don't use TYPE_ARG_TYPES in the Ada frontend
[gcc.git] / gcc / tree-if-conv.c
1 /* If-conversion for vectorizer.
2 Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010
3 Free Software Foundation, Inc.
4 Contributed by Devang Patel <dpatel@apple.com>
5
6 This file is part of GCC.
7
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
12
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
17
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
21
22 /* This pass implements a tree level if-conversion of loops. Its
23 initial goal is to help the vectorizer to vectorize loops with
24 conditions.
25
26 A short description of if-conversion:
27
28 o Decide if a loop is if-convertible or not.
29 o Walk all loop basic blocks in breadth first order (BFS order).
30 o Remove conditional statements (at the end of basic block)
31 and propagate condition into destination basic blocks'
32 predicate list.
33 o Replace modify expression with conditional modify expression
34 using current basic block's condition.
35 o Merge all basic blocks
36 o Replace phi nodes with conditional modify expr
37 o Merge all basic blocks into header
38
39 Sample transformation:
40
41 INPUT
42 -----
43
44 # i_23 = PHI <0(0), i_18(10)>;
45 <L0>:;
46 j_15 = A[i_23];
47 if (j_15 > 41) goto <L1>; else goto <L17>;
48
49 <L17>:;
50 goto <bb 3> (<L3>);
51
52 <L1>:;
53
54 # iftmp.2_4 = PHI <0(8), 42(2)>;
55 <L3>:;
56 A[i_23] = iftmp.2_4;
57 i_18 = i_23 + 1;
58 if (i_18 <= 15) goto <L19>; else goto <L18>;
59
60 <L19>:;
61 goto <bb 1> (<L0>);
62
63 <L18>:;
64
65 OUTPUT
66 ------
67
68 # i_23 = PHI <0(0), i_18(10)>;
69 <L0>:;
70 j_15 = A[i_23];
71
72 <L3>:;
73 iftmp.2_4 = j_15 > 41 ? 42 : 0;
74 A[i_23] = iftmp.2_4;
75 i_18 = i_23 + 1;
76 if (i_18 <= 15) goto <L19>; else goto <L18>;
77
78 <L19>:;
79 goto <bb 1> (<L0>);
80
81 <L18>:;
82 */
83
84 #include "config.h"
85 #include "system.h"
86 #include "coretypes.h"
87 #include "tm.h"
88 #include "tree.h"
89 #include "flags.h"
90 #include "timevar.h"
91 #include "basic-block.h"
92 #include "tree-pretty-print.h"
93 #include "gimple-pretty-print.h"
94 #include "tree-flow.h"
95 #include "tree-dump.h"
96 #include "cfgloop.h"
97 #include "tree-chrec.h"
98 #include "tree-data-ref.h"
99 #include "tree-scalar-evolution.h"
100 #include "tree-pass.h"
101 #include "dbgcnt.h"
102
103 /* List of basic blocks in if-conversion-suitable order. */
104 static basic_block *ifc_bbs;
105
106 /* Structure used to predicate basic blocks. This is attached to the
107 ->aux field of the BBs in the loop to be if-converted. */
108 typedef struct bb_predicate_s {
109
110 /* The condition under which this basic block is executed. */
111 tree predicate;
112
113 /* PREDICATE is gimplified, and the sequence of statements is
114 recorded here, in order to avoid the duplication of computations
115 that occur in previous conditions. See PR44483. */
116 gimple_seq predicate_gimplified_stmts;
117 } *bb_predicate_p;
118
119 /* Returns true when the basic block BB has a predicate. */
120
121 static inline bool
122 bb_has_predicate (basic_block bb)
123 {
124 return bb->aux != NULL;
125 }
126
127 /* Returns the gimplified predicate for basic block BB. */
128
129 static inline tree
130 bb_predicate (basic_block bb)
131 {
132 return ((bb_predicate_p) bb->aux)->predicate;
133 }
134
135 /* Sets the gimplified predicate COND for basic block BB. */
136
137 static inline void
138 set_bb_predicate (basic_block bb, tree cond)
139 {
140 ((bb_predicate_p) bb->aux)->predicate = cond;
141 }
142
143 /* Returns the sequence of statements of the gimplification of the
144 predicate for basic block BB. */
145
146 static inline gimple_seq
147 bb_predicate_gimplified_stmts (basic_block bb)
148 {
149 return ((bb_predicate_p) bb->aux)->predicate_gimplified_stmts;
150 }
151
152 /* Sets the sequence of statements STMTS of the gimplification of the
153 predicate for basic block BB. */
154
155 static inline void
156 set_bb_predicate_gimplified_stmts (basic_block bb, gimple_seq stmts)
157 {
158 ((bb_predicate_p) bb->aux)->predicate_gimplified_stmts = stmts;
159 }
160
161 /* Adds the sequence of statements STMTS to the sequence of statements
162 of the predicate for basic block BB. */
163
164 static inline void
165 add_bb_predicate_gimplified_stmts (basic_block bb, gimple_seq stmts)
166 {
167 gimple_seq_add_seq
168 (&(((bb_predicate_p) bb->aux)->predicate_gimplified_stmts), stmts);
169 }
170
171 /* Initializes to TRUE the predicate of basic block BB. */
172
173 static inline void
174 init_bb_predicate (basic_block bb)
175 {
176 bb->aux = XNEW (struct bb_predicate_s);
177 set_bb_predicate_gimplified_stmts (bb, NULL);
178 set_bb_predicate (bb, boolean_true_node);
179 }
180
181 /* Free the predicate of basic block BB. */
182
183 static inline void
184 free_bb_predicate (basic_block bb)
185 {
186 gimple_seq stmts;
187
188 if (!bb_has_predicate (bb))
189 return;
190
191 /* Release the SSA_NAMEs created for the gimplification of the
192 predicate. */
193 stmts = bb_predicate_gimplified_stmts (bb);
194 if (stmts)
195 {
196 gimple_stmt_iterator i;
197
198 for (i = gsi_start (stmts); !gsi_end_p (i); gsi_next (&i))
199 free_stmt_operands (gsi_stmt (i));
200 }
201
202 free (bb->aux);
203 bb->aux = NULL;
204 }
205
206 /* Free the predicate of BB and reinitialize it with the true
207 predicate. */
208
209 static inline void
210 reset_bb_predicate (basic_block bb)
211 {
212 free_bb_predicate (bb);
213 init_bb_predicate (bb);
214 }
215
216 /* Returns a new SSA_NAME of type TYPE that is assigned the value of
217 the expression EXPR. Inserts the statement created for this
218 computation before GSI and leaves the iterator GSI at the same
219 statement. */
220
221 static tree
222 ifc_temp_var (tree type, tree expr, gimple_stmt_iterator *gsi)
223 {
224 const char *name = "_ifc_";
225 tree var, new_name;
226 gimple stmt;
227
228 /* Create new temporary variable. */
229 var = create_tmp_var (type, name);
230 add_referenced_var (var);
231
232 /* Build new statement to assign EXPR to new variable. */
233 stmt = gimple_build_assign (var, expr);
234
235 /* Get SSA name for the new variable and set make new statement
236 its definition statement. */
237 new_name = make_ssa_name (var, stmt);
238 gimple_assign_set_lhs (stmt, new_name);
239 SSA_NAME_DEF_STMT (new_name) = stmt;
240 update_stmt (stmt);
241
242 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
243 return gimple_assign_lhs (stmt);
244 }
245
