dma: bump man page date
[dragonfly.git] / contrib / gcc-4.4 / gcc / tree-scalar-evolution.c
blob54fb5a880cf891496eaac8cb4a284cba00221f69
1 /* Scalar evolution detector.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009
3 Free Software Foundation, Inc.
4 Contributed by Sebastian Pop <s.pop@laposte.net>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /*
23 Description:
25 This pass analyzes the evolution of scalar variables in loop
26 structures. The algorithm is based on the SSA representation,
27 and on the loop hierarchy tree. This algorithm is not based on
28 the notion of versions of a variable, as it was the case for the
29 previous implementations of the scalar evolution algorithm, but
30 it assumes that each defined name is unique.
32 The notation used in this file is called "chains of recurrences",
33 and has been proposed by Eugene Zima, Robert Van Engelen, and
34 others for describing induction variables in programs. For example
35 "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0
36 when entering in the loop_1 and has a step 2 in this loop, in other
37 words "for (b = 0; b < N; b+=2);". Note that the coefficients of
38 this chain of recurrence (or chrec [shrek]) can contain the name of
39 other variables, in which case they are called parametric chrecs.
40 For example, "b -> {a, +, 2}_1" means that the initial value of "b"
41 is the value of "a". In most of the cases these parametric chrecs
42 are fully instantiated before their use because symbolic names can
43 hide some difficult cases such as self-references described later
44 (see the Fibonacci example).
46 A short sketch of the algorithm is:
48 Given a scalar variable to be analyzed, follow the SSA edge to
49 its definition:
51 - When the definition is a GIMPLE_ASSIGN: if the right hand side
52 (RHS) of the definition cannot be statically analyzed, the answer
53 of the analyzer is: "don't know".
54 Otherwise, for all the variables that are not yet analyzed in the
55 RHS, try to determine their evolution, and finally try to
56 evaluate the operation of the RHS that gives the evolution
57 function of the analyzed variable.
59 - When the definition is a condition-phi-node: determine the
60 evolution function for all the branches of the phi node, and
61 finally merge these evolutions (see chrec_merge).
63 - When the definition is a loop-phi-node: determine its initial
64 condition, that is the SSA edge defined in an outer loop, and
65 keep it symbolic. Then determine the SSA edges that are defined
66 in the body of the loop. Follow the inner edges until ending on
67 another loop-phi-node of the same analyzed loop. If the reached
68 loop-phi-node is not the starting loop-phi-node, then we keep
69 this definition under a symbolic form. If the reached
70 loop-phi-node is the same as the starting one, then we compute a
71 symbolic stride on the return path. The result is then the
72 symbolic chrec {initial_condition, +, symbolic_stride}_loop.
74 Examples:
76 Example 1: Illustration of the basic algorithm.
78 | a = 3
79 | loop_1
80 | b = phi (a, c)
81 | c = b + 1
82 | if (c > 10) exit_loop
83 | endloop
85 Suppose that we want to know the number of iterations of the
86 loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We
87 ask the scalar evolution analyzer two questions: what's the
88 scalar evolution (scev) of "c", and what's the scev of "10". For
89 "10" the answer is "10" since it is a scalar constant. For the
90 scalar variable "c", it follows the SSA edge to its definition,
91 "c = b + 1", and then asks again what's the scev of "b".
92 Following the SSA edge, we end on a loop-phi-node "b = phi (a,
93 c)", where the initial condition is "a", and the inner loop edge
94 is "c". The initial condition is kept under a symbolic form (it
95 may be the case that the copy constant propagation has done its
96 work and we end with the constant "3" as one of the edges of the
97 loop-phi-node). The update edge is followed to the end of the
98 loop, and until reaching again the starting loop-phi-node: b -> c
99 -> b. At this point we have drawn a path from "b" to "b" from
100 which we compute the stride in the loop: in this example it is
101 "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now
102 that the scev for "b" is known, it is possible to compute the
103 scev for "c", that is "c -> {a + 1, +, 1}_1". In order to
104 determine the number of iterations in the loop_1, we have to
105 instantiate_parameters (loop_1, {a + 1, +, 1}_1), that gives after some
106 more analysis the scev {4, +, 1}_1, or in other words, this is
107 the function "f (x) = x + 4", where x is the iteration count of
108 the loop_1. Now we have to solve the inequality "x + 4 > 10",
109 and take the smallest iteration number for which the loop is
110 exited: x = 7. This loop runs from x = 0 to x = 7, and in total
111 there are 8 iterations. In terms of loop normalization, we have
112 created a variable that is implicitly defined, "x" or just "_1",
113 and all the other analyzed scalars of the loop are defined in
114 function of this variable:
116 a -> 3
117 b -> {3, +, 1}_1
118 c -> {4, +, 1}_1
120 or in terms of a C program:
122 | a = 3
123 | for (x = 0; x <= 7; x++)
125 | b = x + 3
126 | c = x + 4
129 Example 2a: Illustration of the algorithm on nested loops.
131 | loop_1
132 | a = phi (1, b)
133 | c = a + 2
134 | loop_2 10 times
135 | b = phi (c, d)
136 | d = b + 3
137 | endloop
138 | endloop
140 For analyzing the scalar evolution of "a", the algorithm follows
141 the SSA edge into the loop's body: "a -> b". "b" is an inner
142 loop-phi-node, and its analysis as in Example 1, gives:
144 b -> {c, +, 3}_2
145 d -> {c + 3, +, 3}_2
147 Following the SSA edge for the initial condition, we end on "c = a
148 + 2", and then on the starting loop-phi-node "a". From this point,
149 the loop stride is computed: back on "c = a + 2" we get a "+2" in
150 the loop_1, then on the loop-phi-node "b" we compute the overall
151 effect of the inner loop that is "b = c + 30", and we get a "+30"
152 in the loop_1. That means that the overall stride in loop_1 is
153 equal to "+32", and the result is:
155 a -> {1, +, 32}_1
156 c -> {3, +, 32}_1
158 Example 2b: Multivariate chains of recurrences.
160 | loop_1
161 | k = phi (0, k + 1)
162 | loop_2 4 times
163 | j = phi (0, j + 1)
164 | loop_3 4 times
165 | i = phi (0, i + 1)
166 | A[j + k] = ...
167 | endloop
168 | endloop
169 | endloop
171 Analyzing the access function of array A with
172 instantiate_parameters (loop_1, "j + k"), we obtain the
173 instantiation and the analysis of the scalar variables "j" and "k"
174 in loop_1. This leads to the scalar evolution {4, +, 1}_1: the end
175 value of loop_2 for "j" is 4, and the evolution of "k" in loop_1 is
176 {0, +, 1}_1. To obtain the evolution function in loop_3 and
177 instantiate the scalar variables up to loop_1, one has to use:
178 instantiate_scev (block_before_loop (loop_1), loop_3, "j + k").
179 The result of this call is {{0, +, 1}_1, +, 1}_2.
181 Example 3: Higher degree polynomials.
183 | loop_1
184 | a = phi (2, b)
185 | c = phi (5, d)
186 | b = a + 1
187 | d = c + a
188 | endloop
190 a -> {2, +, 1}_1
191 b -> {3, +, 1}_1
192 c -> {5, +, a}_1
193 d -> {5 + a, +, a}_1
195 instantiate_parameters (loop_1, {5, +, a}_1) -> {5, +, 2, +, 1}_1
196 instantiate_parameters (loop_1, {5 + a, +, a}_1) -> {7, +, 3, +, 1}_1
198 Example 4: Lucas, Fibonacci, or mixers in general.
200 | loop_1
201 | a = phi (1, b)
202 | c = phi (3, d)
203 | b = c
204 | d = c + a
205 | endloop
207 a -> (1, c)_1
208 c -> {3, +, a}_1
210 The syntax "(1, c)_1" stands for a PEELED_CHREC that has the
211 following semantics: during the first iteration of the loop_1, the
212 variable contains the value 1, and then it contains the value "c".
213 Note that this syntax is close to the syntax of the loop-phi-node:
214 "a -> (1, c)_1" vs. "a = phi (1, c)".
216 The symbolic chrec representation contains all the semantics of the
217 original code. What is more difficult is to use this information.
219 Example 5: Flip-flops, or exchangers.
221 | loop_1
222 | a = phi (1, b)
223 | c = phi (3, d)
224 | b = c
225 | d = a
226 | endloop
228 a -> (1, c)_1
229 c -> (3, a)_1
231 Based on these symbolic chrecs, it is possible to refine this
232 information into the more precise PERIODIC_CHRECs:
234 a -> |1, 3|_1
235 c -> |3, 1|_1
237 This transformation is not yet implemented.
239 Further readings:
241 You can find a more detailed description of the algorithm in:
242 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf
243 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that
244 this is a preliminary report and some of the details of the
245 algorithm have changed. I'm working on a research report that
246 updates the description of the algorithms to reflect the design
247 choices used in this implementation.
249 A set of slides show a high level overview of the algorithm and run
250 an example through the scalar evolution analyzer:
251 http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf
253 The slides that I have presented at the GCC Summit'04 are available
254 at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf
257 #include "config.h"
258 #include "system.h"
259 #include "coretypes.h"
260 #include "tm.h"
261 #include "ggc.h"
262 #include "tree.h"
263 #include "real.h"
265 /* These RTL headers are needed for basic-block.h. */
266 #include "rtl.h"
267 #include "basic-block.h"
268 #include "diagnostic.h"
269 #include "tree-flow.h"
270 #include "tree-dump.h"
271 #include "timevar.h"
272 #include "cfgloop.h"
273 #include "tree-chrec.h"
274 #include "tree-scalar-evolution.h"
275 #include "tree-pass.h"
276 #include "flags.h"
277 #include "params.h"
279 static tree analyze_scalar_evolution_1 (struct loop *, tree, tree);
281 /* The cached information about an SSA name VAR, claiming that below
282 basic block INSTANTIATED_BELOW, the value of VAR can be expressed
283 as CHREC. */
285 struct scev_info_str GTY(())
287 basic_block instantiated_below;
288 tree var;
289 tree chrec;
292 /* Counters for the scev database. */
293 static unsigned nb_set_scev = 0;
294 static unsigned nb_get_scev = 0;
296 /* The following trees are unique elements. Thus the comparison of
297 another element to these elements should be done on the pointer to
298 these trees, and not on their value. */
300 /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
301 tree chrec_not_analyzed_yet;
303 /* Reserved to the cases where the analyzer has detected an
304 undecidable property at compile time. */
305 tree chrec_dont_know;
307 /* When the analyzer has detected that a property will never
308 happen, then it qualifies it with chrec_known. */
309 tree chrec_known;
311 static GTY ((param_is (struct scev_info_str))) htab_t scalar_evolution_info;
314 /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */
316 static inline struct scev_info_str *
317 new_scev_info_str (basic_block instantiated_below, tree var)
319 struct scev_info_str *res;
321 res = GGC_NEW (struct scev_info_str);
322 res->var = var;
323 res->chrec = chrec_not_analyzed_yet;
324 res->instantiated_below = instantiated_below;
326 return res;
329 /* Computes a hash function for database element ELT. */
331 static hashval_t
332 hash_scev_info (const void *elt)
334 return SSA_NAME_VERSION (((const struct scev_info_str *) elt)->var);
337 /* Compares database elements E1 and E2. */
339 static int
340 eq_scev_info (const void *e1, const void *e2)
342 const struct scev_info_str *elt1 = (const struct scev_info_str *) e1;
343 const struct scev_info_str *elt2 = (const struct scev_info_str *) e2;
345 return (elt1->var == elt2->var
346 && elt1->instantiated_below == elt2->instantiated_below);
349 /* Deletes database element E. */
351 static void
352 del_scev_info (void *e)
354 ggc_free (e);
357 /* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block.
