]>
Commit | Line | Data |
---|---|---|
5f40b3cb | 1 | /* Loop autoparallelization. |
5624e564 | 2 | Copyright (C) 2006-2015 Free Software Foundation, Inc. |
70837b71 RL |
3 | Contributed by Sebastian Pop <pop@cri.ensmp.fr> |
4 | Zdenek Dvorak <dvorakz@suse.cz> and Razya Ladelsky <razya@il.ibm.com>. | |
5f40b3cb ZD |
5 | |
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
6da7fc87 | 10 | Software Foundation; either version 3, or (at your option) any later |
5f40b3cb ZD |
11 | version. |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
6da7fc87 NC |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ | |
5f40b3cb ZD |
21 | |
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
60393bbc | 25 | #include "hash-set.h" |
40e23961 | 26 | #include "vec.h" |
40e23961 MC |
27 | #include "input.h" |
28 | #include "alias.h" | |
29 | #include "symtab.h" | |
30 | #include "options.h" | |
40e23961 MC |
31 | #include "inchash.h" |
32 | #include "tree.h" | |
33 | #include "fold-const.h" | |
34 | #include "predict.h" | |
60393bbc AM |
35 | #include "tm.h" |
36 | #include "hard-reg-set.h" | |
37 | #include "input.h" | |
38 | #include "function.h" | |
39 | #include "dominance.h" | |
40 | #include "cfg.h" | |
2fb9a547 AM |
41 | #include "basic-block.h" |
42 | #include "tree-ssa-alias.h" | |
43 | #include "internal-fn.h" | |
44 | #include "gimple-expr.h" | |
45 | #include "is-a.h" | |
18f429e2 | 46 | #include "gimple.h" |
45b0be94 | 47 | #include "gimplify.h" |
5be5c238 | 48 | #include "gimple-iterator.h" |
18f429e2 | 49 | #include "gimplify-me.h" |
5be5c238 | 50 | #include "gimple-walk.h" |
d8a2d370 DN |
51 | #include "stor-layout.h" |
52 | #include "tree-nested.h" | |
442b4905 AM |
53 | #include "gimple-ssa.h" |
54 | #include "tree-cfg.h" | |
55 | #include "tree-phinodes.h" | |
56 | #include "ssa-iterators.h" | |
d8a2d370 | 57 | #include "stringpool.h" |
442b4905 | 58 | #include "tree-ssanames.h" |
e28030cf AM |
59 | #include "tree-ssa-loop-ivopts.h" |
60 | #include "tree-ssa-loop-manip.h" | |
61 | #include "tree-ssa-loop-niter.h" | |
442b4905 AM |
62 | #include "tree-ssa-loop.h" |
63 | #include "tree-into-ssa.h" | |
5f40b3cb | 64 | #include "cfgloop.h" |
5f40b3cb | 65 | #include "tree-data-ref.h" |
1bd6497c | 66 | #include "tree-scalar-evolution.h" |
cf835838 | 67 | #include "gimple-pretty-print.h" |
5f40b3cb | 68 | #include "tree-pass.h" |
5f40b3cb | 69 | #include "langhooks.h" |
a509ebb5 | 70 | #include "tree-vectorizer.h" |
4a8fb1a1 | 71 | #include "tree-hasher.h" |
c1bf2a39 | 72 | #include "tree-parloops.h" |
0645c1a2 | 73 | #include "omp-low.h" |
1fe37220 | 74 | #include "tree-nested.h" |
a79b7ec5 TV |
75 | #include "plugin-api.h" |
76 | #include "ipa-ref.h" | |
77 | #include "cgraph.h" | |
5f40b3cb ZD |
78 | |
79 | /* This pass tries to distribute iterations of loops into several threads. | |
80 | The implementation is straightforward -- for each loop we test whether its | |
81 | iterations are independent, and if it is the case (and some additional | |
82 | conditions regarding profitability and correctness are satisfied), we | |
726a989a RB |
83 | add GIMPLE_OMP_PARALLEL and GIMPLE_OMP_FOR codes and let omp expansion |
84 | machinery do its job. | |
b8698a0f | 85 | |
5f40b3cb ZD |
86 | The most of the complexity is in bringing the code into shape expected |
87 | by the omp expanders: | |
726a989a RB |
88 | -- for GIMPLE_OMP_FOR, ensuring that the loop has only one induction |
89 | variable and that the exit test is at the start of the loop body | |
90 | -- for GIMPLE_OMP_PARALLEL, replacing the references to local addressable | |
5f40b3cb ZD |
91 | variables by accesses through pointers, and breaking up ssa chains |
92 | by storing the values incoming to the parallelized loop to a structure | |
93 | passed to the new function as an argument (something similar is done | |
94 | in omp gimplification, unfortunately only a small part of the code | |
95 | can be shared). | |
96 | ||
97 | TODO: | |
98 | -- if there are several parallelizable loops in a function, it may be | |
99 | possible to generate the threads just once (using synchronization to | |
100 | ensure that cross-loop dependences are obeyed). | |
70837b71 RL |
101 | -- handling of common reduction patterns for outer loops. |
102 | ||
103 | More info can also be found at http://gcc.gnu.org/wiki/AutoParInGCC */ | |
b8698a0f | 104 | /* |
a509ebb5 | 105 | Reduction handling: |
8a9ecffd | 106 | currently we use vect_force_simple_reduction() to detect reduction patterns. |
a509ebb5 | 107 | The code transformation will be introduced by an example. |
b8698a0f L |
108 | |
109 | ||
a509ebb5 RL |
110 | parloop |
111 | { | |
112 | int sum=1; | |
113 | ||
0eb7e7aa | 114 | for (i = 0; i < N; i++) |
a509ebb5 RL |
115 | { |
116 | x[i] = i + 3; | |
117 | sum+=x[i]; | |
118 | } | |
119 | } | |
120 | ||
0eb7e7aa | 121 | gimple-like code: |
a509ebb5 RL |
122 | header_bb: |
123 | ||
0eb7e7aa RL |
124 | # sum_29 = PHI <sum_11(5), 1(3)> |
125 | # i_28 = PHI <i_12(5), 0(3)> | |
126 | D.1795_8 = i_28 + 3; | |
127 | x[i_28] = D.1795_8; | |
128 | sum_11 = D.1795_8 + sum_29; | |
129 | i_12 = i_28 + 1; | |
130 | if (N_6(D) > i_12) | |
131 | goto header_bb; | |
132 | ||
a509ebb5 RL |
133 | |
134 | exit_bb: | |
135 | ||
0eb7e7aa RL |
136 | # sum_21 = PHI <sum_11(4)> |
137 | printf (&"%d"[0], sum_21); | |
a509ebb5 RL |
138 | |
139 | ||
140 | after reduction transformation (only relevant parts): | |
141 | ||
142 | parloop | |
143 | { | |
144 | ||
145 | .... | |
146 | ||
0eb7e7aa | 147 | |
fa10beec | 148 | # Storing the initial value given by the user. # |
0eb7e7aa | 149 | |
ae0bce62 | 150 | .paral_data_store.32.sum.27 = 1; |
b8698a0f L |
151 | |
152 | #pragma omp parallel num_threads(4) | |
a509ebb5 | 153 | |
0eb7e7aa | 154 | #pragma omp for schedule(static) |
ae0bce62 RL |
155 | |
156 | # The neutral element corresponding to the particular | |
157 | reduction's operation, e.g. 0 for PLUS_EXPR, | |
158 | 1 for MULT_EXPR, etc. replaces the user's initial value. # | |
159 | ||
160 | # sum.27_29 = PHI <sum.27_11, 0> | |
161 | ||
0eb7e7aa | 162 | sum.27_11 = D.1827_8 + sum.27_29; |
ae0bce62 | 163 | |
726a989a | 164 | GIMPLE_OMP_CONTINUE |
a509ebb5 | 165 | |
0eb7e7aa RL |
166 | # Adding this reduction phi is done at create_phi_for_local_result() # |
167 | # sum.27_56 = PHI <sum.27_11, 0> | |
726a989a | 168 | GIMPLE_OMP_RETURN |
b8698a0f L |
169 | |
170 | # Creating the atomic operation is done at | |
0eb7e7aa | 171 | create_call_for_reduction_1() # |
a509ebb5 | 172 | |
0eb7e7aa RL |
173 | #pragma omp atomic_load |
174 | D.1839_59 = *&.paral_data_load.33_51->reduction.23; | |
175 | D.1840_60 = sum.27_56 + D.1839_59; | |
176 | #pragma omp atomic_store (D.1840_60); | |
b8698a0f | 177 | |
726a989a | 178 | GIMPLE_OMP_RETURN |
b8698a0f | 179 | |
0eb7e7aa RL |
180 | # collecting the result after the join of the threads is done at |
181 | create_loads_for_reductions(). | |
ae0bce62 RL |
182 | The value computed by the threads is loaded from the |
183 | shared struct. # | |
184 | ||
b8698a0f | 185 | |
0eb7e7aa | 186 | .paral_data_load.33_52 = &.paral_data_store.32; |
ae0bce62 | 187 | sum_37 = .paral_data_load.33_52->sum.27; |
0eb7e7aa RL |
188 | sum_43 = D.1795_41 + sum_37; |
189 | ||
190 | exit bb: | |
191 | # sum_21 = PHI <sum_43, sum_26> | |
192 | printf (&"%d"[0], sum_21); | |
193 | ||
194 | ... | |
195 | ||
a509ebb5 RL |
196 | } |
197 | ||
198 | */ | |
199 | ||
5f40b3cb ZD |
200 | /* Minimal number of iterations of a loop that should be executed in each |
201 | thread. */ | |
202 | #define MIN_PER_THREAD 100 | |
203 | ||
b8698a0f | 204 | /* Element of the hashtable, representing a |
a509ebb5 RL |
205 | reduction in the current loop. */ |
206 | struct reduction_info | |
207 | { | |
726a989a RB |
208 | gimple reduc_stmt; /* reduction statement. */ |
209 | gimple reduc_phi; /* The phi node defining the reduction. */ | |
210 | enum tree_code reduction_code;/* code for the reduction operation. */ | |
5d1fd1de JJ |
211 | unsigned reduc_version; /* SSA_NAME_VERSION of original reduc_phi |
212 | result. */ | |
538dd0b7 | 213 | gphi *keep_res; /* The PHI_RESULT of this phi is the resulting value |
a509ebb5 | 214 | of the reduction variable when existing the loop. */ |
ae0bce62 | 215 | tree initial_value; /* The initial value of the reduction var before entering the loop. */ |
a509ebb5 | 216 | tree field; /* the name of the field in the parloop data structure intended for reduction. */ |
a509ebb5 | 217 | tree init; /* reduction initialization value. */ |
538dd0b7 | 218 | gphi *new_phi; /* (helper field) Newly created phi node whose result |
a509ebb5 RL |
219 | will be passed to the atomic operation. Represents |
220 | the local result each thread computed for the reduction | |
221 | operation. */ | |
222 | }; | |
223 | ||
4a8fb1a1 | 224 | /* Reduction info hashtable helpers. */ |
a509ebb5 | 225 | |
4a8fb1a1 | 226 | struct reduction_hasher : typed_free_remove <reduction_info> |
a509ebb5 | 227 | { |
67f58944 TS |
228 | typedef reduction_info *value_type; |
229 | typedef reduction_info *compare_type; | |
230 | static inline hashval_t hash (const reduction_info *); | |
231 | static inline bool equal (const reduction_info *, const reduction_info *); | |
4a8fb1a1 LC |
232 | }; |
233 | ||
234 | /* Equality and hash functions for hashtab code. */ | |
a509ebb5 | 235 | |
4a8fb1a1 | 236 | inline bool |
67f58944 | 237 | reduction_hasher::equal (const reduction_info *a, const reduction_info *b) |
4a8fb1a1 | 238 | { |
a509ebb5 RL |
239 | return (a->reduc_phi == b->reduc_phi); |
240 | } | |
241 | ||
4a8fb1a1 | 242 | inline hashval_t |
67f58944 | 243 | reduction_hasher::hash (const reduction_info *a) |
a509ebb5 | 244 | { |
5d1fd1de | 245 | return a->reduc_version; |
a509ebb5 RL |
246 | } |
247 | ||
c203e8a7 | 248 | typedef hash_table<reduction_hasher> reduction_info_table_type; |
4a8fb1a1 LC |
249 | |
250 | ||
a509ebb5 | 251 | static struct reduction_info * |
c203e8a7 | 252 | reduction_phi (reduction_info_table_type *reduction_list, gimple phi) |
a509ebb5 RL |
253 | { |
254 | struct reduction_info tmpred, *red; | |
255 | ||
c203e8a7 | 256 | if (reduction_list->elements () == 0 || phi == NULL) |
a509ebb5 RL |
257 | return NULL; |
258 | ||
259 | tmpred.reduc_phi = phi; | |
5d1fd1de | 260 | tmpred.reduc_version = gimple_uid (phi); |
c203e8a7 | 261 | red = reduction_list->find (&tmpred); |
a509ebb5 RL |
262 | |
263 | return red; | |
264 | } | |
265 | ||
5f40b3cb ZD |
266 | /* Element of hashtable of names to copy. */ |
267 | ||
268 | struct name_to_copy_elt | |
269 | { | |
270 | unsigned version; /* The version of the name to copy. */ | |
271 | tree new_name; /* The new name used in the copy. */ | |
272 | tree field; /* The field of the structure used to pass the | |
273 | value. */ | |
274 | }; | |
275 | ||
4a8fb1a1 | 276 | /* Name copies hashtable helpers. */ |
5f40b3cb | 277 | |
4a8fb1a1 | 278 | struct name_to_copy_hasher : typed_free_remove <name_to_copy_elt> |
5f40b3cb | 279 | { |
67f58944 TS |
280 | typedef name_to_copy_elt *value_type; |
281 | typedef name_to_copy_elt *compare_type; | |
282 | static inline hashval_t hash (const name_to_copy_elt *); | |
283 | static inline bool equal (const name_to_copy_elt *, const name_to_copy_elt *); | |
4a8fb1a1 LC |
284 | }; |
285 | ||
286 | /* Equality and hash functions for hashtab code. */ | |
5f40b3cb | 287 | |
4a8fb1a1 | 288 | inline bool |
67f58944 | 289 | name_to_copy_hasher::equal (const name_to_copy_elt *a, const name_to_copy_elt *b) |
4a8fb1a1 | 290 | { |
5f40b3cb ZD |
291 | return a->version == b->version; |
292 | } | |
293 | ||
4a8fb1a1 | 294 | inline hashval_t |
67f58944 | 295 | name_to_copy_hasher::hash (const name_to_copy_elt *a) |
5f40b3cb | 296 | { |
5f40b3cb ZD |
297 | return (hashval_t) a->version; |
298 | } | |
299 | ||
c203e8a7 | 300 | typedef hash_table<name_to_copy_hasher> name_to_copy_table_type; |
4a8fb1a1 | 301 | |
b305e3da SP |
302 | /* A transformation matrix, which is a self-contained ROWSIZE x COLSIZE |
303 | matrix. Rather than use floats, we simply keep a single DENOMINATOR that | |
