]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/tree-scalar-evolution.c
Change copyright header to refer to version 3 of the GNU General Public License and...
[thirdparty/gcc.git] / gcc / tree-scalar-evolution.c
1 /* Scalar evolution detector.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007 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 3, 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 COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
20
21 /*
22 Description:
23
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).
44
45 A short sketch of the algorithm is:
46
47 Given a scalar variable to be analyzed, follow the SSA edge to
48 its definition:
49
50 - When the definition is a GIMPLE_MODIFY_STMT: if the right hand side
51 (RHS) of the definition cannot be statically analyzed, the answer
52 of the analyzer is: "don't know".
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:
74
75 Example 1: Illustration of the basic algorithm.
76
77 | a = 3
78 | loop_1
79 | b = phi (a, c)
80 | c = b + 1
81 | if (c > 10) exit_loop
82 | endloop
83
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
104 instantiate_parameters ({a + 1, +, 1}_1), that gives after some
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:
114
115 a -> 3
116 b -> {3, +, 1}_1
117 c -> {4, +, 1}_1
118
119 or in terms of a C program:
120
121 | a = 3
122 | for (x = 0; x <= 7; x++)
123 | {
124 | b = x + 3
125 | c = x + 4
126 | }
127
128 Example 2: Illustration of the algorithm on nested loops.
129
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
138
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
141 loop-phi-node, and its analysis as in Example 1, gives:
142
143 b -> {c, +, 3}_2
144 d -> {c + 3, +, 3}_2
145
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
152 equal to "+32", and the result is:
153
154 a -> {1, +, 32}_1
155 c -> {3, +, 32}_1
156
157 Example 3: Higher degree polynomials.
158
159 | loop_1
160 | a = phi (2, b)
161 | c = phi (5, d)
162 | b = a + 1
163 | d = c + a
164 | endloop
165
166 a -> {2, +, 1}_1
167 b -> {3, +, 1}_1
168 c -> {5, +, a}_1
169 d -> {5 + a, +, a}_1
170
171 instantiate_parameters ({5, +, a}_1) -> {5, +, 2, +, 1}_1
172 instantiate_parameters ({5 + a, +, a}_1) -> {7, +, 3, +, 1}_1
173
174 Example 4: Lucas, Fibonacci, or mixers in general.
175
176 | loop_1
177 | a = phi (1, b)
178 | c = phi (3, d)
179 | b = c
180 | d = c + a
181 | endloop
182
183 a -> (1, c)_1
184 c -> {3, +, a}_1
185
186 The syntax "(1, c)_1" stands for a PEELED_CHREC that has the
187 following semantics: during the first iteration of the loop_1, the
188 variable contains the value 1, and then it contains the value "c".
189 Note that this syntax is close to the syntax of the loop-phi-node:
190 "a -> (1, c)_1" vs. "a = phi (1, c)".
191
192 The symbolic chrec representation contains all the semantics of the
193 original code. What is more difficult is to use this information.
194
195 Example 5: Flip-flops, or exchangers.
196
197 | loop_1
198 | a = phi (1, b)
199 | c = phi (3, d)
200 | b = c
201 | d = a
202 | endloop
203
204 a -> (1, c)_1
205 c -> (3, a)_1
206
207 Based on these symbolic chrecs, it is possible to refine this
208 information into the more precise PERIODIC_CHRECs:
209
210 a -> |1, 3|_1
211 c -> |3, 1|_1
212
213 This transformation is not yet implemented.
214
215 Further readings:
216
217 You can find a more detailed description of the algorithm in:
218 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf
219 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that
220 this is a preliminary report and some of the details of the
221 algorithm have changed. I'm working on a research report that
222 updates the description of the algorithms to reflect the design
223 choices used in this implementation.
224
225 A set of slides show a high level overview of the algorithm and run
226 an example through the scalar evolution analyzer:
227 http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf
228
229 The slides that I have presented at the GCC Summit'04 are available
230 at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf
231 */
232
233 #include "config.h"
234 #include "system.h"
235 #include "coretypes.h"
236 #include "tm.h"
237 #include "ggc.h"
238 #include "tree.h"
239 #include "real.h"
240
241 /* These RTL headers are needed for basic-block.h. */
242 #include "rtl.h"
243 #include "basic-block.h"
244 #include "diagnostic.h"
245 #include "tree-flow.h"
246 #include "tree-dump.h"
247 #include "timevar.h"
248 #include "cfgloop.h"
249 #include "tree-chrec.h"
250 #include "tree-scalar-evolution.h"
251 #include "tree-pass.h"
252 #include "flags.h"
253 #include "params.h"
254
255 static tree analyze_scalar_evolution_1 (struct loop *, tree, tree);
256
257 /* The cached information about a ssa name VAR, claiming that inside LOOP,
258 the value of VAR can be expressed as CHREC. */
259
260 struct scev_info_str GTY(())
261 {
262 tree var;
263 tree chrec;
264 };
265
266 /* Counters for the scev database. */
267 static unsigned nb_set_scev = 0;
268 static unsigned nb_get_scev = 0;
269
270 /* The following trees are unique elements. Thus the comparison of
271 another element to these elements should be done on the pointer to
272 these trees, and not on their value. */
273
274 /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
275 tree chrec_not_analyzed_yet;
276
277 /* Reserved to the cases where the analyzer has detected an
278 undecidable property at compile time. */
279 tree chrec_dont_know;
280
281 /* When the analyzer has detected that a property will never
282 happen, then it qualifies it with chrec_known. */
283 tree chrec_known;
284
285 static bitmap already_instantiated;
286
287 static GTY ((param_is (struct scev_info_str))) htab_t scalar_evolution_info;
288
289 \f
290 /* Constructs a new SCEV_INFO_STR structure. */
291
292 static inline struct scev_info_str *
293 new_scev_info_str (tree var)
294 {
295 struct scev_info_str *res;
296
297 res = GGC_NEW (struct scev_info_str);
298 res->var = var;
299 res->chrec = chrec_not_analyzed_yet;
300
301 return res;
302 }
303
304 /* Computes a hash function for database element ELT. */
305
306 static hashval_t
307 hash_scev_info (const void *elt)
308 {
309 return SSA_NAME_VERSION (((const struct scev_info_str *) elt)->var);
310 }
311
312 /* Compares database elements E1 and E2. */
313
314 static int
315 eq_scev_info (const void *e1, const void *e2)
316 {
317 const struct scev_info_str *elt1 = (const struct scev_info_str *) e1;
318 const struct scev_info_str *elt2 = (const struct scev_info_str *) e2;
319
320 return elt1->var == elt2->var;
321 }
322
323 /* Deletes database element E. */
324
325 static void
326 del_scev_info (void *e)
327 {
328 ggc_free (e);
329 }
330
331 /* Get the index corresponding to VAR in the current LOOP. If
332 it's the first time we ask for this VAR, then we return
333 chrec_not_analyzed_yet for this VAR and return its index. */
334
335 static tree *
336 find_var_scev_info (tree var)
337 {
338 struct scev_info_str *res;
339 struct scev_info_str tmp;
340 PTR *slot;
341
342 tmp.var = var;
343 slot = htab_find_slot (scalar_evolution_info, &tmp, INSERT);
344
345 if (!*slot)
346 *slot = new_scev_info_str (var);
347 res = (struct scev_info_str *) *slot;
348
349 return &res->chrec;
350 }
351
352 /* Return true when CHREC contains symbolic names defined in
353 LOOP_NB. */
354
355 bool
356 chrec_contains_symbols_defined_in_loop (tree chrec, unsigned loop_nb)
357 {
358 int i, n;
359
360 if (chrec == NULL_TREE)
361 return false;
362
363 if (TREE_INVARIANT (chrec))
364 return false;
365
366 if (TREE_CODE (chrec) == VAR_DECL
367 || TREE_CODE (chrec) == PARM_DECL
368 || TREE_CODE (chrec) == FUNCTION_DECL
369 || TREE_CODE (chrec) == LABEL_DECL
370 || TREE_CODE (chrec) == RESULT_DECL
371 || TREE_CODE (chrec) == FIELD_DECL)
372 return true;
373
374 if (TREE_CODE (chrec) == SSA_NAME)
375 {
376 tree def = SSA_NAME_DEF_STMT (chrec);
377 struct loop *def_loop = loop_containing_stmt (def);
378 struct loop *loop = get_loop (loop_nb);
379
380 if (def_loop == NULL)
381 return false;
382
383 if (loop == def_loop || flow_loop_nested_p (loop, def_loop))
384 return true;
385
386 return false;
387 }
388
389 n = TREE_OPERAND_LENGTH (chrec);
390 for (i = 0; i < n; i++)
391 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, i),
392 loop_nb))
393 return true;
394 return false;
395 }
396
397 /* Return true when PHI is a loop-phi-node. */
398
399 static bool
400 loop_phi_node_p (tree phi)
401 {
402 /* The implementation of this function is based on the following
403 property: "all the loop-phi-nodes of a loop are contained in the
404 loop's header basic block". */
405
406 return loop_containing_stmt (phi)->header == bb_for_stmt (phi);
407 }
408
409 /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
410 In general, in the case of multivariate evolutions we want to get
411 the evolution in different loops. LOOP specifies the level for
412 which to get the evolution.
413
414 Example:
415
416 | for (j = 0; j < 100; j++)
417 | {
418 | for (k = 0; k < 100; k++)
419 | {
420 | i = k + j; - Here the value of i is a function of j, k.
421 | }
422 | ... = i - Here the value of i is a function of j.
423 | }
424 | ... = i - Here the value of i is a scalar.
425
426 Example:
427
428 | i_0 = ...
429 | loop_1 10 times
430 | i_1 = phi (i_0, i_2)