358 A first query on VAR returns chrec_not_analyzed_yet. */
360 static tree *
361 find_var_scev_info (basic_block instantiated_below, tree var)
363 struct scev_info_str *res;
364 struct scev_info_str tmp;
365 PTR *slot;
367 tmp.var = var;
368 tmp.instantiated_below = instantiated_below;
369 slot = htab_find_slot (scalar_evolution_info, &tmp, INSERT);
371 if (!*slot)
372 *slot = new_scev_info_str (instantiated_below, var);
373 res = (struct scev_info_str *) *slot;
375 return &res->chrec;
378 /* Return true when CHREC contains symbolic names defined in
379 LOOP_NB. */
381 bool
382 chrec_contains_symbols_defined_in_loop (const_tree chrec, unsigned loop_nb)
384 int i, n;
386 if (chrec == NULL_TREE)
387 return false;
389 if (is_gimple_min_invariant (chrec))
390 return false;
392 if (TREE_CODE (chrec) == VAR_DECL
393 || TREE_CODE (chrec) == PARM_DECL
394 || TREE_CODE (chrec) == FUNCTION_DECL
395 || TREE_CODE (chrec) == LABEL_DECL
396 || TREE_CODE (chrec) == RESULT_DECL
397 || TREE_CODE (chrec) == FIELD_DECL)
398 return true;
400 if (TREE_CODE (chrec) == SSA_NAME)
402 gimple def = SSA_NAME_DEF_STMT (chrec);
403 struct loop *def_loop = loop_containing_stmt (def);
404 struct loop *loop = get_loop (loop_nb);
406 if (def_loop == NULL)
407 return false;
409 if (loop == def_loop || flow_loop_nested_p (loop, def_loop))
410 return true;
412 return false;
415 n = TREE_OPERAND_LENGTH (chrec);
416 for (i = 0; i < n; i++)
417 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, i),
418 loop_nb))
419 return true;
420 return false;
423 /* Return true when PHI is a loop-phi-node. */
425 static bool
426 loop_phi_node_p (gimple phi)
428 /* The implementation of this function is based on the following
429 property: "all the loop-phi-nodes of a loop are contained in the
430 loop's header basic block". */
432 return loop_containing_stmt (phi)->header == gimple_bb (phi);
435 /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
436 In general, in the case of multivariate evolutions we want to get
437 the evolution in different loops. LOOP specifies the level for
438 which to get the evolution.
440 Example:
442 | for (j = 0; j < 100; j++)
444 | for (k = 0; k < 100; k++)
446 | i = k + j; - Here the value of i is a function of j, k.
448 | ... = i - Here the value of i is a function of j.
450 | ... = i - Here the value of i is a scalar.
452 Example:
454 | i_0 = ...
455 | loop_1 10 times
456 | i_1 = phi (i_0, i_2)
457 | i_2 = i_1 + 2
458 | endloop
460 This loop has the same effect as:
461 LOOP_1 has the same effect as:
463 | i_1 = i_0 + 20
465 The overall effect of the loop, "i_0 + 20" in the previous example,
466 is obtained by passing in the parameters: LOOP = 1,
467 EVOLUTION_FN = {i_0, +, 2}_1.
470 static tree
471 compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn)
473 bool val = false;
475 if (evolution_fn == chrec_dont_know)
476 return chrec_dont_know;
478 else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC)
480 struct loop *inner_loop = get_chrec_loop (evolution_fn);
482 if (inner_loop == loop
483 || flow_loop_nested_p (loop, inner_loop))
485 tree nb_iter = number_of_latch_executions (inner_loop);
487 if (nb_iter == chrec_dont_know)
488 return chrec_dont_know;
489 else
491 tree res;
493 /* evolution_fn is the evolution function in LOOP. Get
494 its value in the nb_iter-th iteration. */
495 res = chrec_apply (inner_loop->num, evolution_fn, nb_iter);
497 /* Continue the computation until ending on a parent of LOOP. */
498 return compute_overall_effect_of_inner_loop (loop, res);
501 else
502 return evolution_fn;
505 /* If the evolution function is an invariant, there is nothing to do. */
506 else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val)
507 return evolution_fn;
509 else
510 return chrec_dont_know;
513 /* Determine whether the CHREC is always positive/negative. If the expression
514 cannot be statically analyzed, return false, otherwise set the answer into
515 VALUE. */
517 bool
518 chrec_is_positive (tree chrec, bool *value)
520 bool value0, value1, value2;
521 tree end_value, nb_iter;
523 switch (TREE_CODE (chrec))
525 case POLYNOMIAL_CHREC:
526 if (!chrec_is_positive (CHREC_LEFT (chrec), &value0)
527 || !chrec_is_positive (CHREC_RIGHT (chrec), &value1))
528 return false;
530 /* FIXME -- overflows. */
531 if (value0 == value1)
533 *value = value0;
534 return true;
537 /* Otherwise the chrec is under the form: "{-197, +, 2}_1",
538 and the proof consists in showing that the sign never
539 changes during the execution of the loop, from 0 to
540 loop->nb_iterations. */
541 if (!evolution_function_is_affine_p (chrec))
542 return false;
544 nb_iter = number_of_latch_executions (get_chrec_loop (chrec));
545 if (chrec_contains_undetermined (nb_iter))
546 return false;
548 #if 0
549 /* TODO -- If the test is after the exit, we may decrease the number of
550 iterations by one. */
551 if (after_exit)
552 nb_iter = chrec_fold_minus (type, nb_iter, build_int_cst (type, 1));
553 #endif
555 end_value = chrec_apply (CHREC_VARIABLE (chrec), chrec, nb_iter);
557 if (!chrec_is_positive (end_value, &value2))
558 return false;
560 *value = value0;
561 return value0 == value1;
563 case INTEGER_CST:
564 *value = (tree_int_cst_sgn (chrec) == 1);
565 return true;
567 default:
568 return false;
572 /* Associate CHREC to SCALAR. */
574 static void
575 set_scalar_evolution (basic_block instantiated_below, tree scalar, tree chrec)
577 tree *scalar_info;
579 if (TREE_CODE (scalar) != SSA_NAME)
580 return;
582 scalar_info = find_var_scev_info (instantiated_below, scalar);
584 if (dump_file)
586 if (dump_flags & TDF_DETAILS)
588 fprintf (dump_file, "(set_scalar_evolution \n");
589 fprintf (dump_file, " instantiated_below = %d \n",
590 instantiated_below->index);
591 fprintf (dump_file, " (scalar = ");
592 print_generic_expr (dump_file, scalar, 0);
593 fprintf (dump_file, ")\n (scalar_evolution = ");
594 print_generic_expr (dump_file, chrec, 0);
595 fprintf (dump_file, "))\n");
597 if (dump_flags & TDF_STATS)
598 nb_set_scev++;
601 *scalar_info = chrec;
604 /* Retrieve the chrec associated to SCALAR instantiated below
605 INSTANTIATED_BELOW block. */
607 static tree
608 get_scalar_evolution (basic_block instantiated_below, tree scalar)
610 tree res;
612 if (dump_file)
614 if (dump_flags & TDF_DETAILS)
616 fprintf (dump_file, "(get_scalar_evolution \n");
617 fprintf (dump_file, " (scalar = ");
618 print_generic_expr (dump_file, scalar, 0);
619 fprintf (dump_file, ")\n");
621 if (dump_flags & TDF_STATS)
622 nb_get_scev++;
625 switch (TREE_CODE (scalar))
627 case SSA_NAME:
628 res = *find_var_scev_info (instantiated_below, scalar);
629 break;
631 case REAL_CST:
632 case FIXED_CST:
633 case INTEGER_CST:
634 res = scalar;
635 break;
637 default:
638 res = chrec_not_analyzed_yet;
639 break;
642 if (dump_file && (dump_flags & TDF_DETAILS))
644 fprintf (dump_file, " (scalar_evolution = ");
645 print_generic_expr (dump_file, res, 0);
646 fprintf (dump_file, "))\n");
649 return res;
652 /* Helper function for add_to_evolution. Returns the evolution
653 function for an assignment of the form "a = b + c", where "a" and
654 "b" are on the strongly connected component. CHREC_BEFORE is the
655 information that we already have collected up to this point.
656 TO_ADD is the evolution of "c".
658 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
659 evolution the expression TO_ADD, otherwise construct an evolution
660 part for this loop. */
662 static tree
663 add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add,
664 gimple at_stmt)
666 tree type, left, right;
667 struct loop *loop = get_loop (loop_nb), *chloop;
669 switch (TREE_CODE (chrec_before))
671 case POLYNOMIAL_CHREC:
672 chloop = get_chrec_loop (chrec_before);
673 if (chloop == loop
674 || flow_loop_nested_p (chloop, loop))
676 unsigned var;
678 type = chrec_type (chrec_before);
680 /* When there is no evolution part in this loop, build it. */
681 if (chloop != loop)
683 var = loop_nb;
684 left = chrec_before;
685 right = SCALAR_FLOAT_TYPE_P (type)
686 ? build_real (type, dconst0)
687 : build_int_cst (type, 0);
689 else
691 var = CHREC_VARIABLE (chrec_before);
692 left = CHREC_LEFT (chrec_before);
693 right = CHREC_RIGHT (chrec_before);
696 to_add = chrec_convert (type, to_add, at_stmt);
697 right = chrec_convert_rhs (type, right, at_stmt);
698 right = chrec_fold_plus (chrec_type (right), right, to_add);
699 return build_polynomial_chrec (var, left, right);
701 else
703 gcc_assert (flow_loop_nested_p (loop, chloop));
705 /* Search the evolution in LOOP_NB. */
706 left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before),
707 to_add, at_stmt);
708 right = CHREC_RIGHT (chrec_before);
709 right = chrec_convert_rhs (chrec_type (left), right, at_stmt);
710 return build_polynomial_chrec (CHREC_VARIABLE (chrec_before),
711 left, right);
714 default:
715 /* These nodes do not depend on a loop. */
716 if (chrec_before == chrec_dont_know)
717 return chrec_dont_know;
719 left = chrec_before;
720 right = chrec_convert_rhs (chrec_type (left), to_add, at_stmt);
721 return build_polynomial_chrec (loop_nb, left, right);
725 /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
726 of LOOP_NB.
728 Description (provided for completeness, for those who read code in
729 a plane, and for my poor 62 bytes brain that would have forgotten
730 all this in the next two or three months):
732 The algorithm of translation of programs from the SSA representation
733 into the chrecs syntax is based on a pattern matching. After having
734 reconstructed the overall tree expression for a loop, there are only
735 two cases that can arise:
737 1. a = loop-phi (init, a + expr)
738 2. a = loop-phi (init, expr)
740 where EXPR is either a scalar constant with respect to the analyzed
741 loop (this is a degree 0 polynomial), or an expression containing
742 other loop-phi definitions (these are higher degree polynomials).
744 Examples:
747 | init = ...
748 | loop_1
749 | a = phi (init, a + 5)
750 | endloop
753 | inita = ...
754 | initb = ...
755 | loop_1
756 | a = phi (inita, 2 * b + 3)
757 | b = phi (initb, b + 1)
758 | endloop
760 For the first case, the semantics of the SSA representation is:
762 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
764 that is, there is a loop index "x" that determines the scalar value
765 of the variable during the loop execution. During the first
766 iteration, the value is that of the initial condition INIT, while
767 during the subsequent iterations, it is the sum of the initial
768 condition with the sum of all the values of EXPR from the initial
769 iteration to the before last considered iteration.
771 For the second case, the semantics of the SSA program is:
773 | a (x) = init, if x = 0;
774 | expr (x - 1), otherwise.
776 The second case corresponds to the PEELED_CHREC, whose syntax is
777 close to the syntax of a loop-phi-node:
779 | phi (init, expr) vs. (init, expr)_x
781 The proof of the translation algorithm for the first case is a
782 proof by structural induction based on the degree of EXPR.
