PR target/65871
[official-gcc.git] / gcc / tree-scalar-evolution.c
blobd89f6314c4dbccf4f35f77f039b6ce184e71e302
1 /* Scalar evolution detector.
2 Copyright (C) 2003-2015 Free Software Foundation, Inc.
3 Contributed by Sebastian Pop <s.pop@laposte.net>
5 This file is part of GCC.
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.
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.
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/>. */
22 Description:
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.
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).
45 A short sketch of the algorithm is:
47 Given a scalar variable to be analyzed, follow the SSA edge to
48 its definition:
50 - When the definition is a GIMPLE_ASSIGN: 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.
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).
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.
73 Examples:
75 Example 1: Illustration of the basic algorithm.
77 | a = 3
78 | loop_1
79 | b = phi (a, c)
80 | c = b + 1
81 | if (c > 10) exit_loop
82 | endloop
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 (loop_1, {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:
115 a -> 3
116 b -> {3, +, 1}_1
117 c -> {4, +, 1}_1
119 or in terms of a C program:
121 | a = 3
122 | for (x = 0; x <= 7; x++)
124 | b = x + 3
125 | c = x + 4
128 Example 2a: Illustration of the algorithm on nested loops.
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
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:
143 b -> {c, +, 3}_2
144 d -> {c + 3, +, 3}_2
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:
154 a -> {1, +, 32}_1
155 c -> {3, +, 32}_1
157 Example 2b: Multivariate chains of recurrences.
159 | loop_1
160 | k = phi (0, k + 1)
161 | loop_2 4 times
162 | j = phi (0, j + 1)
163 | loop_3 4 times
164 | i = phi (0, i + 1)
165 | A[j + k] = ...
166 | endloop
167 | endloop
168 | endloop
170 Analyzing the access function of array A with
171 instantiate_parameters (loop_1, "j + k"), we obtain the
172 instantiation and the analysis of the scalar variables "j" and "k"
173 in loop_1. This leads to the scalar evolution {4, +, 1}_1: the end
174 value of loop_2 for "j" is 4, and the evolution of "k" in loop_1 is
175 {0, +, 1}_1. To obtain the evolution function in loop_3 and
176 instantiate the scalar variables up to loop_1, one has to use:
177 instantiate_scev (block_before_loop (loop_1), loop_3, "j + k").
178 The result of this call is {{0, +, 1}_1, +, 1}_2.
180 Example 3: Higher degree polynomials.
182 | loop_1
183 | a = phi (2, b)
184 | c = phi (5, d)
185 | b = a + 1
186 | d = c + a
187 | endloop
189 a -> {2, +, 1}_1
190 b -> {3, +, 1}_1
191 c -> {5, +, a}_1
192 d -> {5 + a, +, a}_1
194 instantiate_parameters (loop_1, {5, +, a}_1) -> {5, +, 2, +, 1}_1
195 instantiate_parameters (loop_1, {5 + a, +, a}_1) -> {7, +, 3, +, 1}_1
197 Example 4: Lucas, Fibonacci, or mixers in general.
199 | loop_1
200 | a = phi (1, b)
201 | c = phi (3, d)
202 | b = c
203 | d = c + a
204 | endloop
206 a -> (1, c)_1
207 c -> {3, +, a}_1
209 The syntax "(1, c)_1" stands for a PEELED_CHREC that has the
210 following semantics: during the first iteration of the loop_1, the
211 variable contains the value 1, and then it contains the value "c".
212 Note that this syntax is close to the syntax of the loop-phi-node:
213 "a -> (1, c)_1" vs. "a = phi (1, c)".
215 The symbolic chrec representation contains all the semantics of the
216 original code. What is more difficult is to use this information.
218 Example 5: Flip-flops, or exchangers.
220 | loop_1
221 | a = phi (1, b)
222 | c = phi (3, d)
223 | b = c
224 | d = a
225 | endloop
227 a -> (1, c)_1
228 c -> (3, a)_1
230 Based on these symbolic chrecs, it is possible to refine this
231 information into the more precise PERIODIC_CHRECs:
233 a -> |1, 3|_1
234 c -> |3, 1|_1
236 This transformation is not yet implemented.
238 Further readings:
240 You can find a more detailed description of the algorithm in:
241 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf
242 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that
243 this is a preliminary report and some of the details of the
244 algorithm have changed. I'm working on a research report that
245 updates the description of the algorithms to reflect the design
246 choices used in this implementation.
248 A set of slides show a high level overview of the algorithm and run
249 an example through the scalar evolution analyzer:
250 http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf
252 The slides that I have presented at the GCC Summit'04 are available
253 at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf
256 #include "config.h"
257 #include "system.h"
258 #include "coretypes.h"
259 #include "hash-set.h"
260 #include "machmode.h"
261 #include "vec.h"
262 #include "double-int.h"
263 #include "input.h"
264 #include "alias.h"
265 #include "symtab.h"
266 #include "options.h"
267 #include "wide-int.h"
268 #include "inchash.h"
269 #include "tree.h"
270 #include "fold-const.h"
271 #include "hashtab.h"
272 #include "tm.h"
273 #include "hard-reg-set.h"
274 #include "function.h"
275 #include "rtl.h"
276 #include "flags.h"
277 #include "statistics.h"
278 #include "real.h"
279 #include "fixed-value.h"
280 #include "insn-config.h"
281 #include "expmed.h"
282 #include "dojump.h"
283 #include "explow.h"
284 #include "calls.h"
285 #include "emit-rtl.h"
286 #include "varasm.h"
287 #include "stmt.h"
288 #include "expr.h"
289 #include "gimple-pretty-print.h"
290 #include "predict.h"
291 #include "dominance.h"
292 #include "cfg.h"
293 #include "basic-block.h"
294 #include "tree-ssa-alias.h"
295 #include "internal-fn.h"
296 #include "gimple-expr.h"
297 #include "is-a.h"
298 #include "gimple.h"
299 #include "gimplify.h"
300 #include "gimple-iterator.h"
301 #include "gimplify-me.h"
302 #include "gimple-ssa.h"
303 #include "tree-cfg.h"
304 #include "tree-phinodes.h"
305 #include "stringpool.h"
306 #include "tree-ssanames.h"
307 #include "tree-ssa-loop-ivopts.h"
308 #include "tree-ssa-loop-manip.h"
309 #include "tree-ssa-loop-niter.h"
310 #include "tree-ssa-loop.h"
311 #include "tree-ssa.h"
312 #include "cfgloop.h"
313 #include "tree-chrec.h"
314 #include "tree-affine.h"
315 #include "tree-scalar-evolution.h"
316 #include "dumpfile.h"
317 #include "params.h"
318 #include "tree-ssa-propagate.h"
319 #include "gimple-fold.h"
321 static tree analyze_scalar_evolution_1 (struct loop *, tree, tree);
322 static tree analyze_scalar_evolution_for_address_of (struct loop *loop,
323 tree var);
325 /* The cached information about an SSA name with version NAME_VERSION,
326 claiming that below basic block with index INSTANTIATED_BELOW, the
327 value of the SSA name can be expressed as CHREC. */
329 struct GTY((for_user)) scev_info_str {
330 unsigned int name_version;
331 int instantiated_below;
332 tree chrec;
335 /* Counters for the scev database. */
336 static unsigned nb_set_scev = 0;
337 static unsigned nb_get_scev = 0;
339 /* The following trees are unique elements. Thus the comparison of
340 another element to these elements should be done on the pointer to
341 these trees, and not on their value. */
343 /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
344 tree chrec_not_analyzed_yet;
346 /* Reserved to the cases where the analyzer has detected an
347 undecidable property at compile time. */
348 tree chrec_dont_know;
350 /* When the analyzer has detected that a property will never
351 happen, then it qualifies it with chrec_known. */
352 tree chrec_known;
354 struct scev_info_hasher : ggc_hasher<scev_info_str *>
356 static hashval_t hash (scev_info_str *i);
357 static bool equal (const scev_info_str *a, const scev_info_str *b);
360 static GTY (()) hash_table<scev_info_hasher> *scalar_evolution_info;
363 /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */
365 static inline struct scev_info_str *
366 new_scev_info_str (basic_block instantiated_below, tree var)
368 struct scev_info_str *res;
370 res = ggc_alloc<scev_info_str> ();
371 res->name_version = SSA_NAME_VERSION (var);
372 res->chrec = chrec_not_analyzed_yet;
373 res->instantiated_below = instantiated_below->index;
375 return res;
378 /* Computes a hash function for database element ELT. */
380 hashval_t
381 scev_info_hasher::hash (scev_info_str *elt)
383 return elt->name_version ^ elt->instantiated_below;
386 /* Compares database elements E1 and E2. */
388 bool
389 scev_info_hasher::equal (const scev_info_str *elt1, const scev_info_str *elt2)
391 return (elt1->name_version == elt2->name_version
392 && elt1->instantiated_below == elt2->instantiated_below);
395 /* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block.
396 A first query on VAR returns chrec_not_analyzed_yet. */
398 static tree *
399 find_var_scev_info (basic_block instantiated_below, tree var)
401 struct scev_info_str *res;
402 struct scev_info_str tmp;
404 tmp.name_version = SSA_NAME_VERSION (var);
405 tmp.instantiated_below = instantiated_below->index;
406 scev_info_str **slot = scalar_evolution_info->find_slot (&tmp, INSERT);
408 if (!*slot)
409 *slot = new_scev_info_str (instantiated_below, var);
410 res = *slot;
412 return &res->chrec;
415 /* Return true when CHREC contains symbolic names defined in
416 LOOP_NB. */
418 bool
419 chrec_contains_symbols_defined_in_loop (const_tree chrec, unsigned loop_nb)
421 int i, n;
423 if (chrec == NULL_TREE)
424 return false;
426 if (is_gimple_min_invariant (chrec))
427 return false;
429 if (TREE_CODE (chrec) == SSA_NAME)
431 gimple def;
432 loop_p def_loop, loop;
434 if (SSA_NAME_IS_DEFAULT_DEF (chrec))
435 return false;
437 def = SSA_NAME_DEF_STMT (chrec);
438 def_loop = loop_containing_stmt (def);
439 loop = get_loop (cfun, loop_nb);
441 if (def_loop == NULL)
442 return false;
444 if (loop == def_loop || flow_loop_nested_p (loop, def_loop))
445 return true;
447 return false;
450 n = TREE_OPERAND_LENGTH (chrec);
451 for (i = 0; i < n; i++)
452 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, i),
453 loop_nb))
454 return true;
455 return false;
458 /* Return true when PHI is a loop-phi-node. */
460 static bool
461 loop_phi_node_p (gimple phi)
463 /* The implementation of this function is based on the following
464 property: "all the loop-phi-nodes of a loop are contained in the
465 loop's header basic block". */
467 return loop_containing_stmt (phi)->header == gimple_bb (phi);
470 /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
471 In general, in the case of multivariate evolutions we want to get
472 the evolution in different loops. LOOP specifies the level for
473 which to get the evolution.
475 Example:
477 | for (j = 0; j < 100; j++)
479 | for (k = 0; k < 100; k++)
481 | i = k + j; - Here the value of i is a function of j, k.
483 | ... = i - Here the value of i is a function of j.
485 | ... = i - Here the value of i is a scalar.
487 Example:
489 | i_0 = ...
490 | loop_1 10 times
491 | i_1 = phi (i_0, i_2)
492 | i_2 = i_1 + 2
493 | endloop
495 This loop has the same effect as:
496 LOOP_1 has the same effect as:
498 | i_1 = i_0 + 20
500 The overall effect of the loop, "i_0 + 20" in the previous example,
501 is obtained by passing in the parameters: LOOP = 1,
502 EVOLUTION_FN = {i_0, +, 2}_1.
505 tree
506 compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn)
508 bool val = false;
510 if (evolution_fn == chrec_dont_know)
511 return chrec_dont_know;
513 else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC)
515 struct loop *inner_loop = get_chrec_loop (evolution_fn);
517 if (inner_loop == loop
518 || flow_loop_nested_p (loop, inner_loop))
520 tree nb_iter = number_of_latch_executions (inner_loop);
522 if (nb_iter == chrec_dont_know)
523 return chrec_dont_know;
524 else
526 tree res;
528 /* evolution_fn is the evolution function in LOOP. Get
529 its value in the nb_iter-th iteration. */
530 res = chrec_apply (inner_loop->num, evolution_fn, nb_iter);
532 if (chrec_contains_symbols_defined_in_loop (res, loop->num))
533 res = instantiate_parameters (loop, res);
535 /* Continue the computation until ending on a parent of LOOP. */
536 return compute_overall_effect_of_inner_loop (loop, res);
539 else
540 return evolution_fn;
543 /* If the evolution function is an invariant, there is nothing to do. */
544 else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val)
545 return evolution_fn;
547 else
548 return chrec_dont_know;
551 /* Associate CHREC to SCALAR. */
553 static void
554 set_scalar_evolution (basic_block instantiated_below, tree scalar, tree chrec)
556 tree *scalar_info;
558 if (TREE_CODE (scalar) != SSA_NAME)
559 return;
561 scalar_info = find_var_scev_info (instantiated_below, scalar);
563 if (dump_file)
565 if (dump_flags & TDF_SCEV)
567 fprintf (dump_file, "(set_scalar_evolution \n");
568 fprintf (dump_file, " instantiated_below = %d \n",
569 instantiated_below->index);
570 fprintf (dump_file, " (scalar = ");
571 print_generic_expr (dump_file, scalar, 0);
572 fprintf (dump_file, ")\n (scalar_evolution = ");
573 print_generic_expr (dump_file, chrec, 0);
574 fprintf (dump_file, "))\n");
576 if (dump_flags & TDF_STATS)
577 nb_set_scev++;
580 *scalar_info = chrec;
583 /* Retrieve the chrec associated to SCALAR instantiated below
584 INSTANTIATED_BELOW block. */
586 static tree
587 get_scalar_evolution (basic_block instantiated_below, tree scalar)
589 tree res;
591 if (dump_file)
593 if (dump_flags & TDF_SCEV)
595 fprintf (dump_file, "(get_scalar_evolution \n");
596 fprintf (dump_file, " (scalar = ");
597 print_generic_expr (dump_file, scalar, 0);
598 fprintf (dump_file, ")\n");
600 if (dump_flags & TDF_STATS)
601 nb_get_scev++;
604 switch (TREE_CODE (scalar))
606 case SSA_NAME:
607 res = *find_var_scev_info (instantiated_below, scalar);
608 break;
610 case REAL_CST:
611 case FIXED_CST:
612 case INTEGER_CST:
613 res = scalar;
614 break;
616 default:
617 res = chrec_not_analyzed_yet;
618 break;
621 if (dump_file && (dump_flags & TDF_SCEV))
623 fprintf (dump_file, " (scalar_evolution = ");
624 print_generic_expr (dump_file, res, 0);
625 fprintf (dump_file, "))\n");
628 return res;
631 /* Helper function for add_to_evolution. Returns the evolution
632 function for an assignment of the form "a = b + c", where "a" and
633 "b" are on the strongly connected component. CHREC_BEFORE is the
634 information that we already have collected up to this point.
