2010-07-27 Paolo Carlini <paolo.carlini@oracle.com>
[official-gcc/alias-decl.git] / gcc / tree-ssa-reassoc.c
blob6cd3cebfe86559f59bc25d3adc022f326aab8b2e
1 /* Reassociation for trees.
2 Copyright (C) 2005, 2007, 2008, 2009, 2010 Free Software Foundation, Inc.
3 Contributed by Daniel Berlin <dan@dberlin.org>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify
8 it under the terms of the GNU General Public License as published by
9 the Free Software Foundation; either version 3, or (at your option)
10 any later version.
12 GCC is distributed in the hope that it will be useful,
13 but WITHOUT ANY WARRANTY; without even the implied warranty of
14 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
15 GNU General Public License for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "tree.h"
26 #include "basic-block.h"
27 #include "tree-pretty-print.h"
28 #include "gimple-pretty-print.h"
29 #include "tree-inline.h"
30 #include "tree-flow.h"
31 #include "gimple.h"
32 #include "tree-dump.h"
33 #include "timevar.h"
34 #include "tree-iterator.h"
35 #include "tree-pass.h"
36 #include "alloc-pool.h"
37 #include "vec.h"
38 #include "langhooks.h"
39 #include "pointer-set.h"
40 #include "cfgloop.h"
41 #include "flags.h"
43 /* This is a simple global reassociation pass. It is, in part, based
44 on the LLVM pass of the same name (They do some things more/less
45 than we do, in different orders, etc).
47 It consists of five steps:
49 1. Breaking up subtract operations into addition + negate, where
50 it would promote the reassociation of adds.
52 2. Left linearization of the expression trees, so that (A+B)+(C+D)
53 becomes (((A+B)+C)+D), which is easier for us to rewrite later.
54 During linearization, we place the operands of the binary
55 expressions into a vector of operand_entry_t
57 3. Optimization of the operand lists, eliminating things like a +
58 -a, a & a, etc.
60 4. Rewrite the expression trees we linearized and optimized so
61 they are in proper rank order.
63 5. Repropagate negates, as nothing else will clean it up ATM.
65 A bit of theory on #4, since nobody seems to write anything down
66 about why it makes sense to do it the way they do it:
68 We could do this much nicer theoretically, but don't (for reasons
69 explained after how to do it theoretically nice :P).
71 In order to promote the most redundancy elimination, you want
72 binary expressions whose operands are the same rank (or
73 preferably, the same value) exposed to the redundancy eliminator,
74 for possible elimination.
76 So the way to do this if we really cared, is to build the new op
77 tree from the leaves to the roots, merging as you go, and putting the
78 new op on the end of the worklist, until you are left with one
79 thing on the worklist.
81 IE if you have to rewrite the following set of operands (listed with
82 rank in parentheses), with opcode PLUS_EXPR:
84 a (1), b (1), c (1), d (2), e (2)
87 We start with our merge worklist empty, and the ops list with all of
88 those on it.
90 You want to first merge all leaves of the same rank, as much as
91 possible.
93 So first build a binary op of
95 mergetmp = a + b, and put "mergetmp" on the merge worklist.
97 Because there is no three operand form of PLUS_EXPR, c is not going to
98 be exposed to redundancy elimination as a rank 1 operand.
100 So you might as well throw it on the merge worklist (you could also
101 consider it to now be a rank two operand, and merge it with d and e,
102 but in this case, you then have evicted e from a binary op. So at
103 least in this situation, you can't win.)
105 Then build a binary op of d + e
106 mergetmp2 = d + e
108 and put mergetmp2 on the merge worklist.
110 so merge worklist = {mergetmp, c, mergetmp2}
112 Continue building binary ops of these operations until you have only
113 one operation left on the worklist.
115 So we have
117 build binary op
118 mergetmp3 = mergetmp + c
120 worklist = {mergetmp2, mergetmp3}
122 mergetmp4 = mergetmp2 + mergetmp3
124 worklist = {mergetmp4}
126 because we have one operation left, we can now just set the original
127 statement equal to the result of that operation.
129 This will at least expose a + b and d + e to redundancy elimination
130 as binary operations.
132 For extra points, you can reuse the old statements to build the
133 mergetmps, since you shouldn't run out.
135 So why don't we do this?
137 Because it's expensive, and rarely will help. Most trees we are
138 reassociating have 3 or less ops. If they have 2 ops, they already
139 will be written into a nice single binary op. If you have 3 ops, a
140 single simple check suffices to tell you whether the first two are of the
141 same rank. If so, you know to order it
143 mergetmp = op1 + op2
144 newstmt = mergetmp + op3
146 instead of
147 mergetmp = op2 + op3
148 newstmt = mergetmp + op1
150 If all three are of the same rank, you can't expose them all in a
151 single binary operator anyway, so the above is *still* the best you
152 can do.
154 Thus, this is what we do. When we have three ops left, we check to see
155 what order to put them in, and call it a day. As a nod to vector sum
156 reduction, we check if any of the ops are really a phi node that is a
157 destructive update for the associating op, and keep the destructive
158 update together for vector sum reduction recognition. */
161 /* Statistics */
162 static struct
164 int linearized;
165 int constants_eliminated;
166 int ops_eliminated;
167 int rewritten;
168 } reassociate_stats;
170 /* Operator, rank pair. */
171 typedef struct operand_entry
173 unsigned int rank;
174 int id;
175 tree op;
176 } *operand_entry_t;
178 static alloc_pool operand_entry_pool;
180 /* This is used to assign a unique ID to each struct operand_entry
181 so that qsort results are identical on different hosts. */
182 static int next_operand_entry_id;
184 /* Starting rank number for a given basic block, so that we can rank
185 operations using unmovable instructions in that BB based on the bb
186 depth. */
187 static long *bb_rank;
189 /* Operand->rank hashtable. */
190 static struct pointer_map_t *operand_rank;
193 /* Look up the operand rank structure for expression E. */
195 static inline long
196 find_operand_rank (tree e)
198 void **slot = pointer_map_contains (operand_rank, e);
199 return slot ? (long) (intptr_t) *slot : -1;
202 /* Insert {E,RANK} into the operand rank hashtable. */
204 static inline void
205 insert_operand_rank (tree e, long rank)
207 void **slot;
208 gcc_assert (rank > 0);
209 slot = pointer_map_insert (operand_rank, e);
210 gcc_assert (!*slot);
211 *slot = (void *) (intptr_t) rank;
214 /* Given an expression E, return the rank of the expression. */
216 static long
217 get_rank (tree e)
219 /* Constants have rank 0. */
220 if (is_gimple_min_invariant (e))
221 return 0;
223 /* SSA_NAME's have the rank of the expression they are the result
225 For globals and uninitialized values, the rank is 0.
226 For function arguments, use the pre-setup rank.
227 For PHI nodes, stores, asm statements, etc, we use the rank of
228 the BB.
229 For simple operations, the rank is the maximum rank of any of
230 its operands, or the bb_rank, whichever is less.
231 I make no claims that this is optimal, however, it gives good
232 results. */
234 if (TREE_CODE (e) == SSA_NAME)
236 gimple stmt;
237 long rank, maxrank;
238 int i, n;
240 if (TREE_CODE (SSA_NAME_VAR (e)) == PARM_DECL
241 && SSA_NAME_IS_DEFAULT_DEF (e))
242 return find_operand_rank (e);
244 stmt = SSA_NAME_DEF_STMT (e);
245 if (gimple_bb (stmt) == NULL)
246 return 0;
248 if (!is_gimple_assign (stmt)
249 || gimple_vdef (stmt))
250 return bb_rank[gimple_bb (stmt)->index];
252 /* If we already have a rank for this expression, use that. */
253 rank = find_operand_rank (e);
254 if (rank != -1)
255 return rank;
257 /* Otherwise, find the maximum rank for the operands, or the bb
258 rank, whichever is less. */
259 rank = 0;
260 maxrank = bb_rank[gimple_bb(stmt)->index];
261 if (gimple_assign_single_p (stmt))
263 tree rhs = gimple_assign_rhs1 (stmt);
264 n = TREE_OPERAND_LENGTH (rhs);
265 if (n == 0)
266 rank = MAX (rank, get_rank (rhs));
267 else
269 for (i = 0;
270 i < n && TREE_OPERAND (rhs, i) && rank != maxrank; i++)
271 rank = MAX(rank, get_rank (TREE_OPERAND (rhs, i)));
274 else
276 n = gimple_num_ops (stmt);
277 for (i = 1; i < n && rank != maxrank; i++)
279 gcc_assert (gimple_op (stmt, i));
280 rank = MAX(rank, get_rank (gimple_op (stmt, i)));
284 if (dump_file && (dump_flags & TDF_DETAILS))
286 fprintf (dump_file, "Rank for ");
287 print_generic_expr (dump_file, e, 0);
288 fprintf (dump_file, " is %ld\n", (rank + 1));
291 /* Note the rank in the hashtable so we don't recompute it. */
292 insert_operand_rank (e, (rank + 1));
293 return (rank + 1);
296 /* Globals, etc, are rank 0 */
297 return 0;
300 DEF_VEC_P(operand_entry_t);
301 DEF_VEC_ALLOC_P(operand_entry_t, heap);
303 /* We want integer ones to end up last no matter what, since they are
