818 timevar_pop (TV_LSM
);
821 /* Record where pseudo-registers are set. */
822 return run_jump_opt_after_gcse
;
825 /* Misc. utilities. */
827 /* Nonzero for each mode that supports (set (reg) (reg)).
828 This is trivially true for integer and floating point values.
829 It may or may not be true for condition codes. */
830 static char can_copy
[(int) NUM_MACHINE_MODES
];
832 /* Compute which modes support reg/reg copy operations. */
835 compute_can_copy (void)
838 #ifndef AVOID_CCMODE_COPIES
841 memset (can_copy
, 0, NUM_MACHINE_MODES
);
844 for (i
= 0; i
< NUM_MACHINE_MODES
; i
++)
845 if (GET_MODE_CLASS (i
) == MODE_CC
)
847 #ifdef AVOID_CCMODE_COPIES
850 reg
= gen_rtx_REG ((enum machine_mode
) i
, LAST_VIRTUAL_REGISTER
+ 1);
851 insn
= emit_insn (gen_rtx_SET (VOIDmode
, reg
, reg
));
852 if (recog (PATTERN (insn
), insn
, NULL
) >= 0)
862 /* Returns whether the mode supports reg/reg copy operations. */
865 can_copy_p (enum machine_mode mode
)
867 static bool can_copy_init_p
= false;
869 if (! can_copy_init_p
)
872 can_copy_init_p
= true;
875 return can_copy
[mode
] != 0;
878 /* Cover function to xmalloc to record bytes allocated. */
881 gmalloc (size_t size
)
884 return xmalloc (size
);
887 /* Cover function to xcalloc to record bytes allocated. */
890 gcalloc (size_t nelem
, size_t elsize
)
892 bytes_used
+= nelem
* elsize
;
893 return xcalloc (nelem
, elsize
);
896 /* Cover function to xrealloc.
897 We don't record the additional size since we don't know it.
898 It won't affect memory usage stats much anyway. */
901 grealloc (void *ptr
, size_t size
)
903 return xrealloc (ptr
, size
);
906 /* Cover function to obstack_alloc. */
909 gcse_alloc (unsigned long size
)
912 return obstack_alloc (&gcse_obstack
, size
);
915 /* Allocate memory for the cuid mapping array,
916 and reg/memory set tracking tables.
918 This is called at the start of each pass. */
921 alloc_gcse_mem (void)
927 /* Find the largest UID and create a mapping from UIDs to CUIDs.
928 CUIDs are like UIDs except they increase monotonically, have no gaps,
929 and only apply to real insns.
930 (Actually, there are gaps, for insn that are not inside a basic block.
931 but we should never see those anyway, so this is OK.) */
933 max_uid
= get_max_uid ();
934 uid_cuid
= gcalloc (max_uid
+ 1, sizeof (int));
937 FOR_BB_INSNS (bb
, insn
)
940 uid_cuid
[INSN_UID (insn
)] = i
++;
942 uid_cuid
[INSN_UID (insn
)] = i
;
945 /* Create a table mapping cuids to insns. */
948 cuid_insn
= gcalloc (max_cuid
+ 1, sizeof (rtx
));
951 FOR_BB_INSNS (bb
, insn
)
953 CUID_INSN (i
++) = insn
;
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
= gcalloc (last_basic_block
, sizeof (rtx
));
963 canon_modify_mem_list
= gcalloc (last_basic_block
, sizeof (rtx
));
964 modify_mem_list_set
= BITMAP_ALLOC (NULL
);
965 blocks_with_calls
= BITMAP_ALLOC (NULL
);
968 /* Free memory allocated by alloc_gcse_mem. */
976 BITMAP_FREE (reg_set_bitmap
);
978 sbitmap_vector_free (reg_set_in_block
);
979 free_modify_mem_tables ();
980 BITMAP_FREE (modify_mem_list_set
);
981 BITMAP_FREE (blocks_with_calls
);
984 /* Compute the local properties of each recorded expression.
986 Local properties are those that are defined by the block, irrespective of
989 An expression is transparent in a block if its operands are not modified
992 An expression is computed (locally available) in a block if it is computed
993 at least once and expression would contain the same value if the
994 computation was moved to the end of the block.
996 An expression is locally anticipatable in a block if it is computed at
997 least once and expression would contain the same value if the computation
998 was moved to the beginning of the block.
1000 We call this routine for cprop, pre and code hoisting. They all compute
1001 basically the same information and thus can easily share this code.
1003 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
1004 properties. If NULL, then it is not necessary to compute or record that
1005 particular property.
1007 TABLE controls which hash table to look at. If it is set hash table,
1008 additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
1012 compute_local_properties (sbitmap
*transp
, sbitmap
*comp
, sbitmap
*antloc
,
1013 struct hash_table
*table
)
1017 /* Initialize any bitmaps that were passed in. */
1021 sbitmap_vector_zero (transp
, last_basic_block
);
1023 sbitmap_vector_ones (transp
, last_basic_block
);
1027 sbitmap_vector_zero (comp
, last_basic_block
);
1029 sbitmap_vector_zero (antloc
, last_basic_block
);
1031 for (i
= 0; i
< table
->size
; i
++)
1035 for (expr
= table
->table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
1037 int indx
= expr
->bitmap_index
;
1040 /* The expression is transparent in this block if it is not killed.
1041 We start by assuming all are transparent [none are killed], and
1042 then reset the bits for those that are. */
1044 compute_transp (expr
->expr
, indx
, transp
, table
->set_p
);
1046 /* The occurrences recorded in antic_occr are exactly those that
1047 we want to set to nonzero in ANTLOC. */
1049 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
1051 SET_BIT (antloc
[BLOCK_NUM (occr
->insn
)], indx
);
1053 /* While we're scanning the table, this is a good place to
1055 occr
->deleted_p
= 0;
1058 /* The occurrences recorded in avail_occr are exactly those that
1059 we want to set to nonzero in COMP. */
1061 for (occr
= expr
->avail_occr
; occr
!= NULL
; occr
= occr
->next
)
1063 SET_BIT (comp
[BLOCK_NUM (occr
->insn
)], indx
);
1065 /* While we're scanning the table, this is a good place to
1070 /* While we're scanning the table, this is a good place to
1072 expr
->reaching_reg
= 0;
1077 /* Register set information.
1079 `reg_set_table' records where each register is set or otherwise
1082 static struct obstack reg_set_obstack
;
1085 alloc_reg_set_mem (int n_regs
)
1087 reg_set_table_size
= n_regs
+ REG_SET_TABLE_SLOP
;
1088 reg_set_table
= gcalloc (reg_set_table_size
, sizeof (struct reg_set
*));
1090 gcc_obstack_init (®_set_obstack
);
1094 free_reg_set_mem (void)
1096 free (reg_set_table
);
1097 obstack_free (®_set_obstack
, NULL
);
1100 /* Record REGNO in the reg_set table. */
1103 record_one_set (int regno
, rtx insn
)
1105 /* Allocate a new reg_set element and link it onto the list. */
1106 struct reg_set
*new_reg_info
;
1108 /* If the table isn't big enough, enlarge it. */
1109 if (regno
>= reg_set_table_size
)
1111 int new_size
= regno
+ REG_SET_TABLE_SLOP
;
1113 reg_set_table
= grealloc (reg_set_table
,
1114 new_size
* sizeof (struct reg_set
*));
1115 memset (reg_set_table
+ reg_set_table_size
, 0,
1116 (new_size
- reg_set_table_size
) * sizeof (struct reg_set
*));
1117 reg_set_table_size
= new_size
;
1120 new_reg_info
= obstack_alloc (®_set_obstack
, sizeof (struct reg_set
));
1121 bytes_used
+= sizeof (struct reg_set
);
1122 new_reg_info
->bb_index
= BLOCK_NUM (insn
);
1123 new_reg_info
->next
= reg_set_table
[regno
];
1124 reg_set_table
[regno
] = new_reg_info
;
1127 /* Called from compute_sets via note_stores to handle one SET or CLOBBER in
1128 an insn. The DATA is really the instruction in which the SET is
1132 record_set_info (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
, void *data
)
1134 rtx record_set_insn
= (rtx
) data
;
1136 if (REG_P (dest
) && REGNO (dest
) >= FIRST_PSEUDO_REGISTER
)
1137 record_one_set (REGNO (dest
), record_set_insn
);
1140 /* Scan the function and record each set of each pseudo-register.
1142 This is called once, at the start of the gcse pass. See the comments for
1143 `reg_set_table' for further documentation. */
1152 FOR_BB_INSNS (bb
, insn
)
1154 note_stores (PATTERN (insn
), record_set_info
, insn
);
1157 /* Hash table support. */
1159 struct reg_avail_info
1161 basic_block last_bb
;
1166 static struct reg_avail_info
*reg_avail_info
;
1167 static basic_block current_bb
;
1170 /* See whether X, the source of a set, is something we want to consider for
1174 want_to_gcse_p (rtx x
)
1177 /* On register stack architectures, don't GCSE constants from the
1178 constant pool, as the benefits are often swamped by the overhead
1179 of shuffling the register stack between basic blocks. */
1180 if (IS_STACK_MODE (GET_MODE (x
)))
1181 x
= avoid_constant_pool_reference (x
);
1184 switch (GET_CODE (x
))
1196 return can_assign_to_reg_p (x
);
1200 /* Used internally by can_assign_to_reg_p. */
1202 static GTY(()) rtx test_insn
;
1204 /* Return true if we can assign X to a pseudo register. */
1207 can_assign_to_reg_p (rtx x
)
1209 int num_clobbers
= 0;
1212 /* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
1213 if (general_operand (x
, GET_MODE (x
)))
1215 else if (GET_MODE (x
) == VOIDmode
)
1218 /* Otherwise, check if we can make a valid insn from it. First initialize
1219 our test insn if we haven't already. */
1223 = make_insn_raw (gen_rtx_SET (VOIDmode
,
1224 gen_rtx_REG (word_mode
,
1225 FIRST_PSEUDO_REGISTER
* 2),
1227 NEXT_INSN (test_insn
) = PREV_INSN (test_insn
) = 0;
1230 /* Now make an insn like the one we would make when GCSE'ing and see if
1232 PUT_MODE (SET_DEST (PATTERN (test_insn
)), GET_MODE (x
));
1233 SET_SRC (PATTERN (test_insn
)) = x
;
1234 return ((icode
= recog (PATTERN (test_insn
), test_insn
, &num_clobbers
)) >= 0
1235 && (num_clobbers
== 0 || ! added_clobbers_hard_reg_p (icode
)));
1238 /* Return nonzero if the operands of expression X are unchanged from the
1239 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
1240 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
1243 oprs_unchanged_p (const_rtx x
, const_rtx insn
, int avail_p
)
1252 code
= GET_CODE (x
);
1257 struct reg_avail_info
*info
= ®_avail_info
[REGNO (x
)];
1259 if (info
->last_bb
!= current_bb
)
1262 return info
->last_set
< INSN_CUID (insn
);
1264 return info
->first_set
>= INSN_CUID (insn
);
1268 if (load_killed_in_block_p (current_bb
, INSN_CUID (insn
),
1272 return oprs_unchanged_p (XEXP (x
, 0), insn
, avail_p
);
1299 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
1303 /* If we are about to do the last recursive call needed at this
1304 level, change it into iteration. This function is called enough
1307 return oprs_unchanged_p (XEXP (x
, i
), insn
, avail_p
);
1309 else if (! oprs_unchanged_p (XEXP (x
, i
), insn
, avail_p
))
1312 else if (fmt
[i
] == 'E')
1313 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
1314 if (! oprs_unchanged_p (XVECEXP (x
, i
, j
), insn
, avail_p
))
1321 /* Used for communication between mems_conflict_for_gcse_p and
1322 load_killed_in_block_p. Nonzero if mems_conflict_for_gcse_p finds a
1323 conflict between two memory references. */
1324 static int gcse_mems_conflict_p
;
1326 /* Used for communication between mems_conflict_for_gcse_p and
1327 load_killed_in_block_p. A memory reference for a load instruction,
1328 mems_conflict_for_gcse_p will see if a memory store conflicts with
1329 this memory load. */
1330 static const_rtx gcse_mem_operand
;
1332 /* DEST is the output of an instruction. If it is a memory reference, and
1333 possibly conflicts with the load found in gcse_mem_operand, then set
1334 gcse_mems_conflict_p to a nonzero value. */
1337 mems_conflict_for_gcse_p (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
,
1338 void *data ATTRIBUTE_UNUSED
)
1340 while (GET_CODE (dest
) == SUBREG
1341 || GET_CODE (dest
) == ZERO_EXTRACT
1342 || GET_CODE (dest
) == STRICT_LOW_PART
)
1343 dest
= XEXP (dest
, 0);
1345 /* If DEST is not a MEM, then it will not conflict with the load. Note
1346 that function calls are assumed to clobber memory, but are handled
1351 /* If we are setting a MEM in our list of specially recognized MEMs,
1352 don't mark as killed this time. */
1354 if (expr_equiv_p (dest
, gcse_mem_operand
) && pre_ldst_mems
!= NULL
)
1356 if (!find_rtx_in_ldst (dest
))
1357 gcse_mems_conflict_p
= 1;
1361 if (true_dependence (dest
, GET_MODE (dest
), gcse_mem_operand
,
1363 gcse_mems_conflict_p
= 1;
1366 /* Return nonzero if the expression in X (a memory reference) is killed
1367 in block BB before or after the insn with the CUID in UID_LIMIT.
1368 AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills
1371 To check the entire block, set UID_LIMIT to max_uid + 1 and
1375 load_killed_in_block_p (const_basic_block bb
, int uid_limit
, const_rtx x
, int avail_p
)
1377 rtx list_entry
= modify_mem_list
[bb
->index
];
1379 /* If this is a readonly then we aren't going to be changing it. */
1380 if (MEM_READONLY_P (x
))
1386 /* Ignore entries in the list that do not apply. */
1388 && INSN_CUID (XEXP (list_entry
, 0)) < uid_limit
)
1390 && INSN_CUID (XEXP (list_entry
, 0)) > uid_limit
))
1392 list_entry
= XEXP (list_entry
, 1);
1396 setter
= XEXP (list_entry
, 0);
1398 /* If SETTER is a call everything is clobbered. Note that calls
1399 to pure functions are never put on the list, so we need not
1400 worry about them. */
1401 if (CALL_P (setter
))
1404 /* SETTER must be an INSN of some kind that sets memory. Call
1405 note_stores to examine each hunk of memory that is modified.
1407 The note_stores interface is pretty limited, so we have to
1408 communicate via global variables. Yuk. */
1409 gcse_mem_operand
= x
;
1410 gcse_mems_conflict_p
= 0;
1411 note_stores (PATTERN (setter
), mems_conflict_for_gcse_p
, NULL
);
1412 if (gcse_mems_conflict_p
)
1414 list_entry
= XEXP (list_entry
, 1);
1419 /* Return nonzero if the operands of expression X are unchanged from
1420 the start of INSN's basic block up to but not including INSN. */
1423 oprs_anticipatable_p (const_rtx x
, const_rtx insn
)
1425 return oprs_unchanged_p (x
, insn
, 0);
1428 /* Return nonzero if the operands of expression X are unchanged from
1429 INSN to the end of INSN's basic block. */
1432 oprs_available_p (const_rtx x
, const_rtx insn
)
1434 return oprs_unchanged_p (x
, insn
, 1);
1437 /* Hash expression X.
1439 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1440 indicating if a volatile operand is found or if the expression contains
1441 something we don't want to insert in the table. HASH_TABLE_SIZE is
1442 the current size of the hash table to be probed. */
1445 hash_expr (const_rtx x
, enum machine_mode mode
, int *do_not_record_p
,
1446 int hash_table_size
)
1450 *do_not_record_p
= 0;
1452 hash
= hash_rtx (x
, mode
, do_not_record_p
,
1453 NULL
, /*have_reg_qty=*/false);
1454 return hash
% hash_table_size
;
1457 /* Hash a set of register REGNO.
1459 Sets are hashed on the register that is set. This simplifies the PRE copy
1462 ??? May need to make things more elaborate. Later, as necessary. */
1465 hash_set (int regno
, int hash_table_size
)
1470 return hash
% hash_table_size
;
1473 /* Return nonzero if exp1 is equivalent to exp2. */
1476 expr_equiv_p (const_rtx x
, const_rtx y
)
1478 return exp_equiv_p (x
, y
, 0, true);
1481 /* Insert expression X in INSN in the hash TABLE.
1482 If it is already present, record it as the last occurrence in INSN's
1485 MODE is the mode of the value X is being stored into.
1486 It is only used if X is a CONST_INT.
1488 ANTIC_P is nonzero if X is an anticipatable expression.
1489 AVAIL_P is nonzero if X is an available expression. */
1492 insert_expr_in_table (rtx x
, enum machine_mode mode
, rtx insn
, int antic_p
,
1493 int avail_p
, struct hash_table
*table
)
1495 int found
, do_not_record_p
;
1497 struct expr
*cur_expr
, *last_expr
= NULL
;
1498 struct occr
*antic_occr
, *avail_occr
;
1500 hash
= hash_expr (x
, mode
, &do_not_record_p
, table
->size
);
1502 /* Do not insert expression in table if it contains volatile operands,
1503 or if hash_expr determines the expression is something we don't want
1504 to or can't handle. */
1505 if (do_not_record_p
)
1508 cur_expr
= table
->table
[hash
];
1511 while (cur_expr
&& 0 == (found
= expr_equiv_p (cur_expr
->expr
, x
)))
1513 /* If the expression isn't found, save a pointer to the end of
1515 last_expr
= cur_expr
;
1516 cur_expr
= cur_expr
->next_same_hash
;
1521 cur_expr
= gcse_alloc (sizeof (struct expr
));
1522 bytes_used
+= sizeof (struct expr
);
1523 if (table
->table
[hash
] == NULL
)
1524 /* This is the first pattern that hashed to this index. */
1525 table
->table
[hash
] = cur_expr
;
1527 /* Add EXPR to end of this hash chain. */
1528 last_expr
->next_same_hash
= cur_expr
;
1530 /* Set the fields of the expr element. */
1532 cur_expr
->bitmap_index
= table
->n_elems
++;
1533 cur_expr
->next_same_hash
= NULL
;
1534 cur_expr
->antic_occr
= NULL
;
1535 cur_expr
->avail_occr
= NULL
;
1538 /* Now record the occurrence(s). */
1541 antic_occr
= cur_expr
->antic_occr
;
1543 if (antic_occr
&& BLOCK_NUM (antic_occr
->insn
) != BLOCK_NUM (insn
))
1547 /* Found another instance of the expression in the same basic block.
1548 Prefer the currently recorded one. We want the first one in the
1549 block and the block is scanned from start to end. */
1550 ; /* nothing to do */
1553 /* First occurrence of this expression in this basic block. */
1554 antic_occr
= gcse_alloc (sizeof (struct occr
));
1555 bytes_used
+= sizeof (struct occr
);
1556 antic_occr
->insn
= insn
;
1557 antic_occr
->next
= cur_expr
->antic_occr
;
1558 antic_occr
->deleted_p
= 0;
1559 cur_expr
->antic_occr
= antic_occr
;
1565 avail_occr
= cur_expr
->avail_occr
;
1567 if (avail_occr
&& BLOCK_NUM (avail_occr
->insn
) == BLOCK_NUM (insn
))
1569 /* Found another instance of the expression in the same basic block.
1570 Prefer this occurrence to the currently recorded one. We want
1571 the last one in the block and the block is scanned from start
1573 avail_occr
->insn
= insn
;
1577 /* First occurrence of this expression in this basic block. */
1578 avail_occr
= gcse_alloc (sizeof (struct occr
));
1579 bytes_used
+= sizeof (struct occr
);
1580 avail_occr
->insn
= insn
;
1581 avail_occr
->next
= cur_expr
->avail_occr
;
1582 avail_occr
->deleted_p
= 0;
1583 cur_expr
->avail_occr
= avail_occr
;
1588 /* Insert pattern X in INSN in the hash table.
1589 X is a SET of a reg to either another reg or a constant.
1590 If it is already present, record it as the last occurrence in INSN's
1594 insert_set_in_table (rtx x
, rtx insn
, struct hash_table
*table
)
1598 struct expr
*cur_expr
, *last_expr
= NULL
;
1599 struct occr
*cur_occr
;
1601 gcc_assert (GET_CODE (x
) == SET
&& REG_P (SET_DEST (x
)));
1603 hash
= hash_set (REGNO (SET_DEST (x
)), table
->size
);
1605 cur_expr
= table
->table
[hash
];
1608 while (cur_expr
&& 0 == (found
= expr_equiv_p (cur_expr
->expr
, x
)))
1610 /* If the expression isn't found, save a pointer to the end of
1612 last_expr
= cur_expr
;
1613 cur_expr
= cur_expr
->next_same_hash
;
1618 cur_expr
= gcse_alloc (sizeof (struct expr
));
1619 bytes_used
+= sizeof (struct expr
);
1620 if (table
->table
[hash
] == NULL
)
1621 /* This is the first pattern that hashed to this index. */
1622 table
->table
[hash
] = cur_expr
;
1624 /* Add EXPR to end of this hash chain. */
1625 last_expr
->next_same_hash
= cur_expr
;
1627 /* Set the fields of the expr element.
