]>
Commit | Line | Data |
---|---|---|
e9eb809d | 1 | /* Scalar evolution detector. |
ad616de1 | 2 | Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc. |
e9eb809d ZD |
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 2, 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 COPYING. If not, write to the Free | |
366ccddb KC |
19 | Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA |
20 | 02110-1301, USA. */ | |
e9eb809d | 21 | |
9baba81b SP |
22 | /* |
23 | Description: | |
24 | ||
25 | This pass analyzes the evolution of scalar variables in loop | |
26 | structures. The algorithm is based on the SSA representation, | |
27 | and on the loop hierarchy tree. This algorithm is not based on | |
28 | the notion of versions of a variable, as it was the case for the | |
29 | previous implementations of the scalar evolution algorithm, but | |
30 | it assumes that each defined name is unique. | |
31 | ||
32 | The notation used in this file is called "chains of recurrences", | |
33 | and has been proposed by Eugene Zima, Robert Van Engelen, and | |
34 | others for describing induction variables in programs. For example | |
35 | "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0 | |
36 | when entering in the loop_1 and has a step 2 in this loop, in other | |
37 | words "for (b = 0; b < N; b+=2);". Note that the coefficients of | |
38 | this chain of recurrence (or chrec [shrek]) can contain the name of | |
39 | other variables, in which case they are called parametric chrecs. | |
40 | For example, "b -> {a, +, 2}_1" means that the initial value of "b" | |
41 | is the value of "a". In most of the cases these parametric chrecs | |
42 | are fully instantiated before their use because symbolic names can | |
43 | hide some difficult cases such as self-references described later | |
44 | (see the Fibonacci example). | |
45 | ||
46 | A short sketch of the algorithm is: | |
47 | ||
48 | Given a scalar variable to be analyzed, follow the SSA edge to | |
49 | its definition: | |
50 | ||
07beea0d | 51 | - When the definition is a GIMPLE_MODIFY_STMT: if the right hand side |
9baba81b SP |
52 | (RHS) of the definition cannot be statically analyzed, the answer |
53 | of the analyzer is: "don't know". | |
54 | Otherwise, for all the variables that are not yet analyzed in the | |
55 | RHS, try to determine their evolution, and finally try to | |
56 | evaluate the operation of the RHS that gives the evolution | |
57 | function of the analyzed variable. | |
58 | ||
59 | - When the definition is a condition-phi-node: determine the | |
60 | evolution function for all the branches of the phi node, and | |
61 | finally merge these evolutions (see chrec_merge). | |
62 | ||
63 | - When the definition is a loop-phi-node: determine its initial | |
64 | condition, that is the SSA edge defined in an outer loop, and | |
65 | keep it symbolic. Then determine the SSA edges that are defined | |
66 | in the body of the loop. Follow the inner edges until ending on | |
67 | another loop-phi-node of the same analyzed loop. If the reached | |
68 | loop-phi-node is not the starting loop-phi-node, then we keep | |
69 | this definition under a symbolic form. If the reached | |
70 | loop-phi-node is the same as the starting one, then we compute a | |
71 | symbolic stride on the return path. The result is then the | |
72 | symbolic chrec {initial_condition, +, symbolic_stride}_loop. | |
73 | ||
74 | Examples: | |
75 | ||
76 | Example 1: Illustration of the basic algorithm. | |
77 | ||
78 | | a = 3 | |
79 | | loop_1 | |
80 | | b = phi (a, c) | |
81 | | c = b + 1 | |
82 | | if (c > 10) exit_loop | |
83 | | endloop | |
84 | ||
85 | Suppose that we want to know the number of iterations of the | |
86 | loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We | |
87 | ask the scalar evolution analyzer two questions: what's the | |
88 | scalar evolution (scev) of "c", and what's the scev of "10". For | |
89 | "10" the answer is "10" since it is a scalar constant. For the | |
90 | scalar variable "c", it follows the SSA edge to its definition, | |
91 | "c = b + 1", and then asks again what's the scev of "b". | |
92 | Following the SSA edge, we end on a loop-phi-node "b = phi (a, | |
93 | c)", where the initial condition is "a", and the inner loop edge | |
94 | is "c". The initial condition is kept under a symbolic form (it | |
95 | may be the case that the copy constant propagation has done its | |
96 | work and we end with the constant "3" as one of the edges of the | |
97 | loop-phi-node). The update edge is followed to the end of the | |
98 | loop, and until reaching again the starting loop-phi-node: b -> c | |
99 | -> b. At this point we have drawn a path from "b" to "b" from | |
100 | which we compute the stride in the loop: in this example it is | |
101 | "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now | |
102 | that the scev for "b" is known, it is possible to compute the | |
103 | scev for "c", that is "c -> {a + 1, +, 1}_1". In order to | |
104 | determine the number of iterations in the loop_1, we have to | |
105 | instantiate_parameters ({a + 1, +, 1}_1), that gives after some | |
106 | more analysis the scev {4, +, 1}_1, or in other words, this is | |
107 | the function "f (x) = x + 4", where x is the iteration count of | |
108 | the loop_1. Now we have to solve the inequality "x + 4 > 10", | |
109 | and take the smallest iteration number for which the loop is | |
110 | exited: x = 7. This loop runs from x = 0 to x = 7, and in total | |
111 | there are 8 iterations. In terms of loop normalization, we have | |
112 | created a variable that is implicitly defined, "x" or just "_1", | |
113 | and all the other analyzed scalars of the loop are defined in | |
114 | function of this variable: | |
115 | ||
116 | a -> 3 | |
117 | b -> {3, +, 1}_1 | |
118 | c -> {4, +, 1}_1 | |
119 | ||
120 | or in terms of a C program: | |
121 | ||
122 | | a = 3 | |
123 | | for (x = 0; x <= 7; x++) | |
124 | | { | |
125 | | b = x + 3 | |
126 | | c = x + 4 | |
127 | | } | |
128 | ||
129 | Example 2: Illustration of the algorithm on nested loops. | |
130 | ||
131 | | loop_1 | |
132 | | a = phi (1, b) | |
133 | | c = a + 2 | |
134 | | loop_2 10 times | |
135 | | b = phi (c, d) | |
136 | | d = b + 3 | |
137 | | endloop | |
138 | | endloop | |
139 | ||
140 | For analyzing the scalar evolution of "a", the algorithm follows | |
141 | the SSA edge into the loop's body: "a -> b". "b" is an inner | |
142 | loop-phi-node, and its analysis as in Example 1, gives: | |
143 | ||
144 | b -> {c, +, 3}_2 | |
145 | d -> {c + 3, +, 3}_2 | |
146 | ||
147 | Following the SSA edge for the initial condition, we end on "c = a | |
148 | + 2", and then on the starting loop-phi-node "a". From this point, | |
149 | the loop stride is computed: back on "c = a + 2" we get a "+2" in | |
150 | the loop_1, then on the loop-phi-node "b" we compute the overall | |
151 | effect of the inner loop that is "b = c + 30", and we get a "+30" | |
152 | in the loop_1. That means that the overall stride in loop_1 is | |
153 | equal to "+32", and the result is: | |
154 | ||
155 | a -> {1, +, 32}_1 | |
156 | c -> {3, +, 32}_1 | |
157 | ||
158 | Example 3: Higher degree polynomials. | |
159 | ||
160 | | loop_1 | |
161 | | a = phi (2, b) | |
162 | | c = phi (5, d) | |
163 | | b = a + 1 | |
164 | | d = c + a | |
165 | | endloop | |
166 | ||
167 | a -> {2, +, 1}_1 | |
168 | b -> {3, +, 1}_1 | |
169 | c -> {5, +, a}_1 | |
170 | d -> {5 + a, +, a}_1 | |
171 | ||
172 | instantiate_parameters ({5, +, a}_1) -> {5, +, 2, +, 1}_1 | |
173 | instantiate_parameters ({5 + a, +, a}_1) -> {7, +, 3, +, 1}_1 | |
174 | ||
175 | Example 4: Lucas, Fibonacci, or mixers in general. | |
176 | ||
177 | | loop_1 | |
178 | | a = phi (1, b) | |
179 | | c = phi (3, d) | |
180 | | b = c | |
181 | | d = c + a | |
182 | | endloop | |
183 | ||
184 | a -> (1, c)_1 | |
185 | c -> {3, +, a}_1 | |
186 | ||
187 | The syntax "(1, c)_1" stands for a PEELED_CHREC that has the | |
188 | following semantics: during the first iteration of the loop_1, the | |
189 | variable contains the value 1, and then it contains the value "c". | |
190 | Note that this syntax is close to the syntax of the loop-phi-node: | |
191 | "a -> (1, c)_1" vs. "a = phi (1, c)". | |
192 | ||
193 | The symbolic chrec representation contains all the semantics of the | |
194 | original code. What is more difficult is to use this information. | |
195 | ||
196 | Example 5: Flip-flops, or exchangers. | |
197 | ||
198 | | loop_1 | |
199 | | a = phi (1, b) | |
200 | | c = phi (3, d) | |
201 | | b = c | |
202 | | d = a | |
203 | | endloop | |
204 | ||
205 | a -> (1, c)_1 | |
206 | c -> (3, a)_1 | |
207 | ||
208 | Based on these symbolic chrecs, it is possible to refine this | |
209 | information into the more precise PERIODIC_CHRECs: | |
210 | ||
211 | a -> |1, 3|_1 | |
212 | c -> |3, 1|_1 | |
213 | ||
214 | This transformation is not yet implemented. | |
215 | ||
216 | Further readings: | |
217 | ||
218 | You can find a more detailed description of the algorithm in: | |
219 | http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf | |
220 | http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that | |
221 | this is a preliminary report and some of the details of the | |
222 | algorithm have changed. I'm working on a research report that | |
223 | updates the description of the algorithms to reflect the design | |
224 | choices used in this implementation. | |
225 | ||
226 | A set of slides show a high level overview of the algorithm and run | |
227 | an example through the scalar evolution analyzer: | |
228 | http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf | |
229 | ||
230 | The slides that I have presented at the GCC Summit'04 are available | |
231 | at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf | |
232 | */ | |
233 | ||
e9eb809d ZD |
234 | #include "config.h" |
235 | #include "system.h" | |
236 | #include "coretypes.h" | |
237 | #include "tm.h" | |
e9eb809d ZD |
238 | #include "ggc.h" |
239 | #include "tree.h" | |
9d2b0e12 | 240 | #include "real.h" |
9baba81b SP |
241 | |
242 | /* These RTL headers are needed for basic-block.h. */ | |
e9eb809d ZD |
243 | #include "rtl.h" |
244 | #include "basic-block.h" | |
245 | #include "diagnostic.h" | |
246 | #include "tree-flow.h" | |
247 | #include "tree-dump.h" | |
248 | #include "timevar.h" | |
249 | #include "cfgloop.h" | |
250 | #include "tree-chrec.h" | |
251 | #include "tree-scalar-evolution.h" | |
9baba81b SP |
252 | #include "tree-pass.h" |
253 | #include "flags.h" | |
c59dabbe | 254 | #include "params.h" |
9baba81b SP |
255 | |
256 | static tree analyze_scalar_evolution_1 (struct loop *, tree, tree); | |
257 | static tree resolve_mixers (struct loop *, tree); | |
258 | ||
259 | /* The cached information about a ssa name VAR, claiming that inside LOOP, | |
260 | the value of VAR can be expressed as CHREC. */ | |
261 | ||
262 | struct scev_info_str | |
263 | { | |
264 | tree var; | |
265 | tree chrec; | |
266 | }; | |
267 | ||
268 | /* Counters for the scev database. */ | |
269 | static unsigned nb_set_scev = 0; | |
270 | static unsigned nb_get_scev = 0; | |
271 | ||
272 | /* The following trees are unique elements. Thus the comparison of | |
273 | another element to these elements should be done on the pointer to | |
274 | these trees, and not on their value. */ | |
275 | ||
276 | /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */ | |
277 | tree chrec_not_analyzed_yet; | |
278 | ||
279 | /* Reserved to the cases where the analyzer has detected an | |
280 | undecidable property at compile time. */ | |
281 | tree chrec_dont_know; | |
282 | ||
283 | /* When the analyzer has detected that a property will never | |
284 | happen, then it qualifies it with chrec_known. */ | |
285 | tree chrec_known; | |
286 | ||
287 | static bitmap already_instantiated; | |
288 | ||
289 | static htab_t scalar_evolution_info; | |
290 | ||
291 | \f | |
292 | /* Constructs a new SCEV_INFO_STR structure. */ | |
293 | ||
294 | static inline struct scev_info_str * | |
295 | new_scev_info_str (tree var) | |
296 | { | |
297 | struct scev_info_str *res; | |
298 | ||
cceb1885 | 299 | res = XNEW (struct scev_info_str); |
9baba81b SP |
300 | res->var = var; |
301 | res->chrec = chrec_not_analyzed_yet; | |
302 | ||
303 | return res; | |
304 | } | |
305 | ||
306 | /* Computes a hash function for database element ELT. */ | |
307 | ||
308 | static hashval_t | |
309 | hash_scev_info (const void *elt) | |
310 | { | |
311 | return SSA_NAME_VERSION (((struct scev_info_str *) elt)->var); | |
312 | } | |
313 | ||
314 | /* Compares database elements E1 and E2. */ | |
315 | ||
316 | static int | |
317 | eq_scev_info (const void *e1, const void *e2) | |
318 | { | |
cceb1885 GDR |
319 | const struct scev_info_str *elt1 = (const struct scev_info_str *) e1; |
320 | const struct scev_info_str *elt2 = (const struct scev_info_str *) e2; | |
9baba81b SP |
321 | |
322 | return elt1->var == elt2->var; | |
323 | } | |
324 | ||
325 | /* Deletes database element E. */ | |
326 | ||
327 | static void | |
328 | del_scev_info (void *e) | |
329 | { | |
330 | free (e); | |
331 | } | |
332 | ||
333 | /* Get the index corresponding to VAR in the current LOOP. If | |
334 | it's the first time we ask for this VAR, then we return | |
b01d837f | 335 | chrec_not_analyzed_yet for this VAR and return its index. */ |
9baba81b SP |
336 | |
337 | static tree * | |
338 | find_var_scev_info (tree var) | |
339 | { | |
340 | struct scev_info_str *res; | |
341 | struct scev_info_str tmp; | |
342 | PTR *slot; | |
343 | ||
344 | tmp.var = var; | |
345 | slot = htab_find_slot (scalar_evolution_info, &tmp, INSERT); | |
346 | ||
347 | if (!*slot) | |
348 | *slot = new_scev_info_str (var); | |
cceb1885 | 349 | res = (struct scev_info_str *) *slot; |
9baba81b SP |
350 | |
351 | return &res->chrec; | |
352 | } | |
353 | ||
9baba81b SP |
354 | /* Return true when CHREC contains symbolic names defined in |
355 | LOOP_NB. */ | |
356 | ||
357 | bool | |
358 | chrec_contains_symbols_defined_in_loop (tree chrec, unsigned loop_nb) | |
359 | { | |
360 | if (chrec == NULL_TREE) | |
361 | return false; | |
362 | ||
363 | if (TREE_INVARIANT (chrec)) | |
