]> git.ipfire.org Git - thirdparty/gcc.git/blame - gcc/tree-scalar-evolution.c
* tree-ssa-loop-niter.c (nowrap_type_p): Use ANY_INTEGRAL_TYPE_P.
[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;
e6d62b46
BC
1939
1940 case BIT_AND_EXPR:
1941 /* Given int variable A, handle A&0xffff as (int)(unsigned short)A.
1942 If A is SCEV and its value is in the range of representable set
1943 of type unsigned short, the result expression is a (no-overflow)
1944 SCEV. */
1945 res = chrec_dont_know;
1946 if (tree_fits_uhwi_p (rhs2))
1947 {
1948 int precision;
1949 unsigned HOST_WIDE_INT val = tree_to_uhwi (rhs2);
1950
1951 val ++;
1952 /* Skip if value of rhs2 wraps in unsigned HOST_WIDE_INT or
1953 it's not the maximum value of a smaller type than rhs1. */
1954 if (val != 0
1955 && (precision = exact_log2 (val)) > 0
1956 && (unsigned) precision < TYPE_PRECISION (TREE_TYPE (rhs1)))
1957 {
1958 tree utype = build_nonstandard_integer_type (precision, 1);
1959
1960 if (TYPE_PRECISION (utype) < TYPE_PRECISION (TREE_TYPE (rhs1)))
1961 {
1962 chrec1 = analyze_scalar_evolution (loop, rhs1);
1963 chrec1 = chrec_convert (utype, chrec1, at_stmt);
1964 res = chrec_convert (TREE_TYPE (rhs1), chrec1, at_stmt);
1965 }
1966 }
1967 }
1968 break;
b8698a0f 1969
9baba81b
SP
1970 default:
1971 res = chrec_dont_know;
1972 break;
1973 }
b8698a0f 1974
9baba81b
SP
1975 return res;
1976}
1977
726a989a
RB
1978/* Interpret the expression EXPR. */
1979
1980static tree
355fe088 1981interpret_expr (struct loop *loop, gimple *at_stmt, tree expr)
726a989a
RB
1982{
1983 enum tree_code code;
1984 tree type = TREE_TYPE (expr), op0, op1;
1985
1986 if (automatically_generated_chrec_p (expr))
1987 return expr;
1988
4e71066d
RG
1989 if (TREE_CODE (expr) == POLYNOMIAL_CHREC
1990 || get_gimple_rhs_class (TREE_CODE (expr)) == GIMPLE_TERNARY_RHS)
726a989a
RB
1991 return chrec_dont_know;
1992
1993 extract_ops_from_tree (expr, &code, &op0, &op1);
1994
1995 return interpret_rhs_expr (loop, at_stmt, type,
1996 op0, code, op1);
1997}
1998
1999/* Interpret the rhs of the assignment STMT. */
2000
2001static tree
355fe088 2002interpret_gimple_assign (struct loop *loop, gimple *stmt)
726a989a
RB
2003{
2004 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
2005 enum tree_code code = gimple_assign_rhs_code (stmt);
2006
2007 return interpret_rhs_expr (loop, stmt, type,
2008 gimple_assign_rhs1 (stmt), code,
2009 gimple_assign_rhs2 (stmt));
2010}
2011
9baba81b
SP
2012\f
2013
b8698a0f 2014/* This section contains all the entry points:
9baba81b
SP
2015 - number_of_iterations_in_loop,
2016 - analyze_scalar_evolution,
2017 - instantiate_parameters.
2018*/
2019
2020/* Compute and return the evolution function in WRTO_LOOP, the nearest
2021 common ancestor of DEF_LOOP and USE_LOOP. */
2022
b8698a0f
L
2023static tree
2024compute_scalar_evolution_in_loop (struct loop *wrto_loop,
2025 struct loop *def_loop,
9baba81b
SP
2026 tree ev)
2027{
492e5456 2028 bool val;
9baba81b 2029 tree res;
492e5456 2030
9baba81b
SP
2031 if (def_loop == wrto_loop)
2032 return ev;
2033
9ba025a2 2034 def_loop = superloop_at_depth (def_loop, loop_depth (wrto_loop) + 1);
9baba81b
SP
2035 res = compute_overall_effect_of_inner_loop (def_loop, ev);
2036
492e5456
SP
2037 if (no_evolution_in_loop_p (res, wrto_loop->num, &val) && val)
2038 return res;
2039
9baba81b
SP
2040 return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet);
2041}
2042
2043/* Helper recursive function. */
2044
2045static tree
2046analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res)
2047{
726a989a 2048 tree type = TREE_TYPE (var);
355fe088 2049 gimple *def;
9baba81b
SP
2050 basic_block bb;
2051 struct loop *def_loop;
2052
42d375ed 2053 if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE)
9baba81b
SP
2054 return chrec_dont_know;
2055
2056 if (TREE_CODE (var) != SSA_NAME)
726a989a 2057 return interpret_expr (loop, NULL, var);
9baba81b
SP
2058
2059 def = SSA_NAME_DEF_STMT (var);
726a989a 2060 bb = gimple_bb (def);
9baba81b
SP
2061 def_loop = bb ? bb->loop_father : NULL;
2062
2063 if (bb == NULL
2064 || !flow_bb_inside_loop_p (loop, bb))
2065 {
806f2c1b
AL
2066 /* Keep symbolic form, but look through obvious copies for constants. */
2067 res = follow_copies_to_constant (var);
9baba81b
SP
2068 goto set_and_end;
2069 }
2070
2071 if (res != chrec_not_analyzed_yet)
2072 {
2073 if (loop != bb->loop_father)
b8698a0f 2074 res = compute_scalar_evolution_in_loop
9baba81b
SP
2075 (find_common_loop (loop, bb->loop_father), bb->loop_father, res);
2076
2077 goto set_and_end;
2078 }
2079
2080 if (loop != def_loop)
2081 {
2082 res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet);
2083 res = compute_scalar_evolution_in_loop (loop, def_loop, res);
2084
2085 goto set_and_end;
2086 }
2087
726a989a 2088 switch (gimple_code (def))
9baba81b 2089 {
726a989a
RB
2090 case GIMPLE_ASSIGN:
2091 res = interpret_gimple_assign (loop, def);
9baba81b
SP
2092 break;
2093
726a989a 2094 case GIMPLE_PHI:
9baba81b 2095 if (loop_phi_node_p (def))
538dd0b7 2096 res = interpret_loop_phi (loop, as_a <gphi *> (def));
9baba81b 2097 else
538dd0b7 2098 res = interpret_condition_phi (loop, as_a <gphi *> (def));
9baba81b
SP
2099 break;
2100
2101 default:
2102 res = chrec_dont_know;
2103 break;
2104 }
2105
2106 set_and_end:
2107
2108 /* Keep the symbolic form. */
2109 if (res == chrec_dont_know)
2110 res = var;
2111
2112 if (loop == def_loop)
a213b219 2113 set_scalar_evolution (block_before_loop (loop), var, res);
9baba81b
SP
2114
2115 return res;
2116}
2117
52bdd655
SP
2118/* Analyzes and returns the scalar evolution of the ssa_name VAR in
2119 LOOP. LOOP is the loop in which the variable is used.
b8698a0f 2120
9baba81b
SP
2121 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
2122 pointer to the statement that uses this variable, in order to
2123 determine the evolution function of the variable, use the following
2124 calls:
b8698a0f 2125
52bdd655
SP
2126 loop_p loop = loop_containing_stmt (stmt);
2127 tree chrec_with_symbols = analyze_scalar_evolution (loop, var);
3f227a8c 2128 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
9baba81b
SP
2129*/
2130
b8698a0f 2131tree
9baba81b
SP
2132analyze_scalar_evolution (struct loop *loop, tree var)
2133{
2134 tree res;
2135
dfedbe40 2136 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
2137 {
2138 fprintf (dump_file, "(analyze_scalar_evolution \n");
2139 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
2140 fprintf (dump_file, " (scalar = ");
2141 print_generic_expr (dump_file, var, 0);
2142 fprintf (dump_file, ")\n");
2143 }
2144
a213b219
SP
2145 res = get_scalar_evolution (block_before_loop (loop), var);
2146 res = analyze_scalar_evolution_1 (loop, var, res);
9baba81b 2147
dfedbe40 2148 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
2149 fprintf (dump_file, ")\n");
2150
2151 return res;
2152}
2153
bef28ced
JL
2154/* Analyzes and returns the scalar evolution of VAR address in LOOP. */
2155
2156static tree
2157analyze_scalar_evolution_for_address_of (struct loop *loop, tree var)
2158{
2159 return analyze_scalar_evolution (loop, build_fold_addr_expr (var));
2160}
2161
9baba81b 2162/* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
f017bf5e 2163 WRTO_LOOP (which should be a superloop of USE_LOOP)
a6f778b2
ZD
2164
2165 FOLDED_CASTS is set to true if resolve_mixers used
2166 chrec_convert_aggressive (TODO -- not really, we are way too conservative
b8698a0f
L
2167 at the moment in order to keep things simple).