635 TO_ADD is the evolution of "c".
637 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
638 evolution the expression TO_ADD, otherwise construct an evolution
639 part for this loop. */
641 static tree
642 add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add,
643 gimple at_stmt)
645 tree type, left, right;
646 struct loop *loop = get_loop (cfun, loop_nb), *chloop;
648 switch (TREE_CODE (chrec_before))
650 case POLYNOMIAL_CHREC:
651 chloop = get_chrec_loop (chrec_before);
652 if (chloop == loop
653 || flow_loop_nested_p (chloop, loop))
655 unsigned var;
657 type = chrec_type (chrec_before);
659 /* When there is no evolution part in this loop, build it. */
660 if (chloop != loop)
662 var = loop_nb;
663 left = chrec_before;
664 right = SCALAR_FLOAT_TYPE_P (type)
665 ? build_real (type, dconst0)
666 : build_int_cst (type, 0);
668 else
670 var = CHREC_VARIABLE (chrec_before);
671 left = CHREC_LEFT (chrec_before);
672 right = CHREC_RIGHT (chrec_before);
675 to_add = chrec_convert (type, to_add, at_stmt);
676 right = chrec_convert_rhs (type, right, at_stmt);
677 right = chrec_fold_plus (chrec_type (right), right, to_add);
678 return build_polynomial_chrec (var, left, right);
680 else
682 gcc_assert (flow_loop_nested_p (loop, chloop));
684 /* Search the evolution in LOOP_NB. */
685 left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before),
686 to_add, at_stmt);
687 right = CHREC_RIGHT (chrec_before);
688 right = chrec_convert_rhs (chrec_type (left), right, at_stmt);
689 return build_polynomial_chrec (CHREC_VARIABLE (chrec_before),
690 left, right);
693 default:
694 /* These nodes do not depend on a loop. */
695 if (chrec_before == chrec_dont_know)
696 return chrec_dont_know;
698 left = chrec_before;
699 right = chrec_convert_rhs (chrec_type (left), to_add, at_stmt);
700 return build_polynomial_chrec (loop_nb, left, right);
704 /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
705 of LOOP_NB.
707 Description (provided for completeness, for those who read code in
708 a plane, and for my poor 62 bytes brain that would have forgotten
709 all this in the next two or three months):
711 The algorithm of translation of programs from the SSA representation
712 into the chrecs syntax is based on a pattern matching. After having
713 reconstructed the overall tree expression for a loop, there are only
714 two cases that can arise:
716 1. a = loop-phi (init, a + expr)
717 2. a = loop-phi (init, expr)
719 where EXPR is either a scalar constant with respect to the analyzed
720 loop (this is a degree 0 polynomial), or an expression containing
721 other loop-phi definitions (these are higher degree polynomials).
723 Examples:
726 | init = ...
727 | loop_1
728 | a = phi (init, a + 5)
729 | endloop
732 | inita = ...
733 | initb = ...
734 | loop_1
735 | a = phi (inita, 2 * b + 3)
736 | b = phi (initb, b + 1)
737 | endloop
739 For the first case, the semantics of the SSA representation is:
741 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
743 that is, there is a loop index "x" that determines the scalar value
744 of the variable during the loop execution. During the first
745 iteration, the value is that of the initial condition INIT, while
746 during the subsequent iterations, it is the sum of the initial
747 condition with the sum of all the values of EXPR from the initial
748 iteration to the before last considered iteration.
750 For the second case, the semantics of the SSA program is:
752 | a (x) = init, if x = 0;
753 | expr (x - 1), otherwise.
755 The second case corresponds to the PEELED_CHREC, whose syntax is
756 close to the syntax of a loop-phi-node:
758 | phi (init, expr) vs. (init, expr)_x
760 The proof of the translation algorithm for the first case is a
761 proof by structural induction based on the degree of EXPR.
763 Degree 0:
764 When EXPR is a constant with respect to the analyzed loop, or in
765 other words when EXPR is a polynomial of degree 0, the evolution of
766 the variable A in the loop is an affine function with an initial
767 condition INIT, and a step EXPR. In order to show this, we start
768 from the semantics of the SSA representation:
770 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
772 and since "expr (j)" is a constant with respect to "j",
774 f (x) = init + x * expr
776 Finally, based on the semantics of the pure sum chrecs, by
777 identification we get the corresponding chrecs syntax:
779 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
780 f (x) -> {init, +, expr}_x
782 Higher degree:
783 Suppose that EXPR is a polynomial of degree N with respect to the
784 analyzed loop_x for which we have already determined that it is
785 written under the chrecs syntax:
787 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
789 We start from the semantics of the SSA program:
791 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
793 | f (x) = init + \sum_{j = 0}^{x - 1}
794 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
796 | f (x) = init + \sum_{j = 0}^{x - 1}
797 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
799 | f (x) = init + \sum_{k = 0}^{n - 1}
800 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
802 | f (x) = init + \sum_{k = 0}^{n - 1}
803 | (b_k * \binom{x}{k + 1})
805 | f (x) = init + b_0 * \binom{x}{1} + ...
806 | + b_{n-1} * \binom{x}{n}
808 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
809 | + b_{n-1} * \binom{x}{n}
812 And finally from the definition of the chrecs syntax, we identify:
813 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
815 This shows the mechanism that stands behind the add_to_evolution
816 function. An important point is that the use of symbolic
817 parameters avoids the need of an analysis schedule.
819 Example:
821 | inita = ...
822 | initb = ...
823 | loop_1
824 | a = phi (inita, a + 2 + b)
825 | b = phi (initb, b + 1)
826 | endloop
828 When analyzing "a", the algorithm keeps "b" symbolically:
830 | a -> {inita, +, 2 + b}_1
832 Then, after instantiation, the analyzer ends on the evolution:
834 | a -> {inita, +, 2 + initb, +, 1}_1
838 static tree
839 add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code,
840 tree to_add, gimple at_stmt)
842 tree type = chrec_type (to_add);
843 tree res = NULL_TREE;
845 if (to_add == NULL_TREE)
846 return chrec_before;
848 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
849 instantiated at this point. */
850 if (TREE_CODE (to_add) == POLYNOMIAL_CHREC)
851 /* This should not happen. */
852 return chrec_dont_know;
854 if (dump_file && (dump_flags & TDF_SCEV))
856 fprintf (dump_file, "(add_to_evolution \n");
857 fprintf (dump_file, " (loop_nb = %d)\n", loop_nb);
858 fprintf (dump_file, " (chrec_before = ");
859 print_generic_expr (dump_file, chrec_before, 0);
860 fprintf (dump_file, ")\n (to_add = ");
861 print_generic_expr (dump_file, to_add, 0);
862 fprintf (dump_file, ")\n");
865 if (code == MINUS_EXPR)
866 to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type)
867 ? build_real (type, dconstm1)
868 : build_int_cst_type (type, -1));
870 res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt);
872 if (dump_file && (dump_flags & TDF_SCEV))
874 fprintf (dump_file, " (res = ");
875 print_generic_expr (dump_file, res, 0);
876 fprintf (dump_file, "))\n");
879 return res;
884 /* This section selects the loops that will be good candidates for the
885 scalar evolution analysis. For the moment, greedily select all the
886 loop nests we could analyze. */
888 /* For a loop with a single exit edge, return the COND_EXPR that
889 guards the exit edge. If the expression is too difficult to
890 analyze, then give up. */
892 gcond *
893 get_loop_exit_condition (const struct loop *loop)
895 gcond *res = NULL;
896 edge exit_edge = single_exit (loop);
898 if (dump_file && (dump_flags & TDF_SCEV))
899 fprintf (dump_file, "(get_loop_exit_condition \n ");
901 if (exit_edge)
903 gimple stmt;
905 stmt = last_stmt (exit_edge->src);
906 if (gcond *cond_stmt = dyn_cast <gcond *> (stmt))
907 res = cond_stmt;
910 if (dump_file && (dump_flags & TDF_SCEV))
912 print_gimple_stmt (dump_file, res, 0, 0);
913 fprintf (dump_file, ")\n");
916 return res;
920 /* Depth first search algorithm. */
922 typedef enum t_bool {
923 t_false,
924 t_true,
925 t_dont_know
926 } t_bool;
929 static t_bool follow_ssa_edge (struct loop *loop, gimple, gphi *,
930 tree *, int);
932 /* Follow the ssa edge into the binary expression RHS0 CODE RHS1.
933 Return true if the strongly connected component has been found. */
935 static t_bool
936 follow_ssa_edge_binary (struct loop *loop, gimple at_stmt,
937 tree type, tree rhs0, enum tree_code code, tree rhs1,
938 gphi *halting_phi, tree *evolution_of_loop,
939 int limit)
941 t_bool res = t_false;
942 tree evol;
944 switch (code)
946 case POINTER_PLUS_EXPR:
947 case PLUS_EXPR:
948 if (TREE_CODE (rhs0) == SSA_NAME)
950 if (TREE_CODE (rhs1) == SSA_NAME)
952 /* Match an assignment under the form:
953 "a = b + c". */
955 /* We want only assignments of form "name + name" contribute to
956 LIMIT, as the other cases do not necessarily contribute to
957 the complexity of the expression. */
958 limit++;
960 evol = *evolution_of_loop;
961 res = follow_ssa_edge
962 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, &evol, limit);
964 if (res == t_true)
965 *evolution_of_loop = add_to_evolution
966 (loop->num,
967 chrec_convert (type, evol, at_stmt),
968 code, rhs1, at_stmt);
970 else if (res == t_false)
972 res = follow_ssa_edge
973 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
974 evolution_of_loop, limit);
976 if (res == t_true)
977 *evolution_of_loop = add_to_evolution
978 (loop->num,
979 chrec_convert (type, *evolution_of_loop, at_stmt),
980 code, rhs0, at_stmt);
982 else if (res == t_dont_know)
983 *evolution_of_loop = chrec_dont_know;
986 else if (res == t_dont_know)
987 *evolution_of_loop = chrec_dont_know;
990 else
992 /* Match an assignment under the form:
993 "a = b + ...". */
994 res = follow_ssa_edge
995 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
996 evolution_of_loop, limit);
997 if (res == t_true)
998 *evolution_of_loop = add_to_evolution
999 (loop->num, chrec_convert (type, *evolution_of_loop,
1000 at_stmt),
1001 code, rhs1, at_stmt);
1003 else if (res == t_dont_know)
1004 *evolution_of_loop = chrec_dont_know;
1008 else if (TREE_CODE (rhs1) == SSA_NAME)
1010 /* Match an assignment under the form:
1011 "a = ... + c". */
1012 res = follow_ssa_edge
1013 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1014 evolution_of_loop, limit);
1015 if (res == t_true)
1016 *evolution_of_loop = add_to_evolution
1017 (loop->num, chrec_convert (type, *evolution_of_loop,
1018 at_stmt),
1019 code, rhs0, at_stmt);
1021 else if (res == t_dont_know)
1022 *evolution_of_loop = chrec_dont_know;
1025 else
1026 /* Otherwise, match an assignment under the form:
1027 "a = ... + ...". */
1028 /* And there is nothing to do. */
1029 res = t_false;
1030 break;
1032 case MINUS_EXPR:
1033 /* This case is under the form "opnd0 = rhs0 - rhs1". */
1034 if (TREE_CODE (rhs0) == SSA_NAME)
1036 /* Match an assignment under the form:
1037 "a = b - ...". */
1039 /* We want only assignments of form "name - name" contribute to
1040 LIMIT, as the other cases do not necessarily contribute to
1041 the complexity of the expression. */
1042 if (TREE_CODE (rhs1) == SSA_NAME)
1043 limit++;
1045 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1046 evolution_of_loop, limit);
1047 if (res == t_true)
1048 *evolution_of_loop = add_to_evolution
1049 (loop->num, chrec_convert (type, *evolution_of_loop, at_stmt),
1050 MINUS_EXPR, rhs1, at_stmt);
1052 else if (res == t_dont_know)
1053 *evolution_of_loop = chrec_dont_know;
1055 else
1056 /* Otherwise, match an assignment under the form:
1057 "a = ... - ...". */
1058 /* And there is nothing to do. */
1059 res = t_false;
1060 break;
1062 default:
1063 res = t_false;
1066 return res;
1069 /* Follow the ssa edge into the expression EXPR.
