dumping cleanup phase 1 -- Removing TODO_dump_func
[official-gcc.git] / gcc / tree-ssa-phiopt.c
blob79e8e015fda22ce1021f8e5a343289ac1d8d71ab
1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published by the
9 Free Software Foundation; either version 3, or (at your option) any
10 later version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 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 "ggc.h"
26 #include "tree.h"
27 #include "flags.h"
28 #include "tm_p.h"
29 #include "basic-block.h"
30 #include "timevar.h"
31 #include "tree-flow.h"
32 #include "tree-pass.h"
33 #include "tree-dump.h"
34 #include "langhooks.h"
35 #include "pointer-set.h"
36 #include "domwalk.h"
37 #include "cfgloop.h"
38 #include "tree-data-ref.h"
40 static unsigned int tree_ssa_phiopt (void);
41 static unsigned int tree_ssa_phiopt_worker (bool);
42 static bool conditional_replacement (basic_block, basic_block,
43 edge, edge, gimple, tree, tree);
44 static bool value_replacement (basic_block, basic_block,
45 edge, edge, gimple, tree, tree);
46 static bool minmax_replacement (basic_block, basic_block,
47 edge, edge, gimple, tree, tree);
48 static bool abs_replacement (basic_block, basic_block,
49 edge, edge, gimple, tree, tree);
50 static bool cond_store_replacement (basic_block, basic_block, edge, edge,
51 struct pointer_set_t *);
52 static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block);
53 static struct pointer_set_t * get_non_trapping (void);
54 static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree);
56 /* This pass tries to replaces an if-then-else block with an
57 assignment. We have four kinds of transformations. Some of these
58 transformations are also performed by the ifcvt RTL optimizer.
60 Conditional Replacement
61 -----------------------
63 This transformation, implemented in conditional_replacement,
64 replaces
66 bb0:
67 if (cond) goto bb2; else goto bb1;
68 bb1:
69 bb2:
70 x = PHI <0 (bb1), 1 (bb0), ...>;
72 with
74 bb0:
75 x' = cond;
76 goto bb2;
77 bb2:
78 x = PHI <x' (bb0), ...>;
80 We remove bb1 as it becomes unreachable. This occurs often due to
81 gimplification of conditionals.
83 Value Replacement
84 -----------------
86 This transformation, implemented in value_replacement, replaces
88 bb0:
89 if (a != b) goto bb2; else goto bb1;
90 bb1:
91 bb2:
92 x = PHI <a (bb1), b (bb0), ...>;
94 with
96 bb0:
97 bb2:
98 x = PHI <b (bb0), ...>;
100 This opportunity can sometimes occur as a result of other
101 optimizations.
103 ABS Replacement
104 ---------------
106 This transformation, implemented in abs_replacement, replaces
108 bb0:
109 if (a >= 0) goto bb2; else goto bb1;
110 bb1:
111 x = -a;
112 bb2:
113 x = PHI <x (bb1), a (bb0), ...>;
115 with
117 bb0:
118 x' = ABS_EXPR< a >;
119 bb2:
120 x = PHI <x' (bb0), ...>;
122 MIN/MAX Replacement
123 -------------------
125 This transformation, minmax_replacement replaces
127 bb0:
128 if (a <= b) goto bb2; else goto bb1;
129 bb1:
130 bb2:
131 x = PHI <b (bb1), a (bb0), ...>;
133 with
135 bb0:
136 x' = MIN_EXPR (a, b)
137 bb2:
138 x = PHI <x' (bb0), ...>;
140 A similar transformation is done for MAX_EXPR. */
142 static unsigned int
143 tree_ssa_phiopt (void)
145 return tree_ssa_phiopt_worker (false);
148 /* This pass tries to transform conditional stores into unconditional
149 ones, enabling further simplifications with the simpler then and else
150 blocks. In particular it replaces this:
152 bb0:
153 if (cond) goto bb2; else goto bb1;
154 bb1:
155 *p = RHS;
156 bb2:
158 with
160 bb0:
161 if (cond) goto bb1; else goto bb2;
162 bb1:
163 condtmp' = *p;
164 bb2:
165 condtmp = PHI <RHS, condtmp'>
166 *p = condtmp;
168 This transformation can only be done under several constraints,
169 documented below. It also replaces:
171 bb0:
172 if (cond) goto bb2; else goto bb1;
173 bb1:
174 *p = RHS1;
175 goto bb3;
176 bb2:
177 *p = RHS2;
178 bb3:
180 with
182 bb0:
183 if (cond) goto bb3; else goto bb1;
184 bb1:
185 bb3:
186 condtmp = PHI <RHS1, RHS2>
187 *p = condtmp; */
189 static unsigned int
190 tree_ssa_cs_elim (void)
192 return tree_ssa_phiopt_worker (true);
195 /* For conditional store replacement we need a temporary to
196 put the old contents of the memory in. */
197 static tree condstoretemp;
199 /* The core routine of conditional store replacement and normal
200 phi optimizations. Both share much of the infrastructure in how
201 to match applicable basic block patterns. DO_STORE_ELIM is true
202 when we want to do conditional store replacement, false otherwise. */
203 static unsigned int
204 tree_ssa_phiopt_worker (bool do_store_elim)
206 basic_block bb;
207 basic_block *bb_order;
208 unsigned n, i;
209 bool cfgchanged = false;
210 struct pointer_set_t *nontrap = 0;
212 if (do_store_elim)
214 condstoretemp = NULL_TREE;
215 /* Calculate the set of non-trapping memory accesses. */
216 nontrap = get_non_trapping ();
219 /* Search every basic block for COND_EXPR we may be able to optimize.
221 We walk the blocks in order that guarantees that a block with
222 a single predecessor is processed before the predecessor.
