[PATCH] [1/n] Fix minor SSA_NAME leaks
[official-gcc.git] / gcc / tree-ssa-phiopt.c
blobcfa38680249cef2795c2de3de247403162579cf8
1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2015 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 "cfghooks.h"
25 #include "tree.h"
26 #include "gimple.h"
27 #include "rtl.h"
28 #include "ssa.h"
29 #include "alias.h"
30 #include "fold-const.h"
31 #include "stor-layout.h"
32 #include "flags.h"
33 #include "tm_p.h"
34 #include "cfganal.h"
35 #include "internal-fn.h"
36 #include "gimplify.h"
37 #include "gimple-iterator.h"
38 #include "gimplify-me.h"
39 #include "tree-cfg.h"
40 #include "insn-config.h"
41 #include "tree-dfa.h"
42 #include "tree-pass.h"
43 #include "langhooks.h"
44 #include "domwalk.h"
45 #include "cfgloop.h"
46 #include "tree-data-ref.h"
47 #include "gimple-pretty-print.h"
48 #include "insn-codes.h"
49 #include "optabs-tree.h"
50 #include "tree-scalar-evolution.h"
51 #include "tree-inline.h"
52 #include "params.h"
54 static unsigned int tree_ssa_phiopt_worker (bool, bool);
55 static bool conditional_replacement (basic_block, basic_block,
56 edge, edge, gphi *, tree, tree);
57 static bool factor_out_conditional_conversion (edge, edge, gphi *, tree, tree);
58 static int value_replacement (basic_block, basic_block,
59 edge, edge, gimple *, tree, tree);
60 static bool minmax_replacement (basic_block, basic_block,
61 edge, edge, gimple *, tree, tree);
62 static bool abs_replacement (basic_block, basic_block,
63 edge, edge, gimple *, tree, tree);
64 static bool cond_store_replacement (basic_block, basic_block, edge, edge,
65 hash_set<tree> *);
66 static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block);
67 static hash_set<tree> * get_non_trapping ();
68 static void replace_phi_edge_with_variable (basic_block, edge, gimple *, tree);
69 static void hoist_adjacent_loads (basic_block, basic_block,
70 basic_block, basic_block);
71 static bool gate_hoist_loads (void);
73 /* This pass tries to transform conditional stores into unconditional
74 ones, enabling further simplifications with the simpler then and else
75 blocks. In particular it replaces this:
77 bb0:
78 if (cond) goto bb2; else goto bb1;
79 bb1:
80 *p = RHS;
81 bb2:
83 with
85 bb0:
86 if (cond) goto bb1; else goto bb2;
87 bb1:
88 condtmp' = *p;
89 bb2:
90 condtmp = PHI <RHS, condtmp'>
91 *p = condtmp;
93 This transformation can only be done under several constraints,
94 documented below. It also replaces:
96 bb0:
97 if (cond) goto bb2; else goto bb1;
98 bb1:
99 *p = RHS1;
100 goto bb3;
101 bb2:
102 *p = RHS2;
103 bb3:
105 with
107 bb0:
108 if (cond) goto bb3; else goto bb1;
109 bb1:
110 bb3:
111 condtmp = PHI <RHS1, RHS2>
112 *p = condtmp; */
114 static unsigned int
115 tree_ssa_cs_elim (void)
117 unsigned todo;
118 /* ??? We are not interested in loop related info, but the following
119 will create it, ICEing as we didn't init loops with pre-headers.
120 An interfacing issue of find_data_references_in_bb. */
121 loop_optimizer_init (LOOPS_NORMAL);
122 scev_initialize ();
123 todo = tree_ssa_phiopt_worker (true, false);
124 scev_finalize ();
125 loop_optimizer_finalize ();
126 return todo;
129 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
131 static gphi *
132 single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1)
134 gimple_stmt_iterator i;
135 gphi *phi = NULL;
136 if (gimple_seq_singleton_p (seq))
137 return as_a <gphi *> (gsi_stmt (gsi_start (seq)));
138 for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i))
140 gphi *p = as_a <gphi *> (gsi_stmt (i));
141 /* If the PHI arguments are equal then we can skip this PHI. */
142 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx),
143 gimple_phi_arg_def (p, e1->dest_idx)))
144 continue;
146 /* If we already have a PHI that has the two edge arguments are
147 different, then return it is not a singleton for these PHIs. */
148 if (phi)
149 return NULL;
151 phi = p;
153 return phi;
156 /* The core routine of conditional store replacement and normal
157 phi optimizations. Both share much of the infrastructure in how
158 to match applicable basic block patterns. DO_STORE_ELIM is true
159 when we want to do conditional store replacement, false otherwise.
160 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
161 of diamond control flow patterns, false otherwise. */
162 static unsigned int
163 tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads)
165 basic_block bb;
166 basic_block *bb_order;
167 unsigned n, i;
168 bool cfgchanged = false;
169 hash_set<tree> *nontrap = 0;
171 if (do_store_elim)
172 /* Calculate the set of non-trapping memory accesses. */
173 nontrap = get_non_trapping ();
175 /* Search every basic block for COND_EXPR we may be able to optimize.
177 We walk the blocks in order that guarantees that a block with
178 a single predecessor is processed before the predecessor.
179 This ensures that we collapse inner ifs before visiting the
180 outer ones, and also that we do not try to visit a removed
181 block. */
182 bb_order = single_pred_before_succ_order ();
183 n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS;
185 for (i = 0; i < n; i++)
187 gimple *cond_stmt;
188 gphi *phi;
189 basic_block bb1, bb2;
190 edge e1, e2;
191 tree arg0, arg1;
193 bb = bb_order[i];
195 cond_stmt = last_stmt (bb);
196 /* Check to see if the last statement is a GIMPLE_COND. */
197 if (!cond_stmt
198 || gimple_code (cond_stmt) != GIMPLE_COND)
199 continue;
201 e1 = EDGE_SUCC (bb, 0);
202 bb1 = e1->dest;
203 e2 = EDGE_SUCC (bb, 1);
204 bb2 = e2->dest;
206 /* We cannot do the optimization on abnormal edges. */
207 if ((e1->flags & EDGE_ABNORMAL) != 0
208 || (e2->flags & EDGE_ABNORMAL) != 0)
209 continue;
211 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
212 if (EDGE_COUNT (bb1->succs) == 0
213 || bb2 == NULL
214 || EDGE_COUNT (bb2->succs) == 0)
215 continue;
217 /* Find the bb which is the fall through to the other. */
218 if (EDGE_SUCC (bb1, 0)->dest == bb2)
220 else if (EDGE_SUCC (bb2, 0)->dest == bb1)
222 std::swap (bb1, bb2);
223 std::swap (e1, e2);
225 else if (do_store_elim
226 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
228 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
230 if (!single_succ_p (bb1)
231 || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0
232 || !single_succ_p (bb2)
233 || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0
234 || EDGE_COUNT (bb3->preds) != 2)
235 continue;
236 if (cond_if_else_store_replacement (bb1, bb2, bb3))
237 cfgchanged = true;
238 continue;
240 else if (do_hoist_loads
241 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
243 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
245 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt)))
246 && single_succ_p (bb1)
247 && single_succ_p (bb2)
248 && single_pred_p (bb1)
249 && single_pred_p (bb2)
250 && EDGE_COUNT (bb->succs) == 2
251 && EDGE_COUNT (bb3->preds) == 2
252 /* If one edge or the other is dominant, a conditional move
253 is likely to perform worse than the well-predicted branch. */
254 && !predictable_edge_p (EDGE_SUCC (bb, 0))
255 && !predictable_edge_p (EDGE_SUCC (bb, 1)))
256 hoist_adjacent_loads (bb, bb1, bb2, bb3);
257 continue;
259 else
260 continue;
262 e1 = EDGE_SUCC (bb1, 0);
264 /* Make sure that bb1 is just a fall through. */
265 if (!single_succ_p (bb1)
266 || (e1->flags & EDGE_FALLTHRU) == 0)
267 continue;
269 /* Also make sure that bb1 only have one predecessor and that it
270 is bb. */
271 if (!single_pred_p (bb1)
272 || single_pred (bb1) != bb)
273 continue;
275 if (do_store_elim)
277 /* bb1 is the middle block, bb2 the join block, bb the split block,
278 e1 the fallthrough edge from bb1 to bb2. We can't do the
279 optimization if the join block has more than two predecessors. */
280 if (EDGE_COUNT (bb2->preds) > 2)
281 continue;
282 if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
283 cfgchanged = true;
285 else
287 gimple_seq phis = phi_nodes (bb2);
288 gimple_stmt_iterator gsi;
289 bool candorest = true;
291 /* Value replacement can work with more than one PHI
292 so try that first. */
293 for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
295 phi = as_a <gphi *> (gsi_stmt (gsi));
296 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
297 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
298 if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2)
300 candorest = false;
301 cfgchanged = true;
302 break;
306 if (!candorest)
307 continue;
309 phi = single_non_singleton_phi_for_edges (phis, e1, e2);
310 if (!phi)
311 continue;
313 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
314 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
316 /* Something is wrong if we cannot find the arguments in the PHI
317 node. */
318 gcc_assert (arg0 != NULL && arg1 != NULL);
320 if (factor_out_conditional_conversion (e1, e2, phi, arg0, arg1))
322 /* factor_out_conditional_conversion may create a new PHI in
323 BB2 and eliminate an existing PHI in BB2. Recompute values
324 that may be affected by that change. */
325 phis = phi_nodes (bb2);
326 phi = single_non_singleton_phi_for_edges (phis, e1, e2);
327 gcc_assert (phi);
328 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
329 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
330 gcc_assert (arg0 != NULL && arg1 != NULL);
333 /* Do the replacement of conditional if it can be done. */
334 if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
335 cfgchanged = true;
336 else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
337 cfgchanged = true;
338 else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
339 cfgchanged = true;
343 free (bb_order);
345 if (do_store_elim)
346 delete nontrap;
347 /* If the CFG has changed, we should cleanup the CFG. */
348 if (cfgchanged && do_store_elim)
350 /* In cond-store replacement we have added some loads on edges
351 and new VOPS (as we moved the store, and created a load). */
352 gsi_commit_edge_inserts ();
353 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
355 else if (cfgchanged)
356 return TODO_cleanup_cfg;
357 return 0;
360 /* Replace PHI node element whose edge is E in block BB with variable NEW.
