]>
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" | |
40e23961 | 25 | #include "alias.h" |
c7131fb2 | 26 | #include "backend.h" |
40e23961 | 27 | #include "tree.h" |
c7131fb2 | 28 | #include "gimple.h" |
60393bbc | 29 | #include "hard-reg-set.h" |
c7131fb2 AM |
30 | #include "ssa.h" |
31 | #include "options.h" | |
32 | #include "fold-const.h" | |
2fb9a547 | 33 | #include "internal-fn.h" |
45b0be94 | 34 | #include "gimplify.h" |
5be5c238 | 35 | #include "gimple-iterator.h" |
18f429e2 | 36 | #include "gimplify-me.h" |
5be5c238 | 37 | #include "gimple-walk.h" |
d8a2d370 DN |
38 | #include "stor-layout.h" |
39 | #include "tree-nested.h" | |
442b4905 | 40 | #include "tree-cfg.h" |
e28030cf AM |
41 | #include "tree-ssa-loop-ivopts.h" |
42 | #include "tree-ssa-loop-manip.h" | |
43 | #include "tree-ssa-loop-niter.h" | |
442b4905 AM |
44 | #include "tree-ssa-loop.h" |
45 | #include "tree-into-ssa.h" | |
5f40b3cb | 46 | #include "cfgloop.h" |
5f40b3cb | 47 | #include "tree-data-ref.h" |
1bd6497c | 48 | #include "tree-scalar-evolution.h" |
cf835838 | 49 | #include "gimple-pretty-print.h" |
5f40b3cb | 50 | #include "tree-pass.h" |
5f40b3cb | 51 | #include "langhooks.h" |
a509ebb5 | 52 | #include "tree-vectorizer.h" |
4a8fb1a1 | 53 | #include "tree-hasher.h" |
c1bf2a39 | 54 | #include "tree-parloops.h" |
0645c1a2 | 55 | #include "omp-low.h" |
1fe37220 | 56 | #include "tree-nested.h" |
a79b7ec5 | 57 | #include "cgraph.h" |
7c82d827 | 58 | #include "tree-ssa.h" |
5f40b3cb ZD |
59 | |
60 | /* This pass tries to distribute iterations of loops into several threads. | |
61 | The implementation is straightforward -- for each loop we test whether its | |
62 | iterations are independent, and if it is the case (and some additional | |
63 | conditions regarding profitability and correctness are satisfied), we | |
726a989a RB |
64 | add GIMPLE_OMP_PARALLEL and GIMPLE_OMP_FOR codes and let omp expansion |
65 | machinery do its job. | |
b8698a0f | 66 | |
5f40b3cb ZD |
67 | The most of the complexity is in bringing the code into shape expected |
68 | by the omp expanders: | |
726a989a RB |
69 | -- for GIMPLE_OMP_FOR, ensuring that the loop has only one induction |
70 | variable and that the exit test is at the start of the loop body | |
71 | -- for GIMPLE_OMP_PARALLEL, replacing the references to local addressable | |
5f40b3cb ZD |
72 | variables by accesses through pointers, and breaking up ssa chains |
73 | by storing the values incoming to the parallelized loop to a structure | |
74 | passed to the new function as an argument (something similar is done | |
75 | in omp gimplification, unfortunately only a small part of the code | |
76 | can be shared). | |
77 | ||
78 | TODO: | |
79 | -- if there are several parallelizable loops in a function, it may be | |
80 | possible to generate the threads just once (using synchronization to | |
81 | ensure that cross-loop dependences are obeyed). | |
70837b71 RL |
82 | -- handling of common reduction patterns for outer loops. |
83 | ||
84 | More info can also be found at http://gcc.gnu.org/wiki/AutoParInGCC */ | |
b8698a0f | 85 | /* |
a509ebb5 | 86 | Reduction handling: |
8a9ecffd | 87 | currently we use vect_force_simple_reduction() to detect reduction patterns. |
a509ebb5 | 88 | The code transformation will be introduced by an example. |
b8698a0f L |
89 | |
90 | ||
a509ebb5 RL |
91 | parloop |
92 | { | |
93 | int sum=1; | |
94 | ||
0eb7e7aa | 95 | for (i = 0; i < N; i++) |
a509ebb5 RL |
96 | { |
97 | x[i] = i + 3; | |
98 | sum+=x[i]; | |
99 | } | |
100 | } | |
101 | ||
0eb7e7aa | 102 | gimple-like code: |
a509ebb5 RL |
103 | header_bb: |
104 | ||
0eb7e7aa RL |
105 | # sum_29 = PHI <sum_11(5), 1(3)> |
106 | # i_28 = PHI <i_12(5), 0(3)> | |
107 | D.1795_8 = i_28 + 3; | |
108 | x[i_28] = D.1795_8; | |
109 | sum_11 = D.1795_8 + sum_29; | |
110 | i_12 = i_28 + 1; | |
111 | if (N_6(D) > i_12) | |
112 | goto header_bb; | |
113 | ||
a509ebb5 RL |
114 | |
115 | exit_bb: | |
116 | ||
0eb7e7aa RL |
117 | # sum_21 = PHI <sum_11(4)> |
118 | printf (&"%d"[0], sum_21); | |
a509ebb5 RL |
119 | |
120 | ||
121 | after reduction transformation (only relevant parts): | |
122 | ||
123 | parloop | |
124 | { | |
125 | ||
126 | .... | |
127 | ||
0eb7e7aa | 128 | |
fa10beec | 129 | # Storing the initial value given by the user. # |
0eb7e7aa | 130 | |
ae0bce62 | 131 | .paral_data_store.32.sum.27 = 1; |
b8698a0f L |
132 | |
133 | #pragma omp parallel num_threads(4) | |
a509ebb5 | 134 | |
0eb7e7aa | 135 | #pragma omp for schedule(static) |
ae0bce62 RL |
136 | |
137 | # The neutral element corresponding to the particular | |
138 | reduction's operation, e.g. 0 for PLUS_EXPR, | |
139 | 1 for MULT_EXPR, etc. replaces the user's initial value. # | |
140 | ||
141 | # sum.27_29 = PHI <sum.27_11, 0> | |
142 | ||
0eb7e7aa | 143 | sum.27_11 = D.1827_8 + sum.27_29; |
ae0bce62 | 144 | |
726a989a | 145 | GIMPLE_OMP_CONTINUE |
a509ebb5 | 146 | |
0eb7e7aa RL |
147 | # Adding this reduction phi is done at create_phi_for_local_result() # |
148 | # sum.27_56 = PHI <sum.27_11, 0> | |
726a989a | 149 | GIMPLE_OMP_RETURN |
b8698a0f L |
150 | |
151 | # Creating the atomic operation is done at | |
0eb7e7aa | 152 | create_call_for_reduction_1() # |
a509ebb5 | 153 | |
0eb7e7aa RL |
154 | #pragma omp atomic_load |
155 | D.1839_59 = *&.paral_data_load.33_51->reduction.23; | |
156 | D.1840_60 = sum.27_56 + D.1839_59; | |
157 | #pragma omp atomic_store (D.1840_60); | |
b8698a0f | 158 | |
726a989a | 159 | GIMPLE_OMP_RETURN |
b8698a0f | 160 | |
0eb7e7aa RL |
161 | # collecting the result after the join of the threads is done at |
162 | create_loads_for_reductions(). | |
ae0bce62 RL |
163 | The value computed by the threads is loaded from the |
164 | shared struct. # | |
165 | ||
b8698a0f | 166 | |
0eb7e7aa | 167 | .paral_data_load.33_52 = &.paral_data_store.32; |
ae0bce62 | 168 | sum_37 = .paral_data_load.33_52->sum.27; |
0eb7e7aa RL |
169 | sum_43 = D.1795_41 + sum_37; |
170 | ||
171 | exit bb: | |
172 | # sum_21 = PHI <sum_43, sum_26> | |
173 | printf (&"%d"[0], sum_21); | |
174 | ||
175 | ... | |
176 | ||
a509ebb5 RL |
177 | } |
178 | ||
179 | */ | |
180 | ||
5f40b3cb ZD |
181 | /* Minimal number of iterations of a loop that should be executed in each |
182 | thread. */ | |
183 | #define MIN_PER_THREAD 100 | |
184 | ||
b8698a0f | 185 | /* Element of the hashtable, representing a |
a509ebb5 RL |
186 | reduction in the current loop. */ |
187 | struct reduction_info | |
188 | { | |
726a989a RB |
189 | gimple reduc_stmt; /* reduction statement. */ |
190 | gimple reduc_phi; /* The phi node defining the reduction. */ | |
191 | enum tree_code reduction_code;/* code for the reduction operation. */ | |
5d1fd1de JJ |
192 | unsigned reduc_version; /* SSA_NAME_VERSION of original reduc_phi |
193 | result. */ | |
538dd0b7 | 194 | gphi *keep_res; /* The PHI_RESULT of this phi is the resulting value |
a509ebb5 | 195 | of the reduction variable when existing the loop. */ |
ae0bce62 | 196 | tree initial_value; /* The initial value of the reduction var before entering the loop. */ |
a509ebb5 | 197 | tree field; /* the name of the field in the parloop data structure intended for reduction. */ |
a509ebb5 | 198 | tree init; /* reduction initialization value. */ |
538dd0b7 | 199 | gphi *new_phi; /* (helper field) Newly created phi node whose result |
a509ebb5 RL |
200 | will be passed to the atomic operation. Represents |
201 | the local result each thread computed for the reduction | |
202 | operation. */ | |
203 | }; | |
204 | ||
4a8fb1a1 | 205 | /* Reduction info hashtable helpers. */ |
a509ebb5 | 206 | |
95fbe13e | 207 | struct reduction_hasher : free_ptr_hash <reduction_info> |
a509ebb5 | 208 | { |
67f58944 TS |
209 | static inline hashval_t hash (const reduction_info *); |
210 | static inline bool equal (const reduction_info *, const reduction_info *); | |
4a8fb1a1 LC |
211 | }; |
212 | ||
213 | /* Equality and hash functions for hashtab code. */ | |
a509ebb5 | 214 | |
4a8fb1a1 | 215 | inline bool |
67f58944 | 216 | reduction_hasher::equal (const reduction_info *a, const reduction_info *b) |
4a8fb1a1 | 217 | { |
a509ebb5 RL |
218 | return (a->reduc_phi == b->reduc_phi); |
219 | } | |
220 | ||
4a8fb1a1 | 221 | inline hashval_t |
67f58944 | 222 | reduction_hasher::hash (const reduction_info *a) |
a509ebb5 | 223 | { |
5d1fd1de | 224 | return a->reduc_version; |
a509ebb5 RL |
225 | } |
226 | ||
c203e8a7 | 227 | typedef hash_table<reduction_hasher> reduction_info_table_type; |
4a8fb1a1 LC |
228 | |
229 | ||
a509ebb5 | 230 | static struct reduction_info * |
c203e8a7 | 231 | reduction_phi (reduction_info_table_type *reduction_list, gimple phi) |
a509ebb5 RL |
232 | { |
233 | struct reduction_info tmpred, *red; | |
234 | ||
c203e8a7 | 235 | if (reduction_list->elements () == 0 || phi == NULL) |
a509ebb5 RL |
236 | return NULL; |
237 | ||
238 | tmpred.reduc_phi = phi; | |
5d1fd1de | 239 | tmpred.reduc_version = gimple_uid (phi); |
c203e8a7 | 240 | red = reduction_list->find (&tmpred); |
a509ebb5 RL |
241 | |
242 | return red; | |
243 | } | |
244 | ||
5f40b3cb ZD |
245 | /* Element of hashtable of names to copy. */ |
246 | ||
247 | struct name_to_copy_elt | |
248 | { | |
249 | unsigned version; /* The version of the name to copy. */ | |
250 | tree new_name; /* The new name used in the copy. */ | |
251 | tree field; /* The field of the structure used to pass the | |
252 | value. */ | |
253 | }; | |
254 | ||
4a8fb1a1 | 255 | /* Name copies hashtable helpers. */ |
5f40b3cb | 256 | |
95fbe13e | 257 | struct name_to_copy_hasher : free_ptr_hash <name_to_copy_elt> |
5f40b3cb | 258 | { |
67f58944 TS |
259 | static inline hashval_t hash (const name_to_copy_elt *); |
260 | static inline bool equal (const name_to_copy_elt *, const name_to_copy_elt *); | |
4a8fb1a1 LC |
261 | }; |
262 | ||
263 | /* Equality and hash functions for hashtab code. */ | |
5f40b3cb | 264 | |
4a8fb1a1 | 265 | inline bool |
67f58944 | 266 | name_to_copy_hasher::equal (const name_to_copy_elt *a, const name_to_copy_elt *b) |
4a8fb1a1 | 267 | { |
5f40b3cb ZD |
268 | return a->version == b->version; |
269 | } | |
270 | ||
4a8fb1a1 | 271 | inline hashval_t |
67f58944 | 272 | name_to_copy_hasher::hash (const name_to_copy_elt *a) |
5f40b3cb | 273 | { |
5f40b3cb ZD |
274 | return (hashval_t) a->version; |
275 | } | |
276 | ||
c203e8a7 | 277 | typedef hash_table<name_to_copy_hasher> name_to_copy_table_type; |
4a8fb1a1 | 278 | |
b305e3da SP |
279 | /* A transformation matrix, which is a self-contained ROWSIZE x COLSIZE |
280 | matrix. Rather than use floats, we simply keep a single DENOMINATOR that | |
281 | represents the denominator for every element in the matrix. */ | |
282 | typedef struct lambda_trans_matrix_s | |
283 | { | |
284 | lambda_matrix matrix; | |
285 | int rowsize; | |
286 | int colsize; | |
287 | int denominator; | |
288 | } *lambda_trans_matrix; | |
289 | #define LTM_MATRIX(T) ((T)->matrix) | |
290 | #define LTM_ROWSIZE(T) ((T)->rowsize) | |
291 | #define LTM_COLSIZE(T) ((T)->colsize) | |
292 | #define LTM_DENOMINATOR(T) ((T)->denominator) | |
293 | ||
294 | /* Allocate a new transformation matrix. */ | |
295 | ||
296 | static lambda_trans_matrix | |
297 | lambda_trans_matrix_new (int colsize, int rowsize, | |
298 | struct obstack * lambda_obstack) | |
299 | { | |
300 | lambda_trans_matrix ret; | |
301 | ||
302 | ret = (lambda_trans_matrix) | |
303 | obstack_alloc (lambda_obstack, sizeof (struct lambda_trans_matrix_s)); | |
304 | LTM_MATRIX (ret) = lambda_matrix_new (rowsize, colsize, lambda_obstack); | |
305 | LTM_ROWSIZE (ret) = rowsize; | |
306 | LTM_COLSIZE (ret) = colsize; | |
307 | LTM_DENOMINATOR (ret) = 1; | |
308 | return ret; | |
309 | } | |
310 | ||
311 | /* Multiply a vector VEC by a matrix MAT. | |
312 | MAT is an M*N matrix, and VEC is a vector with length N. The result | |
313 | is stored in DEST which must be a vector of length M. */ | |
314 | ||
315 | static void | |
316 | lambda_matrix_vector_mult (lambda_matrix matrix, int m, int n, | |
317 | lambda_vector vec, lambda_vector dest) | |
318 | { | |
319 | int i, j; | |
320 | ||
321 | lambda_vector_clear (dest, m); | |
322 | for (i = 0; i < m; i++) | |
323 | for (j = 0; j < n; j++) | |
324 | dest[i] += matrix[i][j] * vec[j]; | |
325 | } | |
326 | ||
327 | /* Return true if TRANS is a legal transformation matrix that respects | |
328 | the dependence vectors in DISTS and DIRS. The conservative answer | |
329 | is false. | |
330 | ||
331 | "Wolfe proves that a unimodular transformation represented by the | |
332 | matrix T is legal when applied to a loop nest with a set of | |
333 | lexicographically non-negative distance vectors RDG if and only if | |
334 | for each vector d in RDG, (T.d >= 0) is lexicographically positive. | |
335 | i.e.: if and only if it transforms the lexicographically positive | |
336 | distance vectors to lexicographically positive vectors. Note that | |
337 | a unimodular matrix must transform the zero vector (and only it) to | |
338 | the zero vector." S.Muchnick. */ | |
339 | ||
340 | static bool | |
341 | lambda_transform_legal_p (lambda_trans_matrix trans, | |
342 | int nb_loops, | |
9771b263 | 343 | vec<ddr_p> dependence_relations) |
b305e3da SP |
344 | { |
345 | unsigned int i, j; | |
346 | lambda_vector distres; | |
347 | struct data_dependence_relation *ddr; | |
348 | ||
349 | gcc_assert (LTM_COLSIZE (trans) == nb_loops | |
350 | && LTM_ROWSIZE (trans) == nb_loops); | |
351 | ||
352 | /* When there are no dependences, the transformation is correct. */ | |
9771b263 | 353 | if (dependence_relations.length () == 0) |
b305e3da SP |
354 | return true; |
355 | ||
9771b263 | 356 | ddr = dependence_relations[0]; |
b305e3da SP |
357 | if (ddr == NULL) |
358 | return true; | |
359 | ||
