]>
Commit | Line | Data |
---|---|---|
5f40b3cb | 1 | /* Loop autoparallelization. |
5576d6f2 | 2 | Copyright (C) 2006, 2007 Free Software Foundation, Inc. |
5f40b3cb ZD |
3 | Contributed by Sebastian Pop <pop@cri.ensmp.fr> and |
4 | Zdenek Dvorak <dvorakz@suse.cz>. | |
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 | |
10 | Software Foundation; either version 2, or (at your option) any later | |
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 | |
19 | along with GCC; see the file COPYING. If not, write to the Free | |
20 | Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA | |
21 | 02110-1301, USA. */ | |
22 | ||
23 | #include "config.h" | |
24 | #include "system.h" | |
25 | #include "coretypes.h" | |
26 | #include "tm.h" | |
27 | #include "tree.h" | |
28 | #include "rtl.h" | |
29 | #include "tree-flow.h" | |
30 | #include "cfgloop.h" | |
31 | #include "ggc.h" | |
32 | #include "tree-data-ref.h" | |
33 | #include "diagnostic.h" | |
34 | #include "tree-pass.h" | |
35 | #include "tree-scalar-evolution.h" | |
36 | #include "hashtab.h" | |
37 | #include "langhooks.h" | |
a509ebb5 | 38 | #include "tree-vectorizer.h" |
5f40b3cb ZD |
39 | |
40 | /* This pass tries to distribute iterations of loops into several threads. | |
41 | The implementation is straightforward -- for each loop we test whether its | |
42 | iterations are independent, and if it is the case (and some additional | |
43 | conditions regarding profitability and correctness are satisfied), we | |
44 | add OMP_PARALLEL and OMP_FOR codes and let omp expansion machinery do | |
45 | its job. | |
46 | ||
47 | The most of the complexity is in bringing the code into shape expected | |
48 | by the omp expanders: | |
49 | -- for OMP_FOR, ensuring that the loop has only one induction variable | |
50 | and that the exit test is at the start of the loop body | |
51 | -- for OMP_PARALLEL, replacing the references to local addressable | |
52 | variables by accesses through pointers, and breaking up ssa chains | |
53 | by storing the values incoming to the parallelized loop to a structure | |
54 | passed to the new function as an argument (something similar is done | |
55 | in omp gimplification, unfortunately only a small part of the code | |
56 | can be shared). | |
57 | ||
58 | TODO: | |
59 | -- if there are several parallelizable loops in a function, it may be | |
60 | possible to generate the threads just once (using synchronization to | |
61 | ensure that cross-loop dependences are obeyed). | |
62 | -- handling of common scalar dependence patterns (accumulation, ...) | |
63 | -- handling of non-innermost loops */ | |
64 | ||
a509ebb5 RL |
65 | /* |
66 | Reduction handling: | |
67 | currently we use vect_is_simple_reduction() to detect reduction patterns. | |
68 | The code transformation will be introduced by an example. | |
69 | ||
0eb7e7aa | 70 | |
a509ebb5 RL |
71 | parloop |
72 | { | |
73 | int sum=1; | |
74 | ||
0eb7e7aa | 75 | for (i = 0; i < N; i++) |
a509ebb5 RL |
76 | { |
77 | x[i] = i + 3; | |
78 | sum+=x[i]; | |
79 | } | |
80 | } | |
81 | ||
0eb7e7aa | 82 | gimple-like code: |
a509ebb5 RL |
83 | header_bb: |
84 | ||
0eb7e7aa RL |
85 | # sum_29 = PHI <sum_11(5), 1(3)> |
86 | # i_28 = PHI <i_12(5), 0(3)> | |
87 | D.1795_8 = i_28 + 3; | |
88 | x[i_28] = D.1795_8; | |
89 | sum_11 = D.1795_8 + sum_29; | |
90 | i_12 = i_28 + 1; | |
91 | if (N_6(D) > i_12) | |
92 | goto header_bb; | |
93 | ||
a509ebb5 RL |
94 | |
95 | exit_bb: | |
96 | ||
0eb7e7aa RL |
97 | # sum_21 = PHI <sum_11(4)> |
98 | printf (&"%d"[0], sum_21); | |
a509ebb5 RL |
99 | |
100 | ||
101 | after reduction transformation (only relevant parts): | |
102 | ||
103 | parloop | |
104 | { | |
105 | ||
106 | .... | |
107 | ||
0eb7e7aa RL |
108 | |
109 | # A new variable is created for each reduction: | |
110 | "reduction_initial" is the initial value given by the user. | |
111 | It is kept and will be used after the parallel computing is done. # | |
112 | ||
113 | reduction_initial.24_46 = 1; | |
a509ebb5 | 114 | |
0eb7e7aa RL |
115 | # Storing the neutral value of the |
116 | particular reduction's operation, e.g. 0 for PLUS_EXPR, | |
117 | 1 for MULT_EXPR, etc. into the reduction field. | |
118 | This is done in create_stores_for_reduction. # | |
119 | ||
120 | .paral_data_store.32.sum.27 = 0; | |
121 | ||
122 | #pragma omp parallel num_threads(4) | |
a509ebb5 | 123 | |
0eb7e7aa RL |
124 | #pragma omp for schedule(static) |
125 | # sum.27_29 = PHI <sum.27_11, 0> # The neutral element replaces | |
126 | the user's inital value. # | |
127 | sum.27_11 = D.1827_8 + sum.27_29; | |
128 | OMP_CONTINUE | |
a509ebb5 | 129 | |
0eb7e7aa RL |
130 | # Adding this reduction phi is done at create_phi_for_local_result() # |
131 | # sum.27_56 = PHI <sum.27_11, 0> | |
132 | OMP_RETURN | |
133 | ||
134 | # Creating the atomic operation is done at | |
135 | create_call_for_reduction_1() # | |
a509ebb5 | 136 | |
0eb7e7aa RL |
137 | #pragma omp atomic_load |
138 | D.1839_59 = *&.paral_data_load.33_51->reduction.23; | |
139 | D.1840_60 = sum.27_56 + D.1839_59; | |
140 | #pragma omp atomic_store (D.1840_60); | |
a509ebb5 | 141 | |
0eb7e7aa RL |
142 | OMP_RETURN |
143 | ||
144 | # collecting the result after the join of the threads is done at | |
145 | create_loads_for_reductions(). | |
146 | a new variable "reduction_final" is created. It calculates the final | |
147 | value from the initial value and the value computed by the threads # | |
148 | ||
149 | .paral_data_load.33_52 = &.paral_data_store.32; | |
150 | reduction_final.34_53 = .paral_data_load.33_52->sum.27; | |
151 | sum_37 = reduction_initial.24_46 + reduction_final.34_53; | |
152 | sum_43 = D.1795_41 + sum_37; | |
153 | ||
154 | exit bb: | |
155 | # sum_21 = PHI <sum_43, sum_26> | |
156 | printf (&"%d"[0], sum_21); | |
157 | ||
158 | ... | |
159 | ||
a509ebb5 RL |
160 | } |
161 | ||
162 | */ | |
163 | ||
5f40b3cb ZD |
164 | /* Minimal number of iterations of a loop that should be executed in each |
165 | thread. */ | |
166 | #define MIN_PER_THREAD 100 | |
167 | ||
a509ebb5 RL |
168 | /* Element of the hashtable, representing a |
169 | reduction in the current loop. */ | |
170 | struct reduction_info | |
171 | { | |
172 | tree reduc_stmt; /* reduction statement. */ | |
173 | tree reduc_phi; /* The phi node defining the reduction. */ | |
174 | enum tree_code reduction_code; /* code for the reduction operation. */ | |
175 | tree keep_res; /* The PHI_RESULT of this phi is the resulting value | |
176 | of the reduction variable when existing the loop. */ | |
177 | tree initial_value; /* An ssa name representing a new variable holding | |
178 | the initial value of the reduction var before entering the loop. */ | |
179 | tree field; /* the name of the field in the parloop data structure intended for reduction. */ | |
a509ebb5 RL |
180 | tree init; /* reduction initialization value. */ |
181 | tree new_phi; /* (helper field) Newly created phi node whose result | |
182 | will be passed to the atomic operation. Represents | |
183 | the local result each thread computed for the reduction | |
184 | operation. */ | |
185 | }; | |
186 | ||
187 | /* Equality and hash functions for hashtab code. */ | |
188 | ||
189 | static int | |
190 | reduction_info_eq (const void *aa, const void *bb) | |
191 | { | |
192 | const struct reduction_info *a = (const struct reduction_info *) aa; | |
193 | const struct reduction_info *b = (const struct reduction_info *) bb; | |
194 | ||
195 | return (a->reduc_phi == b->reduc_phi); | |
196 | } | |
197 | ||
198 | static hashval_t | |
199 | reduction_info_hash (const void *aa) | |
200 | { | |
201 | const struct reduction_info *a = (const struct reduction_info *) aa; | |
202 | ||
203 | return htab_hash_pointer (a->reduc_phi); | |
204 | } | |
205 | ||
206 | static struct reduction_info * | |
207 | reduction_phi (htab_t reduction_list, tree phi) | |
208 | { | |
209 | struct reduction_info tmpred, *red; | |
210 | ||
211 | if (htab_elements (reduction_list) == 0) | |
212 | return NULL; | |
213 | ||
214 | tmpred.reduc_phi = phi; | |
215 | red = htab_find (reduction_list, &tmpred); | |
216 | ||
217 | return red; | |
218 | } | |
219 | ||
5f40b3cb ZD |
220 | /* Element of hashtable of names to copy. */ |
221 | ||
222 | struct name_to_copy_elt | |
223 | { | |
224 | unsigned version; /* The version of the name to copy. */ | |
225 | tree new_name; /* The new name used in the copy. */ | |
226 | tree field; /* The field of the structure used to pass the | |
227 | value. */ | |
228 | }; | |
229 | ||
230 | /* Equality and hash functions for hashtab code. */ | |
231 | ||
232 | static int | |
