]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/tree-scalar-evolution.c
[multiple changes]
[thirdparty/gcc.git] / gcc / tree-scalar-evolution.c
CommitLineData
e9eb809d 1/* Scalar evolution detector.
818ab71a 2 Copyright (C) 2003-2016 Free Software Foundation, Inc.
e9eb809d
ZD
3 Contributed by Sebastian Pop <s.pop@laposte.net>
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
9dcd6f09 9Software Foundation; either version 3, or (at your option) any later
e9eb809d
ZD
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
9dcd6f09
NC
18along with GCC; see the file COPYING3. If not see
19<http://www.gnu.org/licenses/>. */
e9eb809d 20
b8698a0f
L
21/*
22 Description:
23
9baba81b
SP
24 This pass analyzes the evolution of scalar variables in loop
25 structures. The algorithm is based on the SSA representation,
26 and on the loop hierarchy tree. This algorithm is not based on
27 the notion of versions of a variable, as it was the case for the
28 previous implementations of the scalar evolution algorithm, but
29 it assumes that each defined name is unique.
30
31 The notation used in this file is called "chains of recurrences",
32 and has been proposed by Eugene Zima, Robert Van Engelen, and
33 others for describing induction variables in programs. For example
34 "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0
35 when entering in the loop_1 and has a step 2 in this loop, in other
36 words "for (b = 0; b < N; b+=2);". Note that the coefficients of
37 this chain of recurrence (or chrec [shrek]) can contain the name of
38 other variables, in which case they are called parametric chrecs.
39 For example, "b -> {a, +, 2}_1" means that the initial value of "b"
40 is the value of "a". In most of the cases these parametric chrecs
41 are fully instantiated before their use because symbolic names can
42 hide some difficult cases such as self-references described later
43 (see the Fibonacci example).
b8698a0f 44
9baba81b 45 A short sketch of the algorithm is:
b8698a0f 46
9baba81b
SP
47 Given a scalar variable to be analyzed, follow the SSA edge to
48 its definition:
b8698a0f 49
726a989a 50 - When the definition is a GIMPLE_ASSIGN: if the right hand side
9baba81b 51 (RHS) of the definition cannot be statically analyzed, the answer
b8698a0f 52 of the analyzer is: "don't know".
9baba81b
SP
53 Otherwise, for all the variables that are not yet analyzed in the
54 RHS, try to determine their evolution, and finally try to
55 evaluate the operation of the RHS that gives the evolution
56 function of the analyzed variable.
57
58 - When the definition is a condition-phi-node: determine the
59 evolution function for all the branches of the phi node, and
60 finally merge these evolutions (see chrec_merge).
61
62 - When the definition is a loop-phi-node: determine its initial
63 condition, that is the SSA edge defined in an outer loop, and
64 keep it symbolic. Then determine the SSA edges that are defined
65 in the body of the loop. Follow the inner edges until ending on
66 another loop-phi-node of the same analyzed loop. If the reached
67 loop-phi-node is not the starting loop-phi-node, then we keep
68 this definition under a symbolic form. If the reached
69 loop-phi-node is the same as the starting one, then we compute a
70 symbolic stride on the return path. The result is then the
71 symbolic chrec {initial_condition, +, symbolic_stride}_loop.
72
73 Examples:
b8698a0f 74
9baba81b 75 Example 1: Illustration of the basic algorithm.
b8698a0f 76
9baba81b
SP
77 | a = 3
78 | loop_1
79 | b = phi (a, c)
80 | c = b + 1
81 | if (c > 10) exit_loop
82 | endloop
b8698a0f 83
9baba81b
SP
84 Suppose that we want to know the number of iterations of the
85 loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We
86 ask the scalar evolution analyzer two questions: what's the
87 scalar evolution (scev) of "c", and what's the scev of "10". For
88 "10" the answer is "10" since it is a scalar constant. For the
89 scalar variable "c", it follows the SSA edge to its definition,
90 "c = b + 1", and then asks again what's the scev of "b".
91 Following the SSA edge, we end on a loop-phi-node "b = phi (a,
92 c)", where the initial condition is "a", and the inner loop edge
93 is "c". The initial condition is kept under a symbolic form (it
94 may be the case that the copy constant propagation has done its
95 work and we end with the constant "3" as one of the edges of the
96 loop-phi-node). The update edge is followed to the end of the
97 loop, and until reaching again the starting loop-phi-node: b -> c
98 -> b. At this point we have drawn a path from "b" to "b" from
99 which we compute the stride in the loop: in this example it is
100 "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now
101 that the scev for "b" is known, it is possible to compute the
102 scev for "c", that is "c -> {a + 1, +, 1}_1". In order to
103 determine the number of iterations in the loop_1, we have to
3f227a8c 104 instantiate_parameters (loop_1, {a + 1, +, 1}_1), that gives after some
9baba81b
SP
105 more analysis the scev {4, +, 1}_1, or in other words, this is
106 the function "f (x) = x + 4", where x is the iteration count of
107 the loop_1. Now we have to solve the inequality "x + 4 > 10",
108 and take the smallest iteration number for which the loop is
109 exited: x = 7. This loop runs from x = 0 to x = 7, and in total
110 there are 8 iterations. In terms of loop normalization, we have
111 created a variable that is implicitly defined, "x" or just "_1",
112 and all the other analyzed scalars of the loop are defined in
113 function of this variable:
b8698a0f 114
9baba81b
SP
115 a -> 3
116 b -> {3, +, 1}_1
117 c -> {4, +, 1}_1
b8698a0f
L
118
119 or in terms of a C program:
120
9baba81b
SP
121 | a = 3
122 | for (x = 0; x <= 7; x++)
123 | {
124 | b = x + 3
125 | c = x + 4
126 | }
b8698a0f 127
3f227a8c 128 Example 2a: Illustration of the algorithm on nested loops.
b8698a0f 129
9baba81b
SP
130 | loop_1
131 | a = phi (1, b)
132 | c = a + 2
133 | loop_2 10 times
134 | b = phi (c, d)
135 | d = b + 3
136 | endloop
137 | endloop
b8698a0f 138
9baba81b
SP
139 For analyzing the scalar evolution of "a", the algorithm follows
140 the SSA edge into the loop's body: "a -> b". "b" is an inner
b8698a0f
L
141 loop-phi-node, and its analysis as in Example 1, gives:
142
9baba81b
SP
143 b -> {c, +, 3}_2
144 d -> {c + 3, +, 3}_2
b8698a0f 145
9baba81b
SP
146 Following the SSA edge for the initial condition, we end on "c = a
147 + 2", and then on the starting loop-phi-node "a". From this point,
148 the loop stride is computed: back on "c = a + 2" we get a "+2" in
149 the loop_1, then on the loop-phi-node "b" we compute the overall
150 effect of the inner loop that is "b = c + 30", and we get a "+30"
151 in the loop_1. That means that the overall stride in loop_1 is
b8698a0f
L
152 equal to "+32", and the result is:
153
9baba81b
SP
154 a -> {1, +, 32}_1
155 c -> {3, +, 32}_1
3f227a8c
JS
156
157 Example 2b: Multivariate chains of recurrences.
158
159 | loop_1
160 | k = phi (0, k + 1)
161 | loop_2 4 times
162 | j = phi (0, j + 1)
163 | loop_3 4 times
164 | i = phi (0, i + 1)
165 | A[j + k] = ...
166 | endloop
167 | endloop
168 | endloop
169
170 Analyzing the access function of array A with
171 instantiate_parameters (loop_1, "j + k"), we obtain the
172 instantiation and the analysis of the scalar variables "j" and "k"
173 in loop_1. This leads to the scalar evolution {4, +, 1}_1: the end
174 value of loop_2 for "j" is 4, and the evolution of "k" in loop_1 is
175 {0, +, 1}_1. To obtain the evolution function in loop_3 and
176 instantiate the scalar variables up to loop_1, one has to use:
a213b219
SP
177 instantiate_scev (block_before_loop (loop_1), loop_3, "j + k").
178 The result of this call is {{0, +, 1}_1, +, 1}_2.
3f227a8c 179
9baba81b 180 Example 3: Higher degree polynomials.
b8698a0f 181
9baba81b
SP
182 | loop_1
183 | a = phi (2, b)
184 | c = phi (5, d)
185 | b = a + 1
186 | d = c + a
187 | endloop
b8698a0f 188
9baba81b
SP
189 a -> {2, +, 1}_1
190 b -> {3, +, 1}_1
191 c -> {5, +, a}_1
192 d -> {5 + a, +, a}_1
b8698a0f 193
3f227a8c
JS
194 instantiate_parameters (loop_1, {5, +, a}_1) -> {5, +, 2, +, 1}_1
195 instantiate_parameters (loop_1, {5 + a, +, a}_1) -> {7, +, 3, +, 1}_1
b8698a0f 196
9baba81b 197 Example 4: Lucas, Fibonacci, or mixers in general.
b8698a0f 198
9baba81b
SP
199 | loop_1
200 | a = phi (1, b)
201 | c = phi (3, d)
202 | b = c
203 | d = c + a
204 | endloop
b8698a0f 205
9baba81b
SP
206 a -> (1, c)_1
207 c -> {3, +, a}_1
b8698a0f 208
9baba81b
SP
209 The syntax "(1, c)_1" stands for a PEELED_CHREC that has the
210 following semantics: during the first iteration of the loop_1, the
211 variable contains the value 1, and then it contains the value "c".
212 Note that this syntax is close to the syntax of the loop-phi-node:
213 "a -> (1, c)_1" vs. "a = phi (1, c)".
b8698a0f 214
9baba81b
SP
215 The symbolic chrec representation contains all the semantics of the
216 original code. What is more difficult is to use this information.
b8698a0f 217
9baba81b 218 Example 5: Flip-flops, or exchangers.
b8698a0f 219
9baba81b
SP
220 | loop_1
221 | a = phi (1, b)
222 | c = phi (3, d)
223 | b = c
224 | d = a
225 | endloop
b8698a0f 226
9baba81b
SP
227 a -> (1, c)_1
228 c -> (3, a)_1
b8698a0f 229
9baba81b 230 Based on these symbolic chrecs, it is possible to refine this
b8698a0f
L
231 information into the more precise PERIODIC_CHRECs:
232
9baba81b
SP
233 a -> |1, 3|_1
234 c -> |3, 1|_1
b8698a0f 235
9baba81b 236 This transformation is not yet implemented.
b8698a0f 237
9baba81b 238 Further readings:
b8698a0f 239
9baba81b
SP
240 You can find a more detailed description of the algorithm in:
241 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf
242 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that
243 this is a preliminary report and some of the details of the
244 algorithm have changed. I'm working on a research report that
245 updates the description of the algorithms to reflect the design
246 choices used in this implementation.
b8698a0f 247
9baba81b
SP
248 A set of slides show a high level overview of the algorithm and run
249 an example through the scalar evolution analyzer:
250 http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf
251
252 The slides that I have presented at the GCC Summit'04 are available
253 at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf
254*/
255
e9eb809d
ZD
256#include "config.h"
257#include "system.h"
258#include "coretypes.h"
c7131fb2 259#include "backend.h"
957060b5 260#include "rtl.h"
cf2d1b38 261#include "tree.h"
c7131fb2 262#include "gimple.h"
c7131fb2 263#include "ssa.h"
957060b5 264#include "gimple-pretty-print.h"
c7131fb2 265#include "fold-const.h"
45b0be94 266#include "gimplify.h"
5be5c238 267#include "gimple-iterator.h"
18f429e2 268#include "gimplify-me.h"
442b4905 269#include "tree-cfg.h"
e28030cf
AM
270#include "tree-ssa-loop-ivopts.h"
271#include "tree-ssa-loop-manip.h"
272#include "tree-ssa-loop-niter.h"
442b4905 273#include "tree-ssa-loop.h"
7a300452 274#include "tree-ssa.h"
e9eb809d
ZD
275#include "cfgloop.h"
276#include "tree-chrec.h"
b83b5507 277#include "tree-affine.h"
e9eb809d 278#include "tree-scalar-evolution.h"
7ee2468b 279#include "dumpfile.h"
c59dabbe 280#include "params.h"
744730a4 281#include "tree-ssa-propagate.h"
19e51b40 282#include "gimple-fold.h"
9baba81b
SP
283
284static tree analyze_scalar_evolution_1 (struct loop *, tree, tree);
bef28ced
JL
285static tree analyze_scalar_evolution_for_address_of (struct loop *loop,
286 tree var);
9baba81b 287
a3cc13cc
RB
288/* The cached information about an SSA name with version NAME_VERSION,
289 claiming that below basic block with index INSTANTIATED_BELOW, the
290 value of the SSA name can be expressed as CHREC. */
9baba81b 291
907dadbd 292struct GTY((for_user)) scev_info_str {
a3cc13cc
RB
293 unsigned int name_version;
294 int instantiated_below;
9baba81b
SP
295 tree chrec;
296};
297
298/* Counters for the scev database. */
299static unsigned nb_set_scev = 0;
300static unsigned nb_get_scev = 0;
301
302/* The following trees are unique elements. Thus the comparison of
303 another element to these elements should be done on the pointer to
304 these trees, and not on their value. */
305
306/* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
307tree chrec_not_analyzed_yet;
308
309/* Reserved to the cases where the analyzer has detected an
310 undecidable property at compile time. */
311tree chrec_dont_know;
312
313/* When the analyzer has detected that a property will never
314 happen, then it qualifies it with chrec_known. */
315tree chrec_known;
316
ca752f39 317struct scev_info_hasher : ggc_ptr_hash<scev_info_str>
907dadbd
TS
318{
319 static hashval_t hash (scev_info_str *i);
320 static bool equal (const scev_info_str *a, const scev_info_str *b);
321};
322
323static GTY (()) hash_table<scev_info_hasher> *scalar_evolution_info;
9baba81b
SP
324
325\f
a213b219 326/* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */
9baba81b
SP
327
328static inline struct scev_info_str *
a213b219 329new_scev_info_str (basic_block instantiated_below, tree var)
9baba81b
SP
330{
331 struct scev_info_str *res;
b8698a0f 332
766090c2 333 res = ggc_alloc<scev_info_str> ();
a3cc13cc 334 res->name_version = SSA_NAME_VERSION (var);
9baba81b 335 res->chrec = chrec_not_analyzed_yet;
a3cc13cc 336 res->instantiated_below = instantiated_below->index;
a213b219 337
9baba81b
SP
338 return res;
339}
340
341/* Computes a hash function for database element ELT. */
342
907dadbd
TS
343hashval_t
344scev_info_hasher::hash (scev_info_str *elt)
9baba81b 345{
a3cc13cc 346 return elt->name_version ^ elt->instantiated_below;
9baba81b
SP
347}
348
349/* Compares database elements E1 and E2. */
350
907dadbd
TS
351bool
352scev_info_hasher::equal (const scev_info_str *elt1, const scev_info_str *elt2)
9baba81b 353{
a3cc13cc 354 return (elt1->name_version == elt2->name_version
a213b219 355 && elt1->instantiated_below == elt2->instantiated_below);
9baba81b
SP
356}
357
a213b219
SP
358/* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block.
359 A first query on VAR returns chrec_not_analyzed_yet. */
9baba81b
SP
360
361static tree *
a213b219 362find_var_scev_info (basic_block instantiated_below, tree var)
9baba81b
SP
363{
364 struct scev_info_str *res;
365 struct scev_info_str tmp;
9baba81b 366
a3cc13cc
RB
367 tmp.name_version = SSA_NAME_VERSION (var);
368 tmp.instantiated_below = instantiated_below->index;
907dadbd 369 scev_info_str **slot = scalar_evolution_info->find_slot (&tmp, INSERT);
9baba81b
SP
370
371 if (!*slot)
a213b219 372 *slot = new_scev_info_str (instantiated_below, var);
907dadbd 373 res = *slot;
9baba81b
SP
374
375 return &res->chrec;
376}
377
9baba81b
SP
378/* Return true when CHREC contains symbolic names defined in
379 LOOP_NB. */
380
b8698a0f 381bool
ed7a4b4b 382chrec_contains_symbols_defined_in_loop (const_tree chrec, unsigned loop_nb)
9baba81b 383{
5039610b
SL
384 int i, n;
385
9baba81b
SP
386 if (chrec == NULL_TREE)
387 return false;
388
ad6003f2 389 if (is_gimple_min_invariant (chrec))
9baba81b
SP
390 return false;
391
9baba81b
SP
392 if (TREE_CODE (chrec) == SSA_NAME)
393 {
355fe088 394 gimple *def;
492e5456
SP
395 loop_p def_loop, loop;
396
397 if (SSA_NAME_IS_DEFAULT_DEF (chrec))
398 return false;
399
400 def = SSA_NAME_DEF_STMT (chrec);
401 def_loop = loop_containing_stmt (def);
0fc822d0 402 loop = get_loop (cfun, loop_nb);
9baba81b
SP
403
404 if (def_loop == NULL)
405 return false;
406
407 if (loop == def_loop || flow_loop_nested_p (loop, def_loop))
408 return true;
409
410 return false;
411 }
412
5039610b
SL
413 n = TREE_OPERAND_LENGTH (chrec);
414 for (i = 0; i < n; i++)
b8698a0f 415 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, i),
5039610b
SL
416 loop_nb))
417 return true;
418 return false;
9baba81b
SP
419}
420
421/* Return true when PHI is a loop-phi-node. */
422
423static bool
355fe088 424loop_phi_node_p (gimple *phi)
9baba81b
SP
425{
426 /* The implementation of this function is based on the following
427 property: "all the loop-phi-nodes of a loop are contained in the
428 loop's header basic block". */
429
726a989a 430 return loop_containing_stmt (phi)->header == gimple_bb (phi);
9baba81b
SP
431}
432
433/* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
434 In general, in the case of multivariate evolutions we want to get
435 the evolution in different loops. LOOP specifies the level for
436 which to get the evolution.
b8698a0f 437
9baba81b 438 Example:
b8698a0f 439
9baba81b
SP
440 | for (j = 0; j < 100; j++)
441 | {
442 | for (k = 0; k < 100; k++)
443 | {
b8698a0f 444 | i = k + j; - Here the value of i is a function of j, k.
9baba81b 445 | }
b8698a0f 446 | ... = i - Here the value of i is a function of j.
9baba81b 447 | }
b8698a0f
L
448 | ... = i - Here the value of i is a scalar.
449
450 Example:
451
9baba81b
SP
452 | i_0 = ...
