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"
96 #include "tree-pass.h"
97 #include "alloc-pool.h"
98 #include "basic-block.h"
100 #include "gimple-pretty-print.h"
102 /* FIXME: RTL headers have to be included here for optabs. */
103 #include "rtl.h" /* Because optabs.h wants enum rtx_code. */
104 #include "expr.h" /* Because optabs.h wants sepops. */
107 /* This structure represents one basic block that either computes a
108 division, or is a common dominator for basic block that compute a
111 /* The basic block represented by this structure. */
114 /* If non-NULL, the SSA_NAME holding the definition for a reciprocal
118 /* If non-NULL, the GIMPLE_ASSIGN for a reciprocal computation that
119 was inserted in BB. */
120 gimple recip_def_stmt;
122 /* Pointer to a list of "struct occurrence"s for blocks dominated
124 struct occurrence *children;
126 /* Pointer to the next "struct occurrence"s in the list of blocks
127 sharing a common dominator. */
128 struct occurrence *next;
130 /* The number of divisions that are in BB before compute_merit. The
131 number of divisions that are in BB or post-dominate it after
135 /* True if the basic block has a division, false if it is a common
136 dominator for basic blocks that do. If it is false and trapping
137 math is active, BB is not a candidate for inserting a reciprocal. */
138 bool bb_has_division;
143 /* Number of 1.0/X ops inserted. */
146 /* Number of 1.0/FUNC ops inserted. */
152 /* Number of cexpi calls inserted. */
158 /* Number of hand-written 32-bit bswaps found. */
161 /* Number of hand-written 64-bit bswaps found. */
167 /* Number of widening multiplication ops inserted. */
168 int widen_mults_inserted;
170 /* Number of integer multiply-and-accumulate ops inserted. */
173 /* Number of fp fused multiply-add ops inserted. */
177 /* The instance of "struct occurrence" representing the highest
178 interesting block in the dominator tree. */
179 static struct occurrence *occ_head;
181 /* Allocation pool for getting instances of "struct occurrence". */
182 static alloc_pool occ_pool;
186 /* Allocate and return a new struct occurrence for basic block BB, and
187 whose children list is headed by CHILDREN. */
188 static struct occurrence *
189 occ_new (basic_block bb, struct occurrence *children)
191 struct occurrence *occ;
193 bb->aux = occ = (struct occurrence *) pool_alloc (occ_pool);
194 memset (occ, 0, sizeof (struct occurrence));
197 occ->children = children;
202 /* Insert NEW_OCC into our subset of the dominator tree. P_HEAD points to a
203 list of "struct occurrence"s, one per basic block, having IDOM as
204 their common dominator.
206 We try to insert NEW_OCC as deep as possible in the tree, and we also
207 insert any other block that is a common dominator for BB and one
208 block already in the tree. */
211 insert_bb (struct occurrence *new_occ, basic_block idom,
212 struct occurrence **p_head)
214 struct occurrence *occ, **p_occ;
216 for (p_occ = p_head; (occ = *p_occ) != NULL; )
218 basic_block bb = new_occ->bb, occ_bb = occ->bb;
219 basic_block dom = nearest_common_dominator (CDI_DOMINATORS, occ_bb, bb);
222 /* BB dominates OCC_BB. OCC becomes NEW_OCC's child: remove OCC
225 occ->next = new_occ->children;
226 new_occ->children = occ;
228 /* Try the next block (it may as well be dominated by BB). */
231 else if (dom == occ_bb)
233 /* OCC_BB dominates BB. Tail recurse to look deeper. */
234 insert_bb (new_occ, dom, &occ->children);
238 else if (dom != idom)
240 gcc_assert (!dom->aux);
242 /* There is a dominator between IDOM and BB, add it and make
243 two children out of NEW_OCC and OCC. First, remove OCC from
249 /* None of the previous blocks has DOM as a dominator: if we tail
250 recursed, we would reexamine them uselessly. Just switch BB with
251 DOM, and go on looking for blocks dominated by DOM. */
252 new_occ = occ_new (dom, new_occ);
257 /* Nothing special, go on with the next element. */
262 /* No place was found as a child of IDOM. Make BB a sibling of IDOM. */
263 new_occ->next = *p_head;
267 /* Register that we found a division in BB. */
270 register_division_in (basic_block bb)
272 struct occurrence *occ;
274 occ = (struct occurrence *) bb->aux;
277 occ = occ_new (bb, NULL);
278 insert_bb (occ, ENTRY_BLOCK_PTR, &occ_head);
281 occ->bb_has_division = true;
282 occ->num_divisions++;
286 /* Compute the number of divisions that postdominate each block in OCC and
290 compute_merit (struct occurrence *occ)
292 struct occurrence *occ_child;
293 basic_block dom = occ->bb;
295 for (occ_child = occ->children; occ_child; occ_child = occ_child->next)
298 if (occ_child->children)
299 compute_merit (occ_child);
302 bb = single_noncomplex_succ (dom);
306 if (dominated_by_p (CDI_POST_DOMINATORS, bb, occ_child->bb))
307 occ->num_divisions += occ_child->num_divisions;
312 /* Return whether USE_STMT is a floating-point division by DEF. */
314 is_division_by (gimple use_stmt, tree def)
316 return is_gimple_assign (use_stmt)
317 && gimple_assign_rhs_code (use_stmt) == RDIV_EXPR
318 && gimple_assign_rhs2 (use_stmt) == def
319 /* Do not recognize x / x as valid division, as we are getting
320 confused later by replacing all immediate uses x in such
322 && gimple_assign_rhs1 (use_stmt) != def;
325 /* Walk the subset of the dominator tree rooted at OCC, setting the
326 RECIP_DEF field to a definition of 1.0 / DEF that can be used in
327 the given basic block. The field may be left NULL, of course,
328 if it is not possible or profitable to do the optimization.
