RISC-V: Optimize SELECT_VL codegen when length is known as smaller than VF
[official-gcc.git] / gcc / tree-ssa-phiopt.cc
bloba3b660bb18ac509fed09bbae252c5ec06dddafb0
1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2023 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
9 later version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "insn-codes.h"
25 #include "rtl.h"
26 #include "tree.h"
27 #include "gimple.h"
28 #include "cfghooks.h"
29 #include "tree-pass.h"
30 #include "ssa.h"
31 #include "tree-ssa.h"
32 #include "optabs-tree.h"
33 #include "insn-config.h"
34 #include "gimple-pretty-print.h"
35 #include "fold-const.h"
36 #include "stor-layout.h"
37 #include "cfganal.h"
38 #include "gimplify.h"
39 #include "gimple-iterator.h"
40 #include "gimplify-me.h"
41 #include "tree-cfg.h"
42 #include "tree-dfa.h"
43 #include "domwalk.h"
44 #include "cfgloop.h"
45 #include "tree-data-ref.h"
46 #include "tree-scalar-evolution.h"
47 #include "tree-inline.h"
48 #include "case-cfn-macros.h"
49 #include "tree-eh.h"
50 #include "gimple-fold.h"
51 #include "internal-fn.h"
52 #include "gimple-range.h"
53 #include "gimple-match.h"
54 #include "dbgcnt.h"
55 #include "tree-ssa-propagate.h"
56 #include "tree-ssa-dce.h"
58 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
60 static gphi *
61 single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1)
63 gimple_stmt_iterator i;
64 gphi *phi = NULL;
65 if (gimple_seq_singleton_p (seq))
67 phi = as_a <gphi *> (gsi_stmt (gsi_start (seq)));
68 /* Never return virtual phis. */
69 if (virtual_operand_p (gimple_phi_result (phi)))
70 return NULL;
71 return phi;
73 for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i))
75 gphi *p = as_a <gphi *> (gsi_stmt (i));
76 /* If the PHI arguments are equal then we can skip this PHI. */
77 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx),
78 gimple_phi_arg_def (p, e1->dest_idx)))
79 continue;
81 /* Punt on virtual phis with different arguments from the edges. */
82 if (virtual_operand_p (gimple_phi_result (p)))
83 return NULL;
85 /* If we already have a PHI that has the two edge arguments are
86 different, then return it is not a singleton for these PHIs. */
87 if (phi)
88 return NULL;
90 phi = p;
92 return phi;
95 /* Replace PHI node element whose edge is E in block BB with variable NEW.
96 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
97 is known to have two edges, one of which must reach BB). */
99 static void
100 replace_phi_edge_with_variable (basic_block cond_block,
101 edge e, gphi *phi, tree new_tree,
102 bitmap dce_ssa_names = nullptr)
104 basic_block bb = gimple_bb (phi);
105 gimple_stmt_iterator gsi;
106 tree phi_result = PHI_RESULT (phi);
107 bool deleteboth = false;
109 /* Duplicate range info if they are the only things setting the target PHI.
110 This is needed as later on, the new_tree will be replacing
111 The assignement of the PHI.
112 For an example:
113 bb1:
114 _4 = min<a_1, 255>
115 goto bb2
117 # RANGE [-INF, 255]
118 a_3 = PHI<_4(1)>
119 bb3:
121 use(a_3)
122 And _4 gets propagated into the use of a_3 and losing the range info.
123 This can't be done for more than 2 incoming edges as the propagation
124 won't happen.
125 The new_tree needs to be defined in the same basic block as the conditional. */
126 if (TREE_CODE (new_tree) == SSA_NAME
127 && EDGE_COUNT (gimple_bb (phi)->preds) == 2
128 && INTEGRAL_TYPE_P (TREE_TYPE (phi_result))
129 && !SSA_NAME_RANGE_INFO (new_tree)
130 && SSA_NAME_RANGE_INFO (phi_result)
131 && gimple_bb (SSA_NAME_DEF_STMT (new_tree)) == cond_block
132 && dbg_cnt (phiopt_edge_range))
133 duplicate_ssa_name_range_info (new_tree, phi_result);
135 /* Change the PHI argument to new. */
136 SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree);
138 /* Remove the empty basic block. */
139 edge edge_to_remove = NULL, keep_edge = NULL;
140 if (EDGE_SUCC (cond_block, 0)->dest == bb)
142 edge_to_remove = EDGE_SUCC (cond_block, 1);
143 keep_edge = EDGE_SUCC (cond_block, 0);
145 else if (EDGE_SUCC (cond_block, 1)->dest == bb)
147 edge_to_remove = EDGE_SUCC (cond_block, 0);
148 keep_edge = EDGE_SUCC (cond_block, 1);
150 else if ((keep_edge = find_edge (cond_block, e->src)))
152 basic_block bb1 = EDGE_SUCC (cond_block, 0)->dest;
153 basic_block bb2 = EDGE_SUCC (cond_block, 1)->dest;
154 if (single_pred_p (bb1) && single_pred_p (bb2)
155 && single_succ_p (bb1) && single_succ_p (bb2)
156 && empty_block_p (bb1) && empty_block_p (bb2))
157 deleteboth = true;
159 else
160 gcc_unreachable ();
162 if (edge_to_remove && EDGE_COUNT (edge_to_remove->dest->preds) == 1)
164 e->flags |= EDGE_FALLTHRU;
165 e->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
166 e->probability = profile_probability::always ();
167 delete_basic_block (edge_to_remove->dest);
169 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
170 gsi = gsi_last_bb (cond_block);
171 gsi_remove (&gsi, true);
173 else if (deleteboth)
175 basic_block bb1 = EDGE_SUCC (cond_block, 0)->dest;
176 basic_block bb2 = EDGE_SUCC (cond_block, 1)->dest;
178 edge newedge = redirect_edge_and_branch (keep_edge, bb);
180 /* The new edge should be the same. */
181 gcc_assert (newedge == keep_edge);
183 keep_edge->flags |= EDGE_FALLTHRU;
184 keep_edge->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
185 keep_edge->probability = profile_probability::always ();
187 /* Copy the edge's phi entry from the old one. */
188 copy_phi_arg_into_existing_phi (e, keep_edge);
190 /* Delete the old 2 empty basic blocks */
191 delete_basic_block (bb1);
192 delete_basic_block (bb2);
194 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
195 gsi = gsi_last_bb (cond_block);
196 gsi_remove (&gsi, true);
198 else
200 /* If there are other edges into the middle block make
201 CFG cleanup deal with the edge removal to avoid
202 updating dominators here in a non-trivial way. */
203 gcond *cond = as_a <gcond *> (*gsi_last_bb (cond_block));
204 if (keep_edge->flags & EDGE_FALSE_VALUE)
205 gimple_cond_make_false (cond);
206 else if (keep_edge->flags & EDGE_TRUE_VALUE)
207 gimple_cond_make_true (cond);
210 if (dce_ssa_names)
211 simple_dce_from_worklist (dce_ssa_names);
213 statistics_counter_event (cfun, "Replace PHI with variable", 1);
215 if (dump_file && (dump_flags & TDF_DETAILS))
216 fprintf (dump_file,
217 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
218 cond_block->index,
219 bb->index);
222 /* PR66726: Factor operations out of COND_EXPR. If the arguments of the PHI
223 stmt are CONVERT_STMT, factor out the conversion and perform the conversion
224 to the result of PHI stmt. COND_STMT is the controlling predicate.
225 Return the newly-created PHI, if any. */
227 static gphi *
228 factor_out_conditional_operation (edge e0, edge e1, gphi *phi,
229 tree arg0, tree arg1, gimple *cond_stmt)
231 gimple *arg0_def_stmt = NULL, *arg1_def_stmt = NULL, *new_stmt;
232 tree new_arg0 = NULL_TREE, new_arg1 = NULL_TREE;
233 tree temp, result;
234 gphi *newphi;
235 gimple_stmt_iterator gsi, gsi_for_def;
236 location_t locus = gimple_location (phi);
237 enum tree_code op_code;
239 /* Handle only PHI statements with two arguments. TODO: If all
240 other arguments to PHI are INTEGER_CST or if their defining
241 statement have the same unary operation, we can handle more
242 than two arguments too. */
243 if (gimple_phi_num_args (phi) != 2)
244 return NULL;
246 /* First canonicalize to simplify tests. */
247 if (TREE_CODE (arg0) != SSA_NAME)
249 std::swap (arg0, arg1);
250 std::swap (e0, e1);
253 if (TREE_CODE (arg0) != SSA_NAME
254 || (TREE_CODE (arg1) != SSA_NAME
255 && TREE_CODE (arg1) != INTEGER_CST))
256 return NULL;
258 /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is
259 an unary operation. */
260 arg0_def_stmt = SSA_NAME_DEF_STMT (arg0);
261 if (!is_gimple_assign (arg0_def_stmt)
262 || (gimple_assign_rhs_class (arg0_def_stmt) != GIMPLE_UNARY_RHS
263 && gimple_assign_rhs_code (arg0_def_stmt) != VIEW_CONVERT_EXPR))
264 return NULL;
266 /* Use the RHS as new_arg0. */
267 op_code = gimple_assign_rhs_code (arg0_def_stmt);
268 new_arg0 = gimple_assign_rhs1 (arg0_def_stmt);
269 if (op_code == VIEW_CONVERT_EXPR)
271 new_arg0 = TREE_OPERAND (new_arg0, 0);
272 if (!is_gimple_reg_type (TREE_TYPE (new_arg0)))
273 return NULL;
275 if (TREE_CODE (new_arg0) == SSA_NAME
276 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (new_arg0))
277 return NULL;
279 if (TREE_CODE (arg1) == SSA_NAME)
281 /* Check if arg1 is an SSA_NAME and the stmt which defines arg1
282 is an unary operation. */
283 arg1_def_stmt = SSA_NAME_DEF_STMT (arg1);
284 if (!is_gimple_assign (arg1_def_stmt)
285 || gimple_assign_rhs_code (arg1_def_stmt) != op_code)
286 return NULL;
288 /* Either arg1_def_stmt or arg0_def_stmt should be conditional. */
289 if (dominated_by_p (CDI_DOMINATORS, gimple_bb (phi), gimple_bb (arg0_def_stmt))
290 && dominated_by_p (CDI_DOMINATORS,
291 gimple_bb (phi), gimple_bb (arg1_def_stmt)))
292 return NULL;
294 /* Use the RHS as new_arg1. */
295 new_arg1 = gimple_assign_rhs1 (arg1_def_stmt);
296 if (op_code == VIEW_CONVERT_EXPR)
297 new_arg1 = TREE_OPERAND (new_arg1, 0);
298 if (TREE_CODE (new_arg1) == SSA_NAME
299 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (new_arg1))
300 return NULL;
302 else
304 /* TODO: handle more than just casts here. */
305 if (!gimple_assign_cast_p (arg0_def_stmt))
306 return NULL;
308 /* arg0_def_stmt should be conditional. */
309 if (dominated_by_p (CDI_DOMINATORS, gimple_bb (phi), gimple_bb (arg0_def_stmt)))
310 return NULL;
311 /* If arg1 is an INTEGER_CST, fold it to new type. */
312 if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0))
313 && (int_fits_type_p (arg1, TREE_TYPE (new_arg0))
314 || (TYPE_PRECISION (TREE_TYPE (new_arg0))
315 == TYPE_PRECISION (TREE_TYPE (arg1)))))
317 if (gimple_assign_cast_p (arg0_def_stmt))
319 /* For the INTEGER_CST case, we are just moving the
320 conversion from one place to another, which can often
321 hurt as the conversion moves further away from the
322 statement that computes the value. So, perform this
323 only if new_arg0 is an operand of COND_STMT, or
324 if arg0_def_stmt is the only non-debug stmt in
325 its basic block, because then it is possible this
326 could enable further optimizations (minmax replacement
327 etc.). See PR71016.
328 Note no-op conversions don't have this issue as
329 it will not generate any zero/sign extend in that case. */
330 if ((TYPE_PRECISION (TREE_TYPE (new_arg0))
331 != TYPE_PRECISION (TREE_TYPE (arg1)))
332 && new_arg0 != gimple_cond_lhs (cond_stmt)
333 && new_arg0 != gimple_cond_rhs (cond_stmt)
334 && gimple_bb (arg0_def_stmt) == e0->src)
336 gsi = gsi_for_stmt (arg0_def_stmt);
337 gsi_prev_nondebug (&gsi);
338 if (!gsi_end_p (gsi))
340 if (gassign *assign
341 = dyn_cast <gassign *> (gsi_stmt (gsi)))
343 tree lhs = gimple_assign_lhs (assign);
344 enum tree_code ass_code
345 = gimple_assign_rhs_code (assign);
346 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
347 return NULL;
348 if (lhs != gimple_assign_rhs1 (arg0_def_stmt))
349 return NULL;
350 gsi_prev_nondebug (&gsi);
351 if (!gsi_end_p (gsi))
352 return NULL;
354 else
355 return NULL;
357 gsi = gsi_for_stmt (arg0_def_stmt);
358 gsi_next_nondebug (&gsi);
359 if (!gsi_end_p (gsi))
360 return NULL;
362 new_arg1 = fold_convert (TREE_TYPE (new_arg0), arg1);
364 /* Drop the overlow that fold_convert might add. */
365 if (TREE_OVERFLOW (new_arg1))
366 new_arg1 = drop_tree_overflow (new_arg1);
368 else
369 return NULL;
371 else
372 return NULL;
375 /* If arg0/arg1 have > 1 use, then this transformation actually increases
376 the number of expressions evaluated at runtime. */
377 if (!has_single_use (arg0)
378 || (arg1_def_stmt && !has_single_use (arg1)))
379 return NULL;
381 /* If types of new_arg0 and new_arg1 are different bailout. */
382 if (!types_compatible_p (TREE_TYPE (new_arg0), TREE_TYPE (new_arg1)))
383 return NULL;
385 /* Create a new PHI stmt. */
386 result = PHI_RESULT (phi);
387 temp = make_ssa_name (TREE_TYPE (new_arg0), NULL);
388 newphi = create_phi_node (temp, gimple_bb (phi));
390 if (dump_file && (dump_flags & TDF_DETAILS))
392 fprintf (dump_file, "PHI ");
393 print_generic_expr (dump_file, gimple_phi_result (phi));
394 fprintf (dump_file,
395 " changed to factor operation out from COND_EXPR.\n");
396 fprintf (dump_file, "New stmt with OPERATION that defines ");
397 print_generic_expr (dump_file, result);
398 fprintf (dump_file, ".\n");
401 /* Remove the old operation(s) that has single use. */
402 gsi_for_def = gsi_for_stmt (arg0_def_stmt);
403 gsi_remove (&gsi_for_def, true);
404 release_defs (arg0_def_stmt);
406 if (arg1_def_stmt)
408 gsi_for_def = gsi_for_stmt (arg1_def_stmt);
409 gsi_remove (&gsi_for_def, true);
410 release_defs (arg1_def_stmt);
413 add_phi_arg (newphi, new_arg0, e0, locus);
414 add_phi_arg (newphi, new_arg1, e1, locus);
416 /* Create the operation stmt and insert it. */
417 if (op_code == VIEW_CONVERT_EXPR)
419 temp = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (result), temp);
420 new_stmt = gimple_build_assign (result, temp);
422 else
423 new_stmt = gimple_build_assign (result, op_code, temp);
424 gsi = gsi_after_labels (gimple_bb (phi));
425 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
427 /* Remove the original PHI stmt. */
428 gsi = gsi_for_stmt (phi);
429 gsi_remove (&gsi, true);
431 statistics_counter_event (cfun, "factored out operation", 1);
433 return newphi;
437 /* Return TRUE if SEQ/OP pair should be allowed during early phiopt.
438 Currently this is to allow MIN/MAX and ABS/NEGATE and constants. */
439 static bool
440 phiopt_early_allow (gimple_seq &seq, gimple_match_op &op)
442 /* Don't allow functions. */
443 if (!op.code.is_tree_code ())
444 return false;
445 tree_code code = (tree_code)op.code;
447 /* For non-empty sequence, only allow one statement
448 except for MIN/MAX, allow max 2 statements,
449 each with MIN/MAX. */
450 if (!gimple_seq_empty_p (seq))
452 if (code == MIN_EXPR || code == MAX_EXPR)
454 if (!gimple_seq_singleton_p (seq))
455 return false;
457 gimple *stmt = gimple_seq_first_stmt (seq);
458 /* Only allow assignments. */
459 if (!is_gimple_assign (stmt))
460 return false;
461 code = gimple_assign_rhs_code (stmt);
462 return code == MIN_EXPR || code == MAX_EXPR;
464 /* Check to make sure op was already a SSA_NAME. */
465 if (code != SSA_NAME)
466 return false;
467 if (!gimple_seq_singleton_p (seq))
468 return false;
469 gimple *stmt = gimple_seq_first_stmt (seq);
470 /* Only allow assignments. */
471 if (!is_gimple_assign (stmt))
472 return false;
473 if (gimple_assign_lhs (stmt) != op.ops[0])
474 return false;
475 code = gimple_assign_rhs_code (stmt);
478 switch (code)
480 case MIN_EXPR:
481 case MAX_EXPR:
482 case ABS_EXPR:
483 case ABSU_EXPR:
484 case NEGATE_EXPR:
485 case SSA_NAME:
486 return true;
487 case INTEGER_CST:
488 case REAL_CST:
489 case VECTOR_CST:
490 case FIXED_CST:
491 return true;
492 default:
493 return false;
497 /* gimple_simplify_phiopt is like gimple_simplify but designed for PHIOPT.
498 Return NULL if nothing can be simplified or the resulting simplified value
499 with parts pushed if EARLY_P was true. Also rejects non allowed tree code
500 if EARLY_P is set.
501 Takes the comparison from COMP_STMT and two args, ARG0 and ARG1 and tries
502 to simplify CMP ? ARG0 : ARG1.
503 Also try to simplify (!CMP) ? ARG1 : ARG0 if the non-inverse failed. */
504 static tree
505 gimple_simplify_phiopt (bool early_p, tree type, gimple *comp_stmt,
506 tree arg0, tree arg1,
507 gimple_seq *seq)
509 gimple_seq seq1 = NULL;
510 enum tree_code comp_code = gimple_cond_code (comp_stmt);
511 location_t loc = gimple_location (comp_stmt);
512 tree cmp0 = gimple_cond_lhs (comp_stmt);
513 tree cmp1 = gimple_cond_rhs (comp_stmt);
514 /* To handle special cases like floating point comparison, it is easier and
515 less error-prone to build a tree and gimplify it on the fly though it is
516 less efficient.