233 | name_to_copy_elt_eq (const void *aa, const void *bb) | |
234 | { | |
a509ebb5 RL |
235 | const struct name_to_copy_elt *a = (const struct name_to_copy_elt *) aa; |
236 | const struct name_to_copy_elt *b = (const struct name_to_copy_elt *) bb; | |
5f40b3cb ZD |
237 | |
238 | return a->version == b->version; | |
239 | } | |
240 | ||
241 | static hashval_t | |
242 | name_to_copy_elt_hash (const void *aa) | |
243 | { | |
a509ebb5 | 244 | const struct name_to_copy_elt *a = (const struct name_to_copy_elt *) aa; |
5f40b3cb ZD |
245 | |
246 | return (hashval_t) a->version; | |
247 | } | |
248 | ||
249 | /* Returns true if the iterations of LOOP are independent on each other (that | |
250 | is, if we can execute them in parallel), and if LOOP satisfies other | |
251 | conditions that we need to be able to parallelize it. Description of number | |
a509ebb5 RL |
252 | of iterations is stored to NITER. Reduction analysis is done, if |
253 | reductions are found, they are inserted to the REDUCTION_LIST. */ | |
5f40b3cb ZD |
254 | |
255 | static bool | |
a509ebb5 | 256 | loop_parallel_p (struct loop *loop, htab_t reduction_list, struct tree_niter_desc *niter) |
5f40b3cb ZD |
257 | { |
258 | edge exit = single_dom_exit (loop); | |
a509ebb5 RL |
259 | VEC (ddr_p, heap) * dependence_relations; |
260 | VEC (data_reference_p, heap) * datarefs; | |
5f40b3cb ZD |
261 | lambda_trans_matrix trans; |
262 | bool ret = false; | |
263 | tree phi; | |
a509ebb5 | 264 | loop_vec_info simple_loop_info; |
5f40b3cb ZD |
265 | |
266 | /* Only consider innermost loops with just one exit. The innermost-loop | |
267 | restriction is not necessary, but it makes things simpler. */ | |
268 | if (loop->inner || !exit) | |
269 | return false; | |
270 | ||
271 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
272 | fprintf (dump_file, "\nConsidering loop %d\n", loop->num); | |
273 | ||
274 | /* We need to know # of iterations, and there should be no uses of values | |
275 | defined inside loop outside of it, unless the values are invariants of | |
276 | the loop. */ | |
277 | if (!number_of_iterations_exit (loop, exit, niter, false)) | |
278 | { | |
279 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
280 | fprintf (dump_file, " FAILED: number of iterations not known\n"); | |
281 | return false; | |
282 | } | |
283 | ||
a509ebb5 RL |
284 | simple_loop_info = vect_analyze_loop_form (loop); |
285 | ||
286 | for (phi = phi_nodes (loop->header); phi; phi = PHI_CHAIN (phi)) | |
287 | { | |
288 | tree reduc_stmt = NULL, operation; | |
289 | ||
290 | /* ??? TODO: Change this into a generic function that | |
291 | recognizes reductions. */ | |
292 | if (!is_gimple_reg (PHI_RESULT (phi))) | |
293 | continue; | |
294 | if (simple_loop_info) | |
295 | reduc_stmt = vect_is_simple_reduction (simple_loop_info, phi); | |
296 | ||
297 | /* Create a reduction_info struct, initialize it and insert it to | |
298 | the reduction list. */ | |
299 | ||
300 | if (reduc_stmt) | |
301 | { | |
302 | PTR *slot; | |
303 | struct reduction_info *new_reduction; | |
304 | ||
305 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
306 | { | |
307 | fprintf (dump_file, | |
308 | "Detected reduction. reduction stmt is: \n"); | |
309 | print_generic_stmt (dump_file, reduc_stmt, 0); | |
310 | fprintf (dump_file, "\n"); | |
311 | } | |
312 | ||
313 | new_reduction = XCNEW (struct reduction_info); | |
314 | ||
315 | new_reduction->reduc_stmt = reduc_stmt; | |
316 | new_reduction->reduc_phi = phi; | |
317 | operation = GIMPLE_STMT_OPERAND (reduc_stmt, 1); | |
318 | new_reduction->reduction_code = TREE_CODE (operation); | |
319 | slot = htab_find_slot (reduction_list, new_reduction, INSERT); | |
320 | *slot = new_reduction; | |
321 | } | |
322 | } | |
323 | ||
5f40b3cb ZD |
324 | for (phi = phi_nodes (exit->dest); phi; phi = PHI_CHAIN (phi)) |
325 | { | |
a509ebb5 RL |
326 | struct reduction_info *red; |
327 | imm_use_iterator imm_iter; | |
328 | use_operand_p use_p; | |
329 | tree reduc_phi; | |
330 | ||
5f40b3cb ZD |
331 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
332 | ||
333 | if (is_gimple_reg (val)) | |
334 | { | |
335 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
a509ebb5 RL |
336 | { |
337 | fprintf (dump_file, "phi is "); | |
338 | print_generic_expr (dump_file, phi, 0); | |
339 | fprintf (dump_file, "arg of phi to exit: value "); | |
340 | print_generic_expr (dump_file, val, 0); | |
341 | fprintf (dump_file, " used outside loop\n"); | |
342 | fprintf (dump_file, | |
343 | " checking if it a part of reduction pattern: \n"); | |
344 | } | |
345 | if (htab_elements (reduction_list) == 0) | |
346 | { | |
347 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
348 | fprintf (dump_file, | |
349 | " FAILED: it is not a part of reduction.\n"); | |
350 | return false; | |
351 | } | |
352 | reduc_phi = NULL; | |
353 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, val) | |
354 | { | |
355 | if (flow_bb_inside_loop_p (loop, bb_for_stmt (USE_STMT (use_p)))) | |
356 | { | |
357 | reduc_phi = USE_STMT (use_p); | |
358 | break; | |
359 | } | |
360 | } | |
361 | red = reduction_phi (reduction_list, reduc_phi); | |
362 | if (red == NULL) | |
363 | { | |
364 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
365 | fprintf (dump_file, | |
366 | " FAILED: it is not a part of reduction.\n"); | |
367 | return false; | |
368 | } | |
369 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
370 | { | |
371 | fprintf (dump_file, "reduction phi is "); | |
372 | print_generic_expr (dump_file, red->reduc_phi, 0); | |
373 | fprintf (dump_file, "reduction stmt is "); | |
374 | print_generic_expr (dump_file, red->reduc_stmt, 0); | |
375 | } | |
376 | ||
5f40b3cb ZD |
377 | } |
378 | } | |
379 | ||
380 | /* The iterations of the loop may communicate only through bivs whose | |
381 | iteration space can be distributed efficiently. */ | |
382 | for (phi = phi_nodes (loop->header); phi; phi = PHI_CHAIN (phi)) | |
383 | { | |
384 | tree def = PHI_RESULT (phi); | |
385 | affine_iv iv; | |
386 | ||
a509ebb5 | 387 | if (is_gimple_reg (def) && !simple_iv (loop, phi, def, &iv, true)) |
5f40b3cb | 388 | { |
a509ebb5 RL |
389 | struct reduction_info *red; |
390 | ||
391 | red = reduction_phi (reduction_list, phi); | |
392 | if (red == NULL) | |
393 | { | |
394 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
395 | fprintf (dump_file, | |
396 | " FAILED: scalar dependency between iterations\n"); | |
397 | return false; | |
398 | } | |
5f40b3cb ZD |
399 | } |
400 | } | |
401 | ||
402 | /* We need to version the loop to verify assumptions in runtime. */ | |
403 | if (!can_duplicate_loop_p (loop)) | |
404 | { | |
405 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
406 | fprintf (dump_file, " FAILED: cannot be duplicated\n"); | |
407 | return false; | |
408 | } | |
409 | ||
410 | /* Check for problems with dependences. If the loop can be reversed, | |
411 | the iterations are independent. */ | |
412 | datarefs = VEC_alloc (data_reference_p, heap, 10); | |
413 | dependence_relations = VEC_alloc (ddr_p, heap, 10 * 10); | |
414 | compute_data_dependences_for_loop (loop, true, &datarefs, | |
415 | &dependence_relations); | |
416 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
417 | dump_data_dependence_relations (dump_file, dependence_relations); | |
418 | ||
419 | trans = lambda_trans_matrix_new (1, 1); | |
420 | LTM_MATRIX (trans)[0][0] = -1; | |
421 | ||
422 | if (lambda_transform_legal_p (trans, 1, dependence_relations)) | |
423 | { | |
424 | ret = true; | |
425 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
426 | fprintf (dump_file, " SUCCESS: may be parallelized\n"); | |
427 | } | |
428 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
a509ebb5 RL |
429 | fprintf (dump_file, |
430 | " FAILED: data dependencies exist across iterations\n"); | |
5f40b3cb ZD |
431 | |
432 | free_dependence_relations (dependence_relations); | |
433 | free_data_refs (datarefs); | |
434 | ||
435 | return ret; | |
436 | } | |
437 | ||
438 | /* Assigns the address of VAR in TYPE to an ssa name, and returns this name. | |
439 | The assignment statement is placed before LOOP. DECL_ADDRESS maps decls | |
440 | to their addresses that can be reused. */ | |
441 | ||
442 | static tree | |
443 | take_address_of (tree var, tree type, struct loop *loop, htab_t decl_address) | |
444 | { | |
445 | int uid = DECL_UID (var); | |
446 | void **dslot; | |
447 | struct int_tree_map ielt, *nielt; | |
448 | tree name, bvar, stmt; | |
449 | edge entry = loop_preheader_edge (loop); | |
450 | ||
451 | ielt.uid = uid; | |
452 | dslot = htab_find_slot_with_hash (decl_address, &ielt, uid, INSERT); | |
