1 /* Global, SSA-based optimizations using mathematical identities.
2 Copyright (C) 2005, 2006, 2007, 2008, 2009, 2010, 2011
3 Free Software Foundation, Inc.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it
8 under the terms of the GNU General Public License as published by the
9 Free Software Foundation; either version 3, or (at your option) any
12 GCC is distributed in the hope that it will be useful, but WITHOUT
13 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
21 /* Currently, the only mini-pass in this file tries to CSE reciprocal
22 operations. These are common in sequences such as this one:
24 modulus = sqrt(x*x + y*y + z*z);
29 that can be optimized to
31 modulus = sqrt(x*x + y*y + z*z);
32 rmodulus = 1.0 / modulus;
37 We do this for loop invariant divisors, and with this pass whenever
38 we notice that a division has the same divisor multiple times.
40 Of course, like in PRE, we don't insert a division if a dominator
41 already has one. However, this cannot be done as an extension of
42 PRE for several reasons.
44 First of all, with some experiments it was found out that the
45 transformation is not always useful if there are only two divisions
46 hy the same divisor. This is probably because modern processors
47 can pipeline the divisions; on older, in-order processors it should
48 still be effective to optimize two divisions by the same number.
49 We make this a param, and it shall be called N in the remainder of
52 Second, if trapping math is active, we have less freedom on where
53 to insert divisions: we can only do so in basic blocks that already
54 contain one. (If divisions don't trap, instead, we can insert
55 divisions elsewhere, which will be in blocks that are common dominators
56 of those that have the division).
58 We really don't want to compute the reciprocal unless a division will
59 be found. To do this, we won't insert the division in a basic block
60 that has less than N divisions *post-dominating* it.
62 The algorithm constructs a subset of the dominator tree, holding the
63 blocks containing the divisions and the common dominators to them,
64 and walk it twice. The first walk is in post-order, and it annotates
65 each block with the number of divisions that post-dominate it: this
66 gives information on where divisions can be inserted profitably.
67 The second walk is in pre-order, and it inserts divisions as explained
68 above, and replaces divisions by multiplications.
70 In the best case, the cost of the pass is O(n_statements). In the
71 worst-case, the cost is due to creating the dominator tree subset,
72 with a cost of O(n_basic_blocks ^ 2); however this can only happen
73 for n_statements / n_basic_blocks statements. So, the amortized cost
74 of creating the dominator tree subset is O(n_basic_blocks) and the
75 worst-case cost of the pass is O(n_statements * n_basic_blocks).
77 More practically, the cost will be small because there are few
78 divisions, and they tend to be in the same basic block, so insert_bb
79 is called very few times.
81 If we did this using domwalk.c, an efficient implementation would have
82 to work on all the variables in a single pass, because we could not
83 work on just a subset of the dominator tree, as we do now, and the
84 cost would also be something like O(n_statements * n_basic_blocks).
85 The data structures would be more complex in order to work on all the
86 variables in a single pass. */
90 #include "coretypes.h"
94 #include "tree-flow.h"
95 #include "tree-pass.h"
96 #include "alloc-pool.h"
97 #include "basic-block.h"
99 #include "gimple-pretty-print.h"
101 /* FIXME: RTL headers have to be included here for optabs. */
102 #include "rtl.h" /* Because optabs.h wants enum rtx_code. */
103 #include "expr.h" /* Because optabs.h wants sepops. */
106 /* This structure represents one basic block that either computes a
107 division, or is a common dominator for basic block that compute a
110 /* The basic block represented by this structure. */
113 /* If non-NULL, the SSA_NAME holding the definition for a reciprocal
117 /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that
118 was inserted in BB. */
119 gimple recip_def_stmt
;
121 /* Pointer to a list of "struct occurrence"s for blocks dominated
123 struct occurrence
*children
;
125 /* Pointer to the next "struct occurrence"s in the list of blocks
126 sharing a common dominator. */
127 struct occurrence
*next
;
129 /* The number of divisions that are in BB before compute_merit. The
130 number of divisions that are in BB or post-dominate it after
134 /* True if the basic block has a division, false if it is a common
135 dominator for basic blocks that do. If it is false and trapping
136 math is active, BB is not a candidate for inserting a reciprocal. */
137 bool bb_has_division
;
142 /* Number of 1.0/X ops inserted. */
145 /* Number of 1.0/FUNC ops inserted. */
151 /* Number of cexpi calls inserted. */
157 /* Number of hand-written 32-bit bswaps found. */
160 /* Number of hand-written 64-bit bswaps found. */
166 /* Number of widening multiplication ops inserted. */
167 int widen_mults_inserted
;
169 /* Number of integer multiply-and-accumulate ops inserted. */
172 /* Number of fp fused multiply-add ops inserted. */
176 /* The instance of "struct occurrence" representing the highest
177 interesting block in the dominator tree. */
178 static struct occurrence
*occ_head
;
180 /* Allocation pool for getting instances of "struct occurrence". */
181 static alloc_pool occ_pool
;
185 /* Allocate and return a new struct occurrence for basic block BB, and
186 whose children list is headed by CHILDREN. */
187 static struct occurrence
*
188 occ_new (basic_block bb
, struct occurrence
*children
)
190 struct occurrence
*occ
;
192 bb
->aux
= occ
= (struct occurrence
*) pool_alloc (occ_pool
);
193 memset (occ
, 0, sizeof (struct occurrence
));
196 occ
->children
= children
;
201 /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a
202 list of "struct occurrence"s, one per basic block, having IDOM as
203 their common dominator.
205 We try to insert NEW_OCC as deep as possible in the tree, and we also
206 insert any other block that is a common dominator for BB and one
207 block already in the tree. */
210 insert_bb (struct occurrence
*new_occ
, basic_block idom
,
211 struct occurrence
**p_head
)
213 struct occurrence
*occ
, **p_occ
;
215 for (p_occ
= p_head
; (occ
= *p_occ
) != NULL
; )
217 basic_block bb
= new_occ
->bb
, occ_bb
= occ
->bb
;
218 basic_block dom
= nearest_common_dominator (CDI_DOMINATORS
, occ_bb
, bb
);
221 /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC
224 occ
->next
= new_occ
->children
;
225 new_occ
->children
= occ
;
227 /* Try the next block (it may as well be dominated by BB). */
230 else if (dom
== occ_bb
)
232 /* OCC_BB dominates BB. Tail recurse to look deeper. */
233 insert_bb (new_occ
, dom
, &occ
->children
);
237 else if (dom
!= idom
)
239 gcc_assert (!dom
->aux
);
241 /* There is a dominator between IDOM and BB, add it and make
242 two children out of NEW_OCC and OCC. First, remove OCC from
248 /* None of the previous blocks has DOM as a dominator: if we tail
249 recursed, we would reexamine them uselessly. Just switch BB with
250 DOM, and go on looking for blocks dominated by DOM. */
251 new_occ
= occ_new (dom
, new_occ
);
256 /* Nothing special, go on with the next element. */
261 /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */
262 new_occ
->next
= *p_head
;
266 /* Register that we found a division in BB. */
269 register_division_in (basic_block bb
)
271 struct occurrence
*occ
;
273 occ
= (struct occurrence
*) bb
->aux
;
276 occ
= occ_new (bb
, NULL
);
277 insert_bb (occ
, ENTRY_BLOCK_PTR
, &occ_head
);
280 occ
->bb_has_division
= true;
281 occ
->num_divisions
++;
285 /* Compute the number of divisions that postdominate each block in OCC and
289 compute_merit (struct occurrence
*occ
)
291 struct occurrence
*occ_child
;
292 basic_block dom
= occ
->bb
;
294 for (occ_child
= occ
->children
; occ_child
; occ_child
= occ_child
->next
)
297 if (occ_child
->children
)
298 compute_merit (occ_child
);
301 bb
= single_noncomplex_succ (dom
);
305 if (dominated_by_p (CDI_POST_DOMINATORS
, bb
, occ_child
->bb
))
306 occ
->num_divisions
+= occ_child
->num_divisions
;
311 /* Return whether USE_STMT is a floating-point division by DEF. */
313 is_division_by (gimple use_stmt
, tree def
)
315 return is_gimple_assign (use_stmt
)
316 && gimple_assign_rhs_code (use_stmt
) == RDIV_EXPR
317 && gimple_assign_rhs2 (use_stmt
) == def
318 /* Do not recognize x / x as valid division, as we are getting
319 confused later by replacing all immediate uses x in such
321 && gimple_assign_rhs1 (use_stmt
) != def
;
324 /* Walk the subset of the dominator tree rooted at OCC, setting the
325 RECIP_DEF field to a definition of 1.0 / DEF that can be used in
326 the given basic block. The field may be left NULL, of course,
327 if it is not possible or profitable to do the optimization.
329 DEF_BSI is an iterator pointing at the statement defining DEF.
330 If RECIP_DEF is set, a dominator already has a computation that can
334 insert_reciprocals (gimple_stmt_iterator
*def_gsi
, struct occurrence
*occ
,
335 tree def
, tree recip_def
, int threshold
)
339 gimple_stmt_iterator gsi
;
340 struct occurrence
*occ_child
;
343 && (occ
->bb_has_division
|| !flag_trapping_math
)
344 && occ
->num_divisions
>= threshold
)
346 /* Make a variable with the replacement and substitute it. */
347 type
= TREE_TYPE (def
);
348 recip_def
= create_tmp_reg (type
, "reciptmp");
349 new_stmt
= gimple_build_assign_with_ops (RDIV_EXPR
, recip_def
,
350 build_one_cst (type
), def
);
352 if (occ
->bb_has_division
)
354 /* Case 1: insert before an existing division. */
355 gsi
= gsi_after_labels (occ
->bb
);
356 while (!gsi_end_p (gsi
) && !is_division_by (gsi_stmt (gsi
), def
))
359 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
361 else if (def_gsi
&& occ
->bb
== def_gsi
->bb
)
363 /* Case 2: insert right after the definition. Note that this will
364 never happen if the definition statement can throw, because in
365 that case the sole successor of the statement's basic block will
366 dominate all the uses as well. */
367 gsi_insert_after (def_gsi
, new_stmt
, GSI_NEW_STMT
);
371 /* Case 3: insert in a basic block not containing defs/uses. */
372 gsi
= gsi_after_labels (occ
->bb
);
373 gsi_insert_before (&gsi
, new_stmt
, GSI_SAME_STMT
);
376 reciprocal_stats
.rdivs_inserted
++;
378 occ
->recip_def_stmt
= new_stmt
;
381 occ
->recip_def
= recip_def
;
382 for (occ_child
= occ
->children
; occ_child
; occ_child
= occ_child
->next
)
383 insert_reciprocals (def_gsi
, occ_child
, def
, recip_def
, threshold
);
387 /* Replace the division at USE_P with a multiplication by the reciprocal, if
391 replace_reciprocal (use_operand_p use_p
)
393 gimple use_stmt
= USE_STMT (use_p
);
394 basic_block bb
= gimple_bb (use_stmt
);
395 struct occurrence
*occ
= (struct occurrence
*) bb
->aux
;
397 if (optimize_bb_for_speed_p (bb
)
398 && occ
->recip_def
&& use_stmt
!= occ
->recip_def_stmt
)
400 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
401 gimple_assign_set_rhs_code (use_stmt
, MULT_EXPR
);
402 SET_USE (use_p
, occ
->recip_def
);
403 fold_stmt_inplace (&gsi
);
404 update_stmt (use_stmt
);
409 /* Free OCC and return one more "struct occurrence" to be freed. */
411 static struct occurrence
*
412 free_bb (struct occurrence
*occ
)
414 struct occurrence
*child
, *next
;
416 /* First get the two pointers hanging off OCC. */
418 child
= occ
->children
;
420 pool_free (occ_pool
, occ
);
422 /* Now ensure that we don't recurse unless it is necessary. */
428 next
= free_bb (next
);
435 /* Look for floating-point divisions among DEF's uses, and try to
436 replace them by multiplications with the reciprocal. Add
437 as many statements computing the reciprocal as needed.
