1 | /* Scalar evolution detector. |
2 | Copyright (C) 2003-2017 Free Software Foundation, Inc. |
3 | Contributed by Sebastian Pop <s.pop@laposte.net> |
4 | |
5 | This file is part of GCC. |
6 | |
7 | GCC is free software; you can redistribute it and/or modify it under |
8 | the terms of the GNU General Public License as published by the Free |
9 | Software Foundation; either version 3, or (at your option) any later |
10 | version. |
11 | |
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
15 | for more details. |
16 | |
17 | You should have received a copy of the GNU General Public License |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ |
20 | |
21 | /* |
22 | Description: |
23 | |
24 | This pass analyzes the evolution of scalar variables in loop |
25 | structures. The algorithm is based on the SSA representation, |
26 | and on the loop hierarchy tree. This algorithm is not based on |
27 | the notion of versions of a variable, as it was the case for the |
28 | previous implementations of the scalar evolution algorithm, but |
29 | it assumes that each defined name is unique. |
30 | |
31 | The notation used in this file is called "chains of recurrences", |
32 | and has been proposed by Eugene Zima, Robert Van Engelen, and |
33 | others for describing induction variables in programs. For example |
34 | "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0 |
35 | when entering in the loop_1 and has a step 2 in this loop, in other |
36 | words "for (b = 0; b < N; b+=2);". Note that the coefficients of |
37 | this chain of recurrence (or chrec [shrek]) can contain the name of |
38 | other variables, in which case they are called parametric chrecs. |
39 | For example, "b -> {a, +, 2}_1" means that the initial value of "b" |
40 | is the value of "a". In most of the cases these parametric chrecs |
41 | are fully instantiated before their use because symbolic names can |
42 | hide some difficult cases such as self-references described later |
43 | (see the Fibonacci example). |
44 | |
45 | A short sketch of the algorithm is: |
46 | |
47 | Given a scalar variable to be analyzed, follow the SSA edge to |
48 | its definition: |
49 | |
50 | - When the definition is a GIMPLE_ASSIGN: if the right hand side |
51 | (RHS) of the definition cannot be statically analyzed, the answer |
52 | of the analyzer is: "don't know". |
53 | Otherwise, for all the variables that are not yet analyzed in the |
54 | RHS, try to determine their evolution, and finally try to |
55 | evaluate the operation of the RHS that gives the evolution |
56 | function of the analyzed variable. |
57 | |
58 | - When the definition is a condition-phi-node: determine the |
59 | evolution function for all the branches of the phi node, and |
60 | finally merge these evolutions (see chrec_merge). |
61 | |
62 | - When the definition is a loop-phi-node: determine its initial |
63 | condition, that is the SSA edge defined in an outer loop, and |
64 | keep it symbolic. Then determine the SSA edges that are defined |
65 | in the body of the loop. Follow the inner edges until ending on |
66 | another loop-phi-node of the same analyzed loop. If the reached |
67 | loop-phi-node is not the starting loop-phi-node, then we keep |
68 | this definition under a symbolic form. If the reached |
69 | loop-phi-node is the same as the starting one, then we compute a |
70 | symbolic stride on the return path. The result is then the |
71 | symbolic chrec {initial_condition, +, symbolic_stride}_loop. |
72 | |
73 | Examples: |
74 | |
75 | Example 1: Illustration of the basic algorithm. |
76 | |
77 | | a = 3 |
78 | | loop_1 |
79 | | b = phi (a, c) |
80 | | c = b + 1 |
81 | | if (c > 10) exit_loop |
82 | | endloop |
83 | |
84 | Suppose that we want to know the number of iterations of the |
85 | loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We |
86 | ask the scalar evolution analyzer two questions: what's the |
87 | scalar evolution (scev) of "c", and what's the scev of "10". For |
88 | "10" the answer is "10" since it is a scalar constant. For the |
89 | scalar variable "c", it follows the SSA edge to its definition, |
90 | "c = b + 1", and then asks again what's the scev of "b". |
91 | Following the SSA edge, we end on a loop-phi-node "b = phi (a, |
92 | c)", where the initial condition is "a", and the inner loop edge |
93 | is "c". The initial condition is kept under a symbolic form (it |
94 | may be the case that the copy constant propagation has done its |
95 | work and we end with the constant "3" as one of the edges of the |
96 | loop-phi-node). The update edge is followed to the end of the |
97 | loop, and until reaching again the starting loop-phi-node: b -> c |
98 | -> b. At this point we have drawn a path from "b" to "b" from |
99 | which we compute the stride in the loop: in this example it is |
100 | "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now |
101 | that the scev for "b" is known, it is possible to compute the |
102 | scev for "c", that is "c -> {a + 1, +, 1}_1". In order to |
103 | determine the number of iterations in the loop_1, we have to |
104 | instantiate_parameters (loop_1, {a + 1, +, 1}_1), that gives after some |
105 | more analysis the scev {4, +, 1}_1, or in other words, this is |
106 | the function "f (x) = x + 4", where x is the iteration count of |
107 | the loop_1. Now we have to solve the inequality "x + 4 > 10", |
108 | and take the smallest iteration number for which the loop is |
109 | exited: x = 7. This loop runs from x = 0 to x = 7, and in total |
110 | there are 8 iterations. In terms of loop normalization, we have |
111 | created a variable that is implicitly defined, "x" or just "_1", |
112 | and all the other analyzed scalars of the loop are defined in |
113 | function of this variable: |
114 | |
115 | a -> 3 |
116 | b -> {3, +, 1}_1 |
117 | c -> {4, +, 1}_1 |
118 | |
119 | or in terms of a C program: |
120 | |
121 | | a = 3 |
122 | | for (x = 0; x <= 7; x++) |
123 | | { |
124 | | b = x + 3 |
125 | | c = x + 4 |
126 | | } |
127 | |
128 | Example 2a: Illustration of the algorithm on nested loops. |
129 | |
130 | | loop_1 |
131 | | a = phi (1, b) |
132 | | c = a + 2 |
133 | | loop_2 10 times |
134 | | b = phi (c, d) |
135 | | d = b + 3 |
136 | | endloop |
137 | | endloop |
138 | |
139 | For analyzing the scalar evolution of "a", the algorithm follows |
140 | the SSA edge into the loop's body: "a -> b". "b" is an inner |
141 | loop-phi-node, and its analysis as in Example 1, gives: |
142 | |
143 | b -> {c, +, 3}_2 |
144 | d -> {c + 3, +, 3}_2 |
145 | |
146 | Following the SSA edge for the initial condition, we end on "c = a |
147 | + 2", and then on the starting loop-phi-node "a". From this point, |
148 | the loop stride is computed: back on "c = a + 2" we get a "+2" in |
149 | the loop_1, then on the loop-phi-node "b" we compute the overall |
150 | effect of the inner loop that is "b = c + 30", and we get a "+30" |
151 | in the loop_1. That means that the overall stride in loop_1 is |
152 | equal to "+32", and the result is: |
153 | |
154 | a -> {1, +, 32}_1 |
155 | c -> {3, +, 32}_1 |
156 | |
157 | Example 2b: Multivariate chains of recurrences. |
158 | |
159 | | loop_1 |
160 | | k = phi (0, k + 1) |
161 | | loop_2 4 times |
162 | | j = phi (0, j + 1) |
163 | | loop_3 4 times |
164 | | i = phi (0, i + 1) |
165 | | A[j + k] = ... |
166 | | endloop |
167 | | endloop |
168 | | endloop |
169 | |
170 | Analyzing the access function of array A with |
171 | instantiate_parameters (loop_1, "j + k"), we obtain the |
172 | instantiation and the analysis of the scalar variables "j" and "k" |
173 | in loop_1. This leads to the scalar evolution {4, +, 1}_1: the end |
174 | value of loop_2 for "j" is 4, and the evolution of "k" in loop_1 is |
175 | {0, +, 1}_1. To obtain the evolution function in loop_3 and |
176 | instantiate the scalar variables up to loop_1, one has to use: |
177 | instantiate_scev (block_before_loop (loop_1), loop_3, "j + k"). |
178 | The result of this call is {{0, +, 1}_1, +, 1}_2. |
179 | |
180 | Example 3: Higher degree polynomials. |
181 | |
182 | | loop_1 |
183 | | a = phi (2, b) |
184 | | c = phi (5, d) |
185 | | b = a + 1 |
186 | | d = c + a |
187 | | endloop |
188 | |
189 | a -> {2, +, 1}_1 |
190 | b -> {3, +, 1}_1 |
191 | c -> {5, +, a}_1 |
192 | d -> {5 + a, +, a}_1 |
193 | |
194 | instantiate_parameters (loop_1, {5, +, a}_1) -> {5, +, 2, +, 1}_1 |
195 | instantiate_parameters (loop_1, {5 + a, +, a}_1) -> {7, +, 3, +, 1}_1 |
196 | |
197 | Example 4: Lucas, Fibonacci, or mixers in general. |
198 | |
199 | | loop_1 |
200 | | a = phi (1, b) |
201 | | c = phi (3, d) |
202 | | b = c |
203 | | d = c + a |
204 | | endloop |
205 | |
206 | a -> (1, c)_1 |
207 | c -> {3, +, a}_1 |
208 | |
209 | The syntax "(1, c)_1" stands for a PEELED_CHREC that has the |
210 | following semantics: during the first iteration of the loop_1, the |
211 | variable contains the value 1, and then it contains the value "c". |
212 | Note that this syntax is close to the syntax of the loop-phi-node: |
213 | "a -> (1, c)_1" vs. "a = phi (1, c)". |
214 | |
215 | The symbolic chrec representation contains all the semantics of the |
216 | original code. What is more difficult is to use this information. |
217 | |
218 | Example 5: Flip-flops, or exchangers. |
219 | |
220 | | loop_1 |
221 | | a = phi (1, b) |
222 | | c = phi (3, d) |
223 | | b = c |
224 | | d = a |
225 | | endloop |
226 | |
227 | a -> (1, c)_1 |
228 | c -> (3, a)_1 |
229 | |
230 | Based on these symbolic chrecs, it is possible to refine this |
231 | information into the more precise PERIODIC_CHRECs: |
232 | |
233 | a -> |1, 3|_1 |
234 | c -> |3, 1|_1 |
235 | |
236 | This transformation is not yet implemented. |
237 | |
238 | Further readings: |
239 | |
240 | You can find a more detailed description of the algorithm in: |
241 | http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf |
242 | http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that |
243 | this is a preliminary report and some of the details of the |
244 | algorithm have changed. I'm working on a research report that |
245 | updates the description of the algorithms to reflect the design |
246 | choices used in this implementation. |
247 | |
248 | A set of slides show a high level overview of the algorithm and run |
249 | an example through the scalar evolution analyzer: |
250 | http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf |
251 | |
252 | The slides that I have presented at the GCC Summit'04 are available |
253 | at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf |
254 | */ |
255 | |
256 | #include "config.h" |
257 | #include "system.h" |
258 | #include "coretypes.h" |
259 | #include "backend.h" |
260 | #include "rtl.h" |
261 | #include "tree.h" |
262 | #include "gimple.h" |
263 | #include "ssa.h" |
264 | #include "gimple-pretty-print.h" |
265 | #include "fold-const.h" |
266 | #include "gimplify.h" |
267 | #include "gimple-iterator.h" |
268 | #include "gimplify-me.h" |
269 | #include "tree-cfg.h" |
270 | #include "tree-ssa-loop-ivopts.h" |
271 | #include "tree-ssa-loop-manip.h" |
272 | #include "tree-ssa-loop-niter.h" |
273 | #include "tree-ssa-loop.h" |
274 | #include "tree-ssa.h" |
275 | #include "cfgloop.h" |
276 | #include "tree-chrec.h" |
277 | #include "tree-affine.h" |
278 | #include "tree-scalar-evolution.h" |
279 | #include "dumpfile.h" |
280 | #include "params.h" |
281 | #include "tree-ssa-propagate.h" |
282 | #include "gimple-fold.h" |
283 | |
284 | static tree analyze_scalar_evolution_1 (struct loop *, tree); |
285 | static tree analyze_scalar_evolution_for_address_of (struct loop *loop, |
286 | tree var); |
287 | |
288 | /* The cached information about an SSA name with version NAME_VERSION, |
289 | claiming that below basic block with index INSTANTIATED_BELOW, the |
290 | value of the SSA name can be expressed as CHREC. */ |
291 | |
292 | struct GTY((for_user)) scev_info_str { |
293 | unsigned int name_version; |
294 | int instantiated_below; |
295 | tree chrec; |
296 | }; |
297 | |
298 | /* Counters for the scev database. */ |
299 | static unsigned nb_set_scev = 0; |
300 | static unsigned nb_get_scev = 0; |
301 | |
302 | /* The following trees are unique elements. Thus the comparison of |
303 | another element to these elements should be done on the pointer to |
304 | these trees, and not on their value. */ |
305 | |
306 | /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */ |
307 | tree chrec_not_analyzed_yet; |
308 | |
309 | /* Reserved to the cases where the analyzer has detected an |
310 | undecidable property at compile time. */ |
311 | tree chrec_dont_know; |
312 | |
313 | /* When the analyzer has detected that a property will never |
314 | happen, then it qualifies it with chrec_known. */ |
315 | tree chrec_known; |
316 | |
317 | struct scev_info_hasher : ggc_ptr_hash<scev_info_str> |
318 | { |
319 | static hashval_t hash (scev_info_str *i); |
320 | static bool equal (const scev_info_str *a, const scev_info_str *b); |
321 | }; |
322 | |
323 | static GTY (()) hash_table<scev_info_hasher> *scalar_evolution_info; |
324 | |
325 | |
326 | /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */ |
327 | |
328 | static inline struct scev_info_str * |
329 | new_scev_info_str (basic_block instantiated_below, tree var) |
330 | { |
331 | struct scev_info_str *res; |
332 | |
333 | res = ggc_alloc<scev_info_str> (); |
334 | res->name_version = SSA_NAME_VERSION (var); |
335 | res->chrec = chrec_not_analyzed_yet; |
336 | res->instantiated_below = instantiated_below->index; |
337 | |
338 | return res; |
339 | } |
340 | |
341 | /* Computes a hash function for database element ELT. */ |
342 | |
343 | hashval_t |
344 | scev_info_hasher::hash (scev_info_str *elt) |
345 | { |
346 | return elt->name_version ^ elt->instantiated_below; |
347 | } |
348 | |
349 | /* Compares database elements E1 and E2. */ |
350 | |
351 | bool |
352 | scev_info_hasher::equal (const scev_info_str *elt1, const scev_info_str *elt2) |
353 | { |
354 | return (elt1->name_version == elt2->name_version |
355 | && elt1->instantiated_below == elt2->instantiated_below); |
356 | } |
357 | |
358 | /* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block. |
359 | A first query on VAR returns chrec_not_analyzed_yet. */ |
360 | |
361 | static tree * |
362 | find_var_scev_info (basic_block instantiated_below, tree var) |
363 | { |
364 | struct scev_info_str *res; |
365 | struct scev_info_str tmp; |
366 | |
367 | tmp.name_version = SSA_NAME_VERSION (var); |
368 | tmp.instantiated_below = instantiated_below->index; |
369 | scev_info_str **slot = scalar_evolution_info->find_slot (&tmp, INSERT); |
370 | |
371 | if (!*slot) |
372 | *slot = new_scev_info_str (instantiated_below, var); |
373 | res = *slot; |
374 | |
375 | return &res->chrec; |
376 | } |
377 | |
378 | /* Return true when CHREC contains symbolic names defined in |
379 | LOOP_NB. */ |
380 | |
381 | bool |
382 | chrec_contains_symbols_defined_in_loop (const_tree chrec, unsigned loop_nb) |
383 | { |
384 | int i, n; |
385 | |
386 | if (chrec == NULL_TREE) |
387 | return false; |
388 | |
389 | if (is_gimple_min_invariant (chrec)) |
390 | return false; |
391 | |
392 | if (TREE_CODE (chrec) == SSA_NAME) |
393 | { |
394 | gimple *def; |
395 | loop_p def_loop, loop; |
396 | |
397 | if (SSA_NAME_IS_DEFAULT_DEF (chrec)) |
398 | return false; |
399 | |
400 | def = SSA_NAME_DEF_STMT (chrec); |
401 | def_loop = loop_containing_stmt (def); |
402 | loop = get_loop (cfun, loop_nb); |
403 | |
404 | if (def_loop == NULL) |
405 | return false; |
406 | |
407 | if (loop == def_loop || flow_loop_nested_p (loop, def_loop)) |
408 | return true; |
409 | |
410 | return false; |
411 | } |
412 | |
413 | n = TREE_OPERAND_LENGTH (chrec); |
414 | for (i = 0; i < n; i++) |
415 | if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, i), |
416 | loop_nb)) |
417 | return true; |
418 | return false; |
419 | } |
420 | |
421 | /* Return true when PHI is a loop-phi-node. */ |
422 | |
423 | static bool |
424 | loop_phi_node_p (gimple *phi) |
425 | { |
426 | /* The implementation of this function is based on the following |
427 | property: "all the loop-phi-nodes of a loop are contained in the |
428 | loop's header basic block". */ |
429 | |
430 | return loop_containing_stmt (phi)->header == gimple_bb (phi); |
431 | } |
432 | |
433 | /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP. |