330 DEF_BSI is an iterator pointing at the statement defining DEF.
331 If RECIP_DEF is set, a dominator already has a computation that can
335 insert_reciprocals (gimple_stmt_iterator *def_gsi, struct occurrence *occ,
336 tree def, tree recip_def, int threshold)
340 gimple_stmt_iterator gsi;
341 struct occurrence *occ_child;
344 && (occ->bb_has_division || !flag_trapping_math)
345 && occ->num_divisions >= threshold)
347 /* Make a variable with the replacement and substitute it. */
348 type = TREE_TYPE (def);
349 recip_def = make_rename_temp (type, "reciptmp");
350 new_stmt = gimple_build_assign_with_ops (RDIV_EXPR, recip_def,
351 build_one_cst (type), def);
353 if (occ->bb_has_division)
355 /* Case 1: insert before an existing division. */
356 gsi = gsi_after_labels (occ->bb);
357 while (!gsi_end_p (gsi) && !is_division_by (gsi_stmt (gsi), def))
360 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
362 else if (def_gsi && occ->bb == def_gsi->bb)
364 /* Case 2: insert right after the definition. Note that this will
365 never happen if the definition statement can throw, because in
366 that case the sole successor of the statement's basic block will
367 dominate all the uses as well. */
368 gsi_insert_after (def_gsi, new_stmt, GSI_NEW_STMT);
372 /* Case 3: insert in a basic block not containing defs/uses. */
373 gsi = gsi_after_labels (occ->bb);
374 gsi_insert_before (&gsi, new_stmt, GSI_SAME_STMT);
377 reciprocal_stats.rdivs_inserted++;
379 occ->recip_def_stmt = new_stmt;
382 occ->recip_def = recip_def;
383 for (occ_child = occ->children; occ_child; occ_child = occ_child->next)
384 insert_reciprocals (def_gsi, occ_child, def, recip_def, threshold);
388 /* Replace the division at USE_P with a multiplication by the reciprocal, if
392 replace_reciprocal (use_operand_p use_p)
394 gimple use_stmt = USE_STMT (use_p);
395 basic_block bb = gimple_bb (use_stmt);
396 struct occurrence *occ = (struct occurrence *) bb->aux;
398 if (optimize_bb_for_speed_p (bb)
399 && occ->recip_def && use_stmt != occ->recip_def_stmt)
401 gimple_assign_set_rhs_code (use_stmt, MULT_EXPR);
402 SET_USE (use_p, occ->recip_def);
403 fold_stmt_inplace (use_stmt);
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 (gimple_default_def (cfun, arg)
517 && FLOAT_TYPE_P (TREE_TYPE (arg))
518 && is_gimple_reg (arg))
519 execute_cse_reciprocals_1 (NULL, gimple_default_def (cfun, arg));
523 gimple_stmt_iterator gsi;
527 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
529 phi = gsi_stmt (gsi);
530 def = PHI_RESULT (phi);
531 if (FLOAT_TYPE_P (TREE_TYPE (def))
532 && is_gimple_reg (def))
533 execute_cse_reciprocals_1 (NULL, def);
536 for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
538 gimple stmt = gsi_stmt (gsi);
540 if (gimple_has_lhs (stmt)
541 && (def = SINGLE_SSA_TREE_OPERAND (stmt, SSA_OP_DEF)) != NULL
542 && FLOAT_TYPE_P (TREE_TYPE (def))
543 && TREE_CODE (def) == SSA_NAME)
544 execute_cse_reciprocals_1 (&gsi, def);
547 if (optimize_bb_for_size_p (bb))
550 /* Scan for a/func(b) and convert it to reciprocal a*rfunc(b). */
551 for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
553 gimple stmt = gsi_stmt (gsi);
556 if (is_gimple_assign (stmt)
557 && gimple_assign_rhs_code (stmt) == RDIV_EXPR)
559 tree arg1 = gimple_assign_rhs2 (stmt);
562 if (TREE_CODE (arg1) != SSA_NAME)
565 stmt1 = SSA_NAME_DEF_STMT (arg1);
567 if (is_gimple_call (stmt1)
568 && gimple_call_lhs (stmt1)
569 && (fndecl = gimple_call_fndecl (stmt1))
570 && (DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL
571 || DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD))
573 enum built_in_function code;
578 code = DECL_FUNCTION_CODE (fndecl);
579 md_code = DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_MD;
581 fndecl = targetm.builtin_reciprocal (code, md_code, false);
585 /* Check that all uses of the SSA name are divisions,
586 otherwise replacing the defining statement will do
589 FOR_EACH_IMM_USE_FAST (use_p, ui, arg1)
591 gimple stmt2 = USE_STMT (use_p);
592 if (is_gimple_debug (stmt2))
594 if (!is_gimple_assign (stmt2)
595 || gimple_assign_rhs_code (stmt2) != RDIV_EXPR
596 || gimple_assign_rhs1 (stmt2) == arg1
597 || gimple_assign_rhs2 (stmt2) != arg1)
606 gimple_replace_lhs (stmt1, arg1);
607 gimple_call_set_fndecl (stmt1, fndecl);
609 reciprocal_stats.rfuncs_inserted++;
611 FOR_EACH_IMM_USE_STMT (stmt, ui, arg1)
613 gimple_assign_set_rhs_code (stmt, MULT_EXPR);
614 fold_stmt_inplace (stmt);
622 statistics_counter_event (cfun, "reciprocal divs inserted",
623 reciprocal_stats.rdivs_inserted);
624 statistics_counter_event (cfun, "reciprocal functions inserted",
625 reciprocal_stats.rfuncs_inserted);
627 free_dominance_info (CDI_DOMINATORS);
628 free_dominance_info (CDI_POST_DOMINATORS);
629 free_alloc_pool (occ_pool);
633 struct gimple_opt_pass pass_cse_reciprocals =
638 gate_cse_reciprocals, /* gate */
639 execute_cse_reciprocals, /* execute */
642 0, /* static_pass_number */
644 PROP_ssa, /* properties_required */
645 0, /* properties_provided */
646 0, /* properties_destroyed */
647 0, /* todo_flags_start */
648 TODO_dump_func | TODO_update_ssa | TODO_verify_ssa
649 | TODO_verify_stmts /* todo_flags_finish */
653 /* Records an occurrence at statement USE_STMT in the vector of trees
654 STMTS if it is dominated by *TOP_BB or dominates it or this basic block
655 is not yet initialized. Returns true if the occurrence was pushed on
656 the vector. Adjusts *TOP_BB to be the basic block dominating all
657 statements in the vector. */
660 maybe_record_sincos (VEC(gimple, heap) **stmts,
661 basic_block *top_bb, gimple use_stmt)
663 basic_block use_bb = gimple_bb (use_stmt);
665 && (*top_bb == use_bb
666 || dominated_by_p (CDI_DOMINATORS, use_bb, *top_bb)))
667 VEC_safe_push (gimple, heap, *stmts, use_stmt);
669 || dominated_by_p (CDI_DOMINATORS, *top_bb, use_bb))
671 VEC_safe_push (gimple, heap, *stmts, use_stmt);
680 /* Look for sin, cos and cexpi calls with the same argument NAME and
681 create a single call to cexpi CSEing the result in this case.
