]>
Commit | Line | Data |
---|---|---|
5f40b3cb | 1 | /* Loop autoparallelization. |
6da7fc87 | 2 | Copyright (C) 2006, 2007, 2008, 2009 Free Software Foundation, Inc. |
5f40b3cb ZD |
3 | Contributed by Sebastian Pop <pop@cri.ensmp.fr> and |
4 | Zdenek Dvorak <dvorakz@suse.cz>. | |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify it under | |
9 | the terms of the GNU General Public License as published by the Free | |
6da7fc87 | 10 | Software Foundation; either version 3, or (at your option) any later |
5f40b3cb ZD |
11 | version. |
12 | ||
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
6da7fc87 NC |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ | |
5f40b3cb ZD |
21 | |
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
25 | #include "tm.h" | |
26 | #include "tree.h" | |
27 | #include "rtl.h" | |
28 | #include "tree-flow.h" | |
29 | #include "cfgloop.h" | |
30 | #include "ggc.h" | |
31 | #include "tree-data-ref.h" | |
32 | #include "diagnostic.h" | |
33 | #include "tree-pass.h" | |
34 | #include "tree-scalar-evolution.h" | |
35 | #include "hashtab.h" | |
36 | #include "langhooks.h" | |
a509ebb5 | 37 | #include "tree-vectorizer.h" |
5f40b3cb ZD |
38 | |
39 | /* This pass tries to distribute iterations of loops into several threads. | |
40 | The implementation is straightforward -- for each loop we test whether its | |
41 | iterations are independent, and if it is the case (and some additional | |
42 | conditions regarding profitability and correctness are satisfied), we | |
726a989a RB |
43 | add GIMPLE_OMP_PARALLEL and GIMPLE_OMP_FOR codes and let omp expansion |
44 | machinery do its job. | |
5f40b3cb ZD |
45 | |
46 | The most of the complexity is in bringing the code into shape expected | |
47 | by the omp expanders: | |
726a989a RB |
48 | -- for GIMPLE_OMP_FOR, ensuring that the loop has only one induction |
49 | variable and that the exit test is at the start of the loop body | |
50 | -- for GIMPLE_OMP_PARALLEL, replacing the references to local addressable | |
5f40b3cb ZD |
51 | variables by accesses through pointers, and breaking up ssa chains |
52 | by storing the values incoming to the parallelized loop to a structure | |
53 | passed to the new function as an argument (something similar is done | |
54 | in omp gimplification, unfortunately only a small part of the code | |
55 | can be shared). | |
56 | ||
57 | TODO: | |
58 | -- if there are several parallelizable loops in a function, it may be | |
59 | possible to generate the threads just once (using synchronization to | |
60 | ensure that cross-loop dependences are obeyed). | |
61 | -- handling of common scalar dependence patterns (accumulation, ...) | |
62 | -- handling of non-innermost loops */ | |
63 | ||
a509ebb5 RL |
64 | /* |
65 | Reduction handling: | |
66 | currently we use vect_is_simple_reduction() to detect reduction patterns. | |
67 | The code transformation will be introduced by an example. | |
68 | ||
0eb7e7aa | 69 | |
a509ebb5 RL |
70 | parloop |
71 | { | |
72 | int sum=1; | |
73 | ||
0eb7e7aa | 74 | for (i = 0; i < N; i++) |
a509ebb5 RL |
75 | { |
76 | x[i] = i + 3; | |
77 | sum+=x[i]; | |
78 | } | |
79 | } | |
80 | ||
0eb7e7aa | 81 | gimple-like code: |
a509ebb5 RL |
82 | header_bb: |
83 | ||
0eb7e7aa RL |
84 | # sum_29 = PHI <sum_11(5), 1(3)> |
85 | # i_28 = PHI <i_12(5), 0(3)> | |
86 | D.1795_8 = i_28 + 3; | |
87 | x[i_28] = D.1795_8; | |
88 | sum_11 = D.1795_8 + sum_29; | |
89 | i_12 = i_28 + 1; | |
90 | if (N_6(D) > i_12) | |
91 | goto header_bb; | |
92 | ||
a509ebb5 RL |
93 | |
94 | exit_bb: | |
95 | ||
0eb7e7aa RL |
96 | # sum_21 = PHI <sum_11(4)> |
97 | printf (&"%d"[0], sum_21); | |
a509ebb5 RL |
98 | |
99 | ||
100 | after reduction transformation (only relevant parts): | |
101 | ||
102 | parloop | |
103 | { | |
104 | ||
105 | .... | |
106 | ||
0eb7e7aa | 107 | |
fa10beec | 108 | # Storing the initial value given by the user. # |
0eb7e7aa | 109 | |
ae0bce62 | 110 | .paral_data_store.32.sum.27 = 1; |
0eb7e7aa | 111 | |
0eb7e7aa | 112 | #pragma omp parallel num_threads(4) |
a509ebb5 | 113 | |
0eb7e7aa | 114 | #pragma omp for schedule(static) |
ae0bce62 RL |
115 | |
116 | # The neutral element corresponding to the particular | |
117 | reduction's operation, e.g. 0 for PLUS_EXPR, | |
118 | 1 for MULT_EXPR, etc. replaces the user's initial value. # | |
119 | ||
120 | # sum.27_29 = PHI <sum.27_11, 0> | |
121 | ||
0eb7e7aa | 122 | sum.27_11 = D.1827_8 + sum.27_29; |
ae0bce62 | 123 | |
726a989a | 124 | GIMPLE_OMP_CONTINUE |
a509ebb5 | 125 | |
0eb7e7aa RL |
126 | # Adding this reduction phi is done at create_phi_for_local_result() # |
127 | # sum.27_56 = PHI <sum.27_11, 0> | |
726a989a | 128 | GIMPLE_OMP_RETURN |
0eb7e7aa RL |
129 | |
130 | # Creating the atomic operation is done at | |
131 | create_call_for_reduction_1() # | |
a509ebb5 | 132 | |
0eb7e7aa RL |
133 | #pragma omp atomic_load |
134 | D.1839_59 = *&.paral_data_load.33_51->reduction.23; | |
135 | D.1840_60 = sum.27_56 + D.1839_59; | |
136 | #pragma omp atomic_store (D.1840_60); | |
a509ebb5 | 137 | |
726a989a | 138 | GIMPLE_OMP_RETURN |
0eb7e7aa RL |
139 | |
140 | # collecting the result after the join of the threads is done at | |
141 | create_loads_for_reductions(). | |
ae0bce62 RL |
142 | The value computed by the threads is loaded from the |
143 | shared struct. # | |
144 | ||
0eb7e7aa RL |
145 | |
146 | .paral_data_load.33_52 = &.paral_data_store.32; | |
ae0bce62 | 147 | sum_37 = .paral_data_load.33_52->sum.27; |
0eb7e7aa RL |
148 | sum_43 = D.1795_41 + sum_37; |
149 | ||
150 | exit bb: | |
151 | # sum_21 = PHI <sum_43, sum_26> | |
152 | printf (&"%d"[0], sum_21); | |
153 | ||
154 | ... | |
155 | ||
a509ebb5 RL |
156 | } |
157 | ||
158 | */ | |
159 | ||
5f40b3cb ZD |
160 | /* Minimal number of iterations of a loop that should be executed in each |
161 | thread. */ | |
162 | #define MIN_PER_THREAD 100 | |
163 | ||
a509ebb5 RL |
164 | /* Element of the hashtable, representing a |
165 | reduction in the current loop. */ | |
166 | struct reduction_info | |
167 | { | |
726a989a RB |
168 | gimple reduc_stmt; /* reduction statement. */ |
169 | gimple reduc_phi; /* The phi node defining the reduction. */ | |
170 | enum tree_code reduction_code;/* code for the reduction operation. */ | |
171 | gimple keep_res; /* The PHI_RESULT of this phi is the resulting value | |
a509ebb5 | 172 | of the reduction variable when existing the loop. */ |
ae0bce62 | 173 | tree initial_value; /* The initial value of the reduction var before entering the loop. */ |
a509ebb5 | 174 | tree field; /* the name of the field in the parloop data structure intended for reduction. */ |
a509ebb5 | 175 | tree init; /* reduction initialization value. */ |
726a989a | 176 | gimple new_phi; /* (helper field) Newly created phi node whose result |
a509ebb5 RL |
177 | will be passed to the atomic operation. Represents |
178 | the local result each thread computed for the reduction | |
179 | operation. */ | |
180 | }; | |
181 | ||
182 | /* Equality and hash functions for hashtab code. */ | |
183 | ||
184 | static int | |
185 | reduction_info_eq (const void *aa, const void *bb) | |
186 | { | |
187 | const struct reduction_info *a = (const struct reduction_info *) aa; | |
188 | const struct reduction_info *b = (const struct reduction_info *) bb; | |
189 | ||
190 | return (a->reduc_phi == b->reduc_phi); | |
191 | } | |
192 | ||
193 | static hashval_t | |
194 | reduction_info_hash (const void *aa) | |
195 | { | |
196 | const struct reduction_info *a = (const struct reduction_info *) aa; | |
197 | ||
198 | return htab_hash_pointer (a->reduc_phi); | |
199 | } | |
200 | ||
201 | static struct reduction_info * | |
726a989a | 202 | reduction_phi (htab_t reduction_list, gimple phi) |
a509ebb5 RL |
203 | { |
204 | struct reduction_info tmpred, *red; | |
205 | ||
206 | if (htab_elements (reduction_list) == 0) | |
207 | return NULL; | |
208 | ||
209 | tmpred.reduc_phi = phi; | |
3d9a9f94 | 210 | red = (struct reduction_info *) htab_find (reduction_list, &tmpred); |
a509ebb5 RL |
211 | |
212 | return red; | |
213 | } | |
214 | ||
5f40b3cb ZD |
215 | /* Element of hashtable of names to copy. */ |
216 | ||
217 | struct name_to_copy_elt | |
218 | { | |
219 | unsigned version; /* The version of the name to copy. */ | |
220 | tree new_name; /* The new name used in the copy. */ | |
221 | tree field; /* The field of the structure used to pass the | |
222 | value. */ | |
223 | }; | |
224 | ||
225 | /* Equality and hash functions for hashtab code. */ | |
226 | ||
227 | static int | |
228 | name_to_copy_elt_eq (const void *aa, const void *bb) | |
229 | { | |
a509ebb5 RL |
230 | const struct name_to_copy_elt *a = (const struct name_to_copy_elt *) aa; |
231 | const struct name_to_copy_elt *b = (const struct name_to_copy_elt *) bb; | |
5f40b3cb ZD |
232 | |
233 | return a->version == b->version; | |
234 | } | |
235 | ||
236 | static hashval_t | |
237 | name_to_copy_elt_hash (const void *aa) | |
238 | { | |
a509ebb5 | 239 | const struct name_to_copy_elt *a = (const struct name_to_copy_elt *) aa; |
5f40b3cb ZD |
240 | |
241 | return (hashval_t) a->version; | |
242 | } | |
243 | ||
08dab97a RL |
244 | |
245 | /* Data dependency analysis. Returns true if the iterations of LOOP | |
246 | are independent on each other (that is, if we can execute them | |
247 | in parallel). */ | |
5f40b3cb ZD |
248 | |
249 | static bool | |
08dab97a | 250 | loop_parallel_p (struct loop *loop) |
5f40b3cb | 251 | { |
a509ebb5 | 252 | VEC (ddr_p, heap) * dependence_relations; |
726a989a | 253 | VEC (data_reference_p, heap) *datarefs; |
5f40b3cb ZD |
254 | lambda_trans_matrix trans; |
255 | bool ret = false; | |
5f40b3cb ZD |
256 | |
257 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
48710229 RL |
258 | { |
259 | fprintf (dump_file, "Considering loop %d\n", loop->num); | |
260 | if (!loop->inner) | |
261 | fprintf (dump_file, "loop is innermost\n"); | |
262 | else | |
263 | fprintf (dump_file, "loop NOT innermost\n"); | |
264 | } | |
5f40b3cb | 265 | |
5f40b3cb ZD |
266 | /* Check for problems with dependences. If the loop can be reversed, |
267 | the iterations are independent. */ | |
268 | datarefs = VEC_alloc (data_reference_p, heap, 10); | |
