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.h"
102 #include "gimplify-me.h"
103 #include "stor-layout.h"
104 #include "gimple-ssa.h"
105 #include "tree-cfg.h"
106 #include "tree-phinodes.h"
107 #include "ssa-iterators.h"
108 #include "stringpool.h"
109 #include "tree-ssanames.h"
111 #include "tree-dfa.h"
112 #include "tree-ssa.h"
113 #include "tree-pass.h"
114 #include "alloc-pool.h"
116 #include "gimple-pretty-print.h"
117 #include "builtins.h"
119 /* FIXME: RTL headers have to be included here for optabs. */
120 #include "rtl.h" /* Because optabs.h wants enum rtx_code. */
121 #include "expr.h" /* Because optabs.h wants sepops. */
124 /* This structure represents one basic block that either computes a
125 division, or is a common dominator for basic block that compute a
128 /* The basic block represented by this structure. */
131 /* If non-NULL, the SSA_NAME holding the definition for a reciprocal
135 /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that
136 was inserted in BB. */
137 gimple recip_def_stmt
;
139 /* Pointer to a list of "struct occurrence"s for blocks dominated
141 struct occurrence
*children
;
143 /* Pointer to the next "struct occurrence"s in the list of blocks
144 sharing a common dominator. */
145 struct occurrence
*next
;
147 /* The number of divisions that are in BB before compute_merit. The
148 number of divisions that are in BB or post-dominate it after
152 /* True if the basic block has a division, false if it is a common
153 dominator for basic blocks that do. If it is false and trapping
154 math is active, BB is not a candidate for inserting a reciprocal. */
155 bool bb_has_division
;
160 /* Number of 1.0/X ops inserted. */
163 /* Number of 1.0/FUNC ops inserted. */
169 /* Number of cexpi calls inserted. */
175 /* Number of hand-written 16-bit nop / bswaps found. */
178 /* Number of hand-written 32-bit nop / bswaps found. */
181 /* Number of hand-written 64-bit nop / bswaps found. */
183 } nop_stats
, bswap_stats
;
187 /* Number of widening multiplication ops inserted. */
188 int widen_mults_inserted
;
190 /* Number of integer multiply-and-accumulate ops inserted. */
193 /* Number of fp fused multiply-add ops inserted. */
197 /* The instance of "struct occurrence" representing the highest
198 interesting block in the dominator tree. */
199 static struct occurrence
*occ_head
;
201 /* Allocation pool for getting instances of "struct occurrence". */
202 static alloc_pool occ_pool
;
206 /* Allocate and return a new struct occurrence for basic block BB, and
207 whose children list is headed by CHILDREN. */
208 static struct occurrence
*
209 occ_new (basic_block bb
, struct occurrence
*children
)
211 struct occurrence
*occ
;
213 bb
->aux
= occ
= (struct occurrence
*) pool_alloc (occ_pool
);
214 memset (occ
, 0, sizeof (struct occurrence
));
217 occ
->children
= children
;
222 /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a
223 list of "struct occurrence"s, one per basic block, having IDOM as
224 their common dominator.
226 We try to insert NEW_OCC as deep as possible in the tree, and we also
227 insert any other block that is a common dominator for BB and one
228 block already in the tree. */
231 insert_bb (struct occurrence
*new_occ
, basic_block idom
,
232 struct occurrence
**p_head
)
234 struct occurrence
*occ
, **p_occ
;
236 for (p_occ
= p_head
; (occ
= *p_occ
) != NULL
; )
238 basic_block bb
= new_occ
->bb
, occ_bb
= occ
->bb
;
239 basic_block dom
= nearest_common_dominator (CDI_DOMINATORS
, occ_bb
, bb
);
242 /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC
245 occ
->next
= new_occ
->children
;
246 new_occ
->children
= occ
;
248 /* Try the next block (it may as well be dominated by BB). */
251 else if (dom
== occ_bb
)
253 /* OCC_BB dominates BB. Tail recurse to look deeper. */
254 insert_bb (new_occ
, dom
, &occ
->children
);
258 else if (dom
!= idom
)
260 gcc_assert (!dom
->aux
);
262 /* There is a dominator between IDOM and BB, add it and make
263 two children out of NEW_OCC and OCC. First, remove OCC from
269 /* None of the previous blocks has DOM as a dominator: if we tail
270 recursed, we would reexamine them uselessly. Just switch BB with
271 DOM, and go on looking for blocks dominated by DOM. */
272 new_occ
= occ_new (dom
, new_occ
);
277 /* Nothing special, go on with the next element. */
282 /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */
283 new_occ
->next
= *p_head
;
287 /* Register that we found a division in BB. */
290 register_division_in (basic_block bb
)
292 struct occurrence
*occ
;
294 occ
= (struct occurrence
*) bb
->aux
;
297 occ
= occ_new (bb
, NULL
);
298 insert_bb (occ
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), &occ_head
);
301 occ
->bb_has_division
= true;
302 occ
->num_divisions
++;
306 /* Compute the number of divisions that postdominate each block in OCC and
310 compute_merit (struct occurrence
*occ
)
312 struct occurrence
*occ_child
;
313 basic_block dom
= occ
->bb
;
315 for (occ_child
= occ
->children
; occ_child
; occ_child
= occ_child
->next
)
318 if (occ_child
->children
)
319 compute_merit (occ_child
);
322 bb
= single_noncomplex_succ (dom
);
326 if (dominated_by_p (CDI_POST_DOMINATORS
, bb
, occ_child
->bb
))
327 occ
->num_divisions
+= occ_child
->num_divisions
;
332 /* Return whether USE_STMT is a floating-point division by DEF. */
334 is_division_by (gimple use_stmt
, tree def
)
336 return is_gimple_assign (use_stmt
)
337 && gimple_assign_rhs_code (use_stmt
) == RDIV_EXPR
338 && gimple_assign_rhs2 (use_stmt
) == def
339 /* Do not recognize x / x as valid division, as we are getting
340 confused later by replacing all immediate uses x in such
342 && gimple_assign_rhs1 (use_stmt
) != def
;
345 /* Walk the subset of the dominator tree rooted at OCC, setting the
346 RECIP_DEF field to a definition of 1.0 / DEF that can be used in
347 the given basic block. The field may be left NULL, of course,
348 if it is not possible or profitable to do the optimization.
350 DEF_BSI is an iterator pointing at the statement defining DEF.
351 If RECIP_DEF is set, a dominator already has a computation that can
355 insert_reciprocals (gimple_stmt_iterator
*def_gsi
, struct occurrence
*occ
,
356 tree def
, tree recip_def
, int threshold
)
360 gimple_stmt_iterator gsi
;
361 struct occurrence
*occ_child
;
364 && (occ
->bb_has_division
|| !flag_trapping_math
)
365 && occ
->num_divisions
>= threshold
)
367 /* Make a variable with the replacement and substitute it. */
368 type
= TREE_TYPE (def
);
369 recip_def
= create_tmp_reg (type
, "reciptmp");
370 new_stmt
= gimple_build_assign_with_ops (RDIV_EXPR
, recip_def
,
371 build_one_cst (type
), def
);
373 if (occ
->bb_has_division
)
375 /* Case 1: insert before an existing division. */
376 gsi
= gsi_after_labels (occ
->bb
);
377 while (!gsi_end_p (gsi
) && !is_division_by (gsi_stmt (gsi
), def
))
380 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
382 else if (def_gsi
&& occ
->bb
== def_gsi
->bb
)
384 /* Case 2: insert right after the definition. Note that this will
385 never happen if the definition statement can throw, because in
386 that case the sole successor of the statement's basic block will
387 dominate all the uses as well. */
388 gsi_insert_after (def_gsi
, new_stmt
, GSI_NEW_STMT
);
392 /* Case 3: insert in a basic block not containing defs/uses. */
393 gsi
= gsi_after_labels (occ
->bb
);
394 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
397 reciprocal_stats
.rdivs_inserted
++;
399 occ
->recip_def_stmt
= new_stmt
;
402 occ
->recip_def
= recip_def
;
403 for (occ_child
= occ
->children
; occ_child
; occ_child
= occ_child
->next
)
404 insert_reciprocals (def_gsi
, occ_child
, def
, recip_def
, threshold
);
408 /* Replace the division at USE_P with a multiplication by the reciprocal, if
412 replace_reciprocal (use_operand_p use_p
)
414 gimple use_stmt
= USE_STMT (use_p
);
415 basic_block bb
= gimple_bb (use_stmt
);
416 struct occurrence
*occ
= (struct occurrence
*) bb
->aux
;
418 if (optimize_bb_for_speed_p (bb
)
419 && occ
->recip_def
&& use_stmt
!= occ
->recip_def_stmt
)
421 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
422 gimple_assign_set_rhs_code (use_stmt
, MULT_EXPR
);
423 SET_USE (use_p
, occ
->recip_def
);
424 fold_stmt_inplace (&gsi
);
425 update_stmt (use_stmt
);
430 /* Free OCC and return one more "struct occurrence" to be freed. */
432 static struct occurrence
*
433 free_bb (struct occurrence
*occ
)
435 struct occurrence
*child
, *next
;
437 /* First get the two pointers hanging off OCC. */
439 child
= occ
->children
;
441 pool_free (occ_pool
, occ
);
443 /* Now ensure that we don't recurse unless it is necessary. */
449 next
= free_bb (next
);
456 /* Look for floating-point divisions among DEF's uses, and try to
457 replace them by multiplications with the reciprocal. Add
458 as many statements computing the reciprocal as needed.