364 | return false; | |
365 | ||
366 | if (TREE_CODE (chrec) == VAR_DECL | |
367 | || TREE_CODE (chrec) == PARM_DECL | |
368 | || TREE_CODE (chrec) == FUNCTION_DECL | |
369 | || TREE_CODE (chrec) == LABEL_DECL | |
370 | || TREE_CODE (chrec) == RESULT_DECL | |
371 | || TREE_CODE (chrec) == FIELD_DECL) | |
372 | return true; | |
373 | ||
374 | if (TREE_CODE (chrec) == SSA_NAME) | |
375 | { | |
376 | tree def = SSA_NAME_DEF_STMT (chrec); | |
377 | struct loop *def_loop = loop_containing_stmt (def); | |
42fd6772 | 378 | struct loop *loop = get_loop (loop_nb); |
9baba81b SP |
379 | |
380 | if (def_loop == NULL) | |
381 | return false; | |
382 | ||
383 | if (loop == def_loop || flow_loop_nested_p (loop, def_loop)) | |
384 | return true; | |
385 | ||
386 | return false; | |
387 | } | |
388 | ||
389 | switch (TREE_CODE_LENGTH (TREE_CODE (chrec))) | |
390 | { | |
391 | case 3: | |
392 | if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, 2), | |
393 | loop_nb)) | |
394 | return true; | |
395 | ||
396 | case 2: | |
397 | if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, 1), | |
398 | loop_nb)) | |
399 | return true; | |
400 | ||
401 | case 1: | |
402 | if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, 0), | |
403 | loop_nb)) | |
404 | return true; | |
405 | ||
406 | default: | |
407 | return false; | |
408 | } | |
409 | } | |
410 | ||
411 | /* Return true when PHI is a loop-phi-node. */ | |
412 | ||
413 | static bool | |
414 | loop_phi_node_p (tree phi) | |
415 | { | |
416 | /* The implementation of this function is based on the following | |
417 | property: "all the loop-phi-nodes of a loop are contained in the | |
418 | loop's header basic block". */ | |
419 | ||
420 | return loop_containing_stmt (phi)->header == bb_for_stmt (phi); | |
421 | } | |
422 | ||
423 | /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP. | |
424 | In general, in the case of multivariate evolutions we want to get | |
425 | the evolution in different loops. LOOP specifies the level for | |
426 | which to get the evolution. | |
427 | ||
428 | Example: | |
429 | ||
430 | | for (j = 0; j < 100; j++) | |
431 | | { | |
432 | | for (k = 0; k < 100; k++) | |
433 | | { | |
434 | | i = k + j; - Here the value of i is a function of j, k. | |
435 | | } | |
436 | | ... = i - Here the value of i is a function of j. | |
437 | | } | |
438 | | ... = i - Here the value of i is a scalar. | |
439 | ||
440 | Example: | |
441 | ||
442 | | i_0 = ... | |
443 | | loop_1 10 times | |
444 | | i_1 = phi (i_0, i_2) | |
445 | | i_2 = i_1 + 2 | |
446 | | endloop | |
447 | ||
448 | This loop has the same effect as: | |
449 | LOOP_1 has the same effect as: | |
450 | ||
451 | | i_1 = i_0 + 20 | |
452 | ||
453 | The overall effect of the loop, "i_0 + 20" in the previous example, | |
454 | is obtained by passing in the parameters: LOOP = 1, | |
455 | EVOLUTION_FN = {i_0, +, 2}_1. | |
456 | */ | |
457 | ||
458 | static tree | |
459 | compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn) | |
460 | { | |
461 | bool val = false; | |
462 | ||
463 | if (evolution_fn == chrec_dont_know) | |
464 | return chrec_dont_know; | |
465 | ||
466 | else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC) | |
467 | { | |
468 | if (CHREC_VARIABLE (evolution_fn) >= (unsigned) loop->num) | |
469 | { | |
42fd6772 | 470 | struct loop *inner_loop = get_chrec_loop (evolution_fn); |
a14865db | 471 | tree nb_iter = number_of_latch_executions (inner_loop); |
9baba81b SP |
472 | |
473 | if (nb_iter == chrec_dont_know) | |
474 | return chrec_dont_know; | |
475 | else | |
476 | { | |
477 | tree res; | |
478 | ||
9baba81b SP |
479 | /* evolution_fn is the evolution function in LOOP. Get |
480 | its value in the nb_iter-th iteration. */ | |
481 | res = chrec_apply (inner_loop->num, evolution_fn, nb_iter); | |
482 | ||
8c27b7d4 | 483 | /* Continue the computation until ending on a parent of LOOP. */ |
9baba81b SP |
484 | return compute_overall_effect_of_inner_loop (loop, res); |
485 | } | |
486 | } | |
487 | else | |
488 | return evolution_fn; | |
489 | } | |
490 | ||
491 | /* If the evolution function is an invariant, there is nothing to do. */ | |
492 | else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val) | |
493 | return evolution_fn; | |
494 | ||
495 | else | |
496 | return chrec_dont_know; | |
497 | } | |
498 | ||
499 | /* Determine whether the CHREC is always positive/negative. If the expression | |
500 | cannot be statically analyzed, return false, otherwise set the answer into | |
501 | VALUE. */ | |
502 | ||
503 | bool | |
504 | chrec_is_positive (tree chrec, bool *value) | |
505 | { | |
16a2acea | 506 | bool value0, value1, value2; |
a14865db | 507 | tree end_value, nb_iter; |
9baba81b SP |
508 | |
509 | switch (TREE_CODE (chrec)) | |
510 | { | |
511 | case POLYNOMIAL_CHREC: | |
512 | if (!chrec_is_positive (CHREC_LEFT (chrec), &value0) | |
513 | || !chrec_is_positive (CHREC_RIGHT (chrec), &value1)) | |
514 | return false; | |
515 | ||
516 | /* FIXME -- overflows. */ | |
517 | if (value0 == value1) | |
518 | { | |
519 | *value = value0; | |
520 | return true; | |
521 | } | |
522 | ||
523 | /* Otherwise the chrec is under the form: "{-197, +, 2}_1", | |
524 | and the proof consists in showing that the sign never | |
525 | changes during the execution of the loop, from 0 to | |
526 | loop->nb_iterations. */ | |
527 | if (!evolution_function_is_affine_p (chrec)) | |
528 | return false; | |
529 | ||
a14865db | 530 | nb_iter = number_of_latch_executions (get_chrec_loop (chrec)); |
9baba81b SP |
531 | if (chrec_contains_undetermined (nb_iter)) |
532 | return false; | |
533 | ||
9baba81b SP |
534 | #if 0 |
535 | /* TODO -- If the test is after the exit, we may decrease the number of | |
536 | iterations by one. */ | |
537 | if (after_exit) | |
16a2acea | 538 | nb_iter = chrec_fold_minus (type, nb_iter, build_int_cst (type, 1)); |
9baba81b SP |
539 | #endif |
540 | ||
541 | end_value = chrec_apply (CHREC_VARIABLE (chrec), chrec, nb_iter); | |
542 | ||
543 | if (!chrec_is_positive (end_value, &value2)) | |
544 | return false; | |
545 | ||
546 | *value = value0; | |
547 | return value0 == value1; | |
548 | ||
549 | case INTEGER_CST: | |
550 | *value = (tree_int_cst_sgn (chrec) == 1); | |
551 | return true; | |
552 | ||
553 | default: | |
554 | return false; | |
555 | } | |
556 | } | |
557 | ||
558 | /* Associate CHREC to SCALAR. */ | |
559 | ||
560 | static void | |
561 | set_scalar_evolution (tree scalar, tree chrec) | |
562 | { | |
563 | tree *scalar_info; | |
564 | ||
565 | if (TREE_CODE (scalar) != SSA_NAME) | |
566 | return; | |
567 | ||
568 | scalar_info = find_var_scev_info (scalar); | |
569 | ||
570 | if (dump_file) | |
571 | { | |
572 | if (dump_flags & TDF_DETAILS) | |
573 | { | |
574 | fprintf (dump_file, "(set_scalar_evolution \n"); | |
575 | fprintf (dump_file, " (scalar = "); | |
576 | print_generic_expr (dump_file, scalar, 0); | |
577 | fprintf (dump_file, ")\n (scalar_evolution = "); | |
578 | print_generic_expr (dump_file, chrec, 0); | |
579 | fprintf (dump_file, "))\n"); | |
580 | } | |
581 | if (dump_flags & TDF_STATS) | |
582 | nb_set_scev++; | |
583 | } | |
584 | ||
585 | *scalar_info = chrec; | |
586 | } | |
587 | ||
588 | /* Retrieve the chrec associated to SCALAR in the LOOP. */ | |
589 | ||
590 | static tree | |
591 | get_scalar_evolution (tree scalar) | |
592 | { | |
593 | tree res; | |
594 | ||
595 | if (dump_file) | |
596 | { | |
597 | if (dump_flags & TDF_DETAILS) | |
598 | { | |
599 | fprintf (dump_file, "(get_scalar_evolution \n"); | |
600 | fprintf (dump_file, " (scalar = "); | |
601 | print_generic_expr (dump_file, scalar, 0); | |
602 | fprintf (dump_file, ")\n"); | |
603 | } | |
604 | if (dump_flags & TDF_STATS) | |
605 | nb_get_scev++; | |
606 | } | |
607 | ||
608 | switch (TREE_CODE (scalar)) | |
609 | { | |
610 | case SSA_NAME: | |
611 | res = *find_var_scev_info (scalar); | |
612 | break; | |
613 | ||
614 | case REAL_CST: | |
615 | case INTEGER_CST: | |
616 | res = scalar; | |
617 | break; | |
618 | ||
619 | default: | |
620 | res = chrec_not_analyzed_yet; | |
621 | break; | |
622 | } | |
623 | ||
624 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
625 | { | |
626 | fprintf (dump_file, " (scalar_evolution = "); | |
627 | print_generic_expr (dump_file, res, 0); | |
628 | fprintf (dump_file, "))\n"); | |
629 | } | |
630 | ||
631 | return res; | |
632 | } | |
633 | ||
634 | /* Helper function for add_to_evolution. Returns the evolution | |
635 | function for an assignment of the form "a = b + c", where "a" and | |
636 | "b" are on the strongly connected component. CHREC_BEFORE is the | |
637 | information that we already have collected up to this point. | |
638 | TO_ADD is the evolution of "c". | |
639 | ||
640 | When CHREC_BEFORE has an evolution part in LOOP_NB, add to this | |
641 | evolution the expression TO_ADD, otherwise construct an evolution | |
642 | part for this loop. */ | |
643 | ||
644 | static tree | |
e2157b49 SP |
645 | add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add, |
646 | tree at_stmt) | |
9baba81b | 647 | { |
e2157b49 SP |
648 | tree type, left, right; |
649 | ||
9baba81b SP |
650 | switch (TREE_CODE (chrec_before)) |
651 | { | |
652 | case POLYNOMIAL_CHREC: | |
653 | if (CHREC_VARIABLE (chrec_before) <= loop_nb) | |
654 | { | |
655 | unsigned var; | |
e2157b49 SP |
656 | |
657 | type = chrec_type (chrec_before); | |
9baba81b SP |
658 | |
659 | /* When there is no evolution part in this loop, build it. */ | |
660 | if (CHREC_VARIABLE (chrec_before) < loop_nb) | |
661 | { | |
662 | var = loop_nb; | |
663 | left = chrec_before; | |
7e0923cd SP |
664 | right = SCALAR_FLOAT_TYPE_P (type) |
665 | ? build_real (type, dconst0) | |
666 | : build_int_cst (type, 0); | |
9baba81b SP |
667 | } |
668 | else | |
669 | { | |
670 | var = CHREC_VARIABLE (chrec_before); | |
671 | left = CHREC_LEFT (chrec_before); | |
672 | right = CHREC_RIGHT (chrec_before); | |
673 | } | |
674 | ||
e2157b49 SP |
675 | to_add = chrec_convert (type, to_add, at_stmt); |
676 | right = chrec_convert (type, right, at_stmt); | |
677 | right = chrec_fold_plus (type, right, to_add); | |
678 | return build_polynomial_chrec (var, left, right); | |
9baba81b SP |
679 | } |
680 | else | |
e2157b49 SP |
681 | { |
682 | /* Search the evolution in LOOP_NB. */ | |
683 | left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before), | |
684 | to_add, at_stmt); | |
685 | right = CHREC_RIGHT (chrec_before); | |
686 | right = chrec_convert (chrec_type (left), right, at_stmt); | |
687 | return build_polynomial_chrec (CHREC_VARIABLE (chrec_before), | |
688 | left, right); | |
689 | } | |
9baba81b SP |
690 | |
691 | default: | |
692 | /* These nodes do not depend on a loop. */ | |
693 | if (chrec_before == chrec_dont_know) | |
694 | return chrec_dont_know; | |
e2157b49 SP |
695 | |
696 | left = chrec_before; | |
697 | right = chrec_convert (chrec_type (left), to_add, at_stmt); | |
698 | return build_polynomial_chrec (loop_nb, left, right); | |
9baba81b SP |
699 | } |
700 | } | |
701 | ||
702 | /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension | |
703 | of LOOP_NB. | |
704 | ||
705 | Description (provided for completeness, for those who read code in | |
706 | a plane, and for my poor 62 bytes brain that would have forgotten | |
707 | all this in the next two or three months): | |
708 | ||
709 | The algorithm of translation of programs from the SSA representation | |
710 | into the chrecs syntax is based on a pattern matching. After having | |
711 | reconstructed the overall tree expression for a loop, there are only | |
712 | two cases that can arise: | |
713 | ||
714 | 1. a = loop-phi (init, a + expr) | |
715 | 2. a = loop-phi (init, expr) | |
716 | ||
717 | where EXPR is either a scalar constant with respect to the analyzed | |
718 | loop (this is a degree 0 polynomial), or an expression containing | |
719 | other loop-phi definitions (these are higher degree polynomials). | |
720 | ||
721 | Examples: | |
722 | ||
723 | 1. | |
724 | | init = ... | |
725 | | loop_1 | |
726 | | a = phi (init, a + 5) | |
727 | | endloop | |
728 | ||
729 | 2. | |
730 | | inita = ... | |
731 | | initb = ... | |
732 | | loop_1 | |
733 | | a = phi (inita, 2 * b + 3) | |
734 | | b = phi (initb, b + 1) | |
735 | | endloop | |
736 | ||
737 | For the first case, the semantics of the SSA representation is: | |
738 | ||
739 | | a (x) = init + \sum_{j = 0}^{x - 1} expr (j) | |
740 | ||
741 | that is, there is a loop index "x" that determines the scalar value | |
742 | of the variable during the loop execution. During the first | |
743 | iteration, the value is that of the initial condition INIT, while | |
744 | during the subsequent iterations, it is the sum of the initial | |
745 | condition with the sum of all the values of EXPR from the initial | |
746 | iteration to the before last considered iteration. | |
747 | ||
748 | For the second case, the semantics of the SSA program is: | |
749 | ||
750 | | a (x) = init, if x = 0; | |
751 | | expr (x - 1), otherwise. | |
752 | ||
753 | The second case corresponds to the PEELED_CHREC, whose syntax is | |
754 | close to the syntax of a loop-phi-node: | |
755 | ||
756 | | phi (init, expr) vs. (init, expr)_x | |
757 | ||
758 | The proof of the translation algorithm for the first case is a | |
759 | proof by structural induction based on the degree of EXPR. | |
760 | ||
761 | Degree 0: | |
762 | When EXPR is a constant with respect to the analyzed loop, or in | |
763 | other words when EXPR is a polynomial of degree 0, the evolution of | |
764 | the variable A in the loop is an affine function with an initial | |
765 | condition INIT, and a step EXPR. In order to show this, we start | |
766 | from the semantics of the SSA representation: | |
767 | ||
768 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j) | |
769 | ||
770 | and since "expr (j)" is a constant with respect to "j", | |
771 | ||
772 | f (x) = init + x * expr | |
773 | ||
774 | Finally, based on the semantics of the pure sum chrecs, by | |
775 | identification we get the corresponding chrecs syntax: | |
776 | ||
777 | f (x) = init * \binom{x}{0} + expr * \binom{x}{1} | |
778 | f (x) -> {init, +, expr}_x | |
