Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-parloops.c @ 55:77e2b8dfacca gcc-4.4.5
update it from 4.4.3 to 4.5.0
author | ryoma <e075725@ie.u-ryukyu.ac.jp> |
---|---|
date | Fri, 12 Feb 2010 23:39:51 +0900 |
parents | a06113de4d67 |
children | b7f97abdc517 |
rev | line source |
---|---|
0 | 1 /* Loop autoparallelization. |
2 Copyright (C) 2006, 2007, 2008, 2009 Free Software Foundation, Inc. | |
3 Contributed by Sebastian Pop <pop@cri.ensmp.fr> and | |
4 Zdenek Dvorak <dvorakz@suse.cz>. | |
5 | |
6 This file is part of GCC. | |
7 | |
8 GCC is free software; you can redistribute it and/or modify it under | |
9 the terms of the GNU General Public License as published by the Free | |
10 Software Foundation; either version 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 "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" | |
37 #include "tree-vectorizer.h" | |
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 | |
43 add GIMPLE_OMP_PARALLEL and GIMPLE_OMP_FOR codes and let omp expansion | |
44 machinery do its job. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
45 |
0 | 46 The most of the complexity is in bringing the code into shape expected |
47 by the omp expanders: | |
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 | |
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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
64 /* |
0 | 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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
68 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
69 |
0 | 70 parloop |
71 { | |
72 int sum=1; | |
73 | |
74 for (i = 0; i < N; i++) | |
75 { | |
76 x[i] = i + 3; | |
77 sum+=x[i]; | |
78 } | |
79 } | |
80 | |
81 gimple-like code: | |
82 header_bb: | |
83 | |
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 | |
93 | |
94 exit_bb: | |
95 | |
96 # sum_21 = PHI <sum_11(4)> | |
97 printf (&"%d"[0], sum_21); | |
98 | |
99 | |
100 after reduction transformation (only relevant parts): | |
101 | |
102 parloop | |
103 { | |
104 | |
105 .... | |
106 | |
107 | |
108 # Storing the initial value given by the user. # | |
109 | |
110 .paral_data_store.32.sum.27 = 1; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
111 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
112 #pragma omp parallel num_threads(4) |
0 | 113 |
114 #pragma omp for schedule(static) | |
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 | |
122 sum.27_11 = D.1827_8 + sum.27_29; | |
123 | |
124 GIMPLE_OMP_CONTINUE | |
125 | |
126 # Adding this reduction phi is done at create_phi_for_local_result() # | |
127 # sum.27_56 = PHI <sum.27_11, 0> | |
128 GIMPLE_OMP_RETURN | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
129 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
130 # Creating the atomic operation is done at |
0 | 131 create_call_for_reduction_1() # |
132 | |
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); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
137 |
0 | 138 GIMPLE_OMP_RETURN |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
139 |
0 | 140 # collecting the result after the join of the threads is done at |
141 create_loads_for_reductions(). | |
142 The value computed by the threads is loaded from the | |
143 shared struct. # | |
144 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
145 |
0 | 146 .paral_data_load.33_52 = &.paral_data_store.32; |
147 sum_37 = .paral_data_load.33_52->sum.27; | |
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 | |
156 } | |
157 | |
158 */ | |
159 | |
160 /* Minimal number of iterations of a loop that should be executed in each | |
161 thread. */ | |
162 #define MIN_PER_THREAD 100 | |
163 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
164 /* Element of the hashtable, representing a |
0 | 165 reduction in the current loop. */ |
166 struct reduction_info | |
167 { | |
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. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
171 gimple keep_res; /* The PHI_RESULT of this phi is the resulting value |
0 | 172 of the reduction variable when existing the loop. */ |
173 tree initial_value; /* The initial value of the reduction var before entering the loop. */ | |
174 tree field; /* the name of the field in the parloop data structure intended for reduction. */ | |
175 tree init; /* reduction initialization value. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
176 gimple new_phi; /* (helper field) Newly created phi node whose result |
0 | 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 * | |
202 reduction_phi (htab_t reduction_list, gimple phi) | |
203 { | |
204 struct reduction_info tmpred, *red; | |
205 | |
206 if (htab_elements (reduction_list) == 0) | |
207 return NULL; | |
208 | |
209 tmpred.reduc_phi = phi; | |
210 red = (struct reduction_info *) htab_find (reduction_list, &tmpred); | |
211 | |
212 return red; | |
213 } | |
214 | |
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 { | |
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; | |
232 | |
233 return a->version == b->version; | |
234 } | |
235 | |
236 static hashval_t | |
237 name_to_copy_elt_hash (const void *aa) | |
238 { | |
239 const struct name_to_copy_elt *a = (const struct name_to_copy_elt *) aa; | |
240 | |
241 return (hashval_t) a->version; | |
242 } | |
243 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
244 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
245 /* Data dependency analysis. Returns true if the iterations of LOOP |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
246 are independent on each other (that is, if we can execute them |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
247 in parallel). */ |
0 | 248 |
249 static bool | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
250 loop_parallel_p (struct loop *loop) |
0 | 251 { |
252 VEC (ddr_p, heap) * dependence_relations; | |
253 VEC (data_reference_p, heap) *datarefs; | |
254 lambda_trans_matrix trans; | |
255 bool ret = false; | |
256 | |
257 if (dump_file && (dump_flags & TDF_DETAILS)) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
258 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
259 fprintf (dump_file, "Considering loop %d\n", loop->num); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
260 if (!loop->inner) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
261 fprintf (dump_file, "loop is innermost\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
262 else |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
263 fprintf (dump_file, "loop NOT innermost\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
264 } |
0 | 265 |
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)) | |
285 fprintf (dump_file, | |
286 " FAILED: data dependencies exist across iterations\n"); | |
287 | |
288 free_dependence_relations (dependence_relations); | |
289 free_data_refs (datarefs); | |
290 | |
291 return ret; | |
292 } | |
293 | |
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 | |
314 /* Assigns the address of OBJ in TYPE to an ssa name, and returns this name. | |
315 The assignment statement is placed on edge ENTRY. DECL_ADDRESS maps decls | |
316 to their addresses that can be reused. The address of OBJ is known to | |
317 be invariant in the whole function. */ | |
318 | |
319 static tree | |
320 take_address_of (tree obj, tree type, edge entry, htab_t decl_address) | |
321 { | |
322 int uid; | |
323 void **dslot; | |
324 struct int_tree_map ielt, *nielt; | |
325 tree *var_p, name, bvar, addr; | |
326 gimple stmt; | |
327 gimple_seq stmts; | |
328 | |
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 | |
338 ielt.uid = uid; | |
339 dslot = htab_find_slot_with_hash (decl_address, &ielt, uid, INSERT); | |
340 if (!*dslot) | |
341 { | |
342 addr = build_addr (*var_p, current_function_decl); | |
343 bvar = create_tmp_var (TREE_TYPE (addr), get_name (*var_p)); | |
344 add_referenced_var (bvar); | |
345 stmt = gimple_build_assign (bvar, addr); | |
346 name = make_ssa_name (bvar, stmt); | |
347 gimple_assign_set_lhs (stmt, name); | |
348 gsi_insert_on_edge_immediate (entry, stmt); | |
349 | |
350 nielt = XNEW (struct int_tree_map); | |
351 nielt->uid = uid; | |
352 nielt->to = name; | |
353 *dslot = nielt; | |
354 } | |
355 else | |
356 name = ((struct int_tree_map *) *dslot)->to; | |
357 | |
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), | |
362 &stmts, true, NULL_TREE); | |
363 if (!gimple_seq_empty_p (stmts)) | |
364 gsi_insert_seq_on_edge_immediate (entry, stmts); | |
365 } | |
366 | |
367 if (TREE_TYPE (name) != type) | |
368 { | |
369 name = force_gimple_operand (fold_convert (type, name), &stmts, true, | |
370 NULL_TREE); | |
371 if (!gimple_seq_empty_p (stmts)) | |
372 gsi_insert_seq_on_edge_immediate (entry, stmts); | |
373 } | |
374 | |
375 return name; | |
376 } | |
377 | |
378 /* Callback for htab_traverse. Create the initialization statement | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
379 for reduction described in SLOT, and place it at the preheader of |
0 | 380 the loop described in DATA. */ |
381 | |
382 static int | |
383 initialize_reductions (void **slot, void *data) | |
384 { | |
385 tree init, c; | |
386 tree bvar, type, arg; | |
387 edge e; | |
388 | |
389 struct reduction_info *const reduc = (struct reduction_info *) *slot; | |
390 struct loop *loop = (struct loop *) data; | |
391 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
392 /* Create initialization in preheader: |
0 | 393 reduction_variable = initialization value of reduction. */ |
394 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
395 /* In the phi node at the header, replace the argument coming |
0 | 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 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
403 c = build_omp_clause (gimple_location (reduc->reduc_stmt), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
404 OMP_CLAUSE_REDUCTION); |
0 | 405 OMP_CLAUSE_REDUCTION_CODE (c) = reduc->reduction_code; |
406 OMP_CLAUSE_DECL (c) = SSA_NAME_VAR (gimple_assign_lhs (reduc->reduc_stmt)); | |
407 | |
408 init = omp_reduction_init (c, TREE_TYPE (bvar)); | |
409 reduc->init = init; | |
410 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
411 /* Replace the argument representing the initialization value |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