453 | loop_1 10 times
454 | i_1 = phi (i_0, i_2)
455 | i_2 = i_1 + 2
456 | endloop
b8698a0f 457
9baba81b
SP
458 This loop has the same effect as:
459 LOOP_1 has the same effect as:
b8698a0f 460
9baba81b 461 | i_1 = i_0 + 20
b8698a0f
L
462
463 The overall effect of the loop, "i_0 + 20" in the previous example,
464 is obtained by passing in the parameters: LOOP = 1,
9baba81b
SP
465 EVOLUTION_FN = {i_0, +, 2}_1.
466*/
b8698a0f 467
42e6eec5 468tree
9baba81b
SP
469compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn)
470{
471 bool val = false;
472
473 if (evolution_fn == chrec_dont_know)
474 return chrec_dont_know;
475
476 else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC)
477 {
677cc14d
ZD
478 struct loop *inner_loop = get_chrec_loop (evolution_fn);
479
480 if (inner_loop == loop
481 || flow_loop_nested_p (loop, inner_loop))
9baba81b 482 {
a14865db 483 tree nb_iter = number_of_latch_executions (inner_loop);
9baba81b
SP
484
485 if (nb_iter == chrec_dont_know)
486 return chrec_dont_know;
487 else
488 {
489 tree res;
490
9baba81b
SP
491 /* evolution_fn is the evolution function in LOOP. Get
492 its value in the nb_iter-th iteration. */
493 res = chrec_apply (inner_loop->num, evolution_fn, nb_iter);
42e6eec5
SP
494
495 if (chrec_contains_symbols_defined_in_loop (res, loop->num))
496 res = instantiate_parameters (loop, res);
497
8c27b7d4 498 /* Continue the computation until ending on a parent of LOOP. */
9baba81b
SP
499 return compute_overall_effect_of_inner_loop (loop, res);
500 }
501 }
502 else
503 return evolution_fn;
504 }
b8698a0f 505
9baba81b
SP
506 /* If the evolution function is an invariant, there is nothing to do. */
507 else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val)
508 return evolution_fn;
b8698a0f 509
9baba81b
SP
510 else
511 return chrec_dont_know;
512}
513
9baba81b
SP
514/* Associate CHREC to SCALAR. */
515
516static void
a213b219 517set_scalar_evolution (basic_block instantiated_below, tree scalar, tree chrec)
9baba81b
SP
518{
519 tree *scalar_info;
b8698a0f 520
9baba81b
SP
521 if (TREE_CODE (scalar) != SSA_NAME)
522 return;
523
a213b219 524 scalar_info = find_var_scev_info (instantiated_below, scalar);
b8698a0f 525
9baba81b
SP
526 if (dump_file)
527 {
dfedbe40 528 if (dump_flags & TDF_SCEV)
9baba81b
SP
529 {
530 fprintf (dump_file, "(set_scalar_evolution \n");
a213b219
SP
531 fprintf (dump_file, " instantiated_below = %d \n",
532 instantiated_below->index);
9baba81b
SP
533 fprintf (dump_file, " (scalar = ");
534 print_generic_expr (dump_file, scalar, 0);
535 fprintf (dump_file, ")\n (scalar_evolution = ");
536 print_generic_expr (dump_file, chrec, 0);
537 fprintf (dump_file, "))\n");
538 }
539 if (dump_flags & TDF_STATS)
540 nb_set_scev++;
541 }
b8698a0f 542
9baba81b
SP
543 *scalar_info = chrec;
544}
545
a213b219
SP
546/* Retrieve the chrec associated to SCALAR instantiated below
547 INSTANTIATED_BELOW block. */
9baba81b
SP
548
549static tree
a213b219 550get_scalar_evolution (basic_block instantiated_below, tree scalar)
9baba81b
SP
551{
552 tree res;
b8698a0f 553
9baba81b
SP
554 if (dump_file)
555 {
dfedbe40 556 if (dump_flags & TDF_SCEV)
9baba81b
SP
557 {
558 fprintf (dump_file, "(get_scalar_evolution \n");
559 fprintf (dump_file, " (scalar = ");
560 print_generic_expr (dump_file, scalar, 0);
561 fprintf (dump_file, ")\n");
562 }
563 if (dump_flags & TDF_STATS)
564 nb_get_scev++;
565 }
b8698a0f 566
9baba81b
SP
567 switch (TREE_CODE (scalar))
568 {
569 case SSA_NAME:
a213b219 570 res = *find_var_scev_info (instantiated_below, scalar);
9baba81b
SP
571 break;
572
573 case REAL_CST:
325217ed 574 case FIXED_CST:
9baba81b
SP
575 case INTEGER_CST:
576 res = scalar;
577 break;
578
579 default:
580 res = chrec_not_analyzed_yet;
581 break;
582 }
b8698a0f 583
dfedbe40 584 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
585 {
586 fprintf (dump_file, " (scalar_evolution = ");
587 print_generic_expr (dump_file, res, 0);
588 fprintf (dump_file, "))\n");
589 }
b8698a0f 590
9baba81b
SP
591 return res;
592}
593
594/* Helper function for add_to_evolution. Returns the evolution
595 function for an assignment of the form "a = b + c", where "a" and
596 "b" are on the strongly connected component. CHREC_BEFORE is the
597 information that we already have collected up to this point.
b8698a0f
L
598 TO_ADD is the evolution of "c".
599
9baba81b
SP
600 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
601 evolution the expression TO_ADD, otherwise construct an evolution
602 part for this loop. */
603
604static tree
e2157b49 605add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add,
355fe088 606 gimple *at_stmt)
9baba81b 607{
e2157b49 608 tree type, left, right;
0fc822d0 609 struct loop *loop = get_loop (cfun, loop_nb), *chloop;
e2157b49 610
9baba81b
SP
611 switch (TREE_CODE (chrec_before))
612 {
613 case POLYNOMIAL_CHREC:
677cc14d
ZD
614 chloop = get_chrec_loop (chrec_before);
615 if (chloop == loop
616 || flow_loop_nested_p (chloop, loop))
9baba81b
SP
617 {
618 unsigned var;
e2157b49
SP
619
620 type = chrec_type (chrec_before);
b8698a0f 621
9baba81b 622 /* When there is no evolution part in this loop, build it. */
677cc14d 623 if (chloop != loop)
9baba81b
SP
624 {
625 var = loop_nb;
626 left = chrec_before;
7e0923cd
SP
627 right = SCALAR_FLOAT_TYPE_P (type)
628 ? build_real (type, dconst0)
629 : build_int_cst (type, 0);
9baba81b
SP
630 }
631 else
632 {
633 var = CHREC_VARIABLE (chrec_before);
634 left = CHREC_LEFT (chrec_before);
635 right = CHREC_RIGHT (chrec_before);
636 }
637
e2157b49 638 to_add = chrec_convert (type, to_add, at_stmt);
5be014d5
AP
639 right = chrec_convert_rhs (type, right, at_stmt);
640 right = chrec_fold_plus (chrec_type (right), right, to_add);
e2157b49 641 return build_polynomial_chrec (var, left, right);
9baba81b
SP
642 }
643 else
e2157b49 644 {
677cc14d
ZD
645 gcc_assert (flow_loop_nested_p (loop, chloop));
646
e2157b49
SP
647 /* Search the evolution in LOOP_NB. */
648 left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before),
649 to_add, at_stmt);
650 right = CHREC_RIGHT (chrec_before);
5be014d5 651 right = chrec_convert_rhs (chrec_type (left), right, at_stmt);
e2157b49
SP
652 return build_polynomial_chrec (CHREC_VARIABLE (chrec_before),
653 left, right);
654 }
b8698a0f 655
9baba81b
SP
656 default:
657 /* These nodes do not depend on a loop. */
658 if (chrec_before == chrec_dont_know)
659 return chrec_dont_know;
e2157b49
SP
660
661 left = chrec_before;
5be014d5 662 right = chrec_convert_rhs (chrec_type (left), to_add, at_stmt);
e2157b49 663 return build_polynomial_chrec (loop_nb, left, right);
9baba81b
SP
664 }
665}
666
667/* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
b8698a0f
L
668 of LOOP_NB.
669
9baba81b
SP
670 Description (provided for completeness, for those who read code in
671 a plane, and for my poor 62 bytes brain that would have forgotten
672 all this in the next two or three months):
b8698a0f 673
9baba81b
SP
674 The algorithm of translation of programs from the SSA representation
675 into the chrecs syntax is based on a pattern matching. After having
676 reconstructed the overall tree expression for a loop, there are only
677 two cases that can arise:
b8698a0f 678
9baba81b
SP
679 1. a = loop-phi (init, a + expr)
680 2. a = loop-phi (init, expr)
b8698a0f 681
9baba81b
SP
682 where EXPR is either a scalar constant with respect to the analyzed
683 loop (this is a degree 0 polynomial), or an expression containing
684 other loop-phi definitions (these are higher degree polynomials).
b8698a0f 685
9baba81b 686 Examples:
b8698a0f
L
687
688 1.
9baba81b
SP
689 | init = ...
690 | loop_1
691 | a = phi (init, a + 5)
692 | endloop
b8698a0f
L
693
694 2.
9baba81b
SP
695 | inita = ...
696 | initb = ...
697 | loop_1
698 | a = phi (inita, 2 * b + 3)
699 | b = phi (initb, b + 1)
700 | endloop
b8698a0f
L
701
702 For the first case, the semantics of the SSA representation is:
703
9baba81b 704 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
b8698a0f 705
9baba81b
SP
706 that is, there is a loop index "x" that determines the scalar value
707 of the variable during the loop execution. During the first
708 iteration, the value is that of the initial condition INIT, while
709 during the subsequent iterations, it is the sum of the initial
710 condition with the sum of all the values of EXPR from the initial
b8698a0f
L
711 iteration to the before last considered iteration.
712
9baba81b 713 For the second case, the semantics of the SSA program is:
b8698a0f 714
9baba81b
SP
715 | a (x) = init, if x = 0;
716 | expr (x - 1), otherwise.
b8698a0f 717
9baba81b 718 The second case corresponds to the PEELED_CHREC, whose syntax is
b8698a0f
L
719 close to the syntax of a loop-phi-node:
720
9baba81b 721 | phi (init, expr) vs. (init, expr)_x
b8698a0f 722
9baba81b 723 The proof of the translation algorithm for the first case is a
b8698a0f
L
724 proof by structural induction based on the degree of EXPR.
725
9baba81b
SP
726 Degree 0:
727 When EXPR is a constant with respect to the analyzed loop, or in
728 other words when EXPR is a polynomial of degree 0, the evolution of
729 the variable A in the loop is an affine function with an initial
730 condition INIT, and a step EXPR. In order to show this, we start
731 from the semantics of the SSA representation:
b8698a0f 732
9baba81b 733 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
b8698a0f 734
9baba81b 735 and since "expr (j)" is a constant with respect to "j",
b8698a0f
L
736
737 f (x) = init + x * expr
738
9baba81b
SP
739 Finally, based on the semantics of the pure sum chrecs, by
740 identification we get the corresponding chrecs syntax:
b8698a0f
L
741
742 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
9baba81b 743 f (x) -> {init, +, expr}_x
b8698a0f 744
9baba81b
SP
745 Higher degree:
746 Suppose that EXPR is a polynomial of degree N with respect to the
747 analyzed loop_x for which we have already determined that it is
748 written under the chrecs syntax:
b8698a0f 749
9baba81b 750 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
b8698a0f 751
9baba81b 752 We start from the semantics of the SSA program:
b8698a0f 753
9baba81b
SP
754 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
755 |
b8698a0f 756 | f (x) = init + \sum_{j = 0}^{x - 1}
9baba81b
SP
757 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
758 |
b8698a0f
L
759 | f (x) = init + \sum_{j = 0}^{x - 1}
760 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
9baba81b 761 |
b8698a0f
L
762 | f (x) = init + \sum_{k = 0}^{n - 1}
763 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
9baba81b 764 |
b8698a0f
L
765 | f (x) = init + \sum_{k = 0}^{n - 1}
766 | (b_k * \binom{x}{k + 1})
9baba81b 767 |
b8698a0f
L
768 | f (x) = init + b_0 * \binom{x}{1} + ...
769 | + b_{n-1} * \binom{x}{n}
9baba81b 770 |
b8698a0f
L
771 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
772 | + b_{n-1} * \binom{x}{n}
9baba81b 773 |
b8698a0f 774
9baba81b 775 And finally from the definition of the chrecs syntax, we identify:
b8698a0f
L
776 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
777
9baba81b
SP
778 This shows the mechanism that stands behind the add_to_evolution
779 function. An important point is that the use of symbolic
780 parameters avoids the need of an analysis schedule.
b8698a0f 781
9baba81b 782 Example:
b8698a0f 783
9baba81b
SP
784 | inita = ...
785 | initb = ...
b8698a0f 786 | loop_1
9baba81b
SP
787 | a = phi (inita, a + 2 + b)
788 | b = phi (initb, b + 1)
789 | endloop
b8698a0f 790
9baba81b 791 When analyzing "a", the algorithm keeps "b" symbolically:
b8698a0f 792
9baba81b 793 | a -> {inita, +, 2 + b}_1
b8698a0f 794
9baba81b 795 Then, after instantiation, the analyzer ends on the evolution:
b8698a0f 796
9baba81b
SP
797 | a -> {inita, +, 2 + initb, +, 1}_1
798
799*/
800
b8698a0f 801static tree
e2157b49 802add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code,
355fe088 803 tree to_add, gimple *at_stmt)
9baba81b
SP
804{
805 tree type = chrec_type (to_add);
806 tree res = NULL_TREE;
b8698a0f 807
9baba81b
SP
808 if (to_add == NULL_TREE)
809 return chrec_before;
b8698a0f 810
9baba81b
SP
811 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
812 instantiated at this point. */
813 if (TREE_CODE (to_add) == POLYNOMIAL_CHREC)
814 /* This should not happen. */
815 return chrec_dont_know;
b8698a0f 816
dfedbe40 817 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
818 {
819 fprintf (dump_file, "(add_to_evolution \n");
820 fprintf (dump_file, " (loop_nb = %d)\n", loop_nb);
821 fprintf (dump_file, " (chrec_before = ");
822 print_generic_expr (dump_file, chrec_before, 0);
823 fprintf (dump_file, ")\n (to_add = ");
824 print_generic_expr (dump_file, to_add, 0);
825 fprintf (dump_file, ")\n");
826 }
827
828 if (code == MINUS_EXPR)
9d2b0e12
VR
829 to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type)
830 ? build_real (type, dconstm1)
831 : build_int_cst_type (type, -1));
9baba81b 832
e2157b49 833 res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt);
9baba81b 834
dfedbe40 835 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
836 {
837 fprintf (dump_file, " (res = ");
838 print_generic_expr (dump_file, res, 0);
839 fprintf (dump_file, "))\n");
840 }
841
842 return res;
843}
844
9baba81b
SP
845\f
846
847/* This section selects the loops that will be good candidates for the
848 scalar evolution analysis. For the moment, greedily select all the
849 loop nests we could analyze. */
850
9baba81b
SP
851/* For a loop with a single exit edge, return the COND_EXPR that
852 guards the exit edge. If the expression is too difficult to
853 analyze, then give up. */
854
538dd0b7 855gcond *
22ea9ec0 856get_loop_exit_condition (const struct loop *loop)
9baba81b 857{
538dd0b7 858 gcond *res = NULL;
ac8f6c69 859 edge exit_edge = single_exit (loop);
b8698a0f 860
dfedbe40 861 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b 862 fprintf (dump_file, "(get_loop_exit_condition \n ");
b8698a0f 863
82b85a85 864 if (exit_edge)
9baba81b 865 {
355fe088 866 gimple *stmt;
b8698a0f 867
726a989a 868 stmt = last_stmt (exit_edge->src);
538dd0b7
DM
869 if (gcond *cond_stmt = dyn_cast <gcond *> (stmt))
870 res = cond_stmt;
9baba81b 871 }
b8698a0f 872
dfedbe40 873 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b 874 {
726a989a 875 print_gimple_stmt (dump_file, res, 0, 0);
9baba81b
SP
876 fprintf (dump_file, ")\n");
877 }
b8698a0f 878
9baba81b
SP
879 return res;
880}
881
9baba81b
SP
882\f
883/* Depth first search algorithm. */
884
a79683d5 885enum t_bool {
c59dabbe
SP
886 t_false,
887 t_true,
888 t_dont_know
a79683d5 889};
c59dabbe
SP
890
891
355fe088 892static t_bool follow_ssa_edge (struct loop *loop, gimple *, gphi *,
538dd0b7 893 tree *, int);
9baba81b 894
726a989a 895/* Follow the ssa edge into the binary expression RHS0 CODE RHS1.
