1 /* Global, SSA-based optimizations using mathematical identities.
2 Copyright (C) 2005-2014 Free Software Foundation, Inc.
4 This file is part of GCC.
6 GCC is free software; you can redistribute it and/or modify it
7 under the terms of the GNU General Public License as published by the
8 Free Software Foundation; either version 3, or (at your option) any
11 GCC is distributed in the hope that it will be useful, but WITHOUT
12 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 /* Currently, the only mini-pass in this file tries to CSE reciprocal
21 operations. These are common in sequences such as this one:
23 modulus = sqrt(x*x + y*y + z*z);
28 that can be optimized to
30 modulus = sqrt(x*x + y*y + z*z);
31 rmodulus = 1.0 / modulus;
36 We do this for loop invariant divisors, and with this pass whenever
37 we notice that a division has the same divisor multiple times.
39 Of course, like in PRE, we don't insert a division if a dominator
40 already has one. However, this cannot be done as an extension of
41 PRE for several reasons.
43 First of all, with some experiments it was found out that the
44 transformation is not always useful if there are only two divisions
45 hy the same divisor. This is probably because modern processors
46 can pipeline the divisions; on older, in-order processors it should
47 still be effective to optimize two divisions by the same number.
48 We make this a param, and it shall be called N in the remainder of
51 Second, if trapping math is active, we have less freedom on where
52 to insert divisions: we can only do so in basic blocks that already
53 contain one. (If divisions don't trap, instead, we can insert
54 divisions elsewhere, which will be in blocks that are common dominators
55 of those that have the division).
57 We really don't want to compute the reciprocal unless a division will
58 be found. To do this, we won't insert the division in a basic block
59 that has less than N divisions *post-dominating* it.
61 The algorithm constructs a subset of the dominator tree, holding the
62 blocks containing the divisions and the common dominators to them,
63 and walk it twice. The first walk is in post-order, and it annotates
64 each block with the number of divisions that post-dominate it: this
65 gives information on where divisions can be inserted profitably.
66 The second walk is in pre-order, and it inserts divisions as explained
67 above, and replaces divisions by multiplications.
69 In the best case, the cost of the pass is O(n_statements). In the
70 worst-case, the cost is due to creating the dominator tree subset,
71 with a cost of O(n_basic_blocks ^ 2); however this can only happen
72 for n_statements / n_basic_blocks statements. So, the amortized cost
73 of creating the dominator tree subset is O(n_basic_blocks) and the
74 worst-case cost of the pass is O(n_statements * n_basic_blocks).
76 More practically, the cost will be small because there are few
77 divisions, and they tend to be in the same basic block, so insert_bb
78 is called very few times.
80 If we did this using domwalk.c, an efficient implementation would have
81 to work on all the variables in a single pass, because we could not
82 work on just a subset of the dominator tree, as we do now, and the
83 cost would also be something like O(n_statements * n_basic_blocks).
84 The data structures would be more complex in order to work on all the
85 variables in a single pass. */
89 #include "coretypes.h"
93 #include "basic-block.h"
94 #include "tree-ssa-alias.h"
95 #include "internal-fn.h"
96 #include "gimple-fold.h"
97 #include "gimple-expr.h"
100 #include "gimple-iterator.h"
101 #include "gimplify-me.h"
102 #include "stor-layout.h"
103 #include "gimple-ssa.h"
104 #include "tree-cfg.h"
105 #include "tree-phinodes.h"
106 #include "ssa-iterators.h"
107 #include "stringpool.h"
108 #include "tree-ssanames.h"
110 #include "tree-dfa.h"
111 #include "tree-ssa.h"
112 #include "tree-pass.h"
113 #include "alloc-pool.h"
115 #include "gimple-pretty-print.h"
117 /* FIXME: RTL headers have to be included here for optabs. */
118 #include "rtl.h" /* Because optabs.h wants enum rtx_code. */
119 #include "expr.h" /* Because optabs.h wants sepops. */
122 /* This structure represents one basic block that either computes a
123 division, or is a common dominator for basic block that compute a
126 /* The basic block represented by this structure. */
129 /* If non-NULL, the SSA_NAME holding the definition for a reciprocal
133 /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that
134 was inserted in BB. */
135 gimple recip_def_stmt
;
137 /* Pointer to a list of "struct occurrence"s for blocks dominated
139 struct occurrence
*children
;
141 /* Pointer to the next "struct occurrence"s in the list of blocks
142 sharing a common dominator. */
143 struct occurrence
*next
;
145 /* The number of divisions that are in BB before compute_merit. The
146 number of divisions that are in BB or post-dominate it after
150 /* True if the basic block has a division, false if it is a common
151 dominator for basic blocks that do. If it is false and trapping
152 math is active, BB is not a candidate for inserting a reciprocal. */
153 bool bb_has_division
;
158 /* Number of 1.0/X ops inserted. */
161 /* Number of 1.0/FUNC ops inserted. */
167 /* Number of cexpi calls inserted. */
173 /* Number of hand-written 16-bit bswaps found. */
176 /* Number of hand-written 32-bit bswaps found. */
179 /* Number of hand-written 64-bit bswaps found. */
185 /* Number of widening multiplication ops inserted. */
186 int widen_mults_inserted
;
188 /* Number of integer multiply-and-accumulate ops inserted. */
191 /* Number of fp fused multiply-add ops inserted. */
195 /* The instance of "struct occurrence" representing the highest
196 interesting block in the dominator tree. */
197 static struct occurrence
*occ_head
;
199 /* Allocation pool for getting instances of "struct occurrence". */
200 static alloc_pool occ_pool
;
204 /* Allocate and return a new struct occurrence for basic block BB, and
205 whose children list is headed by CHILDREN. */
206 static struct occurrence
*
207 occ_new (basic_block bb
, struct occurrence
*children
)
209 struct occurrence
*occ
;
211 bb
->aux
= occ
= (struct occurrence
*) pool_alloc (occ_pool
);
212 memset (occ
, 0, sizeof (struct occurrence
));
215 occ
->children
= children
;
220 /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a
221 list of "struct occurrence"s, one per basic block, having IDOM as
222 their common dominator.
224 We try to insert NEW_OCC as deep as possible in the tree, and we also
225 insert any other block that is a common dominator for BB and one
226 block already in the tree. */
229 insert_bb (struct occurrence
*new_occ
, basic_block idom
,
230 struct occurrence
**p_head
)
232 struct occurrence
*occ
, **p_occ
;
234 for (p_occ
= p_head
; (occ
= *p_occ
) != NULL
; )
236 basic_block bb
= new_occ
->bb
, occ_bb
= occ
->bb
;
237 basic_block dom
= nearest_common_dominator (CDI_DOMINATORS
, occ_bb
, bb
);
240 /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC
243 occ
->next
= new_occ
->children
;
244 new_occ
->children
= occ
;
246 /* Try the next block (it may as well be dominated by BB). */
249 else if (dom
== occ_bb
)
251 /* OCC_BB dominates BB. Tail recurse to look deeper. */
252 insert_bb (new_occ
, dom
, &occ
->children
);
256 else if (dom
!= idom
)
258 gcc_assert (!dom
->aux
);
260 /* There is a dominator between IDOM and BB, add it and make
261 two children out of NEW_OCC and OCC. First, remove OCC from
267 /* None of the previous blocks has DOM as a dominator: if we tail
268 recursed, we would reexamine them uselessly. Just switch BB with
269 DOM, and go on looking for blocks dominated by DOM. */
270 new_occ
= occ_new (dom
, new_occ
);
275 /* Nothing special, go on with the next element. */
280 /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */
281 new_occ
->next
= *p_head
;
285 /* Register that we found a division in BB. */
288 register_division_in (basic_block bb
)
290 struct occurrence
*occ
;
292 occ
= (struct occurrence
*) bb
->aux
;
295 occ
= occ_new (bb
, NULL
);
296 insert_bb (occ
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), &occ_head
);
299 occ
->bb_has_division
= true;
300 occ
->num_divisions
++;
304 /* Compute the number of divisions that postdominate each block in OCC and
308 compute_merit (struct occurrence
*occ
)
310 struct occurrence
*occ_child
;
311 basic_block dom
= occ
->bb
;
313 for (occ_child
= occ
->children
; occ_child
; occ_child
= occ_child
->next
)
316 if (occ_child
->children
)
317 compute_merit (occ_child
);
320 bb
= single_noncomplex_succ (dom
);
324 if (dominated_by_p (CDI_POST_DOMINATORS
, bb
, occ_child
->bb
))
325 occ
->num_divisions
+= occ_child
->num_divisions
;
330 /* Return whether USE_STMT is a floating-point division by DEF. */
332 is_division_by (gimple use_stmt
, tree def
)
334 return is_gimple_assign (use_stmt
)
335 && gimple_assign_rhs_code (use_stmt
) == RDIV_EXPR
336 && gimple_assign_rhs2 (use_stmt
) == def
337 /* Do not recognize x / x as valid division, as we are getting
338 confused later by replacing all immediate uses x in such
340 && gimple_assign_rhs1 (use_stmt
) != def
;
343 /* Walk the subset of the dominator tree rooted at OCC, setting the
344 RECIP_DEF field to a definition of 1.0 / DEF that can be used in
345 the given basic block. The field may be left NULL, of course,
346 if it is not possible or profitable to do the optimization.
348 DEF_BSI is an iterator pointing at the statement defining DEF.
349 If RECIP_DEF is set, a dominator already has a computation that can
353 insert_reciprocals (gimple_stmt_iterator
*def_gsi
, struct occurrence
*occ
,
354 tree def
, tree recip_def
, int threshold
)
358 gimple_stmt_iterator gsi
;
359 struct occurrence
*occ_child
;
362 && (occ
->bb_has_division
|| !flag_trapping_math
)
363 && occ
->num_divisions
>= threshold
)
365 /* Make a variable with the replacement and substitute it. */
366 type
= TREE_TYPE (def
);
367 recip_def
= create_tmp_reg (type
, "reciptmp");
368 new_stmt
= gimple_build_assign_with_ops (RDIV_EXPR
, recip_def
,
369 build_one_cst (type
), def
);
371 if (occ
->bb_has_division
)
373 /* Case 1: insert before an existing division. */
374 gsi
= gsi_after_labels (occ
->bb
);
375 while (!gsi_end_p (gsi
) && !is_division_by (gsi_stmt (gsi
), def
))
378 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
380 else if (def_gsi
&& occ
->bb
== def_gsi
->bb
)
382 /* Case 2: insert right after the definition. Note that this will
383 never happen if the definition statement can throw, because in
384 that case the sole successor of the statement's basic block will
385 dominate all the uses as well. */
386 gsi_insert_after (def_gsi
, new_stmt
, GSI_NEW_STMT
);
390 /* Case 3: insert in a basic block not containing defs/uses. */
391 gsi
= gsi_after_labels (occ
->bb
);
392 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
395 reciprocal_stats
.rdivs_inserted
++;
397 occ
->recip_def_stmt
= new_stmt
;
400 occ
->recip_def
= recip_def
;
401 for (occ_child
= occ
->children
; occ_child
; occ_child
= occ_child
->next
)
402 insert_reciprocals (def_gsi
, occ_child
, def
, recip_def
, threshold
);
406 /* Replace the division at USE_P with a multiplication by the reciprocal, if
410 replace_reciprocal (use_operand_p use_p
)
412 gimple use_stmt
= USE_STMT (use_p
);
413 basic_block bb
= gimple_bb (use_stmt
);
414 struct occurrence
*occ
= (struct occurrence
*) bb
->aux
;
416 if (optimize_bb_for_speed_p (bb
)
417 && occ
->recip_def
&& use_stmt
!= occ
->recip_def_stmt
)
419 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
420 gimple_assign_set_rhs_code (use_stmt
, MULT_EXPR
);
421 SET_USE (use_p
, occ
->recip_def
);
422 fold_stmt_inplace (&gsi
);
423 update_stmt (use_stmt
);
428 /* Free OCC and return one more "struct occurrence" to be freed. */
430 static struct occurrence
*
431 free_bb (struct occurrence
*occ
)
433 struct occurrence
*child
, *next
;
435 /* First get the two pointers hanging off OCC. */
437 child
= occ
->children
;
439 pool_free (occ_pool
, occ
);
441 /* Now ensure that we don't recurse unless it is necessary. */
447 next
= free_bb (next
);
454 /* Look for floating-point divisions among DEF's uses, and try to
455 replace them by multiplications with the reciprocal. Add
456 as many statements computing the reciprocal as needed.