1628 We must copy X because it can be modified when copy propagation is
1629 performed on its operands. */
1630 cur_expr
->expr
= copy_rtx (x
);
1631 cur_expr
->bitmap_index
= table
->n_elems
++;
1632 cur_expr
->next_same_hash
= NULL
;
1633 cur_expr
->antic_occr
= NULL
;
1634 cur_expr
->avail_occr
= NULL
;
1637 /* Now record the occurrence. */
1638 cur_occr
= cur_expr
->avail_occr
;
1640 if (cur_occr
&& BLOCK_NUM (cur_occr
->insn
) == BLOCK_NUM (insn
))
1642 /* Found another instance of the expression in the same basic block.
1643 Prefer this occurrence to the currently recorded one. We want
1644 the last one in the block and the block is scanned from start
1646 cur_occr
->insn
= insn
;
1650 /* First occurrence of this expression in this basic block. */
1651 cur_occr
= gcse_alloc (sizeof (struct occr
));
1652 bytes_used
+= sizeof (struct occr
);
1654 cur_occr
->insn
= insn
;
1655 cur_occr
->next
= cur_expr
->avail_occr
;
1656 cur_occr
->deleted_p
= 0;
1657 cur_expr
->avail_occr
= cur_occr
;
1661 /* Determine whether the rtx X should be treated as a constant for
1662 the purposes of GCSE's constant propagation. */
1665 gcse_constant_p (const_rtx x
)
1667 /* Consider a COMPARE of two integers constant. */
1668 if (GET_CODE (x
) == COMPARE
1669 && GET_CODE (XEXP (x
, 0)) == CONST_INT
1670 && GET_CODE (XEXP (x
, 1)) == CONST_INT
)
1673 /* Consider a COMPARE of the same registers is a constant
1674 if they are not floating point registers. */
1675 if (GET_CODE(x
) == COMPARE
1676 && REG_P (XEXP (x
, 0)) && REG_P (XEXP (x
, 1))
1677 && REGNO (XEXP (x
, 0)) == REGNO (XEXP (x
, 1))
1678 && ! FLOAT_MODE_P (GET_MODE (XEXP (x
, 0)))
1679 && ! FLOAT_MODE_P (GET_MODE (XEXP (x
, 1))))
1682 return CONSTANT_P (x
);
1685 /* Scan pattern PAT of INSN and add an entry to the hash TABLE (set or
1689 hash_scan_set (rtx pat
, rtx insn
, struct hash_table
*table
)
1691 rtx src
= SET_SRC (pat
);
1692 rtx dest
= SET_DEST (pat
);
1695 if (GET_CODE (src
) == CALL
)
1696 hash_scan_call (src
, insn
, table
);
1698 else if (REG_P (dest
))
1700 unsigned int regno
= REGNO (dest
);
1703 /* See if a REG_NOTE shows this equivalent to a simpler expression.
1704 This allows us to do a single GCSE pass and still eliminate
1705 redundant constants, addresses or other expressions that are
1706 constructed with multiple instructions. */
1707 note
= find_reg_equal_equiv_note (insn
);
1710 ? gcse_constant_p (XEXP (note
, 0))
1711 : want_to_gcse_p (XEXP (note
, 0))))
1712 src
= XEXP (note
, 0), pat
= gen_rtx_SET (VOIDmode
, dest
, src
);
1714 /* Only record sets of pseudo-regs in the hash table. */
1716 && regno
>= FIRST_PSEUDO_REGISTER
1717 /* Don't GCSE something if we can't do a reg/reg copy. */
1718 && can_copy_p (GET_MODE (dest
))
1719 /* GCSE commonly inserts instruction after the insn. We can't
1720 do that easily for EH_REGION notes so disable GCSE on these
1722 && !find_reg_note (insn
, REG_EH_REGION
, NULL_RTX
)
1723 /* Is SET_SRC something we want to gcse? */
1724 && want_to_gcse_p (src
)
1725 /* Don't CSE a nop. */
1726 && ! set_noop_p (pat
)
1727 /* Don't GCSE if it has attached REG_EQUIV note.
1728 At this point this only function parameters should have
1729 REG_EQUIV notes and if the argument slot is used somewhere
1730 explicitly, it means address of parameter has been taken,
1731 so we should not extend the lifetime of the pseudo. */
1732 && (note
== NULL_RTX
|| ! MEM_P (XEXP (note
, 0))))
1734 /* An expression is not anticipatable if its operands are
1735 modified before this insn or if this is not the only SET in
1736 this insn. The latter condition does not have to mean that
1737 SRC itself is not anticipatable, but we just will not be
1738 able to handle code motion of insns with multiple sets. */
1739 int antic_p
= oprs_anticipatable_p (src
, insn
)
1740 && !multiple_sets (insn
);
1741 /* An expression is not available if its operands are
1742 subsequently modified, including this insn. It's also not
1743 available if this is a branch, because we can't insert
1744 a set after the branch. */
1745 int avail_p
= (oprs_available_p (src
, insn
)
1746 && ! JUMP_P (insn
));
1748 insert_expr_in_table (src
, GET_MODE (dest
), insn
, antic_p
, avail_p
, table
);
1751 /* Record sets for constant/copy propagation. */
1752 else if (table
->set_p
1753 && regno
>= FIRST_PSEUDO_REGISTER
1755 && REGNO (src
) >= FIRST_PSEUDO_REGISTER
1756 && can_copy_p (GET_MODE (dest
))
1757 && REGNO (src
) != regno
)
1758 || gcse_constant_p (src
))
1759 /* A copy is not available if its src or dest is subsequently
1760 modified. Here we want to search from INSN+1 on, but
1761 oprs_available_p searches from INSN on. */
1762 && (insn
== BB_END (BLOCK_FOR_INSN (insn
))
1763 || (tmp
= next_nonnote_insn (insn
)) == NULL_RTX
1764 || BLOCK_FOR_INSN (tmp
) != BLOCK_FOR_INSN (insn
)
1765 || oprs_available_p (pat
, tmp
)))
1766 insert_set_in_table (pat
, insn
, table
);
1768 /* In case of store we want to consider the memory value as available in
1769 the REG stored in that memory. This makes it possible to remove
1770 redundant loads from due to stores to the same location. */
1771 else if (flag_gcse_las
&& REG_P (src
) && MEM_P (dest
))
1773 unsigned int regno
= REGNO (src
);
1775 /* Do not do this for constant/copy propagation. */
1777 /* Only record sets of pseudo-regs in the hash table. */
1778 && regno
>= FIRST_PSEUDO_REGISTER
1779 /* Don't GCSE something if we can't do a reg/reg copy. */
1780 && can_copy_p (GET_MODE (src
))
1781 /* GCSE commonly inserts instruction after the insn. We can't
1782 do that easily for EH_REGION notes so disable GCSE on these
1784 && ! find_reg_note (insn
, REG_EH_REGION
, NULL_RTX
)
1785 /* Is SET_DEST something we want to gcse? */
1786 && want_to_gcse_p (dest
)
1787 /* Don't CSE a nop. */
1788 && ! set_noop_p (pat
)
1789 /* Don't GCSE if it has attached REG_EQUIV note.
1790 At this point this only function parameters should have
1791 REG_EQUIV notes and if the argument slot is used somewhere
1792 explicitly, it means address of parameter has been taken,
1793 so we should not extend the lifetime of the pseudo. */
1794 && ((note
= find_reg_note (insn
, REG_EQUIV
, NULL_RTX
)) == 0
1795 || ! MEM_P (XEXP (note
, 0))))
1797 /* Stores are never anticipatable. */
1799 /* An expression is not available if its operands are
1800 subsequently modified, including this insn. It's also not
1801 available if this is a branch, because we can't insert
1802 a set after the branch. */
1803 int avail_p
= oprs_available_p (dest
, insn
)
1806 /* Record the memory expression (DEST) in the hash table. */
1807 insert_expr_in_table (dest
, GET_MODE (dest
), insn
,
1808 antic_p
, avail_p
, table
);
1814 hash_scan_clobber (rtx x ATTRIBUTE_UNUSED
, rtx insn ATTRIBUTE_UNUSED
,
1815 struct hash_table
*table ATTRIBUTE_UNUSED
)
1817 /* Currently nothing to do. */
1821 hash_scan_call (rtx x ATTRIBUTE_UNUSED
, rtx insn ATTRIBUTE_UNUSED
,
1822 struct hash_table
*table ATTRIBUTE_UNUSED
)
1824 /* Currently nothing to do. */
1827 /* Process INSN and add hash table entries as appropriate.
1829 Only available expressions that set a single pseudo-reg are recorded.
1831 Single sets in a PARALLEL could be handled, but it's an extra complication
1832 that isn't dealt with right now. The trick is handling the CLOBBERs that
1833 are also in the PARALLEL. Later.
1835 If SET_P is nonzero, this is for the assignment hash table,
1836 otherwise it is for the expression hash table.
1837 If IN_LIBCALL_BLOCK nonzero, we are in a libcall block, and should
1838 not record any expressions. */
1841 hash_scan_insn (rtx insn
, struct hash_table
*table
, int in_libcall_block
)
1843 rtx pat
= PATTERN (insn
);
1846 if (in_libcall_block
)
1849 /* Pick out the sets of INSN and for other forms of instructions record
1850 what's been modified. */
1852 if (GET_CODE (pat
) == SET
)
1853 hash_scan_set (pat
, insn
, table
);
1854 else if (GET_CODE (pat
) == PARALLEL
)
1855 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
1857 rtx x
= XVECEXP (pat
, 0, i
);
1859 if (GET_CODE (x
) == SET
)
1860 hash_scan_set (x
, insn
, table
);
1861 else if (GET_CODE (x
) == CLOBBER
)
1862 hash_scan_clobber (x
, insn
, table
);
1863 else if (GET_CODE (x
) == CALL
)
1864 hash_scan_call (x
, insn
, table
);
1867 else if (GET_CODE (pat
) == CLOBBER
)
1868 hash_scan_clobber (pat
, insn
, table
);
1869 else if (GET_CODE (pat
) == CALL
)
1870 hash_scan_call (pat
, insn
, table
);
1874 dump_hash_table (FILE *file
, const char *name
, struct hash_table
*table
)
1877 /* Flattened out table, so it's printed in proper order. */
1878 struct expr
**flat_table
;
1879 unsigned int *hash_val
;
1882 flat_table
= xcalloc (table
->n_elems
, sizeof (struct expr
*));
1883 hash_val
= xmalloc (table
->n_elems
* sizeof (unsigned int));
1885 for (i
= 0; i
< (int) table
->size
; i
++)
1886 for (expr
= table
->table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
1888 flat_table
[expr
->bitmap_index
] = expr
;
1889 hash_val
[expr
->bitmap_index
] = i
;
1892 fprintf (file
, "%s hash table (%d buckets, %d entries)\n",
1893 name
, table
->size
, table
->n_elems
);
1895 for (i
= 0; i
< (int) table
->n_elems
; i
++)
1896 if (flat_table
[i
] != 0)
1898 expr
= flat_table
[i
];
1899 fprintf (file
, "Index %d (hash value %d)\n ",
1900 expr
->bitmap_index
, hash_val
[i
]);
1901 print_rtl (file
, expr
->expr
);
1902 fprintf (file
, "\n");
1905 fprintf (file
, "\n");
1911 /* Record register first/last/block set information for REGNO in INSN.
1913 first_set records the first place in the block where the register
1914 is set and is used to compute "anticipatability".
1916 last_set records the last place in the block where the register
1917 is set and is used to compute "availability".
1919 last_bb records the block for which first_set and last_set are
1920 valid, as a quick test to invalidate them.
1922 reg_set_in_block records whether the register is set in the block
1923 and is used to compute "transparency". */
1926 record_last_reg_set_info (rtx insn
, int regno
)
1928 struct reg_avail_info
*info
= ®_avail_info
[regno
];
1929 int cuid
= INSN_CUID (insn
);
1931 info
->last_set
= cuid
;
1932 if (info
->last_bb
!= current_bb
)
1934 info
->last_bb
= current_bb
;
1935 info
->first_set
= cuid
;
1936 SET_BIT (reg_set_in_block
[current_bb
->index
], regno
);
1941 /* Record all of the canonicalized MEMs of record_last_mem_set_info's insn.
1942 Note we store a pair of elements in the list, so they have to be
1943 taken off pairwise. */
1946 canon_list_insert (rtx dest ATTRIBUTE_UNUSED
, const_rtx unused1 ATTRIBUTE_UNUSED
,
1949 rtx dest_addr
, insn
;
1952 while (GET_CODE (dest
) == SUBREG
1953 || GET_CODE (dest
) == ZERO_EXTRACT
1954 || GET_CODE (dest
) == STRICT_LOW_PART
)
1955 dest
= XEXP (dest
, 0);
1957 /* If DEST is not a MEM, then it will not conflict with a load. Note
1958 that function calls are assumed to clobber memory, but are handled
1964 dest_addr
= get_addr (XEXP (dest
, 0));
1965 dest_addr
= canon_rtx (dest_addr
);
1966 insn
= (rtx
) v_insn
;
1967 bb
= BLOCK_NUM (insn
);
1969 canon_modify_mem_list
[bb
] =
1970 alloc_EXPR_LIST (VOIDmode
, dest_addr
, canon_modify_mem_list
[bb
]);
1971 canon_modify_mem_list
[bb
] =
1972 alloc_EXPR_LIST (VOIDmode
, dest
, canon_modify_mem_list
[bb
]);
1975 /* Record memory modification information for INSN. We do not actually care
1976 about the memory location(s) that are set, or even how they are set (consider
1977 a CALL_INSN). We merely need to record which insns modify memory. */
1980 record_last_mem_set_info (rtx insn
)
1982 int bb
= BLOCK_NUM (insn
);
1984 /* load_killed_in_block_p will handle the case of calls clobbering
1986 modify_mem_list
[bb
] = alloc_INSN_LIST (insn
, modify_mem_list
[bb
]);
1987 bitmap_set_bit (modify_mem_list_set
, bb
);
1991 /* Note that traversals of this loop (other than for free-ing)
1992 will break after encountering a CALL_INSN. So, there's no
1993 need to insert a pair of items, as canon_list_insert does. */
1994 canon_modify_mem_list
[bb
] =
1995 alloc_INSN_LIST (insn
, canon_modify_mem_list
[bb
]);
1996 bitmap_set_bit (blocks_with_calls
, bb
);
1999 note_stores (PATTERN (insn
), canon_list_insert
, (void*) insn
);
2002 /* Called from compute_hash_table via note_stores to handle one
2003 SET or CLOBBER in an insn. DATA is really the instruction in which
2004 the SET is taking place. */
2007 record_last_set_info (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
, void *data
)
2009 rtx last_set_insn
= (rtx
) data
;
2011 if (GET_CODE (dest
) == SUBREG
)
2012 dest
= SUBREG_REG (dest
);
2015 record_last_reg_set_info (last_set_insn
, REGNO (dest
));
2016 else if (MEM_P (dest
)
2017 /* Ignore pushes, they clobber nothing. */
2018 && ! push_operand (dest
, GET_MODE (dest
)))
2019 record_last_mem_set_info (last_set_insn
);
2022 /* Top level function to create an expression or assignment hash table.
2024 Expression entries are placed in the hash table if
2025 - they are of the form (set (pseudo-reg) src),
2026 - src is something we want to perform GCSE on,
2027 - none of the operands are subsequently modified in the block
2029 Assignment entries are placed in the hash table if
2030 - they are of the form (set (pseudo-reg) src),
2031 - src is something we want to perform const/copy propagation on,
2032 - none of the operands or target are subsequently modified in the block
2034 Currently src must be a pseudo-reg or a const_int.
2036 TABLE is the table computed. */
2039 compute_hash_table_work (struct hash_table
*table
)
2043 /* While we compute the hash table we also compute a bit array of which
2044 registers are set in which blocks.
2045 ??? This isn't needed during const/copy propagation, but it's cheap to
2047 sbitmap_vector_zero (reg_set_in_block
, last_basic_block
);
2049 /* re-Cache any INSN_LIST nodes we have allocated. */
2050 clear_modify_mem_tables ();
2051 /* Some working arrays used to track first and last set in each block. */
2052 reg_avail_info
= gmalloc (max_gcse_regno
* sizeof (struct reg_avail_info
));
2054 for (i
= 0; i
< max_gcse_regno
; ++i
)
2055 reg_avail_info
[i
].last_bb
= NULL
;
2057 FOR_EACH_BB (current_bb
)
2061 int in_libcall_block
;
2063 /* First pass over the instructions records information used to
2064 determine when registers and memory are first and last set.
2065 ??? hard-reg reg_set_in_block computation
2066 could be moved to compute_sets since they currently don't change. */
2068 FOR_BB_INSNS (current_bb
, insn
)
2070 if (! INSN_P (insn
))
2075 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
2076 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
2077 record_last_reg_set_info (insn
, regno
);
2082 note_stores (PATTERN (insn
), record_last_set_info
, insn
);
2085 /* Insert implicit sets in the hash table. */
2087 && implicit_sets
[current_bb
->index
] != NULL_RTX
)
2088 hash_scan_set (implicit_sets
[current_bb
->index
],
2089 BB_HEAD (current_bb
), table
);
2091 /* The next pass builds the hash table. */
2092 in_libcall_block
= 0;
2093 FOR_BB_INSNS (current_bb
, insn
)
2096 if (find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
))
2097 in_libcall_block
= 1;
2098 else if (table
->set_p
&& find_reg_note (insn
, REG_RETVAL
, NULL_RTX
))
2099 in_libcall_block
= 0;
2100 hash_scan_insn (insn
, table
, in_libcall_block
);
2101 if (!table
->set_p
&& find_reg_note (insn
, REG_RETVAL
, NULL_RTX
))
2102 in_libcall_block
= 0;
2106 free (reg_avail_info
);
2107 reg_avail_info
= NULL
;
2110 /* Allocate space for the set/expr hash TABLE.
2111 N_INSNS is the number of instructions in the function.
2112 It is used to determine the number of buckets to use.
2113 SET_P determines whether set or expression table will
2117 alloc_hash_table (int n_insns
, struct hash_table
*table
, int set_p
)
2121 table
->size
= n_insns
/ 4;
2122 if (table
->size
< 11)
2125 /* Attempt to maintain efficient use of hash table.
2126 Making it an odd number is simplest for now.
2127 ??? Later take some measurements. */
2129 n
= table
->size
* sizeof (struct expr
*);
2130 table
->table
= gmalloc (n
);
2131 table
->set_p
= set_p
;
2134 /* Free things allocated by alloc_hash_table. */
2137 free_hash_table (struct hash_table
*table
)
2139 free (table
->table
);
2142 /* Compute the hash TABLE for doing copy/const propagation or
2143 expression hash table. */
2146 compute_hash_table (struct hash_table
*table
)
2148 /* Initialize count of number of entries in hash table. */
2150 memset (table
->table
, 0, table
->size
* sizeof (struct expr
*));
2152 compute_hash_table_work (table
);
2155 /* Expression tracking support. */
2157 /* Lookup REGNO in the set TABLE. The result is a pointer to the
2158 table entry, or NULL if not found. */
2160 static struct expr
*
2161 lookup_set (unsigned int regno
, struct hash_table
*table
)
2163 unsigned int hash
= hash_set (regno
, table
->size
);
2166 expr
= table
->table
[hash
];
2168 while (expr
&& REGNO (SET_DEST (expr
->expr
)) != regno
)
2169 expr
= expr
->next_same_hash
;
2174 /* Return the next entry for REGNO in list EXPR. */
2176 static struct expr
*
2177 next_set (unsigned int regno
, struct expr
*expr
)
2180 expr
= expr
->next_same_hash
;
2181 while (expr
&& REGNO (SET_DEST (expr
->expr
)) != regno
);
2186 /* Like free_INSN_LIST_list or free_EXPR_LIST_list, except that the node
2187 types may be mixed. */
2190 free_insn_expr_list_list (rtx
*listp
)
2194 for (list
= *listp
; list
; list
= next
)
2196 next
= XEXP (list
, 1);
2197 if (GET_CODE (list
) == EXPR_LIST
)
2198 free_EXPR_LIST_node (list
);
2200 free_INSN_LIST_node (list
);
2206 /* Clear canon_modify_mem_list and modify_mem_list tables. */
2208 clear_modify_mem_tables (void)
2213 EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set
, 0, i
, bi
)
2215 free_INSN_LIST_list (modify_mem_list
+ i
);
2216 free_insn_expr_list_list (canon_modify_mem_list
+ i
);
2218 bitmap_clear (modify_mem_list_set
);
2219 bitmap_clear (blocks_with_calls
);
2222 /* Release memory used by modify_mem_list_set. */
2225 free_modify_mem_tables (void)
2227 clear_modify_mem_tables ();
2228 free (modify_mem_list
);
2229 free (canon_modify_mem_list
);
2230 modify_mem_list
= 0;
2231 canon_modify_mem_list
= 0;
2234 /* Reset tables used to keep track of what's still available [since the
2235 start of the block]. */
2238 reset_opr_set_tables (void)
2240 /* Maintain a bitmap of which regs have been set since beginning of
2242 CLEAR_REG_SET (reg_set_bitmap
);
2244 /* Also keep a record of the last instruction to modify memory.
