* var-tracking.c (insn_stack_adjust_offset_pre_post): If insn has a
[official-gcc.git] / gcc / tree-scalar-evolution.c
blob52e2c75f0fe210cf13ae0eca2211164f200b0033
1 /* Scalar evolution detector.
2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008 Free Software
3 Foundation, Inc.
4 Contributed by Sebastian Pop <s.pop@laposte.net>
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
22 /*
23 Description:
25 This pass analyzes the evolution of scalar variables in loop
26 structures. The algorithm is based on the SSA representation,
27 and on the loop hierarchy tree. This algorithm is not based on
28 the notion of versions of a variable, as it was the case for the
29 previous implementations of the scalar evolution algorithm, but
30 it assumes that each defined name is unique.
32 The notation used in this file is called "chains of recurrences",
33 and has been proposed by Eugene Zima, Robert Van Engelen, and
34 others for describing induction variables in programs. For example
35 "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0
36 when entering in the loop_1 and has a step 2 in this loop, in other
37 words "for (b = 0; b < N; b+=2);". Note that the coefficients of
38 this chain of recurrence (or chrec [shrek]) can contain the name of
39 other variables, in which case they are called parametric chrecs.
40 For example, "b -> {a, +, 2}_1" means that the initial value of "b"
41 is the value of "a". In most of the cases these parametric chrecs
42 are fully instantiated before their use because symbolic names can
43 hide some difficult cases such as self-references described later
44 (see the Fibonacci example).
46 A short sketch of the algorithm is:
48 Given a scalar variable to be analyzed, follow the SSA edge to
49 its definition:
51 - When the definition is a GIMPLE_ASSIGN: if the right hand side
52 (RHS) of the definition cannot be statically analyzed, the answer
53 of the analyzer is: "don't know".
54 Otherwise, for all the variables that are not yet analyzed in the
55 RHS, try to determine their evolution, and finally try to
56 evaluate the operation of the RHS that gives the evolution
57 function of the analyzed variable.
59 - When the definition is a condition-phi-node: determine the
60 evolution function for all the branches of the phi node, and
61 finally merge these evolutions (see chrec_merge).
63 - When the definition is a loop-phi-node: determine its initial
64 condition, that is the SSA edge defined in an outer loop, and
65 keep it symbolic. Then determine the SSA edges that are defined
66 in the body of the loop. Follow the inner edges until ending on
67 another loop-phi-node of the same analyzed loop. If the reached
68 loop-phi-node is not the starting loop-phi-node, then we keep
69 this definition under a symbolic form. If the reached
70 loop-phi-node is the same as the starting one, then we compute a
71 symbolic stride on the return path. The result is then the
72 symbolic chrec {initial_condition, +, symbolic_stride}_loop.
74 Examples:
76 Example 1: Illustration of the basic algorithm.
78 | a = 3
79 | loop_1
80 | b = phi (a, c)
81 | c = b + 1
82 | if (c > 10) exit_loop
83 | endloop
85 Suppose that we want to know the number of iterations of the
86 loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We
87 ask the scalar evolution analyzer two questions: what's the
88 scalar evolution (scev) of "c", and what's the scev of "10". For
89 "10" the answer is "10" since it is a scalar constant. For the
90 scalar variable "c", it follows the SSA edge to its definition,
91 "c = b + 1", and then asks again what's the scev of "b".
92 Following the SSA edge, we end on a loop-phi-node "b = phi (a,
93 c)", where the initial condition is "a", and the inner loop edge
94 is "c". The initial condition is kept under a symbolic form (it
95 may be the case that the copy constant propagation has done its
96 work and we end with the constant "3" as one of the edges of the
97 loop-phi-node). The update edge is followed to the end of the
98 loop, and until reaching again the starting loop-phi-node: b -> c
99 -> b. At this point we have drawn a path from "b" to "b" from
100 which we compute the stride in the loop: in this example it is
101 "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now
102 that the scev for "b" is known, it is possible to compute the
103 scev for "c", that is "c -> {a + 1, +, 1}_1". In order to
104 determine the number of iterations in the loop_1, we have to
105 instantiate_parameters (loop_1, {a + 1, +, 1}_1), that gives after some
106 more analysis the scev {4, +, 1}_1, or in other words, this is
107 the function "f (x) = x + 4", where x is the iteration count of
108 the loop_1. Now we have to solve the inequality "x + 4 > 10",
109 and take the smallest iteration number for which the loop is
110 exited: x = 7. This loop runs from x = 0 to x = 7, and in total
111 there are 8 iterations. In terms of loop normalization, we have
112 created a variable that is implicitly defined, "x" or just "_1",
113 and all the other analyzed scalars of the loop are defined in
114 function of this variable:
116 a -> 3
117 b -> {3, +, 1}_1
118 c -> {4, +, 1}_1
120 or in terms of a C program:
122 | a = 3
123 | for (x = 0; x <= 7; x++)
125 | b = x + 3
126 | c = x + 4
129 Example 2a: Illustration of the algorithm on nested loops.
131 | loop_1
132 | a = phi (1, b)
133 | c = a + 2
134 | loop_2 10 times
135 | b = phi (c, d)
136 | d = b + 3
137 | endloop
138 | endloop
140 For analyzing the scalar evolution of "a", the algorithm follows
141 the SSA edge into the loop's body: "a -> b". "b" is an inner
142 loop-phi-node, and its analysis as in Example 1, gives:
144 b -> {c, +, 3}_2
145 d -> {c + 3, +, 3}_2
147 Following the SSA edge for the initial condition, we end on "c = a
148 + 2", and then on the starting loop-phi-node "a". From this point,
149 the loop stride is computed: back on "c = a + 2" we get a "+2" in
150 the loop_1, then on the loop-phi-node "b" we compute the overall
151 effect of the inner loop that is "b = c + 30", and we get a "+30"
152 in the loop_1. That means that the overall stride in loop_1 is
153 equal to "+32", and the result is:
155 a -> {1, +, 32}_1
156 c -> {3, +, 32}_1
158 Example 2b: Multivariate chains of recurrences.
160 | loop_1
161 | k = phi (0, k + 1)
162 | loop_2 4 times
163 | j = phi (0, j + 1)
164 | loop_3 4 times
165 | i = phi (0, i + 1)
166 | A[j + k] = ...
167 | endloop
168 | endloop
169 | endloop
171 Analyzing the access function of array A with
172 instantiate_parameters (loop_1, "j + k"), we obtain the
173 instantiation and the analysis of the scalar variables "j" and "k"
174 in loop_1. This leads to the scalar evolution {4, +, 1}_1: the end
175 value of loop_2 for "j" is 4, and the evolution of "k" in loop_1 is
176 {0, +, 1}_1. To obtain the evolution function in loop_3 and
177 instantiate the scalar variables up to loop_1, one has to use:
178 instantiate_scev (block_before_loop (loop_1), loop_3, "j + k").
179 The result of this call is {{0, +, 1}_1, +, 1}_2.
181 Example 3: Higher degree polynomials.
183 | loop_1
184 | a = phi (2, b)
185 | c = phi (5, d)
186 | b = a + 1
187 | d = c + a
188 | endloop
190 a -> {2, +, 1}_1
191 b -> {3, +, 1}_1
192 c -> {5, +, a}_1
193 d -> {5 + a, +, a}_1
195 instantiate_parameters (loop_1, {5, +, a}_1) -> {5, +, 2, +, 1}_1
196 instantiate_parameters (loop_1, {5 + a, +, a}_1) -> {7, +, 3, +, 1}_1
198 Example 4: Lucas, Fibonacci, or mixers in general.
200 | loop_1
201 | a = phi (1, b)
202 | c = phi (3, d)
203 | b = c
204 | d = c + a
205 | endloop
207 a -> (1, c)_1
208 c -> {3, +, a}_1
210 The syntax "(1, c)_1" stands for a PEELED_CHREC that has the
211 following semantics: during the first iteration of the loop_1, the
212 variable contains the value 1, and then it contains the value "c".
213 Note that this syntax is close to the syntax of the loop-phi-node:
214 "a -> (1, c)_1" vs. "a = phi (1, c)".
216 The symbolic chrec representation contains all the semantics of the
217 original code. What is more difficult is to use this information.
219 Example 5: Flip-flops, or exchangers.
221 | loop_1
222 | a = phi (1, b)
223 | c = phi (3, d)
224 | b = c
225 | d = a
226 | endloop
228 a -> (1, c)_1
229 c -> (3, a)_1
231 Based on these symbolic chrecs, it is possible to refine this
232 information into the more precise PERIODIC_CHRECs:
234 a -> |1, 3|_1
235 c -> |3, 1|_1
237 This transformation is not yet implemented.
239 Further readings:
241 You can find a more detailed description of the algorithm in:
242 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf
243 http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that
244 this is a preliminary report and some of the details of the
245 algorithm have changed. I'm working on a research report that
246 updates the description of the algorithms to reflect the design
247 choices used in this implementation.
249 A set of slides show a high level overview of the algorithm and run
250 an example through the scalar evolution analyzer:
251 http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf
253 The slides that I have presented at the GCC Summit'04 are available
254 at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf
257 #include "config.h"
258 #include "system.h"
259 #include "coretypes.h"
260 #include "tm.h"
261 #include "ggc.h"
262 #include "tree.h"
263 #include "real.h"
265 /* These RTL headers are needed for basic-block.h. */
266 #include "rtl.h"
267 #include "basic-block.h"
268 #include "diagnostic.h"
269 #include "tree-flow.h"
270 #include "tree-dump.h"
271 #include "timevar.h"
272 #include "cfgloop.h"
273 #include "tree-chrec.h"
274 #include "tree-scalar-evolution.h"
275 #include "tree-pass.h"
276 #include "flags.h"
277 #include "params.h"
279 static tree analyze_scalar_evolution_1 (struct loop *, tree, tree);
281 /* The cached information about an SSA name VAR, claiming that below
282 basic block INSTANTIATED_BELOW, the value of VAR can be expressed
283 as CHREC. */
285 struct scev_info_str GTY(())
287 basic_block instantiated_below;
288 tree var;
289 tree chrec;
292 /* Counters for the scev database. */
293 static unsigned nb_set_scev = 0;
294 static unsigned nb_get_scev = 0;
296 /* The following trees are unique elements. Thus the comparison of
297 another element to these elements should be done on the pointer to
298 these trees, and not on their value. */
300 /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */
301 tree chrec_not_analyzed_yet;
303 /* Reserved to the cases where the analyzer has detected an
304 undecidable property at compile time. */
305 tree chrec_dont_know;
307 /* When the analyzer has detected that a property will never
308 happen, then it qualifies it with chrec_known. */
309 tree chrec_known;
311 static GTY ((param_is (struct scev_info_str))) htab_t scalar_evolution_info;
314 /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */
316 static inline struct scev_info_str *
317 new_scev_info_str (basic_block instantiated_below, tree var)
319 struct scev_info_str *res;
321 res = GGC_NEW (struct scev_info_str);
322 res->var = var;
323 res->chrec = chrec_not_analyzed_yet;
324 res->instantiated_below = instantiated_below;
326 return res;
329 /* Computes a hash function for database element ELT. */
331 static hashval_t
332 hash_scev_info (const void *elt)
334 return SSA_NAME_VERSION (((const struct scev_info_str *) elt)->var);
337 /* Compares database elements E1 and E2. */
339 static int
340 eq_scev_info (const void *e1, const void *e2)
342 const struct scev_info_str *elt1 = (const struct scev_info_str *) e1;
343 const struct scev_info_str *elt2 = (const struct scev_info_str *) e2;
345 return (elt1->var == elt2->var
346 && elt1->instantiated_below == elt2->instantiated_below);
349 /* Deletes database element E. */
351 static void
352 del_scev_info (void *e)
354 ggc_free (e);
357 /* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block.
