]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/gimple-ssa-strength-reduction.c
re PR rtl-optimization/68376 (wrong code at -O1 and above on x86_64-linux-gnu)
[thirdparty/gcc.git] / gcc / gimple-ssa-strength-reduction.c
CommitLineData
f9453c07 1/* Straight-line strength reduction.
5624e564 2 Copyright (C) 2012-2015 Free Software Foundation, Inc.
f9453c07
BS
3 Contributed by Bill Schmidt, IBM <wschmidt@linux.ibm.com>
4
5This file is part of GCC.
6
7GCC is free software; you can redistribute it and/or modify it under
8the terms of the GNU General Public License as published by the Free
9Software Foundation; either version 3, or (at your option) any later
10version.
11
12GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13WARRANTY; without even the implied warranty of MERCHANTABILITY or
14FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15for more details.
16
17You should have received a copy of the GNU General Public License
18along with GCC; see the file COPYING3. If not see
19<http://www.gnu.org/licenses/>. */
20
21/* There are many algorithms for performing strength reduction on
22 loops. This is not one of them. IVOPTS handles strength reduction
23 of induction variables just fine. This pass is intended to pick
24 up the crumbs it leaves behind, by considering opportunities for
25 strength reduction along dominator paths.
26
9b92d12b
BS
27 Strength reduction addresses explicit multiplies, and certain
28 multiplies implicit in addressing expressions. It would also be
29 possible to apply strength reduction to divisions and modulos,
30 but such opportunities are relatively uncommon.
f9453c07
BS
31
32 Strength reduction is also currently restricted to integer operations.
33 If desired, it could be extended to floating-point operations under
34 control of something like -funsafe-math-optimizations. */
35
36#include "config.h"
37#include "system.h"
38#include "coretypes.h"
c7131fb2 39#include "backend.h"
957060b5 40#include "rtl.h"
40e23961 41#include "tree.h"
c7131fb2 42#include "gimple.h"
957060b5
AM
43#include "cfghooks.h"
44#include "tree-pass.h"
c7131fb2 45#include "ssa.h"
957060b5 46#include "expmed.h"
957060b5 47#include "gimple-pretty-print.h"
40e23961 48#include "fold-const.h"
5be5c238 49#include "gimple-iterator.h"
18f429e2 50#include "gimplify-me.h"
d8a2d370 51#include "stor-layout.h"
f9453c07 52#include "cfgloop.h"
442b4905 53#include "tree-cfg.h"
f9453c07 54#include "domwalk.h"
ccdbfe93 55#include "params.h"
4484a35a 56#include "tree-ssa-address.h"
96d75a2c 57#include "tree-affine.h"
9b2b7279 58#include "builtins.h"
f9453c07
BS
59\f
60/* Information about a strength reduction candidate. Each statement
61 in the candidate table represents an expression of one of the
62 following forms (the special case of CAND_REF will be described
63 later):
64
65 (CAND_MULT) S1: X = (B + i) * S
66 (CAND_ADD) S1: X = B + (i * S)
67
68 Here X and B are SSA names, i is an integer constant, and S is
69 either an SSA name or a constant. We call B the "base," i the
70 "index", and S the "stride."
71
72 Any statement S0 that dominates S1 and is of the form:
73
74 (CAND_MULT) S0: Y = (B + i') * S
75 (CAND_ADD) S0: Y = B + (i' * S)
76
77 is called a "basis" for S1. In both cases, S1 may be replaced by
78
79 S1': X = Y + (i - i') * S,
80
81 where (i - i') * S is folded to the extent possible.
82
83 All gimple statements are visited in dominator order, and each
84 statement that may contribute to one of the forms of S1 above is
85 given at least one entry in the candidate table. Such statements
86 include addition, pointer addition, subtraction, multiplication,
87 negation, copies, and nontrivial type casts. If a statement may
88 represent more than one expression of the forms of S1 above,
89 multiple "interpretations" are stored in the table and chained
90 together. Examples:
91
92 * An add of two SSA names may treat either operand as the base.
93 * A multiply of two SSA names, likewise.
94 * A copy or cast may be thought of as either a CAND_MULT with
95 i = 0 and S = 1, or as a CAND_ADD with i = 0 or S = 0.
96
97 Candidate records are allocated from an obstack. They are addressed
98 both from a hash table keyed on S1, and from a vector of candidate
99 pointers arranged in predominator order.
100
101 Opportunity note
102 ----------------
103 Currently we don't recognize:
104
105 S0: Y = (S * i') - B
106 S1: X = (S * i) - B
107
108 as a strength reduction opportunity, even though this S1 would
109 also be replaceable by the S1' above. This can be added if it
2749c8f6
BS
110 comes up in practice.
111
112 Strength reduction in addressing
113 --------------------------------
114 There is another kind of candidate known as CAND_REF. A CAND_REF
115 describes a statement containing a memory reference having
116 complex addressing that might benefit from strength reduction.
117 Specifically, we are interested in references for which
118 get_inner_reference returns a base address, offset, and bitpos as
119 follows:
120
121 base: MEM_REF (T1, C1)
122 offset: MULT_EXPR (PLUS_EXPR (T2, C2), C3)
123 bitpos: C4 * BITS_PER_UNIT
124
125 Here T1 and T2 are arbitrary trees, and C1, C2, C3, C4 are
126 arbitrary integer constants. Note that C2 may be zero, in which
127 case the offset will be MULT_EXPR (T2, C3).
128
129 When this pattern is recognized, the original memory reference
130 can be replaced with:
131
132 MEM_REF (POINTER_PLUS_EXPR (T1, MULT_EXPR (T2, C3)),
133 C1 + (C2 * C3) + C4)
134
135 which distributes the multiply to allow constant folding. When
136 two or more addressing expressions can be represented by MEM_REFs
137 of this form, differing only in the constants C1, C2, and C4,
138 making this substitution produces more efficient addressing during
139 the RTL phases. When there are not at least two expressions with
140 the same values of T1, T2, and C3, there is nothing to be gained
141 by the replacement.
142
143 Strength reduction of CAND_REFs uses the same infrastructure as
144 that used by CAND_MULTs and CAND_ADDs. We record T1 in the base (B)
145 field, MULT_EXPR (T2, C3) in the stride (S) field, and
146 C1 + (C2 * C3) + C4 in the index (i) field. A basis for a CAND_REF
147 is thus another CAND_REF with the same B and S values. When at
148 least two CAND_REFs are chained together using the basis relation,
149 each of them is replaced as above, resulting in improved code
9b92d12b
BS
150 generation for addressing.
151
152 Conditional candidates
153 ======================
154
155 Conditional candidates are best illustrated with an example.
156 Consider the code sequence:
157
158 (1) x_0 = ...;
159 (2) a_0 = x_0 * 5; MULT (B: x_0; i: 0; S: 5)
160 if (...)
161 (3) x_1 = x_0 + 1; ADD (B: x_0, i: 1; S: 1)
162 (4) x_2 = PHI <x_0, x_1>; PHI (B: x_0, i: 0, S: 1)
163 (5) x_3 = x_2 + 1; ADD (B: x_2, i: 1, S: 1)
164 (6) a_1 = x_3 * 5; MULT (B: x_2, i: 1; S: 5)
165
166 Here strength reduction is complicated by the uncertain value of x_2.
167 A legitimate transformation is:
168
169 (1) x_0 = ...;
170 (2) a_0 = x_0 * 5;
171 if (...)
172 {
173 (3) [x_1 = x_0 + 1;]
174 (3a) t_1 = a_0 + 5;
175 }
176 (4) [x_2 = PHI <x_0, x_1>;]
177 (4a) t_2 = PHI <a_0, t_1>;
178 (5) [x_3 = x_2 + 1;]
179 (6r) a_1 = t_2 + 5;
180
181 where the bracketed instructions may go dead.
182
183 To recognize this opportunity, we have to observe that statement (6)
184 has a "hidden basis" (2). The hidden basis is unlike a normal basis
185 in that the statement and the hidden basis have different base SSA
186 names (x_2 and x_0, respectively). The relationship is established
187 when a statement's base name (x_2) is defined by a phi statement (4),
188 each argument of which (x_0, x_1) has an identical "derived base name."
189 If the argument is defined by a candidate (as x_1 is by (3)) that is a
190 CAND_ADD having a stride of 1, the derived base name of the argument is
191 the base name of the candidate (x_0). Otherwise, the argument itself
192 is its derived base name (as is the case with argument x_0).
193
194 The hidden basis for statement (6) is the nearest dominating candidate
195 whose base name is the derived base name (x_0) of the feeding phi (4),
196 and whose stride is identical to that of the statement. We can then
197 create the new "phi basis" (4a) and feeding adds along incoming arcs (3a),
198 allowing the final replacement of (6) by the strength-reduced (6r).
199
200 To facilitate this, a new kind of candidate (CAND_PHI) is introduced.
201 A CAND_PHI is not a candidate for replacement, but is maintained in the
202 candidate table to ease discovery of hidden bases. Any phi statement
203 whose arguments share a common derived base name is entered into the
204 table with the derived base name, an (arbitrary) index of zero, and a
205 stride of 1. A statement with a hidden basis can then be detected by
206 simply looking up its feeding phi definition in the candidate table,
207 extracting the derived base name, and searching for a basis in the
208 usual manner after substituting the derived base name.
209
210 Note that the transformation is only valid when the original phi and
211 the statements that define the phi's arguments are all at the same
212 position in the loop hierarchy. */
f9453c07
BS
213
214
215/* Index into the candidate vector, offset by 1. VECs are zero-based,
216 while cand_idx's are one-based, with zero indicating null. */
217typedef unsigned cand_idx;
218
219/* The kind of candidate. */
220enum cand_kind
221{
222 CAND_MULT,
2749c8f6 223 CAND_ADD,
9b92d12b
BS
224 CAND_REF,
225 CAND_PHI
f9453c07
BS
226};
227
228struct slsr_cand_d
229{
230 /* The candidate statement S1. */
355fe088 231 gimple *cand_stmt;
f9453c07 232
3cfd4469
BS
233 /* The base expression B: often an SSA name, but not always. */
234 tree base_expr;
f9453c07
BS
235
236 /* The stride S. */
237 tree stride;
238
239 /* The index constant i. */
807e902e 240 widest_int index;
f9453c07 241
3cfd4469 242 /* The type of the candidate. This is normally the type of base_expr,
f9453c07 243 but casts may have occurred when combining feeding instructions.
2749c8f6
BS
244 A candidate can only be a basis for candidates of the same final type.
245 (For CAND_REFs, this is the type to be used for operand 1 of the
246 replacement MEM_REF.) */
f9453c07
BS
247 tree cand_type;
248
249 /* The kind of candidate (CAND_MULT, etc.). */
250 enum cand_kind kind;
251
252 /* Index of this candidate in the candidate vector. */
253 cand_idx cand_num;
254
255 /* Index of the next candidate record for the same statement.
256 A statement may be useful in more than one way (e.g., due to
257 commutativity). So we can have multiple "interpretations"
258 of a statement. */
259 cand_idx next_interp;
260
261 /* Index of the basis statement S0, if any, in the candidate vector. */
262 cand_idx basis;
263
264 /* First candidate for which this candidate is a basis, if one exists. */
265 cand_idx dependent;
266
267 /* Next candidate having the same basis as this one. */
268 cand_idx sibling;
269
9b92d12b
BS
270 /* If this is a conditional candidate, the CAND_PHI candidate
271 that defines the base SSA name B. */
272 cand_idx def_phi;
f9453c07
BS
273
274 /* Savings that can be expected from eliminating dead code if this
275 candidate is replaced. */
276 int dead_savings;
277};
278
279typedef struct slsr_cand_d slsr_cand, *slsr_cand_t;
280typedef const struct slsr_cand_d *const_slsr_cand_t;
281
282/* Pointers to candidates are chained together as part of a mapping
3cfd4469 283 from base expressions to the candidates that use them. */
f9453c07
BS
284
285struct cand_chain_d
286{
3cfd4469
BS
287 /* Base expression for the chain of candidates: often, but not
288 always, an SSA name. */
289 tree base_expr;
f9453c07
BS
290
291 /* Pointer to a candidate. */
292 slsr_cand_t cand;
293
294 /* Chain pointer. */
295 struct cand_chain_d *next;
296
297};
298
299typedef struct cand_chain_d cand_chain, *cand_chain_t;
300typedef const struct cand_chain_d *const_cand_chain_t;
301
88ca9ea1
BS
302/* Information about a unique "increment" associated with candidates
303 having an SSA name for a stride. An increment is the difference
304 between the index of the candidate and the index of its basis,
305 i.e., (i - i') as discussed in the module commentary.
306
307 When we are not going to generate address arithmetic we treat
308 increments that differ only in sign as the same, allowing sharing
309 of the cost of initializers. The absolute value of the increment
310 is stored in the incr_info. */
311
312struct incr_info_d
313{
314 /* The increment that relates a candidate to its basis. */
807e902e 315 widest_int incr;
88ca9ea1
BS
316
317 /* How many times the increment occurs in the candidate tree. */
318 unsigned count;
319
320 /* Cost of replacing candidates using this increment. Negative and
321 zero costs indicate replacement should be performed. */
322 int cost;
323
324 /* If this increment is profitable but is not -1, 0, or 1, it requires
325 an initializer T_0 = stride * incr to be found or introduced in the
326 nearest common dominator of all candidates. This field holds T_0
327 for subsequent use. */
328 tree initializer;
329
330 /* If the initializer was found to already exist, this is the block
331 where it was found. */
332 basic_block init_bb;
333};
334
335typedef struct incr_info_d incr_info, *incr_info_t;
336
f9453c07
BS
337/* Candidates are maintained in a vector. If candidate X dominates
338 candidate Y, then X appears before Y in the vector; but the
339 converse does not necessarily hold. */
9771b263 340static vec<slsr_cand_t> cand_vec;
f9453c07
BS
341
342enum cost_consts
343{
344 COST_NEUTRAL = 0,
345 COST_INFINITE = 1000
346};
347
9b92d12b
BS
348enum stride_status
349{
350 UNKNOWN_STRIDE = 0,
351 KNOWN_STRIDE = 1
352};
353
354enum phi_adjust_status
355{
356 NOT_PHI_ADJUST = 0,
357 PHI_ADJUST = 1
358};
359
360enum count_phis_status
361{
362 DONT_COUNT_PHIS = 0,
363 COUNT_PHIS = 1
364};
365
f9453c07 366/* Pointer map embodying a mapping from statements to candidates. */
355fe088 367static hash_map<gimple *, slsr_cand_t> *stmt_cand_map;
f9453c07
BS
368
369/* Obstack for candidates. */
370static struct obstack cand_obstack;
371
f9453c07
BS
372/* Obstack for candidate chains. */
373static struct obstack chain_obstack;
88ca9ea1
BS
374
375/* An array INCR_VEC of incr_infos is used during analysis of related
376 candidates having an SSA name for a stride. INCR_VEC_LEN describes
7bf55a70
BS
377 its current length. MAX_INCR_VEC_LEN is used to avoid costly
378 pathological cases. */
88ca9ea1
BS
379static incr_info_t incr_vec;
380static unsigned incr_vec_len;
7bf55a70 381const int MAX_INCR_VEC_LEN = 16;
88ca9ea1
BS
382
383/* For a chain of candidates with unknown stride, indicates whether or not
384 we must generate pointer arithmetic when replacing statements. */
385static bool address_arithmetic_p;
9b92d12b
BS
386
387/* Forward function declarations. */
388static slsr_cand_t base_cand_from_table (tree);
a7a7d10e 389static tree introduce_cast_before_cand (slsr_cand_t, tree, tree);
0916f876 390static bool legal_cast_p_1 (tree, tree);
f9453c07
BS
391\f
392/* Produce a pointer to the IDX'th candidate in the candidate vector. */
393
394static slsr_cand_t
395lookup_cand (cand_idx idx)
396{
9771b263 397 return cand_vec[idx - 1];
f9453c07
BS
398}
399
4a8fb1a1 400/* Helper for hashing a candidate chain header. */
2749c8f6 401
8d67ee55 402struct cand_chain_hasher : nofree_ptr_hash <cand_chain>
2749c8f6 403{
67f58944
TS
404 static inline hashval_t hash (const cand_chain *);
405 static inline bool equal (const cand_chain *, const cand_chain *);
4a8fb1a1 406};
2749c8f6 407
4a8fb1a1 408inline hashval_t
67f58944 409cand_chain_hasher::hash (const cand_chain *p)
2749c8f6 410{
4a8fb1a1
LC
411 tree base_expr = p->base_expr;
412 return iterative_hash_expr (base_expr, 0);
2749c8f6
BS
413}
414
4a8fb1a1 415inline bool
67f58944 416cand_chain_hasher::equal (const cand_chain *chain1, const cand_chain *chain2)
2749c8f6 417{
3cfd4469 418 return operand_equal_p (chain1->base_expr, chain2->base_expr, 0);
2749c8f6 419}
4a8fb1a1
LC
420
421/* Hash table embodying a mapping from base exprs to chains of candidates. */
c203e8a7 422static hash_table<cand_chain_hasher> *base_cand_map;
2749c8f6 423\f
96d75a2c 424/* Pointer map used by tree_to_aff_combination_expand. */
39c8aaa4 425static hash_map<tree, name_expansion *> *name_expansions;
96d75a2c 426/* Pointer map embodying a mapping from bases to alternative bases. */
b787e7a2 427static hash_map<tree, tree> *alt_base_map;
96d75a2c
YZ
428
429/* Given BASE, use the tree affine combiniation facilities to
430 find the underlying tree expression for BASE, with any
8fde427f
YZ
431 immediate offset excluded.
432
433 N.B. we should eliminate this backtracking with better forward
434 analysis in a future release. */
96d75a2c
YZ
435
436static tree
437get_alternative_base (tree base)
438{
b787e7a2 439 tree *result = alt_base_map->get (base);
96d75a2c
YZ
440
441 if (result == NULL)
442 {
443 tree expr;
444 aff_tree aff;
445
446 tree_to_aff_combination_expand (base, TREE_TYPE (base),
447 &aff, &name_expansions);
807e902e 448 aff.offset = 0;
96d75a2c
YZ
449 expr = aff_combination_to_tree (&aff);
450
b787e7a2 451 gcc_assert (!alt_base_map->put (base, base == expr ? NULL : expr));
96d75a2c 452
b787e7a2 453 return expr == base ? NULL : expr;
96d75a2c
YZ
454 }
455
456 return *result;
457}
458
9b92d12b 459/* Look in the candidate table for a CAND_PHI that defines BASE and
8b28cf47 460 return it if found; otherwise return NULL. */
f9453c07 461
9b92d12b 462static cand_idx
8b28cf47 463find_phi_def (tree base)
9b92d12b
BS
464{
465 slsr_cand_t c;
466
467 if (TREE_CODE (base) != SSA_NAME)
468 return 0;
29105868 469
9b92d12b
BS
470 c = base_cand_from_table (base);
471
472 if (!c || c->kind != CAND_PHI)
473 return 0;
474
475 return c->cand_num;
476}
477
478/* Helper routine for find_basis_for_candidate. May be called twice:
96d75a2c
YZ
479 once for the candidate's base expr, and optionally again either for
480 the candidate's phi definition or for a CAND_REF's alternative base
481 expression. */
9b92d12b
BS
482
483static slsr_cand_t
484find_basis_for_base_expr (slsr_cand_t c, tree base_expr)
f9453c07 485{
2749c8f6 486 cand_chain mapping_key;
f9453c07
BS
487 cand_chain_t chain;
488 slsr_cand_t basis = NULL;
489
ccdbfe93
BS
490 // Limit potential of N^2 behavior for long candidate chains.