458 DEF must be a GIMPLE register of a floating-point type. */
461 execute_cse_reciprocals_1 (gimple_stmt_iterator
*def_gsi
, tree def
)
464 imm_use_iterator use_iter
;
465 struct occurrence
*occ
;
466 int count
= 0, threshold
;
468 gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def
)) && is_gimple_reg (def
));
470 FOR_EACH_IMM_USE_FAST (use_p
, use_iter
, def
)
472 gimple use_stmt
= USE_STMT (use_p
);
473 if (is_division_by (use_stmt
, def
))
475 register_division_in (gimple_bb (use_stmt
));
480 /* Do the expensive part only if we can hope to optimize something. */
481 threshold
= targetm
.min_divisions_for_recip_mul (TYPE_MODE (TREE_TYPE (def
)));
482 if (count
>= threshold
)
485 for (occ
= occ_head
; occ
; occ
= occ
->next
)
488 insert_reciprocals (def_gsi
, occ
, def
, NULL
, threshold
);
491 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, def
)
493 if (is_division_by (use_stmt
, def
))
495 FOR_EACH_IMM_USE_ON_STMT (use_p
, use_iter
)
496 replace_reciprocal (use_p
);
501 for (occ
= occ_head
; occ
; )
507 /* Go through all the floating-point SSA_NAMEs, and call
508 execute_cse_reciprocals_1 on each of them. */
511 const pass_data pass_data_cse_reciprocals
=
513 GIMPLE_PASS
, /* type */
515 OPTGROUP_NONE
, /* optinfo_flags */
516 true, /* has_execute */
518 PROP_ssa
, /* properties_required */
519 0, /* properties_provided */
520 0, /* properties_destroyed */
521 0, /* todo_flags_start */
522 TODO_update_ssa
, /* todo_flags_finish */
525 class pass_cse_reciprocals
: public gimple_opt_pass
528 pass_cse_reciprocals (gcc::context
*ctxt
)
529 : gimple_opt_pass (pass_data_cse_reciprocals
, ctxt
)
532 /* opt_pass methods: */
533 virtual bool gate (function
*) { return optimize
&& flag_reciprocal_math
; }
534 virtual unsigned int execute (function
*);
536 }; // class pass_cse_reciprocals
539 pass_cse_reciprocals::execute (function
*fun
)
544 occ_pool
= create_alloc_pool ("dominators for recip",
545 sizeof (struct occurrence
),
546 n_basic_blocks_for_fn (fun
) / 3 + 1);
548 memset (&reciprocal_stats
, 0, sizeof (reciprocal_stats
));
549 calculate_dominance_info (CDI_DOMINATORS
);
550 calculate_dominance_info (CDI_POST_DOMINATORS
);
552 #ifdef ENABLE_CHECKING
553 FOR_EACH_BB_FN (bb
, fun
)
554 gcc_assert (!bb
->aux
);
557 for (arg
= DECL_ARGUMENTS (fun
->decl
); arg
; arg
= DECL_CHAIN (arg
))
558 if (FLOAT_TYPE_P (TREE_TYPE (arg
))
559 && is_gimple_reg (arg
))
561 tree name
= ssa_default_def (fun
, arg
);
563 execute_cse_reciprocals_1 (NULL
, name
);
566 FOR_EACH_BB_FN (bb
, fun
)
568 gimple_stmt_iterator gsi
;
572 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
574 phi
= gsi_stmt (gsi
);
575 def
= PHI_RESULT (phi
);
576 if (! virtual_operand_p (def
)
577 && FLOAT_TYPE_P (TREE_TYPE (def
)))
578 execute_cse_reciprocals_1 (NULL
, def
);
581 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
583 gimple stmt
= gsi_stmt (gsi
);
585 if (gimple_has_lhs (stmt
)
586 && (def
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_DEF
)) != NULL
587 && FLOAT_TYPE_P (TREE_TYPE (def
))
588 && TREE_CODE (def
) == SSA_NAME
)
589 execute_cse_reciprocals_1 (&gsi
, def
);
592 if (optimize_bb_for_size_p (bb
))
595 /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */
596 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
598 gimple stmt
= gsi_stmt (gsi
);
601 if (is_gimple_assign (stmt
)
602 && gimple_assign_rhs_code (stmt
) == RDIV_EXPR
)
604 tree arg1
= gimple_assign_rhs2 (stmt
);
607 if (TREE_CODE (arg1
) != SSA_NAME
)
610 stmt1
= SSA_NAME_DEF_STMT (arg1
);
612 if (is_gimple_call (stmt1
)
613 && gimple_call_lhs (stmt1
)
614 && (fndecl
= gimple_call_fndecl (stmt1
))
615 && (DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
616 || DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_MD
))
618 enum built_in_function code
;
623 code
= DECL_FUNCTION_CODE (fndecl
);
624 md_code
= DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_MD
;
626 fndecl
= targetm
.builtin_reciprocal (code
, md_code
, false);
630 /* Check that all uses of the SSA name are divisions,
631 otherwise replacing the defining statement will do
634 FOR_EACH_IMM_USE_FAST (use_p
, ui
, arg1
)
636 gimple stmt2
= USE_STMT (use_p
);
637 if (is_gimple_debug (stmt2
))
639 if (!is_gimple_assign (stmt2
)
640 || gimple_assign_rhs_code (stmt2
) != RDIV_EXPR
641 || gimple_assign_rhs1 (stmt2
) == arg1
642 || gimple_assign_rhs2 (stmt2
) != arg1
)
651 gimple_replace_ssa_lhs (stmt1
, arg1
);
652 gimple_call_set_fndecl (stmt1
, fndecl
);
654 reciprocal_stats
.rfuncs_inserted
++;
656 FOR_EACH_IMM_USE_STMT (stmt
, ui
, arg1
)
658 gimple_stmt_iterator gsi
= gsi_for_stmt (stmt
);
659 gimple_assign_set_rhs_code (stmt
, MULT_EXPR
);
660 fold_stmt_inplace (&gsi
);
668 statistics_counter_event (fun
, "reciprocal divs inserted",
669 reciprocal_stats
.rdivs_inserted
);
670 statistics_counter_event (fun
, "reciprocal functions inserted",
671 reciprocal_stats
.rfuncs_inserted
);
673 free_dominance_info (CDI_DOMINATORS
);
674 free_dominance_info (CDI_POST_DOMINATORS
);
675 free_alloc_pool (occ_pool
);
682 make_pass_cse_reciprocals (gcc::context
*ctxt
)
684 return new pass_cse_reciprocals (ctxt
);
687 /* Records an occurrence at statement USE_STMT in the vector of trees
688 STMTS if it is dominated by *TOP_BB or dominates it or this basic block
689 is not yet initialized. Returns true if the occurrence was pushed on
690 the vector. Adjusts *TOP_BB to be the basic block dominating all
691 statements in the vector. */
694 maybe_record_sincos (vec
<gimple
> *stmts
,
695 basic_block
*top_bb
, gimple use_stmt
)
697 basic_block use_bb
= gimple_bb (use_stmt
);
699 && (*top_bb
== use_bb
700 || dominated_by_p (CDI_DOMINATORS
, use_bb
, *top_bb
)))
701 stmts
->safe_push (use_stmt
);
703 || dominated_by_p (CDI_DOMINATORS
, *top_bb
, use_bb
))
705 stmts
->safe_push (use_stmt
);
714 /* Look for sin, cos and cexpi calls with the same argument NAME and
715 create a single call to cexpi CSEing the result in this case.