304 | represents the denominator for every element in the matrix. */ | |
305 | typedef struct lambda_trans_matrix_s | |
306 | { | |
307 | lambda_matrix matrix; | |
308 | int rowsize; | |
309 | int colsize; | |
310 | int denominator; | |
311 | } *lambda_trans_matrix; | |
312 | #define LTM_MATRIX(T) ((T)->matrix) | |
313 | #define LTM_ROWSIZE(T) ((T)->rowsize) | |
314 | #define LTM_COLSIZE(T) ((T)->colsize) | |
315 | #define LTM_DENOMINATOR(T) ((T)->denominator) | |
316 | ||
317 | /* Allocate a new transformation matrix. */ | |
318 | ||
319 | static lambda_trans_matrix | |
320 | lambda_trans_matrix_new (int colsize, int rowsize, | |
321 | struct obstack * lambda_obstack) | |
322 | { | |
323 | lambda_trans_matrix ret; | |
324 | ||
325 | ret = (lambda_trans_matrix) | |
326 | obstack_alloc (lambda_obstack, sizeof (struct lambda_trans_matrix_s)); | |
327 | LTM_MATRIX (ret) = lambda_matrix_new (rowsize, colsize, lambda_obstack); | |
328 | LTM_ROWSIZE (ret) = rowsize; | |
329 | LTM_COLSIZE (ret) = colsize; | |
330 | LTM_DENOMINATOR (ret) = 1; | |
331 | return ret; | |
332 | } | |
333 | ||
334 | /* Multiply a vector VEC by a matrix MAT. | |
335 | MAT is an M*N matrix, and VEC is a vector with length N. The result | |
336 | is stored in DEST which must be a vector of length M. */ | |
337 | ||
338 | static void | |
339 | lambda_matrix_vector_mult (lambda_matrix matrix, int m, int n, | |
340 | lambda_vector vec, lambda_vector dest) | |
341 | { | |
342 | int i, j; | |
343 | ||
344 | lambda_vector_clear (dest, m); | |
345 | for (i = 0; i < m; i++) | |
346 | for (j = 0; j < n; j++) | |
347 | dest[i] += matrix[i][j] * vec[j]; | |
348 | } | |
349 | ||
350 | /* Return true if TRANS is a legal transformation matrix that respects | |
351 | the dependence vectors in DISTS and DIRS. The conservative answer | |
352 | is false. | |
353 | ||
354 | "Wolfe proves that a unimodular transformation represented by the | |
355 | matrix T is legal when applied to a loop nest with a set of | |
356 | lexicographically non-negative distance vectors RDG if and only if | |
357 | for each vector d in RDG, (T.d >= 0) is lexicographically positive. | |
358 | i.e.: if and only if it transforms the lexicographically positive | |
359 | distance vectors to lexicographically positive vectors. Note that | |
360 | a unimodular matrix must transform the zero vector (and only it) to | |
361 | the zero vector." S.Muchnick. */ | |
362 | ||
363 | static bool | |
364 | lambda_transform_legal_p (lambda_trans_matrix trans, | |
365 | int nb_loops, | |
9771b263 | 366 | vec<ddr_p> dependence_relations) |
b305e3da SP |
367 | { |
368 | unsigned int i, j; | |
369 | lambda_vector distres; | |
370 | struct data_dependence_relation *ddr; | |
371 | ||
372 | gcc_assert (LTM_COLSIZE (trans) == nb_loops | |
373 | && LTM_ROWSIZE (trans) == nb_loops); | |
374 | ||
375 | /* When there are no dependences, the transformation is correct. */ | |
9771b263 | 376 | if (dependence_relations.length () == 0) |
b305e3da SP |
377 | return true; |
378 | ||
9771b263 | 379 | ddr = dependence_relations[0]; |
b305e3da SP |
380 | if (ddr == NULL) |
381 | return true; | |
382 | ||
383 | /* When there is an unknown relation in the dependence_relations, we | |
384 | know that it is no worth looking at this loop nest: give up. */ | |
385 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
386 | return false; | |
387 | ||
388 | distres = lambda_vector_new (nb_loops); | |
389 | ||
390 | /* For each distance vector in the dependence graph. */ | |
9771b263 | 391 | FOR_EACH_VEC_ELT (dependence_relations, i, ddr) |
b305e3da SP |
392 | { |
393 | /* Don't care about relations for which we know that there is no | |
394 | dependence, nor about read-read (aka. output-dependences): | |
395 | these data accesses can happen in any order. */ | |
396 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known | |
397 | || (DR_IS_READ (DDR_A (ddr)) && DR_IS_READ (DDR_B (ddr)))) | |
398 | continue; | |
399 | ||
400 | /* Conservatively answer: "this transformation is not valid". */ | |
401 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
402 | return false; | |
403 | ||
404 | /* If the dependence could not be captured by a distance vector, | |
405 | conservatively answer that the transform is not valid. */ | |
406 | if (DDR_NUM_DIST_VECTS (ddr) == 0) | |
407 | return false; | |
408 | ||
409 | /* Compute trans.dist_vect */ | |
410 | for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++) | |
411 | { | |
412 | lambda_matrix_vector_mult (LTM_MATRIX (trans), nb_loops, nb_loops, | |
413 | DDR_DIST_VECT (ddr, j), distres); | |
414 | ||
415 | if (!lambda_vector_lexico_pos (distres, nb_loops)) | |
416 | return false; | |
417 | } | |
418 | } | |
419 | return true; | |
420 | } | |
08dab97a RL |
421 | |
422 | /* Data dependency analysis. Returns true if the iterations of LOOP | |
423 | are independent on each other (that is, if we can execute them | |
424 | in parallel). */ | |
5f40b3cb ZD |
425 | |
426 | static bool | |
f873b205 | 427 | loop_parallel_p (struct loop *loop, struct obstack * parloop_obstack) |
5f40b3cb | 428 | { |
9771b263 DN |
429 | vec<ddr_p> dependence_relations; |
430 | vec<data_reference_p> datarefs; | |
5f40b3cb ZD |
431 | lambda_trans_matrix trans; |
432 | bool ret = false; | |
5f40b3cb ZD |
433 | |
434 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
48710229 RL |
435 | { |
436 | fprintf (dump_file, "Considering loop %d\n", loop->num); | |
437 | if (!loop->inner) | |
438 | fprintf (dump_file, "loop is innermost\n"); | |
b8698a0f | 439 | else |
48710229 RL |
440 | fprintf (dump_file, "loop NOT innermost\n"); |
441 | } | |
5f40b3cb | 442 | |
5f40b3cb ZD |
443 | /* Check for problems with dependences. If the loop can be reversed, |
444 | the iterations are independent. */ | |
00f96dc9 | 445 | auto_vec<loop_p, 3> loop_nest; |
9771b263 | 446 | datarefs.create (10); |
07687835 | 447 | dependence_relations.create (100); |
9ca3d00e AB |
448 | if (! compute_data_dependences_for_loop (loop, true, &loop_nest, &datarefs, |
449 | &dependence_relations)) | |
450 | { | |
451 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
452 | fprintf (dump_file, " FAILED: cannot analyze data dependencies\n"); | |
453 | ret = false; | |
454 | goto end; | |
455 | } | |
5f40b3cb ZD |
456 | if (dump_file && (dump_flags & TDF_DETAILS)) |
457 | dump_data_dependence_relations (dump_file, dependence_relations); | |
458 | ||
f873b205 | 459 | trans = lambda_trans_matrix_new (1, 1, parloop_obstack); |
5f40b3cb ZD |
460 | LTM_MATRIX (trans)[0][0] = -1; |
461 | ||
462 | if (lambda_transform_legal_p (trans, 1, dependence_relations)) | |
463 | { | |
464 | ret = true; | |
465 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
466 | fprintf (dump_file, " SUCCESS: may be parallelized\n"); | |
467 | } | |
468 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
a509ebb5 RL |
469 | fprintf (dump_file, |
470 | " FAILED: data dependencies exist across iterations\n"); | |
5f40b3cb | 471 | |
9ca3d00e | 472 | end: |
5f40b3cb ZD |
473 | free_dependence_relations (dependence_relations); |
474 | free_data_refs (datarefs); | |
475 | ||
476 | return ret; | |
477 | } | |
478 | ||
1d4af1e8 SP |
479 | /* Return true when LOOP contains basic blocks marked with the |
480 | BB_IRREDUCIBLE_LOOP flag. */ | |
481 | ||
482 | static inline bool | |
483 | loop_has_blocks_with_irreducible_flag (struct loop *loop) | |
484 | { | |
485 | unsigned i; | |
486 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
487 | bool res = true; | |
488 | ||
489 | for (i = 0; i < loop->num_nodes; i++) | |
490 | if (bbs[i]->flags & BB_IRREDUCIBLE_LOOP) | |
491 | goto end; | |
492 | ||
493 | res = false; | |
494 | end: | |
495 | free (bbs); | |
496 | return res; | |
497 | } | |
498 | ||
8a171a59 | 499 | /* Assigns the address of OBJ in TYPE to an ssa name, and returns this name. |
9f9f72aa | 500 | The assignment statement is placed on edge ENTRY. DECL_ADDRESS maps decls |
8a171a59 | 501 | to their addresses that can be reused. The address of OBJ is known to |
cba1eb61 JJ |
502 | be invariant in the whole function. Other needed statements are placed |
503 | right before GSI. */ | |
5f40b3cb ZD |
504 | |
505 | static tree | |
4a8fb1a1 | 506 | take_address_of (tree obj, tree type, edge entry, |
c203e8a7 | 507 | int_tree_htab_type *decl_address, gimple_stmt_iterator *gsi) |
5f40b3cb | 508 | { |
8a171a59 | 509 | int uid; |
83d5977e | 510 | tree *var_p, name, addr; |
538dd0b7 | 511 | gassign *stmt; |
726a989a | 512 | gimple_seq stmts; |
5f40b3cb | 513 | |
8a171a59 ZD |
514 | /* Since the address of OBJ is invariant, the trees may be shared. |
515 | Avoid rewriting unrelated parts of the code. */ | |
516 | obj = unshare_expr (obj); | |
517 | for (var_p = &obj; | |
518 | handled_component_p (*var_p); | |
519 | var_p = &TREE_OPERAND (*var_p, 0)) | |
520 | continue; | |
8a171a59 | 521 | |
c9a410f0 RG |
522 | /* Canonicalize the access to base on a MEM_REF. */ |
523 | if (DECL_P (*var_p)) | |
524 | *var_p = build_simple_mem_ref (build_fold_addr_expr (*var_p)); | |
525 | ||
526 | /* Assign a canonical SSA name to the address of the base decl used | |
527 | in the address and share it for all accesses and addresses based | |
528 | on it. */ | |
529 | uid = DECL_UID (TREE_OPERAND (TREE_OPERAND (*var_p, 0), 0)); | |
84baa4b9 TS |
530 | int_tree_map elt; |
531 | elt.uid = uid; | |
532 | int_tree_map *slot = decl_address->find_slot (elt, INSERT); | |
533 | if (!slot->to) | |
5f40b3cb | 534 | { |
cba1eb61 JJ |
535 | if (gsi == NULL) |
536 | return NULL; | |
c9a410f0 | 537 | addr = TREE_OPERAND (*var_p, 0); |
29b89442 JJ |
538 | const char *obj_name |
539 | = get_name (TREE_OPERAND (TREE_OPERAND (*var_p, 0), 0)); | |
540 | if (obj_name) | |
541 | name = make_temp_ssa_name (TREE_TYPE (addr), NULL, obj_name); | |
542 | else | |
b731b390 | 543 | name = make_ssa_name (TREE_TYPE (addr)); |
83d5977e | 544 | stmt = gimple_build_assign (name, addr); |
726a989a | 545 | gsi_insert_on_edge_immediate (entry, stmt); |
5f40b3cb | 546 | |
84baa4b9 TS |
547 | slot->uid = uid; |
548 | slot->to = name; | |
5f40b3cb | 549 | } |
8a171a59 | 550 | else |
84baa4b9 | 551 | name = slot->to; |
5f40b3cb | 552 | |
c9a410f0 RG |
553 | /* Express the address in terms of the canonical SSA name. */ |
554 | TREE_OPERAND (*var_p, 0) = name; | |
cba1eb61 JJ |
555 | if (gsi == NULL) |
556 | return build_fold_addr_expr_with_type (obj, type); | |
557 | ||
c9a410f0 RG |
558 | name = force_gimple_operand (build_addr (obj, current_function_decl), |
559 | &stmts, true, NULL_TREE); | |
560 | if (!gimple_seq_empty_p (stmts)) | |
cba1eb61 | 561 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
5f40b3cb | 562 | |
c9a410f0 | 563 | if (!useless_type_conversion_p (type, TREE_TYPE (name))) |
8a171a59 | 564 | { |
726a989a | 565 | name = force_gimple_operand (fold_convert (type, name), &stmts, true, |
8a171a59 | 566 | NULL_TREE); |
726a989a | 567 | if (!gimple_seq_empty_p (stmts)) |
cba1eb61 | 568 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
8a171a59 | 569 | } |
5f40b3cb ZD |
570 | |
571 | return name; | |
572 | } | |
573 | ||
a509ebb5 | 574 | /* Callback for htab_traverse. Create the initialization statement |
b8698a0f | 575 | for reduction described in SLOT, and place it at the preheader of |
a509ebb5 RL |
576 | the loop described in DATA. */ |
577 | ||
4a8fb1a1 LC |
578 | int |
579 | initialize_reductions (reduction_info **slot, struct loop *loop) | |
a509ebb5 | 580 | { |
a509ebb5 | 581 | tree init, c; |
a509ebb5 RL |
582 | tree bvar, type, arg; |
583 | edge e; | |
584 | ||
4a8fb1a1 | 585 | struct reduction_info *const reduc = *slot; |
a509ebb5 | 586 | |
b8698a0f | 587 | /* Create initialization in preheader: |
a509ebb5 RL |
588 | reduction_variable = initialization value of reduction. */ |
589 | ||
b8698a0f | 590 | /* In the phi node at the header, replace the argument coming |
a509ebb5 RL |
591 | from the preheader with the reduction initialization value. */ |
592 | ||
593 | /* Create a new variable to initialize the reduction. */ | |
594 | type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); | |
