2013-08-02 Paolo Carlini <paolo.carlini@oracle.com>
[official-gcc.git] / gcc / tree-ssa-phiopt.c
blob5e99678ac01caba423e2b312b533ea9c41ead449
1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2013 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
9 later version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "hash-table.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 "tree-flow.h"
31 #include "tree-pass.h"
32 #include "langhooks.h"
33 #include "pointer-set.h"
34 #include "domwalk.h"
35 #include "cfgloop.h"
36 #include "tree-data-ref.h"
37 #include "gimple-pretty-print.h"
38 #include "insn-config.h"
39 #include "expr.h"
40 #include "optabs.h"
41 #include "tree-scalar-evolution.h"
43 #ifndef HAVE_conditional_move
44 #define HAVE_conditional_move (0)
45 #endif
47 static unsigned int tree_ssa_phiopt (void);
48 static unsigned int tree_ssa_phiopt_worker (bool, bool);
49 static bool conditional_replacement (basic_block, basic_block,
50 edge, edge, gimple, tree, tree);
51 static int value_replacement (basic_block, basic_block,
52 edge, edge, gimple, tree, tree);
53 static bool minmax_replacement (basic_block, basic_block,
54 edge, edge, gimple, tree, tree);
55 static bool abs_replacement (basic_block, basic_block,
56 edge, edge, gimple, tree, tree);
57 static bool cond_store_replacement (basic_block, basic_block, edge, edge,
58 struct pointer_set_t *);
59 static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block);
60 static struct pointer_set_t * get_non_trapping (void);
61 static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree);
62 static void hoist_adjacent_loads (basic_block, basic_block,
63 basic_block, basic_block);
64 static bool gate_hoist_loads (void);
66 /* This pass tries to replaces an if-then-else block with an
67 assignment. We have four kinds of transformations. Some of these
68 transformations are also performed by the ifcvt RTL optimizer.
70 Conditional Replacement
71 -----------------------
73 This transformation, implemented in conditional_replacement,
74 replaces
76 bb0:
77 if (cond) goto bb2; else goto bb1;
78 bb1:
79 bb2:
80 x = PHI <0 (bb1), 1 (bb0), ...>;
82 with
84 bb0:
85 x' = cond;
86 goto bb2;
87 bb2:
88 x = PHI <x' (bb0), ...>;
90 We remove bb1 as it becomes unreachable. This occurs often due to
91 gimplification of conditionals.
93 Value Replacement
94 -----------------
96 This transformation, implemented in value_replacement, replaces
98 bb0:
99 if (a != b) goto bb2; else goto bb1;
100 bb1:
101 bb2:
102 x = PHI <a (bb1), b (bb0), ...>;
104 with
106 bb0:
107 bb2:
108 x = PHI <b (bb0), ...>;
110 This opportunity can sometimes occur as a result of other
111 optimizations.
113 ABS Replacement
114 ---------------
116 This transformation, implemented in abs_replacement, replaces
118 bb0:
119 if (a >= 0) goto bb2; else goto bb1;
120 bb1:
121 x = -a;
122 bb2:
123 x = PHI <x (bb1), a (bb0), ...>;
125 with
127 bb0:
128 x' = ABS_EXPR< a >;
129 bb2:
130 x = PHI <x' (bb0), ...>;
132 MIN/MAX Replacement
133 -------------------
135 This transformation, minmax_replacement replaces
137 bb0:
138 if (a <= b) goto bb2; else goto bb1;
139 bb1:
140 bb2:
141 x = PHI <b (bb1), a (bb0), ...>;
143 with
145 bb0:
146 x' = MIN_EXPR (a, b)
147 bb2:
148 x = PHI <x' (bb0), ...>;
150 A similar transformation is done for MAX_EXPR.
153 This pass also performs a fifth transformation of a slightly different
154 flavor.
156 Adjacent Load Hoisting
157 ----------------------
159 This transformation replaces
161 bb0:
162 if (...) goto bb2; else goto bb1;
163 bb1:
164 x1 = (<expr>).field1;
165 goto bb3;
166 bb2:
167 x2 = (<expr>).field2;
168 bb3:
169 # x = PHI <x1, x2>;
171 with
173 bb0:
174 x1 = (<expr>).field1;
175 x2 = (<expr>).field2;
176 if (...) goto bb2; else goto bb1;
177 bb1:
178 goto bb3;
179 bb2:
180 bb3:
181 # x = PHI <x1, x2>;
183 The purpose of this transformation is to enable generation of conditional
184 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
185 the loads is speculative, the transformation is restricted to very
186 specific cases to avoid introducing a page fault. We are looking for
187 the common idiom:
189 if (...)
190 x = y->left;
191 else
192 x = y->right;
194 where left and right are typically adjacent pointers in a tree structure. */
196 static unsigned int
197 tree_ssa_phiopt (void)
199 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
202 /* This pass tries to transform conditional stores into unconditional
203 ones, enabling further simplifications with the simpler then and else
204 blocks. In particular it replaces this:
206 bb0:
207 if (cond) goto bb2; else goto bb1;
208 bb1:
209 *p = RHS;
210 bb2:
212 with
214 bb0:
215 if (cond) goto bb1; else goto bb2;
216 bb1:
217 condtmp' = *p;
218 bb2:
219 condtmp = PHI <RHS, condtmp'>
220 *p = condtmp;
222 This transformation can only be done under several constraints,
223 documented below. It also replaces:
225 bb0:
226 if (cond) goto bb2; else goto bb1;
227 bb1:
228 *p = RHS1;
229 goto bb3;
230 bb2:
231 *p = RHS2;
232 bb3:
234 with
236 bb0:
237 if (cond) goto bb3; else goto bb1;
238 bb1:
239 bb3:
240 condtmp = PHI <RHS1, RHS2>
241 *p = condtmp; */
243 static unsigned int
244 tree_ssa_cs_elim (void)
246 unsigned todo;
247 /* ??? We are not interested in loop related info, but the following
248 will create it, ICEing as we didn't init loops with pre-headers.
249 An interfacing issue of find_data_references_in_bb. */
250 loop_optimizer_init (LOOPS_NORMAL);
251 scev_initialize ();
252 todo = tree_ssa_phiopt_worker (true, false);
253 scev_finalize ();
254 loop_optimizer_finalize ();
255 return todo;
258 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
260 static gimple
261 single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1)
263 gimple_stmt_iterator i;
264 gimple phi = NULL;
265 if (gimple_seq_singleton_p (seq))
266 return gsi_stmt (gsi_start (seq));
267 for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i))
269 gimple p = gsi_stmt (i);
270 /* If the PHI arguments are equal then we can skip this PHI. */
271 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx),
272 gimple_phi_arg_def (p, e1->dest_idx)))
273 continue;
275 /* If we already have a PHI that has the two edge arguments are
276 different, then return it is not a singleton for these PHIs. */
277 if (phi)
278 return NULL;
280 phi = p;
282 return phi;
285 /* The core routine of conditional store replacement and normal
286 phi optimizations. Both share much of the infrastructure in how
287 to match applicable basic block patterns. DO_STORE_ELIM is true
288 when we want to do conditional store replacement, false otherwise.
289 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
290 of diamond control flow patterns, false otherwise. */
291 static unsigned int
292 tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads)
294 basic_block bb;
295 basic_block *bb_order;
296 unsigned n, i;
297 bool cfgchanged = false;
298 struct pointer_set_t *nontrap = 0;
300 if (do_store_elim)
301 /* Calculate the set of non-trapping memory accesses. */
302 nontrap = get_non_trapping ();
304 /* Search every basic block for COND_EXPR we may be able to optimize.
306 We walk the blocks in order that guarantees that a block with
307 a single predecessor is processed before the predecessor.
