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>
6 This file is part of GCC.
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
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
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/>. */
22 /* This pass implements a tree level if-conversion of loops. Its
23 initial goal is to help the vectorizer to vectorize loops with
26 A short description of if-conversion:
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'
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
39 Sample transformation:
44 # i_23 = PHI <0(0), i_18(10)>;
47 if (j_15 > 41) goto <L1>; else goto <L17>;
54 # iftmp.2_4 = PHI <0(8), 42(2)>;
58 if (i_18 <= 15) goto <L19>; else goto <L18>;
68 # i_23 = PHI <0(0), i_18(10)>;
73 iftmp.2_4 = j_15 > 41 ? 42 : 0;
76 if (i_18 <= 15) goto <L19>; else goto <L18>;
86 #include "coretypes.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"
97 #include "tree-chrec.h"
98 #include "tree-data-ref.h"
99 #include "tree-scalar-evolution.h"
100 #include "tree-pass.h"
103 /* List of basic blocks in if-conversion-suitable order. */
104 static basic_block
*ifc_bbs
;
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
{
110 /* The condition under which this basic block is executed. */
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
;
119 /* Returns true when the basic block BB has a predicate. */
122 bb_has_predicate (basic_block bb
)
124 return bb
->aux
!= NULL
;
127 /* Returns the gimplified predicate for basic block BB. */
130 bb_predicate (basic_block bb
)
132 return ((bb_predicate_p
) bb
->aux
)->predicate
;
135 /* Sets the gimplified predicate COND for basic block BB. */
138 set_bb_predicate (basic_block bb
, tree cond
)
140 ((bb_predicate_p
) bb
->aux
)->predicate
= cond
;
143 /* Returns the sequence of statements of the gimplification of the
144 predicate for basic block BB. */
146 static inline gimple_seq
147 bb_predicate_gimplified_stmts (basic_block bb
)
149 return ((bb_predicate_p
) bb
->aux
)->predicate_gimplified_stmts
;
152 /* Sets the sequence of statements STMTS of the gimplification of the
153 predicate for basic block BB. */
156 set_bb_predicate_gimplified_stmts (basic_block bb
, gimple_seq stmts
)
158 ((bb_predicate_p
) bb
->aux
)->predicate_gimplified_stmts
= stmts
;
161 /* Adds the sequence of statements STMTS to the sequence of statements
162 of the predicate for basic block BB. */
165 add_bb_predicate_gimplified_stmts (basic_block bb
, gimple_seq stmts
)
168 (&(((bb_predicate_p
) bb
->aux
)->predicate_gimplified_stmts
), stmts
);
171 /* Initializes to TRUE the predicate of basic block BB. */
174 init_bb_predicate (basic_block bb
)
176 bb
->aux
= XNEW (struct bb_predicate_s
);
177 set_bb_predicate_gimplified_stmts (bb
, NULL
);
178 set_bb_predicate (bb
, boolean_true_node
);
181 /* Free the predicate of basic block BB. */
184 free_bb_predicate (basic_block bb
)
188 if (!bb_has_predicate (bb
))
191 /* Release the SSA_NAMEs created for the gimplification of the
193 stmts
= bb_predicate_gimplified_stmts (bb
);
196 gimple_stmt_iterator i
;
198 for (i
= gsi_start (stmts
); !gsi_end_p (i
); gsi_next (&i
))
199 free_stmt_operands (gsi_stmt (i
));
206 /* Free the predicate of BB and reinitialize it with the true
210 reset_bb_predicate (basic_block bb
)
212 free_bb_predicate (bb
);
213 init_bb_predicate (bb
);
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
222 ifc_temp_var (tree type
, tree expr
, gimple_stmt_iterator
*gsi
)
224 const char *name
= "_ifc_";
228 /* Create new temporary variable. */
229 var
= create_tmp_var (type
, name
);
230 add_referenced_var (var
);
232 /* Build new statement to assign EXPR to new variable. */
233 stmt
= gimple_build_assign (var
, expr
);
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
;
242 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
243 return gimple_assign_lhs (stmt
);
246 /* Return true when COND is a true predicate. */
249 is_true_predicate (tree cond
)
251 return (cond
== NULL_TREE
252 || cond
== boolean_true_node
253 || integer_onep (cond
));
256 /* Returns true when BB has a predicate that is not trivial: true or
260 is_predicated (basic_block bb
)
262 return !is_true_predicate (bb_predicate (bb
));
265 /* Parses the predicate COND and returns its comparison code and
266 operands OP0 and OP1. */
268 static enum tree_code
269 parse_predicate (tree cond
, tree
*op0
, tree
*op1
)
273 if (TREE_CODE (cond
) == SSA_NAME
274 && is_gimple_assign (s
= SSA_NAME_DEF_STMT (cond
)))
276 if (TREE_CODE_CLASS (gimple_assign_rhs_code (s
)) == tcc_comparison
)
278 *op0
= gimple_assign_rhs1 (s
);
279 *op1
= gimple_assign_rhs2 (s
);
280 return gimple_assign_rhs_code (s
);
283 else if (gimple_assign_rhs_code (s
) == TRUTH_NOT_EXPR
)
285 tree op
= gimple_assign_rhs1 (s
);
286 tree type
= TREE_TYPE (op
);
287 enum tree_code code
= parse_predicate (op
, op0
, op1
);
289 return code
== ERROR_MARK
? ERROR_MARK