269 | dependence_relations = VEC_alloc (ddr_p, heap, 10 * 10); | |
270 | compute_data_dependences_for_loop (loop, true, &datarefs, | |
271 | &dependence_relations); | |
272 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
273 | dump_data_dependence_relations (dump_file, dependence_relations); | |
274 | ||
275 | trans = lambda_trans_matrix_new (1, 1); | |
276 | LTM_MATRIX (trans)[0][0] = -1; | |
277 | ||
278 | if (lambda_transform_legal_p (trans, 1, dependence_relations)) | |
279 | { | |
280 | ret = true; | |
281 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
282 | fprintf (dump_file, " SUCCESS: may be parallelized\n"); | |
283 | } | |
284 | else if (dump_file && (dump_flags & TDF_DETAILS)) | |
a509ebb5 RL |
285 | fprintf (dump_file, |
286 | " FAILED: data dependencies exist across iterations\n"); | |
5f40b3cb ZD |
287 | |
288 | free_dependence_relations (dependence_relations); | |
289 | free_data_refs (datarefs); | |
290 | ||
291 | return ret; | |
292 | } | |
293 | ||
1d4af1e8 SP |
294 | /* Return true when LOOP contains basic blocks marked with the |
295 | BB_IRREDUCIBLE_LOOP flag. */ | |
296 | ||
297 | static inline bool | |
298 | loop_has_blocks_with_irreducible_flag (struct loop *loop) | |
299 | { | |
300 | unsigned i; | |
301 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
302 | bool res = true; | |
303 | ||
304 | for (i = 0; i < loop->num_nodes; i++) | |
305 | if (bbs[i]->flags & BB_IRREDUCIBLE_LOOP) | |
306 | goto end; | |
307 | ||
308 | res = false; | |
309 | end: | |
310 | free (bbs); | |
311 | return res; | |
312 | } | |
313 | ||
8a171a59 | 314 | /* Assigns the address of OBJ in TYPE to an ssa name, and returns this name. |
9f9f72aa | 315 | The assignment statement is placed on edge ENTRY. DECL_ADDRESS maps decls |
8a171a59 ZD |
316 | to their addresses that can be reused. The address of OBJ is known to |
317 | be invariant in the whole function. */ | |
5f40b3cb ZD |
318 | |
319 | static tree | |
9f9f72aa | 320 | take_address_of (tree obj, tree type, edge entry, htab_t decl_address) |
5f40b3cb | 321 | { |
8a171a59 | 322 | int uid; |
5f40b3cb ZD |
323 | void **dslot; |
324 | struct int_tree_map ielt, *nielt; | |
726a989a RB |
325 | tree *var_p, name, bvar, addr; |
326 | gimple stmt; | |
327 | gimple_seq stmts; | |
5f40b3cb | 328 | |
8a171a59 ZD |
329 | /* Since the address of OBJ is invariant, the trees may be shared. |
330 | Avoid rewriting unrelated parts of the code. */ | |
331 | obj = unshare_expr (obj); | |
332 | for (var_p = &obj; | |
333 | handled_component_p (*var_p); | |
334 | var_p = &TREE_OPERAND (*var_p, 0)) | |
335 | continue; | |
336 | uid = DECL_UID (*var_p); | |
337 | ||
5f40b3cb ZD |
338 | ielt.uid = uid; |
339 | dslot = htab_find_slot_with_hash (decl_address, &ielt, uid, INSERT); | |
340 | if (!*dslot) | |
341 | { | |
8a171a59 ZD |
342 | addr = build_addr (*var_p, current_function_decl); |
343 | bvar = create_tmp_var (TREE_TYPE (addr), get_name (*var_p)); | |
5f40b3cb | 344 | add_referenced_var (bvar); |
726a989a | 345 | stmt = gimple_build_assign (bvar, addr); |
5f40b3cb | 346 | name = make_ssa_name (bvar, stmt); |
726a989a RB |
347 | gimple_assign_set_lhs (stmt, name); |
348 | gsi_insert_on_edge_immediate (entry, stmt); | |
5f40b3cb ZD |
349 | |
350 | nielt = XNEW (struct int_tree_map); | |
351 | nielt->uid = uid; | |
352 | nielt->to = name; | |
353 | *dslot = nielt; | |
5f40b3cb | 354 | } |
8a171a59 ZD |
355 | else |
356 | name = ((struct int_tree_map *) *dslot)->to; | |
5f40b3cb | 357 | |
8a171a59 ZD |
358 | if (var_p != &obj) |
359 | { | |
360 | *var_p = build1 (INDIRECT_REF, TREE_TYPE (*var_p), name); | |
361 | name = force_gimple_operand (build_addr (obj, current_function_decl), | |
726a989a RB |
362 | &stmts, true, NULL_TREE); |
363 | if (!gimple_seq_empty_p (stmts)) | |
364 | gsi_insert_seq_on_edge_immediate (entry, stmts); | |
8a171a59 | 365 | } |
5f40b3cb | 366 | |
8a171a59 ZD |
367 | if (TREE_TYPE (name) != type) |
368 | { | |
726a989a | 369 | name = force_gimple_operand (fold_convert (type, name), &stmts, true, |
8a171a59 | 370 | NULL_TREE); |
726a989a RB |
371 | if (!gimple_seq_empty_p (stmts)) |
372 | gsi_insert_seq_on_edge_immediate (entry, stmts); | |
8a171a59 | 373 | } |
5f40b3cb ZD |
374 | |
375 | return name; | |
376 | } | |
377 | ||
a509ebb5 RL |
378 | /* Callback for htab_traverse. Create the initialization statement |
379 | for reduction described in SLOT, and place it at the preheader of | |
380 | the loop described in DATA. */ | |
381 | ||
382 | static int | |
383 | initialize_reductions (void **slot, void *data) | |
384 | { | |
a509ebb5 | 385 | tree init, c; |
a509ebb5 RL |
386 | tree bvar, type, arg; |
387 | edge e; | |
388 | ||
3d9a9f94 | 389 | struct reduction_info *const reduc = (struct reduction_info *) *slot; |
a509ebb5 RL |
390 | struct loop *loop = (struct loop *) data; |
391 | ||
392 | /* Create initialization in preheader: | |
393 | reduction_variable = initialization value of reduction. */ | |
394 | ||
395 | /* In the phi node at the header, replace the argument coming | |
396 | from the preheader with the reduction initialization value. */ | |
397 | ||
398 | /* Create a new variable to initialize the reduction. */ | |
399 | type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); | |
400 | bvar = create_tmp_var (type, "reduction"); | |
401 | add_referenced_var (bvar); | |
402 | ||
c2255bc4 AH |
403 | c = build_omp_clause (gimple_location (reduc->reduc_stmt), |
404 | OMP_CLAUSE_REDUCTION); | |
a509ebb5 | 405 | OMP_CLAUSE_REDUCTION_CODE (c) = reduc->reduction_code; |
726a989a | 406 | OMP_CLAUSE_DECL (c) = SSA_NAME_VAR (gimple_assign_lhs (reduc->reduc_stmt)); |
a509ebb5 RL |
407 | |
408 | init = omp_reduction_init (c, TREE_TYPE (bvar)); | |
409 | reduc->init = init; | |
410 | ||
0eb7e7aa RL |
411 | /* Replace the argument representing the initialization value |
412 | with the initialization value for the reduction (neutral | |
413 | element for the particular operation, e.g. 0 for PLUS_EXPR, | |
414 | 1 for MULT_EXPR, etc). | |
415 | Keep the old value in a new variable "reduction_initial", | |
416 | that will be taken in consideration after the parallel | |
417 | computing is done. */ | |
a509ebb5 RL |
418 | |
419 | e = loop_preheader_edge (loop); | |
420 | arg = PHI_ARG_DEF_FROM_EDGE (reduc->reduc_phi, e); | |
421 | /* Create new variable to hold the initial value. */ | |
a509ebb5 | 422 | |
a509ebb5 | 423 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE |
0eb7e7aa | 424 | (reduc->reduc_phi, loop_preheader_edge (loop)), init); |
ae0bce62 | 425 | reduc->initial_value = arg; |
a509ebb5 RL |
426 | return 1; |
427 | } | |
5f40b3cb ZD |
428 | |
429 | struct elv_data | |
430 | { | |
726a989a | 431 | struct walk_stmt_info info; |
9f9f72aa | 432 | edge entry; |
5f40b3cb ZD |
433 | htab_t decl_address; |
434 | bool changed; | |
435 | }; | |
436 | ||
9f9f72aa AP |
437 | /* Eliminates references to local variables in *TP out of the single |
438 | entry single exit region starting at DTA->ENTRY. | |
439 | DECL_ADDRESS contains addresses of the references that had their | |
440 | address taken already. If the expression is changed, CHANGED is | |
441 | set to true. Callback for walk_tree. */ | |
a509ebb5 | 442 | |
5f40b3cb | 443 | static tree |
8a171a59 | 444 | eliminate_local_variables_1 (tree *tp, int *walk_subtrees, void *data) |
5f40b3cb | 445 | { |
3d9a9f94 | 446 | struct elv_data *const dta = (struct elv_data *) data; |
8a171a59 | 447 | tree t = *tp, var, addr, addr_type, type, obj; |
5f40b3cb ZD |
448 | |
449 | if (DECL_P (t)) | |
450 | { | |
451 | *walk_subtrees = 0; | |
452 | ||
453 | if (!SSA_VAR_P (t) || DECL_EXTERNAL (t)) | |
454 | return NULL_TREE; | |
455 | ||
456 | type = TREE_TYPE (t); | |
457 | addr_type = build_pointer_type (type); | |
9f9f72aa | 458 | addr = take_address_of (t, addr_type, dta->entry, dta->decl_address); |
5f40b3cb ZD |
459 | *tp = build1 (INDIRECT_REF, TREE_TYPE (*tp), addr); |
460 | ||
461 | dta->changed = true; | |
462 | return NULL_TREE; | |
463 | } | |
464 | ||
465 | if (TREE_CODE (t) == ADDR_EXPR) | |
466 | { | |
8a171a59 ZD |
467 | /* ADDR_EXPR may appear in two contexts: |
468 | -- as a gimple operand, when the address taken is a function invariant | |
469 | -- as gimple rhs, when the resulting address in not a function | |
470 | invariant | |
471 | We do not need to do anything special in the latter case (the base of | |
472 | the memory reference whose address is taken may be replaced in the | |
473 | DECL_P case). The former case is more complicated, as we need to | |
474 | ensure that the new address is still a gimple operand. Thus, it | |
475 | is not sufficient to replace just the base of the memory reference -- | |
476 | we need to move the whole computation of the address out of the | |
477 | loop. */ | |
478 | if (!is_gimple_val (t)) | |
5f40b3cb ZD |
479 | return NULL_TREE; |
480 | ||
481 | *walk_subtrees = 0; | |
8a171a59 ZD |
482 | obj = TREE_OPERAND (t, 0); |
483 | var = get_base_address (obj); | |
484 | if (!var || !SSA_VAR_P (var) || DECL_EXTERNAL (var)) | |
5f40b3cb ZD |
485 | return NULL_TREE; |
486 | ||
487 | addr_type = TREE_TYPE (t); | |
9f9f72aa | 488 | addr = take_address_of (obj, addr_type, dta->entry, dta->decl_address); |
5f40b3cb ZD |
489 | *tp = addr; |
490 | ||
491 | dta->changed = true; | |
492 | return NULL_TREE; | |
493 | } | |
494 | ||
726a989a | 495 | if (!EXPR_P (t)) |
5f40b3cb ZD |
496 | *walk_subtrees = 0; |
497 | ||
498 | return NULL_TREE; | |
499 | } | |
500 | ||
9f9f72aa AP |
501 | /* Moves the references to local variables in STMT out of the single |
502 | entry single exit region starting at ENTRY. DECL_ADDRESS contains | |
503 | addresses of the references that had their address taken | |
504 | already. */ | |
5f40b3cb ZD |
505 | |
506 | static void | |
726a989a | 507 | eliminate_local_variables_stmt (edge entry, gimple stmt, |
5f40b3cb ZD |
508 | htab_t decl_address) |
509 | { | |
510 | struct elv_data dta; | |
511 | ||
726a989a | 512 | memset (&dta.info, '\0', sizeof (dta.info)); |
9f9f72aa | 513 | dta.entry = entry; |
5f40b3cb ZD |
514 | dta.decl_address = decl_address; |
515 | dta.changed = false; | |
516 | ||
b5b8b0ac AO |
517 | if (gimple_debug_bind_p (stmt)) |