308 This ensures that we collapse inner ifs before visiting the
309 outer ones, and also that we do not try to visit a removed
310 block. */
311 bb_order = blocks_in_phiopt_order ();
312 n = n_basic_blocks - NUM_FIXED_BLOCKS;
314 for (i = 0; i < n; i++)
316 gimple cond_stmt, phi;
317 basic_block bb1, bb2;
318 edge e1, e2;
319 tree arg0, arg1;
321 bb = bb_order[i];
323 cond_stmt = last_stmt (bb);
324 /* Check to see if the last statement is a GIMPLE_COND. */
325 if (!cond_stmt
326 || gimple_code (cond_stmt) != GIMPLE_COND)
327 continue;
329 e1 = EDGE_SUCC (bb, 0);
330 bb1 = e1->dest;
331 e2 = EDGE_SUCC (bb, 1);
332 bb2 = e2->dest;
334 /* We cannot do the optimization on abnormal edges. */
335 if ((e1->flags & EDGE_ABNORMAL) != 0
336 || (e2->flags & EDGE_ABNORMAL) != 0)
337 continue;
339 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
340 if (EDGE_COUNT (bb1->succs) == 0
341 || bb2 == NULL
342 || EDGE_COUNT (bb2->succs) == 0)
343 continue;
345 /* Find the bb which is the fall through to the other. */
346 if (EDGE_SUCC (bb1, 0)->dest == bb2)
348 else if (EDGE_SUCC (bb2, 0)->dest == bb1)
350 basic_block bb_tmp = bb1;
351 edge e_tmp = e1;
352 bb1 = bb2;
353 bb2 = bb_tmp;
354 e1 = e2;
355 e2 = e_tmp;
357 else if (do_store_elim
358 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
360 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
362 if (!single_succ_p (bb1)
363 || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0
364 || !single_succ_p (bb2)
365 || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0
366 || EDGE_COUNT (bb3->preds) != 2)
367 continue;
368 if (cond_if_else_store_replacement (bb1, bb2, bb3))
369 cfgchanged = true;
370 continue;
372 else if (do_hoist_loads
373 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
375 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
377 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt)))
378 && single_succ_p (bb1)
379 && single_succ_p (bb2)
380 && single_pred_p (bb1)
381 && single_pred_p (bb2)
382 && EDGE_COUNT (bb->succs) == 2
383 && EDGE_COUNT (bb3->preds) == 2
384 /* If one edge or the other is dominant, a conditional move
385 is likely to perform worse than the well-predicted branch. */
386 && !predictable_edge_p (EDGE_SUCC (bb, 0))
387 && !predictable_edge_p (EDGE_SUCC (bb, 1)))
388 hoist_adjacent_loads (bb, bb1, bb2, bb3);
389 continue;
391 else
392 continue;
394 e1 = EDGE_SUCC (bb1, 0);
396 /* Make sure that bb1 is just a fall through. */
397 if (!single_succ_p (bb1)
398 || (e1->flags & EDGE_FALLTHRU) == 0)
399 continue;
401 /* Also make sure that bb1 only have one predecessor and that it
402 is bb. */
403 if (!single_pred_p (bb1)
404 || single_pred (bb1) != bb)
405 continue;
407 if (do_store_elim)
409 /* bb1 is the middle block, bb2 the join block, bb the split block,
410 e1 the fallthrough edge from bb1 to bb2. We can't do the
411 optimization if the join block has more than two predecessors. */
412 if (EDGE_COUNT (bb2->preds) > 2)
413 continue;
414 if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
415 cfgchanged = true;
417 else
419 gimple_seq phis = phi_nodes (bb2);
420 gimple_stmt_iterator gsi;
421 bool candorest = true;
423 /* Value replacement can work with more than one PHI
424 so try that first. */
425 for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
427 phi = gsi_stmt (gsi);
428 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
429 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
430 if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2)
432 candorest = false;
433 cfgchanged = true;
434 break;
438 if (!candorest)
439 continue;
441 phi = single_non_singleton_phi_for_edges (phis, e1, e2);
442 if (!phi)
443 continue;
445 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
446 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
448 /* Something is wrong if we cannot find the arguments in the PHI
449 node. */
450 gcc_assert (arg0 != NULL && arg1 != NULL);
452 /* Do the replacement of conditional if it can be done. */
453 if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
454 cfgchanged = true;
455 else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
456 cfgchanged = true;
457 else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
458 cfgchanged = true;
462 free (bb_order);
464 if (do_store_elim)
465 pointer_set_destroy (nontrap);
466 /* If the CFG has changed, we should cleanup the CFG. */
467 if (cfgchanged && do_store_elim)
469 /* In cond-store replacement we have added some loads on edges
470 and new VOPS (as we moved the store, and created a load). */
471 gsi_commit_edge_inserts ();
472 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
474 else if (cfgchanged)
475 return TODO_cleanup_cfg;
476 return 0;
479 /* Returns the list of basic blocks in the function in an order that guarantees
480 that if a block X has just a single predecessor Y, then Y is after X in the
481 ordering. */
483 basic_block *
484 blocks_in_phiopt_order (void)
486 basic_block x, y;
487 basic_block *order = XNEWVEC (basic_block, n_basic_blocks);
488 unsigned n = n_basic_blocks - NUM_FIXED_BLOCKS;
489 unsigned np, i;
490 sbitmap visited = sbitmap_alloc (last_basic_block);
492 #define MARK_VISITED(BB) (bitmap_set_bit (visited, (BB)->index))
493 #define VISITED_P(BB) (bitmap_bit_p (visited, (BB)->index))
495 bitmap_clear (visited);
497 MARK_VISITED (ENTRY_BLOCK_PTR);
498 FOR_EACH_BB (x)
500 if (VISITED_P (x))
501 continue;
503 /* Walk the predecessors of x as long as they have precisely one
504 predecessor and add them to the list, so that they get stored
505 after x. */
506 for (y = x, np = 1;
507 single_pred_p (y) && !VISITED_P (single_pred (y));
508 y = single_pred (y))
509 np++;
510 for (y = x, i = n - np;
511 single_pred_p (y) && !VISITED_P (single_pred (y));
512 y = single_pred (y), i++)
514 order[i] = y;
515 MARK_VISITED (y);
517 order[i] = y;
518 MARK_VISITED (y);
520 gcc_assert (i == n - 1);
521 n -= np;
524 sbitmap_free (visited);
525 gcc_assert (n == 0);
526 return order;
528 #undef MARK_VISITED
529 #undef VISITED_P
532 /* Replace PHI node element whose edge is E in block BB with variable NEW.
533 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
534 is known to have two edges, one of which must reach BB). */
536 static void
537 replace_phi_edge_with_variable (basic_block cond_block,
538 edge e, gimple phi, tree new_tree)
540 basic_block bb = gimple_bb (phi);
541 basic_block block_to_remove;
542 gimple_stmt_iterator gsi;
544 /* Change the PHI argument to new. */
545 SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree);
547 /* Remove the empty basic block. */
548 if (EDGE_SUCC (cond_block, 0)->dest == bb)
550 EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU;
551 EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
552 EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE;
553 EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count;
555 block_to_remove = EDGE_SUCC (cond_block, 1)->dest;
557 else
559 EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU;
560 EDGE_SUCC (cond_block, 1)->flags
561 &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
562 EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE;
563 EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count;
565 block_to_remove = EDGE_SUCC (cond_block, 0)->dest;
567 delete_basic_block (block_to_remove);
569 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
570 gsi = gsi_last_bb (cond_block);
571 gsi_remove (&gsi, true);
573 if (dump_file && (dump_flags & TDF_DETAILS))
574 fprintf (dump_file,
575 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
576 cond_block->index,
577 bb->index);
580 /* The function conditional_replacement does the main work of doing the
581 conditional replacement. Return true if the replacement is done.
