1 /* Global, SSA-based optimizations using mathematical identities.
2 Copyright (C) 2005-2016 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"
96 #include "alloc-pool.h"
97 #include "tree-pass.h"
99 #include "optabs-tree.h"
100 #include "gimple-pretty-print.h"
102 #include "fold-const.h"
103 #include "gimple-fold.h"
104 #include "gimple-iterator.h"
105 #include "gimplify.h"
106 #include "gimplify-me.h"
107 #include "stor-layout.h"
108 #include "tree-cfg.h"
109 #include "tree-dfa.h"
110 #include "tree-ssa.h"
111 #include "builtins.h"
113 #include "internal-fn.h"
114 #include "case-cfn-macros.h"
116 /* This structure represents one basic block that either computes a
117 division, or is a common dominator for basic block that compute a
120 /* The basic block represented by this structure. */
123 /* If non-NULL, the SSA_NAME holding the definition for a reciprocal
127 /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that
128 was inserted in BB. */
129 gimple
*recip_def_stmt
;
131 /* Pointer to a list of "struct occurrence"s for blocks dominated
133 struct occurrence
*children
;
135 /* Pointer to the next "struct occurrence"s in the list of blocks
136 sharing a common dominator. */
137 struct occurrence
*next
;
139 /* The number of divisions that are in BB before compute_merit. The
140 number of divisions that are in BB or post-dominate it after
144 /* True if the basic block has a division, false if it is a common
145 dominator for basic blocks that do. If it is false and trapping
146 math is active, BB is not a candidate for inserting a reciprocal. */
147 bool bb_has_division
;
152 /* Number of 1.0/X ops inserted. */
155 /* Number of 1.0/FUNC ops inserted. */
161 /* Number of cexpi calls inserted. */
167 /* Number of hand-written 16-bit nop / bswaps found. */
170 /* Number of hand-written 32-bit nop / bswaps found. */
173 /* Number of hand-written 64-bit nop / bswaps found. */
175 } nop_stats
, bswap_stats
;
179 /* Number of widening multiplication ops inserted. */
180 int widen_mults_inserted
;
182 /* Number of integer multiply-and-accumulate ops inserted. */
185 /* Number of fp fused multiply-add ops inserted. */
189 /* The instance of "struct occurrence" representing the highest
190 interesting block in the dominator tree. */
191 static struct occurrence
*occ_head
;
193 /* Allocation pool for getting instances of "struct occurrence". */
194 static object_allocator
<occurrence
> *occ_pool
;
198 /* Allocate and return a new struct occurrence for basic block BB, and
199 whose children list is headed by CHILDREN. */
200 static struct occurrence
*
201 occ_new (basic_block bb
, struct occurrence
*children
)
203 struct occurrence
*occ
;
205 bb
->aux
= occ
= occ_pool
->allocate ();
206 memset (occ
, 0, sizeof (struct occurrence
));
209 occ
->children
= children
;
214 /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a
215 list of "struct occurrence"s, one per basic block, having IDOM as
216 their common dominator.
218 We try to insert NEW_OCC as deep as possible in the tree, and we also
219 insert any other block that is a common dominator for BB and one
220 block already in the tree. */
223 insert_bb (struct occurrence
*new_occ
, basic_block idom
,
224 struct occurrence
**p_head
)
226 struct occurrence
*occ
, **p_occ
;
228 for (p_occ
= p_head
; (occ
= *p_occ
) != NULL
; )
230 basic_block bb
= new_occ
->bb
, occ_bb
= occ
->bb
;
231 basic_block dom
= nearest_common_dominator (CDI_DOMINATORS
, occ_bb
, bb
);
234 /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC
237 occ
->next
= new_occ
->children
;
238 new_occ
->children
= occ
;
240 /* Try the next block (it may as well be dominated by BB). */
243 else if (dom
== occ_bb
)
245 /* OCC_BB dominates BB. Tail recurse to look deeper. */
246 insert_bb (new_occ
, dom
, &occ
->children
);
250 else if (dom
!= idom
)
252 gcc_assert (!dom
->aux
);
254 /* There is a dominator between IDOM and BB, add it and make
255 two children out of NEW_OCC and OCC. First, remove OCC from
261 /* None of the previous blocks has DOM as a dominator: if we tail
262 recursed, we would reexamine them uselessly. Just switch BB with
263 DOM, and go on looking for blocks dominated by DOM. */
264 new_occ
= occ_new (dom
, new_occ
);
269 /* Nothing special, go on with the next element. */
274 /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */
275 new_occ
->next
= *p_head
;
279 /* Register that we found a division in BB. */
282 register_division_in (basic_block bb
)
284 struct occurrence
*occ
;
286 occ
= (struct occurrence
*) bb
->aux
;
289 occ
= occ_new (bb
, NULL
);
290 insert_bb (occ
, ENTRY_BLOCK_PTR_FOR_FN (cfun
), &occ_head
);
293 occ
->bb_has_division
= true;
294 occ
->num_divisions
++;
298 /* Compute the number of divisions that postdominate each block in OCC and
302 compute_merit (struct occurrence
*occ
)
304 struct occurrence
*occ_child
;
305 basic_block dom
= occ
->bb
;
307 for (occ_child
= occ
->children
; occ_child
; occ_child
= occ_child
->next
)
310 if (occ_child
->children
)
311 compute_merit (occ_child
);
314 bb
= single_noncomplex_succ (dom
);
318 if (dominated_by_p (CDI_POST_DOMINATORS
, bb
, occ_child
->bb
))
319 occ
->num_divisions
+= occ_child
->num_divisions
;
324 /* Return whether USE_STMT is a floating-point division by DEF. */
326 is_division_by (gimple
*use_stmt
, tree def
)
328 return is_gimple_assign (use_stmt
)
329 && gimple_assign_rhs_code (use_stmt
) == RDIV_EXPR
330 && gimple_assign_rhs2 (use_stmt
) == def
331 /* Do not recognize x / x as valid division, as we are getting
332 confused later by replacing all immediate uses x in such
334 && gimple_assign_rhs1 (use_stmt
) != def
;
337 /* Walk the subset of the dominator tree rooted at OCC, setting the
338 RECIP_DEF field to a definition of 1.0 / DEF that can be used in
339 the given basic block. The field may be left NULL, of course,
340 if it is not possible or profitable to do the optimization.
342 DEF_BSI is an iterator pointing at the statement defining DEF.
343 If RECIP_DEF is set, a dominator already has a computation that can
347 insert_reciprocals (gimple_stmt_iterator
*def_gsi
, struct occurrence
*occ
,
348 tree def
, tree recip_def
, int threshold
)
352 gimple_stmt_iterator gsi
;
353 struct occurrence
*occ_child
;
356 && (occ
->bb_has_division
|| !flag_trapping_math
)
357 && occ
->num_divisions
>= threshold
)
359 /* Make a variable with the replacement and substitute it. */
360 type
= TREE_TYPE (def
);
361 recip_def
= create_tmp_reg (type
, "reciptmp");
362 new_stmt
= gimple_build_assign (recip_def
, RDIV_EXPR
,
363 build_one_cst (type
), def
);
365 if (occ
->bb_has_division
)
367 /* Case 1: insert before an existing division. */
368 gsi
= gsi_after_labels (occ
->bb
);
369 while (!gsi_end_p (gsi
) && !is_division_by (gsi_stmt (gsi
), def
))
372 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
374 else if (def_gsi
&& occ
->bb
== def_gsi
->bb
)
376 /* Case 2: insert right after the definition. Note that this will
377 never happen if the definition statement can throw, because in
378 that case the sole successor of the statement's basic block will
379 dominate all the uses as well. */
380 gsi_insert_after (def_gsi
, new_stmt
, GSI_NEW_STMT
);
384 /* Case 3: insert in a basic block not containing defs/uses. */
385 gsi
= gsi_after_labels (occ
->bb
);
386 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
389 reciprocal_stats
.rdivs_inserted
++;
391 occ
->recip_def_stmt
= new_stmt
;
394 occ
->recip_def
= recip_def
;
395 for (occ_child
= occ
->children
; occ_child
; occ_child
= occ_child
->next
)
396 insert_reciprocals (def_gsi
, occ_child
, def
, recip_def
, threshold
);
400 /* Replace the division at USE_P with a multiplication by the reciprocal, if
404 replace_reciprocal (use_operand_p use_p
)
406 gimple
*use_stmt
= USE_STMT (use_p
);
407 basic_block bb
= gimple_bb (use_stmt
);
408 struct occurrence
*occ
= (struct occurrence
*) bb
->aux
;
410 if (optimize_bb_for_speed_p (bb
)
411 && occ
->recip_def
&& use_stmt
!= occ
->recip_def_stmt
)
413 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
414 gimple_assign_set_rhs_code (use_stmt
, MULT_EXPR
);
415 SET_USE (use_p
, occ
->recip_def
);
416 fold_stmt_inplace (&gsi
);
417 update_stmt (use_stmt
);
422 /* Free OCC and return one more "struct occurrence" to be freed. */
424 static struct occurrence
*
425 free_bb (struct occurrence
*occ
)
427 struct occurrence
*child
, *next
;
429 /* First get the two pointers hanging off OCC. */
431 child
= occ
->children
;
433 occ_pool
->remove (occ
);
435 /* Now ensure that we don't recurse unless it is necessary. */
441 next
= free_bb (next
);
448 /* Look for floating-point divisions among DEF's uses, and try to
449 replace them by multiplications with the reciprocal. Add
450 as many statements computing the reciprocal as needed.
452 DEF must be a GIMPLE register of a floating-point type. */
455 execute_cse_reciprocals_1 (gimple_stmt_iterator
*def_gsi
, tree def
)
458 imm_use_iterator use_iter
;
459 struct occurrence
*occ
;
460 int count
= 0, threshold
;
462 gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def
)) && is_gimple_reg (def
));
464 FOR_EACH_IMM_USE_FAST (use_p
, use_iter
, def
)
466 gimple
*use_stmt
= USE_STMT (use_p
);
467 if (is_division_by (use_stmt
, def
))
469 register_division_in (gimple_bb (use_stmt
));
474 /* Do the expensive part only if we can hope to optimize something. */
475 threshold
= targetm
.min_divisions_for_recip_mul (TYPE_MODE (TREE_TYPE (def
)));
476 if (count
>= threshold
)
479 for (occ
= occ_head
; occ
; occ
= occ
->next
)
482 insert_reciprocals (def_gsi
, occ
, def
, NULL
, threshold
);
485 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, def
)
487 if (is_division_by (use_stmt
, def
))
489 FOR_EACH_IMM_USE_ON_STMT (use_p
, use_iter
)
490 replace_reciprocal (use_p
);
495 for (occ
= occ_head
; occ
; )
501 /* Return an internal function that implements the reciprocal of CALL,
502 or IFN_LAST if there is no such function that the target supports. */
505 internal_fn_reciprocal (gcall
*call
)
509 switch (gimple_call_combined_fn (call
))
519 tree_pair types
= direct_internal_fn_types (ifn
, call
);
520 if (!direct_internal_fn_supported_p (ifn
, types
, OPTIMIZE_FOR_SPEED
))
526 /* Go through all the floating-point SSA_NAMEs, and call
527 execute_cse_reciprocals_1 on each of them. */
530 const pass_data pass_data_cse_reciprocals
=
532 GIMPLE_PASS
, /* type */
534 OPTGROUP_NONE
, /* optinfo_flags */
536 PROP_ssa
, /* properties_required */
537 0, /* properties_provided */
538 0, /* properties_destroyed */
539 0, /* todo_flags_start */
540 TODO_update_ssa
, /* todo_flags_finish */
543 class pass_cse_reciprocals
: public gimple_opt_pass
546 pass_cse_reciprocals (gcc::context
*ctxt
)
547 : gimple_opt_pass (pass_data_cse_reciprocals
, ctxt
)
550 /* opt_pass methods: */
551 virtual bool gate (function
*) { return optimize
&& flag_reciprocal_math
; }
552 virtual unsigned int execute (function
*);
554 }; // class pass_cse_reciprocals
557 pass_cse_reciprocals::execute (function
*fun
)
562 occ_pool
= new object_allocator
<occurrence
> ("dominators for recip");
564 memset (&reciprocal_stats
, 0, sizeof (reciprocal_stats
));
565 calculate_dominance_info (CDI_DOMINATORS
);
566 calculate_dominance_info (CDI_POST_DOMINATORS
);
569 FOR_EACH_BB_FN (bb
, fun
)
570 gcc_assert (!bb
->aux
);
572 for (arg
= DECL_ARGUMENTS (fun
->decl
); arg
; arg
= DECL_CHAIN (arg
))
573 if (FLOAT_TYPE_P (TREE_TYPE (arg
))
574 && is_gimple_reg (arg
))
576 tree name
= ssa_default_def (fun
, arg
);
578 execute_cse_reciprocals_1 (NULL
, name
);
581 FOR_EACH_BB_FN (bb
, fun
)
585 for (gphi_iterator gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
);
588 gphi
*phi
= gsi
.phi ();
589 def
= PHI_RESULT (phi
);
590 if (! virtual_operand_p (def
)
591 && FLOAT_TYPE_P (TREE_TYPE (def
)))
592 execute_cse_reciprocals_1 (NULL
, def
);
595 for (gimple_stmt_iterator gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
);
598 gimple
*stmt
= gsi_stmt (gsi
);
600 if (gimple_has_lhs (stmt
)
601 && (def
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_DEF
)) != NULL
602 && FLOAT_TYPE_P (TREE_TYPE (def
))
603 && TREE_CODE (def
) == SSA_NAME
)
604 execute_cse_reciprocals_1 (&gsi
, def
);
607 if (optimize_bb_for_size_p (bb
))
610 /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */
611 for (gimple_stmt_iterator gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
);
614 gimple
*stmt
= gsi_stmt (gsi
);
616 if (is_gimple_assign (stmt
)
617 && gimple_assign_rhs_code (stmt
) == RDIV_EXPR
)
619 tree arg1
= gimple_assign_rhs2 (stmt
);
622 if (TREE_CODE (arg1
) != SSA_NAME
)
625 stmt1
= SSA_NAME_DEF_STMT (arg1
);
627 if (is_gimple_call (stmt1
)
628 && gimple_call_lhs (stmt1
))
633 tree fndecl
= NULL_TREE
;
635 gcall
*call
= as_a
<gcall
*> (stmt1
);
636 internal_fn ifn
= internal_fn_reciprocal (call
);
639 fndecl
= gimple_call_fndecl (call
);
641 || DECL_BUILT_IN_CLASS (fndecl
) != BUILT_IN_MD
)
643 fndecl
= targetm
.builtin_reciprocal (fndecl
);
648 /* Check that all uses of the SSA name are divisions,
649 otherwise replacing the defining statement will do
652 FOR_EACH_IMM_USE_FAST (use_p
, ui
, arg1
)
654 gimple
*stmt2
= USE_STMT (use_p
);
655 if (is_gimple_debug (stmt2
))
657 if (!is_gimple_assign (stmt2
)
658 || gimple_assign_rhs_code (stmt2
) != RDIV_EXPR
659 || gimple_assign_rhs1 (stmt2
) == arg1
660 || gimple_assign_rhs2 (stmt2
) != arg1
)
669 gimple_replace_ssa_lhs (call
, arg1
);
670 if (gimple_call_internal_p (call
) != (ifn
!= IFN_LAST
))
672 auto_vec
<tree
, 4> args
;
673 for (unsigned int i
= 0;
674 i
< gimple_call_num_args (call
); i
++)
675 args
.safe_push (gimple_call_arg (call
, i
));
678 stmt2
= gimple_build_call_vec (fndecl
, args
);
680 stmt2
= gimple_build_call_internal_vec (ifn
, args
);
681 gimple_call_set_lhs (stmt2
, arg1
);
682 if (gimple_vdef (call
))
684 gimple_set_vdef (stmt2
, gimple_vdef (call
));
685 SSA_NAME_DEF_STMT (gimple_vdef (stmt2
)) = stmt2
;
687 gimple_set_vuse (stmt2
, gimple_vuse (call
));
688 gimple_stmt_iterator gsi2
= gsi_for_stmt (call
);
689 gsi_replace (&gsi2
, stmt2
, true);
694 gimple_call_set_fndecl (call
, fndecl
);
696 gimple_call_set_internal_fn (call
, ifn
);
699 reciprocal_stats
.rfuncs_inserted
++;
701 FOR_EACH_IMM_USE_STMT (stmt
, ui
, arg1
)
703 gimple_stmt_iterator gsi
= gsi_for_stmt (stmt
);
704 gimple_assign_set_rhs_code (stmt
, MULT_EXPR
);
705 fold_stmt_inplace (&gsi
);
713 statistics_counter_event (fun
, "reciprocal divs inserted",
714 reciprocal_stats
.rdivs_inserted
);
715 statistics_counter_event (fun
, "reciprocal functions inserted",
716 reciprocal_stats
.rfuncs_inserted
);
718 free_dominance_info (CDI_DOMINATORS
);
719 free_dominance_info (CDI_POST_DOMINATORS
);
727 make_pass_cse_reciprocals (gcc::context
*ctxt
)
729 return new pass_cse_reciprocals (ctxt
);
732 /* Records an occurrence at statement USE_STMT in the vector of trees
733 STMTS if it is dominated by *TOP_BB or dominates it or this basic block
734 is not yet initialized. Returns true if the occurrence was pushed on
735 the vector. Adjusts *TOP_BB to be the basic block dominating all
736 statements in the vector. */
739 maybe_record_sincos (vec
<gimple
*> *stmts
,
740 basic_block
*top_bb
, gimple
*use_stmt
)
742 basic_block use_bb
= gimple_bb (use_stmt
);
744 && (*top_bb
== use_bb
745 || dominated_by_p (CDI_DOMINATORS
, use_bb
, *top_bb
)))
746 stmts
->safe_push (use_stmt
);
748 || dominated_by_p (CDI_DOMINATORS
, *top_bb
, use_bb
))
750 stmts
->safe_push (use_stmt
);
759 /* Look for sin, cos and cexpi calls with the same argument NAME and
760 create a single call to cexpi CSEing the result in this case.