431 | i_2 = i_1 + 2
432 | endloop
433
434 This loop has the same effect as:
435 LOOP_1 has the same effect as:
436
437 | i_1 = i_0 + 20
438
439 The overall effect of the loop, "i_0 + 20" in the previous example,
440 is obtained by passing in the parameters: LOOP = 1,
441 EVOLUTION_FN = {i_0, +, 2}_1.
442 */
443
444 static tree
445 compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn)
446 {
447 bool val = false;
448
449 if (evolution_fn == chrec_dont_know)
450 return chrec_dont_know;
451
452 else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC)
453 {
454 struct loop *inner_loop = get_chrec_loop (evolution_fn);
455
456 if (inner_loop == loop
457 || flow_loop_nested_p (loop, inner_loop))
458 {
459 tree nb_iter = number_of_latch_executions (inner_loop);
460
461 if (nb_iter == chrec_dont_know)
462 return chrec_dont_know;
463 else
464 {
465 tree res;
466
467 /* evolution_fn is the evolution function in LOOP. Get
468 its value in the nb_iter-th iteration. */
469 res = chrec_apply (inner_loop->num, evolution_fn, nb_iter);
470
471 /* Continue the computation until ending on a parent of LOOP. */
472 return compute_overall_effect_of_inner_loop (loop, res);
473 }
474 }
475 else
476 return evolution_fn;
477 }
478
479 /* If the evolution function is an invariant, there is nothing to do. */
480 else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val)
481 return evolution_fn;
482
483 else
484 return chrec_dont_know;
485 }
486
487 /* Determine whether the CHREC is always positive/negative. If the expression
488 cannot be statically analyzed, return false, otherwise set the answer into
489 VALUE. */
490
491 bool
492 chrec_is_positive (tree chrec, bool *value)
493 {
494 bool value0, value1, value2;
495 tree end_value, nb_iter;
496
497 switch (TREE_CODE (chrec))
498 {
499 case POLYNOMIAL_CHREC:
500 if (!chrec_is_positive (CHREC_LEFT (chrec), &value0)
501 || !chrec_is_positive (CHREC_RIGHT (chrec), &value1))
502 return false;
503
504 /* FIXME -- overflows. */
505 if (value0 == value1)
506 {
507 *value = value0;
508 return true;
509 }
510
511 /* Otherwise the chrec is under the form: "{-197, +, 2}_1",
512 and the proof consists in showing that the sign never
513 changes during the execution of the loop, from 0 to
514 loop->nb_iterations. */
515 if (!evolution_function_is_affine_p (chrec))
516 return false;
517
518 nb_iter = number_of_latch_executions (get_chrec_loop (chrec));
519 if (chrec_contains_undetermined (nb_iter))
520 return false;
521
522 #if 0
523 /* TODO -- If the test is after the exit, we may decrease the number of
524 iterations by one. */
525 if (after_exit)
526 nb_iter = chrec_fold_minus (type, nb_iter, build_int_cst (type, 1));
527 #endif
528
529 end_value = chrec_apply (CHREC_VARIABLE (chrec), chrec, nb_iter);
530
531 if (!chrec_is_positive (end_value, &value2))
532 return false;
533
534 *value = value0;
535 return value0 == value1;
536
537 case INTEGER_CST:
538 *value = (tree_int_cst_sgn (chrec) == 1);
539 return true;
540
541 default:
542 return false;
543 }
544 }
545
546 /* Associate CHREC to SCALAR. */
547
548 static void
549 set_scalar_evolution (tree scalar, tree chrec)
550 {
551 tree *scalar_info;
552
553 if (TREE_CODE (scalar) != SSA_NAME)
554 return;
555
556 scalar_info = find_var_scev_info (scalar);
557
558 if (dump_file)
559 {
560 if (dump_flags & TDF_DETAILS)
561 {
562 fprintf (dump_file, "(set_scalar_evolution \n");
563 fprintf (dump_file, " (scalar = ");
564 print_generic_expr (dump_file, scalar, 0);
565 fprintf (dump_file, ")\n (scalar_evolution = ");
566 print_generic_expr (dump_file, chrec, 0);
567 fprintf (dump_file, "))\n");
568 }
569 if (dump_flags & TDF_STATS)
570 nb_set_scev++;
571 }
572
573 *scalar_info = chrec;
574 }
575
576 /* Retrieve the chrec associated to SCALAR in the LOOP. */
577
578 static tree
579 get_scalar_evolution (tree scalar)
580 {
581 tree res;
582
583 if (dump_file)
584 {
585 if (dump_flags & TDF_DETAILS)
586 {
587 fprintf (dump_file, "(get_scalar_evolution \n");
588 fprintf (dump_file, " (scalar = ");
589 print_generic_expr (dump_file, scalar, 0);
590 fprintf (dump_file, ")\n");
591 }
592 if (dump_flags & TDF_STATS)
593 nb_get_scev++;
594 }
595
596 switch (TREE_CODE (scalar))
597 {
598 case SSA_NAME:
599 res = *find_var_scev_info (scalar);
600 break;
601
602 case REAL_CST:
603 case INTEGER_CST:
604 res = scalar;
605 break;
606
607 default:
608 res = chrec_not_analyzed_yet;
609 break;
610 }
611
612 if (dump_file && (dump_flags & TDF_DETAILS))
613 {
614 fprintf (dump_file, " (scalar_evolution = ");
615 print_generic_expr (dump_file, res, 0);
616 fprintf (dump_file, "))\n");
617 }
618
619 return res;
620 }
621
622 /* Helper function for add_to_evolution. Returns the evolution
623 function for an assignment of the form "a = b + c", where "a" and
624 "b" are on the strongly connected component. CHREC_BEFORE is the
625 information that we already have collected up to this point.
626 TO_ADD is the evolution of "c".
627
628 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
629 evolution the expression TO_ADD, otherwise construct an evolution
630 part for this loop. */
631
632 static tree
633 add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add,
634 tree at_stmt)
635 {
636 tree type, left, right;
637 struct loop *loop = get_loop (loop_nb), *chloop;
638
639 switch (TREE_CODE (chrec_before))
640 {
641 case POLYNOMIAL_CHREC:
642 chloop = get_chrec_loop (chrec_before);
643 if (chloop == loop
644 || flow_loop_nested_p (chloop, loop))
645 {
646 unsigned var;
647
648 type = chrec_type (chrec_before);
649
650 /* When there is no evolution part in this loop, build it. */
651 if (chloop != loop)
652 {
653 var = loop_nb;
654 left = chrec_before;
655 right = SCALAR_FLOAT_TYPE_P (type)
656 ? build_real (type, dconst0)
657 : build_int_cst (type, 0);
658 }
659 else
660 {
661 var = CHREC_VARIABLE (chrec_before);
662 left = CHREC_LEFT (chrec_before);
663 right = CHREC_RIGHT (chrec_before);
664 }
665
666 to_add = chrec_convert (type, to_add, at_stmt);
667 right = chrec_convert_rhs (type, right, at_stmt);
668 right = chrec_fold_plus (chrec_type (right), right, to_add);
669 return build_polynomial_chrec (var, left, right);
670 }
671 else
672 {
673 gcc_assert (flow_loop_nested_p (loop, chloop));
674
675 /* Search the evolution in LOOP_NB. */
676 left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before),
677 to_add, at_stmt);
678 right = CHREC_RIGHT (chrec_before);
679 right = chrec_convert_rhs (chrec_type (left), right, at_stmt);
680 return build_polynomial_chrec (CHREC_VARIABLE (chrec_before),
681 left, right);
682 }
683
684 default:
685 /* These nodes do not depend on a loop. */
686 if (chrec_before == chrec_dont_know)
687 return chrec_dont_know;
688
689 left = chrec_before;
690 right = chrec_convert_rhs (chrec_type (left), to_add, at_stmt);
691 return build_polynomial_chrec (loop_nb, left, right);
692 }
693 }
694
695 /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
696 of LOOP_NB.
697
698 Description (provided for completeness, for those who read code in
699 a plane, and for my poor 62 bytes brain that would have forgotten
700 all this in the next two or three months):
701
702 The algorithm of translation of programs from the SSA representation
703 into the chrecs syntax is based on a pattern matching. After having
704 reconstructed the overall tree expression for a loop, there are only
705 two cases that can arise:
706
707 1. a = loop-phi (init, a + expr)
708 2. a = loop-phi (init, expr)
709
710 where EXPR is either a scalar constant with respect to the analyzed
711 loop (this is a degree 0 polynomial), or an expression containing
712 other loop-phi definitions (these are higher degree polynomials).
713
714 Examples:
715
716 1.
717 | init = ...
718 | loop_1
719 | a = phi (init, a + 5)
720 | endloop
721
722 2.
723 | inita = ...
724 | initb = ...
725 | loop_1
726 | a = phi (inita, 2 * b + 3)
727 | b = phi (initb, b + 1)
728 | endloop
729
730 For the first case, the semantics of the SSA representation is:
731
732 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
733
734 that is, there is a loop index "x" that determines the scalar value
735 of the variable during the loop execution. During the first
736 iteration, the value is that of the initial condition INIT, while
737 during the subsequent iterations, it is the sum of the initial
738 condition with the sum of all the values of EXPR from the initial
739 iteration to the before last considered iteration.
740
741 For the second case, the semantics of the SSA program is:
742
743 | a (x) = init, if x = 0;
744 | expr (x - 1), otherwise.
745
746 The second case corresponds to the PEELED_CHREC, whose syntax is
747 close to the syntax of a loop-phi-node:
748
749 | phi (init, expr) vs. (init, expr)_x
750
751 The proof of the translation algorithm for the first case is a
752 proof by structural induction based on the degree of EXPR.
753
754 Degree 0:
755 When EXPR is a constant with respect to the analyzed loop, or in
756 other words when EXPR is a polynomial of degree 0, the evolution of
757 the variable A in the loop is an affine function with an initial
758 condition INIT, and a step EXPR. In order to show this, we start
759 from the semantics of the SSA representation:
760
761 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
762
763 and since "expr (j)" is a constant with respect to "j",
764
765 f (x) = init + x * expr
766
767 Finally, based on the semantics of the pure sum chrecs, by
768 identification we get the corresponding chrecs syntax:
769
770 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
771 f (x) -> {init, +, expr}_x
772
773 Higher degree:
774 Suppose that EXPR is a polynomial of degree N with respect to the
775 analyzed loop_x for which we have already determined that it is
776 written under the chrecs syntax:
777
778 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
779
780 We start from the semantics of the SSA program:
781
782 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
783 |
784 | f (x) = init + \sum_{j = 0}^{x - 1}
785 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
786 |
787 | f (x) = init + \sum_{j = 0}^{x - 1}
788 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
789 |
790 | f (x) = init + \sum_{k = 0}^{n - 1}
791 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
792 |
793 | f (x) = init + \sum_{k = 0}^{n - 1}
794 | (b_k * \binom{x}{k + 1})
795 |
796 | f (x) = init + b_0 * \binom{x}{1} + ...
797 | + b_{n-1} * \binom{x}{n}
798 |
799 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
800 | + b_{n-1} * \binom{x}{n}
801 |
802
803 And finally from the definition of the chrecs syntax, we identify:
804 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
805
806 This shows the mechanism that stands behind the add_to_evolution
807 function. An important point is that the use of symbolic
808 parameters avoids the need of an analysis schedule.
809
810 Example:
811
812 | inita = ...
813 | initb = ...