517 Don't use fold_build2 here as that might create (bool)a instead of just
518 "a != 0". */
519 tree cond = build2_loc (loc, comp_code, boolean_type_node,
520 cmp0, cmp1);
522 if (dump_file && (dump_flags & TDF_FOLDING))
524 fprintf (dump_file, "\nphiopt match-simplify trying:\n\t");
525 print_generic_expr (dump_file, cond);
526 fprintf (dump_file, " ? ");
527 print_generic_expr (dump_file, arg0);
528 fprintf (dump_file, " : ");
529 print_generic_expr (dump_file, arg1);
530 fprintf (dump_file, "\n");
533 gimple_match_op op (gimple_match_cond::UNCOND,
534 COND_EXPR, type, cond, arg0, arg1);
536 if (op.resimplify (&seq1, follow_all_ssa_edges))
538 bool allowed = !early_p || phiopt_early_allow (seq1, op);
539 tree result = maybe_push_res_to_seq (&op, &seq1);
540 if (dump_file && (dump_flags & TDF_FOLDING))
542 fprintf (dump_file, "\nphiopt match-simplify back:\n");
543 if (seq1)
544 print_gimple_seq (dump_file, seq1, 0, TDF_VOPS|TDF_MEMSYMS);
545 fprintf (dump_file, "result: ");
546 if (result)
547 print_generic_expr (dump_file, result);
548 else
549 fprintf (dump_file, " (none)");
550 fprintf (dump_file, "\n");
551 if (!allowed)
552 fprintf (dump_file, "rejected because early\n");
554 /* Early we want only to allow some generated tree codes. */
555 if (allowed && result)
557 if (loc != UNKNOWN_LOCATION)
558 annotate_all_with_location (seq1, loc);
559 gimple_seq_add_seq_without_update (seq, seq1);
560 return result;
563 gimple_seq_discard (seq1);
564 seq1 = NULL;
566 /* Try the inverted comparison, that is !COMP ? ARG1 : ARG0. */
567 comp_code = invert_tree_comparison (comp_code, HONOR_NANS (cmp0));
569 if (comp_code == ERROR_MARK)
570 return NULL;
572 cond = build2_loc (loc,
573 comp_code, boolean_type_node,
574 cmp0, cmp1);
576 if (dump_file && (dump_flags & TDF_FOLDING))
578 fprintf (dump_file, "\nphiopt match-simplify trying:\n\t");
579 print_generic_expr (dump_file, cond);
580 fprintf (dump_file, " ? ");
581 print_generic_expr (dump_file, arg1);
582 fprintf (dump_file, " : ");
583 print_generic_expr (dump_file, arg0);
584 fprintf (dump_file, "\n");
587 gimple_match_op op1 (gimple_match_cond::UNCOND,
588 COND_EXPR, type, cond, arg1, arg0);
590 if (op1.resimplify (&seq1, follow_all_ssa_edges))
592 bool allowed = !early_p || phiopt_early_allow (seq1, op1);
593 tree result = maybe_push_res_to_seq (&op1, &seq1);
594 if (dump_file && (dump_flags & TDF_FOLDING))
596 fprintf (dump_file, "\nphiopt match-simplify back:\n");
597 if (seq1)
598 print_gimple_seq (dump_file, seq1, 0, TDF_VOPS|TDF_MEMSYMS);
599 fprintf (dump_file, "result: ");
600 if (result)
601 print_generic_expr (dump_file, result);
602 else
603 fprintf (dump_file, " (none)");
604 fprintf (dump_file, "\n");
605 if (!allowed)
606 fprintf (dump_file, "rejected because early\n");
608 /* Early we want only to allow some generated tree codes. */
609 if (allowed && result)
611 if (loc != UNKNOWN_LOCATION)
612 annotate_all_with_location (seq1, loc);
613 gimple_seq_add_seq_without_update (seq, seq1);
614 return result;
617 gimple_seq_discard (seq1);
619 return NULL;
622 /* empty_bb_or_one_feeding_into_p returns true if bb was empty basic block
623 or it has one cheap preparation statement that feeds into the PHI
624 statement and it sets STMT to that statement. */
625 static bool
626 empty_bb_or_one_feeding_into_p (basic_block bb,
627 gimple *phi,
628 gimple *&stmt)
630 stmt = nullptr;
631 gimple *stmt_to_move = nullptr;
632 tree lhs;
634 if (empty_block_p (bb))
635 return true;
637 if (!single_pred_p (bb))
638 return false;
640 /* The middle bb cannot have phi nodes as we don't
641 move those assignments yet. */
642 if (!gimple_seq_empty_p (phi_nodes (bb)))
643 return false;
645 gimple_stmt_iterator gsi;
647 gsi = gsi_start_nondebug_after_labels_bb (bb);
648 while (!gsi_end_p (gsi))
650 gimple *s = gsi_stmt (gsi);
651 gsi_next_nondebug (&gsi);
652 /* Skip over Predict and nop statements. */
653 if (gimple_code (s) == GIMPLE_PREDICT
654 || gimple_code (s) == GIMPLE_NOP)
655 continue;
656 /* If there is more one statement return false. */
657 if (stmt_to_move)
658 return false;
659 stmt_to_move = s;
662 /* The only statement here was a Predict or a nop statement
663 so return true. */
664 if (!stmt_to_move)
665 return true;
667 if (gimple_vuse (stmt_to_move))
668 return false;
670 if (gimple_could_trap_p (stmt_to_move)
671 || gimple_has_side_effects (stmt_to_move))
672 return false;
674 ssa_op_iter it;
675 tree use;
676 FOR_EACH_SSA_TREE_OPERAND (use, stmt_to_move, it, SSA_OP_USE)
677 if (ssa_name_maybe_undef_p (use))
678 return false;
680 /* Allow assignments but allow some builtin/internal calls.
681 As const calls don't match any of the above, yet they could
682 still have some side-effects - they could contain
683 gimple_could_trap_p statements, like floating point
684 exceptions or integer division by zero. See PR70586.
685 FIXME: perhaps gimple_has_side_effects or gimple_could_trap_p
686 should handle this.
687 Allow some known builtin/internal calls that are known not to
688 trap: logical functions (e.g. bswap and bit counting). */
689 if (!is_gimple_assign (stmt_to_move))
691 if (!is_gimple_call (stmt_to_move))
692 return false;
693 combined_fn cfn = gimple_call_combined_fn (stmt_to_move);
694 switch (cfn)
696 default:
697 return false;
698 case CFN_BUILT_IN_BSWAP16:
699 case CFN_BUILT_IN_BSWAP32:
700 case CFN_BUILT_IN_BSWAP64:
701 case CFN_BUILT_IN_BSWAP128:
702 CASE_CFN_FFS:
703 CASE_CFN_PARITY:
704 CASE_CFN_POPCOUNT:
705 CASE_CFN_CLZ:
706 CASE_CFN_CTZ:
707 case CFN_BUILT_IN_CLRSB:
708 case CFN_BUILT_IN_CLRSBL:
709 case CFN_BUILT_IN_CLRSBLL:
710 lhs = gimple_call_lhs (stmt_to_move);
711 break;
714 else
715 lhs = gimple_assign_lhs (stmt_to_move);
717 gimple *use_stmt;
718 use_operand_p use_p;
720 /* Allow only a statement which feeds into the other stmt. */
721 if (!lhs || TREE_CODE (lhs) != SSA_NAME
722 || !single_imm_use (lhs, &use_p, &use_stmt)
723 || use_stmt != phi)
724 return false;
726 stmt = stmt_to_move;
727 return true;
730 /* Move STMT to before GSI and insert its defining
731 name into INSERTED_EXPRS bitmap. */
732 static void
733 move_stmt (gimple *stmt, gimple_stmt_iterator *gsi, auto_bitmap &inserted_exprs)
735 if (!stmt)
736 return;
737 if (dump_file && (dump_flags & TDF_DETAILS))
739 fprintf (dump_file, "statement un-sinked:\n");
740 print_gimple_stmt (dump_file, stmt, 0,
741 TDF_VOPS|TDF_MEMSYMS);
744 tree name = gimple_get_lhs (stmt);
745 // Mark the name to be renamed if there is one.
746 bitmap_set_bit (inserted_exprs, SSA_NAME_VERSION (name));
747 gimple_stmt_iterator gsi1 = gsi_for_stmt (stmt);
748 gsi_move_before (&gsi1, gsi);
749 reset_flow_sensitive_info (name);
752 /* RAII style class to temporarily remove flow sensitive
753 from ssa names defined by a gimple statement. */
754 class auto_flow_sensitive
756 public:
757 auto_flow_sensitive (gimple *s);
758 ~auto_flow_sensitive ();
759 private:
760 auto_vec<std::pair<tree, flow_sensitive_info_storage>, 2> stack;
763 /* Constructor for auto_flow_sensitive. Saves
764 off the ssa names' flow sensitive information
765 that was defined by gimple statement S and
766 resets it to be non-flow based ones. */
768 auto_flow_sensitive::auto_flow_sensitive (gimple *s)
770 if (!s)
771 return;
772 ssa_op_iter it;
773 tree def;
774 FOR_EACH_SSA_TREE_OPERAND (def, s, it, SSA_OP_DEF)
776 flow_sensitive_info_storage storage;
777 storage.save_and_clear (def);
778 stack.safe_push (std::make_pair (def, storage));
782 /* Deconstructor, restores the flow sensitive information
783 for the SSA names that had been saved off. */
785 auto_flow_sensitive::~auto_flow_sensitive ()
787 for (auto p : stack)
788 p.second.restore (p.first);
791 /* The function match_simplify_replacement does the main work of doing the
792 replacement using match and simplify. Return true if the replacement is done.
793 Otherwise return false.
794 BB is the basic block where the replacement is going to be done on. ARG0
795 is argument 0 from PHI. Likewise for ARG1. */
797 static bool
798 match_simplify_replacement (basic_block cond_bb, basic_block middle_bb,
799 basic_block middle_bb_alt,
800 edge e0, edge e1, gphi *phi,
801 tree arg0, tree arg1, bool early_p,
802 bool threeway_p)
804 gimple *stmt;
805 gimple_stmt_iterator gsi;
806 edge true_edge, false_edge;
807 gimple_seq seq = NULL;
808 tree result;
809 gimple *stmt_to_move = NULL;
810 gimple *stmt_to_move_alt = NULL;
811 tree arg_true, arg_false;
813 /* Special case A ? B : B as this will always simplify to B. */
814 if (operand_equal_for_phi_arg_p (arg0, arg1))
815 return false;
817 /* If the basic block only has a cheap preparation statement,
818 allow it and move it once the transformation is done. */
819 if (!empty_bb_or_one_feeding_into_p (middle_bb, phi, stmt_to_move))
820 return false;
822 if (threeway_p
823 && middle_bb != middle_bb_alt
824 && !empty_bb_or_one_feeding_into_p (middle_bb_alt, phi,
825 stmt_to_move_alt))
826 return false;
828 /* At this point we know we have a GIMPLE_COND with two successors.
829 One successor is BB, the other successor is an empty block which
830 falls through into BB.
832 There is a single PHI node at the join point (BB).
834 So, given the condition COND, and the two PHI arguments, match and simplify
835 can happen on (COND) ? arg0 : arg1. */
837 stmt = last_nondebug_stmt (cond_bb);
839 /* We need to know which is the true edge and which is the false
840 edge so that we know when to invert the condition below. */
841 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
843 /* Forward the edges over the middle basic block. */
844 if (true_edge->dest == middle_bb)
845 true_edge = EDGE_SUCC (true_edge->dest, 0);
846 if (false_edge->dest == middle_bb)
847 false_edge = EDGE_SUCC (false_edge->dest, 0);
849 /* When THREEWAY_P then e1 will point to the edge of the final transition
850 from middle-bb to end. */
851 if (true_edge == e0)
853 if (!threeway_p)
854 gcc_assert (false_edge == e1);
855 arg_true = arg0;
856 arg_false = arg1;
858 else
860 gcc_assert (false_edge == e0);
861 if (!threeway_p)
862 gcc_assert (true_edge == e1);
863 arg_true = arg1;
864 arg_false = arg0;
867 /* Do not make conditional undefs unconditional. */
868 if ((TREE_CODE (arg_true) == SSA_NAME
869 && ssa_name_maybe_undef_p (arg_true))
870 || (TREE_CODE (arg_false) == SSA_NAME
871 && ssa_name_maybe_undef_p (arg_false)))
872 return false;
874 tree type = TREE_TYPE (gimple_phi_result (phi));
876 auto_flow_sensitive s1(stmt_to_move);
877 auto_flow_sensitive s_alt(stmt_to_move_alt);
879 result = gimple_simplify_phiopt (early_p, type, stmt,
880 arg_true, arg_false,
881 &seq);
884 if (!result)
885 return false;
886 if (dump_file && (dump_flags & TDF_FOLDING))
887 fprintf (dump_file, "accepted the phiopt match-simplify.\n");
889 auto_bitmap exprs_maybe_dce;
891 /* Mark the cond statements' lhs/rhs as maybe dce. */
892 if (TREE_CODE (gimple_cond_lhs (stmt)) == SSA_NAME
893 && !SSA_NAME_IS_DEFAULT_DEF (gimple_cond_lhs (stmt)))
894 bitmap_set_bit (exprs_maybe_dce,
895 SSA_NAME_VERSION (gimple_cond_lhs (stmt)));
896 if (TREE_CODE (gimple_cond_rhs (stmt)) == SSA_NAME
897 && !SSA_NAME_IS_DEFAULT_DEF (gimple_cond_rhs (stmt)))
898 bitmap_set_bit (exprs_maybe_dce,
899 SSA_NAME_VERSION (gimple_cond_rhs (stmt)));
901 gsi = gsi_last_bb (cond_bb);
902 /* Insert the sequence generated from gimple_simplify_phiopt. */
903 if (seq)
905 // Mark the lhs of the new statements maybe for dce
906 gimple_stmt_iterator gsi1 = gsi_start (seq);
907 for (; !gsi_end_p (gsi1); gsi_next (&gsi1))
909 gimple *stmt = gsi_stmt (gsi1);
910 tree name = gimple_get_lhs (stmt);
911 if (name && TREE_CODE (name) == SSA_NAME)
912 bitmap_set_bit (exprs_maybe_dce, SSA_NAME_VERSION (name));
914 gsi_insert_seq_before (&gsi, seq, GSI_CONTINUE_LINKING);
917 /* If there was a statement to move, move it to right before
918 the original conditional. */
919 move_stmt (stmt_to_move, &gsi, exprs_maybe_dce);
920 move_stmt (stmt_to_move_alt, &gsi, exprs_maybe_dce);
922 replace_phi_edge_with_variable (cond_bb, e1, phi, result, exprs_maybe_dce);
924 /* Add Statistic here even though replace_phi_edge_with_variable already
925 does it as we want to be able to count when match-simplify happens vs
926 the others. */
927 statistics_counter_event (cfun, "match-simplify PHI replacement", 1);
929 /* Note that we optimized this PHI. */
930 return true;
933 /* Update *ARG which is defined in STMT so that it contains the
934 computed value if that seems profitable. Return true if the
935 statement is made dead by that rewriting. */
937 static bool
938 jump_function_from_stmt (tree *arg, gimple *stmt)
940 enum tree_code code = gimple_assign_rhs_code (stmt);
941 if (code == ADDR_EXPR)
943 /* For arg = &p->i transform it to p, if possible. */
944 tree rhs1 = gimple_assign_rhs1 (stmt);
945 poly_int64 offset;
946 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0),
947 &offset);
948 if (tem
949 && TREE_CODE (tem) == MEM_REF
950 && known_eq (mem_ref_offset (tem) + offset, 0))
952 *arg = TREE_OPERAND (tem, 0);
953 return true;
956 /* TODO: Much like IPA-CP jump-functions we want to handle constant
957 additions symbolically here, and we'd need to update the comparison
958 code that compares the arg + cst tuples in our caller. For now the
959 code above exactly handles the VEC_BASE pattern from vec.h. */
960 return false;
963 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
964 of the form SSA_NAME NE 0.
966 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
967 the two input values of the EQ_EXPR match arg0 and arg1.
969 If so update *code and return TRUE. Otherwise return FALSE. */
971 static bool
972 rhs_is_fed_for_value_replacement (const_tree arg0, const_tree arg1,
973 enum tree_code *code, const_tree rhs)
975 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
976 statement. */
977 if (TREE_CODE (rhs) == SSA_NAME)
979 gimple *def1 = SSA_NAME_DEF_STMT (rhs);
981 /* Verify the defining statement has an EQ_EXPR on the RHS. */
982 if (is_gimple_assign (def1) && gimple_assign_rhs_code (def1) == EQ_EXPR)
984 /* Finally verify the source operands of the EQ_EXPR are equal
985 to arg0 and arg1. */
986 tree op0 = gimple_assign_rhs1 (def1);
987 tree op1 = gimple_assign_rhs2 (def1);
988 if ((operand_equal_for_phi_arg_p (arg0, op0)
989 && operand_equal_for_phi_arg_p (arg1, op1))
990 || (operand_equal_for_phi_arg_p (arg0, op1)
991 && operand_equal_for_phi_arg_p (arg1, op0)))
993 /* We will perform the optimization. */
994 *code = gimple_assign_rhs_code (def1);
995 return true;
999 return false;
1002 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
1004 Also return TRUE if arg0/arg1 are equal to the source arguments of a
1005 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
1007 Return FALSE otherwise. */
1009 static bool
1010 operand_equal_for_value_replacement (const_tree arg0, const_tree arg1,
1011 enum tree_code *code, gimple *cond)
1013 gimple *def;
1014 tree lhs = gimple_cond_lhs (cond);
1015 tree rhs = gimple_cond_rhs (cond);
1017 if ((operand_equal_for_phi_arg_p (arg0, lhs)
1018 && operand_equal_for_phi_arg_p (arg1, rhs))
1019 || (operand_equal_for_phi_arg_p (arg1, lhs)
1020 && operand_equal_for_phi_arg_p (arg0, rhs)))
1021 return true;
1023 /* Now handle more complex case where we have an EQ comparison
1024 which feeds a BIT_AND_EXPR which feeds COND.
1026 First verify that COND is of the form SSA_NAME NE 0. */
1027 if (*code != NE_EXPR || !integer_zerop (rhs)
1028 || TREE_CODE (lhs) != SSA_NAME)
1029 return false;
1031 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
1032 def = SSA_NAME_DEF_STMT (lhs);
1033 if (!is_gimple_assign (def) || gimple_assign_rhs_code (def) != BIT_AND_EXPR)
1034 return false;
1036 /* Now verify arg0/arg1 correspond to the source arguments of an
1037 EQ comparison feeding the BIT_AND_EXPR. */
1039 tree tmp = gimple_assign_rhs1 (def);
1040 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
1041 return true;
1043 tmp = gimple_assign_rhs2 (def);
1044 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
1045 return true;
1047 return false;
1050 /* Returns true if ARG is a neutral element for operation CODE
1051 on the RIGHT side. */
1053 static bool
1054 neutral_element_p (tree_code code, tree arg, bool right)
1056 switch (code)
1058 case PLUS_EXPR:
1059 case BIT_IOR_EXPR:
1060 case BIT_XOR_EXPR:
1061 return integer_zerop (arg);
1063 case LROTATE_EXPR:
1064 case RROTATE_EXPR:
1065 case LSHIFT_EXPR:
1066 case RSHIFT_EXPR:
1067 case MINUS_EXPR:
1068 case POINTER_PLUS_EXPR:
1069 return right && integer_zerop (arg);
1071 case MULT_EXPR:
1072 return integer_onep (arg);
1074 case TRUNC_DIV_EXPR:
1075 case CEIL_DIV_EXPR:
1076 case FLOOR_DIV_EXPR:
1077 case ROUND_DIV_EXPR:
1078 case EXACT_DIV_EXPR:
1079 return right && integer_onep (arg);
1081 case BIT_AND_EXPR:
1082 return integer_all_onesp (arg);
1084 default:
1085 return false;
1089 /* Returns true if ARG is an absorbing element for operation CODE. */
1091 static bool
1092 absorbing_element_p (tree_code code, tree arg, bool right, tree rval)
1094 switch (code)
1096 case BIT_IOR_EXPR:
1097 return integer_all_onesp (arg);
1099 case MULT_EXPR:
1100 case BIT_AND_EXPR:
1101 return integer_zerop (arg);
1103 case LSHIFT_EXPR:
1104 case RSHIFT_EXPR:
1105 case LROTATE_EXPR:
1106 case RROTATE_EXPR:
1107 return !right && integer_zerop (arg);
1109 case TRUNC_DIV_EXPR:
1110 case CEIL_DIV_EXPR:
1111 case FLOOR_DIV_EXPR:
1112 case ROUND_DIV_EXPR:
1113 case EXACT_DIV_EXPR:
1114 case TRUNC_MOD_EXPR:
1115 case CEIL_MOD_EXPR:
1116 case FLOOR_MOD_EXPR:
1117 case ROUND_MOD_EXPR:
1118 return (!right
1119 && integer_zerop (arg)
1120 && tree_single_nonzero_warnv_p (rval, NULL));
1122 default:
1123 return false;
1127 /* The function value_replacement does the main work of doing the value
1128 replacement. Return non-zero if the replacement is done. Otherwise return
1129 0. If we remove the middle basic block, return 2.
