]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/tree-ssa-phiopt.c
2015-06-17 Andrew MacLeod <amacleod@redhat.com>
[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 "tm.h"
24 #include "alias.h"
25 #include "symtab.h"
26 #include "tree.h"
27 #include "fold-const.h"
28 #include "stor-layout.h"
29 #include "flags.h"
30 #include "tm_p.h"
31 #include "predict.h"
32 #include "hard-reg-set.h"
33 #include "function.h"
34 #include "dominance.h"
35 #include "cfg.h"
36 #include "cfganal.h"
37 #include "basic-block.h"
38 #include "tree-ssa-alias.h"
39 #include "internal-fn.h"
40 #include "gimple-expr.h"
41 #include "gimple.h"
42 #include "gimplify.h"
43 #include "gimple-iterator.h"
44 #include "gimplify-me.h"
45 #include "gimple-ssa.h"
46 #include "tree-cfg.h"
47 #include "tree-phinodes.h"
48 #include "ssa-iterators.h"
49 #include "stringpool.h"
50 #include "tree-ssanames.h"
51 #include "rtl.h"
52 #include "insn-config.h"
53 #include "expmed.h"
54 #include "dojump.h"
55 #include "explow.h"
56 #include "calls.h"
57 #include "emit-rtl.h"
58 #include "varasm.h"
59 #include "stmt.h"
60 #include "expr.h"
61 #include "tree-dfa.h"
62 #include "tree-pass.h"
63 #include "langhooks.h"
64 #include "domwalk.h"
65 #include "cfgloop.h"
66 #include "tree-data-ref.h"
67 #include "gimple-pretty-print.h"
68 #include "insn-codes.h"
69 #include "optabs.h"
70 #include "tree-scalar-evolution.h"
71 #include "tree-inline.h"
72
73 static unsigned int tree_ssa_phiopt_worker (bool, bool);
74 static bool conditional_replacement (basic_block, basic_block,
75 edge, edge, gphi *, tree, tree);
76 static int value_replacement (basic_block, basic_block,
77 edge, edge, gimple, tree, tree);
78 static bool minmax_replacement (basic_block, basic_block,
79 edge, edge, gimple, tree, tree);
80 static bool abs_replacement (basic_block, basic_block,
81 edge, edge, gimple, tree, tree);
82 static bool cond_store_replacement (basic_block, basic_block, edge, edge,
83 hash_set<tree> *);
84 static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block);
85 static hash_set<tree> * get_non_trapping ();
86 static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree);
87 static void hoist_adjacent_loads (basic_block, basic_block,
88 basic_block, basic_block);
89 static bool gate_hoist_loads (void);
90
91 /* This pass tries to transform conditional stores into unconditional
92 ones, enabling further simplifications with the simpler then and else
93 blocks. In particular it replaces this:
94
95 bb0:
96 if (cond) goto bb2; else goto bb1;
97 bb1:
98 *p = RHS;
99 bb2:
100
101 with
102
103 bb0:
104 if (cond) goto bb1; else goto bb2;
105 bb1:
106 condtmp' = *p;
107 bb2:
108 condtmp = PHI <RHS, condtmp'>
109 *p = condtmp;
110
111 This transformation can only be done under several constraints,
112 documented below. It also replaces:
113
114 bb0:
115 if (cond) goto bb2; else goto bb1;
116 bb1:
117 *p = RHS1;
118 goto bb3;
119 bb2:
120 *p = RHS2;
121 bb3:
122
123 with
124
125 bb0:
126 if (cond) goto bb3; else goto bb1;
127 bb1:
128 bb3:
129 condtmp = PHI <RHS1, RHS2>
130 *p = condtmp; */
131
132 static unsigned int
133 tree_ssa_cs_elim (void)
134 {
135 unsigned todo;
136 /* ??? We are not interested in loop related info, but the following
137 will create it, ICEing as we didn't init loops with pre-headers.
138 An interfacing issue of find_data_references_in_bb. */
139 loop_optimizer_init (LOOPS_NORMAL);
140 scev_initialize ();
141 todo = tree_ssa_phiopt_worker (true, false);
142 scev_finalize ();
143 loop_optimizer_finalize ();
144 return todo;
145 }
146
147 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
148
149 static gphi *
150 single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1)
151 {
152 gimple_stmt_iterator i;
153 gphi *phi = NULL;
154 if (gimple_seq_singleton_p (seq))
155 return as_a <gphi *> (gsi_stmt (gsi_start (seq)));
156 for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i))
157 {
158 gphi *p = as_a <gphi *> (gsi_stmt (i));
159 /* If the PHI arguments are equal then we can skip this PHI. */
160 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx),
161 gimple_phi_arg_def (p, e1->dest_idx)))
162 continue;
163
164 /* If we already have a PHI that has the two edge arguments are
165 different, then return it is not a singleton for these PHIs. */
166 if (phi)
167 return NULL;
168
169 phi = p;
170 }
171 return phi;
172 }
173
174 /* The core routine of conditional store replacement and normal
175 phi optimizations. Both share much of the infrastructure in how
176 to match applicable basic block patterns. DO_STORE_ELIM is true
177 when we want to do conditional store replacement, false otherwise.
178 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
179 of diamond control flow patterns, false otherwise. */
180 static unsigned int
181 tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads)
182 {
183 basic_block bb;
184 basic_block *bb_order;
185 unsigned n, i;
186 bool cfgchanged = false;
187 hash_set<tree> *nontrap = 0;
188
189 if (do_store_elim)
190 /* Calculate the set of non-trapping memory accesses. */
191 nontrap = get_non_trapping ();
192
193 /* Search every basic block for COND_EXPR we may be able to optimize.
194
195 We walk the blocks in order that guarantees that a block with
196 a single predecessor is processed before the predecessor.
197 This ensures that we collapse inner ifs before visiting the
198 outer ones, and also that we do not try to visit a removed
199 block. */
200 bb_order = single_pred_before_succ_order ();
201 n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS;
202
203 for (i = 0; i < n; i++)
204 {
205 gimple cond_stmt;
206 gphi *phi;
207 basic_block bb1, bb2;
208 edge e1, e2;
209 tree arg0, arg1;
210
211 bb = bb_order[i];
212
213 cond_stmt = last_stmt (bb);
214 /* Check to see if the last statement is a GIMPLE_COND. */
215 if (!cond_stmt
216 || gimple_code (cond_stmt) != GIMPLE_COND)
217 continue;
218
219 e1 = EDGE_SUCC (bb, 0);
220 bb1 = e1->dest;
221 e2 = EDGE_SUCC (bb, 1);
222 bb2 = e2->dest;
223
224 /* We cannot do the optimization on abnormal edges. */
225 if ((e1->flags & EDGE_ABNORMAL) != 0
226 || (e2->flags & EDGE_ABNORMAL) != 0)
227 continue;
228
229 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
230 if (EDGE_COUNT (bb1->succs) == 0
231 || bb2 == NULL
232 || EDGE_COUNT (bb2->succs) == 0)
233 continue;
234
235 /* Find the bb which is the fall through to the other. */
236 if (EDGE_SUCC (bb1, 0)->dest == bb2)
237 ;
238 else if (EDGE_SUCC (bb2, 0)->dest == bb1)
239 {
240 basic_block bb_tmp = bb1;
241 edge e_tmp = e1;
242 bb1 = bb2;
243 bb2 = bb_tmp;
244 e1 = e2;
245 e2 = e_tmp;
246 }
247 else if (do_store_elim
248 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
249 {
250 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
251
252 if (!single_succ_p (bb1)
253 || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0
254 || !single_succ_p (bb2)
255 || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0
256 || EDGE_COUNT (bb3->preds) != 2)
257 continue;
258 if (cond_if_else_store_replacement (bb1, bb2, bb3))
259 cfgchanged = true;
260 continue;
261 }
262 else if (do_hoist_loads
263 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
264 {
265 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
266
267 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt)))
268 && single_succ_p (bb1)
269 && single_succ_p (bb2)
270 && single_pred_p (bb1)
271 && single_pred_p (bb2)
272 && EDGE_COUNT (bb->succs) == 2
273 && EDGE_COUNT (bb3->preds) == 2
274 /* If one edge or the other is dominant, a conditional move
275 is likely to perform worse than the well-predicted branch. */
276 && !predictable_edge_p (EDGE_SUCC (bb, 0))
277 && !predictable_edge_p (EDGE_SUCC (bb, 1)))
278 hoist_adjacent_loads (bb, bb1, bb2, bb3);
279 continue;
280 }
281 else
282 continue;
283
284 e1 = EDGE_SUCC (bb1, 0);
285
286 /* Make sure that bb1 is just a fall through. */
287 if (!single_succ_p (bb1)
288 || (e1->flags & EDGE_FALLTHRU) == 0)
289 continue;
290
291 /* Also make sure that bb1 only have one predecessor and that it
292 is bb. */
293 if (!single_pred_p (bb1)
294 || single_pred (bb1) != bb)
295 continue;
296
297 if (do_store_elim)
298 {
299 /* bb1 is the middle block, bb2 the join block, bb the split block,
300 e1 the fallthrough edge from bb1 to bb2. We can't do the
301 optimization if the join block has more than two predecessors. */
302 if (EDGE_COUNT (bb2->preds) > 2)
303 continue;
304 if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
305 cfgchanged = true;
306 }
307 else
308 {
309 gimple_seq phis = phi_nodes (bb2);
310 gimple_stmt_iterator gsi;
311 bool candorest = true;
312
313 /* Value replacement can work with more than one PHI
314 so try that first. */
315 for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
316 {
317 phi = as_a <gphi *> (gsi_stmt (gsi));
318 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
319 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
320 if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2)
321 {
322 candorest = false;
323 cfgchanged = true;
324 break;
325 }
326 }
327
328 if (!candorest)
329 continue;
330
331 phi = single_non_singleton_phi_for_edges (phis, e1, e2);
332 if (!phi)
333 continue;
334
335 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
336 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
337
338 /* Something is wrong if we cannot find the arguments in the PHI
339 node. */
340 gcc_assert (arg0 != NULL && arg1 != NULL);
341
342 /* Do the replacement of conditional if it can be done. */
343 if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
344 cfgchanged = true;
345 else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
346 cfgchanged = true;
347 else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
348 cfgchanged = true;
349 }
350 }
351
352 free (bb_order);
353
354 if (do_store_elim)
355 delete nontrap;
356 /* If the CFG has changed, we should cleanup the CFG. */
357 if (cfgchanged && do_store_elim)
358 {
359 /* In cond-store replacement we have added some loads on edges
360 and new VOPS (as we moved the store, and created a load). */
361 gsi_commit_edge_inserts ();
362 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
363 }
364 else if (cfgchanged)
365 return TODO_cleanup_cfg;
366 return 0;
367 }
368
369 /* Replace PHI node element whose edge is E in block BB with variable NEW.