304 the ones we can do the most with. */
305 #define INTEGER_CONST_TYPE 1 << 3
306 #define FLOAT_CONST_TYPE 1 << 2
307 #define OTHER_CONST_TYPE 1 << 1
309 /* Classify an invariant tree into integer, float, or other, so that
310 we can sort them to be near other constants of the same type. */
311 static inline int
312 constant_type (tree t)
314 if (INTEGRAL_TYPE_P (TREE_TYPE (t)))
315 return INTEGER_CONST_TYPE;
316 else if (SCALAR_FLOAT_TYPE_P (TREE_TYPE (t)))
317 return FLOAT_CONST_TYPE;
318 else
319 return OTHER_CONST_TYPE;
322 /* qsort comparison function to sort operand entries PA and PB by rank
323 so that the sorted array is ordered by rank in decreasing order. */
324 static int
325 sort_by_operand_rank (const void *pa, const void *pb)
327 const operand_entry_t oea = *(const operand_entry_t *)pa;
328 const operand_entry_t oeb = *(const operand_entry_t *)pb;
330 /* It's nicer for optimize_expression if constants that are likely
331 to fold when added/multiplied//whatever are put next to each
332 other. Since all constants have rank 0, order them by type. */
333 if (oeb->rank == 0 && oea->rank == 0)
335 if (constant_type (oeb->op) != constant_type (oea->op))
336 return constant_type (oeb->op) - constant_type (oea->op);
337 else
338 /* To make sorting result stable, we use unique IDs to determine
339 order. */
340 return oeb->id - oea->id;
343 /* Lastly, make sure the versions that are the same go next to each
344 other. We use SSA_NAME_VERSION because it's stable. */
345 if ((oeb->rank - oea->rank == 0)
346 && TREE_CODE (oea->op) == SSA_NAME
347 && TREE_CODE (oeb->op) == SSA_NAME)
349 if (SSA_NAME_VERSION (oeb->op) != SSA_NAME_VERSION (oea->op))
350 return SSA_NAME_VERSION (oeb->op) - SSA_NAME_VERSION (oea->op);
351 else
352 return oeb->id - oea->id;
355 if (oeb->rank != oea->rank)
356 return oeb->rank - oea->rank;
357 else
358 return oeb->id - oea->id;
361 /* Add an operand entry to *OPS for the tree operand OP. */
363 static void
364 add_to_ops_vec (VEC(operand_entry_t, heap) **ops, tree op)
366 operand_entry_t oe = (operand_entry_t) pool_alloc (operand_entry_pool);
368 oe->op = op;
369 oe->rank = get_rank (op);
370 oe->id = next_operand_entry_id++;
371 VEC_safe_push (operand_entry_t, heap, *ops, oe);
374 /* Return true if STMT is reassociable operation containing a binary
375 operation with tree code CODE, and is inside LOOP. */
377 static bool
378 is_reassociable_op (gimple stmt, enum tree_code code, struct loop *loop)
380 basic_block bb = gimple_bb (stmt);
382 if (gimple_bb (stmt) == NULL)
383 return false;
385 if (!flow_bb_inside_loop_p (loop, bb))
386 return false;
388 if (is_gimple_assign (stmt)
389 && gimple_assign_rhs_code (stmt) == code
390 && has_single_use (gimple_assign_lhs (stmt)))
391 return true;
393 return false;
397 /* Given NAME, if NAME is defined by a unary operation OPCODE, return the
398 operand of the negate operation. Otherwise, return NULL. */
400 static tree
401 get_unary_op (tree name, enum tree_code opcode)
403 gimple stmt = SSA_NAME_DEF_STMT (name);
405 if (!is_gimple_assign (stmt))
406 return NULL_TREE;
408 if (gimple_assign_rhs_code (stmt) == opcode)
409 return gimple_assign_rhs1 (stmt);
410 return NULL_TREE;
413 /* If CURR and LAST are a pair of ops that OPCODE allows us to
414 eliminate through equivalences, do so, remove them from OPS, and
415 return true. Otherwise, return false. */
417 static bool
418 eliminate_duplicate_pair (enum tree_code opcode,
419 VEC (operand_entry_t, heap) **ops,
420 bool *all_done,
421 unsigned int i,
422 operand_entry_t curr,
423 operand_entry_t last)
426 /* If we have two of the same op, and the opcode is & |, min, or max,
427 we can eliminate one of them.
428 If we have two of the same op, and the opcode is ^, we can
429 eliminate both of them. */
431 if (last && last->op == curr->op)
433 switch (opcode)
435 case MAX_EXPR:
436 case MIN_EXPR:
437 case BIT_IOR_EXPR:
438 case BIT_AND_EXPR:
439 if (dump_file && (dump_flags & TDF_DETAILS))
441 fprintf (dump_file, "Equivalence: ");
442 print_generic_expr (dump_file, curr->op, 0);
443 fprintf (dump_file, " [&|minmax] ");
444 print_generic_expr (dump_file, last->op, 0);
445 fprintf (dump_file, " -> ");
446 print_generic_stmt (dump_file, last->op, 0);
449 VEC_ordered_remove (operand_entry_t, *ops, i);
450 reassociate_stats.ops_eliminated ++;
452 return true;
454 case BIT_XOR_EXPR:
455 if (dump_file && (dump_flags & TDF_DETAILS))
457 fprintf (dump_file, "Equivalence: ");
458 print_generic_expr (dump_file, curr->op, 0);
459 fprintf (dump_file, " ^ ");
460 print_generic_expr (dump_file, last->op, 0);
461 fprintf (dump_file, " -> nothing\n");
464 reassociate_stats.ops_eliminated += 2;
466 if (VEC_length (operand_entry_t, *ops) == 2)
468 VEC_free (operand_entry_t, heap, *ops);
469 *ops = NULL;
470 add_to_ops_vec (ops, fold_convert (TREE_TYPE (last->op),
471 integer_zero_node));
472 *all_done = true;
474 else
476 VEC_ordered_remove (operand_entry_t, *ops, i-1);
477 VEC_ordered_remove (operand_entry_t, *ops, i-1);
480 return true;
482 default:
483 break;
486 return false;
489 static VEC(tree, heap) *plus_negates;
491 /* If OPCODE is PLUS_EXPR, CURR->OP is a negate expression or a bitwise not
492 expression, look in OPS for a corresponding positive operation to cancel
493 it out. If we find one, remove the other from OPS, replace
494 OPS[CURRINDEX] with 0 or -1, respectively, and return true. Otherwise,
495 return false. */
497 static bool
498 eliminate_plus_minus_pair (enum tree_code opcode,
499 VEC (operand_entry_t, heap) **ops,
500 unsigned int currindex,
501 operand_entry_t curr)
503 tree negateop;
504 tree notop;
505 unsigned int i;
506 operand_entry_t oe;
508 if (opcode != PLUS_EXPR || TREE_CODE (curr->op) != SSA_NAME)
509 return false;
511 negateop = get_unary_op (curr->op, NEGATE_EXPR);
512 notop = get_unary_op (curr->op, BIT_NOT_EXPR);
513 if (negateop == NULL_TREE && notop == NULL_TREE)
514 return false;
516 /* Any non-negated version will have a rank that is one less than
517 the current rank. So once we hit those ranks, if we don't find
518 one, we can stop. */
520 for (i = currindex + 1;
521 VEC_iterate (operand_entry_t, *ops, i, oe)
522 && oe->rank >= curr->rank - 1 ;
523 i++)
525 if (oe->op == negateop)
528 if (dump_file && (dump_flags & TDF_DETAILS))
530 fprintf (dump_file, "Equivalence: ");
531 print_generic_expr (dump_file, negateop, 0);
532 fprintf (dump_file, " + -");
533 print_generic_expr (dump_file, oe->op, 0);
534 fprintf (dump_file, " -> 0\n");
537 VEC_ordered_remove (operand_entry_t, *ops, i);
538 add_to_ops_vec (ops, fold_convert(TREE_TYPE (oe->op),
539 integer_zero_node));
540 VEC_ordered_remove (operand_entry_t, *ops, currindex);
541 reassociate_stats.ops_eliminated ++;
543 return true;
545 else if (oe->op == notop)
547 tree op_type = TREE_TYPE (oe->op);
549 if (dump_file && (dump_flags & TDF_DETAILS))
551 fprintf (dump_file, "Equivalence: ");
552 print_generic_expr (dump_file, notop, 0);
553 fprintf (dump_file, " + ~");
554 print_generic_expr (dump_file, oe->op, 0);
555 fprintf (dump_file, " -> -1\n");
558 VEC_ordered_remove (operand_entry_t, *ops, i);
559 add_to_ops_vec (ops, build_int_cst_type (op_type, -1));
560 VEC_ordered_remove (operand_entry_t, *ops, currindex);
561 reassociate_stats.ops_eliminated ++;
563 return true;
567 /* CURR->OP is a negate expr in a plus expr: save it for later
568 inspection in repropagate_negates(). */
569 if (negateop != NULL_TREE)
570 VEC_safe_push (tree, heap, plus_negates, curr->op);
572 return false;
575 /* If OPCODE is BIT_IOR_EXPR, BIT_AND_EXPR, and, CURR->OP is really a
576 bitwise not expression, look in OPS for a corresponding operand to
577 cancel it out. If we find one, remove the other from OPS, replace
578 OPS[CURRINDEX] with 0, and return true. Otherwise, return
579 false. */
581 static bool
582 eliminate_not_pairs (enum tree_code opcode,
583 VEC (operand_entry_t, heap) **ops,
584 unsigned int currindex,
585 operand_entry_t curr)
587 tree notop;
588 unsigned int i;
589 operand_entry_t oe;
591 if ((opcode != BIT_IOR_EXPR && opcode != BIT_AND_EXPR)
592 || TREE_CODE (curr->op) != SSA_NAME)
593 return false;
595 notop = get_unary_op (curr->op, BIT_NOT_EXPR);
596 if (notop == NULL_TREE)
597 return false;
599 /* Any non-not version will have a rank that is one less than
600 the current rank. So once we hit those ranks, if we don't find
601 one, we can stop. */
603 for (i = currindex + 1;
604 VEC_iterate (operand_entry_t, *ops, i, oe)
605 && oe->rank >= curr->rank - 1;
606 i++)
608 if (oe->op == notop)
610 if (dump_file && (dump_flags & TDF_DETAILS))
612 fprintf (dump_file, "Equivalence: ");
613 print_generic_expr (dump_file, notop, 0);
614 if (opcode == BIT_AND_EXPR)
615 fprintf (dump_file, " & ~");
616 else if (opcode == BIT_IOR_EXPR)
617 fprintf (dump_file, " | ~");
618 print_generic_expr (dump_file, oe->op, 0);
619 if (opcode == BIT_AND_EXPR)
620 fprintf (dump_file, " -> 0\n");
