]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/tree-ssa-phiopt.c
tree-ssa-phiopt.c (minmax_replacement): Create new ssa name if we're not the only...
[thirdparty/gcc.git] / gcc / tree-ssa-phiopt.c
1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2015 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
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.
10
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.
15
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/>. */
19
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "cfghooks.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "rtl.h"
28 #include "ssa.h"
29 #include "alias.h"
30 #include "fold-const.h"
31 #include "stor-layout.h"
32 #include "flags.h"
33 #include "tm_p.h"
34 #include "cfganal.h"
35 #include "internal-fn.h"
36 #include "gimplify.h"
37 #include "gimple-iterator.h"
38 #include "gimplify-me.h"
39 #include "tree-cfg.h"
40 #include "insn-config.h"
41 #include "expmed.h"
42 #include "dojump.h"
43 #include "explow.h"
44 #include "calls.h"
45 #include "emit-rtl.h"
46 #include "varasm.h"
47 #include "stmt.h"
48 #include "expr.h"
49 #include "tree-dfa.h"
50 #include "tree-pass.h"
51 #include "langhooks.h"
52 #include "domwalk.h"
53 #include "cfgloop.h"
54 #include "tree-data-ref.h"
55 #include "gimple-pretty-print.h"
56 #include "insn-codes.h"
57 #include "optabs.h"
58 #include "tree-scalar-evolution.h"
59 #include "tree-inline.h"
60 #include "params.h"
61
62 static unsigned int tree_ssa_phiopt_worker (bool, bool);
63 static bool conditional_replacement (basic_block, basic_block,
64 edge, edge, gphi *, tree, tree);
65 static bool factor_out_conditional_conversion (edge, edge, gphi *, tree, tree);
66 static int value_replacement (basic_block, basic_block,
67 edge, edge, gimple, tree, tree);
68 static bool minmax_replacement (basic_block, basic_block,
69 edge, edge, gimple, tree, tree);
70 static bool abs_replacement (basic_block, basic_block,
71 edge, edge, gimple, tree, tree);
72 static bool cond_store_replacement (basic_block, basic_block, edge, edge,
73 hash_set<tree> *);
74 static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block);
75 static hash_set<tree> * get_non_trapping ();
76 static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree);
77 static void hoist_adjacent_loads (basic_block, basic_block,
78 basic_block, basic_block);
79 static bool gate_hoist_loads (void);
80
81 /* This pass tries to transform conditional stores into unconditional
82 ones, enabling further simplifications with the simpler then and else
83 blocks. In particular it replaces this:
84
85 bb0:
86 if (cond) goto bb2; else goto bb1;
87 bb1:
88 *p = RHS;
89 bb2:
90
91 with
92
93 bb0:
94 if (cond) goto bb1; else goto bb2;
95 bb1:
96 condtmp' = *p;
97 bb2:
98 condtmp = PHI <RHS, condtmp'>
99 *p = condtmp;
100
101 This transformation can only be done under several constraints,
102 documented below. It also replaces:
103
104 bb0:
105 if (cond) goto bb2; else goto bb1;
106 bb1:
107 *p = RHS1;
108 goto bb3;
109 bb2:
110 *p = RHS2;
111 bb3:
112
113 with
114
115 bb0:
116 if (cond) goto bb3; else goto bb1;
117 bb1:
118 bb3:
119 condtmp = PHI <RHS1, RHS2>
120 *p = condtmp; */
121
122 static unsigned int
123 tree_ssa_cs_elim (void)
124 {
125 unsigned todo;
126 /* ??? We are not interested in loop related info, but the following
127 will create it, ICEing as we didn't init loops with pre-headers.
128 An interfacing issue of find_data_references_in_bb. */
129 loop_optimizer_init (LOOPS_NORMAL);
130 scev_initialize ();
131 todo = tree_ssa_phiopt_worker (true, false);
132 scev_finalize ();
133 loop_optimizer_finalize ();
134 return todo;
135 }
136
137 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
138
139 static gphi *
140 single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1)
141 {
142 gimple_stmt_iterator i;
143 gphi *phi = NULL;
144 if (gimple_seq_singleton_p (seq))
145 return as_a <gphi *> (gsi_stmt (gsi_start (seq)));
146 for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i))
147 {
148 gphi *p = as_a <gphi *> (gsi_stmt (i));
149 /* If the PHI arguments are equal then we can skip this PHI. */
150 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx),
151 gimple_phi_arg_def (p, e1->dest_idx)))
152 continue;
153
154 /* If we already have a PHI that has the two edge arguments are
155 different, then return it is not a singleton for these PHIs. */
156 if (phi)
157 return NULL;
158
159 phi = p;
160 }
161 return phi;
162 }
163
164 /* The core routine of conditional store replacement and normal
165 phi optimizations. Both share much of the infrastructure in how
166 to match applicable basic block patterns. DO_STORE_ELIM is true
167 when we want to do conditional store replacement, false otherwise.
168 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
169 of diamond control flow patterns, false otherwise. */
170 static unsigned int
171 tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads)
172 {
173 basic_block bb;
174 basic_block *bb_order;
175 unsigned n, i;
176 bool cfgchanged = false;
177 hash_set<tree> *nontrap = 0;
178
179 if (do_store_elim)
180 /* Calculate the set of non-trapping memory accesses. */
181 nontrap = get_non_trapping ();
182
183 /* Search every basic block for COND_EXPR we may be able to optimize.
184
185 We walk the blocks in order that guarantees that a block with
186 a single predecessor is processed before the predecessor.
187 This ensures that we collapse inner ifs before visiting the
188 outer ones, and also that we do not try to visit a removed
189 block. */
190 bb_order = single_pred_before_succ_order ();
191 n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS;
192
193 for (i = 0; i < n; i++)
194 {
195 gimple cond_stmt;
196 gphi *phi;
197 basic_block bb1, bb2;
198 edge e1, e2;
199 tree arg0, arg1;
200
201 bb = bb_order[i];
202
203 cond_stmt = last_stmt (bb);
204 /* Check to see if the last statement is a GIMPLE_COND. */
205 if (!cond_stmt
206 || gimple_code (cond_stmt) != GIMPLE_COND)
207 continue;
208
209 e1 = EDGE_SUCC (bb, 0);
210 bb1 = e1->dest;
211 e2 = EDGE_SUCC (bb, 1);
212 bb2 = e2->dest;
213
214 /* We cannot do the optimization on abnormal edges. */
215 if ((e1->flags & EDGE_ABNORMAL) != 0
216 || (e2->flags & EDGE_ABNORMAL) != 0)
217 continue;
218
219 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
220 if (EDGE_COUNT (bb1->succs) == 0
221 || bb2 == NULL
222 || EDGE_COUNT (bb2->succs) == 0)
223 continue;
224
225 /* Find the bb which is the fall through to the other. */
226 if (EDGE_SUCC (bb1, 0)->dest == bb2)
227 ;
228 else if (EDGE_SUCC (bb2, 0)->dest == bb1)
229 {
230 std::swap (bb1, bb2);
231 std::swap (e1, e2);
232 }
233 else if (do_store_elim
234 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
235 {
236 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
237
238 if (!single_succ_p (bb1)
239 || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0
240 || !single_succ_p (bb2)
241 || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0
242 || EDGE_COUNT (bb3->preds) != 2)
243 continue;
244 if (cond_if_else_store_replacement (bb1, bb2, bb3))
245 cfgchanged = true;
246 continue;
247 }
248 else if (do_hoist_loads
249 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
250 {
251 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
252
253 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt)))
254 && single_succ_p (bb1)
255 && single_succ_p (bb2)
256 && single_pred_p (bb1)
257 && single_pred_p (bb2)
258 && EDGE_COUNT (bb->succs) == 2
259 && EDGE_COUNT (bb3->preds) == 2
260 /* If one edge or the other is dominant, a conditional move
261 is likely to perform worse than the well-predicted branch. */
262 && !predictable_edge_p (EDGE_SUCC (bb, 0))
263 && !predictable_edge_p (EDGE_SUCC (bb, 1)))
264 hoist_adjacent_loads (bb, bb1, bb2, bb3);
265 continue;
266 }
267 else
268 continue;
269
270 e1 = EDGE_SUCC (bb1, 0);
271
272 /* Make sure that bb1 is just a fall through. */
273 if (!single_succ_p (bb1)
274 || (e1->flags & EDGE_FALLTHRU) == 0)
275 continue;
276
277 /* Also make sure that bb1 only have one predecessor and that it
278 is bb. */
279 if (!single_pred_p (bb1)
280 || single_pred (bb1) != bb)
281 continue;
282
283 if (do_store_elim)
284 {
285 /* bb1 is the middle block, bb2 the join block, bb the split block,
286 e1 the fallthrough edge from bb1 to bb2. We can't do the
287 optimization if the join block has more than two predecessors. */
288 if (EDGE_COUNT (bb2->preds) > 2)
289 continue;
290 if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
291 cfgchanged = true;
292 }
293 else
294 {
295 gimple_seq phis = phi_nodes (bb2);
296 gimple_stmt_iterator gsi;
297 bool candorest = true;
298
299 /* Value replacement can work with more than one PHI
300 so try that first. */
301 for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
302 {
303 phi = as_a <gphi *> (gsi_stmt (gsi));
304 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
305 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
306 if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2)
307 {
308 candorest = false;
309 cfgchanged = true;
310 break;
311 }
312 }
313
314 if (!candorest)
315 continue;
316
317 phi = single_non_singleton_phi_for_edges (phis, e1, e2);
318 if (!phi)
319 continue;
320
321 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
322 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
323
324 /* Something is wrong if we cannot find the arguments in the PHI
325 node. */
326 gcc_assert (arg0 != NULL && arg1 != NULL);
327
328 if (factor_out_conditional_conversion (e1, e2, phi, arg0, arg1))
329 {
330 /* factor_out_conditional_conversion may create a new PHI in
331 BB2 and eliminate an existing PHI in BB2. Recompute values
332 that may be affected by that change. */
333 phis = phi_nodes (bb2);
334 phi = single_non_singleton_phi_for_edges (phis, e1, e2);
335 gcc_assert (phi);
336 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
337 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
338 gcc_assert (arg0 != NULL && arg1 != NULL);
339 }
340
341 /* Do the replacement of conditional if it can be done. */
342 if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
343 cfgchanged = true;
344 else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
345 cfgchanged = true;
346 else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
347 cfgchanged = true;
348 }
349 }
350
351 free (bb_order);
352
353 if (do_store_elim)
354 delete nontrap;
355 /* If the CFG has changed, we should cleanup the CFG. */
356 if (cfgchanged && do_store_elim)
357 {
358 /* In cond-store replacement we have added some loads on edges
359 and new VOPS (as we moved the store, and created a load). */
360 gsi_commit_edge_inserts ();
361 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
362 }
363 else if (cfgchanged)
364 return TODO_cleanup_cfg;
365 return 0;
366 }
367
368 /* Replace PHI node element whose edge is E in block BB with variable NEW.