595 | bvar = create_tmp_var (type, "reduction"); | |
a509ebb5 | 596 | |
c2255bc4 AH |
597 | c = build_omp_clause (gimple_location (reduc->reduc_stmt), |
598 | OMP_CLAUSE_REDUCTION); | |
a509ebb5 | 599 | OMP_CLAUSE_REDUCTION_CODE (c) = reduc->reduction_code; |
726a989a | 600 | OMP_CLAUSE_DECL (c) = SSA_NAME_VAR (gimple_assign_lhs (reduc->reduc_stmt)); |
a509ebb5 RL |
601 | |
602 | init = omp_reduction_init (c, TREE_TYPE (bvar)); | |
603 | reduc->init = init; | |
604 | ||
b8698a0f L |
605 | /* Replace the argument representing the initialization value |
606 | with the initialization value for the reduction (neutral | |
607 | element for the particular operation, e.g. 0 for PLUS_EXPR, | |
608 | 1 for MULT_EXPR, etc). | |
609 | Keep the old value in a new variable "reduction_initial", | |
610 | that will be taken in consideration after the parallel | |
0eb7e7aa | 611 | computing is done. */ |
a509ebb5 RL |
612 | |
613 | e = loop_preheader_edge (loop); | |
614 | arg = PHI_ARG_DEF_FROM_EDGE (reduc->reduc_phi, e); | |
615 | /* Create new variable to hold the initial value. */ | |
a509ebb5 | 616 | |
a509ebb5 | 617 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE |
0eb7e7aa | 618 | (reduc->reduc_phi, loop_preheader_edge (loop)), init); |
ae0bce62 | 619 | reduc->initial_value = arg; |
a509ebb5 RL |
620 | return 1; |
621 | } | |
5f40b3cb ZD |
622 | |
623 | struct elv_data | |
624 | { | |
726a989a | 625 | struct walk_stmt_info info; |
9f9f72aa | 626 | edge entry; |
c203e8a7 | 627 | int_tree_htab_type *decl_address; |
cba1eb61 | 628 | gimple_stmt_iterator *gsi; |
5f40b3cb | 629 | bool changed; |
cba1eb61 | 630 | bool reset; |
5f40b3cb ZD |
631 | }; |
632 | ||
9f9f72aa AP |
633 | /* Eliminates references to local variables in *TP out of the single |
634 | entry single exit region starting at DTA->ENTRY. | |
635 | DECL_ADDRESS contains addresses of the references that had their | |
636 | address taken already. If the expression is changed, CHANGED is | |
637 | set to true. Callback for walk_tree. */ | |
a509ebb5 | 638 | |
5f40b3cb | 639 | static tree |
8a171a59 | 640 | eliminate_local_variables_1 (tree *tp, int *walk_subtrees, void *data) |
5f40b3cb | 641 | { |
3d9a9f94 | 642 | struct elv_data *const dta = (struct elv_data *) data; |
8a171a59 | 643 | tree t = *tp, var, addr, addr_type, type, obj; |
5f40b3cb ZD |
644 | |
645 | if (DECL_P (t)) | |
646 | { | |
647 | *walk_subtrees = 0; | |
648 | ||
649 | if (!SSA_VAR_P (t) || DECL_EXTERNAL (t)) | |
650 | return NULL_TREE; | |
651 | ||
652 | type = TREE_TYPE (t); | |
653 | addr_type = build_pointer_type (type); | |
cba1eb61 JJ |
654 | addr = take_address_of (t, addr_type, dta->entry, dta->decl_address, |
655 | dta->gsi); | |
656 | if (dta->gsi == NULL && addr == NULL_TREE) | |
657 | { | |
658 | dta->reset = true; | |
659 | return NULL_TREE; | |
660 | } | |
661 | ||
70f34814 | 662 | *tp = build_simple_mem_ref (addr); |
5f40b3cb ZD |
663 | |
664 | dta->changed = true; | |
665 | return NULL_TREE; | |
666 | } | |
667 | ||
668 | if (TREE_CODE (t) == ADDR_EXPR) | |
669 | { | |
8a171a59 ZD |
670 | /* ADDR_EXPR may appear in two contexts: |
671 | -- as a gimple operand, when the address taken is a function invariant | |
672 | -- as gimple rhs, when the resulting address in not a function | |
673 | invariant | |
674 | We do not need to do anything special in the latter case (the base of | |
675 | the memory reference whose address is taken may be replaced in the | |
676 | DECL_P case). The former case is more complicated, as we need to | |
677 | ensure that the new address is still a gimple operand. Thus, it | |
678 | is not sufficient to replace just the base of the memory reference -- | |
679 | we need to move the whole computation of the address out of the | |
680 | loop. */ | |
681 | if (!is_gimple_val (t)) | |
5f40b3cb ZD |
682 | return NULL_TREE; |
683 | ||
684 | *walk_subtrees = 0; | |
8a171a59 ZD |
685 | obj = TREE_OPERAND (t, 0); |
686 | var = get_base_address (obj); | |
687 | if (!var || !SSA_VAR_P (var) || DECL_EXTERNAL (var)) | |
5f40b3cb ZD |
688 | return NULL_TREE; |
689 | ||
690 | addr_type = TREE_TYPE (t); | |
cba1eb61 JJ |
691 | addr = take_address_of (obj, addr_type, dta->entry, dta->decl_address, |
692 | dta->gsi); | |
693 | if (dta->gsi == NULL && addr == NULL_TREE) | |
694 | { | |
695 | dta->reset = true; | |
696 | return NULL_TREE; | |
697 | } | |
5f40b3cb ZD |
698 | *tp = addr; |
699 | ||
700 | dta->changed = true; | |
701 | return NULL_TREE; | |
702 | } | |
703 | ||
726a989a | 704 | if (!EXPR_P (t)) |
5f40b3cb ZD |
705 | *walk_subtrees = 0; |
706 | ||
707 | return NULL_TREE; | |
708 | } | |
709 | ||
cba1eb61 | 710 | /* Moves the references to local variables in STMT at *GSI out of the single |
9f9f72aa AP |
711 | entry single exit region starting at ENTRY. DECL_ADDRESS contains |
712 | addresses of the references that had their address taken | |
713 | already. */ | |
5f40b3cb ZD |
714 | |
715 | static void | |
cba1eb61 | 716 | eliminate_local_variables_stmt (edge entry, gimple_stmt_iterator *gsi, |
c203e8a7 | 717 | int_tree_htab_type *decl_address) |
5f40b3cb ZD |
718 | { |
719 | struct elv_data dta; | |
cba1eb61 | 720 | gimple stmt = gsi_stmt (*gsi); |
5f40b3cb | 721 | |
726a989a | 722 | memset (&dta.info, '\0', sizeof (dta.info)); |
9f9f72aa | 723 | dta.entry = entry; |
5f40b3cb ZD |
724 | dta.decl_address = decl_address; |
725 | dta.changed = false; | |
cba1eb61 | 726 | dta.reset = false; |
5f40b3cb | 727 | |
b5b8b0ac | 728 | if (gimple_debug_bind_p (stmt)) |
cba1eb61 JJ |
729 | { |
730 | dta.gsi = NULL; | |
731 | walk_tree (gimple_debug_bind_get_value_ptr (stmt), | |
732 | eliminate_local_variables_1, &dta.info, NULL); | |
733 | if (dta.reset) | |
734 | { | |
735 | gimple_debug_bind_reset_value (stmt); | |
736 | dta.changed = true; | |
737 | } | |
738 | } | |
29b89442 JJ |
739 | else if (gimple_clobber_p (stmt)) |
740 | { | |
741 | stmt = gimple_build_nop (); | |
742 | gsi_replace (gsi, stmt, false); | |
743 | dta.changed = true; | |
744 | } | |
b5b8b0ac | 745 | else |
cba1eb61 JJ |
746 | { |
747 | dta.gsi = gsi; | |
748 | walk_gimple_op (stmt, eliminate_local_variables_1, &dta.info); | |
749 | } | |
5f40b3cb ZD |
750 | |
751 | if (dta.changed) | |
752 | update_stmt (stmt); | |
753 | } | |
754 | ||
9f9f72aa AP |
755 | /* Eliminates the references to local variables from the single entry |
756 | single exit region between the ENTRY and EXIT edges. | |
b8698a0f | 757 | |
a509ebb5 | 758 | This includes: |
b8698a0f L |
759 | 1) Taking address of a local variable -- these are moved out of the |
760 | region (and temporary variable is created to hold the address if | |
a509ebb5 | 761 | necessary). |
9f9f72aa | 762 | |
5f40b3cb | 763 | 2) Dereferencing a local variable -- these are replaced with indirect |
a509ebb5 | 764 | references. */ |
5f40b3cb ZD |
765 | |
766 | static void | |
9f9f72aa | 767 | eliminate_local_variables (edge entry, edge exit) |
5f40b3cb | 768 | { |
9f9f72aa | 769 | basic_block bb; |
00f96dc9 | 770 | auto_vec<basic_block, 3> body; |
5f40b3cb | 771 | unsigned i; |
726a989a | 772 | gimple_stmt_iterator gsi; |
cba1eb61 | 773 | bool has_debug_stmt = false; |
c203e8a7 | 774 | int_tree_htab_type decl_address (10); |
9f9f72aa AP |
775 | basic_block entry_bb = entry->src; |
776 | basic_block exit_bb = exit->dest; | |
5f40b3cb | 777 | |
9f9f72aa | 778 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); |
5f40b3cb | 779 | |
9771b263 | 780 | FOR_EACH_VEC_ELT (body, i, bb) |
9f9f72aa | 781 | if (bb != entry_bb && bb != exit_bb) |
726a989a | 782 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
ddb555ed JJ |
783 | if (is_gimple_debug (gsi_stmt (gsi))) |
784 | { | |
785 | if (gimple_debug_bind_p (gsi_stmt (gsi))) | |
786 | has_debug_stmt = true; | |
787 | } | |
cba1eb61 | 788 | else |
c203e8a7 | 789 | eliminate_local_variables_stmt (entry, &gsi, &decl_address); |
cba1eb61 JJ |
790 | |
791 | if (has_debug_stmt) | |
9771b263 | 792 | FOR_EACH_VEC_ELT (body, i, bb) |
cba1eb61 JJ |
793 | if (bb != entry_bb && bb != exit_bb) |
794 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
795 | if (gimple_debug_bind_p (gsi_stmt (gsi))) | |
c203e8a7 | 796 | eliminate_local_variables_stmt (entry, &gsi, &decl_address); |
9f9f72aa AP |
797 | } |
798 | ||
799 | /* Returns true if expression EXPR is not defined between ENTRY and | |
800 | EXIT, i.e. if all its operands are defined outside of the region. */ | |
801 | ||
802 | static bool | |
803 | expr_invariant_in_region_p (edge entry, edge exit, tree expr) | |
804 | { | |
805 | basic_block entry_bb = entry->src; | |
806 | basic_block exit_bb = exit->dest; | |
807 | basic_block def_bb; | |
9f9f72aa AP |
808 | |
809 | if (is_gimple_min_invariant (expr)) | |
810 | return true; | |
811 | ||
812 | if (TREE_CODE (expr) == SSA_NAME) | |
813 | { | |
726a989a | 814 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
9f9f72aa AP |
815 | if (def_bb |
816 | && dominated_by_p (CDI_DOMINATORS, def_bb, entry_bb) | |
817 | && !dominated_by_p (CDI_DOMINATORS, def_bb, exit_bb)) | |
818 | return false; | |
819 | ||
820 | return true; | |
821 | } | |
822 | ||
726a989a | 823 | return false; |
5f40b3cb ZD |
824 | } |
825 | ||
826 | /* If COPY_NAME_P is true, creates and returns a duplicate of NAME. | |
827 | The copies are stored to NAME_COPIES, if NAME was already duplicated, | |
828 | its duplicate stored in NAME_COPIES is returned. | |
b8698a0f | 829 | |
5f40b3cb ZD |
830 | Regardless of COPY_NAME_P, the decl used as a base of the ssa name is also |
831 | duplicated, storing the copies in DECL_COPIES. */ | |
832 | ||
833 | static tree | |
c203e8a7 TS |
834 | separate_decls_in_region_name (tree name, name_to_copy_table_type *name_copies, |
835 | int_tree_htab_type *decl_copies, | |
836 | bool copy_name_p) | |
5f40b3cb ZD |
837 | { |
838 | tree copy, var, var_copy; | |
839 | unsigned idx, uid, nuid; | |
84baa4b9 | 840 | struct int_tree_map ielt; |
5f40b3cb | 841 | struct name_to_copy_elt elt, *nelt; |
4a8fb1a1 | 842 | name_to_copy_elt **slot; |
84baa4b9 | 843 | int_tree_map *dslot; |
5f40b3cb ZD |
844 | |
845 | if (TREE_CODE (name) != SSA_NAME) | |
846 | return name; | |
847 | ||
848 | idx = SSA_NAME_VERSION (name); | |
849 | elt.version = idx; | |
c203e8a7 TS |
850 | slot = name_copies->find_slot_with_hash (&elt, idx, |
851 | copy_name_p ? INSERT : NO_INSERT); | |
5f40b3cb | 852 | if (slot && *slot) |
4a8fb1a1 | 853 | return (*slot)->new_name; |
5f40b3cb | 854 | |
70b5e7dc RG |
855 | if (copy_name_p) |
856 | { | |
857 | copy = duplicate_ssa_name (name, NULL); | |
858 | nelt = XNEW (struct name_to_copy_elt); | |
859 | nelt->version = idx; | |
860 | nelt->new_name = copy; | |
861 | nelt->field = NULL_TREE; | |
862 | *slot = nelt; | |
863 | } | |
864 | else | |
865 | { | |
866 | gcc_assert (!slot); | |
867 | copy = name; | |
868 | } | |
869 | ||
5f40b3cb | 870 | var = SSA_NAME_VAR (name); |
70b5e7dc RG |
871 | if (!var) |
872 | return copy; | |
873 | ||
5f40b3cb ZD |
874 | uid = DECL_UID (var); |
875 | ielt.uid = uid; | |
84baa4b9 TS |
876 | dslot = decl_copies->find_slot_with_hash (ielt, uid, INSERT); |
877 | if (!dslot->to) | |
5f40b3cb ZD |
878 | { |
879 | var_copy = create_tmp_var (TREE_TYPE (var), get_name (var)); | |
36ad7922 | 880 | DECL_GIMPLE_REG_P (var_copy) = DECL_GIMPLE_REG_P (var); |
84baa4b9 TS |
881 | dslot->uid = uid; |
882 | dslot->to = var_copy; | |
5f40b3cb ZD |
883 | |
884 | /* Ensure that when we meet this decl next time, we won't duplicate | |
a509ebb5 | 885 | it again. */ |
5f40b3cb ZD |
886 | nuid = DECL_UID (var_copy); |
887 | ielt.uid = nuid; | |
84baa4b9 TS |
888 | dslot = decl_copies->find_slot_with_hash (ielt, nuid, INSERT); |
889 | gcc_assert (!dslot->to); | |
890 | dslot->uid = nuid; | |
891 | dslot->to = var_copy; | |
5f40b3cb ZD |
892 | } |
893 | else | |
84baa4b9 | 894 | var_copy = dslot->to; |
5f40b3cb | 895 | |
b2ec94d4 | 896 | replace_ssa_name_symbol (copy, var_copy); |
5f40b3cb ZD |
897 | return copy; |
898 | } | |
899 | ||
9f9f72aa AP |
900 | /* Finds the ssa names used in STMT that are defined outside the |
