fix pr/45972
[official-gcc.git] / gcc / tree-scalar-evolution.c
blob8a5797eb827f4a8f606cd850c0597c351f6bf4e3
1 /* Scalar evolution detector.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
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/>. */
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 "basic-block.h"
264 #include "tree-pretty-print.h"
265 #include "gimple-pretty-print.h"
266 #include "tree-flow.h"
267 #include "tree-dump.h"
268 #include "timevar.h"
269 #include "cfgloop.h"
270 #include "tree-chrec.h"
271 #include "tree-scalar-evolution.h"
272 #include "tree-pass.h"
273 #include "flags.h"
274 #include "params.h"
276 static tree analyze_scalar_evolution_1 (struct loop *, tree, tree);
278 /* The cached information about an SSA name VAR, claiming that below
279 basic block INSTANTIATED_BELOW, the value of VAR can be expressed
280 as CHREC. */
282 struct GTY(()) scev_info_str {
283 basic_block instantiated_below;
284 tree var;
285 tree chrec;
288 /* Counters for the scev database. */
289 static unsigned nb_set_scev = 0;
290 static unsigned nb_get_scev = 0;
292 /* The following trees are unique elements. Thus the comparison of
293 another element to these elements should be done on the pointer to
294 these trees, and not on their value. */
296 /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
297 tree chrec_not_analyzed_yet;
299 /* Reserved to the cases where the analyzer has detected an
300 undecidable property at compile time. */
301 tree chrec_dont_know;
303 /* When the analyzer has detected that a property will never
304 happen, then it qualifies it with chrec_known. */
305 tree chrec_known;
307 static GTY ((param_is (struct scev_info_str))) htab_t scalar_evolution_info;
310 /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */
312 static inline struct scev_info_str *
313 new_scev_info_str (basic_block instantiated_below, tree var)
315 struct scev_info_str *res;
317 res = ggc_alloc_scev_info_str ();
318 res->var = var;
319 res->chrec = chrec_not_analyzed_yet;
320 res->instantiated_below = instantiated_below;
322 return res;
325 /* Computes a hash function for database element ELT. */
327 static hashval_t
328 hash_scev_info (const void *elt)
330 return SSA_NAME_VERSION (((const struct scev_info_str *) elt)->var);
333 /* Compares database elements E1 and E2. */
335 static int
336 eq_scev_info (const void *e1, const void *e2)
338 const struct scev_info_str *elt1 = (const struct scev_info_str *) e1;
339 const struct scev_info_str *elt2 = (const struct scev_info_str *) e2;
341 return (elt1->var == elt2->var
342 && elt1->instantiated_below == elt2->instantiated_below);
345 /* Deletes database element E. */
347 static void
348 del_scev_info (void *e)
350 ggc_free (e);
353 /* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block.
354 A first query on VAR returns chrec_not_analyzed_yet. */
356 static tree *
357 find_var_scev_info (basic_block instantiated_below, tree var)
359 struct scev_info_str *res;
360 struct scev_info_str tmp;
361 PTR *slot;
363 tmp.var = var;
364 tmp.instantiated_below = instantiated_below;
365 slot = htab_find_slot (scalar_evolution_info, &tmp, INSERT);
367 if (!*slot)
368 *slot = new_scev_info_str (instantiated_below, var);
369 res = (struct scev_info_str *) *slot;
371 return &res->chrec;
374 /* Return true when CHREC contains symbolic names defined in
375 LOOP_NB. */
377 bool
378 chrec_contains_symbols_defined_in_loop (const_tree chrec, unsigned loop_nb)
380 int i, n;
382 if (chrec == NULL_TREE)
383 return false;
385 if (is_gimple_min_invariant (chrec))
386 return false;
388 if (TREE_CODE (chrec) == SSA_NAME)
390 gimple def;
391 loop_p def_loop, loop;
393 if (SSA_NAME_IS_DEFAULT_DEF (chrec))
394 return false;
396 def = SSA_NAME_DEF_STMT (chrec);
397 def_loop = loop_containing_stmt (def);
398 loop = get_loop (loop_nb);
400 if (def_loop == NULL)
401 return false;
403 if (loop == def_loop || flow_loop_nested_p (loop, def_loop))
404 return true;
406 return false;
409 n = TREE_OPERAND_LENGTH (chrec);
410 for (i = 0; i < n; i++)
411 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, i),
412 loop_nb))
413 return true;
414 return false;
417 /* Return true when PHI is a loop-phi-node. */
419 static bool
420 loop_phi_node_p (gimple phi)
422 /* The implementation of this function is based on the following
423 property: "all the loop-phi-nodes of a loop are contained in the
424 loop's header basic block". */
426 return loop_containing_stmt (phi)->header == gimple_bb (phi);
429 /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
430 In general, in the case of multivariate evolutions we want to get
431 the evolution in different loops. LOOP specifies the level for
432 which to get the evolution.
434 Example:
436 | for (j = 0; j < 100; j++)
438 | for (k = 0; k < 100; k++)
440 | i = k + j; - Here the value of i is a function of j, k.
442 | ... = i - Here the value of i is a function of j.
444 | ... = i - Here the value of i is a scalar.
446 Example:
448 | i_0 = ...
449 | loop_1 10 times
450 | i_1 = phi (i_0, i_2)
451 | i_2 = i_1 + 2
452 | endloop
454 This loop has the same effect as:
455 LOOP_1 has the same effect as:
457 | i_1 = i_0 + 20
459 The overall effect of the loop, "i_0 + 20" in the previous example,
460 is obtained by passing in the parameters: LOOP = 1,
461 EVOLUTION_FN = {i_0, +, 2}_1.
464 tree
465 compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn)
467 bool val = false;
469 if (evolution_fn == chrec_dont_know)
470 return chrec_dont_know;
472 else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC)
474 struct loop *inner_loop = get_chrec_loop (evolution_fn);
476 if (inner_loop == loop
477 || flow_loop_nested_p (loop, inner_loop))
479 tree nb_iter = number_of_latch_executions (inner_loop);
481 if (nb_iter == chrec_dont_know)
482 return chrec_dont_know;
483 else
485 tree res;
487 /* evolution_fn is the evolution function in LOOP. Get
488 its value in the nb_iter-th iteration. */
489 res = chrec_apply (inner_loop->num, evolution_fn, nb_iter);
491 if (chrec_contains_symbols_defined_in_loop (res, loop->num))
492 res = instantiate_parameters (loop, res);
494 /* Continue the computation until ending on a parent of LOOP. */
495 return compute_overall_effect_of_inner_loop (loop, res);
498 else
499 return evolution_fn;
502 /* If the evolution function is an invariant, there is nothing to do. */
503 else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val)
504 return evolution_fn;
506 else
507 return chrec_dont_know;
510 /* Determine whether the CHREC is always positive/negative. If the expression
511 cannot be statically analyzed, return false, otherwise set the answer into
512 VALUE. */
514 bool
515 chrec_is_positive (tree chrec, bool *value)
517 bool value0, value1, value2;
518 tree end_value, nb_iter;
520 switch (TREE_CODE (chrec))
522 case POLYNOMIAL_CHREC:
523 if (!chrec_is_positive (CHREC_LEFT (chrec), &value0)
524 || !chrec_is_positive (CHREC_RIGHT (chrec), &value1))
525 return false;
527 /* FIXME -- overflows. */
528 if (value0 == value1)
530 *value = value0;
531 return true;
534 /* Otherwise the chrec is under the form: "{-197, +, 2}_1",
535 and the proof consists in showing that the sign never
536 changes during the execution of the loop, from 0 to
537 loop->nb_iterations. */
538 if (!evolution_function_is_affine_p (chrec))
539 return false;
541 nb_iter = number_of_latch_executions (get_chrec_loop (chrec));
542 if (chrec_contains_undetermined (nb_iter))
543 return false;
545 #if 0
546 /* TODO -- If the test is after the exit, we may decrease the number of
547 iterations by one. */
548 if (after_exit)
549 nb_iter = chrec_fold_minus (type, nb_iter, build_int_cst (type, 1));
550 #endif
552 end_value = chrec_apply (CHREC_VARIABLE (chrec), chrec, nb_iter);
554 if (!chrec_is_positive (end_value, &value2))
555 return false;
557 *value = value0;
558 return value0 == value1;
560 case INTEGER_CST:
561 *value = (tree_int_cst_sgn (chrec) == 1);
562 return true;
564 default:
565 return false;
569 /* Associate CHREC to SCALAR. */
571 static void
572 set_scalar_evolution (basic_block instantiated_below, tree scalar, tree chrec)
574 tree *scalar_info;
576 if (TREE_CODE (scalar) != SSA_NAME)
577 return;
579 scalar_info = find_var_scev_info (instantiated_below, scalar);
581 if (dump_file)
583 if (dump_flags & TDF_DETAILS)
585 fprintf (dump_file, "(set_scalar_evolution \n");
586 fprintf (dump_file, " instantiated_below = %d \n",
587 instantiated_below->index);
588 fprintf (dump_file, " (scalar = ");
589 print_generic_expr (dump_file, scalar, 0);
590 fprintf (dump_file, ")\n (scalar_evolution = ");
591 print_generic_expr (dump_file, chrec, 0);
592 fprintf (dump_file, "))\n");
594 if (dump_flags & TDF_STATS)
595 nb_set_scev++;
598 *scalar_info = chrec;
601 /* Retrieve the chrec associated to SCALAR instantiated below
602 INSTANTIATED_BELOW block. */
604 static tree
605 get_scalar_evolution (basic_block instantiated_below, tree scalar)
607 tree res;
609 if (dump_file)
611 if (dump_flags & TDF_DETAILS)
613 fprintf (dump_file, "(get_scalar_evolution \n");
614 fprintf (dump_file, " (scalar = ");
615 print_generic_expr (dump_file, scalar, 0);
616 fprintf (dump_file, ")\n");
618 if (dump_flags & TDF_STATS)
619 nb_get_scev++;
622 switch (TREE_CODE (scalar))
624 case SSA_NAME:
625 res = *find_var_scev_info (instantiated_below, scalar);
626 break;
628 case REAL_CST:
629 case FIXED_CST:
630 case INTEGER_CST:
631 res = scalar;
632 break;
634 default:
635 res = chrec_not_analyzed_yet;
636 break;
639 if (dump_file && (dump_flags & TDF_DETAILS))
641 fprintf (dump_file, " (scalar_evolution = ");
642 print_generic_expr (dump_file, res, 0);
643 fprintf (dump_file, "))\n");
646 return res;
649 /* Helper function for add_to_evolution. Returns the evolution
650 function for an assignment of the form "a = b + c", where "a" and
651 "b" are on the strongly connected component. CHREC_BEFORE is the
652 information that we already have collected up to this point.
653 TO_ADD is the evolution of "c".
655 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
656 evolution the expression TO_ADD, otherwise construct an evolution
657 part for this loop. */
659 static tree
660 add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add,
661 gimple at_stmt)
663 tree type, left, right;
664 struct loop *loop = get_loop (loop_nb), *chloop;
666 switch (TREE_CODE (chrec_before))
668 case POLYNOMIAL_CHREC:
669 chloop = get_chrec_loop (chrec_before);
670 if (chloop == loop
671 || flow_loop_nested_p (chloop, loop))
673 unsigned var;
675 type = chrec_type (chrec_before);
677 /* When there is no evolution part in this loop, build it. */
678 if (chloop != loop)
680 var = loop_nb;
681 left = chrec_before;
682 right = SCALAR_FLOAT_TYPE_P (type)
683 ? build_real (type, dconst0)
684 : build_int_cst (type, 0);
686 else
688 var = CHREC_VARIABLE (chrec_before);
689 left = CHREC_LEFT (chrec_before);
690 right = CHREC_RIGHT (chrec_before);
693 to_add = chrec_convert (type, to_add, at_stmt);
694 right = chrec_convert_rhs (type, right, at_stmt);
695 right = chrec_fold_plus (chrec_type (right), right, to_add);
696 return build_polynomial_chrec (var, left, right);
698 else
700 gcc_assert (flow_loop_nested_p (loop, chloop));
702 /* Search the evolution in LOOP_NB. */
703 left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before),
704 to_add, at_stmt);
705 right = CHREC_RIGHT (chrec_before);
706 right = chrec_convert_rhs (chrec_type (left), right, at_stmt);
707 return build_polynomial_chrec (CHREC_VARIABLE (chrec_before),
708 left, right);
711 default:
712 /* These nodes do not depend on a loop. */
713 if (chrec_before == chrec_dont_know)
714 return chrec_dont_know;
716 left = chrec_before;
717 right = chrec_convert_rhs (chrec_type (left), to_add, at_stmt);
718 return build_polynomial_chrec (loop_nb, left, right);
722 /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
723 of LOOP_NB.