223 This ensures that we collapse inner ifs before visiting the
224 outer ones, and also that we do not try to visit a removed
225 block. */
226 bb_order = blocks_in_phiopt_order ();
227 n = n_basic_blocks - NUM_FIXED_BLOCKS;
229 for (i = 0; i < n; i++)
231 gimple cond_stmt, phi;
232 basic_block bb1, bb2;
233 edge e1, e2;
234 tree arg0, arg1;
236 bb = bb_order[i];
238 cond_stmt = last_stmt (bb);
239 /* Check to see if the last statement is a GIMPLE_COND. */
240 if (!cond_stmt
241 || gimple_code (cond_stmt) != GIMPLE_COND)
242 continue;
244 e1 = EDGE_SUCC (bb, 0);
245 bb1 = e1->dest;
246 e2 = EDGE_SUCC (bb, 1);
247 bb2 = e2->dest;
249 /* We cannot do the optimization on abnormal edges. */
250 if ((e1->flags & EDGE_ABNORMAL) != 0
251 || (e2->flags & EDGE_ABNORMAL) != 0)
252 continue;
254 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
255 if (EDGE_COUNT (bb1->succs) == 0
256 || bb2 == NULL
257 || EDGE_COUNT (bb2->succs) == 0)
258 continue;
260 /* Find the bb which is the fall through to the other. */
261 if (EDGE_SUCC (bb1, 0)->dest == bb2)
263 else if (EDGE_SUCC (bb2, 0)->dest == bb1)
265 basic_block bb_tmp = bb1;
266 edge e_tmp = e1;
267 bb1 = bb2;
268 bb2 = bb_tmp;
269 e1 = e2;
270 e2 = e_tmp;
272 else if (do_store_elim
273 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
275 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
277 if (!single_succ_p (bb1)
278 || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0
279 || !single_succ_p (bb2)
280 || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0
281 || EDGE_COUNT (bb3->preds) != 2)
282 continue;
283 if (cond_if_else_store_replacement (bb1, bb2, bb3))
284 cfgchanged = true;
285 continue;
287 else
288 continue;
290 e1 = EDGE_SUCC (bb1, 0);
292 /* Make sure that bb1 is just a fall through. */
293 if (!single_succ_p (bb1)
294 || (e1->flags & EDGE_FALLTHRU) == 0)
295 continue;
297 /* Also make sure that bb1 only have one predecessor and that it
298 is bb. */
299 if (!single_pred_p (bb1)
300 || single_pred (bb1) != bb)
301 continue;
303 if (do_store_elim)
305 /* bb1 is the middle block, bb2 the join block, bb the split block,
306 e1 the fallthrough edge from bb1 to bb2. We can't do the
307 optimization if the join block has more than two predecessors. */
308 if (EDGE_COUNT (bb2->preds) > 2)
309 continue;
310 if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
311 cfgchanged = true;
313 else
315 gimple_seq phis = phi_nodes (bb2);
316 gimple_stmt_iterator gsi;
318 /* Check to make sure that there is only one non-virtual PHI node.
319 TODO: we could do it with more than one iff the other PHI nodes
320 have the same elements for these two edges. */
321 phi = NULL;
322 for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
324 if (!is_gimple_reg (gimple_phi_result (gsi_stmt (gsi))))
325 continue;
326 if (phi)
328 phi = NULL;
329 break;
331 phi = gsi_stmt (gsi);
333 if (!phi)
334 continue;
336 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
337 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
339 /* Something is wrong if we cannot find the arguments in the PHI
340 node. */
341 gcc_assert (arg0 != NULL && arg1 != NULL);
343 /* Do the replacement of conditional if it can be done. */
344 if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
345 cfgchanged = true;
346 else if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
347 cfgchanged = true;
348 else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
349 cfgchanged = true;
350 else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
351 cfgchanged = true;
355 free (bb_order);
357 if (do_store_elim)
358 pointer_set_destroy (nontrap);
359 /* If the CFG has changed, we should cleanup the CFG. */
360 if (cfgchanged && do_store_elim)
362 /* In cond-store replacement we have added some loads on edges
363 and new VOPS (as we moved the store, and created a load). */
364 gsi_commit_edge_inserts ();
365 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
367 else if (cfgchanged)
368 return TODO_cleanup_cfg;
369 return 0;
372 /* Returns the list of basic blocks in the function in an order that guarantees
373 that if a block X has just a single predecessor Y, then Y is after X in the
374 ordering. */
376 basic_block *
377 blocks_in_phiopt_order (void)
379 basic_block x, y;
380 basic_block *order = XNEWVEC (basic_block, n_basic_blocks);
381 unsigned n = n_basic_blocks - NUM_FIXED_BLOCKS;
382 unsigned np, i;
383 sbitmap visited = sbitmap_alloc (last_basic_block);
385 #define MARK_VISITED(BB) (SET_BIT (visited, (BB)->index))
386 #define VISITED_P(BB) (TEST_BIT (visited, (BB)->index))
388 sbitmap_zero (visited);
390 MARK_VISITED (ENTRY_BLOCK_PTR);
391 FOR_EACH_BB (x)
393 if (VISITED_P (x))
394 continue;
396 /* Walk the predecessors of x as long as they have precisely one
397 predecessor and add them to the list, so that they get stored
398 after x. */
399 for (y = x, np = 1;
400 single_pred_p (y) && !VISITED_P (single_pred (y));
401 y = single_pred (y))
402 np++;
403 for (y = x, i = n - np;
404 single_pred_p (y) && !VISITED_P (single_pred (y));
405 y = single_pred (y), i++)
407 order[i] = y;
408 MARK_VISITED (y);
410 order[i] = y;
411 MARK_VISITED (y);
413 gcc_assert (i == n - 1);
414 n -= np;
417 sbitmap_free (visited);
418 gcc_assert (n == 0);
419 return order;
421 #undef MARK_VISITED
422 #undef VISITED_P
426 /* Return TRUE if block BB has no executable statements, otherwise return
427 FALSE. */
429 bool
430 empty_block_p (basic_block bb)
432 /* BB must have no executable statements. */
433 gimple_stmt_iterator gsi = gsi_after_labels (bb);
434 if (gsi_end_p (gsi))
435 return true;
436 if (is_gimple_debug (gsi_stmt (gsi)))
437 gsi_next_nondebug (&gsi);
438 return gsi_end_p (gsi);
441 /* Replace PHI node element whose edge is E in block BB with variable NEW.