1130 BB is the basic block where the replacement is going to be done on. ARG0
1131 is argument 0 from the PHI. Likewise for ARG1. */
1133 static int
1134 value_replacement (basic_block cond_bb, basic_block middle_bb,
1135 edge e0, edge e1, gphi *phi, tree arg0, tree arg1)
1137 gimple_stmt_iterator gsi;
1138 edge true_edge, false_edge;
1139 enum tree_code code;
1140 bool empty_or_with_defined_p = true;
1142 /* If the type says honor signed zeros we cannot do this
1143 optimization. */
1144 if (HONOR_SIGNED_ZEROS (arg1))
1145 return 0;
1147 /* If there is a statement in MIDDLE_BB that defines one of the PHI
1148 arguments, then adjust arg0 or arg1. */
1149 gsi = gsi_start_nondebug_after_labels_bb (middle_bb);
1150 while (!gsi_end_p (gsi))
1152 gimple *stmt = gsi_stmt (gsi);
1153 tree lhs;
1154 gsi_next_nondebug (&gsi);
1155 if (!is_gimple_assign (stmt))
1157 if (gimple_code (stmt) != GIMPLE_PREDICT
1158 && gimple_code (stmt) != GIMPLE_NOP)
1159 empty_or_with_defined_p = false;
1160 continue;
1162 /* Now try to adjust arg0 or arg1 according to the computation
1163 in the statement. */
1164 lhs = gimple_assign_lhs (stmt);
1165 if (!(lhs == arg0
1166 && jump_function_from_stmt (&arg0, stmt))
1167 || (lhs == arg1
1168 && jump_function_from_stmt (&arg1, stmt)))
1169 empty_or_with_defined_p = false;
1172 gcond *cond = as_a <gcond *> (*gsi_last_bb (cond_bb));
1173 code = gimple_cond_code (cond);
1175 /* This transformation is only valid for equality comparisons. */
1176 if (code != NE_EXPR && code != EQ_EXPR)
1177 return 0;
1179 /* We need to know which is the true edge and which is the false
1180 edge so that we know if have abs or negative abs. */
1181 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1183 /* At this point we know we have a COND_EXPR with two successors.
1184 One successor is BB, the other successor is an empty block which
1185 falls through into BB.
1187 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
1189 There is a single PHI node at the join point (BB) with two arguments.
1191 We now need to verify that the two arguments in the PHI node match
1192 the two arguments to the equality comparison. */
1194 bool equal_p = operand_equal_for_value_replacement (arg0, arg1, &code, cond);
1195 bool maybe_equal_p = false;
1196 if (!equal_p
1197 && empty_or_with_defined_p
1198 && TREE_CODE (gimple_cond_rhs (cond)) == INTEGER_CST
1199 && (operand_equal_for_phi_arg_p (gimple_cond_lhs (cond), arg0)
1200 ? TREE_CODE (arg1) == INTEGER_CST
1201 : (operand_equal_for_phi_arg_p (gimple_cond_lhs (cond), arg1)
1202 && TREE_CODE (arg0) == INTEGER_CST)))
1203 maybe_equal_p = true;
1204 if (equal_p || maybe_equal_p)
1206 edge e;
1207 tree arg;
1209 /* For NE_EXPR, we want to build an assignment result = arg where
1210 arg is the PHI argument associated with the true edge. For
1211 EQ_EXPR we want the PHI argument associated with the false edge. */
1212 e = (code == NE_EXPR ? true_edge : false_edge);
1214 /* Unfortunately, E may not reach BB (it may instead have gone to
1215 OTHER_BLOCK). If that is the case, then we want the single outgoing
1216 edge from OTHER_BLOCK which reaches BB and represents the desired
1217 path from COND_BLOCK. */
1218 if (e->dest == middle_bb)
1219 e = single_succ_edge (e->dest);
1221 /* Now we know the incoming edge to BB that has the argument for the
1222 RHS of our new assignment statement. */
1223 if (e0 == e)
1224 arg = arg0;
1225 else
1226 arg = arg1;
1228 /* If the middle basic block was empty or is defining the
1229 PHI arguments and this is a single phi where the args are different
1230 for the edges e0 and e1 then we can remove the middle basic block. */
1231 if (empty_or_with_defined_p
1232 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)),
1233 e0, e1) == phi)
1235 use_operand_p use_p;
1236 gimple *use_stmt;
1238 /* Even if arg0/arg1 isn't equal to second operand of cond, we
1239 can optimize away the bb if we can prove it doesn't care whether
1240 phi result is arg0/arg1 or second operand of cond. Consider:
1241 <bb 2> [local count: 118111600]:
1242 if (i_2(D) == 4)
1243 goto <bb 4>; [97.00%]
1244 else
1245 goto <bb 3>; [3.00%]
1247 <bb 3> [local count: 3540129]:
1249 <bb 4> [local count: 118111600]:
1250 # i_6 = PHI <i_2(D)(3), 6(2)>
1251 _3 = i_6 != 0;
1252 Here, carg is 4, oarg is 6, crhs is 0, and because
1253 (4 != 0) == (6 != 0), we don't care if i_6 is 4 or 6, both
1254 have the same outcome. So, can can optimize this to:
1255 _3 = i_2(D) != 0;
1256 If the single imm use of phi result >, >=, < or <=, similarly
1257 we can check if both carg and oarg compare the same against
1258 crhs using ccode. */
1259 if (maybe_equal_p
1260 && TREE_CODE (arg) != INTEGER_CST
1261 && single_imm_use (gimple_phi_result (phi), &use_p, &use_stmt))
1263 enum tree_code ccode = ERROR_MARK;
1264 tree clhs = NULL_TREE, crhs = NULL_TREE;
1265 tree carg = gimple_cond_rhs (cond);
1266 tree oarg = e0 == e ? arg1 : arg0;
1267 if (is_gimple_assign (use_stmt)
1268 && (TREE_CODE_CLASS (gimple_assign_rhs_code (use_stmt))
1269 == tcc_comparison))
1271 ccode = gimple_assign_rhs_code (use_stmt);
1272 clhs = gimple_assign_rhs1 (use_stmt);
1273 crhs = gimple_assign_rhs2 (use_stmt);
1275 else if (gimple_code (use_stmt) == GIMPLE_COND)
1277 ccode = gimple_cond_code (use_stmt);
1278 clhs = gimple_cond_lhs (use_stmt);
1279 crhs = gimple_cond_rhs (use_stmt);
1281 if (ccode != ERROR_MARK
1282 && clhs == gimple_phi_result (phi)
1283 && TREE_CODE (crhs) == INTEGER_CST)
1284 switch (ccode)
1286 case EQ_EXPR:
1287 case NE_EXPR:
1288 if (!tree_int_cst_equal (crhs, carg)
1289 && !tree_int_cst_equal (crhs, oarg))
1290 equal_p = true;
1291 break;
1292 case GT_EXPR:
1293 if (tree_int_cst_lt (crhs, carg)
1294 == tree_int_cst_lt (crhs, oarg))
1295 equal_p = true;
1296 break;
1297 case GE_EXPR:
1298 if (tree_int_cst_le (crhs, carg)
1299 == tree_int_cst_le (crhs, oarg))
1300 equal_p = true;
1301 break;
1302 case LT_EXPR:
1303 if (tree_int_cst_lt (carg, crhs)
1304 == tree_int_cst_lt (oarg, crhs))
1305 equal_p = true;
1306 break;
1307 case LE_EXPR:
1308 if (tree_int_cst_le (carg, crhs)
1309 == tree_int_cst_le (oarg, crhs))
1310 equal_p = true;
1311 break;
1312 default:
1313 break;
1315 if (equal_p)
1317 tree phires = gimple_phi_result (phi);
1318 if (SSA_NAME_RANGE_INFO (phires))
1320 /* After the optimization PHI result can have value
1321 which it couldn't have previously. */
1322 int_range_max r;
1323 if (get_global_range_query ()->range_of_expr (r, phires,
1324 phi))
1326 wide_int warg = wi::to_wide (carg);
1327 int_range<2> tmp (TREE_TYPE (carg), warg, warg);
1328 r.union_ (tmp);
1329 reset_flow_sensitive_info (phires);
1330 set_range_info (phires, r);
1332 else
1333 reset_flow_sensitive_info (phires);
1336 if (equal_p && MAY_HAVE_DEBUG_BIND_STMTS)
1338 imm_use_iterator imm_iter;
1339 tree phires = gimple_phi_result (phi);
1340 tree temp = NULL_TREE;
1341 bool reset_p = false;
1343 /* Add # DEBUG D#1 => arg != carg ? arg : oarg. */
1344 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, phires)
1346 if (!is_gimple_debug (use_stmt))
1347 continue;
1348 if (temp == NULL_TREE)
1350 if (!single_pred_p (middle_bb)
1351 || EDGE_COUNT (gimple_bb (phi)->preds) != 2)
1353 /* But only if middle_bb has a single
1354 predecessor and phi bb has two, otherwise
1355 we could use a SSA_NAME not usable in that
1356 place or wrong-debug. */
1357 reset_p = true;
1358 break;
1360 gimple_stmt_iterator gsi
1361 = gsi_after_labels (gimple_bb (phi));
1362 tree type = TREE_TYPE (phires);
1363 temp = build_debug_expr_decl (type);
1364 tree t = build2 (NE_EXPR, boolean_type_node,
1365 arg, carg);
1366 t = build3 (COND_EXPR, type, t, arg, oarg);
1367 gimple *g = gimple_build_debug_bind (temp, t, phi);
1368 gsi_insert_before (&gsi, g, GSI_SAME_STMT);
1370 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter)
1371 replace_exp (use_p, temp);
1372 update_stmt (use_stmt);
1374 if (reset_p)
1375 reset_debug_uses (phi);
1378 if (equal_p)
1380 replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
1381 /* Note that we optimized this PHI. */
1382 return 2;
1385 else if (equal_p)
1387 if (!single_pred_p (middle_bb))
1388 return 0;
1389 statistics_counter_event (cfun, "Replace PHI with "
1390 "variable/value_replacement", 1);
1392 /* Replace the PHI arguments with arg. */
1393 SET_PHI_ARG_DEF (phi, e0->dest_idx, arg);
1394 SET_PHI_ARG_DEF (phi, e1->dest_idx, arg);
1395 if (dump_file && (dump_flags & TDF_DETAILS))
1397 fprintf (dump_file, "PHI ");
1398 print_generic_expr (dump_file, gimple_phi_result (phi));
1399 fprintf (dump_file, " reduced for COND_EXPR in block %d to ",
1400 cond_bb->index);
1401 print_generic_expr (dump_file, arg);
1402 fprintf (dump_file, ".\n");
1404 return 1;
1408 if (!single_pred_p (middle_bb))
1409 return 0;
1411 /* Now optimize (x != 0) ? x + y : y to just x + y. */
1412 gsi = gsi_last_nondebug_bb (middle_bb);
1413 if (gsi_end_p (gsi))
1414 return 0;
1416 gimple *assign = gsi_stmt (gsi);
1417 if (!is_gimple_assign (assign)
1418 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0))
1419 && !POINTER_TYPE_P (TREE_TYPE (arg0))))
1420 return 0;
1422 if (gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS)
1424 /* If last stmt of the middle_bb is a conversion, handle it like
1425 a preparation statement through constant evaluation with
1426 checking for UB. */
1427 enum tree_code sc = gimple_assign_rhs_code (assign);
1428 if (CONVERT_EXPR_CODE_P (sc))
1429 assign = NULL;
1430 else
1431 return 0;
1434 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
1435 if (!gimple_seq_empty_p (phi_nodes (middle_bb)))
1436 return 0;
1438 /* Allow up to 2 cheap preparation statements that prepare argument
1439 for assign, e.g.:
1440 if (y_4 != 0)
1441 goto <bb 3>;
1442 else
1443 goto <bb 4>;
1444 <bb 3>:
1445 _1 = (int) y_4;
1446 iftmp.0_6 = x_5(D) r<< _1;
1447 <bb 4>:
1448 # iftmp.0_2 = PHI <iftmp.0_6(3), x_5(D)(2)>
1450 if (y_3(D) == 0)
1451 goto <bb 4>;
1452 else
1453 goto <bb 3>;
1454 <bb 3>:
1455 y_4 = y_3(D) & 31;
1456 _1 = (int) y_4;
1457 _6 = x_5(D) r<< _1;
1458 <bb 4>:
1459 # _2 = PHI <x_5(D)(2), _6(3)> */
1460 gimple *prep_stmt[2] = { NULL, NULL };
1461 int prep_cnt;
1462 for (prep_cnt = 0; ; prep_cnt++)
1464 if (prep_cnt || assign)
1465 gsi_prev_nondebug (&gsi);
1466 if (gsi_end_p (gsi))
1467 break;
1469 gimple *g = gsi_stmt (gsi);
1470 if (gimple_code (g) == GIMPLE_LABEL)
1471 break;
1473 if (prep_cnt == 2 || !is_gimple_assign (g))
1474 return 0;
1476 tree lhs = gimple_assign_lhs (g);
1477 tree rhs1 = gimple_assign_rhs1 (g);
1478 use_operand_p use_p;
1479 gimple *use_stmt;
1480 if (TREE_CODE (lhs) != SSA_NAME
1481 || TREE_CODE (rhs1) != SSA_NAME
1482 || !INTEGRAL_TYPE_P (TREE_TYPE (lhs))
1483 || !INTEGRAL_TYPE_P (TREE_TYPE (rhs1))
1484 || !single_imm_use (lhs, &use_p, &use_stmt)
1485 || ((prep_cnt || assign)
1486 && use_stmt != (prep_cnt ? prep_stmt[prep_cnt - 1] : assign)))
1487 return 0;
1488 switch (gimple_assign_rhs_code (g))
1490 CASE_CONVERT:
1491 break;
1492 case PLUS_EXPR:
1493 case BIT_AND_EXPR:
1494 case BIT_IOR_EXPR:
1495 case BIT_XOR_EXPR:
1496 if (TREE_CODE (gimple_assign_rhs2 (g)) != INTEGER_CST)
1497 return 0;
1498 break;
1499 default:
1500 return 0;
1502 prep_stmt[prep_cnt] = g;
1505 /* Only transform if it removes the condition. */
1506 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1))
1507 return 0;
1509 /* Size-wise, this is always profitable. */
1510 if (optimize_bb_for_speed_p (cond_bb)
1511 /* The special case is useless if it has a low probability. */
1512 && profile_status_for_fn (cfun) != PROFILE_ABSENT
1513 && EDGE_PRED (middle_bb, 0)->probability < profile_probability::even ()
1514 /* If assign is cheap, there is no point avoiding it. */
1515 && estimate_num_insns_seq (bb_seq (middle_bb), &eni_time_weights)
1516 >= 3 * estimate_num_insns (cond, &eni_time_weights))
1517 return 0;
1519 tree cond_lhs = gimple_cond_lhs (cond);
1520 tree cond_rhs = gimple_cond_rhs (cond);
1522 /* Propagate the cond_rhs constant through preparation stmts,
1523 make sure UB isn't invoked while doing that. */
1524 for (int i = prep_cnt - 1; i >= 0; --i)
1526 gimple *g = prep_stmt[i];
1527 tree grhs1 = gimple_assign_rhs1 (g);
1528 if (!operand_equal_for_phi_arg_p (cond_lhs, grhs1))
1529 return 0;
1530 cond_lhs = gimple_assign_lhs (g);
1531 cond_rhs = fold_convert (TREE_TYPE (grhs1), cond_rhs);
1532 if (TREE_CODE (cond_rhs) != INTEGER_CST
1533 || TREE_OVERFLOW (cond_rhs))
1534 return 0;
1535 if (gimple_assign_rhs_class (g) == GIMPLE_BINARY_RHS)
1537 cond_rhs = int_const_binop (gimple_assign_rhs_code (g), cond_rhs,
1538 gimple_assign_rhs2 (g));
1539 if (TREE_OVERFLOW (cond_rhs))
1540 return 0;
1542 cond_rhs = fold_convert (TREE_TYPE (cond_lhs), cond_rhs);
1543 if (TREE_CODE (cond_rhs) != INTEGER_CST
1544 || TREE_OVERFLOW (cond_rhs))
1545 return 0;
1548 tree lhs, rhs1, rhs2;
1549 enum tree_code code_def;
1550 if (assign)
1552 lhs = gimple_assign_lhs (assign);
1553 rhs1 = gimple_assign_rhs1 (assign);
1554 rhs2 = gimple_assign_rhs2 (assign);
1555 code_def = gimple_assign_rhs_code (assign);
1557 else
1559 gcc_assert (prep_cnt > 0);
1560 lhs = cond_lhs;
1561 rhs1 = NULL_TREE;
1562 rhs2 = NULL_TREE;
1563 code_def = ERROR_MARK;
1566 if (((code == NE_EXPR && e1 == false_edge)
1567 || (code == EQ_EXPR && e1 == true_edge))
1568 && arg0 == lhs
1569 && ((assign == NULL
1570 && operand_equal_for_phi_arg_p (arg1, cond_rhs))
1571 || (assign
1572 && arg1 == rhs1
1573 && operand_equal_for_phi_arg_p (rhs2, cond_lhs)
1574 && neutral_element_p (code_def, cond_rhs, true))
1575 || (assign
1576 && arg1 == rhs2
1577 && operand_equal_for_phi_arg_p (rhs1, cond_lhs)
1578 && neutral_element_p (code_def, cond_rhs, false))
1579 || (assign
1580 && operand_equal_for_phi_arg_p (arg1, cond_rhs)
1581 && ((operand_equal_for_phi_arg_p (rhs2, cond_lhs)
1582 && absorbing_element_p (code_def, cond_rhs, true, rhs2))
1583 || (operand_equal_for_phi_arg_p (rhs1, cond_lhs)
1584 && absorbing_element_p (code_def,
1585 cond_rhs, false, rhs2))))))
1587 gsi = gsi_for_stmt (cond);
1588 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
1589 def-stmt in:
1590 if (n_5 != 0)
1591 goto <bb 3>;
1592 else
1593 goto <bb 4>;
1595 <bb 3>:
1596 # RANGE [0, 4294967294]
1597 u_6 = n_5 + 4294967295;
1599 <bb 4>:
1600 # u_3 = PHI <u_6(3), 4294967295(2)> */
1601 reset_flow_sensitive_info (lhs);
1602 gimple_stmt_iterator gsi_from;
1603 for (int i = prep_cnt - 1; i >= 0; --i)
1605 tree plhs = gimple_assign_lhs (prep_stmt[i]);
1606 reset_flow_sensitive_info (plhs);
1607 gsi_from = gsi_for_stmt (prep_stmt[i]);
1608 gsi_move_before (&gsi_from, &gsi);
1610 if (assign)
1612 gsi_from = gsi_for_stmt (assign);
1613 gsi_move_before (&gsi_from, &gsi);
1615 replace_phi_edge_with_variable (cond_bb, e1, phi, lhs);
1616 return 2;
1619 return 0;
1622 /* If VAR is an SSA_NAME that points to a BIT_NOT_EXPR then return the TREE for
1623 the value being inverted. */
1625 static tree
1626 strip_bit_not (tree var)
1628 if (TREE_CODE (var) != SSA_NAME)
1629 return NULL_TREE;
1631 gimple *assign = SSA_NAME_DEF_STMT (var);
1632 if (gimple_code (assign) != GIMPLE_ASSIGN)
1633 return NULL_TREE;
1635 if (gimple_assign_rhs_code (assign) != BIT_NOT_EXPR)
1636 return NULL_TREE;
1638 return gimple_assign_rhs1 (assign);
1641 /* Invert a MIN to a MAX or a MAX to a MIN expression CODE. */
1643 enum tree_code
1644 invert_minmax_code (enum tree_code code)
1646 switch (code) {
1647 case MIN_EXPR:
1648 return MAX_EXPR;
1649 case MAX_EXPR:
1650 return MIN_EXPR;
1651 default:
1652 gcc_unreachable ();
1656 /* The function minmax_replacement does the main work of doing the minmax
1657 replacement. Return true if the replacement is done. Otherwise return
1658 false.
1659 BB is the basic block where the replacement is going to be done on. ARG0
1660 is argument 0 from the PHI. Likewise for ARG1.
