]>
Commit | Line | Data |
---|---|---|
60cbb674 TS |
1 | /* OMP constructs' SIMD clone supporting code. |
2 | ||
cbe34bb5 | 3 | Copyright (C) 2005-2017 Free Software Foundation, Inc. |
60cbb674 TS |
4 | |
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
9 | Software Foundation; either version 3, or (at your option) any later | |
10 | version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | #include "config.h" | |
22 | #include "system.h" | |
23 | #include "coretypes.h" | |
24 | #include "backend.h" | |
25 | #include "target.h" | |
26 | #include "tree.h" | |
27 | #include "gimple.h" | |
28 | #include "cfghooks.h" | |
29 | #include "alloc-pool.h" | |
30 | #include "tree-pass.h" | |
31 | #include "ssa.h" | |
32 | #include "cgraph.h" | |
33 | #include "pretty-print.h" | |
34 | #include "diagnostic-core.h" | |
35 | #include "fold-const.h" | |
36 | #include "stor-layout.h" | |
37 | #include "cfganal.h" | |
38 | #include "gimplify.h" | |
39 | #include "gimple-iterator.h" | |
40 | #include "gimplify-me.h" | |
41 | #include "gimple-walk.h" | |
42 | #include "langhooks.h" | |
43 | #include "tree-cfg.h" | |
44 | #include "tree-into-ssa.h" | |
45 | #include "tree-dfa.h" | |
46 | #include "cfgloop.h" | |
47 | #include "symbol-summary.h" | |
4d99a848 | 48 | #include "ipa-param-manipulation.h" |
60cbb674 | 49 | #include "tree-eh.h" |
ad200580 | 50 | #include "varasm.h" |
314e6352 ML |
51 | #include "stringpool.h" |
52 | #include "attribs.h" | |
60cbb674 TS |
53 | |
54 | /* Allocate a fresh `simd_clone' and return it. NARGS is the number | |
55 | of arguments to reserve space for. */ | |
56 | ||
57 | static struct cgraph_simd_clone * | |
58 | simd_clone_struct_alloc (int nargs) | |
59 | { | |
60 | struct cgraph_simd_clone *clone_info; | |
61 | size_t len = (sizeof (struct cgraph_simd_clone) | |
62 | + nargs * sizeof (struct cgraph_simd_clone_arg)); | |
63 | clone_info = (struct cgraph_simd_clone *) | |
64 | ggc_internal_cleared_alloc (len); | |
65 | return clone_info; | |
66 | } | |
67 | ||
68 | /* Make a copy of the `struct cgraph_simd_clone' in FROM to TO. */ | |
69 | ||
70 | static inline void | |
71 | simd_clone_struct_copy (struct cgraph_simd_clone *to, | |
72 | struct cgraph_simd_clone *from) | |
73 | { | |
74 | memcpy (to, from, (sizeof (struct cgraph_simd_clone) | |
75 | + ((from->nargs - from->inbranch) | |
76 | * sizeof (struct cgraph_simd_clone_arg)))); | |
77 | } | |
78 | ||
79 | /* Return vector of parameter types of function FNDECL. This uses | |
80 | TYPE_ARG_TYPES if available, otherwise falls back to types of | |
81 | DECL_ARGUMENTS types. */ | |
82 | ||
83 | static vec<tree> | |
84 | simd_clone_vector_of_formal_parm_types (tree fndecl) | |
85 | { | |
86 | if (TYPE_ARG_TYPES (TREE_TYPE (fndecl))) | |
87 | return ipa_get_vector_of_formal_parm_types (TREE_TYPE (fndecl)); | |
88 | vec<tree> args = ipa_get_vector_of_formal_parms (fndecl); | |
89 | unsigned int i; | |
90 | tree arg; | |
91 | FOR_EACH_VEC_ELT (args, i, arg) | |
92 | args[i] = TREE_TYPE (args[i]); | |
93 | return args; | |
94 | } | |
95 | ||
96 | /* Given a simd function in NODE, extract the simd specific | |
97 | information from the OMP clauses passed in CLAUSES, and return | |
98 | the struct cgraph_simd_clone * if it should be cloned. *INBRANCH_SPECIFIED | |
99 | is set to TRUE if the `inbranch' or `notinbranch' clause specified, | |
100 | otherwise set to FALSE. */ | |
101 | ||
102 | static struct cgraph_simd_clone * | |
103 | simd_clone_clauses_extract (struct cgraph_node *node, tree clauses, | |
104 | bool *inbranch_specified) | |
105 | { | |
106 | vec<tree> args = simd_clone_vector_of_formal_parm_types (node->decl); | |
107 | tree t; | |
108 | int n; | |
109 | *inbranch_specified = false; | |
110 | ||
111 | n = args.length (); | |
112 | if (n > 0 && args.last () == void_type_node) | |
113 | n--; | |
114 | ||
115 | /* To distinguish from an OpenMP simd clone, Cilk Plus functions to | |
116 | be cloned have a distinctive artificial label in addition to "omp | |
117 | declare simd". */ | |
118 | bool cilk_clone | |
119 | = (flag_cilkplus | |
120 | && lookup_attribute ("cilk simd function", | |
121 | DECL_ATTRIBUTES (node->decl))); | |
122 | ||
123 | /* Allocate one more than needed just in case this is an in-branch | |
124 | clone which will require a mask argument. */ | |
125 | struct cgraph_simd_clone *clone_info = simd_clone_struct_alloc (n + 1); | |
126 | clone_info->nargs = n; | |
127 | clone_info->cilk_elemental = cilk_clone; | |
128 | ||
129 | if (!clauses) | |
9dc5773f JJ |
130 | goto out; |
131 | ||
60cbb674 TS |
132 | clauses = TREE_VALUE (clauses); |
133 | if (!clauses || TREE_CODE (clauses) != OMP_CLAUSE) | |
9dc5773f | 134 | goto out; |
60cbb674 TS |
135 | |
136 | for (t = clauses; t; t = OMP_CLAUSE_CHAIN (t)) | |
137 | { | |
138 | switch (OMP_CLAUSE_CODE (t)) | |
139 | { | |
140 | case OMP_CLAUSE_INBRANCH: | |
141 | clone_info->inbranch = 1; | |
142 | *inbranch_specified = true; | |
143 | break; | |
144 | case OMP_CLAUSE_NOTINBRANCH: | |
145 | clone_info->inbranch = 0; | |
146 | *inbranch_specified = true; | |
147 | break; | |
148 | case OMP_CLAUSE_SIMDLEN: | |
149 | clone_info->simdlen | |
150 | = TREE_INT_CST_LOW (OMP_CLAUSE_SIMDLEN_EXPR (t)); | |
151 | break; | |
152 | case OMP_CLAUSE_LINEAR: | |
153 | { | |
154 | tree decl = OMP_CLAUSE_DECL (t); | |
155 | tree step = OMP_CLAUSE_LINEAR_STEP (t); | |
156 | int argno = TREE_INT_CST_LOW (decl); | |
157 | if (OMP_CLAUSE_LINEAR_VARIABLE_STRIDE (t)) | |
158 | { | |
159 | enum cgraph_simd_clone_arg_type arg_type; | |
160 | if (TREE_CODE (args[argno]) == REFERENCE_TYPE) | |
161 | switch (OMP_CLAUSE_LINEAR_KIND (t)) | |
162 | { | |
163 | case OMP_CLAUSE_LINEAR_REF: | |
164 | arg_type | |
165 | = SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP; | |
166 | break; | |
167 | case OMP_CLAUSE_LINEAR_UVAL: | |
168 | arg_type | |
169 | = SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP; | |
170 | break; | |
171 | case OMP_CLAUSE_LINEAR_VAL: | |
172 | case OMP_CLAUSE_LINEAR_DEFAULT: | |
173 | arg_type | |
174 | = SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP; | |
175 | break; | |
176 | default: | |
177 | gcc_unreachable (); | |
178 | } | |
179 | else | |
180 | arg_type = SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP; | |
181 | clone_info->args[argno].arg_type = arg_type; | |
182 | clone_info->args[argno].linear_step = tree_to_shwi (step); | |
183 | gcc_assert (clone_info->args[argno].linear_step >= 0 | |
184 | && clone_info->args[argno].linear_step < n); | |
185 | } | |
186 | else | |
187 | { | |
188 | if (POINTER_TYPE_P (args[argno])) | |
189 | step = fold_convert (ssizetype, step); | |
190 | if (!tree_fits_shwi_p (step)) | |
191 | { | |
192 | warning_at (OMP_CLAUSE_LOCATION (t), 0, | |
193 | "ignoring large linear step"); | |
194 | args.release (); | |
195 | return NULL; | |
196 | } | |
197 | else if (integer_zerop (step)) | |
198 | { | |
199 | warning_at (OMP_CLAUSE_LOCATION (t), 0, | |
200 | "ignoring zero linear step"); | |
201 | args.release (); | |
202 | return NULL; | |
203 | } | |
204 | else | |
205 | { | |
206 | enum cgraph_simd_clone_arg_type arg_type; | |
207 | if (TREE_CODE (args[argno]) == REFERENCE_TYPE) | |
208 | switch (OMP_CLAUSE_LINEAR_KIND (t)) | |
209 | { | |
210 | case OMP_CLAUSE_LINEAR_REF: | |
211 | arg_type | |
212 | = SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP; | |
213 | break; | |
214 | case OMP_CLAUSE_LINEAR_UVAL: | |
215 | arg_type | |
216 | = SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP; | |
217 | break; | |
218 | case OMP_CLAUSE_LINEAR_VAL: | |
219 | case OMP_CLAUSE_LINEAR_DEFAULT: | |
220 | arg_type | |
221 | = SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP; | |
222 | break; | |
223 | default: | |
224 | gcc_unreachable (); | |
225 | } | |
226 | else | |
227 | arg_type = SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP; | |
228 | clone_info->args[argno].arg_type = arg_type; | |
229 | clone_info->args[argno].linear_step = tree_to_shwi (step); | |
230 | } | |
231 | } | |
232 | break; | |
233 | } | |
234 | case OMP_CLAUSE_UNIFORM: | |
235 | { | |
236 | tree decl = OMP_CLAUSE_DECL (t); | |
237 | int argno = tree_to_uhwi (decl); | |
238 | clone_info->args[argno].arg_type | |
239 | = SIMD_CLONE_ARG_TYPE_UNIFORM; | |
240 | break; | |
241 | } | |
242 | case OMP_CLAUSE_ALIGNED: | |
243 | { | |
244 | tree decl = OMP_CLAUSE_DECL (t); | |