779 | ||
780 | Higher degree: | |
781 | Suppose that EXPR is a polynomial of degree N with respect to the | |
782 | analyzed loop_x for which we have already determined that it is | |
783 | written under the chrecs syntax: | |
784 | ||
785 | | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x) | |
786 | ||
787 | We start from the semantics of the SSA program: | |
788 | ||
789 | | f (x) = init + \sum_{j = 0}^{x - 1} expr (j) | |
790 | | | |
791 | | f (x) = init + \sum_{j = 0}^{x - 1} | |
792 | | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1}) | |
793 | | | |
794 | | f (x) = init + \sum_{j = 0}^{x - 1} | |
795 | | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k}) | |
796 | | | |
797 | | f (x) = init + \sum_{k = 0}^{n - 1} | |
798 | | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k}) | |
799 | | | |
800 | | f (x) = init + \sum_{k = 0}^{n - 1} | |
801 | | (b_k * \binom{x}{k + 1}) | |
802 | | | |
803 | | f (x) = init + b_0 * \binom{x}{1} + ... | |
804 | | + b_{n-1} * \binom{x}{n} | |
805 | | | |
806 | | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ... | |
807 | | + b_{n-1} * \binom{x}{n} | |
808 | | | |
809 | ||
810 | And finally from the definition of the chrecs syntax, we identify: | |
811 | | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x | |
812 | ||
813 | This shows the mechanism that stands behind the add_to_evolution | |
814 | function. An important point is that the use of symbolic | |
815 | parameters avoids the need of an analysis schedule. | |
816 | ||
817 | Example: | |
818 | ||
819 | | inita = ... | |
820 | | initb = ... | |
821 | | loop_1 | |
822 | | a = phi (inita, a + 2 + b) | |
823 | | b = phi (initb, b + 1) | |
824 | | endloop | |
825 | ||
826 | When analyzing "a", the algorithm keeps "b" symbolically: | |
827 | ||
828 | | a -> {inita, +, 2 + b}_1 | |
829 | ||
830 | Then, after instantiation, the analyzer ends on the evolution: | |
831 | ||
832 | | a -> {inita, +, 2 + initb, +, 1}_1 | |
833 | ||
834 | */ | |
835 | ||
836 | static tree | |
e2157b49 SP |
837 | add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code, |
838 | tree to_add, tree at_stmt) | |
9baba81b SP |
839 | { |
840 | tree type = chrec_type (to_add); | |
841 | tree res = NULL_TREE; | |
842 | ||
843 | if (to_add == NULL_TREE) | |
844 | return chrec_before; | |
845 | ||
846 | /* TO_ADD is either a scalar, or a parameter. TO_ADD is not | |
847 | instantiated at this point. */ | |
848 | if (TREE_CODE (to_add) == POLYNOMIAL_CHREC) | |
849 | /* This should not happen. */ | |
850 | return chrec_dont_know; | |
851 | ||
852 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
853 | { | |
854 | fprintf (dump_file, "(add_to_evolution \n"); | |
855 | fprintf (dump_file, " (loop_nb = %d)\n", loop_nb); | |
856 | fprintf (dump_file, " (chrec_before = "); | |
857 | print_generic_expr (dump_file, chrec_before, 0); | |
858 | fprintf (dump_file, ")\n (to_add = "); | |
859 | print_generic_expr (dump_file, to_add, 0); | |
860 | fprintf (dump_file, ")\n"); | |
861 | } | |
862 | ||
863 | if (code == MINUS_EXPR) | |
9d2b0e12 VR |
864 | to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type) |
865 | ? build_real (type, dconstm1) | |
866 | : build_int_cst_type (type, -1)); | |
9baba81b | 867 | |
e2157b49 | 868 | res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt); |
9baba81b SP |
869 | |
870 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
871 | { | |
872 | fprintf (dump_file, " (res = "); | |
873 | print_generic_expr (dump_file, res, 0); | |
874 | fprintf (dump_file, "))\n"); | |
875 | } | |
876 | ||
877 | return res; | |
878 | } | |
879 | ||
880 | /* Helper function. */ | |
881 | ||
882 | static inline tree | |
883 | set_nb_iterations_in_loop (struct loop *loop, | |
884 | tree res) | |
885 | { | |
9baba81b SP |
886 | if (dump_file && (dump_flags & TDF_DETAILS)) |
887 | { | |
888 | fprintf (dump_file, " (set_nb_iterations_in_loop = "); | |
889 | print_generic_expr (dump_file, res, 0); | |
890 | fprintf (dump_file, "))\n"); | |
891 | } | |
892 | ||
893 | loop->nb_iterations = res; | |
894 | return res; | |
895 | } | |
896 | ||
897 | \f | |
898 | ||
899 | /* This section selects the loops that will be good candidates for the | |
900 | scalar evolution analysis. For the moment, greedily select all the | |
901 | loop nests we could analyze. */ | |
902 | ||
903 | /* Return true when it is possible to analyze the condition expression | |
904 | EXPR. */ | |
905 | ||
906 | static bool | |
907 | analyzable_condition (tree expr) | |
908 | { | |
909 | tree condition; | |
910 | ||
911 | if (TREE_CODE (expr) != COND_EXPR) | |
912 | return false; | |
913 | ||
914 | condition = TREE_OPERAND (expr, 0); | |
915 | ||
916 | switch (TREE_CODE (condition)) | |
917 | { | |
918 | case SSA_NAME: | |
9baba81b SP |
919 | return true; |
920 | ||
921 | case LT_EXPR: | |
922 | case LE_EXPR: | |
923 | case GT_EXPR: | |
924 | case GE_EXPR: | |
925 | case EQ_EXPR: | |
926 | case NE_EXPR: | |
85022b3f | 927 | return true; |
9baba81b SP |
928 | |
929 | default: | |
930 | return false; | |
931 | } | |
932 | ||
933 | return false; | |
934 | } | |
935 | ||
936 | /* For a loop with a single exit edge, return the COND_EXPR that | |
937 | guards the exit edge. If the expression is too difficult to | |
938 | analyze, then give up. */ | |
939 | ||
940 | tree | |
941 | get_loop_exit_condition (struct loop *loop) | |
942 | { | |
943 | tree res = NULL_TREE; | |
ac8f6c69 | 944 | edge exit_edge = single_exit (loop); |
9baba81b SP |
945 | |
946 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
947 | fprintf (dump_file, "(get_loop_exit_condition \n "); | |
948 | ||
82b85a85 | 949 | if (exit_edge) |
9baba81b | 950 | { |
9baba81b SP |
951 | tree expr; |
952 | ||
9baba81b | 953 | expr = last_stmt (exit_edge->src); |
9baba81b SP |
954 | if (analyzable_condition (expr)) |
955 | res = expr; | |
956 | } | |
957 | ||
958 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
959 | { | |
960 | print_generic_expr (dump_file, res, 0); | |
961 | fprintf (dump_file, ")\n"); | |
962 | } | |
963 | ||
964 | return res; | |
965 | } | |
966 | ||
967 | /* Recursively determine and enqueue the exit conditions for a loop. */ | |
968 | ||
969 | static void | |
970 | get_exit_conditions_rec (struct loop *loop, | |
5310bac6 | 971 | VEC(tree,heap) **exit_conditions) |
9baba81b SP |
972 | { |
973 | if (!loop) | |
974 | return; | |
975 | ||
976 | /* Recurse on the inner loops, then on the next (sibling) loops. */ | |
977 | get_exit_conditions_rec (loop->inner, exit_conditions); | |
978 | get_exit_conditions_rec (loop->next, exit_conditions); | |
979 | ||
ac8f6c69 | 980 | if (single_exit (loop)) |
9baba81b SP |
981 | { |
982 | tree loop_condition = get_loop_exit_condition (loop); | |
983 | ||
984 | if (loop_condition) | |
5310bac6 | 985 | VEC_safe_push (tree, heap, *exit_conditions, loop_condition); |
9baba81b SP |
986 | } |
987 | } | |
988 | ||
989 | /* Select the candidate loop nests for the analysis. This function | |
471854f8 | 990 | initializes the EXIT_CONDITIONS array. */ |
9baba81b SP |
991 | |
992 | static void | |
d73be268 | 993 | select_loops_exit_conditions (VEC(tree,heap) **exit_conditions) |
9baba81b | 994 | { |
d73be268 | 995 | struct loop *function_body = current_loops->tree_root; |
9baba81b SP |
996 | |
997 | get_exit_conditions_rec (function_body->inner, exit_conditions); | |
998 | } | |
999 | ||
1000 | \f | |
1001 | /* Depth first search algorithm. */ | |
1002 | ||
c59dabbe SP |
1003 | typedef enum t_bool { |
1004 | t_false, | |
1005 | t_true, | |
1006 | t_dont_know | |
1007 | } t_bool; | |
1008 | ||
1009 | ||
1010 | static t_bool follow_ssa_edge (struct loop *loop, tree, tree, tree *, int); | |
9baba81b SP |
1011 | |
1012 | /* Follow the ssa edge into the right hand side RHS of an assignment. | |
1013 | Return true if the strongly connected component has been found. */ | |
1014 | ||
c59dabbe SP |
1015 | static t_bool |
1016 | follow_ssa_edge_in_rhs (struct loop *loop, tree at_stmt, tree rhs, | |
1017 | tree halting_phi, tree *evolution_of_loop, int limit) | |
9baba81b | 1018 | { |
c59dabbe | 1019 | t_bool res = t_false; |
9baba81b SP |
1020 | tree rhs0, rhs1; |
1021 | tree type_rhs = TREE_TYPE (rhs); | |
b2a93c0a | 1022 | tree evol; |
9baba81b SP |
1023 | |
1024 | /* The RHS is one of the following cases: | |
1025 | - an SSA_NAME, | |
1026 | - an INTEGER_CST, | |
1027 | - a PLUS_EXPR, | |
1028 | - a MINUS_EXPR, | |
0bca51f0 DN |
1029 | - an ASSERT_EXPR, |
1030 | - other cases are not yet handled. */ | |
9baba81b SP |
1031 | switch (TREE_CODE (rhs)) |
1032 | { | |
1033 | case NOP_EXPR: | |
1034 | /* This assignment is under the form "a_1 = (cast) rhs. */ | |
1e8552eb | 1035 | res = follow_ssa_edge_in_rhs (loop, at_stmt, TREE_OPERAND (rhs, 0), |
c59dabbe | 1036 | halting_phi, evolution_of_loop, limit); |
1e8552eb SP |
1037 | *evolution_of_loop = chrec_convert (TREE_TYPE (rhs), |
1038 | *evolution_of_loop, at_stmt); | |
9baba81b SP |
1039 | break; |
1040 | ||
1041 | case INTEGER_CST: | |
1042 | /* This assignment is under the form "a_1 = 7". */ | |
c59dabbe | 1043 | res = t_false; |
9baba81b SP |
1044 | break; |
1045 | ||
1046 | case SSA_NAME: | |
1047 | /* This assignment is under the form: "a_1 = b_2". */ | |
1048 | res = follow_ssa_edge | |
c59dabbe | 1049 | (loop, SSA_NAME_DEF_STMT (rhs), halting_phi, evolution_of_loop, limit); |
9baba81b SP |
1050 | break; |
1051 | ||
1052 | case PLUS_EXPR: | |
1053 | /* This case is under the form "rhs0 + rhs1". */ | |
1054 | rhs0 = TREE_OPERAND (rhs, 0); | |
1055 | rhs1 = TREE_OPERAND (rhs, 1); | |
1056 | STRIP_TYPE_NOPS (rhs0); | |
1057 | STRIP_TYPE_NOPS (rhs1); | |
1058 | ||
1059 | if (TREE_CODE (rhs0) == SSA_NAME) | |
1060 | { | |
1061 | if (TREE_CODE (rhs1) == SSA_NAME) | |
1062 | { | |
1063 | /* Match an assignment under the form: | |
1064 | "a = b + c". */ | |
b2a93c0a | 1065 | evol = *evolution_of_loop; |
9baba81b SP |
1066 | res = follow_ssa_edge |
1067 | (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, | |
b2a93c0a | 1068 | &evol, limit); |
9baba81b | 1069 | |
c59dabbe | 1070 | if (res == t_true) |
9baba81b SP |
1071 | *evolution_of_loop = add_to_evolution |
1072 | (loop->num, | |
b2a93c0a | 1073 | chrec_convert (type_rhs, evol, at_stmt), |
e2157b49 | 1074 | PLUS_EXPR, rhs1, at_stmt); |
9baba81b | 1075 | |
c59dabbe | 1076 | else if (res == t_false) |
9baba81b SP |
1077 | { |
1078 | res = follow_ssa_edge | |
1079 | (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi, | |
c59dabbe | 1080 | evolution_of_loop, limit); |
9baba81b | 1081 | |
c59dabbe | 1082 | if (res == t_true) |
9baba81b SP |
1083 | *evolution_of_loop = add_to_evolution |
1084 | (loop->num, | |
1e8552eb | 1085 | chrec_convert (type_rhs, *evolution_of_loop, at_stmt), |
e2157b49 | 1086 | PLUS_EXPR, rhs0, at_stmt); |
c59dabbe SP |
1087 | |
1088 | else if (res == t_dont_know) | |
1089 | *evolution_of_loop = chrec_dont_know; | |
9baba81b | 1090 | } |
c59dabbe SP |
1091 | |
1092 | else if (res == t_dont_know) | |
1093 | *evolution_of_loop = chrec_dont_know; | |
9baba81b SP |
1094 | } |
1095 | ||
1096 | else | |
1097 | { | |
1098 | /* Match an assignment under the form: | |
1099 | "a = b + ...". */ | |
1100 | res = follow_ssa_edge | |
1101 | (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, | |
c59dabbe SP |
1102 | evolution_of_loop, limit); |
1103 | if (res == t_true) | |
9baba81b | 1104 | *evolution_of_loop = add_to_evolution |
1e8552eb SP |
1105 | (loop->num, chrec_convert (type_rhs, *evolution_of_loop, |
1106 | at_stmt), | |
e2157b49 | 1107 | PLUS_EXPR, rhs1, at_stmt); |
c59dabbe SP |
1108 | |
1109 | else if (res == t_dont_know) | |
1110 | *evolution_of_loop = chrec_dont_know; | |
9baba81b SP |
1111 | } |
1112 | } | |
1113 | ||
1114 | else if (TREE_CODE (rhs1) == SSA_NAME) | |
1115 | { | |
1116 | /* Match an assignment under the form: | |
1117 | "a = ... + c". */ | |
1118 | res = follow_ssa_edge | |
1119 | (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi, | |
c59dabbe SP |
1120 | evolution_of_loop, limit); |
1121 | if (res == t_true) | |
9baba81b | 1122 | *evolution_of_loop = add_to_evolution |
1e8552eb SP |
1123 | (loop->num, chrec_convert (type_rhs, *evolution_of_loop, |
1124 | at_stmt), | |
e2157b49 | 1125 | PLUS_EXPR, rhs0, at_stmt); |
c59dabbe SP |
1126 | |
1127 | else if (res == t_dont_know) | |
1128 | *evolution_of_loop = chrec_dont_know; | |
9baba81b SP |
1129 | } |
1130 | ||
1131 | else | |
1132 | /* Otherwise, match an assignment under the form: | |
1133 | "a = ... + ...". */ | |
1134 | /* And there is nothing to do. */ | |
c59dabbe | 1135 | res = t_false; |
9baba81b SP |
1136 | |
1137 | break; | |
1138 | ||
1139 | case MINUS_EXPR: | |
1140 | /* This case is under the form "opnd0 = rhs0 - rhs1". */ | |
1141 | rhs0 = TREE_OPERAND (rhs, 0); | |
1142 | rhs1 = TREE_OPERAND (rhs, 1); | |
1143 | STRIP_TYPE_NOPS (rhs0); | |
1144 | STRIP_TYPE_NOPS (rhs1); | |
1145 | ||
1146 | if (TREE_CODE (rhs0) == SSA_NAME) | |
9baba81b SP |
1147 | { |
1148 | /* Match an assignment under the form: | |
f8e9d512 ZD |
1149 | "a = b - ...". */ |
1150 | res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, | |
c59dabbe SP |
1151 | evolution_of_loop, limit); |
1152 | if (res == t_true) | |
9baba81b | 1153 | *evolution_of_loop = add_to_evolution |
c59dabbe | 1154 | (loop->num, chrec_convert (type_rhs, *evolution_of_loop, at_stmt), |
e2157b49 | 1155 | MINUS_EXPR, rhs1, at_stmt); |
c59dabbe SP |
1156 | |
1157 | else if (res == t_dont_know) | |
1158 | *evolution_of_loop = chrec_dont_know; | |
9baba81b | 1159 | } |
9baba81b SP |
1160 | else |
1161 | /* Otherwise, match an assignment under the form: | |
1162 | "a = ... - ...". */ | |
1163 | /* And there is nothing to do. */ | |
c59dabbe | 1164 | res = t_false; |
9baba81b SP |
1165 | |
1166 | break; | |
1167 | ||
0bca51f0 DN |
1168 | case ASSERT_EXPR: |
1169 | { | |
1170 | /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>" | |
1171 | It must be handled as a copy assignment of the form a_1 = a_2. */ | |
1172 | tree op0 = ASSERT_EXPR_VAR (rhs); | |
1173 | if (TREE_CODE (op0) == SSA_NAME) | |
1174 | res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (op0), | |
c59dabbe | 1175 | halting_phi, evolution_of_loop, limit); |