358 A first query on VAR returns chrec_not_analyzed_yet. */
360 static tree *
361 find_var_scev_info (basic_block instantiated_below, tree var)
363 struct scev_info_str *res;
364 struct scev_info_str tmp;
365 PTR *slot;
367 tmp.var = var;
368 tmp.instantiated_below = instantiated_below;
369 slot = htab_find_slot (scalar_evolution_info, &tmp, INSERT);
371 if (!*slot)
372 *slot = new_scev_info_str (instantiated_below, var);
373 res = (struct scev_info_str *) *slot;
375 return &res->chrec;
378 /* Return true when CHREC contains symbolic names defined in
379 LOOP_NB. */
381 bool
382 chrec_contains_symbols_defined_in_loop (const_tree chrec, unsigned loop_nb)
384 int i, n;
386 if (chrec == NULL_TREE)
387 return false;
389 if (is_gimple_min_invariant (chrec))
390 return false;
392 if (TREE_CODE (chrec) == VAR_DECL
393 || TREE_CODE (chrec) == PARM_DECL
394 || TREE_CODE (chrec) == FUNCTION_DECL
395 || TREE_CODE (chrec) == LABEL_DECL
396 || TREE_CODE (chrec) == RESULT_DECL
397 || TREE_CODE (chrec) == FIELD_DECL)
398 return true;
400 if (TREE_CODE (chrec) == SSA_NAME)
402 gimple def = SSA_NAME_DEF_STMT (chrec);
403 struct loop *def_loop = loop_containing_stmt (def);
404 struct loop *loop = get_loop (loop_nb);
406 if (def_loop == NULL)
407 return false;
409 if (loop == def_loop || flow_loop_nested_p (loop, def_loop))
410 return true;
412 return false;
415 n = TREE_OPERAND_LENGTH (chrec);
416 for (i = 0; i < n; i++)
417 if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, i),
418 loop_nb))
419 return true;
420 return false;
423 /* Return true when PHI is a loop-phi-node. */
425 static bool
426 loop_phi_node_p (gimple phi)
428 /* The implementation of this function is based on the following
429 property: "all the loop-phi-nodes of a loop are contained in the
430 loop's header basic block". */
432 return loop_containing_stmt (phi)->header == gimple_bb (phi);
435 /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP.
436 In general, in the case of multivariate evolutions we want to get
437 the evolution in different loops. LOOP specifies the level for
438 which to get the evolution.
440 Example:
442 | for (j = 0; j < 100; j++)
444 | for (k = 0; k < 100; k++)
446 | i = k + j; - Here the value of i is a function of j, k.
448 | ... = i - Here the value of i is a function of j.
450 | ... = i - Here the value of i is a scalar.
452 Example:
454 | i_0 = ...
455 | loop_1 10 times
456 | i_1 = phi (i_0, i_2)
457 | i_2 = i_1 + 2
458 | endloop
460 This loop has the same effect as:
461 LOOP_1 has the same effect as:
463 | i_1 = i_0 + 20
465 The overall effect of the loop, "i_0 + 20" in the previous example,
466 is obtained by passing in the parameters: LOOP = 1,
467 EVOLUTION_FN = {i_0, +, 2}_1.
470 static tree
471 compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn)
473 bool val = false;
475 if (evolution_fn == chrec_dont_know)
476 return chrec_dont_know;
478 else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC)
480 struct loop *inner_loop = get_chrec_loop (evolution_fn);
482 if (inner_loop == loop
483 || flow_loop_nested_p (loop, inner_loop))
485 tree nb_iter = number_of_latch_executions (inner_loop);
487 if (nb_iter == chrec_dont_know)
488 return chrec_dont_know;
489 else
491 tree res;
493 /* evolution_fn is the evolution function in LOOP. Get
494 its value in the nb_iter-th iteration. */
495 res = chrec_apply (inner_loop->num, evolution_fn, nb_iter);
497 /* Continue the computation until ending on a parent of LOOP. */
498 return compute_overall_effect_of_inner_loop (loop, res);
501 else
502 return evolution_fn;
505 /* If the evolution function is an invariant, there is nothing to do. */
506 else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val)
507 return evolution_fn;
509 else
510 return chrec_dont_know;
513 /* Determine whether the CHREC is always positive/negative. If the expression
514 cannot be statically analyzed, return false, otherwise set the answer into
515 VALUE. */
517 bool
518 chrec_is_positive (tree chrec, bool *value)
520 bool value0, value1, value2;
521 tree end_value, nb_iter;
523 switch (TREE_CODE (chrec))
525 case POLYNOMIAL_CHREC:
526 if (!chrec_is_positive (CHREC_LEFT (chrec), &value0)
527 || !chrec_is_positive (CHREC_RIGHT (chrec), &value1))
528 return false;
530 /* FIXME -- overflows. */
531 if (value0 == value1)
533 *value = value0;
534 return true;
537 /* Otherwise the chrec is under the form: "{-197, +, 2}_1",
538 and the proof consists in showing that the sign never
539 changes during the execution of the loop, from 0 to
540 loop->nb_iterations. */
541 if (!evolution_function_is_affine_p (chrec))
542 return false;
544 nb_iter = number_of_latch_executions (get_chrec_loop (chrec));
545 if (chrec_contains_undetermined (nb_iter))
546 return false;
548 #if 0
549 /* TODO -- If the test is after the exit, we may decrease the number of
550 iterations by one. */
551 if (after_exit)
552 nb_iter = chrec_fold_minus (type, nb_iter, build_int_cst (type, 1));
553 #endif
555 end_value = chrec_apply (CHREC_VARIABLE (chrec), chrec, nb_iter);
557 if (!chrec_is_positive (end_value, &value2))
558 return false;
560 *value = value0;
561 return value0 == value1;
563 case INTEGER_CST:
564 *value = (tree_int_cst_sgn (chrec) == 1);
565 return true;
567 default:
568 return false;
572 /* Associate CHREC to SCALAR. */
574 static void
575 set_scalar_evolution (basic_block instantiated_below, tree scalar, tree chrec)
577 tree *scalar_info;
579 if (TREE_CODE (scalar) != SSA_NAME)
580 return;
582 scalar_info = find_var_scev_info (instantiated_below, scalar);
584 if (dump_file)
586 if (dump_flags & TDF_DETAILS)
588 fprintf (dump_file, "(set_scalar_evolution \n");
589 fprintf (dump_file, " instantiated_below = %d \n",
590 instantiated_below->index);
591 fprintf (dump_file, " (scalar = ");
592 print_generic_expr (dump_file, scalar, 0);
593 fprintf (dump_file, ")\n (scalar_evolution = ");
594 print_generic_expr (dump_file, chrec, 0);
595 fprintf (dump_file, "))\n");
597 if (dump_flags & TDF_STATS)
598 nb_set_scev++;
601 *scalar_info = chrec;
604 /* Retrieve the chrec associated to SCALAR instantiated below
605 INSTANTIATED_BELOW block. */
607 static tree
608 get_scalar_evolution (basic_block instantiated_below, tree scalar)
610 tree res;
612 if (dump_file)
614 if (dump_flags & TDF_DETAILS)
616 fprintf (dump_file, "(get_scalar_evolution \n");
617 fprintf (dump_file, " (scalar = ");
618 print_generic_expr (dump_file, scalar, 0);
619 fprintf (dump_file, ")\n");
621 if (dump_flags & TDF_STATS)
622 nb_get_scev++;
625 switch (TREE_CODE (scalar))
627 case SSA_NAME:
628 res = *find_var_scev_info (instantiated_below, scalar);
629 break;
631 case REAL_CST:
632 case FIXED_CST:
633 case INTEGER_CST:
634 res = scalar;
635 break;
637 default:
638 res = chrec_not_analyzed_yet;
639 break;
642 if (dump_file && (dump_flags & TDF_DETAILS))
644 fprintf (dump_file, " (scalar_evolution = ");
645 print_generic_expr (dump_file, res, 0);
646 fprintf (dump_file, "))\n");
649 return res;
652 /* Helper function for add_to_evolution. Returns the evolution
653 function for an assignment of the form "a = b + c", where "a" and
654 "b" are on the strongly connected component. CHREC_BEFORE is the
655 information that we already have collected up to this point.
656 TO_ADD is the evolution of "c".
658 When CHREC_BEFORE has an evolution part in LOOP_NB, add to this
659 evolution the expression TO_ADD, otherwise construct an evolution
660 part for this loop. */
662 static tree
663 add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add,
664 gimple at_stmt)
666 tree type, left, right;
667 struct loop *loop = get_loop (loop_nb), *chloop;
669 switch (TREE_CODE (chrec_before))
671 case POLYNOMIAL_CHREC:
672 chloop = get_chrec_loop (chrec_before);
673 if (chloop == loop
674 || flow_loop_nested_p (chloop, loop))
676 unsigned var;
678 type = chrec_type (chrec_before);
680 /* When there is no evolution part in this loop, build it. */
681 if (chloop != loop)
683 var = loop_nb;
684 left = chrec_before;
685 right = SCALAR_FLOAT_TYPE_P (type)
686 ? build_real (type, dconst0)
687 : build_int_cst (type, 0);
689 else
691 var = CHREC_VARIABLE (chrec_before);
692 left = CHREC_LEFT (chrec_before);
693 right = CHREC_RIGHT (chrec_before);
696 to_add = chrec_convert (type, to_add, at_stmt);
697 right = chrec_convert_rhs (type, right, at_stmt);
698 right = chrec_fold_plus (chrec_type (right), right, to_add);
699 return build_polynomial_chrec (var, left, right);
701 else
703 gcc_assert (flow_loop_nested_p (loop, chloop));
705 /* Search the evolution in LOOP_NB. */
706 left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before),
707 to_add, at_stmt);
708 right = CHREC_RIGHT (chrec_before);
709 right = chrec_convert_rhs (chrec_type (left), right, at_stmt);
710 return build_polynomial_chrec (CHREC_VARIABLE (chrec_before),
711 left, right);
714 default:
715 /* These nodes do not depend on a loop. */
716 if (chrec_before == chrec_dont_know)
717 return chrec_dont_know;
719 left = chrec_before;
720 right = chrec_convert_rhs (chrec_type (left), to_add, at_stmt);
721 return build_polynomial_chrec (loop_nb, left, right);
725 /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension
726 of LOOP_NB.
728 Description (provided for completeness, for those who read code in
729 a plane, and for my poor 62 bytes brain that would have forgotten
730 all this in the next two or three months):
732 The algorithm of translation of programs from the SSA representation
733 into the chrecs syntax is based on a pattern matching. After having
734 reconstructed the overall tree expression for a loop, there are only
735 two cases that can arise:
737 1. a = loop-phi (init, a + expr)
738 2. a = loop-phi (init, expr)
740 where EXPR is either a scalar constant with respect to the analyzed
741 loop (this is a degree 0 polynomial), or an expression containing
742 other loop-phi definitions (these are higher degree polynomials).
744 Examples:
747 | init = ...
748 | loop_1
749 | a = phi (init, a + 5)
750 | endloop
753 | inita = ...
754 | initb = ...