453 | if (!*dslot) | |
454 | { | |
455 | bvar = create_tmp_var (type, get_name (var)); | |
456 | add_referenced_var (bvar); | |
457 | stmt = build_gimple_modify_stmt (bvar, | |
a509ebb5 RL |
458 | fold_convert (type, |
459 | build_addr (var, | |
460 | current_function_decl))); | |
5f40b3cb ZD |
461 | name = make_ssa_name (bvar, stmt); |
462 | GIMPLE_STMT_OPERAND (stmt, 0) = name; | |
463 | bsi_insert_on_edge_immediate (entry, stmt); | |
464 | ||
465 | nielt = XNEW (struct int_tree_map); | |
466 | nielt->uid = uid; | |
467 | nielt->to = name; | |
468 | *dslot = nielt; | |
469 | ||
470 | return name; | |
471 | } | |
472 | ||
473 | name = ((struct int_tree_map *) *dslot)->to; | |
474 | if (TREE_TYPE (name) == type) | |
475 | return name; | |
476 | ||
477 | bvar = SSA_NAME_VAR (name); | |
a509ebb5 | 478 | stmt = build_gimple_modify_stmt (bvar, fold_convert (type, name)); |
5f40b3cb ZD |
479 | name = make_ssa_name (bvar, stmt); |
480 | GIMPLE_STMT_OPERAND (stmt, 0) = name; | |
481 | bsi_insert_on_edge_immediate (entry, stmt); | |
482 | ||
483 | return name; | |
484 | } | |
485 | ||
a509ebb5 RL |
486 | /* Callback for htab_traverse. Create the initialization statement |
487 | for reduction described in SLOT, and place it at the preheader of | |
488 | the loop described in DATA. */ | |
489 | ||
490 | static int | |
491 | initialize_reductions (void **slot, void *data) | |
492 | { | |
0eb7e7aa | 493 | tree stmt; |
a509ebb5 | 494 | tree init, c; |
0eb7e7aa | 495 | tree name1; |
a509ebb5 RL |
496 | tree bvar, type, arg; |
497 | edge e; | |
498 | ||
499 | struct reduction_info *reduc = *slot; | |
500 | struct loop *loop = (struct loop *) data; | |
501 | ||
502 | /* Create initialization in preheader: | |
503 | reduction_variable = initialization value of reduction. */ | |
504 | ||
505 | /* In the phi node at the header, replace the argument coming | |
506 | from the preheader with the reduction initialization value. */ | |
507 | ||
508 | /* Create a new variable to initialize the reduction. */ | |
509 | type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); | |
510 | bvar = create_tmp_var (type, "reduction"); | |
511 | add_referenced_var (bvar); | |
512 | ||
513 | c = build_omp_clause (OMP_CLAUSE_REDUCTION); | |
514 | OMP_CLAUSE_REDUCTION_CODE (c) = reduc->reduction_code; | |
515 | OMP_CLAUSE_DECL (c) = | |
516 | SSA_NAME_VAR (GIMPLE_STMT_OPERAND (reduc->reduc_stmt, 0)); | |
517 | ||
518 | init = omp_reduction_init (c, TREE_TYPE (bvar)); | |
519 | reduc->init = init; | |
520 | ||
0eb7e7aa RL |
521 | /* Replace the argument representing the initialization value |
522 | with the initialization value for the reduction (neutral | |
523 | element for the particular operation, e.g. 0 for PLUS_EXPR, | |
524 | 1 for MULT_EXPR, etc). | |
525 | Keep the old value in a new variable "reduction_initial", | |
526 | that will be taken in consideration after the parallel | |
527 | computing is done. */ | |
a509ebb5 RL |
528 | |
529 | e = loop_preheader_edge (loop); | |
530 | arg = PHI_ARG_DEF_FROM_EDGE (reduc->reduc_phi, e); | |
531 | /* Create new variable to hold the initial value. */ | |
532 | type = TREE_TYPE (bvar); | |
533 | bvar = create_tmp_var (type, "reduction_initial"); | |
534 | add_referenced_var (bvar); | |
535 | ||
536 | stmt = build_gimple_modify_stmt (bvar, arg); | |
537 | name1 = make_ssa_name (bvar, stmt); | |
538 | GIMPLE_STMT_OPERAND (stmt, 0) = name1; | |
539 | SSA_NAME_DEF_STMT (name1) = stmt; | |
540 | ||
541 | bsi_insert_on_edge_immediate (e, stmt); | |
a509ebb5 | 542 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE |
0eb7e7aa | 543 | (reduc->reduc_phi, loop_preheader_edge (loop)), init); |
a509ebb5 | 544 | reduc->initial_value = name1; |
a509ebb5 RL |
545 | return 1; |
546 | } | |
5f40b3cb ZD |
547 | |
548 | struct elv_data | |
549 | { | |
550 | struct loop *loop; | |
551 | htab_t decl_address; | |
552 | bool changed; | |
553 | }; | |
554 | ||
a509ebb5 RL |
555 | /* Eliminates references to local variables in *TP out of LOOP. DECL_ADDRESS |
556 | contains addresses of the references that had their address taken already. | |
557 | If the expression is changed, CHANGED is set to true. Callback for | |
558 | walk_tree. */ | |
559 | ||
5f40b3cb | 560 | static tree |
a509ebb5 | 561 | eliminate_local_variables_1 (tree * tp, int *walk_subtrees, void *data) |
5f40b3cb ZD |
562 | { |
563 | struct elv_data *dta = data; | |
564 | tree t = *tp, var, addr, addr_type, type; | |
565 | ||
566 | if (DECL_P (t)) | |
567 | { | |
568 | *walk_subtrees = 0; | |
569 | ||
570 | if (!SSA_VAR_P (t) || DECL_EXTERNAL (t)) | |
571 | return NULL_TREE; | |
572 | ||
573 | type = TREE_TYPE (t); | |
574 | addr_type = build_pointer_type (type); | |
575 | addr = take_address_of (t, addr_type, dta->loop, dta->decl_address); | |
576 | *tp = build1 (INDIRECT_REF, TREE_TYPE (*tp), addr); | |
577 | ||
578 | dta->changed = true; | |
579 | return NULL_TREE; | |
580 | } | |
581 | ||
582 | if (TREE_CODE (t) == ADDR_EXPR) | |
583 | { | |
584 | var = TREE_OPERAND (t, 0); | |
585 | if (!DECL_P (var)) | |
586 | return NULL_TREE; | |
587 | ||
588 | *walk_subtrees = 0; | |
589 | if (!SSA_VAR_P (var) || DECL_EXTERNAL (var)) | |
590 | return NULL_TREE; | |
591 | ||
592 | addr_type = TREE_TYPE (t); | |
593 | addr = take_address_of (var, addr_type, dta->loop, dta->decl_address); | |
594 | *tp = addr; | |
595 | ||
596 | dta->changed = true; | |
597 | return NULL_TREE; | |
598 | } | |
599 | ||
a509ebb5 | 600 | if (!EXPR_P (t) && !GIMPLE_STMT_P (t)) |
5f40b3cb ZD |
601 | *walk_subtrees = 0; |
602 | ||
603 | return NULL_TREE; | |
604 | } | |
605 | ||
606 | /* Moves the references to local variables in STMT from LOOP. DECL_ADDRESS | |
607 | contains addresses for the references for that we have already taken | |
608 | them. */ | |
609 | ||
610 | static void | |
611 | eliminate_local_variables_stmt (struct loop *loop, tree stmt, | |
612 | htab_t decl_address) | |
613 | { | |
614 | struct elv_data dta; | |
615 | ||
616 | dta.loop = loop; | |
617 | dta.decl_address = decl_address; | |
618 | dta.changed = false; | |
619 | ||
620 | walk_tree (&stmt, eliminate_local_variables_1, &dta, NULL); | |
621 | ||
622 | if (dta.changed) | |
623 | update_stmt (stmt); | |
624 | } | |
625 | ||
a509ebb5 RL |
626 | /* Eliminates the references to local variables from LOOP. |
627 | This includes: | |
628 | 1) Taking address of a local variable -- these are moved out of the | |
629 | loop (and temporary variable is created to hold the address if | |
630 | necessary). | |
5f40b3cb | 631 | 2) Dereferencing a local variable -- these are replaced with indirect |
a509ebb5 | 632 | references. */ |
5f40b3cb ZD |
633 | |
634 | static void | |
635 | eliminate_local_variables (struct loop *loop) | |
636 | { | |
637 | basic_block bb, *body = get_loop_body (loop); | |
638 | unsigned i; | |
639 | block_stmt_iterator bsi; | |
640 | htab_t decl_address = htab_create (10, int_tree_map_hash, int_tree_map_eq, | |
641 | free); | |
642 | ||
643 | /* Find and rename the ssa names defined outside of loop. */ | |
644 | for (i = 0; i < loop->num_nodes; i++) | |
645 | { | |
646 | bb = body[i]; | |
647 | ||
648 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) | |
649 | eliminate_local_variables_stmt (loop, bsi_stmt (bsi), decl_address); | |
650 | } | |
651 | ||
652 | htab_delete (decl_address); | |
653 | } | |
654 | ||
655 | /* If COPY_NAME_P is true, creates and returns a duplicate of NAME. | |
656 | The copies are stored to NAME_COPIES, if NAME was already duplicated, | |
657 | its duplicate stored in NAME_COPIES is returned. | |
658 | ||
659 | Regardless of COPY_NAME_P, the decl used as a base of the ssa name is also | |
660 | duplicated, storing the copies in DECL_COPIES. */ | |
661 | ||
662 | static tree | |
663 | separate_decls_in_loop_name (tree name, | |
664 | htab_t name_copies, htab_t decl_copies, | |
665 | bool copy_name_p) | |
666 | { | |
667 | tree copy, var, var_copy; | |
668 | unsigned idx, uid, nuid; | |
669 | struct int_tree_map ielt, *nielt; | |
670 | struct name_to_copy_elt elt, *nelt; | |
671 | void **slot, **dslot; | |
672 | ||
673 | if (TREE_CODE (name) != SSA_NAME) | |
674 | return name; | |
675 | ||
676 | idx = SSA_NAME_VERSION (name); | |
677 | elt.version = idx; | |
678 | slot = htab_find_slot_with_hash (name_copies, &elt, idx, | |
679 | copy_name_p ? INSERT : NO_INSERT); | |
680 | if (slot && *slot) | |
681 | return ((struct name_to_copy_elt *) *slot)->new_name; | |
682 | ||
683 | var = SSA_NAME_VAR (name); | |
684 | uid = DECL_UID (var); | |
685 | ielt.uid = uid; | |
686 | dslot = htab_find_slot_with_hash (decl_copies, &ielt, uid, INSERT); | |
