1 | /* Bits of OpenMP and OpenACC handling that is specific to device offloading |
2 | and a lowering pass for OpenACC device directives. |
3 | |
4 | Copyright (C) 2005-2023 Free Software Foundation, Inc. |
5 | |
6 | This file is part of GCC. |
7 | |
8 | GCC is free software; you can redistribute it and/or modify it under |
9 | the terms of the GNU General Public License as published by the Free |
10 | Software Foundation; either version 3, or (at your option) any later |
11 | version. |
12 | |
13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
16 | for more details. |
17 | |
18 | You should have received a copy of the GNU General Public License |
19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ |
21 | |
22 | #include "config.h" |
23 | #include "system.h" |
24 | #include "coretypes.h" |
25 | #include "backend.h" |
26 | #include "target.h" |
27 | #include "tree.h" |
28 | #include "gimple.h" |
29 | #include "tree-pass.h" |
30 | #include "ssa.h" |
31 | #include "cgraph.h" |
32 | #include "pretty-print.h" |
33 | #include "diagnostic-core.h" |
34 | #include "fold-const.h" |
35 | #include "internal-fn.h" |
36 | #include "langhooks.h" |
37 | #include "gimplify.h" |
38 | #include "gimple-iterator.h" |
39 | #include "gimplify-me.h" |
40 | #include "gimple-walk.h" |
41 | #include "tree-cfg.h" |
42 | #include "tree-into-ssa.h" |
43 | #include "tree-nested.h" |
44 | #include "stor-layout.h" |
45 | #include "common/common-target.h" |
46 | #include "omp-general.h" |
47 | #include "omp-offload.h" |
48 | #include "lto-section-names.h" |
49 | #include "gomp-constants.h" |
50 | #include "gimple-pretty-print.h" |
51 | #include "intl.h" |
52 | #include "stringpool.h" |
53 | #include "attribs.h" |
54 | #include "cfgloop.h" |
55 | #include "context.h" |
56 | #include "convert.h" |
57 | #include "opts.h" |
58 | |
59 | /* Describe the OpenACC looping structure of a function. The entire |
60 | function is held in a 'NULL' loop. */ |
61 | |
62 | struct oacc_loop |
63 | { |
64 | oacc_loop *parent; /* Containing loop. */ |
65 | |
66 | oacc_loop *child; /* First inner loop. */ |
67 | |
68 | oacc_loop *sibling; /* Next loop within same parent. */ |
69 | |
70 | location_t loc; /* Location of the loop start. */ |
71 | |
72 | gcall *marker; /* Initial head marker. */ |
73 | |
74 | gcall *heads[GOMP_DIM_MAX]; /* Head marker functions. */ |
75 | gcall *tails[GOMP_DIM_MAX]; /* Tail marker functions. */ |
76 | |
77 | tree routine; /* Pseudo-loop enclosing a routine. */ |
78 | |
79 | unsigned mask; /* Partitioning mask. */ |
80 | unsigned e_mask; /* Partitioning of element loops (when tiling). */ |
81 | unsigned inner; /* Partitioning of inner loops. */ |
82 | unsigned flags; /* Partitioning flags. */ |
83 | vec<gcall *> ifns; /* Contained loop abstraction functions. */ |
84 | tree chunk_size; /* Chunk size. */ |
85 | gcall *head_end; /* Final marker of head sequence. */ |
86 | }; |
87 | |
88 | /* Holds offload tables with decls. */ |
89 | vec<tree, va_gc> *offload_funcs, *offload_vars, *offload_ind_funcs; |
90 | |
91 | /* Return level at which oacc routine may spawn a partitioned loop, or |
92 | -1 if it is not a routine (i.e. is an offload fn). */ |
93 | |
94 | int |
95 | oacc_fn_attrib_level (tree attr) |
96 | { |
97 | tree pos = TREE_VALUE (attr); |
98 | |
99 | if (!TREE_PURPOSE (pos)) |
100 | return -1; |
101 | |
102 | int ix = 0; |
103 | for (ix = 0; ix != GOMP_DIM_MAX; |
104 | ix++, pos = TREE_CHAIN (pos)) |
105 | if (!integer_zerop (TREE_PURPOSE (pos))) |
106 | break; |
107 | |
108 | return ix; |
109 | } |
110 | |
111 | /* Helper function for omp_finish_file routine. Takes decls from V_DECLS and |
112 | adds their addresses and sizes to constructor-vector V_CTOR. */ |
113 | |
114 | static void |
115 | add_decls_addresses_to_decl_constructor (vec<tree, va_gc> *v_decls, |
116 | vec<constructor_elt, va_gc> *v_ctor) |
117 | { |
118 | unsigned len = vec_safe_length (v: v_decls); |
119 | for (unsigned i = 0; i < len; i++) |
120 | { |
121 | tree it = (*v_decls)[i]; |
122 | bool is_var = VAR_P (it); |
123 | bool is_link_var |
124 | = is_var |
125 | #ifdef ACCEL_COMPILER |
126 | && DECL_HAS_VALUE_EXPR_P (it) |
127 | #endif |
128 | && lookup_attribute (attr_name: "omp declare target link" , DECL_ATTRIBUTES (it)); |
129 | |
130 | /* See also omp_finish_file and output_offload_tables in lto-cgraph.cc. */ |
131 | if (!in_lto_p && !symtab_node::get (decl: it)) |
132 | continue; |
133 | |
134 | tree size = NULL_TREE; |
135 | if (is_var) |
136 | size = fold_convert (const_ptr_type_node, DECL_SIZE_UNIT (it)); |
137 | |
138 | tree addr; |
139 | if (!is_link_var) |
140 | addr = build_fold_addr_expr (it); |
141 | else |
142 | { |
143 | #ifdef ACCEL_COMPILER |
144 | /* For "omp declare target link" vars add address of the pointer to |
145 | the target table, instead of address of the var. */ |
146 | tree value_expr = DECL_VALUE_EXPR (it); |
147 | tree link_ptr_decl = TREE_OPERAND (value_expr, 0); |
148 | varpool_node::finalize_decl (link_ptr_decl); |
149 | addr = build_fold_addr_expr (link_ptr_decl); |
