PR target/56124
[official-gcc.git] / gcc / tree-if-conv.c
blob95d0394fc51d4cfd582dec119c97e39ae5b59bd1
1 /* If-conversion for vectorizer.
2 Copyright (C) 2004-2013 Free Software Foundation, Inc.
3 Contributed by Devang Patel <dpatel@apple.com>
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* This pass implements a tree level if-conversion of loops. Its
22 initial goal is to help the vectorizer to vectorize loops with
23 conditions.
25 A short description of if-conversion:
27 o Decide if a loop is if-convertible or not.
28 o Walk all loop basic blocks in breadth first order (BFS order).
29 o Remove conditional statements (at the end of basic block)
30 and propagate condition into destination basic blocks'
31 predicate list.
32 o Replace modify expression with conditional modify expression
33 using current basic block's condition.
34 o Merge all basic blocks
35 o Replace phi nodes with conditional modify expr
36 o Merge all basic blocks into header
38 Sample transformation:
40 INPUT
41 -----
43 # i_23 = PHI <0(0), i_18(10)>;
44 <L0>:;
45 j_15 = A[i_23];
46 if (j_15 > 41) goto <L1>; else goto <L17>;
48 <L17>:;
49 goto <bb 3> (<L3>);
51 <L1>:;
53 # iftmp.2_4 = PHI <0(8), 42(2)>;
54 <L3>:;
55 A[i_23] = iftmp.2_4;
56 i_18 = i_23 + 1;
57 if (i_18 <= 15) goto <L19>; else goto <L18>;
59 <L19>:;
60 goto <bb 1> (<L0>);
62 <L18>:;
64 OUTPUT
65 ------
67 # i_23 = PHI <0(0), i_18(10)>;
68 <L0>:;
69 j_15 = A[i_23];
71 <L3>:;
72 iftmp.2_4 = j_15 > 41 ? 42 : 0;
73 A[i_23] = iftmp.2_4;
74 i_18 = i_23 + 1;
75 if (i_18 <= 15) goto <L19>; else goto <L18>;
77 <L19>:;
78 goto <bb 1> (<L0>);
80 <L18>:;
83 #include "config.h"
84 #include "system.h"
85 #include "coretypes.h"
86 #include "tm.h"
87 #include "tree.h"
88 #include "flags.h"
89 #include "basic-block.h"
90 #include "gimple-pretty-print.h"
91 #include "tree-flow.h"
92 #include "cfgloop.h"
93 #include "tree-chrec.h"
94 #include "tree-data-ref.h"
95 #include "tree-scalar-evolution.h"
96 #include "tree-pass.h"
97 #include "dbgcnt.h"
99 /* List of basic blocks in if-conversion-suitable order. */
100 static basic_block *ifc_bbs;
102 /* Structure used to predicate basic blocks. This is attached to the
103 ->aux field of the BBs in the loop to be if-converted. */
104 typedef struct bb_predicate_s {
106 /* The condition under which this basic block is executed. */
107 tree predicate;
109 /* PREDICATE is gimplified, and the sequence of statements is
110 recorded here, in order to avoid the duplication of computations
111 that occur in previous conditions. See PR44483. */
112 gimple_seq predicate_gimplified_stmts;
113 } *bb_predicate_p;
115 /* Returns true when the basic block BB has a predicate. */
117 static inline bool
118 bb_has_predicate (basic_block bb)
120 return bb->aux != NULL;
123 /* Returns the gimplified predicate for basic block BB. */
125 static inline tree
126 bb_predicate (basic_block bb)
128 return ((bb_predicate_p) bb->aux)->predicate;
131 /* Sets the gimplified predicate COND for basic block BB. */
133 static inline void
134 set_bb_predicate (basic_block bb, tree cond)
136 gcc_assert ((TREE_CODE (cond) == TRUTH_NOT_EXPR
137 && is_gimple_condexpr (TREE_OPERAND (cond, 0)))
138 || is_gimple_condexpr (cond));
139 ((bb_predicate_p) bb->aux)->predicate = cond;
142 /* Returns the sequence of statements of the gimplification of the
143 predicate for basic block BB. */
145 static inline gimple_seq
146 bb_predicate_gimplified_stmts (basic_block bb)
148 return ((bb_predicate_p) bb->aux)->predicate_gimplified_stmts;
151 /* Sets the sequence of statements STMTS of the gimplification of the
152 predicate for basic block BB. */
154 static inline void
155 set_bb_predicate_gimplified_stmts (basic_block bb, gimple_seq stmts)
157 ((bb_predicate_p) bb->aux)->predicate_gimplified_stmts = stmts;
160 /* Adds the sequence of statements STMTS to the sequence of statements
161 of the predicate for basic block BB. */
163 static inline void
164 add_bb_predicate_gimplified_stmts (basic_block bb, gimple_seq stmts)
166 gimple_seq_add_seq
167 (&(((bb_predicate_p) bb->aux)->predicate_gimplified_stmts), stmts);
170 /* Initializes to TRUE the predicate of basic block BB. */
172 static inline void
173 init_bb_predicate (basic_block bb)
175 bb->aux = XNEW (struct bb_predicate_s);
176 set_bb_predicate_gimplified_stmts (bb, NULL);
177 set_bb_predicate (bb, boolean_true_node);
180 /* Free the predicate of basic block BB. */
182 static inline void
183 free_bb_predicate (basic_block bb)
185 gimple_seq stmts;
187 if (!bb_has_predicate (bb))
188 return;
190 /* Release the SSA_NAMEs created for the gimplification of the
191 predicate. */
192 stmts = bb_predicate_gimplified_stmts (bb);
193 if (stmts)
195 gimple_stmt_iterator i;
197 for (i = gsi_start (stmts); !gsi_end_p (i); gsi_next (&i))
198 free_stmt_operands (gsi_stmt (i));
201 free (bb->aux);
202 bb->aux = NULL;
205 /* Free the predicate of BB and reinitialize it with the true
206 predicate. */
208 static inline void
209 reset_bb_predicate (basic_block bb)
211 free_bb_predicate (bb);
212 init_bb_predicate (bb);
215 /* Returns a new SSA_NAME of type TYPE that is assigned the value of
216 the expression EXPR. Inserts the statement created for this
217 computation before GSI and leaves the iterator GSI at the same
218 statement. */
220 static tree
221 ifc_temp_var (tree type, tree expr, gimple_stmt_iterator *gsi)