439 DEF must be a GIMPLE register of a floating-point type. */
442 execute_cse_reciprocals_1 (gimple_stmt_iterator
*def_gsi
, tree def
)
445 imm_use_iterator use_iter
;
446 struct occurrence
*occ
;
447 int count
= 0, threshold
;
449 gcc_assert (FLOAT_TYPE_P (TREE_TYPE (def
)) && is_gimple_reg (def
));
451 FOR_EACH_IMM_USE_FAST (use_p
, use_iter
, def
)
453 gimple use_stmt
= USE_STMT (use_p
);
454 if (is_division_by (use_stmt
, def
))
456 register_division_in (gimple_bb (use_stmt
));
461 /* Do the expensive part only if we can hope to optimize something. */
462 threshold
= targetm
.min_divisions_for_recip_mul (TYPE_MODE (TREE_TYPE (def
)));
463 if (count
>= threshold
)
466 for (occ
= occ_head
; occ
; occ
= occ
->next
)
469 insert_reciprocals (def_gsi
, occ
, def
, NULL
, threshold
);
472 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, def
)
474 if (is_division_by (use_stmt
, def
))
476 FOR_EACH_IMM_USE_ON_STMT (use_p
, use_iter
)
477 replace_reciprocal (use_p
);
482 for (occ
= occ_head
; occ
; )
489 gate_cse_reciprocals (void)
491 return optimize
&& flag_reciprocal_math
;
494 /* Go through all the floating-point SSA_NAMEs, and call
495 execute_cse_reciprocals_1 on each of them. */
497 execute_cse_reciprocals (void)
502 occ_pool
= create_alloc_pool ("dominators for recip",
503 sizeof (struct occurrence
),
504 n_basic_blocks
/ 3 + 1);
506 memset (&reciprocal_stats
, 0, sizeof (reciprocal_stats
));
507 calculate_dominance_info (CDI_DOMINATORS
);
508 calculate_dominance_info (CDI_POST_DOMINATORS
);
510 #ifdef ENABLE_CHECKING
512 gcc_assert (!bb
->aux
);
515 for (arg
= DECL_ARGUMENTS (cfun
->decl
); arg
; arg
= DECL_CHAIN (arg
))
516 if (FLOAT_TYPE_P (TREE_TYPE (arg
))
517 && is_gimple_reg (arg
))
519 tree name
= ssa_default_def (cfun
, arg
);
521 execute_cse_reciprocals_1 (NULL
, name
);
526 gimple_stmt_iterator gsi
;
530 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
532 phi
= gsi_stmt (gsi
);
533 def
= PHI_RESULT (phi
);
534 if (! virtual_operand_p (def
)
535 && FLOAT_TYPE_P (TREE_TYPE (def
)))
536 execute_cse_reciprocals_1 (NULL
, def
);
539 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
541 gimple stmt
= gsi_stmt (gsi
);
543 if (gimple_has_lhs (stmt
)
544 && (def
= SINGLE_SSA_TREE_OPERAND (stmt
, SSA_OP_DEF
)) != NULL
545 && FLOAT_TYPE_P (TREE_TYPE (def
))
546 && TREE_CODE (def
) == SSA_NAME
)
547 execute_cse_reciprocals_1 (&gsi
, def
);
550 if (optimize_bb_for_size_p (bb
))
553 /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */
554 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
556 gimple stmt
= gsi_stmt (gsi
);
559 if (is_gimple_assign (stmt
)
560 && gimple_assign_rhs_code (stmt
) == RDIV_EXPR
)
562 tree arg1
= gimple_assign_rhs2 (stmt
);
565 if (TREE_CODE (arg1
) != SSA_NAME
)
568 stmt1
= SSA_NAME_DEF_STMT (arg1
);
570 if (is_gimple_call (stmt1
)
571 && gimple_call_lhs (stmt1
)
572 && (fndecl
= gimple_call_fndecl (stmt1
))
573 && (DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
574 || DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_MD
))
576 enum built_in_function code
;
581 code
= DECL_FUNCTION_CODE (fndecl
);
582 md_code
= DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_MD
;
584 fndecl
= targetm
.builtin_reciprocal (code
, md_code
, false);
588 /* Check that all uses of the SSA name are divisions,
589 otherwise replacing the defining statement will do
592 FOR_EACH_IMM_USE_FAST (use_p
, ui
, arg1
)
594 gimple stmt2
= USE_STMT (use_p
);
595 if (is_gimple_debug (stmt2
))
597 if (!is_gimple_assign (stmt2
)
598 || gimple_assign_rhs_code (stmt2
) != RDIV_EXPR
599 || gimple_assign_rhs1 (stmt2
) == arg1
600 || gimple_assign_rhs2 (stmt2
) != arg1
)
609 gimple_replace_lhs (stmt1
, arg1
);
610 gimple_call_set_fndecl (stmt1
, fndecl
);
612 reciprocal_stats
.rfuncs_inserted
++;
614 FOR_EACH_IMM_USE_STMT (stmt
, ui
, arg1
)
616 gimple_stmt_iterator gsi
= gsi_for_stmt (stmt
);
617 gimple_assign_set_rhs_code (stmt
, MULT_EXPR
);
618 fold_stmt_inplace (&gsi
);
626 statistics_counter_event (cfun
, "reciprocal divs inserted",
627 reciprocal_stats
.rdivs_inserted
);
628 statistics_counter_event (cfun
, "reciprocal functions inserted",
629 reciprocal_stats
.rfuncs_inserted
);
631 free_dominance_info (CDI_DOMINATORS
);
632 free_dominance_info (CDI_POST_DOMINATORS
);
633 free_alloc_pool (occ_pool
);
637 struct gimple_opt_pass pass_cse_reciprocals
=
642 gate_cse_reciprocals
, /* gate */
643 execute_cse_reciprocals
, /* execute */
646 0, /* static_pass_number */
648 PROP_ssa
, /* properties_required */
649 0, /* properties_provided */
650 0, /* properties_destroyed */
651 0, /* todo_flags_start */
652 TODO_update_ssa
| TODO_verify_ssa
653 | TODO_verify_stmts
/* todo_flags_finish */
657 /* Records an occurrence at statement USE_STMT in the vector of trees
658 STMTS if it is dominated by *TOP_BB or dominates it or this basic block
659 is not yet initialized. Returns true if the occurrence was pushed on
660 the vector. Adjusts *TOP_BB to be the basic block dominating all
661 statements in the vector. */
664 maybe_record_sincos (VEC(gimple
, heap
) **stmts
,
665 basic_block
*top_bb
, gimple use_stmt
)
667 basic_block use_bb
= gimple_bb (use_stmt
);
669 && (*top_bb
== use_bb
670 || dominated_by_p (CDI_DOMINATORS
, use_bb
, *top_bb
)))
671 VEC_safe_push (gimple
, heap
, *stmts
, use_stmt
);
673 || dominated_by_p (CDI_DOMINATORS
, *top_bb
, use_bb
))
675 VEC_safe_push (gimple
, heap
, *stmts
, use_stmt
);
684 /* Look for sin, cos and cexpi calls with the same argument NAME and
685 create a single call to cexpi CSEing the result in this case.