290 : invert_tree_comparison (code
, HONOR_NANS (TYPE_MODE (type
)));
296 if (TREE_CODE_CLASS (TREE_CODE (cond
)) == tcc_comparison
)
298 *op0
= TREE_OPERAND (cond
, 0);
299 *op1
= TREE_OPERAND (cond
, 1);
300 return TREE_CODE (cond
);
306 /* Returns the fold of predicate C1 OR C2 at location LOC. */
309 fold_or_predicates (location_t loc
, tree c1
, tree c2
)
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
);
315 if (code1
!= ERROR_MARK
&& code2
!= ERROR_MARK
)
317 tree t
= maybe_fold_or_comparisons (code1
, op1a
, op1b
,
323 return fold_build2_loc (loc
, TRUTH_OR_EXPR
, boolean_type_node
, c1
, c2
);
326 /* Add condition NC to the predicate list of basic block BB. */
329 add_to_predicate_list (basic_block bb
, tree nc
)
333 if (is_true_predicate (nc
))
336 if (!is_predicated (bb
))
340 bc
= bb_predicate (bb
);
341 bc
= fold_or_predicates (EXPR_LOCATION (bc
), nc
, bc
);
344 if (!is_gimple_condexpr (bc
))
347 bc
= force_gimple_operand (bc
, &stmts
, true, NULL_TREE
);
348 add_bb_predicate_gimplified_stmts (bb
, stmts
);
351 if (is_true_predicate (bc
))
352 reset_bb_predicate (bb
);
354 set_bb_predicate (bb
, bc
);
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. */
362 add_to_dst_predicate_list (struct loop
*loop
, edge e
,
363 tree prev_cond
, tree cond
)
365 if (!flow_bb_inside_loop_p (loop
, e
->dest
))
368 if (!is_true_predicate (prev_cond
))
369 cond
= fold_build2 (TRUTH_AND_EXPR
, boolean_type_node
,
372 add_to_predicate_list (e
->dest
, cond
);
375 /* Return true if one of the successor edges of BB exits LOOP. */
378 bb_with_exit_edge_p (struct loop
*loop
, basic_block bb
)
383 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
384 if (loop_exit_edge_p (loop
, e
))
390 /* Return true when PHI is if-convertible. PHI is part of loop LOOP
391 and it belongs to basic block BB.
393 PHI is not if-convertible if:
394 - it has more than 2 arguments.
396 When the flag_tree_loop_if_convert_stores is not set, PHI is not
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. */
402 if_convertible_phi_p (struct loop
*loop
, basic_block bb
, gimple phi
)
404 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
406 fprintf (dump_file
, "-------------------------\n");
407 print_gimple_stmt (dump_file
, phi
, 0, TDF_SLIM
);
410 if (bb
!= loop
->header
&& gimple_phi_num_args (phi
) != 2)
412 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
413 fprintf (dump_file
, "More than two phi node args.\n");
417 if (flag_tree_loop_if_convert_stores
)
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
423 if (!is_gimple_reg (SSA_NAME_VAR (gimple_phi_result (phi
))))
425 imm_use_iterator imm_iter
;
428 if (bb
!= loop
->header
)
430 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
431 fprintf (dump_file
, "Virtual phi not on loop->header.\n");
435 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, gimple_phi_result (phi
))
437 if (gimple_code (USE_STMT (use_p
)) == GIMPLE_PHI
)
439 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
440 fprintf (dump_file
, "Difficult to handle this virtual phi.\n");
449 /* Records the status of a data reference. This struct is attached to
450 each DR->aux field. */
453 /* -1 when not initialized, 0 when false, 1 when true. */
454 int written_at_least_once
;
456 /* -1 when not initialized, 0 when false, 1 when true. */
457 int rw_unconditionally
;
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)
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 the same data reference with predicates that add up (OR-up) to the
468 true predicate: this ensures that the data reference A is touched
469 (read or written) on every iteration of the if-converted loop. */
472 memrefs_read_or_written_unconditionally (gimple stmt
,
473 VEC (data_reference_p
, heap
) *drs
)
476 data_reference_p a
, b
;
477 tree ca
= bb_predicate (gimple_bb (stmt
));
479 for (i
= 0; VEC_iterate (data_reference_p
, drs
, i
, a
); i
++)
480 if (DR_STMT (a
) == stmt
)
483 int x
= DR_RW_UNCONDITIONALLY (a
);
491 for (j
= 0; VEC_iterate (data_reference_p
, drs
, j
, b
); j
++)
492 if (DR_STMT (b
) != stmt
493 && same_data_refs (a
, b
))
495 tree cb
= bb_predicate (gimple_bb (DR_STMT (b
)));
497 if (DR_RW_UNCONDITIONALLY (b
) == 1
498 || is_true_predicate (cb
)
499 || is_true_predicate (ca
= fold_or_predicates (EXPR_LOCATION (cb
),
502 DR_RW_UNCONDITIONALLY (a
) = 1;
503 DR_RW_UNCONDITIONALLY (b
) = 1;
511 DR_RW_UNCONDITIONALLY (a
) = 0;
519 /* Returns true when the memory references of STMT are unconditionally
520 written. In other words, this function returns true when for every
521 data reference A written in STMT, there exist other writes to the
522 same data reference with predicates that add up (OR-up) to the true
523 predicate: this ensures that the data reference A is written on
524 every iteration of the if-converted loop. */
527 write_memrefs_written_at_least_once (gimple stmt
,
528 VEC (data_reference_p
, heap
) *drs
)
