]>
Commit | Line | Data |
---|---|---|
e9eb809d | 1 | /* Scalar evolution detector. |
d1e082c2 | 2 | Copyright (C) 2003-2013 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 | |
9dcd6f09 | 9 | Software Foundation; either version 3, or (at your option) any later |
e9eb809d ZD |
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 | |
9dcd6f09 NC |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ | |
e9eb809d | 20 | |
b8698a0f L |
21 | /* |
22 | Description: | |
23 | ||
9baba81b SP |
24 | This pass analyzes the evolution of scalar variables in loop |
25 | structures. The algorithm is based on the SSA representation, | |
26 | and on the loop hierarchy tree. This algorithm is not based on | |
27 | the notion of versions of a variable, as it was the case for the | |
28 | previous implementations of the scalar evolution algorithm, but | |
29 | it assumes that each defined name is unique. | |
30 | ||
31 | The notation used in this file is called "chains of recurrences", | |
32 | and has been proposed by Eugene Zima, Robert Van Engelen, and | |
33 | others for describing induction variables in programs. For example | |
34 | "b -> {0, +, 2}_1" means that the scalar variable "b" is equal to 0 | |
35 | when entering in the loop_1 and has a step 2 in this loop, in other | |
36 | words "for (b = 0; b < N; b+=2);". Note that the coefficients of | |
37 | this chain of recurrence (or chrec [shrek]) can contain the name of | |
38 | other variables, in which case they are called parametric chrecs. | |
39 | For example, "b -> {a, +, 2}_1" means that the initial value of "b" | |
40 | is the value of "a". In most of the cases these parametric chrecs | |
41 | are fully instantiated before their use because symbolic names can | |
42 | hide some difficult cases such as self-references described later | |
43 | (see the Fibonacci example). | |
b8698a0f | 44 | |
9baba81b | 45 | A short sketch of the algorithm is: |
b8698a0f | 46 | |
9baba81b SP |
47 | Given a scalar variable to be analyzed, follow the SSA edge to |
48 | its definition: | |
b8698a0f | 49 | |
726a989a | 50 | - When the definition is a GIMPLE_ASSIGN: if the right hand side |
9baba81b | 51 | (RHS) of the definition cannot be statically analyzed, the answer |
b8698a0f | 52 | of the analyzer is: "don't know". |
9baba81b SP |
53 | Otherwise, for all the variables that are not yet analyzed in the |
54 | RHS, try to determine their evolution, and finally try to | |
55 | evaluate the operation of the RHS that gives the evolution | |
56 | function of the analyzed variable. | |
57 | ||
58 | - When the definition is a condition-phi-node: determine the | |
59 | evolution function for all the branches of the phi node, and | |
60 | finally merge these evolutions (see chrec_merge). | |
61 | ||
62 | - When the definition is a loop-phi-node: determine its initial | |
63 | condition, that is the SSA edge defined in an outer loop, and | |
64 | keep it symbolic. Then determine the SSA edges that are defined | |
65 | in the body of the loop. Follow the inner edges until ending on | |
66 | another loop-phi-node of the same analyzed loop. If the reached | |
67 | loop-phi-node is not the starting loop-phi-node, then we keep | |
68 | this definition under a symbolic form. If the reached | |
69 | loop-phi-node is the same as the starting one, then we compute a | |
70 | symbolic stride on the return path. The result is then the | |
71 | symbolic chrec {initial_condition, +, symbolic_stride}_loop. | |
72 | ||
73 | Examples: | |
b8698a0f | 74 | |
9baba81b | 75 | Example 1: Illustration of the basic algorithm. |
b8698a0f | 76 | |
9baba81b SP |
77 | | a = 3 |
78 | | loop_1 | |
79 | | b = phi (a, c) | |
80 | | c = b + 1 | |
81 | | if (c > 10) exit_loop | |
82 | | endloop | |
b8698a0f | 83 | |
9baba81b SP |
84 | Suppose that we want to know the number of iterations of the |
85 | loop_1. The exit_loop is controlled by a COND_EXPR (c > 10). We | |
86 | ask the scalar evolution analyzer two questions: what's the | |
87 | scalar evolution (scev) of "c", and what's the scev of "10". For | |
88 | "10" the answer is "10" since it is a scalar constant. For the | |
89 | scalar variable "c", it follows the SSA edge to its definition, | |
90 | "c = b + 1", and then asks again what's the scev of "b". | |
91 | Following the SSA edge, we end on a loop-phi-node "b = phi (a, | |
92 | c)", where the initial condition is "a", and the inner loop edge | |
93 | is "c". The initial condition is kept under a symbolic form (it | |
94 | may be the case that the copy constant propagation has done its | |
95 | work and we end with the constant "3" as one of the edges of the | |
96 | loop-phi-node). The update edge is followed to the end of the | |
97 | loop, and until reaching again the starting loop-phi-node: b -> c | |
98 | -> b. At this point we have drawn a path from "b" to "b" from | |
99 | which we compute the stride in the loop: in this example it is | |
100 | "+1". The resulting scev for "b" is "b -> {a, +, 1}_1". Now | |
101 | that the scev for "b" is known, it is possible to compute the | |
102 | scev for "c", that is "c -> {a + 1, +, 1}_1". In order to | |
103 | determine the number of iterations in the loop_1, we have to | |
3f227a8c | 104 | instantiate_parameters (loop_1, {a + 1, +, 1}_1), that gives after some |
9baba81b SP |
105 | more analysis the scev {4, +, 1}_1, or in other words, this is |
106 | the function "f (x) = x + 4", where x is the iteration count of | |
107 | the loop_1. Now we have to solve the inequality "x + 4 > 10", | |
108 | and take the smallest iteration number for which the loop is | |
109 | exited: x = 7. This loop runs from x = 0 to x = 7, and in total | |
110 | there are 8 iterations. In terms of loop normalization, we have | |
111 | created a variable that is implicitly defined, "x" or just "_1", | |
112 | and all the other analyzed scalars of the loop are defined in | |
113 | function of this variable: | |
b8698a0f | 114 | |
9baba81b SP |
115 | a -> 3 |
116 | b -> {3, +, 1}_1 | |
117 | c -> {4, +, 1}_1 | |
b8698a0f L |
118 | |
119 | or in terms of a C program: | |
120 | ||
9baba81b SP |
121 | | a = 3 |
122 | | for (x = 0; x <= 7; x++) | |
123 | | { | |
124 | | b = x + 3 | |
125 | | c = x + 4 | |
126 | | } | |
b8698a0f | 127 | |
3f227a8c | 128 | Example 2a: Illustration of the algorithm on nested loops. |
b8698a0f | 129 | |
9baba81b SP |
130 | | loop_1 |
131 | | a = phi (1, b) | |
132 | | c = a + 2 | |
133 | | loop_2 10 times | |
134 | | b = phi (c, d) | |
135 | | d = b + 3 | |
136 | | endloop | |
137 | | endloop | |
b8698a0f | 138 | |
9baba81b SP |
139 | For analyzing the scalar evolution of "a", the algorithm follows |
140 | the SSA edge into the loop's body: "a -> b". "b" is an inner | |
b8698a0f L |
141 | loop-phi-node, and its analysis as in Example 1, gives: |
142 | ||
9baba81b SP |
143 | b -> {c, +, 3}_2 |
144 | d -> {c + 3, +, 3}_2 | |
b8698a0f | 145 | |
9baba81b SP |
146 | Following the SSA edge for the initial condition, we end on "c = a |
147 | + 2", and then on the starting loop-phi-node "a". From this point, | |
148 | the loop stride is computed: back on "c = a + 2" we get a "+2" in | |
149 | the loop_1, then on the loop-phi-node "b" we compute the overall | |
150 | effect of the inner loop that is "b = c + 30", and we get a "+30" | |
151 | in the loop_1. That means that the overall stride in loop_1 is | |
b8698a0f L |
152 | equal to "+32", and the result is: |
153 | ||
9baba81b SP |
154 | a -> {1, +, 32}_1 |
155 | c -> {3, +, 32}_1 | |
3f227a8c JS |
156 | |
157 | Example 2b: Multivariate chains of recurrences. | |
158 | ||
159 | | loop_1 | |
160 | | k = phi (0, k + 1) | |
161 | | loop_2 4 times | |
162 | | j = phi (0, j + 1) | |
163 | | loop_3 4 times | |
164 | | i = phi (0, i + 1) | |
165 | | A[j + k] = ... | |
166 | | endloop | |
167 | | endloop | |
168 | | endloop | |
169 | ||
170 | Analyzing the access function of array A with | |
171 | instantiate_parameters (loop_1, "j + k"), we obtain the | |
172 | instantiation and the analysis of the scalar variables "j" and "k" | |
173 | in loop_1. This leads to the scalar evolution {4, +, 1}_1: the end | |
174 | value of loop_2 for "j" is 4, and the evolution of "k" in loop_1 is | |
175 | {0, +, 1}_1. To obtain the evolution function in loop_3 and | |
176 | instantiate the scalar variables up to loop_1, one has to use: | |
a213b219 SP |
177 | instantiate_scev (block_before_loop (loop_1), loop_3, "j + k"). |
178 | The result of this call is {{0, +, 1}_1, +, 1}_2. | |
3f227a8c | 179 | |
9baba81b | 180 | Example 3: Higher degree polynomials. |
b8698a0f | 181 | |
9baba81b SP |
182 | | loop_1 |
183 | | a = phi (2, b) | |
184 | | c = phi (5, d) | |
185 | | b = a + 1 | |
186 | | d = c + a | |
187 | | endloop | |
b8698a0f | 188 | |
9baba81b SP |
189 | a -> {2, +, 1}_1 |
190 | b -> {3, +, 1}_1 | |
191 | c -> {5, +, a}_1 | |
192 | d -> {5 + a, +, a}_1 | |
b8698a0f | 193 | |
3f227a8c JS |
194 | instantiate_parameters (loop_1, {5, +, a}_1) -> {5, +, 2, +, 1}_1 |
195 | instantiate_parameters (loop_1, {5 + a, +, a}_1) -> {7, +, 3, +, 1}_1 | |
b8698a0f | 196 | |
9baba81b | 197 | Example 4: Lucas, Fibonacci, or mixers in general. |
b8698a0f | 198 | |
9baba81b SP |
199 | | loop_1 |
200 | | a = phi (1, b) | |
201 | | c = phi (3, d) | |
202 | | b = c | |
203 | | d = c + a | |
204 | | endloop | |
b8698a0f | 205 | |
9baba81b SP |
206 | a -> (1, c)_1 |
207 | c -> {3, +, a}_1 | |
b8698a0f | 208 | |
9baba81b SP |
209 | The syntax "(1, c)_1" stands for a PEELED_CHREC that has the |
210 | following semantics: during the first iteration of the loop_1, the | |
211 | variable contains the value 1, and then it contains the value "c". | |
212 | Note that this syntax is close to the syntax of the loop-phi-node: | |
213 | "a -> (1, c)_1" vs. "a = phi (1, c)". | |
b8698a0f | 214 | |
9baba81b SP |
215 | The symbolic chrec representation contains all the semantics of the |
216 | original code. What is more difficult is to use this information. | |
b8698a0f | 217 | |
9baba81b | 218 | Example 5: Flip-flops, or exchangers. |
b8698a0f | 219 | |
9baba81b SP |
220 | | loop_1 |
221 | | a = phi (1, b) | |
222 | | c = phi (3, d) | |
223 | | b = c | |
224 | | d = a | |
225 | | endloop | |
b8698a0f | 226 | |
9baba81b SP |
227 | a -> (1, c)_1 |
228 | c -> (3, a)_1 | |
b8698a0f | 229 | |
9baba81b | 230 | Based on these symbolic chrecs, it is possible to refine this |
b8698a0f L |
231 | information into the more precise PERIODIC_CHRECs: |
232 | ||
9baba81b SP |
233 | a -> |1, 3|_1 |
234 | c -> |3, 1|_1 | |
b8698a0f | 235 | |
9baba81b | 236 | This transformation is not yet implemented. |
b8698a0f | 237 | |
9baba81b | 238 | Further readings: |
b8698a0f | 239 | |
9baba81b SP |
240 | You can find a more detailed description of the algorithm in: |
241 | http://icps.u-strasbg.fr/~pop/DEA_03_Pop.pdf | |
242 | http://icps.u-strasbg.fr/~pop/DEA_03_Pop.ps.gz. But note that | |
243 | this is a preliminary report and some of the details of the | |
244 | algorithm have changed. I'm working on a research report that | |
245 | updates the description of the algorithms to reflect the design | |
246 | choices used in this implementation. | |
b8698a0f | 247 | |
9baba81b SP |
248 | A set of slides show a high level overview of the algorithm and run |
249 | an example through the scalar evolution analyzer: | |
250 | http://cri.ensmp.fr/~pop/gcc/mar04/slides.pdf | |
251 | ||
252 | The slides that I have presented at the GCC Summit'04 are available | |
253 | at: http://cri.ensmp.fr/~pop/gcc/20040604/gccsummit-lno-spop.pdf | |
254 | */ | |
255 | ||
e9eb809d ZD |
256 | #include "config.h" |
257 | #include "system.h" | |
258 | #include "coretypes.h" | |
4a8fb1a1 | 259 | #include "hash-table.h" |
cf835838 | 260 | #include "gimple-pretty-print.h" |
e9eb809d | 261 | #include "tree-flow.h" |
e9eb809d ZD |
262 | #include "cfgloop.h" |
263 | #include "tree-chrec.h" | |
264 | #include "tree-scalar-evolution.h" | |
7ee2468b | 265 | #include "dumpfile.h" |
c59dabbe | 266 | #include "params.h" |
9baba81b SP |
267 | |
268 | static tree analyze_scalar_evolution_1 (struct loop *, tree, tree); | |
bef28ced JL |
269 | static tree analyze_scalar_evolution_for_address_of (struct loop *loop, |
270 | tree var); | |
9baba81b | 271 | |
a213b219 SP |
272 | /* The cached information about an SSA name VAR, claiming that below |
273 | basic block INSTANTIATED_BELOW, the value of VAR can be expressed | |
274 | as CHREC. */ | |
9baba81b | 275 | |
d1b38208 | 276 | struct GTY(()) scev_info_str { |
a213b219 | 277 | basic_block instantiated_below; |
9baba81b SP |
278 | tree var; |
279 | tree chrec; | |
280 | }; | |
281 | ||
282 | /* Counters for the scev database. */ | |
283 | static unsigned nb_set_scev = 0; | |
284 | static unsigned nb_get_scev = 0; | |
285 | ||
286 | /* The following trees are unique elements. Thus the comparison of | |
287 | another element to these elements should be done on the pointer to | |
288 | these trees, and not on their value. */ | |
289 | ||
290 | /* The SSA_NAMEs that are not yet analyzed are qualified with NULL_TREE. */ | |
291 | tree chrec_not_analyzed_yet; | |
292 | ||
293 | /* Reserved to the cases where the analyzer has detected an | |
294 | undecidable property at compile time. */ | |
295 | tree chrec_dont_know; | |
296 | ||
297 | /* When the analyzer has detected that a property will never | |
298 | happen, then it qualifies it with chrec_known. */ | |
299 | tree chrec_known; | |
300 | ||
9e2f83a5 | 301 | static GTY ((param_is (struct scev_info_str))) htab_t scalar_evolution_info; |
9baba81b SP |
302 | |
303 | \f | |
a213b219 | 304 | /* Constructs a new SCEV_INFO_STR structure for VAR and INSTANTIATED_BELOW. */ |
9baba81b SP |
305 | |
306 | static inline struct scev_info_str * | |
a213b219 | 307 | new_scev_info_str (basic_block instantiated_below, tree var) |
9baba81b SP |
308 | { |
309 | struct scev_info_str *res; | |
b8698a0f | 310 | |
a9429e29 | 311 | res = ggc_alloc_scev_info_str (); |
9baba81b SP |
312 | res->var = var; |
313 | res->chrec = chrec_not_analyzed_yet; | |
a213b219 SP |
314 | res->instantiated_below = instantiated_below; |
315 | ||
9baba81b SP |
316 | return res; |
317 | } | |
318 | ||
319 | /* Computes a hash function for database element ELT. */ | |
320 | ||
4a8fb1a1 | 321 | static inline hashval_t |
9baba81b SP |
322 | hash_scev_info (const void *elt) |
323 | { | |
741ac903 | 324 | return SSA_NAME_VERSION (((const struct scev_info_str *) elt)->var); |
9baba81b SP |
325 | } |
326 | ||
327 | /* Compares database elements E1 and E2. */ | |
328 | ||
4a8fb1a1 | 329 | static inline int |
9baba81b SP |
330 | eq_scev_info (const void *e1, const void *e2) |
331 | { | |
cceb1885 GDR |
332 | const struct scev_info_str *elt1 = (const struct scev_info_str *) e1; |
333 | const struct scev_info_str *elt2 = (const struct scev_info_str *) e2; | |
9baba81b | 334 | |
a213b219 SP |
335 | return (elt1->var == elt2->var |
336 | && elt1->instantiated_below == elt2->instantiated_below); | |
9baba81b SP |
337 | } |
338 | ||
339 | /* Deletes database element E. */ | |
340 | ||
341 | static void | |
342 | del_scev_info (void *e) | |
343 | { | |
9e2f83a5 | 344 | ggc_free (e); |
9baba81b SP |
345 | } |
346 | ||
4a8fb1a1 | 347 | |
a213b219 SP |
348 | /* Get the scalar evolution of VAR for INSTANTIATED_BELOW basic block. |
349 | A first query on VAR returns chrec_not_analyzed_yet. */ | |
9baba81b SP |
350 | |
351 | static tree * | |
a213b219 | 352 | find_var_scev_info (basic_block instantiated_below, tree var) |
9baba81b SP |
353 | { |
354 | struct scev_info_str *res; | |
355 | struct scev_info_str tmp; | |
356 | PTR *slot; | |
357 | ||
358 | tmp.var = var; | |
a213b219 | 359 | tmp.instantiated_below = instantiated_below; |
9baba81b SP |
360 | slot = htab_find_slot (scalar_evolution_info, &tmp, INSERT); |
361 | ||
362 | if (!*slot) | |
a213b219 | 363 | *slot = new_scev_info_str (instantiated_below, var); |
cceb1885 | 364 | res = (struct scev_info_str *) *slot; |
9baba81b SP |
365 | |
366 | return &res->chrec; | |
367 | } | |
368 | ||
9baba81b SP |
369 | /* Return true when CHREC contains symbolic names defined in |
370 | LOOP_NB. */ | |
371 | ||
b8698a0f | 372 | bool |
ed7a4b4b | 373 | chrec_contains_symbols_defined_in_loop (const_tree chrec, unsigned loop_nb) |
9baba81b | 374 | { |
5039610b SL |
375 | int i, n; |
376 | ||
9baba81b SP |
377 | if (chrec == NULL_TREE) |
378 | return false; | |
379 | ||
ad6003f2 | 380 | if (is_gimple_min_invariant (chrec)) |
9baba81b SP |
381 | return false; |
382 | ||
9baba81b SP |
383 | if (TREE_CODE (chrec) == SSA_NAME) |
384 | { | |
492e5456 SP |
385 | gimple def; |
386 | loop_p def_loop, loop; | |
387 | ||
388 | if (SSA_NAME_IS_DEFAULT_DEF (chrec)) | |
389 | return false; | |
390 | ||
391 | def = SSA_NAME_DEF_STMT (chrec); | |
392 | def_loop = loop_containing_stmt (def); | |
0fc822d0 | 393 | loop = get_loop (cfun, loop_nb); |
9baba81b SP |
394 | |
395 | if (def_loop == NULL) | |
396 | return false; | |
397 | ||
398 | if (loop == def_loop || flow_loop_nested_p (loop, def_loop)) | |
399 | return true; | |
400 | ||
401 | return false; | |
402 | } | |
403 | ||
5039610b SL |
404 | n = TREE_OPERAND_LENGTH (chrec); |
405 | for (i = 0; i < n; i++) | |