682 We first walk over all immediate uses of the argument collecting
683 statements that we can CSE in a vector and in a second pass replace
684 the statement rhs with a REALPART or IMAGPART expression on the
685 result of the cexpi call we insert before the use statement that
686 dominates all other candidates. */
689 execute_cse_sincos_1 (tree name)
691 gimple_stmt_iterator gsi;
692 imm_use_iterator use_iter;
693 tree fndecl, res, type;
694 gimple def_stmt, use_stmt, stmt;
695 int seen_cos = 0, seen_sin = 0, seen_cexpi = 0;
696 VEC(gimple, heap) *stmts = NULL;
697 basic_block top_bb = NULL;
699 bool cfg_changed = false;
701 type = TREE_TYPE (name);
702 FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, name)
704 if (gimple_code (use_stmt) != GIMPLE_CALL
705 || !gimple_call_lhs (use_stmt)
706 || !(fndecl = gimple_call_fndecl (use_stmt))
707 || DECL_BUILT_IN_CLASS (fndecl) != BUILT_IN_NORMAL)
710 switch (DECL_FUNCTION_CODE (fndecl))
712 CASE_FLT_FN (BUILT_IN_COS):
713 seen_cos |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0;
716 CASE_FLT_FN (BUILT_IN_SIN):
717 seen_sin |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0;
720 CASE_FLT_FN (BUILT_IN_CEXPI):
721 seen_cexpi |= maybe_record_sincos (&stmts, &top_bb, use_stmt) ? 1 : 0;
728 if (seen_cos + seen_sin + seen_cexpi <= 1)
730 VEC_free(gimple, heap, stmts);
734 /* Simply insert cexpi at the beginning of top_bb but not earlier than
735 the name def statement. */
736 fndecl = mathfn_built_in (type, BUILT_IN_CEXPI);
739 res = create_tmp_reg (TREE_TYPE (TREE_TYPE (fndecl)), "sincostmp");
740 stmt = gimple_build_call (fndecl, 1, name);
741 res = make_ssa_name (res, stmt);
742 gimple_call_set_lhs (stmt, res);
744 def_stmt = SSA_NAME_DEF_STMT (name);
745 if (!SSA_NAME_IS_DEFAULT_DEF (name)
746 && gimple_code (def_stmt) != GIMPLE_PHI
747 && gimple_bb (def_stmt) == top_bb)
749 gsi = gsi_for_stmt (def_stmt);
750 gsi_insert_after (&gsi, stmt, GSI_SAME_STMT);
754 gsi = gsi_after_labels (top_bb);
755 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
758 sincos_stats.inserted++;
760 /* And adjust the recorded old call sites. */
761 for (i = 0; VEC_iterate(gimple, stmts, i, use_stmt); ++i)
764 fndecl = gimple_call_fndecl (use_stmt);
766 switch (DECL_FUNCTION_CODE (fndecl))
768 CASE_FLT_FN (BUILT_IN_COS):
769 rhs = fold_build1 (REALPART_EXPR, type, res);
772 CASE_FLT_FN (BUILT_IN_SIN):
773 rhs = fold_build1 (IMAGPART_EXPR, type, res);
776 CASE_FLT_FN (BUILT_IN_CEXPI):
784 /* Replace call with a copy. */
785 stmt = gimple_build_assign (gimple_call_lhs (use_stmt), rhs);
787 gsi = gsi_for_stmt (use_stmt);
788 gsi_replace (&gsi, stmt, true);
789 if (gimple_purge_dead_eh_edges (gimple_bb (stmt)))
793 VEC_free(gimple, heap, stmts);
798 /* To evaluate powi(x,n), the floating point value x raised to the
799 constant integer exponent n, we use a hybrid algorithm that
800 combines the "window method" with look-up tables. For an
801 introduction to exponentiation algorithms and "addition chains",
802 see section 4.6.3, "Evaluation of Powers" of Donald E. Knuth,
803 "Seminumerical Algorithms", Vol. 2, "The Art of Computer Programming",
804 3rd Edition, 1998, and Daniel M. Gordon, "A Survey of Fast Exponentiation
805 Methods", Journal of Algorithms, Vol. 27, pp. 129-146, 1998. */
807 /* Provide a default value for POWI_MAX_MULTS, the maximum number of
808 multiplications to inline before calling the system library's pow
809 function. powi(x,n) requires at worst 2*bits(n)-2 multiplications,
810 so this default never requires calling pow, powf or powl. */
812 #ifndef POWI_MAX_MULTS
813 #define POWI_MAX_MULTS (2*HOST_BITS_PER_WIDE_INT-2)
816 /* The size of the "optimal power tree" lookup table. All
817 exponents less than this value are simply looked up in the
818 powi_table below. This threshold is also used to size the
819 cache of pseudo registers that hold intermediate results. */
820 #define POWI_TABLE_SIZE 256
822 /* The size, in bits of the window, used in the "window method"
823 exponentiation algorithm. This is equivalent to a radix of
824 (1<<POWI_WINDOW_SIZE) in the corresponding "m-ary method". */
825 #define POWI_WINDOW_SIZE 3
827 /* The following table is an efficient representation of an
828 "optimal power tree". For each value, i, the corresponding
829 value, j, in the table states than an optimal evaluation
830 sequence for calculating pow(x,i) can be found by evaluating
831 pow(x,j)*pow(x,i-j). An optimal power tree for the first
832 100 integers is given in Knuth's "Seminumerical algorithms". */
834 static const unsigned char powi_table[POWI_TABLE_SIZE] =
836 0, 1, 1, 2, 2, 3, 3, 4, /* 0 - 7 */
837 4, 6, 5, 6, 6, 10, 7, 9, /* 8 - 15 */
838 8, 16, 9, 16, 10, 12, 11, 13, /* 16 - 23 */
839 12, 17, 13, 18, 14, 24, 15, 26, /* 24 - 31 */
840 16, 17, 17, 19, 18, 33, 19, 26, /* 32 - 39 */
841 20, 25, 21, 40, 22, 27, 23, 44, /* 40 - 47 */
842 24, 32, 25, 34, 26, 29, 27, 44, /* 48 - 55 */
843 28, 31, 29, 34, 30, 60, 31, 36, /* 56 - 63 */
844 32, 64, 33, 34, 34, 46, 35, 37, /* 64 - 71 */
845 36, 65, 37, 50, 38, 48, 39, 69, /* 72 - 79 */
846 40, 49, 41, 43, 42, 51, 43, 58, /* 80 - 87 */
847 44, 64, 45, 47, 46, 59, 47, 76, /* 88 - 95 */
848 48, 65, 49, 66, 50, 67, 51, 66, /* 96 - 103 */
849 52, 70, 53, 74, 54, 104, 55, 74, /* 104 - 111 */
850 56, 64, 57, 69, 58, 78, 59, 68, /* 112 - 119 */
851 60, 61, 61, 80, 62, 75, 63, 68, /* 120 - 127 */
852 64, 65, 65, 128, 66, 129, 67, 90, /* 128 - 135 */
853 68, 73, 69, 131, 70, 94, 71, 88, /* 136 - 143 */
854 72, 128, 73, 98, 74, 132, 75, 121, /* 144 - 151 */
855 76, 102, 77, 124, 78, 132, 79, 106, /* 152 - 159 */
856 80, 97, 81, 160, 82, 99, 83, 134, /* 160 - 167 */
857 84, 86, 85, 95, 86, 160, 87, 100, /* 168 - 175 */
858 88, 113, 89, 98, 90, 107, 91, 122, /* 176 - 183 */
859 92, 111, 93, 102, 94, 126, 95, 150, /* 184 - 191 */
860 96, 128, 97, 130, 98, 133, 99, 195, /* 192 - 199 */
861 100, 128, 101, 123, 102, 164, 103, 138, /* 200 - 207 */
862 104, 145, 105, 146, 106, 109, 107, 149, /* 208 - 215 */
863 108, 200, 109, 146, 110, 170, 111, 157, /* 216 - 223 */
864 112, 128, 113, 130, 114, 182, 115, 132, /* 224 - 231 */
865 116, 200, 117, 132, 118, 158, 119, 206, /* 232 - 239 */
866 120, 240, 121, 162, 122, 147, 123, 152, /* 240 - 247 */
867 124, 166, 125, 214, 126, 138, 127, 153, /* 248 - 255 */
871 /* Return the number of multiplications required to calculate
872 powi(x,n) where n is less than POWI_TABLE_SIZE. This is a
873 subroutine of powi_cost. CACHE is an array indicating
874 which exponents have already been calculated. */
877 powi_lookup_cost (unsigned HOST_WIDE_INT n, bool *cache)
879 /* If we've already calculated this exponent, then this evaluation
880 doesn't require any additional multiplications. */
885 return powi_lookup_cost (n - powi_table[n], cache)
886 + powi_lookup_cost (powi_table[n], cache) + 1;
889 /* Return the number of multiplications required to calculate
890 powi(x,n) for an arbitrary x, given the exponent N. This
891 function needs to be kept in sync with powi_as_mults below. */
894 powi_cost (HOST_WIDE_INT n)
896 bool cache[POWI_TABLE_SIZE];
897 unsigned HOST_WIDE_INT digit;
898 unsigned HOST_WIDE_INT val;
904 /* Ignore the reciprocal when calculating the cost. */
905 val = (n < 0) ? -n : n;
907 /* Initialize the exponent cache. */
908 memset (cache, 0, POWI_TABLE_SIZE * sizeof (bool));
913 while (val >= POWI_TABLE_SIZE)
917 digit = val & ((1 << POWI_WINDOW_SIZE) - 1);
918 result += powi_lookup_cost (digit, cache)
919 + POWI_WINDOW_SIZE + 1;
920 val >>= POWI_WINDOW_SIZE;
929 return result + powi_lookup_cost (val, cache);
932 /* Recursive subroutine of powi_as_mults. This function takes the
933 array, CACHE, of already calculated exponents and an exponent N and
934 returns a tree that corresponds to CACHE[1]**N, with type TYPE. */
937 powi_as_mults_1 (gimple_stmt_iterator *gsi, location_t loc, tree type,
938 HOST_WIDE_INT n, tree *cache, tree target)
940 tree op0, op1, ssa_target;
941 unsigned HOST_WIDE_INT digit;
944 if (n < POWI_TABLE_SIZE && cache[n])
947 ssa_target = make_ssa_name (target, NULL);
949 if (n < POWI_TABLE_SIZE)
951 cache[n] = ssa_target;
952 op0 = powi_as_mults_1 (gsi, loc, type, n - powi_table[n], cache, target);
953 op1 = powi_as_mults_1 (gsi, loc, type, powi_table[n], cache, target);
957 digit = n & ((1 << POWI_WINDOW_SIZE) - 1);
958 op0 = powi_as_mults_1 (gsi, loc, type, n - digit, cache, target);
959 op1 = powi_as_mults_1 (gsi, loc, type, digit, cache, target);
963 op0 = powi_as_mults_1 (gsi, loc, type, n >> 1, cache, target);
967 mult_stmt = gimple_build_assign_with_ops (MULT_EXPR, ssa_target, op0, op1);
968 gsi_insert_before (gsi, mult_stmt, GSI_SAME_STMT);
973 /* Convert ARG0**N to a tree of multiplications of ARG0 with itself.