223 tree new_name = make_temp_ssa_name (type, NULL, "_ifc_");
224 gimple stmt = gimple_build_assign (new_name, expr);
225 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
226 return new_name;
229 /* Return true when COND is a true predicate. */
231 static inline bool
232 is_true_predicate (tree cond)
234 return (cond == NULL_TREE
235 || cond == boolean_true_node
236 || integer_onep (cond));
239 /* Returns true when BB has a predicate that is not trivial: true or
240 NULL_TREE. */
242 static inline bool
243 is_predicated (basic_block bb)
245 return !is_true_predicate (bb_predicate (bb));
248 /* Parses the predicate COND and returns its comparison code and
249 operands OP0 and OP1. */
251 static enum tree_code
252 parse_predicate (tree cond, tree *op0, tree *op1)
254 gimple s;
256 if (TREE_CODE (cond) == SSA_NAME
257 && is_gimple_assign (s = SSA_NAME_DEF_STMT (cond)))
259 if (TREE_CODE_CLASS (gimple_assign_rhs_code (s)) == tcc_comparison)
261 *op0 = gimple_assign_rhs1 (s);
262 *op1 = gimple_assign_rhs2 (s);
263 return gimple_assign_rhs_code (s);
266 else if (gimple_assign_rhs_code (s) == TRUTH_NOT_EXPR)
268 tree op = gimple_assign_rhs1 (s);
269 tree type = TREE_TYPE (op);
270 enum tree_code code = parse_predicate (op, op0, op1);
272 return code == ERROR_MARK ? ERROR_MARK
273 : invert_tree_comparison (code, HONOR_NANS (TYPE_MODE (type)));
276 return ERROR_MARK;
279 if (TREE_CODE_CLASS (TREE_CODE (cond)) == tcc_comparison)
281 *op0 = TREE_OPERAND (cond, 0);
282 *op1 = TREE_OPERAND (cond, 1);
283 return TREE_CODE (cond);
286 return ERROR_MARK;
289 /* Returns the fold of predicate C1 OR C2 at location LOC. */
291 static tree
292 fold_or_predicates (location_t loc, tree c1, tree c2)
294 tree op1a, op1b, op2a, op2b;
295 enum tree_code code1 = parse_predicate (c1, &op1a, &op1b);
296 enum tree_code code2 = parse_predicate (c2, &op2a, &op2b);
298 if (code1 != ERROR_MARK && code2 != ERROR_MARK)
300 tree t = maybe_fold_or_comparisons (code1, op1a, op1b,
301 code2, op2a, op2b);
302 if (t)
303 return t;
306 return fold_build2_loc (loc, TRUTH_OR_EXPR, boolean_type_node, c1, c2);
309 /* Returns true if N is either a constant or a SSA_NAME. */
311 static bool
312 constant_or_ssa_name (tree n)
314 switch (TREE_CODE (n))
316 case SSA_NAME:
317 case INTEGER_CST:
318 case REAL_CST:
319 case COMPLEX_CST:
320 case VECTOR_CST:
321 return true;
322 default:
323 return false;
327 /* Returns either a COND_EXPR or the folded expression if the folded
328 expression is a MIN_EXPR, a MAX_EXPR, an ABS_EXPR,
329 a constant or a SSA_NAME. */
331 static tree
332 fold_build_cond_expr (tree type, tree cond, tree rhs, tree lhs)
334 tree rhs1, lhs1, cond_expr;
335 cond_expr = fold_ternary (COND_EXPR, type, cond,
336 rhs, lhs);
338 if (cond_expr == NULL_TREE)
339 return build3 (COND_EXPR, type, cond, rhs, lhs);
341 STRIP_USELESS_TYPE_CONVERSION (cond_expr);
343 if (constant_or_ssa_name (cond_expr))
344 return cond_expr;
346 if (TREE_CODE (cond_expr) == ABS_EXPR)
348 rhs1 = TREE_OPERAND (cond_expr, 1);
349 STRIP_USELESS_TYPE_CONVERSION (rhs1);
350 if (constant_or_ssa_name (rhs1))
351 return build1 (ABS_EXPR, type, rhs1);
354 if (TREE_CODE (cond_expr) == MIN_EXPR
355 || TREE_CODE (cond_expr) == MAX_EXPR)
357 lhs1 = TREE_OPERAND (cond_expr, 0);
358 STRIP_USELESS_TYPE_CONVERSION (lhs1);
359 rhs1 = TREE_OPERAND (cond_expr, 1);
360 STRIP_USELESS_TYPE_CONVERSION (rhs1);
361 if (constant_or_ssa_name (rhs1)
362 && constant_or_ssa_name (lhs1))
363 return build2 (TREE_CODE (cond_expr), type, lhs1, rhs1);
365 return build3 (COND_EXPR, type, cond, rhs, lhs);
368 /* Add condition NC to the predicate list of basic block BB. */
370 static inline void
371 add_to_predicate_list (basic_block bb, tree nc)
373 tree bc, *tp;
375 if (is_true_predicate (nc))
376 return;
378 if (!is_predicated (bb))
379 bc = nc;
380 else
382 bc = bb_predicate (bb);
383 bc = fold_or_predicates (EXPR_LOCATION (bc), nc, bc);
384 if (is_true_predicate (bc))
386 reset_bb_predicate (bb);
387 return;
391 /* Allow a TRUTH_NOT_EXPR around the main predicate. */
392 if (TREE_CODE (bc) == TRUTH_NOT_EXPR)
393 tp = &TREE_OPERAND (bc, 0);
394 else
395 tp = &bc;
396 if (!is_gimple_condexpr (*tp))
398 gimple_seq stmts;
399 *tp = force_gimple_operand_1 (*tp, &stmts, is_gimple_condexpr, NULL_TREE);
400 add_bb_predicate_gimplified_stmts (bb, stmts);
402 set_bb_predicate (bb, bc);
405 /* Add the condition COND to the previous condition PREV_COND, and add
406 this to the predicate list of the destination of edge E. LOOP is
407 the loop to be if-converted. */
409 static void
410 add_to_dst_predicate_list (struct loop *loop, edge e,
411 tree prev_cond, tree cond)
413 if (!flow_bb_inside_loop_p (loop, e->dest))
414 return;
416 if (!is_true_predicate (prev_cond))
417 cond = fold_build2 (TRUTH_AND_EXPR, boolean_type_node,
418 prev_cond, cond);
420 add_to_predicate_list (e->dest, cond);
423 /* Return true if one of the successor edges of BB exits LOOP. */
425 static bool
426 bb_with_exit_edge_p (struct loop *loop, basic_block bb)
428 edge e;
429 edge_iterator ei;
431 FOR_EACH_EDGE (e, ei, bb->succs)
432 if (loop_exit_edge_p (loop, e))
433 return true;
435 return false;
438 /* Return true when PHI is if-convertible. PHI is part of loop LOOP
439 and it belongs to basic block BB.
441 PHI is not if-convertible if:
442 - it has more than 2 arguments.