246 /* Return true when COND is a true predicate. */
247
248 static inline bool
249 is_true_predicate (tree cond)
250 {
251 return (cond == NULL_TREE
252 || cond == boolean_true_node
253 || integer_onep (cond));
254 }
255
256 /* Returns true when BB has a predicate that is not trivial: true or
257 NULL_TREE. */
258
259 static inline bool
260 is_predicated (basic_block bb)
261 {
262 return !is_true_predicate (bb_predicate (bb));
263 }
264
265 /* Parses the predicate COND and returns its comparison code and
266 operands OP0 and OP1. */
267
268 static enum tree_code
269 parse_predicate (tree cond, tree *op0, tree *op1)
270 {
271 gimple s;
272
273 if (TREE_CODE (cond) == SSA_NAME
274 && is_gimple_assign (s = SSA_NAME_DEF_STMT (cond)))
275 {
276 if (TREE_CODE_CLASS (gimple_assign_rhs_code (s)) == tcc_comparison)
277 {
278 *op0 = gimple_assign_rhs1 (s);
279 *op1 = gimple_assign_rhs2 (s);
280 return gimple_assign_rhs_code (s);
281 }
282
283 else if (gimple_assign_rhs_code (s) == TRUTH_NOT_EXPR)
284 {
285 tree op = gimple_assign_rhs1 (s);
286 tree type = TREE_TYPE (op);
287 enum tree_code code = parse_predicate (op, op0, op1);
288
289 return code == ERROR_MARK ? ERROR_MARK
290 : invert_tree_comparison (code, HONOR_NANS (TYPE_MODE (type)));
291 }
292
293 return ERROR_MARK;
294 }
295
296 if (TREE_CODE_CLASS (TREE_CODE (cond)) == tcc_comparison)
297 {
298 *op0 = TREE_OPERAND (cond, 0);
299 *op1 = TREE_OPERAND (cond, 1);
300 return TREE_CODE (cond);
301 }
302
303 return ERROR_MARK;
304 }
305
306 /* Returns the fold of predicate C1 OR C2 at location LOC. */
307
308 static tree
309 fold_or_predicates (location_t loc, tree c1, tree c2)
310 {
311 tree op1a, op1b, op2a, op2b;
312 enum tree_code code1 = parse_predicate (c1, &op1a, &op1b);
313 enum tree_code code2 = parse_predicate (c2, &op2a, &op2b);
314
315 if (code1 != ERROR_MARK && code2 != ERROR_MARK)
316 {
317 tree t = maybe_fold_or_comparisons (code1, op1a, op1b,
318 code2, op2a, op2b);
319 if (t)
320 return t;
321 }
322
323 return fold_build2_loc (loc, TRUTH_OR_EXPR, boolean_type_node, c1, c2);
324 }
325
326 /* Add condition NC to the predicate list of basic block BB. */
327
328 static inline void
329 add_to_predicate_list (basic_block bb, tree nc)
330 {
331 tree bc;
332
333 if (is_true_predicate (nc))
334 return;
335
336 if (!is_predicated (bb))
337 bc = nc;
338 else
339 {
340 bc = bb_predicate (bb);
341 bc = fold_or_predicates (EXPR_LOCATION (bc), nc, bc);
342 }
343
344 if (!is_gimple_condexpr (bc))
345 {
346 gimple_seq stmts;
347 bc = force_gimple_operand (bc, &stmts, true, NULL_TREE);
348 add_bb_predicate_gimplified_stmts (bb, stmts);
349 }
350
351 if (is_true_predicate (bc))
352 reset_bb_predicate (bb);
353 else
354 set_bb_predicate (bb, bc);
355 }
356
357 /* Add the condition COND to the previous condition PREV_COND, and add
358 this to the predicate list of the destination of edge E. LOOP is
359 the loop to be if-converted. */
360
361 static void
362 add_to_dst_predicate_list (struct loop *loop, edge e,
363 tree prev_cond, tree cond)
364 {
365 if (!flow_bb_inside_loop_p (loop, e->dest))
366 return;
367
368 if (!is_true_predicate (prev_cond))
369 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
370 prev_cond, cond);
371
372 add_to_predicate_list (e->dest, cond);
373 }
374
375 /* Return true if one of the successor edges of BB exits LOOP. */
376
377 static bool
378 bb_with_exit_edge_p (struct loop *loop, basic_block bb)
379 {
380 edge e;
381 edge_iterator ei;
382
383 FOR_EACH_EDGE (e, ei, bb->succs)
384 if (loop_exit_edge_p (loop, e))
385 return true;
386
387 return false;
388 }
389
390 /* Return true when PHI is if-convertible. PHI is part of loop LOOP
391 and it belongs to basic block BB.
392
393 PHI is not if-convertible if:
394 - it has more than 2 arguments.
395
396 When the flag_tree_loop_if_convert_stores is not set, PHI is not
397 if-convertible if:
398 - a virtual PHI is immediately used in another PHI node,
399 - there is a virtual PHI in a BB other than the loop->header. */
400
401 static bool
402 if_convertible_phi_p (struct loop *loop, basic_block bb, gimple phi)
403 {
404 if (dump_file && (dump_flags & TDF_DETAILS))
405 {
406 fprintf (dump_file, "-------------------------\n");
407 print_gimple_stmt (dump_file, phi, 0, TDF_SLIM);
408 }
409
410 if (bb != loop->header && gimple_phi_num_args (phi) != 2)
411 {
412 if (dump_file && (dump_flags & TDF_DETAILS))
413 fprintf (dump_file, "More than two phi node args.\n");
414 return false;
415 }
416
417 if (flag_tree_loop_if_convert_stores)
418 return true;
419
420 /* When the flag_tree_loop_if_convert_stores is not set, check
421 that there are no memory writes in the branches of the loop to be
422 if-converted. */
423 if (!is_gimple_reg (SSA_NAME_VAR (gimple_phi_result (phi))))
424 {
425 imm_use_iterator imm_iter;
426 use_operand_p use_p;
427
428 if (bb != loop->header)
429 {
430 if (dump_file && (dump_flags & TDF_DETAILS))
431 fprintf (dump_file, "Virtual phi not on loop->header.\n");
432 return false;
433 }
434
435 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_phi_result (phi))
436 {
437 if (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI)
438 {
439 if (dump_file && (dump_flags & TDF_DETAILS))
440 fprintf (dump_file, "Difficult to handle this virtual phi.\n");
441 return false;
442 }
443 }
444 }
445
446 return true;
447 }
448
449 /* Records the status of a data reference. This struct is attached to
450 each DR->aux field. */
451
452 struct ifc_dr {
453 /* -1 when not initialized, 0 when false, 1 when true. */
454 int written_at_least_once;
455
456 /* -1 when not initialized, 0 when false, 1 when true. */
457 int rw_unconditionally;
458 };
459
460 #define IFC_DR(DR) ((struct ifc_dr *) (DR)->aux)
461 #define DR_WRITTEN_AT_LEAST_ONCE(DR) (IFC_DR (DR)->written_at_least_once)
462 #define DR_RW_UNCONDITIONALLY(DR) (IFC_DR (DR)->rw_unconditionally)
463
464 /* Returns true when the memory references of STMT are read or written
465 unconditionally. In other words, this function returns true when
466 for every data reference A in STMT there exist other accesses to
467 a data reference with the same base with predicates that add up (OR-up) to
468 the true predicate: this ensures that the data reference A is touched
469 (read or written) on every iteration of the if-converted loop. */
470
471 static bool
472 memrefs_read_or_written_unconditionally (gimple stmt,
473 VEC (data_reference_p, heap) *drs)
474 {
475 int i, j;
476 data_reference_p a, b;