1070 Return true if the strongly connected component has been found. */
1072 static t_bool
1073 follow_ssa_edge_expr (struct loop *loop, gimple at_stmt, tree expr,
1074 gphi *halting_phi, tree *evolution_of_loop,
1075 int limit)
1077 enum tree_code code = TREE_CODE (expr);
1078 tree type = TREE_TYPE (expr), rhs0, rhs1;
1079 t_bool res;
1081 /* The EXPR is one of the following cases:
1082 - an SSA_NAME,
1083 - an INTEGER_CST,
1084 - a PLUS_EXPR,
1085 - a POINTER_PLUS_EXPR,
1086 - a MINUS_EXPR,
1087 - an ASSERT_EXPR,
1088 - other cases are not yet handled. */
1090 switch (code)
1092 CASE_CONVERT:
1093 /* This assignment is under the form "a_1 = (cast) rhs. */
1094 res = follow_ssa_edge_expr (loop, at_stmt, TREE_OPERAND (expr, 0),
1095 halting_phi, evolution_of_loop, limit);
1096 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, at_stmt);
1097 break;
1099 case INTEGER_CST:
1100 /* This assignment is under the form "a_1 = 7". */
1101 res = t_false;
1102 break;
1104 case SSA_NAME:
1105 /* This assignment is under the form: "a_1 = b_2". */
1106 res = follow_ssa_edge
1107 (loop, SSA_NAME_DEF_STMT (expr), halting_phi, evolution_of_loop, limit);
1108 break;
1110 case POINTER_PLUS_EXPR:
1111 case PLUS_EXPR:
1112 case MINUS_EXPR:
1113 /* This case is under the form "rhs0 +- rhs1". */
1114 rhs0 = TREE_OPERAND (expr, 0);
1115 rhs1 = TREE_OPERAND (expr, 1);
1116 type = TREE_TYPE (rhs0);
1117 STRIP_USELESS_TYPE_CONVERSION (rhs0);
1118 STRIP_USELESS_TYPE_CONVERSION (rhs1);
1119 res = follow_ssa_edge_binary (loop, at_stmt, type, rhs0, code, rhs1,
1120 halting_phi, evolution_of_loop, limit);
1121 break;
1123 case ADDR_EXPR:
1124 /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */
1125 if (TREE_CODE (TREE_OPERAND (expr, 0)) == MEM_REF)
1127 expr = TREE_OPERAND (expr, 0);
1128 rhs0 = TREE_OPERAND (expr, 0);
1129 rhs1 = TREE_OPERAND (expr, 1);
1130 type = TREE_TYPE (rhs0);
1131 STRIP_USELESS_TYPE_CONVERSION (rhs0);
1132 STRIP_USELESS_TYPE_CONVERSION (rhs1);
1133 res = follow_ssa_edge_binary (loop, at_stmt, type,
1134 rhs0, POINTER_PLUS_EXPR, rhs1,
1135 halting_phi, evolution_of_loop, limit);
1137 else
1138 res = t_false;
1139 break;
1141 case ASSERT_EXPR:
1142 /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>"
1143 It must be handled as a copy assignment of the form a_1 = a_2. */
1144 rhs0 = ASSERT_EXPR_VAR (expr);
1145 if (TREE_CODE (rhs0) == SSA_NAME)
1146 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0),
1147 halting_phi, evolution_of_loop, limit);
1148 else
1149 res = t_false;
1150 break;
1152 default:
1153 res = t_false;
1154 break;
1157 return res;
1160 /* Follow the ssa edge into the right hand side of an assignment STMT.
1161 Return true if the strongly connected component has been found. */
1163 static t_bool
1164 follow_ssa_edge_in_rhs (struct loop *loop, gimple stmt,
1165 gphi *halting_phi, tree *evolution_of_loop,
1166 int limit)
1168 enum tree_code code = gimple_assign_rhs_code (stmt);
1169 tree type = gimple_expr_type (stmt), rhs1, rhs2;
1170 t_bool res;
1172 switch (code)
1174 CASE_CONVERT:
1175 /* This assignment is under the form "a_1 = (cast) rhs. */
1176 res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1177 halting_phi, evolution_of_loop, limit);
1178 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, stmt);
1179 break;
1181 case POINTER_PLUS_EXPR:
1182 case PLUS_EXPR:
1183 case MINUS_EXPR:
1184 rhs1 = gimple_assign_rhs1 (stmt);
1185 rhs2 = gimple_assign_rhs2 (stmt);
1186 type = TREE_TYPE (rhs1);
1187 res = follow_ssa_edge_binary (loop, stmt, type, rhs1, code, rhs2,
1188 halting_phi, evolution_of_loop, limit);
1189 break;
1191 default:
1192 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1193 res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1194 halting_phi, evolution_of_loop, limit);
1195 else
1196 res = t_false;
1197 break;
1200 return res;
1203 /* Checks whether the I-th argument of a PHI comes from a backedge. */
1205 static bool
1206 backedge_phi_arg_p (gphi *phi, int i)
1208 const_edge e = gimple_phi_arg_edge (phi, i);
1210 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1211 about updating it anywhere, and this should work as well most of the
1212 time. */
1213 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1214 return true;
1216 return false;
1219 /* Helper function for one branch of the condition-phi-node. Return
1220 true if the strongly connected component has been found following
1221 this path. */
1223 static inline t_bool
1224 follow_ssa_edge_in_condition_phi_branch (int i,
1225 struct loop *loop,
1226 gphi *condition_phi,
1227 gphi *halting_phi,
1228 tree *evolution_of_branch,
1229 tree init_cond, int limit)
1231 tree branch = PHI_ARG_DEF (condition_phi, i);
1232 *evolution_of_branch = chrec_dont_know;
1234 /* Do not follow back edges (they must belong to an irreducible loop, which
1235 we really do not want to worry about). */
1236 if (backedge_phi_arg_p (condition_phi, i))
1237 return t_false;
1239 if (TREE_CODE (branch) == SSA_NAME)
1241 *evolution_of_branch = init_cond;
1242 return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi,
1243 evolution_of_branch, limit);
1246 /* This case occurs when one of the condition branches sets
1247 the variable to a constant: i.e. a phi-node like
1248 "a_2 = PHI <a_7(5), 2(6)>;".
1250 FIXME: This case have to be refined correctly:
1251 in some cases it is possible to say something better than
1252 chrec_dont_know, for example using a wrap-around notation. */
1253 return t_false;
1256 /* This function merges the branches of a condition-phi-node in a
1257 loop. */
1259 static t_bool
1260 follow_ssa_edge_in_condition_phi (struct loop *loop,
1261 gphi *condition_phi,
1262 gphi *halting_phi,
1263 tree *evolution_of_loop, int limit)
1265 int i, n;
1266 tree init = *evolution_of_loop;
1267 tree evolution_of_branch;
1268 t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi,
1269 halting_phi,
1270 &evolution_of_branch,
1271 init, limit);
1272 if (res == t_false || res == t_dont_know)
1273 return res;
1275 *evolution_of_loop = evolution_of_branch;
1277 n = gimple_phi_num_args (condition_phi);
1278 for (i = 1; i < n; i++)
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;
1285 /* Increase the limit by the PHI argument number to avoid exponential
1286 time and memory complexity. */
1287 res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi,
1288 halting_phi,
1289 &evolution_of_branch,
1290 init, limit + i);
1291 if (res == t_false || res == t_dont_know)
1292 return res;
1294 *evolution_of_loop = chrec_merge (*evolution_of_loop,
1295 evolution_of_branch);
1298 return t_true;
1301 /* Follow an SSA edge in an inner loop. It computes the overall
1302 effect of the loop, and following the symbolic initial conditions,
1303 it follows the edges in the parent loop. The inner loop is
1304 considered as a single statement. */
1306 static t_bool
1307 follow_ssa_edge_inner_loop_phi (struct loop *outer_loop,
1308 gphi *loop_phi_node,
1309 gphi *halting_phi,
1310 tree *evolution_of_loop, int limit)
1312 struct loop *loop = loop_containing_stmt (loop_phi_node);
1313 tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node));
1315 /* Sometimes, the inner loop is too difficult to analyze, and the
1316 result of the analysis is a symbolic parameter. */
1317 if (ev == PHI_RESULT (loop_phi_node))
1319 t_bool res = t_false;
1320 int i, n = gimple_phi_num_args (loop_phi_node);
1322 for (i = 0; i < n; i++)
1324 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1325 basic_block bb;
1327 /* Follow the edges that exit the inner loop. */
1328 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1329 if (!flow_bb_inside_loop_p (loop, bb))
1330 res = follow_ssa_edge_expr (outer_loop, loop_phi_node,
1331 arg, halting_phi,
1332 evolution_of_loop, limit);
1333 if (res == t_true)
1334 break;
1337 /* If the path crosses this loop-phi, give up. */
1338 if (res == t_true)
1339 *evolution_of_loop = chrec_dont_know;
1341 return res;
1344 /* Otherwise, compute the overall effect of the inner loop. */
1345 ev = compute_overall_effect_of_inner_loop (loop, ev);
1346 return follow_ssa_edge_expr (outer_loop, loop_phi_node, ev, halting_phi,
1347 evolution_of_loop, limit);
1350 /* Follow an SSA edge from a loop-phi-node to itself, constructing a
1351 path that is analyzed on the return walk. */
1353 static t_bool
1354 follow_ssa_edge (struct loop *loop, gimple def, gphi *halting_phi,
1355 tree *evolution_of_loop, int limit)
1357 struct loop *def_loop;
1359 if (gimple_nop_p (def))
1360 return t_false;
1362 /* Give up if the path is longer than the MAX that we allow. */
1363 if (limit > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_COMPLEXITY))
1364 return t_dont_know;
1366 def_loop = loop_containing_stmt (def);
1368 switch (gimple_code (def))
1370 case GIMPLE_PHI:
1371 if (!loop_phi_node_p (def))
1372 /* DEF is a condition-phi-node. Follow the branches, and
1373 record their evolutions. Finally, merge the collected
1374 information and set the approximation to the main
1375 variable. */
1376 return follow_ssa_edge_in_condition_phi
1377 (loop, as_a <gphi *> (def), halting_phi, evolution_of_loop,
1378 limit);
1380 /* When the analyzed phi is the halting_phi, the
1381 depth-first search is over: we have found a path from
1382 the halting_phi to itself in the loop. */
1383 if (def == halting_phi)
1384 return t_true;
1386 /* Otherwise, the evolution of the HALTING_PHI depends
1387 on the evolution of another loop-phi-node, i.e. the
1388 evolution function is a higher degree polynomial. */
1389 if (def_loop == loop)
1390 return t_false;
1392 /* Inner loop. */
1393 if (flow_loop_nested_p (loop, def_loop))
1394 return follow_ssa_edge_inner_loop_phi
1395 (loop, as_a <gphi *> (def), halting_phi, evolution_of_loop,
1396 limit + 1);
1398 /* Outer loop. */
1399 return t_false;
1401 case GIMPLE_ASSIGN:
1402 return follow_ssa_edge_in_rhs (loop, def, halting_phi,
1403 evolution_of_loop, limit);
1405 default:
1406 /* At this level of abstraction, the program is just a set
1407 of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no
1408 other node to be handled. */
1409 return t_false;
1414 /* Simplify PEELED_CHREC represented by (init_cond, arg) in LOOP.
1415 Handle below case and return the corresponding POLYNOMIAL_CHREC:
1417 # i_17 = PHI <i_13(5), 0(3)>
1418 # _20 = PHI <_5(5), start_4(D)(3)>
1420 i_13 = i_17 + 1;
1421 _5 = start_4(D) + i_13;
1423 Though variable _20 appears as a PEELED_CHREC in the form of
1424 (start_4, _5)_LOOP, it's a POLYNOMIAL_CHREC like {start_4, 1}_LOOP.