2245 For now this is very trivial, we only record whether any memory
2246 location has been modified. */
2247 clear_modify_mem_tables ();
2250 /* Return nonzero if the operands of X are not set before INSN in
2251 INSN's basic block. */
2254 oprs_not_set_p (const_rtx x
, const_rtx insn
)
2263 code
= GET_CODE (x
);
2280 if (load_killed_in_block_p (BLOCK_FOR_INSN (insn
),
2281 INSN_CUID (insn
), x
, 0))
2284 return oprs_not_set_p (XEXP (x
, 0), insn
);
2287 return ! REGNO_REG_SET_P (reg_set_bitmap
, REGNO (x
));
2293 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
2297 /* If we are about to do the last recursive call
2298 needed at this level, change it into iteration.
2299 This function is called enough to be worth it. */
2301 return oprs_not_set_p (XEXP (x
, i
), insn
);
2303 if (! oprs_not_set_p (XEXP (x
, i
), insn
))
2306 else if (fmt
[i
] == 'E')
2307 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
2308 if (! oprs_not_set_p (XVECEXP (x
, i
, j
), insn
))
2315 /* Mark things set by a CALL. */
2318 mark_call (rtx insn
)
2320 if (! CONST_OR_PURE_CALL_P (insn
))
2321 record_last_mem_set_info (insn
);
2324 /* Mark things set by a SET. */
2327 mark_set (rtx pat
, rtx insn
)
2329 rtx dest
= SET_DEST (pat
);
2331 while (GET_CODE (dest
) == SUBREG
2332 || GET_CODE (dest
) == ZERO_EXTRACT
2333 || GET_CODE (dest
) == STRICT_LOW_PART
)
2334 dest
= XEXP (dest
, 0);
2337 SET_REGNO_REG_SET (reg_set_bitmap
, REGNO (dest
));
2338 else if (MEM_P (dest
))
2339 record_last_mem_set_info (insn
);
2341 if (GET_CODE (SET_SRC (pat
)) == CALL
)
2345 /* Record things set by a CLOBBER. */
2348 mark_clobber (rtx pat
, rtx insn
)
2350 rtx clob
= XEXP (pat
, 0);
2352 while (GET_CODE (clob
) == SUBREG
|| GET_CODE (clob
) == STRICT_LOW_PART
)
2353 clob
= XEXP (clob
, 0);
2356 SET_REGNO_REG_SET (reg_set_bitmap
, REGNO (clob
));
2358 record_last_mem_set_info (insn
);
2361 /* Record things set by INSN.
2362 This data is used by oprs_not_set_p. */
2365 mark_oprs_set (rtx insn
)
2367 rtx pat
= PATTERN (insn
);
2370 if (GET_CODE (pat
) == SET
)
2371 mark_set (pat
, insn
);
2372 else if (GET_CODE (pat
) == PARALLEL
)
2373 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
2375 rtx x
= XVECEXP (pat
, 0, i
);
2377 if (GET_CODE (x
) == SET
)
2379 else if (GET_CODE (x
) == CLOBBER
)
2380 mark_clobber (x
, insn
);
2381 else if (GET_CODE (x
) == CALL
)
2385 else if (GET_CODE (pat
) == CLOBBER
)
2386 mark_clobber (pat
, insn
);
2387 else if (GET_CODE (pat
) == CALL
)
2392 /* Compute copy/constant propagation working variables. */
2394 /* Local properties of assignments. */
2395 static sbitmap
*cprop_pavloc
;
2396 static sbitmap
*cprop_absaltered
;
2398 /* Global properties of assignments (computed from the local properties). */
2399 static sbitmap
*cprop_avin
;
2400 static sbitmap
*cprop_avout
;
2402 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
2403 basic blocks. N_SETS is the number of sets. */
2406 alloc_cprop_mem (int n_blocks
, int n_sets
)
2408 cprop_pavloc
= sbitmap_vector_alloc (n_blocks
, n_sets
);
2409 cprop_absaltered
= sbitmap_vector_alloc (n_blocks
, n_sets
);
2411 cprop_avin
= sbitmap_vector_alloc (n_blocks
, n_sets
);
2412 cprop_avout
= sbitmap_vector_alloc (n_blocks
, n_sets
);
2415 /* Free vars used by copy/const propagation. */
2418 free_cprop_mem (void)
2420 sbitmap_vector_free (cprop_pavloc
);
2421 sbitmap_vector_free (cprop_absaltered
);
2422 sbitmap_vector_free (cprop_avin
);
2423 sbitmap_vector_free (cprop_avout
);
2426 /* For each block, compute whether X is transparent. X is either an
2427 expression or an assignment [though we don't care which, for this context
2428 an assignment is treated as an expression]. For each block where an
2429 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
2433 compute_transp (const_rtx x
, int indx
, sbitmap
*bmap
, int set_p
)
2441 /* repeat is used to turn tail-recursion into iteration since GCC
2442 can't do it when there's no return value. */
2448 code
= GET_CODE (x
);
2454 if (REGNO (x
) < FIRST_PSEUDO_REGISTER
)
2457 if (TEST_BIT (reg_set_in_block
[bb
->index
], REGNO (x
)))
2458 SET_BIT (bmap
[bb
->index
], indx
);
2462 for (r
= reg_set_table
[REGNO (x
)]; r
!= NULL
; r
= r
->next
)
2463 SET_BIT (bmap
[r
->bb_index
], indx
);
2468 if (REGNO (x
) < FIRST_PSEUDO_REGISTER
)
2471 if (TEST_BIT (reg_set_in_block
[bb
->index
], REGNO (x
)))
2472 RESET_BIT (bmap
[bb
->index
], indx
);
2476 for (r
= reg_set_table
[REGNO (x
)]; r
!= NULL
; r
= r
->next
)
2477 RESET_BIT (bmap
[r
->bb_index
], indx
);
2484 if (! MEM_READONLY_P (x
))
2489 /* First handle all the blocks with calls. We don't need to
2490 do any list walking for them. */
2491 EXECUTE_IF_SET_IN_BITMAP (blocks_with_calls
, 0, bb_index
, bi
)
2494 SET_BIT (bmap
[bb_index
], indx
);
2496 RESET_BIT (bmap
[bb_index
], indx
);
2499 /* Now iterate over the blocks which have memory modifications
2500 but which do not have any calls. */
2501 EXECUTE_IF_AND_COMPL_IN_BITMAP (modify_mem_list_set
,
2505 rtx list_entry
= canon_modify_mem_list
[bb_index
];
2509 rtx dest
, dest_addr
;
2511 /* LIST_ENTRY must be an INSN of some kind that sets memory.
2512 Examine each hunk of memory that is modified. */
2514 dest
= XEXP (list_entry
, 0);
2515 list_entry
= XEXP (list_entry
, 1);
2516 dest_addr
= XEXP (list_entry
, 0);
2518 if (canon_true_dependence (dest
, GET_MODE (dest
), dest_addr
,
2519 x
, rtx_addr_varies_p
))
2522 SET_BIT (bmap
[bb_index
], indx
);
2524 RESET_BIT (bmap
[bb_index
], indx
);
2527 list_entry
= XEXP (list_entry
, 1);
2552 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
2556 /* If we are about to do the last recursive call
2557 needed at this level, change it into iteration.
2558 This function is called enough to be worth it. */
2565 compute_transp (XEXP (x
, i
), indx
, bmap
, set_p
);
2567 else if (fmt
[i
] == 'E')
2568 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
2569 compute_transp (XVECEXP (x
, i
, j
), indx
, bmap
, set_p
);
2573 /* Top level routine to do the dataflow analysis needed by copy/const
2577 compute_cprop_data (void)
2579 compute_local_properties (cprop_absaltered
, cprop_pavloc
, NULL
, &set_hash_table
);
2580 compute_available (cprop_pavloc
, cprop_absaltered
,
2581 cprop_avout
, cprop_avin
);
2584 /* Copy/constant propagation. */
2586 /* Maximum number of register uses in an insn that we handle. */
2589 /* Table of uses found in an insn.
2590 Allocated statically to avoid alloc/free complexity and overhead. */
2591 static struct reg_use reg_use_table
[MAX_USES
];
2593 /* Index into `reg_use_table' while building it. */
2594 static int reg_use_count
;
2596 /* Set up a list of register numbers used in INSN. The found uses are stored
2597 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
2598 and contains the number of uses in the table upon exit.
2600 ??? If a register appears multiple times we will record it multiple times.
2601 This doesn't hurt anything but it will slow things down. */
2604 find_used_regs (rtx
*xptr
, void *data ATTRIBUTE_UNUSED
)
2611 /* repeat is used to turn tail-recursion into iteration since GCC
2612 can't do it when there's no return value. */
2617 code
= GET_CODE (x
);
2620 if (reg_use_count
== MAX_USES
)
2623 reg_use_table
[reg_use_count
].reg_rtx
= x
;
2627 /* Recursively scan the operands of this expression. */
2629 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
2633 /* If we are about to do the last recursive call
2634 needed at this level, change it into iteration.
2635 This function is called enough to be worth it. */
2642 find_used_regs (&XEXP (x
, i
), data
);
2644 else if (fmt
[i
] == 'E')
2645 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
2646 find_used_regs (&XVECEXP (x
, i
, j
), data
);
2650 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
2651 Returns nonzero is successful. */
2654 try_replace_reg (rtx from
, rtx to
, rtx insn
)
2656 rtx note
= find_reg_equal_equiv_note (insn
);
2659 rtx set
= single_set (insn
);
2661 /* Usually we substitute easy stuff, so we won't copy everything.
2662 We however need to take care to not duplicate non-trivial CONST
2666 validate_replace_src_group (from
, to
, insn
);
2667 if (num_changes_pending () && apply_change_group ())
2670 /* Try to simplify SET_SRC if we have substituted a constant. */
2671 if (success
&& set
&& CONSTANT_P (to
))
2673 src
= simplify_rtx (SET_SRC (set
));
2676 validate_change (insn
, &SET_SRC (set
), src
, 0);
2679 /* If there is already a REG_EQUAL note, update the expression in it
2680 with our replacement. */
2681 if (note
!= 0 && REG_NOTE_KIND (note
) == REG_EQUAL
)
2682 set_unique_reg_note (insn
, REG_EQUAL
,
2683 simplify_replace_rtx (XEXP (note
, 0), from
, to
));
2684 if (!success
&& set
&& reg_mentioned_p (from
, SET_SRC (set
)))
2686 /* If above failed and this is a single set, try to simplify the source of
2687 the set given our substitution. We could perhaps try this for multiple
2688 SETs, but it probably won't buy us anything. */
2689 src
= simplify_replace_rtx (SET_SRC (set
), from
, to
);
2691 if (!rtx_equal_p (src
, SET_SRC (set
))
2692 && validate_change (insn
, &SET_SRC (set
), src
, 0))
2695 /* If we've failed to do replacement, have a single SET, don't already
2696 have a note, and have no special SET, add a REG_EQUAL note to not
2697 lose information. */
2698 if (!success
&& note
== 0 && set
!= 0
2699 && GET_CODE (SET_DEST (set
)) != ZERO_EXTRACT
2700 && GET_CODE (SET_DEST (set
)) != STRICT_LOW_PART
)
2701 note
= set_unique_reg_note (insn
, REG_EQUAL
, copy_rtx (src
));
2704 /* REG_EQUAL may get simplified into register.
2705 We don't allow that. Remove that note. This code ought
2706 not to happen, because previous code ought to synthesize
2707 reg-reg move, but be on the safe side. */
2708 if (note
&& REG_NOTE_KIND (note
) == REG_EQUAL
&& REG_P (XEXP (note
, 0)))
2709 remove_note (insn
, note
);
2714 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
2715 NULL no such set is found. */
2717 static struct expr
*
2718 find_avail_set (int regno
, rtx insn
)
2720 /* SET1 contains the last set found that can be returned to the caller for
2721 use in a substitution. */
2722 struct expr
*set1
= 0;
2724 /* Loops are not possible here. To get a loop we would need two sets
2725 available at the start of the block containing INSN. i.e. we would
2726 need two sets like this available at the start of the block:
2728 (set (reg X) (reg Y))
2729 (set (reg Y) (reg X))
2731 This can not happen since the set of (reg Y) would have killed the
2732 set of (reg X) making it unavailable at the start of this block. */
2736 struct expr
*set
= lookup_set (regno
, &set_hash_table
);
2738 /* Find a set that is available at the start of the block
2739 which contains INSN. */
2742 if (TEST_BIT (cprop_avin
[BLOCK_NUM (insn
)], set
->bitmap_index
))
2744 set
= next_set (regno
, set
);
2747 /* If no available set was found we've reached the end of the
2748 (possibly empty) copy chain. */
2752 gcc_assert (GET_CODE (set
->expr
) == SET
);
2754 src
= SET_SRC (set
->expr
);
2756 /* We know the set is available.
2757 Now check that SRC is ANTLOC (i.e. none of the source operands
2758 have changed since the start of the block).
2760 If the source operand changed, we may still use it for the next
2761 iteration of this loop, but we may not use it for substitutions. */
2763 if (gcse_constant_p (src
) || oprs_not_set_p (src
, insn
))
2766 /* If the source of the set is anything except a register, then
2767 we have reached the end of the copy chain. */
2771 /* Follow the copy chain, i.e. start another iteration of the loop
2772 and see if we have an available copy into SRC. */
2773 regno
= REGNO (src
);
2776 /* SET1 holds the last set that was available and anticipatable at
2781 /* Subroutine of cprop_insn that tries to propagate constants into
2782 JUMP_INSNS. JUMP must be a conditional jump. If SETCC is non-NULL
2783 it is the instruction that immediately precedes JUMP, and must be a
2784 single SET of a register. FROM is what we will try to replace,
2785 SRC is the constant we will try to substitute for it. Returns nonzero
2786 if a change was made. */
2789 cprop_jump (basic_block bb
, rtx setcc
, rtx jump
, rtx from
, rtx src
)
2791 rtx
new, set_src
, note_src
;
2792 rtx set
= pc_set (jump
);
2793 rtx note
= find_reg_equal_equiv_note (jump
);
2797 note_src
= XEXP (note
, 0);
2798 if (GET_CODE (note_src
) == EXPR_LIST
)
2799 note_src
= NULL_RTX
;
2801 else note_src
= NULL_RTX
;
2803 /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
2804 set_src
= note_src
? note_src
: SET_SRC (set
);
2806 /* First substitute the SETCC condition into the JUMP instruction,
2807 then substitute that given values into this expanded JUMP. */
2808 if (setcc
!= NULL_RTX
2809 && !modified_between_p (from
, setcc
, jump
)
2810 && !modified_between_p (src
, setcc
, jump
))
2813 rtx setcc_set
= single_set (setcc
);
2814 rtx setcc_note
= find_reg_equal_equiv_note (setcc
);
2815 setcc_src
= (setcc_note
&& GET_CODE (XEXP (setcc_note
, 0)) != EXPR_LIST
)
2816 ? XEXP (setcc_note
, 0) : SET_SRC (setcc_set
);
2817 set_src
= simplify_replace_rtx (set_src
, SET_DEST (setcc_set
),
2823 new = simplify_replace_rtx (set_src
, from
, src
);
2825 /* If no simplification can be made, then try the next register. */
2826 if (rtx_equal_p (new, SET_SRC (set
)))
2829 /* If this is now a no-op delete it, otherwise this must be a valid insn. */
2834 /* Ensure the value computed inside the jump insn to be equivalent
2835 to one computed by setcc. */
2836 if (setcc
&& modified_in_p (new, setcc
))
2838 if (! validate_change (jump
, &SET_SRC (set
), new, 0))
2840 /* When (some) constants are not valid in a comparison, and there
2841 are two registers to be replaced by constants before the entire
2842 comparison can be folded into a constant, we need to keep
2843 intermediate information in REG_EQUAL notes. For targets with
2844 separate compare insns, such notes are added by try_replace_reg.
2845 When we have a combined compare-and-branch instruction, however,
2846 we need to attach a note to the branch itself to make this
2847 optimization work. */
2849 if (!rtx_equal_p (new, note_src
))
2850 set_unique_reg_note (jump
, REG_EQUAL
, copy_rtx (new));
2854 /* Remove REG_EQUAL note after simplification. */
2856 remove_note (jump
, note
);
2860 /* Delete the cc0 setter. */
2861 if (setcc
!= NULL
&& CC0_P (SET_DEST (single_set (setcc
))))
2862 delete_insn (setcc
);
2865 run_jump_opt_after_gcse
= 1;
2867 global_const_prop_count
++;
2868 if (dump_file
!= NULL
)
2871 "GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
2872 REGNO (from
), INSN_UID (jump
));
2873 print_rtl (dump_file
, src
);
2874 fprintf (dump_file
, "\n");
2876 purge_dead_edges (bb
);
2882 constprop_register (rtx insn
, rtx from
, rtx to
, bool alter_jumps
)
2886 /* Check for reg or cc0 setting instructions followed by
2887 conditional branch instructions first. */
2889 && (sset
= single_set (insn
)) != NULL
2891 && any_condjump_p (NEXT_INSN (insn
)) && onlyjump_p (NEXT_INSN (insn
)))
2893 rtx dest
= SET_DEST (sset
);
2894 if ((REG_P (dest
) || CC0_P (dest
))
2895 && cprop_jump (BLOCK_FOR_INSN (insn
), insn
, NEXT_INSN (insn
), from
, to
))
2899 /* Handle normal insns next. */
2900 if (NONJUMP_INSN_P (insn
)
2901 && try_replace_reg (from
, to
, insn
))
2904 /* Try to propagate a CONST_INT into a conditional jump.
2905 We're pretty specific about what we will handle in this
2906 code, we can extend this as necessary over time.
2908 Right now the insn in question must look like
2909 (set (pc) (if_then_else ...)) */
2910 else if (alter_jumps
&& any_condjump_p (insn
) && onlyjump_p (insn
))
2911 return cprop_jump (BLOCK_FOR_INSN (insn
), NULL
, insn
, from
, to
);
2915 /* Perform constant and copy propagation on INSN.
2916 The result is nonzero if a change was made. */
2919 cprop_insn (rtx insn
, int alter_jumps
)
2921 struct reg_use
*reg_used
;
2929 note_uses (&PATTERN (insn
), find_used_regs
, NULL
);
2931 note
= find_reg_equal_equiv_note (insn
);
2933 /* We may win even when propagating constants into notes. */
2935 find_used_regs (&XEXP (note
, 0), NULL
);
2937 for (reg_used
= ®_use_table
[0]; reg_use_count
> 0;
2938 reg_used
++, reg_use_count
--)
2940 unsigned int regno
= REGNO (reg_used
->reg_rtx
);
2944 /* Ignore registers created by GCSE.
2945 We do this because ... */
2946 if (regno
>= max_gcse_regno
)
2949 /* If the register has already been set in this block, there's
2950 nothing we can do. */
2951 if (! oprs_not_set_p (reg_used
->reg_rtx
, insn
))
2954 /* Find an assignment that sets reg_used and is available
2955 at the start of the block. */
2956 set
= find_avail_set (regno
, insn
);
2961 /* ??? We might be able to handle PARALLELs. Later. */
2962 gcc_assert (GET_CODE (pat
) == SET
);
2964 src
= SET_SRC (pat
);
2966 /* Constant propagation. */
2967 if (gcse_constant_p (src
))
2969 if (constprop_register (insn
, reg_used
->reg_rtx
, src
, alter_jumps
))
2972 global_const_prop_count
++;
2973 if (dump_file
!= NULL
)
2975 fprintf (dump_file
, "GLOBAL CONST-PROP: Replacing reg %d in ", regno
);
2976 fprintf (dump_file
, "insn %d with constant ", INSN_UID (insn
));
2977 print_rtl (dump_file
, src
);
2978 fprintf (dump_file
, "\n");
2980 if (INSN_DELETED_P (insn
))
2984 else if (REG_P (src
)
2985 && REGNO (src
) >= FIRST_PSEUDO_REGISTER
2986 && REGNO (src
) != regno
)
2988 if (try_replace_reg (reg_used
->reg_rtx
, src
, insn
))
2991 global_copy_prop_count
++;
2992 if (dump_file
!= NULL
)
2994 fprintf (dump_file
, "GLOBAL COPY-PROP: Replacing reg %d in insn %d",
2995 regno
, INSN_UID (insn
));
2996 fprintf (dump_file
, " with reg %d\n", REGNO (src
));
2999 /* The original insn setting reg_used may or may not now be
3000 deletable. We leave the deletion to flow. */
3001 /* FIXME: If it turns out that the insn isn't deletable,
3002 then we may have unnecessarily extended register lifetimes
3003 and made things worse. */
3011 /* Like find_used_regs, but avoid recording uses that appear in
3012 input-output contexts such as zero_extract or pre_dec. This
3013 restricts the cases we consider to those for which local cprop
3014 can legitimately make replacements. */
3017 local_cprop_find_used_regs (rtx
*xptr
, void *data
)
3024 switch (GET_CODE (x
))
3028 case STRICT_LOW_PART
:
3037 /* Can only legitimately appear this early in the context of
3038 stack pushes for function arguments, but handle all of the
3039 codes nonetheless. */
3043 /* Setting a subreg of a register larger than word_mode leaves
3044 the non-written words unchanged. */
3045 if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x
))) > BITS_PER_WORD
)
3053 find_used_regs (xptr
, data
);
3056 /* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
3057 their REG_EQUAL notes need updating. */
3060 do_local_cprop (rtx x
, rtx insn
, bool alter_jumps
, rtx
*libcall_sp
)
3062 rtx newreg
= NULL
, newcnst
= NULL
;
3064 /* Rule out USE instructions and ASM statements as we don't want to
3065 change the hard registers mentioned. */
3067 && (REGNO (x
) >= FIRST_PSEUDO_REGISTER
3068 || (GET_CODE (PATTERN (insn
)) != USE
3069 && asm_noperands (PATTERN (insn
)) < 0)))
3071 cselib_val
*val
= cselib_lookup (x
, GET_MODE (x
), 0);
3072 struct elt_loc_list
*l
;
3076 for (l
= val
->locs
; l
; l
= l
->next
)
3078 rtx this_rtx
= l
->loc
;
3081 /* Don't CSE non-constant values out of libcall blocks. */
3082 if (l
->in_libcall
&& ! CONSTANT_P (this_rtx
))
3085 if (gcse_constant_p (this_rtx
))
3087 if (REG_P (this_rtx
) && REGNO (this_rtx
) >= FIRST_PSEUDO_REGISTER
3088 /* Don't copy propagate if it has attached REG_EQUIV note.