477 tree ca = bb_predicate (gimple_bb (stmt));
478
479 for (i = 0; VEC_iterate (data_reference_p, drs, i, a); i++)
480 if (DR_STMT (a) == stmt)
481 {
482 bool found = false;
483 int x = DR_RW_UNCONDITIONALLY (a);
484
485 if (x == 0)
486 return false;
487
488 if (x == 1)
489 continue;
490
491 for (j = 0; VEC_iterate (data_reference_p, drs, j, b); j++)
492 {
493 tree ref_base_a = DR_REF (a);
494 tree ref_base_b = DR_REF (b);
495
496 if (DR_STMT (b) == stmt)
497 continue;
498
499 while (TREE_CODE (ref_base_a) == COMPONENT_REF
500 || TREE_CODE (ref_base_a) == IMAGPART_EXPR
501 || TREE_CODE (ref_base_a) == REALPART_EXPR)
502 ref_base_a = TREE_OPERAND (ref_base_a, 0);
503
504 while (TREE_CODE (ref_base_b) == COMPONENT_REF
505 || TREE_CODE (ref_base_b) == IMAGPART_EXPR
506 || TREE_CODE (ref_base_b) == REALPART_EXPR)
507 ref_base_b = TREE_OPERAND (ref_base_b, 0);
508
509 if (!operand_equal_p (ref_base_a, ref_base_b, 0))
510 {
511 tree cb = bb_predicate (gimple_bb (DR_STMT (b)));
512
513 if (DR_RW_UNCONDITIONALLY (b) == 1
514 || is_true_predicate (cb)
515 || is_true_predicate (ca
516 = fold_or_predicates (EXPR_LOCATION (cb), ca, cb)))
517 {
518 DR_RW_UNCONDITIONALLY (a) = 1;
519 DR_RW_UNCONDITIONALLY (b) = 1;
520 found = true;
521 break;
522 }
523 }
524 }
525
526 if (!found)
527 {
528 DR_RW_UNCONDITIONALLY (a) = 0;
529 return false;
530 }
531 }
532
533 return true;
534 }
535
536 /* Returns true when the memory references of STMT are unconditionally
537 written. In other words, this function returns true when for every
538 data reference A written in STMT, there exist other writes to the
539 same data reference with predicates that add up (OR-up) to the true
540 predicate: this ensures that the data reference A is written on
541 every iteration of the if-converted loop. */
542
543 static bool
544 write_memrefs_written_at_least_once (gimple stmt,
545 VEC (data_reference_p, heap) *drs)
546 {
547 int i, j;
548 data_reference_p a, b;
549 tree ca = bb_predicate (gimple_bb (stmt));
550
551 for (i = 0; VEC_iterate (data_reference_p, drs, i, a); i++)
552 if (DR_STMT (a) == stmt
553 && DR_IS_WRITE (a))
554 {
555 bool found = false;
556 int x = DR_WRITTEN_AT_LEAST_ONCE (a);
557
558 if (x == 0)
559 return false;
560
561 if (x == 1)
562 continue;
563
564 for (j = 0; VEC_iterate (data_reference_p, drs, j, b); j++)
565 if (DR_STMT (b) != stmt
566 && DR_IS_WRITE (b)
567 && same_data_refs_base_objects (a, b))
568 {
569 tree cb = bb_predicate (gimple_bb (DR_STMT (b)));
570
571 if (DR_WRITTEN_AT_LEAST_ONCE (b) == 1
572 || is_true_predicate (cb)
573 || is_true_predicate (ca = fold_or_predicates (EXPR_LOCATION (cb),
574 ca, cb)))
575 {
576 DR_WRITTEN_AT_LEAST_ONCE (a) = 1;
577 DR_WRITTEN_AT_LEAST_ONCE (b) = 1;
578 found = true;
579 break;
580 }
581 }
582
583 if (!found)
584 {
585 DR_WRITTEN_AT_LEAST_ONCE (a) = 0;
586 return false;
587 }
588 }
589
590 return true;
591 }
592
593 /* Return true when the memory references of STMT won't trap in the
594 if-converted code. There are two things that we have to check for:
595
596 - writes to memory occur to writable memory: if-conversion of
597 memory writes transforms the conditional memory writes into
598 unconditional writes, i.e. "if (cond) A[i] = foo" is transformed
599 into "A[i] = cond ? foo : A[i]", and as the write to memory may not
600 be executed at all in the original code, it may be a readonly
601 memory. To check that A is not const-qualified, we check that
602 there exists at least an unconditional write to A in the current
603 function.
604
605 - reads or writes to memory are valid memory accesses for every
606 iteration. To check that the memory accesses are correctly formed
607 and that we are allowed to read and write in these locations, we
608 check that the memory accesses to be if-converted occur at every
609 iteration unconditionally. */
610
611 static bool
612 ifcvt_memrefs_wont_trap (gimple stmt, VEC (data_reference_p, heap) *refs)
613 {
614 return write_memrefs_written_at_least_once (stmt, refs)
615 && memrefs_read_or_written_unconditionally (stmt, refs);
616 }
617
618 /* Wrapper around gimple_could_trap_p refined for the needs of the
619 if-conversion. Try to prove that the memory accesses of STMT could
620 not trap in the innermost loop containing STMT. */
621
622 static bool
623 ifcvt_could_trap_p (gimple stmt, VEC (data_reference_p, heap) *refs)
624 {
625 if (gimple_vuse (stmt)
626 && !gimple_could_trap_p_1 (stmt, false, false)
627 && ifcvt_memrefs_wont_trap (stmt, refs))
628 return false;
629
630 return gimple_could_trap_p (stmt);
631 }
632
633 /* Return true when STMT is if-convertible.
634
635 GIMPLE_ASSIGN statement is not if-convertible if,
636 - it is not movable,
637 - it could trap,
638 - LHS is not var decl. */
639
640 static bool
641 if_convertible_gimple_assign_stmt_p (gimple stmt,
642 VEC (data_reference_p, heap) *refs)
643 {
644 tree lhs = gimple_assign_lhs (stmt);
645 basic_block bb;
646
647 if (dump_file && (dump_flags & TDF_DETAILS))
648 {
649 fprintf (dump_file, "-------------------------\n");
650 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
651 }
652
653 if (!is_gimple_reg_type (TREE_TYPE (lhs)))
654 return false;
655
656 /* Some of these constrains might be too conservative. */
657 if (stmt_ends_bb_p (stmt)
658 || gimple_has_volatile_ops (stmt)
659 || (TREE_CODE (lhs) == SSA_NAME
660 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
661 || gimple_has_side_effects (stmt))
662 {
663 if (dump_file && (dump_flags & TDF_DETAILS))
664 fprintf (dump_file, "stmt not suitable for ifcvt\n");
665 return false;
666 }
667
668 if (flag_tree_loop_if_convert_stores)
669 {
670 if (ifcvt_could_trap_p (stmt, refs))
671 {
672 if (dump_file && (dump_flags & TDF_DETAILS))
673 fprintf (dump_file, "tree could trap...\n");
674 return false;
675 }
676 return true;
677 }
678
679 if (gimple_assign_rhs_could_trap_p (stmt))
680 {
681 if (dump_file && (dump_flags & TDF_DETAILS))
682 fprintf (dump_file, "tree could trap...\n");
683 return false;
684 }
685
686 bb = gimple_bb (stmt);
687
688 if (TREE_CODE (lhs) != SSA_NAME
689 && bb != bb->loop_father->header
690 && !bb_with_exit_edge_p (bb->loop_father, bb))
691 {
692 if (dump_file && (dump_flags & TDF_DETAILS))
693 {
694 fprintf (dump_file, "LHS is not var\n");
695 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
696 }
697 return false;
698 }
699
700 return true;
701 }
702
703 /* Return true when STMT is if-convertible.