761 We first walk over all immediate uses of the argument collecting
762 statements that we can CSE in a vector and in a second pass replace
763 the statement rhs with a REALPART or IMAGPART expression on the
764 result of the cexpi call we insert before the use statement that
765 dominates all other candidates. */
768 execute_cse_sincos_1 (tree name
)
770 gimple_stmt_iterator gsi
;
771 imm_use_iterator use_iter
;
772 tree fndecl
, res
, type
;
773 gimple
*def_stmt
, *use_stmt
, *stmt
;
774 int seen_cos
= 0, seen_sin
= 0, seen_cexpi
= 0;
775 auto_vec
<gimple
*> stmts
;
776 basic_block top_bb
= NULL
;
778 bool cfg_changed
= false;
780 type
= TREE_TYPE (name
);
781 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, name
)
783 if (gimple_code (use_stmt
) != GIMPLE_CALL
784 || !gimple_call_lhs (use_stmt
))
787 switch (gimple_call_combined_fn (use_stmt
))
790 seen_cos
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
794 seen_sin
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
798 seen_cexpi
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
805 if (seen_cos
+ seen_sin
+ seen_cexpi
<= 1)
808 /* Simply insert cexpi at the beginning of top_bb but not earlier than
809 the name def statement. */
810 fndecl
= mathfn_built_in (type
, BUILT_IN_CEXPI
);
813 stmt
= gimple_build_call (fndecl
, 1, name
);
814 res
= make_temp_ssa_name (TREE_TYPE (TREE_TYPE (fndecl
)), stmt
, "sincostmp");
815 gimple_call_set_lhs (stmt
, res
);
817 def_stmt
= SSA_NAME_DEF_STMT (name
);
818 if (!SSA_NAME_IS_DEFAULT_DEF (name
)
819 && gimple_code (def_stmt
) != GIMPLE_PHI
820 && gimple_bb (def_stmt
) == top_bb
)
822 gsi
= gsi_for_stmt (def_stmt
);
823 gsi_insert_after (&gsi
, stmt
, GSI_SAME_STMT
);
827 gsi
= gsi_after_labels (top_bb
);
828 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
830 sincos_stats
.inserted
++;
832 /* And adjust the recorded old call sites. */
833 for (i
= 0; stmts
.iterate (i
, &use_stmt
); ++i
)
837 switch (gimple_call_combined_fn (use_stmt
))
840 rhs
= fold_build1 (REALPART_EXPR
, type
, res
);
844 rhs
= fold_build1 (IMAGPART_EXPR
, type
, res
);
855 /* Replace call with a copy. */
856 stmt
= gimple_build_assign (gimple_call_lhs (use_stmt
), rhs
);
858 gsi
= gsi_for_stmt (use_stmt
);
859 gsi_replace (&gsi
, stmt
, true);
860 if (gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
867 /* To evaluate powi(x,n), the floating point value x raised to the
868 constant integer exponent n, we use a hybrid algorithm that
869 combines the "window method" with look-up tables. For an
870 introduction to exponentiation algorithms and "addition chains",
871 see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth,
872 "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming",
873 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation
874 Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */
876 /* Provide a default value for POWI_MAX_MULTS, the maximum number of
877 multiplications to inline before calling the system library's pow
878 function. powi(x,n) requires at worst 2*bits(n)-2 multiplications,
879 so this default never requires calling pow, powf or powl. */
881 #ifndef POWI_MAX_MULTS
882 #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2)
885 /* The size of the "optimal power tree" lookup table. All
886 exponents less than this value are simply looked up in the
887 powi_table below. This threshold is also used to size the
888 cache of pseudo registers that hold intermediate results. */
889 #define POWI_TABLE_SIZE 256
891 /* The size, in bits of the window, used in the "window method"
892 exponentiation algorithm. This is equivalent to a radix of
893 (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */
894 #define POWI_WINDOW_SIZE 3
896 /* The following table is an efficient representation of an
897 "optimal power tree". For each value, i, the corresponding
898 value, j, in the table states than an optimal evaluation
899 sequence for calculating pow(x,i) can be found by evaluating
900 pow(x,j)*pow(x,i-j). An optimal power tree for the first
901 100 integers is given in Knuth's "Seminumerical algorithms". */
903 static const unsigned char powi_table
[POWI_TABLE_SIZE
] =
905 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */
906 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */
907 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */
908 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */
909 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */
910 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */
911 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */
912 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */
913 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */
914 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */
915 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */
916 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */
917 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */
918 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */
919 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */
920 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */
921 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */
922 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */
923 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */
924 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */
925 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */
926 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */
927 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */
928 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */
929 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */
930 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */
931 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */
932 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */
933 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */
934 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */
935 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */
936 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */
940 /* Return the number of multiplications required to calculate
941 powi(x,n) where n is less than POWI_TABLE_SIZE. This is a
942 subroutine of powi_cost. CACHE is an array indicating
943 which exponents have already been calculated. */
946 powi_lookup_cost (unsigned HOST_WIDE_INT n
, bool *cache
)
948 /* If we've already calculated this exponent, then this evaluation
949 doesn't require any additional multiplications. */
954 return powi_lookup_cost (n
- powi_table
[n
], cache
)
955 + powi_lookup_cost (powi_table
[n
], cache
) + 1;
958 /* Return the number of multiplications required to calculate
959 powi(x,n) for an arbitrary x, given the exponent N. This
960 function needs to be kept in sync with powi_as_mults below. */
963 powi_cost (HOST_WIDE_INT n
)
965 bool cache
[POWI_TABLE_SIZE
];
966 unsigned HOST_WIDE_INT digit
;
967 unsigned HOST_WIDE_INT val
;
973 /* Ignore the reciprocal when calculating the cost. */
974 val
= (n
< 0) ? -n
: n
;
976 /* Initialize the exponent cache. */
977 memset (cache
, 0, POWI_TABLE_SIZE
* sizeof (bool));
982 while (val
>= POWI_TABLE_SIZE
)
986 digit
= val
& ((1 << POWI_WINDOW_SIZE
) - 1);
987 result
+= powi_lookup_cost (digit
, cache
)
988 + POWI_WINDOW_SIZE
+ 1;
989 val
>>= POWI_WINDOW_SIZE
;
998 return result
+ powi_lookup_cost (val
, cache
);
1001 /* Recursive subroutine of powi_as_mults. This function takes the
1002 array, CACHE, of already calculated exponents and an exponent N and
1003 returns a tree that corresponds to CACHE[1]**N, with type TYPE. */
1006 powi_as_mults_1 (gimple_stmt_iterator
*gsi
, location_t loc
, tree type
,
1007 HOST_WIDE_INT n
, tree
*cache
)
1009 tree op0
, op1
, ssa_target
;
1010 unsigned HOST_WIDE_INT digit
;
1013 if (n
< POWI_TABLE_SIZE
&& cache
[n
])
1016 ssa_target
= make_temp_ssa_name (type
, NULL
, "powmult");
1018 if (n
< POWI_TABLE_SIZE
)
1020 cache
[n
] = ssa_target
;
1021 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
- powi_table
[n
], cache
);
1022 op1
= powi_as_mults_1 (gsi
, loc
, type
, powi_table
[n
], cache
);
1026 digit
= n
& ((1 << POWI_WINDOW_SIZE
) - 1);
1027 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
- digit
, cache
);
1028 op1
= powi_as_mults_1 (gsi
, loc
, type
, digit
, cache
);
1032 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
>> 1, cache
);
1036 mult_stmt
= gimple_build_assign (ssa_target
, MULT_EXPR
, op0
, op1
);
1037 gimple_set_location (mult_stmt
, loc
);
1038 gsi_insert_before (gsi
, mult_stmt
, GSI_SAME_STMT
);
1043 /* Convert ARG0**N to a tree of multiplications of ARG0 with itself.
1044 This function needs to be kept in sync with powi_cost above. */
1047 powi_as_mults (gimple_stmt_iterator
*gsi
, location_t loc
,
1048 tree arg0
, HOST_WIDE_INT n
)
1050 tree cache
[POWI_TABLE_SIZE
], result
, type
= TREE_TYPE (arg0
);
1055 return build_real (type
, dconst1
);
1057 memset (cache
, 0, sizeof (cache
));
1060 result
= powi_as_mults_1 (gsi
, loc
, type
, (n
< 0) ? -n
: n
, cache
);
1064 /* If the original exponent was negative, reciprocate the result. */
1065 target
= make_temp_ssa_name (type
, NULL
, "powmult");
1066 div_stmt
= gimple_build_assign (target
, RDIV_EXPR
,
1067 build_real (type
, dconst1
), result
);
1068 gimple_set_location (div_stmt
, loc
);
1069 gsi_insert_before (gsi
, div_stmt
, GSI_SAME_STMT
);
1074 /* ARG0 and N are the two arguments to a powi builtin in GSI with
1075 location info LOC. If the arguments are appropriate, create an
1076 equivalent sequence of statements prior to GSI using an optimal
1077 number of multiplications, and return an expession holding the
1081 gimple_expand_builtin_powi (gimple_stmt_iterator
*gsi
, location_t loc
,
1082 tree arg0
, HOST_WIDE_INT n
)
1084 /* Avoid largest negative number. */
1086 && ((n
>= -1 && n
<= 2)
1087 || (optimize_function_for_speed_p (cfun
)
1088 && powi_cost (n
) <= POWI_MAX_MULTS
)))
1089 return powi_as_mults (gsi
, loc
, arg0
, n
);
1094 /* Build a gimple call statement that calls FN with argument ARG.
1095 Set the lhs of the call statement to a fresh SSA name. Insert the
1096 statement prior to GSI's current position, and return the fresh
1100 build_and_insert_call (gimple_stmt_iterator
*gsi
, location_t loc
,
1106 call_stmt
= gimple_build_call (fn
, 1, arg
);
1107 ssa_target
= make_temp_ssa_name (TREE_TYPE (arg
), NULL
, "powroot");
1108 gimple_set_lhs (call_stmt
, ssa_target
);
1109 gimple_set_location (call_stmt
, loc
);
1110 gsi_insert_before (gsi
, call_stmt
, GSI_SAME_STMT
);
1115 /* Build a gimple binary operation with the given CODE and arguments
1116 ARG0, ARG1, assigning the result to a new SSA name for variable
1117 TARGET. Insert the statement prior to GSI's current position, and
1118 return the fresh SSA name.*/
1121 build_and_insert_binop (gimple_stmt_iterator
*gsi
, location_t loc
,
1122 const char *name
, enum tree_code code
,
1123 tree arg0
, tree arg1
)
1125 tree result
= make_temp_ssa_name (TREE_TYPE (arg0
), NULL
, name
);
1126 gassign
*stmt
= gimple_build_assign (result
, code
, arg0
, arg1
);
1127 gimple_set_location (stmt
, loc
);
1128 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1132 /* Build a gimple reference operation with the given CODE and argument
1133 ARG, assigning the result to a new SSA name of TYPE with NAME.
1134 Insert the statement prior to GSI's current position, and return
1135 the fresh SSA name. */
1138 build_and_insert_ref (gimple_stmt_iterator
*gsi
, location_t loc
, tree type
,
1139 const char *name
, enum tree_code code
, tree arg0
)
1141 tree result
= make_temp_ssa_name (type
, NULL
, name
);
1142 gimple
*stmt
= gimple_build_assign (result
, build1 (code
, type
, arg0
));
1143 gimple_set_location (stmt
, loc
);
1144 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1148 /* Build a gimple assignment to cast VAL to TYPE. Insert the statement
1149 prior to GSI's current position, and return the fresh SSA name. */
1152 build_and_insert_cast (gimple_stmt_iterator
*gsi
, location_t loc
,
1153 tree type
, tree val
)
1155 tree result
= make_ssa_name (type
);
1156 gassign
*stmt
= gimple_build_assign (result
, NOP_EXPR
, val
);
1157 gimple_set_location (stmt
, loc
);
1158 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1162 struct pow_synth_sqrt_info
1165 unsigned int deepest
;
1166 unsigned int num_mults
;
1169 /* Return true iff the real value C can be represented as a
1170 sum of powers of 0.5 up to N. That is:
1171 C == SUM<i from 1..N> (a[i]*(0.5**i)) where a[i] is either 0 or 1.
1172 Record in INFO the various parameters of the synthesis algorithm such
1173 as the factors a[i], the maximum 0.5 power and the number of
1174 multiplications that will be required. */
1177 representable_as_half_series_p (REAL_VALUE_TYPE c
, unsigned n
,
1178 struct pow_synth_sqrt_info
*info
)
1180 REAL_VALUE_TYPE factor
= dconsthalf
;
1181 REAL_VALUE_TYPE remainder
= c
;
1184 info
->num_mults
= 0;
1185 memset (info
->factors
, 0, n
* sizeof (bool));
1187 for (unsigned i
= 0; i
< n
; i
++)
1189 REAL_VALUE_TYPE res
;
1191 /* If something inexact happened bail out now. */
1192 if (real_arithmetic (&res
, MINUS_EXPR
, &remainder
, &factor
))
1195 /* We have hit zero. The number is representable as a sum
1196 of powers of 0.5. */
1197 if (real_equal (&res
, &dconst0
))
1199 info
->factors
[i
] = true;
1200 info
->deepest
= i
+ 1;
1203 else if (!REAL_VALUE_NEGATIVE (res
))
1206 info
->factors
[i
] = true;
1210 info
->factors
[i
] = false;
1212 real_arithmetic (&factor
, MULT_EXPR
, &factor
, &dconsthalf
);
1217 /* Return the tree corresponding to FN being applied
1218 to ARG N times at GSI and LOC.