442 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
443 is known to have two edges, one of which must reach BB). */
445 static void
446 replace_phi_edge_with_variable (basic_block cond_block,
447 edge e, gimple phi, tree new_tree)
449 basic_block bb = gimple_bb (phi);
450 basic_block block_to_remove;
451 gimple_stmt_iterator gsi;
453 /* Change the PHI argument to new. */
454 SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree);
456 /* Remove the empty basic block. */
457 if (EDGE_SUCC (cond_block, 0)->dest == bb)
459 EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU;
460 EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
461 EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE;
462 EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count;
464 block_to_remove = EDGE_SUCC (cond_block, 1)->dest;
466 else
468 EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU;
469 EDGE_SUCC (cond_block, 1)->flags
470 &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
471 EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE;
472 EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count;
474 block_to_remove = EDGE_SUCC (cond_block, 0)->dest;
476 delete_basic_block (block_to_remove);
478 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
479 gsi = gsi_last_bb (cond_block);
480 gsi_remove (&gsi, true);
482 if (dump_file && (dump_flags & TDF_DETAILS))
483 fprintf (dump_file,
484 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
485 cond_block->index,
486 bb->index);
489 /* The function conditional_replacement does the main work of doing the
490 conditional replacement. Return true if the replacement is done.
491 Otherwise return false.
492 BB is the basic block where the replacement is going to be done on. ARG0
493 is argument 0 from PHI. Likewise for ARG1. */
495 static bool
496 conditional_replacement (basic_block cond_bb, basic_block middle_bb,
497 edge e0, edge e1, gimple phi,
498 tree arg0, tree arg1)
500 tree result;
501 gimple stmt, new_stmt;
502 tree cond;
503 gimple_stmt_iterator gsi;
504 edge true_edge, false_edge;
505 tree new_var, new_var2;
507 /* FIXME: Gimplification of complex type is too hard for now. */
508 if (TREE_CODE (TREE_TYPE (arg0)) == COMPLEX_TYPE
509 || TREE_CODE (TREE_TYPE (arg1)) == COMPLEX_TYPE)
510 return false;
512 /* The PHI arguments have the constants 0 and 1, then convert
513 it to the conditional. */
514 if ((integer_zerop (arg0) && integer_onep (arg1))
515 || (integer_zerop (arg1) && integer_onep (arg0)))
517 else
518 return false;
520 if (!empty_block_p (middle_bb))
521 return false;
523 /* At this point we know we have a GIMPLE_COND with two successors.
524 One successor is BB, the other successor is an empty block which
525 falls through into BB.
527 There is a single PHI node at the join point (BB) and its arguments
528 are constants (0, 1).
530 So, given the condition COND, and the two PHI arguments, we can
531 rewrite this PHI into non-branching code:
533 dest = (COND) or dest = COND'
535 We use the condition as-is if the argument associated with the
536 true edge has the value one or the argument associated with the
537 false edge as the value zero. Note that those conditions are not
538 the same since only one of the outgoing edges from the GIMPLE_COND
539 will directly reach BB and thus be associated with an argument. */
541 stmt = last_stmt (cond_bb);
542 result = PHI_RESULT (phi);
544 /* To handle special cases like floating point comparison, it is easier and
545 less error-prone to build a tree and gimplify it on the fly though it is
546 less efficient. */
547 cond = fold_build2 (gimple_cond_code (stmt), boolean_type_node,
548 gimple_cond_lhs (stmt), gimple_cond_rhs (stmt));
550 /* We need to know which is the true edge and which is the false
551 edge so that we know when to invert the condition below. */
552 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
553 if ((e0 == true_edge && integer_zerop (arg0))
554 || (e0 == false_edge && integer_onep (arg0))
555 || (e1 == true_edge && integer_zerop (arg1))
556 || (e1 == false_edge && integer_onep (arg1)))
557 cond = fold_build1 (TRUTH_NOT_EXPR, TREE_TYPE (cond), cond);
559 /* Insert our new statements at the end of conditional block before the
560 COND_STMT. */
561 gsi = gsi_for_stmt (stmt);
562 new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true,
563 GSI_SAME_STMT);
565 if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var)))
567 source_location locus_0, locus_1;
569 new_var2 = create_tmp_var (TREE_TYPE (result), NULL);
570 add_referenced_var (new_var2);
571 new_stmt = gimple_build_assign_with_ops (CONVERT_EXPR, new_var2,
572 new_var, NULL);
573 new_var2 = make_ssa_name (new_var2, new_stmt);
574 gimple_assign_set_lhs (new_stmt, new_var2);
575 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
576 new_var = new_var2;
578 /* Set the locus to the first argument, unless is doesn't have one. */
579 locus_0 = gimple_phi_arg_location (phi, 0);
580 locus_1 = gimple_phi_arg_location (phi, 1);
581 if (locus_0 == UNKNOWN_LOCATION)
582 locus_0 = locus_1;
583 gimple_set_location (new_stmt, locus_0);
586 replace_phi_edge_with_variable (cond_bb, e1, phi, new_var);
588 /* Note that we optimized this PHI. */
589 return true;
592 /* The function value_replacement does the main work of doing the value
593 replacement. Return true if the replacement is done. Otherwise return
594 false.
595 BB is the basic block where the replacement is going to be done on. ARG0
596 is argument 0 from the PHI. Likewise for ARG1. */
598 static bool
599 value_replacement (basic_block cond_bb, basic_block middle_bb,
600 edge e0, edge e1, gimple phi,
601 tree arg0, tree arg1)
603 gimple cond;
604 edge true_edge, false_edge;
605 enum tree_code code;
607 /* If the type says honor signed zeros we cannot do this
608 optimization. */
609 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
610 return false;
612 if (!empty_block_p (middle_bb))
613 return false;
615 cond = last_stmt (cond_bb);
616 code = gimple_cond_code (cond);
618 /* This transformation is only valid for equality comparisons. */
619 if (code != NE_EXPR && code != EQ_EXPR)
620 return false;
622 /* We need to know which is the true edge and which is the false
623 edge so that we know if have abs or negative abs. */
624 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
626 /* At this point we know we have a COND_EXPR with two successors.
627 One successor is BB, the other successor is an empty block which
628 falls through into BB.