582 Otherwise return false.
583 BB is the basic block where the replacement is going to be done on. ARG0
584 is argument 0 from PHI. Likewise for ARG1. */
586 static bool
587 conditional_replacement (basic_block cond_bb, basic_block middle_bb,
588 edge e0, edge e1, gimple phi,
589 tree arg0, tree arg1)
591 tree result;
592 gimple stmt, new_stmt;
593 tree cond;
594 gimple_stmt_iterator gsi;
595 edge true_edge, false_edge;
596 tree new_var, new_var2;
597 bool neg;
599 /* FIXME: Gimplification of complex type is too hard for now. */
600 /* We aren't prepared to handle vectors either (and it is a question
601 if it would be worthwhile anyway). */
602 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0))
603 || POINTER_TYPE_P (TREE_TYPE (arg0)))
604 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1))
605 || POINTER_TYPE_P (TREE_TYPE (arg1))))
606 return false;
608 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
609 convert it to the conditional. */
610 if ((integer_zerop (arg0) && integer_onep (arg1))
611 || (integer_zerop (arg1) && integer_onep (arg0)))
612 neg = false;
613 else if ((integer_zerop (arg0) && integer_all_onesp (arg1))
614 || (integer_zerop (arg1) && integer_all_onesp (arg0)))
615 neg = true;
616 else
617 return false;
619 if (!empty_block_p (middle_bb))
620 return false;
622 /* At this point we know we have a GIMPLE_COND with two successors.
623 One successor is BB, the other successor is an empty block which
624 falls through into BB.
626 There is a single PHI node at the join point (BB) and its arguments
627 are constants (0, 1) or (0, -1).
629 So, given the condition COND, and the two PHI arguments, we can
630 rewrite this PHI into non-branching code:
632 dest = (COND) or dest = COND'
634 We use the condition as-is if the argument associated with the
635 true edge has the value one or the argument associated with the
636 false edge as the value zero. Note that those conditions are not
637 the same since only one of the outgoing edges from the GIMPLE_COND
638 will directly reach BB and thus be associated with an argument. */
640 stmt = last_stmt (cond_bb);
641 result = PHI_RESULT (phi);
643 /* To handle special cases like floating point comparison, it is easier and
644 less error-prone to build a tree and gimplify it on the fly though it is
645 less efficient. */
646 cond = fold_build2_loc (gimple_location (stmt),
647 gimple_cond_code (stmt), boolean_type_node,
648 gimple_cond_lhs (stmt), gimple_cond_rhs (stmt));
650 /* We need to know which is the true edge and which is the false
651 edge so that we know when to invert the condition below. */
652 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
653 if ((e0 == true_edge && integer_zerop (arg0))
654 || (e0 == false_edge && !integer_zerop (arg0))
655 || (e1 == true_edge && integer_zerop (arg1))
656 || (e1 == false_edge && !integer_zerop (arg1)))
657 cond = fold_build1_loc (gimple_location (stmt),
658 TRUTH_NOT_EXPR, TREE_TYPE (cond), cond);
660 if (neg)
662 cond = fold_convert_loc (gimple_location (stmt),
663 TREE_TYPE (result), cond);
664 cond = fold_build1_loc (gimple_location (stmt),
665 NEGATE_EXPR, TREE_TYPE (cond), cond);
668 /* Insert our new statements at the end of conditional block before the
669 COND_STMT. */
670 gsi = gsi_for_stmt (stmt);
671 new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true,
672 GSI_SAME_STMT);
674 if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var)))
676 source_location locus_0, locus_1;
678 new_var2 = make_ssa_name (TREE_TYPE (result), NULL);
679 new_stmt = gimple_build_assign_with_ops (CONVERT_EXPR, new_var2,
680 new_var, NULL);
681 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
682 new_var = new_var2;
684 /* Set the locus to the first argument, unless is doesn't have one. */
685 locus_0 = gimple_phi_arg_location (phi, 0);
686 locus_1 = gimple_phi_arg_location (phi, 1);
687 if (locus_0 == UNKNOWN_LOCATION)
688 locus_0 = locus_1;
689 gimple_set_location (new_stmt, locus_0);
692 replace_phi_edge_with_variable (cond_bb, e1, phi, new_var);
694 /* Note that we optimized this PHI. */
695 return true;
698 /* Update *ARG which is defined in STMT so that it contains the
699 computed value if that seems profitable. Return true if the
700 statement is made dead by that rewriting. */
702 static bool
703 jump_function_from_stmt (tree *arg, gimple stmt)
705 enum tree_code code = gimple_assign_rhs_code (stmt);
706 if (code == ADDR_EXPR)
708 /* For arg = &p->i transform it to p, if possible. */
709 tree rhs1 = gimple_assign_rhs1 (stmt);
710 HOST_WIDE_INT offset;
711 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0),
712 &offset);
713 if (tem
714 && TREE_CODE (tem) == MEM_REF
715 && (mem_ref_offset (tem) + double_int::from_shwi (offset)).is_zero ())
717 *arg = TREE_OPERAND (tem, 0);
718 return true;
721 /* TODO: Much like IPA-CP jump-functions we want to handle constant
722 additions symbolically here, and we'd need to update the comparison
723 code that compares the arg + cst tuples in our caller. For now the
724 code above exactly handles the VEC_BASE pattern from vec.h. */
725 return false;
728 /* The function value_replacement does the main work of doing the value
729 replacement. Return non-zero if the replacement is done. Otherwise return
730 0. If we remove the middle basic block, return 2.
731 BB is the basic block where the replacement is going to be done on. ARG0
732 is argument 0 from the PHI. Likewise for ARG1. */
734 static int
735 value_replacement (basic_block cond_bb, basic_block middle_bb,
736 edge e0, edge e1, gimple phi,
737 tree arg0, tree arg1)
739 gimple_stmt_iterator gsi;
740 gimple cond;
741 edge true_edge, false_edge;
742 enum tree_code code;
743 bool emtpy_or_with_defined_p = true;
745 /* If the type says honor signed zeros we cannot do this
746 optimization. */
747 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
748 return 0;
750 /* If there is a statement in MIDDLE_BB that defines one of the PHI
751 arguments, then adjust arg0 or arg1. */
752 gsi = gsi_after_labels (middle_bb);
753 if (!gsi_end_p (gsi) && is_gimple_debug (gsi_stmt (gsi)))
754 gsi_next_nondebug (&gsi);
755 while (!gsi_end_p (gsi))
757 gimple stmt = gsi_stmt (gsi);
758 tree lhs;
759 gsi_next_nondebug (&gsi);
760 if (!is_gimple_assign (stmt))
762 emtpy_or_with_defined_p = false;
763 continue;
765 /* Now try to adjust arg0 or arg1 according to the computation
766 in the statement. */
767 lhs = gimple_assign_lhs (stmt);
768 if (!(lhs == arg0
769 && jump_function_from_stmt (&arg0, stmt))
770 || (lhs == arg1
771 && jump_function_from_stmt (&arg1, stmt)))
772 emtpy_or_with_defined_p = false;
775 cond = last_stmt (cond_bb);
776 code = gimple_cond_code (cond);
778 /* This transformation is only valid for equality comparisons. */
779 if (code != NE_EXPR && code != EQ_EXPR)
780 return 0;
782 /* We need to know which is the true edge and which is the false
783 edge so that we know if have abs or negative abs. */
784 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
786 /* At this point we know we have a COND_EXPR with two successors.
787 One successor is BB, the other successor is an empty block which
788 falls through into BB.