412 with the initialization value for the reduction (neutral |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
413 element for the particular operation, e.g. 0 for PLUS_EXPR, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
414 1 for MULT_EXPR, etc). |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
415 Keep the old value in a new variable "reduction_initial", |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
416 that will be taken in consideration after the parallel |
0 | 417 computing is done. */ |
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. */ | |
422 | |
423 SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE | |
424 (reduc->reduc_phi, loop_preheader_edge (loop)), init); | |
425 reduc->initial_value = arg; | |
426 return 1; | |
427 } | |
428 | |
429 struct elv_data | |
430 { | |
431 struct walk_stmt_info info; | |
432 edge entry; | |
433 htab_t decl_address; | |
434 bool changed; | |
435 }; | |
436 | |
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. */ | |
442 | |
443 static tree | |
444 eliminate_local_variables_1 (tree *tp, int *walk_subtrees, void *data) | |
445 { | |
446 struct elv_data *const dta = (struct elv_data *) data; | |
447 tree t = *tp, var, addr, addr_type, type, obj; | |
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); | |
458 addr = take_address_of (t, addr_type, dta->entry, dta->decl_address); | |
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 { | |
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)) | |
479 return NULL_TREE; | |
480 | |
481 *walk_subtrees = 0; | |
482 obj = TREE_OPERAND (t, 0); | |
483 var = get_base_address (obj); | |
484 if (!var || !SSA_VAR_P (var) || DECL_EXTERNAL (var)) | |
485 return NULL_TREE; | |
486 | |
487 addr_type = TREE_TYPE (t); | |
488 addr = take_address_of (obj, addr_type, dta->entry, dta->decl_address); | |
489 *tp = addr; | |
490 | |
491 dta->changed = true; | |
492 return NULL_TREE; | |
493 } | |
494 | |
495 if (!EXPR_P (t)) | |
496 *walk_subtrees = 0; | |
497 | |
498 return NULL_TREE; | |
499 } | |
500 | |
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. */ | |
505 | |
506 static void | |
507 eliminate_local_variables_stmt (edge entry, gimple stmt, | |
508 htab_t decl_address) | |
509 { | |
510 struct elv_data dta; | |
511 | |
512 memset (&dta.info, '\0', sizeof (dta.info)); | |
513 dta.entry = entry; | |
514 dta.decl_address = decl_address; | |
515 dta.changed = false; | |
516 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
517 if (gimple_debug_bind_p (stmt)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
518 walk_tree (gimple_debug_bind_get_value_ptr (stmt), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
519 eliminate_local_variables_1, &dta.info, NULL); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
520 else |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
521 walk_gimple_op (stmt, eliminate_local_variables_1, &dta.info); |
0 | 522 |
523 if (dta.changed) | |
524 update_stmt (stmt); | |
525 } | |
526 | |
527 /* Eliminates the references to local variables from the single entry | |
528 single exit region between the ENTRY and EXIT edges. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
529 |
0 | 530 This includes: |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
531 1) Taking address of a local variable -- these are moved out of the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
532 region (and temporary variable is created to hold the address if |
0 | 533 necessary). |
534 | |
535 2) Dereferencing a local variable -- these are replaced with indirect | |
536 references. */ | |
537 | |
538 static void | |
539 eliminate_local_variables (edge entry, edge exit) | |
540 { | |
541 basic_block bb; | |
542 VEC (basic_block, heap) *body = VEC_alloc (basic_block, heap, 3); | |
543 unsigned i; | |
544 gimple_stmt_iterator gsi; | |
545 htab_t decl_address = htab_create (10, int_tree_map_hash, int_tree_map_eq, | |
546 free); | |
547 basic_block entry_bb = entry->src; | |
548 basic_block exit_bb = exit->dest; | |
549 | |
550 gather_blocks_in_sese_region (entry_bb, exit_bb, &body); | |
551 | |
552 for (i = 0; VEC_iterate (basic_block, body, i, bb); i++) | |
553 if (bb != entry_bb && bb != exit_bb) | |
554 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
555 eliminate_local_variables_stmt (entry, gsi_stmt (gsi), | |
556 decl_address); | |
557 | |
558 htab_delete (decl_address); | |
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; | |
571 | |
572 if (is_gimple_min_invariant (expr)) | |
573 return true; | |
574 | |
575 if (TREE_CODE (expr) == SSA_NAME) | |
576 { | |
577 def_bb = gimple_bb (SSA_NAME_DEF_STMT (expr)); | |
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 | |
586 return false; | |
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. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
592 |
0 | 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 | |
597 separate_decls_in_region_name (tree name, | |
598 htab_t name_copies, htab_t decl_copies, | |
599 bool copy_name_p) | |
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)); | |
624 DECL_GIMPLE_REG_P (var_copy) = DECL_GIMPLE_REG_P (var); | |
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 | |
632 it again. */ | |
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 { | |
647 copy = duplicate_ssa_name (name, NULL); | |
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 | |
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. */ | |
670 | |
671 static void | |
672 separate_decls_in_region_stmt (edge entry, edge exit, gimple stmt, | |
673 htab_t name_copies, htab_t decl_copies) | |
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) | |
684 { | |
685 name = DEF_FROM_PTR (def); | |
686 gcc_assert (TREE_CODE (name) == SSA_NAME); | |
687 copy = separate_decls_in_region_name (name, name_copies, decl_copies, | |
688 false); | |
689 gcc_assert (copy == name); | |
690 } | |
691 | |
692 FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) | |
693 { | |
694 name = USE_FROM_PTR (use); | |
695 if (TREE_CODE (name) != SSA_NAME) | |
696 continue; | |
697 | |
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); | |
701 SET_USE (use, copy); | |
702 } | |
703 } | |
704 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
705 /* Finds the ssa names used in STMT that are defined outside the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
706 region between ENTRY and EXIT and replaces such ssa names with |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
707 their duplicates. The duplicates are stored to NAME_COPIES. Base |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
708 decls of all ssa names used in STMT (including those defined in |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
709 LOOP) are replaced with the new temporary variables; the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
710 replacement decls are stored in DECL_COPIES. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
711 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
712 static bool |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
713 separate_decls_in_region_debug_bind (gimple stmt, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
714 htab_t name_copies, htab_t decl_copies) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
715 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
716 use_operand_p use; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
717 ssa_op_iter oi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
718 tree var, name; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
719 struct int_tree_map ielt; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
720 struct name_to_copy_elt elt; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
721 void **slot, **dslot; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
722 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
723 var = gimple_debug_bind_get_var (stmt); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
724 if (TREE_CODE (var) == DEBUG_EXPR_DECL) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
725 return true; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
726 gcc_assert (DECL_P (var) && SSA_VAR_P (var)); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
727 ielt.uid = DECL_UID (var); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
728 dslot = htab_find_slot_with_hash (decl_copies, &ielt, ielt.uid, NO_INSERT); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
729 if (!dslot) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
730 return true; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
731 gimple_debug_bind_set_var (stmt, ((struct int_tree_map *) *dslot)->to); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
732 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
733 FOR_EACH_PHI_OR_STMT_USE (use, stmt, oi, SSA_OP_USE) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
734 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
735 name = USE_FROM_PTR (use); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
736 if (TREE_CODE (name) != SSA_NAME) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
737 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
738 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
739 elt.version = SSA_NAME_VERSION (name); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
740 slot = htab_find_slot_with_hash (name_copies, &elt, elt.version, NO_INSERT); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
741 if (!slot) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
742 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
743 gimple_debug_bind_reset_value (stmt); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
744 update_stmt (stmt); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
745 break; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
746 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
747 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
748 SET_USE (use, ((struct name_to_copy_elt *) *slot)->new_name); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
749 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
750 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
751 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
752 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
753 |
0 | 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) | |