9baba81b
SP
896 Return true if the strongly connected component has been found. */
897
c59dabbe 898static t_bool
355fe088 899follow_ssa_edge_binary (struct loop *loop, gimple *at_stmt,
726a989a 900 tree type, tree rhs0, enum tree_code code, tree rhs1,
538dd0b7
DM
901 gphi *halting_phi, tree *evolution_of_loop,
902 int limit)
9baba81b 903{
c59dabbe 904 t_bool res = t_false;
b2a93c0a 905 tree evol;
726a989a 906
5be014d5 907 switch (code)
9baba81b 908 {
5be014d5 909 case POINTER_PLUS_EXPR:
9baba81b 910 case PLUS_EXPR:
9baba81b
SP
911 if (TREE_CODE (rhs0) == SSA_NAME)
912 {
913 if (TREE_CODE (rhs1) == SSA_NAME)
914 {
b8698a0f 915 /* Match an assignment under the form:
9baba81b 916 "a = b + c". */
b8698a0f 917
9e824336
ZD
918 /* We want only assignments of form "name + name" contribute to
919 LIMIT, as the other cases do not necessarily contribute to
920 the complexity of the expression. */
921 limit++;
922
b2a93c0a 923 evol = *evolution_of_loop;
b9b79ba4 924 evol = add_to_evolution
b8698a0f
L
925 (loop->num,
926 chrec_convert (type, evol, at_stmt),
5be014d5 927 code, rhs1, at_stmt);
b9b79ba4
RB
928 res = follow_ssa_edge
929 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, &evol, limit);
930 if (res == t_true)
931 *evolution_of_loop = evol;
c59dabbe 932 else if (res == t_false)
9baba81b 933 {
b9b79ba4
RB
934 *evolution_of_loop = add_to_evolution
935 (loop->num,
936 chrec_convert (type, *evolution_of_loop, at_stmt),
937 code, rhs0, at_stmt);
b8698a0f
L
938 res = follow_ssa_edge
939 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
c59dabbe 940 evolution_of_loop, limit);
c59dabbe 941 if (res == t_true)
b9b79ba4 942 ;
c59dabbe
SP
943 else if (res == t_dont_know)
944 *evolution_of_loop = chrec_dont_know;
9baba81b 945 }
c59dabbe
SP
946
947 else if (res == t_dont_know)
948 *evolution_of_loop = chrec_dont_know;
9baba81b 949 }
b8698a0f 950
9baba81b
SP
951 else
952 {
b8698a0f 953 /* Match an assignment under the form:
9baba81b 954 "a = b + ...". */
b9b79ba4
RB
955 *evolution_of_loop = add_to_evolution
956 (loop->num, chrec_convert (type, *evolution_of_loop,
957 at_stmt),
958 code, rhs1, at_stmt);
b8698a0f
L
959 res = follow_ssa_edge
960 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
c59dabbe
SP
961 evolution_of_loop, limit);
962 if (res == t_true)
b9b79ba4 963 ;
c59dabbe
SP
964 else if (res == t_dont_know)
965 *evolution_of_loop = chrec_dont_know;
9baba81b
SP
966 }
967 }
b8698a0f 968
9baba81b
SP
969 else if (TREE_CODE (rhs1) == SSA_NAME)
970 {
b8698a0f 971 /* Match an assignment under the form:
9baba81b 972 "a = ... + c". */
b9b79ba4
RB
973 *evolution_of_loop = add_to_evolution
974 (loop->num, chrec_convert (type, *evolution_of_loop,
975 at_stmt),
976 code, rhs0, at_stmt);
b8698a0f
L
977 res = follow_ssa_edge
978 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
c59dabbe
SP
979 evolution_of_loop, limit);
980 if (res == t_true)
b9b79ba4 981 ;
c59dabbe
SP
982 else if (res == t_dont_know)
983 *evolution_of_loop = chrec_dont_know;
9baba81b
SP
984 }
985
986 else
b8698a0f 987 /* Otherwise, match an assignment under the form:
9baba81b
SP
988 "a = ... + ...". */
989 /* And there is nothing to do. */
c59dabbe 990 res = t_false;
9baba81b 991 break;
b8698a0f 992
9baba81b
SP
993 case MINUS_EXPR:
994 /* This case is under the form "opnd0 = rhs0 - rhs1". */
9baba81b 995 if (TREE_CODE (rhs0) == SSA_NAME)
9baba81b 996 {
b8698a0f 997 /* Match an assignment under the form:
f8e9d512 998 "a = b - ...". */
9e824336
ZD
999
1000 /* We want only assignments of form "name - name" contribute to
1001 LIMIT, as the other cases do not necessarily contribute to
1002 the complexity of the expression. */
1003 if (TREE_CODE (rhs1) == SSA_NAME)
1004 limit++;
1005
b9b79ba4
RB
1006 *evolution_of_loop = add_to_evolution
1007 (loop->num, chrec_convert (type, *evolution_of_loop, at_stmt),
1008 MINUS_EXPR, rhs1, at_stmt);
b8698a0f 1009 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
c59dabbe
SP
1010 evolution_of_loop, limit);
1011 if (res == t_true)
b9b79ba4 1012 ;
c59dabbe
SP
1013 else if (res == t_dont_know)
1014 *evolution_of_loop = chrec_dont_know;
9baba81b 1015 }
9baba81b 1016 else
b8698a0f 1017 /* Otherwise, match an assignment under the form:
9baba81b
SP
1018 "a = ... - ...". */
1019 /* And there is nothing to do. */
c59dabbe 1020 res = t_false;
9baba81b 1021 break;
726a989a
RB
1022
1023 default:
1024 res = t_false;
1025 }
1026
1027 return res;
1028}
b8698a0f 1029
726a989a
RB
1030/* Follow the ssa edge into the expression EXPR.
1031 Return true if the strongly connected component has been found. */
1032
1033static t_bool
355fe088 1034follow_ssa_edge_expr (struct loop *loop, gimple *at_stmt, tree expr,
538dd0b7
DM
1035 gphi *halting_phi, tree *evolution_of_loop,
1036 int limit)
726a989a 1037{
5aefc6a0
EB
1038 enum tree_code code = TREE_CODE (expr);
1039 tree type = TREE_TYPE (expr), rhs0, rhs1;
1040 t_bool res;
1041
726a989a 1042 /* The EXPR is one of the following cases:
b8698a0f 1043 - an SSA_NAME,
726a989a 1044 - an INTEGER_CST,
b8698a0f
L
1045 - a PLUS_EXPR,
1046 - a POINTER_PLUS_EXPR,
726a989a
RB
1047 - a MINUS_EXPR,
1048 - an ASSERT_EXPR,
1049 - other cases are not yet handled. */
5aefc6a0 1050
726a989a
RB
1051 switch (code)
1052 {
5aefc6a0 1053 CASE_CONVERT:
726a989a
RB
1054 /* This assignment is under the form "a_1 = (cast) rhs. */
1055 res = follow_ssa_edge_expr (loop, at_stmt, TREE_OPERAND (expr, 0),
1056 halting_phi, evolution_of_loop, limit);
1057 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, at_stmt);
1058 break;
1059
1060 case INTEGER_CST:
1061 /* This assignment is under the form "a_1 = 7". */
1062 res = t_false;
1063 break;
5aefc6a0 1064
726a989a
RB
1065 case SSA_NAME:
1066 /* This assignment is under the form: "a_1 = b_2". */
b8698a0f 1067 res = follow_ssa_edge
726a989a
RB
1068 (loop, SSA_NAME_DEF_STMT (expr), halting_phi, evolution_of_loop, limit);
1069 break;
5aefc6a0 1070
726a989a
RB
1071 case POINTER_PLUS_EXPR:
1072 case PLUS_EXPR:
1073 case MINUS_EXPR:
1074 /* This case is under the form "rhs0 +- rhs1". */
1075 rhs0 = TREE_OPERAND (expr, 0);
1076 rhs1 = TREE_OPERAND (expr, 1);
5aefc6a0
EB
1077 type = TREE_TYPE (rhs0);
1078 STRIP_USELESS_TYPE_CONVERSION (rhs0);
1079 STRIP_USELESS_TYPE_CONVERSION (rhs1);
1080 res = follow_ssa_edge_binary (loop, at_stmt, type, rhs0, code, rhs1,
1081 halting_phi, evolution_of_loop, limit);
1082 break;
726a989a 1083
70f34814
RG
1084 case ADDR_EXPR:
1085 /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */
1086 if (TREE_CODE (TREE_OPERAND (expr, 0)) == MEM_REF)
1087 {
1088 expr = TREE_OPERAND (expr, 0);
1089 rhs0 = TREE_OPERAND (expr, 0);
1090 rhs1 = TREE_OPERAND (expr, 1);
1091 type = TREE_TYPE (rhs0);
1092 STRIP_USELESS_TYPE_CONVERSION (rhs0);
1093 STRIP_USELESS_TYPE_CONVERSION (rhs1);
1094 res = follow_ssa_edge_binary (loop, at_stmt, type,
1095 rhs0, POINTER_PLUS_EXPR, rhs1,
1096 halting_phi, evolution_of_loop, limit);
1097 }
1098 else
1099 res = t_false;
1100 break;
1101
0bca51f0 1102 case ASSERT_EXPR:
5aefc6a0
EB
1103 /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>"
1104 It must be handled as a copy assignment of the form a_1 = a_2. */
1105 rhs0 = ASSERT_EXPR_VAR (expr);
1106 if (TREE_CODE (rhs0) == SSA_NAME)
1107 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0),
1108 halting_phi, evolution_of_loop, limit);
1109 else
1110 res = t_false;
1111 break;
0bca51f0 1112
9baba81b 1113 default:
c59dabbe 1114 res = t_false;
9baba81b
SP
1115 break;
1116 }
5aefc6a0 1117
9baba81b
SP
1118 return res;
1119}
1120
726a989a
RB
1121/* Follow the ssa edge into the right hand side of an assignment STMT.
1122 Return true if the strongly connected component has been found. */
1123
1124static t_bool
355fe088 1125follow_ssa_edge_in_rhs (struct loop *loop, gimple *stmt,
538dd0b7
DM
1126 gphi *halting_phi, tree *evolution_of_loop,
1127 int limit)
726a989a 1128{
726a989a 1129 enum tree_code code = gimple_assign_rhs_code (stmt);
5aefc6a0
EB
1130 tree type = gimple_expr_type (stmt), rhs1, rhs2;
1131 t_bool res;
726a989a 1132
5aefc6a0 1133 switch (code)
726a989a 1134 {
5aefc6a0
EB
1135 CASE_CONVERT:
1136 /* This assignment is under the form "a_1 = (cast) rhs. */
1137 res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1138 halting_phi, evolution_of_loop, limit);
1139 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, stmt);
1140 break;
1141
1142 case POINTER_PLUS_EXPR:
1143 case PLUS_EXPR:
1144 case MINUS_EXPR:
1145 rhs1 = gimple_assign_rhs1 (stmt);
1146 rhs2 = gimple_assign_rhs2 (stmt);
1147 type = TREE_TYPE (rhs1);
1148 res = follow_ssa_edge_binary (loop, stmt, type, rhs1, code, rhs2,
218d1c24 1149 halting_phi, evolution_of_loop, limit);
5aefc6a0 1150 break;
218d1c24 1151
726a989a 1152 default:
5aefc6a0
EB
1153 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1154 res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1155 halting_phi, evolution_of_loop, limit);
1156 else
1157 res = t_false;
1158 break;
726a989a 1159 }
5aefc6a0
EB
1160
1161 return res;
726a989a
RB
1162}
1163
9baba81b
SP
1164/* Checks whether the I-th argument of a PHI comes from a backedge. */
1165
1166static bool
538dd0b7 1167backedge_phi_arg_p (gphi *phi, int i)
9baba81b 1168{
726a989a 1169 const_edge e = gimple_phi_arg_edge (phi, i);
9baba81b
SP
1170
1171 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1172 about updating it anywhere, and this should work as well most of the
1173 time. */
1174 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1175 return true;
1176
1177 return false;
1178}
1179
1180/* Helper function for one branch of the condition-phi-node. Return
1181 true if the strongly connected component has been found following
1182 this path. */
1183
c59dabbe 1184static inline t_bool
9baba81b 1185follow_ssa_edge_in_condition_phi_branch (int i,
b8698a0f 1186 struct loop *loop,
538dd0b7
DM
1187 gphi *condition_phi,
1188 gphi *halting_phi,
9baba81b 1189 tree *evolution_of_branch,
c59dabbe 1190 tree init_cond, int limit)
9baba81b
SP
1191{
1192 tree branch = PHI_ARG_DEF (condition_phi, i);
1193 *evolution_of_branch = chrec_dont_know;
1194
1195 /* Do not follow back edges (they must belong to an irreducible loop, which
1196 we really do not want to worry about). */
1197 if (backedge_phi_arg_p (condition_phi, i))
c59dabbe 1198 return t_false;
9baba81b
SP
1199
1200 if (TREE_CODE (branch) == SSA_NAME)
1201 {
1202 *evolution_of_branch = init_cond;
b8698a0f 1203 return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi,
c59dabbe 1204 evolution_of_branch, limit);
9baba81b
SP
1205 }
1206
b8698a0f 1207 /* This case occurs when one of the condition branches sets
89dbed81 1208 the variable to a constant: i.e. a phi-node like
b8698a0f
L
1209 "a_2 = PHI <a_7(5), 2(6)>;".
1210
1211 FIXME: This case have to be refined correctly:
9baba81b
SP
1212 in some cases it is possible to say something better than
1213 chrec_dont_know, for example using a wrap-around notation. */
c59dabbe 1214 return t_false;
9baba81b
SP
1215}
1216
1217/* This function merges the branches of a condition-phi-node in a
1218 loop. */
1219
c59dabbe 1220static t_bool
9baba81b 1221follow_ssa_edge_in_condition_phi (struct loop *loop,
538dd0b7
DM
1222 gphi *condition_phi,
1223 gphi *halting_phi,
c59dabbe 1224 tree *evolution_of_loop, int limit)
9baba81b 1225{
726a989a 1226 int i, n;
9baba81b
SP
1227 tree init = *evolution_of_loop;
1228 tree evolution_of_branch;
c59dabbe
SP
1229 t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi,
1230 halting_phi,
1231 &evolution_of_branch,
1232 init, limit);
1233 if (res == t_false || res == t_dont_know)
1234 return res;
9baba81b 1235
9baba81b
SP
1236 *evolution_of_loop = evolution_of_branch;
1237
726a989a 1238 n = gimple_phi_num_args (condition_phi);
726a989a 1239 for (i = 1; i < n; i++)
9baba81b 1240 {
e0afb98a
SP
1241 /* Quickly give up when the evolution of one of the branches is
1242 not known. */
1243 if (*evolution_of_loop == chrec_dont_know)
c59dabbe 1244 return t_true;
e0afb98a 1245
788d3075
RG
1246 /* Increase the limit by the PHI argument number to avoid exponential
1247 time and memory complexity. */
c59dabbe
SP
1248 res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi,
1249 halting_phi,
1250 &evolution_of_branch,
788d3075 1251 init, limit + i);
c59dabbe
SP
1252 if (res == t_false || res == t_dont_know)
1253 return res;
9baba81b
SP
1254
1255 *evolution_of_loop = chrec_merge (*evolution_of_loop,
1256 evolution_of_branch);
1257 }
b8698a0f 1258
c59dabbe 1259 return t_true;
9baba81b
SP
1260}
1261
1262/* Follow an SSA edge in an inner loop. It computes the overall
1263 effect of the loop, and following the symbolic initial conditions,
1264 it follows the edges in the parent loop. The inner loop is
1265 considered as a single statement. */
1266
c59dabbe 1267static t_bool
9baba81b 1268follow_ssa_edge_inner_loop_phi (struct loop *outer_loop,
538dd0b7
DM
1269 gphi *loop_phi_node,
1270 gphi *halting_phi,
c59dabbe 1271 tree *evolution_of_loop, int limit)
9baba81b
SP
1272{
1273 struct loop *loop = loop_containing_stmt (loop_phi_node);
1274 tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node));
1275
1276 /* Sometimes, the inner loop is too difficult to analyze, and the
1277 result of the analysis is a symbolic parameter. */
1278 if (ev == PHI_RESULT (loop_phi_node))
1279 {
c59dabbe 1280 t_bool res = t_false;
726a989a 1281 int i, n = gimple_phi_num_args (loop_phi_node);
9baba81b 1282
726a989a 1283 for (i = 0; i < n; i++)
9baba81b
SP
1284 {
1285 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1286 basic_block bb;
1287
1288 /* Follow the edges that exit the inner loop. */
726a989a 1289 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
9baba81b 1290 if (!flow_bb_inside_loop_p (loop, bb))
726a989a
RB
1291 res = follow_ssa_edge_expr (outer_loop, loop_phi_node,
1292 arg, halting_phi,
1293 evolution_of_loop, limit);
c59dabbe
SP
1294 if (res == t_true)
1295 break;
9baba81b
SP
1296 }
1297
1298 /* If the path crosses this loop-phi, give up. */
c59dabbe 1299 if (res == t_true)
9baba81b
SP
1300 *evolution_of_loop = chrec_dont_know;
1301
1302 return res;
1303 }
1304
1305 /* Otherwise, compute the overall effect of the inner loop. */
1306 ev = compute_overall_effect_of_inner_loop (loop, ev);
726a989a
RB
1307 return follow_ssa_edge_expr (outer_loop, loop_phi_node, ev, halting_phi,
1308 evolution_of_loop, limit);
9baba81b
SP
1309}
1310
1311/* Follow an SSA edge from a loop-phi-node to itself, constructing a
1312 path that is analyzed on the return walk. */
1313
c59dabbe 1314static t_bool
355fe088 1315follow_ssa_edge (struct loop *loop, gimple *def, gphi *halting_phi,
c59dabbe 1316 tree *evolution_of_loop, int limit)
9baba81b
SP
1317{
1318 struct loop *def_loop;
b8698a0f 1319
726a989a 1320 if (gimple_nop_p (def))
c59dabbe 1321 return t_false;
b8698a0f 1322
c59dabbe 1323 /* Give up if the path is longer than the MAX that we allow. */
14dd9aab 1324 if (limit > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_COMPLEXITY))
c59dabbe 1325 return t_dont_know;
b8698a0f 1326
9baba81b 1327 def_loop = loop_containing_stmt (def);
b8698a0f 1328
726a989a 1329 switch (gimple_code (def))
9baba81b 1330 {
726a989a 1331 case GIMPLE_PHI:
9baba81b
SP
1332 if (!loop_phi_node_p (def))
1333 /* DEF is a condition-phi-node. Follow the branches, and
1334 record their evolutions. Finally, merge the collected
1335 information and set the approximation to the main
1336 variable. */
b8698a0f 1337 return follow_ssa_edge_in_condition_phi
538dd0b7
DM
1338 (loop, as_a <gphi *> (def), halting_phi, evolution_of_loop,
1339 limit);
9baba81b
SP
1340
1341 /* When the analyzed phi is the halting_phi, the
1342 depth-first search is over: we have found a path from
1343 the halting_phi to itself in the loop. */
1344 if (def == halting_phi)
c59dabbe 1345 return t_true;
b8698a0f 1346
9baba81b 1347 /* Otherwise, the evolution of the HALTING_PHI depends
89dbed81 1348 on the evolution of another loop-phi-node, i.e. the
9baba81b
SP
1349 evolution function is a higher degree polynomial. */
1350 if (def_loop == loop)
c59dabbe 1351 return t_false;
b8698a0f 1352
9baba81b
SP
1353 /* Inner loop. */
1354 if (flow_loop_nested_p (loop, def_loop))
b8698a0f 1355 return follow_ssa_edge_inner_loop_phi
538dd0b7
DM
1356 (loop, as_a <gphi *> (def), halting_phi, evolution_of_loop,
1357 limit + 1);
9baba81b
SP
1358
1359 /* Outer loop. */
c59dabbe 1360 return t_false;
9baba81b 1361
726a989a 1362 case GIMPLE_ASSIGN:
b8698a0f 1363 return follow_ssa_edge_in_rhs (loop, def, halting_phi,
c59dabbe 1364 evolution_of_loop, limit);
b8698a0f 1365
9baba81b
SP
1366 default:
1367 /* At this level of abstraction, the program is just a set
726a989a 1368 of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no
9baba81b 1369 other node to be handled. */
c59dabbe 1370 return t_false;
9baba81b
SP
1371 }
1372}
1373
1374\f
b83b5507
BC
1375/* Simplify PEELED_CHREC represented by (init_cond, arg) in LOOP.
1376 Handle below case and return the corresponding POLYNOMIAL_CHREC:
1377
1378 # i_17 = PHI <i_13(5), 0(3)>
1379 # _20 = PHI <_5(5), start_4(D)(3)>
1380 ...
1381 i_13 = i_17 + 1;
1382 _5 = start_4(D) + i_13;
1383
1384 Though variable _20 appears as a PEELED_CHREC in the form of
1385 (start_4, _5)_LOOP, it's a POLYNOMIAL_CHREC like {start_4, 1}_LOOP.