531 data_reference_p a
, b
;
532 tree ca
= bb_predicate (gimple_bb (stmt
));
534 for (i
= 0; VEC_iterate (data_reference_p
, drs
, i
, a
); i
++)
535 if (DR_STMT (a
) == stmt
539 int x
= DR_WRITTEN_AT_LEAST_ONCE (a
);
547 for (j
= 0; VEC_iterate (data_reference_p
, drs
, j
, b
); j
++)
548 if (DR_STMT (b
) != stmt
550 && same_data_refs_base_objects (a
, b
))
552 tree cb
= bb_predicate (gimple_bb (DR_STMT (b
)));
554 if (DR_WRITTEN_AT_LEAST_ONCE (b
) == 1
555 || is_true_predicate (cb
)
556 || is_true_predicate (ca
= fold_or_predicates (EXPR_LOCATION (cb
),
559 DR_WRITTEN_AT_LEAST_ONCE (a
) = 1;
560 DR_WRITTEN_AT_LEAST_ONCE (b
) = 1;
568 DR_WRITTEN_AT_LEAST_ONCE (a
) = 0;
576 /* Return true when the memory references of STMT won't trap in the
577 if-converted code. There are two things that we have to check for:
579 - writes to memory occur to writable memory: if-conversion of
580 memory writes transforms the conditional memory writes into
581 unconditional writes, i.e. "if (cond) A[i] = foo" is transformed
582 into "A[i] = cond ? foo : A[i]", and as the write to memory may not
583 be executed at all in the original code, it may be a readonly
584 memory. To check that A is not const-qualified, we check that
585 there exists at least an unconditional write to A in the current
588 - reads or writes to memory are valid memory accesses for every
589 iteration. To check that the memory accesses are correctly formed
590 and that we are allowed to read and write in these locations, we
591 check that the memory accesses to be if-converted occur at every
592 iteration unconditionally. */
595 ifcvt_memrefs_wont_trap (gimple stmt
, VEC (data_reference_p
, heap
) *refs
)
597 return write_memrefs_written_at_least_once (stmt
, refs
)
598 && memrefs_read_or_written_unconditionally (stmt
, refs
);
601 /* Wrapper around gimple_could_trap_p refined for the needs of the
602 if-conversion. Try to prove that the memory accesses of STMT could
603 not trap in the innermost loop containing STMT. */
606 ifcvt_could_trap_p (gimple stmt
, VEC (data_reference_p
, heap
) *refs
)
608 if (gimple_vuse (stmt
)
609 && !gimple_could_trap_p_1 (stmt
, false, false)
610 && ifcvt_memrefs_wont_trap (stmt
, refs
))
613 return gimple_could_trap_p (stmt
);
616 /* Return true when STMT is if-convertible.
618 GIMPLE_ASSIGN statement is not if-convertible if,
621 - LHS is not var decl. */
624 if_convertible_gimple_assign_stmt_p (gimple stmt
,
625 VEC (data_reference_p
, heap
) *refs
)
627 tree lhs
= gimple_assign_lhs (stmt
);
630 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
632 fprintf (dump_file
, "-------------------------\n");
633 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
636 if (!is_gimple_reg_type (TREE_TYPE (lhs
)))
639 /* Some of these constrains might be too conservative. */
640 if (stmt_ends_bb_p (stmt
)
641 || gimple_has_volatile_ops (stmt
)
642 || (TREE_CODE (lhs
) == SSA_NAME
643 && SSA_NAME_OCCURS_IN_ABNORMAL_PHI (lhs
))
644 || gimple_has_side_effects (stmt
))
646 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
647 fprintf (dump_file
, "stmt not suitable for ifcvt\n");
651 if (flag_tree_loop_if_convert_stores
)
653 if (ifcvt_could_trap_p (stmt
, refs
))
655 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
656 fprintf (dump_file
, "tree could trap...\n");
662 if (gimple_assign_rhs_could_trap_p (stmt
))
664 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
665 fprintf (dump_file
, "tree could trap...\n");
669 bb
= gimple_bb (stmt
);
671 if (TREE_CODE (lhs
) != SSA_NAME
672 && bb
!= bb
->loop_father
->header
673 && !bb_with_exit_edge_p (bb
->loop_father
, bb
))
675 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
677 fprintf (dump_file
, "LHS is not var\n");
678 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
686 /* Return true when STMT is if-convertible.
688 A statement is if-convertible if:
689 - it is an if-convertible GIMPLE_ASSGIN,
690 - it is a GIMPLE_LABEL or a GIMPLE_COND. */
693 if_convertible_stmt_p (gimple stmt
, VEC (data_reference_p
, heap
) *refs
)
695 switch (gimple_code (stmt
))
703 return if_convertible_gimple_assign_stmt_p (stmt
, refs
);
706 /* Don't know what to do with 'em so don't do anything. */
707 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
709 fprintf (dump_file
, "don't know what to do\n");
710 print_gimple_stmt (dump_file
, stmt
, 0, TDF_SLIM
);
719 /* Return true when BB post-dominates all its predecessors. */
722 bb_postdominates_preds (basic_block bb
)
726 for (i
= 0; i
< EDGE_COUNT (bb
->preds
); i
++)
727 if (!dominated_by_p (CDI_POST_DOMINATORS
, EDGE_PRED (bb
, i
)->src
, bb
))
733 /* Return true when BB is if-convertible. This routine does not check
734 basic block's statements and phis.
736 A basic block is not if-convertible if:
737 - it is non-empty and it is after the exit block (in BFS order),
738 - it is after the exit block but before the latch,
739 - its edges are not normal.