370 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
371 is known to have two edges, one of which must reach BB). */
372
373 static void
374 replace_phi_edge_with_variable (basic_block cond_block,
375 edge e, gimple phi, tree new_tree)
376 {
377 basic_block bb = gimple_bb (phi);
378 basic_block block_to_remove;
379 gimple_stmt_iterator gsi;
380
381 /* Change the PHI argument to new. */
382 SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree);
383
384 /* Remove the empty basic block. */
385 if (EDGE_SUCC (cond_block, 0)->dest == bb)
386 {
387 EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU;
388 EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
389 EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE;
390 EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count;
391
392 block_to_remove = EDGE_SUCC (cond_block, 1)->dest;
393 }
394 else
395 {
396 EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU;
397 EDGE_SUCC (cond_block, 1)->flags
398 &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
399 EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE;
400 EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count;
401
402 block_to_remove = EDGE_SUCC (cond_block, 0)->dest;
403 }
404 delete_basic_block (block_to_remove);
405
406 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
407 gsi = gsi_last_bb (cond_block);
408 gsi_remove (&gsi, true);
409
410 if (dump_file && (dump_flags & TDF_DETAILS))
411 fprintf (dump_file,
412 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
413 cond_block->index,
414 bb->index);
415 }
416
417 /* The function conditional_replacement does the main work of doing the
418 conditional replacement. Return true if the replacement is done.
419 Otherwise return false.
420 BB is the basic block where the replacement is going to be done on. ARG0
421 is argument 0 from PHI. Likewise for ARG1. */
422
423 static bool
424 conditional_replacement (basic_block cond_bb, basic_block middle_bb,
425 edge e0, edge e1, gphi *phi,
426 tree arg0, tree arg1)
427 {
428 tree result;
429 gimple stmt;
430 gassign *new_stmt;
431 tree cond;
432 gimple_stmt_iterator gsi;
433 edge true_edge, false_edge;
434 tree new_var, new_var2;
435 bool neg;
436
437 /* FIXME: Gimplification of complex type is too hard for now. */
438 /* We aren't prepared to handle vectors either (and it is a question
439 if it would be worthwhile anyway). */
440 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0))
441 || POINTER_TYPE_P (TREE_TYPE (arg0)))
442 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1))
443 || POINTER_TYPE_P (TREE_TYPE (arg1))))
444 return false;
445
446 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
447 convert it to the conditional. */
448 if ((integer_zerop (arg0) && integer_onep (arg1))
449 || (integer_zerop (arg1) && integer_onep (arg0)))
450 neg = false;
451 else if ((integer_zerop (arg0) && integer_all_onesp (arg1))
452 || (integer_zerop (arg1) && integer_all_onesp (arg0)))
453 neg = true;
454 else
455 return false;
456
457 if (!empty_block_p (middle_bb))
458 return false;
459
460 /* At this point we know we have a GIMPLE_COND with two successors.
461 One successor is BB, the other successor is an empty block which
462 falls through into BB.
463
464 There is a single PHI node at the join point (BB) and its arguments
465 are constants (0, 1) or (0, -1).
466
467 So, given the condition COND, and the two PHI arguments, we can
468 rewrite this PHI into non-branching code:
469
470 dest = (COND) or dest = COND'
471
472 We use the condition as-is if the argument associated with the
473 true edge has the value one or the argument associated with the
474 false edge as the value zero. Note that those conditions are not
475 the same since only one of the outgoing edges from the GIMPLE_COND
476 will directly reach BB and thus be associated with an argument. */
477
478 stmt = last_stmt (cond_bb);
479 result = PHI_RESULT (phi);
480
481 /* To handle special cases like floating point comparison, it is easier and
482 less error-prone to build a tree and gimplify it on the fly though it is
483 less efficient. */
484 cond = fold_build2_loc (gimple_location (stmt),
485 gimple_cond_code (stmt), boolean_type_node,
486 gimple_cond_lhs (stmt), gimple_cond_rhs (stmt));
487
488 /* We need to know which is the true edge and which is the false
489 edge so that we know when to invert the condition below. */
490 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
491 if ((e0 == true_edge && integer_zerop (arg0))
492 || (e0 == false_edge && !integer_zerop (arg0))
493 || (e1 == true_edge && integer_zerop (arg1))
494 || (e1 == false_edge && !integer_zerop (arg1)))
495 cond = fold_build1_loc (gimple_location (stmt),
496 TRUTH_NOT_EXPR, TREE_TYPE (cond), cond);
497
498 if (neg)
499 {
500 cond = fold_convert_loc (gimple_location (stmt),
501 TREE_TYPE (result), cond);
502 cond = fold_build1_loc (gimple_location (stmt),
503 NEGATE_EXPR, TREE_TYPE (cond), cond);
504 }
505
506 /* Insert our new statements at the end of conditional block before the
507 COND_STMT. */
508 gsi = gsi_for_stmt (stmt);
509 new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true,
510 GSI_SAME_STMT);
511
512 if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var)))
513 {
514 source_location locus_0, locus_1;
515
516 new_var2 = make_ssa_name (TREE_TYPE (result));
517 new_stmt = gimple_build_assign (new_var2, CONVERT_EXPR, new_var);
518 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
519 new_var = new_var2;
520
521 /* Set the locus to the first argument, unless is doesn't have one. */
522 locus_0 = gimple_phi_arg_location (phi, 0);
523 locus_1 = gimple_phi_arg_location (phi, 1);
524 if (locus_0 == UNKNOWN_LOCATION)
525 locus_0 = locus_1;
526 gimple_set_location (new_stmt, locus_0);
527 }
528
529 replace_phi_edge_with_variable (cond_bb, e1, phi, new_var);
530
531 /* Note that we optimized this PHI. */
532 return true;
533 }
534
535 /* Update *ARG which is defined in STMT so that it contains the
536 computed value if that seems profitable. Return true if the
537 statement is made dead by that rewriting. */
538
539 static bool
540 jump_function_from_stmt (tree *arg, gimple stmt)
541 {
542 enum tree_code code = gimple_assign_rhs_code (stmt);
543 if (code == ADDR_EXPR)
544 {
545 /* For arg = &p->i transform it to p, if possible. */
546 tree rhs1 = gimple_assign_rhs1 (stmt);
547 HOST_WIDE_INT offset;
548 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0),
549 &offset);
550 if (tem
551 && TREE_CODE (tem) == MEM_REF
552 && (mem_ref_offset (tem) + offset) == 0)
553 {
554 *arg = TREE_OPERAND (tem, 0);
555 return true;
556 }
557 }
558 /* TODO: Much like IPA-CP jump-functions we want to handle constant
559 additions symbolically here, and we'd need to update the comparison
560 code that compares the arg + cst tuples in our caller. For now the
561 code above exactly handles the VEC_BASE pattern from vec.h. */
562 return false;
563 }
564
565 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
566 of the form SSA_NAME NE 0.
567
568 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
569 the two input values of the EQ_EXPR match arg0 and arg1.