725 Description (provided for completeness, for those who read code in
726 a plane, and for my poor 62 bytes brain that would have forgotten
727 all this in the next two or three months):
729 The algorithm of translation of programs from the SSA representation
730 into the chrecs syntax is based on a pattern matching. After having
731 reconstructed the overall tree expression for a loop, there are only
732 two cases that can arise:
734 1. a = loop-phi (init, a + expr)
735 2. a = loop-phi (init, expr)
737 where EXPR is either a scalar constant with respect to the analyzed
738 loop (this is a degree 0 polynomial), or an expression containing
739 other loop-phi definitions (these are higher degree polynomials).
741 Examples:
744 | init = ...
745 | loop_1
746 | a = phi (init, a + 5)
747 | endloop
750 | inita = ...
751 | initb = ...
752 | loop_1
753 | a = phi (inita, 2 * b + 3)
754 | b = phi (initb, b + 1)
755 | endloop
757 For the first case, the semantics of the SSA representation is:
759 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
761 that is, there is a loop index "x" that determines the scalar value
762 of the variable during the loop execution. During the first
763 iteration, the value is that of the initial condition INIT, while
764 during the subsequent iterations, it is the sum of the initial
765 condition with the sum of all the values of EXPR from the initial
766 iteration to the before last considered iteration.
768 For the second case, the semantics of the SSA program is:
770 | a (x) = init, if x = 0;
771 | expr (x - 1), otherwise.
773 The second case corresponds to the PEELED_CHREC, whose syntax is
774 close to the syntax of a loop-phi-node:
776 | phi (init, expr) vs. (init, expr)_x
778 The proof of the translation algorithm for the first case is a
779 proof by structural induction based on the degree of EXPR.
781 Degree 0:
782 When EXPR is a constant with respect to the analyzed loop, or in
783 other words when EXPR is a polynomial of degree 0, the evolution of
784 the variable A in the loop is an affine function with an initial
785 condition INIT, and a step EXPR. In order to show this, we start
786 from the semantics of the SSA representation:
788 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
790 and since "expr (j)" is a constant with respect to "j",
792 f (x) = init + x * expr
794 Finally, based on the semantics of the pure sum chrecs, by
795 identification we get the corresponding chrecs syntax:
797 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
798 f (x) -> {init, +, expr}_x
800 Higher degree:
801 Suppose that EXPR is a polynomial of degree N with respect to the
802 analyzed loop_x for which we have already determined that it is
803 written under the chrecs syntax:
805 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
807 We start from the semantics of the SSA program:
809 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
811 | f (x) = init + \sum_{j = 0}^{x - 1}
812 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
814 | f (x) = init + \sum_{j = 0}^{x - 1}
815 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
817 | f (x) = init + \sum_{k = 0}^{n - 1}
818 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
820 | f (x) = init + \sum_{k = 0}^{n - 1}
821 | (b_k * \binom{x}{k + 1})
823 | f (x) = init + b_0 * \binom{x}{1} + ...
824 | + b_{n-1} * \binom{x}{n}
826 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
827 | + b_{n-1} * \binom{x}{n}
830 And finally from the definition of the chrecs syntax, we identify:
831 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
833 This shows the mechanism that stands behind the add_to_evolution
834 function. An important point is that the use of symbolic
835 parameters avoids the need of an analysis schedule.
837 Example:
839 | inita = ...
840 | initb = ...
841 | loop_1
842 | a = phi (inita, a + 2 + b)
843 | b = phi (initb, b + 1)
844 | endloop
846 When analyzing "a", the algorithm keeps "b" symbolically:
848 | a -> {inita, +, 2 + b}_1
850 Then, after instantiation, the analyzer ends on the evolution:
852 | a -> {inita, +, 2 + initb, +, 1}_1
856 static tree
857 add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code,
858 tree to_add, gimple at_stmt)
860 tree type = chrec_type (to_add);
861 tree res = NULL_TREE;
863 if (to_add == NULL_TREE)
864 return chrec_before;
866 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
867 instantiated at this point. */
868 if (TREE_CODE (to_add) == POLYNOMIAL_CHREC)
869 /* This should not happen. */
870 return chrec_dont_know;
872 if (dump_file && (dump_flags & TDF_DETAILS))
874 fprintf (dump_file, "(add_to_evolution \n");
875 fprintf (dump_file, " (loop_nb = %d)\n", loop_nb);
876 fprintf (dump_file, " (chrec_before = ");
877 print_generic_expr (dump_file, chrec_before, 0);
878 fprintf (dump_file, ")\n (to_add = ");
879 print_generic_expr (dump_file, to_add, 0);
880 fprintf (dump_file, ")\n");
883 if (code == MINUS_EXPR)
884 to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type)
885 ? build_real (type, dconstm1)
886 : build_int_cst_type (type, -1));
888 res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt);
890 if (dump_file && (dump_flags & TDF_DETAILS))
892 fprintf (dump_file, " (res = ");
893 print_generic_expr (dump_file, res, 0);
894 fprintf (dump_file, "))\n");
897 return res;
902 /* This section selects the loops that will be good candidates for the
903 scalar evolution analysis. For the moment, greedily select all the
904 loop nests we could analyze. */
906 /* For a loop with a single exit edge, return the COND_EXPR that
907 guards the exit edge. If the expression is too difficult to
908 analyze, then give up. */
910 gimple
911 get_loop_exit_condition (const struct loop *loop)
913 gimple res = NULL;
914 edge exit_edge = single_exit (loop);
916 if (dump_file && (dump_flags & TDF_DETAILS))
917 fprintf (dump_file, "(get_loop_exit_condition \n ");
919 if (exit_edge)
921 gimple stmt;
923 stmt = last_stmt (exit_edge->src);
924 if (gimple_code (stmt) == GIMPLE_COND)
925 res = stmt;
928 if (dump_file && (dump_flags & TDF_DETAILS))
930 print_gimple_stmt (dump_file, res, 0, 0);
931 fprintf (dump_file, ")\n");
934 return res;
937 /* Recursively determine and enqueue the exit conditions for a loop. */
939 static void
940 get_exit_conditions_rec (struct loop *loop,
941 VEC(gimple,heap) **exit_conditions)
943 if (!loop)
944 return;
946 /* Recurse on the inner loops, then on the next (sibling) loops. */
947 get_exit_conditions_rec (loop->inner, exit_conditions);
948 get_exit_conditions_rec (loop->next, exit_conditions);
950 if (single_exit (loop))
952 gimple loop_condition = get_loop_exit_condition (loop);
954 if (loop_condition)
955 VEC_safe_push (gimple, heap, *exit_conditions, loop_condition);
959 /* Select the candidate loop nests for the analysis. This function
960 initializes the EXIT_CONDITIONS array. */
962 static void
963 select_loops_exit_conditions (VEC(gimple,heap) **exit_conditions)
965 struct loop *function_body = current_loops->tree_root;
967 get_exit_conditions_rec (function_body->inner, exit_conditions);
971 /* Depth first search algorithm. */
973 typedef enum t_bool {
974 t_false,
975 t_true,
976 t_dont_know
977 } t_bool;
980 static t_bool follow_ssa_edge (struct loop *loop, gimple, gimple, tree *, int);
982 /* Follow the ssa edge into the binary expression RHS0 CODE RHS1.
983 Return true if the strongly connected component has been found. */
985 static t_bool
986 follow_ssa_edge_binary (struct loop *loop, gimple at_stmt,
987 tree type, tree rhs0, enum tree_code code, tree rhs1,
988 gimple halting_phi, tree *evolution_of_loop, int limit)
990 t_bool res = t_false;
991 tree evol;
993 switch (code)
995 case POINTER_PLUS_EXPR:
996 case PLUS_EXPR:
997 if (TREE_CODE (rhs0) == SSA_NAME)
999 if (TREE_CODE (rhs1) == SSA_NAME)
1001 /* Match an assignment under the form:
1002 "a = b + c". */
1004 /* We want only assignments of form "name + name" contribute to
1005 LIMIT, as the other cases do not necessarily contribute to
1006 the complexity of the expression. */
1007 limit++;
1009 evol = *evolution_of_loop;
1010 res = follow_ssa_edge
1011 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, &evol, limit);
1013 if (res == t_true)
1014 *evolution_of_loop = add_to_evolution
1015 (loop->num,
1016 chrec_convert (type, evol, at_stmt),
1017 code, rhs1, at_stmt);
1019 else if (res == t_false)
1021 res = follow_ssa_edge
1022 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1023 evolution_of_loop, limit);
1025 if (res == t_true)
1026 *evolution_of_loop = add_to_evolution
1027 (loop->num,
1028 chrec_convert (type, *evolution_of_loop, at_stmt),
1029 code, rhs0, at_stmt);
1031 else if (res == t_dont_know)
1032 *evolution_of_loop = chrec_dont_know;
1035 else if (res == t_dont_know)
1036 *evolution_of_loop = chrec_dont_know;
1039 else
1041 /* Match an assignment under the form:
1042 "a = b + ...". */
1043 res = follow_ssa_edge
1044 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1045 evolution_of_loop, limit);
1046 if (res == t_true)
1047 *evolution_of_loop = add_to_evolution
1048 (loop->num, chrec_convert (type, *evolution_of_loop,
1049 at_stmt),
1050 code, rhs1, at_stmt);
1052 else if (res == t_dont_know)
1053 *evolution_of_loop = chrec_dont_know;
1057 else if (TREE_CODE (rhs1) == SSA_NAME)
1059 /* Match an assignment under the form:
1060 "a = ... + c". */
1061 res = follow_ssa_edge
1062 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1063 evolution_of_loop, limit);
1064 if (res == t_true)
1065 *evolution_of_loop = add_to_evolution
1066 (loop->num, chrec_convert (type, *evolution_of_loop,
1067 at_stmt),
1068 code, rhs0, at_stmt);
1070 else if (res == t_dont_know)
1071 *evolution_of_loop = chrec_dont_know;
1074 else
1075 /* Otherwise, match an assignment under the form:
1076 "a = ... + ...". */
1077 /* And there is nothing to do. */
1078 res = t_false;
1079 break;
1081 case MINUS_EXPR:
1082 /* This case is under the form "opnd0 = rhs0 - rhs1". */
1083 if (TREE_CODE (rhs0) == SSA_NAME)
1085 /* Match an assignment under the form:
1086 "a = b - ...". */
1088 /* We want only assignments of form "name - name" contribute to
1089 LIMIT, as the other cases do not necessarily contribute to
1090 the complexity of the expression. */
1091 if (TREE_CODE (rhs1) == SSA_NAME)
1092 limit++;
1094 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1095 evolution_of_loop, limit);
1096 if (res == t_true)
1097 *evolution_of_loop = add_to_evolution
1098 (loop->num, chrec_convert (type, *evolution_of_loop, at_stmt),
1099 MINUS_EXPR, rhs1, at_stmt);
1101 else if (res == t_dont_know)
1102 *evolution_of_loop = chrec_dont_know;
1104 else
1105 /* Otherwise, match an assignment under the form:
1106 "a = ... - ...". */
1107 /* And there is nothing to do. */
1108 res = t_false;
1109 break;
1111 default:
1112 res = t_false;
1115 return res;
1118 /* Follow the ssa edge into the expression EXPR.