0bca51f0 | 1176 | else |
c59dabbe | 1177 | res = t_false; |
0bca51f0 DN |
1178 | break; |
1179 | } | |
1180 | ||
1181 | ||
9baba81b | 1182 | default: |
c59dabbe | 1183 | res = t_false; |
9baba81b SP |
1184 | break; |
1185 | } | |
1186 | ||
1187 | return res; | |
1188 | } | |
1189 | ||
1190 | /* Checks whether the I-th argument of a PHI comes from a backedge. */ | |
1191 | ||
1192 | static bool | |
1193 | backedge_phi_arg_p (tree phi, int i) | |
1194 | { | |
1195 | edge e = PHI_ARG_EDGE (phi, i); | |
1196 | ||
1197 | /* We would in fact like to test EDGE_DFS_BACK here, but we do not care | |
1198 | about updating it anywhere, and this should work as well most of the | |
1199 | time. */ | |
1200 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
1201 | return true; | |
1202 | ||
1203 | return false; | |
1204 | } | |
1205 | ||
1206 | /* Helper function for one branch of the condition-phi-node. Return | |
1207 | true if the strongly connected component has been found following | |
1208 | this path. */ | |
1209 | ||
c59dabbe | 1210 | static inline t_bool |
9baba81b SP |
1211 | follow_ssa_edge_in_condition_phi_branch (int i, |
1212 | struct loop *loop, | |
1213 | tree condition_phi, | |
1214 | tree halting_phi, | |
1215 | tree *evolution_of_branch, | |
c59dabbe | 1216 | tree init_cond, int limit) |
9baba81b SP |
1217 | { |
1218 | tree branch = PHI_ARG_DEF (condition_phi, i); | |
1219 | *evolution_of_branch = chrec_dont_know; | |
1220 | ||
1221 | /* Do not follow back edges (they must belong to an irreducible loop, which | |
1222 | we really do not want to worry about). */ | |
1223 | if (backedge_phi_arg_p (condition_phi, i)) | |
c59dabbe | 1224 | return t_false; |
9baba81b SP |
1225 | |
1226 | if (TREE_CODE (branch) == SSA_NAME) | |
1227 | { | |
1228 | *evolution_of_branch = init_cond; | |
1229 | return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi, | |
c59dabbe | 1230 | evolution_of_branch, limit); |
9baba81b SP |
1231 | } |
1232 | ||
1233 | /* This case occurs when one of the condition branches sets | |
89dbed81 | 1234 | the variable to a constant: i.e. a phi-node like |
9baba81b SP |
1235 | "a_2 = PHI <a_7(5), 2(6)>;". |
1236 | ||
1237 | FIXME: This case have to be refined correctly: | |
1238 | in some cases it is possible to say something better than | |
1239 | chrec_dont_know, for example using a wrap-around notation. */ | |
c59dabbe | 1240 | return t_false; |
9baba81b SP |
1241 | } |
1242 | ||
1243 | /* This function merges the branches of a condition-phi-node in a | |
1244 | loop. */ | |
1245 | ||
c59dabbe | 1246 | static t_bool |
9baba81b SP |
1247 | follow_ssa_edge_in_condition_phi (struct loop *loop, |
1248 | tree condition_phi, | |
1249 | tree halting_phi, | |
c59dabbe | 1250 | tree *evolution_of_loop, int limit) |
9baba81b SP |
1251 | { |
1252 | int i; | |
1253 | tree init = *evolution_of_loop; | |
1254 | tree evolution_of_branch; | |
c59dabbe SP |
1255 | t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi, |
1256 | halting_phi, | |
1257 | &evolution_of_branch, | |
1258 | init, limit); | |
1259 | if (res == t_false || res == t_dont_know) | |
1260 | return res; | |
9baba81b | 1261 | |
9baba81b SP |
1262 | *evolution_of_loop = evolution_of_branch; |
1263 | ||
1264 | for (i = 1; i < PHI_NUM_ARGS (condition_phi); i++) | |
1265 | { | |
e0afb98a SP |
1266 | /* Quickly give up when the evolution of one of the branches is |
1267 | not known. */ | |
1268 | if (*evolution_of_loop == chrec_dont_know) | |
c59dabbe | 1269 | return t_true; |
e0afb98a | 1270 | |
c59dabbe SP |
1271 | res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi, |
1272 | halting_phi, | |
1273 | &evolution_of_branch, | |
1274 | init, limit); | |
1275 | if (res == t_false || res == t_dont_know) | |
1276 | return res; | |
9baba81b SP |
1277 | |
1278 | *evolution_of_loop = chrec_merge (*evolution_of_loop, | |
1279 | evolution_of_branch); | |
1280 | } | |
1281 | ||
c59dabbe | 1282 | return t_true; |
9baba81b SP |
1283 | } |
1284 | ||
1285 | /* Follow an SSA edge in an inner loop. It computes the overall | |
1286 | effect of the loop, and following the symbolic initial conditions, | |
1287 | it follows the edges in the parent loop. The inner loop is | |
1288 | considered as a single statement. */ | |
1289 | ||
c59dabbe | 1290 | static t_bool |
9baba81b SP |
1291 | follow_ssa_edge_inner_loop_phi (struct loop *outer_loop, |
1292 | tree loop_phi_node, | |
1293 | tree halting_phi, | |
c59dabbe | 1294 | tree *evolution_of_loop, int limit) |
9baba81b SP |
1295 | { |
1296 | struct loop *loop = loop_containing_stmt (loop_phi_node); | |
1297 | tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node)); | |
1298 | ||
1299 | /* Sometimes, the inner loop is too difficult to analyze, and the | |
1300 | result of the analysis is a symbolic parameter. */ | |
1301 | if (ev == PHI_RESULT (loop_phi_node)) | |
1302 | { | |
c59dabbe | 1303 | t_bool res = t_false; |
9baba81b SP |
1304 | int i; |
1305 | ||
1306 | for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++) | |
1307 | { | |
1308 | tree arg = PHI_ARG_DEF (loop_phi_node, i); | |
1309 | basic_block bb; | |
1310 | ||
1311 | /* Follow the edges that exit the inner loop. */ | |
1312 | bb = PHI_ARG_EDGE (loop_phi_node, i)->src; | |
1313 | if (!flow_bb_inside_loop_p (loop, bb)) | |
c59dabbe SP |
1314 | res = follow_ssa_edge_in_rhs (outer_loop, loop_phi_node, |
1315 | arg, halting_phi, | |
1316 | evolution_of_loop, limit); | |
1317 | if (res == t_true) | |
1318 | break; | |
9baba81b SP |
1319 | } |
1320 | ||
1321 | /* If the path crosses this loop-phi, give up. */ | |
c59dabbe | 1322 | if (res == t_true) |
9baba81b SP |
1323 | *evolution_of_loop = chrec_dont_know; |
1324 | ||
1325 | return res; | |
1326 | } | |
1327 | ||
1328 | /* Otherwise, compute the overall effect of the inner loop. */ | |
1329 | ev = compute_overall_effect_of_inner_loop (loop, ev); | |
1e8552eb | 1330 | return follow_ssa_edge_in_rhs (outer_loop, loop_phi_node, ev, halting_phi, |
c59dabbe | 1331 | evolution_of_loop, limit); |
9baba81b SP |
1332 | } |
1333 | ||
1334 | /* Follow an SSA edge from a loop-phi-node to itself, constructing a | |
1335 | path that is analyzed on the return walk. */ | |
1336 | ||
c59dabbe SP |
1337 | static t_bool |
1338 | follow_ssa_edge (struct loop *loop, tree def, tree halting_phi, | |
1339 | tree *evolution_of_loop, int limit) | |
9baba81b SP |
1340 | { |
1341 | struct loop *def_loop; | |
1342 | ||
1343 | if (TREE_CODE (def) == NOP_EXPR) | |
c59dabbe SP |
1344 | return t_false; |
1345 | ||
1346 | /* Give up if the path is longer than the MAX that we allow. */ | |
1347 | if (limit++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE)) | |
1348 | return t_dont_know; | |
9baba81b SP |
1349 | |
1350 | def_loop = loop_containing_stmt (def); | |
1351 | ||
1352 | switch (TREE_CODE (def)) | |
1353 | { | |
1354 | case PHI_NODE: | |
1355 | if (!loop_phi_node_p (def)) | |
1356 | /* DEF is a condition-phi-node. Follow the branches, and | |
1357 | record their evolutions. Finally, merge the collected | |
1358 | information and set the approximation to the main | |
1359 | variable. */ | |
1360 | return follow_ssa_edge_in_condition_phi | |
c59dabbe | 1361 | (loop, def, halting_phi, evolution_of_loop, limit); |
9baba81b SP |
1362 | |
1363 | /* When the analyzed phi is the halting_phi, the | |
1364 | depth-first search is over: we have found a path from | |
1365 | the halting_phi to itself in the loop. */ | |
1366 | if (def == halting_phi) | |
c59dabbe | 1367 | return t_true; |
9baba81b SP |
1368 | |
1369 | /* Otherwise, the evolution of the HALTING_PHI depends | |
89dbed81 | 1370 | on the evolution of another loop-phi-node, i.e. the |
9baba81b SP |
1371 | evolution function is a higher degree polynomial. */ |
1372 | if (def_loop == loop) | |
c59dabbe | 1373 | return t_false; |
9baba81b SP |
1374 | |
1375 | /* Inner loop. */ | |
1376 | if (flow_loop_nested_p (loop, def_loop)) | |
1377 | return follow_ssa_edge_inner_loop_phi | |
c59dabbe | 1378 | (loop, def, halting_phi, evolution_of_loop, limit); |
9baba81b SP |
1379 | |
1380 | /* Outer loop. */ | |
c59dabbe | 1381 | return t_false; |
9baba81b | 1382 | |
07beea0d | 1383 | case GIMPLE_MODIFY_STMT: |
1e8552eb | 1384 | return follow_ssa_edge_in_rhs (loop, def, |
07beea0d | 1385 | GIMPLE_STMT_OPERAND (def, 1), |
9baba81b | 1386 | halting_phi, |
c59dabbe | 1387 | evolution_of_loop, limit); |
9baba81b SP |
1388 | |
1389 | default: | |
1390 | /* At this level of abstraction, the program is just a set | |
07beea0d | 1391 | of GIMPLE_MODIFY_STMTs and PHI_NODEs. In principle there is no |
9baba81b | 1392 | other node to be handled. */ |
c59dabbe | 1393 | return t_false; |
9baba81b SP |
1394 | } |
1395 | } | |
1396 | ||
1397 | \f | |
1398 | ||
1399 | /* Given a LOOP_PHI_NODE, this function determines the evolution | |
1400 | function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */ | |
1401 | ||
1402 | static tree | |
1403 | analyze_evolution_in_loop (tree loop_phi_node, | |
1404 | tree init_cond) | |
1405 | { | |
1406 | int i; | |
1407 | tree evolution_function = chrec_not_analyzed_yet; | |
1408 | struct loop *loop = loop_containing_stmt (loop_phi_node); | |
1409 | basic_block bb; | |
1410 | ||
1411 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1412 | { | |
1413 | fprintf (dump_file, "(analyze_evolution_in_loop \n"); | |
1414 | fprintf (dump_file, " (loop_phi_node = "); | |
1415 | print_generic_expr (dump_file, loop_phi_node, 0); | |
1416 | fprintf (dump_file, ")\n"); | |
1417 | } | |
1418 | ||
1419 | for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++) | |
1420 | { | |
1421 | tree arg = PHI_ARG_DEF (loop_phi_node, i); | |
1422 | tree ssa_chain, ev_fn; | |
874caa00 | 1423 | t_bool res; |
9baba81b SP |
1424 | |
1425 | /* Select the edges that enter the loop body. */ | |
1426 | bb = PHI_ARG_EDGE (loop_phi_node, i)->src; | |
1427 | if (!flow_bb_inside_loop_p (loop, bb)) | |
1428 | continue; | |
1429 | ||
1430 | if (TREE_CODE (arg) == SSA_NAME) | |
1431 | { | |
1432 | ssa_chain = SSA_NAME_DEF_STMT (arg); | |
1433 | ||
1434 | /* Pass in the initial condition to the follow edge function. */ | |
1435 | ev_fn = init_cond; | |
c59dabbe | 1436 | res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0); |
9baba81b SP |
1437 | } |
1438 | else | |
874caa00 | 1439 | res = t_false; |
9baba81b SP |
1440 | |
1441 | /* When it is impossible to go back on the same | |
1442 | loop_phi_node by following the ssa edges, the | |
89dbed81 | 1443 | evolution is represented by a peeled chrec, i.e. the |
9baba81b SP |
1444 | first iteration, EV_FN has the value INIT_COND, then |
1445 | all the other iterations it has the value of ARG. | |
1446 | For the moment, PEELED_CHREC nodes are not built. */ | |
874caa00 | 1447 | if (res != t_true) |
9baba81b SP |
1448 | ev_fn = chrec_dont_know; |
1449 | ||
1450 | /* When there are multiple back edges of the loop (which in fact never | |
8c27b7d4 | 1451 | happens currently, but nevertheless), merge their evolutions. */ |
9baba81b SP |
1452 | evolution_function = chrec_merge (evolution_function, ev_fn); |
1453 | } | |
1454 | ||
1455 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1456 | { | |
1457 | fprintf (dump_file, " (evolution_function = "); | |
1458 | print_generic_expr (dump_file, evolution_function, 0); | |
1459 | fprintf (dump_file, "))\n"); | |
1460 | } | |
1461 | ||
1462 | return evolution_function; | |
1463 | } | |
1464 | ||
1465 | /* Given a loop-phi-node, return the initial conditions of the | |
1466 | variable on entry of the loop. When the CCP has propagated | |
1467 | constants into the loop-phi-node, the initial condition is | |
1468 | instantiated, otherwise the initial condition is kept symbolic. | |
1469 | This analyzer does not analyze the evolution outside the current | |
1470 | loop, and leaves this task to the on-demand tree reconstructor. */ | |
1471 | ||
1472 | static tree | |
1473 | analyze_initial_condition (tree loop_phi_node) | |
1474 | { | |
1475 | int i; | |
1476 | tree init_cond = chrec_not_analyzed_yet; | |
1477 | struct loop *loop = bb_for_stmt (loop_phi_node)->loop_father; | |
1478 | ||
1479 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1480 | { | |
1481 | fprintf (dump_file, "(analyze_initial_condition \n"); | |
1482 | fprintf (dump_file, " (loop_phi_node = \n"); | |
1483 | print_generic_expr (dump_file, loop_phi_node, 0); | |
1484 | fprintf (dump_file, ")\n"); | |
1485 | } | |
1486 | ||
1487 | for (i = 0; i < PHI_NUM_ARGS (loop_phi_node); i++) | |
1488 | { | |
1489 | tree branch = PHI_ARG_DEF (loop_phi_node, i); | |
1490 | basic_block bb = PHI_ARG_EDGE (loop_phi_node, i)->src; | |
1491 | ||
1492 | /* When the branch is oriented to the loop's body, it does | |
1493 | not contribute to the initial condition. */ | |
1494 | if (flow_bb_inside_loop_p (loop, bb)) | |
1495 | continue; | |
1496 | ||
1497 | if (init_cond == chrec_not_analyzed_yet) | |
1498 | { | |
1499 | init_cond = branch; | |
1500 | continue; | |
1501 | } | |
1502 | ||
1503 | if (TREE_CODE (branch) == SSA_NAME) | |
1504 | { | |
1505 | init_cond = chrec_dont_know; | |
1506 | break; | |
1507 | } | |
1508 | ||
1509 | init_cond = chrec_merge (init_cond, branch); | |
1510 | } | |
1511 | ||
1512 | /* Ooops -- a loop without an entry??? */ | |
1513 | if (init_cond == chrec_not_analyzed_yet) | |
1514 | init_cond = chrec_dont_know; | |
1515 | ||
1516 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1517 | { | |
1518 | fprintf (dump_file, " (init_cond = "); | |
1519 | print_generic_expr (dump_file, init_cond, 0); | |
1520 | fprintf (dump_file, "))\n"); | |
1521 | } | |
1522 | ||
1523 | return init_cond; | |
1524 | } | |
1525 | ||
1526 | /* Analyze the scalar evolution for LOOP_PHI_NODE. */ | |
1527 | ||
1528 | static tree | |
1529 | interpret_loop_phi (struct loop *loop, tree loop_phi_node) | |
1530 | { | |
1531 | tree res; | |
1532 | struct loop *phi_loop = loop_containing_stmt (loop_phi_node); | |
1533 | tree init_cond; | |
1534 | ||
1535 | if (phi_loop != loop) | |
1536 | { | |
1537 | struct loop *subloop; | |
1538 | tree evolution_fn = analyze_scalar_evolution | |
1539 | (phi_loop, PHI_RESULT (loop_phi_node)); | |
1540 | ||
1541 | /* Dive one level deeper. */ | |