361 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
362 is known to have two edges, one of which must reach BB). */
364 static void
365 replace_phi_edge_with_variable (basic_block cond_block,
366 edge e, gimple *phi, tree new_tree)
368 basic_block bb = gimple_bb (phi);
369 basic_block block_to_remove;
370 gimple_stmt_iterator gsi;
372 /* Change the PHI argument to new. */
373 SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree);
375 /* Remove the empty basic block. */
376 if (EDGE_SUCC (cond_block, 0)->dest == bb)
378 EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU;
379 EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
380 EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE;
381 EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count;
383 block_to_remove = EDGE_SUCC (cond_block, 1)->dest;
385 else
387 EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU;
388 EDGE_SUCC (cond_block, 1)->flags
389 &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
390 EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE;
391 EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count;
393 block_to_remove = EDGE_SUCC (cond_block, 0)->dest;
395 delete_basic_block (block_to_remove);
397 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
398 gsi = gsi_last_bb (cond_block);
399 gsi_remove (&gsi, true);
401 if (dump_file && (dump_flags & TDF_DETAILS))
402 fprintf (dump_file,
403 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
404 cond_block->index,
405 bb->index);
408 /* PR66726: Factor conversion out of COND_EXPR. If the arguments of the PHI
409 stmt are CONVERT_STMT, factor out the conversion and perform the conversion
410 to the result of PHI stmt. */
412 static bool
413 factor_out_conditional_conversion (edge e0, edge e1, gphi *phi,
414 tree arg0, tree arg1)
416 gimple *arg0_def_stmt = NULL, *arg1_def_stmt = NULL, *new_stmt;
417 tree new_arg0 = NULL_TREE, new_arg1 = NULL_TREE;
418 tree temp, result;
419 gphi *newphi;
420 gimple_stmt_iterator gsi, gsi_for_def;
421 source_location locus = gimple_location (phi);
422 enum tree_code convert_code;
424 /* Handle only PHI statements with two arguments. TODO: If all
425 other arguments to PHI are INTEGER_CST or if their defining
426 statement have the same unary operation, we can handle more
427 than two arguments too. */
428 if (gimple_phi_num_args (phi) != 2)
429 return false;
431 /* First canonicalize to simplify tests. */
432 if (TREE_CODE (arg0) != SSA_NAME)
434 std::swap (arg0, arg1);
435 std::swap (e0, e1);
438 if (TREE_CODE (arg0) != SSA_NAME
439 || (TREE_CODE (arg1) != SSA_NAME
440 && TREE_CODE (arg1) != INTEGER_CST))
441 return false;
443 /* Check if arg0 is an SSA_NAME and the stmt which defines arg0 is
444 a conversion. */
445 arg0_def_stmt = SSA_NAME_DEF_STMT (arg0);
446 if (!is_gimple_assign (arg0_def_stmt)
447 || !gimple_assign_cast_p (arg0_def_stmt))
448 return false;
450 /* Use the RHS as new_arg0. */
451 convert_code = gimple_assign_rhs_code (arg0_def_stmt);
452 new_arg0 = gimple_assign_rhs1 (arg0_def_stmt);
453 if (convert_code == VIEW_CONVERT_EXPR)
454 new_arg0 = TREE_OPERAND (new_arg0, 0);
456 if (TREE_CODE (arg1) == SSA_NAME)
458 /* Check if arg1 is an SSA_NAME and the stmt which defines arg1
459 is a conversion. */
460 arg1_def_stmt = SSA_NAME_DEF_STMT (arg1);
461 if (!is_gimple_assign (arg1_def_stmt)
462 || gimple_assign_rhs_code (arg1_def_stmt) != convert_code)
463 return false;
465 /* Use the RHS as new_arg1. */
466 new_arg1 = gimple_assign_rhs1 (arg1_def_stmt);
467 if (convert_code == VIEW_CONVERT_EXPR)
468 new_arg1 = TREE_OPERAND (new_arg1, 0);
470 else
472 /* If arg1 is an INTEGER_CST, fold it to new type. */
473 if (INTEGRAL_TYPE_P (TREE_TYPE (new_arg0))
474 && int_fits_type_p (arg1, TREE_TYPE (new_arg0)))
476 if (gimple_assign_cast_p (arg0_def_stmt))
477 new_arg1 = fold_convert (TREE_TYPE (new_arg0), arg1);
478 else
479 return false;
481 else
482 return false;
485 /* If arg0/arg1 have > 1 use, then this transformation actually increases
486 the number of expressions evaluated at runtime. */
487 if (!has_single_use (arg0)
488 || (arg1_def_stmt && !has_single_use (arg1)))
489 return false;
491 /* If types of new_arg0 and new_arg1 are different bailout. */
492 if (!types_compatible_p (TREE_TYPE (new_arg0), TREE_TYPE (new_arg1)))
493 return false;
495 /* Create a new PHI stmt. */
496 result = PHI_RESULT (phi);
497 temp = make_ssa_name (TREE_TYPE (new_arg0), NULL);
498 newphi = create_phi_node (temp, gimple_bb (phi));
500 if (dump_file && (dump_flags & TDF_DETAILS))
502 fprintf (dump_file, "PHI ");
503 print_generic_expr (dump_file, gimple_phi_result (phi), 0);
504 fprintf (dump_file,
505 " changed to factor conversion out from COND_EXPR.\n");
506 fprintf (dump_file, "New stmt with CAST that defines ");
507 print_generic_expr (dump_file, result, 0);
508 fprintf (dump_file, ".\n");
511 /* Remove the old cast(s) that has single use. */
512 gsi_for_def = gsi_for_stmt (arg0_def_stmt);
513 gsi_remove (&gsi_for_def, true);
514 release_defs (arg0_def_stmt);
516 if (arg1_def_stmt)
518 gsi_for_def = gsi_for_stmt (arg1_def_stmt);
519 gsi_remove (&gsi_for_def, true);
520 release_defs (arg1_def_stmt);
523 add_phi_arg (newphi, new_arg0, e0, locus);
524 add_phi_arg (newphi, new_arg1, e1, locus);
526 /* Create the conversion stmt and insert it. */
527 if (convert_code == VIEW_CONVERT_EXPR)
528 temp = fold_build1 (VIEW_CONVERT_EXPR, TREE_TYPE (result), temp);
529 new_stmt = gimple_build_assign (result, convert_code, temp);
530 gsi = gsi_after_labels (gimple_bb (phi));
531 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
533 /* Remove the original PHI stmt. */
534 gsi = gsi_for_stmt (phi);
535 gsi_remove (&gsi, true);
536 return true;
539 /* The function conditional_replacement does the main work of doing the
540 conditional replacement. Return true if the replacement is done.
541 Otherwise return false.
542 BB is the basic block where the replacement is going to be done on. ARG0
543 is argument 0 from PHI. Likewise for ARG1. */
545 static bool
546 conditional_replacement (basic_block cond_bb, basic_block middle_bb,
547 edge e0, edge e1, gphi *phi,
548 tree arg0, tree arg1)
550 tree result;
551 gimple *stmt;
552 gassign *new_stmt;
553 tree cond;
554 gimple_stmt_iterator gsi;
555 edge true_edge, false_edge;
556 tree new_var, new_var2;
557 bool neg;
559 /* FIXME: Gimplification of complex type is too hard for now. */
560 /* We aren't prepared to handle vectors either (and it is a question
561 if it would be worthwhile anyway). */
562 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0))
563 || POINTER_TYPE_P (TREE_TYPE (arg0)))
564 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1))
565 || POINTER_TYPE_P (TREE_TYPE (arg1))))
566 return false;
568 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
569 convert it to the conditional. */
570 if ((integer_zerop (arg0) && integer_onep (arg1))
571 || (integer_zerop (arg1) && integer_onep (arg0)))
572 neg = false;
573 else if ((integer_zerop (arg0) && integer_all_onesp (arg1))
574 || (integer_zerop (arg1) && integer_all_onesp (arg0)))
575 neg = true;
576 else
577 return false;
579 if (!empty_block_p (middle_bb))
580 return false;
582 /* At this point we know we have a GIMPLE_COND with two successors.
583 One successor is BB, the other successor is an empty block which
584 falls through into BB.
586 There is a single PHI node at the join point (BB) and its arguments
587 are constants (0, 1) or (0, -1).