369 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
370 is known to have two edges, one of which must reach BB). */
371
372 static void
373 replace_phi_edge_with_variable (basic_block cond_block,
374 edge e, gimple phi, tree new_tree)
375 {
376 basic_block bb = gimple_bb (phi);
377 basic_block block_to_remove;
378 gimple_stmt_iterator gsi;
379
380 /* Change the PHI argument to new. */
381 SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree);
382
383 /* Remove the empty basic block. */
384 if (EDGE_SUCC (cond_block, 0)->dest == bb)
385 {
386 EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU;
387 EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
388 EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE;
389 EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count;
390
391 block_to_remove = EDGE_SUCC (cond_block, 1)->dest;
392 }
393 else
394 {
395 EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU;
396 EDGE_SUCC (cond_block, 1)->flags
397 &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
398 EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE;
399 EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count;
400
401 block_to_remove = EDGE_SUCC (cond_block, 0)->dest;
402 }
403 delete_basic_block (block_to_remove);
404
405 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
406 gsi = gsi_last_bb (cond_block);
407 gsi_remove (&gsi, true);
408
409 if (dump_file && (dump_flags & TDF_DETAILS))
410 fprintf (dump_file,
411 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
412 cond_block->index,
413 bb->index);
414 }
415
416 /* PR66726: Factor conversion out of COND_EXPR. If the arguments of the PHI
417 stmt are CONVERT_STMT, factor out the conversion and perform the conversion
418 to the result of PHI stmt. */
419
420 static bool
421 factor_out_conditional_conversion (edge e0, edge e1, gphi *phi,
422 tree arg0, tree arg1)
423 {
424 gimple arg0_def_stmt = NULL, arg1_def_stmt = NULL, new_stmt;
425 tree new_arg0 = NULL_TREE, new_arg1 = NULL_TREE;
426 tree temp, result;
427 gphi *newphi;
428 gimple_stmt_iterator gsi, gsi_for_def;
429 source_location locus = gimple_location (phi);
430 enum tree_code convert_code;
431
432 /* Handle only PHI statements with two arguments. TODO: If all
433 other arguments to PHI are INTEGER_CST or if their defining
434 statement have the same unary operation, we can handle more
435 than two arguments too. */
436 if (gimple_phi_num_args (phi) != 2)
437 return false;
438
439 /* First canonicalize to simplify tests. */
440 if (TREE_CODE (arg0) != SSA_NAME)
441 {
442 std::swap (arg0, arg1);
443 std::swap (e0, e1);
444 }
445
446 if (TREE_CODE (arg0) != SSA_NAME
447 || (TREE_CODE (arg1) != SSA_NAME
448 && TREE_CODE (arg1) != INTEGER_CST))
449 return false;
450
451 /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is
452 a conversion. */
453 arg0_def_stmt = SSA_NAME_DEF_STMT (arg0);
454 if (!is_gimple_assign (arg0_def_stmt)
455 || !gimple_assign_cast_p (arg0_def_stmt))
456 return false;
457
458 /* Use the RHS as new_arg0. */
459 convert_code = gimple_assign_rhs_code (arg0_def_stmt);
460 new_arg0 = gimple_assign_rhs1 (arg0_def_stmt);
461 if (convert_code == VIEW_CONVERT_EXPR)
462 new_arg0 = TREE_OPERAND (new_arg0, 0);
463
464 if (TREE_CODE (arg1) == SSA_NAME)
465 {
466 /* Check if arg1 is an SSA_NAME and the stmt which defines arg1
467 is a conversion. */
468 arg1_def_stmt = SSA_NAME_DEF_STMT (arg1);
469 if (!is_gimple_assign (arg1_def_stmt)
470 || gimple_assign_rhs_code (arg1_def_stmt) != convert_code)
471 return false;
472
473 /* Use the RHS as new_arg1. */
474 new_arg1 = gimple_assign_rhs1 (arg1_def_stmt);
475 if (convert_code == VIEW_CONVERT_EXPR)
476 new_arg1 = TREE_OPERAND (new_arg1, 0);
477 }
478 else
479 {
480 /* If arg1 is an INTEGER_CST, fold it to new type. */
481 if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0))
482 && int_fits_type_p (arg1, TREE_TYPE (new_arg0)))
483 {
484 if (gimple_assign_cast_p (arg0_def_stmt))
485 new_arg1 = fold_convert (TREE_TYPE (new_arg0), arg1);
486 else
487 return false;
488 }
489 else
490 return false;
491 }
492
493 /* If arg0/arg1 have > 1 use, then this transformation actually increases
494 the number of expressions evaluated at runtime. */
495 if (!has_single_use (arg0)
496 || (arg1_def_stmt && !has_single_use (arg1)))
497 return false;
498
499 /* If types of new_arg0 and new_arg1 are different bailout. */
500 if (!types_compatible_p (TREE_TYPE (new_arg0), TREE_TYPE (new_arg1)))
501 return false;
502
503 /* Create a new PHI stmt. */
504 result = PHI_RESULT (phi);
505 temp = make_ssa_name (TREE_TYPE (new_arg0), NULL);
506 newphi = create_phi_node (temp, gimple_bb (phi));
507
508 if (dump_file && (dump_flags & TDF_DETAILS))
509 {
510 fprintf (dump_file, "PHI ");
511 print_generic_expr (dump_file, gimple_phi_result (phi), 0);
512 fprintf (dump_file,
513 " changed to factor conversion out from COND_EXPR.\n");
514 fprintf (dump_file, "New stmt with CAST that defines ");
515 print_generic_expr (dump_file, result, 0);
516 fprintf (dump_file, ".\n");
517 }
518
519 /* Remove the old cast(s) that has single use. */
520 gsi_for_def = gsi_for_stmt (arg0_def_stmt);
521 gsi_remove (&gsi_for_def, true);
522 if (arg1_def_stmt)
523 {
524 gsi_for_def = gsi_for_stmt (arg1_def_stmt);
525 gsi_remove (&gsi_for_def, true);
526 }
527
528 add_phi_arg (newphi, new_arg0, e0, locus);
529 add_phi_arg (newphi, new_arg1, e1, locus);
530
531 /* Create the conversion stmt and insert it. */
532 if (convert_code == VIEW_CONVERT_EXPR)
533 temp = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (result), temp);
534 new_stmt = gimple_build_assign (result, convert_code, temp);
535 gsi = gsi_after_labels (gimple_bb (phi));
536 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
537
538 /* Remove he original PHI stmt. */
539 gsi = gsi_for_stmt (phi);
540 gsi_remove (&gsi, true);
541 return true;
542 }
543
544 /* The function conditional_replacement does the main work of doing the
545 conditional replacement. Return true if the replacement is done.
546 Otherwise return false.
547 BB is the basic block where the replacement is going to be done on. ARG0
548 is argument 0 from PHI. Likewise for ARG1. */
549
550 static bool
551 conditional_replacement (basic_block cond_bb, basic_block middle_bb,
552 edge e0, edge e1, gphi *phi,
553 tree arg0, tree arg1)
554 {
555 tree result;
556 gimple stmt;
557 gassign *new_stmt;
558 tree cond;
559 gimple_stmt_iterator gsi;
560 edge true_edge, false_edge;
561 tree new_var, new_var2;
562 bool neg;
563
564 /* FIXME: Gimplification of complex type is too hard for now. */
565 /* We aren't prepared to handle vectors either (and it is a question
566 if it would be worthwhile anyway). */
567 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0))
568 || POINTER_TYPE_P (TREE_TYPE (arg0)))
569 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1))
570 || POINTER_TYPE_P (TREE_TYPE (arg1))))
571 return false;
572
573 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
574 convert it to the conditional. */
575 if ((integer_zerop (arg0) && integer_onep (arg1))
576 || (integer_zerop (arg1) && integer_onep (arg0)))
577 neg = false;
578 else if ((integer_zerop (arg0) && integer_all_onesp (arg1))
579 || (integer_zerop (arg1) && integer_all_onesp (arg0)))
580 neg = true;
581 else
582 return false;
583
584 if (!empty_block_p (middle_bb))
585 return false;
586
587 /* At this point we know we have a GIMPLE_COND with two successors.
588 One successor is BB, the other successor is an empty block which
589 falls through into BB.
590
591 There is a single PHI node at the join point (BB) and its arguments
592 are constants (0, 1) or (0, -1).