2168
f017bf5e
ZD
2169 To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following
2170 example:
2171
2172 for (i = 0; i < 100; i++) -- loop 1
2173 {
2174 for (j = 0; j < 100; j++) -- loop 2
2175 {
2176 k1 = i;
2177 k2 = j;
2178
2179 use2 (k1, k2);
2180
2181 for (t = 0; t < 100; t++) -- loop 3
2182 use3 (k1, k2);
2183
2184 }
2185 use1 (k1, k2);
2186 }
2187
2188 Both k1 and k2 are invariants in loop3, thus
2189 analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1
2190 analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2
2191
2192 As they are invariant, it does not matter whether we consider their
2193 usage in loop 3 or loop 2, hence
2194 analyze_scalar_evolution_in_loop (loop2, loop3, k1) =
2195 analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i
2196 analyze_scalar_evolution_in_loop (loop2, loop3, k2) =
2197 analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2
2198
2199 Similarly for their evolutions with respect to loop 1. The values of K2
2200 in the use in loop 2 vary independently on loop 1, thus we cannot express
2201 the evolution with respect to loop 1:
2202 analyze_scalar_evolution_in_loop (loop1, loop3, k1) =
2203 analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1
2204 analyze_scalar_evolution_in_loop (loop1, loop3, k2) =
2205 analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know
2206
2207 The value of k2 in the use in loop 1 is known, though:
2208 analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1
2209 analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100
2210 */
9baba81b
SP
2211
2212static tree
2213analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop,
a6f778b2 2214 tree version, bool *folded_casts)
9baba81b
SP
2215{
2216 bool val = false;
a6f778b2 2217 tree ev = version, tmp;
9baba81b 2218
b8698a0f 2219 /* We cannot just do
f017bf5e
ZD
2220
2221 tmp = analyze_scalar_evolution (use_loop, version);
c70ed622 2222 ev = resolve_mixers (wrto_loop, tmp, folded_casts);
f017bf5e
ZD
2223
2224 as resolve_mixers would query the scalar evolution with respect to
2225 wrto_loop. For example, in the situation described in the function
2226 comment, suppose that wrto_loop = loop1, use_loop = loop3 and
2227 version = k2. Then
2228
2229 analyze_scalar_evolution (use_loop, version) = k2
2230
c70ed622
BC
2231 and resolve_mixers (loop1, k2, folded_casts) finds that the value of
2232 k2 in loop 1 is 100, which is a wrong result, since we are interested
2233 in the value in loop 3.
f017bf5e
ZD
2234
2235 Instead, we need to proceed from use_loop to wrto_loop loop by loop,
2236 each time checking that there is no evolution in the inner loop. */
2237
a6f778b2
ZD
2238 if (folded_casts)
2239 *folded_casts = false;
9baba81b
SP
2240 while (1)
2241 {
a6f778b2 2242 tmp = analyze_scalar_evolution (use_loop, ev);
c70ed622 2243 ev = resolve_mixers (use_loop, tmp, folded_casts);
9baba81b
SP
2244
2245 if (use_loop == wrto_loop)
2246 return ev;
2247
2248 /* If the value of the use changes in the inner loop, we cannot express
2249 its value in the outer loop (we might try to return interval chrec,
2250 but we do not have a user for it anyway) */
2251 if (!no_evolution_in_loop_p (ev, use_loop->num, &val)
2252 || !val)
2253 return chrec_dont_know;
2254
9ba025a2 2255 use_loop = loop_outer (use_loop);
9baba81b
SP
2256 }
2257}
2258
eb0bc7af 2259
fdd43ac4
RB
2260/* Hashtable helpers for a temporary hash-table used when
2261 instantiating a CHREC or resolving mixers. For this use
2262 instantiated_below is always the same. */
2263
fdd43ac4 2264struct instantiate_cache_type
eb0bc7af 2265{
a3cc13cc
RB
2266 htab_t map;
2267 vec<scev_info_str> entries;
b8698a0f 2268
c3284718 2269 instantiate_cache_type () : map (NULL), entries (vNULL) {}
fdd43ac4 2270 ~instantiate_cache_type ();
0547c9b6
RB
2271 tree get (unsigned slot) { return entries[slot].chrec; }
2272 void set (unsigned slot, tree chrec) { entries[slot].chrec = chrec; }
fdd43ac4 2273};
eb0bc7af 2274
fdd43ac4
RB
2275instantiate_cache_type::~instantiate_cache_type ()
2276{
0547c9b6 2277 if (map != NULL)
fdd43ac4 2278 {
a3cc13cc 2279 htab_delete (map);
fdd43ac4
RB
2280 entries.release ();
2281 }
eb0bc7af
ZD
2282}
2283
a3cc13cc
RB
2284/* Cache to avoid infinite recursion when instantiating an SSA name.
2285 Live during the outermost instantiate_scev or resolve_mixers call. */
2286static instantiate_cache_type *global_cache;
2287
2288/* Computes a hash function for database element ELT. */
2289
2290static inline hashval_t
2291hash_idx_scev_info (const void *elt_)
2292{
2293 unsigned idx = ((size_t) elt_) - 2;
907dadbd 2294 return scev_info_hasher::hash (&global_cache->entries[idx]);
a3cc13cc
RB
2295}
2296
2297/* Compares database elements E1 and E2. */
2298
2299static inline int
2300eq_idx_scev_info (const void *e1, const void *e2)
2301{
2302 unsigned idx1 = ((size_t) e1) - 2;
907dadbd
TS
2303 return scev_info_hasher::equal (&global_cache->entries[idx1],
2304 (const scev_info_str *) e2);
a3cc13cc
RB
2305}
2306
0547c9b6 2307/* Returns from CACHE the slot number of the cached chrec for NAME. */
fdd43ac4 2308
0547c9b6 2309static unsigned
a3cc13cc
RB
2310get_instantiated_value_entry (instantiate_cache_type &cache,
2311 tree name, basic_block instantiate_below)
fdd43ac4 2312{
0547c9b6 2313 if (!cache.map)
fdd43ac4 2314 {
a3cc13cc 2315 cache.map = htab_create (10, hash_idx_scev_info, eq_idx_scev_info, NULL);
fdd43ac4
RB
2316 cache.entries.create (10);
2317 }
b8698a0f 2318
a3cc13cc
RB
2319 scev_info_str e;
2320 e.name_version = SSA_NAME_VERSION (name);
2321 e.instantiated_below = instantiate_below->index;
2322 void **slot = htab_find_slot_with_hash (cache.map, &e,
907dadbd 2323 scev_info_hasher::hash (&e), INSERT);
a3cc13cc 2324 if (!*slot)
fdd43ac4
RB
2325 {
2326 e.chrec = chrec_not_analyzed_yet;
a3cc13cc 2327 *slot = (void *)(size_t)(cache.entries.length () + 2);
fdd43ac4 2328 cache.entries.safe_push (e);
fdd43ac4
RB
2329 }
2330
a3cc13cc 2331 return ((size_t)*slot) - 2;
eb0bc7af
ZD
2332}
2333
0547c9b6 2334
18aed06a
SP
2335/* Return the closed_loop_phi node for VAR. If there is none, return
2336 NULL_TREE. */
2337
2338static tree
2339loop_closed_phi_def (tree var)
2340{
2341 struct loop *loop;
2342 edge exit;
538dd0b7
DM
2343 gphi *phi;
2344 gphi_iterator psi;
18aed06a
SP
2345
2346 if (var == NULL_TREE
2347 || TREE_CODE (var) != SSA_NAME)
2348 return NULL_TREE;
2349
2350 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var));
ac8f6c69 2351 exit = single_exit (loop);
18aed06a
SP
2352 if (!exit)
2353 return NULL_TREE;
2354
726a989a
RB
2355 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi))
2356 {
538dd0b7 2357 phi = psi.phi ();
726a989a
RB
2358 if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var)
2359 return PHI_RESULT (phi);
2360 }
18aed06a
SP
2361
2362 return NULL_TREE;
2363}
2364
8b679c9b 2365static tree instantiate_scev_r (basic_block, struct loop *, struct loop *,
c70ed622 2366 tree, bool *, int);
320f5a78
SP
2367
2368/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2369 and EVOLUTION_LOOP, that were left under a symbolic form.
2370
2495a183 2371 CHREC is an SSA_NAME to be instantiated.
320f5a78
SP
2372
2373 CACHE is the cache of already instantiated values.
2374
c70ed622
BC
2375 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2376 conversions that may wrap in signed/pointer type are folded, as long
2377 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2378 then we don't do such fold.
