* include/parallel/numeric.h: Do not use default arguments in function
[official-gcc.git] / gcc / tree-ssa-phiopt.c
blobc7c839d29ac5d411deec8a660331e4c95ea8898e
1 /* Optimization of PHI nodes by converting them into straightline code.
2 Copyright (C) 2004-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
9 later version.
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "hash-table.h"
24 #include "tm.h"
25 #include "tree.h"
26 #include "stor-layout.h"
27 #include "flags.h"
28 #include "tm_p.h"
29 #include "predict.h"
30 #include "vec.h"
31 #include "hashtab.h"
32 #include "hash-set.h"
33 #include "machmode.h"
34 #include "hard-reg-set.h"
35 #include "input.h"
36 #include "function.h"
37 #include "dominance.h"
38 #include "cfg.h"
39 #include "cfganal.h"
40 #include "basic-block.h"
41 #include "tree-ssa-alias.h"
42 #include "internal-fn.h"
43 #include "gimple-expr.h"
44 #include "is-a.h"
45 #include "gimple.h"
46 #include "gimplify.h"
47 #include "gimple-iterator.h"
48 #include "gimplify-me.h"
49 #include "gimple-ssa.h"
50 #include "tree-cfg.h"
51 #include "tree-phinodes.h"
52 #include "ssa-iterators.h"
53 #include "stringpool.h"
54 #include "tree-ssanames.h"
55 #include "expr.h"
56 #include "tree-dfa.h"
57 #include "tree-pass.h"
58 #include "langhooks.h"
59 #include "domwalk.h"
60 #include "cfgloop.h"
61 #include "tree-data-ref.h"
62 #include "gimple-pretty-print.h"
63 #include "insn-config.h"
64 #include "expr.h"
65 #include "insn-codes.h"
66 #include "optabs.h"
67 #include "tree-scalar-evolution.h"
68 #include "tree-inline.h"
70 #ifndef HAVE_conditional_move
71 #define HAVE_conditional_move (0)
72 #endif
74 static unsigned int tree_ssa_phiopt_worker (bool, bool);
75 static bool conditional_replacement (basic_block, basic_block,
76 edge, edge, gimple, tree, tree);
77 static int value_replacement (basic_block, basic_block,
78 edge, edge, gimple, tree, tree);
79 static bool minmax_replacement (basic_block, basic_block,
80 edge, edge, gimple, tree, tree);
81 static bool abs_replacement (basic_block, basic_block,
82 edge, edge, gimple, tree, tree);
83 static bool neg_replacement (basic_block, basic_block,
84 edge, edge, gimple, tree, tree);
85 static bool cond_store_replacement (basic_block, basic_block, edge, edge,
86 hash_set<tree> *);
87 static bool cond_if_else_store_replacement (basic_block, basic_block, basic_block);
88 static hash_set<tree> * get_non_trapping ();
89 static void replace_phi_edge_with_variable (basic_block, edge, gimple, tree);
90 static void hoist_adjacent_loads (basic_block, basic_block,
91 basic_block, basic_block);
92 static bool gate_hoist_loads (void);
94 /* This pass tries to transform conditional stores into unconditional
95 ones, enabling further simplifications with the simpler then and else
96 blocks. In particular it replaces this:
98 bb0:
99 if (cond) goto bb2; else goto bb1;
100 bb1:
101 *p = RHS;
102 bb2:
104 with
106 bb0:
107 if (cond) goto bb1; else goto bb2;
108 bb1:
109 condtmp' = *p;
110 bb2:
111 condtmp = PHI <RHS, condtmp'>
112 *p = condtmp;
114 This transformation can only be done under several constraints,
115 documented below. It also replaces:
117 bb0:
118 if (cond) goto bb2; else goto bb1;
119 bb1:
120 *p = RHS1;
121 goto bb3;
122 bb2:
123 *p = RHS2;
124 bb3:
126 with
128 bb0:
129 if (cond) goto bb3; else goto bb1;
130 bb1:
131 bb3:
132 condtmp = PHI <RHS1, RHS2>
133 *p = condtmp; */
135 static unsigned int
136 tree_ssa_cs_elim (void)
138 unsigned todo;
139 /* ??? We are not interested in loop related info, but the following
140 will create it, ICEing as we didn't init loops with pre-headers.
141 An interfacing issue of find_data_references_in_bb. */
142 loop_optimizer_init (LOOPS_NORMAL);
143 scev_initialize ();
144 todo = tree_ssa_phiopt_worker (true, false);
145 scev_finalize ();
146 loop_optimizer_finalize ();
147 return todo;
150 /* Return the singleton PHI in the SEQ of PHIs for edges E0 and E1. */
152 static gimple
153 single_non_singleton_phi_for_edges (gimple_seq seq, edge e0, edge e1)
155 gimple_stmt_iterator i;
156 gimple phi = NULL;
157 if (gimple_seq_singleton_p (seq))
158 return gsi_stmt (gsi_start (seq));
159 for (i = gsi_start (seq); !gsi_end_p (i); gsi_next (&i))
161 gimple p = gsi_stmt (i);
162 /* If the PHI arguments are equal then we can skip this PHI. */
163 if (operand_equal_for_phi_arg_p (gimple_phi_arg_def (p, e0->dest_idx),
164 gimple_phi_arg_def (p, e1->dest_idx)))
165 continue;
167 /* If we already have a PHI that has the two edge arguments are
168 different, then return it is not a singleton for these PHIs. */
169 if (phi)
170 return NULL;
172 phi = p;
174 return phi;
177 /* The core routine of conditional store replacement and normal
178 phi optimizations. Both share much of the infrastructure in how
179 to match applicable basic block patterns. DO_STORE_ELIM is true
180 when we want to do conditional store replacement, false otherwise.
181 DO_HOIST_LOADS is true when we want to hoist adjacent loads out
182 of diamond control flow patterns, false otherwise. */
183 static unsigned int
184 tree_ssa_phiopt_worker (bool do_store_elim, bool do_hoist_loads)
186 basic_block bb;
187 basic_block *bb_order;
188 unsigned n, i;
189 bool cfgchanged = false;
190 hash_set<tree> *nontrap = 0;
192 if (do_store_elim)
193 /* Calculate the set of non-trapping memory accesses. */
194 nontrap = get_non_trapping ();
196 /* The replacement of conditional negation with a non-branching
197 sequence is really only a win when optimizing for speed and we
198 can avoid transformations by gimple if-conversion that result
199 in poor RTL generation.
201 Ideally either gimple if-conversion or the RTL expanders will
202 be improved and the code to emit branchless conditional negation
203 can be removed. */
204 bool replace_conditional_negation = false;
205 if (!do_store_elim)
206 replace_conditional_negation
207 = ((!optimize_size && optimize >= 2)
208 || (((flag_tree_loop_vectorize || cfun->has_force_vectorize_loops)
209 && flag_tree_loop_if_convert != 0)
210 || flag_tree_loop_if_convert == 1
211 || flag_tree_loop_if_convert_stores == 1));
213 /* Search every basic block for COND_EXPR we may be able to optimize.
215 We walk the blocks in order that guarantees that a block with
216 a single predecessor is processed before the predecessor.
217 This ensures that we collapse inner ifs before visiting the
218 outer ones, and also that we do not try to visit a removed
219 block. */
220 bb_order = single_pred_before_succ_order ();
221 n = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS;
223 for (i = 0; i < n; i++)
225 gimple cond_stmt, phi;
226 basic_block bb1, bb2;
227 edge e1, e2;
228 tree arg0, arg1;
230 bb = bb_order[i];
232 cond_stmt = last_stmt (bb);
233 /* Check to see if the last statement is a GIMPLE_COND. */
234 if (!cond_stmt
235 || gimple_code (cond_stmt) != GIMPLE_COND)
236 continue;
238 e1 = EDGE_SUCC (bb, 0);
239 bb1 = e1->dest;
240 e2 = EDGE_SUCC (bb, 1);
241 bb2 = e2->dest;
243 /* We cannot do the optimization on abnormal edges. */
244 if ((e1->flags & EDGE_ABNORMAL) != 0
245 || (e2->flags & EDGE_ABNORMAL) != 0)
246 continue;
248 /* If either bb1's succ or bb2 or bb2's succ is non NULL. */
249 if (EDGE_COUNT (bb1->succs) == 0
250 || bb2 == NULL
251 || EDGE_COUNT (bb2->succs) == 0)
252 continue;
254 /* Find the bb which is the fall through to the other. */
255 if (EDGE_SUCC (bb1, 0)->dest == bb2)
257 else if (EDGE_SUCC (bb2, 0)->dest == bb1)
259 basic_block bb_tmp = bb1;
260 edge e_tmp = e1;
261 bb1 = bb2;
262 bb2 = bb_tmp;
263 e1 = e2;
264 e2 = e_tmp;
266 else if (do_store_elim
267 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
269 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
271 if (!single_succ_p (bb1)
272 || (EDGE_SUCC (bb1, 0)->flags & EDGE_FALLTHRU) == 0
273 || !single_succ_p (bb2)
274 || (EDGE_SUCC (bb2, 0)->flags & EDGE_FALLTHRU) == 0
275 || EDGE_COUNT (bb3->preds) != 2)
276 continue;
277 if (cond_if_else_store_replacement (bb1, bb2, bb3))
278 cfgchanged = true;
279 continue;
281 else if (do_hoist_loads
282 && EDGE_SUCC (bb1, 0)->dest == EDGE_SUCC (bb2, 0)->dest)
284 basic_block bb3 = EDGE_SUCC (bb1, 0)->dest;
286 if (!FLOAT_TYPE_P (TREE_TYPE (gimple_cond_lhs (cond_stmt)))
287 && single_succ_p (bb1)
288 && single_succ_p (bb2)
289 && single_pred_p (bb1)
290 && single_pred_p (bb2)
291 && EDGE_COUNT (bb->succs) == 2
292 && EDGE_COUNT (bb3->preds) == 2
293 /* If one edge or the other is dominant, a conditional move
294 is likely to perform worse than the well-predicted branch. */
295 && !predictable_edge_p (EDGE_SUCC (bb, 0))
296 && !predictable_edge_p (EDGE_SUCC (bb, 1)))
297 hoist_adjacent_loads (bb, bb1, bb2, bb3);
298 continue;
300 else
301 continue;
303 e1 = EDGE_SUCC (bb1, 0);
305 /* Make sure that bb1 is just a fall through. */
306 if (!single_succ_p (bb1)
307 || (e1->flags & EDGE_FALLTHRU) == 0)
308 continue;
310 /* Also make sure that bb1 only have one predecessor and that it
311 is bb. */
312 if (!single_pred_p (bb1)
313 || single_pred (bb1) != bb)
314 continue;
316 if (do_store_elim)
318 /* bb1 is the middle block, bb2 the join block, bb the split block,
319 e1 the fallthrough edge from bb1 to bb2. We can't do the
320 optimization if the join block has more than two predecessors. */
321 if (EDGE_COUNT (bb2->preds) > 2)
322 continue;
323 if (cond_store_replacement (bb1, bb2, e1, e2, nontrap))
324 cfgchanged = true;
326 else
328 gimple_seq phis = phi_nodes (bb2);
329 gimple_stmt_iterator gsi;
330 bool candorest = true;
332 /* Value replacement can work with more than one PHI
333 so try that first. */
334 for (gsi = gsi_start (phis); !gsi_end_p (gsi); gsi_next (&gsi))
336 phi = gsi_stmt (gsi);
337 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
338 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
339 if (value_replacement (bb, bb1, e1, e2, phi, arg0, arg1) == 2)
341 candorest = false;
342 cfgchanged = true;
343 break;
347 if (!candorest)
348 continue;
350 phi = single_non_singleton_phi_for_edges (phis, e1, e2);
351 if (!phi)
352 continue;
354 arg0 = gimple_phi_arg_def (phi, e1->dest_idx);
355 arg1 = gimple_phi_arg_def (phi, e2->dest_idx);
357 /* Something is wrong if we cannot find the arguments in the PHI
358 node. */
359 gcc_assert (arg0 != NULL && arg1 != NULL);
361 /* Do the replacement of conditional if it can be done. */
362 if (conditional_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
363 cfgchanged = true;
364 else if (abs_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
365 cfgchanged = true;
366 else if (replace_conditional_negation
367 && neg_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
368 cfgchanged = true;
369 else if (minmax_replacement (bb, bb1, e1, e2, phi, arg0, arg1))
370 cfgchanged = true;
374 free (bb_order);
376 if (do_store_elim)
377 delete nontrap;
378 /* If the CFG has changed, we should cleanup the CFG. */
379 if (cfgchanged && do_store_elim)
381 /* In cond-store replacement we have added some loads on edges
382 and new VOPS (as we moved the store, and created a load). */
383 gsi_commit_edge_inserts ();
384 return TODO_cleanup_cfg | TODO_update_ssa_only_virtuals;
386 else if (cfgchanged)
387 return TODO_cleanup_cfg;
388 return 0;
391 /* Replace PHI node element whose edge is E in block BB with variable NEW.