901 | region between ENTRY and EXIT and replaces such ssa names with | |
902 | their duplicates. The duplicates are stored to NAME_COPIES. Base | |
903 | decls of all ssa names used in STMT (including those defined in | |
904 | LOOP) are replaced with the new temporary variables; the | |
905 | replacement decls are stored in DECL_COPIES. */ | |
5f40b3cb ZD |
906 | |
907 | static void | |
726a989a | 908 | separate_decls_in_region_stmt (edge entry, edge exit, gimple stmt, |
c203e8a7 TS |
909 | name_to_copy_table_type *name_copies, |
910 | int_tree_htab_type *decl_copies) | |
5f40b3cb ZD |
911 | { |
912 | use_operand_p use; | |
913 | def_operand_p def; | |
914 | ssa_op_iter oi; | |
915 | tree name, copy; | |
916 | bool copy_name_p; | |
917 | ||
5f40b3cb | 918 | FOR_EACH_PHI_OR_STMT_DEF (def, stmt, oi, SSA_OP_DEF) |
a509ebb5 RL |
919 | { |
920 | name = DEF_FROM_PTR (def); | |
921 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
9f9f72aa AP |
922 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, |
923 | false); | |
a509ebb5 RL |
924 | gcc_assert (copy == name); |
925 | } | |
5f40b3cb ZD |
926 | |
927 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
a509ebb5 RL |
928 | { |
929 | name = USE_FROM_PTR (use); | |
930 | if (TREE_CODE (name) != SSA_NAME) | |
931 | continue; | |
932 | ||
9f9f72aa AP |
933 | copy_name_p = expr_invariant_in_region_p (entry, exit, name); |
934 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, | |
935 | copy_name_p); | |
a509ebb5 RL |
936 | SET_USE (use, copy); |
937 | } | |
5f40b3cb ZD |
938 | } |
939 | ||
b5b8b0ac AO |
940 | /* Finds the ssa names used in STMT that are defined outside the |
941 | region between ENTRY and EXIT and replaces such ssa names with | |
942 | their duplicates. The duplicates are stored to NAME_COPIES. Base | |
943 | decls of all ssa names used in STMT (including those defined in | |
944 | LOOP) are replaced with the new temporary variables; the | |
945 | replacement decls are stored in DECL_COPIES. */ | |
946 | ||
947 | static bool | |
4a8fb1a1 | 948 | separate_decls_in_region_debug (gimple stmt, |
c203e8a7 TS |
949 | name_to_copy_table_type *name_copies, |
950 | int_tree_htab_type *decl_copies) | |
b5b8b0ac AO |
951 | { |
952 | use_operand_p use; | |
953 | ssa_op_iter oi; | |
954 | tree var, name; | |
955 | struct int_tree_map ielt; | |
956 | struct name_to_copy_elt elt; | |
4a8fb1a1 | 957 | name_to_copy_elt **slot; |
84baa4b9 | 958 | int_tree_map *dslot; |
b5b8b0ac | 959 | |
ddb555ed JJ |
960 | if (gimple_debug_bind_p (stmt)) |
961 | var = gimple_debug_bind_get_var (stmt); | |
962 | else if (gimple_debug_source_bind_p (stmt)) | |
963 | var = gimple_debug_source_bind_get_var (stmt); | |
964 | else | |
965 | return true; | |
598e67d7 | 966 | if (TREE_CODE (var) == DEBUG_EXPR_DECL || TREE_CODE (var) == LABEL_DECL) |
4f2a9af8 | 967 | return true; |
b5b8b0ac AO |
968 | gcc_assert (DECL_P (var) && SSA_VAR_P (var)); |
969 | ielt.uid = DECL_UID (var); | |
84baa4b9 | 970 | dslot = decl_copies->find_slot_with_hash (ielt, ielt.uid, NO_INSERT); |
b5b8b0ac AO |
971 | if (!dslot) |
972 | return true; | |
ddb555ed | 973 | if (gimple_debug_bind_p (stmt)) |
84baa4b9 | 974 | gimple_debug_bind_set_var (stmt, dslot->to); |
ddb555ed | 975 | else if (gimple_debug_source_bind_p (stmt)) |
84baa4b9 | 976 | gimple_debug_source_bind_set_var (stmt, dslot->to); |
b5b8b0ac AO |
977 | |
978 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
979 | { | |
980 | name = USE_FROM_PTR (use); | |
981 | if (TREE_CODE (name) != SSA_NAME) | |
982 | continue; | |
983 | ||
984 | elt.version = SSA_NAME_VERSION (name); | |
c203e8a7 | 985 | slot = name_copies->find_slot_with_hash (&elt, elt.version, NO_INSERT); |
b5b8b0ac AO |
986 | if (!slot) |
987 | { | |
988 | gimple_debug_bind_reset_value (stmt); | |
989 | update_stmt (stmt); | |
990 | break; | |
991 | } | |
992 | ||
4a8fb1a1 | 993 | SET_USE (use, (*slot)->new_name); |
b5b8b0ac AO |
994 | } |
995 | ||
996 | return false; | |
997 | } | |
998 | ||
0eb7e7aa RL |
999 | /* Callback for htab_traverse. Adds a field corresponding to the reduction |
1000 | specified in SLOT. The type is passed in DATA. */ | |
1001 | ||
4a8fb1a1 LC |
1002 | int |
1003 | add_field_for_reduction (reduction_info **slot, tree type) | |
a509ebb5 | 1004 | { |
b8698a0f | 1005 | |
4a8fb1a1 | 1006 | struct reduction_info *const red = *slot; |
aa06a978 RB |
1007 | tree var = gimple_assign_lhs (red->reduc_stmt); |
1008 | tree field = build_decl (gimple_location (red->reduc_stmt), FIELD_DECL, | |
1009 | SSA_NAME_IDENTIFIER (var), TREE_TYPE (var)); | |
0eb7e7aa RL |
1010 | |
1011 | insert_field_into_struct (type, field); | |
1012 | ||
1013 | red->field = field; | |
1014 | ||
1015 | return 1; | |
1016 | } | |
a509ebb5 | 1017 | |
5f40b3cb | 1018 | /* Callback for htab_traverse. Adds a field corresponding to a ssa name |
b8698a0f | 1019 | described in SLOT. The type is passed in DATA. */ |
5f40b3cb | 1020 | |
4a8fb1a1 LC |
1021 | int |
1022 | add_field_for_name (name_to_copy_elt **slot, tree type) | |
5f40b3cb | 1023 | { |
4a8fb1a1 | 1024 | struct name_to_copy_elt *const elt = *slot; |
5f40b3cb | 1025 | tree name = ssa_name (elt->version); |
70b5e7dc RG |
1026 | tree field = build_decl (UNKNOWN_LOCATION, |
1027 | FIELD_DECL, SSA_NAME_IDENTIFIER (name), | |
1028 | TREE_TYPE (name)); | |
5f40b3cb ZD |
1029 | |
1030 | insert_field_into_struct (type, field); | |
1031 | elt->field = field; | |
a509ebb5 | 1032 | |
5f40b3cb ZD |
1033 | return 1; |
1034 | } | |
1035 | ||
b8698a0f L |
1036 | /* Callback for htab_traverse. A local result is the intermediate result |
1037 | computed by a single | |
fa10beec | 1038 | thread, or the initial value in case no iteration was executed. |
b8698a0f L |
1039 | This function creates a phi node reflecting these values. |
1040 | The phi's result will be stored in NEW_PHI field of the | |
1041 | reduction's data structure. */ | |
a509ebb5 | 1042 | |
4a8fb1a1 LC |
1043 | int |
1044 | create_phi_for_local_result (reduction_info **slot, struct loop *loop) | |
a509ebb5 | 1045 | { |
4a8fb1a1 | 1046 | struct reduction_info *const reduc = *slot; |
a509ebb5 | 1047 | edge e; |
538dd0b7 | 1048 | gphi *new_phi; |
a509ebb5 RL |
1049 | basic_block store_bb; |
1050 | tree local_res; | |
f5045c96 | 1051 | source_location locus; |
a509ebb5 | 1052 | |
b8698a0f L |
1053 | /* STORE_BB is the block where the phi |
1054 | should be stored. It is the destination of the loop exit. | |
726a989a | 1055 | (Find the fallthru edge from GIMPLE_OMP_CONTINUE). */ |
a509ebb5 RL |
1056 | store_bb = FALLTHRU_EDGE (loop->latch)->dest; |
1057 | ||
1058 | /* STORE_BB has two predecessors. One coming from the loop | |
1059 | (the reduction's result is computed at the loop), | |
b8698a0f L |
1060 | and another coming from a block preceding the loop, |
1061 | when no iterations | |
1062 | are executed (the initial value should be taken). */ | |
a509ebb5 RL |
1063 | if (EDGE_PRED (store_bb, 0) == FALLTHRU_EDGE (loop->latch)) |
1064 | e = EDGE_PRED (store_bb, 1); | |
1065 | else | |
1066 | e = EDGE_PRED (store_bb, 0); | |
b731b390 | 1067 | local_res = copy_ssa_name (gimple_assign_lhs (reduc->reduc_stmt)); |
f5045c96 | 1068 | locus = gimple_location (reduc->reduc_stmt); |
a509ebb5 | 1069 | new_phi = create_phi_node (local_res, store_bb); |
9e227d60 | 1070 | add_phi_arg (new_phi, reduc->init, e, locus); |
726a989a | 1071 | add_phi_arg (new_phi, gimple_assign_lhs (reduc->reduc_stmt), |
9e227d60 | 1072 | FALLTHRU_EDGE (loop->latch), locus); |
a509ebb5 RL |
1073 | reduc->new_phi = new_phi; |
1074 | ||
1075 | return 1; | |
1076 | } | |
5f40b3cb ZD |
1077 | |
1078 | struct clsn_data | |
1079 | { | |
1080 | tree store; | |
1081 | tree load; | |
1082 | ||
1083 | basic_block store_bb; | |
1084 | basic_block load_bb; | |
1085 | }; | |
1086 | ||
a509ebb5 | 1087 | /* Callback for htab_traverse. Create an atomic instruction for the |
b8698a0f | 1088 | reduction described in SLOT. |
a509ebb5 RL |
1089 | DATA annotates the place in memory the atomic operation relates to, |
1090 | and the basic block it needs to be generated in. */ | |
1091 | ||
4a8fb1a1 LC |
1092 | int |
1093 | create_call_for_reduction_1 (reduction_info **slot, struct clsn_data *clsn_data) | |
a509ebb5 | 1094 | { |
4a8fb1a1 | 1095 | struct reduction_info *const reduc = *slot; |
726a989a | 1096 | gimple_stmt_iterator gsi; |
a509ebb5 | 1097 | tree type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); |
a509ebb5 RL |
1098 | tree load_struct; |
1099 | basic_block bb; | |
1100 | basic_block new_bb; | |
1101 | edge e; | |
0f900dfa | 1102 | tree t, addr, ref, x; |
726a989a RB |
1103 | tree tmp_load, name; |
1104 | gimple load; | |
a509ebb5 | 1105 | |
70f34814 | 1106 | load_struct = build_simple_mem_ref (clsn_data->load); |
a509ebb5 | 1107 | t = build3 (COMPONENT_REF, type, load_struct, reduc->field, NULL_TREE); |
a509ebb5 RL |
1108 | |
1109 | addr = build_addr (t, current_function_decl); | |
1110 | ||
1111 | /* Create phi node. */ | |
1112 | bb = clsn_data->load_bb; | |
1113 | ||
b13c907a RB |
1114 | gsi = gsi_last_bb (bb); |
1115 | e = split_block (bb, gsi_stmt (gsi)); | |
a509ebb5 RL |
1116 | new_bb = e->dest; |
1117 | ||
b731b390 JJ |
1118 | tmp_load = create_tmp_var (TREE_TYPE (TREE_TYPE (addr))); |
1119 | tmp_load = make_ssa_name (tmp_load); | |
726a989a | 1120 | load = gimple_build_omp_atomic_load (tmp_load, addr); |
a509ebb5 | 1121 | SSA_NAME_DEF_STMT (tmp_load) = load; |
726a989a RB |
1122 | gsi = gsi_start_bb (new_bb); |
1123 | gsi_insert_after (&gsi, load, GSI_NEW_STMT); | |
a509ebb5 RL |
1124 | |
1125 | e = split_block (new_bb, load); | |
1126 | new_bb = e->dest; | |
726a989a | 1127 | gsi = gsi_start_bb (new_bb); |
a509ebb5 | 1128 | ref = tmp_load; |
726a989a RB |
1129 | x = fold_build2 (reduc->reduction_code, |
1130 | TREE_TYPE (PHI_RESULT (reduc->new_phi)), ref, | |
1131 | PHI_RESULT (reduc->new_phi)); | |
a509ebb5 | 1132 | |
726a989a RB |
1133 | name = force_gimple_operand_gsi (&gsi, x, true, NULL_TREE, true, |
1134 | GSI_CONTINUE_LINKING); | |
a509ebb5 | 1135 | |
726a989a | 1136 | gsi_insert_after (&gsi, gimple_build_omp_atomic_store (name), GSI_NEW_STMT); |
a509ebb5 RL |
1137 | return 1; |
1138 | } | |
1139 | ||
b8698a0f L |
1140 | /* Create the atomic operation at the join point of the threads. |
1141 | REDUCTION_LIST describes the reductions in the LOOP. | |
1142 | LD_ST_DATA describes the shared data structure where | |
a509ebb5 RL |
1143 | shared data is stored in and loaded from. */ |
1144 | static void | |
4a8fb1a1 | 1145 | create_call_for_reduction (struct loop *loop, |
c203e8a7 | 1146 | reduction_info_table_type *reduction_list, |
a509ebb5 RL |
1147 | struct clsn_data *ld_st_data) |
1148 | { | |
c203e8a7 | 1149 | reduction_list->traverse <struct loop *, create_phi_for_local_result> (loop); |
726a989a | 1150 | /* Find the fallthru edge from GIMPLE_OMP_CONTINUE. */ |
a509ebb5 | 1151 | ld_st_data->load_bb = FALLTHRU_EDGE (loop->latch)->dest; |
4a8fb1a1 | 1152 | reduction_list |
c203e8a7 | 1153 | ->traverse <struct clsn_data *, create_call_for_reduction_1> (ld_st_data); |
a509ebb5 RL |
1154 | } |
1155 | ||
ae0bce62 RL |
1156 | /* Callback for htab_traverse. Loads the final reduction value at the |
1157 | join point of all threads, and inserts it in the right place. */ | |
a509ebb5 | 1158 | |
4a8fb1a1 LC |
1159 | int |
1160 | create_loads_for_reductions (reduction_info **slot, struct clsn_data *clsn_data) | |
a509ebb5 | 1161 | { |
4a8fb1a1 | 1162 | struct reduction_info *const red = *slot; |
726a989a RB |
1163 | gimple stmt; |
1164 | gimple_stmt_iterator gsi; | |
1165 | tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); | |
a509ebb5 | 1166 | tree load_struct; |
ae0bce62 | 1167 | tree name; |
a509ebb5 RL |
1168 | tree x; |
1169 | ||
726a989a | 1170 | gsi = gsi_after_labels (clsn_data->load_bb); |
70f34814 | 1171 | load_struct = build_simple_mem_ref (clsn_data->load); |
a509ebb5 RL |
1172 | load_struct = build3 (COMPONENT_REF, type, load_struct, red->field, |
1173 | NULL_TREE); | |
a509ebb5 | 1174 | |
ae0bce62 | 1175 | x = load_struct; |