784 Degree 0:
785 When EXPR is a constant with respect to the analyzed loop, or in
786 other words when EXPR is a polynomial of degree 0, the evolution of
787 the variable A in the loop is an affine function with an initial
788 condition INIT, and a step EXPR. In order to show this, we start
789 from the semantics of the SSA representation:
791 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
793 and since "expr (j)" is a constant with respect to "j",
795 f (x) = init + x * expr
797 Finally, based on the semantics of the pure sum chrecs, by
798 identification we get the corresponding chrecs syntax:
800 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
801 f (x) -> {init, +, expr}_x
803 Higher degree:
804 Suppose that EXPR is a polynomial of degree N with respect to the
805 analyzed loop_x for which we have already determined that it is
806 written under the chrecs syntax:
808 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
810 We start from the semantics of the SSA program:
812 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
814 | f (x) = init + \sum_{j = 0}^{x - 1}
815 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
817 | f (x) = init + \sum_{j = 0}^{x - 1}
818 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
820 | f (x) = init + \sum_{k = 0}^{n - 1}
821 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
823 | f (x) = init + \sum_{k = 0}^{n - 1}
824 | (b_k * \binom{x}{k + 1})
826 | f (x) = init + b_0 * \binom{x}{1} + ...
827 | + b_{n-1} * \binom{x}{n}
829 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
830 | + b_{n-1} * \binom{x}{n}
833 And finally from the definition of the chrecs syntax, we identify:
834 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
836 This shows the mechanism that stands behind the add_to_evolution
837 function. An important point is that the use of symbolic
838 parameters avoids the need of an analysis schedule.
840 Example:
842 | inita = ...
843 | initb = ...
844 | loop_1
845 | a = phi (inita, a + 2 + b)
846 | b = phi (initb, b + 1)
847 | endloop
849 When analyzing "a", the algorithm keeps "b" symbolically:
851 | a -> {inita, +, 2 + b}_1
853 Then, after instantiation, the analyzer ends on the evolution:
855 | a -> {inita, +, 2 + initb, +, 1}_1
859 static tree
860 add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code,
861 tree to_add, gimple at_stmt)
863 tree type = chrec_type (to_add);
864 tree res = NULL_TREE;
866 if (to_add == NULL_TREE)
867 return chrec_before;
869 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
870 instantiated at this point. */
871 if (TREE_CODE (to_add) == POLYNOMIAL_CHREC)
872 /* This should not happen. */
873 return chrec_dont_know;
875 if (dump_file && (dump_flags & TDF_DETAILS))
877 fprintf (dump_file, "(add_to_evolution \n");
878 fprintf (dump_file, " (loop_nb = %d)\n", loop_nb);
879 fprintf (dump_file, " (chrec_before = ");
880 print_generic_expr (dump_file, chrec_before, 0);
881 fprintf (dump_file, ")\n (to_add = ");
882 print_generic_expr (dump_file, to_add, 0);
883 fprintf (dump_file, ")\n");
886 if (code == MINUS_EXPR)
887 to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type)
888 ? build_real (type, dconstm1)
889 : build_int_cst_type (type, -1));
891 res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt);
893 if (dump_file && (dump_flags & TDF_DETAILS))
895 fprintf (dump_file, " (res = ");
896 print_generic_expr (dump_file, res, 0);
897 fprintf (dump_file, "))\n");
900 return res;
903 /* Helper function. */
905 static inline tree
906 set_nb_iterations_in_loop (struct loop *loop,
907 tree res)
909 if (dump_file && (dump_flags & TDF_DETAILS))
911 fprintf (dump_file, " (set_nb_iterations_in_loop = ");
912 print_generic_expr (dump_file, res, 0);
913 fprintf (dump_file, "))\n");
916 loop->nb_iterations = res;
917 return res;
922 /* This section selects the loops that will be good candidates for the
923 scalar evolution analysis. For the moment, greedily select all the
924 loop nests we could analyze. */
926 /* For a loop with a single exit edge, return the COND_EXPR that
927 guards the exit edge. If the expression is too difficult to
928 analyze, then give up. */
930 gimple
931 get_loop_exit_condition (const struct loop *loop)
933 gimple res = NULL;
934 edge exit_edge = single_exit (loop);
936 if (dump_file && (dump_flags & TDF_DETAILS))
937 fprintf (dump_file, "(get_loop_exit_condition \n ");
939 if (exit_edge)
941 gimple stmt;
943 stmt = last_stmt (exit_edge->src);
944 if (gimple_code (stmt) == GIMPLE_COND)
945 res = stmt;
948 if (dump_file && (dump_flags & TDF_DETAILS))
950 print_gimple_stmt (dump_file, res, 0, 0);
951 fprintf (dump_file, ")\n");
954 return res;
957 /* Recursively determine and enqueue the exit conditions for a loop. */
959 static void
960 get_exit_conditions_rec (struct loop *loop,
961 VEC(gimple,heap) **exit_conditions)
963 if (!loop)
964 return;
966 /* Recurse on the inner loops, then on the next (sibling) loops. */
967 get_exit_conditions_rec (loop->inner, exit_conditions);
968 get_exit_conditions_rec (loop->next, exit_conditions);
970 if (single_exit (loop))
972 gimple loop_condition = get_loop_exit_condition (loop);
974 if (loop_condition)
975 VEC_safe_push (gimple, heap, *exit_conditions, loop_condition);
979 /* Select the candidate loop nests for the analysis. This function
980 initializes the EXIT_CONDITIONS array. */
982 static void
983 select_loops_exit_conditions (VEC(gimple,heap) **exit_conditions)
985 struct loop *function_body = current_loops->tree_root;
987 get_exit_conditions_rec (function_body->inner, exit_conditions);
991 /* Depth first search algorithm. */
993 typedef enum t_bool {
994 t_false,
995 t_true,
996 t_dont_know
997 } t_bool;
1000 static t_bool follow_ssa_edge (struct loop *loop, gimple, gimple, tree *, int);
1002 /* Follow the ssa edge into the binary expression RHS0 CODE RHS1.
1003 Return true if the strongly connected component has been found. */
1005 static t_bool
1006 follow_ssa_edge_binary (struct loop *loop, gimple at_stmt,
1007 tree type, tree rhs0, enum tree_code code, tree rhs1,
1008 gimple halting_phi, tree *evolution_of_loop, int limit)
1010 t_bool res = t_false;
1011 tree evol;
1013 switch (code)
1015 case POINTER_PLUS_EXPR:
1016 case PLUS_EXPR:
1017 if (TREE_CODE (rhs0) == SSA_NAME)
1019 if (TREE_CODE (rhs1) == SSA_NAME)
1021 /* Match an assignment under the form:
1022 "a = b + c". */
1024 /* We want only assignments of form "name + name" contribute to
1025 LIMIT, as the other cases do not necessarily contribute to
1026 the complexity of the expression. */
1027 limit++;
1029 evol = *evolution_of_loop;
1030 res = follow_ssa_edge
1031 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, &evol, limit);
1033 if (res == t_true)
1034 *evolution_of_loop = add_to_evolution
1035 (loop->num,
1036 chrec_convert (type, evol, at_stmt),
1037 code, rhs1, at_stmt);
1039 else if (res == t_false)
1041 res = follow_ssa_edge
1042 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1043 evolution_of_loop, limit);
1045 if (res == t_true)
1046 *evolution_of_loop = add_to_evolution
1047 (loop->num,
1048 chrec_convert (type, *evolution_of_loop, at_stmt),
1049 code, rhs0, at_stmt);
1051 else if (res == t_dont_know)
1052 *evolution_of_loop = chrec_dont_know;
1055 else if (res == t_dont_know)
1056 *evolution_of_loop = chrec_dont_know;
1059 else
1061 /* Match an assignment under the form:
1062 "a = b + ...". */
1063 res = follow_ssa_edge
1064 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1065 evolution_of_loop, limit);
1066 if (res == t_true)
1067 *evolution_of_loop = add_to_evolution
1068 (loop->num, chrec_convert (type, *evolution_of_loop,
1069 at_stmt),
1070 code, rhs1, at_stmt);
1072 else if (res == t_dont_know)
1073 *evolution_of_loop = chrec_dont_know;
1077 else if (TREE_CODE (rhs1) == SSA_NAME)
1079 /* Match an assignment under the form:
1080 "a = ... + c". */
1081 res = follow_ssa_edge
1082 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1083 evolution_of_loop, limit);
1084 if (res == t_true)
1085 *evolution_of_loop = add_to_evolution
1086 (loop->num, chrec_convert (type, *evolution_of_loop,
1087 at_stmt),
1088 code, rhs0, at_stmt);
1090 else if (res == t_dont_know)
1091 *evolution_of_loop = chrec_dont_know;
1094 else
1095 /* Otherwise, match an assignment under the form:
1096 "a = ... + ...". */
1097 /* And there is nothing to do. */
1098 res = t_false;
1099 break;
1101 case MINUS_EXPR:
1102 /* This case is under the form "opnd0 = rhs0 - rhs1". */
1103 if (TREE_CODE (rhs0) == SSA_NAME)
1105 /* Match an assignment under the form:
1106 "a = b - ...". */
1108 /* We want only assignments of form "name - name" contribute to
1109 LIMIT, as the other cases do not necessarily contribute to
1110 the complexity of the expression. */
1111 if (TREE_CODE (rhs1) == SSA_NAME)
1112 limit++;
1114 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1115 evolution_of_loop, limit);
1116 if (res == t_true)
1117 *evolution_of_loop = add_to_evolution
1118 (loop->num, chrec_convert (type, *evolution_of_loop, at_stmt),
1119 MINUS_EXPR, rhs1, at_stmt);
1121 else if (res == t_dont_know)
1122 *evolution_of_loop = chrec_dont_know;
1124 else
1125 /* Otherwise, match an assignment under the form:
1126 "a = ... - ...". */
1127 /* And there is nothing to do. */
1128 res = t_false;
1129 break;
1131 default:
1132 res = t_false;
1135 return res;
1138 /* Follow the ssa edge into the expression EXPR.
1139 Return true if the strongly connected component has been found. */
1141 static t_bool
1142 follow_ssa_edge_expr (struct loop *loop, gimple at_stmt, tree expr,
1143 gimple halting_phi, tree *evolution_of_loop, int limit)
1145 t_bool res = t_false;
1146 tree rhs0, rhs1;
1147 tree type = TREE_TYPE (expr);
1148 enum tree_code code;
1150 /* The EXPR is one of the following cases:
1151 - an SSA_NAME,
1152 - an INTEGER_CST,
1153 - a PLUS_EXPR,
1154 - a POINTER_PLUS_EXPR,
1155 - a MINUS_EXPR,
1156 - an ASSERT_EXPR,
1157 - other cases are not yet handled. */
1158 code = TREE_CODE (expr);
1159 switch (code)
1161 case NOP_EXPR:
1162 /* This assignment is under the form "a_1 = (cast) rhs. */
1163 res = follow_ssa_edge_expr (loop, at_stmt, TREE_OPERAND (expr, 0),
1164 halting_phi, evolution_of_loop, limit);
1165 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, at_stmt);
1166 break;
1168 case INTEGER_CST:
1169 /* This assignment is under the form "a_1 = 7". */
1170 res = t_false;
1171 break;
1173 case SSA_NAME:
1174 /* This assignment is under the form: "a_1 = b_2". */
1175 res = follow_ssa_edge
1176 (loop, SSA_NAME_DEF_STMT (expr), halting_phi, evolution_of_loop, limit);
1177 break;
1179 case POINTER_PLUS_EXPR:
1180 case PLUS_EXPR:
1181 case MINUS_EXPR:
1182 /* This case is under the form "rhs0 +- rhs1". */
1183 rhs0 = TREE_OPERAND (expr, 0);
1184 rhs1 = TREE_OPERAND (expr, 1);
1185 STRIP_TYPE_NOPS (rhs0);
1186 STRIP_TYPE_NOPS (rhs1);
1187 return follow_ssa_edge_binary (loop, at_stmt, type, rhs0, code, rhs1,
1188 halting_phi, evolution_of_loop, limit);
1190 case ASSERT_EXPR:
1192 /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>"
1193 It must be handled as a copy assignment of the form a_1 = a_2. */
1194 tree op0 = ASSERT_EXPR_VAR (expr);
1195 if (TREE_CODE (op0) == SSA_NAME)
1196 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (op0),
1197 halting_phi, evolution_of_loop, limit);
1198 else
1199 res = t_false;
1200 break;
1204 default:
1205 res = t_false;
1206 break;
1209 return res;
1212 /* Follow the ssa edge into the right hand side of an assignment STMT.