444 When the flag_tree_loop_if_convert_stores is not set, PHI is not
445 if-convertible if:
446 - a virtual PHI is immediately used in another PHI node,
447 - there is a virtual PHI in a BB other than the loop->header. */
449 static bool
450 if_convertible_phi_p (struct loop *loop, basic_block bb, gimple phi)
452 if (dump_file && (dump_flags & TDF_DETAILS))
454 fprintf (dump_file, "-------------------------\n");
455 print_gimple_stmt (dump_file, phi, 0, TDF_SLIM);
458 if (bb != loop->header && gimple_phi_num_args (phi) != 2)
460 if (dump_file && (dump_flags & TDF_DETAILS))
461 fprintf (dump_file, "More than two phi node args.\n");
462 return false;
465 if (flag_tree_loop_if_convert_stores)
466 return true;
468 /* When the flag_tree_loop_if_convert_stores is not set, check
469 that there are no memory writes in the branches of the loop to be
470 if-converted. */
471 if (virtual_operand_p (gimple_phi_result (phi)))
473 imm_use_iterator imm_iter;
474 use_operand_p use_p;
476 if (bb != loop->header)
478 if (dump_file && (dump_flags & TDF_DETAILS))
479 fprintf (dump_file, "Virtual phi not on loop->header.\n");
480 return false;
483 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_phi_result (phi))
485 if (gimple_code (USE_STMT (use_p)) == GIMPLE_PHI)
487 if (dump_file && (dump_flags & TDF_DETAILS))
488 fprintf (dump_file, "Difficult to handle this virtual phi.\n");
489 return false;
494 return true;
497 /* Records the status of a data reference. This struct is attached to
498 each DR->aux field. */
500 struct ifc_dr {
501 /* -1 when not initialized, 0 when false, 1 when true. */
502 int written_at_least_once;
504 /* -1 when not initialized, 0 when false, 1 when true. */
505 int rw_unconditionally;
508 #define IFC_DR(DR) ((struct ifc_dr *) (DR)->aux)
509 #define DR_WRITTEN_AT_LEAST_ONCE(DR) (IFC_DR (DR)->written_at_least_once)
510 #define DR_RW_UNCONDITIONALLY(DR) (IFC_DR (DR)->rw_unconditionally)
512 /* Returns true when the memory references of STMT are read or written
513 unconditionally. In other words, this function returns true when
514 for every data reference A in STMT there exist other accesses to
515 a data reference with the same base with predicates that add up (OR-up) to
516 the true predicate: this ensures that the data reference A is touched
517 (read or written) on every iteration of the if-converted loop. */
519 static bool
520 memrefs_read_or_written_unconditionally (gimple stmt,
521 vec<data_reference_p> drs)
523 int i, j;
524 data_reference_p a, b;
525 tree ca = bb_predicate (gimple_bb (stmt));
527 for (i = 0; drs.iterate (i, &a); i++)
528 if (DR_STMT (a) == stmt)
530 bool found = false;
531 int x = DR_RW_UNCONDITIONALLY (a);
533 if (x == 0)
534 return false;
536 if (x == 1)
537 continue;
539 for (j = 0; drs.iterate (j, &b); j++)
541 tree ref_base_a = DR_REF (a);
542 tree ref_base_b = DR_REF (b);
544 if (DR_STMT (b) == stmt)
545 continue;
547 while (TREE_CODE (ref_base_a) == COMPONENT_REF
548 || TREE_CODE (ref_base_a) == IMAGPART_EXPR
549 || TREE_CODE (ref_base_a) == REALPART_EXPR)
550 ref_base_a = TREE_OPERAND (ref_base_a, 0);
552 while (TREE_CODE (ref_base_b) == COMPONENT_REF
553 || TREE_CODE (ref_base_b) == IMAGPART_EXPR
554 || TREE_CODE (ref_base_b) == REALPART_EXPR)
555 ref_base_b = TREE_OPERAND (ref_base_b, 0);
557 if (!operand_equal_p (ref_base_a, ref_base_b, 0))
559 tree cb = bb_predicate (gimple_bb (DR_STMT (b)));
561 if (DR_RW_UNCONDITIONALLY (b) == 1
562 || is_true_predicate (cb)
563 || is_true_predicate (ca
564 = fold_or_predicates (EXPR_LOCATION (cb), ca, cb)))
566 DR_RW_UNCONDITIONALLY (a) = 1;
567 DR_RW_UNCONDITIONALLY (b) = 1;
568 found = true;
569 break;
574 if (!found)
576 DR_RW_UNCONDITIONALLY (a) = 0;
577 return false;
581 return true;
584 /* Returns true when the memory references of STMT are unconditionally
585 written. In other words, this function returns true when for every
586 data reference A written in STMT, there exist other writes to the
587 same data reference with predicates that add up (OR-up) to the true
588 predicate: this ensures that the data reference A is written on
589 every iteration of the if-converted loop. */
591 static bool
592 write_memrefs_written_at_least_once (gimple stmt,
593 vec<data_reference_p> drs)
595 int i, j;
596 data_reference_p a, b;
597 tree ca = bb_predicate (gimple_bb (stmt));
599 for (i = 0; drs.iterate (i, &a); i++)
600 if (DR_STMT (a) == stmt
601 && DR_IS_WRITE (a))
603 bool found = false;
604 int x = DR_WRITTEN_AT_LEAST_ONCE (a);
606 if (x == 0)
607 return false;
609 if (x == 1)
610 continue;
612 for (j = 0; drs.iterate (j, &b); j++)
613 if (DR_STMT (b) != stmt
614 && DR_IS_WRITE (b)
615 && same_data_refs_base_objects (a, b))
617 tree cb = bb_predicate (gimple_bb (DR_STMT (b)));
619 if (DR_WRITTEN_AT_LEAST_ONCE (b) == 1
620 || is_true_predicate (cb)
621 || is_true_predicate (ca = fold_or_predicates (EXPR_LOCATION (cb),
622 ca, cb)))
624 DR_WRITTEN_AT_LEAST_ONCE (a) = 1;
625 DR_WRITTEN_AT_LEAST_ONCE (b) = 1;
626 found = true;
627 break;
631 if (!found)
633 DR_WRITTEN_AT_LEAST_ONCE (a) = 0;
634 return false;
638 return true;
641 /* Return true when the memory references of STMT won't trap in the
642 if-converted code. There are two things that we have to check for:
644 - writes to memory occur to writable memory: if-conversion of
645 memory writes transforms the conditional memory writes into
646 unconditional writes, i.e. "if (cond) A[i] = foo" is transformed
647 into "A[i] = cond ? foo : A[i]", and as the write to memory may not
648 be executed at all in the original code, it may be a readonly
649 memory. To check that A is not const-qualified, we check that
650 there exists at least an unconditional write to A in the current
651 function.
653 - reads or writes to memory are valid memory accesses for every
654 iteration. To check that the memory accesses are correctly formed
655 and that we are allowed to read and write in these locations, we
656 check that the memory accesses to be if-converted occur at every
657 iteration unconditionally. */
659 static bool
660 ifcvt_memrefs_wont_trap (gimple stmt, vec<data_reference_p> refs)
662 return write_memrefs_written_at_least_once (stmt, refs)
663 && memrefs_read_or_written_unconditionally (stmt, refs);
666 /* Wrapper around gimple_could_trap_p refined for the needs of the
667 if-conversion. Try to prove that the memory accesses of STMT could
668 not trap in the innermost loop containing STMT. */
670 static bool
671 ifcvt_could_trap_p (gimple stmt, vec<data_reference_p> refs)
673 if (gimple_vuse (stmt)
674 && !gimple_could_trap_p_1 (stmt, false, false)
675 && ifcvt_memrefs_wont_trap (stmt, refs))
676 return false;
678 return gimple_could_trap_p (stmt);
681 /* Return true when STMT is if-convertible.