a509ebb5 | 1176 | name = PHI_RESULT (red->keep_res); |
726a989a | 1177 | stmt = gimple_build_assign (name, x); |
a509ebb5 | 1178 | |
726a989a | 1179 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
a509ebb5 | 1180 | |
726a989a RB |
1181 | for (gsi = gsi_start_phis (gimple_bb (red->keep_res)); |
1182 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
1183 | if (gsi_stmt (gsi) == red->keep_res) | |
1184 | { | |
1185 | remove_phi_node (&gsi, false); | |
1186 | return 1; | |
1187 | } | |
1188 | gcc_unreachable (); | |
a509ebb5 RL |
1189 | } |
1190 | ||
b8698a0f | 1191 | /* Load the reduction result that was stored in LD_ST_DATA. |
a509ebb5 | 1192 | REDUCTION_LIST describes the list of reductions that the |
fa10beec | 1193 | loads should be generated for. */ |
a509ebb5 | 1194 | static void |
c203e8a7 | 1195 | create_final_loads_for_reduction (reduction_info_table_type *reduction_list, |
a509ebb5 RL |
1196 | struct clsn_data *ld_st_data) |
1197 | { | |
726a989a | 1198 | gimple_stmt_iterator gsi; |
a509ebb5 | 1199 | tree t; |
726a989a | 1200 | gimple stmt; |
a509ebb5 | 1201 | |
726a989a | 1202 | gsi = gsi_after_labels (ld_st_data->load_bb); |
a509ebb5 | 1203 | t = build_fold_addr_expr (ld_st_data->store); |
726a989a | 1204 | stmt = gimple_build_assign (ld_st_data->load, t); |
a509ebb5 | 1205 | |
726a989a | 1206 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
a509ebb5 | 1207 | |
4a8fb1a1 | 1208 | reduction_list |
c203e8a7 | 1209 | ->traverse <struct clsn_data *, create_loads_for_reductions> (ld_st_data); |
a509ebb5 RL |
1210 | |
1211 | } | |
1212 | ||
0eb7e7aa RL |
1213 | /* Callback for htab_traverse. Store the neutral value for the |
1214 | particular reduction's operation, e.g. 0 for PLUS_EXPR, | |
1215 | 1 for MULT_EXPR, etc. into the reduction field. | |
b8698a0f L |
1216 | The reduction is specified in SLOT. The store information is |
1217 | passed in DATA. */ | |
0eb7e7aa | 1218 | |
4a8fb1a1 LC |
1219 | int |
1220 | create_stores_for_reduction (reduction_info **slot, struct clsn_data *clsn_data) | |
0eb7e7aa | 1221 | { |
4a8fb1a1 | 1222 | struct reduction_info *const red = *slot; |
726a989a RB |
1223 | tree t; |
1224 | gimple stmt; | |
1225 | gimple_stmt_iterator gsi; | |
1226 | tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); | |
1227 | ||
1228 | gsi = gsi_last_bb (clsn_data->store_bb); | |
1229 | t = build3 (COMPONENT_REF, type, clsn_data->store, red->field, NULL_TREE); | |
1230 | stmt = gimple_build_assign (t, red->initial_value); | |
726a989a | 1231 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
0eb7e7aa RL |
1232 | |
1233 | return 1; | |
1234 | } | |
1235 | ||
a509ebb5 RL |
1236 | /* Callback for htab_traverse. Creates loads to a field of LOAD in LOAD_BB and |
1237 | store to a field of STORE in STORE_BB for the ssa name and its duplicate | |
1238 | specified in SLOT. */ | |
1239 | ||
4a8fb1a1 LC |
1240 | int |
1241 | create_loads_and_stores_for_name (name_to_copy_elt **slot, | |
1242 | struct clsn_data *clsn_data) | |
5f40b3cb | 1243 | { |
4a8fb1a1 | 1244 | struct name_to_copy_elt *const elt = *slot; |
726a989a RB |
1245 | tree t; |
1246 | gimple stmt; | |
1247 | gimple_stmt_iterator gsi; | |
5f40b3cb | 1248 | tree type = TREE_TYPE (elt->new_name); |
5f40b3cb ZD |
1249 | tree load_struct; |
1250 | ||
726a989a RB |
1251 | gsi = gsi_last_bb (clsn_data->store_bb); |
1252 | t = build3 (COMPONENT_REF, type, clsn_data->store, elt->field, NULL_TREE); | |
1253 | stmt = gimple_build_assign (t, ssa_name (elt->version)); | |
726a989a | 1254 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb | 1255 | |
726a989a | 1256 | gsi = gsi_last_bb (clsn_data->load_bb); |
70f34814 | 1257 | load_struct = build_simple_mem_ref (clsn_data->load); |
726a989a RB |
1258 | t = build3 (COMPONENT_REF, type, load_struct, elt->field, NULL_TREE); |
1259 | stmt = gimple_build_assign (elt->new_name, t); | |
726a989a | 1260 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb ZD |
1261 | |
1262 | return 1; | |
1263 | } | |
1264 | ||
1265 | /* Moves all the variables used in LOOP and defined outside of it (including | |
1266 | the initial values of loop phi nodes, and *PER_THREAD if it is a ssa | |
1267 | name) to a structure created for this purpose. The code | |
b8698a0f | 1268 | |
5f40b3cb ZD |
1269 | while (1) |
1270 | { | |
1271 | use (a); | |
1272 | use (b); | |
1273 | } | |
1274 | ||
1275 | is transformed this way: | |
1276 | ||
1277 | bb0: | |
1278 | old.a = a; | |
1279 | old.b = b; | |
1280 | ||
1281 | bb1: | |
1282 | a' = new->a; | |
1283 | b' = new->b; | |
1284 | while (1) | |
1285 | { | |
1286 | use (a'); | |
1287 | use (b'); | |
1288 | } | |
1289 | ||
1290 | `old' is stored to *ARG_STRUCT and `new' is stored to NEW_ARG_STRUCT. The | |
1291 | pointer `new' is intentionally not initialized (the loop will be split to a | |
1292 | separate function later, and `new' will be initialized from its arguments). | |
a509ebb5 | 1293 | LD_ST_DATA holds information about the shared data structure used to pass |
b8698a0f L |
1294 | information among the threads. It is initialized here, and |
1295 | gen_parallel_loop will pass it to create_call_for_reduction that | |
1296 | needs this information. REDUCTION_LIST describes the reductions | |
a509ebb5 | 1297 | in LOOP. */ |
5f40b3cb ZD |
1298 | |
1299 | static void | |
4a8fb1a1 | 1300 | separate_decls_in_region (edge entry, edge exit, |
c203e8a7 | 1301 | reduction_info_table_type *reduction_list, |
b8698a0f | 1302 | tree *arg_struct, tree *new_arg_struct, |
9f9f72aa | 1303 | struct clsn_data *ld_st_data) |
a509ebb5 | 1304 | |
5f40b3cb | 1305 | { |
9f9f72aa | 1306 | basic_block bb1 = split_edge (entry); |
5f40b3cb | 1307 | basic_block bb0 = single_pred (bb1); |
c203e8a7 TS |
1308 | name_to_copy_table_type name_copies (10); |
1309 | int_tree_htab_type decl_copies (10); | |
5f40b3cb | 1310 | unsigned i; |
726a989a RB |
1311 | tree type, type_name, nvar; |
1312 | gimple_stmt_iterator gsi; | |
5f40b3cb | 1313 | struct clsn_data clsn_data; |
00f96dc9 | 1314 | auto_vec<basic_block, 3> body; |
9f9f72aa AP |
1315 | basic_block bb; |
1316 | basic_block entry_bb = bb1; | |
1317 | basic_block exit_bb = exit->dest; | |
b5b8b0ac | 1318 | bool has_debug_stmt = false; |
5f40b3cb | 1319 | |
726a989a | 1320 | entry = single_succ_edge (entry_bb); |
9f9f72aa | 1321 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); |
5f40b3cb | 1322 | |
9771b263 | 1323 | FOR_EACH_VEC_ELT (body, i, bb) |
9f9f72aa | 1324 | { |
b8698a0f | 1325 | if (bb != entry_bb && bb != exit_bb) |
9f9f72aa | 1326 | { |
726a989a RB |
1327 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1328 | separate_decls_in_region_stmt (entry, exit, gsi_stmt (gsi), | |
c203e8a7 | 1329 | &name_copies, &decl_copies); |
726a989a RB |
1330 | |
1331 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
b5b8b0ac AO |
1332 | { |
1333 | gimple stmt = gsi_stmt (gsi); | |
1334 | ||
1335 | if (is_gimple_debug (stmt)) | |
1336 | has_debug_stmt = true; | |
1337 | else | |
1338 | separate_decls_in_region_stmt (entry, exit, stmt, | |
c203e8a7 | 1339 | &name_copies, &decl_copies); |
b5b8b0ac | 1340 | } |
9f9f72aa | 1341 | } |
5f40b3cb | 1342 | } |
9f9f72aa | 1343 | |
b5b8b0ac AO |
1344 | /* Now process debug bind stmts. We must not create decls while |
1345 | processing debug stmts, so we defer their processing so as to | |
1346 | make sure we will have debug info for as many variables as | |
1347 | possible (all of those that were dealt with in the loop above), | |
1348 | and discard those for which we know there's nothing we can | |
1349 | do. */ | |
1350 | if (has_debug_stmt) | |
9771b263 | 1351 | FOR_EACH_VEC_ELT (body, i, bb) |
b5b8b0ac AO |
1352 | if (bb != entry_bb && bb != exit_bb) |
1353 | { | |
1354 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) | |
1355 | { | |
1356 | gimple stmt = gsi_stmt (gsi); | |
1357 | ||
ddb555ed | 1358 | if (is_gimple_debug (stmt)) |
b5b8b0ac | 1359 | { |
c203e8a7 TS |
1360 | if (separate_decls_in_region_debug (stmt, &name_copies, |
1361 | &decl_copies)) | |
b5b8b0ac AO |
1362 | { |
1363 | gsi_remove (&gsi, true); | |
1364 | continue; | |
1365 | } | |
1366 | } | |
1367 | ||
1368 | gsi_next (&gsi); | |
1369 | } | |
1370 | } | |
1371 | ||
c203e8a7 | 1372 | if (name_copies.elements () == 0 && reduction_list->elements () == 0) |
5f40b3cb ZD |
1373 | { |
1374 | /* It may happen that there is nothing to copy (if there are only | |
a509ebb5 | 1375 | loop carried and external variables in the loop). */ |
5f40b3cb ZD |
1376 | *arg_struct = NULL; |
1377 | *new_arg_struct = NULL; | |
1378 | } | |
1379 | else | |
1380 | { | |
1381 | /* Create the type for the structure to store the ssa names to. */ | |
1382 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
9ff70652 | 1383 | type_name = build_decl (UNKNOWN_LOCATION, |
c2255bc4 | 1384 | TYPE_DECL, create_tmp_var_name (".paral_data"), |
5f40b3cb ZD |
1385 | type); |
1386 | TYPE_NAME (type) = type_name; | |
1387 | ||
4a8fb1a1 | 1388 | name_copies.traverse <tree, add_field_for_name> (type); |
c203e8a7 | 1389 | if (reduction_list && reduction_list->elements () > 0) |
0eb7e7aa RL |
1390 | { |
1391 | /* Create the fields for reductions. */ | |
c203e8a7 | 1392 | reduction_list->traverse <tree, add_field_for_reduction> (type); |
0eb7e7aa | 1393 | } |
5f40b3cb | 1394 | layout_type (type); |
b8698a0f | 1395 | |
5f40b3cb ZD |
1396 | /* Create the loads and stores. */ |
1397 | *arg_struct = create_tmp_var (type, ".paral_data_store"); | |
5f40b3cb | 1398 | nvar = create_tmp_var (build_pointer_type (type), ".paral_data_load"); |
b731b390 | 1399 | *new_arg_struct = make_ssa_name (nvar); |
5f40b3cb | 1400 | |
a509ebb5 RL |
1401 | ld_st_data->store = *arg_struct; |
1402 | ld_st_data->load = *new_arg_struct; | |
1403 | ld_st_data->store_bb = bb0; | |
1404 | ld_st_data->load_bb = bb1; | |
0eb7e7aa | 1405 | |
4a8fb1a1 LC |
1406 | name_copies |
1407 | .traverse <struct clsn_data *, create_loads_and_stores_for_name> | |
1408 | (ld_st_data); | |
a509ebb5 | 1409 | |
ae0bce62 RL |
1410 | /* Load the calculation from memory (after the join of the threads). */ |
1411 | ||
c203e8a7 | 1412 | if (reduction_list && reduction_list->elements () > 0) |
a509ebb5 | 1413 | { |
4a8fb1a1 | 1414 | reduction_list |
c203e8a7 TS |
1415 | ->traverse <struct clsn_data *, create_stores_for_reduction> |
1416 | (ld_st_data); | |
b731b390 | 1417 | clsn_data.load = make_ssa_name (nvar); |
9f9f72aa | 1418 | clsn_data.load_bb = exit->dest; |
a509ebb5 RL |
1419 | clsn_data.store = ld_st_data->store; |
1420 | create_final_loads_for_reduction (reduction_list, &clsn_data); | |
1421 | } | |
5f40b3cb | 1422 | } |
5f40b3cb ZD |
1423 | } |
1424 | ||
a79b7ec5 | 1425 | /* Returns true if FN was created to run in parallel. */ |
5f40b3cb | 1426 | |
62e0a1ed | 1427 | bool |
a79b7ec5 | 1428 | parallelized_function_p (tree fndecl) |
5f40b3cb | 1429 | { |
a79b7ec5 TV |
1430 | cgraph_node *node = cgraph_node::get (fndecl); |
1431 | gcc_assert (node != NULL); | |
1432 | return node->parallelized_function; | |
5f40b3cb ZD |
1433 | } |
1434 | ||
1435 | /* Creates and returns an empty function that will receive the body of | |
1436 | a parallelized loop. */ | |
1437 | ||
1438 | static tree | |
9ff70652 | 1439 | create_loop_fn (location_t loc) |
5f40b3cb ZD |
1440 | { |
1441 | char buf[100]; | |
1442 | char *tname; | |
1443 | tree decl, type, name, t; | |
1444 | struct function *act_cfun = cfun; | |
1445 | static unsigned loopfn_num; | |
1446 | ||
5368224f | 1447 | loc = LOCATION_LOCUS (loc); |
5f40b3cb ZD |
1448 | snprintf (buf, 100, "%s.$loopfn", current_function_name ()); |
1449 | ASM_FORMAT_PRIVATE_NAME (tname, buf, loopfn_num++); | |
1450 | clean_symbol_name (tname); | |
1451 | name = get_identifier (tname); | |
1452 | type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); | |
1453 | ||
9ff70652 | 1454 | decl = build_decl (loc, FUNCTION_DECL, name, type); |
5f40b3cb ZD |
1455 | TREE_STATIC (decl) = 1; |
1456 | TREE_USED (decl) = 1; | |
1457 | DECL_ARTIFICIAL (decl) = 1; | |
1458 | DECL_IGNORED_P (decl) = 0; | |
1459 | TREE_PUBLIC (decl) = 0; | |