1213 Return true if the strongly connected component has been found. */
1215 static t_bool
1216 follow_ssa_edge_in_rhs (struct loop *loop, gimple stmt,
1217 gimple halting_phi, tree *evolution_of_loop, int limit)
1219 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
1220 enum tree_code code = gimple_assign_rhs_code (stmt);
1222 switch (get_gimple_rhs_class (code))
1224 case GIMPLE_BINARY_RHS:
1225 return follow_ssa_edge_binary (loop, stmt, type,
1226 gimple_assign_rhs1 (stmt), code,
1227 gimple_assign_rhs2 (stmt),
1228 halting_phi, evolution_of_loop, limit);
1229 case GIMPLE_SINGLE_RHS:
1230 return follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1231 halting_phi, evolution_of_loop, limit);
1232 case GIMPLE_UNARY_RHS:
1233 if (code == NOP_EXPR)
1235 /* This assignment is under the form "a_1 = (cast) rhs. */
1236 t_bool res
1237 = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1238 halting_phi, evolution_of_loop, limit);
1239 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, stmt);
1240 return res;
1242 /* FALLTHRU */
1244 default:
1245 return t_false;
1249 /* Checks whether the I-th argument of a PHI comes from a backedge. */
1251 static bool
1252 backedge_phi_arg_p (gimple phi, int i)
1254 const_edge e = gimple_phi_arg_edge (phi, i);
1256 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1257 about updating it anywhere, and this should work as well most of the
1258 time. */
1259 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1260 return true;
1262 return false;
1265 /* Helper function for one branch of the condition-phi-node. Return
1266 true if the strongly connected component has been found following
1267 this path. */
1269 static inline t_bool
1270 follow_ssa_edge_in_condition_phi_branch (int i,
1271 struct loop *loop,
1272 gimple condition_phi,
1273 gimple halting_phi,
1274 tree *evolution_of_branch,
1275 tree init_cond, int limit)
1277 tree branch = PHI_ARG_DEF (condition_phi, i);
1278 *evolution_of_branch = chrec_dont_know;
1280 /* Do not follow back edges (they must belong to an irreducible loop, which
1281 we really do not want to worry about). */
1282 if (backedge_phi_arg_p (condition_phi, i))
1283 return t_false;
1285 if (TREE_CODE (branch) == SSA_NAME)
1287 *evolution_of_branch = init_cond;
1288 return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi,
1289 evolution_of_branch, limit);
1292 /* This case occurs when one of the condition branches sets
1293 the variable to a constant: i.e. a phi-node like
1294 "a_2 = PHI <a_7(5), 2(6)>;".
1296 FIXME: This case have to be refined correctly:
1297 in some cases it is possible to say something better than
1298 chrec_dont_know, for example using a wrap-around notation. */
1299 return t_false;
1302 /* This function merges the branches of a condition-phi-node in a
1303 loop. */
1305 static t_bool
1306 follow_ssa_edge_in_condition_phi (struct loop *loop,
1307 gimple condition_phi,
1308 gimple halting_phi,
1309 tree *evolution_of_loop, int limit)
1311 int i, n;
1312 tree init = *evolution_of_loop;
1313 tree evolution_of_branch;
1314 t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi,
1315 halting_phi,
1316 &evolution_of_branch,
1317 init, limit);
1318 if (res == t_false || res == t_dont_know)
1319 return res;
1321 *evolution_of_loop = evolution_of_branch;
1323 n = gimple_phi_num_args (condition_phi);
1324 for (i = 1; i < n; i++)
1326 /* Quickly give up when the evolution of one of the branches is
1327 not known. */
1328 if (*evolution_of_loop == chrec_dont_know)
1329 return t_true;
1331 /* Increase the limit by the PHI argument number to avoid exponential
1332 time and memory complexity. */
1333 res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi,
1334 halting_phi,
1335 &evolution_of_branch,
1336 init, limit + i);
1337 if (res == t_false || res == t_dont_know)
1338 return res;
1340 *evolution_of_loop = chrec_merge (*evolution_of_loop,
1341 evolution_of_branch);
1344 return t_true;
1347 /* Follow an SSA edge in an inner loop. It computes the overall
1348 effect of the loop, and following the symbolic initial conditions,
1349 it follows the edges in the parent loop. The inner loop is
1350 considered as a single statement. */
1352 static t_bool
1353 follow_ssa_edge_inner_loop_phi (struct loop *outer_loop,
1354 gimple loop_phi_node,
1355 gimple halting_phi,
1356 tree *evolution_of_loop, int limit)
1358 struct loop *loop = loop_containing_stmt (loop_phi_node);
1359 tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node));
1361 /* Sometimes, the inner loop is too difficult to analyze, and the
1362 result of the analysis is a symbolic parameter. */
1363 if (ev == PHI_RESULT (loop_phi_node))
1365 t_bool res = t_false;
1366 int i, n = gimple_phi_num_args (loop_phi_node);
1368 for (i = 0; i < n; i++)
1370 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1371 basic_block bb;
1373 /* Follow the edges that exit the inner loop. */
1374 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1375 if (!flow_bb_inside_loop_p (loop, bb))
1376 res = follow_ssa_edge_expr (outer_loop, loop_phi_node,
1377 arg, halting_phi,
1378 evolution_of_loop, limit);
1379 if (res == t_true)
1380 break;
1383 /* If the path crosses this loop-phi, give up. */
1384 if (res == t_true)
1385 *evolution_of_loop = chrec_dont_know;
1387 return res;
1390 /* Otherwise, compute the overall effect of the inner loop. */
1391 ev = compute_overall_effect_of_inner_loop (loop, ev);
1392 return follow_ssa_edge_expr (outer_loop, loop_phi_node, ev, halting_phi,
1393 evolution_of_loop, limit);
1396 /* Follow an SSA edge from a loop-phi-node to itself, constructing a
1397 path that is analyzed on the return walk. */
1399 static t_bool
1400 follow_ssa_edge (struct loop *loop, gimple def, gimple halting_phi,
1401 tree *evolution_of_loop, int limit)
1403 struct loop *def_loop;
1405 if (gimple_nop_p (def))
1406 return t_false;
1408 /* Give up if the path is longer than the MAX that we allow. */
1409 if (limit > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
1410 return t_dont_know;
1412 def_loop = loop_containing_stmt (def);
1414 switch (gimple_code (def))
1416 case GIMPLE_PHI:
1417 if (!loop_phi_node_p (def))
1418 /* DEF is a condition-phi-node. Follow the branches, and
1419 record their evolutions. Finally, merge the collected
1420 information and set the approximation to the main
1421 variable. */
1422 return follow_ssa_edge_in_condition_phi
1423 (loop, def, halting_phi, evolution_of_loop, limit);
1425 /* When the analyzed phi is the halting_phi, the
1426 depth-first search is over: we have found a path from
1427 the halting_phi to itself in the loop. */
1428 if (def == halting_phi)
1429 return t_true;
1431 /* Otherwise, the evolution of the HALTING_PHI depends
1432 on the evolution of another loop-phi-node, i.e. the
1433 evolution function is a higher degree polynomial. */
1434 if (def_loop == loop)
1435 return t_false;
1437 /* Inner loop. */
1438 if (flow_loop_nested_p (loop, def_loop))
1439 return follow_ssa_edge_inner_loop_phi
1440 (loop, def, halting_phi, evolution_of_loop, limit + 1);
1442 /* Outer loop. */
1443 return t_false;
1445 case GIMPLE_ASSIGN:
1446 return follow_ssa_edge_in_rhs (loop, def, halting_phi,
1447 evolution_of_loop, limit);
1449 default:
1450 /* At this level of abstraction, the program is just a set
1451 of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no
1452 other node to be handled. */
1453 return t_false;
1459 /* Given a LOOP_PHI_NODE, this function determines the evolution
1460 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1462 static tree
1463 analyze_evolution_in_loop (gimple loop_phi_node,
1464 tree init_cond)
1466 int i, n = gimple_phi_num_args (loop_phi_node);
1467 tree evolution_function = chrec_not_analyzed_yet;
1468 struct loop *loop = loop_containing_stmt (loop_phi_node);
1469 basic_block bb;
1471 if (dump_file && (dump_flags & TDF_DETAILS))
1473 fprintf (dump_file, "(analyze_evolution_in_loop \n");
1474 fprintf (dump_file, " (loop_phi_node = ");
1475 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
1476 fprintf (dump_file, ")\n");
1479 for (i = 0; i < n; i++)
1481 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1482 gimple ssa_chain;
1483 tree ev_fn;
1484 t_bool res;
1486 /* Select the edges that enter the loop body. */
1487 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1488 if (!flow_bb_inside_loop_p (loop, bb))
1489 continue;
1491 if (TREE_CODE (arg) == SSA_NAME)
1493 ssa_chain = SSA_NAME_DEF_STMT (arg);
1495 /* Pass in the initial condition to the follow edge function. */
1496 ev_fn = init_cond;
1497 res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0);
1499 else
1500 res = t_false;
1502 /* When it is impossible to go back on the same
1503 loop_phi_node by following the ssa edges, the
1504 evolution is represented by a peeled chrec, i.e. the
1505 first iteration, EV_FN has the value INIT_COND, then
1506 all the other iterations it has the value of ARG.
1507 For the moment, PEELED_CHREC nodes are not built. */
1508 if (res != t_true)
1509 ev_fn = chrec_dont_know;
1511 /* When there are multiple back edges of the loop (which in fact never
1512 happens currently, but nevertheless), merge their evolutions. */
1513 evolution_function = chrec_merge (evolution_function, ev_fn);
1516 if (dump_file && (dump_flags & TDF_DETAILS))
1518 fprintf (dump_file, " (evolution_function = ");
1519 print_generic_expr (dump_file, evolution_function, 0);
1520 fprintf (dump_file, "))\n");
1523 return evolution_function;
1526 /* Given a loop-phi-node, return the initial conditions of the
1527 variable on entry of the loop. When the CCP has propagated
1528 constants into the loop-phi-node, the initial condition is
1529 instantiated, otherwise the initial condition is kept symbolic.