b8698a0f | 406 | if (chrec_contains_symbols_defined_in_loop (TREE_OPERAND (chrec, i), |
5039610b SL |
407 | loop_nb)) |
408 | return true; | |
409 | return false; | |
9baba81b SP |
410 | } |
411 | ||
412 | /* Return true when PHI is a loop-phi-node. */ | |
413 | ||
414 | static bool | |
726a989a | 415 | loop_phi_node_p (gimple phi) |
9baba81b SP |
416 | { |
417 | /* The implementation of this function is based on the following | |
418 | property: "all the loop-phi-nodes of a loop are contained in the | |
419 | loop's header basic block". */ | |
420 | ||
726a989a | 421 | return loop_containing_stmt (phi)->header == gimple_bb (phi); |
9baba81b SP |
422 | } |
423 | ||
424 | /* Compute the scalar evolution for EVOLUTION_FN after crossing LOOP. | |
425 | In general, in the case of multivariate evolutions we want to get | |
426 | the evolution in different loops. LOOP specifies the level for | |
427 | which to get the evolution. | |
b8698a0f | 428 | |
9baba81b | 429 | Example: |
b8698a0f | 430 | |
9baba81b SP |
431 | | for (j = 0; j < 100; j++) |
432 | | { | |
433 | | for (k = 0; k < 100; k++) | |
434 | | { | |
b8698a0f | 435 | | i = k + j; - Here the value of i is a function of j, k. |
9baba81b | 436 | | } |
b8698a0f | 437 | | ... = i - Here the value of i is a function of j. |
9baba81b | 438 | | } |
b8698a0f L |
439 | | ... = i - Here the value of i is a scalar. |
440 | ||
441 | Example: | |
442 | ||
9baba81b SP |
443 | | i_0 = ... |
444 | | loop_1 10 times | |
445 | | i_1 = phi (i_0, i_2) | |
446 | | i_2 = i_1 + 2 | |
447 | | endloop | |
b8698a0f | 448 | |
9baba81b SP |
449 | This loop has the same effect as: |
450 | LOOP_1 has the same effect as: | |
b8698a0f | 451 | |
9baba81b | 452 | | i_1 = i_0 + 20 |
b8698a0f L |
453 | |
454 | The overall effect of the loop, "i_0 + 20" in the previous example, | |
455 | is obtained by passing in the parameters: LOOP = 1, | |
9baba81b SP |
456 | EVOLUTION_FN = {i_0, +, 2}_1. |
457 | */ | |
b8698a0f | 458 | |
42e6eec5 | 459 | tree |
9baba81b SP |
460 | compute_overall_effect_of_inner_loop (struct loop *loop, tree evolution_fn) |
461 | { | |
462 | bool val = false; | |
463 | ||
464 | if (evolution_fn == chrec_dont_know) | |
465 | return chrec_dont_know; | |
466 | ||
467 | else if (TREE_CODE (evolution_fn) == POLYNOMIAL_CHREC) | |
468 | { | |
677cc14d ZD |
469 | struct loop *inner_loop = get_chrec_loop (evolution_fn); |
470 | ||
471 | if (inner_loop == loop | |
472 | || flow_loop_nested_p (loop, inner_loop)) | |
9baba81b | 473 | { |
a14865db | 474 | tree nb_iter = number_of_latch_executions (inner_loop); |
9baba81b SP |
475 | |
476 | if (nb_iter == chrec_dont_know) | |
477 | return chrec_dont_know; | |
478 | else | |
479 | { | |
480 | tree res; | |
481 | ||
9baba81b SP |
482 | /* evolution_fn is the evolution function in LOOP. Get |
483 | its value in the nb_iter-th iteration. */ | |
484 | res = chrec_apply (inner_loop->num, evolution_fn, nb_iter); | |
42e6eec5 SP |
485 | |
486 | if (chrec_contains_symbols_defined_in_loop (res, loop->num)) | |
487 | res = instantiate_parameters (loop, res); | |
488 | ||
8c27b7d4 | 489 | /* Continue the computation until ending on a parent of LOOP. */ |
9baba81b SP |
490 | return compute_overall_effect_of_inner_loop (loop, res); |
491 | } | |
492 | } | |
493 | else | |
494 | return evolution_fn; | |
495 | } | |
b8698a0f | 496 | |
9baba81b SP |
497 | /* If the evolution function is an invariant, there is nothing to do. */ |
498 | else if (no_evolution_in_loop_p (evolution_fn, loop->num, &val) && val) | |
499 | return evolution_fn; | |
b8698a0f | 500 | |
9baba81b SP |
501 | else |
502 | return chrec_dont_know; | |
503 | } | |
504 | ||
9baba81b SP |
505 | /* Associate CHREC to SCALAR. */ |
506 | ||
507 | static void | |
a213b219 | 508 | set_scalar_evolution (basic_block instantiated_below, tree scalar, tree chrec) |
9baba81b SP |
509 | { |
510 | tree *scalar_info; | |
b8698a0f | 511 | |
9baba81b SP |
512 | if (TREE_CODE (scalar) != SSA_NAME) |
513 | return; | |
514 | ||
a213b219 | 515 | scalar_info = find_var_scev_info (instantiated_below, scalar); |
b8698a0f | 516 | |
9baba81b SP |
517 | if (dump_file) |
518 | { | |
dfedbe40 | 519 | if (dump_flags & TDF_SCEV) |
9baba81b SP |
520 | { |
521 | fprintf (dump_file, "(set_scalar_evolution \n"); | |
a213b219 SP |
522 | fprintf (dump_file, " instantiated_below = %d \n", |
523 | instantiated_below->index); | |
9baba81b SP |
524 | fprintf (dump_file, " (scalar = "); |
525 | print_generic_expr (dump_file, scalar, 0); | |
526 | fprintf (dump_file, ")\n (scalar_evolution = "); | |
527 | print_generic_expr (dump_file, chrec, 0); | |
528 | fprintf (dump_file, "))\n"); | |
529 | } | |
530 | if (dump_flags & TDF_STATS) | |
531 | nb_set_scev++; | |
532 | } | |
b8698a0f | 533 | |
9baba81b SP |
534 | *scalar_info = chrec; |
535 | } | |
536 | ||
a213b219 SP |
537 | /* Retrieve the chrec associated to SCALAR instantiated below |
538 | INSTANTIATED_BELOW block. */ | |
9baba81b SP |
539 | |
540 | static tree | |
a213b219 | 541 | get_scalar_evolution (basic_block instantiated_below, tree scalar) |
9baba81b SP |
542 | { |
543 | tree res; | |
b8698a0f | 544 | |
9baba81b SP |
545 | if (dump_file) |
546 | { | |
dfedbe40 | 547 | if (dump_flags & TDF_SCEV) |
9baba81b SP |
548 | { |
549 | fprintf (dump_file, "(get_scalar_evolution \n"); | |
550 | fprintf (dump_file, " (scalar = "); | |
551 | print_generic_expr (dump_file, scalar, 0); | |
552 | fprintf (dump_file, ")\n"); | |
553 | } | |
554 | if (dump_flags & TDF_STATS) | |
555 | nb_get_scev++; | |
556 | } | |
b8698a0f | 557 | |
9baba81b SP |
558 | switch (TREE_CODE (scalar)) |
559 | { | |
560 | case SSA_NAME: | |
a213b219 | 561 | res = *find_var_scev_info (instantiated_below, scalar); |
9baba81b SP |
562 | break; |
563 | ||
564 | case REAL_CST: | |
325217ed | 565 | case FIXED_CST: |
9baba81b SP |
566 | case INTEGER_CST: |
567 | res = scalar; | |
568 | break; | |
569 | ||
570 | default: | |
571 | res = chrec_not_analyzed_yet; | |
572 | break; | |
573 | } | |
b8698a0f | 574 | |
dfedbe40 | 575 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
576 | { |
577 | fprintf (dump_file, " (scalar_evolution = "); | |
578 | print_generic_expr (dump_file, res, 0); | |
579 | fprintf (dump_file, "))\n"); | |
580 | } | |
b8698a0f | 581 | |
9baba81b SP |
582 | return res; |
583 | } | |
584 | ||
585 | /* Helper function for add_to_evolution. Returns the evolution | |
586 | function for an assignment of the form "a = b + c", where "a" and | |
587 | "b" are on the strongly connected component. CHREC_BEFORE is the | |
588 | information that we already have collected up to this point. | |
b8698a0f L |
589 | TO_ADD is the evolution of "c". |
590 | ||
9baba81b SP |
591 | When CHREC_BEFORE has an evolution part in LOOP_NB, add to this |
592 | evolution the expression TO_ADD, otherwise construct an evolution | |
593 | part for this loop. */ | |
594 | ||
595 | static tree | |
e2157b49 | 596 | add_to_evolution_1 (unsigned loop_nb, tree chrec_before, tree to_add, |
726a989a | 597 | gimple at_stmt) |
9baba81b | 598 | { |
e2157b49 | 599 | tree type, left, right; |
0fc822d0 | 600 | struct loop *loop = get_loop (cfun, loop_nb), *chloop; |
e2157b49 | 601 | |
9baba81b SP |
602 | switch (TREE_CODE (chrec_before)) |
603 | { | |
604 | case POLYNOMIAL_CHREC: | |
677cc14d ZD |
605 | chloop = get_chrec_loop (chrec_before); |
606 | if (chloop == loop | |
607 | || flow_loop_nested_p (chloop, loop)) | |
9baba81b SP |
608 | { |
609 | unsigned var; | |
e2157b49 SP |
610 | |
611 | type = chrec_type (chrec_before); | |
b8698a0f | 612 | |
9baba81b | 613 | /* When there is no evolution part in this loop, build it. */ |
677cc14d | 614 | if (chloop != loop) |
9baba81b SP |
615 | { |
616 | var = loop_nb; | |
617 | left = chrec_before; | |
7e0923cd SP |
618 | right = SCALAR_FLOAT_TYPE_P (type) |
619 | ? build_real (type, dconst0) | |
620 | : build_int_cst (type, 0); | |
9baba81b SP |
621 | } |
622 | else | |
623 | { | |
624 | var = CHREC_VARIABLE (chrec_before); | |
625 | left = CHREC_LEFT (chrec_before); | |
626 | right = CHREC_RIGHT (chrec_before); | |
627 | } | |
628 | ||
e2157b49 | 629 | to_add = chrec_convert (type, to_add, at_stmt); |
5be014d5 AP |
630 | right = chrec_convert_rhs (type, right, at_stmt); |
631 | right = chrec_fold_plus (chrec_type (right), right, to_add); | |
e2157b49 | 632 | return build_polynomial_chrec (var, left, right); |
9baba81b SP |
633 | } |
634 | else | |
e2157b49 | 635 | { |
677cc14d ZD |
636 | gcc_assert (flow_loop_nested_p (loop, chloop)); |
637 | ||
e2157b49 SP |
638 | /* Search the evolution in LOOP_NB. */ |
639 | left = add_to_evolution_1 (loop_nb, CHREC_LEFT (chrec_before), | |
640 | to_add, at_stmt); | |
641 | right = CHREC_RIGHT (chrec_before); | |
5be014d5 | 642 | right = chrec_convert_rhs (chrec_type (left), right, at_stmt); |
e2157b49 SP |
643 | return build_polynomial_chrec (CHREC_VARIABLE (chrec_before), |
644 | left, right); | |
645 | } | |
b8698a0f | 646 | |
9baba81b SP |
647 | default: |
648 | /* These nodes do not depend on a loop. */ | |
649 | if (chrec_before == chrec_dont_know) | |
650 | return chrec_dont_know; | |
e2157b49 SP |
651 | |
652 | left = chrec_before; | |
5be014d5 | 653 | right = chrec_convert_rhs (chrec_type (left), to_add, at_stmt); |
e2157b49 | 654 | return build_polynomial_chrec (loop_nb, left, right); |
9baba81b SP |
655 | } |
656 | } | |
657 | ||
658 | /* Add TO_ADD to the evolution part of CHREC_BEFORE in the dimension | |
b8698a0f L |
659 | of LOOP_NB. |
660 | ||
9baba81b SP |
661 | Description (provided for completeness, for those who read code in |
662 | a plane, and for my poor 62 bytes brain that would have forgotten | |
663 | all this in the next two or three months): | |
b8698a0f | 664 | |
9baba81b SP |
665 | The algorithm of translation of programs from the SSA representation |
666 | into the chrecs syntax is based on a pattern matching. After having | |
667 | reconstructed the overall tree expression for a loop, there are only | |
668 | two cases that can arise: | |
b8698a0f | 669 | |
9baba81b SP |
670 | 1. a = loop-phi (init, a + expr) |
671 | 2. a = loop-phi (init, expr) | |
b8698a0f | 672 | |
9baba81b SP |
673 | where EXPR is either a scalar constant with respect to the analyzed |
674 | loop (this is a degree 0 polynomial), or an expression containing | |
675 | other loop-phi definitions (these are higher degree polynomials). | |
b8698a0f | 676 | |
9baba81b | 677 | Examples: |
b8698a0f L |
678 | |
679 | 1. | |
9baba81b SP |
680 | | init = ... |
681 | | loop_1 | |
682 | | a = phi (init, a + 5) | |
683 | | endloop | |
b8698a0f L |
684 | |
685 | 2. | |
9baba81b SP |
686 | | inita = ... |
687 | | initb = ... | |
688 | | loop_1 | |
689 | | a = phi (inita, 2 * b + 3) | |
690 | | b = phi (initb, b + 1) | |
691 | | endloop | |
b8698a0f L |
692 | |
693 | For the first case, the semantics of the SSA representation is: | |
694 | ||
9baba81b | 695 | | a (x) = init + \sum_{j = 0}^{x - 1} expr (j) |
b8698a0f | 696 | |
9baba81b SP |
697 | that is, there is a loop index "x" that determines the scalar value |
698 | of the variable during the loop execution. During the first | |
699 | iteration, the value is that of the initial condition INIT, while | |
700 | during the subsequent iterations, it is the sum of the initial | |
701 | condition with the sum of all the values of EXPR from the initial | |
b8698a0f L |
702 | iteration to the before last considered iteration. |
703 | ||
9baba81b | 704 | For the second case, the semantics of the SSA program is: |
b8698a0f | 705 | |
9baba81b SP |
706 | | a (x) = init, if x = 0; |
707 | | expr (x - 1), otherwise. | |
b8698a0f | 708 | |
9baba81b | 709 | The second case corresponds to the PEELED_CHREC, whose syntax is |
b8698a0f L |
710 | close to the syntax of a loop-phi-node: |
711 | ||
9baba81b | 712 | | phi (init, expr) vs. (init, expr)_x |
b8698a0f | 713 | |
9baba81b | 714 | The proof of the translation algorithm for the first case is a |
b8698a0f L |
715 | proof by structural induction based on the degree of EXPR. |
716 | ||
9baba81b SP |
717 | Degree 0: |
718 | When EXPR is a constant with respect to the analyzed loop, or in | |
719 | other words when EXPR is a polynomial of degree 0, the evolution of | |
720 | the variable A in the loop is an affine function with an initial | |
721 | condition INIT, and a step EXPR. In order to show this, we start | |
722 | from the semantics of the SSA representation: | |
b8698a0f | 723 | |
9baba81b | 724 | f (x) = init + \sum_{j = 0}^{x - 1} expr (j) |
b8698a0f | 725 | |
9baba81b | 726 | and since "expr (j)" is a constant with respect to "j", |
b8698a0f L |
727 | |
728 | f (x) = init + x * expr | |
729 | ||
9baba81b SP |
730 | Finally, based on the semantics of the pure sum chrecs, by |
731 | identification we get the corresponding chrecs syntax: | |
b8698a0f L |
732 | |
733 | f (x) = init * \binom{x}{0} + expr * \binom{x}{1} | |
9baba81b | 734 | f (x) -> {init, +, expr}_x |
b8698a0f | 735 | |
9baba81b SP |
736 | Higher degree: |
737 | Suppose that EXPR is a polynomial of degree N with respect to the | |
738 | analyzed loop_x for which we have already determined that it is | |
739 | written under the chrecs syntax: | |
b8698a0f | 740 | |
9baba81b | 741 | | expr (x) -> {b_0, +, b_1, +, ..., +, b_{n-1}} (x) |
b8698a0f | 742 | |
9baba81b | 743 | We start from the semantics of the SSA program: |
b8698a0f | 744 | |
9baba81b SP |
745 | | f (x) = init + \sum_{j = 0}^{x - 1} expr (j) |
746 | | | |
b8698a0f | 747 | | f (x) = init + \sum_{j = 0}^{x - 1} |
9baba81b SP |
748 | | (b_0 * \binom{j}{0} + ... + b_{n-1} * \binom{j}{n-1}) |
749 | | | |
b8698a0f L |
750 | | f (x) = init + \sum_{j = 0}^{x - 1} |
751 | | \sum_{k = 0}^{n - 1} (b_k * \binom{j}{k}) | |
9baba81b | 752 | | |
b8698a0f L |
753 | | f (x) = init + \sum_{k = 0}^{n - 1} |
754 | | (b_k * \sum_{j = 0}^{x - 1} \binom{j}{k}) | |
9baba81b | 755 | | |
b8698a0f L |
756 | | f (x) = init + \sum_{k = 0}^{n - 1} |
757 | | (b_k * \binom{x}{k + 1}) | |
9baba81b | 758 | | |
b8698a0f L |
759 | | f (x) = init + b_0 * \binom{x}{1} + ... |
760 | | + b_{n-1} * \binom{x}{n} | |
9baba81b | 761 | | |
b8698a0f L |
762 | | f (x) = init * \binom{x}{0} + b_0 * \binom{x}{1} + ... |
763 | | + b_{n-1} * \binom{x}{n} | |
9baba81b | 764 | | |
b8698a0f | 765 | |
9baba81b | 766 | And finally from the definition of the chrecs syntax, we identify: |
b8698a0f L |
767 | | f (x) -> {init, +, b_0, +, ..., +, b_{n-1}}_x |
768 | ||
9baba81b SP |
769 | This shows the mechanism that stands behind the add_to_evolution |
770 | function. An important point is that the use of symbolic | |
771 | parameters avoids the need of an analysis schedule. | |
b8698a0f | 772 | |
9baba81b | 773 | Example: |
b8698a0f | 774 | |
9baba81b SP |
775 | | inita = ... |
776 | | initb = ... | |
b8698a0f | 777 | | loop_1 |
9baba81b SP |
778 | | a = phi (inita, a + 2 + b) |
779 | | b = phi (initb, b + 1) | |
780 | | endloop | |
b8698a0f | 781 | |
9baba81b | 782 | When analyzing "a", the algorithm keeps "b" symbolically: |
b8698a0f | 783 | |
9baba81b | 784 | | a -> {inita, +, 2 + b}_1 |
b8698a0f | 785 | |
9baba81b | 786 | Then, after instantiation, the analyzer ends on the evolution: |
b8698a0f | 787 | |
9baba81b SP |
788 | | a -> {inita, +, 2 + initb, +, 1}_1 |
789 | ||
790 | */ | |
791 | ||
b8698a0f | 792 | static tree |
e2157b49 | 793 | add_to_evolution (unsigned loop_nb, tree chrec_before, enum tree_code code, |
726a989a | 794 | tree to_add, gimple at_stmt) |
9baba81b SP |
795 | { |
796 | tree type = chrec_type (to_add); | |
797 | tree res = NULL_TREE; | |
b8698a0f | 798 | |
9baba81b SP |
799 | if (to_add == NULL_TREE) |
800 | return chrec_before; | |
b8698a0f | 801 | |
9baba81b SP |
802 | /* TO_ADD is either a scalar, or a parameter. TO_ADD is not |
803 | instantiated at this point. */ | |
804 | if (TREE_CODE (to_add) == POLYNOMIAL_CHREC) | |
805 | /* This should not happen. */ | |
806 | return chrec_dont_know; | |
b8698a0f | 807 | |
dfedbe40 | 808 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
809 | { |
810 | fprintf (dump_file, "(add_to_evolution \n"); | |
811 | fprintf (dump_file, " (loop_nb = %d)\n", loop_nb); | |
812 | fprintf (dump_file, " (chrec_before = "); | |
813 | print_generic_expr (dump_file, chrec_before, 0); | |
814 | fprintf (dump_file, ")\n (to_add = "); | |
815 | print_generic_expr (dump_file, to_add, 0); | |
816 | fprintf (dump_file, ")\n"); | |
817 | } | |
818 | ||
819 | if (code == MINUS_EXPR) | |
9d2b0e12 VR |
820 | to_add = chrec_fold_multiply (type, to_add, SCALAR_FLOAT_TYPE_P (type) |
821 | ? build_real (type, dconstm1) | |
822 | : build_int_cst_type (type, -1)); | |
9baba81b | 823 | |
e2157b49 | 824 | res = add_to_evolution_1 (loop_nb, chrec_before, to_add, at_stmt); |
9baba81b | 825 | |
dfedbe40 | 826 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
827 | { |
828 | fprintf (dump_file, " (res = "); | |
829 | print_generic_expr (dump_file, res, 0); | |
830 | fprintf (dump_file, "))\n"); | |
831 | } | |
832 | ||
833 | return res; | |
834 | } | |
835 | ||
9baba81b SP |
836 | \f |
837 | ||
838 | /* This section selects the loops that will be good candidates for the | |
839 | scalar evolution analysis. For the moment, greedily select all the | |
840 | loop nests we could analyze. */ | |