434 | In general, in the case of multivariate evolutions we want to get |
435 | the evolution in different loops. LOOP specifies the level for |
436 | which to get the evolution. |
437 | |
438 | Example: |
439 | |
440 | | for (j = 0; j < 100; j++) |
441 | | { |
442 | | for (k = 0; k < 100; k++) |
443 | | { |
444 | | i = k + j; - Here the value of i is a function of j, k. |
445 | | } |
446 | | ... = i - Here the value of i is a function of j. |
447 | | } |
448 | | ... = i - Here the value of i is a scalar. |
449 | |
450 | Example: |
451 | |
452 | | i_0 = ... |
453 | | loop_1 10 times |
454 | | i_1 = phi (i_0, i_2) |
455 | | i_2 = i_1 + 2 |
456 | | endloop |
457 | |
458 | This loop has the same effect as: |
459 | LOOP_1 has the same effect as: |
460 | |
461 | | i_1 = i_0 + 20 |
462 | |
463 | The overall effect of the loop, "i_0 + 20" in the previous example, |
464 | is obtained by passing in the parameters: LOOP = 1, |
465 | EVOLUTION_FN = {i_0, +, 2}_1. |
466 | */ |
467 | |
468 | tree |
469 | compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn) |
470 | { |
471 | bool val = false; |
472 | |
473 | if (evolution_fn == chrec_dont_know) |
474 | return chrec_dont_know; |
475 | |
476 | else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC) |
477 | { |
478 | struct loop *inner_loop = get_chrec_loop (evolution_fn); |
479 | |
480 | if (inner_loop == loop |
481 | || flow_loop_nested_p (loop, inner_loop)) |
482 | { |
483 | tree nb_iter = number_of_latch_executions (inner_loop); |
484 | |
485 | if (nb_iter == chrec_dont_know) |
486 | return chrec_dont_know; |
487 | else |
488 | { |
489 | tree res; |
490 | |
491 | /* evolution_fn is the evolution function in LOOP. Get |
492 | its value in the nb_iter-th iteration. */ |
493 | res = chrec_apply (inner_loop->num, evolution_fn, nb_iter); |
494 | |
495 | if (chrec_contains_symbols_defined_in_loop (res, loop->num)) |
496 | res = instantiate_parameters (loop, res); |
497 | |
498 | /* Continue the computation until ending on a parent of LOOP. */ |
499 | return compute_overall_effect_of_inner_loop (loop, res); |
500 | } |
501 | } |
502 | else |
503 | return evolution_fn; |
504 | } |
505 | |
506 | /* If the evolution function is an invariant, there is nothing to do. */ |
507 | else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val) |
508 | return evolution_fn; |
509 | |
510 | else |
511 | return chrec_dont_know; |
512 | } |
513 | |
514 | /* Associate CHREC to SCALAR. */ |
515 | |
516 | static void |
517 | set_scalar_evolution (basic_block instantiated_below, tree scalar, tree chrec) |
518 | { |
519 | tree *scalar_info; |
520 | |
521 | if (TREE_CODE (scalar) != SSA_NAME) |
522 | return; |
523 | |
524 | scalar_info = find_var_scev_info (instantiated_below, scalar); |
525 | |
526 | if (dump_file) |
527 | { |
528 | if (dump_flags & TDF_SCEV) |
529 | { |
530 | fprintf (dump_file, "(set_scalar_evolution \n" ); |
531 | fprintf (dump_file, " instantiated_below = %d \n" , |
532 | instantiated_below->index); |
533 | fprintf (dump_file, " (scalar = " ); |
534 | print_generic_expr (dump_file, scalar); |
535 | fprintf (dump_file, ")\n (scalar_evolution = " ); |
536 | print_generic_expr (dump_file, chrec); |
537 | fprintf (dump_file, "))\n" ); |
538 | } |
539 | if (dump_flags & TDF_STATS) |
540 | nb_set_scev++; |
541 | } |
542 | |
543 | *scalar_info = chrec; |
544 | } |
545 | |
546 | /* Retrieve the chrec associated to SCALAR instantiated below |
547 | INSTANTIATED_BELOW block. */ |
548 | |
549 | static tree |
550 | get_scalar_evolution (basic_block instantiated_below, tree scalar) |
551 | { |
552 | tree res; |
553 | |
554 | if (dump_file) |
555 | { |
556 | if (dump_flags & TDF_SCEV) |
557 | { |
558 | fprintf (dump_file, "(get_scalar_evolution \n" ); |
559 | fprintf (dump_file, " (scalar = " ); |
560 | print_generic_expr (dump_file, scalar); |
561 | fprintf (dump_file, ")\n" ); |
562 | } |
563 | if (dump_flags & TDF_STATS) |
564 | nb_get_scev++; |
565 | } |
566 | |
567 | if (VECTOR_TYPE_P (TREE_TYPE (scalar)) |
568 | || TREE_CODE (TREE_TYPE (scalar)) == COMPLEX_TYPE) |
569 | /* For chrec_dont_know we keep the symbolic form. */ |
570 | res = scalar; |
571 | else |
572 | switch (TREE_CODE (scalar)) |
573 | { |
574 | case SSA_NAME: |
575 | if (SSA_NAME_IS_DEFAULT_DEF (scalar)) |
576 | res = scalar; |
577 | else |
578 | res = *find_var_scev_info (instantiated_below, scalar); |
579 | break; |
580 | |
581 | case REAL_CST: |
582 | case FIXED_CST: |
583 | case INTEGER_CST: |
584 | res = scalar; |
585 | break; |
586 | |
587 | default: |
588 | res = chrec_not_analyzed_yet; |
589 | break; |
590 | } |
591 | |
592 | if (dump_file && (dump_flags & TDF_SCEV)) |
593 | { |
594 | fprintf (dump_file, " (scalar_evolution = " ); |
595 | print_generic_expr (dump_file, res); |
596 | fprintf (dump_file, "))\n" ); |
597 | } |
598 | |
599 | return res; |
600 | } |
601 | |
602 | /* Helper function for add_to_evolution. Returns the evolution |
603 | function for an assignment of the form "a = b + c", where "a" and |
604 | "b" are on the strongly connected component. CHREC_BEFORE is the |
605 | information that we already have collected up to this point. |
606 | TO_ADD is the evolution of "c". |
607 | |
608 | When CHREC_BEFORE has an evolution part in LOOP_NB, add to this |
609 | evolution the expression TO_ADD, otherwise construct an evolution |
610 | part for this loop. */ |
611 | |
612 | static tree |
613 | add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add, |
614 | gimple *at_stmt) |
615 | { |
616 | tree type, left, right; |
617 | struct loop *loop = get_loop (cfun, loop_nb), *chloop; |
618 | |
619 | switch (TREE_CODE (chrec_before)) |
620 | { |
621 | case POLYNOMIAL_CHREC: |
622 | chloop = get_chrec_loop (chrec_before); |
623 | if (chloop == loop |
624 | || flow_loop_nested_p (chloop, loop)) |
625 | { |
626 | unsigned var; |
627 | |
628 | type = chrec_type (chrec_before); |
629 | |
630 | /* When there is no evolution part in this loop, build it. */ |
631 | if (chloop != loop) |
632 | { |
633 | var = loop_nb; |
634 | left = chrec_before; |
635 | right = SCALAR_FLOAT_TYPE_P (type) |
636 | ? build_real (type, dconst0) |
637 | : build_int_cst (type, 0); |
638 | } |
639 | else |
640 | { |
641 | var = CHREC_VARIABLE (chrec_before); |
642 | left = CHREC_LEFT (chrec_before); |
643 | right = CHREC_RIGHT (chrec_before); |
644 | } |
645 | |
646 | to_add = chrec_convert (type, to_add, at_stmt); |
647 | right = chrec_convert_rhs (type, right, at_stmt); |
648 | right = chrec_fold_plus (chrec_type (right), right, to_add); |
649 | return build_polynomial_chrec (var, left, right); |
650 | } |
651 | else |
652 | { |
653 | gcc_assert (flow_loop_nested_p (loop, chloop)); |
654 | |
655 | /* Search the evolution in LOOP_NB. */ |
656 | left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before), |
657 | to_add, at_stmt); |
658 | right = CHREC_RIGHT (chrec_before); |
659 | right = chrec_convert_rhs (chrec_type (left), right, at_stmt); |
660 | return build_polynomial_chrec (CHREC_VARIABLE (chrec_before), |
661 | left, right); |
662 | } |
663 | |
664 | default: |
665 | /* These nodes do not depend on a loop. */ |
666 | if (chrec_before == chrec_dont_know) |
667 | return chrec_dont_know; |
668 | |
669 | left = chrec_before; |
670 | right = chrec_convert_rhs (chrec_type (left), to_add, at_stmt); |
671 | return build_polynomial_chrec (loop_nb, left, right); |
672 | } |
673 | } |
674 | |
675 | /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension |
676 | of LOOP_NB. |
677 | |
678 | Description (provided for completeness, for those who read code in |
679 | a plane, and for my poor 62 bytes brain that would have forgotten |
680 | all this in the next two or three months): |
681 | |
682 | The algorithm of translation of programs from the SSA representation |
683 | into the chrecs syntax is based on a pattern matching. After having |
684 | reconstructed the overall tree expression for a loop, there are only |
685 | two cases that can arise: |
686 | |
687 | 1. a = loop-phi (init, a + expr) |
688 | 2. a = loop-phi (init, expr) |
689 | |
690 | where EXPR is either a scalar constant with respect to the analyzed |
691 | loop (this is a degree 0 polynomial), or an expression containing |
692 | other loop-phi definitions (these are higher degree polynomials). |
693 | |
694 | Examples: |
695 | |
696 | 1. |
697 | | init = ... |
698 | | loop_1 |
699 | | a = phi (init, a + 5) |
700 | | endloop |
701 | |
702 | 2. |
703 | | inita = ... |
704 | | initb = ... |
705 | | loop_1 |
706 | | a = phi (inita, 2 * b + 3) |
707 | | b = phi (initb, b + 1) |
708 | | endloop |
709 | |
710 | For the first case, the semantics of the SSA representation is: |
711 | |
712 | | a (x) = init + \sum_{j = 0}^{x - 1} expr (j) |
713 | |
714 | that is, there is a loop index "x" that determines the scalar value |
715 | of the variable during the loop execution. During the first |
716 | iteration, the value is that of the initial condition INIT, while |
717 | during the subsequent iterations, it is the sum of the initial |
718 | condition with the sum of all the values of EXPR from the initial |
719 | iteration to the before last considered iteration. |
720 | |
721 | For the second case, the semantics of the SSA program is: |
722 | |
723 | | a (x) = init, if x = 0; |
724 | | expr (x - 1), otherwise. |
725 | |
726 | The second case corresponds to the PEELED_CHREC, whose syntax is |
727 | close to the syntax of a loop-phi-node: |
728 | |
729 | | phi (init, expr) vs. (init, expr)_x |
730 | |
731 | The proof of the translation algorithm for the first case is a |
732 | proof by structural induction based on the degree of EXPR. |
733 | |
734 | Degree 0: |
735 | When EXPR is a constant with respect to the analyzed loop, or in |
736 | other words when EXPR is a polynomial of degree 0, the evolution of |
737 | the variable A in the loop is an affine function with an initial |
738 | condition INIT, and a step EXPR. In order to show this, we start |
739 | from the semantics of the SSA representation: |
740 | |
741 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j) |
742 | |
743 | and since "expr (j)" is a constant with respect to "j", |
744 | |
745 | f (x) = init + x * expr |
746 | |
747 | Finally, based on the semantics of the pure sum chrecs, by |
748 | identification we get the corresponding chrecs syntax: |
749 | |
750 | f (x) = init * \binom{x}{0} + expr * \binom{x}{1} |
751 | f (x) -> {init, +, expr}_x |
752 | |
753 | Higher degree: |
754 | Suppose that EXPR is a polynomial of degree N with respect to the |
755 | analyzed loop_x for which we have already determined that it is |
756 | written under the chrecs syntax: |
757 | |
758 | | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x) |
759 | |
760 | We start from the semantics of the SSA program: |
761 | |
762 | | f (x) = init + \sum_{j = 0}^{x - 1} expr (j) |
763 | | |
764 | | f (x) = init + \sum_{j = 0}^{x - 1} |
765 | | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1}) |
766 | | |
767 | | f (x) = init + \sum_{j = 0}^{x - 1} |
768 | | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k}) |
769 | | |
770 | | f (x) = init + \sum_{k = 0}^{n - 1} |
771 | | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k}) |
772 | | |
773 | | f (x) = init + \sum_{k = 0}^{n - 1} |
774 | | (b_k * \binom{x}{k + 1}) |
775 | | |
776 | | f (x) = init + b_0 * \binom{x}{1} + ... |
777 | | + b_{n-1} * \binom{x}{n} |
778 | | |
779 | | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ... |
780 | | + b_{n-1} * \binom{x}{n} |
781 | | |
782 | |
783 | And finally from the definition of the chrecs syntax, we identify: |
784 | | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x |
785 | |
786 | This shows the mechanism that stands behind the add_to_evolution |
787 | function. An important point is that the use of symbolic |
788 | parameters avoids the need of an analysis schedule. |
789 | |
790 | Example: |
791 | |
792 | | inita = ... |
793 | | initb = ... |
794 | | loop_1 |
795 | | a = phi (inita, a + 2 + b) |
796 | | b = phi (initb, b + 1) |
797 | | endloop |
798 | |
799 | When analyzing "a", the algorithm keeps "b" symbolically: |
800 | |
801 | | a -> {inita, +, 2 + b}_1 |
802 | |
803 | Then, after instantiation, the analyzer ends on the evolution: |
804 | |
805 | | a -> {inita, +, 2 + initb, +, 1}_1 |
806 | |
807 | */ |
808 | |
809 | static tree |
810 | add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code, |
811 | tree to_add, gimple *at_stmt) |
812 | { |
813 | tree type = chrec_type (to_add); |
814 | tree res = NULL_TREE; |
815 | |
816 | if (to_add == NULL_TREE) |
817 | return chrec_before; |
818 | |
819 | /* TO_ADD is either a scalar, or a parameter. TO_ADD is not |
820 | instantiated at this point. */ |
821 | if (TREE_CODE (to_add) == POLYNOMIAL_CHREC) |
822 | /* This should not happen. */ |
823 | return chrec_dont_know; |
824 | |
825 | if (dump_file && (dump_flags & TDF_SCEV)) |
826 | { |
827 | fprintf (dump_file, "(add_to_evolution \n" ); |
828 | fprintf (dump_file, " (loop_nb = %d)\n" , loop_nb); |
829 | fprintf (dump_file, " (chrec_before = " ); |
830 | print_generic_expr (dump_file, chrec_before); |
831 | fprintf (dump_file, ")\n (to_add = " ); |
832 | print_generic_expr (dump_file, to_add); |
833 | fprintf (dump_file, ")\n" ); |
834 | } |
835 | |
836 | if (code == MINUS_EXPR) |
837 | to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type) |
838 | ? build_real (type, dconstm1) |
839 | : build_int_cst_type (type, -1)); |
840 | |
841 | res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt); |
842 | |
843 | if (dump_file && (dump_flags & TDF_SCEV)) |
844 | { |
845 | fprintf (dump_file, " (res = " ); |
846 | print_generic_expr (dump_file, res); |
847 | fprintf (dump_file, "))\n" ); |
848 | } |
849 | |
850 | return res; |
851 | } |
852 | |
853 | |
854 | |
855 | /* This section selects the loops that will be good candidates for the |
856 | scalar evolution analysis. For the moment, greedily select all the |
857 | loop nests we could analyze. */ |
858 | |
859 | /* For a loop with a single exit edge, return the COND_EXPR that |
860 | guards the exit edge. If the expression is too difficult to |
861 | analyze, then give up. */ |
862 | |
863 | gcond * |
864 | get_loop_exit_condition (const struct loop *loop) |
865 | { |
866 | gcond *res = NULL; |
867 | edge exit_edge = single_exit (loop); |
868 | |
869 | if (dump_file && (dump_flags & TDF_SCEV)) |
870 | fprintf (dump_file, "(get_loop_exit_condition \n " ); |
871 | |
872 | if (exit_edge) |
873 | { |
874 | gimple *stmt; |
875 | |
876 | stmt = last_stmt (exit_edge->src); |
877 | if (gcond *cond_stmt = dyn_cast <gcond *> (stmt)) |
878 | res = cond_stmt; |
879 | } |
880 | |
881 | if (dump_file && (dump_flags & TDF_SCEV)) |
882 | { |
883 | print_gimple_stmt (dump_file, res, 0); |
884 | fprintf (dump_file, ")\n" ); |
885 | } |
886 | |
887 | return res; |
888 | } |
889 | |
890 | |
891 | /* Depth first search algorithm. */ |
892 | |
893 | enum t_bool { |
894 | t_false, |
895 | t_true, |
896 | t_dont_know |
897 | }; |
898 | |
899 | |
900 | static t_bool follow_ssa_edge (struct loop *loop, gimple *, gphi *, |
901 | tree *, int); |
902 | |
903 | /* Follow the ssa edge into the binary expression RHS0 CODE RHS1. |
904 | Return true if the strongly connected component has been found. */ |
905 | |
906 | static t_bool |
907 | follow_ssa_edge_binary (struct loop *loop, gimple *at_stmt, |
908 | tree type, tree rhs0, enum tree_code code, tree rhs1, |
909 | gphi *halting_phi, tree *evolution_of_loop, |
910 | int limit) |
911 | { |
912 | t_bool res = t_false; |
913 | tree evol; |
914 | |
915 | switch (code) |
916 | { |
917 | case POINTER_PLUS_EXPR: |
918 | case PLUS_EXPR: |
919 | if (TREE_CODE (rhs0) == SSA_NAME) |
920 | { |
921 | if (TREE_CODE (rhs1) == SSA_NAME) |
922 | { |
923 | /* Match an assignment under the form: |
924 | "a = b + c". */ |
925 | |
926 | /* We want only assignments of form "name + name" contribute to |
927 | LIMIT, as the other cases do not necessarily contribute to |
928 | the complexity of the expression. */ |
929 | limit++; |
930 | |
931 | evol = *evolution_of_loop; |
932 | evol = add_to_evolution |
933 | (loop->num, |
934 | chrec_convert (type, evol, at_stmt), |
935 | code, rhs1, at_stmt); |
936 | res = follow_ssa_edge |
937 | (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, &evol, limit); |
