]>
Commit | Line | Data |
---|---|---|
1 | /* Loop autoparallelization. | |
2 | Copyright (C) 2006-2021 Free Software Foundation, Inc. | |
3 | Contributed by Sebastian Pop <pop@cri.ensmp.fr> | |
4 | Zdenek Dvorak <dvorakz@suse.cz> and Razya Ladelsky <razya@il.ibm.com>. | |
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 3, 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 COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
25 | #include "backend.h" | |
26 | #include "tree.h" | |
27 | #include "gimple.h" | |
28 | #include "cfghooks.h" | |
29 | #include "tree-pass.h" | |
30 | #include "ssa.h" | |
31 | #include "cgraph.h" | |
32 | #include "gimple-pretty-print.h" | |
33 | #include "fold-const.h" | |
34 | #include "gimplify.h" | |
35 | #include "gimple-iterator.h" | |
36 | #include "gimplify-me.h" | |
37 | #include "gimple-walk.h" | |
38 | #include "stor-layout.h" | |
39 | #include "tree-nested.h" | |
40 | #include "tree-cfg.h" | |
41 | #include "tree-ssa-loop-ivopts.h" | |
42 | #include "tree-ssa-loop-manip.h" | |
43 | #include "tree-ssa-loop-niter.h" | |
44 | #include "tree-ssa-loop.h" | |
45 | #include "tree-into-ssa.h" | |
46 | #include "cfgloop.h" | |
47 | #include "tree-scalar-evolution.h" | |
48 | #include "langhooks.h" | |
49 | #include "tree-vectorizer.h" | |
50 | #include "tree-hasher.h" | |
51 | #include "tree-parloops.h" | |
52 | #include "omp-general.h" | |
53 | #include "omp-low.h" | |
54 | #include "tree-ssa.h" | |
55 | #include "tree-ssa-alias.h" | |
56 | #include "tree-eh.h" | |
57 | #include "gomp-constants.h" | |
58 | #include "tree-dfa.h" | |
59 | #include "stringpool.h" | |
60 | #include "attribs.h" | |
61 | ||
62 | /* This pass tries to distribute iterations of loops into several threads. | |
63 | The implementation is straightforward -- for each loop we test whether its | |
64 | iterations are independent, and if it is the case (and some additional | |
65 | conditions regarding profitability and correctness are satisfied), we | |
66 | add GIMPLE_OMP_PARALLEL and GIMPLE_OMP_FOR codes and let omp expansion | |
67 | machinery do its job. | |
68 | ||
69 | The most of the complexity is in bringing the code into shape expected | |
70 | by the omp expanders: | |
71 | -- for GIMPLE_OMP_FOR, ensuring that the loop has only one induction | |
72 | variable and that the exit test is at the start of the loop body | |
73 | -- for GIMPLE_OMP_PARALLEL, replacing the references to local addressable | |
74 | variables by accesses through pointers, and breaking up ssa chains | |
75 | by storing the values incoming to the parallelized loop to a structure | |
76 | passed to the new function as an argument (something similar is done | |
77 | in omp gimplification, unfortunately only a small part of the code | |
78 | can be shared). | |
79 | ||
80 | TODO: | |
81 | -- if there are several parallelizable loops in a function, it may be | |
82 | possible to generate the threads just once (using synchronization to | |
83 | ensure that cross-loop dependences are obeyed). | |
84 | -- handling of common reduction patterns for outer loops. | |
85 | ||
86 | More info can also be found at http://gcc.gnu.org/wiki/AutoParInGCC */ | |
87 | /* | |
88 | Reduction handling: | |
89 | currently we use code inspired by vect_force_simple_reduction to detect | |
90 | reduction patterns. | |
91 | The code transformation will be introduced by an example. | |
92 | ||
93 | ||
94 | parloop | |
95 | { | |
96 | int sum=1; | |
97 | ||
98 | for (i = 0; i < N; i++) | |
99 | { | |
100 | x[i] = i + 3; | |
101 | sum+=x[i]; | |
102 | } | |
103 | } | |
104 | ||
105 | gimple-like code: | |
106 | header_bb: | |
107 | ||
108 | # sum_29 = PHI <sum_11(5), 1(3)> | |
109 | # i_28 = PHI <i_12(5), 0(3)> | |
110 | D.1795_8 = i_28 + 3; | |
111 | x[i_28] = D.1795_8; | |
112 | sum_11 = D.1795_8 + sum_29; | |
113 | i_12 = i_28 + 1; | |
114 | if (N_6(D) > i_12) | |
115 | goto header_bb; | |
116 | ||
117 | ||
118 | exit_bb: | |
119 | ||
120 | # sum_21 = PHI <sum_11(4)> | |
121 | printf (&"%d"[0], sum_21); | |
122 | ||
123 | ||
124 | after reduction transformation (only relevant parts): | |
125 | ||
126 | parloop | |
127 | { | |
128 | ||
129 | .... | |
130 | ||
131 | ||
132 | # Storing the initial value given by the user. # | |
133 | ||
134 | .paral_data_store.32.sum.27 = 1; | |
135 | ||
136 | #pragma omp parallel num_threads(4) | |
137 | ||
138 | #pragma omp for schedule(static) | |
139 | ||
140 | # The neutral element corresponding to the particular | |
141 | reduction's operation, e.g. 0 for PLUS_EXPR, | |
142 | 1 for MULT_EXPR, etc. replaces the user's initial value. # | |
143 | ||
144 | # sum.27_29 = PHI <sum.27_11, 0> | |
145 | ||
146 | sum.27_11 = D.1827_8 + sum.27_29; | |
147 | ||
148 | GIMPLE_OMP_CONTINUE | |
149 | ||
150 | # Adding this reduction phi is done at create_phi_for_local_result() # | |
151 | # sum.27_56 = PHI <sum.27_11, 0> | |
152 | GIMPLE_OMP_RETURN | |
153 | ||
154 | # Creating the atomic operation is done at | |
155 | create_call_for_reduction_1() # | |
156 | ||
157 | #pragma omp atomic_load | |
158 | D.1839_59 = *&.paral_data_load.33_51->reduction.23; | |
159 | D.1840_60 = sum.27_56 + D.1839_59; | |
160 | #pragma omp atomic_store (D.1840_60); | |
161 | ||
162 | GIMPLE_OMP_RETURN | |
163 | ||
164 | # collecting the result after the join of the threads is done at | |
165 | create_loads_for_reductions(). | |
166 | The value computed by the threads is loaded from the | |
167 | shared struct. # | |
168 | ||
169 | ||
170 | .paral_data_load.33_52 = &.paral_data_store.32; | |
171 | sum_37 = .paral_data_load.33_52->sum.27; | |
172 | sum_43 = D.1795_41 + sum_37; | |
173 | ||
174 | exit bb: | |
175 | # sum_21 = PHI <sum_43, sum_26> | |
176 | printf (&"%d"[0], sum_21); | |
177 | ||
178 | ... | |
179 | ||
180 | } | |
181 | ||
182 | */ | |
183 | ||
184 | /* Error reporting helper for parloops_is_simple_reduction below. GIMPLE | |
185 | statement STMT is printed with a message MSG. */ | |
186 | ||
187 | static void | |
188 | report_ploop_op (dump_flags_t msg_type, gimple *stmt, const char *msg) | |
189 | { | |
190 | dump_printf_loc (msg_type, vect_location, "%s%G", msg, stmt); | |
191 | } | |
192 | ||
193 | /* DEF_STMT_INFO occurs in a loop that contains a potential reduction | |
194 | operation. Return true if the results of DEF_STMT_INFO are something | |
195 | that can be accumulated by such a reduction. */ | |
196 | ||
197 | static bool | |
198 | parloops_valid_reduction_input_p (stmt_vec_info def_stmt_info) | |
199 | { | |
200 | return (is_gimple_assign (def_stmt_info->stmt) | |
201 | || is_gimple_call (def_stmt_info->stmt) | |
202 | || STMT_VINFO_DEF_TYPE (def_stmt_info) == vect_induction_def | |
203 | || (gimple_code (def_stmt_info->stmt) == GIMPLE_PHI | |
204 | && STMT_VINFO_DEF_TYPE (def_stmt_info) == vect_internal_def | |
205 | && !is_loop_header_bb_p (gimple_bb (def_stmt_info->stmt)))); | |
206 | } | |
207 | ||
208 | /* Detect SLP reduction of the form: | |
209 | ||
210 | #a1 = phi <a5, a0> | |
211 | a2 = operation (a1) | |
212 | a3 = operation (a2) | |
213 | a4 = operation (a3) | |
214 | a5 = operation (a4) | |
215 | ||
216 | #a = phi <a5> | |
217 | ||
218 | PHI is the reduction phi node (#a1 = phi <a5, a0> above) | |
219 | FIRST_STMT is the first reduction stmt in the chain | |
220 | (a2 = operation (a1)). | |
221 | ||
222 | Return TRUE if a reduction chain was detected. */ | |
223 | ||
224 | static bool | |
225 | parloops_is_slp_reduction (loop_vec_info loop_info, gimple *phi, | |
226 | gimple *first_stmt) | |
227 | { | |
228 | class loop *loop = (gimple_bb (phi))->loop_father; | |
229 | class loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
230 | enum tree_code code; | |
231 | gimple *loop_use_stmt = NULL; | |
232 | stmt_vec_info use_stmt_info; | |
233 | tree lhs; | |
234 | imm_use_iterator imm_iter; | |
235 | use_operand_p use_p; | |
236 | int nloop_uses, size = 0, n_out_of_loop_uses; | |
237 | bool found = false; | |
238 | ||
239 | if (loop != vect_loop) | |
240 | return false; | |
241 | ||
242 | auto_vec<stmt_vec_info, 8> reduc_chain; | |
243 | lhs = PHI_RESULT (phi); | |
244 | code = gimple_assign_rhs_code (first_stmt); | |
245 | while (1) | |
246 | { | |
247 | nloop_uses = 0; | |
248 | n_out_of_loop_uses = 0; | |
249 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) | |
250 | { | |
251 | gimple *use_stmt = USE_STMT (use_p); | |
252 | if (is_gimple_debug (use_stmt)) | |
253 | continue; | |
254 | ||
255 | /* Check if we got back to the reduction phi. */ | |
256 | if (use_stmt == phi) | |
257 | { | |
258 | loop_use_stmt = use_stmt; | |
259 | found = true; | |
260 | break; | |
261 | } | |
262 | ||
263 | if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) | |
264 | { | |
265 | loop_use_stmt = use_stmt; | |
266 | nloop_uses++; | |
267 | } | |
268 | else | |
269 | n_out_of_loop_uses++; | |
270 | ||
271 | /* There are can be either a single use in the loop or two uses in | |
272 | phi nodes. */ | |
273 | if (nloop_uses > 1 || (n_out_of_loop_uses && nloop_uses)) | |
274 | return false; | |
275 | } | |
276 | ||
277 | if (found) | |
278 | break; | |
279 | ||
280 | /* We reached a statement with no loop uses. */ | |
281 | if (nloop_uses == 0) | |
282 | return false; | |
283 | ||
284 | /* This is a loop exit phi, and we haven't reached the reduction phi. */ | |
285 | if (gimple_code (loop_use_stmt) == GIMPLE_PHI) | |
286 | return false; | |
287 | ||
288 | if (!is_gimple_assign (loop_use_stmt) | |
289 | || code != gimple_assign_rhs_code (loop_use_stmt) | |
290 | || !flow_bb_inside_loop_p (loop, gimple_bb (loop_use_stmt))) | |
291 | return false; | |
292 | ||
293 | /* Insert USE_STMT into reduction chain. */ | |
294 | use_stmt_info = loop_info->lookup_stmt (loop_use_stmt); | |
295 | reduc_chain.safe_push (use_stmt_info); | |
296 | ||
297 | lhs = gimple_assign_lhs (loop_use_stmt); | |
298 | size++; | |
299 | } | |
300 | ||
301 | if (!found || loop_use_stmt != phi || size < 2) | |
302 | return false; | |
303 | ||
304 | /* Swap the operands, if needed, to make the reduction operand be the second | |
305 | operand. */ | |
306 | lhs = PHI_RESULT (phi); | |
307 | for (unsigned i = 0; i < reduc_chain.length (); ++i) | |
308 | { | |
309 | gassign *next_stmt = as_a <gassign *> (reduc_chain[i]->stmt); | |
310 | if (gimple_assign_rhs2 (next_stmt) == lhs) | |
311 | { | |
312 | tree op = gimple_assign_rhs1 (next_stmt); | |
313 | stmt_vec_info def_stmt_info = loop_info->lookup_def (op); | |
314 | ||
315 | /* Check that the other def is either defined in the loop | |
316 | ("vect_internal_def"), or it's an induction (defined by a | |
317 | loop-header phi-node). */ | |
318 | if (def_stmt_info | |
319 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt_info->stmt)) | |
320 | && parloops_valid_reduction_input_p (def_stmt_info)) | |
321 | { | |
322 | lhs = gimple_assign_lhs (next_stmt); | |
323 | continue; | |
324 | } | |
325 | ||
326 | return false; | |
327 | } | |
328 | else | |
329 | { | |
330 | tree op = gimple_assign_rhs2 (next_stmt); | |
331 | stmt_vec_info def_stmt_info = loop_info->lookup_def (op); | |
332 | ||
333 | /* Check that the other def is either defined in the loop | |
334 | ("vect_internal_def"), or it's an induction (defined by a | |
335 | loop-header phi-node). */ | |
336 | if (def_stmt_info | |
337 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt_info->stmt)) | |
338 | && parloops_valid_reduction_input_p (def_stmt_info)) | |
339 | { | |
340 | if (dump_enabled_p ()) | |
341 | dump_printf_loc (MSG_NOTE, vect_location, "swapping oprnds: %G", | |
342 | next_stmt); | |
343 | ||
344 | swap_ssa_operands (next_stmt, | |
345 | gimple_assign_rhs1_ptr (next_stmt), | |
346 | gimple_assign_rhs2_ptr (next_stmt)); | |
347 | update_stmt (next_stmt); | |
348 | } | |
349 | else | |
350 | return false; | |
351 | } | |
352 | ||
353 | lhs = gimple_assign_lhs (next_stmt); | |
354 | } | |
355 | ||
356 | /* Build up the actual chain. */ | |
357 | for (unsigned i = 0; i < reduc_chain.length () - 1; ++i) | |
358 | { | |
359 | REDUC_GROUP_FIRST_ELEMENT (reduc_chain[i]) = reduc_chain[0]; | |
360 | REDUC_GROUP_NEXT_ELEMENT (reduc_chain[i]) = reduc_chain[i+1]; | |
361 | } | |
362 | REDUC_GROUP_FIRST_ELEMENT (reduc_chain.last ()) = reduc_chain[0]; | |
363 | REDUC_GROUP_NEXT_ELEMENT (reduc_chain.last ()) = NULL; | |
364 | ||
365 | /* Save the chain for further analysis in SLP detection. */ | |
366 | LOOP_VINFO_REDUCTION_CHAINS (loop_info).safe_push (reduc_chain[0]); | |
367 | REDUC_GROUP_SIZE (reduc_chain[0]) = size; | |
368 | ||
369 | return true; | |
370 | } | |
371 | ||
372 | /* Return true if we need an in-order reduction for operation CODE | |
373 | on type TYPE. NEED_WRAPPING_INTEGRAL_OVERFLOW is true if integer | |
374 | overflow must wrap. */ | |
375 | ||
376 | static bool | |
377 | parloops_needs_fold_left_reduction_p (tree type, tree_code code, | |
378 | bool need_wrapping_integral_overflow) | |
379 | { | |
380 | /* CHECKME: check for !flag_finite_math_only too? */ | |
381 | if (SCALAR_FLOAT_TYPE_P (type)) | |
382 | switch (code) | |
383 | { | |
384 | case MIN_EXPR: | |
385 | case MAX_EXPR: | |
386 | return false; | |
387 | ||
388 | default: | |
389 | return !flag_associative_math; | |
390 | } | |
391 | ||
392 | if (INTEGRAL_TYPE_P (type)) | |
393 | { | |
394 | if (!operation_no_trapping_overflow (type, code)) | |
395 | return true; | |
396 | if (need_wrapping_integral_overflow | |
397 | && !TYPE_OVERFLOW_WRAPS (type) | |
398 | && operation_can_overflow (code)) | |
399 | return true; | |
400 | return false; | |
401 | } | |
402 | ||
403 | if (SAT_FIXED_POINT_TYPE_P (type)) | |
404 | return true; | |
405 | ||
406 | return false; | |
407 | } | |
408 | ||
409 | ||
410 | /* Function parloops_is_simple_reduction | |
411 | ||
412 | (1) Detect a cross-iteration def-use cycle that represents a simple | |
413 | reduction computation. We look for the following pattern: | |
414 | ||
415 | loop_header: | |
416 | a1 = phi < a0, a2 > | |
417 | a3 = ... | |
418 | a2 = operation (a3, a1) | |
419 | ||
420 | or | |
421 | ||
422 | a3 = ... | |
423 | loop_header: | |
424 | a1 = phi < a0, a2 > | |
425 | a2 = operation (a3, a1) | |
426 | ||
427 | such that: | |
428 | 1. operation is commutative and associative and it is safe to | |
429 | change the order of the computation | |
430 | 2. no uses for a2 in the loop (a2 is used out of the loop) | |
431 | 3. no uses of a1 in the loop besides the reduction operation | |
432 | 4. no uses of a1 outside the loop. | |
433 | ||
434 | Conditions 1,4 are tested here. | |
435 | Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized. | |
436 | ||
437 | (2) Detect a cross-iteration def-use cycle in nested loops, i.e., | |
438 | nested cycles. | |
439 | ||
440 | (3) Detect cycles of phi nodes in outer-loop vectorization, i.e., double | |
441 | reductions: | |
442 | ||
443 | a1 = phi < a0, a2 > | |
444 | inner loop (def of a3) | |
445 | a2 = phi < a3 > | |
446 | ||
447 | (4) Detect condition expressions, ie: | |
448 | for (int i = 0; i < N; i++) | |
449 | if (a[i] < val) | |
450 | ret_val = a[i]; | |
451 | ||
452 | */ | |
453 | ||
454 | static stmt_vec_info | |
455 | parloops_is_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info, | |
456 | bool *double_reduc, | |
457 | bool need_wrapping_integral_overflow, | |
458 | enum vect_reduction_type *v_reduc_type) | |
459 | { | |
460 | gphi *phi = as_a <gphi *> (phi_info->stmt); | |
461 | class loop *loop = (gimple_bb (phi))->loop_father; | |
462 | class loop *vect_loop = LOOP_VINFO_LOOP (loop_info); | |
463 | bool nested_in_vect_loop = flow_loop_nested_p (vect_loop, loop); | |
464 | gimple *phi_use_stmt = NULL; | |
465 | enum tree_code orig_code, code; | |
466 | tree op1, op2, op3 = NULL_TREE, op4 = NULL_TREE; | |
467 | tree type; | |
468 | tree name; | |
469 | imm_use_iterator imm_iter; | |
470 | use_operand_p use_p; | |
471 | bool phi_def; | |
472 | ||
473 | *double_reduc = false; | |
474 | *v_reduc_type = TREE_CODE_REDUCTION; | |
475 | ||
476 | tree phi_name = PHI_RESULT (phi); | |
477 | /* ??? If there are no uses of the PHI result the inner loop reduction | |
478 | won't be detected as possibly double-reduction by vectorizable_reduction | |
479 | because that tries to walk the PHI arg from the preheader edge which | |
480 | can be constant. See PR60382. */ | |
481 | if (has_zero_uses (phi_name)) | |
482 | return NULL; | |
483 | unsigned nphi_def_loop_uses = 0; | |
484 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, phi_name) | |
485 | { | |
486 | gimple *use_stmt = USE_STMT (use_p); | |
487 | if (is_gimple_debug (use_stmt)) | |
488 | continue; | |
489 | ||
490 | if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) | |
491 | { | |
492 | if (dump_enabled_p ()) | |
493 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
494 | "intermediate value used outside loop.\n"); | |
495 | ||
496 | return NULL; | |
497 | } | |
498 | ||
499 | nphi_def_loop_uses++; | |
500 | phi_use_stmt = use_stmt; | |
501 | } | |
502 | ||
503 | edge latch_e = loop_latch_edge (loop); | |
504 | tree loop_arg = PHI_ARG_DEF_FROM_EDGE (phi, latch_e); | |
505 | if (TREE_CODE (loop_arg) != SSA_NAME) | |
506 | { | |
507 | if (dump_enabled_p ()) | |
508 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
509 | "reduction: not ssa_name: %T\n", loop_arg); | |
510 | return NULL; | |
511 | } | |
512 | ||
513 | stmt_vec_info def_stmt_info = loop_info->lookup_def (loop_arg); | |
514 | if (!def_stmt_info | |
515 | || !flow_bb_inside_loop_p (loop, gimple_bb (def_stmt_info->stmt))) | |
516 | return NULL; | |
517 | ||
518 | if (gassign *def_stmt = dyn_cast <gassign *> (def_stmt_info->stmt)) | |
519 | { | |
520 | name = gimple_assign_lhs (def_stmt); | |
521 | phi_def = false; | |
522 | } | |
523 | else if (gphi *def_stmt = dyn_cast <gphi *> (def_stmt_info->stmt)) | |
524 | { | |
525 | name = PHI_RESULT (def_stmt); | |
526 | phi_def = true; | |
527 | } | |
528 | else | |
529 | { | |
530 | if (dump_enabled_p ()) | |
531 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
532 | "reduction: unhandled reduction operation: %G", | |
533 | def_stmt_info->stmt); | |
534 | return NULL; | |
535 | } | |
536 | ||
537 | unsigned nlatch_def_loop_uses = 0; | |
538 | auto_vec<gphi *, 3> lcphis; | |
539 | bool inner_loop_of_double_reduc = false; | |
540 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name) | |
541 | { | |
542 | gimple *use_stmt = USE_STMT (use_p); | |
543 | if (is_gimple_debug (use_stmt)) | |
544 | continue; | |
545 | if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) | |
546 | nlatch_def_loop_uses++; | |
547 | else | |
548 | { | |
549 | /* We can have more than one loop-closed PHI. */ | |
550 | lcphis.safe_push (as_a <gphi *> (use_stmt)); | |
551 | if (nested_in_vect_loop | |
552 | && (STMT_VINFO_DEF_TYPE (loop_info->lookup_stmt (use_stmt)) | |
553 | == vect_double_reduction_def)) | |
554 | inner_loop_of_double_reduc = true; | |
555 | } | |
556 | } | |
557 | ||
558 | /* If this isn't a nested cycle or if the nested cycle reduction value | |
559 | is used ouside of the inner loop we cannot handle uses of the reduction | |
560 | value. */ | |
561 | if ((!nested_in_vect_loop || inner_loop_of_double_reduc) | |
562 | && (nlatch_def_loop_uses > 1 || nphi_def_loop_uses > 1)) | |
563 | { | |
564 | if (dump_enabled_p ()) | |
565 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