1219 Look up previous results from CACHE if need be.
1220 cache[0] should contain just plain ARG i.e. FN applied to ARG 0 times. */
1223 get_fn_chain (tree arg
, unsigned int n
, gimple_stmt_iterator
*gsi
,
1224 tree fn
, location_t loc
, tree
*cache
)
1226 tree res
= cache
[n
];
1229 tree prev
= get_fn_chain (arg
, n
- 1, gsi
, fn
, loc
, cache
);
1230 res
= build_and_insert_call (gsi
, loc
, fn
, prev
);
1237 /* Print to STREAM the repeated application of function FNAME to ARG
1238 N times. So, for FNAME = "foo", ARG = "x", N = 2 it would print:
1242 print_nested_fn (FILE* stream
, const char *fname
, const char* arg
,
1246 fprintf (stream
, "%s", arg
);
1249 fprintf (stream
, "%s (", fname
);
1250 print_nested_fn (stream
, fname
, arg
, n
- 1);
1251 fprintf (stream
, ")");
1255 /* Print to STREAM the fractional sequence of sqrt chains
1256 applied to ARG, described by INFO. Used for the dump file. */
1259 dump_fractional_sqrt_sequence (FILE *stream
, const char *arg
,
1260 struct pow_synth_sqrt_info
*info
)
1262 for (unsigned int i
= 0; i
< info
->deepest
; i
++)
1264 bool is_set
= info
->factors
[i
];
1267 print_nested_fn (stream
, "sqrt", arg
, i
+ 1);
1268 if (i
!= info
->deepest
- 1)
1269 fprintf (stream
, " * ");
1274 /* Print to STREAM a representation of raising ARG to an integer
1275 power N. Used for the dump file. */
1278 dump_integer_part (FILE *stream
, const char* arg
, HOST_WIDE_INT n
)
1281 fprintf (stream
, "powi (%s, " HOST_WIDE_INT_PRINT_DEC
")", arg
, n
);
1283 fprintf (stream
, "%s", arg
);
1286 /* Attempt to synthesize a POW[F] (ARG0, ARG1) call using chains of
1287 square roots. Place at GSI and LOC. Limit the maximum depth
1288 of the sqrt chains to MAX_DEPTH. Return the tree holding the
1289 result of the expanded sequence or NULL_TREE if the expansion failed.
1291 This routine assumes that ARG1 is a real number with a fractional part
1292 (the integer exponent case will have been handled earlier in
1293 gimple_expand_builtin_pow).
1296 * For ARG1 composed of a whole part WHOLE_PART and a fractional part
1297 FRAC_PART i.e. WHOLE_PART == floor (ARG1) and
1298 FRAC_PART == ARG1 - WHOLE_PART:
1299 Produce POWI (ARG0, WHOLE_PART) * POW (ARG0, FRAC_PART) where
1300 POW (ARG0, FRAC_PART) is expanded as a product of square root chains
1301 if it can be expressed as such, that is if FRAC_PART satisfies:
1302 FRAC_PART == <SUM from i = 1 until MAX_DEPTH> (a[i] * (0.5**i))
1303 where integer a[i] is either 0 or 1.
1306 POW (x, 3.625) == POWI (x, 3) * POW (x, 0.625)
1307 --> POWI (x, 3) * SQRT (x) * SQRT (SQRT (SQRT (x)))
1309 For ARG1 < 0.0 there are two approaches:
1310 * (A) Expand to 1.0 / POW (ARG0, -ARG1) where POW (ARG0, -ARG1)
1311 is calculated as above.
1314 POW (x, -5.625) == 1.0 / POW (x, 5.625)
1315 --> 1.0 / (POWI (x, 5) * SQRT (x) * SQRT (SQRT (SQRT (x))))
1317 * (B) : WHOLE_PART := - ceil (abs (ARG1))
1318 FRAC_PART := ARG1 - WHOLE_PART
1319 and expand to POW (x, FRAC_PART) / POWI (x, WHOLE_PART).
1321 POW (x, -5.875) == POW (x, 0.125) / POWI (X, 6)
1322 --> SQRT (SQRT (SQRT (x))) / (POWI (x, 6))
1324 For ARG1 < 0.0 we choose between (A) and (B) depending on
1325 how many multiplications we'd have to do.
1326 So, for the example in (B): POW (x, -5.875), if we were to
1327 follow algorithm (A) we would produce:
1328 1.0 / POWI (X, 5) * SQRT (X) * SQRT (SQRT (X)) * SQRT (SQRT (SQRT (X)))
1329 which contains more multiplications than approach (B).
1331 Hopefully, this approach will eliminate potentially expensive POW library
1332 calls when unsafe floating point math is enabled and allow the compiler to
1333 further optimise the multiplies, square roots and divides produced by this
1337 expand_pow_as_sqrts (gimple_stmt_iterator
*gsi
, location_t loc
,
1338 tree arg0
, tree arg1
, HOST_WIDE_INT max_depth
)
1340 tree type
= TREE_TYPE (arg0
);
1341 machine_mode mode
= TYPE_MODE (type
);
1342 tree sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1343 bool one_over
= true;
1348 if (TREE_CODE (arg1
) != REAL_CST
)
1351 REAL_VALUE_TYPE exp_init
= TREE_REAL_CST (arg1
);
1353 gcc_assert (max_depth
> 0);
1354 tree
*cache
= XALLOCAVEC (tree
, max_depth
+ 1);
1356 struct pow_synth_sqrt_info synth_info
;
1357 synth_info
.factors
= XALLOCAVEC (bool, max_depth
+ 1);
1358 synth_info
.deepest
= 0;
1359 synth_info
.num_mults
= 0;
1361 bool neg_exp
= REAL_VALUE_NEGATIVE (exp_init
);
1362 REAL_VALUE_TYPE exp
= real_value_abs (&exp_init
);
1364 /* The whole and fractional parts of exp. */
1365 REAL_VALUE_TYPE whole_part
;
1366 REAL_VALUE_TYPE frac_part
;
1368 real_floor (&whole_part
, mode
, &exp
);
1369 real_arithmetic (&frac_part
, MINUS_EXPR
, &exp
, &whole_part
);
1372 REAL_VALUE_TYPE ceil_whole
= dconst0
;
1373 REAL_VALUE_TYPE ceil_fract
= dconst0
;
1377 real_ceil (&ceil_whole
, mode
, &exp
);
1378 real_arithmetic (&ceil_fract
, MINUS_EXPR
, &ceil_whole
, &exp
);
1381 if (!representable_as_half_series_p (frac_part
, max_depth
, &synth_info
))
1384 /* Check whether it's more profitable to not use 1.0 / ... */
1387 struct pow_synth_sqrt_info alt_synth_info
;
1388 alt_synth_info
.factors
= XALLOCAVEC (bool, max_depth
+ 1);
1389 alt_synth_info
.deepest
= 0;
1390 alt_synth_info
.num_mults
= 0;
1392 if (representable_as_half_series_p (ceil_fract
, max_depth
,
1394 && alt_synth_info
.deepest
<= synth_info
.deepest
1395 && alt_synth_info
.num_mults
< synth_info
.num_mults
)
1397 whole_part
= ceil_whole
;
1398 frac_part
= ceil_fract
;
1399 synth_info
.deepest
= alt_synth_info
.deepest
;
1400 synth_info
.num_mults
= alt_synth_info
.num_mults
;
1401 memcpy (synth_info
.factors
, alt_synth_info
.factors
,
1402 (max_depth
+ 1) * sizeof (bool));
1407 HOST_WIDE_INT n
= real_to_integer (&whole_part
);
1408 REAL_VALUE_TYPE cint
;
1409 real_from_integer (&cint
, VOIDmode
, n
, SIGNED
);
1411 if (!real_identical (&whole_part
, &cint
))
1414 if (powi_cost (n
) + synth_info
.num_mults
> POWI_MAX_MULTS
)
1417 memset (cache
, 0, (max_depth
+ 1) * sizeof (tree
));
1419 tree integer_res
= n
== 0 ? build_real (type
, dconst1
) : arg0
;
1421 /* Calculate the integer part of the exponent. */
1424 integer_res
= gimple_expand_builtin_powi (gsi
, loc
, arg0
, n
);
1433 real_to_decimal (string
, &exp_init
, sizeof (string
), 0, 1);
1434 fprintf (dump_file
, "synthesizing pow (x, %s) as:\n", string
);
1440 fprintf (dump_file
, "1.0 / (");
1441 dump_integer_part (dump_file
, "x", n
);
1443 fprintf (dump_file
, " * ");
1444 dump_fractional_sqrt_sequence (dump_file
, "x", &synth_info
);
1445 fprintf (dump_file
, ")");
1449 dump_fractional_sqrt_sequence (dump_file
, "x", &synth_info
);
1450 fprintf (dump_file
, " / (");
1451 dump_integer_part (dump_file
, "x", n
);
1452 fprintf (dump_file
, ")");
1457 dump_fractional_sqrt_sequence (dump_file
, "x", &synth_info
);
1459 fprintf (dump_file
, " * ");
1460 dump_integer_part (dump_file
, "x", n
);
1463 fprintf (dump_file
, "\ndeepest sqrt chain: %d\n", synth_info
.deepest
);
1467 tree fract_res
= NULL_TREE
;
1470 /* Calculate the fractional part of the exponent. */
1471 for (unsigned i
= 0; i
< synth_info
.deepest
; i
++)
1473 if (synth_info
.factors
[i
])
1475 tree sqrt_chain
= get_fn_chain (arg0
, i
+ 1, gsi
, sqrtfn
, loc
, cache
);
1478 fract_res
= sqrt_chain
;
1481 fract_res
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1482 fract_res
, sqrt_chain
);
1486 tree res
= NULL_TREE
;
1493 res
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1494 fract_res
, integer_res
);
1498 res
= build_and_insert_binop (gsi
, loc
, "powrootrecip", RDIV_EXPR
,
1499 build_real (type
, dconst1
), res
);
1503 res
= build_and_insert_binop (gsi
, loc
, "powroot", RDIV_EXPR
,
1504 fract_res
, integer_res
);
1508 res
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1509 fract_res
, integer_res
);
1513 /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI
1514 with location info LOC. If possible, create an equivalent and
1515 less expensive sequence of statements prior to GSI, and return an
1516 expession holding the result. */
1519 gimple_expand_builtin_pow (gimple_stmt_iterator
*gsi
, location_t loc
,
1520 tree arg0
, tree arg1
)
1522 REAL_VALUE_TYPE c
, cint
, dconst1_3
, dconst1_4
, dconst1_6
;
1523 REAL_VALUE_TYPE c2
, dconst3
;
1525 tree type
, sqrtfn
, cbrtfn
, sqrt_arg0
, result
, cbrt_x
, powi_cbrt_x
;
1527 bool speed_p
= optimize_bb_for_speed_p (gsi_bb (*gsi
));
1528 bool hw_sqrt_exists
, c_is_int
, c2_is_int
;
1530 dconst1_4
= dconst1
;
1531 SET_REAL_EXP (&dconst1_4
, REAL_EXP (&dconst1_4
) - 2);
1533 /* If the exponent isn't a constant, there's nothing of interest
1535 if (TREE_CODE (arg1
) != REAL_CST
)
1538 /* Don't perform the operation if flag_signaling_nans is on
1539 and the operand is a signaling NaN. */
1540 if (HONOR_SNANS (TYPE_MODE (TREE_TYPE (arg1
)))
1541 && ((TREE_CODE (arg0
) == REAL_CST
1542 && REAL_VALUE_ISSIGNALING_NAN (TREE_REAL_CST (arg0
)))
1543 || REAL_VALUE_ISSIGNALING_NAN (TREE_REAL_CST (arg1
))))
1546 /* If the exponent is equivalent to an integer, expand to an optimal
1547 multiplication sequence when profitable. */
1548 c
= TREE_REAL_CST (arg1
);
1549 n
= real_to_integer (&c
);
1550 real_from_integer (&cint
, VOIDmode
, n
, SIGNED
);
1551 c_is_int
= real_identical (&c
, &cint
);
1554 && ((n
>= -1 && n
<= 2)
1555 || (flag_unsafe_math_optimizations
1557 && powi_cost (n
) <= POWI_MAX_MULTS
)))
1558 return gimple_expand_builtin_powi (gsi
, loc
, arg0
, n
);
1560 /* Attempt various optimizations using sqrt and cbrt. */
1561 type
= TREE_TYPE (arg0
);
1562 mode
= TYPE_MODE (type
);
1563 sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1565 /* Optimize pow(x,0.5) = sqrt(x). This replacement is always safe
1566 unless signed zeros must be maintained. pow(-0,0.5) = +0, while
1569 && real_equal (&c
, &dconsthalf
)
1570 && !HONOR_SIGNED_ZEROS (mode
))
1571 return build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1573 hw_sqrt_exists
= optab_handler (sqrt_optab
, mode
) != CODE_FOR_nothing
;
1575 /* Optimize pow(x,1./3.) = cbrt(x). This requires unsafe math
1576 optimizations since 1./3. is not exactly representable. If x
1577 is negative and finite, the correct value of pow(x,1./3.) is
1578 a NaN with the "invalid" exception raised, because the value
1579 of 1./3. actually has an even denominator. The correct value
1580 of cbrt(x) is a negative real value. */
1581 cbrtfn
= mathfn_built_in (type
, BUILT_IN_CBRT
);
1582 dconst1_3
= real_value_truncate (mode
, dconst_third ());
1584 if (flag_unsafe_math_optimizations
1586 && (!HONOR_NANS (mode
) || tree_expr_nonnegative_p (arg0
))
1587 && real_equal (&c
, &dconst1_3
))
1588 return build_and_insert_call (gsi
, loc
, cbrtfn
, arg0
);
1590 /* Optimize pow(x,1./6.) = cbrt(sqrt(x)). Don't do this optimization
1591 if we don't have a hardware sqrt insn. */
1592 dconst1_6
= dconst1_3
;
1593 SET_REAL_EXP (&dconst1_6
, REAL_EXP (&dconst1_6
) - 1);
1595 if (flag_unsafe_math_optimizations
1598 && (!HONOR_NANS (mode
) || tree_expr_nonnegative_p (arg0
))
1601 && real_equal (&c
, &dconst1_6
))
1604 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1607 return build_and_insert_call (gsi
, loc
, cbrtfn
, sqrt_arg0
);
1611 /* Attempt to expand the POW as a product of square root chains.
1612 Expand the 0.25 case even when otpimising for size. */
1613 if (flag_unsafe_math_optimizations
1616 && (speed_p
|| real_equal (&c
, &dconst1_4
))
1617 && !HONOR_SIGNED_ZEROS (mode
))
1619 unsigned int max_depth
= speed_p
1620 ? PARAM_VALUE (PARAM_MAX_POW_SQRT_DEPTH
)
1623 tree expand_with_sqrts
1624 = expand_pow_as_sqrts (gsi
, loc
, arg0
, arg1
, max_depth
);
1626 if (expand_with_sqrts
)
1627 return expand_with_sqrts
;
1630 real_arithmetic (&c2
, MULT_EXPR
, &c
, &dconst2
);
1631 n
= real_to_integer (&c2
);
1632 real_from_integer (&cint
, VOIDmode
, n
, SIGNED
);
1633 c2_is_int
= real_identical (&c2
, &cint
);
1635 /* Optimize pow(x,c), where 3c = n for some nonzero integer n, into
1637 powi(x, n/3) * powi(cbrt(x), n%3), n > 0;
1638 1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)), n < 0.