1386
1387 See PR41488. */
1388
1389static tree
1390simplify_peeled_chrec (struct loop *loop, tree arg, tree init_cond)
1391{
1392 aff_tree aff1, aff2;
1393 tree ev, left, right, type, step_val;
39c8aaa4 1394 hash_map<tree, name_expansion *> *peeled_chrec_map = NULL;
b83b5507
BC
1395
1396 ev = instantiate_parameters (loop, analyze_scalar_evolution (loop, arg));
1397 if (ev == NULL_TREE || TREE_CODE (ev) != POLYNOMIAL_CHREC)
1398 return chrec_dont_know;
1399
1400 left = CHREC_LEFT (ev);
1401 right = CHREC_RIGHT (ev);
1402 type = TREE_TYPE (left);
1403 step_val = chrec_fold_plus (type, init_cond, right);
1404
1405 /* Transform (init, {left, right}_LOOP)_LOOP to {init, right}_LOOP
1406 if "left" equals to "init + right". */
1407 if (operand_equal_p (left, step_val, 0))
1408 {
1409 if (dump_file && (dump_flags & TDF_SCEV))
1410 fprintf (dump_file, "Simplify PEELED_CHREC into POLYNOMIAL_CHREC.\n");
1411
1412 return build_polynomial_chrec (loop->num, init_cond, right);
1413 }
1414
1415 /* Try harder to check if they are equal. */
1416 tree_to_aff_combination_expand (left, type, &aff1, &peeled_chrec_map);
1417 tree_to_aff_combination_expand (step_val, type, &aff2, &peeled_chrec_map);
1418 free_affine_expand_cache (&peeled_chrec_map);
807e902e 1419 aff_combination_scale (&aff2, -1);
b83b5507
BC
1420 aff_combination_add (&aff1, &aff2);
1421
1422 /* Transform (init, {left, right}_LOOP)_LOOP to {init, right}_LOOP
1423 if "left" equals to "init + right". */
1424 if (aff_combination_zero_p (&aff1))
1425 {
1426 if (dump_file && (dump_flags & TDF_SCEV))
1427 fprintf (dump_file, "Simplify PEELED_CHREC into POLYNOMIAL_CHREC.\n");
1428
1429 return build_polynomial_chrec (loop->num, init_cond, right);
1430 }
1431 return chrec_dont_know;
1432}
9baba81b
SP
1433
1434/* Given a LOOP_PHI_NODE, this function determines the evolution
1435 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1436
1437static tree
538dd0b7 1438analyze_evolution_in_loop (gphi *loop_phi_node,
9baba81b
SP
1439 tree init_cond)
1440{
726a989a 1441 int i, n = gimple_phi_num_args (loop_phi_node);
9baba81b
SP
1442 tree evolution_function = chrec_not_analyzed_yet;
1443 struct loop *loop = loop_containing_stmt (loop_phi_node);
1444 basic_block bb;
b83b5507 1445 static bool simplify_peeled_chrec_p = true;
b8698a0f 1446
dfedbe40 1447 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
1448 {
1449 fprintf (dump_file, "(analyze_evolution_in_loop \n");
1450 fprintf (dump_file, " (loop_phi_node = ");
726a989a 1451 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
9baba81b
SP
1452 fprintf (dump_file, ")\n");
1453 }
b8698a0f 1454
726a989a 1455 for (i = 0; i < n; i++)
9baba81b
SP
1456 {
1457 tree arg = PHI_ARG_DEF (loop_phi_node, i);
355fe088 1458 gimple *ssa_chain;
726a989a 1459 tree ev_fn;
874caa00 1460 t_bool res;
9baba81b
SP
1461
1462 /* Select the edges that enter the loop body. */
726a989a 1463 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
9baba81b
SP
1464 if (!flow_bb_inside_loop_p (loop, bb))
1465 continue;
f29deac9 1466
9baba81b
SP
1467 if (TREE_CODE (arg) == SSA_NAME)
1468 {
f29deac9
SP
1469 bool val = false;
1470
9baba81b
SP
1471 ssa_chain = SSA_NAME_DEF_STMT (arg);
1472
1473 /* Pass in the initial condition to the follow edge function. */
1474 ev_fn = init_cond;
c59dabbe 1475 res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0);
f29deac9
SP
1476
1477 /* If ev_fn has no evolution in the inner loop, and the
1478 init_cond is not equal to ev_fn, then we have an
1479 ambiguity between two possible values, as we cannot know
1480 the number of iterations at this point. */
1481 if (TREE_CODE (ev_fn) != POLYNOMIAL_CHREC
1482 && no_evolution_in_loop_p (ev_fn, loop->num, &val) && val
1483 && !operand_equal_p (init_cond, ev_fn, 0))
1484 ev_fn = chrec_dont_know;
9baba81b
SP
1485 }
1486 else
874caa00 1487 res = t_false;
f29deac9 1488
9baba81b
SP
1489 /* When it is impossible to go back on the same
1490 loop_phi_node by following the ssa edges, the
89dbed81 1491 evolution is represented by a peeled chrec, i.e. the
9baba81b 1492 first iteration, EV_FN has the value INIT_COND, then
b8698a0f 1493 all the other iterations it has the value of ARG.
9baba81b 1494 For the moment, PEELED_CHREC nodes are not built. */
874caa00 1495 if (res != t_true)
b83b5507
BC
1496 {
1497 ev_fn = chrec_dont_know;
1498 /* Try to recognize POLYNOMIAL_CHREC which appears in
1499 the form of PEELED_CHREC, but guard the process with
1500 a bool variable to keep the analyzer from infinite
1501 recurrence for real PEELED_RECs. */
1502 if (simplify_peeled_chrec_p && TREE_CODE (arg) == SSA_NAME)
1503 {
1504 simplify_peeled_chrec_p = false;
1505 ev_fn = simplify_peeled_chrec (loop, arg, init_cond);
1506 simplify_peeled_chrec_p = true;
1507 }
1508 }
b8698a0f 1509
9baba81b 1510 /* When there are multiple back edges of the loop (which in fact never
8c27b7d4 1511 happens currently, but nevertheless), merge their evolutions. */
9baba81b 1512 evolution_function = chrec_merge (evolution_function, ev_fn);
e21401b6
RB
1513
1514 if (evolution_function == chrec_dont_know)
1515 break;
9baba81b 1516 }
b8698a0f 1517
dfedbe40 1518 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
1519 {
1520 fprintf (dump_file, " (evolution_function = ");
1521 print_generic_expr (dump_file, evolution_function, 0);
1522 fprintf (dump_file, "))\n");
1523 }
b8698a0f 1524
9baba81b
SP
1525 return evolution_function;
1526}
1527
806f2c1b
AL
1528/* Looks to see if VAR is a copy of a constant (via straightforward assignments
1529 or degenerate phi's). If so, returns the constant; else, returns VAR. */
1530
1531static tree
1532follow_copies_to_constant (tree var)
1533{
1534 tree res = var;
1535 while (TREE_CODE (res) == SSA_NAME)
1536 {
1537 gimple *def = SSA_NAME_DEF_STMT (res);
1538 if (gphi *phi = dyn_cast <gphi *> (def))
1539 {
1540 if (tree rhs = degenerate_phi_result (phi))
1541 res = rhs;
1542 else
1543 break;
1544 }
1545 else if (gimple_assign_single_p (def))
1546 /* Will exit loop if not an SSA_NAME. */
1547 res = gimple_assign_rhs1 (def);
1548 else
1549 break;
1550 }
1551 if (CONSTANT_CLASS_P (res))
1552 return res;
1553 return var;
1554}
1555
9baba81b
SP
1556/* Given a loop-phi-node, return the initial conditions of the
1557 variable on entry of the loop. When the CCP has propagated
1558 constants into the loop-phi-node, the initial condition is
1559 instantiated, otherwise the initial condition is kept symbolic.
1560 This analyzer does not analyze the evolution outside the current
1561 loop, and leaves this task to the on-demand tree reconstructor. */
1562
b8698a0f 1563static tree
538dd0b7 1564analyze_initial_condition (gphi *loop_phi_node)
9baba81b 1565{
726a989a 1566 int i, n;
9baba81b 1567 tree init_cond = chrec_not_analyzed_yet;
726a989a 1568 struct loop *loop = loop_containing_stmt (loop_phi_node);
b8698a0f 1569
dfedbe40 1570 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
1571 {
1572 fprintf (dump_file, "(analyze_initial_condition \n");
1573 fprintf (dump_file, " (loop_phi_node = \n");
726a989a 1574 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
9baba81b
SP
1575 fprintf (dump_file, ")\n");
1576 }
b8698a0f 1577
726a989a
RB
1578 n = gimple_phi_num_args (loop_phi_node);
1579 for (i = 0; i < n; i++)
9baba81b
SP
1580 {
1581 tree branch = PHI_ARG_DEF (loop_phi_node, i);
726a989a 1582 basic_block bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
b8698a0f 1583
9baba81b
SP
1584 /* When the branch is oriented to the loop's body, it does
1585 not contribute to the initial condition. */
1586 if (flow_bb_inside_loop_p (loop, bb))
1587 continue;
1588
1589 if (init_cond == chrec_not_analyzed_yet)
1590 {
1591 init_cond = branch;
1592 continue;
1593 }
1594
1595 if (TREE_CODE (branch) == SSA_NAME)
1596 {
1597 init_cond = chrec_dont_know;
1598 break;
1599 }
1600
1601 init_cond = chrec_merge (init_cond, branch);
1602 }
1603
1604 /* Ooops -- a loop without an entry??? */
1605 if (init_cond == chrec_not_analyzed_yet)
1606 init_cond = chrec_dont_know;
1607
806f2c1b
AL
1608 /* We may not have fully constant propagated IL. Handle degenerate PHIs here
1609 to not miss important early loop unrollings. */
1610 init_cond = follow_copies_to_constant (init_cond);
bf1cbdc6 1611
dfedbe40 1612 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
1613 {
1614 fprintf (dump_file, " (init_cond = ");
1615 print_generic_expr (dump_file, init_cond, 0);
1616 fprintf (dump_file, "))\n");
1617 }
b8698a0f 1618
9baba81b
SP
1619 return init_cond;
1620}
1621
1622/* Analyze the scalar evolution for LOOP_PHI_NODE. */
1623
b8698a0f 1624static tree
538dd0b7 1625interpret_loop_phi (struct loop *loop, gphi *loop_phi_node)
9baba81b
SP
1626{
1627 tree res;
1628 struct loop *phi_loop = loop_containing_stmt (loop_phi_node);
1629 tree init_cond;
b8698a0f 1630
9baba81b
SP
1631 if (phi_loop != loop)
1632 {
1633 struct loop *subloop;
1634 tree evolution_fn = analyze_scalar_evolution
1635 (phi_loop, PHI_RESULT (loop_phi_node));
1636
1637 /* Dive one level deeper. */
9ba025a2 1638 subloop = superloop_at_depth (phi_loop, loop_depth (loop) + 1);
9baba81b
SP
1639
1640 /* Interpret the subloop. */
1641 res = compute_overall_effect_of_inner_loop (subloop, evolution_fn);
1642 return res;
1643 }
1644
1645 /* Otherwise really interpret the loop phi. */
1646 init_cond = analyze_initial_condition (loop_phi_node);
1647 res = analyze_evolution_in_loop (loop_phi_node, init_cond);
1648
73c865fa
RG
1649 /* Verify we maintained the correct initial condition throughout
1650 possible conversions in the SSA chain. */
1651 if (res != chrec_dont_know)
1652 {
1653 tree new_init = res;
1654 if (CONVERT_EXPR_P (res)
1655 && TREE_CODE (TREE_OPERAND (res, 0)) == POLYNOMIAL_CHREC)
1656 new_init = fold_convert (TREE_TYPE (res),
1657 CHREC_LEFT (TREE_OPERAND (res, 0)));
1658 else if (TREE_CODE (res) == POLYNOMIAL_CHREC)
1659 new_init = CHREC_LEFT (res);
1660 STRIP_USELESS_TYPE_CONVERSION (new_init);
eb723fa3
RG
1661 if (TREE_CODE (new_init) == POLYNOMIAL_CHREC
1662 || !operand_equal_p (init_cond, new_init, 0))
73c865fa
RG
1663 return chrec_dont_know;
1664 }
1665
9baba81b
SP
1666 return res;
1667}
1668
1669/* This function merges the branches of a condition-phi-node,
1670 contained in the outermost loop, and whose arguments are already
1671 analyzed. */
1672
1673static tree
538dd0b7 1674interpret_condition_phi (struct loop *loop, gphi *condition_phi)
9baba81b 1675{
726a989a 1676 int i, n = gimple_phi_num_args (condition_phi);
9baba81b 1677 tree res = chrec_not_analyzed_yet;
b8698a0f 1678
726a989a 1679 for (i = 0; i < n; i++)
9baba81b
SP
1680 {
1681 tree branch_chrec;
b8698a0f 1682
9baba81b
SP
1683 if (backedge_phi_arg_p (condition_phi, i))
1684 {
1685 res = chrec_dont_know;
1686 break;
1687 }
1688
1689 branch_chrec = analyze_scalar_evolution
1690 (loop, PHI_ARG_DEF (condition_phi, i));
b8698a0f 1691
9baba81b 1692 res = chrec_merge (res, branch_chrec);
e21401b6
RB
1693 if (res == chrec_dont_know)
1694 break;
9baba81b
SP
1695 }
1696
1697 return res;
1698}
1699
726a989a 1700/* Interpret the operation RHS1 OP RHS2. If we didn't
29836d07 1701 analyze this node before, follow the definitions until ending
726a989a 1702 either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the
9baba81b
SP
1703 return path, this function propagates evolutions (ala constant copy
1704 propagation). OPND1 is not a GIMPLE expression because we could
1705 analyze the effect of an inner loop: see interpret_loop_phi. */
1706
1707static tree
355fe088 1708interpret_rhs_expr (struct loop *loop, gimple *at_stmt,
726a989a 1709 tree type, tree rhs1, enum tree_code code, tree rhs2)
9baba81b 1710{
f802a424 1711 tree res, chrec1, chrec2, ctype;
355fe088 1712 gimple *def;
726a989a
RB
1713
1714 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1715 {
1716 if (is_gimple_min_invariant (rhs1))
1717 return chrec_convert (type, rhs1, at_stmt);
1718
1719 if (code == SSA_NAME)
1720 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1721 at_stmt);
1e8552eb 1722
726a989a
RB
1723 if (code == ASSERT_EXPR)
1724 {
1725 rhs1 = ASSERT_EXPR_VAR (rhs1);
1726 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1727 at_stmt);
1728 }
726a989a 1729 }
1e8552eb 1730
726a989a 1731 switch (code)
9baba81b 1732 {
6a02a719 1733 case ADDR_EXPR:
bef28ced
JL
1734 if (TREE_CODE (TREE_OPERAND (rhs1, 0)) == MEM_REF
1735 || handled_component_p (TREE_OPERAND (rhs1, 0)))
1736 {
ef4bddc2 1737 machine_mode mode;
bef28ced 1738 HOST_WIDE_INT bitsize, bitpos;
ee45a32d 1739 int unsignedp, reversep;
bef28ced
JL
1740 int volatilep = 0;
1741 tree base, offset;
1742 tree chrec3;
1743 tree unitpos;
1744
1745 base = get_inner_reference (TREE_OPERAND (rhs1, 0),
ee45a32d
EB
1746 &bitsize, &bitpos, &offset, &mode,
1747 &unsignedp, &reversep, &volatilep,
1748 false);
bef28ced
JL
1749
1750 if (TREE_CODE (base) == MEM_REF)
1751 {
1752 rhs2 = TREE_OPERAND (base, 1);
1753 rhs1 = TREE_OPERAND (base, 0);
1754
1755 chrec1 = analyze_scalar_evolution (loop, rhs1);
1756 chrec2 = analyze_scalar_evolution (loop, rhs2);
1757 chrec1 = chrec_convert (type, chrec1, at_stmt);
1758 chrec2 = chrec_convert (TREE_TYPE (rhs2), chrec2, at_stmt);
0547c9b6
RB
1759 chrec1 = instantiate_parameters (loop, chrec1);
1760 chrec2 = instantiate_parameters (loop, chrec2);
bef28ced
JL
1761 res = chrec_fold_plus (type, chrec1, chrec2);
1762 }
1763 else
1764 {
1765 chrec1 = analyze_scalar_evolution_for_address_of (loop, base);
1766 chrec1 = chrec_convert (type, chrec1, at_stmt);
1767 res = chrec1;
1768 }
6a02a719 1769
bef28ced
JL
1770 if (offset != NULL_TREE)
1771 {
1772 chrec2 = analyze_scalar_evolution (loop, offset);
1773 chrec2 = chrec_convert (TREE_TYPE (offset), chrec2, at_stmt);
0547c9b6 1774 chrec2 = instantiate_parameters (loop, chrec2);
bef28ced
JL
1775 res = chrec_fold_plus (type, res, chrec2);
1776 }
1777
1778 if (bitpos != 0)
1779 {
1780 gcc_assert ((bitpos % BITS_PER_UNIT) == 0);
1781
18dae016 1782 unitpos = size_int (bitpos / BITS_PER_UNIT);
bef28ced
JL
1783 chrec3 = analyze_scalar_evolution (loop, unitpos);
1784 chrec3 = chrec_convert (TREE_TYPE (unitpos), chrec3, at_stmt);
0547c9b6 1785 chrec3 = instantiate_parameters (loop, chrec3);
bef28ced
JL
1786 res = chrec_fold_plus (type, res, chrec3);
1787 }
1788 }
1789 else
1790 res = chrec_dont_know;
1791 break;
6a02a719 1792
5be014d5 1793 case POINTER_PLUS_EXPR:
726a989a
RB
1794 chrec1 = analyze_scalar_evolution (loop, rhs1);
1795 chrec2 = analyze_scalar_evolution (loop, rhs2);
1796 chrec1 = chrec_convert (type, chrec1, at_stmt);
0d82a1c8 1797 chrec2 = chrec_convert (TREE_TYPE (rhs2), chrec2, at_stmt);
0547c9b6
RB
1798 chrec1 = instantiate_parameters (loop, chrec1);
1799 chrec2 = instantiate_parameters (loop, chrec2);
726a989a 1800 res = chrec_fold_plus (type, chrec1, chrec2);
5be014d5
AP
1801 break;
1802
9baba81b 1803 case PLUS_EXPR:
726a989a
RB
1804 chrec1 = analyze_scalar_evolution (loop, rhs1);
1805 chrec2 = analyze_scalar_evolution (loop, rhs2);
f802a424
RB
1806 ctype = type;
1807 /* When the stmt is conditionally executed re-write the CHREC
1808 into a form that has well-defined behavior on overflow. */
1809 if (at_stmt
1810 && INTEGRAL_TYPE_P (type)
1811 && ! TYPE_OVERFLOW_WRAPS (type)
1812 && ! dominated_by_p (CDI_DOMINATORS, loop->latch,
1813 gimple_bb (at_stmt)))
1814 ctype = unsigned_type_for (type);
1815 chrec1 = chrec_convert (ctype, chrec1, at_stmt);
1816 chrec2 = chrec_convert (ctype, chrec2, at_stmt);
0547c9b6
RB
1817 chrec1 = instantiate_parameters (loop, chrec1);
1818 chrec2 = instantiate_parameters (loop, chrec2);