566 | "reduction used in loop.\n"); | |
567 | return NULL; | |
568 | } | |
569 | ||
570 | /* If DEF_STMT is a phi node itself, we expect it to have a single argument | |
571 | defined in the inner loop. */ | |
572 | if (phi_def) | |
573 | { | |
574 | gphi *def_stmt = as_a <gphi *> (def_stmt_info->stmt); | |
575 | op1 = PHI_ARG_DEF (def_stmt, 0); | |
576 | ||
577 | if (gimple_phi_num_args (def_stmt) != 1 | |
578 | || TREE_CODE (op1) != SSA_NAME) | |
579 | { | |
580 | if (dump_enabled_p ()) | |
581 | dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, | |
582 | "unsupported phi node definition.\n"); | |
583 | ||
584 | return NULL; | |
585 | } | |
586 | ||
587 | gimple *def1 = SSA_NAME_DEF_STMT (op1); | |
588 | if (gimple_bb (def1) | |
589 | && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) | |
590 | && loop->inner | |
591 | && flow_bb_inside_loop_p (loop->inner, gimple_bb (def1)) | |
592 | && is_gimple_assign (def1) | |
593 | && is_a <gphi *> (phi_use_stmt) | |
594 | && flow_bb_inside_loop_p (loop->inner, gimple_bb (phi_use_stmt))) | |
595 | { | |
596 | if (dump_enabled_p ()) | |
597 | report_ploop_op (MSG_NOTE, def_stmt, | |
598 | "detected double reduction: "); | |
599 | ||
600 | *double_reduc = true; | |
601 | return def_stmt_info; | |
602 | } | |
603 | ||
604 | return NULL; | |
605 | } | |
606 | ||
607 | /* If we are vectorizing an inner reduction we are executing that | |
608 | in the original order only in case we are not dealing with a | |
609 | double reduction. */ | |
610 | bool check_reduction = true; | |
611 | if (flow_loop_nested_p (vect_loop, loop)) | |
612 | { | |
613 | gphi *lcphi; | |
614 | unsigned i; | |
615 | check_reduction = false; | |
616 | FOR_EACH_VEC_ELT (lcphis, i, lcphi) | |
617 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, gimple_phi_result (lcphi)) | |
618 | { | |
619 | gimple *use_stmt = USE_STMT (use_p); | |
620 | if (is_gimple_debug (use_stmt)) | |
621 | continue; | |
622 | if (! flow_bb_inside_loop_p (vect_loop, gimple_bb (use_stmt))) | |
623 | check_reduction = true; | |
624 | } | |
625 | } | |
626 | ||
627 | gassign *def_stmt = as_a <gassign *> (def_stmt_info->stmt); | |
628 | code = orig_code = gimple_assign_rhs_code (def_stmt); | |
629 | ||
630 | if (nested_in_vect_loop && !check_reduction) | |
631 | { | |
632 | /* FIXME: Even for non-reductions code generation is funneled | |
633 | through vectorizable_reduction for the stmt defining the | |
634 | PHI latch value. So we have to artificially restrict ourselves | |
635 | for the supported operations. */ | |
636 | switch (get_gimple_rhs_class (code)) | |
637 | { | |
638 | case GIMPLE_BINARY_RHS: | |
639 | case GIMPLE_TERNARY_RHS: | |
640 | break; | |
641 | default: | |
642 | /* Not supported by vectorizable_reduction. */ | |
643 | if (dump_enabled_p ()) | |
644 | report_ploop_op (MSG_MISSED_OPTIMIZATION, def_stmt, | |
645 | "nested cycle: not handled operation: "); | |
646 | return NULL; | |
647 | } | |
648 | if (dump_enabled_p ()) | |
649 | report_ploop_op (MSG_NOTE, def_stmt, "detected nested cycle: "); | |
650 | return def_stmt_info; | |
651 | } | |
652 | ||
653 | /* We can handle "res -= x[i]", which is non-associative by | |
654 | simply rewriting this into "res += -x[i]". Avoid changing | |
655 | gimple instruction for the first simple tests and only do this | |
656 | if we're allowed to change code at all. */ | |
657 | if (code == MINUS_EXPR && gimple_assign_rhs2 (def_stmt) != phi_name) | |
658 | code = PLUS_EXPR; | |
659 | ||
660 | if (code == COND_EXPR) | |
661 | { | |
662 | if (! nested_in_vect_loop) | |
663 | *v_reduc_type = COND_REDUCTION; | |
664 | ||
665 | op3 = gimple_assign_rhs1 (def_stmt); | |
666 | if (COMPARISON_CLASS_P (op3)) | |
667 | { | |
668 | op4 = TREE_OPERAND (op3, 1); | |
669 | op3 = TREE_OPERAND (op3, 0); | |
670 | } | |
671 | if (op3 == phi_name || op4 == phi_name) | |
672 | { | |
673 | if (dump_enabled_p ()) | |
674 | report_ploop_op (MSG_MISSED_OPTIMIZATION, def_stmt, | |
675 | "reduction: condition depends on previous" | |
676 | " iteration: "); | |
677 | return NULL; | |
678 | } | |
679 | ||
680 | op1 = gimple_assign_rhs2 (def_stmt); | |
681 | op2 = gimple_assign_rhs3 (def_stmt); | |
682 | } | |
683 | else if (!commutative_tree_code (code) || !associative_tree_code (code)) | |
684 | { | |
685 | if (dump_enabled_p ()) | |
686 | report_ploop_op (MSG_MISSED_OPTIMIZATION, def_stmt, | |
687 | "reduction: not commutative/associative: "); | |
688 | return NULL; | |
689 | } | |
690 | else if (get_gimple_rhs_class (code) == GIMPLE_BINARY_RHS) | |
691 | { | |
692 | op1 = gimple_assign_rhs1 (def_stmt); | |
693 | op2 = gimple_assign_rhs2 (def_stmt); | |
694 | } | |
695 | else | |
696 | { | |
697 | if (dump_enabled_p ()) | |
698 | report_ploop_op (MSG_MISSED_OPTIMIZATION, def_stmt, | |
699 | "reduction: not handled operation: "); | |
700 | return NULL; | |
701 | } | |
702 | ||
703 | if (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op2) != SSA_NAME) | |
704 | { | |
705 | if (dump_enabled_p ()) | |
706 | report_ploop_op (MSG_MISSED_OPTIMIZATION, def_stmt, | |
707 | "reduction: both uses not ssa_names: "); | |
708 | ||
709 | return NULL; | |
710 | } | |
711 | ||
712 | type = TREE_TYPE (gimple_assign_lhs (def_stmt)); | |
713 | if ((TREE_CODE (op1) == SSA_NAME | |
714 | && !types_compatible_p (type,TREE_TYPE (op1))) | |
715 | || (TREE_CODE (op2) == SSA_NAME | |
716 | && !types_compatible_p (type, TREE_TYPE (op2))) | |
717 | || (op3 && TREE_CODE (op3) == SSA_NAME | |
718 | && !types_compatible_p (type, TREE_TYPE (op3))) | |
719 | || (op4 && TREE_CODE (op4) == SSA_NAME | |
720 | && !types_compatible_p (type, TREE_TYPE (op4)))) | |
721 | { | |
722 | if (dump_enabled_p ()) | |
723 | { | |
724 | dump_printf_loc (MSG_NOTE, vect_location, | |
725 | "reduction: multiple types: operation type: " | |
726 | "%T, operands types: %T,%T", | |
727 | type, TREE_TYPE (op1), TREE_TYPE (op2)); | |
728 | if (op3) | |
729 | dump_printf (MSG_NOTE, ",%T", TREE_TYPE (op3)); | |
730 | ||
731 | if (op4) | |
732 | dump_printf (MSG_NOTE, ",%T", TREE_TYPE (op4)); | |
733 | dump_printf (MSG_NOTE, "\n"); | |
734 | } | |
735 | ||
736 | return NULL; | |
737 | } | |
738 | ||
739 | /* Check whether it's ok to change the order of the computation. | |
740 | Generally, when vectorizing a reduction we change the order of the | |
741 | computation. This may change the behavior of the program in some | |
742 | cases, so we need to check that this is ok. One exception is when | |
743 | vectorizing an outer-loop: the inner-loop is executed sequentially, | |
744 | and therefore vectorizing reductions in the inner-loop during | |
745 | outer-loop vectorization is safe. */ | |
746 | if (check_reduction | |
747 | && *v_reduc_type == TREE_CODE_REDUCTION | |
748 | && parloops_needs_fold_left_reduction_p (type, code, | |
749 | need_wrapping_integral_overflow)) | |
750 | *v_reduc_type = FOLD_LEFT_REDUCTION; | |
751 | ||
752 | /* Reduction is safe. We're dealing with one of the following: | |
753 | 1) integer arithmetic and no trapv | |
754 | 2) floating point arithmetic, and special flags permit this optimization | |
755 | 3) nested cycle (i.e., outer loop vectorization). */ | |
756 | stmt_vec_info def1_info = loop_info->lookup_def (op1); | |
757 | stmt_vec_info def2_info = loop_info->lookup_def (op2); | |
758 | if (code != COND_EXPR && !def1_info && !def2_info) | |
759 | { | |
760 | if (dump_enabled_p ()) | |
761 | report_ploop_op (MSG_NOTE, def_stmt, | |
762 | "reduction: no defs for operands: "); | |
763 | return NULL; | |
764 | } | |
765 | ||
766 | /* Check that one def is the reduction def, defined by PHI, | |
767 | the other def is either defined in the loop ("vect_internal_def"), | |
768 | or it's an induction (defined by a loop-header phi-node). */ | |
769 | ||
770 | if (def2_info | |
771 | && def2_info->stmt == phi | |
772 | && (code == COND_EXPR | |
773 | || !def1_info | |
774 | || !flow_bb_inside_loop_p (loop, gimple_bb (def1_info->stmt)) | |
775 | || parloops_valid_reduction_input_p (def1_info))) | |
776 | { | |
777 | if (dump_enabled_p ()) | |
778 | report_ploop_op (MSG_NOTE, def_stmt, "detected reduction: "); | |
779 | return def_stmt_info; | |
780 | } | |
781 | ||
782 | if (def1_info | |
783 | && def1_info->stmt == phi | |
784 | && (code == COND_EXPR | |
785 | || !def2_info | |
786 | || !flow_bb_inside_loop_p (loop, gimple_bb (def2_info->stmt)) | |
787 | || parloops_valid_reduction_input_p (def2_info))) | |
788 | { | |
789 | if (! nested_in_vect_loop && orig_code != MINUS_EXPR) | |
790 | { | |
791 | /* Check if we can swap operands (just for simplicity - so that | |
792 | the rest of the code can assume that the reduction variable | |
793 | is always the last (second) argument). */ | |
794 | if (code == COND_EXPR) | |
795 | { | |
796 | /* Swap cond_expr by inverting the condition. */ | |
797 | tree cond_expr = gimple_assign_rhs1 (def_stmt); | |
798 | enum tree_code invert_code = ERROR_MARK; | |
799 | enum tree_code cond_code = TREE_CODE (cond_expr); | |
800 | ||
801 | if (TREE_CODE_CLASS (cond_code) == tcc_comparison) | |
802 | { | |
803 | bool honor_nans = HONOR_NANS (TREE_OPERAND (cond_expr, 0)); | |
804 | invert_code = invert_tree_comparison (cond_code, honor_nans); | |
805 | } | |
806 | if (invert_code != ERROR_MARK) | |
807 | { | |
808 | TREE_SET_CODE (cond_expr, invert_code); | |
809 | swap_ssa_operands (def_stmt, | |
810 | gimple_assign_rhs2_ptr (def_stmt), | |
811 | gimple_assign_rhs3_ptr (def_stmt)); | |
812 | } | |
813 | else | |
814 | { | |
815 | if (dump_enabled_p ()) | |
816 | report_ploop_op (MSG_NOTE, def_stmt, | |
817 | "detected reduction: cannot swap operands " | |
818 | "for cond_expr"); | |
819 | return NULL; | |
820 | } | |
821 | } | |
822 | else | |
823 | swap_ssa_operands (def_stmt, gimple_assign_rhs1_ptr (def_stmt), | |
824 | gimple_assign_rhs2_ptr (def_stmt)); | |
825 | ||
826 | if (dump_enabled_p ()) | |
827 | report_ploop_op (MSG_NOTE, def_stmt, | |
828 | "detected reduction: need to swap operands: "); | |
829 | } | |
830 | else | |
831 | { | |
832 | if (dump_enabled_p ()) | |
833 | report_ploop_op (MSG_NOTE, def_stmt, "detected reduction: "); | |
834 | } | |
835 | ||
836 | return def_stmt_info; | |
837 | } | |
838 | ||
839 | /* Try to find SLP reduction chain. */ | |
840 | if (! nested_in_vect_loop | |
841 | && code != COND_EXPR | |
842 | && orig_code != MINUS_EXPR | |
843 | && parloops_is_slp_reduction (loop_info, phi, def_stmt)) | |
844 | { | |
845 | if (dump_enabled_p ()) | |
846 | report_ploop_op (MSG_NOTE, def_stmt, | |
847 | "reduction: detected reduction chain: "); | |
848 | ||
849 | return def_stmt_info; | |
850 | } | |
851 | ||
852 | /* Look for the expression computing loop_arg from loop PHI result. */ | |
853 | if (check_reduction_path (vect_location, loop, phi, loop_arg, code)) | |
854 | return def_stmt_info; | |
855 | ||
856 | if (dump_enabled_p ()) | |
857 | { | |
858 | report_ploop_op (MSG_MISSED_OPTIMIZATION, def_stmt, | |
859 | "reduction: unknown pattern: "); | |
860 | } | |
861 | ||
862 | return NULL; | |
863 | } | |
864 | ||
865 | /* Wrapper around vect_is_simple_reduction, which will modify code | |
866 | in-place if it enables detection of more reductions. Arguments | |
867 | as there. */ | |
868 | ||
869 | stmt_vec_info | |
870 | parloops_force_simple_reduction (loop_vec_info loop_info, stmt_vec_info phi_info, | |
871 | bool *double_reduc, | |
872 | bool need_wrapping_integral_overflow) | |
873 | { | |
874 | enum vect_reduction_type v_reduc_type; | |
875 | stmt_vec_info def_info | |
876 | = parloops_is_simple_reduction (loop_info, phi_info, double_reduc, | |
877 | need_wrapping_integral_overflow, | |
878 | &v_reduc_type); | |
879 | if (def_info) | |
880 | { | |
881 | STMT_VINFO_REDUC_TYPE (phi_info) = v_reduc_type; | |
882 | STMT_VINFO_REDUC_DEF (phi_info) = def_info; | |
883 | STMT_VINFO_REDUC_TYPE (def_info) = v_reduc_type; | |
884 | STMT_VINFO_REDUC_DEF (def_info) = phi_info; | |
885 | } | |
886 | return def_info; | |
887 | } | |
888 | ||
889 | /* Minimal number of iterations of a loop that should be executed in each | |
890 | thread. */ | |
891 | #define MIN_PER_THREAD param_parloops_min_per_thread | |
892 | ||
893 | /* Element of the hashtable, representing a | |
894 | reduction in the current loop. */ | |
895 | struct reduction_info | |
896 | { | |
897 | gimple *reduc_stmt; /* reduction statement. */ | |
898 | gimple *reduc_phi; /* The phi node defining the reduction. */ | |
899 | enum tree_code reduction_code;/* code for the reduction operation. */ | |
900 | unsigned reduc_version; /* SSA_NAME_VERSION of original reduc_phi | |
901 | result. */ | |
902 | gphi *keep_res; /* The PHI_RESULT of this phi is the resulting value | |
903 | of the reduction variable when existing the loop. */ | |
904 | tree initial_value; /* The initial value of the reduction var before entering the loop. */ | |
905 | tree field; /* the name of the field in the parloop data structure intended for reduction. */ | |
906 | tree reduc_addr; /* The address of the reduction variable for | |
907 | openacc reductions. */ | |
908 | tree init; /* reduction initialization value. */ | |
909 | gphi *new_phi; /* (helper field) Newly created phi node whose result | |
910 | will be passed to the atomic operation. Represents | |
911 | the local result each thread computed for the reduction | |
912 | operation. */ | |
913 | }; | |
914 | ||
915 | /* Reduction info hashtable helpers. */ | |
916 | ||
917 | struct reduction_hasher : free_ptr_hash <reduction_info> | |
918 | { | |
919 | static inline hashval_t hash (const reduction_info *); | |
920 | static inline bool equal (const reduction_info *, const reduction_info *); | |
921 | }; | |
922 | ||
923 | /* Equality and hash functions for hashtab code. */ | |
924 | ||
925 | inline bool | |
926 | reduction_hasher::equal (const reduction_info *a, const reduction_info *b) | |
927 | { | |
928 | return (a->reduc_phi == b->reduc_phi); | |
929 | } | |
930 | ||
931 | inline hashval_t | |
932 | reduction_hasher::hash (const reduction_info *a) | |
933 | { | |
934 | return a->reduc_version; | |
935 | } | |
936 | ||
937 | typedef hash_table<reduction_hasher> reduction_info_table_type; | |
938 | ||
939 | ||
940 | static struct reduction_info * | |
941 | reduction_phi (reduction_info_table_type *reduction_list, gimple *phi) | |
942 | { | |
943 | struct reduction_info tmpred, *red; | |
944 | ||
945 | if (reduction_list->is_empty () || phi == NULL) | |
946 | return NULL; | |
947 | ||
948 | if (gimple_uid (phi) == (unsigned int)-1 | |
949 | || gimple_uid (phi) == 0) | |
950 | return NULL; | |
951 | ||
952 | tmpred.reduc_phi = phi; | |
953 | tmpred.reduc_version = gimple_uid (phi); | |
954 | red = reduction_list->find (&tmpred); | |
955 | gcc_assert (red == NULL || red->reduc_phi == phi); | |
956 | ||
957 | return red; | |
958 | } | |
959 | ||
960 | /* Element of hashtable of names to copy. */ | |
961 | ||
962 | struct name_to_copy_elt | |
963 | { | |
964 | unsigned version; /* The version of the name to copy. */ | |
965 | tree new_name; /* The new name used in the copy. */ | |
966 | tree field; /* The field of the structure used to pass the | |
967 | value. */ | |
968 | }; | |
969 | ||
970 | /* Name copies hashtable helpers. */ | |
971 | ||
972 | struct name_to_copy_hasher : free_ptr_hash <name_to_copy_elt> | |
973 | { | |
974 | static inline hashval_t hash (const name_to_copy_elt *); | |
975 | static inline bool equal (const name_to_copy_elt *, const name_to_copy_elt *); | |
976 | }; | |
977 | ||
978 | /* Equality and hash functions for hashtab code. */ | |
979 | ||
980 | inline bool | |
981 | name_to_copy_hasher::equal (const name_to_copy_elt *a, const name_to_copy_elt *b) | |
982 | { | |
983 | return a->version == b->version; | |
984 | } | |
985 | ||
986 | inline hashval_t | |
987 | name_to_copy_hasher::hash (const name_to_copy_elt *a) | |
988 | { | |
989 | return (hashval_t) a->version; | |
990 | } | |
991 | ||
992 | typedef hash_table<name_to_copy_hasher> name_to_copy_table_type; | |
993 | ||
994 | /* A transformation matrix, which is a self-contained ROWSIZE x COLSIZE | |
995 | matrix. Rather than use floats, we simply keep a single DENOMINATOR that | |
996 | represents the denominator for every element in the matrix. */ | |
997 | typedef struct lambda_trans_matrix_s | |
998 | { | |
999 | lambda_matrix matrix; | |
1000 | int rowsize; | |
1001 | int colsize; | |
1002 | int denominator; | |
1003 | } *lambda_trans_matrix; | |
1004 | #define LTM_MATRIX(T) ((T)->matrix) | |
1005 | #define LTM_ROWSIZE(T) ((T)->rowsize) | |
1006 | #define LTM_COLSIZE(T) ((T)->colsize) | |
1007 | #define LTM_DENOMINATOR(T) ((T)->denominator) | |
1008 | ||
1009 | /* Allocate a new transformation matrix. */ | |
1010 | ||
1011 | static lambda_trans_matrix | |
1012 | lambda_trans_matrix_new (int colsize, int rowsize, | |
1013 | struct obstack * lambda_obstack) | |
1014 | { | |
1015 | lambda_trans_matrix ret; | |
1016 | ||
1017 | ret = (lambda_trans_matrix) | |
1018 | obstack_alloc (lambda_obstack, sizeof (struct lambda_trans_matrix_s)); | |
1019 | LTM_MATRIX (ret) = lambda_matrix_new (rowsize, colsize, lambda_obstack); | |
1020 | LTM_ROWSIZE (ret) = rowsize; | |
1021 | LTM_COLSIZE (ret) = colsize; | |
1022 | LTM_DENOMINATOR (ret) = 1; | |
1023 | return ret; | |
1024 | } | |
1025 | ||
1026 | /* Multiply a vector VEC by a matrix MAT. | |
1027 | MAT is an M*N matrix, and VEC is a vector with length N. The result | |
1028 | is stored in DEST which must be a vector of length M. */ | |
1029 | ||
1030 | static void | |
1031 | lambda_matrix_vector_mult (lambda_matrix matrix, int m, int n, | |
1032 | lambda_vector vec, lambda_vector dest) | |
1033 | { | |
1034 | int i, j; | |
1035 | ||
1036 | lambda_vector_clear (dest, m); | |
1037 | for (i = 0; i < m; i++) | |
1038 | for (j = 0; j < n; j++) | |
1039 | dest[i] += matrix[i][j] * vec[j]; | |
1040 | } | |
1041 | ||
1042 | /* Return true if TRANS is a legal transformation matrix that respects | |
1043 | the dependence vectors in DISTS and DIRS. The conservative answer | |
1044 | is false. | |
1045 | ||
1046 | "Wolfe proves that a unimodular transformation represented by the | |
1047 | matrix T is legal when applied to a loop nest with a set of | |
1048 | lexicographically non-negative distance vectors RDG if and only if | |
1049 | for each vector d in RDG, (T.d >= 0) is lexicographically positive. | |
1050 | i.e.: if and only if it transforms the lexicographically positive | |
1051 | distance vectors to lexicographically positive vectors. Note that | |
1052 | a unimodular matrix must transform the zero vector (and only it) to | |
1053 | the zero vector." S.Muchnick. */ | |
1054 | ||
1055 | static bool | |
1056 | lambda_transform_legal_p (lambda_trans_matrix trans, | |
1057 | int nb_loops, | |
1058 | vec<ddr_p> dependence_relations) | |
1059 | { | |
1060 | unsigned int i, j; | |
1061 | lambda_vector distres; | |
1062 | struct data_dependence_relation *ddr; | |
1063 | ||
1064 | gcc_assert (LTM_COLSIZE (trans) == nb_loops | |
1065 | && LTM_ROWSIZE (trans) == nb_loops); | |
1066 | ||
1067 | /* When there are no dependences, the transformation is correct. */ | |
1068 | if (dependence_relations.length () == 0) | |
1069 | return true; | |
1070 | ||
1071 | ddr = dependence_relations[0]; | |
1072 | if (ddr == NULL) | |
1073 | return true; | |
1074 | ||