150 | #else |
151 | addr = build_fold_addr_expr (it); |
152 | #endif |
153 | |
154 | /* Most significant bit of the size marks "omp declare target link" |
155 | vars in host and target tables. */ |
156 | unsigned HOST_WIDE_INT isize = tree_to_uhwi (size); |
157 | isize |= 1ULL << (int_size_in_bytes (const_ptr_type_node) |
158 | * BITS_PER_UNIT - 1); |
159 | size = wide_int_to_tree (const_ptr_type_node, cst: isize); |
160 | } |
161 | |
162 | CONSTRUCTOR_APPEND_ELT (v_ctor, NULL_TREE, addr); |
163 | if (is_var) |
164 | CONSTRUCTOR_APPEND_ELT (v_ctor, NULL_TREE, size); |
165 | } |
166 | } |
167 | |
168 | /* Return true if DECL is a function for which its references should be |
169 | analyzed. */ |
170 | |
171 | static bool |
172 | omp_declare_target_fn_p (tree decl) |
173 | { |
174 | return (TREE_CODE (decl) == FUNCTION_DECL |
175 | && lookup_attribute (attr_name: "omp declare target" , DECL_ATTRIBUTES (decl)) |
176 | && !lookup_attribute (attr_name: "omp declare target host" , |
177 | DECL_ATTRIBUTES (decl)) |
178 | && (!flag_openacc |
179 | || oacc_get_fn_attrib (fn: decl) == NULL_TREE)); |
180 | } |
181 | |
182 | /* Return true if DECL Is a variable for which its initializer references |
183 | should be analyzed. */ |
184 | |
185 | static bool |
186 | omp_declare_target_var_p (tree decl) |
187 | { |
188 | return (VAR_P (decl) |
189 | && lookup_attribute (attr_name: "omp declare target" , DECL_ATTRIBUTES (decl)) |
190 | && !lookup_attribute (attr_name: "omp declare target link" , |
191 | DECL_ATTRIBUTES (decl))); |
192 | } |
193 | |
194 | /* Helper function for omp_discover_implicit_declare_target, called through |
195 | walk_tree. Mark referenced FUNCTION_DECLs implicitly as |
196 | declare target to. */ |
197 | |
198 | static tree |
199 | omp_discover_declare_target_tgt_fn_r (tree *tp, int *walk_subtrees, void *data) |
200 | { |
201 | if (TREE_CODE (*tp) == CALL_EXPR |
202 | && CALL_EXPR_FN (*tp) |
203 | && TREE_CODE (CALL_EXPR_FN (*tp)) == ADDR_EXPR |
204 | && TREE_CODE (TREE_OPERAND (CALL_EXPR_FN (*tp), 0)) == FUNCTION_DECL |
205 | && lookup_attribute (attr_name: "omp declare variant base" , |
206 | DECL_ATTRIBUTES (TREE_OPERAND (CALL_EXPR_FN (*tp), |
207 | 0)))) |
208 | { |
209 | tree fn = TREE_OPERAND (CALL_EXPR_FN (*tp), 0); |
210 | for (tree attr = DECL_ATTRIBUTES (fn); attr; attr = TREE_CHAIN (attr)) |
211 | { |
212 | attr = lookup_attribute (attr_name: "omp declare variant base" , list: attr); |
213 | if (attr == NULL_TREE) |
214 | break; |
215 | tree purpose = TREE_PURPOSE (TREE_VALUE (attr)); |
216 | if (TREE_CODE (purpose) == FUNCTION_DECL) |
217 | omp_discover_declare_target_tgt_fn_r (tp: &purpose, walk_subtrees, data); |
218 | } |
219 | } |
220 | else if (TREE_CODE (*tp) == FUNCTION_DECL) |
221 | { |
222 | tree decl = *tp; |
223 | tree id = get_identifier ("omp declare target" ); |
224 | symtab_node *node = symtab_node::get (decl: *tp); |
225 | if (node != NULL) |
226 | { |
227 | while (node->alias_target |
228 | && TREE_CODE (node->alias_target) == FUNCTION_DECL) |
229 | { |
230 | if (!omp_declare_target_fn_p (decl: node->decl) |
231 | && !lookup_attribute (attr_name: "omp declare target host" , |
232 | DECL_ATTRIBUTES (node->decl))) |
233 | { |
234 | node->offloadable = 1; |
235 | DECL_ATTRIBUTES (node->decl) |
236 | = tree_cons (id, NULL_TREE, DECL_ATTRIBUTES (node->decl)); |
237 | } |
238 | node = symtab_node::get (decl: node->alias_target); |
239 | } |
240 | symtab_node *new_node = node->ultimate_alias_target (); |
241 | decl = new_node->decl; |
242 | while (node != new_node) |
243 | { |
244 | if (!omp_declare_target_fn_p (decl: node->decl) |
245 | && !lookup_attribute (attr_name: "omp declare target host" , |
246 | DECL_ATTRIBUTES (node->decl))) |
247 | { |
248 | node->offloadable = 1; |
249 | DECL_ATTRIBUTES (node->decl) |
250 | = tree_cons (id, NULL_TREE, DECL_ATTRIBUTES (node->decl)); |
251 | } |
252 | gcc_assert (node->alias && node->analyzed); |
253 | node = node->get_alias_target (); |
254 | } |
255 | node->offloadable = 1; |
256 | if (ENABLE_OFFLOADING) |
257 | g->have_offload = true; |
258 | } |
259 | if (omp_declare_target_fn_p (decl) |
260 | || lookup_attribute (attr_name: "omp declare target host" , |
261 | DECL_ATTRIBUTES (decl))) |
262 | return NULL_TREE; |
263 | |
264 | if (!DECL_EXTERNAL (decl) && DECL_SAVED_TREE (decl)) |
265 | ((vec<tree> *) data)->safe_push (obj: decl); |
266 | DECL_ATTRIBUTES (decl) = tree_cons (id, NULL_TREE, |
267 | DECL_ATTRIBUTES (decl)); |
268 | } |
269 | else if (TYPE_P (*tp)) |
270 | *walk_subtrees = 0; |
271 | else if (TREE_CODE (*tp) == OMP_TARGET) |
272 | { |
273 | tree c = omp_find_clause (OMP_CLAUSES (*tp), kind: OMP_CLAUSE_DEVICE); |
274 | if (c && OMP_CLAUSE_DEVICE_ANCESTOR (c)) |
275 | *walk_subtrees = 0; |
276 | } |
277 | return NULL_TREE; |
278 | } |
279 | |
280 | /* Similarly, but ignore references outside of OMP_TARGET regions. */ |
281 | |
282 | static tree |
283 | omp_discover_declare_target_fn_r (tree *tp, int *walk_subtrees, void *data) |
284 | { |