593
594 So, given the condition COND, and the two PHI arguments, we can
595 rewrite this PHI into non-branching code:
596
597 dest = (COND) or dest = COND'
598
599 We use the condition as-is if the argument associated with the
600 true edge has the value one or the argument associated with the
601 false edge as the value zero. Note that those conditions are not
602 the same since only one of the outgoing edges from the GIMPLE_COND
603 will directly reach BB and thus be associated with an argument. */
604
605 stmt = last_stmt (cond_bb);
606 result = PHI_RESULT (phi);
607
608 /* To handle special cases like floating point comparison, it is easier and
609 less error-prone to build a tree and gimplify it on the fly though it is
610 less efficient. */
611 cond = fold_build2_loc (gimple_location (stmt),
612 gimple_cond_code (stmt), boolean_type_node,
613 gimple_cond_lhs (stmt), gimple_cond_rhs (stmt));
614
615 /* We need to know which is the true edge and which is the false
616 edge so that we know when to invert the condition below. */
617 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
618 if ((e0 == true_edge && integer_zerop (arg0))
619 || (e0 == false_edge && !integer_zerop (arg0))
620 || (e1 == true_edge && integer_zerop (arg1))
621 || (e1 == false_edge && !integer_zerop (arg1)))
622 cond = fold_build1_loc (gimple_location (stmt),
623 TRUTH_NOT_EXPR, TREE_TYPE (cond), cond);
624
625 if (neg)
626 {
627 cond = fold_convert_loc (gimple_location (stmt),
628 TREE_TYPE (result), cond);
629 cond = fold_build1_loc (gimple_location (stmt),
630 NEGATE_EXPR, TREE_TYPE (cond), cond);
631 }
632
633 /* Insert our new statements at the end of conditional block before the
634 COND_STMT. */
635 gsi = gsi_for_stmt (stmt);
636 new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true,
637 GSI_SAME_STMT);
638
639 if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var)))
640 {
641 source_location locus_0, locus_1;
642
643 new_var2 = make_ssa_name (TREE_TYPE (result));
644 new_stmt = gimple_build_assign (new_var2, CONVERT_EXPR, new_var);
645 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
646 new_var = new_var2;
647
648 /* Set the locus to the first argument, unless is doesn't have one. */
649 locus_0 = gimple_phi_arg_location (phi, 0);
650 locus_1 = gimple_phi_arg_location (phi, 1);
651 if (locus_0 == UNKNOWN_LOCATION)
652 locus_0 = locus_1;
653 gimple_set_location (new_stmt, locus_0);
654 }
655
656 replace_phi_edge_with_variable (cond_bb, e1, phi, new_var);
657
658 /* Note that we optimized this PHI. */
659 return true;
660 }
661
662 /* Update *ARG which is defined in STMT so that it contains the
663 computed value if that seems profitable. Return true if the
664 statement is made dead by that rewriting. */
665
666 static bool
667 jump_function_from_stmt (tree *arg, gimple stmt)
668 {
669 enum tree_code code = gimple_assign_rhs_code (stmt);
670 if (code == ADDR_EXPR)
671 {
672 /* For arg = &p->i transform it to p, if possible. */
673 tree rhs1 = gimple_assign_rhs1 (stmt);
674 HOST_WIDE_INT offset;
675 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0),
676 &offset);
677 if (tem
678 && TREE_CODE (tem) == MEM_REF
679 && (mem_ref_offset (tem) + offset) == 0)
680 {
681 *arg = TREE_OPERAND (tem, 0);
682 return true;
683 }
684 }
685 /* TODO: Much like IPA-CP jump-functions we want to handle constant
686 additions symbolically here, and we'd need to update the comparison
687 code that compares the arg + cst tuples in our caller. For now the
688 code above exactly handles the VEC_BASE pattern from vec.h. */
689 return false;
690 }
691
692 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
693 of the form SSA_NAME NE 0.
694
695 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
696 the two input values of the EQ_EXPR match arg0 and arg1.
697
698 If so update *code and return TRUE. Otherwise return FALSE. */
699
700 static bool
701 rhs_is_fed_for_value_replacement (const_tree arg0, const_tree arg1,
702 enum tree_code *code, const_tree rhs)
703 {
704 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
705 statement. */
706 if (TREE_CODE (rhs) == SSA_NAME)
707 {
708 gimple def1 = SSA_NAME_DEF_STMT (rhs);
709
710 /* Verify the defining statement has an EQ_EXPR on the RHS. */
711 if (is_gimple_assign (def1) && gimple_assign_rhs_code (def1) == EQ_EXPR)
712 {
713 /* Finally verify the source operands of the EQ_EXPR are equal
714 to arg0 and arg1. */
715 tree op0 = gimple_assign_rhs1 (def1);
716 tree op1 = gimple_assign_rhs2 (def1);
717 if ((operand_equal_for_phi_arg_p (arg0, op0)
718 && operand_equal_for_phi_arg_p (arg1, op1))
719 || (operand_equal_for_phi_arg_p (arg0, op1)
720 && operand_equal_for_phi_arg_p (arg1, op0)))
721 {
722 /* We will perform the optimization. */
723 *code = gimple_assign_rhs_code (def1);
724 return true;
725 }
726 }
727 }
728 return false;
729 }
730
731 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
732
733 Also return TRUE if arg0/arg1 are equal to the source arguments of a
734 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
735
736 Return FALSE otherwise. */
737
738 static bool
739 operand_equal_for_value_replacement (const_tree arg0, const_tree arg1,
740 enum tree_code *code, gimple cond)
741 {
742 gimple def;
743 tree lhs = gimple_cond_lhs (cond);
744 tree rhs = gimple_cond_rhs (cond);
745
746 if ((operand_equal_for_phi_arg_p (arg0, lhs)
747 && operand_equal_for_phi_arg_p (arg1, rhs))
748 || (operand_equal_for_phi_arg_p (arg1, lhs)
749 && operand_equal_for_phi_arg_p (arg0, rhs)))
750 return true;
751
752 /* Now handle more complex case where we have an EQ comparison
753 which feeds a BIT_AND_EXPR which feeds COND.
754
755 First verify that COND is of the form SSA_NAME NE 0. */
756 if (*code != NE_EXPR || !integer_zerop (rhs)
757 || TREE_CODE (lhs) != SSA_NAME)
758 return false;
759
760 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
761 def = SSA_NAME_DEF_STMT (lhs);
762 if (!is_gimple_assign (def) || gimple_assign_rhs_code (def) != BIT_AND_EXPR)
763 return false;
764
765 /* Now verify arg0/arg1 correspond to the source arguments of an
766 EQ comparison feeding the BIT_AND_EXPR. */
767
768 tree tmp = gimple_assign_rhs1 (def);
769 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
770 return true;
771
772 tmp = gimple_assign_rhs2 (def);
773 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
774 return true;
775
776 return false;
777 }
778
779 /* Returns true if ARG is a neutral element for operation CODE
780 on the RIGHT side. */
781
782 static bool
783 neutral_element_p (tree_code code, tree arg, bool right)
784 {
785 switch (code)
786 {
787 case PLUS_EXPR:
788 case BIT_IOR_EXPR:
789 case BIT_XOR_EXPR:
790 return integer_zerop (arg);
791
792 case LROTATE_EXPR:
793 case RROTATE_EXPR:
794 case LSHIFT_EXPR:
795 case RSHIFT_EXPR:
796 case MINUS_EXPR:
797 case POINTER_PLUS_EXPR:
798 return right && integer_zerop (arg);
799
800 case MULT_EXPR:
801 return integer_onep (arg);
802
803 case TRUNC_DIV_EXPR:
804 case CEIL_DIV_EXPR:
805 case FLOOR_DIV_EXPR:
806 case ROUND_DIV_EXPR:
807 case EXACT_DIV_EXPR:
808 return right && integer_onep (arg);
809
810 case BIT_AND_EXPR:
811 return integer_all_onesp (arg);
812
813 default:
814 return false;
815 }
816 }
817
818 /* Returns true if ARG is an absorbing element for operation CODE. */
819
820 static bool
821 absorbing_element_p (tree_code code, tree arg)
822 {
823 switch (code)
824 {
825 case BIT_IOR_EXPR:
826 return integer_all_onesp (arg);
827
828 case MULT_EXPR:
829 case BIT_AND_EXPR:
830 return integer_zerop (arg);
831
832 default:
833 return false;
834 }
835 }
836
837 /* The function value_replacement does the main work of doing the value
838 replacement. Return non-zero if the replacement is done. Otherwise return
839 0. If we remove the middle basic block, return 2.
840 BB is the basic block where the replacement is going to be done on. ARG0
841 is argument 0 from the PHI. Likewise for ARG1. */
842
843 static int
844 value_replacement (basic_block cond_bb, basic_block middle_bb,
845 edge e0, edge e1, gimple phi,
846 tree arg0, tree arg1)
847 {
848 gimple_stmt_iterator gsi;
849 gimple cond;
850 edge true_edge, false_edge;
851 enum tree_code code;
852 bool emtpy_or_with_defined_p = true;
853
854 /* If the type says honor signed zeros we cannot do this
855 optimization. */
856 if (HONOR_SIGNED_ZEROS (arg1))
857 return 0;
858
859 /* If there is a statement in MIDDLE_BB that defines one of the PHI
860 arguments, then adjust arg0 or arg1. */
861 gsi = gsi_start_nondebug_after_labels_bb (middle_bb);
862 while (!gsi_end_p (gsi))
863 {
864 gimple stmt = gsi_stmt (gsi);
865 tree lhs;
866 gsi_next_nondebug (&gsi);
867 if (!is_gimple_assign (stmt))
868 {
869 emtpy_or_with_defined_p = false;
870 continue;
871 }
872 /* Now try to adjust arg0 or arg1 according to the computation
873 in the statement. */
874 lhs = gimple_assign_lhs (stmt);
875 if (!(lhs == arg0
876 && jump_function_from_stmt (&arg0, stmt))
877 || (lhs == arg1
878 && jump_function_from_stmt (&arg1, stmt)))
879 emtpy_or_with_defined_p = false;
880 }
881
882 cond = last_stmt (cond_bb);
883 code = gimple_cond_code (cond);
884
885 /* This transformation is only valid for equality comparisons. */
886 if (code != NE_EXPR && code != EQ_EXPR)
887 return 0;
888
889 /* We need to know which is the true edge and which is the false
890 edge so that we know if have abs or negative abs. */
891 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
892
893 /* At this point we know we have a COND_EXPR with two successors.
894 One successor is BB, the other successor is an empty block which
895 falls through into BB.