759 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
760 |
0 | 761 struct reduction_info *const red = (struct reduction_info *) *slot; |
762 tree const type = (tree) data; | |
763 tree var = SSA_NAME_VAR (gimple_assign_lhs (red->reduc_stmt)); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
764 tree field = build_decl (gimple_location (red->reduc_stmt), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
765 FIELD_DECL, DECL_NAME (var), TREE_TYPE (var)); |
0 | 766 |
767 insert_field_into_struct (type, field); | |
768 | |
769 red->field = field; | |
770 | |
771 return 1; | |
772 } | |
773 | |
774 /* Callback for htab_traverse. Adds a field corresponding to a ssa name | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
775 described in SLOT. The type is passed in DATA. */ |
0 | 776 |
777 static int | |
778 add_field_for_name (void **slot, void *data) | |
779 { | |
780 struct name_to_copy_elt *const elt = (struct name_to_copy_elt *) *slot; | |
781 tree type = (tree) data; | |
782 tree name = ssa_name (elt->version); | |
783 tree var = SSA_NAME_VAR (name); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
784 tree field = build_decl (DECL_SOURCE_LOCATION (var), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
785 FIELD_DECL, DECL_NAME (var), TREE_TYPE (var)); |
0 | 786 |
787 insert_field_into_struct (type, field); | |
788 elt->field = field; | |
789 | |
790 return 1; | |
791 } | |
792 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
793 /* Callback for htab_traverse. A local result is the intermediate result |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
794 computed by a single |
0 | 795 thread, or the initial value in case no iteration was executed. |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
796 This function creates a phi node reflecting these values. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
797 The phi's result will be stored in NEW_PHI field of the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
798 reduction's data structure. */ |
0 | 799 |
800 static int | |
801 create_phi_for_local_result (void **slot, void *data) | |
802 { | |
803 struct reduction_info *const reduc = (struct reduction_info *) *slot; | |
804 const struct loop *const loop = (const struct loop *) data; | |
805 edge e; | |
806 gimple new_phi; | |
807 basic_block store_bb; | |
808 tree local_res; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
809 source_location locus; |
0 | 810 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
811 /* STORE_BB is the block where the phi |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
812 should be stored. It is the destination of the loop exit. |
0 | 813 (Find the fallthru edge from GIMPLE_OMP_CONTINUE). */ |
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), | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
818 and another coming from a block preceding the loop, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
819 when no iterations |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
820 are executed (the initial value should be taken). */ |
0 | 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); | |
825 local_res | |
826 = make_ssa_name (SSA_NAME_VAR (gimple_assign_lhs (reduc->reduc_stmt)), | |
827 NULL); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
828 locus = gimple_location (reduc->reduc_stmt); |
0 | 829 new_phi = create_phi_node (local_res, store_bb); |
830 SSA_NAME_DEF_STMT (local_res) = new_phi; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
831 add_phi_arg (new_phi, reduc->init, e, locus); |
0 | 832 add_phi_arg (new_phi, gimple_assign_lhs (reduc->reduc_stmt), |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
833 FALLTHRU_EDGE (loop->latch), locus); |
0 | 834 reduc->new_phi = new_phi; |
835 | |
836 return 1; | |
837 } | |
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 | |
848 /* Callback for htab_traverse. Create an atomic instruction for the | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
849 reduction described in SLOT. |
0 | 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 { | |
856 struct reduction_info *const reduc = (struct reduction_info *) *slot; | |
857 struct clsn_data *const clsn_data = (struct clsn_data *) data; | |
858 gimple_stmt_iterator gsi; | |
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; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
865 tree t, addr, ref, x; |
0 | 866 tree tmp_load, name; |
867 gimple load; | |
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 | |
872 addr = build_addr (t, current_function_decl); | |
873 | |
874 /* Create phi node. */ | |
875 bb = clsn_data->load_bb; | |
876 | |
877 e = split_block (bb, t); | |
878 new_bb = e->dest; | |
879 | |
880 tmp_load = create_tmp_var (TREE_TYPE (TREE_TYPE (addr)), NULL); | |
881 add_referenced_var (tmp_load); | |
882 tmp_load = make_ssa_name (tmp_load, NULL); | |
883 load = gimple_build_omp_atomic_load (tmp_load, addr); | |
884 SSA_NAME_DEF_STMT (tmp_load) = load; | |
885 gsi = gsi_start_bb (new_bb); | |
886 gsi_insert_after (&gsi, load, GSI_NEW_STMT); | |
887 | |
888 e = split_block (new_bb, load); | |
889 new_bb = e->dest; | |
890 gsi = gsi_start_bb (new_bb); | |
891 ref = tmp_load; | |
892 x = fold_build2 (reduc->reduction_code, | |
893 TREE_TYPE (PHI_RESULT (reduc->new_phi)), ref, | |
894 PHI_RESULT (reduc->new_phi)); | |
895 | |
896 name = force_gimple_operand_gsi (&gsi, x, true, NULL_TREE, true, | |
897 GSI_CONTINUE_LINKING); | |
898 | |
899 gsi_insert_after (&gsi, gimple_build_omp_atomic_store (name), GSI_NEW_STMT); | |
900 return 1; | |
901 } | |
902 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
903 /* Create the atomic operation at the join point of the threads. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
904 REDUCTION_LIST describes the reductions in the LOOP. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
905 LD_ST_DATA describes the shared data structure where |
0 | 906 shared data is stored in and loaded from. */ |
907 static void | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
908 create_call_for_reduction (struct loop *loop, htab_t reduction_list, |
0 | 909 struct clsn_data *ld_st_data) |
910 { | |
911 htab_traverse (reduction_list, create_phi_for_local_result, loop); | |
912 /* Find the fallthru edge from GIMPLE_OMP_CONTINUE. */ | |
913 ld_st_data->load_bb = FALLTHRU_EDGE (loop->latch)->dest; | |
914 htab_traverse (reduction_list, create_call_for_reduction_1, ld_st_data); | |
915 } | |
916 | |
917 /* Callback for htab_traverse. Loads the final reduction value at the | |
918 join point of all threads, and inserts it in the right place. */ | |
919 | |
920 static int | |
921 create_loads_for_reductions (void **slot, void *data) | |
922 { | |
923 struct reduction_info *const red = (struct reduction_info *) *slot; | |
924 struct clsn_data *const clsn_data = (struct clsn_data *) data; | |
925 gimple stmt; | |
926 gimple_stmt_iterator gsi; | |
927 tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); | |
928 tree struct_type = TREE_TYPE (TREE_TYPE (clsn_data->load)); | |
929 tree load_struct; | |
930 tree name; | |
931 tree x; | |
932 | |
933 gsi = gsi_after_labels (clsn_data->load_bb); | |
934 load_struct = fold_build1 (INDIRECT_REF, struct_type, clsn_data->load); | |
935 load_struct = build3 (COMPONENT_REF, type, load_struct, red->field, | |
936 NULL_TREE); | |
937 | |
938 x = load_struct; | |
939 name = PHI_RESULT (red->keep_res); | |
940 stmt = gimple_build_assign (name, x); | |
941 SSA_NAME_DEF_STMT (name) = stmt; | |
942 | |
943 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
944 | |
945 for (gsi = gsi_start_phis (gimple_bb (red->keep_res)); | |
946 !gsi_end_p (gsi); gsi_next (&gsi)) | |
947 if (gsi_stmt (gsi) == red->keep_res) | |
948 { | |
949 remove_phi_node (&gsi, false); | |
950 return 1; | |
951 } | |
952 gcc_unreachable (); | |
953 } | |
954 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
955 /* Load the reduction result that was stored in LD_ST_DATA. |
0 | 956 REDUCTION_LIST describes the list of reductions that the |
957 loads should be generated for. */ | |
958 static void | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
959 create_final_loads_for_reduction (htab_t reduction_list, |
0 | 960 struct clsn_data *ld_st_data) |
961 { | |
962 gimple_stmt_iterator gsi; | |
963 tree t; | |
964 gimple stmt; | |
965 | |
966 gsi = gsi_after_labels (ld_st_data->load_bb); | |
967 t = build_fold_addr_expr (ld_st_data->store); | |
968 stmt = gimple_build_assign (ld_st_data->load, t); | |
969 | |
970 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); | |
971 SSA_NAME_DEF_STMT (ld_st_data->load) = stmt; | |
972 | |
973 htab_traverse (reduction_list, create_loads_for_reductions, ld_st_data); | |
974 | |
975 } | |
976 | |
977 /* Callback for htab_traverse. Store the neutral value for the | |
978 particular reduction's operation, e.g. 0 for PLUS_EXPR, | |
979 1 for MULT_EXPR, etc. into the reduction field. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
980 The reduction is specified in SLOT. The store information is |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
981 passed in DATA. */ |
0 | 982 |
983 static int | |
984 create_stores_for_reduction (void **slot, void *data) | |
985 { | |
986 struct reduction_info *const red = (struct reduction_info *) *slot; | |
987 struct clsn_data *const clsn_data = (struct clsn_data *) data; | |
988 tree t; | |
989 gimple stmt; | |
990 gimple_stmt_iterator gsi; | |
991 tree type = TREE_TYPE (gimple_assign_lhs (red->reduc_stmt)); | |
992 | |
993 gsi = gsi_last_bb (clsn_data->store_bb); | |
994 t = build3 (COMPONENT_REF, type, clsn_data->store, red->field, NULL_TREE); | |
995 stmt = gimple_build_assign (t, red->initial_value); | |
996 mark_virtual_ops_for_renaming (stmt); | |
997 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
998 | |
999 return 1; | |
1000 } | |
1001 | |
1002 /* Callback for htab_traverse. Creates loads to a field of LOAD in LOAD_BB and | |
1003 store to a field of STORE in STORE_BB for the ssa name and its duplicate | |
1004 specified in SLOT. */ | |
1005 | |
1006 static int | |
1007 create_loads_and_stores_for_name (void **slot, void *data) | |
1008 { | |
1009 struct name_to_copy_elt *const elt = (struct name_to_copy_elt *) *slot; | |
1010 struct clsn_data *const clsn_data = (struct clsn_data *) data; | |