360 | /* When there is an unknown relation in the dependence_relations, we | |
361 | know that it is no worth looking at this loop nest: give up. */ | |
362 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
363 | return false; | |
364 | ||
365 | distres = lambda_vector_new (nb_loops); | |
366 | ||
367 | /* For each distance vector in the dependence graph. */ | |
9771b263 | 368 | FOR_EACH_VEC_ELT (dependence_relations, i, ddr) |
b305e3da SP |
369 | { |
370 | /* Don't care about relations for which we know that there is no | |
371 | dependence, nor about read-read (aka. output-dependences): | |
372 | these data accesses can happen in any order. */ | |
373 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known | |
374 | || (DR_IS_READ (DDR_A (ddr)) && DR_IS_READ (DDR_B (ddr)))) | |
375 | continue; | |
376 | ||
377 | /* Conservatively answer: "this transformation is not valid". */ | |
378 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
379 | return false; | |
380 | ||
381 | /* If the dependence could not be captured by a distance vector, | |
382 | conservatively answer that the transform is not valid. */ | |
383 | if (DDR_NUM_DIST_VECTS (ddr) == 0) | |
384 | return false; | |
385 | ||
386 | /* Compute trans.dist_vect */ | |
387 | for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++) | |
388 | { | |
389 | lambda_matrix_vector_mult (LTM_MATRIX (trans), nb_loops, nb_loops, | |
390 | DDR_DIST_VECT (ddr, j), distres); | |
391 | ||
392 | if (!lambda_vector_lexico_pos (distres, nb_loops)) | |
393 | return false; | |
394 | } | |
395 | } | |
396 | return true; | |
397 | } | |
08dab97a RL |
398 | |
399 | /* Data dependency analysis. Returns true if the iterations of LOOP | |
400 | are independent on each other (that is, if we can execute them | |
401 | in parallel). */ | |
5f40b3cb ZD |
402 | |
403 | static bool | |
f873b205 | 404 | loop_parallel_p (struct loop *loop, struct obstack * parloop_obstack) |
5f40b3cb | 405 | { |
9771b263 DN |
406 | vec<ddr_p> dependence_relations; |
407 | vec<data_reference_p> datarefs; | |
5f40b3cb ZD |
408 | lambda_trans_matrix trans; |
409 | bool ret = false; | |
5f40b3cb ZD |
410 | |
411 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
48710229 RL |
412 | { |
413 | fprintf (dump_file, "Considering loop %d\n", loop->num); | |
414 | if (!loop->inner) | |
415 | fprintf (dump_file, "loop is innermost\n"); | |
b8698a0f | 416 | else |
48710229 RL |
417 | fprintf (dump_file, "loop NOT innermost\n"); |
418 | } | |
5f40b3cb | 419 | |
5f40b3cb ZD |
420 | /* Check for problems with dependences. If the loop can be reversed, |
421 | the iterations are independent. */ | |
00f96dc9 | 422 | auto_vec<loop_p, 3> loop_nest; |
9771b263 | 423 | datarefs.create (10); |
07687835 | 424 | dependence_relations.create (100); |
9ca3d00e AB |
425 | if (! compute_data_dependences_for_loop (loop, true, &loop_nest, &datarefs, |
426 | &dependence_relations)) | |
427 | { | |
428 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
429 | fprintf (dump_file, " FAILED: cannot analyze data dependencies\n"); | |
430 | ret = false; | |
431 | goto end; | |
432 | } | |
5f40b3cb ZD |
433 | if (dump_file && (dump_flags & TDF_DETAILS)) |
434 | dump_data_dependence_relations (dump_file, dependence_relations); | |
435 | ||
f873b205 | 436 | trans = lambda_trans_matrix_new (1, 1, parloop_obstack); |
5f40b3cb ZD |
437 | LTM_MATRIX (trans)[0][0] = -1; |
438 | ||
439 | if (lambda_transform_legal_p (trans, 1, dependence_relations)) | |
440 | { | |
441 | ret = true; | |
442 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
443 | fprintf (dump_file, " SUCCESS: may be parallelized\n"); | |
444 | } | |
445 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
a509ebb5 RL |
446 | fprintf (dump_file, |
447 | " FAILED: data dependencies exist across iterations\n"); | |
5f40b3cb | 448 | |
9ca3d00e | 449 | end: |
5f40b3cb ZD |
450 | free_dependence_relations (dependence_relations); |
451 | free_data_refs (datarefs); | |
452 | ||
453 | return ret; | |
454 | } | |
455 | ||
1d4af1e8 SP |
456 | /* Return true when LOOP contains basic blocks marked with the |
457 | BB_IRREDUCIBLE_LOOP flag. */ | |
458 | ||
459 | static inline bool | |
460 | loop_has_blocks_with_irreducible_flag (struct loop *loop) | |
461 | { | |
462 | unsigned i; | |
463 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
464 | bool res = true; | |
465 | ||
466 | for (i = 0; i < loop->num_nodes; i++) | |
467 | if (bbs[i]->flags & BB_IRREDUCIBLE_LOOP) | |
468 | goto end; | |
469 | ||
470 | res = false; | |
471 | end: | |
472 | free (bbs); | |
473 | return res; | |
474 | } | |
475 | ||
8a171a59 | 476 | /* Assigns the address of OBJ in TYPE to an ssa name, and returns this name. |
9f9f72aa | 477 | The assignment statement is placed on edge ENTRY. DECL_ADDRESS maps decls |
8a171a59 | 478 | to their addresses that can be reused. The address of OBJ is known to |
cba1eb61 JJ |
479 | be invariant in the whole function. Other needed statements are placed |
480 | right before GSI. */ | |
5f40b3cb ZD |
481 | |
482 | static tree | |
4a8fb1a1 | 483 | take_address_of (tree obj, tree type, edge entry, |
c203e8a7 | 484 | int_tree_htab_type *decl_address, gimple_stmt_iterator *gsi) |
5f40b3cb | 485 | { |
8a171a59 | 486 | int uid; |
83d5977e | 487 | tree *var_p, name, addr; |
538dd0b7 | 488 | gassign *stmt; |
726a989a | 489 | gimple_seq stmts; |
5f40b3cb | 490 | |
8a171a59 ZD |
491 | /* Since the address of OBJ is invariant, the trees may be shared. |
492 | Avoid rewriting unrelated parts of the code. */ | |
493 | obj = unshare_expr (obj); | |
494 | for (var_p = &obj; | |
495 | handled_component_p (*var_p); | |
496 | var_p = &TREE_OPERAND (*var_p, 0)) | |
497 | continue; | |
8a171a59 | 498 | |
c9a410f0 RG |
499 | /* Canonicalize the access to base on a MEM_REF. */ |
500 | if (DECL_P (*var_p)) | |
501 | *var_p = build_simple_mem_ref (build_fold_addr_expr (*var_p)); | |
502 | ||
503 | /* Assign a canonical SSA name to the address of the base decl used | |
504 | in the address and share it for all accesses and addresses based | |
505 | on it. */ | |
506 | uid = DECL_UID (TREE_OPERAND (TREE_OPERAND (*var_p, 0), 0)); | |
84baa4b9 TS |
507 | int_tree_map elt; |
508 | elt.uid = uid; | |
509 | int_tree_map *slot = decl_address->find_slot (elt, INSERT); | |
510 | if (!slot->to) | |
5f40b3cb | 511 | { |
cba1eb61 JJ |
512 | if (gsi == NULL) |
513 | return NULL; | |
c9a410f0 | 514 | addr = TREE_OPERAND (*var_p, 0); |
29b89442 JJ |
515 | const char *obj_name |
516 | = get_name (TREE_OPERAND (TREE_OPERAND (*var_p, 0), 0)); | |
517 | if (obj_name) | |
518 | name = make_temp_ssa_name (TREE_TYPE (addr), NULL, obj_name); | |
519 | else | |
b731b390 | 520 | name = make_ssa_name (TREE_TYPE (addr)); |
83d5977e | 521 | stmt = gimple_build_assign (name, addr); |
726a989a | 522 | gsi_insert_on_edge_immediate (entry, stmt); |
5f40b3cb | 523 | |
84baa4b9 TS |
524 | slot->uid = uid; |
525 | slot->to = name; | |
5f40b3cb | 526 | } |
8a171a59 | 527 | else |
84baa4b9 | 528 | name = slot->to; |
5f40b3cb | 529 | |
c9a410f0 RG |
530 | /* Express the address in terms of the canonical SSA name. */ |
531 | TREE_OPERAND (*var_p, 0) = name; | |
cba1eb61 JJ |
532 | if (gsi == NULL) |
533 | return build_fold_addr_expr_with_type (obj, type); | |
534 | ||
c9a410f0 RG |
535 | name = force_gimple_operand (build_addr (obj, current_function_decl), |
536 | &stmts, true, NULL_TREE); | |
537 | if (!gimple_seq_empty_p (stmts)) | |
cba1eb61 | 538 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
5f40b3cb | 539 | |
c9a410f0 | 540 | if (!useless_type_conversion_p (type, TREE_TYPE (name))) |
8a171a59 | 541 | { |
726a989a | 542 | name = force_gimple_operand (fold_convert (type, name), &stmts, true, |
8a171a59 | 543 | NULL_TREE); |
726a989a | 544 | if (!gimple_seq_empty_p (stmts)) |
cba1eb61 | 545 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); |
8a171a59 | 546 | } |
5f40b3cb ZD |
547 | |
548 | return name; | |
549 | } | |
550 | ||
a509ebb5 | 551 | /* Callback for htab_traverse. Create the initialization statement |
b8698a0f | 552 | for reduction described in SLOT, and place it at the preheader of |
a509ebb5 RL |
553 | the loop described in DATA. */ |
554 | ||
4a8fb1a1 LC |
555 | int |
556 | initialize_reductions (reduction_info **slot, struct loop *loop) | |
a509ebb5 | 557 | { |
a509ebb5 | 558 | tree init, c; |
a509ebb5 RL |
559 | tree bvar, type, arg; |
560 | edge e; | |
561 | ||
4a8fb1a1 | 562 | struct reduction_info *const reduc = *slot; |
a509ebb5 | 563 | |
b8698a0f | 564 | /* Create initialization in preheader: |
a509ebb5 RL |
565 | reduction_variable = initialization value of reduction. */ |
566 | ||
b8698a0f | 567 | /* In the phi node at the header, replace the argument coming |
a509ebb5 RL |
568 | from the preheader with the reduction initialization value. */ |
569 | ||
570 | /* Create a new variable to initialize the reduction. */ | |
571 | type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); | |
572 | bvar = create_tmp_var (type, "reduction"); | |
a509ebb5 | 573 | |
c2255bc4 AH |
574 | c = build_omp_clause (gimple_location (reduc->reduc_stmt), |
575 | OMP_CLAUSE_REDUCTION); | |
a509ebb5 | 576 | OMP_CLAUSE_REDUCTION_CODE (c) = reduc->reduction_code; |
726a989a | 577 | OMP_CLAUSE_DECL (c) = SSA_NAME_VAR (gimple_assign_lhs (reduc->reduc_stmt)); |
a509ebb5 RL |
578 | |
579 | init = omp_reduction_init (c, TREE_TYPE (bvar)); | |
580 | reduc->init = init; | |
581 | ||
b8698a0f L |
582 | /* Replace the argument representing the initialization value |
583 | with the initialization value for the reduction (neutral | |
584 | element for the particular operation, e.g. 0 for PLUS_EXPR, | |
585 | 1 for MULT_EXPR, etc). | |
586 | Keep the old value in a new variable "reduction_initial", | |
587 | that will be taken in consideration after the parallel | |
0eb7e7aa | 588 | computing is done. */ |
a509ebb5 RL |
589 | |
590 | e = loop_preheader_edge (loop); | |
591 | arg = PHI_ARG_DEF_FROM_EDGE (reduc->reduc_phi, e); | |
592 | /* Create new variable to hold the initial value. */ | |
a509ebb5 | 593 | |
a509ebb5 | 594 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE |
0eb7e7aa | 595 | (reduc->reduc_phi, loop_preheader_edge (loop)), init); |
ae0bce62 | 596 | reduc->initial_value = arg; |
a509ebb5 RL |
597 | return 1; |
598 | } | |
5f40b3cb ZD |
599 | |
600 | struct elv_data | |
601 | { | |
726a989a | 602 | struct walk_stmt_info info; |
9f9f72aa | 603 | edge entry; |
c203e8a7 | 604 | int_tree_htab_type *decl_address; |
cba1eb61 | 605 | gimple_stmt_iterator *gsi; |
5f40b3cb | 606 | bool changed; |
cba1eb61 | 607 | bool reset; |
5f40b3cb ZD |
608 | }; |
609 | ||
9f9f72aa AP |
610 | /* Eliminates references to local variables in *TP out of the single |
611 | entry single exit region starting at DTA->ENTRY. | |
612 | DECL_ADDRESS contains addresses of the references that had their | |
613 | address taken already. If the expression is changed, CHANGED is | |
614 | set to true. Callback for walk_tree. */ | |
a509ebb5 | 615 | |
5f40b3cb | 616 | static tree |
8a171a59 | 617 | eliminate_local_variables_1 (tree *tp, int *walk_subtrees, void *data) |
5f40b3cb | 618 | { |
3d9a9f94 | 619 | struct elv_data *const dta = (struct elv_data *) data; |
8a171a59 | 620 | tree t = *tp, var, addr, addr_type, type, obj; |
5f40b3cb ZD |
621 | |
622 | if (DECL_P (t)) | |
623 | { | |
624 | *walk_subtrees = 0; | |
625 | ||
626 | if (!SSA_VAR_P (t) || DECL_EXTERNAL (t)) | |
627 | return NULL_TREE; | |
628 | ||
629 | type = TREE_TYPE (t); | |
630 | addr_type = build_pointer_type (type); | |
cba1eb61 JJ |
631 | addr = take_address_of (t, addr_type, dta->entry, dta->decl_address, |
632 | dta->gsi); | |
633 | if (dta->gsi == NULL && addr == NULL_TREE) | |
634 | { | |
635 | dta->reset = true; | |
636 | return NULL_TREE; | |
637 | } | |
638 | ||
70f34814 | 639 | *tp = build_simple_mem_ref (addr); |
5f40b3cb ZD |
640 | |
641 | dta->changed = true; | |
642 | return NULL_TREE; | |
643 | } | |
644 | ||
645 | if (TREE_CODE (t) == ADDR_EXPR) | |
646 | { | |
8a171a59 ZD |
647 | /* ADDR_EXPR may appear in two contexts: |
648 | -- as a gimple operand, when the address taken is a function invariant | |
649 | -- as gimple rhs, when the resulting address in not a function | |
650 | invariant | |
651 | We do not need to do anything special in the latter case (the base of | |
652 | the memory reference whose address is taken may be replaced in the | |
653 | DECL_P case). The former case is more complicated, as we need to | |
654 | ensure that the new address is still a gimple operand. Thus, it | |
655 | is not sufficient to replace just the base of the memory reference -- | |
656 | we need to move the whole computation of the address out of the | |
657 | loop. */ | |
658 | if (!is_gimple_val (t)) | |
5f40b3cb ZD |
659 | return NULL_TREE; |
660 | ||
661 | *walk_subtrees = 0; | |
8a171a59 ZD |
662 | obj = TREE_OPERAND (t, 0); |
663 | var = get_base_address (obj); | |
664 | if (!var || !SSA_VAR_P (var) || DECL_EXTERNAL (var)) | |
5f40b3cb ZD |
665 | return NULL_TREE; |
666 | ||
667 | addr_type = TREE_TYPE (t); | |
cba1eb61 JJ |
668 | addr = take_address_of (obj, addr_type, dta->entry, dta->decl_address, |
669 | dta->gsi); | |
670 | if (dta->gsi == NULL && addr == NULL_TREE) | |
671 | { | |
672 | dta->reset = true; | |
673 | return NULL_TREE; | |
674 | } | |
5f40b3cb ZD |
675 | *tp = addr; |
676 | ||
677 | dta->changed = true; | |
678 | return NULL_TREE; | |
679 | } | |
680 | ||
726a989a | 681 | if (!EXPR_P (t)) |
5f40b3cb ZD |
682 | *walk_subtrees = 0; |
683 | ||
684 | return NULL_TREE; | |
685 | } | |
686 | ||
cba1eb61 | 687 | /* Moves the references to local variables in STMT at *GSI out of the single |
9f9f72aa AP |
688 | entry single exit region starting at ENTRY. DECL_ADDRESS contains |