938 | if (res == t_true) |
939 | *evolution_of_loop = evol; |
940 | else if (res == t_false) |
941 | { |
942 | *evolution_of_loop = add_to_evolution |
943 | (loop->num, |
944 | chrec_convert (type, *evolution_of_loop, at_stmt), |
945 | code, rhs0, at_stmt); |
946 | res = follow_ssa_edge |
947 | (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi, |
948 | evolution_of_loop, limit); |
949 | if (res == t_true) |
950 | ; |
951 | else if (res == t_dont_know) |
952 | *evolution_of_loop = chrec_dont_know; |
953 | } |
954 | |
955 | else if (res == t_dont_know) |
956 | *evolution_of_loop = chrec_dont_know; |
957 | } |
958 | |
959 | else |
960 | { |
961 | /* Match an assignment under the form: |
962 | "a = b + ...". */ |
963 | *evolution_of_loop = add_to_evolution |
964 | (loop->num, chrec_convert (type, *evolution_of_loop, |
965 | at_stmt), |
966 | code, rhs1, at_stmt); |
967 | res = follow_ssa_edge |
968 | (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, |
969 | evolution_of_loop, limit); |
970 | if (res == t_true) |
971 | ; |
972 | else if (res == t_dont_know) |
973 | *evolution_of_loop = chrec_dont_know; |
974 | } |
975 | } |
976 | |
977 | else if (TREE_CODE (rhs1) == SSA_NAME) |
978 | { |
979 | /* Match an assignment under the form: |
980 | "a = ... + c". */ |
981 | *evolution_of_loop = add_to_evolution |
982 | (loop->num, chrec_convert (type, *evolution_of_loop, |
983 | at_stmt), |
984 | code, rhs0, at_stmt); |
985 | res = follow_ssa_edge |
986 | (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi, |
987 | evolution_of_loop, limit); |
988 | if (res == t_true) |
989 | ; |
990 | else if (res == t_dont_know) |
991 | *evolution_of_loop = chrec_dont_know; |
992 | } |
993 | |
994 | else |
995 | /* Otherwise, match an assignment under the form: |
996 | "a = ... + ...". */ |
997 | /* And there is nothing to do. */ |
998 | res = t_false; |
999 | break; |
1000 | |
1001 | case MINUS_EXPR: |
1002 | /* This case is under the form "opnd0 = rhs0 - rhs1". */ |
1003 | if (TREE_CODE (rhs0) == SSA_NAME) |
1004 | { |
1005 | /* Match an assignment under the form: |
1006 | "a = b - ...". */ |
1007 | |
1008 | /* We want only assignments of form "name - name" contribute to |
1009 | LIMIT, as the other cases do not necessarily contribute to |
1010 | the complexity of the expression. */ |
1011 | if (TREE_CODE (rhs1) == SSA_NAME) |
1012 | limit++; |
1013 | |
1014 | *evolution_of_loop = add_to_evolution |
1015 | (loop->num, chrec_convert (type, *evolution_of_loop, at_stmt), |
1016 | MINUS_EXPR, rhs1, at_stmt); |
1017 | res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, |
1018 | evolution_of_loop, limit); |
1019 | if (res == t_true) |
1020 | ; |
1021 | else if (res == t_dont_know) |
1022 | *evolution_of_loop = chrec_dont_know; |
1023 | } |
1024 | else |
1025 | /* Otherwise, match an assignment under the form: |
1026 | "a = ... - ...". */ |
1027 | /* And there is nothing to do. */ |
1028 | res = t_false; |
1029 | break; |
1030 | |
1031 | default: |
1032 | res = t_false; |
1033 | } |
1034 | |
1035 | return res; |
1036 | } |
1037 | |
1038 | /* Follow the ssa edge into the expression EXPR. |
1039 | Return true if the strongly connected component has been found. */ |
1040 | |
1041 | static t_bool |
1042 | follow_ssa_edge_expr (struct loop *loop, gimple *at_stmt, tree expr, |
1043 | gphi *halting_phi, tree *evolution_of_loop, |
1044 | int limit) |
1045 | { |
1046 | enum tree_code code = TREE_CODE (expr); |
1047 | tree type = TREE_TYPE (expr), rhs0, rhs1; |
1048 | t_bool res; |
1049 | |
1050 | /* The EXPR is one of the following cases: |
1051 | - an SSA_NAME, |
1052 | - an INTEGER_CST, |
1053 | - a PLUS_EXPR, |
1054 | - a POINTER_PLUS_EXPR, |
1055 | - a MINUS_EXPR, |
1056 | - an ASSERT_EXPR, |
1057 | - other cases are not yet handled. */ |
1058 | |
1059 | switch (code) |
1060 | { |
1061 | CASE_CONVERT: |
1062 | /* This assignment is under the form "a_1 = (cast) rhs. */ |
1063 | res = follow_ssa_edge_expr (loop, at_stmt, TREE_OPERAND (expr, 0), |
1064 | halting_phi, evolution_of_loop, limit); |
1065 | *evolution_of_loop = chrec_convert (type, *evolution_of_loop, at_stmt); |
1066 | break; |
1067 | |
1068 | case INTEGER_CST: |
1069 | /* This assignment is under the form "a_1 = 7". */ |
1070 | res = t_false; |
1071 | break; |
1072 | |
1073 | case SSA_NAME: |
1074 | /* This assignment is under the form: "a_1 = b_2". */ |
1075 | res = follow_ssa_edge |
1076 | (loop, SSA_NAME_DEF_STMT (expr), halting_phi, evolution_of_loop, limit); |
1077 | break; |
1078 | |
1079 | case POINTER_PLUS_EXPR: |
1080 | case PLUS_EXPR: |
1081 | case MINUS_EXPR: |
1082 | /* This case is under the form "rhs0 +- rhs1". */ |
1083 | rhs0 = TREE_OPERAND (expr, 0); |
1084 | rhs1 = TREE_OPERAND (expr, 1); |
1085 | type = TREE_TYPE (rhs0); |
1086 | STRIP_USELESS_TYPE_CONVERSION (rhs0); |
1087 | STRIP_USELESS_TYPE_CONVERSION (rhs1); |
1088 | res = follow_ssa_edge_binary (loop, at_stmt, type, rhs0, code, rhs1, |
1089 | halting_phi, evolution_of_loop, limit); |
1090 | break; |
1091 | |
1092 | case ADDR_EXPR: |
1093 | /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */ |
1094 | if (TREE_CODE (TREE_OPERAND (expr, 0)) == MEM_REF) |
1095 | { |
1096 | expr = TREE_OPERAND (expr, 0); |
1097 | rhs0 = TREE_OPERAND (expr, 0); |
1098 | rhs1 = TREE_OPERAND (expr, 1); |
1099 | type = TREE_TYPE (rhs0); |
1100 | STRIP_USELESS_TYPE_CONVERSION (rhs0); |
1101 | STRIP_USELESS_TYPE_CONVERSION (rhs1); |
1102 | res = follow_ssa_edge_binary (loop, at_stmt, type, |
1103 | rhs0, POINTER_PLUS_EXPR, rhs1, |
1104 | halting_phi, evolution_of_loop, limit); |
1105 | } |
1106 | else |
1107 | res = t_false; |
1108 | break; |
1109 | |
1110 | case ASSERT_EXPR: |
1111 | /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>" |
1112 | It must be handled as a copy assignment of the form a_1 = a_2. */ |
1113 | rhs0 = ASSERT_EXPR_VAR (expr); |
1114 | if (TREE_CODE (rhs0) == SSA_NAME) |
1115 | res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), |
1116 | halting_phi, evolution_of_loop, limit); |
1117 | else |
1118 | res = t_false; |
1119 | break; |
1120 | |
1121 | default: |
1122 | res = t_false; |
1123 | break; |
1124 | } |
1125 | |
1126 | return res; |
1127 | } |
1128 | |
1129 | /* Follow the ssa edge into the right hand side of an assignment STMT. |
1130 | Return true if the strongly connected component has been found. */ |
1131 | |
1132 | static t_bool |
1133 | follow_ssa_edge_in_rhs (struct loop *loop, gimple *stmt, |
1134 | gphi *halting_phi, tree *evolution_of_loop, |
1135 | int limit) |
1136 | { |
1137 | enum tree_code code = gimple_assign_rhs_code (stmt); |
1138 | tree type = gimple_expr_type (stmt), rhs1, rhs2; |
1139 | t_bool res; |
1140 | |
1141 | switch (code) |
1142 | { |
1143 | CASE_CONVERT: |
1144 | /* This assignment is under the form "a_1 = (cast) rhs. */ |
1145 | res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt), |
1146 | halting_phi, evolution_of_loop, limit); |
1147 | *evolution_of_loop = chrec_convert (type, *evolution_of_loop, stmt); |
1148 | break; |
1149 | |
1150 | case POINTER_PLUS_EXPR: |
1151 | case PLUS_EXPR: |
1152 | case MINUS_EXPR: |
1153 | rhs1 = gimple_assign_rhs1 (stmt); |
1154 | rhs2 = gimple_assign_rhs2 (stmt); |
1155 | type = TREE_TYPE (rhs1); |
1156 | res = follow_ssa_edge_binary (loop, stmt, type, rhs1, code, rhs2, |
1157 | halting_phi, evolution_of_loop, limit); |
1158 | break; |
1159 | |
1160 | default: |
1161 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) |
1162 | res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt), |
1163 | halting_phi, evolution_of_loop, limit); |
1164 | else |
1165 | res = t_false; |
1166 | break; |
1167 | } |
1168 | |
1169 | return res; |
1170 | } |
1171 | |
1172 | /* Checks whether the I-th argument of a PHI comes from a backedge. */ |
1173 | |
1174 | static bool |
1175 | backedge_phi_arg_p (gphi *phi, int i) |
1176 | { |
1177 | const_edge e = gimple_phi_arg_edge (phi, i); |
1178 | |
1179 | /* We would in fact like to test EDGE_DFS_BACK here, but we do not care |
1180 | about updating it anywhere, and this should work as well most of the |
1181 | time. */ |
1182 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) |
1183 | return true; |
1184 | |
1185 | return false; |
1186 | } |
1187 | |
1188 | /* Helper function for one branch of the condition-phi-node. Return |
1189 | true if the strongly connected component has been found following |
1190 | this path. */ |
1191 | |
1192 | static inline t_bool |
1193 | follow_ssa_edge_in_condition_phi_branch (int i, |
1194 | struct loop *loop, |
1195 | gphi *condition_phi, |
1196 | gphi *halting_phi, |
1197 | tree *evolution_of_branch, |
1198 | tree init_cond, int limit) |
1199 | { |
1200 | tree branch = PHI_ARG_DEF (condition_phi, i); |
1201 | *evolution_of_branch = chrec_dont_know; |
1202 | |
1203 | /* Do not follow back edges (they must belong to an irreducible loop, which |
1204 | we really do not want to worry about). */ |
1205 | if (backedge_phi_arg_p (condition_phi, i)) |
1206 | return t_false; |
1207 | |
1208 | if (TREE_CODE (branch) == SSA_NAME) |
1209 | { |
1210 | *evolution_of_branch = init_cond; |
1211 | return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi, |
1212 | evolution_of_branch, limit); |
1213 | } |
1214 | |
1215 | /* This case occurs when one of the condition branches sets |
1216 | the variable to a constant: i.e. a phi-node like |
1217 | "a_2 = PHI <a_7(5), 2(6)>;". |
1218 | |
1219 | FIXME: This case have to be refined correctly: |
1220 | in some cases it is possible to say something better than |
1221 | chrec_dont_know, for example using a wrap-around notation. */ |
1222 | return t_false; |
1223 | } |
1224 | |
1225 | /* This function merges the branches of a condition-phi-node in a |
1226 | loop. */ |
1227 | |
1228 | static t_bool |
1229 | follow_ssa_edge_in_condition_phi (struct loop *loop, |
1230 | gphi *condition_phi, |
1231 | gphi *halting_phi, |
1232 | tree *evolution_of_loop, int limit) |
1233 | { |
1234 | int i, n; |
1235 | tree init = *evolution_of_loop; |
1236 | tree evolution_of_branch; |
1237 | t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi, |
1238 | halting_phi, |
1239 | &evolution_of_branch, |
1240 | init, limit); |
1241 | if (res == t_false || res == t_dont_know) |
1242 | return res; |
1243 | |
1244 | *evolution_of_loop = evolution_of_branch; |
1245 | |
1246 | n = gimple_phi_num_args (condition_phi); |
1247 | for (i = 1; i < n; i++) |
1248 | { |
1249 | /* Quickly give up when the evolution of one of the branches is |
1250 | not known. */ |
1251 | if (*evolution_of_loop == chrec_dont_know) |
1252 | return t_true; |
1253 | |
1254 | /* Increase the limit by the PHI argument number to avoid exponential |
1255 | time and memory complexity. */ |
1256 | res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi, |
1257 | halting_phi, |
1258 | &evolution_of_branch, |
1259 | init, limit + i); |
1260 | if (res == t_false || res == t_dont_know) |
1261 | return res; |
1262 | |
1263 | *evolution_of_loop = chrec_merge (*evolution_of_loop, |
1264 | evolution_of_branch); |
1265 | } |
1266 | |
1267 | return t_true; |
1268 | } |
1269 | |
1270 | /* Follow an SSA edge in an inner loop. It computes the overall |
1271 | effect of the loop, and following the symbolic initial conditions, |
1272 | it follows the edges in the parent loop. The inner loop is |
1273 | considered as a single statement. */ |
1274 | |
1275 | static t_bool |
1276 | follow_ssa_edge_inner_loop_phi (struct loop *outer_loop, |
1277 | gphi *loop_phi_node, |
1278 | gphi *halting_phi, |
1279 | tree *evolution_of_loop, int limit) |
1280 | { |
1281 | struct loop *loop = loop_containing_stmt (loop_phi_node); |
1282 | tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node)); |
1283 | |
1284 | /* Sometimes, the inner loop is too difficult to analyze, and the |
1285 | result of the analysis is a symbolic parameter. */ |
1286 | if (ev == PHI_RESULT (loop_phi_node)) |
1287 | { |
1288 | t_bool res = t_false; |
1289 | int i, n = gimple_phi_num_args (loop_phi_node); |
1290 | |
1291 | for (i = 0; i < n; i++) |
1292 | { |
1293 | tree arg = PHI_ARG_DEF (loop_phi_node, i); |
1294 | basic_block bb; |
1295 | |
1296 | /* Follow the edges that exit the inner loop. */ |
1297 | bb = gimple_phi_arg_edge (loop_phi_node, i)->src; |
1298 | if (!flow_bb_inside_loop_p (loop, bb)) |
1299 | res = follow_ssa_edge_expr (outer_loop, loop_phi_node, |
1300 | arg, halting_phi, |
1301 | evolution_of_loop, limit); |
1302 | if (res == t_true) |
1303 | break; |
1304 | } |
1305 | |
1306 | /* If the path crosses this loop-phi, give up. */ |
1307 | if (res == t_true) |
1308 | *evolution_of_loop = chrec_dont_know; |
1309 | |
1310 | return res; |
1311 | } |
1312 | |
1313 | /* Otherwise, compute the overall effect of the inner loop. */ |
1314 | ev = compute_overall_effect_of_inner_loop (loop, ev); |
1315 | return follow_ssa_edge_expr (outer_loop, loop_phi_node, ev, halting_phi, |
1316 | evolution_of_loop, limit); |
1317 | } |
1318 | |
1319 | /* Follow an SSA edge from a loop-phi-node to itself, constructing a |
1320 | path that is analyzed on the return walk. */ |
1321 | |
1322 | static t_bool |
1323 | follow_ssa_edge (struct loop *loop, gimple *def, gphi *halting_phi, |
1324 | tree *evolution_of_loop, int limit) |
1325 | { |
1326 | struct loop *def_loop; |
1327 | |
1328 | if (gimple_nop_p (def)) |
1329 | return t_false; |
1330 | |
1331 | /* Give up if the path is longer than the MAX that we allow. */ |
1332 | if (limit > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_COMPLEXITY)) |
1333 | return t_dont_know; |
1334 | |
1335 | def_loop = loop_containing_stmt (def); |
1336 | |
1337 | switch (gimple_code (def)) |
1338 | { |
1339 | case GIMPLE_PHI: |
1340 | if (!loop_phi_node_p (def)) |
1341 | /* DEF is a condition-phi-node. Follow the branches, and |
1342 | record their evolutions. Finally, merge the collected |
1343 | information and set the approximation to the main |
1344 | variable. */ |
1345 | return follow_ssa_edge_in_condition_phi |
1346 | (loop, as_a <gphi *> (def), halting_phi, evolution_of_loop, |
1347 | limit); |
1348 | |
1349 | /* When the analyzed phi is the halting_phi, the |
1350 | depth-first search is over: we have found a path from |
1351 | the halting_phi to itself in the loop. */ |
1352 | if (def == halting_phi) |
1353 | return t_true; |
1354 | |
1355 | /* Otherwise, the evolution of the HALTING_PHI depends |
1356 | on the evolution of another loop-phi-node, i.e. the |
1357 | evolution function is a higher degree polynomial. */ |
1358 | if (def_loop == loop) |
1359 | return t_false; |
1360 | |
1361 | /* Inner loop. */ |
1362 | if (flow_loop_nested_p (loop, def_loop)) |
1363 | return follow_ssa_edge_inner_loop_phi |
1364 | (loop, as_a <gphi *> (def), halting_phi, evolution_of_loop, |
1365 | limit + 1); |
1366 | |
1367 | /* Outer loop. */ |
1368 | return t_false; |
1369 | |
1370 | case GIMPLE_ASSIGN: |
1371 | return follow_ssa_edge_in_rhs (loop, def, halting_phi, |
1372 | evolution_of_loop, limit); |
1373 | |
1374 | default: |
1375 | /* At this level of abstraction, the program is just a set |
1376 | of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no |
1377 | other node to be handled. */ |
1378 | return t_false; |
1379 | } |
1380 | } |
1381 | |
1382 | |
1383 | /* Simplify PEELED_CHREC represented by (init_cond, arg) in LOOP. |
1384 | Handle below case and return the corresponding POLYNOMIAL_CHREC: |
1385 | |
1386 | # i_17 = PHI <i_13(5), 0(3)> |
1387 | # _20 = PHI <_5(5), start_4(D)(3)> |
1388 | ... |
1389 | i_13 = i_17 + 1; |
1390 | _5 = start_4(D) + i_13; |
1391 | |
1392 | Though variable _20 appears as a PEELED_CHREC in the form of |
1393 | (start_4, _5)_LOOP, it's a POLYNOMIAL_CHREC like {start_4, 1}_LOOP. |
1394 | |
1395 | See PR41488. */ |
1396 | |
1397 | static tree |
1398 | simplify_peeled_chrec (struct loop *loop, tree arg, tree init_cond) |
1399 | { |
1400 | aff_tree aff1, aff2; |
1401 | tree ev, left, right, type, step_val; |
1402 | hash_map<tree, name_expansion *> *peeled_chrec_map = NULL; |
1403 | |
1404 | ev = instantiate_parameters (loop, analyze_scalar_evolution (loop, arg)); |
1405 | if (ev == NULL_TREE || TREE_CODE (ev) != POLYNOMIAL_CHREC) |
1406 | return chrec_dont_know; |
1407 | |
1408 | left = CHREC_LEFT (ev); |
1409 | right = CHREC_RIGHT (ev); |
1410 | type = TREE_TYPE (left); |
1411 | step_val = chrec_fold_plus (type, init_cond, right); |