460 DEF must be a GIMPLE register of a floating-point type. */
463 execute_cse_reciprocals_1 (gimple_stmt_iterator
*def_gsi
, tree def
)
466 imm_use_iterator use_iter
;
467 struct occurrence
*occ
;
468 int count
= 0, threshold
;
470 gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def
)) && is_gimple_reg (def
));
472 FOR_EACH_IMM_USE_FAST (use_p
, use_iter
, def
)
474 gimple use_stmt
= USE_STMT (use_p
);
475 if (is_division_by (use_stmt
, def
))
477 register_division_in (gimple_bb (use_stmt
));
482 /* Do the expensive part only if we can hope to optimize something. */
483 threshold
= targetm
.min_divisions_for_recip_mul (TYPE_MODE (TREE_TYPE (def
)));
484 if (count
>= threshold
)
487 for (occ
= occ_head
; occ
; occ
= occ
->next
)
490 insert_reciprocals (def_gsi
, occ
, def
, NULL
, threshold
);
493 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, def
)
495 if (is_division_by (use_stmt
, def
))
497 FOR_EACH_IMM_USE_ON_STMT (use_p
, use_iter
)
498 replace_reciprocal (use_p
);
503 for (occ
= occ_head
; occ
; )
509 /* Go through all the floating-point SSA_NAMEs, and call
510 execute_cse_reciprocals_1 on each of them. */
513 const pass_data pass_data_cse_reciprocals
=
515 GIMPLE_PASS
, /* type */
517 OPTGROUP_NONE
, /* optinfo_flags */
518 true, /* has_execute */
520 PROP_ssa
, /* properties_required */
521 0, /* properties_provided */
522 0, /* properties_destroyed */
523 0, /* todo_flags_start */
524 TODO_update_ssa
, /* todo_flags_finish */
527 class pass_cse_reciprocals
: public gimple_opt_pass
530 pass_cse_reciprocals (gcc::context
*ctxt
)
531 : gimple_opt_pass (pass_data_cse_reciprocals
, ctxt
)
534 /* opt_pass methods: */
535 virtual bool gate (function
*) { return optimize
&& flag_reciprocal_math
; }
536 virtual unsigned int execute (function
*);
538 }; // class pass_cse_reciprocals
541 pass_cse_reciprocals::execute (function
*fun
)
546 occ_pool
= create_alloc_pool ("dominators for recip",
547 sizeof (struct occurrence
),
548 n_basic_blocks_for_fn (fun
) / 3 + 1);
550 memset (&reciprocal_stats
, 0, sizeof (reciprocal_stats
));
551 calculate_dominance_info (CDI_DOMINATORS
);
552 calculate_dominance_info (CDI_POST_DOMINATORS
);
554 #ifdef ENABLE_CHECKING
555 FOR_EACH_BB_FN (bb
, fun
)
556 gcc_assert (!bb
->aux
);
559 for (arg
= DECL_ARGUMENTS (fun
->decl
); arg
; arg
= DECL_CHAIN (arg
))
560 if (FLOAT_TYPE_P (TREE_TYPE (arg
))
561 && is_gimple_reg (arg
))
563 tree name
= ssa_default_def (fun
, arg
);
565 execute_cse_reciprocals_1 (NULL
, name
);
568 FOR_EACH_BB_FN (bb
, fun
)
570 gimple_stmt_iterator gsi
;
574 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
576 phi
= gsi_stmt (gsi
);
577 def
= PHI_RESULT (phi
);
578 if (! virtual_operand_p (def
)
579 && FLOAT_TYPE_P (TREE_TYPE (def
)))
580 execute_cse_reciprocals_1 (NULL
, def
);
583 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
585 gimple stmt
= gsi_stmt (gsi
);
587 if (gimple_has_lhs (stmt
)
588 && (def
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_DEF
)) != NULL
589 && FLOAT_TYPE_P (TREE_TYPE (def
))
590 && TREE_CODE (def
) == SSA_NAME
)
591 execute_cse_reciprocals_1 (&gsi
, def
);
594 if (optimize_bb_for_size_p (bb
))
597 /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */
598 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
600 gimple stmt
= gsi_stmt (gsi
);
603 if (is_gimple_assign (stmt
)
604 && gimple_assign_rhs_code (stmt
) == RDIV_EXPR
)
606 tree arg1
= gimple_assign_rhs2 (stmt
);
609 if (TREE_CODE (arg1
) != SSA_NAME
)
612 stmt1
= SSA_NAME_DEF_STMT (arg1
);
614 if (is_gimple_call (stmt1
)
615 && gimple_call_lhs (stmt1
)
616 && (fndecl
= gimple_call_fndecl (stmt1
))
617 && (DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
618 || DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_MD
))
620 enum built_in_function code
;
625 code
= DECL_FUNCTION_CODE (fndecl
);
626 md_code
= DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_MD
;
628 fndecl
= targetm
.builtin_reciprocal (code
, md_code
, false);
632 /* Check that all uses of the SSA name are divisions,
633 otherwise replacing the defining statement will do
636 FOR_EACH_IMM_USE_FAST (use_p
, ui
, arg1
)
638 gimple stmt2
= USE_STMT (use_p
);
639 if (is_gimple_debug (stmt2
))
641 if (!is_gimple_assign (stmt2
)
642 || gimple_assign_rhs_code (stmt2
) != RDIV_EXPR
643 || gimple_assign_rhs1 (stmt2
) == arg1
644 || gimple_assign_rhs2 (stmt2
) != arg1
)
653 gimple_replace_ssa_lhs (stmt1
, arg1
);
654 gimple_call_set_fndecl (stmt1
, fndecl
);
656 reciprocal_stats
.rfuncs_inserted
++;
658 FOR_EACH_IMM_USE_STMT (stmt
, ui
, arg1
)
660 gimple_stmt_iterator gsi
= gsi_for_stmt (stmt
);
661 gimple_assign_set_rhs_code (stmt
, MULT_EXPR
);
662 fold_stmt_inplace (&gsi
);
670 statistics_counter_event (fun
, "reciprocal divs inserted",
671 reciprocal_stats
.rdivs_inserted
);
672 statistics_counter_event (fun
, "reciprocal functions inserted",
673 reciprocal_stats
.rfuncs_inserted
);
675 free_dominance_info (CDI_DOMINATORS
);
676 free_dominance_info (CDI_POST_DOMINATORS
);
677 free_alloc_pool (occ_pool
);
684 make_pass_cse_reciprocals (gcc::context
*ctxt
)
686 return new pass_cse_reciprocals (ctxt
);
689 /* Records an occurrence at statement USE_STMT in the vector of trees
690 STMTS if it is dominated by *TOP_BB or dominates it or this basic block
691 is not yet initialized. Returns true if the occurrence was pushed on
692 the vector. Adjusts *TOP_BB to be the basic block dominating all
693 statements in the vector. */
696 maybe_record_sincos (vec
<gimple
> *stmts
,
697 basic_block
*top_bb
, gimple use_stmt
)
699 basic_block use_bb
= gimple_bb (use_stmt
);
701 && (*top_bb
== use_bb
702 || dominated_by_p (CDI_DOMINATORS
, use_bb
, *top_bb
)))
703 stmts
->safe_push (use_stmt
);
705 || dominated_by_p (CDI_DOMINATORS
, *top_bb
, use_bb
))
707 stmts
->safe_push (use_stmt
);
716 /* Look for sin, cos and cexpi calls with the same argument NAME and
717 create a single call to cexpi CSEing the result in this case.
718 We first walk over all immediate uses of the argument collecting
719 statements that we can CSE in a vector and in a second pass replace
720 the statement rhs with a REALPART or IMAGPART expression on the
721 result of the cexpi call we insert before the use statement that
722 dominates all other candidates. */
725 execute_cse_sincos_1 (tree name
)
727 gimple_stmt_iterator gsi
;
728 imm_use_iterator use_iter
;
729 tree fndecl
, res
, type
;
730 gimple def_stmt
, use_stmt
, stmt
;
731 int seen_cos
= 0, seen_sin
= 0, seen_cexpi
= 0;
732 vec
<gimple
> stmts
= vNULL
;
733 basic_block top_bb
= NULL
;
735 bool cfg_changed
= false;
737 type
= TREE_TYPE (name
);
738 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, name
)
740 if (gimple_code (use_stmt
) != GIMPLE_CALL
741 || !gimple_call_lhs (use_stmt
)
742 || !(fndecl
= gimple_call_fndecl (use_stmt
))
743 || DECL_BUILT_IN_CLASS (fndecl
) != BUILT_IN_NORMAL
)
746 switch (DECL_FUNCTION_CODE (fndecl
))
748 CASE_FLT_FN (BUILT_IN_COS
):
749 seen_cos
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
752 CASE_FLT_FN (BUILT_IN_SIN
):
753 seen_sin
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
756 CASE_FLT_FN (BUILT_IN_CEXPI
):
757 seen_cexpi
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
764 if (seen_cos
+ seen_sin
+ seen_cexpi
<= 1)
770 /* Simply insert cexpi at the beginning of top_bb but not earlier than
771 the name def statement. */
772 fndecl
= mathfn_built_in (type
, BUILT_IN_CEXPI
);
775 stmt
= gimple_build_call (fndecl
, 1, name
);
776 res
= make_temp_ssa_name (TREE_TYPE (TREE_TYPE (fndecl
)), stmt
, "sincostmp");
777 gimple_call_set_lhs (stmt
, res
);
779 def_stmt
= SSA_NAME_DEF_STMT (name
);
780 if (!SSA_NAME_IS_DEFAULT_DEF (name
)
781 && gimple_code (def_stmt
) != GIMPLE_PHI
782 && gimple_bb (def_stmt
) == top_bb
)
784 gsi
= gsi_for_stmt (def_stmt
);
785 gsi_insert_after (&gsi
, stmt
, GSI_SAME_STMT
);
789 gsi
= gsi_after_labels (top_bb
);
790 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
792 sincos_stats
.inserted
++;
794 /* And adjust the recorded old call sites. */
795 for (i
= 0; stmts
.iterate (i
, &use_stmt
); ++i
)
798 fndecl
= gimple_call_fndecl (use_stmt
);
800 switch (DECL_FUNCTION_CODE (fndecl
))
802 CASE_FLT_FN (BUILT_IN_COS
):
803 rhs
= fold_build1 (REALPART_EXPR
, type
, res
);
806 CASE_FLT_FN (BUILT_IN_SIN
):
807 rhs
= fold_build1 (IMAGPART_EXPR
, type
, res
);
810 CASE_FLT_FN (BUILT_IN_CEXPI
):
818 /* Replace call with a copy. */
819 stmt
= gimple_build_assign (gimple_call_lhs (use_stmt
), rhs
);
821 gsi
= gsi_for_stmt (use_stmt
);
822 gsi_replace (&gsi
, stmt
, true);
823 if (gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
832 /* To evaluate powi(x,n), the floating point value x raised to the
833 constant integer exponent n, we use a hybrid algorithm that
834 combines the "window method" with look-up tables. For an
835 introduction to exponentiation algorithms and "addition chains",
836 see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth,
837 "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming",
838 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation
839 Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */
841 /* Provide a default value for POWI_MAX_MULTS, the maximum number of
842 multiplications to inline before calling the system library's pow
843 function. powi(x,n) requires at worst 2*bits(n)-2 multiplications,
844 so this default never requires calling pow, powf or powl. */
846 #ifndef POWI_MAX_MULTS
847 #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2)
850 /* The size of the "optimal power tree" lookup table. All
851 exponents less than this value are simply looked up in the
852 powi_table below. This threshold is also used to size the
853 cache of pseudo registers that hold intermediate results. */
854 #define POWI_TABLE_SIZE 256
856 /* The size, in bits of the window, used in the "window method"
857 exponentiation algorithm. This is equivalent to a radix of
858 (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */
859 #define POWI_WINDOW_SIZE 3
861 /* The following table is an efficient representation of an
862 "optimal power tree". For each value, i, the corresponding
863 value, j, in the table states than an optimal evaluation
864 sequence for calculating pow(x,i) can be found by evaluating
865 pow(x,j)*pow(x,i-j). An optimal power tree for the first
866 100 integers is given in Knuth's "Seminumerical algorithms". */
868 static const unsigned char powi_table
[POWI_TABLE_SIZE
] =
870 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */
871 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */
872 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */
873 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */
874 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */
875 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */
876 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */
877 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */
878 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */
879 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */
880 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */
881 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */
882 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */
883 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */
884 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */
885 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */
886 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */
887 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */
888 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */
889 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */
890 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */
891 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */
892 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */
893 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */
894 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */
895 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */
896 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */
897 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */
898 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */
899 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */
900 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */
901 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */
905 /* Return the number of multiplications required to calculate
906 powi(x,n) where n is less than POWI_TABLE_SIZE. This is a
907 subroutine of powi_cost. CACHE is an array indicating
908 which exponents have already been calculated. */
911 powi_lookup_cost (unsigned HOST_WIDE_INT n
, bool *cache
)
913 /* If we've already calculated this exponent, then this evaluation
914 doesn't require any additional multiplications. */
919 return powi_lookup_cost (n
- powi_table
[n
], cache
)
920 + powi_lookup_cost (powi_table
[n
], cache
) + 1;
923 /* Return the number of multiplications required to calculate
924 powi(x,n) for an arbitrary x, given the exponent N. This
925 function needs to be kept in sync with powi_as_mults below. */
928 powi_cost (HOST_WIDE_INT n
)
930 bool cache
[POWI_TABLE_SIZE
];
931 unsigned HOST_WIDE_INT digit
;
932 unsigned HOST_WIDE_INT val
;
938 /* Ignore the reciprocal when calculating the cost. */
939 val
= (n
< 0) ? -n
: n
;
941 /* Initialize the exponent cache. */
942 memset (cache
, 0, POWI_TABLE_SIZE
* sizeof (bool));
947 while (val
>= POWI_TABLE_SIZE
)
951 digit
= val
& ((1 << POWI_WINDOW_SIZE
) - 1);
952 result
+= powi_lookup_cost (digit
, cache
)
953 + POWI_WINDOW_SIZE
+ 1;
954 val
>>= POWI_WINDOW_SIZE
;
963 return result
+ powi_lookup_cost (val
, cache
);
966 /* Recursive subroutine of powi_as_mults. This function takes the
967 array, CACHE, of already calculated exponents and an exponent N and
968 returns a tree that corresponds to CACHE[1]**N, with type TYPE. */
971 powi_as_mults_1 (gimple_stmt_iterator
*gsi
, location_t loc
, tree type
,
972 HOST_WIDE_INT n
, tree
*cache
)
974 tree op0
, op1
, ssa_target
;
975 unsigned HOST_WIDE_INT digit
;
978 if (n
< POWI_TABLE_SIZE
&& cache
[n
])
981 ssa_target
= make_temp_ssa_name (type
, NULL
, "powmult");
983 if (n
< POWI_TABLE_SIZE
)
985 cache
[n
] = ssa_target
;
986 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
- powi_table
[n
], cache
);
987 op1
= powi_as_mults_1 (gsi
, loc
, type
, powi_table
[n
], cache
);
991 digit
= n
& ((1 << POWI_WINDOW_SIZE
) - 1);
992 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
- digit
, cache
);
993 op1
= powi_as_mults_1 (gsi
, loc
, type
, digit
, cache
);
997 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
>> 1, cache
);
1001 mult_stmt
= gimple_build_assign_with_ops (MULT_EXPR
, ssa_target
, op0
, op1
);
1002 gimple_set_location (mult_stmt
, loc
);
1003 gsi_insert_before (gsi
, mult_stmt
, GSI_SAME_STMT
);
1008 /* Convert ARG0**N to a tree of multiplications of ARG0 with itself.