1075 | /* When there is an unknown relation in the dependence_relations, we | |
1076 | know that it is no worth looking at this loop nest: give up. */ | |
1077 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
1078 | return false; | |
1079 | ||
1080 | distres = lambda_vector_new (nb_loops); | |
1081 | ||
1082 | /* For each distance vector in the dependence graph. */ | |
1083 | FOR_EACH_VEC_ELT (dependence_relations, i, ddr) | |
1084 | { | |
1085 | /* Don't care about relations for which we know that there is no | |
1086 | dependence, nor about read-read (aka. output-dependences): | |
1087 | these data accesses can happen in any order. */ | |
1088 | if (DDR_ARE_DEPENDENT (ddr) == chrec_known | |
1089 | || (DR_IS_READ (DDR_A (ddr)) && DR_IS_READ (DDR_B (ddr)))) | |
1090 | continue; | |
1091 | ||
1092 | /* Conservatively answer: "this transformation is not valid". */ | |
1093 | if (DDR_ARE_DEPENDENT (ddr) == chrec_dont_know) | |
1094 | return false; | |
1095 | ||
1096 | /* If the dependence could not be captured by a distance vector, | |
1097 | conservatively answer that the transform is not valid. */ | |
1098 | if (DDR_NUM_DIST_VECTS (ddr) == 0) | |
1099 | return false; | |
1100 | ||
1101 | /* Compute trans.dist_vect */ | |
1102 | for (j = 0; j < DDR_NUM_DIST_VECTS (ddr); j++) | |
1103 | { | |
1104 | lambda_matrix_vector_mult (LTM_MATRIX (trans), nb_loops, nb_loops, | |
1105 | DDR_DIST_VECT (ddr, j), distres); | |
1106 | ||
1107 | if (!lambda_vector_lexico_pos (distres, nb_loops)) | |
1108 | return false; | |
1109 | } | |
1110 | } | |
1111 | return true; | |
1112 | } | |
1113 | ||
1114 | /* Data dependency analysis. Returns true if the iterations of LOOP | |
1115 | are independent on each other (that is, if we can execute them | |
1116 | in parallel). */ | |
1117 | ||
1118 | static bool | |
1119 | loop_parallel_p (class loop *loop, struct obstack * parloop_obstack) | |
1120 | { | |
1121 | vec<ddr_p> dependence_relations; | |
1122 | vec<data_reference_p> datarefs; | |
1123 | lambda_trans_matrix trans; | |
1124 | bool ret = false; | |
1125 | ||
1126 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1127 | { | |
1128 | fprintf (dump_file, "Considering loop %d\n", loop->num); | |
1129 | if (!loop->inner) | |
1130 | fprintf (dump_file, "loop is innermost\n"); | |
1131 | else | |
1132 | fprintf (dump_file, "loop NOT innermost\n"); | |
1133 | } | |
1134 | ||
1135 | /* Check for problems with dependences. If the loop can be reversed, | |
1136 | the iterations are independent. */ | |
1137 | auto_vec<loop_p, 3> loop_nest; | |
1138 | datarefs.create (10); | |
1139 | dependence_relations.create (100); | |
1140 | if (! compute_data_dependences_for_loop (loop, true, &loop_nest, &datarefs, | |
1141 | &dependence_relations)) | |
1142 | { | |
1143 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1144 | fprintf (dump_file, " FAILED: cannot analyze data dependencies\n"); | |
1145 | ret = false; | |
1146 | goto end; | |
1147 | } | |
1148 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1149 | dump_data_dependence_relations (dump_file, dependence_relations); | |
1150 | ||
1151 | trans = lambda_trans_matrix_new (1, 1, parloop_obstack); | |
1152 | LTM_MATRIX (trans)[0][0] = -1; | |
1153 | ||
1154 | if (lambda_transform_legal_p (trans, 1, dependence_relations)) | |
1155 | { | |
1156 | ret = true; | |
1157 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1158 | fprintf (dump_file, " SUCCESS: may be parallelized\n"); | |
1159 | } | |
1160 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
1161 | fprintf (dump_file, | |
1162 | " FAILED: data dependencies exist across iterations\n"); | |
1163 | ||
1164 | end: | |
1165 | free_dependence_relations (dependence_relations); | |
1166 | free_data_refs (datarefs); | |
1167 | ||
1168 | return ret; | |
1169 | } | |
1170 | ||
1171 | /* Return true when LOOP contains basic blocks marked with the | |
1172 | BB_IRREDUCIBLE_LOOP flag. */ | |
1173 | ||
1174 | static inline bool | |
1175 | loop_has_blocks_with_irreducible_flag (class loop *loop) | |
1176 | { | |
1177 | unsigned i; | |
1178 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
1179 | bool res = true; | |
1180 | ||
1181 | for (i = 0; i < loop->num_nodes; i++) | |
1182 | if (bbs[i]->flags & BB_IRREDUCIBLE_LOOP) | |
1183 | goto end; | |
1184 | ||
1185 | res = false; | |
1186 | end: | |
1187 | free (bbs); | |
1188 | return res; | |
1189 | } | |
1190 | ||
1191 | /* Assigns the address of OBJ in TYPE to an ssa name, and returns this name. | |
1192 | The assignment statement is placed on edge ENTRY. DECL_ADDRESS maps decls | |
1193 | to their addresses that can be reused. The address of OBJ is known to | |
1194 | be invariant in the whole function. Other needed statements are placed | |
1195 | right before GSI. */ | |
1196 | ||
1197 | static tree | |
1198 | take_address_of (tree obj, tree type, edge entry, | |
1199 | int_tree_htab_type *decl_address, gimple_stmt_iterator *gsi) | |
1200 | { | |
1201 | int uid; | |
1202 | tree *var_p, name, addr; | |
1203 | gassign *stmt; | |
1204 | gimple_seq stmts; | |
1205 | ||
1206 | /* Since the address of OBJ is invariant, the trees may be shared. | |
1207 | Avoid rewriting unrelated parts of the code. */ | |
1208 | obj = unshare_expr (obj); | |
1209 | for (var_p = &obj; | |
1210 | handled_component_p (*var_p); | |
1211 | var_p = &TREE_OPERAND (*var_p, 0)) | |
1212 | continue; | |
1213 | ||
1214 | /* Canonicalize the access to base on a MEM_REF. */ | |
1215 | if (DECL_P (*var_p)) | |
1216 | *var_p = build_simple_mem_ref (build_fold_addr_expr (*var_p)); | |
1217 | ||
1218 | /* Assign a canonical SSA name to the address of the base decl used | |
1219 | in the address and share it for all accesses and addresses based | |
1220 | on it. */ | |
1221 | uid = DECL_UID (TREE_OPERAND (TREE_OPERAND (*var_p, 0), 0)); | |
1222 | int_tree_map elt; | |
1223 | elt.uid = uid; | |
1224 | int_tree_map *slot = decl_address->find_slot (elt, INSERT); | |
1225 | if (!slot->to) | |
1226 | { | |
1227 | if (gsi == NULL) | |
1228 | return NULL; | |
1229 | addr = TREE_OPERAND (*var_p, 0); | |
1230 | const char *obj_name | |
1231 | = get_name (TREE_OPERAND (TREE_OPERAND (*var_p, 0), 0)); | |
1232 | if (obj_name) | |
1233 | name = make_temp_ssa_name (TREE_TYPE (addr), NULL, obj_name); | |
1234 | else | |
1235 | name = make_ssa_name (TREE_TYPE (addr)); | |
1236 | stmt = gimple_build_assign (name, addr); | |
1237 | gsi_insert_on_edge_immediate (entry, stmt); | |
1238 | ||
1239 | slot->uid = uid; | |
1240 | slot->to = name; | |
1241 | } | |
1242 | else | |
1243 | name = slot->to; | |
1244 | ||
1245 | /* Express the address in terms of the canonical SSA name. */ | |
1246 | TREE_OPERAND (*var_p, 0) = name; | |
1247 | if (gsi == NULL) | |
1248 | return build_fold_addr_expr_with_type (obj, type); | |
1249 | ||
1250 | name = force_gimple_operand (build_addr (obj), | |
1251 | &stmts, true, NULL_TREE); | |
1252 | if (!gimple_seq_empty_p (stmts)) | |
1253 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); | |
1254 | ||
1255 | if (!useless_type_conversion_p (type, TREE_TYPE (name))) | |
1256 | { | |
1257 | name = force_gimple_operand (fold_convert (type, name), &stmts, true, | |
1258 | NULL_TREE); | |
1259 | if (!gimple_seq_empty_p (stmts)) | |
1260 | gsi_insert_seq_before (gsi, stmts, GSI_SAME_STMT); | |
1261 | } | |
1262 | ||
1263 | return name; | |
1264 | } | |
1265 | ||
1266 | static tree | |
1267 | reduc_stmt_res (gimple *stmt) | |
1268 | { | |
1269 | return (gimple_code (stmt) == GIMPLE_PHI | |
1270 | ? gimple_phi_result (stmt) | |
1271 | : gimple_assign_lhs (stmt)); | |
1272 | } | |
1273 | ||
1274 | /* Callback for htab_traverse. Create the initialization statement | |
1275 | for reduction described in SLOT, and place it at the preheader of | |
1276 | the loop described in DATA. */ | |
1277 | ||
1278 | int | |
1279 | initialize_reductions (reduction_info **slot, class loop *loop) | |
1280 | { | |
1281 | tree init; | |
1282 | tree type, arg; | |
1283 | edge e; | |
1284 | ||
1285 | struct reduction_info *const reduc = *slot; | |
1286 | ||
1287 | /* Create initialization in preheader: | |
1288 | reduction_variable = initialization value of reduction. */ | |
1289 | ||
1290 | /* In the phi node at the header, replace the argument coming | |
1291 | from the preheader with the reduction initialization value. */ | |
1292 | ||
1293 | /* Initialize the reduction. */ | |
1294 | type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); | |
1295 | init = omp_reduction_init_op (gimple_location (reduc->reduc_stmt), | |
1296 | reduc->reduction_code, type); | |
1297 | reduc->init = init; | |
1298 | ||
1299 | /* Replace the argument representing the initialization value | |
1300 | with the initialization value for the reduction (neutral | |
1301 | element for the particular operation, e.g. 0 for PLUS_EXPR, | |
1302 | 1 for MULT_EXPR, etc). | |
1303 | Keep the old value in a new variable "reduction_initial", | |
1304 | that will be taken in consideration after the parallel | |
1305 | computing is done. */ | |
1306 | ||
1307 | e = loop_preheader_edge (loop); | |
1308 | arg = PHI_ARG_DEF_FROM_EDGE (reduc->reduc_phi, e); | |
1309 | /* Create new variable to hold the initial value. */ | |
1310 | ||
1311 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE | |
1312 | (reduc->reduc_phi, loop_preheader_edge (loop)), init); | |
1313 | reduc->initial_value = arg; | |
1314 | return 1; | |
1315 | } | |
1316 | ||
1317 | struct elv_data | |
1318 | { | |
1319 | struct walk_stmt_info info; | |
1320 | edge entry; | |
1321 | int_tree_htab_type *decl_address; | |
1322 | gimple_stmt_iterator *gsi; | |
1323 | bool changed; | |
1324 | bool reset; | |
1325 | }; | |
1326 | ||
1327 | /* Eliminates references to local variables in *TP out of the single | |
1328 | entry single exit region starting at DTA->ENTRY. | |
1329 | DECL_ADDRESS contains addresses of the references that had their | |
1330 | address taken already. If the expression is changed, CHANGED is | |
1331 | set to true. Callback for walk_tree. */ | |
1332 | ||
1333 | static tree | |
1334 | eliminate_local_variables_1 (tree *tp, int *walk_subtrees, void *data) | |
1335 | { | |
1336 | struct elv_data *const dta = (struct elv_data *) data; | |
1337 | tree t = *tp, var, addr, addr_type, type, obj; | |
1338 | ||
1339 | if (DECL_P (t)) | |
1340 | { | |
1341 | *walk_subtrees = 0; | |
1342 | ||
1343 | if (!SSA_VAR_P (t) || DECL_EXTERNAL (t)) | |
1344 | return NULL_TREE; | |
1345 | ||
1346 | type = TREE_TYPE (t); | |
1347 | addr_type = build_pointer_type (type); | |
1348 | addr = take_address_of (t, addr_type, dta->entry, dta->decl_address, | |
1349 | dta->gsi); | |
1350 | if (dta->gsi == NULL && addr == NULL_TREE) | |
1351 | { | |
1352 | dta->reset = true; | |
1353 | return NULL_TREE; | |
1354 | } | |
1355 | ||
1356 | *tp = build_simple_mem_ref (addr); | |
1357 | ||
1358 | dta->changed = true; | |
1359 | return NULL_TREE; | |
1360 | } | |
1361 | ||
1362 | if (TREE_CODE (t) == ADDR_EXPR) | |
1363 | { | |
1364 | /* ADDR_EXPR may appear in two contexts: | |
1365 | -- as a gimple operand, when the address taken is a function invariant | |
1366 | -- as gimple rhs, when the resulting address in not a function | |
1367 | invariant | |
1368 | We do not need to do anything special in the latter case (the base of | |
1369 | the memory reference whose address is taken may be replaced in the | |
1370 | DECL_P case). The former case is more complicated, as we need to | |
1371 | ensure that the new address is still a gimple operand. Thus, it | |
1372 | is not sufficient to replace just the base of the memory reference -- | |
1373 | we need to move the whole computation of the address out of the | |
1374 | loop. */ | |
1375 | if (!is_gimple_val (t)) | |
1376 | return NULL_TREE; | |
1377 | ||
1378 | *walk_subtrees = 0; | |
1379 | obj = TREE_OPERAND (t, 0); | |
1380 | var = get_base_address (obj); | |
1381 | if (!var || !SSA_VAR_P (var) || DECL_EXTERNAL (var)) | |
1382 | return NULL_TREE; | |
1383 | ||
1384 | addr_type = TREE_TYPE (t); | |
1385 | addr = take_address_of (obj, addr_type, dta->entry, dta->decl_address, | |
1386 | dta->gsi); | |
1387 | if (dta->gsi == NULL && addr == NULL_TREE) | |
1388 | { | |
1389 | dta->reset = true; | |
1390 | return NULL_TREE; | |
1391 | } | |
1392 | *tp = addr; | |
1393 | ||
1394 | dta->changed = true; | |
1395 | return NULL_TREE; | |
1396 | } | |
1397 | ||
1398 | if (!EXPR_P (t)) | |
1399 | *walk_subtrees = 0; | |
1400 | ||
1401 | return NULL_TREE; | |
1402 | } | |
1403 | ||
1404 | /* Moves the references to local variables in STMT at *GSI out of the single | |
1405 | entry single exit region starting at ENTRY. DECL_ADDRESS contains | |
1406 | addresses of the references that had their address taken | |
1407 | already. */ | |
1408 | ||
1409 | static void | |
1410 | eliminate_local_variables_stmt (edge entry, gimple_stmt_iterator *gsi, | |
1411 | int_tree_htab_type *decl_address) | |
1412 | { | |
1413 | struct elv_data dta; | |
1414 | gimple *stmt = gsi_stmt (*gsi); | |
1415 | ||
1416 | memset (&dta.info, '\0', sizeof (dta.info)); | |
1417 | dta.entry = entry; | |
1418 | dta.decl_address = decl_address; | |
1419 | dta.changed = false; | |
1420 | dta.reset = false; | |
1421 | ||
1422 | if (gimple_debug_bind_p (stmt)) | |
1423 | { | |
1424 | dta.gsi = NULL; | |
1425 | walk_tree (gimple_debug_bind_get_value_ptr (stmt), | |
1426 | eliminate_local_variables_1, &dta.info, NULL); | |
1427 | if (dta.reset) | |
1428 | { | |
1429 | gimple_debug_bind_reset_value (stmt); | |
1430 | dta.changed = true; | |
1431 | } | |
1432 | } | |
1433 | else if (gimple_clobber_p (stmt)) | |
1434 | { | |
1435 | unlink_stmt_vdef (stmt); | |
1436 | stmt = gimple_build_nop (); | |
1437 | gsi_replace (gsi, stmt, false); | |
1438 | dta.changed = true; | |
1439 | } | |
1440 | else | |
1441 | { | |
1442 | dta.gsi = gsi; | |
1443 | walk_gimple_op (stmt, eliminate_local_variables_1, &dta.info); | |
1444 | } | |
1445 | ||
1446 | if (dta.changed) | |
1447 | update_stmt (stmt); | |
1448 | } | |
1449 | ||
1450 | /* Eliminates the references to local variables from the single entry | |
1451 | single exit region between the ENTRY and EXIT edges. | |
1452 | ||
1453 | This includes: | |
1454 | 1) Taking address of a local variable -- these are moved out of the | |
1455 | region (and temporary variable is created to hold the address if | |
1456 | necessary). | |
1457 | ||
1458 | 2) Dereferencing a local variable -- these are replaced with indirect | |
1459 | references. */ | |
1460 | ||
1461 | static void | |
1462 | eliminate_local_variables (edge entry, edge exit) | |
1463 | { | |
1464 | basic_block bb; | |
1465 | auto_vec<basic_block, 3> body; | |
1466 | unsigned i; | |
1467 | gimple_stmt_iterator gsi; | |
1468 | bool has_debug_stmt = false; | |
1469 | int_tree_htab_type decl_address (10); | |
1470 | basic_block entry_bb = entry->src; | |
1471 | basic_block exit_bb = exit->dest; | |
1472 | ||
1473 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); | |
1474 | ||
1475 | FOR_EACH_VEC_ELT (body, i, bb) | |
1476 | if (bb != entry_bb && bb != exit_bb) | |
1477 | { | |
1478 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1479 | if (is_gimple_debug (gsi_stmt (gsi))) | |
1480 | { | |
1481 | if (gimple_debug_bind_p (gsi_stmt (gsi))) | |
1482 | has_debug_stmt = true; | |
1483 | } | |
1484 | else | |
1485 | eliminate_local_variables_stmt (entry, &gsi, &decl_address); | |
1486 | } | |
1487 | ||
1488 | if (has_debug_stmt) | |
1489 | FOR_EACH_VEC_ELT (body, i, bb) | |
1490 | if (bb != entry_bb && bb != exit_bb) | |
1491 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1492 | if (gimple_debug_bind_p (gsi_stmt (gsi))) | |
1493 | eliminate_local_variables_stmt (entry, &gsi, &decl_address); | |
1494 | } | |
1495 | ||
1496 | /* Returns true if expression EXPR is not defined between ENTRY and | |
1497 | EXIT, i.e. if all its operands are defined outside of the region. */ | |
1498 | ||
1499 | static bool | |
1500 | expr_invariant_in_region_p (edge entry, edge exit, tree expr) | |
1501 | { | |
1502 | basic_block entry_bb = entry->src; | |
1503 | basic_block exit_bb = exit->dest; | |
1504 | basic_block def_bb; | |
1505 | ||
1506 | if (is_gimple_min_invariant (expr)) | |
1507 | return true; | |
1508 | ||
1509 | if (TREE_CODE (expr) == SSA_NAME) | |
1510 | { | |
1511 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); | |
1512 | if (def_bb | |
1513 | && dominated_by_p (CDI_DOMINATORS, def_bb, entry_bb) | |
1514 | && !dominated_by_p (CDI_DOMINATORS, def_bb, exit_bb)) | |
1515 | return false; | |
1516 | ||
1517 | return true; | |
1518 | } | |
1519 | ||
1520 | return false; | |
1521 | } | |
1522 | ||
1523 | /* If COPY_NAME_P is true, creates and returns a duplicate of NAME. | |
1524 | The copies are stored to NAME_COPIES, if NAME was already duplicated, | |
1525 | its duplicate stored in NAME_COPIES is returned. | |
1526 | ||
1527 | Regardless of COPY_NAME_P, the decl used as a base of the ssa name is also | |
1528 | duplicated, storing the copies in DECL_COPIES. */ | |
1529 | ||
1530 | static tree | |
1531 | separate_decls_in_region_name (tree name, name_to_copy_table_type *name_copies, | |
1532 | int_tree_htab_type *decl_copies, | |
1533 | bool copy_name_p) | |
1534 | { | |
1535 | tree copy, var, var_copy; | |
1536 | unsigned idx, uid, nuid; | |
1537 | struct int_tree_map ielt; | |
1538 | struct name_to_copy_elt elt, *nelt; | |
1539 | name_to_copy_elt **slot; | |
1540 | int_tree_map *dslot; | |
1541 | ||
1542 | if (TREE_CODE (name) != SSA_NAME) | |
1543 | return name; | |
1544 | ||
1545 | idx = SSA_NAME_VERSION (name); | |
1546 | elt.version = idx; | |
1547 | slot = name_copies->find_slot_with_hash (&elt, idx, | |
1548 | copy_name_p ? INSERT : NO_INSERT); | |
1549 | if (slot && *slot) | |
1550 | return (*slot)->new_name; | |
1551 | ||
1552 | if (copy_name_p) | |
1553 | { | |
1554 | copy = duplicate_ssa_name (name, NULL); | |
1555 | nelt = XNEW (struct name_to_copy_elt); | |
1556 | nelt->version = idx; | |
1557 | nelt->new_name = copy; | |
1558 | nelt->field = NULL_TREE; | |
1559 | *slot = nelt; | |
1560 | } | |
1561 | else | |
1562 | { | |
1563 | gcc_assert (!slot); | |
1564 | copy = name; | |
1565 | } | |
1566 | ||
1567 | var = SSA_NAME_VAR (name); | |
1568 | if (!var) | |
1569 | return copy; | |
1570 | ||
1571 | uid = DECL_UID (var); | |
1572 | ielt.uid = uid; | |
1573 | dslot = decl_copies->find_slot_with_hash (ielt, uid, INSERT); | |
1574 | if (!dslot->to) | |
1575 | { | |
1576 | var_copy = create_tmp_var (TREE_TYPE (var), get_name (var)); | |
1577 | DECL_NOT_GIMPLE_REG_P (var_copy) = DECL_NOT_GIMPLE_REG_P (var); | |
1578 | dslot->uid = uid; | |
1579 | dslot->to = var_copy; | |
1580 | ||
1581 | /* Ensure that when we meet this decl next time, we won't duplicate | |
1582 | it again. */ | |
1583 | nuid = DECL_UID (var_copy); | |
1584 | ielt.uid = nuid; | |
1585 | dslot = decl_copies->find_slot_with_hash (ielt, nuid, INSERT); | |
1586 | gcc_assert (!dslot->to); | |
1587 | dslot->uid = nuid; | |
1588 | dslot->to = var_copy; | |
1589 | } | |
1590 | else | |
1591 | var_copy = dslot->to; | |
1592 | ||
1593 | replace_ssa_name_symbol (copy, var_copy); | |
1594 | return copy; | |
1595 | } | |
1596 | ||
1597 | /* Finds the ssa names used in STMT that are defined outside the | |
1598 | region between ENTRY and EXIT and replaces such ssa names with | |