1426 See PR41488. */
1428 static tree
1429 simplify_peeled_chrec (struct loop *loop, tree arg, tree init_cond)
1431 aff_tree aff1, aff2;
1432 tree ev, left, right, type, step_val;
1433 hash_map<tree, name_expansion *> *peeled_chrec_map = NULL;
1435 ev = instantiate_parameters (loop, analyze_scalar_evolution (loop, arg));
1436 if (ev == NULL_TREE || TREE_CODE (ev) != POLYNOMIAL_CHREC)
1437 return chrec_dont_know;
1439 left = CHREC_LEFT (ev);
1440 right = CHREC_RIGHT (ev);
1441 type = TREE_TYPE (left);
1442 step_val = chrec_fold_plus (type, init_cond, right);
1444 /* Transform (init, {left, right}_LOOP)_LOOP to {init, right}_LOOP
1445 if "left" equals to "init + right". */
1446 if (operand_equal_p (left, step_val, 0))
1448 if (dump_file && (dump_flags & TDF_SCEV))
1449 fprintf (dump_file, "Simplify PEELED_CHREC into POLYNOMIAL_CHREC.\n");
1451 return build_polynomial_chrec (loop->num, init_cond, right);
1454 /* Try harder to check if they are equal. */
1455 tree_to_aff_combination_expand (left, type, &aff1, &peeled_chrec_map);
1456 tree_to_aff_combination_expand (step_val, type, &aff2, &peeled_chrec_map);
1457 free_affine_expand_cache (&peeled_chrec_map);
1458 aff_combination_scale (&aff2, -1);
1459 aff_combination_add (&aff1, &aff2);
1461 /* Transform (init, {left, right}_LOOP)_LOOP to {init, right}_LOOP
1462 if "left" equals to "init + right". */
1463 if (aff_combination_zero_p (&aff1))
1465 if (dump_file && (dump_flags & TDF_SCEV))
1466 fprintf (dump_file, "Simplify PEELED_CHREC into POLYNOMIAL_CHREC.\n");
1468 return build_polynomial_chrec (loop->num, init_cond, right);
1470 return chrec_dont_know;
1473 /* Given a LOOP_PHI_NODE, this function determines the evolution
1474 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1476 static tree
1477 analyze_evolution_in_loop (gphi *loop_phi_node,
1478 tree init_cond)
1480 int i, n = gimple_phi_num_args (loop_phi_node);
1481 tree evolution_function = chrec_not_analyzed_yet;
1482 struct loop *loop = loop_containing_stmt (loop_phi_node);
1483 basic_block bb;
1484 static bool simplify_peeled_chrec_p = true;
1486 if (dump_file && (dump_flags & TDF_SCEV))
1488 fprintf (dump_file, "(analyze_evolution_in_loop \n");
1489 fprintf (dump_file, " (loop_phi_node = ");
1490 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
1491 fprintf (dump_file, ")\n");
1494 for (i = 0; i < n; i++)
1496 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1497 gimple ssa_chain;
1498 tree ev_fn;
1499 t_bool res;
1501 /* Select the edges that enter the loop body. */
1502 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1503 if (!flow_bb_inside_loop_p (loop, bb))
1504 continue;
1506 if (TREE_CODE (arg) == SSA_NAME)
1508 bool val = false;
1510 ssa_chain = SSA_NAME_DEF_STMT (arg);
1512 /* Pass in the initial condition to the follow edge function. */
1513 ev_fn = init_cond;
1514 res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0);
1516 /* If ev_fn has no evolution in the inner loop, and the
1517 init_cond is not equal to ev_fn, then we have an
1518 ambiguity between two possible values, as we cannot know
1519 the number of iterations at this point. */
1520 if (TREE_CODE (ev_fn) != POLYNOMIAL_CHREC
1521 && no_evolution_in_loop_p (ev_fn, loop->num, &val) && val
1522 && !operand_equal_p (init_cond, ev_fn, 0))
1523 ev_fn = chrec_dont_know;
1525 else
1526 res = t_false;
1528 /* When it is impossible to go back on the same
1529 loop_phi_node by following the ssa edges, the
1530 evolution is represented by a peeled chrec, i.e. the
1531 first iteration, EV_FN has the value INIT_COND, then
1532 all the other iterations it has the value of ARG.
1533 For the moment, PEELED_CHREC nodes are not built. */
1534 if (res != t_true)
1536 ev_fn = chrec_dont_know;
1537 /* Try to recognize POLYNOMIAL_CHREC which appears in
1538 the form of PEELED_CHREC, but guard the process with
1539 a bool variable to keep the analyzer from infinite
1540 recurrence for real PEELED_RECs. */
1541 if (simplify_peeled_chrec_p && TREE_CODE (arg) == SSA_NAME)
1543 simplify_peeled_chrec_p = false;
1544 ev_fn = simplify_peeled_chrec (loop, arg, init_cond);
1545 simplify_peeled_chrec_p = true;
1549 /* When there are multiple back edges of the loop (which in fact never
1550 happens currently, but nevertheless), merge their evolutions. */
1551 evolution_function = chrec_merge (evolution_function, ev_fn);
1554 if (dump_file && (dump_flags & TDF_SCEV))
1556 fprintf (dump_file, " (evolution_function = ");
1557 print_generic_expr (dump_file, evolution_function, 0);
1558 fprintf (dump_file, "))\n");
1561 return evolution_function;
1564 /* Given a loop-phi-node, return the initial conditions of the
1565 variable on entry of the loop. When the CCP has propagated
1566 constants into the loop-phi-node, the initial condition is
1567 instantiated, otherwise the initial condition is kept symbolic.
1568 This analyzer does not analyze the evolution outside the current
1569 loop, and leaves this task to the on-demand tree reconstructor. */
1571 static tree
1572 analyze_initial_condition (gphi *loop_phi_node)
1574 int i, n;
1575 tree init_cond = chrec_not_analyzed_yet;
1576 struct loop *loop = loop_containing_stmt (loop_phi_node);
1578 if (dump_file && (dump_flags & TDF_SCEV))
1580 fprintf (dump_file, "(analyze_initial_condition \n");
1581 fprintf (dump_file, " (loop_phi_node = \n");
1582 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
1583 fprintf (dump_file, ")\n");
1586 n = gimple_phi_num_args (loop_phi_node);
1587 for (i = 0; i < n; i++)
1589 tree branch = PHI_ARG_DEF (loop_phi_node, i);
1590 basic_block bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1592 /* When the branch is oriented to the loop's body, it does
1593 not contribute to the initial condition. */
1594 if (flow_bb_inside_loop_p (loop, bb))
1595 continue;
1597 if (init_cond == chrec_not_analyzed_yet)
1599 init_cond = branch;
1600 continue;
1603 if (TREE_CODE (branch) == SSA_NAME)
1605 init_cond = chrec_dont_know;
1606 break;
1609 init_cond = chrec_merge (init_cond, branch);
1612 /* Ooops -- a loop without an entry??? */
1613 if (init_cond == chrec_not_analyzed_yet)
1614 init_cond = chrec_dont_know;
1616 /* During early loop unrolling we do not have fully constant propagated IL.
1617 Handle degenerate PHIs here to not miss important unrollings. */
1618 if (TREE_CODE (init_cond) == SSA_NAME)
1620 gimple def = SSA_NAME_DEF_STMT (init_cond);
1621 if (gphi *phi = dyn_cast <gphi *> (def))
1623 tree res = degenerate_phi_result (phi);
1624 if (res != NULL_TREE
1625 /* Only allow invariants here, otherwise we may break
1626 loop-closed SSA form. */
1627 && is_gimple_min_invariant (res))
1628 init_cond = res;
1632 if (dump_file && (dump_flags & TDF_SCEV))
1634 fprintf (dump_file, " (init_cond = ");
1635 print_generic_expr (dump_file, init_cond, 0);
1636 fprintf (dump_file, "))\n");
1639 return init_cond;
1642 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1644 static tree
1645 interpret_loop_phi (struct loop *loop, gphi *loop_phi_node)
1647 tree res;
1648 struct loop *phi_loop = loop_containing_stmt (loop_phi_node);
1649 tree init_cond;
1651 if (phi_loop != loop)
1653 struct loop *subloop;
1654 tree evolution_fn = analyze_scalar_evolution
1655 (phi_loop, PHI_RESULT (loop_phi_node));
1657 /* Dive one level deeper. */
1658 subloop = superloop_at_depth (phi_loop, loop_depth (loop) + 1);
1660 /* Interpret the subloop. */
1661 res = compute_overall_effect_of_inner_loop (subloop, evolution_fn);
1662 return res;
1665 /* Otherwise really interpret the loop phi. */
1666 init_cond = analyze_initial_condition (loop_phi_node);
1667 res = analyze_evolution_in_loop (loop_phi_node, init_cond);
1669 /* Verify we maintained the correct initial condition throughout
1670 possible conversions in the SSA chain. */
1671 if (res != chrec_dont_know)
1673 tree new_init = res;
1674 if (CONVERT_EXPR_P (res)
1675 && TREE_CODE (TREE_OPERAND (res, 0)) == POLYNOMIAL_CHREC)
1676 new_init = fold_convert (TREE_TYPE (res),
1677 CHREC_LEFT (TREE_OPERAND (res, 0)));
1678 else if (TREE_CODE (res) == POLYNOMIAL_CHREC)
1679 new_init = CHREC_LEFT (res);
1680 STRIP_USELESS_TYPE_CONVERSION (new_init);
1681 if (TREE_CODE (new_init) == POLYNOMIAL_CHREC
1682 || !operand_equal_p (init_cond, new_init, 0))
1683 return chrec_dont_know;
1686 return res;
1689 /* This function merges the branches of a condition-phi-node,
1690 contained in the outermost loop, and whose arguments are already
1691 analyzed. */
1693 static tree
1694 interpret_condition_phi (struct loop *loop, gphi *condition_phi)
1696 int i, n = gimple_phi_num_args (condition_phi);
1697 tree res = chrec_not_analyzed_yet;
1699 for (i = 0; i < n; i++)
1701 tree branch_chrec;
1703 if (backedge_phi_arg_p (condition_phi, i))
1705 res = chrec_dont_know;
1706 break;
1709 branch_chrec = analyze_scalar_evolution
1710 (loop, PHI_ARG_DEF (condition_phi, i));
1712 res = chrec_merge (res, branch_chrec);
1715 return res;
1718 /* Interpret the operation RHS1 OP RHS2. If we didn't
1719 analyze this node before, follow the definitions until ending
1720 either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the
1721 return path, this function propagates evolutions (ala constant copy
1722 propagation). OPND1 is not a GIMPLE expression because we could
1723 analyze the effect of an inner loop: see interpret_loop_phi. */
1725 static tree
1726 interpret_rhs_expr (struct loop *loop, gimple at_stmt,
1727 tree type, tree rhs1, enum tree_code code, tree rhs2)
1729 tree res, chrec1, chrec2;
1730 gimple def;
1732 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1734 if (is_gimple_min_invariant (rhs1))
1735 return chrec_convert (type, rhs1, at_stmt);
1737 if (code == SSA_NAME)
1738 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1739 at_stmt);
1741 if (code == ASSERT_EXPR)
1743 rhs1 = ASSERT_EXPR_VAR (rhs1);
1744 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1745 at_stmt);
1749 switch (code)
1751 case ADDR_EXPR:
1752 if (TREE_CODE (TREE_OPERAND (rhs1, 0)) == MEM_REF
1753 || handled_component_p (TREE_OPERAND (rhs1, 0)))
1755 machine_mode mode;
1756 HOST_WIDE_INT bitsize, bitpos;
1757 int unsignedp;
1758 int volatilep = 0;
1759 tree base, offset;
1760 tree chrec3;
1761 tree unitpos;
1763 base = get_inner_reference (TREE_OPERAND (rhs1, 0),
1764 &bitsize, &bitpos, &offset,
1765 &mode, &unsignedp, &volatilep, false);
1767 if (TREE_CODE (base) == MEM_REF)
1769 rhs2 = TREE_OPERAND (base, 1);
1770 rhs1 = TREE_OPERAND (base, 0);
1772 chrec1 = analyze_scalar_evolution (loop, rhs1);
1773 chrec2 = analyze_scalar_evolution (loop, rhs2);
1774 chrec1 = chrec_convert (type, chrec1, at_stmt);
1775 chrec2 = chrec_convert (TREE_TYPE (rhs2), chrec2, at_stmt);
1776 chrec1 = instantiate_parameters (loop, chrec1);
1777 chrec2 = instantiate_parameters (loop, chrec2);
1778 res = chrec_fold_plus (type, chrec1, chrec2);
1780 else
1782 chrec1 = analyze_scalar_evolution_for_address_of (loop, base);
1783 chrec1 = chrec_convert (type, chrec1, at_stmt);
1784 res = chrec1;
1787 if (offset != NULL_TREE)
1789 chrec2 = analyze_scalar_evolution (loop, offset);
1790 chrec2 = chrec_convert (TREE_TYPE (offset), chrec2, at_stmt);
1791 chrec2 = instantiate_parameters (loop, chrec2);
1792 res = chrec_fold_plus (type, res, chrec2);
1795 if (bitpos != 0)
1797 gcc_assert ((bitpos % BITS_PER_UNIT) == 0);
1799 unitpos = size_int (bitpos / BITS_PER_UNIT);
1800 chrec3 = analyze_scalar_evolution (loop, unitpos);
1801 chrec3 = chrec_convert (TREE_TYPE (unitpos), chrec3, at_stmt);
1802 chrec3 = instantiate_parameters (loop, chrec3);
1803 res = chrec_fold_plus (type, res, chrec3);
1806 else
1807 res = chrec_dont_know;
1808 break;
1810 case POINTER_PLUS_EXPR:
1811 chrec1 = analyze_scalar_evolution (loop, rhs1);
1812 chrec2 = analyze_scalar_evolution (loop, rhs2);
1813 chrec1 = chrec_convert (type, chrec1, at_stmt);
1814 chrec2 = chrec_convert (TREE_TYPE (rhs2), chrec2, at_stmt);