1530 This analyzer does not analyze the evolution outside the current
1531 loop, and leaves this task to the on-demand tree reconstructor. */
1533 static tree
1534 analyze_initial_condition (gimple loop_phi_node)
1536 int i, n;
1537 tree init_cond = chrec_not_analyzed_yet;
1538 struct loop *loop = loop_containing_stmt (loop_phi_node);
1540 if (dump_file && (dump_flags & TDF_DETAILS))
1542 fprintf (dump_file, "(analyze_initial_condition \n");
1543 fprintf (dump_file, " (loop_phi_node = \n");
1544 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
1545 fprintf (dump_file, ")\n");
1548 n = gimple_phi_num_args (loop_phi_node);
1549 for (i = 0; i < n; i++)
1551 tree branch = PHI_ARG_DEF (loop_phi_node, i);
1552 basic_block bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1554 /* When the branch is oriented to the loop's body, it does
1555 not contribute to the initial condition. */
1556 if (flow_bb_inside_loop_p (loop, bb))
1557 continue;
1559 if (init_cond == chrec_not_analyzed_yet)
1561 init_cond = branch;
1562 continue;
1565 if (TREE_CODE (branch) == SSA_NAME)
1567 init_cond = chrec_dont_know;
1568 break;
1571 init_cond = chrec_merge (init_cond, branch);
1574 /* Ooops -- a loop without an entry??? */
1575 if (init_cond == chrec_not_analyzed_yet)
1576 init_cond = chrec_dont_know;
1578 if (dump_file && (dump_flags & TDF_DETAILS))
1580 fprintf (dump_file, " (init_cond = ");
1581 print_generic_expr (dump_file, init_cond, 0);
1582 fprintf (dump_file, "))\n");
1585 return init_cond;
1588 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1590 static tree
1591 interpret_loop_phi (struct loop *loop, gimple loop_phi_node)
1593 tree res;
1594 struct loop *phi_loop = loop_containing_stmt (loop_phi_node);
1595 tree init_cond;
1597 if (phi_loop != loop)
1599 struct loop *subloop;
1600 tree evolution_fn = analyze_scalar_evolution
1601 (phi_loop, PHI_RESULT (loop_phi_node));
1603 /* Dive one level deeper. */
1604 subloop = superloop_at_depth (phi_loop, loop_depth (loop) + 1);
1606 /* Interpret the subloop. */
1607 res = compute_overall_effect_of_inner_loop (subloop, evolution_fn);
1608 return res;
1611 /* Otherwise really interpret the loop phi. */
1612 init_cond = analyze_initial_condition (loop_phi_node);
1613 res = analyze_evolution_in_loop (loop_phi_node, init_cond);
1615 return res;
1618 /* This function merges the branches of a condition-phi-node,
1619 contained in the outermost loop, and whose arguments are already
1620 analyzed. */
1622 static tree
1623 interpret_condition_phi (struct loop *loop, gimple condition_phi)
1625 int i, n = gimple_phi_num_args (condition_phi);
1626 tree res = chrec_not_analyzed_yet;
1628 for (i = 0; i < n; i++)
1630 tree branch_chrec;
1632 if (backedge_phi_arg_p (condition_phi, i))
1634 res = chrec_dont_know;
1635 break;
1638 branch_chrec = analyze_scalar_evolution
1639 (loop, PHI_ARG_DEF (condition_phi, i));
1641 res = chrec_merge (res, branch_chrec);
1644 return res;
1647 /* Interpret the operation RHS1 OP RHS2. If we didn't
1648 analyze this node before, follow the definitions until ending
1649 either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the
1650 return path, this function propagates evolutions (ala constant copy
1651 propagation). OPND1 is not a GIMPLE expression because we could
1652 analyze the effect of an inner loop: see interpret_loop_phi. */
1654 static tree
1655 interpret_rhs_expr (struct loop *loop, gimple at_stmt,
1656 tree type, tree rhs1, enum tree_code code, tree rhs2)
1658 tree res, chrec1, chrec2;
1660 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1662 if (is_gimple_min_invariant (rhs1))
1663 return chrec_convert (type, rhs1, at_stmt);
1665 if (code == SSA_NAME)
1666 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1667 at_stmt);
1669 if (code == ASSERT_EXPR)
1671 rhs1 = ASSERT_EXPR_VAR (rhs1);
1672 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1673 at_stmt);
1676 return chrec_dont_know;
1679 switch (code)
1681 case POINTER_PLUS_EXPR:
1682 chrec1 = analyze_scalar_evolution (loop, rhs1);
1683 chrec2 = analyze_scalar_evolution (loop, rhs2);
1684 chrec1 = chrec_convert (type, chrec1, at_stmt);
1685 chrec2 = chrec_convert (sizetype, chrec2, at_stmt);
1686 res = chrec_fold_plus (type, chrec1, chrec2);
1687 break;
1689 case PLUS_EXPR:
1690 chrec1 = analyze_scalar_evolution (loop, rhs1);
1691 chrec2 = analyze_scalar_evolution (loop, rhs2);
1692 chrec1 = chrec_convert (type, chrec1, at_stmt);
1693 chrec2 = chrec_convert (type, chrec2, at_stmt);
1694 res = chrec_fold_plus (type, chrec1, chrec2);
1695 break;
1697 case MINUS_EXPR:
1698 chrec1 = analyze_scalar_evolution (loop, rhs1);
1699 chrec2 = analyze_scalar_evolution (loop, rhs2);
1700 chrec1 = chrec_convert (type, chrec1, at_stmt);
1701 chrec2 = chrec_convert (type, chrec2, at_stmt);
1702 res = chrec_fold_minus (type, chrec1, chrec2);
1703 break;
1705 case NEGATE_EXPR:
1706 chrec1 = analyze_scalar_evolution (loop, rhs1);
1707 chrec1 = chrec_convert (type, chrec1, at_stmt);
1708 /* TYPE may be integer, real or complex, so use fold_convert. */
1709 res = chrec_fold_multiply (type, chrec1,
1710 fold_convert (type, integer_minus_one_node));
1711 break;
1713 case BIT_NOT_EXPR:
1714 /* Handle ~X as -1 - X. */
1715 chrec1 = analyze_scalar_evolution (loop, rhs1);
1716 chrec1 = chrec_convert (type, chrec1, at_stmt);
1717 res = chrec_fold_minus (type,
1718 fold_convert (type, integer_minus_one_node),
1719 chrec1);
1720 break;
1722 case MULT_EXPR:
1723 chrec1 = analyze_scalar_evolution (loop, rhs1);
1724 chrec2 = analyze_scalar_evolution (loop, rhs2);
1725 chrec1 = chrec_convert (type, chrec1, at_stmt);
1726 chrec2 = chrec_convert (type, chrec2, at_stmt);
1727 res = chrec_fold_multiply (type, chrec1, chrec2);
1728 break;
1730 CASE_CONVERT:
1731 chrec1 = analyze_scalar_evolution (loop, rhs1);
1732 res = chrec_convert (type, chrec1, at_stmt);
1733 break;
1735 default:
1736 res = chrec_dont_know;
1737 break;
1740 return res;
1743 /* Interpret the expression EXPR. */
1745 static tree
1746 interpret_expr (struct loop *loop, gimple at_stmt, tree expr)
1748 enum tree_code code;
1749 tree type = TREE_TYPE (expr), op0, op1;
1751 if (automatically_generated_chrec_p (expr))
1752 return expr;
1754 if (TREE_CODE (expr) == POLYNOMIAL_CHREC)
1755 return chrec_dont_know;
1757 extract_ops_from_tree (expr, &code, &op0, &op1);
1759 return interpret_rhs_expr (loop, at_stmt, type,
1760 op0, code, op1);
1763 /* Interpret the rhs of the assignment STMT. */
1765 static tree
1766 interpret_gimple_assign (struct loop *loop, gimple stmt)
1768 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
1769 enum tree_code code = gimple_assign_rhs_code (stmt);
1771 return interpret_rhs_expr (loop, stmt, type,
1772 gimple_assign_rhs1 (stmt), code,
1773 gimple_assign_rhs2 (stmt));
1778 /* This section contains all the entry points:
1779 - number_of_iterations_in_loop,
1780 - analyze_scalar_evolution,
1781 - instantiate_parameters.
1784 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1785 common ancestor of DEF_LOOP and USE_LOOP. */
1787 static tree
1788 compute_scalar_evolution_in_loop (struct loop *wrto_loop,
1789 struct loop *def_loop,
1790 tree ev)
1792 tree res;
1793 if (def_loop == wrto_loop)
1794 return ev;
1796 def_loop = superloop_at_depth (def_loop, loop_depth (wrto_loop) + 1);
1797 res = compute_overall_effect_of_inner_loop (def_loop, ev);
1799 return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet);
1802 /* Helper recursive function. */
1804 static tree
1805 analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res)
1807 tree type = TREE_TYPE (var);
1808 gimple def;
1809 basic_block bb;
1810 struct loop *def_loop;
1812 if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE)
1813 return chrec_dont_know;
1815 if (TREE_CODE (var) != SSA_NAME)
1816 return interpret_expr (loop, NULL, var);
1818 def = SSA_NAME_DEF_STMT (var);
1819 bb = gimple_bb (def);
1820 def_loop = bb ? bb->loop_father : NULL;
1822 if (bb == NULL
1823 || !flow_bb_inside_loop_p (loop, bb))
1825 /* Keep the symbolic form. */
1826 res = var;
1827 goto set_and_end;
1830 if (res != chrec_not_analyzed_yet)
1832 if (loop != bb->loop_father)
1833 res = compute_scalar_evolution_in_loop
1834 (find_common_loop (loop, bb->loop_father), bb->loop_father, res);
1836 goto set_and_end;
1839 if (loop != def_loop)
1841 res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet);
1842 res = compute_scalar_evolution_in_loop (loop, def_loop, res);
1844 goto set_and_end;
1847 switch (gimple_code (def))
1849 case GIMPLE_ASSIGN:
1850 res = interpret_gimple_assign (loop, def);
1851 break;
1853 case GIMPLE_PHI:
1854 if (loop_phi_node_p (def))
1855 res = interpret_loop_phi (loop, def);
1856 else
1857 res = interpret_condition_phi (loop, def);
1858 break;
1860 default:
1861 res = chrec_dont_know;
1862 break;
1865 set_and_end:
1867 /* Keep the symbolic form. */
1868 if (res == chrec_dont_know)
1869 res = var;
1871 if (loop == def_loop)
1872 set_scalar_evolution (block_before_loop (loop), var, res);
1874 return res;
1877 /* Entry point for the scalar evolution analyzer.
1878 Analyzes and returns the scalar evolution of the ssa_name VAR.
1879 LOOP_NB is the identifier number of the loop in which the variable
1880 is used.
1882 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
1883 pointer to the statement that uses this variable, in order to
1884 determine the evolution function of the variable, use the following
1885 calls:
1887 unsigned loop_nb = loop_containing_stmt (stmt)->num;
1888 tree chrec_with_symbols = analyze_scalar_evolution (loop_nb, var);
1889 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
1892 tree
1893 analyze_scalar_evolution (struct loop *loop, tree var)
1895 tree res;
1897 if (dump_file && (dump_flags & TDF_DETAILS))
1899 fprintf (dump_file, "(analyze_scalar_evolution \n");
1900 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
1901 fprintf (dump_file, " (scalar = ");
1902 print_generic_expr (dump_file, var, 0);
1903 fprintf (dump_file, ")\n");
1906 res = get_scalar_evolution (block_before_loop (loop), var);
1907 res = analyze_scalar_evolution_1 (loop, var, res);
1909 if (dump_file && (dump_flags & TDF_DETAILS))
1910 fprintf (dump_file, ")\n");
1912 return res;
1915 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
1916 WRTO_LOOP (which should be a superloop of USE_LOOP)
1918 FOLDED_CASTS is set to true if resolve_mixers used
1919 chrec_convert_aggressive (TODO -- not really, we are way too conservative
1920 at the moment in order to keep things simple).
1922 To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following
1923 example:
1925 for (i = 0; i < 100; i++) -- loop 1
1927 for (j = 0; j < 100; j++) -- loop 2
1929 k1 = i;
1930 k2 = j;
1932 use2 (k1, k2);
1934 for (t = 0; t < 100; t++) -- loop 3
1935 use3 (k1, k2);
1938 use1 (k1, k2);
1941 Both k1 and k2 are invariants in loop3, thus
1942 analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1
1943 analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2
1945 As they are invariant, it does not matter whether we consider their
1946 usage in loop 3 or loop 2, hence
1947 analyze_scalar_evolution_in_loop (loop2, loop3, k1) =
1948 analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i
1949 analyze_scalar_evolution_in_loop (loop2, loop3, k2) =
1950 analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2
1952 Similarly for their evolutions with respect to loop 1. The values of K2
1953 in the use in loop 2 vary independently on loop 1, thus we cannot express
1954 the evolution with respect to loop 1:
1955 analyze_scalar_evolution_in_loop (loop1, loop3, k1) =
1956 analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1
1957 analyze_scalar_evolution_in_loop (loop1, loop3, k2) =
1958 analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know
1960 The value of k2 in the use in loop 1 is known, though:
1961 analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1
1962 analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100
1965 static tree
1966 analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop,
1967 tree version, bool *folded_casts)
1969 bool val = false;
1970 tree ev = version, tmp;
1972 /* We cannot just do
1974 tmp = analyze_scalar_evolution (use_loop, version);
1975 ev = resolve_mixers (wrto_loop, tmp);
1977 as resolve_mixers would query the scalar evolution with respect to
1978 wrto_loop. For example, in the situation described in the function
1979 comment, suppose that wrto_loop = loop1, use_loop = loop3 and
1980 version = k2. Then
1982 analyze_scalar_evolution (use_loop, version) = k2
1984 and resolve_mixers (loop1, k2) finds that the value of k2 in loop 1
1985 is 100, which is a wrong result, since we are interested in the
1986 value in loop 3.