755 | loop_1
756 | a = phi (inita, 2 * b + 3)
757 | b = phi (initb, b + 1)
758 | endloop
760 For the first case, the semantics of the SSA representation is:
762 | a (x) = init + \sum_{j = 0}^{x - 1} expr (j)
764 that is, there is a loop index "x" that determines the scalar value
765 of the variable during the loop execution. During the first
766 iteration, the value is that of the initial condition INIT, while
767 during the subsequent iterations, it is the sum of the initial
768 condition with the sum of all the values of EXPR from the initial
769 iteration to the before last considered iteration.
771 For the second case, the semantics of the SSA program is:
773 | a (x) = init, if x = 0;
774 | expr (x - 1), otherwise.
776 The second case corresponds to the PEELED_CHREC, whose syntax is
777 close to the syntax of a loop-phi-node:
779 | phi (init, expr) vs. (init, expr)_x
781 The proof of the translation algorithm for the first case is a
782 proof by structural induction based on the degree of EXPR.
784 Degree 0:
785 When EXPR is a constant with respect to the analyzed loop, or in
786 other words when EXPR is a polynomial of degree 0, the evolution of
787 the variable A in the loop is an affine function with an initial
788 condition INIT, and a step EXPR. In order to show this, we start
789 from the semantics of the SSA representation:
791 f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
793 and since "expr (j)" is a constant with respect to "j",
795 f (x) = init + x * expr
797 Finally, based on the semantics of the pure sum chrecs, by
798 identification we get the corresponding chrecs syntax:
800 f (x) = init * \binom{x}{0} + expr * \binom{x}{1}
801 f (x) -> {init, +, expr}_x
803 Higher degree:
804 Suppose that EXPR is a polynomial of degree N with respect to the
805 analyzed loop_x for which we have already determined that it is
806 written under the chrecs syntax:
808 | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x)
810 We start from the semantics of the SSA program:
812 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j)
814 | f (x) = init + \sum_{j = 0}^{x - 1}
815 | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1})
817 | f (x) = init + \sum_{j = 0}^{x - 1}
818 | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k})
820 | f (x) = init + \sum_{k = 0}^{n - 1}
821 | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k})
823 | f (x) = init + \sum_{k = 0}^{n - 1}
824 | (b_k * \binom{x}{k + 1})
826 | f (x) = init + b_0 * \binom{x}{1} + ...
827 | + b_{n-1} * \binom{x}{n}
829 | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ...
830 | + b_{n-1} * \binom{x}{n}
833 And finally from the definition of the chrecs syntax, we identify:
834 | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x
836 This shows the mechanism that stands behind the add_to_evolution
837 function. An important point is that the use of symbolic
838 parameters avoids the need of an analysis schedule.
840 Example:
842 | inita = ...
843 | initb = ...
844 | loop_1
845 | a = phi (inita, a + 2 + b)
846 | b = phi (initb, b + 1)
847 | endloop
849 When analyzing "a", the algorithm keeps "b" symbolically:
851 | a -> {inita, +, 2 + b}_1
853 Then, after instantiation, the analyzer ends on the evolution:
855 | a -> {inita, +, 2 + initb, +, 1}_1
859 static tree
860 add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code,
861 tree to_add, gimple at_stmt)
863 tree type = chrec_type (to_add);
864 tree res = NULL_TREE;
866 if (to_add == NULL_TREE)
867 return chrec_before;
869 /* TO_ADD is either a scalar, or a parameter. TO_ADD is not
870 instantiated at this point. */
871 if (TREE_CODE (to_add) == POLYNOMIAL_CHREC)
872 /* This should not happen. */
873 return chrec_dont_know;
875 if (dump_file && (dump_flags & TDF_DETAILS))
877 fprintf (dump_file, "(add_to_evolution \n");
878 fprintf (dump_file, " (loop_nb = %d)\n", loop_nb);
879 fprintf (dump_file, " (chrec_before = ");
880 print_generic_expr (dump_file, chrec_before, 0);
881 fprintf (dump_file, ")\n (to_add = ");
882 print_generic_expr (dump_file, to_add, 0);
883 fprintf (dump_file, ")\n");
886 if (code == MINUS_EXPR)
887 to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type)
888 ? build_real (type, dconstm1)
889 : build_int_cst_type (type, -1));
891 res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt);
893 if (dump_file && (dump_flags & TDF_DETAILS))
895 fprintf (dump_file, " (res = ");
896 print_generic_expr (dump_file, res, 0);
897 fprintf (dump_file, "))\n");
900 return res;
903 /* Helper function. */
905 static inline tree
906 set_nb_iterations_in_loop (struct loop *loop,
907 tree res)
909 if (dump_file && (dump_flags & TDF_DETAILS))
911 fprintf (dump_file, " (set_nb_iterations_in_loop = ");
912 print_generic_expr (dump_file, res, 0);
913 fprintf (dump_file, "))\n");
916 loop->nb_iterations = res;
917 return res;
922 /* This section selects the loops that will be good candidates for the
923 scalar evolution analysis. For the moment, greedily select all the
924 loop nests we could analyze. */
926 /* For a loop with a single exit edge, return the COND_EXPR that
927 guards the exit edge. If the expression is too difficult to
928 analyze, then give up. */
930 gimple
931 get_loop_exit_condition (const struct loop *loop)
933 gimple res = NULL;
934 edge exit_edge = single_exit (loop);
936 if (dump_file && (dump_flags & TDF_DETAILS))
937 fprintf (dump_file, "(get_loop_exit_condition \n ");
939 if (exit_edge)
941 gimple stmt;
943 stmt = last_stmt (exit_edge->src);
944 if (gimple_code (stmt) == GIMPLE_COND)
945 res = stmt;
948 if (dump_file && (dump_flags & TDF_DETAILS))
950 print_gimple_stmt (dump_file, res, 0, 0);
951 fprintf (dump_file, ")\n");
954 return res;
957 /* Recursively determine and enqueue the exit conditions for a loop. */
959 static void
960 get_exit_conditions_rec (struct loop *loop,
961 VEC(gimple,heap) **exit_conditions)
963 if (!loop)
964 return;
966 /* Recurse on the inner loops, then on the next (sibling) loops. */
967 get_exit_conditions_rec (loop->inner, exit_conditions);
968 get_exit_conditions_rec (loop->next, exit_conditions);
970 if (single_exit (loop))
972 gimple loop_condition = get_loop_exit_condition (loop);
974 if (loop_condition)
975 VEC_safe_push (gimple, heap, *exit_conditions, loop_condition);
979 /* Select the candidate loop nests for the analysis. This function
980 initializes the EXIT_CONDITIONS array. */
982 static void
983 select_loops_exit_conditions (VEC(gimple,heap) **exit_conditions)
985 struct loop *function_body = current_loops->tree_root;
987 get_exit_conditions_rec (function_body->inner, exit_conditions);
991 /* Depth first search algorithm. */
993 typedef enum t_bool {
994 t_false,
995 t_true,
996 t_dont_know
997 } t_bool;
1000 static t_bool follow_ssa_edge (struct loop *loop, gimple, gimple, tree *, int);
1002 /* Follow the ssa edge into the binary expression RHS0 CODE RHS1.
1003 Return true if the strongly connected component has been found. */
1005 static t_bool
1006 follow_ssa_edge_binary (struct loop *loop, gimple at_stmt,
1007 tree type, tree rhs0, enum tree_code code, tree rhs1,
1008 gimple halting_phi, tree *evolution_of_loop, int limit)
1010 t_bool res = t_false;
1011 tree evol;
1013 switch (code)
1015 case POINTER_PLUS_EXPR:
1016 case PLUS_EXPR:
1017 if (TREE_CODE (rhs0) == SSA_NAME)
1019 if (TREE_CODE (rhs1) == SSA_NAME)
1021 /* Match an assignment under the form:
1022 "a = b + c". */
1024 /* We want only assignments of form "name + name" contribute to
1025 LIMIT, as the other cases do not necessarily contribute to
1026 the complexity of the expression. */
1027 limit++;
1029 evol = *evolution_of_loop;
1030 res = follow_ssa_edge
1031 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, &evol, limit);
1033 if (res == t_true)
1034 *evolution_of_loop = add_to_evolution
1035 (loop->num,
1036 chrec_convert (type, evol, at_stmt),
1037 code, rhs1, at_stmt);
1039 else if (res == t_false)
1041 res = follow_ssa_edge
1042 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1043 evolution_of_loop, limit);
1045 if (res == t_true)
1046 *evolution_of_loop = add_to_evolution
1047 (loop->num,
1048 chrec_convert (type, *evolution_of_loop, at_stmt),
1049 code, rhs0, at_stmt);
1051 else if (res == t_dont_know)
1052 *evolution_of_loop = chrec_dont_know;
1055 else if (res == t_dont_know)
1056 *evolution_of_loop = chrec_dont_know;
1059 else
1061 /* Match an assignment under the form:
1062 "a = b + ...". */
1063 res = follow_ssa_edge
1064 (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1065 evolution_of_loop, limit);
1066 if (res == t_true)
1067 *evolution_of_loop = add_to_evolution
1068 (loop->num, chrec_convert (type, *evolution_of_loop,
1069 at_stmt),
1070 code, rhs1, at_stmt);
1072 else if (res == t_dont_know)
1073 *evolution_of_loop = chrec_dont_know;
1077 else if (TREE_CODE (rhs1) == SSA_NAME)
1079 /* Match an assignment under the form:
1080 "a = ... + c". */
1081 res = follow_ssa_edge
1082 (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi,
1083 evolution_of_loop, limit);
1084 if (res == t_true)
1085 *evolution_of_loop = add_to_evolution
1086 (loop->num, chrec_convert (type, *evolution_of_loop,
1087 at_stmt),
1088 code, rhs0, at_stmt);
1090 else if (res == t_dont_know)
1091 *evolution_of_loop = chrec_dont_know;
1094 else
1095 /* Otherwise, match an assignment under the form:
1096 "a = ... + ...". */
1097 /* And there is nothing to do. */
1098 res = t_false;
1099 break;
1101 case MINUS_EXPR:
1102 /* This case is under the form "opnd0 = rhs0 - rhs1". */
1103 if (TREE_CODE (rhs0) == SSA_NAME)
1105 /* Match an assignment under the form:
1106 "a = b - ...". */
1108 /* We want only assignments of form "name - name" contribute to
1109 LIMIT, as the other cases do not necessarily contribute to
1110 the complexity of the expression. */
1111 if (TREE_CODE (rhs1) == SSA_NAME)
1112 limit++;
1114 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi,
1115 evolution_of_loop, limit);
1116 if (res == t_true)
1117 *evolution_of_loop = add_to_evolution
1118 (loop->num, chrec_convert (type, *evolution_of_loop, at_stmt),
1119 MINUS_EXPR, rhs1, at_stmt);
1121 else if (res == t_dont_know)
1122 *evolution_of_loop = chrec_dont_know;
1124 else
1125 /* Otherwise, match an assignment under the form:
1126 "a = ... - ...". */
1127 /* And there is nothing to do. */
1128 res = t_false;
1129 break;
1131 default:
1132 res = t_false;
1135 return res;
1138 /* Follow the ssa edge into the expression EXPR.