790 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
792 There is a single PHI node at the join point (BB) with two arguments.
794 We now need to verify that the two arguments in the PHI node match
795 the two arguments to the equality comparison. */
797 if ((operand_equal_for_phi_arg_p (arg0, gimple_cond_lhs (cond))
798 && operand_equal_for_phi_arg_p (arg1, gimple_cond_rhs (cond)))
799 || (operand_equal_for_phi_arg_p (arg1, gimple_cond_lhs (cond))
800 && operand_equal_for_phi_arg_p (arg0, gimple_cond_rhs (cond))))
802 edge e;
803 tree arg;
805 /* For NE_EXPR, we want to build an assignment result = arg where
806 arg is the PHI argument associated with the true edge. For
807 EQ_EXPR we want the PHI argument associated with the false edge. */
808 e = (code == NE_EXPR ? true_edge : false_edge);
810 /* Unfortunately, E may not reach BB (it may instead have gone to
811 OTHER_BLOCK). If that is the case, then we want the single outgoing
812 edge from OTHER_BLOCK which reaches BB and represents the desired
813 path from COND_BLOCK. */
814 if (e->dest == middle_bb)
815 e = single_succ_edge (e->dest);
817 /* Now we know the incoming edge to BB that has the argument for the
818 RHS of our new assignment statement. */
819 if (e0 == e)
820 arg = arg0;
821 else
822 arg = arg1;
824 /* If the middle basic block was empty or is defining the
825 PHI arguments and this is a single phi where the args are different
826 for the edges e0 and e1 then we can remove the middle basic block. */
827 if (emtpy_or_with_defined_p
828 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)),
829 e0, e1))
831 replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
832 /* Note that we optimized this PHI. */
833 return 2;
835 else
837 /* Replace the PHI arguments with arg. */
838 SET_PHI_ARG_DEF (phi, e0->dest_idx, arg);
839 SET_PHI_ARG_DEF (phi, e1->dest_idx, arg);
840 if (dump_file && (dump_flags & TDF_DETAILS))
842 fprintf (dump_file, "PHI ");
843 print_generic_expr (dump_file, gimple_phi_result (phi), 0);
844 fprintf (dump_file, " reduced for COND_EXPR in block %d to ",
845 cond_bb->index);
846 print_generic_expr (dump_file, arg, 0);
847 fprintf (dump_file, ".\n");
849 return 1;
853 return 0;
856 /* The function minmax_replacement does the main work of doing the minmax
857 replacement. Return true if the replacement is done. Otherwise return
858 false.
859 BB is the basic block where the replacement is going to be done on. ARG0
860 is argument 0 from the PHI. Likewise for ARG1. */
862 static bool
863 minmax_replacement (basic_block cond_bb, basic_block middle_bb,
864 edge e0, edge e1, gimple phi,
865 tree arg0, tree arg1)
867 tree result, type;
868 gimple cond, new_stmt;
869 edge true_edge, false_edge;
870 enum tree_code cmp, minmax, ass_code;
871 tree smaller, larger, arg_true, arg_false;
872 gimple_stmt_iterator gsi, gsi_from;
874 type = TREE_TYPE (PHI_RESULT (phi));
876 /* The optimization may be unsafe due to NaNs. */
877 if (HONOR_NANS (TYPE_MODE (type)))
878 return false;
880 cond = last_stmt (cond_bb);
881 cmp = gimple_cond_code (cond);
883 /* This transformation is only valid for order comparisons. Record which
884 operand is smaller/larger if the result of the comparison is true. */
885 if (cmp == LT_EXPR || cmp == LE_EXPR)
887 smaller = gimple_cond_lhs (cond);
888 larger = gimple_cond_rhs (cond);
890 else if (cmp == GT_EXPR || cmp == GE_EXPR)
892 smaller = gimple_cond_rhs (cond);
893 larger = gimple_cond_lhs (cond);
895 else
896 return false;
898 /* We need to know which is the true edge and which is the false
899 edge so that we know if have abs or negative abs. */
900 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
902 /* Forward the edges over the middle basic block. */
903 if (true_edge->dest == middle_bb)
904 true_edge = EDGE_SUCC (true_edge->dest, 0);
905 if (false_edge->dest == middle_bb)
906 false_edge = EDGE_SUCC (false_edge->dest, 0);
908 if (true_edge == e0)
910 gcc_assert (false_edge == e1);
911 arg_true = arg0;
912 arg_false = arg1;
914 else
916 gcc_assert (false_edge == e0);
917 gcc_assert (true_edge == e1);
918 arg_true = arg1;
919 arg_false = arg0;
922 if (empty_block_p (middle_bb))
924 if (operand_equal_for_phi_arg_p (arg_true, smaller)
925 && operand_equal_for_phi_arg_p (arg_false, larger))
927 /* Case
929 if (smaller < larger)
930 rslt = smaller;
931 else
932 rslt = larger; */
933 minmax = MIN_EXPR;
935 else if (operand_equal_for_phi_arg_p (arg_false, smaller)
936 && operand_equal_for_phi_arg_p (arg_true, larger))
937 minmax = MAX_EXPR;
938 else
939 return false;
941 else
943 /* Recognize the following case, assuming d <= u:
945 if (a <= u)
946 b = MAX (a, d);
947 x = PHI <b, u>
949 This is equivalent to
951 b = MAX (a, d);
952 x = MIN (b, u); */
954 gimple assign = last_and_only_stmt (middle_bb);
955 tree lhs, op0, op1, bound;
957 if (!assign
958 || gimple_code (assign) != GIMPLE_ASSIGN)
959 return false;
961 lhs = gimple_assign_lhs (assign);
962 ass_code = gimple_assign_rhs_code (assign);
963 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
964 return false;
965 op0 = gimple_assign_rhs1 (assign);
966 op1 = gimple_assign_rhs2 (assign);
968 if (true_edge->src == middle_bb)
970 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
971 if (!operand_equal_for_phi_arg_p (lhs, arg_true))
972 return false;
974 if (operand_equal_for_phi_arg_p (arg_false, larger))
976 /* Case
978 if (smaller < larger)
980 r' = MAX_EXPR (smaller, bound)
982 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
983 if (ass_code != MAX_EXPR)
984 return false;
986 minmax = MIN_EXPR;
987 if (operand_equal_for_phi_arg_p (op0, smaller))
988 bound = op1;
989 else if (operand_equal_for_phi_arg_p (op1, smaller))
990 bound = op0;
991 else
992 return false;
994 /* We need BOUND <= LARGER. */
995 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
996 bound, larger)))
997 return false;
999 else if (operand_equal_for_phi_arg_p (arg_false, smaller))
1001 /* Case
1003 if (smaller < larger)
1005 r' = MIN_EXPR (larger, bound)
1007 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1008 if (ass_code != MIN_EXPR)
1009 return false;
1011 minmax = MAX_EXPR;
1012 if (operand_equal_for_phi_arg_p (op0, larger))
1013 bound = op1;
1014 else if (operand_equal_for_phi_arg_p (op1, larger))
1015 bound = op0;
1016 else
1017 return false;
1019 /* We need BOUND >= SMALLER. */
1020 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1021 bound, smaller)))
1022 return false;
1024 else
1025 return false;
1027 else
1029 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1030 if (!operand_equal_for_phi_arg_p (lhs, arg_false))
1031 return false;
1033 if (operand_equal_for_phi_arg_p (arg_true, larger))
1035 /* Case
1037 if (smaller > larger)
1039 r' = MIN_EXPR (smaller, bound)
1041 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1042 if (ass_code != MIN_EXPR)
1043 return false;
1045 minmax = MAX_EXPR;
1046 if (operand_equal_for_phi_arg_p (op0, smaller))
1047 bound = op1;
1048 else if (operand_equal_for_phi_arg_p (op1, smaller))
1049 bound = op0;
1050 else
1051 return false;
1053 /* We need BOUND >= LARGER. */
1054 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1055 bound, larger)))
1056 return false;
1058 else if (operand_equal_for_phi_arg_p (arg_true, smaller))
1060 /* Case
1062 if (smaller > larger)
1064 r' = MAX_EXPR (larger, bound)
1066 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1067 if (ass_code != MAX_EXPR)
1068 return false;
1070 minmax = MIN_EXPR;
1071 if (operand_equal_for_phi_arg_p (op0, larger))
1072 bound = op1;
1073 else if (operand_equal_for_phi_arg_p (op1, larger))
1074 bound = op0;
1075 else
1076 return false;
1078 /* We need BOUND <= SMALLER. */
1079 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1080 bound, smaller)))
1081 return false;
1083 else
1084 return false;
1087 /* Move the statement from the middle block. */
1088 gsi = gsi_last_bb (cond_bb);
1089 gsi_from = gsi_last_nondebug_bb (middle_bb);
1090 gsi_move_before (&gsi_from, &gsi);
1093 /* Emit the statement to compute min/max. */
1094 result = duplicate_ssa_name (PHI_RESULT (phi), NULL);
1095 new_stmt = gimple_build_assign_with_ops (minmax, result, arg0, arg1);
1096 gsi = gsi_last_bb (cond_bb);
1097 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1099 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1100 return true;
1103 /* The function absolute_replacement does the main work of doing the absolute
1104 replacement. Return true if the replacement is done. Otherwise return
1105 false.