1119 Return true if the strongly connected component has been found. */
1121 static t_bool
1122 follow_ssa_edge_expr (struct loop *loop, gimple at_stmt, tree expr,
1123 gimple halting_phi, tree *evolution_of_loop, int limit)
1125 enum tree_code code = TREE_CODE (expr);
1126 tree type = TREE_TYPE (expr), rhs0, rhs1;
1127 t_bool res;
1129 /* The EXPR is one of the following cases:
1130 - an SSA_NAME,
1131 - an INTEGER_CST,
1132 - a PLUS_EXPR,
1133 - a POINTER_PLUS_EXPR,
1134 - a MINUS_EXPR,
1135 - an ASSERT_EXPR,
1136 - other cases are not yet handled. */
1138 switch (code)
1140 CASE_CONVERT:
1141 /* This assignment is under the form "a_1 = (cast) rhs. */
1142 res = follow_ssa_edge_expr (loop, at_stmt, TREE_OPERAND (expr, 0),
1143 halting_phi, evolution_of_loop, limit);
1144 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, at_stmt);
1145 break;
1147 case INTEGER_CST:
1148 /* This assignment is under the form "a_1 = 7". */
1149 res = t_false;
1150 break;
1152 case SSA_NAME:
1153 /* This assignment is under the form: "a_1 = b_2". */
1154 res = follow_ssa_edge
1155 (loop, SSA_NAME_DEF_STMT (expr), halting_phi, evolution_of_loop, limit);
1156 break;
1158 case POINTER_PLUS_EXPR:
1159 case PLUS_EXPR:
1160 case MINUS_EXPR:
1161 /* This case is under the form "rhs0 +- rhs1". */
1162 rhs0 = TREE_OPERAND (expr, 0);
1163 rhs1 = TREE_OPERAND (expr, 1);
1164 type = TREE_TYPE (rhs0);
1165 STRIP_USELESS_TYPE_CONVERSION (rhs0);
1166 STRIP_USELESS_TYPE_CONVERSION (rhs1);
1167 res = follow_ssa_edge_binary (loop, at_stmt, type, rhs0, code, rhs1,
1168 halting_phi, evolution_of_loop, limit);
1169 break;
1171 case ADDR_EXPR:
1172 /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */
1173 if (TREE_CODE (TREE_OPERAND (expr, 0)) == MEM_REF)
1175 expr = TREE_OPERAND (expr, 0);
1176 rhs0 = TREE_OPERAND (expr, 0);
1177 rhs1 = TREE_OPERAND (expr, 1);
1178 type = TREE_TYPE (rhs0);
1179 STRIP_USELESS_TYPE_CONVERSION (rhs0);
1180 STRIP_USELESS_TYPE_CONVERSION (rhs1);
1181 res = follow_ssa_edge_binary (loop, at_stmt, type,
1182 rhs0, POINTER_PLUS_EXPR, rhs1,
1183 halting_phi, evolution_of_loop, limit);
1185 else
1186 res = t_false;
1187 break;
1189 case ASSERT_EXPR:
1190 /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>"
1191 It must be handled as a copy assignment of the form a_1 = a_2. */
1192 rhs0 = ASSERT_EXPR_VAR (expr);
1193 if (TREE_CODE (rhs0) == SSA_NAME)
1194 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0),
1195 halting_phi, evolution_of_loop, limit);
1196 else
1197 res = t_false;
1198 break;
1200 default:
1201 res = t_false;
1202 break;
1205 return res;
1208 /* Follow the ssa edge into the right hand side of an assignment STMT.
1209 Return true if the strongly connected component has been found. */
1211 static t_bool
1212 follow_ssa_edge_in_rhs (struct loop *loop, gimple stmt,
1213 gimple halting_phi, tree *evolution_of_loop, int limit)
1215 enum tree_code code = gimple_assign_rhs_code (stmt);
1216 tree type = gimple_expr_type (stmt), rhs1, rhs2;
1217 t_bool res;
1219 switch (code)
1221 CASE_CONVERT:
1222 /* This assignment is under the form "a_1 = (cast) rhs. */
1223 res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1224 halting_phi, evolution_of_loop, limit);
1225 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, stmt);
1226 break;
1228 case POINTER_PLUS_EXPR:
1229 case PLUS_EXPR:
1230 case MINUS_EXPR:
1231 rhs1 = gimple_assign_rhs1 (stmt);
1232 rhs2 = gimple_assign_rhs2 (stmt);
1233 type = TREE_TYPE (rhs1);
1234 res = follow_ssa_edge_binary (loop, stmt, type, rhs1, code, rhs2,
1235 halting_phi, evolution_of_loop, limit);
1236 break;
1238 default:
1239 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1240 res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1241 halting_phi, evolution_of_loop, limit);
1242 else
1243 res = t_false;
1244 break;
1247 return res;
1250 /* Checks whether the I-th argument of a PHI comes from a backedge. */
1252 static bool
1253 backedge_phi_arg_p (gimple phi, int i)
1255 const_edge e = gimple_phi_arg_edge (phi, i);
1257 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1258 about updating it anywhere, and this should work as well most of the
1259 time. */
1260 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1261 return true;
1263 return false;
1266 /* Helper function for one branch of the condition-phi-node. Return
1267 true if the strongly connected component has been found following
1268 this path. */
1270 static inline t_bool
1271 follow_ssa_edge_in_condition_phi_branch (int i,
1272 struct loop *loop,
1273 gimple condition_phi,
1274 gimple halting_phi,
1275 tree *evolution_of_branch,
1276 tree init_cond, int limit)
1278 tree branch = PHI_ARG_DEF (condition_phi, i);
1279 *evolution_of_branch = chrec_dont_know;
1281 /* Do not follow back edges (they must belong to an irreducible loop, which
1282 we really do not want to worry about). */
1283 if (backedge_phi_arg_p (condition_phi, i))
1284 return t_false;
1286 if (TREE_CODE (branch) == SSA_NAME)
1288 *evolution_of_branch = init_cond;
1289 return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi,
1290 evolution_of_branch, limit);
1293 /* This case occurs when one of the condition branches sets
1294 the variable to a constant: i.e. a phi-node like
1295 "a_2 = PHI <a_7(5), 2(6)>;".
1297 FIXME: This case have to be refined correctly:
1298 in some cases it is possible to say something better than
1299 chrec_dont_know, for example using a wrap-around notation. */
1300 return t_false;
1303 /* This function merges the branches of a condition-phi-node in a
1304 loop. */
1306 static t_bool
1307 follow_ssa_edge_in_condition_phi (struct loop *loop,
1308 gimple condition_phi,
1309 gimple halting_phi,
1310 tree *evolution_of_loop, int limit)
1312 int i, n;
1313 tree init = *evolution_of_loop;
1314 tree evolution_of_branch;
1315 t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi,
1316 halting_phi,
1317 &evolution_of_branch,
1318 init, limit);
1319 if (res == t_false || res == t_dont_know)
1320 return res;
1322 *evolution_of_loop = evolution_of_branch;
1324 n = gimple_phi_num_args (condition_phi);
1325 for (i = 1; i < n; i++)
1327 /* Quickly give up when the evolution of one of the branches is
1328 not known. */
1329 if (*evolution_of_loop == chrec_dont_know)
1330 return t_true;
1332 /* Increase the limit by the PHI argument number to avoid exponential
1333 time and memory complexity. */
1334 res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi,
1335 halting_phi,
1336 &evolution_of_branch,
1337 init, limit + i);
1338 if (res == t_false || res == t_dont_know)
1339 return res;
1341 *evolution_of_loop = chrec_merge (*evolution_of_loop,
1342 evolution_of_branch);
1345 return t_true;
1348 /* Follow an SSA edge in an inner loop. It computes the overall
1349 effect of the loop, and following the symbolic initial conditions,
1350 it follows the edges in the parent loop. The inner loop is
1351 considered as a single statement. */
1353 static t_bool
1354 follow_ssa_edge_inner_loop_phi (struct loop *outer_loop,
1355 gimple loop_phi_node,
1356 gimple halting_phi,
1357 tree *evolution_of_loop, int limit)
1359 struct loop *loop = loop_containing_stmt (loop_phi_node);
1360 tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node));
1362 /* Sometimes, the inner loop is too difficult to analyze, and the
1363 result of the analysis is a symbolic parameter. */
1364 if (ev == PHI_RESULT (loop_phi_node))
1366 t_bool res = t_false;
1367 int i, n = gimple_phi_num_args (loop_phi_node);
1369 for (i = 0; i < n; i++)
1371 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1372 basic_block bb;
1374 /* Follow the edges that exit the inner loop. */
1375 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1376 if (!flow_bb_inside_loop_p (loop, bb))
1377 res = follow_ssa_edge_expr (outer_loop, loop_phi_node,
1378 arg, halting_phi,
1379 evolution_of_loop, limit);
1380 if (res == t_true)
1381 break;
1384 /* If the path crosses this loop-phi, give up. */
1385 if (res == t_true)
1386 *evolution_of_loop = chrec_dont_know;
1388 return res;
1391 /* Otherwise, compute the overall effect of the inner loop. */
1392 ev = compute_overall_effect_of_inner_loop (loop, ev);
1393 return follow_ssa_edge_expr (outer_loop, loop_phi_node, ev, halting_phi,
1394 evolution_of_loop, limit);
1397 /* Follow an SSA edge from a loop-phi-node to itself, constructing a
1398 path that is analyzed on the return walk. */
1400 static t_bool
1401 follow_ssa_edge (struct loop *loop, gimple def, gimple halting_phi,
1402 tree *evolution_of_loop, int limit)
1404 struct loop *def_loop;
1406 if (gimple_nop_p (def))
1407 return t_false;
1409 /* Give up if the path is longer than the MAX that we allow. */
1410 if (limit > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
1411 return t_dont_know;
1413 def_loop = loop_containing_stmt (def);
1415 switch (gimple_code (def))
1417 case GIMPLE_PHI:
1418 if (!loop_phi_node_p (def))
1419 /* DEF is a condition-phi-node. Follow the branches, and
1420 record their evolutions. Finally, merge the collected
1421 information and set the approximation to the main
1422 variable. */
1423 return follow_ssa_edge_in_condition_phi
1424 (loop, def, halting_phi, evolution_of_loop, limit);
1426 /* When the analyzed phi is the halting_phi, the
1427 depth-first search is over: we have found a path from
1428 the halting_phi to itself in the loop. */
1429 if (def == halting_phi)
1430 return t_true;
1432 /* Otherwise, the evolution of the HALTING_PHI depends
1433 on the evolution of another loop-phi-node, i.e. the
1434 evolution function is a higher degree polynomial. */
1435 if (def_loop == loop)
1436 return t_false;
1438 /* Inner loop. */
1439 if (flow_loop_nested_p (loop, def_loop))
1440 return follow_ssa_edge_inner_loop_phi
1441 (loop, def, halting_phi, evolution_of_loop, limit + 1);
1443 /* Outer loop. */
1444 return t_false;
1446 case GIMPLE_ASSIGN:
1447 return follow_ssa_edge_in_rhs (loop, def, halting_phi,
1448 evolution_of_loop, limit);
1450 default:
1451 /* At this level of abstraction, the program is just a set
1452 of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no
1453 other node to be handled. */
1454 return t_false;
1460 /* Given a LOOP_PHI_NODE, this function determines the evolution
1461 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1463 static tree
1464 analyze_evolution_in_loop (gimple loop_phi_node,
1465 tree init_cond)
1467 int i, n = gimple_phi_num_args (loop_phi_node);
1468 tree evolution_function = chrec_not_analyzed_yet;
1469 struct loop *loop = loop_containing_stmt (loop_phi_node);
1470 basic_block bb;
1472 if (dump_file && (dump_flags & TDF_DETAILS))
1474 fprintf (dump_file, "(analyze_evolution_in_loop \n");
1475 fprintf (dump_file, " (loop_phi_node = ");
1476 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
1477 fprintf (dump_file, ")\n");
1480 for (i = 0; i < n; i++)
1482 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1483 gimple ssa_chain;
1484 tree ev_fn;
1485 t_bool res;
1487 /* Select the edges that enter the loop body. */
1488 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1489 if (!flow_bb_inside_loop_p (loop, bb))
1490 continue;
1492 if (TREE_CODE (arg) == SSA_NAME)
1494 bool val = false;
1496 ssa_chain = SSA_NAME_DEF_STMT (arg);
1498 /* Pass in the initial condition to the follow edge function. */
1499 ev_fn = init_cond;
1500 res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0);
1502 /* If ev_fn has no evolution in the inner loop, and the
1503 init_cond is not equal to ev_fn, then we have an
1504 ambiguity between two possible values, as we cannot know
1505 the number of iterations at this point. */
1506 if (TREE_CODE (ev_fn) != POLYNOMIAL_CHREC
1507 && no_evolution_in_loop_p (ev_fn, loop->num, &val) && val
1508 && !operand_equal_p (init_cond, ev_fn, 0))
1509 ev_fn = chrec_dont_know;
1511 else
1512 res = t_false;
1514 /* When it is impossible to go back on the same
1515 loop_phi_node by following the ssa edges, the
1516 evolution is represented by a peeled chrec, i.e. the
1517 first iteration, EV_FN has the value INIT_COND, then
1518 all the other iterations it has the value of ARG.