1542 | subloop = superloop_at_depth (phi_loop, loop->depth + 1); | |
1543 | ||
1544 | /* Interpret the subloop. */ | |
1545 | res = compute_overall_effect_of_inner_loop (subloop, evolution_fn); | |
1546 | return res; | |
1547 | } | |
1548 | ||
1549 | /* Otherwise really interpret the loop phi. */ | |
1550 | init_cond = analyze_initial_condition (loop_phi_node); | |
1551 | res = analyze_evolution_in_loop (loop_phi_node, init_cond); | |
1552 | ||
1553 | return res; | |
1554 | } | |
1555 | ||
1556 | /* This function merges the branches of a condition-phi-node, | |
1557 | contained in the outermost loop, and whose arguments are already | |
1558 | analyzed. */ | |
1559 | ||
1560 | static tree | |
1561 | interpret_condition_phi (struct loop *loop, tree condition_phi) | |
1562 | { | |
1563 | int i; | |
1564 | tree res = chrec_not_analyzed_yet; | |
1565 | ||
1566 | for (i = 0; i < PHI_NUM_ARGS (condition_phi); i++) | |
1567 | { | |
1568 | tree branch_chrec; | |
1569 | ||
1570 | if (backedge_phi_arg_p (condition_phi, i)) | |
1571 | { | |
1572 | res = chrec_dont_know; | |
1573 | break; | |
1574 | } | |
1575 | ||
1576 | branch_chrec = analyze_scalar_evolution | |
1577 | (loop, PHI_ARG_DEF (condition_phi, i)); | |
1578 | ||
1579 | res = chrec_merge (res, branch_chrec); | |
1580 | } | |
1581 | ||
1582 | return res; | |
1583 | } | |
1584 | ||
07beea0d | 1585 | /* Interpret the right hand side of a GIMPLE_MODIFY_STMT OPND1. If we didn't |
29836d07 | 1586 | analyze this node before, follow the definitions until ending |
07beea0d | 1587 | either on an analyzed GIMPLE_MODIFY_STMT, or on a loop-phi-node. On the |
9baba81b SP |
1588 | return path, this function propagates evolutions (ala constant copy |
1589 | propagation). OPND1 is not a GIMPLE expression because we could | |
1590 | analyze the effect of an inner loop: see interpret_loop_phi. */ | |
1591 | ||
1592 | static tree | |
07beea0d AH |
1593 | interpret_rhs_modify_stmt (struct loop *loop, tree at_stmt, |
1594 | tree opnd1, tree type) | |
9baba81b SP |
1595 | { |
1596 | tree res, opnd10, opnd11, chrec10, chrec11; | |
1e8552eb | 1597 | |
9baba81b | 1598 | if (is_gimple_min_invariant (opnd1)) |
1e8552eb SP |
1599 | return chrec_convert (type, opnd1, at_stmt); |
1600 | ||
9baba81b SP |
1601 | switch (TREE_CODE (opnd1)) |
1602 | { | |
1603 | case PLUS_EXPR: | |
1604 | opnd10 = TREE_OPERAND (opnd1, 0); | |
1605 | opnd11 = TREE_OPERAND (opnd1, 1); | |
1606 | chrec10 = analyze_scalar_evolution (loop, opnd10); | |
1607 | chrec11 = analyze_scalar_evolution (loop, opnd11); | |
1e8552eb SP |
1608 | chrec10 = chrec_convert (type, chrec10, at_stmt); |
1609 | chrec11 = chrec_convert (type, chrec11, at_stmt); | |
9baba81b SP |
1610 | res = chrec_fold_plus (type, chrec10, chrec11); |
1611 | break; | |
1612 | ||
1613 | case MINUS_EXPR: | |
1614 | opnd10 = TREE_OPERAND (opnd1, 0); | |
1615 | opnd11 = TREE_OPERAND (opnd1, 1); | |
1616 | chrec10 = analyze_scalar_evolution (loop, opnd10); | |
1617 | chrec11 = analyze_scalar_evolution (loop, opnd11); | |
1e8552eb SP |
1618 | chrec10 = chrec_convert (type, chrec10, at_stmt); |
1619 | chrec11 = chrec_convert (type, chrec11, at_stmt); | |
9baba81b SP |
1620 | res = chrec_fold_minus (type, chrec10, chrec11); |
1621 | break; | |
1622 | ||
1623 | case NEGATE_EXPR: | |
1624 | opnd10 = TREE_OPERAND (opnd1, 0); | |
1625 | chrec10 = analyze_scalar_evolution (loop, opnd10); | |
1e8552eb | 1626 | chrec10 = chrec_convert (type, chrec10, at_stmt); |
9a75ede0 RS |
1627 | /* TYPE may be integer, real or complex, so use fold_convert. */ |
1628 | res = chrec_fold_multiply (type, chrec10, | |
1629 | fold_convert (type, integer_minus_one_node)); | |
9baba81b SP |
1630 | break; |
1631 | ||
1632 | case MULT_EXPR: | |
1633 | opnd10 = TREE_OPERAND (opnd1, 0); | |
1634 | opnd11 = TREE_OPERAND (opnd1, 1); | |
1635 | chrec10 = analyze_scalar_evolution (loop, opnd10); | |
1636 | chrec11 = analyze_scalar_evolution (loop, opnd11); | |
1e8552eb SP |
1637 | chrec10 = chrec_convert (type, chrec10, at_stmt); |
1638 | chrec11 = chrec_convert (type, chrec11, at_stmt); | |
9baba81b SP |
1639 | res = chrec_fold_multiply (type, chrec10, chrec11); |
1640 | break; | |
1641 | ||
1642 | case SSA_NAME: | |
1e8552eb SP |
1643 | res = chrec_convert (type, analyze_scalar_evolution (loop, opnd1), |
1644 | at_stmt); | |
9baba81b | 1645 | break; |
0bca51f0 DN |
1646 | |
1647 | case ASSERT_EXPR: | |
1648 | opnd10 = ASSERT_EXPR_VAR (opnd1); | |
1e8552eb SP |
1649 | res = chrec_convert (type, analyze_scalar_evolution (loop, opnd10), |
1650 | at_stmt); | |
0bca51f0 | 1651 | break; |
9baba81b SP |
1652 | |
1653 | case NOP_EXPR: | |
1654 | case CONVERT_EXPR: | |
1655 | opnd10 = TREE_OPERAND (opnd1, 0); | |
1656 | chrec10 = analyze_scalar_evolution (loop, opnd10); | |
1e8552eb | 1657 | res = chrec_convert (type, chrec10, at_stmt); |
9baba81b SP |
1658 | break; |
1659 | ||
1660 | default: | |
1661 | res = chrec_dont_know; | |
1662 | break; | |
1663 | } | |
1664 | ||
1665 | return res; | |
1666 | } | |
1667 | ||
1668 | \f | |
1669 | ||
1670 | /* This section contains all the entry points: | |
1671 | - number_of_iterations_in_loop, | |
1672 | - analyze_scalar_evolution, | |
1673 | - instantiate_parameters. | |
1674 | */ | |
1675 | ||
1676 | /* Compute and return the evolution function in WRTO_LOOP, the nearest | |
1677 | common ancestor of DEF_LOOP and USE_LOOP. */ | |
1678 | ||
1679 | static tree | |
1680 | compute_scalar_evolution_in_loop (struct loop *wrto_loop, | |
1681 | struct loop *def_loop, | |
1682 | tree ev) | |
1683 | { | |
1684 | tree res; | |
1685 | if (def_loop == wrto_loop) | |
1686 | return ev; | |
1687 | ||
1688 | def_loop = superloop_at_depth (def_loop, wrto_loop->depth + 1); | |
1689 | res = compute_overall_effect_of_inner_loop (def_loop, ev); | |
1690 | ||
1691 | return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet); | |
1692 | } | |
1693 | ||
20527215 ZD |
1694 | /* Folds EXPR, if it is a cast to pointer, assuming that the created |
1695 | polynomial_chrec does not wrap. */ | |
1696 | ||
1697 | static tree | |
1698 | fold_used_pointer_cast (tree expr) | |
1699 | { | |
1700 | tree op; | |
1701 | tree type, inner_type; | |
1702 | ||
1703 | if (TREE_CODE (expr) != NOP_EXPR && TREE_CODE (expr) != CONVERT_EXPR) | |
1704 | return expr; | |
1705 | ||
1706 | op = TREE_OPERAND (expr, 0); | |
1707 | if (TREE_CODE (op) != POLYNOMIAL_CHREC) | |
1708 | return expr; | |
1709 | ||
1710 | type = TREE_TYPE (expr); | |
1711 | inner_type = TREE_TYPE (op); | |
1712 | ||
1713 | if (!INTEGRAL_TYPE_P (inner_type) | |
1714 | || TYPE_PRECISION (inner_type) != TYPE_PRECISION (type)) | |
1715 | return expr; | |
1716 | ||
1717 | return build_polynomial_chrec (CHREC_VARIABLE (op), | |
1718 | chrec_convert (type, CHREC_LEFT (op), NULL_TREE), | |
1719 | chrec_convert (type, CHREC_RIGHT (op), NULL_TREE)); | |
1720 | } | |
1721 | ||
1722 | /* Returns true if EXPR is an expression corresponding to offset of pointer | |
1723 | in p + offset. */ | |
1724 | ||
1725 | static bool | |
1726 | pointer_offset_p (tree expr) | |
1727 | { | |
1728 | if (TREE_CODE (expr) == INTEGER_CST) | |
1729 | return true; | |
1730 | ||
1731 | if ((TREE_CODE (expr) == NOP_EXPR || TREE_CODE (expr) == CONVERT_EXPR) | |
1732 | && INTEGRAL_TYPE_P (TREE_TYPE (TREE_OPERAND (expr, 0)))) | |
1733 | return true; | |
1734 | ||
1735 | return false; | |
1736 | } | |
1737 | ||
1738 | /* EXPR is a scalar evolution of a pointer that is dereferenced or used in | |
1739 | comparison. This means that it must point to a part of some object in | |
1740 | memory, which enables us to argue about overflows and possibly simplify | |
8a613cae RG |
1741 | the EXPR. AT_STMT is the statement in which this conversion has to be |
1742 | performed. Returns the simplified value. | |
20527215 ZD |
1743 | |
1744 | Currently, for | |
1745 | ||
1746 | int i, n; | |
1747 | int *p; | |
1748 | ||
1749 | for (i = -n; i < n; i++) | |
1750 | *(p + i) = ...; | |
1751 | ||
1752 | We generate the following code (assuming that size of int and size_t is | |
1753 | 4 bytes): | |
1754 | ||
1755 | for (i = -n; i < n; i++) | |
1756 | { | |
1757 | size_t tmp1, tmp2; | |
1758 | int *tmp3, *tmp4; | |
1759 | ||
1760 | tmp1 = (size_t) i; (1) | |
1761 | tmp2 = 4 * tmp1; (2) | |
1762 | tmp3 = (int *) tmp2; (3) | |
1763 | tmp4 = p + tmp3; (4) | |
1764 | ||
1765 | *tmp4 = ...; | |
1766 | } | |
1767 | ||
1768 | We in general assume that pointer arithmetics does not overflow (since its | |
1769 | behavior is undefined in that case). One of the problems is that our | |
1770 | translation does not capture this property very well -- (int *) is | |
1771 | considered unsigned, hence the computation in (4) does overflow if i is | |
1772 | negative. | |
1773 | ||
1774 | This impreciseness creates complications in scev analysis. The scalar | |
1775 | evolution of i is [-n, +, 1]. Since int and size_t have the same precision | |
1776 | (in this example), and size_t is unsigned (so we do not care about | |
1777 | overflows), we succeed to derive that scev of tmp1 is [(size_t) -n, +, 1] | |
1778 | and scev of tmp2 is [4 * (size_t) -n, +, 4]. With tmp3, we run into | |
1779 | problem -- [(int *) (4 * (size_t) -n), +, 4] wraps, and since we on several | |
1780 | places assume that this is not the case for scevs with pointer type, we | |
1781 | cannot use this scev for tmp3; hence, its scev is | |
1782 | (int *) [(4 * (size_t) -n), +, 4], and scev of tmp4 is | |
1783 | p + (int *) [(4 * (size_t) -n), +, 4]. Most of the optimizers are unable to | |
1784 | work with scevs of this shape. | |
1785 | ||
1786 | However, since tmp4 is dereferenced, all its values must belong to a single | |
1787 | object, and taking into account that the precision of int * and size_t is | |
1788 | the same, it is impossible for its scev to wrap. Hence, we can derive that | |
1789 | its evolution is [p + (int *) (4 * (size_t) -n), +, 4], which the optimizers | |
1790 | can work with. | |
1791 | ||
1792 | ??? Maybe we should use different representation for pointer arithmetics, | |
1793 | however that is a long-term project with a lot of potential for creating | |
1794 | bugs. */ | |
1795 | ||
1796 | static tree | |
8a613cae | 1797 | fold_used_pointer (tree expr, tree at_stmt) |
20527215 ZD |
1798 | { |
1799 | tree op0, op1, new0, new1; | |
1800 | enum tree_code code = TREE_CODE (expr); | |
1801 | ||
1802 | if (code == PLUS_EXPR | |
1803 | || code == MINUS_EXPR) | |
1804 | { | |
1805 | op0 = TREE_OPERAND (expr, 0); | |
1806 | op1 = TREE_OPERAND (expr, 1); | |
1807 | ||
1808 | if (pointer_offset_p (op1)) | |
1809 | { | |
8a613cae | 1810 | new0 = fold_used_pointer (op0, at_stmt); |
20527215 ZD |
1811 | new1 = fold_used_pointer_cast (op1); |
1812 | } | |
1813 | else if (code == PLUS_EXPR && pointer_offset_p (op0)) | |
1814 | { | |
1815 | new0 = fold_used_pointer_cast (op0); | |
8a613cae | 1816 | new1 = fold_used_pointer (op1, at_stmt); |
20527215 ZD |
1817 | } |
1818 | else | |
1819 | return expr; | |
1820 | ||
1821 | if (new0 == op0 && new1 == op1) | |
1822 | return expr; | |
1823 | ||
8a613cae RG |
1824 | new0 = chrec_convert (TREE_TYPE (expr), new0, at_stmt); |
1825 | new1 = chrec_convert (TREE_TYPE (expr), new1, at_stmt); | |
1826 | ||
20527215 ZD |
1827 | if (code == PLUS_EXPR) |
1828 | expr = chrec_fold_plus (TREE_TYPE (expr), new0, new1); | |
1829 | else | |
1830 | expr = chrec_fold_minus (TREE_TYPE (expr), new0, new1); | |
1831 | ||
1832 | return expr; | |
1833 | } | |
1834 | else | |
1835 | return fold_used_pointer_cast (expr); | |
1836 | } | |
1837 | ||
1838 | /* Returns true if PTR is dereferenced, or used in comparison. */ | |
1839 | ||
1840 | static bool | |
1841 | pointer_used_p (tree ptr) | |
1842 | { | |
1843 | use_operand_p use_p; | |
1844 | imm_use_iterator imm_iter; | |
1845 | tree stmt, rhs; | |
1846 | struct ptr_info_def *pi = get_ptr_info (ptr); | |
20527215 ZD |
1847 | |
1848 | /* Check whether the pointer has a memory tag; if it does, it is | |
1849 | (or at least used to be) dereferenced. */ | |
1850 | if ((pi != NULL && pi->name_mem_tag != NULL) | |
cfaab3a9 | 1851 | || symbol_mem_tag (SSA_NAME_VAR (ptr))) |
20527215 ZD |
1852 | return true; |
1853 | ||
1854 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, ptr) | |
1855 | { | |
1856 | stmt = USE_STMT (use_p); | |
1857 | if (TREE_CODE (stmt) == COND_EXPR) | |
1858 | return true; | |
1859 | ||
07beea0d | 1860 | if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT) |
20527215 ZD |
1861 | continue; |
1862 | ||
07beea0d | 1863 | rhs = GIMPLE_STMT_OPERAND (stmt, 1); |
20527215 ZD |
1864 | if (!COMPARISON_CLASS_P (rhs)) |
1865 | continue; | |
1866 | ||
07beea0d AH |
1867 | if (GIMPLE_STMT_OPERAND (stmt, 0) == ptr |
1868 | || GIMPLE_STMT_OPERAND (stmt, 1) == ptr) | |
20527215 ZD |
1869 | return true; |
1870 | } | |
1871 | ||
1872 | return false; | |
1873 | } | |
1874 | ||
9baba81b SP |
1875 | /* Helper recursive function. */ |
1876 | ||
1877 | static tree | |
1878 | analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res) | |
1879 | { | |
1880 | tree def, type = TREE_TYPE (var); | |
1881 | basic_block bb; | |
1882 | struct loop *def_loop; | |
1883 | ||
42d375ed | 1884 | if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE) |
9baba81b SP |
1885 | return chrec_dont_know; |
1886 | ||
1887 | if (TREE_CODE (var) != SSA_NAME) | |
07beea0d | 1888 | return interpret_rhs_modify_stmt (loop, NULL_TREE, var, type); |
9baba81b SP |
1889 | |
1890 | def = SSA_NAME_DEF_STMT (var); | |
1891 | bb = bb_for_stmt (def); | |
1892 | def_loop = bb ? bb->loop_father : NULL; | |
1893 | ||
1894 | if (bb == NULL | |
1895 | || !flow_bb_inside_loop_p (loop, bb)) | |
1896 | { | |
1897 | /* Keep the symbolic form. */ | |
1898 | res = var; | |
1899 | goto set_and_end; | |
1900 | } | |
1901 | ||
1902 | if (res != chrec_not_analyzed_yet) | |
1903 | { | |
1904 | if (loop != bb->loop_father) | |
1905 | res = compute_scalar_evolution_in_loop | |
1906 | (find_common_loop (loop, bb->loop_father), bb->loop_father, res); | |
1907 | ||
1908 | goto set_and_end; | |
1909 | } | |
1910 | ||
1911 | if (loop != def_loop) | |
1912 | { | |
1913 | res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet); | |
1914 | res = compute_scalar_evolution_in_loop (loop, def_loop, res); | |