716 We first walk over all immediate uses of the argument collecting
717 statements that we can CSE in a vector and in a second pass replace
718 the statement rhs with a REALPART or IMAGPART expression on the
719 result of the cexpi call we insert before the use statement that
720 dominates all other candidates. */
723 execute_cse_sincos_1 (tree name
)
725 gimple_stmt_iterator gsi
;
726 imm_use_iterator use_iter
;
727 tree fndecl
, res
, type
;
728 gimple def_stmt
, use_stmt
, stmt
;
729 int seen_cos
= 0, seen_sin
= 0, seen_cexpi
= 0;
730 vec
<gimple
> stmts
= vNULL
;
731 basic_block top_bb
= NULL
;
733 bool cfg_changed
= false;
735 type
= TREE_TYPE (name
);
736 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, name
)
738 if (gimple_code (use_stmt
) != GIMPLE_CALL
739 || !gimple_call_lhs (use_stmt
)
740 || !(fndecl
= gimple_call_fndecl (use_stmt
))
741 || DECL_BUILT_IN_CLASS (fndecl
) != BUILT_IN_NORMAL
)
744 switch (DECL_FUNCTION_CODE (fndecl
))
746 CASE_FLT_FN (BUILT_IN_COS
):
747 seen_cos
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
750 CASE_FLT_FN (BUILT_IN_SIN
):
751 seen_sin
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
754 CASE_FLT_FN (BUILT_IN_CEXPI
):
755 seen_cexpi
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
762 if (seen_cos
+ seen_sin
+ seen_cexpi
<= 1)
768 /* Simply insert cexpi at the beginning of top_bb but not earlier than
769 the name def statement. */
770 fndecl
= mathfn_built_in (type
, BUILT_IN_CEXPI
);
773 stmt
= gimple_build_call (fndecl
, 1, name
);
774 res
= make_temp_ssa_name (TREE_TYPE (TREE_TYPE (fndecl
)), stmt
, "sincostmp");
775 gimple_call_set_lhs (stmt
, res
);
777 def_stmt
= SSA_NAME_DEF_STMT (name
);
778 if (!SSA_NAME_IS_DEFAULT_DEF (name
)
779 && gimple_code (def_stmt
) != GIMPLE_PHI
780 && gimple_bb (def_stmt
) == top_bb
)
782 gsi
= gsi_for_stmt (def_stmt
);
783 gsi_insert_after (&gsi
, stmt
, GSI_SAME_STMT
);
787 gsi
= gsi_after_labels (top_bb
);
788 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
790 sincos_stats
.inserted
++;
792 /* And adjust the recorded old call sites. */
793 for (i
= 0; stmts
.iterate (i
, &use_stmt
); ++i
)
796 fndecl
= gimple_call_fndecl (use_stmt
);
798 switch (DECL_FUNCTION_CODE (fndecl
))
800 CASE_FLT_FN (BUILT_IN_COS
):
801 rhs
= fold_build1 (REALPART_EXPR
, type
, res
);
804 CASE_FLT_FN (BUILT_IN_SIN
):
805 rhs
= fold_build1 (IMAGPART_EXPR
, type
, res
);
808 CASE_FLT_FN (BUILT_IN_CEXPI
):
816 /* Replace call with a copy. */
817 stmt
= gimple_build_assign (gimple_call_lhs (use_stmt
), rhs
);
819 gsi
= gsi_for_stmt (use_stmt
);
820 gsi_replace (&gsi
, stmt
, true);
821 if (gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
830 /* To evaluate powi(x,n), the floating point value x raised to the
831 constant integer exponent n, we use a hybrid algorithm that
832 combines the "window method" with look-up tables. For an
833 introduction to exponentiation algorithms and "addition chains",
834 see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth,
835 "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming",
836 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation
837 Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */
839 /* Provide a default value for POWI_MAX_MULTS, the maximum number of
840 multiplications to inline before calling the system library's pow
841 function. powi(x,n) requires at worst 2*bits(n)-2 multiplications,
842 so this default never requires calling pow, powf or powl. */
844 #ifndef POWI_MAX_MULTS
845 #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2)
848 /* The size of the "optimal power tree" lookup table. All
849 exponents less than this value are simply looked up in the
850 powi_table below. This threshold is also used to size the
851 cache of pseudo registers that hold intermediate results. */
852 #define POWI_TABLE_SIZE 256
854 /* The size, in bits of the window, used in the "window method"
855 exponentiation algorithm. This is equivalent to a radix of
856 (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */
857 #define POWI_WINDOW_SIZE 3
859 /* The following table is an efficient representation of an
860 "optimal power tree". For each value, i, the corresponding
861 value, j, in the table states than an optimal evaluation
862 sequence for calculating pow(x,i) can be found by evaluating
863 pow(x,j)*pow(x,i-j). An optimal power tree for the first
864 100 integers is given in Knuth's "Seminumerical algorithms". */
866 static const unsigned char powi_table
[POWI_TABLE_SIZE
] =
868 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */
869 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */
870 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */
871 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */
872 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */
873 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */
874 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */
875 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */
876 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */
877 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */
878 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */
879 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */
880 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */
881 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */
882 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */
883 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */
884 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */
885 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */
886 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */
887 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */
888 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */
889 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */
890 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */
891 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */
892 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */
893 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */
894 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */
895 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */
896 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */
897 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */
898 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */
899 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */
903 /* Return the number of multiplications required to calculate
904 powi(x,n) where n is less than POWI_TABLE_SIZE. This is a
905 subroutine of powi_cost. CACHE is an array indicating
906 which exponents have already been calculated. */
909 powi_lookup_cost (unsigned HOST_WIDE_INT n
, bool *cache
)
911 /* If we've already calculated this exponent, then this evaluation
912 doesn't require any additional multiplications. */
917 return powi_lookup_cost (n
- powi_table
[n
], cache
)
918 + powi_lookup_cost (powi_table
[n
], cache
) + 1;
921 /* Return the number of multiplications required to calculate
922 powi(x,n) for an arbitrary x, given the exponent N. This
923 function needs to be kept in sync with powi_as_mults below. */
926 powi_cost (HOST_WIDE_INT n
)
928 bool cache
[POWI_TABLE_SIZE
];
929 unsigned HOST_WIDE_INT digit
;
930 unsigned HOST_WIDE_INT val
;
936 /* Ignore the reciprocal when calculating the cost. */
937 val
= (n
< 0) ? -n
: n
;
939 /* Initialize the exponent cache. */
940 memset (cache
, 0, POWI_TABLE_SIZE
* sizeof (bool));
945 while (val
>= POWI_TABLE_SIZE
)
949 digit
= val
& ((1 << POWI_WINDOW_SIZE
) - 1);
950 result
+= powi_lookup_cost (digit
, cache
)
951 + POWI_WINDOW_SIZE
+ 1;
952 val
>>= POWI_WINDOW_SIZE
;
961 return result
+ powi_lookup_cost (val
, cache
);
964 /* Recursive subroutine of powi_as_mults. This function takes the
965 array, CACHE, of already calculated exponents and an exponent N and
966 returns a tree that corresponds to CACHE[1]**N, with type TYPE. */
969 powi_as_mults_1 (gimple_stmt_iterator
*gsi
, location_t loc
, tree type
,
970 HOST_WIDE_INT n
, tree
*cache
)
972 tree op0
, op1
, ssa_target
;
973 unsigned HOST_WIDE_INT digit
;
976 if (n
< POWI_TABLE_SIZE
&& cache
[n
])
979 ssa_target
= make_temp_ssa_name (type
, NULL
, "powmult");
981 if (n
< POWI_TABLE_SIZE
)
983 cache
[n
] = ssa_target
;
984 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
- powi_table
[n
], cache
);
985 op1
= powi_as_mults_1 (gsi
, loc
, type
, powi_table
[n
], cache
);
989 digit
= n
& ((1 << POWI_WINDOW_SIZE
) - 1);
990 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
- digit
, cache
);
991 op1
= powi_as_mults_1 (gsi
, loc
, type
, digit
, cache
);
995 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
>> 1, cache
);
999 mult_stmt
= gimple_build_assign_with_ops (MULT_EXPR
, ssa_target
, op0
, op1
);
1000 gimple_set_location (mult_stmt
, loc
);
1001 gsi_insert_before (gsi
, mult_stmt
, GSI_SAME_STMT
);
1006 /* Convert ARG0**N to a tree of multiplications of ARG0 with itself.
1007 This function needs to be kept in sync with powi_cost above. */
1010 powi_as_mults (gimple_stmt_iterator
*gsi
, location_t loc
,
1011 tree arg0
, HOST_WIDE_INT n
)
1013 tree cache
[POWI_TABLE_SIZE
], result
, type
= TREE_TYPE (arg0
);
1018 return build_real (type
, dconst1
);
1020 memset (cache
, 0, sizeof (cache
));
1023 result
= powi_as_mults_1 (gsi
, loc
, type
, (n
< 0) ? -n
: n
, cache
);
1027 /* If the original exponent was negative, reciprocate the result. */
1028 target
= make_temp_ssa_name (type
, NULL
, "powmult");
1029 div_stmt
= gimple_build_assign_with_ops (RDIV_EXPR
, target
,
1030 build_real (type
, dconst1
),
1032 gimple_set_location (div_stmt
, loc
);
1033 gsi_insert_before (gsi
, div_stmt
, GSI_SAME_STMT
);
1038 /* ARG0 and N are the two arguments to a powi builtin in GSI with
1039 location info LOC. If the arguments are appropriate, create an
1040 equivalent sequence of statements prior to GSI using an optimal
1041 number of multiplications, and return an expession holding the
1045 gimple_expand_builtin_powi (gimple_stmt_iterator
*gsi
, location_t loc
,
1046 tree arg0
, HOST_WIDE_INT n
)
1048 /* Avoid largest negative number. */
1050 && ((n
>= -1 && n
<= 2)
1051 || (optimize_function_for_speed_p (cfun
)
1052 && powi_cost (n
) <= POWI_MAX_MULTS
)))
1053 return powi_as_mults (gsi
, loc
, arg0
, n
);
1058 /* Build a gimple call statement that calls FN with argument ARG.
1059 Set the lhs of the call statement to a fresh SSA name. Insert the
1060 statement prior to GSI's current position, and return the fresh
1064 build_and_insert_call (gimple_stmt_iterator
*gsi
, location_t loc
,
1070 call_stmt
= gimple_build_call (fn
, 1, arg
);
1071 ssa_target
= make_temp_ssa_name (TREE_TYPE (arg
), NULL
, "powroot");
1072 gimple_set_lhs (call_stmt
, ssa_target
);
1073 gimple_set_location (call_stmt
, loc
);
1074 gsi_insert_before (gsi
, call_stmt
, GSI_SAME_STMT
);
1079 /* Build a gimple binary operation with the given CODE and arguments
1080 ARG0, ARG1, assigning the result to a new SSA name for variable
1081 TARGET. Insert the statement prior to GSI's current position, and
1082 return the fresh SSA name.*/
1085 build_and_insert_binop (gimple_stmt_iterator
*gsi
, location_t loc
,
1086 const char *name
, enum tree_code code
,
1087 tree arg0
, tree arg1
)
1089 tree result
= make_temp_ssa_name (TREE_TYPE (arg0
), NULL
, name
);
1090 gimple stmt
= gimple_build_assign_with_ops (code
, result
, arg0
, arg1
);
1091 gimple_set_location (stmt
, loc
);
1092 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1096 /* Build a gimple reference operation with the given CODE and argument