1009 This function needs to be kept in sync with powi_cost above. */
1012 powi_as_mults (gimple_stmt_iterator
*gsi
, location_t loc
,
1013 tree arg0
, HOST_WIDE_INT n
)
1015 tree cache
[POWI_TABLE_SIZE
], result
, type
= TREE_TYPE (arg0
);
1020 return build_real (type
, dconst1
);
1022 memset (cache
, 0, sizeof (cache
));
1025 result
= powi_as_mults_1 (gsi
, loc
, type
, (n
< 0) ? -n
: n
, cache
);
1029 /* If the original exponent was negative, reciprocate the result. */
1030 target
= make_temp_ssa_name (type
, NULL
, "powmult");
1031 div_stmt
= gimple_build_assign_with_ops (RDIV_EXPR
, target
,
1032 build_real (type
, dconst1
),
1034 gimple_set_location (div_stmt
, loc
);
1035 gsi_insert_before (gsi
, div_stmt
, GSI_SAME_STMT
);
1040 /* ARG0 and N are the two arguments to a powi builtin in GSI with
1041 location info LOC. If the arguments are appropriate, create an
1042 equivalent sequence of statements prior to GSI using an optimal
1043 number of multiplications, and return an expession holding the
1047 gimple_expand_builtin_powi (gimple_stmt_iterator
*gsi
, location_t loc
,
1048 tree arg0
, HOST_WIDE_INT n
)
1050 /* Avoid largest negative number. */
1052 && ((n
>= -1 && n
<= 2)
1053 || (optimize_function_for_speed_p (cfun
)
1054 && powi_cost (n
) <= POWI_MAX_MULTS
)))
1055 return powi_as_mults (gsi
, loc
, arg0
, n
);
1060 /* Build a gimple call statement that calls FN with argument ARG.
1061 Set the lhs of the call statement to a fresh SSA name. Insert the
1062 statement prior to GSI's current position, and return the fresh
1066 build_and_insert_call (gimple_stmt_iterator
*gsi
, location_t loc
,
1072 call_stmt
= gimple_build_call (fn
, 1, arg
);
1073 ssa_target
= make_temp_ssa_name (TREE_TYPE (arg
), NULL
, "powroot");
1074 gimple_set_lhs (call_stmt
, ssa_target
);
1075 gimple_set_location (call_stmt
, loc
);
1076 gsi_insert_before (gsi
, call_stmt
, GSI_SAME_STMT
);
1081 /* Build a gimple binary operation with the given CODE and arguments
1082 ARG0, ARG1, assigning the result to a new SSA name for variable
1083 TARGET. Insert the statement prior to GSI's current position, and
1084 return the fresh SSA name.*/
1087 build_and_insert_binop (gimple_stmt_iterator
*gsi
, location_t loc
,
1088 const char *name
, enum tree_code code
,
1089 tree arg0
, tree arg1
)
1091 tree result
= make_temp_ssa_name (TREE_TYPE (arg0
), NULL
, name
);
1092 gimple stmt
= gimple_build_assign_with_ops (code
, result
, arg0
, arg1
);
1093 gimple_set_location (stmt
, loc
);
1094 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1098 /* Build a gimple reference operation with the given CODE and argument
1099 ARG, assigning the result to a new SSA name of TYPE with NAME.
1100 Insert the statement prior to GSI's current position, and return
1101 the fresh SSA name. */
1104 build_and_insert_ref (gimple_stmt_iterator
*gsi
, location_t loc
, tree type
,
1105 const char *name
, enum tree_code code
, tree arg0
)
1107 tree result
= make_temp_ssa_name (type
, NULL
, name
);
1108 gimple stmt
= gimple_build_assign (result
, build1 (code
, type
, arg0
));
1109 gimple_set_location (stmt
, loc
);
1110 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1114 /* Build a gimple assignment to cast VAL to TYPE. Insert the statement
1115 prior to GSI's current position, and return the fresh SSA name. */
1118 build_and_insert_cast (gimple_stmt_iterator
*gsi
, location_t loc
,
1119 tree type
, tree val
)
1121 tree result
= make_ssa_name (type
, NULL
);
1122 gimple stmt
= gimple_build_assign_with_ops (NOP_EXPR
, result
, val
, NULL_TREE
);
1123 gimple_set_location (stmt
, loc
);
1124 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1128 /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI
1129 with location info LOC. If possible, create an equivalent and
1130 less expensive sequence of statements prior to GSI, and return an
1131 expession holding the result. */
1134 gimple_expand_builtin_pow (gimple_stmt_iterator
*gsi
, location_t loc
,
1135 tree arg0
, tree arg1
)
1137 REAL_VALUE_TYPE c
, cint
, dconst1_4
, dconst3_4
, dconst1_3
, dconst1_6
;
1138 REAL_VALUE_TYPE c2
, dconst3
;
1140 tree type
, sqrtfn
, cbrtfn
, sqrt_arg0
, sqrt_sqrt
, result
, cbrt_x
, powi_cbrt_x
;
1141 enum machine_mode mode
;
1142 bool hw_sqrt_exists
, c_is_int
, c2_is_int
;
1144 /* If the exponent isn't a constant, there's nothing of interest
1146 if (TREE_CODE (arg1
) != REAL_CST
)
1149 /* If the exponent is equivalent to an integer, expand to an optimal
1150 multiplication sequence when profitable. */
1151 c
= TREE_REAL_CST (arg1
);
1152 n
= real_to_integer (&c
);
1153 real_from_integer (&cint
, VOIDmode
, n
, SIGNED
);
1154 c_is_int
= real_identical (&c
, &cint
);
1157 && ((n
>= -1 && n
<= 2)
1158 || (flag_unsafe_math_optimizations
1159 && optimize_bb_for_speed_p (gsi_bb (*gsi
))
1160 && powi_cost (n
) <= POWI_MAX_MULTS
)))
1161 return gimple_expand_builtin_powi (gsi
, loc
, arg0
, n
);
1163 /* Attempt various optimizations using sqrt and cbrt. */
1164 type
= TREE_TYPE (arg0
);
1165 mode
= TYPE_MODE (type
);
1166 sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1168 /* Optimize pow(x,0.5) = sqrt(x). This replacement is always safe
1169 unless signed zeros must be maintained. pow(-0,0.5) = +0, while
1172 && REAL_VALUES_EQUAL (c
, dconsthalf
)
1173 && !HONOR_SIGNED_ZEROS (mode
))
1174 return build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1176 /* Optimize pow(x,0.25) = sqrt(sqrt(x)). Assume on most machines that
1177 a builtin sqrt instruction is smaller than a call to pow with 0.25,
1178 so do this optimization even if -Os. Don't do this optimization
1179 if we don't have a hardware sqrt insn. */
1180 dconst1_4
= dconst1
;
1181 SET_REAL_EXP (&dconst1_4
, REAL_EXP (&dconst1_4
) - 2);
1182 hw_sqrt_exists
= optab_handler (sqrt_optab
, mode
) != CODE_FOR_nothing
;
1184 if (flag_unsafe_math_optimizations
1186 && REAL_VALUES_EQUAL (c
, dconst1_4
)
1190 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1193 return build_and_insert_call (gsi
, loc
, sqrtfn
, sqrt_arg0
);
1196 /* Optimize pow(x,0.75) = sqrt(x) * sqrt(sqrt(x)) unless we are
1197 optimizing for space. Don't do this optimization if we don't have
1198 a hardware sqrt insn. */
1199 real_from_integer (&dconst3_4
, VOIDmode
, 3, SIGNED
);
1200 SET_REAL_EXP (&dconst3_4
, REAL_EXP (&dconst3_4
) - 2);
1202 if (flag_unsafe_math_optimizations
1204 && optimize_function_for_speed_p (cfun
)
1205 && REAL_VALUES_EQUAL (c
, dconst3_4
)
1209 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1212 sqrt_sqrt
= build_and_insert_call (gsi
, loc
, sqrtfn
, sqrt_arg0
);
1214 /* sqrt(x) * sqrt(sqrt(x)) */
1215 return build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1216 sqrt_arg0
, sqrt_sqrt
);
1219 /* Optimize pow(x,1./3.) = cbrt(x). This requires unsafe math
1220 optimizations since 1./3. is not exactly representable. If x
1221 is negative and finite, the correct value of pow(x,1./3.) is
1222 a NaN with the "invalid" exception raised, because the value
1223 of 1./3. actually has an even denominator. The correct value
1224 of cbrt(x) is a negative real value. */
1225 cbrtfn
= mathfn_built_in (type
, BUILT_IN_CBRT
);
1226 dconst1_3
= real_value_truncate (mode
, dconst_third ());
1228 if (flag_unsafe_math_optimizations
1230 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1231 && REAL_VALUES_EQUAL (c
, dconst1_3
))
1232 return build_and_insert_call (gsi
, loc
, cbrtfn
, arg0
);
1234 /* Optimize pow(x,1./6.) = cbrt(sqrt(x)). Don't do this optimization
1235 if we don't have a hardware sqrt insn. */
1236 dconst1_6
= dconst1_3
;
1237 SET_REAL_EXP (&dconst1_6
, REAL_EXP (&dconst1_6
) - 1);
1239 if (flag_unsafe_math_optimizations
1242 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1243 && optimize_function_for_speed_p (cfun
)
1245 && REAL_VALUES_EQUAL (c
, dconst1_6
))
1248 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1251 return build_and_insert_call (gsi
, loc
, cbrtfn
, sqrt_arg0
);
1254 /* Optimize pow(x,c), where n = 2c for some nonzero integer n
1255 and c not an integer, into
1257 sqrt(x) * powi(x, n/2), n > 0;
1258 1.0 / (sqrt(x) * powi(x, abs(n/2))), n < 0.
1260 Do not calculate the powi factor when n/2 = 0. */
1261 real_arithmetic (&c2
, MULT_EXPR
, &c
, &dconst2
);
1262 n
= real_to_integer (&c2
);
1263 real_from_integer (&cint
, VOIDmode
, n
, SIGNED
);
1264 c2_is_int
= real_identical (&c2
, &cint
);
1266 if (flag_unsafe_math_optimizations
1270 && optimize_function_for_speed_p (cfun
))
1272 tree powi_x_ndiv2
= NULL_TREE
;
1274 /* Attempt to fold powi(arg0, abs(n/2)) into multiplies. If not
1275 possible or profitable, give up. Skip the degenerate case when
1276 n is 1 or -1, where the result is always 1. */
1277 if (absu_hwi (n
) != 1)
1279 powi_x_ndiv2
= gimple_expand_builtin_powi (gsi
, loc
, arg0
,
1285 /* Calculate sqrt(x). When n is not 1 or -1, multiply it by the
1286 result of the optimal multiply sequence just calculated. */
1287 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1289 if (absu_hwi (n
) == 1)
1292 result
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1293 sqrt_arg0
, powi_x_ndiv2
);
1295 /* If n is negative, reciprocate the result. */
1297 result
= build_and_insert_binop (gsi
, loc
, "powroot", RDIV_EXPR
,
1298 build_real (type
, dconst1
), result
);
1302 /* Optimize pow(x,c), where 3c = n for some nonzero integer n, into
1304 powi(x, n/3) * powi(cbrt(x), n%3), n > 0;
1305 1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)), n < 0.