1915 | ||
1916 | goto set_and_end; | |
1917 | } | |
1918 | ||
1919 | switch (TREE_CODE (def)) | |
1920 | { | |
07beea0d AH |
1921 | case GIMPLE_MODIFY_STMT: |
1922 | res = interpret_rhs_modify_stmt (loop, def, | |
1923 | GIMPLE_STMT_OPERAND (def, 1), type); | |
20527215 ZD |
1924 | |
1925 | if (POINTER_TYPE_P (type) | |
1926 | && !automatically_generated_chrec_p (res) | |
1927 | && pointer_used_p (var)) | |
8a613cae | 1928 | res = fold_used_pointer (res, def); |
9baba81b SP |
1929 | break; |
1930 | ||
1931 | case PHI_NODE: | |
1932 | if (loop_phi_node_p (def)) | |
1933 | res = interpret_loop_phi (loop, def); | |
1934 | else | |
1935 | res = interpret_condition_phi (loop, def); | |
1936 | break; | |
1937 | ||
1938 | default: | |
1939 | res = chrec_dont_know; | |
1940 | break; | |
1941 | } | |
1942 | ||
1943 | set_and_end: | |
1944 | ||
1945 | /* Keep the symbolic form. */ | |
1946 | if (res == chrec_dont_know) | |
1947 | res = var; | |
1948 | ||
1949 | if (loop == def_loop) | |
1950 | set_scalar_evolution (var, res); | |
1951 | ||
1952 | return res; | |
1953 | } | |
1954 | ||
1955 | /* Entry point for the scalar evolution analyzer. | |
1956 | Analyzes and returns the scalar evolution of the ssa_name VAR. | |
1957 | LOOP_NB is the identifier number of the loop in which the variable | |
1958 | is used. | |
1959 | ||
1960 | Example of use: having a pointer VAR to a SSA_NAME node, STMT a | |
1961 | pointer to the statement that uses this variable, in order to | |
1962 | determine the evolution function of the variable, use the following | |
1963 | calls: | |
1964 | ||
1965 | unsigned loop_nb = loop_containing_stmt (stmt)->num; | |
1966 | tree chrec_with_symbols = analyze_scalar_evolution (loop_nb, var); | |
1967 | tree chrec_instantiated = instantiate_parameters | |
1968 | (loop_nb, chrec_with_symbols); | |
1969 | */ | |
1970 | ||
1971 | tree | |
1972 | analyze_scalar_evolution (struct loop *loop, tree var) | |
1973 | { | |
1974 | tree res; | |
1975 | ||
1976 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1977 | { | |
1978 | fprintf (dump_file, "(analyze_scalar_evolution \n"); | |
1979 | fprintf (dump_file, " (loop_nb = %d)\n", loop->num); | |
1980 | fprintf (dump_file, " (scalar = "); | |
1981 | print_generic_expr (dump_file, var, 0); | |
1982 | fprintf (dump_file, ")\n"); | |
1983 | } | |
1984 | ||
1985 | res = analyze_scalar_evolution_1 (loop, var, get_scalar_evolution (var)); | |
1986 | ||
1987 | if (TREE_CODE (var) == SSA_NAME && res == chrec_dont_know) | |
1988 | res = var; | |
1989 | ||
1990 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1991 | fprintf (dump_file, ")\n"); | |
1992 | ||
1993 | return res; | |
1994 | } | |
1995 | ||
1996 | /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to | |
1997 | WRTO_LOOP (which should be a superloop of both USE_LOOP and definition | |
a6f778b2 ZD |
1998 | of VERSION). |
1999 | ||
2000 | FOLDED_CASTS is set to true if resolve_mixers used | |
2001 | chrec_convert_aggressive (TODO -- not really, we are way too conservative | |
2002 | at the moment in order to keep things simple). */ | |
9baba81b SP |
2003 | |
2004 | static tree | |
2005 | analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop, | |
a6f778b2 | 2006 | tree version, bool *folded_casts) |
9baba81b SP |
2007 | { |
2008 | bool val = false; | |
a6f778b2 | 2009 | tree ev = version, tmp; |
9baba81b | 2010 | |
a6f778b2 ZD |
2011 | if (folded_casts) |
2012 | *folded_casts = false; | |
9baba81b SP |
2013 | while (1) |
2014 | { | |
a6f778b2 ZD |
2015 | tmp = analyze_scalar_evolution (use_loop, ev); |
2016 | ev = resolve_mixers (use_loop, tmp); | |
2017 | ||
2018 | if (folded_casts && tmp != ev) | |
2019 | *folded_casts = true; | |
9baba81b SP |
2020 | |
2021 | if (use_loop == wrto_loop) | |
2022 | return ev; | |
2023 | ||
2024 | /* If the value of the use changes in the inner loop, we cannot express | |
2025 | its value in the outer loop (we might try to return interval chrec, | |
2026 | but we do not have a user for it anyway) */ | |
2027 | if (!no_evolution_in_loop_p (ev, use_loop->num, &val) | |
2028 | || !val) | |
2029 | return chrec_dont_know; | |
2030 | ||
2031 | use_loop = use_loop->outer; | |
2032 | } | |
2033 | } | |
2034 | ||
eb0bc7af ZD |
2035 | /* Returns instantiated value for VERSION in CACHE. */ |
2036 | ||
2037 | static tree | |
2038 | get_instantiated_value (htab_t cache, tree version) | |
2039 | { | |
2040 | struct scev_info_str *info, pattern; | |
2041 | ||
2042 | pattern.var = version; | |
858904db | 2043 | info = (struct scev_info_str *) htab_find (cache, &pattern); |
eb0bc7af ZD |
2044 | |
2045 | if (info) | |
2046 | return info->chrec; | |
2047 | else | |
2048 | return NULL_TREE; | |
2049 | } | |
2050 | ||
2051 | /* Sets instantiated value for VERSION to VAL in CACHE. */ | |
2052 | ||
2053 | static void | |
2054 | set_instantiated_value (htab_t cache, tree version, tree val) | |
2055 | { | |
2056 | struct scev_info_str *info, pattern; | |
2057 | PTR *slot; | |
2058 | ||
2059 | pattern.var = version; | |
2060 | slot = htab_find_slot (cache, &pattern, INSERT); | |
2061 | ||
cceb1885 GDR |
2062 | if (!*slot) |
2063 | *slot = new_scev_info_str (version); | |
2064 | info = (struct scev_info_str *) *slot; | |
eb0bc7af ZD |
2065 | info->chrec = val; |
2066 | } | |
2067 | ||
18aed06a SP |
2068 | /* Return the closed_loop_phi node for VAR. If there is none, return |
2069 | NULL_TREE. */ | |
2070 | ||
2071 | static tree | |
2072 | loop_closed_phi_def (tree var) | |
2073 | { | |
2074 | struct loop *loop; | |
2075 | edge exit; | |
2076 | tree phi; | |
2077 | ||
2078 | if (var == NULL_TREE | |
2079 | || TREE_CODE (var) != SSA_NAME) | |
2080 | return NULL_TREE; | |
2081 | ||
2082 | loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var)); | |
ac8f6c69 | 2083 | exit = single_exit (loop); |
18aed06a SP |
2084 | if (!exit) |
2085 | return NULL_TREE; | |
2086 | ||
2087 | for (phi = phi_nodes (exit->dest); phi; phi = PHI_CHAIN (phi)) | |
2088 | if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var) | |
2089 | return PHI_RESULT (phi); | |
2090 | ||
2091 | return NULL_TREE; | |
2092 | } | |
2093 | ||
9baba81b | 2094 | /* Analyze all the parameters of the chrec that were left under a symbolic form, |
2282a0e6 ZD |
2095 | with respect to LOOP. CHREC is the chrec to instantiate. CACHE is the cache |
2096 | of already instantiated values. FLAGS modify the way chrecs are | |
47ae9e4c SP |
2097 | instantiated. SIZE_EXPR is used for computing the size of the expression to |
2098 | be instantiated, and to stop if it exceeds some limit. */ | |
9baba81b | 2099 | |
2282a0e6 ZD |
2100 | /* Values for FLAGS. */ |
2101 | enum | |
2102 | { | |
2103 | INSERT_SUPERLOOP_CHRECS = 1, /* Loop invariants are replaced with chrecs | |
2104 | in outer loops. */ | |
2105 | FOLD_CONVERSIONS = 2 /* The conversions that may wrap in | |
2106 | signed/pointer type are folded, as long as the | |
2107 | value of the chrec is preserved. */ | |
2108 | }; | |
2109 | ||
9baba81b | 2110 | static tree |
47ae9e4c SP |
2111 | instantiate_parameters_1 (struct loop *loop, tree chrec, int flags, htab_t cache, |
2112 | int size_expr) | |
9baba81b SP |
2113 | { |
2114 | tree res, op0, op1, op2; | |
2115 | basic_block def_bb; | |
2116 | struct loop *def_loop; | |
16a2acea | 2117 | tree type = chrec_type (chrec); |
2282a0e6 | 2118 | |
47ae9e4c SP |
2119 | /* Give up if the expression is larger than the MAX that we allow. */ |
2120 | if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE)) | |
2121 | return chrec_dont_know; | |
2122 | ||
d7770457 SP |
2123 | if (automatically_generated_chrec_p (chrec) |
2124 | || is_gimple_min_invariant (chrec)) | |
9baba81b SP |
2125 | return chrec; |
2126 | ||
2127 | switch (TREE_CODE (chrec)) | |
2128 | { | |
2129 | case SSA_NAME: | |
2130 | def_bb = bb_for_stmt (SSA_NAME_DEF_STMT (chrec)); | |
2131 | ||
2132 | /* A parameter (or loop invariant and we do not want to include | |
2133 | evolutions in outer loops), nothing to do. */ | |
2134 | if (!def_bb | |
2282a0e6 | 2135 | || (!(flags & INSERT_SUPERLOOP_CHRECS) |
9baba81b SP |
2136 | && !flow_bb_inside_loop_p (loop, def_bb))) |
2137 | return chrec; | |
2138 | ||
eb0bc7af ZD |
2139 | /* We cache the value of instantiated variable to avoid exponential |
2140 | time complexity due to reevaluations. We also store the convenient | |
2141 | value in the cache in order to prevent infinite recursion -- we do | |
2142 | not want to instantiate the SSA_NAME if it is in a mixer | |
9baba81b SP |
2143 | structure. This is used for avoiding the instantiation of |
2144 | recursively defined functions, such as: | |
2145 | ||
2146 | | a_2 -> {0, +, 1, +, a_2}_1 */ | |
eb0bc7af ZD |
2147 | |
2148 | res = get_instantiated_value (cache, chrec); | |
2149 | if (res) | |
2150 | return res; | |
2151 | ||
2152 | /* Store the convenient value for chrec in the structure. If it | |
2153 | is defined outside of the loop, we may just leave it in symbolic | |
2154 | form, otherwise we need to admit that we do not know its behavior | |
2155 | inside the loop. */ | |
2156 | res = !flow_bb_inside_loop_p (loop, def_bb) ? chrec : chrec_dont_know; | |
2157 | set_instantiated_value (cache, chrec, res); | |
2158 | ||
2159 | /* To make things even more complicated, instantiate_parameters_1 | |
2160 | calls analyze_scalar_evolution that may call # of iterations | |
2161 | analysis that may in turn call instantiate_parameters_1 again. | |
2162 | To prevent the infinite recursion, keep also the bitmap of | |
2163 | ssa names that are being instantiated globally. */ | |
9baba81b | 2164 | if (bitmap_bit_p (already_instantiated, SSA_NAME_VERSION (chrec))) |
eb0bc7af | 2165 | return res; |
9baba81b SP |
2166 | |
2167 | def_loop = find_common_loop (loop, def_bb->loop_father); | |
2168 | ||
2169 | /* If the analysis yields a parametric chrec, instantiate the | |
eb0bc7af | 2170 | result again. */ |
9baba81b SP |
2171 | bitmap_set_bit (already_instantiated, SSA_NAME_VERSION (chrec)); |
2172 | res = analyze_scalar_evolution (def_loop, chrec); | |
18aed06a SP |
2173 | |
2174 | /* Don't instantiate loop-closed-ssa phi nodes. */ | |
2175 | if (TREE_CODE (res) == SSA_NAME | |
2176 | && (loop_containing_stmt (SSA_NAME_DEF_STMT (res)) == NULL | |
2177 | || (loop_containing_stmt (SSA_NAME_DEF_STMT (res))->depth | |
2178 | > def_loop->depth))) | |
2179 | { | |
2180 | if (res == chrec) | |
2181 | res = loop_closed_phi_def (chrec); | |
2182 | else | |
2183 | res = chrec; | |
2184 | ||
2185 | if (res == NULL_TREE) | |
2186 | res = chrec_dont_know; | |
2187 | } | |
2188 | ||
2189 | else if (res != chrec_dont_know) | |
47ae9e4c | 2190 | res = instantiate_parameters_1 (loop, res, flags, cache, size_expr); |
18aed06a | 2191 | |
9baba81b | 2192 | bitmap_clear_bit (already_instantiated, SSA_NAME_VERSION (chrec)); |
eb0bc7af ZD |
2193 | |
2194 | /* Store the correct value to the cache. */ | |
2195 | set_instantiated_value (cache, chrec, res); | |
9baba81b SP |
2196 | return res; |
2197 | ||
2198 | case POLYNOMIAL_CHREC: | |
2199 | op0 = instantiate_parameters_1 (loop, CHREC_LEFT (chrec), | |
47ae9e4c | 2200 | flags, cache, size_expr); |
fca81712 SP |
2201 | if (op0 == chrec_dont_know) |
2202 | return chrec_dont_know; | |
2203 | ||
9baba81b | 2204 | op1 = instantiate_parameters_1 (loop, CHREC_RIGHT (chrec), |
47ae9e4c | 2205 | flags, cache, size_expr); |
fca81712 SP |
2206 | if (op1 == chrec_dont_know) |
2207 | return chrec_dont_know; | |
2208 | ||
eac30183 ZD |
2209 | if (CHREC_LEFT (chrec) != op0 |
2210 | || CHREC_RIGHT (chrec) != op1) | |
e2157b49 SP |
2211 | { |
2212 | op1 = chrec_convert (chrec_type (op0), op1, NULL_TREE); | |
2213 | chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1); | |
2214 | } | |
eac30183 | 2215 | return chrec; |
9baba81b SP |
2216 | |
2217 | case PLUS_EXPR: | |
2218 | op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0), | |
47ae9e4c | 2219 | flags, cache, size_expr); |
fca81712 SP |
2220 | if (op0 == chrec_dont_know) |
2221 | return chrec_dont_know; | |
2222 | ||
9baba81b | 2223 | op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1), |
47ae9e4c | 2224 | flags, cache, size_expr); |
fca81712 SP |
2225 | if (op1 == chrec_dont_know) |
2226 | return chrec_dont_know; | |
2227 | ||
eac30183 ZD |
2228 | if (TREE_OPERAND (chrec, 0) != op0 |
2229 | || TREE_OPERAND (chrec, 1) != op1) | |
16a2acea SP |
2230 | { |
2231 | op0 = chrec_convert (type, op0, NULL_TREE); | |
2232 | op1 = chrec_convert (type, op1, NULL_TREE); | |
2233 | chrec = chrec_fold_plus (type, op0, op1); | |
2234 | } | |
eac30183 | 2235 | return chrec; |
9baba81b SP |
2236 | |
2237 | case MINUS_EXPR: | |
2238 | op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0), | |
47ae9e4c | 2239 | flags, cache, size_expr); |
fca81712 SP |
2240 | if (op0 == chrec_dont_know) |
2241 | return chrec_dont_know; | |
2242 | ||
9baba81b | 2243 | op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1), |
47ae9e4c | 2244 | flags, cache, size_expr); |
fca81712 SP |
2245 | if (op1 == chrec_dont_know) |
2246 | return chrec_dont_know; | |
2247 | ||
eac30183 ZD |
2248 | if (TREE_OPERAND (chrec, 0) != op0 |
2249 | || TREE_OPERAND (chrec, 1) != op1) | |
16a2acea SP |
2250 | { |
2251 | op0 = chrec_convert (type, op0, NULL_TREE); | |
2252 | op1 = chrec_convert (type, op1, NULL_TREE); | |
2253 | chrec = chrec_fold_minus (type, op0, op1); | |
2254 | } | |
eac30183 | 2255 | return chrec; |
9baba81b SP |
2256 | |
2257 | case MULT_EXPR: | |
2258 | op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0), | |
47ae9e4c | 2259 | flags, cache, size_expr); |
fca81712 SP |
2260 | if (op0 == chrec_dont_know) |
2261 | return chrec_dont_know; | |
2262 | ||
9baba81b | 2263 | op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1), |
47ae9e4c | 2264 | flags, cache, size_expr); |
fca81712 SP |
2265 | if (op1 == chrec_dont_know) |
2266 | return chrec_dont_know; | |
2267 | ||
eac30183 ZD |
2268 | if (TREE_OPERAND (chrec, 0) != op0 |
2269 | || TREE_OPERAND (chrec, 1) != op1) | |
16a2acea SP |
2270 | { |
2271 | op0 = chrec_convert (type, op0, NULL_TREE); | |
2272 | op1 = chrec_convert (type, op1, NULL_TREE); | |
2273 | chrec = chrec_fold_multiply (type, op0, op1); | |