245 | int argno = tree_to_uhwi (decl); | |
246 | clone_info->args[argno].alignment | |
247 | = TREE_INT_CST_LOW (OMP_CLAUSE_ALIGNED_ALIGNMENT (t)); | |
248 | break; | |
249 | } | |
250 | default: | |
251 | break; | |
252 | } | |
253 | } | |
9dc5773f JJ |
254 | |
255 | out: | |
256 | if (TYPE_ATOMIC (TREE_TYPE (TREE_TYPE (node->decl)))) | |
257 | { | |
258 | warning_at (DECL_SOURCE_LOCATION (node->decl), 0, | |
259 | "ignoring %<#pragma omp declare simd%> on function " | |
260 | "with %<_Atomic%> qualified return type"); | |
261 | args.release (); | |
262 | return NULL; | |
263 | } | |
264 | ||
265 | for (unsigned int argno = 0; argno < clone_info->nargs; argno++) | |
266 | if (TYPE_ATOMIC (args[argno]) | |
267 | && clone_info->args[argno].arg_type != SIMD_CLONE_ARG_TYPE_UNIFORM) | |
268 | { | |
269 | warning_at (DECL_SOURCE_LOCATION (node->decl), 0, | |
270 | "ignoring %<#pragma omp declare simd%> on function " | |
271 | "with %<_Atomic%> qualified non-%<uniform%> argument"); | |
272 | args.release (); | |
273 | return NULL; | |
274 | } | |
275 | ||
60cbb674 TS |
276 | args.release (); |
277 | return clone_info; | |
278 | } | |
279 | ||
280 | /* Given a SIMD clone in NODE, calculate the characteristic data | |
281 | type and return the coresponding type. The characteristic data | |
282 | type is computed as described in the Intel Vector ABI. */ | |
283 | ||
284 | static tree | |
285 | simd_clone_compute_base_data_type (struct cgraph_node *node, | |
286 | struct cgraph_simd_clone *clone_info) | |
287 | { | |
288 | tree type = integer_type_node; | |
289 | tree fndecl = node->decl; | |
290 | ||
291 | /* a) For non-void function, the characteristic data type is the | |
292 | return type. */ | |
293 | if (TREE_CODE (TREE_TYPE (TREE_TYPE (fndecl))) != VOID_TYPE) | |
294 | type = TREE_TYPE (TREE_TYPE (fndecl)); | |
295 | ||
296 | /* b) If the function has any non-uniform, non-linear parameters, | |
297 | then the characteristic data type is the type of the first | |
298 | such parameter. */ | |
299 | else | |
300 | { | |
301 | vec<tree> map = simd_clone_vector_of_formal_parm_types (fndecl); | |
302 | for (unsigned int i = 0; i < clone_info->nargs; ++i) | |
303 | if (clone_info->args[i].arg_type == SIMD_CLONE_ARG_TYPE_VECTOR) | |
304 | { | |
305 | type = map[i]; | |
306 | break; | |
307 | } | |
308 | map.release (); | |
309 | } | |
310 | ||
311 | /* c) If the characteristic data type determined by a) or b) above | |
312 | is struct, union, or class type which is pass-by-value (except | |
313 | for the type that maps to the built-in complex data type), the | |
314 | characteristic data type is int. */ | |
315 | if (RECORD_OR_UNION_TYPE_P (type) | |
316 | && !aggregate_value_p (type, NULL) | |
317 | && TREE_CODE (type) != COMPLEX_TYPE) | |
318 | return integer_type_node; | |
319 | ||
320 | /* d) If none of the above three classes is applicable, the | |
321 | characteristic data type is int. */ | |
322 | ||
323 | return type; | |
324 | ||
325 | /* e) For Intel Xeon Phi native and offload compilation, if the | |
326 | resulting characteristic data type is 8-bit or 16-bit integer | |
327 | data type, the characteristic data type is int. */ | |
328 | /* Well, we don't handle Xeon Phi yet. */ | |
329 | } | |
330 | ||
331 | static tree | |
332 | simd_clone_mangle (struct cgraph_node *node, | |
333 | struct cgraph_simd_clone *clone_info) | |
334 | { | |
335 | char vecsize_mangle = clone_info->vecsize_mangle; | |
336 | char mask = clone_info->inbranch ? 'M' : 'N'; | |
337 | unsigned int simdlen = clone_info->simdlen; | |
338 | unsigned int n; | |
339 | pretty_printer pp; | |
340 | ||
341 | gcc_assert (vecsize_mangle && simdlen); | |
342 | ||
343 | pp_string (&pp, "_ZGV"); | |
344 | pp_character (&pp, vecsize_mangle); | |
345 | pp_character (&pp, mask); | |
346 | pp_decimal_int (&pp, simdlen); | |
347 | ||
348 | for (n = 0; n < clone_info->nargs; ++n) | |
349 | { | |
350 | struct cgraph_simd_clone_arg arg = clone_info->args[n]; | |
351 | ||
352 | switch (arg.arg_type) | |
353 | { | |
354 | case SIMD_CLONE_ARG_TYPE_UNIFORM: | |
355 | pp_character (&pp, 'u'); | |
356 | break; | |
357 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
358 | pp_character (&pp, 'l'); | |
359 | goto mangle_linear; | |
360 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: | |
361 | pp_character (&pp, 'R'); | |
362 | goto mangle_linear; | |
363 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: | |
364 | pp_character (&pp, 'L'); | |
365 | goto mangle_linear; | |
366 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
367 | pp_character (&pp, 'U'); | |
368 | goto mangle_linear; | |
369 | mangle_linear: | |
370 | gcc_assert (arg.linear_step != 0); | |
371 | if (arg.linear_step > 1) | |
372 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
373 | else if (arg.linear_step < 0) | |
374 | { | |
375 | pp_character (&pp, 'n'); | |
376 | pp_unsigned_wide_integer (&pp, (-(unsigned HOST_WIDE_INT) | |
377 | arg.linear_step)); | |
378 | } | |
379 | break; | |
380 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
381 | pp_string (&pp, "ls"); | |
382 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
383 | break; | |
384 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: | |
385 | pp_string (&pp, "Rs"); | |
386 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
387 | break; | |
388 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
389 | pp_string (&pp, "Ls"); | |
390 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
391 | break; | |
392 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
393 | pp_string (&pp, "Us"); | |
394 | pp_unsigned_wide_integer (&pp, arg.linear_step); | |
395 | break; | |
396 | default: | |
397 | pp_character (&pp, 'v'); | |
398 | } | |
399 | if (arg.alignment) | |
400 | { | |
401 | pp_character (&pp, 'a'); | |
402 | pp_decimal_int (&pp, arg.alignment); | |
403 | } | |
404 | } | |
405 | ||
406 | pp_underscore (&pp); | |
407 | const char *str = IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (node->decl)); | |
408 | if (*str == '*') | |
409 | ++str; | |
410 | pp_string (&pp, str); | |
411 | str = pp_formatted_text (&pp); | |
412 | ||
413 | /* If there already is a SIMD clone with the same mangled name, don't | |
414 | add another one. This can happen e.g. for | |
415 | #pragma omp declare simd | |
416 | #pragma omp declare simd simdlen(8) | |
417 | int foo (int, int); | |
418 | if the simdlen is assumed to be 8 for the first one, etc. */ | |
419 | for (struct cgraph_node *clone = node->simd_clones; clone; | |
420 | clone = clone->simdclone->next_clone) | |
a01f151f | 421 | if (id_equal (DECL_ASSEMBLER_NAME (clone->decl), str)) |
60cbb674 TS |
422 | return NULL_TREE; |
423 | ||
424 | return get_identifier (str); | |
425 | } | |
426 | ||
427 | /* Create a simd clone of OLD_NODE and return it. */ | |
428 | ||
429 | static struct cgraph_node * | |
430 | simd_clone_create (struct cgraph_node *old_node) | |
431 | { | |
432 | struct cgraph_node *new_node; | |
433 | if (old_node->definition) | |
434 | { | |
435 | if (!old_node->has_gimple_body_p ()) | |
436 | return NULL; | |
437 | old_node->get_body (); | |
438 | new_node = old_node->create_version_clone_with_body (vNULL, NULL, NULL, | |
439 | false, NULL, NULL, | |
440 | "simdclone"); | |
441 | } | |
442 | else | |
443 | { | |
444 | tree old_decl = old_node->decl; | |
445 | tree new_decl = copy_node (old_node->decl); | |
446 | DECL_NAME (new_decl) = clone_function_name (old_decl, "simdclone"); | |
447 | SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl)); | |
448 | SET_DECL_RTL (new_decl, NULL); | |
449 | DECL_STATIC_CONSTRUCTOR (new_decl) = 0; | |
450 | DECL_STATIC_DESTRUCTOR (new_decl) = 0; | |
451 | new_node = old_node->create_version_clone (new_decl, vNULL, NULL); | |
452 | if (old_node->in_other_partition) | |
453 | new_node->in_other_partition = 1; | |
454 | } | |
455 | if (new_node == NULL) | |
456 | return new_node; | |
457 | ||
458 | TREE_PUBLIC (new_node->decl) = TREE_PUBLIC (old_node->decl); | |
ad200580 JJ |
459 | DECL_COMDAT (new_node->decl) = DECL_COMDAT (old_node->decl); |
460 | DECL_WEAK (new_node->decl) = DECL_WEAK (old_node->decl); | |
461 | DECL_EXTERNAL (new_node->decl) = DECL_EXTERNAL (old_node->decl); | |
462 | DECL_VISIBILITY_SPECIFIED (new_node->decl) | |
463 | = DECL_VISIBILITY_SPECIFIED (old_node->decl); | |