1139 Return true if the strongly connected component has been found. */
1141 static t_bool
1142 follow_ssa_edge_expr (struct loop *loop, gimple at_stmt, tree expr,
1143 gimple halting_phi, tree *evolution_of_loop, int limit)
1145 t_bool res = t_false;
1146 tree rhs0, rhs1;
1147 tree type = TREE_TYPE (expr);
1148 enum tree_code code;
1150 /* The EXPR is one of the following cases:
1151 - an SSA_NAME,
1152 - an INTEGER_CST,
1153 - a PLUS_EXPR,
1154 - a POINTER_PLUS_EXPR,
1155 - a MINUS_EXPR,
1156 - an ASSERT_EXPR,
1157 - other cases are not yet handled. */
1158 code = TREE_CODE (expr);
1159 switch (code)
1161 case NOP_EXPR:
1162 /* This assignment is under the form "a_1 = (cast) rhs. */
1163 res = follow_ssa_edge_expr (loop, at_stmt, TREE_OPERAND (expr, 0),
1164 halting_phi, evolution_of_loop, limit);
1165 *evolution_of_loop = chrec_convert (type, *evolution_of_loop, at_stmt);
1166 break;
1168 case INTEGER_CST:
1169 /* This assignment is under the form "a_1 = 7". */
1170 res = t_false;
1171 break;
1173 case SSA_NAME:
1174 /* This assignment is under the form: "a_1 = b_2". */
1175 res = follow_ssa_edge
1176 (loop, SSA_NAME_DEF_STMT (expr), halting_phi, evolution_of_loop, limit);
1177 break;
1179 case POINTER_PLUS_EXPR:
1180 case PLUS_EXPR:
1181 case MINUS_EXPR:
1182 /* This case is under the form "rhs0 +- rhs1". */
1183 rhs0 = TREE_OPERAND (expr, 0);
1184 rhs1 = TREE_OPERAND (expr, 1);
1185 STRIP_TYPE_NOPS (rhs0);
1186 STRIP_TYPE_NOPS (rhs1);
1187 return follow_ssa_edge_binary (loop, at_stmt, type, rhs0, code, rhs1,
1188 halting_phi, evolution_of_loop, limit);
1190 case ASSERT_EXPR:
1192 /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>"
1193 It must be handled as a copy assignment of the form a_1 = a_2. */
1194 tree op0 = ASSERT_EXPR_VAR (expr);
1195 if (TREE_CODE (op0) == SSA_NAME)
1196 res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (op0),
1197 halting_phi, evolution_of_loop, limit);
1198 else
1199 res = t_false;
1200 break;
1204 default:
1205 res = t_false;
1206 break;
1209 return res;
1212 /* Follow the ssa edge into the right hand side of an assignment STMT.
1213 Return true if the strongly connected component has been found. */
1215 static t_bool
1216 follow_ssa_edge_in_rhs (struct loop *loop, gimple stmt,
1217 gimple halting_phi, tree *evolution_of_loop, int limit)
1219 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
1220 enum tree_code code = gimple_assign_rhs_code (stmt);
1222 switch (get_gimple_rhs_class (code))
1224 case GIMPLE_BINARY_RHS:
1225 return follow_ssa_edge_binary (loop, stmt, type,
1226 gimple_assign_rhs1 (stmt), code,
1227 gimple_assign_rhs2 (stmt),
1228 halting_phi, evolution_of_loop, limit);
1229 case GIMPLE_SINGLE_RHS:
1230 return follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt),
1231 halting_phi, evolution_of_loop, limit);
1232 default:
1233 return t_false;
1237 /* Checks whether the I-th argument of a PHI comes from a backedge. */
1239 static bool
1240 backedge_phi_arg_p (gimple phi, int i)
1242 const_edge e = gimple_phi_arg_edge (phi, i);
1244 /* We would in fact like to test EDGE_DFS_BACK here, but we do not care
1245 about updating it anywhere, and this should work as well most of the
1246 time. */
1247 if (e->flags & EDGE_IRREDUCIBLE_LOOP)
1248 return true;
1250 return false;
1253 /* Helper function for one branch of the condition-phi-node. Return
1254 true if the strongly connected component has been found following
1255 this path. */
1257 static inline t_bool
1258 follow_ssa_edge_in_condition_phi_branch (int i,
1259 struct loop *loop,
1260 gimple condition_phi,
1261 gimple halting_phi,
1262 tree *evolution_of_branch,
1263 tree init_cond, int limit)
1265 tree branch = PHI_ARG_DEF (condition_phi, i);
1266 *evolution_of_branch = chrec_dont_know;
1268 /* Do not follow back edges (they must belong to an irreducible loop, which
1269 we really do not want to worry about). */
1270 if (backedge_phi_arg_p (condition_phi, i))
1271 return t_false;
1273 if (TREE_CODE (branch) == SSA_NAME)
1275 *evolution_of_branch = init_cond;
1276 return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi,
1277 evolution_of_branch, limit);
1280 /* This case occurs when one of the condition branches sets
1281 the variable to a constant: i.e. a phi-node like
1282 "a_2 = PHI <a_7(5), 2(6)>;".
1284 FIXME: This case have to be refined correctly:
1285 in some cases it is possible to say something better than
1286 chrec_dont_know, for example using a wrap-around notation. */
1287 return t_false;
1290 /* This function merges the branches of a condition-phi-node in a
1291 loop. */
1293 static t_bool
1294 follow_ssa_edge_in_condition_phi (struct loop *loop,
1295 gimple condition_phi,
1296 gimple halting_phi,
1297 tree *evolution_of_loop, int limit)
1299 int i, n;
1300 tree init = *evolution_of_loop;
1301 tree evolution_of_branch;
1302 t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi,
1303 halting_phi,
1304 &evolution_of_branch,
1305 init, limit);
1306 if (res == t_false || res == t_dont_know)
1307 return res;
1309 *evolution_of_loop = evolution_of_branch;
1311 /* If the phi node is just a copy, do not increase the limit. */
1312 n = gimple_phi_num_args (condition_phi);
1313 if (n > 1)
1314 limit++;
1316 for (i = 1; i < n; i++)
1318 /* Quickly give up when the evolution of one of the branches is
1319 not known. */
1320 if (*evolution_of_loop == chrec_dont_know)
1321 return t_true;
1323 res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi,
1324 halting_phi,
1325 &evolution_of_branch,
1326 init, limit);
1327 if (res == t_false || res == t_dont_know)
1328 return res;
1330 *evolution_of_loop = chrec_merge (*evolution_of_loop,
1331 evolution_of_branch);
1334 return t_true;
1337 /* Follow an SSA edge in an inner loop. It computes the overall
1338 effect of the loop, and following the symbolic initial conditions,
1339 it follows the edges in the parent loop. The inner loop is
1340 considered as a single statement. */
1342 static t_bool
1343 follow_ssa_edge_inner_loop_phi (struct loop *outer_loop,
1344 gimple loop_phi_node,
1345 gimple halting_phi,
1346 tree *evolution_of_loop, int limit)
1348 struct loop *loop = loop_containing_stmt (loop_phi_node);
1349 tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node));
1351 /* Sometimes, the inner loop is too difficult to analyze, and the
1352 result of the analysis is a symbolic parameter. */
1353 if (ev == PHI_RESULT (loop_phi_node))
1355 t_bool res = t_false;
1356 int i, n = gimple_phi_num_args (loop_phi_node);
1358 for (i = 0; i < n; i++)
1360 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1361 basic_block bb;
1363 /* Follow the edges that exit the inner loop. */
1364 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1365 if (!flow_bb_inside_loop_p (loop, bb))
1366 res = follow_ssa_edge_expr (outer_loop, loop_phi_node,
1367 arg, halting_phi,
1368 evolution_of_loop, limit);
1369 if (res == t_true)
1370 break;
1373 /* If the path crosses this loop-phi, give up. */
1374 if (res == t_true)
1375 *evolution_of_loop = chrec_dont_know;
1377 return res;
1380 /* Otherwise, compute the overall effect of the inner loop. */
1381 ev = compute_overall_effect_of_inner_loop (loop, ev);
1382 return follow_ssa_edge_expr (outer_loop, loop_phi_node, ev, halting_phi,
1383 evolution_of_loop, limit);
1386 /* Follow an SSA edge from a loop-phi-node to itself, constructing a
1387 path that is analyzed on the return walk. */
1389 static t_bool
1390 follow_ssa_edge (struct loop *loop, gimple def, gimple halting_phi,
1391 tree *evolution_of_loop, int limit)
1393 struct loop *def_loop;
1395 if (gimple_nop_p (def))
1396 return t_false;
1398 /* Give up if the path is longer than the MAX that we allow. */
1399 if (limit > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
1400 return t_dont_know;
1402 def_loop = loop_containing_stmt (def);
1404 switch (gimple_code (def))
1406 case GIMPLE_PHI:
1407 if (!loop_phi_node_p (def))
1408 /* DEF is a condition-phi-node. Follow the branches, and
1409 record their evolutions. Finally, merge the collected
1410 information and set the approximation to the main
1411 variable. */
1412 return follow_ssa_edge_in_condition_phi
1413 (loop, def, halting_phi, evolution_of_loop, limit);
1415 /* When the analyzed phi is the halting_phi, the
1416 depth-first search is over: we have found a path from
1417 the halting_phi to itself in the loop. */
1418 if (def == halting_phi)
1419 return t_true;
1421 /* Otherwise, the evolution of the HALTING_PHI depends
1422 on the evolution of another loop-phi-node, i.e. the
1423 evolution function is a higher degree polynomial. */
1424 if (def_loop == loop)
1425 return t_false;
1427 /* Inner loop. */
1428 if (flow_loop_nested_p (loop, def_loop))
1429 return follow_ssa_edge_inner_loop_phi
1430 (loop, def, halting_phi, evolution_of_loop, limit + 1);
1432 /* Outer loop. */
1433 return t_false;
1435 case GIMPLE_ASSIGN:
1436 return follow_ssa_edge_in_rhs (loop, def, halting_phi,
1437 evolution_of_loop, limit);
1439 default:
1440 /* At this level of abstraction, the program is just a set
1441 of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no
1442 other node to be handled. */
1443 return t_false;
1449 /* Given a LOOP_PHI_NODE, this function determines the evolution
1450 function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */
1452 static tree
1453 analyze_evolution_in_loop (gimple loop_phi_node,
1454 tree init_cond)
1456 int i, n = gimple_phi_num_args (loop_phi_node);
1457 tree evolution_function = chrec_not_analyzed_yet;
1458 struct loop *loop = loop_containing_stmt (loop_phi_node);
1459 basic_block bb;
1461 if (dump_file && (dump_flags & TDF_DETAILS))
1463 fprintf (dump_file, "(analyze_evolution_in_loop \n");
1464 fprintf (dump_file, " (loop_phi_node = ");
1465 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
1466 fprintf (dump_file, ")\n");
1469 for (i = 0; i < n; i++)
1471 tree arg = PHI_ARG_DEF (loop_phi_node, i);
1472 gimple ssa_chain;
1473 tree ev_fn;
1474 t_bool res;
1476 /* Select the edges that enter the loop body. */
1477 bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1478 if (!flow_bb_inside_loop_p (loop, bb))
1479 continue;
1481 if (TREE_CODE (arg) == SSA_NAME)
1483 ssa_chain = SSA_NAME_DEF_STMT (arg);
1485 /* Pass in the initial condition to the follow edge function. */
1486 ev_fn = init_cond;
1487 res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0);
1489 else
1490 res = t_false;
1492 /* When it is impossible to go back on the same
1493 loop_phi_node by following the ssa edges, the
1494 evolution is represented by a peeled chrec, i.e. the
1495 first iteration, EV_FN has the value INIT_COND, then
1496 all the other iterations it has the value of ARG.