518 | walk_tree (gimple_debug_bind_get_value_ptr (stmt), | |
519 | eliminate_local_variables_1, &dta.info, NULL); | |
520 | else | |
521 | walk_gimple_op (stmt, eliminate_local_variables_1, &dta.info); | |
5f40b3cb ZD |
522 | |
523 | if (dta.changed) | |
524 | update_stmt (stmt); | |
525 | } | |
526 | ||
9f9f72aa AP |
527 | /* Eliminates the references to local variables from the single entry |
528 | single exit region between the ENTRY and EXIT edges. | |
529 | ||
a509ebb5 RL |
530 | This includes: |
531 | 1) Taking address of a local variable -- these are moved out of the | |
9f9f72aa | 532 | region (and temporary variable is created to hold the address if |
a509ebb5 | 533 | necessary). |
9f9f72aa | 534 | |
5f40b3cb | 535 | 2) Dereferencing a local variable -- these are replaced with indirect |
a509ebb5 | 536 | references. */ |
5f40b3cb ZD |
537 | |
538 | static void | |
9f9f72aa | 539 | eliminate_local_variables (edge entry, edge exit) |
5f40b3cb | 540 | { |
9f9f72aa AP |
541 | basic_block bb; |
542 | VEC (basic_block, heap) *body = VEC_alloc (basic_block, heap, 3); | |
5f40b3cb | 543 | unsigned i; |
726a989a | 544 | gimple_stmt_iterator gsi; |
5f40b3cb ZD |
545 | htab_t decl_address = htab_create (10, int_tree_map_hash, int_tree_map_eq, |
546 | free); | |
9f9f72aa AP |
547 | basic_block entry_bb = entry->src; |
548 | basic_block exit_bb = exit->dest; | |
5f40b3cb | 549 | |
9f9f72aa | 550 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); |
5f40b3cb | 551 | |
9f9f72aa AP |
552 | for (i = 0; VEC_iterate (basic_block, body, i, bb); i++) |
553 | if (bb != entry_bb && bb != exit_bb) | |
726a989a RB |
554 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
555 | eliminate_local_variables_stmt (entry, gsi_stmt (gsi), | |
9f9f72aa | 556 | decl_address); |
5f40b3cb ZD |
557 | |
558 | htab_delete (decl_address); | |
9f9f72aa AP |
559 | VEC_free (basic_block, heap, body); |
560 | } | |
561 | ||
562 | /* Returns true if expression EXPR is not defined between ENTRY and | |
563 | EXIT, i.e. if all its operands are defined outside of the region. */ | |
564 | ||
565 | static bool | |
566 | expr_invariant_in_region_p (edge entry, edge exit, tree expr) | |
567 | { | |
568 | basic_block entry_bb = entry->src; | |
569 | basic_block exit_bb = exit->dest; | |
570 | basic_block def_bb; | |
9f9f72aa AP |
571 | |
572 | if (is_gimple_min_invariant (expr)) | |
573 | return true; | |
574 | ||
575 | if (TREE_CODE (expr) == SSA_NAME) | |
576 | { | |
726a989a | 577 | def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); |
9f9f72aa AP |
578 | if (def_bb |
579 | && dominated_by_p (CDI_DOMINATORS, def_bb, entry_bb) | |
580 | && !dominated_by_p (CDI_DOMINATORS, def_bb, exit_bb)) | |
581 | return false; | |
582 | ||
583 | return true; | |
584 | } | |
585 | ||
726a989a | 586 | return false; |
5f40b3cb ZD |
587 | } |
588 | ||
589 | /* If COPY_NAME_P is true, creates and returns a duplicate of NAME. | |
590 | The copies are stored to NAME_COPIES, if NAME was already duplicated, | |
591 | its duplicate stored in NAME_COPIES is returned. | |
592 | ||
593 | Regardless of COPY_NAME_P, the decl used as a base of the ssa name is also | |
594 | duplicated, storing the copies in DECL_COPIES. */ | |
595 | ||
596 | static tree | |
9f9f72aa AP |
597 | separate_decls_in_region_name (tree name, |
598 | htab_t name_copies, htab_t decl_copies, | |
599 | bool copy_name_p) | |
5f40b3cb ZD |
600 | { |
601 | tree copy, var, var_copy; | |
602 | unsigned idx, uid, nuid; | |
603 | struct int_tree_map ielt, *nielt; | |
604 | struct name_to_copy_elt elt, *nelt; | |
605 | void **slot, **dslot; | |
606 | ||
607 | if (TREE_CODE (name) != SSA_NAME) | |
608 | return name; | |
609 | ||
610 | idx = SSA_NAME_VERSION (name); | |
611 | elt.version = idx; | |
612 | slot = htab_find_slot_with_hash (name_copies, &elt, idx, | |
613 | copy_name_p ? INSERT : NO_INSERT); | |
614 | if (slot && *slot) | |
615 | return ((struct name_to_copy_elt *) *slot)->new_name; | |
616 | ||
617 | var = SSA_NAME_VAR (name); | |
618 | uid = DECL_UID (var); | |
619 | ielt.uid = uid; | |
620 | dslot = htab_find_slot_with_hash (decl_copies, &ielt, uid, INSERT); | |
621 | if (!*dslot) | |
622 | { | |
623 | var_copy = create_tmp_var (TREE_TYPE (var), get_name (var)); | |
36ad7922 | 624 | DECL_GIMPLE_REG_P (var_copy) = DECL_GIMPLE_REG_P (var); |
5f40b3cb ZD |
625 | add_referenced_var (var_copy); |
626 | nielt = XNEW (struct int_tree_map); | |
627 | nielt->uid = uid; | |
628 | nielt->to = var_copy; | |
629 | *dslot = nielt; | |
630 | ||
631 | /* Ensure that when we meet this decl next time, we won't duplicate | |
a509ebb5 | 632 | it again. */ |
5f40b3cb ZD |
633 | nuid = DECL_UID (var_copy); |
634 | ielt.uid = nuid; | |
635 | dslot = htab_find_slot_with_hash (decl_copies, &ielt, nuid, INSERT); | |
636 | gcc_assert (!*dslot); | |
637 | nielt = XNEW (struct int_tree_map); | |
638 | nielt->uid = nuid; | |
639 | nielt->to = var_copy; | |
640 | *dslot = nielt; | |
641 | } | |
642 | else | |
643 | var_copy = ((struct int_tree_map *) *dslot)->to; | |
644 | ||
645 | if (copy_name_p) | |
646 | { | |
726a989a | 647 | copy = duplicate_ssa_name (name, NULL); |
5f40b3cb ZD |
648 | nelt = XNEW (struct name_to_copy_elt); |
649 | nelt->version = idx; | |
650 | nelt->new_name = copy; | |
651 | nelt->field = NULL_TREE; | |
652 | *slot = nelt; | |
653 | } | |
654 | else | |
655 | { | |
656 | gcc_assert (!slot); | |
657 | copy = name; | |
658 | } | |
659 | ||
660 | SSA_NAME_VAR (copy) = var_copy; | |
661 | return copy; | |
662 | } | |
663 | ||
9f9f72aa AP |
664 | /* Finds the ssa names used in STMT that are defined outside the |
665 | region between ENTRY and EXIT and replaces such ssa names with | |
666 | their duplicates. The duplicates are stored to NAME_COPIES. Base | |
667 | decls of all ssa names used in STMT (including those defined in | |
668 | LOOP) are replaced with the new temporary variables; the | |
669 | replacement decls are stored in DECL_COPIES. */ | |
5f40b3cb ZD |
670 | |
671 | static void | |
726a989a | 672 | separate_decls_in_region_stmt (edge entry, edge exit, gimple stmt, |
9f9f72aa | 673 | htab_t name_copies, htab_t decl_copies) |
5f40b3cb ZD |
674 | { |
675 | use_operand_p use; | |
676 | def_operand_p def; | |
677 | ssa_op_iter oi; | |
678 | tree name, copy; | |
679 | bool copy_name_p; | |
680 | ||
681 | mark_virtual_ops_for_renaming (stmt); | |
682 | ||
683 | FOR_EACH_PHI_OR_STMT_DEF (def, stmt, oi, SSA_OP_DEF) | |
a509ebb5 RL |
684 | { |
685 | name = DEF_FROM_PTR (def); | |
686 | gcc_assert (TREE_CODE (name) == SSA_NAME); | |
9f9f72aa AP |
687 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, |
688 | false); | |
a509ebb5 RL |
689 | gcc_assert (copy == name); |
690 | } | |
5f40b3cb ZD |
691 | |
692 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
a509ebb5 RL |
693 | { |
694 | name = USE_FROM_PTR (use); | |
695 | if (TREE_CODE (name) != SSA_NAME) | |
696 | continue; | |
697 | ||
9f9f72aa AP |
698 | copy_name_p = expr_invariant_in_region_p (entry, exit, name); |
699 | copy = separate_decls_in_region_name (name, name_copies, decl_copies, | |
700 | copy_name_p); | |
a509ebb5 RL |
701 | SET_USE (use, copy); |
702 | } | |
5f40b3cb ZD |
703 | } |
704 | ||
b5b8b0ac AO |
705 | /* Finds the ssa names used in STMT that are defined outside the |
706 | region between ENTRY and EXIT and replaces such ssa names with | |
707 | their duplicates. The duplicates are stored to NAME_COPIES. Base | |
708 | decls of all ssa names used in STMT (including those defined in | |
709 | LOOP) are replaced with the new temporary variables; the | |
710 | replacement decls are stored in DECL_COPIES. */ | |
711 | ||
712 | static bool | |
713 | separate_decls_in_region_debug_bind (gimple stmt, | |
714 | htab_t name_copies, htab_t decl_copies) | |
715 | { | |
716 | use_operand_p use; | |
717 | ssa_op_iter oi; | |
718 | tree var, name; | |
719 | struct int_tree_map ielt; | |
720 | struct name_to_copy_elt elt; | |
721 | void **slot, **dslot; | |
722 | ||
723 | var = gimple_debug_bind_get_var (stmt); | |
4f2a9af8 JJ |
724 | if (TREE_CODE (var) == DEBUG_EXPR_DECL) |
725 | return true; | |
b5b8b0ac AO |
726 | gcc_assert (DECL_P (var) && SSA_VAR_P (var)); |
727 | ielt.uid = DECL_UID (var); | |
728 | dslot = htab_find_slot_with_hash (decl_copies, &ielt, ielt.uid, NO_INSERT); | |
729 | if (!dslot) | |
730 | return true; | |
731 | gimple_debug_bind_set_var (stmt, ((struct int_tree_map *) *dslot)->to); | |
732 | ||
733 | FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
734 | { | |
735 | name = USE_FROM_PTR (use); | |
736 | if (TREE_CODE (name) != SSA_NAME) | |
737 | continue; | |
738 | ||
739 | elt.version = SSA_NAME_VERSION (name); | |
740 | slot = htab_find_slot_with_hash (name_copies, &elt, elt.version, NO_INSERT); | |
741 | if (!slot) | |
742 | { | |
743 | gimple_debug_bind_reset_value (stmt); | |
744 | update_stmt (stmt); | |
745 | break; | |
746 | } | |
747 | ||
748 | SET_USE (use, ((struct name_to_copy_elt *) *slot)->new_name); | |
749 | } | |
750 | ||
751 | return false; | |
752 | } | |
753 | ||
0eb7e7aa RL |
754 | /* Callback for htab_traverse. Adds a field corresponding to the reduction |
755 | specified in SLOT. The type is passed in DATA. */ | |
756 | ||
757 | static int | |
758 | add_field_for_reduction (void **slot, void *data) | |
a509ebb5 | 759 | { |
0eb7e7aa | 760 | |
3d9a9f94 KG |
761 | struct reduction_info *const red = (struct reduction_info *) *slot; |
762 | tree const type = (tree) data; | |
726a989a | 763 | tree var = SSA_NAME_VAR (gimple_assign_lhs (red->reduc_stmt)); |
c2255bc4 AH |
764 | tree field = build_decl (gimple_location (red->reduc_stmt), |
765 | FIELD_DECL, DECL_NAME (var), TREE_TYPE (var)); | |
0eb7e7aa RL |
766 | |
767 | insert_field_into_struct (type, field); | |
768 | ||
769 | red->field = field; | |
770 | ||
771 | return 1; | |
772 | } | |
a509ebb5 | 773 | |
5f40b3cb | 774 | /* Callback for htab_traverse. Adds a field corresponding to a ssa name |
0eb7e7aa | 775 | described in SLOT. The type is passed in DATA. */ |
5f40b3cb ZD |
776 | |
777 | static int | |
778 | add_field_for_name (void **slot, void *data) | |
779 | { | |
3d9a9f94 KG |