974 This function needs to be kept in sync with powi_cost above. */
977 powi_as_mults (gimple_stmt_iterator *gsi, location_t loc,
978 tree arg0, HOST_WIDE_INT n)
980 tree cache[POWI_TABLE_SIZE], result, type = TREE_TYPE (arg0), target;
984 return build_real (type, dconst1);
986 memset (cache, 0, sizeof (cache));
989 target = create_tmp_var (type, "powmult");
990 add_referenced_var (target);
992 result = powi_as_mults_1 (gsi, loc, type, (n < 0) ? -n : n, cache, target);
997 /* If the original exponent was negative, reciprocate the result. */
998 target = make_ssa_name (target, NULL);
999 div_stmt = gimple_build_assign_with_ops (RDIV_EXPR, target,
1000 build_real (type, dconst1),
1002 gsi_insert_before (gsi, div_stmt, GSI_SAME_STMT);
1007 /* ARG0 and N are the two arguments to a powi builtin in GSI with
1008 location info LOC. If the arguments are appropriate, create an
1009 equivalent sequence of statements prior to GSI using an optimal
1010 number of multiplications, and return an expession holding the
1014 gimple_expand_builtin_powi (gimple_stmt_iterator *gsi, location_t loc,
1015 tree arg0, HOST_WIDE_INT n)
1017 /* Avoid largest negative number. */
1019 && ((n >= -1 && n <= 2)
1020 || (optimize_function_for_speed_p (cfun)
1021 && powi_cost (n) <= POWI_MAX_MULTS)))
1022 return powi_as_mults (gsi, loc, arg0, n);
1027 /* ARG0 and ARG1 are the two arguments to a pow builtin call in GSI
1028 with location info LOC. If possible, create an equivalent and
1029 less expensive sequence of statements prior to GSI, and return an
1030 expession holding the result. */
1033 gimple_expand_builtin_pow (gimple_stmt_iterator *gsi, location_t loc,
1034 tree arg0, tree arg1)
1036 REAL_VALUE_TYPE c, cint;
1039 /* If the exponent isn't a constant, there's nothing of interest
1041 if (TREE_CODE (arg1) != REAL_CST)
1044 /* If the exponent is equivalent to an integer, expand it into
1045 multiplies when profitable. */
1046 c = TREE_REAL_CST (arg1);
1047 n = real_to_integer (&c);
1048 real_from_integer (&cint, VOIDmode, n, n < 0 ? -1 : 0, 0);
1050 if (real_identical (&c, &cint)
1051 && ((n >= -1 && n <= 2)
1052 || (flag_unsafe_math_optimizations
1053 && optimize_insn_for_speed_p ()
1054 && powi_cost (n) <= POWI_MAX_MULTS)))
1055 return gimple_expand_builtin_powi (gsi, loc, arg0, n);
1060 /* Go through all calls to sin, cos and cexpi and call execute_cse_sincos_1
1061 on the SSA_NAME argument of each of them. Also expand powi(x,n) into
1062 an optimal number of multiplies, when n is a constant. */
1065 execute_cse_sincos (void)
1068 bool cfg_changed = false;
1070 calculate_dominance_info (CDI_DOMINATORS);
1071 memset (&sincos_stats, 0, sizeof (sincos_stats));
1075 gimple_stmt_iterator gsi;
1077 for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1079 gimple stmt = gsi_stmt (gsi);
1082 if (is_gimple_call (stmt)
1083 && gimple_call_lhs (stmt)
1084 && (fndecl = gimple_call_fndecl (stmt))
1085 && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
1087 tree arg, arg0, arg1, result;
1091 switch (DECL_FUNCTION_CODE (fndecl))
1093 CASE_FLT_FN (BUILT_IN_COS):
1094 CASE_FLT_FN (BUILT_IN_SIN):
1095 CASE_FLT_FN (BUILT_IN_CEXPI):
1096 arg = gimple_call_arg (stmt, 0);
1097 if (TREE_CODE (arg) == SSA_NAME)
1098 cfg_changed |= execute_cse_sincos_1 (arg);
1101 CASE_FLT_FN (BUILT_IN_POW):
1102 arg0 = gimple_call_arg (stmt, 0);
1103 arg1 = gimple_call_arg (stmt, 1);
1105 loc = gimple_location (stmt);
1106 result = gimple_expand_builtin_pow (&gsi, loc, arg0, arg1);
1110 tree lhs = gimple_get_lhs (stmt);
1111 gimple new_stmt = gimple_build_assign (lhs, result);
1112 gimple_set_location (new_stmt, loc);
1113 unlink_stmt_vdef (stmt);
1114 gsi_replace (&gsi, new_stmt, true);
1118 CASE_FLT_FN (BUILT_IN_POWI):
1119 arg0 = gimple_call_arg (stmt, 0);
1120 arg1 = gimple_call_arg (stmt, 1);
1121 if (!host_integerp (arg1, 0))
1124 n = TREE_INT_CST_LOW (arg1);
1125 loc = gimple_location (stmt);
1126 result = gimple_expand_builtin_powi (&gsi, loc, arg0, n);
1130 tree lhs = gimple_get_lhs (stmt);
1131 gimple new_stmt = gimple_build_assign (lhs, result);
1132 gimple_set_location (new_stmt, loc);
1133 unlink_stmt_vdef (stmt);
1134 gsi_replace (&gsi, new_stmt, true);
1144 statistics_counter_event (cfun, "sincos statements inserted",
1145 sincos_stats.inserted);
1147 free_dominance_info (CDI_DOMINATORS);
1148 return cfg_changed ? TODO_cleanup_cfg : 0;
1152 gate_cse_sincos (void)
1154 /* We no longer require either sincos or cexp, since powi expansion
1155 piggybacks on this pass. */
1159 struct gimple_opt_pass pass_cse_sincos =
1163 "sincos", /* name */
1164 gate_cse_sincos, /* gate */
1165 execute_cse_sincos, /* execute */
1168 0, /* static_pass_number */
1169 TV_NONE, /* tv_id */
1170 PROP_ssa, /* properties_required */
1171 0, /* properties_provided */
1172 0, /* properties_destroyed */
1173 0, /* todo_flags_start */
1174 TODO_dump_func | TODO_update_ssa | TODO_verify_ssa
1175 | TODO_verify_stmts /* todo_flags_finish */