741 EXIT_BB is the basic block containing the exit of the LOOP. BB is
745 if_convertible_bb_p (struct loop
*loop
, basic_block bb
, basic_block exit_bb
)
750 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
751 fprintf (dump_file
, "----------[%d]-------------\n", bb
->index
);
753 if (EDGE_COUNT (bb
->preds
) > 2
754 || EDGE_COUNT (bb
->succs
) > 2)
759 if (bb
!= loop
->latch
)
761 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
762 fprintf (dump_file
, "basic block after exit bb but before latch\n");
765 else if (!empty_block_p (bb
))
767 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
768 fprintf (dump_file
, "non empty basic block after exit bb\n");
771 else if (bb
== loop
->latch
773 && !dominated_by_p (CDI_DOMINATORS
, bb
, exit_bb
))
775 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
776 fprintf (dump_file
, "latch is not dominated by exit_block\n");
781 /* Be less adventurous and handle only normal edges. */
782 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
784 (EDGE_ABNORMAL_CALL
| EDGE_EH
| EDGE_ABNORMAL
| EDGE_IRREDUCIBLE_LOOP
))
786 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
787 fprintf (dump_file
, "Difficult to handle edges\n");
791 if (EDGE_COUNT (bb
->preds
) == 2
792 && bb
!= loop
->header
793 && !bb_postdominates_preds (bb
))
799 /* Return true when all predecessor blocks of BB are visited. The
800 VISITED bitmap keeps track of the visited blocks. */
803 pred_blocks_visited_p (basic_block bb
, bitmap
*visited
)
807 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
808 if (!bitmap_bit_p (*visited
, e
->src
->index
))
814 /* Get body of a LOOP in suitable order for if-conversion. It is
815 caller's responsibility to deallocate basic block list.
816 If-conversion suitable order is, breadth first sort (BFS) order
817 with an additional constraint: select a block only if all its
818 predecessors are already selected. */
821 get_loop_body_in_if_conv_order (const struct loop
*loop
)
823 basic_block
*blocks
, *blocks_in_bfs_order
;
826 unsigned int index
= 0;
827 unsigned int visited_count
= 0;
829 gcc_assert (loop
->num_nodes
);
830 gcc_assert (loop
->latch
!= EXIT_BLOCK_PTR
);
832 blocks
= XCNEWVEC (basic_block
, loop
->num_nodes
);
833 visited
= BITMAP_ALLOC (NULL
);
835 blocks_in_bfs_order
= get_loop_body_in_bfs_order (loop
);
838 while (index
< loop
->num_nodes
)
840 bb
= blocks_in_bfs_order
[index
];
842 if (bb
->flags
& BB_IRREDUCIBLE_LOOP
)
844 free (blocks_in_bfs_order
);
845 BITMAP_FREE (visited
);
850 if (!bitmap_bit_p (visited
, bb
->index
))
852 if (pred_blocks_visited_p (bb
, &visited
)
853 || bb
== loop
->header
)
855 /* This block is now visited. */
856 bitmap_set_bit (visited
, bb
->index
);
857 blocks
[visited_count
++] = bb
;
863 if (index
== loop
->num_nodes
864 && visited_count
!= loop
->num_nodes
)
868 free (blocks_in_bfs_order
);
869 BITMAP_FREE (visited
);
873 /* Returns true when the analysis of the predicates for all the basic
874 blocks in LOOP succeeded.
876 predicate_bbs first allocates the predicates of the basic blocks.
877 These fields are then initialized with the tree expressions
878 representing the predicates under which a basic block is executed
879 in the LOOP. As the loop->header is executed at each iteration, it
880 has the "true" predicate. Other statements executed under a
881 condition are predicated with that condition, for example
888 S1 will be predicated with "x", and
889 S2 will be predicated with "!x". */
892 predicate_bbs (loop_p loop
)
896 for (i
= 0; i
< loop
->num_nodes
; i
++)
897 init_bb_predicate (ifc_bbs
[i
]);
899 for (i
= 0; i
< loop
->num_nodes
; i
++)
901 basic_block bb
= ifc_bbs
[i
];
903 gimple_stmt_iterator itr
;
905 /* The loop latch is always executed and has no extra conditions
906 to be processed: skip it. */
907 if (bb
== loop
->latch
)
909 reset_bb_predicate (loop
->latch
);
913 cond
= bb_predicate (bb
);
915 && bb
!= loop
->header
)
919 cond
= force_gimple_operand (cond
, &stmts
, true, NULL_TREE
);
920 add_bb_predicate_gimplified_stmts (bb
, stmts
);
923 for (itr
= gsi_start_bb (bb
); !gsi_end_p (itr
); gsi_next (&itr
))
925 gimple stmt
= gsi_stmt (itr
);
927 switch (gimple_code (stmt
))
938 edge true_edge
, false_edge
;
939 location_t loc
= gimple_location (stmt
);
940 tree c
= fold_build2_loc (loc
, gimple_cond_code (stmt
),
942 gimple_cond_lhs (stmt
),
943 gimple_cond_rhs (stmt
));
945 /* Add new condition into destination's predicate list. */
946 extract_true_false_edges_from_block (gimple_bb (stmt
),
947 &true_edge
, &false_edge
);
949 /* If C is true, then TRUE_EDGE is taken. */
950 add_to_dst_predicate_list (loop
, true_edge
, cond
, unshare_expr (c
));
952 /* If C is false, then FALSE_EDGE is taken. */
953 c2
= invert_truthvalue_loc (loc
, unshare_expr (c
));
954 tem
= canonicalize_cond_expr_cond (c2
);
957 add_to_dst_predicate_list (loop
, false_edge
, cond
, c2
);
964 /* Not handled yet in if-conversion. */
969 /* If current bb has only one successor, then consider it as an