630 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
632 There is a single PHI node at the join point (BB) with two arguments.
634 We now need to verify that the two arguments in the PHI node match
635 the two arguments to the equality comparison. */
637 if ((operand_equal_for_phi_arg_p (arg0, gimple_cond_lhs (cond))
638 && operand_equal_for_phi_arg_p (arg1, gimple_cond_rhs (cond)))
639 || (operand_equal_for_phi_arg_p (arg1, gimple_cond_lhs (cond))
640 && operand_equal_for_phi_arg_p (arg0, gimple_cond_rhs (cond))))
642 edge e;
643 tree arg;
645 /* For NE_EXPR, we want to build an assignment result = arg where
646 arg is the PHI argument associated with the true edge. For
647 EQ_EXPR we want the PHI argument associated with the false edge. */
648 e = (code == NE_EXPR ? true_edge : false_edge);
650 /* Unfortunately, E may not reach BB (it may instead have gone to
651 OTHER_BLOCK). If that is the case, then we want the single outgoing
652 edge from OTHER_BLOCK which reaches BB and represents the desired
653 path from COND_BLOCK. */
654 if (e->dest == middle_bb)
655 e = single_succ_edge (e->dest);
657 /* Now we know the incoming edge to BB that has the argument for the
658 RHS of our new assignment statement. */
659 if (e0 == e)
660 arg = arg0;
661 else
662 arg = arg1;
664 replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
666 /* Note that we optimized this PHI. */
667 return true;
669 return false;
672 /* The function minmax_replacement does the main work of doing the minmax
673 replacement. Return true if the replacement is done. Otherwise return
674 false.
675 BB is the basic block where the replacement is going to be done on. ARG0
676 is argument 0 from the PHI. Likewise for ARG1. */
678 static bool
679 minmax_replacement (basic_block cond_bb, basic_block middle_bb,
680 edge e0, edge e1, gimple phi,
681 tree arg0, tree arg1)
683 tree result, type;
684 gimple cond, new_stmt;
685 edge true_edge, false_edge;
686 enum tree_code cmp, minmax, ass_code;
687 tree smaller, larger, arg_true, arg_false;
688 gimple_stmt_iterator gsi, gsi_from;
690 type = TREE_TYPE (PHI_RESULT (phi));
692 /* The optimization may be unsafe due to NaNs. */
693 if (HONOR_NANS (TYPE_MODE (type)))
694 return false;
696 cond = last_stmt (cond_bb);
697 cmp = gimple_cond_code (cond);
699 /* This transformation is only valid for order comparisons. Record which
700 operand is smaller/larger if the result of the comparison is true. */
701 if (cmp == LT_EXPR || cmp == LE_EXPR)
703 smaller = gimple_cond_lhs (cond);
704 larger = gimple_cond_rhs (cond);
706 else if (cmp == GT_EXPR || cmp == GE_EXPR)
708 smaller = gimple_cond_rhs (cond);
709 larger = gimple_cond_lhs (cond);
711 else
712 return false;
714 /* We need to know which is the true edge and which is the false
715 edge so that we know if have abs or negative abs. */
716 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
718 /* Forward the edges over the middle basic block. */
719 if (true_edge->dest == middle_bb)
720 true_edge = EDGE_SUCC (true_edge->dest, 0);
721 if (false_edge->dest == middle_bb)
722 false_edge = EDGE_SUCC (false_edge->dest, 0);
724 if (true_edge == e0)
726 gcc_assert (false_edge == e1);
727 arg_true = arg0;
728 arg_false = arg1;
730 else
732 gcc_assert (false_edge == e0);
733 gcc_assert (true_edge == e1);
734 arg_true = arg1;
735 arg_false = arg0;
738 if (empty_block_p (middle_bb))
740 if (operand_equal_for_phi_arg_p (arg_true, smaller)
741 && operand_equal_for_phi_arg_p (arg_false, larger))
743 /* Case
745 if (smaller < larger)
746 rslt = smaller;
747 else
748 rslt = larger; */
749 minmax = MIN_EXPR;
751 else if (operand_equal_for_phi_arg_p (arg_false, smaller)
752 && operand_equal_for_phi_arg_p (arg_true, larger))
753 minmax = MAX_EXPR;
754 else
755 return false;
757 else
759 /* Recognize the following case, assuming d <= u:
761 if (a <= u)
762 b = MAX (a, d);
763 x = PHI <b, u>
765 This is equivalent to
767 b = MAX (a, d);
768 x = MIN (b, u); */
770 gimple assign = last_and_only_stmt (middle_bb);
771 tree lhs, op0, op1, bound;
773 if (!assign
774 || gimple_code (assign) != GIMPLE_ASSIGN)
775 return false;
777 lhs = gimple_assign_lhs (assign);
778 ass_code = gimple_assign_rhs_code (assign);
779 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
780 return false;
781 op0 = gimple_assign_rhs1 (assign);
782 op1 = gimple_assign_rhs2 (assign);
784 if (true_edge->src == middle_bb)
786 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
787 if (!operand_equal_for_phi_arg_p (lhs, arg_true))
788 return false;
790 if (operand_equal_for_phi_arg_p (arg_false, larger))
792 /* Case
794 if (smaller < larger)
796 r' = MAX_EXPR (smaller, bound)
798 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
799 if (ass_code != MAX_EXPR)
800 return false;
802 minmax = MIN_EXPR;
803 if (operand_equal_for_phi_arg_p (op0, smaller))
804 bound = op1;
805 else if (operand_equal_for_phi_arg_p (op1, smaller))
806 bound = op0;
807 else
808 return false;
810 /* We need BOUND <= LARGER. */
811 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
812 bound, larger)))
813 return false;
815 else if (operand_equal_for_phi_arg_p (arg_false, smaller))
817 /* Case
819 if (smaller < larger)
821 r' = MIN_EXPR (larger, bound)
823 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
824 if (ass_code != MIN_EXPR)
825 return false;
827 minmax = MAX_EXPR;
828 if (operand_equal_for_phi_arg_p (op0, larger))
829 bound = op1;
830 else if (operand_equal_for_phi_arg_p (op1, larger))
831 bound = op0;
832 else
833 return false;
835 /* We need BOUND >= SMALLER. */
836 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
837 bound, smaller)))
838 return false;
840 else
841 return false;
843 else
845 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
846 if (!operand_equal_for_phi_arg_p (lhs, arg_false))
847 return false;
849 if (operand_equal_for_phi_arg_p (arg_true, larger))
851 /* Case
853 if (smaller > larger)
855 r' = MIN_EXPR (smaller, bound)
857 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
858 if (ass_code != MIN_EXPR)
859 return false;
861 minmax = MAX_EXPR;
862 if (operand_equal_for_phi_arg_p (op0, smaller))
863 bound = op1;
864 else if (operand_equal_for_phi_arg_p (op1, smaller))
865 bound = op0;
866 else
867 return false;
869 /* We need BOUND >= LARGER. */
870 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
871 bound, larger)))
872 return false;
874 else if (operand_equal_for_phi_arg_p (arg_true, smaller))
876 /* Case
878 if (smaller > larger)
880 r' = MAX_EXPR (larger, bound)
882 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
883 if (ass_code != MAX_EXPR)
884 return false;
886 minmax = MIN_EXPR;
887 if (operand_equal_for_phi_arg_p (op0, larger))
888 bound = op1;
889 else if (operand_equal_for_phi_arg_p (op1, larger))
890 bound = op0;
891 else
892 return false;
894 /* We need BOUND <= SMALLER. */
895 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
896 bound, smaller)))
897 return false;
899 else
900 return false;
903 /* Move the statement from the middle block. */
904 gsi = gsi_last_bb (cond_bb);
905 gsi_from = gsi_last_nondebug_bb (middle_bb);
906 gsi_move_before (&gsi_from, &gsi);
909 /* Emit the statement to compute min/max. */
910 result = duplicate_ssa_name (PHI_RESULT (phi), NULL);
911 new_stmt = gimple_build_assign_with_ops (minmax, result, arg0, arg1);
912 gsi = gsi_last_bb (cond_bb);
913 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
915 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
916 return true;
919 /* The function absolute_replacement does the main work of doing the absolute
920 replacement. Return true if the replacement is done. Otherwise return
921 false.