686 We first walk over all immediate uses of the argument collecting
687 statements that we can CSE in a vector and in a second pass replace
688 the statement rhs with a REALPART or IMAGPART expression on the
689 result of the cexpi call we insert before the use statement that
690 dominates all other candidates. */
693 execute_cse_sincos_1 (tree name
)
695 gimple_stmt_iterator gsi
;
696 imm_use_iterator use_iter
;
697 tree fndecl
, res
, type
;
698 gimple def_stmt
, use_stmt
, stmt
;
699 int seen_cos
= 0, seen_sin
= 0, seen_cexpi
= 0;
700 VEC(gimple
, heap
) *stmts
= NULL
;
701 basic_block top_bb
= NULL
;
703 bool cfg_changed
= false;
705 type
= TREE_TYPE (name
);
706 FOR_EACH_IMM_USE_STMT (use_stmt
, use_iter
, name
)
708 if (gimple_code (use_stmt
) != GIMPLE_CALL
709 || !gimple_call_lhs (use_stmt
)
710 || !(fndecl
= gimple_call_fndecl (use_stmt
))
711 || DECL_BUILT_IN_CLASS (fndecl
) != BUILT_IN_NORMAL
)
714 switch (DECL_FUNCTION_CODE (fndecl
))
716 CASE_FLT_FN (BUILT_IN_COS
):
717 seen_cos
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
720 CASE_FLT_FN (BUILT_IN_SIN
):
721 seen_sin
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
724 CASE_FLT_FN (BUILT_IN_CEXPI
):
725 seen_cexpi
|= maybe_record_sincos (&stmts
, &top_bb
, use_stmt
) ? 1 : 0;
732 if (seen_cos
+ seen_sin
+ seen_cexpi
<= 1)
734 VEC_free(gimple
, heap
, stmts
);
738 /* Simply insert cexpi at the beginning of top_bb but not earlier than
739 the name def statement. */
740 fndecl
= mathfn_built_in (type
, BUILT_IN_CEXPI
);
743 stmt
= gimple_build_call (fndecl
, 1, name
);
744 res
= make_temp_ssa_name (TREE_TYPE (TREE_TYPE (fndecl
)), stmt
, "sincostmp");
745 gimple_call_set_lhs (stmt
, res
);
747 def_stmt
= SSA_NAME_DEF_STMT (name
);
748 if (!SSA_NAME_IS_DEFAULT_DEF (name
)
749 && gimple_code (def_stmt
) != GIMPLE_PHI
750 && gimple_bb (def_stmt
) == top_bb
)
752 gsi
= gsi_for_stmt (def_stmt
);
753 gsi_insert_after (&gsi
, stmt
, GSI_SAME_STMT
);
757 gsi
= gsi_after_labels (top_bb
);
758 gsi_insert_before (&gsi
, stmt
, GSI_SAME_STMT
);
760 sincos_stats
.inserted
++;
762 /* And adjust the recorded old call sites. */
763 for (i
= 0; VEC_iterate(gimple
, stmts
, i
, use_stmt
); ++i
)
766 fndecl
= gimple_call_fndecl (use_stmt
);
768 switch (DECL_FUNCTION_CODE (fndecl
))
770 CASE_FLT_FN (BUILT_IN_COS
):
771 rhs
= fold_build1 (REALPART_EXPR
, type
, res
);
774 CASE_FLT_FN (BUILT_IN_SIN
):
775 rhs
= fold_build1 (IMAGPART_EXPR
, type
, res
);
778 CASE_FLT_FN (BUILT_IN_CEXPI
):
786 /* Replace call with a copy. */
787 stmt
= gimple_build_assign (gimple_call_lhs (use_stmt
), rhs
);
789 gsi
= gsi_for_stmt (use_stmt
);
790 gsi_replace (&gsi
, stmt
, true);
791 if (gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
795 VEC_free(gimple
, heap
, stmts
);
800 /* To evaluate powi(x,n), the floating point value x raised to the
801 constant integer exponent n, we use a hybrid algorithm that
802 combines the "window method" with look-up tables. For an
803 introduction to exponentiation algorithms and "addition chains",
804 see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth,
805 "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming",
806 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation
807 Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */
809 /* Provide a default value for POWI_MAX_MULTS, the maximum number of
810 multiplications to inline before calling the system library's pow
811 function. powi(x,n) requires at worst 2*bits(n)-2 multiplications,
812 so this default never requires calling pow, powf or powl. */
814 #ifndef POWI_MAX_MULTS
815 #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2)
818 /* The size of the "optimal power tree" lookup table. All
819 exponents less than this value are simply looked up in the
820 powi_table below. This threshold is also used to size the
821 cache of pseudo registers that hold intermediate results. */
822 #define POWI_TABLE_SIZE 256
824 /* The size, in bits of the window, used in the "window method"
825 exponentiation algorithm. This is equivalent to a radix of
826 (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */
827 #define POWI_WINDOW_SIZE 3
829 /* The following table is an efficient representation of an
830 "optimal power tree". For each value, i, the corresponding
831 value, j, in the table states than an optimal evaluation
832 sequence for calculating pow(x,i) can be found by evaluating
833 pow(x,j)*pow(x,i-j). An optimal power tree for the first
834 100 integers is given in Knuth's "Seminumerical algorithms". */
836 static const unsigned char powi_table
[POWI_TABLE_SIZE
] =
838 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */
839 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */
840 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */
841 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */
842 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */
843 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */
844 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */
845 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */
846 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */
847 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */
848 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */
849 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */
850 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */
851 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */
852 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */
853 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */
854 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */
855 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */
856 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */
857 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */
858 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */
859 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */
860 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */
861 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */
862 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */
863 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */
864 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */
865 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */
866 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */
867 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */
868 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */
869 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */
873 /* Return the number of multiplications required to calculate
874 powi(x,n) where n is less than POWI_TABLE_SIZE. This is a
875 subroutine of powi_cost. CACHE is an array indicating
876 which exponents have already been calculated. */
879 powi_lookup_cost (unsigned HOST_WIDE_INT n
, bool *cache
)
881 /* If we've already calculated this exponent, then this evaluation
882 doesn't require any additional multiplications. */
887 return powi_lookup_cost (n
- powi_table
[n
], cache
)
888 + powi_lookup_cost (powi_table
[n
], cache
) + 1;
891 /* Return the number of multiplications required to calculate
892 powi(x,n) for an arbitrary x, given the exponent N. This
893 function needs to be kept in sync with powi_as_mults below. */
896 powi_cost (HOST_WIDE_INT n
)
898 bool cache
[POWI_TABLE_SIZE
];
899 unsigned HOST_WIDE_INT digit
;
900 unsigned HOST_WIDE_INT val
;
906 /* Ignore the reciprocal when calculating the cost. */
907 val
= (n
< 0) ? -n
: n
;
909 /* Initialize the exponent cache. */
910 memset (cache
, 0, POWI_TABLE_SIZE
* sizeof (bool));
915 while (val
>= POWI_TABLE_SIZE
)
919 digit
= val
& ((1 << POWI_WINDOW_SIZE
) - 1);
920 result
+= powi_lookup_cost (digit
, cache
)
921 + POWI_WINDOW_SIZE
+ 1;
922 val
>>= POWI_WINDOW_SIZE
;
931 return result
+ powi_lookup_cost (val
, cache
);
934 /* Recursive subroutine of powi_as_mults. This function takes the
935 array, CACHE, of already calculated exponents and an exponent N and
936 returns a tree that corresponds to CACHE[1]**N, with type TYPE. */
939 powi_as_mults_1 (gimple_stmt_iterator
*gsi
, location_t loc
, tree type
,
940 HOST_WIDE_INT n
, tree
*cache
)
942 tree op0
, op1
, ssa_target
;
943 unsigned HOST_WIDE_INT digit
;
946 if (n
< POWI_TABLE_SIZE
&& cache
[n
])
949 ssa_target
= make_temp_ssa_name (type
, NULL
, "powmult");
951 if (n
< POWI_TABLE_SIZE
)
953 cache
[n
] = ssa_target
;
954 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
- powi_table
[n
], cache
);
955 op1
= powi_as_mults_1 (gsi
, loc
, type
, powi_table
[n
], cache
);
959 digit
= n
& ((1 << POWI_WINDOW_SIZE
) - 1);
960 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
- digit
, cache
);
961 op1
= powi_as_mults_1 (gsi
, loc
, type
, digit
, cache
);
965 op0
= powi_as_mults_1 (gsi
, loc
, type
, n
>> 1, cache
);
969 mult_stmt
= gimple_build_assign_with_ops (MULT_EXPR
, ssa_target
, op0
, op1
);
970 gimple_set_location (mult_stmt
, loc
);
971 gsi_insert_before (gsi
, mult_stmt
, GSI_SAME_STMT
);
976 /* Convert ARG0**N to a tree of multiplications of ARG0 with itself.
977 This function needs to be kept in sync with powi_cost above. */
980 powi_as_mults (gimple_stmt_iterator
*gsi
, location_t loc
,
981 tree arg0
, HOST_WIDE_INT n
)
983 tree cache
[POWI_TABLE_SIZE
], result
, type
= TREE_TYPE (arg0
);
988 return build_real (type
, dconst1
);
990 memset (cache
, 0, sizeof (cache
));
993 result
= powi_as_mults_1 (gsi
, loc
, type
, (n
< 0) ? -n
: n
, cache
);
997 /* If the original exponent was negative, reciprocate the result. */
998 target
= make_temp_ssa_name (type
, NULL
, "powmult");
999 div_stmt
= gimple_build_assign_with_ops (RDIV_EXPR
, target
,
1000 build_real (type
, dconst1
),
1002 gimple_set_location (div_stmt
, loc
);
1003 gsi_insert_before (gsi
, div_stmt
, GSI_SAME_STMT
);
1008 /* ARG0 and N are the two arguments to a powi builtin in GSI with
1009 location info LOC. If the arguments are appropriate, create an
1010 equivalent sequence of statements prior to GSI using an optimal
1011 number of multiplications, and return an expession holding the
1015 gimple_expand_builtin_powi (gimple_stmt_iterator
*gsi
, location_t loc
,
1016 tree arg0
, HOST_WIDE_INT n
)
1018 /* Avoid largest negative number. */
1020 && ((n
>= -1 && n
<= 2)
1021 || (optimize_function_for_speed_p (cfun
)
1022 && powi_cost (n
) <= POWI_MAX_MULTS
)))
1023 return powi_as_mults (gsi
, loc
, arg0
, n
);
1028 /* Build a gimple call statement that calls FN with argument ARG.
1029 Set the lhs of the call statement to a fresh SSA name. Insert the
1030 statement prior to GSI's current position, and return the fresh
1034 build_and_insert_call (gimple_stmt_iterator
*gsi
, location_t loc
,
1040 call_stmt
= gimple_build_call (fn
, 1, arg
);
1041 ssa_target
= make_temp_ssa_name (TREE_TYPE (arg
), NULL
, "powroot");
1042 gimple_set_lhs (call_stmt
, ssa_target
);
1043 gimple_set_location (call_stmt
, loc
);
1044 gsi_insert_before (gsi
, call_stmt
, GSI_SAME_STMT
);
1049 /* Build a gimple binary operation with the given CODE and arguments
1050 ARG0, ARG1, assigning the result to a new SSA name for variable
1051 TARGET. Insert the statement prior to GSI's current position, and
1052 return the fresh SSA name.*/
1055 build_and_insert_binop (gimple_stmt_iterator
*gsi
, location_t loc
,
1056 const char *name
, enum tree_code code
,
1057 tree arg0
, tree arg1
)
1059 tree result
= make_temp_ssa_name (TREE_TYPE (arg0
), NULL
, name
);
1060 gimple stmt
= gimple_build_assign_with_ops (code
, result
, arg0
, arg1
);
1061 gimple_set_location (stmt
, loc
);
1062 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1066 /* Build a gimple reference operation with the given CODE and argument