1599 | their duplicates. The duplicates are stored to NAME_COPIES. Base | |
1600 | decls of all ssa names used in STMT (including those defined in | |
1601 | LOOP) are replaced with the new temporary variables; the | |
1602 | replacement decls are stored in DECL_COPIES. */ | |
1603 | ||
1604 | static void | |
1605 | separate_decls_in_region_stmt (edge entry, edge exit, gimple *stmt, | |
1606 | name_to_copy_table_type *name_copies, | |
1607 | int_tree_htab_type *decl_copies) | |
1608 | { | |
1609 | use_operand_p use; | |
1610 | def_operand_p def; | |
1611 | ssa_op_iter oi; | |
1612 | tree name, copy; | |
1613 | bool copy_name_p; | |
1614 | ||
1615 | FOR_EACH_PHI_OR_STMT_DEF (def, stmt, oi, SSA_OP_DEF) | |
1616 | { | |
1617 | name = DEF_FROM_PTR (def); | |
1618 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
1619 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, | |
1620 | false); | |
1621 | gcc_assert (copy == name); | |
1622 | } | |
1623 | ||
1624 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
1625 | { | |
1626 | name = USE_FROM_PTR (use); | |
1627 | if (TREE_CODE (name) != SSA_NAME) | |
1628 | continue; | |
1629 | ||
1630 | copy_name_p = expr_invariant_in_region_p (entry, exit, name); | |
1631 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, | |
1632 | copy_name_p); | |
1633 | SET_USE (use, copy); | |
1634 | } | |
1635 | } | |
1636 | ||
1637 | /* Finds the ssa names used in STMT that are defined outside the | |
1638 | region between ENTRY and EXIT and replaces such ssa names with | |
1639 | their duplicates. The duplicates are stored to NAME_COPIES. Base | |
1640 | decls of all ssa names used in STMT (including those defined in | |
1641 | LOOP) are replaced with the new temporary variables; the | |
1642 | replacement decls are stored in DECL_COPIES. */ | |
1643 | ||
1644 | static bool | |
1645 | separate_decls_in_region_debug (gimple *stmt, | |
1646 | name_to_copy_table_type *name_copies, | |
1647 | int_tree_htab_type *decl_copies) | |
1648 | { | |
1649 | use_operand_p use; | |
1650 | ssa_op_iter oi; | |
1651 | tree var, name; | |
1652 | struct int_tree_map ielt; | |
1653 | struct name_to_copy_elt elt; | |
1654 | name_to_copy_elt **slot; | |
1655 | int_tree_map *dslot; | |
1656 | ||
1657 | if (gimple_debug_bind_p (stmt)) | |
1658 | var = gimple_debug_bind_get_var (stmt); | |
1659 | else if (gimple_debug_source_bind_p (stmt)) | |
1660 | var = gimple_debug_source_bind_get_var (stmt); | |
1661 | else | |
1662 | return true; | |
1663 | if (TREE_CODE (var) == DEBUG_EXPR_DECL || TREE_CODE (var) == LABEL_DECL) | |
1664 | return true; | |
1665 | gcc_assert (DECL_P (var) && SSA_VAR_P (var)); | |
1666 | ielt.uid = DECL_UID (var); | |
1667 | dslot = decl_copies->find_slot_with_hash (ielt, ielt.uid, NO_INSERT); | |
1668 | if (!dslot) | |
1669 | return true; | |
1670 | if (gimple_debug_bind_p (stmt)) | |
1671 | gimple_debug_bind_set_var (stmt, dslot->to); | |
1672 | else if (gimple_debug_source_bind_p (stmt)) | |
1673 | gimple_debug_source_bind_set_var (stmt, dslot->to); | |
1674 | ||
1675 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
1676 | { | |
1677 | name = USE_FROM_PTR (use); | |
1678 | if (TREE_CODE (name) != SSA_NAME) | |
1679 | continue; | |
1680 | ||
1681 | elt.version = SSA_NAME_VERSION (name); | |
1682 | slot = name_copies->find_slot_with_hash (&elt, elt.version, NO_INSERT); | |
1683 | if (!slot) | |
1684 | { | |
1685 | gimple_debug_bind_reset_value (stmt); | |
1686 | update_stmt (stmt); | |
1687 | break; | |
1688 | } | |
1689 | ||
1690 | SET_USE (use, (*slot)->new_name); | |
1691 | } | |
1692 | ||
1693 | return false; | |
1694 | } | |
1695 | ||
1696 | /* Callback for htab_traverse. Adds a field corresponding to the reduction | |
1697 | specified in SLOT. The type is passed in DATA. */ | |
1698 | ||
1699 | int | |
1700 | add_field_for_reduction (reduction_info **slot, tree type) | |
1701 | { | |
1702 | ||
1703 | struct reduction_info *const red = *slot; | |
1704 | tree var = reduc_stmt_res (red->reduc_stmt); | |
1705 | tree field = build_decl (gimple_location (red->reduc_stmt), FIELD_DECL, | |
1706 | SSA_NAME_IDENTIFIER (var), TREE_TYPE (var)); | |
1707 | ||
1708 | insert_field_into_struct (type, field); | |
1709 | ||
1710 | red->field = field; | |
1711 | ||
1712 | return 1; | |
1713 | } | |
1714 | ||
1715 | /* Callback for htab_traverse. Adds a field corresponding to a ssa name | |
1716 | described in SLOT. The type is passed in DATA. */ | |
1717 | ||
1718 | int | |
1719 | add_field_for_name (name_to_copy_elt **slot, tree type) | |
1720 | { | |
1721 | struct name_to_copy_elt *const elt = *slot; | |
1722 | tree name = ssa_name (elt->version); | |
1723 | tree field = build_decl (UNKNOWN_LOCATION, | |
1724 | FIELD_DECL, SSA_NAME_IDENTIFIER (name), | |
1725 | TREE_TYPE (name)); | |
1726 | ||
1727 | insert_field_into_struct (type, field); | |
1728 | elt->field = field; | |
1729 | ||
1730 | return 1; | |
1731 | } | |
1732 | ||
1733 | /* Callback for htab_traverse. A local result is the intermediate result | |
1734 | computed by a single | |
1735 | thread, or the initial value in case no iteration was executed. | |
1736 | This function creates a phi node reflecting these values. | |
1737 | The phi's result will be stored in NEW_PHI field of the | |
1738 | reduction's data structure. */ | |
1739 | ||
1740 | int | |
1741 | create_phi_for_local_result (reduction_info **slot, class loop *loop) | |
1742 | { | |
1743 | struct reduction_info *const reduc = *slot; | |
1744 | edge e; | |
1745 | gphi *new_phi; | |
1746 | basic_block store_bb, continue_bb; | |
1747 | tree local_res; | |
1748 | location_t locus; | |
1749 | ||
1750 | /* STORE_BB is the block where the phi | |
1751 | should be stored. It is the destination of the loop exit. | |
1752 | (Find the fallthru edge from GIMPLE_OMP_CONTINUE). */ | |
1753 | continue_bb = single_pred (loop->latch); | |
1754 | store_bb = FALLTHRU_EDGE (continue_bb)->dest; | |
1755 | ||
1756 | /* STORE_BB has two predecessors. One coming from the loop | |
1757 | (the reduction's result is computed at the loop), | |
1758 | and another coming from a block preceding the loop, | |
1759 | when no iterations | |
1760 | are executed (the initial value should be taken). */ | |
1761 | if (EDGE_PRED (store_bb, 0) == FALLTHRU_EDGE (continue_bb)) | |
1762 | e = EDGE_PRED (store_bb, 1); | |
1763 | else | |
1764 | e = EDGE_PRED (store_bb, 0); | |
1765 | tree lhs = reduc_stmt_res (reduc->reduc_stmt); | |
1766 | local_res = copy_ssa_name (lhs); | |
1767 | locus = gimple_location (reduc->reduc_stmt); | |
1768 | new_phi = create_phi_node (local_res, store_bb); | |
1769 | add_phi_arg (new_phi, reduc->init, e, locus); | |
1770 | add_phi_arg (new_phi, lhs, FALLTHRU_EDGE (continue_bb), locus); | |
1771 | reduc->new_phi = new_phi; | |
1772 | ||
1773 | return 1; | |
1774 | } | |
1775 | ||
1776 | struct clsn_data | |
1777 | { | |
1778 | tree store; | |
1779 | tree load; | |
1780 | ||
1781 | basic_block store_bb; | |
1782 | basic_block load_bb; | |
1783 | }; | |
1784 | ||
1785 | /* Callback for htab_traverse. Create an atomic instruction for the | |
1786 | reduction described in SLOT. | |
1787 | DATA annotates the place in memory the atomic operation relates to, | |
1788 | and the basic block it needs to be generated in. */ | |
1789 | ||
1790 | int | |
1791 | create_call_for_reduction_1 (reduction_info **slot, struct clsn_data *clsn_data) | |
1792 | { | |
1793 | struct reduction_info *const reduc = *slot; | |
1794 | gimple_stmt_iterator gsi; | |
1795 | tree type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); | |
1796 | tree load_struct; | |
1797 | basic_block bb; | |
1798 | basic_block new_bb; | |
1799 | edge e; | |
1800 | tree t, addr, ref, x; | |
1801 | tree tmp_load, name; | |
1802 | gimple *load; | |
1803 | ||
1804 | if (reduc->reduc_addr == NULL_TREE) | |
1805 | { | |
1806 | load_struct = build_simple_mem_ref (clsn_data->load); | |
1807 | t = build3 (COMPONENT_REF, type, load_struct, reduc->field, NULL_TREE); | |
1808 | ||
1809 | addr = build_addr (t); | |
1810 | } | |
1811 | else | |
1812 | { | |
1813 | /* Set the address for the atomic store. */ | |
1814 | addr = reduc->reduc_addr; | |
1815 | ||
1816 | /* Remove the non-atomic store '*addr = sum'. */ | |
1817 | tree res = PHI_RESULT (reduc->keep_res); | |
1818 | use_operand_p use_p; | |
1819 | gimple *stmt; | |
1820 | bool single_use_p = single_imm_use (res, &use_p, &stmt); | |
1821 | gcc_assert (single_use_p); | |
1822 | replace_uses_by (gimple_vdef (stmt), | |
1823 | gimple_vuse (stmt)); | |
1824 | gimple_stmt_iterator gsi = gsi_for_stmt (stmt); | |
1825 | gsi_remove (&gsi, true); | |
1826 | } | |
1827 | ||
1828 | /* Create phi node. */ | |
1829 | bb = clsn_data->load_bb; | |
1830 | ||
1831 | gsi = gsi_last_bb (bb); | |
1832 | e = split_block (bb, gsi_stmt (gsi)); | |
1833 | new_bb = e->dest; | |
1834 | ||
1835 | tmp_load = create_tmp_var (TREE_TYPE (TREE_TYPE (addr))); | |
1836 | tmp_load = make_ssa_name (tmp_load); | |
1837 | load = gimple_build_omp_atomic_load (tmp_load, addr, | |
1838 | OMP_MEMORY_ORDER_RELAXED); | |
1839 | SSA_NAME_DEF_STMT (tmp_load) = load; | |
1840 | gsi = gsi_start_bb (new_bb); | |
1841 | gsi_insert_after (&gsi, load, GSI_NEW_STMT); | |
1842 | ||
1843 | e = split_block (new_bb, load); | |
1844 | new_bb = e->dest; | |
1845 | gsi = gsi_start_bb (new_bb); | |
1846 | ref = tmp_load; | |
1847 | x = fold_build2 (reduc->reduction_code, | |
1848 | TREE_TYPE (PHI_RESULT (reduc->new_phi)), ref, | |
1849 | PHI_RESULT (reduc->new_phi)); | |
1850 | ||
1851 | name = force_gimple_operand_gsi (&gsi, x, true, NULL_TREE, true, | |
1852 | GSI_CONTINUE_LINKING); | |
1853 | ||
1854 | gimple *store = gimple_build_omp_atomic_store (name, | |
1855 | OMP_MEMORY_ORDER_RELAXED); | |
1856 | gsi_insert_after (&gsi, store, GSI_NEW_STMT); | |
1857 | return 1; | |
1858 | } | |
1859 | ||
1860 | /* Create the atomic operation at the join point of the threads. | |
1861 | REDUCTION_LIST describes the reductions in the LOOP. | |
1862 | LD_ST_DATA describes the shared data structure where | |
1863 | shared data is stored in and loaded from. */ | |
1864 | static void | |
1865 | create_call_for_reduction (class loop *loop, | |
1866 | reduction_info_table_type *reduction_list, | |
1867 | struct clsn_data *ld_st_data) | |
1868 | { | |
1869 | reduction_list->traverse <class loop *, create_phi_for_local_result> (loop); | |
1870 | /* Find the fallthru edge from GIMPLE_OMP_CONTINUE. */ | |
1871 | basic_block continue_bb = single_pred (loop->latch); | |
1872 | ld_st_data->load_bb = FALLTHRU_EDGE (continue_bb)->dest; | |
1873 | reduction_list | |
1874 | ->traverse <struct clsn_data *, create_call_for_reduction_1> (ld_st_data); | |
1875 | } | |
1876 | ||
1877 | /* Callback for htab_traverse. Loads the final reduction value at the | |
1878 | join point of all threads, and inserts it in the right place. */ | |
1879 | ||
1880 | int | |
1881 | create_loads_for_reductions (reduction_info **slot, struct clsn_data *clsn_data) | |
1882 | { | |
1883 | struct reduction_info *const red = *slot; | |
1884 | gimple *stmt; | |
1885 | gimple_stmt_iterator gsi; | |
1886 | tree type = TREE_TYPE (reduc_stmt_res (red->reduc_stmt)); | |
1887 | tree load_struct; | |
1888 | tree name; | |
1889 | tree x; | |
1890 | ||
1891 | /* If there's no exit phi, the result of the reduction is unused. */ | |
1892 | if (red->keep_res == NULL) | |
1893 | return 1; | |
1894 | ||
1895 | gsi = gsi_after_labels (clsn_data->load_bb); | |
1896 | load_struct = build_simple_mem_ref (clsn_data->load); | |
1897 | load_struct = build3 (COMPONENT_REF, type, load_struct, red->field, | |
1898 | NULL_TREE); | |
1899 | ||
1900 | x = load_struct; | |
1901 | name = PHI_RESULT (red->keep_res); | |
1902 | stmt = gimple_build_assign (name, x); | |
1903 | ||
1904 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1905 | ||
1906 | for (gsi = gsi_start_phis (gimple_bb (red->keep_res)); | |
1907 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
1908 | if (gsi_stmt (gsi) == red->keep_res) | |
1909 | { | |
1910 | remove_phi_node (&gsi, false); | |
1911 | return 1; | |
1912 | } | |
1913 | gcc_unreachable (); | |
1914 | } | |
1915 | ||
1916 | /* Load the reduction result that was stored in LD_ST_DATA. | |
1917 | REDUCTION_LIST describes the list of reductions that the | |
1918 | loads should be generated for. */ | |
1919 | static void | |
1920 | create_final_loads_for_reduction (reduction_info_table_type *reduction_list, | |
1921 | struct clsn_data *ld_st_data) | |
1922 | { | |
1923 | gimple_stmt_iterator gsi; | |
1924 | tree t; | |
1925 | gimple *stmt; | |
1926 | ||
1927 | gsi = gsi_after_labels (ld_st_data->load_bb); | |
1928 | t = build_fold_addr_expr (ld_st_data->store); | |
1929 | stmt = gimple_build_assign (ld_st_data->load, t); | |
1930 | ||
1931 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
1932 | ||
1933 | reduction_list | |
1934 | ->traverse <struct clsn_data *, create_loads_for_reductions> (ld_st_data); | |
1935 | ||
1936 | } | |
1937 | ||
1938 | /* Callback for htab_traverse. Store the neutral value for the | |
1939 | particular reduction's operation, e.g. 0 for PLUS_EXPR, | |
1940 | 1 for MULT_EXPR, etc. into the reduction field. | |
1941 | The reduction is specified in SLOT. The store information is | |
1942 | passed in DATA. */ | |
1943 | ||
1944 | int | |
1945 | create_stores_for_reduction (reduction_info **slot, struct clsn_data *clsn_data) | |
1946 | { | |
1947 | struct reduction_info *const red = *slot; | |
1948 | tree t; | |
1949 | gimple *stmt; | |
1950 | gimple_stmt_iterator gsi; | |
1951 | tree type = TREE_TYPE (reduc_stmt_res (red->reduc_stmt)); | |
1952 | ||
1953 | gsi = gsi_last_bb (clsn_data->store_bb); | |
1954 | t = build3 (COMPONENT_REF, type, clsn_data->store, red->field, NULL_TREE); | |
1955 | stmt = gimple_build_assign (t, red->initial_value); | |
1956 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1957 | ||
1958 | return 1; | |
1959 | } | |
1960 | ||
1961 | /* Callback for htab_traverse. Creates loads to a field of LOAD in LOAD_BB and | |
1962 | store to a field of STORE in STORE_BB for the ssa name and its duplicate | |
1963 | specified in SLOT. */ | |
1964 | ||
1965 | int | |
1966 | create_loads_and_stores_for_name (name_to_copy_elt **slot, | |
1967 | struct clsn_data *clsn_data) | |
1968 | { | |
1969 | struct name_to_copy_elt *const elt = *slot; | |
1970 | tree t; | |
1971 | gimple *stmt; | |
1972 | gimple_stmt_iterator gsi; | |
1973 | tree type = TREE_TYPE (elt->new_name); | |
1974 | tree load_struct; | |
1975 | ||
1976 | gsi = gsi_last_bb (clsn_data->store_bb); | |
1977 | t = build3 (COMPONENT_REF, type, clsn_data->store, elt->field, NULL_TREE); | |
1978 | stmt = gimple_build_assign (t, ssa_name (elt->version)); | |
1979 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1980 | ||
1981 | gsi = gsi_last_bb (clsn_data->load_bb); | |
1982 | load_struct = build_simple_mem_ref (clsn_data->load); | |
1983 | t = build3 (COMPONENT_REF, type, load_struct, elt->field, NULL_TREE); | |
1984 | stmt = gimple_build_assign (elt->new_name, t); | |
1985 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1986 | ||
1987 | return 1; | |
1988 | } | |
1989 | ||
1990 | /* Moves all the variables used in LOOP and defined outside of it (including | |
1991 | the initial values of loop phi nodes, and *PER_THREAD if it is a ssa | |
1992 | name) to a structure created for this purpose. The code | |
1993 | ||
1994 | while (1) | |
1995 | { | |
1996 | use (a); | |
1997 | use (b); | |
1998 | } | |
1999 | ||
2000 | is transformed this way: | |
2001 | ||
2002 | bb0: | |
2003 | old.a = a; | |
2004 | old.b = b; | |
2005 | ||
2006 | bb1: | |
2007 | a' = new->a; | |
2008 | b' = new->b; | |
2009 | while (1) | |
2010 | { | |
2011 | use (a'); | |
2012 | use (b'); | |
2013 | } | |
2014 | ||
2015 | `old' is stored to *ARG_STRUCT and `new' is stored to NEW_ARG_STRUCT. The | |
2016 | pointer `new' is intentionally not initialized (the loop will be split to a | |
2017 | separate function later, and `new' will be initialized from its arguments). | |
2018 | LD_ST_DATA holds information about the shared data structure used to pass | |
2019 | information among the threads. It is initialized here, and | |
2020 | gen_parallel_loop will pass it to create_call_for_reduction that | |
2021 | needs this information. REDUCTION_LIST describes the reductions | |
2022 | in LOOP. */ | |
2023 | ||
2024 | static void | |
2025 | separate_decls_in_region (edge entry, edge exit, | |
2026 | reduction_info_table_type *reduction_list, | |
2027 | tree *arg_struct, tree *new_arg_struct, | |
2028 | struct clsn_data *ld_st_data) | |
2029 | ||
2030 | { | |
2031 | basic_block bb1 = split_edge (entry); | |
2032 | basic_block bb0 = single_pred (bb1); | |
2033 | name_to_copy_table_type name_copies (10); | |
2034 | int_tree_htab_type decl_copies (10); | |
2035 | unsigned i; | |
2036 | tree type, type_name, nvar; | |
2037 | gimple_stmt_iterator gsi; | |
2038 | struct clsn_data clsn_data; | |
2039 | auto_vec<basic_block, 3> body; | |
2040 | basic_block bb; | |
2041 | basic_block entry_bb = bb1; | |
2042 | basic_block exit_bb = exit->dest; | |
2043 | bool has_debug_stmt = false; | |
2044 | ||
2045 | entry = single_succ_edge (entry_bb); | |
2046 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); | |
2047 | ||
2048 | FOR_EACH_VEC_ELT (body, i, bb) | |
2049 | { | |
2050 | if (bb != entry_bb && bb != exit_bb) | |
2051 | { | |
2052 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2053 | separate_decls_in_region_stmt (entry, exit, gsi_stmt (gsi), | |
2054 | &name_copies, &decl_copies); | |
2055 | ||
2056 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2057 | { | |
2058 | gimple *stmt = gsi_stmt (gsi); | |
2059 | ||
2060 | if (is_gimple_debug (stmt)) | |
2061 | has_debug_stmt = true; | |
2062 | else | |
2063 | separate_decls_in_region_stmt (entry, exit, stmt, | |
2064 | &name_copies, &decl_copies); | |
2065 | } | |
2066 | } | |
2067 | } | |
2068 | ||
2069 | /* Now process debug bind stmts. We must not create decls while | |
2070 | processing debug stmts, so we defer their processing so as to | |
2071 | make sure we will have debug info for as many variables as | |
2072 | possible (all of those that were dealt with in the loop above), | |
2073 | and discard those for which we know there's nothing we can | |
2074 | do. */ | |
2075 | if (has_debug_stmt) | |
2076 | FOR_EACH_VEC_ELT (body, i, bb) | |
2077 | if (bb != entry_bb && bb != exit_bb) | |
2078 | { | |
2079 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) | |
2080 | { | |
2081 | gimple *stmt = gsi_stmt (gsi); | |
2082 | ||
2083 | if (is_gimple_debug (stmt)) | |
2084 | { | |
2085 | if (separate_decls_in_region_debug (stmt, &name_copies, | |
2086 | &decl_copies)) | |
2087 | { | |
2088 | gsi_remove (&gsi, true); | |
2089 | continue; | |
2090 | } | |
2091 | } | |
2092 | ||
2093 | gsi_next (&gsi); | |
2094 | } | |
2095 | } | |
2096 | ||
2097 | if (name_copies.is_empty () && reduction_list->is_empty ()) | |
2098 | { | |
2099 | /* It may happen that there is nothing to copy (if there are only | |
2100 | loop carried and external variables in the loop). */ | |
2101 | *arg_struct = NULL; | |
2102 | *new_arg_struct = NULL; | |
2103 | } | |
2104 | else | |
2105 | { | |
2106 | /* Create the type for the structure to store the ssa names to. */ | |