814 | loop_1
815 | a = phi (inita, a + 2 + b)
816 | b = phi (initb, b + 1)
817 | endloop
818
819 When analyzing "a", the algorithm keeps "b" symbolically:
820
821 | a -> {inita, +, 2 + b}_1
822
823 Then, after instantiation, the analyzer ends on the evolution:
824
825 | a -> {inita, +, 2 + initb, +, 1}_1
826
827 */
828
829 static tree
830 add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code,
831 tree to_add, tree at_stmt)
832 {
833 tree type = chrec_type (to_add);
834 tree res = NULL_TREE;
835
836 if (to_add == NULL_TREE)
837 return chrec_before;
838
839 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
840 instantiated at this point. */
841 if (TREE_CODE (to_add) == POLYNOMIAL_CHREC)
842 /* This should not happen. */
843 return chrec_dont_know;
844
845 if (dump_file && (dump_flags & TDF_DETAILS))
846 {
847 fprintf (dump_file, "(add_to_evolution \n");
848 fprintf (dump_file, " (loop_nb = %d)\n", loop_nb);
849 fprintf (dump_file, " (chrec_before = ");
850 print_generic_expr (dump_file, chrec_before, 0);
851 fprintf (dump_file, ")\n (to_add = ");
852 print_generic_expr (dump_file, to_add, 0);
853 fprintf (dump_file, ")\n");
854 }
855
856 if (code == MINUS_EXPR)
857 to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type)
858 ? build_real (type, dconstm1)
859 : build_int_cst_type (type, -1));
860
861 res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt);
862
863 if (dump_file && (dump_flags & TDF_DETAILS))
864 {
865 fprintf (dump_file, " (res = ");
866 print_generic_expr (dump_file, res, 0);
867 fprintf (dump_file, "))\n");
868 }
869
870 return res;
871 }
872
873 /* Helper function. */
874
875 static inline tree
876 set_nb_iterations_in_loop (struct loop *loop,
877 tree res)
878 {
879 if (dump_file && (dump_flags & TDF_DETAILS))
880 {
881 fprintf (dump_file, " (set_nb_iterations_in_loop = ");
882 print_generic_expr (dump_file, res, 0);
883 fprintf (dump_file, "))\n");
884 }
885
886 loop->nb_iterations = res;
887 return res;
888 }
889
890 \f
891
892 /* This section selects the loops that will be good candidates for the
893 scalar evolution analysis. For the moment, greedily select all the
894 loop nests we could analyze. */
895
896 /* Return true when it is possible to analyze the condition expression
897 EXPR. */
898
899 static bool
900 analyzable_condition (tree expr)
901 {
902 tree condition;
903
904 if (TREE_CODE (expr) != COND_EXPR)
905 return false;
906
907 condition = TREE_OPERAND (expr, 0);
908
909 switch (TREE_CODE (condition))
910 {
911 case SSA_NAME:
912 return true;
913
914 case LT_EXPR:
915 case LE_EXPR:
916 case GT_EXPR:
917 case GE_EXPR:
918 case EQ_EXPR:
919 case NE_EXPR:
920 return true;
921
922 default:
923 return false;
924 }
925
926 return false;
927 }
928
929 /* For a loop with a single exit edge, return the COND_EXPR that
930 guards the exit edge. If the expression is too difficult to
931 analyze, then give up. */
932
933 tree
934 get_loop_exit_condition (struct loop *loop)
935 {
936 tree res = NULL_TREE;
937 edge exit_edge = single_exit (loop);
938
939 if (dump_file && (dump_flags & TDF_DETAILS))
940 fprintf (dump_file, "(get_loop_exit_condition \n ");
941
942 if (exit_edge)
943 {
944 tree expr;
945
946 expr = last_stmt (exit_edge->src);
947 if (analyzable_condition (expr))
948 res = expr;
949 }
950
951 if (dump_file && (dump_flags & TDF_DETAILS))
952 {
953 print_generic_expr (dump_file, res, 0);
954 fprintf (dump_file, ")\n");
955 }
956
957 return res;
958 }
959
960 /* Recursively determine and enqueue the exit conditions for a loop. */
961
962 static void
963 get_exit_conditions_rec (struct loop *loop,
964 VEC(tree,heap) **exit_conditions)
965 {
966 if (!loop)
967 return;
968
969 /* Recurse on the inner loops, then on the next (sibling) loops. */
970 get_exit_conditions_rec (loop->inner, exit_conditions);
971 get_exit_conditions_rec (loop->next, exit_conditions);
972
973 if (single_exit (loop))
974 {
975 tree loop_condition = get_loop_exit_condition (loop);
976
977 if (loop_condition)
978 VEC_safe_push (tree, heap, *exit_conditions, loop_condition);
979 }
980 }
981
982 /* Select the candidate loop nests for the analysis. This function
983 initializes the EXIT_CONDITIONS array. */
984
985 static void
986 select_loops_exit_conditions (VEC(tree,heap) **exit_conditions)
987 {
988 struct loop *function_body = current_loops->tree_root;
989
990 get_exit_conditions_rec (function_body->inner, exit_conditions);
991 }
992
993 \f
994 /* Depth first search algorithm. */
995
996 typedef enum t_bool {
997 t_false,
998 t_true,
999 t_dont_know
1000 } t_bool;
1001
1002
1003 static t_bool follow_ssa_edge (struct loop *loop, tree, tree, tree *, int);
1004
1005 /* Follow the ssa edge into the right hand side RHS of an assignment.
1006 Return true if the strongly connected component has been found. */
1007
1008 static t_bool
1009 follow_ssa_edge_in_rhs (struct loop *loop, tree at_stmt, tree rhs,
1010 tree halting_phi, tree *evolution_of_loop, int limit)
1011 {
1012 t_bool res = t_false;
1013 tree rhs0, rhs1;
1014 tree type_rhs = TREE_TYPE (rhs);
1015 tree evol;
1016 enum tree_code code;
1017
1018 /* The RHS is one of the following cases:
1019 - an SSA_NAME,
1020 - an INTEGER_CST,
1021 - a PLUS_EXPR,
1022 - a POINTER_PLUS_EXPR,
1023 - a MINUS_EXPR,
1024 - an ASSERT_EXPR,
1025 - other cases are not yet handled. */
1026 code = TREE_CODE (rhs);
1027 switch (code)
1028 {
1029 case NOP_EXPR:
1030 /* This assignment is under the form "a_1 = (cast) rhs. */
1031 res = follow_ssa_edge_in_rhs (loop, at_stmt, TREE_OPERAND (rhs, 0),
1032 halting_phi, evolution_of_loop, limit);
1033 *evolution_of_loop = chrec_convert (TREE_TYPE (rhs),
1034 *evolution_of_loop, at_stmt);
1035 break;
1036
1037 case INTEGER_CST:
1038 /* This assignment is under the form "a_1 = 7". */
1039 res = t_false;
1040 break;
1041
1042 case SSA_NAME:
1043 /* This assignment is under the form: "a_1 = b_2". */
1044 res = follow_ssa_edge
1045 (loop, SSA_NAME_DEF_STMT (rhs), halting_phi, evolution_of_loop, limit);
1046 break;
1047
1048 case POINTER_PLUS_EXPR:
1049 case PLUS_EXPR:
1050 /* This case is under the form "rhs0 + rhs1". */
1051 rhs0 = TREE_OPERAND (rhs, 0);
1052 rhs1 = TREE_OPERAND (rhs, 1);
1053 STRIP_TYPE_NOPS (rhs0);
1054 STRIP_TYPE_NOPS (rhs1);
1055
1056 if (TREE_CODE (rhs0) == SSA_NAME)
1057 {
1058 if (TREE_CODE (rhs1) == SSA_NAME)
1059 {
1060 /* Match an assignment under the form:
1061 "a = b + c". */
1062
1063 /* We want only assignments of form "name + name" contribute to
1064 LIMIT, as the other cases do not necessarily contribute to
1065 the complexity of the expression. */
1066 limit++;
1067
1068 evol = *evolution_of_loop;
1069 res = follow_ssa_edge
1070 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1071 &evol, limit);
1072
1073 if (res == t_true)
1074 *evolution_of_loop = add_to_evolution
1075 (loop->num,
1076 chrec_convert (type_rhs, evol, at_stmt),
1077 code, rhs1, at_stmt);
1078
1079 else if (res == t_false)
1080 {
1081 res = follow_ssa_edge
1082 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1083 evolution_of_loop, limit);
1084
1085 if (res == t_true)
1086 *evolution_of_loop = add_to_evolution
1087 (loop->num,
1088 chrec_convert (type_rhs, *evolution_of_loop, at_stmt),
1089 code, rhs0, at_stmt);
1090
1091 else if (res == t_dont_know)
1092 *evolution_of_loop = chrec_dont_know;
1093 }
1094
1095 else if (res == t_dont_know)
1096 *evolution_of_loop = chrec_dont_know;
1097 }
1098
1099 else
1100 {
1101 /* Match an assignment under the form:
1102 "a = b + ...". */
1103 res = follow_ssa_edge
1104 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1105 evolution_of_loop, limit);
1106 if (res == t_true)
1107 *evolution_of_loop = add_to_evolution
1108 (loop->num, chrec_convert (type_rhs, *evolution_of_loop,
1109 at_stmt),
1110 code, rhs1, at_stmt);
1111
1112 else if (res == t_dont_know)
1113 *evolution_of_loop = chrec_dont_know;
1114 }
1115 }
1116
1117 else if (TREE_CODE (rhs1) == SSA_NAME)
1118 {
1119 /* Match an assignment under the form:
1120 "a = ... + c". */
1121 res = follow_ssa_edge
1122 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1123 evolution_of_loop, limit);
1124 if (res == t_true)
1125 *evolution_of_loop = add_to_evolution
1126 (loop->num, chrec_convert (type_rhs, *evolution_of_loop,
1127 at_stmt),
1128 code, rhs0, at_stmt);
1129
1130 else if (res == t_dont_know)
1131 *evolution_of_loop = chrec_dont_know;
1132 }
1133
1134 else
1135 /* Otherwise, match an assignment under the form:
1136 "a = ... + ...". */
1137 /* And there is nothing to do. */
1138 res = t_false;
1139
1140 break;
1141
1142 case MINUS_EXPR:
1143 /* This case is under the form "opnd0 = rhs0 - rhs1". */
1144 rhs0 = TREE_OPERAND (rhs, 0);
1145 rhs1 = TREE_OPERAND (rhs, 1);
1146 STRIP_TYPE_NOPS (rhs0);
1147 STRIP_TYPE_NOPS (rhs1);
1148
1149 if (TREE_CODE (rhs0) == SSA_NAME)
1150 {
1151 /* Match an assignment under the form:
1152 "a = b - ...". */
1153
1154 /* We want only assignments of form "name - name" contribute to
1155 LIMIT, as the other cases do not necessarily contribute to
1156 the complexity of the expression. */
1157 if (TREE_CODE (rhs1) == SSA_NAME)
1158 limit++;
1159
1160 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1161 evolution_of_loop, limit);
1162 if (res == t_true)
1163 *evolution_of_loop = add_to_evolution
1164 (loop->num, chrec_convert (type_rhs, *evolution_of_loop, at_stmt),
1165 MINUS_EXPR, rhs1, at_stmt);
1166
1167 else if (res == t_dont_know)
1168 *evolution_of_loop = chrec_dont_know;
1169 }
1170 else
1171 /* Otherwise, match an assignment under the form:
1172 "a = ... - ...". */
1173 /* And there is nothing to do. */
1174 res = t_false;
1175
1176 break;
1177
1178 case ASSERT_EXPR:
1179 {
1180 /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>"
1181 It must be handled as a copy assignment of the form a_1 = a_2. */
1182 tree op0 = ASSERT_EXPR_VAR (rhs);
1183 if (TREE_CODE (op0) == SSA_NAME)
1184 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (op0),
1185 halting_phi, evolution_of_loop, limit);
1186 else
1187 res = t_false;
1188 break;
1189 }
1190
1191
1192 default:
1193 res = t_false;
1194 break;
1195 }
1196
1197 return res;
1198 }
1199
1200 /* Checks whether the I-th argument of a PHI comes from a backedge. */
1201
1202 static bool
1203 backedge_phi_arg_p (tree phi, int i)
1204 {
1205 edge e = PHI_ARG_EDGE (phi, i);
1206
1207 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1208 about updating it anywhere, and this should work as well most of the
1209 time. */
1210 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1211 return true;
1212
1213 return false;
1214 }
1215
1216 /* Helper function for one branch of the condition-phi-node. Return
1217 true if the strongly connected component has been found following
1218 this path. */
1219
1220 static inline t_bool
1221 follow_ssa_edge_in_condition_phi_branch (int i,
1222 struct loop *loop,
1223 tree condition_phi,
1224 tree halting_phi,
1225 tree *evolution_of_branch,
1226 tree init_cond, int limit)
1227 {
1228 tree branch = PHI_ARG_DEF (condition_phi, i);
1229 *evolution_of_branch = chrec_dont_know;
1230
1231 /* Do not follow back edges (they must belong to an irreducible loop, which
1232 we really do not want to worry about). */
1233 if (backedge_phi_arg_p (condition_phi, i))
1234 return t_false;
1235
1236 if (TREE_CODE (branch) == SSA_NAME)
1237 {
1238 *evolution_of_branch = init_cond;
1239 return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi,
1240 evolution_of_branch, limit);
1241 }
1242
1243 /* This case occurs when one of the condition branches sets
1244 the variable to a constant: i.e. a phi-node like
1245 "a_2 = PHI <a_7(5), 2(6)>;".