1106 bb is the basic block where the replacement is going to be done on. arg0
1107 is argument 0 from the phi. Likewise for arg1. */
1109 static bool
1110 abs_replacement (basic_block cond_bb, basic_block middle_bb,
1111 edge e0 ATTRIBUTE_UNUSED, edge e1,
1112 gimple phi, tree arg0, tree arg1)
1114 tree result;
1115 gimple new_stmt, cond;
1116 gimple_stmt_iterator gsi;
1117 edge true_edge, false_edge;
1118 gimple assign;
1119 edge e;
1120 tree rhs, lhs;
1121 bool negate;
1122 enum tree_code cond_code;
1124 /* If the type says honor signed zeros we cannot do this
1125 optimization. */
1126 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
1127 return false;
1129 /* OTHER_BLOCK must have only one executable statement which must have the
1130 form arg0 = -arg1 or arg1 = -arg0. */
1132 assign = last_and_only_stmt (middle_bb);
1133 /* If we did not find the proper negation assignment, then we can not
1134 optimize. */
1135 if (assign == NULL)
1136 return false;
1138 /* If we got here, then we have found the only executable statement
1139 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1140 arg1 = -arg0, then we can not optimize. */
1141 if (gimple_code (assign) != GIMPLE_ASSIGN)
1142 return false;
1144 lhs = gimple_assign_lhs (assign);
1146 if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
1147 return false;
1149 rhs = gimple_assign_rhs1 (assign);
1151 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1152 if (!(lhs == arg0 && rhs == arg1)
1153 && !(lhs == arg1 && rhs == arg0))
1154 return false;
1156 cond = last_stmt (cond_bb);
1157 result = PHI_RESULT (phi);
1159 /* Only relationals comparing arg[01] against zero are interesting. */
1160 cond_code = gimple_cond_code (cond);
1161 if (cond_code != GT_EXPR && cond_code != GE_EXPR
1162 && cond_code != LT_EXPR && cond_code != LE_EXPR)
1163 return false;
1165 /* Make sure the conditional is arg[01] OP y. */
1166 if (gimple_cond_lhs (cond) != rhs)
1167 return false;
1169 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond)))
1170 ? real_zerop (gimple_cond_rhs (cond))
1171 : integer_zerop (gimple_cond_rhs (cond)))
1173 else
1174 return false;
1176 /* We need to know which is the true edge and which is the false
1177 edge so that we know if have abs or negative abs. */
1178 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1180 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1181 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1182 the false edge goes to OTHER_BLOCK. */
1183 if (cond_code == GT_EXPR || cond_code == GE_EXPR)
1184 e = true_edge;
1185 else
1186 e = false_edge;
1188 if (e->dest == middle_bb)
1189 negate = true;
1190 else
1191 negate = false;
1193 result = duplicate_ssa_name (result, NULL);
1195 if (negate)
1196 lhs = make_ssa_name (TREE_TYPE (result), NULL);
1197 else
1198 lhs = result;
1200 /* Build the modify expression with abs expression. */
1201 new_stmt = gimple_build_assign_with_ops (ABS_EXPR, lhs, rhs, NULL);
1203 gsi = gsi_last_bb (cond_bb);
1204 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1206 if (negate)
1208 /* Get the right GSI. We want to insert after the recently
1209 added ABS_EXPR statement (which we know is the first statement
1210 in the block. */
1211 new_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, result, lhs, NULL);
1213 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1216 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1218 /* Note that we optimized this PHI. */
1219 return true;
1222 /* Auxiliary functions to determine the set of memory accesses which
1223 can't trap because they are preceded by accesses to the same memory
1224 portion. We do that for MEM_REFs, so we only need to track
1225 the SSA_NAME of the pointer indirectly referenced. The algorithm
1226 simply is a walk over all instructions in dominator order. When
1227 we see an MEM_REF we determine if we've already seen a same
1228 ref anywhere up to the root of the dominator tree. If we do the
1229 current access can't trap. If we don't see any dominating access
1230 the current access might trap, but might also make later accesses
1231 non-trapping, so we remember it. We need to be careful with loads
1232 or stores, for instance a load might not trap, while a store would,
1233 so if we see a dominating read access this doesn't mean that a later
1234 write access would not trap. Hence we also need to differentiate the
1235 type of access(es) seen.
1237 ??? We currently are very conservative and assume that a load might
1238 trap even if a store doesn't (write-only memory). This probably is
1239 overly conservative. */
1241 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1242 through it was seen, which would constitute a no-trap region for
1243 same accesses. */
1244 struct name_to_bb
1246 unsigned int ssa_name_ver;
1247 unsigned int phase;
1248 bool store;
1249 HOST_WIDE_INT offset, size;
1250 basic_block bb;
1253 /* Hashtable helpers. */
1255 struct ssa_names_hasher : typed_free_remove <name_to_bb>
1257 typedef name_to_bb value_type;
1258 typedef name_to_bb compare_type;
1259 static inline hashval_t hash (const value_type *);
1260 static inline bool equal (const value_type *, const compare_type *);
1263 /* Used for quick clearing of the hash-table when we see calls.
1264 Hash entries with phase < nt_call_phase are invalid. */
1265 static unsigned int nt_call_phase;
1267 /* The set of MEM_REFs which can't trap. */
1268 static struct pointer_set_t *nontrap_set;
1270 /* The hash function. */
1272 inline hashval_t
1273 ssa_names_hasher::hash (const value_type *n)
1275 return n->ssa_name_ver ^ (((hashval_t) n->store) << 31)
1276 ^ (n->offset << 6) ^ (n->size << 3);
1279 /* The equality function of *P1 and *P2. */
1281 inline bool
1282 ssa_names_hasher::equal (const value_type *n1, const compare_type *n2)
1284 return n1->ssa_name_ver == n2->ssa_name_ver
1285 && n1->store == n2->store
1286 && n1->offset == n2->offset
1287 && n1->size == n2->size;
1290 /* The hash table for remembering what we've seen. */
1291 static hash_table <ssa_names_hasher> seen_ssa_names;
1293 /* We see the expression EXP in basic block BB. If it's an interesting
1294 expression (an MEM_REF through an SSA_NAME) possibly insert the
1295 expression into the set NONTRAP or the hash table of seen expressions.