780 | struct name_to_copy_elt *const elt = (struct name_to_copy_elt *) *slot; |
781 | tree type = (tree) data; | |
5f40b3cb ZD |
782 | tree name = ssa_name (elt->version); |
783 | tree var = SSA_NAME_VAR (name); | |
c2255bc4 AH |
784 | tree field = build_decl (DECL_SOURCE_LOCATION (var), |
785 | FIELD_DECL, DECL_NAME (var), TREE_TYPE (var)); | |
5f40b3cb ZD |
786 | |
787 | insert_field_into_struct (type, field); | |
788 | elt->field = field; | |
a509ebb5 | 789 | |
5f40b3cb ZD |
790 | return 1; |
791 | } | |
792 | ||
a509ebb5 RL |
793 | /* Callback for htab_traverse. A local result is the intermediate result |
794 | computed by a single | |
fa10beec | 795 | thread, or the initial value in case no iteration was executed. |
a509ebb5 RL |
796 | This function creates a phi node reflecting these values. |
797 | The phi's result will be stored in NEW_PHI field of the | |
798 | reduction's data structure. */ | |
799 | ||
800 | static int | |
801 | create_phi_for_local_result (void **slot, void *data) | |
802 | { | |
3d9a9f94 KG |
803 | struct reduction_info *const reduc = (struct reduction_info *) *slot; |
804 | const struct loop *const loop = (const struct loop *) data; | |
a509ebb5 | 805 | edge e; |
726a989a | 806 | gimple new_phi; |
a509ebb5 RL |
807 | basic_block store_bb; |
808 | tree local_res; | |
f5045c96 | 809 | source_location locus; |
a509ebb5 RL |
810 | |
811 | /* STORE_BB is the block where the phi | |
812 | should be stored. It is the destination of the loop exit. | |
726a989a | 813 | (Find the fallthru edge from GIMPLE_OMP_CONTINUE). */ |
a509ebb5 RL |
814 | store_bb = FALLTHRU_EDGE (loop->latch)->dest; |
815 | ||
816 | /* STORE_BB has two predecessors. One coming from the loop | |
817 | (the reduction's result is computed at the loop), | |
818 | and another coming from a block preceding the loop, | |
819 | when no iterations | |
820 | are executed (the initial value should be taken). */ | |
821 | if (EDGE_PRED (store_bb, 0) == FALLTHRU_EDGE (loop->latch)) | |
822 | e = EDGE_PRED (store_bb, 1); | |
823 | else | |
824 | e = EDGE_PRED (store_bb, 0); | |
726a989a RB |
825 | local_res |
826 | = make_ssa_name (SSA_NAME_VAR (gimple_assign_lhs (reduc->reduc_stmt)), | |
827 | NULL); | |
f5045c96 | 828 | locus = gimple_location (reduc->reduc_stmt); |
a509ebb5 RL |
829 | new_phi = create_phi_node (local_res, store_bb); |
830 | SSA_NAME_DEF_STMT (local_res) = new_phi; | |
f5045c96 | 831 | add_phi_arg (new_phi, reduc->init, e, locus); |
726a989a | 832 | add_phi_arg (new_phi, gimple_assign_lhs (reduc->reduc_stmt), |
f5045c96 | 833 | FALLTHRU_EDGE (loop->latch), locus); |
a509ebb5 RL |
834 | reduc->new_phi = new_phi; |
835 | ||
836 | return 1; | |
837 | } | |
5f40b3cb ZD |
838 | |
839 | struct clsn_data | |
840 | { | |
841 | tree store; | |
842 | tree load; | |
843 | ||
844 | basic_block store_bb; | |
845 | basic_block load_bb; | |
846 | }; | |
847 | ||
a509ebb5 RL |
848 | /* Callback for htab_traverse. Create an atomic instruction for the |
849 | reduction described in SLOT. | |
850 | DATA annotates the place in memory the atomic operation relates to, | |
851 | and the basic block it needs to be generated in. */ | |
852 | ||
853 | static int | |
854 | create_call_for_reduction_1 (void **slot, void *data) | |
855 | { | |
3d9a9f94 KG |
856 | struct reduction_info *const reduc = (struct reduction_info *) *slot; |
857 | struct clsn_data *const clsn_data = (struct clsn_data *) data; | |
726a989a | 858 | gimple_stmt_iterator gsi; |
a509ebb5 RL |
859 | tree type = TREE_TYPE (PHI_RESULT (reduc->reduc_phi)); |
860 | tree struct_type = TREE_TYPE (TREE_TYPE (clsn_data->load)); | |
861 | tree load_struct; | |
862 | basic_block bb; | |
863 | basic_block new_bb; | |
864 | edge e; | |
865 | tree t, addr, addr_type, ref, x; | |
726a989a RB |
866 | tree tmp_load, name; |
867 | gimple load; | |
a509ebb5 RL |
868 | |
869 | load_struct = fold_build1 (INDIRECT_REF, struct_type, clsn_data->load); | |
870 | t = build3 (COMPONENT_REF, type, load_struct, reduc->field, NULL_TREE); | |
871 | addr_type = build_pointer_type (type); | |
872 | ||
873 | addr = build_addr (t, current_function_decl); | |
874 | ||
875 | /* Create phi node. */ | |
876 | bb = clsn_data->load_bb; | |
877 | ||
878 | e = split_block (bb, t); | |
879 | new_bb = e->dest; | |
880 | ||
881 | tmp_load = create_tmp_var (TREE_TYPE (TREE_TYPE (addr)), NULL); | |
882 | add_referenced_var (tmp_load); | |
883 | tmp_load = make_ssa_name (tmp_load, NULL); | |
726a989a | 884 | load = gimple_build_omp_atomic_load (tmp_load, addr); |
a509ebb5 | 885 | SSA_NAME_DEF_STMT (tmp_load) = load; |
726a989a RB |
886 | gsi = gsi_start_bb (new_bb); |
887 | gsi_insert_after (&gsi, load, GSI_NEW_STMT); | |
a509ebb5 RL |
888 | |
889 | e = split_block (new_bb, load); | |
890 | new_bb = e->dest; | |
726a989a | 891 | gsi = gsi_start_bb (new_bb); |
a509ebb5 | 892 | ref = tmp_load; |
726a989a RB |
893 | x = fold_build2 (reduc->reduction_code, |
894 | TREE_TYPE (PHI_RESULT (reduc->new_phi)), ref, | |
895 | PHI_RESULT (reduc->new_phi)); | |
a509ebb5 | 896 | |
726a989a RB |
897 | name = force_gimple_operand_gsi (&gsi, x, true, NULL_TREE, true, |
898 | GSI_CONTINUE_LINKING); | |
a509ebb5 | 899 | |
726a989a | 900 | gsi_insert_after (&gsi, gimple_build_omp_atomic_store (name), GSI_NEW_STMT); |
a509ebb5 RL |
901 | return 1; |
902 | } | |
903 | ||
904 | /* Create the atomic operation at the join point of the threads. | |
905 | REDUCTION_LIST describes the reductions in the LOOP. | |
906 | LD_ST_DATA describes the shared data structure where | |
907 | shared data is stored in and loaded from. */ | |
908 | static void | |
909 | create_call_for_reduction (struct loop *loop, htab_t reduction_list, | |
910 | struct clsn_data *ld_st_data) | |
911 | { | |
912 | htab_traverse (reduction_list, create_phi_for_local_result, loop); | |
726a989a | 913 | /* Find the fallthru edge from GIMPLE_OMP_CONTINUE. */ |
a509ebb5 RL |
914 | ld_st_data->load_bb = FALLTHRU_EDGE (loop->latch)->dest; |
915 | htab_traverse (reduction_list, create_call_for_reduction_1, ld_st_data); | |
916 | } | |
917 | ||
ae0bce62 RL |
918 | /* Callback for htab_traverse. Loads the final reduction value at the |
919 | join point of all threads, and inserts it in the right place. */ | |
a509ebb5 RL |
920 | |
921 | static int | |
922 | create_loads_for_reductions (void **slot, void *data) | |
923 | { | |
3d9a9f94 KG |
924 | struct reduction_info *const red = (struct reduction_info *) *slot; |
925 | struct clsn_data *const clsn_data = (struct clsn_data *) data; | |
726a989a RB |
926 | gimple stmt; |
927 | gimple_stmt_iterator gsi; | |
928 | tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); | |
a509ebb5 RL |
929 | tree struct_type = TREE_TYPE (TREE_TYPE (clsn_data->load)); |
930 | tree load_struct; | |
ae0bce62 | 931 | tree name; |
a509ebb5 RL |
932 | tree x; |
933 | ||
726a989a | 934 | gsi = gsi_after_labels (clsn_data->load_bb); |
a509ebb5 RL |
935 | load_struct = fold_build1 (INDIRECT_REF, struct_type, clsn_data->load); |
936 | load_struct = build3 (COMPONENT_REF, type, load_struct, red->field, | |
937 | NULL_TREE); | |
a509ebb5 | 938 | |
ae0bce62 | 939 | x = load_struct; |
a509ebb5 | 940 | name = PHI_RESULT (red->keep_res); |
726a989a | 941 | stmt = gimple_build_assign (name, x); |
a509ebb5 RL |
942 | SSA_NAME_DEF_STMT (name) = stmt; |
943 | ||
726a989a | 944 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
a509ebb5 | 945 | |
726a989a RB |
946 | for (gsi = gsi_start_phis (gimple_bb (red->keep_res)); |
947 | !gsi_end_p (gsi); gsi_next (&gsi)) | |
948 | if (gsi_stmt (gsi) == red->keep_res) | |
949 | { | |
950 | remove_phi_node (&gsi, false); | |
951 | return 1; | |
952 | } | |
953 | gcc_unreachable (); | |
a509ebb5 RL |
954 | } |
955 | ||
956 | /* Load the reduction result that was stored in LD_ST_DATA. | |
957 | REDUCTION_LIST describes the list of reductions that the | |
fa10beec | 958 | loads should be generated for. */ |
a509ebb5 RL |
959 | static void |
960 | create_final_loads_for_reduction (htab_t reduction_list, | |
961 | struct clsn_data *ld_st_data) | |
962 | { | |
726a989a | 963 | gimple_stmt_iterator gsi; |
a509ebb5 | 964 | tree t; |
726a989a | 965 | gimple stmt; |
a509ebb5 | 966 | |
726a989a | 967 | gsi = gsi_after_labels (ld_st_data->load_bb); |
a509ebb5 | 968 | t = build_fold_addr_expr (ld_st_data->store); |
726a989a | 969 | stmt = gimple_build_assign (ld_st_data->load, t); |
a509ebb5 | 970 | |
726a989a RB |
971 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
972 | SSA_NAME_DEF_STMT (ld_st_data->load) = stmt; | |
a509ebb5 RL |
973 | |
974 | htab_traverse (reduction_list, create_loads_for_reductions, ld_st_data); | |
975 | ||
976 | } | |
977 | ||
0eb7e7aa RL |
978 | /* Callback for htab_traverse. Store the neutral value for the |
979 | particular reduction's operation, e.g. 0 for PLUS_EXPR, | |
980 | 1 for MULT_EXPR, etc. into the reduction field. | |
981 | The reduction is specified in SLOT. The store information is | |
982 | passed in DATA. */ | |
983 | ||
984 | static int | |
985 | create_stores_for_reduction (void **slot, void *data) | |
986 | { | |
3d9a9f94 KG |
987 | struct reduction_info *const red = (struct reduction_info *) *slot; |
988 | struct clsn_data *const clsn_data = (struct clsn_data *) data; | |
726a989a RB |
989 | tree t; |
990 | gimple stmt; | |
991 | gimple_stmt_iterator gsi; | |
992 | tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); | |
993 | ||
994 | gsi = gsi_last_bb (clsn_data->store_bb); | |
995 | t = build3 (COMPONENT_REF, type, clsn_data->store, red->field, NULL_TREE); | |
996 | stmt = gimple_build_assign (t, red->initial_value); | |
0eb7e7aa | 997 | mark_virtual_ops_for_renaming (stmt); |
726a989a | 998 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
0eb7e7aa RL |
999 | |
1000 | return 1; | |
1001 | } | |
1002 | ||
a509ebb5 RL |
1003 | /* Callback for htab_traverse. Creates loads to a field of LOAD in LOAD_BB and |
1004 | store to a field of STORE in STORE_BB for the ssa name and its duplicate | |
1005 | specified in SLOT. */ | |
1006 | ||
5f40b3cb ZD |
1007 | static int |
1008 | create_loads_and_stores_for_name (void **slot, void *data) | |
1009 | { | |
3d9a9f94 KG |
1010 | struct name_to_copy_elt *const elt = (struct name_to_copy_elt *) *slot; |