392 Remove the edge from COND_BLOCK which does not lead to BB (COND_BLOCK
393 is known to have two edges, one of which must reach BB). */
395 static void
396 replace_phi_edge_with_variable (basic_block cond_block,
397 edge e, gimple phi, tree new_tree)
399 basic_block bb = gimple_bb (phi);
400 basic_block block_to_remove;
401 gimple_stmt_iterator gsi;
403 /* Change the PHI argument to new. */
404 SET_USE (PHI_ARG_DEF_PTR (phi, e->dest_idx), new_tree);
406 /* Remove the empty basic block. */
407 if (EDGE_SUCC (cond_block, 0)->dest == bb)
409 EDGE_SUCC (cond_block, 0)->flags |= EDGE_FALLTHRU;
410 EDGE_SUCC (cond_block, 0)->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
411 EDGE_SUCC (cond_block, 0)->probability = REG_BR_PROB_BASE;
412 EDGE_SUCC (cond_block, 0)->count += EDGE_SUCC (cond_block, 1)->count;
414 block_to_remove = EDGE_SUCC (cond_block, 1)->dest;
416 else
418 EDGE_SUCC (cond_block, 1)->flags |= EDGE_FALLTHRU;
419 EDGE_SUCC (cond_block, 1)->flags
420 &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE);
421 EDGE_SUCC (cond_block, 1)->probability = REG_BR_PROB_BASE;
422 EDGE_SUCC (cond_block, 1)->count += EDGE_SUCC (cond_block, 0)->count;
424 block_to_remove = EDGE_SUCC (cond_block, 0)->dest;
426 delete_basic_block (block_to_remove);
428 /* Eliminate the COND_EXPR at the end of COND_BLOCK. */
429 gsi = gsi_last_bb (cond_block);
430 gsi_remove (&gsi, true);
432 if (dump_file && (dump_flags & TDF_DETAILS))
433 fprintf (dump_file,
434 "COND_EXPR in block %d and PHI in block %d converted to straightline code.\n",
435 cond_block->index,
436 bb->index);
439 /* The function conditional_replacement does the main work of doing the
440 conditional replacement. Return true if the replacement is done.
441 Otherwise return false.
442 BB is the basic block where the replacement is going to be done on. ARG0
443 is argument 0 from PHI. Likewise for ARG1. */
445 static bool
446 conditional_replacement (basic_block cond_bb, basic_block middle_bb,
447 edge e0, edge e1, gimple phi,
448 tree arg0, tree arg1)
450 tree result;
451 gimple stmt, new_stmt;
452 tree cond;
453 gimple_stmt_iterator gsi;
454 edge true_edge, false_edge;
455 tree new_var, new_var2;
456 bool neg;
458 /* FIXME: Gimplification of complex type is too hard for now. */
459 /* We aren't prepared to handle vectors either (and it is a question
460 if it would be worthwhile anyway). */
461 if (!(INTEGRAL_TYPE_P (TREE_TYPE (arg0))
462 || POINTER_TYPE_P (TREE_TYPE (arg0)))
463 || !(INTEGRAL_TYPE_P (TREE_TYPE (arg1))
464 || POINTER_TYPE_P (TREE_TYPE (arg1))))
465 return false;
467 /* The PHI arguments have the constants 0 and 1, or 0 and -1, then
468 convert it to the conditional. */
469 if ((integer_zerop (arg0) && integer_onep (arg1))
470 || (integer_zerop (arg1) && integer_onep (arg0)))
471 neg = false;
472 else if ((integer_zerop (arg0) && integer_all_onesp (arg1))
473 || (integer_zerop (arg1) && integer_all_onesp (arg0)))
474 neg = true;
475 else
476 return false;
478 if (!empty_block_p (middle_bb))
479 return false;
481 /* At this point we know we have a GIMPLE_COND with two successors.
482 One successor is BB, the other successor is an empty block which
483 falls through into BB.
485 There is a single PHI node at the join point (BB) and its arguments
486 are constants (0, 1) or (0, -1).
488 So, given the condition COND, and the two PHI arguments, we can
489 rewrite this PHI into non-branching code:
491 dest = (COND) or dest = COND'
493 We use the condition as-is if the argument associated with the
494 true edge has the value one or the argument associated with the
495 false edge as the value zero. Note that those conditions are not
496 the same since only one of the outgoing edges from the GIMPLE_COND
497 will directly reach BB and thus be associated with an argument. */
499 stmt = last_stmt (cond_bb);
500 result = PHI_RESULT (phi);
502 /* To handle special cases like floating point comparison, it is easier and
503 less error-prone to build a tree and gimplify it on the fly though it is
504 less efficient. */
505 cond = fold_build2_loc (gimple_location (stmt),
506 gimple_cond_code (stmt), boolean_type_node,
507 gimple_cond_lhs (stmt), gimple_cond_rhs (stmt));
509 /* We need to know which is the true edge and which is the false
510 edge so that we know when to invert the condition below. */
511 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
512 if ((e0 == true_edge && integer_zerop (arg0))
513 || (e0 == false_edge && !integer_zerop (arg0))
514 || (e1 == true_edge && integer_zerop (arg1))
515 || (e1 == false_edge && !integer_zerop (arg1)))
516 cond = fold_build1_loc (gimple_location (stmt),
517 TRUTH_NOT_EXPR, TREE_TYPE (cond), cond);
519 if (neg)
521 cond = fold_convert_loc (gimple_location (stmt),
522 TREE_TYPE (result), cond);
523 cond = fold_build1_loc (gimple_location (stmt),
524 NEGATE_EXPR, TREE_TYPE (cond), cond);
527 /* Insert our new statements at the end of conditional block before the
528 COND_STMT. */
529 gsi = gsi_for_stmt (stmt);
530 new_var = force_gimple_operand_gsi (&gsi, cond, true, NULL, true,
531 GSI_SAME_STMT);
533 if (!useless_type_conversion_p (TREE_TYPE (result), TREE_TYPE (new_var)))
535 source_location locus_0, locus_1;
537 new_var2 = make_ssa_name (TREE_TYPE (result), NULL);
538 new_stmt = gimple_build_assign_with_ops (CONVERT_EXPR, new_var2,
539 new_var, NULL);
540 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
541 new_var = new_var2;
543 /* Set the locus to the first argument, unless is doesn't have one. */
544 locus_0 = gimple_phi_arg_location (phi, 0);
545 locus_1 = gimple_phi_arg_location (phi, 1);
546 if (locus_0 == UNKNOWN_LOCATION)
547 locus_0 = locus_1;
548 gimple_set_location (new_stmt, locus_0);
551 replace_phi_edge_with_variable (cond_bb, e1, phi, new_var);
553 /* Note that we optimized this PHI. */
554 return true;
557 /* Update *ARG which is defined in STMT so that it contains the
558 computed value if that seems profitable. Return true if the
559 statement is made dead by that rewriting. */
561 static bool
562 jump_function_from_stmt (tree *arg, gimple stmt)
564 enum tree_code code = gimple_assign_rhs_code (stmt);
565 if (code == ADDR_EXPR)
567 /* For arg = &p->i transform it to p, if possible. */
568 tree rhs1 = gimple_assign_rhs1 (stmt);
569 HOST_WIDE_INT offset;
570 tree tem = get_addr_base_and_unit_offset (TREE_OPERAND (rhs1, 0),
571 &offset);
572 if (tem
573 && TREE_CODE (tem) == MEM_REF
574 && (mem_ref_offset (tem) + offset) == 0)
576 *arg = TREE_OPERAND (tem, 0);
577 return true;
580 /* TODO: Much like IPA-CP jump-functions we want to handle constant
581 additions symbolically here, and we'd need to update the comparison
582 code that compares the arg + cst tuples in our caller. For now the
583 code above exactly handles the VEC_BASE pattern from vec.h. */
584 return false;
587 /* RHS is a source argument in a BIT_AND_EXPR which feeds a conditional
588 of the form SSA_NAME NE 0.