704
705 A statement is if-convertible if:
706 - it is an if-convertible GIMPLE_ASSGIN,
707 - it is a GIMPLE_LABEL or a GIMPLE_COND. */
708
709 static bool
710 if_convertible_stmt_p (gimple stmt, VEC (data_reference_p, heap) *refs)
711 {
712 switch (gimple_code (stmt))
713 {
714 case GIMPLE_LABEL:
715 case GIMPLE_DEBUG:
716 case GIMPLE_COND:
717 return true;
718
719 case GIMPLE_ASSIGN:
720 return if_convertible_gimple_assign_stmt_p (stmt, refs);
721
722 case GIMPLE_CALL:
723 {
724 tree fndecl = gimple_call_fndecl (stmt);
725 if (fndecl)
726 {
727 int flags = gimple_call_flags (stmt);
728 if ((flags & ECF_CONST)
729 && !(flags & ECF_LOOPING_CONST_OR_PURE)
730 /* We can only vectorize some builtins at the moment,
731 so restrict if-conversion to those. */
732 && DECL_BUILT_IN (fndecl))
733 return true;
734 }
735 return false;
736 }
737
738 default:
739 /* Don't know what to do with 'em so don't do anything. */
740 if (dump_file && (dump_flags & TDF_DETAILS))
741 {
742 fprintf (dump_file, "don't know what to do\n");
743 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
744 }
745 return false;
746 break;
747 }
748
749 return true;
750 }
751
752 /* Return true when BB post-dominates all its predecessors. */
753
754 static bool
755 bb_postdominates_preds (basic_block bb)
756 {
757 unsigned i;
758
759 for (i = 0; i < EDGE_COUNT (bb->preds); i++)
760 if (!dominated_by_p (CDI_POST_DOMINATORS, EDGE_PRED (bb, i)->src, bb))
761 return false;
762
763 return true;
764 }
765
766 /* Return true when BB is if-convertible. This routine does not check
767 basic block's statements and phis.
768
769 A basic block is not if-convertible if:
770 - it is non-empty and it is after the exit block (in BFS order),
771 - it is after the exit block but before the latch,
772 - its edges are not normal.
773
774 EXIT_BB is the basic block containing the exit of the LOOP. BB is
775 inside LOOP. */
776
777 static bool
778 if_convertible_bb_p (struct loop *loop, basic_block bb, basic_block exit_bb)
779 {
780 edge e;
781 edge_iterator ei;
782
783 if (dump_file && (dump_flags & TDF_DETAILS))
784 fprintf (dump_file, "----------[%d]-------------\n", bb->index);
785
786 if (EDGE_COUNT (bb->preds) > 2
787 || EDGE_COUNT (bb->succs) > 2)
788 return false;
789
790 if (exit_bb)
791 {
792 if (bb != loop->latch)
793 {
794 if (dump_file && (dump_flags & TDF_DETAILS))
795 fprintf (dump_file, "basic block after exit bb but before latch\n");
796 return false;
797 }
798 else if (!empty_block_p (bb))
799 {
800 if (dump_file && (dump_flags & TDF_DETAILS))
801 fprintf (dump_file, "non empty basic block after exit bb\n");
802 return false;
803 }
804 else if (bb == loop->latch
805 && bb != exit_bb
806 && !dominated_by_p (CDI_DOMINATORS, bb, exit_bb))
807 {
808 if (dump_file && (dump_flags & TDF_DETAILS))
809 fprintf (dump_file, "latch is not dominated by exit_block\n");
810 return false;
811 }
812 }
813
814 /* Be less adventurous and handle only normal edges. */
815 FOR_EACH_EDGE (e, ei, bb->succs)
816 if (e->flags &
817 (EDGE_ABNORMAL_CALL | EDGE_EH | EDGE_ABNORMAL | EDGE_IRREDUCIBLE_LOOP))
818 {
819 if (dump_file && (dump_flags & TDF_DETAILS))
820 fprintf (dump_file, "Difficult to handle edges\n");
821 return false;
822 }
823
824 if (EDGE_COUNT (bb->preds) == 2
825 && bb != loop->header
826 && !bb_postdominates_preds (bb))
827 return false;
828
829 return true;
830 }
831
832 /* Return true when all predecessor blocks of BB are visited. The
833 VISITED bitmap keeps track of the visited blocks. */
834
835 static bool
836 pred_blocks_visited_p (basic_block bb, bitmap *visited)
837 {
838 edge e;
839 edge_iterator ei;
840 FOR_EACH_EDGE (e, ei, bb->preds)
841 if (!bitmap_bit_p (*visited, e->src->index))
842 return false;
843
844 return true;
845 }
846
847 /* Get body of a LOOP in suitable order for if-conversion. It is
848 caller's responsibility to deallocate basic block list.
849 If-conversion suitable order is, breadth first sort (BFS) order
850 with an additional constraint: select a block only if all its
851 predecessors are already selected. */
852
853 static basic_block *
854 get_loop_body_in_if_conv_order (const struct loop *loop)
855 {
856 basic_block *blocks, *blocks_in_bfs_order;
857 basic_block bb;
858 bitmap visited;
859 unsigned int index = 0;
860 unsigned int visited_count = 0;
861
862 gcc_assert (loop->num_nodes);
863 gcc_assert (loop->latch != EXIT_BLOCK_PTR);
864
865 blocks = XCNEWVEC (basic_block, loop->num_nodes);
866 visited = BITMAP_ALLOC (NULL);
867
868 blocks_in_bfs_order = get_loop_body_in_bfs_order (loop);
869
870 index = 0;
871 while (index < loop->num_nodes)
872 {
873 bb = blocks_in_bfs_order [index];
874
875 if (bb->flags & BB_IRREDUCIBLE_LOOP)
876 {
877 free (blocks_in_bfs_order);
878 BITMAP_FREE (visited);
879 free (blocks);
880 return NULL;
881 }
882
883 if (!bitmap_bit_p (visited, bb->index))
884 {
885 if (pred_blocks_visited_p (bb, &visited)
886 || bb == loop->header)
887 {
888 /* This block is now visited. */
889 bitmap_set_bit (visited, bb->index);
890 blocks[visited_count++] = bb;
891 }
892 }
893
894 index++;
895
896 if (index == loop->num_nodes
897 && visited_count != loop->num_nodes)
898 /* Not done yet. */
899 index = 0;
900 }
901 free (blocks_in_bfs_order);
902 BITMAP_FREE (visited);
903 return blocks;
904 }
905
906 /* Returns true when the analysis of the predicates for all the basic
907 blocks in LOOP succeeded.
908
909 predicate_bbs first allocates the predicates of the basic blocks.