687 | if (!*dslot) | |
688 | { | |
689 | var_copy = create_tmp_var (TREE_TYPE (var), get_name (var)); | |
690 | add_referenced_var (var_copy); | |
691 | nielt = XNEW (struct int_tree_map); | |
692 | nielt->uid = uid; | |
693 | nielt->to = var_copy; | |
694 | *dslot = nielt; | |
695 | ||
696 | /* Ensure that when we meet this decl next time, we won't duplicate | |
a509ebb5 | 697 | it again. */ |
5f40b3cb ZD |
698 | nuid = DECL_UID (var_copy); |
699 | ielt.uid = nuid; | |
700 | dslot = htab_find_slot_with_hash (decl_copies, &ielt, nuid, INSERT); | |
701 | gcc_assert (!*dslot); | |
702 | nielt = XNEW (struct int_tree_map); | |
703 | nielt->uid = nuid; | |
704 | nielt->to = var_copy; | |
705 | *dslot = nielt; | |
706 | } | |
707 | else | |
708 | var_copy = ((struct int_tree_map *) *dslot)->to; | |
709 | ||
710 | if (copy_name_p) | |
711 | { | |
712 | copy = duplicate_ssa_name (name, NULL_TREE); | |
713 | nelt = XNEW (struct name_to_copy_elt); | |
714 | nelt->version = idx; | |
715 | nelt->new_name = copy; | |
716 | nelt->field = NULL_TREE; | |
717 | *slot = nelt; | |
718 | } | |
719 | else | |
720 | { | |
721 | gcc_assert (!slot); | |
722 | copy = name; | |
723 | } | |
724 | ||
725 | SSA_NAME_VAR (copy) = var_copy; | |
726 | return copy; | |
727 | } | |
728 | ||
729 | /* Finds the ssa names used in STMT that are defined outside of LOOP and | |
730 | replaces such ssa names with their duplicates. The duplicates are stored to | |
731 | NAME_COPIES. Base decls of all ssa names used in STMT | |
732 | (including those defined in LOOP) are replaced with the new temporary | |
733 | variables; the replacement decls are stored in DECL_COPIES. */ | |
734 | ||
735 | static void | |
736 | separate_decls_in_loop_stmt (struct loop *loop, tree stmt, | |
737 | htab_t name_copies, htab_t decl_copies) | |
738 | { | |
739 | use_operand_p use; | |
740 | def_operand_p def; | |
741 | ssa_op_iter oi; | |
742 | tree name, copy; | |
743 | bool copy_name_p; | |
744 | ||
745 | mark_virtual_ops_for_renaming (stmt); | |
746 | ||
747 | FOR_EACH_PHI_OR_STMT_DEF (def, stmt, oi, SSA_OP_DEF) | |
a509ebb5 RL |
748 | { |
749 | name = DEF_FROM_PTR (def); | |
750 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
751 | copy = separate_decls_in_loop_name (name, name_copies, decl_copies, | |
752 | false); | |
753 | gcc_assert (copy == name); | |
754 | } | |
5f40b3cb ZD |
755 | |
756 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
a509ebb5 RL |
757 | { |
758 | name = USE_FROM_PTR (use); | |
759 | if (TREE_CODE (name) != SSA_NAME) | |
760 | continue; | |
761 | ||
762 | copy_name_p = expr_invariant_in_loop_p (loop, name); | |
763 | copy = separate_decls_in_loop_name (name, name_copies, decl_copies, | |
764 | copy_name_p); | |
765 | SET_USE (use, copy); | |
766 | } | |
5f40b3cb ZD |
767 | } |
768 | ||
0eb7e7aa RL |
769 | /* Callback for htab_traverse. Adds a field corresponding to the reduction |
770 | specified in SLOT. The type is passed in DATA. */ | |
771 | ||
772 | static int | |
773 | add_field_for_reduction (void **slot, void *data) | |
a509ebb5 | 774 | { |
0eb7e7aa RL |
775 | |
776 | struct reduction_info *red = *slot; | |
777 | tree type = data; | |
778 | tree var = SSA_NAME_VAR (GIMPLE_STMT_OPERAND (red->reduc_stmt, 0)); | |
779 | tree field = build_decl (FIELD_DECL, DECL_NAME (var), TREE_TYPE (var)); | |
780 | ||
781 | insert_field_into_struct (type, field); | |
782 | ||
783 | red->field = field; | |
784 | ||
785 | return 1; | |
786 | } | |
a509ebb5 | 787 | |
5f40b3cb | 788 | /* Callback for htab_traverse. Adds a field corresponding to a ssa name |
0eb7e7aa | 789 | described in SLOT. The type is passed in DATA. */ |
5f40b3cb ZD |
790 | |
791 | static int | |
792 | add_field_for_name (void **slot, void *data) | |
793 | { | |
794 | struct name_to_copy_elt *elt = *slot; | |
0eb7e7aa | 795 | tree type = data; |
5f40b3cb ZD |
796 | tree name = ssa_name (elt->version); |
797 | tree var = SSA_NAME_VAR (name); | |
798 | tree field = build_decl (FIELD_DECL, DECL_NAME (var), TREE_TYPE (var)); | |
799 | ||
800 | insert_field_into_struct (type, field); | |
801 | elt->field = field; | |
a509ebb5 | 802 | |
5f40b3cb ZD |
803 | return 1; |
804 | } | |
805 | ||
a509ebb5 RL |
806 | /* Callback for htab_traverse. A local result is the intermediate result |
807 | computed by a single | |
808 | thread, or the intial value in case no iteration was executed. | |
809 | This function creates a phi node reflecting these values. | |
810 | The phi's result will be stored in NEW_PHI field of the | |
811 | reduction's data structure. */ | |
812 | ||
813 | static int | |
814 | create_phi_for_local_result (void **slot, void *data) | |
815 | { | |
816 | struct reduction_info *reduc = *slot; | |
817 | struct loop *loop = data; | |
818 | edge e; | |
819 | tree new_phi; | |
820 | basic_block store_bb; | |
821 | tree local_res; | |
822 | ||
823 | /* STORE_BB is the block where the phi | |
824 | should be stored. It is the destination of the loop exit. | |
825 | (Find the fallthru edge from OMP_CONTINUE). */ | |
826 | store_bb = FALLTHRU_EDGE (loop->latch)->dest; | |
827 | ||
828 | /* STORE_BB has two predecessors. One coming from the loop | |
829 | (the reduction's result is computed at the loop), | |
830 | and another coming from a block preceding the loop, | |
831 | when no iterations | |
832 | are executed (the initial value should be taken). */ | |
833 | if (EDGE_PRED (store_bb, 0) == FALLTHRU_EDGE (loop->latch)) | |
834 | e = EDGE_PRED (store_bb, 1); | |
835 | else | |
836 | e = EDGE_PRED (store_bb, 0); | |
0eb7e7aa | 837 | local_res = make_ssa_name (SSA_NAME_VAR (GIMPLE_STMT_OPERAND (reduc->reduc_stmt, 0)), NULL_TREE); |
a509ebb5 RL |
838 | new_phi = create_phi_node (local_res, store_bb); |
839 | SSA_NAME_DEF_STMT (local_res) = new_phi; | |
840 | add_phi_arg (new_phi, reduc->init, e); | |
841 | add_phi_arg (new_phi, GIMPLE_STMT_OPERAND (reduc->reduc_stmt, 0), | |
842 | FALLTHRU_EDGE (loop->latch)); | |
843 | reduc->new_phi = new_phi; | |
844 | ||
845 | return 1; | |
846 | } | |
5f40b3cb ZD |
847 | |
848 | struct clsn_data | |
849 | { | |
850 | tree store; | |
851 | tree load; | |
852 | ||
853 | basic_block store_bb; | |
854 | basic_block load_bb; | |
855 | }; | |
856 | ||
a509ebb5 RL |
857 | /* Callback for htab_traverse. Create an atomic instruction for the |
858 | reduction described in SLOT. | |
859 | DATA annotates the place in memory the atomic operation relates to, | |
860 | and the basic block it needs to be generated in. */ | |
861 | ||
862 | static int | |
863 | create_call_for_reduction_1 (void **slot, void *data) | |
864 | { | |
865 | struct reduction_info *reduc = *slot; | |
866 | struct clsn_data *clsn_data = data; | |
867 | block_stmt_iterator bsi; | |
868 | tree type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); | |
869 | tree struct_type = TREE_TYPE (TREE_TYPE (clsn_data->load)); | |
870 | tree load_struct; | |
871 | basic_block bb; | |
872 | basic_block new_bb; | |
873 | edge e; | |
874 | tree t, addr, addr_type, ref, x; | |
875 | tree tmp_load, load, name; | |
876 | ||
877 | load_struct = fold_build1 (INDIRECT_REF, struct_type, clsn_data->load); | |
878 | t = build3 (COMPONENT_REF, type, load_struct, reduc->field, NULL_TREE); | |
879 | addr_type = build_pointer_type (type); | |
880 | ||
881 | addr = build_addr (t, current_function_decl); | |
882 | ||
883 | /* Create phi node. */ | |
884 | bb = clsn_data->load_bb; | |
885 | ||
886 | e = split_block (bb, t); | |
887 | new_bb = e->dest; | |
888 | ||
889 | tmp_load = create_tmp_var (TREE_TYPE (TREE_TYPE (addr)), NULL); | |
890 | add_referenced_var (tmp_load); | |
891 | tmp_load = make_ssa_name (tmp_load, NULL); | |
892 | load = build2 (OMP_ATOMIC_LOAD, void_type_node, tmp_load, addr); | |
893 | SSA_NAME_DEF_STMT (tmp_load) = load; | |
894 | bsi = bsi_start (new_bb); | |
895 | bsi_insert_after (&bsi, load, BSI_NEW_STMT); | |
896 | ||
897 | e = split_block (new_bb, load); | |
898 | new_bb = e->dest; | |
899 | bsi = bsi_start (new_bb); | |
900 | ref = tmp_load; | |
901 | x = | |
902 | fold_build2 (reduc->reduction_code, | |
903 | TREE_TYPE (PHI_RESULT (reduc->new_phi)), ref, | |
904 | PHI_RESULT (reduc->new_phi)); | |
905 | ||
906 | name = | |
907 | force_gimple_operand_bsi (&bsi, x, true, NULL_TREE, true, | |
908 | BSI_CONTINUE_LINKING); | |
909 | ||
910 | x = build1 (OMP_ATOMIC_STORE, void_type_node, name); | |
911 | ||
912 | bsi_insert_after (&bsi, x, BSI_NEW_STMT); | |
913 | return 1; | |
914 | } | |
915 | ||
916 | /* Create the atomic operation at the join point of the threads. | |
917 | REDUCTION_LIST describes the reductions in the LOOP. | |
918 | LD_ST_DATA describes the shared data structure where | |