1412 | |
1413 | /* Transform (init, {left, right}_LOOP)_LOOP to {init, right}_LOOP |
1414 | if "left" equals to "init + right". */ |
1415 | if (operand_equal_p (left, step_val, 0)) |
1416 | { |
1417 | if (dump_file && (dump_flags & TDF_SCEV)) |
1418 | fprintf (dump_file, "Simplify PEELED_CHREC into POLYNOMIAL_CHREC.\n" ); |
1419 | |
1420 | return build_polynomial_chrec (loop->num, init_cond, right); |
1421 | } |
1422 | |
1423 | /* Try harder to check if they are equal. */ |
1424 | tree_to_aff_combination_expand (left, type, &aff1, &peeled_chrec_map); |
1425 | tree_to_aff_combination_expand (step_val, type, &aff2, &peeled_chrec_map); |
1426 | free_affine_expand_cache (&peeled_chrec_map); |
1427 | aff_combination_scale (&aff2, -1); |
1428 | aff_combination_add (&aff1, &aff2); |
1429 | |
1430 | /* Transform (init, {left, right}_LOOP)_LOOP to {init, right}_LOOP |
1431 | if "left" equals to "init + right". */ |
1432 | if (aff_combination_zero_p (&aff1)) |
1433 | { |
1434 | if (dump_file && (dump_flags & TDF_SCEV)) |
1435 | fprintf (dump_file, "Simplify PEELED_CHREC into POLYNOMIAL_CHREC.\n" ); |
1436 | |
1437 | return build_polynomial_chrec (loop->num, init_cond, right); |
1438 | } |
1439 | return chrec_dont_know; |
1440 | } |
1441 | |
1442 | /* Given a LOOP_PHI_NODE, this function determines the evolution |
1443 | function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */ |
1444 | |
1445 | static tree |
1446 | analyze_evolution_in_loop (gphi *loop_phi_node, |
1447 | tree init_cond) |
1448 | { |
1449 | int i, n = gimple_phi_num_args (loop_phi_node); |
1450 | tree evolution_function = chrec_not_analyzed_yet; |
1451 | struct loop *loop = loop_containing_stmt (loop_phi_node); |
1452 | basic_block bb; |
1453 | static bool simplify_peeled_chrec_p = true; |
1454 | |
1455 | if (dump_file && (dump_flags & TDF_SCEV)) |
1456 | { |
1457 | fprintf (dump_file, "(analyze_evolution_in_loop \n" ); |
1458 | fprintf (dump_file, " (loop_phi_node = " ); |
1459 | print_gimple_stmt (dump_file, loop_phi_node, 0); |
1460 | fprintf (dump_file, ")\n" ); |
1461 | } |
1462 | |
1463 | for (i = 0; i < n; i++) |
1464 | { |
1465 | tree arg = PHI_ARG_DEF (loop_phi_node, i); |
1466 | gimple *ssa_chain; |
1467 | tree ev_fn; |
1468 | t_bool res; |
1469 | |
1470 | /* Select the edges that enter the loop body. */ |
1471 | bb = gimple_phi_arg_edge (loop_phi_node, i)->src; |
1472 | if (!flow_bb_inside_loop_p (loop, bb)) |
1473 | continue; |
1474 | |
1475 | if (TREE_CODE (arg) == SSA_NAME) |
1476 | { |
1477 | bool val = false; |
1478 | |
1479 | ssa_chain = SSA_NAME_DEF_STMT (arg); |
1480 | |
1481 | /* Pass in the initial condition to the follow edge function. */ |
1482 | ev_fn = init_cond; |
1483 | res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0); |
1484 | |
1485 | /* If ev_fn has no evolution in the inner loop, and the |
1486 | init_cond is not equal to ev_fn, then we have an |
1487 | ambiguity between two possible values, as we cannot know |
1488 | the number of iterations at this point. */ |
1489 | if (TREE_CODE (ev_fn) != POLYNOMIAL_CHREC |
1490 | && no_evolution_in_loop_p (ev_fn, loop->num, &val) && val |
1491 | && !operand_equal_p (init_cond, ev_fn, 0)) |
1492 | ev_fn = chrec_dont_know; |
1493 | } |
1494 | else |
1495 | res = t_false; |
1496 | |
1497 | /* When it is impossible to go back on the same |
1498 | loop_phi_node by following the ssa edges, the |
1499 | evolution is represented by a peeled chrec, i.e. the |
1500 | first iteration, EV_FN has the value INIT_COND, then |
1501 | all the other iterations it has the value of ARG. |
1502 | For the moment, PEELED_CHREC nodes are not built. */ |
1503 | if (res != t_true) |
1504 | { |
1505 | ev_fn = chrec_dont_know; |
1506 | /* Try to recognize POLYNOMIAL_CHREC which appears in |
1507 | the form of PEELED_CHREC, but guard the process with |
1508 | a bool variable to keep the analyzer from infinite |
1509 | recurrence for real PEELED_RECs. */ |
1510 | if (simplify_peeled_chrec_p && TREE_CODE (arg) == SSA_NAME) |
1511 | { |
1512 | simplify_peeled_chrec_p = false; |
1513 | ev_fn = simplify_peeled_chrec (loop, arg, init_cond); |
1514 | simplify_peeled_chrec_p = true; |
1515 | } |
1516 | } |
1517 | |
1518 | /* When there are multiple back edges of the loop (which in fact never |
1519 | happens currently, but nevertheless), merge their evolutions. */ |
1520 | evolution_function = chrec_merge (evolution_function, ev_fn); |
1521 | |
1522 | if (evolution_function == chrec_dont_know) |
1523 | break; |
1524 | } |
1525 | |
1526 | if (dump_file && (dump_flags & TDF_SCEV)) |
1527 | { |
1528 | fprintf (dump_file, " (evolution_function = " ); |
1529 | print_generic_expr (dump_file, evolution_function); |
1530 | fprintf (dump_file, "))\n" ); |
1531 | } |
1532 | |
1533 | return evolution_function; |
1534 | } |
1535 | |
1536 | /* Looks to see if VAR is a copy of a constant (via straightforward assignments |
1537 | or degenerate phi's). If so, returns the constant; else, returns VAR. */ |
1538 | |
1539 | static tree |
1540 | follow_copies_to_constant (tree var) |
1541 | { |
1542 | tree res = var; |
1543 | while (TREE_CODE (res) == SSA_NAME) |
1544 | { |
1545 | gimple *def = SSA_NAME_DEF_STMT (res); |
1546 | if (gphi *phi = dyn_cast <gphi *> (def)) |
1547 | { |
1548 | if (tree rhs = degenerate_phi_result (phi)) |
1549 | res = rhs; |
1550 | else |
1551 | break; |
1552 | } |
1553 | else if (gimple_assign_single_p (def)) |
1554 | /* Will exit loop if not an SSA_NAME. */ |
1555 | res = gimple_assign_rhs1 (def); |
1556 | else |
1557 | break; |
1558 | } |
1559 | if (CONSTANT_CLASS_P (res)) |
1560 | return res; |
1561 | return var; |
1562 | } |
1563 | |
1564 | /* Given a loop-phi-node, return the initial conditions of the |
1565 | variable on entry of the loop. When the CCP has propagated |
1566 | constants into the loop-phi-node, the initial condition is |
1567 | instantiated, otherwise the initial condition is kept symbolic. |
1568 | This analyzer does not analyze the evolution outside the current |
1569 | loop, and leaves this task to the on-demand tree reconstructor. */ |
1570 | |
1571 | static tree |
1572 | analyze_initial_condition (gphi *loop_phi_node) |
1573 | { |
1574 | int i, n; |
1575 | tree init_cond = chrec_not_analyzed_yet; |
1576 | struct loop *loop = loop_containing_stmt (loop_phi_node); |
1577 | |
1578 | if (dump_file && (dump_flags & TDF_SCEV)) |
1579 | { |
1580 | fprintf (dump_file, "(analyze_initial_condition \n" ); |
1581 | fprintf (dump_file, " (loop_phi_node = \n" ); |
1582 | print_gimple_stmt (dump_file, loop_phi_node, 0); |
1583 | fprintf (dump_file, ")\n" ); |
1584 | } |
1585 | |
1586 | n = gimple_phi_num_args (loop_phi_node); |
1587 | for (i = 0; i < n; i++) |
1588 | { |
1589 | tree branch = PHI_ARG_DEF (loop_phi_node, i); |
1590 | basic_block bb = gimple_phi_arg_edge (loop_phi_node, i)->src; |
1591 | |
1592 | /* When the branch is oriented to the loop's body, it does |
1593 | not contribute to the initial condition. */ |
1594 | if (flow_bb_inside_loop_p (loop, bb)) |
1595 | continue; |
1596 | |
1597 | if (init_cond == chrec_not_analyzed_yet) |
1598 | { |
1599 | init_cond = branch; |
1600 | continue; |
1601 | } |
1602 | |
1603 | if (TREE_CODE (branch) == SSA_NAME) |
1604 | { |
1605 | init_cond = chrec_dont_know; |
1606 | break; |
1607 | } |
1608 | |
1609 | init_cond = chrec_merge (init_cond, branch); |
1610 | } |
1611 | |
1612 | /* Ooops -- a loop without an entry??? */ |
1613 | if (init_cond == chrec_not_analyzed_yet) |
1614 | init_cond = chrec_dont_know; |
1615 | |
1616 | /* We may not have fully constant propagated IL. Handle degenerate PHIs here |
1617 | to not miss important early loop unrollings. */ |
1618 | init_cond = follow_copies_to_constant (init_cond); |
1619 | |
1620 | if (dump_file && (dump_flags & TDF_SCEV)) |
1621 | { |
1622 | fprintf (dump_file, " (init_cond = " ); |
1623 | print_generic_expr (dump_file, init_cond); |
1624 | fprintf (dump_file, "))\n" ); |
1625 | } |
1626 | |
1627 | return init_cond; |
1628 | } |
1629 | |
1630 | /* Analyze the scalar evolution for LOOP_PHI_NODE. */ |
1631 | |
1632 | static tree |
1633 | interpret_loop_phi (struct loop *loop, gphi *loop_phi_node) |
1634 | { |
1635 | tree res; |
1636 | struct loop *phi_loop = loop_containing_stmt (loop_phi_node); |
1637 | tree init_cond; |
1638 | |
1639 | gcc_assert (phi_loop == loop); |
1640 | |
1641 | /* Otherwise really interpret the loop phi. */ |
1642 | init_cond = analyze_initial_condition (loop_phi_node); |
1643 | res = analyze_evolution_in_loop (loop_phi_node, init_cond); |
1644 | |
1645 | /* Verify we maintained the correct initial condition throughout |
1646 | possible conversions in the SSA chain. */ |
1647 | if (res != chrec_dont_know) |
1648 | { |
1649 | tree new_init = res; |
1650 | if (CONVERT_EXPR_P (res) |
1651 | && TREE_CODE (TREE_OPERAND (res, 0)) == POLYNOMIAL_CHREC) |
1652 | new_init = fold_convert (TREE_TYPE (res), |
1653 | CHREC_LEFT (TREE_OPERAND (res, 0))); |
1654 | else if (TREE_CODE (res) == POLYNOMIAL_CHREC) |
1655 | new_init = CHREC_LEFT (res); |
1656 | STRIP_USELESS_TYPE_CONVERSION (new_init); |
1657 | if (TREE_CODE (new_init) == POLYNOMIAL_CHREC |
1658 | || !operand_equal_p (init_cond, new_init, 0)) |
1659 | return chrec_dont_know; |
1660 | } |
1661 | |
1662 | return res; |
1663 | } |
1664 | |
1665 | /* This function merges the branches of a condition-phi-node, |
1666 | contained in the outermost loop, and whose arguments are already |
1667 | analyzed. */ |
1668 | |
1669 | static tree |
1670 | interpret_condition_phi (struct loop *loop, gphi *condition_phi) |
1671 | { |
1672 | int i, n = gimple_phi_num_args (condition_phi); |
1673 | tree res = chrec_not_analyzed_yet; |
1674 | |
1675 | for (i = 0; i < n; i++) |
1676 | { |
1677 | tree branch_chrec; |
1678 | |
1679 | if (backedge_phi_arg_p (condition_phi, i)) |
1680 | { |
1681 | res = chrec_dont_know; |
1682 | break; |
1683 | } |
1684 | |
1685 | branch_chrec = analyze_scalar_evolution |
1686 | (loop, PHI_ARG_DEF (condition_phi, i)); |
1687 | |
1688 | res = chrec_merge (res, branch_chrec); |
1689 | if (res == chrec_dont_know) |
1690 | break; |
1691 | } |
1692 | |
1693 | return res; |
1694 | } |
1695 | |
1696 | /* Interpret the operation RHS1 OP RHS2. If we didn't |
1697 | analyze this node before, follow the definitions until ending |
1698 | either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the |
1699 | return path, this function propagates evolutions (ala constant copy |
1700 | propagation). OPND1 is not a GIMPLE expression because we could |
1701 | analyze the effect of an inner loop: see interpret_loop_phi. */ |
1702 | |
1703 | static tree |
1704 | interpret_rhs_expr (struct loop *loop, gimple *at_stmt, |
1705 | tree type, tree rhs1, enum tree_code code, tree rhs2) |
1706 | { |
1707 | tree res, chrec1, chrec2, ctype; |
1708 | gimple *def; |
1709 | |
1710 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) |
1711 | { |
1712 | if (is_gimple_min_invariant (rhs1)) |
1713 | return chrec_convert (type, rhs1, at_stmt); |
1714 | |
1715 | if (code == SSA_NAME) |
1716 | return chrec_convert (type, analyze_scalar_evolution (loop, rhs1), |
1717 | at_stmt); |
1718 | |
1719 | if (code == ASSERT_EXPR) |
1720 | { |
1721 | rhs1 = ASSERT_EXPR_VAR (rhs1); |
1722 | return chrec_convert (type, analyze_scalar_evolution (loop, rhs1), |
1723 | at_stmt); |
1724 | } |
1725 | } |
1726 | |
1727 | switch (code) |
1728 | { |
1729 | case ADDR_EXPR: |
1730 | if (TREE_CODE (TREE_OPERAND (rhs1, 0)) == MEM_REF |
1731 | || handled_component_p (TREE_OPERAND (rhs1, 0))) |
1732 | { |
1733 | machine_mode mode; |
1734 | HOST_WIDE_INT bitsize, bitpos; |
1735 | int unsignedp, reversep; |
1736 | int volatilep = 0; |
1737 | tree base, offset; |
1738 | tree chrec3; |
1739 | tree unitpos; |
1740 | |
1741 | base = get_inner_reference (TREE_OPERAND (rhs1, 0), |
1742 | &bitsize, &bitpos, &offset, &mode, |
1743 | &unsignedp, &reversep, &volatilep); |
1744 | |
1745 | if (TREE_CODE (base) == MEM_REF) |
1746 | { |
1747 | rhs2 = TREE_OPERAND (base, 1); |
1748 | rhs1 = TREE_OPERAND (base, 0); |
1749 | |
1750 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1751 | chrec2 = analyze_scalar_evolution (loop, rhs2); |
1752 | chrec1 = chrec_convert (type, chrec1, at_stmt); |
1753 | chrec2 = chrec_convert (TREE_TYPE (rhs2), chrec2, at_stmt); |
1754 | chrec1 = instantiate_parameters (loop, chrec1); |
1755 | chrec2 = instantiate_parameters (loop, chrec2); |
1756 | res = chrec_fold_plus (type, chrec1, chrec2); |
1757 | } |
1758 | else |
1759 | { |
1760 | chrec1 = analyze_scalar_evolution_for_address_of (loop, base); |
1761 | chrec1 = chrec_convert (type, chrec1, at_stmt); |
1762 | res = chrec1; |
1763 | } |
1764 | |
1765 | if (offset != NULL_TREE) |
1766 | { |
1767 | chrec2 = analyze_scalar_evolution (loop, offset); |
1768 | chrec2 = chrec_convert (TREE_TYPE (offset), chrec2, at_stmt); |
1769 | chrec2 = instantiate_parameters (loop, chrec2); |
1770 | res = chrec_fold_plus (type, res, chrec2); |
1771 | } |
1772 | |
1773 | if (bitpos != 0) |
1774 | { |
1775 | gcc_assert ((bitpos % BITS_PER_UNIT) == 0); |
1776 | |
1777 | unitpos = size_int (bitpos / BITS_PER_UNIT); |
1778 | chrec3 = analyze_scalar_evolution (loop, unitpos); |
1779 | chrec3 = chrec_convert (TREE_TYPE (unitpos), chrec3, at_stmt); |
1780 | chrec3 = instantiate_parameters (loop, chrec3); |
1781 | res = chrec_fold_plus (type, res, chrec3); |
1782 | } |
1783 | } |
1784 | else |
1785 | res = chrec_dont_know; |
1786 | break; |
1787 | |
1788 | case POINTER_PLUS_EXPR: |
1789 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1790 | chrec2 = analyze_scalar_evolution (loop, rhs2); |
1791 | chrec1 = chrec_convert (type, chrec1, at_stmt); |
1792 | chrec2 = chrec_convert (TREE_TYPE (rhs2), chrec2, at_stmt); |
1793 | chrec1 = instantiate_parameters (loop, chrec1); |
1794 | chrec2 = instantiate_parameters (loop, chrec2); |
1795 | res = chrec_fold_plus (type, chrec1, chrec2); |
1796 | break; |
1797 | |
1798 | case PLUS_EXPR: |
1799 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1800 | chrec2 = analyze_scalar_evolution (loop, rhs2); |
1801 | ctype = type; |
1802 | /* When the stmt is conditionally executed re-write the CHREC |
1803 | into a form that has well-defined behavior on overflow. */ |
1804 | if (at_stmt |
1805 | && INTEGRAL_TYPE_P (type) |
1806 | && ! TYPE_OVERFLOW_WRAPS (type) |
1807 | && ! dominated_by_p (CDI_DOMINATORS, loop->latch, |
1808 | gimple_bb (at_stmt))) |
1809 | ctype = unsigned_type_for (type); |
1810 | chrec1 = chrec_convert (ctype, chrec1, at_stmt); |
1811 | chrec2 = chrec_convert (ctype, chrec2, at_stmt); |
1812 | chrec1 = instantiate_parameters (loop, chrec1); |
1813 | chrec2 = instantiate_parameters (loop, chrec2); |
1814 | res = chrec_fold_plus (ctype, chrec1, chrec2); |
1815 | if (type != ctype) |
1816 | res = chrec_convert (type, res, at_stmt); |
1817 | break; |
1818 | |
1819 | case MINUS_EXPR: |
1820 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1821 | chrec2 = analyze_scalar_evolution (loop, rhs2); |
1822 | ctype = type; |
1823 | /* When the stmt is conditionally executed re-write the CHREC |
1824 | into a form that has well-defined behavior on overflow. */ |
1825 | if (at_stmt |
1826 | && INTEGRAL_TYPE_P (type) |
1827 | && ! TYPE_OVERFLOW_WRAPS (type) |
1828 | && ! dominated_by_p (CDI_DOMINATORS, |
1829 | loop->latch, gimple_bb (at_stmt))) |
1830 | ctype = unsigned_type_for (type); |
1831 | chrec1 = chrec_convert (ctype, chrec1, at_stmt); |
1832 | chrec2 = chrec_convert (ctype, chrec2, at_stmt); |
1833 | chrec1 = instantiate_parameters (loop, chrec1); |
1834 | chrec2 = instantiate_parameters (loop, chrec2); |
1835 | res = chrec_fold_minus (ctype, chrec1, chrec2); |
1836 | if (type != ctype) |
1837 | res = chrec_convert (type, res, at_stmt); |
1838 | break; |
1839 | |
1840 | case NEGATE_EXPR: |
1841 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1842 | ctype = type; |
1843 | /* When the stmt is conditionally executed re-write the CHREC |
1844 | into a form that has well-defined behavior on overflow. */ |
1845 | if (at_stmt |
1846 | && INTEGRAL_TYPE_P (type) |
1847 | && ! TYPE_OVERFLOW_WRAPS (type) |
1848 | && ! dominated_by_p (CDI_DOMINATORS, |
1849 | loop->latch, gimple_bb (at_stmt))) |
1850 | ctype = unsigned_type_for (type); |
1851 | chrec1 = chrec_convert (ctype, chrec1, at_stmt); |
1852 | /* TYPE may be integer, real or complex, so use fold_convert. */ |