1011 | struct clsn_data *const clsn_data = (struct clsn_data *) data; | |
726a989a RB |
1012 | tree t; |
1013 | gimple stmt; | |
1014 | gimple_stmt_iterator gsi; | |
5f40b3cb ZD |
1015 | tree type = TREE_TYPE (elt->new_name); |
1016 | tree struct_type = TREE_TYPE (TREE_TYPE (clsn_data->load)); | |
1017 | tree load_struct; | |
1018 | ||
726a989a RB |
1019 | gsi = gsi_last_bb (clsn_data->store_bb); |
1020 | t = build3 (COMPONENT_REF, type, clsn_data->store, elt->field, NULL_TREE); | |
1021 | stmt = gimple_build_assign (t, ssa_name (elt->version)); | |
5f40b3cb | 1022 | mark_virtual_ops_for_renaming (stmt); |
726a989a | 1023 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb | 1024 | |
726a989a | 1025 | gsi = gsi_last_bb (clsn_data->load_bb); |
5f40b3cb | 1026 | load_struct = fold_build1 (INDIRECT_REF, struct_type, clsn_data->load); |
726a989a RB |
1027 | t = build3 (COMPONENT_REF, type, load_struct, elt->field, NULL_TREE); |
1028 | stmt = gimple_build_assign (elt->new_name, t); | |
5f40b3cb | 1029 | SSA_NAME_DEF_STMT (elt->new_name) = stmt; |
726a989a | 1030 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb ZD |
1031 | |
1032 | return 1; | |
1033 | } | |
1034 | ||
1035 | /* Moves all the variables used in LOOP and defined outside of it (including | |
1036 | the initial values of loop phi nodes, and *PER_THREAD if it is a ssa | |
1037 | name) to a structure created for this purpose. The code | |
1038 | ||
1039 | while (1) | |
1040 | { | |
1041 | use (a); | |
1042 | use (b); | |
1043 | } | |
1044 | ||
1045 | is transformed this way: | |
1046 | ||
1047 | bb0: | |
1048 | old.a = a; | |
1049 | old.b = b; | |
1050 | ||
1051 | bb1: | |
1052 | a' = new->a; | |
1053 | b' = new->b; | |
1054 | while (1) | |
1055 | { | |
1056 | use (a'); | |
1057 | use (b'); | |
1058 | } | |
1059 | ||
1060 | `old' is stored to *ARG_STRUCT and `new' is stored to NEW_ARG_STRUCT. The | |
1061 | pointer `new' is intentionally not initialized (the loop will be split to a | |
1062 | separate function later, and `new' will be initialized from its arguments). | |
a509ebb5 RL |
1063 | LD_ST_DATA holds information about the shared data structure used to pass |
1064 | information among the threads. It is initialized here, and | |
1065 | gen_parallel_loop will pass it to create_call_for_reduction that | |
1066 | needs this information. REDUCTION_LIST describes the reductions | |
1067 | in LOOP. */ | |
5f40b3cb ZD |
1068 | |
1069 | static void | |
9f9f72aa AP |
1070 | separate_decls_in_region (edge entry, edge exit, htab_t reduction_list, |
1071 | tree *arg_struct, tree *new_arg_struct, | |
1072 | struct clsn_data *ld_st_data) | |
a509ebb5 | 1073 | |
5f40b3cb | 1074 | { |
9f9f72aa | 1075 | basic_block bb1 = split_edge (entry); |
5f40b3cb ZD |
1076 | basic_block bb0 = single_pred (bb1); |
1077 | htab_t name_copies = htab_create (10, name_to_copy_elt_hash, | |
1078 | name_to_copy_elt_eq, free); | |
1079 | htab_t decl_copies = htab_create (10, int_tree_map_hash, int_tree_map_eq, | |
1080 | free); | |
5f40b3cb | 1081 | unsigned i; |
726a989a RB |
1082 | tree type, type_name, nvar; |
1083 | gimple_stmt_iterator gsi; | |
5f40b3cb | 1084 | struct clsn_data clsn_data; |
9f9f72aa AP |
1085 | VEC (basic_block, heap) *body = VEC_alloc (basic_block, heap, 3); |
1086 | basic_block bb; | |
1087 | basic_block entry_bb = bb1; | |
1088 | basic_block exit_bb = exit->dest; | |
b5b8b0ac | 1089 | bool has_debug_stmt = false; |
5f40b3cb | 1090 | |
726a989a | 1091 | entry = single_succ_edge (entry_bb); |
9f9f72aa | 1092 | gather_blocks_in_sese_region (entry_bb, exit_bb, &body); |
5f40b3cb | 1093 | |
9f9f72aa AP |
1094 | for (i = 0; VEC_iterate (basic_block, body, i, bb); i++) |
1095 | { | |
1096 | if (bb != entry_bb && bb != exit_bb) | |
1097 | { | |
726a989a RB |
1098 | for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
1099 | separate_decls_in_region_stmt (entry, exit, gsi_stmt (gsi), | |
1100 | name_copies, decl_copies); | |
1101 | ||
1102 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
b5b8b0ac AO |
1103 | { |
1104 | gimple stmt = gsi_stmt (gsi); | |
1105 | ||
1106 | if (is_gimple_debug (stmt)) | |
1107 | has_debug_stmt = true; | |
1108 | else | |
1109 | separate_decls_in_region_stmt (entry, exit, stmt, | |
1110 | name_copies, decl_copies); | |
1111 | } | |
9f9f72aa | 1112 | } |
5f40b3cb | 1113 | } |
9f9f72aa | 1114 | |
b5b8b0ac AO |
1115 | /* Now process debug bind stmts. We must not create decls while |
1116 | processing debug stmts, so we defer their processing so as to | |
1117 | make sure we will have debug info for as many variables as | |
1118 | possible (all of those that were dealt with in the loop above), | |
1119 | and discard those for which we know there's nothing we can | |
1120 | do. */ | |
1121 | if (has_debug_stmt) | |
1122 | for (i = 0; VEC_iterate (basic_block, body, i, bb); i++) | |
1123 | if (bb != entry_bb && bb != exit_bb) | |
1124 | { | |
1125 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) | |
1126 | { | |
1127 | gimple stmt = gsi_stmt (gsi); | |
1128 | ||
1129 | if (gimple_debug_bind_p (stmt)) | |
1130 | { | |
1131 | if (separate_decls_in_region_debug_bind (stmt, | |
1132 | name_copies, | |
1133 | decl_copies)) | |
1134 | { | |
1135 | gsi_remove (&gsi, true); | |
1136 | continue; | |
1137 | } | |
1138 | } | |
1139 | ||
1140 | gsi_next (&gsi); | |
1141 | } | |
1142 | } | |
1143 | ||
9f9f72aa | 1144 | VEC_free (basic_block, heap, body); |
5f40b3cb | 1145 | |
c9630c40 | 1146 | if (htab_elements (name_copies) == 0 && htab_elements (reduction_list) == 0) |
5f40b3cb ZD |
1147 | { |
1148 | /* It may happen that there is nothing to copy (if there are only | |
a509ebb5 | 1149 | loop carried and external variables in the loop). */ |
5f40b3cb ZD |
1150 | *arg_struct = NULL; |
1151 | *new_arg_struct = NULL; | |
1152 | } | |
1153 | else | |
1154 | { | |
1155 | /* Create the type for the structure to store the ssa names to. */ | |
1156 | type = lang_hooks.types.make_type (RECORD_TYPE); | |
c2255bc4 AH |
1157 | type_name = build_decl (BUILTINS_LOCATION, |
1158 | TYPE_DECL, create_tmp_var_name (".paral_data"), | |
5f40b3cb ZD |
1159 | type); |
1160 | TYPE_NAME (type) = type_name; | |
1161 | ||
0eb7e7aa | 1162 | htab_traverse (name_copies, add_field_for_name, type); |
9f9f72aa | 1163 | if (reduction_list && htab_elements (reduction_list) > 0) |
0eb7e7aa RL |
1164 | { |
1165 | /* Create the fields for reductions. */ | |
1166 | htab_traverse (reduction_list, add_field_for_reduction, | |
1167 | type); | |
1168 | } | |
5f40b3cb | 1169 | layout_type (type); |
0eb7e7aa | 1170 | |
5f40b3cb ZD |
1171 | /* Create the loads and stores. */ |
1172 | *arg_struct = create_tmp_var (type, ".paral_data_store"); | |
1173 | add_referenced_var (*arg_struct); | |
1174 | nvar = create_tmp_var (build_pointer_type (type), ".paral_data_load"); | |
1175 | add_referenced_var (nvar); | |
726a989a | 1176 | *new_arg_struct = make_ssa_name (nvar, NULL); |
5f40b3cb | 1177 | |
a509ebb5 RL |
1178 | ld_st_data->store = *arg_struct; |
1179 | ld_st_data->load = *new_arg_struct; | |
1180 | ld_st_data->store_bb = bb0; | |
1181 | ld_st_data->load_bb = bb1; | |
0eb7e7aa | 1182 | |
5f40b3cb | 1183 | htab_traverse (name_copies, create_loads_and_stores_for_name, |
a509ebb5 RL |
1184 | ld_st_data); |
1185 | ||
ae0bce62 RL |
1186 | /* Load the calculation from memory (after the join of the threads). */ |
1187 | ||
9f9f72aa | 1188 | if (reduction_list && htab_elements (reduction_list) > 0) |
a509ebb5 | 1189 | { |
0eb7e7aa RL |
1190 | htab_traverse (reduction_list, create_stores_for_reduction, |
1191 | ld_st_data); | |
726a989a | 1192 | clsn_data.load = make_ssa_name (nvar, NULL); |
9f9f72aa | 1193 | clsn_data.load_bb = exit->dest; |
a509ebb5 RL |
1194 | clsn_data.store = ld_st_data->store; |
1195 | create_final_loads_for_reduction (reduction_list, &clsn_data); | |
1196 | } | |
5f40b3cb ZD |
1197 | } |
1198 | ||
1199 | htab_delete (decl_copies); | |
1200 | htab_delete (name_copies); | |
1201 | } | |
1202 | ||
1203 | /* Bitmap containing uids of functions created by parallelization. We cannot | |
1204 | allocate it from the default obstack, as it must live across compilation | |
1205 | of several functions; we make it gc allocated instead. */ | |
1206 | ||
1207 | static GTY(()) bitmap parallelized_functions; | |
1208 | ||
1209 | /* Returns true if FN was created by create_loop_fn. */ | |
1210 | ||
1211 | static bool | |
1212 | parallelized_function_p (tree fn) | |
1213 | { | |
1214 | if (!parallelized_functions || !DECL_ARTIFICIAL (fn)) | |
1215 | return false; | |
1216 | ||
1217 | return bitmap_bit_p (parallelized_functions, DECL_UID (fn)); | |
1218 | } | |
1219 | ||
1220 | /* Creates and returns an empty function that will receive the body of | |
1221 | a parallelized loop. */ | |
1222 | ||
1223 | static tree | |
1224 | create_loop_fn (void) | |
1225 | { | |
1226 | char buf[100]; | |
1227 | char *tname; | |
1228 | tree decl, type, name, t; | |
1229 | struct function *act_cfun = cfun; | |
1230 | static unsigned loopfn_num; | |
1231 | ||
1232 | snprintf (buf, 100, "%s.$loopfn", current_function_name ()); | |
1233 | ASM_FORMAT_PRIVATE_NAME (tname, buf, loopfn_num++); | |
1234 | clean_symbol_name (tname); | |
1235 | name = get_identifier (tname); | |
1236 | type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); | |
1237 | ||
c2255bc4 AH |
1238 | decl = build_decl (BUILTINS_LOCATION, |
1239 | FUNCTION_DECL, name, type); | |
5f40b3cb ZD |
1240 | if (!parallelized_functions) |
1241 | parallelized_functions = BITMAP_GGC_ALLOC (); | |
1242 | bitmap_set_bit (parallelized_functions, DECL_UID (decl)); | |
1243 | ||
1244 | TREE_STATIC (decl) = 1; | |
1245 | TREE_USED (decl) = 1; | |
1246 | DECL_ARTIFICIAL (decl) = 1; | |
1247 | DECL_IGNORED_P (decl) = 0; | |
1248 | TREE_PUBLIC (decl) = 0; | |
1249 | DECL_UNINLINABLE (decl) = 1; | |
1250 | DECL_EXTERNAL (decl) = 0; | |
1251 | DECL_CONTEXT (decl) = NULL_TREE; | |
1252 | DECL_INITIAL (decl) = make_node (BLOCK); | |
1253 | ||
c2255bc4 AH |
1254 | t = build_decl (BUILTINS_LOCATION, |
1255 | RESULT_DECL, NULL_TREE, void_type_node); | |
5f40b3cb ZD |
1256 | DECL_ARTIFICIAL (t) = 1; |
1257 | DECL_IGNORED_P (t) = 1; | |