1179 /* A symbolic number is used to detect byte permutation and selection
1180 patterns. Therefore the field N contains an artificial number
1181 consisting of byte size markers:
1183 0 - byte has the value 0
1184 1..size - byte contains the content of the byte
1185 number indexed with that value minus one */
1187 struct symbolic_number {
1188 unsigned HOST_WIDEST_INT n;
1192 /* Perform a SHIFT or ROTATE operation by COUNT bits on symbolic
1193 number N. Return false if the requested operation is not permitted
1194 on a symbolic number. */
1197 do_shift_rotate (enum tree_code code,
1198 struct symbolic_number *n,
1204 /* Zero out the extra bits of N in order to avoid them being shifted
1205 into the significant bits. */
1206 if (n->size < (int)sizeof (HOST_WIDEST_INT))
1207 n->n &= ((unsigned HOST_WIDEST_INT)1 << (n->size * BITS_PER_UNIT)) - 1;
1218 n->n = (n->n << count) | (n->n >> ((n->size * BITS_PER_UNIT) - count));
1221 n->n = (n->n >> count) | (n->n << ((n->size * BITS_PER_UNIT) - count));
1229 /* Perform sanity checking for the symbolic number N and the gimple
1233 verify_symbolic_number_p (struct symbolic_number *n, gimple stmt)
1237 lhs_type = gimple_expr_type (stmt);
1239 if (TREE_CODE (lhs_type) != INTEGER_TYPE)
1242 if (TYPE_PRECISION (lhs_type) != n->size * BITS_PER_UNIT)
1248 /* find_bswap_1 invokes itself recursively with N and tries to perform
1249 the operation given by the rhs of STMT on the result. If the
1250 operation could successfully be executed the function returns the
1251 tree expression of the source operand and NULL otherwise. */
1254 find_bswap_1 (gimple stmt, struct symbolic_number *n, int limit)
1256 enum tree_code code;
1257 tree rhs1, rhs2 = NULL;
1258 gimple rhs1_stmt, rhs2_stmt;
1260 enum gimple_rhs_class rhs_class;
1262 if (!limit || !is_gimple_assign (stmt))
1265 rhs1 = gimple_assign_rhs1 (stmt);
1267 if (TREE_CODE (rhs1) != SSA_NAME)
1270 code = gimple_assign_rhs_code (stmt);
1271 rhs_class = gimple_assign_rhs_class (stmt);
1272 rhs1_stmt = SSA_NAME_DEF_STMT (rhs1);
1274 if (rhs_class == GIMPLE_BINARY_RHS)
1275 rhs2 = gimple_assign_rhs2 (stmt);
1277 /* Handle unary rhs and binary rhs with integer constants as second
1280 if (rhs_class == GIMPLE_UNARY_RHS
1281 || (rhs_class == GIMPLE_BINARY_RHS
1282 && TREE_CODE (rhs2) == INTEGER_CST))
1284 if (code != BIT_AND_EXPR
1285 && code != LSHIFT_EXPR
1286 && code != RSHIFT_EXPR
1287 && code != LROTATE_EXPR
1288 && code != RROTATE_EXPR
1290 && code != CONVERT_EXPR)
1293 source_expr1 = find_bswap_1 (rhs1_stmt, n, limit - 1);
1295 /* If find_bswap_1 returned NULL STMT is a leaf node and we have
1296 to initialize the symbolic number. */
1299 /* Set up the symbolic number N by setting each byte to a
1300 value between 1 and the byte size of rhs1. The highest
1301 order byte is set to n->size and the lowest order
1303 n->size = TYPE_PRECISION (TREE_TYPE (rhs1));
1304 if (n->size % BITS_PER_UNIT != 0)
1306 n->size /= BITS_PER_UNIT;
1307 n->n = (sizeof (HOST_WIDEST_INT) < 8 ? 0 :
1308 (unsigned HOST_WIDEST_INT)0x08070605 << 32 | 0x04030201);
1310 if (n->size < (int)sizeof (HOST_WIDEST_INT))
1311 n->n &= ((unsigned HOST_WIDEST_INT)1 <<
1312 (n->size * BITS_PER_UNIT)) - 1;
1314 source_expr1 = rhs1;
1322 unsigned HOST_WIDEST_INT val = widest_int_cst_value (rhs2);
1323 unsigned HOST_WIDEST_INT tmp = val;
1325 /* Only constants masking full bytes are allowed. */
1326 for (i = 0; i < n->size; i++, tmp >>= BITS_PER_UNIT)
1327 if ((tmp & 0xff) != 0 && (tmp & 0xff) != 0xff)
1337 if (!do_shift_rotate (code, n, (int)TREE_INT_CST_LOW (rhs2)))
1344 type_size = TYPE_PRECISION (gimple_expr_type (stmt));
1345 if (type_size % BITS_PER_UNIT != 0)
1348 if (type_size / BITS_PER_UNIT < (int)(sizeof (HOST_WIDEST_INT)))
1350 /* If STMT casts to a smaller type mask out the bits not
1351 belonging to the target type. */
1352 n->n &= ((unsigned HOST_WIDEST_INT)1 << type_size) - 1;
1354 n->size = type_size / BITS_PER_UNIT;
1360 return verify_symbolic_number_p (n, stmt) ? source_expr1 : NULL;
1363 /* Handle binary rhs. */
1365 if (rhs_class == GIMPLE_BINARY_RHS)
1367 struct symbolic_number n1, n2;
1370 if (code != BIT_IOR_EXPR)
1373 if (TREE_CODE (rhs2) != SSA_NAME)
1376 rhs2_stmt = SSA_NAME_DEF_STMT (rhs2);
1381 source_expr1 = find_bswap_1 (rhs1_stmt, &n1, limit - 1);
1386 source_expr2 = find_bswap_1 (rhs2_stmt, &n2, limit - 1);
1388 if (source_expr1 != source_expr2
1389 || n1.size != n2.size)
1395 if (!verify_symbolic_number_p (n, stmt))
1402 return source_expr1;
1407 /* Check if STMT completes a bswap implementation consisting of ORs,
1408 SHIFTs and ANDs. Return the source tree expression on which the
1409 byte swap is performed and NULL if no bswap was found. */
1412 find_bswap (gimple stmt)
1414 /* The number which the find_bswap result should match in order to
1415 have a full byte swap. The number is shifted to the left according
1416 to the size of the symbolic number before using it. */