896
897 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
898
899 There is a single PHI node at the join point (BB) with two arguments.
900
901 We now need to verify that the two arguments in the PHI node match
902 the two arguments to the equality comparison. */
903
904 if (operand_equal_for_value_replacement (arg0, arg1, &code, cond))
905 {
906 edge e;
907 tree arg;
908
909 /* For NE_EXPR, we want to build an assignment result = arg where
910 arg is the PHI argument associated with the true edge. For
911 EQ_EXPR we want the PHI argument associated with the false edge. */
912 e = (code == NE_EXPR ? true_edge : false_edge);
913
914 /* Unfortunately, E may not reach BB (it may instead have gone to
915 OTHER_BLOCK). If that is the case, then we want the single outgoing
916 edge from OTHER_BLOCK which reaches BB and represents the desired
917 path from COND_BLOCK. */
918 if (e->dest == middle_bb)
919 e = single_succ_edge (e->dest);
920
921 /* Now we know the incoming edge to BB that has the argument for the
922 RHS of our new assignment statement. */
923 if (e0 == e)
924 arg = arg0;
925 else
926 arg = arg1;
927
928 /* If the middle basic block was empty or is defining the
929 PHI arguments and this is a single phi where the args are different
930 for the edges e0 and e1 then we can remove the middle basic block. */
931 if (emtpy_or_with_defined_p
932 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)),
933 e0, e1) == phi)
934 {
935 replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
936 /* Note that we optimized this PHI. */
937 return 2;
938 }
939 else
940 {
941 /* Replace the PHI arguments with arg. */
942 SET_PHI_ARG_DEF (phi, e0->dest_idx, arg);
943 SET_PHI_ARG_DEF (phi, e1->dest_idx, arg);
944 if (dump_file && (dump_flags & TDF_DETAILS))
945 {
946 fprintf (dump_file, "PHI ");
947 print_generic_expr (dump_file, gimple_phi_result (phi), 0);
948 fprintf (dump_file, " reduced for COND_EXPR in block %d to ",
949 cond_bb->index);
950 print_generic_expr (dump_file, arg, 0);
951 fprintf (dump_file, ".\n");
952 }
953 return 1;
954 }
955
956 }
957
958 /* Now optimize (x != 0) ? x + y : y to just y.
959 The following condition is too restrictive, there can easily be another
960 stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */
961 gimple assign = last_and_only_stmt (middle_bb);
962 if (!assign || gimple_code (assign) != GIMPLE_ASSIGN
963 || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS
964 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0))
965 && !POINTER_TYPE_P (TREE_TYPE (arg0))))
966 return 0;
967
968 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
969 if (!gimple_seq_empty_p (phi_nodes (middle_bb)))
970 return 0;
971
972 /* Only transform if it removes the condition. */
973 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1))
974 return 0;
975
976 /* Size-wise, this is always profitable. */
977 if (optimize_bb_for_speed_p (cond_bb)
978 /* The special case is useless if it has a low probability. */
979 && profile_status_for_fn (cfun) != PROFILE_ABSENT
980 && EDGE_PRED (middle_bb, 0)->probability < PROB_EVEN
981 /* If assign is cheap, there is no point avoiding it. */
982 && estimate_num_insns (assign, &eni_time_weights)
983 >= 3 * estimate_num_insns (cond, &eni_time_weights))
984 return 0;
985
986 tree lhs = gimple_assign_lhs (assign);
987 tree rhs1 = gimple_assign_rhs1 (assign);
988 tree rhs2 = gimple_assign_rhs2 (assign);
989 enum tree_code code_def = gimple_assign_rhs_code (assign);
990 tree cond_lhs = gimple_cond_lhs (cond);
991 tree cond_rhs = gimple_cond_rhs (cond);
992
993 if (((code == NE_EXPR && e1 == false_edge)
994 || (code == EQ_EXPR && e1 == true_edge))
995 && arg0 == lhs
996 && ((arg1 == rhs1
997 && operand_equal_for_phi_arg_p (rhs2, cond_lhs)
998 && neutral_element_p (code_def, cond_rhs, true))
999 || (arg1 == rhs2
1000 && operand_equal_for_phi_arg_p (rhs1, cond_lhs)
1001 && neutral_element_p (code_def, cond_rhs, false))
1002 || (operand_equal_for_phi_arg_p (arg1, cond_rhs)
1003 && (operand_equal_for_phi_arg_p (rhs2, cond_lhs)
1004 || operand_equal_for_phi_arg_p (rhs1, cond_lhs))
1005 && absorbing_element_p (code_def, cond_rhs))))
1006 {
1007 gsi = gsi_for_stmt (cond);
1008 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)))
1009 {
1010 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
1011 def-stmt in:
1012 if (n_5 != 0)
1013 goto <bb 3>;
1014 else
1015 goto <bb 4>;
1016
1017 <bb 3>:
1018 # RANGE [0, 4294967294]
1019 u_6 = n_5 + 4294967295;
1020
1021 <bb 4>:
1022 # u_3 = PHI <u_6(3), 4294967295(2)> */
1023 SSA_NAME_RANGE_INFO (lhs) = NULL;
1024 SSA_NAME_ANTI_RANGE_P (lhs) = 0;
1025 /* If available, we can use VR of phi result at least. */
1026 tree phires = gimple_phi_result (phi);
1027 struct range_info_def *phires_range_info
1028 = SSA_NAME_RANGE_INFO (phires);
1029 if (phires_range_info)
1030 duplicate_ssa_name_range_info (lhs, SSA_NAME_RANGE_TYPE (phires),
1031 phires_range_info);
1032 }
1033 gimple_stmt_iterator gsi_from = gsi_for_stmt (assign);
1034 gsi_move_before (&gsi_from, &gsi);
1035 replace_phi_edge_with_variable (cond_bb, e1, phi, lhs);
1036 return 2;
1037 }
1038
1039 return 0;
1040 }
1041
1042 /* The function minmax_replacement does the main work of doing the minmax
1043 replacement. Return true if the replacement is done. Otherwise return
1044 false.
1045 BB is the basic block where the replacement is going to be done on. ARG0
1046 is argument 0 from the PHI. Likewise for ARG1. */
1047
1048 static bool
1049 minmax_replacement (basic_block cond_bb, basic_block middle_bb,
1050 edge e0, edge e1, gimple phi,
1051 tree arg0, tree arg1)
1052 {
1053 tree result, type;
1054 gcond *cond;
1055 gassign *new_stmt;
1056 edge true_edge, false_edge;
1057 enum tree_code cmp, minmax, ass_code;
1058 tree smaller, larger, arg_true, arg_false;
1059 gimple_stmt_iterator gsi, gsi_from;
1060
1061 type = TREE_TYPE (PHI_RESULT (phi));
1062
1063 /* The optimization may be unsafe due to NaNs. */
1064 if (HONOR_NANS (type))
1065 return false;
1066
1067 cond = as_a <gcond *> (last_stmt (cond_bb));
1068 cmp = gimple_cond_code (cond);
1069
1070 /* This transformation is only valid for order comparisons. Record which
1071 operand is smaller/larger if the result of the comparison is true. */
1072 if (cmp == LT_EXPR || cmp == LE_EXPR)
1073 {
1074 smaller = gimple_cond_lhs (cond);
1075 larger = gimple_cond_rhs (cond);
1076 }
1077 else if (cmp == GT_EXPR || cmp == GE_EXPR)
1078 {
1079 smaller = gimple_cond_rhs (cond);
1080 larger = gimple_cond_lhs (cond);
1081 }
1082 else
1083 return false;
1084
1085 /* We need to know which is the true edge and which is the false
1086 edge so that we know if have abs or negative abs. */
1087 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1088
1089 /* Forward the edges over the middle basic block. */
1090 if (true_edge->dest == middle_bb)
1091 true_edge = EDGE_SUCC (true_edge->dest, 0);
1092 if (false_edge->dest == middle_bb)
1093 false_edge = EDGE_SUCC (false_edge->dest, 0);
1094
1095 if (true_edge == e0)
1096 {
1097 gcc_assert (false_edge == e1);
1098 arg_true = arg0;
1099 arg_false = arg1;
1100 }
1101 else
1102 {
1103 gcc_assert (false_edge == e0);
1104 gcc_assert (true_edge == e1);
1105 arg_true = arg1;
1106 arg_false = arg0;
1107 }
1108
1109 if (empty_block_p (middle_bb))
1110 {
1111 if (operand_equal_for_phi_arg_p (arg_true, smaller)
1112 && operand_equal_for_phi_arg_p (arg_false, larger))
1113 {
1114 /* Case
1115
1116 if (smaller < larger)
1117 rslt = smaller;
1118 else
1119 rslt = larger; */
1120 minmax = MIN_EXPR;
1121 }
1122 else if (operand_equal_for_phi_arg_p (arg_false, smaller)
1123 && operand_equal_for_phi_arg_p (arg_true, larger))
1124 minmax = MAX_EXPR;
1125 else
1126 return false;
1127 }
1128 else
1129 {
1130 /* Recognize the following case, assuming d <= u:
1131
1132 if (a <= u)
1133 b = MAX (a, d);
1134 x = PHI <b, u>
1135
1136 This is equivalent to
1137
1138 b = MAX (a, d);
1139 x = MIN (b, u); */
1140
1141 gimple assign = last_and_only_stmt (middle_bb);
1142 tree lhs, op0, op1, bound;
1143
1144 if (!assign
1145 || gimple_code (assign) != GIMPLE_ASSIGN)
1146 return false;
1147
1148 lhs = gimple_assign_lhs (assign);
1149 ass_code = gimple_assign_rhs_code (assign);
1150 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
1151 return false;
1152 op0 = gimple_assign_rhs1 (assign);
1153 op1 = gimple_assign_rhs2 (assign);
1154
1155 if (true_edge->src == middle_bb)
1156 {
1157 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1158 if (!operand_equal_for_phi_arg_p (lhs, arg_true))
1159 return false;
1160
1161 if (operand_equal_for_phi_arg_p (arg_false, larger))
1162 {
1163 /* Case
1164
1165 if (smaller < larger)
1166 {
1167 r' = MAX_EXPR (smaller, bound)
1168 }
1169 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1170 if (ass_code != MAX_EXPR)
1171 return false;
1172
1173 minmax = MIN_EXPR;
1174 if (operand_equal_for_phi_arg_p (op0, smaller))
1175 bound = op1;
1176 else if (operand_equal_for_phi_arg_p (op1, smaller))
1177 bound = op0;
1178 else
1179 return false;
1180
1181 /* We need BOUND <= LARGER. */
1182 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1183 bound, larger)))
1184 return false;
1185 }
1186 else if (operand_equal_for_phi_arg_p (arg_false, smaller))
1187 {
1188 /* Case
1189
1190 if (smaller < larger)
1191 {
1192 r' = MIN_EXPR (larger, bound)
1193 }
1194 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1195 if (ass_code != MIN_EXPR)
1196 return false;