1988 Instead, we need to proceed from use_loop to wrto_loop loop by loop,
1989 each time checking that there is no evolution in the inner loop. */
1991 if (folded_casts)
1992 *folded_casts = false;
1993 while (1)
1995 tmp = analyze_scalar_evolution (use_loop, ev);
1996 ev = resolve_mixers (use_loop, tmp);
1998 if (folded_casts && tmp != ev)
1999 *folded_casts = true;
2001 if (use_loop == wrto_loop)
2002 return ev;
2004 /* If the value of the use changes in the inner loop, we cannot express
2005 its value in the outer loop (we might try to return interval chrec,
2006 but we do not have a user for it anyway) */
2007 if (!no_evolution_in_loop_p (ev, use_loop->num, &val)
2008 || !val)
2009 return chrec_dont_know;
2011 use_loop = loop_outer (use_loop);
2015 /* Returns from CACHE the value for VERSION instantiated below
2016 INSTANTIATED_BELOW block. */
2018 static tree
2019 get_instantiated_value (htab_t cache, basic_block instantiated_below,
2020 tree version)
2022 struct scev_info_str *info, pattern;
2024 pattern.var = version;
2025 pattern.instantiated_below = instantiated_below;
2026 info = (struct scev_info_str *) htab_find (cache, &pattern);
2028 if (info)
2029 return info->chrec;
2030 else
2031 return NULL_TREE;
2034 /* Sets in CACHE the value of VERSION instantiated below basic block
2035 INSTANTIATED_BELOW to VAL. */
2037 static void
2038 set_instantiated_value (htab_t cache, basic_block instantiated_below,
2039 tree version, tree val)
2041 struct scev_info_str *info, pattern;
2042 PTR *slot;
2044 pattern.var = version;
2045 pattern.instantiated_below = instantiated_below;
2046 slot = htab_find_slot (cache, &pattern, INSERT);
2048 if (!*slot)
2049 *slot = new_scev_info_str (instantiated_below, version);
2050 info = (struct scev_info_str *) *slot;
2051 info->chrec = val;
2054 /* Return the closed_loop_phi node for VAR. If there is none, return
2055 NULL_TREE. */
2057 static tree
2058 loop_closed_phi_def (tree var)
2060 struct loop *loop;
2061 edge exit;
2062 gimple phi;
2063 gimple_stmt_iterator psi;
2065 if (var == NULL_TREE
2066 || TREE_CODE (var) != SSA_NAME)
2067 return NULL_TREE;
2069 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var));
2070 exit = single_exit (loop);
2071 if (!exit)
2072 return NULL_TREE;
2074 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi))
2076 phi = gsi_stmt (psi);
2077 if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var)
2078 return PHI_RESULT (phi);
2081 return NULL_TREE;
2084 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2085 and EVOLUTION_LOOP, that were left under a symbolic form.
2087 CHREC is the scalar evolution to instantiate.
2089 CACHE is the cache of already instantiated values.
2091 FOLD_CONVERSIONS should be set to true when the conversions that
2092 may wrap in signed/pointer type are folded, as long as the value of
2093 the chrec is preserved.
2095 SIZE_EXPR is used for computing the size of the expression to be
2096 instantiated, and to stop if it exceeds some limit. */
2098 static tree
2099 instantiate_scev_1 (basic_block instantiate_below,
2100 struct loop *evolution_loop, tree chrec,
2101 bool fold_conversions, htab_t cache, int size_expr)
2103 tree res, op0, op1, op2;
2104 basic_block def_bb;
2105 struct loop *def_loop;
2106 tree type = chrec_type (chrec);
2108 /* Give up if the expression is larger than the MAX that we allow. */
2109 if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
2110 return chrec_dont_know;
2112 if (automatically_generated_chrec_p (chrec)
2113 || is_gimple_min_invariant (chrec))
2114 return chrec;
2116 switch (TREE_CODE (chrec))
2118 case SSA_NAME:
2119 def_bb = gimple_bb (SSA_NAME_DEF_STMT (chrec));
2121 /* A parameter (or loop invariant and we do not want to include
2122 evolutions in outer loops), nothing to do. */
2123 if (!def_bb
2124 || loop_depth (def_bb->loop_father) == 0
2125 || dominated_by_p (CDI_DOMINATORS, instantiate_below, def_bb))
2126 return chrec;
2128 /* We cache the value of instantiated variable to avoid exponential
2129 time complexity due to reevaluations. We also store the convenient
2130 value in the cache in order to prevent infinite recursion -- we do
2131 not want to instantiate the SSA_NAME if it is in a mixer
2132 structure. This is used for avoiding the instantiation of
2133 recursively defined functions, such as:
2135 | a_2 -> {0, +, 1, +, a_2}_1 */
2137 res = get_instantiated_value (cache, instantiate_below, chrec);
2138 if (res)
2139 return res;
2141 res = chrec_dont_know;
2142 set_instantiated_value (cache, instantiate_below, chrec, res);
2144 def_loop = find_common_loop (evolution_loop, def_bb->loop_father);
2146 /* If the analysis yields a parametric chrec, instantiate the
2147 result again. */
2148 res = analyze_scalar_evolution (def_loop, chrec);
2150 /* Don't instantiate loop-closed-ssa phi nodes. */
2151 if (TREE_CODE (res) == SSA_NAME
2152 && (loop_containing_stmt (SSA_NAME_DEF_STMT (res)) == NULL
2153 || (loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res)))
2154 > loop_depth (def_loop))))
2156 if (res == chrec)
2157 res = loop_closed_phi_def (chrec);
2158 else
2159 res = chrec;
2161 if (res == NULL_TREE)
2162 res = chrec_dont_know;
2165 else if (res != chrec_dont_know)
2166 res = instantiate_scev_1 (instantiate_below, evolution_loop, res,
2167 fold_conversions, cache, size_expr);
2169 /* Store the correct value to the cache. */
2170 set_instantiated_value (cache, instantiate_below, chrec, res);
2171 return res;
2173 case POLYNOMIAL_CHREC:
2174 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2175 CHREC_LEFT (chrec), fold_conversions, cache,
2176 size_expr);
2177 if (op0 == chrec_dont_know)
2178 return chrec_dont_know;
2180 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2181 CHREC_RIGHT (chrec), fold_conversions, cache,
2182 size_expr);
2183 if (op1 == chrec_dont_know)
2184 return chrec_dont_know;
2186 if (CHREC_LEFT (chrec) != op0
2187 || CHREC_RIGHT (chrec) != op1)
2189 op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL);
2190 chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1);
2192 return chrec;
2194 case POINTER_PLUS_EXPR:
2195 case PLUS_EXPR:
2196 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2197 TREE_OPERAND (chrec, 0), fold_conversions, cache,
2198 size_expr);
2199 if (op0 == chrec_dont_know)
2200 return chrec_dont_know;
2202 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2203 TREE_OPERAND (chrec, 1), fold_conversions, cache,
2204 size_expr);
2205 if (op1 == chrec_dont_know)
2206 return chrec_dont_know;
2208 if (TREE_OPERAND (chrec, 0) != op0
2209 || TREE_OPERAND (chrec, 1) != op1)
2211 op0 = chrec_convert (type, op0, NULL);
2212 op1 = chrec_convert_rhs (type, op1, NULL);
2213 chrec = chrec_fold_plus (type, op0, op1);
2215 return chrec;
2217 case MINUS_EXPR:
2218 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2219 TREE_OPERAND (chrec, 0), fold_conversions, cache,
2220 size_expr);
2221 if (op0 == chrec_dont_know)
2222 return chrec_dont_know;
2224 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2225 TREE_OPERAND (chrec, 1),
2226 fold_conversions, cache, size_expr);
2227 if (op1 == chrec_dont_know)
2228 return chrec_dont_know;
2230 if (TREE_OPERAND (chrec, 0) != op0
2231 || TREE_OPERAND (chrec, 1) != op1)
2233 op0 = chrec_convert (type, op0, NULL);
2234 op1 = chrec_convert (type, op1, NULL);
2235 chrec = chrec_fold_minus (type, op0, op1);
2237 return chrec;
2239 case MULT_EXPR:
2240 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2241 TREE_OPERAND (chrec, 0),
2242 fold_conversions, cache, size_expr);
2243 if (op0 == chrec_dont_know)
2244 return chrec_dont_know;
2246 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2247 TREE_OPERAND (chrec, 1),
2248 fold_conversions, cache, size_expr);
2249 if (op1 == chrec_dont_know)
2250 return chrec_dont_know;
2252 if (TREE_OPERAND (chrec, 0) != op0
2253 || TREE_OPERAND (chrec, 1) != op1)
2255 op0 = chrec_convert (type, op0, NULL);
2256 op1 = chrec_convert (type, op1, NULL);
2257 chrec = chrec_fold_multiply (type, op0, op1);
2259 return chrec;
2261 CASE_CONVERT:
2262 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2263 TREE_OPERAND (chrec, 0),
2264 fold_conversions, cache, size_expr);
2265 if (op0 == chrec_dont_know)
2266 return chrec_dont_know;
2268 if (fold_conversions)
2270 tree tmp = chrec_convert_aggressive (TREE_TYPE (chrec), op0);
2271 if (tmp)
2272 return tmp;
2275 if (op0 == TREE_OPERAND (chrec, 0))
2276 return chrec;
2278 /* If we used chrec_convert_aggressive, we can no longer assume that
2279 signed chrecs do not overflow, as chrec_convert does, so avoid
2280 calling it in that case. */
2281 if (fold_conversions)
2282 return fold_convert (TREE_TYPE (chrec), op0);
2284 return chrec_convert (TREE_TYPE (chrec), op0, NULL);
2286 case BIT_NOT_EXPR:
2287 /* Handle ~X as -1 - X. */
2288 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2289 TREE_OPERAND (chrec, 0),
2290 fold_conversions, cache, size_expr);
2291 if (op0 == chrec_dont_know)
2292 return chrec_dont_know;
2294 if (TREE_OPERAND (chrec, 0) != op0)
2296 op0 = chrec_convert (type, op0, NULL);
2297 chrec = chrec_fold_minus (type,
2298 fold_convert (type,
2299 integer_minus_one_node),
2300 op0);
2302 return chrec;
2304 case SCEV_NOT_KNOWN:
2305 return chrec_dont_know;
2307 case SCEV_KNOWN:
2308 return chrec_known;
2310 default:
2311 break;
2314 if (VL_EXP_CLASS_P (chrec))
2315 return chrec_dont_know;
2317 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
2319 case 3:
2320 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2321 TREE_OPERAND (chrec, 0),
2322 fold_conversions, cache, size_expr);
2323 if (op0 == chrec_dont_know)
2324 return chrec_dont_know;
2326 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2327 TREE_OPERAND (chrec, 1),
2328 fold_conversions, cache, size_expr);
2329 if (op1 == chrec_dont_know)
2330 return chrec_dont_know;
2332 op2 = instantiate_scev_1 (instantiate_below, evolution_loop,
2333 TREE_OPERAND (chrec, 2),
2334 fold_conversions, cache, size_expr);
2335 if (op2 == chrec_dont_know)
2336 return chrec_dont_know;
2338 if (op0 == TREE_OPERAND (chrec, 0)
2339 && op1 == TREE_OPERAND (chrec, 1)
2340 && op2 == TREE_OPERAND (chrec, 2))
2341 return chrec;
2343 return fold_build3 (TREE_CODE (chrec),
2344 TREE_TYPE (chrec), op0, op1, op2);
2346 case 2:
2347 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2348 TREE_OPERAND (chrec, 0),
2349 fold_conversions, cache, size_expr);
2350 if (op0 == chrec_dont_know)
2351 return chrec_dont_know;
2353 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2354 TREE_OPERAND (chrec, 1),
2355 fold_conversions, cache, size_expr);
2356 if (op1 == chrec_dont_know)
2357 return chrec_dont_know;
2359 if (op0 == TREE_OPERAND (chrec, 0)
2360 && op1 == TREE_OPERAND (chrec, 1))
2361 return chrec;
2362 return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1);
2364 case 1:
2365 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2366 TREE_OPERAND (chrec, 0),
2367 fold_conversions, cache, size_expr);
2368 if (op0 == chrec_dont_know)
2369 return chrec_dont_know;
2370 if (op0 == TREE_OPERAND (chrec, 0))
2371 return chrec;
2372 return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0);
2374 case 0:
2375 return chrec;
2377 default:
2378 break;
2381 /* Too complicated to handle. */
2382 return chrec_dont_know;
2385 /* Analyze all the parameters of the chrec that were left under a
2386 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2387 recursive instantiation of parameters: a parameter is a variable
2388 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2389 a function parameter. */
2391 tree
2392 instantiate_scev (basic_block instantiate_below, struct loop *evolution_loop,
2393 tree chrec)
2395 tree res;
2396 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2398 if (dump_file && (dump_flags & TDF_DETAILS))
2400 fprintf (dump_file, "(instantiate_scev \n");
2401 fprintf (dump_file, " (instantiate_below = %d)\n", instantiate_below->index);
2402 fprintf (dump_file, " (evolution_loop = %d)\n", evolution_loop->num);
2403 fprintf (dump_file, " (chrec = ");
2404 print_generic_expr (dump_file, chrec, 0);
2405 fprintf (dump_file, ")\n");
2408 res = instantiate_scev_1 (instantiate_below, evolution_loop, chrec, false,
2409 cache, 0);
2411 if (dump_file && (dump_flags & TDF_DETAILS))
2413 fprintf (dump_file, " (res = ");
2414 print_generic_expr (dump_file, res, 0);
2415 fprintf (dump_file, "))\n");
2418 htab_delete (cache);
2420 return res;
2423 /* Similar to instantiate_parameters, but does not introduce the
2424 evolutions in outer loops for LOOP invariants in CHREC, and does not
2425 care about causing overflows, as long as they do not affect value
2426 of an expression. */
2428 tree
2429 resolve_mixers (struct loop *loop, tree chrec)
2431 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2432 tree ret = instantiate_scev_1 (block_before_loop (loop), loop, chrec, true,
2433 cache, 0);
2434 htab_delete (cache);
2435 return ret;
2438 /* Entry point for the analysis of the number of iterations pass.