1640 Do not calculate the first factor when n/3 = 0. As cbrt(x) is
1641 different from pow(x, 1./3.) due to rounding and behavior with
1642 negative x, we need to constrain this transformation to unsafe
1643 math and positive x or finite math. */
1644 real_from_integer (&dconst3
, VOIDmode
, 3, SIGNED
);
1645 real_arithmetic (&c2
, MULT_EXPR
, &c
, &dconst3
);
1646 real_round (&c2
, mode
, &c2
);
1647 n
= real_to_integer (&c2
);
1648 real_from_integer (&cint
, VOIDmode
, n
, SIGNED
);
1649 real_arithmetic (&c2
, RDIV_EXPR
, &cint
, &dconst3
);
1650 real_convert (&c2
, mode
, &c2
);
1652 if (flag_unsafe_math_optimizations
1654 && (!HONOR_NANS (mode
) || tree_expr_nonnegative_p (arg0
))
1655 && real_identical (&c2
, &c
)
1657 && optimize_function_for_speed_p (cfun
)
1658 && powi_cost (n
/ 3) <= POWI_MAX_MULTS
)
1660 tree powi_x_ndiv3
= NULL_TREE
;
1662 /* Attempt to fold powi(arg0, abs(n/3)) into multiplies. If not
1663 possible or profitable, give up. Skip the degenerate case when
1664 abs(n) < 3, where the result is always 1. */
1665 if (absu_hwi (n
) >= 3)
1667 powi_x_ndiv3
= gimple_expand_builtin_powi (gsi
, loc
, arg0
,
1673 /* Calculate powi(cbrt(x), n%3). Don't use gimple_expand_builtin_powi
1674 as that creates an unnecessary variable. Instead, just produce
1675 either cbrt(x) or cbrt(x) * cbrt(x). */
1676 cbrt_x
= build_and_insert_call (gsi
, loc
, cbrtfn
, arg0
);
1678 if (absu_hwi (n
) % 3 == 1)
1679 powi_cbrt_x
= cbrt_x
;
1681 powi_cbrt_x
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1684 /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1. */
1685 if (absu_hwi (n
) < 3)
1686 result
= powi_cbrt_x
;
1688 result
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1689 powi_x_ndiv3
, powi_cbrt_x
);
1691 /* If n is negative, reciprocate the result. */
1693 result
= build_and_insert_binop (gsi
, loc
, "powroot", RDIV_EXPR
,
1694 build_real (type
, dconst1
), result
);
1699 /* No optimizations succeeded. */
1703 /* ARG is the argument to a cabs builtin call in GSI with location info
1704 LOC. Create a sequence of statements prior to GSI that calculates
1705 sqrt(R*R + I*I), where R and I are the real and imaginary components
1706 of ARG, respectively. Return an expression holding the result. */
1709 gimple_expand_builtin_cabs (gimple_stmt_iterator
*gsi
, location_t loc
, tree arg
)
1711 tree real_part
, imag_part
, addend1
, addend2
, sum
, result
;
1712 tree type
= TREE_TYPE (TREE_TYPE (arg
));
1713 tree sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1714 machine_mode mode
= TYPE_MODE (type
);
1716 if (!flag_unsafe_math_optimizations
1717 || !optimize_bb_for_speed_p (gimple_bb (gsi_stmt (*gsi
)))
1719 || optab_handler (sqrt_optab
, mode
) == CODE_FOR_nothing
)
1722 real_part
= build_and_insert_ref (gsi
, loc
, type
, "cabs",
1723 REALPART_EXPR
, arg
);
1724 addend1
= build_and_insert_binop (gsi
, loc
, "cabs", MULT_EXPR
,
1725 real_part
, real_part
);
1726 imag_part
= build_and_insert_ref (gsi
, loc
, type
, "cabs",
1727 IMAGPART_EXPR
, arg
);
1728 addend2
= build_and_insert_binop (gsi
, loc
, "cabs", MULT_EXPR
,
1729 imag_part
, imag_part
);
1730 sum
= build_and_insert_binop (gsi
, loc
, "cabs", PLUS_EXPR
, addend1
, addend2
);
1731 result
= build_and_insert_call (gsi
, loc
, sqrtfn
, sum
);
1736 /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1
1737 on the SSA_NAME argument of each of them. Also expand powi(x,n) into
1738 an optimal number of multiplies, when n is a constant. */
1742 const pass_data pass_data_cse_sincos
=
1744 GIMPLE_PASS
, /* type */
1745 "sincos", /* name */
1746 OPTGROUP_NONE
, /* optinfo_flags */
1747 TV_NONE
, /* tv_id */
1748 PROP_ssa
, /* properties_required */
1749 PROP_gimple_opt_math
, /* properties_provided */
1750 0, /* properties_destroyed */
1751 0, /* todo_flags_start */
1752 TODO_update_ssa
, /* todo_flags_finish */
1755 class pass_cse_sincos
: public gimple_opt_pass
1758 pass_cse_sincos (gcc::context
*ctxt
)
1759 : gimple_opt_pass (pass_data_cse_sincos
, ctxt
)
1762 /* opt_pass methods: */
1763 virtual bool gate (function
*)
1765 /* We no longer require either sincos or cexp, since powi expansion
1766 piggybacks on this pass. */
1770 virtual unsigned int execute (function
*);
1772 }; // class pass_cse_sincos
1775 pass_cse_sincos::execute (function
*fun
)
1778 bool cfg_changed
= false;
1780 calculate_dominance_info (CDI_DOMINATORS
);
1781 memset (&sincos_stats
, 0, sizeof (sincos_stats
));
1783 FOR_EACH_BB_FN (bb
, fun
)
1785 gimple_stmt_iterator gsi
;
1786 bool cleanup_eh
= false;
1788 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1790 gimple
*stmt
= gsi_stmt (gsi
);
1792 /* Only the last stmt in a bb could throw, no need to call
1793 gimple_purge_dead_eh_edges if we change something in the middle
1794 of a basic block. */
1797 if (is_gimple_call (stmt
)
1798 && gimple_call_lhs (stmt
))
1800 tree arg
, arg0
, arg1
, result
;
1804 switch (gimple_call_combined_fn (stmt
))
1809 /* Make sure we have either sincos or cexp. */
1810 if (!targetm
.libc_has_function (function_c99_math_complex
)
1811 && !targetm
.libc_has_function (function_sincos
))
1814 arg
= gimple_call_arg (stmt
, 0);
1815 if (TREE_CODE (arg
) == SSA_NAME
)
1816 cfg_changed
|= execute_cse_sincos_1 (arg
);
1820 arg0
= gimple_call_arg (stmt
, 0);
1821 arg1
= gimple_call_arg (stmt
, 1);
1823 loc
= gimple_location (stmt
);
1824 result
= gimple_expand_builtin_pow (&gsi
, loc
, arg0
, arg1
);
1828 tree lhs
= gimple_get_lhs (stmt
);
1829 gassign
*new_stmt
= gimple_build_assign (lhs
, result
);
1830 gimple_set_location (new_stmt
, loc
);
1831 unlink_stmt_vdef (stmt
);
1832 gsi_replace (&gsi
, new_stmt
, true);
1834 if (gimple_vdef (stmt
))
1835 release_ssa_name (gimple_vdef (stmt
));
1840 arg0
= gimple_call_arg (stmt
, 0);
1841 arg1
= gimple_call_arg (stmt
, 1);
1842 loc
= gimple_location (stmt
);
1844 if (real_minus_onep (arg0
))
1846 tree t0
, t1
, cond
, one
, minus_one
;
1849 t0
= TREE_TYPE (arg0
);
1850 t1
= TREE_TYPE (arg1
);
1851 one
= build_real (t0
, dconst1
);
1852 minus_one
= build_real (t0
, dconstm1
);
1854 cond
= make_temp_ssa_name (t1
, NULL
, "powi_cond");
1855 stmt
= gimple_build_assign (cond
, BIT_AND_EXPR
,
1856 arg1
, build_int_cst (t1
, 1));
1857 gimple_set_location (stmt
, loc
);
1858 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1860 result
= make_temp_ssa_name (t0
, NULL
, "powi");
1861 stmt
= gimple_build_assign (result
, COND_EXPR
, cond
,
1863 gimple_set_location (stmt
, loc
);
1864 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
1868 if (!tree_fits_shwi_p (arg1
))
1871 n
= tree_to_shwi (arg1
);
1872 result
= gimple_expand_builtin_powi (&gsi
, loc
, arg0
, n
);
1877 tree lhs
= gimple_get_lhs (stmt
);
1878 gassign
*new_stmt
= gimple_build_assign (lhs
, result
);
1879 gimple_set_location (new_stmt
, loc
);
1880 unlink_stmt_vdef (stmt
);
1881 gsi_replace (&gsi
, new_stmt
, true);
1883 if (gimple_vdef (stmt
))
1884 release_ssa_name (gimple_vdef (stmt
));
1889 arg0
= gimple_call_arg (stmt
, 0);
1890 loc
= gimple_location (stmt
);
1891 result
= gimple_expand_builtin_cabs (&gsi
, loc
, arg0
);
1895 tree lhs
= gimple_get_lhs (stmt
);
1896 gassign
*new_stmt
= gimple_build_assign (lhs
, result
);
1897 gimple_set_location (new_stmt
, loc
);
1898 unlink_stmt_vdef (stmt
);
1899 gsi_replace (&gsi
, new_stmt
, true);
1901 if (gimple_vdef (stmt
))
1902 release_ssa_name (gimple_vdef (stmt
));
1911 cfg_changed
|= gimple_purge_dead_eh_edges (bb
);
1914 statistics_counter_event (fun
, "sincos statements inserted",
1915 sincos_stats
.inserted
);
1917 return cfg_changed
? TODO_cleanup_cfg
: 0;
1923 make_pass_cse_sincos (gcc::context
*ctxt
)
1925 return new pass_cse_sincos (ctxt
);
1928 /* A symbolic number is used to detect byte permutation and selection
1929 patterns. Therefore the field N contains an artificial number
1930 consisting of octet sized markers:
1932 0 - target byte has the value 0
1933 FF - target byte has an unknown value (eg. due to sign extension)
1934 1..size - marker value is the target byte index minus one.
1936 To detect permutations on memory sources (arrays and structures), a symbolic
1937 number is also associated a base address (the array or structure the load is
1938 made from), an offset from the base address and a range which gives the
1939 difference between the highest and lowest accessed memory location to make
1940 such a symbolic number. The range is thus different from size which reflects
1941 the size of the type of current expression. Note that for non memory source,
1942 range holds the same value as size.
1944 For instance, for an array char a[], (short) a[0] | (short) a[3] would have
1945 a size of 2 but a range of 4 while (short) a[0] | ((short) a[0] << 1) would
1946 still have a size of 2 but this time a range of 1. */
1948 struct symbolic_number
{
1953 HOST_WIDE_INT bytepos
;
1956 unsigned HOST_WIDE_INT range
;
1959 #define BITS_PER_MARKER 8
1960 #define MARKER_MASK ((1 << BITS_PER_MARKER) - 1)
1961 #define MARKER_BYTE_UNKNOWN MARKER_MASK
1962 #define HEAD_MARKER(n, size) \
1963 ((n) & ((uint64_t) MARKER_MASK << (((size) - 1) * BITS_PER_MARKER)))
1965 /* The number which the find_bswap_or_nop_1 result should match in
1966 order to have a nop. The number is masked according to the size of
1967 the symbolic number before using it. */
1968 #define CMPNOP (sizeof (int64_t) < 8 ? 0 : \
1969 (uint64_t)0x08070605 << 32 | 0x04030201)
1971 /* The number which the find_bswap_or_nop_1 result should match in
1972 order to have a byte swap. The number is masked according to the
1973 size of the symbolic number before using it. */
1974 #define CMPXCHG (sizeof (int64_t) < 8 ? 0 : \
1975 (uint64_t)0x01020304 << 32 | 0x05060708)
1977 /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic
1978 number N. Return false if the requested operation is not permitted
1979 on a symbolic number. */
1982 do_shift_rotate (enum tree_code code
,
1983 struct symbolic_number
*n
,
1986 int i
, size
= TYPE_PRECISION (n
->type
) / BITS_PER_UNIT
;
1987 unsigned head_marker
;
1989 if (count
% BITS_PER_UNIT
!= 0)
1991 count
= (count
/ BITS_PER_UNIT
) * BITS_PER_MARKER
;
1993 /* Zero out the extra bits of N in order to avoid them being shifted
1994 into the significant bits. */
1995 if (size
< 64 / BITS_PER_MARKER
)
1996 n
->n
&= ((uint64_t) 1 << (size
* BITS_PER_MARKER
)) - 1;
2004 head_marker
= HEAD_MARKER (n
->n
, size
);
2006 /* Arithmetic shift of signed type: result is dependent on the value. */
2007 if (!TYPE_UNSIGNED (n
->type
) && head_marker
)
2008 for (i
= 0; i
< count
/ BITS_PER_MARKER
; i
++)
2009 n
->n
|= (uint64_t) MARKER_BYTE_UNKNOWN
2010 << ((size
- 1 - i
) * BITS_PER_MARKER
);
2013 n
->n
= (n
->n
<< count
) | (n
->n
>> ((size
* BITS_PER_MARKER
) - count
));
2016 n
->n
= (n
->n
>> count
) | (n
->n
<< ((size
* BITS_PER_MARKER
) - count
));
2021 /* Zero unused bits for size. */
2022 if (size
< 64 / BITS_PER_MARKER
)
2023 n
->n
&= ((uint64_t) 1 << (size
* BITS_PER_MARKER
)) - 1;
2027 /* Perform sanity checking for the symbolic number N and the gimple
2031 verify_symbolic_number_p (struct symbolic_number
*n
, gimple
*stmt
)
2035 lhs_type
= gimple_expr_type (stmt
);
2037 if (TREE_CODE (lhs_type
) != INTEGER_TYPE
)
2040 if (TYPE_PRECISION (lhs_type
) != TYPE_PRECISION (n
->type
))
2046 /* Initialize the symbolic number N for the bswap pass from the base element
2047 SRC manipulated by the bitwise OR expression. */
2050 init_symbolic_number (struct symbolic_number
*n
, tree src
)
2054 n
->base_addr
= n
->offset
= n
->alias_set
= n
->vuse
= NULL_TREE
;
2056 /* Set up the symbolic number N by setting each byte to a value between 1 and
2057 the byte size of rhs1. The highest order byte is set to n->size and the
2058 lowest order byte to 1. */
2059 n
->type
= TREE_TYPE (src
);
2060 size
= TYPE_PRECISION (n
->type
);
2061 if (size
% BITS_PER_UNIT
!= 0)
2063 size
/= BITS_PER_UNIT
;
2064 if (size
> 64 / BITS_PER_MARKER
)
2069 if (size
< 64 / BITS_PER_MARKER
)
2070 n
->n
&= ((uint64_t) 1 << (size
* BITS_PER_MARKER
)) - 1;
2075 /* Check if STMT might be a byte swap or a nop from a memory source and returns
2076 the answer. If so, REF is that memory source and the base of the memory area
2077 accessed and the offset of the access from that base are recorded in N. */
2080 find_bswap_or_nop_load (gimple
*stmt
, tree ref
, struct symbolic_number
*n
)
2082 /* Leaf node is an array or component ref. Memorize its base and
2083 offset from base to compare to other such leaf node. */
2084 HOST_WIDE_INT bitsize
, bitpos
;
2086 int unsignedp
, reversep
, volatilep
;
2087 tree offset
, base_addr
;
2089 /* Not prepared to handle PDP endian. */
2090 if (BYTES_BIG_ENDIAN
!= WORDS_BIG_ENDIAN
)
2093 if (!gimple_assign_load_p (stmt
) || gimple_has_volatile_ops (stmt
))
2096 base_addr
= get_inner_reference (ref
, &bitsize
, &bitpos
, &offset
, &mode
,
2097 &unsignedp
, &reversep
, &volatilep
, false);
2099 if (TREE_CODE (base_addr
) == MEM_REF
)
2101 offset_int bit_offset
= 0;
2102 tree off
= TREE_OPERAND (base_addr
, 1);
2104 if (!integer_zerop (off
))
2106 offset_int boff
, coff
= mem_ref_offset (base_addr
);
2107 boff
= wi::lshift (coff
, LOG2_BITS_PER_UNIT
);
2111 base_addr
= TREE_OPERAND (base_addr
, 0);
2113 /* Avoid returning a negative bitpos as this may wreak havoc later. */
2114 if (wi::neg_p (bit_offset
))
2116 offset_int mask
= wi::mask
<offset_int
> (LOG2_BITS_PER_UNIT
, false);
2117 offset_int tem
= bit_offset
.and_not (mask
);
2118 /* TEM is the bitpos rounded to BITS_PER_UNIT towards -Inf.