2274 | } | |
eac30183 | 2275 | return chrec; |
9baba81b SP |
2276 | |
2277 | case NOP_EXPR: | |
2278 | case CONVERT_EXPR: | |
2279 | case NON_LVALUE_EXPR: | |
2280 | op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0), | |
47ae9e4c | 2281 | flags, cache, size_expr); |
9baba81b SP |
2282 | if (op0 == chrec_dont_know) |
2283 | return chrec_dont_know; | |
2284 | ||
2282a0e6 ZD |
2285 | if (flags & FOLD_CONVERSIONS) |
2286 | { | |
2287 | tree tmp = chrec_convert_aggressive (TREE_TYPE (chrec), op0); | |
2288 | if (tmp) | |
2289 | return tmp; | |
2290 | } | |
2291 | ||
eac30183 ZD |
2292 | if (op0 == TREE_OPERAND (chrec, 0)) |
2293 | return chrec; | |
2294 | ||
d7f5de76 ZD |
2295 | /* If we used chrec_convert_aggressive, we can no longer assume that |
2296 | signed chrecs do not overflow, as chrec_convert does, so avoid | |
2297 | calling it in that case. */ | |
2298 | if (flags & FOLD_CONVERSIONS) | |
2299 | return fold_convert (TREE_TYPE (chrec), op0); | |
2300 | ||
1e8552eb | 2301 | return chrec_convert (TREE_TYPE (chrec), op0, NULL_TREE); |
9baba81b SP |
2302 | |
2303 | case SCEV_NOT_KNOWN: | |
2304 | return chrec_dont_know; | |
2305 | ||
2306 | case SCEV_KNOWN: | |
2307 | return chrec_known; | |
2308 | ||
2309 | default: | |
2310 | break; | |
2311 | } | |
2312 | ||
2313 | switch (TREE_CODE_LENGTH (TREE_CODE (chrec))) | |
2314 | { | |
2315 | case 3: | |
2316 | op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0), | |
47ae9e4c | 2317 | flags, cache, size_expr); |
fca81712 SP |
2318 | if (op0 == chrec_dont_know) |
2319 | return chrec_dont_know; | |
2320 | ||
9baba81b | 2321 | op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1), |
47ae9e4c | 2322 | flags, cache, size_expr); |
fca81712 SP |
2323 | if (op1 == chrec_dont_know) |
2324 | return chrec_dont_know; | |
2325 | ||
9baba81b | 2326 | op2 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 2), |
47ae9e4c | 2327 | flags, cache, size_expr); |
fca81712 | 2328 | if (op2 == chrec_dont_know) |
9baba81b | 2329 | return chrec_dont_know; |
eac30183 ZD |
2330 | |
2331 | if (op0 == TREE_OPERAND (chrec, 0) | |
2332 | && op1 == TREE_OPERAND (chrec, 1) | |
2333 | && op2 == TREE_OPERAND (chrec, 2)) | |
2334 | return chrec; | |
2335 | ||
987b67bc KH |
2336 | return fold_build3 (TREE_CODE (chrec), |
2337 | TREE_TYPE (chrec), op0, op1, op2); | |
9baba81b SP |
2338 | |
2339 | case 2: | |
2340 | op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0), | |
47ae9e4c | 2341 | flags, cache, size_expr); |
fca81712 SP |
2342 | if (op0 == chrec_dont_know) |
2343 | return chrec_dont_know; | |
2344 | ||
9baba81b | 2345 | op1 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 1), |
47ae9e4c | 2346 | flags, cache, size_expr); |
fca81712 | 2347 | if (op1 == chrec_dont_know) |
9baba81b | 2348 | return chrec_dont_know; |
eac30183 ZD |
2349 | |
2350 | if (op0 == TREE_OPERAND (chrec, 0) | |
2351 | && op1 == TREE_OPERAND (chrec, 1)) | |
2352 | return chrec; | |
987b67bc | 2353 | return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1); |
9baba81b SP |
2354 | |
2355 | case 1: | |
2356 | op0 = instantiate_parameters_1 (loop, TREE_OPERAND (chrec, 0), | |
47ae9e4c | 2357 | flags, cache, size_expr); |
9baba81b SP |
2358 | if (op0 == chrec_dont_know) |
2359 | return chrec_dont_know; | |
eac30183 ZD |
2360 | if (op0 == TREE_OPERAND (chrec, 0)) |
2361 | return chrec; | |
987b67bc | 2362 | return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0); |
9baba81b SP |
2363 | |
2364 | case 0: | |
2365 | return chrec; | |
2366 | ||
2367 | default: | |
2368 | break; | |
2369 | } | |
2370 | ||
2371 | /* Too complicated to handle. */ | |
2372 | return chrec_dont_know; | |
2373 | } | |
e9eb809d ZD |
2374 | |
2375 | /* Analyze all the parameters of the chrec that were left under a | |
2376 | symbolic form. LOOP is the loop in which symbolic names have to | |
2377 | be analyzed and instantiated. */ | |
2378 | ||
2379 | tree | |
9baba81b | 2380 | instantiate_parameters (struct loop *loop, |
e9eb809d ZD |
2381 | tree chrec) |
2382 | { | |
9baba81b | 2383 | tree res; |
eb0bc7af | 2384 | htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info); |
9baba81b SP |
2385 | |
2386 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2387 | { | |
2388 | fprintf (dump_file, "(instantiate_parameters \n"); | |
2389 | fprintf (dump_file, " (loop_nb = %d)\n", loop->num); | |
2390 | fprintf (dump_file, " (chrec = "); | |
2391 | print_generic_expr (dump_file, chrec, 0); | |
2392 | fprintf (dump_file, ")\n"); | |
2393 | } | |
2394 | ||
47ae9e4c SP |
2395 | res = instantiate_parameters_1 (loop, chrec, INSERT_SUPERLOOP_CHRECS, cache, |
2396 | 0); | |
9baba81b SP |
2397 | |
2398 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2399 | { | |
2400 | fprintf (dump_file, " (res = "); | |
2401 | print_generic_expr (dump_file, res, 0); | |
2402 | fprintf (dump_file, "))\n"); | |
2403 | } | |
eb0bc7af ZD |
2404 | |
2405 | htab_delete (cache); | |
9baba81b SP |
2406 | |
2407 | return res; | |
2408 | } | |
2409 | ||
2410 | /* Similar to instantiate_parameters, but does not introduce the | |
2282a0e6 ZD |
2411 | evolutions in outer loops for LOOP invariants in CHREC, and does not |
2412 | care about causing overflows, as long as they do not affect value | |
2413 | of an expression. */ | |
9baba81b SP |
2414 | |
2415 | static tree | |
2416 | resolve_mixers (struct loop *loop, tree chrec) | |
2417 | { | |
eb0bc7af | 2418 | htab_t cache = htab_create (10, hash_scev_info, eq_scev_info, del_scev_info); |
47ae9e4c | 2419 | tree ret = instantiate_parameters_1 (loop, chrec, FOLD_CONVERSIONS, cache, 0); |
eb0bc7af ZD |
2420 | htab_delete (cache); |
2421 | return ret; | |
9baba81b SP |
2422 | } |
2423 | ||
2424 | /* Entry point for the analysis of the number of iterations pass. | |
2425 | This function tries to safely approximate the number of iterations | |
2426 | the loop will run. When this property is not decidable at compile | |
2427 | time, the result is chrec_dont_know. Otherwise the result is | |
2428 | a scalar or a symbolic parameter. | |
2429 | ||
2430 | Example of analysis: suppose that the loop has an exit condition: | |
2431 | ||
2432 | "if (b > 49) goto end_loop;" | |
2433 | ||
2434 | and that in a previous analysis we have determined that the | |
2435 | variable 'b' has an evolution function: | |
2436 | ||
2437 | "EF = {23, +, 5}_2". | |
2438 | ||
2439 | When we evaluate the function at the point 5, i.e. the value of the | |
2440 | variable 'b' after 5 iterations in the loop, we have EF (5) = 48, | |
2441 | and EF (6) = 53. In this case the value of 'b' on exit is '53' and | |
2442 | the loop body has been executed 6 times. */ | |
2443 | ||
2444 | tree | |
a14865db | 2445 | number_of_latch_executions (struct loop *loop) |
9baba81b SP |
2446 | { |
2447 | tree res, type; | |
2448 | edge exit; | |
2449 | struct tree_niter_desc niter_desc; | |
2450 | ||
2451 | /* Determine whether the number_of_iterations_in_loop has already | |
2452 | been computed. */ | |
2453 | res = loop->nb_iterations; | |
2454 | if (res) | |
2455 | return res; | |
2456 | res = chrec_dont_know; | |
2457 | ||
2458 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2459 | fprintf (dump_file, "(number_of_iterations_in_loop\n"); | |
2460 | ||
ac8f6c69 | 2461 | exit = single_exit (loop); |
82b85a85 | 2462 | if (!exit) |
9baba81b | 2463 | goto end; |
9baba81b | 2464 | |
f9cc1a70 | 2465 | if (!number_of_iterations_exit (loop, exit, &niter_desc, false)) |
9baba81b SP |
2466 | goto end; |
2467 | ||
2468 | type = TREE_TYPE (niter_desc.niter); | |
2469 | if (integer_nonzerop (niter_desc.may_be_zero)) | |
5212068f | 2470 | res = build_int_cst (type, 0); |
9baba81b SP |
2471 | else if (integer_zerop (niter_desc.may_be_zero)) |
2472 | res = niter_desc.niter; | |
2473 | else | |
2474 | res = chrec_dont_know; | |
2475 | ||
2476 | end: | |
2477 | return set_nb_iterations_in_loop (loop, res); | |
2478 | } | |
2479 | ||
a14865db ZD |
2480 | /* Returns the number of executions of the exit condition of LOOP, |
2481 | i.e., the number by one higher than number_of_latch_executions. | |
2482 | Note that unline number_of_latch_executions, this number does | |
2483 | not necessarily fit in the unsigned variant of the type of | |
2484 | the control variable -- if the number of iterations is a constant, | |
2485 | we return chrec_dont_know if adding one to number_of_latch_executions | |
2486 | overflows; however, in case the number of iterations is symbolic | |
2487 | expression, the caller is responsible for dealing with this | |
2488 | the possible overflow. */ | |
2489 | ||
2490 | tree | |
2491 | number_of_exit_cond_executions (struct loop *loop) | |
2492 | { | |
2493 | tree ret = number_of_latch_executions (loop); | |
2494 | tree type = chrec_type (ret); | |
2495 | ||
2496 | if (chrec_contains_undetermined (ret)) | |
2497 | return ret; | |
2498 | ||
2499 | ret = chrec_fold_plus (type, ret, build_int_cst (type, 1)); | |
2500 | if (TREE_CODE (ret) == INTEGER_CST | |
2501 | && TREE_OVERFLOW (ret)) | |
2502 | return chrec_dont_know; | |
2503 | ||
2504 | return ret; | |
2505 | } | |
2506 | ||
9baba81b SP |
2507 | /* One of the drivers for testing the scalar evolutions analysis. |
2508 | This function computes the number of iterations for all the loops | |
2509 | from the EXIT_CONDITIONS array. */ | |
2510 | ||
2511 | static void | |
5310bac6 | 2512 | number_of_iterations_for_all_loops (VEC(tree,heap) **exit_conditions) |
9baba81b SP |
2513 | { |
2514 | unsigned int i; | |
2515 | unsigned nb_chrec_dont_know_loops = 0; | |
2516 | unsigned nb_static_loops = 0; | |
5310bac6 | 2517 | tree cond; |
9baba81b | 2518 | |
5310bac6 | 2519 | for (i = 0; VEC_iterate (tree, *exit_conditions, i, cond); i++) |
9baba81b | 2520 | { |
a14865db | 2521 | tree res = number_of_latch_executions (loop_containing_stmt (cond)); |
9baba81b SP |
2522 | if (chrec_contains_undetermined (res)) |
2523 | nb_chrec_dont_know_loops++; | |
2524 | else | |
2525 | nb_static_loops++; | |
2526 | } | |
2527 | ||
2528 | if (dump_file) | |
2529 | { | |
2530 | fprintf (dump_file, "\n(\n"); | |
2531 | fprintf (dump_file, "-----------------------------------------\n"); | |
2532 | fprintf (dump_file, "%d\tnb_chrec_dont_know_loops\n", nb_chrec_dont_know_loops); | |
2533 | fprintf (dump_file, "%d\tnb_static_loops\n", nb_static_loops); | |
42fd6772 | 2534 | fprintf (dump_file, "%d\tnb_total_loops\n", number_of_loops ()); |
9baba81b SP |
2535 | fprintf (dump_file, "-----------------------------------------\n"); |
2536 | fprintf (dump_file, ")\n\n"); | |
2537 | ||
2538 | print_loop_ir (dump_file); | |
2539 | } | |
2540 | } | |
2541 | ||
2542 | \f | |
2543 | ||
2544 | /* Counters for the stats. */ | |
2545 | ||
2546 | struct chrec_stats | |
2547 | { | |
2548 | unsigned nb_chrecs; | |
2549 | unsigned nb_affine; | |
2550 | unsigned nb_affine_multivar; | |
2551 | unsigned nb_higher_poly; | |
2552 | unsigned nb_chrec_dont_know; | |
2553 | unsigned nb_undetermined; | |
2554 | }; | |
2555 | ||
2556 | /* Reset the counters. */ | |
2557 | ||
2558 | static inline void | |
2559 | reset_chrecs_counters (struct chrec_stats *stats) | |
2560 | { | |
2561 | stats->nb_chrecs = 0; | |
2562 | stats->nb_affine = 0; | |
2563 | stats->nb_affine_multivar = 0; | |
2564 | stats->nb_higher_poly = 0; | |
2565 | stats->nb_chrec_dont_know = 0; | |
2566 | stats->nb_undetermined = 0; | |
2567 | } | |
2568 | ||
2569 | /* Dump the contents of a CHREC_STATS structure. */ | |
2570 | ||
2571 | static void | |
2572 | dump_chrecs_stats (FILE *file, struct chrec_stats *stats) | |
2573 | { | |
2574 | fprintf (file, "\n(\n"); | |
2575 | fprintf (file, "-----------------------------------------\n"); | |
2576 | fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine); | |
2577 | fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar); | |
2578 | fprintf (file, "%d\tdegree greater than 2 polynomials\n", | |
2579 | stats->nb_higher_poly); | |
2580 | fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know); | |
2581 | fprintf (file, "-----------------------------------------\n"); | |
2582 | fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs); | |
2583 | fprintf (file, "%d\twith undetermined coefficients\n", | |
2584 | stats->nb_undetermined); | |
2585 | fprintf (file, "-----------------------------------------\n"); | |
2586 | fprintf (file, "%d\tchrecs in the scev database\n", | |
2587 | (int) htab_elements (scalar_evolution_info)); | |
2588 | fprintf (file, "%d\tsets in the scev database\n", nb_set_scev); | |
2589 | fprintf (file, "%d\tgets in the scev database\n", nb_get_scev); | |
2590 | fprintf (file, "-----------------------------------------\n"); | |
2591 | fprintf (file, ")\n\n"); | |
2592 | } | |
2593 | ||
2594 | /* Gather statistics about CHREC. */ | |
2595 | ||
2596 | static void | |
2597 | gather_chrec_stats (tree chrec, struct chrec_stats *stats) | |
2598 | { | |
2599 | if (dump_file && (dump_flags & TDF_STATS)) | |
2600 | { | |
2601 | fprintf (dump_file, "(classify_chrec "); | |
2602 | print_generic_expr (dump_file, chrec, 0); | |
2603 | fprintf (dump_file, "\n"); | |
2604 | } | |
2605 | ||
2606 | stats->nb_chrecs++; | |
2607 | ||
2608 | if (chrec == NULL_TREE) | |
2609 | { | |
2610 | stats->nb_undetermined++; | |
2611 | return; | |
2612 | } | |
2613 | ||
2614 | switch (TREE_CODE (chrec)) | |
2615 | { | |
2616 | case POLYNOMIAL_CHREC: | |
2617 | if (evolution_function_is_affine_p (chrec)) | |
2618 | { | |
2619 | if (dump_file && (dump_flags & TDF_STATS)) | |
2620 | fprintf (dump_file, " affine_univariate\n"); | |
2621 | stats->nb_affine++; | |
2622 | } | |
2623 | else if (evolution_function_is_affine_multivariate_p (chrec)) | |
2624 | { | |
2625 | if (dump_file && (dump_flags & TDF_STATS)) | |
2626 | fprintf (dump_file, " affine_multivariate\n"); | |
2627 | stats->nb_affine_multivar++; | |
2628 | } | |
2629 | else | |
2630 | { | |
2631 | if (dump_file && (dump_flags & TDF_STATS)) | |
2632 | fprintf (dump_file, " higher_degree_polynomial\n"); | |
2633 | stats->nb_higher_poly++; | |
2634 | } | |
2635 | ||
2636 | break; | |
2637 | ||
2638 | default: | |
2639 | break; | |
2640 | } | |
2641 | ||
2642 | if (chrec_contains_undetermined (chrec)) | |
2643 | { | |
2644 | if (dump_file && (dump_flags & TDF_STATS)) | |