491 int iters = 0;
492 int max_iters = PARAM_VALUE (PARAM_MAX_SLSR_CANDIDATE_SCAN);
493
9b92d12b 494 mapping_key.base_expr = base_expr;
c203e8a7 495 chain = base_cand_map->find (&mapping_key);
f9453c07 496
ccdbfe93 497 for (; chain && iters < max_iters; chain = chain->next, ++iters)
f9453c07
BS
498 {
499 slsr_cand_t one_basis = chain->cand;
500
501 if (one_basis->kind != c->kind
69e1a1a3 502 || one_basis->cand_stmt == c->cand_stmt
f9453c07
BS
503 || !operand_equal_p (one_basis->stride, c->stride, 0)
504 || !types_compatible_p (one_basis->cand_type, c->cand_type)
505 || !dominated_by_p (CDI_DOMINATORS,
506 gimple_bb (c->cand_stmt),
507 gimple_bb (one_basis->cand_stmt)))
508 continue;
509
510 if (!basis || basis->cand_num < one_basis->cand_num)
511 basis = one_basis;
512 }
513
9b92d12b
BS
514 return basis;
515}
516
517/* Use the base expr from candidate C to look for possible candidates
518 that can serve as a basis for C. Each potential basis must also
519 appear in a block that dominates the candidate statement and have
520 the same stride and type. If more than one possible basis exists,
521 the one with highest index in the vector is chosen; this will be
522 the most immediately dominating basis. */
523
524static int
525find_basis_for_candidate (slsr_cand_t c)
526{
527 slsr_cand_t basis = find_basis_for_base_expr (c, c->base_expr);
528
529 /* If a candidate doesn't have a basis using its base expression,
530 it may have a basis hidden by one or more intervening phis. */
531 if (!basis && c->def_phi)
532 {
533 basic_block basis_bb, phi_bb;
534 slsr_cand_t phi_cand = lookup_cand (c->def_phi);
535 basis = find_basis_for_base_expr (c, phi_cand->base_expr);
536
537 if (basis)
538 {
539 /* A hidden basis must dominate the phi-definition of the
540 candidate's base name. */
541 phi_bb = gimple_bb (phi_cand->cand_stmt);
542 basis_bb = gimple_bb (basis->cand_stmt);
543
544 if (phi_bb == basis_bb
545 || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb))
546 {
547 basis = NULL;
548 c->basis = 0;
549 }
550
551 /* If we found a hidden basis, estimate additional dead-code
552 savings if the phi and its feeding statements can be removed. */
553 if (basis && has_single_use (gimple_phi_result (phi_cand->cand_stmt)))
554 c->dead_savings += phi_cand->dead_savings;
555 }
556 }
557
8fde427f 558 if (flag_expensive_optimizations && !basis && c->kind == CAND_REF)
96d75a2c
YZ
559 {
560 tree alt_base_expr = get_alternative_base (c->base_expr);
561 if (alt_base_expr)
562 basis = find_basis_for_base_expr (c, alt_base_expr);
563 }
564
f9453c07
BS
565 if (basis)
566 {
567 c->sibling = basis->dependent;
568 basis->dependent = c->cand_num;
569 return basis->cand_num;
570 }
571
572 return 0;
573}
574
96d75a2c
YZ
575/* Record a mapping from BASE to C, indicating that C may potentially serve
576 as a basis using that base expression. BASE may be the same as
577 C->BASE_EXPR; alternatively BASE can be a different tree that share the
578 underlining expression of C->BASE_EXPR. */
f9453c07
BS
579
580static void
96d75a2c 581record_potential_basis (slsr_cand_t c, tree base)
f9453c07 582{
2749c8f6 583 cand_chain_t node;
4a8fb1a1 584 cand_chain **slot;
f9453c07 585
96d75a2c
YZ
586 gcc_assert (base);
587
f9453c07 588 node = (cand_chain_t) obstack_alloc (&chain_obstack, sizeof (cand_chain));
96d75a2c 589 node->base_expr = base;
f9453c07
BS
590 node->cand = c;
591 node->next = NULL;
c203e8a7 592 slot = base_cand_map->find_slot (node, INSERT);
f9453c07 593
2749c8f6 594 if (*slot)
f9453c07 595 {
2749c8f6 596 cand_chain_t head = (cand_chain_t) (*slot);
f9453c07
BS
597 node->next = head->next;
598 head->next = node;
599 }
600 else
2749c8f6 601 *slot = node;
f9453c07
BS
602}
603
604/* Allocate storage for a new candidate and initialize its fields.
96d75a2c
YZ
605 Attempt to find a basis for the candidate.
606
607 For CAND_REF, an alternative base may also be recorded and used
608 to find a basis. This helps cases where the expression hidden
609 behind BASE (which is usually an SSA_NAME) has immediate offset,
610 e.g.
611
612 a2[i][j] = 1;
613 a2[i + 20][j] = 2; */
f9453c07
BS
614
615static slsr_cand_t
355fe088 616alloc_cand_and_find_basis (enum cand_kind kind, gimple *gs, tree base,
807e902e 617 const widest_int &index, tree stride, tree ctype,
f9453c07
BS
618 unsigned savings)
619{
620 slsr_cand_t c = (slsr_cand_t) obstack_alloc (&cand_obstack,
621 sizeof (slsr_cand));
622 c->cand_stmt = gs;
3cfd4469 623 c->base_expr = base;
f9453c07
BS
624 c->stride = stride;
625 c->index = index;
626 c->cand_type = ctype;
627 c->kind = kind;
9771b263 628 c->cand_num = cand_vec.length () + 1;
f9453c07
BS
629 c->next_interp = 0;
630 c->dependent = 0;
631 c->sibling = 0;
8b28cf47 632 c->def_phi = kind == CAND_MULT ? find_phi_def (base) : 0;
f9453c07
BS
633 c->dead_savings = savings;
634
9771b263 635 cand_vec.safe_push (c);
9b92d12b
BS
636
637 if (kind == CAND_PHI)
638 c->basis = 0;
639 else
640 c->basis = find_basis_for_candidate (c);
641
96d75a2c 642 record_potential_basis (c, base);
8fde427f 643 if (flag_expensive_optimizations && kind == CAND_REF)
96d75a2c
YZ
644 {
645 tree alt_base = get_alternative_base (base);
646 if (alt_base)
647 record_potential_basis (c, alt_base);
648 }
f9453c07
BS
649
650 return c;
651}
652
653/* Determine the target cost of statement GS when compiling according
654 to SPEED. */
655
656static int
355fe088 657stmt_cost (gimple *gs, bool speed)
f9453c07
BS
658{
659 tree lhs, rhs1, rhs2;
ef4bddc2 660 machine_mode lhs_mode;
f9453c07
BS
661
662 gcc_assert (is_gimple_assign (gs));
663 lhs = gimple_assign_lhs (gs);
664 rhs1 = gimple_assign_rhs1 (gs);
665 lhs_mode = TYPE_MODE (TREE_TYPE (lhs));
666
667 switch (gimple_assign_rhs_code (gs))
668 {
669 case MULT_EXPR:
670 rhs2 = gimple_assign_rhs2 (gs);
671
9541ffee 672 if (tree_fits_shwi_p (rhs2))
eb1ce453 673 return mult_by_coeff_cost (tree_to_shwi (rhs2), lhs_mode, speed);
f9453c07
BS
674
675 gcc_assert (TREE_CODE (rhs1) != INTEGER_CST);
5322d07e 676 return mul_cost (speed, lhs_mode);
f9453c07
BS
677
678 case PLUS_EXPR:
679 case POINTER_PLUS_EXPR:
680 case MINUS_EXPR:
5322d07e 681 return add_cost (speed, lhs_mode);
f9453c07
BS
682
683 case NEGATE_EXPR:
5322d07e 684 return neg_cost (speed, lhs_mode);
f9453c07 685
d822570f 686 CASE_CONVERT:
6dd8f4bb 687 return convert_cost (lhs_mode, TYPE_MODE (TREE_TYPE (rhs1)), speed);
f9453c07
BS
688
689 /* Note that we don't assign costs to copies that in most cases
690 will go away. */
691 default:
692 ;
693 }
694
695 gcc_unreachable ();
696 return 0;
697}
698
699/* Look up the defining statement for BASE_IN and return a pointer
700 to its candidate in the candidate table, if any; otherwise NULL.
701 Only CAND_ADD and CAND_MULT candidates are returned. */
702
703static slsr_cand_t
704base_cand_from_table (tree base_in)
705{
706 slsr_cand_t *result;
707
355fe088 708 gimple *def = SSA_NAME_DEF_STMT (base_in);
f9453c07
BS
709 if (!def)
710 return (slsr_cand_t) NULL;
711
b787e7a2 712 result = stmt_cand_map->get (def);
2749c8f6
BS
713
714 if (result && (*result)->kind != CAND_REF)
715 return *result;
f9453c07 716
2749c8f6 717 return (slsr_cand_t) NULL;
f9453c07
BS
718}
719
720/* Add an entry to the statement-to-candidate mapping. */
721
722static void
355fe088 723add_cand_for_stmt (gimple *gs, slsr_cand_t c)
f9453c07 724{
b787e7a2 725 gcc_assert (!stmt_cand_map->put (gs, c));
f9453c07
BS
726}
727\f
9b92d12b
BS
728/* Given PHI which contains a phi statement, determine whether it
729 satisfies all the requirements of a phi candidate. If so, create
730 a candidate. Note that a CAND_PHI never has a basis itself, but
731 is used to help find a basis for subsequent candidates. */
732
733static void
538dd0b7 734slsr_process_phi (gphi *phi, bool speed)
9b92d12b
BS
735{
736 unsigned i;
737 tree arg0_base = NULL_TREE, base_type;
738 slsr_cand_t c;
739 struct loop *cand_loop = gimple_bb (phi)->loop_father;
740 unsigned savings = 0;
741
742 /* A CAND_PHI requires each of its arguments to have the same
743 derived base name. (See the module header commentary for a
744 definition of derived base names.) Furthermore, all feeding
745 definitions must be in the same position in the loop hierarchy
746 as PHI. */
747
748 for (i = 0; i < gimple_phi_num_args (phi); i++)
749 {
750 slsr_cand_t arg_cand;
751 tree arg = gimple_phi_arg_def (phi, i);
752 tree derived_base_name = NULL_TREE;
355fe088 753 gimple *arg_stmt = NULL;
9b92d12b
BS
754 basic_block arg_bb = NULL;
755
756 if (TREE_CODE (arg) != SSA_NAME)
757 return;
758
759 arg_cand = base_cand_from_table (arg);
760
761 if (arg_cand)
762 {
763 while (arg_cand->kind != CAND_ADD && arg_cand->kind != CAND_PHI)
764 {
765 if (!arg_cand->next_interp)
766 return;
767
768 arg_cand = lookup_cand (arg_cand->next_interp);
769 }
770
771 if (!integer_onep (arg_cand->stride))
772 return;
773
774 derived_base_name = arg_cand->base_expr;
775 arg_stmt = arg_cand->cand_stmt;
776 arg_bb = gimple_bb (arg_stmt);
777
778 /* Gather potential dead code savings if the phi statement
779 can be removed later on. */
780 if (has_single_use (arg))
781 {
782 if (gimple_code (arg_stmt) == GIMPLE_PHI)
783 savings += arg_cand->dead_savings;
784 else
785 savings += stmt_cost (arg_stmt, speed);
786 }
787 }
788 else
789 {
790 derived_base_name = arg;
791
792 if (SSA_NAME_IS_DEFAULT_DEF (arg))
fefa31b5 793 arg_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun));
9b92d12b
BS
794 else
795 gimple_bb (SSA_NAME_DEF_STMT (arg));
796 }
797
798 if (!arg_bb || arg_bb->loop_father != cand_loop)
799 return;
800
801 if (i == 0)
802 arg0_base = derived_base_name;
803 else if (!operand_equal_p (derived_base_name, arg0_base, 0))
804 return;
805 }
806
807 /* Create the candidate. "alloc_cand_and_find_basis" is named
808 misleadingly for this case, as no basis will be sought for a
809 CAND_PHI. */
810 base_type = TREE_TYPE (arg0_base);
811
807e902e
KZ
812 c = alloc_cand_and_find_basis (CAND_PHI, phi, arg0_base,
813 0, integer_one_node, base_type, savings);
9b92d12b
BS
814
815 /* Add the candidate to the statement-candidate mapping. */
816 add_cand_for_stmt (phi, c);
817}
818
c068654b
BC
819/* Given PBASE which is a pointer to tree, look up the defining
820 statement for it and check whether the candidate is in the
821 form of:
822
823 X = B + (1 * S), S is integer constant
824 X = B + (i * S), S is integer one
825
826 If so, set PBASE to the candidate's base_expr and return double
827 int (i * S).
828 Otherwise, just return double int zero. */
829
807e902e 830static widest_int
c068654b
BC
831backtrace_base_for_ref (tree *pbase)
832{
833 tree base_in = *pbase;
834 slsr_cand_t base_cand;
835
836 STRIP_NOPS (base_in);
0916f876
YZ
837
838 /* Strip off widening conversion(s) to handle cases where
839 e.g. 'B' is widened from an 'int' in order to calculate
840 a 64-bit address. */
841 if (CONVERT_EXPR_P (base_in)
842 && legal_cast_p_1 (base_in, TREE_OPERAND (base_in, 0)))
843 base_in = get_unwidened (base_in, NULL_TREE);
844
c068654b 845 if (TREE_CODE (base_in) != SSA_NAME)
807e902e 846 return 0;
c068654b
BC
847
848 base_cand = base_cand_from_table (base_in);
849
850 while (base_cand && base_cand->kind != CAND_PHI)
851 {
852 if (base_cand->kind == CAND_ADD
807e902e 853 && base_cand->index == 1
c068654b
BC
854 && TREE_CODE (base_cand->stride) == INTEGER_CST)
855 {
856 /* X = B + (1 * S), S is integer constant. */
857 *pbase = base_cand->base_expr;
807e902e 858 return wi::to_widest (base_cand->stride);
c068654b
BC
859 }
860 else if (base_cand->kind == CAND_ADD
861 && TREE_CODE (base_cand->stride) == INTEGER_CST
862 && integer_onep (base_cand->stride))
1d2151c6 863 {
c068654b
BC
864 /* X = B + (i * S), S is integer one. */
865 *pbase = base_cand->base_expr;
866 return base_cand->index;
867 }
868
869 if (base_cand->next_interp)
870 base_cand = lookup_cand (base_cand->next_interp);
871 else
872 base_cand = NULL;
873 }
874
807e902e 875 return 0;
c068654b
BC
876}
877
2749c8f6
BS
878/* Look for the following pattern:
879
880 *PBASE: MEM_REF (T1, C1)
881
882 *POFFSET: MULT_EXPR (T2, C3) [C2 is zero]