1815 chrec1 = instantiate_parameters (loop, chrec1);
1816 chrec2 = instantiate_parameters (loop, chrec2);
1817 res = chrec_fold_plus (type, chrec1, chrec2);
1818 break;
1820 case PLUS_EXPR:
1821 chrec1 = analyze_scalar_evolution (loop, rhs1);
1822 chrec2 = analyze_scalar_evolution (loop, rhs2);
1823 chrec1 = chrec_convert (type, chrec1, at_stmt);
1824 chrec2 = chrec_convert (type, chrec2, at_stmt);
1825 chrec1 = instantiate_parameters (loop, chrec1);
1826 chrec2 = instantiate_parameters (loop, chrec2);
1827 res = chrec_fold_plus (type, chrec1, chrec2);
1828 break;
1830 case MINUS_EXPR:
1831 chrec1 = analyze_scalar_evolution (loop, rhs1);
1832 chrec2 = analyze_scalar_evolution (loop, rhs2);
1833 chrec1 = chrec_convert (type, chrec1, at_stmt);
1834 chrec2 = chrec_convert (type, chrec2, at_stmt);
1835 chrec1 = instantiate_parameters (loop, chrec1);
1836 chrec2 = instantiate_parameters (loop, chrec2);
1837 res = chrec_fold_minus (type, chrec1, chrec2);
1838 break;
1840 case NEGATE_EXPR:
1841 chrec1 = analyze_scalar_evolution (loop, rhs1);
1842 chrec1 = chrec_convert (type, chrec1, at_stmt);
1843 /* TYPE may be integer, real or complex, so use fold_convert. */
1844 chrec1 = instantiate_parameters (loop, chrec1);
1845 res = chrec_fold_multiply (type, chrec1,
1846 fold_convert (type, integer_minus_one_node));
1847 break;
1849 case BIT_NOT_EXPR:
1850 /* Handle ~X as -1 - X. */
1851 chrec1 = analyze_scalar_evolution (loop, rhs1);
1852 chrec1 = chrec_convert (type, chrec1, at_stmt);
1853 chrec1 = instantiate_parameters (loop, chrec1);
1854 res = chrec_fold_minus (type,
1855 fold_convert (type, integer_minus_one_node),
1856 chrec1);
1857 break;
1859 case MULT_EXPR:
1860 chrec1 = analyze_scalar_evolution (loop, rhs1);
1861 chrec2 = analyze_scalar_evolution (loop, rhs2);
1862 chrec1 = chrec_convert (type, chrec1, at_stmt);
1863 chrec2 = chrec_convert (type, chrec2, at_stmt);
1864 chrec1 = instantiate_parameters (loop, chrec1);
1865 chrec2 = instantiate_parameters (loop, chrec2);
1866 res = chrec_fold_multiply (type, chrec1, chrec2);
1867 break;
1869 CASE_CONVERT:
1870 /* In case we have a truncation of a widened operation that in
1871 the truncated type has undefined overflow behavior analyze
1872 the operation done in an unsigned type of the same precision
1873 as the final truncation. We cannot derive a scalar evolution
1874 for the widened operation but for the truncated result. */
1875 if (TREE_CODE (type) == INTEGER_TYPE
1876 && TREE_CODE (TREE_TYPE (rhs1)) == INTEGER_TYPE
1877 && TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (rhs1))
1878 && TYPE_OVERFLOW_UNDEFINED (type)
1879 && TREE_CODE (rhs1) == SSA_NAME
1880 && (def = SSA_NAME_DEF_STMT (rhs1))
1881 && is_gimple_assign (def)
1882 && TREE_CODE_CLASS (gimple_assign_rhs_code (def)) == tcc_binary
1883 && TREE_CODE (gimple_assign_rhs2 (def)) == INTEGER_CST)
1885 tree utype = unsigned_type_for (type);
1886 chrec1 = interpret_rhs_expr (loop, at_stmt, utype,
1887 gimple_assign_rhs1 (def),
1888 gimple_assign_rhs_code (def),
1889 gimple_assign_rhs2 (def));
1891 else
1892 chrec1 = analyze_scalar_evolution (loop, rhs1);
1893 res = chrec_convert (type, chrec1, at_stmt);
1894 break;
1896 default:
1897 res = chrec_dont_know;
1898 break;
1901 return res;
1904 /* Interpret the expression EXPR. */
1906 static tree
1907 interpret_expr (struct loop *loop, gimple at_stmt, tree expr)
1909 enum tree_code code;
1910 tree type = TREE_TYPE (expr), op0, op1;
1912 if (automatically_generated_chrec_p (expr))
1913 return expr;
1915 if (TREE_CODE (expr) == POLYNOMIAL_CHREC
1916 || get_gimple_rhs_class (TREE_CODE (expr)) == GIMPLE_TERNARY_RHS)
1917 return chrec_dont_know;
1919 extract_ops_from_tree (expr, &code, &op0, &op1);
1921 return interpret_rhs_expr (loop, at_stmt, type,
1922 op0, code, op1);
1925 /* Interpret the rhs of the assignment STMT. */
1927 static tree
1928 interpret_gimple_assign (struct loop *loop, gimple stmt)
1930 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
1931 enum tree_code code = gimple_assign_rhs_code (stmt);
1933 return interpret_rhs_expr (loop, stmt, type,
1934 gimple_assign_rhs1 (stmt), code,
1935 gimple_assign_rhs2 (stmt));
1940 /* This section contains all the entry points:
1941 - number_of_iterations_in_loop,
1942 - analyze_scalar_evolution,
1943 - instantiate_parameters.
1946 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1947 common ancestor of DEF_LOOP and USE_LOOP. */
1949 static tree
1950 compute_scalar_evolution_in_loop (struct loop *wrto_loop,
1951 struct loop *def_loop,
1952 tree ev)
1954 bool val;
1955 tree res;
1957 if (def_loop == wrto_loop)
1958 return ev;
1960 def_loop = superloop_at_depth (def_loop, loop_depth (wrto_loop) + 1);
1961 res = compute_overall_effect_of_inner_loop (def_loop, ev);
1963 if (no_evolution_in_loop_p (res, wrto_loop->num, &val) && val)
1964 return res;
1966 return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet);
1969 /* Helper recursive function. */
1971 static tree
1972 analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res)
1974 tree type = TREE_TYPE (var);
1975 gimple def;
1976 basic_block bb;
1977 struct loop *def_loop;
1979 if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE)
1980 return chrec_dont_know;
1982 if (TREE_CODE (var) != SSA_NAME)
1983 return interpret_expr (loop, NULL, var);
1985 def = SSA_NAME_DEF_STMT (var);
1986 bb = gimple_bb (def);
1987 def_loop = bb ? bb->loop_father : NULL;
1989 if (bb == NULL
1990 || !flow_bb_inside_loop_p (loop, bb))
1992 /* Keep the symbolic form. */
1993 res = var;
1994 goto set_and_end;
1997 if (res != chrec_not_analyzed_yet)
1999 if (loop != bb->loop_father)
2000 res = compute_scalar_evolution_in_loop
2001 (find_common_loop (loop, bb->loop_father), bb->loop_father, res);
2003 goto set_and_end;
2006 if (loop != def_loop)
2008 res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet);
2009 res = compute_scalar_evolution_in_loop (loop, def_loop, res);
2011 goto set_and_end;
2014 switch (gimple_code (def))
2016 case GIMPLE_ASSIGN:
2017 res = interpret_gimple_assign (loop, def);
2018 break;
2020 case GIMPLE_PHI:
2021 if (loop_phi_node_p (def))
2022 res = interpret_loop_phi (loop, as_a <gphi *> (def));
2023 else
2024 res = interpret_condition_phi (loop, as_a <gphi *> (def));
2025 break;
2027 default:
2028 res = chrec_dont_know;
2029 break;
2032 set_and_end:
2034 /* Keep the symbolic form. */
2035 if (res == chrec_dont_know)
2036 res = var;
2038 if (loop == def_loop)
2039 set_scalar_evolution (block_before_loop (loop), var, res);
2041 return res;
2044 /* Analyzes and returns the scalar evolution of the ssa_name VAR in
2045 LOOP. LOOP is the loop in which the variable is used.
2047 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
2048 pointer to the statement that uses this variable, in order to
2049 determine the evolution function of the variable, use the following
2050 calls:
2052 loop_p loop = loop_containing_stmt (stmt);
2053 tree chrec_with_symbols = analyze_scalar_evolution (loop, var);
2054 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
2057 tree
2058 analyze_scalar_evolution (struct loop *loop, tree var)
2060 tree res;
2062 if (dump_file && (dump_flags & TDF_SCEV))
2064 fprintf (dump_file, "(analyze_scalar_evolution \n");
2065 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
2066 fprintf (dump_file, " (scalar = ");
2067 print_generic_expr (dump_file, var, 0);
2068 fprintf (dump_file, ")\n");
2071 res = get_scalar_evolution (block_before_loop (loop), var);
2072 res = analyze_scalar_evolution_1 (loop, var, res);
2074 if (dump_file && (dump_flags & TDF_SCEV))
2075 fprintf (dump_file, ")\n");
2077 return res;
2080 /* Analyzes and returns the scalar evolution of VAR address in LOOP. */
2082 static tree
2083 analyze_scalar_evolution_for_address_of (struct loop *loop, tree var)
2085 return analyze_scalar_evolution (loop, build_fold_addr_expr (var));
2088 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
2089 WRTO_LOOP (which should be a superloop of USE_LOOP)
2091 FOLDED_CASTS is set to true if resolve_mixers used
2092 chrec_convert_aggressive (TODO -- not really, we are way too conservative
2093 at the moment in order to keep things simple).
2095 To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following
2096 example:
2098 for (i = 0; i < 100; i++) -- loop 1
2100 for (j = 0; j < 100; j++) -- loop 2
2102 k1 = i;
2103 k2 = j;
2105 use2 (k1, k2);
2107 for (t = 0; t < 100; t++) -- loop 3
2108 use3 (k1, k2);
2111 use1 (k1, k2);
2114 Both k1 and k2 are invariants in loop3, thus
2115 analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1
2116 analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2
2118 As they are invariant, it does not matter whether we consider their
2119 usage in loop 3 or loop 2, hence
2120 analyze_scalar_evolution_in_loop (loop2, loop3, k1) =
2121 analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i
2122 analyze_scalar_evolution_in_loop (loop2, loop3, k2) =
2123 analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2
2125 Similarly for their evolutions with respect to loop 1. The values of K2
2126 in the use in loop 2 vary independently on loop 1, thus we cannot express
2127 the evolution with respect to loop 1:
2128 analyze_scalar_evolution_in_loop (loop1, loop3, k1) =
2129 analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1
2130 analyze_scalar_evolution_in_loop (loop1, loop3, k2) =
2131 analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know
2133 The value of k2 in the use in loop 1 is known, though:
2134 analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1
2135 analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100
2138 static tree
2139 analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop,
2140 tree version, bool *folded_casts)
2142 bool val = false;
2143 tree ev = version, tmp;
2145 /* We cannot just do
2147 tmp = analyze_scalar_evolution (use_loop, version);
2148 ev = resolve_mixers (wrto_loop, tmp);
2150 as resolve_mixers would query the scalar evolution with respect to
2151 wrto_loop. For example, in the situation described in the function
2152 comment, suppose that wrto_loop = loop1, use_loop = loop3 and
2153 version = k2. Then
2155 analyze_scalar_evolution (use_loop, version) = k2
2157 and resolve_mixers (loop1, k2) finds that the value of k2 in loop 1
2158 is 100, which is a wrong result, since we are interested in the
2159 value in loop 3.
2161 Instead, we need to proceed from use_loop to wrto_loop loop by loop,
2162 each time checking that there is no evolution in the inner loop. */
2164 if (folded_casts)
2165 *folded_casts = false;
2166 while (1)
2168 tmp = analyze_scalar_evolution (use_loop, ev);
2169 ev = resolve_mixers (use_loop, tmp);
2171 if (folded_casts && tmp != ev)
2172 *folded_casts = true;
2174 if (use_loop == wrto_loop)
2175 return ev;
2177 /* If the value of the use changes in the inner loop, we cannot express
2178 its value in the outer loop (we might try to return interval chrec,
2179 but we do not have a user for it anyway) */
2180 if (!no_evolution_in_loop_p (ev, use_loop->num, &val)
2181 || !val)
2182 return chrec_dont_know;
2184 use_loop = loop_outer (use_loop);
2189 /* Hashtable helpers for a temporary hash-table used when
2190 instantiating a CHREC or resolving mixers. For this use
2191 instantiated_below is always the same. */
2193 struct instantiate_cache_type
2195 htab_t map;
2196 vec<scev_info_str> entries;
2198 instantiate_cache_type () : map (NULL), entries (vNULL) {}
2199 ~instantiate_cache_type ();
2200 tree get (unsigned slot) { return entries[slot].chrec; }
2201 void set (unsigned slot, tree chrec) { entries[slot].chrec = chrec; }
2204 instantiate_cache_type::~instantiate_cache_type ()
2206 if (map != NULL)
2208 htab_delete (map);
2209 entries.release ();
2213 /* Cache to avoid infinite recursion when instantiating an SSA name.