590 If RHS is fed by a simple EQ_EXPR comparison of two values, see if
591 the two input values of the EQ_EXPR match arg0 and arg1.
593 If so update *code and return TRUE. Otherwise return FALSE. */
595 static bool
596 rhs_is_fed_for_value_replacement (const_tree arg0, const_tree arg1,
597 enum tree_code *code, const_tree rhs)
599 /* Obviously if RHS is not an SSA_NAME, we can't look at the defining
600 statement. */
601 if (TREE_CODE (rhs) == SSA_NAME)
603 gimple def1 = SSA_NAME_DEF_STMT (rhs);
605 /* Verify the defining statement has an EQ_EXPR on the RHS. */
606 if (is_gimple_assign (def1) && gimple_assign_rhs_code (def1) == EQ_EXPR)
608 /* Finally verify the source operands of the EQ_EXPR are equal
609 to arg0 and arg1. */
610 tree op0 = gimple_assign_rhs1 (def1);
611 tree op1 = gimple_assign_rhs2 (def1);
612 if ((operand_equal_for_phi_arg_p (arg0, op0)
613 && operand_equal_for_phi_arg_p (arg1, op1))
614 || (operand_equal_for_phi_arg_p (arg0, op1)
615 && operand_equal_for_phi_arg_p (arg1, op0)))
617 /* We will perform the optimization. */
618 *code = gimple_assign_rhs_code (def1);
619 return true;
623 return false;
626 /* Return TRUE if arg0/arg1 are equal to the rhs/lhs or lhs/rhs of COND.
628 Also return TRUE if arg0/arg1 are equal to the source arguments of a
629 an EQ comparison feeding a BIT_AND_EXPR which feeds COND.
631 Return FALSE otherwise. */
633 static bool
634 operand_equal_for_value_replacement (const_tree arg0, const_tree arg1,
635 enum tree_code *code, gimple cond)
637 gimple def;
638 tree lhs = gimple_cond_lhs (cond);
639 tree rhs = gimple_cond_rhs (cond);
641 if ((operand_equal_for_phi_arg_p (arg0, lhs)
642 && operand_equal_for_phi_arg_p (arg1, rhs))
643 || (operand_equal_for_phi_arg_p (arg1, lhs)
644 && operand_equal_for_phi_arg_p (arg0, rhs)))
645 return true;
647 /* Now handle more complex case where we have an EQ comparison
648 which feeds a BIT_AND_EXPR which feeds COND.
650 First verify that COND is of the form SSA_NAME NE 0. */
651 if (*code != NE_EXPR || !integer_zerop (rhs)
652 || TREE_CODE (lhs) != SSA_NAME)
653 return false;
655 /* Now ensure that SSA_NAME is set by a BIT_AND_EXPR. */
656 def = SSA_NAME_DEF_STMT (lhs);
657 if (!is_gimple_assign (def) || gimple_assign_rhs_code (def) != BIT_AND_EXPR)
658 return false;
660 /* Now verify arg0/arg1 correspond to the source arguments of an
661 EQ comparison feeding the BIT_AND_EXPR. */
663 tree tmp = gimple_assign_rhs1 (def);
664 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
665 return true;
667 tmp = gimple_assign_rhs2 (def);
668 if (rhs_is_fed_for_value_replacement (arg0, arg1, code, tmp))
669 return true;
671 return false;
674 /* Returns true if ARG is a neutral element for operation CODE
675 on the RIGHT side. */
677 static bool
678 neutral_element_p (tree_code code, tree arg, bool right)
680 switch (code)
682 case PLUS_EXPR:
683 case BIT_IOR_EXPR:
684 case BIT_XOR_EXPR:
685 return integer_zerop (arg);
687 case LROTATE_EXPR:
688 case RROTATE_EXPR:
689 case LSHIFT_EXPR:
690 case RSHIFT_EXPR:
691 case MINUS_EXPR:
692 case POINTER_PLUS_EXPR:
693 return right && integer_zerop (arg);
695 case MULT_EXPR:
696 return integer_onep (arg);
698 case TRUNC_DIV_EXPR:
699 case CEIL_DIV_EXPR:
700 case FLOOR_DIV_EXPR:
701 case ROUND_DIV_EXPR:
702 case EXACT_DIV_EXPR:
703 return right && integer_onep (arg);
705 case BIT_AND_EXPR:
706 return integer_all_onesp (arg);
708 default:
709 return false;
713 /* Returns true if ARG is an absorbing element for operation CODE. */
715 static bool
716 absorbing_element_p (tree_code code, tree arg)
718 switch (code)
720 case BIT_IOR_EXPR:
721 return integer_all_onesp (arg);
723 case MULT_EXPR:
724 case BIT_AND_EXPR:
725 return integer_zerop (arg);
727 default:
728 return false;
732 /* The function value_replacement does the main work of doing the value
733 replacement. Return non-zero if the replacement is done. Otherwise return
734 0. If we remove the middle basic block, return 2.
735 BB is the basic block where the replacement is going to be done on. ARG0
736 is argument 0 from the PHI. Likewise for ARG1. */
738 static int
739 value_replacement (basic_block cond_bb, basic_block middle_bb,
740 edge e0, edge e1, gimple phi,
741 tree arg0, tree arg1)
743 gimple_stmt_iterator gsi;
744 gimple cond;
745 edge true_edge, false_edge;
746 enum tree_code code;
747 bool emtpy_or_with_defined_p = true;
749 /* If the type says honor signed zeros we cannot do this
750 optimization. */
751 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
752 return 0;
754 /* If there is a statement in MIDDLE_BB that defines one of the PHI
755 arguments, then adjust arg0 or arg1. */
756 gsi = gsi_start_nondebug_after_labels_bb (middle_bb);
757 while (!gsi_end_p (gsi))
759 gimple stmt = gsi_stmt (gsi);
760 tree lhs;
761 gsi_next_nondebug (&gsi);
762 if (!is_gimple_assign (stmt))
764 emtpy_or_with_defined_p = false;
765 continue;
767 /* Now try to adjust arg0 or arg1 according to the computation
768 in the statement. */
769 lhs = gimple_assign_lhs (stmt);
770 if (!(lhs == arg0
771 && jump_function_from_stmt (&arg0, stmt))
772 || (lhs == arg1
773 && jump_function_from_stmt (&arg1, stmt)))
774 emtpy_or_with_defined_p = false;
777 cond = last_stmt (cond_bb);
778 code = gimple_cond_code (cond);
780 /* This transformation is only valid for equality comparisons. */
781 if (code != NE_EXPR && code != EQ_EXPR)
782 return 0;
784 /* We need to know which is the true edge and which is the false
785 edge so that we know if have abs or negative abs. */
786 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
788 /* At this point we know we have a COND_EXPR with two successors.
789 One successor is BB, the other successor is an empty block which
790 falls through into BB.
792 The condition for the COND_EXPR is known to be NE_EXPR or EQ_EXPR.
794 There is a single PHI node at the join point (BB) with two arguments.
796 We now need to verify that the two arguments in the PHI node match
797 the two arguments to the equality comparison. */
799 if (operand_equal_for_value_replacement (arg0, arg1, &code, cond))
801 edge e;
802 tree arg;
804 /* For NE_EXPR, we want to build an assignment result = arg where
805 arg is the PHI argument associated with the true edge. For
806 EQ_EXPR we want the PHI argument associated with the false edge. */
807 e = (code == NE_EXPR ? true_edge : false_edge);
809 /* Unfortunately, E may not reach BB (it may instead have gone to
810 OTHER_BLOCK). If that is the case, then we want the single outgoing
811 edge from OTHER_BLOCK which reaches BB and represents the desired
812 path from COND_BLOCK. */
813 if (e->dest == middle_bb)
814 e = single_succ_edge (e->dest);
816 /* Now we know the incoming edge to BB that has the argument for the
817 RHS of our new assignment statement. */
818 if (e0 == e)
819 arg = arg0;
820 else
821 arg = arg1;
823 /* If the middle basic block was empty or is defining the
824 PHI arguments and this is a single phi where the args are different
825 for the edges e0 and e1 then we can remove the middle basic block. */
826 if (emtpy_or_with_defined_p
827 && single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)),
828 e0, e1) == phi)
830 replace_phi_edge_with_variable (cond_bb, e1, phi, arg);
831 /* Note that we optimized this PHI. */
832 return 2;
834 else
836 /* Replace the PHI arguments with arg. */
837 SET_PHI_ARG_DEF (phi, e0->dest_idx, arg);
838 SET_PHI_ARG_DEF (phi, e1->dest_idx, arg);
839 if (dump_file && (dump_flags & TDF_DETAILS))
841 fprintf (dump_file, "PHI ");
842 print_generic_expr (dump_file, gimple_phi_result (phi), 0);
843 fprintf (dump_file, " reduced for COND_EXPR in block %d to ",
844 cond_bb->index);
845 print_generic_expr (dump_file, arg, 0);
846 fprintf (dump_file, ".\n");
848 return 1;
853 /* Now optimize (x != 0) ? x + y : y to just y.