1197
1198 minmax = MAX_EXPR;
1199 if (operand_equal_for_phi_arg_p (op0, larger))
1200 bound = op1;
1201 else if (operand_equal_for_phi_arg_p (op1, larger))
1202 bound = op0;
1203 else
1204 return false;
1205
1206 /* We need BOUND >= SMALLER. */
1207 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1208 bound, smaller)))
1209 return false;
1210 }
1211 else
1212 return false;
1213 }
1214 else
1215 {
1216 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1217 if (!operand_equal_for_phi_arg_p (lhs, arg_false))
1218 return false;
1219
1220 if (operand_equal_for_phi_arg_p (arg_true, larger))
1221 {
1222 /* Case
1223
1224 if (smaller > larger)
1225 {
1226 r' = MIN_EXPR (smaller, bound)
1227 }
1228 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1229 if (ass_code != MIN_EXPR)
1230 return false;
1231
1232 minmax = MAX_EXPR;
1233 if (operand_equal_for_phi_arg_p (op0, smaller))
1234 bound = op1;
1235 else if (operand_equal_for_phi_arg_p (op1, smaller))
1236 bound = op0;
1237 else
1238 return false;
1239
1240 /* We need BOUND >= LARGER. */
1241 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1242 bound, larger)))
1243 return false;
1244 }
1245 else if (operand_equal_for_phi_arg_p (arg_true, smaller))
1246 {
1247 /* Case
1248
1249 if (smaller > larger)
1250 {
1251 r' = MAX_EXPR (larger, bound)
1252 }
1253 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1254 if (ass_code != MAX_EXPR)
1255 return false;
1256
1257 minmax = MIN_EXPR;
1258 if (operand_equal_for_phi_arg_p (op0, larger))
1259 bound = op1;
1260 else if (operand_equal_for_phi_arg_p (op1, larger))
1261 bound = op0;
1262 else
1263 return false;
1264
1265 /* We need BOUND <= SMALLER. */
1266 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1267 bound, smaller)))
1268 return false;
1269 }
1270 else
1271 return false;
1272 }
1273
1274 /* Move the statement from the middle block. */
1275 gsi = gsi_last_bb (cond_bb);
1276 gsi_from = gsi_last_nondebug_bb (middle_bb);
1277 gsi_move_before (&gsi_from, &gsi);
1278 }
1279
1280 /* Create an SSA var to hold the min/max result. If we're the only
1281 things setting the target PHI, then we can clone the PHI
1282 variable. Otherwise we must create a new one. */
1283 result = PHI_RESULT (phi);
1284 if (EDGE_COUNT (gimple_bb (phi)->preds) == 2)
1285 result = duplicate_ssa_name (result, NULL);
1286 else
1287 result = make_ssa_name (TREE_TYPE (result));
1288
1289 /* Emit the statement to compute min/max. */
1290 new_stmt = gimple_build_assign (result, minmax, arg0, arg1);
1291 gsi = gsi_last_bb (cond_bb);
1292 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1293
1294 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1295 return true;
1296 }
1297
1298 /* The function absolute_replacement does the main work of doing the absolute
1299 replacement. Return true if the replacement is done. Otherwise return
1300 false.
1301 bb is the basic block where the replacement is going to be done on. arg0
1302 is argument 0 from the phi. Likewise for arg1. */
1303
1304 static bool
1305 abs_replacement (basic_block cond_bb, basic_block middle_bb,
1306 edge e0 ATTRIBUTE_UNUSED, edge e1,
1307 gimple phi, tree arg0, tree arg1)
1308 {
1309 tree result;
1310 gassign *new_stmt;
1311 gimple cond;
1312 gimple_stmt_iterator gsi;
1313 edge true_edge, false_edge;
1314 gimple assign;
1315 edge e;
1316 tree rhs, lhs;
1317 bool negate;
1318 enum tree_code cond_code;
1319
1320 /* If the type says honor signed zeros we cannot do this
1321 optimization. */
1322 if (HONOR_SIGNED_ZEROS (arg1))
1323 return false;
1324
1325 /* OTHER_BLOCK must have only one executable statement which must have the
1326 form arg0 = -arg1 or arg1 = -arg0. */
1327
1328 assign = last_and_only_stmt (middle_bb);
1329 /* If we did not find the proper negation assignment, then we can not
1330 optimize. */
1331 if (assign == NULL)
1332 return false;
1333
1334 /* If we got here, then we have found the only executable statement
1335 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1336 arg1 = -arg0, then we can not optimize. */
1337 if (gimple_code (assign) != GIMPLE_ASSIGN)
1338 return false;
1339
1340 lhs = gimple_assign_lhs (assign);
1341
1342 if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
1343 return false;
1344
1345 rhs = gimple_assign_rhs1 (assign);
1346
1347 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1348 if (!(lhs == arg0 && rhs == arg1)
1349 && !(lhs == arg1 && rhs == arg0))
1350 return false;
1351
1352 cond = last_stmt (cond_bb);
1353 result = PHI_RESULT (phi);
1354
1355 /* Only relationals comparing arg[01] against zero are interesting. */
1356 cond_code = gimple_cond_code (cond);
1357 if (cond_code != GT_EXPR && cond_code != GE_EXPR
1358 && cond_code != LT_EXPR && cond_code != LE_EXPR)
1359 return false;
1360
1361 /* Make sure the conditional is arg[01] OP y. */
1362 if (gimple_cond_lhs (cond) != rhs)
1363 return false;
1364
1365 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond)))
1366 ? real_zerop (gimple_cond_rhs (cond))
1367 : integer_zerop (gimple_cond_rhs (cond)))
1368 ;
1369 else
1370 return false;
1371
1372 /* We need to know which is the true edge and which is the false
1373 edge so that we know if have abs or negative abs. */
1374 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1375
1376 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1377 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1378 the false edge goes to OTHER_BLOCK. */
1379 if (cond_code == GT_EXPR || cond_code == GE_EXPR)
1380 e = true_edge;
1381 else
1382 e = false_edge;
1383
1384 if (e->dest == middle_bb)
1385 negate = true;
1386 else
1387 negate = false;
1388
1389 result = duplicate_ssa_name (result, NULL);
1390
1391 if (negate)
1392 lhs = make_ssa_name (TREE_TYPE (result));
1393 else
1394 lhs = result;
1395
1396 /* Build the modify expression with abs expression. */
1397 new_stmt = gimple_build_assign (lhs, ABS_EXPR, rhs);
1398
1399 gsi = gsi_last_bb (cond_bb);
1400 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1401
1402 if (negate)
1403 {
1404 /* Get the right GSI. We want to insert after the recently
1405 added ABS_EXPR statement (which we know is the first statement
1406 in the block. */
1407 new_stmt = gimple_build_assign (result, NEGATE_EXPR, lhs);
1408
1409 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1410 }
1411
1412 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1413
1414 /* Note that we optimized this PHI. */
1415 return true;
1416 }
1417
1418 /* Auxiliary functions to determine the set of memory accesses which
1419 can't trap because they are preceded by accesses to the same memory
1420 portion. We do that for MEM_REFs, so we only need to track
1421 the SSA_NAME of the pointer indirectly referenced. The algorithm
1422 simply is a walk over all instructions in dominator order. When
1423 we see an MEM_REF we determine if we've already seen a same
1424 ref anywhere up to the root of the dominator tree. If we do the
1425 current access can't trap. If we don't see any dominating access
1426 the current access might trap, but might also make later accesses
1427 non-trapping, so we remember it. We need to be careful with loads
1428 or stores, for instance a load might not trap, while a store would,
1429 so if we see a dominating read access this doesn't mean that a later
1430 write access would not trap. Hence we also need to differentiate the
1431 type of access(es) seen.
1432
1433 ??? We currently are very conservative and assume that a load might
1434 trap even if a store doesn't (write-only memory). This probably is
1435 overly conservative. */
1436
1437 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1438 through it was seen, which would constitute a no-trap region for
1439 same accesses. */
1440 struct name_to_bb
1441 {
1442 unsigned int ssa_name_ver;
1443 unsigned int phase;
1444 bool store;
1445 HOST_WIDE_INT offset, size;
1446 basic_block bb;
1447 };
1448
1449 /* Hashtable helpers. */
1450
1451 struct ssa_names_hasher : free_ptr_hash <name_to_bb>
1452 {
1453 static inline hashval_t hash (const name_to_bb *);
1454 static inline bool equal (const name_to_bb *, const name_to_bb *);
1455 };
1456
1457 /* Used for quick clearing of the hash-table when we see calls.
1458 Hash entries with phase < nt_call_phase are invalid. */
1459 static unsigned int nt_call_phase;
1460
1461 /* The hash function. */
1462
1463 inline hashval_t
1464 ssa_names_hasher::hash (const name_to_bb *n)
1465 {
1466 return n->ssa_name_ver ^ (((hashval_t) n->store) << 31)
1467 ^ (n->offset << 6) ^ (n->size << 3);
1468 }
1469
1470 /* The equality function of *P1 and *P2. */
1471
1472 inline bool
1473 ssa_names_hasher::equal (const name_to_bb *n1, const name_to_bb *n2)
1474 {
1475 return n1->ssa_name_ver == n2->ssa_name_ver
1476 && n1->store == n2->store
1477 && n1->offset == n2->offset
1478 && n1->size == n2->size;
1479 }
1480
1481 class nontrapping_dom_walker : public dom_walker
1482 {
1483 public:
1484 nontrapping_dom_walker (cdi_direction direction, hash_set<tree> *ps)
1485 : dom_walker (direction), m_nontrapping (ps), m_seen_ssa_names (128) {}
1486
1487 virtual void before_dom_children (basic_block);
1488 virtual void after_dom_children (basic_block);
1489
1490 private:
1491
1492 /* We see the expression EXP in basic block BB. If it's an interesting
1493 expression (an MEM_REF through an SSA_NAME) possibly insert the
1494 expression into the set NONTRAP or the hash table of seen expressions.
