real.h (struct real_format): Split the signbit field into two two fields, signbit_ro...
[gcc.git] / gcc / tree-scalar-evolution.c
1 /* Scalar evolution detector.
2 Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <s.pop@laposte.net>
4
5 This file is part of GCC.
6
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 2, or (at your option) any later
10 version.
11
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
16
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING. If not, write to the Free
19 Software Foundation, 59 Temple Place - Suite 330, Boston, MA
20 02111-1307, USA. */
21
22 /*
23 Description:
24
25 This pass analyzes the evolution of scalar variables in loop
26 structures. The algorithm is based on the SSA representation,
27 and on the loop hierarchy tree. This algorithm is not based on
28 the notion of versions of a variable, as it was the case for the
29 previous implementations of the scalar evolution algorithm, but
30 it assumes that each defined name is unique.
31
32 The notation used in this file is called "chains of recurrences",
33 and has been proposed by Eugene Zima, Robert Van Engelen, and
34 others for describing induction variables in programs. For example
35 "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0
36 when entering in the loop_1 and has a step 2 in this loop, in other
37 words "for (b = 0; b < N; b+=2);". Note that the coefficients of
38 this chain of recurrence (or chrec [shrek]) can contain the name of
39 other variables, in which case they are called parametric chrecs.
40 For example, "b -> {a, +, 2}_1" means that the initial value of "b"
41 is the value of "a". In most of the cases these parametric chrecs
42 are fully instantiated before their use because symbolic names can
43 hide some difficult cases such as self-references described later
44 (see the Fibonacci example).
45
46 A short sketch of the algorithm is:
47
48 Given a scalar variable to be analyzed, follow the SSA edge to
49 its definition:
50
51 - When the definition is a MODIFY_EXPR: if the right hand side
52 (RHS) of the definition cannot be statically analyzed, the answer
53 of the analyzer is: "don't know".
54 Otherwise, for all the variables that are not yet analyzed in the
55 RHS, try to determine their evolution, and finally try to
56 evaluate the operation of the RHS that gives the evolution
57 function of the analyzed variable.
58
59 - When the definition is a condition-phi-node: determine the
60 evolution function for all the branches of the phi node, and
61 finally merge these evolutions (see chrec_merge).
62
63 - When the definition is a loop-phi-node: determine its initial
64 condition, that is the SSA edge defined in an outer loop, and
65 keep it symbolic. Then determine the SSA edges that are defined
66 in the body of the loop. Follow the inner edges until ending on
67 another loop-phi-node of the same analyzed loop. If the reached
68 loop-phi-node is not the starting loop-phi-node, then we keep
69 this definition under a symbolic form. If the reached
70 loop-phi-node is the same as the starting one, then we compute a
71 symbolic stride on the return path. The result is then the
72 symbolic chrec {initial_condition, +, symbolic_stride}_loop.
73
74 Examples:
75
76 Example 1: Illustration of the basic algorithm.
77
78 | a = 3
79 | loop_1
80 | b = phi (a, c)
81 | c = b + 1
82 | if (c > 10) exit_loop
83 | endloop
84
85 Suppose that we want to know the number of iterations of the
86 loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We
87 ask the scalar evolution analyzer two questions: what's the
88 scalar evolution (scev) of "c", and what's the scev of "10". For
89 "10" the answer is "10" since it is a scalar constant. For the
90 scalar variable "c", it follows the SSA edge to its definition,
91 "c = b + 1", and then asks again what's the scev of "b".
92 Following the SSA edge, we end on a loop-phi-node "b = phi (a,
93 c)", where the initial condition is "a", and the inner loop edge
94 is "c". The initial condition is kept under a symbolic form (it
95 may be the case that the copy constant propagation has done its
96 work and we end with the constant "3" as one of the edges of the
97 loop-phi-node). The update edge is followed to the end of the
98 loop, and until reaching again the starting loop-phi-node: b -> c
99 -> b. At this point we have drawn a path from "b" to "b" from
100 which we compute the stride in the loop: in this example it is
101 "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now
102 that the scev for "b" is known, it is possible to compute the
103 scev for "c", that is "c -> {a + 1, +, 1}_1". In order to
104 determine the number of iterations in the loop_1, we have to
105 instantiate_parameters ({a + 1, +, 1}_1), that gives after some
106 more analysis the scev {4, +, 1}_1, or in other words, this is
107 the function "f (x) = x + 4", where x is the iteration count of
108 the loop_1. Now we have to solve the inequality "x + 4 > 10",
109 and take the smallest iteration number for which the loop is
110 exited: x = 7. This loop runs from x = 0 to x = 7, and in total
111 there are 8 iterations. In terms of loop normalization, we have
112 created a variable that is implicitly defined, "x" or just "_1",
113 and all the other analyzed scalars of the loop are defined in
114 function of this variable:
115
116 a -> 3
117 b -> {3, +, 1}_1
118 c -> {4, +, 1}_1
119
120 or in terms of a C program:
121
122 | a = 3
123 | for (x = 0; x <= 7; x++)
124 | {
125 | b = x + 3
126 | c = x + 4
127 | }
128
129 Example 2: Illustration of the algorithm on nested loops.
130
131 | loop_1
132 | a = phi (1, b)
133 | c = a + 2
134 | loop_2 10 times
135 | b = phi (c, d)
136 | d = b + 3
137 | endloop
138 | endloop
139
140 For analyzing the scalar evolution of "a", the algorithm follows
141 the SSA edge into the loop's body: "a -> b". "b" is an inner
142 loop-phi-node, and its analysis as in Example 1, gives:
143
144 b -> {c, +, 3}_2
145 d -> {c + 3, +, 3}_2
146
147 Following the SSA edge for the initial condition, we end on "c = a
148 + 2", and then on the starting loop-phi-node "a". From this point,
149 the loop stride is computed: back on "c = a + 2" we get a "+2" in
150 the loop_1, then on the loop-phi-node "b" we compute the overall
151 effect of the inner loop that is "b = c + 30", and we get a "+30"
152 in the loop_1. That means that the overall stride in loop_1 is
153 equal to "+32", and the result is:
154
155 a -> {1, +, 32}_1
156 c -> {3, +, 32}_1
157
158 Example 3: Higher degree polynomials.
159
160 | loop_1
161 | a = phi (2, b)
162 | c = phi (5, d)
163 | b = a + 1
164 | d = c + a
165 | endloop
166
167 a -> {2, +, 1}_1
168 b -> {3, +, 1}_1
169 c -> {5, +, a}_1
170 d -> {5 + a, +, a}_1
171
172 instantiate_parameters ({5, +, a}_1) -> {5, +, 2, +, 1}_1
173 instantiate_parameters ({5 + a, +, a}_1) -> {7, +, 3, +, 1}_1
174
175 Example 4: Lucas, Fibonacci, or mixers in general.
176
177 | loop_1
178 | a = phi (1, b)
179 | c = phi (3, d)
180 | b = c
181 | d = c + a
182 | endloop
183
184 a -> (1, c)_1
185 c -> {3, +, a}_1
186
187 The syntax "(1, c)_1" stands for a PEELED_CHREC that has the
188 following semantics: during the first iteration of the loop_1, the
189 variable contains the value 1, and then it contains the value "c".
190 Note that this syntax is close to the syntax of the loop-phi-node:
191 "a -> (1, c)_1" vs. "a = phi (1, c)".
192
193 The symbolic chrec representation contains all the semantics of the
194 original code. What is more difficult is to use this information.
195
196 Example 5: Flip-flops, or exchangers.
197
198 | loop_1
199 | a = phi (1, b)
200 | c = phi (3, d)
201 | b = c
202 | d = a
203 | endloop
204
205 a -> (1, c)_1
206 c -> (3, a)_1
207
208 Based on these symbolic chrecs, it is possible to refine this
209 information into the more precise PERIODIC_CHRECs:
210
211 a -> |1, 3|_1
212 c -> |3, 1|_1
213
214 This transformation is not yet implemented.
215
216 Further readings:
217
218 You can find a more detailed description of the algorithm in:
219 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf
220 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that
221 this is a preliminary report and some of the details of the
222 algorithm have changed. I'm working on a research report that
223 updates the description of the algorithms to reflect the design
224 choices used in this implementation.