3089 At this point this only function parameters should have
3090 REG_EQUIV notes and if the argument slot is used somewhere
3091 explicitly, it means address of parameter has been taken,
3092 so we should not extend the lifetime of the pseudo. */
3093 && (!(note
= find_reg_note (l
->setting_insn
, REG_EQUIV
, NULL_RTX
))
3094 || ! MEM_P (XEXP (note
, 0))))
3097 if (newcnst
&& constprop_register (insn
, x
, newcnst
, alter_jumps
))
3099 /* If we find a case where we can't fix the retval REG_EQUAL notes
3100 match the new register, we either have to abandon this replacement
3101 or fix delete_trivially_dead_insns to preserve the setting insn,
3102 or make it delete the REG_EQUAL note, and fix up all passes that
3103 require the REG_EQUAL note there. */
3106 adjusted
= adjust_libcall_notes (x
, newcnst
, insn
, libcall_sp
);
3107 gcc_assert (adjusted
);
3109 if (dump_file
!= NULL
)
3111 fprintf (dump_file
, "LOCAL CONST-PROP: Replacing reg %d in ",
3113 fprintf (dump_file
, "insn %d with constant ",
3115 print_rtl (dump_file
, newcnst
);
3116 fprintf (dump_file
, "\n");
3118 local_const_prop_count
++;
3121 else if (newreg
&& newreg
!= x
&& try_replace_reg (x
, newreg
, insn
))
3123 adjust_libcall_notes (x
, newreg
, insn
, libcall_sp
);
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 /* LIBCALL_SP is a zero-terminated array of insns at the end of a libcall;
3139 their REG_EQUAL notes need updating to reflect that OLDREG has been
3140 replaced with NEWVAL in INSN. Return true if all substitutions could
3143 adjust_libcall_notes (rtx oldreg
, rtx newval
, rtx insn
, rtx
*libcall_sp
)
3147 while ((end
= *libcall_sp
++))
3149 rtx note
= find_reg_equal_equiv_note (end
);
3156 if (reg_set_between_p (newval
, PREV_INSN (insn
), end
))
3160 note
= find_reg_equal_equiv_note (end
);
3163 if (reg_mentioned_p (newval
, XEXP (note
, 0)))
3166 while ((end
= *libcall_sp
++));
3170 XEXP (note
, 0) = simplify_replace_rtx (XEXP (note
, 0), oldreg
, newval
);
3171 df_notes_rescan (end
);
3177 #define MAX_NESTED_LIBCALLS 9
3179 /* Do local const/copy propagation (i.e. within each basic block).
3180 If ALTER_JUMPS is true, allow propagating into jump insns, which
3181 could modify the CFG. */
3184 local_cprop_pass (bool alter_jumps
)
3188 struct reg_use
*reg_used
;
3189 rtx libcall_stack
[MAX_NESTED_LIBCALLS
+ 1], *libcall_sp
;
3190 bool changed
= false;
3192 cselib_init (false);
3193 libcall_sp
= &libcall_stack
[MAX_NESTED_LIBCALLS
];
3197 FOR_BB_INSNS (bb
, insn
)
3201 rtx note
= find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
);
3205 gcc_assert (libcall_sp
!= libcall_stack
);
3206 *--libcall_sp
= XEXP (note
, 0);
3208 note
= find_reg_note (insn
, REG_RETVAL
, NULL_RTX
);
3211 note
= find_reg_equal_equiv_note (insn
);
3215 note_uses (&PATTERN (insn
), local_cprop_find_used_regs
,
3218 local_cprop_find_used_regs (&XEXP (note
, 0), NULL
);
3220 for (reg_used
= ®_use_table
[0]; reg_use_count
> 0;
3221 reg_used
++, reg_use_count
--)
3223 if (do_local_cprop (reg_used
->reg_rtx
, insn
, alter_jumps
,
3230 if (INSN_DELETED_P (insn
))
3233 while (reg_use_count
);
3235 cselib_process_insn (insn
);
3238 /* Forget everything at the end of a basic block. Make sure we are
3239 not inside a libcall, they should never cross basic blocks. */
3240 cselib_clear_table ();
3241 gcc_assert (libcall_sp
== &libcall_stack
[MAX_NESTED_LIBCALLS
]);
3246 /* Global analysis may get into infinite loops for unreachable blocks. */
3247 if (changed
&& alter_jumps
)
3249 delete_unreachable_blocks ();
3250 free_reg_set_mem ();
3251 alloc_reg_set_mem (max_reg_num ());
3256 /* Forward propagate copies. This includes copies and constants. Return
3257 nonzero if a change was made. */
3260 cprop (int alter_jumps
)
3266 /* Note we start at block 1. */
3267 if (ENTRY_BLOCK_PTR
->next_bb
== EXIT_BLOCK_PTR
)
3269 if (dump_file
!= NULL
)
3270 fprintf (dump_file
, "\n");
3275 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
->next_bb
->next_bb
, EXIT_BLOCK_PTR
, next_bb
)
3277 /* Reset tables used to keep track of what's still valid [since the
3278 start of the block]. */
3279 reset_opr_set_tables ();
3281 FOR_BB_INSNS (bb
, insn
)
3284 changed
|= cprop_insn (insn
, alter_jumps
);
3286 /* Keep track of everything modified by this insn. */
3287 /* ??? Need to be careful w.r.t. mods done to INSN. Don't
3288 call mark_oprs_set if we turned the insn into a NOTE. */
3289 if (! NOTE_P (insn
))
3290 mark_oprs_set (insn
);
3294 if (dump_file
!= NULL
)
3295 fprintf (dump_file
, "\n");
3300 /* Similar to get_condition, only the resulting condition must be
3301 valid at JUMP, instead of at EARLIEST.
3303 This differs from noce_get_condition in ifcvt.c in that we prefer not to
3304 settle for the condition variable in the jump instruction being integral.
3305 We prefer to be able to record the value of a user variable, rather than
3306 the value of a temporary used in a condition. This could be solved by
3307 recording the value of *every* register scanned by canonicalize_condition,
3308 but this would require some code reorganization. */
3311 fis_get_condition (rtx jump
)
3313 return get_condition (jump
, NULL
, false, true);
3316 /* Check the comparison COND to see if we can safely form an implicit set from
3317 it. COND is either an EQ or NE comparison. */
3320 implicit_set_cond_p (const_rtx cond
)
3322 const enum machine_mode mode
= GET_MODE (XEXP (cond
, 0));
3323 const_rtx cst
= XEXP (cond
, 1);
3325 /* We can't perform this optimization if either operand might be or might
3326 contain a signed zero. */
3327 if (HONOR_SIGNED_ZEROS (mode
))
3329 /* It is sufficient to check if CST is or contains a zero. We must
3330 handle float, complex, and vector. If any subpart is a zero, then
3331 the optimization can't be performed. */
3332 /* ??? The complex and vector checks are not implemented yet. We just
3333 always return zero for them. */
3334 if (GET_CODE (cst
) == CONST_DOUBLE
)
3337 REAL_VALUE_FROM_CONST_DOUBLE (d
, cst
);
3338 if (REAL_VALUES_EQUAL (d
, dconst0
))
3345 return gcse_constant_p (cst
);
3348 /* Find the implicit sets of a function. An "implicit set" is a constraint
3349 on the value of a variable, implied by a conditional jump. For example,
3350 following "if (x == 2)", the then branch may be optimized as though the
3351 conditional performed an "explicit set", in this example, "x = 2". This
3352 function records the set patterns that are implicit at the start of each
3356 find_implicit_sets (void)
3358 basic_block bb
, dest
;
3364 /* Check for more than one successor. */
3365 if (EDGE_COUNT (bb
->succs
) > 1)
3367 cond
= fis_get_condition (BB_END (bb
));
3370 && (GET_CODE (cond
) == EQ
|| GET_CODE (cond
) == NE
)
3371 && REG_P (XEXP (cond
, 0))
3372 && REGNO (XEXP (cond
, 0)) >= FIRST_PSEUDO_REGISTER
3373 && implicit_set_cond_p (cond
))
3375 dest
= GET_CODE (cond
) == EQ
? BRANCH_EDGE (bb
)->dest
3376 : FALLTHRU_EDGE (bb
)->dest
;
3378 if (dest
&& single_pred_p (dest
)
3379 && dest
!= EXIT_BLOCK_PTR
)
3381 new = gen_rtx_SET (VOIDmode
, XEXP (cond
, 0),
3383 implicit_sets
[dest
->index
] = new;
3386 fprintf(dump_file
, "Implicit set of reg %d in ",
3387 REGNO (XEXP (cond
, 0)));
3388 fprintf(dump_file
, "basic block %d\n", dest
->index
);
3396 fprintf (dump_file
, "Found %d implicit sets\n", count
);
3399 /* Perform one copy/constant propagation pass.
3400 PASS is the pass count. If CPROP_JUMPS is true, perform constant
3401 propagation into conditional jumps. If BYPASS_JUMPS is true,
3402 perform conditional jump bypassing optimizations. */
3405 one_cprop_pass (int pass
, bool cprop_jumps
, bool bypass_jumps
)
3409 global_const_prop_count
= local_const_prop_count
= 0;
3410 global_copy_prop_count
= local_copy_prop_count
= 0;
3413 local_cprop_pass (cprop_jumps
);
3415 /* Determine implicit sets. */
3416 implicit_sets
= XCNEWVEC (rtx
, last_basic_block
);
3417 find_implicit_sets ();
3419 alloc_hash_table (max_cuid
, &set_hash_table
, 1);
3420 compute_hash_table (&set_hash_table
);
3422 /* Free implicit_sets before peak usage. */
3423 free (implicit_sets
);
3424 implicit_sets
= NULL
;
3427 dump_hash_table (dump_file
, "SET", &set_hash_table
);
3428 if (set_hash_table
.n_elems
> 0)
3430 alloc_cprop_mem (last_basic_block
, set_hash_table
.n_elems
);
3431 compute_cprop_data ();
3432 changed
= cprop (cprop_jumps
);
3434 changed
|= bypass_conditional_jumps ();
3438 free_hash_table (&set_hash_table
);
3442 fprintf (dump_file
, "CPROP of %s, pass %d: %d bytes needed, ",
3443 current_function_name (), pass
, bytes_used
);
3444 fprintf (dump_file
, "%d local const props, %d local copy props, ",
3445 local_const_prop_count
, local_copy_prop_count
);
3446 fprintf (dump_file
, "%d global const props, %d global copy props\n\n",
3447 global_const_prop_count
, global_copy_prop_count
);
3449 /* Global analysis may get into infinite loops for unreachable blocks. */
3450 if (changed
&& cprop_jumps
)
3451 delete_unreachable_blocks ();
3456 /* Bypass conditional jumps. */
3458 /* The value of last_basic_block at the beginning of the jump_bypass
3459 pass. The use of redirect_edge_and_branch_force may introduce new
3460 basic blocks, but the data flow analysis is only valid for basic
3461 block indices less than bypass_last_basic_block. */
3463 static int bypass_last_basic_block
;
3465 /* Find a set of REGNO to a constant that is available at the end of basic
3466 block BB. Returns NULL if no such set is found. Based heavily upon
3469 static struct expr
*
3470 find_bypass_set (int regno
, int bb
)
3472 struct expr
*result
= 0;
3477 struct expr
*set
= lookup_set (regno
, &set_hash_table
);
3481 if (TEST_BIT (cprop_avout
[bb
], set
->bitmap_index
))
3483 set
= next_set (regno
, set
);
3489 gcc_assert (GET_CODE (set
->expr
) == SET
);
3491 src
= SET_SRC (set
->expr
);
3492 if (gcse_constant_p (src
))
3498 regno
= REGNO (src
);
3504 /* Subroutine of bypass_block that checks whether a pseudo is killed by
3505 any of the instructions inserted on an edge. Jump bypassing places
3506 condition code setters on CFG edges using insert_insn_on_edge. This
3507 function is required to check that our data flow analysis is still
3508 valid prior to commit_edge_insertions. */
3511 reg_killed_on_edge (const_rtx reg
, const_edge e
)
3515 for (insn
= e
->insns
.r
; insn
; insn
= NEXT_INSN (insn
))
3516 if (INSN_P (insn
) && reg_set_p (reg
, insn
))
3522 /* Subroutine of bypass_conditional_jumps that attempts to bypass the given
3523 basic block BB which has more than one predecessor. If not NULL, SETCC
3524 is the first instruction of BB, which is immediately followed by JUMP_INSN
3525 JUMP. Otherwise, SETCC is NULL, and JUMP is the first insn of BB.
3526 Returns nonzero if a change was made.
3528 During the jump bypassing pass, we may place copies of SETCC instructions
3529 on CFG edges. The following routine must be careful to pay attention to
3530 these inserted insns when performing its transformations. */
3533 bypass_block (basic_block bb
, rtx setcc
, rtx jump
)
3538 int may_be_loop_header
;
3542 insn
= (setcc
!= NULL
) ? setcc
: jump
;
3544 /* Determine set of register uses in INSN. */
3546 note_uses (&PATTERN (insn
), find_used_regs
, NULL
);
3547 note
= find_reg_equal_equiv_note (insn
);
3549 find_used_regs (&XEXP (note
, 0), NULL
);
3551 may_be_loop_header
= false;
3552 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3553 if (e
->flags
& EDGE_DFS_BACK
)
3555 may_be_loop_header
= true;
3560 for (ei
= ei_start (bb
->preds
); (e
= ei_safe_edge (ei
)); )
3564 if (e
->flags
& EDGE_COMPLEX
)
3570 /* We can't redirect edges from new basic blocks. */
3571 if (e
->src
->index
>= bypass_last_basic_block
)
3577 /* The irreducible loops created by redirecting of edges entering the
3578 loop from outside would decrease effectiveness of some of the following
3579 optimizations, so prevent this. */
3580 if (may_be_loop_header
3581 && !(e
->flags
& EDGE_DFS_BACK
))
3587 for (i
= 0; i
< reg_use_count
; i
++)
3589 struct reg_use
*reg_used
= ®_use_table
[i
];
3590 unsigned int regno
= REGNO (reg_used
->reg_rtx
);
3591 basic_block dest
, old_dest
;
3595 if (regno
>= max_gcse_regno
)
3598 set
= find_bypass_set (regno
, e
->src
->index
);
3603 /* Check the data flow is valid after edge insertions. */
3604 if (e
->insns
.r
&& reg_killed_on_edge (reg_used
->reg_rtx
, e
))
3607 src
= SET_SRC (pc_set (jump
));
3610 src
= simplify_replace_rtx (src
,
3611 SET_DEST (PATTERN (setcc
)),
3612 SET_SRC (PATTERN (setcc
)));
3614 new = simplify_replace_rtx (src
, reg_used
->reg_rtx
,
3615 SET_SRC (set
->expr
));
3617 /* Jump bypassing may have already placed instructions on
3618 edges of the CFG. We can't bypass an outgoing edge that
3619 has instructions associated with it, as these insns won't
3620 get executed if the incoming edge is redirected. */
3624 edest
= FALLTHRU_EDGE (bb
);
3625 dest
= edest
->insns
.r
? NULL
: edest
->dest
;
3627 else if (GET_CODE (new) == LABEL_REF
)
3629 dest
= BLOCK_FOR_INSN (XEXP (new, 0));
3630 /* Don't bypass edges containing instructions. */
3631 edest
= find_edge (bb
, dest
);
3632 if (edest
&& edest
->insns
.r
)
3638 /* Avoid unification of the edge with other edges from original
3639 branch. We would end up emitting the instruction on "both"
3642 if (dest
&& setcc
&& !CC0_P (SET_DEST (PATTERN (setcc
)))
3643 && find_edge (e
->src
, dest
))
3649 && dest
!= EXIT_BLOCK_PTR
)
3651 redirect_edge_and_branch_force (e
, dest
);
3653 /* Copy the register setter to the redirected edge.
3654 Don't copy CC0 setters, as CC0 is dead after jump. */
3657 rtx pat
= PATTERN (setcc
);
3658 if (!CC0_P (SET_DEST (pat
)))
3659 insert_insn_on_edge (copy_insn (pat
), e
);
3662 if (dump_file
!= NULL
)
3664 fprintf (dump_file
, "JUMP-BYPASS: Proved reg %d "
3665 "in jump_insn %d equals constant ",
3666 regno
, INSN_UID (jump
));
3667 print_rtl (dump_file
, SET_SRC (set
->expr
));
3668 fprintf (dump_file
, "\nBypass edge from %d->%d to %d\n",
3669 e
->src
->index
, old_dest
->index
, dest
->index
);
3682 /* Find basic blocks with more than one predecessor that only contain a
3683 single conditional jump. If the result of the comparison is known at
3684 compile-time from any incoming edge, redirect that edge to the
3685 appropriate target. Returns nonzero if a change was made.
3687 This function is now mis-named, because we also handle indirect jumps. */
3690 bypass_conditional_jumps (void)
3698 /* Note we start at block 1. */
3699 if (ENTRY_BLOCK_PTR
->next_bb
== EXIT_BLOCK_PTR
)
3702 bypass_last_basic_block
= last_basic_block
;
3703 mark_dfs_back_edges ();
3706 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
->next_bb
->next_bb
,
3707 EXIT_BLOCK_PTR
, next_bb
)
3709 /* Check for more than one predecessor. */
3710 if (!single_pred_p (bb
))
3713 FOR_BB_INSNS (bb
, insn
)
3714 if (NONJUMP_INSN_P (insn
))
3718 if (GET_CODE (PATTERN (insn
)) != SET
)
3721 dest
= SET_DEST (PATTERN (insn
));
3722 if (REG_P (dest
) || CC0_P (dest
))
3727 else if (JUMP_P (insn
))
3729 if ((any_condjump_p (insn
) || computed_jump_p (insn
))
3730 && onlyjump_p (insn
))
3731 changed
|= bypass_block (bb
, setcc
, insn
);
3734 else if (INSN_P (insn
))
3739 /* If we bypassed any register setting insns, we inserted a
3740 copy on the redirected edge. These need to be committed. */
3742 commit_edge_insertions ();
3747 /* Compute PRE+LCM working variables. */
3749 /* Local properties of expressions. */
3750 /* Nonzero for expressions that are transparent in the block. */
3751 static sbitmap
*transp
;
3753 /* Nonzero for expressions that are transparent at the end of the block.