570
571 If so update *code and return TRUE. Otherwise return FALSE. */
572
573 static bool
574 rhs_is_fed_for_value_replacement (const_tree arg0, const_tree arg1,
575 enum tree_code *code, const_tree rhs)
576 {
577 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
578 statement. */
579 if (TREE_CODE (rhs) == SSA_NAME)
580 {
581 gimple def1 = SSA_NAME_DEF_STMT (rhs);
582
583 /* Verify the defining statement has an EQ_EXPR on the RHS. */
584 if (is_gimple_assign (def1) && gimple_assign_rhs_code (def1) == EQ_EXPR)
585 {
586 /* Finally verify the source operands of the EQ_EXPR are equal
587 to arg0 and arg1. */
588 tree op0 = gimple_assign_rhs1 (def1);
589 tree op1 = gimple_assign_rhs2 (def1);
590 if ((operand_equal_for_phi_arg_p (arg0, op0)
591 && operand_equal_for_phi_arg_p (arg1, op1))
592 || (operand_equal_for_phi_arg_p (arg0, op1)
593 && operand_equal_for_phi_arg_p (arg1, op0)))
594 {
595 /* We will perform the optimization. */
596 *code = gimple_assign_rhs_code (def1);
597 return true;
598 }
599 }
600 }
601 return false;
602 }
603
604 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
605
606 Also return TRUE if arg0/arg1 are equal to the source arguments of a
607 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
608
609 Return FALSE otherwise. */
610
611 static bool
612 operand_equal_for_value_replacement (const_tree arg0, const_tree arg1,
613 enum tree_code *code, gimple cond)
614 {
615 gimple def;
616 tree lhs = gimple_cond_lhs (cond);
617 tree rhs = gimple_cond_rhs (cond);
618
619 if ((operand_equal_for_phi_arg_p (arg0, lhs)
620 && operand_equal_for_phi_arg_p (arg1, rhs))
621 || (operand_equal_for_phi_arg_p (arg1, lhs)
622 && operand_equal_for_phi_arg_p (arg0, rhs)))
623 return true;
624
625 /* Now handle more complex case where we have an EQ comparison
626 which feeds a BIT_AND_EXPR which feeds COND.
627
628 First verify that COND is of the form SSA_NAME NE 0. */
629 if (*code != NE_EXPR || !integer_zerop (rhs)
630 || TREE_CODE (lhs) != SSA_NAME)
631 return false;
632
633 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
634 def = SSA_NAME_DEF_STMT (lhs);
635 if (!is_gimple_assign (def) || gimple_assign_rhs_code (def) != BIT_AND_EXPR)
636 return false;
637
638 /* Now verify arg0/arg1 correspond to the source arguments of an
639 EQ comparison feeding the BIT_AND_EXPR. */
640
641 tree tmp = gimple_assign_rhs1 (def);
642 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
643 return true;
644
645 tmp = gimple_assign_rhs2 (def);
646 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
647 return true;
648
649 return false;
650 }
651
652 /* Returns true if ARG is a neutral element for operation CODE
653 on the RIGHT side. */
654
655 static bool
656 neutral_element_p (tree_code code, tree arg, bool right)
657 {
658 switch (code)
659 {
660 case PLUS_EXPR:
661 case BIT_IOR_EXPR:
662 case BIT_XOR_EXPR:
663 return integer_zerop (arg);
664
665 case LROTATE_EXPR:
666 case RROTATE_EXPR:
667 case LSHIFT_EXPR:
668 case RSHIFT_EXPR:
669 case MINUS_EXPR:
670 case POINTER_PLUS_EXPR:
671 return right && integer_zerop (arg);
672
673 case MULT_EXPR:
674 return integer_onep (arg);
675
676 case TRUNC_DIV_EXPR:
677 case CEIL_DIV_EXPR:
678 case FLOOR_DIV_EXPR:
679 case ROUND_DIV_EXPR:
680 case EXACT_DIV_EXPR:
681 return right && integer_onep (arg);
682
683 case BIT_AND_EXPR:
684 return integer_all_onesp (arg);
685
686 default:
687 return false;
688 }
689 }
690
691 /* Returns true if ARG is an absorbing element for operation CODE. */
692
693 static bool
694 absorbing_element_p (tree_code code, tree arg)
695 {
696 switch (code)
697 {
698 case BIT_IOR_EXPR:
699 return integer_all_onesp (arg);
700
701 case MULT_EXPR:
702 case BIT_AND_EXPR:
703 return integer_zerop (arg);
704
705 default:
706 return false;
707 }
708 }
709
710 /* The function value_replacement does the main work of doing the value
711 replacement. Return non-zero if the replacement is done. Otherwise return
712 0. If we remove the middle basic block, return 2.
713 BB is the basic block where the replacement is going to be done on. ARG0
714 is argument 0 from the PHI. Likewise for ARG1. */
715
716 static int
717 value_replacement (basic_block cond_bb, basic_block middle_bb,
718 edge e0, edge e1, gimple phi,
719 tree arg0, tree arg1)
720 {
721 gimple_stmt_iterator gsi;
722 gimple cond;
723 edge true_edge, false_edge;
724 enum tree_code code;
725 bool emtpy_or_with_defined_p = true;
726
727 /* If the type says honor signed zeros we cannot do this
728 optimization. */
729 if (HONOR_SIGNED_ZEROS (arg1))
730 return 0;
731
732 /* If there is a statement in MIDDLE_BB that defines one of the PHI
733 arguments, then adjust arg0 or arg1. */
734 gsi = gsi_start_nondebug_after_labels_bb (middle_bb);
735 while (!gsi_end_p (gsi))
736 {
737 gimple stmt = gsi_stmt (gsi);
738 tree lhs;
739 gsi_next_nondebug (&gsi);
740 if (!is_gimple_assign (stmt))
741 {
742 emtpy_or_with_defined_p = false;
743 continue;
744 }
745 /* Now try to adjust arg0 or arg1 according to the computation
746 in the statement. */
747 lhs = gimple_assign_lhs (stmt);
748 if (!(lhs == arg0
749 && jump_function_from_stmt (&arg0, stmt))
750 || (lhs == arg1
751 && jump_function_from_stmt (&arg1, stmt)))
752 emtpy_or_with_defined_p = false;
753 }
754
755 cond = last_stmt (cond_bb);
756 code = gimple_cond_code (cond);
757
758 /* This transformation is only valid for equality comparisons. */
759 if (code != NE_EXPR && code != EQ_EXPR)
760 return 0;
761
762 /* We need to know which is the true edge and which is the false
763 edge so that we know if have abs or negative abs. */
764 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
765
766 /* At this point we know we have a COND_EXPR with two successors.
767 One successor is BB, the other successor is an empty block which
768 falls through into BB.
769
770 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
771
772 There is a single PHI node at the join point (BB) with two arguments.
773
774 We now need to verify that the two arguments in the PHI node match
775 the two arguments to the equality comparison. */
776
777 if (operand_equal_for_value_replacement (arg0, arg1, &code, cond))
778 {
779 edge e;
780 tree arg;
781
782 /* For NE_EXPR, we want to build an assignment result = arg where
783 arg is the PHI argument associated with the true edge. For
784 EQ_EXPR we want the PHI argument associated with the false edge. */
785 e = (code == NE_EXPR ? true_edge : false_edge);
786
787 /* Unfortunately, E may not reach BB (it may instead have gone to
788 OTHER_BLOCK). If that is the case, then we want the single outgoing
789 edge from OTHER_BLOCK which reaches BB and represents the desired
790 path from COND_BLOCK. */
791 if (e->dest == middle_bb)
792 e = single_succ_edge (e->dest);
793
794 /* Now we know the incoming edge to BB that has the argument for the
795 RHS of our new assignment statement. */
796 if (e0 == e)
797 arg = arg0;
798 else
799 arg = arg1;
800
801 /* If the middle basic block was empty or is defining the
802 PHI arguments and this is a single phi where the args are different
803 for the edges e0 and e1 then we can remove the middle basic block. */
804 if (emtpy_or_with_defined_p
805 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)),
806 e0, e1) == phi)
807 {
808 replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
809 /* Note that we optimized this PHI. */
810 return 2;
811 }
812 else
813 {
814 /* Replace the PHI arguments with arg. */
815 SET_PHI_ARG_DEF (phi, e0->dest_idx, arg);
816 SET_PHI_ARG_DEF (phi, e1->dest_idx, arg);
817 if (dump_file && (dump_flags & TDF_DETAILS))
818 {
819 fprintf (dump_file, "PHI ");
820 print_generic_expr (dump_file, gimple_phi_result (phi), 0);
821 fprintf (dump_file, " reduced for COND_EXPR in block %d to ",
822 cond_bb->index);
823 print_generic_expr (dump_file, arg, 0);
824 fprintf (dump_file, ".\n");
825 }
826 return 1;
827 }
828
829 }
830
831 /* Now optimize (x != 0) ? x + y : y to just y.