2107 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
2108 | type_name = build_decl (UNKNOWN_LOCATION, | |
2109 | TYPE_DECL, create_tmp_var_name (".paral_data"), | |
2110 | type); | |
2111 | TYPE_NAME (type) = type_name; | |
2112 | ||
2113 | name_copies.traverse <tree, add_field_for_name> (type); | |
2114 | if (reduction_list && !reduction_list->is_empty ()) | |
2115 | { | |
2116 | /* Create the fields for reductions. */ | |
2117 | reduction_list->traverse <tree, add_field_for_reduction> (type); | |
2118 | } | |
2119 | layout_type (type); | |
2120 | ||
2121 | /* Create the loads and stores. */ | |
2122 | *arg_struct = create_tmp_var (type, ".paral_data_store"); | |
2123 | nvar = create_tmp_var (build_pointer_type (type), ".paral_data_load"); | |
2124 | *new_arg_struct = make_ssa_name (nvar); | |
2125 | ||
2126 | ld_st_data->store = *arg_struct; | |
2127 | ld_st_data->load = *new_arg_struct; | |
2128 | ld_st_data->store_bb = bb0; | |
2129 | ld_st_data->load_bb = bb1; | |
2130 | ||
2131 | name_copies | |
2132 | .traverse <struct clsn_data *, create_loads_and_stores_for_name> | |
2133 | (ld_st_data); | |
2134 | ||
2135 | /* Load the calculation from memory (after the join of the threads). */ | |
2136 | ||
2137 | if (reduction_list && !reduction_list->is_empty ()) | |
2138 | { | |
2139 | reduction_list | |
2140 | ->traverse <struct clsn_data *, create_stores_for_reduction> | |
2141 | (ld_st_data); | |
2142 | clsn_data.load = make_ssa_name (nvar); | |
2143 | clsn_data.load_bb = exit->dest; | |
2144 | clsn_data.store = ld_st_data->store; | |
2145 | create_final_loads_for_reduction (reduction_list, &clsn_data); | |
2146 | } | |
2147 | } | |
2148 | } | |
2149 | ||
2150 | /* Returns true if FN was created to run in parallel. */ | |
2151 | ||
2152 | bool | |
2153 | parallelized_function_p (tree fndecl) | |
2154 | { | |
2155 | cgraph_node *node = cgraph_node::get (fndecl); | |
2156 | gcc_assert (node != NULL); | |
2157 | return node->parallelized_function; | |
2158 | } | |
2159 | ||
2160 | /* Creates and returns an empty function that will receive the body of | |
2161 | a parallelized loop. */ | |
2162 | ||
2163 | static tree | |
2164 | create_loop_fn (location_t loc) | |
2165 | { | |
2166 | char buf[100]; | |
2167 | char *tname; | |
2168 | tree decl, type, name, t; | |
2169 | struct function *act_cfun = cfun; | |
2170 | static unsigned loopfn_num; | |
2171 | ||
2172 | loc = LOCATION_LOCUS (loc); | |
2173 | snprintf (buf, 100, "%s.$loopfn", current_function_name ()); | |
2174 | ASM_FORMAT_PRIVATE_NAME (tname, buf, loopfn_num++); | |
2175 | clean_symbol_name (tname); | |
2176 | name = get_identifier (tname); | |
2177 | type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); | |
2178 | ||
2179 | decl = build_decl (loc, FUNCTION_DECL, name, type); | |
2180 | TREE_STATIC (decl) = 1; | |
2181 | TREE_USED (decl) = 1; | |
2182 | DECL_ARTIFICIAL (decl) = 1; | |
2183 | DECL_IGNORED_P (decl) = 0; | |
2184 | TREE_PUBLIC (decl) = 0; | |
2185 | DECL_UNINLINABLE (decl) = 1; | |
2186 | DECL_EXTERNAL (decl) = 0; | |
2187 | DECL_CONTEXT (decl) = NULL_TREE; | |
2188 | DECL_INITIAL (decl) = make_node (BLOCK); | |
2189 | BLOCK_SUPERCONTEXT (DECL_INITIAL (decl)) = decl; | |
2190 | ||
2191 | t = build_decl (loc, RESULT_DECL, NULL_TREE, void_type_node); | |
2192 | DECL_ARTIFICIAL (t) = 1; | |
2193 | DECL_IGNORED_P (t) = 1; | |
2194 | DECL_RESULT (decl) = t; | |
2195 | ||
2196 | t = build_decl (loc, PARM_DECL, get_identifier (".paral_data_param"), | |
2197 | ptr_type_node); | |
2198 | DECL_ARTIFICIAL (t) = 1; | |
2199 | DECL_ARG_TYPE (t) = ptr_type_node; | |
2200 | DECL_CONTEXT (t) = decl; | |
2201 | TREE_USED (t) = 1; | |
2202 | DECL_ARGUMENTS (decl) = t; | |
2203 | ||
2204 | allocate_struct_function (decl, false); | |
2205 | ||
2206 | /* The call to allocate_struct_function clobbers CFUN, so we need to restore | |
2207 | it. */ | |
2208 | set_cfun (act_cfun); | |
2209 | ||
2210 | return decl; | |
2211 | } | |
2212 | ||
2213 | /* Replace uses of NAME by VAL in block BB. */ | |
2214 | ||
2215 | static void | |
2216 | replace_uses_in_bb_by (tree name, tree val, basic_block bb) | |
2217 | { | |
2218 | gimple *use_stmt; | |
2219 | imm_use_iterator imm_iter; | |
2220 | ||
2221 | FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, name) | |
2222 | { | |
2223 | if (gimple_bb (use_stmt) != bb) | |
2224 | continue; | |
2225 | ||
2226 | use_operand_p use_p; | |
2227 | FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) | |
2228 | SET_USE (use_p, val); | |
2229 | } | |
2230 | } | |
2231 | ||
2232 | /* Do transformation from: | |
2233 | ||
2234 | <bb preheader>: | |
2235 | ... | |
2236 | goto <bb header> | |
2237 | ||
2238 | <bb header>: | |
2239 | ivtmp_a = PHI <ivtmp_init (preheader), ivtmp_b (latch)> | |
2240 | sum_a = PHI <sum_init (preheader), sum_b (latch)> | |
2241 | ... | |
2242 | use (ivtmp_a) | |
2243 | ... | |
2244 | sum_b = sum_a + sum_update | |
2245 | ... | |
2246 | if (ivtmp_a < n) | |
2247 | goto <bb latch>; | |
2248 | else | |
2249 | goto <bb exit>; | |
2250 | ||
2251 | <bb latch>: | |
2252 | ivtmp_b = ivtmp_a + 1; | |
2253 | goto <bb header> | |
2254 | ||
2255 | <bb exit>: | |
2256 | sum_z = PHI <sum_b (cond[1]), ...> | |
2257 | ||
2258 | [1] Where <bb cond> is single_pred (bb latch); In the simplest case, | |
2259 | that's <bb header>. | |
2260 | ||
2261 | to: | |
2262 | ||
2263 | <bb preheader>: | |
2264 | ... | |
2265 | goto <bb newheader> | |
2266 | ||
2267 | <bb header>: | |
2268 | ivtmp_a = PHI <ivtmp_c (latch)> | |
2269 | sum_a = PHI <sum_c (latch)> | |
2270 | ... | |
2271 | use (ivtmp_a) | |
2272 | ... | |
2273 | sum_b = sum_a + sum_update | |
2274 | ... | |
2275 | goto <bb latch>; | |
2276 | ||
2277 | <bb newheader>: | |
2278 | ivtmp_c = PHI <ivtmp_init (preheader), ivtmp_b (latch)> | |
2279 | sum_c = PHI <sum_init (preheader), sum_b (latch)> | |
2280 | if (ivtmp_c < n + 1) | |
2281 | goto <bb header>; | |
2282 | else | |
2283 | goto <bb newexit>; | |
2284 | ||
2285 | <bb latch>: | |
2286 | ivtmp_b = ivtmp_a + 1; | |
2287 | goto <bb newheader> | |
2288 | ||
2289 | <bb newexit>: | |
2290 | sum_y = PHI <sum_c (newheader)> | |
2291 | ||
2292 | <bb exit>: | |
2293 | sum_z = PHI <sum_y (newexit), ...> | |
2294 | ||
2295 | ||
2296 | In unified diff format: | |
2297 | ||
2298 | <bb preheader>: | |
2299 | ... | |
2300 | - goto <bb header> | |
2301 | + goto <bb newheader> | |
2302 | ||
2303 | <bb header>: | |
2304 | - ivtmp_a = PHI <ivtmp_init (preheader), ivtmp_b (latch)> | |
2305 | - sum_a = PHI <sum_init (preheader), sum_b (latch)> | |
2306 | + ivtmp_a = PHI <ivtmp_c (latch)> | |
2307 | + sum_a = PHI <sum_c (latch)> | |
2308 | ... | |
2309 | use (ivtmp_a) | |
2310 | ... | |
2311 | sum_b = sum_a + sum_update | |
2312 | ... | |
2313 | - if (ivtmp_a < n) | |
2314 | - goto <bb latch>; | |
2315 | + goto <bb latch>; | |
2316 | + | |
2317 | + <bb newheader>: | |
2318 | + ivtmp_c = PHI <ivtmp_init (preheader), ivtmp_b (latch)> | |
2319 | + sum_c = PHI <sum_init (preheader), sum_b (latch)> | |
2320 | + if (ivtmp_c < n + 1) | |
2321 | + goto <bb header>; | |
2322 | else | |
2323 | goto <bb exit>; | |
2324 | ||
2325 | <bb latch>: | |
2326 | ivtmp_b = ivtmp_a + 1; | |
2327 | - goto <bb header> | |
2328 | + goto <bb newheader> | |
2329 | ||
2330 | + <bb newexit>: | |
2331 | + sum_y = PHI <sum_c (newheader)> | |
2332 | ||
2333 | <bb exit>: | |
2334 | - sum_z = PHI <sum_b (cond[1]), ...> | |
2335 | + sum_z = PHI <sum_y (newexit), ...> | |
2336 | ||
2337 | Note: the example does not show any virtual phis, but these are handled more | |
2338 | or less as reductions. | |
2339 | ||
2340 | ||
2341 | Moves the exit condition of LOOP to the beginning of its header. | |
2342 | REDUCTION_LIST describes the reductions in LOOP. BOUND is the new loop | |
2343 | bound. */ | |
2344 | ||
2345 | static void | |
2346 | transform_to_exit_first_loop_alt (class loop *loop, | |
2347 | reduction_info_table_type *reduction_list, | |
2348 | tree bound) | |
2349 | { | |
2350 | basic_block header = loop->header; | |
2351 | basic_block latch = loop->latch; | |
2352 | edge exit = single_dom_exit (loop); | |
2353 | basic_block exit_block = exit->dest; | |
2354 | gcond *cond_stmt = as_a <gcond *> (last_stmt (exit->src)); | |
2355 | tree control = gimple_cond_lhs (cond_stmt); | |
2356 | edge e; | |
2357 | ||
2358 | /* Rewriting virtuals into loop-closed ssa normal form makes this | |
2359 | transformation simpler. It also ensures that the virtuals are in | |
2360 | loop-closed ssa normal from after the transformation, which is required by | |
2361 | create_parallel_loop. */ | |
2362 | rewrite_virtuals_into_loop_closed_ssa (loop); | |
2363 | ||
2364 | /* Create the new_header block. */ | |
2365 | basic_block new_header = split_block_before_cond_jump (exit->src); | |
2366 | edge edge_at_split = single_pred_edge (new_header); | |
2367 | ||
2368 | /* Redirect entry edge to new_header. */ | |
2369 | edge entry = loop_preheader_edge (loop); | |
2370 | e = redirect_edge_and_branch (entry, new_header); | |
2371 | gcc_assert (e == entry); | |
2372 | ||
2373 | /* Redirect post_inc_edge to new_header. */ | |
2374 | edge post_inc_edge = single_succ_edge (latch); | |
2375 | e = redirect_edge_and_branch (post_inc_edge, new_header); | |
2376 | gcc_assert (e == post_inc_edge); | |
2377 | ||
2378 | /* Redirect post_cond_edge to header. */ | |
2379 | edge post_cond_edge = single_pred_edge (latch); | |
2380 | e = redirect_edge_and_branch (post_cond_edge, header); | |
2381 | gcc_assert (e == post_cond_edge); | |
2382 | ||
2383 | /* Redirect edge_at_split to latch. */ | |
2384 | e = redirect_edge_and_branch (edge_at_split, latch); | |
2385 | gcc_assert (e == edge_at_split); | |
2386 | ||
2387 | /* Set the new loop bound. */ | |
2388 | gimple_cond_set_rhs (cond_stmt, bound); | |
2389 | update_stmt (cond_stmt); | |
2390 | ||
2391 | /* Repair the ssa. */ | |
2392 | vec<edge_var_map> *v = redirect_edge_var_map_vector (post_inc_edge); | |
2393 | edge_var_map *vm; | |
2394 | gphi_iterator gsi; | |
2395 | int i; | |
2396 | for (gsi = gsi_start_phis (header), i = 0; | |
2397 | !gsi_end_p (gsi) && v->iterate (i, &vm); | |
2398 | gsi_next (&gsi), i++) | |
2399 | { | |
2400 | gphi *phi = gsi.phi (); | |
2401 | tree res_a = PHI_RESULT (phi); | |
2402 | ||
2403 | /* Create new phi. */ | |
2404 | tree res_c = copy_ssa_name (res_a, phi); | |
2405 | gphi *nphi = create_phi_node (res_c, new_header); | |
2406 | ||
2407 | /* Replace ivtmp_a with ivtmp_c in condition 'if (ivtmp_a < n)'. */ | |
2408 | replace_uses_in_bb_by (res_a, res_c, new_header); | |
2409 | ||
2410 | /* Replace ivtmp/sum_b with ivtmp/sum_c in header phi. */ | |
2411 | add_phi_arg (phi, res_c, post_cond_edge, UNKNOWN_LOCATION); | |
2412 | ||
2413 | /* Replace sum_b with sum_c in exit phi. */ | |
2414 | tree res_b = redirect_edge_var_map_def (vm); | |
2415 | replace_uses_in_bb_by (res_b, res_c, exit_block); | |
2416 | ||
2417 | struct reduction_info *red = reduction_phi (reduction_list, phi); | |
2418 | gcc_assert (virtual_operand_p (res_a) | |
2419 | || res_a == control | |
2420 | || red != NULL); | |
2421 | ||
2422 | if (red) | |
2423 | { | |
2424 | /* Register the new reduction phi. */ | |
2425 | red->reduc_phi = nphi; | |
2426 | gimple_set_uid (red->reduc_phi, red->reduc_version); | |
2427 | } | |
2428 | } | |
2429 | gcc_assert (gsi_end_p (gsi) && !v->iterate (i, &vm)); | |
2430 | ||
2431 | /* Set the preheader argument of the new phis to ivtmp/sum_init. */ | |
2432 | flush_pending_stmts (entry); | |
2433 | ||
2434 | /* Set the latch arguments of the new phis to ivtmp/sum_b. */ | |
2435 | flush_pending_stmts (post_inc_edge); | |
2436 | ||
2437 | ||
2438 | basic_block new_exit_block = NULL; | |
2439 | if (!single_pred_p (exit->dest)) | |
2440 | { | |
2441 | /* Create a new empty exit block, inbetween the new loop header and the | |
2442 | old exit block. The function separate_decls_in_region needs this block | |
2443 | to insert code that is active on loop exit, but not any other path. */ | |
2444 | new_exit_block = split_edge (exit); | |
2445 | } | |
2446 | ||
2447 | /* Insert and register the reduction exit phis. */ | |
2448 | for (gphi_iterator gsi = gsi_start_phis (exit_block); | |
2449 | !gsi_end_p (gsi); | |
2450 | gsi_next (&gsi)) | |
2451 | { | |
2452 | gphi *phi = gsi.phi (); | |
2453 | gphi *nphi = NULL; | |
2454 | tree res_z = PHI_RESULT (phi); | |
2455 | tree res_c; | |
2456 | ||
2457 | if (new_exit_block != NULL) | |
2458 | { | |
2459 | /* Now that we have a new exit block, duplicate the phi of the old | |
2460 | exit block in the new exit block to preserve loop-closed ssa. */ | |
2461 | edge succ_new_exit_block = single_succ_edge (new_exit_block); | |
2462 | edge pred_new_exit_block = single_pred_edge (new_exit_block); | |
2463 | tree res_y = copy_ssa_name (res_z, phi); | |
2464 | nphi = create_phi_node (res_y, new_exit_block); | |
2465 | res_c = PHI_ARG_DEF_FROM_EDGE (phi, succ_new_exit_block); | |
2466 | add_phi_arg (nphi, res_c, pred_new_exit_block, UNKNOWN_LOCATION); | |
2467 | add_phi_arg (phi, res_y, succ_new_exit_block, UNKNOWN_LOCATION); | |
2468 | } | |
2469 | else | |
2470 | res_c = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
2471 | ||
2472 | if (virtual_operand_p (res_z)) | |
2473 | continue; | |
2474 | ||
2475 | gimple *reduc_phi = SSA_NAME_DEF_STMT (res_c); | |
2476 | struct reduction_info *red = reduction_phi (reduction_list, reduc_phi); | |
2477 | if (red != NULL) | |
2478 | red->keep_res = (nphi != NULL | |
2479 | ? nphi | |
2480 | : phi); | |
2481 | } | |
2482 | ||
2483 | /* We're going to cancel the loop at the end of gen_parallel_loop, but until | |
2484 | then we're still using some fields, so only bother about fields that are | |
2485 | still used: header and latch. | |
2486 | The loop has a new header bb, so we update it. The latch bb stays the | |
2487 | same. */ | |
2488 | loop->header = new_header; | |
2489 | ||
2490 | /* Recalculate dominance info. */ | |
2491 | free_dominance_info (CDI_DOMINATORS); | |
2492 | calculate_dominance_info (CDI_DOMINATORS); | |
2493 | ||
2494 | checking_verify_ssa (true, true); | |
2495 | } | |
2496 | ||
2497 | /* Tries to moves the exit condition of LOOP to the beginning of its header | |
2498 | without duplication of the loop body. NIT is the number of iterations of the | |
2499 | loop. REDUCTION_LIST describes the reductions in LOOP. Return true if | |
2500 | transformation is successful. */ | |
2501 | ||
2502 | static bool | |
2503 | try_transform_to_exit_first_loop_alt (class loop *loop, | |
2504 | reduction_info_table_type *reduction_list, | |
2505 | tree nit) | |
2506 | { | |
2507 | /* Check whether the latch contains a single statement. */ | |
2508 | if (!gimple_seq_nondebug_singleton_p (bb_seq (loop->latch))) | |
2509 | return false; | |
2510 | ||
2511 | /* Check whether the latch contains no phis. */ | |
2512 | if (phi_nodes (loop->latch) != NULL) | |
2513 | return false; | |
2514 | ||
2515 | /* Check whether the latch contains the loop iv increment. */ | |
2516 | edge back = single_succ_edge (loop->latch); | |
2517 | edge exit = single_dom_exit (loop); | |
2518 | gcond *cond_stmt = as_a <gcond *> (last_stmt (exit->src)); | |
2519 | tree control = gimple_cond_lhs (cond_stmt); | |
2520 | gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (control)); | |
2521 | tree inc_res = gimple_phi_arg_def (phi, back->dest_idx); | |
2522 | if (gimple_bb (SSA_NAME_DEF_STMT (inc_res)) != loop->latch) | |
2523 | return false; | |
2524 | ||
2525 | /* Check whether there's no code between the loop condition and the latch. */ | |
2526 | if (!single_pred_p (loop->latch) | |
2527 | || single_pred (loop->latch) != exit->src) | |
2528 | return false; | |
2529 | ||
2530 | tree alt_bound = NULL_TREE; | |
2531 | tree nit_type = TREE_TYPE (nit); | |
2532 | ||
2533 | /* Figure out whether nit + 1 overflows. */ | |
2534 | if (TREE_CODE (nit) == INTEGER_CST) | |
2535 | { | |
2536 | if (!tree_int_cst_equal (nit, TYPE_MAX_VALUE (nit_type))) | |
2537 | { | |
2538 | alt_bound = fold_build2_loc (UNKNOWN_LOCATION, PLUS_EXPR, nit_type, | |
2539 | nit, build_one_cst (nit_type)); | |
2540 | ||
2541 | gcc_assert (TREE_CODE (alt_bound) == INTEGER_CST); | |
2542 | transform_to_exit_first_loop_alt (loop, reduction_list, alt_bound); | |
2543 | return true; | |
2544 | } | |
2545 | else | |
2546 | { | |
2547 | /* Todo: Figure out if we can trigger this, if it's worth to handle | |
2548 | optimally, and if we can handle it optimally. */ | |
2549 | return false; | |
2550 | } | |
2551 | } | |
2552 | ||
2553 | gcc_assert (TREE_CODE (nit) == SSA_NAME); | |
2554 | ||
2555 | /* Variable nit is the loop bound as returned by canonicalize_loop_ivs, for an | |
2556 | iv with base 0 and step 1 that is incremented in the latch, like this: | |
2557 | ||
2558 | <bb header>: | |
2559 | # iv_1 = PHI <0 (preheader), iv_2 (latch)> | |
2560 | ... | |
2561 | if (iv_1 < nit) | |
2562 | goto <bb latch>; | |
2563 | else | |
2564 | goto <bb exit>; | |
2565 | ||
2566 | <bb latch>: | |
2567 | iv_2 = iv_1 + 1; | |
2568 | goto <bb header>; | |
2569 | ||
2570 | The range of iv_1 is [0, nit]. The latch edge is taken for | |
2571 | iv_1 == [0, nit - 1] and the exit edge is taken for iv_1 == nit. So the | |
2572 | number of latch executions is equal to nit. | |
2573 | ||
2574 | The function max_loop_iterations gives us the maximum number of latch | |
2575 | executions, so it gives us the maximum value of nit. */ | |
2576 | widest_int nit_max; | |
2577 | if (!max_loop_iterations (loop, &nit_max)) | |
2578 | return false; | |
2579 | ||
2580 | /* Check if nit + 1 overflows. */ | |
2581 | widest_int type_max = wi::to_widest (TYPE_MAX_VALUE (nit_type)); | |
2582 | if (nit_max >= type_max) | |
2583 | return false; | |
2584 | ||
2585 | gimple *def = SSA_NAME_DEF_STMT (nit); | |
2586 | ||
2587 | /* Try to find nit + 1, in the form of n in an assignment nit = n - 1. */ | |
2588 | if (def | |
2589 | && is_gimple_assign (def) | |
2590 | && gimple_assign_rhs_code (def) == PLUS_EXPR) | |
2591 | { | |
2592 | tree op1 = gimple_assign_rhs1 (def); | |
2593 | tree op2 = gimple_assign_rhs2 (def); | |
2594 | if (integer_minus_onep (op1)) | |
2595 | alt_bound = op2; | |
2596 | else if (integer_minus_onep (op2)) | |
2597 | alt_bound = op1; | |
2598 | } | |
2599 | ||
2600 | /* If not found, insert nit + 1. */ | |
2601 | if (alt_bound == NULL_TREE) | |
2602 | { | |
2603 | alt_bound = fold_build2 (PLUS_EXPR, nit_type, nit, | |
2604 | build_int_cst_type (nit_type, 1)); | |
2605 | ||
2606 | gimple_stmt_iterator gsi = gsi_last_bb (loop_preheader_edge (loop)->src); | |
2607 | ||
2608 | alt_bound | |
2609 | = force_gimple_operand_gsi (&gsi, alt_bound, true, NULL_TREE, false, | |
2610 | GSI_CONTINUE_LINKING); | |
2611 | } | |
2612 | ||
2613 | transform_to_exit_first_loop_alt (loop, reduction_list, alt_bound); | |
2614 | return true; | |
2615 | } | |
2616 | ||
2617 | /* Moves the exit condition of LOOP to the beginning of its header. NIT is the | |
2618 | number of iterations of the loop. REDUCTION_LIST describes the reductions in | |
2619 | LOOP. */ | |
2620 | ||
2621 | static void | |
2622 | transform_to_exit_first_loop (class loop *loop, | |
2623 | reduction_info_table_type *reduction_list, | |
2624 | tree nit) | |
2625 | { | |
2626 | basic_block *bbs, *nbbs, ex_bb, orig_header; | |
2627 | unsigned n; | |
2628 | bool ok; | |
2629 | edge exit = single_dom_exit (loop), hpred; | |