320f5a78
SP
2379
2380 SIZE_EXPR is used for computing the size of the expression to be
2381 instantiated, and to stop if it exceeds some limit. */
2382
2383static tree
2495a183 2384instantiate_scev_name (basic_block instantiate_below,
8b679c9b
RB
2385 struct loop *evolution_loop, struct loop *inner_loop,
2386 tree chrec,
c70ed622 2387 bool *fold_conversions,
4a8fb1a1 2388 int size_expr)
320f5a78 2389{
2495a183
SP
2390 tree res;
2391 struct loop *def_loop;
2392 basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (chrec));
20179b0d 2393
2495a183
SP
2394 /* A parameter (or loop invariant and we do not want to include
2395 evolutions in outer loops), nothing to do. */
2396 if (!def_bb
2397 || loop_depth (def_bb->loop_father) == 0
2398 || dominated_by_p (CDI_DOMINATORS, instantiate_below, def_bb))
2399 return chrec;
20179b0d 2400
2495a183
SP
2401 /* We cache the value of instantiated variable to avoid exponential
2402 time complexity due to reevaluations. We also store the convenient
2403 value in the cache in order to prevent infinite recursion -- we do
2404 not want to instantiate the SSA_NAME if it is in a mixer
2405 structure. This is used for avoiding the instantiation of
2406 recursively defined functions, such as:
320f5a78 2407
2495a183 2408 | a_2 -> {0, +, 1, +, a_2}_1 */
20179b0d 2409
a3cc13cc
RB
2410 unsigned si = get_instantiated_value_entry (*global_cache,
2411 chrec, instantiate_below);
0547c9b6
RB
2412 if (global_cache->get (si) != chrec_not_analyzed_yet)
2413 return global_cache->get (si);
20179b0d 2414
fdd43ac4 2415 /* On recursion return chrec_dont_know. */
0547c9b6 2416 global_cache->set (si, chrec_dont_know);
320f5a78 2417
2495a183
SP
2418 def_loop = find_common_loop (evolution_loop, def_bb->loop_father);
2419
320f5a78
SP
2420 /* If the analysis yields a parametric chrec, instantiate the
2421 result again. */
2422 res = analyze_scalar_evolution (def_loop, chrec);
2423
2847388e 2424 /* Don't instantiate default definitions. */
320f5a78 2425 if (TREE_CODE (res) == SSA_NAME
2847388e
SP
2426 && SSA_NAME_IS_DEFAULT_DEF (res))
2427 ;
2428
2429 /* Don't instantiate loop-closed-ssa phi nodes. */
2430 else if (TREE_CODE (res) == SSA_NAME
2431 && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res)))
2432 > loop_depth (def_loop))
320f5a78
SP
2433 {
2434 if (res == chrec)
2435 res = loop_closed_phi_def (chrec);
2436 else
2437 res = chrec;
2438
7472eb13
SP
2439 /* When there is no loop_closed_phi_def, it means that the
2440 variable is not used after the loop: try to still compute the
2441 value of the variable when exiting the loop. */
2442 if (res == NULL_TREE)
2443 {
2444 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (chrec));
2445 res = analyze_scalar_evolution (loop, chrec);
2446 res = compute_overall_effect_of_inner_loop (loop, res);
8b679c9b
RB
2447 res = instantiate_scev_r (instantiate_below, evolution_loop,
2448 inner_loop, res,
0547c9b6 2449 fold_conversions, size_expr);
7472eb13
SP
2450 }
2451 else if (!dominated_by_p (CDI_DOMINATORS, instantiate_below,
2452 gimple_bb (SSA_NAME_DEF_STMT (res))))
320f5a78
SP
2453 res = chrec_dont_know;
2454 }
2455
2456 else if (res != chrec_dont_know)
8b679c9b
RB
2457 {
2458 if (inner_loop
63fdb7be 2459 && def_bb->loop_father != inner_loop
8b679c9b
RB
2460 && !flow_loop_nested_p (def_bb->loop_father, inner_loop))
2461 /* ??? We could try to compute the overall effect of the loop here. */
2462 res = chrec_dont_know;
2463 else
2464 res = instantiate_scev_r (instantiate_below, evolution_loop,
2465 inner_loop, res,
0547c9b6 2466 fold_conversions, size_expr);
8b679c9b 2467 }
320f5a78
SP
2468
2469 /* Store the correct value to the cache. */
0547c9b6 2470 global_cache->set (si, res);
320f5a78 2471 return res;
320f5a78
SP
2472}
2473
ec6636eb
SP
2474/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2475 and EVOLUTION_LOOP, that were left under a symbolic form.
2476
2477 CHREC is a polynomial chain of recurrence to be instantiated.
2478
2479 CACHE is the cache of already instantiated values.
2480
c70ed622
BC
2481 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2482 conversions that may wrap in signed/pointer type are folded, as long
2483 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2484 then we don't do such fold.
ec6636eb
SP
2485
2486 SIZE_EXPR is used for computing the size of the expression to be
2487 instantiated, and to stop if it exceeds some limit. */
2488
2489static tree
2490instantiate_scev_poly (basic_block instantiate_below,
8b679c9b 2491 struct loop *evolution_loop, struct loop *,
c70ed622 2492 tree chrec, bool *fold_conversions, int size_expr)
ec6636eb
SP
2493{
2494 tree op1;
9e5dc77f 2495 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2496 get_chrec_loop (chrec),
0547c9b6 2497 CHREC_LEFT (chrec), fold_conversions,
ec6636eb
SP
2498 size_expr);
2499 if (op0 == chrec_dont_know)
2500 return chrec_dont_know;
2501
9e5dc77f 2502 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2503 get_chrec_loop (chrec),
0547c9b6 2504 CHREC_RIGHT (chrec), fold_conversions,
ec6636eb
SP
2505 size_expr);
2506 if (op1 == chrec_dont_know)
2507 return chrec_dont_know;
2508
2509 if (CHREC_LEFT (chrec) != op0
2510 || CHREC_RIGHT (chrec) != op1)
2511 {
2512 op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL);
8b679c9b 2513 chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1);
ec6636eb 2514 }
4bf4e169 2515
ec6636eb
SP
2516 return chrec;
2517}
2518
15fda317
SP
2519/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2520 and EVOLUTION_LOOP, that were left under a symbolic form.
2521
ffa34f4b 2522 "C0 CODE C1" is a binary expression of type TYPE to be instantiated.
15fda317
SP
2523
2524 CACHE is the cache of already instantiated values.
2525
c70ed622
BC
2526 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2527 conversions that may wrap in signed/pointer type are folded, as long
2528 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2529 then we don't do such fold.
15fda317
SP
2530
2531 SIZE_EXPR is used for computing the size of the expression to be
2532 instantiated, and to stop if it exceeds some limit. */
2533
2534static tree
2535instantiate_scev_binary (basic_block instantiate_below,
8b679c9b
RB
2536 struct loop *evolution_loop, struct loop *inner_loop,
2537 tree chrec, enum tree_code code,
ffa34f4b 2538 tree type, tree c0, tree c1,
c70ed622 2539 bool *fold_conversions, int size_expr)
15fda317
SP
2540{
2541 tree op1;
8b679c9b 2542 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop,
0547c9b6 2543 c0, fold_conversions, size_expr);
15fda317
SP
2544 if (op0 == chrec_dont_know)
2545 return chrec_dont_know;
2546
8b679c9b 2547 op1 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop,
0547c9b6 2548 c1, fold_conversions, size_expr);
15fda317
SP
2549 if (op1 == chrec_dont_know)
2550 return chrec_dont_know;
2551
ffa34f4b
SP
2552 if (c0 != op0
2553 || c1 != op1)
15fda317 2554 {
15fda317
SP
2555 op0 = chrec_convert (type, op0, NULL);
2556 op1 = chrec_convert_rhs (type, op1, NULL);
2557
ffa34f4b 2558 switch (code)
15fda317
SP
2559 {
2560 case POINTER_PLUS_EXPR:
2561 case PLUS_EXPR:
2562 return chrec_fold_plus (type, op0, op1);
2563
2564 case MINUS_EXPR:
2565 return chrec_fold_minus (type, op0, op1);
2566
2567 case MULT_EXPR:
2568 return chrec_fold_multiply (type, op0, op1);
2569
2570 default:
2571 gcc_unreachable ();
2572 }
2573 }
2574
ffa34f4b 2575 return chrec ? chrec : fold_build2 (code, type, c0, c1);
15fda317
SP
2576}
2577
dbc08079
SP
2578/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2579 and EVOLUTION_LOOP, that were left under a symbolic form.
2580
2581 "CHREC" is an array reference to be instantiated.
2582
2583 CACHE is the cache of already instantiated values.
2584
c70ed622
BC
2585 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2586 conversions that may wrap in signed/pointer type are folded, as long
2587 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2588 then we don't do such fold.
dbc08079
SP
2589
2590 SIZE_EXPR is used for computing the size of the expression to be
2591 instantiated, and to stop if it exceeds some limit. */
2592
2593static tree
2594instantiate_array_ref (basic_block instantiate_below,
8b679c9b 2595 struct loop *evolution_loop, struct loop *inner_loop,
c70ed622 2596 tree chrec, bool *fold_conversions, int size_expr)
dbc08079
SP
2597{
2598 tree res;
2599 tree index = TREE_OPERAND (chrec, 1);
8b679c9b
RB
2600 tree op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2601 inner_loop, index,
0547c9b6 2602 fold_conversions, size_expr);
dbc08079
SP
2603
2604 if (op1 == chrec_dont_know)
2605 return chrec_dont_know;
2606
2607 if (chrec && op1 == index)
2608 return chrec;
2609
2610 res = unshare_expr (chrec);
2611 TREE_OPERAND (res, 1) = op1;
2612 return res;
2613}
2614
a213b219 2615/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
9c382ce9
SP
2616 and EVOLUTION_LOOP, that were left under a symbolic form.