1460 | DECL_UNINLINABLE (decl) = 1; | |
1461 | DECL_EXTERNAL (decl) = 0; | |
1462 | DECL_CONTEXT (decl) = NULL_TREE; | |
1463 | DECL_INITIAL (decl) = make_node (BLOCK); | |
1464 | ||
9ff70652 | 1465 | t = build_decl (loc, RESULT_DECL, NULL_TREE, void_type_node); |
5f40b3cb ZD |
1466 | DECL_ARTIFICIAL (t) = 1; |
1467 | DECL_IGNORED_P (t) = 1; | |
1468 | DECL_RESULT (decl) = t; | |
1469 | ||
9ff70652 | 1470 | t = build_decl (loc, PARM_DECL, get_identifier (".paral_data_param"), |
5f40b3cb ZD |
1471 | ptr_type_node); |
1472 | DECL_ARTIFICIAL (t) = 1; | |
1473 | DECL_ARG_TYPE (t) = ptr_type_node; | |
1474 | DECL_CONTEXT (t) = decl; | |
1475 | TREE_USED (t) = 1; | |
1476 | DECL_ARGUMENTS (decl) = t; | |
1477 | ||
182e0d71 | 1478 | allocate_struct_function (decl, false); |
5f40b3cb ZD |
1479 | |
1480 | /* The call to allocate_struct_function clobbers CFUN, so we need to restore | |
1481 | it. */ | |
5576d6f2 | 1482 | set_cfun (act_cfun); |
5f40b3cb ZD |
1483 | |
1484 | return decl; | |
1485 | } | |
1486 | ||
5f40b3cb ZD |
1487 | /* Moves the exit condition of LOOP to the beginning of its header, and |
1488 | duplicates the part of the last iteration that gets disabled to the | |
1489 | exit of the loop. NIT is the number of iterations of the loop | |
1490 | (used to initialize the variables in the duplicated part). | |
b8698a0f | 1491 | |
fa10beec | 1492 | TODO: the common case is that latch of the loop is empty and immediately |
5f40b3cb ZD |
1493 | follows the loop exit. In this case, it would be better not to copy the |
1494 | body of the loop, but only move the entry of the loop directly before the | |
1495 | exit check and increase the number of iterations of the loop by one. | |
b8698a0f | 1496 | This may need some additional preconditioning in case NIT = ~0. |
a509ebb5 | 1497 | REDUCTION_LIST describes the reductions in LOOP. */ |
5f40b3cb ZD |
1498 | |
1499 | static void | |
4a8fb1a1 | 1500 | transform_to_exit_first_loop (struct loop *loop, |
c203e8a7 | 1501 | reduction_info_table_type *reduction_list, |
4a8fb1a1 | 1502 | tree nit) |
5f40b3cb ZD |
1503 | { |
1504 | basic_block *bbs, *nbbs, ex_bb, orig_header; | |
1505 | unsigned n; | |
1506 | bool ok; | |
1507 | edge exit = single_dom_exit (loop), hpred; | |
726a989a | 1508 | tree control, control_name, res, t; |
538dd0b7 DM |
1509 | gphi *phi, *nphi; |
1510 | gassign *stmt; | |
1511 | gcond *cond_stmt, *cond_nit; | |
48710229 | 1512 | tree nit_1; |
5f40b3cb ZD |
1513 | |
1514 | split_block_after_labels (loop->header); | |
1515 | orig_header = single_succ (loop->header); | |
1516 | hpred = single_succ_edge (loop->header); | |
1517 | ||
538dd0b7 | 1518 | cond_stmt = as_a <gcond *> (last_stmt (exit->src)); |
726a989a RB |
1519 | control = gimple_cond_lhs (cond_stmt); |
1520 | gcc_assert (gimple_cond_rhs (cond_stmt) == nit); | |
5f40b3cb ZD |
1521 | |
1522 | /* Make sure that we have phi nodes on exit for all loop header phis | |
1523 | (create_parallel_loop requires that). */ | |
538dd0b7 DM |
1524 | for (gphi_iterator gsi = gsi_start_phis (loop->header); |
1525 | !gsi_end_p (gsi); | |
1526 | gsi_next (&gsi)) | |
5f40b3cb | 1527 | { |
538dd0b7 | 1528 | phi = gsi.phi (); |
5f40b3cb | 1529 | res = PHI_RESULT (phi); |
070ecdfd | 1530 | t = copy_ssa_name (res, phi); |
5f40b3cb | 1531 | SET_PHI_RESULT (phi, t); |
5f40b3cb | 1532 | nphi = create_phi_node (res, orig_header); |
9e227d60 | 1533 | add_phi_arg (nphi, t, hpred, UNKNOWN_LOCATION); |
5f40b3cb ZD |
1534 | |
1535 | if (res == control) | |
1536 | { | |
726a989a | 1537 | gimple_cond_set_lhs (cond_stmt, t); |
5f40b3cb ZD |
1538 | update_stmt (cond_stmt); |
1539 | control = t; | |
1540 | } | |
1541 | } | |
12037899 | 1542 | |
5f40b3cb | 1543 | bbs = get_loop_body_in_dom_order (loop); |
48710229 | 1544 | |
69958396 RL |
1545 | for (n = 0; bbs[n] != exit->src; n++) |
1546 | continue; | |
5f40b3cb | 1547 | nbbs = XNEWVEC (basic_block, n); |
726a989a RB |
1548 | ok = gimple_duplicate_sese_tail (single_succ_edge (loop->header), exit, |
1549 | bbs + 1, n, nbbs); | |
5f40b3cb ZD |
1550 | gcc_assert (ok); |
1551 | free (bbs); | |
1552 | ex_bb = nbbs[0]; | |
1553 | free (nbbs); | |
1554 | ||
b8698a0f | 1555 | /* Other than reductions, the only gimple reg that should be copied |
726a989a | 1556 | out of the loop is the control variable. */ |
69958396 | 1557 | exit = single_dom_exit (loop); |
5f40b3cb | 1558 | control_name = NULL_TREE; |
538dd0b7 DM |
1559 | for (gphi_iterator gsi = gsi_start_phis (ex_bb); |
1560 | !gsi_end_p (gsi); ) | |
5f40b3cb | 1561 | { |
538dd0b7 | 1562 | phi = gsi.phi (); |
5f40b3cb | 1563 | res = PHI_RESULT (phi); |
ea057359 | 1564 | if (virtual_operand_p (res)) |
726a989a RB |
1565 | { |
1566 | gsi_next (&gsi); | |
1567 | continue; | |
1568 | } | |
5f40b3cb | 1569 | |
a509ebb5 | 1570 | /* Check if it is a part of reduction. If it is, |
b8698a0f L |
1571 | keep the phi at the reduction's keep_res field. The |
1572 | PHI_RESULT of this phi is the resulting value of the reduction | |
a509ebb5 RL |
1573 | variable when exiting the loop. */ |
1574 | ||
c203e8a7 | 1575 | if (reduction_list->elements () > 0) |
a509ebb5 RL |
1576 | { |
1577 | struct reduction_info *red; | |
1578 | ||
1579 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
a509ebb5 RL |
1580 | red = reduction_phi (reduction_list, SSA_NAME_DEF_STMT (val)); |
1581 | if (red) | |
726a989a RB |
1582 | { |
1583 | red->keep_res = phi; | |
1584 | gsi_next (&gsi); | |
1585 | continue; | |
1586 | } | |
a509ebb5 | 1587 | } |
726a989a RB |
1588 | gcc_assert (control_name == NULL_TREE |
1589 | && SSA_NAME_VAR (res) == SSA_NAME_VAR (control)); | |
5f40b3cb | 1590 | control_name = res; |
726a989a | 1591 | remove_phi_node (&gsi, false); |
5f40b3cb ZD |
1592 | } |
1593 | gcc_assert (control_name != NULL_TREE); | |
5f40b3cb | 1594 | |
b8698a0f | 1595 | /* Initialize the control variable to number of iterations |
48710229 | 1596 | according to the rhs of the exit condition. */ |
538dd0b7 DM |
1597 | gimple_stmt_iterator gsi = gsi_after_labels (ex_bb); |
1598 | cond_nit = as_a <gcond *> (last_stmt (exit->src)); | |
48710229 RL |
1599 | nit_1 = gimple_cond_rhs (cond_nit); |
1600 | nit_1 = force_gimple_operand_gsi (&gsi, | |
1601 | fold_convert (TREE_TYPE (control_name), nit_1), | |
726a989a | 1602 | false, NULL_TREE, false, GSI_SAME_STMT); |
48710229 | 1603 | stmt = gimple_build_assign (control_name, nit_1); |
726a989a | 1604 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb ZD |
1605 | } |
1606 | ||
1607 | /* Create the parallel constructs for LOOP as described in gen_parallel_loop. | |
726a989a | 1608 | LOOP_FN and DATA are the arguments of GIMPLE_OMP_PARALLEL. |
5f40b3cb ZD |
1609 | NEW_DATA is the variable that should be initialized from the argument |
1610 | of LOOP_FN. N_THREADS is the requested number of threads. Returns the | |
726a989a | 1611 | basic block containing GIMPLE_OMP_PARALLEL tree. */ |
5f40b3cb ZD |
1612 | |
1613 | static basic_block | |
1614 | create_parallel_loop (struct loop *loop, tree loop_fn, tree data, | |
9ff70652 | 1615 | tree new_data, unsigned n_threads, location_t loc) |
5f40b3cb | 1616 | { |
726a989a | 1617 | gimple_stmt_iterator gsi; |
5f40b3cb | 1618 | basic_block bb, paral_bb, for_bb, ex_bb; |
0f900dfa | 1619 | tree t, param; |
538dd0b7 DM |
1620 | gomp_parallel *omp_par_stmt; |
1621 | gimple omp_return_stmt1, omp_return_stmt2; | |
1622 | gimple phi; | |
1623 | gcond *cond_stmt; | |
1624 | gomp_for *for_stmt; | |
1625 | gomp_continue *omp_cont_stmt; | |
726a989a | 1626 | tree cvar, cvar_init, initvar, cvar_next, cvar_base, type; |
5f40b3cb ZD |
1627 | edge exit, nexit, guard, end, e; |
1628 | ||
726a989a | 1629 | /* Prepare the GIMPLE_OMP_PARALLEL statement. */ |
5f40b3cb ZD |
1630 | bb = loop_preheader_edge (loop)->src; |
1631 | paral_bb = single_pred (bb); | |
726a989a | 1632 | gsi = gsi_last_bb (paral_bb); |
5f40b3cb | 1633 | |
9ff70652 | 1634 | t = build_omp_clause (loc, OMP_CLAUSE_NUM_THREADS); |
5f40b3cb | 1635 | OMP_CLAUSE_NUM_THREADS_EXPR (t) |
a509ebb5 | 1636 | = build_int_cst (integer_type_node, n_threads); |
538dd0b7 DM |
1637 | omp_par_stmt = gimple_build_omp_parallel (NULL, t, loop_fn, data); |
1638 | gimple_set_location (omp_par_stmt, loc); | |
5f40b3cb | 1639 | |
538dd0b7 | 1640 | gsi_insert_after (&gsi, omp_par_stmt, GSI_NEW_STMT); |
5f40b3cb ZD |
1641 | |
1642 | /* Initialize NEW_DATA. */ | |
1643 | if (data) | |
1644 | { | |
538dd0b7 DM |
1645 | gassign *assign_stmt; |
1646 | ||
726a989a RB |
1647 | gsi = gsi_after_labels (bb); |
1648 | ||
b731b390 | 1649 | param = make_ssa_name (DECL_ARGUMENTS (loop_fn)); |
538dd0b7 DM |
1650 | assign_stmt = gimple_build_assign (param, build_fold_addr_expr (data)); |
1651 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
726a989a | 1652 | |
538dd0b7 | 1653 | assign_stmt = gimple_build_assign (new_data, |
726a989a | 1654 | fold_convert (TREE_TYPE (new_data), param)); |
538dd0b7 | 1655 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
5f40b3cb ZD |
1656 | } |
1657 | ||
726a989a | 1658 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_PARALLEL. */ |
5f40b3cb | 1659 | bb = split_loop_exit_edge (single_dom_exit (loop)); |
726a989a | 1660 | gsi = gsi_last_bb (bb); |
538dd0b7 DM |
1661 | omp_return_stmt1 = gimple_build_omp_return (false); |
1662 | gimple_set_location (omp_return_stmt1, loc); | |
1663 | gsi_insert_after (&gsi, omp_return_stmt1, GSI_NEW_STMT); | |
5f40b3cb | 1664 | |
726a989a | 1665 | /* Extract data for GIMPLE_OMP_FOR. */ |
5f40b3cb | 1666 | gcc_assert (loop->header == single_dom_exit (loop)->src); |
538dd0b7 | 1667 | cond_stmt = as_a <gcond *> (last_stmt (loop->header)); |
5f40b3cb | 1668 | |
726a989a | 1669 | cvar = gimple_cond_lhs (cond_stmt); |
5f40b3cb ZD |
1670 | cvar_base = SSA_NAME_VAR (cvar); |
1671 | phi = SSA_NAME_DEF_STMT (cvar); | |
1672 | cvar_init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
b731b390 | 1673 | initvar = copy_ssa_name (cvar); |
5f40b3cb ZD |
1674 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, loop_preheader_edge (loop)), |
1675 | initvar); | |
1676 | cvar_next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
1677 | ||
1dff453d | 1678 | gsi = gsi_last_nondebug_bb (loop->latch); |
726a989a RB |
1679 | gcc_assert (gsi_stmt (gsi) == SSA_NAME_DEF_STMT (cvar_next)); |
1680 | gsi_remove (&gsi, true); | |
5f40b3cb ZD |
1681 | |
1682 | /* Prepare cfg. */ | |
1683 | for_bb = split_edge (loop_preheader_edge (loop)); | |
1684 | ex_bb = split_loop_exit_edge (single_dom_exit (loop)); | |
1685 | extract_true_false_edges_from_block (loop->header, &nexit, &exit); | |
1686 | gcc_assert (exit == single_dom_exit (loop)); | |
1687 | ||
1688 | guard = make_edge (for_bb, ex_bb, 0); | |
1689 | single_succ_edge (loop->latch)->flags = 0; | |
1690 | end = make_edge (loop->latch, ex_bb, EDGE_FALLTHRU); | |
538dd0b7 DM |
1691 | for (gphi_iterator gpi = gsi_start_phis (ex_bb); |
1692 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
5f40b3cb | 1693 | { |
f5045c96 AM |
1694 | source_location locus; |
1695 | tree def; | |
538dd0b7 DM |
1696 | gphi *phi = gpi.phi (); |
1697 | gphi *stmt; | |
1698 | ||
1699 | stmt = as_a <gphi *> ( | |
1700 | SSA_NAME_DEF_STMT (PHI_ARG_DEF_FROM_EDGE (phi, exit))); | |
f5045c96 AM |
1701 | |
1702 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_preheader_edge (loop)); | |
b8698a0f | 1703 | locus = gimple_phi_arg_location_from_edge (stmt, |
f5045c96 | 1704 | loop_preheader_edge (loop)); |
9e227d60 | 1705 | add_phi_arg (phi, def, guard, locus); |
f5045c96 AM |
1706 | |
1707 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (loop)); | |
1708 | locus = gimple_phi_arg_location_from_edge (stmt, loop_latch_edge (loop)); | |
9e227d60 | 1709 | add_phi_arg (phi, def, end, locus); |
5f40b3cb ZD |
1710 | } |
1711 | e = redirect_edge_and_branch (exit, nexit->dest); | |
1712 | PENDING_STMT (e) = NULL; | |
1713 | ||
726a989a RB |
1714 | /* Emit GIMPLE_OMP_FOR. */ |
1715 | gimple_cond_set_lhs (cond_stmt, cvar_base); | |
5f40b3cb | 1716 | type = TREE_TYPE (cvar); |
9ff70652 | 1717 | t = build_omp_clause (loc, OMP_CLAUSE_SCHEDULE); |
5f40b3cb ZD |