2439 This function tries to safely approximate the number of iterations
2440 the loop will run. When this property is not decidable at compile
2441 time, the result is chrec_dont_know. Otherwise the result is
2442 a scalar or a symbolic parameter.
2444 Example of analysis: suppose that the loop has an exit condition:
2446 "if (b > 49) goto end_loop;"
2448 and that in a previous analysis we have determined that the
2449 variable 'b' has an evolution function:
2451 "EF = {23, +, 5}_2".
2453 When we evaluate the function at the point 5, i.e. the value of the
2454 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2455 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2456 the loop body has been executed 6 times. */
2458 tree
2459 number_of_latch_executions (struct loop *loop)
2461 tree res, type;
2462 edge exit;
2463 struct tree_niter_desc niter_desc;
2465 /* Determine whether the number_of_iterations_in_loop has already
2466 been computed. */
2467 res = loop->nb_iterations;
2468 if (res)
2469 return res;
2470 res = chrec_dont_know;
2472 if (dump_file && (dump_flags & TDF_DETAILS))
2473 fprintf (dump_file, "(number_of_iterations_in_loop\n");
2475 exit = single_exit (loop);
2476 if (!exit)
2477 goto end;
2479 if (!number_of_iterations_exit (loop, exit, &niter_desc, false))
2480 goto end;
2482 type = TREE_TYPE (niter_desc.niter);
2483 if (integer_nonzerop (niter_desc.may_be_zero))
2484 res = build_int_cst (type, 0);
2485 else if (integer_zerop (niter_desc.may_be_zero))
2486 res = niter_desc.niter;
2487 else
2488 res = chrec_dont_know;
2490 end:
2491 return set_nb_iterations_in_loop (loop, res);
2494 /* Returns the number of executions of the exit condition of LOOP,
2495 i.e., the number by one higher than number_of_latch_executions.
2496 Note that unlike number_of_latch_executions, this number does
2497 not necessarily fit in the unsigned variant of the type of
2498 the control variable -- if the number of iterations is a constant,
2499 we return chrec_dont_know if adding one to number_of_latch_executions
2500 overflows; however, in case the number of iterations is symbolic
2501 expression, the caller is responsible for dealing with this
2502 the possible overflow. */
2504 tree
2505 number_of_exit_cond_executions (struct loop *loop)
2507 tree ret = number_of_latch_executions (loop);
2508 tree type = chrec_type (ret);
2510 if (chrec_contains_undetermined (ret))
2511 return ret;
2513 ret = chrec_fold_plus (type, ret, build_int_cst (type, 1));
2514 if (TREE_CODE (ret) == INTEGER_CST
2515 && TREE_OVERFLOW (ret))
2516 return chrec_dont_know;
2518 return ret;
2521 /* One of the drivers for testing the scalar evolutions analysis.
2522 This function computes the number of iterations for all the loops
2523 from the EXIT_CONDITIONS array. */
2525 static void
2526 number_of_iterations_for_all_loops (VEC(gimple,heap) **exit_conditions)
2528 unsigned int i;
2529 unsigned nb_chrec_dont_know_loops = 0;
2530 unsigned nb_static_loops = 0;
2531 gimple cond;
2533 for (i = 0; VEC_iterate (gimple, *exit_conditions, i, cond); i++)
2535 tree res = number_of_latch_executions (loop_containing_stmt (cond));
2536 if (chrec_contains_undetermined (res))
2537 nb_chrec_dont_know_loops++;
2538 else
2539 nb_static_loops++;
2542 if (dump_file)
2544 fprintf (dump_file, "\n(\n");
2545 fprintf (dump_file, "-----------------------------------------\n");
2546 fprintf (dump_file, "%d\tnb_chrec_dont_know_loops\n", nb_chrec_dont_know_loops);
2547 fprintf (dump_file, "%d\tnb_static_loops\n", nb_static_loops);
2548 fprintf (dump_file, "%d\tnb_total_loops\n", number_of_loops ());
2549 fprintf (dump_file, "-----------------------------------------\n");
2550 fprintf (dump_file, ")\n\n");
2552 print_loops (dump_file, 3);
2558 /* Counters for the stats. */
2560 struct chrec_stats
2562 unsigned nb_chrecs;
2563 unsigned nb_affine;
2564 unsigned nb_affine_multivar;
2565 unsigned nb_higher_poly;
2566 unsigned nb_chrec_dont_know;
2567 unsigned nb_undetermined;
2570 /* Reset the counters. */
2572 static inline void
2573 reset_chrecs_counters (struct chrec_stats *stats)
2575 stats->nb_chrecs = 0;
2576 stats->nb_affine = 0;
2577 stats->nb_affine_multivar = 0;
2578 stats->nb_higher_poly = 0;
2579 stats->nb_chrec_dont_know = 0;
2580 stats->nb_undetermined = 0;
2583 /* Dump the contents of a CHREC_STATS structure. */
2585 static void
2586 dump_chrecs_stats (FILE *file, struct chrec_stats *stats)
2588 fprintf (file, "\n(\n");
2589 fprintf (file, "-----------------------------------------\n");
2590 fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine);
2591 fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar);
2592 fprintf (file, "%d\tdegree greater than 2 polynomials\n",
2593 stats->nb_higher_poly);
2594 fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know);
2595 fprintf (file, "-----------------------------------------\n");
2596 fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs);
2597 fprintf (file, "%d\twith undetermined coefficients\n",
2598 stats->nb_undetermined);
2599 fprintf (file, "-----------------------------------------\n");
2600 fprintf (file, "%d\tchrecs in the scev database\n",
2601 (int) htab_elements (scalar_evolution_info));
2602 fprintf (file, "%d\tsets in the scev database\n", nb_set_scev);
2603 fprintf (file, "%d\tgets in the scev database\n", nb_get_scev);
2604 fprintf (file, "-----------------------------------------\n");
2605 fprintf (file, ")\n\n");
2608 /* Gather statistics about CHREC. */
2610 static void
2611 gather_chrec_stats (tree chrec, struct chrec_stats *stats)
2613 if (dump_file && (dump_flags & TDF_STATS))
2615 fprintf (dump_file, "(classify_chrec ");
2616 print_generic_expr (dump_file, chrec, 0);
2617 fprintf (dump_file, "\n");
2620 stats->nb_chrecs++;
2622 if (chrec == NULL_TREE)
2624 stats->nb_undetermined++;
2625 return;
2628 switch (TREE_CODE (chrec))
2630 case POLYNOMIAL_CHREC:
2631 if (evolution_function_is_affine_p (chrec))
2633 if (dump_file && (dump_flags & TDF_STATS))
2634 fprintf (dump_file, " affine_univariate\n");
2635 stats->nb_affine++;
2637 else if (evolution_function_is_affine_multivariate_p (chrec, 0))
2639 if (dump_file && (dump_flags & TDF_STATS))
2640 fprintf (dump_file, " affine_multivariate\n");
2641 stats->nb_affine_multivar++;
2643 else
2645 if (dump_file && (dump_flags & TDF_STATS))
2646 fprintf (dump_file, " higher_degree_polynomial\n");
2647 stats->nb_higher_poly++;
2650 break;
2652 default:
2653 break;
2656 if (chrec_contains_undetermined (chrec))
2658 if (dump_file && (dump_flags & TDF_STATS))
2659 fprintf (dump_file, " undetermined\n");
2660 stats->nb_undetermined++;
2663 if (dump_file && (dump_flags & TDF_STATS))
2664 fprintf (dump_file, ")\n");
2667 /* One of the drivers for testing the scalar evolutions analysis.