1011 tree t; | |
1012 gimple stmt; | |
1013 gimple_stmt_iterator gsi; | |
1014 tree type = TREE_TYPE (elt->new_name); | |
1015 tree struct_type = TREE_TYPE (TREE_TYPE (clsn_data->load)); | |
1016 tree load_struct; | |
1017 | |
1018 gsi = gsi_last_bb (clsn_data->store_bb); | |
1019 t = build3 (COMPONENT_REF, type, clsn_data->store, elt->field, NULL_TREE); | |
1020 stmt = gimple_build_assign (t, ssa_name (elt->version)); | |
1021 mark_virtual_ops_for_renaming (stmt); | |
1022 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1023 | |
1024 gsi = gsi_last_bb (clsn_data->load_bb); | |
1025 load_struct = fold_build1 (INDIRECT_REF, struct_type, clsn_data->load); | |
1026 t = build3 (COMPONENT_REF, type, load_struct, elt->field, NULL_TREE); | |
1027 stmt = gimple_build_assign (elt->new_name, t); | |
1028 SSA_NAME_DEF_STMT (elt->new_name) = stmt; | |
1029 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1030 | |
1031 return 1; | |
1032 } | |
1033 | |
1034 /* Moves all the variables used in LOOP and defined outside of it (including | |
1035 the initial values of loop phi nodes, and *PER_THREAD if it is a ssa | |
1036 name) to a structure created for this purpose. The code | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1037 |
0 | 1038 while (1) |
1039 { | |
1040 use (a); | |
1041 use (b); | |
1042 } | |
1043 | |
1044 is transformed this way: | |
1045 | |
1046 bb0: | |
1047 old.a = a; | |
1048 old.b = b; | |
1049 | |
1050 bb1: | |
1051 a' = new->a; | |
1052 b' = new->b; | |
1053 while (1) | |
1054 { | |
1055 use (a'); | |
1056 use (b'); | |
1057 } | |
1058 | |
1059 `old' is stored to *ARG_STRUCT and `new' is stored to NEW_ARG_STRUCT. The | |
1060 pointer `new' is intentionally not initialized (the loop will be split to a | |
1061 separate function later, and `new' will be initialized from its arguments). | |
1062 LD_ST_DATA holds information about the shared data structure used to pass | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1063 information among the threads. It is initialized here, and |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1064 gen_parallel_loop will pass it to create_call_for_reduction that |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1065 needs this information. REDUCTION_LIST describes the reductions |
0 | 1066 in LOOP. */ |
1067 | |
1068 static void | |
1069 separate_decls_in_region (edge entry, edge exit, htab_t reduction_list, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1070 tree *arg_struct, tree *new_arg_struct, |
0 | 1071 struct clsn_data *ld_st_data) |
1072 | |
1073 { | |
1074 basic_block bb1 = split_edge (entry); | |
1075 basic_block bb0 = single_pred (bb1); | |
1076 htab_t name_copies = htab_create (10, name_to_copy_elt_hash, | |
1077 name_to_copy_elt_eq, free); | |
1078 htab_t decl_copies = htab_create (10, int_tree_map_hash, int_tree_map_eq, | |
1079 free); | |
1080 unsigned i; | |
1081 tree type, type_name, nvar; | |
1082 gimple_stmt_iterator gsi; | |
1083 struct clsn_data clsn_data; | |
1084 VEC (basic_block, heap) *body = VEC_alloc (basic_block, heap, 3); | |
1085 basic_block bb; | |
1086 basic_block entry_bb = bb1; | |
1087 basic_block exit_bb = exit->dest; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1088 bool has_debug_stmt = false; |
0 | 1089 |
1090 entry = single_succ_edge (entry_bb); | |
1091 gather_blocks_in_sese_region (entry_bb, exit_bb, &body); | |
1092 | |
1093 for (i = 0; VEC_iterate (basic_block, body, i, bb); i++) | |
1094 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1095 if (bb != entry_bb && bb != exit_bb) |
0 | 1096 { |
1097 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1098 separate_decls_in_region_stmt (entry, exit, gsi_stmt (gsi), | |
1099 name_copies, decl_copies); | |
1100 | |
1101 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1102 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1103 gimple stmt = gsi_stmt (gsi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1104 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1105 if (is_gimple_debug (stmt)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1106 has_debug_stmt = true; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1107 else |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1108 separate_decls_in_region_stmt (entry, exit, stmt, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1109 name_copies, decl_copies); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1110 } |
0 | 1111 } |
1112 } | |
1113 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1114 /* Now process debug bind stmts. We must not create decls while |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1115 processing debug stmts, so we defer their processing so as to |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1116 make sure we will have debug info for as many variables as |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1117 possible (all of those that were dealt with in the loop above), |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1118 and discard those for which we know there's nothing we can |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1119 do. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1120 if (has_debug_stmt) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1121 for (i = 0; VEC_iterate (basic_block, body, i, bb); i++) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1122 if (bb != entry_bb && bb != exit_bb) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1123 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1124 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1125 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1126 gimple stmt = gsi_stmt (gsi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1127 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1128 if (gimple_debug_bind_p (stmt)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1129 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1130 if (separate_decls_in_region_debug_bind (stmt, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1131 name_copies, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1132 decl_copies)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1133 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1134 gsi_remove (&gsi, true); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1135 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1136 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1137 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1138 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1139 gsi_next (&gsi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1140 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1141 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1142 |
0 | 1143 VEC_free (basic_block, heap, body); |
1144 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1145 if (htab_elements (name_copies) == 0 && htab_elements (reduction_list) == 0) |
0 | 1146 { |
1147 /* It may happen that there is nothing to copy (if there are only | |
1148 loop carried and external variables in the loop). */ | |
1149 *arg_struct = NULL; | |
1150 *new_arg_struct = NULL; | |
1151 } | |
1152 else | |
1153 { | |
1154 /* Create the type for the structure to store the ssa names to. */ | |
1155 type = lang_hooks.types.make_type (RECORD_TYPE); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1156 type_name = build_decl (BUILTINS_LOCATION, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1157 TYPE_DECL, create_tmp_var_name (".paral_data"), |
0 | 1158 type); |
1159 TYPE_NAME (type) = type_name; | |
1160 | |
1161 htab_traverse (name_copies, add_field_for_name, type); | |
1162 if (reduction_list && htab_elements (reduction_list) > 0) | |
1163 { | |
1164 /* Create the fields for reductions. */ | |
1165 htab_traverse (reduction_list, add_field_for_reduction, | |
1166 type); | |
1167 } | |
1168 layout_type (type); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1169 |
0 | 1170 /* Create the loads and stores. */ |
1171 *arg_struct = create_tmp_var (type, ".paral_data_store"); | |
1172 add_referenced_var (*arg_struct); | |
1173 nvar = create_tmp_var (build_pointer_type (type), ".paral_data_load"); | |
1174 add_referenced_var (nvar); | |
1175 *new_arg_struct = make_ssa_name (nvar, NULL); | |
1176 | |
1177 ld_st_data->store = *arg_struct; | |
1178 ld_st_data->load = *new_arg_struct; | |
1179 ld_st_data->store_bb = bb0; | |
1180 ld_st_data->load_bb = bb1; | |
1181 | |
1182 htab_traverse (name_copies, create_loads_and_stores_for_name, | |
1183 ld_st_data); | |
1184 | |
1185 /* Load the calculation from memory (after the join of the threads). */ | |
1186 | |
1187 if (reduction_list && htab_elements (reduction_list) > 0) | |
1188 { | |
1189 htab_traverse (reduction_list, create_stores_for_reduction, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1190 ld_st_data); |
0 | 1191 clsn_data.load = make_ssa_name (nvar, NULL); |
1192 clsn_data.load_bb = exit->dest; | |
1193 clsn_data.store = ld_st_data->store; | |
1194 create_final_loads_for_reduction (reduction_list, &clsn_data); | |
1195 } | |
1196 } | |
1197 | |
1198 htab_delete (decl_copies); | |
1199 htab_delete (name_copies); | |
1200 } | |
1201 | |
1202 /* Bitmap containing uids of functions created by parallelization. We cannot | |
1203 allocate it from the default obstack, as it must live across compilation | |
1204 of several functions; we make it gc allocated instead. */ | |
1205 | |
1206 static GTY(()) bitmap parallelized_functions; | |
1207 | |
1208 /* Returns true if FN was created by create_loop_fn. */ | |
1209 | |
1210 static bool | |
1211 parallelized_function_p (tree fn) | |
1212 { | |
1213 if (!parallelized_functions || !DECL_ARTIFICIAL (fn)) | |
1214 return false; | |
1215 | |
1216 return bitmap_bit_p (parallelized_functions, DECL_UID (fn)); | |
1217 } | |
1218 | |
1219 /* Creates and returns an empty function that will receive the body of | |
1220 a parallelized loop. */ | |
1221 | |
1222 static tree | |
1223 create_loop_fn (void) | |
1224 { | |
1225 char buf[100]; | |
1226 char *tname; | |
1227 tree decl, type, name, t; | |
1228 struct function *act_cfun = cfun; | |
1229 static unsigned loopfn_num; | |
1230 | |
1231 snprintf (buf, 100, "%s.$loopfn", current_function_name ()); | |
1232 ASM_FORMAT_PRIVATE_NAME (tname, buf, loopfn_num++); | |
1233 clean_symbol_name (tname); | |