683 GIMPLE_ASSIGN statement is not if-convertible if,
684 - it is not movable,
685 - it could trap,
686 - LHS is not var decl. */
688 static bool
689 if_convertible_gimple_assign_stmt_p (gimple stmt,
690 vec<data_reference_p> refs)
692 tree lhs = gimple_assign_lhs (stmt);
693 basic_block bb;
695 if (dump_file && (dump_flags & TDF_DETAILS))
697 fprintf (dump_file, "-------------------------\n");
698 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
701 if (!is_gimple_reg_type (TREE_TYPE (lhs)))
702 return false;
704 /* Some of these constrains might be too conservative. */
705 if (stmt_ends_bb_p (stmt)
706 || gimple_has_volatile_ops (stmt)
707 || (TREE_CODE (lhs) == SSA_NAME
708 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs))
709 || gimple_has_side_effects (stmt))
711 if (dump_file && (dump_flags & TDF_DETAILS))
712 fprintf (dump_file, "stmt not suitable for ifcvt\n");
713 return false;
716 if (flag_tree_loop_if_convert_stores)
718 if (ifcvt_could_trap_p (stmt, refs))
720 if (dump_file && (dump_flags & TDF_DETAILS))
721 fprintf (dump_file, "tree could trap...\n");
722 return false;
724 return true;
727 if (gimple_assign_rhs_could_trap_p (stmt))
729 if (dump_file && (dump_flags & TDF_DETAILS))
730 fprintf (dump_file, "tree could trap...\n");
731 return false;
734 bb = gimple_bb (stmt);
736 if (TREE_CODE (lhs) != SSA_NAME
737 && bb != bb->loop_father->header
738 && !bb_with_exit_edge_p (bb->loop_father, bb))
740 if (dump_file && (dump_flags & TDF_DETAILS))
742 fprintf (dump_file, "LHS is not var\n");
743 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
745 return false;
748 return true;
751 /* Return true when STMT is if-convertible.
753 A statement is if-convertible if:
754 - it is an if-convertible GIMPLE_ASSIGN,
755 - it is a GIMPLE_LABEL or a GIMPLE_COND. */
757 static bool
758 if_convertible_stmt_p (gimple stmt, vec<data_reference_p> refs)
760 switch (gimple_code (stmt))
762 case GIMPLE_LABEL:
763 case GIMPLE_DEBUG:
764 case GIMPLE_COND:
765 return true;
767 case GIMPLE_ASSIGN:
768 return if_convertible_gimple_assign_stmt_p (stmt, refs);
770 case GIMPLE_CALL:
772 tree fndecl = gimple_call_fndecl (stmt);
773 if (fndecl)
775 int flags = gimple_call_flags (stmt);
776 if ((flags & ECF_CONST)
777 && !(flags & ECF_LOOPING_CONST_OR_PURE)
778 /* We can only vectorize some builtins at the moment,
779 so restrict if-conversion to those. */
780 && DECL_BUILT_IN (fndecl))
781 return true;
783 return false;
786 default:
787 /* Don't know what to do with 'em so don't do anything. */
788 if (dump_file && (dump_flags & TDF_DETAILS))
790 fprintf (dump_file, "don't know what to do\n");
791 print_gimple_stmt (dump_file, stmt, 0, TDF_SLIM);
793 return false;
794 break;
797 return true;
800 /* Return true when BB post-dominates all its predecessors. */
802 static bool
803 bb_postdominates_preds (basic_block bb)
805 unsigned i;
807 for (i = 0; i < EDGE_COUNT (bb->preds); i++)
808 if (!dominated_by_p (CDI_POST_DOMINATORS, EDGE_PRED (bb, i)->src, bb))
809 return false;
811 return true;
814 /* Return true when BB is if-convertible. This routine does not check
815 basic block's statements and phis.
817 A basic block is not if-convertible if:
818 - it is non-empty and it is after the exit block (in BFS order),
819 - it is after the exit block but before the latch,
820 - its edges are not normal.
822 EXIT_BB is the basic block containing the exit of the LOOP. BB is
823 inside LOOP. */
825 static bool
826 if_convertible_bb_p (struct loop *loop, basic_block bb, basic_block exit_bb)
828 edge e;
829 edge_iterator ei;
831 if (dump_file && (dump_flags & TDF_DETAILS))
832 fprintf (dump_file, "----------[%d]-------------\n", bb->index);
834 if (EDGE_COUNT (bb->preds) > 2
835 || EDGE_COUNT (bb->succs) > 2)
836 return false;
838 if (exit_bb)
840 if (bb != loop->latch)
842 if (dump_file && (dump_flags & TDF_DETAILS))
843 fprintf (dump_file, "basic block after exit bb but before latch\n");
844 return false;
846 else if (!empty_block_p (bb))
848 if (dump_file && (dump_flags & TDF_DETAILS))
849 fprintf (dump_file, "non empty basic block after exit bb\n");
850 return false;
852 else if (bb == loop->latch
853 && bb != exit_bb
854 && !dominated_by_p (CDI_DOMINATORS, bb, exit_bb))
856 if (dump_file && (dump_flags & TDF_DETAILS))
857 fprintf (dump_file, "latch is not dominated by exit_block\n");
858 return false;
862 /* Be less adventurous and handle only normal edges. */
863 FOR_EACH_EDGE (e, ei, bb->succs)
864 if (e->flags & (EDGE_EH | EDGE_ABNORMAL | EDGE_IRREDUCIBLE_LOOP))
866 if (dump_file && (dump_flags & TDF_DETAILS))
867 fprintf (dump_file, "Difficult to handle edges\n");
868 return false;
871 if (EDGE_COUNT (bb->preds) == 2
872 && bb != loop->header
873 && !bb_postdominates_preds (bb))
874 return false;
876 return true;
879 /* Return true when all predecessor blocks of BB are visited. The
880 VISITED bitmap keeps track of the visited blocks. */
882 static bool
883 pred_blocks_visited_p (basic_block bb, bitmap *visited)
885 edge e;
886 edge_iterator ei;
887 FOR_EACH_EDGE (e, ei, bb->preds)
888 if (!bitmap_bit_p (*visited, e->src->index))
889 return false;
891 return true;
894 /* Get body of a LOOP in suitable order for if-conversion. It is
895 caller's responsibility to deallocate basic block list.
896 If-conversion suitable order is, breadth first sort (BFS) order
897 with an additional constraint: select a block only if all its
898 predecessors are already selected. */
900 static basic_block *
901 get_loop_body_in_if_conv_order (const struct loop *loop)
903 basic_block *blocks, *blocks_in_bfs_order;
904 basic_block bb;
905 bitmap visited;
906 unsigned int index = 0;
907 unsigned int visited_count = 0;
909 gcc_assert (loop->num_nodes);
910 gcc_assert (loop->latch != EXIT_BLOCK_PTR);
912 blocks = XCNEWVEC (basic_block, loop->num_nodes);
913 visited = BITMAP_ALLOC (NULL);
915 blocks_in_bfs_order = get_loop_body_in_bfs_order (loop);
917 index = 0;
918 while (index < loop->num_nodes)
920 bb = blocks_in_bfs_order [index];
922 if (bb->flags & BB_IRREDUCIBLE_LOOP)
924 free (blocks_in_bfs_order);
925 BITMAP_FREE (visited);
926 free (blocks);
927 return NULL;
930 if (!bitmap_bit_p (visited, bb->index))
932 if (pred_blocks_visited_p (bb, &visited)
933 || bb == loop->header)
935 /* This block is now visited. */
936 bitmap_set_bit (visited, bb->index);
937 blocks[visited_count++] = bb;
941 index++;
943 if (index == loop->num_nodes
944 && visited_count != loop->num_nodes)
945 /* Not done yet. */
946 index = 0;
948 free (blocks_in_bfs_order);
949 BITMAP_FREE (visited);
950 return blocks;
953 /* Returns true when the analysis of the predicates for all the basic
954 blocks in LOOP succeeded.