854 The following condition is too restrictive, there can easily be another
855 stmt in middle_bb, for instance a CONVERT_EXPR for the second argument. */
856 gimple assign = last_and_only_stmt (middle_bb);
857 if (!assign || gimple_code (assign) != GIMPLE_ASSIGN
858 || gimple_assign_rhs_class (assign) != GIMPLE_BINARY_RHS
859 || (!INTEGRAL_TYPE_P (TREE_TYPE (arg0))
860 && !POINTER_TYPE_P (TREE_TYPE (arg0))))
861 return 0;
863 /* Punt if there are (degenerate) PHIs in middle_bb, there should not be. */
864 if (!gimple_seq_empty_p (phi_nodes (middle_bb)))
865 return 0;
867 /* Only transform if it removes the condition. */
868 if (!single_non_singleton_phi_for_edges (phi_nodes (gimple_bb (phi)), e0, e1))
869 return 0;
871 /* Size-wise, this is always profitable. */
872 if (optimize_bb_for_speed_p (cond_bb)
873 /* The special case is useless if it has a low probability. */
874 && profile_status_for_fn (cfun) != PROFILE_ABSENT
875 && EDGE_PRED (middle_bb, 0)->probability < PROB_EVEN
876 /* If assign is cheap, there is no point avoiding it. */
877 && estimate_num_insns (assign, &eni_time_weights)
878 >= 3 * estimate_num_insns (cond, &eni_time_weights))
879 return 0;
881 tree lhs = gimple_assign_lhs (assign);
882 tree rhs1 = gimple_assign_rhs1 (assign);
883 tree rhs2 = gimple_assign_rhs2 (assign);
884 enum tree_code code_def = gimple_assign_rhs_code (assign);
885 tree cond_lhs = gimple_cond_lhs (cond);
886 tree cond_rhs = gimple_cond_rhs (cond);
888 if (((code == NE_EXPR && e1 == false_edge)
889 || (code == EQ_EXPR && e1 == true_edge))
890 && arg0 == lhs
891 && ((arg1 == rhs1
892 && operand_equal_for_phi_arg_p (rhs2, cond_lhs)
893 && neutral_element_p (code_def, cond_rhs, true))
894 || (arg1 == rhs2
895 && operand_equal_for_phi_arg_p (rhs1, cond_lhs)
896 && neutral_element_p (code_def, cond_rhs, false))
897 || (operand_equal_for_phi_arg_p (arg1, cond_rhs)
898 && (operand_equal_for_phi_arg_p (rhs2, cond_lhs)
899 || operand_equal_for_phi_arg_p (rhs1, cond_lhs))
900 && absorbing_element_p (code_def, cond_rhs))))
902 gsi = gsi_for_stmt (cond);
903 gimple_stmt_iterator gsi_from = gsi_for_stmt (assign);
904 gsi_move_before (&gsi_from, &gsi);
905 replace_phi_edge_with_variable (cond_bb, e1, phi, lhs);
906 return 2;
909 return 0;
912 /* The function minmax_replacement does the main work of doing the minmax
913 replacement. Return true if the replacement is done. Otherwise return
914 false.
915 BB is the basic block where the replacement is going to be done on. ARG0
916 is argument 0 from the PHI. Likewise for ARG1. */
918 static bool
919 minmax_replacement (basic_block cond_bb, basic_block middle_bb,
920 edge e0, edge e1, gimple phi,
921 tree arg0, tree arg1)
923 tree result, type;
924 gimple cond, new_stmt;
925 edge true_edge, false_edge;
926 enum tree_code cmp, minmax, ass_code;
927 tree smaller, larger, arg_true, arg_false;
928 gimple_stmt_iterator gsi, gsi_from;
930 type = TREE_TYPE (PHI_RESULT (phi));
932 /* The optimization may be unsafe due to NaNs. */
933 if (HONOR_NANS (TYPE_MODE (type)))
934 return false;
936 cond = last_stmt (cond_bb);
937 cmp = gimple_cond_code (cond);
939 /* This transformation is only valid for order comparisons. Record which
940 operand is smaller/larger if the result of the comparison is true. */
941 if (cmp == LT_EXPR || cmp == LE_EXPR)
943 smaller = gimple_cond_lhs (cond);
944 larger = gimple_cond_rhs (cond);
946 else if (cmp == GT_EXPR || cmp == GE_EXPR)
948 smaller = gimple_cond_rhs (cond);
949 larger = gimple_cond_lhs (cond);
951 else
952 return false;
954 /* We need to know which is the true edge and which is the false
955 edge so that we know if have abs or negative abs. */
956 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
958 /* Forward the edges over the middle basic block. */
959 if (true_edge->dest == middle_bb)
960 true_edge = EDGE_SUCC (true_edge->dest, 0);
961 if (false_edge->dest == middle_bb)
962 false_edge = EDGE_SUCC (false_edge->dest, 0);
964 if (true_edge == e0)
966 gcc_assert (false_edge == e1);
967 arg_true = arg0;
968 arg_false = arg1;
970 else
972 gcc_assert (false_edge == e0);
973 gcc_assert (true_edge == e1);
974 arg_true = arg1;
975 arg_false = arg0;
978 if (empty_block_p (middle_bb))
980 if (operand_equal_for_phi_arg_p (arg_true, smaller)
981 && operand_equal_for_phi_arg_p (arg_false, larger))
983 /* Case
985 if (smaller < larger)
986 rslt = smaller;
987 else
988 rslt = larger; */
989 minmax = MIN_EXPR;
991 else if (operand_equal_for_phi_arg_p (arg_false, smaller)
992 && operand_equal_for_phi_arg_p (arg_true, larger))
993 minmax = MAX_EXPR;
994 else
995 return false;
997 else
999 /* Recognize the following case, assuming d <= u:
1001 if (a <= u)
1002 b = MAX (a, d);
1003 x = PHI <b, u>
1005 This is equivalent to
1007 b = MAX (a, d);
1008 x = MIN (b, u); */
1010 gimple assign = last_and_only_stmt (middle_bb);
1011 tree lhs, op0, op1, bound;
1013 if (!assign
1014 || gimple_code (assign) != GIMPLE_ASSIGN)
1015 return false;
1017 lhs = gimple_assign_lhs (assign);
1018 ass_code = gimple_assign_rhs_code (assign);
1019 if (ass_code != MAX_EXPR && ass_code != MIN_EXPR)
1020 return false;
1021 op0 = gimple_assign_rhs1 (assign);
1022 op1 = gimple_assign_rhs2 (assign);
1024 if (true_edge->src == middle_bb)
1026 /* We got here if the condition is true, i.e., SMALLER < LARGER. */
1027 if (!operand_equal_for_phi_arg_p (lhs, arg_true))
1028 return false;
1030 if (operand_equal_for_phi_arg_p (arg_false, larger))
1032 /* Case
1034 if (smaller < larger)
1036 r' = MAX_EXPR (smaller, bound)
1038 r = PHI <r', larger> --> to be turned to MIN_EXPR. */
1039 if (ass_code != MAX_EXPR)
1040 return false;
1042 minmax = MIN_EXPR;
1043 if (operand_equal_for_phi_arg_p (op0, smaller))
1044 bound = op1;
1045 else if (operand_equal_for_phi_arg_p (op1, smaller))
1046 bound = op0;
1047 else
1048 return false;
1050 /* We need BOUND <= LARGER. */
1051 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1052 bound, larger)))
1053 return false;
1055 else if (operand_equal_for_phi_arg_p (arg_false, smaller))
1057 /* Case
1059 if (smaller < larger)
1061 r' = MIN_EXPR (larger, bound)
1063 r = PHI <r', smaller> --> to be turned to MAX_EXPR. */
1064 if (ass_code != MIN_EXPR)
1065 return false;
1067 minmax = MAX_EXPR;
1068 if (operand_equal_for_phi_arg_p (op0, larger))
1069 bound = op1;
1070 else if (operand_equal_for_phi_arg_p (op1, larger))
1071 bound = op0;
1072 else
1073 return false;
1075 /* We need BOUND >= SMALLER. */
1076 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1077 bound, smaller)))
1078 return false;
1080 else
1081 return false;
1083 else
1085 /* We got here if the condition is false, i.e., SMALLER > LARGER. */
1086 if (!operand_equal_for_phi_arg_p (lhs, arg_false))
1087 return false;
1089 if (operand_equal_for_phi_arg_p (arg_true, larger))
1091 /* Case
1093 if (smaller > larger)
1095 r' = MIN_EXPR (smaller, bound)
1097 r = PHI <r', larger> --> to be turned to MAX_EXPR. */
1098 if (ass_code != MIN_EXPR)
1099 return false;
1101 minmax = MAX_EXPR;
1102 if (operand_equal_for_phi_arg_p (op0, smaller))
1103 bound = op1;
1104 else if (operand_equal_for_phi_arg_p (op1, smaller))
1105 bound = op0;
1106 else
1107 return false;
1109 /* We need BOUND >= LARGER. */
1110 if (!integer_nonzerop (fold_build2 (GE_EXPR, boolean_type_node,
1111 bound, larger)))
1112 return false;
1114 else if (operand_equal_for_phi_arg_p (arg_true, smaller))
1116 /* Case
1118 if (smaller > larger)
1120 r' = MAX_EXPR (larger, bound)
1122 r = PHI <r', smaller> --> to be turned to MIN_EXPR. */
1123 if (ass_code != MAX_EXPR)
1124 return false;
1126 minmax = MIN_EXPR;
1127 if (operand_equal_for_phi_arg_p (op0, larger))
1128 bound = op1;
1129 else if (operand_equal_for_phi_arg_p (op1, larger))
1130 bound = op0;
1131 else
1132 return false;
1134 /* We need BOUND <= SMALLER. */
1135 if (!integer_nonzerop (fold_build2 (LE_EXPR, boolean_type_node,
1136 bound, smaller)))
1137 return false;
1139 else
1140 return false;
1143 /* Move the statement from the middle block. */
1144 gsi = gsi_last_bb (cond_bb);
1145 gsi_from = gsi_last_nondebug_bb (middle_bb);
1146 gsi_move_before (&gsi_from, &gsi);
1149 /* Emit the statement to compute min/max. */
1150 result = duplicate_ssa_name (PHI_RESULT (phi), NULL);
1151 new_stmt = gimple_build_assign_with_ops (minmax, result, arg0, arg1);
1152 gsi = gsi_last_bb (cond_bb);
1153 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1155 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1156 return true;
1159 /* The function absolute_replacement does the main work of doing the absolute
1160 replacement. Return true if the replacement is done. Otherwise return
1161 false.