1067 ARG, assigning the result to a new SSA name of TYPE with NAME.
1068 Insert the statement prior to GSI's current position, and return
1069 the fresh SSA name. */
1072 build_and_insert_ref (gimple_stmt_iterator
*gsi
, location_t loc
, tree type
,
1073 const char *name
, enum tree_code code
, tree arg0
)
1075 tree result
= make_temp_ssa_name (type
, NULL
, name
);
1076 gimple stmt
= gimple_build_assign (result
, build1 (code
, type
, arg0
));
1077 gimple_set_location (stmt
, loc
);
1078 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1082 /* Build a gimple assignment to cast VAL to TYPE. Insert the statement
1083 prior to GSI's current position, and return the fresh SSA name. */
1086 build_and_insert_cast (gimple_stmt_iterator
*gsi
, location_t loc
,
1087 tree type
, tree val
)
1089 tree result
= make_ssa_name (type
, NULL
);
1090 gimple stmt
= gimple_build_assign_with_ops (NOP_EXPR
, result
, val
, NULL_TREE
);
1091 gimple_set_location (stmt
, loc
);
1092 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1096 /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI
1097 with location info LOC. If possible, create an equivalent and
1098 less expensive sequence of statements prior to GSI, and return an
1099 expession holding the result. */
1102 gimple_expand_builtin_pow (gimple_stmt_iterator
*gsi
, location_t loc
,
1103 tree arg0
, tree arg1
)
1105 REAL_VALUE_TYPE c
, cint
, dconst1_4
, dconst3_4
, dconst1_3
, dconst1_6
;
1106 REAL_VALUE_TYPE c2
, dconst3
;
1108 tree type
, sqrtfn
, cbrtfn
, sqrt_arg0
, sqrt_sqrt
, result
, cbrt_x
, powi_cbrt_x
;
1109 enum machine_mode mode
;
1110 bool hw_sqrt_exists
;
1112 /* If the exponent isn't a constant, there's nothing of interest
1114 if (TREE_CODE (arg1
) != REAL_CST
)
1117 /* If the exponent is equivalent to an integer, expand to an optimal
1118 multiplication sequence when profitable. */
1119 c
= TREE_REAL_CST (arg1
);
1120 n
= real_to_integer (&c
);
1121 real_from_integer (&cint
, VOIDmode
, n
, n
< 0 ? -1 : 0, 0);
1123 if (real_identical (&c
, &cint
)
1124 && ((n
>= -1 && n
<= 2)
1125 || (flag_unsafe_math_optimizations
1126 && optimize_insn_for_speed_p ()
1127 && powi_cost (n
) <= POWI_MAX_MULTS
)))
1128 return gimple_expand_builtin_powi (gsi
, loc
, arg0
, n
);
1130 /* Attempt various optimizations using sqrt and cbrt. */
1131 type
= TREE_TYPE (arg0
);
1132 mode
= TYPE_MODE (type
);
1133 sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1135 /* Optimize pow(x,0.5) = sqrt(x). This replacement is always safe
1136 unless signed zeros must be maintained. pow(-0,0.5) = +0, while
1139 && REAL_VALUES_EQUAL (c
, dconsthalf
)
1140 && !HONOR_SIGNED_ZEROS (mode
))
1141 return build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1143 /* Optimize pow(x,0.25) = sqrt(sqrt(x)). Assume on most machines that
1144 a builtin sqrt instruction is smaller than a call to pow with 0.25,
1145 so do this optimization even if -Os. Don't do this optimization
1146 if we don't have a hardware sqrt insn. */
1147 dconst1_4
= dconst1
;
1148 SET_REAL_EXP (&dconst1_4
, REAL_EXP (&dconst1_4
) - 2);
1149 hw_sqrt_exists
= optab_handler (sqrt_optab
, mode
) != CODE_FOR_nothing
;
1151 if (flag_unsafe_math_optimizations
1153 && REAL_VALUES_EQUAL (c
, dconst1_4
)
1157 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1160 return build_and_insert_call (gsi
, loc
, sqrtfn
, sqrt_arg0
);
1163 /* Optimize pow(x,0.75) = sqrt(x) * sqrt(sqrt(x)) unless we are
1164 optimizing for space. Don't do this optimization if we don't have
1165 a hardware sqrt insn. */
1166 real_from_integer (&dconst3_4
, VOIDmode
, 3, 0, 0);
1167 SET_REAL_EXP (&dconst3_4
, REAL_EXP (&dconst3_4
) - 2);
1169 if (flag_unsafe_math_optimizations
1171 && optimize_function_for_speed_p (cfun
)
1172 && REAL_VALUES_EQUAL (c
, dconst3_4
)
1176 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1179 sqrt_sqrt
= build_and_insert_call (gsi
, loc
, sqrtfn
, sqrt_arg0
);
1181 /* sqrt(x) * sqrt(sqrt(x)) */
1182 return build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1183 sqrt_arg0
, sqrt_sqrt
);
1186 /* Optimize pow(x,1./3.) = cbrt(x). This requires unsafe math
1187 optimizations since 1./3. is not exactly representable. If x
1188 is negative and finite, the correct value of pow(x,1./3.) is
1189 a NaN with the "invalid" exception raised, because the value
1190 of 1./3. actually has an even denominator. The correct value
1191 of cbrt(x) is a negative real value. */
1192 cbrtfn
= mathfn_built_in (type
, BUILT_IN_CBRT
);
1193 dconst1_3
= real_value_truncate (mode
, dconst_third ());
1195 if (flag_unsafe_math_optimizations
1197 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1198 && REAL_VALUES_EQUAL (c
, dconst1_3
))
1199 return build_and_insert_call (gsi
, loc
, cbrtfn
, arg0
);
1201 /* Optimize pow(x,1./6.) = cbrt(sqrt(x)). Don't do this optimization
1202 if we don't have a hardware sqrt insn. */
1203 dconst1_6
= dconst1_3
;
1204 SET_REAL_EXP (&dconst1_6
, REAL_EXP (&dconst1_6
) - 1);
1206 if (flag_unsafe_math_optimizations
1209 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1210 && optimize_function_for_speed_p (cfun
)
1212 && REAL_VALUES_EQUAL (c
, dconst1_6
))
1215 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1218 return build_and_insert_call (gsi
, loc
, cbrtfn
, sqrt_arg0
);
1221 /* Optimize pow(x,c), where n = 2c for some nonzero integer n, into
1223 sqrt(x) * powi(x, n/2), n > 0;
1224 1.0 / (sqrt(x) * powi(x, abs(n/2))), n < 0.
1226 Do not calculate the powi factor when n/2 = 0. */
1227 real_arithmetic (&c2
, MULT_EXPR
, &c
, &dconst2
);
1228 n
= real_to_integer (&c2
);
1229 real_from_integer (&cint
, VOIDmode
, n
, n
< 0 ? -1 : 0, 0);
1231 if (flag_unsafe_math_optimizations
1233 && real_identical (&c2
, &cint
))
1235 tree powi_x_ndiv2
= NULL_TREE
;
1237 /* Attempt to fold powi(arg0, abs(n/2)) into multiplies. If not
1238 possible or profitable, give up. Skip the degenerate case when
1239 n is 1 or -1, where the result is always 1. */
1240 if (absu_hwi (n
) != 1)
1242 powi_x_ndiv2
= gimple_expand_builtin_powi (gsi
, loc
, arg0
,
1248 /* Calculate sqrt(x). When n is not 1 or -1, multiply it by the
1249 result of the optimal multiply sequence just calculated. */
1250 sqrt_arg0
= build_and_insert_call (gsi
, loc
, sqrtfn
, arg0
);
1252 if (absu_hwi (n
) == 1)
1255 result
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1256 sqrt_arg0
, powi_x_ndiv2
);
1258 /* If n is negative, reciprocate the result. */
1260 result
= build_and_insert_binop (gsi
, loc
, "powroot", RDIV_EXPR
,
1261 build_real (type
, dconst1
), result
);
1265 /* Optimize pow(x,c), where 3c = n for some nonzero integer n, into
1267 powi(x, n/3) * powi(cbrt(x), n%3), n > 0;
1268 1.0 / (powi(x, abs(n)/3) * powi(cbrt(x), abs(n)%3)), n < 0.
1270 Do not calculate the first factor when n/3 = 0. As cbrt(x) is
1271 different from pow(x, 1./3.) due to rounding and behavior with
1272 negative x, we need to constrain this transformation to unsafe
1273 math and positive x or finite math. */
1274 real_from_integer (&dconst3
, VOIDmode
, 3, 0, 0);
1275 real_arithmetic (&c2
, MULT_EXPR
, &c
, &dconst3
);
1276 real_round (&c2
, mode
, &c2
);
1277 n
= real_to_integer (&c2
);
1278 real_from_integer (&cint
, VOIDmode
, n
, n
< 0 ? -1 : 0, 0);
1279 real_arithmetic (&c2
, RDIV_EXPR
, &cint
, &dconst3
);
1280 real_convert (&c2
, mode
, &c2
);
1282 if (flag_unsafe_math_optimizations
1284 && (gimple_val_nonnegative_real_p (arg0
) || !HONOR_NANS (mode
))
1285 && real_identical (&c2
, &c
)
1286 && optimize_function_for_speed_p (cfun
)
1287 && powi_cost (n
/ 3) <= POWI_MAX_MULTS
)
1289 tree powi_x_ndiv3
= NULL_TREE
;
1291 /* Attempt to fold powi(arg0, abs(n/3)) into multiplies. If not
1292 possible or profitable, give up. Skip the degenerate case when
1293 abs(n) < 3, where the result is always 1. */
1294 if (absu_hwi (n
) >= 3)
1296 powi_x_ndiv3
= gimple_expand_builtin_powi (gsi
, loc
, arg0
,
1302 /* Calculate powi(cbrt(x), n%3). Don't use gimple_expand_builtin_powi
1303 as that creates an unnecessary variable. Instead, just produce
1304 either cbrt(x) or cbrt(x) * cbrt(x). */
1305 cbrt_x
= build_and_insert_call (gsi
, loc
, cbrtfn
, arg0
);
1307 if (absu_hwi (n
) % 3 == 1)
1308 powi_cbrt_x
= cbrt_x
;
1310 powi_cbrt_x
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1313 /* Multiply the two subexpressions, unless powi(x,abs(n)/3) = 1. */
1314 if (absu_hwi (n
) < 3)
1315 result
= powi_cbrt_x
;
1317 result
= build_and_insert_binop (gsi
, loc