285 | if (TREE_CODE (*tp) == OMP_TARGET) |
286 | { |
287 | tree c = omp_find_clause (OMP_CLAUSES (*tp), kind: OMP_CLAUSE_DEVICE); |
288 | if (!c || !OMP_CLAUSE_DEVICE_ANCESTOR (c)) |
289 | walk_tree_without_duplicates (&OMP_TARGET_BODY (*tp), |
290 | omp_discover_declare_target_tgt_fn_r, |
291 | data); |
292 | *walk_subtrees = 0; |
293 | } |
294 | else if (TYPE_P (*tp)) |
295 | *walk_subtrees = 0; |
296 | return NULL_TREE; |
297 | } |
298 | |
299 | /* Helper function for omp_discover_implicit_declare_target, called through |
300 | walk_tree. Mark referenced FUNCTION_DECLs implicitly as |
301 | declare target to. */ |
302 | |
303 | static tree |
304 | omp_discover_declare_target_var_r (tree *tp, int *walk_subtrees, void *data) |
305 | { |
306 | if (TREE_CODE (*tp) == FUNCTION_DECL) |
307 | return omp_discover_declare_target_tgt_fn_r (tp, walk_subtrees, data); |
308 | else if (VAR_P (*tp) |
309 | && is_global_var (t: *tp) |
310 | && !omp_declare_target_var_p (decl: *tp)) |
311 | { |
312 | tree id = get_identifier ("omp declare target" ); |
313 | if (lookup_attribute (attr_name: "omp declare target link" , DECL_ATTRIBUTES (*tp))) |
314 | { |
315 | error_at (DECL_SOURCE_LOCATION (*tp), |
316 | "%qD specified both in declare target %<link%> and " |
317 | "implicitly in %<to%> clauses" , *tp); |
318 | DECL_ATTRIBUTES (*tp) |
319 | = remove_attribute ("omp declare target link" , DECL_ATTRIBUTES (*tp)); |
320 | } |
321 | if (TREE_STATIC (*tp) && lang_hooks.decls.omp_get_decl_init (*tp)) |
322 | ((vec<tree> *) data)->safe_push (obj: *tp); |
323 | DECL_ATTRIBUTES (*tp) = tree_cons (id, NULL_TREE, DECL_ATTRIBUTES (*tp)); |
324 | symtab_node *node = symtab_node::get (decl: *tp); |
325 | if (node != NULL && !node->offloadable) |
326 | { |
327 | node->offloadable = 1; |
328 | if (ENABLE_OFFLOADING) |
329 | { |
330 | g->have_offload = true; |
331 | if (is_a <varpool_node *> (p: node)) |
332 | vec_safe_push (v&: offload_vars, obj: node->decl); |
333 | } |
334 | } |
335 | } |
336 | else if (TYPE_P (*tp)) |
337 | *walk_subtrees = 0; |
338 | return NULL_TREE; |
339 | } |
340 | |
341 | /* Perform the OpenMP implicit declare target to discovery. */ |
342 | |
343 | void |
344 | omp_discover_implicit_declare_target (void) |
345 | { |
346 | cgraph_node *node; |
347 | varpool_node *vnode; |
348 | auto_vec<tree> worklist; |
349 | |
350 | FOR_EACH_DEFINED_FUNCTION (node) |
351 | if (DECL_SAVED_TREE (node->decl)) |
352 | { |
353 | struct cgraph_node *cgn; |
354 | if (lookup_attribute (attr_name: "omp declare target indirect" , |
355 | DECL_ATTRIBUTES (node->decl))) |
356 | vec_safe_push (v&: offload_ind_funcs, obj: node->decl); |
357 | if (omp_declare_target_fn_p (decl: node->decl)) |
358 | worklist.safe_push (obj: node->decl); |
359 | else if (DECL_STRUCT_FUNCTION (node->decl) |
360 | && DECL_STRUCT_FUNCTION (node->decl)->has_omp_target) |
361 | worklist.safe_push (obj: node->decl); |
362 | for (cgn = first_nested_function (node); |
363 | cgn; cgn = next_nested_function (node: cgn)) |
364 | if (omp_declare_target_fn_p (decl: cgn->decl)) |
365 | worklist.safe_push (obj: cgn->decl); |
366 | else if (DECL_STRUCT_FUNCTION (cgn->decl) |
367 | && DECL_STRUCT_FUNCTION (cgn->decl)->has_omp_target) |
368 | worklist.safe_push (obj: cgn->decl); |
369 | } |
370 | FOR_EACH_VARIABLE (vnode) |
371 | if (lang_hooks.decls.omp_get_decl_init (vnode->decl) |
372 | && omp_declare_target_var_p (decl: vnode->decl)) |
373 | worklist.safe_push (obj: vnode->decl); |
374 | while (!worklist.is_empty ()) |
375 | { |
376 | tree decl = worklist.pop (); |
377 | if (VAR_P (decl)) |
378 | walk_tree_without_duplicates (lang_hooks.decls.omp_get_decl_init (decl), |
379 | omp_discover_declare_target_var_r, |
380 | &worklist); |
381 | else if (omp_declare_target_fn_p (decl)) |
382 | walk_tree_without_duplicates (&DECL_SAVED_TREE (decl), |
383 | omp_discover_declare_target_tgt_fn_r, |
384 | &worklist); |
385 | else |
386 | walk_tree_without_duplicates (&DECL_SAVED_TREE (decl), |
387 | omp_discover_declare_target_fn_r, |
388 | &worklist); |
389 | } |
390 | |
391 | lang_hooks.decls.omp_finish_decl_inits (); |
392 | } |
393 | |
394 | |
395 | /* Create new symbols containing (address, size) pairs for global variables, |
396 | marked with "omp declare target" attribute, as well as addresses for the |
397 | functions, which are outlined offloading regions. */ |
398 | void |
399 | omp_finish_file (void) |
400 | { |
401 | unsigned num_funcs = vec_safe_length (v: offload_funcs); |
402 | unsigned num_vars = vec_safe_length (v: offload_vars); |
403 | unsigned num_ind_funcs = vec_safe_length (v: offload_ind_funcs); |
404 | |
405 | if (num_funcs == 0 && num_vars == 0 && num_ind_funcs == 0) |
406 | return; |
407 | |
408 | if (targetm_common.have_named_sections) |
409 | { |
410 | vec<constructor_elt, va_gc> *v_f, *v_v, *v_if; |
411 | vec_alloc (v&: v_f, nelems: num_funcs); |
412 | vec_alloc (v&: v_v, nelems: num_vars * 2); |
413 | vec_alloc (v&: v_if, nelems: num_ind_funcs); |
414 | |
415 | add_decls_addresses_to_decl_constructor (v_decls: offload_funcs, v_ctor: v_f); |