3754 This is only zero for expressions killed by abnormal critical edge
3755 created by a calls. */
3756 static sbitmap
*transpout
;
3758 /* Nonzero for expressions that are computed (available) in the block. */
3759 static sbitmap
*comp
;
3761 /* Nonzero for expressions that are locally anticipatable in the block. */
3762 static sbitmap
*antloc
;
3764 /* Nonzero for expressions where this block is an optimal computation
3766 static sbitmap
*pre_optimal
;
3768 /* Nonzero for expressions which are redundant in a particular block. */
3769 static sbitmap
*pre_redundant
;
3771 /* Nonzero for expressions which should be inserted on a specific edge. */
3772 static sbitmap
*pre_insert_map
;
3774 /* Nonzero for expressions which should be deleted in a specific block. */
3775 static sbitmap
*pre_delete_map
;
3777 /* Contains the edge_list returned by pre_edge_lcm. */
3778 static struct edge_list
*edge_list
;
3780 /* Redundant insns. */
3781 static sbitmap pre_redundant_insns
;
3783 /* Allocate vars used for PRE analysis. */
3786 alloc_pre_mem (int n_blocks
, int n_exprs
)
3788 transp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
3789 comp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
3790 antloc
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
3793 pre_redundant
= NULL
;
3794 pre_insert_map
= NULL
;
3795 pre_delete_map
= NULL
;
3796 ae_kill
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
3798 /* pre_insert and pre_delete are allocated later. */
3801 /* Free vars used for PRE analysis. */
3806 sbitmap_vector_free (transp
);
3807 sbitmap_vector_free (comp
);
3809 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
3812 sbitmap_vector_free (pre_optimal
);
3814 sbitmap_vector_free (pre_redundant
);
3816 sbitmap_vector_free (pre_insert_map
);
3818 sbitmap_vector_free (pre_delete_map
);
3820 transp
= comp
= NULL
;
3821 pre_optimal
= pre_redundant
= pre_insert_map
= pre_delete_map
= NULL
;
3824 /* Top level routine to do the dataflow analysis needed by PRE. */
3827 compute_pre_data (void)
3829 sbitmap trapping_expr
;
3833 compute_local_properties (transp
, comp
, antloc
, &expr_hash_table
);
3834 sbitmap_vector_zero (ae_kill
, last_basic_block
);
3836 /* Collect expressions which might trap. */
3837 trapping_expr
= sbitmap_alloc (expr_hash_table
.n_elems
);
3838 sbitmap_zero (trapping_expr
);
3839 for (ui
= 0; ui
< expr_hash_table
.size
; ui
++)
3842 for (e
= expr_hash_table
.table
[ui
]; e
!= NULL
; e
= e
->next_same_hash
)
3843 if (may_trap_p (e
->expr
))
3844 SET_BIT (trapping_expr
, e
->bitmap_index
);
3847 /* Compute ae_kill for each basic block using:
3857 /* If the current block is the destination of an abnormal edge, we
3858 kill all trapping expressions because we won't be able to properly
3859 place the instruction on the edge. So make them neither
3860 anticipatable nor transparent. This is fairly conservative. */
3861 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3862 if (e
->flags
& EDGE_ABNORMAL
)
3864 sbitmap_difference (antloc
[bb
->index
], antloc
[bb
->index
], trapping_expr
);
3865 sbitmap_difference (transp
[bb
->index
], transp
[bb
->index
], trapping_expr
);
3869 sbitmap_a_or_b (ae_kill
[bb
->index
], transp
[bb
->index
], comp
[bb
->index
]);
3870 sbitmap_not (ae_kill
[bb
->index
], ae_kill
[bb
->index
]);
3873 edge_list
= pre_edge_lcm (expr_hash_table
.n_elems
, transp
, comp
, antloc
,
3874 ae_kill
, &pre_insert_map
, &pre_delete_map
);
3875 sbitmap_vector_free (antloc
);
3877 sbitmap_vector_free (ae_kill
);
3879 sbitmap_free (trapping_expr
);
3884 /* Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach
3887 VISITED is a pointer to a working buffer for tracking which BB's have
3888 been visited. It is NULL for the top-level call.
3890 We treat reaching expressions that go through blocks containing the same
3891 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
3892 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
3893 2 as not reaching. The intent is to improve the probability of finding
3894 only one reaching expression and to reduce register lifetimes by picking
3895 the closest such expression. */
3898 pre_expr_reaches_here_p_work (basic_block occr_bb
, struct expr
*expr
, basic_block bb
, char *visited
)
3903 FOR_EACH_EDGE (pred
, ei
, bb
->preds
)
3905 basic_block pred_bb
= pred
->src
;
3907 if (pred
->src
== ENTRY_BLOCK_PTR
3908 /* Has predecessor has already been visited? */
3909 || visited
[pred_bb
->index
])
3910 ;/* Nothing to do. */
3912 /* Does this predecessor generate this expression? */
3913 else if (TEST_BIT (comp
[pred_bb
->index
], expr
->bitmap_index
))
3915 /* Is this the occurrence we're looking for?
3916 Note that there's only one generating occurrence per block
3917 so we just need to check the block number. */
3918 if (occr_bb
== pred_bb
)
3921 visited
[pred_bb
->index
] = 1;
3923 /* Ignore this predecessor if it kills the expression. */
3924 else if (! TEST_BIT (transp
[pred_bb
->index
], expr
->bitmap_index
))
3925 visited
[pred_bb
->index
] = 1;
3927 /* Neither gen nor kill. */
3930 visited
[pred_bb
->index
] = 1;
3931 if (pre_expr_reaches_here_p_work (occr_bb
, expr
, pred_bb
, visited
))
3936 /* All paths have been checked. */
3940 /* The wrapper for pre_expr_reaches_here_work that ensures that any
3941 memory allocated for that function is returned. */
3944 pre_expr_reaches_here_p (basic_block occr_bb
, struct expr
*expr
, basic_block bb
)
3947 char *visited
= XCNEWVEC (char, last_basic_block
);
3949 rval
= pre_expr_reaches_here_p_work (occr_bb
, expr
, bb
, visited
);
3956 /* Given an expr, generate RTL which we can insert at the end of a BB,
3957 or on an edge. Set the block number of any insns generated to
3961 process_insert_insn (struct expr
*expr
)
3963 rtx reg
= expr
->reaching_reg
;
3964 rtx exp
= copy_rtx (expr
->expr
);
3969 /* If the expression is something that's an operand, like a constant,
3970 just copy it to a register. */
3971 if (general_operand (exp
, GET_MODE (reg
)))
3972 emit_move_insn (reg
, exp
);
3974 /* Otherwise, make a new insn to compute this expression and make sure the
3975 insn will be recognized (this also adds any needed CLOBBERs). Copy the
3976 expression to make sure we don't have any sharing issues. */
3979 rtx insn
= emit_insn (gen_rtx_SET (VOIDmode
, reg
, exp
));
3981 if (insn_invalid_p (insn
))
3992 /* Add EXPR to the end of basic block BB.
3994 This is used by both the PRE and code hoisting.
3996 For PRE, we want to verify that the expr is either transparent
3997 or locally anticipatable in the target block. This check makes
3998 no sense for code hoisting. */
4001 insert_insn_end_basic_block (struct expr
*expr
, basic_block bb
, int pre
)
4003 rtx insn
= BB_END (bb
);
4005 rtx reg
= expr
->reaching_reg
;
4006 int regno
= REGNO (reg
);
4009 pat
= process_insert_insn (expr
);
4010 gcc_assert (pat
&& INSN_P (pat
));
4013 while (NEXT_INSN (pat_end
) != NULL_RTX
)
4014 pat_end
= NEXT_INSN (pat_end
);
4016 /* If the last insn is a jump, insert EXPR in front [taking care to
4017 handle cc0, etc. properly]. Similarly we need to care trapping
4018 instructions in presence of non-call exceptions. */
4021 || (NONJUMP_INSN_P (insn
)
4022 && (!single_succ_p (bb
)
4023 || single_succ_edge (bb
)->flags
& EDGE_ABNORMAL
)))
4028 /* It should always be the case that we can put these instructions
4029 anywhere in the basic block with performing PRE optimizations.
4031 gcc_assert (!NONJUMP_INSN_P (insn
) || !pre
4032 || TEST_BIT (antloc
[bb
->index
], expr
->bitmap_index
)
4033 || TEST_BIT (transp
[bb
->index
], expr
->bitmap_index
));
4035 /* If this is a jump table, then we can't insert stuff here. Since
4036 we know the previous real insn must be the tablejump, we insert
4037 the new instruction just before the tablejump. */
4038 if (GET_CODE (PATTERN (insn
)) == ADDR_VEC
4039 || GET_CODE (PATTERN (insn
)) == ADDR_DIFF_VEC
)
4040 insn
= prev_real_insn (insn
);
4043 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
4044 if cc0 isn't set. */
4045 note
= find_reg_note (insn
, REG_CC_SETTER
, NULL_RTX
);
4047 insn
= XEXP (note
, 0);
4050 rtx maybe_cc0_setter
= prev_nonnote_insn (insn
);
4051 if (maybe_cc0_setter
4052 && INSN_P (maybe_cc0_setter
)
4053 && sets_cc0_p (PATTERN (maybe_cc0_setter
)))
4054 insn
= maybe_cc0_setter
;
4057 /* FIXME: What if something in cc0/jump uses value set in new insn? */
4058 new_insn
= emit_insn_before_noloc (pat
, insn
, bb
);
4061 /* Likewise if the last insn is a call, as will happen in the presence
4062 of exception handling. */
4063 else if (CALL_P (insn
)
4064 && (!single_succ_p (bb
)
4065 || single_succ_edge (bb
)->flags
& EDGE_ABNORMAL
))
4067 /* Keeping in mind SMALL_REGISTER_CLASSES and parameters in registers,
4068 we search backward and place the instructions before the first
4069 parameter is loaded. Do this for everyone for consistency and a
4070 presumption that we'll get better code elsewhere as well.
4072 It should always be the case that we can put these instructions
4073 anywhere in the basic block with performing PRE optimizations.
4077 || TEST_BIT (antloc
[bb
->index
], expr
->bitmap_index
)
4078 || TEST_BIT (transp
[bb
->index
], expr
->bitmap_index
));
4080 /* Since different machines initialize their parameter registers
4081 in different orders, assume nothing. Collect the set of all
4082 parameter registers. */
4083 insn
= find_first_parameter_load (insn
, BB_HEAD (bb
));
4085 /* If we found all the parameter loads, then we want to insert
4086 before the first parameter load.
4088 If we did not find all the parameter loads, then we might have
4089 stopped on the head of the block, which could be a CODE_LABEL.
4090 If we inserted before the CODE_LABEL, then we would be putting
4091 the insn in the wrong basic block. In that case, put the insn
4092 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
4093 while (LABEL_P (insn
)
4094 || NOTE_INSN_BASIC_BLOCK_P (insn
))
4095 insn
= NEXT_INSN (insn
);
4097 new_insn
= emit_insn_before_noloc (pat
, insn
, bb
);
4100 new_insn
= emit_insn_after_noloc (pat
, insn
, bb
);
4106 add_label_notes (PATTERN (pat
), new_insn
);
4107 note_stores (PATTERN (pat
), record_set_info
, pat
);
4111 pat
= NEXT_INSN (pat
);
4114 gcse_create_count
++;
4118 fprintf (dump_file
, "PRE/HOIST: end of bb %d, insn %d, ",
4119 bb
->index
, INSN_UID (new_insn
));
4120 fprintf (dump_file
, "copying expression %d to reg %d\n",
4121 expr
->bitmap_index
, regno
);
4125 /* Insert partially redundant expressions on edges in the CFG to make
4126 the expressions fully redundant. */
4129 pre_edge_insert (struct edge_list
*edge_list
, struct expr
**index_map
)
4131 int e
, i
, j
, num_edges
, set_size
, did_insert
= 0;
4134 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
4135 if it reaches any of the deleted expressions. */
4137 set_size
= pre_insert_map
[0]->size
;
4138 num_edges
= NUM_EDGES (edge_list
);
4139 inserted
= sbitmap_vector_alloc (num_edges
, expr_hash_table
.n_elems
);
4140 sbitmap_vector_zero (inserted
, num_edges
);
4142 for (e
= 0; e
< num_edges
; e
++)
4145 basic_block bb
= INDEX_EDGE_PRED_BB (edge_list
, e
);
4147 for (i
= indx
= 0; i
< set_size
; i
++, indx
+= SBITMAP_ELT_BITS
)
4149 SBITMAP_ELT_TYPE insert
= pre_insert_map
[e
]->elms
[i
];
4151 for (j
= indx
; insert
&& j
< (int) expr_hash_table
.n_elems
; j
++, insert
>>= 1)
4152 if ((insert
& 1) != 0 && index_map
[j
]->reaching_reg
!= NULL_RTX
)
4154 struct expr
*expr
= index_map
[j
];
4157 /* Now look at each deleted occurrence of this expression. */
4158 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4160 if (! occr
->deleted_p
)
4163 /* Insert this expression on this edge if it would
4164 reach the deleted occurrence in BB. */
4165 if (!TEST_BIT (inserted
[e
], j
))
4168 edge eg
= INDEX_EDGE (edge_list
, e
);
4170 /* We can't insert anything on an abnormal and
4171 critical edge, so we insert the insn at the end of
4172 the previous block. There are several alternatives
4173 detailed in Morgans book P277 (sec 10.5) for
4174 handling this situation. This one is easiest for
4177 if (eg
->flags
& EDGE_ABNORMAL
)
4178 insert_insn_end_basic_block (index_map
[j
], bb
, 0);
4181 insn
= process_insert_insn (index_map
[j
]);
4182 insert_insn_on_edge (insn
, eg
);
4187 fprintf (dump_file
, "PRE/HOIST: edge (%d,%d), ",
4189 INDEX_EDGE_SUCC_BB (edge_list
, e
)->index
);
4190 fprintf (dump_file
, "copy expression %d\n",
4191 expr
->bitmap_index
);
4194 update_ld_motion_stores (expr
);
4195 SET_BIT (inserted
[e
], j
);
4197 gcse_create_count
++;
4204 sbitmap_vector_free (inserted
);
4208 /* Copy the result of EXPR->EXPR generated by INSN to EXPR->REACHING_REG.
4209 Given "old_reg <- expr" (INSN), instead of adding after it
4210 reaching_reg <- old_reg
4211 it's better to do the following:
4212 reaching_reg <- expr
4213 old_reg <- reaching_reg
4214 because this way copy propagation can discover additional PRE
4215 opportunities. But if this fails, we try the old way.
4216 When "expr" is a store, i.e.
4217 given "MEM <- old_reg", instead of adding after it
4218 reaching_reg <- old_reg
4219 it's better to add it before as follows:
4220 reaching_reg <- old_reg
4221 MEM <- reaching_reg. */
4224 pre_insert_copy_insn (struct expr
*expr
, rtx insn
)
4226 rtx reg
= expr
->reaching_reg
;
4227 int regno
= REGNO (reg
);
4228 int indx
= expr
->bitmap_index
;
4229 rtx pat
= PATTERN (insn
);
4230 rtx set
, first_set
, new_insn
;
4234 /* This block matches the logic in hash_scan_insn. */
4235 switch (GET_CODE (pat
))
4242 /* Search through the parallel looking for the set whose
4243 source was the expression that we're interested in. */
4244 first_set
= NULL_RTX
;
4246 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
4248 rtx x
= XVECEXP (pat
, 0, i
);
4249 if (GET_CODE (x
) == SET
)
4251 /* If the source was a REG_EQUAL or REG_EQUIV note, we
4252 may not find an equivalent expression, but in this
4253 case the PARALLEL will have a single set. */
4254 if (first_set
== NULL_RTX
)
4256 if (expr_equiv_p (SET_SRC (x
), expr
->expr
))
4264 gcc_assert (first_set
);
4265 if (set
== NULL_RTX
)
4273 if (REG_P (SET_DEST (set
)))
4275 old_reg
= SET_DEST (set
);
4276 /* Check if we can modify the set destination in the original insn. */
4277 if (validate_change (insn
, &SET_DEST (set
), reg
, 0))
4279 new_insn
= gen_move_insn (old_reg
, reg
);
4280 new_insn
= emit_insn_after (new_insn
, insn
);
4282 /* Keep register set table up to date. */
4283 record_one_set (regno
, insn
);
4287 new_insn
= gen_move_insn (reg
, old_reg
);
4288 new_insn
= emit_insn_after (new_insn
, insn
);
4290 /* Keep register set table up to date. */
4291 record_one_set (regno
, new_insn
);
4294 else /* This is possible only in case of a store to memory. */
4296 old_reg
= SET_SRC (set
);
4297 new_insn
= gen_move_insn (reg
, old_reg
);
4299 /* Check if we can modify the set source in the original insn. */
4300 if (validate_change (insn
, &SET_SRC (set
), reg
, 0))
4301 new_insn
= emit_insn_before (new_insn
, insn
);
4303 new_insn
= emit_insn_after (new_insn
, insn
);
4305 /* Keep register set table up to date. */
4306 record_one_set (regno
, new_insn
);
4309 gcse_create_count
++;
4313 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
4314 BLOCK_NUM (insn
), INSN_UID (new_insn
), indx
,
4315 INSN_UID (insn
), regno
);
4318 /* Copy available expressions that reach the redundant expression
4319 to `reaching_reg'. */
4322 pre_insert_copies (void)
4324 unsigned int i
, added_copy
;
4329 /* For each available expression in the table, copy the result to
4330 `reaching_reg' if the expression reaches a deleted one.
4332 ??? The current algorithm is rather brute force.
4333 Need to do some profiling. */
4335 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4336 for (expr
= expr_hash_table
.table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4338 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
4339 we don't want to insert a copy here because the expression may not
4340 really be redundant. So only insert an insn if the expression was
4341 deleted. This test also avoids further processing if the
4342 expression wasn't deleted anywhere. */
4343 if (expr
->reaching_reg
== NULL
)
4346 /* Set when we add a copy for that expression. */
4349 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4351 if (! occr
->deleted_p
)
4354 for (avail
= expr
->avail_occr
; avail
!= NULL
; avail
= avail
->next
)
4356 rtx insn
= avail
->insn
;
4358 /* No need to handle this one if handled already. */
4359 if (avail
->copied_p
)
4362 /* Don't handle this one if it's a redundant one. */
4363 if (TEST_BIT (pre_redundant_insns
, INSN_CUID (insn
)))
4366 /* Or if the expression doesn't reach the deleted one. */
4367 if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail
->insn
),
4369 BLOCK_FOR_INSN (occr
->insn
)))
4374 /* Copy the result of avail to reaching_reg. */
4375 pre_insert_copy_insn (expr
, insn
);
4376 avail
->copied_p
= 1;
4381 update_ld_motion_stores (expr
);
4385 /* Emit move from SRC to DEST noting the equivalence with expression computed
4388 gcse_emit_move_after (rtx src
, rtx dest
, rtx insn
)
4391 rtx set
= single_set (insn
), set2
;
4395 /* This should never fail since we're creating a reg->reg copy
4396 we've verified to be valid. */
4398 new = emit_insn_after (gen_move_insn (dest
, src
), insn
);
4400 /* Note the equivalence for local CSE pass. */
4401 set2
= single_set (new);
4402 if (!set2
|| !rtx_equal_p (SET_DEST (set2
), dest
))
4404 if ((note
= find_reg_equal_equiv_note (insn
)))
4405 eqv
= XEXP (note
, 0);
4407 eqv
= SET_SRC (set
);
4409 set_unique_reg_note (new, REG_EQUAL
, copy_insn_1 (eqv
));
4414 /* Delete redundant computations.
4415 Deletion is done by changing the insn to copy the `reaching_reg' of
4416 the expression into the result of the SET. It is left to later passes
4417 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
4419 Returns nonzero if a change is made. */
4430 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4431 for (expr
= expr_hash_table
.table
[i
];
4433 expr
= expr
->next_same_hash
)
4435 int indx
= expr
->bitmap_index
;
4437 /* We only need to search antic_occr since we require
4440 for (occr
= expr
->antic_occr
; occr
!= NULL
; occr
= occr
->next
)
4442 rtx insn
= occr
->insn
;
4444 basic_block bb
= BLOCK_FOR_INSN (insn
);
4446 /* We only delete insns that have a single_set. */
4447 if (TEST_BIT (pre_delete_map
[bb
->index
], indx
)
4448 && (set
= single_set (insn
)) != 0
4449 && dbg_cnt (pre_insn
))
4451 /* Create a pseudo-reg to store the result of reaching
4452 expressions into. Get the mode for the new pseudo from
4453 the mode of the original destination pseudo. */
4454 if (expr
->reaching_reg
== NULL
)
4456 = gen_reg_rtx (GET_MODE (SET_DEST (set
)));
4458 gcse_emit_move_after (expr
->reaching_reg
, SET_DEST (set
), insn
);
4460 occr
->deleted_p
= 1;
4461 SET_BIT (pre_redundant_insns
, INSN_CUID (insn
));
4468 "PRE: redundant insn %d (expression %d) in ",
4469 INSN_UID (insn
), indx
);
4470 fprintf (dump_file
, "bb %d, reaching reg is %d\n",
4471 bb
->index
, REGNO (expr
->reaching_reg
));
4480 /* Perform GCSE optimizations using PRE.
4481 This is called by one_pre_gcse_pass after all the dataflow analysis
4484 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
4485 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
4486 Compiler Design and Implementation.
4488 ??? A new pseudo reg is created to hold the reaching expression. The nice
4489 thing about the classical approach is that it would try to use an existing
4490 reg. If the register can't be adequately optimized [i.e. we introduce
4491 reload problems], one could add a pass here to propagate the new register
4494 ??? We don't handle single sets in PARALLELs because we're [currently] not
4495 able to copy the rest of the parallel when we insert copies to create full
4496 redundancies from partial redundancies. However, there's no reason why we
4497 can't handle PARALLELs in the cases where there are no partial
4504 int did_insert
, changed
;
4505 struct expr
**index_map
;
4508 /* Compute a mapping from expression number (`bitmap_index') to
4509 hash table entry. */
4511 index_map
= XCNEWVEC (struct expr
*, expr_hash_table
.n_elems
);
4512 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4513 for (expr
= expr_hash_table
.table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4514 index_map
[expr
->bitmap_index
] = expr
;
4516 /* Reset bitmap used to track which insns are redundant. */
4517 pre_redundant_insns
= sbitmap_alloc (max_cuid
);
4518 sbitmap_zero (pre_redundant_insns
);
4520 /* Delete the redundant insns first so that
4521 - we know what register to use for the new insns and for the other
4522 ones with reaching expressions
4523 - we know which insns are redundant when we go to create copies */
4525 changed
= pre_delete ();
4526 did_insert
= pre_edge_insert (edge_list
, index_map
);
4528 /* In other places with reaching expressions, copy the expression to the
4529 specially allocated pseudo-reg that reaches the redundant expr. */
4530 pre_insert_copies ();
4533 commit_edge_insertions ();
4538 sbitmap_free (pre_redundant_insns
);
4542 /* Top level routine to perform one PRE GCSE pass.