1246
1247 FIXME: This case have to be refined correctly:
1248 in some cases it is possible to say something better than
1249 chrec_dont_know, for example using a wrap-around notation. */
1250 return t_false;
1251 }
1252
1253 /* This function merges the branches of a condition-phi-node in a
1254 loop. */
1255
1256 static t_bool
1257 follow_ssa_edge_in_condition_phi (struct loop *loop,
1258 tree condition_phi,
1259 tree halting_phi,
1260 tree *evolution_of_loop, int limit)
1261 {
1262 int i;
1263 tree init = *evolution_of_loop;
1264 tree evolution_of_branch;
1265 t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi,
1266 halting_phi,
1267 &evolution_of_branch,
1268 init, limit);
1269 if (res == t_false || res == t_dont_know)
1270 return res;
1271
1272 *evolution_of_loop = evolution_of_branch;
1273
1274 /* If the phi node is just a copy, do not increase the limit. */
1275 if (PHI_NUM_ARGS (condition_phi) > 1)
1276 limit++;
1277
1278 for (i = 1; i < PHI_NUM_ARGS (condition_phi); i++)
1279 {
1280 /* Quickly give up when the evolution of one of the branches is
1281 not known. */
1282 if (*evolution_of_loop == chrec_dont_know)
1283 return t_true;
1284
1285 res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi,
1286 halting_phi,
1287 &evolution_of_branch,
1288 init, limit);
1289 if (res == t_false || res == t_dont_know)
1290 return res;
1291
1292 *evolution_of_loop = chrec_merge (*evolution_of_loop,
1293 evolution_of_branch);
1294 }
1295
1296 return t_true;
1297 }
1298
1299 /* Follow an SSA edge in an inner loop. It computes the overall
1300 effect of the loop, and following the symbolic initial conditions,
1301 it follows the edges in the parent loop. The inner loop is
1302 considered as a single statement. */
1303
1304 static t_bool
1305 follow_ssa_edge_inner_loop_phi (struct loop *outer_loop,
1306 tree loop_phi_node,
1307 tree halting_phi,
1308 tree *evolution_of_loop, int limit)
1309 {
1310 struct loop *loop = loop_containing_stmt (loop_phi_node);
1311 tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node));
1312
1313 /* Sometimes, the inner loop is too difficult to analyze, and the
1314 result of the analysis is a symbolic parameter. */
1315 if (ev == PHI_RESULT (loop_phi_node))
1316 {
1317 t_bool res = t_false;
1318 int i;
1319
1320 for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++)
1321 {
1322 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1323 basic_block bb;
1324
1325 /* Follow the edges that exit the inner loop. */
1326 bb = PHI_ARG_EDGE (loop_phi_node, i)->src;
1327 if (!flow_bb_inside_loop_p (loop, bb))
1328 res = follow_ssa_edge_in_rhs (outer_loop, loop_phi_node,
1329 arg, halting_phi,
1330 evolution_of_loop, limit);
1331 if (res == t_true)
1332 break;
1333 }
1334
1335 /* If the path crosses this loop-phi, give up. */
1336 if (res == t_true)
1337 *evolution_of_loop = chrec_dont_know;
1338
1339 return res;
1340 }
1341
1342 /* Otherwise, compute the overall effect of the inner loop. */
1343 ev = compute_overall_effect_of_inner_loop (loop, ev);
1344 return follow_ssa_edge_in_rhs (outer_loop, loop_phi_node, ev, halting_phi,
1345 evolution_of_loop, limit);
1346 }
1347
1348 /* Follow an SSA edge from a loop-phi-node to itself, constructing a
1349 path that is analyzed on the return walk. */
1350
1351 static t_bool
1352 follow_ssa_edge (struct loop *loop, tree def, tree halting_phi,
1353 tree *evolution_of_loop, int limit)
1354 {
1355 struct loop *def_loop;
1356
1357 if (TREE_CODE (def) == NOP_EXPR)
1358 return t_false;
1359
1360 /* Give up if the path is longer than the MAX that we allow. */
1361 if (limit > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
1362 return t_dont_know;
1363
1364 def_loop = loop_containing_stmt (def);
1365
1366 switch (TREE_CODE (def))
1367 {
1368 case PHI_NODE:
1369 if (!loop_phi_node_p (def))
1370 /* DEF is a condition-phi-node. Follow the branches, and
1371 record their evolutions. Finally, merge the collected
1372 information and set the approximation to the main
1373 variable. */
1374 return follow_ssa_edge_in_condition_phi
1375 (loop, def, halting_phi, evolution_of_loop, limit);
1376
1377 /* When the analyzed phi is the halting_phi, the
1378 depth-first search is over: we have found a path from
1379 the halting_phi to itself in the loop. */
1380 if (def == halting_phi)
1381 return t_true;
1382
1383 /* Otherwise, the evolution of the HALTING_PHI depends
1384 on the evolution of another loop-phi-node, i.e. the
1385 evolution function is a higher degree polynomial. */
1386 if (def_loop == loop)
1387 return t_false;
1388
1389 /* Inner loop. */
1390 if (flow_loop_nested_p (loop, def_loop))
1391 return follow_ssa_edge_inner_loop_phi
1392 (loop, def, halting_phi, evolution_of_loop, limit + 1);
1393
1394 /* Outer loop. */
1395 return t_false;
1396
1397 case GIMPLE_MODIFY_STMT:
1398 return follow_ssa_edge_in_rhs (loop, def,
1399 GIMPLE_STMT_OPERAND (def, 1),
1400 halting_phi,
1401 evolution_of_loop, limit);
1402
1403 default:
1404 /* At this level of abstraction, the program is just a set
1405 of GIMPLE_MODIFY_STMTs and PHI_NODEs. In principle there is no
1406 other node to be handled. */
1407 return t_false;
1408 }
1409 }
1410
1411 \f
1412
1413 /* Given a LOOP_PHI_NODE, this function determines the evolution
1414 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1415
1416 static tree
1417 analyze_evolution_in_loop (tree loop_phi_node,
1418 tree init_cond)
1419 {
1420 int i;
1421 tree evolution_function = chrec_not_analyzed_yet;
1422 struct loop *loop = loop_containing_stmt (loop_phi_node);
1423 basic_block bb;
1424
1425 if (dump_file && (dump_flags & TDF_DETAILS))
1426 {
1427 fprintf (dump_file, "(analyze_evolution_in_loop \n");
1428 fprintf (dump_file, " (loop_phi_node = ");
1429 print_generic_expr (dump_file, loop_phi_node, 0);
1430 fprintf (dump_file, ")\n");
1431 }
1432
1433 for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++)
1434 {
1435 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1436 tree ssa_chain, ev_fn;
1437 t_bool res;
1438
1439 /* Select the edges that enter the loop body. */
1440 bb = PHI_ARG_EDGE (loop_phi_node, i)->src;
1441 if (!flow_bb_inside_loop_p (loop, bb))
1442 continue;
1443
1444 if (TREE_CODE (arg) == SSA_NAME)
1445 {
1446 ssa_chain = SSA_NAME_DEF_STMT (arg);
1447
1448 /* Pass in the initial condition to the follow edge function. */
1449 ev_fn = init_cond;
1450 res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0);
1451 }
1452 else
1453 res = t_false;
1454
1455 /* When it is impossible to go back on the same
1456 loop_phi_node by following the ssa edges, the
1457 evolution is represented by a peeled chrec, i.e. the
1458 first iteration, EV_FN has the value INIT_COND, then
1459 all the other iterations it has the value of ARG.
1460 For the moment, PEELED_CHREC nodes are not built. */
1461 if (res != t_true)
1462 ev_fn = chrec_dont_know;
1463
1464 /* When there are multiple back edges of the loop (which in fact never
1465 happens currently, but nevertheless), merge their evolutions. */
1466 evolution_function = chrec_merge (evolution_function, ev_fn);
1467 }
1468
1469 if (dump_file && (dump_flags & TDF_DETAILS))
1470 {
1471 fprintf (dump_file, " (evolution_function = ");
1472 print_generic_expr (dump_file, evolution_function, 0);
1473 fprintf (dump_file, "))\n");
1474 }
1475
1476 return evolution_function;
1477 }
1478
1479 /* Given a loop-phi-node, return the initial conditions of the
1480 variable on entry of the loop. When the CCP has propagated
1481 constants into the loop-phi-node, the initial condition is
1482 instantiated, otherwise the initial condition is kept symbolic.