1234 name = get_identifier (tname); | |
1235 type = build_function_type_list (void_type_node, ptr_type_node, NULL_TREE); | |
1236 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1237 decl = build_decl (BUILTINS_LOCATION, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1238 FUNCTION_DECL, name, type); |
0 | 1239 if (!parallelized_functions) |
1240 parallelized_functions = BITMAP_GGC_ALLOC (); | |
1241 bitmap_set_bit (parallelized_functions, DECL_UID (decl)); | |
1242 | |
1243 TREE_STATIC (decl) = 1; | |
1244 TREE_USED (decl) = 1; | |
1245 DECL_ARTIFICIAL (decl) = 1; | |
1246 DECL_IGNORED_P (decl) = 0; | |
1247 TREE_PUBLIC (decl) = 0; | |
1248 DECL_UNINLINABLE (decl) = 1; | |
1249 DECL_EXTERNAL (decl) = 0; | |
1250 DECL_CONTEXT (decl) = NULL_TREE; | |
1251 DECL_INITIAL (decl) = make_node (BLOCK); | |
1252 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1253 t = build_decl (BUILTINS_LOCATION, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1254 RESULT_DECL, NULL_TREE, void_type_node); |
0 | 1255 DECL_ARTIFICIAL (t) = 1; |
1256 DECL_IGNORED_P (t) = 1; | |
1257 DECL_RESULT (decl) = t; | |
1258 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1259 t = build_decl (BUILTINS_LOCATION, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1260 PARM_DECL, get_identifier (".paral_data_param"), |
0 | 1261 ptr_type_node); |
1262 DECL_ARTIFICIAL (t) = 1; | |
1263 DECL_ARG_TYPE (t) = ptr_type_node; | |
1264 DECL_CONTEXT (t) = decl; | |
1265 TREE_USED (t) = 1; | |
1266 DECL_ARGUMENTS (decl) = t; | |
1267 | |
1268 allocate_struct_function (decl, false); | |
1269 | |
1270 /* The call to allocate_struct_function clobbers CFUN, so we need to restore | |
1271 it. */ | |
1272 set_cfun (act_cfun); | |
1273 | |
1274 return decl; | |
1275 } | |
1276 | |
1277 /* Moves the exit condition of LOOP to the beginning of its header, and | |
1278 duplicates the part of the last iteration that gets disabled to the | |
1279 exit of the loop. NIT is the number of iterations of the loop | |
1280 (used to initialize the variables in the duplicated part). | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1281 |
0 | 1282 TODO: the common case is that latch of the loop is empty and immediately |
1283 follows the loop exit. In this case, it would be better not to copy the | |
1284 body of the loop, but only move the entry of the loop directly before the | |
1285 exit check and increase the number of iterations of the loop by one. | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1286 This may need some additional preconditioning in case NIT = ~0. |
0 | 1287 REDUCTION_LIST describes the reductions in LOOP. */ |
1288 | |
1289 static void | |
1290 transform_to_exit_first_loop (struct loop *loop, htab_t reduction_list, tree nit) | |
1291 { | |
1292 basic_block *bbs, *nbbs, ex_bb, orig_header; | |
1293 unsigned n; | |
1294 bool ok; | |
1295 edge exit = single_dom_exit (loop), hpred; | |
1296 tree control, control_name, res, t; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1297 gimple phi, nphi, cond_stmt, stmt, cond_nit; |
0 | 1298 gimple_stmt_iterator gsi; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1299 tree nit_1; |
0 | 1300 |
1301 split_block_after_labels (loop->header); | |
1302 orig_header = single_succ (loop->header); | |
1303 hpred = single_succ_edge (loop->header); | |
1304 | |
1305 cond_stmt = last_stmt (exit->src); | |
1306 control = gimple_cond_lhs (cond_stmt); | |
1307 gcc_assert (gimple_cond_rhs (cond_stmt) == nit); | |
1308 | |
1309 /* Make sure that we have phi nodes on exit for all loop header phis | |
1310 (create_parallel_loop requires that). */ | |
1311 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1312 { | |
1313 phi = gsi_stmt (gsi); | |
1314 res = PHI_RESULT (phi); | |
1315 t = make_ssa_name (SSA_NAME_VAR (res), phi); | |
1316 SET_PHI_RESULT (phi, t); | |
1317 nphi = create_phi_node (res, orig_header); | |
1318 SSA_NAME_DEF_STMT (res) = nphi; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1319 add_phi_arg (nphi, t, hpred, UNKNOWN_LOCATION); |
0 | 1320 |
1321 if (res == control) | |
1322 { | |
1323 gimple_cond_set_lhs (cond_stmt, t); | |
1324 update_stmt (cond_stmt); | |
1325 control = t; | |
1326 } | |
1327 } | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1328 bbs = get_loop_body_in_dom_order (loop); |
0 | 1329 |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1330 for (n = 0; bbs[n] != loop->latch; n++) |
0 | 1331 continue; |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1332 n--; |
0 | 1333 nbbs = XNEWVEC (basic_block, n); |
1334 ok = gimple_duplicate_sese_tail (single_succ_edge (loop->header), exit, | |
1335 bbs + 1, n, nbbs); | |
1336 gcc_assert (ok); | |
1337 free (bbs); | |
1338 ex_bb = nbbs[0]; | |
1339 free (nbbs); | |
1340 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1341 /* Other than reductions, the only gimple reg that should be copied |
0 | 1342 out of the loop is the control variable. */ |
1343 | |
1344 control_name = NULL_TREE; | |
1345 for (gsi = gsi_start_phis (ex_bb); !gsi_end_p (gsi); ) | |
1346 { | |
1347 phi = gsi_stmt (gsi); | |
1348 res = PHI_RESULT (phi); | |
1349 if (!is_gimple_reg (res)) | |
1350 { | |
1351 gsi_next (&gsi); | |
1352 continue; | |
1353 } | |
1354 | |
1355 /* Check if it is a part of reduction. If it is, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1356 keep the phi at the reduction's keep_res field. The |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1357 PHI_RESULT of this phi is the resulting value of the reduction |
0 | 1358 variable when exiting the loop. */ |
1359 | |
1360 exit = single_dom_exit (loop); | |
1361 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1362 if (htab_elements (reduction_list) > 0) |
0 | 1363 { |
1364 struct reduction_info *red; | |
1365 | |
1366 tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); | |
1367 red = reduction_phi (reduction_list, SSA_NAME_DEF_STMT (val)); | |
1368 if (red) | |
1369 { | |
1370 red->keep_res = phi; | |
1371 gsi_next (&gsi); | |
1372 continue; | |
1373 } | |
1374 } | |
1375 gcc_assert (control_name == NULL_TREE | |
1376 && SSA_NAME_VAR (res) == SSA_NAME_VAR (control)); | |
1377 control_name = res; | |
1378 remove_phi_node (&gsi, false); | |
1379 } | |
1380 gcc_assert (control_name != NULL_TREE); | |
1381 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1382 /* Initialize the control variable to number of iterations |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1383 according to the rhs of the exit condition. */ |
0 | 1384 gsi = gsi_after_labels (ex_bb); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1385 cond_nit = last_stmt (exit->src); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1386 nit_1 = gimple_cond_rhs (cond_nit); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1387 nit_1 = force_gimple_operand_gsi (&gsi, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1388 fold_convert (TREE_TYPE (control_name), nit_1), |
0 | 1389 false, NULL_TREE, false, GSI_SAME_STMT); |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1390 stmt = gimple_build_assign (control_name, nit_1); |
0 | 1391 gsi_insert_before (&gsi, stmt, GSI_NEW_STMT); |
1392 SSA_NAME_DEF_STMT (control_name) = stmt; | |
1393 } | |
1394 | |
1395 /* Create the parallel constructs for LOOP as described in gen_parallel_loop. | |
1396 LOOP_FN and DATA are the arguments of GIMPLE_OMP_PARALLEL. | |
1397 NEW_DATA is the variable that should be initialized from the argument | |
1398 of LOOP_FN. N_THREADS is the requested number of threads. Returns the | |
1399 basic block containing GIMPLE_OMP_PARALLEL tree. */ | |
1400 | |
1401 static basic_block | |
1402 create_parallel_loop (struct loop *loop, tree loop_fn, tree data, | |
1403 tree new_data, unsigned n_threads) | |
1404 { | |
1405 gimple_stmt_iterator gsi; | |
1406 basic_block bb, paral_bb, for_bb, ex_bb; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1407 tree t, param; |
0 | 1408 gimple stmt, for_stmt, phi, cond_stmt; |
1409 tree cvar, cvar_init, initvar, cvar_next, cvar_base, type; | |
1410 edge exit, nexit, guard, end, e; | |
1411 | |
1412 /* Prepare the GIMPLE_OMP_PARALLEL statement. */ | |
1413 bb = loop_preheader_edge (loop)->src; | |
1414 paral_bb = single_pred (bb); | |
1415 gsi = gsi_last_bb (paral_bb); | |
1416 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1417 t = build_omp_clause (BUILTINS_LOCATION, OMP_CLAUSE_NUM_THREADS); |
0 | 1418 OMP_CLAUSE_NUM_THREADS_EXPR (t) |
1419 = build_int_cst (integer_type_node, n_threads); | |
1420 stmt = gimple_build_omp_parallel (NULL, t, loop_fn, data); | |
1421 | |
1422 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1423 | |
1424 /* Initialize NEW_DATA. */ | |
1425 if (data) | |
1426 { | |
1427 gsi = gsi_after_labels (bb); | |
1428 | |
1429 param = make_ssa_name (DECL_ARGUMENTS (loop_fn), NULL); | |
1430 stmt = gimple_build_assign (param, build_fold_addr_expr (data)); | |
1431 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
1432 SSA_NAME_DEF_STMT (param) = stmt; | |
1433 | |
1434 stmt = gimple_build_assign (new_data, | |
1435 fold_convert (TREE_TYPE (new_data), param)); | |
1436 gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
1437 SSA_NAME_DEF_STMT (new_data) = stmt; | |
1438 } | |
1439 | |
1440 /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_PARALLEL. */ | |
1441 bb = split_loop_exit_edge (single_dom_exit (loop)); | |
1442 gsi = gsi_last_bb (bb); | |
1443 gsi_insert_after (&gsi, gimple_build_omp_return (false), GSI_NEW_STMT); | |
1444 | |
1445 /* Extract data for GIMPLE_OMP_FOR. */ | |
1446 gcc_assert (loop->header == single_dom_exit (loop)->src); | |
1447 cond_stmt = last_stmt (loop->header); | |
1448 | |
1449 cvar = gimple_cond_lhs (cond_stmt); | |
1450 cvar_base = SSA_NAME_VAR (cvar); | |
1451 phi = SSA_NAME_DEF_STMT (cvar); | |
1452 cvar_init = PHI_ARG_DEF_FROM_EDGE (phi, loop_preheader_edge (loop)); | |
1453 initvar = make_ssa_name (cvar_base, NULL); | |
1454 SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, loop_preheader_edge (loop)), | |
1455 initvar); | |
1456 cvar_next = PHI_ARG_DEF_FROM_EDGE (phi, loop_latch_edge (loop)); | |
1457 | |
1458 gsi = gsi_last_bb (loop->latch); | |
1459 gcc_assert (gsi_stmt (gsi) == SSA_NAME_DEF_STMT (cvar_next)); | |
1460 gsi_remove (&gsi, true); | |
1461 | |
1462 /* Prepare cfg. */ | |
1463 for_bb = split_edge (loop_preheader_edge (loop)); | |