f802a424
RB
1819 res = chrec_fold_plus (ctype, chrec1, chrec2);
1820 if (type != ctype)
1821 res = chrec_convert (type, res, at_stmt);
9baba81b 1822 break;
b8698a0f 1823
9baba81b 1824 case MINUS_EXPR:
726a989a
RB
1825 chrec1 = analyze_scalar_evolution (loop, rhs1);
1826 chrec2 = analyze_scalar_evolution (loop, rhs2);
f802a424
RB
1827 ctype = type;
1828 /* When the stmt is conditionally executed re-write the CHREC
1829 into a form that has well-defined behavior on overflow. */
1830 if (at_stmt
1831 && INTEGRAL_TYPE_P (type)
1832 && ! TYPE_OVERFLOW_WRAPS (type)
1833 && ! dominated_by_p (CDI_DOMINATORS,
1834 loop->latch, gimple_bb (at_stmt)))
1835 ctype = unsigned_type_for (type);
1836 chrec1 = chrec_convert (ctype, chrec1, at_stmt);
1837 chrec2 = chrec_convert (ctype, chrec2, at_stmt);
0547c9b6
RB
1838 chrec1 = instantiate_parameters (loop, chrec1);
1839 chrec2 = instantiate_parameters (loop, chrec2);
f802a424
RB
1840 res = chrec_fold_minus (ctype, chrec1, chrec2);
1841 if (type != ctype)
1842 res = chrec_convert (type, res, at_stmt);
9baba81b
SP
1843 break;
1844
1845 case NEGATE_EXPR:
726a989a 1846 chrec1 = analyze_scalar_evolution (loop, rhs1);
f802a424
RB
1847 ctype = type;
1848 /* When the stmt is conditionally executed re-write the CHREC
1849 into a form that has well-defined behavior on overflow. */
1850 if (at_stmt
1851 && INTEGRAL_TYPE_P (type)
1852 && ! TYPE_OVERFLOW_WRAPS (type)
1853 && ! dominated_by_p (CDI_DOMINATORS,
1854 loop->latch, gimple_bb (at_stmt)))
1855 ctype = unsigned_type_for (type);
1856 chrec1 = chrec_convert (ctype, chrec1, at_stmt);
9a75ede0 1857 /* TYPE may be integer, real or complex, so use fold_convert. */
0547c9b6 1858 chrec1 = instantiate_parameters (loop, chrec1);
f802a424
RB
1859 res = chrec_fold_multiply (ctype, chrec1,
1860 fold_convert (ctype, integer_minus_one_node));
1861 if (type != ctype)
1862 res = chrec_convert (type, res, at_stmt);
9baba81b
SP
1863 break;
1864
418df9d7
JJ
1865 case BIT_NOT_EXPR:
1866 /* Handle ~X as -1 - X. */
1867 chrec1 = analyze_scalar_evolution (loop, rhs1);
1868 chrec1 = chrec_convert (type, chrec1, at_stmt);
0547c9b6 1869 chrec1 = instantiate_parameters (loop, chrec1);
418df9d7
JJ
1870 res = chrec_fold_minus (type,
1871 fold_convert (type, integer_minus_one_node),
1872 chrec1);
1873 break;
1874
9baba81b 1875 case MULT_EXPR:
726a989a
RB
1876 chrec1 = analyze_scalar_evolution (loop, rhs1);
1877 chrec2 = analyze_scalar_evolution (loop, rhs2);
f802a424
RB
1878 ctype = type;
1879 /* When the stmt is conditionally executed re-write the CHREC
1880 into a form that has well-defined behavior on overflow. */
1881 if (at_stmt
1882 && INTEGRAL_TYPE_P (type)
1883 && ! TYPE_OVERFLOW_WRAPS (type)
1884 && ! dominated_by_p (CDI_DOMINATORS,
1885 loop->latch, gimple_bb (at_stmt)))
1886 ctype = unsigned_type_for (type);
1887 chrec1 = chrec_convert (ctype, chrec1, at_stmt);
1888 chrec2 = chrec_convert (ctype, chrec2, at_stmt);
0547c9b6
RB
1889 chrec1 = instantiate_parameters (loop, chrec1);
1890 chrec2 = instantiate_parameters (loop, chrec2);
f802a424
RB
1891 res = chrec_fold_multiply (ctype, chrec1, chrec2);
1892 if (type != ctype)
1893 res = chrec_convert (type, res, at_stmt);
0bca51f0 1894 break;
b8698a0f 1895
60f2d2f3
AL
1896 case LSHIFT_EXPR:
1897 {
1898 /* Handle A<<B as A * (1<<B). */
1899 tree uns = unsigned_type_for (type);
1900 chrec1 = analyze_scalar_evolution (loop, rhs1);
1901 chrec2 = analyze_scalar_evolution (loop, rhs2);
1902 chrec1 = chrec_convert (uns, chrec1, at_stmt);
1903 chrec1 = instantiate_parameters (loop, chrec1);
1904 chrec2 = instantiate_parameters (loop, chrec2);
1905
1906 tree one = build_int_cst (uns, 1);
1907 chrec2 = fold_build2 (LSHIFT_EXPR, uns, one, chrec2);
1908 res = chrec_fold_multiply (uns, chrec1, chrec2);
1909 res = chrec_convert (type, res, at_stmt);
1910 }
1911 break;
1912
1043771b 1913 CASE_CONVERT:
195b4c50
RG
1914 /* In case we have a truncation of a widened operation that in
1915 the truncated type has undefined overflow behavior analyze
1916 the operation done in an unsigned type of the same precision
1917 as the final truncation. We cannot derive a scalar evolution
1918 for the widened operation but for the truncated result. */
1919 if (TREE_CODE (type) == INTEGER_TYPE
1920 && TREE_CODE (TREE_TYPE (rhs1)) == INTEGER_TYPE
1921 && TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (rhs1))
1922 && TYPE_OVERFLOW_UNDEFINED (type)
1923 && TREE_CODE (rhs1) == SSA_NAME
1924 && (def = SSA_NAME_DEF_STMT (rhs1))
1925 && is_gimple_assign (def)
1926 && TREE_CODE_CLASS (gimple_assign_rhs_code (def)) == tcc_binary
1927 && TREE_CODE (gimple_assign_rhs2 (def)) == INTEGER_CST)
1928 {
1929 tree utype = unsigned_type_for (type);
1930 chrec1 = interpret_rhs_expr (loop, at_stmt, utype,
1931 gimple_assign_rhs1 (def),
1932 gimple_assign_rhs_code (def),
1933 gimple_assign_rhs2 (def));
1934 }
1935 else
1936 chrec1 = analyze_scalar_evolution (loop, rhs1);
726a989a 1937 res = chrec_convert (type, chrec1, at_stmt);
9baba81b 1938 break;
b8698a0f 1939
9baba81b
SP
1940 default:
1941 res = chrec_dont_know;
1942 break;
1943 }
b8698a0f 1944
9baba81b
SP
1945 return res;
1946}
1947
726a989a
RB
1948/* Interpret the expression EXPR. */
1949
1950static tree
355fe088 1951interpret_expr (struct loop *loop, gimple *at_stmt, tree expr)
726a989a
RB
1952{
1953 enum tree_code code;
1954 tree type = TREE_TYPE (expr), op0, op1;
1955
1956 if (automatically_generated_chrec_p (expr))
1957 return expr;
1958
4e71066d
RG
1959 if (TREE_CODE (expr) == POLYNOMIAL_CHREC
1960 || get_gimple_rhs_class (TREE_CODE (expr)) == GIMPLE_TERNARY_RHS)
726a989a
RB
1961 return chrec_dont_know;
1962
1963 extract_ops_from_tree (expr, &code, &op0, &op1);
1964
1965 return interpret_rhs_expr (loop, at_stmt, type,
1966 op0, code, op1);
1967}
1968
1969/* Interpret the rhs of the assignment STMT. */
1970
1971static tree
355fe088 1972interpret_gimple_assign (struct loop *loop, gimple *stmt)
726a989a
RB
1973{
1974 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
1975 enum tree_code code = gimple_assign_rhs_code (stmt);
1976
1977 return interpret_rhs_expr (loop, stmt, type,
1978 gimple_assign_rhs1 (stmt), code,
1979 gimple_assign_rhs2 (stmt));
1980}
1981
9baba81b
SP
1982\f
1983
b8698a0f 1984/* This section contains all the entry points:
9baba81b
SP
1985 - number_of_iterations_in_loop,
1986 - analyze_scalar_evolution,
1987 - instantiate_parameters.
1988*/
1989
1990/* Compute and return the evolution function in WRTO_LOOP, the nearest
1991 common ancestor of DEF_LOOP and USE_LOOP. */
1992
b8698a0f
L
1993static tree
1994compute_scalar_evolution_in_loop (struct loop *wrto_loop,
1995 struct loop *def_loop,
9baba81b
SP
1996 tree ev)
1997{
492e5456 1998 bool val;
9baba81b 1999 tree res;
492e5456 2000
9baba81b
SP
2001 if (def_loop == wrto_loop)
2002 return ev;
2003
9ba025a2 2004 def_loop = superloop_at_depth (def_loop, loop_depth (wrto_loop) + 1);
9baba81b
SP
2005 res = compute_overall_effect_of_inner_loop (def_loop, ev);
2006
492e5456
SP
2007 if (no_evolution_in_loop_p (res, wrto_loop->num, &val) && val)
2008 return res;
2009
9baba81b
SP
2010 return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet);
2011}
2012
2013/* Helper recursive function. */
2014
2015static tree
2016analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res)
2017{
726a989a 2018 tree type = TREE_TYPE (var);
355fe088 2019 gimple *def;
9baba81b
SP
2020 basic_block bb;
2021 struct loop *def_loop;
2022
42d375ed 2023 if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE)
9baba81b
SP
2024 return chrec_dont_know;
2025
2026 if (TREE_CODE (var) != SSA_NAME)
726a989a 2027 return interpret_expr (loop, NULL, var);
9baba81b
SP
2028
2029 def = SSA_NAME_DEF_STMT (var);
726a989a 2030 bb = gimple_bb (def);
9baba81b
SP
2031 def_loop = bb ? bb->loop_father : NULL;
2032
2033 if (bb == NULL
2034 || !flow_bb_inside_loop_p (loop, bb))
2035 {
806f2c1b
AL
2036 /* Keep symbolic form, but look through obvious copies for constants. */
2037 res = follow_copies_to_constant (var);
9baba81b
SP
2038 goto set_and_end;
2039 }
2040
2041 if (res != chrec_not_analyzed_yet)
2042 {
2043 if (loop != bb->loop_father)
b8698a0f 2044 res = compute_scalar_evolution_in_loop
9baba81b
SP
2045 (find_common_loop (loop, bb->loop_father), bb->loop_father, res);
2046
2047 goto set_and_end;
2048 }
2049
2050 if (loop != def_loop)
2051 {
2052 res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet);
2053 res = compute_scalar_evolution_in_loop (loop, def_loop, res);
2054
2055 goto set_and_end;
2056 }
2057
726a989a 2058 switch (gimple_code (def))
9baba81b 2059 {
726a989a
RB
2060 case GIMPLE_ASSIGN:
2061 res = interpret_gimple_assign (loop, def);
9baba81b
SP
2062 break;
2063
726a989a 2064 case GIMPLE_PHI:
9baba81b 2065 if (loop_phi_node_p (def))
538dd0b7 2066 res = interpret_loop_phi (loop, as_a <gphi *> (def));
9baba81b 2067 else
538dd0b7 2068 res = interpret_condition_phi (loop, as_a <gphi *> (def));
9baba81b
SP
2069 break;
2070
2071 default:
2072 res = chrec_dont_know;
2073 break;
2074 }
2075
2076 set_and_end:
2077
2078 /* Keep the symbolic form. */
2079 if (res == chrec_dont_know)
2080 res = var;
2081
2082 if (loop == def_loop)
a213b219 2083 set_scalar_evolution (block_before_loop (loop), var, res);
9baba81b
SP
2084
2085 return res;
2086}
2087
52bdd655
SP
2088/* Analyzes and returns the scalar evolution of the ssa_name VAR in
2089 LOOP. LOOP is the loop in which the variable is used.
b8698a0f 2090
9baba81b
SP
2091 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
2092 pointer to the statement that uses this variable, in order to
2093 determine the evolution function of the variable, use the following
2094 calls:
b8698a0f 2095
52bdd655
SP
2096 loop_p loop = loop_containing_stmt (stmt);
2097 tree chrec_with_symbols = analyze_scalar_evolution (loop, var);
3f227a8c 2098 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
9baba81b
SP
2099*/
2100
b8698a0f 2101tree
9baba81b
SP
2102analyze_scalar_evolution (struct loop *loop, tree var)
2103{
2104 tree res;
2105
dfedbe40 2106 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
2107 {
2108 fprintf (dump_file, "(analyze_scalar_evolution \n");
2109 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
2110 fprintf (dump_file, " (scalar = ");
2111 print_generic_expr (dump_file, var, 0);
2112 fprintf (dump_file, ")\n");
2113 }
2114
a213b219
SP
2115 res = get_scalar_evolution (block_before_loop (loop), var);
2116 res = analyze_scalar_evolution_1 (loop, var, res);
9baba81b 2117
dfedbe40 2118 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
2119 fprintf (dump_file, ")\n");
2120
2121 return res;
2122}
2123
bef28ced
JL
2124/* Analyzes and returns the scalar evolution of VAR address in LOOP. */
2125
2126static tree
2127analyze_scalar_evolution_for_address_of (struct loop *loop, tree var)
2128{
2129 return analyze_scalar_evolution (loop, build_fold_addr_expr (var));
2130}
2131
9baba81b 2132/* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
f017bf5e 2133 WRTO_LOOP (which should be a superloop of USE_LOOP)
a6f778b2
ZD
2134
2135 FOLDED_CASTS is set to true if resolve_mixers used
2136 chrec_convert_aggressive (TODO -- not really, we are way too conservative
b8698a0f
L
2137 at the moment in order to keep things simple).
2138
f017bf5e
ZD
2139 To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following
2140 example:
2141
2142 for (i = 0; i < 100; i++) -- loop 1
2143 {
2144 for (j = 0; j < 100; j++) -- loop 2
2145 {
2146 k1 = i;
2147 k2 = j;
2148
2149 use2 (k1, k2);
2150
2151 for (t = 0; t < 100; t++) -- loop 3
2152 use3 (k1, k2);
2153
2154 }
2155 use1 (k1, k2);
2156 }
2157
2158 Both k1 and k2 are invariants in loop3, thus
2159 analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1
2160 analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2
2161
2162 As they are invariant, it does not matter whether we consider their
2163 usage in loop 3 or loop 2, hence
2164 analyze_scalar_evolution_in_loop (loop2, loop3, k1) =
2165 analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i
2166 analyze_scalar_evolution_in_loop (loop2, loop3, k2) =
2167 analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2
2168
2169 Similarly for their evolutions with respect to loop 1. The values of K2
2170 in the use in loop 2 vary independently on loop 1, thus we cannot express
2171 the evolution with respect to loop 1:
2172 analyze_scalar_evolution_in_loop (loop1, loop3, k1) =
2173 analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1
2174 analyze_scalar_evolution_in_loop (loop1, loop3, k2) =
2175 analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know
2176
2177 The value of k2 in the use in loop 1 is known, though:
2178 analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1
2179 analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100
2180 */
9baba81b
SP
2181
2182static tree
2183analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop,
a6f778b2 2184 tree version, bool *folded_casts)
9baba81b
SP
2185{
2186 bool val = false;
a6f778b2 2187 tree ev = version, tmp;
9baba81b 2188
b8698a0f 2189 /* We cannot just do
f017bf5e
ZD
2190
2191 tmp = analyze_scalar_evolution (use_loop, version);
c70ed622 2192 ev = resolve_mixers (wrto_loop, tmp, folded_casts);
f017bf5e
ZD
2193
2194 as resolve_mixers would query the scalar evolution with respect to
2195 wrto_loop. For example, in the situation described in the function
2196 comment, suppose that wrto_loop = loop1, use_loop = loop3 and
2197 version = k2. Then
2198
2199 analyze_scalar_evolution (use_loop, version) = k2
2200
c70ed622
BC
2201 and resolve_mixers (loop1, k2, folded_casts) finds that the value of
2202 k2 in loop 1 is 100, which is a wrong result, since we are interested
2203 in the value in loop 3.
f017bf5e
ZD
2204
2205 Instead, we need to proceed from use_loop to wrto_loop loop by loop,
2206 each time checking that there is no evolution in the inner loop. */
2207
a6f778b2
ZD
2208 if (folded_casts)
2209 *folded_casts = false;
9baba81b
SP
2210 while (1)
2211 {
a6f778b2 2212 tmp = analyze_scalar_evolution (use_loop, ev);
c70ed622 2213 ev = resolve_mixers (use_loop, tmp, folded_casts);
9baba81b
SP
2214
2215 if (use_loop == wrto_loop)
2216 return ev;
2217
2218 /* If the value of the use changes in the inner loop, we cannot express
2219 its value in the outer loop (we might try to return interval chrec,
2220 but we do not have a user for it anyway) */
2221 if (!no_evolution_in_loop_p (ev, use_loop->num, &val)
2222 || !val)
2223 return chrec_dont_know;
2224
9ba025a2 2225 use_loop = loop_outer (use_loop);
9baba81b
SP
2226 }
2227}
2228
eb0bc7af 2229
fdd43ac4
RB
2230/* Hashtable helpers for a temporary hash-table used when
2231 instantiating a CHREC or resolving mixers. For this use
2232 instantiated_below is always the same. */
2233
fdd43ac4 2234struct instantiate_cache_type
eb0bc7af 2235{
a3cc13cc
RB
2236 htab_t map;
2237 vec<scev_info_str> entries;
b8698a0f 2238
c3284718 2239 instantiate_cache_type () : map (NULL), entries (vNULL) {}
fdd43ac4 2240 ~instantiate_cache_type ();
0547c9b6
RB
2241 tree get (unsigned slot) { return entries[slot].chrec; }
2242 void set (unsigned slot, tree chrec) { entries[slot].chrec = chrec; }
fdd43ac4 2243};
eb0bc7af 2244
fdd43ac4
RB
2245instantiate_cache_type::~instantiate_cache_type ()
2246{
0547c9b6 2247 if (map != NULL)
fdd43ac4 2248 {
a3cc13cc 2249 htab_delete (map);
fdd43ac4
RB
2250 entries.release ();
2251 }
eb0bc7af
ZD
2252}
2253
a3cc13cc
RB
2254/* Cache to avoid infinite recursion when instantiating an SSA name.