970 unconditional goto. */
971 if (single_succ_p (bb
))
973 basic_block bb_n
= single_succ (bb
);
975 /* The successor bb inherits the predicate of its
976 predecessor. If there is no predicate in the predecessor
977 bb, then consider the successor bb as always executed. */
978 if (cond
== NULL_TREE
)
979 cond
= boolean_true_node
;
981 add_to_predicate_list (bb_n
, cond
);
985 /* The loop header is always executed. */
986 reset_bb_predicate (loop
->header
);
987 gcc_assert (bb_predicate_gimplified_stmts (loop
->header
) == NULL
988 && bb_predicate_gimplified_stmts (loop
->latch
) == NULL
);
993 /* Return true when LOOP is if-convertible. This is a helper function
994 for if_convertible_loop_p. REFS and DDRS are initialized and freed
995 in if_convertible_loop_p. */
998 if_convertible_loop_p_1 (struct loop
*loop
,
999 VEC (loop_p
, heap
) **loop_nest
,
1000 VEC (data_reference_p
, heap
) **refs
,
1001 VEC (ddr_p
, heap
) **ddrs
)
1005 basic_block exit_bb
= NULL
;
1007 /* Don't if-convert the loop when the data dependences cannot be
1008 computed: the loop won't be vectorized in that case. */
1009 res
= compute_data_dependences_for_loop (loop
, true, loop_nest
, refs
, ddrs
);
1013 calculate_dominance_info (CDI_DOMINATORS
);
1014 calculate_dominance_info (CDI_POST_DOMINATORS
);
1016 /* Allow statements that can be handled during if-conversion. */
1017 ifc_bbs
= get_loop_body_in_if_conv_order (loop
);
1020 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1021 fprintf (dump_file
, "Irreducible loop\n");
1025 for (i
= 0; i
< loop
->num_nodes
; i
++)
1027 basic_block bb
= ifc_bbs
[i
];
1029 if (!if_convertible_bb_p (loop
, bb
, exit_bb
))
1032 if (bb_with_exit_edge_p (loop
, bb
))
1036 res
= predicate_bbs (loop
);
1040 if (flag_tree_loop_if_convert_stores
)
1042 data_reference_p dr
;
1044 for (i
= 0; VEC_iterate (data_reference_p
, *refs
, i
, dr
); i
++)
1046 dr
->aux
= XNEW (struct ifc_dr
);
1047 DR_WRITTEN_AT_LEAST_ONCE (dr
) = -1;
1048 DR_RW_UNCONDITIONALLY (dr
) = -1;
1052 for (i
= 0; i
< loop
->num_nodes
; i
++)
1054 basic_block bb
= ifc_bbs
[i
];
1055 gimple_stmt_iterator itr
;
1057 for (itr
= gsi_start_phis (bb
); !gsi_end_p (itr
); gsi_next (&itr
))
1058 if (!if_convertible_phi_p (loop
, bb
, gsi_stmt (itr
)))
1061 /* Check the if-convertibility of statements in predicated BBs. */
1062 if (is_predicated (bb
))
1063 for (itr
= gsi_start_bb (bb
); !gsi_end_p (itr
); gsi_next (&itr
))
1064 if (!if_convertible_stmt_p (gsi_stmt (itr
), *refs
))
1069 fprintf (dump_file
, "Applying if-conversion\n");
1074 /* Return true when LOOP is if-convertible.
1075 LOOP is if-convertible if:
1077 - it has two or more basic blocks,
1078 - it has only one exit,
1079 - loop header is not the exit edge,
1080 - if its basic blocks and phi nodes are if convertible. */
1083 if_convertible_loop_p (struct loop
*loop
)
1088 VEC (data_reference_p
, heap
) *refs
;
1089 VEC (ddr_p
, heap
) *ddrs
;
1090 VEC (loop_p
, heap
) *loop_nest
;
1092 /* Handle only innermost loop. */
1093 if (!loop
|| loop
->inner
)
1095 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1096 fprintf (dump_file
, "not innermost loop\n");
1100 /* If only one block, no need for if-conversion. */
1101 if (loop
->num_nodes
<= 2)
1103 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1104 fprintf (dump_file
, "less than 2 basic blocks\n");
1108 /* More than one loop exit is too much to handle. */
1109 if (!single_exit (loop
))
1111 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1112 fprintf (dump_file
, "multiple exits\n");
1116 /* If one of the loop header's edge is an exit edge then do not
1117 apply if-conversion. */
1118 FOR_EACH_EDGE (e
, ei
, loop
->header
->succs
)
1119 if (loop_exit_edge_p (loop
, e
))
1122 refs
= VEC_alloc (data_reference_p
, heap
, 5);
1123 ddrs
= VEC_alloc (ddr_p
, heap
, 25);
1124 loop_nest
= VEC_alloc (loop_p
, heap
, 3);
1125 res
= if_convertible_loop_p_1 (loop
, &loop_nest
, &refs
, &ddrs
);
1127 if (flag_tree_loop_if_convert_stores
)
1129 data_reference_p dr
;
1132 for (i
= 0; VEC_iterate (data_reference_p
, refs
, i
, dr
); i
++)
1136 VEC_free (loop_p
, heap
, loop_nest
);
1137 free_data_refs (refs
);
1138 free_dependence_relations (ddrs
);
1142 /* Basic block BB has two predecessors. Using predecessor's bb
1143 predicate, set an appropriate condition COND for the PHI node
1144 replacement. Return the true block whose phi arguments are
1145 selected when cond is true. LOOP is the loop containing the
1146 if-converted region, GSI is the place to insert the code for the
1150 find_phi_replacement_condition (struct loop
*loop
,
1151 basic_block bb
, tree
*cond
,
1152 gimple_stmt_iterator
*gsi
)
1154 edge first_edge
, second_edge
;
1157 gcc_assert (EDGE_COUNT (bb
->preds
) == 2);
1158 first_edge
= EDGE_PRED (bb
, 0);
1159 second_edge
= EDGE_PRED (bb
, 1);
1161 /* Use condition based on following criteria:
1167 S2 is preferred over S1. Make 'b' first_bb and use its condition.
1169 2) Do not make loop header first_bb.