1307 Do not calculate the first factor when n/3 = 0. As cbrt(x) is
1308 different from pow(x, 1./3.) due to rounding and behavior with
1309 negative x, we need to constrain this transformation to unsafe
1310 math and positive x or finite math. */
1311 real_from_integer (&dconst3
, VOIDmode
, 3, SIGNED
);
1312 real_arithmetic (&c2
, MULT_EXPR
, &c
, &dconst3
);
1313 real_round (&c2
, mode
, &c2
);
1314 n
= real_to_integer (&c2
);
1315 real_from_integer (&cint
, VOIDmode
, n
, SIGNED
);
1316 real_arithmetic (&c2
, RDIV_EXPR
, &cint
, &dconst3
);
1317 real_convert (&c2
, mode
, &c2
);
1319 if (flag_unsafe_math_optimizations
1321 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1322 && real_identical (&c2
, &c
)
1324 && optimize_function_for_speed_p (cfun
)
1325 && powi_cost (n
/ 3) <= POWI_MAX_MULTS
)
1327 tree powi_x_ndiv3
= NULL_TREE
;
1329 /* Attempt to fold powi(arg0, abs(n/3)) into multiplies. If not
1330 possible or profitable, give up. Skip the degenerate case when
1331 abs(n) < 3, where the result is always 1. */
1332 if (absu_hwi (n
) >= 3)
1334 powi_x_ndiv3
= gimple_expand_builtin_powi (gsi
, loc
, arg0
,
1340 /* Calculate powi(cbrt(x), n%3). Don't use gimple_expand_builtin_powi
1341 as that creates an unnecessary variable. Instead, just produce
1342 either cbrt(x) or cbrt(x) * cbrt(x). */
1343 cbrt_x
= build_and_insert_call (gsi
, loc
, cbrtfn
, arg0
);
1345 if (absu_hwi (n
) % 3 == 1)
1346 powi_cbrt_x
= cbrt_x
;
1348 powi_cbrt_x
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1351 /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1. */
1352 if (absu_hwi (n
) < 3)
1353 result
= powi_cbrt_x
;
1355 result
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1356 powi_x_ndiv3
, powi_cbrt_x
);
1358 /* If n is negative, reciprocate the result. */
1360 result
= build_and_insert_binop (gsi
, loc
, "powroot", RDIV_EXPR
,
1361 build_real (type
, dconst1
), result
);
1366 /* No optimizations succeeded. */
1370 /* ARG is the argument to a cabs builtin call in GSI with location info
1371 LOC. Create a sequence of statements prior to GSI that calculates
1372 sqrt(R*R + I*I), where R and I are the real and imaginary components
1373 of ARG, respectively. Return an expression holding the result. */
1376 gimple_expand_builtin_cabs (gimple_stmt_iterator
*gsi
, location_t loc
, tree arg
)
1378 tree real_part
, imag_part
, addend1
, addend2
, sum
, result
;
1379 tree type
= TREE_TYPE (TREE_TYPE (arg
));
1380 tree sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1381 enum machine_mode mode
= TYPE_MODE (type
);
1383 if (!flag_unsafe_math_optimizations
1384 || !optimize_bb_for_speed_p (gimple_bb (gsi_stmt (*gsi
)))
1386 || optab_handler (sqrt_optab
, mode
) == CODE_FOR_nothing
)
1389 real_part
= build_and_insert_ref (gsi
, loc
, type
, "cabs",
1390 REALPART_EXPR
, arg
);
1391 addend1
= build_and_insert_binop (gsi
, loc
, "cabs", MULT_EXPR
,
1392 real_part
, real_part
);
1393 imag_part
= build_and_insert_ref (gsi
, loc
, type
, "cabs",
1394 IMAGPART_EXPR
, arg
);
1395 addend2
= build_and_insert_binop (gsi
, loc
, "cabs", MULT_EXPR
,
1396 imag_part
, imag_part
);
1397 sum
= build_and_insert_binop (gsi
, loc
, "cabs", PLUS_EXPR
, addend1
, addend2
);
1398 result
= build_and_insert_call (gsi
, loc
, sqrtfn
, sum
);
1403 /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1
1404 on the SSA_NAME argument of each of them. Also expand powi(x,n) into
1405 an optimal number of multiplies, when n is a constant. */
1409 const pass_data pass_data_cse_sincos
=
1411 GIMPLE_PASS
, /* type */
1412 "sincos", /* name */
1413 OPTGROUP_NONE
, /* optinfo_flags */
1414 true, /* has_execute */
1415 TV_NONE
, /* tv_id */
1416 PROP_ssa
, /* properties_required */
1417 0, /* properties_provided */
1418 0, /* properties_destroyed */
1419 0, /* todo_flags_start */
1420 TODO_update_ssa
, /* todo_flags_finish */
1423 class pass_cse_sincos
: public gimple_opt_pass
1426 pass_cse_sincos (gcc::context
*ctxt
)
1427 : gimple_opt_pass (pass_data_cse_sincos
, ctxt
)
1430 /* opt_pass methods: */
1431 virtual bool gate (function
*)
1433 /* We no longer require either sincos or cexp, since powi expansion
1434 piggybacks on this pass. */
1438 virtual unsigned int execute (function
*);
1440 }; // class pass_cse_sincos
1443 pass_cse_sincos::execute (function
*fun
)
1446 bool cfg_changed
= false;
1448 calculate_dominance_info (CDI_DOMINATORS
);
1449 memset (&sincos_stats
, 0, sizeof (sincos_stats
));
1451 FOR_EACH_BB_FN (bb
, fun
)
1453 gimple_stmt_iterator gsi
;
1454 bool cleanup_eh
= false;
1456 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1458 gimple stmt
= gsi_stmt (gsi
);
1461 /* Only the last stmt in a bb could throw, no need to call
1462 gimple_purge_dead_eh_edges if we change something in the middle
1463 of a basic block. */
1466 if (is_gimple_call (stmt
)
1467 && gimple_call_lhs (stmt
)
1468 && (fndecl
= gimple_call_fndecl (stmt
))
1469 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
)
1471 tree arg
, arg0
, arg1
, result
;
1475 switch (DECL_FUNCTION_CODE (fndecl
))
1477 CASE_FLT_FN (BUILT_IN_COS
):
1478 CASE_FLT_FN (BUILT_IN_SIN
):
1479 CASE_FLT_FN (BUILT_IN_CEXPI
):
1480 /* Make sure we have either sincos or cexp. */
1481 if (!targetm
.libc_has_function (function_c99_math_complex
)
1482 && !targetm
.libc_has_function (function_sincos
))
1485 arg
= gimple_call_arg (stmt
, 0);
1486 if (TREE_CODE (arg
) == SSA_NAME
)
1487 cfg_changed
|= execute_cse_sincos_1 (arg
);
1490 CASE_FLT_FN (BUILT_IN_POW
):
1491 arg0
= gimple_call_arg (stmt
, 0);
1492 arg1
= gimple_call_arg (stmt
, 1);
1494 loc
= gimple_location (stmt
);
1495 result
= gimple_expand_builtin_pow (&gsi
, loc
, arg0
, arg1
);
1499 tree lhs
= gimple_get_lhs (stmt
);
1500 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1501 gimple_set_location (new_stmt
, loc
);
1502 unlink_stmt_vdef (stmt
);
1503 gsi_replace (&gsi
, new_stmt
, true);
1505 if (gimple_vdef (stmt
))
1506 release_ssa_name (gimple_vdef (stmt
));
1510 CASE_FLT_FN (BUILT_IN_POWI
):
1511 arg0
= gimple_call_arg (stmt
, 0);
1512 arg1
= gimple_call_arg (stmt
, 1);
1513 loc
= gimple_location (stmt
);
1515 if (real_minus_onep (arg0
))
1517 tree t0
, t1
, cond
, one
, minus_one
;
1520 t0
= TREE_TYPE (arg0
);
1521 t1
= TREE_TYPE (arg1
);
1522 one
= build_real (t0
, dconst1
);
1523 minus_one
= build_real (t0
, dconstm1
);
1525 cond
= make_temp_ssa_name (t1
, NULL
, "powi_cond");
1526 stmt
= gimple_build_assign_with_ops (BIT_AND_EXPR
, cond
,
1530 gimple_set_location (stmt
, loc
);
1531 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1533 result
= make_temp_ssa_name (t0
, NULL
, "powi");
1534 stmt
= gimple_build_assign_with_ops (COND_EXPR
, result
,
1537 gimple_set_location (stmt
, loc
);
1538 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1542 if (!tree_fits_shwi_p (arg1
))
1545 n
= tree_to_shwi (arg1
);
1546 result
= gimple_expand_builtin_powi (&gsi
, loc
, arg0
, n
);
1551 tree lhs
= gimple_get_lhs (stmt
);
1552 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1553 gimple_set_location (new_stmt
, loc
);
1554 unlink_stmt_vdef (stmt
);
1555 gsi_replace (&gsi
, new_stmt
, true);
1557 if (gimple_vdef (stmt
))
1558 release_ssa_name (gimple_vdef (stmt
));
1562 CASE_FLT_FN (BUILT_IN_CABS
):
1563 arg0
= gimple_call_arg (stmt
, 0);
1564 loc
= gimple_location (stmt
);
1565 result
= gimple_expand_builtin_cabs (&gsi
, loc
, arg0
);
1569 tree lhs
= gimple_get_lhs (stmt
);
1570 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1571 gimple_set_location (new_stmt
, loc
);
1572 unlink_stmt_vdef (stmt
);
1573 gsi_replace (&gsi
, new_stmt
, true);
1575 if (gimple_vdef (stmt
))
1576 release_ssa_name (gimple_vdef (stmt
));
1585 cfg_changed
|= gimple_purge_dead_eh_edges (bb
);
1588 statistics_counter_event (fun
, "sincos statements inserted",
1589 sincos_stats
.inserted
);
1591 free_dominance_info (CDI_DOMINATORS
);
1592 return cfg_changed
? TODO_cleanup_cfg
: 0;
1598 make_pass_cse_sincos (gcc::context
*ctxt
)
1600 return new pass_cse_sincos (ctxt
);
1603 /* A symbolic number is used to detect byte permutation and selection
1604 patterns. Therefore the field N contains an artificial number
1605 consisting of byte size markers:
1607 0 - byte has the value 0
1608 1..size - byte contains the content of the byte
1609 number indexed with that value minus one.
1611 To detect permutations on memory sources (arrays and structures), a symbolic
1612 number is also associated a base address (the array or structure the load is
1613 made from), an offset from the base address and a range which gives the
1614 difference between the highest and lowest accessed memory location to make
1615 such a symbolic number. The range is thus different from size which reflects
1616 the size of the type of current expression. Note that for non memory source,
1617 range holds the same value as size.
1619 For instance, for an array char a[], (short) a[0] | (short) a[3] would have
1620 a size of 2 but a range of 4 while (short) a[0] | ((short) a[0] << 1) would
1621 still have a size of 2 but this time a range of 1. */
1623 struct symbolic_number
{
1628 HOST_WIDE_INT bytepos
;
1631 unsigned HOST_WIDE_INT range
;
1634 /* The number which the find_bswap_or_nop_1 result should match in
1635 order to have a nop. The number is masked according to the size of
1636 the symbolic number before using it. */
1637 #define CMPNOP (sizeof (int64_t) < 8 ? 0 : \
1638 (uint64_t)0x08070605 << 32 | 0x04030201)
1640 /* The number which the find_bswap_or_nop_1 result should match in
1641 order to have a byte swap. The number is masked according to the
1642 size of the symbolic number before using it. */
1643 #define CMPXCHG (sizeof (int64_t) < 8 ? 0 : \
1644 (uint64_t)0x01020304 << 32 | 0x05060708)
1646 /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic
1647 number N. Return false if the requested operation is not permitted
1648 on a symbolic number. */
1651 do_shift_rotate (enum tree_code code
,
1652 struct symbolic_number
*n
,
1655 int bitsize
= TYPE_PRECISION (n
->type
);
1660 /* Zero out the extra bits of N in order to avoid them being shifted
1661 into the significant bits. */
1662 if (bitsize
< 8 * (int)sizeof (int64_t))
1663 n
->n
&= ((uint64_t)1 << bitsize
) - 1;
1671 /* Arithmetic shift of signed type: result is dependent on the value. */
1672 if (!TYPE_UNSIGNED (n
->type
)
1673 && (n
->n
& ((uint64_t) 0xff << (bitsize
- 8))))
1678 n
->n
= (n
->n
<< count
) | (n
->n
>> (bitsize
- count
));
1681 n
->n
= (n
->n
>> count
) | (n
->n
<< (bitsize
- count
));
1686 /* Zero unused bits for size. */
1687 if (bitsize
< 8 * (int)sizeof (int64_t))
1688 n
->n
&= ((uint64_t)1 << bitsize
) - 1;
1692 /* Perform sanity checking for the symbolic number N and the gimple
1696 verify_symbolic_number_p (struct symbolic_number
*n
, gimple stmt
)
1700 lhs_type
= gimple_expr_type (stmt
);
1702 if (TREE_CODE (lhs_type
) != INTEGER_TYPE
)
1705 if (TYPE_PRECISION (lhs_type
) != TYPE_PRECISION (n
->type
))
1711 /* Initialize the symbolic number N for the bswap pass from the base element
1712 SRC manipulated by the bitwise OR expression. */
1715 init_symbolic_number (struct symbolic_number
*n
, tree src
)
1719 n
->base_addr
= n
->offset
= n
->alias_set
= n
->vuse
= NULL_TREE
;
1721 /* Set up the symbolic number N by setting each byte to a value between 1 and
1722 the byte size of rhs1. The highest order byte is set to n->size and the
1723 lowest order byte to 1. */
1724 n
->type
= TREE_TYPE (src
);
1725 size
= TYPE_PRECISION (n
->type
);
1726 if (size
% BITS_PER_UNIT
!= 0)
1728 size
/= BITS_PER_UNIT
;
1729 if (size
> (int)sizeof (uint64_t))
1734 if (size
< (int)sizeof (int64_t))
1735 n
->n
&= ((uint64_t)1 << (size
* BITS_PER_UNIT
)) - 1;
1740 /* Check if STMT might be a byte swap or a nop from a memory source and returns
1741 the answer. If so, REF is that memory source and the base of the memory area
1742 accessed and the offset of the access from that base are recorded in N. */
1745 find_bswap_or_nop_load (gimple stmt
, tree ref
, struct symbolic_number
*n
)
1747 /* Leaf node is an array or component ref. Memorize its base and
1748 offset from base to compare to other such leaf node. */
1749 HOST_WIDE_INT bitsize
, bitpos
;
1750 enum machine_mode mode
;
1751 int unsignedp
, volatilep
;
1752 tree offset
, base_addr
;
1754 if (!gimple_assign_load_p (stmt
) || gimple_has_volatile_ops (stmt
))
1757 base_addr
= get_inner_reference (ref
, &bitsize
, &bitpos
, &offset
, &mode
,
1758 &unsignedp
, &volatilep
, false);
1760 if (TREE_CODE (base_addr
) == MEM_REF
)
1762 offset_int bit_offset
= 0;
1763 tree off
= TREE_OPERAND (base_addr
, 1);
1765 if (!integer_zerop (off
))
1767 offset_int boff
, coff
= mem_ref_offset (base_addr
);
1768 boff
= wi::lshift (coff
, LOG2_BITS_PER_UNIT
);
1772 base_addr
= TREE_OPERAND (base_addr
, 0);
1774 /* Avoid returning a negative bitpos as this may wreak havoc later. */
1775 if (wi::neg_p (bit_offset
))
1777 offset_int mask
= wi::mask
<offset_int
> (LOG2_BITS_PER_UNIT
, false);
1778 offset_int tem
= bit_offset
.and_not (mask
);
1779 /* TEM is the bitpos rounded to BITS_PER_UNIT towards -Inf.