, "powroot", MULT_EXPR
,
1318 powi_x_ndiv3
, powi_cbrt_x
);
1320 /* If n is negative, reciprocate the result. */
1322 result
= build_and_insert_binop (gsi
, loc
, "powroot", RDIV_EXPR
,
1323 build_real (type
, dconst1
), result
);
1328 /* No optimizations succeeded. */
1332 /* ARG is the argument to a cabs builtin call in GSI with location info
1333 LOC. Create a sequence of statements prior to GSI that calculates
1334 sqrt(R*R + I*I), where R and I are the real and imaginary components
1335 of ARG, respectively. Return an expression holding the result. */
1338 gimple_expand_builtin_cabs (gimple_stmt_iterator
*gsi
, location_t loc
, tree arg
)
1340 tree real_part
, imag_part
, addend1
, addend2
, sum
, result
;
1341 tree type
= TREE_TYPE (TREE_TYPE (arg
));
1342 tree sqrtfn
= mathfn_built_in (type
, BUILT_IN_SQRT
);
1343 enum machine_mode mode
= TYPE_MODE (type
);
1345 if (!flag_unsafe_math_optimizations
1346 || !optimize_bb_for_speed_p (gimple_bb (gsi_stmt (*gsi
)))
1348 || optab_handler (sqrt_optab
, mode
) == CODE_FOR_nothing
)
1351 real_part
= build_and_insert_ref (gsi
, loc
, type
, "cabs",
1352 REALPART_EXPR
, arg
);
1353 addend1
= build_and_insert_binop (gsi
, loc
, "cabs", MULT_EXPR
,
1354 real_part
, real_part
);
1355 imag_part
= build_and_insert_ref (gsi
, loc
, type
, "cabs",
1356 IMAGPART_EXPR
, arg
);
1357 addend2
= build_and_insert_binop (gsi
, loc
, "cabs", MULT_EXPR
,
1358 imag_part
, imag_part
);
1359 sum
= build_and_insert_binop (gsi
, loc
, "cabs", PLUS_EXPR
, addend1
, addend2
);
1360 result
= build_and_insert_call (gsi
, loc
, sqrtfn
, sum
);
1365 /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1
1366 on the SSA_NAME argument of each of them. Also expand powi(x,n) into
1367 an optimal number of multiplies, when n is a constant. */
1370 execute_cse_sincos (void)
1373 bool cfg_changed
= false;
1375 calculate_dominance_info (CDI_DOMINATORS
);
1376 memset (&sincos_stats
, 0, sizeof (sincos_stats
));
1380 gimple_stmt_iterator gsi
;
1382 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1384 gimple stmt
= gsi_stmt (gsi
);
1387 if (is_gimple_call (stmt
)
1388 && gimple_call_lhs (stmt
)
1389 && (fndecl
= gimple_call_fndecl (stmt
))
1390 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
)
1392 tree arg
, arg0
, arg1
, result
;
1396 switch (DECL_FUNCTION_CODE (fndecl
))
1398 CASE_FLT_FN (BUILT_IN_COS
):
1399 CASE_FLT_FN (BUILT_IN_SIN
):
1400 CASE_FLT_FN (BUILT_IN_CEXPI
):
1401 /* Make sure we have either sincos or cexp. */
1402 if (!TARGET_HAS_SINCOS
&& !TARGET_C99_FUNCTIONS
)
1405 arg
= gimple_call_arg (stmt
, 0);
1406 if (TREE_CODE (arg
) == SSA_NAME
)
1407 cfg_changed
|= execute_cse_sincos_1 (arg
);
1410 CASE_FLT_FN (BUILT_IN_POW
):
1411 arg0
= gimple_call_arg (stmt
, 0);
1412 arg1
= gimple_call_arg (stmt
, 1);
1414 loc
= gimple_location (stmt
);
1415 result
= gimple_expand_builtin_pow (&gsi
, loc
, arg0
, arg1
);
1419 tree lhs
= gimple_get_lhs (stmt
);
1420 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1421 gimple_set_location (new_stmt
, loc
);
1422 unlink_stmt_vdef (stmt
);
1423 gsi_replace (&gsi
, new_stmt
, true);
1424 if (gimple_vdef (stmt
))
1425 release_ssa_name (gimple_vdef (stmt
));
1429 CASE_FLT_FN (BUILT_IN_POWI
):
1430 arg0
= gimple_call_arg (stmt
, 0);
1431 arg1
= gimple_call_arg (stmt
, 1);
1432 if (!host_integerp (arg1
, 0))
1435 n
= TREE_INT_CST_LOW (arg1
);
1436 loc
= gimple_location (stmt
);
1437 result
= gimple_expand_builtin_powi (&gsi
, loc
, arg0
, n
);
1441 tree lhs
= gimple_get_lhs (stmt
);
1442 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1443 gimple_set_location (new_stmt
, loc
);
1444 unlink_stmt_vdef (stmt
);
1445 gsi_replace (&gsi
, new_stmt
, true);
1446 if (gimple_vdef (stmt
))
1447 release_ssa_name (gimple_vdef (stmt
));
1451 CASE_FLT_FN (BUILT_IN_CABS
):
1452 arg0
= gimple_call_arg (stmt
, 0);
1453 loc
= gimple_location (stmt
);
1454 result
= gimple_expand_builtin_cabs (&gsi
, loc
, arg0
);
1458 tree lhs
= gimple_get_lhs (stmt
);
1459 gimple new_stmt
= gimple_build_assign (lhs
, result
);
1460 gimple_set_location (new_stmt
, loc
);
1461 unlink_stmt_vdef (stmt
);
1462 gsi_replace (&gsi
, new_stmt
, true);
1463 if (gimple_vdef (stmt
))
1464 release_ssa_name (gimple_vdef (stmt
));
1474 statistics_counter_event (cfun
, "sincos statements inserted",
1475 sincos_stats
.inserted
);
1477 free_dominance_info (CDI_DOMINATORS
);
1478 return cfg_changed
? TODO_cleanup_cfg
: 0;
1482 gate_cse_sincos (void)
1484 /* We no longer require either sincos or cexp, since powi expansion
1485 piggybacks on this pass. */
1489 struct gimple_opt_pass pass_cse_sincos
=
1493 "sincos", /* name */
1494 gate_cse_sincos
, /* gate */
1495 execute_cse_sincos
, /* execute */
1498 0, /* static_pass_number */
1499 TV_NONE
, /* tv_id */
1500 PROP_ssa
, /* properties_required */
1501 0, /* properties_provided */
1502 0, /* properties_destroyed */
1503 0, /* todo_flags_start */
1504 TODO_update_ssa
| TODO_verify_ssa
1505 | TODO_verify_stmts
/* todo_flags_finish */
1509 /* A symbolic number is used to detect byte permutation and selection
1510 patterns. Therefore the field N contains an artificial number
1511 consisting of byte size markers:
1513 0 - byte has the value 0
1514 1..size - byte contains the content of the byte
1515 number indexed with that value minus one */
1517 struct symbolic_number
{
1518 unsigned HOST_WIDEST_INT n
;
1522 /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic
1523 number N. Return false if the requested operation is not permitted
1524 on a symbolic number. */
1527 do_shift_rotate (enum tree_code code
,
1528 struct symbolic_number
*n
,
1534 /* Zero out the extra bits of N in order to avoid them being shifted
1535 into the significant bits. */
1536 if (n
->size
< (int)sizeof (HOST_WIDEST_INT
))
1537 n
->n
&= ((unsigned HOST_WIDEST_INT
)1 << (n
->size
* BITS_PER_UNIT
)) - 1;
1548 n
->n
= (n
->n
<< count
) | (n
->n
>> ((n
->size
* BITS_PER_UNIT
) - count
));
1551 n
->n
= (n
->n
>> count
) | (n
->n
<< ((n
->size
* BITS_PER_UNIT
) - count
));
1556 /* Zero unused bits for size. */
1557 if (n
->size
< (int)sizeof (HOST_WIDEST_INT
))
1558 n
->n
&= ((unsigned HOST_WIDEST_INT
)1 << (n
->size
* BITS_PER_UNIT
)) - 1;
1562 /* Perform sanity checking for the symbolic number N and the gimple
1566 verify_symbolic_number_p (struct symbolic_number
*n
, gimple stmt
)
1570 lhs_type
= gimple_expr_type (stmt
);
1572 if (TREE_CODE (lhs_type
) != INTEGER_TYPE
)
1575 if (TYPE_PRECISION (lhs_type
) != n
->size
* BITS_PER_UNIT
)
1581 /* find_bswap_1 invokes itself recursively with N and tries to perform
1582 the operation given by the rhs of STMT on the result. If the
1583 operation could successfully be executed the function returns the
1584 tree expression of the source operand and NULL otherwise. */
1587 find_bswap_1 (gimple stmt
, struct symbolic_number
*n
, int limit
)
1589 enum tree_code code
;
1590 tree rhs1
, rhs2
= NULL
;
1591 gimple rhs1_stmt
, rhs2_stmt
;
1593 enum gimple_rhs_class rhs_class
;
1595 if (!limit
|| !is_gimple_assign (stmt
))
1598 rhs1
= gimple_assign_rhs1 (stmt
);
1600 if (TREE_CODE (rhs1
) != SSA_NAME
)
1603 code
= gimple_assign_rhs_code (stmt
);
1604 rhs_class
= gimple_assign_rhs_class (stmt
);
1605 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
1607 if (rhs_class
== GIMPLE_BINARY_RHS
)
1608 rhs2
= gimple_assign_rhs2 (stmt
);
1610 /* Handle unary rhs and binary rhs with integer constants as second
1613 if (rhs_class
== GIMPLE_UNARY_RHS
1614 || (rhs_class
== GIMPLE_BINARY_RHS
1615 && TREE_CODE (rhs2
) == INTEGER_CST
))
1617 if (code
!= BIT_AND_EXPR
1618 && code
!= LSHIFT_EXPR
1619 && code
!= RSHIFT_EXPR
1620 && code
!= LROTATE_EXPR
1621 && code
!= RROTATE_EXPR
1623 && code
!= CONVERT_EXPR
)
1626 source_expr1
= find_bswap_1 (rhs1_stmt
, n
, limit
- 1);
1628 /* If find_bswap_1 returned NULL STMT is a leaf node and we have
1629 to initialize the symbolic number. */
1632 /* Set up the symbolic number N by setting each byte to a
1633 value between 1 and the byte size of rhs1. The highest
1634 order byte is set to n->size and the lowest order
1636 n
->size
= TYPE_PRECISION (TREE_TYPE (rhs1
));
1637 if (n
->size
% BITS_PER_UNIT
!= 0)
1639 n
->size
/= BITS_PER_UNIT
;
1640 n
->n
= (sizeof (HOST_WIDEST_INT
) < 8 ? 0 :
1641 (unsigned HOST_WIDEST_INT
)0x08070605 << 32 | 0x04030201);
1643 if (n
->size
< (int)sizeof (HOST_WIDEST_INT
))
1644 n
->n
&= ((unsigned HOST_WIDEST_INT
)1 <<
1645 (n
->size
* BITS_PER_UNIT
)) - 1;
1647 source_expr1