416 | add_decls_addresses_to_decl_constructor (v_decls: offload_vars, v_ctor: v_v); |
417 | add_decls_addresses_to_decl_constructor (v_decls: offload_ind_funcs, v_ctor: v_if); |
418 | |
419 | tree vars_decl_type = build_array_type_nelts (pointer_sized_int_node, |
420 | vec_safe_length (v: v_v)); |
421 | tree funcs_decl_type = build_array_type_nelts (pointer_sized_int_node, |
422 | num_funcs); |
423 | tree ind_funcs_decl_type = build_array_type_nelts (pointer_sized_int_node, |
424 | num_ind_funcs); |
425 | |
426 | SET_TYPE_ALIGN (vars_decl_type, TYPE_ALIGN (pointer_sized_int_node)); |
427 | SET_TYPE_ALIGN (funcs_decl_type, TYPE_ALIGN (pointer_sized_int_node)); |
428 | SET_TYPE_ALIGN (ind_funcs_decl_type, TYPE_ALIGN (pointer_sized_int_node)); |
429 | tree ctor_v = build_constructor (vars_decl_type, v_v); |
430 | tree ctor_f = build_constructor (funcs_decl_type, v_f); |
431 | tree ctor_if = build_constructor (ind_funcs_decl_type, v_if); |
432 | TREE_CONSTANT (ctor_v) = TREE_CONSTANT (ctor_f) = TREE_CONSTANT (ctor_if) = 1; |
433 | TREE_STATIC (ctor_v) = TREE_STATIC (ctor_f) = TREE_STATIC (ctor_if) = 1; |
434 | tree funcs_decl = build_decl (UNKNOWN_LOCATION, VAR_DECL, |
435 | get_identifier (".offload_func_table" ), |
436 | funcs_decl_type); |
437 | tree vars_decl = build_decl (UNKNOWN_LOCATION, VAR_DECL, |
438 | get_identifier (".offload_var_table" ), |
439 | vars_decl_type); |
440 | tree ind_funcs_decl = build_decl (UNKNOWN_LOCATION, VAR_DECL, |
441 | get_identifier (".offload_ind_func_table" ), |
442 | ind_funcs_decl_type); |
443 | TREE_STATIC (funcs_decl) = TREE_STATIC (ind_funcs_decl) = 1; |
444 | TREE_STATIC (vars_decl) = 1; |
445 | /* Do not align tables more than TYPE_ALIGN (pointer_sized_int_node), |
446 | otherwise a joint table in a binary will contain padding between |
447 | tables from multiple object files. */ |
448 | DECL_USER_ALIGN (funcs_decl) = DECL_USER_ALIGN (ind_funcs_decl) = 1; |
449 | DECL_USER_ALIGN (vars_decl) = 1; |
450 | SET_DECL_ALIGN (funcs_decl, TYPE_ALIGN (funcs_decl_type)); |
451 | SET_DECL_ALIGN (vars_decl, TYPE_ALIGN (vars_decl_type)); |
452 | SET_DECL_ALIGN (ind_funcs_decl, TYPE_ALIGN (ind_funcs_decl_type)); |
453 | DECL_INITIAL (funcs_decl) = ctor_f; |
454 | DECL_INITIAL (vars_decl) = ctor_v; |
455 | DECL_INITIAL (ind_funcs_decl) = ctor_if; |
456 | set_decl_section_name (funcs_decl, OFFLOAD_FUNC_TABLE_SECTION_NAME); |
457 | set_decl_section_name (vars_decl, OFFLOAD_VAR_TABLE_SECTION_NAME); |
458 | set_decl_section_name (ind_funcs_decl, |
459 | OFFLOAD_IND_FUNC_TABLE_SECTION_NAME); |
460 | varpool_node::finalize_decl (decl: vars_decl); |
461 | varpool_node::finalize_decl (decl: funcs_decl); |
462 | varpool_node::finalize_decl (decl: ind_funcs_decl); |
463 | } |
464 | else |
465 | { |
466 | for (unsigned i = 0; i < num_funcs; i++) |
467 | { |
468 | tree it = (*offload_funcs)[i]; |
469 | /* See also add_decls_addresses_to_decl_constructor |
470 | and output_offload_tables in lto-cgraph.cc. */ |
471 | if (!in_lto_p && !symtab_node::get (decl: it)) |
472 | continue; |
473 | targetm.record_offload_symbol (it); |
474 | } |
475 | for (unsigned i = 0; i < num_vars; i++) |
476 | { |
477 | tree it = (*offload_vars)[i]; |
478 | if (!in_lto_p && !symtab_node::get (decl: it)) |
479 | continue; |
480 | #ifdef ACCEL_COMPILER |
481 | if (DECL_HAS_VALUE_EXPR_P (it) |
482 | && lookup_attribute ("omp declare target link" , |
483 | DECL_ATTRIBUTES (it))) |
484 | { |
485 | tree value_expr = DECL_VALUE_EXPR (it); |
486 | tree link_ptr_decl = TREE_OPERAND (value_expr, 0); |
487 | targetm.record_offload_symbol (link_ptr_decl); |
488 | varpool_node::finalize_decl (link_ptr_decl); |
489 | } |
490 | else |
491 | #endif |
492 | targetm.record_offload_symbol (it); |
493 | } |
494 | for (unsigned i = 0; i < num_ind_funcs; i++) |
495 | { |
496 | tree it = (*offload_ind_funcs)[i]; |
497 | /* See also add_decls_addresses_to_decl_constructor |
498 | and output_offload_tables in lto-cgraph.cc. */ |
499 | if (!in_lto_p && !symtab_node::get (decl: it)) |
500 | continue; |
501 | targetm.record_offload_symbol (it); |
502 | } |
503 | } |
504 | } |
505 | |
506 | /* Call dim_pos (POS == true) or dim_size (POS == false) builtins for |
507 | axis DIM. Return a tmp var holding the result. */ |
508 | |
509 | static tree |
510 | oacc_dim_call (bool pos, int dim, gimple_seq *seq) |
511 | { |
512 | tree arg = build_int_cst (unsigned_type_node, dim); |
513 | tree size = create_tmp_var (integer_type_node); |
514 | enum internal_fn fn = pos ? IFN_GOACC_DIM_POS : IFN_GOACC_DIM_SIZE; |
515 | gimple *call = gimple_build_call_internal (fn, 1, arg); |
516 | |
517 | gimple_call_set_lhs (gs: call, lhs: size); |
518 | gimple_seq_add_stmt (seq, call); |
519 | |
520 | return size; |
521 | } |
522 | |
523 | /* Find the number of threads (POS = false), or thread number (POS = |
524 | true) for an OpenACC region partitioned as MASK. Setup code |
525 | required for the calculation is added to SEQ. */ |
526 | |
527 | static tree |