841 | ||
9baba81b SP |
842 | /* For a loop with a single exit edge, return the COND_EXPR that |
843 | guards the exit edge. If the expression is too difficult to | |
844 | analyze, then give up. */ | |
845 | ||
b8698a0f | 846 | gimple |
22ea9ec0 | 847 | get_loop_exit_condition (const struct loop *loop) |
9baba81b | 848 | { |
726a989a | 849 | gimple res = NULL; |
ac8f6c69 | 850 | edge exit_edge = single_exit (loop); |
b8698a0f | 851 | |
dfedbe40 | 852 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b | 853 | fprintf (dump_file, "(get_loop_exit_condition \n "); |
b8698a0f | 854 | |
82b85a85 | 855 | if (exit_edge) |
9baba81b | 856 | { |
726a989a | 857 | gimple stmt; |
b8698a0f | 858 | |
726a989a RB |
859 | stmt = last_stmt (exit_edge->src); |
860 | if (gimple_code (stmt) == GIMPLE_COND) | |
861 | res = stmt; | |
9baba81b | 862 | } |
b8698a0f | 863 | |
dfedbe40 | 864 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b | 865 | { |
726a989a | 866 | print_gimple_stmt (dump_file, res, 0, 0); |
9baba81b SP |
867 | fprintf (dump_file, ")\n"); |
868 | } | |
b8698a0f | 869 | |
9baba81b SP |
870 | return res; |
871 | } | |
872 | ||
9baba81b SP |
873 | \f |
874 | /* Depth first search algorithm. */ | |
875 | ||
c59dabbe SP |
876 | typedef enum t_bool { |
877 | t_false, | |
878 | t_true, | |
879 | t_dont_know | |
880 | } t_bool; | |
881 | ||
882 | ||
726a989a | 883 | static t_bool follow_ssa_edge (struct loop *loop, gimple, gimple, tree *, int); |
9baba81b | 884 | |
726a989a | 885 | /* Follow the ssa edge into the binary expression RHS0 CODE RHS1. |
9baba81b SP |
886 | Return true if the strongly connected component has been found. */ |
887 | ||
c59dabbe | 888 | static t_bool |
726a989a RB |
889 | follow_ssa_edge_binary (struct loop *loop, gimple at_stmt, |
890 | tree type, tree rhs0, enum tree_code code, tree rhs1, | |
891 | gimple halting_phi, tree *evolution_of_loop, int limit) | |
9baba81b | 892 | { |
c59dabbe | 893 | t_bool res = t_false; |
b2a93c0a | 894 | tree evol; |
726a989a | 895 | |
5be014d5 | 896 | switch (code) |
9baba81b | 897 | { |
5be014d5 | 898 | case POINTER_PLUS_EXPR: |
9baba81b | 899 | case PLUS_EXPR: |
9baba81b SP |
900 | if (TREE_CODE (rhs0) == SSA_NAME) |
901 | { | |
902 | if (TREE_CODE (rhs1) == SSA_NAME) | |
903 | { | |
b8698a0f | 904 | /* Match an assignment under the form: |
9baba81b | 905 | "a = b + c". */ |
b8698a0f | 906 | |
9e824336 ZD |
907 | /* We want only assignments of form "name + name" contribute to |
908 | LIMIT, as the other cases do not necessarily contribute to | |
909 | the complexity of the expression. */ | |
910 | limit++; | |
911 | ||
b2a93c0a | 912 | evol = *evolution_of_loop; |
b8698a0f | 913 | res = follow_ssa_edge |
726a989a | 914 | (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, &evol, limit); |
b8698a0f | 915 | |
c59dabbe | 916 | if (res == t_true) |
b8698a0f L |
917 | *evolution_of_loop = add_to_evolution |
918 | (loop->num, | |
919 | chrec_convert (type, evol, at_stmt), | |
5be014d5 | 920 | code, rhs1, at_stmt); |
b8698a0f | 921 | |
c59dabbe | 922 | else if (res == t_false) |
9baba81b | 923 | { |
b8698a0f L |
924 | res = follow_ssa_edge |
925 | (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi, | |
c59dabbe | 926 | evolution_of_loop, limit); |
b8698a0f | 927 | |
c59dabbe | 928 | if (res == t_true) |
b8698a0f L |
929 | *evolution_of_loop = add_to_evolution |
930 | (loop->num, | |
931 | chrec_convert (type, *evolution_of_loop, at_stmt), | |
5be014d5 | 932 | code, rhs0, at_stmt); |
c59dabbe SP |
933 | |
934 | else if (res == t_dont_know) | |
935 | *evolution_of_loop = chrec_dont_know; | |
9baba81b | 936 | } |
c59dabbe SP |
937 | |
938 | else if (res == t_dont_know) | |
939 | *evolution_of_loop = chrec_dont_know; | |
9baba81b | 940 | } |
b8698a0f | 941 | |
9baba81b SP |
942 | else |
943 | { | |
b8698a0f | 944 | /* Match an assignment under the form: |
9baba81b | 945 | "a = b + ...". */ |
b8698a0f L |
946 | res = follow_ssa_edge |
947 | (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, | |
c59dabbe SP |
948 | evolution_of_loop, limit); |
949 | if (res == t_true) | |
b8698a0f | 950 | *evolution_of_loop = add_to_evolution |
726a989a | 951 | (loop->num, chrec_convert (type, *evolution_of_loop, |
1e8552eb | 952 | at_stmt), |
5be014d5 | 953 | code, rhs1, at_stmt); |
c59dabbe SP |
954 | |
955 | else if (res == t_dont_know) | |
956 | *evolution_of_loop = chrec_dont_know; | |
9baba81b SP |
957 | } |
958 | } | |
b8698a0f | 959 | |
9baba81b SP |
960 | else if (TREE_CODE (rhs1) == SSA_NAME) |
961 | { | |
b8698a0f | 962 | /* Match an assignment under the form: |
9baba81b | 963 | "a = ... + c". */ |
b8698a0f L |
964 | res = follow_ssa_edge |
965 | (loop, SSA_NAME_DEF_STMT (rhs1), halting_phi, | |
c59dabbe SP |
966 | evolution_of_loop, limit); |
967 | if (res == t_true) | |
b8698a0f | 968 | *evolution_of_loop = add_to_evolution |
726a989a | 969 | (loop->num, chrec_convert (type, *evolution_of_loop, |
1e8552eb | 970 | at_stmt), |
5be014d5 | 971 | code, rhs0, at_stmt); |
c59dabbe SP |
972 | |
973 | else if (res == t_dont_know) | |
974 | *evolution_of_loop = chrec_dont_know; | |
9baba81b SP |
975 | } |
976 | ||
977 | else | |
b8698a0f | 978 | /* Otherwise, match an assignment under the form: |
9baba81b SP |
979 | "a = ... + ...". */ |
980 | /* And there is nothing to do. */ | |
c59dabbe | 981 | res = t_false; |
9baba81b | 982 | break; |
b8698a0f | 983 | |
9baba81b SP |
984 | case MINUS_EXPR: |
985 | /* This case is under the form "opnd0 = rhs0 - rhs1". */ | |
9baba81b | 986 | if (TREE_CODE (rhs0) == SSA_NAME) |
9baba81b | 987 | { |
b8698a0f | 988 | /* Match an assignment under the form: |
f8e9d512 | 989 | "a = b - ...". */ |
9e824336 ZD |
990 | |
991 | /* We want only assignments of form "name - name" contribute to | |
992 | LIMIT, as the other cases do not necessarily contribute to | |
993 | the complexity of the expression. */ | |
994 | if (TREE_CODE (rhs1) == SSA_NAME) | |
995 | limit++; | |
996 | ||
b8698a0f | 997 | res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), halting_phi, |
c59dabbe SP |
998 | evolution_of_loop, limit); |
999 | if (res == t_true) | |
b8698a0f | 1000 | *evolution_of_loop = add_to_evolution |
726a989a | 1001 | (loop->num, chrec_convert (type, *evolution_of_loop, at_stmt), |
e2157b49 | 1002 | MINUS_EXPR, rhs1, at_stmt); |
c59dabbe SP |
1003 | |
1004 | else if (res == t_dont_know) | |
1005 | *evolution_of_loop = chrec_dont_know; | |
9baba81b | 1006 | } |
9baba81b | 1007 | else |
b8698a0f | 1008 | /* Otherwise, match an assignment under the form: |
9baba81b SP |
1009 | "a = ... - ...". */ |
1010 | /* And there is nothing to do. */ | |
c59dabbe | 1011 | res = t_false; |
9baba81b | 1012 | break; |
726a989a RB |
1013 | |
1014 | default: | |
1015 | res = t_false; | |
1016 | } | |
1017 | ||
1018 | return res; | |
1019 | } | |
b8698a0f | 1020 | |
726a989a RB |
1021 | /* Follow the ssa edge into the expression EXPR. |
1022 | Return true if the strongly connected component has been found. */ | |
1023 | ||
1024 | static t_bool | |
b8698a0f | 1025 | follow_ssa_edge_expr (struct loop *loop, gimple at_stmt, tree expr, |
726a989a RB |
1026 | gimple halting_phi, tree *evolution_of_loop, int limit) |
1027 | { | |
5aefc6a0 EB |
1028 | enum tree_code code = TREE_CODE (expr); |
1029 | tree type = TREE_TYPE (expr), rhs0, rhs1; | |
1030 | t_bool res; | |
1031 | ||
726a989a | 1032 | /* The EXPR is one of the following cases: |
b8698a0f | 1033 | - an SSA_NAME, |
726a989a | 1034 | - an INTEGER_CST, |
b8698a0f L |
1035 | - a PLUS_EXPR, |
1036 | - a POINTER_PLUS_EXPR, | |
726a989a RB |
1037 | - a MINUS_EXPR, |
1038 | - an ASSERT_EXPR, | |
1039 | - other cases are not yet handled. */ | |
5aefc6a0 | 1040 | |
726a989a RB |
1041 | switch (code) |
1042 | { | |
5aefc6a0 | 1043 | CASE_CONVERT: |
726a989a RB |
1044 | /* This assignment is under the form "a_1 = (cast) rhs. */ |
1045 | res = follow_ssa_edge_expr (loop, at_stmt, TREE_OPERAND (expr, 0), | |
1046 | halting_phi, evolution_of_loop, limit); | |
1047 | *evolution_of_loop = chrec_convert (type, *evolution_of_loop, at_stmt); | |
1048 | break; | |
1049 | ||
1050 | case INTEGER_CST: | |
1051 | /* This assignment is under the form "a_1 = 7". */ | |
1052 | res = t_false; | |
1053 | break; | |
5aefc6a0 | 1054 | |
726a989a RB |
1055 | case SSA_NAME: |
1056 | /* This assignment is under the form: "a_1 = b_2". */ | |
b8698a0f | 1057 | res = follow_ssa_edge |
726a989a RB |
1058 | (loop, SSA_NAME_DEF_STMT (expr), halting_phi, evolution_of_loop, limit); |
1059 | break; | |
5aefc6a0 | 1060 | |
726a989a RB |
1061 | case POINTER_PLUS_EXPR: |
1062 | case PLUS_EXPR: | |
1063 | case MINUS_EXPR: | |
1064 | /* This case is under the form "rhs0 +- rhs1". */ | |
1065 | rhs0 = TREE_OPERAND (expr, 0); | |
1066 | rhs1 = TREE_OPERAND (expr, 1); | |
5aefc6a0 EB |
1067 | type = TREE_TYPE (rhs0); |
1068 | STRIP_USELESS_TYPE_CONVERSION (rhs0); | |
1069 | STRIP_USELESS_TYPE_CONVERSION (rhs1); | |
1070 | res = follow_ssa_edge_binary (loop, at_stmt, type, rhs0, code, rhs1, | |
1071 | halting_phi, evolution_of_loop, limit); | |
1072 | break; | |
726a989a | 1073 | |
70f34814 RG |
1074 | case ADDR_EXPR: |
1075 | /* Handle &MEM[ptr + CST] which is equivalent to POINTER_PLUS_EXPR. */ | |
1076 | if (TREE_CODE (TREE_OPERAND (expr, 0)) == MEM_REF) | |
1077 | { | |
1078 | expr = TREE_OPERAND (expr, 0); | |
1079 | rhs0 = TREE_OPERAND (expr, 0); | |
1080 | rhs1 = TREE_OPERAND (expr, 1); | |
1081 | type = TREE_TYPE (rhs0); | |
1082 | STRIP_USELESS_TYPE_CONVERSION (rhs0); | |
1083 | STRIP_USELESS_TYPE_CONVERSION (rhs1); | |
1084 | res = follow_ssa_edge_binary (loop, at_stmt, type, | |
1085 | rhs0, POINTER_PLUS_EXPR, rhs1, | |
1086 | halting_phi, evolution_of_loop, limit); | |
1087 | } | |
1088 | else | |
1089 | res = t_false; | |
1090 | break; | |
1091 | ||
0bca51f0 | 1092 | case ASSERT_EXPR: |
5aefc6a0 EB |
1093 | /* This assignment is of the form: "a_1 = ASSERT_EXPR <a_2, ...>" |
1094 | It must be handled as a copy assignment of the form a_1 = a_2. */ | |
1095 | rhs0 = ASSERT_EXPR_VAR (expr); | |
1096 | if (TREE_CODE (rhs0) == SSA_NAME) | |
1097 | res = follow_ssa_edge (loop, SSA_NAME_DEF_STMT (rhs0), | |
1098 | halting_phi, evolution_of_loop, limit); | |
1099 | else | |
1100 | res = t_false; | |
1101 | break; | |
0bca51f0 | 1102 | |
9baba81b | 1103 | default: |
c59dabbe | 1104 | res = t_false; |
9baba81b SP |
1105 | break; |
1106 | } | |
5aefc6a0 | 1107 | |
9baba81b SP |
1108 | return res; |
1109 | } | |
1110 | ||
726a989a RB |
1111 | /* Follow the ssa edge into the right hand side of an assignment STMT. |
1112 | Return true if the strongly connected component has been found. */ | |
1113 | ||
1114 | static t_bool | |
1115 | follow_ssa_edge_in_rhs (struct loop *loop, gimple stmt, | |
1116 | gimple halting_phi, tree *evolution_of_loop, int limit) | |
1117 | { | |
726a989a | 1118 | enum tree_code code = gimple_assign_rhs_code (stmt); |
5aefc6a0 EB |
1119 | tree type = gimple_expr_type (stmt), rhs1, rhs2; |
1120 | t_bool res; | |
726a989a | 1121 | |
5aefc6a0 | 1122 | switch (code) |
726a989a | 1123 | { |
5aefc6a0 EB |
1124 | CASE_CONVERT: |
1125 | /* This assignment is under the form "a_1 = (cast) rhs. */ | |
1126 | res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt), | |
1127 | halting_phi, evolution_of_loop, limit); | |
1128 | *evolution_of_loop = chrec_convert (type, *evolution_of_loop, stmt); | |
1129 | break; | |
1130 | ||
1131 | case POINTER_PLUS_EXPR: | |
1132 | case PLUS_EXPR: | |
1133 | case MINUS_EXPR: | |
1134 | rhs1 = gimple_assign_rhs1 (stmt); | |
1135 | rhs2 = gimple_assign_rhs2 (stmt); | |
1136 | type = TREE_TYPE (rhs1); | |
1137 | res = follow_ssa_edge_binary (loop, stmt, type, rhs1, code, rhs2, | |
218d1c24 | 1138 | halting_phi, evolution_of_loop, limit); |
5aefc6a0 | 1139 | break; |
218d1c24 | 1140 | |
726a989a | 1141 | default: |
5aefc6a0 EB |
1142 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) |
1143 | res = follow_ssa_edge_expr (loop, stmt, gimple_assign_rhs1 (stmt), | |
1144 | halting_phi, evolution_of_loop, limit); | |
1145 | else | |
1146 | res = t_false; | |
1147 | break; | |
726a989a | 1148 | } |
5aefc6a0 EB |
1149 | |
1150 | return res; | |
726a989a RB |
1151 | } |
1152 | ||
9baba81b SP |
1153 | /* Checks whether the I-th argument of a PHI comes from a backedge. */ |
1154 | ||
1155 | static bool | |
726a989a | 1156 | backedge_phi_arg_p (gimple phi, int i) |
9baba81b | 1157 | { |
726a989a | 1158 | const_edge e = gimple_phi_arg_edge (phi, i); |
9baba81b SP |
1159 | |
1160 | /* We would in fact like to test EDGE_DFS_BACK here, but we do not care | |
1161 | about updating it anywhere, and this should work as well most of the | |
1162 | time. */ | |
1163 | if (e->flags & EDGE_IRREDUCIBLE_LOOP) | |
1164 | return true; | |
1165 | ||
1166 | return false; | |
1167 | } | |
1168 | ||
1169 | /* Helper function for one branch of the condition-phi-node. Return | |
1170 | true if the strongly connected component has been found following | |
1171 | this path. */ | |
1172 | ||
c59dabbe | 1173 | static inline t_bool |
9baba81b | 1174 | follow_ssa_edge_in_condition_phi_branch (int i, |
b8698a0f L |
1175 | struct loop *loop, |
1176 | gimple condition_phi, | |
726a989a | 1177 | gimple halting_phi, |
9baba81b | 1178 | tree *evolution_of_branch, |
c59dabbe | 1179 | tree init_cond, int limit) |
9baba81b SP |
1180 | { |
1181 | tree branch = PHI_ARG_DEF (condition_phi, i); | |
1182 | *evolution_of_branch = chrec_dont_know; | |
1183 | ||
1184 | /* Do not follow back edges (they must belong to an irreducible loop, which | |
1185 | we really do not want to worry about). */ | |
1186 | if (backedge_phi_arg_p (condition_phi, i)) | |
c59dabbe | 1187 | return t_false; |
9baba81b SP |
1188 | |
1189 | if (TREE_CODE (branch) == SSA_NAME) | |
1190 | { | |
1191 | *evolution_of_branch = init_cond; | |
b8698a0f | 1192 | return follow_ssa_edge (loop, SSA_NAME_DEF_STMT (branch), halting_phi, |
c59dabbe | 1193 | evolution_of_branch, limit); |
9baba81b SP |
1194 | } |
1195 | ||
b8698a0f | 1196 | /* This case occurs when one of the condition branches sets |
89dbed81 | 1197 | the variable to a constant: i.e. a phi-node like |
b8698a0f L |
1198 | "a_2 = PHI <a_7(5), 2(6)>;". |
1199 | ||
1200 | FIXME: This case have to be refined correctly: | |
9baba81b SP |
1201 | in some cases it is possible to say something better than |
1202 | chrec_dont_know, for example using a wrap-around notation. */ | |
c59dabbe | 1203 | return t_false; |
9baba81b SP |
1204 | } |
1205 | ||
1206 | /* This function merges the branches of a condition-phi-node in a | |
1207 | loop. */ | |
1208 | ||
c59dabbe | 1209 | static t_bool |
9baba81b | 1210 | follow_ssa_edge_in_condition_phi (struct loop *loop, |
b8698a0f L |
1211 | gimple condition_phi, |
1212 | gimple halting_phi, | |
c59dabbe | 1213 | tree *evolution_of_loop, int limit) |
9baba81b | 1214 | { |
726a989a | 1215 | int i, n; |
9baba81b SP |
1216 | tree init = *evolution_of_loop; |
1217 | tree evolution_of_branch; | |
c59dabbe SP |
1218 | t_bool res = follow_ssa_edge_in_condition_phi_branch (0, loop, condition_phi, |
1219 | halting_phi, | |
1220 | &evolution_of_branch, | |
1221 | init, limit); | |
1222 | if (res == t_false || res == t_dont_know) | |
1223 | return res; | |
9baba81b | 1224 | |
9baba81b SP |
1225 | *evolution_of_loop = evolution_of_branch; |
1226 | ||
726a989a | 1227 | n = gimple_phi_num_args (condition_phi); |
726a989a | 1228 | for (i = 1; i < n; i++) |
9baba81b | 1229 | { |
e0afb98a SP |
1230 | /* Quickly give up when the evolution of one of the branches is |
1231 | not known. */ | |
1232 | if (*evolution_of_loop == chrec_dont_know) | |
c59dabbe | 1233 | return t_true; |
e0afb98a | 1234 | |
788d3075 RG |
1235 | /* Increase the limit by the PHI argument number to avoid exponential |
1236 | time and memory complexity. */ | |
c59dabbe SP |
1237 | res = follow_ssa_edge_in_condition_phi_branch (i, loop, condition_phi, |
1238 | halting_phi, | |
1239 | &evolution_of_branch, | |
788d3075 | 1240 | init, limit + i); |
c59dabbe SP |
1241 | if (res == t_false || res == t_dont_know) |
1242 | return res; | |
9baba81b SP |
1243 | |
1244 | *evolution_of_loop = chrec_merge (*evolution_of_loop, | |
1245 | evolution_of_branch); | |
1246 | } | |
b8698a0f | 1247 | |
c59dabbe | 1248 | return t_true; |
9baba81b SP |
1249 | } |
1250 | ||
1251 | /* Follow an SSA edge in an inner loop. It computes the overall | |
1252 | effect of the loop, and following the symbolic initial conditions, | |
1253 | it follows the edges in the parent loop. The inner loop is | |
1254 | considered as a single statement. */ | |
1255 | ||
c59dabbe | 1256 | static t_bool |
9baba81b | 1257 | follow_ssa_edge_inner_loop_phi (struct loop *outer_loop, |
b8698a0f | 1258 | gimple loop_phi_node, |
726a989a | 1259 | gimple halting_phi, |
c59dabbe | 1260 | tree *evolution_of_loop, int limit) |
9baba81b SP |
1261 | { |
1262 | struct loop *loop = loop_containing_stmt (loop_phi_node); | |
1263 | tree ev = analyze_scalar_evolution (loop, PHI_RESULT (loop_phi_node)); | |
1264 | ||
1265 | /* Sometimes, the inner loop is too difficult to analyze, and the | |
1266 | result of the analysis is a symbolic parameter. */ | |
1267 | if (ev == PHI_RESULT (loop_phi_node)) | |