2119 Subtract it to BIT_OFFSET and add it (scaled) to OFFSET. */
2121 tem
= wi::arshift (tem
, LOG2_BITS_PER_UNIT
);
2123 offset
= size_binop (PLUS_EXPR
, offset
,
2124 wide_int_to_tree (sizetype
, tem
));
2126 offset
= wide_int_to_tree (sizetype
, tem
);
2129 bitpos
+= bit_offset
.to_shwi ();
2132 if (bitpos
% BITS_PER_UNIT
)
2134 if (bitsize
% BITS_PER_UNIT
)
2139 if (!init_symbolic_number (n
, ref
))
2141 n
->base_addr
= base_addr
;
2143 n
->bytepos
= bitpos
/ BITS_PER_UNIT
;
2144 n
->alias_set
= reference_alias_ptr_type (ref
);
2145 n
->vuse
= gimple_vuse (stmt
);
2149 /* Compute the symbolic number N representing the result of a bitwise OR on 2
2150 symbolic number N1 and N2 whose source statements are respectively
2151 SOURCE_STMT1 and SOURCE_STMT2. */
2154 perform_symbolic_merge (gimple
*source_stmt1
, struct symbolic_number
*n1
,
2155 gimple
*source_stmt2
, struct symbolic_number
*n2
,
2156 struct symbolic_number
*n
)
2160 gimple
*source_stmt
;
2161 struct symbolic_number
*n_start
;
2163 /* Sources are different, cancel bswap if they are not memory location with
2164 the same base (array, structure, ...). */
2165 if (gimple_assign_rhs1 (source_stmt1
) != gimple_assign_rhs1 (source_stmt2
))
2168 HOST_WIDE_INT start_sub
, end_sub
, end1
, end2
, end
;
2169 struct symbolic_number
*toinc_n_ptr
, *n_end
;
2171 if (!n1
->base_addr
|| !n2
->base_addr
2172 || !operand_equal_p (n1
->base_addr
, n2
->base_addr
, 0))
2175 if (!n1
->offset
!= !n2
->offset
2176 || (n1
->offset
&& !operand_equal_p (n1
->offset
, n2
->offset
, 0)))
2179 if (n1
->bytepos
< n2
->bytepos
)
2182 start_sub
= n2
->bytepos
- n1
->bytepos
;
2183 source_stmt
= source_stmt1
;
2188 start_sub
= n1
->bytepos
- n2
->bytepos
;
2189 source_stmt
= source_stmt2
;
2192 /* Find the highest address at which a load is performed and
2193 compute related info. */
2194 end1
= n1
->bytepos
+ (n1
->range
- 1);
2195 end2
= n2
->bytepos
+ (n2
->range
- 1);
2199 end_sub
= end2
- end1
;
2204 end_sub
= end1
- end2
;
2206 n_end
= (end2
> end1
) ? n2
: n1
;
2208 /* Find symbolic number whose lsb is the most significant. */
2209 if (BYTES_BIG_ENDIAN
)
2210 toinc_n_ptr
= (n_end
== n1
) ? n2
: n1
;
2212 toinc_n_ptr
= (n_start
== n1
) ? n2
: n1
;
2214 n
->range
= end
- n_start
->bytepos
+ 1;
2216 /* Check that the range of memory covered can be represented by
2217 a symbolic number. */
2218 if (n
->range
> 64 / BITS_PER_MARKER
)
2221 /* Reinterpret byte marks in symbolic number holding the value of
2222 bigger weight according to target endianness. */
2223 inc
= BYTES_BIG_ENDIAN
? end_sub
: start_sub
;
2224 size
= TYPE_PRECISION (n1
->type
) / BITS_PER_UNIT
;
2225 for (i
= 0; i
< size
; i
++, inc
<<= BITS_PER_MARKER
)
2228 = (toinc_n_ptr
->n
>> (i
* BITS_PER_MARKER
)) & MARKER_MASK
;
2229 if (marker
&& marker
!= MARKER_BYTE_UNKNOWN
)
2230 toinc_n_ptr
->n
+= inc
;
2235 n
->range
= n1
->range
;
2237 source_stmt
= source_stmt1
;
2241 || alias_ptr_types_compatible_p (n1
->alias_set
, n2
->alias_set
))
2242 n
->alias_set
= n1
->alias_set
;
2244 n
->alias_set
= ptr_type_node
;
2245 n
->vuse
= n_start
->vuse
;
2246 n
->base_addr
= n_start
->base_addr
;
2247 n
->offset
= n_start
->offset
;
2248 n
->bytepos
= n_start
->bytepos
;
2249 n
->type
= n_start
->type
;
2250 size
= TYPE_PRECISION (n
->type
) / BITS_PER_UNIT
;
2252 for (i
= 0, mask
= MARKER_MASK
; i
< size
; i
++, mask
<<= BITS_PER_MARKER
)
2254 uint64_t masked1
, masked2
;
2256 masked1
= n1
->n
& mask
;
2257 masked2
= n2
->n
& mask
;
2258 if (masked1
&& masked2
&& masked1
!= masked2
)
2261 n
->n
= n1
->n
| n2
->n
;
2266 /* find_bswap_or_nop_1 invokes itself recursively with N and tries to perform
2267 the operation given by the rhs of STMT on the result. If the operation
2268 could successfully be executed the function returns a gimple stmt whose
2269 rhs's first tree is the expression of the source operand and NULL
2273 find_bswap_or_nop_1 (gimple
*stmt
, struct symbolic_number
*n
, int limit
)
2275 enum tree_code code
;
2276 tree rhs1
, rhs2
= NULL
;
2277 gimple
*rhs1_stmt
, *rhs2_stmt
, *source_stmt1
;
2278 enum gimple_rhs_class rhs_class
;
2280 if (!limit
|| !is_gimple_assign (stmt
))
2283 rhs1
= gimple_assign_rhs1 (stmt
);
2285 if (find_bswap_or_nop_load (stmt
, rhs1
, n
))
2288 if (TREE_CODE (rhs1
) != SSA_NAME
)
2291 code
= gimple_assign_rhs_code (stmt
);
2292 rhs_class
= gimple_assign_rhs_class (stmt
);
2293 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
2295 if (rhs_class
== GIMPLE_BINARY_RHS
)
2296 rhs2
= gimple_assign_rhs2 (stmt
);
2298 /* Handle unary rhs and binary rhs with integer constants as second
2301 if (rhs_class
== GIMPLE_UNARY_RHS
2302 || (rhs_class
== GIMPLE_BINARY_RHS
2303 && TREE_CODE (rhs2
) == INTEGER_CST
))
2305 if (code
!= BIT_AND_EXPR
2306 && code
!= LSHIFT_EXPR
2307 && code
!= RSHIFT_EXPR
2308 && code
!= LROTATE_EXPR
2309 && code
!= RROTATE_EXPR
2310 && !CONVERT_EXPR_CODE_P (code
))
2313 source_stmt1
= find_bswap_or_nop_1 (rhs1_stmt
, n
, limit
- 1);
2315 /* If find_bswap_or_nop_1 returned NULL, STMT is a leaf node and
2316 we have to initialize the symbolic number. */
2319 if (gimple_assign_load_p (stmt
)
2320 || !init_symbolic_number (n
, rhs1
))
2322 source_stmt1
= stmt
;
2329 int i
, size
= TYPE_PRECISION (n
->type
) / BITS_PER_UNIT
;
2330 uint64_t val
= int_cst_value (rhs2
), mask
= 0;
2331 uint64_t tmp
= (1 << BITS_PER_UNIT
) - 1;
2333 /* Only constants masking full bytes are allowed. */
2334 for (i
= 0; i
< size
; i
++, tmp
<<= BITS_PER_UNIT
)
2335 if ((val
& tmp
) != 0 && (val
& tmp
) != tmp
)
2338 mask
|= (uint64_t) MARKER_MASK
<< (i
* BITS_PER_MARKER
);
2347 if (!do_shift_rotate (code
, n
, (int) TREE_INT_CST_LOW (rhs2
)))
2352 int i
, type_size
, old_type_size
;
2355 type
= gimple_expr_type (stmt
);
2356 type_size
= TYPE_PRECISION (type
);
2357 if (type_size
% BITS_PER_UNIT
!= 0)
2359 type_size
/= BITS_PER_UNIT
;
2360 if (type_size
> 64 / BITS_PER_MARKER
)
2363 /* Sign extension: result is dependent on the value. */
2364 old_type_size
= TYPE_PRECISION (n
->type
) / BITS_PER_UNIT
;
2365 if (!TYPE_UNSIGNED (n
->type
) && type_size
> old_type_size
2366 && HEAD_MARKER (n
->n
, old_type_size
))
2367 for (i
= 0; i
< type_size
- old_type_size
; i
++)
2368 n
->n
|= (uint64_t) MARKER_BYTE_UNKNOWN
2369 << ((type_size
- 1 - i
) * BITS_PER_MARKER
);
2371 if (type_size
< 64 / BITS_PER_MARKER
)
2373 /* If STMT casts to a smaller type mask out the bits not
2374 belonging to the target type. */
2375 n
->n
&= ((uint64_t) 1 << (type_size
* BITS_PER_MARKER
)) - 1;
2379 n
->range
= type_size
;
2385 return verify_symbolic_number_p (n
, stmt
) ? source_stmt1
: NULL
;
2388 /* Handle binary rhs. */
2390 if (rhs_class
== GIMPLE_BINARY_RHS
)
2392 struct symbolic_number n1
, n2
;
2393 gimple
*source_stmt
, *source_stmt2
;
2395 if (code
!= BIT_IOR_EXPR
)
2398 if (TREE_CODE (rhs2
) != SSA_NAME
)
2401 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
2406 source_stmt1
= find_bswap_or_nop_1 (rhs1_stmt
, &n1
, limit
- 1);
2411 source_stmt2
= find_bswap_or_nop_1 (rhs2_stmt
, &n2
, limit
- 1);
2416 if (TYPE_PRECISION (n1
.type
) != TYPE_PRECISION (n2
.type
))
2419 if (!n1
.vuse
!= !n2
.vuse
2420 || (n1
.vuse
&& !operand_equal_p (n1
.vuse
, n2
.vuse
, 0)))
2424 = perform_symbolic_merge (source_stmt1
, &n1
, source_stmt2
, &n2
, n
);
2429 if (!verify_symbolic_number_p (n
, stmt
))
2441 /* Check if STMT completes a bswap implementation or a read in a given
2442 endianness consisting of ORs, SHIFTs and ANDs and sets *BSWAP
2443 accordingly. It also sets N to represent the kind of operations
2444 performed: size of the resulting expression and whether it works on
2445 a memory source, and if so alias-set and vuse. At last, the
2446 function returns a stmt whose rhs's first tree is the source
2450 find_bswap_or_nop (gimple
*stmt
, struct symbolic_number
*n
, bool *bswap
)
2453 uint64_t tmpn
, mask
;
2454 /* The number which the find_bswap_or_nop_1 result should match in order
2455 to have a full byte swap. The number is shifted to the right
2456 according to the size of the symbolic number before using it. */
2457 uint64_t cmpxchg
= CMPXCHG
;
2458 uint64_t cmpnop
= CMPNOP
;
2460 gimple
*source_stmt
;
2463 /* The last parameter determines the depth search limit. It usually
2464 correlates directly to the number n of bytes to be touched. We
2465 increase that number by log2(n) + 1 here in order to also
2466 cover signed -> unsigned conversions of the src operand as can be seen
2467 in libgcc, and for initial shift/and operation of the src operand. */
2468 limit
= TREE_INT_CST_LOW (TYPE_SIZE_UNIT (gimple_expr_type (stmt
)));
2469 limit
+= 1 + (int) ceil_log2 ((unsigned HOST_WIDE_INT
) limit
);
2470 source_stmt
= find_bswap_or_nop_1 (stmt
, n
, limit
);
2475 /* Find real size of result (highest non-zero byte). */
2477 for (tmpn
= n
->n
, rsize
= 0; tmpn
; tmpn
>>= BITS_PER_MARKER
, rsize
++);
2481 /* Zero out the bits corresponding to untouched bytes in original gimple
2483 if (n
->range
< (int) sizeof (int64_t))
2485 mask
= ((uint64_t) 1 << (n
->range
* BITS_PER_MARKER
)) - 1;
2486 cmpxchg
>>= (64 / BITS_PER_MARKER
- n
->range
) * BITS_PER_MARKER
;
2490 /* Zero out the bits corresponding to unused bytes in the result of the
2491 gimple expression. */
2492 if (rsize
< n
->range
)
2494 if (BYTES_BIG_ENDIAN
)
2496 mask
= ((uint64_t) 1 << (rsize
* BITS_PER_MARKER
)) - 1;
2498 cmpnop
>>= (n
->range
- rsize
) * BITS_PER_MARKER
;
2502 mask
= ((uint64_t) 1 << (rsize
* BITS_PER_MARKER
)) - 1;
2503 cmpxchg
>>= (n
->range
- rsize
) * BITS_PER_MARKER
;
2509 /* A complete byte swap should make the symbolic number to start with
2510 the largest digit in the highest order byte. Unchanged symbolic
2511 number indicates a read with same endianness as target architecture. */
2514 else if (n
->n
== cmpxchg
)
2519 /* Useless bit manipulation performed by code. */
2520 if (!n
->base_addr
&& n
->n
== cmpnop
)
2523 n
->range
*= BITS_PER_UNIT
;
2529 const pass_data pass_data_optimize_bswap
=
2531 GIMPLE_PASS
, /* type */
2533 OPTGROUP_NONE
, /* optinfo_flags */
2534 TV_NONE
, /* tv_id */
2535 PROP_ssa
, /* properties_required */
2536 0, /* properties_provided */
2537 0, /* properties_destroyed */
2538 0, /* todo_flags_start */
2539 0, /* todo_flags_finish */
2542 class pass_optimize_bswap
: public gimple_opt_pass
2545 pass_optimize_bswap (gcc::context
*ctxt
)
2546 : gimple_opt_pass (pass_data_optimize_bswap
, ctxt
)
2549 /* opt_pass methods: */
2550 virtual bool gate (function
*)
2552 return flag_expensive_optimizations
&& optimize
;
2555 virtual unsigned int execute (function
*);
2557 }; // class pass_optimize_bswap
2559 /* Perform the bswap optimization: replace the expression computed in the rhs
2560 of CUR_STMT by an equivalent bswap, load or load + bswap expression.