1417 unsigned HOST_WIDEST_INT cmp =
1418 sizeof (HOST_WIDEST_INT) < 8 ? 0 :
1419 (unsigned HOST_WIDEST_INT)0x01020304 << 32 | 0x05060708;
1421 struct symbolic_number n;
1424 /* The last parameter determines the depth search limit. It usually
1425 correlates directly to the number of bytes to be touched. We
1426 increase that number by one here in order to also cover signed ->
1427 unsigned conversions of the src operand as can be seen in
1429 source_expr = find_bswap_1 (stmt, &n,
1431 TYPE_SIZE_UNIT (gimple_expr_type (stmt))) + 1);
1436 /* Zero out the extra bits of N and CMP. */
1437 if (n.size < (int)sizeof (HOST_WIDEST_INT))
1439 unsigned HOST_WIDEST_INT mask =
1440 ((unsigned HOST_WIDEST_INT)1 << (n.size * BITS_PER_UNIT)) - 1;
1443 cmp >>= (sizeof (HOST_WIDEST_INT) - n.size) * BITS_PER_UNIT;
1446 /* A complete byte swap should make the symbolic number to start
1447 with the largest digit in the highest order byte. */
1454 /* Find manual byte swap implementations and turn them into a bswap
1455 builtin invokation. */
1458 execute_optimize_bswap (void)
1461 bool bswap32_p, bswap64_p;
1462 bool changed = false;
1463 tree bswap32_type = NULL_TREE, bswap64_type = NULL_TREE;
1465 if (BITS_PER_UNIT != 8)
1468 if (sizeof (HOST_WIDEST_INT) < 8)
1471 bswap32_p = (built_in_decls[BUILT_IN_BSWAP32]
1472 && optab_handler (bswap_optab, SImode) != CODE_FOR_nothing);
1473 bswap64_p = (built_in_decls[BUILT_IN_BSWAP64]
1474 && (optab_handler (bswap_optab, DImode) != CODE_FOR_nothing
1475 || (bswap32_p && word_mode == SImode)));
1477 if (!bswap32_p && !bswap64_p)
1480 /* Determine the argument type of the builtins. The code later on
1481 assumes that the return and argument type are the same. */
1484 tree fndecl = built_in_decls[BUILT_IN_BSWAP32];
1485 bswap32_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl)));
1490 tree fndecl = built_in_decls[BUILT_IN_BSWAP64];
1491 bswap64_type = TREE_VALUE (TYPE_ARG_TYPES (TREE_TYPE (fndecl)));
1494 memset (&bswap_stats, 0, sizeof (bswap_stats));
1498 gimple_stmt_iterator gsi;
1500 for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1502 gimple stmt = gsi_stmt (gsi);
1503 tree bswap_src, bswap_type;
1505 tree fndecl = NULL_TREE;
1509 if (!is_gimple_assign (stmt)
1510 || gimple_assign_rhs_code (stmt) != BIT_IOR_EXPR)
1513 type_size = TYPE_PRECISION (gimple_expr_type (stmt));
1520 fndecl = built_in_decls[BUILT_IN_BSWAP32];
1521 bswap_type = bswap32_type;
1527 fndecl = built_in_decls[BUILT_IN_BSWAP64];
1528 bswap_type = bswap64_type;
1538 bswap_src = find_bswap (stmt);
1544 if (type_size == 32)
1545 bswap_stats.found_32bit++;
1547 bswap_stats.found_64bit++;
1549 bswap_tmp = bswap_src;
1551 /* Convert the src expression if necessary. */
1552 if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp), bswap_type))
1554 gimple convert_stmt;
1556 bswap_tmp = create_tmp_var (bswap_type, "bswapsrc");
1557 add_referenced_var (bswap_tmp);
1558 bswap_tmp = make_ssa_name (bswap_tmp, NULL);
1560 convert_stmt = gimple_build_assign_with_ops (
1561 CONVERT_EXPR, bswap_tmp, bswap_src, NULL);
1562 gsi_insert_before (&gsi, convert_stmt, GSI_SAME_STMT);
1565 call = gimple_build_call (fndecl, 1, bswap_tmp);
1567 bswap_tmp = gimple_assign_lhs (stmt);
1569 /* Convert the result if necessary. */
1570 if (!useless_type_conversion_p (TREE_TYPE (bswap_tmp), bswap_type))
1572 gimple convert_stmt;
1574 bswap_tmp = create_tmp_var (bswap_type, "bswapdst");
1575 add_referenced_var (bswap_tmp);
1576 bswap_tmp = make_ssa_name (bswap_tmp, NULL);
1577 convert_stmt = gimple_build_assign_with_ops (
1578 CONVERT_EXPR, gimple_assign_lhs (stmt), bswap_tmp, NULL);
1579 gsi_insert_after (&gsi, convert_stmt, GSI_SAME_STMT);
1582 gimple_call_set_lhs (call, bswap_tmp);
1586 fprintf (dump_file, "%d bit bswap implementation found at: ",
1588 print_gimple_stmt (dump_file, stmt, 0, 0);
1591 gsi_insert_after (&gsi, call, GSI_SAME_STMT);
1592 gsi_remove (&gsi, true);
1596 statistics_counter_event (cfun, "32-bit bswap implementations found",
1597 bswap_stats.found_32bit);
1598 statistics_counter_event (cfun, "64-bit bswap implementations found",
1599 bswap_stats.found_64bit);
1601 return (changed ? TODO_dump_func | TODO_update_ssa | TODO_verify_ssa
1602 | TODO_verify_stmts : 0);
1606 gate_optimize_bswap (void)
1608 return flag_expensive_optimizations && optimize;
1611 struct gimple_opt_pass pass_optimize_bswap =
1616 gate_optimize_bswap, /* gate */
1617 execute_optimize_bswap, /* execute */
1620 0, /* static_pass_number */
1621 TV_NONE, /* tv_id */
1622 PROP_ssa, /* properties_required */
1623 0, /* properties_provided */
1624 0, /* properties_destroyed */
1625 0, /* todo_flags_start */
1626 0 /* todo_flags_finish */
1630 /* Return true if RHS is a suitable operand for a widening multiplication.
1631 There are two cases:
1633 - RHS makes some value twice as wide. Store that value in *NEW_RHS_OUT
1634 if so, and store its type in *TYPE_OUT.