1268 | { | |
c59dabbe | 1269 | t_bool res = t_false; |
726a989a | 1270 | int i, n = gimple_phi_num_args (loop_phi_node); |
9baba81b | 1271 | |
726a989a | 1272 | for (i = 0; i < n; i++) |
9baba81b SP |
1273 | { |
1274 | tree arg = PHI_ARG_DEF (loop_phi_node, i); | |
1275 | basic_block bb; | |
1276 | ||
1277 | /* Follow the edges that exit the inner loop. */ | |
726a989a | 1278 | bb = gimple_phi_arg_edge (loop_phi_node, i)->src; |
9baba81b | 1279 | if (!flow_bb_inside_loop_p (loop, bb)) |
726a989a RB |
1280 | res = follow_ssa_edge_expr (outer_loop, loop_phi_node, |
1281 | arg, halting_phi, | |
1282 | evolution_of_loop, limit); | |
c59dabbe SP |
1283 | if (res == t_true) |
1284 | break; | |
9baba81b SP |
1285 | } |
1286 | ||
1287 | /* If the path crosses this loop-phi, give up. */ | |
c59dabbe | 1288 | if (res == t_true) |
9baba81b SP |
1289 | *evolution_of_loop = chrec_dont_know; |
1290 | ||
1291 | return res; | |
1292 | } | |
1293 | ||
1294 | /* Otherwise, compute the overall effect of the inner loop. */ | |
1295 | ev = compute_overall_effect_of_inner_loop (loop, ev); | |
726a989a RB |
1296 | return follow_ssa_edge_expr (outer_loop, loop_phi_node, ev, halting_phi, |
1297 | evolution_of_loop, limit); | |
9baba81b SP |
1298 | } |
1299 | ||
1300 | /* Follow an SSA edge from a loop-phi-node to itself, constructing a | |
1301 | path that is analyzed on the return walk. */ | |
1302 | ||
c59dabbe | 1303 | static t_bool |
726a989a | 1304 | follow_ssa_edge (struct loop *loop, gimple def, gimple halting_phi, |
c59dabbe | 1305 | tree *evolution_of_loop, int limit) |
9baba81b SP |
1306 | { |
1307 | struct loop *def_loop; | |
b8698a0f | 1308 | |
726a989a | 1309 | if (gimple_nop_p (def)) |
c59dabbe | 1310 | return t_false; |
b8698a0f | 1311 | |
c59dabbe | 1312 | /* Give up if the path is longer than the MAX that we allow. */ |
14dd9aab | 1313 | if (limit > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_COMPLEXITY)) |
c59dabbe | 1314 | return t_dont_know; |
b8698a0f | 1315 | |
9baba81b | 1316 | def_loop = loop_containing_stmt (def); |
b8698a0f | 1317 | |
726a989a | 1318 | switch (gimple_code (def)) |
9baba81b | 1319 | { |
726a989a | 1320 | case GIMPLE_PHI: |
9baba81b SP |
1321 | if (!loop_phi_node_p (def)) |
1322 | /* DEF is a condition-phi-node. Follow the branches, and | |
1323 | record their evolutions. Finally, merge the collected | |
1324 | information and set the approximation to the main | |
1325 | variable. */ | |
b8698a0f | 1326 | return follow_ssa_edge_in_condition_phi |
c59dabbe | 1327 | (loop, def, halting_phi, evolution_of_loop, limit); |
9baba81b SP |
1328 | |
1329 | /* When the analyzed phi is the halting_phi, the | |
1330 | depth-first search is over: we have found a path from | |
1331 | the halting_phi to itself in the loop. */ | |
1332 | if (def == halting_phi) | |
c59dabbe | 1333 | return t_true; |
b8698a0f | 1334 | |
9baba81b | 1335 | /* Otherwise, the evolution of the HALTING_PHI depends |
89dbed81 | 1336 | on the evolution of another loop-phi-node, i.e. the |
9baba81b SP |
1337 | evolution function is a higher degree polynomial. */ |
1338 | if (def_loop == loop) | |
c59dabbe | 1339 | return t_false; |
b8698a0f | 1340 | |
9baba81b SP |
1341 | /* Inner loop. */ |
1342 | if (flow_loop_nested_p (loop, def_loop)) | |
b8698a0f | 1343 | return follow_ssa_edge_inner_loop_phi |
9e824336 | 1344 | (loop, def, halting_phi, evolution_of_loop, limit + 1); |
9baba81b SP |
1345 | |
1346 | /* Outer loop. */ | |
c59dabbe | 1347 | return t_false; |
9baba81b | 1348 | |
726a989a | 1349 | case GIMPLE_ASSIGN: |
b8698a0f | 1350 | return follow_ssa_edge_in_rhs (loop, def, halting_phi, |
c59dabbe | 1351 | evolution_of_loop, limit); |
b8698a0f | 1352 | |
9baba81b SP |
1353 | default: |
1354 | /* At this level of abstraction, the program is just a set | |
726a989a | 1355 | of GIMPLE_ASSIGNs and PHI_NODEs. In principle there is no |
9baba81b | 1356 | other node to be handled. */ |
c59dabbe | 1357 | return t_false; |
9baba81b SP |
1358 | } |
1359 | } | |
1360 | ||
1361 | \f | |
1362 | ||
1363 | /* Given a LOOP_PHI_NODE, this function determines the evolution | |
1364 | function from LOOP_PHI_NODE to LOOP_PHI_NODE in the loop. */ | |
1365 | ||
1366 | static tree | |
b8698a0f | 1367 | analyze_evolution_in_loop (gimple loop_phi_node, |
9baba81b SP |
1368 | tree init_cond) |
1369 | { | |
726a989a | 1370 | int i, n = gimple_phi_num_args (loop_phi_node); |
9baba81b SP |
1371 | tree evolution_function = chrec_not_analyzed_yet; |
1372 | struct loop *loop = loop_containing_stmt (loop_phi_node); | |
1373 | basic_block bb; | |
b8698a0f | 1374 | |
dfedbe40 | 1375 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
1376 | { |
1377 | fprintf (dump_file, "(analyze_evolution_in_loop \n"); | |
1378 | fprintf (dump_file, " (loop_phi_node = "); | |
726a989a | 1379 | print_gimple_stmt (dump_file, loop_phi_node, 0, 0); |
9baba81b SP |
1380 | fprintf (dump_file, ")\n"); |
1381 | } | |
b8698a0f | 1382 | |
726a989a | 1383 | for (i = 0; i < n; i++) |
9baba81b SP |
1384 | { |
1385 | tree arg = PHI_ARG_DEF (loop_phi_node, i); | |
726a989a RB |
1386 | gimple ssa_chain; |
1387 | tree ev_fn; | |
874caa00 | 1388 | t_bool res; |
9baba81b SP |
1389 | |
1390 | /* Select the edges that enter the loop body. */ | |
726a989a | 1391 | bb = gimple_phi_arg_edge (loop_phi_node, i)->src; |
9baba81b SP |
1392 | if (!flow_bb_inside_loop_p (loop, bb)) |
1393 | continue; | |
f29deac9 | 1394 | |
9baba81b SP |
1395 | if (TREE_CODE (arg) == SSA_NAME) |
1396 | { | |
f29deac9 SP |
1397 | bool val = false; |
1398 | ||
9baba81b SP |
1399 | ssa_chain = SSA_NAME_DEF_STMT (arg); |
1400 | ||
1401 | /* Pass in the initial condition to the follow edge function. */ | |
1402 | ev_fn = init_cond; | |
c59dabbe | 1403 | res = follow_ssa_edge (loop, ssa_chain, loop_phi_node, &ev_fn, 0); |
f29deac9 SP |
1404 | |
1405 | /* If ev_fn has no evolution in the inner loop, and the | |
1406 | init_cond is not equal to ev_fn, then we have an | |
1407 | ambiguity between two possible values, as we cannot know | |
1408 | the number of iterations at this point. */ | |
1409 | if (TREE_CODE (ev_fn) != POLYNOMIAL_CHREC | |
1410 | && no_evolution_in_loop_p (ev_fn, loop->num, &val) && val | |
1411 | && !operand_equal_p (init_cond, ev_fn, 0)) | |
1412 | ev_fn = chrec_dont_know; | |
9baba81b SP |
1413 | } |
1414 | else | |
874caa00 | 1415 | res = t_false; |
f29deac9 | 1416 | |
9baba81b SP |
1417 | /* When it is impossible to go back on the same |
1418 | loop_phi_node by following the ssa edges, the | |
89dbed81 | 1419 | evolution is represented by a peeled chrec, i.e. the |
9baba81b | 1420 | first iteration, EV_FN has the value INIT_COND, then |
b8698a0f | 1421 | all the other iterations it has the value of ARG. |
9baba81b | 1422 | For the moment, PEELED_CHREC nodes are not built. */ |
874caa00 | 1423 | if (res != t_true) |
9baba81b | 1424 | ev_fn = chrec_dont_know; |
b8698a0f | 1425 | |
9baba81b | 1426 | /* When there are multiple back edges of the loop (which in fact never |
8c27b7d4 | 1427 | happens currently, but nevertheless), merge their evolutions. */ |
9baba81b SP |
1428 | evolution_function = chrec_merge (evolution_function, ev_fn); |
1429 | } | |
b8698a0f | 1430 | |
dfedbe40 | 1431 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
1432 | { |
1433 | fprintf (dump_file, " (evolution_function = "); | |
1434 | print_generic_expr (dump_file, evolution_function, 0); | |
1435 | fprintf (dump_file, "))\n"); | |
1436 | } | |
b8698a0f | 1437 | |
9baba81b SP |
1438 | return evolution_function; |
1439 | } | |
1440 | ||
1441 | /* Given a loop-phi-node, return the initial conditions of the | |
1442 | variable on entry of the loop. When the CCP has propagated | |
1443 | constants into the loop-phi-node, the initial condition is | |
1444 | instantiated, otherwise the initial condition is kept symbolic. | |
1445 | This analyzer does not analyze the evolution outside the current | |
1446 | loop, and leaves this task to the on-demand tree reconstructor. */ | |
1447 | ||
b8698a0f | 1448 | static tree |
726a989a | 1449 | analyze_initial_condition (gimple loop_phi_node) |
9baba81b | 1450 | { |
726a989a | 1451 | int i, n; |
9baba81b | 1452 | tree init_cond = chrec_not_analyzed_yet; |
726a989a | 1453 | struct loop *loop = loop_containing_stmt (loop_phi_node); |
b8698a0f | 1454 | |
dfedbe40 | 1455 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
1456 | { |
1457 | fprintf (dump_file, "(analyze_initial_condition \n"); | |
1458 | fprintf (dump_file, " (loop_phi_node = \n"); | |
726a989a | 1459 | print_gimple_stmt (dump_file, loop_phi_node, 0, 0); |
9baba81b SP |
1460 | fprintf (dump_file, ")\n"); |
1461 | } | |
b8698a0f | 1462 | |
726a989a RB |
1463 | n = gimple_phi_num_args (loop_phi_node); |
1464 | for (i = 0; i < n; i++) | |
9baba81b SP |
1465 | { |
1466 | tree branch = PHI_ARG_DEF (loop_phi_node, i); | |
726a989a | 1467 | basic_block bb = gimple_phi_arg_edge (loop_phi_node, i)->src; |
b8698a0f | 1468 | |
9baba81b SP |
1469 | /* When the branch is oriented to the loop's body, it does |
1470 | not contribute to the initial condition. */ | |
1471 | if (flow_bb_inside_loop_p (loop, bb)) | |
1472 | continue; | |
1473 | ||
1474 | if (init_cond == chrec_not_analyzed_yet) | |
1475 | { | |
1476 | init_cond = branch; | |
1477 | continue; | |
1478 | } | |
1479 | ||
1480 | if (TREE_CODE (branch) == SSA_NAME) | |
1481 | { | |
1482 | init_cond = chrec_dont_know; | |
1483 | break; | |
1484 | } | |
1485 | ||
1486 | init_cond = chrec_merge (init_cond, branch); | |
1487 | } | |
1488 | ||
1489 | /* Ooops -- a loop without an entry??? */ | |
1490 | if (init_cond == chrec_not_analyzed_yet) | |
1491 | init_cond = chrec_dont_know; | |
1492 | ||
bf1cbdc6 RG |
1493 | /* During early loop unrolling we do not have fully constant propagated IL. |
1494 | Handle degenerate PHIs here to not miss important unrollings. */ | |
1495 | if (TREE_CODE (init_cond) == SSA_NAME) | |
1496 | { | |
1497 | gimple def = SSA_NAME_DEF_STMT (init_cond); | |
1498 | tree res; | |
1499 | if (gimple_code (def) == GIMPLE_PHI | |
1500 | && (res = degenerate_phi_result (def)) != NULL_TREE | |
1501 | /* Only allow invariants here, otherwise we may break | |
1502 | loop-closed SSA form. */ | |
1503 | && is_gimple_min_invariant (res)) | |
1504 | init_cond = res; | |
1505 | } | |
1506 | ||
dfedbe40 | 1507 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
1508 | { |
1509 | fprintf (dump_file, " (init_cond = "); | |
1510 | print_generic_expr (dump_file, init_cond, 0); | |
1511 | fprintf (dump_file, "))\n"); | |
1512 | } | |
b8698a0f | 1513 | |
9baba81b SP |
1514 | return init_cond; |
1515 | } | |
1516 | ||
1517 | /* Analyze the scalar evolution for LOOP_PHI_NODE. */ | |
1518 | ||
b8698a0f | 1519 | static tree |
726a989a | 1520 | interpret_loop_phi (struct loop *loop, gimple loop_phi_node) |
9baba81b SP |
1521 | { |
1522 | tree res; | |
1523 | struct loop *phi_loop = loop_containing_stmt (loop_phi_node); | |
1524 | tree init_cond; | |
b8698a0f | 1525 | |
9baba81b SP |
1526 | if (phi_loop != loop) |
1527 | { | |
1528 | struct loop *subloop; | |
1529 | tree evolution_fn = analyze_scalar_evolution | |
1530 | (phi_loop, PHI_RESULT (loop_phi_node)); | |
1531 | ||
1532 | /* Dive one level deeper. */ | |
9ba025a2 | 1533 | subloop = superloop_at_depth (phi_loop, loop_depth (loop) + 1); |
9baba81b SP |
1534 | |
1535 | /* Interpret the subloop. */ | |
1536 | res = compute_overall_effect_of_inner_loop (subloop, evolution_fn); | |
1537 | return res; | |
1538 | } | |
1539 | ||
1540 | /* Otherwise really interpret the loop phi. */ | |
1541 | init_cond = analyze_initial_condition (loop_phi_node); | |
1542 | res = analyze_evolution_in_loop (loop_phi_node, init_cond); | |
1543 | ||
73c865fa RG |
1544 | /* Verify we maintained the correct initial condition throughout |
1545 | possible conversions in the SSA chain. */ | |
1546 | if (res != chrec_dont_know) | |
1547 | { | |
1548 | tree new_init = res; | |
1549 | if (CONVERT_EXPR_P (res) | |
1550 | && TREE_CODE (TREE_OPERAND (res, 0)) == POLYNOMIAL_CHREC) | |
1551 | new_init = fold_convert (TREE_TYPE (res), | |
1552 | CHREC_LEFT (TREE_OPERAND (res, 0))); | |
1553 | else if (TREE_CODE (res) == POLYNOMIAL_CHREC) | |
1554 | new_init = CHREC_LEFT (res); | |
1555 | STRIP_USELESS_TYPE_CONVERSION (new_init); | |
eb723fa3 RG |
1556 | if (TREE_CODE (new_init) == POLYNOMIAL_CHREC |
1557 | || !operand_equal_p (init_cond, new_init, 0)) | |
73c865fa RG |
1558 | return chrec_dont_know; |
1559 | } | |
1560 | ||
9baba81b SP |
1561 | return res; |
1562 | } | |
1563 | ||
1564 | /* This function merges the branches of a condition-phi-node, | |
1565 | contained in the outermost loop, and whose arguments are already | |
1566 | analyzed. */ | |
1567 | ||
1568 | static tree | |
726a989a | 1569 | interpret_condition_phi (struct loop *loop, gimple condition_phi) |
9baba81b | 1570 | { |
726a989a | 1571 | int i, n = gimple_phi_num_args (condition_phi); |
9baba81b | 1572 | tree res = chrec_not_analyzed_yet; |
b8698a0f | 1573 | |
726a989a | 1574 | for (i = 0; i < n; i++) |
9baba81b SP |
1575 | { |
1576 | tree branch_chrec; | |
b8698a0f | 1577 | |
9baba81b SP |
1578 | if (backedge_phi_arg_p (condition_phi, i)) |
1579 | { | |
1580 | res = chrec_dont_know; | |
1581 | break; | |
1582 | } | |
1583 | ||
1584 | branch_chrec = analyze_scalar_evolution | |
1585 | (loop, PHI_ARG_DEF (condition_phi, i)); | |
b8698a0f | 1586 | |
9baba81b SP |
1587 | res = chrec_merge (res, branch_chrec); |
1588 | } | |
1589 | ||
1590 | return res; | |
1591 | } | |
1592 | ||
726a989a | 1593 | /* Interpret the operation RHS1 OP RHS2. If we didn't |
29836d07 | 1594 | analyze this node before, follow the definitions until ending |
726a989a | 1595 | either on an analyzed GIMPLE_ASSIGN, or on a loop-phi-node. On the |
9baba81b SP |
1596 | return path, this function propagates evolutions (ala constant copy |
1597 | propagation). OPND1 is not a GIMPLE expression because we could | |
1598 | analyze the effect of an inner loop: see interpret_loop_phi. */ | |
1599 | ||
1600 | static tree | |
726a989a RB |
1601 | interpret_rhs_expr (struct loop *loop, gimple at_stmt, |
1602 | tree type, tree rhs1, enum tree_code code, tree rhs2) | |
9baba81b | 1603 | { |
726a989a | 1604 | tree res, chrec1, chrec2; |
195b4c50 | 1605 | gimple def; |
726a989a RB |
1606 | |
1607 | if (get_gimple_rhs_class (code) == GIMPLE_SINGLE_RHS) | |
1608 | { | |
1609 | if (is_gimple_min_invariant (rhs1)) | |
1610 | return chrec_convert (type, rhs1, at_stmt); | |
1611 | ||
1612 | if (code == SSA_NAME) | |
1613 | return chrec_convert (type, analyze_scalar_evolution (loop, rhs1), | |
1614 | at_stmt); | |
1e8552eb | 1615 | |
726a989a RB |
1616 | if (code == ASSERT_EXPR) |
1617 | { | |
1618 | rhs1 = ASSERT_EXPR_VAR (rhs1); | |
1619 | return chrec_convert (type, analyze_scalar_evolution (loop, rhs1), | |
1620 | at_stmt); | |
1621 | } | |
726a989a | 1622 | } |
1e8552eb | 1623 | |
726a989a | 1624 | switch (code) |
9baba81b | 1625 | { |
6a02a719 | 1626 | case ADDR_EXPR: |
bef28ced JL |
1627 | if (TREE_CODE (TREE_OPERAND (rhs1, 0)) == MEM_REF |
1628 | || handled_component_p (TREE_OPERAND (rhs1, 0))) | |
1629 | { | |
1630 | enum machine_mode mode; | |
1631 | HOST_WIDE_INT bitsize, bitpos; | |
1632 | int unsignedp; | |
1633 | int volatilep = 0; | |
1634 | tree base, offset; | |
1635 | tree chrec3; | |
1636 | tree unitpos; | |
1637 | ||
1638 | base = get_inner_reference (TREE_OPERAND (rhs1, 0), | |
1639 | &bitsize, &bitpos, &offset, | |
1640 | &mode, &unsignedp, &volatilep, false); | |
1641 | ||
1642 | if (TREE_CODE (base) == MEM_REF) | |
1643 | { | |
1644 | rhs2 = TREE_OPERAND (base, 1); | |
1645 | rhs1 = TREE_OPERAND (base, 0); | |
1646 | ||
1647 | chrec1 = analyze_scalar_evolution (loop, rhs1); | |
1648 | chrec2 = analyze_scalar_evolution (loop, rhs2); | |
1649 | chrec1 = chrec_convert (type, chrec1, at_stmt); | |
1650 | chrec2 = chrec_convert (TREE_TYPE (rhs2), chrec2, at_stmt); | |
1651 | res = chrec_fold_plus (type, chrec1, chrec2); | |
1652 | } | |
1653 | else | |
1654 | { | |
1655 | chrec1 = analyze_scalar_evolution_for_address_of (loop, base); | |
1656 | chrec1 = chrec_convert (type, chrec1, at_stmt); | |
1657 | res = chrec1; | |
1658 | } | |
6a02a719 | 1659 | |
bef28ced JL |
1660 | if (offset != NULL_TREE) |
1661 | { | |
1662 | chrec2 = analyze_scalar_evolution (loop, offset); | |
1663 | chrec2 = chrec_convert (TREE_TYPE (offset), chrec2, at_stmt); | |
1664 | res = chrec_fold_plus (type, res, chrec2); | |
1665 | } | |
1666 | ||
1667 | if (bitpos != 0) | |
1668 | { | |
1669 | gcc_assert ((bitpos % BITS_PER_UNIT) == 0); | |
1670 | ||
18dae016 | 1671 | unitpos = size_int (bitpos / BITS_PER_UNIT); |
bef28ced JL |
1672 | chrec3 = analyze_scalar_evolution (loop, unitpos); |
1673 | chrec3 = chrec_convert (TREE_TYPE (unitpos), chrec3, at_stmt); | |
1674 | res = chrec_fold_plus (type, res, chrec3); | |
1675 | } | |
1676 | } | |
1677 | else | |
1678 | res = chrec_dont_know; | |
1679 | break; | |
6a02a719 | 1680 | |
5be014d5 | 1681 | case POINTER_PLUS_EXPR: |
726a989a RB |
1682 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1683 | chrec2 = analyze_scalar_evolution (loop, rhs2); | |
1684 | chrec1 = chrec_convert (type, chrec1, at_stmt); | |