1497 For the moment, PEELED_CHREC nodes are not built. */
1498 if (res != t_true)
1499 ev_fn = chrec_dont_know;
1501 /* When there are multiple back edges of the loop (which in fact never
1502 happens currently, but nevertheless), merge their evolutions. */
1503 evolution_function = chrec_merge (evolution_function, ev_fn);
1506 if (dump_file && (dump_flags & TDF_DETAILS))
1508 fprintf (dump_file, " (evolution_function = ");
1509 print_generic_expr (dump_file, evolution_function, 0);
1510 fprintf (dump_file, "))\n");
1513 return evolution_function;
1516 /* Given a loop-phi-node, return the initial conditions of the
1517 variable on entry of the loop. When the CCP has propagated
1518 constants into the loop-phi-node, the initial condition is
1519 instantiated, otherwise the initial condition is kept symbolic.
1520 This analyzer does not analyze the evolution outside the current
1521 loop, and leaves this task to the on-demand tree reconstructor. */
1523 static tree
1524 analyze_initial_condition (gimple loop_phi_node)
1526 int i, n;
1527 tree init_cond = chrec_not_analyzed_yet;
1528 struct loop *loop = loop_containing_stmt (loop_phi_node);
1530 if (dump_file && (dump_flags & TDF_DETAILS))
1532 fprintf (dump_file, "(analyze_initial_condition \n");
1533 fprintf (dump_file, " (loop_phi_node = \n");
1534 print_gimple_stmt (dump_file, loop_phi_node, 0, 0);
1535 fprintf (dump_file, ")\n");
1538 n = gimple_phi_num_args (loop_phi_node);
1539 for (i = 0; i < n; i++)
1541 tree branch = PHI_ARG_DEF (loop_phi_node, i);
1542 basic_block bb = gimple_phi_arg_edge (loop_phi_node, i)->src;
1544 /* When the branch is oriented to the loop's body, it does
1545 not contribute to the initial condition. */
1546 if (flow_bb_inside_loop_p (loop, bb))
1547 continue;
1549 if (init_cond == chrec_not_analyzed_yet)
1551 init_cond = branch;
1552 continue;
1555 if (TREE_CODE (branch) == SSA_NAME)
1557 init_cond = chrec_dont_know;
1558 break;
1561 init_cond = chrec_merge (init_cond, branch);
1564 /* Ooops -- a loop without an entry??? */
1565 if (init_cond == chrec_not_analyzed_yet)
1566 init_cond = chrec_dont_know;
1568 if (dump_file && (dump_flags & TDF_DETAILS))
1570 fprintf (dump_file, " (init_cond = ");
1571 print_generic_expr (dump_file, init_cond, 0);
1572 fprintf (dump_file, "))\n");
1575 return init_cond;
1578 /* Analyze the scalar evolution for LOOP_PHI_NODE. */
1580 static tree
1581 interpret_loop_phi (struct loop *loop, gimple loop_phi_node)
1583 tree res;
1584 struct loop *phi_loop = loop_containing_stmt (loop_phi_node);
1585 tree init_cond;
1587 if (phi_loop != loop)
1589 struct loop *subloop;
1590 tree evolution_fn = analyze_scalar_evolution
1591 (phi_loop, PHI_RESULT (loop_phi_node));
1593 /* Dive one level deeper. */
1594 subloop = superloop_at_depth (phi_loop, loop_depth (loop) + 1);
1596 /* Interpret the subloop. */
1597 res = compute_overall_effect_of_inner_loop (subloop, evolution_fn);
1598 return res;
1601 /* Otherwise really interpret the loop phi. */
1602 init_cond = analyze_initial_condition (loop_phi_node);
1603 res = analyze_evolution_in_loop (loop_phi_node, init_cond);
1605 return res;
1608 /* This function merges the branches of a condition-phi-node,
1609 contained in the outermost loop, and whose arguments are already
1610 analyzed. */
1612 static tree
1613 interpret_condition_phi (struct loop *loop, gimple condition_phi)
1615 int i, n = gimple_phi_num_args (condition_phi);
1616 tree res = chrec_not_analyzed_yet;
1618 for (i = 0; i < n; i++)
1620 tree branch_chrec;
1622 if (backedge_phi_arg_p (condition_phi, i))
1624 res = chrec_dont_know;
1625 break;
1628 branch_chrec = analyze_scalar_evolution
1629 (loop, PHI_ARG_DEF (condition_phi, i));
1631 res = chrec_merge (res, branch_chrec);
1634 return res;
1637 /* Interpret the operation RHS1 OP RHS2. If we didn't
1638 analyze this node before, follow the definitions until ending
1639 either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the
1640 return path, this function propagates evolutions (ala constant copy
1641 propagation). OPND1 is not a GIMPLE expression because we could
1642 analyze the effect of an inner loop: see interpret_loop_phi. */
1644 static tree
1645 interpret_rhs_expr (struct loop *loop, gimple at_stmt,
1646 tree type, tree rhs1, enum tree_code code, tree rhs2)
1648 tree res, chrec1, chrec2;
1650 if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS)
1652 if (is_gimple_min_invariant (rhs1))
1653 return chrec_convert (type, rhs1, at_stmt);
1655 if (code == SSA_NAME)
1656 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1657 at_stmt);
1659 if (code == ASSERT_EXPR)
1661 rhs1 = ASSERT_EXPR_VAR (rhs1);
1662 return chrec_convert (type, analyze_scalar_evolution (loop, rhs1),
1663 at_stmt);
1666 return chrec_dont_know;
1669 switch (code)
1671 case POINTER_PLUS_EXPR:
1672 chrec1 = analyze_scalar_evolution (loop, rhs1);
1673 chrec2 = analyze_scalar_evolution (loop, rhs2);
1674 chrec1 = chrec_convert (type, chrec1, at_stmt);
1675 chrec2 = chrec_convert (sizetype, chrec2, at_stmt);
1676 res = chrec_fold_plus (type, chrec1, chrec2);
1677 break;
1679 case PLUS_EXPR:
1680 chrec1 = analyze_scalar_evolution (loop, rhs1);
1681 chrec2 = analyze_scalar_evolution (loop, rhs2);
1682 chrec1 = chrec_convert (type, chrec1, at_stmt);
1683 chrec2 = chrec_convert (type, chrec2, at_stmt);
1684 res = chrec_fold_plus (type, chrec1, chrec2);
1685 break;
1687 case MINUS_EXPR:
1688 chrec1 = analyze_scalar_evolution (loop, rhs1);
1689 chrec2 = analyze_scalar_evolution (loop, rhs2);
1690 chrec1 = chrec_convert (type, chrec1, at_stmt);
1691 chrec2 = chrec_convert (type, chrec2, at_stmt);
1692 res = chrec_fold_minus (type, chrec1, chrec2);
1693 break;
1695 case NEGATE_EXPR:
1696 chrec1 = analyze_scalar_evolution (loop, rhs1);
1697 chrec1 = chrec_convert (type, chrec1, at_stmt);
1698 /* TYPE may be integer, real or complex, so use fold_convert. */
1699 res = chrec_fold_multiply (type, chrec1,
1700 fold_convert (type, integer_minus_one_node));
1701 break;
1703 case MULT_EXPR:
1704 chrec1 = analyze_scalar_evolution (loop, rhs1);
1705 chrec2 = analyze_scalar_evolution (loop, rhs2);
1706 chrec1 = chrec_convert (type, chrec1, at_stmt);
1707 chrec2 = chrec_convert (type, chrec2, at_stmt);
1708 res = chrec_fold_multiply (type, chrec1, chrec2);
1709 break;
1711 CASE_CONVERT:
1712 chrec1 = analyze_scalar_evolution (loop, rhs1);
1713 res = chrec_convert (type, chrec1, at_stmt);
1714 break;
1716 default:
1717 res = chrec_dont_know;
1718 break;
1721 return res;
1724 /* Interpret the expression EXPR. */
1726 static tree
1727 interpret_expr (struct loop *loop, gimple at_stmt, tree expr)
1729 enum tree_code code;
1730 tree type = TREE_TYPE (expr), op0, op1;
1732 if (automatically_generated_chrec_p (expr))
1733 return expr;
1735 if (TREE_CODE (expr) == POLYNOMIAL_CHREC)
1736 return chrec_dont_know;
1738 extract_ops_from_tree (expr, &code, &op0, &op1);
1740 return interpret_rhs_expr (loop, at_stmt, type,
1741 op0, code, op1);
1744 /* Interpret the rhs of the assignment STMT. */
1746 static tree
1747 interpret_gimple_assign (struct loop *loop, gimple stmt)
1749 tree type = TREE_TYPE (gimple_assign_lhs (stmt));
1750 enum tree_code code = gimple_assign_rhs_code (stmt);
1752 return interpret_rhs_expr (loop, stmt, type,
1753 gimple_assign_rhs1 (stmt), code,
1754 gimple_assign_rhs2 (stmt));
1759 /* This section contains all the entry points:
1760 - number_of_iterations_in_loop,
1761 - analyze_scalar_evolution,
1762 - instantiate_parameters.
1765 /* Compute and return the evolution function in WRTO_LOOP, the nearest
1766 common ancestor of DEF_LOOP and USE_LOOP. */
1768 static tree
1769 compute_scalar_evolution_in_loop (struct loop *wrto_loop,
1770 struct loop *def_loop,
1771 tree ev)
1773 tree res;
1774 if (def_loop == wrto_loop)
1775 return ev;
1777 def_loop = superloop_at_depth (def_loop, loop_depth (wrto_loop) + 1);
1778 res = compute_overall_effect_of_inner_loop (def_loop, ev);
1780 return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet);
1783 /* Helper recursive function. */
1785 static tree
1786 analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res)
1788 tree type = TREE_TYPE (var);
1789 gimple def;
1790 basic_block bb;
1791 struct loop *def_loop;
1793 if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE)
1794 return chrec_dont_know;
1796 if (TREE_CODE (var) != SSA_NAME)
1797 return interpret_expr (loop, NULL, var);
1799 def = SSA_NAME_DEF_STMT (var);
1800 bb = gimple_bb (def);
1801 def_loop = bb ? bb->loop_father : NULL;
1803 if (bb == NULL
1804 || !flow_bb_inside_loop_p (loop, bb))
1806 /* Keep the symbolic form. */
1807 res = var;
1808 goto set_and_end;
1811 if (res != chrec_not_analyzed_yet)
1813 if (loop != bb->loop_father)
1814 res = compute_scalar_evolution_in_loop
1815 (find_common_loop (loop, bb->loop_father), bb->loop_father, res);
1817 goto set_and_end;
1820 if (loop != def_loop)
1822 res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet);
1823 res = compute_scalar_evolution_in_loop (loop, def_loop, res);
1825 goto set_and_end;
1828 switch (gimple_code (def))
1830 case GIMPLE_ASSIGN:
1831 res = interpret_gimple_assign (loop, def);
1832 break;
1834 case GIMPLE_PHI:
1835 if (loop_phi_node_p (def))
1836 res = interpret_loop_phi (loop, def);
1837 else
1838 res = interpret_condition_phi (loop, def);
1839 break;
1841 default:
1842 res = chrec_dont_know;
1843 break;
1846 set_and_end:
1848 /* Keep the symbolic form. */
1849 if (res == chrec_dont_know)
1850 res = var;
1852 if (loop == def_loop)
1853 set_scalar_evolution (block_before_loop (loop), var, res);
1855 return res;
1858 /* Entry point for the scalar evolution analyzer.
1859 Analyzes and returns the scalar evolution of the ssa_name VAR.
1860 LOOP_NB is the identifier number of the loop in which the variable
1861 is used.