956 predicate_bbs first allocates the predicates of the basic blocks.
957 These fields are then initialized with the tree expressions
958 representing the predicates under which a basic block is executed
959 in the LOOP. As the loop->header is executed at each iteration, it
960 has the "true" predicate. Other statements executed under a
961 condition are predicated with that condition, for example
963 | if (x)
964 | S1;
965 | else
966 | S2;
968 S1 will be predicated with "x", and
969 S2 will be predicated with "!x". */
971 static bool
972 predicate_bbs (loop_p loop)
974 unsigned int i;
976 for (i = 0; i < loop->num_nodes; i++)
977 init_bb_predicate (ifc_bbs[i]);
979 for (i = 0; i < loop->num_nodes; i++)
981 basic_block bb = ifc_bbs[i];
982 tree cond;
983 gimple_stmt_iterator itr;
985 /* The loop latch is always executed and has no extra conditions
986 to be processed: skip it. */
987 if (bb == loop->latch)
989 reset_bb_predicate (loop->latch);
990 continue;
993 cond = bb_predicate (bb);
995 for (itr = gsi_start_bb (bb); !gsi_end_p (itr); gsi_next (&itr))
997 gimple stmt = gsi_stmt (itr);
999 switch (gimple_code (stmt))
1001 case GIMPLE_LABEL:
1002 case GIMPLE_ASSIGN:
1003 case GIMPLE_CALL:
1004 case GIMPLE_DEBUG:
1005 break;
1007 case GIMPLE_COND:
1009 tree c2;
1010 edge true_edge, false_edge;
1011 location_t loc = gimple_location (stmt);
1012 tree c = fold_build2_loc (loc, gimple_cond_code (stmt),
1013 boolean_type_node,
1014 gimple_cond_lhs (stmt),
1015 gimple_cond_rhs (stmt));
1017 /* Add new condition into destination's predicate list. */
1018 extract_true_false_edges_from_block (gimple_bb (stmt),
1019 &true_edge, &false_edge);
1021 /* If C is true, then TRUE_EDGE is taken. */
1022 add_to_dst_predicate_list (loop, true_edge,
1023 unshare_expr (cond),
1024 unshare_expr (c));
1026 /* If C is false, then FALSE_EDGE is taken. */
1027 c2 = build1_loc (loc, TRUTH_NOT_EXPR,
1028 boolean_type_node, unshare_expr (c));
1029 add_to_dst_predicate_list (loop, false_edge,
1030 unshare_expr (cond), c2);
1032 cond = NULL_TREE;
1033 break;
1036 default:
1037 /* Not handled yet in if-conversion. */
1038 return false;
1042 /* If current bb has only one successor, then consider it as an
1043 unconditional goto. */
1044 if (single_succ_p (bb))
1046 basic_block bb_n = single_succ (bb);
1048 /* The successor bb inherits the predicate of its
1049 predecessor. If there is no predicate in the predecessor
1050 bb, then consider the successor bb as always executed. */
1051 if (cond == NULL_TREE)
1052 cond = boolean_true_node;
1054 add_to_predicate_list (bb_n, cond);
1058 /* The loop header is always executed. */
1059 reset_bb_predicate (loop->header);
1060 gcc_assert (bb_predicate_gimplified_stmts (loop->header) == NULL
1061 && bb_predicate_gimplified_stmts (loop->latch) == NULL);
1063 return true;
1066 /* Return true when LOOP is if-convertible. This is a helper function
1067 for if_convertible_loop_p. REFS and DDRS are initialized and freed
1068 in if_convertible_loop_p. */
1070 static bool
1071 if_convertible_loop_p_1 (struct loop *loop,
1072 vec<loop_p> *loop_nest,
1073 vec<data_reference_p> *refs,
1074 vec<ddr_p> *ddrs)
1076 bool res;
1077 unsigned int i;
1078 basic_block exit_bb = NULL;
1080 /* Don't if-convert the loop when the data dependences cannot be
1081 computed: the loop won't be vectorized in that case. */
1082 res = compute_data_dependences_for_loop (loop, true, loop_nest, refs, ddrs);
1083 if (!res)
1084 return false;
1086 calculate_dominance_info (CDI_DOMINATORS);
1087 calculate_dominance_info (CDI_POST_DOMINATORS);
1089 /* Allow statements that can be handled during if-conversion. */
1090 ifc_bbs = get_loop_body_in_if_conv_order (loop);
1091 if (!ifc_bbs)
1093 if (dump_file && (dump_flags & TDF_DETAILS))
1094 fprintf (dump_file, "Irreducible loop\n");
1095 return false;
1098 for (i = 0; i < loop->num_nodes; i++)
1100 basic_block bb = ifc_bbs[i];
1102 if (!if_convertible_bb_p (loop, bb, exit_bb))
1103 return false;
1105 if (bb_with_exit_edge_p (loop, bb))
1106 exit_bb = bb;
1109 res = predicate_bbs (loop);
1110 if (!res)
1111 return false;
1113 if (flag_tree_loop_if_convert_stores)
1115 data_reference_p dr;
1117 for (i = 0; refs->iterate (i, &dr); i++)
1119 dr->aux = XNEW (struct ifc_dr);
1120 DR_WRITTEN_AT_LEAST_ONCE (dr) = -1;
1121 DR_RW_UNCONDITIONALLY (dr) = -1;
1125 for (i = 0; i < loop->num_nodes; i++)
1127 basic_block bb = ifc_bbs[i];
1128 gimple_stmt_iterator itr;
1130 for (itr = gsi_start_phis (bb); !gsi_end_p (itr); gsi_next (&itr))
1131 if (!if_convertible_phi_p (loop, bb, gsi_stmt (itr)))
1132 return false;
1134 /* Check the if-convertibility of statements in predicated BBs. */
1135 if (is_predicated (bb))
1136 for (itr = gsi_start_bb (bb); !gsi_end_p (itr); gsi_next (&itr))
1137 if (!if_convertible_stmt_p (gsi_stmt (itr), *refs))
1138 return false;
1141 if (dump_file)
1142 fprintf (dump_file, "Applying if-conversion\n");
1144 return true;
1147 /* Return true when LOOP is if-convertible.