910 These fields are then initialized with the tree expressions
911 representing the predicates under which a basic block is executed
912 in the LOOP. As the loop->header is executed at each iteration, it
913 has the "true" predicate. Other statements executed under a
914 condition are predicated with that condition, for example
915
916 | if (x)
917 | S1;
918 | else
919 | S2;
920
921 S1 will be predicated with "x", and
922 S2 will be predicated with "!x". */
923
924 static bool
925 predicate_bbs (loop_p loop)
926 {
927 unsigned int i;
928
929 for (i = 0; i < loop->num_nodes; i++)
930 init_bb_predicate (ifc_bbs[i]);
931
932 for (i = 0; i < loop->num_nodes; i++)
933 {
934 basic_block bb = ifc_bbs[i];
935 tree cond;
936 gimple_stmt_iterator itr;
937
938 /* The loop latch is always executed and has no extra conditions
939 to be processed: skip it. */
940 if (bb == loop->latch)
941 {
942 reset_bb_predicate (loop->latch);
943 continue;
944 }
945
946 cond = bb_predicate (bb);
947 if (cond
948 && bb != loop->header)
949 {
950 gimple_seq stmts;
951
952 cond = force_gimple_operand (cond, &stmts, true, NULL_TREE);
953 add_bb_predicate_gimplified_stmts (bb, stmts);
954 }
955
956 for (itr = gsi_start_bb (bb); !gsi_end_p (itr); gsi_next (&itr))
957 {
958 gimple stmt = gsi_stmt (itr);
959
960 switch (gimple_code (stmt))
961 {
962 case GIMPLE_LABEL:
963 case GIMPLE_ASSIGN:
964 case GIMPLE_CALL:
965 case GIMPLE_DEBUG:
966 break;
967
968 case GIMPLE_COND:
969 {
970 tree c2, tem;
971 edge true_edge, false_edge;
972 location_t loc = gimple_location (stmt);
973 tree c = fold_build2_loc (loc, gimple_cond_code (stmt),
974 boolean_type_node,
975 gimple_cond_lhs (stmt),
976 gimple_cond_rhs (stmt));
977
978 /* Add new condition into destination's predicate list. */
979 extract_true_false_edges_from_block (gimple_bb (stmt),
980 &true_edge, &false_edge);
981
982 /* If C is true, then TRUE_EDGE is taken. */
983 add_to_dst_predicate_list (loop, true_edge, cond, unshare_expr (c));
984
985 /* If C is false, then FALSE_EDGE is taken. */
986 c2 = invert_truthvalue_loc (loc, unshare_expr (c));
987 tem = canonicalize_cond_expr_cond (c2);
988 if (tem)
989 c2 = tem;
990 add_to_dst_predicate_list (loop, false_edge, cond, c2);
991
992 cond = NULL_TREE;
993 break;
994 }
995
996 default:
997 /* Not handled yet in if-conversion. */
998 return false;
999 }
1000 }
1001
1002 /* If current bb has only one successor, then consider it as an
1003 unconditional goto. */
1004 if (single_succ_p (bb))
1005 {
1006 basic_block bb_n = single_succ (bb);
1007
1008 /* The successor bb inherits the predicate of its
1009 predecessor. If there is no predicate in the predecessor
1010 bb, then consider the successor bb as always executed. */
1011 if (cond == NULL_TREE)
1012 cond = boolean_true_node;
1013
1014 add_to_predicate_list (bb_n, cond);
1015 }
1016 }
1017
1018 /* The loop header is always executed. */
1019 reset_bb_predicate (loop->header);
1020 gcc_assert (bb_predicate_gimplified_stmts (loop->header) == NULL
1021 && bb_predicate_gimplified_stmts (loop->latch) == NULL);
1022
1023 return true;
1024 }
1025
1026 /* Return true when LOOP is if-convertible. This is a helper function
1027 for if_convertible_loop_p. REFS and DDRS are initialized and freed
1028 in if_convertible_loop_p. */
1029
1030 static bool
1031 if_convertible_loop_p_1 (struct loop *loop,
1032 VEC (loop_p, heap) **loop_nest,
1033 VEC (data_reference_p, heap) **refs,
1034 VEC (ddr_p, heap) **ddrs)
1035 {
1036 bool res;
1037 unsigned int i;
1038 basic_block exit_bb = NULL;
1039
1040 /* Don't if-convert the loop when the data dependences cannot be
1041 computed: the loop won't be vectorized in that case. */
1042 res = compute_data_dependences_for_loop (loop, true, loop_nest, refs, ddrs);
1043 if (!res)
1044 return false;
1045
1046 calculate_dominance_info (CDI_DOMINATORS);
1047 calculate_dominance_info (CDI_POST_DOMINATORS);
1048
1049 /* Allow statements that can be handled during if-conversion. */
1050 ifc_bbs = get_loop_body_in_if_conv_order (loop);
1051 if (!ifc_bbs)
1052 {
1053 if (dump_file && (dump_flags & TDF_DETAILS))
1054 fprintf (dump_file, "Irreducible loop\n");
1055 return false;
1056 }
1057
1058 for (i = 0; i < loop->num_nodes; i++)
1059 {
1060 basic_block bb = ifc_bbs[i];
1061
1062 if (!if_convertible_bb_p (loop, bb, exit_bb))
1063 return false;
1064
1065 if (bb_with_exit_edge_p (loop, bb))
1066 exit_bb = bb;
1067 }
1068
1069 res = predicate_bbs (loop);
1070 if (!res)
1071 return false;
1072
1073 if (flag_tree_loop_if_convert_stores)
1074 {
1075 data_reference_p dr;
1076
1077 for (i = 0; VEC_iterate (data_reference_p, *refs, i, dr); i++)
1078 {
1079 dr->aux = XNEW (struct ifc_dr);
1080 DR_WRITTEN_AT_LEAST_ONCE (dr) = -1;
1081 DR_RW_UNCONDITIONALLY (dr) = -1;
1082 }
1083 }
1084
1085 for (i = 0; i < loop->num_nodes; i++)
1086 {
1087 basic_block bb = ifc_bbs[i];
1088 gimple_stmt_iterator itr;
1089
1090 for (itr = gsi_start_phis (bb); !gsi_end_p (itr); gsi_next (&itr))
1091 if (!if_convertible_phi_p (loop, bb, gsi_stmt (itr)))
1092 return false;
1093
1094 /* Check the if-convertibility of statements in predicated BBs. */
1095 if (is_predicated (bb))
1096 for (itr = gsi_start_bb (bb); !gsi_end_p (itr); gsi_next (&itr))
1097 if (!if_convertible_stmt_p (gsi_stmt (itr), *refs))
1098 return false;
1099 }
1100
1101 if (dump_file)
1102 fprintf (dump_file, "Applying if-conversion\n");
1103
1104 return true;
1105 }
1106
1107 /* Return true when LOOP is if-convertible.