225
226 A set of slides show a high level overview of the algorithm and run
227 an example through the scalar evolution analyzer:
228 http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf
229
230 The slides that I have presented at the GCC Summit'04 are available
231 at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf
232 */
233
234 #include "config.h"
235 #include "system.h"
236 #include "coretypes.h"
237 #include "tm.h"
238 #include "errors.h"
239 #include "ggc.h"
240 #include "tree.h"
241
242 /* These RTL headers are needed for basic-block.h. */
243 #include "rtl.h"
244 #include "basic-block.h"
245 #include "diagnostic.h"
246 #include "tree-flow.h"
247 #include "tree-dump.h"
248 #include "timevar.h"
249 #include "cfgloop.h"
250 #include "tree-chrec.h"
251 #include "tree-scalar-evolution.h"
252 #include "tree-pass.h"
253 #include "flags.h"
254
255 static tree analyze_scalar_evolution_1 (struct loop *, tree, tree);
256 static tree resolve_mixers (struct loop *, tree);
257
258 /* The cached information about a ssa name VAR, claiming that inside LOOP,
259 the value of VAR can be expressed as CHREC. */
260
261 struct scev_info_str
262 {
263 tree var;
264 tree chrec;
265 };
266
267 /* Counters for the scev database. */
268 static unsigned nb_set_scev = 0;
269 static unsigned nb_get_scev = 0;
270
271 /* The following trees are unique elements. Thus the comparison of
272 another element to these elements should be done on the pointer to
273 these trees, and not on their value. */
274
275 /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
276 tree chrec_not_analyzed_yet;
277
278 /* Reserved to the cases where the analyzer has detected an
279 undecidable property at compile time. */
280 tree chrec_dont_know;
281
282 /* When the analyzer has detected that a property will never
283 happen, then it qualifies it with chrec_known. */
284 tree chrec_known;
285
286 static bitmap already_instantiated;
287
288 static htab_t scalar_evolution_info;
289
290 \f
291 /* Constructs a new SCEV_INFO_STR structure. */
292
293 static inline struct scev_info_str *
294 new_scev_info_str (tree var)
295 {
296 struct scev_info_str *res;
297
298 res = xmalloc (sizeof (struct scev_info_str));
299 res->var = var;
300 res->chrec = chrec_not_analyzed_yet;
301
302 return res;
303 }
304
305 /* Computes a hash function for database element ELT. */
306
307 static hashval_t
308 hash_scev_info (const void *elt)
309 {
310 return SSA_NAME_VERSION (((struct scev_info_str *) elt)->var);
311 }
312
313 /* Compares database elements E1 and E2. */
314
315 static int
316 eq_scev_info (const void *e1, const void *e2)
317 {
318 const struct scev_info_str *elt1 = e1;
319 const struct scev_info_str *elt2 = e2;
320
321 return elt1->var == elt2->var;
322 }
323
324 /* Deletes database element E. */
325
326 static void
327 del_scev_info (void *e)
328 {
329 free (e);
330 }
331
332 /* Get the index corresponding to VAR in the current LOOP. If
333 it's the first time we ask for this VAR, then we return
334 chrec_not_analyzed_yet for this VAR and return its index. */
335
336 static tree *
337 find_var_scev_info (tree var)
338 {
339 struct scev_info_str *res;
340 struct scev_info_str tmp;
341 PTR *slot;
342
343 tmp.var = var;
344 slot = htab_find_slot (scalar_evolution_info, &tmp, INSERT);
345
346 if (!*slot)
347 *slot = new_scev_info_str (var);
348 res = *slot;
349
350 return &res->chrec;
351 }
352
353 /* Tries to express CHREC in wider type TYPE. */
354
355 tree
356 count_ev_in_wider_type (tree type, tree chrec)
357 {
358 tree base, step;
359 struct loop *loop;
360
361 if (!evolution_function_is_affine_p (chrec))
362 return fold_convert (type, chrec);
363
364 base = CHREC_LEFT (chrec);
365 step = CHREC_RIGHT (chrec);
366 loop = current_loops->parray[CHREC_VARIABLE (chrec)];
367
368 /* TODO -- if we knew the statement at that the conversion occurs,
369 we could pass it to can_count_iv_in_wider_type and get a better
370 result. */
371 step = can_count_iv_in_wider_type (loop, type, base, step, NULL_TREE);
372 if (!step)
373 return fold_convert (type, chrec);
374 base = chrec_convert (type, base);
375
376 return build_polynomial_chrec (CHREC_VARIABLE (chrec),
377 base, step);
378 }
379
380 /* Return true when CHREC contains symbolic names defined in
381 LOOP_NB. */
382
383 bool
384 chrec_contains_symbols_defined_in_loop (tree chrec, unsigned loop_nb)
385 {
386 if (chrec == NULL_TREE)
387 return false;
388
389 if (TREE_INVARIANT (chrec))
390 return false;
391
392 if (TREE_CODE (chrec) == VAR_DECL
393 || TREE_CODE (chrec) == PARM_DECL
394 || TREE_CODE (chrec) == FUNCTION_DECL
395 || TREE_CODE (chrec) == LABEL_DECL
396 || TREE_CODE (chrec) == RESULT_DECL
397 || TREE_CODE (chrec) == FIELD_DECL)
398 return true;
399
400 if (TREE_CODE (chrec) == SSA_NAME)
401 {
402 tree def = SSA_NAME_DEF_STMT (chrec);
403 struct loop *def_loop = loop_containing_stmt (def);
404 struct loop *loop = current_loops->parray[loop_nb];
405
406 if (def_loop == NULL)
407 return false;
408
409 if (loop == def_loop || flow_loop_nested_p (loop, def_loop))
410 return true;
411
412 return false;
413 }
414
415 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
416 {
417 case 3:
418 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, 2),
419 loop_nb))
420 return true;
421
422 case 2:
423 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, 1),
424 loop_nb))
425 return true;
426
427 case 1:
428 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, 0),
429 loop_nb))
430 return true;
431
432 default:
433 return false;
434 }
435 }
436
437 /* Return true when PHI is a loop-phi-node. */
438
439 static bool
440 loop_phi_node_p (tree phi)
441 {
442 /* The implementation of this function is based on the following
443 property: "all the loop-phi-nodes of a loop are contained in the
444 loop's header basic block". */
445
446 return loop_containing_stmt (phi)->header == bb_for_stmt (phi);
447 }
448
449 /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
450 In general, in the case of multivariate evolutions we want to get
451 the evolution in different loops. LOOP specifies the level for
452 which to get the evolution.
453
454 Example:
455
456 | for (j = 0; j < 100; j++)
457 | {
458 | for (k = 0; k < 100; k++)
459 | {
460 | i = k + j; - Here the value of i is a function of j, k.
461 | }
462 | ... = i - Here the value of i is a function of j.
463 | }
464 | ... = i - Here the value of i is a scalar.
465
466 Example:
467
468 | i_0 = ...
469 | loop_1 10 times
470 | i_1 = phi (i_0, i_2)
471 | i_2 = i_1 + 2
472 | endloop
473
474 This loop has the same effect as:
475 LOOP_1 has the same effect as:
476
477 | i_1 = i_0 + 20
478
479 The overall effect of the loop, "i_0 + 20" in the previous example,
480 is obtained by passing in the parameters: LOOP = 1,
481 EVOLUTION_FN = {i_0, +, 2}_1.
482 */
483
484 static tree
485 compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn)
486 {
487 bool val = false;
488
489 if (evolution_fn == chrec_dont_know)
490 return chrec_dont_know;
491
492 else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC)
493 {
494 if (CHREC_VARIABLE (evolution_fn) >= (unsigned) loop->num)
495 {
496 struct loop *inner_loop =
497 current_loops->parray[CHREC_VARIABLE (evolution_fn)];
498 tree nb_iter = number_of_iterations_in_loop (inner_loop);
499
500 if (nb_iter == chrec_dont_know)
501 return chrec_dont_know;
502 else
503 {
504 tree res;
505
506 /* Number of iterations is off by one (the ssa name we
507 analyze must be defined before the exit). */
508 nb_iter = chrec_fold_minus (chrec_type (nb_iter),
509 nb_iter,
510 build_int_cst_type (chrec_type (nb_iter), 1));
511
512 /* evolution_fn is the evolution function in LOOP. Get
513 its value in the nb_iter-th iteration. */
514 res = chrec_apply (inner_loop->num, evolution_fn, nb_iter);
515
516 /* Continue the computation until ending on a parent of LOOP. */
517 return compute_overall_effect_of_inner_loop (loop, res);
518 }
519 }
520 else
521 return evolution_fn;
522 }
523
524 /* If the evolution function is an invariant, there is nothing to do. */
525 else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val)
526 return evolution_fn;
527
528 else
529 return chrec_dont_know;
530 }
531
532 /* Determine whether the CHREC is always positive/negative. If the expression
533 cannot be statically analyzed, return false, otherwise set the answer into
534 VALUE. */
535
536 bool
537 chrec_is_positive (tree chrec, bool *value)
538 {
539 bool value0, value1;
540 bool value2;
541 tree end_value;
542 tree nb_iter;
543
544 switch (TREE_CODE (chrec))
545 {
546 case POLYNOMIAL_CHREC:
547 if (!chrec_is_positive (CHREC_LEFT (chrec), &value0)
548 || !chrec_is_positive (CHREC_RIGHT (chrec), &value1))
549 return false;
550
551 /* FIXME -- overflows. */
552 if (value0 == value1)
553 {
554 *value = value0;
555 return true;
556 }
557
558 /* Otherwise the chrec is under the form: "{-197, +, 2}_1",
559 and the proof consists in showing that the sign never
560 changes during the execution of the loop, from 0 to
561 loop->nb_iterations. */
562 if (!evolution_function_is_affine_p (chrec))
563 return false;
564
565 nb_iter = number_of_iterations_in_loop
566 (current_loops->parray[CHREC_VARIABLE (chrec)]);
567
568 if (chrec_contains_undetermined (nb_iter))
569 return false;
570
571 nb_iter = chrec_fold_minus
572 (chrec_type (nb_iter), nb_iter,
573 build_int_cst (chrec_type (nb_iter), 1));
574
575 #if 0
576 /* TODO -- If the test is after the exit, we may decrease the number of
577 iterations by one. */
578 if (after_exit)
579 nb_iter = chrec_fold_minus
580 (chrec_type (nb_iter), nb_iter,
581 build_int_cst (chrec_type (nb_iter), 1));
582 #endif
583
584 end_value = chrec_apply (CHREC_VARIABLE (chrec), chrec, nb_iter);
585
586 if (!chrec_is_positive (end_value, &value2))
587 return false;
588
589 *value = value0;
590 return value0 == value1;
591
592 case INTEGER_CST:
593 *value = (tree_int_cst_sgn (chrec) == 1);
594 return true;
595
596 default:
597 return false;
598 }
599 }
600
601 /* Associate CHREC to SCALAR. */
602
603 static void
604 set_scalar_evolution (tree scalar, tree chrec)
605 {
606 tree *scalar_info;
607
608 if (TREE_CODE (scalar) != SSA_NAME)
609 return;
610
611 scalar_info = find_var_scev_info (scalar);
612
613 if (dump_file)
614 {
615 if (dump_flags & TDF_DETAILS)
616 {
617 fprintf (dump_file, "(set_scalar_evolution \n");
618 fprintf (dump_file, " (scalar = ");
619 print_generic_expr (dump_file, scalar, 0);
620 fprintf (dump_file, ")\n (scalar_evolution = ");
621 print_generic_expr (dump_file, chrec, 0);
622 fprintf (dump_file, "))\n");
623 }
624 if (dump_flags & TDF_STATS)
625 nb_set_scev++;
626 }
627
628 *scalar_info = chrec;
629 }
630
631 /* Retrieve the chrec associated to SCALAR in the LOOP. */
632
633 static tree
634 get_scalar_evolution (tree scalar)
635 {
636 tree res;
637
638 if (dump_file)
639 {
640 if (dump_flags & TDF_DETAILS)
641 {
642 fprintf (dump_file, "(get_scalar_evolution \n");
643 fprintf (dump_file, " (scalar = ");
644 print_generic_expr (dump_file, scalar, 0);
645 fprintf (dump_file, ")\n");
646 }
647 if (dump_flags & TDF_STATS)
648 nb_get_scev++;
649 }
650
651 switch (TREE_CODE (scalar))
652 {
653 case SSA_NAME:
654 res = *find_var_scev_info (scalar);
655 break;
656
657 case REAL_CST:
658 case INTEGER_CST:
659 res = scalar;
660 break;
661
662 default:
663 res = chrec_not_analyzed_yet;
664 break;
665 }
666
667 if (dump_file && (dump_flags & TDF_DETAILS))
668 {
669 fprintf (dump_file, " (scalar_evolution = ");
670 print_generic_expr (dump_file, res, 0);
671 fprintf (dump_file, "))\n");
672 }
673
674 return res;
675 }
676
677 /* Helper function for add_to_evolution. Returns the evolution
678 function for an assignment of the form "a = b + c", where "a" and
679 "b" are on the strongly connected component. CHREC_BEFORE is the
680 information that we already have collected up to this point.