2561 Which of these alternatives replace the rhs is given by N->base_addr (non
2562 null if a load is needed) and BSWAP. The type, VUSE and set-alias of the
2563 load to perform are also given in N while the builtin bswap invoke is given
2564 in FNDEL. Finally, if a load is involved, SRC_STMT refers to one of the
2565 load statements involved to construct the rhs in CUR_STMT and N->range gives
2566 the size of the rhs expression for maintaining some statistics.
2568 Note that if the replacement involve a load, CUR_STMT is moved just after
2569 SRC_STMT to do the load with the same VUSE which can lead to CUR_STMT
2570 changing of basic block. */
2573 bswap_replace (gimple
*cur_stmt
, gimple
*src_stmt
, tree fndecl
,
2574 tree bswap_type
, tree load_type
, struct symbolic_number
*n
,
2577 gimple_stmt_iterator gsi
;
2581 gsi
= gsi_for_stmt (cur_stmt
);
2582 src
= gimple_assign_rhs1 (src_stmt
);
2583 tgt
= gimple_assign_lhs (cur_stmt
);
2585 /* Need to load the value from memory first. */
2588 gimple_stmt_iterator gsi_ins
= gsi_for_stmt (src_stmt
);
2589 tree addr_expr
, addr_tmp
, val_expr
, val_tmp
;
2590 tree load_offset_ptr
, aligned_load_type
;
2591 gimple
*addr_stmt
, *load_stmt
;
2593 HOST_WIDE_INT load_offset
= 0;
2595 align
= get_object_alignment (src
);
2596 /* If the new access is smaller than the original one, we need
2597 to perform big endian adjustment. */
2598 if (BYTES_BIG_ENDIAN
)
2600 HOST_WIDE_INT bitsize
, bitpos
;
2602 int unsignedp
, reversep
, volatilep
;
2605 get_inner_reference (src
, &bitsize
, &bitpos
, &offset
, &mode
,
2606 &unsignedp
, &reversep
, &volatilep
, false);
2607 if (n
->range
< (unsigned HOST_WIDE_INT
) bitsize
)
2609 load_offset
= (bitsize
- n
->range
) / BITS_PER_UNIT
;
2610 unsigned HOST_WIDE_INT l
2611 = (load_offset
* BITS_PER_UNIT
) & (align
- 1);
2618 && align
< GET_MODE_ALIGNMENT (TYPE_MODE (load_type
))
2619 && SLOW_UNALIGNED_ACCESS (TYPE_MODE (load_type
), align
))
2622 /* Move cur_stmt just before one of the load of the original
2623 to ensure it has the same VUSE. See PR61517 for what could
2625 if (gimple_bb (cur_stmt
) != gimple_bb (src_stmt
))
2626 reset_flow_sensitive_info (gimple_assign_lhs (cur_stmt
));
2627 gsi_move_before (&gsi
, &gsi_ins
);
2628 gsi
= gsi_for_stmt (cur_stmt
);
2630 /* Compute address to load from and cast according to the size
2632 addr_expr
= build_fold_addr_expr (unshare_expr (src
));
2633 if (is_gimple_mem_ref_addr (addr_expr
))
2634 addr_tmp
= addr_expr
;
2637 addr_tmp
= make_temp_ssa_name (TREE_TYPE (addr_expr
), NULL
,
2639 addr_stmt
= gimple_build_assign (addr_tmp
, addr_expr
);
2640 gsi_insert_before (&gsi
, addr_stmt
, GSI_SAME_STMT
);
2643 /* Perform the load. */
2644 aligned_load_type
= load_type
;
2645 if (align
< TYPE_ALIGN (load_type
))
2646 aligned_load_type
= build_aligned_type (load_type
, align
);
2647 load_offset_ptr
= build_int_cst (n
->alias_set
, load_offset
);
2648 val_expr
= fold_build2 (MEM_REF
, aligned_load_type
, addr_tmp
,
2654 nop_stats
.found_16bit
++;
2655 else if (n
->range
== 32)
2656 nop_stats
.found_32bit
++;
2659 gcc_assert (n
->range
== 64);
2660 nop_stats
.found_64bit
++;
2663 /* Convert the result of load if necessary. */
2664 if (!useless_type_conversion_p (TREE_TYPE (tgt
), load_type
))
2666 val_tmp
= make_temp_ssa_name (aligned_load_type
, NULL
,
2668 load_stmt
= gimple_build_assign (val_tmp
, val_expr
);
2669 gimple_set_vuse (load_stmt
, n
->vuse
);
2670 gsi_insert_before (&gsi
, load_stmt
, GSI_SAME_STMT
);
2671 gimple_assign_set_rhs_with_ops (&gsi
, NOP_EXPR
, val_tmp
);
2675 gimple_assign_set_rhs_with_ops (&gsi
, MEM_REF
, val_expr
);
2676 gimple_set_vuse (cur_stmt
, n
->vuse
);
2678 update_stmt (cur_stmt
);
2683 "%d bit load in target endianness found at: ",
2685 print_gimple_stmt (dump_file
, cur_stmt
, 0, 0);
2691 val_tmp
= make_temp_ssa_name (aligned_load_type
, NULL
, "load_dst");
2692 load_stmt
= gimple_build_assign (val_tmp
, val_expr
);
2693 gimple_set_vuse (load_stmt
, n
->vuse
);
2694 gsi_insert_before (&gsi
, load_stmt
, GSI_SAME_STMT
);
2700 bswap_stats
.found_16bit
++;
2701 else if (n
->range
== 32)
2702 bswap_stats
.found_32bit
++;
2705 gcc_assert (n
->range
== 64);
2706 bswap_stats
.found_64bit
++;
2711 /* Convert the src expression if necessary. */
2712 if (!useless_type_conversion_p (TREE_TYPE (tmp
), bswap_type
))
2714 gimple
*convert_stmt
;
2716 tmp
= make_temp_ssa_name (bswap_type
, NULL
, "bswapsrc");
2717 convert_stmt
= gimple_build_assign (tmp
, NOP_EXPR
, src
);
2718 gsi_insert_before (&gsi
, convert_stmt
, GSI_SAME_STMT
);
2721 /* Canonical form for 16 bit bswap is a rotate expression. Only 16bit values
2722 are considered as rotation of 2N bit values by N bits is generally not
2723 equivalent to a bswap. Consider for instance 0x01020304 r>> 16 which
2724 gives 0x03040102 while a bswap for that value is 0x04030201. */
2725 if (bswap
&& n
->range
== 16)
2727 tree count
= build_int_cst (NULL
, BITS_PER_UNIT
);
2728 src
= fold_build2 (LROTATE_EXPR
, bswap_type
, tmp
, count
);
2729 bswap_stmt
= gimple_build_assign (NULL
, src
);
2732 bswap_stmt
= gimple_build_call (fndecl
, 1, tmp
);
2736 /* Convert the result if necessary. */
2737 if (!useless_type_conversion_p (TREE_TYPE (tgt
), bswap_type
))
2739 gimple
*convert_stmt
;
2741 tmp
= make_temp_ssa_name (bswap_type
, NULL
, "bswapdst");
2742 convert_stmt
= gimple_build_assign (tgt
, NOP_EXPR
, tmp
);
2743 gsi_insert_after (&gsi
, convert_stmt
, GSI_SAME_STMT
);
2746 gimple_set_lhs (bswap_stmt
, tmp
);
2750 fprintf (dump_file
, "%d bit bswap implementation found at: ",
2752 print_gimple_stmt (dump_file
, cur_stmt
, 0, 0);
2755 gsi_insert_after (&gsi
, bswap_stmt
, GSI_SAME_STMT
);
2756 gsi_remove (&gsi
, true);
2760 /* Find manual byte swap implementations as well as load in a given
2761 endianness. Byte swaps are turned into a bswap builtin invokation
2762 while endian loads are converted to bswap builtin invokation or
2763 simple load according to the target endianness. */
2766 pass_optimize_bswap::execute (function
*fun
)
2769 bool bswap32_p
, bswap64_p
;
2770 bool changed
= false;
2771 tree bswap32_type
= NULL_TREE
, bswap64_type
= NULL_TREE
;
2773 if (BITS_PER_UNIT
!= 8)
2776 bswap32_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP32
)
2777 && optab_handler (bswap_optab
, SImode
) != CODE_FOR_nothing
);
2778 bswap64_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP64
)
2779 && (optab_handler (bswap_optab
, DImode
) != CODE_FOR_nothing
2780 || (bswap32_p
&& word_mode
== SImode
)));
2782 /* Determine the argument type of the builtins. The code later on
2783 assumes that the return and argument type are the same. */
2786 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP32
);
2787 bswap32_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
2792 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP64
);
2793 bswap64_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
2796 memset (&nop_stats
, 0, sizeof (nop_stats
));
2797 memset (&bswap_stats
, 0, sizeof (bswap_stats
));
2799 FOR_EACH_BB_FN (bb
, fun
)
2801 gimple_stmt_iterator gsi
;
2803 /* We do a reverse scan for bswap patterns to make sure we get the
2804 widest match. As bswap pattern matching doesn't handle previously
2805 inserted smaller bswap replacements as sub-patterns, the wider
2806 variant wouldn't be detected. */
2807 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
);)
2809 gimple
*src_stmt
, *cur_stmt
= gsi_stmt (gsi
);
2810 tree fndecl
= NULL_TREE
, bswap_type
= NULL_TREE
, load_type
;
2811 enum tree_code code
;
2812 struct symbolic_number n
;
2815 /* This gsi_prev (&gsi) is not part of the for loop because cur_stmt
2816 might be moved to a different basic block by bswap_replace and gsi
2817 must not points to it if that's the case. Moving the gsi_prev
2818 there make sure that gsi points to the statement previous to
2819 cur_stmt while still making sure that all statements are
2820 considered in this basic block. */
2823 if (!is_gimple_assign (cur_stmt
))
2826 code
= gimple_assign_rhs_code (cur_stmt
);
2831 if (!tree_fits_uhwi_p (gimple_assign_rhs2 (cur_stmt
))
2832 || tree_to_uhwi (gimple_assign_rhs2 (cur_stmt
))
2842 src_stmt
= find_bswap_or_nop (cur_stmt
, &n
, &bswap
);
2850 /* Already in canonical form, nothing to do. */
2851 if (code
== LROTATE_EXPR
|| code
== RROTATE_EXPR
)
2853 load_type
= bswap_type
= uint16_type_node
;
2856 load_type
= uint32_type_node
;
2859 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP32
);
2860 bswap_type
= bswap32_type
;
2864 load_type
= uint64_type_node
;
2867 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP64
);
2868 bswap_type
= bswap64_type
;
2875 if (bswap
&& !fndecl
&& n
.range
!= 16)
2878 if (bswap_replace (cur_stmt
, src_stmt
, fndecl
, bswap_type
, load_type
,
2884 statistics_counter_event (fun
, "16-bit nop implementations found",
2885 nop_stats
.found_16bit
);
2886 statistics_counter_event (fun
, "32-bit nop implementations found",
2887 nop_stats
.found_32bit
);
2888 statistics_counter_event (fun
, "64-bit nop implementations found",
2889 nop_stats
.found_64bit
);
2890 statistics_counter_event (fun
, "16-bit bswap implementations found",
2891 bswap_stats
.found_16bit
);
2892 statistics_counter_event (fun
, "32-bit bswap implementations found",
2893 bswap_stats
.found_32bit
);
2894 statistics_counter_event (fun
, "64-bit bswap implementations found",
2895 bswap_stats
.found_64bit
);
2897 return (changed
? TODO_update_ssa
: 0);
2903 make_pass_optimize_bswap (gcc::context
*ctxt
)
2905 return new pass_optimize_bswap (ctxt
);
2908 /* Return true if stmt is a type conversion operation that can be stripped
2909 when used in a widening multiply operation. */
2911 widening_mult_conversion_strippable_p (tree result_type
, gimple
*stmt
)
2913 enum tree_code rhs_code
= gimple_assign_rhs_code (stmt
);
2915 if (TREE_CODE (result_type
) == INTEGER_TYPE
)
2920 if (!CONVERT_EXPR_CODE_P (rhs_code
))
2923 op_type
= TREE_TYPE (gimple_assign_lhs (stmt
));
2925 /* If the type of OP has the same precision as the result, then
2926 we can strip this conversion. The multiply operation will be
2927 selected to create the correct extension as a by-product. */
2928 if (TYPE_PRECISION (result_type
) == TYPE_PRECISION (op_type
))
2931 /* We can also strip a conversion if it preserves the signed-ness of
2932 the operation and doesn't narrow the range. */
2933 inner_op_type
= TREE_TYPE (gimple_assign_rhs1 (stmt
));
2935 /* If the inner-most type is unsigned, then we can strip any
2936 intermediate widening operation. If it's signed, then the
2937 intermediate widening operation must also be signed. */
2938 if ((TYPE_UNSIGNED (inner_op_type
)
2939 || TYPE_UNSIGNED (op_type
) == TYPE_UNSIGNED (inner_op_type
))
2940 && TYPE_PRECISION (op_type
) > TYPE_PRECISION (inner_op_type
))
2946 return rhs_code
== FIXED_CONVERT_EXPR
;
2949 /* Return true if RHS is a suitable operand for a widening multiplication,
2950 assuming a target type of TYPE.
2951 There are two cases:
2953 - RHS makes some value at least twice as wide. Store that value
2954 in *NEW_RHS_OUT if so, and store its type in *TYPE_OUT.