1097 ARG, assigning the result to a new SSA name of TYPE with NAME.
1098 Insert the statement prior to GSI's current position, and return
1099 the fresh SSA name. */
1102 build_and_insert_ref (gimple_stmt_iterator
*gsi
, location_t loc
, tree type
,
1103 const char *name
, enum tree_code code
, tree arg0
)
1105 tree result
= make_temp_ssa_name (type
, NULL
, name
);
1106 gimple stmt
= gimple_build_assign (result
, build1 (code
, type
, arg0
));
1107 gimple_set_location (stmt
, loc
);
1108 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1112 /* Build a gimple assignment to cast VAL to TYPE. Insert the statement
1113 prior to GSI's current position, and return the fresh SSA name. */
1116 build_and_insert_cast (gimple_stmt_iterator
*gsi
, location_t loc
,
1117 tree type
, tree val
)
1119 tree result
= make_ssa_name (type
, NULL
);
1120 gimple stmt
= gimple_build_assign_with_ops (NOP_EXPR
, result
, val
, NULL_TREE
);
1121 gimple_set_location (stmt
, loc
);
1122 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1126 /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI
1127 with location info LOC. If possible, create an equivalent and
1128 less expensive sequence of statements prior to GSI, and return an
1129 expession holding the result. */
1132 gimple_expand_builtin_pow (gimple_stmt_iterator
*gsi
, location_t loc
,
1133 tree arg0
, tree arg1
)
1135 REAL_VALUE_TYPE c
, cint
, dconst1_4
, dconst3_4
, dconst1_3
, dconst1_6
;
1136 REAL_VALUE_TYPE c2
, dconst3
;
1138 tree type
, sqrtfn
, cbrtfn
, sqrt_arg0
, sqrt_sqrt
, result
, cbrt_x
, powi_cbrt_x
;
1139 enum machine_mode mode
;
1140 bool hw_sqrt_exists
, c_is_int
, c2_is_int
;
1142 /* If the exponent isn't a constant, there's nothing of interest
1144 if (TREE_CODE (arg1
) != REAL_CST
)
1147 /* If the exponent is equivalent to an integer, expand to an optimal
1148 multiplication sequence when profitable. */
1149 c
= TREE_REAL_CST (arg1
);
1150 n
= real_to_integer (&c
);
1151 real_from_integer (&cint
, VOIDmode
, n
, n
< 0 ? -1 : 0, 0);
1152 c_is_int
= real_identical (&c
, &cint
);
1155 && ((n
>= -1 && n
<= 2)
1156 || (flag_unsafe_math_optimizations
1157 && optimize_bb_for_speed_p (gsi_bb (*gsi
))
1158 && powi_cost (n
) <= POWI_MAX_MULTS
)))
1159 return gimple_expand_builtin_powi (gsi
, loc
, arg0
, n
);
1161 /* Attempt various optimizations using sqrt and cbrt. */
1162 type
= TREE_TYPE (arg0
);
1163 mode
= TYPE_MODE (type
);
1164 sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1166 /* Optimize pow(x,0.5) = sqrt(x). This replacement is always safe
1167 unless signed zeros must be maintained. pow(-0,0.5) = +0, while
1170 && REAL_VALUES_EQUAL (c
, dconsthalf
)
1171 && !HONOR_SIGNED_ZEROS (mode
))
1172 return build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1174 /* Optimize pow(x,0.25) = sqrt(sqrt(x)). Assume on most machines that
1175 a builtin sqrt instruction is smaller than a call to pow with 0.25,
1176 so do this optimization even if -Os. Don't do this optimization
1177 if we don't have a hardware sqrt insn. */
1178 dconst1_4
= dconst1
;
1179 SET_REAL_EXP (&dconst1_4
, REAL_EXP (&dconst1_4
) - 2);
1180 hw_sqrt_exists
= optab_handler (sqrt_optab
, mode
) != CODE_FOR_nothing
;
1182 if (flag_unsafe_math_optimizations
1184 && REAL_VALUES_EQUAL (c
, dconst1_4
)
1188 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1191 return build_and_insert_call (gsi
, loc
, sqrtfn
, sqrt_arg0
);
1194 /* Optimize pow(x,0.75) = sqrt(x) * sqrt(sqrt(x)) unless we are
1195 optimizing for space. Don't do this optimization if we don't have
1196 a hardware sqrt insn. */
1197 real_from_integer (&dconst3_4
, VOIDmode
, 3, 0, 0);
1198 SET_REAL_EXP (&dconst3_4
, REAL_EXP (&dconst3_4
) - 2);
1200 if (flag_unsafe_math_optimizations
1202 && optimize_function_for_speed_p (cfun
)
1203 && REAL_VALUES_EQUAL (c
, dconst3_4
)
1207 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1210 sqrt_sqrt
= build_and_insert_call (gsi
, loc
, sqrtfn
, sqrt_arg0
);
1212 /* sqrt(x) * sqrt(sqrt(x)) */
1213 return build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1214 sqrt_arg0
, sqrt_sqrt
);
1217 /* Optimize pow(x,1./3.) = cbrt(x). This requires unsafe math
1218 optimizations since 1./3. is not exactly representable. If x
1219 is negative and finite, the correct value of pow(x,1./3.) is
1220 a NaN with the "invalid" exception raised, because the value
1221 of 1./3. actually has an even denominator. The correct value
1222 of cbrt(x) is a negative real value. */
1223 cbrtfn
= mathfn_built_in (type
, BUILT_IN_CBRT
);
1224 dconst1_3
= real_value_truncate (mode
, dconst_third ());
1226 if (flag_unsafe_math_optimizations
1228 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1229 && REAL_VALUES_EQUAL (c
, dconst1_3
))
1230 return build_and_insert_call (gsi
, loc
, cbrtfn
, arg0
);
1232 /* Optimize pow(x,1./6.) = cbrt(sqrt(x)). Don't do this optimization
1233 if we don't have a hardware sqrt insn. */
1234 dconst1_6
= dconst1_3
;
1235 SET_REAL_EXP (&dconst1_6
, REAL_EXP (&dconst1_6
) - 1);
1237 if (flag_unsafe_math_optimizations
1240 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1241 && optimize_function_for_speed_p (cfun
)
1243 && REAL_VALUES_EQUAL (c
, dconst1_6
))
1246 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1249 return build_and_insert_call (gsi
, loc
, cbrtfn
, sqrt_arg0
);
1252 /* Optimize pow(x,c), where n = 2c for some nonzero integer n
1253 and c not an integer, into
1255 sqrt(x) * powi(x, n/2), n > 0;
1256 1.0 / (sqrt(x) * powi(x, abs(n/2))), n < 0.
1258 Do not calculate the powi factor when n/2 = 0. */
1259 real_arithmetic (&c2
, MULT_EXPR
, &c
, &dconst2
);
1260 n
= real_to_integer (&c2
);
1261 real_from_integer (&cint
, VOIDmode
, n
, n
< 0 ? -1 : 0, 0);
1262 c2_is_int
= real_identical (&c2
, &cint
);
1264 if (flag_unsafe_math_optimizations
1268 && optimize_function_for_speed_p (cfun
))
1270 tree powi_x_ndiv2
= NULL_TREE
;
1272 /* Attempt to fold powi(arg0, abs(n/2)) into multiplies. If not
1273 possible or profitable, give up. Skip the degenerate case when
1274 n is 1 or -1, where the result is always 1. */
1275 if (absu_hwi (n
) != 1)
1277 powi_x_ndiv2
= gimple_expand_builtin_powi (gsi
, loc
, arg0
,
1283 /* Calculate sqrt(x). When n is not 1 or -1, multiply it by the
1284 result of the optimal multiply sequence just calculated. */
1285 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1287 if (absu_hwi (n
) == 1)
1290 result
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1291 sqrt_arg0
, powi_x_ndiv2
);
1293 /* If n is negative, reciprocate the result. */
1295 result
= build_and_insert_binop (gsi
, loc
, "powroot", RDIV_EXPR
,
1296 build_real (type
, dconst1
), result
);
1300 /* Optimize pow(x,c), where 3c = n for some nonzero integer n, into
1302 powi(x, n/3) * powi(cbrt(x), n%3), n > 0;
1303 1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)), n < 0.
1305 Do not calculate the first factor when n/3 = 0. As cbrt(x) is
1306 different from pow(x, 1./3.) due to rounding and behavior with
1307 negative x, we need to constrain this transformation to unsafe
1308 math and positive x or finite math. */
1309 real_from_integer (&dconst3
, VOIDmode
, 3, 0, 0);
1310 real_arithmetic (&c2
, MULT_EXPR
, &c
, &dconst3
);
1311 real_round (&c2
, mode
, &c2
);
1312 n
= real_to_integer (&c2
);
1313 real_from_integer (&cint
, VOIDmode
, n
, n
< 0 ? -1 : 0, 0);
1314 real_arithmetic (&c2
, RDIV_EXPR
, &cint
, &dconst3
);
1315 real_convert (&c2
, mode
, &c2
);
1317 if (flag_unsafe_math_optimizations
1319 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1320 && real_identical (&c2
, &c
)
1322 && optimize_function_for_speed_p (cfun
)
1323 && powi_cost (n
/ 3) <= POWI_MAX_MULTS
)
1325 tree powi_x_ndiv3
= NULL_TREE
;
1327 /* Attempt to fold powi(arg0, abs(n/3)) into multiplies. If not
1328 possible or profitable, give up. Skip the degenerate case when
1329 abs(n) < 3, where the result is always 1. */
1330 if (absu_hwi (n
) >= 3)
1332 powi_x_ndiv3
= gimple_expand_builtin_powi (gsi
, loc
, arg0
,
1338 /* Calculate powi(cbrt(x), n%3). Don't use gimple_expand_builtin_powi
1339 as that creates an unnecessary variable. Instead, just produce
1340 either cbrt(x) or cbrt(x) * cbrt(x). */
1341 cbrt_x
= build_and_insert_call (gsi
, loc
, cbrtfn
, arg0
);
1343 if (absu_hwi (n
) % 3 == 1)
1344 powi_cbrt_x
= cbrt_x
;
1346 powi_cbrt_x
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1349 /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1. */
1350 if (absu_hwi (n
) < 3)
1351 result
= powi_cbrt_x
;
1353 result
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1354 powi_x_ndiv3
, powi_cbrt_x
);
1356 /* If n is negative, reciprocate the result. */
1358 result
= build_and_insert_binop (gsi
, loc
, "powroot", RDIV_EXPR
,
1359 build_real (type
, dconst1
), result
);
1364 /* No optimizations succeeded. */
1368 /* ARG is the argument to a cabs builtin call in GSI with location info
1369 LOC. Create a sequence of statements prior to GSI that calculates
1370 sqrt(R*R + I*I), where R and I are the real and imaginary components
1371 of ARG, respectively. Return an expression holding the result. */
1374 gimple_expand_builtin_cabs (gimple_stmt_iterator
*gsi
, location_t loc
, tree arg
)
1376 tree real_part
, imag_part
, addend1
, addend2
, sum
, result
;
1377 tree type
= TREE_TYPE (TREE_TYPE (arg
));