1464 ex_bb = split_loop_exit_edge (single_dom_exit (loop)); | |
1465 extract_true_false_edges_from_block (loop->header, &nexit, &exit); | |
1466 gcc_assert (exit == single_dom_exit (loop)); | |
1467 | |
1468 guard = make_edge (for_bb, ex_bb, 0); | |
1469 single_succ_edge (loop->latch)->flags = 0; | |
1470 end = make_edge (loop->latch, ex_bb, EDGE_FALLTHRU); | |
1471 for (gsi = gsi_start_phis (ex_bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1472 { | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1473 source_location locus; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1474 tree def; |
0 | 1475 phi = gsi_stmt (gsi); |
1476 stmt = SSA_NAME_DEF_STMT (PHI_ARG_DEF_FROM_EDGE (phi, exit)); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1477 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1478 def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_preheader_edge (loop)); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1479 locus = gimple_phi_arg_location_from_edge (stmt, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1480 loop_preheader_edge (loop)); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1481 add_phi_arg (phi, def, guard, locus); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1482 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1483 def = PHI_ARG_DEF_FROM_EDGE (stmt, loop_latch_edge (loop)); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1484 locus = gimple_phi_arg_location_from_edge (stmt, loop_latch_edge (loop)); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1485 add_phi_arg (phi, def, end, locus); |
0 | 1486 } |
1487 e = redirect_edge_and_branch (exit, nexit->dest); | |
1488 PENDING_STMT (e) = NULL; | |
1489 | |
1490 /* Emit GIMPLE_OMP_FOR. */ | |
1491 gimple_cond_set_lhs (cond_stmt, cvar_base); | |
1492 type = TREE_TYPE (cvar); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1493 t = build_omp_clause (BUILTINS_LOCATION, OMP_CLAUSE_SCHEDULE); |
0 | 1494 OMP_CLAUSE_SCHEDULE_KIND (t) = OMP_CLAUSE_SCHEDULE_STATIC; |
1495 | |
1496 for_stmt = gimple_build_omp_for (NULL, t, 1, NULL); | |
1497 gimple_omp_for_set_index (for_stmt, 0, initvar); | |
1498 gimple_omp_for_set_initial (for_stmt, 0, cvar_init); | |
1499 gimple_omp_for_set_final (for_stmt, 0, gimple_cond_rhs (cond_stmt)); | |
1500 gimple_omp_for_set_cond (for_stmt, 0, gimple_cond_code (cond_stmt)); | |
1501 gimple_omp_for_set_incr (for_stmt, 0, build2 (PLUS_EXPR, type, | |
1502 cvar_base, | |
1503 build_int_cst (type, 1))); | |
1504 | |
1505 gsi = gsi_last_bb (for_bb); | |
1506 gsi_insert_after (&gsi, for_stmt, GSI_NEW_STMT); | |
1507 SSA_NAME_DEF_STMT (initvar) = for_stmt; | |
1508 | |
1509 /* Emit GIMPLE_OMP_CONTINUE. */ | |
1510 gsi = gsi_last_bb (loop->latch); | |
1511 stmt = gimple_build_omp_continue (cvar_next, cvar); | |
1512 gsi_insert_after (&gsi, stmt, GSI_NEW_STMT); | |
1513 SSA_NAME_DEF_STMT (cvar_next) = stmt; | |
1514 | |
1515 /* Emit GIMPLE_OMP_RETURN for GIMPLE_OMP_FOR. */ | |
1516 gsi = gsi_last_bb (ex_bb); | |
1517 gsi_insert_after (&gsi, gimple_build_omp_return (true), GSI_NEW_STMT); | |
1518 | |
1519 return paral_bb; | |
1520 } | |
1521 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1522 /* Generates code to execute the iterations of LOOP in N_THREADS |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1523 threads in parallel. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1524 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1525 NITER describes number of iterations of LOOP. |
0 | 1526 REDUCTION_LIST describes the reductions existent in the LOOP. */ |
1527 | |
1528 static void | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1529 gen_parallel_loop (struct loop *loop, htab_t reduction_list, |
0 | 1530 unsigned n_threads, struct tree_niter_desc *niter) |
1531 { | |
1532 loop_iterator li; | |
1533 tree many_iterations_cond, type, nit; | |
1534 tree arg_struct, new_arg_struct; | |
1535 gimple_seq stmts; | |
1536 basic_block parallel_head; | |
1537 edge entry, exit; | |
1538 struct clsn_data clsn_data; | |
1539 unsigned prob; | |
1540 | |
1541 /* From | |
1542 | |
1543 --------------------------------------------------------------------- | |
1544 loop | |
1545 { | |
1546 IV = phi (INIT, IV + STEP) | |
1547 BODY1; | |
1548 if (COND) | |
1549 break; | |
1550 BODY2; | |
1551 } | |
1552 --------------------------------------------------------------------- | |
1553 | |
1554 with # of iterations NITER (possibly with MAY_BE_ZERO assumption), | |
1555 we generate the following code: | |
1556 | |
1557 --------------------------------------------------------------------- | |
1558 | |
1559 if (MAY_BE_ZERO | |
1560 || NITER < MIN_PER_THREAD * N_THREADS) | |
1561 goto original; | |
1562 | |
1563 BODY1; | |
1564 store all local loop-invariant variables used in body of the loop to DATA. | |
1565 GIMPLE_OMP_PARALLEL (OMP_CLAUSE_NUM_THREADS (N_THREADS), LOOPFN, DATA); | |
1566 load the variables from DATA. | |
1567 GIMPLE_OMP_FOR (IV = INIT; COND; IV += STEP) (OMP_CLAUSE_SCHEDULE (static)) | |
1568 BODY2; | |
1569 BODY1; | |
1570 GIMPLE_OMP_CONTINUE; | |
1571 GIMPLE_OMP_RETURN -- GIMPLE_OMP_FOR | |
1572 GIMPLE_OMP_RETURN -- GIMPLE_OMP_PARALLEL | |
1573 goto end; | |
1574 | |
1575 original: | |
1576 loop | |
1577 { | |
1578 IV = phi (INIT, IV + STEP) | |
1579 BODY1; | |
1580 if (COND) | |
1581 break; | |
1582 BODY2; | |
1583 } | |
1584 | |
1585 end: | |
1586 | |
1587 */ | |
1588 | |
1589 /* Create two versions of the loop -- in the old one, we know that the | |
1590 number of iterations is large enough, and we will transform it into the | |
1591 loop that will be split to loop_fn, the new one will be used for the | |
1592 remaining iterations. */ | |
1593 | |
1594 type = TREE_TYPE (niter->niter); | |
1595 nit = force_gimple_operand (unshare_expr (niter->niter), &stmts, true, | |
1596 NULL_TREE); | |
1597 if (stmts) | |
1598 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
1599 | |
1600 many_iterations_cond = | |
1601 fold_build2 (GE_EXPR, boolean_type_node, | |
1602 nit, build_int_cst (type, MIN_PER_THREAD * n_threads)); | |
1603 many_iterations_cond | |
1604 = fold_build2 (TRUTH_AND_EXPR, boolean_type_node, | |
1605 invert_truthvalue (unshare_expr (niter->may_be_zero)), | |
1606 many_iterations_cond); | |
1607 many_iterations_cond | |
1608 = force_gimple_operand (many_iterations_cond, &stmts, false, NULL_TREE); | |
1609 if (stmts) | |
1610 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
1611 if (!is_gimple_condexpr (many_iterations_cond)) | |
1612 { | |
1613 many_iterations_cond | |
1614 = force_gimple_operand (many_iterations_cond, &stmts, | |
1615 true, NULL_TREE); | |
1616 if (stmts) | |
1617 gsi_insert_seq_on_edge_immediate (loop_preheader_edge (loop), stmts); | |
1618 } | |
1619 | |
1620 initialize_original_copy_tables (); | |
1621 | |
1622 /* We assume that the loop usually iterates a lot. */ | |
1623 prob = 4 * REG_BR_PROB_BASE / 5; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1624 loop_version (loop, many_iterations_cond, NULL, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1625 prob, prob, REG_BR_PROB_BASE - prob, true); |
0 | 1626 update_ssa (TODO_update_ssa); |
1627 free_original_copy_tables (); | |
1628 | |
1629 /* Base all the induction variables in LOOP on a single control one. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1630 canonicalize_loop_ivs (loop, &nit); |
0 | 1631 |
1632 /* Ensure that the exit condition is the first statement in the loop. */ | |
1633 transform_to_exit_first_loop (loop, reduction_list, nit); | |
1634 | |
1635 /* Generate initializations for reductions. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1636 if (htab_elements (reduction_list) > 0) |
0 | 1637 htab_traverse (reduction_list, initialize_reductions, loop); |
1638 | |
1639 /* Eliminate the references to local variables from the loop. */ | |
1640 gcc_assert (single_exit (loop)); | |
1641 entry = loop_preheader_edge (loop); | |
1642 exit = single_dom_exit (loop); | |
1643 | |
1644 eliminate_local_variables (entry, exit); | |
1645 /* In the old loop, move all variables non-local to the loop to a structure | |
1646 and back, and create separate decls for the variables used in loop. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1647 separate_decls_in_region (entry, exit, reduction_list, &arg_struct, |
0 | 1648 &new_arg_struct, &clsn_data); |
1649 | |
1650 /* Create the parallel constructs. */ | |
1651 parallel_head = create_parallel_loop (loop, create_loop_fn (), arg_struct, | |
1652 new_arg_struct, n_threads); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1653 if (htab_elements (reduction_list) > 0) |
0 | 1654 create_call_for_reduction (loop, reduction_list, &clsn_data); |
1655 | |
1656 scev_reset (); | |
1657 | |
1658 /* Cancel the loop (it is simpler to do it here rather than to teach the | |
1659 expander to do it). */ | |
1660 cancel_loop_tree (loop); | |
1661 | |
1662 /* Free loop bound estimations that could contain references to | |
1663 removed statements. */ | |
1664 FOR_EACH_LOOP (li, loop, 0) | |
1665 free_numbers_of_iterations_estimates_loop (loop); | |
1666 | |
1667 /* Expand the parallel constructs. We do it directly here instead of running | |
1668 a separate expand_omp pass, since it is more efficient, and less likely to | |
1669 cause troubles with further analyses not being able to deal with the | |
1670 OMP trees. */ | |
1671 | |
1672 omp_expand_local (parallel_head); | |
1673 } | |
1674 | |
1675 /* Returns true when LOOP contains vector phi nodes. */ | |
1676 | |
1677 static bool | |
1678 loop_has_vector_phi_nodes (struct loop *loop ATTRIBUTE_UNUSED) | |
1679 { | |
1680 unsigned i; | |
1681 basic_block *bbs = get_loop_body_in_dom_order (loop); | |
1682 gimple_stmt_iterator gsi; | |
1683 bool res = true; | |
1684 | |
1685 for (i = 0; i < loop->num_nodes; i++) | |
1686 for (gsi = gsi_start_phis (bbs[i]); !gsi_end_p (gsi); gsi_next (&gsi)) | |
1687 if (TREE_CODE (TREE_TYPE (PHI_RESULT (gsi_stmt (gsi)))) == VECTOR_TYPE) | |
1688 goto end; | |
1689 | |
1690 res = false; | |
1691 end: | |
1692 free (bbs); | |
1693 return res; | |
1694 } | |