1162 bb is the basic block where the replacement is going to be done on. arg0
1163 is argument 0 from the phi. Likewise for arg1. */
1165 static bool
1166 abs_replacement (basic_block cond_bb, basic_block middle_bb,
1167 edge e0 ATTRIBUTE_UNUSED, edge e1,
1168 gimple phi, tree arg0, tree arg1)
1170 tree result;
1171 gimple new_stmt, cond;
1172 gimple_stmt_iterator gsi;
1173 edge true_edge, false_edge;
1174 gimple assign;
1175 edge e;
1176 tree rhs, lhs;
1177 bool negate;
1178 enum tree_code cond_code;
1180 /* If the type says honor signed zeros we cannot do this
1181 optimization. */
1182 if (HONOR_SIGNED_ZEROS (TYPE_MODE (TREE_TYPE (arg1))))
1183 return false;
1185 /* OTHER_BLOCK must have only one executable statement which must have the
1186 form arg0 = -arg1 or arg1 = -arg0. */
1188 assign = last_and_only_stmt (middle_bb);
1189 /* If we did not find the proper negation assignment, then we can not
1190 optimize. */
1191 if (assign == NULL)
1192 return false;
1194 /* If we got here, then we have found the only executable statement
1195 in OTHER_BLOCK. If it is anything other than arg = -arg1 or
1196 arg1 = -arg0, then we can not optimize. */
1197 if (gimple_code (assign) != GIMPLE_ASSIGN)
1198 return false;
1200 lhs = gimple_assign_lhs (assign);
1202 if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
1203 return false;
1205 rhs = gimple_assign_rhs1 (assign);
1207 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1208 if (!(lhs == arg0 && rhs == arg1)
1209 && !(lhs == arg1 && rhs == arg0))
1210 return false;
1212 cond = last_stmt (cond_bb);
1213 result = PHI_RESULT (phi);
1215 /* Only relationals comparing arg[01] against zero are interesting. */
1216 cond_code = gimple_cond_code (cond);
1217 if (cond_code != GT_EXPR && cond_code != GE_EXPR
1218 && cond_code != LT_EXPR && cond_code != LE_EXPR)
1219 return false;
1221 /* Make sure the conditional is arg[01] OP y. */
1222 if (gimple_cond_lhs (cond) != rhs)
1223 return false;
1225 if (FLOAT_TYPE_P (TREE_TYPE (gimple_cond_rhs (cond)))
1226 ? real_zerop (gimple_cond_rhs (cond))
1227 : integer_zerop (gimple_cond_rhs (cond)))
1229 else
1230 return false;
1232 /* We need to know which is the true edge and which is the false
1233 edge so that we know if have abs or negative abs. */
1234 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1236 /* For GT_EXPR/GE_EXPR, if the true edge goes to OTHER_BLOCK, then we
1237 will need to negate the result. Similarly for LT_EXPR/LE_EXPR if
1238 the false edge goes to OTHER_BLOCK. */
1239 if (cond_code == GT_EXPR || cond_code == GE_EXPR)
1240 e = true_edge;
1241 else
1242 e = false_edge;
1244 if (e->dest == middle_bb)
1245 negate = true;
1246 else
1247 negate = false;
1249 result = duplicate_ssa_name (result, NULL);
1251 if (negate)
1252 lhs = make_ssa_name (TREE_TYPE (result), NULL);
1253 else
1254 lhs = result;
1256 /* Build the modify expression with abs expression. */
1257 new_stmt = gimple_build_assign_with_ops (ABS_EXPR, lhs, rhs, NULL);
1259 gsi = gsi_last_bb (cond_bb);
1260 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1262 if (negate)
1264 /* Get the right GSI. We want to insert after the recently
1265 added ABS_EXPR statement (which we know is the first statement
1266 in the block. */
1267 new_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, result, lhs, NULL);
1269 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1272 replace_phi_edge_with_variable (cond_bb, e1, phi, result);
1274 /* Note that we optimized this PHI. */
1275 return true;
1278 /* The function neg_replacement replaces conditional negation with
1279 equivalent straight line code. Returns TRUE if replacement is done,
1280 otherwise returns FALSE.
1282 COND_BB branches around negation occuring in MIDDLE_BB.
1284 E0 and E1 are edges out of COND_BB. E0 reaches MIDDLE_BB and
1285 E1 reaches the other successor which should contain PHI with
1286 arguments ARG0 and ARG1.
1288 Assuming negation is to occur when the condition is true,
1289 then the non-branching sequence is:
1291 result = (rhs ^ -cond) + cond
1293 Inverting the condition or its result gives us negation
1294 when the original condition is false. */
1296 static bool
1297 neg_replacement (basic_block cond_bb, basic_block middle_bb,
1298 edge e0 ATTRIBUTE_UNUSED, edge e1,
1299 gimple phi, tree arg0, tree arg1)
1301 gimple new_stmt, cond;
1302 gimple_stmt_iterator gsi;
1303 gimple assign;
1304 edge true_edge, false_edge;
1305 tree rhs, lhs;
1306 enum tree_code cond_code;
1307 bool invert = false;
1309 /* This transformation performs logical operations on the
1310 incoming arguments. So force them to be integral types. */
1311 if (!INTEGRAL_TYPE_P (TREE_TYPE (arg0)))
1312 return false;
1314 /* OTHER_BLOCK must have only one executable statement which must have the
1315 form arg0 = -arg1 or arg1 = -arg0. */
1317 assign = last_and_only_stmt (middle_bb);
1318 /* If we did not find the proper negation assignment, then we can not
1319 optimize. */
1320 if (assign == NULL)
1321 return false;
1323 /* If we got here, then we have found the only executable statement
1324 in OTHER_BLOCK. If it is anything other than arg0 = -arg1 or
1325 arg1 = -arg0, then we can not optimize. */
1326 if (gimple_code (assign) != GIMPLE_ASSIGN)
1327 return false;
1329 lhs = gimple_assign_lhs (assign);
1331 if (gimple_assign_rhs_code (assign) != NEGATE_EXPR)
1332 return false;
1334 rhs = gimple_assign_rhs1 (assign);
1336 /* The assignment has to be arg0 = -arg1 or arg1 = -arg0. */
1337 if (!(lhs == arg0 && rhs == arg1)
1338 && !(lhs == arg1 && rhs == arg0))
1339 return false;
1341 /* The basic sequence assumes we negate when the condition is true.
1342 If we need the opposite, then we will either need to invert the
1343 condition or its result. */
1344 extract_true_false_edges_from_block (cond_bb, &true_edge, &false_edge);
1345 invert = false_edge->dest == middle_bb;
1347 /* Unlike abs_replacement, we can handle arbitrary conditionals here. */
1348 cond = last_stmt (cond_bb);
1349 cond_code = gimple_cond_code (cond);
1351 /* If inversion is needed, first try to invert the test since
1352 that's cheapest. */
1353 if (invert)
1355 bool honor_nans
1356 = HONOR_NANS (TYPE_MODE (TREE_TYPE (gimple_cond_lhs (cond))));
1357 enum tree_code new_code = invert_tree_comparison (cond_code, honor_nans);
1359 /* If invert_tree_comparison was successful, then use its return
1360 value as the new code and note that inversion is no longer
1361 needed. */
1362 if (new_code != ERROR_MARK)
1364 cond_code = new_code;
1365 invert = false;
1369 tree cond_val = make_ssa_name (boolean_type_node, NULL);
1370 new_stmt = gimple_build_assign_with_ops (cond_code, cond_val,
1371 gimple_cond_lhs (cond),
1372 gimple_cond_rhs (cond));
1373 gsi = gsi_last_bb (cond_bb);
1374 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1376 /* If we still need inversion, then invert the result of the
1377 condition. */
1378 if (invert)
1380 tree tmp = make_ssa_name (boolean_type_node, NULL);
1381 new_stmt = gimple_build_assign_with_ops (BIT_XOR_EXPR, tmp,
1382 cond_val, boolean_true_node);
1383 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1384 cond_val = tmp;
1387 /* Get the condition in the right type so that we can perform
1388 logical and arithmetic operations on it. */
1389 tree cond_val_converted = make_ssa_name (TREE_TYPE (rhs), NULL);
1390 new_stmt = gimple_build_assign_with_ops (NOP_EXPR, cond_val_converted,
1391 cond_val, NULL_TREE);
1392 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1394 tree neg_cond_val_converted = make_ssa_name (TREE_TYPE (rhs), NULL);
1395 new_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, neg_cond_val_converted,
1396 cond_val_converted, NULL_TREE);
1397 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1399 tree tmp = make_ssa_name (TREE_TYPE (rhs), NULL);
1400 new_stmt = gimple_build_assign_with_ops (BIT_XOR_EXPR, tmp,
1401 rhs, neg_cond_val_converted);
1402 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1404 tree new_lhs = make_ssa_name (TREE_TYPE (rhs), NULL);
1405 new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, new_lhs,
1406 tmp, cond_val_converted);
1407 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1409 replace_phi_edge_with_variable (cond_bb, e1, phi, new_lhs);
1411 /* Note that we optimized this PHI. */
1412 return true;
1415 /* Auxiliary functions to determine the set of memory accesses which
1416 can't trap because they are preceded by accesses to the same memory
1417 portion. We do that for MEM_REFs, so we only need to track
1418 the SSA_NAME of the pointer indirectly referenced. The algorithm
1419 simply is a walk over all instructions in dominator order. When
1420 we see an MEM_REF we determine if we've already seen a same
1421 ref anywhere up to the root of the dominator tree. If we do the
1422 current access can't trap. If we don't see any dominating access
1423 the current access might trap, but might also make later accesses
1424 non-trapping, so we remember it. We need to be careful with loads
1425 or stores, for instance a load might not trap, while a store would,
1426 so if we see a dominating read access this doesn't mean that a later
1427 write access would not trap. Hence we also need to differentiate the
1428 type of access(es) seen.