1519 For the moment, PEELED_CHREC nodes are not built. */
1520 if (res != t_true)
1521 ev_fn = chrec_dont_know;
1523 /* When there are multiple back edges of the loop (which in fact never
1524 happens currently, but nevertheless), merge their evolutions. */
1525 evolution_function = chrec_merge (evolution_function, ev_fn);
1528 if (dump_file && (dump_flags & TDF_DETAILS))
1530 fprintf (dump_file, " (evolution_function = ");
1531 print_generic_expr (dump_file, evolution_function, 0);
1532 fprintf (dump_file, "))\n");
1535 return evolution_function;
1538 /* Given a loop-phi-node, return the initial conditions of the
1539 variable on entry of the loop. When the CCP has propagated
1540 constants into the loop-phi-node, the initial condition is
1541 instantiated, otherwise the initial condition is kept symbolic.
1542 This analyzer does not analyze the evolution outside the current
1543 loop, and leaves this task to the on-demand tree reconstructor. */
1545 static tree
1546 analyze_initial_condition (gimple loop_phi_node)
1548 int i, n;
1549 tree init_cond = chrec_not_analyzed_yet;
1550 struct loop *loop = loop_containing_stmt (loop_phi_node);
1552 if (dump_file && (dump_flags & TDF_DETAILS))
1554 fprintf (dump_file, "(analyze_initial_condition \n");
1555 fprintf (dump_file, " (loop_phi_node = \n");
1556 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
1557 fprintf (dump_file, ")\n");
1560 n = gimple_phi_num_args (loop_phi_node);
1561 for (i = 0; i < n; i++)
1563 tree branch = PHI_ARG_DEF (loop_phi_node, i);
1564 basic_block bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1566 /* When the branch is oriented to the loop's body, it does
1567 not contribute to the initial condition. */
1568 if (flow_bb_inside_loop_p (loop, bb))
1569 continue;
1571 if (init_cond == chrec_not_analyzed_yet)
1573 init_cond = branch;
1574 continue;
1577 if (TREE_CODE (branch) == SSA_NAME)
1579 init_cond = chrec_dont_know;
1580 break;
1583 init_cond = chrec_merge (init_cond, branch);
1586 /* Ooops -- a loop without an entry??? */
1587 if (init_cond == chrec_not_analyzed_yet)
1588 init_cond = chrec_dont_know;
1590 /* During early loop unrolling we do not have fully constant propagated IL.
1591 Handle degenerate PHIs here to not miss important unrollings. */
1592 if (TREE_CODE (init_cond) == SSA_NAME)
1594 gimple def = SSA_NAME_DEF_STMT (init_cond);
1595 tree res;
1596 if (gimple_code (def) == GIMPLE_PHI
1597 && (res = degenerate_phi_result (def)) != NULL_TREE
1598 /* Only allow invariants here, otherwise we may break
1599 loop-closed SSA form. */
1600 && is_gimple_min_invariant (res))
1601 init_cond = res;
1604 if (dump_file && (dump_flags & TDF_DETAILS))
1606 fprintf (dump_file, " (init_cond = ");
1607 print_generic_expr (dump_file, init_cond, 0);
1608 fprintf (dump_file, "))\n");
1611 return init_cond;
1614 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1616 static tree
1617 interpret_loop_phi (struct loop *loop, gimple loop_phi_node)
1619 tree res;
1620 struct loop *phi_loop = loop_containing_stmt (loop_phi_node);
1621 tree init_cond;
1623 if (phi_loop != loop)
1625 struct loop *subloop;
1626 tree evolution_fn = analyze_scalar_evolution
1627 (phi_loop, PHI_RESULT (loop_phi_node));
1629 /* Dive one level deeper. */
1630 subloop = superloop_at_depth (phi_loop, loop_depth (loop) + 1);
1632 /* Interpret the subloop. */
1633 res = compute_overall_effect_of_inner_loop (subloop, evolution_fn);
1634 return res;
1637 /* Otherwise really interpret the loop phi. */
1638 init_cond = analyze_initial_condition (loop_phi_node);
1639 res = analyze_evolution_in_loop (loop_phi_node, init_cond);
1641 /* Verify we maintained the correct initial condition throughout
1642 possible conversions in the SSA chain. */
1643 if (res != chrec_dont_know)
1645 tree new_init = res;
1646 if (CONVERT_EXPR_P (res)
1647 && TREE_CODE (TREE_OPERAND (res, 0)) == POLYNOMIAL_CHREC)
1648 new_init = fold_convert (TREE_TYPE (res),
1649 CHREC_LEFT (TREE_OPERAND (res, 0)));
1650 else if (TREE_CODE (res) == POLYNOMIAL_CHREC)
1651 new_init = CHREC_LEFT (res);
1652 STRIP_USELESS_TYPE_CONVERSION (new_init);
1653 gcc_assert (TREE_CODE (new_init) != POLYNOMIAL_CHREC);
1654 if (!operand_equal_p (init_cond, new_init, 0))
1655 return chrec_dont_know;
1658 return res;
1661 /* This function merges the branches of a condition-phi-node,
1662 contained in the outermost loop, and whose arguments are already
1663 analyzed. */
1665 static tree
1666 interpret_condition_phi (struct loop *loop, gimple condition_phi)
1668 int i, n = gimple_phi_num_args (condition_phi);
1669 tree res = chrec_not_analyzed_yet;
1671 for (i = 0; i < n; i++)
1673 tree branch_chrec;
1675 if (backedge_phi_arg_p (condition_phi, i))
1677 res = chrec_dont_know;
1678 break;
1681 branch_chrec = analyze_scalar_evolution
1682 (loop, PHI_ARG_DEF (condition_phi, i));
1684 res = chrec_merge (res, branch_chrec);
1687 return res;
1690 /* Interpret the operation RHS1 OP RHS2. If we didn't
1691 analyze this node before, follow the definitions until ending
1692 either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the
1693 return path, this function propagates evolutions (ala constant copy
1694 propagation). OPND1 is not a GIMPLE expression because we could
1695 analyze the effect of an inner loop: see interpret_loop_phi. */
1697 static tree
1698 interpret_rhs_expr (struct loop *loop, gimple at_stmt,
1699 tree type, tree rhs1, enum tree_code code, tree rhs2)
1701 tree res, chrec1, chrec2;
1703 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1705 if (is_gimple_min_invariant (rhs1))
1706 return chrec_convert (type, rhs1, at_stmt);
1708 if (code == SSA_NAME)
1709 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1710 at_stmt);
1712 if (code == ASSERT_EXPR)
1714 rhs1 = ASSERT_EXPR_VAR (rhs1);
1715 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1716 at_stmt);
1719 return chrec_dont_know;
1722 switch (code)
1724 case POINTER_PLUS_EXPR:
1725 chrec1 = analyze_scalar_evolution (loop, rhs1);
1726 chrec2 = analyze_scalar_evolution (loop, rhs2);
1727 chrec1 = chrec_convert (type, chrec1, at_stmt);
1728 chrec2 = chrec_convert (sizetype, chrec2, at_stmt);
1729 res = chrec_fold_plus (type, chrec1, chrec2);
1730 break;
1732 case PLUS_EXPR:
1733 chrec1 = analyze_scalar_evolution (loop, rhs1);
1734 chrec2 = analyze_scalar_evolution (loop, rhs2);
1735 chrec1 = chrec_convert (type, chrec1, at_stmt);
1736 chrec2 = chrec_convert (type, chrec2, at_stmt);
1737 res = chrec_fold_plus (type, chrec1, chrec2);
1738 break;
1740 case MINUS_EXPR:
1741 chrec1 = analyze_scalar_evolution (loop, rhs1);
1742 chrec2 = analyze_scalar_evolution (loop, rhs2);
1743 chrec1 = chrec_convert (type, chrec1, at_stmt);
1744 chrec2 = chrec_convert (type, chrec2, at_stmt);
1745 res = chrec_fold_minus (type, chrec1, chrec2);
1746 break;
1748 case NEGATE_EXPR:
1749 chrec1 = analyze_scalar_evolution (loop, rhs1);
1750 chrec1 = chrec_convert (type, chrec1, at_stmt);
1751 /* TYPE may be integer, real or complex, so use fold_convert. */
1752 res = chrec_fold_multiply (type, chrec1,
1753 fold_convert (type, integer_minus_one_node));
1754 break;
1756 case BIT_NOT_EXPR:
1757 /* Handle ~X as -1 - X. */
1758 chrec1 = analyze_scalar_evolution (loop, rhs1);
1759 chrec1 = chrec_convert (type, chrec1, at_stmt);
1760 res = chrec_fold_minus (type,
1761 fold_convert (type, integer_minus_one_node),
1762 chrec1);
1763 break;
1765 case MULT_EXPR:
1766 chrec1 = analyze_scalar_evolution (loop, rhs1);
1767 chrec2 = analyze_scalar_evolution (loop, rhs2);
1768 chrec1 = chrec_convert (type, chrec1, at_stmt);
1769 chrec2 = chrec_convert (type, chrec2, at_stmt);
1770 res = chrec_fold_multiply (type, chrec1, chrec2);
1771 break;
1773 CASE_CONVERT:
1774 chrec1 = analyze_scalar_evolution (loop, rhs1);
1775 res = chrec_convert (type, chrec1, at_stmt);
1776 break;
1778 default:
1779 res = chrec_dont_know;
1780 break;
1783 return res;
1786 /* Interpret the expression EXPR. */
1788 static tree
1789 interpret_expr (struct loop *loop, gimple at_stmt, tree expr)
1791 enum tree_code code;
1792 tree type = TREE_TYPE (expr), op0, op1;
1794 if (automatically_generated_chrec_p (expr))
1795 return expr;
1797 if (TREE_CODE (expr) == POLYNOMIAL_CHREC)
1798 return chrec_dont_know;
1800 extract_ops_from_tree (expr, &code, &op0, &op1);
1802 return interpret_rhs_expr (loop, at_stmt, type,
1803 op0, code, op1);
1806 /* Interpret the rhs of the assignment STMT. */
1808 static tree
1809 interpret_gimple_assign (struct loop *loop, gimple stmt)
1811 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
1812 enum tree_code code = gimple_assign_rhs_code (stmt);
1814 return interpret_rhs_expr (loop, stmt, type,
1815 gimple_assign_rhs1 (stmt), code,
1816 gimple_assign_rhs2 (stmt));
1821 /* This section contains all the entry points:
1822 - number_of_iterations_in_loop,
1823 - analyze_scalar_evolution,
1824 - instantiate_parameters.