1172 S1: x = !(c == d)? a : b;
1175 S22: x = t1 ? b : a;
1177 S3: x = (c == d) ? b : a;
1179 S3 is preferred over S1 and S2*, Make 'b' first_bb and use
1182 4) If pred B is dominated by pred A then use pred B's condition.
1185 /* Select condition that is not TRUTH_NOT_EXPR. */
1186 tmp_cond
= bb_predicate (first_edge
->src
);
1187 gcc_assert (tmp_cond
);
1189 if (TREE_CODE (tmp_cond
) == TRUTH_NOT_EXPR
)
1193 tmp_edge
= first_edge
;
1194 first_edge
= second_edge
;
1195 second_edge
= tmp_edge
;
1198 /* Check if FIRST_BB is loop header or not and make sure that
1199 FIRST_BB does not dominate SECOND_BB. */
1200 if (first_edge
->src
== loop
->header
1201 || dominated_by_p (CDI_DOMINATORS
,
1202 second_edge
->src
, first_edge
->src
))
1204 *cond
= bb_predicate (second_edge
->src
);
1206 if (TREE_CODE (*cond
) == TRUTH_NOT_EXPR
)
1207 *cond
= invert_truthvalue (*cond
);
1209 /* Select non loop header bb. */
1210 first_edge
= second_edge
;
1213 *cond
= bb_predicate (first_edge
->src
);
1215 /* Gimplify the condition: the vectorizer prefers to have gimple
1216 values as conditions. Various targets use different means to
1217 communicate conditions in vector compare operations. Using a
1218 gimple value allows the compiler to emit vector compare and
1219 select RTL without exposing compare's result. */
1220 *cond
= force_gimple_operand_gsi (gsi
, unshare_expr (*cond
),
1222 true, GSI_SAME_STMT
);
1223 if (!is_gimple_reg (*cond
) && !is_gimple_condexpr (*cond
))
1224 *cond
= ifc_temp_var (TREE_TYPE (*cond
), unshare_expr (*cond
), gsi
);
1228 return first_edge
->src
;
1231 /* Replace a scalar PHI node with a COND_EXPR using COND as condition.
1232 This routine does not handle PHI nodes with more than two
1236 S1: A = PHI <x1(1), x2(5)
1238 S2: A = cond ? x1 : x2;
1240 The generated code is inserted at GSI that points to the top of
1241 basic block's statement list. When COND is true, phi arg from
1242 TRUE_BB is selected. */
1245 predicate_scalar_phi (gimple phi
, tree cond
,
1246 basic_block true_bb
,
1247 gimple_stmt_iterator
*gsi
)
1251 tree rhs
, res
, arg
, scev
;
1253 gcc_assert (gimple_code (phi
) == GIMPLE_PHI
1254 && gimple_phi_num_args (phi
) == 2);
1256 res
= gimple_phi_result (phi
);
1257 /* Do not handle virtual phi nodes. */
1258 if (!is_gimple_reg (SSA_NAME_VAR (res
)))
1261 bb
= gimple_bb (phi
);
1263 if ((arg
= degenerate_phi_result (phi
))
1264 || ((scev
= analyze_scalar_evolution (gimple_bb (phi
)->loop_father
,
1266 && !chrec_contains_undetermined (scev
)
1268 && (arg
= gimple_phi_arg_def (phi
, 0))))
1273 /* Use condition that is not TRUTH_NOT_EXPR in conditional modify expr. */
1274 if (EDGE_PRED (bb
, 1)->src
== true_bb
)
1276 arg_0
= gimple_phi_arg_def (phi
, 1);
1277 arg_1
= gimple_phi_arg_def (phi
, 0);
1281 arg_0
= gimple_phi_arg_def (phi
, 0);
1282 arg_1
= gimple_phi_arg_def (phi
, 1);
1285 gcc_checking_assert (bb
== bb
->loop_father
->header
1286 || bb_postdominates_preds (bb
));
1288 /* Build new RHS using selected condition and arguments. */
1289 rhs
= build3 (COND_EXPR
, TREE_TYPE (res
),
1290 unshare_expr (cond
), arg_0
, arg_1
);
1293 new_stmt
= gimple_build_assign (res
, rhs
);
1294 SSA_NAME_DEF_STMT (gimple_phi_result (phi
)) = new_stmt
;
1295 gsi_insert_before (gsi
, new_stmt
, GSI_SAME_STMT
);
1296 update_stmt (new_stmt
);
1298 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1300 fprintf (dump_file
, "new phi replacement stmt\n");
1301 print_gimple_stmt (dump_file
, new_stmt
, 0, TDF_SLIM
);
1305 /* Replaces in LOOP all the scalar phi nodes other than those in the
1306 LOOP->header block with conditional modify expressions. */
1309 predicate_all_scalar_phis (struct loop
*loop
)
1312 unsigned int orig_loop_num_nodes
= loop
->num_nodes
;
1315 for (i
= 1; i
< orig_loop_num_nodes
; i
++)
1318 tree cond
= NULL_TREE
;
1319 gimple_stmt_iterator gsi
, phi_gsi
;
1320 basic_block true_bb
= NULL
;
1323 if (bb
== loop
->header
)
1326 phi_gsi
= gsi_start_phis (bb
);
1327 if (gsi_end_p (phi_gsi
))
1330 /* BB has two predecessors. Using predecessor's aux field, set
1331 appropriate condition for the PHI node replacement. */
1332 gsi
= gsi_after_labels (bb
);
1333 true_bb
= find_phi_replacement_condition (loop
, bb
, &cond
, &gsi
);