1863 Example of use: having a pointer VAR to a SSA_NAME node, STMT a
1864 pointer to the statement that uses this variable, in order to
1865 determine the evolution function of the variable, use the following
1866 calls:
1868 unsigned loop_nb = loop_containing_stmt (stmt)->num;
1869 tree chrec_with_symbols = analyze_scalar_evolution (loop_nb, var);
1870 tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols);
1873 tree
1874 analyze_scalar_evolution (struct loop *loop, tree var)
1876 tree res;
1878 if (dump_file && (dump_flags & TDF_DETAILS))
1880 fprintf (dump_file, "(analyze_scalar_evolution \n");
1881 fprintf (dump_file, " (loop_nb = %d)\n", loop->num);
1882 fprintf (dump_file, " (scalar = ");
1883 print_generic_expr (dump_file, var, 0);
1884 fprintf (dump_file, ")\n");
1887 res = get_scalar_evolution (block_before_loop (loop), var);
1888 res = analyze_scalar_evolution_1 (loop, var, res);
1890 if (dump_file && (dump_flags & TDF_DETAILS))
1891 fprintf (dump_file, ")\n");
1893 return res;
1896 /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to
1897 WRTO_LOOP (which should be a superloop of both USE_LOOP and definition
1898 of VERSION).
1900 FOLDED_CASTS is set to true if resolve_mixers used
1901 chrec_convert_aggressive (TODO -- not really, we are way too conservative
1902 at the moment in order to keep things simple). */
1904 static tree
1905 analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop,
1906 tree version, bool *folded_casts)
1908 bool val = false;
1909 tree ev = version, tmp;
1911 if (folded_casts)
1912 *folded_casts = false;
1913 while (1)
1915 tmp = analyze_scalar_evolution (use_loop, ev);
1916 ev = resolve_mixers (use_loop, tmp);
1918 if (folded_casts && tmp != ev)
1919 *folded_casts = true;
1921 if (use_loop == wrto_loop)
1922 return ev;
1924 /* If the value of the use changes in the inner loop, we cannot express
1925 its value in the outer loop (we might try to return interval chrec,
1926 but we do not have a user for it anyway) */
1927 if (!no_evolution_in_loop_p (ev, use_loop->num, &val)
1928 || !val)
1929 return chrec_dont_know;
1931 use_loop = loop_outer (use_loop);
1935 /* Returns from CACHE the value for VERSION instantiated below
1936 INSTANTIATED_BELOW block. */
1938 static tree
1939 get_instantiated_value (htab_t cache, basic_block instantiated_below,
1940 tree version)
1942 struct scev_info_str *info, pattern;
1944 pattern.var = version;
1945 pattern.instantiated_below = instantiated_below;
1946 info = (struct scev_info_str *) htab_find (cache, &pattern);
1948 if (info)
1949 return info->chrec;
1950 else
1951 return NULL_TREE;
1954 /* Sets in CACHE the value of VERSION instantiated below basic block
1955 INSTANTIATED_BELOW to VAL. */
1957 static void
1958 set_instantiated_value (htab_t cache, basic_block instantiated_below,
1959 tree version, tree val)
1961 struct scev_info_str *info, pattern;
1962 PTR *slot;
1964 pattern.var = version;
1965 pattern.instantiated_below = instantiated_below;
1966 slot = htab_find_slot (cache, &pattern, INSERT);
1968 if (!*slot)
1969 *slot = new_scev_info_str (instantiated_below, version);
1970 info = (struct scev_info_str *) *slot;
1971 info->chrec = val;
1974 /* Return the closed_loop_phi node for VAR. If there is none, return
1975 NULL_TREE. */
1977 static tree
1978 loop_closed_phi_def (tree var)
1980 struct loop *loop;
1981 edge exit;
1982 gimple phi;
1983 gimple_stmt_iterator psi;
1985 if (var == NULL_TREE
1986 || TREE_CODE (var) != SSA_NAME)
1987 return NULL_TREE;
1989 loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var));
1990 exit = single_exit (loop);
1991 if (!exit)
1992 return NULL_TREE;
1994 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi))
1996 phi = gsi_stmt (psi);
1997 if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var)
1998 return PHI_RESULT (phi);
2001 return NULL_TREE;
2004 /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW
2005 and EVOLUTION_LOOP, that were left under a symbolic form.
2007 CHREC is the scalar evolution to instantiate.
2009 CACHE is the cache of already instantiated values.
2011 FOLD_CONVERSIONS should be set to true when the conversions that
2012 may wrap in signed/pointer type are folded, as long as the value of
2013 the chrec is preserved.
2015 SIZE_EXPR is used for computing the size of the expression to be
2016 instantiated, and to stop if it exceeds some limit. */
2018 static tree
2019 instantiate_scev_1 (basic_block instantiate_below,
2020 struct loop *evolution_loop, tree chrec,
2021 bool fold_conversions, htab_t cache, int size_expr)
2023 tree res, op0, op1, op2;
2024 basic_block def_bb;
2025 struct loop *def_loop;
2026 tree type = chrec_type (chrec);
2028 /* Give up if the expression is larger than the MAX that we allow. */
2029 if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE))
2030 return chrec_dont_know;
2032 if (automatically_generated_chrec_p (chrec)
2033 || is_gimple_min_invariant (chrec))
2034 return chrec;
2036 switch (TREE_CODE (chrec))
2038 case SSA_NAME:
2039 def_bb = gimple_bb (SSA_NAME_DEF_STMT (chrec));
2041 /* A parameter (or loop invariant and we do not want to include
2042 evolutions in outer loops), nothing to do. */
2043 if (!def_bb
2044 || loop_depth (def_bb->loop_father) == 0
2045 || dominated_by_p (CDI_DOMINATORS, instantiate_below, def_bb))
2046 return chrec;
2048 /* We cache the value of instantiated variable to avoid exponential
2049 time complexity due to reevaluations. We also store the convenient
2050 value in the cache in order to prevent infinite recursion -- we do
2051 not want to instantiate the SSA_NAME if it is in a mixer
2052 structure. This is used for avoiding the instantiation of
2053 recursively defined functions, such as:
2055 | a_2 -> {0, +, 1, +, a_2}_1 */
2057 res = get_instantiated_value (cache, instantiate_below, chrec);
2058 if (res)
2059 return res;
2061 res = chrec_dont_know;
2062 set_instantiated_value (cache, instantiate_below, chrec, res);
2064 def_loop = find_common_loop (evolution_loop, def_bb->loop_father);
2066 /* If the analysis yields a parametric chrec, instantiate the
2067 result again. */
2068 res = analyze_scalar_evolution (def_loop, chrec);
2070 /* Don't instantiate loop-closed-ssa phi nodes. */
2071 if (TREE_CODE (res) == SSA_NAME
2072 && (loop_containing_stmt (SSA_NAME_DEF_STMT (res)) == NULL
2073 || (loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res)))
2074 > loop_depth (def_loop))))
2076 if (res == chrec)
2077 res = loop_closed_phi_def (chrec);
2078 else
2079 res = chrec;
2081 if (res == NULL_TREE)
2082 res = chrec_dont_know;
2085 else if (res != chrec_dont_know)
2086 res = instantiate_scev_1 (instantiate_below, evolution_loop, res,
2087 fold_conversions, cache, size_expr);
2089 /* Store the correct value to the cache. */
2090 set_instantiated_value (cache, instantiate_below, chrec, res);
2091 return res;
2093 case POLYNOMIAL_CHREC:
2094 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2095 CHREC_LEFT (chrec), fold_conversions, cache,
2096 size_expr);
2097 if (op0 == chrec_dont_know)
2098 return chrec_dont_know;
2100 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2101 CHREC_RIGHT (chrec), fold_conversions, cache,
2102 size_expr);
2103 if (op1 == chrec_dont_know)
2104 return chrec_dont_know;
2106 if (CHREC_LEFT (chrec) != op0
2107 || CHREC_RIGHT (chrec) != op1)
2109 op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL);
2110 chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1);
2112 return chrec;
2114 case POINTER_PLUS_EXPR:
2115 case PLUS_EXPR:
2116 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2117 TREE_OPERAND (chrec, 0), fold_conversions, cache,
2118 size_expr);
2119 if (op0 == chrec_dont_know)
2120 return chrec_dont_know;
2122 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2123 TREE_OPERAND (chrec, 1), fold_conversions, cache,
2124 size_expr);
2125 if (op1 == chrec_dont_know)
2126 return chrec_dont_know;
2128 if (TREE_OPERAND (chrec, 0) != op0
2129 || TREE_OPERAND (chrec, 1) != op1)
2131 op0 = chrec_convert (type, op0, NULL);
2132 op1 = chrec_convert_rhs (type, op1, NULL);
2133 chrec = chrec_fold_plus (type, op0, op1);
2135 return chrec;
2137 case MINUS_EXPR:
2138 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2139 TREE_OPERAND (chrec, 0), fold_conversions, cache,
2140 size_expr);
2141 if (op0 == chrec_dont_know)
2142 return chrec_dont_know;
2144 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2145 TREE_OPERAND (chrec, 1),
2146 fold_conversions, cache, size_expr);
2147 if (op1 == chrec_dont_know)
2148 return chrec_dont_know;
2150 if (TREE_OPERAND (chrec, 0) != op0
2151 || TREE_OPERAND (chrec, 1) != op1)
2153 op0 = chrec_convert (type, op0, NULL);
2154 op1 = chrec_convert (type, op1, NULL);
2155 chrec = chrec_fold_minus (type, op0, op1);
2157 return chrec;
2159 case MULT_EXPR:
2160 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2161 TREE_OPERAND (chrec, 0),
2162 fold_conversions, cache, size_expr);
2163 if (op0 == chrec_dont_know)
2164 return chrec_dont_know;
2166 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2167 TREE_OPERAND (chrec, 1),
2168 fold_conversions, cache, size_expr);
2169 if (op1 == chrec_dont_know)
2170 return chrec_dont_know;
2172 if (TREE_OPERAND (chrec, 0) != op0
2173 || TREE_OPERAND (chrec, 1) != op1)
2175 op0 = chrec_convert (type, op0, NULL);
2176 op1 = chrec_convert (type, op1, NULL);
2177 chrec = chrec_fold_multiply (type, op0, op1);
2179 return chrec;
2181 CASE_CONVERT:
2182 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2183 TREE_OPERAND (chrec, 0),
2184 fold_conversions, cache, size_expr);
2185 if (op0 == chrec_dont_know)
2186 return chrec_dont_know;
2188 if (fold_conversions)
2190 tree tmp = chrec_convert_aggressive (TREE_TYPE (chrec), op0);
2191 if (tmp)
2192 return tmp;
2195 if (op0 == TREE_OPERAND (chrec, 0))
2196 return chrec;
2198 /* If we used chrec_convert_aggressive, we can no longer assume that
2199 signed chrecs do not overflow, as chrec_convert does, so avoid
2200 calling it in that case. */
2201 if (fold_conversions)
2202 return fold_convert (TREE_TYPE (chrec), op0);
2204 return chrec_convert (TREE_TYPE (chrec), op0, NULL);
2206 case SCEV_NOT_KNOWN:
2207 return chrec_dont_know;
2209 case SCEV_KNOWN:
2210 return chrec_known;
2212 default:
2213 break;
2216 gcc_assert (!VL_EXP_CLASS_P (chrec));
2217 switch (TREE_CODE_LENGTH (TREE_CODE (chrec)))
2219 case 3:
2220 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2221 TREE_OPERAND (chrec, 0),
2222 fold_conversions, cache, size_expr);
2223 if (op0 == chrec_dont_know)
2224 return chrec_dont_know;
2226 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2227 TREE_OPERAND (chrec, 1),
2228 fold_conversions, cache, size_expr);
2229 if (op1 == chrec_dont_know)
2230 return chrec_dont_know;
2232 op2 = instantiate_scev_1 (instantiate_below, evolution_loop,
2233 TREE_OPERAND (chrec, 2),
2234 fold_conversions, cache, size_expr);
2235 if (op2 == chrec_dont_know)
2236 return chrec_dont_know;
2238 if (op0 == TREE_OPERAND (chrec, 0)
2239 && op1 == TREE_OPERAND (chrec, 1)
2240 && op2 == TREE_OPERAND (chrec, 2))
2241 return chrec;
2243 return fold_build3 (TREE_CODE (chrec),
2244 TREE_TYPE (chrec), op0, op1, op2);
2246 case 2:
2247 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2248 TREE_OPERAND (chrec, 0),
2249 fold_conversions, cache, size_expr);
2250 if (op0 == chrec_dont_know)
2251 return chrec_dont_know;
2253 op1 = instantiate_scev_1 (instantiate_below, evolution_loop,
2254 TREE_OPERAND (chrec, 1),
2255 fold_conversions, cache, size_expr);
2256 if (op1 == chrec_dont_know)
2257 return chrec_dont_know;
2259 if (op0 == TREE_OPERAND (chrec, 0)
2260 && op1 == TREE_OPERAND (chrec, 1))
2261 return chrec;
2262 return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1);
2264 case 1:
2265 op0 = instantiate_scev_1 (instantiate_below, evolution_loop,
2266 TREE_OPERAND (chrec, 0),
2267 fold_conversions, cache, size_expr);
2268 if (op0 == chrec_dont_know)
2269 return chrec_dont_know;
2270 if (op0 == TREE_OPERAND (chrec, 0))
2271 return chrec;
2272 return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0);
2274 case 0:
2275 return chrec;
2277 default:
2278 break;
2281 /* Too complicated to handle. */
2282 return chrec_dont_know;
2285 /* Analyze all the parameters of the chrec that were left under a
2286 symbolic form. INSTANTIATE_BELOW is the basic block that stops the
2287 recursive instantiation of parameters: a parameter is a variable
2288 that is defined in a basic block that dominates INSTANTIATE_BELOW or
2289 a function parameter. */
2291 tree
2292 instantiate_scev (basic_block instantiate_below, struct loop *evolution_loop,
2293 tree chrec)
2295 tree res;
2296 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2298 if (dump_file && (dump_flags & TDF_DETAILS))
2300 fprintf (dump_file, "(instantiate_scev \n");
2301 fprintf (dump_file, " (instantiate_below = %d)\n", instantiate_below->index);
2302 fprintf (dump_file, " (evolution_loop = %d)\n", evolution_loop->num);
2303 fprintf (dump_file, " (chrec = ");
2304 print_generic_expr (dump_file, chrec, 0);
2305 fprintf (dump_file, ")\n");
2308 res = instantiate_scev_1 (instantiate_below, evolution_loop, chrec, false,
2309 cache, 0);
2311 if (dump_file && (dump_flags & TDF_DETAILS))
2313 fprintf (dump_file, " (res = ");
2314 print_generic_expr (dump_file, res, 0);
2315 fprintf (dump_file, "))\n");
2318 htab_delete (cache);
2320 return res;
2323 /* Similar to instantiate_parameters, but does not introduce the
2324 evolutions in outer loops for LOOP invariants in CHREC, and does not
2325 care about causing overflows, as long as they do not affect value
2326 of an expression. */
2328 tree
2329 resolve_mixers (struct loop *loop, tree chrec)
2331 htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info);
2332 tree ret = instantiate_scev_1 (block_before_loop (loop), loop, chrec, true,
2333 cache, 0);
2334 htab_delete (cache);
2335 return ret;
2338 /* Entry point for the analysis of the number of iterations pass.
2339 This function tries to safely approximate the number of iterations
2340 the loop will run. When this property is not decidable at compile
2341 time, the result is chrec_dont_know. Otherwise the result is
2342 a scalar or a symbolic parameter.
2344 Example of analysis: suppose that the loop has an exit condition:
2346 "if (b > 49) goto end_loop;"
2348 and that in a previous analysis we have determined that the
2349 variable 'b' has an evolution function:
2351 "EF = {23, +, 5}_2".
2353 When we evaluate the function at the point 5, i.e. the value of the
2354 variable 'b' after 5 iterations in the loop, we have EF (5) = 48,
2355 and EF (6) = 53. In this case the value of 'b' on exit is '53' and
2356 the loop body has been executed 6 times. */
2358 tree
2359 number_of_latch_executions (struct loop *loop)
2361 tree res, type;
2362 edge exit;
2363 struct tree_niter_desc niter_desc;
2365 /* Determine whether the number_of_iterations_in_loop has already
2366 been computed. */
2367 res = loop->nb_iterations;
2368 if (res)
2369 return res;
2370 res = chrec_dont_know;
2372 if (dump_file && (dump_flags & TDF_DETAILS))
2373 fprintf (dump_file, "(number_of_iterations_in_loop\n");
2375 exit = single_exit (loop);
2376 if (!exit)
2377 goto end;
2379 if (!number_of_iterations_exit (loop, exit, &niter_desc, false))
2380 goto end;
2382 type = TREE_TYPE (niter_desc.niter);
2383 if (integer_nonzerop (niter_desc.may_be_zero))
2384 res = build_int_cst (type, 0);
2385 else if (integer_zerop (niter_desc.may_be_zero))
2386 res = niter_desc.niter;
2387 else
2388 res = chrec_dont_know;
2390 end:
2391 return set_nb_iterations_in_loop (loop, res);
2394 /* Returns the number of executions of the exit condition of LOOP,
2395 i.e., the number by one higher than number_of_latch_executions.
2396 Note that unlike number_of_latch_executions, this number does
2397 not necessarily fit in the unsigned variant of the type of
2398 the control variable -- if the number of iterations is a constant,
2399 we return chrec_dont_know if adding one to number_of_latch_executions
2400 overflows; however, in case the number of iterations is symbolic
2401 expression, the caller is responsible for dealing with this
2402 the possible overflow. */
2404 tree
2405 number_of_exit_cond_executions (struct loop *loop)
2407 tree ret = number_of_latch_executions (loop);
2408 tree type = chrec_type (ret);
2410 if (chrec_contains_undetermined (ret))
2411 return ret;
2413 ret = chrec_fold_plus (type, ret, build_int_cst (type, 1));
2414 if (TREE_CODE (ret) == INTEGER_CST
2415 && TREE_OVERFLOW (ret))
2416 return chrec_dont_know;
2418 return ret;
2421 /* One of the drivers for testing the scalar evolutions analysis.
2422 This function computes the number of iterations for all the loops
2423 from the EXIT_CONDITIONS array. */
2425 static void
2426 number_of_iterations_for_all_loops (VEC(gimple,heap) **exit_conditions)
2428 unsigned int i;
2429 unsigned nb_chrec_dont_know_loops = 0;
2430 unsigned nb_static_loops = 0;
2431 gimple cond;
2433 for (i = 0; VEC_iterate (gimple, *exit_conditions, i, cond); i++)
2435 tree res = number_of_latch_executions (loop_containing_stmt (cond));
2436 if (chrec_contains_undetermined (res))
2437 nb_chrec_dont_know_loops++;
2438 else
2439 nb_static_loops++;
2442 if (dump_file)
2444 fprintf (dump_file, "\n(\n");
2445 fprintf (dump_file, "-----------------------------------------\n");
2446 fprintf (dump_file, "%d\tnb_chrec_dont_know_loops\n", nb_chrec_dont_know_loops);
2447 fprintf (dump_file, "%d\tnb_static_loops\n", nb_static_loops);
2448 fprintf (dump_file, "%d\tnb_total_loops\n", number_of_loops ());
2449 fprintf (dump_file, "-----------------------------------------\n");
2450 fprintf (dump_file, ")\n\n");
2452 print_loops (dump_file, 3);
2458 /* Counters for the stats. */
2460 struct chrec_stats
2462 unsigned nb_chrecs;
2463 unsigned nb_affine;
2464 unsigned nb_affine_multivar;
2465 unsigned nb_higher_poly;
2466 unsigned nb_chrec_dont_know;
2467 unsigned nb_undetermined;
2470 /* Reset the counters. */
2472 static inline void
2473 reset_chrecs_counters (struct chrec_stats *stats)
2475 stats->nb_chrecs = 0;
2476 stats->nb_affine = 0;
2477 stats->nb_affine_multivar = 0;
2478 stats->nb_higher_poly = 0;
2479 stats->nb_chrec_dont_know = 0;
2480 stats->nb_undetermined = 0;
2483 /* Dump the contents of a CHREC_STATS structure. */
2485 static void
2486 dump_chrecs_stats (FILE *file, struct chrec_stats *stats)
2488 fprintf (file, "\n(\n");
2489 fprintf (file, "-----------------------------------------\n");
2490 fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine);
2491 fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar);
2492 fprintf (file, "%d\tdegree greater than 2 polynomials\n",
2493 stats->nb_higher_poly);
2494 fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know);
2495 fprintf (file, "-----------------------------------------\n");
2496 fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs);
2497 fprintf (file, "%d\twith undetermined coefficients\n",
2498 stats->nb_undetermined);
2499 fprintf (file, "-----------------------------------------\n");
2500 fprintf (file, "%d\tchrecs in the scev database\n",
2501 (int) htab_elements (scalar_evolution_info));
2502 fprintf (file, "%d\tsets in the scev database\n", nb_set_scev);
2503 fprintf (file, "%d\tgets in the scev database\n", nb_get_scev);
2504 fprintf (file, "-----------------------------------------\n");
2505 fprintf (file, ")\n\n");
2508 /* Gather statistics about CHREC. */
2510 static void
2511 gather_chrec_stats (tree chrec, struct chrec_stats *stats)
2513 if (dump_file && (dump_flags & TDF_STATS))
2515 fprintf (dump_file, "(classify_chrec ");
2516 print_generic_expr (dump_file, chrec, 0);
2517 fprintf (dump_file, "\n");
2520 stats->nb_chrecs++;
2522 if (chrec == NULL_TREE)
2524 stats->nb_undetermined++;
2525 return;
2528 switch (TREE_CODE (chrec))
2530 case POLYNOMIAL_CHREC:
2531 if (evolution_function_is_affine_p (chrec))
2533 if (dump_file && (dump_flags & TDF_STATS))
2534 fprintf (dump_file, " affine_univariate\n");
2535 stats->nb_affine++;
2537 else if (evolution_function_is_affine_multivariate_p (chrec, 0))
2539 if (dump_file && (dump_flags & TDF_STATS))
2540 fprintf (dump_file, " affine_multivariate\n");
2541 stats->nb_affine_multivar++;
2543 else
2545 if (dump_file && (dump_flags & TDF_STATS))
2546 fprintf (dump_file, " higher_degree_polynomial\n");
2547 stats->nb_higher_poly++;
2550 break;
2552 default:
2553 break;
2556 if (chrec_contains_undetermined (chrec))
2558 if (dump_file && (dump_flags & TDF_STATS))
2559 fprintf (dump_file, " undetermined\n");
2560 stats->nb_undetermined++;
2563 if (dump_file && (dump_flags & TDF_STATS))
2564 fprintf (dump_file, ")\n");
2567 /* One of the drivers for testing the scalar evolutions analysis.