= rhs1
;
1655 unsigned HOST_WIDEST_INT val
= widest_int_cst_value (rhs2
);
1656 unsigned HOST_WIDEST_INT tmp
= val
;
1658 /* Only constants masking full bytes are allowed. */
1659 for (i
= 0; i
< n
->size
; i
++, tmp
>>= BITS_PER_UNIT
)
1660 if ((tmp
& 0xff) != 0 && (tmp
& 0xff) != 0xff)
1670 if (!do_shift_rotate (code
, n
, (int)TREE_INT_CST_LOW (rhs2
)))
1677 type_size
= TYPE_PRECISION (gimple_expr_type (stmt
));
1678 if (type_size
% BITS_PER_UNIT
!= 0)
1681 if (type_size
/ BITS_PER_UNIT
< (int)(sizeof (HOST_WIDEST_INT
)))
1683 /* If STMT casts to a smaller type mask out the bits not
1684 belonging to the target type. */
1685 n
->n
&= ((unsigned HOST_WIDEST_INT
)1 << type_size
) - 1;
1687 n
->size
= type_size
/ BITS_PER_UNIT
;
1693 return verify_symbolic_number_p (n
, stmt
) ? source_expr1
: NULL
;
1696 /* Handle binary rhs. */
1698 if (rhs_class
== GIMPLE_BINARY_RHS
)
1700 struct symbolic_number n1
, n2
;
1703 if (code
!= BIT_IOR_EXPR
)
1706 if (TREE_CODE (rhs2
) != SSA_NAME
)
1709 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
1714 source_expr1
= find_bswap_1 (rhs1_stmt
, &n1
, limit
- 1);
1719 source_expr2
= find_bswap_1 (rhs2_stmt
, &n2
, limit
- 1);
1721 if (source_expr1
!= source_expr2
1722 || n1
.size
!= n2
.size
)
1728 if (!verify_symbolic_number_p (n
, stmt
))
1735 return source_expr1
;
1740 /* Check if STMT completes a bswap implementation consisting of ORs,
1741 SHIFTs and ANDs. Return the source tree expression on which the
1742 byte swap is performed and NULL if no bswap was found. */
1745 find_bswap (gimple stmt
)
1747 /* The number which the find_bswap result should match in order to
1748 have a full byte swap. The number is shifted to the left according
1749 to the size of the symbolic number before using it. */
1750 unsigned HOST_WIDEST_INT cmp
=
1751 sizeof (HOST_WIDEST_INT
) < 8 ? 0 :
1752 (unsigned HOST_WIDEST_INT
)0x01020304 << 32 | 0x05060708;
1754 struct symbolic_number n
;
1758 /* The last parameter determines the depth search limit. It usually
1759 correlates directly to the number of bytes to be touched. We
1760 increase that number by three here in order to also
1761 cover signed -> unsigned converions of the src operand as can be seen
1762 in libgcc, and for initial shift/and operation of the src operand. */
1763 limit
= TREE_INT_CST_LOW (TYPE_SIZE_UNIT (gimple_expr_type (stmt
)));
1764 limit
+= 1 + (int) ceil_log2 ((unsigned HOST_WIDE_INT
) limit
);
1765 source_expr
= find_bswap_1 (stmt
, &n
, limit
);
1770 /* Zero out the extra bits of N and CMP. */
1771 if (n
.size
< (int)sizeof (HOST_WIDEST_INT
))
1773 unsigned HOST_WIDEST_INT mask
=
1774 ((unsigned HOST_WIDEST_INT
)1 << (n
.size
* BITS_PER_UNIT
)) - 1;
1777 cmp
>>= (sizeof (HOST_WIDEST_INT
) - n
.size
) * BITS_PER_UNIT
;
1780 /* A complete byte swap should make the symbolic number to start
1781 with the largest digit in the highest order byte. */
1788 /* Find manual byte swap implementations and turn them into a bswap
1789 builtin invokation. */
1792 execute_optimize_bswap (void)
1795 bool bswap32_p
, bswap64_p
;
1796 bool changed
= false;
1797 tree bswap32_type
= NULL_TREE
, bswap64_type
= NULL_TREE
;
1799 if (BITS_PER_UNIT
!= 8)
1802 if (sizeof (HOST_WIDEST_INT
) < 8)
1805 bswap32_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP32
)
1806 && optab_handler (bswap_optab
, SImode
) != CODE_FOR_nothing
);
1807 bswap64_p
= (builtin_decl_explicit_p (BUILT_IN_BSWAP64
)
1808 && (optab_handler (bswap_optab
, DImode
) != CODE_FOR_nothing
1809 || (bswap32_p
&& word_mode
== SImode
)));
1811 if (!bswap32_p
&& !bswap64_p
)
1814 /* Determine the argument type of the builtins. The code later on
1815 assumes that the return and argument type are the same. */
1818 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP32
);
1819 bswap32_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
1824 tree fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP64
);
1825 bswap64_type
= TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl
)));
1828 memset (&bswap_stats
, 0, sizeof (bswap_stats
));
1832 gimple_stmt_iterator gsi
;
1834 /* We do a reverse scan for bswap patterns to make sure we get the
1835 widest match. As bswap pattern matching doesn't handle
1836 previously inserted smaller bswap replacements as sub-
1837 patterns, the wider variant wouldn't be detected. */
1838 for (gsi
= gsi_last_bb (bb
); !gsi_end_p (gsi
); gsi_prev (&gsi
))
1840 gimple stmt
= gsi_stmt (gsi
);
1841 tree bswap_src
, bswap_type
;
1843 tree fndecl
= NULL_TREE
;
1847 if (!is_gimple_assign (stmt
)
1848 || gimple_assign_rhs_code (stmt
) != BIT_IOR_EXPR
)
1851 type_size
= TYPE_PRECISION (gimple_expr_type (stmt
));
1858 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP32
);
1859 bswap_type
= bswap32_type
;
1865 fndecl
= builtin_decl_explicit (BUILT_IN_BSWAP64
);
1866 bswap_type
= bswap64_type
;
1876 bswap_src
= find_bswap (stmt
);
1882 if (type_size
== 32)
1883 bswap_stats
.found_32bit
++;
1885 bswap_stats
.found_64bit
++;
1887 bswap_tmp
= bswap_src
;
1889 /* Convert the src expression if necessary. */
1890 if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp
), bswap_type
))
1892 gimple convert_stmt
;
1893 bswap_tmp
= make_temp_ssa_name (bswap_type
, NULL
, "bswapsrc");
1894 convert_stmt
= gimple_build_assign_with_ops
1895 (NOP_EXPR
, bswap_tmp
, bswap_src
, NULL
);
1896 gsi_insert_before (&gsi
, convert_stmt
, GSI_SAME_STMT
);
1899 call
= gimple_build_call (fndecl
, 1, bswap_tmp
);
1901 bswap_tmp
= gimple_assign_lhs (stmt
);
1903 /* Convert the result if necessary. */
1904 if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp
), bswap_type
))
1906 gimple convert_stmt
;
1907 bswap_tmp
= make_temp_ssa_name (bswap_type
, NULL
, "bswapdst");
1908 convert_stmt
= gimple_build_assign_with_ops
1909 (NOP_EXPR
, gimple_assign_lhs (stmt
), bswap_tmp
, NULL
);
1910 gsi_insert_after (&gsi
, convert_stmt
, GSI_SAME_STMT
);
1913 gimple_call_set_lhs (call
, bswap_tmp
);
1917 fprintf (dump_file
, "%d bit bswap implementation found at: ",
1919 print_gimple_stmt (dump_file
, stmt
, 0, 0);
1922 gsi_insert_after (&gsi
, call
, GSI_SAME_STMT
);
1923 gsi_remove (&gsi
, true);
1927 statistics_counter_event (cfun
, "32-bit bswap implementations found",
1928 bswap_stats
.found_32bit
);
1929 statistics_counter_event (cfun
, "64-bit bswap implementations found",
1930 bswap_stats
.found_64bit
);
1932 return (changed
? TODO_update_ssa
| TODO_verify_ssa
1933 | TODO_verify_stmts
: 0);
1937 gate_optimize_bswap (void)
1939 return flag_expensive_optimizations
&& optimize
;
1942 struct gimple_opt_pass pass_optimize_bswap
=
1947 gate_optimize_bswap
, /* gate */
1948 execute_optimize_bswap
, /* execute */
1951 0, /* static_pass_number */
1952 TV_NONE
, /* tv_id */
1953 PROP_ssa
, /* properties_required */
1954 0, /* properties_provided */
1955 0, /* properties_destroyed */
1956 0, /* todo_flags_start */
1957 0 /* todo_flags_finish */
1961 /* Return true if stmt is a type conversion operation that can be stripped
1962 when used in a widening multiply operation. */
1964 widening_mult_conversion_strippable_p (tree result_type
, gimple stmt
)
1966 enum tree_code rhs_code
= gimple_assign_rhs_code (stmt
);
1968 if (TREE_CODE (result_type
) == INTEGER_TYPE
)
1973 if (!CONVERT_EXPR_CODE_P (rhs_code
))
1976 op_type
= TREE_TYPE (gimple_assign_lhs (stmt
));
1978 /* If the type of OP has the same precision as the result, then
1979 we can strip this conversion. The multiply operation will be
1980 selected to create the correct extension as a by-product. */
1981 if (TYPE_PRECISION (result_type
) == TYPE_PRECISION (op_type
))
1984 /* We can also strip a conversion if it preserves the signed-ness of
1985 the operation and doesn't narrow the range. */
1986 inner_op_type
= TREE_TYPE (gimple_assign_rhs1 (stmt
));
1988 /* If the inner-most type is unsigned, then we can strip any
1989 intermediate widening operation. If it's signed, then the
1990 intermediate widening operation must also be signed. */
1991 if ((TYPE_UNSIGNED (inner_op_type
)
1992 || TYPE_UNSIGNED (op_type
) == TYPE_UNSIGNED (inner_op_type
))
1993 && TYPE_PRECISION (op_type
) > TYPE_PRECISION (inner_op_type
))
1999 return rhs_code
== FIXED_CONVERT_EXPR
;
2002 /* Return true if RHS is a suitable operand for a widening multiplication,
2003 assuming a target type of TYPE.
2004 There are two cases:
2006 - RHS makes some value at least twice as wide. Store that value
2007 in *NEW_RHS_OUT if so, and store its type in *TYPE_OUT.