883 or
884 MULT_EXPR (PLUS_EXPR (T2, C2), C3)
885 or
886 MULT_EXPR (MINUS_EXPR (T2, -C2), C3)
887
888 *PINDEX: C4 * BITS_PER_UNIT
889
890 If not present, leave the input values unchanged and return FALSE.
891 Otherwise, modify the input values as follows and return TRUE:
892
893 *PBASE: T1
894 *POFFSET: MULT_EXPR (T2, C3)
c068654b
BC
895 *PINDEX: C1 + (C2 * C3) + C4
896
897 When T2 is recorded by a CAND_ADD in the form of (T2' + C5), it
898 will be further restructured to:
899
900 *PBASE: T1
901 *POFFSET: MULT_EXPR (T2', C3)
902 *PINDEX: C1 + (C2 * C3) + C4 + (C5 * C3) */
2749c8f6
BS
903
904static bool
807e902e 905restructure_reference (tree *pbase, tree *poffset, widest_int *pindex,
2749c8f6
BS
906 tree *ptype)
907{
908 tree base = *pbase, offset = *poffset;
807e902e
KZ
909 widest_int index = *pindex;
910 tree mult_op0, t1, t2, type;
911 widest_int c1, c2, c3, c4, c5;
2749c8f6
BS
912
913 if (!base
914 || !offset
915 || TREE_CODE (base) != MEM_REF
916 || TREE_CODE (offset) != MULT_EXPR
917 || TREE_CODE (TREE_OPERAND (offset, 1)) != INTEGER_CST
807e902e 918 || wi::umod_floor (index, BITS_PER_UNIT) != 0)
2749c8f6
BS
919 return false;
920
921 t1 = TREE_OPERAND (base, 0);
807e902e 922 c1 = widest_int::from (mem_ref_offset (base), SIGNED);
2749c8f6
BS
923 type = TREE_TYPE (TREE_OPERAND (base, 1));
924
925 mult_op0 = TREE_OPERAND (offset, 0);
807e902e 926 c3 = wi::to_widest (TREE_OPERAND (offset, 1));
2749c8f6
BS
927
928 if (TREE_CODE (mult_op0) == PLUS_EXPR)
929
930 if (TREE_CODE (TREE_OPERAND (mult_op0, 1)) == INTEGER_CST)
931 {
932 t2 = TREE_OPERAND (mult_op0, 0);
807e902e 933 c2 = wi::to_widest (TREE_OPERAND (mult_op0, 1));
2749c8f6
BS
934 }
935 else
936 return false;
937
938 else if (TREE_CODE (mult_op0) == MINUS_EXPR)
939
940 if (TREE_CODE (TREE_OPERAND (mult_op0, 1)) == INTEGER_CST)
941 {
942 t2 = TREE_OPERAND (mult_op0, 0);
807e902e 943 c2 = -wi::to_widest (TREE_OPERAND (mult_op0, 1));
2749c8f6
BS
944 }
945 else
946 return false;
947
948 else
949 {
950 t2 = mult_op0;
807e902e 951 c2 = 0;
2749c8f6
BS
952 }
953
807e902e 954 c4 = wi::lrshift (index, LOG2_BITS_PER_UNIT);
c068654b 955 c5 = backtrace_base_for_ref (&t2);
2749c8f6
BS
956
957 *pbase = t1;
c068654b 958 *poffset = fold_build2 (MULT_EXPR, sizetype, fold_convert (sizetype, t2),
807e902e 959 wide_int_to_tree (sizetype, c3));
c068654b 960 *pindex = c1 + c2 * c3 + c4 + c5 * c3;
2749c8f6
BS
961 *ptype = type;
962
963 return true;
964}
965
966/* Given GS which contains a data reference, create a CAND_REF entry in
967 the candidate table and attempt to find a basis. */
968
969static void
355fe088 970slsr_process_ref (gimple *gs)
2749c8f6
BS
971{
972 tree ref_expr, base, offset, type;
973 HOST_WIDE_INT bitsize, bitpos;
ef4bddc2 974 machine_mode mode;
ee45a32d 975 int unsignedp, reversep, volatilep;
2749c8f6
BS
976 slsr_cand_t c;
977
978 if (gimple_vdef (gs))
979 ref_expr = gimple_assign_lhs (gs);
980 else
981 ref_expr = gimple_assign_rhs1 (gs);
982
983 if (!handled_component_p (ref_expr)
984 || TREE_CODE (ref_expr) == BIT_FIELD_REF
985 || (TREE_CODE (ref_expr) == COMPONENT_REF
986 && DECL_BIT_FIELD (TREE_OPERAND (ref_expr, 1))))
987 return;
988
989 base = get_inner_reference (ref_expr, &bitsize, &bitpos, &offset, &mode,
ee45a32d
EB
990 &unsignedp, &reversep, &volatilep, false);
991 if (reversep)
992 return;
807e902e 993 widest_int index = bitpos;
2749c8f6
BS
994
995 if (!restructure_reference (&base, &offset, &index, &type))
996 return;
997
998 c = alloc_cand_and_find_basis (CAND_REF, gs, base, index, offset,
999 type, 0);
1000
1001 /* Add the candidate to the statement-candidate mapping. */
1002 add_cand_for_stmt (gs, c);
1003}
1004
f9453c07
BS
1005/* Create a candidate entry for a statement GS, where GS multiplies
1006 two SSA names BASE_IN and STRIDE_IN. Propagate any known information
1007 about the two SSA names into the new candidate. Return the new
1008 candidate. */
1009
1010static slsr_cand_t
355fe088 1011create_mul_ssa_cand (gimple *gs, tree base_in, tree stride_in, bool speed)
f9453c07
BS
1012{
1013 tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE;
807e902e 1014 widest_int index;
f9453c07
BS
1015 unsigned savings = 0;
1016 slsr_cand_t c;
1017 slsr_cand_t base_cand = base_cand_from_table (base_in);
1018
1019 /* Look at all interpretations of the base candidate, if necessary,
1020 to find information to propagate into this candidate. */
9b92d12b 1021 while (base_cand && !base && base_cand->kind != CAND_PHI)
f9453c07
BS
1022 {
1023
9b92d12b 1024 if (base_cand->kind == CAND_MULT && integer_onep (base_cand->stride))
f9453c07
BS
1025 {
1026 /* Y = (B + i') * 1
1027 X = Y * Z
1028 ================
1029 X = (B + i') * Z */
3cfd4469 1030 base = base_cand->base_expr;
f9453c07
BS
1031 index = base_cand->index;
1032 stride = stride_in;
1033 ctype = base_cand->cand_type;
1034 if (has_single_use (base_in))
1035 savings = (base_cand->dead_savings
1036 + stmt_cost (base_cand->cand_stmt, speed));
1037 }
1038 else if (base_cand->kind == CAND_ADD
1039 && TREE_CODE (base_cand->stride) == INTEGER_CST)
1040 {
1041 /* Y = B + (i' * S), S constant
1042 X = Y * Z
1043 ============================
1044 X = B + ((i' * S) * Z) */
3cfd4469 1045 base = base_cand->base_expr;
807e902e 1046 index = base_cand->index * wi::to_widest (base_cand->stride);
f9453c07
BS
1047 stride = stride_in;
1048 ctype = base_cand->cand_type;
1049 if (has_single_use (base_in))
1050 savings = (base_cand->dead_savings
1051 + stmt_cost (base_cand->cand_stmt, speed));
1052 }
1053
1054 if (base_cand->next_interp)
1055 base_cand = lookup_cand (base_cand->next_interp);
1056 else
1057 base_cand = NULL;
1058 }
1059
1060 if (!base)
1061 {
1062 /* No interpretations had anything useful to propagate, so
1063 produce X = (Y + 0) * Z. */
1064 base = base_in;
807e902e 1065 index = 0;
f9453c07 1066 stride = stride_in;
b2ec94d4 1067 ctype = TREE_TYPE (base_in);
f9453c07
BS
1068 }
1069
1070 c = alloc_cand_and_find_basis (CAND_MULT, gs, base, index, stride,
1071 ctype, savings);
1072 return c;
1073}
1074
1075/* Create a candidate entry for a statement GS, where GS multiplies
1076 SSA name BASE_IN by constant STRIDE_IN. Propagate any known
1077 information about BASE_IN into the new candidate. Return the new
1078 candidate. */
1079
1080static slsr_cand_t
355fe088 1081create_mul_imm_cand (gimple *gs, tree base_in, tree stride_in, bool speed)
f9453c07
BS
1082{
1083 tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE;
807e902e 1084 widest_int index, temp;
f9453c07
BS
1085 unsigned savings = 0;
1086 slsr_cand_t c;
1087 slsr_cand_t base_cand = base_cand_from_table (base_in);
1088
1089 /* Look at all interpretations of the base candidate, if necessary,
1090 to find information to propagate into this candidate. */
9b92d12b 1091 while (base_cand && !base && base_cand->kind != CAND_PHI)
f9453c07
BS
1092 {
1093 if (base_cand->kind == CAND_MULT
1094 && TREE_CODE (base_cand->stride) == INTEGER_CST)
1095 {
1096 /* Y = (B + i') * S, S constant
1097 X = Y * c
1098 ============================
1099 X = (B + i') * (S * c) */
807e902e
KZ
1100 temp = wi::to_widest (base_cand->stride) * wi::to_widest (stride_in);
1101 if (wi::fits_to_tree_p (temp, TREE_TYPE (stride_in)))
61ba7329
BS
1102 {
1103 base = base_cand->base_expr;
1104 index = base_cand->index;
807e902e 1105 stride = wide_int_to_tree (TREE_TYPE (stride_in), temp);
61ba7329
BS
1106 ctype = base_cand->cand_type;
1107 if (has_single_use (base_in))
1108 savings = (base_cand->dead_savings
1109 + stmt_cost (base_cand->cand_stmt, speed));
1110 }
f9453c07 1111 }
9b92d12b 1112 else if (base_cand->kind == CAND_ADD && integer_onep (base_cand->stride))
f9453c07
BS
1113 {
1114 /* Y = B + (i' * 1)
1115 X = Y * c
1116 ===========================
1117 X = (B + i') * c */
3cfd4469 1118 base = base_cand->base_expr;
f9453c07
BS
1119 index = base_cand->index;
1120 stride = stride_in;
1121 ctype = base_cand->cand_type;
1122 if (has_single_use (base_in))
1123 savings = (base_cand->dead_savings
1124 + stmt_cost (base_cand->cand_stmt, speed));
1125 }
1126 else if (base_cand->kind == CAND_ADD
807e902e 1127 && base_cand->index == 1
f9453c07
BS
1128 && TREE_CODE (base_cand->stride) == INTEGER_CST)
1129 {
1130 /* Y = B + (1 * S), S constant
1131 X = Y * c
1132 ===========================
1133 X = (B + S) * c */
3cfd4469 1134 base = base_cand->base_expr;
807e902e 1135 index = wi::to_widest (base_cand->stride);
f9453c07
BS
1136 stride = stride_in;
1137 ctype = base_cand->cand_type;
1138 if (has_single_use (base_in))
1139 savings = (base_cand->dead_savings
1140 + stmt_cost (base_cand->cand_stmt, speed));
1141 }
1142
1143 if (base_cand->next_interp)
1144 base_cand = lookup_cand (base_cand->next_interp);
1145 else
1146 base_cand = NULL;
1147 }
1148
1149 if (!base)
1150 {
1151 /* No interpretations had anything useful to propagate, so
1152 produce X = (Y + 0) * c. */
1153 base = base_in;
807e902e 1154 index = 0;
f9453c07 1155 stride = stride_in;
b2ec94d4 1156 ctype = TREE_TYPE (base_in);
f9453c07
BS
1157 }
1158
1159 c = alloc_cand_and_find_basis (CAND_MULT, gs, base, index, stride,
1160 ctype, savings);
1161 return c;
1162}
1163
1164/* Given GS which is a multiply of scalar integers, make an appropriate
1165 entry in the candidate table. If this is a multiply of two SSA names,
1166 create two CAND_MULT interpretations and attempt to find a basis for
1167 each of them. Otherwise, create a single CAND_MULT and attempt to
1168 find a basis. */
1169
1170static void
355fe088 1171slsr_process_mul (gimple *gs, tree rhs1, tree rhs2, bool speed)
f9453c07
BS
1172{
1173 slsr_cand_t c, c2;
1174
1175 /* If this is a multiply of an SSA name with itself, it is highly
1176 unlikely that we will get a strength reduction opportunity, so
1177 don't record it as a candidate. This simplifies the logic for
1178 finding a basis, so if this is removed that must be considered. */
1179 if (rhs1 == rhs2)
1180 return;
1181
1182 if (TREE_CODE (rhs2) == SSA_NAME)
1183 {
1184 /* Record an interpretation of this statement in the candidate table
3cfd4469 1185 assuming RHS1 is the base expression and RHS2 is the stride. */
f9453c07
BS
1186 c = create_mul_ssa_cand (gs, rhs1, rhs2, speed);
1187
1188 /* Add the first interpretation to the statement-candidate mapping. */
1189 add_cand_for_stmt (gs, c);
1190
1191 /* Record another interpretation of this statement assuming RHS1
3cfd4469 1192 is the stride and RHS2 is the base expression. */
f9453c07
BS
1193 c2 = create_mul_ssa_cand (gs, rhs2, rhs1, speed);
1194 c->next_interp = c2->cand_num;
1195 }
1196 else
1197 {
1198 /* Record an interpretation for the multiply-immediate. */
1199 c = create_mul_imm_cand (gs, rhs1, rhs2, speed);
1200
1201 /* Add the interpretation to the statement-candidate mapping. */
1202 add_cand_for_stmt (gs, c);
1203 }
1204}
1205
1206/* Create a candidate entry for a statement GS, where GS adds two
1207 SSA names BASE_IN and ADDEND_IN if SUBTRACT_P is false, and
1208 subtracts ADDEND_IN from BASE_IN otherwise. Propagate any known
1209 information about the two SSA names into the new candidate.
1210 Return the new candidate. */
1211
1212static slsr_cand_t
355fe088 1213create_add_ssa_cand (gimple *gs, tree base_in, tree addend_in,
f9453c07
BS
1214 bool subtract_p, bool speed)
1215{
1216 tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL;
807e902e 1217 widest_int index;
f9453c07
BS
1218 unsigned savings = 0;
1219 slsr_cand_t c;
1220 slsr_cand_t base_cand = base_cand_from_table (base_in);
1221 slsr_cand_t addend_cand = base_cand_from_table (addend_in);
1222
1223 /* The most useful transformation is a multiply-immediate feeding
1224 an add or subtract. Look for that first. */
9b92d12b 1225 while (addend_cand && !base && addend_cand->kind != CAND_PHI)
f9453c07
BS
1226 {
1227 if (addend_cand->kind == CAND_MULT
807e902e 1228 && addend_cand->index == 0
f9453c07
BS
1229 && TREE_CODE (addend_cand->stride) == INTEGER_CST)
1230 {
1231 /* Z = (B + 0) * S, S constant
1232 X = Y +/- Z
1233 ===========================
1234 X = Y + ((+/-1 * S) * B) */
1235 base = base_in;
807e902e 1236 index = wi::to_widest (addend_cand->stride);
f9453c07 1237 if (subtract_p)
27bcd47c 1238 index = -index;
3cfd4469 1239 stride = addend_cand->base_expr;
b2ec94d4 1240 ctype = TREE_TYPE (base_in);
f9453c07
BS
1241 if (has_single_use (addend_in))
1242 savings = (addend_cand->dead_savings
1243 + stmt_cost (addend_cand->cand_stmt, speed));
1244 }
1245
1246 if (addend_cand->next_interp)
1247 addend_cand = lookup_cand (addend_cand->next_interp);
1248 else
1249 addend_cand = NULL;
1250 }
1251
9b92d12b 1252 while (base_cand && !base && base_cand->kind != CAND_PHI)
f9453c07
BS
1253 {
1254 if (base_cand->kind == CAND_ADD
807e902e 1255 && (base_cand->index == 0
f9453c07
BS
1256 || operand_equal_p (base_cand->stride,
1257 integer_zero_node, 0)))
1258 {
1259 /* Y = B + (i' * S), i' * S = 0
1260 X = Y +/- Z
1261 ============================
1262 X = B + (+/-1 * Z) */
3cfd4469 1263 base = base_cand->base_expr;
807e902e 1264 index = subtract_p ? -1 : 1;
f9453c07
BS
1265 stride = addend_in;
1266 ctype = base_cand->cand_type;
1267 if (has_single_use (base_in))
1268 savings = (base_cand->dead_savings
1269 + stmt_cost (base_cand->cand_stmt, speed));
1270 }
1271 else if (subtract_p)
1272 {
1273 slsr_cand_t subtrahend_cand = base_cand_from_table (addend_in);
1274
9b92d12b 1275 while (subtrahend_cand && !base && subtrahend_cand->kind != CAND_PHI)
f9453c07
BS
1276 {
1277 if (subtrahend_cand->kind == CAND_MULT
807e902e 1278 && subtrahend_cand->index == 0
f9453c07
BS
1279 && TREE_CODE (subtrahend_cand->stride) == INTEGER_CST)
1280 {
1281 /* Z = (B + 0) * S, S constant
1282 X = Y - Z
1283 ===========================
1284 Value: X = Y + ((-1 * S) * B) */
1285 base = base_in;
807e902e 1286 index = wi::to_widest (subtrahend_cand->stride);
27bcd47c 1287 index = -index;
3cfd4469 1288 stride = subtrahend_cand->base_expr;
b2ec94d4 1289 ctype = TREE_TYPE (base_in);
f9453c07
BS
1290 if (has_single_use (addend_in))
1291 savings = (subtrahend_cand->dead_savings
1292 + stmt_cost (subtrahend_cand->cand_stmt, speed));
1293 }
1294
1295 if (subtrahend_cand->next_interp)
1296 subtrahend_cand = lookup_cand (subtrahend_cand->next_interp);
1297 else
1298 subtrahend_cand = NULL;
1299 }
1300 }
1301
1302 if (base_cand->next_interp)
1303 base_cand = lookup_cand (base_cand->next_interp);
1304 else
1305 base_cand = NULL;
1306 }
1307
1308 if (!base)
1309 {
1310 /* No interpretations had anything useful to propagate, so
1311 produce X = Y + (1 * Z). */
1312 base = base_in;
807e902e 1313 index = subtract_p ? -1 : 1;
f9453c07 1314 stride = addend_in;
b2ec94d4 1315 ctype = TREE_TYPE (base_in);
f9453c07
BS
1316 }
1317
1318 c = alloc_cand_and_find_basis (CAND_ADD, gs, base, index, stride,
1319 ctype, savings);
1320 return c;
1321}
1322
1323/* Create a candidate entry for a statement GS, where GS adds SSA
1324 name BASE_IN to constant INDEX_IN. Propagate any known information
1325 about BASE_IN into the new candidate. Return the new candidate. */
1326
1327static slsr_cand_t
355fe088 1328create_add_imm_cand (gimple *gs, tree base_in, const widest_int &index_in,
807e902e 1329 bool speed)
f9453c07
BS
1330{
1331 enum cand_kind kind = CAND_ADD;
1332 tree base = NULL_TREE, stride = NULL_TREE, ctype = NULL_TREE;
807e902e 1333 widest_int index, multiple;
f9453c07
BS
1334 unsigned savings = 0;
1335 slsr_cand_t c;
1336 slsr_cand_t base_cand = base_cand_from_table (base_in);
1337
9b92d12b 1338 while (base_cand && !base && base_cand->kind != CAND_PHI)
f9453c07 1339 {
807e902e 1340 signop sign = TYPE_SIGN (TREE_TYPE (base_cand->stride));
f9453c07
BS
1341
1342 if (TREE_CODE (base_cand->stride) == INTEGER_CST
807e902e
KZ
1343 && wi::multiple_of_p (index_in, wi::to_widest (base_cand->stride),
1344 sign, &multiple))
f9453c07
BS
1345 {
1346 /* Y = (B + i') * S, S constant, c = kS for some integer k
1347 X = Y + c
1348 ============================
1349 X = (B + (i'+ k)) * S
1350 OR
1351 Y = B + (i' * S), S constant, c = kS for some integer k
1352 X = Y + c
1353 ============================
1354 X = (B + (i'+ k)) * S */
1355 kind = base_cand->kind;
3cfd4469 1356 base = base_cand->base_expr;
27bcd47c 1357 index = base_cand->index + multiple;
f9453c07
BS
1358 stride = base_cand->stride;
1359 ctype = base_cand->cand_type;
1360 if (has_single_use (base_in))
1361 savings = (base_cand->dead_savings
1362 + stmt_cost (base_cand->cand_stmt, speed));
1363 }
1364
1365 if (base_cand->next_interp)
1366 base_cand = lookup_cand (base_cand->next_interp);
1367 else
1368 base_cand = NULL;
1369 }
1370
1371 if (!base)
1372 {
1373 /* No interpretations had anything useful to propagate, so
1374 produce X = Y + (c * 1). */
1375 kind = CAND_ADD;
1376 base = base_in;
1377 index = index_in;
1378 stride = integer_one_node;
b2ec94d4 1379 ctype = TREE_TYPE (base_in);
f9453c07
BS
1380 }
1381
1382 c = alloc_cand_and_find_basis (kind, gs, base, index, stride,
1383 ctype, savings);
1384 return c;
1385}
1386
1387/* Given GS which is an add or subtract of scalar integers or pointers,
1388 make at least one appropriate entry in the candidate table. */
1389
1390static void
355fe088 1391slsr_process_add (gimple *gs, tree rhs1, tree rhs2, bool speed)
f9453c07
BS
1392{
1393 bool subtract_p = gimple_assign_rhs_code (gs) == MINUS_EXPR;
1394 slsr_cand_t c = NULL, c2;
1395
1396 if (TREE_CODE (rhs2) == SSA_NAME)
1397 {
3cfd4469 1398 /* First record an interpretation assuming RHS1 is the base expression
f9453c07
BS
1399 and RHS2 is the stride. But it doesn't make sense for the
1400 stride to be a pointer, so don't record a candidate in that case. */
b2ec94d4 1401 if (!POINTER_TYPE_P (TREE_TYPE (rhs2)))
f9453c07
BS
1402 {
1403 c = create_add_ssa_cand (gs, rhs1, rhs2, subtract_p, speed);
1404
1405 /* Add the first interpretation to the statement-candidate
1406 mapping. */
1407 add_cand_for_stmt (gs, c);
1408 }
1409
1410 /* If the two RHS operands are identical, or this is a subtract,
1411 we're done. */
1412 if (operand_equal_p (rhs1, rhs2, 0) || subtract_p)
1413 return;
1414
1415 /* Otherwise, record another interpretation assuming RHS2 is the
3cfd4469 1416 base expression and RHS1 is the stride, again provided that the
f9453c07 1417 stride is not a pointer. */
b2ec94d4 1418 if (!POINTER_TYPE_P (TREE_TYPE (rhs1)))
f9453c07
BS
1419 {
1420 c2 = create_add_ssa_cand (gs, rhs2, rhs1, false, speed);
1421 if (c)
1422 c->next_interp = c2->cand_num;
1423 else
1424 add_cand_for_stmt (gs, c2);
1425 }
1426 }
1427 else
1428 {
f9453c07 1429 /* Record an interpretation for the add-immediate. */
807e902e 1430 widest_int index = wi::to_widest (rhs2);
f9453c07 1431 if (subtract_p)
27bcd47c 1432 index = -index;
f9453c07
BS
1433
1434 c = create_add_imm_cand (gs, rhs1, index, speed);
1435
1436 /* Add the interpretation to the statement-candidate mapping. */
1437 add_cand_for_stmt (gs, c);
1438 }
1439}
1440
1441/* Given GS which is a negate of a scalar integer, make an appropriate
1442 entry in the candidate table. A negate is equivalent to a multiply
1443 by -1. */
1444
1445static void
355fe088 1446slsr_process_neg (gimple *gs, tree rhs1, bool speed)
f9453c07
BS
1447{
1448 /* Record a CAND_MULT interpretation for the multiply by -1. */
1449 slsr_cand_t c = create_mul_imm_cand (gs, rhs1, integer_minus_one_node, speed);
1450
1451 /* Add the interpretation to the statement-candidate mapping. */
1452 add_cand_for_stmt (gs, c);
1453}
1454
6b5eea61
BS
1455/* Help function for legal_cast_p, operating on two trees. Checks
1456 whether it's allowable to cast from RHS to LHS. See legal_cast_p
1457 for more details. */
1458
1459static bool
1460legal_cast_p_1 (tree lhs, tree rhs)
1461{
1462 tree lhs_type, rhs_type;
1463 unsigned lhs_size, rhs_size;
1464 bool lhs_wraps, rhs_wraps;
1465
1466 lhs_type = TREE_TYPE (lhs);
1467 rhs_type = TREE_TYPE (rhs);
1468 lhs_size = TYPE_PRECISION (lhs_type);
1469 rhs_size = TYPE_PRECISION (rhs_type);
20bd649a
MP
1470 lhs_wraps = ANY_INTEGRAL_TYPE_P (lhs_type) && TYPE_OVERFLOW_WRAPS (lhs_type);
1471 rhs_wraps = ANY_INTEGRAL_TYPE_P (rhs_type) && TYPE_OVERFLOW_WRAPS (rhs_type);
6b5eea61
BS
1472
1473 if (lhs_size < rhs_size
1474 || (rhs_wraps && !lhs_wraps)
1475 || (rhs_wraps && lhs_wraps && rhs_size != lhs_size))
1476 return false;
1477
1478 return true;
1479}
1480
f9453c07
BS
1481/* Return TRUE if GS is a statement that defines an SSA name from
1482 a conversion and is legal for us to combine with an add and multiply
1483 in the candidate table. For example, suppose we have:
1484
1485 A = B + i;
1486 C = (type) A;
1487 D = C * S;
1488
1489 Without the type-cast, we would create a CAND_MULT for D with base B,
1490 index i, and stride S. We want to record this candidate only if it
1491 is equivalent to apply the type cast following the multiply:
1492
1493 A = B + i;
1494 E = A * S;
1495 D = (type) E;
1496
1497 We will record the type with the candidate for D. This allows us
1498 to use a similar previous candidate as a basis. If we have earlier seen
1499
1500 A' = B + i';
1501 C' = (type) A';
1502 D' = C' * S;
1503
1504 we can replace D with
1505
1506 D = D' + (i - i') * S;
1507
1508 But if moving the type-cast would change semantics, we mustn't do this.
1509
1510 This is legitimate for casts from a non-wrapping integral type to
1511 any integral type of the same or larger size. It is not legitimate
1512 to convert a wrapping type to a non-wrapping type, or to a wrapping
1513 type of a different size. I.e., with a wrapping type, we must
1514 assume that the addition B + i could wrap, in which case performing
1515 the multiply before or after one of the "illegal" type casts will
1516 have different semantics. */
1517
1518static bool
355fe088 1519legal_cast_p (gimple *gs, tree rhs)
f9453c07 1520{
f9453c07
BS
1521 if (!is_gimple_assign (gs)
1522 || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (gs)))
1523 return false;
1524
6b5eea61 1525 return legal_cast_p_1 (gimple_assign_lhs (gs), rhs);
f9453c07
BS
1526}
1527
1528/* Given GS which is a cast to a scalar integer type, determine whether
1529 the cast is legal for strength reduction. If so, make at least one
1530 appropriate entry in the candidate table. */
1531
1532static void
355fe088 1533slsr_process_cast (gimple *gs, tree rhs1, bool speed)
f9453c07
BS
1534{
1535 tree lhs, ctype;
1536 slsr_cand_t base_cand, c, c2;
1537 unsigned savings = 0;
1538
1539 if (!legal_cast_p (gs, rhs1))
1540 return;
1541
1542 lhs = gimple_assign_lhs (gs);
1543 base_cand = base_cand_from_table (rhs1);
1544 ctype = TREE_TYPE (lhs);
1545
9b92d12b 1546 if (base_cand && base_cand->kind != CAND_PHI)
f9453c07
BS
1547 {
1548 while (base_cand)
1549 {
1550 /* Propagate all data from the base candidate except the type,
1551 which comes from the cast, and the base candidate's cast,
1552 which is no longer applicable. */
1553 if (has_single_use (rhs1))
1554 savings = (base_cand->dead_savings
1555 + stmt_cost (base_cand->cand_stmt, speed));
1556
1557 c = alloc_cand_and_find_basis (base_cand->kind, gs,
3cfd4469 1558 base_cand->base_expr,
f9453c07
BS
1559 base_cand->index, base_cand->stride,
1560 ctype, savings);
1561 if (base_cand->next_interp)
1562 base_cand = lookup_cand (base_cand->next_interp);
1563 else
1564 base_cand = NULL;
1565 }
1566 }
1567 else
1568 {
1569 /* If nothing is known about the RHS, create fresh CAND_ADD and
1570 CAND_MULT interpretations:
1571
1572 X = Y + (0 * 1)
1573 X = (Y + 0) * 1
1574
1575 The first of these is somewhat arbitrary, but the choice of
1576 1 for the stride simplifies the logic for propagating casts
1577 into their uses. */
807e902e
KZ
1578 c = alloc_cand_and_find_basis (CAND_ADD, gs, rhs1,
1579 0, integer_one_node, ctype, 0);
1580 c2 = alloc_cand_and_find_basis (CAND_MULT, gs, rhs1,
1581 0, integer_one_node, ctype, 0);
f9453c07
BS
1582 c->next_interp = c2->cand_num;
1583 }
1584
1585 /* Add the first (or only) interpretation to the statement-candidate
1586 mapping. */
1587 add_cand_for_stmt (gs, c);
1588}
1589
1590/* Given GS which is a copy of a scalar integer type, make at least one
1591 appropriate entry in the candidate table.
1592
1593 This interface is included for completeness, but is unnecessary
1594 if this pass immediately follows a pass that performs copy
1595 propagation, such as DOM. */
1596
1597static void
355fe088 1598slsr_process_copy (gimple *gs, tree rhs1, bool speed)
f9453c07
BS
1599{
1600 slsr_cand_t base_cand, c, c2;
1601 unsigned savings = 0;
1602
1603 base_cand = base_cand_from_table (rhs1);
1604
9b92d12b 1605 if (base_cand && base_cand->kind != CAND_PHI)
f9453c07
BS
1606 {
1607 while (base_cand)
1608 {
1609 /* Propagate all data from the base candidate. */
1610 if (has_single_use (rhs1))
1611 savings = (base_cand->dead_savings
1612 + stmt_cost (base_cand->cand_stmt, speed));
1613
1614 c = alloc_cand_and_find_basis (base_cand->kind, gs,
3cfd4469 1615 base_cand->base_expr,
f9453c07
BS
1616 base_cand->index, base_cand->stride,
1617 base_cand->cand_type, savings);
1618 if (base_cand->next_interp)
1619 base_cand = lookup_cand (base_cand->next_interp);
1620 else
1621 base_cand = NULL;
1622 }
1623 }
1624 else
1625 {
1626 /* If nothing is known about the RHS, create fresh CAND_ADD and
1627 CAND_MULT interpretations:
1628
1629 X = Y + (0 * 1)
1630 X = (Y + 0) * 1
1631
1632 The first of these is somewhat arbitrary, but the choice of
1633 1 for the stride simplifies the logic for propagating casts
1634 into their uses. */
807e902e
KZ
1635 c = alloc_cand_and_find_basis (CAND_ADD, gs, rhs1,
1636 0, integer_one_node, TREE_TYPE (rhs1), 0);
1637 c2 = alloc_cand_and_find_basis (CAND_MULT, gs, rhs1,
1638 0, integer_one_node, TREE_TYPE (rhs1), 0);
f9453c07
BS
1639 c->next_interp = c2->cand_num;
1640 }
1641
1642 /* Add the first (or only) interpretation to the statement-candidate
1643 mapping. */
1644 add_cand_for_stmt (gs, c);
1645}
1646\f
4d9192b5
TS
1647class find_candidates_dom_walker : public dom_walker
1648{
1649public:
1650 find_candidates_dom_walker (cdi_direction direction)
1651 : dom_walker (direction) {}
1652 virtual void before_dom_children (basic_block);
1653};
1654
f9453c07
BS
1655/* Find strength-reduction candidates in block BB. */
1656
4d9192b5
TS
1657void
1658find_candidates_dom_walker::before_dom_children (basic_block bb)
f9453c07
BS
1659{
1660 bool speed = optimize_bb_for_speed_p (bb);
f9453c07 1661
538dd0b7
DM
1662 for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
1663 gsi_next (&gsi))
1664 slsr_process_phi (gsi.phi (), speed);
9b92d12b 1665
538dd0b7
DM
1666 for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1667 gsi_next (&gsi))
f9453c07 1668 {
355fe088 1669 gimple *gs = gsi_stmt (gsi);
f9453c07 1670
2749c8f6
BS
1671 if (gimple_vuse (gs) && gimple_assign_single_p (gs))
1672 slsr_process_ref (gs);
1673
1674 else if (is_gimple_assign (gs)
1675 && SCALAR_INT_MODE_P
1676 (TYPE_MODE (TREE_TYPE (gimple_assign_lhs (gs)))))
f9453c07
BS
1677 {
1678 tree rhs1 = NULL_TREE, rhs2 = NULL_TREE;
1679
1680 switch (gimple_assign_rhs_code (gs))
1681 {
1682 case MULT_EXPR:
1683 case PLUS_EXPR:
1684 rhs1 = gimple_assign_rhs1 (gs);
1685 rhs2 = gimple_assign_rhs2 (gs);
1686 /* Should never happen, but currently some buggy situations
1687 in earlier phases put constants in rhs1. */
1688 if (TREE_CODE (rhs1) != SSA_NAME)
1689 continue;
1690 break;
1691
1692 /* Possible future opportunity: rhs1 of a ptr+ can be
1693 an ADDR_EXPR. */
1694 case POINTER_PLUS_EXPR:
1695 case MINUS_EXPR:
1696 rhs2 = gimple_assign_rhs2 (gs);
1697 /* Fall-through. */
1698
d822570f 1699 CASE_CONVERT:
f9453c07
BS
1700 case MODIFY_EXPR:
1701 case NEGATE_EXPR:
1702 rhs1 = gimple_assign_rhs1 (gs);
1703 if (TREE_CODE (rhs1) != SSA_NAME)
1704 continue;
1705 break;
1706
1707 default:
1708 ;
1709 }
1710
1711 switch (gimple_assign_rhs_code (gs))
1712 {
1713 case MULT_EXPR:
1714 slsr_process_mul (gs, rhs1, rhs2, speed);
1715 break;
1716
1717 case PLUS_EXPR:
1718 case POINTER_PLUS_EXPR:
1719 case MINUS_EXPR:
1720 slsr_process_add (gs, rhs1, rhs2, speed);
1721 break;
1722
1723 case NEGATE_EXPR:
1724 slsr_process_neg (gs, rhs1, speed);
1725 break;
1726
d822570f 1727 CASE_CONVERT:
f9453c07
BS
1728 slsr_process_cast (gs, rhs1, speed);
1729 break;
1730
1731 case MODIFY_EXPR:
1732 slsr_process_copy (gs, rhs1, speed);
1733 break;
1734
1735 default:
1736 ;
1737 }
1738 }
1739 }
1740}
1741\f
1742/* Dump a candidate for debug. */
1743
1744static void
1745dump_candidate (slsr_cand_t c)
1746{
1747 fprintf (dump_file, "%3d [%d] ", c->cand_num,
1748 gimple_bb (c->cand_stmt)->index);
1749 print_gimple_stmt (dump_file, c->cand_stmt, 0, 0);
1750 switch (c->kind)
1751 {
1752 case CAND_MULT:
1753 fputs (" MULT : (", dump_file);
3cfd4469 1754 print_generic_expr (dump_file, c->base_expr, 0);
f9453c07 1755 fputs (" + ", dump_file);
807e902e 1756 print_decs (c->index, dump_file);
f9453c07
BS
1757 fputs (") * ", dump_file);
1758 print_generic_expr (dump_file, c->stride, 0);
1759 fputs (" : ", dump_file);
1760 break;
1761 case CAND_ADD:
1762 fputs (" ADD : ", dump_file);
3cfd4469 1763 print_generic_expr (dump_file, c->base_expr, 0);
f9453c07 1764 fputs (" + (", dump_file);
807e902e 1765 print_decs (c->index, dump_file);
f9453c07
BS
1766 fputs (" * ", dump_file);
1767 print_generic_expr (dump_file, c->stride, 0);
1768 fputs (") : ", dump_file);
1769 break;
2749c8f6
BS
1770 case CAND_REF:
1771 fputs (" REF : ", dump_file);
3cfd4469 1772 print_generic_expr (dump_file, c->base_expr, 0);
2749c8f6
BS
1773 fputs (" + (", dump_file);
1774 print_generic_expr (dump_file, c->stride, 0);
1775 fputs (") + ", dump_file);
807e902e 1776 print_decs (c->index, dump_file);
2749c8f6
BS
1777 fputs (" : ", dump_file);
1778 break;
9b92d12b
BS
1779 case CAND_PHI:
1780 fputs (" PHI : ", dump_file);
1781 print_generic_expr (dump_file, c->base_expr, 0);
1782 fputs (" + (unknown * ", dump_file);
1783 print_generic_expr (dump_file, c->stride, 0);
1784 fputs (") : ", dump_file);
1785 break;
f9453c07
BS
1786 default:
1787 gcc_unreachable ();
1788 }
1789 print_generic_expr (dump_file, c->cand_type, 0);
1790 fprintf (dump_file, "\n basis: %d dependent: %d sibling: %d\n",
1791 c->basis, c->dependent, c->sibling);
1792 fprintf (dump_file, " next-interp: %d dead-savings: %d\n",
1793 c->next_interp, c->dead_savings);
1794 if (c->def_phi)
9b92d12b 1795 fprintf (dump_file, " phi: %d\n", c->def_phi);
f9453c07
BS
1796 fputs ("\n", dump_file);
1797}
1798
1799/* Dump the candidate vector for debug. */
1800
1801static void
1802dump_cand_vec (void)
1803{
1804 unsigned i;
1805 slsr_cand_t c;
1806
1807 fprintf (dump_file, "\nStrength reduction candidate vector:\n\n");
1808
9771b263 1809 FOR_EACH_VEC_ELT (cand_vec, i, c)
f9453c07
BS
1810 dump_candidate (c);
1811}
1812
2749c8f6 1813/* Callback used to dump the candidate chains hash table. */
f9453c07 1814
4a8fb1a1
LC
1815int
1816ssa_base_cand_dump_callback (cand_chain **slot, void *ignored ATTRIBUTE_UNUSED)
f9453c07 1817{
4a8fb1a1 1818 const_cand_chain_t chain = *slot;
2749c8f6 1819 cand_chain_t p;
f9453c07 1820
3cfd4469 1821 print_generic_expr (dump_file, chain->base_expr, 0);
2749c8f6 1822 fprintf (dump_file, " -> %d", chain->cand->cand_num);
f9453c07 1823
2749c8f6
BS
1824 for (p = chain->next; p; p = p->next)
1825 fprintf (dump_file, " -> %d", p->cand->cand_num);
f9453c07 1826
2749c8f6
BS
1827 fputs ("\n", dump_file);
1828 return 1;
1829}
f9453c07 1830
2749c8f6 1831/* Dump the candidate chains. */
f9453c07 1832
2749c8f6
BS
1833static void
1834dump_cand_chains (void)
1835{
1836 fprintf (dump_file, "\nStrength reduction candidate chains:\n\n");
c203e8a7
TS
1837 base_cand_map->traverse_noresize <void *, ssa_base_cand_dump_callback>
1838 (NULL);
f9453c07
BS
1839 fputs ("\n", dump_file);
1840}
88ca9ea1
BS
1841
1842/* Dump the increment vector for debug. */
1843
1844static void
1845dump_incr_vec (void)
1846{
1847 if (dump_file && (dump_flags & TDF_DETAILS))
1848 {
1849 unsigned i;
1850
1851 fprintf (dump_file, "\nIncrement vector:\n\n");
1852
1853 for (i = 0; i < incr_vec_len; i++)
1854 {
1855 fprintf (dump_file, "%3d increment: ", i);
807e902e 1856 print_decs (incr_vec[i].incr, dump_file);
88ca9ea1
BS
1857 fprintf (dump_file, "\n count: %d", incr_vec[i].count);
1858 fprintf (dump_file, "\n cost: %d", incr_vec[i].cost);
1859 fputs ("\n initializer: ", dump_file);
1860 print_generic_expr (dump_file, incr_vec[i].initializer, 0);
1861 fputs ("\n\n", dump_file);
1862 }
1863 }
1864}
f9453c07 1865\f
2749c8f6
BS
1866/* Replace *EXPR in candidate C with an equivalent strength-reduced
1867 data reference. */
1868
1869static void
1870replace_ref (tree *expr, slsr_cand_t c)
1871{
78f6dd68
MJ
1872 tree add_expr, mem_ref, acc_type = TREE_TYPE (*expr);
1873 unsigned HOST_WIDE_INT misalign;
1874 unsigned align;
1875
1876 /* Ensure the memory reference carries the minimum alignment
1877 requirement for the data type. See PR58041. */
1878 get_object_alignment_1 (*expr, &align, &misalign);
1879 if (misalign != 0)
1880 align = (misalign & -misalign);
1881 if (align < TYPE_ALIGN (acc_type))
1882 acc_type = build_aligned_type (acc_type, align);
1883
1884 add_expr = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (c->base_expr),
1885 c->base_expr, c->stride);
1886 mem_ref = fold_build2 (MEM_REF, acc_type, add_expr,
807e902e 1887 wide_int_to_tree (c->cand_type, c->index));
78f6dd68 1888
2749c8f6
BS
1889 /* Gimplify the base addressing expression for the new MEM_REF tree. */
1890 gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
1891 TREE_OPERAND (mem_ref, 0)
1892 = force_gimple_operand_gsi (&gsi, TREE_OPERAND (mem_ref, 0),
1893 /*simple_p=*/true, NULL,
1894 /*before=*/true, GSI_SAME_STMT);
1895 copy_ref_info (mem_ref, *expr);
1896 *expr = mem_ref;
1897 update_stmt (c->cand_stmt);
1898}
1899
1900/* Replace CAND_REF candidate C, each sibling of candidate C, and each
1901 dependent of candidate C with an equivalent strength-reduced data
1902 reference. */
1903
1904static void
1905replace_refs (slsr_cand_t c)
1906{
96d75a2c
YZ
1907 if (dump_file && (dump_flags & TDF_DETAILS))
1908 {
1909 fputs ("Replacing reference: ", dump_file);
1910 print_gimple_stmt (dump_file, c->cand_stmt, 0, 0);
1911 }
1912
2749c8f6
BS
1913 if (gimple_vdef (c->cand_stmt))
1914 {
1915 tree *lhs = gimple_assign_lhs_ptr (c->cand_stmt);
1916 replace_ref (lhs, c);
1917 }
1918 else
1919 {
1920 tree *rhs = gimple_assign_rhs1_ptr (c->cand_stmt);
1921 replace_ref (rhs, c);
1922 }
1923
96d75a2c
YZ
1924 if (dump_file && (dump_flags & TDF_DETAILS))
1925 {
1926 fputs ("With: ", dump_file);
1927 print_gimple_stmt (dump_file, c->cand_stmt, 0, 0);
1928 fputs ("\n", dump_file);
1929 }
1930
2749c8f6
BS
1931 if (c->sibling)
1932 replace_refs (lookup_cand (c->sibling));
1933
1934 if (c->dependent)
1935 replace_refs (lookup_cand (c->dependent));
1936}
1937
9b92d12b
BS
1938/* Return TRUE if candidate C is dependent upon a PHI. */
1939
1940static bool
1941phi_dependent_cand_p (slsr_cand_t c)
1942{
1943 /* A candidate is not necessarily dependent upon a PHI just because
1944 it has a phi definition for its base name. It may have a basis
1945 that relies upon the same phi definition, in which case the PHI
1946 is irrelevant to this candidate. */
1947 return (c->def_phi
1948 && c->basis
1949 && lookup_cand (c->basis)->def_phi != c->def_phi);
1950}
1951
f9453c07
BS
1952/* Calculate the increment required for candidate C relative to
1953 its basis. */
1954
807e902e 1955static widest_int
f9453c07
BS
1956cand_increment (slsr_cand_t c)
1957{
1958 slsr_cand_t basis;
1959
1960 /* If the candidate doesn't have a basis, just return its own
1961 index. This is useful in record_increments to help us find
9b92d12b
BS
1962 an existing initializer. Also, if the candidate's basis is
1963 hidden by a phi, then its own index will be the increment
1964 from the newly introduced phi basis. */
1965 if (!c->basis || phi_dependent_cand_p (c))
f9453c07
BS
1966 return c->index;
1967
1968 basis = lookup_cand (c->basis);
3cfd4469 1969 gcc_assert (operand_equal_p (c->base_expr, basis->base_expr, 0));
27bcd47c 1970 return c->index - basis->index;