1258 | DECL_RESULT (decl) = t; | |
1259 | ||
c2255bc4 AH |
1260 | t = build_decl (BUILTINS_LOCATION, |
1261 | PARM_DECL, get_identifier (".paral_data_param"), | |
5f40b3cb ZD |
1262 | ptr_type_node); |
1263 | DECL_ARTIFICIAL (t) = 1; | |
1264 | DECL_ARG_TYPE (t) = ptr_type_node; | |
1265 | DECL_CONTEXT (t) = decl; | |
1266 | TREE_USED (t) = 1; | |
1267 | DECL_ARGUMENTS (decl) = t; | |
1268 | ||
182e0d71 | 1269 | allocate_struct_function (decl, false); |
5f40b3cb ZD |
1270 | |
1271 | /* The call to allocate_struct_function clobbers CFUN, so we need to restore | |
1272 | it. */ | |
5576d6f2 | 1273 | set_cfun (act_cfun); |
5f40b3cb ZD |
1274 | |
1275 | return decl; | |
1276 | } | |
1277 | ||
5f40b3cb ZD |
1278 | /* Moves the exit condition of LOOP to the beginning of its header, and |
1279 | duplicates the part of the last iteration that gets disabled to the | |
1280 | exit of the loop. NIT is the number of iterations of the loop | |
1281 | (used to initialize the variables in the duplicated part). | |
1282 | ||
fa10beec | 1283 | TODO: the common case is that latch of the loop is empty and immediately |
5f40b3cb ZD |
1284 | follows the loop exit. In this case, it would be better not to copy the |
1285 | body of the loop, but only move the entry of the loop directly before the | |
1286 | exit check and increase the number of iterations of the loop by one. | |
a509ebb5 RL |
1287 | This may need some additional preconditioning in case NIT = ~0. |
1288 | REDUCTION_LIST describes the reductions in LOOP. */ | |
5f40b3cb ZD |
1289 | |
1290 | static void | |
a509ebb5 | 1291 | transform_to_exit_first_loop (struct loop *loop, htab_t reduction_list, tree nit) |
5f40b3cb ZD |
1292 | { |
1293 | basic_block *bbs, *nbbs, ex_bb, orig_header; | |
1294 | unsigned n; | |
1295 | bool ok; | |
1296 | edge exit = single_dom_exit (loop), hpred; | |
726a989a | 1297 | tree control, control_name, res, t; |
48710229 | 1298 | gimple phi, nphi, cond_stmt, stmt, cond_nit; |
726a989a | 1299 | gimple_stmt_iterator gsi; |
48710229 | 1300 | tree nit_1; |
5f40b3cb ZD |
1301 | |
1302 | split_block_after_labels (loop->header); | |
1303 | orig_header = single_succ (loop->header); | |
1304 | hpred = single_succ_edge (loop->header); | |
1305 | ||
1306 | cond_stmt = last_stmt (exit->src); | |
726a989a RB |
1307 | control = gimple_cond_lhs (cond_stmt); |
1308 | gcc_assert (gimple_cond_rhs (cond_stmt) == nit); | |
5f40b3cb ZD |
1309 | |
1310 | /* Make sure that we have phi nodes on exit for all loop header phis | |
1311 | (create_parallel_loop requires that). */ | |
726a989a | 1312 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) |
5f40b3cb | 1313 | { |
726a989a | 1314 | phi = gsi_stmt (gsi); |
5f40b3cb ZD |
1315 | res = PHI_RESULT (phi); |
1316 | t = make_ssa_name (SSA_NAME_VAR (res), phi); | |
1317 | SET_PHI_RESULT (phi, t); | |
5f40b3cb ZD |
1318 | nphi = create_phi_node (res, orig_header); |
1319 | SSA_NAME_DEF_STMT (res) = nphi; | |
f5045c96 | 1320 | add_phi_arg (nphi, t, hpred, UNKNOWN_LOCATION); |
5f40b3cb ZD |
1321 | |
1322 | if (res == control) | |
1323 | { | |
726a989a | 1324 | gimple_cond_set_lhs (cond_stmt, t); |
5f40b3cb ZD |
1325 | update_stmt (cond_stmt); |
1326 | control = t; | |
1327 | } | |
1328 | } | |
5f40b3cb | 1329 | bbs = get_loop_body_in_dom_order (loop); |
48710229 RL |
1330 | |
1331 | for (n = 0; bbs[n] != loop->latch; n++) | |
5f40b3cb | 1332 | continue; |
48710229 | 1333 | n--; |
5f40b3cb | 1334 | nbbs = XNEWVEC (basic_block, n); |
726a989a RB |
1335 | ok = gimple_duplicate_sese_tail (single_succ_edge (loop->header), exit, |
1336 | bbs + 1, n, nbbs); | |
5f40b3cb ZD |
1337 | gcc_assert (ok); |
1338 | free (bbs); | |
1339 | ex_bb = nbbs[0]; | |
1340 | free (nbbs); | |
1341 | ||
a509ebb5 | 1342 | /* Other than reductions, the only gimple reg that should be copied |
726a989a | 1343 | out of the loop is the control variable. */ |
a509ebb5 | 1344 | |
5f40b3cb | 1345 | control_name = NULL_TREE; |
726a989a | 1346 | for (gsi = gsi_start_phis (ex_bb); !gsi_end_p (gsi); ) |
5f40b3cb | 1347 | { |
726a989a | 1348 | phi = gsi_stmt (gsi); |
5f40b3cb ZD |
1349 | res = PHI_RESULT (phi); |
1350 | if (!is_gimple_reg (res)) | |
726a989a RB |
1351 | { |
1352 | gsi_next (&gsi); | |
1353 | continue; | |
1354 | } | |
5f40b3cb | 1355 | |
a509ebb5 RL |
1356 | /* Check if it is a part of reduction. If it is, |
1357 | keep the phi at the reduction's keep_res field. The | |
1358 | PHI_RESULT of this phi is the resulting value of the reduction | |
1359 | variable when exiting the loop. */ | |
1360 | ||
1361 | exit = single_dom_exit (loop); | |
1362 | ||
1363 | if (htab_elements (reduction_list) > 0) | |
1364 | { | |
1365 | struct reduction_info *red; | |
1366 | ||
1367 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
a509ebb5 RL |
1368 | red = reduction_phi (reduction_list, SSA_NAME_DEF_STMT (val)); |
1369 | if (red) | |
726a989a RB |
1370 | { |
1371 | red->keep_res = phi; | |
1372 | gsi_next (&gsi); | |
1373 | continue; | |
1374 | } | |
a509ebb5 | 1375 | } |
726a989a RB |
1376 | gcc_assert (control_name == NULL_TREE |
1377 | && SSA_NAME_VAR (res) == SSA_NAME_VAR (control)); | |
5f40b3cb | 1378 | control_name = res; |
726a989a | 1379 | remove_phi_node (&gsi, false); |
5f40b3cb ZD |
1380 | } |
1381 | gcc_assert (control_name != NULL_TREE); | |
5f40b3cb | 1382 | |
48710229 RL |
1383 | /* Initialize the control variable to number of iterations |
1384 | according to the rhs of the exit condition. */ | |
726a989a | 1385 | gsi = gsi_after_labels (ex_bb); |
48710229 RL |
1386 | cond_nit = last_stmt (exit->src); |
1387 | nit_1 = gimple_cond_rhs (cond_nit); | |
1388 | nit_1 = force_gimple_operand_gsi (&gsi, | |
1389 | fold_convert (TREE_TYPE (control_name), nit_1), | |
726a989a | 1390 | false, NULL_TREE, false, GSI_SAME_STMT); |
48710229 | 1391 | stmt = gimple_build_assign (control_name, nit_1); |
726a989a RB |
1392 | gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
1393 | SSA_NAME_DEF_STMT (control_name) = stmt; | |
5f40b3cb ZD |
1394 | } |
1395 | ||
1396 | /* Create the parallel constructs for LOOP as described in gen_parallel_loop. | |
726a989a | 1397 | LOOP_FN and DATA are the arguments of GIMPLE_OMP_PARALLEL. |
5f40b3cb ZD |
1398 | NEW_DATA is the variable that should be initialized from the argument |
1399 | of LOOP_FN. N_THREADS is the requested number of threads. Returns the | |
726a989a | 1400 | basic block containing GIMPLE_OMP_PARALLEL tree. */ |
5f40b3cb ZD |
1401 | |
1402 | static basic_block | |
1403 | create_parallel_loop (struct loop *loop, tree loop_fn, tree data, | |
1404 | tree new_data, unsigned n_threads) | |
1405 | { | |
726a989a | 1406 | gimple_stmt_iterator gsi; |
5f40b3cb | 1407 | basic_block bb, paral_bb, for_bb, ex_bb; |
726a989a RB |
1408 | tree t, param, res; |
1409 | gimple stmt, for_stmt, phi, cond_stmt; | |
1410 | tree cvar, cvar_init, initvar, cvar_next, cvar_base, type; | |
5f40b3cb ZD |
1411 | edge exit, nexit, guard, end, e; |
1412 | ||
726a989a | 1413 | /* Prepare the GIMPLE_OMP_PARALLEL statement. */ |
5f40b3cb ZD |
1414 | bb = loop_preheader_edge (loop)->src; |
1415 | paral_bb = single_pred (bb); | |
726a989a | 1416 | gsi = gsi_last_bb (paral_bb); |
5f40b3cb | 1417 | |
c2255bc4 | 1418 | t = build_omp_clause (BUILTINS_LOCATION, OMP_CLAUSE_NUM_THREADS); |
5f40b3cb | 1419 | OMP_CLAUSE_NUM_THREADS_EXPR (t) |
a509ebb5 | 1420 | = build_int_cst (integer_type_node, n_threads); |
726a989a | 1421 | stmt = gimple_build_omp_parallel (NULL, t, loop_fn, data); |
5f40b3cb | 1422 | |
726a989a | 1423 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); |
5f40b3cb ZD |
1424 | |
1425 | /* Initialize NEW_DATA. */ | |
1426 | if (data) | |
1427 | { | |
726a989a RB |
1428 | gsi = gsi_after_labels (bb); |
1429 | ||
1430 | param = make_ssa_name (DECL_ARGUMENTS (loop_fn), NULL); | |
1431 | stmt = gimple_build_assign (param, build_fold_addr_expr (data)); | |
1432 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
1433 | SSA_NAME_DEF_STMT (param) = stmt; | |
1434 | ||
1435 | stmt = gimple_build_assign (new_data, | |
1436 | fold_convert (TREE_TYPE (new_data), param)); | |
1437 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
1438 | SSA_NAME_DEF_STMT (new_data) = stmt; | |
5f40b3cb ZD |
1439 | } |
1440 | ||
726a989a | 1441 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_PARALLEL. */ |
5f40b3cb | 1442 | bb = split_loop_exit_edge (single_dom_exit (loop)); |
726a989a RB |
1443 | gsi = gsi_last_bb (bb); |
1444 | gsi_insert_after (&gsi, gimple_build_omp_return (false), GSI_NEW_STMT); | |
5f40b3cb | 1445 | |
726a989a | 1446 | /* Extract data for GIMPLE_OMP_FOR. */ |
5f40b3cb | 1447 | gcc_assert (loop->header == single_dom_exit (loop)->src); |
726a989a | 1448 | cond_stmt = last_stmt (loop->header); |
5f40b3cb | 1449 | |
726a989a | 1450 | cvar = gimple_cond_lhs (cond_stmt); |
5f40b3cb ZD |
1451 | cvar_base = SSA_NAME_VAR (cvar); |
1452 | phi = SSA_NAME_DEF_STMT (cvar); | |
1453 | cvar_init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
726a989a | 1454 | initvar = make_ssa_name (cvar_base, NULL); |
5f40b3cb ZD |
1455 | SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, loop_preheader_edge (loop)), |
1456 | initvar); | |
1457 | cvar_next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
1458 | ||
726a989a RB |
1459 | gsi = gsi_last_bb (loop->latch); |
1460 | gcc_assert (gsi_stmt (gsi) == SSA_NAME_DEF_STMT (cvar_next)); | |
1461 | gsi_remove (&gsi, true); | |
5f40b3cb ZD |
1462 | |
1463 | /* Prepare cfg. */ | |
1464 | for_bb = split_edge (loop_preheader_edge (loop)); | |
1465 | ex_bb = split_loop_exit_edge (single_dom_exit (loop)); | |
1466 | extract_true_false_edges_from_block (loop->header, &nexit, &exit); | |
1467 | gcc_assert (exit == single_dom_exit (loop)); | |
1468 | ||
1469 | guard = make_edge (for_bb, ex_bb, 0); | |
1470 | single_succ_edge (loop->latch)->flags = 0; | |
1471 | end = make_edge (loop->latch, ex_bb, EDGE_FALLTHRU); | |
726a989a | 1472 | for (gsi = gsi_start_phis (ex_bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
5f40b3cb | 1473 | { |
f5045c96 AM |
1474 | source_location locus; |
1475 | tree def; | |
726a989a | 1476 | phi = gsi_stmt (gsi); |
5f40b3cb | 1477 | res = PHI_RESULT (phi); |