1296 STORE is true if this expression is on the LHS, otherwise it's on
1297 the RHS. */
1298 static void
1299 add_or_mark_expr (basic_block bb, tree exp,
1300 struct pointer_set_t *nontrap, bool store)
1302 HOST_WIDE_INT size;
1304 if (TREE_CODE (exp) == MEM_REF
1305 && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME
1306 && host_integerp (TREE_OPERAND (exp, 1), 0)
1307 && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0)
1309 tree name = TREE_OPERAND (exp, 0);
1310 struct name_to_bb map;
1311 name_to_bb **slot;
1312 struct name_to_bb *n2bb;
1313 basic_block found_bb = 0;
1315 /* Try to find the last seen MEM_REF through the same
1316 SSA_NAME, which can trap. */
1317 map.ssa_name_ver = SSA_NAME_VERSION (name);
1318 map.phase = 0;
1319 map.bb = 0;
1320 map.store = store;
1321 map.offset = tree_low_cst (TREE_OPERAND (exp, 1), 0);
1322 map.size = size;
1324 slot = seen_ssa_names.find_slot (&map, INSERT);
1325 n2bb = *slot;
1326 if (n2bb && n2bb->phase >= nt_call_phase)
1327 found_bb = n2bb->bb;
1329 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1330 (it's in a basic block on the path from us to the dominator root)
1331 then we can't trap. */
1332 if (found_bb && (((size_t)found_bb->aux) & 1) == 1)
1334 pointer_set_insert (nontrap, exp);
1336 else
1338 /* EXP might trap, so insert it into the hash table. */
1339 if (n2bb)
1341 n2bb->phase = nt_call_phase;
1342 n2bb->bb = bb;
1344 else
1346 n2bb = XNEW (struct name_to_bb);
1347 n2bb->ssa_name_ver = SSA_NAME_VERSION (name);
1348 n2bb->phase = nt_call_phase;
1349 n2bb->bb = bb;
1350 n2bb->store = store;
1351 n2bb->offset = map.offset;
1352 n2bb->size = size;
1353 *slot = n2bb;
1359 /* Return true when CALL is a call stmt that definitely doesn't
1360 free any memory or makes it unavailable otherwise. */
1361 bool
1362 nonfreeing_call_p (gimple call)
1364 if (gimple_call_builtin_p (call, BUILT_IN_NORMAL)
1365 && gimple_call_flags (call) & ECF_LEAF)
1366 switch (DECL_FUNCTION_CODE (gimple_call_fndecl (call)))
1368 /* Just in case these become ECF_LEAF in the future. */
1369 case BUILT_IN_FREE:
1370 case BUILT_IN_TM_FREE:
1371 case BUILT_IN_REALLOC:
1372 case BUILT_IN_STACK_RESTORE:
1373 return false;
1374 default:
1375 return true;
1378 return false;
1381 /* Called by walk_dominator_tree, when entering the block BB. */
1382 static void
1383 nt_init_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb)
1385 edge e;
1386 edge_iterator ei;
1387 gimple_stmt_iterator gsi;
1389 /* If we haven't seen all our predecessors, clear the hash-table. */
1390 FOR_EACH_EDGE (e, ei, bb->preds)
1391 if ((((size_t)e->src->aux) & 2) == 0)
1393 nt_call_phase++;
1394 break;
1397 /* Mark this BB as being on the path to dominator root and as visited. */
1398 bb->aux = (void*)(1 | 2);
1400 /* And walk the statements in order. */
1401 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1403 gimple stmt = gsi_stmt (gsi);
1405 if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt))
1406 nt_call_phase++;
1407 else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt))
1409 add_or_mark_expr (bb, gimple_assign_lhs (stmt), nontrap_set, true);
1410 add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), nontrap_set, false);
1415 /* Called by walk_dominator_tree, when basic block BB is exited. */
1416 static void
1417 nt_fini_block (struct dom_walk_data *data ATTRIBUTE_UNUSED, basic_block bb)
1419 /* This BB isn't on the path to dominator root anymore. */
1420 bb->aux = (void*)2;
1423 /* This is the entry point of gathering non trapping memory accesses.
1424 It will do a dominator walk over the whole function, and it will
1425 make use of the bb->aux pointers. It returns a set of trees
1426 (the MEM_REFs itself) which can't trap. */
1427 static struct pointer_set_t *
1428 get_non_trapping (void)
1430 struct pointer_set_t *nontrap;
1431 struct dom_walk_data walk_data;
1433 nt_call_phase = 0;
1434 nontrap = pointer_set_create ();
1435 seen_ssa_names.create (128);
1436 /* We're going to do a dominator walk, so ensure that we have
1437 dominance information. */
1438 calculate_dominance_info (CDI_DOMINATORS);
1440 /* Setup callbacks for the generic dominator tree walker. */
1441 nontrap_set = nontrap;
1442 walk_data.dom_direction = CDI_DOMINATORS;
1443 walk_data.initialize_block_local_data = NULL;
1444 walk_data.before_dom_children = nt_init_block;
1445 walk_data.after_dom_children = nt_fini_block;
1446 walk_data.global_data = NULL;
1447 walk_data.block_local_data_size = 0;
1449 init_walk_dominator_tree (&walk_data);
1450 walk_dominator_tree (&walk_data, ENTRY_BLOCK_PTR);
1451 fini_walk_dominator_tree (&walk_data);
1452 seen_ssa_names.dispose ();
1454 clear_aux_for_blocks ();
1455 return nontrap;
1458 /* Do the main work of conditional store replacement. We already know
1459 that the recognized pattern looks like so:
1461 split:
1462 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1463 MIDDLE_BB:
1464 something
1465 fallthrough (edge E0)
1466 JOIN_BB:
1467 some more
1469 We check that MIDDLE_BB contains only one store, that that store
1470 doesn't trap (not via NOTRAP, but via checking if an access to the same
1471 memory location dominates us) and that the store has a "simple" RHS. */
1473 static bool
1474 cond_store_replacement (basic_block middle_bb, basic_block join_bb,
1475 edge e0, edge e1, struct pointer_set_t *nontrap)
1477 gimple assign = last_and_only_stmt (middle_bb);
1478 tree lhs, rhs, name, name2;
1479 gimple newphi, new_stmt;
1480 gimple_stmt_iterator gsi;
1481 source_location locus;
1483 /* Check if middle_bb contains of only one store. */
1484 if (!assign
1485 || !gimple_assign_single_p (assign)
1486 || gimple_has_volatile_ops (assign))
1487 return false;
1489 locus = gimple_location (assign);
1490 lhs = gimple_assign_lhs (assign);
1491 rhs = gimple_assign_rhs1 (assign);
1492 if (TREE_CODE (lhs) != MEM_REF
1493 || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME
1494 || !is_gimple_reg_type (TREE_TYPE (lhs)))
1495 return false;
1497 /* Prove that we can move the store down. We could also check
1498 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1499 whose value is not available readily, which we want to avoid. */
1500 if (!pointer_set_contains (nontrap, lhs))
1501 return false;
1503 /* Now we've checked the constraints, so do the transformation:
1504 1) Remove the single store. */
1505 gsi = gsi_for_stmt (assign);
1506 unlink_stmt_vdef (assign);
1507 gsi_remove (&gsi, true);
1508 release_defs (assign);
1510 /* 2) Insert a load from the memory of the store to the temporary
1511 on the edge which did not contain the store. */
1512 lhs = unshare_expr (lhs);
1513 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1514 new_stmt = gimple_build_assign (name, lhs);
1515 gimple_set_location (new_stmt, locus);
1516 gsi_insert_on_edge (e1, new_stmt);
1518 /* 3) Create a PHI node at the join block, with one argument
1519 holding the old RHS, and the other holding the temporary
1520 where we stored the old memory contents. */
1521 name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1522 newphi = create_phi_node (name2, join_bb);
1523 add_phi_arg (newphi, rhs, e0, locus);
1524 add_phi_arg (newphi, name, e1, locus);
1526 lhs = unshare_expr (lhs);
1527 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1529 /* 4) Insert that PHI node. */
1530 gsi = gsi_after_labels (join_bb);
1531 if (gsi_end_p (gsi))
1533 gsi = gsi_last_bb (join_bb);
1534 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1536 else
1537 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1539 return true;
1542 /* Do the main work of conditional store replacement. */
1544 static bool
1545 cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb,
1546 basic_block join_bb, gimple then_assign,
1547 gimple else_assign)
1549 tree lhs_base, lhs, then_rhs, else_rhs, name;
1550 source_location then_locus, else_locus;
1551 gimple_stmt_iterator gsi;
1552 gimple newphi, new_stmt;
1554 if (then_assign == NULL
1555 || !gimple_assign_single_p (then_assign)
1556 || gimple_clobber_p (then_assign)
1557 || gimple_has_volatile_ops (then_assign)
1558 || else_assign == NULL
1559 || !gimple_assign_single_p (else_assign)