464 | DECL_VISIBILITY (new_node->decl) = DECL_VISIBILITY (old_node->decl); | |
465 | DECL_DLLIMPORT_P (new_node->decl) = DECL_DLLIMPORT_P (old_node->decl); | |
466 | if (DECL_ONE_ONLY (old_node->decl)) | |
467 | make_decl_one_only (new_node->decl, DECL_ASSEMBLER_NAME (new_node->decl)); | |
468 | ||
469 | /* The method cgraph_version_clone_with_body () will force the new | |
470 | symbol local. Undo this, and inherit external visibility from | |
60cbb674 TS |
471 | the old node. */ |
472 | new_node->local.local = old_node->local.local; | |
473 | new_node->externally_visible = old_node->externally_visible; | |
474 | ||
475 | return new_node; | |
476 | } | |
477 | ||
478 | /* Adjust the return type of the given function to its appropriate | |
479 | vector counterpart. Returns a simd array to be used throughout the | |
480 | function as a return value. */ | |
481 | ||
482 | static tree | |
483 | simd_clone_adjust_return_type (struct cgraph_node *node) | |
484 | { | |
485 | tree fndecl = node->decl; | |
486 | tree orig_rettype = TREE_TYPE (TREE_TYPE (fndecl)); | |
487 | unsigned int veclen; | |
488 | tree t; | |
489 | ||
490 | /* Adjust the function return type. */ | |
491 | if (orig_rettype == void_type_node) | |
492 | return NULL_TREE; | |
493 | TREE_TYPE (fndecl) = build_distinct_type_copy (TREE_TYPE (fndecl)); | |
494 | t = TREE_TYPE (TREE_TYPE (fndecl)); | |
495 | if (INTEGRAL_TYPE_P (t) || POINTER_TYPE_P (t)) | |
496 | veclen = node->simdclone->vecsize_int; | |
497 | else | |
498 | veclen = node->simdclone->vecsize_float; | |
b397965c | 499 | veclen /= GET_MODE_BITSIZE (SCALAR_TYPE_MODE (t)); |
60cbb674 TS |
500 | if (veclen > node->simdclone->simdlen) |
501 | veclen = node->simdclone->simdlen; | |
502 | if (POINTER_TYPE_P (t)) | |
503 | t = pointer_sized_int_node; | |
504 | if (veclen == node->simdclone->simdlen) | |
505 | t = build_vector_type (t, node->simdclone->simdlen); | |
506 | else | |
507 | { | |
508 | t = build_vector_type (t, veclen); | |
509 | t = build_array_type_nelts (t, node->simdclone->simdlen / veclen); | |
510 | } | |
511 | TREE_TYPE (TREE_TYPE (fndecl)) = t; | |
512 | if (!node->definition) | |
513 | return NULL_TREE; | |
514 | ||
515 | t = DECL_RESULT (fndecl); | |
516 | /* Adjust the DECL_RESULT. */ | |
517 | gcc_assert (TREE_TYPE (t) != void_type_node); | |
518 | TREE_TYPE (t) = TREE_TYPE (TREE_TYPE (fndecl)); | |
519 | relayout_decl (t); | |
520 | ||
521 | tree atype = build_array_type_nelts (orig_rettype, | |
522 | node->simdclone->simdlen); | |
523 | if (veclen != node->simdclone->simdlen) | |
524 | return build1 (VIEW_CONVERT_EXPR, atype, t); | |
525 | ||
526 | /* Set up a SIMD array to use as the return value. */ | |
527 | tree retval = create_tmp_var_raw (atype, "retval"); | |
528 | gimple_add_tmp_var (retval); | |
529 | return retval; | |
530 | } | |
531 | ||
532 | /* Each vector argument has a corresponding array to be used locally | |
533 | as part of the eventual loop. Create such temporary array and | |
534 | return it. | |
535 | ||
536 | PREFIX is the prefix to be used for the temporary. | |
537 | ||
538 | TYPE is the inner element type. | |
539 | ||
540 | SIMDLEN is the number of elements. */ | |
541 | ||
542 | static tree | |
543 | create_tmp_simd_array (const char *prefix, tree type, int simdlen) | |
544 | { | |
545 | tree atype = build_array_type_nelts (type, simdlen); | |
546 | tree avar = create_tmp_var_raw (atype, prefix); | |
547 | gimple_add_tmp_var (avar); | |
548 | return avar; | |
549 | } | |
550 | ||
551 | /* Modify the function argument types to their corresponding vector | |
552 | counterparts if appropriate. Also, create one array for each simd | |
553 | argument to be used locally when using the function arguments as | |
554 | part of the loop. | |
555 | ||
556 | NODE is the function whose arguments are to be adjusted. | |
557 | ||
558 | Returns an adjustment vector that will be filled describing how the | |
559 | argument types will be adjusted. */ | |
560 | ||
561 | static ipa_parm_adjustment_vec | |
562 | simd_clone_adjust_argument_types (struct cgraph_node *node) | |
563 | { | |
564 | vec<tree> args; | |
565 | ipa_parm_adjustment_vec adjustments; | |
566 | ||
567 | if (node->definition) | |
568 | args = ipa_get_vector_of_formal_parms (node->decl); | |
569 | else | |
570 | args = simd_clone_vector_of_formal_parm_types (node->decl); | |
571 | adjustments.create (args.length ()); | |
572 | unsigned i, j, veclen; | |
573 | struct ipa_parm_adjustment adj; | |
574 | struct cgraph_simd_clone *sc = node->simdclone; | |
575 | ||
576 | for (i = 0; i < sc->nargs; ++i) | |
577 | { | |
578 | memset (&adj, 0, sizeof (adj)); | |
579 | tree parm = args[i]; | |
580 | tree parm_type = node->definition ? TREE_TYPE (parm) : parm; | |
581 | adj.base_index = i; | |
582 | adj.base = parm; | |
583 | ||
584 | sc->args[i].orig_arg = node->definition ? parm : NULL_TREE; | |
585 | sc->args[i].orig_type = parm_type; | |
586 | ||
587 | switch (sc->args[i].arg_type) | |
588 | { | |
589 | default: | |
590 | /* No adjustment necessary for scalar arguments. */ | |
591 | adj.op = IPA_PARM_OP_COPY; | |
592 | break; | |
593 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
594 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
595 | if (node->definition) | |
596 | sc->args[i].simd_array | |
597 | = create_tmp_simd_array (IDENTIFIER_POINTER (DECL_NAME (parm)), | |
598 | TREE_TYPE (parm_type), | |
599 | sc->simdlen); | |
600 | adj.op = IPA_PARM_OP_COPY; | |
601 | break; | |
602 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: | |
603 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
604 | case SIMD_CLONE_ARG_TYPE_VECTOR: | |
605 | if (INTEGRAL_TYPE_P (parm_type) || POINTER_TYPE_P (parm_type)) | |
606 | veclen = sc->vecsize_int; | |
607 | else | |
608 | veclen = sc->vecsize_float; | |
b397965c | 609 | veclen /= GET_MODE_BITSIZE (SCALAR_TYPE_MODE (parm_type)); |
60cbb674 TS |
610 | if (veclen > sc->simdlen) |
611 | veclen = sc->simdlen; | |
612 | adj.arg_prefix = "simd"; | |
613 | if (POINTER_TYPE_P (parm_type)) | |
614 | adj.type = build_vector_type (pointer_sized_int_node, veclen); | |
615 | else | |
616 | adj.type = build_vector_type (parm_type, veclen); | |
617 | sc->args[i].vector_type = adj.type; | |
618 | for (j = veclen; j < sc->simdlen; j += veclen) | |
619 | { | |
620 | adjustments.safe_push (adj); | |
621 | if (j == veclen) | |
622 | { | |
623 | memset (&adj, 0, sizeof (adj)); | |
624 | adj.op = IPA_PARM_OP_NEW; | |
625 | adj.arg_prefix = "simd"; | |
626 | adj.base_index = i; | |
627 | adj.type = sc->args[i].vector_type; | |
628 | } | |
629 | } | |
630 | ||
631 | if (node->definition) | |
632 | sc->args[i].simd_array | |
699e8cb7 JJ |
633 | = create_tmp_simd_array (DECL_NAME (parm) |
634 | ? IDENTIFIER_POINTER (DECL_NAME (parm)) | |
635 | : NULL, parm_type, sc->simdlen); | |
60cbb674 TS |
636 | } |
637 | adjustments.safe_push (adj); | |
638 | } | |
639 | ||
640 | if (sc->inbranch) | |
641 | { | |
642 | tree base_type = simd_clone_compute_base_data_type (sc->origin, sc); | |
643 | ||
644 | memset (&adj, 0, sizeof (adj)); | |
645 | adj.op = IPA_PARM_OP_NEW; | |
646 | adj.arg_prefix = "mask"; | |
647 | ||
648 | adj.base_index = i; | |
649 | if (INTEGRAL_TYPE_P (base_type) || POINTER_TYPE_P (base_type)) | |
650 | veclen = sc->vecsize_int; | |
651 | else | |
652 | veclen = sc->vecsize_float; | |
b397965c | 653 | veclen /= GET_MODE_BITSIZE (SCALAR_TYPE_MODE (base_type)); |
60cbb674 TS |
654 | if (veclen > sc->simdlen) |
655 | veclen = sc->simdlen; | |
656 | if (sc->mask_mode != VOIDmode) | |
657 | adj.type | |
658 | = lang_hooks.types.type_for_mode (sc->mask_mode, 1); | |
659 | else if (POINTER_TYPE_P (base_type)) | |
660 | adj.type = build_vector_type (pointer_sized_int_node, veclen); | |
661 | else | |
662 | adj.type = build_vector_type (base_type, veclen); | |
663 | adjustments.safe_push (adj); | |
664 | ||
665 | for (j = veclen; j < sc->simdlen; j += veclen) | |
666 | adjustments.safe_push (adj); | |
667 | ||
668 | /* We have previously allocated one extra entry for the mask. Use | |
669 | it and fill it. */ | |
670 | sc->nargs++; | |
671 | if (sc->mask_mode != VOIDmode) | |
672 | base_type = boolean_type_node; | |
673 | if (node->definition) | |
674 | { | |
675 | sc->args[i].orig_arg | |
676 | = build_decl (UNKNOWN_LOCATION, PARM_DECL, NULL, base_type); | |
677 | if (sc->mask_mode == VOIDmode) | |
678 | sc->args[i].simd_array | |