1495 STORE is true if this expression is on the LHS, otherwise it's on
1496 the RHS. */
1497 void add_or_mark_expr (basic_block, tree, bool);
1498
1499 hash_set<tree> *m_nontrapping;
1500
1501 /* The hash table for remembering what we've seen. */
1502 hash_table<ssa_names_hasher> m_seen_ssa_names;
1503 };
1504
1505 /* Called by walk_dominator_tree, when entering the block BB. */
1506 void
1507 nontrapping_dom_walker::before_dom_children (basic_block bb)
1508 {
1509 edge e;
1510 edge_iterator ei;
1511 gimple_stmt_iterator gsi;
1512
1513 /* If we haven't seen all our predecessors, clear the hash-table. */
1514 FOR_EACH_EDGE (e, ei, bb->preds)
1515 if ((((size_t)e->src->aux) & 2) == 0)
1516 {
1517 nt_call_phase++;
1518 break;
1519 }
1520
1521 /* Mark this BB as being on the path to dominator root and as visited. */
1522 bb->aux = (void*)(1 | 2);
1523
1524 /* And walk the statements in order. */
1525 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1526 {
1527 gimple stmt = gsi_stmt (gsi);
1528
1529 if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt))
1530 nt_call_phase++;
1531 else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt))
1532 {
1533 add_or_mark_expr (bb, gimple_assign_lhs (stmt), true);
1534 add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), false);
1535 }
1536 }
1537 }
1538
1539 /* Called by walk_dominator_tree, when basic block BB is exited. */
1540 void
1541 nontrapping_dom_walker::after_dom_children (basic_block bb)
1542 {
1543 /* This BB isn't on the path to dominator root anymore. */
1544 bb->aux = (void*)2;
1545 }
1546
1547 /* We see the expression EXP in basic block BB. If it's an interesting
1548 expression (an MEM_REF through an SSA_NAME) possibly insert the
1549 expression into the set NONTRAP or the hash table of seen expressions.
1550 STORE is true if this expression is on the LHS, otherwise it's on
1551 the RHS. */
1552 void
1553 nontrapping_dom_walker::add_or_mark_expr (basic_block bb, tree exp, bool store)
1554 {
1555 HOST_WIDE_INT size;
1556
1557 if (TREE_CODE (exp) == MEM_REF
1558 && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME
1559 && tree_fits_shwi_p (TREE_OPERAND (exp, 1))
1560 && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0)
1561 {
1562 tree name = TREE_OPERAND (exp, 0);
1563 struct name_to_bb map;
1564 name_to_bb **slot;
1565 struct name_to_bb *n2bb;
1566 basic_block found_bb = 0;
1567
1568 /* Try to find the last seen MEM_REF through the same
1569 SSA_NAME, which can trap. */
1570 map.ssa_name_ver = SSA_NAME_VERSION (name);
1571 map.phase = 0;
1572 map.bb = 0;
1573 map.store = store;
1574 map.offset = tree_to_shwi (TREE_OPERAND (exp, 1));
1575 map.size = size;
1576
1577 slot = m_seen_ssa_names.find_slot (&map, INSERT);
1578 n2bb = *slot;
1579 if (n2bb && n2bb->phase >= nt_call_phase)
1580 found_bb = n2bb->bb;
1581
1582 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1583 (it's in a basic block on the path from us to the dominator root)
1584 then we can't trap. */
1585 if (found_bb && (((size_t)found_bb->aux) & 1) == 1)
1586 {
1587 m_nontrapping->add (exp);
1588 }
1589 else
1590 {
1591 /* EXP might trap, so insert it into the hash table. */
1592 if (n2bb)
1593 {
1594 n2bb->phase = nt_call_phase;
1595 n2bb->bb = bb;
1596 }
1597 else
1598 {
1599 n2bb = XNEW (struct name_to_bb);
1600 n2bb->ssa_name_ver = SSA_NAME_VERSION (name);
1601 n2bb->phase = nt_call_phase;
1602 n2bb->bb = bb;
1603 n2bb->store = store;
1604 n2bb->offset = map.offset;
1605 n2bb->size = size;
1606 *slot = n2bb;
1607 }
1608 }
1609 }
1610 }
1611
1612 /* This is the entry point of gathering non trapping memory accesses.
1613 It will do a dominator walk over the whole function, and it will
1614 make use of the bb->aux pointers. It returns a set of trees
1615 (the MEM_REFs itself) which can't trap. */
1616 static hash_set<tree> *
1617 get_non_trapping (void)
1618 {
1619 nt_call_phase = 0;
1620 hash_set<tree> *nontrap = new hash_set<tree>;
1621 /* We're going to do a dominator walk, so ensure that we have
1622 dominance information. */
1623 calculate_dominance_info (CDI_DOMINATORS);
1624
1625 nontrapping_dom_walker (CDI_DOMINATORS, nontrap)
1626 .walk (cfun->cfg->x_entry_block_ptr);
1627
1628 clear_aux_for_blocks ();
1629 return nontrap;
1630 }
1631
1632 /* Do the main work of conditional store replacement. We already know
1633 that the recognized pattern looks like so:
1634
1635 split:
1636 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1637 MIDDLE_BB:
1638 something
1639 fallthrough (edge E0)
1640 JOIN_BB:
1641 some more
1642
1643 We check that MIDDLE_BB contains only one store, that that store
1644 doesn't trap (not via NOTRAP, but via checking if an access to the same
1645 memory location dominates us) and that the store has a "simple" RHS. */
1646
1647 static bool
1648 cond_store_replacement (basic_block middle_bb, basic_block join_bb,
1649 edge e0, edge e1, hash_set<tree> *nontrap)
1650 {
1651 gimple assign = last_and_only_stmt (middle_bb);
1652 tree lhs, rhs, name, name2;
1653 gphi *newphi;
1654 gassign *new_stmt;
1655 gimple_stmt_iterator gsi;
1656 source_location locus;
1657
1658 /* Check if middle_bb contains of only one store. */
1659 if (!assign
1660 || !gimple_assign_single_p (assign)
1661 || gimple_has_volatile_ops (assign))
1662 return false;
1663
1664 locus = gimple_location (assign);
1665 lhs = gimple_assign_lhs (assign);
1666 rhs = gimple_assign_rhs1 (assign);
1667 if (TREE_CODE (lhs) != MEM_REF
1668 || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME
1669 || !is_gimple_reg_type (TREE_TYPE (lhs)))
1670 return false;
1671
1672 /* Prove that we can move the store down. We could also check
1673 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1674 whose value is not available readily, which we want to avoid. */
1675 if (!nontrap->contains (lhs))
1676 return false;
1677
1678 /* Now we've checked the constraints, so do the transformation:
1679 1) Remove the single store. */
1680 gsi = gsi_for_stmt (assign);
1681 unlink_stmt_vdef (assign);
1682 gsi_remove (&gsi, true);
1683 release_defs (assign);
1684
1685 /* 2) Insert a load from the memory of the store to the temporary
1686 on the edge which did not contain the store. */
1687 lhs = unshare_expr (lhs);
1688 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1689 new_stmt = gimple_build_assign (name, lhs);
1690 gimple_set_location (new_stmt, locus);
1691 gsi_insert_on_edge (e1, new_stmt);
1692
1693 /* 3) Create a PHI node at the join block, with one argument
1694 holding the old RHS, and the other holding the temporary
1695 where we stored the old memory contents. */
1696 name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1697 newphi = create_phi_node (name2, join_bb);
1698 add_phi_arg (newphi, rhs, e0, locus);
1699 add_phi_arg (newphi, name, e1, locus);
1700
1701 lhs = unshare_expr (lhs);
1702 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1703
1704 /* 4) Insert that PHI node. */
1705 gsi = gsi_after_labels (join_bb);
1706 if (gsi_end_p (gsi))
1707 {
1708 gsi = gsi_last_bb (join_bb);
1709 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1710 }
1711 else
1712 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1713
1714 return true;
1715 }
1716
1717 /* Do the main work of conditional store replacement. */
1718
1719 static bool
1720 cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb,
1721 basic_block join_bb, gimple then_assign,
1722 gimple else_assign)
1723 {
1724 tree lhs_base, lhs, then_rhs, else_rhs, name;
1725 source_location then_locus, else_locus;
1726 gimple_stmt_iterator gsi;
1727 gphi *newphi;
1728 gassign *new_stmt;
1729
1730 if (then_assign == NULL
1731 || !gimple_assign_single_p (then_assign)
1732 || gimple_clobber_p (then_assign)
1733 || gimple_has_volatile_ops (then_assign)
1734 || else_assign == NULL
1735 || !gimple_assign_single_p (else_assign)
1736 || gimple_clobber_p (else_assign)
1737 || gimple_has_volatile_ops (else_assign))
1738 return false;
1739
1740 lhs = gimple_assign_lhs (then_assign);
1741 if (!is_gimple_reg_type (TREE_TYPE (lhs))
1742 || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0))
1743 return false;
1744
1745 lhs_base = get_base_address (lhs);
1746 if (lhs_base == NULL_TREE
1747 || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF))
1748 return false;
1749
1750 then_rhs = gimple_assign_rhs1 (then_assign);
1751 else_rhs = gimple_assign_rhs1 (else_assign);
1752 then_locus = gimple_location (then_assign);
1753 else_locus = gimple_location (else_assign);
1754
1755 /* Now we've checked the constraints, so do the transformation:
1756 1) Remove the stores. */
1757 gsi = gsi_for_stmt (then_assign);
1758 unlink_stmt_vdef (then_assign);
1759 gsi_remove (&gsi, true);
1760 release_defs (then_assign);
1761
1762 gsi = gsi_for_stmt (else_assign);
1763 unlink_stmt_vdef (else_assign);
1764 gsi_remove (&gsi, true);
1765 release_defs (else_assign);
1766
1767 /* 2) Create a PHI node at the join block, with one argument
1768 holding the old RHS, and the other holding the temporary
1769 where we stored the old memory contents. */
1770 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1771 newphi = create_phi_node (name, join_bb);
1772 add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus);
1773 add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus);
1774
1775 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1776
1777 /* 3) Insert that PHI node. */
1778 gsi = gsi_after_labels (join_bb);
1779 if (gsi_end_p (gsi))
1780 {
1781 gsi = gsi_last_bb (join_bb);
1782 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1783 }
1784 else
1785 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1786
1787 return true;
1788 }
1789
1790 /* Conditional store replacement. We already know
1791 that the recognized pattern looks like so:
1792
1793 split:
1794 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1795 THEN_BB:
1796 ...