1483 This analyzer does not analyze the evolution outside the current
1484 loop, and leaves this task to the on-demand tree reconstructor. */
1485
1486 static tree
1487 analyze_initial_condition (tree loop_phi_node)
1488 {
1489 int i;
1490 tree init_cond = chrec_not_analyzed_yet;
1491 struct loop *loop = bb_for_stmt (loop_phi_node)->loop_father;
1492
1493 if (dump_file && (dump_flags & TDF_DETAILS))
1494 {
1495 fprintf (dump_file, "(analyze_initial_condition \n");
1496 fprintf (dump_file, " (loop_phi_node = \n");
1497 print_generic_expr (dump_file, loop_phi_node, 0);
1498 fprintf (dump_file, ")\n");
1499 }
1500
1501 for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++)
1502 {
1503 tree branch = PHI_ARG_DEF (loop_phi_node, i);
1504 basic_block bb = PHI_ARG_EDGE (loop_phi_node, i)->src;
1505
1506 /* When the branch is oriented to the loop's body, it does
1507 not contribute to the initial condition. */
1508 if (flow_bb_inside_loop_p (loop, bb))
1509 continue;
1510
1511 if (init_cond == chrec_not_analyzed_yet)
1512 {
1513 init_cond = branch;
1514 continue;
1515 }
1516
1517 if (TREE_CODE (branch) == SSA_NAME)
1518 {
1519 init_cond = chrec_dont_know;
1520 break;
1521 }
1522
1523 init_cond = chrec_merge (init_cond, branch);
1524 }
1525
1526 /* Ooops -- a loop without an entry??? */
1527 if (init_cond == chrec_not_analyzed_yet)
1528 init_cond = chrec_dont_know;
1529
1530 if (dump_file && (dump_flags & TDF_DETAILS))
1531 {
1532 fprintf (dump_file, " (init_cond = ");
1533 print_generic_expr (dump_file, init_cond, 0);
1534 fprintf (dump_file, "))\n");
1535 }
1536
1537 return init_cond;
1538 }
1539
1540 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1541
1542 static tree
1543 interpret_loop_phi (struct loop *loop, tree loop_phi_node)
1544 {
1545 tree res;
1546 struct loop *phi_loop = loop_containing_stmt (loop_phi_node);
1547 tree init_cond;
1548
1549 if (phi_loop != loop)
1550 {
1551 struct loop *subloop;
1552 tree evolution_fn = analyze_scalar_evolution
1553 (phi_loop, PHI_RESULT (loop_phi_node));
1554
1555 /* Dive one level deeper. */
1556 subloop = superloop_at_depth (phi_loop, loop_depth (loop) + 1);
1557
1558 /* Interpret the subloop. */
1559 res = compute_overall_effect_of_inner_loop (subloop, evolution_fn);
1560 return res;
1561 }
1562
1563 /* Otherwise really interpret the loop phi. */
1564 init_cond = analyze_initial_condition (loop_phi_node);
1565 res = analyze_evolution_in_loop (loop_phi_node, init_cond);
1566
1567 return res;
1568 }
1569
1570 /* This function merges the branches of a condition-phi-node,
1571 contained in the outermost loop, and whose arguments are already
1572 analyzed. */
1573
1574 static tree
1575 interpret_condition_phi (struct loop *loop, tree condition_phi)
1576 {
1577 int i;
1578 tree res = chrec_not_analyzed_yet;
1579
1580 for (i = 0; i < PHI_NUM_ARGS (condition_phi); i++)
1581 {
1582 tree branch_chrec;
1583
1584 if (backedge_phi_arg_p (condition_phi, i))
1585 {
1586 res = chrec_dont_know;
1587 break;
1588 }
1589
1590 branch_chrec = analyze_scalar_evolution
1591 (loop, PHI_ARG_DEF (condition_phi, i));
1592
1593 res = chrec_merge (res, branch_chrec);
1594 }
1595
1596 return res;
1597 }
1598
1599 /* Interpret the right hand side of a GIMPLE_MODIFY_STMT OPND1. If we didn't
1600 analyze this node before, follow the definitions until ending
1601 either on an analyzed GIMPLE_MODIFY_STMT, or on a loop-phi-node. On the
1602 return path, this function propagates evolutions (ala constant copy
1603 propagation). OPND1 is not a GIMPLE expression because we could
1604 analyze the effect of an inner loop: see interpret_loop_phi. */
1605
1606 static tree
1607 interpret_rhs_modify_stmt (struct loop *loop, tree at_stmt,
1608 tree opnd1, tree type)
1609 {
1610 tree res, opnd10, opnd11, chrec10, chrec11;
1611
1612 if (is_gimple_min_invariant (opnd1))
1613 return chrec_convert (type, opnd1, at_stmt);
1614
1615 switch (TREE_CODE (opnd1))
1616 {
1617 case POINTER_PLUS_EXPR:
1618 opnd10 = TREE_OPERAND (opnd1, 0);
1619 opnd11 = TREE_OPERAND (opnd1, 1);
1620 chrec10 = analyze_scalar_evolution (loop, opnd10);
1621 chrec11 = analyze_scalar_evolution (loop, opnd11);
1622 chrec10 = chrec_convert (type, chrec10, at_stmt);
1623 chrec11 = chrec_convert (sizetype, chrec11, at_stmt);
1624 res = chrec_fold_plus (type, chrec10, chrec11);
1625 break;
1626
1627 case PLUS_EXPR:
1628 opnd10 = TREE_OPERAND (opnd1, 0);
1629 opnd11 = TREE_OPERAND (opnd1, 1);
1630 chrec10 = analyze_scalar_evolution (loop, opnd10);
1631 chrec11 = analyze_scalar_evolution (loop, opnd11);
1632 chrec10 = chrec_convert (type, chrec10, at_stmt);
1633 chrec11 = chrec_convert (type, chrec11, at_stmt);
1634 res = chrec_fold_plus (type, chrec10, chrec11);
1635 break;
1636
1637 case MINUS_EXPR:
1638 opnd10 = TREE_OPERAND (opnd1, 0);
1639 opnd11 = TREE_OPERAND (opnd1, 1);
1640 chrec10 = analyze_scalar_evolution (loop, opnd10);
1641 chrec11 = analyze_scalar_evolution (loop, opnd11);
1642 chrec10 = chrec_convert (type, chrec10, at_stmt);
1643 chrec11 = chrec_convert (type, chrec11, at_stmt);
1644 res = chrec_fold_minus (type, chrec10, chrec11);
1645 break;
1646
1647 case NEGATE_EXPR:
1648 opnd10 = TREE_OPERAND (opnd1, 0);
1649 chrec10 = analyze_scalar_evolution (loop, opnd10);
1650 chrec10 = chrec_convert (type, chrec10, at_stmt);
1651 /* TYPE may be integer, real or complex, so use fold_convert. */
1652 res = chrec_fold_multiply (type, chrec10,
1653 fold_convert (type, integer_minus_one_node));
1654 break;
1655
1656 case MULT_EXPR:
1657 opnd10 = TREE_OPERAND (opnd1, 0);
1658 opnd11 = TREE_OPERAND (opnd1, 1);
1659 chrec10 = analyze_scalar_evolution (loop, opnd10);
1660 chrec11 = analyze_scalar_evolution (loop, opnd11);
1661 chrec10 = chrec_convert (type, chrec10, at_stmt);
1662 chrec11 = chrec_convert (type, chrec11, at_stmt);
1663 res = chrec_fold_multiply (type, chrec10, chrec11);
1664 break;
1665
1666 case SSA_NAME:
1667 res = chrec_convert (type, analyze_scalar_evolution (loop, opnd1),
1668 at_stmt);
1669 break;
1670
1671 case ASSERT_EXPR:
1672 opnd10 = ASSERT_EXPR_VAR (opnd1);
1673 res = chrec_convert (type, analyze_scalar_evolution (loop, opnd10),
1674 at_stmt);
1675 break;
1676
1677 case NOP_EXPR:
1678 case CONVERT_EXPR:
1679 opnd10 = TREE_OPERAND (opnd1, 0);
1680 chrec10 = analyze_scalar_evolution (loop, opnd10);
1681 res = chrec_convert (type, chrec10, at_stmt);
1682 break;
1683
1684 default:
1685 res = chrec_dont_know;
1686 break;
1687 }
1688
1689 return res;
1690 }
1691
1692 \f
1693
1694 /* This section contains all the entry points:
1695 - number_of_iterations_in_loop,
1696 - analyze_scalar_evolution,
1697 - instantiate_parameters.
1698 */
1699
1700 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1701 common ancestor of DEF_LOOP and USE_LOOP. */
1702
1703 static tree
1704 compute_scalar_evolution_in_loop (struct loop *wrto_loop,
1705 struct loop *def_loop,
1706 tree ev)
1707 {
1708 tree res;
1709 if (def_loop == wrto_loop)
1710 return ev;
1711
1712 def_loop = superloop_at_depth (def_loop, loop_depth (wrto_loop) + 1);
1713 res = compute_overall_effect_of_inner_loop (def_loop, ev);
1714
1715 return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet);
1716 }
1717
1718 /* Helper recursive function. */
1719
1720 static tree
1721 analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res)
1722 {
1723 tree def, type = TREE_TYPE (var);
1724 basic_block bb;
1725 struct loop *def_loop;
1726
1727 if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE)
1728 return chrec_dont_know;
1729
1730 if (TREE_CODE (var) != SSA_NAME)
1731 return interpret_rhs_modify_stmt (loop, NULL_TREE, var, type);
1732
1733 def = SSA_NAME_DEF_STMT (var);
1734 bb = bb_for_stmt (def);
1735 def_loop = bb ? bb->loop_father : NULL;
1736
1737 if (bb == NULL
1738 || !flow_bb_inside_loop_p (loop, bb))
1739 {
1740 /* Keep the symbolic form. */
1741 res = var;
1742 goto set_and_end;
1743 }
1744
1745 if (res != chrec_not_analyzed_yet)
1746 {
1747 if (loop != bb->loop_father)
1748 res = compute_scalar_evolution_in_loop
1749 (find_common_loop (loop, bb->loop_father), bb->loop_father, res);
1750
1751 goto set_and_end;
1752 }
1753
1754 if (loop != def_loop)
1755 {
1756 res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet);
1757 res = compute_scalar_evolution_in_loop (loop, def_loop, res);
1758
1759 goto set_and_end;
1760 }
1761
1762 switch (TREE_CODE (def))
1763 {
1764 case GIMPLE_MODIFY_STMT:
1765 res = interpret_rhs_modify_stmt (loop, def,
1766 GIMPLE_STMT_OPERAND (def, 1), type);
1767 break;
1768
1769 case PHI_NODE:
1770 if (loop_phi_node_p (def))
1771 res = interpret_loop_phi (loop, def);
1772 else
1773 res = interpret_condition_phi (loop, def);
1774 break;
1775
1776 default:
1777 res = chrec_dont_know;
1778 break;
1779 }
1780
1781 set_and_end:
1782
1783 /* Keep the symbolic form. */
1784 if (res == chrec_dont_know)
1785 res = var;
1786
1787 if (loop == def_loop)
1788 set_scalar_evolution (var, res);
1789
1790 return res;
1791 }
1792
1793 /* Entry point for the scalar evolution analyzer.
1794 Analyzes and returns the scalar evolution of the ssa_name VAR.
1795 LOOP_NB is the identifier number of the loop in which the variable
1796 is used.
1797
1798 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
1799 pointer to the statement that uses this variable, in order to
1800 determine the evolution function of the variable, use the following
1801 calls:
1802
1803 unsigned loop_nb = loop_containing_stmt (stmt)->num;
1804 tree chrec_with_symbols = analyze_scalar_evolution (loop_nb, var);
1805 tree chrec_instantiated = instantiate_parameters
1806 (loop_nb, chrec_with_symbols);
1807 */
1808
1809 tree
1810 analyze_scalar_evolution (struct loop *loop, tree var)
1811 {
1812 tree res;
1813
1814 if (dump_file && (dump_flags & TDF_DETAILS))
1815 {
1816 fprintf (dump_file, "(analyze_scalar_evolution \n");
1817 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
1818 fprintf (dump_file, " (scalar = ");
1819 print_generic_expr (dump_file, var, 0);
1820 fprintf (dump_file, ")\n");
1821 }
1822
1823 res = analyze_scalar_evolution_1 (loop, var, get_scalar_evolution (var));
1824
1825 if (TREE_CODE (var) == SSA_NAME && res == chrec_dont_know)
1826 res = var;
1827
1828 if (dump_file && (dump_flags & TDF_DETAILS))
1829 fprintf (dump_file, ")\n");
1830
1831 return res;
1832 }
1833
1834 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
1835 WRTO_LOOP (which should be a superloop of both USE_LOOP and definition
1836 of VERSION).