919 | shared data is stored in and loaded from. */ | |
920 | static void | |
921 | create_call_for_reduction (struct loop *loop, htab_t reduction_list, | |
922 | struct clsn_data *ld_st_data) | |
923 | { | |
924 | htab_traverse (reduction_list, create_phi_for_local_result, loop); | |
925 | /* Find the fallthru edge from OMP_CONTINUE. */ | |
926 | ld_st_data->load_bb = FALLTHRU_EDGE (loop->latch)->dest; | |
927 | htab_traverse (reduction_list, create_call_for_reduction_1, ld_st_data); | |
928 | } | |
929 | ||
930 | /* Callback for htab_traverse. Create a new variable that loads the | |
931 | final reduction value at the | |
932 | join point of all threads, adds the initial value the reduction | |
933 | variable had before the parallel computation started, and | |
934 | inserts it in the right place. */ | |
935 | ||
936 | static int | |
937 | create_loads_for_reductions (void **slot, void *data) | |
938 | { | |
939 | struct reduction_info *red = *slot; | |
940 | struct clsn_data *clsn_data = data; | |
941 | tree stmt; | |
942 | block_stmt_iterator bsi; | |
0eb7e7aa | 943 | tree type = TREE_TYPE (GIMPLE_STMT_OPERAND (red->reduc_stmt, 0)); |
a509ebb5 RL |
944 | tree struct_type = TREE_TYPE (TREE_TYPE (clsn_data->load)); |
945 | tree load_struct; | |
946 | tree bvar, name; | |
947 | tree x; | |
948 | ||
949 | bsi = bsi_after_labels (clsn_data->load_bb); | |
950 | load_struct = fold_build1 (INDIRECT_REF, struct_type, clsn_data->load); | |
951 | load_struct = build3 (COMPONENT_REF, type, load_struct, red->field, | |
952 | NULL_TREE); | |
953 | bvar = create_tmp_var (type, "reduction_final"); | |
954 | add_referenced_var (bvar); | |
955 | ||
956 | /* Apply operation between the new variable which is the result | |
957 | of computation all threads, and the initial value which is kept | |
958 | at reduction->inital_value. */ | |
959 | ||
960 | stmt = build_gimple_modify_stmt (bvar, load_struct); | |
961 | name = make_ssa_name (bvar, stmt); | |
962 | GIMPLE_STMT_OPERAND (stmt, 0) = name; | |
963 | SSA_NAME_DEF_STMT (name) = stmt; | |
a509ebb5 | 964 | bsi_insert_after (&bsi, stmt, BSI_NEW_STMT); |
a509ebb5 RL |
965 | x = |
966 | fold_build2 (red->reduction_code, TREE_TYPE (load_struct), | |
967 | name, red->initial_value); | |
968 | name = PHI_RESULT (red->keep_res); | |
969 | stmt = build_gimple_modify_stmt (name, x); | |
970 | GIMPLE_STMT_OPERAND (stmt, 0) = name; | |
971 | SSA_NAME_DEF_STMT (name) = stmt; | |
972 | ||
973 | bsi_insert_after (&bsi, stmt, BSI_NEW_STMT); | |
974 | ||
975 | remove_phi_node (red->keep_res, NULL_TREE, false); | |
976 | ||
977 | return 1; | |
978 | } | |
979 | ||
980 | /* Load the reduction result that was stored in LD_ST_DATA. | |
981 | REDUCTION_LIST describes the list of reductions that the | |
982 | loades should be generated for. */ | |
983 | static void | |
984 | create_final_loads_for_reduction (htab_t reduction_list, | |
985 | struct clsn_data *ld_st_data) | |
986 | { | |
987 | block_stmt_iterator bsi; | |
988 | tree t; | |
989 | ||
990 | bsi = bsi_after_labels (ld_st_data->load_bb); | |
991 | t = build_fold_addr_expr (ld_st_data->store); | |
992 | t = | |
993 | build_gimple_modify_stmt (ld_st_data->load, | |
994 | build_fold_addr_expr (ld_st_data->store)); | |
995 | ||
996 | bsi_insert_before (&bsi, t, BSI_NEW_STMT); | |
997 | SSA_NAME_DEF_STMT (ld_st_data->load) = t; | |
998 | GIMPLE_STMT_OPERAND (t, 0) = ld_st_data->load; | |
999 | ||
1000 | htab_traverse (reduction_list, create_loads_for_reductions, ld_st_data); | |
1001 | ||
1002 | } | |
1003 | ||
0eb7e7aa RL |
1004 | /* Callback for htab_traverse. Store the neutral value for the |
1005 | particular reduction's operation, e.g. 0 for PLUS_EXPR, | |
1006 | 1 for MULT_EXPR, etc. into the reduction field. | |
1007 | The reduction is specified in SLOT. The store information is | |
1008 | passed in DATA. */ | |
1009 | ||
1010 | static int | |
1011 | create_stores_for_reduction (void **slot, void *data) | |
1012 | { | |
1013 | struct reduction_info *red = *slot; | |
1014 | struct clsn_data *clsn_data = data; | |
1015 | tree stmt; | |
1016 | block_stmt_iterator bsi; | |
1017 | tree type = TREE_TYPE (GIMPLE_STMT_OPERAND (red->reduc_stmt, 0)); | |
1018 | ||
1019 | bsi = bsi_last (clsn_data->store_bb); | |
1020 | stmt = | |
1021 | build_gimple_modify_stmt (build3 | |
1022 | (COMPONENT_REF, type, clsn_data->store, | |
1023 | red->field, NULL_TREE), | |
1024 | red->init ); | |
1025 | mark_virtual_ops_for_renaming (stmt); | |
1026 | bsi_insert_after (&bsi, stmt, BSI_NEW_STMT); | |
1027 | ||
1028 | return 1; | |
1029 | } | |
1030 | ||
a509ebb5 RL |
1031 | /* Callback for htab_traverse. Creates loads to a field of LOAD in LOAD_BB and |
1032 | store to a field of STORE in STORE_BB for the ssa name and its duplicate | |
1033 | specified in SLOT. */ | |
1034 | ||
5f40b3cb ZD |
1035 | static int |
1036 | create_loads_and_stores_for_name (void **slot, void *data) | |
1037 | { | |
1038 | struct name_to_copy_elt *elt = *slot; | |
1039 | struct clsn_data *clsn_data = data; | |
1040 | tree stmt; | |
1041 | block_stmt_iterator bsi; | |
1042 | tree type = TREE_TYPE (elt->new_name); | |
1043 | tree struct_type = TREE_TYPE (TREE_TYPE (clsn_data->load)); | |
1044 | tree load_struct; | |
1045 | ||
1046 | bsi = bsi_last (clsn_data->store_bb); | |
a509ebb5 RL |
1047 | stmt = |
1048 | build_gimple_modify_stmt (build3 | |
1049 | (COMPONENT_REF, type, clsn_data->store, | |
1050 | elt->field, NULL_TREE), | |
1051 | ssa_name (elt->version)); | |
5f40b3cb ZD |
1052 | mark_virtual_ops_for_renaming (stmt); |
1053 | bsi_insert_after (&bsi, stmt, BSI_NEW_STMT); | |
1054 | ||
1055 | bsi = bsi_last (clsn_data->load_bb); | |
1056 | load_struct = fold_build1 (INDIRECT_REF, struct_type, clsn_data->load); | |
a509ebb5 RL |
1057 | stmt = build_gimple_modify_stmt (elt->new_name, |
1058 | build3 (COMPONENT_REF, type, load_struct, | |
1059 | elt->field, NULL_TREE)); | |
5f40b3cb ZD |
1060 | SSA_NAME_DEF_STMT (elt->new_name) = stmt; |
1061 | bsi_insert_after (&bsi, stmt, BSI_NEW_STMT); | |
1062 | ||
1063 | return 1; | |
1064 | } | |
1065 | ||
1066 | /* Moves all the variables used in LOOP and defined outside of it (including | |
1067 | the initial values of loop phi nodes, and *PER_THREAD if it is a ssa | |
1068 | name) to a structure created for this purpose. The code | |
1069 | ||
1070 | while (1) | |
1071 | { | |
1072 | use (a); | |
1073 | use (b); | |
1074 | } | |
1075 | ||
1076 | is transformed this way: | |
1077 | ||
1078 | bb0: | |
1079 | old.a = a; | |
1080 | old.b = b; | |
1081 | ||
1082 | bb1: | |
1083 | a' = new->a; | |
1084 | b' = new->b; | |
1085 | while (1) | |
1086 | { | |
1087 | use (a'); | |
1088 | use (b'); | |
1089 | } | |
1090 | ||
1091 | `old' is stored to *ARG_STRUCT and `new' is stored to NEW_ARG_STRUCT. The | |
1092 | pointer `new' is intentionally not initialized (the loop will be split to a | |
1093 | separate function later, and `new' will be initialized from its arguments). | |
a509ebb5 RL |
1094 | LD_ST_DATA holds information about the shared data structure used to pass |
1095 | information among the threads. It is initialized here, and | |
1096 | gen_parallel_loop will pass it to create_call_for_reduction that | |
1097 | needs this information. REDUCTION_LIST describes the reductions | |
1098 | in LOOP. */ | |
5f40b3cb ZD |
1099 | |
1100 | static void | |
a509ebb5 RL |
1101 | separate_decls_in_loop (struct loop *loop, htab_t reduction_list, |
1102 | tree * arg_struct, tree * new_arg_struct, | |
1103 | struct clsn_data *ld_st_data) | |
1104 | ||
5f40b3cb ZD |
1105 | { |
1106 | basic_block bb1 = split_edge (loop_preheader_edge (loop)); | |
1107 | basic_block bb0 = single_pred (bb1); | |
1108 | htab_t name_copies = htab_create (10, name_to_copy_elt_hash, | |
1109 | name_to_copy_elt_eq, free); | |
1110 | htab_t decl_copies = htab_create (10, int_tree_map_hash, int_tree_map_eq, | |
1111 | free); | |
1112 | basic_block bb, *body = get_loop_body (loop); | |
1113 | unsigned i; | |
1114 | tree phi, type, type_name, nvar; | |
1115 | block_stmt_iterator bsi; | |
1116 | struct clsn_data clsn_data; | |
1117 | ||
1118 | /* Find and rename the ssa names defined outside of loop. */ | |
1119 | for (i = 0; i < loop->num_nodes; i++) | |
1120 | { | |
1121 | bb = body[i]; | |
1122 | ||
1123 | for (phi = phi_nodes (bb); phi; phi = PHI_CHAIN (phi)) | |
1124 | separate_decls_in_loop_stmt (loop, phi, name_copies, decl_copies); | |
1125 | ||
1126 | for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi)) | |
1127 | separate_decls_in_loop_stmt (loop, bsi_stmt (bsi), name_copies, | |
1128 | decl_copies); | |
1129 | } | |