1797 X = Y;
1798 ...
1799 goto JOIN_BB;
1800 ELSE_BB:
1801 ...
1802 X = Z;
1803 ...
1804 fallthrough (edge E0)
1805 JOIN_BB:
1806 some more
1807
1808 We check that it is safe to sink the store to JOIN_BB by verifying that
1809 there are no read-after-write or write-after-write dependencies in
1810 THEN_BB and ELSE_BB. */
1811
1812 static bool
1813 cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
1814 basic_block join_bb)
1815 {
1816 gimple then_assign = last_and_only_stmt (then_bb);
1817 gimple else_assign = last_and_only_stmt (else_bb);
1818 vec<data_reference_p> then_datarefs, else_datarefs;
1819 vec<ddr_p> then_ddrs, else_ddrs;
1820 gimple then_store, else_store;
1821 bool found, ok = false, res;
1822 struct data_dependence_relation *ddr;
1823 data_reference_p then_dr, else_dr;
1824 int i, j;
1825 tree then_lhs, else_lhs;
1826 basic_block blocks[3];
1827
1828 if (MAX_STORES_TO_SINK == 0)
1829 return false;
1830
1831 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1832 if (then_assign && else_assign)
1833 return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1834 then_assign, else_assign);
1835
1836 /* Find data references. */
1837 then_datarefs.create (1);
1838 else_datarefs.create (1);
1839 if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs)
1840 == chrec_dont_know)
1841 || !then_datarefs.length ()
1842 || (find_data_references_in_bb (NULL, else_bb, &else_datarefs)
1843 == chrec_dont_know)
1844 || !else_datarefs.length ())
1845 {
1846 free_data_refs (then_datarefs);
1847 free_data_refs (else_datarefs);
1848 return false;
1849 }
1850
1851 /* Find pairs of stores with equal LHS. */
1852 auto_vec<gimple, 1> then_stores, else_stores;
1853 FOR_EACH_VEC_ELT (then_datarefs, i, then_dr)
1854 {
1855 if (DR_IS_READ (then_dr))
1856 continue;
1857
1858 then_store = DR_STMT (then_dr);
1859 then_lhs = gimple_get_lhs (then_store);
1860 if (then_lhs == NULL_TREE)
1861 continue;
1862 found = false;
1863
1864 FOR_EACH_VEC_ELT (else_datarefs, j, else_dr)
1865 {
1866 if (DR_IS_READ (else_dr))
1867 continue;
1868
1869 else_store = DR_STMT (else_dr);
1870 else_lhs = gimple_get_lhs (else_store);
1871 if (else_lhs == NULL_TREE)
1872 continue;
1873
1874 if (operand_equal_p (then_lhs, else_lhs, 0))
1875 {
1876 found = true;
1877 break;
1878 }
1879 }
1880
1881 if (!found)
1882 continue;
1883
1884 then_stores.safe_push (then_store);
1885 else_stores.safe_push (else_store);
1886 }
1887
1888 /* No pairs of stores found. */
1889 if (!then_stores.length ()
1890 || then_stores.length () > (unsigned) MAX_STORES_TO_SINK)
1891 {
1892 free_data_refs (then_datarefs);
1893 free_data_refs (else_datarefs);
1894 return false;
1895 }
1896
1897 /* Compute and check data dependencies in both basic blocks. */
1898 then_ddrs.create (1);
1899 else_ddrs.create (1);
1900 if (!compute_all_dependences (then_datarefs, &then_ddrs,
1901 vNULL, false)
1902 || !compute_all_dependences (else_datarefs, &else_ddrs,
1903 vNULL, false))
1904 {
1905 free_dependence_relations (then_ddrs);
1906 free_dependence_relations (else_ddrs);
1907 free_data_refs (then_datarefs);
1908 free_data_refs (else_datarefs);
1909 return false;
1910 }
1911 blocks[0] = then_bb;
1912 blocks[1] = else_bb;
1913 blocks[2] = join_bb;
1914 renumber_gimple_stmt_uids_in_blocks (blocks, 3);
1915
1916 /* Check that there are no read-after-write or write-after-write dependencies
1917 in THEN_BB. */
1918 FOR_EACH_VEC_ELT (then_ddrs, i, ddr)
1919 {
1920 struct data_reference *dra = DDR_A (ddr);
1921 struct data_reference *drb = DDR_B (ddr);
1922
1923 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1924 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1925 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1926 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1927 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1928 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1929 {
1930 free_dependence_relations (then_ddrs);
1931 free_dependence_relations (else_ddrs);
1932 free_data_refs (then_datarefs);
1933 free_data_refs (else_datarefs);
1934 return false;
1935 }
1936 }
1937
1938 /* Check that there are no read-after-write or write-after-write dependencies
1939 in ELSE_BB. */
1940 FOR_EACH_VEC_ELT (else_ddrs, i, ddr)
1941 {
1942 struct data_reference *dra = DDR_A (ddr);
1943 struct data_reference *drb = DDR_B (ddr);
1944
1945 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1946 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1947 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1948 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1949 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1950 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1951 {
1952 free_dependence_relations (then_ddrs);
1953 free_dependence_relations (else_ddrs);
1954 free_data_refs (then_datarefs);
1955 free_data_refs (else_datarefs);
1956 return false;
1957 }
1958 }
1959
1960 /* Sink stores with same LHS. */
1961 FOR_EACH_VEC_ELT (then_stores, i, then_store)
1962 {
1963 else_store = else_stores[i];
1964 res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1965 then_store, else_store);
1966 ok = ok || res;
1967 }
1968
1969 free_dependence_relations (then_ddrs);
1970 free_dependence_relations (else_ddrs);
1971 free_data_refs (then_datarefs);
1972 free_data_refs (else_datarefs);
1973
1974 return ok;
1975 }
1976
1977 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1978
1979 static bool
1980 local_mem_dependence (gimple stmt, basic_block bb)
1981 {
1982 tree vuse = gimple_vuse (stmt);
1983 gimple def;
1984
1985 if (!vuse)
1986 return false;
1987
1988 def = SSA_NAME_DEF_STMT (vuse);
1989 return (def && gimple_bb (def) == bb);
1990 }
1991
1992 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1993 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1994 and BB3 rejoins control flow following BB1 and BB2, look for
1995 opportunities to hoist loads as follows. If BB3 contains a PHI of
1996 two loads, one each occurring in BB1 and BB2, and the loads are
1997 provably of adjacent fields in the same structure, then move both
1998 loads into BB0. Of course this can only be done if there are no
1999 dependencies preventing such motion.
2000
2001 One of the hoisted loads will always be speculative, so the
2002 transformation is currently conservative:
2003
2004 - The fields must be strictly adjacent.
2005 - The two fields must occupy a single memory block that is
2006 guaranteed to not cross a page boundary.
2007
2008 The last is difficult to prove, as such memory blocks should be
2009 aligned on the minimum of the stack alignment boundary and the
2010 alignment guaranteed by heap allocation interfaces. Thus we rely
2011 on a parameter for the alignment value.
