1 /* Global common subexpression elimination/Partial redundancy elimination
2 and global constant/copy propagation for GNU compiler.
3 Copyright (C) 1997, 1998, 1999, 2000, 2001, 2002, 2003, 2004, 2005,
4 2006, 2007, 2008, 2009, 2010 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.
33 /* References searched while implementing this.
35 Compilers Principles, Techniques and Tools
39 Global Optimization by Suppression of Partial Redundancies
41 communications of the acm, Vol. 22, Num. 2, Feb. 1979
43 A Portable Machine-Independent Global Optimizer - Design and Measurements
45 Stanford Ph.D. thesis, Dec. 1983
47 A Fast Algorithm for Code Movement Optimization
49 SIGPLAN Notices, Vol. 23, Num. 10, Oct. 1988
51 A Solution to a Problem with Morel and Renvoise's
52 Global Optimization by Suppression of Partial Redundancies
53 K-H Drechsler, M.P. Stadel
54 ACM TOPLAS, Vol. 10, Num. 4, Oct. 1988
56 Practical Adaptation of the Global Optimization
57 Algorithm of Morel and Renvoise
59 ACM TOPLAS, Vol. 13, Num. 2. Apr. 1991
61 Efficiently Computing Static Single Assignment Form and the Control
63 R. Cytron, J. Ferrante, B.K. Rosen, M.N. Wegman, and F.K. Zadeck
64 ACM TOPLAS, Vol. 13, Num. 4, Oct. 1991
67 J. Knoop, O. Ruthing, B. Steffen
68 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
70 What's In a Region? Or Computing Control Dependence Regions in Near-Linear
71 Time for Reducible Flow Control
73 ACM Letters on Programming Languages and Systems,
74 Vol. 2, Num. 1-4, Mar-Dec 1993
76 An Efficient Representation for Sparse Sets
77 Preston Briggs, Linda Torczon
78 ACM Letters on Programming Languages and Systems,
79 Vol. 2, Num. 1-4, Mar-Dec 1993
81 A Variation of Knoop, Ruthing, and Steffen's Lazy Code Motion
82 K-H Drechsler, M.P. Stadel
83 ACM SIGPLAN Notices, Vol. 28, Num. 5, May 1993
85 Partial Dead Code Elimination
86 J. Knoop, O. Ruthing, B. Steffen
87 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
89 Effective Partial Redundancy Elimination
90 P. Briggs, K.D. Cooper
91 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
93 The Program Structure Tree: Computing Control Regions in Linear Time
94 R. Johnson, D. Pearson, K. Pingali
95 ACM SIGPLAN Notices, Vol. 29, Num. 6, Jun. 1994
97 Optimal Code Motion: Theory and Practice
98 J. Knoop, O. Ruthing, B. Steffen
99 ACM TOPLAS, Vol. 16, Num. 4, Jul. 1994
101 The power of assignment motion
102 J. Knoop, O. Ruthing, B. Steffen
103 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
105 Global code motion / global value numbering
107 ACM SIGPLAN Notices Vol. 30, Num. 6, Jun. 1995, '95 Conference on PLDI
109 Value Driven Redundancy Elimination
111 Rice University Ph.D. thesis, Apr. 1996
115 Massively Scalar Compiler Project, Rice University, Sep. 1996
117 High Performance Compilers for Parallel Computing
121 Advanced Compiler Design and Implementation
123 Morgan Kaufmann, 1997
125 Building an Optimizing Compiler
129 People wishing to speed up the code here should read:
130 Elimination Algorithms for Data Flow Analysis
131 B.G. Ryder, M.C. Paull
132 ACM Computing Surveys, Vol. 18, Num. 3, Sep. 1986
134 How to Analyze Large Programs Efficiently and Informatively
135 D.M. Dhamdhere, B.K. Rosen, F.K. Zadeck
136 ACM SIGPLAN Notices Vol. 27, Num. 7, Jul. 1992, '92 Conference on PLDI
138 People wishing to do something different can find various possibilities
139 in the above papers and elsewhere.
144 #include "coretypes.h"
146 #include "diagnostic-core.h"
153 #include "hard-reg-set.h"
155 #include "insn-config.h"
157 #include "basic-block.h"
159 #include "function.h"
168 #include "tree-pass.h"
175 /* We support GCSE via Partial Redundancy Elimination. PRE optimizations
176 are a superset of those done by classic GCSE.
178 We perform the following steps:
180 1) Compute table of places where registers are set.
182 2) Perform copy/constant propagation.
184 3) Perform global cse using lazy code motion if not optimizing
185 for size, or code hoisting if we are.
187 4) Perform another pass of copy/constant propagation. Try to bypass
188 conditional jumps if the condition can be computed from a value of
191 Two passes of copy/constant propagation are done because the first one
192 enables more GCSE and the second one helps to clean up the copies that
193 GCSE creates. This is needed more for PRE than for Classic because Classic
194 GCSE will try to use an existing register containing the common
195 subexpression rather than create a new one. This is harder to do for PRE
196 because of the code motion (which Classic GCSE doesn't do).
198 Expressions we are interested in GCSE-ing are of the form
199 (set (pseudo-reg) (expression)).
200 Function want_to_gcse_p says what these are.
202 In addition, expressions in REG_EQUAL notes are candidates for GCSE-ing.
203 This allows PRE to hoist expressions that are expressed in multiple insns,
204 such as complex address calculations (e.g. for PIC code, or loads with a
205 high part and a low part).
207 PRE handles moving invariant expressions out of loops (by treating them as
208 partially redundant).
210 **********************
212 We used to support multiple passes but there are diminishing returns in
213 doing so. The first pass usually makes 90% of the changes that are doable.
214 A second pass can make a few more changes made possible by the first pass.
215 Experiments show any further passes don't make enough changes to justify
218 A study of spec92 using an unlimited number of passes:
219 [1 pass] = 1208 substitutions, [2] = 577, [3] = 202, [4] = 192, [5] = 83,
220 [6] = 34, [7] = 17, [8] = 9, [9] = 4, [10] = 4, [11] = 2,
221 [12] = 2, [13] = 1, [15] = 1, [16] = 2, [41] = 1
223 It was found doing copy propagation between each pass enables further
226 This study was done before expressions in REG_EQUAL notes were added as
227 candidate expressions for optimization, and before the GIMPLE optimizers
228 were added. Probably, multiple passes is even less efficient now than
229 at the time when the study was conducted.
231 PRE is quite expensive in complicated functions because the DFA can take
232 a while to converge. Hence we only perform one pass.
234 **********************
236 The steps for PRE are:
238 1) Build the hash table of expressions we wish to GCSE (expr_hash_table).
240 2) Perform the data flow analysis for PRE.
242 3) Delete the redundant instructions
244 4) Insert the required copies [if any] that make the partially
245 redundant instructions fully redundant.
247 5) For other reaching expressions, insert an instruction to copy the value
248 to a newly created pseudo that will reach the redundant instruction.
250 The deletion is done first so that when we do insertions we
251 know which pseudo reg to use.
253 Various papers have argued that PRE DFA is expensive (O(n^2)) and others
254 argue it is not. The number of iterations for the algorithm to converge
255 is typically 2-4 so I don't view it as that expensive (relatively speaking).
257 PRE GCSE depends heavily on the second CPROP pass to clean up the copies
258 we create. To make an expression reach the place where it's redundant,
259 the result of the expression is copied to a new register, and the redundant
260 expression is deleted by replacing it with this new register. Classic GCSE
261 doesn't have this problem as much as it computes the reaching defs of
262 each register in each block and thus can try to use an existing
265 /* GCSE global vars. */
267 struct target_gcse default_target_gcse;
268 #if SWITCHABLE_TARGET
269 struct target_gcse *this_target_gcse = &default_target_gcse;
272 /* Set to non-zero if CSE should run after all GCSE optimizations are done. */
273 int flag_rerun_cse_after_global_opts;
275 /* An obstack for our working variables. */
276 static struct obstack gcse_obstack;
278 struct reg_use {rtx reg_rtx; };
280 /* Hash table of expressions. */
284 /* The expression (SET_SRC for expressions, PATTERN for assignments). */
286 /* Index in the available expression bitmaps. */
288 /* Next entry with the same hash. */
289 struct expr *next_same_hash;
290 /* List of anticipatable occurrences in basic blocks in the function.
291 An "anticipatable occurrence" is one that is the first occurrence in the
292 basic block, the operands are not modified in the basic block prior
293 to the occurrence and the output is not used between the start of
294 the block and the occurrence. */
295 struct occr *antic_occr;
296 /* List of available occurrence in basic blocks in the function.
297 An "available occurrence" is one that is the last occurrence in the
298 basic block and the operands are not modified by following statements in
299 the basic block [including this insn]. */
300 struct occr *avail_occr;
301 /* Non-null if the computation is PRE redundant.
302 The value is the newly created pseudo-reg to record a copy of the
303 expression in all the places that reach the redundant copy. */
307 /* Occurrence of an expression.
308 There is one per basic block. If a pattern appears more than once the
309 last appearance is used [or first for anticipatable expressions]. */
313 /* Next occurrence of this expression. */
315 /* The insn that computes the expression. */
317 /* Nonzero if this [anticipatable] occurrence has been deleted. */
319 /* Nonzero if this [available] occurrence has been copied to
321 /* ??? This is mutually exclusive with deleted_p, so they could share
326 /* Expression and copy propagation hash tables.
327 Each hash table is an array of buckets.
328 ??? It is known that if it were an array of entries, structure elements
329 `next_same_hash' and `bitmap_index' wouldn't be necessary. However, it is
330 not clear whether in the final analysis a sufficient amount of memory would
331 be saved as the size of the available expression bitmaps would be larger
332 [one could build a mapping table without holes afterwards though].
333 Someday I'll perform the computation and figure it out. */
338 This is an array of `expr_hash_table_size' elements. */
341 /* Size of the hash table, in elements. */
344 /* Number of hash table elements. */
345 unsigned int n_elems;
347 /* Whether the table is expression of copy propagation one. */
351 /* Expression hash table. */
352 static struct hash_table_d expr_hash_table;
354 /* Copy propagation hash table. */
355 static struct hash_table_d set_hash_table;
357 /* This is a list of expressions which are MEMs and will be used by load
359 Load motion tracks MEMs which aren't killed by
360 anything except itself. (i.e., loads and stores to a single location).
361 We can then allow movement of these MEM refs with a little special
362 allowance. (all stores copy the same value to the reaching reg used
363 for the loads). This means all values used to store into memory must have
364 no side effects so we can re-issue the setter value.
365 Store Motion uses this structure as an expression table to track stores
366 which look interesting, and might be moveable towards the exit block. */
370 struct expr * expr; /* Gcse expression reference for LM. */
371 rtx pattern; /* Pattern of this mem. */
372 rtx pattern_regs; /* List of registers mentioned by the mem. */
373 rtx loads; /* INSN list of loads seen. */
374 rtx stores; /* INSN list of stores seen. */
375 struct ls_expr * next; /* Next in the list. */
376 int invalid; /* Invalid for some reason. */
377 int index; /* If it maps to a bitmap index. */
378 unsigned int hash_index; /* Index when in a hash table. */
379 rtx reaching_reg; /* Register to use when re-writing. */
382 /* Array of implicit set patterns indexed by basic block index. */
383 static rtx *implicit_sets;
385 /* Head of the list of load/store memory refs. */
386 static struct ls_expr * pre_ldst_mems = NULL;
388 /* Hashtable for the load/store memory refs. */
389 static htab_t pre_ldst_table = NULL;
391 /* Bitmap containing one bit for each register in the program.
392 Used when performing GCSE to track which registers have been set since
393 the start of the basic block. */
394 static regset reg_set_bitmap;
396 /* Array, indexed by basic block number for a list of insns which modify
397 memory within that block. */
398 static rtx * modify_mem_list;
399 static bitmap modify_mem_list_set;
401 /* This array parallels modify_mem_list, but is kept canonicalized. */
402 static rtx * canon_modify_mem_list;
404 /* Bitmap indexed by block numbers to record which blocks contain
406 static bitmap blocks_with_calls;
408 /* Various variables for statistics gathering. */
410 /* Memory used in a pass.
411 This isn't intended to be absolutely precise. Its intent is only
412 to keep an eye on memory usage. */
413 static int bytes_used;
415 /* GCSE substitutions made. */
416 static int gcse_subst_count;
417 /* Number of copy instructions created. */
418 static int gcse_create_count;
419 /* Number of local constants propagated. */
420 static int local_const_prop_count;
421 /* Number of local copies propagated. */
422 static int local_copy_prop_count;
423 /* Number of global constants propagated. */
424 static int global_const_prop_count;
425 /* Number of global copies propagated. */
426 static int global_copy_prop_count;
428 /* For available exprs */
429 static sbitmap *ae_kill;
431 static void compute_can_copy (void);
432 static void *gmalloc (size_t) ATTRIBUTE_MALLOC;
433 static void *gcalloc (size_t, size_t) ATTRIBUTE_MALLOC;
434 static void *gcse_alloc (unsigned long);
435 static void alloc_gcse_mem (void);
436 static void free_gcse_mem (void);
437 static void hash_scan_insn (rtx, struct hash_table_d *);
438 static void hash_scan_set (rtx, rtx, struct hash_table_d *);
439 static void hash_scan_clobber (rtx, rtx, struct hash_table_d *);
440 static void hash_scan_call (rtx, rtx, struct hash_table_d *);
441 static int want_to_gcse_p (rtx);
442 static bool gcse_constant_p (const_rtx);
443 static int oprs_unchanged_p (const_rtx, const_rtx, int);
444 static int oprs_anticipatable_p (const_rtx, const_rtx);
445 static int oprs_available_p (const_rtx, const_rtx);
446 static void insert_expr_in_table (rtx, enum machine_mode, rtx, int, int,
447 struct hash_table_d *);
448 static void insert_set_in_table (rtx, rtx, struct hash_table_d *);
449 static unsigned int hash_expr (const_rtx, enum machine_mode, int *, int);
450 static unsigned int hash_set (int, int);
451 static int expr_equiv_p (const_rtx, const_rtx);
452 static void record_last_reg_set_info (rtx, int);
453 static void record_last_mem_set_info (rtx);
454 static void record_last_set_info (rtx, const_rtx, void *);
455 static void compute_hash_table (struct hash_table_d *);
456 static void alloc_hash_table (struct hash_table_d *, int);
457 static void free_hash_table (struct hash_table_d *);
458 static void compute_hash_table_work (struct hash_table_d *);
459 static void dump_hash_table (FILE *, const char *, struct hash_table_d *);
460 static struct expr *lookup_set (unsigned int, struct hash_table_d *);
461 static struct expr *next_set (unsigned int, struct expr *);
462 static void reset_opr_set_tables (void);
463 static int oprs_not_set_p (const_rtx, const_rtx);
464 static void mark_call (rtx);
465 static void mark_set (rtx, rtx);
466 static void mark_clobber (rtx, rtx);
467 static void mark_oprs_set (rtx);
468 static void alloc_cprop_mem (int, int);
469 static void free_cprop_mem (void);
470 static void compute_transp (const_rtx, int, sbitmap *, int);
471 static void compute_transpout (void);
472 static void compute_local_properties (sbitmap *, sbitmap *, sbitmap *,
473 struct hash_table_d *);
474 static void compute_cprop_data (void);
475 static void find_used_regs (rtx *, void *);
476 static int try_replace_reg (rtx, rtx, rtx);
477 static struct expr *find_avail_set (int, rtx);
478 static int cprop_jump (basic_block, rtx, rtx, rtx, rtx);
479 static void mems_conflict_for_gcse_p (rtx, const_rtx, void *);
480 static int load_killed_in_block_p (const_basic_block, int, const_rtx, int);
481 static void canon_list_insert (rtx, const_rtx, void *);
482 static int cprop_insn (rtx);
483 static void find_implicit_sets (void);
484 static int one_cprop_pass (void);
485 static bool constprop_register (rtx, rtx, rtx);
486 static struct expr *find_bypass_set (int, int);
487 static bool reg_killed_on_edge (const_rtx, const_edge);
488 static int bypass_block (basic_block, rtx, rtx);
489 static int bypass_conditional_jumps (void);
490 static void alloc_pre_mem (int, int);
491 static void free_pre_mem (void);
492 static void compute_pre_data (void);
493 static int pre_expr_reaches_here_p (basic_block, struct expr *,
495 static void insert_insn_end_basic_block (struct expr *, basic_block, int);
496 static void pre_insert_copy_insn (struct expr *, rtx);
497 static void pre_insert_copies (void);
498 static int pre_delete (void);
499 static int pre_gcse (void);
500 static int one_pre_gcse_pass (void);
501 static void add_label_notes (rtx, rtx);
502 static void alloc_code_hoist_mem (int, int);
503 static void free_code_hoist_mem (void);
504 static void compute_code_hoist_vbeinout (void);
505 static void compute_code_hoist_data (void);
506 static int hoist_expr_reaches_here_p (basic_block, int, basic_block, char *);
507 static int hoist_code (void);
508 static int one_code_hoisting_pass (void);
509 static rtx process_insert_insn (struct expr *);
510 static int pre_edge_insert (struct edge_list *, struct expr **);
511 static int pre_expr_reaches_here_p_work (basic_block, struct expr *,
512 basic_block, char *);
513 static struct ls_expr * ldst_entry (rtx);
514 static void free_ldst_entry (struct ls_expr *);
515 static void free_ldst_mems (void);
516 static void print_ldst_list (FILE *);
517 static struct ls_expr * find_rtx_in_ldst (rtx);
518 static inline struct ls_expr * first_ls_expr (void);
519 static inline struct ls_expr * next_ls_expr (struct ls_expr *);
520 static int simple_mem (const_rtx);
521 static void invalidate_any_buried_refs (rtx);
522 static void compute_ld_motion_mems (void);
523 static void trim_ld_motion_mems (void);
524 static void update_ld_motion_stores (struct expr *);
525 static void free_insn_expr_list_list (rtx *);
526 static void clear_modify_mem_tables (void);
527 static void free_modify_mem_tables (void);
528 static rtx gcse_emit_move_after (rtx, rtx, rtx);
529 static void local_cprop_find_used_regs (rtx *, void *);
530 static bool do_local_cprop (rtx, rtx);
531 static int local_cprop_pass (void);
532 static bool is_too_expensive (const char *);
534 #define GNEW(T) ((T *) gmalloc (sizeof (T)))
535 #define GCNEW(T) ((T *) gcalloc (1, sizeof (T)))
537 #define GNEWVEC(T, N) ((T *) gmalloc (sizeof (T) * (N)))
538 #define GCNEWVEC(T, N) ((T *) gcalloc ((N), sizeof (T)))
540 #define GNEWVAR(T, S) ((T *) gmalloc ((S)))
541 #define GCNEWVAR(T, S) ((T *) gcalloc (1, (S)))
543 #define GOBNEW(T) ((T *) gcse_alloc (sizeof (T)))
544 #define GOBNEWVAR(T, S) ((T *) gcse_alloc ((S)))
546 /* Misc. utilities. */
549 (this_target_gcse->x_can_copy)
550 #define can_copy_init_p \
551 (this_target_gcse->x_can_copy_init_p)
553 /* Compute which modes support reg/reg copy operations. */
556 compute_can_copy (void)
559 #ifndef AVOID_CCMODE_COPIES
562 memset (can_copy, 0, NUM_MACHINE_MODES);
565 for (i = 0; i < NUM_MACHINE_MODES; i++)
566 if (GET_MODE_CLASS (i) == MODE_CC)
568 #ifdef AVOID_CCMODE_COPIES
571 reg = gen_rtx_REG ((enum machine_mode) i, LAST_VIRTUAL_REGISTER + 1);
572 insn = emit_insn (gen_rtx_SET (VOIDmode, reg, reg));
573 if (recog (PATTERN (insn), insn, NULL) >= 0)
583 /* Returns whether the mode supports reg/reg copy operations. */
586 can_copy_p (enum machine_mode mode)
588 if (! can_copy_init_p)
591 can_copy_init_p = true;
594 return can_copy[mode] != 0;
598 /* Cover function to xmalloc to record bytes allocated. */
601 gmalloc (size_t size)
604 return xmalloc (size);
607 /* Cover function to xcalloc to record bytes allocated. */
610 gcalloc (size_t nelem, size_t elsize)
612 bytes_used += nelem * elsize;
613 return xcalloc (nelem, elsize);
616 /* Cover function to obstack_alloc. */
619 gcse_alloc (unsigned long size)
622 return obstack_alloc (&gcse_obstack, size);
625 /* Allocate memory for the reg/memory set tracking tables.