1827 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1828 common ancestor of DEF_LOOP and USE_LOOP. */
1830 static tree
1831 compute_scalar_evolution_in_loop (struct loop *wrto_loop,
1832 struct loop *def_loop,
1833 tree ev)
1835 bool val;
1836 tree res;
1838 if (def_loop == wrto_loop)
1839 return ev;
1841 def_loop = superloop_at_depth (def_loop, loop_depth (wrto_loop) + 1);
1842 res = compute_overall_effect_of_inner_loop (def_loop, ev);
1844 if (no_evolution_in_loop_p (res, wrto_loop->num, &val) && val)
1845 return res;
1847 return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet);
1850 /* Helper recursive function. */
1852 static tree
1853 analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res)
1855 tree type = TREE_TYPE (var);
1856 gimple def;
1857 basic_block bb;
1858 struct loop *def_loop;
1860 if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE)
1861 return chrec_dont_know;
1863 if (TREE_CODE (var) != SSA_NAME)
1864 return interpret_expr (loop, NULL, var);
1866 def = SSA_NAME_DEF_STMT (var);
1867 bb = gimple_bb (def);
1868 def_loop = bb ? bb->loop_father : NULL;
1870 if (bb == NULL
1871 || !flow_bb_inside_loop_p (loop, bb))
1873 /* Keep the symbolic form. */
1874 res = var;
1875 goto set_and_end;
1878 if (res != chrec_not_analyzed_yet)
1880 if (loop != bb->loop_father)
1881 res = compute_scalar_evolution_in_loop
1882 (find_common_loop (loop, bb->loop_father), bb->loop_father, res);
1884 goto set_and_end;
1887 if (loop != def_loop)
1889 res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet);
1890 res = compute_scalar_evolution_in_loop (loop, def_loop, res);
1892 goto set_and_end;
1895 switch (gimple_code (def))
1897 case GIMPLE_ASSIGN:
1898 res = interpret_gimple_assign (loop, def);
1899 break;
1901 case GIMPLE_PHI:
1902 if (loop_phi_node_p (def))
1903 res = interpret_loop_phi (loop, def);
1904 else
1905 res = interpret_condition_phi (loop, def);
1906 break;
1908 default:
1909 res = chrec_dont_know;
1910 break;
1913 set_and_end:
1915 /* Keep the symbolic form. */
1916 if (res == chrec_dont_know)
1917 res = var;
1919 if (loop == def_loop)
1920 set_scalar_evolution (block_before_loop (loop), var, res);
1922 return res;
1925 /* Analyzes and returns the scalar evolution of the ssa_name VAR in
1926 LOOP. LOOP is the loop in which the variable is used.
1928 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
1929 pointer to the statement that uses this variable, in order to
1930 determine the evolution function of the variable, use the following
1931 calls:
1933 loop_p loop = loop_containing_stmt (stmt);
1934 tree chrec_with_symbols = analyze_scalar_evolution (loop, var);
1935 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
1938 tree
1939 analyze_scalar_evolution (struct loop *loop, tree var)
1941 tree res;
1943 if (dump_file && (dump_flags & TDF_DETAILS))
1945 fprintf (dump_file, "(analyze_scalar_evolution \n");
1946 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
1947 fprintf (dump_file, " (scalar = ");
1948 print_generic_expr (dump_file, var, 0);
1949 fprintf (dump_file, ")\n");
1952 res = get_scalar_evolution (block_before_loop (loop), var);
1953 res = analyze_scalar_evolution_1 (loop, var, res);
1955 if (dump_file && (dump_flags & TDF_DETAILS))
1956 fprintf (dump_file, ")\n");
1958 return res;
1961 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
1962 WRTO_LOOP (which should be a superloop of USE_LOOP)
1964 FOLDED_CASTS is set to true if resolve_mixers used
1965 chrec_convert_aggressive (TODO -- not really, we are way too conservative
1966 at the moment in order to keep things simple).
1968 To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following
1969 example:
1971 for (i = 0; i < 100; i++) -- loop 1
1973 for (j = 0; j < 100; j++) -- loop 2
1975 k1 = i;
1976 k2 = j;
1978 use2 (k1, k2);
1980 for (t = 0; t < 100; t++) -- loop 3
1981 use3 (k1, k2);
1984 use1 (k1, k2);
1987 Both k1 and k2 are invariants in loop3, thus
1988 analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1
1989 analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2
1991 As they are invariant, it does not matter whether we consider their
1992 usage in loop 3 or loop 2, hence
1993 analyze_scalar_evolution_in_loop (loop2, loop3, k1) =
1994 analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i
1995 analyze_scalar_evolution_in_loop (loop2, loop3, k2) =
1996 analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2
1998 Similarly for their evolutions with respect to loop 1. The values of K2
1999 in the use in loop 2 vary independently on loop 1, thus we cannot express
2000 the evolution with respect to loop 1:
2001 analyze_scalar_evolution_in_loop (loop1, loop3, k1) =
2002 analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1
2003 analyze_scalar_evolution_in_loop (loop1, loop3, k2) =
2004 analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know
2006 The value of k2 in the use in loop 1 is known, though:
2007 analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1
2008 analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100
2011 static tree
2012 analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop,
2013 tree version, bool *folded_casts)
2015 bool val = false;
2016 tree ev = version, tmp;
2018 /* We cannot just do
2020 tmp = analyze_scalar_evolution (use_loop, version);
2021 ev = resolve_mixers (wrto_loop, tmp);
2023 as resolve_mixers would query the scalar evolution with respect to
2024 wrto_loop. For example, in the situation described in the function
2025 comment, suppose that wrto_loop = loop1, use_loop = loop3 and
2026 version = k2. Then
2028 analyze_scalar_evolution (use_loop, version) = k2
2030 and resolve_mixers (loop1, k2) finds that the value of k2 in loop 1
2031 is 100, which is a wrong result, since we are interested in the
2032 value in loop 3.
2034 Instead, we need to proceed from use_loop to wrto_loop loop by loop,
2035 each time checking that there is no evolution in the inner loop. */
2037 if (folded_casts)
2038 *folded_casts = false;
2039 while (1)
2041 tmp = analyze_scalar_evolution (use_loop, ev);
2042 ev = resolve_mixers (use_loop, tmp);
2044 if (folded_casts && tmp != ev)
2045 *folded_casts = true;
2047 if (use_loop == wrto_loop)
2048 return ev;
2050 /* If the value of the use changes in the inner loop, we cannot express
2051 its value in the outer loop (we might try to return interval chrec,
2052 but we do not have a user for it anyway) */
2053 if (!no_evolution_in_loop_p (ev, use_loop->num, &val)
2054 || !val)
2055 return chrec_dont_know;
2057 use_loop = loop_outer (use_loop);
2061 /* Returns from CACHE the value for VERSION instantiated below
2062 INSTANTIATED_BELOW block. */
2064 static tree
2065 get_instantiated_value (htab_t cache, basic_block instantiated_below,
2066 tree version)
2068 struct scev_info_str *info, pattern;
2070 pattern.var = version;
2071 pattern.instantiated_below = instantiated_below;
2072 info = (struct scev_info_str *) htab_find (cache, &pattern);
2074 if (info)
2075 return info->chrec;
2076 else
2077 return NULL_TREE;
2080 /* Sets in CACHE the value of VERSION instantiated below basic block
2081 INSTANTIATED_BELOW to VAL. */
2083 static void
2084 set_instantiated_value (htab_t cache, basic_block instantiated_below,
2085 tree version, tree val)
2087 struct scev_info_str *info, pattern;
2088 PTR *slot;
2090 pattern.var = version;
2091 pattern.instantiated_below = instantiated_below;
2092 slot = htab_find_slot (cache, &pattern, INSERT);
2094 if (!*slot)
2095 *slot = new_scev_info_str (instantiated_below, version);
2096 info = (struct scev_info_str *) *slot;
2097 info->chrec = val;
2100 /* Return the closed_loop_phi node for VAR. If there is none, return
2101 NULL_TREE. */
2103 static tree
2104 loop_closed_phi_def (tree var)
2106 struct loop *loop;
2107 edge exit;
2108 gimple phi;
2109 gimple_stmt_iterator psi;
2111 if (var == NULL_TREE
2112 || TREE_CODE (var) != SSA_NAME)
2113 return NULL_TREE;
2115 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var));
2116 exit = single_exit (loop);
2117 if (!exit)
2118 return NULL_TREE;
2120 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi))
2122 phi = gsi_stmt (psi);
2123 if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var)
2124 return PHI_RESULT (phi);
2127 return NULL_TREE;
2130 static tree instantiate_scev_r (basic_block, struct loop *, tree, bool,
2131 htab_t, int);
2133 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2134 and EVOLUTION_LOOP, that were left under a symbolic form.
2136 CHREC is an SSA_NAME to be instantiated.
2138 CACHE is the cache of already instantiated values.
2140 FOLD_CONVERSIONS should be set to true when the conversions that
2141 may wrap in signed/pointer type are folded, as long as the value of
2142 the chrec is preserved.
2144 SIZE_EXPR is used for computing the size of the expression to be
2145 instantiated, and to stop if it exceeds some limit. */
2147 static tree
2148 instantiate_scev_name (basic_block instantiate_below,
2149 struct loop *evolution_loop, tree chrec,
2150 bool fold_conversions, htab_t cache, int size_expr)
2152 tree res;
2153 struct loop *def_loop;
2154 basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (chrec));
2156 /* A parameter (or loop invariant and we do not want to include
2157 evolutions in outer loops), nothing to do. */
2158 if (!def_bb
2159 || loop_depth (def_bb->loop_father) == 0
2160 || dominated_by_p (CDI_DOMINATORS, instantiate_below, def_bb))
2161 return chrec;
2163 /* We cache the value of instantiated variable to avoid exponential
2164 time complexity due to reevaluations. We also store the convenient
2165 value in the cache in order to prevent infinite recursion -- we do
2166 not want to instantiate the SSA_NAME if it is in a mixer
2167 structure. This is used for avoiding the instantiation of
2168 recursively defined functions, such as:
2170 | a_2 -> {0, +, 1, +, a_2}_1 */
2172 res = get_instantiated_value (cache, instantiate_below, chrec);
2173 if (res)
2174 return res;
2176 res = chrec_dont_know;
2177 set_instantiated_value (cache, instantiate_below, chrec, res);
2179 def_loop = find_common_loop (evolution_loop, def_bb->loop_father);
2181 /* If the analysis yields a parametric chrec, instantiate the
2182 result again. */
2183 res = analyze_scalar_evolution (def_loop, chrec);
2185 /* Don't instantiate default definitions. */
2186 if (TREE_CODE (res) == SSA_NAME
2187 && SSA_NAME_IS_DEFAULT_DEF (res))
2190 /* Don't instantiate loop-closed-ssa phi nodes. */
2191 else if (TREE_CODE (res) == SSA_NAME
2192 && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res)))
2193 > loop_depth (def_loop))
2195 if (res == chrec)
2196 res = loop_closed_phi_def (chrec);
2197 else
2198 res = chrec;
2200 /* When there is no loop_closed_phi_def, it means that the
2201 variable is not used after the loop: try to still compute the
2202 value of the variable when exiting the loop. */
2203 if (res == NULL_TREE)
2205 loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (chrec));
2206 res = analyze_scalar_evolution (loop, chrec);
2207 res = compute_overall_effect_of_inner_loop (loop, res);
2208 res = instantiate_scev_r (instantiate_below, evolution_loop, res,
2209 fold_conversions, cache, size_expr);
2211 else if (!dominated_by_p (CDI_DOMINATORS, instantiate_below,
2212 gimple_bb (SSA_NAME_DEF_STMT (res))))
2213 res = chrec_dont_know;
2216 else if (res != chrec_dont_know)
2217 res = instantiate_scev_r (instantiate_below, evolution_loop, res,
2218 fold_conversions, cache, size_expr);
2220 /* Store the correct value to the cache. */
2221 set_instantiated_value (cache, instantiate_below, chrec, res);
2222 return res;
2225 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2226 and EVOLUTION_LOOP, that were left under a symbolic form.
2228 CHREC is a polynomial chain of recurrence to be instantiated.
2230 CACHE is the cache of already instantiated values.
2232 FOLD_CONVERSIONS should be set to true when the conversions that
2233 may wrap in signed/pointer type are folded, as long as the value of
2234 the chrec is preserved.
2236 SIZE_EXPR is used for computing the size of the expression to be
2237 instantiated, and to stop if it exceeds some limit. */
2239 static tree
2240 instantiate_scev_poly (basic_block instantiate_below,
2241 struct loop *evolution_loop, tree chrec,
2242 bool fold_conversions, htab_t cache, int size_expr)
2244 tree op1;
2245 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2246 CHREC_LEFT (chrec), fold_conversions, cache,
2247 size_expr);
2248 if (op0 == chrec_dont_know)
2249 return chrec_dont_know;
2251 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2252 CHREC_RIGHT (chrec), fold_conversions, cache,
2253 size_expr);
2254 if (op1 == chrec_dont_know)
2255 return chrec_dont_know;
2257 if (CHREC_LEFT (chrec) != op0
2258 || CHREC_RIGHT (chrec) != op1)
2260 unsigned var = CHREC_VARIABLE (chrec);
2262 /* When the instantiated stride or base has an evolution in an
2263 innermost loop, return chrec_dont_know, as this is not a
2264 valid SCEV representation. In the reduced testcase for
2265 PR40281 we would have {0, +, {1, +, 1}_2}_1 that has no
2266 meaning. */
2267 if ((tree_is_chrec (op0) && CHREC_VARIABLE (op0) > var)
2268 || (tree_is_chrec (op1) && CHREC_VARIABLE (op1) > var))
2269 return chrec_dont_know;
2271 op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL);
2272 chrec = build_polynomial_chrec (var, op0, op1);
2275 return chrec;
2278 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2279 and EVOLUTION_LOOP, that were left under a symbolic form.