1636 - RHS is an integer constant. Store that value in *NEW_RHS_OUT if so,
1637 but leave *TYPE_OUT untouched. */
1640 is_widening_mult_rhs_p (tree rhs, tree *type_out, tree *new_rhs_out)
1643 tree type, type1, rhs1;
1644 enum tree_code rhs_code;
1646 if (TREE_CODE (rhs) == SSA_NAME)
1648 type = TREE_TYPE (rhs);
1649 stmt = SSA_NAME_DEF_STMT (rhs);
1650 if (!is_gimple_assign (stmt))
1653 rhs_code = gimple_assign_rhs_code (stmt);
1654 if (TREE_CODE (type) == INTEGER_TYPE
1655 ? !CONVERT_EXPR_CODE_P (rhs_code)
1656 : rhs_code != FIXED_CONVERT_EXPR)
1659 rhs1 = gimple_assign_rhs1 (stmt);
1660 type1 = TREE_TYPE (rhs1);
1661 if (TREE_CODE (type1) != TREE_CODE (type)
1662 || TYPE_PRECISION (type1) * 2 != TYPE_PRECISION (type))
1665 *new_rhs_out = rhs1;
1670 if (TREE_CODE (rhs) == INTEGER_CST)
1680 /* Return true if STMT performs a widening multiplication. If so,
1681 store the unwidened types of the operands in *TYPE1_OUT and *TYPE2_OUT
1682 respectively. Also fill *RHS1_OUT and *RHS2_OUT such that converting
1683 those operands to types *TYPE1_OUT and *TYPE2_OUT would give the
1684 operands of the multiplication. */
1687 is_widening_mult_p (gimple stmt,
1688 tree *type1_out, tree *rhs1_out,
1689 tree *type2_out, tree *rhs2_out)
1693 type = TREE_TYPE (gimple_assign_lhs (stmt));
1694 if (TREE_CODE (type) != INTEGER_TYPE
1695 && TREE_CODE (type) != FIXED_POINT_TYPE)
1698 if (!is_widening_mult_rhs_p (gimple_assign_rhs1 (stmt), type1_out, rhs1_out))
1701 if (!is_widening_mult_rhs_p (gimple_assign_rhs2 (stmt), type2_out, rhs2_out))
1704 if (*type1_out == NULL)
1706 if (*type2_out == NULL || !int_fits_type_p (*rhs1_out, *type2_out))
1708 *type1_out = *type2_out;
1711 if (*type2_out == NULL)
1713 if (!int_fits_type_p (*rhs2_out, *type1_out))
1715 *type2_out = *type1_out;
1721 /* Process a single gimple statement STMT, which has a MULT_EXPR as
1722 its rhs, and try to convert it into a WIDEN_MULT_EXPR. The return
1723 value is true iff we converted the statement. */
1726 convert_mult_to_widen (gimple stmt)
1728 tree lhs, rhs1, rhs2, type, type1, type2;
1729 enum insn_code handler;
1731 lhs = gimple_assign_lhs (stmt);
1732 type = TREE_TYPE (lhs);
1733 if (TREE_CODE (type) != INTEGER_TYPE)
1736 if (!is_widening_mult_p (stmt, &type1, &rhs1, &type2, &rhs2))
1739 if (TYPE_UNSIGNED (type1) && TYPE_UNSIGNED (type2))
1740 handler = optab_handler (umul_widen_optab, TYPE_MODE (type));
1741 else if (!TYPE_UNSIGNED (type1) && !TYPE_UNSIGNED (type2))
1742 handler = optab_handler (smul_widen_optab, TYPE_MODE (type));
1744 handler = optab_handler (usmul_widen_optab, TYPE_MODE (type));
1746 if (handler == CODE_FOR_nothing)
1749 gimple_assign_set_rhs1 (stmt, fold_convert (type1, rhs1));
1750 gimple_assign_set_rhs2 (stmt, fold_convert (type2, rhs2));
1751 gimple_assign_set_rhs_code (stmt, WIDEN_MULT_EXPR);
1753 widen_mul_stats.widen_mults_inserted++;
1757 /* Process a single gimple statement STMT, which is found at the
1758 iterator GSI and has a either a PLUS_EXPR or a MINUS_EXPR as its
1759 rhs (given by CODE), and try to convert it into a
1760 WIDEN_MULT_PLUS_EXPR or a WIDEN_MULT_MINUS_EXPR. The return value
1761 is true iff we converted the statement. */
1764 convert_plusminus_to_widen (gimple_stmt_iterator *gsi, gimple stmt,
1765 enum tree_code code)
1767 gimple rhs1_stmt = NULL, rhs2_stmt = NULL;
1768 tree type, type1, type2;
1769 tree lhs, rhs1, rhs2, mult_rhs1, mult_rhs2, add_rhs;
1770 enum tree_code rhs1_code = ERROR_MARK, rhs2_code = ERROR_MARK;
1772 enum tree_code wmult_code;
1774 lhs = gimple_assign_lhs (stmt);
1775 type = TREE_TYPE (lhs);
1776 if (TREE_CODE (type) != INTEGER_TYPE
1777 && TREE_CODE (type) != FIXED_POINT_TYPE)
1780 if (code == MINUS_EXPR)
1781 wmult_code = WIDEN_MULT_MINUS_EXPR;
1783 wmult_code = WIDEN_MULT_PLUS_EXPR;
1785 rhs1 = gimple_assign_rhs1 (stmt);
1786 rhs2 = gimple_assign_rhs2 (stmt);
1788 if (TREE_CODE (rhs1) == SSA_NAME)
1790 rhs1_stmt = SSA_NAME_DEF_STMT (rhs1);
1791 if (is_gimple_assign (rhs1_stmt))
1792 rhs1_code = gimple_assign_rhs_code (rhs1_stmt);
1797 if (TREE_CODE (rhs2) == SSA_NAME)
1799 rhs2_stmt = SSA_NAME_DEF_STMT (rhs2);
1800 if (is_gimple_assign (rhs2_stmt))
1801 rhs2_code = gimple_assign_rhs_code (rhs2_stmt);
1806 if (code == PLUS_EXPR && rhs1_code == MULT_EXPR)
1808 if (!is_widening_mult_p (rhs1_stmt, &type1, &mult_rhs1,
1809 &type2, &mult_rhs2))
1813 else if (rhs2_code == MULT_EXPR)
1815 if (!is_widening_mult_p (rhs2_stmt, &type1, &mult_rhs1,
1816 &type2, &mult_rhs2))
1820 else if (code == PLUS_EXPR && rhs1_code == WIDEN_MULT_EXPR)
1822 mult_rhs1 = gimple_assign_rhs1 (rhs1_stmt);
1823 mult_rhs2 = gimple_assign_rhs2 (rhs1_stmt);
1824 type1 = TREE_TYPE (mult_rhs1);
1825 type2 = TREE_TYPE (mult_rhs2);
1828 else if (rhs2_code == WIDEN_MULT_EXPR)
1830 mult_rhs1 = gimple_assign_rhs1 (rhs2_stmt);
1831 mult_rhs2 = gimple_assign_rhs2 (rhs2_stmt);
1832 type1 = TREE_TYPE (mult_rhs1);
1833 type2 = TREE_TYPE (mult_rhs2);
1839 if (TYPE_UNSIGNED (type1) != TYPE_UNSIGNED (type2))
1842 /* Verify that the machine can perform a widening multiply
1843 accumulate in this mode/signedness combination, otherwise
1844 this transformation is likely to pessimize code. */
1845 this_optab = optab_for_tree_code (wmult_code, type1, optab_default);
1846 if (optab_handler (this_optab, TYPE_MODE (type)) == CODE_FOR_nothing)
1849 /* ??? May need some type verification here? */
1851 gimple_assign_set_rhs_with_ops_1 (gsi, wmult_code,
1852 fold_convert (type1, mult_rhs1),
1853 fold_convert (type2, mult_rhs2),
1855 update_stmt (gsi_stmt (*gsi));
1856 widen_mul_stats.maccs_inserted++;
1860 /* Combine the multiplication at MUL_STMT with operands MULOP1 and MULOP2
1861 with uses in additions and subtractions to form fused multiply-add
1862 operations. Returns true if successful and MUL_STMT should be removed. */
1865 convert_mult_to_fma (gimple mul_stmt, tree op1, tree op2)
1867 tree mul_result = gimple_get_lhs (mul_stmt);
1868 tree type = TREE_TYPE (mul_result);
1869 gimple use_stmt, neguse_stmt, fma_stmt;
1870 use_operand_p use_p;
1871 imm_use_iterator imm_iter;
1873 if (FLOAT_TYPE_P (type)
1874 && flag_fp_contract_mode == FP_CONTRACT_OFF)
1877 /* We don't want to do bitfield reduction ops. */
1878 if (INTEGRAL_TYPE_P (type)
1879 && (TYPE_PRECISION (type)
1880 != GET_MODE_PRECISION (TYPE_MODE (type))))
1883 /* If the target doesn't support it, don't generate it. We assume that
1884 if fma isn't available then fms, fnma or fnms are not either. */
1885 if (optab_handler (fma_optab, TYPE_MODE (type)) == CODE_FOR_nothing)
1888 /* Make sure that the multiplication statement becomes dead after
1889 the transformation, thus that all uses are transformed to FMAs.