626 This is called at the start of each pass. */
629 alloc_gcse_mem (void)
631 /* Allocate vars to track sets of regs. */
632 reg_set_bitmap = ALLOC_REG_SET (NULL);
634 /* Allocate array to keep a list of insns which modify memory in each
636 modify_mem_list = GCNEWVEC (rtx, last_basic_block);
637 canon_modify_mem_list = GCNEWVEC (rtx, last_basic_block);
638 modify_mem_list_set = BITMAP_ALLOC (NULL);
639 blocks_with_calls = BITMAP_ALLOC (NULL);
642 /* Free memory allocated by alloc_gcse_mem. */
647 free_modify_mem_tables ();
648 BITMAP_FREE (modify_mem_list_set);
649 BITMAP_FREE (blocks_with_calls);
652 /* Compute the local properties of each recorded expression.
654 Local properties are those that are defined by the block, irrespective of
657 An expression is transparent in a block if its operands are not modified
660 An expression is computed (locally available) in a block if it is computed
661 at least once and expression would contain the same value if the
662 computation was moved to the end of the block.
664 An expression is locally anticipatable in a block if it is computed at
665 least once and expression would contain the same value if the computation
666 was moved to the beginning of the block.
668 We call this routine for cprop, pre and code hoisting. They all compute
669 basically the same information and thus can easily share this code.
671 TRANSP, COMP, and ANTLOC are destination sbitmaps for recording local
672 properties. If NULL, then it is not necessary to compute or record that
675 TABLE controls which hash table to look at. If it is set hash table,
676 additionally, TRANSP is computed as ~TRANSP, since this is really cprop's
680 compute_local_properties (sbitmap *transp, sbitmap *comp, sbitmap *antloc,
681 struct hash_table_d *table)
685 /* Initialize any bitmaps that were passed in. */
689 sbitmap_vector_zero (transp, last_basic_block);
691 sbitmap_vector_ones (transp, last_basic_block);
695 sbitmap_vector_zero (comp, last_basic_block);
697 sbitmap_vector_zero (antloc, last_basic_block);
699 for (i = 0; i < table->size; i++)
703 for (expr = table->table[i]; expr != NULL; expr = expr->next_same_hash)
705 int indx = expr->bitmap_index;
708 /* The expression is transparent in this block if it is not killed.
709 We start by assuming all are transparent [none are killed], and
710 then reset the bits for those that are. */
712 compute_transp (expr->expr, indx, transp, table->set_p);
714 /* The occurrences recorded in antic_occr are exactly those that
715 we want to set to nonzero in ANTLOC. */
717 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
719 SET_BIT (antloc[BLOCK_FOR_INSN (occr->insn)->index], indx);
721 /* While we're scanning the table, this is a good place to
726 /* The occurrences recorded in avail_occr are exactly those that
727 we want to set to nonzero in COMP. */
729 for (occr = expr->avail_occr; occr != NULL; occr = occr->next)
731 SET_BIT (comp[BLOCK_FOR_INSN (occr->insn)->index], indx);
733 /* While we're scanning the table, this is a good place to
738 /* While we're scanning the table, this is a good place to
740 expr->reaching_reg = 0;
745 /* Hash table support. */
747 struct reg_avail_info
754 static struct reg_avail_info *reg_avail_info;
755 static basic_block current_bb;
758 /* See whether X, the source of a set, is something we want to consider for
762 want_to_gcse_p (rtx x)
765 /* On register stack architectures, don't GCSE constants from the
766 constant pool, as the benefits are often swamped by the overhead
767 of shuffling the register stack between basic blocks. */
768 if (IS_STACK_MODE (GET_MODE (x)))
769 x = avoid_constant_pool_reference (x);
772 switch (GET_CODE (x))
784 return can_assign_to_reg_without_clobbers_p (x);
788 /* Used internally by can_assign_to_reg_without_clobbers_p. */
790 static GTY(()) rtx test_insn;
792 /* Return true if we can assign X to a pseudo register such that the
793 resulting insn does not result in clobbering a hard register as a
796 Additionally, if the target requires it, check that the resulting insn
797 can be copied. If it cannot, this means that X is special and probably
798 has hidden side-effects we don't want to mess with.
800 This function is typically used by code motion passes, to verify
801 that it is safe to insert an insn without worrying about clobbering
802 maybe live hard regs. */
805 can_assign_to_reg_without_clobbers_p (rtx x)
807 int num_clobbers = 0;
810 /* If this is a valid operand, we are OK. If it's VOIDmode, we aren't. */
811 if (general_operand (x, GET_MODE (x)))
813 else if (GET_MODE (x) == VOIDmode)
816 /* Otherwise, check if we can make a valid insn from it. First initialize
817 our test insn if we haven't already. */
821 = make_insn_raw (gen_rtx_SET (VOIDmode,
822 gen_rtx_REG (word_mode,
823 FIRST_PSEUDO_REGISTER * 2),
825 NEXT_INSN (test_insn) = PREV_INSN (test_insn) = 0;
828 /* Now make an insn like the one we would make when GCSE'ing and see if
830 PUT_MODE (SET_DEST (PATTERN (test_insn)), GET_MODE (x));
831 SET_SRC (PATTERN (test_insn)) = x;
833 icode = recog (PATTERN (test_insn), test_insn, &num_clobbers);
837 if (num_clobbers > 0 && added_clobbers_hard_reg_p (icode))
840 if (targetm.cannot_copy_insn_p && targetm.cannot_copy_insn_p (test_insn))
846 /* Return nonzero if the operands of expression X are unchanged from the
847 start of INSN's basic block up to but not including INSN (if AVAIL_P == 0),
848 or from INSN to the end of INSN's basic block (if AVAIL_P != 0). */
851 oprs_unchanged_p (const_rtx x, const_rtx insn, int avail_p)
865 struct reg_avail_info *info = ®_avail_info[REGNO (x)];
867 if (info->last_bb != current_bb)
870 return info->last_set < DF_INSN_LUID (insn);
872 return info->first_set >= DF_INSN_LUID (insn);
876 if (load_killed_in_block_p (current_bb, DF_INSN_LUID (insn),
880 return oprs_unchanged_p (XEXP (x, 0), insn, avail_p);
907 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
911 /* If we are about to do the last recursive call needed at this
912 level, change it into iteration. This function is called enough
915 return oprs_unchanged_p (XEXP (x, i), insn, avail_p);
917 else if (! oprs_unchanged_p (XEXP (x, i), insn, avail_p))
920 else if (fmt[i] == 'E')
921 for (j = 0; j < XVECLEN (x, i); j++)
922 if (! oprs_unchanged_p (XVECEXP (x, i, j), insn, avail_p))
929 /* Used for communication between mems_conflict_for_gcse_p and
930 load_killed_in_block_p. Nonzero if mems_conflict_for_gcse_p finds a
931 conflict between two memory references. */
932 static int gcse_mems_conflict_p;
934 /* Used for communication between mems_conflict_for_gcse_p and
935 load_killed_in_block_p. A memory reference for a load instruction,
936 mems_conflict_for_gcse_p will see if a memory store conflicts with
938 static const_rtx gcse_mem_operand;
940 /* DEST is the output of an instruction. If it is a memory reference, and
941 possibly conflicts with the load found in gcse_mem_operand, then set
942 gcse_mems_conflict_p to a nonzero value. */
945 mems_conflict_for_gcse_p (rtx dest, const_rtx setter ATTRIBUTE_UNUSED,
946 void *data ATTRIBUTE_UNUSED)
948 while (GET_CODE (dest) == SUBREG
949 || GET_CODE (dest) == ZERO_EXTRACT
950 || GET_CODE (dest) == STRICT_LOW_PART)
951 dest = XEXP (dest, 0);
953 /* If DEST is not a MEM, then it will not conflict with the load. Note
954 that function calls are assumed to clobber memory, but are handled
959 /* If we are setting a MEM in our list of specially recognized MEMs,
960 don't mark as killed this time. */
962 if (expr_equiv_p (dest, gcse_mem_operand) && pre_ldst_mems != NULL)
964 if (!find_rtx_in_ldst (dest))
965 gcse_mems_conflict_p = 1;
969 if (true_dependence (dest, GET_MODE (dest), gcse_mem_operand,
971 gcse_mems_conflict_p = 1;
974 /* Return nonzero if the expression in X (a memory reference) is killed
975 in block BB before or after the insn with the LUID in UID_LIMIT.
976 AVAIL_P is nonzero for kills after UID_LIMIT, and zero for kills
979 To check the entire block, set UID_LIMIT to max_uid + 1 and
983 load_killed_in_block_p (const_basic_block bb, int uid_limit, const_rtx x, int avail_p)
985 rtx list_entry = modify_mem_list[bb->index];
987 /* If this is a readonly then we aren't going to be changing it. */
988 if (MEM_READONLY_P (x))
994 /* Ignore entries in the list that do not apply. */
996 && DF_INSN_LUID (XEXP (list_entry, 0)) < uid_limit)
998 && DF_INSN_LUID (XEXP (list_entry, 0)) > uid_limit))
1000 list_entry = XEXP (list_entry, 1);
1004 setter = XEXP (list_entry, 0);
1006 /* If SETTER is a call everything is clobbered. Note that calls
1007 to pure functions are never put on the list, so we need not
1008 worry about them. */
1009 if (CALL_P (setter))
1012 /* SETTER must be an INSN of some kind that sets memory. Call
1013 note_stores to examine each hunk of memory that is modified.
1015 The note_stores interface is pretty limited, so we have to
1016 communicate via global variables. Yuk. */
1017 gcse_mem_operand = x;
1018 gcse_mems_conflict_p = 0;
1019 note_stores (PATTERN (setter), mems_conflict_for_gcse_p, NULL);
1020 if (gcse_mems_conflict_p)
1022 list_entry = XEXP (list_entry, 1);
1027 /* Return nonzero if the operands of expression X are unchanged from
1028 the start of INSN's basic block up to but not including INSN. */
1031 oprs_anticipatable_p (const_rtx x, const_rtx insn)
1033 return oprs_unchanged_p (x, insn, 0);
1036 /* Return nonzero if the operands of expression X are unchanged from
1037 INSN to the end of INSN's basic block. */
1040 oprs_available_p (const_rtx x, const_rtx insn)
1042 return oprs_unchanged_p (x, insn, 1);
1045 /* Hash expression X.
1047 MODE is only used if X is a CONST_INT. DO_NOT_RECORD_P is a boolean
1048 indicating if a volatile operand is found or if the expression contains
1049 something we don't want to insert in the table. HASH_TABLE_SIZE is
1050 the current size of the hash table to be probed. */
1053 hash_expr (const_rtx x, enum machine_mode mode, int *do_not_record_p,
1054 int hash_table_size)
1058 *do_not_record_p = 0;
1060 hash = hash_rtx (x, mode, do_not_record_p,
1061 NULL, /*have_reg_qty=*/false);
1062 return hash % hash_table_size;
1065 /* Hash a set of register REGNO.
1067 Sets are hashed on the register that is set. This simplifies the PRE copy
1070 ??? May need to make things more elaborate. Later, as necessary. */
1073 hash_set (int regno, int hash_table_size)
1078 return hash % hash_table_size;
1081 /* Return nonzero if exp1 is equivalent to exp2. */
1084 expr_equiv_p (const_rtx x, const_rtx y)
1086 return exp_equiv_p (x, y, 0, true);
1089 /* Insert expression X in INSN in the hash TABLE.
1090 If it is already present, record it as the last occurrence in INSN's
1093 MODE is the mode of the value X is being stored into.
1094 It is only used if X is a CONST_INT.
1096 ANTIC_P is nonzero if X is an anticipatable expression.
1097 AVAIL_P is nonzero if X is an available expression. */
1100 insert_expr_in_table (rtx x, enum machine_mode mode, rtx insn, int antic_p,
1101 int avail_p, struct hash_table_d *table)
1103 int found, do_not_record_p;
1105 struct expr *cur_expr, *last_expr = NULL;
1106 struct occr *antic_occr, *avail_occr;
1108 hash = hash_expr (x, mode, &do_not_record_p, table->size);
1110 /* Do not insert expression in table if it contains volatile operands,
1111 or if hash_expr determines the expression is something we don't want
1112 to or can't handle. */
1113 if (do_not_record_p)
1116 cur_expr = table->table[hash];
1119 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1121 /* If the expression isn't found, save a pointer to the end of
1123 last_expr = cur_expr;
1124 cur_expr = cur_expr->next_same_hash;
1129 cur_expr = GOBNEW (struct expr);
1130 bytes_used += sizeof (struct expr);
1131 if (table->table[hash] == NULL)
1132 /* This is the first pattern that hashed to this index. */
1133 table->table[hash] = cur_expr;
1135 /* Add EXPR to end of this hash chain. */
1136 last_expr->next_same_hash = cur_expr;
1138 /* Set the fields of the expr element. */
1140 cur_expr->bitmap_index = table->n_elems++;
1141 cur_expr->next_same_hash = NULL;
1142 cur_expr->antic_occr = NULL;
1143 cur_expr->avail_occr = NULL;
1146 /* Now record the occurrence(s). */
1149 antic_occr = cur_expr->antic_occr;
1152 && BLOCK_FOR_INSN (antic_occr->insn) != BLOCK_FOR_INSN (insn))
1156 /* Found another instance of the expression in the same basic block.
1157 Prefer the currently recorded one. We want the first one in the
1158 block and the block is scanned from start to end. */
1159 ; /* nothing to do */
1162 /* First occurrence of this expression in this basic block. */
1163 antic_occr = GOBNEW (struct occr);
1164 bytes_used += sizeof (struct occr);
1165 antic_occr->insn = insn;
1166 antic_occr->next = cur_expr->antic_occr;
1167 antic_occr->deleted_p = 0;
1168 cur_expr->antic_occr = antic_occr;
1174 avail_occr = cur_expr->avail_occr;
1177 && BLOCK_FOR_INSN (avail_occr->insn) == BLOCK_FOR_INSN (insn))
1179 /* Found another instance of the expression in the same basic block.
1180 Prefer this occurrence to the currently recorded one. We want
1181 the last one in the block and the block is scanned from start
1183 avail_occr->insn = insn;
1187 /* First occurrence of this expression in this basic block. */
1188 avail_occr = GOBNEW (struct occr);
1189 bytes_used += sizeof (struct occr);
1190 avail_occr->insn = insn;
1191 avail_occr->next = cur_expr->avail_occr;
1192 avail_occr->deleted_p = 0;
1193 cur_expr->avail_occr = avail_occr;
1198 /* Insert pattern X in INSN in the hash table.
1199 X is a SET of a reg to either another reg or a constant.
1200 If it is already present, record it as the last occurrence in INSN's
1204 insert_set_in_table (rtx x, rtx insn, struct hash_table_d *table)
1208 struct expr *cur_expr, *last_expr = NULL;
1209 struct occr *cur_occr;
1211 gcc_assert (GET_CODE (x) == SET && REG_P (SET_DEST (x)));
1213 hash = hash_set (REGNO (SET_DEST (x)), table->size);
1215 cur_expr = table->table[hash];
1218 while (cur_expr && 0 == (found = expr_equiv_p (cur_expr->expr, x)))
1220 /* If the expression isn't found, save a pointer to the end of
1222 last_expr = cur_expr;
1223 cur_expr = cur_expr->next_same_hash;
1228 cur_expr = GOBNEW (struct expr);
1229 bytes_used += sizeof (struct expr);
1230 if (table->table[hash] == NULL)
1231 /* This is the first pattern that hashed to this index. */
1232 table->table[hash] = cur_expr;
1234 /* Add EXPR to end of this hash chain. */
1235 last_expr->next_same_hash = cur_expr;
1237 /* Set the fields of the expr element.