589 So, given the condition COND, and the two PHI arguments, we can
590 rewrite this PHI into non-branching code:
592 dest = (COND) or dest = COND'
594 We use the condition as-is if the argument associated with the
595 true edge has the value one or the argument associated with the
596 false edge as the value zero. Note that those conditions are not
597 the same since only one of the outgoing edges from the GIMPLE_COND
598 will directly reach BB and thus be associated with an argument. */
600 stmt = last_stmt (cond_bb);
601 result = PHI_RESULT (phi);
603 /* To handle special cases like floating point comparison, it is easier and
604 less error-prone to build a tree and gimplify it on the fly though it is
605 less efficient. */
606 cond = fold_build2_loc (gimple_location (stmt),
607 gimple_cond_code (stmt), boolean_type_node,
608 gimple_cond_lhs (stmt), gimple_cond_rhs (stmt));
610 /* We need to know which is the true edge and which is the false
611 edge so that we know when to invert the condition below. */
612 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
613 if ((e0 == true_edge && integer_zerop (arg0))
614 || (e0 == false_edge && !integer_zerop (arg0))
615 || (e1 == true_edge && integer_zerop (arg1))
616 || (e1 == false_edge && !integer_zerop (arg1)))
617 cond = fold_build1_loc (gimple_location (stmt),
618 TRUTH_NOT_EXPR, TREE_TYPE (cond), cond);
620 if (neg)
622 cond = fold_convert_loc (gimple_location (stmt),
623 TREE_TYPE (result), cond);
624 cond = fold_build1_loc (gimple_location (stmt),
625 NEGATE_EXPR, TREE_TYPE (cond), cond);
628 /* Insert our new statements at the end of conditional block before the
629 COND_STMT. */
630 gsi = gsi_for_stmt (stmt);
631 new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true,
632 GSI_SAME_STMT);
634 if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var)))
636 source_location locus_0, locus_1;
638 new_var2 = make_ssa_name (TREE_TYPE (result));
639 new_stmt = gimple_build_assign (new_var2, CONVERT_EXPR, new_var);
640 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
641 new_var = new_var2;
643 /* Set the locus to the first argument, unless is doesn't have one. */
644 locus_0 = gimple_phi_arg_location (phi, 0);
645 locus_1 = gimple_phi_arg_location (phi, 1);
646 if (locus_0 == UNKNOWN_LOCATION)
647 locus_0 = locus_1;
648 gimple_set_location (new_stmt, locus_0);
651 replace_phi_edge_with_variable (cond_bb, e1, phi, new_var);
652 reset_flow_sensitive_info_in_bb (cond_bb);
654 /* Note that we optimized this PHI. */
655 return true;
658 /* Update *ARG which is defined in STMT so that it contains the
659 computed value if that seems profitable. Return true if the
660 statement is made dead by that rewriting. */
662 static bool
663 jump_function_from_stmt (tree *arg, gimple *stmt)
665 enum tree_code code = gimple_assign_rhs_code (stmt);
666 if (code == ADDR_EXPR)
668 /* For arg = &p->i transform it to p, if possible. */
669 tree rhs1 = gimple_assign_rhs1 (stmt);
670 HOST_WIDE_INT offset;
671 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0),
672 &offset);
673 if (tem
674 && TREE_CODE (tem) == MEM_REF
675 && (mem_ref_offset (tem) + offset) == 0)
677 *arg = TREE_OPERAND (tem, 0);
678 return true;
681 /* TODO: Much like IPA-CP jump-functions we want to handle constant
682 additions symbolically here, and we'd need to update the comparison
683 code that compares the arg + cst tuples in our caller. For now the
684 code above exactly handles the VEC_BASE pattern from vec.h. */
685 return false;
688 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
689 of the form SSA_NAME NE 0.
691 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
692 the two input values of the EQ_EXPR match arg0 and arg1.
694 If so update *code and return TRUE. Otherwise return FALSE. */
696 static bool
697 rhs_is_fed_for_value_replacement (const_tree arg0, const_tree arg1,
698 enum tree_code *code, const_tree rhs)
700 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
701 statement. */
702 if (TREE_CODE (rhs) == SSA_NAME)
704 gimple *def1 = SSA_NAME_DEF_STMT (rhs);
706 /* Verify the defining statement has an EQ_EXPR on the RHS. */
707 if (is_gimple_assign (def1) && gimple_assign_rhs_code (def1) == EQ_EXPR)
709 /* Finally verify the source operands of the EQ_EXPR are equal
710 to arg0 and arg1. */
711 tree op0 = gimple_assign_rhs1 (def1);
712 tree op1 = gimple_assign_rhs2 (def1);
713 if ((operand_equal_for_phi_arg_p (arg0, op0)
714 && operand_equal_for_phi_arg_p (arg1, op1))
715 || (operand_equal_for_phi_arg_p (arg0, op1)
716 && operand_equal_for_phi_arg_p (arg1, op0)))
718 /* We will perform the optimization. */
719 *code = gimple_assign_rhs_code (def1);
720 return true;
724 return false;
727 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
729 Also return TRUE if arg0/arg1 are equal to the source arguments of a
730 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
732 Return FALSE otherwise. */
734 static bool
735 operand_equal_for_value_replacement (const_tree arg0, const_tree arg1,
736 enum tree_code *code, gimple *cond)
738 gimple *def;
739 tree lhs = gimple_cond_lhs (cond);
740 tree rhs = gimple_cond_rhs (cond);
742 if ((operand_equal_for_phi_arg_p (arg0, lhs)
743 && operand_equal_for_phi_arg_p (arg1, rhs))
744 || (operand_equal_for_phi_arg_p (arg1, lhs)
745 && operand_equal_for_phi_arg_p (arg0, rhs)))
746 return true;
748 /* Now handle more complex case where we have an EQ comparison
749 which feeds a BIT_AND_EXPR which feeds COND.
751 First verify that COND is of the form SSA_NAME NE 0. */
752 if (*code != NE_EXPR || !integer_zerop (rhs)
753 || TREE_CODE (lhs) != SSA_NAME)
754 return false;
756 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
757 def = SSA_NAME_DEF_STMT (lhs);
758 if (!is_gimple_assign (def) || gimple_assign_rhs_code (def) != BIT_AND_EXPR)
759 return false;
761 /* Now verify arg0/arg1 correspond to the source arguments of an
762 EQ comparison feeding the BIT_AND_EXPR. */
764 tree tmp = gimple_assign_rhs1 (def);
765 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
766 return true;
768 tmp = gimple_assign_rhs2 (def);
769 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
770 return true;
772 return false;
775 /* Returns true if ARG is a neutral element for operation CODE
776 on the RIGHT side. */
778 static bool
779 neutral_element_p (tree_code code, tree arg, bool right)
781 switch (code)
783 case PLUS_EXPR:
784 case BIT_IOR_EXPR:
785 case BIT_XOR_EXPR:
786 return integer_zerop (arg);
788 case LROTATE_EXPR:
789 case RROTATE_EXPR:
790 case LSHIFT_EXPR:
791 case RSHIFT_EXPR:
792 case MINUS_EXPR:
793 case POINTER_PLUS_EXPR:
794 return right && integer_zerop (arg);
796 case MULT_EXPR:
797 return integer_onep (arg);
799 case TRUNC_DIV_EXPR:
800 case CEIL_DIV_EXPR:
801 case FLOOR_DIV_EXPR:
802 case ROUND_DIV_EXPR:
803 case EXACT_DIV_EXPR:
804 return right && integer_onep (arg);
806 case BIT_AND_EXPR:
807 return integer_all_onesp (arg);
809 default:
810 return false;
814 /* Returns true if ARG is an absorbing element for operation CODE. */
816 static bool
817 absorbing_element_p (tree_code code, tree arg)
819 switch (code)
821 case BIT_IOR_EXPR:
822 return integer_all_onesp (arg);
824 case MULT_EXPR:
825 case BIT_AND_EXPR:
826 return integer_zerop (arg);
828 default:
829 return false;
833 /* The function value_replacement does the main work of doing the value
834 replacement. Return non-zero if the replacement is done. Otherwise return
835 0. If we remove the middle basic block, return 2.
836 BB is the basic block where the replacement is going to be done on. ARG0
837 is argument 0 from the PHI. Likewise for ARG1. */
839 static int
840 value_replacement (basic_block cond_bb, basic_block middle_bb,
841 edge e0, edge e1, gimple *phi,
842 tree arg0, tree arg1)
844 gimple_stmt_iterator gsi;
845 gimple *cond;
846 edge true_edge, false_edge;
847 enum tree_code code;
848 bool emtpy_or_with_defined_p = true;
850 /* If the type says honor signed zeros we cannot do this
851 optimization. */
852 if (HONOR_SIGNED_ZEROS (arg1))
853 return 0;
855 /* If there is a statement in MIDDLE_BB that defines one of the PHI
856 arguments, then adjust arg0 or arg1. */
857 gsi = gsi_start_nondebug_after_labels_bb (middle_bb);
858 while (!gsi_end_p (gsi))
860 gimple *stmt = gsi_stmt (gsi);
861 tree lhs;
862 gsi_next_nondebug (&gsi);
863 if (!is_gimple_assign (stmt))
865 emtpy_or_with_defined_p = false;
866 continue;
868 /* Now try to adjust arg0 or arg1 according to the computation
869 in the statement. */
870 lhs = gimple_assign_lhs (stmt);
871 if (!(lhs == arg0
872 && jump_function_from_stmt (&arg0, stmt))
873 || (lhs == arg1
874 && jump_function_from_stmt (&arg1, stmt)))
875 emtpy_or_with_defined_p = false;
878 cond = last_stmt (cond_bb);
879 code = gimple_cond_code (cond);
881 /* This transformation is only valid for equality comparisons. */
882 if (code != NE_EXPR && code != EQ_EXPR)
883 return 0;
885 /* We need to know which is the true edge and which is the false
886 edge so that we know if have abs or negative abs. */
887 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
889 /* At this point we know we have a COND_EXPR with two successors.
890 One successor is BB, the other successor is an empty block which
891 falls through into BB.