1780 Subtract it to BIT_OFFSET and add it (scaled) to OFFSET. */
1782 tem
= wi::arshift (tem
, LOG2_BITS_PER_UNIT
);
1784 offset
= size_binop (PLUS_EXPR
, offset
,
1785 wide_int_to_tree (sizetype
, tem
));
1787 offset
= wide_int_to_tree (sizetype
, tem
);
1790 bitpos
+= bit_offset
.to_shwi ();
1793 if (bitpos
% BITS_PER_UNIT
)
1795 if (bitsize
% BITS_PER_UNIT
)
1798 if (!init_symbolic_number (n
, ref
))
1800 n
->base_addr
= base_addr
;
1802 n
->bytepos
= bitpos
/ BITS_PER_UNIT
;
1803 n
->alias_set
= reference_alias_ptr_type (ref
);
1804 n
->vuse
= gimple_vuse (stmt
);
1808 /* find_bswap_or_nop_1 invokes itself recursively with N and tries to perform
1809 the operation given by the rhs of STMT on the result. If the operation
1810 could successfully be executed the function returns a gimple stmt whose
1811 rhs's first tree is the expression of the source operand and NULL
1815 find_bswap_or_nop_1 (gimple stmt
, struct symbolic_number
*n
, int limit
)
1817 enum tree_code code
;
1818 tree rhs1
, rhs2
= NULL
;
1819 gimple rhs1_stmt
, rhs2_stmt
, source_stmt1
;
1820 enum gimple_rhs_class rhs_class
;
1822 if (!limit
|| !is_gimple_assign (stmt
))
1825 rhs1
= gimple_assign_rhs1 (stmt
);
1827 if (find_bswap_or_nop_load (stmt
, rhs1
, n
))
1830 if (TREE_CODE (rhs1
) != SSA_NAME
)
1833 code
= gimple_assign_rhs_code (stmt
);
1834 rhs_class
= gimple_assign_rhs_class (stmt
);
1835 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
1837 if (rhs_class
== GIMPLE_BINARY_RHS
)
1838 rhs2
= gimple_assign_rhs2 (stmt
);
1840 /* Handle unary rhs and binary rhs with integer constants as second
1843 if (rhs_class
== GIMPLE_UNARY_RHS
1844 || (rhs_class
== GIMPLE_BINARY_RHS
1845 && TREE_CODE (rhs2
) == INTEGER_CST
))
1847 if (code
!= BIT_AND_EXPR
1848 && code
!= LSHIFT_EXPR
1849 && code
!= RSHIFT_EXPR
1850 && code
!= LROTATE_EXPR
1851 && code
!= RROTATE_EXPR
1853 && code
!= CONVERT_EXPR
)
1856 source_stmt1
= find_bswap_or_nop_1 (rhs1_stmt
, n
, limit
- 1);
1858 /* If find_bswap_or_nop_1 returned NULL, STMT is a leaf node and
1859 we have to initialize the symbolic number. */
1862 if (gimple_assign_load_p (stmt
)
1863 || !init_symbolic_number (n
, rhs1
))
1865 source_stmt1
= stmt
;
1872 int i
, size
= TYPE_PRECISION (n
->type
) / BITS_PER_UNIT
;
1873 uint64_t val
= int_cst_value (rhs2
);
1876 /* Only constants masking full bytes are allowed. */
1877 for (i
= 0; i
< size
; i
++, tmp
>>= BITS_PER_UNIT
)
1878 if ((tmp
& 0xff) != 0 && (tmp
& 0xff) != 0xff)
1888 if (!do_shift_rotate (code
, n
, (int)TREE_INT_CST_LOW (rhs2
)))
1893 int type_size
, old_type_size
;
1896 type
= gimple_expr_type (stmt
);
1897 type_size
= TYPE_PRECISION (type
);
1898 if (type_size
% BITS_PER_UNIT
!= 0)
1900 if (type_size
> (int)sizeof (uint64_t) * 8)
1903 /* Sign extension: result is dependent on the value. */
1904 old_type_size
= TYPE_PRECISION (n
->type
);
1905 if (!TYPE_UNSIGNED (n
->type
)
1906 && type_size
> old_type_size
1907 && n
->n
& ((uint64_t) 0xff << (old_type_size
- 8)))
1910 if (type_size
/ BITS_PER_UNIT
< (int)(sizeof (int64_t)))
1912 /* If STMT casts to a smaller type mask out the bits not
1913 belonging to the target type. */
1914 n
->n
&= ((uint64_t)1 << type_size
) - 1;
1918 n
->range
= type_size
/ BITS_PER_UNIT
;
1924 return verify_symbolic_number_p (n
, stmt
) ? source_stmt1
: NULL
;
1927 /* Handle binary rhs. */
1929 if (rhs_class
== GIMPLE_BINARY_RHS
)
1932 struct symbolic_number n1
, n2
;
1934 gimple source_stmt2
;
1936 if (code
!= BIT_IOR_EXPR
)
1939 if (TREE_CODE (rhs2
) != SSA_NAME
)
1942 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
1947 source_stmt1
= find_bswap_or_nop_1 (rhs1_stmt
, &n1
, limit
- 1);
1952 source_stmt2
= find_bswap_or_nop_1 (rhs2_stmt
, &n2
, limit
- 1);
1957 if (TYPE_PRECISION (n1
.type
) != TYPE_PRECISION (n2
.type
))
1960 if (!n1
.vuse
!= !n2
.vuse
||
1961 (n1
.vuse
&& !operand_equal_p (n1
.vuse
, n2
.vuse
, 0)))
1964 if (gimple_assign_rhs1 (source_stmt1
)
1965 != gimple_assign_rhs1 (source_stmt2
))
1969 HOST_WIDE_INT off_sub
;
1970 struct symbolic_number
*n_ptr
;
1972 if (!n1
.base_addr
|| !n2
.base_addr
1973 || !operand_equal_p (n1
.base_addr
, n2
.base_addr
, 0))
1975 if (!n1
.offset
!= !n2
.offset
||
1976 (n1
.offset
&& !operand_equal_p (n1
.offset
, n2
.offset
, 0)))
1979 /* We swap n1 with n2 to have n1 < n2. */
1980 if (n2
.bytepos
< n1
.bytepos
)
1982 struct symbolic_number tmpn
;
1987 source_stmt1
= source_stmt2
;
1990 off_sub
= n2
.bytepos
- n1
.bytepos
;
1992 /* Check that the range of memory covered < biggest int size. */
1993 if (off_sub
+ n2
.range
> (int) sizeof (int64_t))
1995 n
->range
= n2
.range
+ off_sub
;
1997 /* Reinterpret byte marks in symbolic number holding the value of
1998 bigger weight according to target endianness. */
1999 inc
= BYTES_BIG_ENDIAN
? off_sub
+ n2
.range
- n1
.range
: off_sub
;
2001 if (BYTES_BIG_ENDIAN
)
2005 for (i
= 0; i
< sizeof (int64_t); i
++, inc
<<= 8,
2008 if (n_ptr
->n
& mask
)
2013 n
->range
= n1
.range
;
2016 || alias_ptr_types_compatible_p (n1
.alias_set
, n2
.alias_set
))
2017 n
->alias_set
= n1
.alias_set
;
2019 n
->alias_set
= ptr_type_node
;
2021 n
->base_addr
= n1
.base_addr
;
2022 n
->offset
= n1
.offset
;
2023 n
->bytepos
= n1
.bytepos
;
2025 size
= TYPE_PRECISION (n
->type
) / BITS_PER_UNIT
;
2026 for (i
= 0, mask
= 0xff; i
< size
; i
++, mask
<<= BITS_PER_UNIT
)
2028 uint64_t masked1
, masked2
;
2030 masked1
= n1
.n
& mask
;
2031 masked2
= n2
.n
& mask
;
2032 if (masked1
&& masked2
&& masked1
!= masked2
)
2037 if (!verify_symbolic_number_p (n
, stmt
))
2044 return source_stmt1
;
2049 /* Check if STMT completes a bswap implementation or a read in a given
2050 endianness consisting of ORs, SHIFTs and ANDs and sets *BSWAP
2051 accordingly. It also sets N to represent the kind of operations
2052 performed: size of the resulting expression and whether it works on
2053 a memory source, and if so alias-set and vuse. At last, the
2054 function returns a stmt whose rhs's first tree is the source
2058 find_bswap_or_nop (gimple stmt
, struct symbolic_number
*n
, bool *bswap
)
2060 /* The number which the find_bswap_or_nop_1 result should match in order
2061 to have a full byte swap. The number is shifted to the right
2062 according to the size of the symbolic number before using it. */
2063 uint64_t cmpxchg
= CMPXCHG
;
2064 uint64_t cmpnop
= CMPNOP
;
2069 /* The last parameter determines the depth search limit. It usually
2070 correlates directly to the number n of bytes to be touched. We
2071 increase that number by log2(n) + 1 here in order to also
2072 cover signed -> unsigned conversions of the src operand as can be seen
2073 in libgcc, and for initial shift/and operation of the src operand. */
2074 limit
= TREE_INT_CST_LOW (TYPE_SIZE_UNIT (gimple_expr_type (stmt
)));
2075 limit
+= 1 + (int) ceil_log2 ((unsigned HOST_WIDE_INT
) limit
);
2076 source_stmt
= find_bswap_or_nop_1 (stmt
, n
, limit
);
2081 /* Find real size of result (highest non zero byte). */
2087 for (tmpn
= n
->n
, rsize
= 0; tmpn
; tmpn
>>= BITS_PER_UNIT
, rsize
++);
2091 /* Zero out the extra bits of N and CMP*. */
2092 if (n
->range
< (int)sizeof (int64_t))
2096 mask
= ((uint64_t)1 << (n
->range
* BITS_PER_UNIT
)) - 1;
2097 cmpxchg
>>= (sizeof (int64_t) - n
->range
) * BITS_PER_UNIT
;
2101 /* A complete byte swap should make the symbolic number to start with
2102 the largest digit in the highest order byte. Unchanged symbolic
2103 number indicates a read with same endianness as target architecture. */
2106 else if (n
->n
== cmpxchg
)
2111 /* Useless bit manipulation performed by code. */
2112 if (!n
->base_addr
&& n
->n
== cmpnop
)
2115 n
->range
*= BITS_PER_UNIT
;
2121 const pass_data pass_data_optimize_bswap
=
2123 GIMPLE_PASS
, /* type */
2125 OPTGROUP_NONE
, /* optinfo_flags */
2126 true, /* has_execute */
2127 TV_NONE
, /* tv_id */
2128 PROP_ssa
, /* properties_required */
2129 0, /* properties_provided */
2130 0, /* properties_destroyed */
2131 0, /* todo_flags_start */
2132 0, /* todo_flags_finish */
2135 class pass_optimize_bswap
: public gimple_opt_pass
2138 pass_optimize_bswap (gcc::context
*ctxt
)
2139 : gimple_opt_pass (pass_data_optimize_bswap
, ctxt
)
2142 /* opt_pass methods: */
2143 virtual bool gate (function
*)
2145 return flag_expensive_optimizations
&& optimize
;
2148 virtual unsigned int execute (function
*);
2150 }; // class pass_optimize_bswap
2152 /* Perform the bswap optimization: replace the statement CUR_STMT at
2153 GSI with a load of type, VUSE and set-alias as described by N if a
2154 memory source is involved (N->base_addr is non null), followed by
2155 the builtin bswap invocation in FNDECL if BSWAP is true. SRC_STMT
2156 gives where should the replacement be made. It also gives the
2157 source on which CUR_STMT is operating via its rhs's first tree nad
2158 N->range gives the size of the expression involved for maintaining
2162 bswap_replace (gimple cur_stmt