1837
1838 FOLDED_CASTS is set to true if resolve_mixers used
1839 chrec_convert_aggressive (TODO -- not really, we are way too conservative
1840 at the moment in order to keep things simple). */
1841
1842 static tree
1843 analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop,
1844 tree version, bool *folded_casts)
1845 {
1846 bool val = false;
1847 tree ev = version, tmp;
1848
1849 if (folded_casts)
1850 *folded_casts = false;
1851 while (1)
1852 {
1853 tmp = analyze_scalar_evolution (use_loop, ev);
1854 ev = resolve_mixers (use_loop, tmp);
1855
1856 if (folded_casts && tmp != ev)
1857 *folded_casts = true;
1858
1859 if (use_loop == wrto_loop)
1860 return ev;
1861
1862 /* If the value of the use changes in the inner loop, we cannot express
1863 its value in the outer loop (we might try to return interval chrec,
1864 but we do not have a user for it anyway) */
1865 if (!no_evolution_in_loop_p (ev, use_loop->num, &val)
1866 || !val)
1867 return chrec_dont_know;
1868
1869 use_loop = loop_outer (use_loop);
1870 }
1871 }
1872
1873 /* Returns instantiated value for VERSION in CACHE. */
1874
1875 static tree
1876 get_instantiated_value (htab_t cache, tree version)
1877 {
1878 struct scev_info_str *info, pattern;
1879
1880 pattern.var = version;
1881 info = (struct scev_info_str *) htab_find (cache, &pattern);
1882
1883 if (info)
1884 return info->chrec;
1885 else
1886 return NULL_TREE;
1887 }
1888
1889 /* Sets instantiated value for VERSION to VAL in CACHE. */
1890
1891 static void
1892 set_instantiated_value (htab_t cache, tree version, tree val)
1893 {
1894 struct scev_info_str *info, pattern;
1895 PTR *slot;
1896
1897 pattern.var = version;
1898 slot = htab_find_slot (cache, &pattern, INSERT);
1899
1900 if (!*slot)
1901 *slot = new_scev_info_str (version);
1902 info = (struct scev_info_str *) *slot;
1903 info->chrec = val;
1904 }
1905
1906 /* Return the closed_loop_phi node for VAR. If there is none, return
1907 NULL_TREE. */
1908
1909 static tree
1910 loop_closed_phi_def (tree var)
1911 {
1912 struct loop *loop;
1913 edge exit;
1914 tree phi;
1915
1916 if (var == NULL_TREE
1917 || TREE_CODE (var) != SSA_NAME)
1918 return NULL_TREE;
1919
1920 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var));
1921 exit = single_exit (loop);
1922 if (!exit)
1923 return NULL_TREE;
1924
1925 for (phi = phi_nodes (exit->dest); phi; phi = PHI_CHAIN (phi))
1926 if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var)
1927 return PHI_RESULT (phi);
1928
1929 return NULL_TREE;
1930 }
1931
1932 /* Analyze all the parameters of the chrec that were left under a symbolic form,
1933 with respect to LOOP. CHREC is the chrec to instantiate. CACHE is the cache
1934 of already instantiated values. FLAGS modify the way chrecs are
1935 instantiated. SIZE_EXPR is used for computing the size of the expression to
1936 be instantiated, and to stop if it exceeds some limit. */
1937
1938 /* Values for FLAGS. */
1939 enum
1940 {
1941 INSERT_SUPERLOOP_CHRECS = 1, /* Loop invariants are replaced with chrecs
1942 in outer loops. */
1943 FOLD_CONVERSIONS = 2 /* The conversions that may wrap in
1944 signed/pointer type are folded, as long as the
1945 value of the chrec is preserved. */
1946 };
1947
1948 static tree
1949 instantiate_parameters_1 (struct loop *loop, tree chrec, int flags, htab_t cache,
1950 int size_expr)
1951 {
1952 tree res, op0, op1, op2;
1953 basic_block def_bb;
1954 struct loop *def_loop;
1955 tree type = chrec_type (chrec);
1956
1957 /* Give up if the expression is larger than the MAX that we allow. */
1958 if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
1959 return chrec_dont_know;
1960
1961 if (automatically_generated_chrec_p (chrec)
1962 || is_gimple_min_invariant (chrec))
1963 return chrec;
1964
1965 switch (TREE_CODE (chrec))
1966 {
1967 case SSA_NAME:
1968 def_bb = bb_for_stmt (SSA_NAME_DEF_STMT (chrec));
1969
1970 /* A parameter (or loop invariant and we do not want to include
1971 evolutions in outer loops), nothing to do. */
1972 if (!def_bb
1973 || (!(flags & INSERT_SUPERLOOP_CHRECS)
1974 && !flow_bb_inside_loop_p (loop, def_bb)))
1975 return chrec;
1976
1977 /* We cache the value of instantiated variable to avoid exponential
1978 time complexity due to reevaluations. We also store the convenient
1979 value in the cache in order to prevent infinite recursion -- we do
1980 not want to instantiate the SSA_NAME if it is in a mixer
1981 structure. This is used for avoiding the instantiation of
1982 recursively defined functions, such as:
1983
1984 | a_2 -> {0, +, 1, +, a_2}_1 */
1985
1986 res = get_instantiated_value (cache, chrec);
1987 if (res)
1988 return res;
1989
1990 /* Store the convenient value for chrec in the structure. If it
1991 is defined outside of the loop, we may just leave it in symbolic
1992 form, otherwise we need to admit that we do not know its behavior
1993 inside the loop. */
1994 res = !flow_bb_inside_loop_p (loop, def_bb) ? chrec : chrec_dont_know;
1995 set_instantiated_value (cache, chrec, res);
1996
1997 /* To make things even more complicated, instantiate_parameters_1
1998 calls analyze_scalar_evolution that may call # of iterations
1999 analysis that may in turn call instantiate_parameters_1 again.
2000 To prevent the infinite recursion, keep also the bitmap of
2001 ssa names that are being instantiated globally. */
2002 if (bitmap_bit_p (already_instantiated, SSA_NAME_VERSION (chrec)))
2003 return res;
2004
2005 def_loop = find_common_loop (loop, def_bb->loop_father);
2006
2007 /* If the analysis yields a parametric chrec, instantiate the
2008 result again. */
2009 bitmap_set_bit (already_instantiated, SSA_NAME_VERSION (chrec));
2010 res = analyze_scalar_evolution (def_loop, chrec);
2011
2012 /* Don't instantiate loop-closed-ssa phi nodes. */
2013 if (TREE_CODE (res) == SSA_NAME
2014 && (loop_containing_stmt (SSA_NAME_DEF_STMT (res)) == NULL
2015 || (loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res)))
2016 > loop_depth (def_loop))))
2017 {
2018 if (res == chrec)
2019 res = loop_closed_phi_def (chrec);
2020 else
2021 res = chrec;
2022
2023 if (res == NULL_TREE)
2024 res = chrec_dont_know;
2025 }
2026
2027 else if (res != chrec_dont_know)
2028 res = instantiate_parameters_1 (loop, res, flags, cache, size_expr);
2029
2030 bitmap_clear_bit (already_instantiated, SSA_NAME_VERSION (chrec));
2031
2032 /* Store the correct value to the cache. */
2033 set_instantiated_value (cache, chrec, res);
2034 return res;
2035
2036 case POLYNOMIAL_CHREC:
2037 op0 = instantiate_parameters_1 (loop, CHREC_LEFT (chrec),
2038 flags, cache, size_expr);
2039 if (op0 == chrec_dont_know)
2040 return chrec_dont_know;
2041
2042 op1 = instantiate_parameters_1 (loop, CHREC_RIGHT (chrec),
2043 flags, cache, size_expr);
2044 if (op1 == chrec_dont_know)
2045 return chrec_dont_know;
2046
2047 if (CHREC_LEFT (chrec) != op0
2048 || CHREC_RIGHT (chrec) != op1)
2049 {
2050 op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL_TREE);
2051 chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1);
2052 }
2053 return chrec;
2054
2055 case POINTER_PLUS_EXPR:
2056 case PLUS_EXPR:
2057 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2058 flags, cache, size_expr);
2059 if (op0 == chrec_dont_know)
2060 return chrec_dont_know;
2061
2062 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2063 flags, cache, size_expr);
2064 if (op1 == chrec_dont_know)
2065 return chrec_dont_know;
2066
2067 if (TREE_OPERAND (chrec, 0) != op0
2068 || TREE_OPERAND (chrec, 1) != op1)
2069 {
2070 op0 = chrec_convert (type, op0, NULL_TREE);
2071 op1 = chrec_convert_rhs (type, op1, NULL_TREE);
2072 chrec = chrec_fold_plus (type, op0, op1);
2073 }
2074 return chrec;
2075
2076 case MINUS_EXPR:
2077 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2078 flags, cache, size_expr);
2079 if (op0 == chrec_dont_know)
2080 return chrec_dont_know;
2081
2082 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2083 flags, cache, size_expr);
2084 if (op1 == chrec_dont_know)
2085 return chrec_dont_know;
2086
2087 if (TREE_OPERAND (chrec, 0) != op0
2088 || TREE_OPERAND (chrec, 1) != op1)
2089 {
2090 op0 = chrec_convert (type, op0, NULL_TREE);
2091 op1 = chrec_convert (type, op1, NULL_TREE);
2092 chrec = chrec_fold_minus (type, op0, op1);
2093 }
2094 return chrec;
2095
2096 case MULT_EXPR:
2097 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2098 flags, cache, size_expr);
2099 if (op0 == chrec_dont_know)
2100 return chrec_dont_know;
2101
2102 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2103 flags, cache, size_expr);
2104 if (op1 == chrec_dont_know)
2105 return chrec_dont_know;
2106
2107 if (TREE_OPERAND (chrec, 0) != op0
2108 || TREE_OPERAND (chrec, 1) != op1)
2109 {
2110 op0 = chrec_convert (type, op0, NULL_TREE);
2111 op1 = chrec_convert (type, op1, NULL_TREE);
2112 chrec = chrec_fold_multiply (type, op0, op1);
2113 }
2114 return chrec;
2115
2116 case NOP_EXPR:
2117 case CONVERT_EXPR:
2118 case NON_LVALUE_EXPR:
2119 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2120 flags, cache, size_expr);
2121 if (op0 == chrec_dont_know)
2122 return chrec_dont_know;
2123
2124 if (flags & FOLD_CONVERSIONS)
2125 {
2126 tree tmp = chrec_convert_aggressive (TREE_TYPE (chrec), op0);
2127 if (tmp)
2128 return tmp;
2129 }
2130
2131 if (op0 == TREE_OPERAND (chrec, 0))
2132 return chrec;
2133
2134 /* If we used chrec_convert_aggressive, we can no longer assume that
2135 signed chrecs do not overflow, as chrec_convert does, so avoid
2136 calling it in that case. */
2137 if (flags & FOLD_CONVERSIONS)
2138 return fold_convert (TREE_TYPE (chrec), op0);
2139
2140 return chrec_convert (TREE_TYPE (chrec), op0, NULL_TREE);
2141
2142 case SCEV_NOT_KNOWN:
2143 return chrec_dont_know;
2144
2145 case SCEV_KNOWN:
2146 return chrec_known;
2147
2148 default:
2149 break;
2150 }
2151
2152 gcc_assert (!VL_EXP_CLASS_P (chrec));
2153 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
2154 {
2155 case 3:
2156 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2157 flags, cache, size_expr);
2158 if (op0 == chrec_dont_know)
2159 return chrec_dont_know;
2160
2161 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2162 flags, cache, size_expr);
2163 if (op1 == chrec_dont_know)
2164 return chrec_dont_know;
2165
2166 op2 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 2),
2167 flags, cache, size_expr);
2168 if (op2 == chrec_dont_know)
2169 return chrec_dont_know;
2170
2171 if (op0 == TREE_OPERAND (chrec, 0)
2172 && op1 == TREE_OPERAND (chrec, 1)
2173 && op2 == TREE_OPERAND (chrec, 2))
2174 return chrec;
2175
2176 return fold_build3 (TREE_CODE (chrec),
2177 TREE_TYPE (chrec), op0, op1, op2);
2178
2179 case 2:
2180 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2181 flags, cache, size_expr);
2182 if (op0 == chrec_dont_know)
2183 return chrec_dont_know;
2184
2185 op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1),
2186 flags, cache, size_expr);
2187 if (op1 == chrec_dont_know)
2188 return chrec_dont_know;
2189
2190 if (op0 == TREE_OPERAND (chrec, 0)
2191 && op1 == TREE_OPERAND (chrec, 1))
2192 return chrec;
2193 return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1);
2194
2195 case 1:
2196 op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0),
2197 flags, cache, size_expr);
2198 if (op0 == chrec_dont_know)
2199 return chrec_dont_know;
2200 if (op0 == TREE_OPERAND (chrec, 0))
2201 return chrec;
2202 return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0);
2203
2204 case 0:
2205 return chrec;
2206
2207 default:
2208 break;
2209 }
2210
2211 /* Too complicated to handle. */
2212 return chrec_dont_know;
2213 }
2214
2215 /* Analyze all the parameters of the chrec that were left under a
2216 symbolic form. LOOP is the loop in which symbolic names have to
2217 be analyzed and instantiated. */
2218
2219 tree
2220 instantiate_parameters (struct loop *loop,
2221 tree chrec)
2222 {
2223 tree res;
2224 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2225
2226 if (dump_file && (dump_flags & TDF_DETAILS))
2227 {
2228 fprintf (dump_file, "(instantiate_parameters \n");
2229 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
2230 fprintf (dump_file, " (chrec = ");
2231 print_generic_expr (dump_file, chrec, 0);
2232 fprintf (dump_file, ")\n");
2233 }
2234
2235 res = instantiate_parameters_1 (loop, chrec, INSERT_SUPERLOOP_CHRECS, cache,
2236 0);
2237
2238 if (dump_file && (dump_flags & TDF_DETAILS))
2239 {
2240 fprintf (dump_file, " (res = ");
2241 print_generic_expr (dump_file, res, 0);
2242 fprintf (dump_file, "))\n");
2243 }
2244
2245 htab_delete (cache);
2246
2247 return res;
2248 }
2249
2250 /* Similar to instantiate_parameters, but does not introduce the
2251 evolutions in outer loops for LOOP invariants in CHREC, and does not
2252 care about causing overflows, as long as they do not affect value
2253 of an expression. */
2254
2255 tree
2256 resolve_mixers (struct loop *loop, tree chrec)
2257 {
2258 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2259 tree ret = instantiate_parameters_1 (loop, chrec, FOLD_CONVERSIONS, cache, 0);
2260 htab_delete (cache);
2261 return ret;
2262 }
2263
2264 /* Entry point for the analysis of the number of iterations pass.