621 else if (opcode == BIT_IOR_EXPR)
622 fprintf (dump_file, " -> -1\n");
625 if (opcode == BIT_AND_EXPR)
626 oe->op = fold_convert (TREE_TYPE (oe->op), integer_zero_node);
627 else if (opcode == BIT_IOR_EXPR)
628 oe->op = build_low_bits_mask (TREE_TYPE (oe->op),
629 TYPE_PRECISION (TREE_TYPE (oe->op)));
631 reassociate_stats.ops_eliminated
632 += VEC_length (operand_entry_t, *ops) - 1;
633 VEC_free (operand_entry_t, heap, *ops);
634 *ops = NULL;
635 VEC_safe_push (operand_entry_t, heap, *ops, oe);
636 return true;
640 return false;
643 /* Use constant value that may be present in OPS to try to eliminate
644 operands. Note that this function is only really used when we've
645 eliminated ops for other reasons, or merged constants. Across
646 single statements, fold already does all of this, plus more. There
647 is little point in duplicating logic, so I've only included the
648 identities that I could ever construct testcases to trigger. */
650 static void
651 eliminate_using_constants (enum tree_code opcode,
652 VEC(operand_entry_t, heap) **ops)
654 operand_entry_t oelast = VEC_last (operand_entry_t, *ops);
655 tree type = TREE_TYPE (oelast->op);
657 if (oelast->rank == 0
658 && (INTEGRAL_TYPE_P (type) || FLOAT_TYPE_P (type)))
660 switch (opcode)
662 case BIT_AND_EXPR:
663 if (integer_zerop (oelast->op))
665 if (VEC_length (operand_entry_t, *ops) != 1)
667 if (dump_file && (dump_flags & TDF_DETAILS))
668 fprintf (dump_file, "Found & 0, removing all other ops\n");
670 reassociate_stats.ops_eliminated
671 += VEC_length (operand_entry_t, *ops) - 1;
673 VEC_free (operand_entry_t, heap, *ops);
674 *ops = NULL;
675 VEC_safe_push (operand_entry_t, heap, *ops, oelast);
676 return;
679 else if (integer_all_onesp (oelast->op))
681 if (VEC_length (operand_entry_t, *ops) != 1)
683 if (dump_file && (dump_flags & TDF_DETAILS))
684 fprintf (dump_file, "Found & -1, removing\n");
685 VEC_pop (operand_entry_t, *ops);
686 reassociate_stats.ops_eliminated++;
689 break;
690 case BIT_IOR_EXPR:
691 if (integer_all_onesp (oelast->op))
693 if (VEC_length (operand_entry_t, *ops) != 1)
695 if (dump_file && (dump_flags & TDF_DETAILS))
696 fprintf (dump_file, "Found | -1, removing all other ops\n");
698 reassociate_stats.ops_eliminated
699 += VEC_length (operand_entry_t, *ops) - 1;
701 VEC_free (operand_entry_t, heap, *ops);
702 *ops = NULL;
703 VEC_safe_push (operand_entry_t, heap, *ops, oelast);
704 return;
707 else if (integer_zerop (oelast->op))
709 if (VEC_length (operand_entry_t, *ops) != 1)
711 if (dump_file && (dump_flags & TDF_DETAILS))
712 fprintf (dump_file, "Found | 0, removing\n");
713 VEC_pop (operand_entry_t, *ops);
714 reassociate_stats.ops_eliminated++;
717 break;
718 case MULT_EXPR:
719 if (integer_zerop (oelast->op)
720 || (FLOAT_TYPE_P (type)
721 && !HONOR_NANS (TYPE_MODE (type))
722 && !HONOR_SIGNED_ZEROS (TYPE_MODE (type))
723 && real_zerop (oelast->op)))
725 if (VEC_length (operand_entry_t, *ops) != 1)
727 if (dump_file && (dump_flags & TDF_DETAILS))
728 fprintf (dump_file, "Found * 0, removing all other ops\n");
730 reassociate_stats.ops_eliminated
731 += VEC_length (operand_entry_t, *ops) - 1;
732 VEC_free (operand_entry_t, heap, *ops);
733 *ops = NULL;
734 VEC_safe_push (operand_entry_t, heap, *ops, oelast);
735 return;
738 else if (integer_onep (oelast->op)
739 || (FLOAT_TYPE_P (type)
740 && !HONOR_SNANS (TYPE_MODE (type))
741 && real_onep (oelast->op)))
743 if (VEC_length (operand_entry_t, *ops) != 1)
745 if (dump_file && (dump_flags & TDF_DETAILS))
746 fprintf (dump_file, "Found * 1, removing\n");
747 VEC_pop (operand_entry_t, *ops);
748 reassociate_stats.ops_eliminated++;
749 return;
752 break;
753 case BIT_XOR_EXPR:
754 case PLUS_EXPR:
755 case MINUS_EXPR:
756 if (integer_zerop (oelast->op)
757 || (FLOAT_TYPE_P (type)
758 && (opcode == PLUS_EXPR || opcode == MINUS_EXPR)
759 && fold_real_zero_addition_p (type, oelast->op,
760 opcode == MINUS_EXPR)))
762 if (VEC_length (operand_entry_t, *ops) != 1)
764 if (dump_file && (dump_flags & TDF_DETAILS))
765 fprintf (dump_file, "Found [|^+] 0, removing\n");
766 VEC_pop (operand_entry_t, *ops);
767 reassociate_stats.ops_eliminated++;
768 return;
771 break;
772 default:
773 break;
779 static void linearize_expr_tree (VEC(operand_entry_t, heap) **, gimple,
780 bool, bool);
782 /* Structure for tracking and counting operands. */
783 typedef struct oecount_s {
784 int cnt;
785 int id;
786 enum tree_code oecode;
787 tree op;
788 } oecount;
790 DEF_VEC_O(oecount);
791 DEF_VEC_ALLOC_O(oecount,heap);
793 /* The heap for the oecount hashtable and the sorted list of operands. */
794 static VEC (oecount, heap) *cvec;
796 /* Hash function for oecount. */
798 static hashval_t
799 oecount_hash (const void *p)
801 const oecount *c = VEC_index (oecount, cvec, (size_t)p - 42);
802 return htab_hash_pointer (c->op) ^ (hashval_t)c->oecode;
805 /* Comparison function for oecount. */
807 static int
808 oecount_eq (const void *p1, const void *p2)
810 const oecount *c1 = VEC_index (oecount, cvec, (size_t)p1 - 42);
811 const oecount *c2 = VEC_index (oecount, cvec, (size_t)p2 - 42);
812 return (c1->oecode == c2->oecode
813 && c1->op == c2->op);
816 /* Comparison function for qsort sorting oecount elements by count. */
818 static int
819 oecount_cmp (const void *p1, const void *p2)
821 const oecount *c1 = (const oecount *)p1;
822 const oecount *c2 = (const oecount *)p2;
823 if (c1->cnt != c2->cnt)
824 return c1->cnt - c2->cnt;
825 else
826 /* If counts are identical, use unique IDs to stabilize qsort. */
827 return c1->id - c2->id;
830 /* Walks the linear chain with result *DEF searching for an operation
831 with operand OP and code OPCODE removing that from the chain. *DEF
832 is updated if there is only one operand but no operation left. */
834 static void
835 zero_one_operation (tree *def, enum tree_code opcode, tree op)
837 gimple stmt = SSA_NAME_DEF_STMT (*def);
841 tree name = gimple_assign_rhs1 (stmt);
843 /* If this is the operation we look for and one of the operands
844 is ours simply propagate the other operand into the stmts
845 single use. */
846 if (gimple_assign_rhs_code (stmt) == opcode
847 && (name == op
848 || gimple_assign_rhs2 (stmt) == op))
850 gimple use_stmt;
851 use_operand_p use;
852 gimple_stmt_iterator gsi;
853 if (name == op)
854 name = gimple_assign_rhs2 (stmt);
855 gcc_assert (has_single_use (gimple_assign_lhs (stmt)));
856 single_imm_use (gimple_assign_lhs (stmt), &use, &use_stmt);
857 if (gimple_assign_lhs (stmt) == *def)
858 *def = name;
859 SET_USE (use, name);
860 if (TREE_CODE (name) != SSA_NAME)
861 update_stmt (use_stmt);
862 gsi = gsi_for_stmt (stmt);
863 gsi_remove (&gsi, true);
864 release_defs (stmt);
865 return;
868 /* Continue walking the chain. */
869 gcc_assert (name != op
870 && TREE_CODE (name) == SSA_NAME);
871 stmt = SSA_NAME_DEF_STMT (name);
873 while (1);
876 /* Builds one statement performing OP1 OPCODE OP2 using TMPVAR for
877 the result. Places the statement after the definition of either
878 OP1 or OP2. Returns the new statement. */
880 static gimple
881 build_and_add_sum (tree tmpvar, tree op1, tree op2, enum tree_code opcode)
883 gimple op1def = NULL, op2def = NULL;
884 gimple_stmt_iterator gsi;
885 tree op;
886 gimple sum;
888 /* Create the addition statement. */
889 sum = gimple_build_assign_with_ops (opcode, tmpvar, op1, op2);
890 op = make_ssa_name (tmpvar, sum);
891 gimple_assign_set_lhs (sum, op);
893 /* Find an insertion place and insert. */
894 if (TREE_CODE (op1) == SSA_NAME)
895 op1def = SSA_NAME_DEF_STMT (op1);
896 if (TREE_CODE (op2) == SSA_NAME)
897 op2def = SSA_NAME_DEF_STMT (op2);
898 if ((!op1def || gimple_nop_p (op1def))
899 && (!op2def || gimple_nop_p (op2def)))
901 gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR));
902 gsi_insert_before (&gsi, sum, GSI_NEW_STMT);
904 else if ((!op1def || gimple_nop_p (op1def))
905 || (op2def && !gimple_nop_p (op2def)
906 && stmt_dominates_stmt_p (op1def, op2def)))
908 if (gimple_code (op2def) == GIMPLE_PHI)
910 gsi = gsi_after_labels (gimple_bb (op2def));
911 gsi_insert_before (&gsi, sum, GSI_NEW_STMT);
913 else
915 if (!stmt_ends_bb_p (op2def))
917 gsi = gsi_for_stmt (op2def);
918 gsi_insert_after (&gsi, sum, GSI_NEW_STMT);
920 else
922 edge e;
923 edge_iterator ei;
925 FOR_EACH_EDGE (e, ei, gimple_bb (op2def)->succs)
926 if (e->flags & EDGE_FALLTHRU)
927 gsi_insert_on_edge_immediate (e, sum);
931 else
933 if (gimple_code (op1def) == GIMPLE_PHI)
935 gsi = gsi_after_labels (gimple_bb (op1def));
936 gsi_insert_before (&gsi, sum, GSI_NEW_STMT);
938 else
940 if (!stmt_ends_bb_p (op1def))
942 gsi = gsi_for_stmt (op1def);
943 gsi_insert_after (&gsi, sum, GSI_NEW_STMT);
945 else
947 edge e;
948 edge_iterator ei;
950 FOR_EACH_EDGE (e, ei, gimple_bb (op1def)->succs)
951 if (e->flags & EDGE_FALLTHRU)
952 gsi_insert_on_edge_immediate (e, sum);
956 update_stmt (sum);
958 return sum;
961 /* Perform un-distribution of divisions and multiplications.