1662 If THREEWAY_P then expect the BB to be laid out in diamond shape with each
1663 BB containing only a MIN or MAX expression. */
1665 static bool
1666 minmax_replacement (basic_block cond_bb, basic_block middle_bb, basic_block alt_middle_bb,
1667 edge e0, edge e1, gphi *phi, tree arg0, tree arg1, bool threeway_p)
1669 tree result;
1670 edge true_edge, false_edge;
1671 enum tree_code minmax, ass_code;
1672 tree smaller, larger, arg_true, arg_false;
1673 gimple_stmt_iterator gsi, gsi_from;
1675 tree type = TREE_TYPE (PHI_RESULT (phi));
1677 gcond *cond = as_a <gcond *> (*gsi_last_bb (cond_bb));
1678 enum tree_code cmp = gimple_cond_code (cond);
1679 tree rhs = gimple_cond_rhs (cond);
1681 /* Turn EQ/NE of extreme values to order comparisons. */
1682 if ((cmp == NE_EXPR || cmp == EQ_EXPR)
1683 && TREE_CODE (rhs) == INTEGER_CST
1684 && INTEGRAL_TYPE_P (TREE_TYPE (rhs)))
1686 if (wi::eq_p (wi::to_wide (rhs), wi::min_value (TREE_TYPE (rhs))))
1688 cmp = (cmp == EQ_EXPR) ? LT_EXPR : GE_EXPR;
1689 rhs = wide_int_to_tree (TREE_TYPE (rhs),
1690 wi::min_value (TREE_TYPE (rhs)) + 1);
1692 else if (wi::eq_p (wi::to_wide (rhs), wi::max_value (TREE_TYPE (rhs))))
1694 cmp = (cmp == EQ_EXPR) ? GT_EXPR : LE_EXPR;
1695 rhs = wide_int_to_tree (TREE_TYPE (rhs),
1696 wi::max_value (TREE_TYPE (rhs)) - 1);
1700 /* This transformation is only valid for order comparisons. Record which
1701 operand is smaller/larger if the result of the comparison is true. */
1702 tree alt_smaller = NULL_TREE;
1703 tree alt_larger = NULL_TREE;
1704 if (cmp == LT_EXPR || cmp == LE_EXPR)
1706 smaller = gimple_cond_lhs (cond);
1707 larger = rhs;
1708 /* If we have smaller < CST it is equivalent to smaller <= CST-1.
1709 Likewise smaller <= CST is equivalent to smaller < CST+1. */
1710 if (TREE_CODE (larger) == INTEGER_CST
1711 && INTEGRAL_TYPE_P (TREE_TYPE (larger)))
1713 if (cmp == LT_EXPR)
1715 wi::overflow_type overflow;
1716 wide_int alt = wi::sub (wi::to_wide (larger), 1,
1717 TYPE_SIGN (TREE_TYPE (larger)),
1718 &overflow);
1719 if (! overflow)
1720 alt_larger = wide_int_to_tree (TREE_TYPE (larger), alt);
1722 else
1724 wi::overflow_type overflow;
1725 wide_int alt = wi::add (wi::to_wide (larger), 1,
1726 TYPE_SIGN (TREE_TYPE (larger)),
1727 &overflow);
1728 if (! overflow)
1729 alt_larger = wide_int_to_tree (TREE_TYPE (larger), alt);
1733 else if (cmp == GT_EXPR || cmp == GE_EXPR)
1735 smaller = rhs;
1736 larger = gimple_cond_lhs (cond);
1737 /* If we have larger > CST it is equivalent to larger >= CST+1.
1738 Likewise larger >= CST is equivalent to larger > CST-1. */
1739 if (TREE_CODE (smaller) == INTEGER_CST
1740 && INTEGRAL_TYPE_P (TREE_TYPE (smaller)))
1742 wi::overflow_type overflow;
1743 if (cmp == GT_EXPR)
1745 wide_int alt = wi::add (wi::to_wide (smaller), 1,
1746 TYPE_SIGN (TREE_TYPE (smaller)),
1747 &overflow);
1748 if (! overflow)
1749 alt_smaller = wide_int_to_tree (TREE_TYPE (smaller), alt);
1751 else
1753 wide_int alt = wi::sub (wi::to_wide (smaller), 1,
1754 TYPE_SIGN (TREE_TYPE (smaller)),
1755 &overflow);
1756 if (! overflow)
1757 alt_smaller = wide_int_to_tree (TREE_TYPE (smaller), alt);
1761 else
1762 return false;
1764 /* Handle the special case of (signed_type)x < 0 being equivalent
1765 to x > MAX_VAL(signed_type) and (signed_type)x >= 0 equivalent
1766 to x <= MAX_VAL(signed_type). */
1767 if ((cmp == GE_EXPR || cmp == LT_EXPR)
1768 && INTEGRAL_TYPE_P (type)
1769 && TYPE_UNSIGNED (type)
1770 && integer_zerop (rhs))
1772 tree op = gimple_cond_lhs (cond);
1773 if (TREE_CODE (op) == SSA_NAME
1774 && INTEGRAL_TYPE_P (TREE_TYPE (op))
1775 && !TYPE_UNSIGNED (TREE_TYPE (op)))
1777 gimple *def_stmt = SSA_NAME_DEF_STMT (op);
1778 if (gimple_assign_cast_p (def_stmt))
1780 tree op1 = gimple_assign_rhs1 (def_stmt);
1781 if (INTEGRAL_TYPE_P (TREE_TYPE (op1))
1782 && TYPE_UNSIGNED (TREE_TYPE (op1))
1783 && (TYPE_PRECISION (TREE_TYPE (op))
1784 == TYPE_PRECISION (TREE_TYPE (op1)))
1785 && useless_type_conversion_p (type, TREE_TYPE (op1)))
1787 wide_int w1 = wi::max_value (TREE_TYPE (op));
1788 wide_int w2 = wi::add (w1, 1);
1789 if (cmp == LT_EXPR)
1791 larger = op1;
1792 smaller = wide_int_to_tree (TREE_TYPE (op1), w1);
1793 alt_smaller = wide_int_to_tree (TREE_TYPE (op1), w2);
1794 alt_larger = NULL_TREE;
1796 else
1798 smaller = op1;
1799 larger = wide_int_to_tree (TREE_TYPE (op1), w1);
1800 alt_larger = wide_int_to_tree (TREE_TYPE (op1), w2);
1801 alt_smaller = NULL_TREE;
1808 /* We need to know which is the true edge and which is the false
1809 edge so that we know if have abs or negative abs. */
1810 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1812 /* Forward the edges over the middle basic block. */
1813 if (true_edge->dest == middle_bb)
1814 true_edge = EDGE_SUCC (true_edge->dest, 0);
1815 if (false_edge->dest == middle_bb)
1816 false_edge = EDGE_SUCC (false_edge->dest, 0);
1818 /* When THREEWAY_P then e1 will point to the edge of the final transition
1819 from middle-bb to end. */
1820 if (true_edge == e0)
1822 if (!threeway_p)
1823 gcc_assert (false_edge == e1);
1824 arg_true = arg0;
1825 arg_false = arg1;
1827 else
1829 gcc_assert (false_edge == e0);
1830 if (!threeway_p)
1831 gcc_assert (true_edge == e1);
1832 arg_true = arg1;
1833 arg_false = arg0;
1836 if (empty_block_p (middle_bb)
1837 && (!threeway_p
1838 || empty_block_p (alt_middle_bb)))
1840 if ((operand_equal_for_phi_arg_p (arg_true, smaller)
1841 || (alt_smaller
1842 && operand_equal_for_phi_arg_p (arg_true, alt_smaller)))
1843 && (operand_equal_for_phi_arg_p (arg_false, larger)
1844 || (alt_larger
1845 && operand_equal_for_phi_arg_p (arg_true, alt_larger))))
1847 /* Case
1849 if (smaller < larger)
1850 rslt = smaller;
1851 else
1852 rslt = larger; */
1853 minmax = MIN_EXPR;
1855 else if ((operand_equal_for_phi_arg_p (arg_false, smaller)
1856 || (alt_smaller
1857 && operand_equal_for_phi_arg_p (arg_false, alt_smaller)))
1858 && (operand_equal_for_phi_arg_p (arg_true, larger)
1859 || (alt_larger
1860 && operand_equal_for_phi_arg_p (arg_true, alt_larger))))
1861 minmax = MAX_EXPR;
1862 else
1863 return false;
1865 else if (HONOR_NANS (type) || HONOR_SIGNED_ZEROS (type))
1866 /* The optimization may be unsafe due to NaNs. */
1867 return false;
1868 else if (middle_bb != alt_middle_bb && threeway_p)
1870 /* Recognize the following case:
1872 if (smaller < larger)
1873 a = MIN (smaller, c);
1874 else
1875 b = MIN (larger, c);
1876 x = PHI <a, b>
1878 This is equivalent to
1880 a = MIN (smaller, c);
1881 x = MIN (larger, a); */
1883 gimple *assign = last_and_only_stmt (middle_bb);
1884 tree lhs, op0, op1, bound;
1885 tree alt_lhs, alt_op0, alt_op1;
1886 bool invert = false;
1888 /* When THREEWAY_P then e1 will point to the edge of the final transition
1889 from middle-bb to end. */
1890 if (true_edge == e0)
1891 gcc_assert (false_edge == EDGE_PRED (e1->src, 0));
1892 else
1893 gcc_assert (true_edge == EDGE_PRED (e1->src, 0));
1895 bool valid_minmax_p = false;
1896 gimple_stmt_iterator it1
1897 = gsi_start_nondebug_after_labels_bb (middle_bb);
1898 gimple_stmt_iterator it2
1899 = gsi_start_nondebug_after_labels_bb (alt_middle_bb);
1900 if (gsi_one_nondebug_before_end_p (it1)
1901 && gsi_one_nondebug_before_end_p (it2))
1903 gimple *stmt1 = gsi_stmt (it1);
1904 gimple *stmt2 = gsi_stmt (it2);
1905 if (is_gimple_assign (stmt1) && is_gimple_assign (stmt2))
1907 enum tree_code code1 = gimple_assign_rhs_code (stmt1);
1908 enum tree_code code2 = gimple_assign_rhs_code (stmt2);
1909 valid_minmax_p = (code1 == MIN_EXPR || code1 == MAX_EXPR)
1910 && (code2 == MIN_EXPR || code2 == MAX_EXPR);
1914 if (!valid_minmax_p)
1915 return false;
1917 if (!assign
1918 || gimple_code (assign) != GIMPLE_ASSIGN)
1919 return false;
1921 lhs = gimple_assign_lhs (assign);
1922 ass_code = gimple_assign_rhs_code (assign);
1923 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
1924 return false;
1926 op0 = gimple_assign_rhs1 (assign);
1927 op1 = gimple_assign_rhs2 (assign);
1929 assign = last_and_only_stmt (alt_middle_bb);
1930 if (!assign
1931 || gimple_code (assign) != GIMPLE_ASSIGN)
1932 return false;
1934 alt_lhs = gimple_assign_lhs (assign);
1935 if (ass_code != gimple_assign_rhs_code (assign))
1936 return false;
1938 if (!operand_equal_for_phi_arg_p (lhs, arg_true)
1939 || !operand_equal_for_phi_arg_p (alt_lhs, arg_false))
1940 return false;
1942 alt_op0 = gimple_assign_rhs1 (assign);
1943 alt_op1 = gimple_assign_rhs2 (assign);
1945 if ((operand_equal_for_phi_arg_p (op0, smaller)
1946 || (alt_smaller
1947 && operand_equal_for_phi_arg_p (op0, alt_smaller)))
1948 && (operand_equal_for_phi_arg_p (alt_op0, larger)
1949 || (alt_larger
1950 && operand_equal_for_phi_arg_p (alt_op0, alt_larger))))
1952 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1953 if (!operand_equal_for_phi_arg_p (op1, alt_op1))
1954 return false;
1956 if ((arg0 = strip_bit_not (op0)) != NULL
1957 && (arg1 = strip_bit_not (alt_op0)) != NULL
1958 && (bound = strip_bit_not (op1)) != NULL)
1960 minmax = MAX_EXPR;
1961 ass_code = invert_minmax_code (ass_code);
1962 invert = true;
1964 else
1966 bound = op1;
1967 minmax = MIN_EXPR;
1968 arg0 = op0;
1969 arg1 = alt_op0;
1972 else if ((operand_equal_for_phi_arg_p (op0, larger)
1973 || (alt_larger
1974 && operand_equal_for_phi_arg_p (op0, alt_larger)))
1975 && (operand_equal_for_phi_arg_p (alt_op0, smaller)
1976 || (alt_smaller
1977 && operand_equal_for_phi_arg_p (alt_op0, alt_smaller))))
1979 /* We got here if the condition is true, i.e., SMALLER > LARGER. */
1980 if (!operand_equal_for_phi_arg_p (op1, alt_op1))
1981 return false;
1983 if ((arg0 = strip_bit_not (op0)) != NULL
1984 && (arg1 = strip_bit_not (alt_op0)) != NULL
1985 && (bound = strip_bit_not (op1)) != NULL)
1987 minmax = MIN_EXPR;
1988 ass_code = invert_minmax_code (ass_code);
1989 invert = true;
1991 else
1993 bound = op1;
1994 minmax = MAX_EXPR;
1995 arg0 = op0;
1996 arg1 = alt_op0;
1999 else
2000 return false;
2002 /* Emit the statement to compute min/max. */
2003 location_t locus = gimple_location (last_nondebug_stmt (cond_bb));
2004 gimple_seq stmts = NULL;
2005 tree phi_result = PHI_RESULT (phi);
2006 result = gimple_build (&stmts, locus, minmax, TREE_TYPE (phi_result),
2007 arg0, arg1);
2008 result = gimple_build (&stmts, locus, ass_code, TREE_TYPE (phi_result),
2009 result, bound);
2010 if (invert)
2011 result = gimple_build (&stmts, locus, BIT_NOT_EXPR, TREE_TYPE (phi_result),
2012 result);
2014 gsi = gsi_last_bb (cond_bb);
2015 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2017 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
2019 return true;
2021 else if (!threeway_p
2022 || empty_block_p (alt_middle_bb))
2024 /* Recognize the following case, assuming d <= u:
2026 if (a <= u)
2027 b = MAX (a, d);
2028 x = PHI <b, u>
2030 This is equivalent to
2032 b = MAX (a, d);
2033 x = MIN (b, u); */
2035 gimple *assign = last_and_only_stmt (middle_bb);
2036 tree lhs, op0, op1, bound;
2038 if (!single_pred_p (middle_bb))
2039 return false;
2041 if (!assign
2042 || gimple_code (assign) != GIMPLE_ASSIGN)
2043 return false;
2045 lhs = gimple_assign_lhs (assign);
2046 ass_code = gimple_assign_rhs_code (assign);
2047 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
2048 return false;
2049 op0 = gimple_assign_rhs1 (assign);
2050 op1 = gimple_assign_rhs2 (assign);
2052 if (true_edge->src == middle_bb)
2054 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
2055 if (!operand_equal_for_phi_arg_p (lhs, arg_true))
2056 return false;
2058 if (operand_equal_for_phi_arg_p (arg_false, larger)
2059 || (alt_larger
2060 && operand_equal_for_phi_arg_p (arg_false, alt_larger)))
2062 /* Case
2064 if (smaller < larger)
2066 r' = MAX_EXPR (smaller, bound)
2068 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
2069 if (ass_code != MAX_EXPR)
2070 return false;
2072 minmax = MIN_EXPR;
2073 if (operand_equal_for_phi_arg_p (op0, smaller)
2074 || (alt_smaller
2075 && operand_equal_for_phi_arg_p (op0, alt_smaller)))
2076 bound = op1;
2077 else if (operand_equal_for_phi_arg_p (op1, smaller)
2078 || (alt_smaller
2079 && operand_equal_for_phi_arg_p (op1, alt_smaller)))
2080 bound = op0;
2081 else
2082 return false;
2084 /* We need BOUND <= LARGER. */
2085 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
2086 bound, arg_false)))
2087 return false;
2089 else if (operand_equal_for_phi_arg_p (arg_false, smaller)
2090 || (alt_smaller
2091 && operand_equal_for_phi_arg_p (arg_false, alt_smaller)))
2093 /* Case
2095 if (smaller < larger)
2097 r' = MIN_EXPR (larger, bound)
2099 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
2100 if (ass_code != MIN_EXPR)
2101 return false;
2103 minmax = MAX_EXPR;
2104 if (operand_equal_for_phi_arg_p (op0, larger)
2105 || (alt_larger
2106 && operand_equal_for_phi_arg_p (op0, alt_larger)))
2107 bound = op1;
2108 else if (operand_equal_for_phi_arg_p (op1, larger)
2109 || (alt_larger
2110 && operand_equal_for_phi_arg_p (op1, alt_larger)))
2111 bound = op0;
2112 else
2113 return false;
2115 /* We need BOUND >= SMALLER. */
2116 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
2117 bound, arg_false)))
2118 return false;
2120 else
2121 return false;
2123 else
2125 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
2126 if (!operand_equal_for_phi_arg_p (lhs, arg_false))
2127 return false;
2129 if (operand_equal_for_phi_arg_p (arg_true, larger)
2130 || (alt_larger
2131 && operand_equal_for_phi_arg_p (arg_true, alt_larger)))
2133 /* Case
2135 if (smaller > larger)
2137 r' = MIN_EXPR (smaller, bound)
2139 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
2140 if (ass_code != MIN_EXPR)
2141 return false;
2143 minmax = MAX_EXPR;
2144 if (operand_equal_for_phi_arg_p (op0, smaller)
2145 || (alt_smaller
2146 && operand_equal_for_phi_arg_p (op0, alt_smaller)))
2147 bound = op1;
2148 else if (operand_equal_for_phi_arg_p (op1, smaller)
2149 || (alt_smaller
2150 && operand_equal_for_phi_arg_p (op1, alt_smaller)))
2151 bound = op0;
2152 else
2153 return false;
2155 /* We need BOUND >= LARGER. */
2156 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
2157 bound, arg_true)))
2158 return false;
2160 else if (operand_equal_for_phi_arg_p (arg_true, smaller)
2161 || (alt_smaller
2162 && operand_equal_for_phi_arg_p (arg_true, alt_smaller)))
2164 /* Case
2166 if (smaller > larger)
2168 r' = MAX_EXPR (larger, bound)
2170 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
2171 if (ass_code != MAX_EXPR)
2172 return false;
2174 minmax = MIN_EXPR;
2175 if (operand_equal_for_phi_arg_p (op0, larger))
2176 bound = op1;
2177 else if (operand_equal_for_phi_arg_p (op1, larger))
2178 bound = op0;
2179 else
2180 return false;
2182 /* We need BOUND <= SMALLER. */
2183 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
2184 bound, arg_true)))
2185 return false;
2187 else
2188 return false;
2191 /* Move the statement from the middle block. */
2192 gsi = gsi_last_bb (cond_bb);
2193 gsi_from = gsi_last_nondebug_bb (middle_bb);
2194 reset_flow_sensitive_info (SINGLE_SSA_TREE_OPERAND (gsi_stmt (gsi_from),
2195 SSA_OP_DEF));
2196 gsi_move_before (&gsi_from, &gsi);
2198 else
2199 return false;
2201 /* Emit the statement to compute min/max. */
2202 gimple_seq stmts = NULL;
2203 tree phi_result = PHI_RESULT (phi);
2205 /* When we can't use a MIN/MAX_EXPR still make sure the expression
2206 stays in a form to be recognized by ISA that map to IEEE x > y ? x : y
2207 semantics (that's not IEEE max semantics). */
2208 if (HONOR_NANS (type) || HONOR_SIGNED_ZEROS (type))
2210 result = gimple_build (&stmts, cmp, boolean_type_node,
2211 gimple_cond_lhs (cond), rhs);
2212 result = gimple_build (&stmts, COND_EXPR, TREE_TYPE (phi_result),
2213 result, arg_true, arg_false);
2215 else
2216 result = gimple_build (&stmts, minmax, TREE_TYPE (phi_result), arg0, arg1);
2218 gsi = gsi_last_bb (cond_bb);
2219 gsi_insert_seq_before (&gsi, stmts, GSI_NEW_STMT);
2221 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
2223 return true;
2226 /* Attempt to optimize (x <=> y) cmp 0 and similar comparisons.