1853 | chrec1 = instantiate_parameters (loop, chrec1); |
1854 | res = chrec_fold_multiply (ctype, chrec1, |
1855 | fold_convert (ctype, integer_minus_one_node)); |
1856 | if (type != ctype) |
1857 | res = chrec_convert (type, res, at_stmt); |
1858 | break; |
1859 | |
1860 | case BIT_NOT_EXPR: |
1861 | /* Handle ~X as -1 - X. */ |
1862 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1863 | chrec1 = chrec_convert (type, chrec1, at_stmt); |
1864 | chrec1 = instantiate_parameters (loop, chrec1); |
1865 | res = chrec_fold_minus (type, |
1866 | fold_convert (type, integer_minus_one_node), |
1867 | chrec1); |
1868 | break; |
1869 | |
1870 | case MULT_EXPR: |
1871 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1872 | chrec2 = analyze_scalar_evolution (loop, rhs2); |
1873 | ctype = type; |
1874 | /* When the stmt is conditionally executed re-write the CHREC |
1875 | into a form that has well-defined behavior on overflow. */ |
1876 | if (at_stmt |
1877 | && INTEGRAL_TYPE_P (type) |
1878 | && ! TYPE_OVERFLOW_WRAPS (type) |
1879 | && ! dominated_by_p (CDI_DOMINATORS, |
1880 | loop->latch, gimple_bb (at_stmt))) |
1881 | ctype = unsigned_type_for (type); |
1882 | chrec1 = chrec_convert (ctype, chrec1, at_stmt); |
1883 | chrec2 = chrec_convert (ctype, chrec2, at_stmt); |
1884 | chrec1 = instantiate_parameters (loop, chrec1); |
1885 | chrec2 = instantiate_parameters (loop, chrec2); |
1886 | res = chrec_fold_multiply (ctype, chrec1, chrec2); |
1887 | if (type != ctype) |
1888 | res = chrec_convert (type, res, at_stmt); |
1889 | break; |
1890 | |
1891 | case LSHIFT_EXPR: |
1892 | { |
1893 | /* Handle A<<B as A * (1<<B). */ |
1894 | tree uns = unsigned_type_for (type); |
1895 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1896 | chrec2 = analyze_scalar_evolution (loop, rhs2); |
1897 | chrec1 = chrec_convert (uns, chrec1, at_stmt); |
1898 | chrec1 = instantiate_parameters (loop, chrec1); |
1899 | chrec2 = instantiate_parameters (loop, chrec2); |
1900 | |
1901 | tree one = build_int_cst (uns, 1); |
1902 | chrec2 = fold_build2 (LSHIFT_EXPR, uns, one, chrec2); |
1903 | res = chrec_fold_multiply (uns, chrec1, chrec2); |
1904 | res = chrec_convert (type, res, at_stmt); |
1905 | } |
1906 | break; |
1907 | |
1908 | CASE_CONVERT: |
1909 | /* In case we have a truncation of a widened operation that in |
1910 | the truncated type has undefined overflow behavior analyze |
1911 | the operation done in an unsigned type of the same precision |
1912 | as the final truncation. We cannot derive a scalar evolution |
1913 | for the widened operation but for the truncated result. */ |
1914 | if (TREE_CODE (type) == INTEGER_TYPE |
1915 | && TREE_CODE (TREE_TYPE (rhs1)) == INTEGER_TYPE |
1916 | && TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (rhs1)) |
1917 | && TYPE_OVERFLOW_UNDEFINED (type) |
1918 | && TREE_CODE (rhs1) == SSA_NAME |
1919 | && (def = SSA_NAME_DEF_STMT (rhs1)) |
1920 | && is_gimple_assign (def) |
1921 | && TREE_CODE_CLASS (gimple_assign_rhs_code (def)) == tcc_binary |
1922 | && TREE_CODE (gimple_assign_rhs2 (def)) == INTEGER_CST) |
1923 | { |
1924 | tree utype = unsigned_type_for (type); |
1925 | chrec1 = interpret_rhs_expr (loop, at_stmt, utype, |
1926 | gimple_assign_rhs1 (def), |
1927 | gimple_assign_rhs_code (def), |
1928 | gimple_assign_rhs2 (def)); |
1929 | } |
1930 | else |
1931 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1932 | res = chrec_convert (type, chrec1, at_stmt, true, rhs1); |
1933 | break; |
1934 | |
1935 | case BIT_AND_EXPR: |
1936 | /* Given int variable A, handle A&0xffff as (int)(unsigned short)A. |
1937 | If A is SCEV and its value is in the range of representable set |
1938 | of type unsigned short, the result expression is a (no-overflow) |
1939 | SCEV. */ |
1940 | res = chrec_dont_know; |
1941 | if (tree_fits_uhwi_p (rhs2)) |
1942 | { |
1943 | int precision; |
1944 | unsigned HOST_WIDE_INT val = tree_to_uhwi (rhs2); |
1945 | |
1946 | val ++; |
1947 | /* Skip if value of rhs2 wraps in unsigned HOST_WIDE_INT or |
1948 | it's not the maximum value of a smaller type than rhs1. */ |
1949 | if (val != 0 |
1950 | && (precision = exact_log2 (val)) > 0 |
1951 | && (unsigned) precision < TYPE_PRECISION (TREE_TYPE (rhs1))) |
1952 | { |
1953 | tree utype = build_nonstandard_integer_type (precision, 1); |
1954 | |
1955 | if (TYPE_PRECISION (utype) < TYPE_PRECISION (TREE_TYPE (rhs1))) |
1956 | { |
1957 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1958 | chrec1 = chrec_convert (utype, chrec1, at_stmt); |
1959 | res = chrec_convert (TREE_TYPE (rhs1), chrec1, at_stmt); |
1960 | } |
1961 | } |
1962 | } |
1963 | break; |
1964 | |
1965 | default: |
1966 | res = chrec_dont_know; |
1967 | break; |
1968 | } |
1969 | |
1970 | return res; |
1971 | } |
1972 | |
1973 | /* Interpret the expression EXPR. */ |
1974 | |
1975 | static tree |
1976 | interpret_expr (struct loop *loop, gimple *at_stmt, tree expr) |
1977 | { |
1978 | enum tree_code code; |
1979 | tree type = TREE_TYPE (expr), op0, op1; |
1980 | |
1981 | if (automatically_generated_chrec_p (expr)) |
1982 | return expr; |
1983 | |
1984 | if (TREE_CODE (expr) == POLYNOMIAL_CHREC |
1985 | || get_gimple_rhs_class (TREE_CODE (expr)) == GIMPLE_TERNARY_RHS) |
1986 | return chrec_dont_know; |
1987 | |
1988 | extract_ops_from_tree (expr, &code, &op0, &op1); |
1989 | |
1990 | return interpret_rhs_expr (loop, at_stmt, type, |
1991 | op0, code, op1); |
1992 | } |
1993 | |
1994 | /* Interpret the rhs of the assignment STMT. */ |
1995 | |
1996 | static tree |
1997 | interpret_gimple_assign (struct loop *loop, gimple *stmt) |
1998 | { |
1999 | tree type = TREE_TYPE (gimple_assign_lhs (stmt)); |
2000 | enum tree_code code = gimple_assign_rhs_code (stmt); |
2001 | |
2002 | return interpret_rhs_expr (loop, stmt, type, |
2003 | gimple_assign_rhs1 (stmt), code, |
2004 | gimple_assign_rhs2 (stmt)); |
2005 | } |
2006 | |
2007 | |
2008 | |
2009 | /* This section contains all the entry points: |
2010 | - number_of_iterations_in_loop, |
2011 | - analyze_scalar_evolution, |
2012 | - instantiate_parameters. |
2013 | */ |
2014 | |
2015 | /* Helper recursive function. */ |
2016 | |
2017 | static tree |
2018 | analyze_scalar_evolution_1 (struct loop *loop, tree var) |
2019 | { |
2020 | gimple *def; |
2021 | basic_block bb; |
2022 | struct loop *def_loop; |
2023 | tree res; |
2024 | |
2025 | if (TREE_CODE (var) != SSA_NAME) |
2026 | return interpret_expr (loop, NULL, var); |
2027 | |
2028 | def = SSA_NAME_DEF_STMT (var); |
2029 | bb = gimple_bb (def); |
2030 | def_loop = bb->loop_father; |
2031 | |
2032 | if (!flow_bb_inside_loop_p (loop, bb)) |
2033 | { |
2034 | /* Keep symbolic form, but look through obvious copies for constants. */ |
2035 | res = follow_copies_to_constant (var); |
2036 | goto set_and_end; |
2037 | } |
2038 | |
2039 | if (loop != def_loop) |
2040 | { |
2041 | res = analyze_scalar_evolution_1 (def_loop, var); |
2042 | struct loop *loop_to_skip = superloop_at_depth (def_loop, |
2043 | loop_depth (loop) + 1); |
2044 | res = compute_overall_effect_of_inner_loop (loop_to_skip, res); |
2045 | if (chrec_contains_symbols_defined_in_loop (res, loop->num)) |
2046 | res = analyze_scalar_evolution_1 (loop, res); |
2047 | goto set_and_end; |
2048 | } |
2049 | |
2050 | switch (gimple_code (def)) |
2051 | { |
2052 | case GIMPLE_ASSIGN: |
2053 | res = interpret_gimple_assign (loop, def); |
2054 | break; |
2055 | |
2056 | case GIMPLE_PHI: |
2057 | if (loop_phi_node_p (def)) |
2058 | res = interpret_loop_phi (loop, as_a <gphi *> (def)); |
2059 | else |
2060 | res = interpret_condition_phi (loop, as_a <gphi *> (def)); |
2061 | break; |
2062 | |
2063 | default: |
2064 | res = chrec_dont_know; |
2065 | break; |
2066 | } |
2067 | |
2068 | set_and_end: |
2069 | |
2070 | /* Keep the symbolic form. */ |
2071 | if (res == chrec_dont_know) |
2072 | res = var; |
2073 | |
2074 | if (loop == def_loop) |
2075 | set_scalar_evolution (block_before_loop (loop), var, res); |
2076 | |
2077 | return res; |
2078 | } |
2079 | |
2080 | /* Analyzes and returns the scalar evolution of the ssa_name VAR in |
2081 | LOOP. LOOP is the loop in which the variable is used. |
2082 | |
2083 | Example of use: having a pointer VAR to a SSA_NAME node, STMT a |
2084 | pointer to the statement that uses this variable, in order to |
2085 | determine the evolution function of the variable, use the following |
2086 | calls: |
2087 | |
2088 | loop_p loop = loop_containing_stmt (stmt); |
2089 | tree chrec_with_symbols = analyze_scalar_evolution (loop, var); |
2090 | tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols); |
2091 | */ |
2092 | |
2093 | tree |
2094 | analyze_scalar_evolution (struct loop *loop, tree var) |
2095 | { |
2096 | tree res; |
2097 | |
2098 | /* ??? Fix callers. */ |
2099 | if (! loop) |
2100 | return var; |
2101 | |
2102 | if (dump_file && (dump_flags & TDF_SCEV)) |
2103 | { |
2104 | fprintf (dump_file, "(analyze_scalar_evolution \n" ); |
2105 | fprintf (dump_file, " (loop_nb = %d)\n" , loop->num); |
2106 | fprintf (dump_file, " (scalar = " ); |
2107 | print_generic_expr (dump_file, var); |
2108 | fprintf (dump_file, ")\n" ); |
2109 | } |
2110 | |
2111 | res = get_scalar_evolution (block_before_loop (loop), var); |
2112 | if (res == chrec_not_analyzed_yet) |
2113 | res = analyze_scalar_evolution_1 (loop, var); |
2114 | |
2115 | if (dump_file && (dump_flags & TDF_SCEV)) |
2116 | fprintf (dump_file, ")\n" ); |
2117 | |
2118 | return res; |
2119 | } |
2120 | |
2121 | /* Analyzes and returns the scalar evolution of VAR address in LOOP. */ |
2122 | |
2123 | static tree |
2124 | analyze_scalar_evolution_for_address_of (struct loop *loop, tree var) |
2125 | { |
2126 | return analyze_scalar_evolution (loop, build_fold_addr_expr (var)); |
2127 | } |
2128 | |
2129 | /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to |
2130 | WRTO_LOOP (which should be a superloop of USE_LOOP) |
2131 | |
2132 | FOLDED_CASTS is set to true if resolve_mixers used |
2133 | chrec_convert_aggressive (TODO -- not really, we are way too conservative |
2134 | at the moment in order to keep things simple). |
2135 | |
2136 | To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following |
2137 | example: |
2138 | |
2139 | for (i = 0; i < 100; i++) -- loop 1 |
2140 | { |
2141 | for (j = 0; j < 100; j++) -- loop 2 |
2142 | { |
2143 | k1 = i; |
2144 | k2 = j; |
2145 | |
2146 | use2 (k1, k2); |
2147 | |
2148 | for (t = 0; t < 100; t++) -- loop 3 |
2149 | use3 (k1, k2); |
2150 | |
2151 | } |
2152 | use1 (k1, k2); |
2153 | } |
2154 | |
2155 | Both k1 and k2 are invariants in loop3, thus |
2156 | analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1 |
2157 | analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2 |
2158 | |
2159 | As they are invariant, it does not matter whether we consider their |
2160 | usage in loop 3 or loop 2, hence |
2161 | analyze_scalar_evolution_in_loop (loop2, loop3, k1) = |
2162 | analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i |
2163 | analyze_scalar_evolution_in_loop (loop2, loop3, k2) = |
2164 | analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2 |
2165 | |
2166 | Similarly for their evolutions with respect to loop 1. The values of K2 |
2167 | in the use in loop 2 vary independently on loop 1, thus we cannot express |
2168 | the evolution with respect to loop 1: |
2169 | analyze_scalar_evolution_in_loop (loop1, loop3, k1) = |
2170 | analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1 |
2171 | analyze_scalar_evolution_in_loop (loop1, loop3, k2) = |
2172 | analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know |
2173 | |
2174 | The value of k2 in the use in loop 1 is known, though: |
2175 | analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1 |
2176 | analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100 |
2177 | */ |
2178 | |
2179 | static tree |
2180 | analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop, |
2181 | tree version, bool *folded_casts) |
2182 | { |
2183 | bool val = false; |
2184 | tree ev = version, tmp; |
2185 | |
2186 | /* We cannot just do |
2187 | |
2188 | tmp = analyze_scalar_evolution (use_loop, version); |
2189 | ev = resolve_mixers (wrto_loop, tmp, folded_casts); |
2190 | |
2191 | as resolve_mixers would query the scalar evolution with respect to |
2192 | wrto_loop. For example, in the situation described in the function |
2193 | comment, suppose that wrto_loop = loop1, use_loop = loop3 and |
2194 | version = k2. Then |
2195 | |
2196 | analyze_scalar_evolution (use_loop, version) = k2 |
2197 | |
2198 | and resolve_mixers (loop1, k2, folded_casts) finds that the value of |
2199 | k2 in loop 1 is 100, which is a wrong result, since we are interested |
2200 | in the value in loop 3. |
2201 | |
2202 | Instead, we need to proceed from use_loop to wrto_loop loop by loop, |
2203 | each time checking that there is no evolution in the inner loop. */ |
2204 | |
2205 | if (folded_casts) |
2206 | *folded_casts = false; |
2207 | while (1) |
2208 | { |
2209 | tmp = analyze_scalar_evolution (use_loop, ev); |
2210 | ev = resolve_mixers (use_loop, tmp, folded_casts); |
2211 | |
2212 | if (use_loop == wrto_loop) |
2213 | return ev; |
2214 | |
2215 | /* If the value of the use changes in the inner loop, we cannot express |
2216 | its value in the outer loop (we might try to return interval chrec, |
2217 | but we do not have a user for it anyway) */ |
2218 | if (!no_evolution_in_loop_p (ev, use_loop->num, &val) |
2219 | || !val) |
2220 | return chrec_dont_know; |
2221 | |
2222 | use_loop = loop_outer (use_loop); |
2223 | } |
2224 | } |
2225 | |
2226 | |
2227 | /* Hashtable helpers for a temporary hash-table used when |
2228 | instantiating a CHREC or resolving mixers. For this use |
2229 | instantiated_below is always the same. */ |
2230 | |
2231 | struct instantiate_cache_type |
2232 | { |
2233 | htab_t map; |
2234 | vec<scev_info_str> entries; |
2235 | |
2236 | instantiate_cache_type () : map (NULL), entries (vNULL) {} |
2237 | ~instantiate_cache_type (); |
2238 | tree get (unsigned slot) { return entries[slot].chrec; } |
2239 | void set (unsigned slot, tree chrec) { entries[slot].chrec = chrec; } |
2240 | }; |
2241 | |
2242 | instantiate_cache_type::~instantiate_cache_type () |
2243 | { |
2244 | if (map != NULL) |
2245 | { |
2246 | htab_delete (map); |
2247 | entries.release (); |
2248 | } |
2249 | } |
2250 | |
2251 | /* Cache to avoid infinite recursion when instantiating an SSA name. |
2252 | Live during the outermost instantiate_scev or resolve_mixers call. */ |
2253 | static instantiate_cache_type *global_cache; |
2254 | |
2255 | /* Computes a hash function for database element ELT. */ |
2256 | |
2257 | static inline hashval_t |
2258 | hash_idx_scev_info (const void *elt_) |
2259 | { |
2260 | unsigned idx = ((size_t) elt_) - 2; |
2261 | return scev_info_hasher::hash (&global_cache->entries[idx]); |
2262 | } |
2263 | |
2264 | /* Compares database elements E1 and E2. */ |
2265 | |
2266 | static inline int |
2267 | eq_idx_scev_info (const void *e1, const void *e2) |
2268 | { |
2269 | unsigned idx1 = ((size_t) e1) - 2; |
2270 | return scev_info_hasher::equal (&global_cache->entries[idx1], |
2271 | (const scev_info_str *) e2); |
2272 | } |
2273 | |
2274 | /* Returns from CACHE the slot number of the cached chrec for NAME. */ |
2275 | |
2276 | static unsigned |
2277 | get_instantiated_value_entry (instantiate_cache_type &cache, |
2278 | tree name, edge instantiate_below) |
2279 | { |
2280 | if (!cache.map) |
2281 | { |
2282 | cache.map = htab_create (10, hash_idx_scev_info, eq_idx_scev_info, NULL); |
2283 | cache.entries.create (10); |
2284 | } |
2285 | |
2286 | scev_info_str e; |
2287 | e.name_version = SSA_NAME_VERSION (name); |
2288 | e.instantiated_below = instantiate_below->dest->index; |
2289 | void **slot = htab_find_slot_with_hash (cache.map, &e, |
2290 | scev_info_hasher::hash (&e), INSERT); |
2291 | if (!*slot) |
2292 | { |
2293 | e.chrec = chrec_not_analyzed_yet; |
2294 | *slot = (void *)(size_t)(cache.entries.length () + 2); |
2295 | cache.entries.safe_push (e); |
2296 | } |
2297 | |
2298 | return ((size_t)*slot) - 2; |
2299 | } |
2300 | |
2301 | |
2302 | /* Return the closed_loop_phi node for VAR. If there is none, return |
2303 | NULL_TREE. */ |
2304 | |
2305 | static tree |
2306 | loop_closed_phi_def (tree var) |
2307 | { |
2308 | struct loop *loop; |
2309 | edge exit; |