2214 Live during the outermost instantiate_scev or resolve_mixers call. */
2215 static instantiate_cache_type *global_cache;
2217 /* Computes a hash function for database element ELT. */
2219 static inline hashval_t
2220 hash_idx_scev_info (const void *elt_)
2222 unsigned idx = ((size_t) elt_) - 2;
2223 return scev_info_hasher::hash (&global_cache->entries[idx]);
2226 /* Compares database elements E1 and E2. */
2228 static inline int
2229 eq_idx_scev_info (const void *e1, const void *e2)
2231 unsigned idx1 = ((size_t) e1) - 2;
2232 return scev_info_hasher::equal (&global_cache->entries[idx1],
2233 (const scev_info_str *) e2);
2236 /* Returns from CACHE the slot number of the cached chrec for NAME. */
2238 static unsigned
2239 get_instantiated_value_entry (instantiate_cache_type &cache,
2240 tree name, basic_block instantiate_below)
2242 if (!cache.map)
2244 cache.map = htab_create (10, hash_idx_scev_info, eq_idx_scev_info, NULL);
2245 cache.entries.create (10);
2248 scev_info_str e;
2249 e.name_version = SSA_NAME_VERSION (name);
2250 e.instantiated_below = instantiate_below->index;
2251 void **slot = htab_find_slot_with_hash (cache.map, &e,
2252 scev_info_hasher::hash (&e), INSERT);
2253 if (!*slot)
2255 e.chrec = chrec_not_analyzed_yet;
2256 *slot = (void *)(size_t)(cache.entries.length () + 2);
2257 cache.entries.safe_push (e);
2260 return ((size_t)*slot) - 2;
2264 /* Return the closed_loop_phi node for VAR. If there is none, return
2265 NULL_TREE. */
2267 static tree
2268 loop_closed_phi_def (tree var)
2270 struct loop *loop;
2271 edge exit;
2272 gphi *phi;
2273 gphi_iterator psi;
2275 if (var == NULL_TREE
2276 || TREE_CODE (var) != SSA_NAME)
2277 return NULL_TREE;
2279 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var));
2280 exit = single_exit (loop);
2281 if (!exit)
2282 return NULL_TREE;
2284 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi))
2286 phi = psi.phi ();
2287 if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var)
2288 return PHI_RESULT (phi);
2291 return NULL_TREE;
2294 static tree instantiate_scev_r (basic_block, struct loop *, struct loop *,
2295 tree, bool, int);
2297 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2298 and EVOLUTION_LOOP, that were left under a symbolic form.
2300 CHREC is an SSA_NAME to be instantiated.
2302 CACHE is the cache of already instantiated values.
2304 FOLD_CONVERSIONS should be set to true when the conversions that
2305 may wrap in signed/pointer type are folded, as long as the value of
2306 the chrec is preserved.
2308 SIZE_EXPR is used for computing the size of the expression to be
2309 instantiated, and to stop if it exceeds some limit. */
2311 static tree
2312 instantiate_scev_name (basic_block instantiate_below,
2313 struct loop *evolution_loop, struct loop *inner_loop,
2314 tree chrec,
2315 bool fold_conversions,
2316 int size_expr)
2318 tree res;
2319 struct loop *def_loop;
2320 basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (chrec));
2322 /* A parameter (or loop invariant and we do not want to include
2323 evolutions in outer loops), nothing to do. */
2324 if (!def_bb
2325 || loop_depth (def_bb->loop_father) == 0
2326 || dominated_by_p (CDI_DOMINATORS, instantiate_below, def_bb))
2327 return chrec;
2329 /* We cache the value of instantiated variable to avoid exponential
2330 time complexity due to reevaluations. We also store the convenient
2331 value in the cache in order to prevent infinite recursion -- we do
2332 not want to instantiate the SSA_NAME if it is in a mixer
2333 structure. This is used for avoiding the instantiation of
2334 recursively defined functions, such as:
2336 | a_2 -> {0, +, 1, +, a_2}_1 */
2338 unsigned si = get_instantiated_value_entry (*global_cache,
2339 chrec, instantiate_below);
2340 if (global_cache->get (si) != chrec_not_analyzed_yet)
2341 return global_cache->get (si);
2343 /* On recursion return chrec_dont_know. */
2344 global_cache->set (si, chrec_dont_know);
2346 def_loop = find_common_loop (evolution_loop, def_bb->loop_father);
2348 /* If the analysis yields a parametric chrec, instantiate the
2349 result again. */
2350 res = analyze_scalar_evolution (def_loop, chrec);
2352 /* Don't instantiate default definitions. */
2353 if (TREE_CODE (res) == SSA_NAME
2354 && SSA_NAME_IS_DEFAULT_DEF (res))
2357 /* Don't instantiate loop-closed-ssa phi nodes. */
2358 else if (TREE_CODE (res) == SSA_NAME
2359 && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res)))
2360 > loop_depth (def_loop))
2362 if (res == chrec)
2363 res = loop_closed_phi_def (chrec);
2364 else
2365 res = chrec;
2367 /* When there is no loop_closed_phi_def, it means that the
2368 variable is not used after the loop: try to still compute the
2369 value of the variable when exiting the loop. */
2370 if (res == NULL_TREE)
2372 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (chrec));
2373 res = analyze_scalar_evolution (loop, chrec);
2374 res = compute_overall_effect_of_inner_loop (loop, res);
2375 res = instantiate_scev_r (instantiate_below, evolution_loop,
2376 inner_loop, res,
2377 fold_conversions, size_expr);
2379 else if (!dominated_by_p (CDI_DOMINATORS, instantiate_below,
2380 gimple_bb (SSA_NAME_DEF_STMT (res))))
2381 res = chrec_dont_know;
2384 else if (res != chrec_dont_know)
2386 if (inner_loop
2387 && def_bb->loop_father != inner_loop
2388 && !flow_loop_nested_p (def_bb->loop_father, inner_loop))
2389 /* ??? We could try to compute the overall effect of the loop here. */
2390 res = chrec_dont_know;
2391 else
2392 res = instantiate_scev_r (instantiate_below, evolution_loop,
2393 inner_loop, res,
2394 fold_conversions, size_expr);
2397 /* Store the correct value to the cache. */
2398 global_cache->set (si, res);
2399 return res;
2402 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2403 and EVOLUTION_LOOP, that were left under a symbolic form.
2405 CHREC is a polynomial chain of recurrence to be instantiated.
2407 CACHE is the cache of already instantiated values.
2409 FOLD_CONVERSIONS should be set to true when the conversions that
2410 may wrap in signed/pointer type are folded, as long as the value of
2411 the chrec is preserved.
2413 SIZE_EXPR is used for computing the size of the expression to be
2414 instantiated, and to stop if it exceeds some limit. */
2416 static tree
2417 instantiate_scev_poly (basic_block instantiate_below,
2418 struct loop *evolution_loop, struct loop *,
2419 tree chrec, bool fold_conversions, int size_expr)
2421 tree op1;
2422 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2423 get_chrec_loop (chrec),
2424 CHREC_LEFT (chrec), fold_conversions,
2425 size_expr);
2426 if (op0 == chrec_dont_know)
2427 return chrec_dont_know;
2429 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2430 get_chrec_loop (chrec),
2431 CHREC_RIGHT (chrec), fold_conversions,
2432 size_expr);
2433 if (op1 == chrec_dont_know)
2434 return chrec_dont_know;
2436 if (CHREC_LEFT (chrec) != op0
2437 || CHREC_RIGHT (chrec) != op1)
2439 op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL);
2440 chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1);
2443 return chrec;
2446 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2447 and EVOLUTION_LOOP, that were left under a symbolic form.
2449 "C0 CODE C1" is a binary expression of type TYPE to be instantiated.
2451 CACHE is the cache of already instantiated values.
2453 FOLD_CONVERSIONS should be set to true when the conversions that
2454 may wrap in signed/pointer type are folded, as long as the value of
2455 the chrec is preserved.
2457 SIZE_EXPR is used for computing the size of the expression to be
2458 instantiated, and to stop if it exceeds some limit. */
2460 static tree
2461 instantiate_scev_binary (basic_block instantiate_below,
2462 struct loop *evolution_loop, struct loop *inner_loop,
2463 tree chrec, enum tree_code code,
2464 tree type, tree c0, tree c1,
2465 bool fold_conversions, int size_expr)
2467 tree op1;
2468 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop,
2469 c0, fold_conversions, size_expr);
2470 if (op0 == chrec_dont_know)
2471 return chrec_dont_know;
2473 op1 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop,
2474 c1, fold_conversions, size_expr);
2475 if (op1 == chrec_dont_know)
2476 return chrec_dont_know;
2478 if (c0 != op0
2479 || c1 != op1)
2481 op0 = chrec_convert (type, op0, NULL);
2482 op1 = chrec_convert_rhs (type, op1, NULL);
2484 switch (code)
2486 case POINTER_PLUS_EXPR:
2487 case PLUS_EXPR:
2488 return chrec_fold_plus (type, op0, op1);
2490 case MINUS_EXPR:
2491 return chrec_fold_minus (type, op0, op1);
2493 case MULT_EXPR:
2494 return chrec_fold_multiply (type, op0, op1);
2496 default:
2497 gcc_unreachable ();
2501 return chrec ? chrec : fold_build2 (code, type, c0, c1);
2504 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2505 and EVOLUTION_LOOP, that were left under a symbolic form.
2507 "CHREC" is an array reference to be instantiated.
2509 CACHE is the cache of already instantiated values.
2511 FOLD_CONVERSIONS should be set to true when the conversions that
2512 may wrap in signed/pointer type are folded, as long as the value of
2513 the chrec is preserved.
2515 SIZE_EXPR is used for computing the size of the expression to be
2516 instantiated, and to stop if it exceeds some limit. */
2518 static tree
2519 instantiate_array_ref (basic_block instantiate_below,
2520 struct loop *evolution_loop, struct loop *inner_loop,
2521 tree chrec, bool fold_conversions, int size_expr)
2523 tree res;
2524 tree index = TREE_OPERAND (chrec, 1);
2525 tree op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2526 inner_loop, index,
2527 fold_conversions, size_expr);
2529 if (op1 == chrec_dont_know)
2530 return chrec_dont_know;
2532 if (chrec && op1 == index)
2533 return chrec;
2535 res = unshare_expr (chrec);
2536 TREE_OPERAND (res, 1) = op1;
2537 return res;
2540 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2541 and EVOLUTION_LOOP, that were left under a symbolic form.
2543 "CHREC" that stands for a convert expression "(TYPE) OP" is to be
2544 instantiated.
2546 CACHE is the cache of already instantiated values.
2548 FOLD_CONVERSIONS should be set to true when the conversions that
2549 may wrap in signed/pointer type are folded, as long as the value of
2550 the chrec is preserved.
2552 SIZE_EXPR is used for computing the size of the expression to be
2553 instantiated, and to stop if it exceeds some limit. */
2555 static tree
2556 instantiate_scev_convert (basic_block instantiate_below,
2557 struct loop *evolution_loop, struct loop *inner_loop,
2558 tree chrec, tree type, tree op,
2559 bool fold_conversions, int size_expr)
2561 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2562 inner_loop, op,
2563 fold_conversions, size_expr);
2565 if (op0 == chrec_dont_know)
2566 return chrec_dont_know;
2568 if (fold_conversions)
2570 tree tmp = chrec_convert_aggressive (type, op0);
2571 if (tmp)
2572 return tmp;
2575 if (chrec && op0 == op)
2576 return chrec;
2578 /* If we used chrec_convert_aggressive, we can no longer assume that
2579 signed chrecs do not overflow, as chrec_convert does, so avoid
2580 calling it in that case. */
2581 if (fold_conversions)
2582 return fold_convert (type, op0);
2584 return chrec_convert (type, op0, NULL);
2587 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2588 and EVOLUTION_LOOP, that were left under a symbolic form.
2590 CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated.
2591 Handle ~X as -1 - X.
2592 Handle -X as -1 * X.
2594 CACHE is the cache of already instantiated values.
2596 FOLD_CONVERSIONS should be set to true when the conversions that
2597 may wrap in signed/pointer type are folded, as long as the value of
2598 the chrec is preserved.
2600 SIZE_EXPR is used for computing the size of the expression to be
2601 instantiated, and to stop if it exceeds some limit. */
2603 static tree
2604 instantiate_scev_not (basic_block instantiate_below,
2605 struct loop *evolution_loop, struct loop *inner_loop,
2606 tree chrec,
2607 enum tree_code code, tree type, tree op,
2608 bool fold_conversions, int size_expr)
2610 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2611 inner_loop, op,
2612 fold_conversions, size_expr);
2614 if (op0 == chrec_dont_know)
2615 return chrec_dont_know;
2617 if (op != op0)
2619 op0 = chrec_convert (type, op0, NULL);
2621 switch (code)
2623 case BIT_NOT_EXPR:
2624 return chrec_fold_minus
2625 (type, fold_convert (type, integer_minus_one_node), op0);
2627 case NEGATE_EXPR:
2628 return chrec_fold_multiply
2629 (type, fold_convert (type, integer_minus_one_node), op0);
2631 default:
2632 gcc_unreachable ();
2636 return chrec ? chrec : fold_build1 (code, type, op0);
2639 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2640 and EVOLUTION_LOOP, that were left under a symbolic form.
2642 CHREC is an expression with 3 operands to be instantiated.
2644 CACHE is the cache of already instantiated values.
2646 FOLD_CONVERSIONS should be set to true when the conversions that
2647 may wrap in signed/pointer type are folded, as long as the value of
2648 the chrec is preserved.