1130 | free (body); | |
1131 | ||
1132 | if (htab_elements (name_copies) == 0) | |
1133 | { | |
1134 | /* It may happen that there is nothing to copy (if there are only | |
a509ebb5 | 1135 | loop carried and external variables in the loop). */ |
5f40b3cb ZD |
1136 | *arg_struct = NULL; |
1137 | *new_arg_struct = NULL; | |
1138 | } | |
1139 | else | |
1140 | { | |
1141 | /* Create the type for the structure to store the ssa names to. */ | |
1142 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
1143 | type_name = build_decl (TYPE_DECL, create_tmp_var_name (".paral_data"), | |
1144 | type); | |
1145 | TYPE_NAME (type) = type_name; | |
1146 | ||
0eb7e7aa RL |
1147 | htab_traverse (name_copies, add_field_for_name, type); |
1148 | if (htab_elements (reduction_list) > 0) | |
1149 | { | |
1150 | /* Create the fields for reductions. */ | |
1151 | htab_traverse (reduction_list, add_field_for_reduction, | |
1152 | type); | |
1153 | } | |
5f40b3cb | 1154 | layout_type (type); |
0eb7e7aa | 1155 | |
5f40b3cb ZD |
1156 | /* Create the loads and stores. */ |
1157 | *arg_struct = create_tmp_var (type, ".paral_data_store"); | |
1158 | add_referenced_var (*arg_struct); | |
1159 | nvar = create_tmp_var (build_pointer_type (type), ".paral_data_load"); | |
1160 | add_referenced_var (nvar); | |
1161 | *new_arg_struct = make_ssa_name (nvar, NULL_TREE); | |
1162 | ||
a509ebb5 RL |
1163 | ld_st_data->store = *arg_struct; |
1164 | ld_st_data->load = *new_arg_struct; | |
1165 | ld_st_data->store_bb = bb0; | |
1166 | ld_st_data->load_bb = bb1; | |
0eb7e7aa | 1167 | |
5f40b3cb | 1168 | htab_traverse (name_copies, create_loads_and_stores_for_name, |
a509ebb5 RL |
1169 | ld_st_data); |
1170 | ||
1171 | /* Load the calculation from memory into a new | |
1172 | reduction variable (after the join of the threads). */ | |
1173 | if (htab_elements (reduction_list) > 0) | |
1174 | { | |
0eb7e7aa RL |
1175 | htab_traverse (reduction_list, create_stores_for_reduction, |
1176 | ld_st_data); | |
a509ebb5 RL |
1177 | clsn_data.load = make_ssa_name (nvar, NULL_TREE); |
1178 | clsn_data.load_bb = single_dom_exit (loop)->dest; | |
1179 | clsn_data.store = ld_st_data->store; | |
1180 | create_final_loads_for_reduction (reduction_list, &clsn_data); | |
1181 | } | |
5f40b3cb ZD |
1182 | } |
1183 | ||
1184 | htab_delete (decl_copies); | |
1185 | htab_delete (name_copies); | |
1186 | } | |
1187 | ||
1188 | /* Bitmap containing uids of functions created by parallelization. We cannot | |
1189 | allocate it from the default obstack, as it must live across compilation | |
1190 | of several functions; we make it gc allocated instead. */ | |
1191 | ||
1192 | static GTY(()) bitmap parallelized_functions; | |
1193 | ||
1194 | /* Returns true if FN was created by create_loop_fn. */ | |
1195 | ||
1196 | static bool | |
1197 | parallelized_function_p (tree fn) | |
1198 | { | |
1199 | if (!parallelized_functions || !DECL_ARTIFICIAL (fn)) | |
1200 | return false; | |
1201 | ||
1202 | return bitmap_bit_p (parallelized_functions, DECL_UID (fn)); | |
1203 | } | |
1204 | ||
1205 | /* Creates and returns an empty function that will receive the body of | |
1206 | a parallelized loop. */ | |
1207 | ||
1208 | static tree | |
1209 | create_loop_fn (void) | |
1210 | { | |
1211 | char buf[100]; | |
1212 | char *tname; | |
1213 | tree decl, type, name, t; | |
1214 | struct function *act_cfun = cfun; | |
1215 | static unsigned loopfn_num; | |
1216 | ||
1217 | snprintf (buf, 100, "%s.$loopfn", current_function_name ()); | |
1218 | ASM_FORMAT_PRIVATE_NAME (tname, buf, loopfn_num++); | |
1219 | clean_symbol_name (tname); | |
1220 | name = get_identifier (tname); | |
1221 | type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); | |
1222 | ||
1223 | decl = build_decl (FUNCTION_DECL, name, type); | |
1224 | if (!parallelized_functions) | |
1225 | parallelized_functions = BITMAP_GGC_ALLOC (); | |
1226 | bitmap_set_bit (parallelized_functions, DECL_UID (decl)); | |
1227 | ||
1228 | TREE_STATIC (decl) = 1; | |
1229 | TREE_USED (decl) = 1; | |
1230 | DECL_ARTIFICIAL (decl) = 1; | |
1231 | DECL_IGNORED_P (decl) = 0; | |
1232 | TREE_PUBLIC (decl) = 0; | |
1233 | DECL_UNINLINABLE (decl) = 1; | |
1234 | DECL_EXTERNAL (decl) = 0; | |
1235 | DECL_CONTEXT (decl) = NULL_TREE; | |
1236 | DECL_INITIAL (decl) = make_node (BLOCK); | |
1237 | ||
1238 | t = build_decl (RESULT_DECL, NULL_TREE, void_type_node); | |
1239 | DECL_ARTIFICIAL (t) = 1; | |
1240 | DECL_IGNORED_P (t) = 1; | |
1241 | DECL_RESULT (decl) = t; | |
1242 | ||
1243 | t = build_decl (PARM_DECL, get_identifier (".paral_data_param"), | |
1244 | ptr_type_node); | |
1245 | DECL_ARTIFICIAL (t) = 1; | |
1246 | DECL_ARG_TYPE (t) = ptr_type_node; | |
1247 | DECL_CONTEXT (t) = decl; | |
1248 | TREE_USED (t) = 1; | |
1249 | DECL_ARGUMENTS (decl) = t; | |
1250 | ||
182e0d71 | 1251 | allocate_struct_function (decl, false); |
5f40b3cb ZD |
1252 | |
1253 | /* The call to allocate_struct_function clobbers CFUN, so we need to restore | |
1254 | it. */ | |
5576d6f2 | 1255 | set_cfun (act_cfun); |
5f40b3cb ZD |
1256 | |
1257 | return decl; | |
1258 | } | |
1259 | ||
1260 | /* Bases all the induction variables in LOOP on a single induction variable | |
1261 | (unsigned with base 0 and step 1), whose final value is compared with | |
a509ebb5 RL |
1262 | NIT. The induction variable is incremented in the loop latch. |
1263 | REDUCTION_LIST describes the reductions in LOOP. */ | |
5f40b3cb ZD |
1264 | |
1265 | static void | |
a509ebb5 | 1266 | canonicalize_loop_ivs (struct loop *loop, htab_t reduction_list, tree nit) |
5f40b3cb ZD |
1267 | { |
1268 | unsigned precision = TYPE_PRECISION (TREE_TYPE (nit)); | |
1269 | tree phi, prev, res, type, var_before, val, atype, t, next; | |
1270 | block_stmt_iterator bsi; | |
1271 | bool ok; | |
1272 | affine_iv iv; | |
1273 | edge exit = single_dom_exit (loop); | |
a509ebb5 | 1274 | struct reduction_info *red; |
5f40b3cb ZD |
1275 | |
1276 | for (phi = phi_nodes (loop->header); phi; phi = PHI_CHAIN (phi)) | |
1277 | { | |
1278 | res = PHI_RESULT (phi); | |
1279 | ||
a509ebb5 | 1280 | if (is_gimple_reg (res) && TYPE_PRECISION (TREE_TYPE (res)) > precision) |
5f40b3cb ZD |
1281 | precision = TYPE_PRECISION (TREE_TYPE (res)); |
1282 | } | |
1283 | ||
1284 | type = lang_hooks.types.type_for_size (precision, 1); | |
1285 | ||
1286 | bsi = bsi_last (loop->latch); | |
1287 | create_iv (build_int_cst_type (type, 0), build_int_cst (type, 1), NULL_TREE, | |
1288 | loop, &bsi, true, &var_before, NULL); | |
1289 | ||
1290 | bsi = bsi_after_labels (loop->header); | |
1291 | prev = NULL; | |
1292 | for (phi = phi_nodes (loop->header); phi; phi = next) | |
1293 | { | |
1294 | next = PHI_CHAIN (phi); | |
1295 | res = PHI_RESULT (phi); | |
1296 | ||
a509ebb5 | 1297 | if (!is_gimple_reg (res) || res == var_before) |
5f40b3cb ZD |
1298 | { |
1299 | prev = phi; | |
1300 | continue; | |
1301 | } | |
5f40b3cb | 1302 | |
a509ebb5 RL |
1303 | ok = simple_iv (loop, phi, res, &iv, true); |
1304 | red = reduction_phi (reduction_list, phi); | |
1305 | /* We preserve the reduction phi nodes. */ | |
1306 | if (!ok && red) | |
1307 | { | |
1308 | prev = phi; | |
1309 | continue; | |
1310 | } | |
1311 | else | |
1312 | gcc_assert (ok); | |
5f40b3cb ZD |
1313 | remove_phi_node (phi, prev, false); |
1314 | ||
1315 | atype = TREE_TYPE (res); | |
1316 | val = fold_build2 (PLUS_EXPR, atype, | |
1317 | unshare_expr (iv.base), | |
1318 | fold_build2 (MULT_EXPR, atype, | |
1319 | unshare_expr (iv.step), | |
1320 | fold_convert (atype, var_before))); | |
1321 | val = force_gimple_operand_bsi (&bsi, val, false, NULL_TREE, true, | |
1322 | BSI_SAME_STMT); | |
1323 | t = build_gimple_modify_stmt (res, val); | |
1324 | bsi_insert_before (&bsi, t, BSI_SAME_STMT); | |
1325 | SSA_NAME_DEF_STMT (res) = t; | |
1326 | } | |
1327 | ||
1328 | t = last_stmt (exit->src); | |
1329 | /* Make the loop exit if the control condition is not satisfied. */ | |
1330 | if (exit->flags & EDGE_TRUE_VALUE) | |
1331 | { | |
1332 | edge te, fe; | |
1333 | ||
1334 | extract_true_false_edges_from_block (exit->src, &te, &fe); | |
1335 | te->flags = EDGE_FALSE_VALUE; | |
1336 | fe->flags = EDGE_TRUE_VALUE; | |
1337 | } | |
1338 | COND_EXPR_COND (t) = build2 (LT_EXPR, boolean_type_node, var_before, nit); | |
1339 | } | |
1340 | ||
1341 | /* Moves the exit condition of LOOP to the beginning of its header, and | |
1342 | duplicates the part of the last iteration that gets disabled to the | |
1343 | exit of the loop. NIT is the number of iterations of the loop | |
1344 | (used to initialize the variables in the duplicated part). | |
1345 | ||