1238 We must copy X because it can be modified when copy propagation is
1239 performed on its operands. */
1240 cur_expr->expr = copy_rtx (x);
1241 cur_expr->bitmap_index = table->n_elems++;
1242 cur_expr->next_same_hash = NULL;
1243 cur_expr->antic_occr = NULL;
1244 cur_expr->avail_occr = NULL;
1247 /* Now record the occurrence. */
1248 cur_occr = cur_expr->avail_occr;
1251 && BLOCK_FOR_INSN (cur_occr->insn) == BLOCK_FOR_INSN (insn))
1253 /* Found another instance of the expression in the same basic block.
1254 Prefer this occurrence to the currently recorded one. We want
1255 the last one in the block and the block is scanned from start
1257 cur_occr->insn = insn;
1261 /* First occurrence of this expression in this basic block. */
1262 cur_occr = GOBNEW (struct occr);
1263 bytes_used += sizeof (struct occr);
1264 cur_occr->insn = insn;
1265 cur_occr->next = cur_expr->avail_occr;
1266 cur_occr->deleted_p = 0;
1267 cur_expr->avail_occr = cur_occr;
1271 /* Determine whether the rtx X should be treated as a constant for
1272 the purposes of GCSE's constant propagation. */
1275 gcse_constant_p (const_rtx x)
1277 /* Consider a COMPARE of two integers constant. */
1278 if (GET_CODE (x) == COMPARE
1279 && CONST_INT_P (XEXP (x, 0))
1280 && CONST_INT_P (XEXP (x, 1)))
1283 /* Consider a COMPARE of the same registers is a constant
1284 if they are not floating point registers. */
1285 if (GET_CODE(x) == COMPARE
1286 && REG_P (XEXP (x, 0)) && REG_P (XEXP (x, 1))
1287 && REGNO (XEXP (x, 0)) == REGNO (XEXP (x, 1))
1288 && ! FLOAT_MODE_P (GET_MODE (XEXP (x, 0)))
1289 && ! FLOAT_MODE_P (GET_MODE (XEXP (x, 1))))
1292 /* Since X might be inserted more than once we have to take care that it
1294 return CONSTANT_P (x) && (GET_CODE (x) != CONST || shared_const_p (x));
1297 /* Scan pattern PAT of INSN and add an entry to the hash TABLE (set or
1301 hash_scan_set (rtx pat, rtx insn, struct hash_table_d *table)
1303 rtx src = SET_SRC (pat);
1304 rtx dest = SET_DEST (pat);
1307 if (GET_CODE (src) == CALL)
1308 hash_scan_call (src, insn, table);
1310 else if (REG_P (dest))
1312 unsigned int regno = REGNO (dest);
1315 /* See if a REG_EQUAL note shows this equivalent to a simpler expression.
1317 This allows us to do a single GCSE pass and still eliminate
1318 redundant constants, addresses or other expressions that are
1319 constructed with multiple instructions.
1321 However, keep the original SRC if INSN is a simple reg-reg move. In
1322 In this case, there will almost always be a REG_EQUAL note on the
1323 insn that sets SRC. By recording the REG_EQUAL value here as SRC
1324 for INSN, we miss copy propagation opportunities and we perform the
1325 same PRE GCSE operation repeatedly on the same REG_EQUAL value if we
1326 do more than one PRE GCSE pass.
1328 Note that this does not impede profitable constant propagations. We
1329 "look through" reg-reg sets in lookup_avail_set. */
1330 note = find_reg_equal_equiv_note (insn);
1332 && REG_NOTE_KIND (note) == REG_EQUAL
1335 ? gcse_constant_p (XEXP (note, 0))
1336 : want_to_gcse_p (XEXP (note, 0))))
1337 src = XEXP (note, 0), pat = gen_rtx_SET (VOIDmode, dest, src);
1339 /* Only record sets of pseudo-regs in the hash table. */
1341 && regno >= FIRST_PSEUDO_REGISTER
1342 /* Don't GCSE something if we can't do a reg/reg copy. */
1343 && can_copy_p (GET_MODE (dest))
1344 /* GCSE commonly inserts instruction after the insn. We can't
1345 do that easily for EH edges so disable GCSE on these for now. */
1346 /* ??? We can now easily create new EH landing pads at the
1347 gimple level, for splitting edges; there's no reason we
1348 can't do the same thing at the rtl level. */
1349 && !can_throw_internal (insn)
1350 /* Is SET_SRC something we want to gcse? */
1351 && want_to_gcse_p (src)
1352 /* Don't CSE a nop. */
1353 && ! set_noop_p (pat)
1354 /* Don't GCSE if it has attached REG_EQUIV note.
1355 At this point this only function parameters should have
1356 REG_EQUIV notes and if the argument slot is used somewhere
1357 explicitly, it means address of parameter has been taken,
1358 so we should not extend the lifetime of the pseudo. */
1359 && (note == NULL_RTX || ! MEM_P (XEXP (note, 0))))
1361 /* An expression is not anticipatable if its operands are
1362 modified before this insn or if this is not the only SET in
1363 this insn. The latter condition does not have to mean that
1364 SRC itself is not anticipatable, but we just will not be
1365 able to handle code motion of insns with multiple sets. */
1366 int antic_p = oprs_anticipatable_p (src, insn)
1367 && !multiple_sets (insn);
1368 /* An expression is not available if its operands are
1369 subsequently modified, including this insn. It's also not
1370 available if this is a branch, because we can't insert
1371 a set after the branch. */
1372 int avail_p = (oprs_available_p (src, insn)
1373 && ! JUMP_P (insn));
1375 insert_expr_in_table (src, GET_MODE (dest), insn, antic_p, avail_p, table);
1378 /* Record sets for constant/copy propagation. */
1379 else if (table->set_p
1380 && regno >= FIRST_PSEUDO_REGISTER
1382 && REGNO (src) >= FIRST_PSEUDO_REGISTER
1383 && can_copy_p (GET_MODE (dest))
1384 && REGNO (src) != regno)
1385 || gcse_constant_p (src))
1386 /* A copy is not available if its src or dest is subsequently
1387 modified. Here we want to search from INSN+1 on, but
1388 oprs_available_p searches from INSN on. */
1389 && (insn == BB_END (BLOCK_FOR_INSN (insn))
1390 || (tmp = next_nonnote_insn (insn)) == NULL_RTX
1391 || BLOCK_FOR_INSN (tmp) != BLOCK_FOR_INSN (insn)
1392 || oprs_available_p (pat, tmp)))
1393 insert_set_in_table (pat, insn, table);
1395 /* In case of store we want to consider the memory value as available in
1396 the REG stored in that memory. This makes it possible to remove
1397 redundant loads from due to stores to the same location. */
1398 else if (flag_gcse_las && REG_P (src) && MEM_P (dest))
1400 unsigned int regno = REGNO (src);
1402 /* Do not do this for constant/copy propagation. */
1404 /* Only record sets of pseudo-regs in the hash table. */
1405 && regno >= FIRST_PSEUDO_REGISTER
1406 /* Don't GCSE something if we can't do a reg/reg copy. */
1407 && can_copy_p (GET_MODE (src))
1408 /* GCSE commonly inserts instruction after the insn. We can't
1409 do that easily for EH edges so disable GCSE on these for now. */
1410 && !can_throw_internal (insn)
1411 /* Is SET_DEST something we want to gcse? */
1412 && want_to_gcse_p (dest)
1413 /* Don't CSE a nop. */
1414 && ! set_noop_p (pat)
1415 /* Don't GCSE if it has attached REG_EQUIV note.
1416 At this point this only function parameters should have
1417 REG_EQUIV notes and if the argument slot is used somewhere
1418 explicitly, it means address of parameter has been taken,
1419 so we should not extend the lifetime of the pseudo. */
1420 && ((note = find_reg_note (insn, REG_EQUIV, NULL_RTX)) == 0
1421 || ! MEM_P (XEXP (note, 0))))
1423 /* Stores are never anticipatable. */
1425 /* An expression is not available if its operands are
1426 subsequently modified, including this insn. It's also not
1427 available if this is a branch, because we can't insert
1428 a set after the branch. */
1429 int avail_p = oprs_available_p (dest, insn)
1432 /* Record the memory expression (DEST) in the hash table. */
1433 insert_expr_in_table (dest, GET_MODE (dest), insn,
1434 antic_p, avail_p, table);
1440 hash_scan_clobber (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
1441 struct hash_table_d *table ATTRIBUTE_UNUSED)
1443 /* Currently nothing to do. */
1447 hash_scan_call (rtx x ATTRIBUTE_UNUSED, rtx insn ATTRIBUTE_UNUSED,
1448 struct hash_table_d *table ATTRIBUTE_UNUSED)
1450 /* Currently nothing to do. */
1453 /* Process INSN and add hash table entries as appropriate.
1455 Only available expressions that set a single pseudo-reg are recorded.
1457 Single sets in a PARALLEL could be handled, but it's an extra complication
1458 that isn't dealt with right now. The trick is handling the CLOBBERs that
1459 are also in the PARALLEL. Later.
1461 If SET_P is nonzero, this is for the assignment hash table,
1462 otherwise it is for the expression hash table. */
1465 hash_scan_insn (rtx insn, struct hash_table_d *table)
1467 rtx pat = PATTERN (insn);
1470 /* Pick out the sets of INSN and for other forms of instructions record
1471 what's been modified. */
1473 if (GET_CODE (pat) == SET)
1474 hash_scan_set (pat, insn, table);
1475 else if (GET_CODE (pat) == PARALLEL)
1476 for (i = 0; i < XVECLEN (pat, 0); i++)
1478 rtx x = XVECEXP (pat, 0, i);
1480 if (GET_CODE (x) == SET)
1481 hash_scan_set (x, insn, table);
1482 else if (GET_CODE (x) == CLOBBER)
1483 hash_scan_clobber (x, insn, table);
1484 else if (GET_CODE (x) == CALL)
1485 hash_scan_call (x, insn, table);
1488 else if (GET_CODE (pat) == CLOBBER)
1489 hash_scan_clobber (pat, insn, table);
1490 else if (GET_CODE (pat) == CALL)
1491 hash_scan_call (pat, insn, table);
1495 dump_hash_table (FILE *file, const char *name, struct hash_table_d *table)
1498 /* Flattened out table, so it's printed in proper order. */
1499 struct expr **flat_table;
1500 unsigned int *hash_val;
1503 flat_table = XCNEWVEC (struct expr *, table->n_elems);
1504 hash_val = XNEWVEC (unsigned int, table->n_elems);
1506 for (i = 0; i < (int) table->size; i++)
1507 for (expr = table->table[i]; expr != NULL; expr = expr->next_same_hash)
1509 flat_table[expr->bitmap_index] = expr;
1510 hash_val[expr->bitmap_index] = i;
1513 fprintf (file, "%s hash table (%d buckets, %d entries)\n",
1514 name, table->size, table->n_elems);
1516 for (i = 0; i < (int) table->n_elems; i++)
1517 if (flat_table[i] != 0)
1519 expr = flat_table[i];
1520 fprintf (file, "Index %d (hash value %d)\n ",
1521 expr->bitmap_index, hash_val[i]);
1522 print_rtl (file, expr->expr);
1523 fprintf (file, "\n");
1526 fprintf (file, "\n");
1532 /* Record register first/last/block set information for REGNO in INSN.
1534 first_set records the first place in the block where the register
1535 is set and is used to compute "anticipatability".
1537 last_set records the last place in the block where the register
1538 is set and is used to compute "availability".
1540 last_bb records the block for which first_set and last_set are
1541 valid, as a quick test to invalidate them. */
1544 record_last_reg_set_info (rtx insn, int regno)
1546 struct reg_avail_info *info = ®_avail_info[regno];
1547 int luid = DF_INSN_LUID (insn);
1549 info->last_set = luid;
1550 if (info->last_bb != current_bb)
1552 info->last_bb = current_bb;
1553 info->first_set = luid;
1558 /* Record all of the canonicalized MEMs of record_last_mem_set_info's insn.
1559 Note we store a pair of elements in the list, so they have to be
1560 taken off pairwise. */
1563 canon_list_insert (rtx dest ATTRIBUTE_UNUSED, const_rtx unused1 ATTRIBUTE_UNUSED,
1566 rtx dest_addr, insn;
1569 while (GET_CODE (dest) == SUBREG
1570 || GET_CODE (dest) == ZERO_EXTRACT
1571 || GET_CODE (dest) == STRICT_LOW_PART)
1572 dest = XEXP (dest, 0);
1574 /* If DEST is not a MEM, then it will not conflict with a load. Note
1575 that function calls are assumed to clobber memory, but are handled
1581 dest_addr = get_addr (XEXP (dest, 0));
1582 dest_addr = canon_rtx (dest_addr);
1583 insn = (rtx) v_insn;
1584 bb = BLOCK_FOR_INSN (insn)->index;
1586 canon_modify_mem_list[bb] =
1587 alloc_EXPR_LIST (VOIDmode, dest_addr, canon_modify_mem_list[bb]);
1588 canon_modify_mem_list[bb] =
1589 alloc_EXPR_LIST (VOIDmode, dest, canon_modify_mem_list[bb]);
1592 /* Record memory modification information for INSN. We do not actually care
1593 about the memory location(s) that are set, or even how they are set (consider
1594 a CALL_INSN). We merely need to record which insns modify memory. */
1597 record_last_mem_set_info (rtx insn)
1599 int bb = BLOCK_FOR_INSN (insn)->index;
1601 /* load_killed_in_block_p will handle the case of calls clobbering
1603 modify_mem_list[bb] = alloc_INSN_LIST (insn, modify_mem_list[bb]);
1604 bitmap_set_bit (modify_mem_list_set, bb);
1608 /* Note that traversals of this loop (other than for free-ing)
1609 will break after encountering a CALL_INSN. So, there's no
1610 need to insert a pair of items, as canon_list_insert does. */
1611 canon_modify_mem_list[bb] =
1612 alloc_INSN_LIST (insn, canon_modify_mem_list[bb]);
1613 bitmap_set_bit (blocks_with_calls, bb);
1616 note_stores (PATTERN (insn), canon_list_insert, (void*) insn);
1619 /* Called from compute_hash_table via note_stores to handle one
1620 SET or CLOBBER in an insn. DATA is really the instruction in which
1621 the SET is taking place. */
1624 record_last_set_info (rtx dest, const_rtx setter ATTRIBUTE_UNUSED, void *data)
1626 rtx last_set_insn = (rtx) data;
1628 if (GET_CODE (dest) == SUBREG)
1629 dest = SUBREG_REG (dest);
1632 record_last_reg_set_info (last_set_insn, REGNO (dest));
1633 else if (MEM_P (dest)
1634 /* Ignore pushes, they clobber nothing. */
1635 && ! push_operand (dest, GET_MODE (dest)))
1636 record_last_mem_set_info (last_set_insn);
1639 /* Top level function to create an expression or assignment hash table.
1641 Expression entries are placed in the hash table if
1642 - they are of the form (set (pseudo-reg) src),
1643 - src is something we want to perform GCSE on,
1644 - none of the operands are subsequently modified in the block
1646 Assignment entries are placed in the hash table if
1647 - they are of the form (set (pseudo-reg) src),
1648 - src is something we want to perform const/copy propagation on,
1649 - none of the operands or target are subsequently modified in the block
1651 Currently src must be a pseudo-reg or a const_int.
1653 TABLE is the table computed. */
1656 compute_hash_table_work (struct hash_table_d *table)
1660 /* re-Cache any INSN_LIST nodes we have allocated. */
1661 clear_modify_mem_tables ();
1662 /* Some working arrays used to track first and last set in each block. */
1663 reg_avail_info = GNEWVEC (struct reg_avail_info, max_reg_num ());
1665 for (i = 0; i < max_reg_num (); ++i)
1666 reg_avail_info[i].last_bb = NULL;
1668 FOR_EACH_BB (current_bb)
1673 /* First pass over the instructions records information used to
1674 determine when registers and memory are first and last set. */
1675 FOR_BB_INSNS (current_bb, insn)
1677 if (! INSN_P (insn))
1682 for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
1683 if (TEST_HARD_REG_BIT (regs_invalidated_by_call, regno))
1684 record_last_reg_set_info (insn, regno);
1689 note_stores (PATTERN (insn), record_last_set_info, insn);
1692 /* Insert implicit sets in the hash table. */
1694 && implicit_sets[current_bb->index] != NULL_RTX)
1695 hash_scan_set (implicit_sets[current_bb->index],
1696 BB_HEAD (current_bb), table);
1698 /* The next pass builds the hash table. */
1699 FOR_BB_INSNS (current_bb, insn)
1701 hash_scan_insn (insn, table);
1704 free (reg_avail_info);
1705 reg_avail_info = NULL;
1708 /* Allocate space for the set/expr hash TABLE.
1709 It is used to determine the number of buckets to use.
1710 SET_P determines whether set or expression table will
1714 alloc_hash_table (struct hash_table_d *table, int set_p)
1718 n = get_max_insn_count ();
1720 table->size = n / 4;
1721 if (table->size < 11)
1724 /* Attempt to maintain efficient use of hash table.