689 | addresses of the references that had their address taken | |
690 | already. */ | |
5f40b3cb ZD |
691 | |
692 | static void | |
cba1eb61 | 693 | eliminate_local_variables_stmt (edge entry, gimple_stmt_iterator *gsi, |
c203e8a7 | 694 | int_tree_htab_type *decl_address) |
5f40b3cb ZD |
695 | { |
696 | struct elv_data dta; | |
cba1eb61 | 697 | gimple stmt = gsi_stmt (*gsi); |
5f40b3cb | 698 | |
726a989a | 699 | memset (&dta.info, '\0', sizeof (dta.info)); |
9f9f72aa | 700 | dta.entry = entry; |
5f40b3cb ZD |
701 | dta.decl_address = decl_address; |
702 | dta.changed = false; | |
cba1eb61 | 703 | dta.reset = false; |
5f40b3cb | 704 | |
b5b8b0ac | 705 | if (gimple_debug_bind_p (stmt)) |
cba1eb61 JJ |
706 | { |
707 | dta.gsi = NULL; | |
708 | walk_tree (gimple_debug_bind_get_value_ptr (stmt), | |
709 | eliminate_local_variables_1, &dta.info, NULL); | |
710 | if (dta.reset) | |
711 | { | |
712 | gimple_debug_bind_reset_value (stmt); | |
713 | dta.changed = true; | |
714 | } | |
715 | } | |
29b89442 JJ |
716 | else if (gimple_clobber_p (stmt)) |
717 | { | |
718 | stmt = gimple_build_nop (); | |
719 | gsi_replace (gsi, stmt, false); | |
720 | dta.changed = true; | |
721 | } | |
b5b8b0ac | 722 | else |
cba1eb61 JJ |
723 | { |
724 | dta.gsi = gsi; | |
725 | walk_gimple_op (stmt, eliminate_local_variables_1, &dta.info); | |
726 | } | |
5f40b3cb ZD |
727 | |
728 | if (dta.changed) | |
729 | update_stmt (stmt); | |
730 | } | |
731 | ||
9f9f72aa AP |
732 | /* Eliminates the references to local variables from the single entry |
733 | single exit region between the ENTRY and EXIT edges. | |
b8698a0f | 734 | |
a509ebb5 | 735 | This includes: |
b8698a0f L |
736 | 1) Taking address of a local variable -- these are moved out of the |
737 | region (and temporary variable is created to hold the address if | |
a509ebb5 | 738 | necessary). |
9f9f72aa | 739 | |
5f40b3cb | 740 | 2) Dereferencing a local variable -- these are replaced with indirect |
a509ebb5 | 741 | references. */ |
5f40b3cb ZD |
742 | |
743 | static void | |
9f9f72aa | 744 | eliminate_local_variables (edge entry, edge exit) |
5f40b3cb | 745 | { |
9f9f72aa | 746 | basic_block bb; |
00f96dc9 | 747 | auto_vec<basic_block, 3> body; |
5f40b3cb | 748 | unsigned i; |
726a989a | 749 | gimple_stmt_iterator gsi; |
cba1eb61 | 750 | bool has_debug_stmt = false; |
c203e8a7 | 751 | int_tree_htab_type decl_address (10); |
9f9f72aa AP |
752 | basic_block entry_bb = entry->src; |
753 | basic_block exit_bb = exit->dest; | |
5f40b3cb | 754 | |
9f9f72aa | 755 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); |
5f40b3cb | 756 | |
9771b263 | 757 | FOR_EACH_VEC_ELT (body, i, bb) |
9f9f72aa | 758 | if (bb != entry_bb && bb != exit_bb) |
726a989a | 759 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
ddb555ed JJ |
760 | if (is_gimple_debug (gsi_stmt (gsi))) |
761 | { | |
762 | if (gimple_debug_bind_p (gsi_stmt (gsi))) | |
763 | has_debug_stmt = true; | |
764 | } | |
cba1eb61 | 765 | else |
c203e8a7 | 766 | eliminate_local_variables_stmt (entry, &gsi, &decl_address); |
cba1eb61 JJ |
767 | |
768 | if (has_debug_stmt) | |
9771b263 | 769 | FOR_EACH_VEC_ELT (body, i, bb) |
cba1eb61 JJ |
770 | if (bb != entry_bb && bb != exit_bb) |
771 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
772 | if (gimple_debug_bind_p (gsi_stmt (gsi))) | |
c203e8a7 | 773 | eliminate_local_variables_stmt (entry, &gsi, &decl_address); |
9f9f72aa AP |
774 | } |
775 | ||
776 | /* Returns true if expression EXPR is not defined between ENTRY and | |
777 | EXIT, i.e. if all its operands are defined outside of the region. */ | |
778 | ||
779 | static bool | |
780 | expr_invariant_in_region_p (edge entry, edge exit, tree expr) | |
781 | { | |
782 | basic_block entry_bb = entry->src; | |
783 | basic_block exit_bb = exit->dest; | |
784 | basic_block def_bb; | |
9f9f72aa AP |
785 | |
786 | if (is_gimple_min_invariant (expr)) | |
787 | return true; | |
788 | ||
789 | if (TREE_CODE (expr) == SSA_NAME) | |
790 | { | |
726a989a | 791 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
9f9f72aa AP |
792 | if (def_bb |
793 | && dominated_by_p (CDI_DOMINATORS, def_bb, entry_bb) | |
794 | && !dominated_by_p (CDI_DOMINATORS, def_bb, exit_bb)) | |
795 | return false; | |
796 | ||
797 | return true; | |
798 | } | |
799 | ||
726a989a | 800 | return false; |
5f40b3cb ZD |
801 | } |
802 | ||
803 | /* If COPY_NAME_P is true, creates and returns a duplicate of NAME. | |
804 | The copies are stored to NAME_COPIES, if NAME was already duplicated, | |
805 | its duplicate stored in NAME_COPIES is returned. | |
b8698a0f | 806 | |
5f40b3cb ZD |
807 | Regardless of COPY_NAME_P, the decl used as a base of the ssa name is also |
808 | duplicated, storing the copies in DECL_COPIES. */ | |
809 | ||
810 | static tree | |
c203e8a7 TS |
811 | separate_decls_in_region_name (tree name, name_to_copy_table_type *name_copies, |
812 | int_tree_htab_type *decl_copies, | |
813 | bool copy_name_p) | |
5f40b3cb ZD |
814 | { |
815 | tree copy, var, var_copy; | |
816 | unsigned idx, uid, nuid; | |
84baa4b9 | 817 | struct int_tree_map ielt; |
5f40b3cb | 818 | struct name_to_copy_elt elt, *nelt; |
4a8fb1a1 | 819 | name_to_copy_elt **slot; |
84baa4b9 | 820 | int_tree_map *dslot; |
5f40b3cb ZD |
821 | |
822 | if (TREE_CODE (name) != SSA_NAME) | |
823 | return name; | |
824 | ||
825 | idx = SSA_NAME_VERSION (name); | |
826 | elt.version = idx; | |
c203e8a7 TS |
827 | slot = name_copies->find_slot_with_hash (&elt, idx, |
828 | copy_name_p ? INSERT : NO_INSERT); | |
5f40b3cb | 829 | if (slot && *slot) |
4a8fb1a1 | 830 | return (*slot)->new_name; |
5f40b3cb | 831 | |
70b5e7dc RG |
832 | if (copy_name_p) |
833 | { | |
834 | copy = duplicate_ssa_name (name, NULL); | |
835 | nelt = XNEW (struct name_to_copy_elt); | |
836 | nelt->version = idx; | |
837 | nelt->new_name = copy; | |
838 | nelt->field = NULL_TREE; | |
839 | *slot = nelt; | |
840 | } | |
841 | else | |
842 | { | |
843 | gcc_assert (!slot); | |
844 | copy = name; | |
845 | } | |
846 | ||
5f40b3cb | 847 | var = SSA_NAME_VAR (name); |
70b5e7dc RG |
848 | if (!var) |
849 | return copy; | |
850 | ||
5f40b3cb ZD |
851 | uid = DECL_UID (var); |
852 | ielt.uid = uid; | |
84baa4b9 TS |
853 | dslot = decl_copies->find_slot_with_hash (ielt, uid, INSERT); |
854 | if (!dslot->to) | |
5f40b3cb ZD |
855 | { |
856 | var_copy = create_tmp_var (TREE_TYPE (var), get_name (var)); | |
36ad7922 | 857 | DECL_GIMPLE_REG_P (var_copy) = DECL_GIMPLE_REG_P (var); |
84baa4b9 TS |
858 | dslot->uid = uid; |
859 | dslot->to = var_copy; | |
5f40b3cb ZD |
860 | |
861 | /* Ensure that when we meet this decl next time, we won't duplicate | |
a509ebb5 | 862 | it again. */ |
5f40b3cb ZD |
863 | nuid = DECL_UID (var_copy); |
864 | ielt.uid = nuid; | |
84baa4b9 TS |
865 | dslot = decl_copies->find_slot_with_hash (ielt, nuid, INSERT); |
866 | gcc_assert (!dslot->to); | |
867 | dslot->uid = nuid; | |
868 | dslot->to = var_copy; | |
5f40b3cb ZD |
869 | } |
870 | else | |
84baa4b9 | 871 | var_copy = dslot->to; |
5f40b3cb | 872 | |
b2ec94d4 | 873 | replace_ssa_name_symbol (copy, var_copy); |
5f40b3cb ZD |
874 | return copy; |
875 | } | |
876 | ||
9f9f72aa AP |
877 | /* Finds the ssa names used in STMT that are defined outside the |
878 | region between ENTRY and EXIT and replaces such ssa names with | |
879 | their duplicates. The duplicates are stored to NAME_COPIES. Base | |
880 | decls of all ssa names used in STMT (including those defined in | |
881 | LOOP) are replaced with the new temporary variables; the | |
882 | replacement decls are stored in DECL_COPIES. */ | |
5f40b3cb ZD |
883 | |
884 | static void | |
726a989a | 885 | separate_decls_in_region_stmt (edge entry, edge exit, gimple stmt, |
c203e8a7 TS |
886 | name_to_copy_table_type *name_copies, |
887 | int_tree_htab_type *decl_copies) | |
5f40b3cb ZD |
888 | { |
889 | use_operand_p use; | |
890 | def_operand_p def; | |
891 | ssa_op_iter oi; | |
892 | tree name, copy; | |
893 | bool copy_name_p; | |
894 | ||
5f40b3cb | 895 | FOR_EACH_PHI_OR_STMT_DEF (def, stmt, oi, SSA_OP_DEF) |
a509ebb5 RL |
896 | { |
897 | name = DEF_FROM_PTR (def); | |
898 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
9f9f72aa AP |
899 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, |
900 | false); | |
a509ebb5 RL |
901 | gcc_assert (copy == name); |
902 | } | |
5f40b3cb ZD |
903 | |
904 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
a509ebb5 RL |
905 | { |
906 | name = USE_FROM_PTR (use); | |
907 | if (TREE_CODE (name) != SSA_NAME) | |
908 | continue; | |
909 | ||
9f9f72aa AP |
910 | copy_name_p = expr_invariant_in_region_p (entry, exit, name); |
911 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, | |
912 | copy_name_p); | |
a509ebb5 RL |
913 | SET_USE (use, copy); |
914 | } | |
5f40b3cb ZD |
915 | } |
916 | ||
b5b8b0ac AO |
917 | /* Finds the ssa names used in STMT that are defined outside the |
918 | region between ENTRY and EXIT and replaces such ssa names with | |
919 | their duplicates. The duplicates are stored to NAME_COPIES. Base | |
920 | decls of all ssa names used in STMT (including those defined in | |
921 | LOOP) are replaced with the new temporary variables; the | |
922 | replacement decls are stored in DECL_COPIES. */ | |
923 | ||
924 | static bool | |
4a8fb1a1 | 925 | separate_decls_in_region_debug (gimple stmt, |
c203e8a7 TS |
926 | name_to_copy_table_type *name_copies, |
927 | int_tree_htab_type *decl_copies) | |
b5b8b0ac AO |
928 | { |
929 | use_operand_p use; | |
930 | ssa_op_iter oi; | |
931 | tree var, name; | |
932 | struct int_tree_map ielt; | |
933 | struct name_to_copy_elt elt; | |
4a8fb1a1 | 934 | name_to_copy_elt **slot; |
84baa4b9 | 935 | int_tree_map *dslot; |
b5b8b0ac | 936 | |
ddb555ed JJ |
937 | if (gimple_debug_bind_p (stmt)) |
938 | var = gimple_debug_bind_get_var (stmt); | |
939 | else if (gimple_debug_source_bind_p (stmt)) | |
940 | var = gimple_debug_source_bind_get_var (stmt); | |
941 | else | |
942 | return true; | |
598e67d7 | 943 | if (TREE_CODE (var) == DEBUG_EXPR_DECL || TREE_CODE (var) == LABEL_DECL) |
4f2a9af8 | 944 | return true; |
b5b8b0ac AO |
945 | gcc_assert (DECL_P (var) && SSA_VAR_P (var)); |
946 | ielt.uid = DECL_UID (var); | |
84baa4b9 | 947 | dslot = decl_copies->find_slot_with_hash (ielt, ielt.uid, NO_INSERT); |
b5b8b0ac AO |
948 | if (!dslot) |
949 | return true; | |
ddb555ed | 950 | if (gimple_debug_bind_p (stmt)) |
84baa4b9 | 951 | gimple_debug_bind_set_var (stmt, dslot->to); |
ddb555ed | 952 | else if (gimple_debug_source_bind_p (stmt)) |
84baa4b9 | 953 | gimple_debug_source_bind_set_var (stmt, dslot->to); |
b5b8b0ac AO |
954 | |
955 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
956 | { | |
957 | name = USE_FROM_PTR (use); | |
958 | if (TREE_CODE (name) != SSA_NAME) | |
959 | continue; | |
960 | ||
961 | elt.version = SSA_NAME_VERSION (name); | |
c203e8a7 | 962 | slot = name_copies->find_slot_with_hash (&elt, elt.version, NO_INSERT); |
b5b8b0ac AO |
963 | if (!slot) |
964 | { | |
965 | gimple_debug_bind_reset_value (stmt); | |
966 | update_stmt (stmt); | |
967 | break; | |
968 | } | |
969 | ||
4a8fb1a1 | 970 | SET_USE (use, (*slot)->new_name); |
b5b8b0ac AO |
971 | } |
972 | ||
973 | return false; | |
974 | } | |
975 | ||
0eb7e7aa RL |
976 | /* Callback for htab_traverse. Adds a field corresponding to the reduction |
977 | specified in SLOT. The type is passed in DATA. */ | |
978 | ||
4a8fb1a1 LC |
979 | int |
980 | add_field_for_reduction (reduction_info **slot, tree type) | |
a509ebb5 | 981 | { |
b8698a0f | 982 | |
4a8fb1a1 | 983 | struct reduction_info *const red = *slot; |
aa06a978 RB |
984 | tree var = gimple_assign_lhs (red->reduc_stmt); |
985 | tree field = build_decl (gimple_location (red->reduc_stmt), FIELD_DECL, | |
986 | SSA_NAME_IDENTIFIER (var), TREE_TYPE (var)); | |
0eb7e7aa RL |
987 | |
988 | insert_field_into_struct (type, field); | |
989 | ||
990 | red->field = field; | |
991 | ||
992 | return 1; | |
993 | } | |
a509ebb5 | 994 | |
5f40b3cb | 995 | /* Callback for htab_traverse. Adds a field corresponding to a ssa name |
b8698a0f | 996 | described in SLOT. The type is passed in DATA. */ |
5f40b3cb | 997 | |
4a8fb1a1 LC |
998 | int |
999 | add_field_for_name (name_to_copy_elt **slot, tree type) | |
5f40b3cb | 1000 | { |
4a8fb1a1 | 1001 | struct name_to_copy_elt *const elt = *slot; |
5f40b3cb | 1002 | tree name = ssa_name (elt->version); |
70b5e7dc RG |
1003 | tree field = build_decl (UNKNOWN_LOCATION, |
1004 | FIELD_DECL, SSA_NAME_IDENTIFIER (name), | |
1005 | TREE_TYPE (name)); | |
5f40b3cb ZD |
1006 | |
1007 | insert_field_into_struct (type, field); | |
1008 | elt->field = field; | |
a509ebb5 | 1009 | |
5f40b3cb ZD |
1010 | return 1; |
1011 | } | |
1012 | ||
b8698a0f L |
1013 | /* Callback for htab_traverse. A local result is the intermediate result |
1014 | computed by a single | |
fa10beec | 1015 | thread, or the initial value in case no iteration was executed. |
b8698a0f L |
1016 | This function creates a phi node reflecting these values. |
1017 | The phi's result will be stored in NEW_PHI field of the | |
1018 | reduction's data structure. */ | |
a509ebb5 | 1019 | |
4a8fb1a1 LC |
1020 | int |
1021 | create_phi_for_local_result (reduction_info **slot, struct loop *loop) | |
a509ebb5 | 1022 | { |
4a8fb1a1 | 1023 | struct reduction_info *const reduc = *slot; |
a509ebb5 | 1024 | edge e; |
538dd0b7 | 1025 | gphi *new_phi; |
a509ebb5 RL |
1026 | basic_block store_bb; |
1027 | tree local_res; | |
f5045c96 | 1028 | source_location locus; |
a509ebb5 | 1029 | |
b8698a0f L |
1030 | /* STORE_BB is the block where the phi |
1031 | should be stored. It is the destination of the loop exit. | |
726a989a | 1032 | (Find the fallthru edge from GIMPLE_OMP_CONTINUE). */ |
a509ebb5 RL |
1033 | store_bb = FALLTHRU_EDGE (loop->latch)->dest; |