962 A * X + B * X is transformed into (A + B) * X and A / X + B / X
963 to (A + B) / X for real X.
965 The algorithm is organized as follows.
967 - First we walk the addition chain *OPS looking for summands that
968 are defined by a multiplication or a real division. This results
969 in the candidates bitmap with relevant indices into *OPS.
971 - Second we build the chains of multiplications or divisions for
972 these candidates, counting the number of occurences of (operand, code)
973 pairs in all of the candidates chains.
975 - Third we sort the (operand, code) pairs by number of occurence and
976 process them starting with the pair with the most uses.
978 * For each such pair we walk the candidates again to build a
979 second candidate bitmap noting all multiplication/division chains
980 that have at least one occurence of (operand, code).
982 * We build an alternate addition chain only covering these
983 candidates with one (operand, code) operation removed from their
984 multiplication/division chain.
986 * The first candidate gets replaced by the alternate addition chain
987 multiplied/divided by the operand.
989 * All candidate chains get disabled for further processing and
990 processing of (operand, code) pairs continues.
992 The alternate addition chains built are re-processed by the main
993 reassociation algorithm which allows optimizing a * x * y + b * y * x
994 to (a + b ) * x * y in one invocation of the reassociation pass. */
996 static bool
997 undistribute_ops_list (enum tree_code opcode,
998 VEC (operand_entry_t, heap) **ops, struct loop *loop)
1000 unsigned int length = VEC_length (operand_entry_t, *ops);
1001 operand_entry_t oe1;
1002 unsigned i, j;
1003 sbitmap candidates, candidates2;
1004 unsigned nr_candidates, nr_candidates2;
1005 sbitmap_iterator sbi0;
1006 VEC (operand_entry_t, heap) **subops;
1007 htab_t ctable;
1008 bool changed = false;
1009 int next_oecount_id = 0;
1011 if (length <= 1
1012 || opcode != PLUS_EXPR)
1013 return false;
1015 /* Build a list of candidates to process. */
1016 candidates = sbitmap_alloc (length);
1017 sbitmap_zero (candidates);
1018 nr_candidates = 0;
1019 for (i = 0; VEC_iterate (operand_entry_t, *ops, i, oe1); ++i)
1021 enum tree_code dcode;
1022 gimple oe1def;
1024 if (TREE_CODE (oe1->op) != SSA_NAME)
1025 continue;
1026 oe1def = SSA_NAME_DEF_STMT (oe1->op);
1027 if (!is_gimple_assign (oe1def))
1028 continue;
1029 dcode = gimple_assign_rhs_code (oe1def);
1030 if ((dcode != MULT_EXPR
1031 && dcode != RDIV_EXPR)
1032 || !is_reassociable_op (oe1def, dcode, loop))
1033 continue;
1035 SET_BIT (candidates, i);
1036 nr_candidates++;
1039 if (nr_candidates < 2)
1041 sbitmap_free (candidates);
1042 return false;
1045 if (dump_file && (dump_flags & TDF_DETAILS))
1047 fprintf (dump_file, "searching for un-distribute opportunities ");
1048 print_generic_expr (dump_file,
1049 VEC_index (operand_entry_t, *ops,
1050 sbitmap_first_set_bit (candidates))->op, 0);
1051 fprintf (dump_file, " %d\n", nr_candidates);
1054 /* Build linearized sub-operand lists and the counting table. */
1055 cvec = NULL;
1056 ctable = htab_create (15, oecount_hash, oecount_eq, NULL);
1057 subops = XCNEWVEC (VEC (operand_entry_t, heap) *,
1058 VEC_length (operand_entry_t, *ops));
1059 EXECUTE_IF_SET_IN_SBITMAP (candidates, 0, i, sbi0)
1061 gimple oedef;
1062 enum tree_code oecode;
1063 unsigned j;
1065 oedef = SSA_NAME_DEF_STMT (VEC_index (operand_entry_t, *ops, i)->op);
1066 oecode = gimple_assign_rhs_code (oedef);
1067 linearize_expr_tree (&subops[i], oedef,
1068 associative_tree_code (oecode), false);
1070 for (j = 0; VEC_iterate (operand_entry_t, subops[i], j, oe1); ++j)
1072 oecount c;
1073 void **slot;
1074 size_t idx;
1075 c.oecode = oecode;
1076 c.cnt = 1;
1077 c.id = next_oecount_id++;
1078 c.op = oe1->op;
1079 VEC_safe_push (oecount, heap, cvec, &c);
1080 idx = VEC_length (oecount, cvec) + 41;
1081 slot = htab_find_slot (ctable, (void *)idx, INSERT);
1082 if (!*slot)
1084 *slot = (void *)idx;
1086 else
1088 VEC_pop (oecount, cvec);
1089 VEC_index (oecount, cvec, (size_t)*slot - 42)->cnt++;
1093 htab_delete (ctable);
1095 /* Sort the counting table. */
1096 qsort (VEC_address (oecount, cvec), VEC_length (oecount, cvec),
1097 sizeof (oecount), oecount_cmp);
1099 if (dump_file && (dump_flags & TDF_DETAILS))
1101 oecount *c;
1102 fprintf (dump_file, "Candidates:\n");
1103 for (j = 0; VEC_iterate (oecount, cvec, j, c); ++j)
1105 fprintf (dump_file, " %u %s: ", c->cnt,
1106 c->oecode == MULT_EXPR
1107 ? "*" : c->oecode == RDIV_EXPR ? "/" : "?");
1108 print_generic_expr (dump_file, c->op, 0);
1109 fprintf (dump_file, "\n");
1113 /* Process the (operand, code) pairs in order of most occurence. */
1114 candidates2 = sbitmap_alloc (length);
1115 while (!VEC_empty (oecount, cvec))
1117 oecount *c = VEC_last (oecount, cvec);
1118 if (c->cnt < 2)
1119 break;
1121 /* Now collect the operands in the outer chain that contain
1122 the common operand in their inner chain. */
1123 sbitmap_zero (candidates2);
1124 nr_candidates2 = 0;
1125 EXECUTE_IF_SET_IN_SBITMAP (candidates, 0, i, sbi0)
1127 gimple oedef;
1128 enum tree_code oecode;
1129 unsigned j;
1130 tree op = VEC_index (operand_entry_t, *ops, i)->op;
1132 /* If we undistributed in this chain already this may be
1133 a constant. */
1134 if (TREE_CODE (op) != SSA_NAME)
1135 continue;
1137 oedef = SSA_NAME_DEF_STMT (op);
1138 oecode = gimple_assign_rhs_code (oedef);
1139 if (oecode != c->oecode)
1140 continue;
1142 for (j = 0; VEC_iterate (operand_entry_t, subops[i], j, oe1); ++j)
1144 if (oe1->op == c->op)
1146 SET_BIT (candidates2, i);
1147 ++nr_candidates2;
1148 break;
1153 if (nr_candidates2 >= 2)
1155 operand_entry_t oe1, oe2;
1156 tree tmpvar;
1157 gimple prod;
1158 int first = sbitmap_first_set_bit (candidates2);
1160 /* Build the new addition chain. */
1161 oe1 = VEC_index (operand_entry_t, *ops, first);
1162 if (dump_file && (dump_flags & TDF_DETAILS))
1164 fprintf (dump_file, "Building (");
1165 print_generic_expr (dump_file, oe1->op, 0);
1167 tmpvar = create_tmp_reg (TREE_TYPE (oe1->op), NULL);
1168 add_referenced_var (tmpvar);
1169 zero_one_operation (&oe1->op, c->oecode, c->op);
1170 EXECUTE_IF_SET_IN_SBITMAP (candidates2, first+1, i, sbi0)
1172 gimple sum;
1173 oe2 = VEC_index (operand_entry_t, *ops, i);
1174 if (dump_file && (dump_flags & TDF_DETAILS))
1176 fprintf (dump_file, " + ");
1177 print_generic_expr (dump_file, oe2->op, 0);
1179 zero_one_operation (&oe2->op, c->oecode, c->op);
1180 sum = build_and_add_sum (tmpvar, oe1->op, oe2->op, opcode);
1181 oe2->op = fold_convert (TREE_TYPE (oe2->op), integer_zero_node);
1182 oe2->rank = 0;
1183 oe1->op = gimple_get_lhs (sum);
1186 /* Apply the multiplication/division. */
1187 prod = build_and_add_sum (tmpvar, oe1->op, c->op, c->oecode);
1188 if (dump_file && (dump_flags & TDF_DETAILS))
1190 fprintf (dump_file, ") %s ", c->oecode == MULT_EXPR ? "*" : "/");
1191 print_generic_expr (dump_file, c->op, 0);
1192 fprintf (dump_file, "\n");
1195 /* Record it in the addition chain and disable further
1196 undistribution with this op. */
1197 oe1->op = gimple_assign_lhs (prod);
1198 oe1->rank = get_rank (oe1->op);
1199 VEC_free (operand_entry_t, heap, subops[first]);
1201 changed = true;
1204 VEC_pop (oecount, cvec);
1207 for (i = 0; i < VEC_length (operand_entry_t, *ops); ++i)
1208 VEC_free (operand_entry_t, heap, subops[i]);
1209 free (subops);
1210 VEC_free (oecount, heap, cvec);
1211 sbitmap_free (candidates);
1212 sbitmap_free (candidates2);
1214 return changed;
1217 /* If OPCODE is BIT_IOR_EXPR or BIT_AND_EXPR and CURR is a comparison
1218 expression, examine the other OPS to see if any of them are comparisons
1219 of the same values, which we may be able to combine or eliminate.