1725 Making it an odd number is simplest for now.
1726 ??? Later take some measurements. */
1728 n = table->size * sizeof (struct expr *);
1729 table->table = GNEWVAR (struct expr *, n);
1730 table->set_p = set_p;
1733 /* Free things allocated by alloc_hash_table. */
1736 free_hash_table (struct hash_table_d *table)
1738 free (table->table);
1741 /* Compute the hash TABLE for doing copy/const propagation or
1742 expression hash table. */
1745 compute_hash_table (struct hash_table_d *table)
1747 /* Initialize count of number of entries in hash table. */
1749 memset (table->table, 0, table->size * sizeof (struct expr *));
1751 compute_hash_table_work (table);
1754 /* Expression tracking support. */
1756 /* Lookup REGNO in the set TABLE. The result is a pointer to the
1757 table entry, or NULL if not found. */
1759 static struct expr *
1760 lookup_set (unsigned int regno, struct hash_table_d *table)
1762 unsigned int hash = hash_set (regno, table->size);
1765 expr = table->table[hash];
1767 while (expr && REGNO (SET_DEST (expr->expr)) != regno)
1768 expr = expr->next_same_hash;
1773 /* Return the next entry for REGNO in list EXPR. */
1775 static struct expr *
1776 next_set (unsigned int regno, struct expr *expr)
1779 expr = expr->next_same_hash;
1780 while (expr && REGNO (SET_DEST (expr->expr)) != regno);
1785 /* Like free_INSN_LIST_list or free_EXPR_LIST_list, except that the node
1786 types may be mixed. */
1789 free_insn_expr_list_list (rtx *listp)
1793 for (list = *listp; list ; list = next)
1795 next = XEXP (list, 1);
1796 if (GET_CODE (list) == EXPR_LIST)
1797 free_EXPR_LIST_node (list);
1799 free_INSN_LIST_node (list);
1805 /* Clear canon_modify_mem_list and modify_mem_list tables. */
1807 clear_modify_mem_tables (void)
1812 EXECUTE_IF_SET_IN_BITMAP (modify_mem_list_set, 0, i, bi)
1814 free_INSN_LIST_list (modify_mem_list + i);
1815 free_insn_expr_list_list (canon_modify_mem_list + i);
1817 bitmap_clear (modify_mem_list_set);
1818 bitmap_clear (blocks_with_calls);
1821 /* Release memory used by modify_mem_list_set. */
1824 free_modify_mem_tables (void)
1826 clear_modify_mem_tables ();
1827 free (modify_mem_list);
1828 free (canon_modify_mem_list);
1829 modify_mem_list = 0;
1830 canon_modify_mem_list = 0;
1833 /* Reset tables used to keep track of what's still available [since the
1834 start of the block]. */
1837 reset_opr_set_tables (void)
1839 /* Maintain a bitmap of which regs have been set since beginning of
1841 CLEAR_REG_SET (reg_set_bitmap);
1843 /* Also keep a record of the last instruction to modify memory.
1844 For now this is very trivial, we only record whether any memory
1845 location has been modified. */
1846 clear_modify_mem_tables ();
1849 /* Return nonzero if the operands of X are not set before INSN in
1850 INSN's basic block. */
1853 oprs_not_set_p (const_rtx x, const_rtx insn)
1862 code = GET_CODE (x);
1879 if (load_killed_in_block_p (BLOCK_FOR_INSN (insn),
1880 DF_INSN_LUID (insn), x, 0))
1883 return oprs_not_set_p (XEXP (x, 0), insn);
1886 return ! REGNO_REG_SET_P (reg_set_bitmap, REGNO (x));
1892 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
1896 /* If we are about to do the last recursive call
1897 needed at this level, change it into iteration.
1898 This function is called enough to be worth it. */
1900 return oprs_not_set_p (XEXP (x, i), insn);
1902 if (! oprs_not_set_p (XEXP (x, i), insn))
1905 else if (fmt[i] == 'E')
1906 for (j = 0; j < XVECLEN (x, i); j++)
1907 if (! oprs_not_set_p (XVECEXP (x, i, j), insn))
1914 /* Mark things set by a CALL. */
1917 mark_call (rtx insn)
1919 if (! RTL_CONST_OR_PURE_CALL_P (insn))
1920 record_last_mem_set_info (insn);
1923 /* Mark things set by a SET. */
1926 mark_set (rtx pat, rtx insn)
1928 rtx dest = SET_DEST (pat);
1930 while (GET_CODE (dest) == SUBREG
1931 || GET_CODE (dest) == ZERO_EXTRACT
1932 || GET_CODE (dest) == STRICT_LOW_PART)
1933 dest = XEXP (dest, 0);
1936 SET_REGNO_REG_SET (reg_set_bitmap, REGNO (dest));
1937 else if (MEM_P (dest))
1938 record_last_mem_set_info (insn);
1940 if (GET_CODE (SET_SRC (pat)) == CALL)
1944 /* Record things set by a CLOBBER. */
1947 mark_clobber (rtx pat, rtx insn)
1949 rtx clob = XEXP (pat, 0);
1951 while (GET_CODE (clob) == SUBREG || GET_CODE (clob) == STRICT_LOW_PART)
1952 clob = XEXP (clob, 0);
1955 SET_REGNO_REG_SET (reg_set_bitmap, REGNO (clob));
1957 record_last_mem_set_info (insn);
1960 /* Record things set by INSN.
1961 This data is used by oprs_not_set_p. */
1964 mark_oprs_set (rtx insn)
1966 rtx pat = PATTERN (insn);
1969 if (GET_CODE (pat) == SET)
1970 mark_set (pat, insn);
1971 else if (GET_CODE (pat) == PARALLEL)
1972 for (i = 0; i < XVECLEN (pat, 0); i++)
1974 rtx x = XVECEXP (pat, 0, i);
1976 if (GET_CODE (x) == SET)
1978 else if (GET_CODE (x) == CLOBBER)
1979 mark_clobber (x, insn);
1980 else if (GET_CODE (x) == CALL)
1984 else if (GET_CODE (pat) == CLOBBER)
1985 mark_clobber (pat, insn);
1986 else if (GET_CODE (pat) == CALL)
1991 /* Compute copy/constant propagation working variables. */
1993 /* Local properties of assignments. */
1994 static sbitmap *cprop_pavloc;
1995 static sbitmap *cprop_absaltered;
1997 /* Global properties of assignments (computed from the local properties). */
1998 static sbitmap *cprop_avin;
1999 static sbitmap *cprop_avout;
2001 /* Allocate vars used for copy/const propagation. N_BLOCKS is the number of
2002 basic blocks. N_SETS is the number of sets. */
2005 alloc_cprop_mem (int n_blocks, int n_sets)
2007 cprop_pavloc = sbitmap_vector_alloc (n_blocks, n_sets);
2008 cprop_absaltered = sbitmap_vector_alloc (n_blocks, n_sets);
2010 cprop_avin = sbitmap_vector_alloc (n_blocks, n_sets);
2011 cprop_avout = sbitmap_vector_alloc (n_blocks, n_sets);
2014 /* Free vars used by copy/const propagation. */
2017 free_cprop_mem (void)
2019 sbitmap_vector_free (cprop_pavloc);
2020 sbitmap_vector_free (cprop_absaltered);
2021 sbitmap_vector_free (cprop_avin);
2022 sbitmap_vector_free (cprop_avout);
2025 /* For each block, compute whether X is transparent. X is either an
2026 expression or an assignment [though we don't care which, for this context
2027 an assignment is treated as an expression]. For each block where an
2028 element of X is modified, set (SET_P == 1) or reset (SET_P == 0) the INDX
2032 compute_transp (const_rtx x, int indx, sbitmap *bmap, int set_p)
2038 /* repeat is used to turn tail-recursion into iteration since GCC
2039 can't do it when there's no return value. */
2045 code = GET_CODE (x);
2052 for (def = DF_REG_DEF_CHAIN (REGNO (x));
2054 def = DF_REF_NEXT_REG (def))
2055 SET_BIT (bmap[DF_REF_BB (def)->index], indx);
2060 for (def = DF_REG_DEF_CHAIN (REGNO (x));
2062 def = DF_REF_NEXT_REG (def))
2063 RESET_BIT (bmap[DF_REF_BB (def)->index], indx);
2069 if (! MEM_READONLY_P (x))
2074 /* First handle all the blocks with calls. We don't need to
2075 do any list walking for them. */
2076 EXECUTE_IF_SET_IN_BITMAP (blocks_with_calls, 0, bb_index, bi)
2079 SET_BIT (bmap[bb_index], indx);
2081 RESET_BIT (bmap[bb_index], indx);
2084 /* Now iterate over the blocks which have memory modifications
2085 but which do not have any calls. */
2086 EXECUTE_IF_AND_COMPL_IN_BITMAP (modify_mem_list_set,
2090 rtx list_entry = canon_modify_mem_list[bb_index];
2094 rtx dest, dest_addr;
2096 /* LIST_ENTRY must be an INSN of some kind that sets memory.
2097 Examine each hunk of memory that is modified. */
2099 dest = XEXP (list_entry, 0);
2100 list_entry = XEXP (list_entry, 1);
2101 dest_addr = XEXP (list_entry, 0);
2103 if (canon_true_dependence (dest, GET_MODE (dest), dest_addr,
2104 x, NULL_RTX, rtx_addr_varies_p))
2107 SET_BIT (bmap[bb_index], indx);
2109 RESET_BIT (bmap[bb_index], indx);
2112 list_entry = XEXP (list_entry, 1);
2137 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2141 /* If we are about to do the last recursive call
2142 needed at this level, change it into iteration.
2143 This function is called enough to be worth it. */
2150 compute_transp (XEXP (x, i), indx, bmap, set_p);
2152 else if (fmt[i] == 'E')
2153 for (j = 0; j < XVECLEN (x, i); j++)
2154 compute_transp (XVECEXP (x, i, j), indx, bmap, set_p);
2158 /* Top level routine to do the dataflow analysis needed by copy/const
2162 compute_cprop_data (void)
2164 compute_local_properties (cprop_absaltered, cprop_pavloc, NULL, &set_hash_table);
2165 compute_available (cprop_pavloc, cprop_absaltered,
2166 cprop_avout, cprop_avin);
2169 /* Copy/constant propagation. */
2171 /* Maximum number of register uses in an insn that we handle. */
2174 /* Table of uses found in an insn.
2175 Allocated statically to avoid alloc/free complexity and overhead. */
2176 static struct reg_use reg_use_table[MAX_USES];
2178 /* Index into `reg_use_table' while building it. */
2179 static int reg_use_count;
2181 /* Set up a list of register numbers used in INSN. The found uses are stored
2182 in `reg_use_table'. `reg_use_count' is initialized to zero before entry,
2183 and contains the number of uses in the table upon exit.
2185 ??? If a register appears multiple times we will record it multiple times.
2186 This doesn't hurt anything but it will slow things down. */
2189 find_used_regs (rtx *xptr, void *data ATTRIBUTE_UNUSED)
2196 /* repeat is used to turn tail-recursion into iteration since GCC
2197 can't do it when there's no return value. */
2202 code = GET_CODE (x);
2205 if (reg_use_count == MAX_USES)
2208 reg_use_table[reg_use_count].reg_rtx = x;
2212 /* Recursively scan the operands of this expression. */
2214 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
2218 /* If we are about to do the last recursive call
2219 needed at this level, change it into iteration.
2220 This function is called enough to be worth it. */
2227 find_used_regs (&XEXP (x, i), data);
2229 else if (fmt[i] == 'E')
2230 for (j = 0; j < XVECLEN (x, i); j++)
2231 find_used_regs (&XVECEXP (x, i, j), data);
2235 /* Try to replace all non-SET_DEST occurrences of FROM in INSN with TO.
2236 Returns nonzero is successful. */
2239 try_replace_reg (rtx from, rtx to, rtx insn)
2241 rtx note = find_reg_equal_equiv_note (insn);
2244 rtx set = single_set (insn);
2246 /* Usually we substitute easy stuff, so we won't copy everything.
2247 We however need to take care to not duplicate non-trivial CONST
2251 validate_replace_src_group (from, to, insn);
2252 if (num_changes_pending () && apply_change_group ())
2255 /* Try to simplify SET_SRC if we have substituted a constant. */
2256 if (success && set && CONSTANT_P (to))
2258 src = simplify_rtx (SET_SRC (set));
2261 validate_change (insn, &SET_SRC (set), src, 0);
2264 /* If there is already a REG_EQUAL note, update the expression in it
2265 with our replacement. */
2266 if (note != 0 && REG_NOTE_KIND (note) == REG_EQUAL)
2267 set_unique_reg_note (insn, REG_EQUAL,
2268 simplify_replace_rtx (XEXP (note, 0), from, to));
2269 if (!success && set && reg_mentioned_p (from, SET_SRC (set)))
2271 /* If above failed and this is a single set, try to simplify the source of
2272 the set given our substitution. We could perhaps try this for multiple
2273 SETs, but it probably won't buy us anything. */
2274 src = simplify_replace_rtx (SET_SRC (set), from, to);
2276 if (!rtx_equal_p (src, SET_SRC (set))
2277 && validate_change (insn, &SET_SRC (set), src, 0))
2280 /* If we've failed to do replacement, have a single SET, don't already
2281 have a note, and have no special SET, add a REG_EQUAL note to not
2282 lose information. */
2283 if (!success && note == 0 && set != 0
2284 && GET_CODE (SET_DEST (set)) != ZERO_EXTRACT
2285 && GET_CODE (SET_DEST (set)) != STRICT_LOW_PART)
2286 note = set_unique_reg_note (insn, REG_EQUAL, copy_rtx (src));
2289 /* REG_EQUAL may get simplified into register.
2290 We don't allow that. Remove that note. This code ought
2291 not to happen, because previous code ought to synthesize
2292 reg-reg move, but be on the safe side. */
2293 if (note && REG_NOTE_KIND (note) == REG_EQUAL && REG_P (XEXP (note, 0)))
2294 remove_note (insn, note);
2299 /* Find a set of REGNOs that are available on entry to INSN's block. Returns
2300 NULL no such set is found. */
2302 static struct expr *
2303 find_avail_set (int regno, rtx insn)
2305 /* SET1 contains the last set found that can be returned to the caller for
2306 use in a substitution. */
2307 struct expr *set1 = 0;
2309 /* Loops are not possible here. To get a loop we would need two sets
2310 available at the start of the block containing INSN. i.e. we would
2311 need two sets like this available at the start of the block:
2313 (set (reg X) (reg Y))
2314 (set (reg Y) (reg X))
2316 This can not happen since the set of (reg Y) would have killed the
2317 set of (reg X) making it unavailable at the start of this block. */
2321 struct expr *set = lookup_set (regno, &set_hash_table);
2323 /* Find a set that is available at the start of the block
2324 which contains INSN. */
2327 if (TEST_BIT (cprop_avin[BLOCK_FOR_INSN (insn)->index],
2330 set = next_set (regno, set);
2333 /* If no available set was found we've reached the end of the
2334 (possibly empty) copy chain. */
2338 gcc_assert (GET_CODE (set->expr) == SET);
2340 src = SET_SRC (set->expr);
2342 /* We know the set is available.
2343 Now check that SRC is ANTLOC (i.e. none of the source operands
2344 have changed since the start of the block).
2346 If the source operand changed, we may still use it for the next
2347 iteration of this loop, but we may not use it for substitutions. */
2349 if (gcse_constant_p (src) || oprs_not_set_p (src, insn))
2352 /* If the source of the set is anything except a register, then
2353 we have reached the end of the copy chain. */
2357 /* Follow the copy chain, i.e. start another iteration of the loop
2358 and see if we have an available copy into SRC. */
2359 regno = REGNO (src);
2362 /* SET1 holds the last set that was available and anticipatable at
2367 /* Subroutine of cprop_insn that tries to propagate constants into
2368 JUMP_INSNS. JUMP must be a conditional jump. If SETCC is non-NULL
2369 it is the instruction that immediately precedes JUMP, and must be a
2370 single SET of a register. FROM is what we will try to replace,
2371 SRC is the constant we will try to substitute for it. Returns nonzero
2372 if a change was made. */
2375 cprop_jump (basic_block bb, rtx setcc, rtx jump, rtx from, rtx src)
2377 rtx new_rtx, set_src, note_src;
2378 rtx set = pc_set (jump);
2379 rtx note = find_reg_equal_equiv_note (jump);
2383 note_src = XEXP (note, 0);
2384 if (GET_CODE (note_src) == EXPR_LIST)
2385 note_src = NULL_RTX;
2387 else note_src = NULL_RTX;
2389 /* Prefer REG_EQUAL notes except those containing EXPR_LISTs. */
2390 set_src = note_src ? note_src : SET_SRC (set);
2392 /* First substitute the SETCC condition into the JUMP instruction,
2393 then substitute that given values into this expanded JUMP. */
2394 if (setcc != NULL_RTX
2395 && !modified_between_p (from, setcc, jump)
2396 && !modified_between_p (src, setcc, jump))
2399 rtx setcc_set = single_set (setcc);
2400 rtx setcc_note = find_reg_equal_equiv_note (setcc);
2401 setcc_src = (setcc_note && GET_CODE (XEXP (setcc_note, 0)) != EXPR_LIST)
2402 ? XEXP (setcc_note, 0) : SET_SRC (setcc_set);
2403 set_src = simplify_replace_rtx (set_src, SET_DEST (setcc_set),
2409 new_rtx = simplify_replace_rtx (set_src, from, src);
2411 /* If no simplification can be made, then try the next register. */
2412 if (rtx_equal_p (new_rtx, SET_SRC (set)))
2415 /* If this is now a no-op delete it, otherwise this must be a valid insn. */
2416 if (new_rtx == pc_rtx)
2420 /* Ensure the value computed inside the jump insn to be equivalent
2421 to one computed by setcc. */
2422 if (setcc && modified_in_p (new_rtx, setcc))
2424 if (! validate_unshare_change (jump, &SET_SRC (set), new_rtx, 0))
2426 /* When (some) constants are not valid in a comparison, and there
2427 are two registers to be replaced by constants before the entire
2428 comparison can be folded into a constant, we need to keep
2429 intermediate information in REG_EQUAL notes. For targets with
2430 separate compare insns, such notes are added by try_replace_reg.
2431 When we have a combined compare-and-branch instruction, however,
2432 we need to attach a note to the branch itself to make this
2433 optimization work. */
2435 if (!rtx_equal_p (new_rtx, note_src))
2436 set_unique_reg_note (jump, REG_EQUAL, copy_rtx (new_rtx));
2440 /* Remove REG_EQUAL note after simplification. */
2442 remove_note (jump, note);
2446 /* Delete the cc0 setter. */
2447 if (setcc != NULL && CC0_P (SET_DEST (single_set (setcc))))
2448 delete_insn (setcc);
2451 global_const_prop_count++;
2452 if (dump_file != NULL)
2455 "GLOBAL CONST-PROP: Replacing reg %d in jump_insn %d with constant ",
2456 REGNO (from), INSN_UID (jump));
2457 print_rtl (dump_file, src);
2458 fprintf (dump_file, "\n");
2460 purge_dead_edges (bb);
2462 /* If a conditional jump has been changed into unconditional jump, remove
2463 the jump and make the edge fallthru - this is always called in
2465 if (new_rtx != pc_rtx && simplejump_p (jump))
2470 for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); ei_next (&ei))
2471 if (e->dest != EXIT_BLOCK_PTR
2472 && BB_HEAD (e->dest) == JUMP_LABEL (jump))
2474 e->flags |= EDGE_FALLTHRU;
2484 constprop_register (rtx insn, rtx from, rtx to)
2488 /* Check for reg or cc0 setting instructions followed by
2489 conditional branch instructions first. */
2490 if ((sset = single_set (insn)) != NULL
2492 && any_condjump_p (NEXT_INSN (insn)) && onlyjump_p (NEXT_INSN (insn)))
2494 rtx dest = SET_DEST (sset);
2495 if ((REG_P (dest) || CC0_P (dest))
2496 && cprop_jump (BLOCK_FOR_INSN (insn), insn, NEXT_INSN (insn), from, to))
2500 /* Handle normal insns next. */
2501 if (NONJUMP_INSN_P (insn)
2502 && try_replace_reg (from, to, insn))
2505 /* Try to propagate a CONST_INT into a conditional jump.