1034 | ||
1035 | /* STORE_BB has two predecessors. One coming from the loop | |
1036 | (the reduction's result is computed at the loop), | |
b8698a0f L |
1037 | and another coming from a block preceding the loop, |
1038 | when no iterations | |
1039 | are executed (the initial value should be taken). */ | |
a509ebb5 RL |
1040 | if (EDGE_PRED (store_bb, 0) == FALLTHRU_EDGE (loop->latch)) |
1041 | e = EDGE_PRED (store_bb, 1); | |
1042 | else | |
1043 | e = EDGE_PRED (store_bb, 0); | |
b731b390 | 1044 | local_res = copy_ssa_name (gimple_assign_lhs (reduc->reduc_stmt)); |
f5045c96 | 1045 | locus = gimple_location (reduc->reduc_stmt); |
a509ebb5 | 1046 | new_phi = create_phi_node (local_res, store_bb); |
9e227d60 | 1047 | add_phi_arg (new_phi, reduc->init, e, locus); |
726a989a | 1048 | add_phi_arg (new_phi, gimple_assign_lhs (reduc->reduc_stmt), |
9e227d60 | 1049 | FALLTHRU_EDGE (loop->latch), locus); |
a509ebb5 RL |
1050 | reduc->new_phi = new_phi; |
1051 | ||
1052 | return 1; | |
1053 | } | |
5f40b3cb ZD |
1054 | |
1055 | struct clsn_data | |
1056 | { | |
1057 | tree store; | |
1058 | tree load; | |
1059 | ||
1060 | basic_block store_bb; | |
1061 | basic_block load_bb; | |
1062 | }; | |
1063 | ||
a509ebb5 | 1064 | /* Callback for htab_traverse. Create an atomic instruction for the |
b8698a0f | 1065 | reduction described in SLOT. |
a509ebb5 RL |
1066 | DATA annotates the place in memory the atomic operation relates to, |
1067 | and the basic block it needs to be generated in. */ | |
1068 | ||
4a8fb1a1 LC |
1069 | int |
1070 | create_call_for_reduction_1 (reduction_info **slot, struct clsn_data *clsn_data) | |
a509ebb5 | 1071 | { |
4a8fb1a1 | 1072 | struct reduction_info *const reduc = *slot; |
726a989a | 1073 | gimple_stmt_iterator gsi; |
a509ebb5 | 1074 | tree type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); |
a509ebb5 RL |
1075 | tree load_struct; |
1076 | basic_block bb; | |
1077 | basic_block new_bb; | |
1078 | edge e; | |
0f900dfa | 1079 | tree t, addr, ref, x; |
726a989a RB |
1080 | tree tmp_load, name; |
1081 | gimple load; | |
a509ebb5 | 1082 | |
70f34814 | 1083 | load_struct = build_simple_mem_ref (clsn_data->load); |
a509ebb5 | 1084 | t = build3 (COMPONENT_REF, type, load_struct, reduc->field, NULL_TREE); |
a509ebb5 RL |
1085 | |
1086 | addr = build_addr (t, current_function_decl); | |
1087 | ||
1088 | /* Create phi node. */ | |
1089 | bb = clsn_data->load_bb; | |
1090 | ||
b13c907a RB |
1091 | gsi = gsi_last_bb (bb); |
1092 | e = split_block (bb, gsi_stmt (gsi)); | |
a509ebb5 RL |
1093 | new_bb = e->dest; |
1094 | ||
b731b390 JJ |
1095 | tmp_load = create_tmp_var (TREE_TYPE (TREE_TYPE (addr))); |
1096 | tmp_load = make_ssa_name (tmp_load); | |
726a989a | 1097 | load = gimple_build_omp_atomic_load (tmp_load, addr); |
a509ebb5 | 1098 | SSA_NAME_DEF_STMT (tmp_load) = load; |
726a989a RB |
1099 | gsi = gsi_start_bb (new_bb); |
1100 | gsi_insert_after (&gsi, load, GSI_NEW_STMT); | |
a509ebb5 RL |
1101 | |
1102 | e = split_block (new_bb, load); | |
1103 | new_bb = e->dest; | |
726a989a | 1104 | gsi = gsi_start_bb (new_bb); |
a509ebb5 | 1105 | ref = tmp_load; |
726a989a RB |
1106 | x = fold_build2 (reduc->reduction_code, |
1107 | TREE_TYPE (PHI_RESULT (reduc->new_phi)), ref, | |
1108 | PHI_RESULT (reduc->new_phi)); | |
a509ebb5 | 1109 | |
726a989a RB |
1110 | name = force_gimple_operand_gsi (&gsi, x, true, NULL_TREE, true, |
1111 | GSI_CONTINUE_LINKING); | |
a509ebb5 | 1112 | |
726a989a | 1113 | gsi_insert_after (&gsi, gimple_build_omp_atomic_store (name), GSI_NEW_STMT); |
a509ebb5 RL |
1114 | return 1; |
1115 | } | |
1116 | ||
b8698a0f L |
1117 | /* Create the atomic operation at the join point of the threads. |
1118 | REDUCTION_LIST describes the reductions in the LOOP. | |
1119 | LD_ST_DATA describes the shared data structure where | |
a509ebb5 RL |
1120 | shared data is stored in and loaded from. */ |
1121 | static void | |
4a8fb1a1 | 1122 | create_call_for_reduction (struct loop *loop, |
c203e8a7 | 1123 | reduction_info_table_type *reduction_list, |
a509ebb5 RL |
1124 | struct clsn_data *ld_st_data) |
1125 | { | |
c203e8a7 | 1126 | reduction_list->traverse <struct loop *, create_phi_for_local_result> (loop); |
726a989a | 1127 | /* Find the fallthru edge from GIMPLE_OMP_CONTINUE. */ |
a509ebb5 | 1128 | ld_st_data->load_bb = FALLTHRU_EDGE (loop->latch)->dest; |
4a8fb1a1 | 1129 | reduction_list |
c203e8a7 | 1130 | ->traverse <struct clsn_data *, create_call_for_reduction_1> (ld_st_data); |
a509ebb5 RL |
1131 | } |
1132 | ||
ae0bce62 RL |
1133 | /* Callback for htab_traverse. Loads the final reduction value at the |
1134 | join point of all threads, and inserts it in the right place. */ | |
a509ebb5 | 1135 | |
4a8fb1a1 LC |
1136 | int |
1137 | create_loads_for_reductions (reduction_info **slot, struct clsn_data *clsn_data) | |
a509ebb5 | 1138 | { |
4a8fb1a1 | 1139 | struct reduction_info *const red = *slot; |
726a989a RB |
1140 | gimple stmt; |
1141 | gimple_stmt_iterator gsi; | |
1142 | tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); | |
a509ebb5 | 1143 | tree load_struct; |
ae0bce62 | 1144 | tree name; |
a509ebb5 RL |
1145 | tree x; |
1146 | ||
726a989a | 1147 | gsi = gsi_after_labels (clsn_data->load_bb); |
70f34814 | 1148 | load_struct = build_simple_mem_ref (clsn_data->load); |
a509ebb5 RL |
1149 | load_struct = build3 (COMPONENT_REF, type, load_struct, red->field, |
1150 | NULL_TREE); | |
a509ebb5 | 1151 | |
ae0bce62 | 1152 | x = load_struct; |
a509ebb5 | 1153 | name = PHI_RESULT (red->keep_res); |
726a989a | 1154 | stmt = gimple_build_assign (name, x); |
a509ebb5 | 1155 | |
726a989a | 1156 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
a509ebb5 | 1157 | |
726a989a RB |
1158 | for (gsi = gsi_start_phis (gimple_bb (red->keep_res)); |
1159 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
1160 | if (gsi_stmt (gsi) == red->keep_res) | |
1161 | { | |
1162 | remove_phi_node (&gsi, false); | |
1163 | return 1; | |
1164 | } | |
1165 | gcc_unreachable (); | |
a509ebb5 RL |
1166 | } |
1167 | ||
b8698a0f | 1168 | /* Load the reduction result that was stored in LD_ST_DATA. |
a509ebb5 | 1169 | REDUCTION_LIST describes the list of reductions that the |
fa10beec | 1170 | loads should be generated for. */ |
a509ebb5 | 1171 | static void |
c203e8a7 | 1172 | create_final_loads_for_reduction (reduction_info_table_type *reduction_list, |
a509ebb5 RL |
1173 | struct clsn_data *ld_st_data) |
1174 | { | |
726a989a | 1175 | gimple_stmt_iterator gsi; |
a509ebb5 | 1176 | tree t; |
726a989a | 1177 | gimple stmt; |
a509ebb5 | 1178 | |
726a989a | 1179 | gsi = gsi_after_labels (ld_st_data->load_bb); |
a509ebb5 | 1180 | t = build_fold_addr_expr (ld_st_data->store); |
726a989a | 1181 | stmt = gimple_build_assign (ld_st_data->load, t); |
a509ebb5 | 1182 | |
726a989a | 1183 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
a509ebb5 | 1184 | |
4a8fb1a1 | 1185 | reduction_list |
c203e8a7 | 1186 | ->traverse <struct clsn_data *, create_loads_for_reductions> (ld_st_data); |
a509ebb5 RL |
1187 | |
1188 | } | |
1189 | ||
0eb7e7aa RL |
1190 | /* Callback for htab_traverse. Store the neutral value for the |
1191 | particular reduction's operation, e.g. 0 for PLUS_EXPR, | |
1192 | 1 for MULT_EXPR, etc. into the reduction field. | |
b8698a0f L |
1193 | The reduction is specified in SLOT. The store information is |
1194 | passed in DATA. */ | |
0eb7e7aa | 1195 | |
4a8fb1a1 LC |
1196 | int |
1197 | create_stores_for_reduction (reduction_info **slot, struct clsn_data *clsn_data) | |
0eb7e7aa | 1198 | { |
4a8fb1a1 | 1199 | struct reduction_info *const red = *slot; |
726a989a RB |
1200 | tree t; |
1201 | gimple stmt; | |
1202 | gimple_stmt_iterator gsi; | |
1203 | tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); | |
1204 | ||
1205 | gsi = gsi_last_bb (clsn_data->store_bb); | |
1206 | t = build3 (COMPONENT_REF, type, clsn_data->store, red->field, NULL_TREE); | |
1207 | stmt = gimple_build_assign (t, red->initial_value); | |
726a989a | 1208 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
0eb7e7aa RL |
1209 | |
1210 | return 1; | |
1211 | } | |
1212 | ||
a509ebb5 RL |
1213 | /* Callback for htab_traverse. Creates loads to a field of LOAD in LOAD_BB and |
1214 | store to a field of STORE in STORE_BB for the ssa name and its duplicate | |
1215 | specified in SLOT. */ | |
1216 | ||
4a8fb1a1 LC |
1217 | int |
1218 | create_loads_and_stores_for_name (name_to_copy_elt **slot, | |
1219 | struct clsn_data *clsn_data) | |
5f40b3cb | 1220 | { |
4a8fb1a1 | 1221 | struct name_to_copy_elt *const elt = *slot; |
726a989a RB |
1222 | tree t; |
1223 | gimple stmt; | |
1224 | gimple_stmt_iterator gsi; | |
5f40b3cb | 1225 | tree type = TREE_TYPE (elt->new_name); |
5f40b3cb ZD |
1226 | tree load_struct; |
1227 | ||
726a989a RB |
1228 | gsi = gsi_last_bb (clsn_data->store_bb); |
1229 | t = build3 (COMPONENT_REF, type, clsn_data->store, elt->field, NULL_TREE); | |
1230 | stmt = gimple_build_assign (t, ssa_name (elt->version)); | |
726a989a | 1231 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb | 1232 | |
726a989a | 1233 | gsi = gsi_last_bb (clsn_data->load_bb); |
70f34814 | 1234 | load_struct = build_simple_mem_ref (clsn_data->load); |
726a989a RB |
1235 | t = build3 (COMPONENT_REF, type, load_struct, elt->field, NULL_TREE); |
1236 | stmt = gimple_build_assign (elt->new_name, t); | |
726a989a | 1237 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb ZD |
1238 | |
1239 | return 1; | |
1240 | } | |
1241 | ||
1242 | /* Moves all the variables used in LOOP and defined outside of it (including | |
1243 | the initial values of loop phi nodes, and *PER_THREAD if it is a ssa | |
1244 | name) to a structure created for this purpose. The code | |
b8698a0f | 1245 | |
5f40b3cb ZD |
1246 | while (1) |
1247 | { | |
1248 | use (a); | |
1249 | use (b); | |
1250 | } | |
1251 | ||
1252 | is transformed this way: | |
1253 | ||
1254 | bb0: | |
1255 | old.a = a; | |
1256 | old.b = b; | |
1257 | ||
1258 | bb1: | |
1259 | a' = new->a; | |
1260 | b' = new->b; | |
1261 | while (1) | |
1262 | { | |
1263 | use (a'); | |
1264 | use (b'); | |
1265 | } | |
1266 | ||
1267 | `old' is stored to *ARG_STRUCT and `new' is stored to NEW_ARG_STRUCT. The | |
1268 | pointer `new' is intentionally not initialized (the loop will be split to a | |
1269 | separate function later, and `new' will be initialized from its arguments). | |
a509ebb5 | 1270 | LD_ST_DATA holds information about the shared data structure used to pass |
b8698a0f L |
1271 | information among the threads. It is initialized here, and |
1272 | gen_parallel_loop will pass it to create_call_for_reduction that | |
1273 | needs this information. REDUCTION_LIST describes the reductions | |
a509ebb5 | 1274 | in LOOP. */ |
5f40b3cb ZD |
1275 | |
1276 | static void | |
4a8fb1a1 | 1277 | separate_decls_in_region (edge entry, edge exit, |
c203e8a7 | 1278 | reduction_info_table_type *reduction_list, |
b8698a0f | 1279 | tree *arg_struct, tree *new_arg_struct, |
9f9f72aa | 1280 | struct clsn_data *ld_st_data) |
a509ebb5 | 1281 | |
5f40b3cb | 1282 | { |
9f9f72aa | 1283 | basic_block bb1 = split_edge (entry); |
5f40b3cb | 1284 | basic_block bb0 = single_pred (bb1); |
c203e8a7 TS |
1285 | name_to_copy_table_type name_copies (10); |
1286 | int_tree_htab_type decl_copies (10); | |
5f40b3cb | 1287 | unsigned i; |
726a989a RB |
1288 | tree type, type_name, nvar; |
1289 | gimple_stmt_iterator gsi; | |
5f40b3cb | 1290 | struct clsn_data clsn_data; |
00f96dc9 | 1291 | auto_vec<basic_block, 3> body; |
9f9f72aa AP |
1292 | basic_block bb; |
1293 | basic_block entry_bb = bb1; | |
1294 | basic_block exit_bb = exit->dest; | |
b5b8b0ac | 1295 | bool has_debug_stmt = false; |
5f40b3cb | 1296 | |
726a989a | 1297 | entry = single_succ_edge (entry_bb); |
9f9f72aa | 1298 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); |
5f40b3cb | 1299 | |
9771b263 | 1300 | FOR_EACH_VEC_ELT (body, i, bb) |
9f9f72aa | 1301 | { |
b8698a0f | 1302 | if (bb != entry_bb && bb != exit_bb) |
9f9f72aa | 1303 | { |
726a989a RB |
1304 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1305 | separate_decls_in_region_stmt (entry, exit, gsi_stmt (gsi), | |
c203e8a7 | 1306 | &name_copies, &decl_copies); |
726a989a RB |
1307 | |
1308 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
b5b8b0ac AO |
1309 | { |
1310 | gimple stmt = gsi_stmt (gsi); | |
1311 | ||
1312 | if (is_gimple_debug (stmt)) | |
1313 | has_debug_stmt = true; | |
1314 | else | |
1315 | separate_decls_in_region_stmt (entry, exit, stmt, | |
c203e8a7 | 1316 | &name_copies, &decl_copies); |
b5b8b0ac | 1317 | } |
9f9f72aa | 1318 | } |
5f40b3cb | 1319 | } |
9f9f72aa | 1320 | |
b5b8b0ac AO |
1321 | /* Now process debug bind stmts. We must not create decls while |
1322 | processing debug stmts, so we defer their processing so as to | |
1323 | make sure we will have debug info for as many variables as | |
1324 | possible (all of those that were dealt with in the loop above), | |
1325 | and discard those for which we know there's nothing we can | |
1326 | do. */ | |
1327 | if (has_debug_stmt) | |
9771b263 | 1328 | FOR_EACH_VEC_ELT (body, i, bb) |
b5b8b0ac AO |
1329 | if (bb != entry_bb && bb != exit_bb) |
1330 | { | |
1331 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) | |
1332 | { | |
1333 | gimple stmt = gsi_stmt (gsi); | |
1334 | ||
ddb555ed | 1335 | if (is_gimple_debug (stmt)) |
b5b8b0ac | 1336 | { |
c203e8a7 TS |
1337 | if (separate_decls_in_region_debug (stmt, &name_copies, |
1338 | &decl_copies)) | |
b5b8b0ac AO |
1339 | { |
1340 | gsi_remove (&gsi, true); | |
1341 | continue; | |
1342 | } | |
1343 | } | |
1344 | ||
1345 | gsi_next (&gsi); | |
1346 | } | |
1347 | } | |
1348 | ||
c203e8a7 | 1349 | if (name_copies.elements () == 0 && reduction_list->elements () == 0) |