1148 LOOP is if-convertible if:
1149 - it is innermost,
1150 - it has two or more basic blocks,
1151 - it has only one exit,
1152 - loop header is not the exit edge,
1153 - if its basic blocks and phi nodes are if convertible. */
1155 static bool
1156 if_convertible_loop_p (struct loop *loop)
1158 edge e;
1159 edge_iterator ei;
1160 bool res = false;
1161 vec<data_reference_p> refs;
1162 vec<ddr_p> ddrs;
1163 vec<loop_p> loop_nest;
1165 /* Handle only innermost loop. */
1166 if (!loop || loop->inner)
1168 if (dump_file && (dump_flags & TDF_DETAILS))
1169 fprintf (dump_file, "not innermost loop\n");
1170 return false;
1173 /* If only one block, no need for if-conversion. */
1174 if (loop->num_nodes <= 2)
1176 if (dump_file && (dump_flags & TDF_DETAILS))
1177 fprintf (dump_file, "less than 2 basic blocks\n");
1178 return false;
1181 /* More than one loop exit is too much to handle. */
1182 if (!single_exit (loop))
1184 if (dump_file && (dump_flags & TDF_DETAILS))
1185 fprintf (dump_file, "multiple exits\n");
1186 return false;
1189 /* If one of the loop header's edge is an exit edge then do not
1190 apply if-conversion. */
1191 FOR_EACH_EDGE (e, ei, loop->header->succs)
1192 if (loop_exit_edge_p (loop, e))
1193 return false;
1195 refs.create (5);
1196 ddrs.create (25);
1197 loop_nest.create (3);
1198 res = if_convertible_loop_p_1 (loop, &loop_nest, &refs, &ddrs);
1200 if (flag_tree_loop_if_convert_stores)
1202 data_reference_p dr;
1203 unsigned int i;
1205 for (i = 0; refs.iterate (i, &dr); i++)
1206 free (dr->aux);
1209 loop_nest.release ();
1210 free_data_refs (refs);
1211 free_dependence_relations (ddrs);
1212 return res;
1215 /* Basic block BB has two predecessors. Using predecessor's bb
1216 predicate, set an appropriate condition COND for the PHI node
1217 replacement. Return the true block whose phi arguments are
1218 selected when cond is true. LOOP is the loop containing the
1219 if-converted region, GSI is the place to insert the code for the
1220 if-conversion. */
1222 static basic_block
1223 find_phi_replacement_condition (struct loop *loop,
1224 basic_block bb, tree *cond,
1225 gimple_stmt_iterator *gsi)
1227 edge first_edge, second_edge;
1228 tree tmp_cond;
1230 gcc_assert (EDGE_COUNT (bb->preds) == 2);
1231 first_edge = EDGE_PRED (bb, 0);
1232 second_edge = EDGE_PRED (bb, 1);
1234 /* Use condition based on following criteria:
1236 S1: x = !c ? a : b;
1238 S2: x = c ? b : a;
1240 S2 is preferred over S1. Make 'b' first_bb and use its condition.
1242 2) Do not make loop header first_bb.
1245 S1: x = !(c == d)? a : b;
1247 S21: t1 = c == d;
1248 S22: x = t1 ? b : a;
1250 S3: x = (c == d) ? b : a;
1252 S3 is preferred over S1 and S2*, Make 'b' first_bb and use
1253 its condition.
1255 4) If pred B is dominated by pred A then use pred B's condition.
1256 See PR23115. */
1258 /* Select condition that is not TRUTH_NOT_EXPR. */
1259 tmp_cond = bb_predicate (first_edge->src);
1260 gcc_assert (tmp_cond);
1262 if (TREE_CODE (tmp_cond) == TRUTH_NOT_EXPR)
1264 edge tmp_edge;
1266 tmp_edge = first_edge;
1267 first_edge = second_edge;
1268 second_edge = tmp_edge;
1271 /* Check if FIRST_BB is loop header or not and make sure that
1272 FIRST_BB does not dominate SECOND_BB. */
1273 if (first_edge->src == loop->header
1274 || dominated_by_p (CDI_DOMINATORS,
1275 second_edge->src, first_edge->src))
1277 *cond = bb_predicate (second_edge->src);
1279 if (TREE_CODE (*cond) == TRUTH_NOT_EXPR)
1280 *cond = TREE_OPERAND (*cond, 0);
1281 else
1282 /* Select non loop header bb. */
1283 first_edge = second_edge;
1285 else
1286 *cond = bb_predicate (first_edge->src);
1288 /* Gimplify the condition to a valid cond-expr conditonal operand. */
1289 *cond = force_gimple_operand_gsi_1 (gsi, unshare_expr (*cond),
1290 is_gimple_condexpr, NULL_TREE,
1291 true, GSI_SAME_STMT);
1293 return first_edge->src;
1296 /* Replace a scalar PHI node with a COND_EXPR using COND as condition.
1297 This routine does not handle PHI nodes with more than two
1298 arguments.
1300 For example,
1301 S1: A = PHI <x1(1), x2(5)>
1302 is converted into,
1303 S2: A = cond ? x1 : x2;
1305 The generated code is inserted at GSI that points to the top of
1306 basic block's statement list. When COND is true, phi arg from
1307 TRUE_BB is selected. */
1309 static void
1310 predicate_scalar_phi (gimple phi, tree cond,
1311 basic_block true_bb,
1312 gimple_stmt_iterator *gsi)
1314 gimple new_stmt;
1315 basic_block bb;
1316 tree rhs, res, arg, scev;
1318 gcc_assert (gimple_code (phi) == GIMPLE_PHI
1319 && gimple_phi_num_args (phi) == 2);
1321 res = gimple_phi_result (phi);
1322 /* Do not handle virtual phi nodes. */
1323 if (virtual_operand_p (res))
1324 return;
1326 bb = gimple_bb (phi);
1328 if ((arg = degenerate_phi_result (phi))
1329 || ((scev = analyze_scalar_evolution (gimple_bb (phi)->loop_father,
1330 res))
1331 && !chrec_contains_undetermined (scev)
1332 && scev != res
1333 && (arg = gimple_phi_arg_def (phi, 0))))
1334 rhs = arg;
1335 else
1337 tree arg_0, arg_1;
1338 /* Use condition that is not TRUTH_NOT_EXPR in conditional modify expr. */
1339 if (EDGE_PRED (bb, 1)->src == true_bb)
1341 arg_0 = gimple_phi_arg_def (phi, 1);
1342 arg_1 = gimple_phi_arg_def (phi, 0);
1344 else
1346 arg_0 = gimple_phi_arg_def (phi, 0);
1347 arg_1 = gimple_phi_arg_def (phi, 1);
1350 gcc_checking_assert (bb == bb->loop_father->header
1351 || bb_postdominates_preds (bb));
1353 /* Build new RHS using selected condition and arguments. */
1354 rhs = fold_build_cond_expr (TREE_TYPE (res), unshare_expr (cond),
1355 arg_0, arg_1);
1358 new_stmt = gimple_build_assign (res, rhs);
1359 SSA_NAME_DEF_STMT (gimple_phi_result (phi)) = new_stmt;
1360 gsi_insert_before (gsi, new_stmt, GSI_SAME_STMT);
1361 update_stmt (new_stmt);
1363 if (dump_file && (dump_flags & TDF_DETAILS))
1365 fprintf (dump_file, "new phi replacement stmt\n");
1366 print_gimple_stmt (dump_file, new_stmt, 0, TDF_SLIM);