, gimple_stmt_iterator gsi
, gimple src_stmt
,
2163 tree fndecl
, tree bswap_type
, tree load_type
,
2164 struct symbolic_number
*n
, bool bswap
)
2169 src
= gimple_assign_rhs1 (src_stmt
);
2170 tgt
= gimple_assign_lhs (cur_stmt
);
2172 /* Need to load the value from memory first. */
2175 gimple_stmt_iterator gsi_ins
= gsi_for_stmt (src_stmt
);
2176 tree addr_expr
, addr_tmp
, val_expr
, val_tmp
;
2177 tree load_offset_ptr
, aligned_load_type
;
2178 gimple addr_stmt
, load_stmt
;
2181 align
= get_object_alignment (src
);
2182 if (bswap
&& SLOW_UNALIGNED_ACCESS (TYPE_MODE (load_type
), align
))
2185 gsi_move_before (&gsi
, &gsi_ins
);
2186 gsi
= gsi_for_stmt (cur_stmt
);
2188 /* Compute address to load from and cast according to the size
2190 addr_expr
= build_fold_addr_expr (unshare_expr (src
));
2191 if (is_gimple_min_invariant (addr_expr
))
2192 addr_tmp
= addr_expr
;
2195 addr_tmp
= make_temp_ssa_name (TREE_TYPE (addr_expr
), NULL
,
2197 addr_stmt
= gimple_build_assign (addr_tmp
, addr_expr
);
2198 gsi_insert_before (&gsi
, addr_stmt
, GSI_SAME_STMT
);
2201 /* Perform the load. */
2202 aligned_load_type
= load_type
;
2203 if (align
< TYPE_ALIGN (load_type
))
2204 aligned_load_type
= build_aligned_type (load_type
, align
);
2205 load_offset_ptr
= build_int_cst (n
->alias_set
, 0);
2206 val_expr
= fold_build2 (MEM_REF
, aligned_load_type
, addr_tmp
,
2212 nop_stats
.found_16bit
++;
2213 else if (n
->range
== 32)
2214 nop_stats
.found_32bit
++;
2217 gcc_assert (n
->range
== 64);
2218 nop_stats
.found_64bit
++;
2221 /* Convert the result of load if necessary. */
2222 if (!useless_type_conversion_p (TREE_TYPE (tgt
), load_type
))
2224 val_tmp
= make_temp_ssa_name (aligned_load_type
, NULL
,
2226 load_stmt
= gimple_build_assign (val_tmp
, val_expr
);
2227 gimple_set_vuse (load_stmt
, n
->vuse
);
2228 gsi_insert_before (&gsi
, load_stmt
, GSI_SAME_STMT
);
2229 gimple_assign_set_rhs_with_ops_1 (&gsi
, NOP_EXPR
, val_tmp
,
2230 NULL_TREE
, NULL_TREE
);
2234 gimple_assign_set_rhs_with_ops_1 (&gsi
, MEM_REF
, val_expr
,
2235 NULL_TREE
, NULL_TREE
);
2236 gimple_set_vuse (cur_stmt
, n
->vuse
);
2238 update_stmt (cur_stmt
);
2243 "%d bit load in target endianness found at: ",
2245 print_gimple_stmt (dump_file
, cur_stmt
, 0, 0);
2251 val_tmp
= make_temp_ssa_name (aligned_load_type
, NULL
, "load_dst");
2252 load_stmt
= gimple_build_assign (val_tmp
, val_expr
);
2253 gimple_set_vuse (load_stmt
, n
->vuse
);
2254 gsi_insert_before (&gsi
, load_stmt
, GSI_SAME_STMT
);
2260 bswap_stats
.found_16bit
++;
2261 else if (n
->range
== 32)
2262 bswap_stats
.found_32bit
++;
2265 gcc_assert (n
->range
== 64);
2266 bswap_stats
.found_64bit
++;
2271 /* Convert the src expression if necessary. */
2272 if (!useless_type_conversion_p (TREE_TYPE (tmp
), bswap_type
))
2274 gimple convert_stmt
;
2275 tmp
= make_temp_ssa_name (bswap_type
, NULL
, "bswapsrc");
2276 convert_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, tmp
, src
, NULL
);
2277 gsi_insert_before (&gsi
, convert_stmt
, GSI_SAME_STMT
);
2280 call
= gimple_build_call (fndecl
, 1, tmp
);
2284 /* Convert the result if necessary. */
2285 if (!useless_type_conversion_p (TREE_TYPE (tgt
), bswap_type
))
2287 gimple convert_stmt
;
2288 tmp
= make_temp_ssa_name (bswap_type
, NULL
, "bswapdst");
2289 convert_stmt
= gimple_build_assign_with_ops (NOP_EXPR
, tgt
, tmp
, NULL
);
2290 gsi_insert_after (&gsi
, convert_stmt
, GSI_SAME_STMT
);
2293 gimple_call_set_lhs (call
, tmp
);
2297 fprintf (dump_file
, "%d bit bswap implementation found at: ",
2299 print_gimple_stmt (dump_file
, cur_stmt
, 0, 0);
2302 gsi_insert_after (&gsi
, call
, GSI_SAME_STMT
);
2303 gsi_remove (&gsi
, true);
2307 /* Find manual byte swap implementations as well as load in a given
2308 endianness. Byte swaps are turned into a bswap builtin invokation
2309 while endian loads are converted to bswap builtin invokation or
2310 simple load according to the target endianness. */
2313 pass_optimize_bswap::execute (function
*fun
)
2316 bool bswap16_p
, bswap32_p
, bswap64_p
;
2317 bool changed
= false;
2318 tree bswap16_type
= NULL_TREE
, bswap32_type
= NULL_TREE
, bswap64_type
= NULL_TREE
;
2320 if (BITS_PER_UNIT
!= 8)
2323 bswap16_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP16
)
2324 && optab_handler (bswap_optab
, HImode
) != CODE_FOR_nothing
);
2325 bswap32_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP32
)
2326 && optab_handler (bswap_optab
, SImode
) != CODE_FOR_nothing
);
2327 bswap64_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP64
)
2328 && (optab_handler (bswap_optab
, DImode
) != CODE_FOR_nothing
2329 || (bswap32_p
&& word_mode
== SImode
)));
2331 /* Determine the argument type of the builtins. The code later on
2332 assumes that the return and argument type are the same. */
2335 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP16
);
2336 bswap16_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
2341 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP32
);
2342 bswap32_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
2347 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP64
);
2348 bswap64_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
2351 memset (&nop_stats
, 0, sizeof (nop_stats
));
2352 memset (&bswap_stats
, 0, sizeof (bswap_stats
));
2354 FOR_EACH_BB_FN (bb
, fun
)
2356 gimple_stmt_iterator gsi
;
2358 /* We do a reverse scan for bswap patterns to make sure we get the
2359 widest match. As bswap pattern matching doesn't handle
2360 previously inserted smaller bswap replacements as sub-
2361 patterns, the wider variant wouldn't be detected. */
2362 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
); gsi_prev (&gsi
))
2364 gimple src_stmt
, cur_stmt
= gsi_stmt (gsi
);
2365 tree fndecl
= NULL_TREE
, bswap_type
= NULL_TREE
, load_type
;
2366 struct symbolic_number n
;
2369 if (!is_gimple_assign (cur_stmt
)
2370 || gimple_assign_rhs_code (cur_stmt
) != BIT_IOR_EXPR
)
2373 src_stmt
= find_bswap_or_nop (cur_stmt
, &n
, &bswap
);
2381 load_type
= uint16_type_node
;
2384 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP16
);
2385 bswap_type
= bswap16_type
;
2389 load_type
= uint32_type_node
;
2392 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP32
);
2393 bswap_type
= bswap32_type
;
2397 load_type
= uint64_type_node
;
2400 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP64
);
2401 bswap_type
= bswap64_type
;
2408 if (bswap
&& !fndecl
)
2411 if (bswap_replace (cur_stmt
, gsi
, src_stmt
, fndecl
, bswap_type
,
2412 load_type
, &n
, bswap
))
2417 statistics_counter_event (fun
, "16-bit nop implementations found",
2418 nop_stats
.found_16bit
);
2419 statistics_counter_event (fun
, "32-bit nop implementations found",
2420 nop_stats
.found_32bit
);
2421 statistics_counter_event (fun
, "64-bit nop implementations found",
2422 nop_stats
.found_64bit
);
2423 statistics_counter_event (fun
, "16-bit bswap implementations found",
2424 bswap_stats
.found_16bit
);
2425 statistics_counter_event (fun
, "32-bit bswap implementations found",
2426 bswap_stats
.found_32bit
);
2427 statistics_counter_event (fun
, "64-bit bswap implementations found",
2428 bswap_stats
.found_64bit
);
2430 return (changed
? TODO_update_ssa
: 0);
2436 make_pass_optimize_bswap (gcc::context
*ctxt
)
2438 return new pass_optimize_bswap (ctxt
);
2441 /* Return true if stmt is a type conversion operation that can be stripped
2442 when used in a widening multiply operation. */
2444 widening_mult_conversion_strippable_p (tree result_type
, gimple stmt
)
2446 enum tree_code rhs_code
= gimple_assign_rhs_code (stmt
);
2448 if (TREE_CODE (result_type
) == INTEGER_TYPE
)
2453 if (!CONVERT_EXPR_CODE_P (rhs_code
))
2456 op_type
= TREE_TYPE (gimple_assign_lhs (stmt
));
2458 /* If the type of OP has the same precision as the result, then
2459 we can strip this conversion. The multiply operation will be
2460 selected to create the correct extension as a by-product. */
2461 if (TYPE_PRECISION (result_type
) == TYPE_PRECISION (op_type
))
2464 /* We can also strip a conversion if it preserves the signed-ness of
2465 the operation and doesn't narrow the range. */
2466 inner_op_type
= TREE_TYPE (gimple_assign_rhs1 (stmt
));
2468 /* If the inner-most type is unsigned, then we can strip any
2469 intermediate widening operation. If it's signed, then the
2470 intermediate widening operation must also be signed. */
2471 if ((TYPE_UNSIGNED (inner_op_type
)
2472 || TYPE_UNSIGNED (op_type
) == TYPE_UNSIGNED (inner_op_type
))
2473 && TYPE_PRECISION (op_type
) > TYPE_PRECISION (inner_op_type
))
2479 return rhs_code
== FIXED_CONVERT_EXPR
;
2482 /* Return true if RHS is a suitable operand for a widening multiplication,
2483 assuming a target type of TYPE.
2484 There are two cases:
2486 - RHS makes some value at least twice as wide. Store that value
2487 in *NEW_RHS_OUT if so, and store its type in *TYPE_OUT.