528 | oacc_thread_numbers (bool pos, int mask, gimple_seq *seq) |
529 | { |
530 | tree res = pos ? NULL_TREE : build_int_cst (unsigned_type_node, 1); |
531 | unsigned ix; |
532 | |
533 | /* Start at gang level, and examine relevant dimension indices. */ |
534 | for (ix = GOMP_DIM_GANG; ix != GOMP_DIM_MAX; ix++) |
535 | if (GOMP_DIM_MASK (ix) & mask) |
536 | { |
537 | if (res) |
538 | { |
539 | /* We had an outer index, so scale that by the size of |
540 | this dimension. */ |
541 | tree n = oacc_dim_call (pos: false, dim: ix, seq); |
542 | res = fold_build2 (MULT_EXPR, integer_type_node, res, n); |
543 | } |
544 | if (pos) |
545 | { |
546 | /* Determine index in this dimension. */ |
547 | tree id = oacc_dim_call (pos: true, dim: ix, seq); |
548 | if (res) |
549 | res = fold_build2 (PLUS_EXPR, integer_type_node, res, id); |
550 | else |
551 | res = id; |
552 | } |
553 | } |
554 | |
555 | if (res == NULL_TREE) |
556 | res = integer_zero_node; |
557 | |
558 | return res; |
559 | } |
560 | |
561 | /* Transform IFN_GOACC_LOOP calls to actual code. See |
562 | expand_oacc_for for where these are generated. At the vector |
563 | level, we stride loops, such that each member of a warp will |
564 | operate on adjacent iterations. At the worker and gang level, |
565 | each gang/warp executes a set of contiguous iterations. Chunking |
566 | can override this such that each iteration engine executes a |
567 | contiguous chunk, and then moves on to stride to the next chunk. */ |
568 | |
569 | static void |
570 | oacc_xform_loop (gcall *call) |
571 | { |
572 | gimple_stmt_iterator gsi = gsi_for_stmt (call); |
573 | enum ifn_goacc_loop_kind code |
574 | = (enum ifn_goacc_loop_kind) TREE_INT_CST_LOW (gimple_call_arg (call, 0)); |
575 | tree dir = gimple_call_arg (gs: call, index: 1); |
576 | tree range = gimple_call_arg (gs: call, index: 2); |
577 | tree step = gimple_call_arg (gs: call, index: 3); |
578 | tree chunk_size = NULL_TREE; |
579 | unsigned mask = (unsigned) TREE_INT_CST_LOW (gimple_call_arg (call, 5)); |
580 | tree lhs = gimple_call_lhs (gs: call); |
581 | tree type = NULL_TREE; |
582 | tree diff_type = TREE_TYPE (range); |
583 | tree r = NULL_TREE; |
584 | gimple_seq seq = NULL; |
585 | bool chunking = false, striding = true; |
586 | unsigned outer_mask = mask & (~mask + 1); // Outermost partitioning |
587 | unsigned inner_mask = mask & ~outer_mask; // Inner partitioning (if any) |
588 | |
589 | /* Skip lowering if return value of IFN_GOACC_LOOP call is not used. */ |
590 | if (!lhs) |
591 | { |
592 | gsi_replace_with_seq (&gsi, seq, true); |
593 | return; |
594 | } |
595 | |
596 | type = TREE_TYPE (lhs); |
597 | |
598 | #ifdef ACCEL_COMPILER |
599 | chunk_size = gimple_call_arg (call, 4); |
600 | if (integer_minus_onep (chunk_size) /* Force static allocation. */ |
601 | || integer_zerop (chunk_size)) /* Default (also static). */ |
602 | { |
603 | /* If we're at the gang level, we want each to execute a |
604 | contiguous run of iterations. Otherwise we want each element |
605 | to stride. */ |
606 | striding = !(outer_mask & GOMP_DIM_MASK (GOMP_DIM_GANG)); |
607 | chunking = false; |
608 | } |
609 | else |
610 | { |
611 | /* Chunk of size 1 is striding. */ |
612 | striding = integer_onep (chunk_size); |
613 | chunking = !striding; |
614 | } |
615 | #endif |
616 | |
617 | /* striding=true, chunking=true |
618 | -> invalid. |
619 | striding=true, chunking=false |
620 | -> chunks=1 |
621 | striding=false,chunking=true |
622 | -> chunks=ceil (range/(chunksize*threads*step)) |
623 | striding=false,chunking=false |
624 | -> chunk_size=ceil(range/(threads*step)),chunks=1 */ |
625 | push_gimplify_context (in_ssa: true); |
626 | |
627 | switch (code) |
628 | { |
629 | default: gcc_unreachable (); |
630 | |
631 | case IFN_GOACC_LOOP_CHUNKS: |
632 | if (!chunking) |
633 | r = build_int_cst (type, 1); |
634 | else |
635 | { |
636 | /* chunk_max |
637 | = (range - dir) / (chunks * step * num_threads) + dir */ |
638 | tree per = oacc_thread_numbers (pos: false, mask, seq: &seq); |
639 | per = fold_convert (type, per); |
640 | chunk_size = fold_convert (type, chunk_size); |
641 | per = fold_build2 (MULT_EXPR, type, per, chunk_size); |
642 | per = fold_build2 (MULT_EXPR, type, per, step); |
643 | r = build2 (MINUS_EXPR, type, range, dir); |
644 | r = build2 (PLUS_EXPR, type, r, per); |
645 | r = build2 (TRUNC_DIV_EXPR, type, r, per); |
646 | } |
647 | break; |
648 | |
649 | case IFN_GOACC_LOOP_STEP: |
650 | { |
651 | /* If striding, step by the entire compute volume, otherwise |
652 | step by the inner volume. */ |
653 | unsigned volume = striding ? mask : inner_mask; |
654 | |
655 | r = oacc_thread_numbers (pos: false, mask: volume, seq: &seq); |
656 | r = build2 (MULT_EXPR, type, fold_convert (type, r), step); |
657 | } |
658 | break; |
659 | |
660 | case IFN_GOACC_LOOP_OFFSET: |
661 | /* Enable vectorization on non-SIMT targets. */ |
662 | if (!targetm.simt.vf |
663 | && outer_mask == GOMP_DIM_MASK (GOMP_DIM_VECTOR) |
664 | /* If not -fno-tree-loop-vectorize, hint that we want to vectorize |