922 bb is the basic block where the replacement is going to be done on. arg0
923 is argument 0 from the phi. Likewise for arg1. */
925 static bool
926 abs_replacement (basic_block cond_bb, basic_block middle_bb,
927 edge e0 ATTRIBUTE_UNUSED, edge e1,
928 gimple phi, tree arg0, tree arg1)
930 tree result;
931 gimple new_stmt, cond;
932 gimple_stmt_iterator gsi;
933 edge true_edge, false_edge;
934 gimple assign;
935 edge e;
936 tree rhs, lhs;
937 bool negate;
938 enum tree_code cond_code;
940 /* If the type says honor signed zeros we cannot do this
941 optimization. */
942 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
943 return false;
945 /* OTHER_BLOCK must have only one executable statement which must have the
946 form arg0 = -arg1 or arg1 = -arg0. */
948 assign = last_and_only_stmt (middle_bb);
949 /* If we did not find the proper negation assignment, then we can not
950 optimize. */
951 if (assign == NULL)
952 return false;
954 /* If we got here, then we have found the only executable statement
955 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
956 arg1 = -arg0, then we can not optimize. */
957 if (gimple_code (assign) != GIMPLE_ASSIGN)
958 return false;
960 lhs = gimple_assign_lhs (assign);
962 if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
963 return false;
965 rhs = gimple_assign_rhs1 (assign);
967 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
968 if (!(lhs == arg0 && rhs == arg1)
969 && !(lhs == arg1 && rhs == arg0))
970 return false;
972 cond = last_stmt (cond_bb);
973 result = PHI_RESULT (phi);
975 /* Only relationals comparing arg[01] against zero are interesting. */
976 cond_code = gimple_cond_code (cond);
977 if (cond_code != GT_EXPR && cond_code != GE_EXPR
978 && cond_code != LT_EXPR && cond_code != LE_EXPR)
979 return false;
981 /* Make sure the conditional is arg[01] OP y. */
982 if (gimple_cond_lhs (cond) != rhs)
983 return false;
985 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond)))
986 ? real_zerop (gimple_cond_rhs (cond))
987 : integer_zerop (gimple_cond_rhs (cond)))
989 else
990 return false;
992 /* We need to know which is the true edge and which is the false
993 edge so that we know if have abs or negative abs. */
994 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
996 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
997 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
998 the false edge goes to OTHER_BLOCK. */
999 if (cond_code == GT_EXPR || cond_code == GE_EXPR)
1000 e = true_edge;
1001 else
1002 e = false_edge;
1004 if (e->dest == middle_bb)
1005 negate = true;
1006 else
1007 negate = false;
1009 result = duplicate_ssa_name (result, NULL);
1011 if (negate)
1013 tree tmp = create_tmp_var (TREE_TYPE (result), NULL);
1014 add_referenced_var (tmp);
1015 lhs = make_ssa_name (tmp, NULL);
1017 else
1018 lhs = result;
1020 /* Build the modify expression with abs expression. */
1021 new_stmt = gimple_build_assign_with_ops (ABS_EXPR, lhs, rhs, NULL);
1023 gsi = gsi_last_bb (cond_bb);
1024 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1026 if (negate)
1028 /* Get the right GSI. We want to insert after the recently
1029 added ABS_EXPR statement (which we know is the first statement
1030 in the block. */
1031 new_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, result, lhs, NULL);
1033 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1036 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1038 /* Note that we optimized this PHI. */
1039 return true;
1042 /* Auxiliary functions to determine the set of memory accesses which
1043 can't trap because they are preceded by accesses to the same memory
1044 portion. We do that for MEM_REFs, so we only need to track
1045 the SSA_NAME of the pointer indirectly referenced. The algorithm
1046 simply is a walk over all instructions in dominator order. When
1047 we see an MEM_REF we determine if we've already seen a same
1048 ref anywhere up to the root of the dominator tree. If we do the
1049 current access can't trap. If we don't see any dominating access
1050 the current access might trap, but might also make later accesses
1051 non-trapping, so we remember it. We need to be careful with loads
1052 or stores, for instance a load might not trap, while a store would,
1053 so if we see a dominating read access this doesn't mean that a later
1054 write access would not trap. Hence we also need to differentiate the
1055 type of access(es) seen.