2489 - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so,
2490 but leave *TYPE_OUT untouched. */
2493 is_widening_mult_rhs_p (tree type
, tree rhs
, tree
*type_out
,
2499 if (TREE_CODE (rhs
) == SSA_NAME
)
2501 stmt
= SSA_NAME_DEF_STMT (rhs
);
2502 if (is_gimple_assign (stmt
))
2504 if (! widening_mult_conversion_strippable_p (type
, stmt
))
2508 rhs1
= gimple_assign_rhs1 (stmt
);
2510 if (TREE_CODE (rhs1
) == INTEGER_CST
)
2512 *new_rhs_out
= rhs1
;
2521 type1
= TREE_TYPE (rhs1
);
2523 if (TREE_CODE (type1
) != TREE_CODE (type
)
2524 || TYPE_PRECISION (type1
) * 2 > TYPE_PRECISION (type
))
2527 *new_rhs_out
= rhs1
;
2532 if (TREE_CODE (rhs
) == INTEGER_CST
)
2542 /* Return true if STMT performs a widening multiplication, assuming the
2543 output type is TYPE. If so, store the unwidened types of the operands
2544 in *TYPE1_OUT and *TYPE2_OUT respectively. Also fill *RHS1_OUT and
2545 *RHS2_OUT such that converting those operands to types *TYPE1_OUT
2546 and *TYPE2_OUT would give the operands of the multiplication. */
2549 is_widening_mult_p (gimple stmt
,
2550 tree
*type1_out
, tree
*rhs1_out
,
2551 tree
*type2_out
, tree
*rhs2_out
)
2553 tree type
= TREE_TYPE (gimple_assign_lhs (stmt
));
2555 if (TREE_CODE (type
) != INTEGER_TYPE
2556 && TREE_CODE (type
) != FIXED_POINT_TYPE
)
2559 if (!is_widening_mult_rhs_p (type
, gimple_assign_rhs1 (stmt
), type1_out
,
2563 if (!is_widening_mult_rhs_p (type
, gimple_assign_rhs2 (stmt
), type2_out
,
2567 if (*type1_out
== NULL
)
2569 if (*type2_out
== NULL
|| !int_fits_type_p (*rhs1_out
, *type2_out
))
2571 *type1_out
= *type2_out
;
2574 if (*type2_out
== NULL
)
2576 if (!int_fits_type_p (*rhs2_out
, *type1_out
))
2578 *type2_out
= *type1_out
;
2581 /* Ensure that the larger of the two operands comes first. */
2582 if (TYPE_PRECISION (*type1_out
) < TYPE_PRECISION (*type2_out
))
2586 *type1_out
= *type2_out
;
2589 *rhs1_out
= *rhs2_out
;
2596 /* Process a single gimple statement STMT, which has a MULT_EXPR as
2597 its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return
2598 value is true iff we converted the statement. */
2601 convert_mult_to_widen (gimple stmt
, gimple_stmt_iterator
*gsi
)
2603 tree lhs
, rhs1
, rhs2
, type
, type1
, type2
;
2604 enum insn_code handler
;
2605 enum machine_mode to_mode
, from_mode
, actual_mode
;
2607 int actual_precision
;
2608 location_t loc
= gimple_location (stmt
);
2609 bool from_unsigned1
, from_unsigned2
;
2611 lhs
= gimple_assign_lhs (stmt
);
2612 type
= TREE_TYPE (lhs
);
2613 if (TREE_CODE (type
) != INTEGER_TYPE
)
2616 if (!is_widening_mult_p (stmt
, &type1
, &rhs1
, &type2
, &rhs2
))
2619 to_mode
= TYPE_MODE (type
);
2620 from_mode
= TYPE_MODE (type1
);
2621 from_unsigned1
= TYPE_UNSIGNED (type1
);
2622 from_unsigned2
= TYPE_UNSIGNED (type2
);
2624 if (from_unsigned1
&& from_unsigned2
)
2625 op
= umul_widen_optab
;
2626 else if (!from_unsigned1
&& !from_unsigned2
)
2627 op
= smul_widen_optab
;
2629 op
= usmul_widen_optab
;
2631 handler
= find_widening_optab_handler_and_mode (op
, to_mode
, from_mode
,
2634 if (handler
== CODE_FOR_nothing
)
2636 if (op
!= smul_widen_optab
)
2638 /* We can use a signed multiply with unsigned types as long as
2639 there is a wider mode to use, or it is the smaller of the two
2640 types that is unsigned. Note that type1 >= type2, always. */
2641 if ((TYPE_UNSIGNED (type1
)
2642 && TYPE_PRECISION (type1
) == GET_MODE_PRECISION (from_mode
))
2643 || (TYPE_UNSIGNED (type2
)
2644 && TYPE_PRECISION (type2
) == GET_MODE_PRECISION (from_mode
)))
2646 from_mode
= GET_MODE_WIDER_MODE (from_mode
);
2647 if (GET_MODE_SIZE (to_mode
) <= GET_MODE_SIZE (from_mode
))
2651 op
= smul_widen_optab
;
2652 handler
= find_widening_optab_handler_and_mode (op
, to_mode
,
2656 if (handler
== CODE_FOR_nothing
)
2659 from_unsigned1
= from_unsigned2
= false;
2665 /* Ensure that the inputs to the handler are in the correct precison
2666 for the opcode. This will be the full mode size. */
2667 actual_precision
= GET_MODE_PRECISION (actual_mode
);
2668 if (2 * actual_precision
> TYPE_PRECISION (type
))
2670 if (actual_precision
!= TYPE_PRECISION (type1
)
2671 || from_unsigned1
!= TYPE_UNSIGNED (type1
))
2672 rhs1
= build_and_insert_cast (gsi
, loc
,
2673 build_nonstandard_integer_type
2674 (actual_precision
, from_unsigned1
), rhs1
);
2675 if (actual_precision
!= TYPE_PRECISION (type2
)
2676 || from_unsigned2
!= TYPE_UNSIGNED (type2
))
2677 rhs2
= build_and_insert_cast (gsi
, loc
,
2678 build_nonstandard_integer_type
2679 (actual_precision
, from_unsigned2
), rhs2
);
2681 /* Handle constants. */
2682 if (TREE_CODE (rhs1
) == INTEGER_CST
)
2683 rhs1
= fold_convert (type1
, rhs1
);
2684 if (TREE_CODE (rhs2
) == INTEGER_CST
)
2685 rhs2
= fold_convert (type2
, rhs2
);
2687 gimple_assign_set_rhs1 (stmt
, rhs1
);
2688 gimple_assign_set_rhs2 (stmt
, rhs2
);
2689 gimple_assign_set_rhs_code (stmt
, WIDEN_MULT_EXPR
);
2691 widen_mul_stats
.widen_mults_inserted
++;
2695 /* Process a single gimple statement STMT, which is found at the
2696 iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its
2697 rhs (given by CODE), and try to convert it into a
2698 WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value
2699 is true iff we converted the statement. */
2702 convert_plusminus_to_widen (gimple_stmt_iterator
*gsi
, gimple stmt
,
2703 enum tree_code code
)
2705 gimple rhs1_stmt
= NULL
, rhs2_stmt
= NULL
;
2706 gimple conv1_stmt
= NULL
, conv2_stmt
= NULL
, conv_stmt
;
2707 tree type
, type1
, type2
, optype
;
2708 tree lhs
, rhs1
, rhs2
, mult_rhs1
, mult_rhs2
, add_rhs
;
2709 enum tree_code rhs1_code
= ERROR_MARK
, rhs2_code
= ERROR_MARK
;
2711 enum tree_code wmult_code
;
2712 enum insn_code handler
;
2713 enum machine_mode to_mode
, from_mode
, actual_mode
;
2714 location_t loc
= gimple_location (stmt
);
2715 int actual_precision
;
2716 bool from_unsigned1
, from_unsigned2
;
2718 lhs
= gimple_assign_lhs (stmt
);
2719 type
= TREE_TYPE (lhs
);
2720 if (TREE_CODE (type
) != INTEGER_TYPE
2721 && TREE_CODE (type
) != FIXED_POINT_TYPE
)
2724 if (code
== MINUS_EXPR
)
2725 wmult_code
= WIDEN_MULT_MINUS_EXPR
;
2727 wmult_code
= WIDEN_MULT_PLUS_EXPR
;
2729 rhs1
= gimple_assign_rhs1 (stmt
);
2730 rhs2
= gimple_assign_rhs2 (stmt
);
2732 if (TREE_CODE (rhs1
) == SSA_NAME
)
2734 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
2735 if (is_gimple_assign (rhs1_stmt
))
2736 rhs1_code
= gimple_assign_rhs_code (rhs1_stmt
);
2739 if (TREE_CODE (rhs2
) == SSA_NAME
)
2741 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
2742 if (is_gimple_assign (rhs2_stmt
))
2743 rhs2_code
= gimple_assign_rhs_code (rhs2_stmt
);
2746 /* Allow for one conversion statement between the multiply
2747 and addition/subtraction statement. If there are more than
2748 one conversions then we assume they would invalidate this
2749 transformation. If that's not the case then they should have
2750 been folded before now. */
2751 if (CONVERT_EXPR_CODE_P (rhs1_code
))
2753 conv1_stmt
= rhs1_stmt
;
2754 rhs1
= gimple_assign_rhs1 (rhs1_stmt
);
2755 if (TREE_CODE (rhs1
) == SSA_NAME
)
2757 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
2758 if (is_gimple_assign (rhs1_stmt
))
2759 rhs1_code
= gimple_assign_rhs_code (rhs1_stmt
);
2764 if (CONVERT_EXPR_CODE_P (rhs2_code
))
2766 conv2_stmt
= rhs2_stmt
;
2767 rhs2
= gimple_assign_rhs1 (rhs2_stmt
);
2768 if (TREE_CODE (rhs2
) == SSA_NAME
)
2770 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
2771 if (is_gimple_assign (rhs2_stmt
))
2772 rhs2_code
= gimple_assign_rhs_code (rhs2_stmt
);
2778 /* If code is WIDEN_MULT_EXPR then it would seem unnecessary to call
2779 is_widening_mult_p, but we still need the rhs returns.
2781 It might also appear that it would be sufficient to use the existing
2782 operands of the widening multiply, but that would limit the choice of
2783 multiply-and-accumulate instructions.
2785 If the widened-multiplication result has more than one uses, it is
2786 probably wiser not to do the conversion. */
2787 if (code
== PLUS_EXPR
2788 && (rhs1_code
== MULT_EXPR
|| rhs1_code
== WIDEN_MULT_EXPR
))
2790 if (!has_single_use (rhs1
)
2791 || !is_widening_mult_p (rhs1_stmt
, &type1
, &mult_rhs1
,
2792 &type2
, &mult_rhs2
))
2795 conv_stmt
= conv1_stmt
;
2797 else if (rhs2_code
== MULT_EXPR
|| rhs2_code
== WIDEN_MULT_EXPR
)
2799 if (!has_single_use (rhs2
)
2800 || !is_widening_mult_p (rhs2_stmt
, &type1
, &mult_rhs1
,
2801 &type2
, &mult_rhs2
))
2804 conv_stmt
= conv2_stmt
;
2809 to_mode
= TYPE_MODE (type
);
2810 from_mode
= TYPE_MODE (type1
);
2811 from_unsigned1
= TYPE_UNSIGNED (type1
);
2812 from_unsigned2
= TYPE_UNSIGNED (type2
);
2815 /* There's no such thing as a mixed sign madd yet, so use a wider mode. */
2816 if (from_unsigned1
!= from_unsigned2
)
2818 if (!INTEGRAL_TYPE_P (type
))
2820 /* We can use a signed multiply with unsigned types as long as
2821 there is a wider mode to use, or it is the smaller of the two
2822 types that is unsigned. Note that type1 >= type2, always. */
2824 && TYPE_PRECISION (type1
) == GET_MODE_PRECISION (from_mode
))
2826 && TYPE_PRECISION (type2
) == GET_MODE_PRECISION (from_mode
)))
2828 from_mode
= GET_MODE_WIDER_MODE (from_mode
);
2829 if (GET_MODE_SIZE (from_mode
) >= GET_MODE_SIZE (to_mode
))
2833 from_unsigned1
= from_unsigned2
= false;
2834 optype
= build_nonstandard_integer_type (GET_MODE_PRECISION (from_mode
),
2838 /* If there was a conversion between the multiply and addition
2839 then we need to make sure it fits a multiply-and-accumulate.