f9453c07
BS
1971}
1972
88ca9ea1
BS
1973/* Calculate the increment required for candidate C relative to
1974 its basis. If we aren't going to generate pointer arithmetic
1975 for this candidate, return the absolute value of that increment
1976 instead. */
1977
807e902e 1978static inline widest_int
88ca9ea1
BS
1979cand_abs_increment (slsr_cand_t c)
1980{
807e902e 1981 widest_int increment = cand_increment (c);
88ca9ea1 1982
807e902e 1983 if (!address_arithmetic_p && wi::neg_p (increment))
27bcd47c 1984 increment = -increment;
88ca9ea1
BS
1985
1986 return increment;
1987}
1988
f9453c07
BS
1989/* Return TRUE iff candidate C has already been replaced under
1990 another interpretation. */
1991
1992static inline bool
1993cand_already_replaced (slsr_cand_t c)
1994{
1995 return (gimple_bb (c->cand_stmt) == 0);
1996}
1997
9b92d12b
BS
1998/* Common logic used by replace_unconditional_candidate and
1999 replace_conditional_candidate. */
f9453c07
BS
2000
2001static void
807e902e 2002replace_mult_candidate (slsr_cand_t c, tree basis_name, widest_int bump)
f9453c07 2003{
9b92d12b
BS
2004 tree target_type = TREE_TYPE (gimple_assign_lhs (c->cand_stmt));
2005 enum tree_code cand_code = gimple_assign_rhs_code (c->cand_stmt);
2006
f9453c07
BS
2007 /* It is highly unlikely, but possible, that the resulting
2008 bump doesn't fit in a HWI. Abandon the replacement
9b92d12b
BS
2009 in this case. This does not affect siblings or dependents
2010 of C. Restriction to signed HWI is conservative for unsigned
2011 types but allows for safe negation without twisted logic. */
807e902e 2012 if (wi::fits_shwi_p (bump)
9b92d12b
BS
2013 && bump.to_shwi () != HOST_WIDE_INT_MIN
2014 /* It is not useful to replace casts, copies, or adds of
2015 an SSA name and a constant. */
2016 && cand_code != MODIFY_EXPR
d822570f 2017 && !CONVERT_EXPR_CODE_P (cand_code)
9b92d12b
BS
2018 && cand_code != PLUS_EXPR
2019 && cand_code != POINTER_PLUS_EXPR
2020 && cand_code != MINUS_EXPR)
2021 {
2022 enum tree_code code = PLUS_EXPR;
2023 tree bump_tree;
355fe088 2024 gimple *stmt_to_print = NULL;
9b92d12b
BS
2025
2026 /* If the basis name and the candidate's LHS have incompatible
2027 types, introduce a cast. */
2028 if (!useless_type_conversion_p (target_type, TREE_TYPE (basis_name)))
a7a7d10e 2029 basis_name = introduce_cast_before_cand (c, target_type, basis_name);
807e902e 2030 if (wi::neg_p (bump))
9b92d12b
BS
2031 {
2032 code = MINUS_EXPR;
2033 bump = -bump;
2034 }
2035
807e902e 2036 bump_tree = wide_int_to_tree (target_type, bump);
9b92d12b
BS
2037
2038 if (dump_file && (dump_flags & TDF_DETAILS))
2039 {
2040 fputs ("Replacing: ", dump_file);
2041 print_gimple_stmt (dump_file, c->cand_stmt, 0, 0);
2042 }
2043
807e902e 2044 if (bump == 0)
9b92d12b
BS
2045 {
2046 tree lhs = gimple_assign_lhs (c->cand_stmt);
538dd0b7 2047 gassign *copy_stmt = gimple_build_assign (lhs, basis_name);
9b92d12b
BS
2048 gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
2049 gimple_set_location (copy_stmt, gimple_location (c->cand_stmt));
2050 gsi_replace (&gsi, copy_stmt, false);
0100cd3f 2051 c->cand_stmt = copy_stmt;
9b92d12b
BS
2052 if (dump_file && (dump_flags & TDF_DETAILS))
2053 stmt_to_print = copy_stmt;
2054 }
2055 else
2056 {
2057 tree rhs1, rhs2;
2058 if (cand_code != NEGATE_EXPR) {
2059 rhs1 = gimple_assign_rhs1 (c->cand_stmt);
2060 rhs2 = gimple_assign_rhs2 (c->cand_stmt);
2061 }
2062 if (cand_code != NEGATE_EXPR
2063 && ((operand_equal_p (rhs1, basis_name, 0)
2064 && operand_equal_p (rhs2, bump_tree, 0))
2065 || (operand_equal_p (rhs1, bump_tree, 0)
2066 && operand_equal_p (rhs2, basis_name, 0))))
2067 {
2068 if (dump_file && (dump_flags & TDF_DETAILS))
2069 {
2070 fputs ("(duplicate, not actually replacing)", dump_file);
2071 stmt_to_print = c->cand_stmt;
2072 }
2073 }
2074 else
2075 {
2076 gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
2077 gimple_assign_set_rhs_with_ops (&gsi, code,
2078 basis_name, bump_tree);
2079 update_stmt (gsi_stmt (gsi));
bb0d2039 2080 c->cand_stmt = gsi_stmt (gsi);
9b92d12b
BS
2081 if (dump_file && (dump_flags & TDF_DETAILS))
2082 stmt_to_print = gsi_stmt (gsi);
2083 }
2084 }
2085
2086 if (dump_file && (dump_flags & TDF_DETAILS))
2087 {
2088 fputs ("With: ", dump_file);
2089 print_gimple_stmt (dump_file, stmt_to_print, 0, 0);
2090 fputs ("\n", dump_file);
2091 }
2092 }
2093}
2094
2095/* Replace candidate C with an add or subtract. Note that we only
2096 operate on CAND_MULTs with known strides, so we will never generate
2097 a POINTER_PLUS_EXPR. Each candidate X = (B + i) * S is replaced by
2098 X = Y + ((i - i') * S), as described in the module commentary. The
2099 folded value ((i - i') * S) is referred to here as the "bump." */
2100
2101static void
2102replace_unconditional_candidate (slsr_cand_t c)
2103{
2104 slsr_cand_t basis;
9b92d12b
BS
2105
2106 if (cand_already_replaced (c))
f9453c07
BS
2107 return;
2108
2109 basis = lookup_cand (c->basis);
807e902e 2110 widest_int bump = cand_increment (c) * wi::to_widest (c->stride);
9b92d12b 2111
a7a7d10e 2112 replace_mult_candidate (c, gimple_assign_lhs (basis->cand_stmt), bump);
9b92d12b
BS
2113}
2114\f
7bf55a70
BS
2115/* Return the index in the increment vector of the given INCREMENT,
2116 or -1 if not found. The latter can occur if more than
2117 MAX_INCR_VEC_LEN increments have been found. */
9b92d12b 2118
7bf55a70 2119static inline int
807e902e 2120incr_vec_index (const widest_int &increment)
9b92d12b
BS
2121{
2122 unsigned i;
2123
2124 for (i = 0; i < incr_vec_len && increment != incr_vec[i].incr; i++)
2125 ;
2126
7bf55a70
BS
2127 if (i < incr_vec_len)
2128 return i;
2129 else
2130 return -1;
9b92d12b
BS
2131}
2132
2133/* Create a new statement along edge E to add BASIS_NAME to the product
2134 of INCREMENT and the stride of candidate C. Create and return a new
2135 SSA name from *VAR to be used as the LHS of the new statement.
2136 KNOWN_STRIDE is true iff C's stride is a constant. */
2137
2138static tree
2139create_add_on_incoming_edge (slsr_cand_t c, tree basis_name,
807e902e 2140 widest_int increment, edge e, location_t loc,
9b92d12b
BS
2141 bool known_stride)
2142{
2143 basic_block insert_bb;
2144 gimple_stmt_iterator gsi;
2145 tree lhs, basis_type;
538dd0b7 2146 gassign *new_stmt;
9b92d12b
BS
2147
2148 /* If the add candidate along this incoming edge has the same
2149 index as C's hidden basis, the hidden basis represents this
2150 edge correctly. */
807e902e 2151 if (increment == 0)
9b92d12b
BS
2152 return basis_name;
2153
2154 basis_type = TREE_TYPE (basis_name);
2155 lhs = make_temp_ssa_name (basis_type, NULL, "slsr");
2156
2157 if (known_stride)
7139194b 2158 {
9b92d12b
BS
2159 tree bump_tree;
2160 enum tree_code code = PLUS_EXPR;
807e902e
KZ
2161 widest_int bump = increment * wi::to_widest (c->stride);
2162 if (wi::neg_p (bump))
9b92d12b
BS
2163 {
2164 code = MINUS_EXPR;
2165 bump = -bump;
2166 }
2167
807e902e 2168 bump_tree = wide_int_to_tree (basis_type, bump);
0d0e4a03 2169 new_stmt = gimple_build_assign (lhs, code, basis_name, bump_tree);
7139194b
AH
2170 }
2171 else
2172 {
7bf55a70 2173 int i;
807e902e 2174 bool negate_incr = (!address_arithmetic_p && wi::neg_p (increment));
9b92d12b 2175 i = incr_vec_index (negate_incr ? -increment : increment);
7bf55a70 2176 gcc_assert (i >= 0);
f9453c07 2177
9b92d12b
BS
2178 if (incr_vec[i].initializer)
2179 {
2180 enum tree_code code = negate_incr ? MINUS_EXPR : PLUS_EXPR;
0d0e4a03
JJ
2181 new_stmt = gimple_build_assign (lhs, code, basis_name,
2182 incr_vec[i].initializer);
9b92d12b 2183 }
807e902e 2184 else if (increment == 1)
0d0e4a03 2185 new_stmt = gimple_build_assign (lhs, PLUS_EXPR, basis_name, c->stride);
807e902e 2186 else if (increment == -1)
0d0e4a03
JJ
2187 new_stmt = gimple_build_assign (lhs, MINUS_EXPR, basis_name,
2188 c->stride);
9b92d12b
BS
2189 else
2190 gcc_unreachable ();
f9453c07
BS
2191 }
2192
9b92d12b
BS
2193 insert_bb = single_succ_p (e->src) ? e->src : split_edge (e);
2194 gsi = gsi_last_bb (insert_bb);
2195
2196 if (!gsi_end_p (gsi) && is_ctrl_stmt (gsi_stmt (gsi)))
2197 gsi_insert_before (&gsi, new_stmt, GSI_NEW_STMT);
2198 else
2199 gsi_insert_after (&gsi, new_stmt, GSI_NEW_STMT);
2200
2201 gimple_set_location (new_stmt, loc);
f9453c07
BS
2202
2203 if (dump_file && (dump_flags & TDF_DETAILS))
2204 {
9b92d12b
BS
2205 fprintf (dump_file, "Inserting in block %d: ", insert_bb->index);
2206 print_gimple_stmt (dump_file, new_stmt, 0, 0);
f9453c07
BS
2207 }
2208
9b92d12b
BS
2209 return lhs;
2210}
2211
2212/* Given a candidate C with BASIS_NAME being the LHS of C's basis which
2213 is hidden by the phi node FROM_PHI, create a new phi node in the same
2214 block as FROM_PHI. The new phi is suitable for use as a basis by C,
2215 with its phi arguments representing conditional adjustments to the
2216 hidden basis along conditional incoming paths. Those adjustments are
2217 made by creating add statements (and sometimes recursively creating
2218 phis) along those incoming paths. LOC is the location to attach to
2219 the introduced statements. KNOWN_STRIDE is true iff C's stride is a
2220 constant. */
2221
2222static tree
355fe088 2223create_phi_basis (slsr_cand_t c, gimple *from_phi, tree basis_name,
9b92d12b
BS
2224 location_t loc, bool known_stride)
2225{
2226 int i;
2227 tree name, phi_arg;
538dd0b7 2228 gphi *phi;
9b92d12b
BS
2229 slsr_cand_t basis = lookup_cand (c->basis);
2230 int nargs = gimple_phi_num_args (from_phi);
2231 basic_block phi_bb = gimple_bb (from_phi);
2232 slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (from_phi));
c8189787 2233 auto_vec<tree> phi_args (nargs);
9b92d12b
BS
2234
2235 /* Process each argument of the existing phi that represents
2236 conditionally-executed add candidates. */
2237 for (i = 0; i < nargs; i++)
f9453c07 2238 {
9b92d12b
BS
2239 edge e = (*phi_bb->preds)[i];
2240 tree arg = gimple_phi_arg_def (from_phi, i);
2241 tree feeding_def;
2242
2243 /* If the phi argument is the base name of the CAND_PHI, then
2244 this incoming arc should use the hidden basis. */
2245 if (operand_equal_p (arg, phi_cand->base_expr, 0))
807e902e 2246 if (basis->index == 0)
9b92d12b
BS
2247 feeding_def = gimple_assign_lhs (basis->cand_stmt);
2248 else
2249 {
807e902e 2250 widest_int incr = -basis->index;
9b92d12b
BS
2251 feeding_def = create_add_on_incoming_edge (c, basis_name, incr,
2252 e, loc, known_stride);
2253 }
2254 else
f9453c07 2255 {
355fe088 2256 gimple *arg_def = SSA_NAME_DEF_STMT (arg);
9b92d12b
BS
2257
2258 /* If there is another phi along this incoming edge, we must
2259 process it in the same fashion to ensure that all basis
2260 adjustments are made along its incoming edges. */
2261 if (gimple_code (arg_def) == GIMPLE_PHI)
2262 feeding_def = create_phi_basis (c, arg_def, basis_name,
2263 loc, known_stride);
2264 else
f9453c07 2265 {
9b92d12b 2266 slsr_cand_t arg_cand = base_cand_from_table (arg);
807e902e 2267 widest_int diff = arg_cand->index - basis->index;
9b92d12b
BS
2268 feeding_def = create_add_on_incoming_edge (c, basis_name, diff,
2269 e, loc, known_stride);
f9453c07
BS
2270 }
2271 }
9b92d12b
BS
2272
2273 /* Because of recursion, we need to save the arguments in a vector
2274 so we can create the PHI statement all at once. Otherwise the
2275 storage for the half-created PHI can be reclaimed. */
2276 phi_args.safe_push (feeding_def);
f9453c07 2277 }
9b92d12b
BS
2278
2279 /* Create the new phi basis. */
2280 name = make_temp_ssa_name (TREE_TYPE (basis_name), NULL, "slsr");
2281 phi = create_phi_node (name, phi_bb);
2282 SSA_NAME_DEF_STMT (name) = phi;
2283
2284 FOR_EACH_VEC_ELT (phi_args, i, phi_arg)
2285 {
2286 edge e = (*phi_bb->preds)[i];
2287 add_phi_arg (phi, phi_arg, e, loc);
2288 }
2289
2290 update_stmt (phi);
2291
f9453c07
BS
2292 if (dump_file && (dump_flags & TDF_DETAILS))
2293 {
9b92d12b
BS
2294 fputs ("Introducing new phi basis: ", dump_file);
2295 print_gimple_stmt (dump_file, phi, 0, 0);
f9453c07 2296 }
9b92d12b
BS
2297
2298 return name;
f9453c07
BS
2299}
2300
9b92d12b
BS
2301/* Given a candidate C whose basis is hidden by at least one intervening
2302 phi, introduce a matching number of new phis to represent its basis
2303 adjusted by conditional increments along possible incoming paths. Then
2304 replace C as though it were an unconditional candidate, using the new
2305 basis. */
f9453c07
BS
2306
2307static void
9b92d12b 2308replace_conditional_candidate (slsr_cand_t c)
f9453c07 2309{
a7a7d10e 2310 tree basis_name, name;
9b92d12b
BS
2311 slsr_cand_t basis;
2312 location_t loc;
f9453c07 2313
9b92d12b
BS
2314 /* Look up the LHS SSA name from C's basis. This will be the
2315 RHS1 of the adds we will introduce to create new phi arguments. */
2316 basis = lookup_cand (c->basis);
2317 basis_name = gimple_assign_lhs (basis->cand_stmt);
f9453c07 2318
9b92d12b
BS
2319 /* Create a new phi statement which will represent C's true basis
2320 after the transformation is complete. */
2321 loc = gimple_location (c->cand_stmt);
2322 name = create_phi_basis (c, lookup_cand (c->def_phi)->cand_stmt,
2323 basis_name, loc, KNOWN_STRIDE);
2324 /* Replace C with an add of the new basis phi and a constant. */
807e902e 2325 widest_int bump = c->index * wi::to_widest (c->stride);
f9453c07 2326
a7a7d10e 2327 replace_mult_candidate (c, name, bump);
f9453c07 2328}
88ca9ea1 2329
9b92d12b
BS
2330/* Compute the expected costs of inserting basis adjustments for
2331 candidate C with phi-definition PHI. The cost of inserting
2332 one adjustment is given by ONE_ADD_COST. If PHI has arguments
2333 which are themselves phi results, recursively calculate costs
2334 for those phis as well. */
2335
2336static int
355fe088 2337phi_add_costs (gimple *phi, slsr_cand_t c, int one_add_cost)
88ca9ea1
BS
2338{
2339 unsigned i;
9b92d12b
BS
2340 int cost = 0;
2341 slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi));
88ca9ea1 2342
0100cd3f
BS
2343 /* If we work our way back to a phi that isn't dominated by the hidden
2344 basis, this isn't a candidate for replacement. Indicate this by
2345 returning an unreasonably high cost. It's not easy to detect
2346 these situations when determining the basis, so we defer the
2347 decision until now. */
2348 basic_block phi_bb = gimple_bb (phi);
2349 slsr_cand_t basis = lookup_cand (c->basis);
2350 basic_block basis_bb = gimple_bb (basis->cand_stmt);
2351
2352 if (phi_bb == basis_bb || !dominated_by_p (CDI_DOMINATORS, phi_bb, basis_bb))
2353 return COST_INFINITE;
2354
9b92d12b
BS
2355 for (i = 0; i < gimple_phi_num_args (phi); i++)
2356 {
2357 tree arg = gimple_phi_arg_def (phi, i);
2358
2359 if (arg != phi_cand->base_expr)
2360 {
355fe088 2361 gimple *arg_def = SSA_NAME_DEF_STMT (arg);
9b92d12b
BS
2362
2363 if (gimple_code (arg_def) == GIMPLE_PHI)
2364 cost += phi_add_costs (arg_def, c, one_add_cost);
2365 else
2366 {
2367 slsr_cand_t arg_cand = base_cand_from_table (arg);
2368
2369 if (arg_cand->index != c->index)
2370 cost += one_add_cost;
2371 }
2372 }
2373 }
2374
2375 return cost;
88ca9ea1
BS
2376}
2377
9b92d12b
BS
2378/* For candidate C, each sibling of candidate C, and each dependent of
2379 candidate C, determine whether the candidate is dependent upon a
2380 phi that hides its basis. If not, replace the candidate unconditionally.