2617
2618 "CHREC" that stands for a convert expression "(TYPE) OP" is to be
2619 instantiated.
2620
2621 CACHE is the cache of already instantiated values.
2622
c70ed622
BC
2623 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2624 conversions that may wrap in signed/pointer type are folded, as long
2625 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2626 then we don't do such fold.
9c382ce9
SP
2627
2628 SIZE_EXPR is used for computing the size of the expression to be
2629 instantiated, and to stop if it exceeds some limit. */
2630
2631static tree
2632instantiate_scev_convert (basic_block instantiate_below,
8b679c9b 2633 struct loop *evolution_loop, struct loop *inner_loop,
0547c9b6 2634 tree chrec, tree type, tree op,
c70ed622 2635 bool *fold_conversions, int size_expr)
9c382ce9 2636{
8b679c9b
RB
2637 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2638 inner_loop, op,
0547c9b6 2639 fold_conversions, size_expr);
9c382ce9
SP
2640
2641 if (op0 == chrec_dont_know)
2642 return chrec_dont_know;
2643
2644 if (fold_conversions)
2645 {
c70ed622 2646 tree tmp = chrec_convert_aggressive (type, op0, fold_conversions);
9c382ce9
SP
2647 if (tmp)
2648 return tmp;
9c382ce9 2649
c70ed622
BC
2650 /* If we used chrec_convert_aggressive, we can no longer assume that
2651 signed chrecs do not overflow, as chrec_convert does, so avoid
2652 calling it in that case. */
2653 if (*fold_conversions)
2654 {
2655 if (chrec && op0 == op)
2656 return chrec;
9c382ce9 2657
c70ed622
BC
2658 return fold_convert (type, op0);
2659 }
2660 }
9c382ce9
SP
2661
2662 return chrec_convert (type, op0, NULL);
2663}
2664
7ec0665d
SP
2665/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2666 and EVOLUTION_LOOP, that were left under a symbolic form.
2667
4b9d48a1 2668 CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated.
7ec0665d 2669 Handle ~X as -1 - X.
4b9d48a1 2670 Handle -X as -1 * X.
7ec0665d
SP
2671
2672 CACHE is the cache of already instantiated values.
2673
c70ed622
BC
2674 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2675 conversions that may wrap in signed/pointer type are folded, as long
2676 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2677 then we don't do such fold.
7ec0665d
SP
2678
2679 SIZE_EXPR is used for computing the size of the expression to be
2680 instantiated, and to stop if it exceeds some limit. */
2681
2682static tree
4b9d48a1 2683instantiate_scev_not (basic_block instantiate_below,
8b679c9b
RB
2684 struct loop *evolution_loop, struct loop *inner_loop,
2685 tree chrec,
20179b0d 2686 enum tree_code code, tree type, tree op,
c70ed622 2687 bool *fold_conversions, int size_expr)
7ec0665d 2688{
8b679c9b
RB
2689 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2690 inner_loop, op,
0547c9b6 2691 fold_conversions, size_expr);
20179b0d 2692
7ec0665d
SP
2693 if (op0 == chrec_dont_know)
2694 return chrec_dont_know;
2695
20179b0d 2696 if (op != op0)
7ec0665d
SP
2697 {
2698 op0 = chrec_convert (type, op0, NULL);
4b9d48a1 2699
20179b0d 2700 switch (code)
4b9d48a1
SP
2701 {
2702 case BIT_NOT_EXPR:
2703 return chrec_fold_minus
2704 (type, fold_convert (type, integer_minus_one_node), op0);
2705
2706 case NEGATE_EXPR:
2707 return chrec_fold_multiply
2708 (type, fold_convert (type, integer_minus_one_node), op0);
2709
2710 default:
2711 gcc_unreachable ();
2712 }
7ec0665d 2713 }
4b9d48a1 2714
20179b0d 2715 return chrec ? chrec : fold_build1 (code, type, op0);
7ec0665d
SP
2716}
2717
d814176c
SP
2718/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2719 and EVOLUTION_LOOP, that were left under a symbolic form.
2720
2721 CHREC is an expression with 3 operands to be instantiated.
2722
2723 CACHE is the cache of already instantiated values.
2724
c70ed622
BC
2725 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2726 conversions that may wrap in signed/pointer type are folded, as long
2727 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2728 then we don't do such fold.
d814176c
SP
2729
2730 SIZE_EXPR is used for computing the size of the expression to be
2731 instantiated, and to stop if it exceeds some limit. */
2732
2733static tree
2734instantiate_scev_3 (basic_block instantiate_below,
8b679c9b
RB
2735 struct loop *evolution_loop, struct loop *inner_loop,
2736 tree chrec,
c70ed622 2737 bool *fold_conversions, int size_expr)
d814176c
SP
2738{
2739 tree op1, op2;
9e5dc77f 2740 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2741 inner_loop, TREE_OPERAND (chrec, 0),
0547c9b6 2742 fold_conversions, size_expr);
d814176c
SP
2743 if (op0 == chrec_dont_know)
2744 return chrec_dont_know;
2745
9e5dc77f 2746 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2747 inner_loop, TREE_OPERAND (chrec, 1),
0547c9b6 2748 fold_conversions, size_expr);
d814176c
SP
2749 if (op1 == chrec_dont_know)
2750 return chrec_dont_know;
2751
9e5dc77f 2752 op2 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2753 inner_loop, TREE_OPERAND (chrec, 2),
0547c9b6 2754 fold_conversions, size_expr);
d814176c
SP
2755 if (op2 == chrec_dont_know)
2756 return chrec_dont_know;
2757
2758 if (op0 == TREE_OPERAND (chrec, 0)
2759 && op1 == TREE_OPERAND (chrec, 1)
2760 && op2 == TREE_OPERAND (chrec, 2))
2761 return chrec;
2762
2763 return fold_build3 (TREE_CODE (chrec),
2764 TREE_TYPE (chrec), op0, op1, op2);
2765}
2766
9c382ce9
SP
2767/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2768 and EVOLUTION_LOOP, that were left under a symbolic form.
5b78fc3e 2769
9e5dc77f
SP
2770 CHREC is an expression with 2 operands to be instantiated.
2771
2772 CACHE is the cache of already instantiated values.
2773
c70ed622
BC
2774 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2775 conversions that may wrap in signed/pointer type are folded, as long
2776 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2777 then we don't do such fold.
9e5dc77f
SP
2778
2779 SIZE_EXPR is used for computing the size of the expression to be
2780 instantiated, and to stop if it exceeds some limit. */
2781
2782static tree
2783instantiate_scev_2 (basic_block instantiate_below,
8b679c9b
RB
2784 struct loop *evolution_loop, struct loop *inner_loop,
2785 tree chrec,
c70ed622 2786 bool *fold_conversions, int size_expr)
9e5dc77f
SP
2787{
2788 tree op1;
2789 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2790 inner_loop, TREE_OPERAND (chrec, 0),
0547c9b6 2791 fold_conversions, size_expr);
9e5dc77f
SP
2792 if (op0 == chrec_dont_know)
2793 return chrec_dont_know;
2794
2795 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2796 inner_loop, TREE_OPERAND (chrec, 1),
0547c9b6 2797 fold_conversions, size_expr);
9e5dc77f
SP
2798 if (op1 == chrec_dont_know)
2799 return chrec_dont_know;
2800
2801 if (op0 == TREE_OPERAND (chrec, 0)
2802 && op1 == TREE_OPERAND (chrec, 1))
2803 return chrec;
2804
2805 return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1);
2806}
2807
2808/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2809 and EVOLUTION_LOOP, that were left under a symbolic form.
2810
2811 CHREC is an expression with 2 operands to be instantiated.
5b78fc3e
JS
2812
2813 CACHE is the cache of already instantiated values.
2814
c70ed622
BC
2815 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2816 conversions that may wrap in signed/pointer type are folded, as long
2817 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2818 then we don't do such fold.
5b78fc3e 2819
3f227a8c
JS
2820 SIZE_EXPR is used for computing the size of the expression to be
2821 instantiated, and to stop if it exceeds some limit. */
9c382ce9 2822
9baba81b 2823static tree
a213b219 2824instantiate_scev_1 (basic_block instantiate_below,
8b679c9b
RB
2825 struct loop *evolution_loop, struct loop *inner_loop,
2826 tree chrec,
c70ed622 2827 bool *fold_conversions, int size_expr)
9baba81b 2828{
9e5dc77f 2829 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
8b679c9b 2830 inner_loop, TREE_OPERAND (chrec, 0),
0547c9b6 2831 fold_conversions, size_expr);
9e5dc77f
SP
2832
2833 if (op0 == chrec_dont_know)
2834 return chrec_dont_know;
2835
2836 if (op0 == TREE_OPERAND (chrec, 0))
2837 return chrec;
2838
2839 return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0);
2840}
2841
2842/* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2843 and EVOLUTION_LOOP, that were left under a symbolic form.