1370 /* Replaces in LOOP all the scalar phi nodes other than those in the
1371 LOOP->header block with conditional modify expressions. */
1373 static void
1374 predicate_all_scalar_phis (struct loop *loop)
1376 basic_block bb;
1377 unsigned int orig_loop_num_nodes = loop->num_nodes;
1378 unsigned int i;
1380 for (i = 1; i < orig_loop_num_nodes; i++)
1382 gimple phi;
1383 tree cond = NULL_TREE;
1384 gimple_stmt_iterator gsi, phi_gsi;
1385 basic_block true_bb = NULL;
1386 bb = ifc_bbs[i];
1388 if (bb == loop->header)
1389 continue;
1391 phi_gsi = gsi_start_phis (bb);
1392 if (gsi_end_p (phi_gsi))
1393 continue;
1395 /* BB has two predecessors. Using predecessor's aux field, set
1396 appropriate condition for the PHI node replacement. */
1397 gsi = gsi_after_labels (bb);
1398 true_bb = find_phi_replacement_condition (loop, bb, &cond, &gsi);
1400 while (!gsi_end_p (phi_gsi))
1402 phi = gsi_stmt (phi_gsi);
1403 predicate_scalar_phi (phi, cond, true_bb, &gsi);
1404 release_phi_node (phi);
1405 gsi_next (&phi_gsi);
1408 set_phi_nodes (bb, NULL);
1412 /* Insert in each basic block of LOOP the statements produced by the
1413 gimplification of the predicates. */
1415 static void
1416 insert_gimplified_predicates (loop_p loop)
1418 unsigned int i;
1420 for (i = 0; i < loop->num_nodes; i++)
1422 basic_block bb = ifc_bbs[i];
1423 gimple_seq stmts;
1425 if (!is_predicated (bb))
1427 /* Do not insert statements for a basic block that is not
1428 predicated. Also make sure that the predicate of the
1429 basic block is set to true. */
1430 reset_bb_predicate (bb);
1431 continue;
1434 stmts = bb_predicate_gimplified_stmts (bb);
1435 if (stmts)
1437 if (flag_tree_loop_if_convert_stores)
1439 /* Insert the predicate of the BB just after the label,
1440 as the if-conversion of memory writes will use this
1441 predicate. */
1442 gimple_stmt_iterator gsi = gsi_after_labels (bb);
1443 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
1445 else
1447 /* Insert the predicate of the BB at the end of the BB
1448 as this would reduce the register pressure: the only
1449 use of this predicate will be in successor BBs. */
1450 gimple_stmt_iterator gsi = gsi_last_bb (bb);
1452 if (gsi_end_p (gsi)
1453 || stmt_ends_bb_p (gsi_stmt (gsi)))
1454 gsi_insert_seq_before (&gsi, stmts, GSI_SAME_STMT);
1455 else
1456 gsi_insert_seq_after (&gsi, stmts, GSI_SAME_STMT);
1459 /* Once the sequence is code generated, set it to NULL. */
1460 set_bb_predicate_gimplified_stmts (bb, NULL);
1465 /* Predicate each write to memory in LOOP.
1467 This function transforms control flow constructs containing memory
1468 writes of the form:
1470 | for (i = 0; i < N; i++)
1471 | if (cond)
1472 | A[i] = expr;
1474 into the following form that does not contain control flow:
1476 | for (i = 0; i < N; i++)
1477 | A[i] = cond ? expr : A[i];
1479 The original CFG looks like this:
1481 | bb_0
1482 | i = 0
1483 | end_bb_0
1485 | bb_1
1486 | if (i < N) goto bb_5 else goto bb_2
1487 | end_bb_1
1489 | bb_2
1490 | cond = some_computation;
1491 | if (cond) goto bb_3 else goto bb_4
1492 | end_bb_2
1494 | bb_3
1495 | A[i] = expr;
1496 | goto bb_4
1497 | end_bb_3
1499 | bb_4
1500 | goto bb_1
1501 | end_bb_4
1503 insert_gimplified_predicates inserts the computation of the COND
1504 expression at the beginning of the destination basic block:
1506 | bb_0
1507 | i = 0
1508 | end_bb_0
1510 | bb_1
1511 | if (i < N) goto bb_5 else goto bb_2
1512 | end_bb_1
1514 | bb_2
1515 | cond = some_computation;
1516 | if (cond) goto bb_3 else goto bb_4
1517 | end_bb_2
1519 | bb_3
1520 | cond = some_computation;
1521 | A[i] = expr;
1522 | goto bb_4
1523 | end_bb_3
1525 | bb_4
1526 | goto bb_1
1527 | end_bb_4
1529 predicate_mem_writes is then predicating the memory write as follows:
1531 | bb_0
1532 | i = 0
1533 | end_bb_0
1535 | bb_1
1536 | if (i < N) goto bb_5 else goto bb_2
1537 | end_bb_1
1539 | bb_2
1540 | if (cond) goto bb_3 else goto bb_4
1541 | end_bb_2
1543 | bb_3
1544 | cond = some_computation;
1545 | A[i] = cond ? expr : A[i];
1546 | goto bb_4
1547 | end_bb_3
1549 | bb_4
1550 | goto bb_1
1551 | end_bb_4
1553 and finally combine_blocks removes the basic block boundaries making
1554 the loop vectorizable:
1556 | bb_0
1557 | i = 0
1558 | if (i < N) goto bb_5 else goto bb_1
1559 | end_bb_0
1561 | bb_1
1562 | cond = some_computation;
1563 | A[i] = cond ? expr : A[i];
1564 | if (i < N) goto bb_5 else goto bb_4
1565 | end_bb_1
1567 | bb_4
1568 | goto bb_1
1569 | end_bb_4
1572 static void
1573 predicate_mem_writes (loop_p loop)
1575 unsigned int i, orig_loop_num_nodes = loop->num_nodes;
1577 for (i = 1; i < orig_loop_num_nodes; i++)
1579 gimple_stmt_iterator gsi;
1580 basic_block bb = ifc_bbs[i];
1581 tree cond = bb_predicate (bb);
1582 bool swap;
1583 gimple stmt;
1585 if (is_true_predicate (cond))
1586 continue;
1588 swap = false;
1589 if (TREE_CODE (cond) == TRUTH_NOT_EXPR)
1591 swap = true;
1592 cond = TREE_OPERAND (cond, 0);
1595 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1596 if ((stmt = gsi_stmt (gsi))
1597 && gimple_assign_single_p (stmt)
1598 && gimple_vdef (stmt))
1600 tree lhs = gimple_assign_lhs (stmt);
1601 tree rhs = gimple_assign_rhs1 (stmt);
1602 tree type = TREE_TYPE (lhs);
1604 lhs = ifc_temp_var (type, unshare_expr (lhs), &gsi);
1605 rhs = ifc_temp_var (type, unshare_expr (rhs), &gsi);
1606 if (swap)
1608 tree tem = lhs;
1609 lhs = rhs;
1610 rhs = tem;
1612 cond = force_gimple_operand_gsi_1 (&gsi, unshare_expr (cond),
1613 is_gimple_condexpr, NULL_TREE,
1614 true, GSI_SAME_STMT);
1615 rhs = fold_build_cond_expr (type, unshare_expr (cond), rhs, lhs);
1616 gimple_assign_set_rhs1 (stmt, ifc_temp_var (type, rhs, &gsi));
1617 update_stmt (stmt);
1622 /* Remove all GIMPLE_CONDs and GIMPLE_LABELs of all the basic blocks
1623 other than the exit and latch of the LOOP. Also resets the
1624 GIMPLE_DEBUG information. */
1626 static void
1627 remove_conditions_and_labels (loop_p loop)