1108 LOOP is if-convertible if:
1109 - it is innermost,
1110 - it has two or more basic blocks,
1111 - it has only one exit,
1112 - loop header is not the exit edge,
1113 - if its basic blocks and phi nodes are if convertible. */
1114
1115 static bool
1116 if_convertible_loop_p (struct loop *loop)
1117 {
1118 edge e;
1119 edge_iterator ei;
1120 bool res = false;
1121 VEC (data_reference_p, heap) *refs;
1122 VEC (ddr_p, heap) *ddrs;
1123 VEC (loop_p, heap) *loop_nest;
1124
1125 /* Handle only innermost loop. */
1126 if (!loop || loop->inner)
1127 {
1128 if (dump_file && (dump_flags & TDF_DETAILS))
1129 fprintf (dump_file, "not innermost loop\n");
1130 return false;
1131 }
1132
1133 /* If only one block, no need for if-conversion. */
1134 if (loop->num_nodes <= 2)
1135 {
1136 if (dump_file && (dump_flags & TDF_DETAILS))
1137 fprintf (dump_file, "less than 2 basic blocks\n");
1138 return false;
1139 }
1140
1141 /* More than one loop exit is too much to handle. */
1142 if (!single_exit (loop))
1143 {
1144 if (dump_file && (dump_flags & TDF_DETAILS))
1145 fprintf (dump_file, "multiple exits\n");
1146 return false;
1147 }
1148
1149 /* If one of the loop header's edge is an exit edge then do not
1150 apply if-conversion. */
1151 FOR_EACH_EDGE (e, ei, loop->header->succs)
1152 if (loop_exit_edge_p (loop, e))
1153 return false;
1154
1155 refs = VEC_alloc (data_reference_p, heap, 5);
1156 ddrs = VEC_alloc (ddr_p, heap, 25);
1157 loop_nest = VEC_alloc (loop_p, heap, 3);
1158 res = if_convertible_loop_p_1 (loop, &loop_nest, &refs, &ddrs);
1159
1160 if (flag_tree_loop_if_convert_stores)
1161 {
1162 data_reference_p dr;
1163 unsigned int i;
1164
1165 for (i = 0; VEC_iterate (data_reference_p, refs, i, dr); i++)
1166 free (dr->aux);
1167 }
1168
1169 VEC_free (loop_p, heap, loop_nest);
1170 free_data_refs (refs);
1171 free_dependence_relations (ddrs);
1172 return res;
1173 }
1174
1175 /* Basic block BB has two predecessors. Using predecessor's bb
1176 predicate, set an appropriate condition COND for the PHI node
1177 replacement. Return the true block whose phi arguments are
1178 selected when cond is true. LOOP is the loop containing the
1179 if-converted region, GSI is the place to insert the code for the
1180 if-conversion. */
1181
1182 static basic_block
1183 find_phi_replacement_condition (struct loop *loop,
1184 basic_block bb, tree *cond,
1185 gimple_stmt_iterator *gsi)
1186 {
1187 edge first_edge, second_edge;
1188 tree tmp_cond;
1189
1190 gcc_assert (EDGE_COUNT (bb->preds) == 2);
1191 first_edge = EDGE_PRED (bb, 0);
1192 second_edge = EDGE_PRED (bb, 1);
1193
1194 /* Use condition based on following criteria:
1195 1)
1196 S1: x = !c ? a : b;
1197
1198 S2: x = c ? b : a;
1199
1200 S2 is preferred over S1. Make 'b' first_bb and use its condition.
1201
1202 2) Do not make loop header first_bb.
1203
1204 3)
1205 S1: x = !(c == d)? a : b;
1206
1207 S21: t1 = c == d;
1208 S22: x = t1 ? b : a;
1209
1210 S3: x = (c == d) ? b : a;
1211
1212 S3 is preferred over S1 and S2*, Make 'b' first_bb and use
1213 its condition.
1214
1215 4) If pred B is dominated by pred A then use pred B's condition.
1216 See PR23115. */
1217
1218 /* Select condition that is not TRUTH_NOT_EXPR. */
1219 tmp_cond = bb_predicate (first_edge->src);
1220 gcc_assert (tmp_cond);
1221
1222 if (TREE_CODE (tmp_cond) == TRUTH_NOT_EXPR)
1223 {
1224 edge tmp_edge;
1225
1226 tmp_edge = first_edge;
1227 first_edge = second_edge;
1228 second_edge = tmp_edge;
1229 }
1230
1231 /* Check if FIRST_BB is loop header or not and make sure that
1232 FIRST_BB does not dominate SECOND_BB. */
1233 if (first_edge->src == loop->header
1234 || dominated_by_p (CDI_DOMINATORS,
1235 second_edge->src, first_edge->src))
1236 {
1237 *cond = bb_predicate (second_edge->src);
1238
1239 if (TREE_CODE (*cond) == TRUTH_NOT_EXPR)
1240 *cond = invert_truthvalue (*cond);
1241 else
1242 /* Select non loop header bb. */
1243 first_edge = second_edge;
1244 }
1245 else
1246 *cond = bb_predicate (first_edge->src);
1247
1248 /* Gimplify the condition: the vectorizer prefers to have gimple
1249 values as conditions. Various targets use different means to
1250 communicate conditions in vector compare operations. Using a
1251 gimple value allows the compiler to emit vector compare and
1252 select RTL without exposing compare's result. */
1253 *cond = force_gimple_operand_gsi (gsi, unshare_expr (*cond),
1254 false, NULL_TREE,
1255 true, GSI_SAME_STMT);
1256 if (!is_gimple_reg (*cond) && !is_gimple_condexpr (*cond))
1257 *cond = ifc_temp_var (TREE_TYPE (*cond), unshare_expr (*cond), gsi);
1258
1259 gcc_assert (*cond);
1260
1261 return first_edge->src;
1262 }
1263
1264 /* Replace a scalar PHI node with a COND_EXPR using COND as condition.
1265 This routine does not handle PHI nodes with more than two
1266 arguments.
1267
1268 For example,
1269 S1: A = PHI <x1(1), x2(5)
1270 is converted into,
1271 S2: A = cond ? x1 : x2;
1272
1273 The generated code is inserted at GSI that points to the top of
1274 basic block's statement list. When COND is true, phi arg from
1275 TRUE_BB is selected. */
1276
1277 static void
1278 predicate_scalar_phi (gimple phi, tree cond,
1279 basic_block true_bb,
1280 gimple_stmt_iterator *gsi)
1281 {
1282 gimple new_stmt;
1283 basic_block bb;
1284 tree rhs, res, arg, scev;
1285
1286 gcc_assert (gimple_code (phi) == GIMPLE_PHI
1287 && gimple_phi_num_args (phi) == 2);
1288
1289 res = gimple_phi_result (phi);
1290 /* Do not handle virtual phi nodes. */
1291 if (!is_gimple_reg (SSA_NAME_VAR (res)))
1292 return;
1293
1294 bb = gimple_bb (phi);
1295
1296 if ((arg = degenerate_phi_result (phi))
1297 || ((scev = analyze_scalar_evolution (gimple_bb (phi)->loop_father,
1298 res))
1299 && !chrec_contains_undetermined (scev)
1300 && scev != res
1301 && (arg = gimple_phi_arg_def (phi, 0))))
1302 rhs = arg;
1303 else
1304 {
1305 tree arg_0, arg_1;
1306 /* Use condition that is not TRUTH_NOT_EXPR in conditional modify expr. */
1307 if (EDGE_PRED (bb, 1)->src == true_bb)
1308 {
1309 arg_0 = gimple_phi_arg_def (phi, 1);
1310 arg_1 = gimple_phi_arg_def (phi, 0);
1311 }
1312 else
1313 {
1314 arg_0 = gimple_phi_arg_def (phi, 0);
1315 arg_1 = gimple_phi_arg_def (phi, 1);
1316 }
1317
1318 gcc_checking_assert (bb == bb->loop_father->header
1319 || bb_postdominates_preds (bb));
1320