2506 We're pretty specific about what we will handle in this
2507 code, we can extend this as necessary over time.
2509 Right now the insn in question must look like
2510 (set (pc) (if_then_else ...)) */
2511 else if (any_condjump_p (insn) && onlyjump_p (insn))
2512 return cprop_jump (BLOCK_FOR_INSN (insn), NULL, insn, from, to);
2516 /* Perform constant and copy propagation on INSN.
2517 The result is nonzero if a change was made. */
2520 cprop_insn (rtx insn)
2522 struct reg_use *reg_used;
2530 note_uses (&PATTERN (insn), find_used_regs, NULL);
2532 note = find_reg_equal_equiv_note (insn);
2534 /* We may win even when propagating constants into notes. */
2536 find_used_regs (&XEXP (note, 0), NULL);
2538 for (reg_used = ®_use_table[0]; reg_use_count > 0;
2539 reg_used++, reg_use_count--)
2541 unsigned int regno = REGNO (reg_used->reg_rtx);
2545 /* If the register has already been set in this block, there's
2546 nothing we can do. */
2547 if (! oprs_not_set_p (reg_used->reg_rtx, insn))
2550 /* Find an assignment that sets reg_used and is available
2551 at the start of the block. */
2552 set = find_avail_set (regno, insn);
2557 /* ??? We might be able to handle PARALLELs. Later. */
2558 gcc_assert (GET_CODE (pat) == SET);
2560 src = SET_SRC (pat);
2562 /* Constant propagation. */
2563 if (gcse_constant_p (src))
2565 if (constprop_register (insn, reg_used->reg_rtx, src))
2568 global_const_prop_count++;
2569 if (dump_file != NULL)
2571 fprintf (dump_file, "GLOBAL CONST-PROP: Replacing reg %d in ", regno);
2572 fprintf (dump_file, "insn %d with constant ", INSN_UID (insn));
2573 print_rtl (dump_file, src);
2574 fprintf (dump_file, "\n");
2576 if (INSN_DELETED_P (insn))
2580 else if (REG_P (src)
2581 && REGNO (src) >= FIRST_PSEUDO_REGISTER
2582 && REGNO (src) != regno)
2584 if (try_replace_reg (reg_used->reg_rtx, src, insn))
2587 global_copy_prop_count++;
2588 if (dump_file != NULL)
2590 fprintf (dump_file, "GLOBAL COPY-PROP: Replacing reg %d in insn %d",
2591 regno, INSN_UID (insn));
2592 fprintf (dump_file, " with reg %d\n", REGNO (src));
2595 /* The original insn setting reg_used may or may not now be
2596 deletable. We leave the deletion to flow. */
2597 /* FIXME: If it turns out that the insn isn't deletable,
2598 then we may have unnecessarily extended register lifetimes
2599 and made things worse. */
2604 if (changed && DEBUG_INSN_P (insn))
2610 /* Like find_used_regs, but avoid recording uses that appear in
2611 input-output contexts such as zero_extract or pre_dec. This
2612 restricts the cases we consider to those for which local cprop
2613 can legitimately make replacements. */
2616 local_cprop_find_used_regs (rtx *xptr, void *data)
2623 switch (GET_CODE (x))
2627 case STRICT_LOW_PART:
2636 /* Can only legitimately appear this early in the context of
2637 stack pushes for function arguments, but handle all of the
2638 codes nonetheless. */
2642 /* Setting a subreg of a register larger than word_mode leaves
2643 the non-written words unchanged. */
2644 if (GET_MODE_BITSIZE (GET_MODE (SUBREG_REG (x))) > BITS_PER_WORD)
2652 find_used_regs (xptr, data);
2655 /* Try to perform local const/copy propagation on X in INSN. */
2658 do_local_cprop (rtx x, rtx insn)
2660 rtx newreg = NULL, newcnst = NULL;
2662 /* Rule out USE instructions and ASM statements as we don't want to
2663 change the hard registers mentioned. */
2665 && (REGNO (x) >= FIRST_PSEUDO_REGISTER
2666 || (GET_CODE (PATTERN (insn)) != USE
2667 && asm_noperands (PATTERN (insn)) < 0)))
2669 cselib_val *val = cselib_lookup (x, GET_MODE (x), 0);
2670 struct elt_loc_list *l;
2674 for (l = val->locs; l; l = l->next)
2676 rtx this_rtx = l->loc;
2679 if (gcse_constant_p (this_rtx))
2681 if (REG_P (this_rtx) && REGNO (this_rtx) >= FIRST_PSEUDO_REGISTER
2682 /* Don't copy propagate if it has attached REG_EQUIV note.
2683 At this point this only function parameters should have
2684 REG_EQUIV notes and if the argument slot is used somewhere
2685 explicitly, it means address of parameter has been taken,
2686 so we should not extend the lifetime of the pseudo. */
2687 && (!(note = find_reg_note (l->setting_insn, REG_EQUIV, NULL_RTX))
2688 || ! MEM_P (XEXP (note, 0))))
2691 if (newcnst && constprop_register (insn, x, newcnst))
2693 if (dump_file != NULL)
2695 fprintf (dump_file, "LOCAL CONST-PROP: Replacing reg %d in ",
2697 fprintf (dump_file, "insn %d with constant ",
2699 print_rtl (dump_file, newcnst);
2700 fprintf (dump_file, "\n");
2702 local_const_prop_count++;
2705 else if (newreg && newreg != x && try_replace_reg (x, newreg, insn))
2707 if (dump_file != NULL)
2710 "LOCAL COPY-PROP: Replacing reg %d in insn %d",
2711 REGNO (x), INSN_UID (insn));
2712 fprintf (dump_file, " with reg %d\n", REGNO (newreg));
2714 local_copy_prop_count++;
2721 /* Do local const/copy propagation (i.e. within each basic block). */
2724 local_cprop_pass (void)
2728 struct reg_use *reg_used;
2729 bool changed = false;
2734 FOR_BB_INSNS (bb, insn)
2738 rtx note = find_reg_equal_equiv_note (insn);
2742 note_uses (&PATTERN (insn), local_cprop_find_used_regs,
2745 local_cprop_find_used_regs (&XEXP (note, 0), NULL);
2747 for (reg_used = ®_use_table[0]; reg_use_count > 0;
2748 reg_used++, reg_use_count--)
2750 if (do_local_cprop (reg_used->reg_rtx, insn))
2756 if (INSN_DELETED_P (insn))
2759 while (reg_use_count);
2761 cselib_process_insn (insn);
2764 /* Forget everything at the end of a basic block. */
2765 cselib_clear_table ();
2773 /* Similar to get_condition, only the resulting condition must be
2774 valid at JUMP, instead of at EARLIEST.
2776 This differs from noce_get_condition in ifcvt.c in that we prefer not to
2777 settle for the condition variable in the jump instruction being integral.
2778 We prefer to be able to record the value of a user variable, rather than
2779 the value of a temporary used in a condition. This could be solved by
2780 recording the value of *every* register scanned by canonicalize_condition,
2781 but this would require some code reorganization. */
2784 fis_get_condition (rtx jump)
2786 return get_condition (jump, NULL, false, true);
2789 /* Check the comparison COND to see if we can safely form an implicit set from
2790 it. COND is either an EQ or NE comparison. */
2793 implicit_set_cond_p (const_rtx cond)
2795 const enum machine_mode mode = GET_MODE (XEXP (cond, 0));
2796 const_rtx cst = XEXP (cond, 1);
2798 /* We can't perform this optimization if either operand might be or might
2799 contain a signed zero. */
2800 if (HONOR_SIGNED_ZEROS (mode))
2802 /* It is sufficient to check if CST is or contains a zero. We must
2803 handle float, complex, and vector. If any subpart is a zero, then
2804 the optimization can't be performed. */
2805 /* ??? The complex and vector checks are not implemented yet. We just
2806 always return zero for them. */
2807 if (GET_CODE (cst) == CONST_DOUBLE)
2810 REAL_VALUE_FROM_CONST_DOUBLE (d, cst);
2811 if (REAL_VALUES_EQUAL (d, dconst0))
2818 return gcse_constant_p (cst);
2821 /* Find the implicit sets of a function. An "implicit set" is a constraint
2822 on the value of a variable, implied by a conditional jump. For example,
2823 following "if (x == 2)", the then branch may be optimized as though the
2824 conditional performed an "explicit set", in this example, "x = 2". This
2825 function records the set patterns that are implicit at the start of each
2828 FIXME: This would be more effective if critical edges are pre-split. As
2829 it is now, we can't record implicit sets for blocks that have
2830 critical successor edges. This results in missed optimizations
2831 and in more (unnecessary) work in cfgcleanup.c:thread_jump(). */
2834 find_implicit_sets (void)
2836 basic_block bb, dest;
2842 /* Check for more than one successor. */
2843 if (EDGE_COUNT (bb->succs) > 1)
2845 cond = fis_get_condition (BB_END (bb));
2848 && (GET_CODE (cond) == EQ || GET_CODE (cond) == NE)
2849 && REG_P (XEXP (cond, 0))
2850 && REGNO (XEXP (cond, 0)) >= FIRST_PSEUDO_REGISTER
2851 && implicit_set_cond_p (cond))
2853 dest = GET_CODE (cond) == EQ ? BRANCH_EDGE (bb)->dest
2854 : FALLTHRU_EDGE (bb)->dest;
2857 /* Record nothing for a critical edge. */
2858 && single_pred_p (dest)
2859 && dest != EXIT_BLOCK_PTR)
2861 new_rtx = gen_rtx_SET (VOIDmode, XEXP (cond, 0),
2863 implicit_sets[dest->index] = new_rtx;
2866 fprintf(dump_file, "Implicit set of reg %d in ",
2867 REGNO (XEXP (cond, 0)));
2868 fprintf(dump_file, "basic block %d\n", dest->index);
2876 fprintf (dump_file, "Found %d implicit sets\n", count);
2879 /* Bypass conditional jumps. */
2881 /* The value of last_basic_block at the beginning of the jump_bypass
2882 pass. The use of redirect_edge_and_branch_force may introduce new
2883 basic blocks, but the data flow analysis is only valid for basic
2884 block indices less than bypass_last_basic_block. */
2886 static int bypass_last_basic_block;
2888 /* Find a set of REGNO to a constant that is available at the end of basic
2889 block BB. Returns NULL if no such set is found. Based heavily upon
2892 static struct expr *
2893 find_bypass_set (int regno, int bb)
2895 struct expr *result = 0;
2900 struct expr *set = lookup_set (regno, &set_hash_table);
2904 if (TEST_BIT (cprop_avout[bb], set->bitmap_index))
2906 set = next_set (regno, set);
2912 gcc_assert (GET_CODE (set->expr) == SET);
2914 src = SET_SRC (set->expr);
2915 if (gcse_constant_p (src))
2921 regno = REGNO (src);
2927 /* Subroutine of bypass_block that checks whether a pseudo is killed by
2928 any of the instructions inserted on an edge. Jump bypassing places
2929 condition code setters on CFG edges using insert_insn_on_edge. This
2930 function is required to check that our data flow analysis is still
2931 valid prior to commit_edge_insertions. */
2934 reg_killed_on_edge (const_rtx reg, const_edge e)
2938 for (insn = e->insns.r; insn; insn = NEXT_INSN (insn))
2939 if (INSN_P (insn) && reg_set_p (reg, insn))
2945 /* Subroutine of bypass_conditional_jumps that attempts to bypass the given
2946 basic block BB which has more than one predecessor. If not NULL, SETCC
2947 is the first instruction of BB, which is immediately followed by JUMP_INSN
2948 JUMP. Otherwise, SETCC is NULL, and JUMP is the first insn of BB.
2949 Returns nonzero if a change was made.
2951 During the jump bypassing pass, we may place copies of SETCC instructions
2952 on CFG edges. The following routine must be careful to pay attention to
2953 these inserted insns when performing its transformations. */
2956 bypass_block (basic_block bb, rtx setcc, rtx jump)
2961 int may_be_loop_header;
2965 insn = (setcc != NULL) ? setcc : jump;
2967 /* Determine set of register uses in INSN. */
2969 note_uses (&PATTERN (insn), find_used_regs, NULL);
2970 note = find_reg_equal_equiv_note (insn);
2972 find_used_regs (&XEXP (note, 0), NULL);
2974 may_be_loop_header = false;
2975 FOR_EACH_EDGE (e, ei, bb->preds)
2976 if (e->flags & EDGE_DFS_BACK)
2978 may_be_loop_header = true;
2983 for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); )
2987 if (e->flags & EDGE_COMPLEX)
2993 /* We can't redirect edges from new basic blocks. */
2994 if (e->src->index >= bypass_last_basic_block)
3000 /* The irreducible loops created by redirecting of edges entering the
3001 loop from outside would decrease effectiveness of some of the following
3002 optimizations, so prevent this. */
3003 if (may_be_loop_header
3004 && !(e->flags & EDGE_DFS_BACK))
3010 for (i = 0; i < reg_use_count; i++)
3012 struct reg_use *reg_used = ®_use_table[i];
3013 unsigned int regno = REGNO (reg_used->reg_rtx);
3014 basic_block dest, old_dest;
3018 set = find_bypass_set (regno, e->src->index);
3023 /* Check the data flow is valid after edge insertions. */
3024 if (e->insns.r && reg_killed_on_edge (reg_used->reg_rtx, e))
3027 src = SET_SRC (pc_set (jump));
3030 src = simplify_replace_rtx (src,
3031 SET_DEST (PATTERN (setcc)),
3032 SET_SRC (PATTERN (setcc)));
3034 new_rtx = simplify_replace_rtx (src, reg_used->reg_rtx,
3035 SET_SRC (set->expr));
3037 /* Jump bypassing may have already placed instructions on
3038 edges of the CFG. We can't bypass an outgoing edge that
3039 has instructions associated with it, as these insns won't
3040 get executed if the incoming edge is redirected. */
3042 if (new_rtx == pc_rtx)
3044 edest = FALLTHRU_EDGE (bb);
3045 dest = edest->insns.r ? NULL : edest->dest;
3047 else if (GET_CODE (new_rtx) == LABEL_REF)
3049 dest = BLOCK_FOR_INSN (XEXP (new_rtx, 0));
3050 /* Don't bypass edges containing instructions. */
3051 edest = find_edge (bb, dest);
3052 if (edest && edest->insns.r)
3058 /* Avoid unification of the edge with other edges from original
3059 branch. We would end up emitting the instruction on "both"
3062 if (dest && setcc && !CC0_P (SET_DEST (PATTERN (setcc)))
3063 && find_edge (e->src, dest))
3069 && dest != EXIT_BLOCK_PTR)
3071 redirect_edge_and_branch_force (e, dest);
3073 /* Copy the register setter to the redirected edge.
3074 Don't copy CC0 setters, as CC0 is dead after jump. */
3077 rtx pat = PATTERN (setcc);
3078 if (!CC0_P (SET_DEST (pat)))
3079 insert_insn_on_edge (copy_insn (pat), e);
3082 if (dump_file != NULL)
3084 fprintf (dump_file, "JUMP-BYPASS: Proved reg %d "
3085 "in jump_insn %d equals constant ",
3086 regno, INSN_UID (jump));
3087 print_rtl (dump_file, SET_SRC (set->expr));
3088 fprintf (dump_file, "\nBypass edge from %d->%d to %d\n",
3089 e->src->index, old_dest->index, dest->index);
3102 /* Find basic blocks with more than one predecessor that only contain a
3103 single conditional jump. If the result of the comparison is known at
3104 compile-time from any incoming edge, redirect that edge to the
3105 appropriate target. Returns nonzero if a change was made.
3107 This function is now mis-named, because we also handle indirect jumps. */
3110 bypass_conditional_jumps (void)
3118 /* Note we start at block 1. */
3119 if (ENTRY_BLOCK_PTR->next_bb == EXIT_BLOCK_PTR)
3122 bypass_last_basic_block = last_basic_block;
3123 mark_dfs_back_edges ();
3126 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb,
3127 EXIT_BLOCK_PTR, next_bb)
3129 /* Check for more than one predecessor. */
3130 if (!single_pred_p (bb))
3133 FOR_BB_INSNS (bb, insn)
3134 if (DEBUG_INSN_P (insn))
3136 else if (NONJUMP_INSN_P (insn))
3140 if (GET_CODE (PATTERN (insn)) != SET)
3143 dest = SET_DEST (PATTERN (insn));
3144 if (REG_P (dest) || CC0_P (dest))
3149 else if (JUMP_P (insn))
3151 if ((any_condjump_p (insn) || computed_jump_p (insn))
3152 && onlyjump_p (insn))
3153 changed |= bypass_block (bb, setcc, insn);
3156 else if (INSN_P (insn))
3161 /* If we bypassed any register setting insns, we inserted a
3162 copy on the redirected edge. These need to be committed. */
3164 commit_edge_insertions ();
3169 /* Compute PRE+LCM working variables. */
3171 /* Local properties of expressions. */
3172 /* Nonzero for expressions that are transparent in the block. */
3173 static sbitmap *transp;
3175 /* Nonzero for expressions that are transparent at the end of the block.