893 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
895 There is a single PHI node at the join point (BB) with two arguments.
897 We now need to verify that the two arguments in the PHI node match
898 the two arguments to the equality comparison. */
900 if (operand_equal_for_value_replacement (arg0, arg1, &code, cond))
902 edge e;
903 tree arg;
905 /* For NE_EXPR, we want to build an assignment result = arg where
906 arg is the PHI argument associated with the true edge. For
907 EQ_EXPR we want the PHI argument associated with the false edge. */
908 e = (code == NE_EXPR ? true_edge : false_edge);
910 /* Unfortunately, E may not reach BB (it may instead have gone to
911 OTHER_BLOCK). If that is the case, then we want the single outgoing
912 edge from OTHER_BLOCK which reaches BB and represents the desired
913 path from COND_BLOCK. */
914 if (e->dest == middle_bb)
915 e = single_succ_edge (e->dest);
917 /* Now we know the incoming edge to BB that has the argument for the
918 RHS of our new assignment statement. */
919 if (e0 == e)
920 arg = arg0;
921 else
922 arg = arg1;
924 /* If the middle basic block was empty or is defining the
925 PHI arguments and this is a single phi where the args are different
926 for the edges e0 and e1 then we can remove the middle basic block. */
927 if (emtpy_or_with_defined_p
928 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)),
929 e0, e1) == phi)
931 replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
932 /* Note that we optimized this PHI. */
933 return 2;
935 else
937 /* Replace the PHI arguments with arg. */
938 SET_PHI_ARG_DEF (phi, e0->dest_idx, arg);
939 SET_PHI_ARG_DEF (phi, e1->dest_idx, arg);
940 if (dump_file && (dump_flags & TDF_DETAILS))
942 fprintf (dump_file, "PHI ");
943 print_generic_expr (dump_file, gimple_phi_result (phi), 0);
944 fprintf (dump_file, " reduced for COND_EXPR in block %d to ",
945 cond_bb->index);
946 print_generic_expr (dump_file, arg, 0);
947 fprintf (dump_file, ".\n");
949 return 1;
954 /* Now optimize (x != 0) ? x + y : y to just y.
955 The following condition is too restrictive, there can easily be another
956 stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */
957 gimple *assign = last_and_only_stmt (middle_bb);
958 if (!assign || gimple_code (assign) != GIMPLE_ASSIGN
959 || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS
960 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0))
961 && !POINTER_TYPE_P (TREE_TYPE (arg0))))
962 return 0;
964 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
965 if (!gimple_seq_empty_p (phi_nodes (middle_bb)))
966 return 0;
968 /* Only transform if it removes the condition. */
969 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1))
970 return 0;
972 /* Size-wise, this is always profitable. */
973 if (optimize_bb_for_speed_p (cond_bb)
974 /* The special case is useless if it has a low probability. */
975 && profile_status_for_fn (cfun) != PROFILE_ABSENT
976 && EDGE_PRED (middle_bb, 0)->probability < PROB_EVEN
977 /* If assign is cheap, there is no point avoiding it. */
978 && estimate_num_insns (assign, &eni_time_weights)
979 >= 3 * estimate_num_insns (cond, &eni_time_weights))
980 return 0;
982 tree lhs = gimple_assign_lhs (assign);
983 tree rhs1 = gimple_assign_rhs1 (assign);
984 tree rhs2 = gimple_assign_rhs2 (assign);
985 enum tree_code code_def = gimple_assign_rhs_code (assign);
986 tree cond_lhs = gimple_cond_lhs (cond);
987 tree cond_rhs = gimple_cond_rhs (cond);
989 if (((code == NE_EXPR && e1 == false_edge)
990 || (code == EQ_EXPR && e1 == true_edge))
991 && arg0 == lhs
992 && ((arg1 == rhs1
993 && operand_equal_for_phi_arg_p (rhs2, cond_lhs)
994 && neutral_element_p (code_def, cond_rhs, true))
995 || (arg1 == rhs2
996 && operand_equal_for_phi_arg_p (rhs1, cond_lhs)
997 && neutral_element_p (code_def, cond_rhs, false))
998 || (operand_equal_for_phi_arg_p (arg1, cond_rhs)
999 && (operand_equal_for_phi_arg_p (rhs2, cond_lhs)
1000 || operand_equal_for_phi_arg_p (rhs1, cond_lhs))
1001 && absorbing_element_p (code_def, cond_rhs))))
1003 gsi = gsi_for_stmt (cond);
1004 if (INTEGRAL_TYPE_P (TREE_TYPE (lhs)))
1006 /* Moving ASSIGN might change VR of lhs, e.g. when moving u_6
1007 def-stmt in:
1008 if (n_5 != 0)
1009 goto <bb 3>;
1010 else
1011 goto <bb 4>;
1013 <bb 3>:
1014 # RANGE [0, 4294967294]
1015 u_6 = n_5 + 4294967295;
1017 <bb 4>:
1018 # u_3 = PHI <u_6(3), 4294967295(2)> */
1019 SSA_NAME_RANGE_INFO (lhs) = NULL;
1020 /* If available, we can use VR of phi result at least. */
1021 tree phires = gimple_phi_result (phi);
1022 struct range_info_def *phires_range_info
1023 = SSA_NAME_RANGE_INFO (phires);
1024 if (phires_range_info)
1025 duplicate_ssa_name_range_info (lhs, SSA_NAME_RANGE_TYPE (phires),
1026 phires_range_info);
1028 gimple_stmt_iterator gsi_from = gsi_for_stmt (assign);
1029 gsi_move_before (&gsi_from, &gsi);
1030 replace_phi_edge_with_variable (cond_bb, e1, phi, lhs);
1031 return 2;
1034 return 0;
1037 /* The function minmax_replacement does the main work of doing the minmax
1038 replacement. Return true if the replacement is done. Otherwise return
1039 false.
1040 BB is the basic block where the replacement is going to be done on. ARG0
1041 is argument 0 from the PHI. Likewise for ARG1. */
1043 static bool
1044 minmax_replacement (basic_block cond_bb, basic_block middle_bb,
1045 edge e0, edge e1, gimple *phi,
1046 tree arg0, tree arg1)
1048 tree result, type;
1049 gcond *cond;
1050 gassign *new_stmt;
1051 edge true_edge, false_edge;
1052 enum tree_code cmp, minmax, ass_code;
1053 tree smaller, larger, arg_true, arg_false;
1054 gimple_stmt_iterator gsi, gsi_from;
1056 type = TREE_TYPE (PHI_RESULT (phi));
1058 /* The optimization may be unsafe due to NaNs. */
1059 if (HONOR_NANS (type))
1060 return false;
1062 cond = as_a <gcond *> (last_stmt (cond_bb));
1063 cmp = gimple_cond_code (cond);
1065 /* This transformation is only valid for order comparisons. Record which
1066 operand is smaller/larger if the result of the comparison is true. */
1067 if (cmp == LT_EXPR || cmp == LE_EXPR)
1069 smaller = gimple_cond_lhs (cond);
1070 larger = gimple_cond_rhs (cond);
1072 else if (cmp == GT_EXPR || cmp == GE_EXPR)
1074 smaller = gimple_cond_rhs (cond);
1075 larger = gimple_cond_lhs (cond);
1077 else
1078 return false;
1080 /* We need to know which is the true edge and which is the false
1081 edge so that we know if have abs or negative abs. */
1082 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1084 /* Forward the edges over the middle basic block. */
1085 if (true_edge->dest == middle_bb)
1086 true_edge = EDGE_SUCC (true_edge->dest, 0);
1087 if (false_edge->dest == middle_bb)
1088 false_edge = EDGE_SUCC (false_edge->dest, 0);
1090 if (true_edge == e0)
1092 gcc_assert (false_edge == e1);
1093 arg_true = arg0;
1094 arg_false = arg1;
1096 else
1098 gcc_assert (false_edge == e0);
1099 gcc_assert (true_edge == e1);
1100 arg_true = arg1;
1101 arg_false = arg0;
1104 if (empty_block_p (middle_bb))
1106 if (operand_equal_for_phi_arg_p (arg_true, smaller)
1107 && operand_equal_for_phi_arg_p (arg_false, larger))
1109 /* Case
1111 if (smaller < larger)
1112 rslt = smaller;
1113 else
1114 rslt = larger; */
1115 minmax = MIN_EXPR;
1117 else if (operand_equal_for_phi_arg_p (arg_false, smaller)
1118 && operand_equal_for_phi_arg_p (arg_true, larger))
1119 minmax = MAX_EXPR;
1120 else
1121 return false;
1123 else
1125 /* Recognize the following case, assuming d <= u:
1127 if (a <= u)
1128 b = MAX (a, d);
1129 x = PHI <b, u>
1131 This is equivalent to
1133 b = MAX (a, d);
1134 x = MIN (b, u); */
1136 gimple *assign = last_and_only_stmt (middle_bb);
1137 tree lhs, op0, op1, bound;
1139 if (!assign
1140 || gimple_code (assign) != GIMPLE_ASSIGN)
1141 return false;
1143 lhs = gimple_assign_lhs (assign);
1144 ass_code = gimple_assign_rhs_code (assign);
1145 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
1146 return false;
1147 op0 = gimple_assign_rhs1 (assign);
1148 op1 = gimple_assign_rhs2 (assign);
1150 if (true_edge->src == middle_bb)
1152 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1153 if (!operand_equal_for_phi_arg_p (lhs, arg_true))
1154 return false;
1156 if (operand_equal_for_phi_arg_p (arg_false, larger))
1158 /* Case
1160 if (smaller < larger)
1162 r' = MAX_EXPR (smaller, bound)
1164 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1165 if (ass_code != MAX_EXPR)
1166 return false;
1168 minmax = MIN_EXPR;
1169 if (operand_equal_for_phi_arg_p (op0, smaller))
1170 bound = op1;
1171 else if (operand_equal_for_phi_arg_p (op1, smaller))
1172 bound = op0;
1173 else
1174 return false;
1176 /* We need BOUND <= LARGER. */
1177 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1178 bound, larger)))
1179 return false;
1181 else if (operand_equal_for_phi_arg_p (arg_false, smaller))
1183 /* Case
1185 if (smaller < larger)
1187 r' = MIN_EXPR (larger, bound)
1189 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1190 if (ass_code != MIN_EXPR)
1191 return false;
1193 minmax = MAX_EXPR;
1194 if (operand_equal_for_phi_arg_p (op0, larger))