2281 "C0 CODE C1" is a binary expression of type TYPE to be instantiated.
2283 CACHE is the cache of already instantiated values.
2285 FOLD_CONVERSIONS should be set to true when the conversions that
2286 may wrap in signed/pointer type are folded, as long as the value of
2287 the chrec is preserved.
2289 SIZE_EXPR is used for computing the size of the expression to be
2290 instantiated, and to stop if it exceeds some limit. */
2292 static tree
2293 instantiate_scev_binary (basic_block instantiate_below,
2294 struct loop *evolution_loop, tree chrec, enum tree_code code,
2295 tree type, tree c0, tree c1,
2296 bool fold_conversions, htab_t cache, int size_expr)
2298 tree op1;
2299 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2300 c0, fold_conversions, cache,
2301 size_expr);
2302 if (op0 == chrec_dont_know)
2303 return chrec_dont_know;
2305 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2306 c1, fold_conversions, cache,
2307 size_expr);
2308 if (op1 == chrec_dont_know)
2309 return chrec_dont_know;
2311 if (c0 != op0
2312 || c1 != op1)
2314 op0 = chrec_convert (type, op0, NULL);
2315 op1 = chrec_convert_rhs (type, op1, NULL);
2317 switch (code)
2319 case POINTER_PLUS_EXPR:
2320 case PLUS_EXPR:
2321 return chrec_fold_plus (type, op0, op1);
2323 case MINUS_EXPR:
2324 return chrec_fold_minus (type, op0, op1);
2326 case MULT_EXPR:
2327 return chrec_fold_multiply (type, op0, op1);
2329 default:
2330 gcc_unreachable ();
2334 return chrec ? chrec : fold_build2 (code, type, c0, c1);
2337 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2338 and EVOLUTION_LOOP, that were left under a symbolic form.
2340 "CHREC" is an array reference to be instantiated.
2342 CACHE is the cache of already instantiated values.
2344 FOLD_CONVERSIONS should be set to true when the conversions that
2345 may wrap in signed/pointer type are folded, as long as the value of
2346 the chrec is preserved.
2348 SIZE_EXPR is used for computing the size of the expression to be
2349 instantiated, and to stop if it exceeds some limit. */
2351 static tree
2352 instantiate_array_ref (basic_block instantiate_below,
2353 struct loop *evolution_loop, tree chrec,
2354 bool fold_conversions, htab_t cache, int size_expr)
2356 tree res;
2357 tree index = TREE_OPERAND (chrec, 1);
2358 tree op1 = instantiate_scev_r (instantiate_below, evolution_loop, index,
2359 fold_conversions, cache, size_expr);
2361 if (op1 == chrec_dont_know)
2362 return chrec_dont_know;
2364 if (chrec && op1 == index)
2365 return chrec;
2367 res = unshare_expr (chrec);
2368 TREE_OPERAND (res, 1) = op1;
2369 return res;
2372 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2373 and EVOLUTION_LOOP, that were left under a symbolic form.
2375 "CHREC" that stands for a convert expression "(TYPE) OP" is to be
2376 instantiated.
2378 CACHE is the cache of already instantiated values.
2380 FOLD_CONVERSIONS should be set to true when the conversions that
2381 may wrap in signed/pointer type are folded, as long as the value of
2382 the chrec is preserved.
2384 SIZE_EXPR is used for computing the size of the expression to be
2385 instantiated, and to stop if it exceeds some limit. */
2387 static tree
2388 instantiate_scev_convert (basic_block instantiate_below,
2389 struct loop *evolution_loop, tree chrec,
2390 tree type, tree op,
2391 bool fold_conversions, htab_t cache, int size_expr)
2393 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, op,
2394 fold_conversions, cache, size_expr);
2396 if (op0 == chrec_dont_know)
2397 return chrec_dont_know;
2399 if (fold_conversions)
2401 tree tmp = chrec_convert_aggressive (type, op0);
2402 if (tmp)
2403 return tmp;
2406 if (chrec && op0 == op)
2407 return chrec;
2409 /* If we used chrec_convert_aggressive, we can no longer assume that
2410 signed chrecs do not overflow, as chrec_convert does, so avoid
2411 calling it in that case. */
2412 if (fold_conversions)
2413 return fold_convert (type, op0);
2415 return chrec_convert (type, op0, NULL);
2418 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2419 and EVOLUTION_LOOP, that were left under a symbolic form.
2421 CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated.
2422 Handle ~X as -1 - X.
2423 Handle -X as -1 * X.
2425 CACHE is the cache of already instantiated values.
2427 FOLD_CONVERSIONS should be set to true when the conversions that
2428 may wrap in signed/pointer type are folded, as long as the value of
2429 the chrec is preserved.
2431 SIZE_EXPR is used for computing the size of the expression to be
2432 instantiated, and to stop if it exceeds some limit. */
2434 static tree
2435 instantiate_scev_not (basic_block instantiate_below,
2436 struct loop *evolution_loop, tree chrec,
2437 enum tree_code code, tree type, tree op,
2438 bool fold_conversions, htab_t cache, int size_expr)
2440 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, op,
2441 fold_conversions, cache, size_expr);
2443 if (op0 == chrec_dont_know)
2444 return chrec_dont_know;
2446 if (op != op0)
2448 op0 = chrec_convert (type, op0, NULL);
2450 switch (code)
2452 case BIT_NOT_EXPR:
2453 return chrec_fold_minus
2454 (type, fold_convert (type, integer_minus_one_node), op0);
2456 case NEGATE_EXPR:
2457 return chrec_fold_multiply
2458 (type, fold_convert (type, integer_minus_one_node), op0);
2460 default:
2461 gcc_unreachable ();
2465 return chrec ? chrec : fold_build1 (code, type, op0);
2468 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2469 and EVOLUTION_LOOP, that were left under a symbolic form.
2471 CHREC is an expression with 3 operands to be instantiated.
2473 CACHE is the cache of already instantiated values.
2475 FOLD_CONVERSIONS should be set to true when the conversions that
2476 may wrap in signed/pointer type are folded, as long as the value of
2477 the chrec is preserved.
2479 SIZE_EXPR is used for computing the size of the expression to be
2480 instantiated, and to stop if it exceeds some limit. */
2482 static tree
2483 instantiate_scev_3 (basic_block instantiate_below,
2484 struct loop *evolution_loop, tree chrec,
2485 bool fold_conversions, htab_t cache, int size_expr)
2487 tree op1, op2;
2488 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2489 TREE_OPERAND (chrec, 0),
2490 fold_conversions, cache, size_expr);
2491 if (op0 == chrec_dont_know)
2492 return chrec_dont_know;
2494 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2495 TREE_OPERAND (chrec, 1),
2496 fold_conversions, cache, size_expr);
2497 if (op1 == chrec_dont_know)
2498 return chrec_dont_know;
2500 op2 = instantiate_scev_r (instantiate_below, evolution_loop,
2501 TREE_OPERAND (chrec, 2),
2502 fold_conversions, cache, size_expr);
2503 if (op2 == chrec_dont_know)
2504 return chrec_dont_know;
2506 if (op0 == TREE_OPERAND (chrec, 0)
2507 && op1 == TREE_OPERAND (chrec, 1)
2508 && op2 == TREE_OPERAND (chrec, 2))
2509 return chrec;
2511 return fold_build3 (TREE_CODE (chrec),
2512 TREE_TYPE (chrec), op0, op1, op2);
2515 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2516 and EVOLUTION_LOOP, that were left under a symbolic form.
2518 CHREC is an expression with 2 operands to be instantiated.
2520 CACHE is the cache of already instantiated values.
2522 FOLD_CONVERSIONS should be set to true when the conversions that
2523 may wrap in signed/pointer type are folded, as long as the value of
2524 the chrec is preserved.
2526 SIZE_EXPR is used for computing the size of the expression to be
2527 instantiated, and to stop if it exceeds some limit. */
2529 static tree
2530 instantiate_scev_2 (basic_block instantiate_below,
2531 struct loop *evolution_loop, tree chrec,
2532 bool fold_conversions, htab_t cache, int size_expr)
2534 tree op1;
2535 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2536 TREE_OPERAND (chrec, 0),
2537 fold_conversions, cache, size_expr);
2538 if (op0 == chrec_dont_know)
2539 return chrec_dont_know;
2541 op1 = instantiate_scev_r (instantiate_below, evolution_loop,
2542 TREE_OPERAND (chrec, 1),
2543 fold_conversions, cache, size_expr);
2544 if (op1 == chrec_dont_know)
2545 return chrec_dont_know;
2547 if (op0 == TREE_OPERAND (chrec, 0)
2548 && op1 == TREE_OPERAND (chrec, 1))
2549 return chrec;
2551 return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1);
2554 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2555 and EVOLUTION_LOOP, that were left under a symbolic form.
2557 CHREC is an expression with 2 operands to be instantiated.
2559 CACHE is the cache of already instantiated values.
2561 FOLD_CONVERSIONS should be set to true when the conversions that
2562 may wrap in signed/pointer type are folded, as long as the value of
2563 the chrec is preserved.
2565 SIZE_EXPR is used for computing the size of the expression to be
2566 instantiated, and to stop if it exceeds some limit. */
2568 static tree
2569 instantiate_scev_1 (basic_block instantiate_below,
2570 struct loop *evolution_loop, tree chrec,
2571 bool fold_conversions, htab_t cache, int size_expr)
2573 tree op0 = instantiate_scev_r (instantiate_below, evolution_loop,
2574 TREE_OPERAND (chrec, 0),
2575 fold_conversions, cache, size_expr);
2577 if (op0 == chrec_dont_know)
2578 return chrec_dont_know;
2580 if (op0 == TREE_OPERAND (chrec, 0))
2581 return chrec;
2583 return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0);
2586 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2587 and EVOLUTION_LOOP, that were left under a symbolic form.
2589 CHREC is the scalar evolution to instantiate.
2591 CACHE is the cache of already instantiated values.
2593 FOLD_CONVERSIONS should be set to true when the conversions that
2594 may wrap in signed/pointer type are folded, as long as the value of
2595 the chrec is preserved.