2650 SIZE_EXPR is used for computing the size of the expression to be
2651 instantiated, and to stop if it exceeds some limit. */
2653 static tree
2654 instantiate_scev_3 (basic_block instantiate_below,
2655 struct loop *evolution_loop, struct loop *inner_loop,
2656 tree chrec,
2657 bool fold_conversions, int size_expr)
2659 tree op1, op2;
2660 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2661 inner_loop, TREE_OPERAND (chrec, 0),
2662 fold_conversions, size_expr);
2663 if (op0 == chrec_dont_know)
2664 return chrec_dont_know;
2666 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2667 inner_loop, TREE_OPERAND (chrec, 1),
2668 fold_conversions, size_expr);
2669 if (op1 == chrec_dont_know)
2670 return chrec_dont_know;
2672 op2 = instantiate_scev_r (instantiate_below, evolution_loop,
2673 inner_loop, TREE_OPERAND (chrec, 2),
2674 fold_conversions, size_expr);
2675 if (op2 == chrec_dont_know)
2676 return chrec_dont_know;
2678 if (op0 == TREE_OPERAND (chrec, 0)
2679 && op1 == TREE_OPERAND (chrec, 1)
2680 && op2 == TREE_OPERAND (chrec, 2))
2681 return chrec;
2683 return fold_build3 (TREE_CODE (chrec),
2684 TREE_TYPE (chrec), op0, op1, op2);
2687 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2688 and EVOLUTION_LOOP, that were left under a symbolic form.
2690 CHREC is an expression with 2 operands to be instantiated.
2692 CACHE is the cache of already instantiated values.
2694 FOLD_CONVERSIONS should be set to true when the conversions that
2695 may wrap in signed/pointer type are folded, as long as the value of
2696 the chrec is preserved.
2698 SIZE_EXPR is used for computing the size of the expression to be
2699 instantiated, and to stop if it exceeds some limit. */
2701 static tree
2702 instantiate_scev_2 (basic_block instantiate_below,
2703 struct loop *evolution_loop, struct loop *inner_loop,
2704 tree chrec,
2705 bool fold_conversions, int size_expr)
2707 tree op1;
2708 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2709 inner_loop, TREE_OPERAND (chrec, 0),
2710 fold_conversions, size_expr);
2711 if (op0 == chrec_dont_know)
2712 return chrec_dont_know;
2714 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2715 inner_loop, TREE_OPERAND (chrec, 1),
2716 fold_conversions, size_expr);
2717 if (op1 == chrec_dont_know)
2718 return chrec_dont_know;
2720 if (op0 == TREE_OPERAND (chrec, 0)
2721 && op1 == TREE_OPERAND (chrec, 1))
2722 return chrec;
2724 return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1);
2727 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2728 and EVOLUTION_LOOP, that were left under a symbolic form.
2730 CHREC is an expression with 2 operands to be instantiated.
2732 CACHE is the cache of already instantiated values.
2734 FOLD_CONVERSIONS should be set to true when the conversions that
2735 may wrap in signed/pointer type are folded, as long as the value of
2736 the chrec is preserved.
2738 SIZE_EXPR is used for computing the size of the expression to be
2739 instantiated, and to stop if it exceeds some limit. */
2741 static tree
2742 instantiate_scev_1 (basic_block instantiate_below,
2743 struct loop *evolution_loop, struct loop *inner_loop,
2744 tree chrec,
2745 bool fold_conversions, int size_expr)
2747 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2748 inner_loop, TREE_OPERAND (chrec, 0),
2749 fold_conversions, size_expr);
2751 if (op0 == chrec_dont_know)
2752 return chrec_dont_know;
2754 if (op0 == TREE_OPERAND (chrec, 0))
2755 return chrec;
2757 return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0);
2760 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2761 and EVOLUTION_LOOP, that were left under a symbolic form.
2763 CHREC is the scalar evolution to instantiate.
2765 CACHE is the cache of already instantiated values.
2767 FOLD_CONVERSIONS should be set to true when the conversions that
2768 may wrap in signed/pointer type are folded, as long as the value of
2769 the chrec is preserved.
2771 SIZE_EXPR is used for computing the size of the expression to be
2772 instantiated, and to stop if it exceeds some limit. */
2774 static tree
2775 instantiate_scev_r (basic_block instantiate_below,
2776 struct loop *evolution_loop, struct loop *inner_loop,
2777 tree chrec,
2778 bool fold_conversions, int size_expr)
2780 /* Give up if the expression is larger than the MAX that we allow. */
2781 if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
2782 return chrec_dont_know;
2784 if (chrec == NULL_TREE
2785 || automatically_generated_chrec_p (chrec)
2786 || is_gimple_min_invariant (chrec))
2787 return chrec;
2789 switch (TREE_CODE (chrec))
2791 case SSA_NAME:
2792 return instantiate_scev_name (instantiate_below, evolution_loop,
2793 inner_loop, chrec,
2794 fold_conversions, size_expr);
2796 case POLYNOMIAL_CHREC:
2797 return instantiate_scev_poly (instantiate_below, evolution_loop,
2798 inner_loop, chrec,
2799 fold_conversions, size_expr);
2801 case POINTER_PLUS_EXPR:
2802 case PLUS_EXPR:
2803 case MINUS_EXPR:
2804 case MULT_EXPR:
2805 return instantiate_scev_binary (instantiate_below, evolution_loop,
2806 inner_loop, chrec,
2807 TREE_CODE (chrec), chrec_type (chrec),
2808 TREE_OPERAND (chrec, 0),
2809 TREE_OPERAND (chrec, 1),
2810 fold_conversions, size_expr);
2812 CASE_CONVERT:
2813 return instantiate_scev_convert (instantiate_below, evolution_loop,
2814 inner_loop, chrec,
2815 TREE_TYPE (chrec), TREE_OPERAND (chrec, 0),
2816 fold_conversions, size_expr);
2818 case NEGATE_EXPR:
2819 case BIT_NOT_EXPR:
2820 return instantiate_scev_not (instantiate_below, evolution_loop,
2821 inner_loop, chrec,
2822 TREE_CODE (chrec), TREE_TYPE (chrec),
2823 TREE_OPERAND (chrec, 0),
2824 fold_conversions, size_expr);
2826 case ADDR_EXPR:
2827 case SCEV_NOT_KNOWN:
2828 return chrec_dont_know;
2830 case SCEV_KNOWN:
2831 return chrec_known;
2833 case ARRAY_REF:
2834 return instantiate_array_ref (instantiate_below, evolution_loop,
2835 inner_loop, chrec,
2836 fold_conversions, size_expr);
2838 default:
2839 break;
2842 if (VL_EXP_CLASS_P (chrec))
2843 return chrec_dont_know;
2845 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
2847 case 3:
2848 return instantiate_scev_3 (instantiate_below, evolution_loop,
2849 inner_loop, chrec,
2850 fold_conversions, size_expr);
2852 case 2:
2853 return instantiate_scev_2 (instantiate_below, evolution_loop,
2854 inner_loop, chrec,
2855 fold_conversions, size_expr);
2857 case 1:
2858 return instantiate_scev_1 (instantiate_below, evolution_loop,
2859 inner_loop, chrec,
2860 fold_conversions, size_expr);
2862 case 0:
2863 return chrec;
2865 default:
2866 break;
2869 /* Too complicated to handle. */
2870 return chrec_dont_know;
2873 /* Analyze all the parameters of the chrec that were left under a
2874 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2875 recursive instantiation of parameters: a parameter is a variable
2876 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2877 a function parameter. */
2879 tree
2880 instantiate_scev (basic_block instantiate_below, struct loop *evolution_loop,
2881 tree chrec)
2883 tree res;
2885 if (dump_file && (dump_flags & TDF_SCEV))
2887 fprintf (dump_file, "(instantiate_scev \n");
2888 fprintf (dump_file, " (instantiate_below = %d)\n", instantiate_below->index);
2889 fprintf (dump_file, " (evolution_loop = %d)\n", evolution_loop->num);
2890 fprintf (dump_file, " (chrec = ");
2891 print_generic_expr (dump_file, chrec, 0);
2892 fprintf (dump_file, ")\n");
2895 bool destr = false;
2896 if (!global_cache)
2898 global_cache = new instantiate_cache_type;
2899 destr = true;
2902 res = instantiate_scev_r (instantiate_below, evolution_loop,
2903 NULL, chrec, false, 0);
2905 if (destr)
2907 delete global_cache;
2908 global_cache = NULL;
2911 if (dump_file && (dump_flags & TDF_SCEV))
2913 fprintf (dump_file, " (res = ");
2914 print_generic_expr (dump_file, res, 0);
2915 fprintf (dump_file, "))\n");
2918 return res;
2921 /* Similar to instantiate_parameters, but does not introduce the
2922 evolutions in outer loops for LOOP invariants in CHREC, and does not
2923 care about causing overflows, as long as they do not affect value
2924 of an expression. */
2926 tree
2927 resolve_mixers (struct loop *loop, tree chrec)
2929 bool destr = false;
2930 if (!global_cache)
2932 global_cache = new instantiate_cache_type;
2933 destr = true;
2936 tree ret = instantiate_scev_r (block_before_loop (loop), loop, NULL,
2937 chrec, true, 0);
2939 if (destr)
2941 delete global_cache;
2942 global_cache = NULL;
2945 return ret;
2948 /* Entry point for the analysis of the number of iterations pass.
2949 This function tries to safely approximate the number of iterations
2950 the loop will run. When this property is not decidable at compile
2951 time, the result is chrec_dont_know. Otherwise the result is a
2952 scalar or a symbolic parameter. When the number of iterations may
2953 be equal to zero and the property cannot be determined at compile
2954 time, the result is a COND_EXPR that represents in a symbolic form
2955 the conditions under which the number of iterations is not zero.
2957 Example of analysis: suppose that the loop has an exit condition:
2959 "if (b > 49) goto end_loop;"
2961 and that in a previous analysis we have determined that the
2962 variable 'b' has an evolution function:
2964 "EF = {23, +, 5}_2".
2966 When we evaluate the function at the point 5, i.e. the value of the
2967 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2968 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2969 the loop body has been executed 6 times. */
2971 tree
2972 number_of_latch_executions (struct loop *loop)
2974 edge exit;
2975 struct tree_niter_desc niter_desc;
2976 tree may_be_zero;
2977 tree res;
2979 /* Determine whether the number of iterations in loop has already
2980 been computed. */
2981 res = loop->nb_iterations;
2982 if (res)
2983 return res;
2985 may_be_zero = NULL_TREE;
2987 if (dump_file && (dump_flags & TDF_SCEV))
2988 fprintf (dump_file, "(number_of_iterations_in_loop = \n");
2990 res = chrec_dont_know;
2991 exit = single_exit (loop);
2993 if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false))
2995 may_be_zero = niter_desc.may_be_zero;
2996 res = niter_desc.niter;
2999 if (res == chrec_dont_know
3000 || !may_be_zero
3001 || integer_zerop (may_be_zero))
3003 else if (integer_nonzerop (may_be_zero))
3004 res = build_int_cst (TREE_TYPE (res), 0);
3006 else if (COMPARISON_CLASS_P (may_be_zero))
3007 res = fold_build3 (COND_EXPR, TREE_TYPE (res), may_be_zero,
3008 build_int_cst (TREE_TYPE (res), 0), res);
3009 else
3010 res = chrec_dont_know;
3012 if (dump_file && (dump_flags & TDF_SCEV))
3014 fprintf (dump_file, " (set_nb_iterations_in_loop = ");
3015 print_generic_expr (dump_file, res, 0);
3016 fprintf (dump_file, "))\n");
3019 loop->nb_iterations = res;
3020 return res;
3024 /* Counters for the stats. */
3026 struct chrec_stats
3028 unsigned nb_chrecs;
3029 unsigned nb_affine;
3030 unsigned nb_affine_multivar;
3031 unsigned nb_higher_poly;
3032 unsigned nb_chrec_dont_know;
3033 unsigned nb_undetermined;
3036 /* Reset the counters. */
3038 static inline void
3039 reset_chrecs_counters (struct chrec_stats *stats)
3041 stats->nb_chrecs = 0;
3042 stats->nb_affine = 0;
3043 stats->nb_affine_multivar = 0;
3044 stats->nb_higher_poly = 0;
3045 stats->nb_chrec_dont_know = 0;
3046 stats->nb_undetermined = 0;
3049 /* Dump the contents of a CHREC_STATS structure. */
3051 static void
3052 dump_chrecs_stats (FILE *file, struct chrec_stats *stats)
3054 fprintf (file, "\n(\n");
3055 fprintf (file, "-----------------------------------------\n");
3056 fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine);
3057 fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar);
3058 fprintf (file, "%d\tdegree greater than 2 polynomials\n",
3059 stats->nb_higher_poly);
3060 fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know);
3061 fprintf (file, "-----------------------------------------\n");
3062 fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs);
3063 fprintf (file, "%d\twith undetermined coefficients\n",
3064 stats->nb_undetermined);
3065 fprintf (file, "-----------------------------------------\n");
3066 fprintf (file, "%d\tchrecs in the scev database\n",
3067 (int) scalar_evolution_info->elements ());
3068 fprintf (file, "%d\tsets in the scev database\n", nb_set_scev);
3069 fprintf (file, "%d\tgets in the scev database\n", nb_get_scev);
3070 fprintf (file, "-----------------------------------------\n");
3071 fprintf (file, ")\n\n");
3074 /* Gather statistics about CHREC. */
3076 static void
3077 gather_chrec_stats (tree chrec, struct chrec_stats *stats)
3079 if (dump_file && (dump_flags & TDF_STATS))
3081 fprintf (dump_file, "(classify_chrec ");
3082 print_generic_expr (dump_file, chrec, 0);
3083 fprintf (dump_file, "\n");
3086 stats->nb_chrecs++;
3088 if (chrec == NULL_TREE)
3090 stats->nb_undetermined++;
3091 return;
3094 switch (TREE_CODE (chrec))
3096 case POLYNOMIAL_CHREC:
3097 if (evolution_function_is_affine_p (chrec))
3099 if (dump_file && (dump_flags & TDF_STATS))
3100 fprintf (dump_file, " affine_univariate\n");
3101 stats->nb_affine++;
3103 else if (evolution_function_is_affine_multivariate_p (chrec, 0))
3105 if (dump_file && (dump_flags & TDF_STATS))
3106 fprintf (dump_file, " affine_multivariate\n");
3107 stats->nb_affine_multivar++;
3109 else
3111 if (dump_file && (dump_flags & TDF_STATS))
3112 fprintf (dump_file, " higher_degree_polynomial\n");
3113 stats->nb_higher_poly++;
3116 break;
3118 default:
3119 break;
3122 if (chrec_contains_undetermined (chrec))
3124 if (dump_file && (dump_flags & TDF_STATS))
3125 fprintf (dump_file, " undetermined\n");
3126 stats->nb_undetermined++;
3129 if (dump_file && (dump_flags & TDF_STATS))
3130 fprintf (dump_file, ")\n");
3133 /* Classify the chrecs of the whole database. */
3135 void
3136 gather_stats_on_scev_database (void)
3138 struct chrec_stats stats;
3140 if (!dump_file)
3141 return;
3143 reset_chrecs_counters (&stats);
3145 hash_table<scev_info_hasher>::iterator iter;
3146 scev_info_str *elt;
3147 FOR_EACH_HASH_TABLE_ELEMENT (*scalar_evolution_info, elt, scev_info_str *,
3148 iter)
3149 gather_chrec_stats (elt->chrec, &stats);
3151 dump_chrecs_stats (dump_file, &stats);
3156 /* Initializer. */
3158 static void
3159 initialize_scalar_evolutions_analyzer (void)
3161 /* The elements below are unique. */
3162 if (chrec_dont_know == NULL_TREE)
3164 chrec_not_analyzed_yet = NULL_TREE;
3165 chrec_dont_know = make_node (SCEV_NOT_KNOWN);
3166 chrec_known = make_node (SCEV_KNOWN);
3167 TREE_TYPE (chrec_dont_know) = void_type_node;
3168 TREE_TYPE (chrec_known) = void_type_node;
3172 /* Initialize the analysis of scalar evolutions for LOOPS. */
3174 void
3175 scev_initialize (void)
3177 struct loop *loop;
3179 scalar_evolution_info = hash_table<scev_info_hasher>::create_ggc (100);
3181 initialize_scalar_evolutions_analyzer ();
3183 FOR_EACH_LOOP (loop, 0)
3185 loop->nb_iterations = NULL_TREE;
3189 /* Return true if SCEV is initialized. */
3191 bool
3192 scev_initialized_p (void)
3194 return scalar_evolution_info != NULL;
3197 /* Cleans up the information cached by the scalar evolutions analysis
3198 in the hash table. */
3200 void
3201 scev_reset_htab (void)
3203 if (!scalar_evolution_info)
3204 return;
3206 scalar_evolution_info->empty ();
3209 /* Cleans up the information cached by the scalar evolutions analysis
3210 in the hash table and in the loop->nb_iterations. */
3212 void
3213 scev_reset (void)
3215 struct loop *loop;
3217 scev_reset_htab ();
3219 FOR_EACH_LOOP (loop, 0)
3221 loop->nb_iterations = NULL_TREE;
3225 /* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with
3226 respect to WRTO_LOOP and returns its base and step in IV if possible
3227 (see analyze_scalar_evolution_in_loop for more details on USE_LOOP
3228 and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be
3229 invariant in LOOP. Otherwise we require it to be an integer constant.