1195 bound = op1;
1196 else if (operand_equal_for_phi_arg_p (op1, larger))
1197 bound = op0;
1198 else
1199 return false;
1201 /* We need BOUND >= SMALLER. */
1202 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1203 bound, smaller)))
1204 return false;
1206 else
1207 return false;
1209 else
1211 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1212 if (!operand_equal_for_phi_arg_p (lhs, arg_false))
1213 return false;
1215 if (operand_equal_for_phi_arg_p (arg_true, larger))
1217 /* Case
1219 if (smaller > larger)
1221 r' = MIN_EXPR (smaller, bound)
1223 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1224 if (ass_code != MIN_EXPR)
1225 return false;
1227 minmax = MAX_EXPR;
1228 if (operand_equal_for_phi_arg_p (op0, smaller))
1229 bound = op1;
1230 else if (operand_equal_for_phi_arg_p (op1, smaller))
1231 bound = op0;
1232 else
1233 return false;
1235 /* We need BOUND >= LARGER. */
1236 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1237 bound, larger)))
1238 return false;
1240 else if (operand_equal_for_phi_arg_p (arg_true, smaller))
1242 /* Case
1244 if (smaller > larger)
1246 r' = MAX_EXPR (larger, bound)
1248 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1249 if (ass_code != MAX_EXPR)
1250 return false;
1252 minmax = MIN_EXPR;
1253 if (operand_equal_for_phi_arg_p (op0, larger))
1254 bound = op1;
1255 else if (operand_equal_for_phi_arg_p (op1, larger))
1256 bound = op0;
1257 else
1258 return false;
1260 /* We need BOUND <= SMALLER. */
1261 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1262 bound, smaller)))
1263 return false;
1265 else
1266 return false;
1269 /* Move the statement from the middle block. */
1270 gsi = gsi_last_bb (cond_bb);
1271 gsi_from = gsi_last_nondebug_bb (middle_bb);
1272 gsi_move_before (&gsi_from, &gsi);
1275 /* Create an SSA var to hold the min/max result. If we're the only
1276 things setting the target PHI, then we can clone the PHI
1277 variable. Otherwise we must create a new one. */
1278 result = PHI_RESULT (phi);
1279 if (EDGE_COUNT (gimple_bb (phi)->preds) == 2)
1280 result = duplicate_ssa_name (result, NULL);
1281 else
1282 result = make_ssa_name (TREE_TYPE (result));
1284 /* Emit the statement to compute min/max. */
1285 new_stmt = gimple_build_assign (result, minmax, arg0, arg1);
1286 gsi = gsi_last_bb (cond_bb);
1287 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1289 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1290 reset_flow_sensitive_info_in_bb (cond_bb);
1292 return true;
1295 /* The function absolute_replacement does the main work of doing the absolute
1296 replacement. Return true if the replacement is done. Otherwise return
1297 false.
1298 bb is the basic block where the replacement is going to be done on. arg0
1299 is argument 0 from the phi. Likewise for arg1. */
1301 static bool
1302 abs_replacement (basic_block cond_bb, basic_block middle_bb,
1303 edge e0 ATTRIBUTE_UNUSED, edge e1,
1304 gimple *phi, tree arg0, tree arg1)
1306 tree result;
1307 gassign *new_stmt;
1308 gimple *cond;
1309 gimple_stmt_iterator gsi;
1310 edge true_edge, false_edge;
1311 gimple *assign;
1312 edge e;
1313 tree rhs, lhs;
1314 bool negate;
1315 enum tree_code cond_code;
1317 /* If the type says honor signed zeros we cannot do this
1318 optimization. */
1319 if (HONOR_SIGNED_ZEROS (arg1))
1320 return false;
1322 /* OTHER_BLOCK must have only one executable statement which must have the
1323 form arg0 = -arg1 or arg1 = -arg0. */
1325 assign = last_and_only_stmt (middle_bb);
1326 /* If we did not find the proper negation assignment, then we can not
1327 optimize. */
1328 if (assign == NULL)
1329 return false;
1331 /* If we got here, then we have found the only executable statement
1332 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1333 arg1 = -arg0, then we can not optimize. */
1334 if (gimple_code (assign) != GIMPLE_ASSIGN)
1335 return false;
1337 lhs = gimple_assign_lhs (assign);
1339 if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
1340 return false;
1342 rhs = gimple_assign_rhs1 (assign);
1344 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1345 if (!(lhs == arg0 && rhs == arg1)
1346 && !(lhs == arg1 && rhs == arg0))
1347 return false;
1349 cond = last_stmt (cond_bb);
1350 result = PHI_RESULT (phi);
1352 /* Only relationals comparing arg[01] against zero are interesting. */
1353 cond_code = gimple_cond_code (cond);
1354 if (cond_code != GT_EXPR && cond_code != GE_EXPR
1355 && cond_code != LT_EXPR && cond_code != LE_EXPR)
1356 return false;
1358 /* Make sure the conditional is arg[01] OP y. */
1359 if (gimple_cond_lhs (cond) != rhs)
1360 return false;
1362 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond)))
1363 ? real_zerop (gimple_cond_rhs (cond))
1364 : integer_zerop (gimple_cond_rhs (cond)))
1366 else
1367 return false;
1369 /* We need to know which is the true edge and which is the false
1370 edge so that we know if have abs or negative abs. */
1371 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1373 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1374 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1375 the false edge goes to OTHER_BLOCK. */
1376 if (cond_code == GT_EXPR || cond_code == GE_EXPR)
1377 e = true_edge;
1378 else
1379 e = false_edge;
1381 if (e->dest == middle_bb)
1382 negate = true;
1383 else
1384 negate = false;
1386 result = duplicate_ssa_name (result, NULL);
1388 if (negate)
1389 lhs = make_ssa_name (TREE_TYPE (result));
1390 else
1391 lhs = result;
1393 /* Build the modify expression with abs expression. */
1394 new_stmt = gimple_build_assign (lhs, ABS_EXPR, rhs);
1396 gsi = gsi_last_bb (cond_bb);
1397 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1399 if (negate)
1401 /* Get the right GSI. We want to insert after the recently
1402 added ABS_EXPR statement (which we know is the first statement
1403 in the block. */
1404 new_stmt = gimple_build_assign (result, NEGATE_EXPR, lhs);
1406 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1409 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1410 reset_flow_sensitive_info_in_bb (cond_bb);
1412 /* Note that we optimized this PHI. */
1413 return true;
1416 /* Auxiliary functions to determine the set of memory accesses which
1417 can't trap because they are preceded by accesses to the same memory
1418 portion. We do that for MEM_REFs, so we only need to track
1419 the SSA_NAME of the pointer indirectly referenced. The algorithm
1420 simply is a walk over all instructions in dominator order. When
1421 we see an MEM_REF we determine if we've already seen a same
1422 ref anywhere up to the root of the dominator tree. If we do the
1423 current access can't trap. If we don't see any dominating access
1424 the current access might trap, but might also make later accesses
1425 non-trapping, so we remember it. We need to be careful with loads
1426 or stores, for instance a load might not trap, while a store would,
1427 so if we see a dominating read access this doesn't mean that a later
1428 write access would not trap. Hence we also need to differentiate the
1429 type of access(es) seen.
1431 ??? We currently are very conservative and assume that a load might
1432 trap even if a store doesn't (write-only memory). This probably is
1433 overly conservative. */
1435 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1436 through it was seen, which would constitute a no-trap region for
1437 same accesses. */
1438 struct name_to_bb
1440 unsigned int ssa_name_ver;
1441 unsigned int phase;
1442 bool store;
1443 HOST_WIDE_INT offset, size;
1444 basic_block bb;
1447 /* Hashtable helpers. */
1449 struct ssa_names_hasher : free_ptr_hash <name_to_bb>
1451 static inline hashval_t hash (const name_to_bb *);
1452 static inline bool equal (const name_to_bb *, const name_to_bb *);
1455 /* Used for quick clearing of the hash-table when we see calls.
1456 Hash entries with phase < nt_call_phase are invalid. */
1457 static unsigned int nt_call_phase;
1459 /* The hash function. */
1461 inline hashval_t
1462 ssa_names_hasher::hash (const name_to_bb *n)
1464 return n->ssa_name_ver ^ (((hashval_t) n->store) << 31)
1465 ^ (n->offset << 6) ^ (n->size << 3);
1468 /* The equality function of *P1 and *P2. */
1470 inline bool
1471 ssa_names_hasher::equal (const name_to_bb *n1, const name_to_bb *n2)
1473 return n1->ssa_name_ver == n2->ssa_name_ver
1474 && n1->store == n2->store
1475 && n1->offset == n2->offset
1476 && n1->size == n2->size;
1479 class nontrapping_dom_walker : public dom_walker
1481 public:
1482 nontrapping_dom_walker (cdi_direction direction, hash_set<tree> *ps)
1483 : dom_walker (direction), m_nontrapping (ps), m_seen_ssa_names (128) {}
1485 virtual void before_dom_children (basic_block);
1486 virtual void after_dom_children (basic_block);
1488 private:
1490 /* We see the expression EXP in basic block BB. If it's an interesting
1491 expression (an MEM_REF through an SSA_NAME) possibly insert the
1492 expression into the set NONTRAP or the hash table of seen expressions.