2227 For strong ordering <=> try to match something like:
2228 <bb 2> : // cond3_bb (== cond2_bb)
2229 if (x_4(D) != y_5(D))
2230 goto <bb 3>; [INV]
2231 else
2232 goto <bb 6>; [INV]
2234 <bb 3> : // cond_bb
2235 if (x_4(D) < y_5(D))
2236 goto <bb 6>; [INV]
2237 else
2238 goto <bb 4>; [INV]
2240 <bb 4> : // middle_bb
2242 <bb 6> : // phi_bb
2243 # iftmp.0_2 = PHI <1(4), 0(2), -1(3)>
2244 _1 = iftmp.0_2 == 0;
2246 and for partial ordering <=> something like:
2248 <bb 2> : // cond3_bb
2249 if (a_3(D) == b_5(D))
2250 goto <bb 6>; [50.00%]
2251 else
2252 goto <bb 3>; [50.00%]
2254 <bb 3> [local count: 536870913]: // cond2_bb
2255 if (a_3(D) < b_5(D))
2256 goto <bb 6>; [50.00%]
2257 else
2258 goto <bb 4>; [50.00%]
2260 <bb 4> [local count: 268435456]: // cond_bb
2261 if (a_3(D) > b_5(D))
2262 goto <bb 6>; [50.00%]
2263 else
2264 goto <bb 5>; [50.00%]
2266 <bb 5> [local count: 134217728]: // middle_bb
2268 <bb 6> [local count: 1073741824]: // phi_bb
2269 # SR.27_4 = PHI <0(2), -1(3), 1(4), 2(5)>
2270 _2 = SR.27_4 > 0; */
2272 static bool
2273 spaceship_replacement (basic_block cond_bb, basic_block middle_bb,
2274 edge e0, edge e1, gphi *phi,
2275 tree arg0, tree arg1)
2277 tree phires = PHI_RESULT (phi);
2278 if (!INTEGRAL_TYPE_P (TREE_TYPE (phires))
2279 || TYPE_UNSIGNED (TREE_TYPE (phires))
2280 || !tree_fits_shwi_p (arg0)
2281 || !tree_fits_shwi_p (arg1)
2282 || !IN_RANGE (tree_to_shwi (arg0), -1, 2)
2283 || !IN_RANGE (tree_to_shwi (arg1), -1, 2))
2284 return false;
2286 basic_block phi_bb = gimple_bb (phi);
2287 gcc_assert (phi_bb == e0->dest && phi_bb == e1->dest);
2288 if (!IN_RANGE (EDGE_COUNT (phi_bb->preds), 3, 4))
2289 return false;
2291 use_operand_p use_p;
2292 gimple *use_stmt;
2293 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (phires))
2294 return false;
2295 if (!single_imm_use (phires, &use_p, &use_stmt))
2296 return false;
2297 enum tree_code cmp;
2298 tree lhs, rhs;
2299 gimple *orig_use_stmt = use_stmt;
2300 tree orig_use_lhs = NULL_TREE;
2301 int prec = TYPE_PRECISION (TREE_TYPE (phires));
2302 bool is_cast = false;
2304 /* Deal with the case when match.pd has rewritten the (res & ~1) == 0
2305 into res <= 1 and has left a type-cast for signed types. */
2306 if (gimple_assign_cast_p (use_stmt))
2308 orig_use_lhs = gimple_assign_lhs (use_stmt);
2309 /* match.pd would have only done this for a signed type,
2310 so the conversion must be to an unsigned one. */
2311 tree ty1 = TREE_TYPE (gimple_assign_rhs1 (use_stmt));
2312 tree ty2 = TREE_TYPE (orig_use_lhs);
2314 if (!TYPE_UNSIGNED (ty2) || !INTEGRAL_TYPE_P (ty2))
2315 return false;
2316 if (TYPE_PRECISION (ty1) > TYPE_PRECISION (ty2))
2317 return false;
2318 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (orig_use_lhs))
2319 return false;
2320 if (!single_imm_use (orig_use_lhs, &use_p, &use_stmt))
2321 return false;
2323 is_cast = true;
2325 else if (is_gimple_assign (use_stmt)
2326 && gimple_assign_rhs_code (use_stmt) == BIT_AND_EXPR
2327 && TREE_CODE (gimple_assign_rhs2 (use_stmt)) == INTEGER_CST
2328 && (wi::to_wide (gimple_assign_rhs2 (use_stmt))
2329 == wi::shifted_mask (1, prec - 1, false, prec)))
2331 /* For partial_ordering result operator>= with unspec as second
2332 argument is (res & 1) == res, folded by match.pd into
2333 (res & ~1) == 0. */
2334 orig_use_lhs = gimple_assign_lhs (use_stmt);
2335 if (SSA_NAME_OCCURS_IN_ABNORMAL_PHI (orig_use_lhs))
2336 return false;
2337 if (!single_imm_use (orig_use_lhs, &use_p, &use_stmt))
2338 return false;
2340 if (gimple_code (use_stmt) == GIMPLE_COND)
2342 cmp = gimple_cond_code (use_stmt);
2343 lhs = gimple_cond_lhs (use_stmt);
2344 rhs = gimple_cond_rhs (use_stmt);
2346 else if (is_gimple_assign (use_stmt))
2348 if (gimple_assign_rhs_class (use_stmt) == GIMPLE_BINARY_RHS)
2350 cmp = gimple_assign_rhs_code (use_stmt);
2351 lhs = gimple_assign_rhs1 (use_stmt);
2352 rhs = gimple_assign_rhs2 (use_stmt);
2354 else if (gimple_assign_rhs_code (use_stmt) == COND_EXPR)
2356 tree cond = gimple_assign_rhs1 (use_stmt);
2357 if (!COMPARISON_CLASS_P (cond))
2358 return false;
2359 cmp = TREE_CODE (cond);
2360 lhs = TREE_OPERAND (cond, 0);
2361 rhs = TREE_OPERAND (cond, 1);
2363 else
2364 return false;
2366 else
2367 return false;
2368 switch (cmp)
2370 case EQ_EXPR:
2371 case NE_EXPR:
2372 case LT_EXPR:
2373 case GT_EXPR:
2374 case LE_EXPR:
2375 case GE_EXPR:
2376 break;
2377 default:
2378 return false;
2380 if (lhs != (orig_use_lhs ? orig_use_lhs : phires)
2381 || !tree_fits_shwi_p (rhs)
2382 || !IN_RANGE (tree_to_shwi (rhs), -1, 1))
2383 return false;
2385 if (is_cast)
2387 if (TREE_CODE (rhs) != INTEGER_CST)
2388 return false;
2389 /* As for -ffast-math we assume the 2 return to be
2390 impossible, canonicalize (unsigned) res <= 1U or
2391 (unsigned) res < 2U into res >= 0 and (unsigned) res > 1U
2392 or (unsigned) res >= 2U as res < 0. */
2393 switch (cmp)
2395 case LE_EXPR:
2396 if (!integer_onep (rhs))
2397 return false;
2398 cmp = GE_EXPR;
2399 break;
2400 case LT_EXPR:
2401 if (wi::ne_p (wi::to_widest (rhs), 2))
2402 return false;
2403 cmp = GE_EXPR;
2404 break;
2405 case GT_EXPR:
2406 if (!integer_onep (rhs))
2407 return false;
2408 cmp = LT_EXPR;
2409 break;
2410 case GE_EXPR:
2411 if (wi::ne_p (wi::to_widest (rhs), 2))
2412 return false;
2413 cmp = LT_EXPR;
2414 break;
2415 default:
2416 return false;
2418 rhs = build_zero_cst (TREE_TYPE (phires));
2420 else if (orig_use_lhs)
2422 if ((cmp != EQ_EXPR && cmp != NE_EXPR) || !integer_zerop (rhs))
2423 return false;
2424 /* As for -ffast-math we assume the 2 return to be
2425 impossible, canonicalize (res & ~1) == 0 into
2426 res >= 0 and (res & ~1) != 0 as res < 0. */
2427 cmp = cmp == EQ_EXPR ? GE_EXPR : LT_EXPR;
2430 if (!empty_block_p (middle_bb))
2431 return false;
2433 gcond *cond1 = as_a <gcond *> (*gsi_last_bb (cond_bb));
2434 enum tree_code cmp1 = gimple_cond_code (cond1);
2435 switch (cmp1)
2437 case LT_EXPR:
2438 case LE_EXPR:
2439 case GT_EXPR:
2440 case GE_EXPR:
2441 break;
2442 default:
2443 return false;
2445 tree lhs1 = gimple_cond_lhs (cond1);
2446 tree rhs1 = gimple_cond_rhs (cond1);
2447 /* The optimization may be unsafe due to NaNs. */
2448 if (HONOR_NANS (TREE_TYPE (lhs1)))
2449 return false;
2450 if (TREE_CODE (lhs1) == SSA_NAME && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs1))
2451 return false;
2452 if (TREE_CODE (rhs1) == SSA_NAME && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (rhs1))
2453 return false;
2455 if (!single_pred_p (cond_bb) || !cond_only_block_p (cond_bb))
2456 return false;
2458 basic_block cond2_bb = single_pred (cond_bb);
2459 if (EDGE_COUNT (cond2_bb->succs) != 2)
2460 return false;
2461 edge cond2_phi_edge;
2462 if (EDGE_SUCC (cond2_bb, 0)->dest == cond_bb)
2464 if (EDGE_SUCC (cond2_bb, 1)->dest != phi_bb)
2465 return false;
2466 cond2_phi_edge = EDGE_SUCC (cond2_bb, 1);
2468 else if (EDGE_SUCC (cond2_bb, 0)->dest != phi_bb)
2469 return false;
2470 else
2471 cond2_phi_edge = EDGE_SUCC (cond2_bb, 0);
2472 tree arg2 = gimple_phi_arg_def (phi, cond2_phi_edge->dest_idx);
2473 if (!tree_fits_shwi_p (arg2))
2474 return false;
2475 gcond *cond2 = safe_dyn_cast <gcond *> (*gsi_last_bb (cond2_bb));
2476 if (!cond2)
2477 return false;
2478 enum tree_code cmp2 = gimple_cond_code (cond2);
2479 tree lhs2 = gimple_cond_lhs (cond2);
2480 tree rhs2 = gimple_cond_rhs (cond2);
2481 if (lhs2 == lhs1)
2483 if (!operand_equal_p (rhs2, rhs1, 0))
2485 if ((cmp2 == EQ_EXPR || cmp2 == NE_EXPR)
2486 && TREE_CODE (rhs1) == INTEGER_CST
2487 && TREE_CODE (rhs2) == INTEGER_CST)
2489 /* For integers, we can have cond2 x == 5
2490 and cond1 x < 5, x <= 4, x <= 5, x < 6,
2491 x > 5, x >= 6, x >= 5 or x > 4. */
2492 if (tree_int_cst_lt (rhs1, rhs2))
2494 if (wi::ne_p (wi::to_wide (rhs1) + 1, wi::to_wide (rhs2)))
2495 return false;
2496 if (cmp1 == LE_EXPR)
2497 cmp1 = LT_EXPR;
2498 else if (cmp1 == GT_EXPR)
2499 cmp1 = GE_EXPR;
2500 else
2501 return false;
2503 else
2505 gcc_checking_assert (tree_int_cst_lt (rhs2, rhs1));
2506 if (wi::ne_p (wi::to_wide (rhs2) + 1, wi::to_wide (rhs1)))
2507 return false;
2508 if (cmp1 == LT_EXPR)
2509 cmp1 = LE_EXPR;
2510 else if (cmp1 == GE_EXPR)
2511 cmp1 = GT_EXPR;
2512 else
2513 return false;
2515 rhs1 = rhs2;
2517 else
2518 return false;
2521 else if (lhs2 == rhs1)
2523 if (rhs2 != lhs1)
2524 return false;
2526 else
2527 return false;
2529 tree arg3 = arg2;
2530 basic_block cond3_bb = cond2_bb;
2531 edge cond3_phi_edge = cond2_phi_edge;
2532 gcond *cond3 = cond2;
2533 enum tree_code cmp3 = cmp2;
2534 tree lhs3 = lhs2;
2535 tree rhs3 = rhs2;
2536 if (EDGE_COUNT (phi_bb->preds) == 4)
2538 if (absu_hwi (tree_to_shwi (arg2)) != 1)
2539 return false;
2540 if (e1->flags & EDGE_TRUE_VALUE)
2542 if (tree_to_shwi (arg0) != 2
2543 || absu_hwi (tree_to_shwi (arg1)) != 1
2544 || wi::to_widest (arg1) == wi::to_widest (arg2))
2545 return false;
2547 else if (tree_to_shwi (arg1) != 2
2548 || absu_hwi (tree_to_shwi (arg0)) != 1
2549 || wi::to_widest (arg0) == wi::to_widest (arg1))
2550 return false;
2551 switch (cmp2)
2553 case LT_EXPR:
2554 case LE_EXPR:
2555 case GT_EXPR:
2556 case GE_EXPR:
2557 break;
2558 default:
2559 return false;
2561 /* if (x < y) goto phi_bb; else fallthru;
2562 if (x > y) goto phi_bb; else fallthru;
2563 bbx:;
2564 phi_bb:;
2565 is ok, but if x and y are swapped in one of the comparisons,
2566 or the comparisons are the same and operands not swapped,
2567 or the true and false edges are swapped, it is not. */
2568 if ((lhs2 == lhs1)
2569 ^ (((cond2_phi_edge->flags
2570 & ((cmp2 == LT_EXPR || cmp2 == LE_EXPR)
2571 ? EDGE_TRUE_VALUE : EDGE_FALSE_VALUE)) != 0)
2572 != ((e1->flags
2573 & ((cmp1 == LT_EXPR || cmp1 == LE_EXPR)
2574 ? EDGE_TRUE_VALUE : EDGE_FALSE_VALUE)) != 0)))
2575 return false;
2576 if (!single_pred_p (cond2_bb) || !cond_only_block_p (cond2_bb))
2577 return false;
2578 cond3_bb = single_pred (cond2_bb);
2579 if (EDGE_COUNT (cond2_bb->succs) != 2)
2580 return false;
2581 if (EDGE_SUCC (cond3_bb, 0)->dest == cond2_bb)
2583 if (EDGE_SUCC (cond3_bb, 1)->dest != phi_bb)
2584 return false;
2585 cond3_phi_edge = EDGE_SUCC (cond3_bb, 1);
2587 else if (EDGE_SUCC (cond3_bb, 0)->dest != phi_bb)
2588 return false;
2589 else
2590 cond3_phi_edge = EDGE_SUCC (cond3_bb, 0);
2591 arg3 = gimple_phi_arg_def (phi, cond3_phi_edge->dest_idx);
2592 cond3 = safe_dyn_cast <gcond *> (*gsi_last_bb (cond3_bb));
2593 if (!cond3)
2594 return false;
2595 cmp3 = gimple_cond_code (cond3);
2596 lhs3 = gimple_cond_lhs (cond3);
2597 rhs3 = gimple_cond_rhs (cond3);
2598 if (lhs3 == lhs1)
2600 if (!operand_equal_p (rhs3, rhs1, 0))
2601 return false;
2603 else if (lhs3 == rhs1)
2605 if (rhs3 != lhs1)
2606 return false;
2608 else
2609 return false;
2611 else if (absu_hwi (tree_to_shwi (arg0)) != 1
2612 || absu_hwi (tree_to_shwi (arg1)) != 1
2613 || wi::to_widest (arg0) == wi::to_widest (arg1))
2614 return false;
2616 if (!integer_zerop (arg3) || (cmp3 != EQ_EXPR && cmp3 != NE_EXPR))
2617 return false;
2618 if ((cond3_phi_edge->flags & (cmp3 == EQ_EXPR
2619 ? EDGE_TRUE_VALUE : EDGE_FALSE_VALUE)) == 0)
2620 return false;
2622 /* lhs1 one_cmp rhs1 results in phires of 1. */
2623 enum tree_code one_cmp;
2624 if ((cmp1 == LT_EXPR || cmp1 == LE_EXPR)
2625 ^ (!integer_onep ((e1->flags & EDGE_TRUE_VALUE) ? arg1 : arg0)))
2626 one_cmp = LT_EXPR;
2627 else
2628 one_cmp = GT_EXPR;
2630 enum tree_code res_cmp;
2631 switch (cmp)
2633 case EQ_EXPR:
2634 if (integer_zerop (rhs))
2635 res_cmp = EQ_EXPR;
2636 else if (integer_minus_onep (rhs))
2637 res_cmp = one_cmp == LT_EXPR ? GT_EXPR : LT_EXPR;
2638 else if (integer_onep (rhs))
2639 res_cmp = one_cmp;
2640 else
2641 return false;
2642 break;
2643 case NE_EXPR:
2644 if (integer_zerop (rhs))
2645 res_cmp = NE_EXPR;
2646 else if (integer_minus_onep (rhs))
2647 res_cmp = one_cmp == LT_EXPR ? LE_EXPR : GE_EXPR;
2648 else if (integer_onep (rhs))
2649 res_cmp = one_cmp == LT_EXPR ? GE_EXPR : LE_EXPR;
2650 else
2651 return false;
2652 break;
2653 case LT_EXPR:
2654 if (integer_onep (rhs))
2655 res_cmp = one_cmp == LT_EXPR ? GE_EXPR : LE_EXPR;
2656 else if (integer_zerop (rhs))
2657 res_cmp = one_cmp == LT_EXPR ? GT_EXPR : LT_EXPR;
2658 else
2659 return false;
2660 break;
2661 case LE_EXPR:
2662 if (integer_zerop (rhs))
2663 res_cmp = one_cmp == LT_EXPR ? GE_EXPR : LE_EXPR;
2664 else if (integer_minus_onep (rhs))
2665 res_cmp = one_cmp == LT_EXPR ? GT_EXPR : LT_EXPR;
2666 else
2667 return false;
2668 break;
2669 case GT_EXPR:
2670 if (integer_minus_onep (rhs))
2671 res_cmp = one_cmp == LT_EXPR ? LE_EXPR : GE_EXPR;
2672 else if (integer_zerop (rhs))
2673 res_cmp = one_cmp;
2674 else
2675 return false;
2676 break;
2677 case GE_EXPR:
2678 if (integer_zerop (rhs))
2679 res_cmp = one_cmp == LT_EXPR ? LE_EXPR : GE_EXPR;
2680 else if (integer_onep (rhs))
2681 res_cmp = one_cmp;
2682 else
2683 return false;
2684 break;
2685 default:
2686 gcc_unreachable ();
2689 if (gimple_code (use_stmt) == GIMPLE_COND)
2691 gcond *use_cond = as_a <gcond *> (use_stmt);
2692 gimple_cond_set_code (use_cond, res_cmp);
2693 gimple_cond_set_lhs (use_cond, lhs1);
2694 gimple_cond_set_rhs (use_cond, rhs1);
2696 else if (gimple_assign_rhs_class (use_stmt) == GIMPLE_BINARY_RHS)
2698 gimple_assign_set_rhs_code (use_stmt, res_cmp);
2699 gimple_assign_set_rhs1 (use_stmt, lhs1);
2700 gimple_assign_set_rhs2 (use_stmt, rhs1);
2702 else
2704 tree cond = build2 (res_cmp, TREE_TYPE (gimple_assign_rhs1 (use_stmt)),
2705 lhs1, rhs1);
2706 gimple_assign_set_rhs1 (use_stmt, cond);
2708 update_stmt (use_stmt);
2710 if (MAY_HAVE_DEBUG_BIND_STMTS)
2712 use_operand_p use_p;
2713 imm_use_iterator iter;
2714 bool has_debug_uses = false;
2715 bool has_cast_debug_uses = false;
2716 FOR_EACH_IMM_USE_FAST (use_p, iter, phires)
2718 gimple *use_stmt = USE_STMT (use_p);
2719 if (orig_use_lhs && use_stmt == orig_use_stmt)
2720 continue;
2721 gcc_assert (is_gimple_debug (use_stmt));
2722 has_debug_uses = true;
2723 break;
2725 if (orig_use_lhs)
2727 if (!has_debug_uses || is_cast)
2728 FOR_EACH_IMM_USE_FAST (use_p, iter, orig_use_lhs)
2730 gimple *use_stmt = USE_STMT (use_p);
2731 gcc_assert (is_gimple_debug (use_stmt));
2732 has_debug_uses = true;
2733 if (is_cast)
2734 has_cast_debug_uses = true;
2736 gimple_stmt_iterator gsi = gsi_for_stmt (orig_use_stmt);
2737 tree zero = build_zero_cst (TREE_TYPE (orig_use_lhs));
2738 gimple_assign_set_rhs_with_ops (&gsi, INTEGER_CST, zero);
2739 update_stmt (orig_use_stmt);
2742 if (has_debug_uses)
2744 /* If there are debug uses, emit something like:
2745 # DEBUG D#1 => i_2(D) > j_3(D) ? 1 : -1
2746 # DEBUG D#2 => i_2(D) == j_3(D) ? 0 : D#1
2747 where > stands for the comparison that yielded 1
2748 and replace debug uses of phi result with that D#2.