2956 - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so,
2957 but leave *TYPE_OUT untouched. */
2960 is_widening_mult_rhs_p (tree type
, tree rhs
, tree
*type_out
,
2966 if (TREE_CODE (rhs
) == SSA_NAME
)
2968 stmt
= SSA_NAME_DEF_STMT (rhs
);
2969 if (is_gimple_assign (stmt
))
2971 if (! widening_mult_conversion_strippable_p (type
, stmt
))
2975 rhs1
= gimple_assign_rhs1 (stmt
);
2977 if (TREE_CODE (rhs1
) == INTEGER_CST
)
2979 *new_rhs_out
= rhs1
;
2988 type1
= TREE_TYPE (rhs1
);
2990 if (TREE_CODE (type1
) != TREE_CODE (type
)
2991 || TYPE_PRECISION (type1
) * 2 > TYPE_PRECISION (type
))
2994 *new_rhs_out
= rhs1
;
2999 if (TREE_CODE (rhs
) == INTEGER_CST
)
3009 /* Return true if STMT performs a widening multiplication, assuming the
3010 output type is TYPE. If so, store the unwidened types of the operands
3011 in *TYPE1_OUT and *TYPE2_OUT respectively. Also fill *RHS1_OUT and
3012 *RHS2_OUT such that converting those operands to types *TYPE1_OUT
3013 and *TYPE2_OUT would give the operands of the multiplication. */
3016 is_widening_mult_p (gimple
*stmt
,
3017 tree
*type1_out
, tree
*rhs1_out
,
3018 tree
*type2_out
, tree
*rhs2_out
)
3020 tree type
= TREE_TYPE (gimple_assign_lhs (stmt
));
3022 if (TREE_CODE (type
) != INTEGER_TYPE
3023 && TREE_CODE (type
) != FIXED_POINT_TYPE
)
3026 if (!is_widening_mult_rhs_p (type
, gimple_assign_rhs1 (stmt
), type1_out
,
3030 if (!is_widening_mult_rhs_p (type
, gimple_assign_rhs2 (stmt
), type2_out
,
3034 if (*type1_out
== NULL
)
3036 if (*type2_out
== NULL
|| !int_fits_type_p (*rhs1_out
, *type2_out
))
3038 *type1_out
= *type2_out
;
3041 if (*type2_out
== NULL
)
3043 if (!int_fits_type_p (*rhs2_out
, *type1_out
))
3045 *type2_out
= *type1_out
;
3048 /* Ensure that the larger of the two operands comes first. */
3049 if (TYPE_PRECISION (*type1_out
) < TYPE_PRECISION (*type2_out
))
3051 std::swap (*type1_out
, *type2_out
);
3052 std::swap (*rhs1_out
, *rhs2_out
);
3058 /* Process a single gimple statement STMT, which has a MULT_EXPR as
3059 its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return
3060 value is true iff we converted the statement. */
3063 convert_mult_to_widen (gimple
*stmt
, gimple_stmt_iterator
*gsi
)
3065 tree lhs
, rhs1
, rhs2
, type
, type1
, type2
;
3066 enum insn_code handler
;
3067 machine_mode to_mode
, from_mode
, actual_mode
;
3069 int actual_precision
;
3070 location_t loc
= gimple_location (stmt
);
3071 bool from_unsigned1
, from_unsigned2
;
3073 lhs
= gimple_assign_lhs (stmt
);
3074 type
= TREE_TYPE (lhs
);
3075 if (TREE_CODE (type
) != INTEGER_TYPE
)
3078 if (!is_widening_mult_p (stmt
, &type1
, &rhs1
, &type2
, &rhs2
))
3081 to_mode
= TYPE_MODE (type
);
3082 from_mode
= TYPE_MODE (type1
);
3083 from_unsigned1
= TYPE_UNSIGNED (type1
);
3084 from_unsigned2
= TYPE_UNSIGNED (type2
);
3086 if (from_unsigned1
&& from_unsigned2
)
3087 op
= umul_widen_optab
;
3088 else if (!from_unsigned1
&& !from_unsigned2
)
3089 op
= smul_widen_optab
;
3091 op
= usmul_widen_optab
;
3093 handler
= find_widening_optab_handler_and_mode (op
, to_mode
, from_mode
,
3096 if (handler
== CODE_FOR_nothing
)
3098 if (op
!= smul_widen_optab
)
3100 /* We can use a signed multiply with unsigned types as long as
3101 there is a wider mode to use, or it is the smaller of the two
3102 types that is unsigned. Note that type1 >= type2, always. */
3103 if ((TYPE_UNSIGNED (type1
)
3104 && TYPE_PRECISION (type1
) == GET_MODE_PRECISION (from_mode
))
3105 || (TYPE_UNSIGNED (type2
)
3106 && TYPE_PRECISION (type2
) == GET_MODE_PRECISION (from_mode
)))
3108 from_mode
= GET_MODE_WIDER_MODE (from_mode
);
3109 if (GET_MODE_SIZE (to_mode
) <= GET_MODE_SIZE (from_mode
))
3113 op
= smul_widen_optab
;
3114 handler
= find_widening_optab_handler_and_mode (op
, to_mode
,
3118 if (handler
== CODE_FOR_nothing
)
3121 from_unsigned1
= from_unsigned2
= false;
3127 /* Ensure that the inputs to the handler are in the correct precison
3128 for the opcode. This will be the full mode size. */
3129 actual_precision
= GET_MODE_PRECISION (actual_mode
);
3130 if (2 * actual_precision
> TYPE_PRECISION (type
))
3132 if (actual_precision
!= TYPE_PRECISION (type1
)
3133 || from_unsigned1
!= TYPE_UNSIGNED (type1
))
3134 rhs1
= build_and_insert_cast (gsi
, loc
,
3135 build_nonstandard_integer_type
3136 (actual_precision
, from_unsigned1
), rhs1
);
3137 if (actual_precision
!= TYPE_PRECISION (type2
)
3138 || from_unsigned2
!= TYPE_UNSIGNED (type2
))
3139 rhs2
= build_and_insert_cast (gsi
, loc
,
3140 build_nonstandard_integer_type
3141 (actual_precision
, from_unsigned2
), rhs2
);
3143 /* Handle constants. */
3144 if (TREE_CODE (rhs1
) == INTEGER_CST
)
3145 rhs1
= fold_convert (type1
, rhs1
);
3146 if (TREE_CODE (rhs2
) == INTEGER_CST
)
3147 rhs2
= fold_convert (type2
, rhs2
);
3149 gimple_assign_set_rhs1 (stmt
, rhs1
);
3150 gimple_assign_set_rhs2 (stmt
, rhs2
);
3151 gimple_assign_set_rhs_code (stmt
, WIDEN_MULT_EXPR
);
3153 widen_mul_stats
.widen_mults_inserted
++;
3157 /* Process a single gimple statement STMT, which is found at the
3158 iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its
3159 rhs (given by CODE), and try to convert it into a
3160 WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value
3161 is true iff we converted the statement. */
3164 convert_plusminus_to_widen (gimple_stmt_iterator
*gsi
, gimple
*stmt
,
3165 enum tree_code code
)
3167 gimple
*rhs1_stmt
= NULL
, *rhs2_stmt
= NULL
;
3168 gimple
*conv1_stmt
= NULL
, *conv2_stmt
= NULL
, *conv_stmt
;
3169 tree type
, type1
, type2
, optype
;
3170 tree lhs
, rhs1
, rhs2
, mult_rhs1
, mult_rhs2
, add_rhs
;
3171 enum tree_code rhs1_code
= ERROR_MARK
, rhs2_code
= ERROR_MARK
;
3173 enum tree_code wmult_code
;
3174 enum insn_code handler
;
3175 machine_mode to_mode
, from_mode
, actual_mode
;
3176 location_t loc
= gimple_location (stmt
);
3177 int actual_precision
;
3178 bool from_unsigned1
, from_unsigned2
;
3180 lhs
= gimple_assign_lhs (stmt
);
3181 type
= TREE_TYPE (lhs
);
3182 if (TREE_CODE (type
) != INTEGER_TYPE
3183 && TREE_CODE (type
) != FIXED_POINT_TYPE
)
3186 if (code
== MINUS_EXPR
)
3187 wmult_code
= WIDEN_MULT_MINUS_EXPR
;
3189 wmult_code
= WIDEN_MULT_PLUS_EXPR
;
3191 rhs1
= gimple_assign_rhs1 (stmt
);
3192 rhs2
= gimple_assign_rhs2 (stmt
);
3194 if (TREE_CODE (rhs1
) == SSA_NAME
)
3196 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
3197 if (is_gimple_assign (rhs1_stmt
))
3198 rhs1_code
= gimple_assign_rhs_code (rhs1_stmt
);
3201 if (TREE_CODE (rhs2
) == SSA_NAME
)
3203 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
3204 if (is_gimple_assign (rhs2_stmt
))
3205 rhs2_code
= gimple_assign_rhs_code (rhs2_stmt
);
3208 /* Allow for one conversion statement between the multiply
3209 and addition/subtraction statement. If there are more than
3210 one conversions then we assume they would invalidate this
3211 transformation. If that's not the case then they should have
3212 been folded before now. */
3213 if (CONVERT_EXPR_CODE_P (rhs1_code
))
3215 conv1_stmt
= rhs1_stmt
;
3216 rhs1
= gimple_assign_rhs1 (rhs1_stmt
);
3217 if (TREE_CODE (rhs1
) == SSA_NAME
)
3219 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
3220 if (is_gimple_assign (rhs1_stmt
))
3221 rhs1_code
= gimple_assign_rhs_code (rhs1_stmt
);
3226 if (CONVERT_EXPR_CODE_P (rhs2_code
))
3228 conv2_stmt
= rhs2_stmt
;
3229 rhs2
= gimple_assign_rhs1 (rhs2_stmt
);
3230 if (TREE_CODE (rhs2
) == SSA_NAME
)
3232 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
3233 if (is_gimple_assign (rhs2_stmt
))
3234 rhs2_code
= gimple_assign_rhs_code (rhs2_stmt
);
3240 /* If code is WIDEN_MULT_EXPR then it would seem unnecessary to call
3241 is_widening_mult_p, but we still need the rhs returns.
3243 It might also appear that it would be sufficient to use the existing
3244 operands of the widening multiply, but that would limit the choice of
3245 multiply-and-accumulate instructions.
3247 If the widened-multiplication result has more than one uses, it is
3248 probably wiser not to do the conversion. */
3249 if (code
== PLUS_EXPR
3250 && (rhs1_code
== MULT_EXPR
|| rhs1_code
== WIDEN_MULT_EXPR
))
3252 if (!has_single_use (rhs1
)
3253 || !is_widening_mult_p (rhs1_stmt
, &type1
, &mult_rhs1
,
3254 &type2
, &mult_rhs2
))
3257 conv_stmt
= conv1_stmt
;
3259 else if (rhs2_code
== MULT_EXPR
|| rhs2_code
== WIDEN_MULT_EXPR
)
3261 if (!has_single_use (rhs2
)
3262 || !is_widening_mult_p (rhs2_stmt
, &type1
, &mult_rhs1
,
3263 &type2
, &mult_rhs2
))
3266 conv_stmt
= conv2_stmt
;
3271 to_mode
= TYPE_MODE (type
);
3272 from_mode
= TYPE_MODE (type1
);
3273 from_unsigned1
= TYPE_UNSIGNED (type1
);
3274 from_unsigned2
= TYPE_UNSIGNED (type2
);
3277 /* There's no such thing as a mixed sign madd yet, so use a wider mode. */
3278 if (from_unsigned1
!= from_unsigned2
)
3280 if (!INTEGRAL_TYPE_P (type
))
3282 /* We can use a signed multiply with unsigned types as long as
3283 there is a wider mode to use, or it is the smaller of the two
3284 types that is unsigned. Note that type1 >= type2, always. */
3286 && TYPE_PRECISION (type1
) == GET_MODE_PRECISION (from_mode
))
3288 && TYPE_PRECISION (type2
) == GET_MODE_PRECISION (from_mode
)))
3290 from_mode
= GET_MODE_WIDER_MODE (from_mode
);
3291 if (GET_MODE_SIZE (from_mode
) >= GET_MODE_SIZE (to_mode
))
3295 from_unsigned1
= from_unsigned2
= false;
3296 optype
= build_nonstandard_integer_type (GET_MODE_PRECISION (from_mode
),
3300 /* If there was a conversion between the multiply and addition
3301 then we need to make sure it fits a multiply-and-accumulate.
3302 The should be a single mode change which does not change the
3306 /* We use the original, unmodified data types for this. */
3307 tree from_type
= TREE_TYPE (gimple_assign_rhs1 (conv_stmt
));
3308 tree to_type
= TREE_TYPE (gimple_assign_lhs (conv_stmt
));
3309 int data_size
= TYPE_PRECISION (type1
) + TYPE_PRECISION (type2
);
3310 bool is_unsigned
= TYPE_UNSIGNED (type1
) && TYPE_UNSIGNED (type2
);
3312 if (TYPE_PRECISION (from_type
) > TYPE_PRECISION (to_type
))
3314 /* Conversion is a truncate. */
3315 if (TYPE_PRECISION (to_type
) < data_size
)
3318 else if (TYPE_PRECISION (from_type
) < TYPE_PRECISION (to_type
))
3320 /* Conversion is an extend. Check it's the right sort. */
3321 if (TYPE_UNSIGNED (from_type
) != is_unsigned
3322 && !(is_unsigned
&& TYPE_PRECISION (from_type
) > data_size
))
3325 /* else convert is a no-op for our purposes. */
3328 /* Verify that the machine can perform a widening multiply
3329 accumulate in this mode/signedness combination, otherwise
3330 this transformation is likely to pessimize code. */
3331 this_optab
= optab_for_tree_code (wmult_code
, optype
, optab_default
);
3332 handler
= find_widening_optab_handler_and_mode (this_optab
, to_mode
,
3333 from_mode
, 0, &actual_mode
);
3335 if (handler
== CODE_FOR_nothing
)
3338 /* Ensure that the inputs to the handler are in the correct precison
3339 for the opcode. This will be the full mode size. */
3340 actual_precision
= GET_MODE_PRECISION (actual_mode
);
3341 if (actual_precision
!= TYPE_PRECISION (type1
)
3342 || from_unsigned1
!= TYPE_UNSIGNED (type1
))
3343 mult_rhs1
= build_and_insert_cast (gsi
, loc
,
3344 build_nonstandard_integer_type
3345 (actual_precision
, from_unsigned1
),
3347 if (actual_precision
!= TYPE_PRECISION (type2
)
3348 || from_unsigned2
!= TYPE_UNSIGNED (type2
))
3349 mult_rhs2
= build_and_insert_cast (gsi
, loc
,
3350 build_nonstandard_integer_type
3351 (actual_precision
, from_unsigned2
),
3354 if (!useless_type_conversion_p (type
, TREE_TYPE (add_rhs
)))
3355 add_rhs
= build_and_insert_cast (gsi
, loc
, type
, add_rhs
);
3357 /* Handle constants. */
3358 if (TREE_CODE (mult_rhs1
) == INTEGER_CST
)
3359 mult_rhs1
= fold_convert (type1
, mult_rhs1
);
3360 if (TREE_CODE (mult_rhs2
) == INTEGER_CST
)
3361 mult_rhs2
= fold_convert (type2
, mult_rhs2
);
3363 gimple_assign_set_rhs_with_ops (gsi
, wmult_code
, mult_rhs1
, mult_rhs2
,
3365 update_stmt (gsi_stmt (*gsi
));
3366 widen_mul_stats
.maccs_inserted
++;
3370 /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2
3371 with uses in additions and subtractions to form fused multiply-add
3372 operations. Returns true if successful and MUL_STMT should be removed. */
3375 convert_mult_to_fma (gimple
*mul_stmt
, tree op1
, tree op2
)
3377 tree mul_result
= gimple_get_lhs (mul_stmt
);
3378 tree type
= TREE_TYPE (mul_result
);
3379 gimple
*use_stmt
, *neguse_stmt
;
3381 use_operand_p use_p
;
3382 imm_use_iterator imm_iter
;
3384 if (FLOAT_TYPE_P (type
)
3385 && flag_fp_contract_mode
== FP_CONTRACT_OFF
)
3388 /* We don't want to do bitfield reduction ops. */
3389 if (INTEGRAL_TYPE_P (type
)
3390 && (TYPE_PRECISION (type
)
3391 != GET_MODE_PRECISION (TYPE_MODE (type
))))
3394 /* If the target doesn't support it, don't generate it. We assume that
3395 if fma isn't available then fms, fnma or fnms are not either. */
3396 if (optab_handler (fma_optab
, TYPE_MODE (type
)) == CODE_FOR_nothing
)
3399 /* If the multiplication has zero uses, it is kept around probably because
3400 of -fnon-call-exceptions. Don't optimize it away in that case,
3402 if (has_zero_uses (mul_result
))
3405 /* Make sure that the multiplication statement becomes dead after
3406 the transformation, thus that all uses are transformed to FMAs.