1695 | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1696 /* Create a reduction_info struct, initialize it with REDUC_STMT |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1697 and PHI, insert it to the REDUCTION_LIST. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1698 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1699 static void |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1700 build_new_reduction (htab_t reduction_list, gimple reduc_stmt, gimple phi) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1701 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1702 PTR *slot; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1703 struct reduction_info *new_reduction; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1704 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1705 gcc_assert (reduc_stmt); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1706 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1707 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1708 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1709 fprintf (dump_file, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1710 "Detected reduction. reduction stmt is: \n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1711 print_gimple_stmt (dump_file, reduc_stmt, 0, 0); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1712 fprintf (dump_file, "\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1713 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1714 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1715 new_reduction = XCNEW (struct reduction_info); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1716 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1717 new_reduction->reduc_stmt = reduc_stmt; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1718 new_reduction->reduc_phi = phi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1719 new_reduction->reduction_code = gimple_assign_rhs_code (reduc_stmt); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1720 slot = htab_find_slot (reduction_list, new_reduction, INSERT); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1721 *slot = new_reduction; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1722 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1723 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1724 /* Detect all reductions in the LOOP, insert them into REDUCTION_LIST. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1725 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1726 static void |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1727 gather_scalar_reductions (loop_p loop, htab_t reduction_list) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1728 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1729 gimple_stmt_iterator gsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1730 loop_vec_info simple_loop_info; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1731 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1732 vect_dump = NULL; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1733 simple_loop_info = vect_analyze_loop_form (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1734 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1735 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1736 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1737 gimple phi = gsi_stmt (gsi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1738 affine_iv iv; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1739 tree res = PHI_RESULT (phi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1740 bool double_reduc; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1741 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1742 if (!is_gimple_reg (res)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1743 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1744 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1745 if (!simple_iv (loop, loop, res, &iv, true) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1746 && simple_loop_info) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1747 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1748 gimple reduc_stmt = vect_is_simple_reduction (simple_loop_info, phi, true, &double_reduc); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1749 if (reduc_stmt && !double_reduc) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1750 build_new_reduction (reduction_list, reduc_stmt, phi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1751 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1752 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1753 destroy_loop_vec_info (simple_loop_info, true); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1754 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1755 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1756 /* Try to initialize NITER for code generation part. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1757 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1758 static bool |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1759 try_get_loop_niter (loop_p loop, struct tree_niter_desc *niter) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1760 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1761 edge exit = single_dom_exit (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1762 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1763 gcc_assert (exit); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1764 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1765 /* We need to know # of iterations, and there should be no uses of values |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1766 defined inside loop outside of it, unless the values are invariants of |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1767 the loop. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1768 if (!number_of_iterations_exit (loop, exit, niter, false)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1769 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1770 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1771 fprintf (dump_file, " FAILED: number of iterations not known\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1772 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1773 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1774 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1775 return true; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1776 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1777 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1778 /* Try to initialize REDUCTION_LIST for code generation part. |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1779 REDUCTION_LIST describes the reductions. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1780 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1781 static bool |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1782 try_create_reduction_list (loop_p loop, htab_t reduction_list) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1783 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1784 edge exit = single_dom_exit (loop); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1785 gimple_stmt_iterator gsi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1786 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1787 gcc_assert (exit); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1788 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1789 gather_scalar_reductions (loop, reduction_list); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1790 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1791 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1792 for (gsi = gsi_start_phis (exit->dest); !gsi_end_p (gsi); gsi_next (&gsi)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1793 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1794 gimple phi = gsi_stmt (gsi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1795 struct reduction_info *red; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1796 imm_use_iterator imm_iter; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1797 use_operand_p use_p; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1798 gimple reduc_phi; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1799 tree val = PHI_ARG_DEF_FROM_EDGE (phi, exit); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1800 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1801 if (is_gimple_reg (val)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1802 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1803 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1804 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1805 fprintf (dump_file, "phi is "); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1806 print_gimple_stmt (dump_file, phi, 0, 0); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1807 fprintf (dump_file, "arg of phi to exit: value "); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1808 print_generic_expr (dump_file, val, 0); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1809 fprintf (dump_file, " used outside loop\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1810 fprintf (dump_file, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1811 " checking if it a part of reduction pattern: \n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1812 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1813 if (htab_elements (reduction_list) == 0) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1814 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1815 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1816 fprintf (dump_file, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1817 " FAILED: it is not a part of reduction.\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1818 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1819 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1820 reduc_phi = NULL; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1821 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, val) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1822 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1823 if (flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p)))) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1824 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1825 reduc_phi = USE_STMT (use_p); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1826 break; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1827 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1828 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1829 red = reduction_phi (reduction_list, reduc_phi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1830 if (red == NULL) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1831 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1832 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1833 fprintf (dump_file, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1834 " FAILED: it is not a part of reduction.