0d82a1c8 | 1685 | chrec2 = chrec_convert (TREE_TYPE (rhs2), chrec2, at_stmt); |
726a989a | 1686 | res = chrec_fold_plus (type, chrec1, chrec2); |
5be014d5 AP |
1687 | break; |
1688 | ||
9baba81b | 1689 | case PLUS_EXPR: |
726a989a RB |
1690 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1691 | chrec2 = analyze_scalar_evolution (loop, rhs2); | |
1692 | chrec1 = chrec_convert (type, chrec1, at_stmt); | |
1693 | chrec2 = chrec_convert (type, chrec2, at_stmt); | |
1694 | res = chrec_fold_plus (type, chrec1, chrec2); | |
9baba81b | 1695 | break; |
b8698a0f | 1696 | |
9baba81b | 1697 | case MINUS_EXPR: |
726a989a RB |
1698 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1699 | chrec2 = analyze_scalar_evolution (loop, rhs2); | |
1700 | chrec1 = chrec_convert (type, chrec1, at_stmt); | |
1701 | chrec2 = chrec_convert (type, chrec2, at_stmt); | |
1702 | res = chrec_fold_minus (type, chrec1, chrec2); | |
9baba81b SP |
1703 | break; |
1704 | ||
1705 | case NEGATE_EXPR: | |
726a989a RB |
1706 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1707 | chrec1 = chrec_convert (type, chrec1, at_stmt); | |
9a75ede0 | 1708 | /* TYPE may be integer, real or complex, so use fold_convert. */ |
726a989a | 1709 | res = chrec_fold_multiply (type, chrec1, |
9a75ede0 | 1710 | fold_convert (type, integer_minus_one_node)); |
9baba81b SP |
1711 | break; |
1712 | ||
418df9d7 JJ |
1713 | case BIT_NOT_EXPR: |
1714 | /* Handle ~X as -1 - X. */ | |
1715 | chrec1 = analyze_scalar_evolution (loop, rhs1); | |
1716 | chrec1 = chrec_convert (type, chrec1, at_stmt); | |
1717 | res = chrec_fold_minus (type, | |
1718 | fold_convert (type, integer_minus_one_node), | |
1719 | chrec1); | |
1720 | break; | |
1721 | ||
9baba81b | 1722 | case MULT_EXPR: |
726a989a RB |
1723 | chrec1 = analyze_scalar_evolution (loop, rhs1); |
1724 | chrec2 = analyze_scalar_evolution (loop, rhs2); | |
1725 | chrec1 = chrec_convert (type, chrec1, at_stmt); | |
1726 | chrec2 = chrec_convert (type, chrec2, at_stmt); | |
1727 | res = chrec_fold_multiply (type, chrec1, chrec2); | |
0bca51f0 | 1728 | break; |
b8698a0f | 1729 | |
1043771b | 1730 | CASE_CONVERT: |
195b4c50 RG |
1731 | /* In case we have a truncation of a widened operation that in |
1732 | the truncated type has undefined overflow behavior analyze | |
1733 | the operation done in an unsigned type of the same precision | |
1734 | as the final truncation. We cannot derive a scalar evolution | |
1735 | for the widened operation but for the truncated result. */ | |
1736 | if (TREE_CODE (type) == INTEGER_TYPE | |
1737 | && TREE_CODE (TREE_TYPE (rhs1)) == INTEGER_TYPE | |
1738 | && TYPE_PRECISION (type) < TYPE_PRECISION (TREE_TYPE (rhs1)) | |
1739 | && TYPE_OVERFLOW_UNDEFINED (type) | |
1740 | && TREE_CODE (rhs1) == SSA_NAME | |
1741 | && (def = SSA_NAME_DEF_STMT (rhs1)) | |
1742 | && is_gimple_assign (def) | |
1743 | && TREE_CODE_CLASS (gimple_assign_rhs_code (def)) == tcc_binary | |
1744 | && TREE_CODE (gimple_assign_rhs2 (def)) == INTEGER_CST) | |
1745 | { | |
1746 | tree utype = unsigned_type_for (type); | |
1747 | chrec1 = interpret_rhs_expr (loop, at_stmt, utype, | |
1748 | gimple_assign_rhs1 (def), | |
1749 | gimple_assign_rhs_code (def), | |
1750 | gimple_assign_rhs2 (def)); | |
1751 | } | |
1752 | else | |
1753 | chrec1 = analyze_scalar_evolution (loop, rhs1); | |
726a989a | 1754 | res = chrec_convert (type, chrec1, at_stmt); |
9baba81b | 1755 | break; |
b8698a0f | 1756 | |
9baba81b SP |
1757 | default: |
1758 | res = chrec_dont_know; | |
1759 | break; | |
1760 | } | |
b8698a0f | 1761 | |
9baba81b SP |
1762 | return res; |
1763 | } | |
1764 | ||
726a989a RB |
1765 | /* Interpret the expression EXPR. */ |
1766 | ||
1767 | static tree | |
1768 | interpret_expr (struct loop *loop, gimple at_stmt, tree expr) | |
1769 | { | |
1770 | enum tree_code code; | |
1771 | tree type = TREE_TYPE (expr), op0, op1; | |
1772 | ||
1773 | if (automatically_generated_chrec_p (expr)) | |
1774 | return expr; | |
1775 | ||
4e71066d RG |
1776 | if (TREE_CODE (expr) == POLYNOMIAL_CHREC |
1777 | || get_gimple_rhs_class (TREE_CODE (expr)) == GIMPLE_TERNARY_RHS) | |
726a989a RB |
1778 | return chrec_dont_know; |
1779 | ||
1780 | extract_ops_from_tree (expr, &code, &op0, &op1); | |
1781 | ||
1782 | return interpret_rhs_expr (loop, at_stmt, type, | |
1783 | op0, code, op1); | |
1784 | } | |
1785 | ||
1786 | /* Interpret the rhs of the assignment STMT. */ | |
1787 | ||
1788 | static tree | |
1789 | interpret_gimple_assign (struct loop *loop, gimple stmt) | |
1790 | { | |
1791 | tree type = TREE_TYPE (gimple_assign_lhs (stmt)); | |
1792 | enum tree_code code = gimple_assign_rhs_code (stmt); | |
1793 | ||
1794 | return interpret_rhs_expr (loop, stmt, type, | |
1795 | gimple_assign_rhs1 (stmt), code, | |
1796 | gimple_assign_rhs2 (stmt)); | |
1797 | } | |
1798 | ||
9baba81b SP |
1799 | \f |
1800 | ||
b8698a0f | 1801 | /* This section contains all the entry points: |
9baba81b SP |
1802 | - number_of_iterations_in_loop, |
1803 | - analyze_scalar_evolution, | |
1804 | - instantiate_parameters. | |
1805 | */ | |
1806 | ||
1807 | /* Compute and return the evolution function in WRTO_LOOP, the nearest | |
1808 | common ancestor of DEF_LOOP and USE_LOOP. */ | |
1809 | ||
b8698a0f L |
1810 | static tree |
1811 | compute_scalar_evolution_in_loop (struct loop *wrto_loop, | |
1812 | struct loop *def_loop, | |
9baba81b SP |
1813 | tree ev) |
1814 | { | |
492e5456 | 1815 | bool val; |
9baba81b | 1816 | tree res; |
492e5456 | 1817 | |
9baba81b SP |
1818 | if (def_loop == wrto_loop) |
1819 | return ev; | |
1820 | ||
9ba025a2 | 1821 | def_loop = superloop_at_depth (def_loop, loop_depth (wrto_loop) + 1); |
9baba81b SP |
1822 | res = compute_overall_effect_of_inner_loop (def_loop, ev); |
1823 | ||
492e5456 SP |
1824 | if (no_evolution_in_loop_p (res, wrto_loop->num, &val) && val) |
1825 | return res; | |
1826 | ||
9baba81b SP |
1827 | return analyze_scalar_evolution_1 (wrto_loop, res, chrec_not_analyzed_yet); |
1828 | } | |
1829 | ||
1830 | /* Helper recursive function. */ | |
1831 | ||
1832 | static tree | |
1833 | analyze_scalar_evolution_1 (struct loop *loop, tree var, tree res) | |
1834 | { | |
726a989a RB |
1835 | tree type = TREE_TYPE (var); |
1836 | gimple def; | |
9baba81b SP |
1837 | basic_block bb; |
1838 | struct loop *def_loop; | |
1839 | ||
42d375ed | 1840 | if (loop == NULL || TREE_CODE (type) == VECTOR_TYPE) |
9baba81b SP |
1841 | return chrec_dont_know; |
1842 | ||
1843 | if (TREE_CODE (var) != SSA_NAME) | |
726a989a | 1844 | return interpret_expr (loop, NULL, var); |
9baba81b SP |
1845 | |
1846 | def = SSA_NAME_DEF_STMT (var); | |
726a989a | 1847 | bb = gimple_bb (def); |
9baba81b SP |
1848 | def_loop = bb ? bb->loop_father : NULL; |
1849 | ||
1850 | if (bb == NULL | |
1851 | || !flow_bb_inside_loop_p (loop, bb)) | |
1852 | { | |
1853 | /* Keep the symbolic form. */ | |
1854 | res = var; | |
1855 | goto set_and_end; | |
1856 | } | |
1857 | ||
1858 | if (res != chrec_not_analyzed_yet) | |
1859 | { | |
1860 | if (loop != bb->loop_father) | |
b8698a0f | 1861 | res = compute_scalar_evolution_in_loop |
9baba81b SP |
1862 | (find_common_loop (loop, bb->loop_father), bb->loop_father, res); |
1863 | ||
1864 | goto set_and_end; | |
1865 | } | |
1866 | ||
1867 | if (loop != def_loop) | |
1868 | { | |
1869 | res = analyze_scalar_evolution_1 (def_loop, var, chrec_not_analyzed_yet); | |
1870 | res = compute_scalar_evolution_in_loop (loop, def_loop, res); | |
1871 | ||
1872 | goto set_and_end; | |
1873 | } | |
1874 | ||
726a989a | 1875 | switch (gimple_code (def)) |
9baba81b | 1876 | { |
726a989a RB |
1877 | case GIMPLE_ASSIGN: |
1878 | res = interpret_gimple_assign (loop, def); | |
9baba81b SP |
1879 | break; |
1880 | ||
726a989a | 1881 | case GIMPLE_PHI: |
9baba81b SP |
1882 | if (loop_phi_node_p (def)) |
1883 | res = interpret_loop_phi (loop, def); | |
1884 | else | |
1885 | res = interpret_condition_phi (loop, def); | |
1886 | break; | |
1887 | ||
1888 | default: | |
1889 | res = chrec_dont_know; | |
1890 | break; | |
1891 | } | |
1892 | ||
1893 | set_and_end: | |
1894 | ||
1895 | /* Keep the symbolic form. */ | |
1896 | if (res == chrec_dont_know) | |
1897 | res = var; | |
1898 | ||
1899 | if (loop == def_loop) | |
a213b219 | 1900 | set_scalar_evolution (block_before_loop (loop), var, res); |
9baba81b SP |
1901 | |
1902 | return res; | |
1903 | } | |
1904 | ||
52bdd655 SP |
1905 | /* Analyzes and returns the scalar evolution of the ssa_name VAR in |
1906 | LOOP. LOOP is the loop in which the variable is used. | |
b8698a0f | 1907 | |
9baba81b SP |
1908 | Example of use: having a pointer VAR to a SSA_NAME node, STMT a |
1909 | pointer to the statement that uses this variable, in order to | |
1910 | determine the evolution function of the variable, use the following | |
1911 | calls: | |
b8698a0f | 1912 | |
52bdd655 SP |
1913 | loop_p loop = loop_containing_stmt (stmt); |
1914 | tree chrec_with_symbols = analyze_scalar_evolution (loop, var); | |
3f227a8c | 1915 | tree chrec_instantiated = instantiate_parameters (loop, chrec_with_symbols); |
9baba81b SP |
1916 | */ |
1917 | ||
b8698a0f | 1918 | tree |
9baba81b SP |
1919 | analyze_scalar_evolution (struct loop *loop, tree var) |
1920 | { | |
1921 | tree res; | |
1922 | ||
dfedbe40 | 1923 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
1924 | { |
1925 | fprintf (dump_file, "(analyze_scalar_evolution \n"); | |
1926 | fprintf (dump_file, " (loop_nb = %d)\n", loop->num); | |
1927 | fprintf (dump_file, " (scalar = "); | |
1928 | print_generic_expr (dump_file, var, 0); | |
1929 | fprintf (dump_file, ")\n"); | |
1930 | } | |
1931 | ||
a213b219 SP |
1932 | res = get_scalar_evolution (block_before_loop (loop), var); |
1933 | res = analyze_scalar_evolution_1 (loop, var, res); | |
9baba81b | 1934 | |
dfedbe40 | 1935 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
1936 | fprintf (dump_file, ")\n"); |
1937 | ||
1938 | return res; | |
1939 | } | |
1940 | ||
bef28ced JL |
1941 | /* Analyzes and returns the scalar evolution of VAR address in LOOP. */ |
1942 | ||
1943 | static tree | |
1944 | analyze_scalar_evolution_for_address_of (struct loop *loop, tree var) | |
1945 | { | |
1946 | return analyze_scalar_evolution (loop, build_fold_addr_expr (var)); | |
1947 | } | |
1948 | ||
9baba81b | 1949 | /* Analyze scalar evolution of use of VERSION in USE_LOOP with respect to |
f017bf5e | 1950 | WRTO_LOOP (which should be a superloop of USE_LOOP) |
a6f778b2 ZD |
1951 | |
1952 | FOLDED_CASTS is set to true if resolve_mixers used | |
1953 | chrec_convert_aggressive (TODO -- not really, we are way too conservative | |
b8698a0f L |
1954 | at the moment in order to keep things simple). |
1955 | ||
f017bf5e ZD |
1956 | To illustrate the meaning of USE_LOOP and WRTO_LOOP, consider the following |
1957 | example: | |
1958 | ||
1959 | for (i = 0; i < 100; i++) -- loop 1 | |
1960 | { | |
1961 | for (j = 0; j < 100; j++) -- loop 2 | |
1962 | { | |
1963 | k1 = i; | |
1964 | k2 = j; | |
1965 | ||
1966 | use2 (k1, k2); | |
1967 | ||
1968 | for (t = 0; t < 100; t++) -- loop 3 | |
1969 | use3 (k1, k2); | |
1970 | ||
1971 | } | |
1972 | use1 (k1, k2); | |
1973 | } | |
1974 | ||
1975 | Both k1 and k2 are invariants in loop3, thus | |
1976 | analyze_scalar_evolution_in_loop (loop3, loop3, k1) = k1 | |
1977 | analyze_scalar_evolution_in_loop (loop3, loop3, k2) = k2 | |
1978 | ||
1979 | As they are invariant, it does not matter whether we consider their | |
1980 | usage in loop 3 or loop 2, hence | |
1981 | analyze_scalar_evolution_in_loop (loop2, loop3, k1) = | |
1982 | analyze_scalar_evolution_in_loop (loop2, loop2, k1) = i | |
1983 | analyze_scalar_evolution_in_loop (loop2, loop3, k2) = | |
1984 | analyze_scalar_evolution_in_loop (loop2, loop2, k2) = [0,+,1]_2 | |
1985 | ||
1986 | Similarly for their evolutions with respect to loop 1. The values of K2 | |
1987 | in the use in loop 2 vary independently on loop 1, thus we cannot express | |
1988 | the evolution with respect to loop 1: | |
1989 | analyze_scalar_evolution_in_loop (loop1, loop3, k1) = | |
1990 | analyze_scalar_evolution_in_loop (loop1, loop2, k1) = [0,+,1]_1 | |
1991 | analyze_scalar_evolution_in_loop (loop1, loop3, k2) = | |
1992 | analyze_scalar_evolution_in_loop (loop1, loop2, k2) = dont_know | |
1993 | ||
1994 | The value of k2 in the use in loop 1 is known, though: | |
1995 | analyze_scalar_evolution_in_loop (loop1, loop1, k1) = [0,+,1]_1 | |
1996 | analyze_scalar_evolution_in_loop (loop1, loop1, k2) = 100 | |
1997 | */ | |
9baba81b SP |
1998 | |
1999 | static tree | |
2000 | analyze_scalar_evolution_in_loop (struct loop *wrto_loop, struct loop *use_loop, | |
a6f778b2 | 2001 | tree version, bool *folded_casts) |
9baba81b SP |
2002 | { |
2003 | bool val = false; | |
a6f778b2 | 2004 | tree ev = version, tmp; |
9baba81b | 2005 | |
b8698a0f | 2006 | /* We cannot just do |
f017bf5e ZD |
2007 | |
2008 | tmp = analyze_scalar_evolution (use_loop, version); | |
2009 | ev = resolve_mixers (wrto_loop, tmp); | |
2010 | ||
2011 | as resolve_mixers would query the scalar evolution with respect to | |
2012 | wrto_loop. For example, in the situation described in the function | |
2013 | comment, suppose that wrto_loop = loop1, use_loop = loop3 and | |
2014 | version = k2. Then | |
2015 | ||
2016 | analyze_scalar_evolution (use_loop, version) = k2 | |
2017 | ||
2018 | and resolve_mixers (loop1, k2) finds that the value of k2 in loop 1 | |
2019 | is 100, which is a wrong result, since we are interested in the | |
2020 | value in loop 3. | |
2021 | ||
2022 | Instead, we need to proceed from use_loop to wrto_loop loop by loop, | |
2023 | each time checking that there is no evolution in the inner loop. */ | |
2024 | ||
a6f778b2 ZD |
2025 | if (folded_casts) |
2026 | *folded_casts = false; | |
9baba81b SP |
2027 | while (1) |
2028 | { | |
a6f778b2 ZD |
2029 | tmp = analyze_scalar_evolution (use_loop, ev); |
2030 | ev = resolve_mixers (use_loop, tmp); | |
2031 | ||
2032 | if (folded_casts && tmp != ev) | |
2033 | *folded_casts = true; | |
9baba81b SP |
2034 | |
2035 | if (use_loop == wrto_loop) | |
2036 | return ev; | |
2037 | ||
2038 | /* If the value of the use changes in the inner loop, we cannot express | |
2039 | its value in the outer loop (we might try to return interval chrec, | |
2040 | but we do not have a user for it anyway) */ | |
2041 | if (!no_evolution_in_loop_p (ev, use_loop->num, &val) | |
2042 | || !val) | |
2043 | return chrec_dont_know; | |
2044 | ||
9ba025a2 | 2045 | use_loop = loop_outer (use_loop); |
9baba81b SP |
2046 | } |
2047 | } | |
2048 | ||
eb0bc7af | 2049 | |
fdd43ac4 RB |
2050 | /* Hashtable helpers for a temporary hash-table used when |
2051 | instantiating a CHREC or resolving mixers. For this use | |
2052 | instantiated_below is always the same. */ | |
2053 | ||
2054 | struct instantiate_cache_entry | |
2055 | { | |
2056 | tree name; | |
2057 | tree chrec; | |
2058 | }; | |
2059 | ||
2060 | struct instantiate_cache_entry_hasher : typed_noop_remove <uintptr_t> | |
2061 | { | |
2062 | typedef uintptr_t value_type; | |
2063 | typedef instantiate_cache_entry compare_type; | |
2064 | static inline hashval_t hash (const value_type *); | |
2065 | static inline bool equal (const value_type *, const compare_type *); | |
2066 | }; | |
2067 | ||
2068 | struct instantiate_cache_type | |
eb0bc7af | 2069 | { |
fdd43ac4 RB |
2070 | hash_table <instantiate_cache_entry_hasher> htab; |
2071 | vec<instantiate_cache_entry> entries; | |
b8698a0f | 2072 | |
fdd43ac4 RB |
2073 | ~instantiate_cache_type (); |
2074 | }; | |
eb0bc7af | 2075 | |
fdd43ac4 RB |
2076 | instantiate_cache_type::~instantiate_cache_type () |
2077 | { | |
2078 | if (htab.is_created ()) | |
2079 | { | |
2080 | htab.dispose (); | |
2081 | entries.release (); | |
2082 | } | |
eb0bc7af ZD |
2083 | } |
2084 | ||
fdd43ac4 | 2085 | static instantiate_cache_type *ctbl; |
eb0bc7af | 2086 | |
fdd43ac4 RB |
2087 | inline hashval_t |
2088 | instantiate_cache_entry_hasher::hash (const value_type *idx) | |
2089 | { | |
2090 | instantiate_cache_entry *elt | |
2091 | = &ctbl->entries[reinterpret_cast <uintptr_t> (idx) - 2]; | |
2092 | return SSA_NAME_VERSION (elt->name); | |
2093 | } | |
2094 | ||
2095 | inline bool | |
2096 | instantiate_cache_entry_hasher::equal (const value_type *idx1, | |
2097 | const compare_type *elt2) | |
eb0bc7af | 2098 | { |
fdd43ac4 RB |
2099 | compare_type *elt1 = &ctbl->entries[reinterpret_cast <uintptr_t> (idx1) - 2]; |
2100 | return elt1->name == elt2->name; | |
2101 | } | |
2102 | ||
2103 | /* Returns from CACHE a pointer to the cached chrec for NAME. */ | |
2104 | ||
2105 | static tree * | |
2106 | get_instantiated_value_entry (instantiate_cache_type &cache, tree name) | |
2107 | { | |
2108 | struct instantiate_cache_entry e; | |
2109 | uintptr_t **slot; | |
2110 | ||
2111 | if (!cache.htab.is_created ()) | |
2112 | { | |
2113 | cache.htab.create (10); | |
2114 | cache.entries.create (10); | |
2115 | } | |
b8698a0f | 2116 | |
fdd43ac4 | 2117 | ctbl = &cache; |
eb0bc7af | 2118 | |
fdd43ac4 RB |
2119 | e.name = name; |
2120 | slot = cache.htab.find_slot_with_hash (&e, SSA_NAME_VERSION (name), INSERT); | |
cceb1885 | 2121 | if (!*slot) |
fdd43ac4 RB |
2122 | { |