2844
2845 CHREC is the scalar evolution to instantiate.
2846
2847 CACHE is the cache of already instantiated values.
2282a0e6 2848
c70ed622
BC
2849 Variable pointed by FOLD_CONVERSIONS is set to TRUE when the
2850 conversions that may wrap in signed/pointer type are folded, as long
2851 as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL
2852 then we don't do such fold.
9e5dc77f
SP
2853
2854 SIZE_EXPR is used for computing the size of the expression to be
2855 instantiated, and to stop if it exceeds some limit. */
2856
2857static tree
2858instantiate_scev_r (basic_block instantiate_below,
8b679c9b
RB
2859 struct loop *evolution_loop, struct loop *inner_loop,
2860 tree chrec,
c70ed622 2861 bool *fold_conversions, int size_expr)
9e5dc77f 2862{
47ae9e4c
SP
2863 /* Give up if the expression is larger than the MAX that we allow. */
2864 if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
2865 return chrec_dont_know;
2866
81fada9a
JJ
2867 if (chrec == NULL_TREE
2868 || automatically_generated_chrec_p (chrec)
d7770457 2869 || is_gimple_min_invariant (chrec))
9baba81b
SP
2870 return chrec;
2871
2872 switch (TREE_CODE (chrec))
2873 {
2874 case SSA_NAME:
8b679c9b
RB
2875 return instantiate_scev_name (instantiate_below, evolution_loop,
2876 inner_loop, chrec,
0547c9b6 2877 fold_conversions, size_expr);
9baba81b
SP
2878
2879 case POLYNOMIAL_CHREC:
8b679c9b
RB
2880 return instantiate_scev_poly (instantiate_below, evolution_loop,
2881 inner_loop, chrec,
0547c9b6 2882 fold_conversions, size_expr);
9baba81b 2883
5be014d5 2884 case POINTER_PLUS_EXPR:
9baba81b 2885 case PLUS_EXPR:
9baba81b 2886 case MINUS_EXPR:
9baba81b 2887 case MULT_EXPR:
8b679c9b
RB
2888 return instantiate_scev_binary (instantiate_below, evolution_loop,
2889 inner_loop, chrec,
ffa34f4b
SP
2890 TREE_CODE (chrec), chrec_type (chrec),
2891 TREE_OPERAND (chrec, 0),
2892 TREE_OPERAND (chrec, 1),
0547c9b6 2893 fold_conversions, size_expr);
9baba81b 2894
1043771b 2895 CASE_CONVERT:
8b679c9b
RB
2896 return instantiate_scev_convert (instantiate_below, evolution_loop,
2897 inner_loop, chrec,
9c382ce9 2898 TREE_TYPE (chrec), TREE_OPERAND (chrec, 0),
0547c9b6 2899 fold_conversions, size_expr);
9baba81b 2900
4b9d48a1 2901 case NEGATE_EXPR:
418df9d7 2902 case BIT_NOT_EXPR:
8b679c9b
RB
2903 return instantiate_scev_not (instantiate_below, evolution_loop,
2904 inner_loop, chrec,
20179b0d
SP
2905 TREE_CODE (chrec), TREE_TYPE (chrec),
2906 TREE_OPERAND (chrec, 0),
0547c9b6 2907 fold_conversions, size_expr);
418df9d7 2908
4c7d6755 2909 case ADDR_EXPR:
9baba81b
SP
2910 case SCEV_NOT_KNOWN:
2911 return chrec_dont_know;
2912
2913 case SCEV_KNOWN:
2914 return chrec_known;
15fda317 2915
dbc08079 2916 case ARRAY_REF:
8b679c9b
RB
2917 return instantiate_array_ref (instantiate_below, evolution_loop,
2918 inner_loop, chrec,
0547c9b6 2919 fold_conversions, size_expr);
dbc08079 2920
9baba81b
SP
2921 default:
2922 break;
2923 }
2924
0dfb0dc6
SP
2925 if (VL_EXP_CLASS_P (chrec))
2926 return chrec_dont_know;
2927
9baba81b
SP
2928 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
2929 {
2930 case 3:
8b679c9b
RB
2931 return instantiate_scev_3 (instantiate_below, evolution_loop,
2932 inner_loop, chrec,
0547c9b6 2933 fold_conversions, size_expr);
9baba81b
SP
2934
2935 case 2:
8b679c9b
RB
2936 return instantiate_scev_2 (instantiate_below, evolution_loop,
2937 inner_loop, chrec,
0547c9b6 2938 fold_conversions, size_expr);
7ec0665d 2939
9baba81b 2940 case 1:
8b679c9b
RB
2941 return instantiate_scev_1 (instantiate_below, evolution_loop,
2942 inner_loop, chrec,
0547c9b6 2943 fold_conversions, size_expr);
9baba81b
SP
2944
2945 case 0:
2946 return chrec;
2947
2948 default:
2949 break;
2950 }
2951
2952 /* Too complicated to handle. */
2953 return chrec_dont_know;
2954}
e9eb809d
ZD
2955
2956/* Analyze all the parameters of the chrec that were left under a
a213b219
SP
2957 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2958 recursive instantiation of parameters: a parameter is a variable
2959 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2960 a function parameter. */
e9eb809d
ZD
2961
2962tree
a213b219 2963instantiate_scev (basic_block instantiate_below, struct loop *evolution_loop,
3f227a8c 2964 tree chrec)
e9eb809d 2965{
9baba81b
SP
2966 tree res;
2967
dfedbe40 2968 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b 2969 {
3f227a8c 2970 fprintf (dump_file, "(instantiate_scev \n");
a213b219 2971 fprintf (dump_file, " (instantiate_below = %d)\n", instantiate_below->index);
3f227a8c 2972 fprintf (dump_file, " (evolution_loop = %d)\n", evolution_loop->num);
9baba81b
SP
2973 fprintf (dump_file, " (chrec = ");
2974 print_generic_expr (dump_file, chrec, 0);
2975 fprintf (dump_file, ")\n");
2976 }
b8698a0f 2977
0547c9b6
RB
2978 bool destr = false;
2979 if (!global_cache)
2980 {
2981 global_cache = new instantiate_cache_type;
2982 destr = true;
2983 }
2984
8b679c9b 2985 res = instantiate_scev_r (instantiate_below, evolution_loop,
c70ed622 2986 NULL, chrec, NULL, 0);
0547c9b6
RB
2987
2988 if (destr)
2989 {
2990 delete global_cache;
2991 global_cache = NULL;
2992 }
9baba81b 2993
dfedbe40 2994 if (dump_file && (dump_flags & TDF_SCEV))
9baba81b
SP
2995 {
2996 fprintf (dump_file, " (res = ");
2997 print_generic_expr (dump_file, res, 0);
2998 fprintf (dump_file, "))\n");
2999 }
eb0bc7af 3000
9baba81b
SP
3001 return res;
3002}
3003
3004/* Similar to instantiate_parameters, but does not introduce the
2282a0e6
ZD
3005 evolutions in outer loops for LOOP invariants in CHREC, and does not
3006 care about causing overflows, as long as they do not affect value
3007 of an expression. */
9baba81b 3008
3cb960c7 3009tree
c70ed622 3010resolve_mixers (struct loop *loop, tree chrec, bool *folded_casts)
9baba81b 3011{
0547c9b6 3012 bool destr = false;
c70ed622 3013 bool fold_conversions = false;
0547c9b6
RB
3014 if (!global_cache)
3015 {
3016 global_cache = new instantiate_cache_type;
3017 destr = true;
3018 }
3019
8b679c9b 3020 tree ret = instantiate_scev_r (block_before_loop (loop), loop, NULL,
c70ed622
BC
3021 chrec, &fold_conversions, 0);
3022
3023 if (folded_casts && !*folded_casts)
3024 *folded_casts = fold_conversions;
0547c9b6
RB
3025
3026 if (destr)
3027 {
3028 delete global_cache;
3029 global_cache = NULL;
3030 }
3031
eb0bc7af 3032 return ret;
9baba81b
SP
3033}
3034
b8698a0f 3035/* Entry point for the analysis of the number of iterations pass.