681 TO_ADD is the evolution of "c".
682
683 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
684 evolution the expression TO_ADD, otherwise construct an evolution
685 part for this loop. */
686
687 static tree
688 add_to_evolution_1 (unsigned loop_nb,
689 tree chrec_before,
690 tree to_add)
691 {
692 switch (TREE_CODE (chrec_before))
693 {
694 case POLYNOMIAL_CHREC:
695 if (CHREC_VARIABLE (chrec_before) <= loop_nb)
696 {
697 unsigned var;
698 tree left, right;
699 tree type = chrec_type (chrec_before);
700
701 /* When there is no evolution part in this loop, build it. */
702 if (CHREC_VARIABLE (chrec_before) < loop_nb)
703 {
704 var = loop_nb;
705 left = chrec_before;
706 right = build_int_cst (type, 0);
707 }
708 else
709 {
710 var = CHREC_VARIABLE (chrec_before);
711 left = CHREC_LEFT (chrec_before);
712 right = CHREC_RIGHT (chrec_before);
713 }
714
715 return build_polynomial_chrec
716 (var, left, chrec_fold_plus (type, right, to_add));
717 }
718 else
719 /* Search the evolution in LOOP_NB. */
720 return build_polynomial_chrec
721 (CHREC_VARIABLE (chrec_before),
722 add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before), to_add),
723 CHREC_RIGHT (chrec_before));
724
725 default:
726 /* These nodes do not depend on a loop. */
727 if (chrec_before == chrec_dont_know)
728 return chrec_dont_know;
729 return build_polynomial_chrec (loop_nb, chrec_before, to_add);
730 }
731 }
732
733 /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
734 of LOOP_NB.
735
736 Description (provided for completeness, for those who read code in
737 a plane, and for my poor 62 bytes brain that would have forgotten
738 all this in the next two or three months):
739
740 The algorithm of translation of programs from the SSA representation
741 into the chrecs syntax is based on a pattern matching. After having
742 reconstructed the overall tree expression for a loop, there are only
743 two cases that can arise:
744
745 1. a = loop-phi (init, a + expr)
746 2. a = loop-phi (init, expr)
747
748 where EXPR is either a scalar constant with respect to the analyzed
749 loop (this is a degree 0 polynomial), or an expression containing
750 other loop-phi definitions (these are higher degree polynomials).
751
752 Examples:
753
754 1.
755 | init = ...
756 | loop_1
757 | a = phi (init, a + 5)
758 | endloop
759
760 2.
761 | inita = ...
762 | initb = ...
763 | loop_1
764 | a = phi (inita, 2 * b + 3)
765 | b = phi (initb, b + 1)
766 | endloop
767
768 For the first case, the semantics of the SSA representation is:
769
770 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
771
772 that is, there is a loop index "x" that determines the scalar value
773 of the variable during the loop execution. During the first
774 iteration, the value is that of the initial condition INIT, while
775 during the subsequent iterations, it is the sum of the initial
776 condition with the sum of all the values of EXPR from the initial
777 iteration to the before last considered iteration.
778
779 For the second case, the semantics of the SSA program is:
780
781 | a (x) = init, if x = 0;
782 | expr (x - 1), otherwise.
783
784 The second case corresponds to the PEELED_CHREC, whose syntax is
785 close to the syntax of a loop-phi-node:
786
787 | phi (init, expr) vs. (init, expr)_x
788
789 The proof of the translation algorithm for the first case is a
790 proof by structural induction based on the degree of EXPR.
791
792 Degree 0:
793 When EXPR is a constant with respect to the analyzed loop, or in
794 other words when EXPR is a polynomial of degree 0, the evolution of
795 the variable A in the loop is an affine function with an initial
796 condition INIT, and a step EXPR. In order to show this, we start
797 from the semantics of the SSA representation:
798
799 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
800
801 and since "expr (j)" is a constant with respect to "j",
802
803 f (x) = init + x * expr
804
805 Finally, based on the semantics of the pure sum chrecs, by
806 identification we get the corresponding chrecs syntax:
807
808 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
809 f (x) -> {init, +, expr}_x
810
811 Higher degree:
812 Suppose that EXPR is a polynomial of degree N with respect to the
813 analyzed loop_x for which we have already determined that it is
814 written under the chrecs syntax:
815
816 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
817
818 We start from the semantics of the SSA program:
819
820 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
821 |
822 | f (x) = init + \sum_{j = 0}^{x - 1}
823 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
824 |
825 | f (x) = init + \sum_{j = 0}^{x - 1}
826 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
827 |
828 | f (x) = init + \sum_{k = 0}^{n - 1}
829 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
830 |
831 | f (x) = init + \sum_{k = 0}^{n - 1}
832 | (b_k * \binom{x}{k + 1})
833 |
834 | f (x) = init + b_0 * \binom{x}{1} + ...
835 | + b_{n-1} * \binom{x}{n}
836 |
837 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
838 | + b_{n-1} * \binom{x}{n}
839 |
840
841 And finally from the definition of the chrecs syntax, we identify:
842 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
843
844 This shows the mechanism that stands behind the add_to_evolution
845 function. An important point is that the use of symbolic
846 parameters avoids the need of an analysis schedule.
847
848 Example:
849
850 | inita = ...
851 | initb = ...
852 | loop_1
853 | a = phi (inita, a + 2 + b)
854 | b = phi (initb, b + 1)
855 | endloop
856
857 When analyzing "a", the algorithm keeps "b" symbolically:
858
859 | a -> {inita, +, 2 + b}_1
860
861 Then, after instantiation, the analyzer ends on the evolution:
862
863 | a -> {inita, +, 2 + initb, +, 1}_1
864
865 */
866
867 static tree
868 add_to_evolution (unsigned loop_nb,
869 tree chrec_before,
870 enum tree_code code,
871 tree to_add)
872 {
873 tree type = chrec_type (to_add);
874 tree res = NULL_TREE;
875
876 if (to_add == NULL_TREE)
877 return chrec_before;
878
879 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
880 instantiated at this point. */
881 if (TREE_CODE (to_add) == POLYNOMIAL_CHREC)
882 /* This should not happen. */
883 return chrec_dont_know;
884
885 if (dump_file && (dump_flags & TDF_DETAILS))
886 {
887 fprintf (dump_file, "(add_to_evolution \n");
888 fprintf (dump_file, " (loop_nb = %d)\n", loop_nb);
889 fprintf (dump_file, " (chrec_before = ");
890 print_generic_expr (dump_file, chrec_before, 0);
891 fprintf (dump_file, ")\n (to_add = ");
892 print_generic_expr (dump_file, to_add, 0);
893 fprintf (dump_file, ")\n");
894 }
895
896 if (code == MINUS_EXPR)
897 to_add = chrec_fold_multiply (type, to_add,
898 build_int_cst_type (type, -1));
899
900 res = add_to_evolution_1 (loop_nb, chrec_before, to_add);
901
902 if (dump_file && (dump_flags & TDF_DETAILS))
903 {
904 fprintf (dump_file, " (res = ");
905 print_generic_expr (dump_file, res, 0);
906 fprintf (dump_file, "))\n");
907 }
908
909 return res;
910 }
911
912 /* Helper function. */
913
914 static inline tree
915 set_nb_iterations_in_loop (struct loop *loop,
916 tree res)
917 {
918 res = chrec_fold_plus (chrec_type (res), res,
919 build_int_cst_type (chrec_type (res), 1));
920
921 /* FIXME HWI: However we want to store one iteration less than the
922 count of the loop in order to be compatible with the other
923 nb_iter computations in loop-iv. This also allows the
924 representation of nb_iters that are equal to MAX_INT. */
925 if ((TREE_CODE (res) == INTEGER_CST && TREE_INT_CST_LOW (res) == 0)
926 || TREE_OVERFLOW (res))
927 res = chrec_dont_know;
928
929 if (dump_file && (dump_flags & TDF_DETAILS))
930 {
931 fprintf (dump_file, " (set_nb_iterations_in_loop = ");
932 print_generic_expr (dump_file, res, 0);
933 fprintf (dump_file, "))\n");
934 }
935
936 loop->nb_iterations = res;
937 return res;
938 }
939
940 \f
941
942 /* This section selects the loops that will be good candidates for the
943 scalar evolution analysis. For the moment, greedily select all the
944 loop nests we could analyze. */
945
946 /* Return true when it is possible to analyze the condition expression
947 EXPR. */
948
949 static bool
950 analyzable_condition (tree expr)
951 {
952 tree condition;
953
954 if (TREE_CODE (expr) != COND_EXPR)
955 return false;
956
957 condition = TREE_OPERAND (expr, 0);
958
959 switch (TREE_CODE (condition))
960 {
961 case SSA_NAME:
962 return true;
963
964 case LT_EXPR:
965 case LE_EXPR:
966 case GT_EXPR:
967 case GE_EXPR:
968 case EQ_EXPR:
969 case NE_EXPR:
970 return true;
971
972 default:
973 return false;
974 }
975
976 return false;
977 }
978
979 /* For a loop with a single exit edge, return the COND_EXPR that
980 guards the exit edge. If the expression is too difficult to
981 analyze, then give up. */
982
983 tree
984 get_loop_exit_condition (struct loop *loop)
985 {
986 tree res = NULL_TREE;
987 edge exit_edge = loop->single_exit;
988
989
990 if (dump_file && (dump_flags & TDF_DETAILS))
991 fprintf (dump_file, "(get_loop_exit_condition \n ");
992
993 if (exit_edge)
994 {
995 tree expr;
996
997 expr = last_stmt (exit_edge->src);
998 if (analyzable_condition (expr))
999 res = expr;
1000 }
1001
1002 if (dump_file && (dump_flags & TDF_DETAILS))
1003 {
1004 print_generic_expr (dump_file, res, 0);
1005 fprintf (dump_file, ")\n");
1006 }
1007
1008 return res;
1009 }
1010
1011 /* Recursively determine and enqueue the exit conditions for a loop. */
1012
1013 static void
1014 get_exit_conditions_rec (struct loop *loop,
1015 varray_type *exit_conditions)
1016 {
1017 if (!loop)
1018 return;
1019
1020 /* Recurse on the inner loops, then on the next (sibling) loops. */
1021 get_exit_conditions_rec (loop->inner, exit_conditions);
1022 get_exit_conditions_rec (loop->next, exit_conditions);
1023
1024 if (loop->single_exit)
1025 {
1026 tree loop_condition = get_loop_exit_condition (loop);
1027
1028 if (loop_condition)
1029 VARRAY_PUSH_TREE (*exit_conditions, loop_condition);
1030 }
1031 }
1032
1033 /* Select the candidate loop nests for the analysis. This function
1034 initializes the EXIT_CONDITIONS array. */
1035
1036 static void
1037 select_loops_exit_conditions (struct loops *loops,
1038 varray_type *exit_conditions)
1039 {
1040 struct loop *function_body = loops->parray[0];
1041
1042 get_exit_conditions_rec (function_body->inner, exit_conditions);
1043 }
1044
1045 \f
1046 /* Depth first search algorithm. */
1047
1048 static bool follow_ssa_edge (struct loop *loop, tree, tree, tree *);
1049
1050 /* Follow the ssa edge into the right hand side RHS of an assignment.