2255 Live during the outermost instantiate_scev or resolve_mixers call. */
2256static instantiate_cache_type *global_cache;
2257
2258/* Computes a hash function for database element ELT. */
2259
2260static inline hashval_t
2261hash_idx_scev_info (const void *elt_)
2262{
2263 unsigned idx = ((size_t) elt_) - 2;
907dadbd 2264 return scev_info_hasher::hash (&global_cache->entries[idx]);
a3cc13cc
RB
2265}
2266
2267/* Compares database elements E1 and E2. */
2268
2269static inline int
2270eq_idx_scev_info (const void *e1, const void *e2)
2271{
2272 unsigned idx1 = ((size_t) e1) - 2;
907dadbd
TS
2273 return scev_info_hasher::equal (&global_cache->entries[idx1],
2274 (const scev_info_str *) e2);
a3cc13cc
RB
2275}
2276
0547c9b6 2277/* Returns from CACHE the slot number of the cached chrec for NAME. */
fdd43ac4 2278
0547c9b6 2279static unsigned
a3cc13cc
RB
2280get_instantiated_value_entry (instantiate_cache_type &cache,
2281 tree name, basic_block instantiate_below)
fdd43ac4 2282{
0547c9b6 2283 if (!cache.map)
fdd43ac4 2284 {
a3cc13cc 2285 cache.map = htab_create (10, hash_idx_scev_info, eq_idx_scev_info, NULL);
fdd43ac4
RB
2286 cache.entries.create (10);
2287 }
b8698a0f 2288
a3cc13cc
RB
2289 scev_info_str e;
2290 e.name_version = SSA_NAME_VERSION (name);
2291 e.instantiated_below = instantiate_below->index;
2292 void **slot = htab_find_slot_with_hash (cache.map, &e,
907dadbd 2293 scev_info_hasher::hash (&e), INSERT);
a3cc13cc 2294 if (!*slot)
fdd43ac4
RB
2295 {
2296 e.chrec = chrec_not_analyzed_yet;
a3cc13cc 2297 *slot = (void *)(size_t)(cache.entries.length () + 2);
fdd43ac4 2298 cache.entries.safe_push (e);
fdd43ac4
RB
2299 }
2300
a3cc13cc 2301 return ((size_t)*slot) - 2;
eb0bc7af
ZD
2302}
2303
0547c9b6 2304
18aed06a
SP
2305/* Return the closed_loop_phi node for VAR. If there is none, return
2306 NULL_TREE. */
2307
2308static tree
2309loop_closed_phi_def (tree var)
2310{
2311 struct loop *loop;
2312 edge exit;
538dd0b7
DM
2313 gphi *phi;
2314 gphi_iterator psi;
18aed06a
SP
2315
2316 if (var == NULL_TREE
2317 || TREE_CODE (var) != SSA_NAME)
2318 return NULL_TREE;
2319
2320 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var));
ac8f6c69 2321 exit = single_exit (loop);
18aed06a
SP
2322 if (!exit)
2323 return NULL_TREE;
2324
726a989a
RB
2325 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi))
2326 {
538dd0b7 2327 phi = psi.phi ();
726a989a
RB
2328 if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var)
2329 return PHI_RESULT (phi);
2330 }
18aed06a
SP
2331
2332 return NULL_TREE;
2333}
2334
8b679c9b 2335static tree instantiate_scev_r (basic_block, struct loop *, struct loop *,
c70ed622 2336 tree, bool *, int);
320f5a78
SP
2337
2338/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2339 and EVOLUTION_LOOP, that were left under a symbolic form.
2340
2495a183 2341 CHREC is an SSA_NAME to be instantiated.
320f5a78
SP
2342
2343 CACHE is the cache of already instantiated values.
2344
c70ed622
BC
2345 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2346 conversions that may wrap in signed/pointer type are folded, as long
2347 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2348 then we don't do such fold.
320f5a78
SP
2349
2350 SIZE_EXPR is used for computing the size of the expression to be
2351 instantiated, and to stop if it exceeds some limit. */
2352
2353static tree
2495a183 2354instantiate_scev_name (basic_block instantiate_below,
8b679c9b
RB
2355 struct loop *evolution_loop, struct loop *inner_loop,
2356 tree chrec,
c70ed622 2357 bool *fold_conversions,
4a8fb1a1 2358 int size_expr)
320f5a78 2359{
2495a183
SP
2360 tree res;
2361 struct loop *def_loop;
2362 basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (chrec));
20179b0d 2363
2495a183
SP
2364 /* A parameter (or loop invariant and we do not want to include
2365 evolutions in outer loops), nothing to do. */
2366 if (!def_bb
2367 || loop_depth (def_bb->loop_father) == 0
2368 || dominated_by_p (CDI_DOMINATORS, instantiate_below, def_bb))
2369 return chrec;
20179b0d 2370
2495a183
SP
2371 /* We cache the value of instantiated variable to avoid exponential
2372 time complexity due to reevaluations. We also store the convenient
2373 value in the cache in order to prevent infinite recursion -- we do
2374 not want to instantiate the SSA_NAME if it is in a mixer
2375 structure. This is used for avoiding the instantiation of
2376 recursively defined functions, such as:
320f5a78 2377
2495a183 2378 | a_2 -> {0, +, 1, +, a_2}_1 */
20179b0d 2379
a3cc13cc
RB
2380 unsigned si = get_instantiated_value_entry (*global_cache,
2381 chrec, instantiate_below);
0547c9b6
RB
2382 if (global_cache->get (si) != chrec_not_analyzed_yet)
2383 return global_cache->get (si);
20179b0d 2384
fdd43ac4 2385 /* On recursion return chrec_dont_know. */
0547c9b6 2386 global_cache->set (si, chrec_dont_know);
320f5a78 2387
2495a183
SP
2388 def_loop = find_common_loop (evolution_loop, def_bb->loop_father);
2389
320f5a78
SP
2390 /* If the analysis yields a parametric chrec, instantiate the
2391 result again. */
2392 res = analyze_scalar_evolution (def_loop, chrec);
2393
2847388e 2394 /* Don't instantiate default definitions. */
320f5a78 2395 if (TREE_CODE (res) == SSA_NAME
2847388e
SP
2396 && SSA_NAME_IS_DEFAULT_DEF (res))
2397 ;
2398
2399 /* Don't instantiate loop-closed-ssa phi nodes. */
2400 else if (TREE_CODE (res) == SSA_NAME
2401 && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res)))
2402 > loop_depth (def_loop))
320f5a78
SP
2403 {
2404 if (res == chrec)
2405 res = loop_closed_phi_def (chrec);
2406 else
2407 res = chrec;
2408
7472eb13
SP
2409 /* When there is no loop_closed_phi_def, it means that the
2410 variable is not used after the loop: try to still compute the
2411 value of the variable when exiting the loop. */
2412 if (res == NULL_TREE)
2413 {
2414 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (chrec));
2415 res = analyze_scalar_evolution (loop, chrec);
2416 res = compute_overall_effect_of_inner_loop (loop, res);
8b679c9b
RB
2417 res = instantiate_scev_r (instantiate_below, evolution_loop,
2418 inner_loop, res,
0547c9b6 2419 fold_conversions, size_expr);
7472eb13
SP
2420 }
2421 else if (!dominated_by_p (CDI_DOMINATORS, instantiate_below,
2422 gimple_bb (SSA_NAME_DEF_STMT (res))))
320f5a78
SP
2423 res = chrec_dont_know;
2424 }
2425
2426 else if (res != chrec_dont_know)
8b679c9b
RB
2427 {
2428 if (inner_loop
63fdb7be 2429 && def_bb->loop_father != inner_loop
8b679c9b
RB
2430 && !flow_loop_nested_p (def_bb->loop_father, inner_loop))
2431 /* ??? We could try to compute the overall effect of the loop here. */
2432 res = chrec_dont_know;
2433 else
2434 res = instantiate_scev_r (instantiate_below, evolution_loop,
2435 inner_loop, res,
0547c9b6 2436 fold_conversions, size_expr);
8b679c9b 2437 }
320f5a78
SP
2438
2439 /* Store the correct value to the cache. */
0547c9b6 2440 global_cache->set (si, res);
320f5a78 2441 return res;
320f5a78
SP
2442}
2443
ec6636eb
SP
2444/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2445 and EVOLUTION_LOOP, that were left under a symbolic form.
2446
2447 CHREC is a polynomial chain of recurrence to be instantiated.
2448
2449 CACHE is the cache of already instantiated values.
2450
c70ed622
BC
2451 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2452 conversions that may wrap in signed/pointer type are folded, as long
2453 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2454 then we don't do such fold.
ec6636eb
SP
2455
2456 SIZE_EXPR is used for computing the size of the expression to be
2457 instantiated, and to stop if it exceeds some limit. */
2458
2459static tree
2460instantiate_scev_poly (basic_block instantiate_below,
8b679c9b 2461 struct loop *evolution_loop, struct loop *,
c70ed622 2462 tree chrec, bool *fold_conversions, int size_expr)
ec6636eb
SP
2463{
2464 tree op1;
9e5dc77f 2465 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2466 get_chrec_loop (chrec),
0547c9b6 2467 CHREC_LEFT (chrec), fold_conversions,
ec6636eb
SP
2468 size_expr);
2469 if (op0 == chrec_dont_know)
2470 return chrec_dont_know;
2471
9e5dc77f 2472 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2473 get_chrec_loop (chrec),
0547c9b6 2474 CHREC_RIGHT (chrec), fold_conversions,
ec6636eb
SP
2475 size_expr);
2476 if (op1 == chrec_dont_know)
2477 return chrec_dont_know;
2478
2479 if (CHREC_LEFT (chrec) != op0
2480 || CHREC_RIGHT (chrec) != op1)
2481 {
2482 op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL);
8b679c9b 2483 chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1);
ec6636eb 2484 }
4bf4e169 2485
ec6636eb
SP
2486 return chrec;
2487}
2488
15fda317
SP
2489/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2490 and EVOLUTION_LOOP, that were left under a symbolic form.
2491
ffa34f4b 2492 "C0 CODE C1" is a binary expression of type TYPE to be instantiated.
15fda317
SP
2493
2494 CACHE is the cache of already instantiated values.
2495
c70ed622
BC
2496 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2497 conversions that may wrap in signed/pointer type are folded, as long
2498 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2499 then we don't do such fold.
15fda317
SP
2500
2501 SIZE_EXPR is used for computing the size of the expression to be
2502 instantiated, and to stop if it exceeds some limit. */
2503
2504static tree
2505instantiate_scev_binary (basic_block instantiate_below,
8b679c9b
RB
2506 struct loop *evolution_loop, struct loop *inner_loop,
2507 tree chrec, enum tree_code code,
ffa34f4b 2508 tree type, tree c0, tree c1,
c70ed622 2509 bool *fold_conversions, int size_expr)
15fda317
SP
2510{
2511 tree op1;
8b679c9b 2512 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop,
0547c9b6 2513 c0, fold_conversions, size_expr);
15fda317
SP
2514 if (op0 == chrec_dont_know)
2515 return chrec_dont_know;
2516
8b679c9b 2517 op1 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop,
0547c9b6 2518 c1, fold_conversions, size_expr);
15fda317
SP
2519 if (op1 == chrec_dont_know)
2520 return chrec_dont_know;
2521
ffa34f4b
SP
2522 if (c0 != op0
2523 || c1 != op1)
15fda317 2524 {
15fda317
SP
2525 op0 = chrec_convert (type, op0, NULL);
2526 op1 = chrec_convert_rhs (type, op1, NULL);
2527
ffa34f4b 2528 switch (code)
15fda317
SP
2529 {
2530 case POINTER_PLUS_EXPR:
2531 case PLUS_EXPR:
2532 return chrec_fold_plus (type, op0, op1);
2533
2534 case MINUS_EXPR:
2535 return chrec_fold_minus (type, op0, op1);
2536
2537 case MULT_EXPR:
2538 return chrec_fold_multiply (type, op0, op1);
2539
2540 default:
2541 gcc_unreachable ();
2542 }
2543 }
2544
ffa34f4b 2545 return chrec ? chrec : fold_build2 (code, type, c0, c1);
15fda317
SP
2546}
2547
dbc08079
SP
2548/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2549 and EVOLUTION_LOOP, that were left under a symbolic form.
2550
2551 "CHREC" is an array reference to be instantiated.
2552
2553 CACHE is the cache of already instantiated values.
2554
c70ed622
BC
2555 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2556 conversions that may wrap in signed/pointer type are folded, as long
2557 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2558 then we don't do such fold.
dbc08079
SP
2559
2560 SIZE_EXPR is used for computing the size of the expression to be
2561 instantiated, and to stop if it exceeds some limit. */
2562
2563static tree
2564instantiate_array_ref (basic_block instantiate_below,
8b679c9b 2565 struct loop *evolution_loop, struct loop *inner_loop,
c70ed622 2566 tree chrec, bool *fold_conversions, int size_expr)
dbc08079
SP
2567{
2568 tree res;
2569 tree index = TREE_OPERAND (chrec, 1);
8b679c9b
RB
2570 tree op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2571 inner_loop, index,
0547c9b6 2572 fold_conversions, size_expr);
dbc08079
SP
2573
2574 if (op1 == chrec_dont_know)
2575 return chrec_dont_know;
2576
2577 if (chrec && op1 == index)
2578 return chrec;
2579
2580 res = unshare_expr (chrec);
2581 TREE_OPERAND (res, 1) = op1;
2582 return res;
2583}
2584
a213b219 2585/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
9c382ce9
SP
2586 and EVOLUTION_LOOP, that were left under a symbolic form.
2587
2588 "CHREC" that stands for a convert expression "(TYPE) OP" is to be
2589 instantiated.
2590
2591 CACHE is the cache of already instantiated values.
2592
c70ed622
BC
2593 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2594 conversions that may wrap in signed/pointer type are folded, as long
2595 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2596 then we don't do such fold.
9c382ce9
SP
2597
2598 SIZE_EXPR is used for computing the size of the expression to be
2599 instantiated, and to stop if it exceeds some limit. */
2600
2601static tree
2602instantiate_scev_convert (basic_block instantiate_below,
8b679c9b 2603 struct loop *evolution_loop, struct loop *inner_loop,
0547c9b6 2604 tree chrec, tree type, tree op,
c70ed622 2605 bool *fold_conversions, int size_expr)
9c382ce9 2606{
8b679c9b
RB
2607 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2608 inner_loop, op,
0547c9b6 2609 fold_conversions, size_expr);
9c382ce9
SP
2610
2611 if (op0 == chrec_dont_know)
2612 return chrec_dont_know;
2613
2614 if (fold_conversions)
2615 {
c70ed622 2616 tree tmp = chrec_convert_aggressive (type, op0, fold_conversions);
9c382ce9
SP
2617 if (tmp)
2618 return tmp;
9c382ce9 2619
c70ed622
BC
2620 /* If we used chrec_convert_aggressive, we can no longer assume that
2621 signed chrecs do not overflow, as chrec_convert does, so avoid
2622 calling it in that case. */
2623 if (*fold_conversions)
2624 {
2625 if (chrec && op0 == op)
2626 return chrec;
9c382ce9 2627
c70ed622
BC
2628 return fold_convert (type, op0);
2629 }
2630 }
9c382ce9
SP
2631
2632 return chrec_convert (type, op0, NULL);
2633}
2634
7ec0665d
SP
2635/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2636 and EVOLUTION_LOOP, that were left under a symbolic form.
2637
4b9d48a1 2638 CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated.
7ec0665d 2639 Handle ~X as -1 - X.
4b9d48a1 2640 Handle -X as -1 * X.
7ec0665d
SP
2641
2642 CACHE is the cache of already instantiated values.
2643
c70ed622
BC
2644 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2645 conversions that may wrap in signed/pointer type are folded, as long
2646 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2647 then we don't do such fold.
7ec0665d
SP
2648
2649 SIZE_EXPR is used for computing the size of the expression to be
2650 instantiated, and to stop if it exceeds some limit. */
2651
2652static tree
4b9d48a1 2653instantiate_scev_not (basic_block instantiate_below,
8b679c9b
RB
2654 struct loop *evolution_loop, struct loop *inner_loop,
2655 tree chrec,
20179b0d 2656 enum tree_code code, tree type, tree op,
c70ed622 2657 bool *fold_conversions, int size_expr)
7ec0665d 2658{
8b679c9b
RB
2659 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2660 inner_loop, op,
0547c9b6 2661 fold_conversions, size_expr);
20179b0d 2662
7ec0665d
SP
2663 if (op0 == chrec_dont_know)
2664 return chrec_dont_know;
2665
20179b0d 2666 if (op != op0)
7ec0665d
SP
2667 {
2668 op0 = chrec_convert (type, op0, NULL);
4b9d48a1 2669
20179b0d 2670 switch (code)
4b9d48a1
SP
2671 {
2672 case BIT_NOT_EXPR:
2673 return chrec_fold_minus
2674 (type, fold_convert (type, integer_minus_one_node), op0);
2675
2676 case NEGATE_EXPR:
2677 return chrec_fold_multiply
2678 (type, fold_convert (type, integer_minus_one_node), op0);
2679
2680 default:
2681 gcc_unreachable ();
2682 }
7ec0665d 2683 }
4b9d48a1 2684
20179b0d 2685 return chrec ? chrec : fold_build1 (code, type, op0);
7ec0665d
SP
2686}
2687
d814176c
SP
2688/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2689 and EVOLUTION_LOOP, that were left under a symbolic form.
2690
2691 CHREC is an expression with 3 operands to be instantiated.
2692
2693 CACHE is the cache of already instantiated values.
2694
c70ed622
BC
2695 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2696 conversions that may wrap in signed/pointer type are folded, as long
2697 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2698 then we don't do such fold.
d814176c
SP
2699
2700 SIZE_EXPR is used for computing the size of the expression to be
2701 instantiated, and to stop if it exceeds some limit. */
2702
2703static tree
2704instantiate_scev_3 (basic_block instantiate_below,
8b679c9b
RB
2705 struct loop *evolution_loop, struct loop *inner_loop,
2706 tree chrec,
c70ed622 2707 bool *fold_conversions, int size_expr)
d814176c
SP
2708{
2709 tree op1, op2;
9e5dc77f 2710 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2711 inner_loop, TREE_OPERAND (chrec, 0),
0547c9b6 2712 fold_conversions, size_expr);
d814176c
SP
2713 if (op0 == chrec_dont_know)
2714 return chrec_dont_know;
2715
9e5dc77f 2716 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2717 inner_loop, TREE_OPERAND (chrec, 1),
0547c9b6 2718 fold_conversions, size_expr);
d814176c
SP
2719 if (op1 == chrec_dont_know)
2720 return chrec_dont_know;
2721
9e5dc77f 2722 op2 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2723 inner_loop, TREE_OPERAND (chrec, 2),
0547c9b6 2724 fold_conversions, size_expr);
d814176c
SP
2725 if (op2 == chrec_dont_know)
2726 return chrec_dont_know;
2727
2728 if (op0 == TREE_OPERAND (chrec, 0)
2729 && op1 == TREE_OPERAND (chrec, 1)
2730 && op2 == TREE_OPERAND (chrec, 2))
2731 return chrec;
2732
2733 return fold_build3 (TREE_CODE (chrec),
2734 TREE_TYPE (chrec), op0, op1, op2);
2735}
2736
9c382ce9
SP
2737/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2738 and EVOLUTION_LOOP, that were left under a symbolic form.