2123 | e.chrec = chrec_not_analyzed_yet; | |
2124 | cache.entries.safe_push (e); | |
2125 | *slot = reinterpret_cast <uintptr_t *> | |
2126 | ((uintptr_t) cache.entries.length () + 1); | |
2127 | } | |
2128 | ||
2129 | return &cache.entries[reinterpret_cast <uintptr_t> (*slot) - 2].chrec; | |
eb0bc7af ZD |
2130 | } |
2131 | ||
18aed06a SP |
2132 | /* Return the closed_loop_phi node for VAR. If there is none, return |
2133 | NULL_TREE. */ | |
2134 | ||
2135 | static tree | |
2136 | loop_closed_phi_def (tree var) | |
2137 | { | |
2138 | struct loop *loop; | |
2139 | edge exit; | |
726a989a RB |
2140 | gimple phi; |
2141 | gimple_stmt_iterator psi; | |
18aed06a SP |
2142 | |
2143 | if (var == NULL_TREE | |
2144 | || TREE_CODE (var) != SSA_NAME) | |
2145 | return NULL_TREE; | |
2146 | ||
2147 | loop = loop_containing_stmt (SSA_NAME_DEF_STMT (var)); | |
ac8f6c69 | 2148 | exit = single_exit (loop); |
18aed06a SP |
2149 | if (!exit) |
2150 | return NULL_TREE; | |
2151 | ||
726a989a RB |
2152 | for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); gsi_next (&psi)) |
2153 | { | |
2154 | phi = gsi_stmt (psi); | |
2155 | if (PHI_ARG_DEF_FROM_EDGE (phi, exit) == var) | |
2156 | return PHI_RESULT (phi); | |
2157 | } | |
18aed06a SP |
2158 | |
2159 | return NULL_TREE; | |
2160 | } | |
2161 | ||
8b679c9b | 2162 | static tree instantiate_scev_r (basic_block, struct loop *, struct loop *, |
fdd43ac4 | 2163 | tree, bool, instantiate_cache_type &, int); |
320f5a78 SP |
2164 | |
2165 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW | |
2166 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
2167 | ||
2495a183 | 2168 | CHREC is an SSA_NAME to be instantiated. |
320f5a78 SP |
2169 | |
2170 | CACHE is the cache of already instantiated values. | |
2171 | ||
2172 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2173 | may wrap in signed/pointer type are folded, as long as the value of | |
2174 | the chrec is preserved. | |
2175 | ||
2176 | SIZE_EXPR is used for computing the size of the expression to be | |
2177 | instantiated, and to stop if it exceeds some limit. */ | |
2178 | ||
2179 | static tree | |
2495a183 | 2180 | instantiate_scev_name (basic_block instantiate_below, |
8b679c9b RB |
2181 | struct loop *evolution_loop, struct loop *inner_loop, |
2182 | tree chrec, | |
fdd43ac4 | 2183 | bool fold_conversions, instantiate_cache_type &cache, |
4a8fb1a1 | 2184 | int size_expr) |
320f5a78 | 2185 | { |
2495a183 SP |
2186 | tree res; |
2187 | struct loop *def_loop; | |
2188 | basic_block def_bb = gimple_bb (SSA_NAME_DEF_STMT (chrec)); | |
20179b0d | 2189 | |
2495a183 SP |
2190 | /* A parameter (or loop invariant and we do not want to include |
2191 | evolutions in outer loops), nothing to do. */ | |
2192 | if (!def_bb | |
2193 | || loop_depth (def_bb->loop_father) == 0 | |
2194 | || dominated_by_p (CDI_DOMINATORS, instantiate_below, def_bb)) | |
2195 | return chrec; | |
20179b0d | 2196 | |
2495a183 SP |
2197 | /* We cache the value of instantiated variable to avoid exponential |
2198 | time complexity due to reevaluations. We also store the convenient | |
2199 | value in the cache in order to prevent infinite recursion -- we do | |
2200 | not want to instantiate the SSA_NAME if it is in a mixer | |
2201 | structure. This is used for avoiding the instantiation of | |
2202 | recursively defined functions, such as: | |
320f5a78 | 2203 | |
2495a183 | 2204 | | a_2 -> {0, +, 1, +, a_2}_1 */ |
20179b0d | 2205 | |
fdd43ac4 RB |
2206 | tree *si; |
2207 | si = get_instantiated_value_entry (cache, chrec); | |
2208 | if (*si != chrec_not_analyzed_yet) | |
2209 | return *si; | |
20179b0d | 2210 | |
fdd43ac4 RB |
2211 | /* On recursion return chrec_dont_know. */ |
2212 | *si = chrec_dont_know; | |
320f5a78 | 2213 | |
2495a183 SP |
2214 | def_loop = find_common_loop (evolution_loop, def_bb->loop_father); |
2215 | ||
320f5a78 SP |
2216 | /* If the analysis yields a parametric chrec, instantiate the |
2217 | result again. */ | |
2218 | res = analyze_scalar_evolution (def_loop, chrec); | |
2219 | ||
2847388e | 2220 | /* Don't instantiate default definitions. */ |
320f5a78 | 2221 | if (TREE_CODE (res) == SSA_NAME |
2847388e SP |
2222 | && SSA_NAME_IS_DEFAULT_DEF (res)) |
2223 | ; | |
2224 | ||
2225 | /* Don't instantiate loop-closed-ssa phi nodes. */ | |
2226 | else if (TREE_CODE (res) == SSA_NAME | |
2227 | && loop_depth (loop_containing_stmt (SSA_NAME_DEF_STMT (res))) | |
2228 | > loop_depth (def_loop)) | |
320f5a78 SP |
2229 | { |
2230 | if (res == chrec) | |
2231 | res = loop_closed_phi_def (chrec); | |
2232 | else | |
2233 | res = chrec; | |
2234 | ||
7472eb13 SP |
2235 | /* When there is no loop_closed_phi_def, it means that the |
2236 | variable is not used after the loop: try to still compute the | |
2237 | value of the variable when exiting the loop. */ | |
2238 | if (res == NULL_TREE) | |
2239 | { | |
2240 | loop_p loop = loop_containing_stmt (SSA_NAME_DEF_STMT (chrec)); | |
2241 | res = analyze_scalar_evolution (loop, chrec); | |
2242 | res = compute_overall_effect_of_inner_loop (loop, res); | |
8b679c9b RB |
2243 | res = instantiate_scev_r (instantiate_below, evolution_loop, |
2244 | inner_loop, res, | |
7472eb13 SP |
2245 | fold_conversions, cache, size_expr); |
2246 | } | |
2247 | else if (!dominated_by_p (CDI_DOMINATORS, instantiate_below, | |
2248 | gimple_bb (SSA_NAME_DEF_STMT (res)))) | |
320f5a78 SP |
2249 | res = chrec_dont_know; |
2250 | } | |
2251 | ||
2252 | else if (res != chrec_dont_know) | |
8b679c9b RB |
2253 | { |
2254 | if (inner_loop | |
2255 | && !flow_loop_nested_p (def_bb->loop_father, inner_loop)) | |
2256 | /* ??? We could try to compute the overall effect of the loop here. */ | |
2257 | res = chrec_dont_know; | |
2258 | else | |
2259 | res = instantiate_scev_r (instantiate_below, evolution_loop, | |
2260 | inner_loop, res, | |
2261 | fold_conversions, cache, size_expr); | |
2262 | } | |
320f5a78 SP |
2263 | |
2264 | /* Store the correct value to the cache. */ | |
fdd43ac4 | 2265 | *si = res; |
320f5a78 | 2266 | return res; |
320f5a78 SP |
2267 | } |
2268 | ||
ec6636eb SP |
2269 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2270 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
2271 | ||
2272 | CHREC is a polynomial chain of recurrence to be instantiated. | |
2273 | ||
2274 | CACHE is the cache of already instantiated values. | |
2275 | ||
2276 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2277 | may wrap in signed/pointer type are folded, as long as the value of | |
2278 | the chrec is preserved. | |
2279 | ||
2280 | SIZE_EXPR is used for computing the size of the expression to be | |
2281 | instantiated, and to stop if it exceeds some limit. */ | |
2282 | ||
2283 | static tree | |
2284 | instantiate_scev_poly (basic_block instantiate_below, | |
8b679c9b RB |
2285 | struct loop *evolution_loop, struct loop *, |
2286 | tree chrec, | |
fdd43ac4 | 2287 | bool fold_conversions, instantiate_cache_type &cache, |
4a8fb1a1 | 2288 | int size_expr) |
ec6636eb SP |
2289 | { |
2290 | tree op1; | |
9e5dc77f | 2291 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
8b679c9b | 2292 | get_chrec_loop (chrec), |
ec6636eb SP |
2293 | CHREC_LEFT (chrec), fold_conversions, cache, |
2294 | size_expr); | |
2295 | if (op0 == chrec_dont_know) | |
2296 | return chrec_dont_know; | |
2297 | ||
9e5dc77f | 2298 | op1 = instantiate_scev_r (instantiate_below, evolution_loop, |
8b679c9b | 2299 | get_chrec_loop (chrec), |
ec6636eb SP |
2300 | CHREC_RIGHT (chrec), fold_conversions, cache, |
2301 | size_expr); | |
2302 | if (op1 == chrec_dont_know) | |
2303 | return chrec_dont_know; | |
2304 | ||
2305 | if (CHREC_LEFT (chrec) != op0 | |
2306 | || CHREC_RIGHT (chrec) != op1) | |
2307 | { | |
2308 | op1 = chrec_convert_rhs (chrec_type (op0), op1, NULL); | |
8b679c9b | 2309 | chrec = build_polynomial_chrec (CHREC_VARIABLE (chrec), op0, op1); |
ec6636eb | 2310 | } |
4bf4e169 | 2311 | |
ec6636eb SP |
2312 | return chrec; |
2313 | } | |
2314 | ||
15fda317 SP |
2315 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2316 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
2317 | ||
ffa34f4b | 2318 | "C0 CODE C1" is a binary expression of type TYPE to be instantiated. |
15fda317 SP |
2319 | |
2320 | CACHE is the cache of already instantiated values. | |
2321 | ||
2322 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2323 | may wrap in signed/pointer type are folded, as long as the value of | |
2324 | the chrec is preserved. | |
2325 | ||
2326 | SIZE_EXPR is used for computing the size of the expression to be | |
2327 | instantiated, and to stop if it exceeds some limit. */ | |
2328 | ||
2329 | static tree | |
2330 | instantiate_scev_binary (basic_block instantiate_below, | |
8b679c9b RB |
2331 | struct loop *evolution_loop, struct loop *inner_loop, |
2332 | tree chrec, enum tree_code code, | |
ffa34f4b | 2333 | tree type, tree c0, tree c1, |
fdd43ac4 RB |
2334 | bool fold_conversions, |
2335 | instantiate_cache_type &cache, | |
4a8fb1a1 | 2336 | int size_expr) |
15fda317 SP |
2337 | { |
2338 | tree op1; | |
8b679c9b | 2339 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop, |
ffa34f4b | 2340 | c0, fold_conversions, cache, |
15fda317 SP |
2341 | size_expr); |
2342 | if (op0 == chrec_dont_know) | |
2343 | return chrec_dont_know; | |
2344 | ||
8b679c9b | 2345 | op1 = instantiate_scev_r (instantiate_below, evolution_loop, inner_loop, |
ffa34f4b | 2346 | c1, fold_conversions, cache, |
15fda317 SP |
2347 | size_expr); |
2348 | if (op1 == chrec_dont_know) | |
2349 | return chrec_dont_know; | |
2350 | ||
ffa34f4b SP |
2351 | if (c0 != op0 |
2352 | || c1 != op1) | |
15fda317 | 2353 | { |
15fda317 SP |
2354 | op0 = chrec_convert (type, op0, NULL); |
2355 | op1 = chrec_convert_rhs (type, op1, NULL); | |
2356 | ||
ffa34f4b | 2357 | switch (code) |
15fda317 SP |
2358 | { |
2359 | case POINTER_PLUS_EXPR: | |
2360 | case PLUS_EXPR: | |
2361 | return chrec_fold_plus (type, op0, op1); | |
2362 | ||
2363 | case MINUS_EXPR: | |
2364 | return chrec_fold_minus (type, op0, op1); | |
2365 | ||
2366 | case MULT_EXPR: | |
2367 | return chrec_fold_multiply (type, op0, op1); | |
2368 | ||
2369 | default: | |
2370 | gcc_unreachable (); | |
2371 | } | |
2372 | } | |
2373 | ||
ffa34f4b | 2374 | return chrec ? chrec : fold_build2 (code, type, c0, c1); |
15fda317 SP |
2375 | } |
2376 | ||
dbc08079 SP |
2377 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2378 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
2379 | ||
2380 | "CHREC" is an array reference to be instantiated. | |
2381 | ||
2382 | CACHE is the cache of already instantiated values. | |
2383 | ||
2384 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2385 | may wrap in signed/pointer type are folded, as long as the value of | |
2386 | the chrec is preserved. | |
2387 | ||
2388 | SIZE_EXPR is used for computing the size of the expression to be | |
2389 | instantiated, and to stop if it exceeds some limit. */ | |
2390 | ||
2391 | static tree | |
2392 | instantiate_array_ref (basic_block instantiate_below, | |
8b679c9b RB |
2393 | struct loop *evolution_loop, struct loop *inner_loop, |
2394 | tree chrec, | |
fdd43ac4 | 2395 | bool fold_conversions, instantiate_cache_type &cache, |
4a8fb1a1 | 2396 | int size_expr) |
dbc08079 SP |
2397 | { |
2398 | tree res; | |
2399 | tree index = TREE_OPERAND (chrec, 1); | |
8b679c9b RB |
2400 | tree op1 = instantiate_scev_r (instantiate_below, evolution_loop, |
2401 | inner_loop, index, | |
dbc08079 SP |
2402 | fold_conversions, cache, size_expr); |
2403 | ||
2404 | if (op1 == chrec_dont_know) | |
2405 | return chrec_dont_know; | |
2406 | ||
2407 | if (chrec && op1 == index) | |
2408 | return chrec; | |
2409 | ||
2410 | res = unshare_expr (chrec); | |
2411 | TREE_OPERAND (res, 1) = op1; | |
2412 | return res; | |
2413 | } | |
2414 | ||
a213b219 | 2415 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
9c382ce9 SP |
2416 | and EVOLUTION_LOOP, that were left under a symbolic form. |
2417 | ||
2418 | "CHREC" that stands for a convert expression "(TYPE) OP" is to be | |
2419 | instantiated. | |
2420 | ||
2421 | CACHE is the cache of already instantiated values. | |
2422 | ||
2423 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2424 | may wrap in signed/pointer type are folded, as long as the value of | |
2425 | the chrec is preserved. | |
2426 | ||
2427 | SIZE_EXPR is used for computing the size of the expression to be | |
2428 | instantiated, and to stop if it exceeds some limit. */ | |
2429 | ||
2430 | static tree | |
2431 | instantiate_scev_convert (basic_block instantiate_below, | |
8b679c9b RB |
2432 | struct loop *evolution_loop, struct loop *inner_loop, |
2433 | tree chrec, | |
9c382ce9 | 2434 | tree type, tree op, |
4a8fb1a1 | 2435 | bool fold_conversions, |
fdd43ac4 | 2436 | instantiate_cache_type &cache, int size_expr) |
9c382ce9 | 2437 | { |
8b679c9b RB |
2438 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
2439 | inner_loop, op, | |
9c382ce9 SP |
2440 | fold_conversions, cache, size_expr); |
2441 | ||
2442 | if (op0 == chrec_dont_know) | |
2443 | return chrec_dont_know; | |
2444 | ||
2445 | if (fold_conversions) | |
2446 | { | |
2447 | tree tmp = chrec_convert_aggressive (type, op0); | |
2448 | if (tmp) | |
2449 | return tmp; | |
2450 | } | |
2451 | ||
2452 | if (chrec && op0 == op) | |
2453 | return chrec; | |
2454 | ||
2455 | /* If we used chrec_convert_aggressive, we can no longer assume that | |
2456 | signed chrecs do not overflow, as chrec_convert does, so avoid | |
2457 | calling it in that case. */ | |
2458 | if (fold_conversions) | |
2459 | return fold_convert (type, op0); | |
2460 | ||
2461 | return chrec_convert (type, op0, NULL); | |
2462 | } | |
2463 | ||
7ec0665d SP |
2464 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2465 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
2466 | ||
4b9d48a1 | 2467 | CHREC is a BIT_NOT_EXPR or a NEGATE_EXPR expression to be instantiated. |
7ec0665d | 2468 | Handle ~X as -1 - X. |
4b9d48a1 | 2469 | Handle -X as -1 * X. |
7ec0665d SP |
2470 | |
2471 | CACHE is the cache of already instantiated values. | |
2472 | ||
2473 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2474 | may wrap in signed/pointer type are folded, as long as the value of | |
2475 | the chrec is preserved. | |
2476 | ||
2477 | SIZE_EXPR is used for computing the size of the expression to be | |
2478 | instantiated, and to stop if it exceeds some limit. */ | |
2479 | ||
2480 | static tree | |
4b9d48a1 | 2481 | instantiate_scev_not (basic_block instantiate_below, |
8b679c9b RB |
2482 | struct loop *evolution_loop, struct loop *inner_loop, |
2483 | tree chrec, | |
20179b0d | 2484 | enum tree_code code, tree type, tree op, |
fdd43ac4 | 2485 | bool fold_conversions, instantiate_cache_type &cache, |
4a8fb1a1 | 2486 | int size_expr) |
7ec0665d | 2487 | { |
8b679c9b RB |
2488 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
2489 | inner_loop, op, | |
7ec0665d | 2490 | fold_conversions, cache, size_expr); |
20179b0d | 2491 | |
7ec0665d SP |
2492 | if (op0 == chrec_dont_know) |
2493 | return chrec_dont_know; | |
2494 | ||
20179b0d | 2495 | if (op != op0) |
7ec0665d SP |
2496 | { |
2497 | op0 = chrec_convert (type, op0, NULL); | |
4b9d48a1 | 2498 | |
20179b0d | 2499 | switch (code) |
4b9d48a1 SP |
2500 | { |
2501 | case BIT_NOT_EXPR: | |
2502 | return chrec_fold_minus | |
2503 | (type, fold_convert (type, integer_minus_one_node), op0); | |
2504 | ||
2505 | case NEGATE_EXPR: | |
2506 | return chrec_fold_multiply | |
2507 | (type, fold_convert (type, integer_minus_one_node), op0); | |
2508 | ||
2509 | default: | |
2510 | gcc_unreachable (); | |
2511 | } | |
7ec0665d | 2512 | } |
4b9d48a1 | 2513 | |
20179b0d | 2514 | return chrec ? chrec : fold_build1 (code, type, op0); |
7ec0665d SP |
2515 | } |
2516 | ||
d814176c SP |
2517 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2518 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
2519 | ||
2520 | CHREC is an expression with 3 operands to be instantiated. | |
2521 | ||
2522 | CACHE is the cache of already instantiated values. | |
2523 | ||
2524 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2525 | may wrap in signed/pointer type are folded, as long as the value of | |
2526 | the chrec is preserved. | |
2527 | ||
2528 | SIZE_EXPR is used for computing the size of the expression to be | |
2529 | instantiated, and to stop if it exceeds some limit. */ | |
2530 | ||
2531 | static tree | |
2532 | instantiate_scev_3 (basic_block instantiate_below, | |
8b679c9b RB |
2533 | struct loop *evolution_loop, struct loop *inner_loop, |
2534 | tree chrec, | |
fdd43ac4 | 2535 | bool fold_conversions, instantiate_cache_type &cache, |
4a8fb1a1 | 2536 | int size_expr) |
d814176c SP |
2537 | { |
2538 | tree op1, op2; | |
9e5dc77f | 2539 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
8b679c9b | 2540 | inner_loop, TREE_OPERAND (chrec, 0), |
d814176c SP |
2541 | fold_conversions, cache, size_expr); |
2542 | if (op0 == chrec_dont_know) | |
2543 | return chrec_dont_know; | |
2544 | ||
9e5dc77f | 2545 | op1 = instantiate_scev_r (instantiate_below, evolution_loop, |
8b679c9b | 2546 | inner_loop, TREE_OPERAND (chrec, 1), |
d814176c SP |
2547 | fold_conversions, cache, size_expr); |