1051 Return true if the strongly connected component has been found. */
1052
1053 static bool
1054 follow_ssa_edge_in_rhs (struct loop *loop,
1055 tree rhs,
1056 tree halting_phi,
1057 tree *evolution_of_loop)
1058 {
1059 bool res = false;
1060 tree rhs0, rhs1;
1061 tree type_rhs = TREE_TYPE (rhs);
1062
1063 /* The RHS is one of the following cases:
1064 - an SSA_NAME,
1065 - an INTEGER_CST,
1066 - a PLUS_EXPR,
1067 - a MINUS_EXPR,
1068 - other cases are not yet handled.
1069 */
1070 switch (TREE_CODE (rhs))
1071 {
1072 case NOP_EXPR:
1073 /* This assignment is under the form "a_1 = (cast) rhs. */
1074 res = follow_ssa_edge_in_rhs (loop, TREE_OPERAND (rhs, 0), halting_phi,
1075 evolution_of_loop);
1076 *evolution_of_loop = chrec_convert (TREE_TYPE (rhs), *evolution_of_loop);
1077 break;
1078
1079 case INTEGER_CST:
1080 /* This assignment is under the form "a_1 = 7". */
1081 res = false;
1082 break;
1083
1084 case SSA_NAME:
1085 /* This assignment is under the form: "a_1 = b_2". */
1086 res = follow_ssa_edge
1087 (loop, SSA_NAME_DEF_STMT (rhs), halting_phi, evolution_of_loop);
1088 break;
1089
1090 case PLUS_EXPR:
1091 /* This case is under the form "rhs0 + rhs1". */
1092 rhs0 = TREE_OPERAND (rhs, 0);
1093 rhs1 = TREE_OPERAND (rhs, 1);
1094 STRIP_TYPE_NOPS (rhs0);
1095 STRIP_TYPE_NOPS (rhs1);
1096
1097 if (TREE_CODE (rhs0) == SSA_NAME)
1098 {
1099 if (TREE_CODE (rhs1) == SSA_NAME)
1100 {
1101 /* Match an assignment under the form:
1102 "a = b + c". */
1103 res = follow_ssa_edge
1104 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1105 evolution_of_loop);
1106
1107 if (res)
1108 *evolution_of_loop = add_to_evolution
1109 (loop->num,
1110 chrec_convert (type_rhs, *evolution_of_loop),
1111 PLUS_EXPR, rhs1);
1112
1113 else
1114 {
1115 res = follow_ssa_edge
1116 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1117 evolution_of_loop);
1118
1119 if (res)
1120 *evolution_of_loop = add_to_evolution
1121 (loop->num,
1122 chrec_convert (type_rhs, *evolution_of_loop),
1123 PLUS_EXPR, rhs0);
1124 }
1125 }
1126
1127 else
1128 {
1129 /* Match an assignment under the form:
1130 "a = b + ...". */
1131 res = follow_ssa_edge
1132 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1133 evolution_of_loop);
1134 if (res)
1135 *evolution_of_loop = add_to_evolution
1136 (loop->num, chrec_convert (type_rhs, *evolution_of_loop),
1137 PLUS_EXPR, rhs1);
1138 }
1139 }
1140
1141 else if (TREE_CODE (rhs1) == SSA_NAME)
1142 {
1143 /* Match an assignment under the form:
1144 "a = ... + c". */
1145 res = follow_ssa_edge
1146 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1147 evolution_of_loop);
1148 if (res)
1149 *evolution_of_loop = add_to_evolution
1150 (loop->num, chrec_convert (type_rhs, *evolution_of_loop),
1151 PLUS_EXPR, rhs0);
1152 }
1153
1154 else
1155 /* Otherwise, match an assignment under the form:
1156 "a = ... + ...". */
1157 /* And there is nothing to do. */
1158 res = false;
1159
1160 break;
1161
1162 case MINUS_EXPR:
1163 /* This case is under the form "opnd0 = rhs0 - rhs1". */
1164 rhs0 = TREE_OPERAND (rhs, 0);
1165 rhs1 = TREE_OPERAND (rhs, 1);
1166 STRIP_TYPE_NOPS (rhs0);
1167 STRIP_TYPE_NOPS (rhs1);
1168
1169 if (TREE_CODE (rhs0) == SSA_NAME)
1170 {
1171 /* Match an assignment under the form:
1172 "a = b - ...". */
1173 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1174 evolution_of_loop);
1175 if (res)
1176 *evolution_of_loop = add_to_evolution
1177 (loop->num, chrec_convert (type_rhs, *evolution_of_loop),
1178 MINUS_EXPR, rhs1);
1179 }
1180 else
1181 /* Otherwise, match an assignment under the form:
1182 "a = ... - ...". */
1183 /* And there is nothing to do. */
1184 res = false;
1185
1186 break;
1187
1188 case MULT_EXPR:
1189 /* This case is under the form "opnd0 = rhs0 * rhs1". */
1190 rhs0 = TREE_OPERAND (rhs, 0);
1191 rhs1 = TREE_OPERAND (rhs, 1);
1192 STRIP_TYPE_NOPS (rhs0);
1193 STRIP_TYPE_NOPS (rhs1);
1194
1195 if (TREE_CODE (rhs0) == SSA_NAME)
1196 {
1197 if (TREE_CODE (rhs1) == SSA_NAME)
1198 {
1199 /* Match an assignment under the form:
1200 "a = b * c". */
1201 res = follow_ssa_edge
1202 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1203 evolution_of_loop);
1204
1205 if (res)
1206 *evolution_of_loop = chrec_dont_know;
1207
1208 else
1209 {
1210 res = follow_ssa_edge
1211 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1212 evolution_of_loop);
1213
1214 if (res)
1215 *evolution_of_loop = chrec_dont_know;
1216 }
1217 }
1218
1219 else
1220 {
1221 /* Match an assignment under the form:
1222 "a = b * ...". */
1223 res = follow_ssa_edge
1224 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1225 evolution_of_loop);
1226 if (res)
1227 *evolution_of_loop = chrec_dont_know;
1228 }
1229 }
1230
1231 else if (TREE_CODE (rhs1) == SSA_NAME)
1232 {
1233 /* Match an assignment under the form:
1234 "a = ... * c". */
1235 res = follow_ssa_edge
1236 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1237 evolution_of_loop);
1238 if (res)
1239 *evolution_of_loop = chrec_dont_know;
1240 }
1241
1242 else
1243 /* Otherwise, match an assignment under the form:
1244 "a = ... * ...". */
1245 /* And there is nothing to do. */
1246 res = false;
1247
1248 break;
1249
1250 default:
1251 res = false;
1252 break;
1253 }
1254
1255 return res;
1256 }
1257
1258 /* Checks whether the I-th argument of a PHI comes from a backedge. */
1259
1260 static bool
1261 backedge_phi_arg_p (tree phi, int i)
1262 {
1263 edge e = PHI_ARG_EDGE (phi, i);
1264
1265 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1266 about updating it anywhere, and this should work as well most of the
1267 time. */
1268 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1269 return true;
1270
1271 return false;
1272 }
1273
1274 /* Helper function for one branch of the condition-phi-node. Return
1275 true if the strongly connected component has been found following
1276 this path. */
1277
1278 static inline bool
1279 follow_ssa_edge_in_condition_phi_branch (int i,
1280 struct loop *loop,
1281 tree condition_phi,
1282 tree halting_phi,
1283 tree *evolution_of_branch,
1284 tree init_cond)
1285 {
1286 tree branch = PHI_ARG_DEF (condition_phi, i);
1287 *evolution_of_branch = chrec_dont_know;
1288
1289 /* Do not follow back edges (they must belong to an irreducible loop, which
1290 we really do not want to worry about). */
1291 if (backedge_phi_arg_p (condition_phi, i))
1292 return false;
1293
1294 if (TREE_CODE (branch) == SSA_NAME)
1295 {
1296 *evolution_of_branch = init_cond;
1297 return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi,
1298 evolution_of_branch);
1299 }
1300
1301 /* This case occurs when one of the condition branches sets
1302 the variable to a constant: i.e. a phi-node like
1303 "a_2 = PHI <a_7(5), 2(6)>;".