679 | = create_tmp_simd_array ("mask", base_type, sc->simdlen); | |
680 | else if (veclen < sc->simdlen) | |
681 | sc->args[i].simd_array | |
682 | = create_tmp_simd_array ("mask", adj.type, sc->simdlen / veclen); | |
683 | else | |
684 | sc->args[i].simd_array = NULL_TREE; | |
685 | } | |
686 | sc->args[i].orig_type = base_type; | |
687 | sc->args[i].arg_type = SIMD_CLONE_ARG_TYPE_MASK; | |
688 | } | |
689 | ||
690 | if (node->definition) | |
691 | ipa_modify_formal_parameters (node->decl, adjustments); | |
692 | else | |
693 | { | |
694 | tree new_arg_types = NULL_TREE, new_reversed; | |
695 | bool last_parm_void = false; | |
696 | if (args.length () > 0 && args.last () == void_type_node) | |
697 | last_parm_void = true; | |
698 | ||
699 | gcc_assert (TYPE_ARG_TYPES (TREE_TYPE (node->decl))); | |
700 | j = adjustments.length (); | |
701 | for (i = 0; i < j; i++) | |
702 | { | |
703 | struct ipa_parm_adjustment *adj = &adjustments[i]; | |
704 | tree ptype; | |
705 | if (adj->op == IPA_PARM_OP_COPY) | |
706 | ptype = args[adj->base_index]; | |
707 | else | |
708 | ptype = adj->type; | |
709 | new_arg_types = tree_cons (NULL_TREE, ptype, new_arg_types); | |
710 | } | |
711 | new_reversed = nreverse (new_arg_types); | |
712 | if (last_parm_void) | |
713 | { | |
714 | if (new_reversed) | |
715 | TREE_CHAIN (new_arg_types) = void_list_node; | |
716 | else | |
717 | new_reversed = void_list_node; | |
718 | } | |
719 | ||
720 | tree new_type = build_distinct_type_copy (TREE_TYPE (node->decl)); | |
721 | TYPE_ARG_TYPES (new_type) = new_reversed; | |
722 | TREE_TYPE (node->decl) = new_type; | |
723 | ||
724 | adjustments.release (); | |
725 | } | |
726 | args.release (); | |
727 | return adjustments; | |
728 | } | |
729 | ||
730 | /* Initialize and copy the function arguments in NODE to their | |
731 | corresponding local simd arrays. Returns a fresh gimple_seq with | |
732 | the instruction sequence generated. */ | |
733 | ||
734 | static gimple_seq | |
735 | simd_clone_init_simd_arrays (struct cgraph_node *node, | |
736 | ipa_parm_adjustment_vec adjustments) | |
737 | { | |
738 | gimple_seq seq = NULL; | |
739 | unsigned i = 0, j = 0, k; | |
740 | ||
741 | for (tree arg = DECL_ARGUMENTS (node->decl); | |
742 | arg; | |
743 | arg = DECL_CHAIN (arg), i++, j++) | |
744 | { | |
745 | if (adjustments[j].op == IPA_PARM_OP_COPY | |
746 | || POINTER_TYPE_P (TREE_TYPE (arg))) | |
747 | continue; | |
748 | ||
749 | node->simdclone->args[i].vector_arg = arg; | |
750 | ||
751 | tree array = node->simdclone->args[i].simd_array; | |
752 | if (node->simdclone->mask_mode != VOIDmode | |
753 | && node->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_MASK) | |
754 | { | |
755 | if (array == NULL_TREE) | |
756 | continue; | |
757 | unsigned int l | |
758 | = tree_to_uhwi (TYPE_MAX_VALUE (TYPE_DOMAIN (TREE_TYPE (array)))); | |
759 | for (k = 0; k <= l; k++) | |
760 | { | |
761 | if (k) | |
762 | { | |
763 | arg = DECL_CHAIN (arg); | |
764 | j++; | |
765 | } | |
766 | tree t = build4 (ARRAY_REF, TREE_TYPE (TREE_TYPE (array)), | |
767 | array, size_int (k), NULL, NULL); | |
768 | t = build2 (MODIFY_EXPR, TREE_TYPE (t), t, arg); | |
769 | gimplify_and_add (t, &seq); | |
770 | } | |
771 | continue; | |
772 | } | |
773 | if (TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg)) == node->simdclone->simdlen) | |
774 | { | |
775 | tree ptype = build_pointer_type (TREE_TYPE (TREE_TYPE (array))); | |
776 | tree ptr = build_fold_addr_expr (array); | |
777 | tree t = build2 (MEM_REF, TREE_TYPE (arg), ptr, | |
778 | build_int_cst (ptype, 0)); | |
779 | t = build2 (MODIFY_EXPR, TREE_TYPE (t), t, arg); | |
780 | gimplify_and_add (t, &seq); | |
781 | } | |
782 | else | |
783 | { | |
784 | unsigned int simdlen = TYPE_VECTOR_SUBPARTS (TREE_TYPE (arg)); | |
785 | tree ptype = build_pointer_type (TREE_TYPE (TREE_TYPE (array))); | |
786 | for (k = 0; k < node->simdclone->simdlen; k += simdlen) | |
787 | { | |
788 | tree ptr = build_fold_addr_expr (array); | |
789 | int elemsize; | |
790 | if (k) | |
791 | { | |
792 | arg = DECL_CHAIN (arg); | |
793 | j++; | |
794 | } | |
b397965c RS |
795 | tree elemtype = TREE_TYPE (TREE_TYPE (arg)); |
796 | elemsize = GET_MODE_SIZE (SCALAR_TYPE_MODE (elemtype)); | |
60cbb674 TS |
797 | tree t = build2 (MEM_REF, TREE_TYPE (arg), ptr, |
798 | build_int_cst (ptype, k * elemsize)); | |
799 | t = build2 (MODIFY_EXPR, TREE_TYPE (t), t, arg); | |
800 | gimplify_and_add (t, &seq); | |
801 | } | |
802 | } | |
803 | } | |
804 | return seq; | |
805 | } | |
806 | ||
807 | /* Callback info for ipa_simd_modify_stmt_ops below. */ | |
808 | ||
809 | struct modify_stmt_info { | |
810 | ipa_parm_adjustment_vec adjustments; | |
811 | gimple *stmt; | |
812 | /* True if the parent statement was modified by | |
813 | ipa_simd_modify_stmt_ops. */ | |
814 | bool modified; | |
815 | }; | |
816 | ||
817 | /* Callback for walk_gimple_op. | |
818 | ||
819 | Adjust operands from a given statement as specified in the | |
820 | adjustments vector in the callback data. */ | |
821 | ||
822 | static tree | |
823 | ipa_simd_modify_stmt_ops (tree *tp, int *walk_subtrees, void *data) | |
824 | { | |
825 | struct walk_stmt_info *wi = (struct walk_stmt_info *) data; | |
826 | struct modify_stmt_info *info = (struct modify_stmt_info *) wi->info; | |
827 | tree *orig_tp = tp; | |
828 | if (TREE_CODE (*tp) == ADDR_EXPR) | |
829 | tp = &TREE_OPERAND (*tp, 0); | |
830 | struct ipa_parm_adjustment *cand = NULL; | |
831 | if (TREE_CODE (*tp) == PARM_DECL) | |
832 | cand = ipa_get_adjustment_candidate (&tp, NULL, info->adjustments, true); | |
833 | else | |
834 | { | |
835 | if (TYPE_P (*tp)) | |
836 | *walk_subtrees = 0; | |
837 | } | |
838 | ||
839 | tree repl = NULL_TREE; | |
840 | if (cand) | |
841 | repl = unshare_expr (cand->new_decl); | |
842 | else | |
843 | { | |
844 | if (tp != orig_tp) | |
845 | { | |
846 | *walk_subtrees = 0; | |
847 | bool modified = info->modified; | |
848 | info->modified = false; | |
849 | walk_tree (tp, ipa_simd_modify_stmt_ops, wi, wi->pset); | |
850 | if (!info->modified) | |
851 | { | |
852 | info->modified = modified; | |
853 | return NULL_TREE; | |
854 | } | |
855 | info->modified = modified; | |
856 | repl = *tp; | |
857 | } | |
858 | else | |
859 | return NULL_TREE; | |
860 | } | |
861 | ||
862 | if (tp != orig_tp) | |
863 | { | |
864 | repl = build_fold_addr_expr (repl); | |
865 | gimple *stmt; | |
866 | if (is_gimple_debug (info->stmt)) | |
867 | { | |
868 | tree vexpr = make_node (DEBUG_EXPR_DECL); | |
869 | stmt = gimple_build_debug_source_bind (vexpr, repl, NULL); | |
870 | DECL_ARTIFICIAL (vexpr) = 1; | |
871 | TREE_TYPE (vexpr) = TREE_TYPE (repl); | |
899ca90e | 872 | SET_DECL_MODE (vexpr, TYPE_MODE (TREE_TYPE (repl))); |
60cbb674 TS |
873 | repl = vexpr; |
874 | } | |
875 | else | |
876 | { | |
877 | stmt = gimple_build_assign (make_ssa_name (TREE_TYPE (repl)), repl); | |
878 | repl = gimple_assign_lhs (stmt); | |
879 | } | |
880 | gimple_stmt_iterator gsi = gsi_for_stmt (info->stmt); | |
881 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
882 | *orig_tp = repl; | |
883 | } | |
884 | else if (!useless_type_conversion_p (TREE_TYPE (*tp), TREE_TYPE (repl))) | |
885 | { | |
886 | tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*tp), repl); | |
887 | *tp = vce; | |
888 | } | |
889 | else | |
890 | *tp = repl; | |
891 | ||
892 | info->modified = true; | |
893 | return NULL_TREE; | |
894 | } | |
895 | ||
896 | /* Traverse the function body and perform all modifications as | |
897 | described in ADJUSTMENTS. At function return, ADJUSTMENTS will be | |
898 | modified such that the replacement/reduction value will now be an | |
899 | offset into the corresponding simd_array. | |
900 | ||
901 | This function will replace all function argument uses with their | |
902 | corresponding simd array elements, and ajust the return values | |
903 | accordingly. */ | |
904 | ||
905 | static void | |
906 | ipa_simd_modify_function_body (struct cgraph_node *node, | |
907 | ipa_parm_adjustment_vec adjustments, | |
908 | tree retval_array, tree iter) | |
909 | { | |
910 | basic_block bb; | |
911 | unsigned int i, j, l; | |
912 | ||
913 | /* Re-use the adjustments array, but this time use it to replace | |
914 | every function argument use to an offset into the corresponding | |
915 | simd_array. */ | |
916 | for (i = 0, j = 0; i < node->simdclone->nargs; ++i, ++j) | |