2840 The should be a single mode change which does not change the
2844 /* We use the original, unmodified data types for this. */
2845 tree from_type
= TREE_TYPE (gimple_assign_rhs1 (conv_stmt
));
2846 tree to_type
= TREE_TYPE (gimple_assign_lhs (conv_stmt
));
2847 int data_size
= TYPE_PRECISION (type1
) + TYPE_PRECISION (type2
);
2848 bool is_unsigned
= TYPE_UNSIGNED (type1
) && TYPE_UNSIGNED (type2
);
2850 if (TYPE_PRECISION (from_type
) > TYPE_PRECISION (to_type
))
2852 /* Conversion is a truncate. */
2853 if (TYPE_PRECISION (to_type
) < data_size
)
2856 else if (TYPE_PRECISION (from_type
) < TYPE_PRECISION (to_type
))
2858 /* Conversion is an extend. Check it's the right sort. */
2859 if (TYPE_UNSIGNED (from_type
) != is_unsigned
2860 && !(is_unsigned
&& TYPE_PRECISION (from_type
) > data_size
))
2863 /* else convert is a no-op for our purposes. */
2866 /* Verify that the machine can perform a widening multiply
2867 accumulate in this mode/signedness combination, otherwise
2868 this transformation is likely to pessimize code. */
2869 this_optab
= optab_for_tree_code (wmult_code
, optype
, optab_default
);
2870 handler
= find_widening_optab_handler_and_mode (this_optab
, to_mode
,
2871 from_mode
, 0, &actual_mode
);
2873 if (handler
== CODE_FOR_nothing
)
2876 /* Ensure that the inputs to the handler are in the correct precison
2877 for the opcode. This will be the full mode size. */
2878 actual_precision
= GET_MODE_PRECISION (actual_mode
);
2879 if (actual_precision
!= TYPE_PRECISION (type1
)
2880 || from_unsigned1
!= TYPE_UNSIGNED (type1
))
2881 mult_rhs1
= build_and_insert_cast (gsi
, loc
,
2882 build_nonstandard_integer_type
2883 (actual_precision
, from_unsigned1
),
2885 if (actual_precision
!= TYPE_PRECISION (type2
)
2886 || from_unsigned2
!= TYPE_UNSIGNED (type2
))
2887 mult_rhs2
= build_and_insert_cast (gsi
, loc
,
2888 build_nonstandard_integer_type
2889 (actual_precision
, from_unsigned2
),
2892 if (!useless_type_conversion_p (type
, TREE_TYPE (add_rhs
)))
2893 add_rhs
= build_and_insert_cast (gsi
, loc
, type
, add_rhs
);
2895 /* Handle constants. */
2896 if (TREE_CODE (mult_rhs1
) == INTEGER_CST
)
2897 mult_rhs1
= fold_convert (type1
, mult_rhs1
);
2898 if (TREE_CODE (mult_rhs2
) == INTEGER_CST
)
2899 mult_rhs2
= fold_convert (type2
, mult_rhs2
);
2901 gimple_assign_set_rhs_with_ops_1 (gsi
, wmult_code
, mult_rhs1
, mult_rhs2
,
2903 update_stmt (gsi_stmt (*gsi
));
2904 widen_mul_stats
.maccs_inserted
++;
2908 /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2
2909 with uses in additions and subtractions to form fused multiply-add
2910 operations. Returns true if successful and MUL_STMT should be removed. */
2913 convert_mult_to_fma (gimple mul_stmt
, tree op1
, tree op2
)
2915 tree mul_result
= gimple_get_lhs (mul_stmt
);
2916 tree type
= TREE_TYPE (mul_result
);
2917 gimple use_stmt
, neguse_stmt
, fma_stmt
;
2918 use_operand_p use_p
;
2919 imm_use_iterator imm_iter
;
2921 if (FLOAT_TYPE_P (type
)
2922 && flag_fp_contract_mode
== FP_CONTRACT_OFF
)
2925 /* We don't want to do bitfield reduction ops. */
2926 if (INTEGRAL_TYPE_P (type
)
2927 && (TYPE_PRECISION (type
)
2928 != GET_MODE_PRECISION (TYPE_MODE (type
))))
2931 /* If the target doesn't support it, don't generate it. We assume that
2932 if fma isn't available then fms, fnma or fnms are not either. */
2933 if (optab_handler (fma_optab
, TYPE_MODE (type
)) == CODE_FOR_nothing
)
2936 /* If the multiplication has zero uses, it is kept around probably because
2937 of -fnon-call-exceptions. Don't optimize it away in that case,
2939 if (has_zero_uses (mul_result
))
2942 /* Make sure that the multiplication statement becomes dead after
2943 the transformation, thus that all uses are transformed to FMAs.
2944 This means we assume that an FMA operation has the same cost
2946 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, mul_result
)
2948 enum tree_code use_code
;
2949 tree result
= mul_result
;
2950 bool negate_p
= false;
2952 use_stmt
= USE_STMT (use_p
);
2954 if (is_gimple_debug (use_stmt
))
2957 /* For now restrict this operations to single basic blocks. In theory
2958 we would want to support sinking the multiplication in
2964 to form a fma in the then block and sink the multiplication to the
2966 if (gimple_bb (use_stmt
) != gimple_bb (mul_stmt
))
2969 if (!is_gimple_assign (use_stmt
))
2972 use_code
= gimple_assign_rhs_code (use_stmt
);
2974 /* A negate on the multiplication leads to FNMA. */
2975 if (use_code
== NEGATE_EXPR
)
2980 result
= gimple_assign_lhs (use_stmt
);
2982 /* Make sure the negate statement becomes dead with this
2983 single transformation. */
2984 if (!single_imm_use (gimple_assign_lhs (use_stmt
),
2985 &use_p
, &neguse_stmt
))
2988 /* Make sure the multiplication isn't also used on that stmt. */
2989 FOR_EACH_PHI_OR_STMT_USE (usep
, neguse_stmt
, iter
, SSA_OP_USE
)
2990 if (USE_FROM_PTR (usep
) == mul_result
)
2994 use_stmt
= neguse_stmt
;
2995 if (gimple_bb (use_stmt
) != gimple_bb (mul_stmt
))
2997 if (!is_gimple_assign (use_stmt
))
3000 use_code
= gimple_assign_rhs_code (use_stmt
);
3007 if (gimple_assign_rhs2 (use_stmt
) == result
)
3008 negate_p
= !negate_p
;
3013 /* FMA can only be formed from PLUS and MINUS. */
3017 /* If the subtrahend (gimple_assign_rhs2 (use_stmt)) is computed
3018 by a MULT_EXPR that we'll visit later, we might be able to
3019 get a more profitable match with fnma.
3020 OTOH, if we don't, a negate / fma pair has likely lower latency
3021 that a mult / subtract pair. */
3022 if (use_code
== MINUS_EXPR
&& !negate_p
3023 && gimple_assign_rhs1 (use_stmt
) == result
3024 && optab_handler (fms_optab
, TYPE_MODE (type
)) == CODE_FOR_nothing
3025 && optab_handler (fnma_optab
, TYPE_MODE (type
)) != CODE_FOR_nothing
)
3027 tree rhs2
= gimple_assign_rhs2 (use_stmt
);
3029 if (TREE_CODE (rhs2
) == SSA_NAME
)
3031 gimple stmt2
= SSA_NAME_DEF_STMT (rhs2
);
3032 if (has_single_use (rhs2
)
3033 && is_gimple_assign (stmt2
)
3034 && gimple_assign_rhs_code (stmt2
) == MULT_EXPR
)
3039 /* We can't handle a * b + a * b. */
3040 if (gimple_assign_rhs1 (use_stmt
) == gimple_assign_rhs2 (use_stmt
))
3043 /* While it is possible to validate whether or not the exact form
3044 that we've recognized is available in the backend, the assumption
3045 is that the transformation is never a loss. For instance, suppose
3046 the target only has the plain FMA pattern available. Consider
3047 a*b-c -> fma(a,b,-c): we've exchanged MUL+SUB for FMA+NEG, which
3048 is still two operations. Consider -(a*b)-c -> fma(-a,b,-c): we
3049 still have 3 operations, but in the FMA form the two NEGs are
3050 independent and could be run in parallel. */
3053 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, mul_result
)
3055 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
3056 enum tree_code use_code
;
3057 tree addop
, mulop1
= op1
, result
= mul_result
;
3058 bool negate_p
= false;
3060 if (is_gimple_debug (use_stmt
))
3063 use_code
= gimple_assign_rhs_code (use_stmt
);
3064 if (use_code
== NEGATE_EXPR
)
3066 result
= gimple_assign_lhs (use_stmt
);
3067 single_imm_use (gimple_assign_lhs (use_stmt
), &use_p
, &neguse_stmt
);
3068 gsi_remove (&gsi
, true);
3069 release_defs (use_stmt
);
3071 use_stmt
= neguse_stmt
;
3072 gsi
= gsi_for_stmt (use_stmt
);
3073 use_code
= gimple_assign_rhs_code (use_stmt
);
3077 if (gimple_assign_rhs1 (use_stmt
) == result
)
3079 addop
= gimple_assign_rhs2 (use_stmt
);
3080 /* a * b - c -> a * b + (-c) */
3081 if (gimple_assign_rhs_code (use_stmt
) == MINUS_EXPR
)
3082 addop
= force_gimple_operand_gsi (&gsi
,
3083 build1 (NEGATE_EXPR
,
3085 true, NULL_TREE
, true,
3090 addop
= gimple_assign_rhs1 (use_stmt
);
3091 /* a - b * c -> (-b) * c + a */
3092 if (gimple_assign_rhs_code (use_stmt
) == MINUS_EXPR
)
3093 negate_p
= !negate_p
;
3097 mulop1
= force_gimple_operand_gsi (&gsi
,
3098 build1 (NEGATE_EXPR
,
3100 true, NULL_TREE
, true,
3103 fma_stmt
= gimple_build_assign_with_ops (FMA_EXPR
,
3104 gimple_assign_lhs (use_stmt
),
3107 gsi_replace (&gsi
, fma_stmt
, true);
3108 widen_mul_stats
.fmas_inserted
++;
3114 /* Find integer multiplications where the operands are extended from
3115 smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR
3116 where appropriate. */
3120 const pass_data pass_data_optimize_widening_mul
=
3122 GIMPLE_PASS
, /* type */
3123 "widening_mul", /* name */
3124 OPTGROUP_NONE
, /* optinfo_flags */
3125 true, /* has_execute */
3126 TV_NONE
, /* tv_id */
3127 PROP_ssa
, /* properties_required */
3128 0, /* properties_provided */
3129 0, /* properties_destroyed */
3130 0, /* todo_flags_start */
3131 TODO_update_ssa
, /* todo_flags_finish */
3134 class pass_optimize_widening_mul
: public gimple_opt_pass
3137 pass_optimize_widening_mul (gcc::context
*ctxt
)
3138 : gimple_opt_pass (pass_data_optimize_widening_mul
, ctxt
)
3141 /* opt_pass methods: */
3142 virtual bool gate (function
*)
3144 return flag_expensive_optimizations
&& optimize
;
3147 virtual unsigned int execute (function
*);
3149 }; // class pass_optimize_widening_mul
3152 pass_optimize_widening_mul::execute (function
*fun
)
3155 bool cfg_changed
= false;
3157 memset (&widen_mul_stats
, 0, sizeof (widen_mul_stats
));
3159 FOR_EACH_BB_FN (bb
, fun
)
3161 gimple_stmt_iterator gsi
;
3163 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
);)
3165 gimple stmt
= gsi_stmt (gsi
);
3166 enum tree_code code
;
3168 if (is_gimple_assign (stmt
))
3170 code
= gimple_assign_rhs_code (stmt
);
3174 if (!convert_mult_to_widen (stmt
, &gsi
)
3175 && convert_mult_to_fma (stmt
,
3176 gimple_assign_rhs1 (stmt
),
3177 gimple_assign_rhs2 (stmt
)))
3179 gsi_remove (&gsi
, true);
3180 release_defs (stmt
);
3187 convert_plusminus_to_widen (&gsi
, stmt
, code
);
3193 else if (is_gimple_call (stmt
)
3194 && gimple_call_lhs (stmt
))
3196 tree fndecl
= gimple_call_fndecl (stmt
);
3198 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
)
3200 switch (DECL_FUNCTION_CODE (fndecl
))
3205 if (TREE_CODE (gimple_call_arg (stmt
, 1)) == REAL_CST
3206 && REAL_VALUES_EQUAL
3207 (TREE_REAL_CST (gimple_call_arg (stmt
, 1)),
3209 && convert_mult_to_fma (stmt
,
3210 gimple_call_arg (stmt
, 0),
3211 gimple_call_arg (stmt
, 0)))
3213 unlink_stmt_vdef (stmt
);
3214 if (gsi_remove (&gsi
, true)
3215 && gimple_purge_dead_eh_edges (bb
))
3217 release_defs (stmt
);
3230 statistics_counter_event (fun
, "widening multiplications inserted",
3231 widen_mul_stats
.widen_mults_inserted
);
3232 statistics_counter_event (fun
, "widening maccs inserted",
3233 widen_mul_stats
.maccs_inserted
);
3234 statistics_counter_event (fun
, "fused multiply-adds inserted",
3235 widen_mul_stats
.fmas_inserted
);
3237 return cfg_changed
? TODO_cleanup_cfg
: 0;
3243 make_pass_optimize_widening_mul (gcc::context
*ctxt
)
3245 return new pass_optimize_widening_mul (ctxt
);