726a989a | 1478 | stmt = SSA_NAME_DEF_STMT (PHI_ARG_DEF_FROM_EDGE (phi, exit)); |
f5045c96 AM |
1479 | |
1480 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_preheader_edge (loop)); | |
1481 | locus = gimple_phi_arg_location_from_edge (stmt, | |
1482 | loop_preheader_edge (loop)); | |
1483 | add_phi_arg (phi, def, guard, locus); | |
1484 | ||
1485 | def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (loop)); | |
1486 | locus = gimple_phi_arg_location_from_edge (stmt, loop_latch_edge (loop)); | |
1487 | add_phi_arg (phi, def, end, locus); | |
5f40b3cb ZD |
1488 | } |
1489 | e = redirect_edge_and_branch (exit, nexit->dest); | |
1490 | PENDING_STMT (e) = NULL; | |
1491 | ||
726a989a RB |
1492 | /* Emit GIMPLE_OMP_FOR. */ |
1493 | gimple_cond_set_lhs (cond_stmt, cvar_base); | |
5f40b3cb | 1494 | type = TREE_TYPE (cvar); |
c2255bc4 | 1495 | t = build_omp_clause (BUILTINS_LOCATION, OMP_CLAUSE_SCHEDULE); |
5f40b3cb ZD |
1496 | OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_STATIC; |
1497 | ||
726a989a RB |
1498 | for_stmt = gimple_build_omp_for (NULL, t, 1, NULL); |
1499 | gimple_omp_for_set_index (for_stmt, 0, initvar); | |
1500 | gimple_omp_for_set_initial (for_stmt, 0, cvar_init); | |
1501 | gimple_omp_for_set_final (for_stmt, 0, gimple_cond_rhs (cond_stmt)); | |
1502 | gimple_omp_for_set_cond (for_stmt, 0, gimple_cond_code (cond_stmt)); | |
1503 | gimple_omp_for_set_incr (for_stmt, 0, build2 (PLUS_EXPR, type, | |
1504 | cvar_base, | |
1505 | build_int_cst (type, 1))); | |
1506 | ||
1507 | gsi = gsi_last_bb (for_bb); | |
1508 | gsi_insert_after (&gsi, for_stmt, GSI_NEW_STMT); | |
5f40b3cb ZD |
1509 | SSA_NAME_DEF_STMT (initvar) = for_stmt; |
1510 | ||
726a989a RB |
1511 | /* Emit GIMPLE_OMP_CONTINUE. */ |
1512 | gsi = gsi_last_bb (loop->latch); | |
1513 | stmt = gimple_build_omp_continue (cvar_next, cvar); | |
1514 | gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1515 | SSA_NAME_DEF_STMT (cvar_next) = stmt; | |
5f40b3cb | 1516 | |
726a989a RB |
1517 | /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_FOR. */ |
1518 | gsi = gsi_last_bb (ex_bb); | |
1519 | gsi_insert_after (&gsi, gimple_build_omp_return (true), GSI_NEW_STMT); | |
5f40b3cb ZD |
1520 | |
1521 | return paral_bb; | |
1522 | } | |
1523 | ||
08dab97a RL |
1524 | /* Generates code to execute the iterations of LOOP in N_THREADS |
1525 | threads in parallel. | |
1526 | ||
1527 | NITER describes number of iterations of LOOP. | |
fa10beec | 1528 | REDUCTION_LIST describes the reductions existent in the LOOP. */ |
5f40b3cb ZD |
1529 | |
1530 | static void | |
08dab97a | 1531 | gen_parallel_loop (struct loop *loop, htab_t reduction_list, |
a509ebb5 | 1532 | unsigned n_threads, struct tree_niter_desc *niter) |
5f40b3cb ZD |
1533 | { |
1534 | struct loop *nloop; | |
9326236d | 1535 | loop_iterator li; |
5f40b3cb | 1536 | tree many_iterations_cond, type, nit; |
726a989a RB |
1537 | tree arg_struct, new_arg_struct; |
1538 | gimple_seq stmts; | |
5f40b3cb | 1539 | basic_block parallel_head; |
9f9f72aa | 1540 | edge entry, exit; |
a509ebb5 | 1541 | struct clsn_data clsn_data; |
5f40b3cb ZD |
1542 | unsigned prob; |
1543 | ||
1544 | /* From | |
1545 | ||
1546 | --------------------------------------------------------------------- | |
1547 | loop | |
1548 | { | |
1549 | IV = phi (INIT, IV + STEP) | |
1550 | BODY1; | |
1551 | if (COND) | |
1552 | break; | |
1553 | BODY2; | |
1554 | } | |
1555 | --------------------------------------------------------------------- | |
1556 | ||
1557 | with # of iterations NITER (possibly with MAY_BE_ZERO assumption), | |
1558 | we generate the following code: | |
1559 | ||
1560 | --------------------------------------------------------------------- | |
1561 | ||
1562 | if (MAY_BE_ZERO | |
a509ebb5 RL |
1563 | || NITER < MIN_PER_THREAD * N_THREADS) |
1564 | goto original; | |
5f40b3cb ZD |
1565 | |
1566 | BODY1; | |
1567 | store all local loop-invariant variables used in body of the loop to DATA. | |
726a989a | 1568 | GIMPLE_OMP_PARALLEL (OMP_CLAUSE_NUM_THREADS (N_THREADS), LOOPFN, DATA); |
5f40b3cb | 1569 | load the variables from DATA. |
726a989a | 1570 | GIMPLE_OMP_FOR (IV = INIT; COND; IV += STEP) (OMP_CLAUSE_SCHEDULE (static)) |
5f40b3cb ZD |
1571 | BODY2; |
1572 | BODY1; | |
726a989a RB |
1573 | GIMPLE_OMP_CONTINUE; |
1574 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_FOR | |
1575 | GIMPLE_OMP_RETURN -- GIMPLE_OMP_PARALLEL | |
5f40b3cb ZD |
1576 | goto end; |
1577 | ||
1578 | original: | |
1579 | loop | |
1580 | { | |
1581 | IV = phi (INIT, IV + STEP) | |
1582 | BODY1; | |
1583 | if (COND) | |
1584 | break; | |
1585 | BODY2; | |
1586 | } | |
1587 | ||
1588 | end: | |
1589 | ||
1590 | */ | |
1591 | ||
1592 | /* Create two versions of the loop -- in the old one, we know that the | |
1593 | number of iterations is large enough, and we will transform it into the | |
1594 | loop that will be split to loop_fn, the new one will be used for the | |
1595 | remaining iterations. */ | |
a509ebb5 | 1596 | |
5f40b3cb ZD |
1597 | type = TREE_TYPE (niter->niter); |
1598 | nit = force_gimple_operand (unshare_expr (niter->niter), &stmts, true, | |
1599 | NULL_TREE); | |
1600 | if (stmts) | |
726a989a | 1601 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb ZD |
1602 | |
1603 | many_iterations_cond = | |
a509ebb5 RL |
1604 | fold_build2 (GE_EXPR, boolean_type_node, |
1605 | nit, build_int_cst (type, MIN_PER_THREAD * n_threads)); | |
5f40b3cb | 1606 | many_iterations_cond |
a509ebb5 RL |
1607 | = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, |
1608 | invert_truthvalue (unshare_expr (niter->may_be_zero)), | |
1609 | many_iterations_cond); | |
5f40b3cb | 1610 | many_iterations_cond |
a509ebb5 | 1611 | = force_gimple_operand (many_iterations_cond, &stmts, false, NULL_TREE); |
5f40b3cb | 1612 | if (stmts) |
726a989a | 1613 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb ZD |
1614 | if (!is_gimple_condexpr (many_iterations_cond)) |
1615 | { | |
1616 | many_iterations_cond | |
a509ebb5 RL |
1617 | = force_gimple_operand (many_iterations_cond, &stmts, |
1618 | true, NULL_TREE); | |
5f40b3cb | 1619 | if (stmts) |
726a989a | 1620 | gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); |
5f40b3cb ZD |
1621 | } |
1622 | ||
1623 | initialize_original_copy_tables (); | |
1624 | ||
1625 | /* We assume that the loop usually iterates a lot. */ | |
1626 | prob = 4 * REG_BR_PROB_BASE / 5; | |
1627 | nloop = loop_version (loop, many_iterations_cond, NULL, | |
1628 | prob, prob, REG_BR_PROB_BASE - prob, true); | |
1629 | update_ssa (TODO_update_ssa); | |
1630 | free_original_copy_tables (); | |
1631 | ||
1632 | /* Base all the induction variables in LOOP on a single control one. */ | |
08dab97a | 1633 | canonicalize_loop_ivs (loop, &nit); |
5f40b3cb ZD |
1634 | |
1635 | /* Ensure that the exit condition is the first statement in the loop. */ | |
a509ebb5 RL |
1636 | transform_to_exit_first_loop (loop, reduction_list, nit); |
1637 | ||
fa10beec | 1638 | /* Generate initializations for reductions. */ |
a509ebb5 RL |
1639 | if (htab_elements (reduction_list) > 0) |
1640 | htab_traverse (reduction_list, initialize_reductions, loop); | |
5f40b3cb ZD |
1641 | |
1642 | /* Eliminate the references to local variables from the loop. */ | |
9f9f72aa AP |
1643 | gcc_assert (single_exit (loop)); |
1644 | entry = loop_preheader_edge (loop); | |
1645 | exit = single_dom_exit (loop); | |
5f40b3cb | 1646 | |
9f9f72aa | 1647 | eliminate_local_variables (entry, exit); |
5f40b3cb ZD |
1648 | /* In the old loop, move all variables non-local to the loop to a structure |
1649 | and back, and create separate decls for the variables used in loop. */ | |
9f9f72aa AP |
1650 | separate_decls_in_region (entry, exit, reduction_list, &arg_struct, |
1651 | &new_arg_struct, &clsn_data); | |
5f40b3cb ZD |
1652 | |
1653 | /* Create the parallel constructs. */ | |
1654 | parallel_head = create_parallel_loop (loop, create_loop_fn (), arg_struct, | |
1655 | new_arg_struct, n_threads); | |
a509ebb5 RL |
1656 | if (htab_elements (reduction_list) > 0) |
1657 | create_call_for_reduction (loop, reduction_list, &clsn_data); | |
5f40b3cb ZD |
1658 | |
1659 | scev_reset (); | |
1660 | ||
1661 | /* Cancel the loop (it is simpler to do it here rather than to teach the | |
1662 | expander to do it). */ | |
1663 | cancel_loop_tree (loop); | |
1664 | ||
92a6bdbd SP |
1665 | /* Free loop bound estimations that could contain references to |
1666 | removed statements. */ | |
1667 | FOR_EACH_LOOP (li, loop, 0) | |
1668 | free_numbers_of_iterations_estimates_loop (loop); | |
1669 | ||
5f40b3cb ZD |
1670 | /* Expand the parallel constructs. We do it directly here instead of running |
1671 | a separate expand_omp pass, since it is more efficient, and less likely to | |
1672 | cause troubles with further analyses not being able to deal with the | |
1673 | OMP trees. */ | |
a509ebb5 | 1674 | |
5f40b3cb ZD |
1675 | omp_expand_local (parallel_head); |
1676 | } | |
1677 | ||
9857228c SP |
1678 | /* Returns true when LOOP contains vector phi nodes. */ |
1679 | ||
1680 | static bool | |
726a989a | 1681 | loop_has_vector_phi_nodes (struct loop *loop ATTRIBUTE_UNUSED) |
9857228c SP |
1682 | { |
1683 | unsigned i; | |
1684 | basic_block *bbs = get_loop_body_in_dom_order (loop); | |
726a989a | 1685 | gimple_stmt_iterator gsi; |
9857228c | 1686 | bool res = true; |
9857228c SP |
1687 | |
1688 | for (i = 0; i < loop->num_nodes; i++) | |
726a989a RB |
1689 | for (gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) |
1690 | if (TREE_CODE (TREE_TYPE (PHI_RESULT (gsi_stmt (gsi)))) == VECTOR_TYPE) | |
9857228c SP |
1691 | goto end; |
1692 | ||
1693 | res = false; | |
1694 | end: | |
1695 | free (bbs); | |
1696 | return res; | |
1697 | } | |
1698 | ||
08dab97a RL |
1699 | /* Create a reduction_info struct, initialize it with REDUC_STMT |
1700 | and PHI, insert it to the REDUCTION_LIST. */ | |
1701 | ||
1702 | static void | |
1703 | build_new_reduction (htab_t reduction_list, gimple reduc_stmt, gimple phi) | |
1704 | { | |
1705 | PTR *slot; | |
1706 | struct reduction_info *new_reduction; | |
1707 | ||
1708 | gcc_assert (reduc_stmt); | |
1709 | ||
1710 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1711 | { | |