1560 || gimple_clobber_p (else_assign)
1561 || gimple_has_volatile_ops (else_assign))
1562 return false;
1564 lhs = gimple_assign_lhs (then_assign);
1565 if (!is_gimple_reg_type (TREE_TYPE (lhs))
1566 || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0))
1567 return false;
1569 lhs_base = get_base_address (lhs);
1570 if (lhs_base == NULL_TREE
1571 || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF))
1572 return false;
1574 then_rhs = gimple_assign_rhs1 (then_assign);
1575 else_rhs = gimple_assign_rhs1 (else_assign);
1576 then_locus = gimple_location (then_assign);
1577 else_locus = gimple_location (else_assign);
1579 /* Now we've checked the constraints, so do the transformation:
1580 1) Remove the stores. */
1581 gsi = gsi_for_stmt (then_assign);
1582 unlink_stmt_vdef (then_assign);
1583 gsi_remove (&gsi, true);
1584 release_defs (then_assign);
1586 gsi = gsi_for_stmt (else_assign);
1587 unlink_stmt_vdef (else_assign);
1588 gsi_remove (&gsi, true);
1589 release_defs (else_assign);
1591 /* 2) Create a PHI node at the join block, with one argument
1592 holding the old RHS, and the other holding the temporary
1593 where we stored the old memory contents. */
1594 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1595 newphi = create_phi_node (name, join_bb);
1596 add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus);
1597 add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus);
1599 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1601 /* 3) Insert that PHI node. */
1602 gsi = gsi_after_labels (join_bb);
1603 if (gsi_end_p (gsi))
1605 gsi = gsi_last_bb (join_bb);
1606 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1608 else
1609 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1611 return true;
1614 /* Conditional store replacement. We already know
1615 that the recognized pattern looks like so:
1617 split:
1618 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1619 THEN_BB:
1621 X = Y;
1623 goto JOIN_BB;
1624 ELSE_BB:
1626 X = Z;
1628 fallthrough (edge E0)
1629 JOIN_BB:
1630 some more
1632 We check that it is safe to sink the store to JOIN_BB by verifying that
1633 there are no read-after-write or write-after-write dependencies in
1634 THEN_BB and ELSE_BB. */
1636 static bool
1637 cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
1638 basic_block join_bb)
1640 gimple then_assign = last_and_only_stmt (then_bb);
1641 gimple else_assign = last_and_only_stmt (else_bb);
1642 vec<data_reference_p> then_datarefs, else_datarefs;
1643 vec<ddr_p> then_ddrs, else_ddrs;
1644 gimple then_store, else_store;
1645 bool found, ok = false, res;
1646 struct data_dependence_relation *ddr;
1647 data_reference_p then_dr, else_dr;
1648 int i, j;
1649 tree then_lhs, else_lhs;
1650 vec<gimple> then_stores, else_stores;
1651 basic_block blocks[3];
1653 if (MAX_STORES_TO_SINK == 0)
1654 return false;
1656 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1657 if (then_assign && else_assign)
1658 return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1659 then_assign, else_assign);
1661 /* Find data references. */
1662 then_datarefs.create (1);
1663 else_datarefs.create (1);
1664 if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs)
1665 == chrec_dont_know)
1666 || !then_datarefs.length ()
1667 || (find_data_references_in_bb (NULL, else_bb, &else_datarefs)
1668 == chrec_dont_know)
1669 || !else_datarefs.length ())
1671 free_data_refs (then_datarefs);
1672 free_data_refs (else_datarefs);
1673 return false;
1676 /* Find pairs of stores with equal LHS. */
1677 then_stores.create (1);
1678 else_stores.create (1);
1679 FOR_EACH_VEC_ELT (then_datarefs, i, then_dr)
1681 if (DR_IS_READ (then_dr))
1682 continue;
1684 then_store = DR_STMT (then_dr);
1685 then_lhs = gimple_get_lhs (then_store);
1686 found = false;
1688 FOR_EACH_VEC_ELT (else_datarefs, j, else_dr)
1690 if (DR_IS_READ (else_dr))
1691 continue;
1693 else_store = DR_STMT (else_dr);
1694 else_lhs = gimple_get_lhs (else_store);
1696 if (operand_equal_p (then_lhs, else_lhs, 0))
1698 found = true;
1699 break;
1703 if (!found)
1704 continue;
1706 then_stores.safe_push (then_store);
1707 else_stores.safe_push (else_store);
1710 /* No pairs of stores found. */
1711 if (!then_stores.length ()
1712 || then_stores.length () > (unsigned) MAX_STORES_TO_SINK)
1714 free_data_refs (then_datarefs);
1715 free_data_refs (else_datarefs);
1716 then_stores.release ();
1717 else_stores.release ();
1718 return false;
1721 /* Compute and check data dependencies in both basic blocks. */
1722 then_ddrs.create (1);
1723 else_ddrs.create (1);
1724 if (!compute_all_dependences (then_datarefs, &then_ddrs,
1725 vNULL, false)
1726 || !compute_all_dependences (else_datarefs, &else_ddrs,
1727 vNULL, false))
1729 free_dependence_relations (then_ddrs);
1730 free_dependence_relations (else_ddrs);
1731 free_data_refs (then_datarefs);
1732 free_data_refs (else_datarefs);
1733 then_stores.release ();
1734 else_stores.release ();
1735 return false;
1737 blocks[0] = then_bb;
1738 blocks[1] = else_bb;
1739 blocks[2] = join_bb;
1740 renumber_gimple_stmt_uids_in_blocks (blocks, 3);
1742 /* Check that there are no read-after-write or write-after-write dependencies
1743 in THEN_BB. */
1744 FOR_EACH_VEC_ELT (then_ddrs, i, ddr)
1746 struct data_reference *dra = DDR_A (ddr);
1747 struct data_reference *drb = DDR_B (ddr);
1749 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1750 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1751 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1752 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1753 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1754 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1756 free_dependence_relations (then_ddrs);
1757 free_dependence_relations (else_ddrs);
1758 free_data_refs (then_datarefs);
1759 free_data_refs (else_datarefs);
1760 then_stores.release ();
1761 else_stores.release ();
1762 return false;
1766 /* Check that there are no read-after-write or write-after-write dependencies
1767 in ELSE_BB. */
1768 FOR_EACH_VEC_ELT (else_ddrs, i, ddr)
1770 struct data_reference *dra = DDR_A (ddr);
1771 struct data_reference *drb = DDR_B (ddr);
1773 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1774 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1775 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1776 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1777 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1778 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1780 free_dependence_relations (then_ddrs);
1781 free_dependence_relations (else_ddrs);
1782 free_data_refs (then_datarefs);
1783 free_data_refs (else_datarefs);
1784 then_stores.release ();
1785 else_stores.release ();
1786 return false;
1790 /* Sink stores with same LHS. */
1791 FOR_EACH_VEC_ELT (then_stores, i, then_store)
1793 else_store = else_stores[i];
1794 res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1795 then_store, else_store);
1796 ok = ok || res;
1799 free_dependence_relations (then_ddrs);
1800 free_dependence_relations (else_ddrs);
1801 free_data_refs (then_datarefs);
1802 free_data_refs (else_datarefs);
1803 then_stores.release ();
1804 else_stores.release ();
1806 return ok;
1809 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1811 static bool
1812 local_mem_dependence (gimple stmt, basic_block bb)
1814 tree vuse = gimple_vuse (stmt);
1815 gimple def;
1817 if (!vuse)
1818 return false;
1820 def = SSA_NAME_DEF_STMT (vuse);
1821 return (def && gimple_bb (def) == bb);
1824 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1825 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1826 and BB3 rejoins control flow following BB1 and BB2, look for
1827 opportunities to hoist loads as follows. If BB3 contains a PHI of
1828 two loads, one each occurring in BB1 and BB2, and the loads are
1829 provably of adjacent fields in the same structure, then move both
1830 loads into BB0. Of course this can only be done if there are no
1831 dependencies preventing such motion.
1833 One of the hoisted loads will always be speculative, so the
1834 transformation is currently conservative:
1836 - The fields must be strictly adjacent.