4544 Return nonzero if a change was made. */
4547 one_pre_gcse_pass (int pass
)
4551 gcse_subst_count
= 0;
4552 gcse_create_count
= 0;
4554 alloc_hash_table (max_cuid
, &expr_hash_table
, 0);
4555 add_noreturn_fake_exit_edges ();
4557 compute_ld_motion_mems ();
4559 compute_hash_table (&expr_hash_table
);
4560 trim_ld_motion_mems ();
4562 dump_hash_table (dump_file
, "Expression", &expr_hash_table
);
4564 if (expr_hash_table
.n_elems
> 0)
4566 alloc_pre_mem (last_basic_block
, expr_hash_table
.n_elems
);
4567 compute_pre_data ();
4568 changed
|= pre_gcse ();
4569 free_edge_list (edge_list
);
4574 remove_fake_exit_edges ();
4575 free_hash_table (&expr_hash_table
);
4579 fprintf (dump_file
, "\nPRE GCSE of %s, pass %d: %d bytes needed, ",
4580 current_function_name (), pass
, bytes_used
);
4581 fprintf (dump_file
, "%d substs, %d insns created\n",
4582 gcse_subst_count
, gcse_create_count
);
4588 /* If X contains any LABEL_REF's, add REG_LABEL_OPERAND notes for them
4589 to INSN. If such notes are added to an insn which references a
4590 CODE_LABEL, the LABEL_NUSES count is incremented. We have to add
4591 that note, because the following loop optimization pass requires
4594 /* ??? If there was a jump optimization pass after gcse and before loop,
4595 then we would not need to do this here, because jump would add the
4596 necessary REG_LABEL_OPERAND and REG_LABEL_TARGET notes. */
4599 add_label_notes (rtx x
, rtx insn
)
4601 enum rtx_code code
= GET_CODE (x
);
4605 if (code
== LABEL_REF
&& !LABEL_REF_NONLOCAL_P (x
))
4607 /* This code used to ignore labels that referred to dispatch tables to
4608 avoid flow generating (slightly) worse code.
4610 We no longer ignore such label references (see LABEL_REF handling in
4611 mark_jump_label for additional information). */
4613 if (reg_mentioned_p (XEXP (x
, 0), insn
))
4615 /* There's no reason for current users to emit jump-insns
4616 with such a LABEL_REF, so we don't have to handle
4617 REG_LABEL_TARGET notes. */
4618 gcc_assert (!JUMP_P (insn
));
4620 = gen_rtx_INSN_LIST (REG_LABEL_OPERAND
, XEXP (x
, 0),
4622 if (LABEL_P (XEXP (x
, 0)))
4623 LABEL_NUSES (XEXP (x
, 0))++;
4628 for (i
= GET_RTX_LENGTH (code
) - 1, fmt
= GET_RTX_FORMAT (code
); i
>= 0; i
--)
4631 add_label_notes (XEXP (x
, i
), insn
);
4632 else if (fmt
[i
] == 'E')
4633 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
4634 add_label_notes (XVECEXP (x
, i
, j
), insn
);
4638 /* Compute transparent outgoing information for each block.
4640 An expression is transparent to an edge unless it is killed by
4641 the edge itself. This can only happen with abnormal control flow,
4642 when the edge is traversed through a call. This happens with
4643 non-local labels and exceptions.
4645 This would not be necessary if we split the edge. While this is
4646 normally impossible for abnormal critical edges, with some effort
4647 it should be possible with exception handling, since we still have
4648 control over which handler should be invoked. But due to increased
4649 EH table sizes, this may not be worthwhile. */
4652 compute_transpout (void)
4658 sbitmap_vector_ones (transpout
, last_basic_block
);
4662 /* Note that flow inserted a nop a the end of basic blocks that
4663 end in call instructions for reasons other than abnormal
4665 if (! CALL_P (BB_END (bb
)))
4668 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4669 for (expr
= expr_hash_table
.table
[i
]; expr
; expr
= expr
->next_same_hash
)
4670 if (MEM_P (expr
->expr
))
4672 if (GET_CODE (XEXP (expr
->expr
, 0)) == SYMBOL_REF
4673 && CONSTANT_POOL_ADDRESS_P (XEXP (expr
->expr
, 0)))
4676 /* ??? Optimally, we would use interprocedural alias
4677 analysis to determine if this mem is actually killed
4679 RESET_BIT (transpout
[bb
->index
], expr
->bitmap_index
);
4684 /* Code Hoisting variables and subroutines. */
4686 /* Very busy expressions. */
4687 static sbitmap
*hoist_vbein
;
4688 static sbitmap
*hoist_vbeout
;
4690 /* Hoistable expressions. */
4691 static sbitmap
*hoist_exprs
;
4693 /* ??? We could compute post dominators and run this algorithm in
4694 reverse to perform tail merging, doing so would probably be
4695 more effective than the tail merging code in jump.c.
4697 It's unclear if tail merging could be run in parallel with
4698 code hoisting. It would be nice. */
4700 /* Allocate vars used for code hoisting analysis. */
4703 alloc_code_hoist_mem (int n_blocks
, int n_exprs
)
4705 antloc
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4706 transp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4707 comp
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4709 hoist_vbein
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4710 hoist_vbeout
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4711 hoist_exprs
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4712 transpout
= sbitmap_vector_alloc (n_blocks
, n_exprs
);
4715 /* Free vars used for code hoisting analysis. */
4718 free_code_hoist_mem (void)
4720 sbitmap_vector_free (antloc
);
4721 sbitmap_vector_free (transp
);
4722 sbitmap_vector_free (comp
);
4724 sbitmap_vector_free (hoist_vbein
);
4725 sbitmap_vector_free (hoist_vbeout
);
4726 sbitmap_vector_free (hoist_exprs
);
4727 sbitmap_vector_free (transpout
);
4729 free_dominance_info (CDI_DOMINATORS
);
4732 /* Compute the very busy expressions at entry/exit from each block.
4734 An expression is very busy if all paths from a given point
4735 compute the expression. */
4738 compute_code_hoist_vbeinout (void)
4740 int changed
, passes
;
4743 sbitmap_vector_zero (hoist_vbeout
, last_basic_block
);
4744 sbitmap_vector_zero (hoist_vbein
, last_basic_block
);
4753 /* We scan the blocks in the reverse order to speed up
4755 FOR_EACH_BB_REVERSE (bb
)
4757 changed
|= sbitmap_a_or_b_and_c_cg (hoist_vbein
[bb
->index
], antloc
[bb
->index
],
4758 hoist_vbeout
[bb
->index
], transp
[bb
->index
]);
4759 if (bb
->next_bb
!= EXIT_BLOCK_PTR
)
4760 sbitmap_intersection_of_succs (hoist_vbeout
[bb
->index
], hoist_vbein
, bb
->index
);
4767 fprintf (dump_file
, "hoisting vbeinout computation: %d passes\n", passes
);
4770 /* Top level routine to do the dataflow analysis needed by code hoisting. */
4773 compute_code_hoist_data (void)
4775 compute_local_properties (transp
, comp
, antloc
, &expr_hash_table
);
4776 compute_transpout ();
4777 compute_code_hoist_vbeinout ();
4778 calculate_dominance_info (CDI_DOMINATORS
);
4780 fprintf (dump_file
, "\n");
4783 /* Determine if the expression identified by EXPR_INDEX would
4784 reach BB unimpared if it was placed at the end of EXPR_BB.
4786 It's unclear exactly what Muchnick meant by "unimpared". It seems
4787 to me that the expression must either be computed or transparent in
4788 *every* block in the path(s) from EXPR_BB to BB. Any other definition
4789 would allow the expression to be hoisted out of loops, even if
4790 the expression wasn't a loop invariant.
4792 Contrast this to reachability for PRE where an expression is
4793 considered reachable if *any* path reaches instead of *all*
4797 hoist_expr_reaches_here_p (basic_block expr_bb
, int expr_index
, basic_block bb
, char *visited
)
4801 int visited_allocated_locally
= 0;
4804 if (visited
== NULL
)
4806 visited_allocated_locally
= 1;
4807 visited
= XCNEWVEC (char, last_basic_block
);
4810 FOR_EACH_EDGE (pred
, ei
, bb
->preds
)
4812 basic_block pred_bb
= pred
->src
;
4814 if (pred
->src
== ENTRY_BLOCK_PTR
)
4816 else if (pred_bb
== expr_bb
)
4818 else if (visited
[pred_bb
->index
])
4821 /* Does this predecessor generate this expression? */
4822 else if (TEST_BIT (comp
[pred_bb
->index
], expr_index
))
4824 else if (! TEST_BIT (transp
[pred_bb
->index
], expr_index
))
4830 visited
[pred_bb
->index
] = 1;
4831 if (! hoist_expr_reaches_here_p (expr_bb
, expr_index
,
4836 if (visited_allocated_locally
)
4839 return (pred
== NULL
);
4842 /* Actually perform code hoisting. */
4847 basic_block bb
, dominated
;
4848 VEC (basic_block
, heap
) *domby
;
4850 struct expr
**index_map
;
4853 sbitmap_vector_zero (hoist_exprs
, last_basic_block
);
4855 /* Compute a mapping from expression number (`bitmap_index') to
4856 hash table entry. */
4858 index_map
= XCNEWVEC (struct expr
*, expr_hash_table
.n_elems
);
4859 for (i
= 0; i
< expr_hash_table
.size
; i
++)
4860 for (expr
= expr_hash_table
.table
[i
]; expr
!= NULL
; expr
= expr
->next_same_hash
)
4861 index_map
[expr
->bitmap_index
] = expr
;
4863 /* Walk over each basic block looking for potentially hoistable
4864 expressions, nothing gets hoisted from the entry block. */
4868 int insn_inserted_p
;
4870 domby
= get_dominated_by (CDI_DOMINATORS
, bb
);
4871 /* Examine each expression that is very busy at the exit of this
4872 block. These are the potentially hoistable expressions. */
4873 for (i
= 0; i
< hoist_vbeout
[bb
->index
]->n_bits
; i
++)
4877 if (TEST_BIT (hoist_vbeout
[bb
->index
], i
)
4878 && TEST_BIT (transpout
[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
)
4888 /* We've found a dominated block, now see if it computes
4889 the busy expression and whether or not moving that
4890 expression to the "beginning" of that block is safe. */
4891 if (!TEST_BIT (antloc
[dominated
->index
], i
))
4894 /* Note if the expression would reach the dominated block
4895 unimpared if it was placed at the end of BB.
4897 Keep track of how many times this expression is hoistable
4898 from a dominated block into BB. */
4899 if (hoist_expr_reaches_here_p (bb
, i
, dominated
, NULL
))
4903 /* If we found more than one hoistable occurrence of this
4904 expression, then note it in the bitmap of expressions to
4905 hoist. It makes no sense to hoist things which are computed
4906 in only one BB, and doing so tends to pessimize register
4907 allocation. One could increase this value to try harder
4908 to avoid any possible code expansion due to register
4909 allocation issues; however experiments have shown that
4910 the vast majority of hoistable expressions are only movable
4911 from two successors, so raising this threshold is likely
4912 to nullify any benefit we get from code hoisting. */
4915 SET_BIT (hoist_exprs
[bb
->index
], i
);
4920 /* If we found nothing to hoist, then quit now. */
4923 VEC_free (basic_block
, heap
, domby
);
4927 /* Loop over all the hoistable expressions. */
4928 for (i
= 0; i
< hoist_exprs
[bb
->index
]->n_bits
; i
++)
4930 /* We want to insert the expression into BB only once, so
4931 note when we've inserted it. */
4932 insn_inserted_p
= 0;
4934 /* These tests should be the same as the tests above. */
4935 if (TEST_BIT (hoist_exprs
[bb
->index
], i
))
4937 /* We've found a potentially hoistable expression, now
4938 we look at every block BB dominates to see if it
4939 computes the expression. */
4940 for (j
= 0; VEC_iterate (basic_block
, domby
, j
, dominated
); j
++)
4942 /* Ignore self dominance. */
4943 if (bb
== dominated
)
4946 /* We've found a dominated block, now see if it computes
4947 the busy expression and whether or not moving that
4948 expression to the "beginning" of that block is safe. */
4949 if (!TEST_BIT (antloc
[dominated
->index
], i
))
4952 /* The expression is computed in the dominated block and
4953 it would be safe to compute it at the start of the
4954 dominated block. Now we have to determine if the
4955 expression would reach the dominated block if it was
4956 placed at the end of BB. */
4957 if (hoist_expr_reaches_here_p (bb
, i
, dominated
, NULL
))
4959 struct expr
*expr
= index_map
[i
];
4960 struct occr
*occr
= expr
->antic_occr
;
4964 /* Find the right occurrence of this expression. */
4965 while (BLOCK_FOR_INSN (occr
->insn
) != dominated
&& occr
)
4970 set
= single_set (insn
);
4973 /* Create a pseudo-reg to store the result of reaching
4974 expressions into. Get the mode for the new pseudo
4975 from the mode of the original destination pseudo. */
4976 if (expr
->reaching_reg
== NULL
)
4978 = gen_reg_rtx (GET_MODE (SET_DEST (set
)));
4980 gcse_emit_move_after (expr
->reaching_reg
, SET_DEST (set
), insn
);
4982 occr
->deleted_p
= 1;
4983 if (!insn_inserted_p
)
4985 insert_insn_end_basic_block (index_map
[i
], bb
, 0);
4986 insn_inserted_p
= 1;
4992 VEC_free (basic_block
, heap
, domby
);
4998 /* Top level routine to perform one code hoisting (aka unification) pass
5000 Return nonzero if a change was made. */
5003 one_code_hoisting_pass (void)
5007 alloc_hash_table (max_cuid
, &expr_hash_table
, 0);
5008 compute_hash_table (&expr_hash_table
);
5010 dump_hash_table (dump_file
, "Code Hosting Expressions", &expr_hash_table
);
5012 if (expr_hash_table
.n_elems
> 0)
5014 alloc_code_hoist_mem (last_basic_block
, expr_hash_table
.n_elems
);
5015 compute_code_hoist_data ();
5017 free_code_hoist_mem ();
5020 free_hash_table (&expr_hash_table
);
5025 /* Here we provide the things required to do store motion towards
5026 the exit. In order for this to be effective, gcse also needed to
5027 be taught how to move a load when it is kill only by a store to itself.
5032 void foo(float scale)
5034 for (i=0; i<10; i++)
5038 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
5039 the load out since its live around the loop, and stored at the bottom
5042 The 'Load Motion' referred to and implemented in this file is
5043 an enhancement to gcse which when using edge based lcm, recognizes
5044 this situation and allows gcse to move the load out of the loop.
5046 Once gcse has hoisted the load, store motion can then push this
5047 load towards the exit, and we end up with no loads or stores of 'i'
5051 pre_ldst_expr_hash (const void *p
)
5053 int do_not_record_p
= 0;
5054 const struct ls_expr
*x
= p
;
5055 return hash_rtx (x
->pattern
, GET_MODE (x
->pattern
), &do_not_record_p
, NULL
, false);
5059 pre_ldst_expr_eq (const void *p1
, const void *p2
)
5061 const struct ls_expr
*ptr1
= p1
, *ptr2
= p2
;
5062 return expr_equiv_p (ptr1
->pattern
, ptr2
->pattern
);
5065 /* This will search the ldst list for a matching expression. If it
5066 doesn't find one, we create one and initialize it. */
5068 static struct ls_expr
*
5071 int do_not_record_p
= 0;
5072 struct ls_expr
* ptr
;
5077 hash
= hash_rtx (x
, GET_MODE (x
), &do_not_record_p
,
5078 NULL
, /*have_reg_qty=*/false);
5081 slot
= htab_find_slot_with_hash (pre_ldst_table
, &e
, hash
, INSERT
);
5083 return (struct ls_expr
*)*slot
;
5085 ptr
= XNEW (struct ls_expr
);
5087 ptr
->next
= pre_ldst_mems
;
5090 ptr
->pattern_regs
= NULL_RTX
;
5091 ptr
->loads
= NULL_RTX
;
5092 ptr
->stores
= NULL_RTX
;
5093 ptr
->reaching_reg
= NULL_RTX
;
5096 ptr
->hash_index
= hash
;
5097 pre_ldst_mems
= ptr
;
5103 /* Free up an individual ldst entry. */
5106 free_ldst_entry (struct ls_expr
* ptr
)
5108 free_INSN_LIST_list (& ptr
->loads
);
5109 free_INSN_LIST_list (& ptr
->stores
);
5114 /* Free up all memory associated with the ldst list. */
5117 free_ldst_mems (void)
5120 htab_delete (pre_ldst_table
);
5121 pre_ldst_table
= NULL
;
5123 while (pre_ldst_mems
)
5125 struct ls_expr
* tmp
= pre_ldst_mems
;
5127 pre_ldst_mems
= pre_ldst_mems
->next
;
5129 free_ldst_entry (tmp
);
5132 pre_ldst_mems
= NULL
;
5135 /* Dump debugging info about the ldst list. */
5138 print_ldst_list (FILE * file
)
5140 struct ls_expr
* ptr
;
5142 fprintf (file
, "LDST list: \n");
5144 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
5146 fprintf (file
, " Pattern (%3d): ", ptr
->index
);
5148 print_rtl (file
, ptr
->pattern
);
5150 fprintf (file
, "\n Loads : ");
5153 print_rtl (file
, ptr
->loads
);
5155 fprintf (file
, "(nil)");
5157 fprintf (file
, "\n Stores : ");
5160 print_rtl (file
, ptr
->stores
);
5162 fprintf (file
, "(nil)");
5164 fprintf (file
, "\n\n");
5167 fprintf (file
, "\n");
5170 /* Returns 1 if X is in the list of ldst only expressions. */
5172 static struct ls_expr
*
5173 find_rtx_in_ldst (rtx x
)
5177 if (!pre_ldst_table
)
5180 slot
= htab_find_slot (pre_ldst_table
, &e
, NO_INSERT
);
5181 if (!slot
|| ((struct ls_expr
*)*slot
)->invalid
)
5186 /* Assign each element of the list of mems a monotonically increasing value. */
5189 enumerate_ldsts (void)
5191 struct ls_expr
* ptr
;
5194 for (ptr
= pre_ldst_mems
; ptr
!= NULL
; ptr
= ptr
->next
)
5200 /* Return first item in the list. */
5202 static inline struct ls_expr
*
5203 first_ls_expr (void)
5205 return pre_ldst_mems
;
5208 /* Return the next item in the list after the specified one. */
5210 static inline struct ls_expr
*
5211 next_ls_expr (struct ls_expr
* ptr
)
5216 /* Load Motion for loads which only kill themselves. */
5218 /* Return true if x is a simple MEM operation, with no registers or
5219 side effects. These are the types of loads we consider for the
5220 ld_motion list, otherwise we let the usual aliasing take care of it. */
5223 simple_mem (const_rtx x
)
5228 if (MEM_VOLATILE_P (x
))
5231 if (GET_MODE (x
) == BLKmode
)
5234 /* If we are handling exceptions, we must be careful with memory references
5235 that may trap. If we are not, the behavior is undefined, so we may just
5237 if (flag_non_call_exceptions
&& may_trap_p (x
))
5240 if (side_effects_p (x
))
5243 /* Do not consider function arguments passed on stack. */
5244 if (reg_mentioned_p (stack_pointer_rtx
, x
))
5247 if (flag_float_store
&& FLOAT_MODE_P (GET_MODE (x
)))
5253 /* Make sure there isn't a buried reference in this pattern anywhere.