1378 tree sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1379 enum machine_mode mode
= TYPE_MODE (type
);
1381 if (!flag_unsafe_math_optimizations
1382 || !optimize_bb_for_speed_p (gimple_bb (gsi_stmt (*gsi
)))
1384 || optab_handler (sqrt_optab
, mode
) == CODE_FOR_nothing
)
1387 real_part
= build_and_insert_ref (gsi
, loc
, type
, "cabs",
1388 REALPART_EXPR
, arg
);
1389 addend1
= build_and_insert_binop (gsi
, loc
, "cabs", MULT_EXPR
,
1390 real_part
, real_part
);
1391 imag_part
= build_and_insert_ref (gsi
, loc
, type
, "cabs",
1392 IMAGPART_EXPR
, arg
);
1393 addend2
= build_and_insert_binop (gsi
, loc
, "cabs", MULT_EXPR
,
1394 imag_part
, imag_part
);
1395 sum
= build_and_insert_binop (gsi
, loc
, "cabs", PLUS_EXPR
, addend1
, addend2
);
1396 result
= build_and_insert_call (gsi
, loc
, sqrtfn
, sum
);
1401 /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1
1402 on the SSA_NAME argument of each of them. Also expand powi(x,n) into
1403 an optimal number of multiplies, when n is a constant. */
1407 const pass_data pass_data_cse_sincos
=
1409 GIMPLE_PASS
, /* type */
1410 "sincos", /* name */
1411 OPTGROUP_NONE
, /* optinfo_flags */
1412 true, /* has_execute */
1413 TV_NONE
, /* tv_id */
1414 PROP_ssa
, /* properties_required */
1415 0, /* properties_provided */
1416 0, /* properties_destroyed */
1417 0, /* todo_flags_start */
1418 TODO_update_ssa
, /* todo_flags_finish */
1421 class pass_cse_sincos
: public gimple_opt_pass
1424 pass_cse_sincos (gcc::context
*ctxt
)
1425 : gimple_opt_pass (pass_data_cse_sincos
, ctxt
)
1428 /* opt_pass methods: */
1429 virtual bool gate (function
*)
1431 /* We no longer require either sincos or cexp, since powi expansion
1432 piggybacks on this pass. */
1436 virtual unsigned int execute (function
*);
1438 }; // class pass_cse_sincos
1441 pass_cse_sincos::execute (function
*fun
)
1444 bool cfg_changed
= false;
1446 calculate_dominance_info (CDI_DOMINATORS
);
1447 memset (&sincos_stats
, 0, sizeof (sincos_stats
));
1449 FOR_EACH_BB_FN (bb
, fun
)
1451 gimple_stmt_iterator gsi
;
1452 bool cleanup_eh
= false;
1454 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1456 gimple stmt
= gsi_stmt (gsi
);
1459 /* Only the last stmt in a bb could throw, no need to call
1460 gimple_purge_dead_eh_edges if we change something in the middle
1461 of a basic block. */
1464 if (is_gimple_call (stmt
)
1465 && gimple_call_lhs (stmt
)
1466 && (fndecl
= gimple_call_fndecl (stmt
))
1467 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
)
1469 tree arg
, arg0
, arg1
, result
;
1473 switch (DECL_FUNCTION_CODE (fndecl
))
1475 CASE_FLT_FN (BUILT_IN_COS
):
1476 CASE_FLT_FN (BUILT_IN_SIN
):
1477 CASE_FLT_FN (BUILT_IN_CEXPI
):
1478 /* Make sure we have either sincos or cexp. */
1479 if (!targetm
.libc_has_function (function_c99_math_complex
)
1480 && !targetm
.libc_has_function (function_sincos
))
1483 arg
= gimple_call_arg (stmt
, 0);
1484 if (TREE_CODE (arg
) == SSA_NAME
)
1485 cfg_changed
|= execute_cse_sincos_1 (arg
);
1488 CASE_FLT_FN (BUILT_IN_POW
):
1489 arg0
= gimple_call_arg (stmt
, 0);
1490 arg1
= gimple_call_arg (stmt
, 1);
1492 loc
= gimple_location (stmt
);
1493 result
= gimple_expand_builtin_pow (&gsi
, loc
, arg0
, arg1
);
1497 tree lhs
= gimple_get_lhs (stmt
);
1498 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1499 gimple_set_location (new_stmt
, loc
);
1500 unlink_stmt_vdef (stmt
);
1501 gsi_replace (&gsi
, new_stmt
, true);
1503 if (gimple_vdef (stmt
))
1504 release_ssa_name (gimple_vdef (stmt
));
1508 CASE_FLT_FN (BUILT_IN_POWI
):
1509 arg0
= gimple_call_arg (stmt
, 0);
1510 arg1
= gimple_call_arg (stmt
, 1);
1511 loc
= gimple_location (stmt
);
1513 if (real_minus_onep (arg0
))
1515 tree t0
, t1
, cond
, one
, minus_one
;
1518 t0
= TREE_TYPE (arg0
);
1519 t1
= TREE_TYPE (arg1
);
1520 one
= build_real (t0
, dconst1
);
1521 minus_one
= build_real (t0
, dconstm1
);
1523 cond
= make_temp_ssa_name (t1
, NULL
, "powi_cond");
1524 stmt
= gimple_build_assign_with_ops (BIT_AND_EXPR
, cond
,
1528 gimple_set_location (stmt
, loc
);
1529 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1531 result
= make_temp_ssa_name (t0
, NULL
, "powi");
1532 stmt
= gimple_build_assign_with_ops (COND_EXPR
, result
,
1535 gimple_set_location (stmt
, loc
);
1536 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1540 if (!tree_fits_shwi_p (arg1
))
1543 n
= tree_to_shwi (arg1
);
1544 result
= gimple_expand_builtin_powi (&gsi
, loc
, arg0
, n
);
1549 tree lhs
= gimple_get_lhs (stmt
);
1550 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1551 gimple_set_location (new_stmt
, loc
);
1552 unlink_stmt_vdef (stmt
);
1553 gsi_replace (&gsi
, new_stmt
, true);
1555 if (gimple_vdef (stmt
))
1556 release_ssa_name (gimple_vdef (stmt
));
1560 CASE_FLT_FN (BUILT_IN_CABS
):
1561 arg0
= gimple_call_arg (stmt
, 0);
1562 loc
= gimple_location (stmt
);
1563 result
= gimple_expand_builtin_cabs (&gsi
, loc
, arg0
);
1567 tree lhs
= gimple_get_lhs (stmt
);
1568 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1569 gimple_set_location (new_stmt
, loc
);
1570 unlink_stmt_vdef (stmt
);
1571 gsi_replace (&gsi
, new_stmt
, true);
1573 if (gimple_vdef (stmt
))
1574 release_ssa_name (gimple_vdef (stmt
));
1583 cfg_changed
|= gimple_purge_dead_eh_edges (bb
);
1586 statistics_counter_event (fun
, "sincos statements inserted",
1587 sincos_stats
.inserted
);
1589 free_dominance_info (CDI_DOMINATORS
);
1590 return cfg_changed
? TODO_cleanup_cfg
: 0;
1596 make_pass_cse_sincos (gcc::context
*ctxt
)
1598 return new pass_cse_sincos (ctxt
);
1601 /* A symbolic number is used to detect byte permutation and selection
1602 patterns. Therefore the field N contains an artificial number
1603 consisting of byte size markers:
1605 0 - byte has the value 0
1606 1..size - byte contains the content of the byte
1607 number indexed with that value minus one */
1609 struct symbolic_number
{
1610 unsigned HOST_WIDEST_INT n
;
1614 /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic
1615 number N. Return false if the requested operation is not permitted
1616 on a symbolic number. */
1619 do_shift_rotate (enum tree_code code
,
1620 struct symbolic_number
*n
,
1626 /* Zero out the extra bits of N in order to avoid them being shifted
1627 into the significant bits. */
1628 if (n
->size
< (int)sizeof (HOST_WIDEST_INT
))
1629 n
->n
&= ((unsigned HOST_WIDEST_INT
)1 << (n
->size
* BITS_PER_UNIT
)) - 1;
1640 n
->n
= (n
->n
<< count
) | (n
->n
>> ((n
->size
* BITS_PER_UNIT
) - count
));
1643 n
->n
= (n
->n
>> count
) | (n
->n
<< ((n
->size
* BITS_PER_UNIT
) - count
));
1648 /* Zero unused bits for size. */
1649 if (n
->size
< (int)sizeof (HOST_WIDEST_INT
))
1650 n
->n
&= ((unsigned HOST_WIDEST_INT
)1 << (n
->size
* BITS_PER_UNIT
)) - 1;
1654 /* Perform sanity checking for the symbolic number N and the gimple
1658 verify_symbolic_number_p (struct symbolic_number
*n
, gimple stmt
)
1662 lhs_type
= gimple_expr_type (stmt
);
1664 if (TREE_CODE (lhs_type
) != INTEGER_TYPE
)
1667 if (TYPE_PRECISION (lhs_type
) != n
->size
* BITS_PER_UNIT
)
1673 /* find_bswap_1 invokes itself recursively with N and tries to perform
1674 the operation given by the rhs of STMT on the result. If the
1675 operation could successfully be executed the function returns the
1676 tree expression of the source operand and NULL otherwise. */
1679 find_bswap_1 (gimple stmt
, struct symbolic_number
*n
, int limit
)
1681 enum tree_code code
;
1682 tree rhs1
, rhs2
= NULL
;
1683 gimple rhs1_stmt
, rhs2_stmt
;
1685 enum gimple_rhs_class rhs_class
;
1687 if (!limit
|| !is_gimple_assign (stmt
))
1690 rhs1
= gimple_assign_rhs1 (stmt
);
1692 if (TREE_CODE (rhs1
) != SSA_NAME
)
1695 code
= gimple_assign_rhs_code (stmt
);
1696 rhs_class
= gimple_assign_rhs_class (stmt
);
1697 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
1699 if (rhs_class
== GIMPLE_BINARY_RHS
)
1700 rhs2
= gimple_assign_rhs2 (stmt
);
1702 /* Handle unary rhs and binary rhs with integer constants as second
1705 if (rhs_class
== GIMPLE_UNARY_RHS
1706 || (rhs_class
== GIMPLE_BINARY_RHS
1707 && TREE_CODE (rhs2
) == INTEGER_CST
))
1709 if (code
!= BIT_AND_EXPR
1710 && code
!= LSHIFT_EXPR
1711 && code
!= RSHIFT_EXPR
1712 && code
!= LROTATE_EXPR
1713 && code
!= RROTATE_EXPR
1715 && code
!= CONVERT_EXPR
)
1718 source_expr1
= find_bswap_1 (rhs1_stmt
, n
, limit
- 1);
1720 /* If find_bswap_1 returned NULL STMT is a leaf node and we have
1721 to initialize the symbolic number. */
1724 /* Set up the symbolic number N by setting each byte to a
1725 value between 1 and the byte size of rhs1. The highest
1726 order byte is set to n->size and the lowest order
1728 n
->size
= TYPE_PRECISION (TREE_TYPE (rhs1
));
1729 if (n
->size
% BITS_PER_UNIT
!= 0)
1731 n
->size
/= BITS_PER_UNIT
;
1732 n
->n
= (sizeof (HOST_WIDEST_INT
) < 8 ? 0 :
1733 (unsigned HOST_WIDEST_INT
)0x08070605 << 32 | 0x04030201);
1735 if (n
->size
< (int)sizeof (HOST_WIDEST_INT
))
1736 n
->n
&= ((unsigned HOST_WIDEST_INT
)1 <<
1737 (n
->size
* BITS_PER_UNIT
)) - 1;
1739 source_expr1
= rhs1
;
1747 unsigned HOST_WIDEST_INT val
= widest_int_cst_value (rhs2
);
1748 unsigned HOST_WIDEST_INT tmp
= val
;
1750 /* Only constants masking full bytes are allowed. */
1751 for (i
= 0; i
< n
->size
; i
++, tmp
>>= BITS_PER_UNIT
)
1752 if ((tmp
& 0xff) != 0 && (tmp
& 0xff) != 0xff)
1762 if (!do_shift_rotate (code
, n
, (int)TREE_INT_CST_LOW (rhs2
)))