1220 For example, we can rewrite (a < b) | (a == b) as (a <= b). */
1222 static bool
1223 eliminate_redundant_comparison (enum tree_code opcode,
1224 VEC (operand_entry_t, heap) **ops,
1225 unsigned int currindex,
1226 operand_entry_t curr)
1228 tree op1, op2;
1229 enum tree_code lcode, rcode;
1230 gimple def1, def2;
1231 int i;
1232 operand_entry_t oe;
1234 if (opcode != BIT_IOR_EXPR && opcode != BIT_AND_EXPR)
1235 return false;
1237 /* Check that CURR is a comparison. */
1238 if (TREE_CODE (curr->op) != SSA_NAME)
1239 return false;
1240 def1 = SSA_NAME_DEF_STMT (curr->op);
1241 if (!is_gimple_assign (def1))
1242 return false;
1243 lcode = gimple_assign_rhs_code (def1);
1244 if (TREE_CODE_CLASS (lcode) != tcc_comparison)
1245 return false;
1246 op1 = gimple_assign_rhs1 (def1);
1247 op2 = gimple_assign_rhs2 (def1);
1249 /* Now look for a similar comparison in the remaining OPS. */
1250 for (i = currindex + 1;
1251 VEC_iterate (operand_entry_t, *ops, i, oe);
1252 i++)
1254 tree t;
1256 if (TREE_CODE (oe->op) != SSA_NAME)
1257 continue;
1258 def2 = SSA_NAME_DEF_STMT (oe->op);
1259 if (!is_gimple_assign (def2))
1260 continue;
1261 rcode = gimple_assign_rhs_code (def2);
1262 if (TREE_CODE_CLASS (rcode) != tcc_comparison)
1263 continue;
1265 /* If we got here, we have a match. See if we can combine the
1266 two comparisons. */
1267 if (opcode == BIT_IOR_EXPR)
1268 t = maybe_fold_or_comparisons (lcode, op1, op2,
1269 rcode, gimple_assign_rhs1 (def2),
1270 gimple_assign_rhs2 (def2));
1271 else
1272 t = maybe_fold_and_comparisons (lcode, op1, op2,
1273 rcode, gimple_assign_rhs1 (def2),
1274 gimple_assign_rhs2 (def2));
1275 if (!t)
1276 continue;
1278 /* maybe_fold_and_comparisons and maybe_fold_or_comparisons
1279 always give us a boolean_type_node value back. If the original
1280 BIT_AND_EXPR or BIT_IOR_EXPR was of a wider integer type,
1281 we need to convert. */
1282 if (!useless_type_conversion_p (TREE_TYPE (curr->op), TREE_TYPE (t)))
1283 t = fold_convert (TREE_TYPE (curr->op), t);
1285 if (dump_file && (dump_flags & TDF_DETAILS))
1287 fprintf (dump_file, "Equivalence: ");
1288 print_generic_expr (dump_file, curr->op, 0);
1289 fprintf (dump_file, " %s ", op_symbol_code (opcode));
1290 print_generic_expr (dump_file, oe->op, 0);
1291 fprintf (dump_file, " -> ");
1292 print_generic_expr (dump_file, t, 0);
1293 fprintf (dump_file, "\n");
1296 /* Now we can delete oe, as it has been subsumed by the new combined
1297 expression t. */
1298 VEC_ordered_remove (operand_entry_t, *ops, i);
1299 reassociate_stats.ops_eliminated ++;
1301 /* If t is the same as curr->op, we're done. Otherwise we must
1302 replace curr->op with t. Special case is if we got a constant
1303 back, in which case we add it to the end instead of in place of
1304 the current entry. */
1305 if (TREE_CODE (t) == INTEGER_CST)
1307 VEC_ordered_remove (operand_entry_t, *ops, currindex);
1308 add_to_ops_vec (ops, t);
1310 else if (!operand_equal_p (t, curr->op, 0))
1312 tree tmpvar;
1313 gimple sum;
1314 enum tree_code subcode;
1315 tree newop1;
1316 tree newop2;
1317 tmpvar = create_tmp_var (TREE_TYPE (t), NULL);
1318 add_referenced_var (tmpvar);
1319 extract_ops_from_tree (t, &subcode, &newop1, &newop2);
1320 sum = build_and_add_sum (tmpvar, newop1, newop2, subcode);
1321 curr->op = gimple_get_lhs (sum);
1323 return true;
1326 return false;
1329 /* Perform various identities and other optimizations on the list of
1330 operand entries, stored in OPS. The tree code for the binary
1331 operation between all the operands is OPCODE. */
1333 static void
1334 optimize_ops_list (enum tree_code opcode,
1335 VEC (operand_entry_t, heap) **ops)
1337 unsigned int length = VEC_length (operand_entry_t, *ops);
1338 unsigned int i;
1339 operand_entry_t oe;
1340 operand_entry_t oelast = NULL;
1341 bool iterate = false;
1343 if (length == 1)
1344 return;
1346 oelast = VEC_last (operand_entry_t, *ops);
1348 /* If the last two are constants, pop the constants off, merge them
1349 and try the next two. */
1350 if (oelast->rank == 0 && is_gimple_min_invariant (oelast->op))
1352 operand_entry_t oelm1 = VEC_index (operand_entry_t, *ops, length - 2);
1354 if (oelm1->rank == 0
1355 && is_gimple_min_invariant (oelm1->op)
1356 && useless_type_conversion_p (TREE_TYPE (oelm1->op),
1357 TREE_TYPE (oelast->op)))
1359 tree folded = fold_binary (opcode, TREE_TYPE (oelm1->op),
1360 oelm1->op, oelast->op);
1362 if (folded && is_gimple_min_invariant (folded))
1364 if (dump_file && (dump_flags & TDF_DETAILS))
1365 fprintf (dump_file, "Merging constants\n");
1367 VEC_pop (operand_entry_t, *ops);
1368 VEC_pop (operand_entry_t, *ops);
1370 add_to_ops_vec (ops, folded);
1371 reassociate_stats.constants_eliminated++;
1373 optimize_ops_list (opcode, ops);
1374 return;
1379 eliminate_using_constants (opcode, ops);
1380 oelast = NULL;
1382 for (i = 0; VEC_iterate (operand_entry_t, *ops, i, oe);)
1384 bool done = false;
1386 if (eliminate_not_pairs (opcode, ops, i, oe))
1387 return;
1388 if (eliminate_duplicate_pair (opcode, ops, &done, i, oe, oelast)
1389 || (!done && eliminate_plus_minus_pair (opcode, ops, i, oe))
1390 || (!done && eliminate_redundant_comparison (opcode, ops, i, oe)))
1392 if (done)
1393 return;
1394 iterate = true;
1395 oelast = NULL;
1396 continue;
1398 oelast = oe;
1399 i++;
1402 length = VEC_length (operand_entry_t, *ops);
1403 oelast = VEC_last (operand_entry_t, *ops);
1405 if (iterate)
1406 optimize_ops_list (opcode, ops);
1409 /* Return true if OPERAND is defined by a PHI node which uses the LHS
1410 of STMT in it's operands. This is also known as a "destructive
1411 update" operation. */
1413 static bool
1414 is_phi_for_stmt (gimple stmt, tree operand)
1416 gimple def_stmt;
1417 tree lhs;
1418 use_operand_p arg_p;
1419 ssa_op_iter i;
1421 if (TREE_CODE (operand) != SSA_NAME)
1422 return false;
1424 lhs = gimple_assign_lhs (stmt);
1426 def_stmt = SSA_NAME_DEF_STMT (operand);
1427 if (gimple_code (def_stmt) != GIMPLE_PHI)
1428 return false;
1430 FOR_EACH_PHI_ARG (arg_p, def_stmt, i, SSA_OP_USE)
1431 if (lhs == USE_FROM_PTR (arg_p))
1432 return true;
1433 return false;
1436 /* Remove def stmt of VAR if VAR has zero uses and recurse
1437 on rhs1 operand if so. */
1439 static void
1440 remove_visited_stmt_chain (tree var)
1442 gimple stmt;
1443 gimple_stmt_iterator gsi;
1445 while (1)
1447 if (TREE_CODE (var) != SSA_NAME || !has_zero_uses (var))
1448 return;
1449 stmt = SSA_NAME_DEF_STMT (var);
1450 if (!is_gimple_assign (stmt)
1451 || !gimple_visited_p (stmt))
1452 return;
1453 var = gimple_assign_rhs1 (stmt);
1454 gsi = gsi_for_stmt (stmt);
1455 gsi_remove (&gsi, true);
1456 release_defs (stmt);
1460 /* Recursively rewrite our linearized statements so that the operators
1461 match those in OPS[OPINDEX], putting the computation in rank
1462 order. */
1464 static void
1465 rewrite_expr_tree (gimple stmt, unsigned int opindex,
1466 VEC(operand_entry_t, heap) * ops, bool moved)
1468 tree rhs1 = gimple_assign_rhs1 (stmt);
1469 tree rhs2 = gimple_assign_rhs2 (stmt);
1470 operand_entry_t oe;
1472 /* If we have three operands left, then we want to make sure the one
1473 that gets the double binary op are the ones with the same rank.