1321 /* Build new RHS using selected condition and arguments. */
1322 rhs = build3 (COND_EXPR, TREE_TYPE (res),
1323 unshare_expr (cond), arg_0, arg_1);
1324 }
1325
1326 new_stmt = gimple_build_assign (res, rhs);
1327 SSA_NAME_DEF_STMT (gimple_phi_result (phi)) = new_stmt;
1328 gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
1329 update_stmt (new_stmt);
1330
1331 if (dump_file && (dump_flags & TDF_DETAILS))
1332 {
1333 fprintf (dump_file, "new phi replacement stmt\n");
1334 print_gimple_stmt (dump_file, new_stmt, 0, TDF_SLIM);
1335 }
1336 }
1337
1338 /* Replaces in LOOP all the scalar phi nodes other than those in the
1339 LOOP->header block with conditional modify expressions. */
1340
1341 static void
1342 predicate_all_scalar_phis (struct loop *loop)
1343 {
1344 basic_block bb;
1345 unsigned int orig_loop_num_nodes = loop->num_nodes;
1346 unsigned int i;
1347
1348 for (i = 1; i < orig_loop_num_nodes; i++)
1349 {
1350 gimple phi;
1351 tree cond = NULL_TREE;
1352 gimple_stmt_iterator gsi, phi_gsi;
1353 basic_block true_bb = NULL;
1354 bb = ifc_bbs[i];
1355
1356 if (bb == loop->header)
1357 continue;
1358
1359 phi_gsi = gsi_start_phis (bb);
1360 if (gsi_end_p (phi_gsi))
1361 continue;
1362
1363 /* BB has two predecessors. Using predecessor's aux field, set
1364 appropriate condition for the PHI node replacement. */
1365 gsi = gsi_after_labels (bb);
1366 true_bb = find_phi_replacement_condition (loop, bb, &cond, &gsi);
1367
1368 while (!gsi_end_p (phi_gsi))
1369 {
1370 phi = gsi_stmt (phi_gsi);
1371 predicate_scalar_phi (phi, cond, true_bb, &gsi);
1372 release_phi_node (phi);
1373 gsi_next (&phi_gsi);
1374 }
1375
1376 set_phi_nodes (bb, NULL);
1377 }
1378 }
1379
1380 /* Insert in each basic block of LOOP the statements produced by the
1381 gimplification of the predicates. */
1382
1383 static void
1384 insert_gimplified_predicates (loop_p loop)
1385 {
1386 unsigned int i;
1387
1388 for (i = 0; i < loop->num_nodes; i++)
1389 {
1390 basic_block bb = ifc_bbs[i];
1391 gimple_seq stmts;
1392
1393 if (!is_predicated (bb))
1394 {
1395 /* Do not insert statements for a basic block that is not
1396 predicated. Also make sure that the predicate of the
1397 basic block is set to true. */
1398 reset_bb_predicate (bb);
1399 continue;
1400 }
1401
1402 stmts = bb_predicate_gimplified_stmts (bb);
1403 if (stmts)
1404 {
1405 if (flag_tree_loop_if_convert_stores)
1406 {
1407 /* Insert the predicate of the BB just after the label,
1408 as the if-conversion of memory writes will use this
1409 predicate. */
1410 gimple_stmt_iterator gsi = gsi_after_labels (bb);
1411 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
1412 }
1413 else
1414 {
1415 /* Insert the predicate of the BB at the end of the BB
1416 as this would reduce the register pressure: the only
1417 use of this predicate will be in successor BBs. */
1418 gimple_stmt_iterator gsi = gsi_last_bb (bb);
1419
1420 if (gsi_end_p (gsi)
1421 || stmt_ends_bb_p (gsi_stmt (gsi)))
1422 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
1423 else
1424 gsi_insert_seq_after (&gsi, stmts, GSI_SAME_STMT);
1425 }
1426
1427 /* Once the sequence is code generated, set it to NULL. */
1428 set_bb_predicate_gimplified_stmts (bb, NULL);
1429 }
1430 }
1431 }
1432
1433 /* Predicate each write to memory in LOOP.
1434
1435 This function transforms control flow constructs containing memory
1436 writes of the form:
1437
1438 | for (i = 0; i < N; i++)
1439 | if (cond)
1440 | A[i] = expr;
1441
1442 into the following form that does not contain control flow:
1443
1444 | for (i = 0; i < N; i++)
1445 | A[i] = cond ? expr : A[i];
1446
1447 The original CFG looks like this:
1448
1449 | bb_0
1450 | i = 0
1451 | end_bb_0
1452 |
1453 | bb_1
1454 | if (i < N) goto bb_5 else goto bb_2
1455 | end_bb_1
1456 |
1457 | bb_2
1458 | cond = some_computation;
1459 | if (cond) goto bb_3 else goto bb_4
1460 | end_bb_2
1461 |
1462 | bb_3
1463 | A[i] = expr;
1464 | goto bb_4
1465 | end_bb_3
1466 |
1467 | bb_4
1468 | goto bb_1
1469 | end_bb_4
1470
1471 insert_gimplified_predicates inserts the computation of the COND
1472 expression at the beginning of the destination basic block:
1473
1474 | bb_0
1475 | i = 0
1476 | end_bb_0
1477 |
1478 | bb_1
1479 | if (i < N) goto bb_5 else goto bb_2
1480 | end_bb_1
1481 |
1482 | bb_2
1483 | cond = some_computation;
1484 | if (cond) goto bb_3 else goto bb_4
1485 | end_bb_2
1486 |
1487 | bb_3
1488 | cond = some_computation;
1489 | A[i] = expr;
1490 | goto bb_4
1491 | end_bb_3
1492 |
1493 | bb_4
1494 | goto bb_1
1495 | end_bb_4
1496
1497 predicate_mem_writes is then predicating the memory write as follows:
1498
1499 | bb_0
1500 | i = 0
1501 | end_bb_0
1502 |
1503 | bb_1
1504 | if (i < N) goto bb_5 else goto bb_2
1505 | end_bb_1
1506 |
1507 | bb_2
1508 | if (cond) goto bb_3 else goto bb_4
1509 | end_bb_2
1510 |
1511 | bb_3
1512 | cond = some_computation;
1513 | A[i] = cond ? expr : A[i];
1514 | goto bb_4
1515 | end_bb_3
1516 |
1517 | bb_4
1518 | goto bb_1
1519 | end_bb_4
1520
1521 and finally combine_blocks removes the basic block boundaries making
1522 the loop vectorizable:
1523
1524 | bb_0
1525 | i = 0
1526 | if (i < N) goto bb_5 else goto bb_1
1527 | end_bb_0
1528 |
1529 | bb_1
1530 | cond = some_computation;
1531 | A[i] = cond ? expr : A[i];
1532 | if (i < N) goto bb_5 else goto bb_4
1533 | end_bb_1
1534 |
1535 | bb_4
1536 | goto bb_1
1537 | end_bb_4
1538 */
1539
1540 static void
1541 predicate_mem_writes (loop_p loop)
1542 {
1543 unsigned int i, orig_loop_num_nodes = loop->num_nodes;
1544
1545 for (i = 1; i < orig_loop_num_nodes; i++)
1546 {
1547 gimple_stmt_iterator gsi;
1548 basic_block bb = ifc_bbs[i];
1549 tree cond = bb_predicate (bb);
1550 gimple stmt;
1551
1552 if (is_true_predicate (cond))
1553 continue;
1554
1555 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1556 if ((stmt = gsi_stmt (gsi))
1557 && gimple_assign_single_p (stmt)
1558 && gimple_vdef (stmt))
1559 {
1560 tree lhs = gimple_assign_lhs (stmt);
1561 tree rhs = gimple_assign_rhs1 (stmt);
1562 tree type = TREE_TYPE (lhs);
1563
1564 lhs = ifc_temp_var (type, unshare_expr (lhs), &gsi);
1565 rhs = ifc_temp_var (type, unshare_expr (rhs), &gsi);
1566 rhs = build3 (COND_EXPR, type, unshare_expr (cond), rhs, lhs);
1567 gimple_assign_set_rhs1 (stmt, ifc_temp_var (type, rhs, &gsi));
1568 update_stmt (stmt);
1569 }
1570 }
1571 }