1304
1305 FIXME: This case have to be refined correctly:
1306 in some cases it is possible to say something better than
1307 chrec_dont_know, for example using a wrap-around notation. */
1308 return false;
1309 }
1310
1311 /* This function merges the branches of a condition-phi-node in a
1312 loop. */
1313
1314 static bool
1315 follow_ssa_edge_in_condition_phi (struct loop *loop,
1316 tree condition_phi,
1317 tree halting_phi,
1318 tree *evolution_of_loop)
1319 {
1320 int i;
1321 tree init = *evolution_of_loop;
1322 tree evolution_of_branch;
1323
1324 if (!follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi,
1325 halting_phi,
1326 &evolution_of_branch,
1327 init))
1328 return false;
1329 *evolution_of_loop = evolution_of_branch;
1330
1331 for (i = 1; i < PHI_NUM_ARGS (condition_phi); i++)
1332 {
1333 /* Quickly give up when the evolution of one of the branches is
1334 not known. */
1335 if (*evolution_of_loop == chrec_dont_know)
1336 return true;
1337
1338 if (!follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi,
1339 halting_phi,
1340 &evolution_of_branch,
1341 init))
1342 return false;
1343
1344 *evolution_of_loop = chrec_merge (*evolution_of_loop,
1345 evolution_of_branch);
1346 }
1347
1348 return true;
1349 }
1350
1351 /* Follow an SSA edge in an inner loop. It computes the overall
1352 effect of the loop, and following the symbolic initial conditions,
1353 it follows the edges in the parent loop. The inner loop is
1354 considered as a single statement. */
1355
1356 static bool
1357 follow_ssa_edge_inner_loop_phi (struct loop *outer_loop,
1358 tree loop_phi_node,
1359 tree halting_phi,
1360 tree *evolution_of_loop)
1361 {
1362 struct loop *loop = loop_containing_stmt (loop_phi_node);
1363 tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node));
1364
1365 /* Sometimes, the inner loop is too difficult to analyze, and the
1366 result of the analysis is a symbolic parameter. */
1367 if (ev == PHI_RESULT (loop_phi_node))
1368 {
1369 bool res = false;
1370 int i;
1371
1372 for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++)
1373 {
1374 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1375 basic_block bb;
1376
1377 /* Follow the edges that exit the inner loop. */
1378 bb = PHI_ARG_EDGE (loop_phi_node, i)->src;
1379 if (!flow_bb_inside_loop_p (loop, bb))
1380 res = res || follow_ssa_edge_in_rhs (outer_loop, arg, halting_phi,
1381 evolution_of_loop);
1382 }
1383
1384 /* If the path crosses this loop-phi, give up. */
1385 if (res == true)
1386 *evolution_of_loop = chrec_dont_know;
1387
1388 return res;
1389 }
1390
1391 /* Otherwise, compute the overall effect of the inner loop. */
1392 ev = compute_overall_effect_of_inner_loop (loop, ev);
1393 return follow_ssa_edge_in_rhs (outer_loop, ev, halting_phi,
1394 evolution_of_loop);
1395 }
1396
1397 /* Follow an SSA edge from a loop-phi-node to itself, constructing a
1398 path that is analyzed on the return walk. */
1399
1400 static bool
1401 follow_ssa_edge (struct loop *loop,
1402 tree def,
1403 tree halting_phi,
1404 tree *evolution_of_loop)
1405 {
1406 struct loop *def_loop;
1407
1408 if (TREE_CODE (def) == NOP_EXPR)
1409 return false;
1410
1411 def_loop = loop_containing_stmt (def);
1412
1413 switch (TREE_CODE (def))
1414 {
1415 case PHI_NODE:
1416 if (!loop_phi_node_p (def))
1417 /* DEF is a condition-phi-node. Follow the branches, and
1418 record their evolutions. Finally, merge the collected
1419 information and set the approximation to the main
1420 variable. */
1421 return follow_ssa_edge_in_condition_phi
1422 (loop, def, halting_phi, evolution_of_loop);
1423
1424 /* When the analyzed phi is the halting_phi, the
1425 depth-first search is over: we have found a path from
1426 the halting_phi to itself in the loop. */
1427 if (def == halting_phi)
1428 return true;
1429
1430 /* Otherwise, the evolution of the HALTING_PHI depends
1431 on the evolution of another loop-phi-node, i.e. the
1432 evolution function is a higher degree polynomial. */
1433 if (def_loop == loop)
1434 return false;
1435
1436 /* Inner loop. */
1437 if (flow_loop_nested_p (loop, def_loop))
1438 return follow_ssa_edge_inner_loop_phi
1439 (loop, def, halting_phi, evolution_of_loop);
1440
1441 /* Outer loop. */
1442 return false;
1443
1444 case MODIFY_EXPR:
1445 return follow_ssa_edge_in_rhs (loop,
1446 TREE_OPERAND (def, 1),
1447 halting_phi,
1448 evolution_of_loop);
1449
1450 default:
1451 /* At this level of abstraction, the program is just a set
1452 of MODIFY_EXPRs and PHI_NODEs. In principle there is no
1453 other node to be handled. */
1454 return false;
1455 }
1456 }
1457
1458 \f
1459
1460 /* Given a LOOP_PHI_NODE, this function determines the evolution
1461 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1462
1463 static tree
1464 analyze_evolution_in_loop (tree loop_phi_node,
1465 tree init_cond)
1466 {
1467 int i;
1468 tree evolution_function = chrec_not_analyzed_yet;
1469 struct loop *loop = loop_containing_stmt (loop_phi_node);
1470 basic_block bb;
1471
1472 if (dump_file && (dump_flags & TDF_DETAILS))
1473 {
1474 fprintf (dump_file, "(analyze_evolution_in_loop \n");
1475 fprintf (dump_file, " (loop_phi_node = ");
1476 print_generic_expr (dump_file, loop_phi_node, 0);
1477 fprintf (dump_file, ")\n");
1478 }
1479
1480 for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++)
1481 {
1482 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1483 tree ssa_chain, ev_fn;
1484 bool res;
1485
1486 /* Select the edges that enter the loop body. */
1487 bb = PHI_ARG_EDGE (loop_phi_node, i)->src;
1488 if (!flow_bb_inside_loop_p (loop, bb))
1489 continue;
1490
1491 if (TREE_CODE (arg) == SSA_NAME)
1492 {
1493 ssa_chain = SSA_NAME_DEF_STMT (arg);
1494
1495 /* Pass in the initial condition to the follow edge function. */
1496 ev_fn = init_cond;
1497 res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn);
1498 }
1499 else
1500 res = false;
1501
1502 /* When it is impossible to go back on the same
1503 loop_phi_node by following the ssa edges, the
1504 evolution is represented by a peeled chrec, i.e. the
1505 first iteration, EV_FN has the value INIT_COND, then
1506 all the other iterations it has the value of ARG.
1507 For the moment, PEELED_CHREC nodes are not built. */
1508 if (!res)
1509 ev_fn = chrec_dont_know;
1510
1511 /* When there are multiple back edges of the loop (which in fact never
1512 happens currently, but nevertheless), merge their evolutions. */
1513 evolution_function = chrec_merge (evolution_function, ev_fn);
1514 }
1515
1516 if (dump_file && (dump_flags & TDF_DETAILS))
1517 {
1518 fprintf (dump_file, " (evolution_function = ");
1519 print_generic_expr (dump_file, evolution_function, 0);
1520 fprintf (dump_file, "))\n");
1521 }
1522
1523 return evolution_function;
1524 }
1525
1526 /* Given a loop-phi-node, return the initial conditions of the
1527 variable on entry of the loop. When the CCP has propagated
1528 constants into the loop-phi-node, the initial condition is
1529 instantiated, otherwise the initial condition is kept symbolic.
1530 This analyzer does not analyze the evolution outside the current
1531 loop, and leaves this task to the on-demand tree reconstructor. */
1532
1533 static tree
1534 analyze_initial_condition (tree loop_phi_node)
1535 {
1536 int i;
1537 tree init_cond = chrec_not_analyzed_yet;
1538 struct loop *loop = bb_for_stmt (loop_phi_node)->loop_father;
1539
1540 if (dump_file && (dump_flags & TDF_DETAILS))
1541 {
1542 fprintf (dump_file, "(analyze_initial_condition \n");
1543 fprintf (dump_file, " (loop_phi_node = \n");
1544 print_generic_expr (dump_file, loop_phi_node, 0);
1545 fprintf (dump_file, ")\n");
1546 }
1547
1548 for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++)
1549 {
1550 tree branch = PHI_ARG_DEF (loop_phi_node, i);
1551 basic_block bb = PHI_ARG_EDGE (loop_phi_node, i)->src;
1552
1553 /* When the branch is oriented to the loop's body, it does
1554 not contribute to the initial condition. */
1555 if (flow_bb_inside_loop_p (loop, bb))
1556 continue;
1557
1558 if (init_cond == chrec_not_analyzed_yet)
1559 {
1560 init_cond = branch;
1561 continue;
1562 }
1563
1564 if (TREE_CODE (branch) == SSA_NAME)
1565 {
1566 init_cond = chrec_dont_know;
1567 break;
1568 }
1569
1570 init_cond = chrec_merge (init_cond, branch);
1571 }
1572
1573 /* Ooops -- a loop without an entry??? */
1574 if (init_cond == chrec_not_analyzed_yet)
1575 init_cond = chrec_dont_know;
1576
1577 if (dump_file && (dump_flags & TDF_DETAILS))
1578 {
1579 fprintf (dump_file, " (init_cond = ");
1580 print_generic_expr (dump_file, init_cond, 0);
1581 fprintf (dump_file, "))\n");
1582 }
1583
1584 return init_cond;
1585 }
1586
1587 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1588
1589 static tree
1590 interpret_loop_phi (struct loop *loop, tree loop_phi_node)
1591 {
1592 tree res;
1593 struct loop *phi_loop = loop_containing_stmt (loop_phi_node);
1594 tree init_cond;
1595
1596 if (phi_loop != loop)
1597 {
1598 struct loop *subloop;
1599 tree evolution_fn = analyze_scalar_evolution
1600 (phi_loop, PHI_RESULT (loop_phi_node));
1601
1602 /* Dive one level deeper. */
1603 subloop = superloop_at_depth (phi_loop, loop->depth + 1);
1604
1605 /* Interpret the subloop. */
1606 res = compute_overall_effect_of_inner_loop (subloop, evolution_fn);
1607 return res;
1608 }
1609
1610 /* Otherwise really interpret the loop phi. */
1611 init_cond = analyze_initial_condition (loop_phi_node);
1612 res = analyze_evolution_in_loop (loop_phi_node, init_cond);
1613
1614 return res;
1615 }
1616
1617 /* This function merges the branches of a condition-phi-node,
1618 contained in the outermost loop, and whose arguments are already
1619 analyzed. */
1620
1621 static tree
1622 interpret_condition_phi (struct loop *loop, tree condition_phi)
1623 {
1624 int i;
1625 tree res = chrec_not_analyzed_yet;
1626
1627 for (i = 0; i < PHI_NUM_ARGS (condition_phi); i++)
1628 {
1629 tree branch_chrec;
1630
1631 if (backedge_phi_arg_p (condition_phi, i))
1632 {
1633 res = chrec_dont_know;
1634 break;
1635 }
1636
1637 branch_chrec = analyze_scalar_evolution
1638 (loop, PHI_ARG_DEF (condition_phi, i));
1639
1640 res = chrec_merge (res, branch_chrec);
1641 }
1642
1643 return res;
1644 }
1645
1646 /* Interpret the right hand side of a modify_expr OPND1. If we didn't
1647 analyzed this node before, follow the definitions until ending
1648 either on an analyzed modify_expr, or on a loop-phi-node. On the
1649 return path, this function propagates evolutions (ala constant copy
1650 propagation). OPND1 is not a GIMPLE expression because we could
1651 analyze the effect of an inner loop: see interpret_loop_phi. */
1652
1653 static tree
1654 interpret_rhs_modify_expr (struct loop *loop,
1655 tree opnd1, tree type)
1656 {
1657 tree res, opnd10, opnd11, chrec10, chrec11;
1658
1659 if (is_gimple_min_invariant (opnd1))
1660 return chrec_convert (type, opnd1);
1661
1662 switch (TREE_CODE (opnd1))
1663 {
1664 case PLUS_EXPR:
1665 opnd10 = TREE_OPERAND (opnd1, 0);
1666 opnd11 = TREE_OPERAND (opnd1, 1);
1667 chrec10 = analyze_scalar_evolution (loop, opnd10);
1668 chrec11 = analyze_scalar_evolution (loop, opnd11);
1669 chrec10 = chrec_convert (type, chrec10);
1670 chrec11 = chrec_convert (type, chrec11);
1671 res = chrec_fold_plus (type, chrec10, chrec11);
1672 break;
1673
1674 case MINUS_EXPR:
1675 opnd10 = TREE_OPERAND (opnd1, 0);
1676 opnd11 = TREE_OPERAND (opnd1, 1);
1677 chrec10 = analyze_scalar_evolution (loop, opnd10);
1678 chrec11 = analyze_scalar_evolution (loop, opnd11);
1679 chrec10 = chrec_convert (type, chrec10);
1680 chrec11 = chrec_convert (type, chrec11);
1681 res = chrec_fold_minus (type, chrec10, chrec11);
1682 break;
1683
1684 case NEGATE_EXPR:
1685 opnd10 = TREE_OPERAND (opnd1, 0);
1686 chrec10 = analyze_scalar_evolution (loop, opnd10);
1687 chrec10 = chrec_convert (type, chrec10);
1688 res = chrec_fold_minus (type, build_int_cst (type, 0), chrec10);
1689 break;
1690
1691 case MULT_EXPR:
1692 opnd10 = TREE_OPERAND (opnd1, 0);
1693 opnd11 = TREE_OPERAND (opnd1, 1);
1694 chrec10 = analyze_scalar_evolution (loop, opnd10);
1695 chrec11 = analyze_scalar_evolution (loop, opnd11);
1696 chrec10 = chrec_convert (type, chrec10);
1697 chrec11 = chrec_convert (type, chrec11);
1698 res = chrec_fold_multiply (type, chrec10, chrec11);
1699 break;
1700
1701 case SSA_NAME:
1702 res = chrec_convert (type, analyze_scalar_evolution (loop, opnd1));
1703 break;
1704
1705 case NOP_EXPR:
1706 case CONVERT_EXPR:
1707 opnd10 = TREE_OPERAND (opnd1, 0);
1708 chrec10 = analyze_scalar_evolution (loop, opnd10);
1709 res = chrec_convert (type, chrec10);
1710 break;
1711
1712 default:
1713 res = chrec_dont_know;
1714 break;
1715 }
1716
1717 return res;
1718 }
1719
1720 \f
1721
1722 /* This section contains all the entry points:
1723 - number_of_iterations_in_loop,
1724 - analyze_scalar_evolution,
1725 - instantiate_parameters.