2381 Otherwise, determine whether the cost of introducing compensation code
2382 for the candidate is offset by the gains from strength reduction. If
2383 so, replace the candidate and introduce the compensation code. */
2384
2385static void
2386replace_uncond_cands_and_profitable_phis (slsr_cand_t c)
2387{
2388 if (phi_dependent_cand_p (c))
2389 {
2390 if (c->kind == CAND_MULT)
2391 {
2392 /* A candidate dependent upon a phi will replace a multiply by
2393 a constant with an add, and will insert at most one add for
2394 each phi argument. Add these costs with the potential dead-code
2395 savings to determine profitability. */
2396 bool speed = optimize_bb_for_speed_p (gimple_bb (c->cand_stmt));
2397 int mult_savings = stmt_cost (c->cand_stmt, speed);
355fe088 2398 gimple *phi = lookup_cand (c->def_phi)->cand_stmt;
9b92d12b
BS
2399 tree phi_result = gimple_phi_result (phi);
2400 int one_add_cost = add_cost (speed,
2401 TYPE_MODE (TREE_TYPE (phi_result)));
2402 int add_costs = one_add_cost + phi_add_costs (phi, c, one_add_cost);
2403 int cost = add_costs - mult_savings - c->dead_savings;
2404
2405 if (dump_file && (dump_flags & TDF_DETAILS))
2406 {
2407 fprintf (dump_file, " Conditional candidate %d:\n", c->cand_num);
2408 fprintf (dump_file, " add_costs = %d\n", add_costs);
2409 fprintf (dump_file, " mult_savings = %d\n", mult_savings);
2410 fprintf (dump_file, " dead_savings = %d\n", c->dead_savings);
2411 fprintf (dump_file, " cost = %d\n", cost);
2412 if (cost <= COST_NEUTRAL)
2413 fputs (" Replacing...\n", dump_file);
2414 else
2415 fputs (" Not replaced.\n", dump_file);
2416 }
2417
2418 if (cost <= COST_NEUTRAL)
2419 replace_conditional_candidate (c);
2420 }
2421 }
2422 else
2423 replace_unconditional_candidate (c);
2424
2425 if (c->sibling)
2426 replace_uncond_cands_and_profitable_phis (lookup_cand (c->sibling));
2427
2428 if (c->dependent)
2429 replace_uncond_cands_and_profitable_phis (lookup_cand (c->dependent));
2430}
2431\f
88ca9ea1
BS
2432/* Count the number of candidates in the tree rooted at C that have
2433 not already been replaced under other interpretations. */
2434
1dc3d6e9 2435static int
88ca9ea1
BS
2436count_candidates (slsr_cand_t c)
2437{
2438 unsigned count = cand_already_replaced (c) ? 0 : 1;
2439
2440 if (c->sibling)
2441 count += count_candidates (lookup_cand (c->sibling));
2442
2443 if (c->dependent)
2444 count += count_candidates (lookup_cand (c->dependent));
2445
2446 return count;
2447}
2448
2449/* Increase the count of INCREMENT by one in the increment vector.
9b92d12b
BS
2450 INCREMENT is associated with candidate C. If INCREMENT is to be
2451 conditionally executed as part of a conditional candidate replacement,
2452 IS_PHI_ADJUST is true, otherwise false. If an initializer
88ca9ea1
BS
2453 T_0 = stride * I is provided by a candidate that dominates all
2454 candidates with the same increment, also record T_0 for subsequent use. */
2455
2456static void
807e902e 2457record_increment (slsr_cand_t c, widest_int increment, bool is_phi_adjust)
88ca9ea1
BS
2458{
2459 bool found = false;
2460 unsigned i;
2461
2462 /* Treat increments that differ only in sign as identical so as to
2463 share initializers, unless we are generating pointer arithmetic. */
807e902e 2464 if (!address_arithmetic_p && wi::neg_p (increment))
27bcd47c 2465 increment = -increment;
88ca9ea1
BS
2466
2467 for (i = 0; i < incr_vec_len; i++)
2468 {
27bcd47c 2469 if (incr_vec[i].incr == increment)
88ca9ea1
BS
2470 {
2471 incr_vec[i].count++;
2472 found = true;
2473
2474 /* If we previously recorded an initializer that doesn't
2475 dominate this candidate, it's not going to be useful to
2476 us after all. */
2477 if (incr_vec[i].initializer
2478 && !dominated_by_p (CDI_DOMINATORS,
2479 gimple_bb (c->cand_stmt),
2480 incr_vec[i].init_bb))
2481 {
2482 incr_vec[i].initializer = NULL_TREE;
2483 incr_vec[i].init_bb = NULL;
2484 }
2485
2486 break;
2487 }
2488 }
2489
7bf55a70 2490 if (!found && incr_vec_len < MAX_INCR_VEC_LEN - 1)
88ca9ea1
BS
2491 {
2492 /* The first time we see an increment, create the entry for it.
2493 If this is the root candidate which doesn't have a basis, set
2494 the count to zero. We're only processing it so it can possibly
2495 provide an initializer for other candidates. */
2496 incr_vec[incr_vec_len].incr = increment;
9b92d12b 2497 incr_vec[incr_vec_len].count = c->basis || is_phi_adjust ? 1 : 0;
88ca9ea1
BS
2498 incr_vec[incr_vec_len].cost = COST_INFINITE;
2499
2500 /* Optimistically record the first occurrence of this increment
2501 as providing an initializer (if it does); we will revise this
2502 opinion later if it doesn't dominate all other occurrences.
9b92d12b
BS
2503 Exception: increments of -1, 0, 1 never need initializers;
2504 and phi adjustments don't ever provide initializers. */
88ca9ea1 2505 if (c->kind == CAND_ADD
9b92d12b 2506 && !is_phi_adjust
27bcd47c 2507 && c->index == increment
807e902e
KZ
2508 && (wi::gts_p (increment, 1)
2509 || wi::lts_p (increment, -1))
7b8265ba
JJ
2510 && (gimple_assign_rhs_code (c->cand_stmt) == PLUS_EXPR
2511 || gimple_assign_rhs_code (c->cand_stmt) == POINTER_PLUS_EXPR))
88ca9ea1 2512 {
7b8265ba 2513 tree t0 = NULL_TREE;
88ca9ea1
BS
2514 tree rhs1 = gimple_assign_rhs1 (c->cand_stmt);
2515 tree rhs2 = gimple_assign_rhs2 (c->cand_stmt);
2516 if (operand_equal_p (rhs1, c->base_expr, 0))
2517 t0 = rhs2;
7b8265ba 2518 else if (operand_equal_p (rhs2, c->base_expr, 0))
88ca9ea1 2519 t0 = rhs1;
7b8265ba
JJ
2520 if (t0
2521 && SSA_NAME_DEF_STMT (t0)
2522 && gimple_bb (SSA_NAME_DEF_STMT (t0)))
88ca9ea1
BS
2523 {
2524 incr_vec[incr_vec_len].initializer = t0;
2525 incr_vec[incr_vec_len++].init_bb
2526 = gimple_bb (SSA_NAME_DEF_STMT (t0));
2527 }
2528 else
2529 {
2530 incr_vec[incr_vec_len].initializer = NULL_TREE;
2531 incr_vec[incr_vec_len++].init_bb = NULL;
2532 }
2533 }
2534 else
2535 {
2536 incr_vec[incr_vec_len].initializer = NULL_TREE;
2537 incr_vec[incr_vec_len++].init_bb = NULL;
2538 }
2539 }
2540}
2541
9b92d12b
BS
2542/* Given phi statement PHI that hides a candidate from its BASIS, find
2543 the increments along each incoming arc (recursively handling additional
2544 phis that may be present) and record them. These increments are the
2545 difference in index between the index-adjusting statements and the
2546 index of the basis. */
2547
2548static void
355fe088 2549record_phi_increments (slsr_cand_t basis, gimple *phi)
9b92d12b
BS
2550{
2551 unsigned i;
2552 slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi));
2553
2554 for (i = 0; i < gimple_phi_num_args (phi); i++)
2555 {
2556 tree arg = gimple_phi_arg_def (phi, i);
2557
2558 if (!operand_equal_p (arg, phi_cand->base_expr, 0))
2559 {
355fe088 2560 gimple *arg_def = SSA_NAME_DEF_STMT (arg);
9b92d12b
BS
2561
2562 if (gimple_code (arg_def) == GIMPLE_PHI)
2563 record_phi_increments (basis, arg_def);
2564 else
2565 {
2566 slsr_cand_t arg_cand = base_cand_from_table (arg);
807e902e 2567 widest_int diff = arg_cand->index - basis->index;
9b92d12b
BS
2568 record_increment (arg_cand, diff, PHI_ADJUST);
2569 }
2570 }
2571 }
2572}
2573
88ca9ea1
BS
2574/* Determine how many times each unique increment occurs in the set
2575 of candidates rooted at C's parent, recording the data in the
2576 increment vector. For each unique increment I, if an initializer
2577 T_0 = stride * I is provided by a candidate that dominates all
2578 candidates with the same increment, also record T_0 for subsequent
2579 use. */
2580
2581static void
2582record_increments (slsr_cand_t c)
2583{
2584 if (!cand_already_replaced (c))
9b92d12b
BS
2585 {
2586 if (!phi_dependent_cand_p (c))
2587 record_increment (c, cand_increment (c), NOT_PHI_ADJUST);
2588 else
2589 {
2590 /* A candidate with a basis hidden by a phi will have one
2591 increment for its relationship to the index represented by
2592 the phi, and potentially additional increments along each
2593 incoming edge. For the root of the dependency tree (which
2594 has no basis), process just the initial index in case it has
2595 an initializer that can be used by subsequent candidates. */
2596 record_increment (c, c->index, NOT_PHI_ADJUST);
2597
2598 if (c->basis)
2599 record_phi_increments (lookup_cand (c->basis),
2600 lookup_cand (c->def_phi)->cand_stmt);
2601 }
2602 }
88ca9ea1
BS
2603
2604 if (c->sibling)
2605 record_increments (lookup_cand (c->sibling));
2606
2607 if (c->dependent)
2608 record_increments (lookup_cand (c->dependent));
2609}
2610
9b92d12b
BS
2611/* Add up and return the costs of introducing add statements that
2612 require the increment INCR on behalf of candidate C and phi
2613 statement PHI. Accumulate into *SAVINGS the potential savings
2614 from removing existing statements that feed PHI and have no other
2615 uses. */
2616
2617static int
355fe088
TS
2618phi_incr_cost (slsr_cand_t c, const widest_int &incr, gimple *phi,
2619 int *savings)
9b92d12b
BS
2620{
2621 unsigned i;
2622 int cost = 0;
2623 slsr_cand_t basis = lookup_cand (c->basis);
2624 slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi));
2625
2626 for (i = 0; i < gimple_phi_num_args (phi); i++)
2627 {
2628 tree arg = gimple_phi_arg_def (phi, i);
2629
2630 if (!operand_equal_p (arg, phi_cand->base_expr, 0))
2631 {
355fe088 2632 gimple *arg_def = SSA_NAME_DEF_STMT (arg);
9b92d12b
BS
2633
2634 if (gimple_code (arg_def) == GIMPLE_PHI)
2635 {
2636 int feeding_savings = 0;
2637 cost += phi_incr_cost (c, incr, arg_def, &feeding_savings);
2638 if (has_single_use (gimple_phi_result (arg_def)))
2639 *savings += feeding_savings;
2640 }
2641 else
2642 {
2643 slsr_cand_t arg_cand = base_cand_from_table (arg);
807e902e 2644 widest_int diff = arg_cand->index - basis->index;
9b92d12b
BS
2645
2646 if (incr == diff)
2647 {
2648 tree basis_lhs = gimple_assign_lhs (basis->cand_stmt);
2649 tree lhs = gimple_assign_lhs (arg_cand->cand_stmt);
2650 cost += add_cost (true, TYPE_MODE (TREE_TYPE (basis_lhs)));
2651 if (has_single_use (lhs))
2652 *savings += stmt_cost (arg_cand->cand_stmt, true);
2653 }
2654 }
2655 }
2656 }
2657
2658 return cost;
2659}
2660
88ca9ea1
BS
2661/* Return the first candidate in the tree rooted at C that has not
2662 already been replaced, favoring siblings over dependents. */
2663
2664static slsr_cand_t
2665unreplaced_cand_in_tree (slsr_cand_t c)
2666{
2667 if (!cand_already_replaced (c))
2668 return c;
2669
2670 if (c->sibling)
2671 {
2672 slsr_cand_t sib = unreplaced_cand_in_tree (lookup_cand (c->sibling));
2673 if (sib)
2674 return sib;
2675 }
2676
2677 if (c->dependent)
2678 {
2679 slsr_cand_t dep = unreplaced_cand_in_tree (lookup_cand (c->dependent));
2680 if (dep)
2681 return dep;
2682 }
2683
2684 return NULL;
2685}
2686
2687/* Return TRUE if the candidates in the tree rooted at C should be
2688 optimized for speed, else FALSE. We estimate this based on the block
2689 containing the most dominant candidate in the tree that has not yet
2690 been replaced. */
2691
2692static bool
2693optimize_cands_for_speed_p (slsr_cand_t c)
2694{
2695 slsr_cand_t c2 = unreplaced_cand_in_tree (c);
2696 gcc_assert (c2);
2697 return optimize_bb_for_speed_p (gimple_bb (c2->cand_stmt));
2698}
2699
2700/* Add COST_IN to the lowest cost of any dependent path starting at
2701 candidate C or any of its siblings, counting only candidates along
2702 such paths with increment INCR. Assume that replacing a candidate
2703 reduces cost by REPL_SAVINGS. Also account for savings from any
9b92d12b
BS
2704 statements that would go dead. If COUNT_PHIS is true, include
2705 costs of introducing feeding statements for conditional candidates. */
88ca9ea1
BS
2706
2707static int
9b92d12b 2708lowest_cost_path (int cost_in, int repl_savings, slsr_cand_t c,
807e902e 2709 const widest_int &incr, bool count_phis)
88ca9ea1 2710{
9b92d12b 2711 int local_cost, sib_cost, savings = 0;
807e902e 2712 widest_int cand_incr = cand_abs_increment (c);
88ca9ea1
BS
2713
2714 if (cand_already_replaced (c))
2715 local_cost = cost_in;
27bcd47c 2716 else if (incr == cand_incr)
88ca9ea1
BS
2717 local_cost = cost_in - repl_savings - c->dead_savings;
2718 else
2719 local_cost = cost_in - c->dead_savings;
2720
9b92d12b
BS
2721 if (count_phis
2722 && phi_dependent_cand_p (c)
2723 && !cand_already_replaced (c))
2724 {
355fe088 2725 gimple *phi = lookup_cand (c->def_phi)->cand_stmt;
9b92d12b
BS
2726 local_cost += phi_incr_cost (c, incr, phi, &savings);
2727
2728 if (has_single_use (gimple_phi_result (phi)))
2729 local_cost -= savings;
2730 }
2731
88ca9ea1
BS
2732 if (c->dependent)
2733 local_cost = lowest_cost_path (local_cost, repl_savings,
9b92d12b
BS
2734 lookup_cand (c->dependent), incr,
2735 count_phis);
88ca9ea1
BS
2736
2737 if (c->sibling)
2738 {
2739 sib_cost = lowest_cost_path (cost_in, repl_savings,
9b92d12b
BS
2740 lookup_cand (c->sibling), incr,
2741 count_phis);
88ca9ea1
BS
2742 local_cost = MIN (local_cost, sib_cost);
2743 }
2744
2745 return local_cost;
2746}
2747
2748/* Compute the total savings that would accrue from all replacements
2749 in the candidate tree rooted at C, counting only candidates with
2750 increment INCR. Assume that replacing a candidate reduces cost
2751 by REPL_SAVINGS. Also account for savings from statements that
2752 would go dead. */
2753
2754static int
807e902e 2755total_savings (int repl_savings, slsr_cand_t c, const widest_int &incr,
9b92d12b 2756 bool count_phis)
88ca9ea1
BS
2757{
2758 int savings = 0;
807e902e 2759 widest_int cand_incr = cand_abs_increment (c);
88ca9ea1 2760
27bcd47c 2761 if (incr == cand_incr && !cand_already_replaced (c))
88ca9ea1
BS
2762 savings += repl_savings + c->dead_savings;
2763
9b92d12b
BS
2764 if (count_phis
2765 && phi_dependent_cand_p (c)
2766 && !cand_already_replaced (c))
2767 {
2768 int phi_savings = 0;
355fe088 2769 gimple *phi = lookup_cand (c->def_phi)->cand_stmt;
9b92d12b
BS
2770 savings -= phi_incr_cost (c, incr, phi, &phi_savings);
2771
2772 if (has_single_use (gimple_phi_result (phi)))
2773 savings += phi_savings;
2774 }
2775
88ca9ea1 2776 if (c->dependent)
9b92d12b
BS
2777 savings += total_savings (repl_savings, lookup_cand (c->dependent), incr,
2778 count_phis);
88ca9ea1
BS
2779
2780 if (c->sibling)
9b92d12b
BS
2781 savings += total_savings (repl_savings, lookup_cand (c->sibling), incr,
2782 count_phis);
88ca9ea1
BS
2783
2784 return savings;
2785}
2786
2787/* Use target-specific costs to determine and record which increments
2788 in the current candidate tree are profitable to replace, assuming
2789 MODE and SPEED. FIRST_DEP is the first dependent of the root of
2790 the candidate tree.