1769 type_size
= TYPE_PRECISION (gimple_expr_type (stmt
));
1770 if (type_size
% BITS_PER_UNIT
!= 0)
1773 if (type_size
/ BITS_PER_UNIT
< (int)(sizeof (HOST_WIDEST_INT
)))
1775 /* If STMT casts to a smaller type mask out the bits not
1776 belonging to the target type. */
1777 n
->n
&= ((unsigned HOST_WIDEST_INT
)1 << type_size
) - 1;
1779 n
->size
= type_size
/ BITS_PER_UNIT
;
1785 return verify_symbolic_number_p (n
, stmt
) ? source_expr1
: NULL
;
1788 /* Handle binary rhs. */
1790 if (rhs_class
== GIMPLE_BINARY_RHS
)
1793 struct symbolic_number n1
, n2
;
1794 unsigned HOST_WIDEST_INT mask
;
1797 if (code
!= BIT_IOR_EXPR
)
1800 if (TREE_CODE (rhs2
) != SSA_NAME
)
1803 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
1808 source_expr1
= find_bswap_1 (rhs1_stmt
, &n1
, limit
- 1);
1813 source_expr2
= find_bswap_1 (rhs2_stmt
, &n2
, limit
- 1);
1815 if (source_expr1
!= source_expr2
1816 || n1
.size
!= n2
.size
)
1820 for (i
= 0, mask
= 0xff; i
< n
->size
; i
++, mask
<<= BITS_PER_UNIT
)
1822 unsigned HOST_WIDEST_INT masked1
, masked2
;
1824 masked1
= n1
.n
& mask
;
1825 masked2
= n2
.n
& mask
;
1826 if (masked1
&& masked2
&& masked1
!= masked2
)
1831 if (!verify_symbolic_number_p (n
, stmt
))
1838 return source_expr1
;
1843 /* Check if STMT completes a bswap implementation consisting of ORs,
1844 SHIFTs and ANDs. Return the source tree expression on which the
1845 byte swap is performed and NULL if no bswap was found. */
1848 find_bswap (gimple stmt
)
1850 /* The number which the find_bswap result should match in order to
1851 have a full byte swap. The number is shifted to the left according
1852 to the size of the symbolic number before using it. */
1853 unsigned HOST_WIDEST_INT cmp
=
1854 sizeof (HOST_WIDEST_INT
) < 8 ? 0 :
1855 (unsigned HOST_WIDEST_INT
)0x01020304 << 32 | 0x05060708;
1857 struct symbolic_number n
;
1861 /* The last parameter determines the depth search limit. It usually
1862 correlates directly to the number of bytes to be touched. We
1863 increase that number by three here in order to also
1864 cover signed -> unsigned converions of the src operand as can be seen
1865 in libgcc, and for initial shift/and operation of the src operand. */
1866 limit
= TREE_INT_CST_LOW (TYPE_SIZE_UNIT (gimple_expr_type (stmt
)));
1867 limit
+= 1 + (int) ceil_log2 ((unsigned HOST_WIDE_INT
) limit
);
1868 source_expr
= find_bswap_1 (stmt
, &n
, limit
);
1873 /* Zero out the extra bits of N and CMP. */
1874 if (n
.size
< (int)sizeof (HOST_WIDEST_INT
))
1876 unsigned HOST_WIDEST_INT mask
=
1877 ((unsigned HOST_WIDEST_INT
)1 << (n
.size
* BITS_PER_UNIT
)) - 1;
1880 cmp
>>= (sizeof (HOST_WIDEST_INT
) - n
.size
) * BITS_PER_UNIT
;
1883 /* A complete byte swap should make the symbolic number to start
1884 with the largest digit in the highest order byte. */
1891 /* Find manual byte swap implementations and turn them into a bswap
1892 builtin invokation. */
1896 const pass_data pass_data_optimize_bswap
=
1898 GIMPLE_PASS
, /* type */
1900 OPTGROUP_NONE
, /* optinfo_flags */
1901 true, /* has_execute */
1902 TV_NONE
, /* tv_id */
1903 PROP_ssa
, /* properties_required */
1904 0, /* properties_provided */
1905 0, /* properties_destroyed */
1906 0, /* todo_flags_start */
1907 0, /* todo_flags_finish */
1910 class pass_optimize_bswap
: public gimple_opt_pass
1913 pass_optimize_bswap (gcc::context
*ctxt
)
1914 : gimple_opt_pass (pass_data_optimize_bswap
, ctxt
)
1917 /* opt_pass methods: */
1918 virtual bool gate (function
*)
1920 return flag_expensive_optimizations
&& optimize
;
1923 virtual unsigned int execute (function
*);
1925 }; // class pass_optimize_bswap
1928 pass_optimize_bswap::execute (function
*fun
)
1931 bool bswap16_p
, bswap32_p
, bswap64_p
;
1932 bool changed
= false;
1933 tree bswap16_type
= NULL_TREE
, bswap32_type
= NULL_TREE
, bswap64_type
= NULL_TREE
;
1935 if (BITS_PER_UNIT
!= 8)
1938 if (sizeof (HOST_WIDEST_INT
) < 8)
1941 bswap16_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP16
)
1942 && optab_handler (bswap_optab
, HImode
) != CODE_FOR_nothing
);
1943 bswap32_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP32
)
1944 && optab_handler (bswap_optab
, SImode
) != CODE_FOR_nothing
);
1945 bswap64_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP64
)
1946 && (optab_handler (bswap_optab
, DImode
) != CODE_FOR_nothing
1947 || (bswap32_p
&& word_mode
== SImode
)));
1949 if (!bswap16_p
&& !bswap32_p
&& !bswap64_p
)
1952 /* Determine the argument type of the builtins. The code later on
1953 assumes that the return and argument type are the same. */
1956 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP16
);
1957 bswap16_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
1962 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP32
);
1963 bswap32_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
1968 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP64
);
1969 bswap64_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
1972 memset (&bswap_stats
, 0, sizeof (bswap_stats
));
1974 FOR_EACH_BB_FN (bb
, fun
)
1976 gimple_stmt_iterator gsi
;
1978 /* We do a reverse scan for bswap patterns to make sure we get the
1979 widest match. As bswap pattern matching doesn't handle
1980 previously inserted smaller bswap replacements as sub-
1981 patterns, the wider variant wouldn't be detected. */
1982 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
); gsi_prev (&gsi
))
1984 gimple stmt
= gsi_stmt (gsi
);
1985 tree bswap_src
, bswap_type
;
1987 tree fndecl
= NULL_TREE
;
1991 if (!is_gimple_assign (stmt
)
1992 || gimple_assign_rhs_code (stmt
) != BIT_IOR_EXPR
)
1995 type_size
= TYPE_PRECISION (gimple_expr_type (stmt
));
2002 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP16
);
2003 bswap_type
= bswap16_type
;
2009 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP32
);
2010 bswap_type
= bswap32_type
;
2016 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP64
);
2017 bswap_type
= bswap64_type
;
2027 bswap_src
= find_bswap (stmt
);
2033 if (type_size
== 16)
2034 bswap_stats
.found_16bit
++;
2035 else if (type_size
== 32)
2036 bswap_stats
.found_32bit
++;
2038 bswap_stats
.found_64bit
++;
2040 bswap_tmp
= bswap_src
;
2042 /* Convert the src expression if necessary. */
2043 if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp
), bswap_type
))
2045 gimple convert_stmt
;
2046 bswap_tmp
= make_temp_ssa_name (bswap_type
, NULL
, "bswapsrc");
2047 convert_stmt
= gimple_build_assign_with_ops
2048 (NOP_EXPR
, bswap_tmp
, bswap_src
, NULL
);
2049 gsi_insert_before (&gsi
, convert_stmt
, GSI_SAME_STMT
);
2052 call
= gimple_build_call (fndecl
, 1, bswap_tmp
);
2054 bswap_tmp
= gimple_assign_lhs (stmt
);
2056 /* Convert the result if necessary. */
2057 if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp
), bswap_type
))
2059 gimple convert_stmt
;
2060 bswap_tmp
= make_temp_ssa_name (bswap_type
, NULL
, "bswapdst");
2061 convert_stmt
= gimple_build_assign_with_ops
2062 (NOP_EXPR
, gimple_assign_lhs (stmt
), bswap_tmp
, NULL
);
2063 gsi_insert_after (&gsi
, convert_stmt
, GSI_SAME_STMT
);
2066 gimple_call_set_lhs (call
, bswap_tmp
);
2070 fprintf (dump_file
, "%d bit bswap implementation found at: ",
2072 print_gimple_stmt (dump_file
, stmt
, 0, 0);
2075 gsi_insert_after (&gsi
, call
, GSI_SAME_STMT
);
2076 gsi_remove (&gsi
, true);
2080 statistics_counter_event (fun
, "16-bit bswap implementations found",
2081 bswap_stats
.found_16bit
);
2082 statistics_counter_event (fun
, "32-bit bswap implementations found",
2083 bswap_stats
.found_32bit
);
2084 statistics_counter_event (fun
, "64-bit bswap implementations found",
2085 bswap_stats
.found_64bit
);
2087 return (changed
? TODO_update_ssa
: 0);
2093 make_pass_optimize_bswap (gcc::context
*ctxt
)
2095 return new pass_optimize_bswap (ctxt
);
2098 /* Return true if stmt is a type conversion operation that can be stripped
2099 when used in a widening multiply operation. */
2101 widening_mult_conversion_strippable_p (tree result_type
, gimple stmt
)
2103 enum tree_code rhs_code
= gimple_assign_rhs_code (stmt
);
2105 if (TREE_CODE (result_type
) == INTEGER_TYPE
)
2110 if (!CONVERT_EXPR_CODE_P (rhs_code
))
2113 op_type
= TREE_TYPE (gimple_assign_lhs (stmt
));
2115 /* If the type of OP has the same precision as the result, then
2116 we can strip this conversion. The multiply operation will be
2117 selected to create the correct extension as a by-product. */
2118 if (TYPE_PRECISION (result_type
) == TYPE_PRECISION (op_type
))
2121 /* We can also strip a conversion if it preserves the signed-ness of
2122 the operation and doesn't narrow the range. */
2123 inner_op_type
= TREE_TYPE (gimple_assign_rhs1 (stmt
));
2125 /* If the inner-most type is unsigned, then we can strip any
2126 intermediate widening operation. If it's signed, then the
2127 intermediate widening operation must also be signed. */
2128 if ((TYPE_UNSIGNED (inner_op_type
)
2129 || TYPE_UNSIGNED (op_type
) == TYPE_UNSIGNED (inner_op_type
))
2130 && TYPE_PRECISION (op_type
) > TYPE_PRECISION (inner_op_type
))
2136 return rhs_code
== FIXED_CONVERT_EXPR
;
2139 /* Return true if RHS is a suitable operand for a widening multiplication,
2140 assuming a target type of TYPE.
2141 There are two cases:
2143 - RHS makes some value at least twice as wide. Store that value
2144 in *NEW_RHS_OUT if so, and store its type in *TYPE_OUT.