832 The following condition is too restrictive, there can easily be another
833 stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */
834 gimple assign = last_and_only_stmt (middle_bb);
835 if (!assign || gimple_code (assign) != GIMPLE_ASSIGN
836 || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS
837 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0))
838 && !POINTER_TYPE_P (TREE_TYPE (arg0))))
839 return 0;
840
841 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
842 if (!gimple_seq_empty_p (phi_nodes (middle_bb)))
843 return 0;
844
845 /* Only transform if it removes the condition. */
846 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1))
847 return 0;
848
849 /* Size-wise, this is always profitable. */
850 if (optimize_bb_for_speed_p (cond_bb)
851 /* The special case is useless if it has a low probability. */
852 && profile_status_for_fn (cfun) != PROFILE_ABSENT
853 && EDGE_PRED (middle_bb, 0)->probability < PROB_EVEN
854 /* If assign is cheap, there is no point avoiding it. */
855 && estimate_num_insns (assign, &eni_time_weights)
856 >= 3 * estimate_num_insns (cond, &eni_time_weights))
857 return 0;
858
859 tree lhs = gimple_assign_lhs (assign);
860 tree rhs1 = gimple_assign_rhs1 (assign);
861 tree rhs2 = gimple_assign_rhs2 (assign);
862 enum tree_code code_def = gimple_assign_rhs_code (assign);
863 tree cond_lhs = gimple_cond_lhs (cond);
864 tree cond_rhs = gimple_cond_rhs (cond);
865
866 if (((code == NE_EXPR && e1 == false_edge)
867 || (code == EQ_EXPR && e1 == true_edge))
868 && arg0 == lhs
869 && ((arg1 == rhs1
870 && operand_equal_for_phi_arg_p (rhs2, cond_lhs)
871 && neutral_element_p (code_def, cond_rhs, true))
872 || (arg1 == rhs2
873 && operand_equal_for_phi_arg_p (rhs1, cond_lhs)
874 && neutral_element_p (code_def, cond_rhs, false))
875 || (operand_equal_for_phi_arg_p (arg1, cond_rhs)
876 && (operand_equal_for_phi_arg_p (rhs2, cond_lhs)
877 || operand_equal_for_phi_arg_p (rhs1, cond_lhs))
878 && absorbing_element_p (code_def, cond_rhs))))
879 {
880 gsi = gsi_for_stmt (cond);
881 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)))
882 {
883 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
884 def-stmt in:
885 if (n_5 != 0)
886 goto <bb 3>;
887 else
888 goto <bb 4>;
889
890 <bb 3>:
891 # RANGE [0, 4294967294]
892 u_6 = n_5 + 4294967295;
893
894 <bb 4>:
895 # u_3 = PHI <u_6(3), 4294967295(2)> */
896 SSA_NAME_RANGE_INFO (lhs) = NULL;
897 SSA_NAME_ANTI_RANGE_P (lhs) = 0;
898 /* If available, we can use VR of phi result at least. */
899 tree phires = gimple_phi_result (phi);
900 struct range_info_def *phires_range_info
901 = SSA_NAME_RANGE_INFO (phires);
902 if (phires_range_info)
903 duplicate_ssa_name_range_info (lhs, SSA_NAME_RANGE_TYPE (phires),
904 phires_range_info);
905 }
906 gimple_stmt_iterator gsi_from = gsi_for_stmt (assign);
907 gsi_move_before (&gsi_from, &gsi);
908 replace_phi_edge_with_variable (cond_bb, e1, phi, lhs);
909 return 2;
910 }
911
912 return 0;
913 }
914
915 /* The function minmax_replacement does the main work of doing the minmax
916 replacement. Return true if the replacement is done. Otherwise return
917 false.
918 BB is the basic block where the replacement is going to be done on. ARG0
919 is argument 0 from the PHI. Likewise for ARG1. */
920
921 static bool
922 minmax_replacement (basic_block cond_bb, basic_block middle_bb,
923 edge e0, edge e1, gimple phi,
924 tree arg0, tree arg1)
925 {
926 tree result, type;
927 gcond *cond;
928 gassign *new_stmt;
929 edge true_edge, false_edge;
930 enum tree_code cmp, minmax, ass_code;
931 tree smaller, larger, arg_true, arg_false;
932 gimple_stmt_iterator gsi, gsi_from;
933
934 type = TREE_TYPE (PHI_RESULT (phi));
935
936 /* The optimization may be unsafe due to NaNs. */
937 if (HONOR_NANS (type))
938 return false;
939
940 cond = as_a <gcond *> (last_stmt (cond_bb));
941 cmp = gimple_cond_code (cond);
942
943 /* This transformation is only valid for order comparisons. Record which
944 operand is smaller/larger if the result of the comparison is true. */
945 if (cmp == LT_EXPR || cmp == LE_EXPR)
946 {
947 smaller = gimple_cond_lhs (cond);
948 larger = gimple_cond_rhs (cond);
949 }
950 else if (cmp == GT_EXPR || cmp == GE_EXPR)
951 {
952 smaller = gimple_cond_rhs (cond);
953 larger = gimple_cond_lhs (cond);
954 }
955 else
956 return false;
957
958 /* We need to know which is the true edge and which is the false
959 edge so that we know if have abs or negative abs. */
960 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
961
962 /* Forward the edges over the middle basic block. */
963 if (true_edge->dest == middle_bb)
964 true_edge = EDGE_SUCC (true_edge->dest, 0);
965 if (false_edge->dest == middle_bb)
966 false_edge = EDGE_SUCC (false_edge->dest, 0);
967
968 if (true_edge == e0)
969 {
970 gcc_assert (false_edge == e1);
971 arg_true = arg0;
972 arg_false = arg1;
973 }
974 else
975 {
976 gcc_assert (false_edge == e0);
977 gcc_assert (true_edge == e1);
978 arg_true = arg1;
979 arg_false = arg0;
980 }
981
982 if (empty_block_p (middle_bb))
983 {
984 if (operand_equal_for_phi_arg_p (arg_true, smaller)
985 && operand_equal_for_phi_arg_p (arg_false, larger))
986 {
987 /* Case
988
989 if (smaller < larger)
990 rslt = smaller;
991 else
992 rslt = larger; */
993 minmax = MIN_EXPR;
994 }
995 else if (operand_equal_for_phi_arg_p (arg_false, smaller)
996 && operand_equal_for_phi_arg_p (arg_true, larger))
997 minmax = MAX_EXPR;
998 else
999 return false;
1000 }
1001 else
1002 {
1003 /* Recognize the following case, assuming d <= u:
1004
1005 if (a <= u)
1006 b = MAX (a, d);
1007 x = PHI <b, u>
1008
1009 This is equivalent to
1010
1011 b = MAX (a, d);
1012 x = MIN (b, u); */
1013
1014 gimple assign = last_and_only_stmt (middle_bb);
1015 tree lhs, op0, op1, bound;
1016
1017 if (!assign
1018 || gimple_code (assign) != GIMPLE_ASSIGN)
1019 return false;
1020
1021 lhs = gimple_assign_lhs (assign);
1022 ass_code = gimple_assign_rhs_code (assign);
1023 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
1024 return false;
1025 op0 = gimple_assign_rhs1 (assign);
1026 op1 = gimple_assign_rhs2 (assign);
1027
1028 if (true_edge->src == middle_bb)
1029 {
1030 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1031 if (!operand_equal_for_phi_arg_p (lhs, arg_true))
1032 return false;
1033
1034 if (operand_equal_for_phi_arg_p (arg_false, larger))
1035 {
1036 /* Case
1037
1038 if (smaller < larger)
1039 {
1040 r' = MAX_EXPR (smaller, bound)
1041 }
1042 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1043 if (ass_code != MAX_EXPR)
1044 return false;
1045
1046 minmax = MIN_EXPR;
1047 if (operand_equal_for_phi_arg_p (op0, smaller))
1048 bound = op1;
1049 else if (operand_equal_for_phi_arg_p (op1, smaller))
1050 bound = op0;
1051 else
1052 return false;
1053
1054 /* We need BOUND <= LARGER. */
1055 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1056 bound, larger)))
1057 return false;
1058 }
1059 else if (operand_equal_for_phi_arg_p (arg_false, smaller))
1060 {
1061 /* Case
1062
1063 if (smaller < larger)
1064 {
1065 r' = MIN_EXPR (larger, bound)
1066 }
1067 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1068 if (ass_code != MIN_EXPR)
1069 return false;
1070
1071 minmax = MAX_EXPR;
1072 if (operand_equal_for_phi_arg_p (op0, larger))
1073 bound = op1;
1074 else if (operand_equal_for_phi_arg_p (op1, larger))
1075 bound = op0;
1076 else
1077 return false;
1078
1079 /* We need BOUND >= SMALLER. */
1080 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1081 bound, smaller)))
1082 return false;
1083 }
1084 else
1085 return false;
1086 }
1087 else
1088 {
1089 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1090 if (!operand_equal_for_phi_arg_p (lhs, arg_false))
1091 return false;
1092
1093 if (operand_equal_for_phi_arg_p (arg_true, larger))
1094 {
1095 /* Case
1096
1097 if (smaller > larger)
1098 {
1099 r' = MIN_EXPR (smaller, bound)
1100 }
1101 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1102 if (ass_code != MIN_EXPR)
1103 return false;
1104
1105 minmax = MAX_EXPR;
1106 if (operand_equal_for_phi_arg_p (op0, smaller))
1107 bound = op1;
1108 else if (operand_equal_for_phi_arg_p (op1, smaller))
1109 bound = op0;
1110 else
1111 return false;
1112
1113 /* We need BOUND >= LARGER. */
1114 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1115 bound, larger)))
1116 return false;
1117 }
1118 else if (operand_equal_for_phi_arg_p (arg_true, smaller))
1119 {
1120 /* Case
1121
1122 if (smaller > larger)
1123 {
1124 r' = MAX_EXPR (larger, bound)
1125 }
1126 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1127 if (ass_code != MAX_EXPR)
1128 return false;
1129
1130 minmax = MIN_EXPR;
1131 if (operand_equal_for_phi_arg_p (op0, larger))
1132 bound = op1;
1133 else if (operand_equal_for_phi_arg_p (op1, larger))
1134 bound = op0;
1135 else
1136 return false;
1137
1138 /* We need BOUND <= SMALLER. */
1139 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1140 bound, smaller)))
1141 return false;
1142 }
1143 else
1144 return false;
1145 }
1146
1147 /* Move the statement from the middle block. */
1148 gsi = gsi_last_bb (cond_bb);
1149 gsi_from = gsi_last_nondebug_bb (middle_bb);
1150 gsi_move_before (&gsi_from, &gsi);
1151 }
1152
1153 /* Emit the statement to compute min/max. */
1154 result = duplicate_ssa_name (PHI_RESULT (phi), NULL);
1155 new_stmt = gimple_build_assign (result, minmax, arg0, arg1);
1156 gsi = gsi_last_bb (cond_bb);
1157 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1158
1159 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1160 return true;
1161 }
1162
1163 /* The function absolute_replacement does the main work of doing the absolute
1164 replacement. Return true if the replacement is done. Otherwise return
1165 false.