\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1835 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1836 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1837 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1838 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1839 fprintf (dump_file, "reduction phi is "); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1840 print_gimple_stmt (dump_file, red->reduc_phi, 0, 0); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1841 fprintf (dump_file, "reduction stmt is "); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1842 print_gimple_stmt (dump_file, red->reduc_stmt, 0, 0); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1843 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1844 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1845 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1846 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1847 /* The iterations of the loop may communicate only through bivs whose |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1848 iteration space can be distributed efficiently. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1849 for (gsi = gsi_start_phis (loop->header); !gsi_end_p (gsi); gsi_next (&gsi)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1850 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1851 gimple phi = gsi_stmt (gsi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1852 tree def = PHI_RESULT (phi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1853 affine_iv iv; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1854 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1855 if (is_gimple_reg (def) && !simple_iv (loop, loop, def, &iv, true)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1856 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1857 struct reduction_info *red; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1858 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1859 red = reduction_phi (reduction_list, phi); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1860 if (red == NULL) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1861 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1862 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1863 fprintf (dump_file, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1864 " FAILED: scalar dependency between iterations\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1865 return false; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1866 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1867 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1868 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1869 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1870 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1871 return true; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1872 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1873 |
0 | 1874 /* Detect parallel loops and generate parallel code using libgomp |
1875 primitives. Returns true if some loop was parallelized, false | |
1876 otherwise. */ | |
1877 | |
1878 bool | |
1879 parallelize_loops (void) | |
1880 { | |
1881 unsigned n_threads = flag_tree_parallelize_loops; | |
1882 bool changed = false; | |
1883 struct loop *loop; | |
1884 struct tree_niter_desc niter_desc; | |
1885 loop_iterator li; | |
1886 htab_t reduction_list; | |
1887 | |
1888 /* Do not parallelize loops in the functions created by parallelization. */ | |
1889 if (parallelized_function_p (cfun->decl)) | |
1890 return false; | |
1891 | |
1892 reduction_list = htab_create (10, reduction_info_hash, | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1893 reduction_info_eq, free); |
0 | 1894 init_stmt_vec_info_vec (); |
1895 | |
1896 FOR_EACH_LOOP (li, loop, 0) | |
1897 { | |
1898 htab_empty (reduction_list); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1899 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1900 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1901 fprintf (dump_file, "Trying loop %d as candidate\n",loop->num); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1902 if (loop->inner) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1903 fprintf (dump_file, "loop %d is not innermost\n",loop->num); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1904 else |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1905 fprintf (dump_file, "loop %d is innermost\n",loop->num); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1906 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1907 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1908 /* If we use autopar in graphite pass, we use its marked dependency |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1909 checking results. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1910 if (flag_loop_parallelize_all && !loop->can_be_parallel) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1911 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1912 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1913 fprintf (dump_file, "loop is not parallel according to graphite\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1914 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1915 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1916 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1917 if (!single_dom_exit (loop)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1918 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1919 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1920 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1921 fprintf (dump_file, "loop is !single_dom_exit\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1922 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1923 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1924 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1925 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1926 if (/* And of course, the loop must be parallelizable. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1927 !can_duplicate_loop_p (loop) |
0 | 1928 || loop_has_blocks_with_irreducible_flag (loop) |
1929 /* FIXME: the check for vector phi nodes could be removed. */ | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1930 || loop_has_vector_phi_nodes (loop)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1931 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1932 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1933 /* FIXME: Bypass this check as graphite doesn't update the |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1934 count and frequency correctly now. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1935 if (!flag_loop_parallelize_all |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1936 && ((estimated_loop_iterations_int (loop, false) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1937 <= (HOST_WIDE_INT) n_threads * MIN_PER_THREAD) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1938 /* Do not bother with loops in cold areas. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1939 || optimize_loop_nest_for_size_p (loop))) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1940 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1941 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1942 if (!try_get_loop_niter (loop, &niter_desc)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1943 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1944 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1945 if (!try_create_reduction_list (loop, reduction_list)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1946 continue; |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1947 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1948 if (!flag_loop_parallelize_all && !loop_parallel_p (loop)) |
0 | 1949 continue; |
1950 | |
1951 changed = true; | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1952 if (dump_file && (dump_flags & TDF_DETAILS)) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1953 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1954 fprintf (dump_file, "parallelizing "); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1955 if (loop->inner) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1956 fprintf (dump_file, "outer loop\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1957 else |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1958 fprintf (dump_file, "inner loop\n"); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1959 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1960 gen_parallel_loop (loop, reduction_list, |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1961 n_threads, &niter_desc); |
0 | 1962 verify_flow_info (); |
1963 verify_dominators (CDI_DOMINATORS); | |
1964 verify_loop_structure (); | |
1965 verify_loop_closed_ssa (); | |
1966 } | |
1967 | |
1968 free_stmt_vec_info_vec (); | |
1969 htab_delete (reduction_list); | |
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1970 |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1971 /* Parallelization will cause new function calls to be inserted through |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1972 which local variables will escape. Reset the points-to solutions |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1973 for ESCAPED and CALLUSED. */ |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1974 if (changed) |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1975 { |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1976 pt_solution_reset (&cfun->gimple_df->escaped); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1977 pt_solution_reset (&cfun->gimple_df->callused); |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1978 } |
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1979 |
0 | 1980 return changed; |
1981 } | |
1982 | |
1983 #include "gt-tree-parloops.h" |