1057 ??? We currently are very conservative and assume that a load might
1058 trap even if a store doesn't (write-only memory). This probably is
1059 overly conservative. */
1061 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1062 through it was seen, which would constitute a no-trap region for
1063 same accesses. */
1064 struct name_to_bb
1066 tree ssa_name;
1067 basic_block bb;
1068 unsigned store : 1;
1071 /* The hash table for remembering what we've seen. */
1072 static htab_t seen_ssa_names;
1074 /* The set of MEM_REFs which can't trap. */
1075 static struct pointer_set_t *nontrap_set;
1077 /* The hash function, based on the pointer to the pointer SSA_NAME. */
1078 static hashval_t
1079 name_to_bb_hash (const void *p)
1081 const_tree n = ((const struct name_to_bb *)p)->ssa_name;
1082 return htab_hash_pointer (n) ^ ((const struct name_to_bb *)p)->store;
1085 /* The equality function of *P1 and *P2. SSA_NAMEs are shared, so
1086 it's enough to simply compare them for equality. */
1087 static int
1088 name_to_bb_eq (const void *p1, const void *p2)
1090 const struct name_to_bb *n1 = (const struct name_to_bb *)p1;
1091 const struct name_to_bb *n2 = (const struct name_to_bb *)p2;
1093 return n1->ssa_name == n2->ssa_name && n1->store == n2->store;
1096 /* We see the expression EXP in basic block BB. If it's an interesting
1097 expression (an MEM_REF through an SSA_NAME) possibly insert the
1098 expression into the set NONTRAP or the hash table of seen expressions.
1099 STORE is true if this expression is on the LHS, otherwise it's on
1100 the RHS. */
1101 static void
1102 add_or_mark_expr (basic_block bb, tree exp,
1103 struct pointer_set_t *nontrap, bool store)
1105 if (TREE_CODE (exp) == MEM_REF
1106 && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME)
1108 tree name = TREE_OPERAND (exp, 0);
1109 struct name_to_bb map;
1110 void **slot;
1111 struct name_to_bb *n2bb;
1112 basic_block found_bb = 0;
1114 /* Try to find the last seen MEM_REF through the same
1115 SSA_NAME, which can trap. */
1116 map.ssa_name = name;
1117 map.bb = 0;
1118 map.store = store;
1119 slot = htab_find_slot (seen_ssa_names, &map, INSERT);
1120 n2bb = (struct name_to_bb *) *slot;
1121 if (n2bb)
1122 found_bb = n2bb->bb;
1124 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1125 (it's in a basic block on the path from us to the dominator root)
1126 then we can't trap. */
1127 if (found_bb && found_bb->aux == (void *)1)
1129 pointer_set_insert (nontrap, exp);
1131 else
1133 /* EXP might trap, so insert it into the hash table. */
1134 if (n2bb)
1136 n2bb->bb = bb;
1138 else
1140 n2bb = XNEW (struct name_to_bb);
1141 n2bb->ssa_name = name;
1142 n2bb->bb = bb;
1143 n2bb->store = store;
1144 *slot = n2bb;
1150 /* Called by walk_dominator_tree, when entering the block BB. */
1151 static void
1152 nt_init_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb)
1154 gimple_stmt_iterator gsi;
1155 /* Mark this BB as being on the path to dominator root. */
1156 bb->aux = (void*)1;
1158 /* And walk the statements in order. */
1159 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1161 gimple stmt = gsi_stmt (gsi);
1163 if (is_gimple_assign (stmt))
1165 add_or_mark_expr (bb, gimple_assign_lhs (stmt), nontrap_set, true);
1166 add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), nontrap_set, false);
1167 if (get_gimple_rhs_num_ops (gimple_assign_rhs_code (stmt)) > 1)
1168 add_or_mark_expr (bb, gimple_assign_rhs2 (stmt), nontrap_set,
1169 false);
1174 /* Called by walk_dominator_tree, when basic block BB is exited. */
1175 static void
1176 nt_fini_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb)
1178 /* This BB isn't on the path to dominator root anymore. */
1179 bb->aux = NULL;
1182 /* This is the entry point of gathering non trapping memory accesses.
1183 It will do a dominator walk over the whole function, and it will
1184 make use of the bb->aux pointers. It returns a set of trees
1185 (the MEM_REFs itself) which can't trap. */
1186 static struct pointer_set_t *
1187 get_non_trapping (void)
1189 struct pointer_set_t *nontrap;
1190 struct dom_walk_data walk_data;
1192 nontrap = pointer_set_create ();
1193 seen_ssa_names = htab_create (128, name_to_bb_hash, name_to_bb_eq,
1194 free);
1195 /* We're going to do a dominator walk, so ensure that we have
1196 dominance information. */
1197 calculate_dominance_info (CDI_DOMINATORS);
1199 /* Setup callbacks for the generic dominator tree walker. */
1200 nontrap_set = nontrap;
1201 walk_data.dom_direction = CDI_DOMINATORS;
1202 walk_data.initialize_block_local_data = NULL;
1203 walk_data.before_dom_children = nt_init_block;
1204 walk_data.after_dom_children = nt_fini_block;
1205 walk_data.global_data = NULL;
1206 walk_data.block_local_data_size = 0;
1208 init_walk_dominator_tree (&walk_data);
1209 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
1210 fini_walk_dominator_tree (&walk_data);
1211 htab_delete (seen_ssa_names);
1213 return nontrap;
1216 /* Do the main work of conditional store replacement. We already know
1217 that the recognized pattern looks like so:
1219 split:
1220 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1221 MIDDLE_BB:
1222 something
1223 fallthrough (edge E0)
1224 JOIN_BB:
1225 some more
1227 We check that MIDDLE_BB contains only one store, that that store
1228 doesn't trap (not via NOTRAP, but via checking if an access to the same
1229 memory location dominates us) and that the store has a "simple" RHS. */
1231 static bool
1232 cond_store_replacement (basic_block middle_bb, basic_block join_bb,
1233 edge e0, edge e1, struct pointer_set_t *nontrap)
1235 gimple assign = last_and_only_stmt (middle_bb);
1236 tree lhs, rhs, name;
1237 gimple newphi, new_stmt;
1238 gimple_stmt_iterator gsi;
1239 source_location locus;
1241 /* Check if middle_bb contains of only one store. */