1166 bb is the basic block where the replacement is going to be done on. arg0
1167 is argument 0 from the phi. Likewise for arg1. */
1168
1169 static bool
1170 abs_replacement (basic_block cond_bb, basic_block middle_bb,
1171 edge e0 ATTRIBUTE_UNUSED, edge e1,
1172 gimple phi, tree arg0, tree arg1)
1173 {
1174 tree result;
1175 gassign *new_stmt;
1176 gimple cond;
1177 gimple_stmt_iterator gsi;
1178 edge true_edge, false_edge;
1179 gimple assign;
1180 edge e;
1181 tree rhs, lhs;
1182 bool negate;
1183 enum tree_code cond_code;
1184
1185 /* If the type says honor signed zeros we cannot do this
1186 optimization. */
1187 if (HONOR_SIGNED_ZEROS (arg1))
1188 return false;
1189
1190 /* OTHER_BLOCK must have only one executable statement which must have the
1191 form arg0 = -arg1 or arg1 = -arg0. */
1192
1193 assign = last_and_only_stmt (middle_bb);
1194 /* If we did not find the proper negation assignment, then we can not
1195 optimize. */
1196 if (assign == NULL)
1197 return false;
1198
1199 /* If we got here, then we have found the only executable statement
1200 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1201 arg1 = -arg0, then we can not optimize. */
1202 if (gimple_code (assign) != GIMPLE_ASSIGN)
1203 return false;
1204
1205 lhs = gimple_assign_lhs (assign);
1206
1207 if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
1208 return false;
1209
1210 rhs = gimple_assign_rhs1 (assign);
1211
1212 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1213 if (!(lhs == arg0 && rhs == arg1)
1214 && !(lhs == arg1 && rhs == arg0))
1215 return false;
1216
1217 cond = last_stmt (cond_bb);
1218 result = PHI_RESULT (phi);
1219
1220 /* Only relationals comparing arg[01] against zero are interesting. */
1221 cond_code = gimple_cond_code (cond);
1222 if (cond_code != GT_EXPR && cond_code != GE_EXPR
1223 && cond_code != LT_EXPR && cond_code != LE_EXPR)
1224 return false;
1225
1226 /* Make sure the conditional is arg[01] OP y. */
1227 if (gimple_cond_lhs (cond) != rhs)
1228 return false;
1229
1230 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond)))
1231 ? real_zerop (gimple_cond_rhs (cond))
1232 : integer_zerop (gimple_cond_rhs (cond)))
1233 ;
1234 else
1235 return false;
1236
1237 /* We need to know which is the true edge and which is the false
1238 edge so that we know if have abs or negative abs. */
1239 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1240
1241 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1242 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1243 the false edge goes to OTHER_BLOCK. */
1244 if (cond_code == GT_EXPR || cond_code == GE_EXPR)
1245 e = true_edge;
1246 else
1247 e = false_edge;
1248
1249 if (e->dest == middle_bb)
1250 negate = true;
1251 else
1252 negate = false;
1253
1254 result = duplicate_ssa_name (result, NULL);
1255
1256 if (negate)
1257 lhs = make_ssa_name (TREE_TYPE (result));
1258 else
1259 lhs = result;
1260
1261 /* Build the modify expression with abs expression. */
1262 new_stmt = gimple_build_assign (lhs, ABS_EXPR, rhs);
1263
1264 gsi = gsi_last_bb (cond_bb);
1265 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1266
1267 if (negate)
1268 {
1269 /* Get the right GSI. We want to insert after the recently
1270 added ABS_EXPR statement (which we know is the first statement
1271 in the block. */
1272 new_stmt = gimple_build_assign (result, NEGATE_EXPR, lhs);
1273
1274 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1275 }
1276
1277 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1278
1279 /* Note that we optimized this PHI. */
1280 return true;
1281 }
1282
1283 /* Auxiliary functions to determine the set of memory accesses which
1284 can't trap because they are preceded by accesses to the same memory
1285 portion. We do that for MEM_REFs, so we only need to track
1286 the SSA_NAME of the pointer indirectly referenced. The algorithm
1287 simply is a walk over all instructions in dominator order. When
1288 we see an MEM_REF we determine if we've already seen a same
1289 ref anywhere up to the root of the dominator tree. If we do the
1290 current access can't trap. If we don't see any dominating access
1291 the current access might trap, but might also make later accesses
1292 non-trapping, so we remember it. We need to be careful with loads
1293 or stores, for instance a load might not trap, while a store would,
1294 so if we see a dominating read access this doesn't mean that a later
1295 write access would not trap. Hence we also need to differentiate the
1296 type of access(es) seen.
1297
1298 ??? We currently are very conservative and assume that a load might
1299 trap even if a store doesn't (write-only memory). This probably is
1300 overly conservative. */
1301
1302 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1303 through it was seen, which would constitute a no-trap region for
1304 same accesses. */
1305 struct name_to_bb
1306 {
1307 unsigned int ssa_name_ver;
1308 unsigned int phase;
1309 bool store;
1310 HOST_WIDE_INT offset, size;
1311 basic_block bb;
1312 };
1313
1314 /* Hashtable helpers. */
1315
1316 struct ssa_names_hasher : typed_free_remove <name_to_bb>
1317 {
1318 typedef name_to_bb *value_type;
1319 typedef name_to_bb *compare_type;
1320 static inline hashval_t hash (const name_to_bb *);
1321 static inline bool equal (const name_to_bb *, const name_to_bb *);
1322 };
1323
1324 /* Used for quick clearing of the hash-table when we see calls.
1325 Hash entries with phase < nt_call_phase are invalid. */
1326 static unsigned int nt_call_phase;
1327
1328 /* The hash function. */
1329
1330 inline hashval_t
1331 ssa_names_hasher::hash (const name_to_bb *n)
1332 {
1333 return n->ssa_name_ver ^ (((hashval_t) n->store) << 31)
1334 ^ (n->offset << 6) ^ (n->size << 3);
1335 }
1336
1337 /* The equality function of *P1 and *P2. */
1338
1339 inline bool
1340 ssa_names_hasher::equal (const name_to_bb *n1, const name_to_bb *n2)
1341 {
1342 return n1->ssa_name_ver == n2->ssa_name_ver
1343 && n1->store == n2->store
1344 && n1->offset == n2->offset
1345 && n1->size == n2->size;
1346 }
1347
1348 class nontrapping_dom_walker : public dom_walker
1349 {
1350 public:
1351 nontrapping_dom_walker (cdi_direction direction, hash_set<tree> *ps)
1352 : dom_walker (direction), m_nontrapping (ps), m_seen_ssa_names (128) {}
1353
1354 virtual void before_dom_children (basic_block);
1355 virtual void after_dom_children (basic_block);
1356
1357 private:
1358
1359 /* We see the expression EXP in basic block BB. If it's an interesting
1360 expression (an MEM_REF through an SSA_NAME) possibly insert the
1361 expression into the set NONTRAP or the hash table of seen expressions.
1362 STORE is true if this expression is on the LHS, otherwise it's on
1363 the RHS. */
1364 void add_or_mark_expr (basic_block, tree, bool);
1365
1366 hash_set<tree> *m_nontrapping;
1367
1368 /* The hash table for remembering what we've seen. */
1369 hash_table<ssa_names_hasher> m_seen_ssa_names;
1370 };
1371
1372 /* Called by walk_dominator_tree, when entering the block BB. */
1373 void
1374 nontrapping_dom_walker::before_dom_children (basic_block bb)
1375 {
1376 edge e;
1377 edge_iterator ei;
1378 gimple_stmt_iterator gsi;
1379
1380 /* If we haven't seen all our predecessors, clear the hash-table. */
1381 FOR_EACH_EDGE (e, ei, bb->preds)
1382 if ((((size_t)e->src->aux) & 2) == 0)
1383 {
1384 nt_call_phase++;
1385 break;
1386 }
1387
1388 /* Mark this BB as being on the path to dominator root and as visited. */
1389 bb->aux = (void*)(1 | 2);
1390
1391 /* And walk the statements in order. */
1392 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1393 {
1394 gimple stmt = gsi_stmt (gsi);
1395
1396 if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt))
1397 nt_call_phase++;
1398 else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt))
1399 {
1400 add_or_mark_expr (bb, gimple_assign_lhs (stmt), true);
1401 add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), false);
1402 }
1403 }
1404 }
1405
1406 /* Called by walk_dominator_tree, when basic block BB is exited. */
1407 void
1408 nontrapping_dom_walker::after_dom_children (basic_block bb)
1409 {
1410 /* This BB isn't on the path to dominator root anymore. */
1411 bb->aux = (void*)2;
1412 }
1413
1414 /* We see the expression EXP in basic block BB. If it's an interesting
1415 expression (an MEM_REF through an SSA_NAME) possibly insert the
1416 expression into the set NONTRAP or the hash table of seen expressions.