1572
1573 /* Remove all GIMPLE_CONDs and GIMPLE_LABELs of all the basic blocks
1574 other than the exit and latch of the LOOP. Also resets the
1575 GIMPLE_DEBUG information. */
1576
1577 static void
1578 remove_conditions_and_labels (loop_p loop)
1579 {
1580 gimple_stmt_iterator gsi;
1581 unsigned int i;
1582
1583 for (i = 0; i < loop->num_nodes; i++)
1584 {
1585 basic_block bb = ifc_bbs[i];
1586
1587 if (bb_with_exit_edge_p (loop, bb)
1588 || bb == loop->latch)
1589 continue;
1590
1591 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); )
1592 switch (gimple_code (gsi_stmt (gsi)))
1593 {
1594 case GIMPLE_COND:
1595 case GIMPLE_LABEL:
1596 gsi_remove (&gsi, true);
1597 break;
1598
1599 case GIMPLE_DEBUG:
1600 /* ??? Should there be conditional GIMPLE_DEBUG_BINDs? */
1601 if (gimple_debug_bind_p (gsi_stmt (gsi)))
1602 {
1603 gimple_debug_bind_reset_value (gsi_stmt (gsi));
1604 update_stmt (gsi_stmt (gsi));
1605 }
1606 gsi_next (&gsi);
1607 break;
1608
1609 default:
1610 gsi_next (&gsi);
1611 }
1612 }
1613 }
1614
1615 /* Combine all the basic blocks from LOOP into one or two super basic
1616 blocks. Replace PHI nodes with conditional modify expressions. */
1617
1618 static void
1619 combine_blocks (struct loop *loop)
1620 {
1621 basic_block bb, exit_bb, merge_target_bb;
1622 unsigned int orig_loop_num_nodes = loop->num_nodes;
1623 unsigned int i;
1624 edge e;
1625 edge_iterator ei;
1626
1627 remove_conditions_and_labels (loop);
1628 insert_gimplified_predicates (loop);
1629 predicate_all_scalar_phis (loop);
1630
1631 if (flag_tree_loop_if_convert_stores)
1632 predicate_mem_writes (loop);
1633
1634 /* Merge basic blocks: first remove all the edges in the loop,
1635 except for those from the exit block. */
1636 exit_bb = NULL;
1637 for (i = 0; i < orig_loop_num_nodes; i++)
1638 {
1639 bb = ifc_bbs[i];
1640 if (bb_with_exit_edge_p (loop, bb))
1641 {
1642 exit_bb = bb;
1643 break;
1644 }
1645 }
1646 gcc_assert (exit_bb != loop->latch);
1647
1648 for (i = 1; i < orig_loop_num_nodes; i++)
1649 {
1650 bb = ifc_bbs[i];
1651
1652 for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei));)
1653 {
1654 if (e->src == exit_bb)
1655 ei_next (&ei);
1656 else
1657 remove_edge (e);
1658 }
1659 }
1660
1661 if (exit_bb != NULL)
1662 {
1663 if (exit_bb != loop->header)
1664 {
1665 /* Connect this node to loop header. */
1666 make_edge (loop->header, exit_bb, EDGE_FALLTHRU);
1667 set_immediate_dominator (CDI_DOMINATORS, exit_bb, loop->header);
1668 }
1669
1670 /* Redirect non-exit edges to loop->latch. */
1671 FOR_EACH_EDGE (e, ei, exit_bb->succs)
1672 {
1673 if (!loop_exit_edge_p (loop, e))
1674 redirect_edge_and_branch (e, loop->latch);
1675 }
1676 set_immediate_dominator (CDI_DOMINATORS, loop->latch, exit_bb);
1677 }
1678 else
1679 {
1680 /* If the loop does not have an exit, reconnect header and latch. */
1681 make_edge (loop->header, loop->latch, EDGE_FALLTHRU);
1682 set_immediate_dominator (CDI_DOMINATORS, loop->latch, loop->header);
1683 }
1684
1685 merge_target_bb = loop->header;
1686 for (i = 1; i < orig_loop_num_nodes; i++)
1687 {
1688 gimple_stmt_iterator gsi;
1689 gimple_stmt_iterator last;
1690
1691 bb = ifc_bbs[i];
1692
1693 if (bb == exit_bb || bb == loop->latch)
1694 continue;
1695
1696 /* Make stmts member of loop->header. */
1697 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1698 gimple_set_bb (gsi_stmt (gsi), merge_target_bb);
1699
1700 /* Update stmt list. */
1701 last = gsi_last_bb (merge_target_bb);
1702 gsi_insert_seq_after (&last, bb_seq (bb), GSI_NEW_STMT);
1703 set_bb_seq (bb, NULL);
1704
1705 delete_basic_block (bb);
1706 }
1707
1708 /* If possible, merge loop header to the block with the exit edge.
1709 This reduces the number of basic blocks to two, to please the
1710 vectorizer that handles only loops with two nodes. */
1711 if (exit_bb
1712 && exit_bb != loop->header
1713 && can_merge_blocks_p (loop->header, exit_bb))
1714 merge_blocks (loop->header, exit_bb);
1715 }
1716
1717 /* If-convert LOOP when it is legal. For the moment this pass has no
1718 profitability analysis. Returns true when something changed. */
1719
1720 static bool
1721 tree_if_conversion (struct loop *loop)
1722 {
1723 bool changed = false;
1724 ifc_bbs = NULL;
1725
1726 if (!if_convertible_loop_p (loop)
1727 || !dbg_cnt (if_conversion_tree))
1728 goto cleanup;
1729
1730 /* Now all statements are if-convertible. Combine all the basic
1731 blocks into one huge basic block doing the if-conversion
1732 on-the-fly. */
1733 combine_blocks (loop);
1734
1735 if (flag_tree_loop_if_convert_stores)
1736 mark_sym_for_renaming (gimple_vop (cfun));
1737
1738 changed = true;
1739
1740 cleanup:
1741 if (ifc_bbs)
1742 {
1743 unsigned int i;
1744
1745 for (i = 0; i < loop->num_nodes; i++)
1746 free_bb_predicate (ifc_bbs[i]);
1747
1748 free (ifc_bbs);
1749 ifc_bbs = NULL;
1750 }
1751
1752 return changed;
1753 }
1754
1755 /* Tree if-conversion pass management. */
1756
1757 static unsigned int
1758 main_tree_if_conversion (void)
1759 {
1760 loop_iterator li;
1761 struct loop *loop;
1762 bool changed = false;
1763 unsigned todo = 0;
1764
1765 if (number_of_loops () <= 1)
1766 return 0;
1767
1768 FOR_EACH_LOOP (li, loop, 0)
1769 changed |= tree_if_conversion (loop);
1770
1771 if (changed)
1772 todo |= TODO_cleanup_cfg;
1773
1774 if (changed && flag_tree_loop_if_convert_stores)
1775 todo |= TODO_update_ssa_only_virtuals;
1776
1777 free_dominance_info (CDI_POST_DOMINATORS);
1778
1779 return todo;
1780 }
1781
1782 /* Returns true when the if-conversion pass is enabled. */
1783
1784 static bool
1785 gate_tree_if_conversion (void)
1786 {
1787 return ((flag_tree_vectorize && flag_tree_loop_if_convert != 0)
1788 || flag_tree_loop_if_convert == 1
1789 || flag_tree_loop_if_convert_stores == 1);
1790 }
1791
1792 struct gimple_opt_pass pass_if_conversion =
1793 {
1794 {
1795 GIMPLE_PASS,
1796 "ifcvt", /* name */
1797 gate_tree_if_conversion, /* gate */
1798 main_tree_if_conversion, /* execute */
1799 NULL, /* sub */
1800 NULL, /* next */
1801 0, /* static_pass_number */
1802 TV_NONE, /* tv_id */
1803 PROP_cfg | PROP_ssa, /* properties_required */
1804 0, /* properties_provided */
1805 0, /* properties_destroyed */
1806 0, /* todo_flags_start */
1807 TODO_dump_func | TODO_verify_stmts | TODO_verify_flow
1808 /* todo_flags_finish */
1809 }
1810 };