1242 if (!assign
1243 || !gimple_assign_single_p (assign))
1244 return false;
1246 locus = gimple_location (assign);
1247 lhs = gimple_assign_lhs (assign);
1248 rhs = gimple_assign_rhs1 (assign);
1249 if (TREE_CODE (lhs) != MEM_REF
1250 || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME
1251 || !is_gimple_reg_type (TREE_TYPE (lhs)))
1252 return false;
1254 /* Prove that we can move the store down. We could also check
1255 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1256 whose value is not available readily, which we want to avoid. */
1257 if (!pointer_set_contains (nontrap, lhs))
1258 return false;
1260 /* Now we've checked the constraints, so do the transformation:
1261 1) Remove the single store. */
1262 gsi = gsi_for_stmt (assign);
1263 unlink_stmt_vdef (assign);
1264 gsi_remove (&gsi, true);
1265 release_defs (assign);
1267 /* 2) Create a temporary where we can store the old content
1268 of the memory touched by the store, if we need to. */
1269 if (!condstoretemp || TREE_TYPE (lhs) != TREE_TYPE (condstoretemp))
1271 condstoretemp = create_tmp_reg (TREE_TYPE (lhs), "cstore");
1272 get_var_ann (condstoretemp);
1274 add_referenced_var (condstoretemp);
1276 /* 3) Insert a load from the memory of the store to the temporary
1277 on the edge which did not contain the store. */
1278 lhs = unshare_expr (lhs);
1279 new_stmt = gimple_build_assign (condstoretemp, lhs);
1280 name = make_ssa_name (condstoretemp, new_stmt);
1281 gimple_assign_set_lhs (new_stmt, name);
1282 gimple_set_location (new_stmt, locus);
1283 gsi_insert_on_edge (e1, new_stmt);
1285 /* 4) Create a PHI node at the join block, with one argument
1286 holding the old RHS, and the other holding the temporary
1287 where we stored the old memory contents. */
1288 newphi = create_phi_node (condstoretemp, join_bb);
1289 add_phi_arg (newphi, rhs, e0, locus);
1290 add_phi_arg (newphi, name, e1, locus);
1292 lhs = unshare_expr (lhs);
1293 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1295 /* 5) Insert that PHI node. */
1296 gsi = gsi_after_labels (join_bb);
1297 if (gsi_end_p (gsi))
1299 gsi = gsi_last_bb (join_bb);
1300 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1302 else
1303 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1305 return true;
1308 /* Do the main work of conditional store replacement. */
1310 static bool
1311 cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb,
1312 basic_block join_bb, gimple then_assign,
1313 gimple else_assign)
1315 tree lhs_base, lhs, then_rhs, else_rhs;
1316 source_location then_locus, else_locus;
1317 gimple_stmt_iterator gsi;
1318 gimple newphi, new_stmt;
1320 if (then_assign == NULL
1321 || !gimple_assign_single_p (then_assign)
1322 || else_assign == NULL
1323 || !gimple_assign_single_p (else_assign))
1324 return false;
1326 lhs = gimple_assign_lhs (then_assign);
1327 if (!is_gimple_reg_type (TREE_TYPE (lhs))
1328 || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0))
1329 return false;
1331 lhs_base = get_base_address (lhs);
1332 if (lhs_base == NULL_TREE
1333 || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF))
1334 return false;
1336 then_rhs = gimple_assign_rhs1 (then_assign);
1337 else_rhs = gimple_assign_rhs1 (else_assign);
1338 then_locus = gimple_location (then_assign);
1339 else_locus = gimple_location (else_assign);
1341 /* Now we've checked the constraints, so do the transformation:
1342 1) Remove the stores. */
1343 gsi = gsi_for_stmt (then_assign);
1344 unlink_stmt_vdef (then_assign);
1345 gsi_remove (&gsi, true);
1346 release_defs (then_assign);
1348 gsi = gsi_for_stmt (else_assign);
1349 unlink_stmt_vdef (else_assign);
1350 gsi_remove (&gsi, true);
1351 release_defs (else_assign);
1353 /* 2) Create a temporary where we can store the old content
1354 of the memory touched by the store, if we need to. */
1355 if (!condstoretemp || TREE_TYPE (lhs) != TREE_TYPE (condstoretemp))
1357 condstoretemp = create_tmp_reg (TREE_TYPE (lhs), "cstore");
1358 get_var_ann (condstoretemp);
1360 add_referenced_var (condstoretemp);
1362 /* 3) Create a PHI node at the join block, with one argument
1363 holding the old RHS, and the other holding the temporary
1364 where we stored the old memory contents. */
1365 newphi = create_phi_node (condstoretemp, join_bb);
1366 add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus);
1367 add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus);
1369 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1371 /* 4) Insert that PHI node. */
1372 gsi = gsi_after_labels (join_bb);
1373 if (gsi_end_p (gsi))
1375 gsi = gsi_last_bb (join_bb);
1376 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1378 else
1379 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1381 return true;
1384 /* Conditional store replacement. We already know
1385 that the recognized pattern looks like so:
1387 split:
1388 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1389 THEN_BB:
1391 X = Y;
1393 goto JOIN_BB;
1394 ELSE_BB:
1396 X = Z;
1398 fallthrough (edge E0)
1399 JOIN_BB:
1400 some more
1402 We check that it is safe to sink the store to JOIN_BB by verifying that
1403 there are no read-after-write or write-after-write dependencies in
1404 THEN_BB and ELSE_BB. */
1406 static bool
1407 cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
1408 basic_block join_bb)
1410 gimple then_assign = last_and_only_stmt (then_bb);
1411 gimple else_assign = last_and_only_stmt (else_bb);
1412 VEC (data_reference_p, heap) *then_datarefs, *else_datarefs;
1413 VEC (ddr_p, heap) *then_ddrs, *else_ddrs;
1414 gimple then_store, else_store;
1415 bool found, ok = false, res;
1416 struct data_dependence_relation *ddr;
1417 data_reference_p then_dr, else_dr;
1418 int i, j;
1419 tree then_lhs, else_lhs;
1420 VEC (gimple, heap) *then_stores, *else_stores;
1421 basic_block blocks[3];