3176 This is only zero for expressions killed by abnormal critical edge
3177 created by a calls. */
3178 static sbitmap *transpout;
3180 /* Nonzero for expressions that are computed (available) in the block. */
3181 static sbitmap *comp;
3183 /* Nonzero for expressions that are locally anticipatable in the block. */
3184 static sbitmap *antloc;
3186 /* Nonzero for expressions where this block is an optimal computation
3188 static sbitmap *pre_optimal;
3190 /* Nonzero for expressions which are redundant in a particular block. */
3191 static sbitmap *pre_redundant;
3193 /* Nonzero for expressions which should be inserted on a specific edge. */
3194 static sbitmap *pre_insert_map;
3196 /* Nonzero for expressions which should be deleted in a specific block. */
3197 static sbitmap *pre_delete_map;
3199 /* Contains the edge_list returned by pre_edge_lcm. */
3200 static struct edge_list *edge_list;
3202 /* Allocate vars used for PRE analysis. */
3205 alloc_pre_mem (int n_blocks, int n_exprs)
3207 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
3208 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
3209 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
3212 pre_redundant = NULL;
3213 pre_insert_map = NULL;
3214 pre_delete_map = NULL;
3215 ae_kill = sbitmap_vector_alloc (n_blocks, n_exprs);
3217 /* pre_insert and pre_delete are allocated later. */
3220 /* Free vars used for PRE analysis. */
3225 sbitmap_vector_free (transp);
3226 sbitmap_vector_free (comp);
3228 /* ANTLOC and AE_KILL are freed just after pre_lcm finishes. */
3231 sbitmap_vector_free (pre_optimal);
3233 sbitmap_vector_free (pre_redundant);
3235 sbitmap_vector_free (pre_insert_map);
3237 sbitmap_vector_free (pre_delete_map);
3239 transp = comp = NULL;
3240 pre_optimal = pre_redundant = pre_insert_map = pre_delete_map = NULL;
3243 /* Top level routine to do the dataflow analysis needed by PRE. */
3246 compute_pre_data (void)
3248 sbitmap trapping_expr;
3252 compute_local_properties (transp, comp, antloc, &expr_hash_table);
3253 sbitmap_vector_zero (ae_kill, last_basic_block);
3255 /* Collect expressions which might trap. */
3256 trapping_expr = sbitmap_alloc (expr_hash_table.n_elems);
3257 sbitmap_zero (trapping_expr);
3258 for (ui = 0; ui < expr_hash_table.size; ui++)
3261 for (e = expr_hash_table.table[ui]; e != NULL; e = e->next_same_hash)
3262 if (may_trap_p (e->expr))
3263 SET_BIT (trapping_expr, e->bitmap_index);
3266 /* Compute ae_kill for each basic block using:
3276 /* If the current block is the destination of an abnormal edge, we
3277 kill all trapping expressions because we won't be able to properly
3278 place the instruction on the edge. So make them neither
3279 anticipatable nor transparent. This is fairly conservative. */
3280 FOR_EACH_EDGE (e, ei, bb->preds)
3281 if (e->flags & EDGE_ABNORMAL)
3283 sbitmap_difference (antloc[bb->index], antloc[bb->index], trapping_expr);
3284 sbitmap_difference (transp[bb->index], transp[bb->index], trapping_expr);
3288 sbitmap_a_or_b (ae_kill[bb->index], transp[bb->index], comp[bb->index]);
3289 sbitmap_not (ae_kill[bb->index], ae_kill[bb->index]);
3292 edge_list = pre_edge_lcm (expr_hash_table.n_elems, transp, comp, antloc,
3293 ae_kill, &pre_insert_map, &pre_delete_map);
3294 sbitmap_vector_free (antloc);
3296 sbitmap_vector_free (ae_kill);
3298 sbitmap_free (trapping_expr);
3303 /* Return nonzero if an occurrence of expression EXPR in OCCR_BB would reach
3306 VISITED is a pointer to a working buffer for tracking which BB's have
3307 been visited. It is NULL for the top-level call.
3309 We treat reaching expressions that go through blocks containing the same
3310 reaching expression as "not reaching". E.g. if EXPR is generated in blocks
3311 2 and 3, INSN is in block 4, and 2->3->4, we treat the expression in block
3312 2 as not reaching. The intent is to improve the probability of finding
3313 only one reaching expression and to reduce register lifetimes by picking
3314 the closest such expression. */
3317 pre_expr_reaches_here_p_work (basic_block occr_bb, struct expr *expr, basic_block bb, char *visited)
3322 FOR_EACH_EDGE (pred, ei, bb->preds)
3324 basic_block pred_bb = pred->src;
3326 if (pred->src == ENTRY_BLOCK_PTR
3327 /* Has predecessor has already been visited? */
3328 || visited[pred_bb->index])
3329 ;/* Nothing to do. */
3331 /* Does this predecessor generate this expression? */
3332 else if (TEST_BIT (comp[pred_bb->index], expr->bitmap_index))
3334 /* Is this the occurrence we're looking for?
3335 Note that there's only one generating occurrence per block
3336 so we just need to check the block number. */
3337 if (occr_bb == pred_bb)
3340 visited[pred_bb->index] = 1;
3342 /* Ignore this predecessor if it kills the expression. */
3343 else if (! TEST_BIT (transp[pred_bb->index], expr->bitmap_index))
3344 visited[pred_bb->index] = 1;
3346 /* Neither gen nor kill. */
3349 visited[pred_bb->index] = 1;
3350 if (pre_expr_reaches_here_p_work (occr_bb, expr, pred_bb, visited))
3355 /* All paths have been checked. */
3359 /* The wrapper for pre_expr_reaches_here_work that ensures that any
3360 memory allocated for that function is returned. */
3363 pre_expr_reaches_here_p (basic_block occr_bb, struct expr *expr, basic_block bb)
3366 char *visited = XCNEWVEC (char, last_basic_block);
3368 rval = pre_expr_reaches_here_p_work (occr_bb, expr, bb, visited);
3375 /* Given an expr, generate RTL which we can insert at the end of a BB,
3376 or on an edge. Set the block number of any insns generated to
3380 process_insert_insn (struct expr *expr)
3382 rtx reg = expr->reaching_reg;
3383 rtx exp = copy_rtx (expr->expr);
3388 /* If the expression is something that's an operand, like a constant,
3389 just copy it to a register. */
3390 if (general_operand (exp, GET_MODE (reg)))
3391 emit_move_insn (reg, exp);
3393 /* Otherwise, make a new insn to compute this expression and make sure the
3394 insn will be recognized (this also adds any needed CLOBBERs). Copy the
3395 expression to make sure we don't have any sharing issues. */
3398 rtx insn = emit_insn (gen_rtx_SET (VOIDmode, reg, exp));
3400 if (insn_invalid_p (insn))
3411 /* Add EXPR to the end of basic block BB.
3413 This is used by both the PRE and code hoisting.
3415 For PRE, we want to verify that the expr is either transparent
3416 or locally anticipatable in the target block. This check makes
3417 no sense for code hoisting. */
3420 insert_insn_end_basic_block (struct expr *expr, basic_block bb, int pre)
3422 rtx insn = BB_END (bb);
3424 rtx reg = expr->reaching_reg;
3425 int regno = REGNO (reg);
3428 pat = process_insert_insn (expr);
3429 gcc_assert (pat && INSN_P (pat));
3432 while (NEXT_INSN (pat_end) != NULL_RTX)
3433 pat_end = NEXT_INSN (pat_end);
3435 /* If the last insn is a jump, insert EXPR in front [taking care to
3436 handle cc0, etc. properly]. Similarly we need to care trapping
3437 instructions in presence of non-call exceptions. */
3440 || (NONJUMP_INSN_P (insn)
3441 && (!single_succ_p (bb)
3442 || single_succ_edge (bb)->flags & EDGE_ABNORMAL)))
3447 /* It should always be the case that we can put these instructions
3448 anywhere in the basic block with performing PRE optimizations.
3450 gcc_assert (!NONJUMP_INSN_P (insn) || !pre
3451 || TEST_BIT (antloc[bb->index], expr->bitmap_index)
3452 || TEST_BIT (transp[bb->index], expr->bitmap_index));
3454 /* If this is a jump table, then we can't insert stuff here. Since
3455 we know the previous real insn must be the tablejump, we insert
3456 the new instruction just before the tablejump. */
3457 if (GET_CODE (PATTERN (insn)) == ADDR_VEC
3458 || GET_CODE (PATTERN (insn)) == ADDR_DIFF_VEC)
3459 insn = prev_real_insn (insn);
3462 /* FIXME: 'twould be nice to call prev_cc0_setter here but it aborts
3463 if cc0 isn't set. */
3464 note = find_reg_note (insn, REG_CC_SETTER, NULL_RTX);
3466 insn = XEXP (note, 0);
3469 rtx maybe_cc0_setter = prev_nonnote_insn (insn);
3470 if (maybe_cc0_setter
3471 && INSN_P (maybe_cc0_setter)
3472 && sets_cc0_p (PATTERN (maybe_cc0_setter)))
3473 insn = maybe_cc0_setter;
3476 /* FIXME: What if something in cc0/jump uses value set in new insn? */
3477 new_insn = emit_insn_before_noloc (pat, insn, bb);
3480 /* Likewise if the last insn is a call, as will happen in the presence
3481 of exception handling. */
3482 else if (CALL_P (insn)
3483 && (!single_succ_p (bb)
3484 || single_succ_edge (bb)->flags & EDGE_ABNORMAL))
3486 /* Keeping in mind targets with small register classes and parameters
3487 in registers, we search backward and place the instructions before
3488 the first parameter is loaded. Do this for everyone for consistency
3489 and a presumption that we'll get better code elsewhere as well.
3491 It should always be the case that we can put these instructions
3492 anywhere in the basic block with performing PRE optimizations.
3496 || TEST_BIT (antloc[bb->index], expr->bitmap_index)
3497 || TEST_BIT (transp[bb->index], expr->bitmap_index));
3499 /* Since different machines initialize their parameter registers
3500 in different orders, assume nothing. Collect the set of all
3501 parameter registers. */
3502 insn = find_first_parameter_load (insn, BB_HEAD (bb));
3504 /* If we found all the parameter loads, then we want to insert
3505 before the first parameter load.
3507 If we did not find all the parameter loads, then we might have
3508 stopped on the head of the block, which could be a CODE_LABEL.
3509 If we inserted before the CODE_LABEL, then we would be putting
3510 the insn in the wrong basic block. In that case, put the insn
3511 after the CODE_LABEL. Also, respect NOTE_INSN_BASIC_BLOCK. */
3512 while (LABEL_P (insn)
3513 || NOTE_INSN_BASIC_BLOCK_P (insn))
3514 insn = NEXT_INSN (insn);
3516 new_insn = emit_insn_before_noloc (pat, insn, bb);
3519 new_insn = emit_insn_after_noloc (pat, insn, bb);
3524 add_label_notes (PATTERN (pat), new_insn);
3527 pat = NEXT_INSN (pat);
3530 gcse_create_count++;
3534 fprintf (dump_file, "PRE/HOIST: end of bb %d, insn %d, ",
3535 bb->index, INSN_UID (new_insn));
3536 fprintf (dump_file, "copying expression %d to reg %d\n",
3537 expr->bitmap_index, regno);
3541 /* Insert partially redundant expressions on edges in the CFG to make
3542 the expressions fully redundant. */
3545 pre_edge_insert (struct edge_list *edge_list, struct expr **index_map)
3547 int e, i, j, num_edges, set_size, did_insert = 0;
3550 /* Where PRE_INSERT_MAP is nonzero, we add the expression on that edge
3551 if it reaches any of the deleted expressions. */
3553 set_size = pre_insert_map[0]->size;
3554 num_edges = NUM_EDGES (edge_list);
3555 inserted = sbitmap_vector_alloc (num_edges, expr_hash_table.n_elems);
3556 sbitmap_vector_zero (inserted, num_edges);
3558 for (e = 0; e < num_edges; e++)
3561 basic_block bb = INDEX_EDGE_PRED_BB (edge_list, e);
3563 for (i = indx = 0; i < set_size; i++, indx += SBITMAP_ELT_BITS)
3565 SBITMAP_ELT_TYPE insert = pre_insert_map[e]->elms[i];
3567 for (j = indx; insert && j < (int) expr_hash_table.n_elems; j++, insert >>= 1)
3568 if ((insert & 1) != 0 && index_map[j]->reaching_reg != NULL_RTX)
3570 struct expr *expr = index_map[j];
3573 /* Now look at each deleted occurrence of this expression. */
3574 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
3576 if (! occr->deleted_p)
3579 /* Insert this expression on this edge if it would
3580 reach the deleted occurrence in BB. */
3581 if (!TEST_BIT (inserted[e], j))
3584 edge eg = INDEX_EDGE (edge_list, e);
3586 /* We can't insert anything on an abnormal and
3587 critical edge, so we insert the insn at the end of
3588 the previous block. There are several alternatives
3589 detailed in Morgans book P277 (sec 10.5) for
3590 handling this situation. This one is easiest for
3593 if (eg->flags & EDGE_ABNORMAL)
3594 insert_insn_end_basic_block (index_map[j], bb, 0);
3597 insn = process_insert_insn (index_map[j]);
3598 insert_insn_on_edge (insn, eg);
3603 fprintf (dump_file, "PRE: edge (%d,%d), ",
3605 INDEX_EDGE_SUCC_BB (edge_list, e)->index);
3606 fprintf (dump_file, "copy expression %d\n",
3607 expr->bitmap_index);
3610 update_ld_motion_stores (expr);
3611 SET_BIT (inserted[e], j);
3613 gcse_create_count++;
3620 sbitmap_vector_free (inserted);
3624 /* Copy the result of EXPR->EXPR generated by INSN to EXPR->REACHING_REG.
3625 Given "old_reg <- expr" (INSN), instead of adding after it
3626 reaching_reg <- old_reg
3627 it's better to do the following:
3628 reaching_reg <- expr
3629 old_reg <- reaching_reg
3630 because this way copy propagation can discover additional PRE
3631 opportunities. But if this fails, we try the old way.
3632 When "expr" is a store, i.e.
3633 given "MEM <- old_reg", instead of adding after it
3634 reaching_reg <- old_reg
3635 it's better to add it before as follows:
3636 reaching_reg <- old_reg
3637 MEM <- reaching_reg. */
3640 pre_insert_copy_insn (struct expr *expr, rtx insn)
3642 rtx reg = expr->reaching_reg;
3643 int regno = REGNO (reg);
3644 int indx = expr->bitmap_index;
3645 rtx pat = PATTERN (insn);
3646 rtx set, first_set, new_insn;
3650 /* This block matches the logic in hash_scan_insn. */
3651 switch (GET_CODE (pat))
3658 /* Search through the parallel looking for the set whose
3659 source was the expression that we're interested in. */
3660 first_set = NULL_RTX;
3662 for (i = 0; i < XVECLEN (pat, 0); i++)
3664 rtx x = XVECEXP (pat, 0, i);
3665 if (GET_CODE (x) == SET)
3667 /* If the source was a REG_EQUAL or REG_EQUIV note, we
3668 may not find an equivalent expression, but in this
3669 case the PARALLEL will have a single set. */
3670 if (first_set == NULL_RTX)
3672 if (expr_equiv_p (SET_SRC (x), expr->expr))
3680 gcc_assert (first_set);
3681 if (set == NULL_RTX)
3689 if (REG_P (SET_DEST (set)))
3691 old_reg = SET_DEST (set);
3692 /* Check if we can modify the set destination in the original insn. */
3693 if (validate_change (insn, &SET_DEST (set), reg, 0))
3695 new_insn = gen_move_insn (old_reg, reg);
3696 new_insn = emit_insn_after (new_insn, insn);
3700 new_insn = gen_move_insn (reg, old_reg);
3701 new_insn = emit_insn_after (new_insn, insn);
3704 else /* This is possible only in case of a store to memory. */
3706 old_reg = SET_SRC (set);
3707 new_insn = gen_move_insn (reg, old_reg);
3709 /* Check if we can modify the set source in the original insn. */
3710 if (validate_change (insn, &SET_SRC (set), reg, 0))
3711 new_insn = emit_insn_before (new_insn, insn);
3713 new_insn = emit_insn_after (new_insn, insn);
3716 gcse_create_count++;
3720 "PRE: bb %d, insn %d, copy expression %d in insn %d to reg %d\n",
3721 BLOCK_FOR_INSN (insn)->index, INSN_UID (new_insn), indx,
3722 INSN_UID (insn), regno);
3725 /* Copy available expressions that reach the redundant expression
3726 to `reaching_reg'. */
3729 pre_insert_copies (void)
3731 unsigned int i, added_copy;
3736 /* For each available expression in the table, copy the result to
3737 `reaching_reg' if the expression reaches a deleted one.
3739 ??? The current algorithm is rather brute force.
3740 Need to do some profiling. */
3742 for (i = 0; i < expr_hash_table.size; i++)
3743 for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
3745 /* If the basic block isn't reachable, PPOUT will be TRUE. However,
3746 we don't want to insert a copy here because the expression may not
3747 really be redundant. So only insert an insn if the expression was
3748 deleted. This test also avoids further processing if the
3749 expression wasn't deleted anywhere. */
3750 if (expr->reaching_reg == NULL)
3753 /* Set when we add a copy for that expression. */
3756 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
3758 if (! occr->deleted_p)
3761 for (avail = expr->avail_occr; avail != NULL; avail = avail->next)
3763 rtx insn = avail->insn;
3765 /* No need to handle this one if handled already. */
3766 if (avail->copied_p)
3769 /* Don't handle this one if it's a redundant one. */
3770 if (INSN_DELETED_P (insn))
3773 /* Or if the expression doesn't reach the deleted one. */
3774 if (! pre_expr_reaches_here_p (BLOCK_FOR_INSN (avail->insn),
3776 BLOCK_FOR_INSN (occr->insn)))
3781 /* Copy the result of avail to reaching_reg. */
3782 pre_insert_copy_insn (expr, insn);
3783 avail->copied_p = 1;
3788 update_ld_motion_stores (expr);
3792 /* Emit move from SRC to DEST noting the equivalence with expression computed
3795 gcse_emit_move_after (rtx src, rtx dest, rtx insn)
3798 rtx set = single_set (insn), set2;
3802 /* This should never fail since we're creating a reg->reg copy
3803 we've verified to be valid. */
3805 new_rtx = emit_insn_after (gen_move_insn (dest, src), insn);
3807 /* Note the equivalence for local CSE pass. */
3808 set2 = single_set (new_rtx);
3809 if (!set2 || !rtx_equal_p (SET_DEST (set2), dest))
3811 if ((note = find_reg_equal_equiv_note (insn)))
3812 eqv = XEXP (note, 0);
3814 eqv = SET_SRC (set);
3816 set_unique_reg_note (new_rtx, REG_EQUAL, copy_insn_1 (eqv));
3821 /* Delete redundant computations.