5f40b3cb ZD |
1350 | { |
1351 | /* It may happen that there is nothing to copy (if there are only | |
a509ebb5 | 1352 | loop carried and external variables in the loop). */ |
5f40b3cb ZD |
1353 | *arg_struct = NULL; |
1354 | *new_arg_struct = NULL; | |
1355 | } | |
1356 | else | |
1357 | { | |
1358 | /* Create the type for the structure to store the ssa names to. */ | |
1359 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
9ff70652 | 1360 | type_name = build_decl (UNKNOWN_LOCATION, |
c2255bc4 | 1361 | TYPE_DECL, create_tmp_var_name (".paral_data"), |
5f40b3cb ZD |
1362 | type); |
1363 | TYPE_NAME (type) = type_name; | |
1364 | ||
4a8fb1a1 | 1365 | name_copies.traverse <tree, add_field_for_name> (type); |
c203e8a7 | 1366 | if (reduction_list && reduction_list->elements () > 0) |
0eb7e7aa RL |
1367 | { |
1368 | /* Create the fields for reductions. */ | |
c203e8a7 | 1369 | reduction_list->traverse <tree, add_field_for_reduction> (type); |
0eb7e7aa | 1370 | } |
5f40b3cb | 1371 | layout_type (type); |
b8698a0f | 1372 | |
5f40b3cb ZD |
1373 | /* Create the loads and stores. */ |
1374 | *arg_struct = create_tmp_var (type, ".paral_data_store"); | |
5f40b3cb | 1375 | nvar = create_tmp_var (build_pointer_type (type), ".paral_data_load"); |
b731b390 | 1376 | *new_arg_struct = make_ssa_name (nvar); |
5f40b3cb | 1377 | |
a509ebb5 RL |
1378 | ld_st_data->store = *arg_struct; |
1379 | ld_st_data->load = *new_arg_struct; | |
1380 | ld_st_data->store_bb = bb0; | |
1381 | ld_st_data->load_bb = bb1; | |
0eb7e7aa | 1382 | |
4a8fb1a1 LC |
1383 | name_copies |
1384 | .traverse <struct clsn_data *, create_loads_and_stores_for_name> | |
1385 | (ld_st_data); | |
a509ebb5 | 1386 | |
ae0bce62 RL |
1387 | /* Load the calculation from memory (after the join of the threads). */ |
1388 | ||
c203e8a7 | 1389 | if (reduction_list && reduction_list->elements () > 0) |
a509ebb5 | 1390 | { |
4a8fb1a1 | 1391 | reduction_list |
c203e8a7 TS |
1392 | ->traverse <struct clsn_data *, create_stores_for_reduction> |
1393 | (ld_st_data); | |
b731b390 | 1394 | clsn_data.load = make_ssa_name (nvar); |
9f9f72aa | 1395 | clsn_data.load_bb = exit->dest; |
a509ebb5 RL |
1396 | clsn_data.store = ld_st_data->store; |
1397 | create_final_loads_for_reduction (reduction_list, &clsn_data); | |
1398 | } | |
5f40b3cb | 1399 | } |
5f40b3cb ZD |
1400 | } |
1401 | ||
a79b7ec5 | 1402 | /* Returns true if FN was created to run in parallel. */ |
5f40b3cb | 1403 | |
62e0a1ed | 1404 | bool |
a79b7ec5 | 1405 | parallelized_function_p (tree fndecl) |
5f40b3cb | 1406 | { |
a79b7ec5 TV |
1407 | cgraph_node *node = cgraph_node::get (fndecl); |
1408 | gcc_assert (node != NULL); | |
1409 | return node->parallelized_function; | |
5f40b3cb ZD |
1410 | } |
1411 | ||
1412 | /* Creates and returns an empty function that will receive the body of | |
1413 | a parallelized loop. */ | |
1414 | ||
1415 | static tree | |
9ff70652 | 1416 | create_loop_fn (location_t loc) |
5f40b3cb ZD |
1417 | { |
1418 | char buf[100]; | |
1419 | char *tname; | |
1420 | tree decl, type, name, t; | |
1421 | struct function *act_cfun = cfun; | |
1422 | static unsigned loopfn_num; | |
1423 | ||
5368224f | 1424 | loc = LOCATION_LOCUS (loc); |
5f40b3cb ZD |
1425 | snprintf (buf, 100, "%s.$loopfn", current_function_name ()); |
1426 | ASM_FORMAT_PRIVATE_NAME (tname, buf, loopfn_num++); | |
1427 | clean_symbol_name (tname); | |
1428 | name = get_identifier (tname); | |
1429 | type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); | |
1430 | ||
9ff70652 | 1431 | decl = build_decl (loc, FUNCTION_DECL, name, type); |
5f40b3cb ZD |
1432 | TREE_STATIC (decl) = 1; |
1433 | TREE_USED (decl) = 1; | |
1434 | DECL_ARTIFICIAL (decl) = 1; | |
1435 | DECL_IGNORED_P (decl) = 0; | |
1436 | TREE_PUBLIC (decl) = 0; | |
1437 | DECL_UNINLINABLE (decl) = 1; | |
1438 | DECL_EXTERNAL (decl) = 0; | |
1439 | DECL_CONTEXT (decl) = NULL_TREE; | |
1440 | DECL_INITIAL (decl) = make_node (BLOCK); | |
1441 | ||
9ff70652 | 1442 | t = build_decl (loc, RESULT_DECL, NULL_TREE, void_type_node); |
5f40b3cb ZD |
1443 | DECL_ARTIFICIAL (t) = 1; |
1444 | DECL_IGNORED_P (t) = 1; | |
1445 | DECL_RESULT (decl) = t; | |
1446 | ||
9ff70652 | 1447 | t = build_decl (loc, PARM_DECL, get_identifier (".paral_data_param"), |
5f40b3cb ZD |
1448 | ptr_type_node); |
1449 | DECL_ARTIFICIAL (t) = 1; | |
1450 | DECL_ARG_TYPE (t) = ptr_type_node; | |
1451 | DECL_CONTEXT (t) = decl; | |
1452 | TREE_USED (t) = 1; | |
1453 | DECL_ARGUMENTS (decl) = t; | |
1454 | ||
182e0d71 | 1455 | allocate_struct_function (decl, false); |
5f40b3cb ZD |
1456 | |
1457 | /* The call to allocate_struct_function clobbers CFUN, so we need to restore | |
1458 | it. */ | |
5576d6f2 | 1459 | set_cfun (act_cfun); |
5f40b3cb ZD |
1460 | |
1461 | return decl; | |
1462 | } | |
1463 | ||
7c82d827 TV |
1464 | /* Replace uses of NAME by VAL in block BB. */ |
1465 | ||
1466 | static void | |
1467 | replace_uses_in_bb_by (tree name, tree val, basic_block bb) | |
1468 | { | |
1469 | gimple use_stmt; | |
1470 | imm_use_iterator imm_iter; | |
1471 | ||
1472 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, name) | |
1473 | { | |
1474 | if (gimple_bb (use_stmt) != bb) | |
1475 | continue; | |
1476 | ||
1477 | use_operand_p use_p; | |
1478 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
1479 | SET_USE (use_p, val); | |
1480 | } | |
1481 | } | |
1482 | ||
7c82d827 TV |
1483 | /* Do transformation from: |
1484 | ||
1485 | <bb preheader>: | |
1486 | ... | |
1487 | goto <bb header> | |
1488 | ||
1489 | <bb header>: | |
1490 | ivtmp_a = PHI <ivtmp_init (preheader), ivtmp_b (latch)> | |
1491 | sum_a = PHI <sum_init (preheader), sum_b (latch)> | |
1492 | ... | |
1493 | use (ivtmp_a) | |
1494 | ... | |
1495 | sum_b = sum_a + sum_update | |
1496 | ... | |
1497 | if (ivtmp_a < n) | |
1498 | goto <bb latch>; | |
1499 | else | |
1500 | goto <bb exit>; | |
1501 | ||
1502 | <bb latch>: | |
1503 | ivtmp_b = ivtmp_a + 1; | |
1504 | goto <bb header> | |
1505 | ||
1506 | <bb exit>: | |
712cb0bb | 1507 | sum_z = PHI <sum_b (cond[1]), ...> |
7c82d827 TV |
1508 | |
1509 | [1] Where <bb cond> is single_pred (bb latch); In the simplest case, | |
1510 | that's <bb header>. | |
1511 | ||
1512 | to: | |
1513 | ||
1514 | <bb preheader>: | |
1515 | ... | |
1516 | goto <bb newheader> | |
1517 | ||
1518 | <bb header>: | |
1519 | ivtmp_a = PHI <ivtmp_c (latch)> | |
1520 | sum_a = PHI <sum_c (latch)> | |
1521 | ... | |
1522 | use (ivtmp_a) | |
1523 | ... | |
1524 | sum_b = sum_a + sum_update | |
1525 | ... | |
1526 | goto <bb latch>; | |
1527 | ||
1528 | <bb newheader>: | |
1529 | ivtmp_c = PHI <ivtmp_init (preheader), ivtmp_b (latch)> | |
1530 | sum_c = PHI <sum_init (preheader), sum_b (latch)> | |
1531 | if (ivtmp_c < n + 1) | |
1532 | goto <bb header>; | |
1533 | else | |
712cb0bb | 1534 | goto <bb newexit>; |
7c82d827 TV |
1535 | |
1536 | <bb latch>: | |
1537 | ivtmp_b = ivtmp_a + 1; | |
1538 | goto <bb newheader> | |
1539 | ||
712cb0bb TV |
1540 | <bb newexit>: |
1541 | sum_y = PHI <sum_c (newheader)> | |
1542 | ||
7c82d827 | 1543 | <bb exit>: |
712cb0bb | 1544 | sum_z = PHI <sum_y (newexit), ...> |
7c82d827 TV |
1545 | |
1546 | ||
1547 | In unified diff format: | |
1548 | ||
1549 | <bb preheader>: | |
1550 | ... | |
1551 | - goto <bb header> | |
1552 | + goto <bb newheader> | |
1553 | ||
1554 | <bb header>: | |
1555 | - ivtmp_a = PHI <ivtmp_init (preheader), ivtmp_b (latch)> | |
1556 | - sum_a = PHI <sum_init (preheader), sum_b (latch)> | |
1557 | + ivtmp_a = PHI <ivtmp_c (latch)> | |
1558 | + sum_a = PHI <sum_c (latch)> | |
1559 | ... | |
1560 | use (ivtmp_a) | |
1561 | ... | |
1562 | sum_b = sum_a + sum_update | |
1563 | ... | |
1564 | - if (ivtmp_a < n) | |
1565 | - goto <bb latch>; | |
1566 | + goto <bb latch>; | |
1567 | + | |
1568 | + <bb newheader>: | |
1569 | + ivtmp_c = PHI <ivtmp_init (preheader), ivtmp_b (latch)> | |
1570 | + sum_c = PHI <sum_init (preheader), sum_b (latch)> | |
1571 | + if (ivtmp_c < n + 1) | |
1572 | + goto <bb header>; | |
1573 | else | |
1574 | goto <bb exit>; | |
1575 | ||
1576 | <bb latch>: | |
1577 | ivtmp_b = ivtmp_a + 1; | |
1578 | - goto <bb header> | |
1579 | + goto <bb newheader> | |
1580 | ||
712cb0bb TV |
1581 | + <bb newexit>: |
1582 | + sum_y = PHI <sum_c (newheader)> | |
1583 | ||
7c82d827 | 1584 | <bb exit>: |
712cb0bb TV |
1585 | - sum_z = PHI <sum_b (cond[1]), ...> |
1586 | + sum_z = PHI <sum_y (newexit), ...> | |
7c82d827 TV |
1587 | |
1588 | Note: the example does not show any virtual phis, but these are handled more | |
1589 | or less as reductions. | |
b8698a0f | 1590 | |
7c82d827 TV |
1591 | |
1592 | Moves the exit condition of LOOP to the beginning of its header. | |
1593 | REDUCTION_LIST describes the reductions in LOOP. BOUND is the new loop | |
1594 | bound. */ | |
1595 | ||
1596 | static void | |
1597 | transform_to_exit_first_loop_alt (struct loop *loop, | |
1598 | reduction_info_table_type *reduction_list, | |
1599 | tree bound) | |
1600 | { | |
1601 | basic_block header = loop->header; | |
1602 | basic_block latch = loop->latch; | |
1603 | edge exit = single_dom_exit (loop); | |
1604 | basic_block exit_block = exit->dest; | |
1605 | gcond *cond_stmt = as_a <gcond *> (last_stmt (exit->src)); | |
1606 | tree control = gimple_cond_lhs (cond_stmt); | |
1607 | edge e; | |
1608 | ||
338392ed TV |
1609 | /* Rewriting virtuals into loop-closed ssa normal form makes this |
1610 | transformation simpler. It also ensures that the virtuals are in | |
1611 | loop-closed ssa normal from after the transformation, which is required by | |
1612 | create_parallel_loop. */ | |
1613 | rewrite_virtuals_into_loop_closed_ssa (loop); | |
7c82d827 TV |
1614 | |
1615 | /* Create the new_header block. */ | |
1616 | basic_block new_header = split_block_before_cond_jump (exit->src); | |
712cb0bb | 1617 | edge edge_at_split = single_pred_edge (new_header); |
7c82d827 TV |
1618 | |
1619 | /* Redirect entry edge to new_header. */ | |
1620 | edge entry = loop_preheader_edge (loop); | |
1621 | e = redirect_edge_and_branch (entry, new_header); | |
1622 | gcc_assert (e == entry); | |
1623 | ||
1624 | /* Redirect post_inc_edge to new_header. */ | |
1625 | edge post_inc_edge = single_succ_edge (latch); | |
1626 | e = redirect_edge_and_branch (post_inc_edge, new_header); | |
1627 | gcc_assert (e == post_inc_edge); | |
1628 | ||
1629 | /* Redirect post_cond_edge to header. */ | |
1630 | edge post_cond_edge = single_pred_edge (latch); | |
1631 | e = redirect_edge_and_branch (post_cond_edge, header); | |
1632 | gcc_assert (e == post_cond_edge); | |
1633 | ||
712cb0bb TV |
1634 | /* Redirect edge_at_split to latch. */ |
1635 | e = redirect_edge_and_branch (edge_at_split, latch); | |
1636 | gcc_assert (e == edge_at_split); | |
7c82d827 TV |
1637 | |
1638 | /* Set the new loop bound. */ | |
1639 | gimple_cond_set_rhs (cond_stmt, bound); | |
5a5fd951 | 1640 | update_stmt (cond_stmt); |
7c82d827 TV |
1641 | |
1642 | /* Repair the ssa. */ | |
1643 | vec<edge_var_map> *v = redirect_edge_var_map_vector (post_inc_edge); | |
1644 | edge_var_map *vm; | |
1645 | gphi_iterator gsi; | |
338392ed | 1646 | int i; |
7c82d827 TV |
1647 | for (gsi = gsi_start_phis (header), i = 0; |
1648 | !gsi_end_p (gsi) && v->iterate (i, &vm); | |
1649 | gsi_next (&gsi), i++) | |
1650 | { | |
1651 | gphi *phi = gsi.phi (); | |
1652 | tree res_a = PHI_RESULT (phi); | |
1653 | ||
1654 | /* Create new phi. */ | |
1655 | tree res_c = copy_ssa_name (res_a, phi); | |
1656 | gphi *nphi = create_phi_node (res_c, new_header); | |
1657 | ||
1658 | /* Replace ivtmp_a with ivtmp_c in condition 'if (ivtmp_a < n)'. */ | |
1659 | replace_uses_in_bb_by (res_a, res_c, new_header); | |
1660 | ||
1661 | /* Replace ivtmp/sum_b with ivtmp/sum_c in header phi. */ | |
1662 | add_phi_arg (phi, res_c, post_cond_edge, UNKNOWN_LOCATION); | |
1663 | ||
338392ed | 1664 | /* Replace sum_b with sum_c in exit phi. */ |
7c82d827 | 1665 | tree res_b = redirect_edge_var_map_def (vm); |
338392ed | 1666 | replace_uses_in_bb_by (res_b, res_c, exit_block); |
7c82d827 TV |
1667 | |
1668 | struct reduction_info *red = reduction_phi (reduction_list, phi); | |
1669 | gcc_assert (virtual_operand_p (res_a) | |
1670 | || res_a == control | |
1671 | || red != NULL); | |
1672 | ||
1673 | if (red) | |
1674 | { | |
1675 | /* Register the new reduction phi. */ | |
1676 | red->reduc_phi = nphi; | |
1677 | gimple_set_uid (red->reduc_phi, red->reduc_version); | |
1678 | } | |
1679 | } | |
1680 | gcc_assert (gsi_end_p (gsi) && !v->iterate (i, &vm)); | |
7c82d827 TV |
1681 | |
1682 | /* Set the preheader argument of the new phis to ivtmp/sum_init. */ | |
1683 | flush_pending_stmts (entry); | |
1684 | ||
1685 | /* Set the latch arguments of the new phis to ivtmp/sum_b. */ | |
1686 | flush_pending_stmts (post_inc_edge); | |
1687 | ||
712cb0bb TV |
1688 | /* Create a new empty exit block, inbetween the new loop header and the old |
1689 | exit block. The function separate_decls_in_region needs this block to | |
1690 | insert code that is active on loop exit, but not any other path. */ | |
1691 | basic_block new_exit_block = split_edge (exit); | |
1692 | ||
1693 | /* Insert and register the reduction exit phis. */ | |
7c82d827 TV |
1694 | for (gphi_iterator gsi = gsi_start_phis (exit_block); |
1695 | !gsi_end_p (gsi); | |
1696 | gsi_next (&gsi)) | |
1697 | { | |
1698 | gphi *phi = gsi.phi (); | |
1699 | tree res_z = PHI_RESULT (phi); | |
712cb0bb TV |
1700 | |
1701 | /* Now that we have a new exit block, duplicate the phi of the old exit | |
1702 | block in the new exit block to preserve loop-closed ssa. */ | |
1703 | edge succ_new_exit_block = single_succ_edge (new_exit_block); | |
1704 | edge pred_new_exit_block = single_pred_edge (new_exit_block); | |