2265 This function tries to safely approximate the number of iterations
2266 the loop will run. When this property is not decidable at compile
2267 time, the result is chrec_dont_know. Otherwise the result is
2268 a scalar or a symbolic parameter.
2269
2270 Example of analysis: suppose that the loop has an exit condition:
2271
2272 "if (b > 49) goto end_loop;"
2273
2274 and that in a previous analysis we have determined that the
2275 variable 'b' has an evolution function:
2276
2277 "EF = {23, +, 5}_2".
2278
2279 When we evaluate the function at the point 5, i.e. the value of the
2280 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2281 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2282 the loop body has been executed 6 times. */
2283
2284 tree
2285 number_of_latch_executions (struct loop *loop)
2286 {
2287 tree res, type;
2288 edge exit;
2289 struct tree_niter_desc niter_desc;
2290
2291 /* Determine whether the number_of_iterations_in_loop has already
2292 been computed. */
2293 res = loop->nb_iterations;
2294 if (res)
2295 return res;
2296 res = chrec_dont_know;
2297
2298 if (dump_file && (dump_flags & TDF_DETAILS))
2299 fprintf (dump_file, "(number_of_iterations_in_loop\n");
2300
2301 exit = single_exit (loop);
2302 if (!exit)
2303 goto end;
2304
2305 if (!number_of_iterations_exit (loop, exit, &niter_desc, false))
2306 goto end;
2307
2308 type = TREE_TYPE (niter_desc.niter);
2309 if (integer_nonzerop (niter_desc.may_be_zero))
2310 res = build_int_cst (type, 0);
2311 else if (integer_zerop (niter_desc.may_be_zero))
2312 res = niter_desc.niter;
2313 else
2314 res = chrec_dont_know;
2315
2316 end:
2317 return set_nb_iterations_in_loop (loop, res);
2318 }
2319
2320 /* Returns the number of executions of the exit condition of LOOP,
2321 i.e., the number by one higher than number_of_latch_executions.
2322 Note that unline number_of_latch_executions, this number does
2323 not necessarily fit in the unsigned variant of the type of
2324 the control variable -- if the number of iterations is a constant,
2325 we return chrec_dont_know if adding one to number_of_latch_executions
2326 overflows; however, in case the number of iterations is symbolic
2327 expression, the caller is responsible for dealing with this
2328 the possible overflow. */
2329
2330 tree
2331 number_of_exit_cond_executions (struct loop *loop)
2332 {
2333 tree ret = number_of_latch_executions (loop);
2334 tree type = chrec_type (ret);
2335
2336 if (chrec_contains_undetermined (ret))
2337 return ret;
2338
2339 ret = chrec_fold_plus (type, ret, build_int_cst (type, 1));
2340 if (TREE_CODE (ret) == INTEGER_CST
2341 && TREE_OVERFLOW (ret))
2342 return chrec_dont_know;
2343
2344 return ret;
2345 }
2346
2347 /* One of the drivers for testing the scalar evolutions analysis.
2348 This function computes the number of iterations for all the loops
2349 from the EXIT_CONDITIONS array. */
2350
2351 static void
2352 number_of_iterations_for_all_loops (VEC(tree,heap) **exit_conditions)
2353 {
2354 unsigned int i;
2355 unsigned nb_chrec_dont_know_loops = 0;
2356 unsigned nb_static_loops = 0;
2357 tree cond;
2358
2359 for (i = 0; VEC_iterate (tree, *exit_conditions, i, cond); i++)
2360 {
2361 tree res = number_of_latch_executions (loop_containing_stmt (cond));
2362 if (chrec_contains_undetermined (res))
2363 nb_chrec_dont_know_loops++;
2364 else
2365 nb_static_loops++;
2366 }
2367
2368 if (dump_file)
2369 {
2370 fprintf (dump_file, "\n(\n");
2371 fprintf (dump_file, "-----------------------------------------\n");
2372 fprintf (dump_file, "%d\tnb_chrec_dont_know_loops\n", nb_chrec_dont_know_loops);
2373 fprintf (dump_file, "%d\tnb_static_loops\n", nb_static_loops);
2374 fprintf (dump_file, "%d\tnb_total_loops\n", number_of_loops ());
2375 fprintf (dump_file, "-----------------------------------------\n");
2376 fprintf (dump_file, ")\n\n");
2377
2378 print_loop_ir (dump_file);
2379 }
2380 }
2381
2382 \f
2383
2384 /* Counters for the stats. */
2385
2386 struct chrec_stats
2387 {
2388 unsigned nb_chrecs;
2389 unsigned nb_affine;
2390 unsigned nb_affine_multivar;
2391 unsigned nb_higher_poly;
2392 unsigned nb_chrec_dont_know;
2393 unsigned nb_undetermined;
2394 };
2395
2396 /* Reset the counters. */
2397
2398 static inline void
2399 reset_chrecs_counters (struct chrec_stats *stats)
2400 {
2401 stats->nb_chrecs = 0;
2402 stats->nb_affine = 0;
2403 stats->nb_affine_multivar = 0;
2404 stats->nb_higher_poly = 0;
2405 stats->nb_chrec_dont_know = 0;
2406 stats->nb_undetermined = 0;
2407 }
2408
2409 /* Dump the contents of a CHREC_STATS structure. */
2410
2411 static void
2412 dump_chrecs_stats (FILE *file, struct chrec_stats *stats)
2413 {
2414 fprintf (file, "\n(\n");
2415 fprintf (file, "-----------------------------------------\n");
2416 fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine);
2417 fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar);
2418 fprintf (file, "%d\tdegree greater than 2 polynomials\n",
2419 stats->nb_higher_poly);
2420 fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know);
2421 fprintf (file, "-----------------------------------------\n");
2422 fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs);
2423 fprintf (file, "%d\twith undetermined coefficients\n",
2424 stats->nb_undetermined);
2425 fprintf (file, "-----------------------------------------\n");
2426 fprintf (file, "%d\tchrecs in the scev database\n",
2427 (int) htab_elements (scalar_evolution_info));
2428 fprintf (file, "%d\tsets in the scev database\n", nb_set_scev);
2429 fprintf (file, "%d\tgets in the scev database\n", nb_get_scev);
2430 fprintf (file, "-----------------------------------------\n");
2431 fprintf (file, ")\n\n");
2432 }
2433
2434 /* Gather statistics about CHREC. */
2435
2436 static void
2437 gather_chrec_stats (tree chrec, struct chrec_stats *stats)
2438 {
2439 if (dump_file && (dump_flags & TDF_STATS))
2440 {
2441 fprintf (dump_file, "(classify_chrec ");
2442 print_generic_expr (dump_file, chrec, 0);
2443 fprintf (dump_file, "\n");
2444 }
2445
2446 stats->nb_chrecs++;
2447
2448 if (chrec == NULL_TREE)
2449 {
2450 stats->nb_undetermined++;
2451 return;
2452 }
2453
2454 switch (TREE_CODE (chrec))
2455 {
2456 case POLYNOMIAL_CHREC:
2457 if (evolution_function_is_affine_p (chrec))
2458 {
2459 if (dump_file && (dump_flags & TDF_STATS))
2460 fprintf (dump_file, " affine_univariate\n");
2461 stats->nb_affine++;
2462 }
2463 else if (evolution_function_is_affine_multivariate_p (chrec, 0))
2464 {
2465 if (dump_file && (dump_flags & TDF_STATS))
2466 fprintf (dump_file, " affine_multivariate\n");
2467 stats->nb_affine_multivar++;
2468 }
2469 else
2470 {
2471 if (dump_file && (dump_flags & TDF_STATS))
2472 fprintf (dump_file, " higher_degree_polynomial\n");
2473 stats->nb_higher_poly++;
2474 }
2475
2476 break;
2477
2478 default:
2479 break;
2480 }
2481
2482 if (chrec_contains_undetermined (chrec))
2483 {
2484 if (dump_file && (dump_flags & TDF_STATS))
2485 fprintf (dump_file, " undetermined\n");
2486 stats->nb_undetermined++;
2487 }
2488
2489 if (dump_file && (dump_flags & TDF_STATS))
2490 fprintf (dump_file, ")\n");
2491 }
2492
2493 /* One of the drivers for testing the scalar evolutions analysis.