2749 Ignore the value of 2, because if NaNs aren't expected,
2750 all floating point numbers should be comparable. */
2751 gimple_stmt_iterator gsi = gsi_after_labels (gimple_bb (phi));
2752 tree type = TREE_TYPE (phires);
2753 tree temp1 = build_debug_expr_decl (type);
2754 tree t = build2 (one_cmp, boolean_type_node, lhs1, rhs2);
2755 t = build3 (COND_EXPR, type, t, build_one_cst (type),
2756 build_int_cst (type, -1));
2757 gimple *g = gimple_build_debug_bind (temp1, t, phi);
2758 gsi_insert_before (&gsi, g, GSI_SAME_STMT);
2759 tree temp2 = build_debug_expr_decl (type);
2760 t = build2 (EQ_EXPR, boolean_type_node, lhs1, rhs2);
2761 t = build3 (COND_EXPR, type, t, build_zero_cst (type), temp1);
2762 g = gimple_build_debug_bind (temp2, t, phi);
2763 gsi_insert_before (&gsi, g, GSI_SAME_STMT);
2764 replace_uses_by (phires, temp2);
2765 if (orig_use_lhs)
2767 if (has_cast_debug_uses)
2769 tree temp3 = make_node (DEBUG_EXPR_DECL);
2770 DECL_ARTIFICIAL (temp3) = 1;
2771 TREE_TYPE (temp3) = TREE_TYPE (orig_use_lhs);
2772 SET_DECL_MODE (temp3, TYPE_MODE (type));
2773 t = fold_convert (TREE_TYPE (temp3), temp2);
2774 g = gimple_build_debug_bind (temp3, t, phi);
2775 gsi_insert_before (&gsi, g, GSI_SAME_STMT);
2776 replace_uses_by (orig_use_lhs, temp3);
2778 else
2779 replace_uses_by (orig_use_lhs, temp2);
2784 if (orig_use_lhs)
2786 gimple_stmt_iterator gsi = gsi_for_stmt (orig_use_stmt);
2787 gsi_remove (&gsi, true);
2790 gimple_stmt_iterator psi = gsi_for_stmt (phi);
2791 remove_phi_node (&psi, true);
2792 statistics_counter_event (cfun, "spaceship replacement", 1);
2794 return true;
2797 /* Optimize x ? __builtin_fun (x) : C, where C is __builtin_fun (0).
2798 Convert
2800 <bb 2>
2801 if (b_4(D) != 0)
2802 goto <bb 3>
2803 else
2804 goto <bb 4>
2806 <bb 3>
2807 _2 = (unsigned long) b_4(D);
2808 _9 = __builtin_popcountl (_2);
2810 _9 = __builtin_popcountl (b_4(D));
2812 <bb 4>
2813 c_12 = PHI <0(2), _9(3)>
2815 Into
2816 <bb 2>
2817 _2 = (unsigned long) b_4(D);
2818 _9 = __builtin_popcountl (_2);
2820 _9 = __builtin_popcountl (b_4(D));
2822 <bb 4>
2823 c_12 = PHI <_9(2)>
2825 Similarly for __builtin_clz or __builtin_ctz if
2826 C?Z_DEFINED_VALUE_AT_ZERO is 2, optab is present and
2827 instead of 0 above it uses the value from that macro. */
2829 static bool
2830 cond_removal_in_builtin_zero_pattern (basic_block cond_bb,
2831 basic_block middle_bb,
2832 edge e1, edge e2, gphi *phi,
2833 tree arg0, tree arg1)
2835 gimple_stmt_iterator gsi, gsi_from;
2836 gimple *call;
2837 gimple *cast = NULL;
2838 tree lhs, arg;
2840 /* Check that
2841 _2 = (unsigned long) b_4(D);
2842 _9 = __builtin_popcountl (_2);
2844 _9 = __builtin_popcountl (b_4(D));
2845 are the only stmts in the middle_bb. */
2847 gsi = gsi_start_nondebug_after_labels_bb (middle_bb);
2848 if (gsi_end_p (gsi))
2849 return false;
2850 cast = gsi_stmt (gsi);
2851 gsi_next_nondebug (&gsi);
2852 if (!gsi_end_p (gsi))
2854 call = gsi_stmt (gsi);
2855 gsi_next_nondebug (&gsi);
2856 if (!gsi_end_p (gsi))
2857 return false;
2859 else
2861 call = cast;
2862 cast = NULL;
2865 /* Check that we have a popcount/clz/ctz builtin. */
2866 if (!is_gimple_call (call))
2867 return false;
2869 lhs = gimple_get_lhs (call);
2871 if (lhs == NULL_TREE)
2872 return false;
2874 combined_fn cfn = gimple_call_combined_fn (call);
2875 if (gimple_call_num_args (call) != 1
2876 && (gimple_call_num_args (call) != 2
2877 || cfn == CFN_CLZ
2878 || cfn == CFN_CTZ))
2879 return false;
2881 arg = gimple_call_arg (call, 0);
2883 internal_fn ifn = IFN_LAST;
2884 int val = 0;
2885 bool any_val = false;
2886 switch (cfn)
2888 case CFN_BUILT_IN_BSWAP16:
2889 case CFN_BUILT_IN_BSWAP32:
2890 case CFN_BUILT_IN_BSWAP64:
2891 case CFN_BUILT_IN_BSWAP128:
2892 CASE_CFN_FFS:
2893 CASE_CFN_PARITY:
2894 CASE_CFN_POPCOUNT:
2895 break;
2896 CASE_CFN_CLZ:
2897 if (INTEGRAL_TYPE_P (TREE_TYPE (arg)))
2899 tree type = TREE_TYPE (arg);
2900 if (TREE_CODE (type) == BITINT_TYPE)
2902 if (gimple_call_num_args (call) == 1)
2904 any_val = true;
2905 ifn = IFN_CLZ;
2906 break;
2908 if (!tree_fits_shwi_p (gimple_call_arg (call, 1)))
2909 return false;
2910 HOST_WIDE_INT at_zero = tree_to_shwi (gimple_call_arg (call, 1));
2911 if ((int) at_zero != at_zero)
2912 return false;
2913 ifn = IFN_CLZ;
2914 val = at_zero;
2915 break;
2917 if (direct_internal_fn_supported_p (IFN_CLZ, type, OPTIMIZE_FOR_BOTH)
2918 && CLZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (type),
2919 val) == 2)
2921 ifn = IFN_CLZ;
2922 break;
2925 return false;
2926 CASE_CFN_CTZ:
2927 if (INTEGRAL_TYPE_P (TREE_TYPE (arg)))
2929 tree type = TREE_TYPE (arg);
2930 if (TREE_CODE (type) == BITINT_TYPE)
2932 if (gimple_call_num_args (call) == 1)
2934 any_val = true;
2935 ifn = IFN_CTZ;
2936 break;
2938 if (!tree_fits_shwi_p (gimple_call_arg (call, 1)))
2939 return false;
2940 HOST_WIDE_INT at_zero = tree_to_shwi (gimple_call_arg (call, 1));
2941 if ((int) at_zero != at_zero)
2942 return false;
2943 ifn = IFN_CTZ;
2944 val = at_zero;
2945 break;
2947 if (direct_internal_fn_supported_p (IFN_CTZ, type, OPTIMIZE_FOR_BOTH)
2948 && CTZ_DEFINED_VALUE_AT_ZERO (SCALAR_INT_TYPE_MODE (type),
2949 val) == 2)
2951 ifn = IFN_CTZ;
2952 break;
2955 return false;
2956 case CFN_BUILT_IN_CLRSB:
2957 val = TYPE_PRECISION (integer_type_node) - 1;
2958 break;
2959 case CFN_BUILT_IN_CLRSBL:
2960 val = TYPE_PRECISION (long_integer_type_node) - 1;
2961 break;
2962 case CFN_BUILT_IN_CLRSBLL:
2963 val = TYPE_PRECISION (long_long_integer_type_node) - 1;
2964 break;
2965 default:
2966 return false;
2969 if (cast)
2971 /* We have a cast stmt feeding popcount/clz/ctz builtin. */
2972 /* Check that we have a cast prior to that. */
2973 if (gimple_code (cast) != GIMPLE_ASSIGN
2974 || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (cast)))
2975 return false;
2976 /* Result of the cast stmt is the argument to the builtin. */
2977 if (arg != gimple_assign_lhs (cast))
2978 return false;
2979 arg = gimple_assign_rhs1 (cast);
2982 gcond *cond = dyn_cast <gcond *> (*gsi_last_bb (cond_bb));
2984 /* Cond_bb has a check for b_4 [!=|==] 0 before calling the popcount/clz/ctz
2985 builtin. */
2986 if (!cond
2987 || (gimple_cond_code (cond) != NE_EXPR
2988 && gimple_cond_code (cond) != EQ_EXPR)
2989 || !integer_zerop (gimple_cond_rhs (cond))
2990 || arg != gimple_cond_lhs (cond))
2991 return false;
2993 /* Canonicalize. */
2994 if ((e2->flags & EDGE_TRUE_VALUE
2995 && gimple_cond_code (cond) == NE_EXPR)
2996 || (e1->flags & EDGE_TRUE_VALUE
2997 && gimple_cond_code (cond) == EQ_EXPR))
2999 std::swap (arg0, arg1);
3000 std::swap (e1, e2);
3003 /* Check PHI arguments. */
3004 if (lhs != arg0
3005 || TREE_CODE (arg1) != INTEGER_CST)
3006 return false;
3007 if (any_val)
3009 if (!tree_fits_shwi_p (arg1))
3010 return false;
3011 HOST_WIDE_INT at_zero = tree_to_shwi (arg1);
3012 if ((int) at_zero != at_zero)
3013 return false;
3014 val = at_zero;
3016 else if (wi::to_wide (arg1) != val)
3017 return false;
3019 /* And insert the popcount/clz/ctz builtin and cast stmt before the
3020 cond_bb. */
3021 gsi = gsi_last_bb (cond_bb);
3022 if (cast)
3024 gsi_from = gsi_for_stmt (cast);
3025 gsi_move_before (&gsi_from, &gsi);
3026 reset_flow_sensitive_info (gimple_get_lhs (cast));
3028 gsi_from = gsi_for_stmt (call);
3029 if (ifn == IFN_LAST
3030 || (gimple_call_internal_p (call) && gimple_call_num_args (call) == 2))
3031 gsi_move_before (&gsi_from, &gsi);
3032 else
3034 /* For __builtin_c[lt]z* force .C[LT]Z ifn, because only
3035 the latter is well defined at zero. */
3036 call = gimple_build_call_internal (ifn, 2, gimple_call_arg (call, 0),
3037 build_int_cst (integer_type_node, val));
3038 gimple_call_set_lhs (call, lhs);
3039 gsi_insert_before (&gsi, call, GSI_SAME_STMT);
3040 gsi_remove (&gsi_from, true);
3042 reset_flow_sensitive_info (lhs);
3044 /* Now update the PHI and remove unneeded bbs. */
3045 replace_phi_edge_with_variable (cond_bb, e2, phi, lhs);
3046 return true;
3049 /* Auxiliary functions to determine the set of memory accesses which
3050 can't trap because they are preceded by accesses to the same memory
3051 portion. We do that for MEM_REFs, so we only need to track
3052 the SSA_NAME of the pointer indirectly referenced. The algorithm
3053 simply is a walk over all instructions in dominator order. When
3054 we see an MEM_REF we determine if we've already seen a same
3055 ref anywhere up to the root of the dominator tree. If we do the
3056 current access can't trap. If we don't see any dominating access
3057 the current access might trap, but might also make later accesses
3058 non-trapping, so we remember it. We need to be careful with loads
3059 or stores, for instance a load might not trap, while a store would,
3060 so if we see a dominating read access this doesn't mean that a later
3061 write access would not trap. Hence we also need to differentiate the
3062 type of access(es) seen.
3064 ??? We currently are very conservative and assume that a load might
3065 trap even if a store doesn't (write-only memory). This probably is
3066 overly conservative.
3068 We currently support a special case that for !TREE_ADDRESSABLE automatic
3069 variables, it could ignore whether something is a load or store because the
3070 local stack should be always writable. */
3072 /* A hash-table of references (MEM_REF/ARRAY_REF/COMPONENT_REF), and in which
3073 basic block an *_REF through it was seen, which would constitute a
3074 no-trap region for same accesses.
3076 Size is needed to support 2 MEM_REFs of different types, like
3077 MEM<double>(s_1) and MEM<long>(s_1), which would compare equal with
3078 OEP_ADDRESS_OF. */
3079 struct ref_to_bb
3081 tree exp;
3082 HOST_WIDE_INT size;
3083 unsigned int phase;
3084 basic_block bb;
3087 /* Hashtable helpers. */
3089 struct refs_hasher : free_ptr_hash<ref_to_bb>
3091 static inline hashval_t hash (const ref_to_bb *);
3092 static inline bool equal (const ref_to_bb *, const ref_to_bb *);
3095 /* Used for quick clearing of the hash-table when we see calls.
3096 Hash entries with phase < nt_call_phase are invalid. */
3097 static unsigned int nt_call_phase;
3099 /* The hash function. */
3101 inline hashval_t
3102 refs_hasher::hash (const ref_to_bb *n)
3104 inchash::hash hstate;
3105 inchash::add_expr (n->exp, hstate, OEP_ADDRESS_OF);
3106 hstate.add_hwi (n->size);
3107 return hstate.end ();
3110 /* The equality function of *P1 and *P2. */
3112 inline bool
3113 refs_hasher::equal (const ref_to_bb *n1, const ref_to_bb *n2)
3115 return operand_equal_p (n1->exp, n2->exp, OEP_ADDRESS_OF)
3116 && n1->size == n2->size;
3119 class nontrapping_dom_walker : public dom_walker
3121 public:
3122 nontrapping_dom_walker (cdi_direction direction, hash_set<tree> *ps)
3123 : dom_walker (direction), m_nontrapping (ps), m_seen_refs (128)
3126 edge before_dom_children (basic_block) final override;
3127 void after_dom_children (basic_block) final override;
3129 private:
3131 /* We see the expression EXP in basic block BB. If it's an interesting
3132 expression (an MEM_REF through an SSA_NAME) possibly insert the
3133 expression into the set NONTRAP or the hash table of seen expressions.
3134 STORE is true if this expression is on the LHS, otherwise it's on
3135 the RHS. */
3136 void add_or_mark_expr (basic_block, tree, bool);
3138 hash_set<tree> *m_nontrapping;
3140 /* The hash table for remembering what we've seen. */
3141 hash_table<refs_hasher> m_seen_refs;
3144 /* Called by walk_dominator_tree, when entering the block BB. */
3145 edge
3146 nontrapping_dom_walker::before_dom_children (basic_block bb)
3148 edge e;
3149 edge_iterator ei;
3150 gimple_stmt_iterator gsi;
3152 /* If we haven't seen all our predecessors, clear the hash-table. */
3153 FOR_EACH_EDGE (e, ei, bb->preds)
3154 if ((((size_t)e->src->aux) & 2) == 0)
3156 nt_call_phase++;
3157 break;
3160 /* Mark this BB as being on the path to dominator root and as visited. */
3161 bb->aux = (void*)(1 | 2);
3163 /* And walk the statements in order. */
3164 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
3166 gimple *stmt = gsi_stmt (gsi);
3168 if ((gimple_code (stmt) == GIMPLE_ASM && gimple_vdef (stmt))
3169 || (is_gimple_call (stmt)
3170 && (!nonfreeing_call_p (stmt) || !nonbarrier_call_p (stmt))))
3171 nt_call_phase++;
3172 else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt))
3174 add_or_mark_expr (bb, gimple_assign_lhs (stmt), true);
3175 add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), false);
3178 return NULL;
3181 /* Called by walk_dominator_tree, when basic block BB is exited. */
3182 void
3183 nontrapping_dom_walker::after_dom_children (basic_block bb)
3185 /* This BB isn't on the path to dominator root anymore. */
3186 bb->aux = (void*)2;
3189 /* We see the expression EXP in basic block BB. If it's an interesting
3190 expression of:
3191 1) MEM_REF
3192 2) ARRAY_REF
3193 3) COMPONENT_REF
3194 possibly insert the expression into the set NONTRAP or the hash table
3195 of seen expressions. STORE is true if this expression is on the LHS,
3196 otherwise it's on the RHS. */
3197 void
3198 nontrapping_dom_walker::add_or_mark_expr (basic_block bb, tree exp, bool store)
3200 HOST_WIDE_INT size;
3202 if ((TREE_CODE (exp) == MEM_REF || TREE_CODE (exp) == ARRAY_REF
3203 || TREE_CODE (exp) == COMPONENT_REF)
3204 && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0)
3206 struct ref_to_bb map;
3207 ref_to_bb **slot;
3208 struct ref_to_bb *r2bb;
3209 basic_block found_bb = 0;
3211 if (!store)
3213 tree base = get_base_address (exp);
3214 /* Only record a LOAD of a local variable without address-taken, as
3215 the local stack is always writable. This allows cselim on a STORE
3216 with a dominating LOAD. */
3217 if (!auto_var_p (base) || TREE_ADDRESSABLE (base))
3218 return;
3221 /* Try to find the last seen *_REF, which can trap. */
3222 map.exp = exp;
3223 map.size = size;
3224 slot = m_seen_refs.find_slot (&map, INSERT);
3225 r2bb = *slot;
3226 if (r2bb && r2bb->phase >= nt_call_phase)
3227 found_bb = r2bb->bb;
3229 /* If we've found a trapping *_REF, _and_ it dominates EXP
3230 (it's in a basic block on the path from us to the dominator root)
3231 then we can't trap. */
3232 if (found_bb && (((size_t)found_bb->aux) & 1) == 1)
3234 m_nontrapping->add (exp);
3236 else
3238 /* EXP might trap, so insert it into the hash table. */
3239 if (r2bb)
3241 r2bb->phase = nt_call_phase;
3242 r2bb->bb = bb;
3244 else
3246 r2bb = XNEW (struct ref_to_bb);
3247 r2bb->phase = nt_call_phase;
3248 r2bb->bb = bb;
3249 r2bb->exp = exp;
3250 r2bb->size = size;
3251 *slot = r2bb;
3257 /* This is the entry point of gathering non trapping memory accesses.