1417 STORE is true if this expression is on the LHS, otherwise it's on
1418 the RHS. */
1419 void
1420 nontrapping_dom_walker::add_or_mark_expr (basic_block bb, tree exp, bool store)
1421 {
1422 HOST_WIDE_INT size;
1423
1424 if (TREE_CODE (exp) == MEM_REF
1425 && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME
1426 && tree_fits_shwi_p (TREE_OPERAND (exp, 1))
1427 && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0)
1428 {
1429 tree name = TREE_OPERAND (exp, 0);
1430 struct name_to_bb map;
1431 name_to_bb **slot;
1432 struct name_to_bb *n2bb;
1433 basic_block found_bb = 0;
1434
1435 /* Try to find the last seen MEM_REF through the same
1436 SSA_NAME, which can trap. */
1437 map.ssa_name_ver = SSA_NAME_VERSION (name);
1438 map.phase = 0;
1439 map.bb = 0;
1440 map.store = store;
1441 map.offset = tree_to_shwi (TREE_OPERAND (exp, 1));
1442 map.size = size;
1443
1444 slot = m_seen_ssa_names.find_slot (&map, INSERT);
1445 n2bb = *slot;
1446 if (n2bb && n2bb->phase >= nt_call_phase)
1447 found_bb = n2bb->bb;
1448
1449 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1450 (it's in a basic block on the path from us to the dominator root)
1451 then we can't trap. */
1452 if (found_bb && (((size_t)found_bb->aux) & 1) == 1)
1453 {
1454 m_nontrapping->add (exp);
1455 }
1456 else
1457 {
1458 /* EXP might trap, so insert it into the hash table. */
1459 if (n2bb)
1460 {
1461 n2bb->phase = nt_call_phase;
1462 n2bb->bb = bb;
1463 }
1464 else
1465 {
1466 n2bb = XNEW (struct name_to_bb);
1467 n2bb->ssa_name_ver = SSA_NAME_VERSION (name);
1468 n2bb->phase = nt_call_phase;
1469 n2bb->bb = bb;
1470 n2bb->store = store;
1471 n2bb->offset = map.offset;
1472 n2bb->size = size;
1473 *slot = n2bb;
1474 }
1475 }
1476 }
1477 }
1478
1479 /* This is the entry point of gathering non trapping memory accesses.
1480 It will do a dominator walk over the whole function, and it will
1481 make use of the bb->aux pointers. It returns a set of trees
1482 (the MEM_REFs itself) which can't trap. */
1483 static hash_set<tree> *
1484 get_non_trapping (void)
1485 {
1486 nt_call_phase = 0;
1487 hash_set<tree> *nontrap = new hash_set<tree>;
1488 /* We're going to do a dominator walk, so ensure that we have
1489 dominance information. */
1490 calculate_dominance_info (CDI_DOMINATORS);
1491
1492 nontrapping_dom_walker (CDI_DOMINATORS, nontrap)
1493 .walk (cfun->cfg->x_entry_block_ptr);
1494
1495 clear_aux_for_blocks ();
1496 return nontrap;
1497 }
1498
1499 /* Do the main work of conditional store replacement. We already know
1500 that the recognized pattern looks like so:
1501
1502 split:
1503 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1504 MIDDLE_BB:
1505 something
1506 fallthrough (edge E0)
1507 JOIN_BB:
1508 some more
1509
1510 We check that MIDDLE_BB contains only one store, that that store
1511 doesn't trap (not via NOTRAP, but via checking if an access to the same
1512 memory location dominates us) and that the store has a "simple" RHS. */
1513
1514 static bool
1515 cond_store_replacement (basic_block middle_bb, basic_block join_bb,
1516 edge e0, edge e1, hash_set<tree> *nontrap)
1517 {
1518 gimple assign = last_and_only_stmt (middle_bb);
1519 tree lhs, rhs, name, name2;
1520 gphi *newphi;
1521 gassign *new_stmt;
1522 gimple_stmt_iterator gsi;
1523 source_location locus;
1524
1525 /* Check if middle_bb contains of only one store. */
1526 if (!assign
1527 || !gimple_assign_single_p (assign)
1528 || gimple_has_volatile_ops (assign))
1529 return false;
1530
1531 locus = gimple_location (assign);
1532 lhs = gimple_assign_lhs (assign);
1533 rhs = gimple_assign_rhs1 (assign);
1534 if (TREE_CODE (lhs) != MEM_REF
1535 || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME
1536 || !is_gimple_reg_type (TREE_TYPE (lhs)))
1537 return false;
1538
1539 /* Prove that we can move the store down. We could also check
1540 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1541 whose value is not available readily, which we want to avoid. */
1542 if (!nontrap->contains (lhs))
1543 return false;
1544
1545 /* Now we've checked the constraints, so do the transformation:
1546 1) Remove the single store. */
1547 gsi = gsi_for_stmt (assign);
1548 unlink_stmt_vdef (assign);
1549 gsi_remove (&gsi, true);
1550 release_defs (assign);
1551
1552 /* 2) Insert a load from the memory of the store to the temporary
1553 on the edge which did not contain the store. */
1554 lhs = unshare_expr (lhs);
1555 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1556 new_stmt = gimple_build_assign (name, lhs);
1557 gimple_set_location (new_stmt, locus);
1558 gsi_insert_on_edge (e1, new_stmt);
1559
1560 /* 3) Create a PHI node at the join block, with one argument
1561 holding the old RHS, and the other holding the temporary
1562 where we stored the old memory contents. */
1563 name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1564 newphi = create_phi_node (name2, join_bb);
1565 add_phi_arg (newphi, rhs, e0, locus);
1566 add_phi_arg (newphi, name, e1, locus);
1567
1568 lhs = unshare_expr (lhs);
1569 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1570
1571 /* 4) Insert that PHI node. */
1572 gsi = gsi_after_labels (join_bb);
1573 if (gsi_end_p (gsi))
1574 {
1575 gsi = gsi_last_bb (join_bb);
1576 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1577 }
1578 else
1579 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1580
1581 return true;
1582 }
1583
1584 /* Do the main work of conditional store replacement. */
1585
1586 static bool
1587 cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb,
1588 basic_block join_bb, gimple then_assign,
1589 gimple else_assign)
1590 {
1591 tree lhs_base, lhs, then_rhs, else_rhs, name;
1592 source_location then_locus, else_locus;
1593 gimple_stmt_iterator gsi;
1594 gphi *newphi;
1595 gassign *new_stmt;
1596
1597 if (then_assign == NULL
1598 || !gimple_assign_single_p (then_assign)
1599 || gimple_clobber_p (then_assign)
1600 || gimple_has_volatile_ops (then_assign)
1601 || else_assign == NULL
1602 || !gimple_assign_single_p (else_assign)
1603 || gimple_clobber_p (else_assign)
1604 || gimple_has_volatile_ops (else_assign))
1605 return false;
1606
1607 lhs = gimple_assign_lhs (then_assign);
1608 if (!is_gimple_reg_type (TREE_TYPE (lhs))
1609 || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0))
1610 return false;
1611
1612 lhs_base = get_base_address (lhs);
1613 if (lhs_base == NULL_TREE
1614 || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF))
1615 return false;
1616
1617 then_rhs = gimple_assign_rhs1 (then_assign);
1618 else_rhs = gimple_assign_rhs1 (else_assign);
1619 then_locus = gimple_location (then_assign);
1620 else_locus = gimple_location (else_assign);
1621
1622 /* Now we've checked the constraints, so do the transformation:
1623 1) Remove the stores. */
1624 gsi = gsi_for_stmt (then_assign);
1625 unlink_stmt_vdef (then_assign);
1626 gsi_remove (&gsi, true);
1627 release_defs (then_assign);
1628
1629 gsi = gsi_for_stmt (else_assign);
1630 unlink_stmt_vdef (else_assign);
1631 gsi_remove (&gsi, true);
1632 release_defs (else_assign);
1633
1634 /* 2) Create a PHI node at the join block, with one argument
1635 holding the old RHS, and the other holding the temporary
1636 where we stored the old memory contents. */
1637 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1638 newphi = create_phi_node (name, join_bb);
1639 add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus);
1640 add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus);
1641
1642 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1643
1644 /* 3) Insert that PHI node. */
1645 gsi = gsi_after_labels (join_bb);
1646 if (gsi_end_p (gsi))
1647 {
1648 gsi = gsi_last_bb (join_bb);
1649 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1650 }
1651 else
1652 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1653
1654 return true;
1655 }
1656
1657 /* Conditional store replacement. We already know
1658 that the recognized pattern looks like so:
1659
1660 split:
1661 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1662 THEN_BB:
1663 ...
1664 X = Y;
1665 ...
1666 goto JOIN_BB;
1667 ELSE_BB:
1668 ...
1669 X = Z;
1670 ...