1335 while (!gsi_end_p (phi_gsi
))
1337 phi
= gsi_stmt (phi_gsi
);
1338 predicate_scalar_phi (phi
, cond
, true_bb
, &gsi
);
1339 release_phi_node (phi
);
1340 gsi_next (&phi_gsi
);
1343 set_phi_nodes (bb
, NULL
);
1347 /* Insert in each basic block of LOOP the statements produced by the
1348 gimplification of the predicates. */
1351 insert_gimplified_predicates (loop_p loop
)
1355 for (i
= 0; i
< loop
->num_nodes
; i
++)
1357 basic_block bb
= ifc_bbs
[i
];
1360 if (!is_predicated (bb
))
1362 /* Do not insert statements for a basic block that is not
1363 predicated. Also make sure that the predicate of the
1364 basic block is set to true. */
1365 reset_bb_predicate (bb
);
1369 stmts
= bb_predicate_gimplified_stmts (bb
);
1372 if (flag_tree_loop_if_convert_stores
)
1374 /* Insert the predicate of the BB just after the label,
1375 as the if-conversion of memory writes will use this
1377 gimple_stmt_iterator gsi
= gsi_after_labels (bb
);
1378 gsi_insert_seq_before (&gsi
, stmts
, GSI_SAME_STMT
);
1382 /* Insert the predicate of the BB at the end of the BB
1383 as this would reduce the register pressure: the only
1384 use of this predicate will be in successor BBs. */
1385 gimple_stmt_iterator gsi
= gsi_last_bb (bb
);
1388 || stmt_ends_bb_p (gsi_stmt (gsi
)))
1389 gsi_insert_seq_before (&gsi
, stmts
, GSI_SAME_STMT
);
1391 gsi_insert_seq_after (&gsi
, stmts
, GSI_SAME_STMT
);
1394 /* Once the sequence is code generated, set it to NULL. */
1395 set_bb_predicate_gimplified_stmts (bb
, NULL
);
1400 /* Predicate each write to memory in LOOP.
1402 This function transforms control flow constructs containing memory
1405 | for (i = 0; i < N; i++)
1409 into the following form that does not contain control flow:
1411 | for (i = 0; i < N; i++)
1412 | A[i] = cond ? expr : A[i];
1414 The original CFG looks like this:
1421 | if (i < N) goto bb_5 else goto bb_2
1425 | cond = some_computation;
1426 | if (cond) goto bb_3 else goto bb_4
1438 insert_gimplified_predicates inserts the computation of the COND
1439 expression at the beginning of the destination basic block:
1446 | if (i < N) goto bb_5 else goto bb_2
1450 | cond = some_computation;
1451 | if (cond) goto bb_3 else goto bb_4
1455 | cond = some_computation;
1464 predicate_mem_writes is then predicating the memory write as follows:
1471 | if (i < N) goto bb_5 else goto bb_2
1475 | if (cond) goto bb_3 else goto bb_4
1479 | cond = some_computation;
1480 | A[i] = cond ? expr : A[i];
1488 and finally combine_blocks removes the basic block boundaries making
1489 the loop vectorizable:
1493 | if (i < N) goto bb_5 else goto bb_1
1497 | cond = some_computation;
1498 | A[i] = cond ? expr : A[i];
1499 | if (i < N) goto bb_5 else goto bb_4
1508 predicate_mem_writes (loop_p loop
)
1510 unsigned int i
, orig_loop_num_nodes
= loop
->num_nodes
;
1512 for (i
= 1; i
< orig_loop_num_nodes
; i
++)
1514 gimple_stmt_iterator gsi
;
1515 basic_block bb
= ifc_bbs
[i
];
1516 tree cond
= bb_predicate (bb
);
1519 if (is_true_predicate (cond
))
1522 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1523 if ((stmt
= gsi_stmt (gsi
))
1524 && gimple_assign_single_p (stmt
)
1525 && gimple_vdef (stmt
))
1527 tree lhs
= gimple_assign_lhs (stmt
);
1528 tree rhs
= gimple_assign_rhs1 (stmt
);
1529 tree type
= TREE_TYPE (lhs
);
1531 lhs
= ifc_temp_var (type
, unshare_expr (lhs
), &gsi
);
1532 rhs
= ifc_temp_var (type
, unshare_expr (rhs
), &gsi
);
1533 rhs
= build3 (COND_EXPR
, type
, unshare_expr (cond
), rhs
, lhs
);
1534 gimple_assign_set_rhs1 (stmt
, ifc_temp_var (type
, rhs
, &gsi
));
1540 /* Remove all GIMPLE_CONDs and GIMPLE_LABELs of all the basic blocks
1541 other than the exit and latch of the LOOP. Also resets the
1542 GIMPLE_DEBUG information. */
1545 remove_conditions_and_labels (loop_p loop
)
1547 gimple_stmt_iterator gsi
;
1550 for (i
= 0; i
< loop
->num_nodes
; i
++)
1552 basic_block bb
= ifc_bbs
[i
];
1554 if (bb_with_exit_edge_p (loop
, bb
)
1555 || bb
== loop
->latch
)
1558 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); )
1559 switch (gimple_code (gsi_stmt (gsi
)))
1563 gsi_remove (&gsi
, true);
1567 /* ??? Should there be conditional GIMPLE_DEBUG_BINDs? */
1568 if (gimple_debug_bind_p (gsi_stmt (gsi
)))
1570 gimple_debug_bind_reset_value (gsi_stmt (gsi