3258 It will do a dominator walk over the whole function, and it will
3259 make use of the bb->aux pointers. It returns a set of trees
3260 (the MEM_REFs itself) which can't trap. */
3261 static hash_set<tree> *
3262 get_non_trapping (void)
3264 nt_call_phase = 0;
3265 hash_set<tree> *nontrap = new hash_set<tree>;
3267 nontrapping_dom_walker (CDI_DOMINATORS, nontrap)
3268 .walk (cfun->cfg->x_entry_block_ptr);
3270 clear_aux_for_blocks ();
3271 return nontrap;
3274 /* Do the main work of conditional store replacement. We already know
3275 that the recognized pattern looks like so:
3277 split:
3278 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
3279 MIDDLE_BB:
3280 something
3281 fallthrough (edge E0)
3282 JOIN_BB:
3283 some more
3285 We check that MIDDLE_BB contains only one store, that that store
3286 doesn't trap (not via NOTRAP, but via checking if an access to the same
3287 memory location dominates us, or the store is to a local addressable
3288 object) and that the store has a "simple" RHS. */
3290 static bool
3291 cond_store_replacement (basic_block middle_bb, basic_block join_bb,
3292 edge e0, edge e1, hash_set<tree> *nontrap)
3294 gimple *assign = last_and_only_stmt (middle_bb);
3295 tree lhs, rhs, name, name2;
3296 gphi *newphi;
3297 gassign *new_stmt;
3298 gimple_stmt_iterator gsi;
3299 location_t locus;
3301 /* Check if middle_bb contains of only one store. */
3302 if (!assign
3303 || !gimple_assign_single_p (assign)
3304 || gimple_has_volatile_ops (assign))
3305 return false;
3307 /* And no PHI nodes so all uses in the single stmt are also
3308 available where we insert to. */
3309 if (!gimple_seq_empty_p (phi_nodes (middle_bb)))
3310 return false;
3312 locus = gimple_location (assign);
3313 lhs = gimple_assign_lhs (assign);
3314 rhs = gimple_assign_rhs1 (assign);
3315 if ((!REFERENCE_CLASS_P (lhs)
3316 && !DECL_P (lhs))
3317 || !is_gimple_reg_type (TREE_TYPE (lhs)))
3318 return false;
3320 /* Prove that we can move the store down. We could also check
3321 TREE_THIS_NOTRAP here, but in that case we also could move stores,
3322 whose value is not available readily, which we want to avoid. */
3323 if (!nontrap->contains (lhs))
3325 /* If LHS is an access to a local variable without address-taken
3326 (or when we allow data races) and known not to trap, we could
3327 always safely move down the store. */
3328 tree base = get_base_address (lhs);
3329 if (!auto_var_p (base)
3330 || (TREE_ADDRESSABLE (base) && !flag_store_data_races)
3331 || tree_could_trap_p (lhs))
3332 return false;
3335 /* Now we've checked the constraints, so do the transformation:
3336 1) Remove the single store. */
3337 gsi = gsi_for_stmt (assign);
3338 unlink_stmt_vdef (assign);
3339 gsi_remove (&gsi, true);
3340 release_defs (assign);
3342 /* Make both store and load use alias-set zero as we have to
3343 deal with the case of the store being a conditional change
3344 of the dynamic type. */
3345 lhs = unshare_expr (lhs);
3346 tree *basep = &lhs;
3347 while (handled_component_p (*basep))
3348 basep = &TREE_OPERAND (*basep, 0);
3349 if (TREE_CODE (*basep) == MEM_REF
3350 || TREE_CODE (*basep) == TARGET_MEM_REF)
3351 TREE_OPERAND (*basep, 1)
3352 = fold_convert (ptr_type_node, TREE_OPERAND (*basep, 1));
3353 else
3354 *basep = build2 (MEM_REF, TREE_TYPE (*basep),
3355 build_fold_addr_expr (*basep),
3356 build_zero_cst (ptr_type_node));
3358 /* 2) Insert a load from the memory of the store to the temporary
3359 on the edge which did not contain the store. */
3360 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
3361 new_stmt = gimple_build_assign (name, lhs);
3362 gimple_set_location (new_stmt, locus);
3363 lhs = unshare_expr (lhs);
3365 /* Set the no-warning bit on the rhs of the load to avoid uninit
3366 warnings. */
3367 tree rhs1 = gimple_assign_rhs1 (new_stmt);
3368 suppress_warning (rhs1, OPT_Wuninitialized);
3370 gsi_insert_on_edge (e1, new_stmt);
3372 /* 3) Create a PHI node at the join block, with one argument
3373 holding the old RHS, and the other holding the temporary
3374 where we stored the old memory contents. */
3375 name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
3376 newphi = create_phi_node (name2, join_bb);
3377 add_phi_arg (newphi, rhs, e0, locus);
3378 add_phi_arg (newphi, name, e1, locus);
3380 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
3382 /* 4) Insert that PHI node. */
3383 gsi = gsi_after_labels (join_bb);
3384 if (gsi_end_p (gsi))
3386 gsi = gsi_last_bb (join_bb);
3387 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
3389 else
3390 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
3392 if (dump_file && (dump_flags & TDF_DETAILS))
3394 fprintf (dump_file, "\nConditional store replacement happened!");
3395 fprintf (dump_file, "\nReplaced the store with a load.");
3396 fprintf (dump_file, "\nInserted a new PHI statement in joint block:\n");
3397 print_gimple_stmt (dump_file, new_stmt, 0, TDF_VOPS|TDF_MEMSYMS);
3399 statistics_counter_event (cfun, "conditional store replacement", 1);
3401 return true;
3404 /* Do the main work of conditional store replacement. */
3406 static bool
3407 cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb,
3408 basic_block join_bb, gimple *then_assign,
3409 gimple *else_assign)
3411 tree lhs_base, lhs, then_rhs, else_rhs, name;
3412 location_t then_locus, else_locus;
3413 gimple_stmt_iterator gsi;
3414 gphi *newphi;
3415 gassign *new_stmt;
3417 if (then_assign == NULL
3418 || !gimple_assign_single_p (then_assign)
3419 || gimple_clobber_p (then_assign)
3420 || gimple_has_volatile_ops (then_assign)
3421 || else_assign == NULL
3422 || !gimple_assign_single_p (else_assign)
3423 || gimple_clobber_p (else_assign)
3424 || gimple_has_volatile_ops (else_assign))
3425 return false;
3427 lhs = gimple_assign_lhs (then_assign);
3428 if (!is_gimple_reg_type (TREE_TYPE (lhs))
3429 || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0))
3430 return false;
3432 lhs_base = get_base_address (lhs);
3433 if (lhs_base == NULL_TREE
3434 || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF))
3435 return false;
3437 then_rhs = gimple_assign_rhs1 (then_assign);
3438 else_rhs = gimple_assign_rhs1 (else_assign);
3439 then_locus = gimple_location (then_assign);
3440 else_locus = gimple_location (else_assign);
3442 /* Now we've checked the constraints, so do the transformation:
3443 1) Remove the stores. */
3444 gsi = gsi_for_stmt (then_assign);
3445 unlink_stmt_vdef (then_assign);
3446 gsi_remove (&gsi, true);
3447 release_defs (then_assign);
3449 gsi = gsi_for_stmt (else_assign);
3450 unlink_stmt_vdef (else_assign);
3451 gsi_remove (&gsi, true);
3452 release_defs (else_assign);
3454 /* 2) Create a PHI node at the join block, with one argument
3455 holding the old RHS, and the other holding the temporary
3456 where we stored the old memory contents. */
3457 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
3458 newphi = create_phi_node (name, join_bb);
3459 add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus);
3460 add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus);
3462 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
3464 /* 3) Insert that PHI node. */
3465 gsi = gsi_after_labels (join_bb);
3466 if (gsi_end_p (gsi))
3468 gsi = gsi_last_bb (join_bb);
3469 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
3471 else
3472 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
3474 statistics_counter_event (cfun, "if-then-else store replacement", 1);
3476 return true;
3479 /* Return the single store in BB with VDEF or NULL if there are
3480 other stores in the BB or loads following the store. */
3482 static gimple *
3483 single_trailing_store_in_bb (basic_block bb, tree vdef)
3485 if (SSA_NAME_IS_DEFAULT_DEF (vdef))
3486 return NULL;
3487 gimple *store = SSA_NAME_DEF_STMT (vdef);
3488 if (gimple_bb (store) != bb
3489 || gimple_code (store) == GIMPLE_PHI)
3490 return NULL;
3492 /* Verify there is no other store in this BB. */
3493 if (!SSA_NAME_IS_DEFAULT_DEF (gimple_vuse (store))
3494 && gimple_bb (SSA_NAME_DEF_STMT (gimple_vuse (store))) == bb
3495 && gimple_code (SSA_NAME_DEF_STMT (gimple_vuse (store))) != GIMPLE_PHI)
3496 return NULL;
3498 /* Verify there is no load or store after the store. */
3499 use_operand_p use_p;
3500 imm_use_iterator imm_iter;
3501 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_vdef (store))
3502 if (USE_STMT (use_p) != store
3503 && gimple_bb (USE_STMT (use_p)) == bb)
3504 return NULL;
3506 return store;
3509 /* Conditional store replacement. We already know
3510 that the recognized pattern looks like so:
3512 split:
3513 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
3514 THEN_BB:
3516 X = Y;
3518 goto JOIN_BB;
3519 ELSE_BB:
3521 X = Z;
3523 fallthrough (edge E0)
3524 JOIN_BB:
3525 some more
3527 We check that it is safe to sink the store to JOIN_BB by verifying that
3528 there are no read-after-write or write-after-write dependencies in
3529 THEN_BB and ELSE_BB. */
3531 static bool
3532 cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
3533 basic_block join_bb)
3535 vec<data_reference_p> then_datarefs, else_datarefs;
3536 vec<ddr_p> then_ddrs, else_ddrs;
3537 gimple *then_store, *else_store;
3538 bool found, ok = false, res;
3539 struct data_dependence_relation *ddr;
3540 data_reference_p then_dr, else_dr;
3541 int i, j;
3542 tree then_lhs, else_lhs;
3543 basic_block blocks[3];
3545 /* Handle the case with single store in THEN_BB and ELSE_BB. That is
3546 cheap enough to always handle as it allows us to elide dependence
3547 checking. */
3548 gphi *vphi = NULL;
3549 for (gphi_iterator si = gsi_start_phis (join_bb); !gsi_end_p (si);
3550 gsi_next (&si))
3551 if (virtual_operand_p (gimple_phi_result (si.phi ())))
3553 vphi = si.phi ();
3554 break;
3556 if (!vphi)
3557 return false;
3558 tree then_vdef = PHI_ARG_DEF_FROM_EDGE (vphi, single_succ_edge (then_bb));
3559 tree else_vdef = PHI_ARG_DEF_FROM_EDGE (vphi, single_succ_edge (else_bb));
3560 gimple *then_assign = single_trailing_store_in_bb (then_bb, then_vdef);
3561 if (then_assign)
3563 gimple *else_assign = single_trailing_store_in_bb (else_bb, else_vdef);
3564 if (else_assign)
3565 return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
3566 then_assign, else_assign);
3569 /* If either vectorization or if-conversion is disabled then do
3570 not sink any stores. */
3571 if (param_max_stores_to_sink == 0
3572 || (!flag_tree_loop_vectorize && !flag_tree_slp_vectorize)
3573 || !flag_tree_loop_if_convert)
3574 return false;
3576 /* Find data references. */
3577 then_datarefs.create (1);
3578 else_datarefs.create (1);
3579 if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs)
3580 == chrec_dont_know)
3581 || !then_datarefs.length ()
3582 || (find_data_references_in_bb (NULL, else_bb, &else_datarefs)
3583 == chrec_dont_know)
3584 || !else_datarefs.length ())
3586 free_data_refs (then_datarefs);
3587 free_data_refs (else_datarefs);
3588 return false;
3591 /* Find pairs of stores with equal LHS. */
3592 auto_vec<gimple *, 1> then_stores, else_stores;
3593 FOR_EACH_VEC_ELT (then_datarefs, i, then_dr)
3595 if (DR_IS_READ (then_dr))
3596 continue;
3598 then_store = DR_STMT (then_dr);
3599 then_lhs = gimple_get_lhs (then_store);
3600 if (then_lhs == NULL_TREE)
3601 continue;
3602 found = false;
3604 FOR_EACH_VEC_ELT (else_datarefs, j, else_dr)
3606 if (DR_IS_READ (else_dr))
3607 continue;
3609 else_store = DR_STMT (else_dr);
3610 else_lhs = gimple_get_lhs (else_store);
3611 if (else_lhs == NULL_TREE)
3612 continue;
3614 if (operand_equal_p (then_lhs, else_lhs, 0))
3616 found = true;
3617 break;
3621 if (!found)
3622 continue;
3624 then_stores.safe_push (then_store);
3625 else_stores.safe_push (else_store);
3628 /* No pairs of stores found. */
3629 if (!then_stores.length ()
3630 || then_stores.length () > (unsigned) param_max_stores_to_sink)
3632 free_data_refs (then_datarefs);
3633 free_data_refs (else_datarefs);
3634 return false;
3637 /* Compute and check data dependencies in both basic blocks. */
3638 then_ddrs.create (1);
3639 else_ddrs.create (1);
3640 if (!compute_all_dependences (then_datarefs, &then_ddrs,
3641 vNULL, false)
3642 || !compute_all_dependences (else_datarefs, &else_ddrs,
3643 vNULL, false))
3645 free_dependence_relations (then_ddrs);
3646 free_dependence_relations (else_ddrs);
3647 free_data_refs (then_datarefs);
3648 free_data_refs (else_datarefs);
3649 return false;
3651 blocks[0] = then_bb;
3652 blocks[1] = else_bb;
3653 blocks[2] = join_bb;
3654 renumber_gimple_stmt_uids_in_blocks (blocks, 3);
3656 /* Check that there are no read-after-write or write-after-write dependencies
3657 in THEN_BB. */
3658 FOR_EACH_VEC_ELT (then_ddrs, i, ddr)
3660 struct data_reference *dra = DDR_A (ddr);
3661 struct data_reference *drb = DDR_B (ddr);
3663 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
3664 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
3665 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
3666 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
3667 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
3668 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
3670 free_dependence_relations (then_ddrs);
3671 free_dependence_relations (else_ddrs);
3672 free_data_refs (then_datarefs);
3673 free_data_refs (else_datarefs);
3674 return false;
3678 /* Check that there are no read-after-write or write-after-write dependencies
3679 in ELSE_BB. */
3680 FOR_EACH_VEC_ELT (else_ddrs, i, ddr)
3682 struct data_reference *dra = DDR_A (ddr);
3683 struct data_reference *drb = DDR_B (ddr);
3685 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
3686 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
3687 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
3688 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
3689 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
3690 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
3692 free_dependence_relations (then_ddrs);
3693 free_dependence_relations (else_ddrs);
3694 free_data_refs (then_datarefs);
3695 free_data_refs (else_datarefs);
3696 return false;
3700 /* Sink stores with same LHS. */
3701 FOR_EACH_VEC_ELT (then_stores, i, then_store)
3703 else_store = else_stores[i];
3704 res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
3705 then_store, else_store);
3706 ok = ok || res;
3709 free_dependence_relations (then_ddrs);
3710 free_dependence_relations (else_ddrs);
3711 free_data_refs (then_datarefs);
3712 free_data_refs (else_datarefs);
3714 return ok;
3717 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
3719 static bool
3720 local_mem_dependence (gimple *stmt, basic_block bb)
3722 tree vuse = gimple_vuse (stmt);
3723 gimple *def;
3725 if (!vuse)
3726 return false;
3728 def = SSA_NAME_DEF_STMT (vuse);
3729 return (def && gimple_bb (def) == bb);
3732 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
3733 BB1 and BB2 are "then" and "else" blocks dependent on this test,
3734 and BB3 rejoins control flow following BB1 and BB2, look for
3735 opportunities to hoist loads as follows. If BB3 contains a PHI of
3736 two loads, one each occurring in BB1 and BB2, and the loads are
3737 provably of adjacent fields in the same structure, then move both
3738 loads into BB0. Of course this can only be done if there are no
3739 dependencies preventing such motion.
3741 One of the hoisted loads will always be speculative, so the
3742 transformation is currently conservative:
3744 - The fields must be strictly adjacent.
3745 - The two fields must occupy a single memory block that is
3746 guaranteed to not cross a page boundary.
3748 The last is difficult to prove, as such memory blocks should be
3749 aligned on the minimum of the stack alignment boundary and the
3750 alignment guaranteed by heap allocation interfaces. Thus we rely
3751 on a parameter for the alignment value.
3753 Provided a good value is used for the last case, the first
3754 restriction could possibly be relaxed. */
3756 static void
3757 hoist_adjacent_loads (basic_block bb0, basic_block bb1,
3758 basic_block bb2, basic_block bb3)
3760 unsigned HOST_WIDE_INT param_align = param_l1_cache_line_size;
3761 unsigned HOST_WIDE_INT param_align_bits = param_align * BITS_PER_UNIT;
3762 gphi_iterator gsi;
3764 /* Walk the phis in bb3 looking for an opportunity. We are looking
3765 for phis of two SSA names, one each of which is defined in bb1 and
3766 bb2. */
3767 for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi))
3769 gphi *phi_stmt = gsi.phi ();
3770 gimple *def1, *def2;
3771 tree arg1, arg2, ref1, ref2, field1, field2;
3772 tree tree_offset1, tree_offset2, tree_size2, next;
3773 unsigned HOST_WIDE_INT offset1, offset2, size2, align1;
3774 gimple_stmt_iterator gsi2;
3775 basic_block bb_for_def1, bb_for_def2;
3777 if (gimple_phi_num_args (phi_stmt) != 2
3778 || virtual_operand_p (gimple_phi_result (phi_stmt)))
3779 continue;
3781 arg1 = gimple_phi_arg_def (phi_stmt, 0);
3782 arg2 = gimple_phi_arg_def (phi_stmt, 1);
3784 if (TREE_CODE (arg1) != SSA_NAME
3785 || TREE_CODE (arg2) != SSA_NAME
3786 || SSA_NAME_IS_DEFAULT_DEF (arg1)
3787 || SSA_NAME_IS_DEFAULT_DEF (arg2))
3788 continue;
3790 def1 = SSA_NAME_DEF_STMT (arg1);
3791 def2 = SSA_NAME_DEF_STMT (arg2);
3793 if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2)
3794 && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2))
3795 continue;
3797 /* Check the mode of the arguments to be sure a conditional move
3798 can be generated for it. */
3799 if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1)))
3800 == CODE_FOR_nothing)
3801 continue;
3803 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
3804 if (!gimple_assign_single_p (def1)
3805 || !gimple_assign_single_p (def2)
3806 || gimple_has_volatile_ops (def1)
3807 || gimple_has_volatile_ops (def2))
3808 continue;
3810 ref1 = gimple_assign_rhs1 (def1);
3811 ref2 = gimple_assign_rhs1 (def2);
3813 if (TREE_CODE (ref1) != COMPONENT_REF
3814 || TREE_CODE (ref2) != COMPONENT_REF)
3815 continue;
3817 /* The zeroth operand of the two component references must be
3818 identical. It is not sufficient to compare get_base_address of
3819 the two references, because this could allow for different
3820 elements of the same array in the two trees. It is not safe to
3821 assume that the existence of one array element implies the
3822 existence of a different one. */
3823 if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0))
3824 continue;
3826 field1 = TREE_OPERAND (ref1, 1);
3827 field2 = TREE_OPERAND (ref2, 1);
3829 /* Check for field adjacency, and ensure field1 comes first. */
3830 for (next = DECL_CHAIN (field1);
3831 next && TREE_CODE (next) != FIELD_DECL;
3832 next = DECL_CHAIN (next))
3835 if (next != field2)
3837 for (next = DECL_CHAIN (field2);
3838 next && TREE_CODE (next) != FIELD_DECL;
3839 next = DECL_CHAIN (next))
3842 if (next != field1)
3843 continue;
3845 std::swap (field1, field2);
3846 std::swap (def1, def2);
3849 bb_for_def1 = gimple_bb (def1);
3850 bb_for_def2 = gimple_bb (def2);
3852 /* Check for proper alignment of the first field. */
3853 tree_offset1 = bit_position (field1);
3854 tree_offset2 = bit_position (field2);
3855 tree_size2 = DECL_SIZE (field2);
3857 if (!tree_fits_uhwi_p (tree_offset1)
3858 || !tree_fits_uhwi_p (tree_offset2)
3859 || !tree_fits_uhwi_p (tree_size2))
3860 continue;
3862 offset1 = tree_to_uhwi (tree_offset1);
3863 offset2 = tree_to_uhwi (tree_offset2);
3864 size2 = tree_to_uhwi (tree_size2);
3865 align1 = DECL_ALIGN (field1) % param_align_bits;
3867 if (offset1 % BITS_PER_UNIT != 0)
3868 continue;
3870 /* For profitability, the two field references should fit within
3871 a single cache line. */
3872 if (align1 + offset2 - offset1 + size2 > param_align_bits)
3873 continue;
3875 /* The two expressions cannot be dependent upon vdefs defined
3876 in bb1/bb2. */
3877 if (local_mem_dependence (def1, bb_for_def1)
3878 || local_mem_dependence (def2, bb_for_def2))
3879 continue;
3881 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
3882 bb0. We hoist the first one first so that a cache miss is handled
3883 efficiently regardless of hardware cache-fill policy. */
3884 gsi2 = gsi_for_stmt (def1);
3885 gsi_move_to_bb_end (&gsi2, bb0);
3886 gsi2 = gsi_for_stmt (def2);
3887 gsi_move_to_bb_end (&gsi2, bb0);
3888 statistics_counter_event (cfun, "hoisted loads", 1);
3890 if (dump_file && (dump_flags & TDF_DETAILS))
3892 fprintf (dump_file,
3893 "\nHoisting adjacent loads from %d and %d into %d: \n",
3894 bb_for_def1->index, bb_for_def2->index, bb0->index);
3895 print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS);
3896 print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS);
3901 /* Determine whether we should attempt to hoist adjacent loads out of
3902 diamond patterns in pass_phiopt. Always hoist loads if
3903 -fhoist-adjacent-loads is specified and the target machine has
3904 both a conditional move instruction and a defined cache line size. */
3906 static bool
3907 gate_hoist_loads (void)
3909 return (flag_hoist_adjacent_loads == 1
3910 && param_l1_cache_line_size
3911 && HAVE_conditional_move);
3914 /* This pass tries to replaces an if-then-else block with an
3915 assignment. We have different kinds of transformations.