1629 gimple_stmt_iterator gsi;
1630 unsigned int i;
1632 for (i = 0; i < loop->num_nodes; i++)
1634 basic_block bb = ifc_bbs[i];
1636 if (bb_with_exit_edge_p (loop, bb)
1637 || bb == loop->latch)
1638 continue;
1640 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); )
1641 switch (gimple_code (gsi_stmt (gsi)))
1643 case GIMPLE_COND:
1644 case GIMPLE_LABEL:
1645 gsi_remove (&gsi, true);
1646 break;
1648 case GIMPLE_DEBUG:
1649 /* ??? Should there be conditional GIMPLE_DEBUG_BINDs? */
1650 if (gimple_debug_bind_p (gsi_stmt (gsi)))
1652 gimple_debug_bind_reset_value (gsi_stmt (gsi));
1653 update_stmt (gsi_stmt (gsi));
1655 gsi_next (&gsi);
1656 break;
1658 default:
1659 gsi_next (&gsi);
1664 /* Combine all the basic blocks from LOOP into one or two super basic
1665 blocks. Replace PHI nodes with conditional modify expressions. */
1667 static void
1668 combine_blocks (struct loop *loop)
1670 basic_block bb, exit_bb, merge_target_bb;
1671 unsigned int orig_loop_num_nodes = loop->num_nodes;
1672 unsigned int i;
1673 edge e;
1674 edge_iterator ei;
1676 remove_conditions_and_labels (loop);
1677 insert_gimplified_predicates (loop);
1678 predicate_all_scalar_phis (loop);
1680 if (flag_tree_loop_if_convert_stores)
1681 predicate_mem_writes (loop);
1683 /* Merge basic blocks: first remove all the edges in the loop,
1684 except for those from the exit block. */
1685 exit_bb = NULL;
1686 for (i = 0; i < orig_loop_num_nodes; i++)
1688 bb = ifc_bbs[i];
1689 free_bb_predicate (bb);
1690 if (bb_with_exit_edge_p (loop, bb))
1692 gcc_assert (exit_bb == NULL);
1693 exit_bb = bb;
1696 gcc_assert (exit_bb != loop->latch);
1698 for (i = 1; i < orig_loop_num_nodes; i++)
1700 bb = ifc_bbs[i];
1702 for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei));)
1704 if (e->src == exit_bb)
1705 ei_next (&ei);
1706 else
1707 remove_edge (e);
1711 if (exit_bb != NULL)
1713 if (exit_bb != loop->header)
1715 /* Connect this node to loop header. */
1716 make_edge (loop->header, exit_bb, EDGE_FALLTHRU);
1717 set_immediate_dominator (CDI_DOMINATORS, exit_bb, loop->header);
1720 /* Redirect non-exit edges to loop->latch. */
1721 FOR_EACH_EDGE (e, ei, exit_bb->succs)
1723 if (!loop_exit_edge_p (loop, e))
1724 redirect_edge_and_branch (e, loop->latch);
1726 set_immediate_dominator (CDI_DOMINATORS, loop->latch, exit_bb);
1728 else
1730 /* If the loop does not have an exit, reconnect header and latch. */
1731 make_edge (loop->header, loop->latch, EDGE_FALLTHRU);
1732 set_immediate_dominator (CDI_DOMINATORS, loop->latch, loop->header);
1735 merge_target_bb = loop->header;
1736 for (i = 1; i < orig_loop_num_nodes; i++)
1738 gimple_stmt_iterator gsi;
1739 gimple_stmt_iterator last;
1741 bb = ifc_bbs[i];
1743 if (bb == exit_bb || bb == loop->latch)
1744 continue;
1746 /* Make stmts member of loop->header. */
1747 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1748 gimple_set_bb (gsi_stmt (gsi), merge_target_bb);
1750 /* Update stmt list. */
1751 last = gsi_last_bb (merge_target_bb);
1752 gsi_insert_seq_after (&last, bb_seq (bb), GSI_NEW_STMT);
1753 set_bb_seq (bb, NULL);
1755 delete_basic_block (bb);
1758 /* If possible, merge loop header to the block with the exit edge.
1759 This reduces the number of basic blocks to two, to please the
1760 vectorizer that handles only loops with two nodes. */
1761 if (exit_bb
1762 && exit_bb != loop->header
1763 && can_merge_blocks_p (loop->header, exit_bb))
1764 merge_blocks (loop->header, exit_bb);
1766 free (ifc_bbs);
1767 ifc_bbs = NULL;
1769 /* Post-dominators are corrupt now. */
1770 free_dominance_info (CDI_POST_DOMINATORS);
1773 /* If-convert LOOP when it is legal. For the moment this pass has no
1774 profitability analysis. Returns true when something changed. */
1776 static bool
1777 tree_if_conversion (struct loop *loop)
1779 bool changed = false;
1780 ifc_bbs = NULL;
1782 if (!if_convertible_loop_p (loop)
1783 || !dbg_cnt (if_conversion_tree))
1784 goto cleanup;
1786 /* Now all statements are if-convertible. Combine all the basic
1787 blocks into one huge basic block doing the if-conversion
1788 on-the-fly. */
1789 combine_blocks (loop);
1791 if (flag_tree_loop_if_convert_stores)
1792 mark_virtual_operands_for_renaming (cfun);
1794 changed = true;
1796 cleanup:
1797 if (ifc_bbs)
1799 unsigned int i;
1801 for (i = 0; i < loop->num_nodes; i++)
1802 free_bb_predicate (ifc_bbs[i]);
1804 free (ifc_bbs);
1805 ifc_bbs = NULL;
1808 return changed;
1811 /* Tree if-conversion pass management. */
1813 static unsigned int
1814 main_tree_if_conversion (void)
1816 loop_iterator li;
1817 struct loop *loop;
1818 bool changed = false;
1819 unsigned todo = 0;
1821 if (number_of_loops () <= 1)
1822 return 0;
1824 FOR_EACH_LOOP (li, loop, 0)
1825 changed |= tree_if_conversion (loop);
1827 if (changed)
1828 todo |= TODO_cleanup_cfg;
1830 if (changed && flag_tree_loop_if_convert_stores)
1831 todo |= TODO_update_ssa_only_virtuals;
1833 free_dominance_info (CDI_POST_DOMINATORS);
1835 #ifdef ENABLE_CHECKING
1837 basic_block bb;
1838 FOR_EACH_BB (bb)
1839 gcc_assert (!bb->aux);
1841 #endif
1843 return todo;
1846 /* Returns true when the if-conversion pass is enabled. */
1848 static bool
1849 gate_tree_if_conversion (void)
1851 return ((flag_tree_vectorize && flag_tree_loop_if_convert != 0)
1852 || flag_tree_loop_if_convert == 1
1853 || flag_tree_loop_if_convert_stores == 1);
1856 struct gimple_opt_pass pass_if_conversion =
1859 GIMPLE_PASS,
1860 "ifcvt", /* name */
1861 OPTGROUP_NONE, /* optinfo_flags */
1862 gate_tree_if_conversion, /* gate */
1863 main_tree_if_conversion, /* execute */
1864 NULL, /* sub */
1865 NULL, /* next */
1866 0, /* static_pass_number */
1867 TV_NONE, /* tv_id */
1868 PROP_cfg | PROP_ssa, /* properties_required */
1869 0, /* properties_provided */
1870 0, /* properties_destroyed */
1871 0, /* todo_flags_start */
1872 TODO_verify_stmts | TODO_verify_flow
1873 /* todo_flags_finish */