2568 This function analyzes the scalar evolution of all the scalars
2569 defined as loop phi nodes in one of the loops from the
2570 EXIT_CONDITIONS array.
2572 TODO Optimization: A loop is in canonical form if it contains only
2573 a single scalar loop phi node. All the other scalars that have an
2574 evolution in the loop are rewritten in function of this single
2575 index. This allows the parallelization of the loop. */
2577 static void
2578 analyze_scalar_evolution_for_all_loop_phi_nodes (VEC(gimple,heap) **exit_conditions)
2580 unsigned int i;
2581 struct chrec_stats stats;
2582 gimple cond, phi;
2583 gimple_stmt_iterator psi;
2585 reset_chrecs_counters (&stats);
2587 for (i = 0; VEC_iterate (gimple, *exit_conditions, i, cond); i++)
2589 struct loop *loop;
2590 basic_block bb;
2591 tree chrec;
2593 loop = loop_containing_stmt (cond);
2594 bb = loop->header;
2596 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
2598 phi = gsi_stmt (psi);
2599 if (is_gimple_reg (PHI_RESULT (phi)))
2601 chrec = instantiate_parameters
2602 (loop,
2603 analyze_scalar_evolution (loop, PHI_RESULT (phi)));
2605 if (dump_file && (dump_flags & TDF_STATS))
2606 gather_chrec_stats (chrec, &stats);
2611 if (dump_file && (dump_flags & TDF_STATS))
2612 dump_chrecs_stats (dump_file, &stats);
2615 /* Callback for htab_traverse, gathers information on chrecs in the
2616 hashtable. */
2618 static int
2619 gather_stats_on_scev_database_1 (void **slot, void *stats)
2621 struct scev_info_str *entry = (struct scev_info_str *) *slot;
2623 gather_chrec_stats (entry->chrec, (struct chrec_stats *) stats);
2625 return 1;
2628 /* Classify the chrecs of the whole database. */
2630 void
2631 gather_stats_on_scev_database (void)
2633 struct chrec_stats stats;
2635 if (!dump_file)
2636 return;
2638 reset_chrecs_counters (&stats);
2640 htab_traverse (scalar_evolution_info, gather_stats_on_scev_database_1,
2641 &stats);
2643 dump_chrecs_stats (dump_file, &stats);
2648 /* Initializer. */
2650 static void
2651 initialize_scalar_evolutions_analyzer (void)
2653 /* The elements below are unique. */
2654 if (chrec_dont_know == NULL_TREE)
2656 chrec_not_analyzed_yet = NULL_TREE;
2657 chrec_dont_know = make_node (SCEV_NOT_KNOWN);
2658 chrec_known = make_node (SCEV_KNOWN);
2659 TREE_TYPE (chrec_dont_know) = void_type_node;
2660 TREE_TYPE (chrec_known) = void_type_node;
2664 /* Initialize the analysis of scalar evolutions for LOOPS. */
2666 void
2667 scev_initialize (void)
2669 loop_iterator li;
2670 struct loop *loop;
2672 scalar_evolution_info = htab_create_alloc (100,
2673 hash_scev_info,
2674 eq_scev_info,
2675 del_scev_info,
2676 ggc_calloc,
2677 ggc_free);
2679 initialize_scalar_evolutions_analyzer ();
2681 FOR_EACH_LOOP (li, loop, 0)
2683 loop->nb_iterations = NULL_TREE;
2687 /* Cleans up the information cached by the scalar evolutions analysis. */
2689 void
2690 scev_reset (void)
2692 loop_iterator li;
2693 struct loop *loop;
2695 if (!scalar_evolution_info || !current_loops)
2696 return;
2698 htab_empty (scalar_evolution_info);
2699 FOR_EACH_LOOP (li, loop, 0)
2701 loop->nb_iterations = NULL_TREE;
2705 /* Checks whether OP behaves as a simple affine iv of LOOP in STMT and returns
2706 its base and step in IV if possible. If ALLOW_NONCONSTANT_STEP is true, we
2707 want step to be invariant in LOOP. Otherwise we require it to be an
2708 integer constant. IV->no_overflow is set to true if we are sure the iv cannot
2709 overflow (e.g. because it is computed in signed arithmetics). */
2711 bool
2712 simple_iv (struct loop *loop, gimple stmt, tree op, affine_iv *iv,
2713 bool allow_nonconstant_step)
2715 basic_block bb = gimple_bb (stmt);
2716 tree type, ev;
2717 bool folded_casts;
2719 iv->base = NULL_TREE;
2720 iv->step = NULL_TREE;
2721 iv->no_overflow = false;
2723 type = TREE_TYPE (op);
2724 if (TREE_CODE (type) != INTEGER_TYPE
2725 && TREE_CODE (type) != POINTER_TYPE)
2726 return false;
2728 ev = analyze_scalar_evolution_in_loop (loop, bb->loop_father, op,
2729 &folded_casts);
2730 if (chrec_contains_undetermined (ev))
2731 return false;
2733 if (tree_does_not_contain_chrecs (ev)
2734 && !chrec_contains_symbols_defined_in_loop (ev, loop->num))
2736 iv->base = ev;
2737 iv->step = build_int_cst (TREE_TYPE (ev), 0);
2738 iv->no_overflow = true;
2739 return true;
2742 if (TREE_CODE (ev) != POLYNOMIAL_CHREC
2743 || CHREC_VARIABLE (ev) != (unsigned) loop->num)
2744 return false;
2746 iv->step = CHREC_RIGHT (ev);
2747 if (allow_nonconstant_step)
2749 if (tree_contains_chrecs (iv->step, NULL)
2750 || chrec_contains_symbols_defined_in_loop (iv->step, loop->num))
2751 return false;
2753 else if (TREE_CODE (iv->step) != INTEGER_CST)
2754 return false;
2756 iv->base = CHREC_LEFT (ev);
2757 if (tree_contains_chrecs (iv->base, NULL)
2758 || chrec_contains_symbols_defined_in_loop (iv->base, loop->num))
2759 return false;
2761 iv->no_overflow = !folded_casts && TYPE_OVERFLOW_UNDEFINED (type);
2763 return true;
2766 /* Runs the analysis of scalar evolutions. */
2768 void
2769 scev_analysis (void)
2771 VEC(gimple,heap) *exit_conditions;
2773 exit_conditions = VEC_alloc (gimple, heap, 37);
2774 select_loops_exit_conditions (&exit_conditions);
2776 if (dump_file && (dump_flags & TDF_STATS))
2777 analyze_scalar_evolution_for_all_loop_phi_nodes (&exit_conditions);
2779 number_of_iterations_for_all_loops (&exit_conditions);
2780 VEC_free (gimple, heap, exit_conditions);
2783 /* Finalize the scalar evolution analysis. */
2785 void
2786 scev_finalize (void)
2788 if (!scalar_evolution_info)
2789 return;
2790 htab_delete (scalar_evolution_info);
2791 scalar_evolution_info = NULL;
2794 /* Replace ssa names for that scev can prove they are constant by the
2795 appropriate constants. Also perform final value replacement in loops,
2796 in case the replacement expressions are cheap.
2798 We only consider SSA names defined by phi nodes; rest is left to the
2799 ordinary constant propagation pass. */
2801 unsigned int
2802 scev_const_prop (void)
2804 basic_block bb;
2805 tree name, type, ev;
2806 gimple phi, ass;
2807 struct loop *loop, *ex_loop;
2808 bitmap ssa_names_to_remove = NULL;
2809 unsigned i;
2810 loop_iterator li;
2811 gimple_stmt_iterator psi;
2813 if (number_of_loops () <= 1)
2814 return 0;
2816 FOR_EACH_BB (bb)
2818 loop = bb->loop_father;
2820 for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi))
2822 phi = gsi_stmt (psi);
2823 name = PHI_RESULT (phi);
2825 if (!is_gimple_reg (name))
2826 continue;
2828 type = TREE_TYPE (name);
2830 if (!POINTER_TYPE_P (type)
2831 && !INTEGRAL_TYPE_P (type))
2832 continue;
2834 ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name));
2835 if (!is_gimple_min_invariant (ev)
2836 || !may_propagate_copy (name, ev))
2837 continue;
2839 /* Replace the uses of the name. */
2840 if (name != ev)
2841 replace_uses_by (name, ev);
2843 if (!ssa_names_to_remove)
2844 ssa_names_to_remove = BITMAP_ALLOC (NULL);
2845 bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name));
2849 /* Remove the ssa names that were replaced by constants. We do not
2850 remove them directly in the previous cycle, since this
2851 invalidates scev cache. */
2852 if (ssa_names_to_remove)
2854 bitmap_iterator bi;
2856 EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi)
2858 gimple_stmt_iterator psi;
2859 name = ssa_name (i);
2860 phi = SSA_NAME_DEF_STMT (name);
2862 gcc_assert (gimple_code (phi) == GIMPLE_PHI);
2863 psi = gsi_for_stmt (phi);
2864 remove_phi_node (&psi, true);
2867 BITMAP_FREE (ssa_names_to_remove);
2868 scev_reset ();
2871 /* Now the regular final value replacement. */
2872 FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST)
2874 edge exit;
2875 tree def, rslt, niter;
2876 gimple_stmt_iterator bsi;
2878 /* If we do not know exact number of iterations of the loop, we cannot
2879 replace the final value. */
2880 exit = single_exit (loop);
2881 if (!exit)
2882 continue;
2884 niter = number_of_latch_executions (loop);
2885 /* We used to check here whether the computation of NITER is expensive,
2886 and avoided final value elimination if that is the case. The problem
2887 is that it is hard to evaluate whether the expression is too
2888 expensive, as we do not know what optimization opportunities the
2889 elimination of the final value may reveal. Therefore, we now
2890 eliminate the final values of induction variables unconditionally. */
2891 if (niter == chrec_dont_know)
2892 continue;
2894 /* Ensure that it is possible to insert new statements somewhere. */
2895 if (!single_pred_p (exit->dest))
2896 split_loop_exit_edge (exit);
2897 bsi = gsi_after_labels (exit->dest);
2899 ex_loop = superloop_at_depth (loop,
2900 loop_depth (exit->dest->loop_father) + 1);
2902 for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); )
2904 phi = gsi_stmt (psi);
2905 rslt = PHI_RESULT (phi);
2906 def = PHI_ARG_DEF_FROM_EDGE (phi, exit);
2907 if (!is_gimple_reg (def))
2909 gsi_next (&psi);
2910 continue;
2913 if (!POINTER_TYPE_P (TREE_TYPE (def))
2914 && !INTEGRAL_TYPE_P (TREE_TYPE (def)))
2916 gsi_next (&psi);
2917 continue;
2920 def = analyze_scalar_evolution_in_loop (ex_loop, loop, def, NULL);
2921 def = compute_overall_effect_of_inner_loop (ex_loop, def);
2922 if (!tree_does_not_contain_chrecs (def)
2923 || chrec_contains_symbols_defined_in_loop (def, ex_loop->num)
2924 /* Moving the computation from the loop may prolong life range
2925 of some ssa names, which may cause problems if they appear
2926 on abnormal edges. */
2927 || contains_abnormal_ssa_name_p (def))
2929 gsi_next (&psi);
2930 continue;
2933 /* Eliminate the PHI node and replace it by a computation outside
2934 the loop. */
2935 def = unshare_expr (def);
2936 remove_phi_node (&psi, false);
2938 def = force_gimple_operand_gsi (&bsi, def, false, NULL_TREE,
2939 true, GSI_SAME_STMT);
2940 ass = gimple_build_assign (rslt, def);
2941 gsi_insert_before (&bsi, ass, GSI_SAME_STMT);
2944 return 0;
2947 #include "gt-tree-scalar-evolution.h"