2630 | tree control, control_name, res, t; | |
2631 | gphi *phi, *nphi; | |
2632 | gassign *stmt; | |
2633 | gcond *cond_stmt, *cond_nit; | |
2634 | tree nit_1; | |
2635 | ||
2636 | split_block_after_labels (loop->header); | |
2637 | orig_header = single_succ (loop->header); | |
2638 | hpred = single_succ_edge (loop->header); | |
2639 | ||
2640 | cond_stmt = as_a <gcond *> (last_stmt (exit->src)); | |
2641 | control = gimple_cond_lhs (cond_stmt); | |
2642 | gcc_assert (gimple_cond_rhs (cond_stmt) == nit); | |
2643 | ||
2644 | /* Make sure that we have phi nodes on exit for all loop header phis | |
2645 | (create_parallel_loop requires that). */ | |
2646 | for (gphi_iterator gsi = gsi_start_phis (loop->header); | |
2647 | !gsi_end_p (gsi); | |
2648 | gsi_next (&gsi)) | |
2649 | { | |
2650 | phi = gsi.phi (); | |
2651 | res = PHI_RESULT (phi); | |
2652 | t = copy_ssa_name (res, phi); | |
2653 | SET_PHI_RESULT (phi, t); | |
2654 | nphi = create_phi_node (res, orig_header); | |
2655 | add_phi_arg (nphi, t, hpred, UNKNOWN_LOCATION); | |
2656 | ||
2657 | if (res == control) | |
2658 | { | |
2659 | gimple_cond_set_lhs (cond_stmt, t); | |
2660 | update_stmt (cond_stmt); | |
2661 | control = t; | |
2662 | } | |
2663 | } | |
2664 | ||
2665 | bbs = get_loop_body_in_dom_order (loop); | |
2666 | ||
2667 | for (n = 0; bbs[n] != exit->src; n++) | |
2668 | continue; | |
2669 | nbbs = XNEWVEC (basic_block, n); | |
2670 | ok = gimple_duplicate_sese_tail (single_succ_edge (loop->header), exit, | |
2671 | bbs + 1, n, nbbs); | |
2672 | gcc_assert (ok); | |
2673 | free (bbs); | |
2674 | ex_bb = nbbs[0]; | |
2675 | free (nbbs); | |
2676 | ||
2677 | /* Other than reductions, the only gimple reg that should be copied | |
2678 | out of the loop is the control variable. */ | |
2679 | exit = single_dom_exit (loop); | |
2680 | control_name = NULL_TREE; | |
2681 | for (gphi_iterator gsi = gsi_start_phis (ex_bb); | |
2682 | !gsi_end_p (gsi); ) | |
2683 | { | |
2684 | phi = gsi.phi (); | |
2685 | res = PHI_RESULT (phi); | |
2686 | if (virtual_operand_p (res)) | |
2687 | { | |
2688 | gsi_next (&gsi); | |
2689 | continue; | |
2690 | } | |
2691 | ||
2692 | /* Check if it is a part of reduction. If it is, | |
2693 | keep the phi at the reduction's keep_res field. The | |
2694 | PHI_RESULT of this phi is the resulting value of the reduction | |
2695 | variable when exiting the loop. */ | |
2696 | ||
2697 | if (!reduction_list->is_empty ()) | |
2698 | { | |
2699 | struct reduction_info *red; | |
2700 | ||
2701 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
2702 | red = reduction_phi (reduction_list, SSA_NAME_DEF_STMT (val)); | |
2703 | if (red) | |
2704 | { | |
2705 | red->keep_res = phi; | |
2706 | gsi_next (&gsi); | |
2707 | continue; | |
2708 | } | |
2709 | } | |
2710 | gcc_assert (control_name == NULL_TREE | |
2711 | && SSA_NAME_VAR (res) == SSA_NAME_VAR (control)); | |
2712 | control_name = res; | |
2713 | remove_phi_node (&gsi, false); | |
2714 | } | |
2715 | gcc_assert (control_name != NULL_TREE); | |
2716 | ||
2717 | /* Initialize the control variable to number of iterations | |
2718 | according to the rhs of the exit condition. */ | |
2719 | gimple_stmt_iterator gsi = gsi_after_labels (ex_bb); | |
2720 | cond_nit = as_a <gcond *> (last_stmt (exit->src)); | |
2721 | nit_1 = gimple_cond_rhs (cond_nit); | |
2722 | nit_1 = force_gimple_operand_gsi (&gsi, | |
2723 | fold_convert (TREE_TYPE (control_name), nit_1), | |
2724 | false, NULL_TREE, false, GSI_SAME_STMT); | |
2725 | stmt = gimple_build_assign (control_name, nit_1); | |
2726 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
2727 | } | |
2728 | ||
2729 | /* Create the parallel constructs for LOOP as described in gen_parallel_loop. | |
2730 | LOOP_FN and DATA are the arguments of GIMPLE_OMP_PARALLEL. | |
2731 | NEW_DATA is the variable that should be initialized from the argument | |
2732 | of LOOP_FN. N_THREADS is the requested number of threads, which can be 0 if | |
2733 | that number is to be determined later. */ | |
2734 | ||
2735 | static void | |
2736 | create_parallel_loop (class loop *loop, tree loop_fn, tree data, | |
2737 | tree new_data, unsigned n_threads, location_t loc, | |
2738 | bool oacc_kernels_p) | |
2739 | { | |
2740 | gimple_stmt_iterator gsi; | |
2741 | basic_block for_bb, ex_bb, continue_bb; | |
2742 | tree t, param; | |
2743 | gomp_parallel *omp_par_stmt; | |
2744 | gimple *omp_return_stmt1, *omp_return_stmt2; | |
2745 | gimple *phi; | |
2746 | gcond *cond_stmt; | |
2747 | gomp_for *for_stmt; | |
2748 | gomp_continue *omp_cont_stmt; | |
2749 | tree cvar, cvar_init, initvar, cvar_next, cvar_base, type; | |
2750 | edge exit, nexit, guard, end, e; | |
2751 | ||
2752 | if (oacc_kernels_p) | |
2753 | { | |
2754 | gcc_checking_assert (lookup_attribute ("oacc kernels", | |
2755 | DECL_ATTRIBUTES (cfun->decl))); | |
2756 | /* Indicate to later processing that this is a parallelized OpenACC | |
2757 | kernels construct. */ | |
2758 | DECL_ATTRIBUTES (cfun->decl) | |
2759 | = tree_cons (get_identifier ("oacc kernels parallelized"), | |
2760 | NULL_TREE, DECL_ATTRIBUTES (cfun->decl)); | |
2761 | } | |
2762 | else | |
2763 | { | |
2764 | /* Prepare the GIMPLE_OMP_PARALLEL statement. */ | |
2765 | ||
2766 | basic_block bb = loop_preheader_edge (loop)->src; | |
2767 | basic_block paral_bb = single_pred (bb); | |
2768 | gsi = gsi_last_bb (paral_bb); | |
2769 | ||
2770 | gcc_checking_assert (n_threads != 0); | |
2771 | t = build_omp_clause (loc, OMP_CLAUSE_NUM_THREADS); | |
2772 | OMP_CLAUSE_NUM_THREADS_EXPR (t) | |
2773 | = build_int_cst (integer_type_node, n_threads); | |
2774 | omp_par_stmt = gimple_build_omp_parallel (NULL, t, loop_fn, data); | |
2775 | gimple_set_location (omp_par_stmt, loc); | |
2776 | ||
2777 | gsi_insert_after (&gsi, omp_par_stmt, GSI_NEW_STMT); | |
2778 | ||
2779 | /* Initialize NEW_DATA. */ | |
2780 | if (data) | |
2781 | { | |
2782 | gassign *assign_stmt; | |
2783 | ||
2784 | gsi = gsi_after_labels (bb); | |
2785 | ||
2786 | param = make_ssa_name (DECL_ARGUMENTS (loop_fn)); | |
2787 | assign_stmt = gimple_build_assign (param, build_fold_addr_expr (data)); | |
2788 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
2789 | ||
2790 | assign_stmt = gimple_build_assign (new_data, | |
2791 | fold_convert (TREE_TYPE (new_data), param)); | |
2792 | gsi_insert_before (&gsi, assign_stmt, GSI_SAME_STMT); | |
2793 | } | |
2794 | ||
2795 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_PARALLEL. */ | |
2796 | bb = split_loop_exit_edge (single_dom_exit (loop)); | |
2797 | gsi = gsi_last_bb (bb); | |
2798 | omp_return_stmt1 = gimple_build_omp_return (false); | |
2799 | gimple_set_location (omp_return_stmt1, loc); | |
2800 | gsi_insert_after (&gsi, omp_return_stmt1, GSI_NEW_STMT); | |
2801 | } | |
2802 | ||
2803 | /* Extract data for GIMPLE_OMP_FOR. */ | |
2804 | gcc_assert (loop->header == single_dom_exit (loop)->src); | |
2805 | cond_stmt = as_a <gcond *> (last_stmt (loop->header)); | |
2806 | ||
2807 | cvar = gimple_cond_lhs (cond_stmt); | |
2808 | cvar_base = SSA_NAME_VAR (cvar); | |
2809 | phi = SSA_NAME_DEF_STMT (cvar); | |
2810 | cvar_init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
2811 | initvar = copy_ssa_name (cvar); | |
2812 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, loop_preheader_edge (loop)), | |
2813 | initvar); | |
2814 | cvar_next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
2815 | ||
2816 | gsi = gsi_last_nondebug_bb (loop->latch); | |
2817 | gcc_assert (gsi_stmt (gsi) == SSA_NAME_DEF_STMT (cvar_next)); | |
2818 | gsi_remove (&gsi, true); | |
2819 | ||
2820 | /* Prepare cfg. */ | |
2821 | for_bb = split_edge (loop_preheader_edge (loop)); | |
2822 | ex_bb = split_loop_exit_edge (single_dom_exit (loop)); | |
2823 | extract_true_false_edges_from_block (loop->header, &nexit, &exit); | |
2824 | gcc_assert (exit == single_dom_exit (loop)); | |
2825 | ||
2826 | guard = make_edge (for_bb, ex_bb, 0); | |
2827 | /* FIXME: What is the probability? */ | |
2828 | guard->probability = profile_probability::guessed_never (); | |
2829 | /* Split the latch edge, so LOOPS_HAVE_SIMPLE_LATCHES is still valid. */ | |
2830 | loop->latch = split_edge (single_succ_edge (loop->latch)); | |
2831 | single_pred_edge (loop->latch)->flags = 0; | |
2832 | end = make_single_succ_edge (single_pred (loop->latch), ex_bb, EDGE_FALLTHRU); | |
2833 | rescan_loop_exit (end, true, false); | |
2834 | ||
2835 | for (gphi_iterator gpi = gsi_start_phis (ex_bb); | |
2836 | !gsi_end_p (gpi); gsi_next (&gpi)) | |
2837 | { | |
2838 | location_t locus; | |
2839 | gphi *phi = gpi.phi (); | |
2840 | tree def = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
2841 | gimple *def_stmt = SSA_NAME_DEF_STMT (def); | |
2842 | ||
2843 | /* If the exit phi is not connected to a header phi in the same loop, this | |
2844 | value is not modified in the loop, and we're done with this phi. */ | |
2845 | if (!(gimple_code (def_stmt) == GIMPLE_PHI | |
2846 | && gimple_bb (def_stmt) == loop->header)) | |
2847 | { | |
2848 | locus = gimple_phi_arg_location_from_edge (phi, exit); | |
2849 | add_phi_arg (phi, def, guard, locus); | |
2850 | add_phi_arg (phi, def, end, locus); | |
2851 | continue; | |
2852 | } | |
2853 | ||
2854 | gphi *stmt = as_a <gphi *> (def_stmt); | |
2855 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_preheader_edge (loop)); | |
2856 | locus = gimple_phi_arg_location_from_edge (stmt, | |
2857 | loop_preheader_edge (loop)); | |
2858 | add_phi_arg (phi, def, guard, locus); | |
2859 | ||
2860 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (loop)); | |
2861 | locus = gimple_phi_arg_location_from_edge (stmt, loop_latch_edge (loop)); | |
2862 | add_phi_arg (phi, def, end, locus); | |
2863 | } | |
2864 | e = redirect_edge_and_branch (exit, nexit->dest); | |
2865 | PENDING_STMT (e) = NULL; | |
2866 | ||
2867 | /* Emit GIMPLE_OMP_FOR. */ | |
2868 | if (oacc_kernels_p) | |
2869 | /* Parallelized OpenACC kernels constructs use gang parallelism. See also | |
2870 | omp-offload.c:execute_oacc_device_lower. */ | |
2871 | t = build_omp_clause (loc, OMP_CLAUSE_GANG); | |
2872 | else | |
2873 | { | |
2874 | t = build_omp_clause (loc, OMP_CLAUSE_SCHEDULE); | |
2875 | int chunk_size = param_parloops_chunk_size; | |
2876 | switch (param_parloops_schedule) | |
2877 | { | |
2878 | case PARLOOPS_SCHEDULE_STATIC: | |
2879 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_STATIC; | |
2880 | break; | |
2881 | case PARLOOPS_SCHEDULE_DYNAMIC: | |
2882 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_DYNAMIC; | |
2883 | break; | |
2884 | case PARLOOPS_SCHEDULE_GUIDED: | |
2885 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_GUIDED; | |
2886 | break; | |
2887 | case PARLOOPS_SCHEDULE_AUTO: | |
2888 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_AUTO; | |
2889 | chunk_size = 0; | |
2890 | break; | |
2891 | case PARLOOPS_SCHEDULE_RUNTIME: | |
2892 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_RUNTIME; | |
2893 | chunk_size = 0; | |
2894 | break; | |
2895 | default: | |
2896 | gcc_unreachable (); | |
2897 | } | |
2898 | if (chunk_size != 0) | |
2899 | OMP_CLAUSE_SCHEDULE_CHUNK_EXPR (t) | |
2900 | = build_int_cst (integer_type_node, chunk_size); | |
2901 | } | |
2902 | ||
2903 | for_stmt = gimple_build_omp_for (NULL, | |
2904 | (oacc_kernels_p | |
2905 | ? GF_OMP_FOR_KIND_OACC_LOOP | |
2906 | : GF_OMP_FOR_KIND_FOR), | |
2907 | t, 1, NULL); | |
2908 | ||
2909 | gimple_cond_set_lhs (cond_stmt, cvar_base); | |
2910 | type = TREE_TYPE (cvar); | |
2911 | gimple_set_location (for_stmt, loc); | |
2912 | gimple_omp_for_set_index (for_stmt, 0, initvar); | |
2913 | gimple_omp_for_set_initial (for_stmt, 0, cvar_init); | |
2914 | gimple_omp_for_set_final (for_stmt, 0, gimple_cond_rhs (cond_stmt)); | |
2915 | gimple_omp_for_set_cond (for_stmt, 0, gimple_cond_code (cond_stmt)); | |
2916 | gimple_omp_for_set_incr (for_stmt, 0, build2 (PLUS_EXPR, type, | |
2917 | cvar_base, | |
2918 | build_int_cst (type, 1))); | |
2919 | ||
2920 | gsi = gsi_last_bb (for_bb); | |
2921 | gsi_insert_after (&gsi, for_stmt, GSI_NEW_STMT); | |
2922 | SSA_NAME_DEF_STMT (initvar) = for_stmt; | |
2923 | ||
2924 | /* Emit GIMPLE_OMP_CONTINUE. */ | |
2925 | continue_bb = single_pred (loop->latch); | |
2926 | gsi = gsi_last_bb (continue_bb); | |
2927 | omp_cont_stmt = gimple_build_omp_continue (cvar_next, cvar); | |
2928 | gimple_set_location (omp_cont_stmt, loc); | |
2929 | gsi_insert_after (&gsi, omp_cont_stmt, GSI_NEW_STMT); | |
2930 | SSA_NAME_DEF_STMT (cvar_next) = omp_cont_stmt; | |
2931 | ||
2932 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_FOR. */ | |
2933 | gsi = gsi_last_bb (ex_bb); | |
2934 | omp_return_stmt2 = gimple_build_omp_return (true); | |
2935 | gimple_set_location (omp_return_stmt2, loc); | |
2936 | gsi_insert_after (&gsi, omp_return_stmt2, GSI_NEW_STMT); | |
2937 | ||
2938 | /* After the above dom info is hosed. Re-compute it. */ | |
2939 | free_dominance_info (CDI_DOMINATORS); | |
2940 | calculate_dominance_info (CDI_DOMINATORS); | |
2941 | } | |
2942 | ||
2943 | /* Return number of phis in bb. If COUNT_VIRTUAL_P is false, don't count the | |
2944 | virtual phi. */ | |
2945 | ||
2946 | static unsigned int | |
2947 | num_phis (basic_block bb, bool count_virtual_p) | |
2948 | { | |
2949 | unsigned int nr_phis = 0; | |
2950 | gphi_iterator gsi; | |
2951 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2952 | { | |
2953 | if (!count_virtual_p && virtual_operand_p (PHI_RESULT (gsi.phi ()))) | |
2954 | continue; | |
2955 | ||
2956 | nr_phis++; | |
2957 | } | |
2958 | ||
2959 | return nr_phis; | |
2960 | } | |
2961 | ||
2962 | /* Generates code to execute the iterations of LOOP in N_THREADS | |
2963 | threads in parallel, which can be 0 if that number is to be determined | |
2964 | later. | |
2965 | ||
2966 | NITER describes number of iterations of LOOP. | |
2967 | REDUCTION_LIST describes the reductions existent in the LOOP. */ | |
2968 | ||
2969 | static void | |
2970 | gen_parallel_loop (class loop *loop, | |
2971 | reduction_info_table_type *reduction_list, | |
2972 | unsigned n_threads, class tree_niter_desc *niter, | |
2973 | bool oacc_kernels_p) | |
2974 | { | |
2975 | tree many_iterations_cond, type, nit; | |
2976 | tree arg_struct, new_arg_struct; | |
2977 | gimple_seq stmts; | |
2978 | edge entry, exit; | |
2979 | struct clsn_data clsn_data; | |
2980 | location_t loc; | |
2981 | gimple *cond_stmt; | |
2982 | unsigned int m_p_thread=2; | |
2983 | ||
2984 | /* From | |
2985 | ||
2986 | --------------------------------------------------------------------- | |
2987 | loop | |
2988 | { | |
2989 | IV = phi (INIT, IV + STEP) | |
2990 | BODY1; | |
2991 | if (COND) | |
2992 | break; | |
2993 | BODY2; | |
2994 | } | |
2995 | --------------------------------------------------------------------- | |
2996 | ||
2997 | with # of iterations NITER (possibly with MAY_BE_ZERO assumption), | |
2998 | we generate the following code: | |
2999 | ||
3000 | --------------------------------------------------------------------- | |
3001 | ||
3002 | if (MAY_BE_ZERO | |
3003 | || NITER < MIN_PER_THREAD * N_THREADS) | |
3004 | goto original; | |
3005 | ||
3006 | BODY1; | |
3007 | store all local loop-invariant variables used in body of the loop to DATA. | |
3008 | GIMPLE_OMP_PARALLEL (OMP_CLAUSE_NUM_THREADS (N_THREADS), LOOPFN, DATA); | |
3009 | load the variables from DATA. | |
3010 | GIMPLE_OMP_FOR (IV = INIT; COND; IV += STEP) (OMP_CLAUSE_SCHEDULE (static)) | |
3011 | BODY2; | |
3012 | BODY1; | |
3013 | GIMPLE_OMP_CONTINUE; | |
3014 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_FOR | |
3015 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_PARALLEL | |
3016 | goto end; | |
3017 | ||
3018 | original: | |
3019 | loop | |
3020 | { | |
3021 | IV = phi (INIT, IV + STEP) | |
3022 | BODY1; | |
3023 | if (COND) | |
3024 | break; | |
3025 | BODY2; | |
3026 | } | |
3027 | ||
3028 | end: | |
3029 | ||
3030 | */ | |
3031 | ||
3032 | /* Create two versions of the loop -- in the old one, we know that the | |
3033 | number of iterations is large enough, and we will transform it into the | |
3034 | loop that will be split to loop_fn, the new one will be used for the | |
3035 | remaining iterations. */ | |
3036 | ||
3037 | /* We should compute a better number-of-iterations value for outer loops. | |
3038 | That is, if we have | |
3039 | ||
3040 | for (i = 0; i < n; ++i) | |
3041 | for (j = 0; j < m; ++j) | |
3042 | ... | |
3043 | ||
3044 | we should compute nit = n * m, not nit = n. | |
3045 | Also may_be_zero handling would need to be adjusted. */ | |
3046 | ||
3047 | type = TREE_TYPE (niter->niter); | |
3048 | nit = force_gimple_operand (unshare_expr (niter->niter), &stmts, true, | |
3049 | NULL_TREE); | |
3050 | if (stmts) | |
3051 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
3052 | ||
3053 | if (!oacc_kernels_p) | |
3054 | { | |
3055 | if (loop->inner) | |
3056 | m_p_thread=2; | |
3057 | else | |
3058 | m_p_thread=MIN_PER_THREAD; | |
3059 | ||
3060 | gcc_checking_assert (n_threads != 0); | |
3061 | many_iterations_cond = | |
3062 | fold_build2 (GE_EXPR, boolean_type_node, | |
3063 | nit, build_int_cst (type, m_p_thread * n_threads - 1)); | |
3064 | ||
3065 | many_iterations_cond | |
3066 | = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
3067 | invert_truthvalue (unshare_expr (niter->may_be_zero)), | |
3068 | many_iterations_cond); | |
3069 | many_iterations_cond | |
3070 | = force_gimple_operand (many_iterations_cond, &stmts, false, NULL_TREE); | |
3071 | if (stmts) | |
3072 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
3073 | if (!is_gimple_condexpr (many_iterations_cond)) | |
3074 | { | |
3075 | many_iterations_cond | |
3076 | = force_gimple_operand (many_iterations_cond, &stmts, | |
3077 | true, NULL_TREE); | |
3078 | if (stmts) | |
3079 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), | |
3080 | stmts); | |
3081 | } | |
3082 | ||
3083 | initialize_original_copy_tables (); | |
3084 | ||
3085 | /* We assume that the loop usually iterates a lot. */ | |
3086 | loop_version (loop, many_iterations_cond, NULL, | |
3087 | profile_probability::likely (), | |
3088 | profile_probability::unlikely (), | |
3089 | profile_probability::likely (), | |
3090 | profile_probability::unlikely (), true); | |
3091 | update_ssa (TODO_update_ssa); | |
3092 | free_original_copy_tables (); | |
3093 | } | |
3094 | ||
3095 | /* Base all the induction variables in LOOP on a single control one. */ | |
3096 | canonicalize_loop_ivs (loop, &nit, true); | |
3097 | if (num_phis (loop->header, false) != reduction_list->elements () + 1) | |
3098 | { | |
3099 | /* The call to canonicalize_loop_ivs above failed to "base all the | |
3100 | induction variables in LOOP on a single control one". Do damage | |
3101 | control. */ | |
3102 | basic_block preheader = loop_preheader_edge (loop)->src; | |
3103 | basic_block cond_bb = single_pred (preheader); | |
3104 | gcond *cond = as_a <gcond *> (gsi_stmt (gsi_last_bb (cond_bb))); | |
3105 | gimple_cond_make_true (cond); | |
3106 | update_stmt (cond); | |
3107 | /* We've gotten rid of the duplicate loop created by loop_version, but | |
3108 | we can't undo whatever canonicalize_loop_ivs has done. | |
3109 | TODO: Fix this properly by ensuring that the call to | |
3110 | canonicalize_loop_ivs succeeds. */ | |