1493 STORE is true if this expression is on the LHS, otherwise it's on
1494 the RHS. */
1495 void add_or_mark_expr (basic_block, tree, bool);
1497 hash_set<tree> *m_nontrapping;
1499 /* The hash table for remembering what we've seen. */
1500 hash_table<ssa_names_hasher> m_seen_ssa_names;
1503 /* Called by walk_dominator_tree, when entering the block BB. */
1504 void
1505 nontrapping_dom_walker::before_dom_children (basic_block bb)
1507 edge e;
1508 edge_iterator ei;
1509 gimple_stmt_iterator gsi;
1511 /* If we haven't seen all our predecessors, clear the hash-table. */
1512 FOR_EACH_EDGE (e, ei, bb->preds)
1513 if ((((size_t)e->src->aux) & 2) == 0)
1515 nt_call_phase++;
1516 break;
1519 /* Mark this BB as being on the path to dominator root and as visited. */
1520 bb->aux = (void*)(1 | 2);
1522 /* And walk the statements in order. */
1523 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1525 gimple *stmt = gsi_stmt (gsi);
1527 if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt))
1528 nt_call_phase++;
1529 else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt))
1531 add_or_mark_expr (bb, gimple_assign_lhs (stmt), true);
1532 add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), false);
1537 /* Called by walk_dominator_tree, when basic block BB is exited. */
1538 void
1539 nontrapping_dom_walker::after_dom_children (basic_block bb)
1541 /* This BB isn't on the path to dominator root anymore. */
1542 bb->aux = (void*)2;
1545 /* We see the expression EXP in basic block BB. If it's an interesting
1546 expression (an MEM_REF through an SSA_NAME) possibly insert the
1547 expression into the set NONTRAP or the hash table of seen expressions.
1548 STORE is true if this expression is on the LHS, otherwise it's on
1549 the RHS. */
1550 void
1551 nontrapping_dom_walker::add_or_mark_expr (basic_block bb, tree exp, bool store)
1553 HOST_WIDE_INT size;
1555 if (TREE_CODE (exp) == MEM_REF
1556 && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME
1557 && tree_fits_shwi_p (TREE_OPERAND (exp, 1))
1558 && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0)
1560 tree name = TREE_OPERAND (exp, 0);
1561 struct name_to_bb map;
1562 name_to_bb **slot;
1563 struct name_to_bb *n2bb;
1564 basic_block found_bb = 0;
1566 /* Try to find the last seen MEM_REF through the same
1567 SSA_NAME, which can trap. */
1568 map.ssa_name_ver = SSA_NAME_VERSION (name);
1569 map.phase = 0;
1570 map.bb = 0;
1571 map.store = store;
1572 map.offset = tree_to_shwi (TREE_OPERAND (exp, 1));
1573 map.size = size;
1575 slot = m_seen_ssa_names.find_slot (&map, INSERT);
1576 n2bb = *slot;
1577 if (n2bb && n2bb->phase >= nt_call_phase)
1578 found_bb = n2bb->bb;
1580 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1581 (it's in a basic block on the path from us to the dominator root)
1582 then we can't trap. */
1583 if (found_bb && (((size_t)found_bb->aux) & 1) == 1)
1585 m_nontrapping->add (exp);
1587 else
1589 /* EXP might trap, so insert it into the hash table. */
1590 if (n2bb)
1592 n2bb->phase = nt_call_phase;
1593 n2bb->bb = bb;
1595 else
1597 n2bb = XNEW (struct name_to_bb);
1598 n2bb->ssa_name_ver = SSA_NAME_VERSION (name);
1599 n2bb->phase = nt_call_phase;
1600 n2bb->bb = bb;
1601 n2bb->store = store;
1602 n2bb->offset = map.offset;
1603 n2bb->size = size;
1604 *slot = n2bb;
1610 /* This is the entry point of gathering non trapping memory accesses.
1611 It will do a dominator walk over the whole function, and it will
1612 make use of the bb->aux pointers. It returns a set of trees
1613 (the MEM_REFs itself) which can't trap. */
1614 static hash_set<tree> *
1615 get_non_trapping (void)
1617 nt_call_phase = 0;
1618 hash_set<tree> *nontrap = new hash_set<tree>;
1619 /* We're going to do a dominator walk, so ensure that we have
1620 dominance information. */
1621 calculate_dominance_info (CDI_DOMINATORS);
1623 nontrapping_dom_walker (CDI_DOMINATORS, nontrap)
1624 .walk (cfun->cfg->x_entry_block_ptr);
1626 clear_aux_for_blocks ();
1627 return nontrap;
1630 /* Do the main work of conditional store replacement. We already know
1631 that the recognized pattern looks like so:
1633 split:
1634 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1635 MIDDLE_BB:
1636 something
1637 fallthrough (edge E0)
1638 JOIN_BB:
1639 some more
1641 We check that MIDDLE_BB contains only one store, that that store
1642 doesn't trap (not via NOTRAP, but via checking if an access to the same
1643 memory location dominates us) and that the store has a "simple" RHS. */
1645 static bool
1646 cond_store_replacement (basic_block middle_bb, basic_block join_bb,
1647 edge e0, edge e1, hash_set<tree> *nontrap)
1649 gimple *assign = last_and_only_stmt (middle_bb);
1650 tree lhs, rhs, name, name2;
1651 gphi *newphi;
1652 gassign *new_stmt;
1653 gimple_stmt_iterator gsi;
1654 source_location locus;
1656 /* Check if middle_bb contains of only one store. */
1657 if (!assign
1658 || !gimple_assign_single_p (assign)
1659 || gimple_has_volatile_ops (assign))
1660 return false;
1662 locus = gimple_location (assign);
1663 lhs = gimple_assign_lhs (assign);
1664 rhs = gimple_assign_rhs1 (assign);
1665 if (TREE_CODE (lhs) != MEM_REF
1666 || TREE_CODE (TREE_OPERAND (lhs, 0)) != SSA_NAME
1667 || !is_gimple_reg_type (TREE_TYPE (lhs)))
1668 return false;
1670 /* Prove that we can move the store down. We could also check
1671 TREE_THIS_NOTRAP here, but in that case we also could move stores,
1672 whose value is not available readily, which we want to avoid. */
1673 if (!nontrap->contains (lhs))
1674 return false;
1676 /* Now we've checked the constraints, so do the transformation:
1677 1) Remove the single store. */
1678 gsi = gsi_for_stmt (assign);
1679 unlink_stmt_vdef (assign);
1680 gsi_remove (&gsi, true);
1681 release_defs (assign);
1683 /* 2) Insert a load from the memory of the store to the temporary
1684 on the edge which did not contain the store. */
1685 lhs = unshare_expr (lhs);
1686 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1687 new_stmt = gimple_build_assign (name, lhs);
1688 gimple_set_location (new_stmt, locus);
1689 gsi_insert_on_edge (e1, new_stmt);
1691 /* 3) Create a PHI node at the join block, with one argument
1692 holding the old RHS, and the other holding the temporary
1693 where we stored the old memory contents. */
1694 name2 = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1695 newphi = create_phi_node (name2, join_bb);
1696 add_phi_arg (newphi, rhs, e0, locus);
1697 add_phi_arg (newphi, name, e1, locus);
1699 lhs = unshare_expr (lhs);
1700 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1702 /* 4) Insert that PHI node. */
1703 gsi = gsi_after_labels (join_bb);
1704 if (gsi_end_p (gsi))
1706 gsi = gsi_last_bb (join_bb);
1707 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1709 else
1710 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1712 return true;
1715 /* Do the main work of conditional store replacement. */
1717 static bool
1718 cond_if_else_store_replacement_1 (basic_block then_bb, basic_block else_bb,
1719 basic_block join_bb, gimple *then_assign,
1720 gimple *else_assign)
1722 tree lhs_base, lhs, then_rhs, else_rhs, name;
1723 source_location then_locus, else_locus;
1724 gimple_stmt_iterator gsi;
1725 gphi *newphi;
1726 gassign *new_stmt;
1728 if (then_assign == NULL
1729 || !gimple_assign_single_p (then_assign)
1730 || gimple_clobber_p (then_assign)
1731 || gimple_has_volatile_ops (then_assign)
1732 || else_assign == NULL
1733 || !gimple_assign_single_p (else_assign)
1734 || gimple_clobber_p (else_assign)
1735 || gimple_has_volatile_ops (else_assign))
1736 return false;
1738 lhs = gimple_assign_lhs (then_assign);
1739 if (!is_gimple_reg_type (TREE_TYPE (lhs))
1740 || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0))
1741 return false;
1743 lhs_base = get_base_address (lhs);
1744 if (lhs_base == NULL_TREE
1745 || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF))