2494 This function analyzes the scalar evolution of all the scalars
2495 defined as loop phi nodes in one of the loops from the
2496 EXIT_CONDITIONS array.
2497
2498 TODO Optimization: A loop is in canonical form if it contains only
2499 a single scalar loop phi node. All the other scalars that have an
2500 evolution in the loop are rewritten in function of this single
2501 index. This allows the parallelization of the loop. */
2502
2503 static void
2504 analyze_scalar_evolution_for_all_loop_phi_nodes (VEC(tree,heap) **exit_conditions)
2505 {
2506 unsigned int i;
2507 struct chrec_stats stats;
2508 tree cond;
2509
2510 reset_chrecs_counters (&stats);
2511
2512 for (i = 0; VEC_iterate (tree, *exit_conditions, i, cond); i++)
2513 {
2514 struct loop *loop;
2515 basic_block bb;
2516 tree phi, chrec;
2517
2518 loop = loop_containing_stmt (cond);
2519 bb = loop->header;
2520
2521 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
2522 if (is_gimple_reg (PHI_RESULT (phi)))
2523 {
2524 chrec = instantiate_parameters
2525 (loop,
2526 analyze_scalar_evolution (loop, PHI_RESULT (phi)));
2527
2528 if (dump_file && (dump_flags & TDF_STATS))
2529 gather_chrec_stats (chrec, &stats);
2530 }
2531 }
2532
2533 if (dump_file && (dump_flags & TDF_STATS))
2534 dump_chrecs_stats (dump_file, &stats);
2535 }
2536
2537 /* Callback for htab_traverse, gathers information on chrecs in the
2538 hashtable. */
2539
2540 static int
2541 gather_stats_on_scev_database_1 (void **slot, void *stats)
2542 {
2543 struct scev_info_str *entry = (struct scev_info_str *) *slot;
2544
2545 gather_chrec_stats (entry->chrec, (struct chrec_stats *) stats);
2546
2547 return 1;
2548 }
2549
2550 /* Classify the chrecs of the whole database. */
2551
2552 void
2553 gather_stats_on_scev_database (void)
2554 {
2555 struct chrec_stats stats;
2556
2557 if (!dump_file)
2558 return;
2559
2560 reset_chrecs_counters (&stats);
2561
2562 htab_traverse (scalar_evolution_info, gather_stats_on_scev_database_1,
2563 &stats);
2564
2565 dump_chrecs_stats (dump_file, &stats);
2566 }
2567
2568 \f
2569
2570 /* Initializer. */
2571
2572 static void
2573 initialize_scalar_evolutions_analyzer (void)
2574 {
2575 /* The elements below are unique. */
2576 if (chrec_dont_know == NULL_TREE)
2577 {
2578 chrec_not_analyzed_yet = NULL_TREE;
2579 chrec_dont_know = make_node (SCEV_NOT_KNOWN);
2580 chrec_known = make_node (SCEV_KNOWN);
2581 TREE_TYPE (chrec_dont_know) = void_type_node;
2582 TREE_TYPE (chrec_known) = void_type_node;
2583 }
2584 }
2585
2586 /* Initialize the analysis of scalar evolutions for LOOPS. */
2587
2588 void
2589 scev_initialize (void)
2590 {
2591 loop_iterator li;
2592 struct loop *loop;
2593
2594 scalar_evolution_info = htab_create_alloc (100,
2595 hash_scev_info,
2596 eq_scev_info,
2597 del_scev_info,
2598 ggc_calloc,
2599 ggc_free);
2600 already_instantiated = BITMAP_ALLOC (NULL);
2601
2602 initialize_scalar_evolutions_analyzer ();
2603
2604 FOR_EACH_LOOP (li, loop, 0)
2605 {
2606 loop->nb_iterations = NULL_TREE;
2607 }
2608 }
2609
2610 /* Cleans up the information cached by the scalar evolutions analysis. */
2611
2612 void
2613 scev_reset (void)
2614 {
2615 loop_iterator li;
2616 struct loop *loop;
2617
2618 if (!scalar_evolution_info || !current_loops)
2619 return;
2620
2621 htab_empty (scalar_evolution_info);
2622 FOR_EACH_LOOP (li, loop, 0)
2623 {
2624 loop->nb_iterations = NULL_TREE;
2625 }
2626 }
2627
2628 /* Checks whether OP behaves as a simple affine iv of LOOP in STMT and returns
2629 its base and step in IV if possible. If ALLOW_NONCONSTANT_STEP is true, we
2630 want step to be invariant in LOOP. Otherwise we require it to be an
2631 integer constant. IV->no_overflow is set to true if we are sure the iv cannot
2632 overflow (e.g. because it is computed in signed arithmetics). */
2633
2634 bool
2635 simple_iv (struct loop *loop, tree stmt, tree op, affine_iv *iv,
2636 bool allow_nonconstant_step)
2637 {
2638 basic_block bb = bb_for_stmt (stmt);
2639 tree type, ev;
2640 bool folded_casts;
2641
2642 iv->base = NULL_TREE;
2643 iv->step = NULL_TREE;
2644 iv->no_overflow = false;
2645
2646 type = TREE_TYPE (op);
2647 if (TREE_CODE (type) != INTEGER_TYPE
2648 && TREE_CODE (type) != POINTER_TYPE)
2649 return false;
2650
2651 ev = analyze_scalar_evolution_in_loop (loop, bb->loop_father, op,
2652 &folded_casts);
2653 if (chrec_contains_undetermined (ev))
2654 return false;
2655
2656 if (tree_does_not_contain_chrecs (ev)
2657 && !chrec_contains_symbols_defined_in_loop (ev, loop->num))
2658 {
2659 iv->base = ev;
2660 iv->step = build_int_cst (TREE_TYPE (ev), 0);
2661 iv->no_overflow = true;
2662 return true;
2663 }
2664
2665 if (TREE_CODE (ev) != POLYNOMIAL_CHREC
2666 || CHREC_VARIABLE (ev) != (unsigned) loop->num)
2667 return false;
2668
2669 iv->step = CHREC_RIGHT (ev);
2670 if (allow_nonconstant_step)
2671 {
2672 if (tree_contains_chrecs (iv->step, NULL)
2673 || chrec_contains_symbols_defined_in_loop (iv->step, loop->num))
2674 return false;
2675 }
2676 else if (TREE_CODE (iv->step) != INTEGER_CST)
2677 return false;
2678
2679 iv->base = CHREC_LEFT (ev);
2680 if (tree_contains_chrecs (iv->base, NULL)
2681 || chrec_contains_symbols_defined_in_loop (iv->base, loop->num))
2682 return false;
2683
2684 iv->no_overflow = !folded_casts && TYPE_OVERFLOW_UNDEFINED (type);
2685
2686 return true;
2687 }
2688
2689 /* Runs the analysis of scalar evolutions. */
2690
2691 void
2692 scev_analysis (void)
2693 {
2694 VEC(tree,heap) *exit_conditions;
2695
2696 exit_conditions = VEC_alloc (tree, heap, 37);
2697 select_loops_exit_conditions (&exit_conditions);
2698
2699 if (dump_file && (dump_flags & TDF_STATS))
2700 analyze_scalar_evolution_for_all_loop_phi_nodes (&exit_conditions);
2701
2702 number_of_iterations_for_all_loops (&exit_conditions);
2703 VEC_free (tree, heap, exit_conditions);
2704 }
2705
2706 /* Finalize the scalar evolution analysis. */
2707
2708 void
2709 scev_finalize (void)
2710 {
2711 if (!scalar_evolution_info)
2712 return;
2713 htab_delete (scalar_evolution_info);
2714 BITMAP_FREE (already_instantiated);
2715 scalar_evolution_info = NULL;
2716 }
2717
2718 /* Replace ssa names for that scev can prove they are constant by the
2719 appropriate constants. Also perform final value replacement in loops,
2720 in case the replacement expressions are cheap.
2721
2722 We only consider SSA names defined by phi nodes; rest is left to the
2723 ordinary constant propagation pass. */
2724
2725 unsigned int
2726 scev_const_prop (void)
2727 {
2728 basic_block bb;
2729 tree name, phi, next_phi, type, ev;
2730 struct loop *loop, *ex_loop;
2731 bitmap ssa_names_to_remove = NULL;
2732 unsigned i;
2733 loop_iterator li;
2734
2735 if (number_of_loops () <= 1)
2736 return 0;
2737
2738 FOR_EACH_BB (bb)
2739 {
2740 loop = bb->loop_father;
2741
2742 for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi))
2743 {
2744 name = PHI_RESULT (phi);
2745
2746 if (!is_gimple_reg (name))
2747 continue;
2748
2749 type = TREE_TYPE (name);
2750
2751 if (!POINTER_TYPE_P (type)
2752 && !INTEGRAL_TYPE_P (type))
2753 continue;
2754
2755 ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name));
2756 if (!is_gimple_min_invariant (ev)
2757 || !may_propagate_copy (name, ev))
2758 continue;
2759
2760 /* Replace the uses of the name. */
2761 if (name != ev)
2762 replace_uses_by (name, ev);
2763
2764 if (!ssa_names_to_remove)
2765 ssa_names_to_remove = BITMAP_ALLOC (NULL);
2766 bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name));
2767 }
2768 }
2769
2770 /* Remove the ssa names that were replaced by constants. We do not
2771 remove them directly in the previous cycle, since this
2772 invalidates scev cache. */
2773 if (ssa_names_to_remove)
2774 {
2775 bitmap_iterator bi;
2776
2777 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi)
2778 {
2779 name = ssa_name (i);
2780 phi = SSA_NAME_DEF_STMT (name);
2781
2782 gcc_assert (TREE_CODE (phi) == PHI_NODE);
2783 remove_phi_node (phi, NULL, true);
2784 }
2785
2786 BITMAP_FREE (ssa_names_to_remove);
2787 scev_reset ();
2788 }
2789
2790 /* Now the regular final value replacement. */
2791 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
2792 {
2793 edge exit;
2794 tree def, rslt, ass, niter;
2795 block_stmt_iterator bsi;
2796
2797 /* If we do not know exact number of iterations of the loop, we cannot
2798 replace the final value. */
2799 exit = single_exit (loop);
2800 if (!exit)
2801 continue;
2802
2803 niter = number_of_latch_executions (loop);
2804 /* We used to check here whether the computation of NITER is expensive,
2805 and avoided final value elimination if that is the case. The problem
2806 is that it is hard to evaluate whether the expression is too
2807 expensive, as we do not know what optimization opportunities the
2808 the elimination of the final value may reveal. Therefore, we now
2809 eliminate the final values of induction variables unconditionally. */
2810 if (niter == chrec_dont_know)
2811 continue;
2812
2813 /* Ensure that it is possible to insert new statements somewhere. */
2814 if (!single_pred_p (exit->dest))
2815 split_loop_exit_edge (exit);
2816 bsi = bsi_after_labels (exit->dest);
2817
2818 ex_loop = superloop_at_depth (loop,
2819 loop_depth (exit->dest->loop_father) + 1);
2820
2821 for (phi = phi_nodes (exit->dest); phi; phi = next_phi)
2822 {
2823 next_phi = PHI_CHAIN (phi);
2824 rslt = PHI_RESULT (phi);
2825 def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
2826 if (!is_gimple_reg (def))
2827 continue;
2828
2829 if (!POINTER_TYPE_P (TREE_TYPE (def))
2830 && !INTEGRAL_TYPE_P (TREE_TYPE (def)))
2831 continue;
2832
2833 def = analyze_scalar_evolution_in_loop (ex_loop, loop, def, NULL);
2834 def = compute_overall_effect_of_inner_loop (ex_loop, def);
2835 if (!tree_does_not_contain_chrecs (def)
2836 || chrec_contains_symbols_defined_in_loop (def, ex_loop->num)
2837 /* Moving the computation from the loop may prolong life range
2838 of some ssa names, which may cause problems if they appear
2839 on abnormal edges. */
2840 || contains_abnormal_ssa_name_p (def))
2841 continue;
2842
2843 /* Eliminate the PHI node and replace it by a computation outside
2844 the loop. */
2845 def = unshare_expr (def);
2846 remove_phi_node (phi, NULL_TREE, false);
2847
2848 ass = build_gimple_modify_stmt (rslt, NULL_TREE);
2849 SSA_NAME_DEF_STMT (rslt) = ass;
2850 {
2851 block_stmt_iterator dest = bsi;
2852 bsi_insert_before (&dest, ass, BSI_NEW_STMT);
2853 def = force_gimple_operand_bsi (&dest, def, false, NULL_TREE,
2854 true, BSI_SAME_STMT);
2855 }
2856 GIMPLE_STMT_OPERAND (ass, 1) = def;
2857 update_stmt (ass);
2858 }
2859 }
2860 return 0;
2861 }
2862
2863 #include "gt-tree-scalar-evolution.h"