3111 | if (dump_file | |
3112 | && (dump_flags & TDF_DETAILS)) | |
3113 | fprintf (dump_file, "canonicalize_loop_ivs failed for loop %d," | |
3114 | " aborting transformation\n", loop->num); | |
3115 | return; | |
3116 | } | |
3117 | ||
3118 | /* Ensure that the exit condition is the first statement in the loop. | |
3119 | The common case is that latch of the loop is empty (apart from the | |
3120 | increment) and immediately follows the loop exit test. Attempt to move the | |
3121 | entry of the loop directly before the exit check and increase the number of | |
3122 | iterations of the loop by one. */ | |
3123 | if (try_transform_to_exit_first_loop_alt (loop, reduction_list, nit)) | |
3124 | { | |
3125 | if (dump_file | |
3126 | && (dump_flags & TDF_DETAILS)) | |
3127 | fprintf (dump_file, | |
3128 | "alternative exit-first loop transform succeeded" | |
3129 | " for loop %d\n", loop->num); | |
3130 | } | |
3131 | else | |
3132 | { | |
3133 | if (oacc_kernels_p) | |
3134 | n_threads = 1; | |
3135 | ||
3136 | /* Fall back on the method that handles more cases, but duplicates the | |
3137 | loop body: move the exit condition of LOOP to the beginning of its | |
3138 | header, and duplicate the part of the last iteration that gets disabled | |
3139 | to the exit of the loop. */ | |
3140 | transform_to_exit_first_loop (loop, reduction_list, nit); | |
3141 | } | |
3142 | ||
3143 | /* Generate initializations for reductions. */ | |
3144 | if (!reduction_list->is_empty ()) | |
3145 | reduction_list->traverse <class loop *, initialize_reductions> (loop); | |
3146 | ||
3147 | /* Eliminate the references to local variables from the loop. */ | |
3148 | gcc_assert (single_exit (loop)); | |
3149 | entry = loop_preheader_edge (loop); | |
3150 | exit = single_dom_exit (loop); | |
3151 | ||
3152 | /* This rewrites the body in terms of new variables. This has already | |
3153 | been done for oacc_kernels_p in pass_lower_omp/lower_omp (). */ | |
3154 | if (!oacc_kernels_p) | |
3155 | { | |
3156 | eliminate_local_variables (entry, exit); | |
3157 | /* In the old loop, move all variables non-local to the loop to a | |
3158 | structure and back, and create separate decls for the variables used in | |
3159 | loop. */ | |
3160 | separate_decls_in_region (entry, exit, reduction_list, &arg_struct, | |
3161 | &new_arg_struct, &clsn_data); | |
3162 | } | |
3163 | else | |
3164 | { | |
3165 | arg_struct = NULL_TREE; | |
3166 | new_arg_struct = NULL_TREE; | |
3167 | clsn_data.load = NULL_TREE; | |
3168 | clsn_data.load_bb = exit->dest; | |
3169 | clsn_data.store = NULL_TREE; | |
3170 | clsn_data.store_bb = NULL; | |
3171 | } | |
3172 | ||
3173 | /* Create the parallel constructs. */ | |
3174 | loc = UNKNOWN_LOCATION; | |
3175 | cond_stmt = last_stmt (loop->header); | |
3176 | if (cond_stmt) | |
3177 | loc = gimple_location (cond_stmt); | |
3178 | create_parallel_loop (loop, create_loop_fn (loc), arg_struct, new_arg_struct, | |
3179 | n_threads, loc, oacc_kernels_p); | |
3180 | if (!reduction_list->is_empty ()) | |
3181 | create_call_for_reduction (loop, reduction_list, &clsn_data); | |
3182 | ||
3183 | scev_reset (); | |
3184 | ||
3185 | /* Free loop bound estimations that could contain references to | |
3186 | removed statements. */ | |
3187 | free_numbers_of_iterations_estimates (cfun); | |
3188 | } | |
3189 | ||
3190 | /* Returns true when LOOP contains vector phi nodes. */ | |
3191 | ||
3192 | static bool | |
3193 | loop_has_vector_phi_nodes (class loop *loop ATTRIBUTE_UNUSED) | |
3194 | { | |
3195 | unsigned i; | |
3196 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
3197 | gphi_iterator gsi; | |
3198 | bool res = true; | |
3199 | ||
3200 | for (i = 0; i < loop->num_nodes; i++) | |
3201 | for (gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3202 | if (TREE_CODE (TREE_TYPE (PHI_RESULT (gsi.phi ()))) == VECTOR_TYPE) | |
3203 | goto end; | |
3204 | ||
3205 | res = false; | |
3206 | end: | |
3207 | free (bbs); | |
3208 | return res; | |
3209 | } | |
3210 | ||
3211 | /* Create a reduction_info struct, initialize it with REDUC_STMT | |
3212 | and PHI, insert it to the REDUCTION_LIST. */ | |
3213 | ||
3214 | static void | |
3215 | build_new_reduction (reduction_info_table_type *reduction_list, | |
3216 | gimple *reduc_stmt, gphi *phi) | |
3217 | { | |
3218 | reduction_info **slot; | |
3219 | struct reduction_info *new_reduction; | |
3220 | enum tree_code reduction_code; | |
3221 | ||
3222 | gcc_assert (reduc_stmt); | |
3223 | ||
3224 | if (gimple_code (reduc_stmt) == GIMPLE_PHI) | |
3225 | { | |
3226 | tree op1 = PHI_ARG_DEF (reduc_stmt, 0); | |
3227 | gimple *def1 = SSA_NAME_DEF_STMT (op1); | |
3228 | reduction_code = gimple_assign_rhs_code (def1); | |
3229 | } | |
3230 | else | |
3231 | reduction_code = gimple_assign_rhs_code (reduc_stmt); | |
3232 | /* Check for OpenMP supported reduction. */ | |
3233 | switch (reduction_code) | |
3234 | { | |
3235 | case PLUS_EXPR: | |
3236 | case MULT_EXPR: | |
3237 | case MAX_EXPR: | |
3238 | case MIN_EXPR: | |
3239 | case BIT_IOR_EXPR: | |
3240 | case BIT_XOR_EXPR: | |
3241 | case BIT_AND_EXPR: | |
3242 | case TRUTH_OR_EXPR: | |
3243 | case TRUTH_XOR_EXPR: | |
3244 | case TRUTH_AND_EXPR: | |
3245 | break; | |
3246 | default: | |
3247 | return; | |
3248 | } | |
3249 | ||
3250 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3251 | { | |
3252 | fprintf (dump_file, | |
3253 | "Detected reduction. reduction stmt is:\n"); | |
3254 | print_gimple_stmt (dump_file, reduc_stmt, 0); | |
3255 | fprintf (dump_file, "\n"); | |
3256 | } | |
3257 | ||
3258 | new_reduction = XCNEW (struct reduction_info); | |
3259 | ||
3260 | new_reduction->reduc_stmt = reduc_stmt; | |
3261 | new_reduction->reduc_phi = phi; | |
3262 | new_reduction->reduc_version = SSA_NAME_VERSION (gimple_phi_result (phi)); | |
3263 | new_reduction->reduction_code = reduction_code; | |
3264 | slot = reduction_list->find_slot (new_reduction, INSERT); | |
3265 | *slot = new_reduction; | |
3266 | } | |
3267 | ||
3268 | /* Callback for htab_traverse. Sets gimple_uid of reduc_phi stmts. */ | |
3269 | ||
3270 | int | |
3271 | set_reduc_phi_uids (reduction_info **slot, void *data ATTRIBUTE_UNUSED) | |
3272 | { | |
3273 | struct reduction_info *const red = *slot; | |
3274 | gimple_set_uid (red->reduc_phi, red->reduc_version); | |
3275 | return 1; | |
3276 | } | |
3277 | ||
3278 | /* Return true if the type of reduction performed by STMT_INFO is suitable | |
3279 | for this pass. */ | |
3280 | ||
3281 | static bool | |
3282 | valid_reduction_p (stmt_vec_info stmt_info) | |
3283 | { | |
3284 | /* Parallelization would reassociate the operation, which isn't | |
3285 | allowed for in-order reductions. */ | |
3286 | vect_reduction_type reduc_type = STMT_VINFO_REDUC_TYPE (stmt_info); | |
3287 | return reduc_type != FOLD_LEFT_REDUCTION; | |
3288 | } | |
3289 | ||
3290 | /* Detect all reductions in the LOOP, insert them into REDUCTION_LIST. */ | |
3291 | ||
3292 | static void | |
3293 | gather_scalar_reductions (loop_p loop, reduction_info_table_type *reduction_list) | |
3294 | { | |
3295 | gphi_iterator gsi; | |
3296 | loop_vec_info simple_loop_info; | |
3297 | auto_vec<gphi *, 4> double_reduc_phis; | |
3298 | auto_vec<gimple *, 4> double_reduc_stmts; | |
3299 | ||
3300 | vec_info_shared shared; | |
3301 | simple_loop_info = vect_analyze_loop_form (loop, &shared); | |
3302 | if (simple_loop_info == NULL) | |
3303 | goto gather_done; | |
3304 | ||
3305 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3306 | { | |
3307 | gphi *phi = gsi.phi (); | |
3308 | affine_iv iv; | |
3309 | tree res = PHI_RESULT (phi); | |
3310 | bool double_reduc; | |
3311 | ||
3312 | if (virtual_operand_p (res)) | |
3313 | continue; | |
3314 | ||
3315 | if (simple_iv (loop, loop, res, &iv, true)) | |
3316 | continue; | |
3317 | ||
3318 | stmt_vec_info reduc_stmt_info | |
3319 | = parloops_force_simple_reduction (simple_loop_info, | |
3320 | simple_loop_info->lookup_stmt (phi), | |
3321 | &double_reduc, true); | |
3322 | if (!reduc_stmt_info || !valid_reduction_p (reduc_stmt_info)) | |
3323 | continue; | |
3324 | ||
3325 | if (double_reduc) | |
3326 | { | |
3327 | if (loop->inner->inner != NULL) | |
3328 | continue; | |
3329 | ||
3330 | double_reduc_phis.safe_push (phi); | |
3331 | double_reduc_stmts.safe_push (reduc_stmt_info->stmt); | |
3332 | continue; | |
3333 | } | |
3334 | ||
3335 | build_new_reduction (reduction_list, reduc_stmt_info->stmt, phi); | |
3336 | } | |
3337 | delete simple_loop_info; | |
3338 | ||
3339 | if (!double_reduc_phis.is_empty ()) | |
3340 | { | |
3341 | vec_info_shared shared; | |
3342 | simple_loop_info = vect_analyze_loop_form (loop->inner, &shared); | |
3343 | if (simple_loop_info) | |
3344 | { | |
3345 | gphi *phi; | |
3346 | unsigned int i; | |
3347 | ||
3348 | FOR_EACH_VEC_ELT (double_reduc_phis, i, phi) | |
3349 | { | |
3350 | affine_iv iv; | |
3351 | tree res = PHI_RESULT (phi); | |
3352 | bool double_reduc; | |
3353 | ||
3354 | use_operand_p use_p; | |
3355 | gimple *inner_stmt; | |
3356 | bool single_use_p = single_imm_use (res, &use_p, &inner_stmt); | |
3357 | gcc_assert (single_use_p); | |
3358 | if (gimple_code (inner_stmt) != GIMPLE_PHI) | |
3359 | continue; | |
3360 | gphi *inner_phi = as_a <gphi *> (inner_stmt); | |
3361 | if (simple_iv (loop->inner, loop->inner, PHI_RESULT (inner_phi), | |
3362 | &iv, true)) | |
3363 | continue; | |
3364 | ||
3365 | stmt_vec_info inner_phi_info | |
3366 | = simple_loop_info->lookup_stmt (inner_phi); | |
3367 | stmt_vec_info inner_reduc_stmt_info | |
3368 | = parloops_force_simple_reduction (simple_loop_info, | |
3369 | inner_phi_info, | |
3370 | &double_reduc, true); | |
3371 | gcc_assert (!double_reduc); | |
3372 | if (!inner_reduc_stmt_info | |
3373 | || !valid_reduction_p (inner_reduc_stmt_info)) | |
3374 | continue; | |
3375 | ||
3376 | build_new_reduction (reduction_list, double_reduc_stmts[i], phi); | |
3377 | } | |
3378 | delete simple_loop_info; | |
3379 | } | |
3380 | } | |
3381 | ||
3382 | gather_done: | |
3383 | if (reduction_list->is_empty ()) | |
3384 | return; | |
3385 | ||
3386 | /* As gimple_uid is used by the vectorizer in between vect_analyze_loop_form | |
3387 | and delete simple_loop_info, we can set gimple_uid of reduc_phi stmts only | |
3388 | now. */ | |
3389 | basic_block bb; | |
3390 | FOR_EACH_BB_FN (bb, cfun) | |
3391 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3392 | gimple_set_uid (gsi_stmt (gsi), (unsigned int)-1); | |
3393 | reduction_list->traverse <void *, set_reduc_phi_uids> (NULL); | |
3394 | } | |
3395 | ||
3396 | /* Try to initialize NITER for code generation part. */ | |
3397 | ||
3398 | static bool | |
3399 | try_get_loop_niter (loop_p loop, class tree_niter_desc *niter) | |
3400 | { | |
3401 | edge exit = single_dom_exit (loop); | |
3402 | ||
3403 | gcc_assert (exit); | |
3404 | ||
3405 | /* We need to know # of iterations, and there should be no uses of values | |
3406 | defined inside loop outside of it, unless the values are invariants of | |
3407 | the loop. */ | |
3408 | if (!number_of_iterations_exit (loop, exit, niter, false)) | |
3409 | { | |
3410 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3411 | fprintf (dump_file, " FAILED: number of iterations not known\n"); | |
3412 | return false; | |
3413 | } | |
3414 | ||
3415 | return true; | |
3416 | } | |
3417 | ||
3418 | /* Return the default def of the first function argument. */ | |
3419 | ||
3420 | static tree | |
3421 | get_omp_data_i_param (void) | |
3422 | { | |
3423 | tree decl = DECL_ARGUMENTS (cfun->decl); | |
3424 | gcc_assert (DECL_CHAIN (decl) == NULL_TREE); | |
3425 | return ssa_default_def (cfun, decl); | |
3426 | } | |
3427 | ||
3428 | /* For PHI in loop header of LOOP, look for pattern: | |
3429 | ||
3430 | <bb preheader> | |
3431 | .omp_data_i = &.omp_data_arr; | |
3432 | addr = .omp_data_i->sum; | |
3433 | sum_a = *addr; | |
3434 | ||
3435 | <bb header>: | |
3436 | sum_b = PHI <sum_a (preheader), sum_c (latch)> | |
3437 | ||
3438 | and return addr. Otherwise, return NULL_TREE. */ | |
3439 | ||
3440 | static tree | |
3441 | find_reduc_addr (class loop *loop, gphi *phi) | |
3442 | { | |
3443 | edge e = loop_preheader_edge (loop); | |
3444 | tree arg = PHI_ARG_DEF_FROM_EDGE (phi, e); | |
3445 | gimple *stmt = SSA_NAME_DEF_STMT (arg); | |
3446 | if (!gimple_assign_single_p (stmt)) | |
3447 | return NULL_TREE; | |
3448 | tree memref = gimple_assign_rhs1 (stmt); | |
3449 | if (TREE_CODE (memref) != MEM_REF) | |
3450 | return NULL_TREE; | |
3451 | tree addr = TREE_OPERAND (memref, 0); | |
3452 | ||
3453 | gimple *stmt2 = SSA_NAME_DEF_STMT (addr); | |
3454 | if (!gimple_assign_single_p (stmt2)) | |
3455 | return NULL_TREE; | |
3456 | tree compref = gimple_assign_rhs1 (stmt2); | |
3457 | if (TREE_CODE (compref) != COMPONENT_REF) | |
3458 | return NULL_TREE; | |
3459 | tree addr2 = TREE_OPERAND (compref, 0); | |
3460 | if (TREE_CODE (addr2) != MEM_REF) | |
3461 | return NULL_TREE; | |
3462 | addr2 = TREE_OPERAND (addr2, 0); | |
3463 | if (TREE_CODE (addr2) != SSA_NAME | |
3464 | || addr2 != get_omp_data_i_param ()) | |
3465 | return NULL_TREE; | |
3466 | ||
3467 | return addr; | |
3468 | } | |
3469 | ||
3470 | /* Try to initialize REDUCTION_LIST for code generation part. | |
3471 | REDUCTION_LIST describes the reductions. */ | |
3472 | ||
3473 | static bool | |
3474 | try_create_reduction_list (loop_p loop, | |
3475 | reduction_info_table_type *reduction_list, | |
3476 | bool oacc_kernels_p) | |
3477 | { | |
3478 | edge exit = single_dom_exit (loop); | |
3479 | gphi_iterator gsi; | |
3480 | ||
3481 | gcc_assert (exit); | |
3482 | ||
3483 | /* Try to get rid of exit phis. */ | |
3484 | final_value_replacement_loop (loop); | |
3485 | ||
3486 | gather_scalar_reductions (loop, reduction_list); | |
3487 | ||
3488 | ||
3489 | for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3490 | { | |
3491 | gphi *phi = gsi.phi (); | |
3492 | struct reduction_info *red; | |
3493 | imm_use_iterator imm_iter; | |
3494 | use_operand_p use_p; | |
3495 | gimple *reduc_phi; | |
3496 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
3497 | ||
3498 | if (!virtual_operand_p (val)) | |
3499 | { | |
3500 | if (TREE_CODE (val) != SSA_NAME) | |
3501 | { | |
3502 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3503 | fprintf (dump_file, | |
3504 | " FAILED: exit PHI argument invariant.\n"); | |
3505 | return false; | |
3506 | } | |
3507 | ||
3508 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3509 | { | |
3510 | fprintf (dump_file, "phi is "); | |
3511 | print_gimple_stmt (dump_file, phi, 0); | |
3512 | fprintf (dump_file, "arg of phi to exit: value "); | |
3513 | print_generic_expr (dump_file, val); | |
3514 | fprintf (dump_file, " used outside loop\n"); | |
3515 | fprintf (dump_file, | |
3516 | " checking if it is part of reduction pattern:\n"); | |
3517 | } | |
3518 | if (reduction_list->is_empty ()) | |
3519 | { | |
3520 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3521 | fprintf (dump_file, | |
3522 | " FAILED: it is not a part of reduction.\n"); | |
3523 | return false; | |
3524 | } | |
3525 | reduc_phi = NULL; | |
3526 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, val) | |
3527 | { | |
3528 | if (!gimple_debug_bind_p (USE_STMT (use_p)) | |
3529 | && flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p)))) | |
3530 | { | |
3531 | reduc_phi = USE_STMT (use_p); | |
3532 | break; | |
3533 | } | |
3534 | } | |
3535 | red = reduction_phi (reduction_list, reduc_phi); | |
3536 | if (red == NULL) | |
3537 | { | |
3538 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3539 | fprintf (dump_file, | |
3540 | " FAILED: it is not a part of reduction.\n"); | |
3541 | return false; | |
3542 | } | |
3543 | if (red->keep_res != NULL) | |
3544 | { | |
3545 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3546 | fprintf (dump_file, | |
3547 | " FAILED: reduction has multiple exit phis.\n"); | |
3548 | return false; | |
3549 | } | |
3550 | red->keep_res = phi; | |
3551 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3552 | { | |
3553 | fprintf (dump_file, "reduction phi is "); | |
3554 | print_gimple_stmt (dump_file, red->reduc_phi, 0); | |
3555 | fprintf (dump_file, "reduction stmt is "); | |
3556 | print_gimple_stmt (dump_file, red->reduc_stmt, 0); | |
3557 | } | |
3558 | } | |
3559 | } | |
3560 | ||
3561 | /* The iterations of the loop may communicate only through bivs whose | |
3562 | iteration space can be distributed efficiently. */ | |
3563 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3564 | { | |
3565 | gphi *phi = gsi.phi (); | |
3566 | tree def = PHI_RESULT (phi); | |
3567 | affine_iv iv; | |
3568 | ||
3569 | if (!virtual_operand_p (def) && !simple_iv (loop, loop, def, &iv, true)) | |
3570 | { | |
3571 | struct reduction_info *red; | |
3572 | ||
3573 | red = reduction_phi (reduction_list, phi); | |
3574 | if (red == NULL) | |
3575 | { | |
3576 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
3577 | fprintf (dump_file, | |
3578 | " FAILED: scalar dependency between iterations\n"); | |
3579 | return false; | |
3580 | } | |
3581 | } | |
3582 | } | |
3583 | ||
3584 | if (oacc_kernels_p) | |
3585 | { | |
3586 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); | |
3587 | gsi_next (&gsi)) | |
3588 | { | |
3589 | gphi *phi = gsi.phi (); | |
3590 | tree def = PHI_RESULT (phi); | |
3591 | affine_iv iv; | |
3592 | ||
3593 | if (!virtual_operand_p (def) | |
3594 | && !simple_iv (loop, loop, def, &iv, true)) | |
3595 | { | |
3596 | tree addr = find_reduc_addr (loop, phi); | |
3597 | if (addr == NULL_TREE) | |
3598 | return false; | |
3599 | struct reduction_info *red = reduction_phi (reduction_list, phi); | |
3600 | red->reduc_addr = addr; | |
3601 | } | |
3602 | } | |
3603 | } | |
3604 | ||
3605 | return true; | |
3606 | } | |
3607 | ||
3608 | /* Return true if LOOP contains phis with ADDR_EXPR in args. */ | |
3609 | ||
3610 | static bool | |
3611 | loop_has_phi_with_address_arg (class loop *loop) | |
3612 | { | |
3613 | basic_block *bbs = get_loop_body (loop); | |
3614 | bool res = false; | |
3615 | ||
3616 | unsigned i, j; | |
3617 | gphi_iterator gsi; | |
3618 | for (i = 0; i < loop->num_nodes; i++) | |
3619 | for (gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) | |
3620 | { | |
3621 | gphi *phi = gsi.phi (); | |
3622 | for (j = 0; j < gimple_phi_num_args (phi); j++) | |
3623 | { | |
3624 | tree arg = gimple_phi_arg_def (phi, j); | |
3625 | if (TREE_CODE (arg) == ADDR_EXPR) | |
3626 | { | |
3627 | /* This should be handled by eliminate_local_variables, but that | |
3628 | function currently ignores phis. */ | |
3629 | res = true; | |
3630 | goto end; | |
3631 | } | |
3632 | } | |
3633 | } | |
3634 | end: | |
3635 | free (bbs); | |
3636 | ||
3637 | return res; | |
3638 | } | |
3639 | ||
3640 | /* Return true if memory ref REF (corresponding to the stmt at GSI in | |
3641 | REGIONS_BB[I]) conflicts with the statements in REGIONS_BB[I] after gsi, | |
3642 | or the statements in REGIONS_BB[I + n]. REF_IS_STORE indicates if REF is a | |
3643 | store. Ignore conflicts with SKIP_STMT. */ | |
3644 | ||
3645 | static bool | |
3646 | ref_conflicts_with_region (gimple_stmt_iterator gsi, ao_ref *ref, | |
3647 | bool ref_is_store, vec<basic_block> region_bbs, | |
3648 | unsigned int i, gimple *skip_stmt) | |
3649 | { | |
3650 | basic_block bb = region_bbs[i]; | |