1726 */
1727
1728 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1729 common ancestor of DEF_LOOP and USE_LOOP. */
1730
1731 static tree
1732 compute_scalar_evolution_in_loop (struct loop *wrto_loop,
1733 struct loop *def_loop,
1734 tree ev)
1735 {
1736 tree res;
1737 if (def_loop == wrto_loop)
1738 return ev;
1739
1740 def_loop = superloop_at_depth (def_loop, wrto_loop->depth + 1);
1741 res = compute_overall_effect_of_inner_loop (def_loop, ev);
1742
1743 return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet);
1744 }
1745
1746 /* Helper recursive function. */
1747
1748 static tree
1749 analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res)
1750 {
1751 tree def, type = TREE_TYPE (var);
1752 basic_block bb;
1753 struct loop *def_loop;
1754
1755 if (loop == NULL)
1756 return chrec_dont_know;
1757
1758 if (TREE_CODE (var) != SSA_NAME)
1759 return interpret_rhs_modify_expr (loop, var, type);
1760
1761 def = SSA_NAME_DEF_STMT (var);
1762 bb = bb_for_stmt (def);
1763 def_loop = bb ? bb->loop_father : NULL;
1764
1765 if (bb == NULL
1766 || !flow_bb_inside_loop_p (loop, bb))
1767 {
1768 /* Keep the symbolic form. */
1769 res = var;
1770 goto set_and_end;
1771 }
1772
1773 if (res != chrec_not_analyzed_yet)
1774 {
1775 if (loop != bb->loop_father)
1776 res = compute_scalar_evolution_in_loop
1777 (find_common_loop (loop, bb->loop_father), bb->loop_father, res);
1778
1779 goto set_and_end;
1780 }
1781
1782 if (loop != def_loop)
1783 {
1784 res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet);
1785 res = compute_scalar_evolution_in_loop (loop, def_loop, res);
1786
1787 goto set_and_end;
1788 }
1789
1790 switch (TREE_CODE (def))
1791 {
1792 case MODIFY_EXPR:
1793 res = interpret_rhs_modify_expr (loop, TREE_OPERAND (def, 1), type);
1794 break;
1795
1796 case PHI_NODE:
1797 if (loop_phi_node_p (def))
1798 res = interpret_loop_phi (loop, def);
1799 else
1800 res = interpret_condition_phi (loop, def);
1801 break;
1802
1803 default:
1804 res = chrec_dont_know;
1805 break;
1806 }
1807
1808 set_and_end:
1809
1810 /* Keep the symbolic form. */
1811 if (res == chrec_dont_know)
1812 res = var;
1813
1814 if (loop == def_loop)
1815 set_scalar_evolution (var, res);
1816
1817 return res;
1818 }
1819
1820 /* Entry point for the scalar evolution analyzer.
1821 Analyzes and returns the scalar evolution of the ssa_name VAR.
1822 LOOP_NB is the identifier number of the loop in which the variable
1823 is used.
1824
1825 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
1826 pointer to the statement that uses this variable, in order to
1827 determine the evolution function of the variable, use the following
1828 calls:
1829
1830 unsigned loop_nb = loop_containing_stmt (stmt)->num;
1831 tree chrec_with_symbols = analyze_scalar_evolution (loop_nb, var);
1832 tree chrec_instantiated = instantiate_parameters
1833 (loop_nb, chrec_with_symbols);
1834 */
1835
1836 tree
1837 analyze_scalar_evolution (struct loop *loop, tree var)
1838 {
1839 tree res;
1840
1841 if (dump_file && (dump_flags & TDF_DETAILS))
1842 {
1843 fprintf (dump_file, "(analyze_scalar_evolution \n");
1844 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
1845 fprintf (dump_file, " (scalar = ");
1846 print_generic_expr (dump_file, var, 0);
1847 fprintf (dump_file, ")\n");
1848 }
1849
1850 res = analyze_scalar_evolution_1 (loop, var, get_scalar_evolution (var));
1851
1852 if (TREE_CODE (var) == SSA_NAME && res == chrec_dont_know)
1853 res = var;
1854
1855 if (dump_file && (dump_flags & TDF_DETAILS))
1856 fprintf (dump_file, ")\n");
1857
1858 return res;
1859 }
1860
1861 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
1862 WRTO_LOOP (which should be a superloop of both USE_LOOP and definition
1863 of VERSION). */
1864
1865 static tree
1866 analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop,
1867 tree version)
1868 {
1869 bool val = false;
1870 tree ev = version;
1871
1872 while (1)
1873 {
1874 ev = analyze_scalar_evolution (use_loop, ev);
1875 ev = resolve_mixers (use_loop, ev);
1876
1877 if (use_loop == wrto_loop)
1878 return ev;
1879
1880 /* If the value of the use changes in the inner loop, we cannot express
1881 its value in the outer loop (we might try to return interval chrec,
1882 but we do not have a user for it anyway) */
1883 if (!no_evolution_in_loop_p (ev, use_loop->num, &val)
1884 || !val)
1885 return chrec_dont_know;
1886
1887 use_loop = use_loop->outer;
1888 }
1889 }
1890
1891 /* Returns instantiated value for VERSION in CACHE. */
1892
1893 static tree
1894 get_instantiated_value (htab_t cache, tree version)
1895 {
1896 struct scev_info_str *info, pattern;
1897
1898 pattern.var = version;
1899 info = htab_find (cache, &pattern);
1900
1901 if (info)
1902 return info->chrec;
1903 else
1904 return NULL_TREE;
1905 }
1906
1907 /* Sets instantiated value for VERSION to VAL in CACHE. */
1908
1909 static void
1910 set_instantiated_value (htab_t cache, tree version, tree val)
1911 {
1912 struct scev_info_str *info, pattern;
1913 PTR *slot;
1914
1915 pattern.var = version;
1916 slot = htab_find_slot (cache, &pattern, INSERT);
1917
1918 if (*slot)
1919 info = *slot;
1920 else
1921 info = *slot = new_scev_info_str (version);
1922 info->chrec = val;
1923 }
1924
1925 /* Analyze all the parameters of the chrec that were left under a symbolic form,
1926 with respect to LOOP. CHREC is the chrec to instantiate. If
1927 ALLOW_SUPERLOOP_CHRECS is true, replacing loop invariants with
1928 outer loop chrecs is done. CACHE is the cache of already instantiated
1929 values. */
1930
1931 static tree
1932 instantiate_parameters_1 (struct loop *loop, tree chrec,
1933 bool allow_superloop_chrecs,
1934 htab_t cache)
1935 {
1936 tree res, op0, op1, op2;
1937 basic_block def_bb;
1938 struct loop *def_loop;
1939
1940 if (chrec == NULL_TREE
1941 || automatically_generated_chrec_p (chrec))
1942 return chrec;
1943
1944 if (is_gimple_min_invariant (chrec))
1945 return chrec;
1946
1947 switch (TREE_CODE (chrec))
1948 {
1949 case SSA_NAME:
1950 def_bb = bb_for_stmt (SSA_NAME_DEF_STMT (chrec));
1951
1952 /* A parameter (or loop invariant and we do not want to include
1953 evolutions in outer loops), nothing to do. */
1954 if (!def_bb
1955 || (!allow_superloop_chrecs
1956 && !flow_bb_inside_loop_p (loop, def_bb)))
1957 return chrec;
1958
1959 /* We cache the value of instantiated variable to avoid exponential
1960 time complexity due to reevaluations. We also store the convenient
1961 value in the cache in order to prevent infinite recursion -- we do
1962 not want to instantiate the SSA_NAME if it is in a mixer
1963 structure. This is used for avoiding the instantiation of
1964 recursively defined functions, such as:
1965
1966 | a_2 -> {0, +, 1, +, a_2}_1 */
1967
1968 res = get_instantiated_value (cache, chrec);
1969 if (res)
1970 return res;
1971
1972 /* Store the convenient value for chrec in the structure. If it
1973 is defined outside of the loop, we may just leave it in symbolic
1974 form, otherwise we need to admit that we do not know its behavior
1975 inside the loop. */
1976 res = !flow_bb_inside_loop_p (loop, def_bb) ? chrec : chrec_dont_know;
1977 set_instantiated_value (cache, chrec, res);
1978
1979 /* To make things even more complicated, instantiate_parameters_1
1980 calls analyze_scalar_evolution that may call # of iterations
1981 analysis that may in turn call instantiate_parameters_1 again.