1712 | fprintf (dump_file, | |
1713 | "Detected reduction. reduction stmt is: \n"); | |
1714 | print_gimple_stmt (dump_file, reduc_stmt, 0, 0); | |
1715 | fprintf (dump_file, "\n"); | |
1716 | } | |
1717 | ||
1718 | new_reduction = XCNEW (struct reduction_info); | |
1719 | ||
1720 | new_reduction->reduc_stmt = reduc_stmt; | |
1721 | new_reduction->reduc_phi = phi; | |
1722 | new_reduction->reduction_code = gimple_assign_rhs_code (reduc_stmt); | |
1723 | slot = htab_find_slot (reduction_list, new_reduction, INSERT); | |
1724 | *slot = new_reduction; | |
1725 | } | |
1726 | ||
1727 | /* Detect all reductions in the LOOP, insert them into REDUCTION_LIST. */ | |
1728 | ||
1729 | static void | |
1730 | gather_scalar_reductions (loop_p loop, htab_t reduction_list) | |
1731 | { | |
1732 | gimple_stmt_iterator gsi; | |
1733 | loop_vec_info simple_loop_info; | |
1734 | ||
1735 | vect_dump = NULL; | |
1736 | simple_loop_info = vect_analyze_loop_form (loop); | |
1737 | ||
1738 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1739 | { | |
1740 | gimple phi = gsi_stmt (gsi); | |
1741 | affine_iv iv; | |
1742 | tree res = PHI_RESULT (phi); | |
1743 | bool double_reduc; | |
1744 | ||
1745 | if (!is_gimple_reg (res)) | |
1746 | continue; | |
1747 | ||
1748 | if (!simple_iv (loop, loop, res, &iv, true) | |
1749 | && simple_loop_info) | |
1750 | { | |
1751 | gimple reduc_stmt = vect_is_simple_reduction (simple_loop_info, phi, true, &double_reduc); | |
48710229 | 1752 | if (reduc_stmt && !double_reduc) |
08dab97a RL |
1753 | build_new_reduction (reduction_list, reduc_stmt, phi); |
1754 | } | |
1755 | } | |
1756 | destroy_loop_vec_info (simple_loop_info, true); | |
1757 | } | |
1758 | ||
1759 | /* Try to initialize NITER for code generation part. */ | |
1760 | ||
1761 | static bool | |
1762 | try_get_loop_niter (loop_p loop, struct tree_niter_desc *niter) | |
1763 | { | |
1764 | edge exit = single_dom_exit (loop); | |
1765 | ||
1766 | gcc_assert (exit); | |
1767 | ||
1768 | /* We need to know # of iterations, and there should be no uses of values | |
1769 | defined inside loop outside of it, unless the values are invariants of | |
1770 | the loop. */ | |
1771 | if (!number_of_iterations_exit (loop, exit, niter, false)) | |
1772 | { | |
1773 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1774 | fprintf (dump_file, " FAILED: number of iterations not known\n"); | |
1775 | return false; | |
1776 | } | |
1777 | ||
1778 | return true; | |
1779 | } | |
1780 | ||
1781 | /* Try to initialize REDUCTION_LIST for code generation part. | |
1782 | REDUCTION_LIST describes the reductions. */ | |
1783 | ||
1784 | static bool | |
1785 | try_create_reduction_list (loop_p loop, htab_t reduction_list) | |
1786 | { | |
1787 | edge exit = single_dom_exit (loop); | |
1788 | gimple_stmt_iterator gsi; | |
1789 | ||
1790 | gcc_assert (exit); | |
1791 | ||
1792 | gather_scalar_reductions (loop, reduction_list); | |
1793 | ||
1794 | ||
1795 | for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1796 | { | |
1797 | gimple phi = gsi_stmt (gsi); | |
1798 | struct reduction_info *red; | |
1799 | imm_use_iterator imm_iter; | |
1800 | use_operand_p use_p; | |
1801 | gimple reduc_phi; | |
1802 | tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
1803 | ||
1804 | if (is_gimple_reg (val)) | |
1805 | { | |
1806 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1807 | { | |
1808 | fprintf (dump_file, "phi is "); | |
1809 | print_gimple_stmt (dump_file, phi, 0, 0); | |
1810 | fprintf (dump_file, "arg of phi to exit: value "); | |
1811 | print_generic_expr (dump_file, val, 0); | |
1812 | fprintf (dump_file, " used outside loop\n"); | |
1813 | fprintf (dump_file, | |
1814 | " checking if it a part of reduction pattern: \n"); | |
1815 | } | |
1816 | if (htab_elements (reduction_list) == 0) | |
1817 | { | |
1818 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1819 | fprintf (dump_file, | |
1820 | " FAILED: it is not a part of reduction.\n"); | |
1821 | return false; | |
1822 | } | |
1823 | reduc_phi = NULL; | |
1824 | FOR_EACH_IMM_USE_FAST (use_p, imm_iter, val) | |
1825 | { | |
1826 | if (flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p)))) | |
1827 | { | |
1828 | reduc_phi = USE_STMT (use_p); | |
1829 | break; | |
1830 | } | |
1831 | } | |
1832 | red = reduction_phi (reduction_list, reduc_phi); | |
1833 | if (red == NULL) | |
1834 | { | |
1835 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1836 | fprintf (dump_file, | |
1837 | " FAILED: it is not a part of reduction.\n"); | |
1838 | return false; | |
1839 | } | |
1840 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1841 | { | |
1842 | fprintf (dump_file, "reduction phi is "); | |
1843 | print_gimple_stmt (dump_file, red->reduc_phi, 0, 0); | |
1844 | fprintf (dump_file, "reduction stmt is "); | |
1845 | print_gimple_stmt (dump_file, red->reduc_stmt, 0, 0); | |
1846 | } | |
1847 | } | |
1848 | } | |
1849 | ||
1850 | /* The iterations of the loop may communicate only through bivs whose | |
1851 | iteration space can be distributed efficiently. */ | |
1852 | for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1853 | { | |
1854 | gimple phi = gsi_stmt (gsi); | |
1855 | tree def = PHI_RESULT (phi); | |
1856 | affine_iv iv; | |
1857 | ||
1858 | if (is_gimple_reg (def) && !simple_iv (loop, loop, def, &iv, true)) | |
1859 | { | |
1860 | struct reduction_info *red; | |
1861 | ||
1862 | red = reduction_phi (reduction_list, phi); | |
1863 | if (red == NULL) | |
1864 | { | |
1865 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1866 | fprintf (dump_file, | |
1867 | " FAILED: scalar dependency between iterations\n"); | |
1868 | return false; | |
1869 | } | |
1870 | } | |
1871 | } | |
1872 | ||
1873 | ||
1874 | return true; | |
1875 | } | |
1876 | ||
5f40b3cb ZD |
1877 | /* Detect parallel loops and generate parallel code using libgomp |
1878 | primitives. Returns true if some loop was parallelized, false | |
1879 | otherwise. */ | |
1880 | ||
1881 | bool | |
1882 | parallelize_loops (void) | |
1883 | { | |
1884 | unsigned n_threads = flag_tree_parallelize_loops; | |
1885 | bool changed = false; | |
1886 | struct loop *loop; | |
1887 | struct tree_niter_desc niter_desc; | |
1888 | loop_iterator li; | |
a509ebb5 | 1889 | htab_t reduction_list; |
5f40b3cb ZD |
1890 | |
1891 | /* Do not parallelize loops in the functions created by parallelization. */ | |
1892 | if (parallelized_function_p (cfun->decl)) | |
1893 | return false; | |
1894 | ||
a509ebb5 | 1895 | reduction_list = htab_create (10, reduction_info_hash, |
08dab97a | 1896 | reduction_info_eq, free); |
726a989a | 1897 | init_stmt_vec_info_vec (); |
a509ebb5 | 1898 | |
5f40b3cb ZD |
1899 | FOR_EACH_LOOP (li, loop, 0) |
1900 | { | |
a509ebb5 | 1901 | htab_empty (reduction_list); |
48710229 RL |
1902 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1903 | { | |
1904 | fprintf (dump_file, "Trying loop %d as candidate\n",loop->num); | |
1905 | if (loop->inner) | |
1906 | fprintf (dump_file, "loop %d is not innermost\n",loop->num); | |
1907 | else | |
1908 | fprintf (dump_file, "loop %d is innermost\n",loop->num); | |
1909 | } | |
1910 | ||
1911 | /* If we use autopar in graphite pass, we use its marked dependency | |
87d4d0ee SP |
1912 | checking results. */ |
1913 | if (flag_loop_parallelize_all && !loop->can_be_parallel) | |
48710229 RL |
1914 | { |
1915 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1916 | fprintf (dump_file, "loop is not parallel according to graphite\n"); | |
87d4d0ee | 1917 | continue; |
48710229 | 1918 | } |
87d4d0ee | 1919 | |
48710229 RL |
1920 | if (!single_dom_exit (loop)) |
1921 | { | |
1922 | ||
1923 | if (dump_file && (dump_flags & TDF_DETAILS)) | |
1924 | fprintf (dump_file, "loop is !single_dom_exit\n"); | |
1925 | ||
08dab97a | 1926 | continue; |
48710229 | 1927 | } |
08dab97a RL |
1928 | |
1929 | if (/* And of course, the loop must be parallelizable. */ | |
1930 | !can_duplicate_loop_p (loop) | |
1d4af1e8 | 1931 | || loop_has_blocks_with_irreducible_flag (loop) |
9857228c | 1932 | /* FIXME: the check for vector phi nodes could be removed. */ |
08dab97a RL |
1933 | || loop_has_vector_phi_nodes (loop)) |
1934 | continue; | |
87d4d0ee SP |
1935 | |
1936 | /* FIXME: Bypass this check as graphite doesn't update the | |
1937 | count and frequency correctly now. */ | |
1938 | if (!flag_loop_parallelize_all | |
8e094aa2 L |
1939 | && ((estimated_loop_iterations_int (loop, false) |
1940 | <= (HOST_WIDE_INT) n_threads * MIN_PER_THREAD) | |
87d4d0ee SP |
1941 | /* Do not bother with loops in cold areas. */ |
1942 | || optimize_loop_nest_for_size_p (loop))) | |
08dab97a | 1943 | continue; |
48710229 | 1944 | |
08dab97a RL |
1945 | if (!try_get_loop_niter (loop, &niter_desc)) |
1946 | continue; | |
1947 | ||
1948 | if (!try_create_reduction_list (loop, reduction_list)) | |
1949 | continue; | |
1950 | ||
87d4d0ee | 1951 | if (!flag_loop_parallelize_all && !loop_parallel_p (loop)) |
5f40b3cb ZD |
1952 | continue; |
1953 | ||
1954 | changed = true; | |
48710229 RL |
1955 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1956 | { | |
1957 | fprintf (dump_file, "parallelizing "); | |
1958 | if (loop->inner) | |
1959 | fprintf (dump_file, "outer loop\n"); | |
1960 | else | |
1961 | fprintf (dump_file, "inner loop\n"); | |
1962 | } | |
08dab97a RL |
1963 | gen_parallel_loop (loop, reduction_list, |
1964 | n_threads, &niter_desc); | |
5f40b3cb ZD |
1965 | verify_flow_info (); |
1966 | verify_dominators (CDI_DOMINATORS); | |
1967 | verify_loop_structure (); | |
1968 | verify_loop_closed_ssa (); | |
1969 | } | |
1970 | ||
726a989a | 1971 | free_stmt_vec_info_vec (); |
a509ebb5 | 1972 | htab_delete (reduction_list); |
6b8ed145 RG |
1973 | |
1974 | /* Parallelization will cause new function calls to be inserted through | |
1975 | which local variables will escape. Reset the points-to solutions | |
1976 | for ESCAPED and CALLUSED. */ | |
1977 | if (changed) | |
1978 | { | |
1979 | pt_solution_reset (&cfun->gimple_df->escaped); | |
1980 | pt_solution_reset (&cfun->gimple_df->callused); | |
1981 | } | |
1982 | ||
5f40b3cb ZD |
1983 | return changed; |
1984 | } | |
1985 | ||
1986 | #include "gt-tree-parloops.h" |