5254 If there is, invalidate the entry for it since we're not capable
5255 of fixing it up just yet.. We have to be sure we know about ALL
5256 loads since the aliasing code will allow all entries in the
5257 ld_motion list to not-alias itself. If we miss a load, we will get
5258 the wrong value since gcse might common it and we won't know to
5262 invalidate_any_buried_refs (rtx x
)
5266 struct ls_expr
* ptr
;
5268 /* Invalidate it in the list. */
5269 if (MEM_P (x
) && simple_mem (x
))
5271 ptr
= ldst_entry (x
);
5275 /* Recursively process the insn. */
5276 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5278 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0; i
--)
5281 invalidate_any_buried_refs (XEXP (x
, i
));
5282 else if (fmt
[i
] == 'E')
5283 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
5284 invalidate_any_buried_refs (XVECEXP (x
, i
, j
));
5288 /* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
5289 being defined as MEM loads and stores to symbols, with no side effects
5290 and no registers in the expression. For a MEM destination, we also
5291 check that the insn is still valid if we replace the destination with a
5292 REG, as is done in update_ld_motion_stores. If there are any uses/defs
5293 which don't match this criteria, they are invalidated and trimmed out
5297 compute_ld_motion_mems (void)
5299 struct ls_expr
* ptr
;
5303 pre_ldst_mems
= NULL
;
5304 pre_ldst_table
= htab_create (13, pre_ldst_expr_hash
,
5305 pre_ldst_expr_eq
, NULL
);
5309 FOR_BB_INSNS (bb
, insn
)
5313 if (GET_CODE (PATTERN (insn
)) == SET
)
5315 rtx src
= SET_SRC (PATTERN (insn
));
5316 rtx dest
= SET_DEST (PATTERN (insn
));
5318 /* Check for a simple LOAD... */
5319 if (MEM_P (src
) && simple_mem (src
))
5321 ptr
= ldst_entry (src
);
5323 ptr
->loads
= alloc_INSN_LIST (insn
, ptr
->loads
);
5329 /* Make sure there isn't a buried load somewhere. */
5330 invalidate_any_buried_refs (src
);
5333 /* Check for stores. Don't worry about aliased ones, they
5334 will block any movement we might do later. We only care
5335 about this exact pattern since those are the only
5336 circumstance that we will ignore the aliasing info. */
5337 if (MEM_P (dest
) && simple_mem (dest
))
5339 ptr
= ldst_entry (dest
);
5342 && GET_CODE (src
) != ASM_OPERANDS
5343 /* Check for REG manually since want_to_gcse_p
5344 returns 0 for all REGs. */
5345 && can_assign_to_reg_p (src
))
5346 ptr
->stores
= alloc_INSN_LIST (insn
, ptr
->stores
);
5352 invalidate_any_buried_refs (PATTERN (insn
));
5358 /* Remove any references that have been either invalidated or are not in the
5359 expression list for pre gcse. */
5362 trim_ld_motion_mems (void)
5364 struct ls_expr
* * last
= & pre_ldst_mems
;
5365 struct ls_expr
* ptr
= pre_ldst_mems
;
5371 /* Delete if entry has been made invalid. */
5374 /* Delete if we cannot find this mem in the expression list. */
5375 unsigned int hash
= ptr
->hash_index
% expr_hash_table
.size
;
5377 for (expr
= expr_hash_table
.table
[hash
];
5379 expr
= expr
->next_same_hash
)
5380 if (expr_equiv_p (expr
->expr
, ptr
->pattern
))
5384 expr
= (struct expr
*) 0;
5388 /* Set the expression field if we are keeping it. */
5396 htab_remove_elt_with_hash (pre_ldst_table
, ptr
, ptr
->hash_index
);
5397 free_ldst_entry (ptr
);
5402 /* Show the world what we've found. */
5403 if (dump_file
&& pre_ldst_mems
!= NULL
)
5404 print_ldst_list (dump_file
);
5407 /* This routine will take an expression which we are replacing with
5408 a reaching register, and update any stores that are needed if
5409 that expression is in the ld_motion list. Stores are updated by
5410 copying their SRC to the reaching register, and then storing
5411 the reaching register into the store location. These keeps the
5412 correct value in the reaching register for the loads. */
5415 update_ld_motion_stores (struct expr
* expr
)
5417 struct ls_expr
* mem_ptr
;
5419 if ((mem_ptr
= find_rtx_in_ldst (expr
->expr
)))
5421 /* We can try to find just the REACHED stores, but is shouldn't
5422 matter to set the reaching reg everywhere... some might be
5423 dead and should be eliminated later. */
5425 /* We replace (set mem expr) with (set reg expr) (set mem reg)
5426 where reg is the reaching reg used in the load. We checked in
5427 compute_ld_motion_mems that we can replace (set mem expr) with
5428 (set reg expr) in that insn. */
5429 rtx list
= mem_ptr
->stores
;
5431 for ( ; list
!= NULL_RTX
; list
= XEXP (list
, 1))
5433 rtx insn
= XEXP (list
, 0);
5434 rtx pat
= PATTERN (insn
);
5435 rtx src
= SET_SRC (pat
);
5436 rtx reg
= expr
->reaching_reg
;
5439 /* If we've already copied it, continue. */
5440 if (expr
->reaching_reg
== src
)
5445 fprintf (dump_file
, "PRE: store updated with reaching reg ");
5446 print_rtl (dump_file
, expr
->reaching_reg
);
5447 fprintf (dump_file
, ":\n ");
5448 print_inline_rtx (dump_file
, insn
, 8);
5449 fprintf (dump_file
, "\n");
5452 copy
= gen_move_insn ( reg
, copy_rtx (SET_SRC (pat
)));
5453 new = emit_insn_before (copy
, insn
);
5454 record_one_set (REGNO (reg
), new);
5455 SET_SRC (pat
) = reg
;
5456 df_insn_rescan (insn
);
5458 /* un-recognize this pattern since it's probably different now. */
5459 INSN_CODE (insn
) = -1;
5460 gcse_create_count
++;
5465 /* Store motion code. */
5467 #define ANTIC_STORE_LIST(x) ((x)->loads)
5468 #define AVAIL_STORE_LIST(x) ((x)->stores)
5469 #define LAST_AVAIL_CHECK_FAILURE(x) ((x)->reaching_reg)
5471 /* This is used to communicate the target bitvector we want to use in the
5472 reg_set_info routine when called via the note_stores mechanism. */
5473 static int * regvec
;
5475 /* And current insn, for the same routine. */
5476 static rtx compute_store_table_current_insn
;
5478 /* Used in computing the reverse edge graph bit vectors. */
5479 static sbitmap
* st_antloc
;
5481 /* Global holding the number of store expressions we are dealing with. */
5482 static int num_stores
;
5484 /* Checks to set if we need to mark a register set. Called from
5488 reg_set_info (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
,
5491 sbitmap bb_reg
= data
;
5493 if (GET_CODE (dest
) == SUBREG
)
5494 dest
= SUBREG_REG (dest
);
5498 regvec
[REGNO (dest
)] = INSN_UID (compute_store_table_current_insn
);
5500 SET_BIT (bb_reg
, REGNO (dest
));
5504 /* Clear any mark that says that this insn sets dest. Called from
5508 reg_clear_last_set (rtx dest
, const_rtx setter ATTRIBUTE_UNUSED
,
5511 int *dead_vec
= data
;
5513 if (GET_CODE (dest
) == SUBREG
)
5514 dest
= SUBREG_REG (dest
);
5517 dead_vec
[REGNO (dest
)] == INSN_UID (compute_store_table_current_insn
))
5518 dead_vec
[REGNO (dest
)] = 0;
5521 /* Return zero if some of the registers in list X are killed
5522 due to set of registers in bitmap REGS_SET. */
5525 store_ops_ok (const_rtx x
, int *regs_set
)
5529 for (; x
; x
= XEXP (x
, 1))
5532 if (regs_set
[REGNO(reg
)])
5539 /* Returns a list of registers mentioned in X. */
5541 extract_mentioned_regs (rtx x
)
5543 return extract_mentioned_regs_helper (x
, NULL_RTX
);
5546 /* Helper for extract_mentioned_regs; ACCUM is used to accumulate used
5549 extract_mentioned_regs_helper (rtx x
, rtx accum
)
5555 /* Repeat is used to turn tail-recursion into iteration. */
5561 code
= GET_CODE (x
);
5565 return alloc_EXPR_LIST (0, x
, accum
);
5577 /* We do not run this function with arguments having side effects. */
5597 i
= GET_RTX_LENGTH (code
) - 1;
5598 fmt
= GET_RTX_FORMAT (code
);
5604 rtx tem
= XEXP (x
, i
);
5606 /* If we are about to do the last recursive call
5607 needed at this level, change it into iteration. */
5614 accum
= extract_mentioned_regs_helper (tem
, accum
);
5616 else if (fmt
[i
] == 'E')
5620 for (j
= 0; j
< XVECLEN (x
, i
); j
++)
5621 accum
= extract_mentioned_regs_helper (XVECEXP (x
, i
, j
), accum
);
5628 /* Determine whether INSN is MEM store pattern that we will consider moving.
5629 REGS_SET_BEFORE is bitmap of registers set before (and including) the
5630 current insn, REGS_SET_AFTER is bitmap of registers set after (and
5631 including) the insn in this basic block. We must be passing through BB from
5632 head to end, as we are using this fact to speed things up.
5634 The results are stored this way:
5636 -- the first anticipatable expression is added into ANTIC_STORE_LIST
5637 -- if the processed expression is not anticipatable, NULL_RTX is added
5638 there instead, so that we can use it as indicator that no further
5639 expression of this type may be anticipatable
5640 -- if the expression is available, it is added as head of AVAIL_STORE_LIST;
5641 consequently, all of them but this head are dead and may be deleted.
5642 -- if the expression is not available, the insn due to that it fails to be
5643 available is stored in reaching_reg.
5645 The things are complicated a bit by fact that there already may be stores
5646 to the same MEM from other blocks; also caller must take care of the
5647 necessary cleanup of the temporary markers after end of the basic block.
5651 find_moveable_store (rtx insn
, int *regs_set_before
, int *regs_set_after
)
5653 struct ls_expr
* ptr
;
5655 int check_anticipatable
, check_available
;
5656 basic_block bb
= BLOCK_FOR_INSN (insn
);
5658 set
= single_set (insn
);
5662 dest
= SET_DEST (set
);
5664 if (! MEM_P (dest
) || MEM_VOLATILE_P (dest
)
5665 || GET_MODE (dest
) == BLKmode
)
5668 if (side_effects_p (dest
))
5671 /* If we are handling exceptions, we must be careful with memory references
5672 that may trap. If we are not, the behavior is undefined, so we may just
5674 if (flag_non_call_exceptions
&& may_trap_p (dest
))
5677 /* Even if the destination cannot trap, the source may. In this case we'd
5678 need to handle updating the REG_EH_REGION note. */
5679 if (find_reg_note (insn
, REG_EH_REGION
, NULL_RTX
))
5682 /* Make sure that the SET_SRC of this store insns can be assigned to
5683 a register, or we will fail later on in replace_store_insn, which
5684 assumes that we can do this. But sometimes the target machine has
5685 oddities like MEM read-modify-write instruction. See for example
5687 if (!can_assign_to_reg_p (SET_SRC (set
)))
5690 ptr
= ldst_entry (dest
);
5691 if (!ptr
->pattern_regs
)
5692 ptr
->pattern_regs
= extract_mentioned_regs (dest
);
5694 /* Do not check for anticipatability if we either found one anticipatable
5695 store already, or tested for one and found out that it was killed. */
5696 check_anticipatable
= 0;
5697 if (!ANTIC_STORE_LIST (ptr
))
5698 check_anticipatable
= 1;
5701 tmp
= XEXP (ANTIC_STORE_LIST (ptr
), 0);
5703 && BLOCK_FOR_INSN (tmp
) != bb
)
5704 check_anticipatable
= 1;
5706 if (check_anticipatable
)
5708 if (store_killed_before (dest
, ptr
->pattern_regs
, insn
, bb
, regs_set_before
))
5712 ANTIC_STORE_LIST (ptr
) = alloc_INSN_LIST (tmp
,
5713 ANTIC_STORE_LIST (ptr
));
5716 /* It is not necessary to check whether store is available if we did
5717 it successfully before; if we failed before, do not bother to check
5718 until we reach the insn that caused us to fail. */
5719 check_available
= 0;
5720 if (!AVAIL_STORE_LIST (ptr
))
5721 check_available
= 1;
5724 tmp
= XEXP (AVAIL_STORE_LIST (ptr
), 0);
5725 if (BLOCK_FOR_INSN (tmp
) != bb
)
5726 check_available
= 1;
5728 if (check_available
)
5730 /* Check that we have already reached the insn at that the check
5731 failed last time. */
5732 if (LAST_AVAIL_CHECK_FAILURE (ptr
))
5734 for (tmp
= BB_END (bb
);
5735 tmp
!= insn
&& tmp
!= LAST_AVAIL_CHECK_FAILURE (ptr
);
5736 tmp
= PREV_INSN (tmp
))
5739 check_available
= 0;
5742 check_available
= store_killed_after (dest
, ptr
->pattern_regs
, insn
,
5744 &LAST_AVAIL_CHECK_FAILURE (ptr
));
5746 if (!check_available
)
5747 AVAIL_STORE_LIST (ptr
) = alloc_INSN_LIST (insn
, AVAIL_STORE_LIST (ptr
));
5750 /* Find available and anticipatable stores. */
5753 compute_store_table (void)
5759 int *last_set_in
, *already_set
;
5760 struct ls_expr
* ptr
, **prev_next_ptr_ptr
;
5762 max_gcse_regno
= max_reg_num ();
5764 reg_set_in_block
= sbitmap_vector_alloc (last_basic_block
,
5766 sbitmap_vector_zero (reg_set_in_block
, last_basic_block
);
5768 pre_ldst_table
= htab_create (13, pre_ldst_expr_hash
,
5769 pre_ldst_expr_eq
, NULL
);
5770 last_set_in
= XCNEWVEC (int, max_gcse_regno
);
5771 already_set
= XNEWVEC (int, max_gcse_regno
);
5773 /* Find all the stores we care about. */
5776 /* First compute the registers set in this block. */
5777 regvec
= last_set_in
;
5779 FOR_BB_INSNS (bb
, insn
)
5781 if (! INSN_P (insn
))
5786 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
5787 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
5789 last_set_in
[regno
] = INSN_UID (insn
);
5790 SET_BIT (reg_set_in_block
[bb
->index
], regno
);
5794 pat
= PATTERN (insn
);
5795 compute_store_table_current_insn
= insn
;
5796 note_stores (pat
, reg_set_info
, reg_set_in_block
[bb
->index
]);
5799 /* Now find the stores. */
5800 memset (already_set
, 0, sizeof (int) * max_gcse_regno
);
5801 regvec
= already_set
;
5802 FOR_BB_INSNS (bb
, insn
)
5804 if (! INSN_P (insn
))
5809 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
5810 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
))
5811 already_set
[regno
] = 1;
5814 pat
= PATTERN (insn
);
5815 note_stores (pat
, reg_set_info
, NULL
);
5817 /* Now that we've marked regs, look for stores. */
5818 find_moveable_store (insn
, already_set
, last_set_in
);
5820 /* Unmark regs that are no longer set. */
5821 compute_store_table_current_insn
= insn
;
5822 note_stores (pat
, reg_clear_last_set
, last_set_in
);
5825 for (regno
= 0; regno
< FIRST_PSEUDO_REGISTER
; regno
++)
5826 if (TEST_HARD_REG_BIT (regs_invalidated_by_call
, regno
)
5827 && last_set_in
[regno
] == INSN_UID (insn
))
5828 last_set_in
[regno
] = 0;
5832 #ifdef ENABLE_CHECKING
5833 /* last_set_in should now be all-zero. */
5834 for (regno
= 0; regno
< max_gcse_regno
; regno
++)
5835 gcc_assert (!last_set_in
[regno
]);
5838 /* Clear temporary marks. */
5839 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
5841 LAST_AVAIL_CHECK_FAILURE(ptr
) = NULL_RTX
;
5842 if (ANTIC_STORE_LIST (ptr
)
5843 && (tmp
= XEXP (ANTIC_STORE_LIST (ptr
), 0)) == NULL_RTX
)
5844 ANTIC_STORE_LIST (ptr
) = XEXP (ANTIC_STORE_LIST (ptr
), 1);
5848 /* Remove the stores that are not available anywhere, as there will
5849 be no opportunity to optimize them. */
5850 for (ptr
= pre_ldst_mems
, prev_next_ptr_ptr
= &pre_ldst_mems
;
5852 ptr
= *prev_next_ptr_ptr
)
5854 if (!AVAIL_STORE_LIST (ptr
))
5856 *prev_next_ptr_ptr
= ptr
->next
;
5857 htab_remove_elt_with_hash (pre_ldst_table
, ptr
, ptr
->hash_index
);
5858 free_ldst_entry (ptr
);
5861 prev_next_ptr_ptr
= &ptr
->next
;
5864 ret
= enumerate_ldsts ();
5868 fprintf (dump_file
, "ST_avail and ST_antic (shown under loads..)\n");
5869 print_ldst_list (dump_file
);
5877 /* Check to see if the load X is aliased with STORE_PATTERN.
5878 AFTER is true if we are checking the case when STORE_PATTERN occurs
5882 load_kills_store (const_rtx x
, const_rtx store_pattern
, int after
)
5885 return anti_dependence (x
, store_pattern
);
5887 return true_dependence (store_pattern
, GET_MODE (store_pattern
), x
,
5891 /* Go through the entire insn X, looking for any loads which might alias
5892 STORE_PATTERN. Return true if found.
5893 AFTER is true if we are checking the case when STORE_PATTERN occurs
5894 after the insn X. */
5897 find_loads (const_rtx x
, const_rtx store_pattern
, int after
)
5906 if (GET_CODE (x
) == SET
)
5911 if (load_kills_store (x
, store_pattern
, after
))
5915 /* Recursively process the insn. */
5916 fmt
= GET_RTX_FORMAT (GET_CODE (x
));
5918 for (i
= GET_RTX_LENGTH (GET_CODE (x
)) - 1; i
>= 0 && !ret
; i
--)
5921 ret
|= find_loads (XEXP (x
, i
), store_pattern
, after
);
5922 else if (fmt
[i
] == 'E')
5923 for (j
= XVECLEN (x
, i
) - 1; j
>= 0; j
--)
5924 ret
|= find_loads (XVECEXP (x
, i
, j
), store_pattern
, after
);
5930 store_killed_in_pat (const_rtx x
, const_rtx pat
, int after
)
5932 if (GET_CODE (pat
) == SET
)
5934 rtx dest
= SET_DEST (pat
);
5936 if (GET_CODE (dest
) == ZERO_EXTRACT
)
5937 dest
= XEXP (dest
, 0);
5939 /* Check for memory stores to aliased objects. */
5941 && !expr_equiv_p (dest
, x
))
5945 if (output_dependence (dest
, x
))
5950 if (output_dependence (x
, dest
))
5956 if (find_loads (pat
, x
, after
))
5962 /* Check if INSN kills the store pattern X (is aliased with it).
5963 AFTER is true if we are checking the case when store X occurs
5964 after the insn. Return true if it does. */
5967 store_killed_in_insn (const_rtx x
, const_rtx x_regs
, const_rtx insn
, int after
)
5969 const_rtx reg
, base
, note
, pat
;
5976 /* A normal or pure call might read from pattern,
5977 but a const call will not. */
5978 if (! CONST_OR_PURE_CALL_P (insn
) || pure_call_p (insn
))
5981 /* But even a const call reads its parameters. Check whether the
5982 base of some of registers used in mem is stack pointer. */
5983 for (reg
= x_regs
; reg
; reg
= XEXP (reg
, 1))
5985 base
= find_base_term (XEXP (reg
, 0));
5987 || (GET_CODE (base
) == ADDRESS
5988 && GET_MODE (base
) == Pmode
5989 && XEXP (base
, 0) == stack_pointer_rtx
))
5996 pat
= PATTERN (insn
);
5997 if (GET_CODE (pat
) == SET
)
5999 if (store_killed_in_pat (x
, pat
, after
))
6002 else if (GET_CODE (pat
) == PARALLEL
)
6006 for (i
= 0; i
< XVECLEN (pat
, 0); i
++)
6007 if (store_killed_in_pat (x
, XVECEXP (pat
, 0, i
), after
))
6010 else if (find_loads (PATTERN (insn
), x
, after
))
6013 /* If this insn has a REG_EQUAL or REG_EQUIV note referencing a memory
6014 location aliased with X, then this insn kills X. */
6015 note
= find_reg_equal_equiv_note (insn
);
6018 note
= XEXP (note
, 0);
6020 /* However, if the note represents a must alias rather than a may
6021 alias relationship, then it does not kill X. */
6022 if (expr_equiv_p (note
, x
))
6025 /* See if there are any aliased loads in the note. */
6026 return find_loads (note
, x
, after
);
6029 /* Returns true if the expression X is loaded or clobbered on or after INSN
6030 within basic block BB. REGS_SET_AFTER is bitmap of registers set in
6031 or after the insn. X_REGS is list of registers mentioned in X. If the store
6032 is killed, return the last insn in that it occurs in FAIL_INSN. */
6035 store_killed_after (const_rtx x
, const_rtx x_regs
, const_rtx insn
, const_basic_block bb
,
6036 int *regs_set_after
, rtx
*fail_insn
)
6038 rtx last
= BB_END (bb
), act
;
6040 if (!store_ops_ok (x_regs
, regs_set_after
))
6042 /* We do not know where it will happen. */
6044 *fail_insn
= NULL_RTX
;
6048 /* Scan from the end, so that fail_insn is determined correctly. */
6049 for (act
= last
; act
!= PREV_INSN (insn
); act
= PREV_INSN (act
))
6050 if (store_killed_in_insn (x
, x_regs
, act
, false))
6060 /* Returns true if the expression X is loaded or clobbered on or before INSN
6061 within basic block BB. X_REGS is list of registers mentioned in X.