917 | { | |
918 | if (!node->simdclone->args[i].vector_arg) | |
919 | continue; | |
920 | ||
921 | tree basetype = TREE_TYPE (node->simdclone->args[i].orig_arg); | |
922 | tree vectype = TREE_TYPE (node->simdclone->args[i].vector_arg); | |
923 | adjustments[j].new_decl | |
924 | = build4 (ARRAY_REF, | |
925 | basetype, | |
926 | node->simdclone->args[i].simd_array, | |
927 | iter, | |
928 | NULL_TREE, NULL_TREE); | |
929 | if (adjustments[j].op == IPA_PARM_OP_NONE | |
930 | && TYPE_VECTOR_SUBPARTS (vectype) < node->simdclone->simdlen) | |
931 | j += node->simdclone->simdlen / TYPE_VECTOR_SUBPARTS (vectype) - 1; | |
932 | } | |
933 | ||
934 | l = adjustments.length (); | |
46aa019a KV |
935 | tree name; |
936 | ||
937 | FOR_EACH_SSA_NAME (i, name, cfun) | |
60cbb674 | 938 | { |
46aa019a | 939 | if (SSA_NAME_VAR (name) |
60cbb674 TS |
940 | && TREE_CODE (SSA_NAME_VAR (name)) == PARM_DECL) |
941 | { | |
942 | for (j = 0; j < l; j++) | |
943 | if (SSA_NAME_VAR (name) == adjustments[j].base | |
944 | && adjustments[j].new_decl) | |
945 | { | |
946 | tree base_var; | |
947 | if (adjustments[j].new_ssa_base == NULL_TREE) | |
948 | { | |
949 | base_var | |
950 | = copy_var_decl (adjustments[j].base, | |
951 | DECL_NAME (adjustments[j].base), | |
952 | TREE_TYPE (adjustments[j].base)); | |
953 | adjustments[j].new_ssa_base = base_var; | |
954 | } | |
955 | else | |
956 | base_var = adjustments[j].new_ssa_base; | |
957 | if (SSA_NAME_IS_DEFAULT_DEF (name)) | |
958 | { | |
959 | bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
960 | gimple_stmt_iterator gsi = gsi_after_labels (bb); | |
961 | tree new_decl = unshare_expr (adjustments[j].new_decl); | |
962 | set_ssa_default_def (cfun, adjustments[j].base, NULL_TREE); | |
963 | SET_SSA_NAME_VAR_OR_IDENTIFIER (name, base_var); | |
964 | SSA_NAME_IS_DEFAULT_DEF (name) = 0; | |
965 | gimple *stmt = gimple_build_assign (name, new_decl); | |
966 | gsi_insert_before (&gsi, stmt, GSI_SAME_STMT); | |
967 | } | |
968 | else | |
969 | SET_SSA_NAME_VAR_OR_IDENTIFIER (name, base_var); | |
970 | } | |
971 | } | |
972 | } | |
973 | ||
974 | struct modify_stmt_info info; | |
975 | info.adjustments = adjustments; | |
976 | ||
977 | FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (node->decl)) | |
978 | { | |
979 | gimple_stmt_iterator gsi; | |
980 | ||
981 | gsi = gsi_start_bb (bb); | |
982 | while (!gsi_end_p (gsi)) | |
983 | { | |
984 | gimple *stmt = gsi_stmt (gsi); | |
985 | info.stmt = stmt; | |
986 | struct walk_stmt_info wi; | |
987 | ||
988 | memset (&wi, 0, sizeof (wi)); | |
989 | info.modified = false; | |
990 | wi.info = &info; | |
991 | walk_gimple_op (stmt, ipa_simd_modify_stmt_ops, &wi); | |
992 | ||
993 | if (greturn *return_stmt = dyn_cast <greturn *> (stmt)) | |
994 | { | |
995 | tree retval = gimple_return_retval (return_stmt); | |
996 | if (!retval) | |
997 | { | |
998 | gsi_remove (&gsi, true); | |
999 | continue; | |
1000 | } | |
1001 | ||
1002 | /* Replace `return foo' with `retval_array[iter] = foo'. */ | |
1003 | tree ref = build4 (ARRAY_REF, TREE_TYPE (retval), | |
1004 | retval_array, iter, NULL, NULL); | |
1005 | stmt = gimple_build_assign (ref, retval); | |
1006 | gsi_replace (&gsi, stmt, true); | |
1007 | info.modified = true; | |
1008 | } | |
1009 | ||
1010 | if (info.modified) | |
1011 | { | |
1012 | update_stmt (stmt); | |
1013 | if (maybe_clean_eh_stmt (stmt)) | |
1014 | gimple_purge_dead_eh_edges (gimple_bb (stmt)); | |
1015 | } | |
1016 | gsi_next (&gsi); | |
1017 | } | |
1018 | } | |
1019 | } | |
1020 | ||
1021 | /* Helper function of simd_clone_adjust, return linear step addend | |
1022 | of Ith argument. */ | |
1023 | ||
1024 | static tree | |
1025 | simd_clone_linear_addend (struct cgraph_node *node, unsigned int i, | |
1026 | tree addtype, basic_block entry_bb) | |
1027 | { | |
1028 | tree ptype = NULL_TREE; | |
1029 | switch (node->simdclone->args[i].arg_type) | |
1030 | { | |
1031 | case SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP: | |
1032 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP: | |
1033 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_CONSTANT_STEP: | |
1034 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP: | |
1035 | return build_int_cst (addtype, node->simdclone->args[i].linear_step); | |
1036 | case SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP: | |
1037 | case SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP: | |
1038 | ptype = TREE_TYPE (node->simdclone->args[i].orig_arg); | |
1039 | break; | |
1040 | case SIMD_CLONE_ARG_TYPE_LINEAR_VAL_VARIABLE_STEP: | |
1041 | case SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP: | |
1042 | ptype = TREE_TYPE (TREE_TYPE (node->simdclone->args[i].orig_arg)); | |
1043 | break; | |
1044 | default: | |
1045 | gcc_unreachable (); | |
1046 | } | |
1047 | ||
1048 | unsigned int idx = node->simdclone->args[i].linear_step; | |
1049 | tree arg = node->simdclone->args[idx].orig_arg; | |
1050 | gcc_assert (is_gimple_reg_type (TREE_TYPE (arg))); | |
1051 | gimple_stmt_iterator gsi = gsi_after_labels (entry_bb); | |
1052 | gimple *g; | |
1053 | tree ret; | |
1054 | if (is_gimple_reg (arg)) | |
1055 | ret = get_or_create_ssa_default_def (cfun, arg); | |
1056 | else | |
1057 | { | |
1058 | g = gimple_build_assign (make_ssa_name (TREE_TYPE (arg)), arg); | |
1059 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1060 | ret = gimple_assign_lhs (g); | |
1061 | } | |
1062 | if (TREE_CODE (TREE_TYPE (arg)) == REFERENCE_TYPE) | |
1063 | { | |
1064 | g = gimple_build_assign (make_ssa_name (TREE_TYPE (TREE_TYPE (arg))), | |
1065 | build_simple_mem_ref (ret)); | |
1066 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1067 | ret = gimple_assign_lhs (g); | |
1068 | } | |
1069 | if (!useless_type_conversion_p (addtype, TREE_TYPE (ret))) | |
1070 | { | |
1071 | g = gimple_build_assign (make_ssa_name (addtype), NOP_EXPR, ret); | |
1072 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1073 | ret = gimple_assign_lhs (g); | |
1074 | } | |
1075 | if (POINTER_TYPE_P (ptype)) | |
1076 | { | |
1077 | tree size = TYPE_SIZE_UNIT (TREE_TYPE (ptype)); | |
1078 | if (size && TREE_CODE (size) == INTEGER_CST) | |
1079 | { | |
1080 | g = gimple_build_assign (make_ssa_name (addtype), MULT_EXPR, | |
1081 | ret, fold_convert (addtype, size)); | |
1082 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1083 | ret = gimple_assign_lhs (g); | |
1084 | } | |
1085 | } | |
1086 | return ret; | |
1087 | } | |
1088 | ||
1089 | /* Adjust the argument types in NODE to their appropriate vector | |
1090 | counterparts. */ | |
1091 | ||
1092 | static void | |
1093 | simd_clone_adjust (struct cgraph_node *node) | |
1094 | { | |
1095 | push_cfun (DECL_STRUCT_FUNCTION (node->decl)); | |
1096 | ||
1097 | targetm.simd_clone.adjust (node); | |
1098 | ||
1099 | tree retval = simd_clone_adjust_return_type (node); | |
1100 | ipa_parm_adjustment_vec adjustments | |
1101 | = simd_clone_adjust_argument_types (node); | |
1102 | ||
1103 | push_gimplify_context (); | |
1104 | ||
1105 | gimple_seq seq = simd_clone_init_simd_arrays (node, adjustments); | |
1106 | ||
1107 | /* Adjust all uses of vector arguments accordingly. Adjust all | |
1108 | return values accordingly. */ | |
1109 | tree iter = create_tmp_var (unsigned_type_node, "iter"); | |
1110 | tree iter1 = make_ssa_name (iter); | |
4ce71579 | 1111 | tree iter2 = NULL_TREE; |
60cbb674 | 1112 | ipa_simd_modify_function_body (node, adjustments, retval, iter1); |
d76815f4 | 1113 | adjustments.release (); |
60cbb674 TS |
1114 | |
1115 | /* Initialize the iteration variable. */ | |
1116 | basic_block entry_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
1117 | basic_block body_bb = split_block_after_labels (entry_bb)->dest; | |
1118 | gimple_stmt_iterator gsi = gsi_after_labels (entry_bb); | |
1119 | /* Insert the SIMD array and iv initialization at function | |
1120 | entry. */ | |
1121 | gsi_insert_seq_before (&gsi, seq, GSI_NEW_STMT); | |
1122 | ||
1123 | pop_gimplify_context (NULL); | |
1124 | ||
4ce71579 JJ |
1125 | gimple *g; |
1126 | basic_block incr_bb = NULL; | |
1127 | struct loop *loop = NULL; | |
1128 | ||
60cbb674 TS |
1129 | /* Create a new BB right before the original exit BB, to hold the |
1130 | iteration increment and the condition/branch. */ | |
4ce71579 JJ |
1131 | if (EDGE_COUNT (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds)) |