9baba81b
SP
3036 This function tries to safely approximate the number of iterations
3037 the loop will run. When this property is not decidable at compile
0a74c758
SP
3038 time, the result is chrec_dont_know. Otherwise the result is a
3039 scalar or a symbolic parameter. When the number of iterations may
3040 be equal to zero and the property cannot be determined at compile
3041 time, the result is a COND_EXPR that represents in a symbolic form
3042 the conditions under which the number of iterations is not zero.
b8698a0f 3043
9baba81b 3044 Example of analysis: suppose that the loop has an exit condition:
b8698a0f 3045
9baba81b 3046 "if (b > 49) goto end_loop;"
b8698a0f 3047
9baba81b
SP
3048 and that in a previous analysis we have determined that the
3049 variable 'b' has an evolution function:
b8698a0f
L
3050
3051 "EF = {23, +, 5}_2".
3052
9baba81b
SP
3053 When we evaluate the function at the point 5, i.e. the value of the
3054 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
3055 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
3056 the loop body has been executed 6 times. */
3057
b8698a0f 3058tree
a14865db 3059number_of_latch_executions (struct loop *loop)
9baba81b 3060{
9baba81b
SP
3061 edge exit;
3062 struct tree_niter_desc niter_desc;
0a74c758
SP
3063 tree may_be_zero;
3064 tree res;
9baba81b 3065
0a74c758 3066 /* Determine whether the number of iterations in loop has already
9baba81b
SP
3067 been computed. */
3068 res = loop->nb_iterations;
3069 if (res)
3070 return res;
0a74c758
SP
3071
3072 may_be_zero = NULL_TREE;
9baba81b 3073
dfedbe40 3074 if (dump_file && (dump_flags & TDF_SCEV))
0a74c758 3075 fprintf (dump_file, "(number_of_iterations_in_loop = \n");
b8698a0f 3076
0a74c758 3077 res = chrec_dont_know;
ac8f6c69 3078 exit = single_exit (loop);
9baba81b 3079
0a74c758
SP
3080 if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false))
3081 {
3082 may_be_zero = niter_desc.may_be_zero;
3083 res = niter_desc.niter;
3084 }
3085
3086 if (res == chrec_dont_know
3087 || !may_be_zero
3088 || integer_zerop (may_be_zero))
3089 ;
3090 else if (integer_nonzerop (may_be_zero))
3091 res = build_int_cst (TREE_TYPE (res), 0);
9baba81b 3092
0a74c758
SP
3093 else if (COMPARISON_CLASS_P (may_be_zero))
3094 res = fold_build3 (COND_EXPR, TREE_TYPE (res), may_be_zero,
3095 build_int_cst (TREE_TYPE (res), 0), res);
9baba81b
SP
3096 else
3097 res = chrec_dont_know;
3098
dfedbe40 3099 if (dump_file && (dump_flags & TDF_SCEV))
0a74c758
SP
3100 {
3101 fprintf (dump_file, " (set_nb_iterations_in_loop = ");
3102 print_generic_expr (dump_file, res, 0);
3103 fprintf (dump_file, "))\n");
3104 }
3105
3106 loop->nb_iterations = res;
3107 return res;
9baba81b 3108}
9baba81b
SP
3109\f
3110
3111/* Counters for the stats. */
3112
b8698a0f 3113struct chrec_stats
9baba81b
SP
3114{
3115 unsigned nb_chrecs;
3116 unsigned nb_affine;
3117 unsigned nb_affine_multivar;
3118 unsigned nb_higher_poly;
3119 unsigned nb_chrec_dont_know;
3120 unsigned nb_undetermined;
3121};
3122
3123/* Reset the counters. */
3124
3125static inline void
3126reset_chrecs_counters (struct chrec_stats *stats)
3127{
3128 stats->nb_chrecs = 0;
3129 stats->nb_affine = 0;
3130 stats->nb_affine_multivar = 0;
3131 stats->nb_higher_poly = 0;
3132 stats->nb_chrec_dont_know = 0;
3133 stats->nb_undetermined = 0;
3134}
3135
3136/* Dump the contents of a CHREC_STATS structure. */
3137
3138static void
3139dump_chrecs_stats (FILE *file, struct chrec_stats *stats)
3140{
3141 fprintf (file, "\n(\n");
3142 fprintf (file, "-----------------------------------------\n");
3143 fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine);
3144 fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar);
b8698a0f 3145 fprintf (file, "%d\tdegree greater than 2 polynomials\n",
9baba81b
SP
3146 stats->nb_higher_poly);
3147 fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know);
3148 fprintf (file, "-----------------------------------------\n");
3149 fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs);
b8698a0f 3150 fprintf (file, "%d\twith undetermined coefficients\n",
9baba81b
SP
3151 stats->nb_undetermined);
3152 fprintf (file, "-----------------------------------------\n");
b8698a0f 3153 fprintf (file, "%d\tchrecs in the scev database\n",
907dadbd 3154 (int) scalar_evolution_info->elements ());
9baba81b
SP
3155 fprintf (file, "%d\tsets in the scev database\n", nb_set_scev);
3156 fprintf (file, "%d\tgets in the scev database\n", nb_get_scev);
3157 fprintf (file, "-----------------------------------------\n");
3158 fprintf (file, ")\n\n");
3159}
3160
3161/* Gather statistics about CHREC. */
3162
3163static void
3164gather_chrec_stats (tree chrec, struct chrec_stats *stats)
3165{
3166 if (dump_file && (dump_flags & TDF_STATS))
3167 {
3168 fprintf (dump_file, "(classify_chrec ");
3169 print_generic_expr (dump_file, chrec, 0);
3170 fprintf (dump_file, "\n");
3171 }
b8698a0f 3172
9baba81b 3173 stats->nb_chrecs++;
b8698a0f 3174
9baba81b
SP
3175 if (chrec == NULL_TREE)
3176 {
3177 stats->nb_undetermined++;
3178 return;
3179 }
b8698a0f 3180
9baba81b
SP
3181 switch (TREE_CODE (chrec))
3182 {
3183 case POLYNOMIAL_CHREC:
3184 if (evolution_function_is_affine_p (chrec))
3185 {
3186 if (dump_file && (dump_flags & TDF_STATS))
3187 fprintf (dump_file, " affine_univariate\n");
3188 stats->nb_affine++;
3189 }
a50411de 3190 else if (evolution_function_is_affine_multivariate_p (chrec, 0))
9baba81b
SP
3191 {
3192 if (dump_file && (dump_flags & TDF_STATS))
3193 fprintf (dump_file, " affine_multivariate\n");
3194 stats->nb_affine_multivar++;
3195 }
3196 else
3197 {
3198 if (dump_file && (dump_flags & TDF_STATS))
3199 fprintf (dump_file, " higher_degree_polynomial\n");
3200 stats->nb_higher_poly++;
3201 }
b8698a0f 3202
9baba81b
SP
3203 break;
3204
3205 default:
3206 break;
3207 }
b8698a0f 3208
9baba81b
SP
3209 if (chrec_contains_undetermined (chrec))
3210 {
3211 if (dump_file && (dump_flags & TDF_STATS))
3212 fprintf (dump_file, " undetermined\n");
3213 stats->nb_undetermined++;
3214 }
b8698a0f 3215
9baba81b
SP
3216 if (dump_file && (dump_flags & TDF_STATS))
3217 fprintf (dump_file, ")\n");
3218}
3219
9baba81b
SP
3220/* Classify the chrecs of the whole database. */
3221
b8698a0f 3222void
9baba81b
SP
3223gather_stats_on_scev_database (void)
3224{
3225 struct chrec_stats stats;
b8698a0f 3226
9baba81b
SP
3227 if (!dump_file)
3228 return;
b8698a0f 3229
9baba81b 3230 reset_chrecs_counters (&stats);
b8698a0f 3231
907dadbd
TS
3232 hash_table<scev_info_hasher>::iterator iter;
3233 scev_info_str *elt;
3234 FOR_EACH_HASH_TABLE_ELEMENT (*scalar_evolution_info, elt, scev_info_str *,
3235 iter)
3236 gather_chrec_stats (elt->chrec, &stats);
9baba81b
SP
3237
3238 dump_chrecs_stats (dump_file, &stats);
3239}
3240
3241\f
3242
3243/* Initializer. */
3244
3245static void
3246initialize_scalar_evolutions_analyzer (void)
3247{
3248 /* The elements below are unique. */
3249 if (chrec_dont_know == NULL_TREE)
3250 {
3251 chrec_not_analyzed_yet = NULL_TREE;
3252 chrec_dont_know = make_node (SCEV_NOT_KNOWN);
3253 chrec_known = make_node (SCEV_KNOWN);
d5ab5675
ZD
3254 TREE_TYPE (chrec_dont_know) = void_type_node;
3255 TREE_TYPE (chrec_known) = void_type_node;
9baba81b
SP
3256 }
3257}
3258
3259/* Initialize the analysis of scalar evolutions for LOOPS. */
3260
3261void
d73be268 3262scev_initialize (void)
9baba81b 3263{
42fd6772 3264 struct loop *loop;
9baba81b 3265
907dadbd 3266 scalar_evolution_info = hash_table<scev_info_hasher>::create_ggc (100);
b8698a0f 3267
9baba81b
SP
3268 initialize_scalar_evolutions_analyzer ();
3269
f0bd40b1 3270 FOR_EACH_LOOP (loop, 0)
42fd6772
ZD
3271 {
3272 loop->nb_iterations = NULL_TREE;
3273 }
9baba81b
SP
3274}
3275
e3a8f1fa
JH
3276/* Return true if SCEV is initialized. */
3277
3278bool
3279scev_initialized_p (void)
3280{
3281 return scalar_evolution_info != NULL;
3282}
3283
a7bf45de
SP
3284/* Cleans up the information cached by the scalar evolutions analysis
3285 in the hash table. */
3286
3287void
3288scev_reset_htab (void)
3289{
3290 if (!scalar_evolution_info)
3291 return;
3292
907dadbd 3293 scalar_evolution_info->empty ();
a7bf45de
SP
3294}
3295
3296/* Cleans up the information cached by the scalar evolutions analysis
3297 in the hash table and in the loop->nb_iterations. */
9baba81b
SP
3298
3299void
3300scev_reset (void)
3301{
9baba81b
SP
3302 struct loop *loop;
3303
a7bf45de
SP
3304 scev_reset_htab ();
3305
f0bd40b1 3306 FOR_EACH_LOOP (loop, 0)
9baba81b 3307 {
42fd6772 3308 loop->nb_iterations = NULL_TREE;
9baba81b 3309 }
e9eb809d
ZD
3310}
3311
f017bf5e
ZD
3312/* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with
3313 respect to WRTO_LOOP and returns its base and step in IV if possible
3314 (see analyze_scalar_evolution_in_loop for more details on USE_LOOP
3315 and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be
3316 invariant in LOOP. Otherwise we require it to be an integer constant.