1718 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_STATIC; |
1719 | ||
74bf76ed | 1720 | for_stmt = gimple_build_omp_for (NULL, GF_OMP_FOR_KIND_FOR, t, 1, NULL); |
9ff70652 | 1721 | gimple_set_location (for_stmt, loc); |
726a989a RB |
1722 | gimple_omp_for_set_index (for_stmt, 0, initvar); |
1723 | gimple_omp_for_set_initial (for_stmt, 0, cvar_init); | |
1724 | gimple_omp_for_set_final (for_stmt, 0, gimple_cond_rhs (cond_stmt)); | |
1725 | gimple_omp_for_set_cond (for_stmt, 0, gimple_cond_code (cond_stmt)); | |
1726 | gimple_omp_for_set_incr (for_stmt, 0, build2 (PLUS_EXPR, type, | |
1727 | cvar_base, | |
1728 | build_int_cst (type, 1))); | |
1729 | ||
1730 | gsi = gsi_last_bb (for_bb); | |
1731 | gsi_insert_after (&gsi, for_stmt, GSI_NEW_STMT); | |
5f40b3cb ZD |
1732 | SSA_NAME_DEF_STMT (initvar) = for_stmt; |
1733 | ||
726a989a RB |
1734 | /* Emit GIMPLE_OMP_CONTINUE. */ |
1735 | gsi = gsi_last_bb (loop->latch); | |
538dd0b7 DM |
1736 | omp_cont_stmt = gimple_build_omp_continue (cvar_next, cvar); |
1737 | gimple_set_location (omp_cont_stmt, loc); | |
1738 | gsi_insert_after (&gsi, omp_cont_stmt, GSI_NEW_STMT); | |
1739 | SSA_NAME_DEF_STMT (cvar_next) = omp_cont_stmt; | |
5f40b3cb | 1740 | |
726a989a RB |
1741 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_FOR. */ |
1742 | gsi = gsi_last_bb (ex_bb); | |
538dd0b7 DM |
1743 | omp_return_stmt2 = gimple_build_omp_return (true); |
1744 | gimple_set_location (omp_return_stmt2, loc); | |
1745 | gsi_insert_after (&gsi, omp_return_stmt2, GSI_NEW_STMT); | |
5f40b3cb | 1746 | |
cd7d9fd7 RG |
1747 | /* After the above dom info is hosed. Re-compute it. */ |
1748 | free_dominance_info (CDI_DOMINATORS); | |
1749 | calculate_dominance_info (CDI_DOMINATORS); | |
1750 | ||
5f40b3cb ZD |
1751 | return paral_bb; |
1752 | } | |
1753 | ||
08dab97a RL |
1754 | /* Generates code to execute the iterations of LOOP in N_THREADS |
1755 | threads in parallel. | |
1756 | ||
1757 | NITER describes number of iterations of LOOP. | |
fa10beec | 1758 | REDUCTION_LIST describes the reductions existent in the LOOP. */ |
5f40b3cb ZD |
1759 | |
1760 | static void | |
c203e8a7 TS |
1761 | gen_parallel_loop (struct loop *loop, |
1762 | reduction_info_table_type *reduction_list, | |
a509ebb5 | 1763 | unsigned n_threads, struct tree_niter_desc *niter) |
5f40b3cb | 1764 | { |
5f40b3cb | 1765 | tree many_iterations_cond, type, nit; |
726a989a RB |
1766 | tree arg_struct, new_arg_struct; |
1767 | gimple_seq stmts; | |
9f9f72aa | 1768 | edge entry, exit; |
a509ebb5 | 1769 | struct clsn_data clsn_data; |
5f40b3cb | 1770 | unsigned prob; |
9ff70652 JJ |
1771 | location_t loc; |
1772 | gimple cond_stmt; | |
768da0da | 1773 | unsigned int m_p_thread=2; |
5f40b3cb ZD |
1774 | |
1775 | /* From | |
1776 | ||
1777 | --------------------------------------------------------------------- | |
1778 | loop | |
1779 | { | |
1780 | IV = phi (INIT, IV + STEP) | |
1781 | BODY1; | |
1782 | if (COND) | |
1783 | break; | |
1784 | BODY2; | |
1785 | } | |
1786 | --------------------------------------------------------------------- | |
1787 | ||
1788 | with # of iterations NITER (possibly with MAY_BE_ZERO assumption), | |
1789 | we generate the following code: | |
1790 | ||
1791 | --------------------------------------------------------------------- | |
1792 | ||
1793 | if (MAY_BE_ZERO | |
a509ebb5 RL |
1794 | || NITER < MIN_PER_THREAD * N_THREADS) |
1795 | goto original; | |
5f40b3cb ZD |
1796 | |
1797 | BODY1; | |
1798 | store all local loop-invariant variables used in body of the loop to DATA. | |
726a989a | 1799 | GIMPLE_OMP_PARALLEL (OMP_CLAUSE_NUM_THREADS (N_THREADS), LOOPFN, DATA); |
5f40b3cb | 1800 | load the variables from DATA. |
726a989a | 1801 | GIMPLE_OMP_FOR (IV = INIT; COND; IV += STEP) (OMP_CLAUSE_SCHEDULE (static)) |
5f40b3cb ZD |
1802 | BODY2; |
1803 | BODY1; | |
726a989a RB |
1804 | GIMPLE_OMP_CONTINUE; |
1805 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_FOR | |
1806 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_PARALLEL | |
5f40b3cb ZD |
1807 | goto end; |
1808 | ||
1809 | original: | |
1810 | loop | |
1811 | { | |
1812 | IV = phi (INIT, IV + STEP) | |
1813 | BODY1; | |
1814 | if (COND) | |
1815 | break; | |
1816 | BODY2; | |
1817 | } | |
1818 | ||
1819 | end: | |
1820 | ||
1821 | */ | |
1822 | ||
1823 | /* Create two versions of the loop -- in the old one, we know that the | |
1824 | number of iterations is large enough, and we will transform it into the | |
1825 | loop that will be split to loop_fn, the new one will be used for the | |
1826 | remaining iterations. */ | |
a509ebb5 | 1827 | |
768da0da RL |
1828 | /* We should compute a better number-of-iterations value for outer loops. |
1829 | That is, if we have | |
1830 | ||
1831 | for (i = 0; i < n; ++i) | |
1832 | for (j = 0; j < m; ++j) | |
1833 | ... | |
1834 | ||
1835 | we should compute nit = n * m, not nit = n. | |
1836 | Also may_be_zero handling would need to be adjusted. */ | |
1837 | ||
5f40b3cb ZD |
1838 | type = TREE_TYPE (niter->niter); |
1839 | nit = force_gimple_operand (unshare_expr (niter->niter), &stmts, true, | |
1840 | NULL_TREE); | |
1841 | if (stmts) | |
726a989a | 1842 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb | 1843 | |
768da0da RL |
1844 | if (loop->inner) |
1845 | m_p_thread=2; | |
1846 | else | |
1847 | m_p_thread=MIN_PER_THREAD; | |
1848 | ||
1849 | many_iterations_cond = | |
1850 | fold_build2 (GE_EXPR, boolean_type_node, | |
1851 | nit, build_int_cst (type, m_p_thread * n_threads)); | |
1852 | ||
5f40b3cb | 1853 | many_iterations_cond |
a509ebb5 RL |
1854 | = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1855 | invert_truthvalue (unshare_expr (niter->may_be_zero)), | |
1856 | many_iterations_cond); | |
5f40b3cb | 1857 | many_iterations_cond |
a509ebb5 | 1858 | = force_gimple_operand (many_iterations_cond, &stmts, false, NULL_TREE); |
5f40b3cb | 1859 | if (stmts) |
726a989a | 1860 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb ZD |
1861 | if (!is_gimple_condexpr (many_iterations_cond)) |
1862 | { | |
1863 | many_iterations_cond | |
a509ebb5 RL |
1864 | = force_gimple_operand (many_iterations_cond, &stmts, |
1865 | true, NULL_TREE); | |
5f40b3cb | 1866 | if (stmts) |
726a989a | 1867 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb ZD |
1868 | } |
1869 | ||
1870 | initialize_original_copy_tables (); | |
1871 | ||
1872 | /* We assume that the loop usually iterates a lot. */ | |
1873 | prob = 4 * REG_BR_PROB_BASE / 5; | |
0f900dfa JJ |
1874 | loop_version (loop, many_iterations_cond, NULL, |
1875 | prob, prob, REG_BR_PROB_BASE - prob, true); | |
5f40b3cb ZD |
1876 | update_ssa (TODO_update_ssa); |
1877 | free_original_copy_tables (); | |
1878 | ||
1879 | /* Base all the induction variables in LOOP on a single control one. */ | |
c80a5403 | 1880 | canonicalize_loop_ivs (loop, &nit, true); |
5f40b3cb ZD |
1881 | |
1882 | /* Ensure that the exit condition is the first statement in the loop. */ | |
a509ebb5 RL |
1883 | transform_to_exit_first_loop (loop, reduction_list, nit); |
1884 | ||
fa10beec | 1885 | /* Generate initializations for reductions. */ |
c203e8a7 TS |
1886 | if (reduction_list->elements () > 0) |
1887 | reduction_list->traverse <struct loop *, initialize_reductions> (loop); | |
5f40b3cb ZD |
1888 | |
1889 | /* Eliminate the references to local variables from the loop. */ | |
9f9f72aa AP |
1890 | gcc_assert (single_exit (loop)); |
1891 | entry = loop_preheader_edge (loop); | |
1892 | exit = single_dom_exit (loop); | |
5f40b3cb | 1893 | |
9f9f72aa | 1894 | eliminate_local_variables (entry, exit); |
5f40b3cb ZD |
1895 | /* In the old loop, move all variables non-local to the loop to a structure |
1896 | and back, and create separate decls for the variables used in loop. */ | |
b8698a0f | 1897 | separate_decls_in_region (entry, exit, reduction_list, &arg_struct, |
9f9f72aa | 1898 | &new_arg_struct, &clsn_data); |
5f40b3cb ZD |
1899 | |
1900 | /* Create the parallel constructs. */ | |
9ff70652 JJ |
1901 | loc = UNKNOWN_LOCATION; |
1902 | cond_stmt = last_stmt (loop->header); | |
1903 | if (cond_stmt) | |
1904 | loc = gimple_location (cond_stmt); | |
18751894 TV |
1905 | create_parallel_loop (loop, create_loop_fn (loc), arg_struct, |
1906 | new_arg_struct, n_threads, loc); | |
c203e8a7 | 1907 | if (reduction_list->elements () > 0) |
a509ebb5 | 1908 | create_call_for_reduction (loop, reduction_list, &clsn_data); |
5f40b3cb ZD |
1909 | |
1910 | scev_reset (); | |
1911 | ||
1912 | /* Cancel the loop (it is simpler to do it here rather than to teach the | |
1913 | expander to do it). */ | |
1914 | cancel_loop_tree (loop); | |
1915 | ||
92a6bdbd SP |
1916 | /* Free loop bound estimations that could contain references to |
1917 | removed statements. */ | |
f0bd40b1 | 1918 | FOR_EACH_LOOP (loop, 0) |
92a6bdbd | 1919 | free_numbers_of_iterations_estimates_loop (loop); |
5f40b3cb ZD |
1920 | } |
1921 | ||
9857228c SP |
1922 | /* Returns true when LOOP contains vector phi nodes. */ |
1923 | ||
1924 | static bool | |
726a989a | 1925 | loop_has_vector_phi_nodes (struct loop *loop ATTRIBUTE_UNUSED) |
9857228c SP |
1926 | { |
1927 | unsigned i; | |
1928 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
538dd0b7 | 1929 | gphi_iterator gsi; |
9857228c | 1930 | bool res = true; |
9857228c SP |
1931 | |
1932 | for (i = 0; i < loop->num_nodes; i++) | |
726a989a | 1933 | for (gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) |
538dd0b7 | 1934 | if (TREE_CODE (TREE_TYPE (PHI_RESULT (gsi.phi ()))) == VECTOR_TYPE) |
9857228c SP |
1935 | goto end; |
1936 | ||
1937 | res = false; | |
1938 | end: | |
1939 | free (bbs); | |
1940 | return res; | |
1941 | } | |
1942 | ||
08dab97a RL |
1943 | /* Create a reduction_info struct, initialize it with REDUC_STMT |
1944 | and PHI, insert it to the REDUCTION_LIST. */ | |
1945 | ||
1946 | static void | |
c203e8a7 | 1947 | build_new_reduction (reduction_info_table_type *reduction_list, |
538dd0b7 | 1948 | gimple reduc_stmt, gphi *phi) |
08dab97a | 1949 | { |
4a8fb1a1 | 1950 | reduction_info **slot; |
08dab97a RL |
1951 | struct reduction_info *new_reduction; |
1952 | ||
1953 | gcc_assert (reduc_stmt); | |
b8698a0f | 1954 | |
08dab97a RL |
1955 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1956 | { | |
1957 | fprintf (dump_file, | |
1958 | "Detected reduction. reduction stmt is: \n"); | |
1959 | print_gimple_stmt (dump_file, reduc_stmt, 0, 0); | |
1960 | fprintf (dump_file, "\n"); | |
1961 | } | |
b8698a0f | 1962 | |
08dab97a | 1963 | new_reduction = XCNEW (struct reduction_info); |
b8698a0f | 1964 | |
08dab97a RL |
1965 | new_reduction->reduc_stmt = reduc_stmt; |
1966 | new_reduction->reduc_phi = phi; | |
5d1fd1de | 1967 | new_reduction->reduc_version = SSA_NAME_VERSION (gimple_phi_result (phi)); |
08dab97a | 1968 | new_reduction->reduction_code = gimple_assign_rhs_code (reduc_stmt); |
c203e8a7 | 1969 | slot = reduction_list->find_slot (new_reduction, INSERT); |
08dab97a RL |
1970 | *slot = new_reduction; |
1971 | } | |
1972 | ||
5d1fd1de JJ |
1973 | /* Callback for htab_traverse. Sets gimple_uid of reduc_phi stmts. */ |
1974 | ||
4a8fb1a1 LC |
1975 | int |
1976 | set_reduc_phi_uids (reduction_info **slot, void *data ATTRIBUTE_UNUSED) | |
5d1fd1de | 1977 | { |
4a8fb1a1 | 1978 | struct reduction_info *const red = *slot; |
5d1fd1de JJ |
1979 | gimple_set_uid (red->reduc_phi, red->reduc_version); |
1980 | return 1; | |
1981 | } | |
1982 | ||
08dab97a RL |
1983 | /* Detect all reductions in the LOOP, insert them into REDUCTION_LIST. */ |
1984 | ||
1985 | static void | |
c203e8a7 | 1986 | gather_scalar_reductions (loop_p loop, reduction_info_table_type *reduction_list) |
08dab97a | 1987 | { |
538dd0b7 | 1988 | gphi_iterator gsi; |
08dab97a RL |
1989 | loop_vec_info simple_loop_info; |
1990 | ||
08dab97a RL |
1991 | simple_loop_info = vect_analyze_loop_form (loop); |
1992 | ||
1993 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1994 | { | |
538dd0b7 | 1995 | gphi *phi = gsi.phi (); |
08dab97a RL |
1996 | affine_iv iv; |
1997 | tree res = PHI_RESULT (phi); | |
1998 | bool double_reduc; | |