665 | the loop. */ |
666 | && (flag_tree_loop_vectorize |
667 | || !OPTION_SET_P (flag_tree_loop_vectorize))) |
668 | { |
669 | basic_block bb = gsi_bb (i: gsi); |
670 | class loop *parent = bb->loop_father; |
671 | class loop *body = parent->inner; |
672 | |
673 | parent->force_vectorize = true; |
674 | parent->safelen = INT_MAX; |
675 | |
676 | /* "Chunking loops" may have inner loops. */ |
677 | if (parent->inner) |
678 | { |
679 | body->force_vectorize = true; |
680 | body->safelen = INT_MAX; |
681 | } |
682 | |
683 | cfun->has_force_vectorize_loops = true; |
684 | } |
685 | if (striding) |
686 | { |
687 | r = oacc_thread_numbers (pos: true, mask, seq: &seq); |
688 | r = fold_convert (diff_type, r); |
689 | } |
690 | else |
691 | { |
692 | tree inner_size = oacc_thread_numbers (pos: false, mask: inner_mask, seq: &seq); |
693 | tree outer_size = oacc_thread_numbers (pos: false, mask: outer_mask, seq: &seq); |
694 | tree volume = fold_build2 (MULT_EXPR, TREE_TYPE (inner_size), |
695 | inner_size, outer_size); |
696 | |
697 | volume = fold_convert (diff_type, volume); |
698 | if (chunking) |
699 | chunk_size = fold_convert (diff_type, chunk_size); |
700 | else |
701 | { |
702 | tree per = fold_build2 (MULT_EXPR, diff_type, volume, step); |
703 | |
704 | chunk_size = build2 (MINUS_EXPR, diff_type, range, dir); |
705 | chunk_size = build2 (PLUS_EXPR, diff_type, chunk_size, per); |
706 | chunk_size = build2 (TRUNC_DIV_EXPR, diff_type, chunk_size, per); |
707 | } |
708 | |
709 | tree span = build2 (MULT_EXPR, diff_type, chunk_size, |
710 | fold_convert (diff_type, inner_size)); |
711 | r = oacc_thread_numbers (pos: true, mask: outer_mask, seq: &seq); |
712 | r = fold_convert (diff_type, r); |
713 | r = build2 (MULT_EXPR, diff_type, r, span); |
714 | |
715 | tree inner = oacc_thread_numbers (pos: true, mask: inner_mask, seq: &seq); |
716 | inner = fold_convert (diff_type, inner); |
717 | r = fold_build2 (PLUS_EXPR, diff_type, r, inner); |
718 | |
719 | if (chunking) |
720 | { |
721 | tree chunk = fold_convert (diff_type, gimple_call_arg (call, 6)); |
722 | tree per |
723 | = fold_build2 (MULT_EXPR, diff_type, volume, chunk_size); |
724 | per = build2 (MULT_EXPR, diff_type, per, chunk); |
725 | |
726 | r = build2 (PLUS_EXPR, diff_type, r, per); |
727 | } |
728 | } |
729 | r = fold_build2 (MULT_EXPR, diff_type, r, step); |
730 | if (type != diff_type) |
731 | r = fold_convert (type, r); |
732 | break; |
733 | |
734 | case IFN_GOACC_LOOP_BOUND: |
735 | if (striding) |
736 | r = range; |
737 | else |
738 | { |
739 | tree inner_size = oacc_thread_numbers (pos: false, mask: inner_mask, seq: &seq); |
740 | tree outer_size = oacc_thread_numbers (pos: false, mask: outer_mask, seq: &seq); |
741 | tree volume = fold_build2 (MULT_EXPR, TREE_TYPE (inner_size), |
742 | inner_size, outer_size); |
743 | |
744 | volume = fold_convert (diff_type, volume); |
745 | if (chunking) |
746 | chunk_size = fold_convert (diff_type, chunk_size); |
747 | else |
748 | { |
749 | tree per = fold_build2 (MULT_EXPR, diff_type, volume, step); |
750 | |
751 | chunk_size = build2 (MINUS_EXPR, diff_type, range, dir); |
752 | chunk_size = build2 (PLUS_EXPR, diff_type, chunk_size, per); |
753 | chunk_size = build2 (TRUNC_DIV_EXPR, diff_type, chunk_size, per); |
754 | } |
755 | |
756 | tree span = build2 (MULT_EXPR, diff_type, chunk_size, |
757 | fold_convert (diff_type, inner_size)); |
758 | |
759 | r = fold_build2 (MULT_EXPR, diff_type, span, step); |
760 | |
761 | tree offset = gimple_call_arg (gs: call, index: 6); |
762 | r = build2 (PLUS_EXPR, diff_type, r, |
763 | fold_convert (diff_type, offset)); |
764 | r = build2 (integer_onep (dir) ? MIN_EXPR : MAX_EXPR, |
765 | diff_type, r, range); |
766 | } |
767 | if (diff_type != type) |
768 | r = fold_convert (type, r); |
769 | break; |
770 | } |
771 | |
772 | gimplify_assign (lhs, r, &seq); |
773 | |
774 | pop_gimplify_context (NULL); |
775 | |
776 | gsi_replace_with_seq (&gsi, seq, true); |
777 | } |
778 | |
779 | /* Transform a GOACC_TILE call. Determines the element loop span for |
780 | the specified loop of the nest. This is 1 if we're not tiling. |
781 | |
782 | GOACC_TILE (collapse_count, loop_no, tile_arg, gwv_tile, gwv_element); */ |
783 | |
784 | static void |
785 | oacc_xform_tile (gcall *call) |
786 | { |
787 | gimple_stmt_iterator gsi = gsi_for_stmt (call); |
788 | unsigned collapse = tree_to_uhwi (gimple_call_arg (gs: call, index: 0)); |
789 | /* Inner loops have higher loop_nos. */ |
790 | unsigned loop_no = tree_to_uhwi (gimple_call_arg (gs: call, index: 1)); |
791 | tree tile_size = gimple_call_arg (gs: call, index: 2); |
792 | unsigned e_mask = tree_to_uhwi (gimple_call_arg (gs: call, index: 4)); |
793 | tree lhs = gimple_call_lhs (gs: call); |
794 | tree type = TREE_TYPE (lhs); |
795 | gimple_seq seq = NULL; |
796 | tree span = build_int_cst (type, 1); |
797 | |
798 | gcc_assert (!(e_mask |
799 | & ~(GOMP_DIM_MASK (GOMP_DIM_VECTOR) |
800 | | GOMP_DIM_MASK (GOMP_DIM_WORKER)))); |
801 | push_gimplify_context (in_ssa: !seen_error ()); |