2146 - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so,
2147 but leave *TYPE_OUT untouched. */
2150 is_widening_mult_rhs_p (tree type
, tree rhs
, tree
*type_out
,
2156 if (TREE_CODE (rhs
) == SSA_NAME
)
2158 stmt
= SSA_NAME_DEF_STMT (rhs
);
2159 if (is_gimple_assign (stmt
))
2161 if (! widening_mult_conversion_strippable_p (type
, stmt
))
2165 rhs1
= gimple_assign_rhs1 (stmt
);
2167 if (TREE_CODE (rhs1
) == INTEGER_CST
)
2169 *new_rhs_out
= rhs1
;
2178 type1
= TREE_TYPE (rhs1
);
2180 if (TREE_CODE (type1
) != TREE_CODE (type
)
2181 || TYPE_PRECISION (type1
) * 2 > TYPE_PRECISION (type
))
2184 *new_rhs_out
= rhs1
;
2189 if (TREE_CODE (rhs
) == INTEGER_CST
)
2199 /* Return true if STMT performs a widening multiplication, assuming the
2200 output type is TYPE. If so, store the unwidened types of the operands
2201 in *TYPE1_OUT and *TYPE2_OUT respectively. Also fill *RHS1_OUT and
2202 *RHS2_OUT such that converting those operands to types *TYPE1_OUT
2203 and *TYPE2_OUT would give the operands of the multiplication. */
2206 is_widening_mult_p (gimple stmt
,
2207 tree
*type1_out
, tree
*rhs1_out
,
2208 tree
*type2_out
, tree
*rhs2_out
)
2210 tree type
= TREE_TYPE (gimple_assign_lhs (stmt
));
2212 if (TREE_CODE (type
) != INTEGER_TYPE
2213 && TREE_CODE (type
) != FIXED_POINT_TYPE
)
2216 if (!is_widening_mult_rhs_p (type
, gimple_assign_rhs1 (stmt
), type1_out
,
2220 if (!is_widening_mult_rhs_p (type
, gimple_assign_rhs2 (stmt
), type2_out
,
2224 if (*type1_out
== NULL
)
2226 if (*type2_out
== NULL
|| !int_fits_type_p (*rhs1_out
, *type2_out
))
2228 *type1_out
= *type2_out
;
2231 if (*type2_out
== NULL
)
2233 if (!int_fits_type_p (*rhs2_out
, *type1_out
))
2235 *type2_out
= *type1_out
;
2238 /* Ensure that the larger of the two operands comes first. */
2239 if (TYPE_PRECISION (*type1_out
) < TYPE_PRECISION (*type2_out
))
2243 *type1_out
= *type2_out
;
2246 *rhs1_out
= *rhs2_out
;
2253 /* Process a single gimple statement STMT, which has a MULT_EXPR as
2254 its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return
2255 value is true iff we converted the statement. */
2258 convert_mult_to_widen (gimple stmt
, gimple_stmt_iterator
*gsi
)
2260 tree lhs
, rhs1
, rhs2
, type
, type1
, type2
;
2261 enum insn_code handler
;
2262 enum machine_mode to_mode
, from_mode
, actual_mode
;
2264 int actual_precision
;
2265 location_t loc
= gimple_location (stmt
);
2266 bool from_unsigned1
, from_unsigned2
;
2268 lhs
= gimple_assign_lhs (stmt
);
2269 type
= TREE_TYPE (lhs
);
2270 if (TREE_CODE (type
) != INTEGER_TYPE
)
2273 if (!is_widening_mult_p (stmt
, &type1
, &rhs1
, &type2
, &rhs2
))
2276 to_mode
= TYPE_MODE (type
);
2277 from_mode
= TYPE_MODE (type1
);
2278 from_unsigned1
= TYPE_UNSIGNED (type1
);
2279 from_unsigned2
= TYPE_UNSIGNED (type2
);
2281 if (from_unsigned1
&& from_unsigned2
)
2282 op
= umul_widen_optab
;
2283 else if (!from_unsigned1
&& !from_unsigned2
)
2284 op
= smul_widen_optab
;
2286 op
= usmul_widen_optab
;
2288 handler
= find_widening_optab_handler_and_mode (op
, to_mode
, from_mode
,
2291 if (handler
== CODE_FOR_nothing
)
2293 if (op
!= smul_widen_optab
)
2295 /* We can use a signed multiply with unsigned types as long as
2296 there is a wider mode to use, or it is the smaller of the two
2297 types that is unsigned. Note that type1 >= type2, always. */
2298 if ((TYPE_UNSIGNED (type1
)
2299 && TYPE_PRECISION (type1
) == GET_MODE_PRECISION (from_mode
))
2300 || (TYPE_UNSIGNED (type2
)
2301 && TYPE_PRECISION (type2
) == GET_MODE_PRECISION (from_mode
)))
2303 from_mode
= GET_MODE_WIDER_MODE (from_mode
);
2304 if (GET_MODE_SIZE (to_mode
) <= GET_MODE_SIZE (from_mode
))
2308 op
= smul_widen_optab
;
2309 handler
= find_widening_optab_handler_and_mode (op
, to_mode
,
2313 if (handler
== CODE_FOR_nothing
)
2316 from_unsigned1
= from_unsigned2
= false;
2322 /* Ensure that the inputs to the handler are in the correct precison
2323 for the opcode. This will be the full mode size. */
2324 actual_precision
= GET_MODE_PRECISION (actual_mode
);
2325 if (2 * actual_precision
> TYPE_PRECISION (type
))
2327 if (actual_precision
!= TYPE_PRECISION (type1
)
2328 || from_unsigned1
!= TYPE_UNSIGNED (type1
))
2329 rhs1
= build_and_insert_cast (gsi
, loc
,
2330 build_nonstandard_integer_type
2331 (actual_precision
, from_unsigned1
), rhs1
);
2332 if (actual_precision
!= TYPE_PRECISION (type2
)
2333 || from_unsigned2
!= TYPE_UNSIGNED (type2
))
2334 rhs2
= build_and_insert_cast (gsi
, loc
,
2335 build_nonstandard_integer_type
2336 (actual_precision
, from_unsigned2
), rhs2
);
2338 /* Handle constants. */
2339 if (TREE_CODE (rhs1
) == INTEGER_CST
)
2340 rhs1
= fold_convert (type1
, rhs1
);
2341 if (TREE_CODE (rhs2
) == INTEGER_CST
)
2342 rhs2
= fold_convert (type2
, rhs2
);
2344 gimple_assign_set_rhs1 (stmt
, rhs1
);
2345 gimple_assign_set_rhs2 (stmt
, rhs2
);
2346 gimple_assign_set_rhs_code (stmt
, WIDEN_MULT_EXPR
);
2348 widen_mul_stats
.widen_mults_inserted
++;
2352 /* Process a single gimple statement STMT, which is found at the
2353 iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its
2354 rhs (given by CODE), and try to convert it into a
2355 WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value
2356 is true iff we converted the statement. */
2359 convert_plusminus_to_widen (gimple_stmt_iterator
*gsi
, gimple stmt
,
2360 enum tree_code code
)
2362 gimple rhs1_stmt
= NULL
, rhs2_stmt
= NULL
;
2363 gimple conv1_stmt
= NULL
, conv2_stmt
= NULL
, conv_stmt
;
2364 tree type
, type1
, type2
, optype
;
2365 tree lhs
, rhs1
, rhs2
, mult_rhs1
, mult_rhs2
, add_rhs
;
2366 enum tree_code rhs1_code
= ERROR_MARK
, rhs2_code
= ERROR_MARK
;
2368 enum tree_code wmult_code
;
2369 enum insn_code handler
;
2370 enum machine_mode to_mode
, from_mode
, actual_mode
;
2371 location_t loc
= gimple_location (stmt
);
2372 int actual_precision
;
2373 bool from_unsigned1
, from_unsigned2
;
2375 lhs
= gimple_assign_lhs (stmt
);
2376 type
= TREE_TYPE (lhs
);
2377 if (TREE_CODE (type
) != INTEGER_TYPE
2378 && TREE_CODE (type
) != FIXED_POINT_TYPE
)
2381 if (code
== MINUS_EXPR
)
2382 wmult_code
= WIDEN_MULT_MINUS_EXPR
;
2384 wmult_code
= WIDEN_MULT_PLUS_EXPR
;
2386 rhs1
= gimple_assign_rhs1 (stmt
);
2387 rhs2
= gimple_assign_rhs2 (stmt
);
2389 if (TREE_CODE (rhs1
) == SSA_NAME
)
2391 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
2392 if (is_gimple_assign (rhs1_stmt
))
2393 rhs1_code
= gimple_assign_rhs_code (rhs1_stmt
);
2396 if (TREE_CODE (rhs2
) == SSA_NAME
)
2398 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
2399 if (is_gimple_assign (rhs2_stmt
))
2400 rhs2_code
= gimple_assign_rhs_code (rhs2_stmt
);
2403 /* Allow for one conversion statement between the multiply
2404 and addition/subtraction statement. If there are more than
2405 one conversions then we assume they would invalidate this
2406 transformation. If that's not the case then they should have
2407 been folded before now. */
2408 if (CONVERT_EXPR_CODE_P (rhs1_code
))
2410 conv1_stmt
= rhs1_stmt
;
2411 rhs1
= gimple_assign_rhs1 (rhs1_stmt
);
2412 if (TREE_CODE (rhs1
) == SSA_NAME
)
2414 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
2415 if (is_gimple_assign (rhs1_stmt
))
2416 rhs1_code
= gimple_assign_rhs_code (rhs1_stmt
);
2421 if (CONVERT_EXPR_CODE_P (rhs2_code
))
2423 conv2_stmt
= rhs2_stmt
;
2424 rhs2
= gimple_assign_rhs1 (rhs2_stmt
);
2425 if (TREE_CODE (rhs2
) == SSA_NAME
)
2427 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
2428 if (is_gimple_assign (rhs2_stmt
))
2429 rhs2_code
= gimple_assign_rhs_code (rhs2_stmt
);
2435 /* If code is WIDEN_MULT_EXPR then it would seem unnecessary to call
2436 is_widening_mult_p, but we still need the rhs returns.
2438 It might also appear that it would be sufficient to use the existing
2439 operands of the widening multiply, but that would limit the choice of
2440 multiply-and-accumulate instructions.
2442 If the widened-multiplication result has more than one uses, it is
2443 probably wiser not to do the conversion. */
2444 if (code
== PLUS_EXPR
2445 && (rhs1_code
== MULT_EXPR
|| rhs1_code
== WIDEN_MULT_EXPR
))
2447 if (!has_single_use (rhs1
)
2448 || !is_widening_mult_p (rhs1_stmt
, &type1
, &mult_rhs1
,
2449 &type2
, &mult_rhs2
))
2452 conv_stmt
= conv1_stmt
;
2454 else if (rhs2_code
== MULT_EXPR
|| rhs2_code
== WIDEN_MULT_EXPR
)
2456 if (!has_single_use (rhs2
)
2457 || !is_widening_mult_p (rhs2_stmt
, &type1
, &mult_rhs1
,
2458 &type2
, &mult_rhs2
))
2461 conv_stmt
= conv2_stmt
;
2466 to_mode
= TYPE_MODE (type
);
2467 from_mode
= TYPE_MODE (type1
);
2468 from_unsigned1
= TYPE_UNSIGNED (type1
);
2469 from_unsigned2
= TYPE_UNSIGNED (type2
);
2472 /* There's no such thing as a mixed sign madd yet, so use a wider mode. */
2473 if (from_unsigned1
!= from_unsigned2
)
2475 if (!INTEGRAL_TYPE_P (type
))
2477 /* We can use a signed multiply with unsigned types as long as
2478 there is a wider mode to use, or it is the smaller of the two
2479 types that is unsigned. Note that type1 >= type2, always. */
2481 && TYPE_PRECISION (type1
) == GET_MODE_PRECISION (from_mode
))
2483 && TYPE_PRECISION (type2
) == GET_MODE_PRECISION (from_mode
)))
2485 from_mode
= GET_MODE_WIDER_MODE (from_mode
);
2486 if (GET_MODE_SIZE (from_mode
) >= GET_MODE_SIZE (to_mode
))
2490 from_unsigned1
= from_unsigned2
= false;
2491 optype
= build_nonstandard_integer_type (GET_MODE_PRECISION (from_mode
),
2495 /* If there was a conversion between the multiply and addition
2496 then we need to make sure it fits a multiply-and-accumulate.