2009 - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so,
2010 but leave *TYPE_OUT untouched. */
2013 is_widening_mult_rhs_p (tree type
, tree rhs
, tree
*type_out
,
2019 if (TREE_CODE (rhs
) == SSA_NAME
)
2021 stmt
= SSA_NAME_DEF_STMT (rhs
);
2022 if (is_gimple_assign (stmt
))
2024 if (! widening_mult_conversion_strippable_p (type
, stmt
))
2028 rhs1
= gimple_assign_rhs1 (stmt
);
2030 if (TREE_CODE (rhs1
) == INTEGER_CST
)
2032 *new_rhs_out
= rhs1
;
2041 type1
= TREE_TYPE (rhs1
);
2043 if (TREE_CODE (type1
) != TREE_CODE (type
)
2044 || TYPE_PRECISION (type1
) * 2 > TYPE_PRECISION (type
))
2047 *new_rhs_out
= rhs1
;
2052 if (TREE_CODE (rhs
) == INTEGER_CST
)
2062 /* Return true if STMT performs a widening multiplication, assuming the
2063 output type is TYPE. If so, store the unwidened types of the operands
2064 in *TYPE1_OUT and *TYPE2_OUT respectively. Also fill *RHS1_OUT and
2065 *RHS2_OUT such that converting those operands to types *TYPE1_OUT
2066 and *TYPE2_OUT would give the operands of the multiplication. */
2069 is_widening_mult_p (gimple stmt
,
2070 tree
*type1_out
, tree
*rhs1_out
,
2071 tree
*type2_out
, tree
*rhs2_out
)
2073 tree type
= TREE_TYPE (gimple_assign_lhs (stmt
));
2075 if (TREE_CODE (type
) != INTEGER_TYPE
2076 && TREE_CODE (type
) != FIXED_POINT_TYPE
)
2079 if (!is_widening_mult_rhs_p (type
, gimple_assign_rhs1 (stmt
), type1_out
,
2083 if (!is_widening_mult_rhs_p (type
, gimple_assign_rhs2 (stmt
), type2_out
,
2087 if (*type1_out
== NULL
)
2089 if (*type2_out
== NULL
|| !int_fits_type_p (*rhs1_out
, *type2_out
))
2091 *type1_out
= *type2_out
;
2094 if (*type2_out
== NULL
)
2096 if (!int_fits_type_p (*rhs2_out
, *type1_out
))
2098 *type2_out
= *type1_out
;
2101 /* Ensure that the larger of the two operands comes first. */
2102 if (TYPE_PRECISION (*type1_out
) < TYPE_PRECISION (*type2_out
))
2106 *type1_out
= *type2_out
;
2109 *rhs1_out
= *rhs2_out
;
2116 /* Process a single gimple statement STMT, which has a MULT_EXPR as
2117 its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return
2118 value is true iff we converted the statement. */
2121 convert_mult_to_widen (gimple stmt
, gimple_stmt_iterator
*gsi
)
2123 tree lhs
, rhs1
, rhs2
, type
, type1
, type2
;
2124 enum insn_code handler
;
2125 enum machine_mode to_mode
, from_mode
, actual_mode
;
2127 int actual_precision
;
2128 location_t loc
= gimple_location (stmt
);
2129 bool from_unsigned1
, from_unsigned2
;
2131 lhs
= gimple_assign_lhs (stmt
);
2132 type
= TREE_TYPE (lhs
);
2133 if (TREE_CODE (type
) != INTEGER_TYPE
)
2136 if (!is_widening_mult_p (stmt
, &type1
, &rhs1
, &type2
, &rhs2
))
2139 to_mode
= TYPE_MODE (type
);
2140 from_mode
= TYPE_MODE (type1
);
2141 from_unsigned1
= TYPE_UNSIGNED (type1
);
2142 from_unsigned2
= TYPE_UNSIGNED (type2
);
2144 if (from_unsigned1
&& from_unsigned2
)
2145 op
= umul_widen_optab
;
2146 else if (!from_unsigned1
&& !from_unsigned2
)
2147 op
= smul_widen_optab
;
2149 op
= usmul_widen_optab
;
2151 handler
= find_widening_optab_handler_and_mode (op
, to_mode
, from_mode
,
2154 if (handler
== CODE_FOR_nothing
)
2156 if (op
!= smul_widen_optab
)
2158 /* We can use a signed multiply with unsigned types as long as
2159 there is a wider mode to use, or it is the smaller of the two
2160 types that is unsigned. Note that type1 >= type2, always. */
2161 if ((TYPE_UNSIGNED (type1
)
2162 && TYPE_PRECISION (type1
) == GET_MODE_PRECISION (from_mode
))
2163 || (TYPE_UNSIGNED (type2
)
2164 && TYPE_PRECISION (type2
) == GET_MODE_PRECISION (from_mode
)))
2166 from_mode
= GET_MODE_WIDER_MODE (from_mode
);
2167 if (GET_MODE_SIZE (to_mode
) <= GET_MODE_SIZE (from_mode
))
2171 op
= smul_widen_optab
;
2172 handler
= find_widening_optab_handler_and_mode (op
, to_mode
,
2176 if (handler
== CODE_FOR_nothing
)
2179 from_unsigned1
= from_unsigned2
= false;
2185 /* Ensure that the inputs to the handler are in the correct precison
2186 for the opcode. This will be the full mode size. */
2187 actual_precision
= GET_MODE_PRECISION (actual_mode
);
2188 if (2 * actual_precision
> TYPE_PRECISION (type
))
2190 if (actual_precision
!= TYPE_PRECISION (type1
)
2191 || from_unsigned1
!= TYPE_UNSIGNED (type1
))
2192 rhs1
= build_and_insert_cast (gsi
, loc
,
2193 build_nonstandard_integer_type
2194 (actual_precision
, from_unsigned1
), rhs1
);
2195 if (actual_precision
!= TYPE_PRECISION (type2
)
2196 || from_unsigned2
!= TYPE_UNSIGNED (type2
))
2197 rhs2
= build_and_insert_cast (gsi
, loc
,
2198 build_nonstandard_integer_type
2199 (actual_precision
, from_unsigned2
), rhs2
);
2201 /* Handle constants. */
2202 if (TREE_CODE (rhs1
) == INTEGER_CST
)
2203 rhs1
= fold_convert (type1
, rhs1
);
2204 if (TREE_CODE (rhs2
) == INTEGER_CST
)
2205 rhs2
= fold_convert (type2
, rhs2
);
2207 gimple_assign_set_rhs1 (stmt
, rhs1
);
2208 gimple_assign_set_rhs2 (stmt
, rhs2
);
2209 gimple_assign_set_rhs_code (stmt
, WIDEN_MULT_EXPR
);
2211 widen_mul_stats
.widen_mults_inserted
++;
2215 /* Process a single gimple statement STMT, which is found at the
2216 iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its
2217 rhs (given by CODE), and try to convert it into a
2218 WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value
2219 is true iff we converted the statement. */
2222 convert_plusminus_to_widen (gimple_stmt_iterator
*gsi
, gimple stmt
,
2223 enum tree_code code
)
2225 gimple rhs1_stmt
= NULL
, rhs2_stmt
= NULL
;
2226 gimple conv1_stmt
= NULL
, conv2_stmt
= NULL
, conv_stmt
;
2227 tree type
, type1
, type2
, optype
;
2228 tree lhs
, rhs1
, rhs2
, mult_rhs1
, mult_rhs2
, add_rhs
;
2229 enum tree_code rhs1_code
= ERROR_MARK
, rhs2_code
= ERROR_MARK
;
2231 enum tree_code wmult_code
;
2232 enum insn_code handler
;
2233 enum machine_mode to_mode
, from_mode
, actual_mode
;
2234 location_t loc
= gimple_location (stmt
);
2235 int actual_precision
;
2236 bool from_unsigned1
, from_unsigned2
;
2238 lhs
= gimple_assign_lhs (stmt
);
2239 type
= TREE_TYPE (lhs
);
2240 if (TREE_CODE (type
) != INTEGER_TYPE
2241 && TREE_CODE (type
) != FIXED_POINT_TYPE
)
2244 if (code
== MINUS_EXPR
)
2245 wmult_code
= WIDEN_MULT_MINUS_EXPR
;
2247 wmult_code
= WIDEN_MULT_PLUS_EXPR
;
2249 rhs1
= gimple_assign_rhs1 (stmt
);
2250 rhs2
= gimple_assign_rhs2 (stmt
);
2252 if (TREE_CODE (rhs1
) == SSA_NAME
)
2254 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
2255 if (is_gimple_assign (rhs1_stmt
))
2256 rhs1_code
= gimple_assign_rhs_code (rhs1_stmt
);
2259 if (TREE_CODE (rhs2
) == SSA_NAME
)
2261 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
2262 if (is_gimple_assign (rhs2_stmt
))
2263 rhs2_code
= gimple_assign_rhs_code (rhs2_stmt
);
2266 /* Allow for one conversion statement between the multiply
2267 and addition/subtraction statement. If there are more than
2268 one conversions then we assume they would invalidate this
2269 transformation. If that's not the case then they should have
2270 been folded before now. */
2271 if (CONVERT_EXPR_CODE_P (rhs1_code
))
2273 conv1_stmt
= rhs1_stmt
;
2274 rhs1
= gimple_assign_rhs1 (rhs1_stmt
);
2275 if (TREE_CODE (rhs1
) == SSA_NAME
)
2277 rhs1_stmt
= SSA_NAME_DEF_STMT (rhs1
);
2278 if (is_gimple_assign (rhs1_stmt
))
2279 rhs1_code
= gimple_assign_rhs_code (rhs1_stmt
);
2284 if (CONVERT_EXPR_CODE_P (rhs2_code
))
2286 conv2_stmt
= rhs2_stmt
;
2287 rhs2
= gimple_assign_rhs1 (rhs2_stmt
);
2288 if (TREE_CODE (rhs2
) == SSA_NAME
)
2290 rhs2_stmt
= SSA_NAME_DEF_STMT (rhs2
);
2291 if (is_gimple_assign (rhs2_stmt
))
2292 rhs2_code
= gimple_assign_rhs_code (rhs2_stmt
);
2298 /* If code is WIDEN_MULT_EXPR then it would seem unnecessary to call
2299 is_widening_mult_p, but we still need the rhs returns.
2301 It might also appear that it would be sufficient to use the existing
2302 operands of the widening multiply, but that would limit the choice of
2303 multiply-and-accumulate instructions. */
2304 if (code
== PLUS_EXPR
2305 && (rhs1_code
== MULT_EXPR
|| rhs1_code
== WIDEN_MULT_EXPR
))
2307 if (!is_widening_mult_p (rhs1_stmt
, &type1
, &mult_rhs1
,
2308 &type2
, &mult_rhs2
))
2311 conv_stmt
= conv1_stmt
;
2313 else if (rhs2_code
== MULT_EXPR
|| rhs2_code
== WIDEN_MULT_EXPR
)
2315 if (!is_widening_mult_p (rhs2_stmt
, &type1
, &mult_rhs1
,
2316 &type2
, &mult_rhs2
))
2319 conv_stmt
= conv2_stmt
;
2324 to_mode
= TYPE_MODE (type
);
2325 from_mode
= TYPE_MODE (type1
);
2326 from_unsigned1
= TYPE_UNSIGNED (type1
);
2327 from_unsigned2
= TYPE_UNSIGNED (type2
);
2330 /* There's no such thing as a mixed sign madd yet, so use a wider mode. */
2331 if (from_unsigned1
!= from_unsigned2
)
2333 if (!INTEGRAL_TYPE_P (type
))
2335 /* We can use a signed multiply with unsigned types as long as
2336 there is a wider mode to use, or it is the smaller of the two
2337 types that is unsigned. Note that type1 >= type2, always. */
2339 && TYPE_PRECISION (type1
) == GET_MODE_PRECISION (from_mode
))
2341 && TYPE_PRECISION (type2
) == GET_MODE_PRECISION (from_mode
)))
2343 from_mode
= GET_MODE_WIDER_MODE (from_mode
);
2344 if (GET_MODE_SIZE (from_mode
) >= GET_MODE_SIZE (to_mode
))
2348 from_unsigned1
= from_unsigned2
= false;
2349 optype
= build_nonstandard_integer_type (GET_MODE_PRECISION (from_mode
),
2353 /* If there was a conversion between the multiply and addition
2354 then we need to make sure it fits a multiply-and-accumulate.