1475 The alternative we try is to see if this is a destructive
1476 update style statement, which is like:
1477 b = phi (a, ...)
1478 a = c + b;
1479 In that case, we want to use the destructive update form to
1480 expose the possible vectorizer sum reduction opportunity.
1481 In that case, the third operand will be the phi node.
1483 We could, of course, try to be better as noted above, and do a
1484 lot of work to try to find these opportunities in >3 operand
1485 cases, but it is unlikely to be worth it. */
1486 if (opindex + 3 == VEC_length (operand_entry_t, ops))
1488 operand_entry_t oe1, oe2, oe3;
1490 oe1 = VEC_index (operand_entry_t, ops, opindex);
1491 oe2 = VEC_index (operand_entry_t, ops, opindex + 1);
1492 oe3 = VEC_index (operand_entry_t, ops, opindex + 2);
1494 if ((oe1->rank == oe2->rank
1495 && oe2->rank != oe3->rank)
1496 || (is_phi_for_stmt (stmt, oe3->op)
1497 && !is_phi_for_stmt (stmt, oe1->op)
1498 && !is_phi_for_stmt (stmt, oe2->op)))
1500 struct operand_entry temp = *oe3;
1501 oe3->op = oe1->op;
1502 oe3->rank = oe1->rank;
1503 oe1->op = temp.op;
1504 oe1->rank= temp.rank;
1506 else if ((oe1->rank == oe3->rank
1507 && oe2->rank != oe3->rank)
1508 || (is_phi_for_stmt (stmt, oe2->op)
1509 && !is_phi_for_stmt (stmt, oe1->op)
1510 && !is_phi_for_stmt (stmt, oe3->op)))
1512 struct operand_entry temp = *oe2;
1513 oe2->op = oe1->op;
1514 oe2->rank = oe1->rank;
1515 oe1->op = temp.op;
1516 oe1->rank= temp.rank;
1520 /* The final recursion case for this function is that you have
1521 exactly two operations left.
1522 If we had one exactly one op in the entire list to start with, we
1523 would have never called this function, and the tail recursion
1524 rewrites them one at a time. */
1525 if (opindex + 2 == VEC_length (operand_entry_t, ops))
1527 operand_entry_t oe1, oe2;
1529 oe1 = VEC_index (operand_entry_t, ops, opindex);
1530 oe2 = VEC_index (operand_entry_t, ops, opindex + 1);
1532 if (rhs1 != oe1->op || rhs2 != oe2->op)
1534 if (dump_file && (dump_flags & TDF_DETAILS))
1536 fprintf (dump_file, "Transforming ");
1537 print_gimple_stmt (dump_file, stmt, 0, 0);
1540 gimple_assign_set_rhs1 (stmt, oe1->op);
1541 gimple_assign_set_rhs2 (stmt, oe2->op);
1542 update_stmt (stmt);
1543 if (rhs1 != oe1->op && rhs1 != oe2->op)
1544 remove_visited_stmt_chain (rhs1);
1546 if (dump_file && (dump_flags & TDF_DETAILS))
1548 fprintf (dump_file, " into ");
1549 print_gimple_stmt (dump_file, stmt, 0, 0);
1553 return;
1556 /* If we hit here, we should have 3 or more ops left. */
1557 gcc_assert (opindex + 2 < VEC_length (operand_entry_t, ops));
1559 /* Rewrite the next operator. */
1560 oe = VEC_index (operand_entry_t, ops, opindex);
1562 if (oe->op != rhs2)
1564 if (!moved)
1566 gimple_stmt_iterator gsinow, gsirhs1;
1567 gimple stmt1 = stmt, stmt2;
1568 unsigned int count;
1570 gsinow = gsi_for_stmt (stmt);
1571 count = VEC_length (operand_entry_t, ops) - opindex - 2;
1572 while (count-- != 0)
1574 stmt2 = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt1));
1575 gsirhs1 = gsi_for_stmt (stmt2);
1576 gsi_move_before (&gsirhs1, &gsinow);
1577 gsi_prev (&gsinow);
1578 stmt1 = stmt2;
1580 moved = true;
1583 if (dump_file && (dump_flags & TDF_DETAILS))
1585 fprintf (dump_file, "Transforming ");
1586 print_gimple_stmt (dump_file, stmt, 0, 0);
1589 gimple_assign_set_rhs2 (stmt, oe->op);
1590 update_stmt (stmt);
1592 if (dump_file && (dump_flags & TDF_DETAILS))
1594 fprintf (dump_file, " into ");
1595 print_gimple_stmt (dump_file, stmt, 0, 0);
1598 /* Recurse on the LHS of the binary operator, which is guaranteed to
1599 be the non-leaf side. */
1600 rewrite_expr_tree (SSA_NAME_DEF_STMT (rhs1), opindex + 1, ops, moved);
1603 /* Transform STMT, which is really (A +B) + (C + D) into the left
1604 linear form, ((A+B)+C)+D.
1605 Recurse on D if necessary. */
1607 static void
1608 linearize_expr (gimple stmt)
1610 gimple_stmt_iterator gsinow, gsirhs;
1611 gimple binlhs = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (stmt));
1612 gimple binrhs = SSA_NAME_DEF_STMT (gimple_assign_rhs2 (stmt));
1613 enum tree_code rhscode = gimple_assign_rhs_code (stmt);
1614 gimple newbinrhs = NULL;
1615 struct loop *loop = loop_containing_stmt (stmt);
1617 gcc_assert (is_reassociable_op (binlhs, rhscode, loop)
1618 && is_reassociable_op (binrhs, rhscode, loop));
1620 gsinow = gsi_for_stmt (stmt);
1621 gsirhs = gsi_for_stmt (binrhs);
1622 gsi_move_before (&gsirhs, &gsinow);
1624 gimple_assign_set_rhs2 (stmt, gimple_assign_rhs1 (binrhs));
1625 gimple_assign_set_rhs1 (binrhs, gimple_assign_lhs (binlhs));
1626 gimple_assign_set_rhs1 (stmt, gimple_assign_lhs (binrhs));
1628 if (TREE_CODE (gimple_assign_rhs2 (stmt)) == SSA_NAME)
1629 newbinrhs = SSA_NAME_DEF_STMT (gimple_assign_rhs2 (stmt));
1631 if (dump_file && (dump_flags & TDF_DETAILS))
1633 fprintf (dump_file, "Linearized: ");
1634 print_gimple_stmt (dump_file, stmt, 0, 0);
1637 reassociate_stats.linearized++;
1638 update_stmt (binrhs);
1639 update_stmt (binlhs);
1640 update_stmt (stmt);
1642 gimple_set_visited (stmt, true);
1643 gimple_set_visited (binlhs, true);
1644 gimple_set_visited (binrhs, true);
1646 /* Tail recurse on the new rhs if it still needs reassociation. */
1647 if (newbinrhs && is_reassociable_op (newbinrhs, rhscode, loop))
1648 /* ??? This should probably be linearize_expr (newbinrhs) but I don't
1649 want to change the algorithm while converting to tuples. */
1650 linearize_expr (stmt);
1653 /* If LHS has a single immediate use that is a GIMPLE_ASSIGN statement, return
1654 it. Otherwise, return NULL. */
1656 static gimple
1657 get_single_immediate_use (tree lhs)
1659 use_operand_p immuse;
1660 gimple immusestmt;
1662 if (TREE_CODE (lhs) == SSA_NAME
1663 && single_imm_use (lhs, &immuse, &immusestmt)
1664 && is_gimple_assign (immusestmt))
1665 return immusestmt;
1667 return NULL;
1670 /* Recursively negate the value of TONEGATE, and return the SSA_NAME
1671 representing the negated value. Insertions of any necessary
1672 instructions go before GSI.
1673 This function is recursive in that, if you hand it "a_5" as the
1674 value to negate, and a_5 is defined by "a_5 = b_3 + b_4", it will
1675 transform b_3 + b_4 into a_5 = -b_3 + -b_4. */
1677 static tree
1678 negate_value (tree tonegate, gimple_stmt_iterator *gsi)
1680 gimple negatedefstmt= NULL;
1681 tree resultofnegate;
1683 /* If we are trying to negate a name, defined by an add, negate the
1684 add operands instead. */
1685 if (TREE_CODE (tonegate) == SSA_NAME)
1686 negatedefstmt = SSA_NAME_DEF_STMT (tonegate);
1687 if (TREE_CODE (tonegate) == SSA_NAME
1688 && is_gimple_assign (negatedefstmt)
1689 && TREE_CODE (gimple_assign_lhs (negatedefstmt)) == SSA_NAME
1690 && has_single_use (gimple_assign_lhs (negatedefstmt))
1691 && gimple_assign_rhs_code (negatedefstmt) == PLUS_EXPR)
1693 gimple_stmt_iterator gsi;
1694 tree rhs1 = gimple_assign_rhs1 (negatedefstmt);
1695 tree rhs2 = gimple_assign_rhs2 (negatedefstmt);
1697 gsi = gsi_for_stmt (negatedefstmt);
1698 rhs1 = negate_value (rhs1, &gsi);
1699 gimple_assign_set_rhs1 (negatedefstmt, rhs1);
1701 gsi = gsi_for_stmt (negatedefstmt);
1702 rhs2 = negate_value (rhs2, &gsi);
1703 gimple_assign_set_rhs2 (negatedefstmt, rhs2);
1705 update_stmt (negatedefstmt);
1706 return gimple_assign_lhs (negatedefstmt);
1709 tonegate = fold_build1 (NEGATE_EXPR, TREE_TYPE (tonegate), tonegate);
1710 resultofnegate = force_gimple_operand_gsi (gsi, tonegate, true,
1711 NULL_TREE, true, GSI_SAME_STMT);
1712 return resultofnegate;
1715 /* Return true if we should break up the subtract in STMT into an add
1716 with negate. This is true when we the subtract operands are really
1717 adds, or the subtract itself is used in an add expression. In
1718 either case, breaking up the subtract into an add with negate
1719 exposes the adds to reassociation. */
1721 static bool
1722 should_break_up_subtract (gimple stmt)
1724 tree lhs = gimple_assign_lhs (stmt);
1725 tree binlhs = gimple_assign_rhs1 (stmt);
1726 tree binrhs = gimple_assign_rhs2 (stmt);
1727 gimple immusestmt;
1728 struct loop *loop = loop_containing_stmt (stmt);
1730 if (TREE_CODE (binlhs) == SSA_NAME
1731 && is_reassociable_op (SSA_NAME_DEF_STMT (binlhs), PLUS_EXPR, loop))
1732 return true;
1734 if (TREE_CODE (binrhs) == SSA_NAME
1735 && is_reassociable_op (SSA_NAME_DEF_STMT (binrhs), PLUS_EXPR, loop))
1736 return true;
1738 if (TREE_CODE (lhs) == SSA_NAME
1739 && (immusestmt = get_single_immediate_use (lhs))
1740 && is_gimple_assign (immusestmt)
1741 && (gimple_assign_rhs_code (immusestmt) == PLUS_EXPR
1742 || gimple_assign_rhs_code (immusestmt) == MULT_EXPR))
1743 return true;
1744 return false;
1747 /* Transform STMT from A - B into A + -B. */
1749 static void
1750 break_up_subtract (gimple stmt, gimple_stmt_iterator *gsip)
1752 tree rhs1 = gimple_assign_rhs1 (stmt);
1753 tree rhs2 = gimple_assign_rhs2 (stmt);
1755 if (dump_file && (dump_flags & TDF_DETAILS))
1757 fprintf (dump_file, "Breaking up subtract ");
1758 print_gimple_stmt (dump_file, stmt, 0, 0);
1761 rhs2 = negate_value (rhs2, gsip);
1762 gimple_assign_set_rhs_with_ops (gsip, PLUS_EXPR, rhs1, rhs2);
1763 update_stmt (stmt);
1766 /* Recursively linearize a binary expression that is the RHS of STMT.