1671 fallthrough (edge E0)
1672 JOIN_BB:
1673 some more
1674
1675 We check that it is safe to sink the store to JOIN_BB by verifying that
1676 there are no read-after-write or write-after-write dependencies in
1677 THEN_BB and ELSE_BB. */
1678
1679 static bool
1680 cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
1681 basic_block join_bb)
1682 {
1683 gimple then_assign = last_and_only_stmt (then_bb);
1684 gimple else_assign = last_and_only_stmt (else_bb);
1685 vec<data_reference_p> then_datarefs, else_datarefs;
1686 vec<ddr_p> then_ddrs, else_ddrs;
1687 gimple then_store, else_store;
1688 bool found, ok = false, res;
1689 struct data_dependence_relation *ddr;
1690 data_reference_p then_dr, else_dr;
1691 int i, j;
1692 tree then_lhs, else_lhs;
1693 basic_block blocks[3];
1694
1695 if (MAX_STORES_TO_SINK == 0)
1696 return false;
1697
1698 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1699 if (then_assign && else_assign)
1700 return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1701 then_assign, else_assign);
1702
1703 /* Find data references. */
1704 then_datarefs.create (1);
1705 else_datarefs.create (1);
1706 if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs)
1707 == chrec_dont_know)
1708 || !then_datarefs.length ()
1709 || (find_data_references_in_bb (NULL, else_bb, &else_datarefs)
1710 == chrec_dont_know)
1711 || !else_datarefs.length ())
1712 {
1713 free_data_refs (then_datarefs);
1714 free_data_refs (else_datarefs);
1715 return false;
1716 }
1717
1718 /* Find pairs of stores with equal LHS. */
1719 auto_vec<gimple, 1> then_stores, else_stores;
1720 FOR_EACH_VEC_ELT (then_datarefs, i, then_dr)
1721 {
1722 if (DR_IS_READ (then_dr))
1723 continue;
1724
1725 then_store = DR_STMT (then_dr);
1726 then_lhs = gimple_get_lhs (then_store);
1727 if (then_lhs == NULL_TREE)
1728 continue;
1729 found = false;
1730
1731 FOR_EACH_VEC_ELT (else_datarefs, j, else_dr)
1732 {
1733 if (DR_IS_READ (else_dr))
1734 continue;
1735
1736 else_store = DR_STMT (else_dr);
1737 else_lhs = gimple_get_lhs (else_store);
1738 if (else_lhs == NULL_TREE)
1739 continue;
1740
1741 if (operand_equal_p (then_lhs, else_lhs, 0))
1742 {
1743 found = true;
1744 break;
1745 }
1746 }
1747
1748 if (!found)
1749 continue;
1750
1751 then_stores.safe_push (then_store);
1752 else_stores.safe_push (else_store);
1753 }
1754
1755 /* No pairs of stores found. */
1756 if (!then_stores.length ()
1757 || then_stores.length () > (unsigned) MAX_STORES_TO_SINK)
1758 {
1759 free_data_refs (then_datarefs);
1760 free_data_refs (else_datarefs);
1761 return false;
1762 }
1763
1764 /* Compute and check data dependencies in both basic blocks. */
1765 then_ddrs.create (1);
1766 else_ddrs.create (1);
1767 if (!compute_all_dependences (then_datarefs, &then_ddrs,
1768 vNULL, false)
1769 || !compute_all_dependences (else_datarefs, &else_ddrs,
1770 vNULL, false))
1771 {
1772 free_dependence_relations (then_ddrs);
1773 free_dependence_relations (else_ddrs);
1774 free_data_refs (then_datarefs);
1775 free_data_refs (else_datarefs);
1776 return false;
1777 }
1778 blocks[0] = then_bb;
1779 blocks[1] = else_bb;
1780 blocks[2] = join_bb;
1781 renumber_gimple_stmt_uids_in_blocks (blocks, 3);
1782
1783 /* Check that there are no read-after-write or write-after-write dependencies
1784 in THEN_BB. */
1785 FOR_EACH_VEC_ELT (then_ddrs, i, ddr)
1786 {
1787 struct data_reference *dra = DDR_A (ddr);
1788 struct data_reference *drb = DDR_B (ddr);
1789
1790 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1791 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1792 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1793 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1794 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1795 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1796 {
1797 free_dependence_relations (then_ddrs);
1798 free_dependence_relations (else_ddrs);
1799 free_data_refs (then_datarefs);
1800 free_data_refs (else_datarefs);
1801 return false;
1802 }
1803 }
1804
1805 /* Check that there are no read-after-write or write-after-write dependencies
1806 in ELSE_BB. */
1807 FOR_EACH_VEC_ELT (else_ddrs, i, ddr)
1808 {
1809 struct data_reference *dra = DDR_A (ddr);
1810 struct data_reference *drb = DDR_B (ddr);
1811
1812 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1813 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1814 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1815 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1816 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1817 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1818 {
1819 free_dependence_relations (then_ddrs);
1820 free_dependence_relations (else_ddrs);
1821 free_data_refs (then_datarefs);
1822 free_data_refs (else_datarefs);
1823 return false;
1824 }
1825 }
1826
1827 /* Sink stores with same LHS. */
1828 FOR_EACH_VEC_ELT (then_stores, i, then_store)
1829 {
1830 else_store = else_stores[i];
1831 res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1832 then_store, else_store);
1833 ok = ok || res;
1834 }
1835
1836 free_dependence_relations (then_ddrs);
1837 free_dependence_relations (else_ddrs);
1838 free_data_refs (then_datarefs);
1839 free_data_refs (else_datarefs);
1840
1841 return ok;
1842 }
1843
1844 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1845
1846 static bool
1847 local_mem_dependence (gimple stmt, basic_block bb)
1848 {
1849 tree vuse = gimple_vuse (stmt);
1850 gimple def;
1851
1852 if (!vuse)
1853 return false;
1854
1855 def = SSA_NAME_DEF_STMT (vuse);
1856 return (def && gimple_bb (def) == bb);
1857 }
1858
1859 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1860 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1861 and BB3 rejoins control flow following BB1 and BB2, look for
1862 opportunities to hoist loads as follows. If BB3 contains a PHI of
1863 two loads, one each occurring in BB1 and BB2, and the loads are
1864 provably of adjacent fields in the same structure, then move both
1865 loads into BB0. Of course this can only be done if there are no
1866 dependencies preventing such motion.
1867
1868 One of the hoisted loads will always be speculative, so the
1869 transformation is currently conservative:
1870
1871 - The fields must be strictly adjacent.
1872 - The two fields must occupy a single memory block that is
1873 guaranteed to not cross a page boundary.
1874
1875 The last is difficult to prove, as such memory blocks should be
1876 aligned on the minimum of the stack alignment boundary and the
1877 alignment guaranteed by heap allocation interfaces. Thus we rely
1878 on a parameter for the alignment value.
1879
1880 Provided a good value is used for the last case, the first
1881 restriction could possibly be relaxed. */
1882
1883 static void
1884 hoist_adjacent_loads (basic_block bb0, basic_block bb1,
1885 basic_block bb2, basic_block bb3)
1886 {
1887 int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE);
1888 unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT);
1889 gphi_iterator gsi;
1890
1891 /* Walk the phis in bb3 looking for an opportunity. We are looking
1892 for phis of two SSA names, one each of which is defined in bb1 and
1893 bb2. */
1894 for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi))
1895 {
1896 gphi *phi_stmt = gsi.phi ();
1897 gimple def1, def2;
1898 tree arg1, arg2, ref1, ref2, field1, field2;
1899 tree tree_offset1, tree_offset2, tree_size2, next;
1900 int offset1, offset2, size2;
1901 unsigned align1;
1902 gimple_stmt_iterator gsi2;
1903 basic_block bb_for_def1, bb_for_def2;
1904
1905 if (gimple_phi_num_args (phi_stmt) != 2
1906 || virtual_operand_p (gimple_phi_result (phi_stmt)))
1907 continue;
1908
1909 arg1 = gimple_phi_arg_def (phi_stmt, 0);
1910 arg2 = gimple_phi_arg_def (phi_stmt, 1);
1911
1912 if (TREE_CODE (arg1) != SSA_NAME
1913 || TREE_CODE (arg2) != SSA_NAME
1914 || SSA_NAME_IS_DEFAULT_DEF (arg1)
1915 || SSA_NAME_IS_DEFAULT_DEF (arg2))
1916 continue;
1917
1918 def1 = SSA_NAME_DEF_STMT (arg1);
1919 def2 = SSA_NAME_DEF_STMT (arg2);
1920
1921 if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2)
1922 && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2))
1923 continue;
1924
1925 /* Check the mode of the arguments to be sure a conditional move
1926 can be generated for it. */
1927 if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1)))
1928 == CODE_FOR_nothing)
1929 continue;
1930
1931 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
1932 if (!gimple_assign_single_p (def1)
1933 || !gimple_assign_single_p (def2)
1934 || gimple_has_volatile_ops (def1)
1935 || gimple_has_volatile_ops (def2))