2597 SIZE_EXPR is used for computing the size of the expression to be
2598 instantiated, and to stop if it exceeds some limit. */
2600 static tree
2601 instantiate_scev_r (basic_block instantiate_below,
2602 struct loop *evolution_loop, tree chrec,
2603 bool fold_conversions, htab_t cache, int size_expr)
2605 /* Give up if the expression is larger than the MAX that we allow. */
2606 if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
2607 return chrec_dont_know;
2609 if (automatically_generated_chrec_p (chrec)
2610 || is_gimple_min_invariant (chrec))
2611 return chrec;
2613 switch (TREE_CODE (chrec))
2615 case SSA_NAME:
2616 return instantiate_scev_name (instantiate_below, evolution_loop, chrec,
2617 fold_conversions, cache, size_expr);
2619 case POLYNOMIAL_CHREC:
2620 return instantiate_scev_poly (instantiate_below, evolution_loop, chrec,
2621 fold_conversions, cache, size_expr);
2623 case POINTER_PLUS_EXPR:
2624 case PLUS_EXPR:
2625 case MINUS_EXPR:
2626 case MULT_EXPR:
2627 return instantiate_scev_binary (instantiate_below, evolution_loop, chrec,
2628 TREE_CODE (chrec), chrec_type (chrec),
2629 TREE_OPERAND (chrec, 0),
2630 TREE_OPERAND (chrec, 1),
2631 fold_conversions, cache, size_expr);
2633 CASE_CONVERT:
2634 return instantiate_scev_convert (instantiate_below, evolution_loop, chrec,
2635 TREE_TYPE (chrec), TREE_OPERAND (chrec, 0),
2636 fold_conversions, cache, size_expr);
2638 case NEGATE_EXPR:
2639 case BIT_NOT_EXPR:
2640 return instantiate_scev_not (instantiate_below, evolution_loop, chrec,
2641 TREE_CODE (chrec), TREE_TYPE (chrec),
2642 TREE_OPERAND (chrec, 0),
2643 fold_conversions, cache, size_expr);
2645 case SCEV_NOT_KNOWN:
2646 return chrec_dont_know;
2648 case SCEV_KNOWN:
2649 return chrec_known;
2651 case ARRAY_REF:
2652 return instantiate_array_ref (instantiate_below, evolution_loop, chrec,
2653 fold_conversions, cache, size_expr);
2655 default:
2656 break;
2659 if (VL_EXP_CLASS_P (chrec))
2660 return chrec_dont_know;
2662 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
2664 case 3:
2665 return instantiate_scev_3 (instantiate_below, evolution_loop, chrec,
2666 fold_conversions, cache, size_expr);
2668 case 2:
2669 return instantiate_scev_2 (instantiate_below, evolution_loop, chrec,
2670 fold_conversions, cache, size_expr);
2672 case 1:
2673 return instantiate_scev_1 (instantiate_below, evolution_loop, chrec,
2674 fold_conversions, cache, size_expr);
2676 case 0:
2677 return chrec;
2679 default:
2680 break;
2683 /* Too complicated to handle. */
2684 return chrec_dont_know;
2687 /* Analyze all the parameters of the chrec that were left under a
2688 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2689 recursive instantiation of parameters: a parameter is a variable
2690 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2691 a function parameter. */
2693 tree
2694 instantiate_scev (basic_block instantiate_below, struct loop *evolution_loop,
2695 tree chrec)
2697 tree res;
2698 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2700 if (dump_file && (dump_flags & TDF_DETAILS))
2702 fprintf (dump_file, "(instantiate_scev \n");
2703 fprintf (dump_file, " (instantiate_below = %d)\n", instantiate_below->index);
2704 fprintf (dump_file, " (evolution_loop = %d)\n", evolution_loop->num);
2705 fprintf (dump_file, " (chrec = ");
2706 print_generic_expr (dump_file, chrec, 0);
2707 fprintf (dump_file, ")\n");
2710 res = instantiate_scev_r (instantiate_below, evolution_loop, chrec, false,
2711 cache, 0);
2713 if (dump_file && (dump_flags & TDF_DETAILS))
2715 fprintf (dump_file, " (res = ");
2716 print_generic_expr (dump_file, res, 0);
2717 fprintf (dump_file, "))\n");
2720 htab_delete (cache);
2722 return res;
2725 /* Similar to instantiate_parameters, but does not introduce the
2726 evolutions in outer loops for LOOP invariants in CHREC, and does not
2727 care about causing overflows, as long as they do not affect value
2728 of an expression. */
2730 tree
2731 resolve_mixers (struct loop *loop, tree chrec)
2733 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2734 tree ret = instantiate_scev_r (block_before_loop (loop), loop, chrec, true,
2735 cache, 0);
2736 htab_delete (cache);
2737 return ret;
2740 /* Entry point for the analysis of the number of iterations pass.
2741 This function tries to safely approximate the number of iterations
2742 the loop will run. When this property is not decidable at compile
2743 time, the result is chrec_dont_know. Otherwise the result is a
2744 scalar or a symbolic parameter. When the number of iterations may
2745 be equal to zero and the property cannot be determined at compile
2746 time, the result is a COND_EXPR that represents in a symbolic form
2747 the conditions under which the number of iterations is not zero.
2749 Example of analysis: suppose that the loop has an exit condition:
2751 "if (b > 49) goto end_loop;"
2753 and that in a previous analysis we have determined that the
2754 variable 'b' has an evolution function:
2756 "EF = {23, +, 5}_2".
2758 When we evaluate the function at the point 5, i.e. the value of the
2759 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2760 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2761 the loop body has been executed 6 times. */
2763 tree
2764 number_of_latch_executions (struct loop *loop)
2766 edge exit;
2767 struct tree_niter_desc niter_desc;
2768 tree may_be_zero;
2769 tree res;
2771 /* Determine whether the number of iterations in loop has already
2772 been computed. */
2773 res = loop->nb_iterations;
2774 if (res)
2775 return res;
2777 may_be_zero = NULL_TREE;
2779 if (dump_file && (dump_flags & TDF_DETAILS))
2780 fprintf (dump_file, "(number_of_iterations_in_loop = \n");
2782 res = chrec_dont_know;
2783 exit = single_exit (loop);
2785 if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false))
2787 may_be_zero = niter_desc.may_be_zero;
2788 res = niter_desc.niter;
2791 if (res == chrec_dont_know
2792 || !may_be_zero
2793 || integer_zerop (may_be_zero))
2795 else if (integer_nonzerop (may_be_zero))
2796 res = build_int_cst (TREE_TYPE (res), 0);
2798 else if (COMPARISON_CLASS_P (may_be_zero))
2799 res = fold_build3 (COND_EXPR, TREE_TYPE (res), may_be_zero,
2800 build_int_cst (TREE_TYPE (res), 0), res);
2801 else
2802 res = chrec_dont_know;
2804 if (dump_file && (dump_flags & TDF_DETAILS))
2806 fprintf (dump_file, " (set_nb_iterations_in_loop = ");
2807 print_generic_expr (dump_file, res, 0);
2808 fprintf (dump_file, "))\n");
2811 loop->nb_iterations = res;
2812 return res;
2815 /* Returns the number of executions of the exit condition of LOOP,
2816 i.e., the number by one higher than number_of_latch_executions.
2817 Note that unlike number_of_latch_executions, this number does
2818 not necessarily fit in the unsigned variant of the type of
2819 the control variable -- if the number of iterations is a constant,
2820 we return chrec_dont_know if adding one to number_of_latch_executions
2821 overflows; however, in case the number of iterations is symbolic
2822 expression, the caller is responsible for dealing with this
2823 the possible overflow. */
2825 tree
2826 number_of_exit_cond_executions (struct loop *loop)
2828 tree ret = number_of_latch_executions (loop);
2829 tree type = chrec_type (ret);
2831 if (chrec_contains_undetermined (ret))
2832 return ret;
2834 ret = chrec_fold_plus (type, ret, build_int_cst (type, 1));
2835 if (TREE_CODE (ret) == INTEGER_CST
2836 && TREE_OVERFLOW (ret))
2837 return chrec_dont_know;
2839 return ret;
2842 /* One of the drivers for testing the scalar evolutions analysis.
2843 This function computes the number of iterations for all the loops
2844 from the EXIT_CONDITIONS array. */
2846 static void
2847 number_of_iterations_for_all_loops (VEC(gimple,heap) **exit_conditions)
2849 unsigned int i;
2850 unsigned nb_chrec_dont_know_loops = 0;
2851 unsigned nb_static_loops = 0;
2852 gimple cond;
2854 FOR_EACH_VEC_ELT (gimple, *exit_conditions, i, cond)
2856 tree res = number_of_latch_executions (loop_containing_stmt (cond));
2857 if (chrec_contains_undetermined (res))
2858 nb_chrec_dont_know_loops++;
2859 else
2860 nb_static_loops++;
2863 if (dump_file)
2865 fprintf (dump_file, "\n(\n");
2866 fprintf (dump_file, "-----------------------------------------\n");
2867 fprintf (dump_file, "%d\tnb_chrec_dont_know_loops\n", nb_chrec_dont_know_loops);
2868 fprintf (dump_file, "%d\tnb_static_loops\n", nb_static_loops);
2869 fprintf (dump_file, "%d\tnb_total_loops\n", number_of_loops ());
2870 fprintf (dump_file, "-----------------------------------------\n");
2871 fprintf (dump_file, ")\n\n");
2873 print_loops (dump_file, 3);
2879 /* Counters for the stats. */
2881 struct chrec_stats
2883 unsigned nb_chrecs;
2884 unsigned nb_affine;
2885 unsigned nb_affine_multivar;
2886 unsigned nb_higher_poly;
2887 unsigned nb_chrec_dont_know;
2888 unsigned nb_undetermined;
2891 /* Reset the counters. */
2893 static inline void
2894 reset_chrecs_counters (struct chrec_stats *stats)
2896 stats->nb_chrecs = 0;
2897 stats->nb_affine = 0;
2898 stats->nb_affine_multivar = 0;
2899 stats->nb_higher_poly = 0;
2900 stats->nb_chrec_dont_know = 0;
2901 stats->nb_undetermined = 0;
2904 /* Dump the contents of a CHREC_STATS structure. */
2906 static void
2907 dump_chrecs_stats (FILE *file, struct chrec_stats *stats)
2909 fprintf (file, "\n(\n");
2910 fprintf (file, "-----------------------------------------\n");
2911 fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine);
2912 fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar);
2913 fprintf (file, "%d\tdegree greater than 2 polynomials\n",
2914 stats->nb_higher_poly);
2915 fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know);
2916 fprintf (file, "-----------------------------------------\n");
2917 fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs);
2918 fprintf (file, "%d\twith undetermined coefficients\n",
2919 stats->nb_undetermined);
2920 fprintf (file, "-----------------------------------------\n");
2921 fprintf (file, "%d\tchrecs in the scev database\n",
2922 (int) htab_elements (scalar_evolution_info));
2923 fprintf (file, "%d\tsets in the scev database\n", nb_set_scev);
2924 fprintf (file, "%d\tgets in the scev database\n", nb_get_scev);
2925 fprintf (file, "-----------------------------------------\n");
2926 fprintf (file, ")\n\n");
2929 /* Gather statistics about CHREC. */
2931 static void
2932 gather_chrec_stats (tree chrec, struct chrec_stats *stats)
2934 if (dump_file && (dump_flags & TDF_STATS))
2936 fprintf (dump_file, "(classify_chrec ");
2937 print_generic_expr (dump_file, chrec, 0);
2938 fprintf (dump_file, "\n");
2941 stats->nb_chrecs++;
2943 if (chrec == NULL_TREE)
2945 stats->nb_undetermined++;
2946 return;
2949 switch (TREE_CODE (chrec))
2951 case POLYNOMIAL_CHREC:
2952 if (evolution_function_is_affine_p (chrec))
2954 if (dump_file && (dump_flags & TDF_STATS))
2955 fprintf (dump_file, " affine_univariate\n");
2956 stats->nb_affine++;
2958 else if (evolution_function_is_affine_multivariate_p (chrec, 0))
2960 if (dump_file && (dump_flags & TDF_STATS))
2961 fprintf (dump_file, " affine_multivariate\n");
2962 stats->nb_affine_multivar++;
2964 else
2966 if (dump_file && (dump_flags & TDF_STATS))
2967 fprintf (dump_file, " higher_degree_polynomial\n");
2968 stats->nb_higher_poly++;
2971 break;
2973 default:
2974 break;
2977 if (chrec_contains_undetermined (chrec))
2979 if (dump_file && (dump_flags & TDF_STATS))
2980 fprintf (dump_file, " undetermined\n");
2981 stats->nb_undetermined++;
2984 if (dump_file && (dump_flags & TDF_STATS))
2985 fprintf (dump_file, ")\n");
2988 /* One of the drivers for testing the scalar evolutions analysis.