1430 ??? We currently are very conservative and assume that a load might
1431 trap even if a store doesn't (write-only memory). This probably is
1432 overly conservative. */
1434 /* A hash-table of SSA_NAMEs, and in which basic block an MEM_REF
1435 through it was seen, which would constitute a no-trap region for
1436 same accesses. */
1437 struct name_to_bb
1439 unsigned int ssa_name_ver;
1440 unsigned int phase;
1441 bool store;
1442 HOST_WIDE_INT offset, size;
1443 basic_block bb;
1446 /* Hashtable helpers. */
1448 struct ssa_names_hasher : typed_free_remove <name_to_bb>
1450 typedef name_to_bb value_type;
1451 typedef name_to_bb compare_type;
1452 static inline hashval_t hash (const value_type *);
1453 static inline bool equal (const value_type *, const compare_type *);
1456 /* Used for quick clearing of the hash-table when we see calls.
1457 Hash entries with phase < nt_call_phase are invalid. */
1458 static unsigned int nt_call_phase;
1460 /* The hash function. */
1462 inline hashval_t
1463 ssa_names_hasher::hash (const value_type *n)
1465 return n->ssa_name_ver ^ (((hashval_t) n->store) << 31)
1466 ^ (n->offset << 6) ^ (n->size << 3);
1469 /* The equality function of *P1 and *P2. */
1471 inline bool
1472 ssa_names_hasher::equal (const value_type *n1, const compare_type *n2)
1474 return n1->ssa_name_ver == n2->ssa_name_ver
1475 && n1->store == n2->store
1476 && n1->offset == n2->offset
1477 && n1->size == n2->size;
1480 class nontrapping_dom_walker : public dom_walker
1482 public:
1483 nontrapping_dom_walker (cdi_direction direction, hash_set<tree> *ps)
1484 : dom_walker (direction), m_nontrapping (ps), m_seen_ssa_names (128) {}
1486 virtual void before_dom_children (basic_block);
1487 virtual void after_dom_children (basic_block);
1489 private:
1491 /* We see the expression EXP in basic block BB. If it's an interesting
1492 expression (an MEM_REF through an SSA_NAME) possibly insert the
1493 expression into the set NONTRAP or the hash table of seen expressions.
1494 STORE is true if this expression is on the LHS, otherwise it's on
1495 the RHS. */
1496 void add_or_mark_expr (basic_block, tree, bool);
1498 hash_set<tree> *m_nontrapping;
1500 /* The hash table for remembering what we've seen. */
1501 hash_table<ssa_names_hasher> m_seen_ssa_names;
1504 /* Called by walk_dominator_tree, when entering the block BB. */
1505 void
1506 nontrapping_dom_walker::before_dom_children (basic_block bb)
1508 edge e;
1509 edge_iterator ei;
1510 gimple_stmt_iterator gsi;
1512 /* If we haven't seen all our predecessors, clear the hash-table. */
1513 FOR_EACH_EDGE (e, ei, bb->preds)
1514 if ((((size_t)e->src->aux) & 2) == 0)
1516 nt_call_phase++;
1517 break;
1520 /* Mark this BB as being on the path to dominator root and as visited. */
1521 bb->aux = (void*)(1 | 2);
1523 /* And walk the statements in order. */
1524 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1526 gimple stmt = gsi_stmt (gsi);
1528 if (is_gimple_call (stmt) && !nonfreeing_call_p (stmt))
1529 nt_call_phase++;
1530 else if (gimple_assign_single_p (stmt) && !gimple_has_volatile_ops (stmt))
1532 add_or_mark_expr (bb, gimple_assign_lhs (stmt), true);
1533 add_or_mark_expr (bb, gimple_assign_rhs1 (stmt), false);
1538 /* Called by walk_dominator_tree, when basic block BB is exited. */
1539 void
1540 nontrapping_dom_walker::after_dom_children (basic_block bb)
1542 /* This BB isn't on the path to dominator root anymore. */
1543 bb->aux = (void*)2;
1546 /* We see the expression EXP in basic block BB. If it's an interesting
1547 expression (an MEM_REF through an SSA_NAME) possibly insert the
1548 expression into the set NONTRAP or the hash table of seen expressions.
1549 STORE is true if this expression is on the LHS, otherwise it's on
1550 the RHS. */
1551 void
1552 nontrapping_dom_walker::add_or_mark_expr (basic_block bb, tree exp, bool store)
1554 HOST_WIDE_INT size;
1556 if (TREE_CODE (exp) == MEM_REF
1557 && TREE_CODE (TREE_OPERAND (exp, 0)) == SSA_NAME
1558 && tree_fits_shwi_p (TREE_OPERAND (exp, 1))
1559 && (size = int_size_in_bytes (TREE_TYPE (exp))) > 0)
1561 tree name = TREE_OPERAND (exp, 0);
1562 struct name_to_bb map;
1563 name_to_bb **slot;
1564 struct name_to_bb *n2bb;
1565 basic_block found_bb = 0;
1567 /* Try to find the last seen MEM_REF through the same
1568 SSA_NAME, which can trap. */
1569 map.ssa_name_ver = SSA_NAME_VERSION (name);
1570 map.phase = 0;
1571 map.bb = 0;
1572 map.store = store;
1573 map.offset = tree_to_shwi (TREE_OPERAND (exp, 1));
1574 map.size = size;
1576 slot = m_seen_ssa_names.find_slot (&map, INSERT);
1577 n2bb = *slot;
1578 if (n2bb && n2bb->phase >= nt_call_phase)
1579 found_bb = n2bb->bb;
1581 /* If we've found a trapping MEM_REF, _and_ it dominates EXP
1582 (it's in a basic block on the path from us to the dominator root)
1583 then we can't trap. */
1584 if (found_bb && (((size_t)found_bb->aux) & 1) == 1)
1586 m_nontrapping->add (exp);
1588 else
1590 /* EXP might trap, so insert it into the hash table. */
1591 if (n2bb)
1593 n2bb->phase = nt_call_phase;
1594 n2bb->bb = bb;
1596 else
1598 n2bb = XNEW (struct name_to_bb);
1599 n2bb->ssa_name_ver = SSA_NAME_VERSION (name);
1600 n2bb->phase = nt_call_phase;
1601 n2bb->bb = bb;
1602 n2bb->store = store;
1603 n2bb->offset = map.offset;
1604 n2bb->size = size;
1605 *slot = n2bb;
1611 /* This is the entry point of gathering non trapping memory accesses.
1612 It will do a dominator walk over the whole function, and it will
1613 make use of the bb->aux pointers. It returns a set of trees
1614 (the MEM_REFs itself) which can't trap. */
1615 static hash_set<tree> *
1616 get_non_trapping (void)
1618 nt_call_phase = 0;
1619 hash_set<tree> *nontrap = new hash_set<tree>;
1620 /* We're going to do a dominator walk, so ensure that we have
1621 dominance information. */
1622 calculate_dominance_info (CDI_DOMINATORS);
1624 nontrapping_dom_walker (CDI_DOMINATORS, nontrap)
1625 .walk (cfun->cfg->x_entry_block_ptr);
1627 clear_aux_for_blocks ();
1628 return nontrap;
1631 /* Do the main work of conditional store replacement. We already know
1632 that the recognized pattern looks like so:
1634 split:
1635 if (cond) goto MIDDLE_BB; else goto JOIN_BB (edge E1)
1636 MIDDLE_BB:
1637 something
1638 fallthrough (edge E0)
1639 JOIN_BB:
1640 some more
1642 We check that MIDDLE_BB contains only one store, that that store
1643 doesn't trap (not via NOTRAP, but via checking if an access to the same
1644 memory location dominates us) and that the store has a "simple" RHS. */
1646 static bool
1647 cond_store_replacement (basic_block middle_bb, basic_block join_bb,
1648 edge e0, edge e1, hash_set<tree> *nontrap)
1650 gimple assign = last_and_only_stmt (middle_bb);
1651 tree lhs, rhs, name, name2;
1652 gimple newphi, 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 gimple newphi, new_stmt;
1727 if (then_assign == NULL
1728 || !gimple_assign_single_p (then_assign)
1729 || gimple_clobber_p (then_assign)
1730 || gimple_has_volatile_ops (then_assign)
1731 || else_assign == NULL
1732 || !gimple_assign_single_p (else_assign)
1733 || gimple_clobber_p (else_assign)
1734 || gimple_has_volatile_ops (else_assign))
1735 return false;
1737 lhs = gimple_assign_lhs (then_assign);
1738 if (!is_gimple_reg_type (TREE_TYPE (lhs))
1739 || !operand_equal_p (lhs, gimple_assign_lhs (else_assign), 0))
1740 return false;
1742 lhs_base = get_base_address (lhs);
1743 if (lhs_base == NULL_TREE
1744 || (!DECL_P (lhs_base) && TREE_CODE (lhs_base) != MEM_REF))
1745 return false;
1747 then_rhs = gimple_assign_rhs1 (then_assign);
1748 else_rhs = gimple_assign_rhs1 (else_assign);
1749 then_locus = gimple_location (then_assign);
1750 else_locus = gimple_location (else_assign);
1752 /* Now we've checked the constraints, so do the transformation:
1753 1) Remove the stores. */
1754 gsi = gsi_for_stmt (then_assign);
1755 unlink_stmt_vdef (then_assign);
1756 gsi_remove (&gsi, true);
1757 release_defs (then_assign);
1759 gsi = gsi_for_stmt (else_assign);
1760 unlink_stmt_vdef (else_assign);
1761 gsi_remove (&gsi, true);
1762 release_defs (else_assign);
1764 /* 2) Create a PHI node at the join block, with one argument
1765 holding the old RHS, and the other holding the temporary
1766 where we stored the old memory contents. */
1767 name = make_temp_ssa_name (TREE_TYPE (lhs), NULL, "cstore");
1768 newphi = create_phi_node (name, join_bb);
1769 add_phi_arg (newphi, then_rhs, EDGE_SUCC (then_bb, 0), then_locus);
1770 add_phi_arg (newphi, else_rhs, EDGE_SUCC (else_bb, 0), else_locus);
1772 new_stmt = gimple_build_assign (lhs, PHI_RESULT (newphi));
1774 /* 3) Insert that PHI node. */
1775 gsi = gsi_after_labels (join_bb);
1776 if (gsi_end_p (gsi))
1778 gsi = gsi_last_bb (join_bb);
1779 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
1781 else
1782 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
1784 return true;
1787 /* Conditional store replacement. We already know
1788 that the recognized pattern looks like so:
1790 split:
1791 if (cond) goto THEN_BB; else goto ELSE_BB (edge E1)
1792 THEN_BB:
1794 X = Y;
1796 goto JOIN_BB;
1797 ELSE_BB:
1799 X = Z;
1801 fallthrough (edge E0)
1802 JOIN_BB:
1803 some more
1805 We check that it is safe to sink the store to JOIN_BB by verifying that
1806 there are no read-after-write or write-after-write dependencies in
1807 THEN_BB and ELSE_BB. */
1809 static bool
1810 cond_if_else_store_replacement (basic_block then_bb, basic_block else_bb,
1811 basic_block join_bb)
1813 gimple then_assign = last_and_only_stmt (then_bb);
1814 gimple else_assign = last_and_only_stmt (else_bb);
1815 vec<data_reference_p> then_datarefs, else_datarefs;
1816 vec<ddr_p> then_ddrs, else_ddrs;
1817 gimple then_store, else_store;
1818 bool found, ok = false, res;
1819 struct data_dependence_relation *ddr;
1820 data_reference_p then_dr, else_dr;
1821 int i, j;
1822 tree then_lhs, else_lhs;
1823 basic_block blocks[3];
1825 if (MAX_STORES_TO_SINK == 0)
1826 return false;
1828 /* Handle the case with single statement in THEN_BB and ELSE_BB. */
1829 if (then_assign && else_assign)
1830 return cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1831 then_assign, else_assign);
1833 /* Find data references. */
1834 then_datarefs.create (1);
1835 else_datarefs.create (1);
1836 if ((find_data_references_in_bb (NULL, then_bb, &then_datarefs)
1837 == chrec_dont_know)
1838 || !then_datarefs.length ()
1839 || (find_data_references_in_bb (NULL, else_bb, &else_datarefs)
1840 == chrec_dont_know)
1841 || !else_datarefs.length ())
1843 free_data_refs (then_datarefs);
1844 free_data_refs (else_datarefs);
1845 return false;
1848 /* Find pairs of stores with equal LHS. */
1849 auto_vec<gimple, 1> then_stores, else_stores;
1850 FOR_EACH_VEC_ELT (then_datarefs, i, then_dr)
1852 if (DR_IS_READ (then_dr))
1853 continue;
1855 then_store = DR_STMT (then_dr);
1856 then_lhs = gimple_get_lhs (then_store);
1857 if (then_lhs == NULL_TREE)
1858 continue;
1859 found = false;
1861 FOR_EACH_VEC_ELT (else_datarefs, j, else_dr)
1863 if (DR_IS_READ (else_dr))
1864 continue;
1866 else_store = DR_STMT (else_dr);
1867 else_lhs = gimple_get_lhs (else_store);
1868 if (else_lhs == NULL_TREE)
1869 continue;
1871 if (operand_equal_p (then_lhs, else_lhs, 0))
1873 found = true;
1874 break;
1878 if (!found)
1879 continue;
1881 then_stores.safe_push (then_store);
1882 else_stores.safe_push (else_store);
1885 /* No pairs of stores found. */
1886 if (!then_stores.length ()
1887 || then_stores.length () > (unsigned) MAX_STORES_TO_SINK)
1889 free_data_refs (then_datarefs);
1890 free_data_refs (else_datarefs);
1891 return false;
1894 /* Compute and check data dependencies in both basic blocks. */
1895 then_ddrs.create (1);
1896 else_ddrs.create (1);
1897 if (!compute_all_dependences (then_datarefs, &then_ddrs,
1898 vNULL, false)
1899 || !compute_all_dependences (else_datarefs, &else_ddrs,
1900 vNULL, false))
1902 free_dependence_relations (then_ddrs);
1903 free_dependence_relations (else_ddrs);
1904 free_data_refs (then_datarefs);
1905 free_data_refs (else_datarefs);
1906 return false;
1908 blocks[0] = then_bb;
1909 blocks[1] = else_bb;
1910 blocks[2] = join_bb;
1911 renumber_gimple_stmt_uids_in_blocks (blocks, 3);
1913 /* Check that there are no read-after-write or write-after-write dependencies
1914 in THEN_BB. */
1915 FOR_EACH_VEC_ELT (then_ddrs, i, ddr)
1917 struct data_reference *dra = DDR_A (ddr);
1918 struct data_reference *drb = DDR_B (ddr);
1920 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1921 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1922 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1923 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1924 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1925 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1927 free_dependence_relations (then_ddrs);
1928 free_dependence_relations (else_ddrs);
1929 free_data_refs (then_datarefs);
1930 free_data_refs (else_datarefs);
1931 return false;
1935 /* Check that there are no read-after-write or write-after-write dependencies
1936 in ELSE_BB. */
1937 FOR_EACH_VEC_ELT (else_ddrs, i, ddr)
1939 struct data_reference *dra = DDR_A (ddr);
1940 struct data_reference *drb = DDR_B (ddr);
1942 if (DDR_ARE_DEPENDENT (ddr) != chrec_known
1943 && ((DR_IS_READ (dra) && DR_IS_WRITE (drb)
1944 && gimple_uid (DR_STMT (dra)) > gimple_uid (DR_STMT (drb)))
1945 || (DR_IS_READ (drb) && DR_IS_WRITE (dra)
1946 && gimple_uid (DR_STMT (drb)) > gimple_uid (DR_STMT (dra)))
1947 || (DR_IS_WRITE (dra) && DR_IS_WRITE (drb))))
1949 free_dependence_relations (then_ddrs);
1950 free_dependence_relations (else_ddrs);
1951 free_data_refs (then_datarefs);
1952 free_data_refs (else_datarefs);
1953 return false;
1957 /* Sink stores with same LHS. */
1958 FOR_EACH_VEC_ELT (then_stores, i, then_store)
1960 else_store = else_stores[i];
1961 res = cond_if_else_store_replacement_1 (then_bb, else_bb, join_bb,
1962 then_store, else_store);
1963 ok = ok || res;
1966 free_dependence_relations (then_ddrs);
1967 free_dependence_relations (else_ddrs);
1968 free_data_refs (then_datarefs);
1969 free_data_refs (else_datarefs);
1971 return ok;
1974 /* Return TRUE if STMT has a VUSE whose corresponding VDEF is in BB. */
1976 static bool
1977 local_mem_dependence (gimple stmt, basic_block bb)
1979 tree vuse = gimple_vuse (stmt);
1980 gimple def;
1982 if (!vuse)
1983 return false;
1985 def = SSA_NAME_DEF_STMT (vuse);
1986 return (def && gimple_bb (def) == bb);
1989 /* Given a "diamond" control-flow pattern where BB0 tests a condition,
1990 BB1 and BB2 are "then" and "else" blocks dependent on this test,
1991 and BB3 rejoins control flow following BB1 and BB2, look for
1992 opportunities to hoist loads as follows. If BB3 contains a PHI of
1993 two loads, one each occurring in BB1 and BB2, and the loads are
1994 provably of adjacent fields in the same structure, then move both
1995 loads into BB0. Of course this can only be done if there are no
1996 dependencies preventing such motion.
1998 One of the hoisted loads will always be speculative, so the
1999 transformation is currently conservative:
2001 - The fields must be strictly adjacent.
2002 - The two fields must occupy a single memory block that is
2003 guaranteed to not cross a page boundary.
2005 The last is difficult to prove, as such memory blocks should be
2006 aligned on the minimum of the stack alignment boundary and the
2007 alignment guaranteed by heap allocation interfaces. Thus we rely
2008 on a parameter for the alignment value.