1936 continue;
1937
1938 ref1 = gimple_assign_rhs1 (def1);
1939 ref2 = gimple_assign_rhs1 (def2);
1940
1941 if (TREE_CODE (ref1) != COMPONENT_REF
1942 || TREE_CODE (ref2) != COMPONENT_REF)
1943 continue;
1944
1945 /* The zeroth operand of the two component references must be
1946 identical. It is not sufficient to compare get_base_address of
1947 the two references, because this could allow for different
1948 elements of the same array in the two trees. It is not safe to
1949 assume that the existence of one array element implies the
1950 existence of a different one. */
1951 if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0))
1952 continue;
1953
1954 field1 = TREE_OPERAND (ref1, 1);
1955 field2 = TREE_OPERAND (ref2, 1);
1956
1957 /* Check for field adjacency, and ensure field1 comes first. */
1958 for (next = DECL_CHAIN (field1);
1959 next && TREE_CODE (next) != FIELD_DECL;
1960 next = DECL_CHAIN (next))
1961 ;
1962
1963 if (next != field2)
1964 {
1965 for (next = DECL_CHAIN (field2);
1966 next && TREE_CODE (next) != FIELD_DECL;
1967 next = DECL_CHAIN (next))
1968 ;
1969
1970 if (next != field1)
1971 continue;
1972
1973 std::swap (field1, field2);
1974 std::swap (def1, def2);
1975 }
1976
1977 bb_for_def1 = gimple_bb (def1);
1978 bb_for_def2 = gimple_bb (def2);
1979
1980 /* Check for proper alignment of the first field. */
1981 tree_offset1 = bit_position (field1);
1982 tree_offset2 = bit_position (field2);
1983 tree_size2 = DECL_SIZE (field2);
1984
1985 if (!tree_fits_uhwi_p (tree_offset1)
1986 || !tree_fits_uhwi_p (tree_offset2)
1987 || !tree_fits_uhwi_p (tree_size2))
1988 continue;
1989
1990 offset1 = tree_to_uhwi (tree_offset1);
1991 offset2 = tree_to_uhwi (tree_offset2);
1992 size2 = tree_to_uhwi (tree_size2);
1993 align1 = DECL_ALIGN (field1) % param_align_bits;
1994
1995 if (offset1 % BITS_PER_UNIT != 0)
1996 continue;
1997
1998 /* For profitability, the two field references should fit within
1999 a single cache line. */
2000 if (align1 + offset2 - offset1 + size2 > param_align_bits)
2001 continue;
2002
2003 /* The two expressions cannot be dependent upon vdefs defined
2004 in bb1/bb2. */
2005 if (local_mem_dependence (def1, bb_for_def1)
2006 || local_mem_dependence (def2, bb_for_def2))
2007 continue;
2008
2009 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2010 bb0. We hoist the first one first so that a cache miss is handled
2011 efficiently regardless of hardware cache-fill policy. */
2012 gsi2 = gsi_for_stmt (def1);
2013 gsi_move_to_bb_end (&gsi2, bb0);
2014 gsi2 = gsi_for_stmt (def2);
2015 gsi_move_to_bb_end (&gsi2, bb0);
2016
2017 if (dump_file && (dump_flags & TDF_DETAILS))
2018 {
2019 fprintf (dump_file,
2020 "\nHoisting adjacent loads from %d and %d into %d: \n",
2021 bb_for_def1->index, bb_for_def2->index, bb0->index);
2022 print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS);
2023 print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS);
2024 }
2025 }
2026 }
2027
2028 /* Determine whether we should attempt to hoist adjacent loads out of
2029 diamond patterns in pass_phiopt. Always hoist loads if
2030 -fhoist-adjacent-loads is specified and the target machine has
2031 both a conditional move instruction and a defined cache line size. */
2032
2033 static bool
2034 gate_hoist_loads (void)
2035 {
2036 return (flag_hoist_adjacent_loads == 1
2037 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE)
2038 && HAVE_conditional_move);
2039 }
2040
2041 /* This pass tries to replaces an if-then-else block with an
2042 assignment. We have four kinds of transformations. Some of these
2043 transformations are also performed by the ifcvt RTL optimizer.
2044
2045 Conditional Replacement
2046 -----------------------
2047
2048 This transformation, implemented in conditional_replacement,
2049 replaces
2050
2051 bb0:
2052 if (cond) goto bb2; else goto bb1;
2053 bb1:
2054 bb2:
2055 x = PHI <0 (bb1), 1 (bb0), ...>;
2056
2057 with
2058
2059 bb0:
2060 x' = cond;
2061 goto bb2;
2062 bb2:
2063 x = PHI <x' (bb0), ...>;
2064
2065 We remove bb1 as it becomes unreachable. This occurs often due to
2066 gimplification of conditionals.
2067
2068 Value Replacement
2069 -----------------
2070
2071 This transformation, implemented in value_replacement, replaces
2072
2073 bb0:
2074 if (a != b) goto bb2; else goto bb1;
2075 bb1:
2076 bb2:
2077 x = PHI <a (bb1), b (bb0), ...>;
2078
2079 with
2080
2081 bb0:
2082 bb2:
2083 x = PHI <b (bb0), ...>;
2084
2085 This opportunity can sometimes occur as a result of other
2086 optimizations.
2087
2088
2089 Another case caught by value replacement looks like this:
2090
2091 bb0:
2092 t1 = a == CONST;
2093 t2 = b > c;
2094 t3 = t1 & t2;
2095 if (t3 != 0) goto bb1; else goto bb2;
2096 bb1:
2097 bb2:
2098 x = PHI (CONST, a)
2099
2100 Gets replaced with:
2101 bb0:
2102 bb2:
2103 t1 = a == CONST;
2104 t2 = b > c;
2105 t3 = t1 & t2;
2106 x = a;
2107
2108 ABS Replacement
2109 ---------------
2110
2111 This transformation, implemented in abs_replacement, replaces
2112
2113 bb0:
2114 if (a >= 0) goto bb2; else goto bb1;
2115 bb1:
2116 x = -a;
2117 bb2:
2118 x = PHI <x (bb1), a (bb0), ...>;
2119
2120 with
2121
2122 bb0:
2123 x' = ABS_EXPR< a >;
2124 bb2:
2125 x = PHI <x' (bb0), ...>;
2126
2127 MIN/MAX Replacement
2128 -------------------
2129
2130 This transformation, minmax_replacement replaces
2131
2132 bb0:
2133 if (a <= b) goto bb2; else goto bb1;
2134 bb1:
2135 bb2:
2136 x = PHI <b (bb1), a (bb0), ...>;
2137
2138 with
2139
2140 bb0:
2141 x' = MIN_EXPR (a, b)
2142 bb2:
2143 x = PHI <x' (bb0), ...>;
2144
2145 A similar transformation is done for MAX_EXPR.
2146
2147
2148 This pass also performs a fifth transformation of a slightly different
2149 flavor.
2150
2151 Adjacent Load Hoisting
2152 ----------------------
2153
2154 This transformation replaces
2155
2156 bb0:
2157 if (...) goto bb2; else goto bb1;
2158 bb1:
2159 x1 = (<expr>).field1;
2160 goto bb3;
2161 bb2:
2162 x2 = (<expr>).field2;
2163 bb3:
2164 # x = PHI <x1, x2>;
2165
2166 with
2167
2168 bb0:
2169 x1 = (<expr>).field1;
2170 x2 = (<expr>).field2;
2171 if (...) goto bb2; else goto bb1;
2172 bb1:
2173 goto bb3;
2174 bb2:
2175 bb3:
2176 # x = PHI <x1, x2>;
2177
2178 The purpose of this transformation is to enable generation of conditional
2179 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2180 the loads is speculative, the transformation is restricted to very
2181 specific cases to avoid introducing a page fault. We are looking for
2182 the common idiom:
2183
2184 if (...)
2185 x = y->left;
2186 else
2187 x = y->right;
2188
2189 where left and right are typically adjacent pointers in a tree structure. */
2190
2191 namespace {
2192
2193 const pass_data pass_data_phiopt =
2194 {
2195 GIMPLE_PASS, /* type */
2196 "phiopt", /* name */
2197 OPTGROUP_NONE, /* optinfo_flags */
2198 TV_TREE_PHIOPT, /* tv_id */
2199 ( PROP_cfg | PROP_ssa ), /* properties_required */
2200 0, /* properties_provided */
2201 0, /* properties_destroyed */
2202 0, /* todo_flags_start */
2203 0, /* todo_flags_finish */
2204 };
2205
2206 class pass_phiopt : public gimple_opt_pass
2207 {
2208 public:
2209 pass_phiopt (gcc::context *ctxt)
2210 : gimple_opt_pass (pass_data_phiopt, ctxt)
2211 {}
2212
2213 /* opt_pass methods: */
2214 opt_pass * clone () { return new pass_phiopt (m_ctxt); }
2215 virtual bool gate (function *) { return flag_ssa_phiopt; }
2216 virtual unsigned int execute (function *)
2217 {
2218 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2219 }
2220
2221 }; // class pass_phiopt
2222
2223 } // anon namespace
2224
2225 gimple_opt_pass *
2226 make_pass_phiopt (gcc::context *ctxt)
2227 {
2228 return new pass_phiopt (ctxt);
2229 }
2230
2231 namespace {
2232
2233 const pass_data pass_data_cselim =
2234 {
2235 GIMPLE_PASS, /* type */
2236 "cselim", /* name */
2237 OPTGROUP_NONE, /* optinfo_flags */
2238 TV_TREE_PHIOPT, /* tv_id */
2239 ( PROP_cfg | PROP_ssa ), /* properties_required */
2240 0, /* properties_provided */
2241 0, /* properties_destroyed */
2242 0, /* todo_flags_start */
2243 0, /* todo_flags_finish */
2244 };
2245
2246 class pass_cselim : public gimple_opt_pass
2247 {
2248 public:
2249 pass_cselim (gcc::context *ctxt)
2250 : gimple_opt_pass (pass_data_cselim, ctxt)
2251 {}
2252
2253 /* opt_pass methods: */
2254 virtual bool gate (function *) { return flag_tree_cselim; }
2255 virtual unsigned int execute (function *) { return tree_ssa_cs_elim (); }
2256
2257 }; // class pass_cselim
2258
2259 } // anon namespace
2260
2261 gimple_opt_pass *
2262 make_pass_cselim (gcc::context *ctxt)
2263 {
2264 return new pass_cselim (ctxt);
2265 }