6062 REGS_SET_BEFORE is bitmap of registers set before or in this insn. */
6064 store_killed_before (const_rtx x
, const_rtx x_regs
, const_rtx insn
, const_basic_block bb
,
6065 int *regs_set_before
)
6067 rtx first
= BB_HEAD (bb
);
6069 if (!store_ops_ok (x_regs
, regs_set_before
))
6072 for ( ; insn
!= PREV_INSN (first
); insn
= PREV_INSN (insn
))
6073 if (store_killed_in_insn (x
, x_regs
, insn
, true))
6079 /* Fill in available, anticipatable, transparent and kill vectors in
6080 STORE_DATA, based on lists of available and anticipatable stores. */
6082 build_store_vectors (void)
6085 int *regs_set_in_block
;
6087 struct ls_expr
* ptr
;
6090 /* Build the gen_vector. This is any store in the table which is not killed
6091 by aliasing later in its block. */
6092 ae_gen
= sbitmap_vector_alloc (last_basic_block
, num_stores
);
6093 sbitmap_vector_zero (ae_gen
, last_basic_block
);
6095 st_antloc
= sbitmap_vector_alloc (last_basic_block
, num_stores
);
6096 sbitmap_vector_zero (st_antloc
, last_basic_block
);
6098 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6100 for (st
= AVAIL_STORE_LIST (ptr
); st
!= NULL
; st
= XEXP (st
, 1))
6102 insn
= XEXP (st
, 0);
6103 bb
= BLOCK_FOR_INSN (insn
);
6105 /* If we've already seen an available expression in this block,
6106 we can delete this one (It occurs earlier in the block). We'll
6107 copy the SRC expression to an unused register in case there
6108 are any side effects. */
6109 if (TEST_BIT (ae_gen
[bb
->index
], ptr
->index
))
6111 rtx r
= gen_reg_rtx (GET_MODE (ptr
->pattern
));
6113 fprintf (dump_file
, "Removing redundant store:\n");
6114 replace_store_insn (r
, XEXP (st
, 0), bb
, ptr
);
6117 SET_BIT (ae_gen
[bb
->index
], ptr
->index
);
6120 for (st
= ANTIC_STORE_LIST (ptr
); st
!= NULL
; st
= XEXP (st
, 1))
6122 insn
= XEXP (st
, 0);
6123 bb
= BLOCK_FOR_INSN (insn
);
6124 SET_BIT (st_antloc
[bb
->index
], ptr
->index
);
6128 ae_kill
= sbitmap_vector_alloc (last_basic_block
, num_stores
);
6129 sbitmap_vector_zero (ae_kill
, last_basic_block
);
6131 transp
= sbitmap_vector_alloc (last_basic_block
, num_stores
);
6132 sbitmap_vector_zero (transp
, last_basic_block
);
6133 regs_set_in_block
= XNEWVEC (int, max_gcse_regno
);
6137 for (regno
= 0; regno
< max_gcse_regno
; regno
++)
6138 regs_set_in_block
[regno
] = TEST_BIT (reg_set_in_block
[bb
->index
], regno
);
6140 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6142 if (store_killed_after (ptr
->pattern
, ptr
->pattern_regs
, BB_HEAD (bb
),
6143 bb
, regs_set_in_block
, NULL
))
6145 /* It should not be necessary to consider the expression
6146 killed if it is both anticipatable and available. */
6147 if (!TEST_BIT (st_antloc
[bb
->index
], ptr
->index
)
6148 || !TEST_BIT (ae_gen
[bb
->index
], ptr
->index
))
6149 SET_BIT (ae_kill
[bb
->index
], ptr
->index
);
6152 SET_BIT (transp
[bb
->index
], ptr
->index
);
6156 free (regs_set_in_block
);
6160 dump_sbitmap_vector (dump_file
, "st_antloc", "", st_antloc
, last_basic_block
);
6161 dump_sbitmap_vector (dump_file
, "st_kill", "", ae_kill
, last_basic_block
);
6162 dump_sbitmap_vector (dump_file
, "Transpt", "", transp
, last_basic_block
);
6163 dump_sbitmap_vector (dump_file
, "st_avloc", "", ae_gen
, last_basic_block
);
6167 /* Insert an instruction at the beginning of a basic block, and update
6168 the BB_HEAD if needed. */
6171 insert_insn_start_basic_block (rtx insn
, basic_block bb
)
6173 /* Insert at start of successor block. */
6174 rtx prev
= PREV_INSN (BB_HEAD (bb
));
6175 rtx before
= BB_HEAD (bb
);
6178 if (! LABEL_P (before
)
6179 && !NOTE_INSN_BASIC_BLOCK_P (before
))
6182 if (prev
== BB_END (bb
))
6184 before
= NEXT_INSN (before
);
6187 insn
= emit_insn_after_noloc (insn
, prev
, bb
);
6191 fprintf (dump_file
, "STORE_MOTION insert store at start of BB %d:\n",
6193 print_inline_rtx (dump_file
, insn
, 6);
6194 fprintf (dump_file
, "\n");
6198 /* This routine will insert a store on an edge. EXPR is the ldst entry for
6199 the memory reference, and E is the edge to insert it on. Returns nonzero
6200 if an edge insertion was performed. */
6203 insert_store (struct ls_expr
* expr
, edge e
)
6210 /* We did all the deleted before this insert, so if we didn't delete a
6211 store, then we haven't set the reaching reg yet either. */
6212 if (expr
->reaching_reg
== NULL_RTX
)
6215 if (e
->flags
& EDGE_FAKE
)
6218 reg
= expr
->reaching_reg
;
6219 insn
= gen_move_insn (copy_rtx (expr
->pattern
), reg
);
6221 /* If we are inserting this expression on ALL predecessor edges of a BB,
6222 insert it at the start of the BB, and reset the insert bits on the other
6223 edges so we don't try to insert it on the other edges. */
6225 FOR_EACH_EDGE (tmp
, ei
, e
->dest
->preds
)
6226 if (!(tmp
->flags
& EDGE_FAKE
))
6228 int index
= EDGE_INDEX (edge_list
, tmp
->src
, tmp
->dest
);
6230 gcc_assert (index
!= EDGE_INDEX_NO_EDGE
);
6231 if (! TEST_BIT (pre_insert_map
[index
], expr
->index
))
6235 /* If tmp is NULL, we found an insertion on every edge, blank the
6236 insertion vector for these edges, and insert at the start of the BB. */
6237 if (!tmp
&& bb
!= EXIT_BLOCK_PTR
)
6239 FOR_EACH_EDGE (tmp
, ei
, e
->dest
->preds
)
6241 int index
= EDGE_INDEX (edge_list
, tmp
->src
, tmp
->dest
);
6242 RESET_BIT (pre_insert_map
[index
], expr
->index
);
6244 insert_insn_start_basic_block (insn
, bb
);
6248 /* We can't put stores in the front of blocks pointed to by abnormal
6249 edges since that may put a store where one didn't used to be. */
6250 gcc_assert (!(e
->flags
& EDGE_ABNORMAL
));
6252 insert_insn_on_edge (insn
, e
);
6256 fprintf (dump_file
, "STORE_MOTION insert insn on edge (%d, %d):\n",
6257 e
->src
->index
, e
->dest
->index
);
6258 print_inline_rtx (dump_file
, insn
, 6);
6259 fprintf (dump_file
, "\n");
6265 /* Remove any REG_EQUAL or REG_EQUIV notes containing a reference to the
6266 memory location in SMEXPR set in basic block BB.
6268 This could be rather expensive. */
6271 remove_reachable_equiv_notes (basic_block bb
, struct ls_expr
*smexpr
)
6273 edge_iterator
*stack
, ei
;
6276 sbitmap visited
= sbitmap_alloc (last_basic_block
);
6277 rtx last
, insn
, note
;
6278 rtx mem
= smexpr
->pattern
;
6280 stack
= XNEWVEC (edge_iterator
, n_basic_blocks
);
6282 ei
= ei_start (bb
->succs
);
6284 sbitmap_zero (visited
);
6286 act
= (EDGE_COUNT (ei_container (ei
)) > 0 ? EDGE_I (ei_container (ei
), 0) : NULL
);
6294 sbitmap_free (visited
);
6297 act
= ei_edge (stack
[--sp
]);
6301 if (bb
== EXIT_BLOCK_PTR
6302 || TEST_BIT (visited
, bb
->index
))
6306 act
= (! ei_end_p (ei
)) ? ei_edge (ei
) : NULL
;
6309 SET_BIT (visited
, bb
->index
);
6311 if (TEST_BIT (st_antloc
[bb
->index
], smexpr
->index
))
6313 for (last
= ANTIC_STORE_LIST (smexpr
);
6314 BLOCK_FOR_INSN (XEXP (last
, 0)) != bb
;
6315 last
= XEXP (last
, 1))
6317 last
= XEXP (last
, 0);
6320 last
= NEXT_INSN (BB_END (bb
));
6322 for (insn
= BB_HEAD (bb
); insn
!= last
; insn
= NEXT_INSN (insn
))
6325 note
= find_reg_equal_equiv_note (insn
);
6326 if (!note
|| !expr_equiv_p (XEXP (note
, 0), mem
))
6330 fprintf (dump_file
, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
6332 remove_note (insn
, note
);
6337 act
= (! ei_end_p (ei
)) ? ei_edge (ei
) : NULL
;
6339 if (EDGE_COUNT (bb
->succs
) > 0)
6343 ei
= ei_start (bb
->succs
);
6344 act
= (EDGE_COUNT (ei_container (ei
)) > 0 ? EDGE_I (ei_container (ei
), 0) : NULL
);
6349 /* This routine will replace a store with a SET to a specified register. */
6352 replace_store_insn (rtx reg
, rtx del
, basic_block bb
, struct ls_expr
*smexpr
)
6354 rtx insn
, mem
, note
, set
, ptr
, pair
;
6356 mem
= smexpr
->pattern
;
6357 insn
= gen_move_insn (reg
, SET_SRC (single_set (del
)));
6359 for (ptr
= ANTIC_STORE_LIST (smexpr
); ptr
; ptr
= XEXP (ptr
, 1))
6360 if (XEXP (ptr
, 0) == del
)
6362 XEXP (ptr
, 0) = insn
;
6366 /* Move the notes from the deleted insn to its replacement, and patch
6367 up the LIBCALL notes. */
6368 REG_NOTES (insn
) = REG_NOTES (del
);
6370 note
= find_reg_note (insn
, REG_RETVAL
, NULL_RTX
);
6373 pair
= XEXP (note
, 0);
6374 note
= find_reg_note (pair
, REG_LIBCALL
, NULL_RTX
);
6375 XEXP (note
, 0) = insn
;
6377 note
= find_reg_note (insn
, REG_LIBCALL
, NULL_RTX
);
6380 pair
= XEXP (note
, 0);
6381 note
= find_reg_note (pair
, REG_RETVAL
, NULL_RTX
);
6382 XEXP (note
, 0) = insn
;
6385 /* Emit the insn AFTER all the notes are transferred.
6386 This is cheaper since we avoid df rescanning for the note change. */
6387 insn
= emit_insn_after (insn
, del
);
6392 "STORE_MOTION delete insn in BB %d:\n ", bb
->index
);
6393 print_inline_rtx (dump_file
, del
, 6);
6394 fprintf (dump_file
, "\nSTORE MOTION replaced with insn:\n ");
6395 print_inline_rtx (dump_file
, insn
, 6);
6396 fprintf (dump_file
, "\n");
6401 /* Now we must handle REG_EQUAL notes whose contents is equal to the mem;
6402 they are no longer accurate provided that they are reached by this
6403 definition, so drop them. */
6404 for (; insn
!= NEXT_INSN (BB_END (bb
)); insn
= NEXT_INSN (insn
))
6407 set
= single_set (insn
);
6410 if (expr_equiv_p (SET_DEST (set
), mem
))
6412 note
= find_reg_equal_equiv_note (insn
);
6413 if (!note
|| !expr_equiv_p (XEXP (note
, 0), mem
))
6417 fprintf (dump_file
, "STORE_MOTION drop REG_EQUAL note at insn %d:\n",
6419 remove_note (insn
, note
);
6421 remove_reachable_equiv_notes (bb
, smexpr
);
6425 /* Delete a store, but copy the value that would have been stored into
6426 the reaching_reg for later storing. */
6429 delete_store (struct ls_expr
* expr
, basic_block bb
)
6433 if (expr
->reaching_reg
== NULL_RTX
)
6434 expr
->reaching_reg
= gen_reg_rtx (GET_MODE (expr
->pattern
));
6436 reg
= expr
->reaching_reg
;
6438 for (i
= AVAIL_STORE_LIST (expr
); i
; i
= XEXP (i
, 1))
6441 if (BLOCK_FOR_INSN (del
) == bb
)
6443 /* We know there is only one since we deleted redundant
6444 ones during the available computation. */
6445 replace_store_insn (reg
, del
, bb
, expr
);
6451 /* Free memory used by store motion. */
6454 free_store_memory (void)
6459 sbitmap_vector_free (ae_gen
);
6461 sbitmap_vector_free (ae_kill
);
6463 sbitmap_vector_free (transp
);
6465 sbitmap_vector_free (st_antloc
);
6467 sbitmap_vector_free (pre_insert_map
);
6469 sbitmap_vector_free (pre_delete_map
);
6470 if (reg_set_in_block
)
6471 sbitmap_vector_free (reg_set_in_block
);
6473 ae_gen
= ae_kill
= transp
= st_antloc
= NULL
;
6474 pre_insert_map
= pre_delete_map
= reg_set_in_block
= NULL
;
6477 /* Perform store motion. Much like gcse, except we move expressions the
6478 other way by looking at the flowgraph in reverse. */
6485 struct ls_expr
* ptr
;
6486 int update_flow
= 0;
6490 fprintf (dump_file
, "before store motion\n");
6491 print_rtl (dump_file
, get_insns ());
6494 init_alias_analysis ();
6496 /* Find all the available and anticipatable stores. */
6497 num_stores
= compute_store_table ();
6498 if (num_stores
== 0)
6500 htab_delete (pre_ldst_table
);
6501 pre_ldst_table
= NULL
;
6502 sbitmap_vector_free (reg_set_in_block
);
6503 end_alias_analysis ();
6507 /* Now compute kill & transp vectors. */
6508 build_store_vectors ();
6509 add_noreturn_fake_exit_edges ();
6510 connect_infinite_loops_to_exit ();
6512 edge_list
= pre_edge_rev_lcm (num_stores
, transp
, ae_gen
,
6513 st_antloc
, ae_kill
, &pre_insert_map
,
6516 /* Now we want to insert the new stores which are going to be needed. */
6517 for (ptr
= first_ls_expr (); ptr
!= NULL
; ptr
= next_ls_expr (ptr
))
6519 /* If any of the edges we have above are abnormal, we can't move this
6521 for (x
= NUM_EDGES (edge_list
) - 1; x
>= 0; x
--)
6522 if (TEST_BIT (pre_insert_map
[x
], ptr
->index
)
6523 && (INDEX_EDGE (edge_list
, x
)->flags
& EDGE_ABNORMAL
))
6528 if (dump_file
!= NULL
)
6530 "Can't replace store %d: abnormal edge from %d to %d\n",
6531 ptr
->index
, INDEX_EDGE (edge_list
, x
)->src
->index
,
6532 INDEX_EDGE (edge_list
, x
)->dest
->index
);
6536 /* Now we want to insert the new stores which are going to be needed. */
6539 if (TEST_BIT (pre_delete_map
[bb
->index
], ptr
->index
))
6540 delete_store (ptr
, bb
);
6542 for (x
= 0; x
< NUM_EDGES (edge_list
); x
++)
6543 if (TEST_BIT (pre_insert_map
[x
], ptr
->index
))
6544 update_flow
|= insert_store (ptr
, INDEX_EDGE (edge_list
, x
));
6548 commit_edge_insertions ();
6550 free_store_memory ();
6551 free_edge_list (edge_list
);
6552 remove_fake_exit_edges ();
6553 end_alias_analysis ();
6557 /* Entry point for jump bypassing optimization pass. */
6564 /* We do not construct an accurate cfg in functions which call
6565 setjmp, so just punt to be safe. */
6566 if (current_function_calls_setjmp
)
6569 /* Identify the basic block information for this function, including
6570 successors and predecessors. */
6571 max_gcse_regno
= max_reg_num ();
6574 dump_flow_info (dump_file
, dump_flags
);
6576 /* Return if there's nothing to do, or it is too expensive. */
6577 if (n_basic_blocks
<= NUM_FIXED_BLOCKS
+ 1
6578 || is_too_expensive (_ ("jump bypassing disabled")))
6581 gcc_obstack_init (&gcse_obstack
);
6584 /* We need alias. */
6585 init_alias_analysis ();
6587 /* Record where pseudo-registers are set. This data is kept accurate
6588 during each pass. ??? We could also record hard-reg information here
6589 [since it's unchanging], however it is currently done during hash table
6592 It may be tempting to compute MEM set information here too, but MEM sets
6593 will be subject to code motion one day and thus we need to compute
6594 information about memory sets when we build the hash tables. */
6596 alloc_reg_set_mem (max_gcse_regno
);
6599 max_gcse_regno
= max_reg_num ();
6601 changed
= one_cprop_pass (MAX_GCSE_PASSES
+ 2, true, true);
6606 fprintf (dump_file
, "BYPASS of %s: %d basic blocks, ",
6607 current_function_name (), n_basic_blocks
);
6608 fprintf (dump_file
, "%d bytes\n\n", bytes_used
);
6611 obstack_free (&gcse_obstack
, NULL
);
6612 free_reg_set_mem ();
6614 /* We are finished with alias. */
6615 end_alias_analysis ();
6620 /* Return true if the graph is too expensive to optimize. PASS is the
6621 optimization about to be performed. */
6624 is_too_expensive (const char *pass
)
6626 /* Trying to perform global optimizations on flow graphs which have
6627 a high connectivity will take a long time and is unlikely to be
6628 particularly useful.
6630 In normal circumstances a cfg should have about twice as many
6631 edges as blocks. But we do not want to punish small functions
6632 which have a couple switch statements. Rather than simply
6633 threshold the number of blocks, uses something with a more
6634 graceful degradation. */
6635 if (n_edges
> 20000 + n_basic_blocks
* 4)
6637 warning (OPT_Wdisabled_optimization
,
6638 "%s: %d basic blocks and %d edges/basic block",
6639 pass
, n_basic_blocks
, n_edges
/ n_basic_blocks
);
6644 /* If allocating memory for the cprop bitmap would take up too much
6645 storage it's better just to disable the optimization. */
6647 * SBITMAP_SET_SIZE (max_reg_num ())
6648 * sizeof (SBITMAP_ELT_TYPE
)) > MAX_GCSE_MEMORY
)
6650 warning (OPT_Wdisabled_optimization
,
6651 "%s: %d basic blocks and %d registers",
6652 pass
, n_basic_blocks
, max_reg_num ());
6661 gate_handle_jump_bypass (void)
6663 return optimize
> 0 && flag_gcse
;
6666 /* Perform jump bypassing and control flow optimizations. */
6668 rest_of_handle_jump_bypass (void)
6670 delete_unreachable_blocks ();
6671 if (bypass_jumps ())
6673 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6674 rebuild_jump_labels (get_insns ());
6680 struct tree_opt_pass pass_jump_bypass
=
6682 "bypass", /* name */
6683 gate_handle_jump_bypass
, /* gate */
6684 rest_of_handle_jump_bypass
, /* execute */
6687 0, /* static_pass_number */
6688 TV_BYPASS
, /* tv_id */
6689 0, /* properties_required */
6690 0, /* properties_provided */
6691 0, /* properties_destroyed */
6692 0, /* todo_flags_start */
6694 TODO_ggc_collect
| TODO_verify_flow
, /* todo_flags_finish */
6700 gate_handle_gcse (void)
6702 return optimize
> 0 && flag_gcse
;
6707 rest_of_handle_gcse (void)
6709 int save_csb
, save_cfj
;
6711 tem
= gcse_main (get_insns ());
6712 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6713 rebuild_jump_labels (get_insns ());
6714 save_csb
= flag_cse_skip_blocks
;
6715 save_cfj
= flag_cse_follow_jumps
;
6716 flag_cse_skip_blocks
= flag_cse_follow_jumps
= 0;
6718 /* If -fexpensive-optimizations, re-run CSE to clean up things done
6720 if (flag_expensive_optimizations
)
6722 timevar_push (TV_CSE
);
6723 tem2
= cse_main (get_insns (), max_reg_num ());
6724 df_finish_pass (false);
6725 purge_all_dead_edges ();
6726 delete_trivially_dead_insns (get_insns (), max_reg_num ());
6727 timevar_pop (TV_CSE
);
6728 cse_not_expected
= !flag_rerun_cse_after_loop
;
6731 /* If gcse or cse altered any jumps, rerun jump optimizations to clean
6735 timevar_push (TV_JUMP
);
6736 rebuild_jump_labels (get_insns ());
6738 timevar_pop (TV_JUMP
);
6741 flag_cse_skip_blocks
= save_csb
;
6742 flag_cse_follow_jumps
= save_cfj
;
6746 struct tree_opt_pass pass_gcse
=
6749 gate_handle_gcse
, /* gate */
6750 rest_of_handle_gcse
, /* execute */
6753 0, /* static_pass_number */
6754 TV_GCSE
, /* tv_id */
6755 0, /* properties_required */
6756 0, /* properties_provided */
6757 0, /* properties_destroyed */
6758 0, /* todo_flags_start */
6759 TODO_df_finish
| TODO_verify_rtl_sharing
|
6761 TODO_verify_flow
| TODO_ggc_collect
, /* todo_flags_finish */
6766 #include "gt-gcse.h"