5b78fc3e 2739
9e5dc77f
SP
2740 CHREC is an expression with 2 operands to be instantiated.
2741
2742 CACHE is the cache of already instantiated values.
2743
c70ed622
BC
2744 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2745 conversions that may wrap in signed/pointer type are folded, as long
2746 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2747 then we don't do such fold.
9e5dc77f
SP
2748
2749 SIZE_EXPR is used for computing the size of the expression to be
2750 instantiated, and to stop if it exceeds some limit. */
2751
2752static tree
2753instantiate_scev_2 (basic_block instantiate_below,
8b679c9b
RB
2754 struct loop *evolution_loop, struct loop *inner_loop,
2755 tree chrec,
c70ed622 2756 bool *fold_conversions, int size_expr)
9e5dc77f
SP
2757{
2758 tree op1;
2759 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2760 inner_loop, TREE_OPERAND (chrec, 0),
0547c9b6 2761 fold_conversions, size_expr);
9e5dc77f
SP
2762 if (op0 == chrec_dont_know)
2763 return chrec_dont_know;
2764
2765 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2766 inner_loop, TREE_OPERAND (chrec, 1),
0547c9b6 2767 fold_conversions, size_expr);
9e5dc77f
SP
2768 if (op1 == chrec_dont_know)
2769 return chrec_dont_know;
2770
2771 if (op0 == TREE_OPERAND (chrec, 0)
2772 && op1 == TREE_OPERAND (chrec, 1))
2773 return chrec;
2774
2775 return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1);
2776}
2777
2778/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2779 and EVOLUTION_LOOP, that were left under a symbolic form.
2780
2781 CHREC is an expression with 2 operands to be instantiated.
5b78fc3e
JS
2782
2783 CACHE is the cache of already instantiated values.
2784
c70ed622
BC
2785 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2786 conversions that may wrap in signed/pointer type are folded, as long
2787 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2788 then we don't do such fold.
5b78fc3e 2789
3f227a8c
JS
2790 SIZE_EXPR is used for computing the size of the expression to be
2791 instantiated, and to stop if it exceeds some limit. */
9c382ce9 2792
9baba81b 2793static tree
a213b219 2794instantiate_scev_1 (basic_block instantiate_below,
8b679c9b
RB
2795 struct loop *evolution_loop, struct loop *inner_loop,
2796 tree chrec,
c70ed622 2797 bool *fold_conversions, int size_expr)
9baba81b 2798{
9e5dc77f 2799 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2800 inner_loop, TREE_OPERAND (chrec, 0),
0547c9b6 2801 fold_conversions, size_expr);
9e5dc77f
SP
2802
2803 if (op0 == chrec_dont_know)
2804 return chrec_dont_know;
2805
2806 if (op0 == TREE_OPERAND (chrec, 0))
2807 return chrec;
2808
2809 return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0);
2810}
2811
2812/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2813 and EVOLUTION_LOOP, that were left under a symbolic form.
2814
2815 CHREC is the scalar evolution to instantiate.
2816
2817 CACHE is the cache of already instantiated values.
2282a0e6 2818
c70ed622
BC
2819 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2820 conversions that may wrap in signed/pointer type are folded, as long
2821 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2822 then we don't do such fold.
9e5dc77f
SP
2823
2824 SIZE_EXPR is used for computing the size of the expression to be
2825 instantiated, and to stop if it exceeds some limit. */
2826
2827static tree
2828instantiate_scev_r (basic_block instantiate_below,
8b679c9b
RB
2829 struct loop *evolution_loop, struct loop *inner_loop,
2830 tree chrec,
c70ed622 2831 bool *fold_conversions, int size_expr)
9e5dc77f 2832{
47ae9e4c
SP
2833 /* Give up if the expression is larger than the MAX that we allow. */
2834 if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
2835 return chrec_dont_know;
2836
81fada9a
JJ
2837 if (chrec == NULL_TREE
2838 || automatically_generated_chrec_p (chrec)
d7770457 2839 || is_gimple_min_invariant (chrec))
9baba81b
SP
2840 return chrec;
2841
2842 switch (TREE_CODE (chrec))
2843 {
2844 case SSA_NAME:
8b679c9b
RB
2845 return instantiate_scev_name (instantiate_below, evolution_loop,
2846 inner_loop, chrec,
0547c9b6 2847 fold_conversions, size_expr);
9baba81b
SP
2848
2849 case POLYNOMIAL_CHREC:
8b679c9b
RB
2850 return instantiate_scev_poly (instantiate_below, evolution_loop,
2851 inner_loop, chrec,
0547c9b6 2852 fold_conversions, size_expr);
9baba81b 2853
5be014d5 2854 case POINTER_PLUS_EXPR:
9baba81b 2855 case PLUS_EXPR:
9baba81b 2856 case MINUS_EXPR:
9baba81b 2857 case MULT_EXPR:
8b679c9b
RB
2858 return instantiate_scev_binary (instantiate_below, evolution_loop,
2859 inner_loop, chrec,
ffa34f4b
SP
2860 TREE_CODE (chrec), chrec_type (chrec),
2861 TREE_OPERAND (chrec, 0),
2862 TREE_OPERAND (chrec, 1),
0547c9b6 2863 fold_conversions, size_expr);
9baba81b 2864
1043771b 2865 CASE_CONVERT:
8b679c9b
RB
2866 return instantiate_scev_convert (instantiate_below, evolution_loop,
2867 inner_loop, chrec,
9c382ce9 2868 TREE_TYPE (chrec), TREE_OPERAND (chrec, 0),
0547c9b6 2869 fold_conversions, size_expr);
9baba81b 2870
4b9d48a1 2871 case NEGATE_EXPR:
418df9d7 2872 case BIT_NOT_EXPR:
8b679c9b
RB
2873 return instantiate_scev_not (instantiate_below, evolution_loop,
2874 inner_loop, chrec,
20179b0d
SP
2875 TREE_CODE (chrec), TREE_TYPE (chrec),
2876 TREE_OPERAND (chrec, 0),
0547c9b6 2877 fold_conversions, size_expr);
418df9d7 2878
4c7d6755 2879 case ADDR_EXPR:
9baba81b
SP
2880 case SCEV_NOT_KNOWN:
2881 return chrec_dont_know;
2882
2883 case SCEV_KNOWN:
2884 return chrec_known;
15fda317 2885
dbc08079 2886 case ARRAY_REF:
8b679c9b
RB
2887 return instantiate_array_ref (instantiate_below, evolution_loop,
2888 inner_loop, chrec,
0547c9b6 2889 fold_conversions, size_expr);
dbc08079 2890
9baba81b
SP
2891 default:
2892 break;
2893 }
2894
0dfb0dc6
SP
2895 if (VL_EXP_CLASS_P (chrec))
2896 return chrec_dont_know;
2897
9baba81b
SP
2898 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
2899 {
2900 case 3:
8b679c9b
RB
2901 return instantiate_scev_3 (instantiate_below, evolution_loop,
2902 inner_loop, chrec,
0547c9b6 2903 fold_conversions, size_expr);
9baba81b
SP
2904
2905 case 2:
8b679c9b
RB
2906 return instantiate_scev_2 (instantiate_below, evolution_loop,
2907 inner_loop, chrec,
0547c9b6 2908 fold_conversions, size_expr);
7ec0665d 2909
9baba81b 2910 case 1:
8b679c9b
RB
2911 return instantiate_scev_1 (instantiate_below, evolution_loop,
2912 inner_loop, chrec,
0547c9b6 2913 fold_conversions, size_expr);
9baba81b
SP
2914
2915 case 0:
2916 return chrec;
2917
2918 default:
2919 break;
2920 }
2921
2922 /* Too complicated to handle. */
2923 return chrec_dont_know;
2924}
e9eb809d
ZD
2925
2926/* Analyze all the parameters of the chrec that were left under a
a213b219
SP
2927 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2928 recursive instantiation of parameters: a parameter is a variable
2929 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2930 a function parameter. */
e9eb809d
ZD
2931
2932tree
a213b219 2933instantiate_scev (basic_block instantiate_below, struct loop *evolution_loop,
3f227a8c 2934 tree chrec)
e9eb809d 2935{
9baba81b
SP
2936 tree res;
2937
dfedbe40 2938 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b 2939 {
3f227a8c 2940 fprintf (dump_file, "(instantiate_scev \n");
a213b219 2941 fprintf (dump_file, " (instantiate_below = %d)\n", instantiate_below->index);
3f227a8c 2942 fprintf (dump_file, " (evolution_loop = %d)\n", evolution_loop->num);
9baba81b
SP
2943 fprintf (dump_file, " (chrec = ");
2944 print_generic_expr (dump_file, chrec, 0);
2945 fprintf (dump_file, ")\n");
2946 }
b8698a0f 2947
0547c9b6
RB
2948 bool destr = false;
2949 if (!global_cache)
2950 {
2951 global_cache = new instantiate_cache_type;
2952 destr = true;
2953 }
2954
8b679c9b 2955 res = instantiate_scev_r (instantiate_below, evolution_loop,
c70ed622 2956 NULL, chrec, NULL, 0);
0547c9b6
RB
2957
2958 if (destr)
2959 {
2960 delete global_cache;
2961 global_cache = NULL;
2962 }
9baba81b 2963
dfedbe40 2964 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
2965 {
2966 fprintf (dump_file, " (res = ");
2967 print_generic_expr (dump_file, res, 0);
2968 fprintf (dump_file, "))\n");
2969 }
eb0bc7af 2970
9baba81b
SP
2971 return res;
2972}
2973
2974/* Similar to instantiate_parameters, but does not introduce the
2282a0e6
ZD
2975 evolutions in outer loops for LOOP invariants in CHREC, and does not
2976 care about causing overflows, as long as they do not affect value
2977 of an expression. */
9baba81b 2978
3cb960c7 2979tree
c70ed622 2980resolve_mixers (struct loop *loop, tree chrec, bool *folded_casts)
9baba81b 2981{
0547c9b6 2982 bool destr = false;
c70ed622 2983 bool fold_conversions = false;
0547c9b6
RB
2984 if (!global_cache)
2985 {
2986 global_cache = new instantiate_cache_type;
2987 destr = true;
2988 }
2989
8b679c9b 2990 tree ret = instantiate_scev_r (block_before_loop (loop), loop, NULL,
c70ed622
BC
2991 chrec, &fold_conversions, 0);
2992
2993 if (folded_casts && !*folded_casts)
2994 *folded_casts = fold_conversions;
0547c9b6
RB
2995
2996 if (destr)
2997 {
2998 delete global_cache;
2999 global_cache = NULL;
3000 }
3001
eb0bc7af 3002 return ret;
9baba81b
SP
3003}
3004
b8698a0f 3005/* Entry point for the analysis of the number of iterations pass.
9baba81b
SP
3006 This function tries to safely approximate the number of iterations
3007 the loop will run. When this property is not decidable at compile
0a74c758
SP
3008 time, the result is chrec_dont_know. Otherwise the result is a
3009 scalar or a symbolic parameter. When the number of iterations may
3010 be equal to zero and the property cannot be determined at compile
3011 time, the result is a COND_EXPR that represents in a symbolic form
3012 the conditions under which the number of iterations is not zero.
b8698a0f 3013
9baba81b 3014 Example of analysis: suppose that the loop has an exit condition:
b8698a0f 3015
9baba81b 3016 "if (b > 49) goto end_loop;"
b8698a0f 3017
9baba81b
SP
3018 and that in a previous analysis we have determined that the
3019 variable 'b' has an evolution function:
b8698a0f
L
3020
3021 "EF = {23, +, 5}_2".
3022
9baba81b
SP
3023 When we evaluate the function at the point 5, i.e. the value of the
3024 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
3025 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
3026 the loop body has been executed 6 times. */
3027
b8698a0f 3028tree
a14865db 3029number_of_latch_executions (struct loop *loop)
9baba81b 3030{
9baba81b
SP
3031 edge exit;
3032 struct tree_niter_desc niter_desc;
0a74c758
SP
3033 tree may_be_zero;
3034 tree res;
9baba81b 3035
0a74c758 3036 /* Determine whether the number of iterations in loop has already
9baba81b
SP
3037 been computed. */
3038 res = loop->nb_iterations;
3039 if (res)
3040 return res;
0a74c758
SP
3041
3042 may_be_zero = NULL_TREE;
9baba81b 3043
dfedbe40 3044 if (dump_file && (dump_flags & TDF_SCEV))
0a74c758 3045 fprintf (dump_file, "(number_of_iterations_in_loop = \n");
b8698a0f 3046
0a74c758 3047 res = chrec_dont_know;
ac8f6c69 3048 exit = single_exit (loop);
9baba81b 3049
0a74c758
SP
3050 if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false))
3051 {
3052 may_be_zero = niter_desc.may_be_zero;
3053 res = niter_desc.niter;
3054 }
3055
3056 if (res == chrec_dont_know
3057 || !may_be_zero
3058 || integer_zerop (may_be_zero))
3059 ;
3060 else if (integer_nonzerop (may_be_zero))
3061 res = build_int_cst (TREE_TYPE (res), 0);
9baba81b 3062
0a74c758
SP
3063 else if (COMPARISON_CLASS_P (may_be_zero))
3064 res = fold_build3 (COND_EXPR, TREE_TYPE (res), may_be_zero,
3065 build_int_cst (TREE_TYPE (res), 0), res);
9baba81b
SP
3066 else
3067 res = chrec_dont_know;
3068
dfedbe40 3069 if (dump_file && (dump_flags & TDF_SCEV))
0a74c758
SP
3070 {
3071 fprintf (dump_file, " (set_nb_iterations_in_loop = ");
3072 print_generic_expr (dump_file, res, 0);
3073 fprintf (dump_file, "))\n");
3074 }
3075
3076 loop->nb_iterations = res;
3077 return res;
9baba81b 3078}
9baba81b
SP
3079\f
3080
3081/* Counters for the stats. */
3082
b8698a0f 3083struct chrec_stats
9baba81b
SP
3084{
3085 unsigned nb_chrecs;
3086 unsigned nb_affine;
3087 unsigned nb_affine_multivar;
3088 unsigned nb_higher_poly;
3089 unsigned nb_chrec_dont_know;
3090 unsigned nb_undetermined;
3091};
3092
3093/* Reset the counters. */
3094
3095static inline void
3096reset_chrecs_counters (struct chrec_stats *stats)
3097{
3098 stats->nb_chrecs = 0;
3099 stats->nb_affine = 0;
3100 stats->nb_affine_multivar = 0;
3101 stats->nb_higher_poly = 0;
3102 stats->nb_chrec_dont_know = 0;
3103 stats->nb_undetermined = 0;
3104}
3105
3106/* Dump the contents of a CHREC_STATS structure. */
3107
3108static void
3109dump_chrecs_stats (FILE *file, struct chrec_stats *stats)
3110{
3111 fprintf (file, "\n(\n");
3112 fprintf (file, "-----------------------------------------\n");
3113 fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine);
3114 fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar);
b8698a0f 3115 fprintf (file, "%d\tdegree greater than 2 polynomials\n",
9baba81b
SP
3116 stats->nb_higher_poly);
3117 fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know);
3118 fprintf (file, "-----------------------------------------\n");
3119 fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs);
b8698a0f 3120 fprintf (file, "%d\twith undetermined coefficients\n",
9baba81b
SP
3121 stats->nb_undetermined);
3122 fprintf (file, "-----------------------------------------\n");
b8698a0f 3123 fprintf (file, "%d\tchrecs in the scev database\n",
907dadbd 3124 (int) scalar_evolution_info->elements ());
9baba81b
SP
3125 fprintf (file, "%d\tsets in the scev database\n", nb_set_scev);
3126 fprintf (file, "%d\tgets in the scev database\n", nb_get_scev);
3127 fprintf (file, "-----------------------------------------\n");
3128 fprintf (file, ")\n\n");
3129}
3130
3131/* Gather statistics about CHREC. */
3132
3133static void
3134gather_chrec_stats (tree chrec, struct chrec_stats *stats)
3135{
3136 if (dump_file && (dump_flags & TDF_STATS))
3137 {
3138 fprintf (dump_file, "(classify_chrec ");
3139 print_generic_expr (dump_file, chrec, 0);
3140 fprintf (dump_file, "\n");
3141 }
b8698a0f 3142
9baba81b 3143 stats->nb_chrecs++;
b8698a0f 3144
9baba81b
SP
3145 if (chrec == NULL_TREE)
3146 {
3147 stats->nb_undetermined++;
3148 return;
3149 }
b8698a0f 3150
9baba81b
SP
3151 switch (TREE_CODE (chrec))
3152 {
3153 case POLYNOMIAL_CHREC:
3154 if (evolution_function_is_affine_p (chrec))
3155 {
3156 if (dump_file && (dump_flags & TDF_STATS))
3157 fprintf (dump_file, " affine_univariate\n");
3158 stats->nb_affine++;
3159 }
a50411de 3160 else if (evolution_function_is_affine_multivariate_p (chrec, 0))
9baba81b
SP
3161 {
3162 if (dump_file && (dump_flags & TDF_STATS))
3163 fprintf (dump_file, " affine_multivariate\n");
3164 stats->nb_affine_multivar++;
3165 }
3166 else
3167 {
3168 if (dump_file && (dump_flags & TDF_STATS))
3169 fprintf (dump_file, " higher_degree_polynomial\n");
3170 stats->nb_higher_poly++;
3171 }
b8698a0f 3172
9baba81b
SP
3173 break;
3174
3175 default:
3176 break;
3177 }
b8698a0f 3178
9baba81b
SP
3179 if (chrec_contains_undetermined (chrec))
3180 {
3181 if (dump_file && (dump_flags & TDF_STATS))
3182 fprintf (dump_file, " undetermined\n");
3183 stats->nb_undetermined++;
3184 }
b8698a0f 3185
9baba81b
SP
3186 if (dump_file && (dump_flags & TDF_STATS))
3187 fprintf (dump_file, ")\n");
3188}
3189
9baba81b
SP
3190/* Classify the chrecs of the whole database. */
3191
b8698a0f 3192void
9baba81b
SP
3193gather_stats_on_scev_database (void)
3194{
3195 struct chrec_stats stats;
b8698a0f 3196
9baba81b
SP
3197 if (!dump_file)
3198 return;
b8698a0f 3199
9baba81b 3200 reset_chrecs_counters (&stats);
b8698a0f 3201
907dadbd
TS
3202 hash_table<scev_info_hasher>::iterator iter;
3203 scev_info_str *elt;
3204 FOR_EACH_HASH_TABLE_ELEMENT (*scalar_evolution_info, elt, scev_info_str *,
3205 iter)
3206 gather_chrec_stats (elt->chrec, &stats);
9baba81b
SP
3207
3208 dump_chrecs_stats (dump_file, &stats);
3209}
3210
3211\f
3212
3213/* Initializer. */
3214
3215static void
3216initialize_scalar_evolutions_analyzer (void)
3217{
3218 /* The elements below are unique. */
3219 if (chrec_dont_know == NULL_TREE)
3220 {
3221 chrec_not_analyzed_yet = NULL_TREE;
3222 chrec_dont_know = make_node (SCEV_NOT_KNOWN);
3223 chrec_known = make_node (SCEV_KNOWN);
d5ab5675
ZD
3224 TREE_TYPE (chrec_dont_know) = void_type_node;
3225 TREE_TYPE (chrec_known) = void_type_node;
9baba81b
SP
3226 }
3227}
3228
3229/* Initialize the analysis of scalar evolutions for LOOPS. */
3230
3231void
d73be268 3232scev_initialize (void)
9baba81b 3233{
42fd6772 3234 struct loop *loop;
9baba81b 3235
907dadbd 3236 scalar_evolution_info = hash_table<scev_info_hasher>::create_ggc (100);
b8698a0f 3237
9baba81b
SP
3238 initialize_scalar_evolutions_analyzer ();
3239
f0bd40b1 3240 FOR_EACH_LOOP (loop, 0)
42fd6772
ZD
3241 {
3242 loop->nb_iterations = NULL_TREE;
3243 }
9baba81b
SP
3244}
3245
e3a8f1fa
JH
3246/* Return true if SCEV is initialized. */
3247
3248bool
3249scev_initialized_p (void)
3250{
3251 return scalar_evolution_info != NULL;
3252}
3253
a7bf45de
SP
3254/* Cleans up the information cached by the scalar evolutions analysis
3255 in the hash table. */
3256
3257void
3258scev_reset_htab (void)
3259{
3260 if (!scalar_evolution_info)
3261 return;
3262
907dadbd 3263 scalar_evolution_info->empty ();
a7bf45de
SP
3264}
3265
3266/* Cleans up the information cached by the scalar evolutions analysis
3267 in the hash table and in the loop->nb_iterations. */
9baba81b
SP
3268
3269void
3270scev_reset (void)
3271{
9baba81b
SP
3272 struct loop *loop;
3273
a7bf45de
SP
3274 scev_reset_htab ();
3275
f0bd40b1 3276 FOR_EACH_LOOP (loop, 0)
9baba81b 3277 {
42fd6772 3278 loop->nb_iterations = NULL_TREE;
9baba81b 3279 }
e9eb809d
ZD
3280}
3281
f017bf5e
ZD
3282/* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with
3283 respect to WRTO_LOOP and returns its base and step in IV if possible
3284 (see analyze_scalar_evolution_in_loop for more details on USE_LOOP
3285 and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be
3286 invariant in LOOP. Otherwise we require it to be an integer constant.