1767 Place the operands of the expression tree in the vector named OPS. */
1769 static void
1770 linearize_expr_tree (VEC(operand_entry_t, heap) **ops, gimple stmt,
1771 bool is_associative, bool set_visited)
1773 tree binlhs = gimple_assign_rhs1 (stmt);
1774 tree binrhs = gimple_assign_rhs2 (stmt);
1775 gimple binlhsdef, binrhsdef;
1776 bool binlhsisreassoc = false;
1777 bool binrhsisreassoc = false;
1778 enum tree_code rhscode = gimple_assign_rhs_code (stmt);
1779 struct loop *loop = loop_containing_stmt (stmt);
1781 if (set_visited)
1782 gimple_set_visited (stmt, true);
1784 if (TREE_CODE (binlhs) == SSA_NAME)
1786 binlhsdef = SSA_NAME_DEF_STMT (binlhs);
1787 binlhsisreassoc = is_reassociable_op (binlhsdef, rhscode, loop);
1790 if (TREE_CODE (binrhs) == SSA_NAME)
1792 binrhsdef = SSA_NAME_DEF_STMT (binrhs);
1793 binrhsisreassoc = is_reassociable_op (binrhsdef, rhscode, loop);
1796 /* If the LHS is not reassociable, but the RHS is, we need to swap
1797 them. If neither is reassociable, there is nothing we can do, so
1798 just put them in the ops vector. If the LHS is reassociable,
1799 linearize it. If both are reassociable, then linearize the RHS
1800 and the LHS. */
1802 if (!binlhsisreassoc)
1804 tree temp;
1806 /* If this is not a associative operation like division, give up. */
1807 if (!is_associative)
1809 add_to_ops_vec (ops, binrhs);
1810 return;
1813 if (!binrhsisreassoc)
1815 add_to_ops_vec (ops, binrhs);
1816 add_to_ops_vec (ops, binlhs);
1817 return;
1820 if (dump_file && (dump_flags & TDF_DETAILS))
1822 fprintf (dump_file, "swapping operands of ");
1823 print_gimple_stmt (dump_file, stmt, 0, 0);
1826 swap_tree_operands (stmt,
1827 gimple_assign_rhs1_ptr (stmt),
1828 gimple_assign_rhs2_ptr (stmt));
1829 update_stmt (stmt);
1831 if (dump_file && (dump_flags & TDF_DETAILS))
1833 fprintf (dump_file, " is now ");
1834 print_gimple_stmt (dump_file, stmt, 0, 0);
1837 /* We want to make it so the lhs is always the reassociative op,
1838 so swap. */
1839 temp = binlhs;
1840 binlhs = binrhs;
1841 binrhs = temp;
1843 else if (binrhsisreassoc)
1845 linearize_expr (stmt);
1846 binlhs = gimple_assign_rhs1 (stmt);
1847 binrhs = gimple_assign_rhs2 (stmt);
1850 gcc_assert (TREE_CODE (binrhs) != SSA_NAME
1851 || !is_reassociable_op (SSA_NAME_DEF_STMT (binrhs),
1852 rhscode, loop));
1853 linearize_expr_tree (ops, SSA_NAME_DEF_STMT (binlhs),
1854 is_associative, set_visited);
1855 add_to_ops_vec (ops, binrhs);
1858 /* Repropagate the negates back into subtracts, since no other pass
1859 currently does it. */
1861 static void
1862 repropagate_negates (void)
1864 unsigned int i = 0;
1865 tree negate;
1867 for (i = 0; VEC_iterate (tree, plus_negates, i, negate); i++)
1869 gimple user = get_single_immediate_use (negate);
1871 if (!user || !is_gimple_assign (user))
1872 continue;
1874 /* The negate operand can be either operand of a PLUS_EXPR
1875 (it can be the LHS if the RHS is a constant for example).
1877 Force the negate operand to the RHS of the PLUS_EXPR, then
1878 transform the PLUS_EXPR into a MINUS_EXPR. */
1879 if (gimple_assign_rhs_code (user) == PLUS_EXPR)
1881 /* If the negated operand appears on the LHS of the
1882 PLUS_EXPR, exchange the operands of the PLUS_EXPR
1883 to force the negated operand to the RHS of the PLUS_EXPR. */
1884 if (gimple_assign_rhs1 (user) == negate)
1886 swap_tree_operands (user,
1887 gimple_assign_rhs1_ptr (user),
1888 gimple_assign_rhs2_ptr (user));
1891 /* Now transform the PLUS_EXPR into a MINUS_EXPR and replace
1892 the RHS of the PLUS_EXPR with the operand of the NEGATE_EXPR. */
1893 if (gimple_assign_rhs2 (user) == negate)
1895 tree rhs1 = gimple_assign_rhs1 (user);
1896 tree rhs2 = get_unary_op (negate, NEGATE_EXPR);
1897 gimple_stmt_iterator gsi = gsi_for_stmt (user);
1898 gimple_assign_set_rhs_with_ops (&gsi, MINUS_EXPR, rhs1, rhs2);
1899 update_stmt (user);
1902 else if (gimple_assign_rhs_code (user) == MINUS_EXPR)
1904 if (gimple_assign_rhs1 (user) == negate)
1906 /* We have
1907 x = -a
1908 y = x - b
1909 which we transform into
1910 x = a + b
1911 y = -x .
1912 This pushes down the negate which we possibly can merge
1913 into some other operation, hence insert it into the
1914 plus_negates vector. */
1915 gimple feed = SSA_NAME_DEF_STMT (negate);
1916 tree a = gimple_assign_rhs1 (feed);
1917 tree rhs2 = gimple_assign_rhs2 (user);
1918 gimple_stmt_iterator gsi = gsi_for_stmt (feed), gsi2;
1919 gimple_replace_lhs (feed, negate);
1920 gimple_assign_set_rhs_with_ops (&gsi, PLUS_EXPR, a, rhs2);
1921 update_stmt (gsi_stmt (gsi));
1922 gsi2 = gsi_for_stmt (user);
1923 gimple_assign_set_rhs_with_ops (&gsi2, NEGATE_EXPR, negate, NULL);
1924 update_stmt (gsi_stmt (gsi2));
1925 gsi_move_before (&gsi, &gsi2);
1926 VEC_safe_push (tree, heap, plus_negates,
1927 gimple_assign_lhs (gsi_stmt (gsi2)));
1929 else
1931 /* Transform "x = -a; y = b - x" into "y = b + a", getting
1932 rid of one operation. */
1933 gimple feed = SSA_NAME_DEF_STMT (negate);
1934 tree a = gimple_assign_rhs1 (feed);
1935 tree rhs1 = gimple_assign_rhs1 (user);
1936 gimple_stmt_iterator gsi = gsi_for_stmt (user);
1937 gimple_assign_set_rhs_with_ops (&gsi, PLUS_EXPR, rhs1, a);
1938 update_stmt (gsi_stmt (gsi));
1944 /* Returns true if OP is of a type for which we can do reassociation.
1945 That is for integral or non-saturating fixed-point types, and for
1946 floating point type when associative-math is enabled. */
1948 static bool
1949 can_reassociate_p (tree op)
1951 tree type = TREE_TYPE (op);
1952 if (INTEGRAL_TYPE_P (type)
1953 || NON_SAT_FIXED_POINT_TYPE_P (type)
1954 || (flag_associative_math && FLOAT_TYPE_P (type)))
1955 return true;
1956 return false;
1959 /* Break up subtract operations in block BB.
1961 We do this top down because we don't know whether the subtract is
1962 part of a possible chain of reassociation except at the top.
1964 IE given
1965 d = f + g
1966 c = a + e
1967 b = c - d
1968 q = b - r
1969 k = t - q
1971 we want to break up k = t - q, but we won't until we've transformed q
1972 = b - r, which won't be broken up until we transform b = c - d.