3407 This means we assume that an FMA operation has the same cost
3409 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, mul_result
)
3411 enum tree_code use_code
;
3412 tree result
= mul_result
;
3413 bool negate_p
= false;
3415 use_stmt
= USE_STMT (use_p
);
3417 if (is_gimple_debug (use_stmt
))
3420 /* For now restrict this operations to single basic blocks. In theory
3421 we would want to support sinking the multiplication in
3427 to form a fma in the then block and sink the multiplication to the
3429 if (gimple_bb (use_stmt
) != gimple_bb (mul_stmt
))
3432 if (!is_gimple_assign (use_stmt
))
3435 use_code
= gimple_assign_rhs_code (use_stmt
);
3437 /* A negate on the multiplication leads to FNMA. */
3438 if (use_code
== NEGATE_EXPR
)
3443 result
= gimple_assign_lhs (use_stmt
);
3445 /* Make sure the negate statement becomes dead with this
3446 single transformation. */
3447 if (!single_imm_use (gimple_assign_lhs (use_stmt
),
3448 &use_p
, &neguse_stmt
))
3451 /* Make sure the multiplication isn't also used on that stmt. */
3452 FOR_EACH_PHI_OR_STMT_USE (usep
, neguse_stmt
, iter
, SSA_OP_USE
)
3453 if (USE_FROM_PTR (usep
) == mul_result
)
3457 use_stmt
= neguse_stmt
;
3458 if (gimple_bb (use_stmt
) != gimple_bb (mul_stmt
))
3460 if (!is_gimple_assign (use_stmt
))
3463 use_code
= gimple_assign_rhs_code (use_stmt
);
3470 if (gimple_assign_rhs2 (use_stmt
) == result
)
3471 negate_p
= !negate_p
;
3476 /* FMA can only be formed from PLUS and MINUS. */
3480 /* If the subtrahend (gimple_assign_rhs2 (use_stmt)) is computed
3481 by a MULT_EXPR that we'll visit later, we might be able to
3482 get a more profitable match with fnma.
3483 OTOH, if we don't, a negate / fma pair has likely lower latency
3484 that a mult / subtract pair. */
3485 if (use_code
== MINUS_EXPR
&& !negate_p
3486 && gimple_assign_rhs1 (use_stmt
) == result
3487 && optab_handler (fms_optab
, TYPE_MODE (type
)) == CODE_FOR_nothing
3488 && optab_handler (fnma_optab
, TYPE_MODE (type
)) != CODE_FOR_nothing
)
3490 tree rhs2
= gimple_assign_rhs2 (use_stmt
);
3492 if (TREE_CODE (rhs2
) == SSA_NAME
)
3494 gimple
*stmt2
= SSA_NAME_DEF_STMT (rhs2
);
3495 if (has_single_use (rhs2
)
3496 && is_gimple_assign (stmt2
)
3497 && gimple_assign_rhs_code (stmt2
) == MULT_EXPR
)
3502 /* We can't handle a * b + a * b. */
3503 if (gimple_assign_rhs1 (use_stmt
) == gimple_assign_rhs2 (use_stmt
))
3506 /* While it is possible to validate whether or not the exact form
3507 that we've recognized is available in the backend, the assumption
3508 is that the transformation is never a loss. For instance, suppose
3509 the target only has the plain FMA pattern available. Consider
3510 a*b-c -> fma(a,b,-c): we've exchanged MUL+SUB for FMA+NEG, which
3511 is still two operations. Consider -(a*b)-c -> fma(-a,b,-c): we
3512 still have 3 operations, but in the FMA form the two NEGs are
3513 independent and could be run in parallel. */
3516 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, mul_result
)
3518 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
3519 enum tree_code use_code
;
3520 tree addop
, mulop1
= op1
, result
= mul_result
;
3521 bool negate_p
= false;
3523 if (is_gimple_debug (use_stmt
))
3526 use_code
= gimple_assign_rhs_code (use_stmt
);
3527 if (use_code
== NEGATE_EXPR
)
3529 result
= gimple_assign_lhs (use_stmt
);
3530 single_imm_use (gimple_assign_lhs (use_stmt
), &use_p
, &neguse_stmt
);
3531 gsi_remove (&gsi
, true);
3532 release_defs (use_stmt
);
3534 use_stmt
= neguse_stmt
;
3535 gsi
= gsi_for_stmt (use_stmt
);
3536 use_code
= gimple_assign_rhs_code (use_stmt
);
3540 if (gimple_assign_rhs1 (use_stmt
) == result
)
3542 addop
= gimple_assign_rhs2 (use_stmt
);
3543 /* a * b - c -> a * b + (-c) */
3544 if (gimple_assign_rhs_code (use_stmt
) == MINUS_EXPR
)
3545 addop
= force_gimple_operand_gsi (&gsi
,
3546 build1 (NEGATE_EXPR
,
3548 true, NULL_TREE
, true,
3553 addop
= gimple_assign_rhs1 (use_stmt
);
3554 /* a - b * c -> (-b) * c + a */
3555 if (gimple_assign_rhs_code (use_stmt
) == MINUS_EXPR
)
3556 negate_p
= !negate_p
;
3560 mulop1
= force_gimple_operand_gsi (&gsi
,
3561 build1 (NEGATE_EXPR
,
3563 true, NULL_TREE
, true,
3566 fma_stmt
= gimple_build_assign (gimple_assign_lhs (use_stmt
),
3567 FMA_EXPR
, mulop1
, op2
, addop
);
3568 gsi_replace (&gsi
, fma_stmt
, true);
3569 widen_mul_stats
.fmas_inserted
++;
3576 /* Helper function of match_uaddsub_overflow. Return 1
3577 if USE_STMT is unsigned overflow check ovf != 0 for
3578 STMT, -1 if USE_STMT is unsigned overflow check ovf == 0
3582 uaddsub_overflow_check_p (gimple
*stmt
, gimple
*use_stmt
)
3584 enum tree_code ccode
= ERROR_MARK
;
3585 tree crhs1
= NULL_TREE
, crhs2
= NULL_TREE
;
3586 if (gimple_code (use_stmt
) == GIMPLE_COND
)
3588 ccode
= gimple_cond_code (use_stmt
);
3589 crhs1
= gimple_cond_lhs (use_stmt
);
3590 crhs2
= gimple_cond_rhs (use_stmt
);
3592 else if (is_gimple_assign (use_stmt
))
3594 if (gimple_assign_rhs_class (use_stmt
) == GIMPLE_BINARY_RHS
)
3596 ccode
= gimple_assign_rhs_code (use_stmt
);
3597 crhs1
= gimple_assign_rhs1 (use_stmt
);
3598 crhs2
= gimple_assign_rhs2 (use_stmt
);
3600 else if (gimple_assign_rhs_code (use_stmt
) == COND_EXPR
)
3602 tree cond
= gimple_assign_rhs1 (use_stmt
);
3603 if (COMPARISON_CLASS_P (cond
))
3605 ccode
= TREE_CODE (cond
);
3606 crhs1
= TREE_OPERAND (cond
, 0);
3607 crhs2
= TREE_OPERAND (cond
, 1);
3618 if (TREE_CODE_CLASS (ccode
) != tcc_comparison
)
3621 enum tree_code code
= gimple_assign_rhs_code (stmt
);
3622 tree lhs
= gimple_assign_lhs (stmt
);
3623 tree rhs1
= gimple_assign_rhs1 (stmt
);
3624 tree rhs2
= gimple_assign_rhs2 (stmt
);
3630 /* r = a - b; r > a or r <= a
3631 r = a + b; a > r or a <= r or b > r or b <= r. */
3632 if ((code
== MINUS_EXPR
&& crhs1
== lhs
&& crhs2
== rhs1
)
3633 || (code
== PLUS_EXPR
&& (crhs1
== rhs1
|| crhs1
== rhs2
)
3635 return ccode
== GT_EXPR
? 1 : -1;
3639 /* r = a - b; a < r or a >= r
3640 r = a + b; r < a or r >= a or r < b or r >= b. */
3641 if ((code
== MINUS_EXPR
&& crhs1
== rhs1
&& crhs2
== lhs
)
3642 || (code
== PLUS_EXPR
&& crhs1
== lhs
3643 && (crhs2
== rhs1
|| crhs2
== rhs2
)))
3644 return ccode
== LT_EXPR
? 1 : -1;
3652 /* Recognize for unsigned x
3655 where there are other uses of x and replace it with
3656 _7 = SUB_OVERFLOW (y, z);
3657 x = REALPART_EXPR <_7>;
3658 _8 = IMAGPART_EXPR <_7>;
3660 and similarly for addition. */
3663 match_uaddsub_overflow (gimple_stmt_iterator
*gsi
, gimple
*stmt
,
3664 enum tree_code code
)
3666 tree lhs
= gimple_assign_lhs (stmt
);
3667 tree type
= TREE_TYPE (lhs
);
3668 use_operand_p use_p
;
3669 imm_use_iterator iter
;
3670 bool use_seen
= false;
3671 bool ovf_use_seen
= false;
3674 gcc_checking_assert (code
== PLUS_EXPR
|| code
== MINUS_EXPR
);
3675 if (!INTEGRAL_TYPE_P (type
)
3676 || !TYPE_UNSIGNED (type
)
3677 || has_zero_uses (lhs
)
3678 || has_single_use (lhs
)
3679 || optab_handler (code
== PLUS_EXPR
? uaddv4_optab
: usubv4_optab
,
3680 TYPE_MODE (type
)) == CODE_FOR_nothing
)
3683 FOR_EACH_IMM_USE_FAST (use_p
, iter
, lhs
)
3685 use_stmt
= USE_STMT (use_p
);
3686 if (is_gimple_debug (use_stmt
))
3689 if (uaddsub_overflow_check_p (stmt
, use_stmt
))
3690 ovf_use_seen
= true;
3693 if (ovf_use_seen
&& use_seen
)
3697 if (!ovf_use_seen
|| !use_seen
)
3700 tree ctype
= build_complex_type (type
);
3701 tree rhs1
= gimple_assign_rhs1 (stmt
);
3702 tree rhs2
= gimple_assign_rhs2 (stmt
);
3703 gcall
*g
= gimple_build_call_internal (code
== PLUS_EXPR
3704 ? IFN_ADD_OVERFLOW
: IFN_SUB_OVERFLOW
,
3706 tree ctmp
= make_ssa_name (ctype
);
3707 gimple_call_set_lhs (g
, ctmp
);
3708 gsi_insert_before (gsi
, g
, GSI_SAME_STMT
);
3709 gassign
*g2
= gimple_build_assign (lhs
, REALPART_EXPR
,
3710 build1 (REALPART_EXPR
, type
, ctmp
));
3711 gsi_replace (gsi
, g2
, true);
3712 tree ovf
= make_ssa_name (type
);
3713 g2
= gimple_build_assign (ovf
, IMAGPART_EXPR
,
3714 build1 (IMAGPART_EXPR
, type
, ctmp
));
3715 gsi_insert_after (gsi
, g2
, GSI_NEW_STMT
);
3717 FOR_EACH_IMM_USE_STMT (use_stmt
, iter
, lhs
)
3719 if (is_gimple_debug (use_stmt
))
3722 int ovf_use
= uaddsub_overflow_check_p (stmt
, use_stmt
);
3725 if (gimple_code (use_stmt
) == GIMPLE_COND
)
3727 gcond
*cond_stmt
= as_a
<gcond
*> (use_stmt
);
3728 gimple_cond_set_lhs (cond_stmt
, ovf
);
3729 gimple_cond_set_rhs (cond_stmt
, build_int_cst (type
, 0));
3730 gimple_cond_set_code (cond_stmt
, ovf_use
== 1 ? NE_EXPR
: EQ_EXPR
);
3734 gcc_checking_assert (is_gimple_assign (use_stmt
));
3735 if (gimple_assign_rhs_class (use_stmt
) == GIMPLE_BINARY_RHS
)
3737 gimple_assign_set_rhs1 (use_stmt
, ovf
);
3738 gimple_assign_set_rhs2 (use_stmt
, build_int_cst (type
, 0));
3739 gimple_assign_set_rhs_code (use_stmt
,
3740 ovf_use
== 1 ? NE_EXPR
: EQ_EXPR
);
3744 gcc_checking_assert (gimple_assign_rhs_code (use_stmt
)
3746 tree cond
= build2 (ovf_use
== 1 ? NE_EXPR
: EQ_EXPR
,
3747 boolean_type_node
, ovf
,
3748 build_int_cst (type
, 0));
3749 gimple_assign_set_rhs1 (use_stmt
, cond
);
3752 update_stmt (use_stmt
);
3758 /* Find integer multiplications where the operands are extended from
3759 smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR
3760 where appropriate. */
3764 const pass_data pass_data_optimize_widening_mul
=
3766 GIMPLE_PASS
, /* type */
3767 "widening_mul", /* name */
3768 OPTGROUP_NONE
, /* optinfo_flags */
3769 TV_NONE
, /* tv_id */
3770 PROP_ssa
, /* properties_required */
3771 0, /* properties_provided */
3772 0, /* properties_destroyed */
3773 0, /* todo_flags_start */
3774 TODO_update_ssa
, /* todo_flags_finish */
3777 class pass_optimize_widening_mul
: public gimple_opt_pass
3780 pass_optimize_widening_mul (gcc::context
*ctxt
)
3781 : gimple_opt_pass (pass_data_optimize_widening_mul
, ctxt
)
3784 /* opt_pass methods: */
3785 virtual bool gate (function
*)
3787 return flag_expensive_optimizations
&& optimize
;
3790 virtual unsigned int execute (function
*);
3792 }; // class pass_optimize_widening_mul
3795 pass_optimize_widening_mul::execute (function
*fun
)
3798 bool cfg_changed
= false;
3800 memset (&widen_mul_stats
, 0, sizeof (widen_mul_stats
));
3802 FOR_EACH_BB_FN (bb
, fun
)
3804 gimple_stmt_iterator gsi
;
3806 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
);)
3808 gimple
*stmt
= gsi_stmt (gsi
);
3809 enum tree_code code
;
3811 if (is_gimple_assign (stmt
))
3813 code
= gimple_assign_rhs_code (stmt
);
3817 if (!convert_mult_to_widen (stmt
, &gsi
)
3818 && convert_mult_to_fma (stmt
,
3819 gimple_assign_rhs1 (stmt
),
3820 gimple_assign_rhs2 (stmt
)))
3822 gsi_remove (&gsi
, true);
3823 release_defs (stmt
);
3830 if (!convert_plusminus_to_widen (&gsi
, stmt
, code
))
3831 match_uaddsub_overflow (&gsi
, stmt
, code
);
3837 else if (is_gimple_call (stmt
)
3838 && gimple_call_lhs (stmt
))
3840 tree fndecl
= gimple_call_fndecl (stmt
);
3842 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
)
3844 switch (DECL_FUNCTION_CODE (fndecl
))
3849 if (TREE_CODE (gimple_call_arg (stmt
, 1)) == REAL_CST
3851 (&TREE_REAL_CST (gimple_call_arg (stmt
, 1)),
3853 && convert_mult_to_fma (stmt
,
3854 gimple_call_arg (stmt
, 0),
3855 gimple_call_arg (stmt
, 0)))
3857 unlink_stmt_vdef (stmt
);
3858 if (gsi_remove (&gsi
, true)
3859 && gimple_purge_dead_eh_edges (bb
))
3861 release_defs (stmt
);
3874 statistics_counter_event (fun
, "widening multiplications inserted",
3875 widen_mul_stats
.widen_mults_inserted
);
3876 statistics_counter_event (fun
, "widening maccs inserted",
3877 widen_mul_stats
.maccs_inserted
);
3878 statistics_counter_event (fun
, "fused multiply-adds inserted",
3879 widen_mul_stats
.fmas_inserted
);
3881 return cfg_changed
? TODO_cleanup_cfg
: 0;
3887 make_pass_optimize_widening_mul (gcc::context
*ctxt
)
3889 return new pass_optimize_widening_mul (ctxt
);