1705 | tree res_y = copy_ssa_name (res_z, phi); | |
1706 | gphi *nphi = create_phi_node (res_y, new_exit_block); | |
1707 | tree res_c = PHI_ARG_DEF_FROM_EDGE (phi, succ_new_exit_block); | |
1708 | add_phi_arg (nphi, res_c, pred_new_exit_block, UNKNOWN_LOCATION); | |
1709 | add_phi_arg (phi, res_y, succ_new_exit_block, UNKNOWN_LOCATION); | |
1710 | ||
7c82d827 TV |
1711 | if (virtual_operand_p (res_z)) |
1712 | continue; | |
1713 | ||
7c82d827 TV |
1714 | gimple reduc_phi = SSA_NAME_DEF_STMT (res_c); |
1715 | struct reduction_info *red = reduction_phi (reduction_list, reduc_phi); | |
1716 | if (red != NULL) | |
712cb0bb | 1717 | red->keep_res = nphi; |
7c82d827 TV |
1718 | } |
1719 | ||
1720 | /* We're going to cancel the loop at the end of gen_parallel_loop, but until | |
1721 | then we're still using some fields, so only bother about fields that are | |
1722 | still used: header and latch. | |
1723 | The loop has a new header bb, so we update it. The latch bb stays the | |
1724 | same. */ | |
1725 | loop->header = new_header; | |
1726 | ||
1727 | /* Recalculate dominance info. */ | |
1728 | free_dominance_info (CDI_DOMINATORS); | |
1729 | calculate_dominance_info (CDI_DOMINATORS); | |
1730 | } | |
1731 | ||
1732 | /* Tries to moves the exit condition of LOOP to the beginning of its header | |
1733 | without duplication of the loop body. NIT is the number of iterations of the | |
1734 | loop. REDUCTION_LIST describes the reductions in LOOP. Return true if | |
1735 | transformation is successful. */ | |
1736 | ||
1737 | static bool | |
1738 | try_transform_to_exit_first_loop_alt (struct loop *loop, | |
1739 | reduction_info_table_type *reduction_list, | |
1740 | tree nit) | |
1741 | { | |
1742 | /* Check whether the latch contains a single statement. */ | |
1b7f61eb TV |
1743 | if (!gimple_seq_nondebug_singleton_p (bb_seq (loop->latch))) |
1744 | return false; | |
7c82d827 TV |
1745 | |
1746 | /* Check whether the latch contains the loop iv increment. */ | |
1747 | edge back = single_succ_edge (loop->latch); | |
1748 | edge exit = single_dom_exit (loop); | |
1749 | gcond *cond_stmt = as_a <gcond *> (last_stmt (exit->src)); | |
1750 | tree control = gimple_cond_lhs (cond_stmt); | |
1751 | gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (control)); | |
1752 | tree inc_res = gimple_phi_arg_def (phi, back->dest_idx); | |
1753 | if (gimple_bb (SSA_NAME_DEF_STMT (inc_res)) != loop->latch) | |
1754 | return false; | |
1755 | ||
1756 | /* Check whether there's no code between the loop condition and the latch. */ | |
1757 | if (!single_pred_p (loop->latch) | |
1758 | || single_pred (loop->latch) != exit->src) | |
1759 | return false; | |
1760 | ||
1761 | tree alt_bound = NULL_TREE; | |
1762 | tree nit_type = TREE_TYPE (nit); | |
1763 | ||
1764 | /* Figure out whether nit + 1 overflows. */ | |
1765 | if (TREE_CODE (nit) == INTEGER_CST) | |
1766 | { | |
1767 | if (!tree_int_cst_equal (nit, TYPE_MAXVAL (nit_type))) | |
1768 | { | |
1769 | alt_bound = fold_build2_loc (UNKNOWN_LOCATION, PLUS_EXPR, nit_type, | |
1770 | nit, build_one_cst (nit_type)); | |
1771 | ||
1772 | gcc_assert (TREE_CODE (alt_bound) == INTEGER_CST); | |
fd7b3ef5 TV |
1773 | transform_to_exit_first_loop_alt (loop, reduction_list, alt_bound); |
1774 | return true; | |
7c82d827 TV |
1775 | } |
1776 | else | |
1777 | { | |
1778 | /* Todo: Figure out if we can trigger this, if it's worth to handle | |
1779 | optimally, and if we can handle it optimally. */ | |
fd7b3ef5 | 1780 | return false; |
7c82d827 TV |
1781 | } |
1782 | } | |
7c82d827 | 1783 | |
fd7b3ef5 | 1784 | gcc_assert (TREE_CODE (nit) == SSA_NAME); |
7c82d827 | 1785 | |
4f75d608 TV |
1786 | /* Variable nit is the loop bound as returned by canonicalize_loop_ivs, for an |
1787 | iv with base 0 and step 1 that is incremented in the latch, like this: | |
1788 | ||
1789 | <bb header>: | |
1790 | # iv_1 = PHI <0 (preheader), iv_2 (latch)> | |
1791 | ... | |
1792 | if (iv_1 < nit) | |
1793 | goto <bb latch>; | |
1794 | else | |
1795 | goto <bb exit>; | |
1796 | ||
1797 | <bb latch>: | |
1798 | iv_2 = iv_1 + 1; | |
1799 | goto <bb header>; | |
1800 | ||
1801 | The range of iv_1 is [0, nit]. The latch edge is taken for | |
1802 | iv_1 == [0, nit - 1] and the exit edge is taken for iv_1 == nit. So the | |
1803 | number of latch executions is equal to nit. | |
1804 | ||
1805 | The function max_loop_iterations gives us the maximum number of latch | |
1806 | executions, so it gives us the maximum value of nit. */ | |
1807 | widest_int nit_max; | |
1808 | if (!max_loop_iterations (loop, &nit_max)) | |
1809 | return false; | |
1810 | ||
1811 | /* Check if nit + 1 overflows. */ | |
1812 | widest_int type_max = wi::to_widest (TYPE_MAXVAL (nit_type)); | |
1813 | if (!wi::lts_p (nit_max, type_max)) | |
1814 | return false; | |
1815 | ||
fd7b3ef5 | 1816 | gimple def = SSA_NAME_DEF_STMT (nit); |
7c82d827 | 1817 | |
4f75d608 | 1818 | /* Try to find nit + 1, in the form of n in an assignment nit = n - 1. */ |
fd7b3ef5 TV |
1819 | if (def |
1820 | && is_gimple_assign (def) | |
1821 | && gimple_assign_rhs_code (def) == PLUS_EXPR) | |
1822 | { | |
1823 | tree op1 = gimple_assign_rhs1 (def); | |
1824 | tree op2 = gimple_assign_rhs2 (def); | |
1825 | if (integer_minus_onep (op1)) | |
1826 | alt_bound = op2; | |
1827 | else if (integer_minus_onep (op2)) | |
1828 | alt_bound = op1; | |
7c82d827 TV |
1829 | } |
1830 | ||
9f620bf1 | 1831 | /* If not found, insert nit + 1. */ |
7c82d827 | 1832 | if (alt_bound == NULL_TREE) |
9f620bf1 TV |
1833 | { |
1834 | alt_bound = fold_build2 (PLUS_EXPR, nit_type, nit, | |
1835 | build_int_cst_type (nit_type, 1)); | |
1836 | ||
1837 | gimple_stmt_iterator gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
1838 | ||
1839 | alt_bound | |
1840 | = force_gimple_operand_gsi (&gsi, alt_bound, true, NULL_TREE, false, | |
1841 | GSI_CONTINUE_LINKING); | |
1842 | } | |
7c82d827 TV |
1843 | |
1844 | transform_to_exit_first_loop_alt (loop, reduction_list, alt_bound); | |
1845 | return true; | |
1846 | } | |
1847 | ||
1848 | /* Moves the exit condition of LOOP to the beginning of its header. NIT is the | |
1849 | number of iterations of the loop. REDUCTION_LIST describes the reductions in | |
1850 | LOOP. */ | |
5f40b3cb ZD |
1851 | |
1852 | static void | |
4a8fb1a1 | 1853 | transform_to_exit_first_loop (struct loop *loop, |
c203e8a7 | 1854 | reduction_info_table_type *reduction_list, |
4a8fb1a1 | 1855 | tree nit) |
5f40b3cb ZD |
1856 | { |
1857 | basic_block *bbs, *nbbs, ex_bb, orig_header; | |
1858 | unsigned n; | |
1859 | bool ok; | |
1860 | edge exit = single_dom_exit (loop), hpred; | |
726a989a | 1861 | tree control, control_name, res, t; |
538dd0b7 DM |
1862 | gphi *phi, *nphi; |
1863 | gassign *stmt; | |
1864 | gcond *cond_stmt, *cond_nit; | |
48710229 | 1865 | tree nit_1; |
5f40b3cb ZD |
1866 | |
1867 | split_block_after_labels (loop->header); | |
1868 | orig_header = single_succ (loop->header); | |
1869 | hpred = single_succ_edge (loop->header); | |
1870 | ||
538dd0b7 | 1871 | cond_stmt = as_a <gcond *> (last_stmt (exit->src)); |
726a989a RB |
1872 | control = gimple_cond_lhs (cond_stmt); |
1873 | gcc_assert (gimple_cond_rhs (cond_stmt) == nit); | |
5f40b3cb ZD |
1874 | |
1875 | /* Make sure that we have phi nodes on exit for all loop header phis | |
1876 | (create_parallel_loop requires that). */ | |
538dd0b7 DM |
1877 | for (gphi_iterator gsi = gsi_start_phis (loop->header); |
1878 | !gsi_end_p (gsi); | |
1879 | gsi_next (&gsi)) | |
5f40b3cb | 1880 | { |
538dd0b7 | 1881 | phi = gsi.phi (); |
5f40b3cb | 1882 | res = PHI_RESULT (phi); |
070ecdfd | 1883 | t = copy_ssa_name (res, phi); |
5f40b3cb | 1884 | SET_PHI_RESULT (phi, t); |
5f40b3cb | 1885 | nphi = create_phi_node (res, orig_header); |
9e227d60 | 1886 | add_phi_arg (nphi, t, hpred, UNKNOWN_LOCATION); |
5f40b3cb ZD |
1887 | |
1888 | if (res == control) | |
1889 | { | |
726a989a | 1890 | gimple_cond_set_lhs (cond_stmt, t); |
5f40b3cb ZD |
1891 | update_stmt (cond_stmt); |
1892 | control = t; | |
1893 | } | |
1894 | } | |
12037899 | 1895 | |
5f40b3cb | 1896 | bbs = get_loop_body_in_dom_order (loop); |
48710229 | 1897 | |
69958396 RL |
1898 | for (n = 0; bbs[n] != exit->src; n++) |
1899 | continue; | |
5f40b3cb | 1900 | nbbs = XNEWVEC (basic_block, n); |
726a989a RB |
1901 | ok = gimple_duplicate_sese_tail (single_succ_edge (loop->header), exit, |
1902 | bbs + 1, n, nbbs); | |
5f40b3cb ZD |
1903 | gcc_assert (ok); |
1904 | free (bbs); | |
1905 | ex_bb = nbbs[0]; | |
1906 | free (nbbs); | |
1907 | ||
b8698a0f | 1908 | /* Other than reductions, the only gimple reg that should be copied |
726a989a | 1909 | out of the loop is the control variable. */ |
69958396 | 1910 | exit = single_dom_exit (loop); |
5f40b3cb | 1911 | control_name = NULL_TREE; |
538dd0b7 DM |
1912 | for (gphi_iterator gsi = gsi_start_phis (ex_bb); |
1913 | !gsi_end_p (gsi); ) | |
5f40b3cb | 1914 | { |
538dd0b7 | 1915 | phi = gsi.phi (); |
5f40b3cb | 1916 | res = PHI_RESULT (phi); |
ea057359 | 1917 | if (virtual_operand_p (res)) |
726a989a RB |
1918 | { |
1919 | gsi_next (&gsi); | |
1920 | continue; | |
1921 | } | |
5f40b3cb | 1922 | |
a509ebb5 | 1923 | /* Check if it is a part of reduction. If it is, |
b8698a0f L |
1924 | keep the phi at the reduction's keep_res field. The |
1925 | PHI_RESULT of this phi is the resulting value of the reduction | |
a509ebb5 RL |
1926 | variable when exiting the loop. */ |
1927 | ||
c203e8a7 | 1928 | if (reduction_list->elements () > 0) |
a509ebb5 RL |
1929 | { |
1930 | struct reduction_info *red; | |
1931 | ||
1932 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
a509ebb5 RL |
1933 | red = reduction_phi (reduction_list, SSA_NAME_DEF_STMT (val)); |
1934 | if (red) | |
726a989a RB |
1935 | { |
1936 | red->keep_res = phi; | |
1937 | gsi_next (&gsi); | |
1938 | continue; | |
1939 | } | |
a509ebb5 | 1940 | } |
726a989a RB |
1941 | gcc_assert (control_name == NULL_TREE |
1942 | && SSA_NAME_VAR (res) == SSA_NAME_VAR (control)); | |
5f40b3cb | 1943 | control_name = res; |
726a989a | 1944 | remove_phi_node (&gsi, false); |
5f40b3cb ZD |
1945 | } |
1946 | gcc_assert (control_name != NULL_TREE); | |
5f40b3cb | 1947 | |
b8698a0f | 1948 | /* Initialize the control variable to number of iterations |
48710229 | 1949 | according to the rhs of the exit condition. */ |
538dd0b7 DM |
1950 | gimple_stmt_iterator gsi = gsi_after_labels (ex_bb); |
1951 | cond_nit = as_a <gcond *> (last_stmt (exit->src)); | |
48710229 RL |
1952 | nit_1 = gimple_cond_rhs (cond_nit); |
1953 | nit_1 = force_gimple_operand_gsi (&gsi, | |
1954 | fold_convert (TREE_TYPE (control_name), nit_1), | |
726a989a | 1955 | false, NULL_TREE, false, GSI_SAME_STMT); |
48710229 | 1956 | stmt = gimple_build_assign (control_name, nit_1); |
726a989a | 1957 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb ZD |
1958 | } |
1959 | ||
1960 | /* Create the parallel constructs for LOOP as described in gen_parallel_loop. | |
726a989a | 1961 | LOOP_FN and DATA are the arguments of GIMPLE_OMP_PARALLEL. |
5f40b3cb ZD |
1962 | NEW_DATA is the variable that should be initialized from the argument |
1963 | of LOOP_FN. N_THREADS is the requested number of threads. Returns the | |
726a989a | 1964 | basic block containing GIMPLE_OMP_PARALLEL tree. */ |
5f40b3cb ZD |
1965 | |
1966 | static basic_block | |
1967 | create_parallel_loop (struct loop *loop, tree loop_fn, tree data, | |
9ff70652 | 1968 | tree new_data, unsigned n_threads, location_t loc) |
5f40b3cb | 1969 | { |
726a989a | 1970 | gimple_stmt_iterator gsi; |
5f40b3cb | 1971 | basic_block bb, paral_bb, for_bb, ex_bb; |
0f900dfa | 1972 | tree t, param; |
538dd0b7 DM |
1973 | gomp_parallel *omp_par_stmt; |
1974 | gimple omp_return_stmt1, omp_return_stmt2; | |
1975 | gimple phi; | |
1976 | gcond *cond_stmt; | |
1977 | gomp_for *for_stmt; | |
1978 | gomp_continue *omp_cont_stmt; | |
726a989a | 1979 | tree cvar, cvar_init, initvar, cvar_next, cvar_base, type; |
5f40b3cb ZD |
1980 | edge exit, nexit, guard, end, e; |
1981 | ||
726a989a | 1982 | /* Prepare the GIMPLE_OMP_PARALLEL statement. */ |
5f40b3cb ZD |
1983 | bb = loop_preheader_edge (loop)->src; |
1984 | paral_bb = single_pred (bb); | |
726a989a | 1985 | gsi = gsi_last_bb (paral_bb); |
5f40b3cb | 1986 | |
9ff70652 | 1987 | t = build_omp_clause (loc, OMP_CLAUSE_NUM_THREADS); |
5f40b3cb | 1988 | OMP_CLAUSE_NUM_THREADS_EXPR (t) |
a509ebb5 | 1989 | = build_int_cst (integer_type_node, n_threads); |
538dd0b7 DM |
1990 | omp_par_stmt = gimple_build_omp_parallel (NULL, t, loop_fn, data); |
1991 | gimple_set_location (omp_par_stmt, loc); | |
5f40b3cb | 1992 | |
538dd0b7 | 1993 | gsi_insert_after (&gsi, omp_par_stmt, GSI_NEW_STMT); |
5f40b3cb ZD |
1994 | |
1995 | /* Initialize NEW_DATA. */ | |
1996 | if (data) | |
1997 | { | |
538dd0b7 DM |
1998 | gassign *assign_stmt; |
1999 | ||
726a989a RB |
2000 | gsi = gsi_after_labels (bb); |
2001 | ||
b731b390 | 2002 | param = make_ssa_name (DECL_ARGUMENTS (loop_fn)); |
538dd0b7 DM |
2003 | assign_stmt = gimple_build_assign (param, build_fold_addr_expr (data)); |
2004 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
726a989a | 2005 | |
538dd0b7 | 2006 | assign_stmt = gimple_build_assign (new_data, |
726a989a | 2007 | fold_convert (TREE_TYPE (new_data), param)); |
538dd0b7 | 2008 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); |
5f40b3cb ZD |
2009 | } |
2010 | ||
726a989a | 2011 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_PARALLEL. */ |
5f40b3cb | 2012 | bb = split_loop_exit_edge (single_dom_exit (loop)); |
726a989a | 2013 | gsi = gsi_last_bb (bb); |
538dd0b7 DM |
2014 | omp_return_stmt1 = gimple_build_omp_return (false); |
2015 | gimple_set_location (omp_return_stmt1, loc); | |
2016 | gsi_insert_after (&gsi, omp_return_stmt1, GSI_NEW_STMT); | |
5f40b3cb | 2017 | |