));
1571 update_stmt (gsi_stmt (gsi
));
1582 /* Combine all the basic blocks from LOOP into one or two super basic
1583 blocks. Replace PHI nodes with conditional modify expressions. */
1586 combine_blocks (struct loop
*loop
)
1588 basic_block bb
, exit_bb
, merge_target_bb
;
1589 unsigned int orig_loop_num_nodes
= loop
->num_nodes
;
1594 remove_conditions_and_labels (loop
);
1595 insert_gimplified_predicates (loop
);
1596 predicate_all_scalar_phis (loop
);
1598 if (flag_tree_loop_if_convert_stores
)
1599 predicate_mem_writes (loop
);
1601 /* Merge basic blocks: first remove all the edges in the loop,
1602 except for those from the exit block. */
1604 for (i
= 0; i
< orig_loop_num_nodes
; i
++)
1607 if (bb_with_exit_edge_p (loop
, bb
))
1613 gcc_assert (exit_bb
!= loop
->latch
);
1615 for (i
= 1; i
< orig_loop_num_nodes
; i
++)
1619 for (ei
= ei_start (bb
->preds
); (e
= ei_safe_edge (ei
));)
1621 if (e
->src
== exit_bb
)
1628 if (exit_bb
!= NULL
)
1630 if (exit_bb
!= loop
->header
)
1632 /* Connect this node to loop header. */
1633 make_edge (loop
->header
, exit_bb
, EDGE_FALLTHRU
);
1634 set_immediate_dominator (CDI_DOMINATORS
, exit_bb
, loop
->header
);
1637 /* Redirect non-exit edges to loop->latch. */
1638 FOR_EACH_EDGE (e
, ei
, exit_bb
->succs
)
1640 if (!loop_exit_edge_p (loop
, e
))
1641 redirect_edge_and_branch (e
, loop
->latch
);
1643 set_immediate_dominator (CDI_DOMINATORS
, loop
->latch
, exit_bb
);
1647 /* If the loop does not have an exit, reconnect header and latch. */
1648 make_edge (loop
->header
, loop
->latch
, EDGE_FALLTHRU
);
1649 set_immediate_dominator (CDI_DOMINATORS
, loop
->latch
, loop
->header
);
1652 merge_target_bb
= loop
->header
;
1653 for (i
= 1; i
< orig_loop_num_nodes
; i
++)
1655 gimple_stmt_iterator gsi
;
1656 gimple_stmt_iterator last
;
1660 if (bb
== exit_bb
|| bb
== loop
->latch
)
1663 /* Make stmts member of loop->header. */
1664 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1665 gimple_set_bb (gsi_stmt (gsi
), merge_target_bb
);
1667 /* Update stmt list. */
1668 last
= gsi_last_bb (merge_target_bb
);
1669 gsi_insert_seq_after (&last
, bb_seq (bb
), GSI_NEW_STMT
);
1670 set_bb_seq (bb
, NULL
);
1672 delete_basic_block (bb
);
1675 /* If possible, merge loop header to the block with the exit edge.
1676 This reduces the number of basic blocks to two, to please the
1677 vectorizer that handles only loops with two nodes. */
1679 && exit_bb
!= loop
->header
1680 && can_merge_blocks_p (loop
->header
, exit_bb
))
1681 merge_blocks (loop
->header
, exit_bb
);
1684 /* If-convert LOOP when it is legal. For the moment this pass has no
1685 profitability analysis. Returns true when something changed. */
1688 tree_if_conversion (struct loop
*loop
)
1690 bool changed
= false;
1693 if (!if_convertible_loop_p (loop
)
1694 || !dbg_cnt (if_conversion_tree
))
1697 /* Now all statements are if-convertible. Combine all the basic
1698 blocks into one huge basic block doing the if-conversion
1700 combine_blocks (loop
);
1702 if (flag_tree_loop_if_convert_stores
)
1703 mark_sym_for_renaming (gimple_vop (cfun
));
1712 for (i
= 0; i
< loop
->num_nodes
; i
++)
1713 free_bb_predicate (ifc_bbs
[i
]);
1722 /* Tree if-conversion pass management. */
1725 main_tree_if_conversion (void)
1729 bool changed
= false;
1732 if (number_of_loops () <= 1)
1735 FOR_EACH_LOOP (li
, loop
, 0)
1736 changed
|= tree_if_conversion (loop
);
1739 todo
|= TODO_cleanup_cfg
;
1741 if (changed
&& flag_tree_loop_if_convert_stores
)
1742 todo
|= TODO_update_ssa_only_virtuals
;
1747 /* Returns true when the if-conversion pass is enabled. */
1750 gate_tree_if_conversion (void)
1752 return ((flag_tree_vectorize
&& flag_tree_loop_if_convert
!= 0)
1753 || flag_tree_loop_if_convert
== 1
1754 || flag_tree_loop_if_convert_stores
== 1);
1757 struct gimple_opt_pass pass_if_conversion
=
1762 gate_tree_if_conversion
, /* gate */
1763 main_tree_if_conversion
, /* execute */
1766 0, /* static_pass_number */
1767 TV_NONE
, /* tv_id */
1768 PROP_cfg
| PROP_ssa
, /* properties_required */
1769 0, /* properties_provided */
1770 0, /* properties_destroyed */
1771 0, /* todo_flags_start */
1772 TODO_dump_func
| TODO_verify_stmts
| TODO_verify_flow
1773 /* todo_flags_finish */