1746 return false;
1748 then_rhs = gimple_assign_rhs1 (then_assign);
1749 else_rhs = gimple_assign_rhs1 (else_assign);
1750 then_locus = gimple_location (then_assign);
1751 else_locus = gimple_location (else_assign);
1753 /* Now we've checked the constraints, so do the transformation:
1754 1) Remove the stores. */
1755 gsi = gsi_for_stmt (then_assign);
1756 unlink_stmt_vdef (then_assign);
1757 gsi_remove (&gsi, true);
1758 release_defs (then_assign);
1760 gsi = gsi_for_stmt (else_assign);
1761 unlink_stmt_vdef (else_assign);
1762 gsi_remove (&gsi, true);
1763 release_defs (else_assign);
1765 /* 2) Create a PHI node at the join block, with one argument
1766 holding the old RHS, and the other holding the temporary
1767 where we stored the old memory contents. */
1768 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1769 newphi = create_phi_node (name, join_bb);
1770 add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus);
1771 add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus);
1773 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1775 /* 3) Insert that PHI node. */
1776 gsi = gsi_after_labels (join_bb);
1777 if (gsi_end_p (gsi))
1779 gsi = gsi_last_bb (join_bb);
1780 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1782 else
1783 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1785 return true;
1788 /* Conditional store replacement. We already know
1789 that the recognized pattern looks like so:
1791 split:
1792 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1793 THEN_BB:
1795 X = Y;
1797 goto JOIN_BB;
1798 ELSE_BB:
1800 X = Z;
1802 fallthrough (edge E0)
1803 JOIN_BB:
1804 some more
1806 We check that it is safe to sink the store to JOIN_BB by verifying that
1807 there are no read-after-write or write-after-write dependencies in
1808 THEN_BB and ELSE_BB. */
1810 static bool
1811 cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
1812 basic_block join_bb)
1814 gimple *then_assign = last_and_only_stmt (then_bb);
1815 gimple *else_assign = last_and_only_stmt (else_bb);
1816 vec<data_reference_p> then_datarefs, else_datarefs;
1817 vec<ddr_p> then_ddrs, else_ddrs;
1818 gimple *then_store, *else_store;
1819 bool found, ok = false, res;
1820 struct data_dependence_relation *ddr;
1821 data_reference_p then_dr, else_dr;
1822 int i, j;
1823 tree then_lhs, else_lhs;
1824 basic_block blocks[3];
1826 if (MAX_STORES_TO_SINK == 0)
1827 return false;
1829 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1830 if (then_assign && else_assign)
1831 return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1832 then_assign, else_assign);
1834 /* Find data references. */
1835 then_datarefs.create (1);
1836 else_datarefs.create (1);
1837 if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs)
1838 == chrec_dont_know)
1839 || !then_datarefs.length ()
1840 || (find_data_references_in_bb (NULL, else_bb, &else_datarefs)
1841 == chrec_dont_know)
1842 || !else_datarefs.length ())
1844 free_data_refs (then_datarefs);
1845 free_data_refs (else_datarefs);
1846 return false;
1849 /* Find pairs of stores with equal LHS. */
1850 auto_vec<gimple *, 1> then_stores, else_stores;
1851 FOR_EACH_VEC_ELT (then_datarefs, i, then_dr)
1853 if (DR_IS_READ (then_dr))
1854 continue;
1856 then_store = DR_STMT (then_dr);
1857 then_lhs = gimple_get_lhs (then_store);
1858 if (then_lhs == NULL_TREE)
1859 continue;
1860 found = false;
1862 FOR_EACH_VEC_ELT (else_datarefs, j, else_dr)
1864 if (DR_IS_READ (else_dr))
1865 continue;
1867 else_store = DR_STMT (else_dr);
1868 else_lhs = gimple_get_lhs (else_store);
1869 if (else_lhs == NULL_TREE)
1870 continue;
1872 if (operand_equal_p (then_lhs, else_lhs, 0))
1874 found = true;
1875 break;
1879 if (!found)
1880 continue;
1882 then_stores.safe_push (then_store);
1883 else_stores.safe_push (else_store);
1886 /* No pairs of stores found. */
1887 if (!then_stores.length ()
1888 || then_stores.length () > (unsigned) MAX_STORES_TO_SINK)
1890 free_data_refs (then_datarefs);
1891 free_data_refs (else_datarefs);
1892 return false;
1895 /* Compute and check data dependencies in both basic blocks. */
1896 then_ddrs.create (1);
1897 else_ddrs.create (1);
1898 if (!compute_all_dependences (then_datarefs, &then_ddrs,
1899 vNULL, false)
1900 || !compute_all_dependences (else_datarefs, &else_ddrs,
1901 vNULL, false))
1903 free_dependence_relations (then_ddrs);
1904 free_dependence_relations (else_ddrs);
1905 free_data_refs (then_datarefs);
1906 free_data_refs (else_datarefs);
1907 return false;
1909 blocks[0] = then_bb;
1910 blocks[1] = else_bb;
1911 blocks[2] = join_bb;
1912 renumber_gimple_stmt_uids_in_blocks (blocks, 3);
1914 /* Check that there are no read-after-write or write-after-write dependencies
1915 in THEN_BB. */
1916 FOR_EACH_VEC_ELT (then_ddrs, i, ddr)
1918 struct data_reference *dra = DDR_A (ddr);
1919 struct data_reference *drb = DDR_B (ddr);
1921 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1922 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1923 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1924 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1925 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1926 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1928 free_dependence_relations (then_ddrs);
1929 free_dependence_relations (else_ddrs);
1930 free_data_refs (then_datarefs);
1931 free_data_refs (else_datarefs);
1932 return false;
1936 /* Check that there are no read-after-write or write-after-write dependencies
1937 in ELSE_BB. */
1938 FOR_EACH_VEC_ELT (else_ddrs, i, ddr)
1940 struct data_reference *dra = DDR_A (ddr);
1941 struct data_reference *drb = DDR_B (ddr);
1943 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1944 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1945 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1946 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1947 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1948 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1950 free_dependence_relations (then_ddrs);
1951 free_dependence_relations (else_ddrs);
1952 free_data_refs (then_datarefs);
1953 free_data_refs (else_datarefs);
1954 return false;
1958 /* Sink stores with same LHS. */
1959 FOR_EACH_VEC_ELT (then_stores, i, then_store)
1961 else_store = else_stores[i];
1962 res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1963 then_store, else_store);
1964 ok = ok || res;
1967 free_dependence_relations (then_ddrs);
1968 free_dependence_relations (else_ddrs);
1969 free_data_refs (then_datarefs);
1970 free_data_refs (else_datarefs);
1972 return ok;
1975 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1977 static bool
1978 local_mem_dependence (gimple *stmt, basic_block bb)
1980 tree vuse = gimple_vuse (stmt);
1981 gimple *def;
1983 if (!vuse)
1984 return false;
1986 def = SSA_NAME_DEF_STMT (vuse);
1987 return (def && gimple_bb (def) == bb);
1990 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1991 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1992 and BB3 rejoins control flow following BB1 and BB2, look for
1993 opportunities to hoist loads as follows. If BB3 contains a PHI of
1994 two loads, one each occurring in BB1 and BB2, and the loads are
1995 provably of adjacent fields in the same structure, then move both
1996 loads into BB0. Of course this can only be done if there are no
1997 dependencies preventing such motion.
1999 One of the hoisted loads will always be speculative, so the
2000 transformation is currently conservative:
2002 - The fields must be strictly adjacent.
2003 - The two fields must occupy a single memory block that is
2004 guaranteed to not cross a page boundary.
2006 The last is difficult to prove, as such memory blocks should be
2007 aligned on the minimum of the stack alignment boundary and the
2008 alignment guaranteed by heap allocation interfaces. Thus we rely
2009 on a parameter for the alignment value.