b8698a0f 3317
f017bf5e
ZD
3318 IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g.
3319 because it is computed in signed arithmetics). Consequently, adding an
3320 induction variable
b8698a0f 3321
f017bf5e
ZD
3322 for (i = IV->base; ; i += IV->step)
3323
3324 is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is
3325 false for the type of the induction variable, or you can prove that i does
3326 not wrap by some other argument. Otherwise, this might introduce undefined
3327 behavior, and
b8698a0f 3328
f017bf5e
ZD
3329 for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step))
3330
3331 must be used instead. */
e9eb809d
ZD
3332
3333bool
f017bf5e
ZD
3334simple_iv (struct loop *wrto_loop, struct loop *use_loop, tree op,
3335 affine_iv *iv, bool allow_nonconstant_step)
e9eb809d 3336{
f3c5f3a3
BC
3337 enum tree_code code;
3338 tree type, ev, base, e, stop;
3339 wide_int extreme;
3340 bool folded_casts, overflow;
9baba81b 3341
a6f778b2
ZD
3342 iv->base = NULL_TREE;
3343 iv->step = NULL_TREE;
3344 iv->no_overflow = false;
9baba81b
SP
3345
3346 type = TREE_TYPE (op);
1ee0d660
EB
3347 if (!POINTER_TYPE_P (type)
3348 && !INTEGRAL_TYPE_P (type))
9baba81b
SP
3349 return false;
3350
f017bf5e 3351 ev = analyze_scalar_evolution_in_loop (wrto_loop, use_loop, op,
a6f778b2 3352 &folded_casts);
f017bf5e
ZD
3353 if (chrec_contains_undetermined (ev)
3354 || chrec_contains_symbols_defined_in_loop (ev, wrto_loop->num))
9baba81b
SP
3355 return false;
3356
f017bf5e 3357 if (tree_does_not_contain_chrecs (ev))
9baba81b 3358 {
a6f778b2 3359 iv->base = ev;
6e42ce54 3360 iv->step = build_int_cst (TREE_TYPE (ev), 0);
a6f778b2 3361 iv->no_overflow = true;
9baba81b
SP
3362 return true;
3363 }
3364
3365 if (TREE_CODE (ev) != POLYNOMIAL_CHREC
f017bf5e 3366 || CHREC_VARIABLE (ev) != (unsigned) wrto_loop->num)
9baba81b
SP
3367 return false;
3368
a6f778b2 3369 iv->step = CHREC_RIGHT (ev);
f017bf5e
ZD
3370 if ((!allow_nonconstant_step && TREE_CODE (iv->step) != INTEGER_CST)
3371 || tree_contains_chrecs (iv->step, NULL))
9baba81b 3372 return false;
9be872b7 3373
a6f778b2 3374 iv->base = CHREC_LEFT (ev);
f017bf5e 3375 if (tree_contains_chrecs (iv->base, NULL))
9baba81b
SP
3376 return false;
3377
1210573b 3378 iv->no_overflow = !folded_casts && nowrap_type_p (type);
eeef0e45 3379
f3c5f3a3
BC
3380 /* Try to simplify iv base:
3381
3382 (signed T) ((unsigned T)base + step) ;; TREE_TYPE (base) == signed T
3383 == (signed T)(unsigned T)base + step
3384 == base + step
3385
3386 If we can prove operation (base + step) doesn't overflow or underflow.
3387 Specifically, we try to prove below conditions are satisfied:
3388
3389 base <= UPPER_BOUND (type) - step ;;step > 0
3390 base >= LOWER_BOUND (type) - step ;;step < 0
3391
3392 This is done by proving the reverse conditions are false using loop's
3393 initial conditions.
3394
3395 The is necessary to make loop niter, or iv overflow analysis easier
3396 for below example:
3397
3398 int foo (int *a, signed char s, signed char l)
3399 {
3400 signed char i;
3401 for (i = s; i < l; i++)
3402 a[i] = 0;
3403 return 0;
3404 }
3405
3406 Note variable I is firstly converted to type unsigned char, incremented,
3407 then converted back to type signed char. */
3408
3409 if (wrto_loop->num != use_loop->num)
3410 return true;
3411
3412 if (!CONVERT_EXPR_P (iv->base) || TREE_CODE (iv->step) != INTEGER_CST)
3413 return true;
3414
3415 type = TREE_TYPE (iv->base);
3416 e = TREE_OPERAND (iv->base, 0);
3417 if (TREE_CODE (e) != PLUS_EXPR
3418 || TREE_CODE (TREE_OPERAND (e, 1)) != INTEGER_CST
3419 || !tree_int_cst_equal (iv->step,
3420 fold_convert (type, TREE_OPERAND (e, 1))))
3421 return true;
3422 e = TREE_OPERAND (e, 0);
3423 if (!CONVERT_EXPR_P (e))
3424 return true;
3425 base = TREE_OPERAND (e, 0);
3426 if (!useless_type_conversion_p (type, TREE_TYPE (base)))
3427 return true;
3428
3429 if (tree_int_cst_sign_bit (iv->step))
3430 {
3431 code = LT_EXPR;
3432 extreme = wi::min_value (type);
3433 }
3434 else
3435 {
3436 code = GT_EXPR;
3437 extreme = wi::max_value (type);
3438 }
3439 overflow = false;
3440 extreme = wi::sub (extreme, iv->step, TYPE_SIGN (type), &overflow);
3441 if (overflow)
3442 return true;
3443 e = fold_build2 (code, boolean_type_node, base,
3444 wide_int_to_tree (type, extreme));
3445 stop = (TREE_CODE (base) == SSA_NAME) ? base : NULL;
3446 e = simplify_using_initial_conditions (use_loop, e, stop);
3447 if (!integer_zerop (e))
3448 return true;
3449
3450 if (POINTER_TYPE_P (TREE_TYPE (base)))
3451 code = POINTER_PLUS_EXPR;
3452 else
3453 code = PLUS_EXPR;
3454
3455 iv->base = fold_build2 (code, TREE_TYPE (base), base, iv->step);
9baba81b
SP
3456 return true;
3457}
3458
9baba81b
SP
3459/* Finalize the scalar evolution analysis. */
3460
3461void
3462scev_finalize (void)
3463{
d51157de
ZD
3464 if (!scalar_evolution_info)
3465 return;
907dadbd 3466 scalar_evolution_info->empty ();
c7b852c8 3467 scalar_evolution_info = NULL;
9baba81b
SP
3468}
3469
771f882e
ZD
3470/* Returns true if the expression EXPR is considered to be too expensive
3471 for scev_const_prop. */
3472
3473bool
3474expression_expensive_p (tree expr)
3475{
3476 enum tree_code code;
3477
3478 if (is_gimple_val (expr))
3479 return false;
3480
3481 code = TREE_CODE (expr);
3482 if (code == TRUNC_DIV_EXPR
3483 || code == CEIL_DIV_EXPR
3484 || code == FLOOR_DIV_EXPR
3485 || code == ROUND_DIV_EXPR
3486 || code == TRUNC_MOD_EXPR
3487 || code == CEIL_MOD_EXPR
3488 || code == FLOOR_MOD_EXPR
3489 || code == ROUND_MOD_EXPR
3490 || code == EXACT_DIV_EXPR)
3491 {
3492 /* Division by power of two is usually cheap, so we allow it.