2791
2792 One slight limitation here is that we don't account for the possible
2793 introduction of casts in some cases. See replace_one_candidate for
2794 the cases where these are introduced. This should probably be cleaned
2795 up sometime. */
2796
2797static void
ef4bddc2 2798analyze_increments (slsr_cand_t first_dep, machine_mode mode, bool speed)
88ca9ea1
BS
2799{
2800 unsigned i;
2801
2802 for (i = 0; i < incr_vec_len; i++)
2803 {
27bcd47c 2804 HOST_WIDE_INT incr = incr_vec[i].incr.to_shwi ();
88ca9ea1
BS
2805
2806 /* If somehow this increment is bigger than a HWI, we won't
2807 be optimizing candidates that use it. And if the increment
2808 has a count of zero, nothing will be done with it. */
807e902e 2809 if (!wi::fits_shwi_p (incr_vec[i].incr) || !incr_vec[i].count)
88ca9ea1
BS
2810 incr_vec[i].cost = COST_INFINITE;
2811
2812 /* Increments of 0, 1, and -1 are always profitable to replace,
2813 because they always replace a multiply or add with an add or
2814 copy, and may cause one or more existing instructions to go
2815 dead. Exception: -1 can't be assumed to be profitable for
2816 pointer addition. */
2817 else if (incr == 0
2818 || incr == 1
2819 || (incr == -1
2820 && (gimple_assign_rhs_code (first_dep->cand_stmt)
2821 != POINTER_PLUS_EXPR)))
2822 incr_vec[i].cost = COST_NEUTRAL;
2823
6b5eea61
BS
2824 /* FORNOW: If we need to add an initializer, give up if a cast from
2825 the candidate's type to its stride's type can lose precision.
2826 This could eventually be handled better by expressly retaining the
2827 result of a cast to a wider type in the stride. Example:
2828
2829 short int _1;
2830 _2 = (int) _1;
2831 _3 = _2 * 10;
2832 _4 = x + _3; ADD: x + (10 * _1) : int
2833 _5 = _2 * 15;
2834 _6 = x + _3; ADD: x + (15 * _1) : int
2835
2836 Right now replacing _6 would cause insertion of an initializer
2837 of the form "short int T = _1 * 5;" followed by a cast to
2838 int, which could overflow incorrectly. Had we recorded _2 or
2839 (int)_1 as the stride, this wouldn't happen. However, doing
2840 this breaks other opportunities, so this will require some
2841 care. */
2842 else if (!incr_vec[i].initializer
2843 && TREE_CODE (first_dep->stride) != INTEGER_CST
2844 && !legal_cast_p_1 (first_dep->stride,
2845 gimple_assign_lhs (first_dep->cand_stmt)))
2846
2847 incr_vec[i].cost = COST_INFINITE;
2848
50251425
BS
2849 /* If we need to add an initializer, make sure we don't introduce
2850 a multiply by a pointer type, which can happen in certain cast
2851 scenarios. FIXME: When cleaning up these cast issues, we can
2852 afford to introduce the multiply provided we cast out to an
2853 unsigned int of appropriate size. */
2854 else if (!incr_vec[i].initializer
2855 && TREE_CODE (first_dep->stride) != INTEGER_CST
2856 && POINTER_TYPE_P (TREE_TYPE (first_dep->stride)))
2857
2858 incr_vec[i].cost = COST_INFINITE;
2859
88ca9ea1
BS
2860 /* For any other increment, if this is a multiply candidate, we
2861 must introduce a temporary T and initialize it with
2862 T_0 = stride * increment. When optimizing for speed, walk the
2863 candidate tree to calculate the best cost reduction along any
2864 path; if it offsets the fixed cost of inserting the initializer,
2865 replacing the increment is profitable. When optimizing for
2866 size, instead calculate the total cost reduction from replacing
2867 all candidates with this increment. */
2868 else if (first_dep->kind == CAND_MULT)
2869 {
2870 int cost = mult_by_coeff_cost (incr, mode, speed);
2871 int repl_savings = mul_cost (speed, mode) - add_cost (speed, mode);
2872 if (speed)
2873 cost = lowest_cost_path (cost, repl_savings, first_dep,
9b92d12b 2874 incr_vec[i].incr, COUNT_PHIS);
88ca9ea1 2875 else
9b92d12b
BS
2876 cost -= total_savings (repl_savings, first_dep, incr_vec[i].incr,
2877 COUNT_PHIS);
88ca9ea1
BS
2878
2879 incr_vec[i].cost = cost;
2880 }
2881
2882 /* If this is an add candidate, the initializer may already
2883 exist, so only calculate the cost of the initializer if it
2884 doesn't. We are replacing one add with another here, so the
2885 known replacement savings is zero. We will account for removal
2886 of dead instructions in lowest_cost_path or total_savings. */
2887 else
2888 {
2889 int cost = 0;
2890 if (!incr_vec[i].initializer)
2891 cost = mult_by_coeff_cost (incr, mode, speed);
2892
2893 if (speed)
9b92d12b
BS
2894 cost = lowest_cost_path (cost, 0, first_dep, incr_vec[i].incr,
2895 DONT_COUNT_PHIS);
88ca9ea1 2896 else
9b92d12b
BS
2897 cost -= total_savings (0, first_dep, incr_vec[i].incr,
2898 DONT_COUNT_PHIS);
88ca9ea1
BS
2899
2900 incr_vec[i].cost = cost;
2901 }
2902 }
2903}
2904
2905/* Return the nearest common dominator of BB1 and BB2. If the blocks
2906 are identical, return the earlier of C1 and C2 in *WHERE. Otherwise,
2907 if the NCD matches BB1, return C1 in *WHERE; if the NCD matches BB2,
2908 return C2 in *WHERE; and if the NCD matches neither, return NULL in
2909 *WHERE. Note: It is possible for one of C1 and C2 to be NULL. */
2910
2911static basic_block
2912ncd_for_two_cands (basic_block bb1, basic_block bb2,
2913 slsr_cand_t c1, slsr_cand_t c2, slsr_cand_t *where)
2914{
2915 basic_block ncd;
2916
2917 if (!bb1)
2918 {
2919 *where = c2;
2920 return bb2;
2921 }
2922
2923 if (!bb2)
2924 {
2925 *where = c1;
2926 return bb1;
2927 }
2928
2929 ncd = nearest_common_dominator (CDI_DOMINATORS, bb1, bb2);
2930
2931 /* If both candidates are in the same block, the earlier
2932 candidate wins. */
2933 if (bb1 == ncd && bb2 == ncd)
2934 {
2935 if (!c1 || (c2 && c2->cand_num < c1->cand_num))
2936 *where = c2;
2937 else
2938 *where = c1;
2939 }
2940
2941 /* Otherwise, if one of them produced a candidate in the
2942 dominator, that one wins. */
2943 else if (bb1 == ncd)
2944 *where = c1;
2945
2946 else if (bb2 == ncd)
2947 *where = c2;
2948
2949 /* If neither matches the dominator, neither wins. */
2950 else
2951 *where = NULL;
2952
2953 return ncd;
2954}
2955
9b92d12b
BS
2956/* Consider all candidates that feed PHI. Find the nearest common
2957 dominator of those candidates requiring the given increment INCR.
2958 Further find and return the nearest common dominator of this result
2959 with block NCD. If the returned block contains one or more of the
2960 candidates, return the earliest candidate in the block in *WHERE. */
2961
2962static basic_block
538dd0b7 2963ncd_with_phi (slsr_cand_t c, const widest_int &incr, gphi *phi,
9b92d12b
BS
2964 basic_block ncd, slsr_cand_t *where)
2965{
2966 unsigned i;
2967 slsr_cand_t basis = lookup_cand (c->basis);
2968 slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi));
2969
2970 for (i = 0; i < gimple_phi_num_args (phi); i++)
2971 {
2972 tree arg = gimple_phi_arg_def (phi, i);
2973
2974 if (!operand_equal_p (arg, phi_cand->base_expr, 0))
2975 {
355fe088 2976 gimple *arg_def = SSA_NAME_DEF_STMT (arg);
9b92d12b
BS
2977
2978 if (gimple_code (arg_def) == GIMPLE_PHI)
538dd0b7
DM
2979 ncd = ncd_with_phi (c, incr, as_a <gphi *> (arg_def), ncd,
2980 where);
9b92d12b
BS
2981 else
2982 {
2983 slsr_cand_t arg_cand = base_cand_from_table (arg);
807e902e 2984 widest_int diff = arg_cand->index - basis->index;
1e386bb8 2985 basic_block pred = gimple_phi_arg_edge (phi, i)->src;
9b92d12b
BS
2986
2987 if ((incr == diff) || (!address_arithmetic_p && incr == -diff))
1e386bb8 2988 ncd = ncd_for_two_cands (ncd, pred, *where, NULL, where);
9b92d12b
BS
2989 }
2990 }
2991 }
2992
2993 return ncd;
2994}
2995
2996/* Consider the candidate C together with any candidates that feed
2997 C's phi dependence (if any). Find and return the nearest common
2998 dominator of those candidates requiring the given increment INCR.
2999 If the returned block contains one or more of the candidates,
3000 return the earliest candidate in the block in *WHERE. */
3001
3002static basic_block
807e902e 3003ncd_of_cand_and_phis (slsr_cand_t c, const widest_int &incr, slsr_cand_t *where)
9b92d12b
BS
3004{
3005 basic_block ncd = NULL;
3006
3007 if (cand_abs_increment (c) == incr)
3008 {
3009 ncd = gimple_bb (c->cand_stmt);
3010 *where = c;
3011 }
3012
3013 if (phi_dependent_cand_p (c))
538dd0b7
DM
3014 ncd = ncd_with_phi (c, incr,
3015 as_a <gphi *> (lookup_cand (c->def_phi)->cand_stmt),
9b92d12b
BS
3016 ncd, where);
3017
3018 return ncd;
3019}
3020
88ca9ea1
BS
3021/* Consider all candidates in the tree rooted at C for which INCR
3022 represents the required increment of C relative to its basis.
3023 Find and return the basic block that most nearly dominates all
3024 such candidates. If the returned block contains one or more of
3025 the candidates, return the earliest candidate in the block in
3026 *WHERE. */
3027
3028static basic_block
807e902e 3029nearest_common_dominator_for_cands (slsr_cand_t c, const widest_int &incr,
88ca9ea1
BS
3030 slsr_cand_t *where)
3031{
3032 basic_block sib_ncd = NULL, dep_ncd = NULL, this_ncd = NULL, ncd;
3033 slsr_cand_t sib_where = NULL, dep_where = NULL, this_where = NULL, new_where;
88ca9ea1
BS
3034
3035 /* First find the NCD of all siblings and dependents. */
3036 if (c->sibling)
3037 sib_ncd = nearest_common_dominator_for_cands (lookup_cand (c->sibling),
3038 incr, &sib_where);
3039 if (c->dependent)
3040 dep_ncd = nearest_common_dominator_for_cands (lookup_cand (c->dependent),
3041 incr, &dep_where);
3042 if (!sib_ncd && !dep_ncd)
3043 {
3044 new_where = NULL;
3045 ncd = NULL;
3046 }
3047 else if (sib_ncd && !dep_ncd)
3048 {
3049 new_where = sib_where;
3050 ncd = sib_ncd;
3051 }
3052 else if (dep_ncd && !sib_ncd)
3053 {
3054 new_where = dep_where;
3055 ncd = dep_ncd;
3056 }
3057 else
3058 ncd = ncd_for_two_cands (sib_ncd, dep_ncd, sib_where,
3059 dep_where, &new_where);
3060
3061 /* If the candidate's increment doesn't match the one we're interested
9b92d12b
BS
3062 in (and nor do any increments for feeding defs of a phi-dependence),
3063 then the result depends only on siblings and dependents. */
3064 this_ncd = ncd_of_cand_and_phis (c, incr, &this_where);
88ca9ea1 3065
9b92d12b 3066 if (!this_ncd || cand_already_replaced (c))
88ca9ea1
BS
3067 {
3068 *where = new_where;
3069 return ncd;
3070 }
3071
3072 /* Otherwise, compare this candidate with the result from all siblings
3073 and dependents. */
88ca9ea1
BS
3074 ncd = ncd_for_two_cands (ncd, this_ncd, new_where, this_where, where);
3075
3076 return ncd;
3077}
3078
3079/* Return TRUE if the increment indexed by INDEX is profitable to replace. */
3080
3081static inline bool
3082profitable_increment_p (unsigned index)
3083{
3084 return (incr_vec[index].cost <= COST_NEUTRAL);
3085}
3086
3087/* For each profitable increment in the increment vector not equal to
3088 0 or 1 (or -1, for non-pointer arithmetic), find the nearest common
3089 dominator of all statements in the candidate chain rooted at C
3090 that require that increment, and insert an initializer
3091 T_0 = stride * increment at that location. Record T_0 with the
3092 increment record. */
3093
3094static void
3095insert_initializers (slsr_cand_t c)
3096{
3097 unsigned i;
88ca9ea1
BS
3098
3099 for (i = 0; i < incr_vec_len; i++)
3100 {
3101 basic_block bb;
3102 slsr_cand_t where = NULL;
538dd0b7 3103 gassign *init_stmt;
88ca9ea1 3104 tree stride_type, new_name, incr_tree;
807e902e 3105 widest_int incr = incr_vec[i].incr;
88ca9ea1
BS
3106
3107 if (!profitable_increment_p (i)
807e902e
KZ
3108 || incr == 1
3109 || (incr == -1
88ca9ea1 3110 && gimple_assign_rhs_code (c->cand_stmt) != POINTER_PLUS_EXPR)
807e902e 3111 || incr == 0)
88ca9ea1
BS
3112 continue;
3113
3114 /* We may have already identified an existing initializer that
3115 will suffice. */
3116 if (incr_vec[i].initializer)
3117 {
3118 if (dump_file && (dump_flags & TDF_DETAILS))
3119 {
3120 fputs ("Using existing initializer: ", dump_file);
3121 print_gimple_stmt (dump_file,
3122 SSA_NAME_DEF_STMT (incr_vec[i].initializer),
3123 0, 0);
3124 }
3125 continue;
3126 }
3127
3128 /* Find the block that most closely dominates all candidates
3129 with this increment. If there is at least one candidate in
3130 that block, the earliest one will be returned in WHERE. */
3131 bb = nearest_common_dominator_for_cands (c, incr, &where);
3132
3133 /* Create a new SSA name to hold the initializer's value. */
3134 stride_type = TREE_TYPE (c->stride);
a7a7d10e 3135 new_name = make_temp_ssa_name (stride_type, NULL, "slsr");
88ca9ea1
BS
3136 incr_vec[i].initializer = new_name;
3137
3138 /* Create the initializer and insert it in the latest possible
3139 dominating position. */
807e902e 3140 incr_tree = wide_int_to_tree (stride_type, incr);
0d0e4a03
JJ
3141 init_stmt = gimple_build_assign (new_name, MULT_EXPR,
3142 c->stride, incr_tree);
88ca9ea1
BS
3143 if (where)
3144 {
3145 gimple_stmt_iterator gsi = gsi_for_stmt (where->cand_stmt);
3146 gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT);
3147 gimple_set_location (init_stmt, gimple_location (where->cand_stmt));
3148 }
3149 else
3150 {
3151 gimple_stmt_iterator gsi = gsi_last_bb (bb);
355fe088 3152 gimple *basis_stmt = lookup_cand (c->basis)->cand_stmt;
88ca9ea1
BS
3153
3154 if (!gsi_end_p (gsi) && is_ctrl_stmt (gsi_stmt (gsi)))
3155 gsi_insert_before (&gsi, init_stmt, GSI_SAME_STMT);
3156 else
3157 gsi_insert_after (&gsi, init_stmt, GSI_SAME_STMT);
3158
3159 gimple_set_location (init_stmt, gimple_location (basis_stmt));
3160 }
3161
3162 if (dump_file && (dump_flags & TDF_DETAILS))
3163 {
3164 fputs ("Inserting initializer: ", dump_file);
3165 print_gimple_stmt (dump_file, init_stmt, 0, 0);
3166 }
3167 }
3168}
3169
9b92d12b
BS
3170/* Return TRUE iff all required increments for candidates feeding PHI
3171 are profitable to replace on behalf of candidate C. */
3172
3173static bool
355fe088 3174all_phi_incrs_profitable (slsr_cand_t c, gimple *phi)
9b92d12b
BS
3175{
3176 unsigned i;
3177 slsr_cand_t basis = lookup_cand (c->basis);
3178 slsr_cand_t phi_cand = base_cand_from_table (gimple_phi_result (phi));
3179
3180 for (i = 0; i < gimple_phi_num_args (phi); i++)
3181 {
3182 tree arg = gimple_phi_arg_def (phi, i);
3183
3184 if (!operand_equal_p (arg, phi_cand->base_expr, 0))
3185 {
355fe088 3186 gimple *arg_def = SSA_NAME_DEF_STMT (arg);
9b92d12b
BS
3187
3188 if (gimple_code (arg_def) == GIMPLE_PHI)
3189 {
3190 if (!all_phi_incrs_profitable (c, arg_def))
3191 return false;
3192 }
3193 else
3194 {
7bf55a70 3195 int j;
9b92d12b 3196 slsr_cand_t arg_cand = base_cand_from_table (arg);
807e902e 3197 widest_int increment = arg_cand->index - basis->index;
9b92d12b 3198
807e902e 3199 if (!address_arithmetic_p && wi::neg_p (increment))
9b92d12b
BS
3200 increment = -increment;
3201
3202 j = incr_vec_index (increment);
3203
3204 if (dump_file && (dump_flags & TDF_DETAILS))
3205 {
3206 fprintf (dump_file, " Conditional candidate %d, phi: ",
3207 c->cand_num);
3208 print_gimple_stmt (dump_file, phi, 0, 0);
3209 fputs (" increment: ", dump_file);
807e902e 3210 print_decs (increment, dump_file);
7bf55a70
BS
3211 if (j < 0)
3212 fprintf (dump_file,
3213 "\n Not replaced; incr_vec overflow.\n");
3214 else {
3215 fprintf (dump_file, "\n cost: %d\n", incr_vec[j].cost);
3216 if (profitable_increment_p (j))
3217 fputs (" Replacing...\n", dump_file);
3218 else
3219 fputs (" Not replaced.\n", dump_file);
3220 }
9b92d12b
BS
3221 }
3222
7bf55a70 3223 if (j < 0 || !profitable_increment_p (j))
9b92d12b
BS
3224 return false;
3225 }
3226 }
3227 }
3228
3229 return true;
3230}
3231
88ca9ea1
BS
3232/* Create a NOP_EXPR that copies FROM_EXPR into a new SSA name of
3233 type TO_TYPE, and insert it in front of the statement represented
3234 by candidate C. Use *NEW_VAR to create the new SSA name. Return
3235 the new SSA name. */
3236
3237static tree
a7a7d10e 3238introduce_cast_before_cand (slsr_cand_t c, tree to_type, tree from_expr)
88ca9ea1
BS
3239{
3240 tree cast_lhs;
538dd0b7 3241 gassign *cast_stmt;
88ca9ea1
BS
3242 gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
3243
a7a7d10e 3244 cast_lhs = make_temp_ssa_name (to_type, NULL, "slsr");
0d0e4a03 3245 cast_stmt = gimple_build_assign (cast_lhs, NOP_EXPR, from_expr);
88ca9ea1
BS
3246 gimple_set_location (cast_stmt, gimple_location (c->cand_stmt));
3247 gsi_insert_before (&gsi, cast_stmt, GSI_SAME_STMT);
3248
3249 if (dump_file && (dump_flags & TDF_DETAILS))
3250 {
3251 fputs (" Inserting: ", dump_file);
3252 print_gimple_stmt (dump_file, cast_stmt, 0, 0);
3253 }
3254
3255 return cast_lhs;
3256}
3257
3258/* Replace the RHS of the statement represented by candidate C with
3259 NEW_CODE, NEW_RHS1, and NEW_RHS2, provided that to do so doesn't
3260 leave C unchanged or just interchange its operands. The original
3261 operation and operands are in OLD_CODE, OLD_RHS1, and OLD_RHS2.