2989 This function analyzes the scalar evolution of all the scalars
2990 defined as loop phi nodes in one of the loops from the
2991 EXIT_CONDITIONS array.
2993 TODO Optimization: A loop is in canonical form if it contains only
2994 a single scalar loop phi node. All the other scalars that have an
2995 evolution in the loop are rewritten in function of this single
2996 index. This allows the parallelization of the loop. */
2998 static void
2999 analyze_scalar_evolution_for_all_loop_phi_nodes (VEC(gimple,heap) **exit_conditions)
3001 unsigned int i;
3002 struct chrec_stats stats;
3003 gimple cond, phi;
3004 gimple_stmt_iterator psi;
3006 reset_chrecs_counters (&stats);
3008 FOR_EACH_VEC_ELT (gimple, *exit_conditions, i, cond)
3010 struct loop *loop;
3011 basic_block bb;
3012 tree chrec;
3014 loop = loop_containing_stmt (cond);
3015 bb = loop->header;
3017 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
3019 phi = gsi_stmt (psi);
3020 if (is_gimple_reg (PHI_RESULT (phi)))
3022 chrec = instantiate_parameters
3023 (loop,
3024 analyze_scalar_evolution (loop, PHI_RESULT (phi)));
3026 if (dump_file && (dump_flags & TDF_STATS))
3027 gather_chrec_stats (chrec, &stats);
3032 if (dump_file && (dump_flags & TDF_STATS))
3033 dump_chrecs_stats (dump_file, &stats);
3036 /* Callback for htab_traverse, gathers information on chrecs in the
3037 hashtable. */
3039 static int
3040 gather_stats_on_scev_database_1 (void **slot, void *stats)
3042 struct scev_info_str *entry = (struct scev_info_str *) *slot;
3044 gather_chrec_stats (entry->chrec, (struct chrec_stats *) stats);
3046 return 1;
3049 /* Classify the chrecs of the whole database. */
3051 void
3052 gather_stats_on_scev_database (void)
3054 struct chrec_stats stats;
3056 if (!dump_file)
3057 return;
3059 reset_chrecs_counters (&stats);
3061 htab_traverse (scalar_evolution_info, gather_stats_on_scev_database_1,
3062 &stats);
3064 dump_chrecs_stats (dump_file, &stats);
3069 /* Initializer. */
3071 static void
3072 initialize_scalar_evolutions_analyzer (void)
3074 /* The elements below are unique. */
3075 if (chrec_dont_know == NULL_TREE)
3077 chrec_not_analyzed_yet = NULL_TREE;
3078 chrec_dont_know = make_node (SCEV_NOT_KNOWN);
3079 chrec_known = make_node (SCEV_KNOWN);
3080 TREE_TYPE (chrec_dont_know) = void_type_node;
3081 TREE_TYPE (chrec_known) = void_type_node;
3085 /* Initialize the analysis of scalar evolutions for LOOPS. */
3087 void
3088 scev_initialize (void)
3090 loop_iterator li;
3091 struct loop *loop;
3094 scalar_evolution_info = htab_create_ggc (100, hash_scev_info, eq_scev_info,
3095 del_scev_info);
3097 initialize_scalar_evolutions_analyzer ();
3099 FOR_EACH_LOOP (li, loop, 0)
3101 loop->nb_iterations = NULL_TREE;
3105 /* Cleans up the information cached by the scalar evolutions analysis
3106 in the hash table. */
3108 void
3109 scev_reset_htab (void)
3111 if (!scalar_evolution_info)
3112 return;
3114 htab_empty (scalar_evolution_info);
3117 /* Cleans up the information cached by the scalar evolutions analysis
3118 in the hash table and in the loop->nb_iterations. */
3120 void
3121 scev_reset (void)
3123 loop_iterator li;
3124 struct loop *loop;
3126 scev_reset_htab ();
3128 if (!current_loops)
3129 return;
3131 FOR_EACH_LOOP (li, loop, 0)
3133 loop->nb_iterations = NULL_TREE;
3137 /* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with
3138 respect to WRTO_LOOP and returns its base and step in IV if possible
3139 (see analyze_scalar_evolution_in_loop for more details on USE_LOOP
3140 and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be
3141 invariant in LOOP. Otherwise we require it to be an integer constant.
3143 IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g.
3144 because it is computed in signed arithmetics). Consequently, adding an
3145 induction variable
3147 for (i = IV->base; ; i += IV->step)
3149 is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is
3150 false for the type of the induction variable, or you can prove that i does
3151 not wrap by some other argument. Otherwise, this might introduce undefined
3152 behavior, and
3154 for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step))
3156 must be used instead. */
3158 bool
3159 simple_iv (struct loop *wrto_loop, struct loop *use_loop, tree op,
3160 affine_iv *iv, bool allow_nonconstant_step)
3162 tree type, ev;
3163 bool folded_casts;
3165 iv->base = NULL_TREE;
3166 iv->step = NULL_TREE;
3167 iv->no_overflow = false;
3169 type = TREE_TYPE (op);
3170 if (TREE_CODE (type) != INTEGER_TYPE
3171 && TREE_CODE (type) != POINTER_TYPE)
3172 return false;
3174 ev = analyze_scalar_evolution_in_loop (wrto_loop, use_loop, op,
3175 &folded_casts);
3176 if (chrec_contains_undetermined (ev)
3177 || chrec_contains_symbols_defined_in_loop (ev, wrto_loop->num))
3178 return false;
3180 if (tree_does_not_contain_chrecs (ev))
3182 iv->base = ev;
3183 iv->step = build_int_cst (TREE_TYPE (ev), 0);
3184 iv->no_overflow = true;
3185 return true;
3188 if (TREE_CODE (ev) != POLYNOMIAL_CHREC
3189 || CHREC_VARIABLE (ev) != (unsigned) wrto_loop->num)
3190 return false;
3192 iv->step = CHREC_RIGHT (ev);
3193 if ((!allow_nonconstant_step && TREE_CODE (iv->step) != INTEGER_CST)
3194 || tree_contains_chrecs (iv->step, NULL))
3195 return false;
3197 iv->base = CHREC_LEFT (ev);
3198 if (tree_contains_chrecs (iv->base, NULL))
3199 return false;
3201 iv->no_overflow = !folded_casts && TYPE_OVERFLOW_UNDEFINED (type);
3203 return true;
3206 /* Runs the analysis of scalar evolutions. */
3208 void
3209 scev_analysis (void)
3211 VEC(gimple,heap) *exit_conditions;
3213 exit_conditions = VEC_alloc (gimple, heap, 37);
3214 select_loops_exit_conditions (&exit_conditions);
3216 if (dump_file && (dump_flags & TDF_STATS))
3217 analyze_scalar_evolution_for_all_loop_phi_nodes (&exit_conditions);
3219 number_of_iterations_for_all_loops (&exit_conditions);
3220 VEC_free (gimple, heap, exit_conditions);
3223 /* Finalize the scalar evolution analysis. */
3225 void
3226 scev_finalize (void)
3228 if (!scalar_evolution_info)
3229 return;
3230 htab_delete (scalar_evolution_info);
3231 scalar_evolution_info = NULL;
3234 /* Returns true if the expression EXPR is considered to be too expensive
3235 for scev_const_prop. */
3237 bool
3238 expression_expensive_p (tree expr)
3240 enum tree_code code;
3242 if (is_gimple_val (expr))
3243 return false;
3245 code = TREE_CODE (expr);
3246 if (code == TRUNC_DIV_EXPR
3247 || code == CEIL_DIV_EXPR
3248 || code == FLOOR_DIV_EXPR
3249 || code == ROUND_DIV_EXPR
3250 || code == TRUNC_MOD_EXPR
3251 || code == CEIL_MOD_EXPR
3252 || code == FLOOR_MOD_EXPR
3253 || code == ROUND_MOD_EXPR
3254 || code == EXACT_DIV_EXPR)
3256 /* Division by power of two is usually cheap, so we allow it.
3257 Forbid anything else. */
3258 if (!integer_pow2p (TREE_OPERAND (expr, 1)))
3259 return true;
3262 switch (TREE_CODE_CLASS (code))
3264 case tcc_binary:
3265 case tcc_comparison:
3266 if (expression_expensive_p (TREE_OPERAND (expr, 1)))
3267 return true;
3269 /* Fallthru. */
3270 case tcc_unary:
3271 return expression_expensive_p (TREE_OPERAND (expr, 0));
3273 default:
3274 return true;
3278 /* Replace ssa names for that scev can prove they are constant by the
3279 appropriate constants. Also perform final value replacement in loops,
3280 in case the replacement expressions are cheap.
3282 We only consider SSA names defined by phi nodes; rest is left to the
3283 ordinary constant propagation pass. */
3285 unsigned int
3286 scev_const_prop (void)
3288 basic_block bb;
3289 tree name, type, ev;
3290 gimple phi, ass;
3291 struct loop *loop, *ex_loop;
3292 bitmap ssa_names_to_remove = NULL;
3293 unsigned i;
3294 loop_iterator li;
3295 gimple_stmt_iterator psi;
3297 if (number_of_loops () <= 1)
3298 return 0;
3300 FOR_EACH_BB (bb)
3302 loop = bb->loop_father;
3304 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
3306 phi = gsi_stmt (psi);
3307 name = PHI_RESULT (phi);
3309 if (!is_gimple_reg (name))
3310 continue;
3312 type = TREE_TYPE (name);
3314 if (!POINTER_TYPE_P (type)
3315 && !INTEGRAL_TYPE_P (type))
3316 continue;
3318 ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name));
3319 if (!is_gimple_min_invariant (ev)
3320 || !may_propagate_copy (name, ev))
3321 continue;
3323 /* Replace the uses of the name. */
3324 if (name != ev)
3325 replace_uses_by (name, ev);
3327 if (!ssa_names_to_remove)
3328 ssa_names_to_remove = BITMAP_ALLOC (NULL);
3329 bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name));
3333 /* Remove the ssa names that were replaced by constants. We do not
3334 remove them directly in the previous cycle, since this
3335 invalidates scev cache. */
3336 if (ssa_names_to_remove)
3338 bitmap_iterator bi;
3340 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi)
3342 gimple_stmt_iterator psi;
3343 name = ssa_name (i);
3344 phi = SSA_NAME_DEF_STMT (name);
3346 gcc_assert (gimple_code (phi) == GIMPLE_PHI);
3347 psi = gsi_for_stmt (phi);
3348 remove_phi_node (&psi, true);
3351 BITMAP_FREE (ssa_names_to_remove);
3352 scev_reset ();
3355 /* Now the regular final value replacement. */
3356 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
3358 edge exit;
3359 tree def, rslt, niter;
3360 gimple_stmt_iterator bsi;
3362 /* If we do not know exact number of iterations of the loop, we cannot
3363 replace the final value. */
3364 exit = single_exit (loop);
3365 if (!exit)
3366 continue;
3368 niter = number_of_latch_executions (loop);
3369 if (niter == chrec_dont_know)
3370 continue;
3372 /* Ensure that it is possible to insert new statements somewhere. */
3373 if (!single_pred_p (exit->dest))
3374 split_loop_exit_edge (exit);
3375 bsi = gsi_after_labels (exit->dest);
3377 ex_loop = superloop_at_depth (loop,
3378 loop_depth (exit->dest->loop_father) + 1);
3380 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); )
3382 phi = gsi_stmt (psi);
3383 rslt = PHI_RESULT (phi);
3384 def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
3385 if (!is_gimple_reg (def))
3387 gsi_next (&psi);
3388 continue;
3391 if (!POINTER_TYPE_P (TREE_TYPE (def))
3392 && !INTEGRAL_TYPE_P (TREE_TYPE (def)))
3394 gsi_next (&psi);
3395 continue;
3398 def = analyze_scalar_evolution_in_loop (ex_loop, loop, def, NULL);
3399 def = compute_overall_effect_of_inner_loop (ex_loop, def);
3400 if (!tree_does_not_contain_chrecs (def)
3401 || chrec_contains_symbols_defined_in_loop (def, ex_loop->num)
3402 /* Moving the computation from the loop may prolong life range
3403 of some ssa names, which may cause problems if they appear
3404 on abnormal edges. */
3405 || contains_abnormal_ssa_name_p (def)
3406 /* Do not emit expensive expressions. The rationale is that
3407 when someone writes a code like
3409 while (n > 45) n -= 45;
3411 he probably knows that n is not large, and does not want it
3412 to be turned into n %= 45. */
3413 || expression_expensive_p (def))
3415 gsi_next (&psi);
3416 continue;
3419 /* Eliminate the PHI node and replace it by a computation outside
3420 the loop. */
3421 def = unshare_expr (def);
3422 remove_phi_node (&psi, false);
3424 def = force_gimple_operand_gsi (&bsi, def, false, NULL_TREE,
3425 true, GSI_SAME_STMT);
3426 ass = gimple_build_assign (rslt, def);
3427 gsi_insert_before (&bsi, ass, GSI_SAME_STMT);
3430 return 0;
3433 #include "gt-tree-scalar-evolution.h"