1423 if (MAX_STORES_TO_SINK == 0)
1424 return false;
1426 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1427 if (then_assign && else_assign)
1428 return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1429 then_assign, else_assign);
1431 /* Find data references. */
1432 then_datarefs = VEC_alloc (data_reference_p, heap, 1);
1433 else_datarefs = VEC_alloc (data_reference_p, heap, 1);
1434 if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs)
1435 == chrec_dont_know)
1436 || !VEC_length (data_reference_p, then_datarefs)
1437 || (find_data_references_in_bb (NULL, else_bb, &else_datarefs)
1438 == chrec_dont_know)
1439 || !VEC_length (data_reference_p, else_datarefs))
1441 free_data_refs (then_datarefs);
1442 free_data_refs (else_datarefs);
1443 return false;
1446 /* Find pairs of stores with equal LHS. */
1447 then_stores = VEC_alloc (gimple, heap, 1);
1448 else_stores = VEC_alloc (gimple, heap, 1);
1449 FOR_EACH_VEC_ELT (data_reference_p, then_datarefs, i, then_dr)
1451 if (DR_IS_READ (then_dr))
1452 continue;
1454 then_store = DR_STMT (then_dr);
1455 then_lhs = gimple_assign_lhs (then_store);
1456 found = false;
1458 FOR_EACH_VEC_ELT (data_reference_p, else_datarefs, j, else_dr)
1460 if (DR_IS_READ (else_dr))
1461 continue;
1463 else_store = DR_STMT (else_dr);
1464 else_lhs = gimple_assign_lhs (else_store);
1466 if (operand_equal_p (then_lhs, else_lhs, 0))
1468 found = true;
1469 break;
1473 if (!found)
1474 continue;
1476 VEC_safe_push (gimple, heap, then_stores, then_store);
1477 VEC_safe_push (gimple, heap, else_stores, else_store);
1480 /* No pairs of stores found. */
1481 if (!VEC_length (gimple, then_stores)
1482 || VEC_length (gimple, then_stores) > (unsigned) MAX_STORES_TO_SINK)
1484 free_data_refs (then_datarefs);
1485 free_data_refs (else_datarefs);
1486 VEC_free (gimple, heap, then_stores);
1487 VEC_free (gimple, heap, else_stores);
1488 return false;
1491 /* Compute and check data dependencies in both basic blocks. */
1492 then_ddrs = VEC_alloc (ddr_p, heap, 1);
1493 else_ddrs = VEC_alloc (ddr_p, heap, 1);
1494 compute_all_dependences (then_datarefs, &then_ddrs, NULL, false);
1495 compute_all_dependences (else_datarefs, &else_ddrs, NULL, false);
1496 blocks[0] = then_bb;
1497 blocks[1] = else_bb;
1498 blocks[2] = join_bb;
1499 renumber_gimple_stmt_uids_in_blocks (blocks, 3);
1501 /* Check that there are no read-after-write or write-after-write dependencies
1502 in THEN_BB. */
1503 FOR_EACH_VEC_ELT (ddr_p, then_ddrs, i, ddr)
1505 struct data_reference *dra = DDR_A (ddr);
1506 struct data_reference *drb = DDR_B (ddr);
1508 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1509 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1510 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1511 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1512 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1513 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1515 free_dependence_relations (then_ddrs);
1516 free_dependence_relations (else_ddrs);
1517 free_data_refs (then_datarefs);
1518 free_data_refs (else_datarefs);
1519 VEC_free (gimple, heap, then_stores);
1520 VEC_free (gimple, heap, else_stores);
1521 return false;
1525 /* Check that there are no read-after-write or write-after-write dependencies
1526 in ELSE_BB. */
1527 FOR_EACH_VEC_ELT (ddr_p, else_ddrs, i, ddr)
1529 struct data_reference *dra = DDR_A (ddr);
1530 struct data_reference *drb = DDR_B (ddr);
1532 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1533 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1534 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1535 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1536 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1537 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1539 free_dependence_relations (then_ddrs);
1540 free_dependence_relations (else_ddrs);
1541 free_data_refs (then_datarefs);
1542 free_data_refs (else_datarefs);
1543 VEC_free (gimple, heap, then_stores);
1544 VEC_free (gimple, heap, else_stores);
1545 return false;
1549 /* Sink stores with same LHS. */
1550 FOR_EACH_VEC_ELT (gimple, then_stores, i, then_store)
1552 else_store = VEC_index (gimple, else_stores, i);
1553 res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1554 then_store, else_store);
1555 ok = ok || res;
1558 free_dependence_relations (then_ddrs);
1559 free_dependence_relations (else_ddrs);
1560 free_data_refs (then_datarefs);
1561 free_data_refs (else_datarefs);
1562 VEC_free (gimple, heap, then_stores);
1563 VEC_free (gimple, heap, else_stores);
1565 return ok;
1568 /* Always do these optimizations if we have SSA
1569 trees to work on. */
1570 static bool
1571 gate_phiopt (void)
1573 return 1;
1576 struct gimple_opt_pass pass_phiopt =
1579 GIMPLE_PASS,
1580 "phiopt", /* name */
1581 gate_phiopt, /* gate */
1582 tree_ssa_phiopt, /* execute */
1583 NULL, /* sub */
1584 NULL, /* next */
1585 0, /* static_pass_number */
1586 TV_TREE_PHIOPT, /* tv_id */
1587 PROP_cfg | PROP_ssa, /* properties_required */
1588 0, /* properties_provided */
1589 0, /* properties_destroyed */
1590 0, /* todo_flags_start */
1591 TODO_ggc_collect
1592 | TODO_verify_ssa
1593 | TODO_verify_flow
1594 | TODO_verify_stmts /* todo_flags_finish */
1598 static bool
1599 gate_cselim (void)
1601 return flag_tree_cselim;
1604 struct gimple_opt_pass pass_cselim =
1607 GIMPLE_PASS,
1608 "cselim", /* name */
1609 gate_cselim, /* gate */
1610 tree_ssa_cs_elim, /* execute */
1611 NULL, /* sub */
1612 NULL, /* next */
1613 0, /* static_pass_number */
1614 TV_TREE_PHIOPT, /* tv_id */
1615 PROP_cfg | PROP_ssa, /* properties_required */
1616 0, /* properties_provided */
1617 0, /* properties_destroyed */
1618 0, /* todo_flags_start */
1619 TODO_ggc_collect
1620 | TODO_verify_ssa
1621 | TODO_verify_flow
1622 | TODO_verify_stmts /* todo_flags_finish */