1346 | TODO: the common case is that latch of the loop is empty and immediatelly | |
1347 | follows the loop exit. In this case, it would be better not to copy the | |
1348 | body of the loop, but only move the entry of the loop directly before the | |
1349 | exit check and increase the number of iterations of the loop by one. | |
a509ebb5 RL |
1350 | This may need some additional preconditioning in case NIT = ~0. |
1351 | REDUCTION_LIST describes the reductions in LOOP. */ | |
5f40b3cb ZD |
1352 | |
1353 | static void | |
a509ebb5 | 1354 | transform_to_exit_first_loop (struct loop *loop, htab_t reduction_list, tree nit) |
5f40b3cb ZD |
1355 | { |
1356 | basic_block *bbs, *nbbs, ex_bb, orig_header; | |
1357 | unsigned n; | |
1358 | bool ok; | |
1359 | edge exit = single_dom_exit (loop), hpred; | |
1360 | tree phi, nphi, cond, control, control_name, res, t, cond_stmt; | |
1361 | block_stmt_iterator bsi; | |
1362 | ||
1363 | split_block_after_labels (loop->header); | |
1364 | orig_header = single_succ (loop->header); | |
1365 | hpred = single_succ_edge (loop->header); | |
1366 | ||
1367 | cond_stmt = last_stmt (exit->src); | |
1368 | cond = COND_EXPR_COND (cond_stmt); | |
1369 | control = TREE_OPERAND (cond, 0); | |
1370 | gcc_assert (TREE_OPERAND (cond, 1) == nit); | |
1371 | ||
1372 | /* Make sure that we have phi nodes on exit for all loop header phis | |
1373 | (create_parallel_loop requires that). */ | |
1374 | for (phi = phi_nodes (loop->header); phi; phi = PHI_CHAIN (phi)) | |
1375 | { | |
1376 | res = PHI_RESULT (phi); | |
1377 | t = make_ssa_name (SSA_NAME_VAR (res), phi); | |
1378 | SET_PHI_RESULT (phi, t); | |
1379 | ||
1380 | nphi = create_phi_node (res, orig_header); | |
1381 | SSA_NAME_DEF_STMT (res) = nphi; | |
1382 | add_phi_arg (nphi, t, hpred); | |
1383 | ||
1384 | if (res == control) | |
1385 | { | |
1386 | TREE_OPERAND (cond, 0) = t; | |
1387 | update_stmt (cond_stmt); | |
1388 | control = t; | |
1389 | } | |
1390 | } | |
1391 | ||
1392 | bbs = get_loop_body_in_dom_order (loop); | |
1393 | for (n = 0; bbs[n] != exit->src; n++) | |
1394 | continue; | |
1395 | nbbs = XNEWVEC (basic_block, n); | |
1396 | ok = tree_duplicate_sese_tail (single_succ_edge (loop->header), exit, | |
1397 | bbs + 1, n, nbbs); | |
1398 | gcc_assert (ok); | |
1399 | free (bbs); | |
1400 | ex_bb = nbbs[0]; | |
1401 | free (nbbs); | |
1402 | ||
a509ebb5 RL |
1403 | /* Other than reductions, the only gimple reg that should be copied |
1404 | out of the loop is the control variable. */ | |
1405 | ||
5f40b3cb ZD |
1406 | control_name = NULL_TREE; |
1407 | for (phi = phi_nodes (ex_bb); phi; phi = PHI_CHAIN (phi)) | |
1408 | { | |
1409 | res = PHI_RESULT (phi); | |
1410 | if (!is_gimple_reg (res)) | |
1411 | continue; | |
1412 | ||
a509ebb5 RL |
1413 | /* Check if it is a part of reduction. If it is, |
1414 | keep the phi at the reduction's keep_res field. The | |
1415 | PHI_RESULT of this phi is the resulting value of the reduction | |
1416 | variable when exiting the loop. */ | |
1417 | ||
1418 | exit = single_dom_exit (loop); | |
1419 | ||
1420 | if (htab_elements (reduction_list) > 0) | |
1421 | { | |
1422 | struct reduction_info *red; | |
1423 | ||
1424 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
1425 | ||
1426 | red = reduction_phi (reduction_list, SSA_NAME_DEF_STMT (val)); | |
1427 | if (red) | |
1428 | red->keep_res = phi; | |
1429 | } | |
1430 | else | |
1431 | gcc_assert (control_name == NULL_TREE | |
1432 | && SSA_NAME_VAR (res) == SSA_NAME_VAR (control)); | |
5f40b3cb ZD |
1433 | control_name = res; |
1434 | } | |
1435 | gcc_assert (control_name != NULL_TREE); | |
1436 | phi = SSA_NAME_DEF_STMT (control_name); | |
1437 | remove_phi_node (phi, NULL_TREE, false); | |
1438 | ||
1439 | /* Initialize the control variable to NIT. */ | |
1440 | bsi = bsi_after_labels (ex_bb); | |
1441 | t = build_gimple_modify_stmt (control_name, nit); | |
1442 | bsi_insert_before (&bsi, t, BSI_NEW_STMT); | |
1443 | SSA_NAME_DEF_STMT (control_name) = t; | |
1444 | } | |
1445 | ||
1446 | /* Create the parallel constructs for LOOP as described in gen_parallel_loop. | |
1447 | LOOP_FN and DATA are the arguments of OMP_PARALLEL. | |
1448 | NEW_DATA is the variable that should be initialized from the argument | |
1449 | of LOOP_FN. N_THREADS is the requested number of threads. Returns the | |
1450 | basic block containing OMP_PARALLEL tree. */ | |
1451 | ||
1452 | static basic_block | |
1453 | create_parallel_loop (struct loop *loop, tree loop_fn, tree data, | |
1454 | tree new_data, unsigned n_threads) | |
1455 | { | |
1456 | block_stmt_iterator bsi; | |
1457 | basic_block bb, paral_bb, for_bb, ex_bb; | |
1458 | tree t, param, res, for_stmt; | |
1459 | tree cvar, cvar_init, initvar, cvar_next, cvar_base, cond, phi, type; | |
1460 | edge exit, nexit, guard, end, e; | |
1461 | ||
1462 | /* Prepare the OMP_PARALLEL statement. */ | |
1463 | bb = loop_preheader_edge (loop)->src; | |
1464 | paral_bb = single_pred (bb); | |
1465 | bsi = bsi_last (paral_bb); | |
1466 | ||
1467 | t = build_omp_clause (OMP_CLAUSE_NUM_THREADS); | |
1468 | OMP_CLAUSE_NUM_THREADS_EXPR (t) | |
a509ebb5 RL |
1469 | = build_int_cst (integer_type_node, n_threads); |
1470 | t = build4 (OMP_PARALLEL, void_type_node, NULL_TREE, t, loop_fn, data); | |
5f40b3cb ZD |
1471 | |
1472 | bsi_insert_after (&bsi, t, BSI_NEW_STMT); | |
1473 | ||
1474 | /* Initialize NEW_DATA. */ | |
1475 | if (data) | |
1476 | { | |
1477 | bsi = bsi_after_labels (bb); | |
1478 | ||
1479 | param = make_ssa_name (DECL_ARGUMENTS (loop_fn), NULL_TREE); | |
1480 | t = build_gimple_modify_stmt (param, build_fold_addr_expr (data)); | |
1481 | bsi_insert_before (&bsi, t, BSI_SAME_STMT); | |
1482 | SSA_NAME_DEF_STMT (param) = t; | |
1483 | ||
1484 | t = build_gimple_modify_stmt (new_data, | |
a509ebb5 RL |
1485 | fold_convert (TREE_TYPE (new_data), |
1486 | param)); | |
5f40b3cb ZD |
1487 | bsi_insert_before (&bsi, t, BSI_SAME_STMT); |
1488 | SSA_NAME_DEF_STMT (new_data) = t; | |
1489 | } | |
1490 | ||
1491 | /* Emit OMP_RETURN for OMP_PARALLEL. */ | |
1492 | bb = split_loop_exit_edge (single_dom_exit (loop)); | |
1493 | bsi = bsi_last (bb); | |
1494 | bsi_insert_after (&bsi, make_node (OMP_RETURN), BSI_NEW_STMT); | |
1495 | ||
1496 | /* Extract data for OMP_FOR. */ | |
1497 | gcc_assert (loop->header == single_dom_exit (loop)->src); | |
1498 | cond = COND_EXPR_COND (last_stmt (loop->header)); | |
1499 | ||
1500 | cvar = TREE_OPERAND (cond, 0); | |
1501 | cvar_base = SSA_NAME_VAR (cvar); | |
1502 | phi = SSA_NAME_DEF_STMT (cvar); | |
1503 | cvar_init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
1504 | initvar = make_ssa_name (cvar_base, NULL_TREE); | |
1505 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, loop_preheader_edge (loop)), | |
1506 | initvar); | |
1507 | cvar_next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
1508 | ||
1509 | bsi = bsi_last (loop->latch); | |
1510 | gcc_assert (bsi_stmt (bsi) == SSA_NAME_DEF_STMT (cvar_next)); | |
1511 | bsi_remove (&bsi, true); | |
1512 | ||
1513 | /* Prepare cfg. */ | |
1514 | for_bb = split_edge (loop_preheader_edge (loop)); | |
1515 | ex_bb = split_loop_exit_edge (single_dom_exit (loop)); | |
1516 | extract_true_false_edges_from_block (loop->header, &nexit, &exit); | |
1517 | gcc_assert (exit == single_dom_exit (loop)); | |
1518 | ||
1519 | guard = make_edge (for_bb, ex_bb, 0); | |
1520 | single_succ_edge (loop->latch)->flags = 0; | |
1521 | end = make_edge (loop->latch, ex_bb, EDGE_FALLTHRU); | |
1522 | for (phi = phi_nodes (ex_bb); phi; phi = PHI_CHAIN (phi)) | |
1523 | { | |
1524 | res = PHI_RESULT (phi); | |
1525 | gcc_assert (!is_gimple_reg (phi)); | |
1526 | t = SSA_NAME_DEF_STMT (PHI_ARG_DEF_FROM_EDGE (phi, exit)); | |
1527 | add_phi_arg (phi, PHI_ARG_DEF_FROM_EDGE (t, loop_preheader_edge (loop)), | |
1528 | guard); | |
1529 | add_phi_arg (phi, PHI_ARG_DEF_FROM_EDGE (t, loop_latch_edge (loop)), | |
1530 | end); | |
1531 | } | |
1532 | e = redirect_edge_and_branch (exit, nexit->dest); | |
1533 | PENDING_STMT (e) = NULL; | |
1534 | ||
1535 | /* Emit OMP_FOR. */ | |
1536 | TREE_OPERAND (cond, 0) = cvar_base; | |
1537 | type = TREE_TYPE (cvar); | |
1538 | t = build_omp_clause (OMP_CLAUSE_SCHEDULE); | |
1539 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_STATIC; | |
1540 | ||
1541 | for_stmt = make_node (OMP_FOR); | |
1542 | TREE_TYPE (for_stmt) = void_type_node; | |
1543 | OMP_FOR_CLAUSES (for_stmt) = t; | |
1544 | OMP_FOR_INIT (for_stmt) = build_gimple_modify_stmt (initvar, cvar_init); | |
1545 | OMP_FOR_COND (for_stmt) = cond; | |
a509ebb5 RL |
1546 | OMP_FOR_INCR (for_stmt) = build_gimple_modify_stmt (cvar_base, |