726a989a | 2018 | /* Extract data for GIMPLE_OMP_FOR. */ |
5f40b3cb | 2019 | gcc_assert (loop->header == single_dom_exit (loop)->src); |
538dd0b7 | 2020 | cond_stmt = as_a <gcond *> (last_stmt (loop->header)); |
5f40b3cb | 2021 | |
726a989a | 2022 | cvar = gimple_cond_lhs (cond_stmt); |
5f40b3cb ZD |
2023 | cvar_base = SSA_NAME_VAR (cvar); |
2024 | phi = SSA_NAME_DEF_STMT (cvar); | |
2025 | cvar_init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
b731b390 | 2026 | initvar = copy_ssa_name (cvar); |
5f40b3cb ZD |
2027 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, loop_preheader_edge (loop)), |
2028 | initvar); | |
2029 | cvar_next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
2030 | ||
1dff453d | 2031 | gsi = gsi_last_nondebug_bb (loop->latch); |
726a989a RB |
2032 | gcc_assert (gsi_stmt (gsi) == SSA_NAME_DEF_STMT (cvar_next)); |
2033 | gsi_remove (&gsi, true); | |
5f40b3cb ZD |
2034 | |
2035 | /* Prepare cfg. */ | |
2036 | for_bb = split_edge (loop_preheader_edge (loop)); | |
2037 | ex_bb = split_loop_exit_edge (single_dom_exit (loop)); | |
2038 | extract_true_false_edges_from_block (loop->header, &nexit, &exit); | |
2039 | gcc_assert (exit == single_dom_exit (loop)); | |
2040 | ||
2041 | guard = make_edge (for_bb, ex_bb, 0); | |
2042 | single_succ_edge (loop->latch)->flags = 0; | |
2043 | end = make_edge (loop->latch, ex_bb, EDGE_FALLTHRU); | |
538dd0b7 DM |
2044 | for (gphi_iterator gpi = gsi_start_phis (ex_bb); |
2045 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
5f40b3cb | 2046 | { |
f5045c96 AM |
2047 | source_location locus; |
2048 | tree def; | |
538dd0b7 DM |
2049 | gphi *phi = gpi.phi (); |
2050 | gphi *stmt; | |
2051 | ||
2052 | stmt = as_a <gphi *> ( | |
2053 | SSA_NAME_DEF_STMT (PHI_ARG_DEF_FROM_EDGE (phi, exit))); | |
f5045c96 AM |
2054 | |
2055 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_preheader_edge (loop)); | |
b8698a0f | 2056 | locus = gimple_phi_arg_location_from_edge (stmt, |
f5045c96 | 2057 | loop_preheader_edge (loop)); |
9e227d60 | 2058 | add_phi_arg (phi, def, guard, locus); |
f5045c96 AM |
2059 | |
2060 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (loop)); | |
2061 | locus = gimple_phi_arg_location_from_edge (stmt, loop_latch_edge (loop)); | |
9e227d60 | 2062 | add_phi_arg (phi, def, end, locus); |
5f40b3cb ZD |
2063 | } |
2064 | e = redirect_edge_and_branch (exit, nexit->dest); | |
2065 | PENDING_STMT (e) = NULL; | |
2066 | ||
726a989a RB |
2067 | /* Emit GIMPLE_OMP_FOR. */ |
2068 | gimple_cond_set_lhs (cond_stmt, cvar_base); | |
5f40b3cb | 2069 | type = TREE_TYPE (cvar); |
9ff70652 | 2070 | t = build_omp_clause (loc, OMP_CLAUSE_SCHEDULE); |
5f40b3cb ZD |
2071 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_STATIC; |
2072 | ||
74bf76ed | 2073 | for_stmt = gimple_build_omp_for (NULL, GF_OMP_FOR_KIND_FOR, t, 1, NULL); |
9ff70652 | 2074 | gimple_set_location (for_stmt, loc); |
726a989a RB |
2075 | gimple_omp_for_set_index (for_stmt, 0, initvar); |
2076 | gimple_omp_for_set_initial (for_stmt, 0, cvar_init); | |
2077 | gimple_omp_for_set_final (for_stmt, 0, gimple_cond_rhs (cond_stmt)); | |
2078 | gimple_omp_for_set_cond (for_stmt, 0, gimple_cond_code (cond_stmt)); | |
2079 | gimple_omp_for_set_incr (for_stmt, 0, build2 (PLUS_EXPR, type, | |
2080 | cvar_base, | |
2081 | build_int_cst (type, 1))); | |
2082 | ||
2083 | gsi = gsi_last_bb (for_bb); | |
2084 | gsi_insert_after (&gsi, for_stmt, GSI_NEW_STMT); | |
5f40b3cb ZD |
2085 | SSA_NAME_DEF_STMT (initvar) = for_stmt; |
2086 | ||
726a989a RB |
2087 | /* Emit GIMPLE_OMP_CONTINUE. */ |
2088 | gsi = gsi_last_bb (loop->latch); | |
538dd0b7 DM |
2089 | omp_cont_stmt = gimple_build_omp_continue (cvar_next, cvar); |
2090 | gimple_set_location (omp_cont_stmt, loc); | |
2091 | gsi_insert_after (&gsi, omp_cont_stmt, GSI_NEW_STMT); | |
2092 | SSA_NAME_DEF_STMT (cvar_next) = omp_cont_stmt; | |
5f40b3cb | 2093 | |
726a989a RB |
2094 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_FOR. */ |
2095 | gsi = gsi_last_bb (ex_bb); | |
538dd0b7 DM |
2096 | omp_return_stmt2 = gimple_build_omp_return (true); |
2097 | gimple_set_location (omp_return_stmt2, loc); | |
2098 | gsi_insert_after (&gsi, omp_return_stmt2, GSI_NEW_STMT); | |
5f40b3cb | 2099 | |
cd7d9fd7 RG |
2100 | /* After the above dom info is hosed. Re-compute it. */ |
2101 | free_dominance_info (CDI_DOMINATORS); | |
2102 | calculate_dominance_info (CDI_DOMINATORS); | |
2103 | ||
5f40b3cb ZD |
2104 | return paral_bb; |
2105 | } | |
2106 | ||
08dab97a RL |
2107 | /* Generates code to execute the iterations of LOOP in N_THREADS |
2108 | threads in parallel. | |
2109 | ||
2110 | NITER describes number of iterations of LOOP. | |
fa10beec | 2111 | REDUCTION_LIST describes the reductions existent in the LOOP. */ |
5f40b3cb ZD |
2112 | |
2113 | static void | |
c203e8a7 TS |
2114 | gen_parallel_loop (struct loop *loop, |
2115 | reduction_info_table_type *reduction_list, | |
a509ebb5 | 2116 | unsigned n_threads, struct tree_niter_desc *niter) |
5f40b3cb | 2117 | { |
5f40b3cb | 2118 | tree many_iterations_cond, type, nit; |
726a989a RB |
2119 | tree arg_struct, new_arg_struct; |
2120 | gimple_seq stmts; | |
9f9f72aa | 2121 | edge entry, exit; |
a509ebb5 | 2122 | struct clsn_data clsn_data; |
5f40b3cb | 2123 | unsigned prob; |
9ff70652 JJ |
2124 | location_t loc; |
2125 | gimple cond_stmt; | |
768da0da | 2126 | unsigned int m_p_thread=2; |
5f40b3cb ZD |
2127 | |
2128 | /* From | |
2129 | ||
2130 | --------------------------------------------------------------------- | |
2131 | loop | |
2132 | { | |
2133 | IV = phi (INIT, IV + STEP) | |
2134 | BODY1; | |
2135 | if (COND) | |
2136 | break; | |
2137 | BODY2; | |
2138 | } | |
2139 | --------------------------------------------------------------------- | |
2140 | ||
2141 | with # of iterations NITER (possibly with MAY_BE_ZERO assumption), | |
2142 | we generate the following code: | |
2143 | ||
2144 | --------------------------------------------------------------------- | |
2145 | ||
2146 | if (MAY_BE_ZERO | |
a509ebb5 RL |
2147 | || NITER < MIN_PER_THREAD * N_THREADS) |
2148 | goto original; | |
5f40b3cb ZD |
2149 | |
2150 | BODY1; | |
2151 | store all local loop-invariant variables used in body of the loop to DATA. | |
726a989a | 2152 | GIMPLE_OMP_PARALLEL (OMP_CLAUSE_NUM_THREADS (N_THREADS), LOOPFN, DATA); |
5f40b3cb | 2153 | load the variables from DATA. |
726a989a | 2154 | GIMPLE_OMP_FOR (IV = INIT; COND; IV += STEP) (OMP_CLAUSE_SCHEDULE (static)) |
5f40b3cb ZD |
2155 | BODY2; |
2156 | BODY1; | |
726a989a RB |
2157 | GIMPLE_OMP_CONTINUE; |
2158 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_FOR | |
2159 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_PARALLEL | |
5f40b3cb ZD |
2160 | goto end; |
2161 | ||
2162 | original: | |
2163 | loop | |
2164 | { | |
2165 | IV = phi (INIT, IV + STEP) | |
2166 | BODY1; | |
2167 | if (COND) | |
2168 | break; | |
2169 | BODY2; | |
2170 | } | |
2171 | ||
2172 | end: | |
2173 | ||
2174 | */ | |
2175 | ||
2176 | /* Create two versions of the loop -- in the old one, we know that the | |
2177 | number of iterations is large enough, and we will transform it into the | |
2178 | loop that will be split to loop_fn, the new one will be used for the | |
2179 | remaining iterations. */ | |
a509ebb5 | 2180 | |
768da0da RL |
2181 | /* We should compute a better number-of-iterations value for outer loops. |
2182 | That is, if we have | |
2183 | ||
2184 | for (i = 0; i < n; ++i) | |
2185 | for (j = 0; j < m; ++j) | |
2186 | ... | |
2187 | ||
2188 | we should compute nit = n * m, not nit = n. | |
2189 | Also may_be_zero handling would need to be adjusted. */ | |
2190 | ||
5f40b3cb ZD |
2191 | type = TREE_TYPE (niter->niter); |
2192 | nit = force_gimple_operand (unshare_expr (niter->niter), &stmts, true, | |
2193 | NULL_TREE); | |
2194 | if (stmts) | |
726a989a | 2195 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb | 2196 | |
768da0da RL |
2197 | if (loop->inner) |
2198 | m_p_thread=2; | |
2199 | else | |
2200 | m_p_thread=MIN_PER_THREAD; | |
2201 | ||
2202 | many_iterations_cond = | |
2203 | fold_build2 (GE_EXPR, boolean_type_node, | |
2204 | nit, build_int_cst (type, m_p_thread * n_threads)); | |
2205 | ||
5f40b3cb | 2206 | many_iterations_cond |
a509ebb5 RL |
2207 | = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
2208 | invert_truthvalue (unshare_expr (niter->may_be_zero)), | |
2209 | many_iterations_cond); | |
5f40b3cb | 2210 | many_iterations_cond |
a509ebb5 | 2211 | = force_gimple_operand (many_iterations_cond, &stmts, false, NULL_TREE); |
5f40b3cb | 2212 | if (stmts) |
726a989a | 2213 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb ZD |
2214 | if (!is_gimple_condexpr (many_iterations_cond)) |
2215 | { | |
2216 | many_iterations_cond | |
a509ebb5 RL |
2217 | = force_gimple_operand (many_iterations_cond, &stmts, |
2218 | true, NULL_TREE); | |
5f40b3cb | 2219 | if (stmts) |
726a989a | 2220 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb ZD |
2221 | } |
2222 | ||
2223 | initialize_original_copy_tables (); | |
2224 | ||
2225 | /* We assume that the loop usually iterates a lot. */ | |
2226 | prob = 4 * REG_BR_PROB_BASE / 5; | |
0f900dfa JJ |
2227 | loop_version (loop, many_iterations_cond, NULL, |
2228 | prob, prob, REG_BR_PROB_BASE - prob, true); | |
5f40b3cb ZD |
2229 | update_ssa (TODO_update_ssa); |
2230 | free_original_copy_tables (); | |
2231 | ||
2232 | /* Base all the induction variables in LOOP on a single control one. */ | |
c80a5403 | 2233 | canonicalize_loop_ivs (loop, &nit, true); |
5f40b3cb | 2234 | |
7c82d827 TV |
2235 | /* Ensure that the exit condition is the first statement in the loop. |
2236 | The common case is that latch of the loop is empty (apart from the | |
2237 | increment) and immediately follows the loop exit test. Attempt to move the | |
2238 | entry of the loop directly before the exit check and increase the number of | |
2239 | iterations of the loop by one. */ | |
2240 | if (!try_transform_to_exit_first_loop_alt (loop, reduction_list, nit)) | |
2241 | { | |
2242 | /* Fall back on the method that handles more cases, but duplicates the | |
2243 | loop body: move the exit condition of LOOP to the beginning of its | |
2244 | header, and duplicate the part of the last iteration that gets disabled | |
2245 | to the exit of the loop. */ | |
2246 | transform_to_exit_first_loop (loop, reduction_list, nit); | |
2247 | } | |
a509ebb5 | 2248 | |
fa10beec | 2249 | /* Generate initializations for reductions. */ |
c203e8a7 TS |
2250 | if (reduction_list->elements () > 0) |
2251 | reduction_list->traverse <struct loop *, initialize_reductions> (loop); | |
5f40b3cb ZD |
2252 | |
2253 | /* Eliminate the references to local variables from the loop. */ | |
9f9f72aa AP |
2254 | gcc_assert (single_exit (loop)); |
2255 | entry = loop_preheader_edge (loop); | |
2256 | exit = single_dom_exit (loop); | |
5f40b3cb | 2257 | |
9f9f72aa | 2258 | eliminate_local_variables (entry, exit); |
5f40b3cb ZD |
2259 | /* In the old loop, move all variables non-local to the loop to a structure |
2260 | and back, and create separate decls for the variables used in loop. */ | |
b8698a0f | 2261 | separate_decls_in_region (entry, exit, reduction_list, &arg_struct, |
9f9f72aa | 2262 | &new_arg_struct, &clsn_data); |
5f40b3cb ZD |
2263 | |
2264 | /* Create the parallel constructs. */ | |
9ff70652 JJ |
2265 | loc = UNKNOWN_LOCATION; |
2266 | cond_stmt = last_stmt (loop->header); | |
2267 | if (cond_stmt) | |
2268 | loc = gimple_location (cond_stmt); | |
18751894 TV |
2269 | create_parallel_loop (loop, create_loop_fn (loc), arg_struct, |
2270 | new_arg_struct, n_threads, loc); | |
c203e8a7 | 2271 | if (reduction_list->elements () > 0) |
a509ebb5 | 2272 | create_call_for_reduction (loop, reduction_list, &clsn_data); |
5f40b3cb ZD |
2273 | |
2274 | scev_reset (); | |
2275 | ||
2276 | /* Cancel the loop (it is simpler to do it here rather than to teach the | |
2277 | expander to do it). */ | |
2278 | cancel_loop_tree (loop); | |
2279 | ||
92a6bdbd SP |
2280 | /* Free loop bound estimations that could contain references to |
2281 | removed statements. */ | |
f0bd40b1 | 2282 | FOR_EACH_LOOP (loop, 0) |
92a6bdbd | 2283 | free_numbers_of_iterations_estimates_loop (loop); |
5f40b3cb ZD |
2284 | } |
2285 | ||
9857228c SP |
2286 | /* Returns true when LOOP contains vector phi nodes. */ |
2287 | ||
2288 | static bool | |
726a989a | 2289 | loop_has_vector_phi_nodes (struct loop *loop ATTRIBUTE_UNUSED) |
9857228c SP |
2290 | { |
2291 | unsigned i; | |
2292 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
538dd0b7 | 2293 | gphi_iterator gsi; |
9857228c | 2294 | bool res = true; |
9857228c SP |
2295 | |
2296 | for (i = 0; i < loop->num_nodes; i++) | |
726a989a | 2297 | for (gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) |
538dd0b7 | 2298 | if (TREE_CODE (TREE_TYPE (PHI_RESULT (gsi.phi ()))) == VECTOR_TYPE) |
9857228c SP |
2299 | goto end; |
2300 | ||
2301 | res = false; | |
2302 | end: | |
2303 | free (bbs); | |
2304 | return res; | |
2305 | } | |
2306 | ||
08dab97a RL |
2307 | /* Create a reduction_info struct, initialize it with REDUC_STMT |
2308 | and PHI, insert it to the REDUCTION_LIST. */ | |
2309 | ||
2310 | static void | |
c203e8a7 | 2311 | build_new_reduction (reduction_info_table_type *reduction_list, |
538dd0b7 | 2312 | gimple reduc_stmt, gphi *phi) |
08dab97a | 2313 | { |
4a8fb1a1 | 2314 | reduction_info **slot; |
08dab97a RL |
2315 | struct reduction_info *new_reduction; |
2316 | ||
2317 | gcc_assert (reduc_stmt); | |
b8698a0f | 2318 | |
08dab97a RL |
2319 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2320 | { | |
2321 | fprintf (dump_file, | |
2322 | "Detected reduction. reduction stmt is: \n"); | |
2323 | print_gimple_stmt (dump_file, reduc_stmt, 0, 0); | |