1982 To prevent the infinite recursion, keep also the bitmap of
1983 ssa names that are being instantiated globally. */
1984 if (bitmap_bit_p (already_instantiated, SSA_NAME_VERSION (chrec)))
1985 return res;
1986
1987 def_loop = find_common_loop (loop, def_bb->loop_father);
1988
1989 /* If the analysis yields a parametric chrec, instantiate the
1990 result again. */
1991 bitmap_set_bit (already_instantiated, SSA_NAME_VERSION (chrec));
1992 res = analyze_scalar_evolution (def_loop, chrec);
1993 res = instantiate_parameters_1 (loop, res, allow_superloop_chrecs, cache);
1994 bitmap_clear_bit (already_instantiated, SSA_NAME_VERSION (chrec));
1995
1996 /* Store the correct value to the cache. */
1997 set_instantiated_value (cache, chrec, res);
1998 return res;
1999
2000 case POLYNOMIAL_CHREC:
2001 op0 = instantiate_parameters_1 (loop, CHREC_LEFT (chrec),
2002 allow_superloop_chrecs, cache);
2003 op1 = instantiate_parameters_1 (loop, CHREC_RIGHT (chrec),
2004 allow_superloop_chrecs, cache);
2005 if (CHREC_LEFT (chrec) != op0
2006 || CHREC_RIGHT (chrec) != op1)
2007 chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1);
2008 return chrec;
2009
2010 case PLUS_EXPR:
2011 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2012 allow_superloop_chrecs, cache);
2013 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2014 allow_superloop_chrecs, cache);
2015 if (TREE_OPERAND (chrec, 0) != op0
2016 || TREE_OPERAND (chrec, 1) != op1)
2017 chrec = chrec_fold_plus (TREE_TYPE (chrec), op0, op1);
2018 return chrec;
2019
2020 case MINUS_EXPR:
2021 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2022 allow_superloop_chrecs, cache);
2023 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2024 allow_superloop_chrecs, cache);
2025 if (TREE_OPERAND (chrec, 0) != op0
2026 || TREE_OPERAND (chrec, 1) != op1)
2027 chrec = chrec_fold_minus (TREE_TYPE (chrec), op0, op1);
2028 return chrec;
2029
2030 case MULT_EXPR:
2031 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2032 allow_superloop_chrecs, cache);
2033 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2034 allow_superloop_chrecs, cache);
2035 if (TREE_OPERAND (chrec, 0) != op0
2036 || TREE_OPERAND (chrec, 1) != op1)
2037 chrec = chrec_fold_multiply (TREE_TYPE (chrec), op0, op1);
2038 return chrec;
2039
2040 case NOP_EXPR:
2041 case CONVERT_EXPR:
2042 case NON_LVALUE_EXPR:
2043 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2044 allow_superloop_chrecs, cache);
2045 if (op0 == chrec_dont_know)
2046 return chrec_dont_know;
2047
2048 if (op0 == TREE_OPERAND (chrec, 0))
2049 return chrec;
2050
2051 return chrec_convert (TREE_TYPE (chrec), op0);
2052
2053 case SCEV_NOT_KNOWN:
2054 return chrec_dont_know;
2055
2056 case SCEV_KNOWN:
2057 return chrec_known;
2058
2059 default:
2060 break;
2061 }
2062
2063 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
2064 {
2065 case 3:
2066 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2067 allow_superloop_chrecs, cache);
2068 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2069 allow_superloop_chrecs, cache);
2070 op2 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 2),
2071 allow_superloop_chrecs, cache);
2072 if (op0 == chrec_dont_know
2073 || op1 == chrec_dont_know
2074 || op2 == chrec_dont_know)
2075 return chrec_dont_know;
2076
2077 if (op0 == TREE_OPERAND (chrec, 0)
2078 && op1 == TREE_OPERAND (chrec, 1)
2079 && op2 == TREE_OPERAND (chrec, 2))
2080 return chrec;
2081
2082 return fold (build (TREE_CODE (chrec),
2083 TREE_TYPE (chrec), op0, op1, op2));
2084
2085 case 2:
2086 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2087 allow_superloop_chrecs, cache);
2088 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2089 allow_superloop_chrecs, cache);
2090 if (op0 == chrec_dont_know
2091 || op1 == chrec_dont_know)
2092 return chrec_dont_know;
2093
2094 if (op0 == TREE_OPERAND (chrec, 0)
2095 && op1 == TREE_OPERAND (chrec, 1))
2096 return chrec;
2097 return fold (build (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1));
2098
2099 case 1:
2100 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2101 allow_superloop_chrecs, cache);
2102 if (op0 == chrec_dont_know)
2103 return chrec_dont_know;
2104 if (op0 == TREE_OPERAND (chrec, 0))
2105 return chrec;
2106 return fold (build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0));
2107
2108 case 0:
2109 return chrec;
2110
2111 default:
2112 break;
2113 }
2114
2115 /* Too complicated to handle. */
2116 return chrec_dont_know;
2117 }
2118
2119 /* Analyze all the parameters of the chrec that were left under a
2120 symbolic form. LOOP is the loop in which symbolic names have to
2121 be analyzed and instantiated. */
2122
2123 tree
2124 instantiate_parameters (struct loop *loop,
2125 tree chrec)
2126 {
2127 tree res;
2128 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2129
2130 if (dump_file && (dump_flags & TDF_DETAILS))
2131 {
2132 fprintf (dump_file, "(instantiate_parameters \n");
2133 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
2134 fprintf (dump_file, " (chrec = ");
2135 print_generic_expr (dump_file, chrec, 0);
2136 fprintf (dump_file, ")\n");
2137 }
2138
2139 res = instantiate_parameters_1 (loop, chrec, true, cache);
2140
2141 if (dump_file && (dump_flags & TDF_DETAILS))
2142 {
2143 fprintf (dump_file, " (res = ");
2144 print_generic_expr (dump_file, res, 0);
2145 fprintf (dump_file, "))\n");
2146 }
2147
2148 htab_delete (cache);
2149
2150 return res;
2151 }
2152
2153 /* Similar to instantiate_parameters, but does not introduce the
2154 evolutions in outer loops for LOOP invariants in CHREC. */
2155
2156 static tree
2157 resolve_mixers (struct loop *loop, tree chrec)
2158 {
2159 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2160 tree ret = instantiate_parameters_1 (loop, chrec, false, cache);
2161 htab_delete (cache);
2162 return ret;
2163 }
2164
2165 /* Entry point for the analysis of the number of iterations pass.
2166 This function tries to safely approximate the number of iterations
2167 the loop will run. When this property is not decidable at compile
2168 time, the result is chrec_dont_know. Otherwise the result is
2169 a scalar or a symbolic parameter.
2170
2171 Example of analysis: suppose that the loop has an exit condition:
2172
2173 "if (b > 49) goto end_loop;"
2174
2175 and that in a previous analysis we have determined that the
2176 variable 'b' has an evolution function:
2177
2178 "EF = {23, +, 5}_2".
2179
2180 When we evaluate the function at the point 5, i.e. the value of the
2181 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2182 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2183 the loop body has been executed 6 times. */
2184
2185 tree
2186 number_of_iterations_in_loop (struct loop *loop)
2187 {
2188 tree res, type;
2189 edge exit;
2190 struct tree_niter_desc niter_desc;
2191
2192 /* Determine whether the number_of_iterations_in_loop has already
2193 been computed. */
2194 res = loop->nb_iterations;
2195 if (res)
2196 return res;
2197 res = chrec_dont_know;
2198
2199 if (dump_file && (dump_flags & TDF_DETAILS))
2200 fprintf (dump_file, "(number_of_iterations_in_loop\n");
2201
2202 exit = loop->single_exit;
2203 if (!exit)
2204 goto end;
2205
2206 if (!number_of_iterations_exit (loop, exit, &niter_desc))
2207 goto end;
2208
2209 type = TREE_TYPE (niter_desc.niter);
2210 if (integer_nonzerop (niter_desc.may_be_zero))
2211 res = build_int_cst (type, 0);
2212 else if (integer_zerop (niter_desc.may_be_zero))
2213 res = niter_desc.niter;
2214 else
2215 res = chrec_dont_know;
2216
2217 end:
2218 return set_nb_iterations_in_loop (loop, res);
2219 }
2220
2221 /* One of the drivers for testing the scalar evolutions analysis.