2668 This function analyzes the scalar evolution of all the scalars
2669 defined as loop phi nodes in one of the loops from the
2670 EXIT_CONDITIONS array.
2672 TODO Optimization: A loop is in canonical form if it contains only
2673 a single scalar loop phi node. All the other scalars that have an
2674 evolution in the loop are rewritten in function of this single
2675 index. This allows the parallelization of the loop. */
2677 static void
2678 analyze_scalar_evolution_for_all_loop_phi_nodes (VEC(gimple,heap) **exit_conditions)
2680 unsigned int i;
2681 struct chrec_stats stats;
2682 gimple cond, phi;
2683 gimple_stmt_iterator psi;
2685 reset_chrecs_counters (&stats);
2687 for (i = 0; VEC_iterate (gimple, *exit_conditions, i, cond); i++)
2689 struct loop *loop;
2690 basic_block bb;
2691 tree chrec;
2693 loop = loop_containing_stmt (cond);
2694 bb = loop->header;
2696 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
2698 phi = gsi_stmt (psi);
2699 if (is_gimple_reg (PHI_RESULT (phi)))
2701 chrec = instantiate_parameters
2702 (loop,
2703 analyze_scalar_evolution (loop, PHI_RESULT (phi)));
2705 if (dump_file && (dump_flags & TDF_STATS))
2706 gather_chrec_stats (chrec, &stats);
2711 if (dump_file && (dump_flags & TDF_STATS))
2712 dump_chrecs_stats (dump_file, &stats);
2715 /* Callback for htab_traverse, gathers information on chrecs in the
2716 hashtable. */
2718 static int
2719 gather_stats_on_scev_database_1 (void **slot, void *stats)
2721 struct scev_info_str *entry = (struct scev_info_str *) *slot;
2723 gather_chrec_stats (entry->chrec, (struct chrec_stats *) stats);
2725 return 1;
2728 /* Classify the chrecs of the whole database. */
2730 void
2731 gather_stats_on_scev_database (void)
2733 struct chrec_stats stats;
2735 if (!dump_file)
2736 return;
2738 reset_chrecs_counters (&stats);
2740 htab_traverse (scalar_evolution_info, gather_stats_on_scev_database_1,
2741 &stats);
2743 dump_chrecs_stats (dump_file, &stats);
2748 /* Initializer. */
2750 static void
2751 initialize_scalar_evolutions_analyzer (void)
2753 /* The elements below are unique. */
2754 if (chrec_dont_know == NULL_TREE)
2756 chrec_not_analyzed_yet = NULL_TREE;
2757 chrec_dont_know = make_node (SCEV_NOT_KNOWN);
2758 chrec_known = make_node (SCEV_KNOWN);
2759 TREE_TYPE (chrec_dont_know) = void_type_node;
2760 TREE_TYPE (chrec_known) = void_type_node;
2764 /* Initialize the analysis of scalar evolutions for LOOPS. */
2766 void
2767 scev_initialize (void)
2769 loop_iterator li;
2770 struct loop *loop;
2772 scalar_evolution_info = htab_create_alloc (100,
2773 hash_scev_info,
2774 eq_scev_info,
2775 del_scev_info,
2776 ggc_calloc,
2777 ggc_free);
2779 initialize_scalar_evolutions_analyzer ();
2781 FOR_EACH_LOOP (li, loop, 0)
2783 loop->nb_iterations = NULL_TREE;
2787 /* Cleans up the information cached by the scalar evolutions analysis. */
2789 void
2790 scev_reset (void)
2792 loop_iterator li;
2793 struct loop *loop;
2795 if (!scalar_evolution_info || !current_loops)
2796 return;
2798 htab_empty (scalar_evolution_info);
2799 FOR_EACH_LOOP (li, loop, 0)
2801 loop->nb_iterations = NULL_TREE;
2805 /* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with
2806 respect to WRTO_LOOP and returns its base and step in IV if possible
2807 (see analyze_scalar_evolution_in_loop for more details on USE_LOOP
2808 and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be
2809 invariant in LOOP. Otherwise we require it to be an integer constant.
2811 IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g.
2812 because it is computed in signed arithmetics). Consequently, adding an
2813 induction variable
2815 for (i = IV->base; ; i += IV->step)
2817 is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is
2818 false for the type of the induction variable, or you can prove that i does
2819 not wrap by some other argument. Otherwise, this might introduce undefined
2820 behavior, and
2822 for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step))
2824 must be used instead. */
2826 bool
2827 simple_iv (struct loop *wrto_loop, struct loop *use_loop, tree op,
2828 affine_iv *iv, bool allow_nonconstant_step)
2830 tree type, ev;
2831 bool folded_casts;
2833 iv->base = NULL_TREE;
2834 iv->step = NULL_TREE;
2835 iv->no_overflow = false;
2837 type = TREE_TYPE (op);
2838 if (TREE_CODE (type) != INTEGER_TYPE
2839 && TREE_CODE (type) != POINTER_TYPE)
2840 return false;
2842 ev = analyze_scalar_evolution_in_loop (wrto_loop, use_loop, op,
2843 &folded_casts);
2844 if (chrec_contains_undetermined (ev)
2845 || chrec_contains_symbols_defined_in_loop (ev, wrto_loop->num))
2846 return false;
2848 if (tree_does_not_contain_chrecs (ev))
2850 iv->base = ev;
2851 iv->step = build_int_cst (TREE_TYPE (ev), 0);
2852 iv->no_overflow = true;
2853 return true;
2856 if (TREE_CODE (ev) != POLYNOMIAL_CHREC
2857 || CHREC_VARIABLE (ev) != (unsigned) wrto_loop->num)
2858 return false;
2860 iv->step = CHREC_RIGHT (ev);
2861 if ((!allow_nonconstant_step && TREE_CODE (iv->step) != INTEGER_CST)
2862 || tree_contains_chrecs (iv->step, NULL))
2863 return false;
2865 iv->base = CHREC_LEFT (ev);
2866 if (tree_contains_chrecs (iv->base, NULL))
2867 return false;
2869 iv->no_overflow = !folded_casts && TYPE_OVERFLOW_UNDEFINED (type);
2871 return true;
2874 /* Runs the analysis of scalar evolutions. */
2876 void
2877 scev_analysis (void)
2879 VEC(gimple,heap) *exit_conditions;
2881 exit_conditions = VEC_alloc (gimple, heap, 37);
2882 select_loops_exit_conditions (&exit_conditions);
2884 if (dump_file && (dump_flags & TDF_STATS))
2885 analyze_scalar_evolution_for_all_loop_phi_nodes (&exit_conditions);
2887 number_of_iterations_for_all_loops (&exit_conditions);
2888 VEC_free (gimple, heap, exit_conditions);
2891 /* Finalize the scalar evolution analysis. */
2893 void
2894 scev_finalize (void)
2896 if (!scalar_evolution_info)
2897 return;
2898 htab_delete (scalar_evolution_info);
2899 scalar_evolution_info = NULL;
2902 /* Returns true if the expression EXPR is considered to be too expensive
2903 for scev_const_prop. */
2905 bool
2906 expression_expensive_p (tree expr)
2908 enum tree_code code;
2910 if (is_gimple_val (expr))
2911 return false;
2913 code = TREE_CODE (expr);
2914 if (code == TRUNC_DIV_EXPR
2915 || code == CEIL_DIV_EXPR
2916 || code == FLOOR_DIV_EXPR
2917 || code == ROUND_DIV_EXPR
2918 || code == TRUNC_MOD_EXPR
2919 || code == CEIL_MOD_EXPR
2920 || code == FLOOR_MOD_EXPR
2921 || code == ROUND_MOD_EXPR
2922 || code == EXACT_DIV_EXPR)
2924 /* Division by power of two is usually cheap, so we allow it.
2925 Forbid anything else. */
2926 if (!integer_pow2p (TREE_OPERAND (expr, 1)))
2927 return true;
2930 switch (TREE_CODE_CLASS (code))
2932 case tcc_binary:
2933 case tcc_comparison:
2934 if (expression_expensive_p (TREE_OPERAND (expr, 1)))
2935 return true;
2937 /* Fallthru. */
2938 case tcc_unary:
2939 return expression_expensive_p (TREE_OPERAND (expr, 0));
2941 default:
2942 return true;
2946 /* Replace ssa names for that scev can prove they are constant by the
2947 appropriate constants. Also perform final value replacement in loops,
2948 in case the replacement expressions are cheap.
2950 We only consider SSA names defined by phi nodes; rest is left to the
2951 ordinary constant propagation pass. */
2953 unsigned int
2954 scev_const_prop (void)
2956 basic_block bb;
2957 tree name, type, ev;
2958 gimple phi, ass;
2959 struct loop *loop, *ex_loop;
2960 bitmap ssa_names_to_remove = NULL;
2961 unsigned i;
2962 loop_iterator li;
2963 gimple_stmt_iterator psi;
2965 if (number_of_loops () <= 1)
2966 return 0;
2968 FOR_EACH_BB (bb)
2970 loop = bb->loop_father;
2972 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
2974 phi = gsi_stmt (psi);
2975 name = PHI_RESULT (phi);
2977 if (!is_gimple_reg (name))
2978 continue;
2980 type = TREE_TYPE (name);
2982 if (!POINTER_TYPE_P (type)
2983 && !INTEGRAL_TYPE_P (type))
2984 continue;
2986 ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name));
2987 if (!is_gimple_min_invariant (ev)
2988 || !may_propagate_copy (name, ev))
2989 continue;
2991 /* Replace the uses of the name. */
2992 if (name != ev)
2993 replace_uses_by (name, ev);
2995 if (!ssa_names_to_remove)
2996 ssa_names_to_remove = BITMAP_ALLOC (NULL);
2997 bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name));
3001 /* Remove the ssa names that were replaced by constants. We do not
3002 remove them directly in the previous cycle, since this
3003 invalidates scev cache. */
3004 if (ssa_names_to_remove)
3006 bitmap_iterator bi;
3008 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi)
3010 gimple_stmt_iterator psi;
3011 name = ssa_name (i);
3012 phi = SSA_NAME_DEF_STMT (name);
3014 gcc_assert (gimple_code (phi) == GIMPLE_PHI);
3015 psi = gsi_for_stmt (phi);
3016 remove_phi_node (&psi, true);
3019 BITMAP_FREE (ssa_names_to_remove);
3020 scev_reset ();
3023 /* Now the regular final value replacement. */
3024 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
3026 edge exit;
3027 tree def, rslt, niter;
3028 gimple_stmt_iterator bsi;
3030 /* If we do not know exact number of iterations of the loop, we cannot
3031 replace the final value. */
3032 exit = single_exit (loop);
3033 if (!exit)
3034 continue;
3036 niter = number_of_latch_executions (loop);
3037 if (niter == chrec_dont_know)
3038 continue;
3040 /* Ensure that it is possible to insert new statements somewhere. */
3041 if (!single_pred_p (exit->dest))
3042 split_loop_exit_edge (exit);
3043 bsi = gsi_after_labels (exit->dest);
3045 ex_loop = superloop_at_depth (loop,
3046 loop_depth (exit->dest->loop_father) + 1);
3048 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); )
3050 phi = gsi_stmt (psi);
3051 rslt = PHI_RESULT (phi);
3052 def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
3053 if (!is_gimple_reg (def))
3055 gsi_next (&psi);
3056 continue;
3059 if (!POINTER_TYPE_P (TREE_TYPE (def))
3060 && !INTEGRAL_TYPE_P (TREE_TYPE (def)))
3062 gsi_next (&psi);
3063 continue;
3066 def = analyze_scalar_evolution_in_loop (ex_loop, loop, def, NULL);
3067 def = compute_overall_effect_of_inner_loop (ex_loop, def);
3068 if (!tree_does_not_contain_chrecs (def)
3069 || chrec_contains_symbols_defined_in_loop (def, ex_loop->num)
3070 /* Moving the computation from the loop may prolong life range
3071 of some ssa names, which may cause problems if they appear
3072 on abnormal edges. */
3073 || contains_abnormal_ssa_name_p (def)
3074 /* Do not emit expensive expressions. The rationale is that
3075 when someone writes a code like
3077 while (n > 45) n -= 45;
3079 he probably knows that n is not large, and does not want it
3080 to be turned into n %= 45. */
3081 || expression_expensive_p (def))
3083 gsi_next (&psi);
3084 continue;
3087 /* Eliminate the PHI node and replace it by a computation outside
3088 the loop. */
3089 def = unshare_expr (def);
3090 remove_phi_node (&psi, false);
3092 def = force_gimple_operand_gsi (&bsi, def, false, NULL_TREE,
3093 true, GSI_SAME_STMT);
3094 ass = gimple_build_assign (rslt, def);
3095 gsi_insert_before (&bsi, ass, GSI_SAME_STMT);
3098 return 0;
3101 #include "gt-tree-scalar-evolution.h"