1890 This means we assume that an FMA operation has the same cost
1892 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, mul_result)
1894 enum tree_code use_code;
1895 tree result = mul_result;
1896 bool negate_p = false;
1898 use_stmt = USE_STMT (use_p);
1900 if (is_gimple_debug (use_stmt))
1903 /* For now restrict this operations to single basic blocks. In theory
1904 we would want to support sinking the multiplication in
1910 to form a fma in the then block and sink the multiplication to the
1912 if (gimple_bb (use_stmt) != gimple_bb (mul_stmt))
1915 if (!is_gimple_assign (use_stmt))
1918 use_code = gimple_assign_rhs_code (use_stmt);
1920 /* A negate on the multiplication leads to FNMA. */
1921 if (use_code == NEGATE_EXPR)
1926 result = gimple_assign_lhs (use_stmt);
1928 /* Make sure the negate statement becomes dead with this
1929 single transformation. */
1930 if (!single_imm_use (gimple_assign_lhs (use_stmt),
1931 &use_p, &neguse_stmt))
1934 /* Make sure the multiplication isn't also used on that stmt. */
1935 FOR_EACH_SSA_TREE_OPERAND (use, neguse_stmt, iter, SSA_OP_USE)
1936 if (use == mul_result)
1940 use_stmt = neguse_stmt;
1941 if (gimple_bb (use_stmt) != gimple_bb (mul_stmt))
1943 if (!is_gimple_assign (use_stmt))
1946 use_code = gimple_assign_rhs_code (use_stmt);
1953 if (gimple_assign_rhs2 (use_stmt) == result)
1954 negate_p = !negate_p;
1959 /* FMA can only be formed from PLUS and MINUS. */
1963 /* We can't handle a * b + a * b. */
1964 if (gimple_assign_rhs1 (use_stmt) == gimple_assign_rhs2 (use_stmt))
1967 /* While it is possible to validate whether or not the exact form
1968 that we've recognized is available in the backend, the assumption
1969 is that the transformation is never a loss. For instance, suppose
1970 the target only has the plain FMA pattern available. Consider
1971 a*b-c -> fma(a,b,-c): we've exchanged MUL+SUB for FMA+NEG, which
1972 is still two operations. Consider -(a*b)-c -> fma(-a,b,-c): we
1973 still have 3 operations, but in the FMA form the two NEGs are
1974 independant and could be run in parallel. */
1977 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, mul_result)
1979 gimple_stmt_iterator gsi = gsi_for_stmt (use_stmt);
1980 enum tree_code use_code;
1981 tree addop, mulop1 = op1, result = mul_result;
1982 bool negate_p = false;
1984 if (is_gimple_debug (use_stmt))
1987 use_code = gimple_assign_rhs_code (use_stmt);
1988 if (use_code == NEGATE_EXPR)
1990 result = gimple_assign_lhs (use_stmt);
1991 single_imm_use (gimple_assign_lhs (use_stmt), &use_p, &neguse_stmt);
1992 gsi_remove (&gsi, true);
1993 release_defs (use_stmt);
1995 use_stmt = neguse_stmt;
1996 gsi = gsi_for_stmt (use_stmt);
1997 use_code = gimple_assign_rhs_code (use_stmt);
2001 if (gimple_assign_rhs1 (use_stmt) == result)
2003 addop = gimple_assign_rhs2 (use_stmt);
2004 /* a * b - c -> a * b + (-c) */
2005 if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR)
2006 addop = force_gimple_operand_gsi (&gsi,
2007 build1 (NEGATE_EXPR,
2009 true, NULL_TREE, true,
2014 addop = gimple_assign_rhs1 (use_stmt);
2015 /* a - b * c -> (-b) * c + a */
2016 if (gimple_assign_rhs_code (use_stmt) == MINUS_EXPR)
2017 negate_p = !negate_p;
2021 mulop1 = force_gimple_operand_gsi (&gsi,
2022 build1 (NEGATE_EXPR,
2024 true, NULL_TREE, true,
2027 fma_stmt = gimple_build_assign_with_ops3 (FMA_EXPR,
2028 gimple_assign_lhs (use_stmt),
2031 gsi_replace (&gsi, fma_stmt, true);
2032 widen_mul_stats.fmas_inserted++;
2038 /* Find integer multiplications where the operands are extended from
2039 smaller types, and replace the MULT_EXPR with a WIDEN_MULT_EXPR
2040 where appropriate. */
2043 execute_optimize_widening_mul (void)
2046 bool cfg_changed = false;
2048 memset (&widen_mul_stats, 0, sizeof (widen_mul_stats));
2052 gimple_stmt_iterator gsi;
2054 for (gsi = gsi_after_labels (bb); !gsi_end_p (gsi);)
2056 gimple stmt = gsi_stmt (gsi);
2057 enum tree_code code;
2059 if (is_gimple_assign (stmt))
2061 code = gimple_assign_rhs_code (stmt);
2065 if (!convert_mult_to_widen (stmt)
2066 && convert_mult_to_fma (stmt,
2067 gimple_assign_rhs1 (stmt),
2068 gimple_assign_rhs2 (stmt)))
2070 gsi_remove (&gsi, true);
2071 release_defs (stmt);
2078 convert_plusminus_to_widen (&gsi, stmt, code);
2084 else if (is_gimple_call (stmt)
2085 && gimple_call_lhs (stmt))
2087 tree fndecl = gimple_call_fndecl (stmt);
2089 && DECL_BUILT_IN_CLASS (fndecl) == BUILT_IN_NORMAL)
2091 switch (DECL_FUNCTION_CODE (fndecl))
2096 if (TREE_CODE (gimple_call_arg (stmt, 1)) == REAL_CST
2097 && REAL_VALUES_EQUAL
2098 (TREE_REAL_CST (gimple_call_arg (stmt, 1)),
2100 && convert_mult_to_fma (stmt,
2101 gimple_call_arg (stmt, 0),
2102 gimple_call_arg (stmt, 0)))
2104 unlink_stmt_vdef (stmt);
2105 gsi_remove (&gsi, true);
2106 release_defs (stmt);
2107 if (gimple_purge_dead_eh_edges (bb))
2121 statistics_counter_event (cfun, "widening multiplications inserted",
2122 widen_mul_stats.widen_mults_inserted);
2123 statistics_counter_event (cfun, "widening maccs inserted",
2124 widen_mul_stats.maccs_inserted);
2125 statistics_counter_event (cfun, "fused multiply-adds inserted",
2126 widen_mul_stats.fmas_inserted);
2128 return cfg_changed ? TODO_cleanup_cfg : 0;
2132 gate_optimize_widening_mul (void)
2134 return flag_expensive_optimizations && optimize;
2137 struct gimple_opt_pass pass_optimize_widening_mul =
2141 "widening_mul", /* name */
2142 gate_optimize_widening_mul, /* gate */
2143 execute_optimize_widening_mul, /* execute */
2146 0, /* static_pass_number */
2147 TV_NONE, /* tv_id */
2148 PROP_ssa, /* properties_required */
2149 0, /* properties_provided */
2150 0, /* properties_destroyed */
2151 0, /* todo_flags_start */
2155 | TODO_update_ssa /* todo_flags_finish */