2324 | fprintf (dump_file, "\n"); | |
2325 | } | |
b8698a0f | 2326 | |
08dab97a | 2327 | new_reduction = XCNEW (struct reduction_info); |
b8698a0f | 2328 | |
08dab97a RL |
2329 | new_reduction->reduc_stmt = reduc_stmt; |
2330 | new_reduction->reduc_phi = phi; | |
5d1fd1de | 2331 | new_reduction->reduc_version = SSA_NAME_VERSION (gimple_phi_result (phi)); |
08dab97a | 2332 | new_reduction->reduction_code = gimple_assign_rhs_code (reduc_stmt); |
c203e8a7 | 2333 | slot = reduction_list->find_slot (new_reduction, INSERT); |
08dab97a RL |
2334 | *slot = new_reduction; |
2335 | } | |
2336 | ||
5d1fd1de JJ |
2337 | /* Callback for htab_traverse. Sets gimple_uid of reduc_phi stmts. */ |
2338 | ||
4a8fb1a1 LC |
2339 | int |
2340 | set_reduc_phi_uids (reduction_info **slot, void *data ATTRIBUTE_UNUSED) | |
5d1fd1de | 2341 | { |
4a8fb1a1 | 2342 | struct reduction_info *const red = *slot; |
5d1fd1de JJ |
2343 | gimple_set_uid (red->reduc_phi, red->reduc_version); |
2344 | return 1; | |
2345 | } | |
2346 | ||
08dab97a RL |
2347 | /* Detect all reductions in the LOOP, insert them into REDUCTION_LIST. */ |
2348 | ||
2349 | static void | |
c203e8a7 | 2350 | gather_scalar_reductions (loop_p loop, reduction_info_table_type *reduction_list) |
08dab97a | 2351 | { |
538dd0b7 | 2352 | gphi_iterator gsi; |
08dab97a RL |
2353 | loop_vec_info simple_loop_info; |
2354 | ||
08dab97a RL |
2355 | simple_loop_info = vect_analyze_loop_form (loop); |
2356 | ||
2357 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2358 | { | |
538dd0b7 | 2359 | gphi *phi = gsi.phi (); |
08dab97a RL |
2360 | affine_iv iv; |
2361 | tree res = PHI_RESULT (phi); | |
2362 | bool double_reduc; | |
2363 | ||
ea057359 | 2364 | if (virtual_operand_p (res)) |
08dab97a RL |
2365 | continue; |
2366 | ||
2367 | if (!simple_iv (loop, loop, res, &iv, true) | |
2368 | && simple_loop_info) | |
2369 | { | |
8a9ecffd MM |
2370 | gimple reduc_stmt = vect_force_simple_reduction (simple_loop_info, |
2371 | phi, true, | |
2372 | &double_reduc); | |
48710229 | 2373 | if (reduc_stmt && !double_reduc) |
08dab97a RL |
2374 | build_new_reduction (reduction_list, reduc_stmt, phi); |
2375 | } | |
2376 | } | |
5d1fd1de JJ |
2377 | destroy_loop_vec_info (simple_loop_info, true); |
2378 | ||
2379 | /* As gimple_uid is used by the vectorizer in between vect_analyze_loop_form | |
2380 | and destroy_loop_vec_info, we can set gimple_uid of reduc_phi stmts | |
2381 | only now. */ | |
c203e8a7 | 2382 | reduction_list->traverse <void *, set_reduc_phi_uids> (NULL); |
08dab97a RL |
2383 | } |
2384 | ||
2385 | /* Try to initialize NITER for code generation part. */ | |
2386 | ||
2387 | static bool | |
2388 | try_get_loop_niter (loop_p loop, struct tree_niter_desc *niter) | |
2389 | { | |
2390 | edge exit = single_dom_exit (loop); | |
2391 | ||
2392 | gcc_assert (exit); | |
2393 | ||
2394 | /* We need to know # of iterations, and there should be no uses of values | |
2395 | defined inside loop outside of it, unless the values are invariants of | |
2396 | the loop. */ | |
2397 | if (!number_of_iterations_exit (loop, exit, niter, false)) | |
2398 | { | |
2399 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2400 | fprintf (dump_file, " FAILED: number of iterations not known\n"); | |
2401 | return false; | |
2402 | } | |
2403 | ||
2404 | return true; | |
2405 | } | |
2406 | ||
2407 | /* Try to initialize REDUCTION_LIST for code generation part. | |
2408 | REDUCTION_LIST describes the reductions. */ | |
2409 | ||
2410 | static bool | |
4a8fb1a1 | 2411 | try_create_reduction_list (loop_p loop, |
c203e8a7 | 2412 | reduction_info_table_type *reduction_list) |
08dab97a RL |
2413 | { |
2414 | edge exit = single_dom_exit (loop); | |
538dd0b7 | 2415 | gphi_iterator gsi; |
08dab97a RL |
2416 | |
2417 | gcc_assert (exit); | |
2418 | ||
2419 | gather_scalar_reductions (loop, reduction_list); | |
2420 | ||
b8698a0f | 2421 | |
08dab97a RL |
2422 | for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) |
2423 | { | |
538dd0b7 | 2424 | gphi *phi = gsi.phi (); |
08dab97a RL |
2425 | struct reduction_info *red; |
2426 | imm_use_iterator imm_iter; | |
2427 | use_operand_p use_p; | |
2428 | gimple reduc_phi; | |
2429 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
2430 | ||
ea057359 | 2431 | if (!virtual_operand_p (val)) |
08dab97a RL |
2432 | { |
2433 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2434 | { | |
2435 | fprintf (dump_file, "phi is "); | |
2436 | print_gimple_stmt (dump_file, phi, 0, 0); | |
2437 | fprintf (dump_file, "arg of phi to exit: value "); | |
2438 | print_generic_expr (dump_file, val, 0); | |
2439 | fprintf (dump_file, " used outside loop\n"); | |
2440 | fprintf (dump_file, | |
2441 | " checking if it a part of reduction pattern: \n"); | |
2442 | } | |
c203e8a7 | 2443 | if (reduction_list->elements () == 0) |
08dab97a RL |
2444 | { |
2445 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2446 | fprintf (dump_file, | |
2447 | " FAILED: it is not a part of reduction.\n"); | |
2448 | return false; | |
2449 | } | |
2450 | reduc_phi = NULL; | |
2451 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, val) | |
2452 | { | |
4942af9b JJ |
2453 | if (!gimple_debug_bind_p (USE_STMT (use_p)) |
2454 | && flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p)))) | |
08dab97a RL |
2455 | { |
2456 | reduc_phi = USE_STMT (use_p); | |
2457 | break; | |
2458 | } | |
2459 | } | |
2460 | red = reduction_phi (reduction_list, reduc_phi); | |
2461 | if (red == NULL) | |
2462 | { | |
2463 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2464 | fprintf (dump_file, | |
2465 | " FAILED: it is not a part of reduction.\n"); | |
2466 | return false; | |
2467 | } | |
2468 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2469 | { | |
2470 | fprintf (dump_file, "reduction phi is "); | |
2471 | print_gimple_stmt (dump_file, red->reduc_phi, 0, 0); | |
2472 | fprintf (dump_file, "reduction stmt is "); | |
2473 | print_gimple_stmt (dump_file, red->reduc_stmt, 0, 0); | |
2474 | } | |
2475 | } | |
2476 | } | |
2477 | ||
2478 | /* The iterations of the loop may communicate only through bivs whose | |
2479 | iteration space can be distributed efficiently. */ | |
2480 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2481 | { | |
538dd0b7 | 2482 | gphi *phi = gsi.phi (); |
08dab97a RL |
2483 | tree def = PHI_RESULT (phi); |
2484 | affine_iv iv; | |
2485 | ||
ea057359 | 2486 | if (!virtual_operand_p (def) && !simple_iv (loop, loop, def, &iv, true)) |
08dab97a RL |
2487 | { |
2488 | struct reduction_info *red; | |
2489 | ||
2490 | red = reduction_phi (reduction_list, phi); | |
2491 | if (red == NULL) | |
2492 | { | |
2493 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2494 | fprintf (dump_file, | |
2495 | " FAILED: scalar dependency between iterations\n"); | |
2496 | return false; | |
2497 | } | |
2498 | } | |
2499 | } | |
2500 | ||
2501 | ||
2502 | return true; | |
2503 | } | |
2504 | ||
5f40b3cb ZD |
2505 | /* Detect parallel loops and generate parallel code using libgomp |
2506 | primitives. Returns true if some loop was parallelized, false | |
2507 | otherwise. */ | |
2508 | ||
09489eb8 | 2509 | static bool |
5f40b3cb ZD |
2510 | parallelize_loops (void) |
2511 | { | |
2512 | unsigned n_threads = flag_tree_parallelize_loops; | |
2513 | bool changed = false; | |
2514 | struct loop *loop; | |
2515 | struct tree_niter_desc niter_desc; | |
f873b205 | 2516 | struct obstack parloop_obstack; |
8adfe01d | 2517 | HOST_WIDE_INT estimated; |
b05e0233 | 2518 | source_location loop_loc; |
f873b205 | 2519 | |
5f40b3cb ZD |
2520 | /* Do not parallelize loops in the functions created by parallelization. */ |
2521 | if (parallelized_function_p (cfun->decl)) | |
2522 | return false; | |
8adfe01d RL |
2523 | if (cfun->has_nonlocal_label) |
2524 | return false; | |
5f40b3cb | 2525 | |
f873b205 | 2526 | gcc_obstack_init (&parloop_obstack); |
c203e8a7 | 2527 | reduction_info_table_type reduction_list (10); |
726a989a | 2528 | init_stmt_vec_info_vec (); |
a509ebb5 | 2529 | |
f0bd40b1 | 2530 | FOR_EACH_LOOP (loop, 0) |
5f40b3cb | 2531 | { |
4a8fb1a1 | 2532 | reduction_list.empty (); |
48710229 RL |
2533 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2534 | { | |
2535 | fprintf (dump_file, "Trying loop %d as candidate\n",loop->num); | |
2536 | if (loop->inner) | |
2537 | fprintf (dump_file, "loop %d is not innermost\n",loop->num); | |
2538 | else | |
2539 | fprintf (dump_file, "loop %d is innermost\n",loop->num); | |
2540 | } | |
b8698a0f | 2541 | |
48710229 | 2542 | /* If we use autopar in graphite pass, we use its marked dependency |
87d4d0ee SP |
2543 | checking results. */ |
2544 | if (flag_loop_parallelize_all && !loop->can_be_parallel) | |
48710229 RL |
2545 | { |
2546 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
2547 | fprintf (dump_file, "loop is not parallel according to graphite\n"); | |
87d4d0ee | 2548 | continue; |
48710229 | 2549 | } |
87d4d0ee | 2550 | |
48710229 RL |
2551 | if (!single_dom_exit (loop)) |
2552 | { | |
b8698a0f | 2553 | |
48710229 RL |
2554 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2555 | fprintf (dump_file, "loop is !single_dom_exit\n"); | |
b8698a0f | 2556 | |
08dab97a | 2557 | continue; |
48710229 | 2558 | } |
08dab97a RL |
2559 | |
2560 | if (/* And of course, the loop must be parallelizable. */ | |
2561 | !can_duplicate_loop_p (loop) | |
1d4af1e8 | 2562 | || loop_has_blocks_with_irreducible_flag (loop) |
8adfe01d | 2563 | || (loop_preheader_edge (loop)->src->flags & BB_IRREDUCIBLE_LOOP) |
9857228c | 2564 | /* FIXME: the check for vector phi nodes could be removed. */ |
69958396 | 2565 | || loop_has_vector_phi_nodes (loop)) |
08dab97a | 2566 | continue; |
e5b332cd | 2567 | |
652c4c71 | 2568 | estimated = estimated_stmt_executions_int (loop); |
e5b332cd RG |
2569 | if (estimated == -1) |
2570 | estimated = max_stmt_executions_int (loop); | |
87d4d0ee | 2571 | /* FIXME: Bypass this check as graphite doesn't update the |
e5b332cd | 2572 | count and frequency correctly now. */ |
87d4d0ee | 2573 | if (!flag_loop_parallelize_all |
e5b332cd RG |
2574 | && ((estimated != -1 |
2575 | && estimated <= (HOST_WIDE_INT) n_threads * MIN_PER_THREAD) | |
87d4d0ee SP |
2576 | /* Do not bother with loops in cold areas. */ |
2577 | || optimize_loop_nest_for_size_p (loop))) | |
08dab97a | 2578 | continue; |
b8698a0f | 2579 | |
08dab97a RL |
2580 | if (!try_get_loop_niter (loop, &niter_desc)) |
2581 | continue; | |
2582 | ||
c203e8a7 | 2583 | if (!try_create_reduction_list (loop, &reduction_list)) |
08dab97a RL |
2584 | continue; |
2585 | ||
f873b205 LB |
2586 | if (!flag_loop_parallelize_all |
2587 | && !loop_parallel_p (loop, &parloop_obstack)) | |
5f40b3cb ZD |
2588 | continue; |
2589 | ||
2590 | changed = true; | |
48710229 RL |
2591 | if (dump_file && (dump_flags & TDF_DETAILS)) |
2592 | { | |
48710229 | 2593 | if (loop->inner) |
8adfe01d | 2594 | fprintf (dump_file, "parallelizing outer loop %d\n",loop->header->index); |
48710229 | 2595 | else |
8adfe01d RL |
2596 | fprintf (dump_file, "parallelizing inner loop %d\n",loop->header->index); |
2597 | loop_loc = find_loop_location (loop); | |
b05e0233 | 2598 | if (loop_loc != UNKNOWN_LOCATION) |
8adfe01d | 2599 | fprintf (dump_file, "\nloop at %s:%d: ", |
b05e0233 | 2600 | LOCATION_FILE (loop_loc), LOCATION_LINE (loop_loc)); |
b8698a0f | 2601 | } |
c203e8a7 | 2602 | gen_parallel_loop (loop, &reduction_list, |
08dab97a | 2603 | n_threads, &niter_desc); |
5f40b3cb ZD |
2604 | } |
2605 | ||
726a989a | 2606 | free_stmt_vec_info_vec (); |
f873b205 | 2607 | obstack_free (&parloop_obstack, NULL); |
6b8ed145 RG |
2608 | |
2609 | /* Parallelization will cause new function calls to be inserted through | |
d086d311 RG |
2610 | which local variables will escape. Reset the points-to solution |
2611 | for ESCAPED. */ | |
6b8ed145 | 2612 | if (changed) |
d086d311 | 2613 | pt_solution_reset (&cfun->gimple_df->escaped); |
6b8ed145 | 2614 | |
5f40b3cb ZD |
2615 | return changed; |
2616 | } | |
2617 | ||
c1bf2a39 AM |
2618 | /* Parallelization. */ |
2619 | ||
c1bf2a39 AM |
2620 | namespace { |
2621 | ||
2622 | const pass_data pass_data_parallelize_loops = | |
2623 | { | |
2624 | GIMPLE_PASS, /* type */ | |
2625 | "parloops", /* name */ | |
2626 | OPTGROUP_LOOP, /* optinfo_flags */ | |
c1bf2a39 AM |
2627 | TV_TREE_PARALLELIZE_LOOPS, /* tv_id */ |
2628 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
2629 | 0, /* properties_provided */ | |
2630 | 0, /* properties_destroyed */ | |
2631 | 0, /* todo_flags_start */ | |
3bea341f | 2632 | 0, /* todo_flags_finish */ |
c1bf2a39 AM |
2633 | }; |
2634 | ||
2635 | class pass_parallelize_loops : public gimple_opt_pass | |
2636 | { | |
2637 | public: | |
2638 | pass_parallelize_loops (gcc::context *ctxt) | |
2639 | : gimple_opt_pass (pass_data_parallelize_loops, ctxt) | |
2640 | {} | |
2641 | ||
2642 | /* opt_pass methods: */ | |
1a3d085c | 2643 | virtual bool gate (function *) { return flag_tree_parallelize_loops > 1; } |
be55bfe6 | 2644 | virtual unsigned int execute (function *); |
c1bf2a39 AM |
2645 | |
2646 | }; // class pass_parallelize_loops | |
2647 | ||
be55bfe6 TS |
2648 | unsigned |
2649 | pass_parallelize_loops::execute (function *fun) | |
2650 | { | |
2651 | if (number_of_loops (fun) <= 1) | |
2652 | return 0; | |
2653 | ||
2654 | if (parallelize_loops ()) | |
18751894 TV |
2655 | { |
2656 | fun->curr_properties &= ~(PROP_gimple_eomp); | |
2657 | return TODO_update_ssa; | |
2658 | } | |
2659 | ||
be55bfe6 TS |
2660 | return 0; |
2661 | } | |
2662 | ||
c1bf2a39 AM |
2663 | } // anon namespace |
2664 | ||
2665 | gimple_opt_pass * | |
2666 | make_pass_parallelize_loops (gcc::context *ctxt) | |
2667 | { | |
2668 | return new pass_parallelize_loops (ctxt); | |
2669 | } |