2310 | gphi *phi; |
2311 | gphi_iterator psi; |
2312 | |
2313 | if (var == NULL_TREE |
2314 | || TREE_CODE (var) != SSA_NAME) |
2315 | return NULL_TREE; |
2316 | |
2317 | loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var)); |
2318 | exit = single_exit (loop); |
2319 | if (!exit) |
2320 | return NULL_TREE; |
2321 | |
2322 | for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi)) |
2323 | { |
2324 | phi = psi.phi (); |
2325 | if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var) |
2326 | return PHI_RESULT (phi); |
2327 | } |
2328 | |
2329 | return NULL_TREE; |
2330 | } |
2331 | |
2332 | static tree instantiate_scev_r (edge, struct loop *, struct loop *, |
2333 | tree, bool *, int); |
2334 | |
2335 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2336 | and EVOLUTION_LOOP, that were left under a symbolic form. |
2337 | |
2338 | CHREC is an SSA_NAME to be instantiated. |
2339 | |
2340 | CACHE is the cache of already instantiated values. |
2341 | |
2342 | Variable pointed by FOLD_CONVERSIONS is set to TRUE when the |
2343 | conversions that may wrap in signed/pointer type are folded, as long |
2344 | as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL |
2345 | then we don't do such fold. |
2346 | |
2347 | SIZE_EXPR is used for computing the size of the expression to be |
2348 | instantiated, and to stop if it exceeds some limit. */ |
2349 | |
2350 | static tree |
2351 | instantiate_scev_name (edge instantiate_below, |
2352 | struct loop *evolution_loop, struct loop *inner_loop, |
2353 | tree chrec, |
2354 | bool *fold_conversions, |
2355 | int size_expr) |
2356 | { |
2357 | tree res; |
2358 | struct loop *def_loop; |
2359 | basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (chrec)); |
2360 | |
2361 | /* A parameter, nothing to do. */ |
2362 | if (!def_bb |
2363 | || !dominated_by_p (CDI_DOMINATORS, def_bb, instantiate_below->dest)) |
2364 | return chrec; |
2365 | |
2366 | /* We cache the value of instantiated variable to avoid exponential |
2367 | time complexity due to reevaluations. We also store the convenient |
2368 | value in the cache in order to prevent infinite recursion -- we do |
2369 | not want to instantiate the SSA_NAME if it is in a mixer |
2370 | structure. This is used for avoiding the instantiation of |
2371 | recursively defined functions, such as: |
2372 | |
2373 | | a_2 -> {0, +, 1, +, a_2}_1 */ |
2374 | |
2375 | unsigned si = get_instantiated_value_entry (*global_cache, |
2376 | chrec, instantiate_below); |
2377 | if (global_cache->get (si) != chrec_not_analyzed_yet) |
2378 | return global_cache->get (si); |
2379 | |
2380 | /* On recursion return chrec_dont_know. */ |
2381 | global_cache->set (si, chrec_dont_know); |
2382 | |
2383 | def_loop = find_common_loop (evolution_loop, def_bb->loop_father); |
2384 | |
2385 | if (! dominated_by_p (CDI_DOMINATORS, |
2386 | def_loop->header, instantiate_below->dest)) |
2387 | { |
2388 | gimple *def = SSA_NAME_DEF_STMT (chrec); |
2389 | if (gassign *ass = dyn_cast <gassign *> (def)) |
2390 | { |
2391 | switch (gimple_assign_rhs_class (ass)) |
2392 | { |
2393 | case GIMPLE_UNARY_RHS: |
2394 | { |
2395 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
2396 | inner_loop, gimple_assign_rhs1 (ass), |
2397 | fold_conversions, size_expr); |
2398 | if (op0 == chrec_dont_know) |
2399 | return chrec_dont_know; |
2400 | res = fold_build1 (gimple_assign_rhs_code (ass), |
2401 | TREE_TYPE (chrec), op0); |
2402 | break; |
2403 | } |
2404 | case GIMPLE_BINARY_RHS: |
2405 | { |
2406 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
2407 | inner_loop, gimple_assign_rhs1 (ass), |
2408 | fold_conversions, size_expr); |
2409 | if (op0 == chrec_dont_know) |
2410 | return chrec_dont_know; |
2411 | tree op1 = instantiate_scev_r (instantiate_below, evolution_loop, |
2412 | inner_loop, gimple_assign_rhs2 (ass), |
2413 | fold_conversions, size_expr); |
2414 | if (op1 == chrec_dont_know) |
2415 | return chrec_dont_know; |
2416 | res = fold_build2 (gimple_assign_rhs_code (ass), |
2417 | TREE_TYPE (chrec), op0, op1); |
2418 | break; |
2419 | } |
2420 | default: |
2421 | res = chrec_dont_know; |
2422 | } |
2423 | } |
2424 | else |
2425 | res = chrec_dont_know; |
2426 | global_cache->set (si, res); |
2427 | return res; |
2428 | } |
2429 | |
2430 | /* If the analysis yields a parametric chrec, instantiate the |
2431 | result again. */ |
2432 | res = analyze_scalar_evolution (def_loop, chrec); |
2433 | |
2434 | /* Don't instantiate default definitions. */ |
2435 | if (TREE_CODE (res) == SSA_NAME |
2436 | && SSA_NAME_IS_DEFAULT_DEF (res)) |
2437 | ; |
2438 | |
2439 | /* Don't instantiate loop-closed-ssa phi nodes. */ |
2440 | else if (TREE_CODE (res) == SSA_NAME |
2441 | && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res))) |
2442 | > loop_depth (def_loop)) |
2443 | { |
2444 | if (res == chrec) |
2445 | res = loop_closed_phi_def (chrec); |
2446 | else |
2447 | res = chrec; |
2448 | |
2449 | /* When there is no loop_closed_phi_def, it means that the |
2450 | variable is not used after the loop: try to still compute the |
2451 | value of the variable when exiting the loop. */ |
2452 | if (res == NULL_TREE) |
2453 | { |
2454 | loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (chrec)); |
2455 | res = analyze_scalar_evolution (loop, chrec); |
2456 | res = compute_overall_effect_of_inner_loop (loop, res); |
2457 | res = instantiate_scev_r (instantiate_below, evolution_loop, |
2458 | inner_loop, res, |
2459 | fold_conversions, size_expr); |
2460 | } |
2461 | else if (dominated_by_p (CDI_DOMINATORS, |
2462 | gimple_bb (SSA_NAME_DEF_STMT (res)), |
2463 | instantiate_below->dest)) |
2464 | res = chrec_dont_know; |
2465 | } |
2466 | |
2467 | else if (res != chrec_dont_know) |
2468 | { |
2469 | if (inner_loop |
2470 | && def_bb->loop_father != inner_loop |
2471 | && !flow_loop_nested_p (def_bb->loop_father, inner_loop)) |
2472 | /* ??? We could try to compute the overall effect of the loop here. */ |
2473 | res = chrec_dont_know; |
2474 | else |
2475 | res = instantiate_scev_r (instantiate_below, evolution_loop, |
2476 | inner_loop, res, |
2477 | fold_conversions, size_expr); |
2478 | } |
2479 | |
2480 | /* Store the correct value to the cache. */ |
2481 | global_cache->set (si, res); |
2482 | return res; |
2483 | } |
2484 | |
2485 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2486 | and EVOLUTION_LOOP, that were left under a symbolic form. |
2487 | |
2488 | CHREC is a polynomial chain of recurrence to be instantiated. |
2489 | |
2490 | CACHE is the cache of already instantiated values. |
2491 | |
2492 | Variable pointed by FOLD_CONVERSIONS is set to TRUE when the |
2493 | conversions that may wrap in signed/pointer type are folded, as long |
2494 | as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL |
2495 | then we don't do such fold. |
2496 | |
2497 | SIZE_EXPR is used for computing the size of the expression to be |
2498 | instantiated, and to stop if it exceeds some limit. */ |
2499 | |
2500 | static tree |
2501 | instantiate_scev_poly (edge instantiate_below, |
2502 | struct loop *evolution_loop, struct loop *, |
2503 | tree chrec, bool *fold_conversions, int size_expr) |
2504 | { |
2505 | tree op1; |
2506 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
2507 | get_chrec_loop (chrec), |
2508 | CHREC_LEFT (chrec), fold_conversions, |
2509 | size_expr); |
2510 | if (op0 == chrec_dont_know) |
2511 | return chrec_dont_know; |
2512 | |
2513 | op1 = instantiate_scev_r (instantiate_below, evolution_loop, |
2514 | get_chrec_loop (chrec), |
2515 | CHREC_RIGHT (chrec), fold_conversions, |
2516 | size_expr); |
2517 | if (op1 == chrec_dont_know) |
2518 | return chrec_dont_know; |
2519 | |
2520 | if (CHREC_LEFT (chrec) != op0 |
2521 | || CHREC_RIGHT (chrec) != op1) |
2522 | { |
2523 | op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL); |
2524 | chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1); |
2525 | } |
2526 | |
2527 | return chrec; |
2528 | } |
2529 | |
2530 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2531 | and EVOLUTION_LOOP, that were left under a symbolic form. |
2532 | |
2533 | "C0 CODE C1" is a binary expression of type TYPE to be instantiated. |
2534 | |
2535 | CACHE is the cache of already instantiated values. |
2536 | |
2537 | Variable pointed by FOLD_CONVERSIONS is set to TRUE when the |
2538 | conversions that may wrap in signed/pointer type are folded, as long |
2539 | as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL |
2540 | then we don't do such fold. |
2541 | |
2542 | SIZE_EXPR is used for computing the size of the expression to be |
2543 | instantiated, and to stop if it exceeds some limit. */ |
2544 | |
2545 | static tree |
2546 | instantiate_scev_binary (edge instantiate_below, |
2547 | struct loop *evolution_loop, struct loop *inner_loop, |
2548 | tree chrec, enum tree_code code, |
2549 | tree type, tree c0, tree c1, |
2550 | bool *fold_conversions, int size_expr) |
2551 | { |
2552 | tree op1; |
2553 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop, |
2554 | c0, fold_conversions, size_expr); |
2555 | if (op0 == chrec_dont_know) |
2556 | return chrec_dont_know; |
2557 | |
2558 | op1 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop, |
2559 | c1, fold_conversions, size_expr); |
2560 | if (op1 == chrec_dont_know) |
2561 | return chrec_dont_know; |
2562 | |
2563 | if (c0 != op0 |
2564 | || c1 != op1) |
2565 | { |
2566 | op0 = chrec_convert (type, op0, NULL); |
2567 | op1 = chrec_convert_rhs (type, op1, NULL); |
2568 | |
2569 | switch (code) |
2570 | { |
2571 | case POINTER_PLUS_EXPR: |
2572 | case PLUS_EXPR: |
2573 | return chrec_fold_plus (type, op0, op1); |
2574 | |
2575 | case MINUS_EXPR: |
2576 | return chrec_fold_minus (type, op0, op1); |
2577 | |
2578 | case MULT_EXPR: |
2579 | return chrec_fold_multiply (type, op0, op1); |
2580 | |
2581 | default: |
2582 | gcc_unreachable (); |
2583 | } |
2584 | } |
2585 | |
2586 | return chrec ? chrec : fold_build2 (code, type, c0, c1); |
2587 | } |
2588 | |
2589 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2590 | and EVOLUTION_LOOP, that were left under a symbolic form. |
2591 | |
2592 | "CHREC" that stands for a convert expression "(TYPE) OP" is to be |
2593 | instantiated. |
2594 | |
2595 | CACHE is the cache of already instantiated values. |
2596 | |
2597 | Variable pointed by FOLD_CONVERSIONS is set to TRUE when the |
2598 | conversions that may wrap in signed/pointer type are folded, as long |
2599 | as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL |
2600 | then we don't do such fold. |
2601 | |
2602 | SIZE_EXPR is used for computing the size of the expression to be |
2603 | instantiated, and to stop if it exceeds some limit. */ |
2604 | |
2605 | static tree |
2606 | instantiate_scev_convert (edge instantiate_below, |
2607 | struct loop *evolution_loop, struct loop *inner_loop, |
2608 | tree chrec, tree type, tree op, |
2609 | bool *fold_conversions, int size_expr) |
2610 | { |
2611 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
2612 | inner_loop, op, |
2613 | fold_conversions, size_expr); |
2614 | |
2615 | if (op0 == chrec_dont_know) |
2616 | return chrec_dont_know; |
2617 | |
2618 | if (fold_conversions) |
2619 | { |
2620 | tree tmp = chrec_convert_aggressive (type, op0, fold_conversions); |
2621 | if (tmp) |
2622 | return tmp; |
2623 | |
2624 | /* If we used chrec_convert_aggressive, we can no longer assume that |
2625 | signed chrecs do not overflow, as chrec_convert does, so avoid |
2626 | calling it in that case. */ |
2627 | if (*fold_conversions) |
2628 | { |
2629 | if (chrec && op0 == op) |
2630 | return chrec; |
2631 | |
2632 | return fold_convert (type, op0); |
2633 | } |
2634 | } |
2635 | |
2636 | return chrec_convert (type, op0, NULL); |
2637 | } |
2638 | |
2639 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2640 | and EVOLUTION_LOOP, that were left under a symbolic form. |
2641 | |
2642 | CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated. |
2643 | Handle ~X as -1 - X. |
2644 | Handle -X as -1 * X. |
2645 | |
2646 | CACHE is the cache of already instantiated values. |
2647 | |
2648 | Variable pointed by FOLD_CONVERSIONS is set to TRUE when the |
2649 | conversions that may wrap in signed/pointer type are folded, as long |
2650 | as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL |
2651 | then we don't do such fold. |
2652 | |
2653 | SIZE_EXPR is used for computing the size of the expression to be |
2654 | instantiated, and to stop if it exceeds some limit. */ |
2655 | |
2656 | static tree |
2657 | instantiate_scev_not (edge instantiate_below, |
2658 | struct loop *evolution_loop, struct loop *inner_loop, |
2659 | tree chrec, |
2660 | enum tree_code code, tree type, tree op, |
2661 | bool *fold_conversions, int size_expr) |
2662 | { |
2663 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
2664 | inner_loop, op, |
2665 | fold_conversions, size_expr); |
2666 | |
2667 | if (op0 == chrec_dont_know) |
2668 | return chrec_dont_know; |
2669 | |
2670 | if (op != op0) |
2671 | { |
2672 | op0 = chrec_convert (type, op0, NULL); |
2673 | |
2674 | switch (code) |
2675 | { |
2676 | case BIT_NOT_EXPR: |
2677 | return chrec_fold_minus |
2678 | (type, fold_convert (type, integer_minus_one_node), op0); |
2679 | |
2680 | case NEGATE_EXPR: |
2681 | return chrec_fold_multiply |
2682 | (type, fold_convert (type, integer_minus_one_node), op0); |
2683 | |
2684 | default: |
2685 | gcc_unreachable (); |
2686 | } |
2687 | } |
2688 | |
2689 | return chrec ? chrec : fold_build1 (code, type, op0); |
2690 | } |
2691 | |
2692 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2693 | and EVOLUTION_LOOP, that were left under a symbolic form. |
2694 | |
2695 | CHREC is the scalar evolution to instantiate. |
2696 | |
2697 | CACHE is the cache of already instantiated values. |
2698 | |
2699 | Variable pointed by FOLD_CONVERSIONS is set to TRUE when the |
2700 | conversions that may wrap in signed/pointer type are folded, as long |
2701 | as the value of the chrec is preserved. If FOLD_CONVERSIONS is NULL |
2702 | then we don't do such fold. |
2703 | |
2704 | SIZE_EXPR is used for computing the size of the expression to be |
2705 | instantiated, and to stop if it exceeds some limit. */ |
2706 | |
2707 | static tree |
2708 | instantiate_scev_r (edge instantiate_below, |
2709 | struct loop *evolution_loop, struct loop *inner_loop, |
2710 | tree chrec, |
2711 | bool *fold_conversions, int size_expr) |
2712 | { |
2713 | /* Give up if the expression is larger than the MAX that we allow. */ |
2714 | if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE)) |
2715 | return chrec_dont_know; |
2716 | |
2717 | if (chrec == NULL_TREE |
2718 | || automatically_generated_chrec_p (chrec) |
2719 | || is_gimple_min_invariant (chrec)) |
2720 | return chrec; |
2721 | |
2722 | switch (TREE_CODE (chrec)) |
2723 | { |
2724 | case SSA_NAME: |
2725 | return instantiate_scev_name (instantiate_below, evolution_loop, |
2726 | inner_loop, chrec, |
2727 | fold_conversions, size_expr); |
2728 | |
2729 | case POLYNOMIAL_CHREC: |
2730 | return instantiate_scev_poly (instantiate_below, evolution_loop, |
2731 | inner_loop, chrec, |
2732 | fold_conversions, size_expr); |
2733 | |
2734 | case POINTER_PLUS_EXPR: |
2735 | case PLUS_EXPR: |
2736 | case MINUS_EXPR: |
2737 | case MULT_EXPR: |
2738 | return instantiate_scev_binary (instantiate_below, evolution_loop, |
2739 | inner_loop, chrec, |
2740 | TREE_CODE (chrec), chrec_type (chrec), |
2741 | TREE_OPERAND (chrec, 0), |
2742 | TREE_OPERAND (chrec, 1), |
2743 | fold_conversions, size_expr); |
2744 | |
2745 | CASE_CONVERT: |
2746 | return instantiate_scev_convert (instantiate_below, evolution_loop, |
2747 | inner_loop, chrec, |
2748 | TREE_TYPE (chrec), TREE_OPERAND (chrec, 0), |
2749 | fold_conversions, size_expr); |
2750 | |
2751 | case NEGATE_EXPR: |
2752 | case BIT_NOT_EXPR: |
2753 | return instantiate_scev_not (instantiate_below, evolution_loop, |
2754 | inner_loop, chrec, |
2755 | TREE_CODE (chrec), TREE_TYPE (chrec), |
2756 | TREE_OPERAND (chrec, 0), |
2757 | fold_conversions, size_expr); |
2758 | |
2759 | case ADDR_EXPR: |
2760 | if (is_gimple_min_invariant (chrec)) |
2761 | return chrec; |
2762 | /* Fallthru. */ |
2763 | case SCEV_NOT_KNOWN: |
2764 | return chrec_dont_know; |
2765 | |
2766 | case SCEV_KNOWN: |
2767 | return chrec_known; |