2548 | if (op1 == chrec_dont_know) | |
2549 | return chrec_dont_know; | |
2550 | ||
9e5dc77f | 2551 | op2 = instantiate_scev_r (instantiate_below, evolution_loop, |
8b679c9b | 2552 | inner_loop, TREE_OPERAND (chrec, 2), |
d814176c SP |
2553 | fold_conversions, cache, size_expr); |
2554 | if (op2 == chrec_dont_know) | |
2555 | return chrec_dont_know; | |
2556 | ||
2557 | if (op0 == TREE_OPERAND (chrec, 0) | |
2558 | && op1 == TREE_OPERAND (chrec, 1) | |
2559 | && op2 == TREE_OPERAND (chrec, 2)) | |
2560 | return chrec; | |
2561 | ||
2562 | return fold_build3 (TREE_CODE (chrec), | |
2563 | TREE_TYPE (chrec), op0, op1, op2); | |
2564 | } | |
2565 | ||
9c382ce9 SP |
2566 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW |
2567 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
5b78fc3e | 2568 | |
9e5dc77f SP |
2569 | CHREC is an expression with 2 operands to be instantiated. |
2570 | ||
2571 | CACHE is the cache of already instantiated values. | |
2572 | ||
2573 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2574 | may wrap in signed/pointer type are folded, as long as the value of | |
2575 | the chrec is preserved. | |
2576 | ||
2577 | SIZE_EXPR is used for computing the size of the expression to be | |
2578 | instantiated, and to stop if it exceeds some limit. */ | |
2579 | ||
2580 | static tree | |
2581 | instantiate_scev_2 (basic_block instantiate_below, | |
8b679c9b RB |
2582 | struct loop *evolution_loop, struct loop *inner_loop, |
2583 | tree chrec, | |
fdd43ac4 | 2584 | bool fold_conversions, instantiate_cache_type &cache, |
4a8fb1a1 | 2585 | int size_expr) |
9e5dc77f SP |
2586 | { |
2587 | tree op1; | |
2588 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, | |
8b679c9b | 2589 | inner_loop, TREE_OPERAND (chrec, 0), |
9e5dc77f SP |
2590 | fold_conversions, cache, size_expr); |
2591 | if (op0 == chrec_dont_know) | |
2592 | return chrec_dont_know; | |
2593 | ||
2594 | op1 = instantiate_scev_r (instantiate_below, evolution_loop, | |
8b679c9b | 2595 | inner_loop, TREE_OPERAND (chrec, 1), |
9e5dc77f SP |
2596 | fold_conversions, cache, size_expr); |
2597 | if (op1 == chrec_dont_know) | |
2598 | return chrec_dont_know; | |
2599 | ||
2600 | if (op0 == TREE_OPERAND (chrec, 0) | |
2601 | && op1 == TREE_OPERAND (chrec, 1)) | |
2602 | return chrec; | |
2603 | ||
2604 | return fold_build2 (TREE_CODE (chrec), TREE_TYPE (chrec), op0, op1); | |
2605 | } | |
2606 | ||
2607 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW | |
2608 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
2609 | ||
2610 | CHREC is an expression with 2 operands to be instantiated. | |
5b78fc3e JS |
2611 | |
2612 | CACHE is the cache of already instantiated values. | |
2613 | ||
2614 | FOLD_CONVERSIONS should be set to true when the conversions that | |
2615 | may wrap in signed/pointer type are folded, as long as the value of | |
2616 | the chrec is preserved. | |
2617 | ||
3f227a8c JS |
2618 | SIZE_EXPR is used for computing the size of the expression to be |
2619 | instantiated, and to stop if it exceeds some limit. */ | |
9c382ce9 | 2620 | |
9baba81b | 2621 | static tree |
a213b219 | 2622 | instantiate_scev_1 (basic_block instantiate_below, |
8b679c9b RB |
2623 | struct loop *evolution_loop, struct loop *inner_loop, |
2624 | tree chrec, | |
fdd43ac4 | 2625 | bool fold_conversions, instantiate_cache_type &cache, |
4a8fb1a1 | 2626 | int size_expr) |
9baba81b | 2627 | { |
9e5dc77f | 2628 | tree op0 = instantiate_scev_r (instantiate_below, evolution_loop, |
8b679c9b | 2629 | inner_loop, TREE_OPERAND (chrec, 0), |
9e5dc77f SP |
2630 | fold_conversions, cache, size_expr); |
2631 | ||
2632 | if (op0 == chrec_dont_know) | |
2633 | return chrec_dont_know; | |
2634 | ||
2635 | if (op0 == TREE_OPERAND (chrec, 0)) | |
2636 | return chrec; | |
2637 | ||
2638 | return fold_build1 (TREE_CODE (chrec), TREE_TYPE (chrec), op0); | |
2639 | } | |
2640 | ||
2641 | /* Analyze all the parameters of the chrec, between INSTANTIATE_BELOW | |
2642 | and EVOLUTION_LOOP, that were left under a symbolic form. | |
2643 | ||
2644 | CHREC is the scalar evolution to instantiate. | |
2645 | ||
2646 | CACHE is the cache of already instantiated values. | |
2282a0e6 | 2647 | |
9e5dc77f SP |
2648 | FOLD_CONVERSIONS should be set to true when the conversions that |
2649 | may wrap in signed/pointer type are folded, as long as the value of | |
2650 | the chrec is preserved. | |
2651 | ||
2652 | SIZE_EXPR is used for computing the size of the expression to be | |
2653 | instantiated, and to stop if it exceeds some limit. */ | |
2654 | ||
2655 | static tree | |
2656 | instantiate_scev_r (basic_block instantiate_below, | |
8b679c9b RB |
2657 | struct loop *evolution_loop, struct loop *inner_loop, |
2658 | tree chrec, | |
fdd43ac4 | 2659 | bool fold_conversions, instantiate_cache_type &cache, |
4a8fb1a1 | 2660 | int size_expr) |
9e5dc77f | 2661 | { |
47ae9e4c SP |
2662 | /* Give up if the expression is larger than the MAX that we allow. */ |
2663 | if (size_expr++ > PARAM_VALUE (PARAM_SCEV_MAX_EXPR_SIZE)) | |
2664 | return chrec_dont_know; | |
2665 | ||
81fada9a JJ |
2666 | if (chrec == NULL_TREE |
2667 | || automatically_generated_chrec_p (chrec) | |
d7770457 | 2668 | || is_gimple_min_invariant (chrec)) |
9baba81b SP |
2669 | return chrec; |
2670 | ||
2671 | switch (TREE_CODE (chrec)) | |
2672 | { | |
2673 | case SSA_NAME: | |
8b679c9b RB |
2674 | return instantiate_scev_name (instantiate_below, evolution_loop, |
2675 | inner_loop, chrec, | |
320f5a78 | 2676 | fold_conversions, cache, size_expr); |
9baba81b SP |
2677 | |
2678 | case POLYNOMIAL_CHREC: | |
8b679c9b RB |
2679 | return instantiate_scev_poly (instantiate_below, evolution_loop, |
2680 | inner_loop, chrec, | |
ec6636eb | 2681 | fold_conversions, cache, size_expr); |
9baba81b | 2682 | |
5be014d5 | 2683 | case POINTER_PLUS_EXPR: |
9baba81b | 2684 | case PLUS_EXPR: |
9baba81b | 2685 | case MINUS_EXPR: |
9baba81b | 2686 | case MULT_EXPR: |
8b679c9b RB |
2687 | return instantiate_scev_binary (instantiate_below, evolution_loop, |
2688 | inner_loop, chrec, | |
ffa34f4b SP |
2689 | TREE_CODE (chrec), chrec_type (chrec), |
2690 | TREE_OPERAND (chrec, 0), | |
2691 | TREE_OPERAND (chrec, 1), | |
15fda317 | 2692 | fold_conversions, cache, size_expr); |
9baba81b | 2693 | |
1043771b | 2694 | CASE_CONVERT: |
8b679c9b RB |
2695 | return instantiate_scev_convert (instantiate_below, evolution_loop, |
2696 | inner_loop, chrec, | |
9c382ce9 SP |
2697 | TREE_TYPE (chrec), TREE_OPERAND (chrec, 0), |
2698 | fold_conversions, cache, size_expr); | |
9baba81b | 2699 | |
4b9d48a1 | 2700 | case NEGATE_EXPR: |
418df9d7 | 2701 | case BIT_NOT_EXPR: |
8b679c9b RB |
2702 | return instantiate_scev_not (instantiate_below, evolution_loop, |
2703 | inner_loop, chrec, | |
20179b0d SP |
2704 | TREE_CODE (chrec), TREE_TYPE (chrec), |
2705 | TREE_OPERAND (chrec, 0), | |
4b9d48a1 | 2706 | fold_conversions, cache, size_expr); |
418df9d7 | 2707 | |
4c7d6755 | 2708 | case ADDR_EXPR: |
9baba81b SP |
2709 | case SCEV_NOT_KNOWN: |
2710 | return chrec_dont_know; | |
2711 | ||
2712 | case SCEV_KNOWN: | |
2713 | return chrec_known; | |
15fda317 | 2714 | |
dbc08079 | 2715 | case ARRAY_REF: |
8b679c9b RB |
2716 | return instantiate_array_ref (instantiate_below, evolution_loop, |
2717 | inner_loop, chrec, | |
dbc08079 SP |
2718 | fold_conversions, cache, size_expr); |
2719 | ||
9baba81b SP |
2720 | default: |
2721 | break; | |
2722 | } | |
2723 | ||
0dfb0dc6 SP |
2724 | if (VL_EXP_CLASS_P (chrec)) |
2725 | return chrec_dont_know; | |
2726 | ||
9baba81b SP |
2727 | switch (TREE_CODE_LENGTH (TREE_CODE (chrec))) |
2728 | { | |
2729 | case 3: | |
8b679c9b RB |
2730 | return instantiate_scev_3 (instantiate_below, evolution_loop, |
2731 | inner_loop, chrec, | |
d814176c | 2732 | fold_conversions, cache, size_expr); |
9baba81b SP |
2733 | |
2734 | case 2: | |
8b679c9b RB |
2735 | return instantiate_scev_2 (instantiate_below, evolution_loop, |
2736 | inner_loop, chrec, | |
9e5dc77f | 2737 | fold_conversions, cache, size_expr); |
7ec0665d | 2738 | |
9baba81b | 2739 | case 1: |
8b679c9b RB |
2740 | return instantiate_scev_1 (instantiate_below, evolution_loop, |
2741 | inner_loop, chrec, | |
9e5dc77f | 2742 | fold_conversions, cache, size_expr); |
9baba81b SP |
2743 | |
2744 | case 0: | |
2745 | return chrec; | |
2746 | ||
2747 | default: | |
2748 | break; | |
2749 | } | |
2750 | ||
2751 | /* Too complicated to handle. */ | |
2752 | return chrec_dont_know; | |
2753 | } | |
e9eb809d ZD |
2754 | |
2755 | /* Analyze all the parameters of the chrec that were left under a | |
a213b219 SP |
2756 | symbolic form. INSTANTIATE_BELOW is the basic block that stops the |
2757 | recursive instantiation of parameters: a parameter is a variable | |
2758 | that is defined in a basic block that dominates INSTANTIATE_BELOW or | |
2759 | a function parameter. */ | |
e9eb809d ZD |
2760 | |
2761 | tree | |
a213b219 | 2762 | instantiate_scev (basic_block instantiate_below, struct loop *evolution_loop, |
3f227a8c | 2763 | tree chrec) |
e9eb809d | 2764 | { |
9baba81b | 2765 | tree res; |
fdd43ac4 | 2766 | instantiate_cache_type cache; |
9baba81b | 2767 | |
dfedbe40 | 2768 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b | 2769 | { |
3f227a8c | 2770 | fprintf (dump_file, "(instantiate_scev \n"); |
a213b219 | 2771 | fprintf (dump_file, " (instantiate_below = %d)\n", instantiate_below->index); |
3f227a8c | 2772 | fprintf (dump_file, " (evolution_loop = %d)\n", evolution_loop->num); |
9baba81b SP |
2773 | fprintf (dump_file, " (chrec = "); |
2774 | print_generic_expr (dump_file, chrec, 0); | |
2775 | fprintf (dump_file, ")\n"); | |
2776 | } | |
b8698a0f | 2777 | |
8b679c9b RB |
2778 | res = instantiate_scev_r (instantiate_below, evolution_loop, |
2779 | NULL, chrec, false, cache, 0); | |
9baba81b | 2780 | |
dfedbe40 | 2781 | if (dump_file && (dump_flags & TDF_SCEV)) |
9baba81b SP |
2782 | { |
2783 | fprintf (dump_file, " (res = "); | |
2784 | print_generic_expr (dump_file, res, 0); | |
2785 | fprintf (dump_file, "))\n"); | |
2786 | } | |
eb0bc7af | 2787 | |
9baba81b SP |
2788 | return res; |
2789 | } | |
2790 | ||
2791 | /* Similar to instantiate_parameters, but does not introduce the | |
2282a0e6 ZD |
2792 | evolutions in outer loops for LOOP invariants in CHREC, and does not |
2793 | care about causing overflows, as long as they do not affect value | |
2794 | of an expression. */ | |
9baba81b | 2795 | |
3cb960c7 | 2796 | tree |
9baba81b SP |
2797 | resolve_mixers (struct loop *loop, tree chrec) |
2798 | { | |
fdd43ac4 | 2799 | instantiate_cache_type cache; |
8b679c9b RB |
2800 | tree ret = instantiate_scev_r (block_before_loop (loop), loop, NULL, |
2801 | chrec, true, cache, 0); | |
eb0bc7af | 2802 | return ret; |
9baba81b SP |
2803 | } |
2804 | ||
b8698a0f | 2805 | /* Entry point for the analysis of the number of iterations pass. |
9baba81b SP |
2806 | This function tries to safely approximate the number of iterations |
2807 | the loop will run. When this property is not decidable at compile | |
0a74c758 SP |
2808 | time, the result is chrec_dont_know. Otherwise the result is a |
2809 | scalar or a symbolic parameter. When the number of iterations may | |
2810 | be equal to zero and the property cannot be determined at compile | |
2811 | time, the result is a COND_EXPR that represents in a symbolic form | |
2812 | the conditions under which the number of iterations is not zero. | |
b8698a0f | 2813 | |
9baba81b | 2814 | Example of analysis: suppose that the loop has an exit condition: |
b8698a0f | 2815 | |
9baba81b | 2816 | "if (b > 49) goto end_loop;" |
b8698a0f | 2817 | |
9baba81b SP |
2818 | and that in a previous analysis we have determined that the |
2819 | variable 'b' has an evolution function: | |
b8698a0f L |
2820 | |
2821 | "EF = {23, +, 5}_2". | |
2822 | ||
9baba81b SP |
2823 | When we evaluate the function at the point 5, i.e. the value of the |
2824 | variable 'b' after 5 iterations in the loop, we have EF (5) = 48, | |
2825 | and EF (6) = 53. In this case the value of 'b' on exit is '53' and | |
2826 | the loop body has been executed 6 times. */ | |
2827 | ||
b8698a0f | 2828 | tree |
a14865db | 2829 | number_of_latch_executions (struct loop *loop) |
9baba81b | 2830 | { |
9baba81b SP |
2831 | edge exit; |
2832 | struct tree_niter_desc niter_desc; | |
0a74c758 SP |
2833 | tree may_be_zero; |
2834 | tree res; | |
9baba81b | 2835 | |
0a74c758 | 2836 | /* Determine whether the number of iterations in loop has already |
9baba81b SP |
2837 | been computed. */ |
2838 | res = loop->nb_iterations; | |
2839 | if (res) | |
2840 | return res; | |
0a74c758 SP |
2841 | |
2842 | may_be_zero = NULL_TREE; | |
9baba81b | 2843 | |
dfedbe40 | 2844 | if (dump_file && (dump_flags & TDF_SCEV)) |
0a74c758 | 2845 | fprintf (dump_file, "(number_of_iterations_in_loop = \n"); |
b8698a0f | 2846 | |
0a74c758 | 2847 | res = chrec_dont_know; |
ac8f6c69 | 2848 | exit = single_exit (loop); |
9baba81b | 2849 | |
0a74c758 SP |
2850 | if (exit && number_of_iterations_exit (loop, exit, &niter_desc, false)) |
2851 | { | |
2852 | may_be_zero = niter_desc.may_be_zero; | |
2853 | res = niter_desc.niter; | |
2854 | } | |
2855 | ||
2856 | if (res == chrec_dont_know | |
2857 | || !may_be_zero | |
2858 | || integer_zerop (may_be_zero)) | |
2859 | ; | |
2860 | else if (integer_nonzerop (may_be_zero)) | |
2861 | res = build_int_cst (TREE_TYPE (res), 0); | |
9baba81b | 2862 | |
0a74c758 SP |
2863 | else if (COMPARISON_CLASS_P (may_be_zero)) |
2864 | res = fold_build3 (COND_EXPR, TREE_TYPE (res), may_be_zero, | |
2865 | build_int_cst (TREE_TYPE (res), 0), res); | |
9baba81b SP |
2866 | else |
2867 | res = chrec_dont_know; | |
2868 | ||
dfedbe40 | 2869 | if (dump_file && (dump_flags & TDF_SCEV)) |
0a74c758 SP |
2870 | { |
2871 | fprintf (dump_file, " (set_nb_iterations_in_loop = "); | |
2872 | print_generic_expr (dump_file, res, 0); | |
2873 | fprintf (dump_file, "))\n"); | |
2874 | } | |
2875 | ||
2876 | loop->nb_iterations = res; | |
2877 | return res; | |
9baba81b SP |
2878 | } |
2879 | ||
a14865db ZD |
2880 | /* Returns the number of executions of the exit condition of LOOP, |
2881 | i.e., the number by one higher than number_of_latch_executions. | |
fa10beec | 2882 | Note that unlike number_of_latch_executions, this number does |
a14865db ZD |
2883 | not necessarily fit in the unsigned variant of the type of |
2884 | the control variable -- if the number of iterations is a constant, | |
2885 | we return chrec_dont_know if adding one to number_of_latch_executions | |
2886 | overflows; however, in case the number of iterations is symbolic | |
2887 | expression, the caller is responsible for dealing with this | |
2888 | the possible overflow. */ | |
2889 | ||
b8698a0f | 2890 | tree |
a14865db ZD |
2891 | number_of_exit_cond_executions (struct loop *loop) |
2892 | { | |
2893 | tree ret = number_of_latch_executions (loop); | |
2894 | tree type = chrec_type (ret); | |
2895 | ||
2896 | if (chrec_contains_undetermined (ret)) | |
2897 | return ret; | |
2898 | ||
2899 | ret = chrec_fold_plus (type, ret, build_int_cst (type, 1)); | |
2900 | if (TREE_CODE (ret) == INTEGER_CST | |
2901 | && TREE_OVERFLOW (ret)) | |
2902 | return chrec_dont_know; | |
2903 | ||
2904 | return ret; | |
2905 | } | |
2906 | ||
9baba81b SP |
2907 | \f |
2908 | ||
2909 | /* Counters for the stats. */ | |
2910 | ||
b8698a0f | 2911 | struct chrec_stats |
9baba81b SP |
2912 | { |
2913 | unsigned nb_chrecs; | |
2914 | unsigned nb_affine; | |
2915 | unsigned nb_affine_multivar; | |
2916 | unsigned nb_higher_poly; | |
2917 | unsigned nb_chrec_dont_know; | |
2918 | unsigned nb_undetermined; | |
2919 | }; | |
2920 | ||
2921 | /* Reset the counters. */ | |
2922 | ||
2923 | static inline void | |
2924 | reset_chrecs_counters (struct chrec_stats *stats) | |
2925 | { | |
2926 | stats->nb_chrecs = 0; | |
2927 | stats->nb_affine = 0; | |
2928 | stats->nb_affine_multivar = 0; | |
2929 | stats->nb_higher_poly = 0; | |
2930 | stats->nb_chrec_dont_know = 0; | |
2931 | stats->nb_undetermined = 0; | |
2932 | } | |
2933 | ||
2934 | /* Dump the contents of a CHREC_STATS structure. */ | |
2935 | ||
2936 | static void | |
2937 | dump_chrecs_stats (FILE *file, struct chrec_stats *stats) | |
2938 | { | |
2939 | fprintf (file, "\n(\n"); | |
2940 | fprintf (file, "-----------------------------------------\n"); | |
2941 | fprintf (file, "%d\taffine univariate chrecs\n", stats->nb_affine); | |
2942 | fprintf (file, "%d\taffine multivariate chrecs\n", stats->nb_affine_multivar); | |
b8698a0f | 2943 | fprintf (file, "%d\tdegree greater than 2 polynomials\n", |
9baba81b SP |
2944 | stats->nb_higher_poly); |
2945 | fprintf (file, "%d\tchrec_dont_know chrecs\n", stats->nb_chrec_dont_know); | |
2946 | fprintf (file, "-----------------------------------------\n"); | |
2947 | fprintf (file, "%d\ttotal chrecs\n", stats->nb_chrecs); | |
b8698a0f | 2948 | fprintf (file, "%d\twith undetermined coefficients\n", |
9baba81b SP |
2949 | stats->nb_undetermined); |
2950 | fprintf (file, "-----------------------------------------\n"); | |
b8698a0f | 2951 | fprintf (file, "%d\tchrecs in the scev database\n", |
9baba81b SP |
2952 | (int) htab_elements (scalar_evolution_info)); |