1132 | { | |
1133 | basic_block orig_exit = EDGE_PRED (EXIT_BLOCK_PTR_FOR_FN (cfun), 0)->src; | |
1134 | incr_bb = create_empty_bb (orig_exit); | |
e7a74006 | 1135 | incr_bb->count = profile_count::zero (); |
4ce71579 JJ |
1136 | add_bb_to_loop (incr_bb, body_bb->loop_father); |
1137 | /* The succ of orig_exit was EXIT_BLOCK_PTR_FOR_FN (cfun), with an empty | |
1138 | flag. Set it now to be a FALLTHRU_EDGE. */ | |
1139 | gcc_assert (EDGE_COUNT (orig_exit->succs) == 1); | |
1140 | EDGE_SUCC (orig_exit, 0)->flags |= EDGE_FALLTHRU; | |
1141 | for (unsigned i = 0; | |
1142 | i < EDGE_COUNT (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds); ++i) | |
1143 | { | |
1144 | edge e = EDGE_PRED (EXIT_BLOCK_PTR_FOR_FN (cfun), i); | |
1145 | redirect_edge_succ (e, incr_bb); | |
e7a74006 | 1146 | incr_bb->count += e->count (); |
4ce71579 JJ |
1147 | } |
1148 | } | |
1149 | else if (node->simdclone->inbranch) | |
1150 | { | |
1151 | incr_bb = create_empty_bb (entry_bb); | |
e7a74006 | 1152 | incr_bb->count = profile_count::zero (); |
4ce71579 JJ |
1153 | add_bb_to_loop (incr_bb, body_bb->loop_father); |
1154 | } | |
1155 | ||
1156 | if (incr_bb) | |
60cbb674 | 1157 | { |
357067f2 | 1158 | make_single_succ_edge (incr_bb, EXIT_BLOCK_PTR_FOR_FN (cfun), 0); |
4ce71579 JJ |
1159 | gsi = gsi_last_bb (incr_bb); |
1160 | iter2 = make_ssa_name (iter); | |
1161 | g = gimple_build_assign (iter2, PLUS_EXPR, iter1, | |
1162 | build_int_cst (unsigned_type_node, 1)); | |
1163 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
1164 | ||
1165 | /* Mostly annotate the loop for the vectorizer (the rest is done | |
1166 | below). */ | |
1167 | loop = alloc_loop (); | |
1168 | cfun->has_force_vectorize_loops = true; | |
1169 | loop->safelen = node->simdclone->simdlen; | |
1170 | loop->force_vectorize = true; | |
1171 | loop->header = body_bb; | |
60cbb674 | 1172 | } |
60cbb674 TS |
1173 | |
1174 | /* Branch around the body if the mask applies. */ | |
1175 | if (node->simdclone->inbranch) | |
1176 | { | |
4ce71579 | 1177 | gsi = gsi_last_bb (loop->header); |
60cbb674 TS |
1178 | tree mask_array |
1179 | = node->simdclone->args[node->simdclone->nargs - 1].simd_array; | |
1180 | tree mask; | |
1181 | if (node->simdclone->mask_mode != VOIDmode) | |
1182 | { | |
1183 | tree shift_cnt; | |
1184 | if (mask_array == NULL_TREE) | |
1185 | { | |
1186 | tree arg = node->simdclone->args[node->simdclone->nargs | |
1187 | - 1].vector_arg; | |
1188 | mask = get_or_create_ssa_default_def (cfun, arg); | |
1189 | shift_cnt = iter1; | |
1190 | } | |
1191 | else | |
1192 | { | |
1193 | tree maskt = TREE_TYPE (mask_array); | |
1194 | int c = tree_to_uhwi (TYPE_MAX_VALUE (TYPE_DOMAIN (maskt))); | |
1195 | c = node->simdclone->simdlen / (c + 1); | |
1196 | int s = exact_log2 (c); | |
1197 | gcc_assert (s > 0); | |
1198 | c--; | |
1199 | tree idx = make_ssa_name (TREE_TYPE (iter1)); | |
1200 | g = gimple_build_assign (idx, RSHIFT_EXPR, iter1, | |
1201 | build_int_cst (NULL_TREE, s)); | |
1202 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
1203 | mask = make_ssa_name (TREE_TYPE (TREE_TYPE (mask_array))); | |
1204 | tree aref = build4 (ARRAY_REF, | |
1205 | TREE_TYPE (TREE_TYPE (mask_array)), | |
1206 | mask_array, idx, NULL, NULL); | |
1207 | g = gimple_build_assign (mask, aref); | |
1208 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
1209 | shift_cnt = make_ssa_name (TREE_TYPE (iter1)); | |
1210 | g = gimple_build_assign (shift_cnt, BIT_AND_EXPR, iter1, | |
1211 | build_int_cst (TREE_TYPE (iter1), c)); | |
1212 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
1213 | } | |
1214 | g = gimple_build_assign (make_ssa_name (TREE_TYPE (mask)), | |
1215 | RSHIFT_EXPR, mask, shift_cnt); | |
1216 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
1217 | mask = gimple_assign_lhs (g); | |
1218 | g = gimple_build_assign (make_ssa_name (TREE_TYPE (mask)), | |
1219 | BIT_AND_EXPR, mask, | |
1220 | build_int_cst (TREE_TYPE (mask), 1)); | |
1221 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
1222 | mask = gimple_assign_lhs (g); | |
1223 | } | |
1224 | else | |
1225 | { | |
1226 | mask = make_ssa_name (TREE_TYPE (TREE_TYPE (mask_array))); | |
1227 | tree aref = build4 (ARRAY_REF, | |
1228 | TREE_TYPE (TREE_TYPE (mask_array)), | |
1229 | mask_array, iter1, NULL, NULL); | |
1230 | g = gimple_build_assign (mask, aref); | |
1231 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
b397965c | 1232 | int bitsize = GET_MODE_BITSIZE (SCALAR_TYPE_MODE (TREE_TYPE (aref))); |
60cbb674 TS |
1233 | if (!INTEGRAL_TYPE_P (TREE_TYPE (aref))) |
1234 | { | |
1235 | aref = build1 (VIEW_CONVERT_EXPR, | |
1236 | build_nonstandard_integer_type (bitsize, 0), | |
1237 | mask); | |
1238 | mask = make_ssa_name (TREE_TYPE (aref)); | |
1239 | g = gimple_build_assign (mask, aref); | |
1240 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
1241 | } | |
1242 | } | |
1243 | ||
1244 | g = gimple_build_cond (EQ_EXPR, mask, build_zero_cst (TREE_TYPE (mask)), | |
1245 | NULL, NULL); | |
1246 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
54cb4e20 TV |
1247 | edge e = make_edge (loop->header, incr_bb, EDGE_TRUE_VALUE); |
1248 | e->probability = profile_probability::unlikely ().guessed (); | |
e7a74006 | 1249 | incr_bb->count += e->count (); |
54cb4e20 TV |
1250 | edge fallthru = FALLTHRU_EDGE (loop->header); |
1251 | fallthru->flags = EDGE_FALSE_VALUE; | |
1252 | fallthru->probability = profile_probability::likely ().guessed (); | |
60cbb674 TS |
1253 | } |
1254 | ||
4ce71579 JJ |
1255 | basic_block latch_bb = NULL; |
1256 | basic_block new_exit_bb = NULL; | |
1257 | ||
60cbb674 | 1258 | /* Generate the condition. */ |
4ce71579 JJ |
1259 | if (incr_bb) |
1260 | { | |
1261 | gsi = gsi_last_bb (incr_bb); | |
1262 | g = gimple_build_cond (LT_EXPR, iter2, | |
1263 | build_int_cst (unsigned_type_node, | |
1264 | node->simdclone->simdlen), | |
1265 | NULL, NULL); | |
1266 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
1267 | edge e = split_block (incr_bb, gsi_stmt (gsi)); | |
1268 | latch_bb = e->dest; | |
1269 | new_exit_bb = split_block_after_labels (latch_bb)->dest; | |
1270 | loop->latch = latch_bb; | |
1271 | ||
1272 | redirect_edge_succ (FALLTHRU_EDGE (latch_bb), body_bb); | |
1273 | ||
357067f2 JH |
1274 | edge new_e = make_edge (incr_bb, new_exit_bb, EDGE_FALSE_VALUE); |
1275 | ||
1276 | /* FIXME: Do we need to distribute probabilities for the conditional? */ | |
1277 | new_e->probability = profile_probability::guessed_never (); | |
4ce71579 JJ |
1278 | /* The successor of incr_bb is already pointing to latch_bb; just |
1279 | change the flags. | |
1280 | make_edge (incr_bb, latch_bb, EDGE_TRUE_VALUE); */ | |
1281 | FALLTHRU_EDGE (incr_bb)->flags = EDGE_TRUE_VALUE; | |
1282 | } | |
60cbb674 TS |
1283 | |
1284 | gphi *phi = create_phi_node (iter1, body_bb); | |
1285 | edge preheader_edge = find_edge (entry_bb, body_bb); | |
4ce71579 | 1286 | edge latch_edge = NULL; |
60cbb674 TS |
1287 | add_phi_arg (phi, build_zero_cst (unsigned_type_node), preheader_edge, |
1288 | UNKNOWN_LOCATION); | |
4ce71579 | 1289 | if (incr_bb) |
60cbb674 | 1290 | { |
4ce71579 JJ |
1291 | latch_edge = single_succ_edge (latch_bb); |
1292 | add_phi_arg (phi, iter2, latch_edge, UNKNOWN_LOCATION); | |
1293 | ||
1294 | /* Generate the new return. */ | |
1295 | gsi = gsi_last_bb (new_exit_bb); | |
1296 | if (retval | |
1297 | && TREE_CODE (retval) == VIEW_CONVERT_EXPR | |
1298 | && TREE_CODE (TREE_OPERAND (retval, 0)) == RESULT_DECL) | |
1299 | retval = TREE_OPERAND (retval, 0); | |
1300 | else if (retval) | |
1301 | { | |
1302 | retval = build1 (VIEW_CONVERT_EXPR, | |
1303 | TREE_TYPE (TREE_TYPE (node->decl)), | |
1304 | retval); | |
1305 | retval = force_gimple_operand_gsi (&gsi, retval, true, NULL, | |
1306 | false, GSI_CONTINUE_LINKING); | |
1307 | } | |
1308 | g = gimple_build_return (retval); | |
1309 | gsi_insert_after (&gsi, g, GSI_CONTINUE_LINKING); | |
60cbb674 | 1310 | } |
60cbb674 TS |
1311 | |
1312 | /* Handle aligned clauses by replacing default defs of the aligned | |
1313 | uniform args with __builtin_assume_aligned (arg_N(D), alignment) | |
1314 | lhs. Handle linear by adding PHIs. */ | |
1315 | for (unsigned i = 0; i < node->simdclone->nargs; i++) | |
1316 | if (node->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_UNIFORM | |
1317 | && (TREE_ADDRESSABLE (node->simdclone->args[i].orig_arg) | |
1318 | || !is_gimple_reg_type | |