2768 | |
2769 | default: |
2770 | if (CONSTANT_CLASS_P (chrec)) |
2771 | return chrec; |
2772 | return chrec_dont_know; |
2773 | } |
2774 | } |
2775 | |
2776 | /* Analyze all the parameters of the chrec that were left under a |
2777 | symbolic form. INSTANTIATE_BELOW is the basic block that stops the |
2778 | recursive instantiation of parameters: a parameter is a variable |
2779 | that is defined in a basic block that dominates INSTANTIATE_BELOW or |
2780 | a function parameter. */ |
2781 | |
2782 | tree |
2783 | instantiate_scev (edge instantiate_below, struct loop *evolution_loop, |
2784 | tree chrec) |
2785 | { |
2786 | tree res; |
2787 | |
2788 | if (dump_file && (dump_flags & TDF_SCEV)) |
2789 | { |
2790 | fprintf (dump_file, "(instantiate_scev \n" ); |
2791 | fprintf (dump_file, " (instantiate_below = %d -> %d)\n" , |
2792 | instantiate_below->src->index, instantiate_below->dest->index); |
2793 | if (evolution_loop) |
2794 | fprintf (dump_file, " (evolution_loop = %d)\n" , evolution_loop->num); |
2795 | fprintf (dump_file, " (chrec = " ); |
2796 | print_generic_expr (dump_file, chrec); |
2797 | fprintf (dump_file, ")\n" ); |
2798 | } |
2799 | |
2800 | bool destr = false; |
2801 | if (!global_cache) |
2802 | { |
2803 | global_cache = new instantiate_cache_type; |
2804 | destr = true; |
2805 | } |
2806 | |
2807 | res = instantiate_scev_r (instantiate_below, evolution_loop, |
2808 | NULL, chrec, NULL, 0); |
2809 | |
2810 | if (destr) |
2811 | { |
2812 | delete global_cache; |
2813 | global_cache = NULL; |
2814 | } |
2815 | |
2816 | if (dump_file && (dump_flags & TDF_SCEV)) |
2817 | { |
2818 | fprintf (dump_file, " (res = " ); |
2819 | print_generic_expr (dump_file, res); |
2820 | fprintf (dump_file, "))\n" ); |
2821 | } |
2822 | |
2823 | return res; |
2824 | } |
2825 | |
2826 | /* Similar to instantiate_parameters, but does not introduce the |
2827 | evolutions in outer loops for LOOP invariants in CHREC, and does not |
2828 | care about causing overflows, as long as they do not affect value |
2829 | of an expression. */ |
2830 | |
2831 | tree |
2832 | resolve_mixers (struct loop *loop, tree chrec, bool *folded_casts) |
2833 | { |
2834 | bool destr = false; |
2835 | bool fold_conversions = false; |
2836 | if (!global_cache) |
2837 | { |
2838 | global_cache = new instantiate_cache_type; |
2839 | destr = true; |
2840 | } |
2841 | |
2842 | tree ret = instantiate_scev_r (loop_preheader_edge (loop), loop, NULL, |
2843 | chrec, &fold_conversions, 0); |
2844 | |
2845 | if (folded_casts && !*folded_casts) |
2846 | *folded_casts = fold_conversions; |
2847 | |
2848 | if (destr) |
2849 | { |
2850 | delete global_cache; |
2851 | global_cache = NULL; |
2852 | } |
2853 | |
2854 | return ret; |
2855 | } |
2856 | |
2857 | /* Entry point for the analysis of the number of iterations pass. |
2858 | This function tries to safely approximate the number of iterations |
2859 | the loop will run. When this property is not decidable at compile |
2860 | time, the result is chrec_dont_know. Otherwise the result is a |
2861 | scalar or a symbolic parameter. When the number of iterations may |
2862 | be equal to zero and the property cannot be determined at compile |
2863 | time, the result is a COND_EXPR that represents in a symbolic form |
2864 | the conditions under which the number of iterations is not zero. |
2865 | |
2866 | Example of analysis: suppose that the loop has an exit condition: |
2867 | |
2868 | "if (b > 49) goto end_loop;" |
2869 | |
2870 | and that in a previous analysis we have determined that the |
2871 | variable 'b' has an evolution function: |
2872 | |
2873 | "EF = {23, +, 5}_2". |
2874 | |
2875 | When we evaluate the function at the point 5, i.e. the value of the |
2876 | variable 'b' after 5 iterations in the loop, we have EF (5) = 48, |
2877 | and EF (6) = 53. In this case the value of 'b' on exit is '53' and |
2878 | the loop body has been executed 6 times. */ |
2879 | |
2880 | tree |
2881 | number_of_latch_executions (struct loop *loop) |
2882 | { |
2883 | edge exit; |
2884 | struct tree_niter_desc niter_desc; |
2885 | tree may_be_zero; |
2886 | tree res; |
2887 | |
2888 | /* Determine whether the number of iterations in loop has already |
2889 | been computed. */ |
2890 | res = loop->nb_iterations; |
2891 | if (res) |
2892 | return res; |
2893 | |
2894 | may_be_zero = NULL_TREE; |
2895 | |
2896 | if (dump_file && (dump_flags & TDF_SCEV)) |
2897 | fprintf (dump_file, "(number_of_iterations_in_loop = \n" ); |
2898 | |
2899 | res = chrec_dont_know; |
2900 | exit = single_exit (loop); |
2901 | |
2902 | if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false)) |
2903 | { |
2904 | may_be_zero = niter_desc.may_be_zero; |
2905 | res = niter_desc.niter; |
2906 | } |
2907 | |
2908 | if (res == chrec_dont_know |
2909 | || !may_be_zero |
2910 | || integer_zerop (may_be_zero)) |
2911 | ; |
2912 | else if (integer_nonzerop (may_be_zero)) |
2913 | res = build_int_cst (TREE_TYPE (res), 0); |
2914 | |
2915 | else if (COMPARISON_CLASS_P (may_be_zero)) |
2916 | res = fold_build3 (COND_EXPR, TREE_TYPE (res), may_be_zero, |
2917 | build_int_cst (TREE_TYPE (res), 0), res); |
2918 | else |
2919 | res = chrec_dont_know; |
2920 | |
2921 | if (dump_file && (dump_flags & TDF_SCEV)) |
2922 | { |
2923 | fprintf (dump_file, " (set_nb_iterations_in_loop = " ); |
2924 | print_generic_expr (dump_file, res); |
2925 | fprintf (dump_file, "))\n" ); |
2926 | } |
2927 | |
2928 | loop->nb_iterations = res; |
2929 | return res; |
2930 | } |
2931 | |
2932 | |
2933 | /* Counters for the stats. */ |
2934 | |
2935 | struct chrec_stats |
2936 | { |
2937 | unsigned nb_chrecs; |
2938 | unsigned nb_affine; |
2939 | unsigned nb_affine_multivar; |
2940 | unsigned nb_higher_poly; |
2941 | unsigned nb_chrec_dont_know; |
2942 | unsigned nb_undetermined; |
2943 | }; |
2944 | |
2945 | /* Reset the counters. */ |
2946 | |
2947 | static inline void |
2948 | reset_chrecs_counters (struct chrec_stats *stats) |
2949 | { |
2950 | stats->nb_chrecs = 0; |
2951 | stats->nb_affine = 0; |
2952 | stats->nb_affine_multivar = 0; |
2953 | stats->nb_higher_poly = 0; |
2954 | stats->nb_chrec_dont_know = 0; |
2955 | stats->nb_undetermined = 0; |
2956 | } |
2957 | |
2958 | /* Dump the contents of a CHREC_STATS structure. */ |
2959 | |
2960 | static void |
2961 | dump_chrecs_stats (FILE *file, struct chrec_stats *stats) |
2962 | { |
2963 | fprintf (file, "\n(\n" ); |
2964 | fprintf (file, "-----------------------------------------\n" ); |
2965 | fprintf (file, "%d\taffine univariate chrecs\n" , stats->nb_affine); |
2966 | fprintf (file, "%d\taffine multivariate chrecs\n" , stats->nb_affine_multivar); |
2967 | fprintf (file, "%d\tdegree greater than 2 polynomials\n" , |
2968 | stats->nb_higher_poly); |
2969 | fprintf (file, "%d\tchrec_dont_know chrecs\n" , stats->nb_chrec_dont_know); |
2970 | fprintf (file, "-----------------------------------------\n" ); |
2971 | fprintf (file, "%d\ttotal chrecs\n" , stats->nb_chrecs); |
2972 | fprintf (file, "%d\twith undetermined coefficients\n" , |
2973 | stats->nb_undetermined); |
2974 | fprintf (file, "-----------------------------------------\n" ); |
2975 | fprintf (file, "%d\tchrecs in the scev database\n" , |
2976 | (int) scalar_evolution_info->elements ()); |
2977 | fprintf (file, "%d\tsets in the scev database\n" , nb_set_scev); |
2978 | fprintf (file, "%d\tgets in the scev database\n" , nb_get_scev); |
2979 | fprintf (file, "-----------------------------------------\n" ); |
2980 | fprintf (file, ")\n\n" ); |
2981 | } |
2982 | |
2983 | /* Gather statistics about CHREC. */ |
2984 | |
2985 | static void |
2986 | gather_chrec_stats (tree chrec, struct chrec_stats *stats) |
2987 | { |
2988 | if (dump_file && (dump_flags & TDF_STATS)) |
2989 | { |
2990 | fprintf (dump_file, "(classify_chrec " ); |
2991 | print_generic_expr (dump_file, chrec); |
2992 | fprintf (dump_file, "\n" ); |
2993 | } |
2994 | |
2995 | stats->nb_chrecs++; |
2996 | |
2997 | if (chrec == NULL_TREE) |
2998 | { |
2999 | stats->nb_undetermined++; |
3000 | return; |
3001 | } |
3002 | |
3003 | switch (TREE_CODE (chrec)) |
3004 | { |
3005 | case POLYNOMIAL_CHREC: |
3006 | if (evolution_function_is_affine_p (chrec)) |
3007 | { |
3008 | if (dump_file && (dump_flags & TDF_STATS)) |
3009 | fprintf (dump_file, " affine_univariate\n" ); |
3010 | stats->nb_affine++; |
3011 | } |
3012 | else if (evolution_function_is_affine_multivariate_p (chrec, 0)) |
3013 | { |
3014 | if (dump_file && (dump_flags & TDF_STATS)) |
3015 | fprintf (dump_file, " affine_multivariate\n" ); |
3016 | stats->nb_affine_multivar++; |
3017 | } |
3018 | else |
3019 | { |
3020 | if (dump_file && (dump_flags & TDF_STATS)) |
3021 | fprintf (dump_file, " higher_degree_polynomial\n" ); |
3022 | stats->nb_higher_poly++; |
3023 | } |
3024 | |
3025 | break; |
3026 | |
3027 | default: |
3028 | break; |
3029 | } |
3030 | |
3031 | if (chrec_contains_undetermined (chrec)) |
3032 | { |
3033 | if (dump_file && (dump_flags & TDF_STATS)) |
3034 | fprintf (dump_file, " undetermined\n" ); |
3035 | stats->nb_undetermined++; |
3036 | } |
3037 | |
3038 | if (dump_file && (dump_flags & TDF_STATS)) |
3039 | fprintf (dump_file, ")\n" ); |
3040 | } |
3041 | |
3042 | /* Classify the chrecs of the whole database. */ |
3043 | |
3044 | void |
3045 | gather_stats_on_scev_database (void) |
3046 | { |
3047 | struct chrec_stats stats; |
3048 | |
3049 | if (!dump_file) |
3050 | return; |
3051 | |
3052 | reset_chrecs_counters (&stats); |
3053 | |
3054 | hash_table<scev_info_hasher>::iterator iter; |
3055 | scev_info_str *elt; |
3056 | FOR_EACH_HASH_TABLE_ELEMENT (*scalar_evolution_info, elt, scev_info_str *, |
3057 | iter) |
3058 | gather_chrec_stats (elt->chrec, &stats); |
3059 | |
3060 | dump_chrecs_stats (dump_file, &stats); |
3061 | } |
3062 | |
3063 | |
3064 | |
3065 | /* Initializer. */ |
3066 | |
3067 | static void |
3068 | initialize_scalar_evolutions_analyzer (void) |
3069 | { |
3070 | /* The elements below are unique. */ |
3071 | if (chrec_dont_know == NULL_TREE) |
3072 | { |
3073 | chrec_not_analyzed_yet = NULL_TREE; |
3074 | chrec_dont_know = make_node (SCEV_NOT_KNOWN); |
3075 | chrec_known = make_node (SCEV_KNOWN); |
3076 | TREE_TYPE (chrec_dont_know) = void_type_node; |
3077 | TREE_TYPE (chrec_known) = void_type_node; |
3078 | } |
3079 | } |
3080 | |
3081 | /* Initialize the analysis of scalar evolutions for LOOPS. */ |
3082 | |
3083 | void |
3084 | scev_initialize (void) |
3085 | { |
3086 | struct loop *loop; |
3087 | |
3088 | gcc_assert (! scev_initialized_p ()); |
3089 | |
3090 | scalar_evolution_info = hash_table<scev_info_hasher>::create_ggc (100); |
3091 | |
3092 | initialize_scalar_evolutions_analyzer (); |
3093 | |
3094 | FOR_EACH_LOOP (loop, 0) |
3095 | { |
3096 | loop->nb_iterations = NULL_TREE; |
3097 | } |
3098 | } |
3099 | |
3100 | /* Return true if SCEV is initialized. */ |
3101 | |
3102 | bool |
3103 | scev_initialized_p (void) |
3104 | { |
3105 | return scalar_evolution_info != NULL; |
3106 | } |
3107 | |
3108 | /* Cleans up the information cached by the scalar evolutions analysis |
3109 | in the hash table. */ |
3110 | |
3111 | void |
3112 | scev_reset_htab (void) |
3113 | { |
3114 | if (!scalar_evolution_info) |
3115 | return; |
3116 | |
3117 | scalar_evolution_info->empty (); |
3118 | } |
3119 | |
3120 | /* Cleans up the information cached by the scalar evolutions analysis |
3121 | in the hash table and in the loop->nb_iterations. */ |
3122 | |
3123 | void |
3124 | scev_reset (void) |
3125 | { |
3126 | struct loop *loop; |
3127 | |
3128 | scev_reset_htab (); |
3129 | |
3130 | FOR_EACH_LOOP (loop, 0) |
3131 | { |
3132 | loop->nb_iterations = NULL_TREE; |
3133 | } |
3134 | } |
3135 | |
3136 | /* Return true if the IV calculation in TYPE can overflow based on the knowledge |
3137 | of the upper bound on the number of iterations of LOOP, the BASE and STEP |
3138 | of IV. |
3139 | |
3140 | We do not use information whether TYPE can overflow so it is safe to |
3141 | use this test even for derived IVs not computed every iteration or |
3142 | hypotetical IVs to be inserted into code. */ |
3143 | |
3144 | bool |
3145 | iv_can_overflow_p (struct loop *loop, tree type, tree base, tree step) |
3146 | { |
3147 | widest_int nit; |
3148 | wide_int base_min, base_max, step_min, step_max, type_min, type_max; |
3149 | signop sgn = TYPE_SIGN (type); |
3150 | |
3151 | if (integer_zerop (step)) |
3152 | return false; |
3153 | |
3154 | if (TREE_CODE (base) == INTEGER_CST) |
3155 | base_min = base_max = wi::to_wide (base); |
3156 | else if (TREE_CODE (base) == SSA_NAME |
3157 | && INTEGRAL_TYPE_P (TREE_TYPE (base)) |
3158 | && get_range_info (base, &base_min, &base_max) == VR_RANGE) |
3159 | ; |
3160 | else |
3161 | return true; |
3162 | |
3163 | if (TREE_CODE (step) == INTEGER_CST) |
3164 | step_min = step_max = wi::to_wide (step); |
3165 | else if (TREE_CODE (step) == SSA_NAME |
3166 | && INTEGRAL_TYPE_P (TREE_TYPE (step)) |
3167 | && get_range_info (step, &step_min, &step_max) == VR_RANGE) |
3168 | ; |
3169 | else |
3170 | return true; |
3171 | |
3172 | if (!get_max_loop_iterations (loop, &nit)) |
3173 | return true; |
3174 | |
3175 | type_min = wi::min_value (type); |
3176 | type_max = wi::max_value (type); |
3177 | |
3178 | /* Just sanity check that we don't see values out of the range of the type. |
3179 | In this case the arithmetics bellow would overflow. */ |
3180 | gcc_checking_assert (wi::ge_p (base_min, type_min, sgn) |
3181 | && wi::le_p (base_max, type_max, sgn)); |
3182 | |
3183 | /* Account the possible increment in the last ieration. */ |
3184 | bool overflow = false; |
3185 | nit = wi::add (nit, 1, SIGNED, &overflow); |
3186 | if (overflow) |
3187 | return true; |
3188 | |
3189 | /* NIT is typeless and can exceed the precision of the type. In this case |
3190 | overflow is always possible, because we know STEP is non-zero. */ |
3191 | if (wi::min_precision (nit, UNSIGNED) > TYPE_PRECISION (type)) |
3192 | return true; |
3193 | wide_int nit2 = wide_int::from (nit, TYPE_PRECISION (type), UNSIGNED); |
3194 | |
3195 | /* If step can be positive, check that nit*step <= type_max-base. |
3196 | This can be done by unsigned arithmetic and we only need to watch overflow |
3197 | in the multiplication. The right hand side can always be represented in |
3198 | the type. */ |
3199 | if (sgn == UNSIGNED || !wi::neg_p (step_max)) |
3200 | { |
3201 | bool overflow = false; |
3202 | if (wi::gtu_p (wi::mul (step_max, nit2, UNSIGNED, &overflow), |
3203 | type_max - base_max) |
3204 | || overflow) |
3205 | return true; |
3206 | } |
3207 | /* If step can be negative, check that nit*(-step) <= base_min-type_min. */ |
3208 | if (sgn == SIGNED && wi::neg_p (step_min)) |
3209 | { |
3210 | bool overflow = false, overflow2 = false; |
3211 | if (wi::gtu_p (wi::mul (wi::neg (step_min, &overflow2), |
3212 | nit2, UNSIGNED, &overflow), |
3213 | base_min - type_min) |
3214 | || overflow || overflow2) |
3215 | return true; |
3216 | } |
3217 | |
3218 | return false; |
3219 | } |
3220 | |
3221 | /* Given EV with form of "(type) {inner_base, inner_step}_loop", this |
3222 | function tries to derive condition under which it can be simplified |
3223 | into "{(type)inner_base, (type)inner_step}_loop". The condition is |
3224 | the maximum number that inner iv can iterate. */ |
3225 | |
3226 | static tree |
3227 | derive_simple_iv_with_niters (tree ev, tree *niters) |
3228 | { |
3229 | if (!CONVERT_EXPR_P (ev)) |
3230 | return ev; |
3231 | |
3232 | tree inner_ev = TREE_OPERAND (ev, 0); |
3233 | if (TREE_CODE (inner_ev) != POLYNOMIAL_CHREC) |
3234 | return ev; |
3235 | |
3236 | tree init = CHREC_LEFT (inner_ev); |
3237 | tree step = CHREC_RIGHT (inner_ev); |
3238 | if (TREE_CODE (init) != INTEGER_CST |
3239 | || TREE_CODE (step) != INTEGER_CST || integer_zerop (step)) |
3240 | return ev; |
3241 | |
3242 | tree type = TREE_TYPE (ev); |
3243 | tree inner_type = TREE_TYPE (inner_ev); |
3244 | if (TYPE_PRECISION (inner_type) >= TYPE_PRECISION (type)) |
3245 | return ev; |
3246 | |
3247 | /* Type conversion in "(type) {inner_base, inner_step}_loop" can be |
3248 | folded only if inner iv won't overflow. We compute the maximum |
3249 | number the inner iv can iterate before overflowing and return the |
3250 | simplified affine iv. */ |
3251 | tree delta; |
3252 | init = fold_convert (type, init); |
3253 | step = fold_convert (type, step); |