3916 Some of these transformations are also performed by the ifcvt
3917 RTL optimizer.
3919 PHI-OPT using Match-and-simplify infrastructure
3920 -----------------------
3922 The PHI-OPT pass will try to use match-and-simplify infrastructure
3923 (gimple_simplify) to do transformations. This is implemented in
3924 match_simplify_replacement.
3926 The way it works is it replaces:
3927 bb0:
3928 if (cond) goto bb2; else goto bb1;
3929 bb1:
3930 bb2:
3931 x = PHI <a (bb1), b (bb0), ...>;
3933 with a statement if it gets simplified from `cond ? b : a`.
3935 bb0:
3936 x1 = cond ? b : a;
3937 bb2:
3938 x = PHI <a (bb1), x1 (bb0), ...>;
3939 Bb1 might be removed as it becomes unreachable when doing the replacement.
3940 Though bb1 does not have to be considered a forwarding basic block from bb0.
3942 Will try to see if `(!cond) ? a : b` gets simplified (iff !cond simplifies);
3943 this is done not to have an explosion of patterns in match.pd.
3944 Note bb1 does not need to be completely empty, it can contain
3945 one statement which is known not to trap.
3947 It also can handle the case where we have two forwarding bbs (diamond):
3948 bb0:
3949 if (cond) goto bb2; else goto bb1;
3950 bb1: goto bb3;
3951 bb2: goto bb3;
3952 bb3:
3953 x = PHI <a (bb1), b (bb2), ...>;
3954 And that is replaced with a statement if it is simplified
3955 from `cond ? b : a`.
3956 Again bb1 and bb2 does not have to be completely empty but
3957 each can contain one statement which is known not to trap.
3958 But in this case bb1/bb2 can only be forwarding basic blocks.
3960 This fully replaces the old "Conditional Replacement",
3961 "ABS Replacement" transformations as they are now
3962 implmeneted in match.pd.
3963 Some parts of the "MIN/MAX Replacement" are re-implemented in match.pd.
3965 Value Replacement
3966 -----------------
3968 This transformation, implemented in value_replacement, replaces
3970 bb0:
3971 if (a != b) goto bb2; else goto bb1;
3972 bb1:
3973 bb2:
3974 x = PHI <a (bb1), b (bb0), ...>;
3976 with
3978 bb0:
3979 bb2:
3980 x = PHI <b (bb0), ...>;
3982 This opportunity can sometimes occur as a result of other
3983 optimizations.
3986 Another case caught by value replacement looks like this:
3988 bb0:
3989 t1 = a == CONST;
3990 t2 = b > c;
3991 t3 = t1 & t2;
3992 if (t3 != 0) goto bb1; else goto bb2;
3993 bb1:
3994 bb2:
3995 x = PHI (CONST, a)
3997 Gets replaced with:
3998 bb0:
3999 bb2:
4000 t1 = a == CONST;
4001 t2 = b > c;
4002 t3 = t1 & t2;
4003 x = a;
4005 MIN/MAX Replacement
4006 -------------------
4008 This transformation, minmax_replacement replaces
4010 bb0:
4011 if (a <= b) goto bb2; else goto bb1;
4012 bb1:
4013 bb2:
4014 x = PHI <b (bb1), a (bb0), ...>;
4016 with
4018 bb0:
4019 x' = MIN_EXPR (a, b)
4020 bb2:
4021 x = PHI <x' (bb0), ...>;
4023 A similar transformation is done for MAX_EXPR.
4026 This pass also performs a fifth transformation of a slightly different
4027 flavor.
4029 Factor operations in COND_EXPR
4030 ------------------------------
4032 This transformation factors the unary operations out of COND_EXPR with
4033 factor_out_conditional_operation.
4035 For example:
4036 if (a <= CST) goto <bb 3>; else goto <bb 4>;
4037 <bb 3>:
4038 tmp = (int) a;
4039 <bb 4>:
4040 tmp = PHI <tmp, CST>
4042 Into:
4043 if (a <= CST) goto <bb 3>; else goto <bb 4>;
4044 <bb 3>:
4045 <bb 4>:
4046 a = PHI <a, CST>
4047 tmp = (int) a;
4049 Adjacent Load Hoisting
4050 ----------------------
4052 This transformation replaces
4054 bb0:
4055 if (...) goto bb2; else goto bb1;
4056 bb1:
4057 x1 = (<expr>).field1;
4058 goto bb3;
4059 bb2:
4060 x2 = (<expr>).field2;
4061 bb3:
4062 # x = PHI <x1, x2>;
4064 with
4066 bb0:
4067 x1 = (<expr>).field1;
4068 x2 = (<expr>).field2;
4069 if (...) goto bb2; else goto bb1;
4070 bb1:
4071 goto bb3;
4072 bb2:
4073 bb3:
4074 # x = PHI <x1, x2>;
4076 The purpose of this transformation is to enable generation of conditional
4077 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
4078 the loads is speculative, the transformation is restricted to very
4079 specific cases to avoid introducing a page fault. We are looking for
4080 the common idiom:
4082 if (...)
4083 x = y->left;
4084 else
4085 x = y->right;
4087 where left and right are typically adjacent pointers in a tree structure. */
4089 namespace {
4091 const pass_data pass_data_phiopt =
4093 GIMPLE_PASS, /* type */
4094 "phiopt", /* name */
4095 OPTGROUP_NONE, /* optinfo_flags */
4096 TV_TREE_PHIOPT, /* tv_id */
4097 ( PROP_cfg | PROP_ssa ), /* properties_required */
4098 0, /* properties_provided */
4099 0, /* properties_destroyed */
4100 0, /* todo_flags_start */
4101 0, /* todo_flags_finish */
4104 class pass_phiopt : public gimple_opt_pass
4106 public:
4107 pass_phiopt (gcc::context *ctxt)
4108 : gimple_opt_pass (pass_data_phiopt, ctxt), early_p (false)
4111 /* opt_pass methods: */
4112 opt_pass * clone () final override { return new pass_phiopt (m_ctxt); }
4113 void set_pass_param (unsigned n, bool param) final override
4115 gcc_assert (n == 0);
4116 early_p = param;
4118 bool gate (function *) final override { return flag_ssa_phiopt; }
4119 unsigned int execute (function *) final override;
4121 private:
4122 bool early_p;
4123 }; // class pass_phiopt
4125 } // anon namespace
4127 gimple_opt_pass *
4128 make_pass_phiopt (gcc::context *ctxt)
4130 return new pass_phiopt (ctxt);
4133 unsigned int
4134 pass_phiopt::execute (function *)
4136 bool do_hoist_loads = !early_p ? gate_hoist_loads () : false;
4137 basic_block bb;
4138 basic_block *bb_order;
4139 unsigned n, i;
4140 bool cfgchanged = false;
4142 calculate_dominance_info (CDI_DOMINATORS);
4143 mark_ssa_maybe_undefs ();
4145 /* Search every basic block for COND_EXPR we may be able to optimize.
4147 We walk the blocks in order that guarantees that a block with
4148 a single predecessor is processed before the predecessor.
4149 This ensures that we collapse inner ifs before visiting the
4150 outer ones, and also that we do not try to visit a removed
4151 block. */
4152 bb_order = single_pred_before_succ_order ();
4153 n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS;
4155 for (i = 0; i < n; i++)
4157 gphi *phi;
4158 basic_block bb1, bb2;
4159 edge e1, e2;
4160 tree arg0, arg1;
4161 bool diamond_p = false;
4163 bb = bb_order[i];
4165 /* Check to see if the last statement is a GIMPLE_COND. */
4166 gcond *cond_stmt = safe_dyn_cast <gcond *> (*gsi_last_bb (bb));
4167 if (!cond_stmt)
4168 continue;
4170 e1 = EDGE_SUCC (bb, 0);
4171 bb1 = e1->dest;
4172 e2 = EDGE_SUCC (bb, 1);
4173 bb2 = e2->dest;
4175 /* We cannot do the optimization on abnormal edges. */
4176 if ((e1->flags & EDGE_ABNORMAL) != 0
4177 || (e2->flags & EDGE_ABNORMAL) != 0)
4178 continue;
4180 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
4181 if (EDGE_COUNT (bb1->succs) == 0
4182 || EDGE_COUNT (bb2->succs) == 0)
4183 continue;
4185 /* Find the bb which is the fall through to the other. */
4186 if (EDGE_SUCC (bb1, 0)->dest == bb2)
4188 else if (EDGE_SUCC (bb2, 0)->dest == bb1)
4190 std::swap (bb1, bb2);
4191 std::swap (e1, e2);
4193 else if (EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest
4194 && single_succ_p (bb2))
4196 diamond_p = true;
4197 e2 = EDGE_SUCC (bb2, 0);
4198 /* Make sure bb2 is just a fall through. */
4199 if ((e2->flags & EDGE_FALLTHRU) == 0)
4200 continue;
4202 else
4203 continue;
4205 e1 = EDGE_SUCC (bb1, 0);
4207 /* Make sure that bb1 is just a fall through. */
4208 if (!single_succ_p (bb1)
4209 || (e1->flags & EDGE_FALLTHRU) == 0)
4210 continue;
4212 if (diamond_p)
4214 basic_block bb3 = e1->dest;
4216 if (!single_pred_p (bb1)
4217 || !single_pred_p (bb2))
4218 continue;
4220 if (do_hoist_loads
4221 && !FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt)))
4222 && EDGE_COUNT (bb->succs) == 2
4223 && EDGE_COUNT (bb3->preds) == 2
4224 /* If one edge or the other is dominant, a conditional move
4225 is likely to perform worse than the well-predicted branch. */
4226 && !predictable_edge_p (EDGE_SUCC (bb, 0))
4227 && !predictable_edge_p (EDGE_SUCC (bb, 1)))
4228 hoist_adjacent_loads (bb, bb1, bb2, bb3);
4231 gimple_stmt_iterator gsi;
4232 bool candorest = true;
4234 /* Check that we're looking for nested phis. */
4235 basic_block merge = diamond_p ? EDGE_SUCC (bb2, 0)->dest : bb2;
4236 gimple_seq phis = phi_nodes (merge);
4238 /* Value replacement can work with more than one PHI
4239 so try that first. */
4240 if (!early_p && !diamond_p)
4241 for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
4243 phi = as_a <gphi *> (gsi_stmt (gsi));
4244 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
4245 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
4246 if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2)
4248 candorest = false;
4249 cfgchanged = true;
4250 break;
4254 if (!candorest)
4255 continue;
4257 phi = single_non_singleton_phi_for_edges (phis, e1, e2);
4258 if (!phi)
4259 continue;
4261 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
4262 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
4264 /* Something is wrong if we cannot find the arguments in the PHI
4265 node. */
4266 gcc_assert (arg0 != NULL_TREE && arg1 != NULL_TREE);
4268 if (single_pred_p (bb1)
4269 && EDGE_COUNT (merge->preds) == 2)
4271 gphi *newphi = phi;
4272 while (newphi)
4274 phi = newphi;
4275 /* factor_out_conditional_operation may create a new PHI in
4276 BB2 and eliminate an existing PHI in BB2. Recompute values
4277 that may be affected by that change. */
4278 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
4279 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
4280 gcc_assert (arg0 != NULL_TREE && arg1 != NULL_TREE);
4281 newphi = factor_out_conditional_operation (e1, e2, phi,
4282 arg0, arg1,
4283 cond_stmt);
4287 /* Do the replacement of conditional if it can be done. */
4288 if (match_simplify_replacement (bb, bb1, bb2, e1, e2, phi,
4289 arg0, arg1, early_p, diamond_p))
4290 cfgchanged = true;
4291 else if (!early_p
4292 && !diamond_p
4293 && single_pred_p (bb1)
4294 && cond_removal_in_builtin_zero_pattern (bb, bb1, e1, e2,
4295 phi, arg0, arg1))
4296 cfgchanged = true;
4297 else if (minmax_replacement (bb, bb1, bb2, e1, e2, phi, arg0, arg1,
4298 diamond_p))
4299 cfgchanged = true;
4300 else if (single_pred_p (bb1)
4301 && !diamond_p
4302 && spaceship_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
4303 cfgchanged = true;
4306 free (bb_order);
4308 if (cfgchanged)
4309 return TODO_cleanup_cfg;
4310 return 0;
4313 /* This pass tries to transform conditional stores into unconditional
4314 ones, enabling further simplifications with the simpler then and else
4315 blocks. In particular it replaces this:
4317 bb0:
4318 if (cond) goto bb2; else goto bb1;
4319 bb1:
4320 *p = RHS;
4321 bb2:
4323 with
4325 bb0:
4326 if (cond) goto bb1; else goto bb2;
4327 bb1:
4328 condtmp' = *p;
4329 bb2:
4330 condtmp = PHI <RHS, condtmp'>
4331 *p = condtmp;
4333 This transformation can only be done under several constraints,
4334 documented below. It also replaces:
4336 bb0:
4337 if (cond) goto bb2; else goto bb1;
4338 bb1:
4339 *p = RHS1;
4340 goto bb3;
4341 bb2:
4342 *p = RHS2;
4343 bb3:
4345 with
4347 bb0:
4348 if (cond) goto bb3; else goto bb1;
4349 bb1:
4350 bb3:
4351 condtmp = PHI <RHS1, RHS2>
4352 *p = condtmp; */
4354 namespace {
4356 const pass_data pass_data_cselim =
4358 GIMPLE_PASS, /* type */
4359 "cselim", /* name */
4360 OPTGROUP_NONE, /* optinfo_flags */
4361 TV_TREE_PHIOPT, /* tv_id */
4362 ( PROP_cfg | PROP_ssa ), /* properties_required */
4363 0, /* properties_provided */
4364 0, /* properties_destroyed */
4365 0, /* todo_flags_start */
4366 0, /* todo_flags_finish */
4369 class pass_cselim : public gimple_opt_pass
4371 public:
4372 pass_cselim (gcc::context *ctxt)
4373 : gimple_opt_pass (pass_data_cselim, ctxt)
4376 /* opt_pass methods: */
4377 bool gate (function *) final override { return flag_tree_cselim; }
4378 unsigned int execute (function *) final override;
4380 }; // class pass_cselim
4382 } // anon namespace
4384 gimple_opt_pass *
4385 make_pass_cselim (gcc::context *ctxt)
4387 return new pass_cselim (ctxt);
4390 unsigned int
4391 pass_cselim::execute (function *)
4393 basic_block bb;
4394 basic_block *bb_order;
4395 unsigned n, i;
4396 bool cfgchanged = false;
4397 hash_set<tree> *nontrap = 0;
4398 unsigned todo = 0;
4400 /* ??? We are not interested in loop related info, but the following
4401 will create it, ICEing as we didn't init loops with pre-headers.
4402 An interfacing issue of find_data_references_in_bb. */
4403 loop_optimizer_init (LOOPS_NORMAL);
4404 scev_initialize ();
4406 calculate_dominance_info (CDI_DOMINATORS);
4408 /* Calculate the set of non-trapping memory accesses. */
4409 nontrap = get_non_trapping ();
4411 /* Search every basic block for COND_EXPR we may be able to optimize.
4413 We walk the blocks in order that guarantees that a block with
4414 a single predecessor is processed before the predecessor.
4415 This ensures that we collapse inner ifs before visiting the
4416 outer ones, and also that we do not try to visit a removed
4417 block. */
4418 bb_order = single_pred_before_succ_order ();
4419 n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS;
4421 for (i = 0; i < n; i++)
4423 basic_block bb1, bb2;
4424 edge e1, e2;
4425 bool diamond_p = false;
4427 bb = bb_order[i];
4429 /* Check to see if the last statement is a GIMPLE_COND. */
4430 gcond *cond_stmt = safe_dyn_cast <gcond *> (*gsi_last_bb (bb));
4431 if (!cond_stmt)
4432 continue;
4434 e1 = EDGE_SUCC (bb, 0);
4435 bb1 = e1->dest;
4436 e2 = EDGE_SUCC (bb, 1);
4437 bb2 = e2->dest;
4439 /* We cannot do the optimization on abnormal edges. */
4440 if ((e1->flags & EDGE_ABNORMAL) != 0
4441 || (e2->flags & EDGE_ABNORMAL) != 0)
4442 continue;
4444 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
4445 if (EDGE_COUNT (bb1->succs) == 0
4446 || EDGE_COUNT (bb2->succs) == 0)
4447 continue;
4449 /* Find the bb which is the fall through to the other. */
4450 if (EDGE_SUCC (bb1, 0)->dest == bb2)
4452 else if (EDGE_SUCC (bb2, 0)->dest == bb1)
4454 std::swap (bb1, bb2);
4455 std::swap (e1, e2);
4457 else if (EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest
4458 && single_succ_p (bb2))
4460 diamond_p = true;
4461 e2 = EDGE_SUCC (bb2, 0);
4462 /* Make sure bb2 is just a fall through. */
4463 if ((e2->flags & EDGE_FALLTHRU) == 0)
4464 continue;
4466 else
4467 continue;
4469 e1 = EDGE_SUCC (bb1, 0);
4471 /* Make sure that bb1 is just a fall through. */
4472 if (!single_succ_p (bb1)
4473 || (e1->flags & EDGE_FALLTHRU) == 0)
4474 continue;
4476 if (diamond_p)
4478 basic_block bb3 = e1->dest;
4480 /* Only handle sinking of store from 2 bbs only,
4481 The middle bbs don't need to come from the
4482 if always since we are sinking rather than
4483 hoisting. */
4484 if (EDGE_COUNT (bb3->preds) != 2)
4485 continue;
4486 if (cond_if_else_store_replacement (bb1, bb2, bb3))
4487 cfgchanged = true;
4488 continue;
4491 /* Also make sure that bb1 only have one predecessor and that it
4492 is bb. */
4493 if (!single_pred_p (bb1)
4494 || single_pred (bb1) != bb)
4495 continue;
4497 /* bb1 is the middle block, bb2 the join block, bb the split block,
4498 e1 the fallthrough edge from bb1 to bb2. We can't do the
4499 optimization if the join block has more than two predecessors. */
4500 if (EDGE_COUNT (bb2->preds) > 2)
4501 continue;
4502 if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
4503 cfgchanged = true;
4506 free (bb_order);
4508 delete nontrap;
4509 /* If the CFG has changed, we should cleanup the CFG. */
4510 if (cfgchanged)
4512 /* In cond-store replacement we have added some loads on edges
4513 and new VOPS (as we moved the store, and created a load). */
4514 gsi_commit_edge_inserts ();
4515 todo = TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
4517 scev_finalize ();
4518 loop_optimizer_finalize ();
4519 return todo;