3822 Deletion is done by changing the insn to copy the `reaching_reg' of
3823 the expression into the result of the SET. It is left to later passes
3824 (cprop, cse2, flow, combine, regmove) to propagate the copy or eliminate it.
3826 Returns nonzero if a change is made. */
3837 for (i = 0; i < expr_hash_table.size; i++)
3838 for (expr = expr_hash_table.table[i];
3840 expr = expr->next_same_hash)
3842 int indx = expr->bitmap_index;
3844 /* We only need to search antic_occr since we require
3847 for (occr = expr->antic_occr; occr != NULL; occr = occr->next)
3849 rtx insn = occr->insn;
3851 basic_block bb = BLOCK_FOR_INSN (insn);
3853 /* We only delete insns that have a single_set. */
3854 if (TEST_BIT (pre_delete_map[bb->index], indx)
3855 && (set = single_set (insn)) != 0
3856 && dbg_cnt (pre_insn))
3858 /* Create a pseudo-reg to store the result of reaching
3859 expressions into. Get the mode for the new pseudo from
3860 the mode of the original destination pseudo. */
3861 if (expr->reaching_reg == NULL)
3862 expr->reaching_reg = gen_reg_rtx_and_attrs (SET_DEST (set));
3864 gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
3866 occr->deleted_p = 1;
3873 "PRE: redundant insn %d (expression %d) in ",
3874 INSN_UID (insn), indx);
3875 fprintf (dump_file, "bb %d, reaching reg is %d\n",
3876 bb->index, REGNO (expr->reaching_reg));
3885 /* Perform GCSE optimizations using PRE.
3886 This is called by one_pre_gcse_pass after all the dataflow analysis
3889 This is based on the original Morel-Renvoise paper Fred Chow's thesis, and
3890 lazy code motion from Knoop, Ruthing and Steffen as described in Advanced
3891 Compiler Design and Implementation.
3893 ??? A new pseudo reg is created to hold the reaching expression. The nice
3894 thing about the classical approach is that it would try to use an existing
3895 reg. If the register can't be adequately optimized [i.e. we introduce
3896 reload problems], one could add a pass here to propagate the new register
3899 ??? We don't handle single sets in PARALLELs because we're [currently] not
3900 able to copy the rest of the parallel when we insert copies to create full
3901 redundancies from partial redundancies. However, there's no reason why we
3902 can't handle PARALLELs in the cases where there are no partial
3909 int did_insert, changed;
3910 struct expr **index_map;
3913 /* Compute a mapping from expression number (`bitmap_index') to
3914 hash table entry. */
3916 index_map = XCNEWVEC (struct expr *, expr_hash_table.n_elems);
3917 for (i = 0; i < expr_hash_table.size; i++)
3918 for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
3919 index_map[expr->bitmap_index] = expr;
3921 /* Delete the redundant insns first so that
3922 - we know what register to use for the new insns and for the other
3923 ones with reaching expressions
3924 - we know which insns are redundant when we go to create copies */
3926 changed = pre_delete ();
3927 did_insert = pre_edge_insert (edge_list, index_map);
3929 /* In other places with reaching expressions, copy the expression to the
3930 specially allocated pseudo-reg that reaches the redundant expr. */
3931 pre_insert_copies ();
3934 commit_edge_insertions ();
3942 /* Top level routine to perform one PRE GCSE pass.
3944 Return nonzero if a change was made. */
3947 one_pre_gcse_pass (void)
3951 gcse_subst_count = 0;
3952 gcse_create_count = 0;
3954 /* Return if there's nothing to do, or it is too expensive. */
3955 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
3956 || is_too_expensive (_("PRE disabled")))
3959 /* We need alias. */
3960 init_alias_analysis ();
3963 gcc_obstack_init (&gcse_obstack);
3966 alloc_hash_table (&expr_hash_table, 0);
3967 add_noreturn_fake_exit_edges ();
3969 compute_ld_motion_mems ();
3971 compute_hash_table (&expr_hash_table);
3972 trim_ld_motion_mems ();
3974 dump_hash_table (dump_file, "Expression", &expr_hash_table);
3976 if (expr_hash_table.n_elems > 0)
3978 alloc_pre_mem (last_basic_block, expr_hash_table.n_elems);
3979 compute_pre_data ();
3980 changed |= pre_gcse ();
3981 free_edge_list (edge_list);
3986 remove_fake_exit_edges ();
3987 free_hash_table (&expr_hash_table);
3990 obstack_free (&gcse_obstack, NULL);
3992 /* We are finished with alias. */
3993 end_alias_analysis ();
3997 fprintf (dump_file, "PRE GCSE of %s, %d basic blocks, %d bytes needed, ",
3998 current_function_name (), n_basic_blocks, bytes_used);
3999 fprintf (dump_file, "%d substs, %d insns created\n",
4000 gcse_subst_count, gcse_create_count);
4006 /* If X contains any LABEL_REF's, add REG_LABEL_OPERAND notes for them
4007 to INSN. If such notes are added to an insn which references a
4008 CODE_LABEL, the LABEL_NUSES count is incremented. We have to add
4009 that note, because the following loop optimization pass requires
4012 /* ??? If there was a jump optimization pass after gcse and before loop,
4013 then we would not need to do this here, because jump would add the
4014 necessary REG_LABEL_OPERAND and REG_LABEL_TARGET notes. */
4017 add_label_notes (rtx x, rtx insn)
4019 enum rtx_code code = GET_CODE (x);
4023 if (code == LABEL_REF && !LABEL_REF_NONLOCAL_P (x))
4025 /* This code used to ignore labels that referred to dispatch tables to
4026 avoid flow generating (slightly) worse code.
4028 We no longer ignore such label references (see LABEL_REF handling in
4029 mark_jump_label for additional information). */
4031 /* There's no reason for current users to emit jump-insns with
4032 such a LABEL_REF, so we don't have to handle REG_LABEL_TARGET
4034 gcc_assert (!JUMP_P (insn));
4035 add_reg_note (insn, REG_LABEL_OPERAND, XEXP (x, 0));
4037 if (LABEL_P (XEXP (x, 0)))
4038 LABEL_NUSES (XEXP (x, 0))++;
4043 for (i = GET_RTX_LENGTH (code) - 1, fmt = GET_RTX_FORMAT (code); i >= 0; i--)
4046 add_label_notes (XEXP (x, i), insn);
4047 else if (fmt[i] == 'E')
4048 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4049 add_label_notes (XVECEXP (x, i, j), insn);
4053 /* Compute transparent outgoing information for each block.
4055 An expression is transparent to an edge unless it is killed by
4056 the edge itself. This can only happen with abnormal control flow,
4057 when the edge is traversed through a call. This happens with
4058 non-local labels and exceptions.
4060 This would not be necessary if we split the edge. While this is
4061 normally impossible for abnormal critical edges, with some effort
4062 it should be possible with exception handling, since we still have
4063 control over which handler should be invoked. But due to increased
4064 EH table sizes, this may not be worthwhile. */
4067 compute_transpout (void)
4073 sbitmap_vector_ones (transpout, last_basic_block);
4077 /* Note that flow inserted a nop at the end of basic blocks that
4078 end in call instructions for reasons other than abnormal
4080 if (! CALL_P (BB_END (bb)))
4083 for (i = 0; i < expr_hash_table.size; i++)
4084 for (expr = expr_hash_table.table[i]; expr ; expr = expr->next_same_hash)
4085 if (MEM_P (expr->expr))
4087 if (GET_CODE (XEXP (expr->expr, 0)) == SYMBOL_REF
4088 && CONSTANT_POOL_ADDRESS_P (XEXP (expr->expr, 0)))
4091 /* ??? Optimally, we would use interprocedural alias
4092 analysis to determine if this mem is actually killed
4094 RESET_BIT (transpout[bb->index], expr->bitmap_index);
4099 /* Code Hoisting variables and subroutines. */
4101 /* Very busy expressions. */
4102 static sbitmap *hoist_vbein;
4103 static sbitmap *hoist_vbeout;
4105 /* Hoistable expressions. */
4106 static sbitmap *hoist_exprs;
4108 /* ??? We could compute post dominators and run this algorithm in
4109 reverse to perform tail merging, doing so would probably be
4110 more effective than the tail merging code in jump.c.
4112 It's unclear if tail merging could be run in parallel with
4113 code hoisting. It would be nice. */
4115 /* Allocate vars used for code hoisting analysis. */
4118 alloc_code_hoist_mem (int n_blocks, int n_exprs)
4120 antloc = sbitmap_vector_alloc (n_blocks, n_exprs);
4121 transp = sbitmap_vector_alloc (n_blocks, n_exprs);
4122 comp = sbitmap_vector_alloc (n_blocks, n_exprs);
4124 hoist_vbein = sbitmap_vector_alloc (n_blocks, n_exprs);
4125 hoist_vbeout = sbitmap_vector_alloc (n_blocks, n_exprs);
4126 hoist_exprs = sbitmap_vector_alloc (n_blocks, n_exprs);
4127 transpout = sbitmap_vector_alloc (n_blocks, n_exprs);
4130 /* Free vars used for code hoisting analysis. */
4133 free_code_hoist_mem (void)
4135 sbitmap_vector_free (antloc);
4136 sbitmap_vector_free (transp);
4137 sbitmap_vector_free (comp);
4139 sbitmap_vector_free (hoist_vbein);
4140 sbitmap_vector_free (hoist_vbeout);
4141 sbitmap_vector_free (hoist_exprs);
4142 sbitmap_vector_free (transpout);
4144 free_dominance_info (CDI_DOMINATORS);
4147 /* Compute the very busy expressions at entry/exit from each block.
4149 An expression is very busy if all paths from a given point
4150 compute the expression. */
4153 compute_code_hoist_vbeinout (void)
4155 int changed, passes;
4158 sbitmap_vector_zero (hoist_vbeout, last_basic_block);
4159 sbitmap_vector_zero (hoist_vbein, last_basic_block);
4168 /* We scan the blocks in the reverse order to speed up
4170 FOR_EACH_BB_REVERSE (bb)
4172 if (bb->next_bb != EXIT_BLOCK_PTR)
4173 sbitmap_intersection_of_succs (hoist_vbeout[bb->index],
4174 hoist_vbein, bb->index);
4176 changed |= sbitmap_a_or_b_and_c_cg (hoist_vbein[bb->index],
4178 hoist_vbeout[bb->index],
4186 fprintf (dump_file, "hoisting vbeinout computation: %d passes\n", passes);
4189 /* Top level routine to do the dataflow analysis needed by code hoisting. */
4192 compute_code_hoist_data (void)
4194 compute_local_properties (transp, comp, antloc, &expr_hash_table);
4195 compute_transpout ();
4196 compute_code_hoist_vbeinout ();
4197 calculate_dominance_info (CDI_DOMINATORS);
4199 fprintf (dump_file, "\n");
4202 /* Determine if the expression identified by EXPR_INDEX would
4203 reach BB unimpared if it was placed at the end of EXPR_BB.
4205 It's unclear exactly what Muchnick meant by "unimpared". It seems
4206 to me that the expression must either be computed or transparent in
4207 *every* block in the path(s) from EXPR_BB to BB. Any other definition
4208 would allow the expression to be hoisted out of loops, even if
4209 the expression wasn't a loop invariant.
4211 Contrast this to reachability for PRE where an expression is
4212 considered reachable if *any* path reaches instead of *all*
4216 hoist_expr_reaches_here_p (basic_block expr_bb, int expr_index, basic_block bb, char *visited)
4220 int visited_allocated_locally = 0;
4223 if (visited == NULL)
4225 visited_allocated_locally = 1;
4226 visited = XCNEWVEC (char, last_basic_block);
4229 FOR_EACH_EDGE (pred, ei, bb->preds)
4231 basic_block pred_bb = pred->src;
4233 if (pred->src == ENTRY_BLOCK_PTR)
4235 else if (pred_bb == expr_bb)
4237 else if (visited[pred_bb->index])
4240 /* Does this predecessor generate this expression? */
4241 else if (TEST_BIT (comp[pred_bb->index], expr_index))
4243 else if (! TEST_BIT (transp[pred_bb->index], expr_index))
4249 visited[pred_bb->index] = 1;
4250 if (! hoist_expr_reaches_here_p (expr_bb, expr_index,
4255 if (visited_allocated_locally)
4258 return (pred == NULL);
4261 /* Actually perform code hoisting. */
4266 basic_block bb, dominated;
4267 VEC (basic_block, heap) *domby;
4269 struct expr **index_map;
4273 sbitmap_vector_zero (hoist_exprs, last_basic_block);
4275 /* Compute a mapping from expression number (`bitmap_index') to
4276 hash table entry. */
4278 index_map = XCNEWVEC (struct expr *, expr_hash_table.n_elems);
4279 for (i = 0; i < expr_hash_table.size; i++)
4280 for (expr = expr_hash_table.table[i]; expr != NULL; expr = expr->next_same_hash)
4281 index_map[expr->bitmap_index] = expr;
4283 /* Walk over each basic block looking for potentially hoistable
4284 expressions, nothing gets hoisted from the entry block. */
4288 int insn_inserted_p;
4290 domby = get_dominated_by (CDI_DOMINATORS, bb);
4291 /* Examine each expression that is very busy at the exit of this
4292 block. These are the potentially hoistable expressions. */
4293 for (i = 0; i < hoist_vbeout[bb->index]->n_bits; i++)
4297 if (TEST_BIT (hoist_vbeout[bb->index], i)
4298 && TEST_BIT (transpout[bb->index], i))
4300 /* We've found a potentially hoistable expression, now
4301 we look at every block BB dominates to see if it
4302 computes the expression. */
4303 for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
4305 /* Ignore self dominance. */
4306 if (bb == dominated)
4308 /* We've found a dominated block, now see if it computes
4309 the busy expression and whether or not moving that
4310 expression to the "beginning" of that block is safe. */
4311 if (!TEST_BIT (antloc[dominated->index], i))
4314 /* Note if the expression would reach the dominated block
4315 unimpared if it was placed at the end of BB.
4317 Keep track of how many times this expression is hoistable
4318 from a dominated block into BB. */
4319 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
4323 /* If we found more than one hoistable occurrence of this
4324 expression, then note it in the bitmap of expressions to
4325 hoist. It makes no sense to hoist things which are computed
4326 in only one BB, and doing so tends to pessimize register
4327 allocation. One could increase this value to try harder
4328 to avoid any possible code expansion due to register
4329 allocation issues; however experiments have shown that
4330 the vast majority of hoistable expressions are only movable
4331 from two successors, so raising this threshold is likely
4332 to nullify any benefit we get from code hoisting. */
4335 SET_BIT (hoist_exprs[bb->index], i);
4340 /* If we found nothing to hoist, then quit now. */
4343 VEC_free (basic_block, heap, domby);
4347 /* Loop over all the hoistable expressions. */
4348 for (i = 0; i < hoist_exprs[bb->index]->n_bits; i++)
4350 /* We want to insert the expression into BB only once, so
4351 note when we've inserted it. */
4352 insn_inserted_p = 0;
4354 /* These tests should be the same as the tests above. */
4355 if (TEST_BIT (hoist_exprs[bb->index], i))
4357 /* We've found a potentially hoistable expression, now
4358 we look at every block BB dominates to see if it
4359 computes the expression. */
4360 for (j = 0; VEC_iterate (basic_block, domby, j, dominated); j++)
4362 /* Ignore self dominance. */
4363 if (bb == dominated)
4366 /* We've found a dominated block, now see if it computes
4367 the busy expression and whether or not moving that
4368 expression to the "beginning" of that block is safe. */
4369 if (!TEST_BIT (antloc[dominated->index], i))
4372 /* The expression is computed in the dominated block and
4373 it would be safe to compute it at the start of the
4374 dominated block. Now we have to determine if the
4375 expression would reach the dominated block if it was
4376 placed at the end of BB. */
4377 if (hoist_expr_reaches_here_p (bb, i, dominated, NULL))
4379 struct expr *expr = index_map[i];
4380 struct occr *occr = expr->antic_occr;
4384 /* Find the right occurrence of this expression. */
4385 while (BLOCK_FOR_INSN (occr->insn) != dominated && occr)
4390 set = single_set (insn);
4393 /* Create a pseudo-reg to store the result of reaching
4394 expressions into. Get the mode for the new pseudo
4395 from the mode of the original destination pseudo. */
4396 if (expr->reaching_reg == NULL)
4398 = gen_reg_rtx_and_attrs (SET_DEST (set));
4400 gcse_emit_move_after (expr->reaching_reg, SET_DEST (set), insn);
4402 occr->deleted_p = 1;
4406 if (!insn_inserted_p)
4408 insert_insn_end_basic_block (index_map[i], bb, 0);
4409 insn_inserted_p = 1;
4415 VEC_free (basic_block, heap, domby);
4423 /* Top level routine to perform one code hoisting (aka unification) pass
4425 Return nonzero if a change was made. */
4428 one_code_hoisting_pass (void)
4432 gcse_subst_count = 0;
4433 gcse_create_count = 0;
4435 /* Return if there's nothing to do, or it is too expensive. */
4436 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
4437 || is_too_expensive (_("GCSE disabled")))
4440 /* We need alias. */
4441 init_alias_analysis ();
4444 gcc_obstack_init (&gcse_obstack);
4447 alloc_hash_table (&expr_hash_table, 0);
4448 compute_hash_table (&expr_hash_table);
4450 dump_hash_table (dump_file, "Code Hosting Expressions", &expr_hash_table);
4452 if (expr_hash_table.n_elems > 0)
4454 alloc_code_hoist_mem (last_basic_block, expr_hash_table.n_elems);
4455 compute_code_hoist_data ();
4456 changed = hoist_code ();
4457 free_code_hoist_mem ();
4460 free_hash_table (&expr_hash_table);
4462 obstack_free (&gcse_obstack, NULL);
4464 /* We are finished with alias. */
4465 end_alias_analysis ();
4469 fprintf (dump_file, "HOIST of %s, %d basic blocks, %d bytes needed, ",
4470 current_function_name (), n_basic_blocks, bytes_used);
4471 fprintf (dump_file, "%d substs, %d insns created\n",
4472 gcse_subst_count, gcse_create_count);
4478 /* Here we provide the things required to do store motion towards
4479 the exit. In order for this to be effective, gcse also needed to
4480 be taught how to move a load when it is kill only by a store to itself.