3231 IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g.
3232 because it is computed in signed arithmetics). Consequently, adding an
3233 induction variable
3235 for (i = IV->base; ; i += IV->step)
3237 is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is
3238 false for the type of the induction variable, or you can prove that i does
3239 not wrap by some other argument. Otherwise, this might introduce undefined
3240 behavior, and
3242 for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step))
3244 must be used instead. */
3246 bool
3247 simple_iv (struct loop *wrto_loop, struct loop *use_loop, tree op,
3248 affine_iv *iv, bool allow_nonconstant_step)
3250 tree type, ev;
3251 bool folded_casts;
3253 iv->base = NULL_TREE;
3254 iv->step = NULL_TREE;
3255 iv->no_overflow = false;
3257 type = TREE_TYPE (op);
3258 if (!POINTER_TYPE_P (type)
3259 && !INTEGRAL_TYPE_P (type))
3260 return false;
3262 ev = analyze_scalar_evolution_in_loop (wrto_loop, use_loop, op,
3263 &folded_casts);
3264 if (chrec_contains_undetermined (ev)
3265 || chrec_contains_symbols_defined_in_loop (ev, wrto_loop->num))
3266 return false;
3268 if (tree_does_not_contain_chrecs (ev))
3270 iv->base = ev;
3271 iv->step = build_int_cst (TREE_TYPE (ev), 0);
3272 iv->no_overflow = true;
3273 return true;
3276 if (TREE_CODE (ev) != POLYNOMIAL_CHREC
3277 || CHREC_VARIABLE (ev) != (unsigned) wrto_loop->num)
3278 return false;
3280 iv->step = CHREC_RIGHT (ev);
3281 if ((!allow_nonconstant_step && TREE_CODE (iv->step) != INTEGER_CST)
3282 || tree_contains_chrecs (iv->step, NULL))
3283 return false;
3285 iv->base = CHREC_LEFT (ev);
3286 if (tree_contains_chrecs (iv->base, NULL))
3287 return false;
3289 iv->no_overflow = (!folded_casts && ANY_INTEGRAL_TYPE_P (type)
3290 && TYPE_OVERFLOW_UNDEFINED (type));
3292 return true;
3295 /* Finalize the scalar evolution analysis. */
3297 void
3298 scev_finalize (void)
3300 if (!scalar_evolution_info)
3301 return;
3302 scalar_evolution_info->empty ();
3303 scalar_evolution_info = NULL;
3306 /* Returns true if the expression EXPR is considered to be too expensive
3307 for scev_const_prop. */
3309 bool
3310 expression_expensive_p (tree expr)
3312 enum tree_code code;
3314 if (is_gimple_val (expr))
3315 return false;
3317 code = TREE_CODE (expr);
3318 if (code == TRUNC_DIV_EXPR
3319 || code == CEIL_DIV_EXPR
3320 || code == FLOOR_DIV_EXPR
3321 || code == ROUND_DIV_EXPR
3322 || code == TRUNC_MOD_EXPR
3323 || code == CEIL_MOD_EXPR
3324 || code == FLOOR_MOD_EXPR
3325 || code == ROUND_MOD_EXPR
3326 || code == EXACT_DIV_EXPR)
3328 /* Division by power of two is usually cheap, so we allow it.
3329 Forbid anything else. */
3330 if (!integer_pow2p (TREE_OPERAND (expr, 1)))
3331 return true;
3334 switch (TREE_CODE_CLASS (code))
3336 case tcc_binary:
3337 case tcc_comparison:
3338 if (expression_expensive_p (TREE_OPERAND (expr, 1)))
3339 return true;
3341 /* Fallthru. */
3342 case tcc_unary:
3343 return expression_expensive_p (TREE_OPERAND (expr, 0));
3345 default:
3346 return true;
3350 /* Replace ssa names for that scev can prove they are constant by the
3351 appropriate constants. Also perform final value replacement in loops,
3352 in case the replacement expressions are cheap.
3354 We only consider SSA names defined by phi nodes; rest is left to the
3355 ordinary constant propagation pass. */
3357 unsigned int
3358 scev_const_prop (void)
3360 basic_block bb;
3361 tree name, type, ev;
3362 gphi *phi;
3363 gassign *ass;
3364 struct loop *loop, *ex_loop;
3365 bitmap ssa_names_to_remove = NULL;
3366 unsigned i;
3367 gphi_iterator psi;
3369 if (number_of_loops (cfun) <= 1)
3370 return 0;
3372 FOR_EACH_BB_FN (bb, cfun)
3374 loop = bb->loop_father;
3376 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
3378 phi = psi.phi ();
3379 name = PHI_RESULT (phi);
3381 if (virtual_operand_p (name))
3382 continue;
3384 type = TREE_TYPE (name);
3386 if (!POINTER_TYPE_P (type)
3387 && !INTEGRAL_TYPE_P (type))
3388 continue;
3390 ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name));
3391 if (!is_gimple_min_invariant (ev)
3392 || !may_propagate_copy (name, ev))
3393 continue;
3395 /* Replace the uses of the name. */
3396 if (name != ev)
3397 replace_uses_by (name, ev);
3399 if (!ssa_names_to_remove)
3400 ssa_names_to_remove = BITMAP_ALLOC (NULL);
3401 bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name));
3405 /* Remove the ssa names that were replaced by constants. We do not
3406 remove them directly in the previous cycle, since this
3407 invalidates scev cache. */
3408 if (ssa_names_to_remove)
3410 bitmap_iterator bi;
3412 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi)
3414 gimple_stmt_iterator psi;
3415 name = ssa_name (i);
3416 phi = as_a <gphi *> (SSA_NAME_DEF_STMT (name));
3418 gcc_assert (gimple_code (phi) == GIMPLE_PHI);
3419 psi = gsi_for_stmt (phi);
3420 remove_phi_node (&psi, true);
3423 BITMAP_FREE (ssa_names_to_remove);
3424 scev_reset ();
3427 /* Now the regular final value replacement. */
3428 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST)
3430 edge exit;
3431 tree def, rslt, niter;
3432 gimple_stmt_iterator gsi;
3434 /* If we do not know exact number of iterations of the loop, we cannot
3435 replace the final value. */
3436 exit = single_exit (loop);
3437 if (!exit)
3438 continue;
3440 niter = number_of_latch_executions (loop);
3441 if (niter == chrec_dont_know)
3442 continue;
3444 /* Ensure that it is possible to insert new statements somewhere. */
3445 if (!single_pred_p (exit->dest))
3446 split_loop_exit_edge (exit);
3447 gsi = gsi_after_labels (exit->dest);
3449 ex_loop = superloop_at_depth (loop,
3450 loop_depth (exit->dest->loop_father) + 1);
3452 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); )
3454 phi = psi.phi ();
3455 rslt = PHI_RESULT (phi);
3456 def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
3457 if (virtual_operand_p (def))
3459 gsi_next (&psi);
3460 continue;
3463 if (!POINTER_TYPE_P (TREE_TYPE (def))
3464 && !INTEGRAL_TYPE_P (TREE_TYPE (def)))
3466 gsi_next (&psi);
3467 continue;
3470 bool folded_casts;
3471 def = analyze_scalar_evolution_in_loop (ex_loop, loop, def,
3472 &folded_casts);
3473 def = compute_overall_effect_of_inner_loop (ex_loop, def);
3474 if (!tree_does_not_contain_chrecs (def)
3475 || chrec_contains_symbols_defined_in_loop (def, ex_loop->num)
3476 /* Moving the computation from the loop may prolong life range
3477 of some ssa names, which may cause problems if they appear
3478 on abnormal edges. */
3479 || contains_abnormal_ssa_name_p (def)
3480 /* Do not emit expensive expressions. The rationale is that
3481 when someone writes a code like
3483 while (n > 45) n -= 45;
3485 he probably knows that n is not large, and does not want it
3486 to be turned into n %= 45. */
3487 || expression_expensive_p (def))
3489 if (dump_file && (dump_flags & TDF_DETAILS))
3491 fprintf (dump_file, "not replacing:\n ");
3492 print_gimple_stmt (dump_file, phi, 0, 0);
3493 fprintf (dump_file, "\n");
3495 gsi_next (&psi);
3496 continue;
3499 /* Eliminate the PHI node and replace it by a computation outside
3500 the loop. */
3501 if (dump_file)
3503 fprintf (dump_file, "\nfinal value replacement:\n ");
3504 print_gimple_stmt (dump_file, phi, 0, 0);
3505 fprintf (dump_file, " with\n ");
3507 def = unshare_expr (def);
3508 remove_phi_node (&psi, false);
3510 /* If def's type has undefined overflow and there were folded
3511 casts, rewrite all stmts added for def into arithmetics
3512 with defined overflow behavior. */
3513 if (folded_casts && ANY_INTEGRAL_TYPE_P (TREE_TYPE (def))
3514 && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (def)))
3516 gimple_seq stmts;
3517 gimple_stmt_iterator gsi2;
3518 def = force_gimple_operand (def, &stmts, true, NULL_TREE);
3519 gsi2 = gsi_start (stmts);
3520 while (!gsi_end_p (gsi2))
3522 gimple stmt = gsi_stmt (gsi2);
3523 gimple_stmt_iterator gsi3 = gsi2;
3524 gsi_next (&gsi2);
3525 gsi_remove (&gsi3, false);
3526 if (is_gimple_assign (stmt)
3527 && arith_code_with_undefined_signed_overflow
3528 (gimple_assign_rhs_code (stmt)))
3529 gsi_insert_seq_before (&gsi,
3530 rewrite_to_defined_overflow (stmt),
3531 GSI_SAME_STMT);
3532 else
3533 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT);
3536 else
3537 def = force_gimple_operand_gsi (&gsi, def, false, NULL_TREE,
3538 true, GSI_SAME_STMT);
3540 ass = gimple_build_assign (rslt, def);
3541 gsi_insert_before (&gsi, ass, GSI_SAME_STMT);
3542 if (dump_file)
3544 print_gimple_stmt (dump_file, ass, 0, 0);
3545 fprintf (dump_file, "\n");
3549 return 0;
3552 #include "gt-tree-scalar-evolution.h"