1319 | (TREE_TYPE (node->simdclone->args[i].orig_arg)))) | |
1320 | { | |
1321 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
1322 | if (is_gimple_reg_type (TREE_TYPE (orig_arg))) | |
1323 | iter1 = make_ssa_name (TREE_TYPE (orig_arg)); | |
1324 | else | |
1325 | { | |
1326 | iter1 = create_tmp_var_raw (TREE_TYPE (orig_arg)); | |
1327 | gimple_add_tmp_var (iter1); | |
1328 | } | |
1329 | gsi = gsi_after_labels (entry_bb); | |
1330 | g = gimple_build_assign (iter1, orig_arg); | |
1331 | gsi_insert_before (&gsi, g, GSI_NEW_STMT); | |
1332 | gsi = gsi_after_labels (body_bb); | |
1333 | g = gimple_build_assign (orig_arg, iter1); | |
1334 | gsi_insert_before (&gsi, g, GSI_NEW_STMT); | |
1335 | } | |
1336 | else if (node->simdclone->args[i].arg_type == SIMD_CLONE_ARG_TYPE_UNIFORM | |
1337 | && DECL_BY_REFERENCE (node->simdclone->args[i].orig_arg) | |
1338 | && TREE_CODE (TREE_TYPE (node->simdclone->args[i].orig_arg)) | |
1339 | == REFERENCE_TYPE | |
1340 | && TREE_ADDRESSABLE | |
1341 | (TREE_TYPE (TREE_TYPE (node->simdclone->args[i].orig_arg)))) | |
1342 | { | |
1343 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
1344 | tree def = ssa_default_def (cfun, orig_arg); | |
1345 | if (def && !has_zero_uses (def)) | |
1346 | { | |
1347 | iter1 = create_tmp_var_raw (TREE_TYPE (TREE_TYPE (orig_arg))); | |
1348 | gimple_add_tmp_var (iter1); | |
1349 | gsi = gsi_after_labels (entry_bb); | |
1350 | g = gimple_build_assign (iter1, build_simple_mem_ref (def)); | |
1351 | gsi_insert_before (&gsi, g, GSI_NEW_STMT); | |
1352 | gsi = gsi_after_labels (body_bb); | |
1353 | g = gimple_build_assign (build_simple_mem_ref (def), iter1); | |
1354 | gsi_insert_before (&gsi, g, GSI_NEW_STMT); | |
1355 | } | |
1356 | } | |
1357 | else if (node->simdclone->args[i].alignment | |
1358 | && node->simdclone->args[i].arg_type | |
1359 | == SIMD_CLONE_ARG_TYPE_UNIFORM | |
1360 | && (node->simdclone->args[i].alignment | |
1361 | & (node->simdclone->args[i].alignment - 1)) == 0 | |
1362 | && TREE_CODE (TREE_TYPE (node->simdclone->args[i].orig_arg)) | |
1363 | == POINTER_TYPE) | |
1364 | { | |
1365 | unsigned int alignment = node->simdclone->args[i].alignment; | |
1366 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
1367 | tree def = ssa_default_def (cfun, orig_arg); | |
1368 | if (def && !has_zero_uses (def)) | |
1369 | { | |
1370 | tree fn = builtin_decl_explicit (BUILT_IN_ASSUME_ALIGNED); | |
1371 | gimple_seq seq = NULL; | |
1372 | bool need_cvt = false; | |
1373 | gcall *call | |
1374 | = gimple_build_call (fn, 2, def, size_int (alignment)); | |
1375 | g = call; | |
1376 | if (!useless_type_conversion_p (TREE_TYPE (orig_arg), | |
1377 | ptr_type_node)) | |
1378 | need_cvt = true; | |
1379 | tree t = make_ssa_name (need_cvt ? ptr_type_node : orig_arg); | |
1380 | gimple_call_set_lhs (g, t); | |
1381 | gimple_seq_add_stmt_without_update (&seq, g); | |
1382 | if (need_cvt) | |
1383 | { | |
1384 | t = make_ssa_name (orig_arg); | |
1385 | g = gimple_build_assign (t, NOP_EXPR, gimple_call_lhs (g)); | |
1386 | gimple_seq_add_stmt_without_update (&seq, g); | |
1387 | } | |
1388 | gsi_insert_seq_on_edge_immediate | |
1389 | (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)), seq); | |
1390 | ||
1391 | entry_bb = single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
1392 | int freq = compute_call_stmt_bb_frequency (current_function_decl, | |
1393 | entry_bb); | |
1394 | node->create_edge (cgraph_node::get_create (fn), | |
1395 | call, entry_bb->count, freq); | |
1396 | ||
1397 | imm_use_iterator iter; | |
1398 | use_operand_p use_p; | |
1399 | gimple *use_stmt; | |
1400 | tree repl = gimple_get_lhs (g); | |
1401 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, def) | |
1402 | if (is_gimple_debug (use_stmt) || use_stmt == call) | |
1403 | continue; | |
1404 | else | |
1405 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
1406 | SET_USE (use_p, repl); | |
1407 | } | |
1408 | } | |
1409 | else if ((node->simdclone->args[i].arg_type | |
1410 | == SIMD_CLONE_ARG_TYPE_LINEAR_CONSTANT_STEP) | |
1411 | || (node->simdclone->args[i].arg_type | |
1412 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_CONSTANT_STEP) | |
1413 | || (node->simdclone->args[i].arg_type | |
1414 | == SIMD_CLONE_ARG_TYPE_LINEAR_VARIABLE_STEP) | |
1415 | || (node->simdclone->args[i].arg_type | |
1416 | == SIMD_CLONE_ARG_TYPE_LINEAR_REF_VARIABLE_STEP)) | |
1417 | { | |
1418 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
1419 | gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
1420 | || POINTER_TYPE_P (TREE_TYPE (orig_arg))); | |
1421 | tree def = NULL_TREE; | |
1422 | if (TREE_ADDRESSABLE (orig_arg)) | |
1423 | { | |
1424 | def = make_ssa_name (TREE_TYPE (orig_arg)); | |
1425 | iter1 = make_ssa_name (TREE_TYPE (orig_arg)); | |
4ce71579 JJ |
1426 | if (incr_bb) |
1427 | iter2 = make_ssa_name (TREE_TYPE (orig_arg)); | |
60cbb674 TS |
1428 | gsi = gsi_after_labels (entry_bb); |
1429 | g = gimple_build_assign (def, orig_arg); | |
1430 | gsi_insert_before (&gsi, g, GSI_NEW_STMT); | |
1431 | } | |
1432 | else | |
1433 | { | |
1434 | def = ssa_default_def (cfun, orig_arg); | |
1435 | if (!def || has_zero_uses (def)) | |
1436 | def = NULL_TREE; | |
1437 | else | |
1438 | { | |
1439 | iter1 = make_ssa_name (orig_arg); | |
4ce71579 JJ |
1440 | if (incr_bb) |
1441 | iter2 = make_ssa_name (orig_arg); | |
60cbb674 TS |
1442 | } |
1443 | } | |
1444 | if (def) | |
1445 | { | |
1446 | phi = create_phi_node (iter1, body_bb); | |
1447 | add_phi_arg (phi, def, preheader_edge, UNKNOWN_LOCATION); | |
4ce71579 JJ |
1448 | if (incr_bb) |
1449 | { | |
1450 | add_phi_arg (phi, iter2, latch_edge, UNKNOWN_LOCATION); | |
1451 | enum tree_code code = INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
1452 | ? PLUS_EXPR : POINTER_PLUS_EXPR; | |
1453 | tree addtype = INTEGRAL_TYPE_P (TREE_TYPE (orig_arg)) | |
1454 | ? TREE_TYPE (orig_arg) : sizetype; | |
1455 | tree addcst = simd_clone_linear_addend (node, i, addtype, | |
1456 | entry_bb); | |
1457 | gsi = gsi_last_bb (incr_bb); | |
1458 | g = gimple_build_assign (iter2, code, iter1, addcst); | |
1459 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1460 | } | |
60cbb674 TS |
1461 | |
1462 | imm_use_iterator iter; | |
1463 | use_operand_p use_p; | |
1464 | gimple *use_stmt; | |
1465 | if (TREE_ADDRESSABLE (orig_arg)) | |
1466 | { | |
1467 | gsi = gsi_after_labels (body_bb); | |
1468 | g = gimple_build_assign (orig_arg, iter1); | |
1469 | gsi_insert_before (&gsi, g, GSI_NEW_STMT); | |
1470 | } | |
1471 | else | |
1472 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, def) | |
1473 | if (use_stmt == phi) | |
1474 | continue; | |
1475 | else | |
1476 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
1477 | SET_USE (use_p, iter1); | |
1478 | } | |
1479 | } | |
1480 | else if (node->simdclone->args[i].arg_type | |
1481 | == SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_CONSTANT_STEP | |
1482 | || (node->simdclone->args[i].arg_type | |
1483 | == SIMD_CLONE_ARG_TYPE_LINEAR_UVAL_VARIABLE_STEP)) | |
1484 | { | |
1485 | tree orig_arg = node->simdclone->args[i].orig_arg; | |
1486 | tree def = ssa_default_def (cfun, orig_arg); | |
1487 | gcc_assert (!TREE_ADDRESSABLE (orig_arg) | |
1488 | && TREE_CODE (TREE_TYPE (orig_arg)) == REFERENCE_TYPE); | |
1489 | if (def && !has_zero_uses (def)) | |
1490 | { | |
1491 | tree rtype = TREE_TYPE (TREE_TYPE (orig_arg)); | |
1492 | iter1 = make_ssa_name (orig_arg); | |
4ce71579 JJ |
1493 | if (incr_bb) |
1494 | iter2 = make_ssa_name (orig_arg); | |
60cbb674 TS |
1495 | tree iter3 = make_ssa_name (rtype); |
1496 | tree iter4 = make_ssa_name (rtype); | |
4ce71579 | 1497 | tree iter5 = incr_bb ? make_ssa_name (rtype) : NULL_TREE; |
60cbb674 TS |
1498 | gsi = gsi_after_labels (entry_bb); |
1499 | gimple *load | |
1500 | = gimple_build_assign (iter3, build_simple_mem_ref (def)); | |
1501 | gsi_insert_before (&gsi, load, GSI_NEW_STMT); | |
1502 | ||
1503 | tree array = node->simdclone->args[i].simd_array; | |
1504 | TREE_ADDRESSABLE (array) = 1; | |
1505 | tree ptr = build_fold_addr_expr (array); | |
1506 | phi = create_phi_node (iter1, body_bb); | |
1507 | add_phi_arg (phi, ptr, preheader_edge, UNKNOWN_LOCATION); | |
4ce71579 JJ |
1508 | if (incr_bb) |
1509 | { | |
1510 | add_phi_arg (phi, iter2, latch_edge, UNKNOWN_LOCATION); | |
1511 | g = gimple_build_assign (iter2, POINTER_PLUS_EXPR, iter1, | |