1974 En passant, clear the GIMPLE visited flag on every statement. */
1976 static void
1977 break_up_subtract_bb (basic_block bb)
1979 gimple_stmt_iterator gsi;
1980 basic_block son;
1982 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1984 gimple stmt = gsi_stmt (gsi);
1985 gimple_set_visited (stmt, false);
1987 if (!is_gimple_assign (stmt)
1988 || !can_reassociate_p (gimple_assign_lhs (stmt)))
1989 continue;
1991 /* Look for simple gimple subtract operations. */
1992 if (gimple_assign_rhs_code (stmt) == MINUS_EXPR)
1994 if (!can_reassociate_p (gimple_assign_rhs1 (stmt))
1995 || !can_reassociate_p (gimple_assign_rhs2 (stmt)))
1996 continue;
1998 /* Check for a subtract used only in an addition. If this
1999 is the case, transform it into add of a negate for better
2000 reassociation. IE transform C = A-B into C = A + -B if C
2001 is only used in an addition. */
2002 if (should_break_up_subtract (stmt))
2003 break_up_subtract (stmt, &gsi);
2005 else if (gimple_assign_rhs_code (stmt) == NEGATE_EXPR
2006 && can_reassociate_p (gimple_assign_rhs1 (stmt)))
2007 VEC_safe_push (tree, heap, plus_negates, gimple_assign_lhs (stmt));
2009 for (son = first_dom_son (CDI_DOMINATORS, bb);
2010 son;
2011 son = next_dom_son (CDI_DOMINATORS, son))
2012 break_up_subtract_bb (son);
2015 /* Reassociate expressions in basic block BB and its post-dominator as
2016 children. */
2018 static void
2019 reassociate_bb (basic_block bb)
2021 gimple_stmt_iterator gsi;
2022 basic_block son;
2024 for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi))
2026 gimple stmt = gsi_stmt (gsi);
2028 if (is_gimple_assign (stmt))
2030 tree lhs, rhs1, rhs2;
2031 enum tree_code rhs_code = gimple_assign_rhs_code (stmt);
2033 /* If this is not a gimple binary expression, there is
2034 nothing for us to do with it. */
2035 if (get_gimple_rhs_class (rhs_code) != GIMPLE_BINARY_RHS)
2036 continue;
2038 /* If this was part of an already processed statement,
2039 we don't need to touch it again. */
2040 if (gimple_visited_p (stmt))
2042 /* This statement might have become dead because of previous
2043 reassociations. */
2044 if (has_zero_uses (gimple_get_lhs (stmt)))
2046 gsi_remove (&gsi, true);
2047 release_defs (stmt);
2048 /* We might end up removing the last stmt above which
2049 places the iterator to the end of the sequence.
2050 Reset it to the last stmt in this case which might
2051 be the end of the sequence as well if we removed
2052 the last statement of the sequence. In which case
2053 we need to bail out. */
2054 if (gsi_end_p (gsi))
2056 gsi = gsi_last_bb (bb);
2057 if (gsi_end_p (gsi))
2058 break;
2061 continue;
2064 lhs = gimple_assign_lhs (stmt);
2065 rhs1 = gimple_assign_rhs1 (stmt);
2066 rhs2 = gimple_assign_rhs2 (stmt);
2068 if (!can_reassociate_p (lhs)
2069 || !can_reassociate_p (rhs1)
2070 || !can_reassociate_p (rhs2))
2071 continue;
2073 if (associative_tree_code (rhs_code))
2075 VEC(operand_entry_t, heap) *ops = NULL;
2077 /* There may be no immediate uses left by the time we
2078 get here because we may have eliminated them all. */
2079 if (TREE_CODE (lhs) == SSA_NAME && has_zero_uses (lhs))
2080 continue;
2082 gimple_set_visited (stmt, true);
2083 linearize_expr_tree (&ops, stmt, true, true);
2084 qsort (VEC_address (operand_entry_t, ops),
2085 VEC_length (operand_entry_t, ops),
2086 sizeof (operand_entry_t),
2087 sort_by_operand_rank);
2088 optimize_ops_list (rhs_code, &ops);
2089 if (undistribute_ops_list (rhs_code, &ops,
2090 loop_containing_stmt (stmt)))
2092 qsort (VEC_address (operand_entry_t, ops),
2093 VEC_length (operand_entry_t, ops),
2094 sizeof (operand_entry_t),
2095 sort_by_operand_rank);
2096 optimize_ops_list (rhs_code, &ops);
2099 if (VEC_length (operand_entry_t, ops) == 1)
2101 if (dump_file && (dump_flags & TDF_DETAILS))
2103 fprintf (dump_file, "Transforming ");
2104 print_gimple_stmt (dump_file, stmt, 0, 0);
2107 rhs1 = gimple_assign_rhs1 (stmt);
2108 gimple_assign_set_rhs_from_tree (&gsi,
2109 VEC_last (operand_entry_t,
2110 ops)->op);
2111 update_stmt (stmt);
2112 remove_visited_stmt_chain (rhs1);
2114 if (dump_file && (dump_flags & TDF_DETAILS))
2116 fprintf (dump_file, " into ");
2117 print_gimple_stmt (dump_file, stmt, 0, 0);
2120 else
2121 rewrite_expr_tree (stmt, 0, ops, false);
2123 VEC_free (operand_entry_t, heap, ops);
2127 for (son = first_dom_son (CDI_POST_DOMINATORS, bb);
2128 son;
2129 son = next_dom_son (CDI_POST_DOMINATORS, son))
2130 reassociate_bb (son);
2133 void dump_ops_vector (FILE *file, VEC (operand_entry_t, heap) *ops);
2134 void debug_ops_vector (VEC (operand_entry_t, heap) *ops);
2136 /* Dump the operand entry vector OPS to FILE. */
2138 void
2139 dump_ops_vector (FILE *file, VEC (operand_entry_t, heap) *ops)
2141 operand_entry_t oe;
2142 unsigned int i;
2144 for (i = 0; VEC_iterate (operand_entry_t, ops, i, oe); i++)
2146 fprintf (file, "Op %d -> rank: %d, tree: ", i, oe->rank);
2147 print_generic_expr (file, oe->op, 0);
2151 /* Dump the operand entry vector OPS to STDERR. */
2153 DEBUG_FUNCTION void
2154 debug_ops_vector (VEC (operand_entry_t, heap) *ops)
2156 dump_ops_vector (stderr, ops);
2159 static void
2160 do_reassoc (void)
2162 break_up_subtract_bb (ENTRY_BLOCK_PTR);
2163 reassociate_bb (EXIT_BLOCK_PTR);
2166 /* Initialize the reassociation pass. */
2168 static void
2169 init_reassoc (void)
2171 int i;
2172 long rank = 2;
2173 tree param;
2174 int *bbs = XNEWVEC (int, last_basic_block + 1);
2176 /* Find the loops, so that we can prevent moving calculations in
2177 them. */
2178 loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
2180 memset (&reassociate_stats, 0, sizeof (reassociate_stats));
2182 operand_entry_pool = create_alloc_pool ("operand entry pool",
2183 sizeof (struct operand_entry), 30);
2184 next_operand_entry_id = 0;
2186 /* Reverse RPO (Reverse Post Order) will give us something where
2187 deeper loops come later. */
2188 pre_and_rev_post_order_compute (NULL, bbs, false);
2189 bb_rank = XCNEWVEC (long, last_basic_block + 1);
2190 operand_rank = pointer_map_create ();
2192 /* Give each argument a distinct rank. */
2193 for (param = DECL_ARGUMENTS (current_function_decl);
2194 param;
2195 param = DECL_CHAIN (param))
2197 if (gimple_default_def (cfun, param) != NULL)
2199 tree def = gimple_default_def (cfun, param);
2200 insert_operand_rank (def, ++rank);
2204 /* Give the chain decl a distinct rank. */
2205 if (cfun->static_chain_decl != NULL)
2207 tree def = gimple_default_def (cfun, cfun->static_chain_decl);
2208 if (def != NULL)
2209 insert_operand_rank (def, ++rank);
2212 /* Set up rank for each BB */
2213 for (i = 0; i < n_basic_blocks - NUM_FIXED_BLOCKS; i++)
2214 bb_rank[bbs[i]] = ++rank << 16;
2216 free (bbs);
2217 calculate_dominance_info (CDI_POST_DOMINATORS);
2218 plus_negates = NULL;
2221 /* Cleanup after the reassociation pass, and print stats if
2222 requested. */
2224 static void
2225 fini_reassoc (void)
2227 statistics_counter_event (cfun, "Linearized",
2228 reassociate_stats.linearized);
2229 statistics_counter_event (cfun, "Constants eliminated",
2230 reassociate_stats.constants_eliminated);
2231 statistics_counter_event (cfun, "Ops eliminated",
2232 reassociate_stats.ops_eliminated);
2233 statistics_counter_event (cfun, "Statements rewritten",
2234 reassociate_stats.rewritten);
2236 pointer_map_destroy (operand_rank);
2237 free_alloc_pool (operand_entry_pool);
2238 free (bb_rank);
2239 VEC_free (tree, heap, plus_negates);
2240 free_dominance_info (CDI_POST_DOMINATORS);
2241 loop_optimizer_finalize ();
2244 /* Gate and execute functions for Reassociation. */
2246 static unsigned int
2247 execute_reassoc (void)
2249 init_reassoc ();
2251 do_reassoc ();
2252 repropagate_negates ();
2254 fini_reassoc ();
2255 return 0;
2258 static bool
2259 gate_tree_ssa_reassoc (void)
2261 return flag_tree_reassoc != 0;
2264 struct gimple_opt_pass pass_reassoc =
2267 GIMPLE_PASS,
2268 "reassoc", /* name */
2269 gate_tree_ssa_reassoc, /* gate */
2270 execute_reassoc, /* execute */
2271 NULL, /* sub */
2272 NULL, /* next */
2273 0, /* static_pass_number */
2274 TV_TREE_REASSOC, /* tv_id */
2275 PROP_cfg | PROP_ssa, /* properties_required */
2276 0, /* properties_provided */
2277 0, /* properties_destroyed */
2278 0, /* todo_flags_start */
2279 TODO_dump_func | TODO_ggc_collect | TODO_verify_ssa /* todo_flags_finish */