1837 - The two fields must occupy a single memory block that is
1838 guaranteed to not cross a page boundary.
1840 The last is difficult to prove, as such memory blocks should be
1841 aligned on the minimum of the stack alignment boundary and the
1842 alignment guaranteed by heap allocation interfaces. Thus we rely
1843 on a parameter for the alignment value.
1845 Provided a good value is used for the last case, the first
1846 restriction could possibly be relaxed. */
1848 static void
1849 hoist_adjacent_loads (basic_block bb0, basic_block bb1,
1850 basic_block bb2, basic_block bb3)
1852 int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE);
1853 unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT);
1854 gimple_stmt_iterator gsi;
1856 /* Walk the phis in bb3 looking for an opportunity. We are looking
1857 for phis of two SSA names, one each of which is defined in bb1 and
1858 bb2. */
1859 for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi))
1861 gimple phi_stmt = gsi_stmt (gsi);
1862 gimple def1, def2, defswap;
1863 tree arg1, arg2, ref1, ref2, field1, field2, fieldswap;
1864 tree tree_offset1, tree_offset2, tree_size2, next;
1865 int offset1, offset2, size2;
1866 unsigned align1;
1867 gimple_stmt_iterator gsi2;
1868 basic_block bb_for_def1, bb_for_def2;
1870 if (gimple_phi_num_args (phi_stmt) != 2
1871 || virtual_operand_p (gimple_phi_result (phi_stmt)))
1872 continue;
1874 arg1 = gimple_phi_arg_def (phi_stmt, 0);
1875 arg2 = gimple_phi_arg_def (phi_stmt, 1);
1877 if (TREE_CODE (arg1) != SSA_NAME
1878 || TREE_CODE (arg2) != SSA_NAME
1879 || SSA_NAME_IS_DEFAULT_DEF (arg1)
1880 || SSA_NAME_IS_DEFAULT_DEF (arg2))
1881 continue;
1883 def1 = SSA_NAME_DEF_STMT (arg1);
1884 def2 = SSA_NAME_DEF_STMT (arg2);
1886 if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2)
1887 && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2))
1888 continue;
1890 /* Check the mode of the arguments to be sure a conditional move
1891 can be generated for it. */
1892 if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1)))
1893 == CODE_FOR_nothing)
1894 continue;
1896 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
1897 if (!gimple_assign_single_p (def1)
1898 || !gimple_assign_single_p (def2)
1899 || gimple_has_volatile_ops (def1)
1900 || gimple_has_volatile_ops (def2))
1901 continue;
1903 ref1 = gimple_assign_rhs1 (def1);
1904 ref2 = gimple_assign_rhs1 (def2);
1906 if (TREE_CODE (ref1) != COMPONENT_REF
1907 || TREE_CODE (ref2) != COMPONENT_REF)
1908 continue;
1910 /* The zeroth operand of the two component references must be
1911 identical. It is not sufficient to compare get_base_address of
1912 the two references, because this could allow for different
1913 elements of the same array in the two trees. It is not safe to
1914 assume that the existence of one array element implies the
1915 existence of a different one. */
1916 if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0))
1917 continue;
1919 field1 = TREE_OPERAND (ref1, 1);
1920 field2 = TREE_OPERAND (ref2, 1);
1922 /* Check for field adjacency, and ensure field1 comes first. */
1923 for (next = DECL_CHAIN (field1);
1924 next && TREE_CODE (next) != FIELD_DECL;
1925 next = DECL_CHAIN (next))
1928 if (next != field2)
1930 for (next = DECL_CHAIN (field2);
1931 next && TREE_CODE (next) != FIELD_DECL;
1932 next = DECL_CHAIN (next))
1935 if (next != field1)
1936 continue;
1938 fieldswap = field1;
1939 field1 = field2;
1940 field2 = fieldswap;
1941 defswap = def1;
1942 def1 = def2;
1943 def2 = defswap;
1946 bb_for_def1 = gimple_bb (def1);
1947 bb_for_def2 = gimple_bb (def2);
1949 /* Check for proper alignment of the first field. */
1950 tree_offset1 = bit_position (field1);
1951 tree_offset2 = bit_position (field2);
1952 tree_size2 = DECL_SIZE (field2);
1954 if (!host_integerp (tree_offset1, 1)
1955 || !host_integerp (tree_offset2, 1)
1956 || !host_integerp (tree_size2, 1))
1957 continue;
1959 offset1 = TREE_INT_CST_LOW (tree_offset1);
1960 offset2 = TREE_INT_CST_LOW (tree_offset2);
1961 size2 = TREE_INT_CST_LOW (tree_size2);
1962 align1 = DECL_ALIGN (field1) % param_align_bits;
1964 if (offset1 % BITS_PER_UNIT != 0)
1965 continue;
1967 /* For profitability, the two field references should fit within
1968 a single cache line. */
1969 if (align1 + offset2 - offset1 + size2 > param_align_bits)
1970 continue;
1972 /* The two expressions cannot be dependent upon vdefs defined
1973 in bb1/bb2. */
1974 if (local_mem_dependence (def1, bb_for_def1)
1975 || local_mem_dependence (def2, bb_for_def2))
1976 continue;
1978 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
1979 bb0. We hoist the first one first so that a cache miss is handled
1980 efficiently regardless of hardware cache-fill policy. */
1981 gsi2 = gsi_for_stmt (def1);
1982 gsi_move_to_bb_end (&gsi2, bb0);
1983 gsi2 = gsi_for_stmt (def2);
1984 gsi_move_to_bb_end (&gsi2, bb0);
1986 if (dump_file && (dump_flags & TDF_DETAILS))
1988 fprintf (dump_file,
1989 "\nHoisting adjacent loads from %d and %d into %d: \n",
1990 bb_for_def1->index, bb_for_def2->index, bb0->index);
1991 print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS);
1992 print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS);
1997 /* Determine whether we should attempt to hoist adjacent loads out of
1998 diamond patterns in pass_phiopt. Always hoist loads if
1999 -fhoist-adjacent-loads is specified and the target machine has
2000 both a conditional move instruction and a defined cache line size. */
2002 static bool
2003 gate_hoist_loads (void)
2005 return (flag_hoist_adjacent_loads == 1
2006 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE)
2007 && HAVE_conditional_move);
2010 /* Always do these optimizations if we have SSA
2011 trees to work on. */
2012 static bool
2013 gate_phiopt (void)
2015 return 1;
2018 struct gimple_opt_pass pass_phiopt =
2021 GIMPLE_PASS,
2022 "phiopt", /* name */
2023 OPTGROUP_NONE, /* optinfo_flags */
2024 gate_phiopt, /* gate */
2025 tree_ssa_phiopt, /* execute */
2026 NULL, /* sub */
2027 NULL, /* next */
2028 0, /* static_pass_number */
2029 TV_TREE_PHIOPT, /* tv_id */
2030 PROP_cfg | PROP_ssa, /* properties_required */
2031 0, /* properties_provided */
2032 0, /* properties_destroyed */
2033 0, /* todo_flags_start */
2034 TODO_verify_ssa
2035 | TODO_verify_flow
2036 | TODO_verify_stmts /* todo_flags_finish */
2040 static bool
2041 gate_cselim (void)
2043 return flag_tree_cselim;
2046 struct gimple_opt_pass pass_cselim =
2049 GIMPLE_PASS,
2050 "cselim", /* name */
2051 OPTGROUP_NONE, /* optinfo_flags */
2052 gate_cselim, /* gate */
2053 tree_ssa_cs_elim, /* execute */
2054 NULL, /* sub */
2055 NULL, /* next */
2056 0, /* static_pass_number */
2057 TV_TREE_PHIOPT, /* tv_id */
2058 PROP_cfg | PROP_ssa, /* properties_required */
2059 0, /* properties_provided */
2060 0, /* properties_destroyed */
2061 0, /* todo_flags_start */
2062 TODO_verify_ssa
2063 | TODO_verify_flow
2064 | TODO_verify_stmts /* todo_flags_finish */