3262 If the replacement was made and we are doing a details dump,
3263 return the revised statement, else NULL. */
3264
355fe088 3265static gimple *
88ca9ea1
BS
3266replace_rhs_if_not_dup (enum tree_code new_code, tree new_rhs1, tree new_rhs2,
3267 enum tree_code old_code, tree old_rhs1, tree old_rhs2,
3268 slsr_cand_t c)
3269{
3270 if (new_code != old_code
3271 || ((!operand_equal_p (new_rhs1, old_rhs1, 0)
3272 || !operand_equal_p (new_rhs2, old_rhs2, 0))
3273 && (!operand_equal_p (new_rhs1, old_rhs2, 0)
3274 || !operand_equal_p (new_rhs2, old_rhs1, 0))))
3275 {
3276 gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
3277 gimple_assign_set_rhs_with_ops (&gsi, new_code, new_rhs1, new_rhs2);
3278 update_stmt (gsi_stmt (gsi));
bb0d2039 3279 c->cand_stmt = gsi_stmt (gsi);
88ca9ea1
BS
3280
3281 if (dump_file && (dump_flags & TDF_DETAILS))
3282 return gsi_stmt (gsi);
3283 }
3284
3285 else if (dump_file && (dump_flags & TDF_DETAILS))
3286 fputs (" (duplicate, not actually replacing)\n", dump_file);
3287
3288 return NULL;
3289}
3290
3291/* Strength-reduce the statement represented by candidate C by replacing
3292 it with an equivalent addition or subtraction. I is the index into
3293 the increment vector identifying C's increment. NEW_VAR is used to
3294 create a new SSA name if a cast needs to be introduced. BASIS_NAME
3295 is the rhs1 to use in creating the add/subtract. */
3296
3297static void
a7a7d10e 3298replace_one_candidate (slsr_cand_t c, unsigned i, tree basis_name)
88ca9ea1 3299{
355fe088 3300 gimple *stmt_to_print = NULL;
88ca9ea1
BS
3301 tree orig_rhs1, orig_rhs2;
3302 tree rhs2;
3303 enum tree_code orig_code, repl_code;
807e902e 3304 widest_int cand_incr;
88ca9ea1
BS
3305
3306 orig_code = gimple_assign_rhs_code (c->cand_stmt);
3307 orig_rhs1 = gimple_assign_rhs1 (c->cand_stmt);
3308 orig_rhs2 = gimple_assign_rhs2 (c->cand_stmt);
3309 cand_incr = cand_increment (c);
3310
3311 if (dump_file && (dump_flags & TDF_DETAILS))
3312 {
3313 fputs ("Replacing: ", dump_file);
3314 print_gimple_stmt (dump_file, c->cand_stmt, 0, 0);
3315 stmt_to_print = c->cand_stmt;
3316 }
3317
3318 if (address_arithmetic_p)
3319 repl_code = POINTER_PLUS_EXPR;
3320 else
3321 repl_code = PLUS_EXPR;
3322
3323 /* If the increment has an initializer T_0, replace the candidate
3324 statement with an add of the basis name and the initializer. */
3325 if (incr_vec[i].initializer)
3326 {
3327 tree init_type = TREE_TYPE (incr_vec[i].initializer);
3328 tree orig_type = TREE_TYPE (orig_rhs2);
3329
3330 if (types_compatible_p (orig_type, init_type))
3331 rhs2 = incr_vec[i].initializer;
3332 else
3333 rhs2 = introduce_cast_before_cand (c, orig_type,
a7a7d10e 3334 incr_vec[i].initializer);
88ca9ea1 3335
27bcd47c 3336 if (incr_vec[i].incr != cand_incr)
88ca9ea1
BS
3337 {
3338 gcc_assert (repl_code == PLUS_EXPR);
3339 repl_code = MINUS_EXPR;
3340 }
3341
3342 stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2,
3343 orig_code, orig_rhs1, orig_rhs2,
3344 c);
3345 }
3346
3347 /* Otherwise, the increment is one of -1, 0, and 1. Replace
3348 with a subtract of the stride from the basis name, a copy
3349 from the basis name, or an add of the stride to the basis
3350 name, respectively. It may be necessary to introduce a
3351 cast (or reuse an existing cast). */
807e902e 3352 else if (cand_incr == 1)
88ca9ea1
BS
3353 {
3354 tree stride_type = TREE_TYPE (c->stride);
3355 tree orig_type = TREE_TYPE (orig_rhs2);
3356
3357 if (types_compatible_p (orig_type, stride_type))
3358 rhs2 = c->stride;
3359 else
a7a7d10e 3360 rhs2 = introduce_cast_before_cand (c, orig_type, c->stride);
88ca9ea1
BS
3361
3362 stmt_to_print = replace_rhs_if_not_dup (repl_code, basis_name, rhs2,
3363 orig_code, orig_rhs1, orig_rhs2,
3364 c);
3365 }
3366
807e902e 3367 else if (cand_incr == -1)
88ca9ea1
BS
3368 {
3369 tree stride_type = TREE_TYPE (c->stride);
3370 tree orig_type = TREE_TYPE (orig_rhs2);
3371 gcc_assert (repl_code != POINTER_PLUS_EXPR);
3372
3373 if (types_compatible_p (orig_type, stride_type))
3374 rhs2 = c->stride;
3375 else
a7a7d10e 3376 rhs2 = introduce_cast_before_cand (c, orig_type, c->stride);
88ca9ea1
BS
3377
3378 if (orig_code != MINUS_EXPR
3379 || !operand_equal_p (basis_name, orig_rhs1, 0)
3380 || !operand_equal_p (rhs2, orig_rhs2, 0))
3381 {
3382 gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
3383 gimple_assign_set_rhs_with_ops (&gsi, MINUS_EXPR, basis_name, rhs2);
3384 update_stmt (gsi_stmt (gsi));
bb0d2039 3385 c->cand_stmt = gsi_stmt (gsi);
88ca9ea1
BS
3386
3387 if (dump_file && (dump_flags & TDF_DETAILS))
3388 stmt_to_print = gsi_stmt (gsi);
3389 }
3390 else if (dump_file && (dump_flags & TDF_DETAILS))
3391 fputs (" (duplicate, not actually replacing)\n", dump_file);
3392 }
3393
807e902e 3394 else if (cand_incr == 0)
88ca9ea1
BS
3395 {
3396 tree lhs = gimple_assign_lhs (c->cand_stmt);
3397 tree lhs_type = TREE_TYPE (lhs);
3398 tree basis_type = TREE_TYPE (basis_name);
3399
3400 if (types_compatible_p (lhs_type, basis_type))
3401 {
538dd0b7 3402 gassign *copy_stmt = gimple_build_assign (lhs, basis_name);
88ca9ea1
BS
3403 gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
3404 gimple_set_location (copy_stmt, gimple_location (c->cand_stmt));
3405 gsi_replace (&gsi, copy_stmt, false);
0100cd3f 3406 c->cand_stmt = copy_stmt;
88ca9ea1
BS
3407
3408 if (dump_file && (dump_flags & TDF_DETAILS))
3409 stmt_to_print = copy_stmt;
3410 }
3411 else
3412 {
3413 gimple_stmt_iterator gsi = gsi_for_stmt (c->cand_stmt);
0d0e4a03 3414 gassign *cast_stmt = gimple_build_assign (lhs, NOP_EXPR, basis_name);
88ca9ea1
BS
3415 gimple_set_location (cast_stmt, gimple_location (c->cand_stmt));
3416 gsi_replace (&gsi, cast_stmt, false);
0100cd3f 3417 c->cand_stmt = cast_stmt;
88ca9ea1
BS
3418
3419 if (dump_file && (dump_flags & TDF_DETAILS))
3420 stmt_to_print = cast_stmt;
3421 }
3422 }
3423 else
3424 gcc_unreachable ();
3425
3426 if (dump_file && (dump_flags & TDF_DETAILS) && stmt_to_print)
3427 {
3428 fputs ("With: ", dump_file);
3429 print_gimple_stmt (dump_file, stmt_to_print, 0, 0);
3430 fputs ("\n", dump_file);
3431 }
3432}
3433
3434/* For each candidate in the tree rooted at C, replace it with
3435 an increment if such has been shown to be profitable. */
3436
3437static void
3438replace_profitable_candidates (slsr_cand_t c)
3439{
3440 if (!cand_already_replaced (c))
3441 {
807e902e 3442 widest_int increment = cand_abs_increment (c);
88ca9ea1 3443 enum tree_code orig_code = gimple_assign_rhs_code (c->cand_stmt);
7bf55a70 3444 int i;
88ca9ea1
BS
3445
3446 i = incr_vec_index (increment);
3447
3448 /* Only process profitable increments. Nothing useful can be done
3449 to a cast or copy. */
7bf55a70
BS
3450 if (i >= 0
3451 && profitable_increment_p (i)
88ca9ea1 3452 && orig_code != MODIFY_EXPR
d822570f 3453 && !CONVERT_EXPR_CODE_P (orig_code))
88ca9ea1 3454 {
9b92d12b
BS
3455 if (phi_dependent_cand_p (c))
3456 {
355fe088 3457 gimple *phi = lookup_cand (c->def_phi)->cand_stmt;
9b92d12b
BS
3458
3459 if (all_phi_incrs_profitable (c, phi))
3460 {
3461 /* Look up the LHS SSA name from C's basis. This will be
3462 the RHS1 of the adds we will introduce to create new
3463 phi arguments. */
3464 slsr_cand_t basis = lookup_cand (c->basis);
3465 tree basis_name = gimple_assign_lhs (basis->cand_stmt);
3466
3467 /* Create a new phi statement that will represent C's true
3468 basis after the transformation is complete. */
3469 location_t loc = gimple_location (c->cand_stmt);
3470 tree name = create_phi_basis (c, phi, basis_name,
3471 loc, UNKNOWN_STRIDE);
3472
3473 /* Replace C with an add of the new basis phi and the
3474 increment. */
a7a7d10e 3475 replace_one_candidate (c, i, name);
9b92d12b
BS
3476 }
3477 }
3478 else
3479 {
3480 slsr_cand_t basis = lookup_cand (c->basis);
3481 tree basis_name = gimple_assign_lhs (basis->cand_stmt);
a7a7d10e 3482 replace_one_candidate (c, i, basis_name);
9b92d12b 3483 }
88ca9ea1
BS
3484 }
3485 }
3486
3487 if (c->sibling)
3488 replace_profitable_candidates (lookup_cand (c->sibling));
3489
3490 if (c->dependent)
3491 replace_profitable_candidates (lookup_cand (c->dependent));
3492}
3493\f
f9453c07
BS
3494/* Analyze costs of related candidates in the candidate vector,
3495 and make beneficial replacements. */
3496
3497static void
3498analyze_candidates_and_replace (void)
3499{
3500 unsigned i;
3501 slsr_cand_t c;
3502
3503 /* Each candidate that has a null basis and a non-null
3504 dependent is the root of a tree of related statements.
3505 Analyze each tree to determine a subset of those
3506 statements that can be replaced with maximum benefit. */
9771b263 3507 FOR_EACH_VEC_ELT (cand_vec, i, c)
f9453c07
BS
3508 {
3509 slsr_cand_t first_dep;
3510
3511 if (c->basis != 0 || c->dependent == 0)
3512 continue;
3513
3514 if (dump_file && (dump_flags & TDF_DETAILS))
3515 fprintf (dump_file, "\nProcessing dependency tree rooted at %d.\n",
3516 c->cand_num);
3517
3518 first_dep = lookup_cand (c->dependent);
3519
2749c8f6
BS
3520 /* If this is a chain of CAND_REFs, unconditionally replace
3521 each of them with a strength-reduced data reference. */
3522 if (c->kind == CAND_REF)
3523 replace_refs (c);
3524
9b92d12b
BS
3525 /* If the common stride of all related candidates is a known
3526 constant, each candidate without a phi-dependence can be
3527 profitably replaced. Each replaces a multiply by a single
3528 add, with the possibility that a feeding add also goes dead.
3529 A candidate with a phi-dependence is replaced only if the
3530 compensation code it requires is offset by the strength
3531 reduction savings. */
3532 else if (TREE_CODE (c->stride) == INTEGER_CST)
3533 replace_uncond_cands_and_profitable_phis (first_dep);
f9453c07 3534
88ca9ea1
BS
3535 /* When the stride is an SSA name, it may still be profitable
3536 to replace some or all of the dependent candidates, depending
3537 on whether the introduced increments can be reused, or are
3538 less expensive to calculate than the replaced statements. */
3539 else
3540 {
ef4bddc2 3541 machine_mode mode;
88ca9ea1
BS
3542 bool speed;
3543
3544 /* Determine whether we'll be generating pointer arithmetic
3545 when replacing candidates. */
3546 address_arithmetic_p = (c->kind == CAND_ADD
99cababb 3547 && POINTER_TYPE_P (c->cand_type));
88ca9ea1
BS
3548
3549 /* If all candidates have already been replaced under other
3550 interpretations, nothing remains to be done. */
4b847da9 3551 if (!count_candidates (c))
88ca9ea1
BS
3552 continue;
3553
3554 /* Construct an array of increments for this candidate chain. */
4b847da9 3555 incr_vec = XNEWVEC (incr_info, MAX_INCR_VEC_LEN);
88ca9ea1
BS
3556 incr_vec_len = 0;
3557 record_increments (c);
3558
3559 /* Determine which increments are profitable to replace. */
3560 mode = TYPE_MODE (TREE_TYPE (gimple_assign_lhs (c->cand_stmt)));
3561 speed = optimize_cands_for_speed_p (c);
3562 analyze_increments (first_dep, mode, speed);
3563
3564 /* Insert initializers of the form T_0 = stride * increment
3565 for use in profitable replacements. */
3566 insert_initializers (first_dep);
3567 dump_incr_vec ();
3568
3569 /* Perform the replacements. */
3570 replace_profitable_candidates (first_dep);
3571 free (incr_vec);
3572 }
f9453c07
BS
3573 }
3574}
3575
17795822
TS
3576namespace {
3577
3578const pass_data pass_data_strength_reduction =
be55bfe6
TS
3579{
3580 GIMPLE_PASS, /* type */
3581 "slsr", /* name */
3582 OPTGROUP_NONE, /* optinfo_flags */
be55bfe6
TS
3583 TV_GIMPLE_SLSR, /* tv_id */
3584 ( PROP_cfg | PROP_ssa ), /* properties_required */
3585 0, /* properties_provided */
3586 0, /* properties_destroyed */
3587 0, /* todo_flags_start */
3bea341f 3588 0, /* todo_flags_finish */
be55bfe6
TS
3589};
3590
17795822 3591class pass_strength_reduction : public gimple_opt_pass
be55bfe6
TS
3592{
3593public:
3594 pass_strength_reduction (gcc::context *ctxt)
3595 : gimple_opt_pass (pass_data_strength_reduction, ctxt)
3596 {}
3597
3598 /* opt_pass methods: */
3599 virtual bool gate (function *) { return flag_tree_slsr; }
3600 virtual unsigned int execute (function *);
3601
3602}; // class pass_strength_reduction
3603
3604unsigned
3605pass_strength_reduction::execute (function *fun)
f9453c07 3606{
f9453c07
BS
3607 /* Create the obstack where candidates will reside. */
3608 gcc_obstack_init (&cand_obstack);
3609
3610 /* Allocate the candidate vector. */
9771b263 3611 cand_vec.create (128);
f9453c07
BS
3612
3613 /* Allocate the mapping from statements to candidate indices. */
355fe088 3614 stmt_cand_map = new hash_map<gimple *, slsr_cand_t>;
f9453c07
BS
3615
3616 /* Create the obstack where candidate chains will reside. */
3617 gcc_obstack_init (&chain_obstack);
3618
3cfd4469 3619 /* Allocate the mapping from base expressions to candidate chains. */
c203e8a7 3620 base_cand_map = new hash_table<cand_chain_hasher> (500);
f9453c07 3621
96d75a2c 3622 /* Allocate the mapping from bases to alternative bases. */
b787e7a2 3623 alt_base_map = new hash_map<tree, tree>;
96d75a2c 3624
f9453c07
BS
3625 /* Initialize the loop optimizer. We need to detect flow across
3626 back edges, and this gives us dominator information as well. */
3627 loop_optimizer_init (AVOID_CFG_MODIFICATIONS);
3628
f9453c07
BS
3629 /* Walk the CFG in predominator order looking for strength reduction
3630 candidates. */
4d9192b5 3631 find_candidates_dom_walker (CDI_DOMINATORS)
be55bfe6 3632 .walk (fun->cfg->x_entry_block_ptr);
f9453c07
BS
3633
3634 if (dump_file && (dump_flags & TDF_DETAILS))
3635 {
3636 dump_cand_vec ();
3637 dump_cand_chains ();
3638 }
3639
b787e7a2 3640 delete alt_base_map;
96d75a2c
YZ
3641 free_affine_expand_cache (&name_expansions);
3642
f9453c07
BS
3643 /* Analyze costs and make appropriate replacements. */
3644 analyze_candidates_and_replace ();
3645
f9453c07 3646 loop_optimizer_finalize ();
c203e8a7
TS
3647 delete base_cand_map;
3648 base_cand_map = NULL;
f9453c07 3649 obstack_free (&chain_obstack, NULL);
b787e7a2 3650 delete stmt_cand_map;
9771b263 3651 cand_vec.release ();
f9453c07 3652 obstack_free (&cand_obstack, NULL);
f9453c07
BS
3653
3654 return 0;
3655}
3656
17795822
TS
3657} // anon namespace
3658
27a4cd48
DM
3659gimple_opt_pass *
3660make_pass_strength_reduction (gcc::context *ctxt)
3661{
3662 return new pass_strength_reduction (ctxt);
3663}