3254 | ev = build_polynomial_chrec (CHREC_VARIABLE (inner_ev), init, step); |
3255 | if (tree_int_cst_sign_bit (step)) |
3256 | { |
3257 | tree bound = lower_bound_in_type (inner_type, inner_type); |
3258 | delta = fold_build2 (MINUS_EXPR, type, init, fold_convert (type, bound)); |
3259 | step = fold_build1 (NEGATE_EXPR, type, step); |
3260 | } |
3261 | else |
3262 | { |
3263 | tree bound = upper_bound_in_type (inner_type, inner_type); |
3264 | delta = fold_build2 (MINUS_EXPR, type, fold_convert (type, bound), init); |
3265 | } |
3266 | *niters = fold_build2 (FLOOR_DIV_EXPR, type, delta, step); |
3267 | return ev; |
3268 | } |
3269 | |
3270 | /* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with |
3271 | respect to WRTO_LOOP and returns its base and step in IV if possible |
3272 | (see analyze_scalar_evolution_in_loop for more details on USE_LOOP |
3273 | and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be |
3274 | invariant in LOOP. Otherwise we require it to be an integer constant. |
3275 | |
3276 | IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g. |
3277 | because it is computed in signed arithmetics). Consequently, adding an |
3278 | induction variable |
3279 | |
3280 | for (i = IV->base; ; i += IV->step) |
3281 | |
3282 | is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is |
3283 | false for the type of the induction variable, or you can prove that i does |
3284 | not wrap by some other argument. Otherwise, this might introduce undefined |
3285 | behavior, and |
3286 | |
3287 | i = iv->base; |
3288 | for (; ; i = (type) ((unsigned type) i + (unsigned type) iv->step)) |
3289 | |
3290 | must be used instead. |
3291 | |
3292 | When IV_NITERS is not NULL, this function also checks case in which OP |
3293 | is a conversion of an inner simple iv of below form: |
3294 | |
3295 | (outer_type){inner_base, inner_step}_loop. |
3296 | |
3297 | If type of inner iv has smaller precision than outer_type, it can't be |
3298 | folded into {(outer_type)inner_base, (outer_type)inner_step}_loop because |
3299 | the inner iv could overflow/wrap. In this case, we derive a condition |
3300 | under which the inner iv won't overflow/wrap and do the simplification. |
3301 | The derived condition normally is the maximum number the inner iv can |
3302 | iterate, and will be stored in IV_NITERS. This is useful in loop niter |
3303 | analysis, to derive break conditions when a loop must terminate, when is |
3304 | infinite. */ |
3305 | |
3306 | bool |
3307 | simple_iv_with_niters (struct loop *wrto_loop, struct loop *use_loop, |
3308 | tree op, affine_iv *iv, tree *iv_niters, |
3309 | bool allow_nonconstant_step) |
3310 | { |
3311 | enum tree_code code; |
3312 | tree type, ev, base, e; |
3313 | wide_int extreme; |
3314 | bool folded_casts, overflow; |
3315 | |
3316 | iv->base = NULL_TREE; |
3317 | iv->step = NULL_TREE; |
3318 | iv->no_overflow = false; |
3319 | |
3320 | type = TREE_TYPE (op); |
3321 | if (!POINTER_TYPE_P (type) |
3322 | && !INTEGRAL_TYPE_P (type)) |
3323 | return false; |
3324 | |
3325 | ev = analyze_scalar_evolution_in_loop (wrto_loop, use_loop, op, |
3326 | &folded_casts); |
3327 | if (chrec_contains_undetermined (ev) |
3328 | || chrec_contains_symbols_defined_in_loop (ev, wrto_loop->num)) |
3329 | return false; |
3330 | |
3331 | if (tree_does_not_contain_chrecs (ev)) |
3332 | { |
3333 | iv->base = ev; |
3334 | iv->step = build_int_cst (TREE_TYPE (ev), 0); |
3335 | iv->no_overflow = true; |
3336 | return true; |
3337 | } |
3338 | |
3339 | /* If we can derive valid scalar evolution with assumptions. */ |
3340 | if (iv_niters && TREE_CODE (ev) != POLYNOMIAL_CHREC) |
3341 | ev = derive_simple_iv_with_niters (ev, iv_niters); |
3342 | |
3343 | if (TREE_CODE (ev) != POLYNOMIAL_CHREC) |
3344 | return false; |
3345 | |
3346 | if (CHREC_VARIABLE (ev) != (unsigned) wrto_loop->num) |
3347 | return false; |
3348 | |
3349 | iv->step = CHREC_RIGHT (ev); |
3350 | if ((!allow_nonconstant_step && TREE_CODE (iv->step) != INTEGER_CST) |
3351 | || tree_contains_chrecs (iv->step, NULL)) |
3352 | return false; |
3353 | |
3354 | iv->base = CHREC_LEFT (ev); |
3355 | if (tree_contains_chrecs (iv->base, NULL)) |
3356 | return false; |
3357 | |
3358 | iv->no_overflow = !folded_casts && nowrap_type_p (type); |
3359 | |
3360 | if (!iv->no_overflow |
3361 | && !iv_can_overflow_p (wrto_loop, type, iv->base, iv->step)) |
3362 | iv->no_overflow = true; |
3363 | |
3364 | /* Try to simplify iv base: |
3365 | |
3366 | (signed T) ((unsigned T)base + step) ;; TREE_TYPE (base) == signed T |
3367 | == (signed T)(unsigned T)base + step |
3368 | == base + step |
3369 | |
3370 | If we can prove operation (base + step) doesn't overflow or underflow. |
3371 | Specifically, we try to prove below conditions are satisfied: |
3372 | |
3373 | base <= UPPER_BOUND (type) - step ;;step > 0 |
3374 | base >= LOWER_BOUND (type) - step ;;step < 0 |
3375 | |
3376 | This is done by proving the reverse conditions are false using loop's |
3377 | initial conditions. |
3378 | |
3379 | The is necessary to make loop niter, or iv overflow analysis easier |
3380 | for below example: |
3381 | |
3382 | int foo (int *a, signed char s, signed char l) |
3383 | { |
3384 | signed char i; |
3385 | for (i = s; i < l; i++) |
3386 | a[i] = 0; |
3387 | return 0; |
3388 | } |
3389 | |
3390 | Note variable I is firstly converted to type unsigned char, incremented, |
3391 | then converted back to type signed char. */ |
3392 | |
3393 | if (wrto_loop->num != use_loop->num) |
3394 | return true; |
3395 | |
3396 | if (!CONVERT_EXPR_P (iv->base) || TREE_CODE (iv->step) != INTEGER_CST) |
3397 | return true; |
3398 | |
3399 | type = TREE_TYPE (iv->base); |
3400 | e = TREE_OPERAND (iv->base, 0); |
3401 | if (TREE_CODE (e) != PLUS_EXPR |
3402 | || TREE_CODE (TREE_OPERAND (e, 1)) != INTEGER_CST |
3403 | || !tree_int_cst_equal (iv->step, |
3404 | fold_convert (type, TREE_OPERAND (e, 1)))) |
3405 | return true; |
3406 | e = TREE_OPERAND (e, 0); |
3407 | if (!CONVERT_EXPR_P (e)) |
3408 | return true; |
3409 | base = TREE_OPERAND (e, 0); |
3410 | if (!useless_type_conversion_p (type, TREE_TYPE (base))) |
3411 | return true; |
3412 | |
3413 | if (tree_int_cst_sign_bit (iv->step)) |
3414 | { |
3415 | code = LT_EXPR; |
3416 | extreme = wi::min_value (type); |
3417 | } |
3418 | else |
3419 | { |
3420 | code = GT_EXPR; |
3421 | extreme = wi::max_value (type); |
3422 | } |
3423 | overflow = false; |
3424 | extreme = wi::sub (extreme, wi::to_wide (iv->step), |
3425 | TYPE_SIGN (type), &overflow); |
3426 | if (overflow) |
3427 | return true; |
3428 | e = fold_build2 (code, boolean_type_node, base, |
3429 | wide_int_to_tree (type, extreme)); |
3430 | e = simplify_using_initial_conditions (use_loop, e); |
3431 | if (!integer_zerop (e)) |
3432 | return true; |
3433 | |
3434 | if (POINTER_TYPE_P (TREE_TYPE (base))) |
3435 | code = POINTER_PLUS_EXPR; |
3436 | else |
3437 | code = PLUS_EXPR; |
3438 | |
3439 | iv->base = fold_build2 (code, TREE_TYPE (base), base, iv->step); |
3440 | return true; |
3441 | } |
3442 | |
3443 | /* Like simple_iv_with_niters, but return TRUE when OP behaves as a simple |
3444 | affine iv unconditionally. */ |
3445 | |
3446 | bool |
3447 | simple_iv (struct loop *wrto_loop, struct loop *use_loop, tree op, |
3448 | affine_iv *iv, bool allow_nonconstant_step) |
3449 | { |
3450 | return simple_iv_with_niters (wrto_loop, use_loop, op, iv, |
3451 | NULL, allow_nonconstant_step); |
3452 | } |
3453 | |
3454 | /* Finalize the scalar evolution analysis. */ |
3455 | |
3456 | void |
3457 | scev_finalize (void) |
3458 | { |
3459 | if (!scalar_evolution_info) |
3460 | return; |
3461 | scalar_evolution_info->empty (); |
3462 | scalar_evolution_info = NULL; |
3463 | free_numbers_of_iterations_estimates (cfun); |
3464 | } |
3465 | |
3466 | /* Returns true if the expression EXPR is considered to be too expensive |
3467 | for scev_const_prop. */ |
3468 | |
3469 | bool |
3470 | expression_expensive_p (tree expr) |
3471 | { |
3472 | enum tree_code code; |
3473 | |
3474 | if (is_gimple_val (expr)) |
3475 | return false; |
3476 | |
3477 | code = TREE_CODE (expr); |
3478 | if (code == TRUNC_DIV_EXPR |
3479 | || code == CEIL_DIV_EXPR |
3480 | || code == FLOOR_DIV_EXPR |
3481 | || code == ROUND_DIV_EXPR |
3482 | || code == TRUNC_MOD_EXPR |
3483 | || code == CEIL_MOD_EXPR |
3484 | || code == FLOOR_MOD_EXPR |
3485 | || code == ROUND_MOD_EXPR |
3486 | || code == EXACT_DIV_EXPR) |
3487 | { |
3488 | /* Division by power of two is usually cheap, so we allow it. |
3489 | Forbid anything else. */ |
3490 | if (!integer_pow2p (TREE_OPERAND (expr, 1))) |
3491 | return true; |
3492 | } |
3493 | |
3494 | switch (TREE_CODE_CLASS (code)) |
3495 | { |
3496 | case tcc_binary: |
3497 | case tcc_comparison: |
3498 | if (expression_expensive_p (TREE_OPERAND (expr, 1))) |
3499 | return true; |
3500 | |
3501 | /* Fallthru. */ |
3502 | case tcc_unary: |
3503 | return expression_expensive_p (TREE_OPERAND (expr, 0)); |
3504 | |
3505 | default: |
3506 | return true; |
3507 | } |
3508 | } |
3509 | |
3510 | /* Do final value replacement for LOOP. */ |
3511 | |
3512 | void |
3513 | final_value_replacement_loop (struct loop *loop) |
3514 | { |
3515 | /* If we do not know exact number of iterations of the loop, we cannot |
3516 | replace the final value. */ |
3517 | edge exit = single_exit (loop); |
3518 | if (!exit) |
3519 | return; |
3520 | |
3521 | tree niter = number_of_latch_executions (loop); |
3522 | if (niter == chrec_dont_know) |
3523 | return; |
3524 | |
3525 | /* Ensure that it is possible to insert new statements somewhere. */ |
3526 | if (!single_pred_p (exit->dest)) |
3527 | split_loop_exit_edge (exit); |
3528 | |
3529 | /* Set stmt insertion pointer. All stmts are inserted before this point. */ |
3530 | gimple_stmt_iterator gsi = gsi_after_labels (exit->dest); |
3531 | |
3532 | struct loop *ex_loop |
3533 | = superloop_at_depth (loop, |
3534 | loop_depth (exit->dest->loop_father) + 1); |
3535 | |
3536 | gphi_iterator psi; |
3537 | for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); ) |
3538 | { |
3539 | gphi *phi = psi.phi (); |
3540 | tree rslt = PHI_RESULT (phi); |
3541 | tree def = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
3542 | if (virtual_operand_p (def)) |
3543 | { |
3544 | gsi_next (&psi); |
3545 | continue; |
3546 | } |
3547 | |
3548 | if (!POINTER_TYPE_P (TREE_TYPE (def)) |
3549 | && !INTEGRAL_TYPE_P (TREE_TYPE (def))) |
3550 | { |
3551 | gsi_next (&psi); |
3552 | continue; |
3553 | } |
3554 | |
3555 | bool folded_casts; |
3556 | def = analyze_scalar_evolution_in_loop (ex_loop, loop, def, |
3557 | &folded_casts); |
3558 | def = compute_overall_effect_of_inner_loop (ex_loop, def); |
3559 | if (!tree_does_not_contain_chrecs (def) |
3560 | || chrec_contains_symbols_defined_in_loop (def, ex_loop->num) |
3561 | /* Moving the computation from the loop may prolong life range |
3562 | of some ssa names, which may cause problems if they appear |
3563 | on abnormal edges. */ |
3564 | || contains_abnormal_ssa_name_p (def) |
3565 | /* Do not emit expensive expressions. The rationale is that |
3566 | when someone writes a code like |
3567 | |
3568 | while (n > 45) n -= 45; |
3569 | |
3570 | he probably knows that n is not large, and does not want it |
3571 | to be turned into n %= 45. */ |
3572 | || expression_expensive_p (def)) |
3573 | { |
3574 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3575 | { |
3576 | fprintf (dump_file, "not replacing:\n " ); |
3577 | print_gimple_stmt (dump_file, phi, 0); |
3578 | fprintf (dump_file, "\n" ); |
3579 | } |
3580 | gsi_next (&psi); |
3581 | continue; |
3582 | } |
3583 | |
3584 | /* Eliminate the PHI node and replace it by a computation outside |
3585 | the loop. */ |
3586 | if (dump_file) |
3587 | { |
3588 | fprintf (dump_file, "\nfinal value replacement:\n " ); |
3589 | print_gimple_stmt (dump_file, phi, 0); |
3590 | fprintf (dump_file, " with\n " ); |
3591 | } |
3592 | def = unshare_expr (def); |
3593 | remove_phi_node (&psi, false); |
3594 | |
3595 | /* If def's type has undefined overflow and there were folded |
3596 | casts, rewrite all stmts added for def into arithmetics |
3597 | with defined overflow behavior. */ |
3598 | if (folded_casts && ANY_INTEGRAL_TYPE_P (TREE_TYPE (def)) |
3599 | && TYPE_OVERFLOW_UNDEFINED (TREE_TYPE (def))) |
3600 | { |
3601 | gimple_seq stmts; |
3602 | gimple_stmt_iterator gsi2; |
3603 | def = force_gimple_operand (def, &stmts, true, NULL_TREE); |
3604 | gsi2 = gsi_start (stmts); |
3605 | while (!gsi_end_p (gsi2)) |
3606 | { |
3607 | gimple *stmt = gsi_stmt (gsi2); |
3608 | gimple_stmt_iterator gsi3 = gsi2; |
3609 | gsi_next (&gsi2); |
3610 | gsi_remove (&gsi3, false); |
3611 | if (is_gimple_assign (stmt) |
3612 | && arith_code_with_undefined_signed_overflow |
3613 | (gimple_assign_rhs_code (stmt))) |
3614 | gsi_insert_seq_before (&gsi, |
3615 | rewrite_to_defined_overflow (stmt), |
3616 | GSI_SAME_STMT); |
3617 | else |
3618 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); |
3619 | } |
3620 | } |
3621 | else |
3622 | def = force_gimple_operand_gsi (&gsi, def, false, NULL_TREE, |
3623 | true, GSI_SAME_STMT); |
3624 | |
3625 | gassign *ass = gimple_build_assign (rslt, def); |
3626 | gsi_insert_before (&gsi, ass, GSI_SAME_STMT); |
3627 | if (dump_file) |
3628 | { |
3629 | print_gimple_stmt (dump_file, ass, 0); |
3630 | fprintf (dump_file, "\n" ); |
3631 | } |
3632 | } |
3633 | } |
3634 | |
3635 | /* Replace ssa names for that scev can prove they are constant by the |
3636 | appropriate constants. Also perform final value replacement in loops, |
3637 | in case the replacement expressions are cheap. |
3638 | |
3639 | We only consider SSA names defined by phi nodes; rest is left to the |
3640 | ordinary constant propagation pass. */ |
3641 | |
3642 | unsigned int |
3643 | scev_const_prop (void) |
3644 | { |
3645 | basic_block bb; |
3646 | tree name, type, ev; |
3647 | gphi *phi; |
3648 | struct loop *loop; |
3649 | bitmap ssa_names_to_remove = NULL; |
3650 | unsigned i; |
3651 | gphi_iterator psi; |
3652 | |
3653 | if (number_of_loops (cfun) <= 1) |
3654 | return 0; |
3655 | |
3656 | FOR_EACH_BB_FN (bb, cfun) |
3657 | { |
3658 | loop = bb->loop_father; |
3659 | |
3660 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi)) |
3661 | { |
3662 | phi = psi.phi (); |
3663 | name = PHI_RESULT (phi); |
3664 | |
3665 | if (virtual_operand_p (name)) |
3666 | continue; |
3667 | |
3668 | type = TREE_TYPE (name); |
3669 | |
3670 | if (!POINTER_TYPE_P (type) |
3671 | && !INTEGRAL_TYPE_P (type)) |
3672 | continue; |
3673 | |
3674 | ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name), |
3675 | NULL); |
3676 | if (!is_gimple_min_invariant (ev) |
3677 | || !may_propagate_copy (name, ev)) |
3678 | continue; |
3679 | |
3680 | /* Replace the uses of the name. */ |
3681 | if (name != ev) |
3682 | { |
3683 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3684 | { |
3685 | fprintf (dump_file, "Replacing uses of: " ); |
3686 | print_generic_expr (dump_file, name); |
3687 | fprintf (dump_file, " with: " ); |
3688 | print_generic_expr (dump_file, ev); |
3689 | fprintf (dump_file, "\n" ); |
3690 | } |
3691 | replace_uses_by (name, ev); |
3692 | } |
3693 | |
3694 | if (!ssa_names_to_remove) |
3695 | ssa_names_to_remove = BITMAP_ALLOC (NULL); |
3696 | bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name)); |
3697 | } |
3698 | } |
3699 | |
3700 | /* Remove the ssa names that were replaced by constants. We do not |
3701 | remove them directly in the previous cycle, since this |
3702 | invalidates scev cache. */ |
3703 | if (ssa_names_to_remove) |
3704 | { |
3705 | bitmap_iterator bi; |
3706 | |
3707 | EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi) |
3708 | { |
3709 | gimple_stmt_iterator psi; |
3710 | name = ssa_name (i); |
3711 | phi = as_a <gphi *> (SSA_NAME_DEF_STMT (name)); |
3712 | |
3713 | gcc_assert (gimple_code (phi) == GIMPLE_PHI); |
3714 | psi = gsi_for_stmt (phi); |
3715 | remove_phi_node (&psi, true); |
3716 | } |
3717 | |
3718 | BITMAP_FREE (ssa_names_to_remove); |
3719 | scev_reset (); |
3720 | } |
3721 | |
3722 | /* Now the regular final value replacement. */ |
3723 | FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
3724 | final_value_replacement_loop (loop); |
3725 | |
3726 | return 0; |
3727 | } |
3728 | |
3729 | #include "gt-tree-scalar-evolution.h" |
3730 | |