1512 | TYPE_SIZE_UNIT (TREE_TYPE (iter3))); | |
1513 | gsi = gsi_last_bb (incr_bb); | |
1514 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1515 | } | |
60cbb674 TS |
1516 | |
1517 | phi = create_phi_node (iter4, body_bb); | |
1518 | add_phi_arg (phi, iter3, preheader_edge, UNKNOWN_LOCATION); | |
4ce71579 JJ |
1519 | if (incr_bb) |
1520 | { | |
1521 | add_phi_arg (phi, iter5, latch_edge, UNKNOWN_LOCATION); | |
1522 | enum tree_code code = INTEGRAL_TYPE_P (TREE_TYPE (iter3)) | |
1523 | ? PLUS_EXPR : POINTER_PLUS_EXPR; | |
1524 | tree addtype = INTEGRAL_TYPE_P (TREE_TYPE (iter3)) | |
1525 | ? TREE_TYPE (iter3) : sizetype; | |
1526 | tree addcst = simd_clone_linear_addend (node, i, addtype, | |
1527 | entry_bb); | |
1528 | g = gimple_build_assign (iter5, code, iter4, addcst); | |
1529 | gsi = gsi_last_bb (incr_bb); | |
1530 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1531 | } | |
60cbb674 TS |
1532 | |
1533 | g = gimple_build_assign (build_simple_mem_ref (iter1), iter4); | |
1534 | gsi = gsi_after_labels (body_bb); | |
1535 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1536 | ||
1537 | imm_use_iterator iter; | |
1538 | use_operand_p use_p; | |
1539 | gimple *use_stmt; | |
1540 | FOR_EACH_IMM_USE_STMT (use_stmt, iter, def) | |
1541 | if (use_stmt == load) | |
1542 | continue; | |
1543 | else | |
1544 | FOR_EACH_IMM_USE_ON_STMT (use_p, iter) | |
1545 | SET_USE (use_p, iter1); | |
1546 | ||
4ce71579 | 1547 | if (!TYPE_READONLY (rtype) && incr_bb) |
60cbb674 TS |
1548 | { |
1549 | tree v = make_ssa_name (rtype); | |
1550 | tree aref = build4 (ARRAY_REF, rtype, array, | |
1551 | size_zero_node, NULL_TREE, | |
1552 | NULL_TREE); | |
1553 | gsi = gsi_after_labels (new_exit_bb); | |
1554 | g = gimple_build_assign (v, aref); | |
1555 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1556 | g = gimple_build_assign (build_simple_mem_ref (def), v); | |
1557 | gsi_insert_before (&gsi, g, GSI_SAME_STMT); | |
1558 | } | |
1559 | } | |
1560 | } | |
1561 | ||
1562 | calculate_dominance_info (CDI_DOMINATORS); | |
4ce71579 JJ |
1563 | if (loop) |
1564 | add_loop (loop, loop->header->loop_father); | |
60cbb674 TS |
1565 | update_ssa (TODO_update_ssa); |
1566 | ||
1567 | pop_cfun (); | |
1568 | } | |
1569 | ||
1570 | /* If the function in NODE is tagged as an elemental SIMD function, | |
1571 | create the appropriate SIMD clones. */ | |
1572 | ||
1573 | static void | |
1574 | expand_simd_clones (struct cgraph_node *node) | |
1575 | { | |
1576 | tree attr = lookup_attribute ("omp declare simd", | |
1577 | DECL_ATTRIBUTES (node->decl)); | |
1578 | if (attr == NULL_TREE | |
1579 | || node->global.inlined_to | |
1580 | || lookup_attribute ("noclone", DECL_ATTRIBUTES (node->decl))) | |
1581 | return; | |
1582 | ||
1583 | /* Ignore | |
1584 | #pragma omp declare simd | |
1585 | extern int foo (); | |
1586 | in C, there we don't know the argument types at all. */ | |
1587 | if (!node->definition | |
1588 | && TYPE_ARG_TYPES (TREE_TYPE (node->decl)) == NULL_TREE) | |
1589 | return; | |
1590 | ||
1591 | /* Call this before creating clone_info, as it might ggc_collect. */ | |
1592 | if (node->definition && node->has_gimple_body_p ()) | |
1593 | node->get_body (); | |
1594 | ||
1595 | do | |
1596 | { | |
1597 | /* Start with parsing the "omp declare simd" attribute(s). */ | |
1598 | bool inbranch_clause_specified; | |
1599 | struct cgraph_simd_clone *clone_info | |
1600 | = simd_clone_clauses_extract (node, TREE_VALUE (attr), | |
1601 | &inbranch_clause_specified); | |
1602 | if (clone_info == NULL) | |
1603 | continue; | |
1604 | ||
1605 | int orig_simdlen = clone_info->simdlen; | |
1606 | tree base_type = simd_clone_compute_base_data_type (node, clone_info); | |
1607 | /* The target can return 0 (no simd clones should be created), | |
1608 | 1 (just one ISA of simd clones should be created) or higher | |
1609 | count of ISA variants. In that case, clone_info is initialized | |
1610 | for the first ISA variant. */ | |
1611 | int count | |
1612 | = targetm.simd_clone.compute_vecsize_and_simdlen (node, clone_info, | |
1613 | base_type, 0); | |
1614 | if (count == 0) | |
1615 | continue; | |
1616 | ||
1617 | /* Loop over all COUNT ISA variants, and if !INBRANCH_CLAUSE_SPECIFIED, | |
1618 | also create one inbranch and one !inbranch clone of it. */ | |
1619 | for (int i = 0; i < count * 2; i++) | |
1620 | { | |
1621 | struct cgraph_simd_clone *clone = clone_info; | |
1622 | if (inbranch_clause_specified && (i & 1) != 0) | |
1623 | continue; | |
1624 | ||
1625 | if (i != 0) | |
1626 | { | |
1627 | clone = simd_clone_struct_alloc (clone_info->nargs | |
1628 | + ((i & 1) != 0)); | |
1629 | simd_clone_struct_copy (clone, clone_info); | |
1630 | /* Undo changes targetm.simd_clone.compute_vecsize_and_simdlen | |
1631 | and simd_clone_adjust_argument_types did to the first | |
1632 | clone's info. */ | |
1633 | clone->nargs -= clone_info->inbranch; | |
1634 | clone->simdlen = orig_simdlen; | |
1635 | /* And call the target hook again to get the right ISA. */ | |
1636 | targetm.simd_clone.compute_vecsize_and_simdlen (node, clone, | |
1637 | base_type, | |
1638 | i / 2); | |
1639 | if ((i & 1) != 0) | |
1640 | clone->inbranch = 1; | |
1641 | } | |
1642 | ||
1643 | /* simd_clone_mangle might fail if such a clone has been created | |
1644 | already. */ | |
1645 | tree id = simd_clone_mangle (node, clone); | |
1646 | if (id == NULL_TREE) | |
1647 | continue; | |
1648 | ||
1649 | /* Only when we are sure we want to create the clone actually | |
1650 | clone the function (or definitions) or create another | |
1651 | extern FUNCTION_DECL (for prototypes without definitions). */ | |
1652 | struct cgraph_node *n = simd_clone_create (node); | |
1653 | if (n == NULL) | |
1654 | continue; | |
1655 | ||
1656 | n->simdclone = clone; | |
1657 | clone->origin = node; | |
1658 | clone->next_clone = NULL; | |
1659 | if (node->simd_clones == NULL) | |
1660 | { | |
1661 | clone->prev_clone = n; | |
1662 | node->simd_clones = n; | |
1663 | } | |
1664 | else | |
1665 | { | |
1666 | clone->prev_clone = node->simd_clones->simdclone->prev_clone; | |
1667 | clone->prev_clone->simdclone->next_clone = n; | |
1668 | node->simd_clones->simdclone->prev_clone = n; | |
1669 | } | |
1670 | symtab->change_decl_assembler_name (n->decl, id); | |
1671 | /* And finally adjust the return type, parameters and for | |
1672 | definitions also function body. */ | |
1673 | if (node->definition) | |
1674 | simd_clone_adjust (n); | |
1675 | else | |
1676 | { | |
1677 | simd_clone_adjust_return_type (n); | |
1678 | simd_clone_adjust_argument_types (n); | |
1679 | } | |
1680 | } | |
1681 | } | |
1682 | while ((attr = lookup_attribute ("omp declare simd", TREE_CHAIN (attr)))); | |
1683 | } | |
1684 | ||
1685 | /* Entry point for IPA simd clone creation pass. */ | |
1686 | ||
1687 | static unsigned int | |
1688 | ipa_omp_simd_clone (void) | |
1689 | { | |
1690 | struct cgraph_node *node; | |
1691 | FOR_EACH_FUNCTION (node) | |
1692 | expand_simd_clones (node); | |
1693 | return 0; | |
1694 | } | |
1695 | ||
1696 | namespace { | |
1697 | ||
1698 | const pass_data pass_data_omp_simd_clone = | |
1699 | { | |
1700 | SIMPLE_IPA_PASS, /* type */ | |
1701 | "simdclone", /* name */ | |
d03958cf | 1702 | OPTGROUP_OMP, /* optinfo_flags */ |
60cbb674 TS |
1703 | TV_NONE, /* tv_id */ |
1704 | ( PROP_ssa | PROP_cfg ), /* properties_required */ | |
1705 | 0, /* properties_provided */ | |
1706 | 0, /* properties_destroyed */ | |
1707 | 0, /* todo_flags_start */ | |
1708 | 0, /* todo_flags_finish */ | |
1709 | }; | |
1710 | ||
1711 | class pass_omp_simd_clone : public simple_ipa_opt_pass | |
1712 | { | |
1713 | public: | |
1714 | pass_omp_simd_clone(gcc::context *ctxt) | |
1715 | : simple_ipa_opt_pass(pass_data_omp_simd_clone, ctxt) | |
1716 | {} | |
1717 | ||
1718 | /* opt_pass methods: */ | |
1719 | virtual bool gate (function *); | |
1720 | virtual unsigned int execute (function *) { return ipa_omp_simd_clone (); } | |
1721 | }; | |
1722 | ||
1723 | bool | |
1724 | pass_omp_simd_clone::gate (function *) | |
1725 | { | |
1726 | return targetm.simd_clone.compute_vecsize_and_simdlen != NULL; | |
1727 | } | |
1728 | ||
1729 | } // anon namespace | |
1730 | ||
1731 | simple_ipa_opt_pass * | |
1732 | make_pass_omp_simd_clone (gcc::context *ctxt) | |
1733 | { | |
1734 | return new pass_omp_simd_clone (ctxt); | |
1735 | } |