802 | |
803 | #ifndef ACCEL_COMPILER |
804 | /* Partitioning disabled on host compilers. */ |
805 | e_mask = 0; |
806 | #endif |
807 | if (!e_mask) |
808 | /* Not paritioning. */ |
809 | span = integer_one_node; |
810 | else if (!integer_zerop (tile_size)) |
811 | /* User explicitly specified size. */ |
812 | span = tile_size; |
813 | else |
814 | { |
815 | /* Pick a size based on the paritioning of the element loop and |
816 | the number of loop nests. */ |
817 | tree first_size = NULL_TREE; |
818 | tree second_size = NULL_TREE; |
819 | |
820 | if (e_mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR)) |
821 | first_size = oacc_dim_call (pos: false, GOMP_DIM_VECTOR, seq: &seq); |
822 | if (e_mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)) |
823 | second_size = oacc_dim_call (pos: false, GOMP_DIM_WORKER, seq: &seq); |
824 | |
825 | if (!first_size) |
826 | { |
827 | first_size = second_size; |
828 | second_size = NULL_TREE; |
829 | } |
830 | |
831 | if (loop_no + 1 == collapse) |
832 | { |
833 | span = first_size; |
834 | if (!loop_no && second_size) |
835 | span = fold_build2 (MULT_EXPR, TREE_TYPE (span), |
836 | span, second_size); |
837 | } |
838 | else if (loop_no + 2 == collapse) |
839 | span = second_size; |
840 | else |
841 | span = NULL_TREE; |
842 | |
843 | if (!span) |
844 | /* There's no obvious element size for this loop. Options |
845 | are 1, first_size or some non-unity constant (32 is my |
846 | favourite). We should gather some statistics. */ |
847 | span = first_size; |
848 | } |
849 | |
850 | span = fold_convert (type, span); |
851 | gimplify_assign (lhs, span, &seq); |
852 | |
853 | pop_gimplify_context (NULL); |
854 | |
855 | gsi_replace_with_seq (&gsi, seq, true); |
856 | } |
857 | |
858 | /* Default partitioned and minimum partitioned dimensions. */ |
859 | |
860 | static int oacc_default_dims[GOMP_DIM_MAX]; |
861 | static int oacc_min_dims[GOMP_DIM_MAX]; |
862 | |
863 | int |
864 | oacc_get_default_dim (int dim) |
865 | { |
866 | gcc_assert (0 <= dim && dim < GOMP_DIM_MAX); |
867 | return oacc_default_dims[dim]; |
868 | } |
869 | |
870 | int |
871 | oacc_get_min_dim (int dim) |
872 | { |
873 | gcc_assert (0 <= dim && dim < GOMP_DIM_MAX); |
874 | return oacc_min_dims[dim]; |
875 | } |
876 | |
877 | /* Parse the default dimension parameter. This is a set of |
878 | :-separated optional compute dimensions. Each specified dimension |
879 | is a positive integer. When device type support is added, it is |
880 | planned to be a comma separated list of such compute dimensions, |
881 | with all but the first prefixed by the colon-terminated device |
882 | type. */ |
883 | |
884 | static void |
885 | oacc_parse_default_dims (const char *dims) |
886 | { |
887 | int ix; |
888 | |
889 | for (ix = GOMP_DIM_MAX; ix--;) |
890 | { |
891 | oacc_default_dims[ix] = -1; |
892 | oacc_min_dims[ix] = 1; |
893 | } |
894 | |
895 | #ifndef ACCEL_COMPILER |
896 | /* Cannot be overridden on the host. */ |
897 | dims = NULL; |
898 | #endif |
899 | if (dims) |
900 | { |
901 | const char *pos = dims; |
902 | |
903 | for (ix = 0; *pos && ix != GOMP_DIM_MAX; ix++) |
904 | { |
905 | if (ix) |
906 | { |
907 | if (*pos != ':') |
908 | goto malformed; |
909 | pos++; |
910 | } |
911 | |
912 | if (*pos != ':') |
913 | { |
914 | long val; |
915 | const char *eptr; |
916 | |
917 | errno = 0; |
918 | val = strtol (nptr: pos, CONST_CAST (char **, &eptr), base: 10); |
919 | if (errno || val <= 0 || (int) val != val) |
920 | goto malformed; |
921 | pos = eptr; |
922 | oacc_default_dims[ix] = (int) val; |
923 | } |
924 | } |
925 | if (*pos) |
926 | { |
927 | malformed: |
928 | error_at (UNKNOWN_LOCATION, |
929 | "%<-fopenacc-dim%> operand is malformed at %qs" , pos); |
930 | } |
931 | } |
932 | |
933 | /* Allow the backend to validate the dimensions. */ |
934 | targetm.goacc.validate_dims (NULL_TREE, oacc_default_dims, -1, 0); |
935 | targetm.goacc.validate_dims (NULL_TREE, oacc_min_dims, -2, 0); |
936 | } |
937 | |
938 | /* Validate and update the dimensions for offloaded FN. ATTRS is the |
939 | raw attribute. DIMS is an array of dimensions, which is filled in. |
940 | LEVEL is the partitioning level of a routine, or -1 for an offload |
941 | region itself. USED is the mask of partitioned execution in the |
942 | function. */ |
943 | |
944 | static void |
945 | oacc_validate_dims (tree fn, tree attrs, int *dims, int level, unsigned used) |
946 | { |
947 | tree purpose[GOMP_DIM_MAX]; |
948 | unsigned ix; |
949 | tree pos = TREE_VALUE (attrs); |
950 | |
951 | /* Make sure the attribute creator attached the dimension |
952 | information. */ |
953 | gcc_assert (pos); |
954 | |
955 | for (ix = 0; ix != GOMP_DIM_MAX; ix++) |
956 | { |
957 | purpose[ix] = TREE_PURPOSE (pos); |
958 | tree val = TREE_VALUE (pos); |
959 | dims[ix] = val ? TREE_INT_CST_LOW (val) : -1; |
960 | pos = TREE_CHAIN (pos); |
961 | } |
962 | |
963 | bool check = true; |
964 | #ifdef ACCEL_COMPILER |
965 | check = false; |
966 | #endif |
967 | if (check |
968 | && warn_openacc_parallelism |