2645 | fprintf (dump_file, " undetermined\n"); | |
2646 | stats->nb_undetermined++; | |
2647 | } | |
2648 | ||
2649 | if (dump_file && (dump_flags & TDF_STATS)) | |
2650 | fprintf (dump_file, ")\n"); | |
2651 | } | |
2652 | ||
2653 | /* One of the drivers for testing the scalar evolutions analysis. | |
2654 | This function analyzes the scalar evolution of all the scalars | |
2655 | defined as loop phi nodes in one of the loops from the | |
2656 | EXIT_CONDITIONS array. | |
2657 | ||
2658 | TODO Optimization: A loop is in canonical form if it contains only | |
2659 | a single scalar loop phi node. All the other scalars that have an | |
2660 | evolution in the loop are rewritten in function of this single | |
2661 | index. This allows the parallelization of the loop. */ | |
2662 | ||
2663 | static void | |
5310bac6 | 2664 | analyze_scalar_evolution_for_all_loop_phi_nodes (VEC(tree,heap) **exit_conditions) |
9baba81b SP |
2665 | { |
2666 | unsigned int i; | |
2667 | struct chrec_stats stats; | |
5310bac6 | 2668 | tree cond; |
9baba81b SP |
2669 | |
2670 | reset_chrecs_counters (&stats); | |
2671 | ||
5310bac6 | 2672 | for (i = 0; VEC_iterate (tree, *exit_conditions, i, cond); i++) |
9baba81b SP |
2673 | { |
2674 | struct loop *loop; | |
2675 | basic_block bb; | |
2676 | tree phi, chrec; | |
2677 | ||
5310bac6 | 2678 | loop = loop_containing_stmt (cond); |
9baba81b SP |
2679 | bb = loop->header; |
2680 | ||
bb29d951 | 2681 | for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi)) |
9baba81b SP |
2682 | if (is_gimple_reg (PHI_RESULT (phi))) |
2683 | { | |
2684 | chrec = instantiate_parameters | |
2685 | (loop, | |
2686 | analyze_scalar_evolution (loop, PHI_RESULT (phi))); | |
2687 | ||
2688 | if (dump_file && (dump_flags & TDF_STATS)) | |
2689 | gather_chrec_stats (chrec, &stats); | |
2690 | } | |
2691 | } | |
2692 | ||
2693 | if (dump_file && (dump_flags & TDF_STATS)) | |
2694 | dump_chrecs_stats (dump_file, &stats); | |
2695 | } | |
2696 | ||
2697 | /* Callback for htab_traverse, gathers information on chrecs in the | |
2698 | hashtable. */ | |
2699 | ||
2700 | static int | |
2701 | gather_stats_on_scev_database_1 (void **slot, void *stats) | |
2702 | { | |
cceb1885 | 2703 | struct scev_info_str *entry = (struct scev_info_str *) *slot; |
9baba81b | 2704 | |
cceb1885 | 2705 | gather_chrec_stats (entry->chrec, (struct chrec_stats *) stats); |
9baba81b SP |
2706 | |
2707 | return 1; | |
2708 | } | |
2709 | ||
2710 | /* Classify the chrecs of the whole database. */ | |
2711 | ||
2712 | void | |
2713 | gather_stats_on_scev_database (void) | |
2714 | { | |
2715 | struct chrec_stats stats; | |
2716 | ||
2717 | if (!dump_file) | |
2718 | return; | |
2719 | ||
2720 | reset_chrecs_counters (&stats); | |
2721 | ||
2722 | htab_traverse (scalar_evolution_info, gather_stats_on_scev_database_1, | |
2723 | &stats); | |
2724 | ||
2725 | dump_chrecs_stats (dump_file, &stats); | |
2726 | } | |
2727 | ||
2728 | \f | |
2729 | ||
2730 | /* Initializer. */ | |
2731 | ||
2732 | static void | |
2733 | initialize_scalar_evolutions_analyzer (void) | |
2734 | { | |
2735 | /* The elements below are unique. */ | |
2736 | if (chrec_dont_know == NULL_TREE) | |
2737 | { | |
2738 | chrec_not_analyzed_yet = NULL_TREE; | |
2739 | chrec_dont_know = make_node (SCEV_NOT_KNOWN); | |
2740 | chrec_known = make_node (SCEV_KNOWN); | |
d5ab5675 ZD |
2741 | TREE_TYPE (chrec_dont_know) = void_type_node; |
2742 | TREE_TYPE (chrec_known) = void_type_node; | |
9baba81b SP |
2743 | } |
2744 | } | |
2745 | ||
2746 | /* Initialize the analysis of scalar evolutions for LOOPS. */ | |
2747 | ||
2748 | void | |
d73be268 | 2749 | scev_initialize (void) |
9baba81b | 2750 | { |
42fd6772 ZD |
2751 | loop_iterator li; |
2752 | struct loop *loop; | |
9baba81b SP |
2753 | |
2754 | scalar_evolution_info = htab_create (100, hash_scev_info, | |
2755 | eq_scev_info, del_scev_info); | |
8bdbfff5 | 2756 | already_instantiated = BITMAP_ALLOC (NULL); |
9baba81b SP |
2757 | |
2758 | initialize_scalar_evolutions_analyzer (); | |
2759 | ||
42fd6772 ZD |
2760 | FOR_EACH_LOOP (li, loop, 0) |
2761 | { | |
2762 | loop->nb_iterations = NULL_TREE; | |
2763 | } | |
9baba81b SP |
2764 | } |
2765 | ||
2766 | /* Cleans up the information cached by the scalar evolutions analysis. */ | |
2767 | ||
2768 | void | |
2769 | scev_reset (void) | |
2770 | { | |
42fd6772 | 2771 | loop_iterator li; |
9baba81b SP |
2772 | struct loop *loop; |
2773 | ||
2774 | if (!scalar_evolution_info || !current_loops) | |
2775 | return; | |
2776 | ||
2777 | htab_empty (scalar_evolution_info); | |
42fd6772 | 2778 | FOR_EACH_LOOP (li, loop, 0) |
9baba81b | 2779 | { |
42fd6772 | 2780 | loop->nb_iterations = NULL_TREE; |
9baba81b | 2781 | } |
e9eb809d ZD |
2782 | } |
2783 | ||
2784 | /* Checks whether OP behaves as a simple affine iv of LOOP in STMT and returns | |
a6f778b2 ZD |
2785 | its base and step in IV if possible. If ALLOW_NONCONSTANT_STEP is true, we |
2786 | want step to be invariant in LOOP. Otherwise we require it to be an | |
2787 | integer constant. IV->no_overflow is set to true if we are sure the iv cannot | |
2788 | overflow (e.g. because it is computed in signed arithmetics). */ | |
e9eb809d ZD |
2789 | |
2790 | bool | |
a6f778b2 | 2791 | simple_iv (struct loop *loop, tree stmt, tree op, affine_iv *iv, |
9be872b7 | 2792 | bool allow_nonconstant_step) |
e9eb809d | 2793 | { |
9baba81b SP |
2794 | basic_block bb = bb_for_stmt (stmt); |
2795 | tree type, ev; | |
a6f778b2 | 2796 | bool folded_casts; |
9baba81b | 2797 | |
a6f778b2 ZD |
2798 | iv->base = NULL_TREE; |
2799 | iv->step = NULL_TREE; | |
2800 | iv->no_overflow = false; | |
9baba81b SP |
2801 | |
2802 | type = TREE_TYPE (op); | |
2803 | if (TREE_CODE (type) != INTEGER_TYPE | |
2804 | && TREE_CODE (type) != POINTER_TYPE) | |
2805 | return false; | |
2806 | ||
a6f778b2 ZD |
2807 | ev = analyze_scalar_evolution_in_loop (loop, bb->loop_father, op, |
2808 | &folded_casts); | |
9baba81b SP |
2809 | if (chrec_contains_undetermined (ev)) |
2810 | return false; | |
2811 | ||
2812 | if (tree_does_not_contain_chrecs (ev) | |
2813 | && !chrec_contains_symbols_defined_in_loop (ev, loop->num)) | |
2814 | { | |
a6f778b2 | 2815 | iv->base = ev; |
6e42ce54 | 2816 | iv->step = build_int_cst (TREE_TYPE (ev), 0); |
a6f778b2 | 2817 | iv->no_overflow = true; |
9baba81b SP |
2818 | return true; |
2819 | } | |
2820 | ||
2821 | if (TREE_CODE (ev) != POLYNOMIAL_CHREC | |
2822 | || CHREC_VARIABLE (ev) != (unsigned) loop->num) | |
2823 | return false; | |
2824 | ||
a6f778b2 | 2825 | iv->step = CHREC_RIGHT (ev); |
9be872b7 ZD |
2826 | if (allow_nonconstant_step) |
2827 | { | |
a6f778b2 ZD |
2828 | if (tree_contains_chrecs (iv->step, NULL) |
2829 | || chrec_contains_symbols_defined_in_loop (iv->step, loop->num)) | |
9be872b7 ZD |
2830 | return false; |
2831 | } | |
a6f778b2 | 2832 | else if (TREE_CODE (iv->step) != INTEGER_CST) |
9baba81b | 2833 | return false; |
9be872b7 | 2834 | |
a6f778b2 ZD |
2835 | iv->base = CHREC_LEFT (ev); |
2836 | if (tree_contains_chrecs (iv->base, NULL) | |
2837 | || chrec_contains_symbols_defined_in_loop (iv->base, loop->num)) | |
9baba81b SP |
2838 | return false; |
2839 | ||
a6f778b2 ZD |
2840 | iv->no_overflow = (!folded_casts |
2841 | && !flag_wrapv | |
2842 | && !TYPE_UNSIGNED (type)); | |
9baba81b SP |
2843 | return true; |
2844 | } | |
2845 | ||
2846 | /* Runs the analysis of scalar evolutions. */ | |
2847 | ||
2848 | void | |
2849 | scev_analysis (void) | |
2850 | { | |
5310bac6 | 2851 | VEC(tree,heap) *exit_conditions; |
9baba81b | 2852 | |
5310bac6 | 2853 | exit_conditions = VEC_alloc (tree, heap, 37); |
d73be268 | 2854 | select_loops_exit_conditions (&exit_conditions); |
9baba81b SP |
2855 | |
2856 | if (dump_file && (dump_flags & TDF_STATS)) | |
5310bac6 | 2857 | analyze_scalar_evolution_for_all_loop_phi_nodes (&exit_conditions); |
9baba81b | 2858 | |
5310bac6 KH |
2859 | number_of_iterations_for_all_loops (&exit_conditions); |
2860 | VEC_free (tree, heap, exit_conditions); | |
e9eb809d | 2861 | } |
9baba81b SP |
2862 | |
2863 | /* Finalize the scalar evolution analysis. */ | |
2864 | ||
2865 | void | |
2866 | scev_finalize (void) | |
2867 | { | |
2868 | htab_delete (scalar_evolution_info); | |
8bdbfff5 | 2869 | BITMAP_FREE (already_instantiated); |
9baba81b SP |
2870 | } |
2871 | ||
925196ed ZD |
2872 | /* Returns true if EXPR looks expensive. */ |
2873 | ||
2874 | static bool | |
2875 | expression_expensive_p (tree expr) | |
2876 | { | |
2877 | return force_expr_to_var_cost (expr) >= target_spill_cost; | |
2878 | } | |
2879 | ||
684aaf29 | 2880 | /* Replace ssa names for that scev can prove they are constant by the |
3ac01fde ZD |
2881 | appropriate constants. Also perform final value replacement in loops, |
2882 | in case the replacement expressions are cheap. | |
684aaf29 ZD |
2883 | |
2884 | We only consider SSA names defined by phi nodes; rest is left to the | |
2885 | ordinary constant propagation pass. */ | |
2886 | ||
c2924966 | 2887 | unsigned int |
684aaf29 ZD |
2888 | scev_const_prop (void) |
2889 | { | |
2890 | basic_block bb; | |
3ac01fde ZD |
2891 | tree name, phi, next_phi, type, ev; |
2892 | struct loop *loop, *ex_loop; | |
684aaf29 | 2893 | bitmap ssa_names_to_remove = NULL; |
3ac01fde | 2894 | unsigned i; |
42fd6772 | 2895 | loop_iterator li; |
684aaf29 ZD |
2896 | |
2897 | if (!current_loops) | |
c2924966 | 2898 | return 0; |
684aaf29 ZD |
2899 | |
2900 | FOR_EACH_BB (bb) | |
2901 | { | |
2902 | loop = bb->loop_father; | |
2903 | ||
2904 | for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi)) | |
2905 | { | |
2906 | name = PHI_RESULT (phi); | |
2907 | ||
2908 | if (!is_gimple_reg (name)) | |
2909 | continue; | |
2910 | ||
2911 | type = TREE_TYPE (name); | |
2912 | ||
2913 | if (!POINTER_TYPE_P (type) | |
2914 | && !INTEGRAL_TYPE_P (type)) | |
2915 | continue; | |
2916 | ||
2917 | ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name)); | |
2918 | if (!is_gimple_min_invariant (ev) | |
2919 | || !may_propagate_copy (name, ev)) | |
2920 | continue; | |
2921 | ||
2922 | /* Replace the uses of the name. */ | |
18aed06a SP |
2923 | if (name != ev) |
2924 | replace_uses_by (name, ev); | |
684aaf29 ZD |
2925 | |
2926 | if (!ssa_names_to_remove) | |
2927 | ssa_names_to_remove = BITMAP_ALLOC (NULL); | |
2928 | bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name)); | |
2929 | } | |
2930 | } | |
2931 | ||
9b3b55a1 DN |
2932 | /* Remove the ssa names that were replaced by constants. We do not |
2933 | remove them directly in the previous cycle, since this | |
2934 | invalidates scev cache. */ | |
684aaf29 ZD |
2935 | if (ssa_names_to_remove) |
2936 | { | |
2937 | bitmap_iterator bi; | |
684aaf29 ZD |
2938 | |
2939 | EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi) | |
2940 | { | |
2941 | name = ssa_name (i); | |
2942 | phi = SSA_NAME_DEF_STMT (name); | |
2943 | ||
2944 | gcc_assert (TREE_CODE (phi) == PHI_NODE); | |
9b3b55a1 | 2945 | remove_phi_node (phi, NULL, true); |
684aaf29 ZD |
2946 | } |
2947 | ||
2948 | BITMAP_FREE (ssa_names_to_remove); | |
2949 | scev_reset (); | |
2950 | } | |
3ac01fde ZD |
2951 | |
2952 | /* Now the regular final value replacement. */ | |
42fd6772 | 2953 | FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) |
3ac01fde ZD |
2954 | { |
2955 | edge exit; | |
a6f778b2 | 2956 | tree def, rslt, ass, niter; |
925196ed | 2957 | block_stmt_iterator bsi; |
3ac01fde | 2958 | |
3ac01fde ZD |
2959 | /* If we do not know exact number of iterations of the loop, we cannot |
2960 | replace the final value. */ | |
ac8f6c69 | 2961 | exit = single_exit (loop); |
a6f778b2 ZD |
2962 | if (!exit) |
2963 | continue; | |
2964 | ||
a14865db | 2965 | niter = number_of_latch_executions (loop); |
a6f778b2 ZD |
2966 | if (niter == chrec_dont_know |
2967 | /* If computing the number of iterations is expensive, it may be | |
2968 | better not to introduce computations involving it. */ | |
2969 | || expression_expensive_p (niter)) | |
3ac01fde | 2970 | continue; |
925196ed ZD |
2971 | |
2972 | /* Ensure that it is possible to insert new statements somewhere. */ | |
2973 | if (!single_pred_p (exit->dest)) | |
2974 | split_loop_exit_edge (exit); | |
2975 | tree_block_label (exit->dest); | |
2976 | bsi = bsi_after_labels (exit->dest); | |
2977 | ||
2978 | ex_loop = superloop_at_depth (loop, exit->dest->loop_father->depth + 1); | |
3ac01fde ZD |
2979 | |
2980 | for (phi = phi_nodes (exit->dest); phi; phi = next_phi) | |
2981 | { | |
2982 | next_phi = PHI_CHAIN (phi); | |
925196ed | 2983 | rslt = PHI_RESULT (phi); |
3ac01fde | 2984 | def = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
925196ed | 2985 | if (!is_gimple_reg (def)) |
3ac01fde ZD |
2986 | continue; |
2987 | ||
2988 | if (!POINTER_TYPE_P (TREE_TYPE (def)) | |
2989 | && !INTEGRAL_TYPE_P (TREE_TYPE (def))) | |
2990 | continue; | |
2991 | ||
a6f778b2 | 2992 | def = analyze_scalar_evolution_in_loop (ex_loop, loop, def, NULL); |
925196ed | 2993 | def = compute_overall_effect_of_inner_loop (ex_loop, def); |
3ac01fde | 2994 | if (!tree_does_not_contain_chrecs (def) |
e5db3515 ZD |
2995 | || chrec_contains_symbols_defined_in_loop (def, ex_loop->num) |
2996 | /* Moving the computation from the loop may prolong life range | |
2997 | of some ssa names, which may cause problems if they appear | |
2998 | on abnormal edges. */ | |
2999 | || contains_abnormal_ssa_name_p (def)) | |
3ac01fde ZD |
3000 | continue; |
3001 | ||
9b3b55a1 | 3002 | /* Eliminate the PHI node and replace it by a computation outside |
925196ed ZD |
3003 | the loop. */ |
3004 | def = unshare_expr (def); | |
9b3b55a1 | 3005 | remove_phi_node (phi, NULL_TREE, false); |
925196ed | 3006 | |
07beea0d | 3007 | ass = build2 (GIMPLE_MODIFY_STMT, void_type_node, rslt, NULL_TREE); |
925196ed | 3008 | SSA_NAME_DEF_STMT (rslt) = ass; |
35771d34 PB |
3009 | { |
3010 | block_stmt_iterator dest = bsi; | |
3011 | bsi_insert_before (&dest, ass, BSI_NEW_STMT); | |
3012 | def = force_gimple_operand_bsi (&dest, def, false, NULL_TREE); | |
3013 | } | |
07beea0d | 3014 | GIMPLE_STMT_OPERAND (ass, 1) = def; |
925196ed | 3015 | update_stmt (ass); |
3ac01fde ZD |
3016 | } |
3017 | } | |
c2924966 | 3018 | return 0; |
684aaf29 | 3019 | } |