2011 Provided a good value is used for the last case, the first
2012 restriction could possibly be relaxed. */
2014 static void
2015 hoist_adjacent_loads (basic_block bb0, basic_block bb1,
2016 basic_block bb2, basic_block bb3)
2018 int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE);
2019 unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT);
2020 gphi_iterator gsi;
2022 /* Walk the phis in bb3 looking for an opportunity. We are looking
2023 for phis of two SSA names, one each of which is defined in bb1 and
2024 bb2. */
2025 for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi))
2027 gphi *phi_stmt = gsi.phi ();
2028 gimple *def1, *def2;
2029 tree arg1, arg2, ref1, ref2, field1, field2;
2030 tree tree_offset1, tree_offset2, tree_size2, next;
2031 int offset1, offset2, size2;
2032 unsigned align1;
2033 gimple_stmt_iterator gsi2;
2034 basic_block bb_for_def1, bb_for_def2;
2036 if (gimple_phi_num_args (phi_stmt) != 2
2037 || virtual_operand_p (gimple_phi_result (phi_stmt)))
2038 continue;
2040 arg1 = gimple_phi_arg_def (phi_stmt, 0);
2041 arg2 = gimple_phi_arg_def (phi_stmt, 1);
2043 if (TREE_CODE (arg1) != SSA_NAME
2044 || TREE_CODE (arg2) != SSA_NAME
2045 || SSA_NAME_IS_DEFAULT_DEF (arg1)
2046 || SSA_NAME_IS_DEFAULT_DEF (arg2))
2047 continue;
2049 def1 = SSA_NAME_DEF_STMT (arg1);
2050 def2 = SSA_NAME_DEF_STMT (arg2);
2052 if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2)
2053 && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2))
2054 continue;
2056 /* Check the mode of the arguments to be sure a conditional move
2057 can be generated for it. */
2058 if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1)))
2059 == CODE_FOR_nothing)
2060 continue;
2062 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2063 if (!gimple_assign_single_p (def1)
2064 || !gimple_assign_single_p (def2)
2065 || gimple_has_volatile_ops (def1)
2066 || gimple_has_volatile_ops (def2))
2067 continue;
2069 ref1 = gimple_assign_rhs1 (def1);
2070 ref2 = gimple_assign_rhs1 (def2);
2072 if (TREE_CODE (ref1) != COMPONENT_REF
2073 || TREE_CODE (ref2) != COMPONENT_REF)
2074 continue;
2076 /* The zeroth operand of the two component references must be
2077 identical. It is not sufficient to compare get_base_address of
2078 the two references, because this could allow for different
2079 elements of the same array in the two trees. It is not safe to
2080 assume that the existence of one array element implies the
2081 existence of a different one. */
2082 if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0))
2083 continue;
2085 field1 = TREE_OPERAND (ref1, 1);
2086 field2 = TREE_OPERAND (ref2, 1);
2088 /* Check for field adjacency, and ensure field1 comes first. */
2089 for (next = DECL_CHAIN (field1);
2090 next && TREE_CODE (next) != FIELD_DECL;
2091 next = DECL_CHAIN (next))
2094 if (next != field2)
2096 for (next = DECL_CHAIN (field2);
2097 next && TREE_CODE (next) != FIELD_DECL;
2098 next = DECL_CHAIN (next))
2101 if (next != field1)
2102 continue;
2104 std::swap (field1, field2);
2105 std::swap (def1, def2);
2108 bb_for_def1 = gimple_bb (def1);
2109 bb_for_def2 = gimple_bb (def2);
2111 /* Check for proper alignment of the first field. */
2112 tree_offset1 = bit_position (field1);
2113 tree_offset2 = bit_position (field2);
2114 tree_size2 = DECL_SIZE (field2);
2116 if (!tree_fits_uhwi_p (tree_offset1)
2117 || !tree_fits_uhwi_p (tree_offset2)
2118 || !tree_fits_uhwi_p (tree_size2))
2119 continue;
2121 offset1 = tree_to_uhwi (tree_offset1);
2122 offset2 = tree_to_uhwi (tree_offset2);
2123 size2 = tree_to_uhwi (tree_size2);
2124 align1 = DECL_ALIGN (field1) % param_align_bits;
2126 if (offset1 % BITS_PER_UNIT != 0)
2127 continue;
2129 /* For profitability, the two field references should fit within
2130 a single cache line. */
2131 if (align1 + offset2 - offset1 + size2 > param_align_bits)
2132 continue;
2134 /* The two expressions cannot be dependent upon vdefs defined
2135 in bb1/bb2. */
2136 if (local_mem_dependence (def1, bb_for_def1)
2137 || local_mem_dependence (def2, bb_for_def2))
2138 continue;
2140 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2141 bb0. We hoist the first one first so that a cache miss is handled
2142 efficiently regardless of hardware cache-fill policy. */
2143 gsi2 = gsi_for_stmt (def1);
2144 gsi_move_to_bb_end (&gsi2, bb0);
2145 gsi2 = gsi_for_stmt (def2);
2146 gsi_move_to_bb_end (&gsi2, bb0);
2148 if (dump_file && (dump_flags & TDF_DETAILS))
2150 fprintf (dump_file,
2151 "\nHoisting adjacent loads from %d and %d into %d: \n",
2152 bb_for_def1->index, bb_for_def2->index, bb0->index);
2153 print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS);
2154 print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS);
2159 /* Determine whether we should attempt to hoist adjacent loads out of
2160 diamond patterns in pass_phiopt. Always hoist loads if
2161 -fhoist-adjacent-loads is specified and the target machine has
2162 both a conditional move instruction and a defined cache line size. */
2164 static bool
2165 gate_hoist_loads (void)
2167 return (flag_hoist_adjacent_loads == 1
2168 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE)
2169 && HAVE_conditional_move);
2172 /* This pass tries to replaces an if-then-else block with an
2173 assignment. We have four kinds of transformations. Some of these
2174 transformations are also performed by the ifcvt RTL optimizer.
2176 Conditional Replacement
2177 -----------------------
2179 This transformation, implemented in conditional_replacement,
2180 replaces
2182 bb0:
2183 if (cond) goto bb2; else goto bb1;
2184 bb1:
2185 bb2:
2186 x = PHI <0 (bb1), 1 (bb0), ...>;
2188 with
2190 bb0:
2191 x' = cond;
2192 goto bb2;
2193 bb2:
2194 x = PHI <x' (bb0), ...>;
2196 We remove bb1 as it becomes unreachable. This occurs often due to
2197 gimplification of conditionals.
2199 Value Replacement
2200 -----------------
2202 This transformation, implemented in value_replacement, replaces
2204 bb0:
2205 if (a != b) goto bb2; else goto bb1;
2206 bb1:
2207 bb2:
2208 x = PHI <a (bb1), b (bb0), ...>;
2210 with
2212 bb0:
2213 bb2:
2214 x = PHI <b (bb0), ...>;
2216 This opportunity can sometimes occur as a result of other
2217 optimizations.
2220 Another case caught by value replacement looks like this:
2222 bb0:
2223 t1 = a == CONST;
2224 t2 = b > c;
2225 t3 = t1 & t2;
2226 if (t3 != 0) goto bb1; else goto bb2;
2227 bb1:
2228 bb2:
2229 x = PHI (CONST, a)
2231 Gets replaced with:
2232 bb0:
2233 bb2:
2234 t1 = a == CONST;
2235 t2 = b > c;
2236 t3 = t1 & t2;
2237 x = a;
2239 ABS Replacement
2240 ---------------
2242 This transformation, implemented in abs_replacement, replaces
2244 bb0:
2245 if (a >= 0) goto bb2; else goto bb1;
2246 bb1:
2247 x = -a;
2248 bb2:
2249 x = PHI <x (bb1), a (bb0), ...>;
2251 with
2253 bb0:
2254 x' = ABS_EXPR< a >;
2255 bb2:
2256 x = PHI <x' (bb0), ...>;
2258 MIN/MAX Replacement
2259 -------------------
2261 This transformation, minmax_replacement replaces
2263 bb0:
2264 if (a <= b) goto bb2; else goto bb1;
2265 bb1:
2266 bb2:
2267 x = PHI <b (bb1), a (bb0), ...>;
2269 with
2271 bb0:
2272 x' = MIN_EXPR (a, b)
2273 bb2:
2274 x = PHI <x' (bb0), ...>;
2276 A similar transformation is done for MAX_EXPR.
2279 This pass also performs a fifth transformation of a slightly different
2280 flavor.
2282 Factor conversion in COND_EXPR
2283 ------------------------------
2285 This transformation factors the conversion out of COND_EXPR with
2286 factor_out_conditional_conversion.
2288 For example:
2289 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2290 <bb 3>:
2291 tmp = (int) a;
2292 <bb 4>:
2293 tmp = PHI <tmp, CST>
2295 Into:
2296 if (a <= CST) goto <bb 3>; else goto <bb 4>;
2297 <bb 3>:
2298 <bb 4>:
2299 a = PHI <a, CST>
2300 tmp = (int) a;
2302 Adjacent Load Hoisting
2303 ----------------------
2305 This transformation replaces
2307 bb0:
2308 if (...) goto bb2; else goto bb1;
2309 bb1:
2310 x1 = (<expr>).field1;
2311 goto bb3;
2312 bb2:
2313 x2 = (<expr>).field2;
2314 bb3:
2315 # x = PHI <x1, x2>;
2317 with
2319 bb0:
2320 x1 = (<expr>).field1;
2321 x2 = (<expr>).field2;
2322 if (...) goto bb2; else goto bb1;
2323 bb1:
2324 goto bb3;
2325 bb2:
2326 bb3:
2327 # x = PHI <x1, x2>;
2329 The purpose of this transformation is to enable generation of conditional
2330 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2331 the loads is speculative, the transformation is restricted to very
2332 specific cases to avoid introducing a page fault. We are looking for
2333 the common idiom:
2335 if (...)
2336 x = y->left;
2337 else
2338 x = y->right;
2340 where left and right are typically adjacent pointers in a tree structure. */
2342 namespace {
2344 const pass_data pass_data_phiopt =
2346 GIMPLE_PASS, /* type */
2347 "phiopt", /* name */
2348 OPTGROUP_NONE, /* optinfo_flags */
2349 TV_TREE_PHIOPT, /* tv_id */
2350 ( PROP_cfg | PROP_ssa ), /* properties_required */
2351 0, /* properties_provided */
2352 0, /* properties_destroyed */
2353 0, /* todo_flags_start */
2354 0, /* todo_flags_finish */
2357 class pass_phiopt : public gimple_opt_pass
2359 public:
2360 pass_phiopt (gcc::context *ctxt)
2361 : gimple_opt_pass (pass_data_phiopt, ctxt)
2364 /* opt_pass methods: */
2365 opt_pass * clone () { return new pass_phiopt (m_ctxt); }
2366 virtual bool gate (function *) { return flag_ssa_phiopt; }
2367 virtual unsigned int execute (function *)
2369 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2372 }; // class pass_phiopt
2374 } // anon namespace
2376 gimple_opt_pass *
2377 make_pass_phiopt (gcc::context *ctxt)
2379 return new pass_phiopt (ctxt);
2382 namespace {
2384 const pass_data pass_data_cselim =
2386 GIMPLE_PASS, /* type */
2387 "cselim", /* name */
2388 OPTGROUP_NONE, /* optinfo_flags */
2389 TV_TREE_PHIOPT, /* tv_id */
2390 ( PROP_cfg | PROP_ssa ), /* properties_required */
2391 0, /* properties_provided */
2392 0, /* properties_destroyed */
2393 0, /* todo_flags_start */
2394 0, /* todo_flags_finish */
2397 class pass_cselim : public gimple_opt_pass
2399 public:
2400 pass_cselim (gcc::context *ctxt)
2401 : gimple_opt_pass (pass_data_cselim, ctxt)
2404 /* opt_pass methods: */
2405 virtual bool gate (function *) { return flag_tree_cselim; }
2406 virtual unsigned int execute (function *) { return tree_ssa_cs_elim (); }
2408 }; // class pass_cselim
2410 } // anon namespace
2412 gimple_opt_pass *
2413 make_pass_cselim (gcc::context *ctxt)
2415 return new pass_cselim (ctxt);