2222 This function computes the number of iterations for all the loops
2223 from the EXIT_CONDITIONS array. */
2224
2225 static void
2226 number_of_iterations_for_all_loops (varray_type exit_conditions)
2227 {
2228 unsigned int i;
2229 unsigned nb_chrec_dont_know_loops = 0;
2230 unsigned nb_static_loops = 0;
2231
2232 for (i = 0; i < VARRAY_ACTIVE_SIZE (exit_conditions); i++)
2233 {
2234 tree res = number_of_iterations_in_loop
2235 (loop_containing_stmt (VARRAY_TREE (exit_conditions, i)));
2236 if (chrec_contains_undetermined (res))
2237 nb_chrec_dont_know_loops++;
2238 else
2239 nb_static_loops++;
2240 }
2241
2242 if (dump_file)
2243 {
2244 fprintf (dump_file, "\n(\n");
2245 fprintf (dump_file, "-----------------------------------------\n");
2246 fprintf (dump_file, "%d\tnb_chrec_dont_know_loops\n", nb_chrec_dont_know_loops);
2247 fprintf (dump_file, "%d\tnb_static_loops\n", nb_static_loops);
2248 fprintf (dump_file, "%d\tnb_total_loops\n", current_loops->num);
2249 fprintf (dump_file, "-----------------------------------------\n");
2250 fprintf (dump_file, ")\n\n");
2251
2252 print_loop_ir (dump_file);
2253 }
2254 }
2255
2256 \f
2257
2258 /* Counters for the stats. */
2259
2260 struct chrec_stats
2261 {
2262 unsigned nb_chrecs;
2263 unsigned nb_affine;
2264 unsigned nb_affine_multivar;
2265 unsigned nb_higher_poly;
2266 unsigned nb_chrec_dont_know;
2267 unsigned nb_undetermined;
2268 };
2269
2270 /* Reset the counters. */
2271
2272 static inline void
2273 reset_chrecs_counters (struct chrec_stats *stats)
2274 {
2275 stats->nb_chrecs = 0;
2276 stats->nb_affine = 0;
2277 stats->nb_affine_multivar = 0;
2278 stats->nb_higher_poly = 0;
2279 stats->nb_chrec_dont_know = 0;
2280 stats->nb_undetermined = 0;
2281 }
2282
2283 /* Dump the contents of a CHREC_STATS structure. */
2284
2285 static void
2286 dump_chrecs_stats (FILE *file, struct chrec_stats *stats)
2287 {
2288 fprintf (file, "\n(\n");
2289 fprintf (file, "-----------------------------------------\n");
2290 fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine);
2291 fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar);
2292 fprintf (file, "%d\tdegree greater than 2 polynomials\n",
2293 stats->nb_higher_poly);
2294 fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know);
2295 fprintf (file, "-----------------------------------------\n");
2296 fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs);
2297 fprintf (file, "%d\twith undetermined coefficients\n",
2298 stats->nb_undetermined);
2299 fprintf (file, "-----------------------------------------\n");
2300 fprintf (file, "%d\tchrecs in the scev database\n",
2301 (int) htab_elements (scalar_evolution_info));
2302 fprintf (file, "%d\tsets in the scev database\n", nb_set_scev);
2303 fprintf (file, "%d\tgets in the scev database\n", nb_get_scev);
2304 fprintf (file, "-----------------------------------------\n");
2305 fprintf (file, ")\n\n");
2306 }
2307
2308 /* Gather statistics about CHREC. */
2309
2310 static void
2311 gather_chrec_stats (tree chrec, struct chrec_stats *stats)
2312 {
2313 if (dump_file && (dump_flags & TDF_STATS))
2314 {
2315 fprintf (dump_file, "(classify_chrec ");
2316 print_generic_expr (dump_file, chrec, 0);
2317 fprintf (dump_file, "\n");
2318 }
2319
2320 stats->nb_chrecs++;
2321
2322 if (chrec == NULL_TREE)
2323 {
2324 stats->nb_undetermined++;
2325 return;
2326 }
2327
2328 switch (TREE_CODE (chrec))
2329 {
2330 case POLYNOMIAL_CHREC:
2331 if (evolution_function_is_affine_p (chrec))
2332 {
2333 if (dump_file && (dump_flags & TDF_STATS))
2334 fprintf (dump_file, " affine_univariate\n");
2335 stats->nb_affine++;
2336 }
2337 else if (evolution_function_is_affine_multivariate_p (chrec))
2338 {
2339 if (dump_file && (dump_flags & TDF_STATS))
2340 fprintf (dump_file, " affine_multivariate\n");
2341 stats->nb_affine_multivar++;
2342 }
2343 else
2344 {
2345 if (dump_file && (dump_flags & TDF_STATS))
2346 fprintf (dump_file, " higher_degree_polynomial\n");
2347 stats->nb_higher_poly++;
2348 }
2349
2350 break;
2351
2352 default:
2353 break;
2354 }
2355
2356 if (chrec_contains_undetermined (chrec))
2357 {
2358 if (dump_file && (dump_flags & TDF_STATS))
2359 fprintf (dump_file, " undetermined\n");
2360 stats->nb_undetermined++;
2361 }
2362
2363 if (dump_file && (dump_flags & TDF_STATS))
2364 fprintf (dump_file, ")\n");
2365 }
2366
2367 /* One of the drivers for testing the scalar evolutions analysis.
2368 This function analyzes the scalar evolution of all the scalars
2369 defined as loop phi nodes in one of the loops from the
2370 EXIT_CONDITIONS array.
2371
2372 TODO Optimization: A loop is in canonical form if it contains only
2373 a single scalar loop phi node. All the other scalars that have an
2374 evolution in the loop are rewritten in function of this single
2375 index. This allows the parallelization of the loop. */
2376
2377 static void
2378 analyze_scalar_evolution_for_all_loop_phi_nodes (varray_type exit_conditions)
2379 {
2380 unsigned int i;
2381 struct chrec_stats stats;
2382
2383 reset_chrecs_counters (&stats);
2384
2385 for (i = 0; i < VARRAY_ACTIVE_SIZE (exit_conditions); i++)
2386 {
2387 struct loop *loop;
2388 basic_block bb;
2389 tree phi, chrec;
2390
2391 loop = loop_containing_stmt (VARRAY_TREE (exit_conditions, i));
2392 bb = loop->header;
2393
2394 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
2395 if (is_gimple_reg (PHI_RESULT (phi)))
2396 {
2397 chrec = instantiate_parameters
2398 (loop,
2399 analyze_scalar_evolution (loop, PHI_RESULT (phi)));
2400
2401 if (dump_file && (dump_flags & TDF_STATS))
2402 gather_chrec_stats (chrec, &stats);
2403 }
2404 }
2405
2406 if (dump_file && (dump_flags & TDF_STATS))
2407 dump_chrecs_stats (dump_file, &stats);
2408 }
2409
2410 /* Callback for htab_traverse, gathers information on chrecs in the
2411 hashtable. */
2412
2413 static int
2414 gather_stats_on_scev_database_1 (void **slot, void *stats)
2415 {
2416 struct scev_info_str *entry = *slot;
2417
2418 gather_chrec_stats (entry->chrec, stats);
2419
2420 return 1;
2421 }
2422
2423 /* Classify the chrecs of the whole database. */
2424
2425 void
2426 gather_stats_on_scev_database (void)
2427 {
2428 struct chrec_stats stats;
2429
2430 if (!dump_file)
2431 return;
2432
2433 reset_chrecs_counters (&stats);
2434
2435 htab_traverse (scalar_evolution_info, gather_stats_on_scev_database_1,
2436 &stats);
2437
2438 dump_chrecs_stats (dump_file, &stats);
2439 }
2440
2441 \f
2442
2443 /* Initializer. */
2444
2445 static void
2446 initialize_scalar_evolutions_analyzer (void)
2447 {
2448 /* The elements below are unique. */
2449 if (chrec_dont_know == NULL_TREE)
2450 {
2451 chrec_not_analyzed_yet = NULL_TREE;
2452 chrec_dont_know = make_node (SCEV_NOT_KNOWN);
2453 chrec_known = make_node (SCEV_KNOWN);
2454 TREE_TYPE (chrec_dont_know) = NULL_TREE;
2455 TREE_TYPE (chrec_known) = NULL_TREE;
2456 }
2457 }
2458
2459 /* Initialize the analysis of scalar evolutions for LOOPS. */
2460
2461 void
2462 scev_initialize (struct loops *loops)
2463 {
2464 unsigned i;
2465 current_loops = loops;
2466
2467 scalar_evolution_info = htab_create (100, hash_scev_info,
2468 eq_scev_info, del_scev_info);
2469 already_instantiated = BITMAP_ALLOC (NULL);
2470
2471 initialize_scalar_evolutions_analyzer ();
2472
2473 for (i = 1; i < loops->num; i++)
2474 if (loops->parray[i])
2475 loops->parray[i]->nb_iterations = NULL_TREE;
2476 }
2477
2478 /* Cleans up the information cached by the scalar evolutions analysis. */
2479
2480 void
2481 scev_reset (void)
2482 {
2483 unsigned i;
2484 struct loop *loop;
2485
2486 if (!scalar_evolution_info || !current_loops)
2487 return;
2488
2489 htab_empty (scalar_evolution_info);
2490 for (i = 1; i < current_loops->num; i++)
2491 {
2492 loop = current_loops->parray[i];
2493 if (loop)
2494 loop->nb_iterations = NULL_TREE;
2495 }
2496 }
2497
2498 /* Checks whether OP behaves as a simple affine iv of LOOP in STMT and returns
2499 its BASE and STEP if possible. */
2500
2501 bool
2502 simple_iv (struct loop *loop, tree stmt, tree op, tree *base, tree *step)
2503 {
2504 basic_block bb = bb_for_stmt (stmt);
2505 tree type, ev;
2506
2507 *base = NULL_TREE;
2508 *step = NULL_TREE;
2509
2510 type = TREE_TYPE (op);
2511 if (TREE_CODE (type) != INTEGER_TYPE
2512 && TREE_CODE (type) != POINTER_TYPE)
2513 return false;
2514
2515 ev = analyze_scalar_evolution_in_loop (loop, bb->loop_father, op);
2516 if (chrec_contains_undetermined (ev))
2517 return false;
2518
2519 if (tree_does_not_contain_chrecs (ev)
2520 && !chrec_contains_symbols_defined_in_loop (ev, loop->num))
2521 {
2522 *base = ev;
2523 return true;
2524 }
2525
2526 if (TREE_CODE (ev) != POLYNOMIAL_CHREC
2527 || CHREC_VARIABLE (ev) != (unsigned) loop->num)
2528 return false;
2529
2530 *step = CHREC_RIGHT (ev);
2531 if (TREE_CODE (*step) != INTEGER_CST)
2532 return false;
2533 *base = CHREC_LEFT (ev);
2534 if (tree_contains_chrecs (*base)
2535 || chrec_contains_symbols_defined_in_loop (*base, loop->num))
2536 return false;
2537
2538 return true;
2539 }
2540
2541 /* Runs the analysis of scalar evolutions. */
2542
2543 void
2544 scev_analysis (void)
2545 {
2546 varray_type exit_conditions;
2547
2548 VARRAY_GENERIC_PTR_INIT (exit_conditions, 37, "exit_conditions");
2549 select_loops_exit_conditions (current_loops, &exit_conditions);
2550
2551 if (dump_file && (dump_flags & TDF_STATS))
2552 analyze_scalar_evolution_for_all_loop_phi_nodes (exit_conditions);
2553
2554 number_of_iterations_for_all_loops (exit_conditions);
2555 VARRAY_CLEAR (exit_conditions);
2556 }
2557
2558 /* Finalize the scalar evolution analysis. */
2559
2560 void
2561 scev_finalize (void)
2562 {
2563 htab_delete (scalar_evolution_info);
2564 BITMAP_FREE (already_instantiated);
2565 }
2566