3651 | gsi_next (&gsi); | |
3652 | ||
3653 | while (true) | |
3654 | { | |
3655 | for (; !gsi_end_p (gsi); | |
3656 | gsi_next (&gsi)) | |
3657 | { | |
3658 | gimple *stmt = gsi_stmt (gsi); | |
3659 | if (stmt == skip_stmt) | |
3660 | { | |
3661 | if (dump_file) | |
3662 | { | |
3663 | fprintf (dump_file, "skipping reduction store: "); | |
3664 | print_gimple_stmt (dump_file, stmt, 0); | |
3665 | } | |
3666 | continue; | |
3667 | } | |
3668 | ||
3669 | if (!gimple_vdef (stmt) | |
3670 | && !gimple_vuse (stmt)) | |
3671 | continue; | |
3672 | ||
3673 | if (gimple_code (stmt) == GIMPLE_RETURN) | |
3674 | continue; | |
3675 | ||
3676 | if (ref_is_store) | |
3677 | { | |
3678 | if (ref_maybe_used_by_stmt_p (stmt, ref)) | |
3679 | { | |
3680 | if (dump_file) | |
3681 | { | |
3682 | fprintf (dump_file, "Stmt "); | |
3683 | print_gimple_stmt (dump_file, stmt, 0); | |
3684 | } | |
3685 | return true; | |
3686 | } | |
3687 | } | |
3688 | else | |
3689 | { | |
3690 | if (stmt_may_clobber_ref_p_1 (stmt, ref)) | |
3691 | { | |
3692 | if (dump_file) | |
3693 | { | |
3694 | fprintf (dump_file, "Stmt "); | |
3695 | print_gimple_stmt (dump_file, stmt, 0); | |
3696 | } | |
3697 | return true; | |
3698 | } | |
3699 | } | |
3700 | } | |
3701 | i++; | |
3702 | if (i == region_bbs.length ()) | |
3703 | break; | |
3704 | bb = region_bbs[i]; | |
3705 | gsi = gsi_start_bb (bb); | |
3706 | } | |
3707 | ||
3708 | return false; | |
3709 | } | |
3710 | ||
3711 | /* Return true if the bbs in REGION_BBS but not in in_loop_bbs can be executed | |
3712 | in parallel with REGION_BBS containing the loop. Return the stores of | |
3713 | reduction results in REDUCTION_STORES. */ | |
3714 | ||
3715 | static bool | |
3716 | oacc_entry_exit_ok_1 (bitmap in_loop_bbs, vec<basic_block> region_bbs, | |
3717 | reduction_info_table_type *reduction_list, | |
3718 | bitmap reduction_stores) | |
3719 | { | |
3720 | tree omp_data_i = get_omp_data_i_param (); | |
3721 | ||
3722 | unsigned i; | |
3723 | basic_block bb; | |
3724 | FOR_EACH_VEC_ELT (region_bbs, i, bb) | |
3725 | { | |
3726 | if (bitmap_bit_p (in_loop_bbs, bb->index)) | |
3727 | continue; | |
3728 | ||
3729 | gimple_stmt_iterator gsi; | |
3730 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); | |
3731 | gsi_next (&gsi)) | |
3732 | { | |
3733 | gimple *stmt = gsi_stmt (gsi); | |
3734 | gimple *skip_stmt = NULL; | |
3735 | ||
3736 | if (is_gimple_debug (stmt) | |
3737 | || gimple_code (stmt) == GIMPLE_COND) | |
3738 | continue; | |
3739 | ||
3740 | ao_ref ref; | |
3741 | bool ref_is_store = false; | |
3742 | if (gimple_assign_load_p (stmt)) | |
3743 | { | |
3744 | tree rhs = gimple_assign_rhs1 (stmt); | |
3745 | tree base = get_base_address (rhs); | |
3746 | if (TREE_CODE (base) == MEM_REF | |
3747 | && operand_equal_p (TREE_OPERAND (base, 0), omp_data_i, 0)) | |
3748 | continue; | |
3749 | ||
3750 | tree lhs = gimple_assign_lhs (stmt); | |
3751 | if (TREE_CODE (lhs) == SSA_NAME | |
3752 | && has_single_use (lhs)) | |
3753 | { | |
3754 | use_operand_p use_p; | |
3755 | gimple *use_stmt; | |
3756 | struct reduction_info *red; | |
3757 | single_imm_use (lhs, &use_p, &use_stmt); | |
3758 | if (gimple_code (use_stmt) == GIMPLE_PHI | |
3759 | && (red = reduction_phi (reduction_list, use_stmt))) | |
3760 | { | |
3761 | tree val = PHI_RESULT (red->keep_res); | |
3762 | if (has_single_use (val)) | |
3763 | { | |
3764 | single_imm_use (val, &use_p, &use_stmt); | |
3765 | if (gimple_store_p (use_stmt)) | |
3766 | { | |
3767 | unsigned int id | |
3768 | = SSA_NAME_VERSION (gimple_vdef (use_stmt)); | |
3769 | bitmap_set_bit (reduction_stores, id); | |
3770 | skip_stmt = use_stmt; | |
3771 | if (dump_file) | |
3772 | { | |
3773 | fprintf (dump_file, "found reduction load: "); | |
3774 | print_gimple_stmt (dump_file, stmt, 0); | |
3775 | } | |
3776 | } | |
3777 | } | |
3778 | } | |
3779 | } | |
3780 | ||
3781 | ao_ref_init (&ref, rhs); | |
3782 | } | |
3783 | else if (gimple_store_p (stmt)) | |
3784 | { | |
3785 | ao_ref_init (&ref, gimple_assign_lhs (stmt)); | |
3786 | ref_is_store = true; | |
3787 | } | |
3788 | else if (gimple_code (stmt) == GIMPLE_OMP_RETURN) | |
3789 | continue; | |
3790 | else if (!gimple_has_side_effects (stmt) | |
3791 | && !gimple_could_trap_p (stmt) | |
3792 | && !stmt_could_throw_p (cfun, stmt) | |
3793 | && !gimple_vdef (stmt) | |
3794 | && !gimple_vuse (stmt)) | |
3795 | continue; | |
3796 | else if (gimple_call_internal_p (stmt, IFN_GOACC_DIM_POS)) | |
3797 | continue; | |
3798 | else if (gimple_code (stmt) == GIMPLE_RETURN) | |
3799 | continue; | |
3800 | else | |
3801 | { | |
3802 | if (dump_file) | |
3803 | { | |
3804 | fprintf (dump_file, "Unhandled stmt in entry/exit: "); | |
3805 | print_gimple_stmt (dump_file, stmt, 0); | |
3806 | } | |
3807 | return false; | |
3808 | } | |
3809 | ||
3810 | if (ref_conflicts_with_region (gsi, &ref, ref_is_store, region_bbs, | |
3811 | i, skip_stmt)) | |
3812 | { | |
3813 | if (dump_file) | |
3814 | { | |
3815 | fprintf (dump_file, "conflicts with entry/exit stmt: "); | |
3816 | print_gimple_stmt (dump_file, stmt, 0); | |
3817 | } | |
3818 | return false; | |
3819 | } | |
3820 | } | |
3821 | } | |
3822 | ||
3823 | return true; | |
3824 | } | |
3825 | ||
3826 | /* Find stores inside REGION_BBS and outside IN_LOOP_BBS, and guard them with | |
3827 | gang_pos == 0, except when the stores are REDUCTION_STORES. Return true | |
3828 | if any changes were made. */ | |
3829 | ||
3830 | static bool | |
3831 | oacc_entry_exit_single_gang (bitmap in_loop_bbs, vec<basic_block> region_bbs, | |
3832 | bitmap reduction_stores) | |
3833 | { | |
3834 | tree gang_pos = NULL_TREE; | |
3835 | bool changed = false; | |
3836 | ||
3837 | unsigned i; | |
3838 | basic_block bb; | |
3839 | FOR_EACH_VEC_ELT (region_bbs, i, bb) | |
3840 | { | |
3841 | if (bitmap_bit_p (in_loop_bbs, bb->index)) | |
3842 | continue; | |
3843 | ||
3844 | gimple_stmt_iterator gsi; | |
3845 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) | |
3846 | { | |
3847 | gimple *stmt = gsi_stmt (gsi); | |
3848 | ||
3849 | if (!gimple_store_p (stmt)) | |
3850 | { | |
3851 | /* Update gsi to point to next stmt. */ | |
3852 | gsi_next (&gsi); | |
3853 | continue; | |
3854 | } | |
3855 | ||
3856 | if (bitmap_bit_p (reduction_stores, | |
3857 | SSA_NAME_VERSION (gimple_vdef (stmt)))) | |
3858 | { | |
3859 | if (dump_file) | |
3860 | { | |
3861 | fprintf (dump_file, | |
3862 | "skipped reduction store for single-gang" | |
3863 | " neutering: "); | |
3864 | print_gimple_stmt (dump_file, stmt, 0); | |
3865 | } | |
3866 | ||
3867 | /* Update gsi to point to next stmt. */ | |
3868 | gsi_next (&gsi); | |
3869 | continue; | |
3870 | } | |
3871 | ||
3872 | changed = true; | |
3873 | ||
3874 | if (gang_pos == NULL_TREE) | |
3875 | { | |
3876 | tree arg = build_int_cst (integer_type_node, GOMP_DIM_GANG); | |
3877 | gcall *gang_single | |
3878 | = gimple_build_call_internal (IFN_GOACC_DIM_POS, 1, arg); | |
3879 | gang_pos = make_ssa_name (integer_type_node); | |
3880 | gimple_call_set_lhs (gang_single, gang_pos); | |
3881 | gimple_stmt_iterator start | |
3882 | = gsi_start_bb (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))); | |
3883 | tree vuse = ssa_default_def (cfun, gimple_vop (cfun)); | |
3884 | gimple_set_vuse (gang_single, vuse); | |
3885 | gsi_insert_before (&start, gang_single, GSI_SAME_STMT); | |
3886 | } | |
3887 | ||
3888 | if (dump_file) | |
3889 | { | |
3890 | fprintf (dump_file, | |
3891 | "found store that needs single-gang neutering: "); | |
3892 | print_gimple_stmt (dump_file, stmt, 0); | |
3893 | } | |
3894 | ||
3895 | { | |
3896 | /* Split block before store. */ | |
3897 | gimple_stmt_iterator gsi2 = gsi; | |
3898 | gsi_prev (&gsi2); | |
3899 | edge e; | |
3900 | if (gsi_end_p (gsi2)) | |
3901 | { | |
3902 | e = split_block_after_labels (bb); | |
3903 | gsi2 = gsi_last_bb (bb); | |
3904 | } | |
3905 | else | |
3906 | e = split_block (bb, gsi_stmt (gsi2)); | |
3907 | basic_block bb2 = e->dest; | |
3908 | ||
3909 | /* Split block after store. */ | |
3910 | gimple_stmt_iterator gsi3 = gsi_start_bb (bb2); | |
3911 | edge e2 = split_block (bb2, gsi_stmt (gsi3)); | |
3912 | basic_block bb3 = e2->dest; | |
3913 | ||
3914 | gimple *cond | |
3915 | = gimple_build_cond (EQ_EXPR, gang_pos, integer_zero_node, | |
3916 | NULL_TREE, NULL_TREE); | |
3917 | gsi_insert_after (&gsi2, cond, GSI_NEW_STMT); | |
3918 | ||
3919 | edge e3 = make_edge (bb, bb3, EDGE_FALSE_VALUE); | |
3920 | /* FIXME: What is the probability? */ | |
3921 | e3->probability = profile_probability::guessed_never (); | |
3922 | e->flags = EDGE_TRUE_VALUE; | |
3923 | ||
3924 | tree vdef = gimple_vdef (stmt); | |
3925 | tree vuse = gimple_vuse (stmt); | |
3926 | ||
3927 | tree phi_res = copy_ssa_name (vdef); | |
3928 | gphi *new_phi = create_phi_node (phi_res, bb3); | |
3929 | replace_uses_by (vdef, phi_res); | |
3930 | add_phi_arg (new_phi, vuse, e3, UNKNOWN_LOCATION); | |
3931 | add_phi_arg (new_phi, vdef, e2, UNKNOWN_LOCATION); | |
3932 | ||
3933 | /* Update gsi to point to next stmt. */ | |
3934 | bb = bb3; | |
3935 | gsi = gsi_start_bb (bb); | |
3936 | } | |
3937 | } | |
3938 | } | |
3939 | ||
3940 | return changed; | |
3941 | } | |
3942 | ||
3943 | /* Return true if the statements before and after the LOOP can be executed in | |
3944 | parallel with the function containing the loop. Resolve conflicting stores | |
3945 | outside LOOP by guarding them such that only a single gang executes them. */ | |
3946 | ||
3947 | static bool | |
3948 | oacc_entry_exit_ok (class loop *loop, | |
3949 | reduction_info_table_type *reduction_list) | |
3950 | { | |
3951 | basic_block *loop_bbs = get_loop_body_in_dom_order (loop); | |
3952 | vec<basic_block> region_bbs | |
3953 | = get_all_dominated_blocks (CDI_DOMINATORS, ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
3954 | ||
3955 | bitmap in_loop_bbs = BITMAP_ALLOC (NULL); | |
3956 | bitmap_clear (in_loop_bbs); | |
3957 | for (unsigned int i = 0; i < loop->num_nodes; i++) | |
3958 | bitmap_set_bit (in_loop_bbs, loop_bbs[i]->index); | |
3959 | ||
3960 | bitmap reduction_stores = BITMAP_ALLOC (NULL); | |
3961 | bool res = oacc_entry_exit_ok_1 (in_loop_bbs, region_bbs, reduction_list, | |
3962 | reduction_stores); | |
3963 | ||
3964 | if (res) | |
3965 | { | |
3966 | bool changed = oacc_entry_exit_single_gang (in_loop_bbs, region_bbs, | |
3967 | reduction_stores); | |
3968 | if (changed) | |
3969 | { | |
3970 | free_dominance_info (CDI_DOMINATORS); | |
3971 | calculate_dominance_info (CDI_DOMINATORS); | |
3972 | } | |
3973 | } | |
3974 | ||
3975 | region_bbs.release (); | |
3976 | free (loop_bbs); | |
3977 | ||
3978 | BITMAP_FREE (in_loop_bbs); | |
3979 | BITMAP_FREE (reduction_stores); | |
3980 | ||
3981 | return res; | |
3982 | } | |
3983 | ||
3984 | /* Detect parallel loops and generate parallel code using libgomp | |
3985 | primitives. Returns true if some loop was parallelized, false | |
3986 | otherwise. */ | |
3987 | ||
3988 | static bool | |
3989 | parallelize_loops (bool oacc_kernels_p) | |
3990 | { | |
3991 | unsigned n_threads; | |
3992 | bool changed = false; | |
3993 | class loop *loop; | |
3994 | class loop *skip_loop = NULL; | |
3995 | class tree_niter_desc niter_desc; | |
3996 | struct obstack parloop_obstack; | |
3997 | HOST_WIDE_INT estimated; | |
3998 | ||
3999 | /* Do not parallelize loops in the functions created by parallelization. */ | |
4000 | if (!oacc_kernels_p | |
4001 | && parallelized_function_p (cfun->decl)) | |
4002 | return false; | |
4003 | ||
4004 | /* Do not parallelize loops in offloaded functions. */ | |
4005 | if (!oacc_kernels_p | |
4006 | && oacc_get_fn_attrib (cfun->decl) != NULL) | |
4007 | return false; | |
4008 | ||
4009 | if (cfun->has_nonlocal_label) | |
4010 | return false; | |
4011 | ||
4012 | /* For OpenACC kernels, n_threads will be determined later; otherwise, it's | |
4013 | the argument to -ftree-parallelize-loops. */ | |
4014 | if (oacc_kernels_p) | |
4015 | n_threads = 0; | |
4016 | else | |
4017 | n_threads = flag_tree_parallelize_loops; | |
4018 | ||
4019 | gcc_obstack_init (&parloop_obstack); | |
4020 | reduction_info_table_type reduction_list (10); | |
4021 | ||
4022 | calculate_dominance_info (CDI_DOMINATORS); | |
4023 | ||
4024 | FOR_EACH_LOOP (loop, 0) | |
4025 | { | |
4026 | if (loop == skip_loop) | |
4027 | { | |
4028 | if (!loop->in_oacc_kernels_region | |
4029 | && dump_file && (dump_flags & TDF_DETAILS)) | |
4030 | fprintf (dump_file, | |
4031 | "Skipping loop %d as inner loop of parallelized loop\n", | |
4032 | loop->num); | |
4033 | ||
4034 | skip_loop = loop->inner; | |
4035 | continue; | |
4036 | } | |
4037 | else | |
4038 | skip_loop = NULL; | |
4039 | ||
4040 | reduction_list.empty (); | |
4041 | ||
4042 | if (oacc_kernels_p) | |
4043 | { | |
4044 | if (!loop->in_oacc_kernels_region) | |
4045 | continue; | |
4046 | ||
4047 | /* Don't try to parallelize inner loops in an oacc kernels region. */ | |
4048 | if (loop->inner) | |
4049 | skip_loop = loop->inner; | |
4050 | ||
4051 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4052 | fprintf (dump_file, | |
4053 | "Trying loop %d with header bb %d in oacc kernels" | |
4054 | " region\n", loop->num, loop->header->index); | |
4055 | } | |
4056 | ||
4057 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4058 | { | |
4059 | fprintf (dump_file, "Trying loop %d as candidate\n",loop->num); | |
4060 | if (loop->inner) | |
4061 | fprintf (dump_file, "loop %d is not innermost\n",loop->num); | |
4062 | else | |
4063 | fprintf (dump_file, "loop %d is innermost\n",loop->num); | |
4064 | } | |
4065 | ||
4066 | if (!single_dom_exit (loop)) | |
4067 | { | |
4068 | ||
4069 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
4070 | fprintf (dump_file, "loop is !single_dom_exit\n"); | |
4071 | ||
4072 | continue; | |
4073 | } | |
4074 | ||
4075 | if (/* And of course, the loop must be parallelizable. */ | |
4076 | !can_duplicate_loop_p (loop) | |
4077 | || loop_has_blocks_with_irreducible_flag (loop) | |
4078 | || (loop_preheader_edge (loop)->src->flags & BB_IRREDUCIBLE_LOOP) | |
4079 | /* FIXME: the check for vector phi nodes could be removed. */ | |
4080 | || loop_has_vector_phi_nodes (loop)) | |
4081 | continue; | |
4082 | ||
4083 | estimated = estimated_loop_iterations_int (loop); | |
4084 | if (estimated == -1) | |
4085 | estimated = get_likely_max_loop_iterations_int (loop); | |
4086 | /* FIXME: Bypass this check as graphite doesn't update the | |
4087 | count and frequency correctly now. */ | |
4088 | if (!flag_loop_parallelize_all | |
4089 | && !oacc_kernels_p | |
4090 | && ((estimated != -1 | |
4091 | && (estimated | |
4092 | < ((HOST_WIDE_INT) n_threads | |
4093 | * (loop->inner ? 2 : MIN_PER_THREAD) - 1))) | |
4094 | /* Do not bother with loops in cold areas. */ | |
4095 | || optimize_loop_nest_for_size_p (loop))) | |
4096 | continue; | |
4097 | ||
4098 | if (!try_get_loop_niter (loop, &niter_desc)) | |
4099 | continue; | |
4100 | ||
4101 | if (!try_create_reduction_list (loop, &reduction_list, oacc_kernels_p)) | |
4102 | continue; | |
4103 | ||
4104 | if (loop_has_phi_with_address_arg (loop)) | |
4105 | continue; | |
4106 | ||
4107 | if (!loop->can_be_parallel | |
4108 | && !loop_parallel_p (loop, &parloop_obstack)) | |
4109 | continue; | |
4110 | ||
4111 | if (oacc_kernels_p | |
4112 | && !oacc_entry_exit_ok (loop, &reduction_list)) | |
4113 | { | |
4114 | if (dump_file) | |
4115 | fprintf (dump_file, "entry/exit not ok: FAILED\n"); | |
4116 | continue; | |
4117 | } | |
4118 | ||
4119 | changed = true; | |
4120 | skip_loop = loop->inner; | |
4121 | ||
4122 | if (dump_enabled_p ()) | |
4123 | { | |
4124 | dump_user_location_t loop_loc = find_loop_location (loop); | |
4125 | if (loop->inner) | |
4126 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loop_loc, | |
4127 | "parallelizing outer loop %d\n", loop->num); | |
4128 | else | |
4129 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loop_loc, | |
4130 | "parallelizing inner loop %d\n", loop->num); | |
4131 | } | |
4132 | ||
4133 | gen_parallel_loop (loop, &reduction_list, | |
4134 | n_threads, &niter_desc, oacc_kernels_p); | |
4135 | } | |
4136 | ||
4137 | obstack_free (&parloop_obstack, NULL); | |
4138 | ||
4139 | /* Parallelization will cause new function calls to be inserted through | |
4140 | which local variables will escape. Reset the points-to solution | |
4141 | for ESCAPED. */ | |
4142 | if (changed) | |
4143 | pt_solution_reset (&cfun->gimple_df->escaped); | |
4144 | ||
4145 | return changed; | |
4146 | } | |
4147 | ||
4148 | /* Parallelization. */ | |
4149 | ||
4150 | namespace { | |
4151 | ||
4152 | const pass_data pass_data_parallelize_loops = | |
4153 | { | |
4154 | GIMPLE_PASS, /* type */ | |
4155 | "parloops", /* name */ | |
4156 | OPTGROUP_LOOP, /* optinfo_flags */ | |
4157 | TV_TREE_PARALLELIZE_LOOPS, /* tv_id */ | |
4158 | ( PROP_cfg | PROP_ssa ), /* properties_required */ | |
4159 | 0, /* properties_provided */ | |
4160 | 0, /* properties_destroyed */ | |
4161 | 0, /* todo_flags_start */ | |
4162 | 0, /* todo_flags_finish */ | |
4163 | }; | |
4164 | ||
4165 | class pass_parallelize_loops : public gimple_opt_pass | |
4166 | { | |
4167 | public: | |
4168 | pass_parallelize_loops (gcc::context *ctxt) | |
4169 | : gimple_opt_pass (pass_data_parallelize_loops, ctxt), | |
4170 | oacc_kernels_p (false) | |
4171 | {} | |
4172 | ||
4173 | /* opt_pass methods: */ | |
4174 | virtual bool gate (function *) | |
4175 | { | |
4176 | if (oacc_kernels_p) | |
4177 | return flag_openacc; | |
4178 | else | |
4179 | return flag_tree_parallelize_loops > 1; | |
4180 | } | |
4181 | virtual unsigned int execute (function *); | |
4182 | opt_pass * clone () { return new pass_parallelize_loops (m_ctxt); } | |
4183 | void set_pass_param (unsigned int n, bool param) | |
4184 | { | |
4185 | gcc_assert (n == 0); | |
4186 | oacc_kernels_p = param; | |
4187 | } | |
4188 | ||
4189 | private: | |
4190 | bool oacc_kernels_p; | |
4191 | }; // class pass_parallelize_loops | |
4192 | ||
4193 | unsigned | |
4194 | pass_parallelize_loops::execute (function *fun) | |
4195 | { | |
4196 | tree nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS); | |
4197 | if (nthreads == NULL_TREE) | |
4198 | return 0; | |
4199 | ||
4200 | bool in_loop_pipeline = scev_initialized_p (); | |
4201 | if (!in_loop_pipeline) | |
4202 | loop_optimizer_init (LOOPS_NORMAL | |
4203 | | LOOPS_HAVE_RECORDED_EXITS); | |
4204 | ||
4205 | if (number_of_loops (fun) <= 1) | |
4206 | return 0; | |
4207 | ||
4208 | if (!in_loop_pipeline) | |
4209 | { | |
4210 | rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa); | |
4211 | scev_initialize (); | |
4212 | } | |
4213 | ||
4214 | unsigned int todo = 0; | |
4215 | if (parallelize_loops (oacc_kernels_p)) | |
4216 | { | |
4217 | fun->curr_properties &= ~(PROP_gimple_eomp); | |
4218 | ||
4219 | checking_verify_loop_structure (); | |
4220 | ||
4221 | todo |= TODO_update_ssa; | |
4222 | } | |
4223 | ||
4224 | if (!in_loop_pipeline) | |
4225 | { | |
4226 | scev_finalize (); | |
4227 | loop_optimizer_finalize (); | |
4228 | } | |
4229 | ||
4230 | return todo; | |
4231 | } | |
4232 | ||
4233 | } // anon namespace | |
4234 | ||
4235 | gimple_opt_pass * | |
4236 | make_pass_parallelize_loops (gcc::context *ctxt) | |
4237 | { | |
4238 | return new pass_parallelize_loops (ctxt); | |
4239 | } |