2355 The should be a single mode change which does not change the
2359 /* We use the original, unmodified data types for this. */
2360 tree from_type
= TREE_TYPE (gimple_assign_rhs1 (conv_stmt
));
2361 tree to_type
= TREE_TYPE (gimple_assign_lhs (conv_stmt
));
2362 int data_size
= TYPE_PRECISION (type1
) + TYPE_PRECISION (type2
);
2363 bool is_unsigned
= TYPE_UNSIGNED (type1
) && TYPE_UNSIGNED (type2
);
2365 if (TYPE_PRECISION (from_type
) > TYPE_PRECISION (to_type
))
2367 /* Conversion is a truncate. */
2368 if (TYPE_PRECISION (to_type
) < data_size
)
2371 else if (TYPE_PRECISION (from_type
) < TYPE_PRECISION (to_type
))
2373 /* Conversion is an extend. Check it's the right sort. */
2374 if (TYPE_UNSIGNED (from_type
) != is_unsigned
2375 && !(is_unsigned
&& TYPE_PRECISION (from_type
) > data_size
))
2378 /* else convert is a no-op for our purposes. */
2381 /* Verify that the machine can perform a widening multiply
2382 accumulate in this mode/signedness combination, otherwise
2383 this transformation is likely to pessimize code. */
2384 this_optab
= optab_for_tree_code (wmult_code
, optype
, optab_default
);
2385 handler
= find_widening_optab_handler_and_mode (this_optab
, to_mode
,
2386 from_mode
, 0, &actual_mode
);
2388 if (handler
== CODE_FOR_nothing
)
2391 /* Ensure that the inputs to the handler are in the correct precison
2392 for the opcode. This will be the full mode size. */
2393 actual_precision
= GET_MODE_PRECISION (actual_mode
);
2394 if (actual_precision
!= TYPE_PRECISION (type1
)
2395 || from_unsigned1
!= TYPE_UNSIGNED (type1
))
2396 mult_rhs1
= build_and_insert_cast (gsi
, loc
,
2397 build_nonstandard_integer_type
2398 (actual_precision
, from_unsigned1
),
2400 if (actual_precision
!= TYPE_PRECISION (type2
)
2401 || from_unsigned2
!= TYPE_UNSIGNED (type2
))
2402 mult_rhs2
= build_and_insert_cast (gsi
, loc
,
2403 build_nonstandard_integer_type
2404 (actual_precision
, from_unsigned2
),
2407 if (!useless_type_conversion_p (type
, TREE_TYPE (add_rhs
)))
2408 add_rhs
= build_and_insert_cast (gsi
, loc
, type
, add_rhs
);
2410 /* Handle constants. */
2411 if (TREE_CODE (mult_rhs1
) == INTEGER_CST
)
2412 mult_rhs1
= fold_convert (type1
, mult_rhs1
);
2413 if (TREE_CODE (mult_rhs2
) == INTEGER_CST
)
2414 mult_rhs2
= fold_convert (type2
, mult_rhs2
);
2416 gimple_assign_set_rhs_with_ops_1 (gsi
, wmult_code
, mult_rhs1
, mult_rhs2
,
2418 update_stmt (gsi_stmt (*gsi
));
2419 widen_mul_stats
.maccs_inserted
++;
2423 /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2
2424 with uses in additions and subtractions to form fused multiply-add
2425 operations. Returns true if successful and MUL_STMT should be removed. */
2428 convert_mult_to_fma (gimple mul_stmt
, tree op1
, tree op2
)
2430 tree mul_result
= gimple_get_lhs (mul_stmt
);
2431 tree type
= TREE_TYPE (mul_result
);
2432 gimple use_stmt
, neguse_stmt
, fma_stmt
;
2433 use_operand_p use_p
;
2434 imm_use_iterator imm_iter
;
2436 if (FLOAT_TYPE_P (type
)
2437 && flag_fp_contract_mode
== FP_CONTRACT_OFF
)
2440 /* We don't want to do bitfield reduction ops. */
2441 if (INTEGRAL_TYPE_P (type
)
2442 && (TYPE_PRECISION (type
)
2443 != GET_MODE_PRECISION (TYPE_MODE (type
))))
2446 /* If the target doesn't support it, don't generate it. We assume that
2447 if fma isn't available then fms, fnma or fnms are not either. */
2448 if (optab_handler (fma_optab
, TYPE_MODE (type
)) == CODE_FOR_nothing
)
2451 /* If the multiplication has zero uses, it is kept around probably because
2452 of -fnon-call-exceptions. Don't optimize it away in that case,
2454 if (has_zero_uses (mul_result
))
2457 /* Make sure that the multiplication statement becomes dead after
2458 the transformation, thus that all uses are transformed to FMAs.
2459 This means we assume that an FMA operation has the same cost
2461 FOR_EACH_IMM_USE_FAST (use_p
, imm_iter
, mul_result
)
2463 enum tree_code use_code
;
2464 tree result
= mul_result
;
2465 bool negate_p
= false;
2467 use_stmt
= USE_STMT (use_p
);
2469 if (is_gimple_debug (use_stmt
))
2472 /* For now restrict this operations to single basic blocks. In theory
2473 we would want to support sinking the multiplication in
2479 to form a fma in the then block and sink the multiplication to the
2481 if (gimple_bb (use_stmt
) != gimple_bb (mul_stmt
))
2484 if (!is_gimple_assign (use_stmt
))
2487 use_code
= gimple_assign_rhs_code (use_stmt
);
2489 /* A negate on the multiplication leads to FNMA. */
2490 if (use_code
== NEGATE_EXPR
)
2495 result
= gimple_assign_lhs (use_stmt
);
2497 /* Make sure the negate statement becomes dead with this
2498 single transformation. */
2499 if (!single_imm_use (gimple_assign_lhs (use_stmt
),
2500 &use_p
, &neguse_stmt
))
2503 /* Make sure the multiplication isn't also used on that stmt. */
2504 FOR_EACH_PHI_OR_STMT_USE (usep
, neguse_stmt
, iter
, SSA_OP_USE
)
2505 if (USE_FROM_PTR (usep
) == mul_result
)
2509 use_stmt
= neguse_stmt
;
2510 if (gimple_bb (use_stmt
) != gimple_bb (mul_stmt
))
2512 if (!is_gimple_assign (use_stmt
))
2515 use_code
= gimple_assign_rhs_code (use_stmt
);
2522 if (gimple_assign_rhs2 (use_stmt
) == result
)
2523 negate_p
= !negate_p
;
2528 /* FMA can only be formed from PLUS and MINUS. */
2532 /* We can't handle a * b + a * b. */
2533 if (gimple_assign_rhs1 (use_stmt
) == gimple_assign_rhs2 (use_stmt
))
2536 /* While it is possible to validate whether or not the exact form
2537 that we've recognized is available in the backend, the assumption
2538 is that the transformation is never a loss. For instance, suppose
2539 the target only has the plain FMA pattern available. Consider
2540 a*b-c -> fma(a,b,-c): we've exchanged MUL+SUB for FMA+NEG, which
2541 is still two operations. Consider -(a*b)-c -> fma(-a,b,-c): we
2542 still have 3 operations, but in the FMA form the two NEGs are
2543 independent and could be run in parallel. */
2546 FOR_EACH_IMM_USE_STMT (use_stmt
, imm_iter
, mul_result
)
2548 gimple_stmt_iterator gsi
= gsi_for_stmt (use_stmt
);
2549 enum tree_code use_code
;
2550 tree addop
, mulop1
= op1
, result
= mul_result
;
2551 bool negate_p
= false;
2553 if (is_gimple_debug (use_stmt
))
2556 use_code
= gimple_assign_rhs_code (use_stmt
);
2557 if (use_code
== NEGATE_EXPR
)
2559 result
= gimple_assign_lhs (use_stmt
);
2560 single_imm_use (gimple_assign_lhs (use_stmt
), &use_p
, &neguse_stmt
);
2561 gsi_remove (&gsi
, true);
2562 release_defs (use_stmt
);
2564 use_stmt
= neguse_stmt
;
2565 gsi
= gsi_for_stmt (use_stmt
);
2566 use_code
= gimple_assign_rhs_code (use_stmt
);
2570 if (gimple_assign_rhs1 (use_stmt
) == result
)
2572 addop
= gimple_assign_rhs2 (use_stmt
);
2573 /* a * b - c -> a * b + (-c) */
2574 if (gimple_assign_rhs_code (use_stmt
) == MINUS_EXPR
)
2575 addop
= force_gimple_operand_gsi (&gsi
,
2576 build1 (NEGATE_EXPR
,
2578 true, NULL_TREE
, true,
2583 addop
= gimple_assign_rhs1 (use_stmt
);
2584 /* a - b * c -> (-b) * c + a */
2585 if (gimple_assign_rhs_code (use_stmt
) == MINUS_EXPR
)
2586 negate_p
= !negate_p
;
2590 mulop1
= force_gimple_operand_gsi (&gsi
,
2591 build1 (NEGATE_EXPR
,
2593 true, NULL_TREE
, true,
2596 fma_stmt
= gimple_build_assign_with_ops3 (FMA_EXPR
,
2597 gimple_assign_lhs (use_stmt
),
2600 gsi_replace (&gsi
, fma_stmt
, true);
2601 widen_mul_stats
.fmas_inserted
++;
2607 /* Find integer multiplications where the operands are extended from
2608 smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR
2609 where appropriate. */
2612 execute_optimize_widening_mul (void)
2615 bool cfg_changed
= false;
2617 memset (&widen_mul_stats
, 0, sizeof (widen_mul_stats
));
2621 gimple_stmt_iterator gsi
;
2623 for (gsi
= gsi_after_labels (bb
); !gsi_end_p (gsi
);)
2625 gimple stmt
= gsi_stmt (gsi
);
2626 enum tree_code code
;
2628 if (is_gimple_assign (stmt
))
2630 code
= gimple_assign_rhs_code (stmt
);
2634 if (!convert_mult_to_widen (stmt
, &gsi
)
2635 && convert_mult_to_fma (stmt
,
2636 gimple_assign_rhs1 (stmt
),
2637 gimple_assign_rhs2 (stmt
)))
2639 gsi_remove (&gsi
, true);
2640 release_defs (stmt
);
2647 convert_plusminus_to_widen (&gsi
, stmt
, code
);
2653 else if (is_gimple_call (stmt
)
2654 && gimple_call_lhs (stmt
))
2656 tree fndecl
= gimple_call_fndecl (stmt
);
2658 && DECL_BUILT_IN_CLASS (fndecl
) == BUILT_IN_NORMAL
)
2660 switch (DECL_FUNCTION_CODE (fndecl
))
2665 if (TREE_CODE (gimple_call_arg (stmt
, 1)) == REAL_CST
2666 && REAL_VALUES_EQUAL
2667 (TREE_REAL_CST (gimple_call_arg (stmt
, 1)),
2669 && convert_mult_to_fma (stmt
,
2670 gimple_call_arg (stmt
, 0),
2671 gimple_call_arg (stmt
, 0)))
2673 unlink_stmt_vdef (stmt
);
2674 if (gsi_remove (&gsi
, true)
2675 && gimple_purge_dead_eh_edges (bb
))
2677 release_defs (stmt
);
2690 statistics_counter_event (cfun
, "widening multiplications inserted",
2691 widen_mul_stats
.widen_mults_inserted
);
2692 statistics_counter_event (cfun
, "widening maccs inserted",
2693 widen_mul_stats
.maccs_inserted
);
2694 statistics_counter_event (cfun
, "fused multiply-adds inserted",
2695 widen_mul_stats
.fmas_inserted
);
2697 return cfg_changed
? TODO_cleanup_cfg
: 0;
2701 gate_optimize_widening_mul (void)
2703 return flag_expensive_optimizations
&& optimize
;
2706 struct gimple_opt_pass pass_optimize_widening_mul
=
2710 "widening_mul", /* name */
2711 gate_optimize_widening_mul
, /* gate */
2712 execute_optimize_widening_mul
, /* execute */
2715 0, /* static_pass_number */
2716 TV_NONE
, /* tv_id */
2717 PROP_ssa
, /* properties_required */
2718 0, /* properties_provided */
2719 0, /* properties_destroyed */
2720 0, /* todo_flags_start */
2723 | TODO_update_ssa
/* todo_flags_finish */