1999 | ||
ea057359 | 2000 | if (virtual_operand_p (res)) |
08dab97a RL |
2001 | continue; |
2002 | ||
2003 | if (!simple_iv (loop, loop, res, &iv, true) | |
2004 | && simple_loop_info) | |
2005 | { | |
8a9ecffd MM |
2006 | gimple reduc_stmt = vect_force_simple_reduction (simple_loop_info, |
2007 | phi, true, | |
2008 | &double_reduc); | |
48710229 | 2009 | if (reduc_stmt && !double_reduc) |
08dab97a RL |
2010 | build_new_reduction (reduction_list, reduc_stmt, phi); |
2011 | } | |
2012 | } | |
5d1fd1de JJ |
2013 | destroy_loop_vec_info (simple_loop_info, true); |
2014 | ||
2015 | /* As gimple_uid is used by the vectorizer in between vect_analyze_loop_form | |
2016 | and destroy_loop_vec_info, we can set gimple_uid of reduc_phi stmts | |
2017 | only now. */ | |
c203e8a7 | 2018 | reduction_list->traverse <void *, set_reduc_phi_uids> (NULL); |
08dab97a RL |
2019 | } |
2020 | ||
2021 | /* Try to initialize NITER for code generation part. */ | |
2022 | ||
2023 | static bool | |
2024 | try_get_loop_niter (loop_p loop, struct tree_niter_desc *niter) | |
2025 | { | |
2026 | edge exit = single_dom_exit (loop); | |
2027 | ||
2028 | gcc_assert (exit); | |
2029 | ||
2030 | /* We need to know # of iterations, and there should be no uses of values | |
2031 | defined inside loop outside of it, unless the values are invariants of | |
2032 | the loop. */ | |
2033 | if (!number_of_iterations_exit (loop, exit, niter, false)) | |
2034 | { | |
2035 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2036 | fprintf (dump_file, " FAILED: number of iterations not known\n"); | |
2037 | return false; | |
2038 | } | |
2039 | ||
2040 | return true; | |
2041 | } | |
2042 | ||
2043 | /* Try to initialize REDUCTION_LIST for code generation part. | |
2044 | REDUCTION_LIST describes the reductions. */ | |
2045 | ||
2046 | static bool | |
4a8fb1a1 | 2047 | try_create_reduction_list (loop_p loop, |
c203e8a7 | 2048 | reduction_info_table_type *reduction_list) |
08dab97a RL |
2049 | { |
2050 | edge exit = single_dom_exit (loop); | |
538dd0b7 | 2051 | gphi_iterator gsi; |
08dab97a RL |
2052 | |
2053 | gcc_assert (exit); | |
2054 | ||
2055 | gather_scalar_reductions (loop, reduction_list); | |
2056 | ||
b8698a0f | 2057 | |
08dab97a RL |
2058 | for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) |
2059 | { | |
538dd0b7 | 2060 | gphi *phi = gsi.phi (); |
08dab97a RL |
2061 | struct reduction_info *red; |
2062 | imm_use_iterator imm_iter; | |
2063 | use_operand_p use_p; | |
2064 | gimple reduc_phi; | |
2065 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
2066 | ||
ea057359 | 2067 | if (!virtual_operand_p (val)) |
08dab97a RL |
2068 | { |
2069 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2070 | { | |
2071 | fprintf (dump_file, "phi is "); | |
2072 | print_gimple_stmt (dump_file, phi, 0, 0); | |
2073 | fprintf (dump_file, "arg of phi to exit: value "); | |
2074 | print_generic_expr (dump_file, val, 0); | |
2075 | fprintf (dump_file, " used outside loop\n"); | |
2076 | fprintf (dump_file, | |
2077 | " checking if it a part of reduction pattern: \n"); | |
2078 | } | |
c203e8a7 | 2079 | if (reduction_list->elements () == 0) |
08dab97a RL |
2080 | { |
2081 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2082 | fprintf (dump_file, | |
2083 | " FAILED: it is not a part of reduction.\n"); | |
2084 | return false; | |
2085 | } | |
2086 | reduc_phi = NULL; | |
2087 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, val) | |
2088 | { | |
4942af9b JJ |
2089 | if (!gimple_debug_bind_p (USE_STMT (use_p)) |
2090 | && flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p)))) | |
08dab97a RL |
2091 | { |
2092 | reduc_phi = USE_STMT (use_p); | |
2093 | break; | |
2094 | } | |
2095 | } | |
2096 | red = reduction_phi (reduction_list, reduc_phi); | |
2097 | if (red == NULL) | |
2098 | { | |
2099 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2100 | fprintf (dump_file, | |
2101 | " FAILED: it is not a part of reduction.\n"); | |
2102 | return false; | |
2103 | } | |
2104 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2105 | { | |
2106 | fprintf (dump_file, "reduction phi is "); | |
2107 | print_gimple_stmt (dump_file, red->reduc_phi, 0, 0); | |
2108 | fprintf (dump_file, "reduction stmt is "); | |
2109 | print_gimple_stmt (dump_file, red->reduc_stmt, 0, 0); | |
2110 | } | |
2111 | } | |
2112 | } | |
2113 | ||
2114 | /* The iterations of the loop may communicate only through bivs whose | |
2115 | iteration space can be distributed efficiently. */ | |
2116 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2117 | { | |
538dd0b7 | 2118 | gphi *phi = gsi.phi (); |
08dab97a RL |
2119 | tree def = PHI_RESULT (phi); |
2120 | affine_iv iv; | |
2121 | ||
ea057359 | 2122 | if (!virtual_operand_p (def) && !simple_iv (loop, loop, def, &iv, true)) |
08dab97a RL |
2123 | { |
2124 | struct reduction_info *red; | |
2125 | ||
2126 | red = reduction_phi (reduction_list, phi); | |
2127 | if (red == NULL) | |
2128 | { | |
2129 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2130 | fprintf (dump_file, | |
2131 | " FAILED: scalar dependency between iterations\n"); | |
2132 | return false; | |
2133 | } | |
2134 | } | |
2135 | } | |
2136 | ||
2137 | ||
2138 | return true; | |
2139 | } | |
2140 | ||
5f40b3cb ZD |
2141 | /* Detect parallel loops and generate parallel code using libgomp |
2142 | primitives. Returns true if some loop was parallelized, false | |
2143 | otherwise. */ | |
2144 | ||
09489eb8 | 2145 | static bool |
5f40b3cb ZD |
2146 | parallelize_loops (void) |
2147 | { | |
2148 | unsigned n_threads = flag_tree_parallelize_loops; | |
2149 | bool changed = false; | |
2150 | struct loop *loop; | |
2151 | struct tree_niter_desc niter_desc; | |
f873b205 | 2152 | struct obstack parloop_obstack; |
8adfe01d | 2153 | HOST_WIDE_INT estimated; |
b05e0233 | 2154 | source_location loop_loc; |
f873b205 | 2155 | |
5f40b3cb ZD |
2156 | /* Do not parallelize loops in the functions created by parallelization. */ |
2157 | if (parallelized_function_p (cfun->decl)) | |
2158 | return false; | |
8adfe01d RL |
2159 | if (cfun->has_nonlocal_label) |
2160 | return false; | |
5f40b3cb | 2161 | |
f873b205 | 2162 | gcc_obstack_init (&parloop_obstack); |
c203e8a7 | 2163 | reduction_info_table_type reduction_list (10); |
726a989a | 2164 | init_stmt_vec_info_vec (); |
a509ebb5 | 2165 | |
f0bd40b1 | 2166 | FOR_EACH_LOOP (loop, 0) |
5f40b3cb | 2167 | { |
4a8fb1a1 | 2168 | reduction_list.empty (); |
48710229 RL |
2169 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2170 | { | |
2171 | fprintf (dump_file, "Trying loop %d as candidate\n",loop->num); | |
2172 | if (loop->inner) | |
2173 | fprintf (dump_file, "loop %d is not innermost\n",loop->num); | |
2174 | else | |
2175 | fprintf (dump_file, "loop %d is innermost\n",loop->num); | |
2176 | } | |
b8698a0f | 2177 | |
48710229 | 2178 | /* If we use autopar in graphite pass, we use its marked dependency |
87d4d0ee SP |
2179 | checking results. */ |
2180 | if (flag_loop_parallelize_all && !loop->can_be_parallel) | |
48710229 RL |
2181 | { |
2182 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2183 | fprintf (dump_file, "loop is not parallel according to graphite\n"); | |
87d4d0ee | 2184 | continue; |
48710229 | 2185 | } |
87d4d0ee | 2186 | |
48710229 RL |
2187 | if (!single_dom_exit (loop)) |
2188 | { | |
b8698a0f | 2189 | |
48710229 RL |
2190 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2191 | fprintf (dump_file, "loop is !single_dom_exit\n"); | |
b8698a0f | 2192 | |
08dab97a | 2193 | continue; |
48710229 | 2194 | } |
08dab97a RL |
2195 | |
2196 | if (/* And of course, the loop must be parallelizable. */ | |
2197 | !can_duplicate_loop_p (loop) | |
1d4af1e8 | 2198 | || loop_has_blocks_with_irreducible_flag (loop) |
8adfe01d | 2199 | || (loop_preheader_edge (loop)->src->flags & BB_IRREDUCIBLE_LOOP) |
9857228c | 2200 | /* FIXME: the check for vector phi nodes could be removed. */ |
69958396 | 2201 | || loop_has_vector_phi_nodes (loop)) |
08dab97a | 2202 | continue; |
e5b332cd | 2203 | |
652c4c71 | 2204 | estimated = estimated_stmt_executions_int (loop); |
e5b332cd RG |
2205 | if (estimated == -1) |
2206 | estimated = max_stmt_executions_int (loop); | |
87d4d0ee | 2207 | /* FIXME: Bypass this check as graphite doesn't update the |
e5b332cd | 2208 | count and frequency correctly now. */ |
87d4d0ee | 2209 | if (!flag_loop_parallelize_all |
e5b332cd RG |
2210 | && ((estimated != -1 |
2211 | && estimated <= (HOST_WIDE_INT) n_threads * MIN_PER_THREAD) | |
87d4d0ee SP |
2212 | /* Do not bother with loops in cold areas. */ |
2213 | || optimize_loop_nest_for_size_p (loop))) | |
08dab97a | 2214 | continue; |
b8698a0f | 2215 | |
08dab97a RL |
2216 | if (!try_get_loop_niter (loop, &niter_desc)) |
2217 | continue; | |
2218 | ||
c203e8a7 | 2219 | if (!try_create_reduction_list (loop, &reduction_list)) |
08dab97a RL |
2220 | continue; |
2221 | ||
f873b205 LB |
2222 | if (!flag_loop_parallelize_all |
2223 | && !loop_parallel_p (loop, &parloop_obstack)) | |
5f40b3cb ZD |
2224 | continue; |
2225 | ||
2226 | changed = true; | |
48710229 RL |
2227 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2228 | { | |
48710229 | 2229 | if (loop->inner) |
8adfe01d | 2230 | fprintf (dump_file, "parallelizing outer loop %d\n",loop->header->index); |
48710229 | 2231 | else |
8adfe01d RL |
2232 | fprintf (dump_file, "parallelizing inner loop %d\n",loop->header->index); |
2233 | loop_loc = find_loop_location (loop); | |
b05e0233 | 2234 | if (loop_loc != UNKNOWN_LOCATION) |
8adfe01d | 2235 | fprintf (dump_file, "\nloop at %s:%d: ", |
b05e0233 | 2236 | LOCATION_FILE (loop_loc), LOCATION_LINE (loop_loc)); |
b8698a0f | 2237 | } |
c203e8a7 | 2238 | gen_parallel_loop (loop, &reduction_list, |
08dab97a | 2239 | n_threads, &niter_desc); |
5f40b3cb ZD |
2240 | } |
2241 | ||
726a989a | 2242 | free_stmt_vec_info_vec (); |
f873b205 | 2243 | obstack_free (&parloop_obstack, NULL); |
6b8ed145 RG |
2244 | |
2245 | /* Parallelization will cause new function calls to be inserted through | |
d086d311 RG |
2246 | which local variables will escape. Reset the points-to solution |
2247 | for ESCAPED. */ | |
6b8ed145 | 2248 | if (changed) |
d086d311 | 2249 | pt_solution_reset (&cfun->gimple_df->escaped); |
6b8ed145 | 2250 | |
5f40b3cb ZD |
2251 | return changed; |
2252 | } | |
2253 | ||
c1bf2a39 AM |
2254 | /* Parallelization. */ |
2255 | ||
c1bf2a39 AM |
2256 | namespace { |
2257 | ||
2258 | const pass_data pass_data_parallelize_loops = | |
2259 | { | |
2260 | GIMPLE_PASS, /* type */ | |
2261 | "parloops", /* name */ | |
2262 | OPTGROUP_LOOP, /* optinfo_flags */ | |
c1bf2a39 AM |
2263 | TV_TREE_PARALLELIZE_LOOPS, /* tv_id */ |
2264 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2265 | 0, /* properties_provided */ | |
2266 | 0, /* properties_destroyed */ | |
2267 | 0, /* todo_flags_start */ | |
3bea341f | 2268 | 0, /* todo_flags_finish */ |
c1bf2a39 AM |
2269 | }; |
2270 | ||
2271 | class pass_parallelize_loops : public gimple_opt_pass | |
2272 | { | |
2273 | public: | |
2274 | pass_parallelize_loops (gcc::context *ctxt) | |
2275 | : gimple_opt_pass (pass_data_parallelize_loops, ctxt) | |
2276 | {} | |
2277 | ||
2278 | /* opt_pass methods: */ | |
1a3d085c | 2279 | virtual bool gate (function *) { return flag_tree_parallelize_loops > 1; } |
be55bfe6 | 2280 | virtual unsigned int execute (function *); |
c1bf2a39 AM |
2281 | |
2282 | }; // class pass_parallelize_loops | |
2283 | ||
be55bfe6 TS |
2284 | unsigned |
2285 | pass_parallelize_loops::execute (function *fun) | |
2286 | { | |
2287 | if (number_of_loops (fun) <= 1) | |
2288 | return 0; | |
2289 | ||
2290 | if (parallelize_loops ()) | |
18751894 TV |
2291 | { |
2292 | fun->curr_properties &= ~(PROP_gimple_eomp); | |
2293 | return TODO_update_ssa; | |
2294 | } | |
2295 | ||
be55bfe6 TS |
2296 | return 0; |
2297 | } | |
2298 | ||
c1bf2a39 AM |
2299 | } // anon namespace |
2300 | ||
2301 | gimple_opt_pass * | |
2302 | make_pass_parallelize_loops (gcc::context *ctxt) | |
2303 | { | |
2304 | return new pass_parallelize_loops (ctxt); | |
2305 | } |