3493 Forbid anything else. */
3494 if (!integer_pow2p (TREE_OPERAND (expr, 1)))
3495 return true;
3496 }
3497
3498 switch (TREE_CODE_CLASS (code))
3499 {
3500 case tcc_binary:
3501 case tcc_comparison:
3502 if (expression_expensive_p (TREE_OPERAND (expr, 1)))
3503 return true;
3504
3505 /* Fallthru. */
3506 case tcc_unary:
3507 return expression_expensive_p (TREE_OPERAND (expr, 0));
3508
3509 default:
3510 return true;
3511 }
3512}
3513
f993a853
TV
3514/* Do final value replacement for LOOP. */
3515
3516void
3517final_value_replacement_loop (struct loop *loop)
3518{
3519 /* If we do not know exact number of iterations of the loop, we cannot
3520 replace the final value. */
3521 edge exit = single_exit (loop);
3522 if (!exit)
3523 return;
3524
3525 tree niter = number_of_latch_executions (loop);
3526 if (niter == chrec_dont_know)
3527 return;
3528
3529 /* Ensure that it is possible to insert new statements somewhere. */
3530 if (!single_pred_p (exit->dest))
3531 split_loop_exit_edge (exit);
3532
3533 /* Set stmt insertion pointer. All stmts are inserted before this point. */
3534 gimple_stmt_iterator gsi = gsi_after_labels (exit->dest);
3535
3536 struct loop *ex_loop
3537 = superloop_at_depth (loop,
3538 loop_depth (exit->dest->loop_father) + 1);
3539
3540 gphi_iterator psi;
3541 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); )
3542 {
3543 gphi *phi = psi.phi ();
3544 tree rslt = PHI_RESULT (phi);
3545 tree def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
3546 if (virtual_operand_p (def))
3547 {
3548 gsi_next (&psi);
3549 continue;
3550 }
3551
3552 if (!POINTER_TYPE_P (TREE_TYPE (def))
3553 && !INTEGRAL_TYPE_P (TREE_TYPE (def)))
3554 {
3555 gsi_next (&psi);
3556 continue;
3557 }
3558
3559 bool folded_casts;
3560 def = analyze_scalar_evolution_in_loop (ex_loop, loop, def,
3561 &folded_casts);
3562 def = compute_overall_effect_of_inner_loop (ex_loop, def);
3563 if (!tree_does_not_contain_chrecs (def)
3564 || chrec_contains_symbols_defined_in_loop (def, ex_loop->num)
3565 /* Moving the computation from the loop may prolong life range
3566 of some ssa names, which may cause problems if they appear
3567 on abnormal edges. */
3568 || contains_abnormal_ssa_name_p (def)
3569 /* Do not emit expensive expressions. The rationale is that
3570 when someone writes a code like
3571
3572 while (n > 45) n -= 45;
3573
3574 he probably knows that n is not large, and does not want it
3575 to be turned into n %= 45. */
3576 || expression_expensive_p (def))
3577 {
3578 if (dump_file && (dump_flags & TDF_DETAILS))
3579 {
3580 fprintf (dump_file, "not replacing:\n ");
3581 print_gimple_stmt (dump_file, phi, 0, 0);
3582 fprintf (dump_file, "\n");
3583 }
3584 gsi_next (&psi);
3585 continue;
3586 }
3587
3588 /* Eliminate the PHI node and replace it by a computation outside
3589 the loop. */
3590 if (dump_file)
3591 {
3592 fprintf (dump_file, "\nfinal value replacement:\n ");
3593 print_gimple_stmt (dump_file, phi, 0, 0);
3594 fprintf (dump_file, " with\n ");
3595 }
3596 def = unshare_expr (def);
3597 remove_phi_node (&psi, false);
3598
3599 /* If def's type has undefined overflow and there were folded
3600 casts, rewrite all stmts added for def into arithmetics
3601 with defined overflow behavior. */
3602 if (folded_casts && ANY_INTEGRAL_TYPE_P (TREE_TYPE (def))
3603 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (def)))
3604 {
3605 gimple_seq stmts;
3606 gimple_stmt_iterator gsi2;
3607 def = force_gimple_operand (def, &stmts, true, NULL_TREE);
3608 gsi2 = gsi_start (stmts);
3609 while (!gsi_end_p (gsi2))
3610 {
3611 gimple *stmt = gsi_stmt (gsi2);
3612 gimple_stmt_iterator gsi3 = gsi2;
3613 gsi_next (&gsi2);
3614 gsi_remove (&gsi3, false);
3615 if (is_gimple_assign (stmt)
3616 && arith_code_with_undefined_signed_overflow
3617 (gimple_assign_rhs_code (stmt)))
3618 gsi_insert_seq_before (&gsi,
3619 rewrite_to_defined_overflow (stmt),
3620 GSI_SAME_STMT);
3621 else
3622 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
3623 }
3624 }
3625 else
3626 def = force_gimple_operand_gsi (&gsi, def, false, NULL_TREE,
3627 true, GSI_SAME_STMT);
3628
3629 gassign *ass = gimple_build_assign (rslt, def);
3630 gsi_insert_before (&gsi, ass, GSI_SAME_STMT);
3631 if (dump_file)
3632 {
3633 print_gimple_stmt (dump_file, ass, 0, 0);
3634 fprintf (dump_file, "\n");
3635 }
3636 }
3637}
3638
684aaf29 3639/* Replace ssa names for that scev can prove they are constant by the
3ac01fde
ZD
3640 appropriate constants. Also perform final value replacement in loops,
3641 in case the replacement expressions are cheap.
b8698a0f 3642
684aaf29
ZD
3643 We only consider SSA names defined by phi nodes; rest is left to the
3644 ordinary constant propagation pass. */
3645
c2924966 3646unsigned int
684aaf29
ZD
3647scev_const_prop (void)
3648{
3649 basic_block bb;
726a989a 3650 tree name, type, ev;
538dd0b7 3651 gphi *phi;
f993a853 3652 struct loop *loop;
684aaf29 3653 bitmap ssa_names_to_remove = NULL;
3ac01fde 3654 unsigned i;
538dd0b7 3655 gphi_iterator psi;
684aaf29 3656
0fc822d0 3657 if (number_of_loops (cfun) <= 1)
c2924966 3658 return 0;
684aaf29 3659
11cd3bed 3660 FOR_EACH_BB_FN (bb, cfun)
684aaf29
ZD
3661 {
3662 loop = bb->loop_father;
3663
726a989a 3664 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
684aaf29 3665 {
538dd0b7 3666 phi = psi.phi ();
684aaf29
ZD
3667 name = PHI_RESULT (phi);
3668
ea057359 3669 if (virtual_operand_p (name))
684aaf29
ZD
3670 continue;
3671
3672 type = TREE_TYPE (name);
3673
3674 if (!POINTER_TYPE_P (type)
3675 && !INTEGRAL_TYPE_P (type))
3676 continue;
3677
c70ed622
BC
3678 ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name),
3679 NULL);
684aaf29
ZD
3680 if (!is_gimple_min_invariant (ev)
3681 || !may_propagate_copy (name, ev))
3682 continue;
3683
3684 /* Replace the uses of the name. */
18aed06a 3685 if (name != ev)
ed22b76f
TV
3686 {
3687 if (dump_file && (dump_flags & TDF_DETAILS))
3688 {
3689 fprintf (dump_file, "Replacing uses of: ");
3690 print_generic_expr (dump_file, name, 0);
3691 fprintf (dump_file, " with: ");
3692 print_generic_expr (dump_file, ev, 0);
3693 fprintf (dump_file, "\n");
3694 }
3695 replace_uses_by (name, ev);
3696 }
684aaf29
ZD
3697
3698 if (!ssa_names_to_remove)
3699 ssa_names_to_remove = BITMAP_ALLOC (NULL);
3700 bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name));
3701 }
3702 }
3703
9b3b55a1
DN
3704 /* Remove the ssa names that were replaced by constants. We do not
3705 remove them directly in the previous cycle, since this
3706 invalidates scev cache. */
684aaf29
ZD
3707 if (ssa_names_to_remove)
3708 {
3709 bitmap_iterator bi;
684aaf29
ZD
3710
3711 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi)
3712 {
726a989a 3713 gimple_stmt_iterator psi;
684aaf29 3714 name = ssa_name (i);
538dd0b7 3715 phi = as_a <gphi *> (SSA_NAME_DEF_STMT (name));
684aaf29 3716
726a989a
RB
3717 gcc_assert (gimple_code (phi) == GIMPLE_PHI);
3718 psi = gsi_for_stmt (phi);
3719 remove_phi_node (&psi, true);
684aaf29
ZD
3720 }
3721
3722 BITMAP_FREE (ssa_names_to_remove);
3723 scev_reset ();
3724 }
3ac01fde
ZD
3725
3726 /* Now the regular final value replacement. */
f0bd40b1 3727 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
f993a853 3728 final_value_replacement_loop (loop);
925196ed 3729
c2924966 3730 return 0;
684aaf29 3731}
9e2f83a5
ZD
3732
3733#include "gt-tree-scalar-evolution.h"