4485 void foo(float scale)
4487 for (i=0; i<10; i++)
4491 'i' is both loaded and stored to in the loop. Normally, gcse cannot move
4492 the load out since its live around the loop, and stored at the bottom
4495 The 'Load Motion' referred to and implemented in this file is
4496 an enhancement to gcse which when using edge based lcm, recognizes
4497 this situation and allows gcse to move the load out of the loop.
4499 Once gcse has hoisted the load, store motion can then push this
4500 load towards the exit, and we end up with no loads or stores of 'i'
4504 pre_ldst_expr_hash (const void *p)
4506 int do_not_record_p = 0;
4507 const struct ls_expr *const x = (const struct ls_expr *) p;
4508 return hash_rtx (x->pattern, GET_MODE (x->pattern), &do_not_record_p, NULL, false);
4512 pre_ldst_expr_eq (const void *p1, const void *p2)
4514 const struct ls_expr *const ptr1 = (const struct ls_expr *) p1,
4515 *const ptr2 = (const struct ls_expr *) p2;
4516 return expr_equiv_p (ptr1->pattern, ptr2->pattern);
4519 /* This will search the ldst list for a matching expression. If it
4520 doesn't find one, we create one and initialize it. */
4522 static struct ls_expr *
4525 int do_not_record_p = 0;
4526 struct ls_expr * ptr;
4531 hash = hash_rtx (x, GET_MODE (x), &do_not_record_p,
4532 NULL, /*have_reg_qty=*/false);
4535 slot = htab_find_slot_with_hash (pre_ldst_table, &e, hash, INSERT);
4537 return (struct ls_expr *)*slot;
4539 ptr = XNEW (struct ls_expr);
4541 ptr->next = pre_ldst_mems;
4544 ptr->pattern_regs = NULL_RTX;
4545 ptr->loads = NULL_RTX;
4546 ptr->stores = NULL_RTX;
4547 ptr->reaching_reg = NULL_RTX;
4550 ptr->hash_index = hash;
4551 pre_ldst_mems = ptr;
4557 /* Free up an individual ldst entry. */
4560 free_ldst_entry (struct ls_expr * ptr)
4562 free_INSN_LIST_list (& ptr->loads);
4563 free_INSN_LIST_list (& ptr->stores);
4568 /* Free up all memory associated with the ldst list. */
4571 free_ldst_mems (void)
4574 htab_delete (pre_ldst_table);
4575 pre_ldst_table = NULL;
4577 while (pre_ldst_mems)
4579 struct ls_expr * tmp = pre_ldst_mems;
4581 pre_ldst_mems = pre_ldst_mems->next;
4583 free_ldst_entry (tmp);
4586 pre_ldst_mems = NULL;
4589 /* Dump debugging info about the ldst list. */
4592 print_ldst_list (FILE * file)
4594 struct ls_expr * ptr;
4596 fprintf (file, "LDST list: \n");
4598 for (ptr = first_ls_expr (); ptr != NULL; ptr = next_ls_expr (ptr))
4600 fprintf (file, " Pattern (%3d): ", ptr->index);
4602 print_rtl (file, ptr->pattern);
4604 fprintf (file, "\n Loads : ");
4607 print_rtl (file, ptr->loads);
4609 fprintf (file, "(nil)");
4611 fprintf (file, "\n Stores : ");
4614 print_rtl (file, ptr->stores);
4616 fprintf (file, "(nil)");
4618 fprintf (file, "\n\n");
4621 fprintf (file, "\n");
4624 /* Returns 1 if X is in the list of ldst only expressions. */
4626 static struct ls_expr *
4627 find_rtx_in_ldst (rtx x)
4631 if (!pre_ldst_table)
4634 slot = htab_find_slot (pre_ldst_table, &e, NO_INSERT);
4635 if (!slot || ((struct ls_expr *)*slot)->invalid)
4637 return (struct ls_expr *) *slot;
4640 /* Return first item in the list. */
4642 static inline struct ls_expr *
4643 first_ls_expr (void)
4645 return pre_ldst_mems;
4648 /* Return the next item in the list after the specified one. */
4650 static inline struct ls_expr *
4651 next_ls_expr (struct ls_expr * ptr)
4656 /* Load Motion for loads which only kill themselves. */
4658 /* Return true if x is a simple MEM operation, with no registers or
4659 side effects. These are the types of loads we consider for the
4660 ld_motion list, otherwise we let the usual aliasing take care of it. */
4663 simple_mem (const_rtx x)
4668 if (MEM_VOLATILE_P (x))
4671 if (GET_MODE (x) == BLKmode)
4674 /* If we are handling exceptions, we must be careful with memory references
4675 that may trap. If we are not, the behavior is undefined, so we may just
4677 if (cfun->can_throw_non_call_exceptions && may_trap_p (x))
4680 if (side_effects_p (x))
4683 /* Do not consider function arguments passed on stack. */
4684 if (reg_mentioned_p (stack_pointer_rtx, x))
4687 if (flag_float_store && FLOAT_MODE_P (GET_MODE (x)))
4693 /* Make sure there isn't a buried reference in this pattern anywhere.
4694 If there is, invalidate the entry for it since we're not capable
4695 of fixing it up just yet.. We have to be sure we know about ALL
4696 loads since the aliasing code will allow all entries in the
4697 ld_motion list to not-alias itself. If we miss a load, we will get
4698 the wrong value since gcse might common it and we won't know to
4702 invalidate_any_buried_refs (rtx x)
4706 struct ls_expr * ptr;
4708 /* Invalidate it in the list. */
4709 if (MEM_P (x) && simple_mem (x))
4711 ptr = ldst_entry (x);
4715 /* Recursively process the insn. */
4716 fmt = GET_RTX_FORMAT (GET_CODE (x));
4718 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
4721 invalidate_any_buried_refs (XEXP (x, i));
4722 else if (fmt[i] == 'E')
4723 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4724 invalidate_any_buried_refs (XVECEXP (x, i, j));
4728 /* Find all the 'simple' MEMs which are used in LOADs and STORES. Simple
4729 being defined as MEM loads and stores to symbols, with no side effects
4730 and no registers in the expression. For a MEM destination, we also
4731 check that the insn is still valid if we replace the destination with a
4732 REG, as is done in update_ld_motion_stores. If there are any uses/defs
4733 which don't match this criteria, they are invalidated and trimmed out
4737 compute_ld_motion_mems (void)
4739 struct ls_expr * ptr;
4743 pre_ldst_mems = NULL;
4744 pre_ldst_table = htab_create (13, pre_ldst_expr_hash,
4745 pre_ldst_expr_eq, NULL);
4749 FOR_BB_INSNS (bb, insn)
4751 if (NONDEBUG_INSN_P (insn))
4753 if (GET_CODE (PATTERN (insn)) == SET)
4755 rtx src = SET_SRC (PATTERN (insn));
4756 rtx dest = SET_DEST (PATTERN (insn));
4758 /* Check for a simple LOAD... */
4759 if (MEM_P (src) && simple_mem (src))
4761 ptr = ldst_entry (src);
4763 ptr->loads = alloc_INSN_LIST (insn, ptr->loads);
4769 /* Make sure there isn't a buried load somewhere. */
4770 invalidate_any_buried_refs (src);
4773 /* Check for stores. Don't worry about aliased ones, they
4774 will block any movement we might do later. We only care
4775 about this exact pattern since those are the only
4776 circumstance that we will ignore the aliasing info. */
4777 if (MEM_P (dest) && simple_mem (dest))
4779 ptr = ldst_entry (dest);
4782 && GET_CODE (src) != ASM_OPERANDS
4783 /* Check for REG manually since want_to_gcse_p
4784 returns 0 for all REGs. */
4785 && can_assign_to_reg_without_clobbers_p (src))
4786 ptr->stores = alloc_INSN_LIST (insn, ptr->stores);
4792 invalidate_any_buried_refs (PATTERN (insn));
4798 /* Remove any references that have been either invalidated or are not in the
4799 expression list for pre gcse. */
4802 trim_ld_motion_mems (void)
4804 struct ls_expr * * last = & pre_ldst_mems;
4805 struct ls_expr * ptr = pre_ldst_mems;
4811 /* Delete if entry has been made invalid. */
4814 /* Delete if we cannot find this mem in the expression list. */
4815 unsigned int hash = ptr->hash_index % expr_hash_table.size;
4817 for (expr = expr_hash_table.table[hash];
4819 expr = expr->next_same_hash)
4820 if (expr_equiv_p (expr->expr, ptr->pattern))
4824 expr = (struct expr *) 0;
4828 /* Set the expression field if we are keeping it. */
4836 htab_remove_elt_with_hash (pre_ldst_table, ptr, ptr->hash_index);
4837 free_ldst_entry (ptr);
4842 /* Show the world what we've found. */
4843 if (dump_file && pre_ldst_mems != NULL)
4844 print_ldst_list (dump_file);
4847 /* This routine will take an expression which we are replacing with
4848 a reaching register, and update any stores that are needed if
4849 that expression is in the ld_motion list. Stores are updated by
4850 copying their SRC to the reaching register, and then storing
4851 the reaching register into the store location. These keeps the
4852 correct value in the reaching register for the loads. */
4855 update_ld_motion_stores (struct expr * expr)
4857 struct ls_expr * mem_ptr;
4859 if ((mem_ptr = find_rtx_in_ldst (expr->expr)))
4861 /* We can try to find just the REACHED stores, but is shouldn't
4862 matter to set the reaching reg everywhere... some might be
4863 dead and should be eliminated later. */
4865 /* We replace (set mem expr) with (set reg expr) (set mem reg)
4866 where reg is the reaching reg used in the load. We checked in
4867 compute_ld_motion_mems that we can replace (set mem expr) with
4868 (set reg expr) in that insn. */
4869 rtx list = mem_ptr->stores;
4871 for ( ; list != NULL_RTX; list = XEXP (list, 1))
4873 rtx insn = XEXP (list, 0);
4874 rtx pat = PATTERN (insn);
4875 rtx src = SET_SRC (pat);
4876 rtx reg = expr->reaching_reg;
4879 /* If we've already copied it, continue. */
4880 if (expr->reaching_reg == src)
4885 fprintf (dump_file, "PRE: store updated with reaching reg ");
4886 print_rtl (dump_file, expr->reaching_reg);
4887 fprintf (dump_file, ":\n ");
4888 print_inline_rtx (dump_file, insn, 8);
4889 fprintf (dump_file, "\n");
4892 copy = gen_move_insn (reg, copy_rtx (SET_SRC (pat)));
4893 emit_insn_before (copy, insn);
4894 SET_SRC (pat) = reg;
4895 df_insn_rescan (insn);
4897 /* un-recognize this pattern since it's probably different now. */
4898 INSN_CODE (insn) = -1;
4899 gcse_create_count++;
4904 /* Return true if the graph is too expensive to optimize. PASS is the
4905 optimization about to be performed. */
4908 is_too_expensive (const char *pass)
4910 /* Trying to perform global optimizations on flow graphs which have
4911 a high connectivity will take a long time and is unlikely to be
4912 particularly useful.
4914 In normal circumstances a cfg should have about twice as many
4915 edges as blocks. But we do not want to punish small functions
4916 which have a couple switch statements. Rather than simply
4917 threshold the number of blocks, uses something with a more
4918 graceful degradation. */
4919 if (n_edges > 20000 + n_basic_blocks * 4)
4921 warning (OPT_Wdisabled_optimization,
4922 "%s: %d basic blocks and %d edges/basic block",
4923 pass, n_basic_blocks, n_edges / n_basic_blocks);
4928 /* If allocating memory for the cprop bitmap would take up too much
4929 storage it's better just to disable the optimization. */
4931 * SBITMAP_SET_SIZE (max_reg_num ())
4932 * sizeof (SBITMAP_ELT_TYPE)) > MAX_GCSE_MEMORY)
4934 warning (OPT_Wdisabled_optimization,
4935 "%s: %d basic blocks and %d registers",
4936 pass, n_basic_blocks, max_reg_num ());
4945 /* Main function for the CPROP pass. */
4948 one_cprop_pass (void)
4952 /* Return if there's nothing to do, or it is too expensive. */
4953 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1
4954 || is_too_expensive (_ ("const/copy propagation disabled")))
4957 global_const_prop_count = local_const_prop_count = 0;
4958 global_copy_prop_count = local_copy_prop_count = 0;
4961 gcc_obstack_init (&gcse_obstack);
4964 /* Do a local const/copy propagation pass first. The global pass
4965 only handles global opportunities.
4966 If the local pass changes something, remove any unreachable blocks
4967 because the CPROP global dataflow analysis may get into infinite
4968 loops for CFGs with unreachable blocks.
4970 FIXME: This local pass should not be necessary after CSE (but for
4971 some reason it still is). It is also (proven) not necessary
4972 to run the local pass right after FWPWOP.
4974 FIXME: The global analysis would not get into infinite loops if it
4975 would use the DF solver (via df_simple_dataflow) instead of
4976 the solver implemented in this file. */
4977 if (local_cprop_pass ())
4979 delete_unreachable_blocks ();
4983 /* Determine implicit sets. */
4984 implicit_sets = XCNEWVEC (rtx, last_basic_block);
4985 find_implicit_sets ();
4987 alloc_hash_table (&set_hash_table, 1);
4988 compute_hash_table (&set_hash_table);
4990 /* Free implicit_sets before peak usage. */
4991 free (implicit_sets);
4992 implicit_sets = NULL;
4995 dump_hash_table (dump_file, "SET", &set_hash_table);
4996 if (set_hash_table.n_elems > 0)
5001 alloc_cprop_mem (last_basic_block, set_hash_table.n_elems);
5002 compute_cprop_data ();
5004 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR->next_bb->next_bb, EXIT_BLOCK_PTR, next_bb)
5006 /* Reset tables used to keep track of what's still valid [since
5007 the start of the block]. */
5008 reset_opr_set_tables ();
5010 FOR_BB_INSNS (bb, insn)
5013 changed |= cprop_insn (insn);
5015 /* Keep track of everything modified by this insn. */
5016 /* ??? Need to be careful w.r.t. mods done to INSN.
5017 Don't call mark_oprs_set if we turned the
5018 insn into a NOTE. */
5019 if (! NOTE_P (insn))
5020 mark_oprs_set (insn);
5024 changed |= bypass_conditional_jumps ();
5028 free_hash_table (&set_hash_table);
5030 obstack_free (&gcse_obstack, NULL);
5034 fprintf (dump_file, "CPROP of %s, %d basic blocks, %d bytes needed, ",
5035 current_function_name (), n_basic_blocks, bytes_used);
5036 fprintf (dump_file, "%d local const props, %d local copy props, ",
5037 local_const_prop_count, local_copy_prop_count);
5038 fprintf (dump_file, "%d global const props, %d global copy props\n\n",
5039 global_const_prop_count, global_copy_prop_count);
5046 /* All the passes implemented in this file. Each pass has its
5047 own gate and execute function, and at the end of the file a
5048 pass definition for passes.c.
5050 We do not construct an accurate cfg in functions which call
5051 setjmp, so none of these passes runs if the function calls
5053 FIXME: Should just handle setjmp via REG_SETJMP notes. */
5056 gate_rtl_cprop (void)
5058 return optimize > 0 && flag_gcse
5059 && !cfun->calls_setjmp
5064 execute_rtl_cprop (void)
5066 delete_unreachable_blocks ();
5067 df_set_flags (DF_LR_RUN_DCE);
5069 flag_rerun_cse_after_global_opts |= one_cprop_pass ();
5076 return optimize > 0 && flag_gcse
5077 && !cfun->calls_setjmp
5078 && optimize_function_for_speed_p (cfun)
5083 execute_rtl_pre (void)
5085 delete_unreachable_blocks ();
5087 flag_rerun_cse_after_global_opts |= one_pre_gcse_pass ();
5092 gate_rtl_hoist (void)
5094 return optimize > 0 && flag_gcse
5095 && !cfun->calls_setjmp
5096 /* It does not make sense to run code hoisting unless we are optimizing
5097 for code size -- it rarely makes programs faster, and can make then
5098 bigger if we did PRE (when optimizing for space, we don't run PRE). */
5099 && optimize_function_for_size_p (cfun)
5104 execute_rtl_hoist (void)
5106 delete_unreachable_blocks ();
5108 flag_rerun_cse_after_global_opts |= one_code_hoisting_pass ();
5112 struct rtl_opt_pass pass_rtl_cprop =
5117 gate_rtl_cprop, /* gate */
5118 execute_rtl_cprop, /* execute */
5121 0, /* static_pass_number */
5122 TV_CPROP, /* tv_id */
5123 PROP_cfglayout, /* properties_required */
5124 0, /* properties_provided */
5125 0, /* properties_destroyed */
5126 0, /* todo_flags_start */
5127 TODO_df_finish | TODO_verify_rtl_sharing |
5129 TODO_verify_flow | TODO_ggc_collect /* todo_flags_finish */
5133 struct rtl_opt_pass pass_rtl_pre =
5137 "rtl pre", /* name */
5138 gate_rtl_pre, /* gate */
5139 execute_rtl_pre, /* execute */
5142 0, /* static_pass_number */
5144 PROP_cfglayout, /* properties_required */
5145 0, /* properties_provided */
5146 0, /* properties_destroyed */
5147 0, /* todo_flags_start */
5148 TODO_df_finish | TODO_verify_rtl_sharing |
5150 TODO_verify_flow | TODO_ggc_collect /* todo_flags_finish */
5154 struct rtl_opt_pass pass_rtl_hoist =
5159 gate_rtl_hoist, /* gate */
5160 execute_rtl_hoist, /* execute */
5163 0, /* static_pass_number */
5164 TV_HOIST, /* tv_id */
5165 PROP_cfglayout, /* properties_required */
5166 0, /* properties_provided */
5167 0, /* properties_destroyed */
5168 0, /* todo_flags_start */
5169 TODO_df_finish | TODO_verify_rtl_sharing |
5171 TODO_verify_flow | TODO_ggc_collect /* todo_flags_finish */
5175 #include "gt-gcse.h"