b8698a0f 3287
f017bf5e
ZD
3288 IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g.
3289 because it is computed in signed arithmetics). Consequently, adding an
3290 induction variable
b8698a0f 3291
f017bf5e
ZD
3292 for (i = IV->base; ; i += IV->step)
3293
3294 is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is
3295 false for the type of the induction variable, or you can prove that i does
3296 not wrap by some other argument. Otherwise, this might introduce undefined
3297 behavior, and
b8698a0f 3298
f017bf5e
ZD
3299 for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step))
3300
3301 must be used instead. */
e9eb809d
ZD
3302
3303bool
f017bf5e
ZD
3304simple_iv (struct loop *wrto_loop, struct loop *use_loop, tree op,
3305 affine_iv *iv, bool allow_nonconstant_step)
e9eb809d 3306{
f3c5f3a3
BC
3307 enum tree_code code;
3308 tree type, ev, base, e, stop;
3309 wide_int extreme;
3310 bool folded_casts, overflow;
9baba81b 3311
a6f778b2
ZD
3312 iv->base = NULL_TREE;
3313 iv->step = NULL_TREE;
3314 iv->no_overflow = false;
9baba81b
SP
3315
3316 type = TREE_TYPE (op);
1ee0d660
EB
3317 if (!POINTER_TYPE_P (type)
3318 && !INTEGRAL_TYPE_P (type))
9baba81b
SP
3319 return false;
3320
f017bf5e 3321 ev = analyze_scalar_evolution_in_loop (wrto_loop, use_loop, op,
a6f778b2 3322 &folded_casts);
f017bf5e
ZD
3323 if (chrec_contains_undetermined (ev)
3324 || chrec_contains_symbols_defined_in_loop (ev, wrto_loop->num))
9baba81b
SP
3325 return false;
3326
f017bf5e 3327 if (tree_does_not_contain_chrecs (ev))
9baba81b 3328 {
a6f778b2 3329 iv->base = ev;
6e42ce54 3330 iv->step = build_int_cst (TREE_TYPE (ev), 0);
a6f778b2 3331 iv->no_overflow = true;
9baba81b
SP
3332 return true;
3333 }
3334
3335 if (TREE_CODE (ev) != POLYNOMIAL_CHREC
f017bf5e 3336 || CHREC_VARIABLE (ev) != (unsigned) wrto_loop->num)
9baba81b
SP
3337 return false;
3338
a6f778b2 3339 iv->step = CHREC_RIGHT (ev);
f017bf5e
ZD
3340 if ((!allow_nonconstant_step && TREE_CODE (iv->step) != INTEGER_CST)
3341 || tree_contains_chrecs (iv->step, NULL))
9baba81b 3342 return false;
9be872b7 3343
a6f778b2 3344 iv->base = CHREC_LEFT (ev);
f017bf5e 3345 if (tree_contains_chrecs (iv->base, NULL))
9baba81b
SP
3346 return false;
3347
20bd649a
MP
3348 iv->no_overflow = (!folded_casts && ANY_INTEGRAL_TYPE_P (type)
3349 && TYPE_OVERFLOW_UNDEFINED (type));
eeef0e45 3350
f3c5f3a3
BC
3351 /* Try to simplify iv base:
3352
3353 (signed T) ((unsigned T)base + step) ;; TREE_TYPE (base) == signed T
3354 == (signed T)(unsigned T)base + step
3355 == base + step
3356
3357 If we can prove operation (base + step) doesn't overflow or underflow.
3358 Specifically, we try to prove below conditions are satisfied:
3359
3360 base <= UPPER_BOUND (type) - step ;;step > 0
3361 base >= LOWER_BOUND (type) - step ;;step < 0
3362
3363 This is done by proving the reverse conditions are false using loop's
3364 initial conditions.
3365
3366 The is necessary to make loop niter, or iv overflow analysis easier
3367 for below example:
3368
3369 int foo (int *a, signed char s, signed char l)
3370 {
3371 signed char i;
3372 for (i = s; i < l; i++)
3373 a[i] = 0;
3374 return 0;
3375 }
3376
3377 Note variable I is firstly converted to type unsigned char, incremented,
3378 then converted back to type signed char. */
3379
3380 if (wrto_loop->num != use_loop->num)
3381 return true;
3382
3383 if (!CONVERT_EXPR_P (iv->base) || TREE_CODE (iv->step) != INTEGER_CST)
3384 return true;
3385
3386 type = TREE_TYPE (iv->base);
3387 e = TREE_OPERAND (iv->base, 0);
3388 if (TREE_CODE (e) != PLUS_EXPR
3389 || TREE_CODE (TREE_OPERAND (e, 1)) != INTEGER_CST
3390 || !tree_int_cst_equal (iv->step,
3391 fold_convert (type, TREE_OPERAND (e, 1))))
3392 return true;
3393 e = TREE_OPERAND (e, 0);
3394 if (!CONVERT_EXPR_P (e))
3395 return true;
3396 base = TREE_OPERAND (e, 0);
3397 if (!useless_type_conversion_p (type, TREE_TYPE (base)))
3398 return true;
3399
3400 if (tree_int_cst_sign_bit (iv->step))
3401 {
3402 code = LT_EXPR;
3403 extreme = wi::min_value (type);
3404 }
3405 else
3406 {
3407 code = GT_EXPR;
3408 extreme = wi::max_value (type);
3409 }
3410 overflow = false;
3411 extreme = wi::sub (extreme, iv->step, TYPE_SIGN (type), &overflow);
3412 if (overflow)
3413 return true;
3414 e = fold_build2 (code, boolean_type_node, base,
3415 wide_int_to_tree (type, extreme));
3416 stop = (TREE_CODE (base) == SSA_NAME) ? base : NULL;
3417 e = simplify_using_initial_conditions (use_loop, e, stop);
3418 if (!integer_zerop (e))
3419 return true;
3420
3421 if (POINTER_TYPE_P (TREE_TYPE (base)))
3422 code = POINTER_PLUS_EXPR;
3423 else
3424 code = PLUS_EXPR;
3425
3426 iv->base = fold_build2 (code, TREE_TYPE (base), base, iv->step);
9baba81b
SP
3427 return true;
3428}
3429
9baba81b
SP
3430/* Finalize the scalar evolution analysis. */
3431
3432void
3433scev_finalize (void)
3434{
d51157de
ZD
3435 if (!scalar_evolution_info)
3436 return;
907dadbd 3437 scalar_evolution_info->empty ();
c7b852c8 3438 scalar_evolution_info = NULL;
9baba81b
SP
3439}
3440
771f882e
ZD
3441/* Returns true if the expression EXPR is considered to be too expensive
3442 for scev_const_prop. */
3443
3444bool
3445expression_expensive_p (tree expr)
3446{
3447 enum tree_code code;
3448
3449 if (is_gimple_val (expr))
3450 return false;
3451
3452 code = TREE_CODE (expr);
3453 if (code == TRUNC_DIV_EXPR
3454 || code == CEIL_DIV_EXPR
3455 || code == FLOOR_DIV_EXPR
3456 || code == ROUND_DIV_EXPR
3457 || code == TRUNC_MOD_EXPR
3458 || code == CEIL_MOD_EXPR
3459 || code == FLOOR_MOD_EXPR
3460 || code == ROUND_MOD_EXPR
3461 || code == EXACT_DIV_EXPR)
3462 {
3463 /* Division by power of two is usually cheap, so we allow it.
3464 Forbid anything else. */
3465 if (!integer_pow2p (TREE_OPERAND (expr, 1)))
3466 return true;
3467 }
3468
3469 switch (TREE_CODE_CLASS (code))
3470 {
3471 case tcc_binary:
3472 case tcc_comparison:
3473 if (expression_expensive_p (TREE_OPERAND (expr, 1)))
3474 return true;
3475
3476 /* Fallthru. */
3477 case tcc_unary:
3478 return expression_expensive_p (TREE_OPERAND (expr, 0));
3479
3480 default:
3481 return true;
3482 }
3483}
3484
f993a853
TV
3485/* Do final value replacement for LOOP. */
3486
3487void
3488final_value_replacement_loop (struct loop *loop)
3489{
3490 /* If we do not know exact number of iterations of the loop, we cannot
3491 replace the final value. */
3492 edge exit = single_exit (loop);
3493 if (!exit)
3494 return;
3495
3496 tree niter = number_of_latch_executions (loop);
3497 if (niter == chrec_dont_know)
3498 return;
3499
3500 /* Ensure that it is possible to insert new statements somewhere. */
3501 if (!single_pred_p (exit->dest))
3502 split_loop_exit_edge (exit);
3503
3504 /* Set stmt insertion pointer. All stmts are inserted before this point. */
3505 gimple_stmt_iterator gsi = gsi_after_labels (exit->dest);
3506
3507 struct loop *ex_loop
3508 = superloop_at_depth (loop,
3509 loop_depth (exit->dest->loop_father) + 1);
3510
3511 gphi_iterator psi;
3512 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); )
3513 {
3514 gphi *phi = psi.phi ();
3515 tree rslt = PHI_RESULT (phi);
3516 tree def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
3517 if (virtual_operand_p (def))
3518 {
3519 gsi_next (&psi);
3520 continue;
3521 }
3522
3523 if (!POINTER_TYPE_P (TREE_TYPE (def))
3524 && !INTEGRAL_TYPE_P (TREE_TYPE (def)))
3525 {
3526 gsi_next (&psi);
3527 continue;
3528 }
3529
3530 bool folded_casts;
3531 def = analyze_scalar_evolution_in_loop (ex_loop, loop, def,
3532 &folded_casts);
3533 def = compute_overall_effect_of_inner_loop (ex_loop, def);
3534 if (!tree_does_not_contain_chrecs (def)
3535 || chrec_contains_symbols_defined_in_loop (def, ex_loop->num)
3536 /* Moving the computation from the loop may prolong life range
3537 of some ssa names, which may cause problems if they appear
3538 on abnormal edges. */
3539 || contains_abnormal_ssa_name_p (def)
3540 /* Do not emit expensive expressions. The rationale is that
3541 when someone writes a code like
3542
3543 while (n > 45) n -= 45;
3544
3545 he probably knows that n is not large, and does not want it
3546 to be turned into n %= 45. */
3547 || expression_expensive_p (def))
3548 {
3549 if (dump_file && (dump_flags & TDF_DETAILS))
3550 {
3551 fprintf (dump_file, "not replacing:\n ");
3552 print_gimple_stmt (dump_file, phi, 0, 0);
3553 fprintf (dump_file, "\n");
3554 }
3555 gsi_next (&psi);
3556 continue;
3557 }
3558
3559 /* Eliminate the PHI node and replace it by a computation outside
3560 the loop. */
3561 if (dump_file)
3562 {
3563 fprintf (dump_file, "\nfinal value replacement:\n ");
3564 print_gimple_stmt (dump_file, phi, 0, 0);
3565 fprintf (dump_file, " with\n ");
3566 }
3567 def = unshare_expr (def);
3568 remove_phi_node (&psi, false);
3569
3570 /* If def's type has undefined overflow and there were folded
3571 casts, rewrite all stmts added for def into arithmetics
3572 with defined overflow behavior. */
3573 if (folded_casts && ANY_INTEGRAL_TYPE_P (TREE_TYPE (def))
3574 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (def)))
3575 {
3576 gimple_seq stmts;
3577 gimple_stmt_iterator gsi2;
3578 def = force_gimple_operand (def, &stmts, true, NULL_TREE);
3579 gsi2 = gsi_start (stmts);
3580 while (!gsi_end_p (gsi2))
3581 {
3582 gimple *stmt = gsi_stmt (gsi2);
3583 gimple_stmt_iterator gsi3 = gsi2;
3584 gsi_next (&gsi2);
3585 gsi_remove (&gsi3, false);
3586 if (is_gimple_assign (stmt)
3587 && arith_code_with_undefined_signed_overflow
3588 (gimple_assign_rhs_code (stmt)))
3589 gsi_insert_seq_before (&gsi,
3590 rewrite_to_defined_overflow (stmt),
3591 GSI_SAME_STMT);
3592 else
3593 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
3594 }
3595 }
3596 else
3597 def = force_gimple_operand_gsi (&gsi, def, false, NULL_TREE,
3598 true, GSI_SAME_STMT);
3599
3600 gassign *ass = gimple_build_assign (rslt, def);
3601 gsi_insert_before (&gsi, ass, GSI_SAME_STMT);
3602 if (dump_file)
3603 {
3604 print_gimple_stmt (dump_file, ass, 0, 0);
3605 fprintf (dump_file, "\n");
3606 }
3607 }
3608}
3609
684aaf29 3610/* Replace ssa names for that scev can prove they are constant by the
3ac01fde
ZD
3611 appropriate constants. Also perform final value replacement in loops,
3612 in case the replacement expressions are cheap.
b8698a0f 3613
684aaf29
ZD
3614 We only consider SSA names defined by phi nodes; rest is left to the
3615 ordinary constant propagation pass. */
3616
c2924966 3617unsigned int
684aaf29
ZD
3618scev_const_prop (void)
3619{
3620 basic_block bb;
726a989a 3621 tree name, type, ev;
538dd0b7 3622 gphi *phi;
f993a853 3623 struct loop *loop;
684aaf29 3624 bitmap ssa_names_to_remove = NULL;
3ac01fde 3625 unsigned i;
538dd0b7 3626 gphi_iterator psi;
684aaf29 3627
0fc822d0 3628 if (number_of_loops (cfun) <= 1)
c2924966 3629 return 0;
684aaf29 3630
11cd3bed 3631 FOR_EACH_BB_FN (bb, cfun)
684aaf29
ZD
3632 {
3633 loop = bb->loop_father;
3634
726a989a 3635 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
684aaf29 3636 {
538dd0b7 3637 phi = psi.phi ();
684aaf29
ZD
3638 name = PHI_RESULT (phi);
3639
ea057359 3640 if (virtual_operand_p (name))
684aaf29
ZD
3641 continue;
3642
3643 type = TREE_TYPE (name);
3644
3645 if (!POINTER_TYPE_P (type)
3646 && !INTEGRAL_TYPE_P (type))
3647 continue;
3648
c70ed622
BC
3649 ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name),
3650 NULL);
684aaf29
ZD
3651 if (!is_gimple_min_invariant (ev)
3652 || !may_propagate_copy (name, ev))
3653 continue;
3654
3655 /* Replace the uses of the name. */
18aed06a 3656 if (name != ev)
ed22b76f
TV
3657 {
3658 if (dump_file && (dump_flags & TDF_DETAILS))
3659 {
3660 fprintf (dump_file, "Replacing uses of: ");
3661 print_generic_expr (dump_file, name, 0);
3662 fprintf (dump_file, " with: ");
3663 print_generic_expr (dump_file, ev, 0);
3664 fprintf (dump_file, "\n");
3665 }
3666 replace_uses_by (name, ev);
3667 }
684aaf29
ZD
3668
3669 if (!ssa_names_to_remove)
3670 ssa_names_to_remove = BITMAP_ALLOC (NULL);
3671 bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name));
3672 }
3673 }
3674
9b3b55a1
DN
3675 /* Remove the ssa names that were replaced by constants. We do not
3676 remove them directly in the previous cycle, since this
3677 invalidates scev cache. */
684aaf29
ZD
3678 if (ssa_names_to_remove)
3679 {
3680 bitmap_iterator bi;
684aaf29
ZD
3681
3682 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi)
3683 {
726a989a 3684 gimple_stmt_iterator psi;
684aaf29 3685 name = ssa_name (i);
538dd0b7 3686 phi = as_a <gphi *> (SSA_NAME_DEF_STMT (name));
684aaf29 3687
726a989a
RB
3688 gcc_assert (gimple_code (phi) == GIMPLE_PHI);
3689 psi = gsi_for_stmt (phi);
3690 remove_phi_node (&psi, true);
684aaf29
ZD
3691 }
3692
3693 BITMAP_FREE (ssa_names_to_remove);
3694 scev_reset ();
3695 }
3ac01fde
ZD
3696
3697 /* Now the regular final value replacement. */
f0bd40b1 3698 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
f993a853 3699 final_value_replacement_loop (loop);
925196ed 3700
c2924966 3701 return 0;
684aaf29 3702}
9e2f83a5
ZD
3703
3704#include "gt-tree-scalar-evolution.h"