1547 | build2 (PLUS_EXPR, type, | |
1548 | cvar_base, | |
1549 | build_int_cst | |
1550 | (type, 1))); | |
5f40b3cb ZD |
1551 | OMP_FOR_BODY (for_stmt) = NULL_TREE; |
1552 | OMP_FOR_PRE_BODY (for_stmt) = NULL_TREE; | |
1553 | ||
1554 | bsi = bsi_last (for_bb); | |
1555 | bsi_insert_after (&bsi, for_stmt, BSI_NEW_STMT); | |
1556 | SSA_NAME_DEF_STMT (initvar) = for_stmt; | |
1557 | ||
1558 | /* Emit OMP_CONTINUE. */ | |
1559 | bsi = bsi_last (loop->latch); | |
1560 | t = build2 (OMP_CONTINUE, void_type_node, cvar_next, cvar); | |
1561 | bsi_insert_after (&bsi, t, BSI_NEW_STMT); | |
1562 | SSA_NAME_DEF_STMT (cvar_next) = t; | |
1563 | ||
1564 | /* Emit OMP_RETURN for OMP_FOR. */ | |
1565 | bsi = bsi_last (ex_bb); | |
1566 | bsi_insert_after (&bsi, make_node (OMP_RETURN), BSI_NEW_STMT); | |
1567 | ||
1568 | return paral_bb; | |
1569 | } | |
1570 | ||
1571 | /* Generates code to execute the iterations of LOOP in N_THREADS threads in | |
a509ebb5 RL |
1572 | parallel. NITER describes number of iterations of LOOP. |
1573 | REDUCTION_LIST describes the reductions existant in the LOOP. */ | |
5f40b3cb ZD |
1574 | |
1575 | static void | |
a509ebb5 RL |
1576 | gen_parallel_loop (struct loop *loop, htab_t reduction_list, |
1577 | unsigned n_threads, struct tree_niter_desc *niter) | |
5f40b3cb ZD |
1578 | { |
1579 | struct loop *nloop; | |
1580 | tree many_iterations_cond, type, nit; | |
1581 | tree stmts, arg_struct, new_arg_struct; | |
1582 | basic_block parallel_head; | |
a509ebb5 | 1583 | struct clsn_data clsn_data; |
5f40b3cb ZD |
1584 | unsigned prob; |
1585 | ||
1586 | /* From | |
1587 | ||
1588 | --------------------------------------------------------------------- | |
1589 | loop | |
1590 | { | |
1591 | IV = phi (INIT, IV + STEP) | |
1592 | BODY1; | |
1593 | if (COND) | |
1594 | break; | |
1595 | BODY2; | |
1596 | } | |
1597 | --------------------------------------------------------------------- | |
1598 | ||
1599 | with # of iterations NITER (possibly with MAY_BE_ZERO assumption), | |
1600 | we generate the following code: | |
1601 | ||
1602 | --------------------------------------------------------------------- | |
1603 | ||
1604 | if (MAY_BE_ZERO | |
a509ebb5 RL |
1605 | || NITER < MIN_PER_THREAD * N_THREADS) |
1606 | goto original; | |
5f40b3cb ZD |
1607 | |
1608 | BODY1; | |
1609 | store all local loop-invariant variables used in body of the loop to DATA. | |
1610 | OMP_PARALLEL (OMP_CLAUSE_NUM_THREADS (N_THREADS), LOOPFN, DATA); | |
1611 | load the variables from DATA. | |
1612 | OMP_FOR (IV = INIT; COND; IV += STEP) (OMP_CLAUSE_SCHEDULE (static)) | |
1613 | BODY2; | |
1614 | BODY1; | |
1615 | OMP_CONTINUE; | |
a509ebb5 RL |
1616 | OMP_RETURN -- OMP_FOR |
1617 | OMP_RETURN -- OMP_PARALLEL | |
5f40b3cb ZD |
1618 | goto end; |
1619 | ||
1620 | original: | |
1621 | loop | |
1622 | { | |
1623 | IV = phi (INIT, IV + STEP) | |
1624 | BODY1; | |
1625 | if (COND) | |
1626 | break; | |
1627 | BODY2; | |
1628 | } | |
1629 | ||
1630 | end: | |
1631 | ||
1632 | */ | |
1633 | ||
1634 | /* Create two versions of the loop -- in the old one, we know that the | |
1635 | number of iterations is large enough, and we will transform it into the | |
1636 | loop that will be split to loop_fn, the new one will be used for the | |
1637 | remaining iterations. */ | |
a509ebb5 | 1638 | |
5f40b3cb ZD |
1639 | type = TREE_TYPE (niter->niter); |
1640 | nit = force_gimple_operand (unshare_expr (niter->niter), &stmts, true, | |
1641 | NULL_TREE); | |
1642 | if (stmts) | |
1643 | bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
1644 | ||
1645 | many_iterations_cond = | |
a509ebb5 RL |
1646 | fold_build2 (GE_EXPR, boolean_type_node, |
1647 | nit, build_int_cst (type, MIN_PER_THREAD * n_threads)); | |
5f40b3cb | 1648 | many_iterations_cond |
a509ebb5 RL |
1649 | = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1650 | invert_truthvalue (unshare_expr (niter->may_be_zero)), | |
1651 | many_iterations_cond); | |
5f40b3cb | 1652 | many_iterations_cond |
a509ebb5 | 1653 | = force_gimple_operand (many_iterations_cond, &stmts, false, NULL_TREE); |
5f40b3cb ZD |
1654 | if (stmts) |
1655 | bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
1656 | if (!is_gimple_condexpr (many_iterations_cond)) | |
1657 | { | |
1658 | many_iterations_cond | |
a509ebb5 RL |
1659 | = force_gimple_operand (many_iterations_cond, &stmts, |
1660 | true, NULL_TREE); | |
5f40b3cb ZD |
1661 | if (stmts) |
1662 | bsi_insert_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
1663 | } | |
1664 | ||
1665 | initialize_original_copy_tables (); | |
1666 | ||
1667 | /* We assume that the loop usually iterates a lot. */ | |
1668 | prob = 4 * REG_BR_PROB_BASE / 5; | |
1669 | nloop = loop_version (loop, many_iterations_cond, NULL, | |
1670 | prob, prob, REG_BR_PROB_BASE - prob, true); | |
1671 | update_ssa (TODO_update_ssa); | |
1672 | free_original_copy_tables (); | |
1673 | ||
1674 | /* Base all the induction variables in LOOP on a single control one. */ | |
a509ebb5 | 1675 | canonicalize_loop_ivs (loop, reduction_list, nit); |
5f40b3cb ZD |
1676 | |
1677 | /* Ensure that the exit condition is the first statement in the loop. */ | |
a509ebb5 RL |
1678 | transform_to_exit_first_loop (loop, reduction_list, nit); |
1679 | ||
1680 | ||
1681 | /* Generate intializations for reductions. */ | |
1682 | ||
1683 | if (htab_elements (reduction_list) > 0) | |
1684 | htab_traverse (reduction_list, initialize_reductions, loop); | |
5f40b3cb ZD |
1685 | |
1686 | /* Eliminate the references to local variables from the loop. */ | |
1687 | eliminate_local_variables (loop); | |
1688 | ||
1689 | /* In the old loop, move all variables non-local to the loop to a structure | |
1690 | and back, and create separate decls for the variables used in loop. */ | |
a509ebb5 | 1691 | separate_decls_in_loop (loop, reduction_list, &arg_struct, &new_arg_struct, &clsn_data); |
5f40b3cb ZD |
1692 | |
1693 | /* Create the parallel constructs. */ | |
1694 | parallel_head = create_parallel_loop (loop, create_loop_fn (), arg_struct, | |
1695 | new_arg_struct, n_threads); | |
a509ebb5 RL |
1696 | if (htab_elements (reduction_list) > 0) |
1697 | create_call_for_reduction (loop, reduction_list, &clsn_data); | |
5f40b3cb ZD |
1698 | |
1699 | scev_reset (); | |
1700 | ||
1701 | /* Cancel the loop (it is simpler to do it here rather than to teach the | |
1702 | expander to do it). */ | |
1703 | cancel_loop_tree (loop); | |
1704 | ||
1705 | /* Expand the parallel constructs. We do it directly here instead of running | |
1706 | a separate expand_omp pass, since it is more efficient, and less likely to | |
1707 | cause troubles with further analyses not being able to deal with the | |
1708 | OMP trees. */ | |
a509ebb5 | 1709 | |
5f40b3cb ZD |
1710 | omp_expand_local (parallel_head); |
1711 | } | |
1712 | ||
1713 | /* Detect parallel loops and generate parallel code using libgomp | |
1714 | primitives. Returns true if some loop was parallelized, false | |
1715 | otherwise. */ | |
1716 | ||
1717 | bool | |
1718 | parallelize_loops (void) | |
1719 | { | |
1720 | unsigned n_threads = flag_tree_parallelize_loops; | |
1721 | bool changed = false; | |
1722 | struct loop *loop; | |
1723 | struct tree_niter_desc niter_desc; | |
1724 | loop_iterator li; | |
a509ebb5 | 1725 | htab_t reduction_list; |
5f40b3cb ZD |
1726 | |
1727 | /* Do not parallelize loops in the functions created by parallelization. */ | |
1728 | if (parallelized_function_p (cfun->decl)) | |
1729 | return false; | |
1730 | ||
a509ebb5 RL |
1731 | reduction_list = htab_create (10, reduction_info_hash, |
1732 | reduction_info_eq, free); | |
1733 | ||
5f40b3cb ZD |
1734 | FOR_EACH_LOOP (li, loop, 0) |
1735 | { | |
a509ebb5 | 1736 | htab_empty (reduction_list); |
5f40b3cb ZD |
1737 | if (/* Do not bother with loops in cold areas. */ |
1738 | !maybe_hot_bb_p (loop->header) | |
1739 | /* Or loops that roll too little. */ | |
1740 | || expected_loop_iterations (loop) <= n_threads | |
1741 | /* And of course, the loop must be parallelizable. */ | |
1742 | || !can_duplicate_loop_p (loop) | |
a509ebb5 | 1743 | || !loop_parallel_p (loop, reduction_list, &niter_desc)) |
5f40b3cb ZD |
1744 | continue; |
1745 | ||
1746 | changed = true; | |
a509ebb5 | 1747 | gen_parallel_loop (loop, reduction_list, n_threads, &niter_desc); |
5f40b3cb ZD |
1748 | verify_flow_info (); |
1749 | verify_dominators (CDI_DOMINATORS); | |
1750 | verify_loop_structure (); | |
1751 | verify_loop_closed_ssa (); | |
1752 | } | |
1753 | ||
a509ebb5 | 1754 | htab_delete (reduction_list); |
5f40b3cb ZD |
1755 | return changed; |
1756 | } | |
1757 | ||
1758 | #include "gt-tree-parloops.h" |