2010 Provided a good value is used for the last case, the first
2011 restriction could possibly be relaxed. */
2013 static void
2014 hoist_adjacent_loads (basic_block bb0, basic_block bb1,
2015 basic_block bb2, basic_block bb3)
2017 int param_align = PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE);
2018 unsigned param_align_bits = (unsigned) (param_align * BITS_PER_UNIT);
2019 gimple_stmt_iterator gsi;
2021 /* Walk the phis in bb3 looking for an opportunity. We are looking
2022 for phis of two SSA names, one each of which is defined in bb1 and
2023 bb2. */
2024 for (gsi = gsi_start_phis (bb3); !gsi_end_p (gsi); gsi_next (&gsi))
2026 gimple phi_stmt = gsi_stmt (gsi);
2027 gimple def1, def2, defswap;
2028 tree arg1, arg2, ref1, ref2, field1, field2, fieldswap;
2029 tree tree_offset1, tree_offset2, tree_size2, next;
2030 int offset1, offset2, size2;
2031 unsigned align1;
2032 gimple_stmt_iterator gsi2;
2033 basic_block bb_for_def1, bb_for_def2;
2035 if (gimple_phi_num_args (phi_stmt) != 2
2036 || virtual_operand_p (gimple_phi_result (phi_stmt)))
2037 continue;
2039 arg1 = gimple_phi_arg_def (phi_stmt, 0);
2040 arg2 = gimple_phi_arg_def (phi_stmt, 1);
2042 if (TREE_CODE (arg1) != SSA_NAME
2043 || TREE_CODE (arg2) != SSA_NAME
2044 || SSA_NAME_IS_DEFAULT_DEF (arg1)
2045 || SSA_NAME_IS_DEFAULT_DEF (arg2))
2046 continue;
2048 def1 = SSA_NAME_DEF_STMT (arg1);
2049 def2 = SSA_NAME_DEF_STMT (arg2);
2051 if ((gimple_bb (def1) != bb1 || gimple_bb (def2) != bb2)
2052 && (gimple_bb (def2) != bb1 || gimple_bb (def1) != bb2))
2053 continue;
2055 /* Check the mode of the arguments to be sure a conditional move
2056 can be generated for it. */
2057 if (optab_handler (movcc_optab, TYPE_MODE (TREE_TYPE (arg1)))
2058 == CODE_FOR_nothing)
2059 continue;
2061 /* Both statements must be assignments whose RHS is a COMPONENT_REF. */
2062 if (!gimple_assign_single_p (def1)
2063 || !gimple_assign_single_p (def2)
2064 || gimple_has_volatile_ops (def1)
2065 || gimple_has_volatile_ops (def2))
2066 continue;
2068 ref1 = gimple_assign_rhs1 (def1);
2069 ref2 = gimple_assign_rhs1 (def2);
2071 if (TREE_CODE (ref1) != COMPONENT_REF
2072 || TREE_CODE (ref2) != COMPONENT_REF)
2073 continue;
2075 /* The zeroth operand of the two component references must be
2076 identical. It is not sufficient to compare get_base_address of
2077 the two references, because this could allow for different
2078 elements of the same array in the two trees. It is not safe to
2079 assume that the existence of one array element implies the
2080 existence of a different one. */
2081 if (!operand_equal_p (TREE_OPERAND (ref1, 0), TREE_OPERAND (ref2, 0), 0))
2082 continue;
2084 field1 = TREE_OPERAND (ref1, 1);
2085 field2 = TREE_OPERAND (ref2, 1);
2087 /* Check for field adjacency, and ensure field1 comes first. */
2088 for (next = DECL_CHAIN (field1);
2089 next && TREE_CODE (next) != FIELD_DECL;
2090 next = DECL_CHAIN (next))
2093 if (next != field2)
2095 for (next = DECL_CHAIN (field2);
2096 next && TREE_CODE (next) != FIELD_DECL;
2097 next = DECL_CHAIN (next))
2100 if (next != field1)
2101 continue;
2103 fieldswap = field1;
2104 field1 = field2;
2105 field2 = fieldswap;
2106 defswap = def1;
2107 def1 = def2;
2108 def2 = defswap;
2111 bb_for_def1 = gimple_bb (def1);
2112 bb_for_def2 = gimple_bb (def2);
2114 /* Check for proper alignment of the first field. */
2115 tree_offset1 = bit_position (field1);
2116 tree_offset2 = bit_position (field2);
2117 tree_size2 = DECL_SIZE (field2);
2119 if (!tree_fits_uhwi_p (tree_offset1)
2120 || !tree_fits_uhwi_p (tree_offset2)
2121 || !tree_fits_uhwi_p (tree_size2))
2122 continue;
2124 offset1 = tree_to_uhwi (tree_offset1);
2125 offset2 = tree_to_uhwi (tree_offset2);
2126 size2 = tree_to_uhwi (tree_size2);
2127 align1 = DECL_ALIGN (field1) % param_align_bits;
2129 if (offset1 % BITS_PER_UNIT != 0)
2130 continue;
2132 /* For profitability, the two field references should fit within
2133 a single cache line. */
2134 if (align1 + offset2 - offset1 + size2 > param_align_bits)
2135 continue;
2137 /* The two expressions cannot be dependent upon vdefs defined
2138 in bb1/bb2. */
2139 if (local_mem_dependence (def1, bb_for_def1)
2140 || local_mem_dependence (def2, bb_for_def2))
2141 continue;
2143 /* The conditions are satisfied; hoist the loads from bb1 and bb2 into
2144 bb0. We hoist the first one first so that a cache miss is handled
2145 efficiently regardless of hardware cache-fill policy. */
2146 gsi2 = gsi_for_stmt (def1);
2147 gsi_move_to_bb_end (&gsi2, bb0);
2148 gsi2 = gsi_for_stmt (def2);
2149 gsi_move_to_bb_end (&gsi2, bb0);
2151 if (dump_file && (dump_flags & TDF_DETAILS))
2153 fprintf (dump_file,
2154 "\nHoisting adjacent loads from %d and %d into %d: \n",
2155 bb_for_def1->index, bb_for_def2->index, bb0->index);
2156 print_gimple_stmt (dump_file, def1, 0, TDF_VOPS|TDF_MEMSYMS);
2157 print_gimple_stmt (dump_file, def2, 0, TDF_VOPS|TDF_MEMSYMS);
2162 /* Determine whether we should attempt to hoist adjacent loads out of
2163 diamond patterns in pass_phiopt. Always hoist loads if
2164 -fhoist-adjacent-loads is specified and the target machine has
2165 both a conditional move instruction and a defined cache line size. */
2167 static bool
2168 gate_hoist_loads (void)
2170 return (flag_hoist_adjacent_loads == 1
2171 && PARAM_VALUE (PARAM_L1_CACHE_LINE_SIZE)
2172 && HAVE_conditional_move);
2175 /* This pass tries to replaces an if-then-else block with an
2176 assignment. We have four kinds of transformations. Some of these
2177 transformations are also performed by the ifcvt RTL optimizer.
2179 Conditional Replacement
2180 -----------------------
2182 This transformation, implemented in conditional_replacement,
2183 replaces
2185 bb0:
2186 if (cond) goto bb2; else goto bb1;
2187 bb1:
2188 bb2:
2189 x = PHI <0 (bb1), 1 (bb0), ...>;
2191 with
2193 bb0:
2194 x' = cond;
2195 goto bb2;
2196 bb2:
2197 x = PHI <x' (bb0), ...>;
2199 We remove bb1 as it becomes unreachable. This occurs often due to
2200 gimplification of conditionals.
2202 Value Replacement
2203 -----------------
2205 This transformation, implemented in value_replacement, replaces
2207 bb0:
2208 if (a != b) goto bb2; else goto bb1;
2209 bb1:
2210 bb2:
2211 x = PHI <a (bb1), b (bb0), ...>;
2213 with
2215 bb0:
2216 bb2:
2217 x = PHI <b (bb0), ...>;
2219 This opportunity can sometimes occur as a result of other
2220 optimizations.
2223 Another case caught by value replacement looks like this:
2225 bb0:
2226 t1 = a == CONST;
2227 t2 = b > c;
2228 t3 = t1 & t2;
2229 if (t3 != 0) goto bb1; else goto bb2;
2230 bb1:
2231 bb2:
2232 x = PHI (CONST, a)
2234 Gets replaced with:
2235 bb0:
2236 bb2:
2237 t1 = a == CONST;
2238 t2 = b > c;
2239 t3 = t1 & t2;
2240 x = a;
2242 ABS Replacement
2243 ---------------
2245 This transformation, implemented in abs_replacement, replaces
2247 bb0:
2248 if (a >= 0) goto bb2; else goto bb1;
2249 bb1:
2250 x = -a;
2251 bb2:
2252 x = PHI <x (bb1), a (bb0), ...>;
2254 with
2256 bb0:
2257 x' = ABS_EXPR< a >;
2258 bb2:
2259 x = PHI <x' (bb0), ...>;
2261 MIN/MAX Replacement
2262 -------------------
2264 This transformation, minmax_replacement replaces
2266 bb0:
2267 if (a <= b) goto bb2; else goto bb1;
2268 bb1:
2269 bb2:
2270 x = PHI <b (bb1), a (bb0), ...>;
2272 with
2274 bb0:
2275 x' = MIN_EXPR (a, b)
2276 bb2:
2277 x = PHI <x' (bb0), ...>;
2279 A similar transformation is done for MAX_EXPR.
2282 This pass also performs a fifth transformation of a slightly different
2283 flavor.
2285 Adjacent Load Hoisting
2286 ----------------------
2288 This transformation replaces
2290 bb0:
2291 if (...) goto bb2; else goto bb1;
2292 bb1:
2293 x1 = (<expr>).field1;
2294 goto bb3;
2295 bb2:
2296 x2 = (<expr>).field2;
2297 bb3:
2298 # x = PHI <x1, x2>;
2300 with
2302 bb0:
2303 x1 = (<expr>).field1;
2304 x2 = (<expr>).field2;
2305 if (...) goto bb2; else goto bb1;
2306 bb1:
2307 goto bb3;
2308 bb2:
2309 bb3:
2310 # x = PHI <x1, x2>;
2312 The purpose of this transformation is to enable generation of conditional
2313 move instructions such as Intel CMOVE or PowerPC ISEL. Because one of
2314 the loads is speculative, the transformation is restricted to very
2315 specific cases to avoid introducing a page fault. We are looking for
2316 the common idiom:
2318 if (...)
2319 x = y->left;
2320 else
2321 x = y->right;
2323 where left and right are typically adjacent pointers in a tree structure. */
2325 namespace {
2327 const pass_data pass_data_phiopt =
2329 GIMPLE_PASS, /* type */
2330 "phiopt", /* name */
2331 OPTGROUP_NONE, /* optinfo_flags */
2332 TV_TREE_PHIOPT, /* tv_id */
2333 ( PROP_cfg | PROP_ssa ), /* properties_required */
2334 0, /* properties_provided */
2335 0, /* properties_destroyed */
2336 0, /* todo_flags_start */
2337 0, /* todo_flags_finish */
2340 class pass_phiopt : public gimple_opt_pass
2342 public:
2343 pass_phiopt (gcc::context *ctxt)
2344 : gimple_opt_pass (pass_data_phiopt, ctxt)
2347 /* opt_pass methods: */
2348 opt_pass * clone () { return new pass_phiopt (m_ctxt); }
2349 virtual bool gate (function *) { return flag_ssa_phiopt; }
2350 virtual unsigned int execute (function *)
2352 return tree_ssa_phiopt_worker (false, gate_hoist_loads ());
2355 }; // class pass_phiopt
2357 } // anon namespace
2359 gimple_opt_pass *
2360 make_pass_phiopt (gcc::context *ctxt)
2362 return new pass_phiopt (ctxt);
2365 namespace {
2367 const pass_data pass_data_cselim =
2369 GIMPLE_PASS, /* type */
2370 "cselim", /* name */
2371 OPTGROUP_NONE, /* optinfo_flags */
2372 TV_TREE_PHIOPT, /* tv_id */
2373 ( PROP_cfg | PROP_ssa ), /* properties_required */
2374 0, /* properties_provided */
2375 0, /* properties_destroyed */
2376 0, /* todo_flags_start */
2377 0, /* todo_flags_finish */
2380 class pass_cselim : public gimple_opt_pass
2382 public:
2383 pass_cselim (gcc::context *ctxt)
2384 : gimple_opt_pass (pass_data_cselim, ctxt)
2387 /* opt_pass methods: */
2388 virtual bool gate (function *) { return flag_tree_cselim; }
2389 virtual unsigned int execute (function *) { return tree_ssa_cs_elim (); }
2391 }; // class pass_cselim
2393 } // anon namespace
2395 gimple_opt_pass *
2396 make_pass_cselim (gcc::context *ctxt)
2398 return new pass_cselim (ctxt);