2497 The should be a single mode change which does not change the
2501 /* We use the original, unmodified data types for this. */
2502 tree from_type
= TREE_TYPE (gimple_assign_rhs1 (conv_stmt
));
2503 tree to_type
= TREE_TYPE (gimple_assign_lhs (conv_stmt
));
2504 int data_size
= TYPE_PRECISION (type1
) + TYPE_PRECISION (type2
);
2505 bool is_unsigned
= TYPE_UNSIGNED (type1
) && TYPE_UNSIGNED (type2
);
2507 if (TYPE_PRECISION (from_type
) > TYPE_PRECISION (to_type
))
2509 /* Conversion is a truncate. */
2510 if (TYPE_PRECISION (to_type
) < data_size
)
2513 else if (TYPE_PRECISION (from_type
) < TYPE_PRECISION (to_type
))
2515 /* Conversion is an extend. Check it's the right sort. */
2516 if (TYPE_UNSIGNED (from_type
) != is_unsigned
2517 && !(is_unsigned
&& TYPE_PRECISION (from_type
) > data_size
))
2520 /* else convert is a no-op for our purposes. */
2523 /* Verify that the machine can perform a widening multiply
2524 accumulate in this mode/signedness combination, otherwise
2525 this transformation is likely to pessimize code. */
2526 this_optab
= optab_for_tree_code (wmult_code
, optype
, optab_default
);
2527 handler
= find_widening_optab_handler_and_mode (this_optab
, to_mode
,
2528 from_mode
, 0, &actual_mode
);
2530 if (handler
== CODE_FOR_nothing
)
2533 /* Ensure that the inputs to the handler are in the correct precison
2534 for the opcode. This will be the full mode size. */
2535 actual_precision
= GET_MODE_PRECISION (actual_mode
);
2536 if (actual_precision
!= TYPE_PRECISION (type1
)
2537 || from_unsigned1
!= TYPE_UNSIGNED (type1
))
2538 mult_rhs1
= build_and_insert_cast (gsi
, loc
,
2539 build_nonstandard_integer_type
2540 (actual_precision
, from_unsigned1
),
2542 if (actual_precision
!= TYPE_PRECISION (type2
)
2543 || from_unsigned2
!= TYPE_UNSIGNED (type2
))
2544 mult_rhs2
= build_and_insert_cast (gsi
, loc
,
2545 build_nonstandard_integer_type
2546 (actual_precision
, from_unsigned2
),
2549 if (!useless_type_conversion_p (type
, TREE_TYPE (add_rhs
)))
2550 add_rhs
= build_and_insert_cast (gsi
, loc
, type
, add_rhs
);
2552 /* Handle constants. */
2553 if (TREE_CODE (mult_rhs1
) == INTEGER_CST
)
2554 mult_rhs1
= fold_convert (type1
, mult_rhs1
);
2555 if (TREE_CODE (mult_rhs2
) == INTEGER_CST
)
2556 mult_rhs2
= fold_convert (type2
, mult_rhs2
);
2558 gimple_assign_set_rhs_with_ops_1 (gsi
, wmult_code
, mult_rhs1
, mult_rhs2
,
2560 update_stmt (gsi_stmt (*gsi
));
2561 widen_mul_stats
.maccs_inserted
++;
2565 /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2
2566 with uses in additions and subtractions to form fused multiply-add
2567 operations. Returns true if successful and MUL_STMT should be removed. */
2570 convert_mult_to_fma (gimple mul_stmt
, tree op1
, tree op2
)
2572 tree mul_result
= gimple_get_lhs (mul_stmt
);
2573 tree type
= TREE_TYPE (mul_result
);
2574 gimple use_stmt
, neguse_stmt
, fma_stmt
;
2575 use_operand_p use_p
;
2576 imm_use_iterator imm_iter
;
2578 if (FLOAT_TYPE_P (type
)
2579 && flag_fp_contract_mode
== FP_CONTRACT_OFF
)
2582 /* We don't want to do bitfield reduction ops. */
2583 if (INTEGRAL_TYPE_P (type
)
2584 && (TYPE_PRECISION (type
)
2585 != GET_MODE_PRECISION (TYPE_MODE (type
))))
2588 /* If the target doesn't support it, don't generate it. We assume that
2589 if fma isn't available then fms, fnma or fnms are not either. */
2590 if (optab_handler (fma_optab
, TYPE_MODE (type
)) == CODE_FOR_nothing
)
2593 /* If the multiplication has zero uses, it is kept around probably because
2594 of -fnon-call-exceptions. Don't optimize it away in that case,
2596 if (has_zero_uses (mul_result
))
2599 /* Make sure that the multiplication statement becomes dead after
2600 the transformation, thus that all uses are transformed to FMAs.
2601 This means we assume that an FMA operation has the same cost
2603 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, mul_result
)
2605 enum tree_code use_code
;
2606 tree result
= mul_result
;
2607 bool negate_p
= false;
2609 use_stmt
= USE_STMT (use_p
);
2611 if (is_gimple_debug (use_stmt
))
2614 /* For now restrict this operations to single basic blocks. In theory
2615 we would want to support sinking the multiplication in
2621 to form a fma in the then block and sink the multiplication to the
2623 if (gimple_bb (use_stmt
) != gimple_bb (mul_stmt
))
2626 if (!is_gimple_assign (use_stmt
))
2629 use_code
= gimple_assign_rhs_code (use_stmt
);
2631 /* A negate on the multiplication leads to FNMA. */
2632 if (use_code
== NEGATE_EXPR
)
2637 result
= gimple_assign_lhs (use_stmt
);
2639 /* Make sure the negate statement becomes dead with this
2640 single transformation. */
2641 if (!single_imm_use (gimple_assign_lhs (use_stmt
),
2642 &use_p
, &neguse_stmt
))
2645 /* Make sure the multiplication isn't also used on that stmt. */
2646 FOR_EACH_PHI_OR_STMT_USE (usep
, neguse_stmt
, iter
, SSA_OP_USE
)
2647 if (USE_FROM_PTR (usep
) == mul_result
)
2651 use_stmt
= neguse_stmt
;
2652 if (gimple_bb (use_stmt
) != gimple_bb (mul_stmt
))
2654 if (!is_gimple_assign (use_stmt
))
2657 use_code
= gimple_assign_rhs_code (use_stmt
);
2664 if (gimple_assign_rhs2 (use_stmt
) == result
)
2665 negate_p
= !negate_p
;
2670 /* FMA can only be formed from PLUS and MINUS. */
2674 /* If the subtrahend (gimple_assign_rhs2 (use_stmt)) is computed
2675 by a MULT_EXPR that we'll visit later, we might be able to
2676 get a more profitable match with fnma.
2677 OTOH, if we don't, a negate / fma pair has likely lower latency
2678 that a mult / subtract pair. */
2679 if (use_code
== MINUS_EXPR
&& !negate_p
2680 && gimple_assign_rhs1 (use_stmt
) == result
2681 && optab_handler (fms_optab
, TYPE_MODE (type
)) == CODE_FOR_nothing
2682 && optab_handler (fnma_optab
, TYPE_MODE (type
)) != CODE_FOR_nothing
)
2684 tree rhs2
= gimple_assign_rhs2 (use_stmt
);
2686 if (TREE_CODE (rhs2
) == SSA_NAME
)
2688 gimple stmt2
= SSA_NAME_DEF_STMT (rhs2
);
2689 if (has_single_use (rhs2
)
2690 && is_gimple_assign (stmt2
)
2691 && gimple_assign_rhs_code (stmt2
) == MULT_EXPR
)
2696 /* We can't handle a * b + a * b. */
2697 if (gimple_assign_rhs1 (use_stmt
) == gimple_assign_rhs2 (use_stmt
))
2700 /* While it is possible to validate whether or not the exact form
2701 that we've recognized is available in the backend, the assumption
2702 is that the transformation is never a loss. For instance, suppose
2703 the target only has the plain FMA pattern available. Consider
2704 a*b-c -> fma(a,b,-c): we've exchanged MUL+SUB for FMA+NEG, which
2705 is still two operations. Consider -(a*b)-c -> fma(-a,b,-c): we
2706 still have 3 operations, but in the FMA form the two NEGs are
2707 independent and could be run in parallel. */
2710 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, mul_result
)
2712 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
2713 enum tree_code use_code
;
2714 tree addop
, mulop1
= op1
, result
= mul_result
;
2715 bool negate_p
= false;
2717 if (is_gimple_debug (use_stmt
))
2720 use_code
= gimple_assign_rhs_code (use_stmt
);
2721 if (use_code
== NEGATE_EXPR
)
2723 result
= gimple_assign_lhs (use_stmt
);
2724 single_imm_use (gimple_assign_lhs (use_stmt
), &use_p
, &neguse_stmt
);
2725 gsi_remove (&gsi
, true);
2726 release_defs (use_stmt
);
2728 use_stmt
= neguse_stmt
;
2729 gsi
= gsi_for_stmt (use_stmt
);
2730 use_code
= gimple_assign_rhs_code (use_stmt
);
2734 if (gimple_assign_rhs1 (use_stmt
) == result
)
2736 addop
= gimple_assign_rhs2 (use_stmt
);
2737 /* a * b - c -> a * b + (-c) */
2738 if (gimple_assign_rhs_code (use_stmt
) == MINUS_EXPR
)
2739 addop
= force_gimple_operand_gsi (&gsi
,
2740 build1 (NEGATE_EXPR
,
2742 true, NULL_TREE
, true,
2747 addop
= gimple_assign_rhs1 (use_stmt
);
2748 /* a - b * c -> (-b) * c + a */
2749 if (gimple_assign_rhs_code (use_stmt
) == MINUS_EXPR
)
2750 negate_p
= !negate_p
;
2754 mulop1
= force_gimple_operand_gsi (&gsi
,
2755 build1 (NEGATE_EXPR
,
2757 true, NULL_TREE
, true,
2760 fma_stmt
= gimple_build_assign_with_ops (FMA_EXPR
,
2761 gimple_assign_lhs (use_stmt
),
2764 gsi_replace (&gsi
, fma_stmt
, true);
2765 widen_mul_stats
.fmas_inserted
++;
2771 /* Find integer multiplications where the operands are extended from
2772 smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR
2773 where appropriate. */
2777 const pass_data pass_data_optimize_widening_mul
=
2779 GIMPLE_PASS
, /* type */
2780 "widening_mul", /* name */
2781 OPTGROUP_NONE
, /* optinfo_flags */
2782 true, /* has_execute */
2783 TV_NONE
, /* tv_id */
2784 PROP_ssa
, /* properties_required */
2785 0, /* properties_provided */
2786 0, /* properties_destroyed */
2787 0, /* todo_flags_start */
2788 TODO_update_ssa
, /* todo_flags_finish */
2791 class pass_optimize_widening_mul
: public gimple_opt_pass
2794 pass_optimize_widening_mul (gcc::context
*ctxt
)
2795 : gimple_opt_pass (pass_data_optimize_widening_mul
, ctxt
)
2798 /* opt_pass methods: */
2799 virtual bool gate (function
*)
2801 return flag_expensive_optimizations
&& optimize
;
2804 virtual unsigned int execute (function
*);
2806 }; // class pass_optimize_widening_mul
2809 pass_optimize_widening_mul::execute (function
*fun
)
2812 bool cfg_changed
= false;
2814 memset (&widen_mul_stats
, 0, sizeof (widen_mul_stats
));
2816 FOR_EACH_BB_FN (bb
, fun
)
2818 gimple_stmt_iterator gsi
;
2820 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
);)
2822 gimple stmt
= gsi_stmt (gsi
);
2823 enum tree_code code
;
2825 if (is_gimple_assign (stmt
))
2827 code
= gimple_assign_rhs_code (stmt
);
2831 if (!convert_mult_to_widen (stmt
, &gsi
)
2832 && convert_mult_to_fma (stmt
,
2833 gimple_assign_rhs1 (stmt
),
2834 gimple_assign_rhs2 (stmt
)))
2836 gsi_remove (&gsi
, true);
2837 release_defs (stmt
);
2844 convert_plusminus_to_widen (&gsi
, stmt
, code
);
2850 else if (is_gimple_call (stmt
)
2851 && gimple_call_lhs (stmt
))
2853 tree fndecl
= gimple_call_fndecl (stmt
);
2855 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
)
2857 switch (DECL_FUNCTION_CODE (fndecl
))
2862 if (TREE_CODE (gimple_call_arg (stmt
, 1)) == REAL_CST
2863 && REAL_VALUES_EQUAL
2864 (TREE_REAL_CST (gimple_call_arg (stmt
, 1)),
2866 && convert_mult_to_fma (stmt
,
2867 gimple_call_arg (stmt
, 0),
2868 gimple_call_arg (stmt
, 0)))
2870 unlink_stmt_vdef (stmt
);
2871 if (gsi_remove (&gsi
, true)
2872 && gimple_purge_dead_eh_edges (bb
))
2874 release_defs (stmt
);
2887 statistics_counter_event (fun
, "widening multiplications inserted",
2888 widen_mul_stats
.widen_mults_inserted
);
2889 statistics_counter_event (fun
, "widening maccs inserted",
2890 widen_mul_stats
.maccs_inserted
);
2891 statistics_counter_event (fun
, "fused multiply-adds inserted",
2892 widen_mul_stats
.fmas_inserted
);
2894 return cfg_changed
? TODO_cleanup_cfg
: 0;
2900 make_pass_optimize_widening_mul (gcc::context
*ctxt
)
2902 return new pass_optimize_widening_mul (ctxt
);