2012
2013 Provided a good value is used for the last case, the first
2014 restriction could possibly be relaxed. */
2015
2016 static void
2017 hoist_adjacent_loads (basic_block bb0, basic_block bb1,
2018 basic_block bb2, basic_block bb3)
2019 {
2020 int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE);
2021 unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT);
2022 gphi_iterator gsi;
2023
2024 /* Walk the phis in bb3 looking for an opportunity. We are looking
2025 for phis of two SSA names, one each of which is defined in bb1 and
2026 bb2. */
2027 for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi))
2028 {
2029 gphi *phi_stmt = gsi.phi ();
2030 gimple def1, def2;
2031 tree arg1, arg2, ref1, ref2, field1, field2;
2032 tree tree_offset1, tree_offset2, tree_size2, next;
2033 int offset1, offset2, size2;
2034 unsigned align1;
2035 gimple_stmt_iterator gsi2;
2036 basic_block bb_for_def1, bb_for_def2;
2037
2038 if (gimple_phi_num_args (phi_stmt) != 2
2039 || virtual_operand_p (gimple_phi_result (phi_stmt)))
2040 continue;
2041
2042 arg1 = gimple_phi_arg_def (phi_stmt, 0);
2043 arg2 = gimple_phi_arg_def (phi_stmt, 1);
2044
2045 if (TREE_CODE (arg1) != SSA_NAME
2046 || TREE_CODE (arg2) != SSA_NAME
2047 || SSA_NAME_IS_DEFAULT_DEF (arg1)
2048 || SSA_NAME_IS_DEFAULT_DEF (arg2))
2049 continue;
2050
2051 def1 = SSA_NAME_DEF_STMT (arg1);
2052 def2 = SSA_NAME_DEF_STMT (arg2);
2053
2054 if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2)
2055 && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2))
2056 continue;
2057
2058 /* Check the mode of the arguments to be sure a conditional move
2059 can be generated for it. */
2060 if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1)))
2061 == CODE_FOR_nothing)
2062 continue;
2063
2064 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2065 if (!gimple_assign_single_p (def1)
2066 || !gimple_assign_single_p (def2)
2067 || gimple_has_volatile_ops (def1)
2068 || gimple_has_volatile_ops (def2))
2069 continue;
2070
2071 ref1 = gimple_assign_rhs1 (def1);
2072 ref2 = gimple_assign_rhs1 (def2);
2073
2074 if (TREE_CODE (ref1) != COMPONENT_REF
2075 || TREE_CODE (ref2) != COMPONENT_REF)
2076 continue;
2077
2078 /* The zeroth operand of the two component references must be
2079 identical. It is not sufficient to compare get_base_address of
2080 the two references, because this could allow for different
2081 elements of the same array in the two trees. It is not safe to
2082 assume that the existence of one array element implies the
2083 existence of a different one. */
2084 if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0))
2085 continue;
2086
2087 field1 = TREE_OPERAND (ref1, 1);
2088 field2 = TREE_OPERAND (ref2, 1);
2089
2090 /* Check for field adjacency, and ensure field1 comes first. */
2091 for (next = DECL_CHAIN (field1);
2092 next && TREE_CODE (next) != FIELD_DECL;
2093 next = DECL_CHAIN (next))
2094 ;
2095
2096 if (next != field2)
2097 {
2098 for (next = DECL_CHAIN (field2);
2099 next && TREE_CODE (next) != FIELD_DECL;
2100 next = DECL_CHAIN (next))
2101 ;
2102
2103 if (next != field1)
2104 continue;
2105
2106 std::swap (field1, field2);
2107 std::swap (def1, def2);
2108 }
2109
2110 bb_for_def1 = gimple_bb (def1);
2111 bb_for_def2 = gimple_bb (def2);
2112
2113 /* Check for proper alignment of the first field. */
2114 tree_offset1 = bit_position (field1);
2115 tree_offset2 = bit_position (field2);
2116 tree_size2 = DECL_SIZE (field2);
2117
2118 if (!tree_fits_uhwi_p (tree_offset1)
2119 || !tree_fits_uhwi_p (tree_offset2)
2120 || !tree_fits_uhwi_p (tree_size2))
2121 continue;
2122
2123 offset1 = tree_to_uhwi (tree_offset1);
2124 offset2 = tree_to_uhwi (tree_offset2);
2125 size2 = tree_to_uhwi (tree_size2);
2126 align1 = DECL_ALIGN (field1) % param_align_bits;
2127
2128 if (offset1 % BITS_PER_UNIT != 0)
2129 continue;
2130
2131 /* For profitability, the two field references should fit within
2132 a single cache line. */
2133 if (align1 + offset2 - offset1 + size2 > param_align_bits)
2134 continue;
2135
2136 /* The two expressions cannot be dependent upon vdefs defined
2137 in bb1/bb2. */
2138 if (local_mem_dependence (def1, bb_for_def1)
2139 || local_mem_dependence (def2, bb_for_def2))
2140 continue;
2141
2142 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2143 bb0. We hoist the first one first so that a cache miss is handled
2144 efficiently regardless of hardware cache-fill policy. */
2145 gsi2 = gsi_for_stmt (def1);
2146 gsi_move_to_bb_end (&gsi2, bb0);
2147 gsi2 = gsi_for_stmt (def2);
2148 gsi_move_to_bb_end (&gsi2, bb0);
2149
2150 if (dump_file && (dump_flags & TDF_DETAILS))
2151 {
2152 fprintf (dump_file,
2153 "\nHoisting adjacent loads from %d and %d into %d: \n",
2154 bb_for_def1->index, bb_for_def2->index, bb0->index);
2155 print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS);
2156 print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS);
2157 }
2158 }
2159 }
2160
2161 /* Determine whether we should attempt to hoist adjacent loads out of
2162 diamond patterns in pass_phiopt. Always hoist loads if
2163 -fhoist-adjacent-loads is specified and the target machine has
2164 both a conditional move instruction and a defined cache line size. */
2165
2166 static bool
2167 gate_hoist_loads (void)
2168 {
2169 return (flag_hoist_adjacent_loads == 1
2170 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE)
2171 && HAVE_conditional_move);
2172 }
2173
2174 /* This pass tries to replaces an if-then-else block with an
2175 assignment. We have four kinds of transformations. Some of these
2176 transformations are also performed by the ifcvt RTL optimizer.
2177
2178 Conditional Replacement
2179 -----------------------
2180
2181 This transformation, implemented in conditional_replacement,
2182 replaces
2183
2184 bb0:
2185 if (cond) goto bb2; else goto bb1;
2186 bb1:
2187 bb2:
2188 x = PHI <0 (bb1), 1 (bb0), ...>;
2189
2190 with
2191
2192 bb0:
2193 x' = cond;
2194 goto bb2;
2195 bb2:
2196 x = PHI <x' (bb0), ...>;
2197
2198 We remove bb1 as it becomes unreachable. This occurs often due to
2199 gimplification of conditionals.
2200
2201 Value Replacement
2202 -----------------
2203
2204 This transformation, implemented in value_replacement, replaces
2205
2206 bb0:
2207 if (a != b) goto bb2; else goto bb1;
2208 bb1:
2209 bb2:
2210 x = PHI <a (bb1), b (bb0), ...>;
2211
2212 with
2213
2214 bb0:
2215 bb2:
2216 x = PHI <b (bb0), ...>;
2217
2218 This opportunity can sometimes occur as a result of other
2219 optimizations.
2220
2221
2222 Another case caught by value replacement looks like this:
2223
2224 bb0:
2225 t1 = a == CONST;
2226 t2 = b > c;
2227 t3 = t1 & t2;
2228 if (t3 != 0) goto bb1; else goto bb2;
2229 bb1:
2230 bb2:
2231 x = PHI (CONST, a)
2232
2233 Gets replaced with:
2234 bb0:
2235 bb2:
2236 t1 = a == CONST;
2237 t2 = b > c;
2238 t3 = t1 & t2;
2239 x = a;
2240
2241 ABS Replacement
2242 ---------------
2243
2244 This transformation, implemented in abs_replacement, replaces
2245
2246 bb0:
2247 if (a >= 0) goto bb2; else goto bb1;
2248 bb1:
2249 x = -a;
2250 bb2:
2251 x = PHI <x (bb1), a (bb0), ...>;
2252
2253 with
2254
2255 bb0:
2256 x' = ABS_EXPR< a >;
2257 bb2:
2258 x = PHI <x' (bb0), ...>;
2259
2260 MIN/MAX Replacement
2261 -------------------
2262
2263 This transformation, minmax_replacement replaces
2264
2265 bb0:
2266 if (a <= b) goto bb2; else goto bb1;
2267 bb1:
2268 bb2:
2269 x = PHI <b (bb1), a (bb0), ...>;
2270
2271 with
2272
2273 bb0:
2274 x' = MIN_EXPR (a, b)
2275 bb2:
2276 x = PHI <x' (bb0), ...>;
2277
2278 A similar transformation is done for MAX_EXPR.
2279
2280
2281 This pass also performs a fifth transformation of a slightly different
2282 flavor.
2283
2284 Factor conversion in COND_EXPR
2285 ------------------------------
2286
2287 This transformation factors the conversion out of COND_EXPR with
2288 factor_out_conditional_conversion.
2289
2290 For example:
2291 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2292 <bb 3>:
2293 tmp = (int) a;
2294 <bb 4>:
2295 tmp = PHI <tmp, CST>
2296
2297 Into:
2298 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2299 <bb 3>:
2300 <bb 4>:
2301 a = PHI <a, CST>
2302 tmp = (int) a;
2303
2304 Adjacent Load Hoisting
2305 ----------------------
2306
2307 This transformation replaces
2308
2309 bb0:
2310 if (...) goto bb2; else goto bb1;
2311 bb1:
2312 x1 = (<expr>).field1;
2313 goto bb3;
2314 bb2:
2315 x2 = (<expr>).field2;
2316 bb3:
2317 # x = PHI <x1, x2>;
2318
2319 with
2320
2321 bb0:
2322 x1 = (<expr>).field1;
2323 x2 = (<expr>).field2;
2324 if (...) goto bb2; else goto bb1;
2325 bb1:
2326 goto bb3;
2327 bb2:
2328 bb3:
2329 # x = PHI <x1, x2>;
2330
2331 The purpose of this transformation is to enable generation of conditional
2332 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2333 the loads is speculative, the transformation is restricted to very
2334 specific cases to avoid introducing a page fault. We are looking for
2335 the common idiom:
2336
2337 if (...)
2338 x = y->left;
2339 else
2340 x = y->right;
2341
2342 where left and right are typically adjacent pointers in a tree structure. */
2343
2344 namespace {
2345
2346 const pass_data pass_data_phiopt =
2347 {
2348 GIMPLE_PASS, /* type */
2349 "phiopt", /* name */
2350 OPTGROUP_NONE, /* optinfo_flags */
2351 TV_TREE_PHIOPT, /* tv_id */
2352 ( PROP_cfg | PROP_ssa ), /* properties_required */
2353 0, /* properties_provided */
2354 0, /* properties_destroyed */
2355 0, /* todo_flags_start */
2356 0, /* todo_flags_finish */
2357 };
2358
2359 class pass_phiopt : public gimple_opt_pass
2360 {
2361 public:
2362 pass_phiopt (gcc::context *ctxt)
2363 : gimple_opt_pass (pass_data_phiopt, ctxt)
2364 {}
2365
2366 /* opt_pass methods: */
2367 opt_pass * clone () { return new pass_phiopt (m_ctxt); }
2368 virtual bool gate (function *) { return flag_ssa_phiopt; }
2369 virtual unsigned int execute (function *)
2370 {
2371 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2372 }
2373
2374 }; // class pass_phiopt
2375
2376 } // anon namespace
2377
2378 gimple_opt_pass *
2379 make_pass_phiopt (gcc::context *ctxt)
2380 {
2381 return new pass_phiopt (ctxt);
2382 }
2383
2384 namespace {
2385
2386 const pass_data pass_data_cselim =
2387 {
2388 GIMPLE_PASS, /* type */
2389 "cselim", /* name */
2390 OPTGROUP_NONE, /* optinfo_flags */
2391 TV_TREE_PHIOPT, /* tv_id */
2392 ( PROP_cfg | PROP_ssa ), /* properties_required */
2393 0, /* properties_provided */
2394 0, /* properties_destroyed */
2395 0, /* todo_flags_start */
2396 0, /* todo_flags_finish */
2397 };
2398
2399 class pass_cselim : public gimple_opt_pass
2400 {
2401 public:
2402 pass_cselim (gcc::context *ctxt)
2403 : gimple_opt_pass (pass_data_cselim, ctxt)
2404 {}
2405
2406 /* opt_pass methods: */
2407 virtual bool gate (function *) { return flag_tree_cselim; }
2408 virtual unsigned int execute (function *) { return tree_ssa_cs_elim (); }
2409
2410 }; // class pass_cselim
2411
2412 } // anon namespace
2413
2414 gimple_opt_pass *
2415 make_pass_cselim (gcc::context *ctxt)
2416 {
2417 return new pass_cselim (ctxt);
2418 }