2953 | fprintf (file, "%d\tsets in the scev database\n", nb_set_scev); | |
2954 | fprintf (file, "%d\tgets in the scev database\n", nb_get_scev); | |
2955 | fprintf (file, "-----------------------------------------\n"); | |
2956 | fprintf (file, ")\n\n"); | |
2957 | } | |
2958 | ||
2959 | /* Gather statistics about CHREC. */ | |
2960 | ||
2961 | static void | |
2962 | gather_chrec_stats (tree chrec, struct chrec_stats *stats) | |
2963 | { | |
2964 | if (dump_file && (dump_flags & TDF_STATS)) | |
2965 | { | |
2966 | fprintf (dump_file, "(classify_chrec "); | |
2967 | print_generic_expr (dump_file, chrec, 0); | |
2968 | fprintf (dump_file, "\n"); | |
2969 | } | |
b8698a0f | 2970 | |
9baba81b | 2971 | stats->nb_chrecs++; |
b8698a0f | 2972 | |
9baba81b SP |
2973 | if (chrec == NULL_TREE) |
2974 | { | |
2975 | stats->nb_undetermined++; | |
2976 | return; | |
2977 | } | |
b8698a0f | 2978 | |
9baba81b SP |
2979 | switch (TREE_CODE (chrec)) |
2980 | { | |
2981 | case POLYNOMIAL_CHREC: | |
2982 | if (evolution_function_is_affine_p (chrec)) | |
2983 | { | |
2984 | if (dump_file && (dump_flags & TDF_STATS)) | |
2985 | fprintf (dump_file, " affine_univariate\n"); | |
2986 | stats->nb_affine++; | |
2987 | } | |
a50411de | 2988 | else if (evolution_function_is_affine_multivariate_p (chrec, 0)) |
9baba81b SP |
2989 | { |
2990 | if (dump_file && (dump_flags & TDF_STATS)) | |
2991 | fprintf (dump_file, " affine_multivariate\n"); | |
2992 | stats->nb_affine_multivar++; | |
2993 | } | |
2994 | else | |
2995 | { | |
2996 | if (dump_file && (dump_flags & TDF_STATS)) | |
2997 | fprintf (dump_file, " higher_degree_polynomial\n"); | |
2998 | stats->nb_higher_poly++; | |
2999 | } | |
b8698a0f | 3000 | |
9baba81b SP |
3001 | break; |
3002 | ||
3003 | default: | |
3004 | break; | |
3005 | } | |
b8698a0f | 3006 | |
9baba81b SP |
3007 | if (chrec_contains_undetermined (chrec)) |
3008 | { | |
3009 | if (dump_file && (dump_flags & TDF_STATS)) | |
3010 | fprintf (dump_file, " undetermined\n"); | |
3011 | stats->nb_undetermined++; | |
3012 | } | |
b8698a0f | 3013 | |
9baba81b SP |
3014 | if (dump_file && (dump_flags & TDF_STATS)) |
3015 | fprintf (dump_file, ")\n"); | |
3016 | } | |
3017 | ||
9baba81b SP |
3018 | /* Callback for htab_traverse, gathers information on chrecs in the |
3019 | hashtable. */ | |
3020 | ||
3021 | static int | |
3022 | gather_stats_on_scev_database_1 (void **slot, void *stats) | |
3023 | { | |
cceb1885 | 3024 | struct scev_info_str *entry = (struct scev_info_str *) *slot; |
9baba81b | 3025 | |
cceb1885 | 3026 | gather_chrec_stats (entry->chrec, (struct chrec_stats *) stats); |
9baba81b SP |
3027 | |
3028 | return 1; | |
3029 | } | |
3030 | ||
3031 | /* Classify the chrecs of the whole database. */ | |
3032 | ||
b8698a0f | 3033 | void |
9baba81b SP |
3034 | gather_stats_on_scev_database (void) |
3035 | { | |
3036 | struct chrec_stats stats; | |
b8698a0f | 3037 | |
9baba81b SP |
3038 | if (!dump_file) |
3039 | return; | |
b8698a0f | 3040 | |
9baba81b | 3041 | reset_chrecs_counters (&stats); |
b8698a0f | 3042 | |
9baba81b SP |
3043 | htab_traverse (scalar_evolution_info, gather_stats_on_scev_database_1, |
3044 | &stats); | |
3045 | ||
3046 | dump_chrecs_stats (dump_file, &stats); | |
3047 | } | |
3048 | ||
3049 | \f | |
3050 | ||
3051 | /* Initializer. */ | |
3052 | ||
3053 | static void | |
3054 | initialize_scalar_evolutions_analyzer (void) | |
3055 | { | |
3056 | /* The elements below are unique. */ | |
3057 | if (chrec_dont_know == NULL_TREE) | |
3058 | { | |
3059 | chrec_not_analyzed_yet = NULL_TREE; | |
3060 | chrec_dont_know = make_node (SCEV_NOT_KNOWN); | |
3061 | chrec_known = make_node (SCEV_KNOWN); | |
d5ab5675 ZD |
3062 | TREE_TYPE (chrec_dont_know) = void_type_node; |
3063 | TREE_TYPE (chrec_known) = void_type_node; | |
9baba81b SP |
3064 | } |
3065 | } | |
3066 | ||
3067 | /* Initialize the analysis of scalar evolutions for LOOPS. */ | |
3068 | ||
3069 | void | |
d73be268 | 3070 | scev_initialize (void) |
9baba81b | 3071 | { |
42fd6772 ZD |
3072 | loop_iterator li; |
3073 | struct loop *loop; | |
9baba81b | 3074 | |
a9429e29 LB |
3075 | |
3076 | scalar_evolution_info = htab_create_ggc (100, hash_scev_info, eq_scev_info, | |
3077 | del_scev_info); | |
b8698a0f | 3078 | |
9baba81b SP |
3079 | initialize_scalar_evolutions_analyzer (); |
3080 | ||
42fd6772 ZD |
3081 | FOR_EACH_LOOP (li, loop, 0) |
3082 | { | |
3083 | loop->nb_iterations = NULL_TREE; | |
3084 | } | |
9baba81b SP |
3085 | } |
3086 | ||
e3a8f1fa JH |
3087 | /* Return true if SCEV is initialized. */ |
3088 | ||
3089 | bool | |
3090 | scev_initialized_p (void) | |
3091 | { | |
3092 | return scalar_evolution_info != NULL; | |
3093 | } | |
3094 | ||
a7bf45de SP |
3095 | /* Cleans up the information cached by the scalar evolutions analysis |
3096 | in the hash table. */ | |
3097 | ||
3098 | void | |
3099 | scev_reset_htab (void) | |
3100 | { | |
3101 | if (!scalar_evolution_info) | |
3102 | return; | |
3103 | ||
3104 | htab_empty (scalar_evolution_info); | |
3105 | } | |
3106 | ||
3107 | /* Cleans up the information cached by the scalar evolutions analysis | |
3108 | in the hash table and in the loop->nb_iterations. */ | |
9baba81b SP |
3109 | |
3110 | void | |
3111 | scev_reset (void) | |
3112 | { | |
42fd6772 | 3113 | loop_iterator li; |
9baba81b SP |
3114 | struct loop *loop; |
3115 | ||
a7bf45de SP |
3116 | scev_reset_htab (); |
3117 | ||
3118 | if (!current_loops) | |
9baba81b SP |
3119 | return; |
3120 | ||
42fd6772 | 3121 | FOR_EACH_LOOP (li, loop, 0) |
9baba81b | 3122 | { |
42fd6772 | 3123 | loop->nb_iterations = NULL_TREE; |
9baba81b | 3124 | } |
e9eb809d ZD |
3125 | } |
3126 | ||
f017bf5e ZD |
3127 | /* Checks whether use of OP in USE_LOOP behaves as a simple affine iv with |
3128 | respect to WRTO_LOOP and returns its base and step in IV if possible | |
3129 | (see analyze_scalar_evolution_in_loop for more details on USE_LOOP | |
3130 | and WRTO_LOOP). If ALLOW_NONCONSTANT_STEP is true, we want step to be | |
3131 | invariant in LOOP. Otherwise we require it to be an integer constant. | |
b8698a0f | 3132 | |
f017bf5e ZD |
3133 | IV->no_overflow is set to true if we are sure the iv cannot overflow (e.g. |
3134 | because it is computed in signed arithmetics). Consequently, adding an | |
3135 | induction variable | |
b8698a0f | 3136 | |
f017bf5e ZD |
3137 | for (i = IV->base; ; i += IV->step) |
3138 | ||
3139 | is only safe if IV->no_overflow is false, or TYPE_OVERFLOW_UNDEFINED is | |
3140 | false for the type of the induction variable, or you can prove that i does | |
3141 | not wrap by some other argument. Otherwise, this might introduce undefined | |
3142 | behavior, and | |
b8698a0f | 3143 | |
f017bf5e ZD |
3144 | for (i = iv->base; ; i = (type) ((unsigned type) i + (unsigned type) iv->step)) |
3145 | ||
3146 | must be used instead. */ | |
e9eb809d ZD |
3147 | |
3148 | bool | |
f017bf5e ZD |
3149 | simple_iv (struct loop *wrto_loop, struct loop *use_loop, tree op, |
3150 | affine_iv *iv, bool allow_nonconstant_step) | |
e9eb809d | 3151 | { |
9baba81b | 3152 | tree type, ev; |
a6f778b2 | 3153 | bool folded_casts; |
9baba81b | 3154 | |
a6f778b2 ZD |
3155 | iv->base = NULL_TREE; |
3156 | iv->step = NULL_TREE; | |
3157 | iv->no_overflow = false; | |
9baba81b SP |
3158 | |
3159 | type = TREE_TYPE (op); | |
1ee0d660 EB |
3160 | if (!POINTER_TYPE_P (type) |
3161 | && !INTEGRAL_TYPE_P (type)) | |
9baba81b SP |
3162 | return false; |
3163 | ||
f017bf5e | 3164 | ev = analyze_scalar_evolution_in_loop (wrto_loop, use_loop, op, |
a6f778b2 | 3165 | &folded_casts); |
f017bf5e ZD |
3166 | if (chrec_contains_undetermined (ev) |
3167 | || chrec_contains_symbols_defined_in_loop (ev, wrto_loop->num)) | |
9baba81b SP |
3168 | return false; |
3169 | ||
f017bf5e | 3170 | if (tree_does_not_contain_chrecs (ev)) |
9baba81b | 3171 | { |
a6f778b2 | 3172 | iv->base = ev; |
6e42ce54 | 3173 | iv->step = build_int_cst (TREE_TYPE (ev), 0); |
a6f778b2 | 3174 | iv->no_overflow = true; |
9baba81b SP |
3175 | return true; |
3176 | } | |
3177 | ||
3178 | if (TREE_CODE (ev) != POLYNOMIAL_CHREC | |
f017bf5e | 3179 | || CHREC_VARIABLE (ev) != (unsigned) wrto_loop->num) |
9baba81b SP |
3180 | return false; |
3181 | ||
a6f778b2 | 3182 | iv->step = CHREC_RIGHT (ev); |
f017bf5e ZD |
3183 | if ((!allow_nonconstant_step && TREE_CODE (iv->step) != INTEGER_CST) |
3184 | || tree_contains_chrecs (iv->step, NULL)) | |
9baba81b | 3185 | return false; |
9be872b7 | 3186 | |
a6f778b2 | 3187 | iv->base = CHREC_LEFT (ev); |
f017bf5e | 3188 | if (tree_contains_chrecs (iv->base, NULL)) |
9baba81b SP |
3189 | return false; |
3190 | ||
eeef0e45 ILT |
3191 | iv->no_overflow = !folded_casts && TYPE_OVERFLOW_UNDEFINED (type); |
3192 | ||
9baba81b SP |
3193 | return true; |
3194 | } | |
3195 | ||
9baba81b SP |
3196 | /* Finalize the scalar evolution analysis. */ |
3197 | ||
3198 | void | |
3199 | scev_finalize (void) | |
3200 | { | |
d51157de ZD |
3201 | if (!scalar_evolution_info) |
3202 | return; | |
9baba81b | 3203 | htab_delete (scalar_evolution_info); |
c7b852c8 | 3204 | scalar_evolution_info = NULL; |
9baba81b SP |
3205 | } |
3206 | ||
771f882e ZD |
3207 | /* Returns true if the expression EXPR is considered to be too expensive |
3208 | for scev_const_prop. */ | |
3209 | ||
3210 | bool | |
3211 | expression_expensive_p (tree expr) | |
3212 | { | |
3213 | enum tree_code code; | |
3214 | ||
3215 | if (is_gimple_val (expr)) | |
3216 | return false; | |
3217 | ||
3218 | code = TREE_CODE (expr); | |
3219 | if (code == TRUNC_DIV_EXPR | |
3220 | || code == CEIL_DIV_EXPR | |
3221 | || code == FLOOR_DIV_EXPR | |
3222 | || code == ROUND_DIV_EXPR | |
3223 | || code == TRUNC_MOD_EXPR | |
3224 | || code == CEIL_MOD_EXPR | |
3225 | || code == FLOOR_MOD_EXPR | |
3226 | || code == ROUND_MOD_EXPR | |
3227 | || code == EXACT_DIV_EXPR) | |
3228 | { | |
3229 | /* Division by power of two is usually cheap, so we allow it. | |
3230 | Forbid anything else. */ | |
3231 | if (!integer_pow2p (TREE_OPERAND (expr, 1))) | |
3232 | return true; | |
3233 | } | |
3234 | ||
3235 | switch (TREE_CODE_CLASS (code)) | |
3236 | { | |
3237 | case tcc_binary: | |
3238 | case tcc_comparison: | |
3239 | if (expression_expensive_p (TREE_OPERAND (expr, 1))) | |
3240 | return true; | |
3241 | ||
3242 | /* Fallthru. */ | |
3243 | case tcc_unary: | |
3244 | return expression_expensive_p (TREE_OPERAND (expr, 0)); | |
3245 | ||
3246 | default: | |
3247 | return true; | |
3248 | } | |
3249 | } | |
3250 | ||
684aaf29 | 3251 | /* Replace ssa names for that scev can prove they are constant by the |
3ac01fde ZD |
3252 | appropriate constants. Also perform final value replacement in loops, |
3253 | in case the replacement expressions are cheap. | |
b8698a0f | 3254 | |
684aaf29 ZD |
3255 | We only consider SSA names defined by phi nodes; rest is left to the |
3256 | ordinary constant propagation pass. */ | |
3257 | ||
c2924966 | 3258 | unsigned int |
684aaf29 ZD |
3259 | scev_const_prop (void) |
3260 | { | |
3261 | basic_block bb; | |
726a989a RB |
3262 | tree name, type, ev; |
3263 | gimple phi, ass; | |
3ac01fde | 3264 | struct loop *loop, *ex_loop; |
684aaf29 | 3265 | bitmap ssa_names_to_remove = NULL; |
3ac01fde | 3266 | unsigned i; |
42fd6772 | 3267 | loop_iterator li; |
726a989a | 3268 | gimple_stmt_iterator psi; |
684aaf29 | 3269 | |
0fc822d0 | 3270 | if (number_of_loops (cfun) <= 1) |
c2924966 | 3271 | return 0; |
684aaf29 ZD |
3272 | |
3273 | FOR_EACH_BB (bb) | |
3274 | { | |
3275 | loop = bb->loop_father; | |
3276 | ||
726a989a | 3277 | for (psi = gsi_start_phis (bb); !gsi_end_p (psi); gsi_next (&psi)) |
684aaf29 | 3278 | { |
726a989a | 3279 | phi = gsi_stmt (psi); |
684aaf29 ZD |
3280 | name = PHI_RESULT (phi); |
3281 | ||
ea057359 | 3282 | if (virtual_operand_p (name)) |
684aaf29 ZD |
3283 | continue; |
3284 | ||
3285 | type = TREE_TYPE (name); | |
3286 | ||
3287 | if (!POINTER_TYPE_P (type) | |
3288 | && !INTEGRAL_TYPE_P (type)) | |
3289 | continue; | |
3290 | ||
3291 | ev = resolve_mixers (loop, analyze_scalar_evolution (loop, name)); | |
3292 | if (!is_gimple_min_invariant (ev) | |
3293 | || !may_propagate_copy (name, ev)) | |
3294 | continue; | |
3295 | ||
3296 | /* Replace the uses of the name. */ | |
18aed06a SP |
3297 | if (name != ev) |
3298 | replace_uses_by (name, ev); | |
684aaf29 ZD |
3299 | |
3300 | if (!ssa_names_to_remove) | |
3301 | ssa_names_to_remove = BITMAP_ALLOC (NULL); | |
3302 | bitmap_set_bit (ssa_names_to_remove, SSA_NAME_VERSION (name)); | |
3303 | } | |
3304 | } | |
3305 | ||
9b3b55a1 DN |
3306 | /* Remove the ssa names that were replaced by constants. We do not |
3307 | remove them directly in the previous cycle, since this | |
3308 | invalidates scev cache. */ | |
684aaf29 ZD |
3309 | if (ssa_names_to_remove) |
3310 | { | |
3311 | bitmap_iterator bi; | |
684aaf29 ZD |
3312 | |
3313 | EXECUTE_IF_SET_IN_BITMAP (ssa_names_to_remove, 0, i, bi) | |
3314 | { | |
726a989a | 3315 | gimple_stmt_iterator psi; |
684aaf29 ZD |
3316 | name = ssa_name (i); |
3317 | phi = SSA_NAME_DEF_STMT (name); | |
3318 | ||
726a989a RB |
3319 | gcc_assert (gimple_code (phi) == GIMPLE_PHI); |
3320 | psi = gsi_for_stmt (phi); | |
3321 | remove_phi_node (&psi, true); | |
684aaf29 ZD |
3322 | } |
3323 | ||
3324 | BITMAP_FREE (ssa_names_to_remove); | |
3325 | scev_reset (); | |
3326 | } | |
3ac01fde ZD |
3327 | |
3328 | /* Now the regular final value replacement. */ | |
42fd6772 | 3329 | FOR_EACH_LOOP (li, loop, LI_FROM_INNERMOST) |
3ac01fde ZD |
3330 | { |
3331 | edge exit; | |
726a989a RB |
3332 | tree def, rslt, niter; |
3333 | gimple_stmt_iterator bsi; | |
3ac01fde | 3334 | |
3ac01fde ZD |
3335 | /* If we do not know exact number of iterations of the loop, we cannot |
3336 | replace the final value. */ | |
ac8f6c69 | 3337 | exit = single_exit (loop); |
a6f778b2 ZD |
3338 | if (!exit) |
3339 | continue; | |
3340 | ||
a14865db | 3341 | niter = number_of_latch_executions (loop); |
b3ce5b6e | 3342 | if (niter == chrec_dont_know) |
3ac01fde | 3343 | continue; |
925196ed ZD |
3344 | |
3345 | /* Ensure that it is possible to insert new statements somewhere. */ | |
3346 | if (!single_pred_p (exit->dest)) | |
3347 | split_loop_exit_edge (exit); | |
726a989a | 3348 | bsi = gsi_after_labels (exit->dest); |
925196ed | 3349 | |
9ba025a2 ZD |
3350 | ex_loop = superloop_at_depth (loop, |
3351 | loop_depth (exit->dest->loop_father) + 1); | |
3ac01fde | 3352 | |
726a989a | 3353 | for (psi = gsi_start_phis (exit->dest); !gsi_end_p (psi); ) |
3ac01fde | 3354 | { |
726a989a | 3355 | phi = gsi_stmt (psi); |
925196ed | 3356 | rslt = PHI_RESULT (phi); |
3ac01fde | 3357 | def = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
ea057359 | 3358 | if (virtual_operand_p (def)) |
726a989a RB |
3359 | { |
3360 | gsi_next (&psi); | |
3361 | continue; | |
3362 | } | |
3ac01fde ZD |
3363 | |
3364 | if (!POINTER_TYPE_P (TREE_TYPE (def)) | |
3365 | && !INTEGRAL_TYPE_P (TREE_TYPE (def))) | |
726a989a RB |
3366 | { |
3367 | gsi_next (&psi); | |
3368 | continue; | |
3369 | } | |
3ac01fde | 3370 | |
a6f778b2 | 3371 | def = analyze_scalar_evolution_in_loop (ex_loop, loop, def, NULL); |
925196ed | 3372 | def = compute_overall_effect_of_inner_loop (ex_loop, def); |
3ac01fde | 3373 | if (!tree_does_not_contain_chrecs (def) |
e5db3515 ZD |
3374 | || chrec_contains_symbols_defined_in_loop (def, ex_loop->num) |
3375 | /* Moving the computation from the loop may prolong life range | |
3376 | of some ssa names, which may cause problems if they appear | |
3377 | on abnormal edges. */ | |
771f882e ZD |
3378 | || contains_abnormal_ssa_name_p (def) |
3379 | /* Do not emit expensive expressions. The rationale is that | |
3380 | when someone writes a code like | |
3381 | ||
3382 | while (n > 45) n -= 45; | |
3383 | ||
3384 | he probably knows that n is not large, and does not want it | |
3385 | to be turned into n %= 45. */ | |
3386 | || expression_expensive_p (def)) | |
726a989a | 3387 | { |
3571dde6 MP |
3388 | if (dump_file && (dump_flags & TDF_DETAILS)) |
3389 | { | |
3390 | fprintf (dump_file, "not replacing:\n "); | |
3391 | print_gimple_stmt (dump_file, phi, 0, 0); | |
3392 | fprintf (dump_file, "\n"); | |
3393 | } | |
726a989a RB |
3394 | gsi_next (&psi); |
3395 | continue; | |
3396 | } | |
3ac01fde | 3397 | |
9b3b55a1 | 3398 | /* Eliminate the PHI node and replace it by a computation outside |
925196ed | 3399 | the loop. */ |
3571dde6 MP |
3400 | if (dump_file) |
3401 | { | |
3402 | fprintf (dump_file, "\nfinal value replacement:\n "); | |
3403 | print_gimple_stmt (dump_file, phi, 0, 0); | |
3404 | fprintf (dump_file, " with\n "); | |
3405 | } | |
925196ed | 3406 | def = unshare_expr (def); |
726a989a RB |
3407 | remove_phi_node (&psi, false); |
3408 | ||
3409 | def = force_gimple_operand_gsi (&bsi, def, false, NULL_TREE, | |
3410 | true, GSI_SAME_STMT); | |
3411 | ass = gimple_build_assign (rslt, def); | |
3412 | gsi_insert_before (&bsi, ass, GSI_SAME_STMT); | |
3571dde6 MP |
3413 | if (dump_file) |
3414 | { | |
3415 | print_gimple_stmt (dump_file, ass, 0, 0); | |
3416 | fprintf (dump_file, "\n"); | |
3417 | } | |
3ac01fde ZD |
3418 | } |
3419 | } | |
c2924966 | 3420 | return 0; |
684aaf29 | 3421 | } |
9e2f83a5 ZD |
3422 | |
3423 | #include "gt-tree-scalar-evolution.h" |