969 | && !lookup_attribute (attr_name: "oacc kernels" , DECL_ATTRIBUTES (fn))) |
970 | { |
971 | static char const *const axes[] = |
972 | /* Must be kept in sync with GOMP_DIM enumeration. */ |
973 | { "gang" , "worker" , "vector" }; |
974 | for (ix = level >= 0 ? level : 0; ix != GOMP_DIM_MAX; ix++) |
975 | if (dims[ix] < 0) |
976 | ; /* Defaulting axis. */ |
977 | else if ((used & GOMP_DIM_MASK (ix)) && dims[ix] == 1) |
978 | /* There is partitioned execution, but the user requested a |
979 | dimension size of 1. They're probably confused. */ |
980 | warning_at (DECL_SOURCE_LOCATION (fn), OPT_Wopenacc_parallelism, |
981 | "region contains %s partitioned code but" |
982 | " is not %s partitioned" , axes[ix], axes[ix]); |
983 | else if (!(used & GOMP_DIM_MASK (ix)) && dims[ix] != 1) |
984 | /* The dimension is explicitly partitioned to non-unity, but |
985 | no use is made within the region. */ |
986 | warning_at (DECL_SOURCE_LOCATION (fn), OPT_Wopenacc_parallelism, |
987 | "region is %s partitioned but" |
988 | " does not contain %s partitioned code" , |
989 | axes[ix], axes[ix]); |
990 | } |
991 | |
992 | bool changed = targetm.goacc.validate_dims (fn, dims, level, used); |
993 | |
994 | /* Default anything left to 1 or a partitioned default. */ |
995 | for (ix = 0; ix != GOMP_DIM_MAX; ix++) |
996 | if (dims[ix] < 0) |
997 | { |
998 | /* The OpenACC spec says 'If the [num_gangs] clause is not |
999 | specified, an implementation-defined default will be used; |
1000 | the default may depend on the code within the construct.' |
1001 | (2.5.6). Thus an implementation is free to choose |
1002 | non-unity default for a parallel region that doesn't have |
1003 | any gang-partitioned loops. However, it appears that there |
1004 | is a sufficient body of user code that expects non-gang |
1005 | partitioned regions to not execute in gang-redundant mode. |
1006 | So we (a) don't warn about the non-portability and (b) pick |
1007 | the minimum permissible dimension size when there is no |
1008 | partitioned execution. Otherwise we pick the global |
1009 | default for the dimension, which the user can control. The |
1010 | same wording and logic applies to num_workers and |
1011 | vector_length, however the worker- or vector- single |
1012 | execution doesn't have the same impact as gang-redundant |
1013 | execution. (If the minimum gang-level partioning is not 1, |
1014 | the target is probably too confusing.) */ |
1015 | dims[ix] = (used & GOMP_DIM_MASK (ix) |
1016 | ? oacc_default_dims[ix] : oacc_min_dims[ix]); |
1017 | changed = true; |
1018 | } |
1019 | |
1020 | if (changed) |
1021 | { |
1022 | /* Replace the attribute with new values. */ |
1023 | pos = NULL_TREE; |
1024 | for (ix = GOMP_DIM_MAX; ix--;) |
1025 | pos = tree_cons (purpose[ix], |
1026 | build_int_cst (integer_type_node, dims[ix]), pos); |
1027 | oacc_replace_fn_attrib (fn, dims: pos); |
1028 | } |
1029 | } |
1030 | |
1031 | /* Create an empty OpenACC loop structure at LOC. */ |
1032 | |
1033 | static oacc_loop * |
1034 | new_oacc_loop_raw (oacc_loop *parent, location_t loc) |
1035 | { |
1036 | oacc_loop *loop = XCNEW (oacc_loop); |
1037 | |
1038 | loop->parent = parent; |
1039 | |
1040 | if (parent) |
1041 | { |
1042 | loop->sibling = parent->child; |
1043 | parent->child = loop; |
1044 | } |
1045 | |
1046 | loop->loc = loc; |
1047 | return loop; |
1048 | } |
1049 | |
1050 | /* Create an outermost, dummy OpenACC loop for offloaded function |
1051 | DECL. */ |
1052 | |
1053 | static oacc_loop * |
1054 | new_oacc_loop_outer (tree decl) |
1055 | { |
1056 | return new_oacc_loop_raw (NULL, DECL_SOURCE_LOCATION (decl)); |
1057 | } |
1058 | |
1059 | /* Start a new OpenACC loop structure beginning at head marker HEAD. |
1060 | Link into PARENT loop. Return the new loop. */ |
1061 | |
1062 | static oacc_loop * |
1063 | new_oacc_loop (oacc_loop *parent, gcall *marker) |
1064 | { |
1065 | oacc_loop *loop = new_oacc_loop_raw (parent, loc: gimple_location (g: marker)); |
1066 | |
1067 | loop->marker = marker; |
1068 | |
1069 | /* TODO: This is where device_type flattening would occur for the loop |
1070 | flags. */ |
1071 | |
1072 | loop->flags = TREE_INT_CST_LOW (gimple_call_arg (marker, 3)); |
1073 | |
1074 | tree chunk_size = integer_zero_node; |
1075 | if (loop->flags & OLF_GANG_STATIC) |
1076 | chunk_size = gimple_call_arg (gs: marker, index: 4); |
1077 | loop->chunk_size = chunk_size; |
1078 | |
1079 | return loop; |
1080 | } |
1081 | |
1082 | /* Create a dummy loop encompassing a call to a openACC routine. |
1083 | Extract the routine's partitioning requirements. */ |
1084 | |
1085 | static void |
1086 | new_oacc_loop_routine (oacc_loop *parent, gcall *call, tree decl, tree attrs) |
1087 | { |
1088 | oacc_loop *loop = new_oacc_loop_raw (parent, loc: gimple_location (g: call)); |
1089 | int level = oacc_fn_attrib_level (attr: attrs); |
1090 | |
1091 | gcc_assert (level >= 0); |
1092 | |
1093 | loop->marker = call; |
1094 | loop->routine = decl; |
1095 | loop->mask = ((GOMP_DIM_MASK (GOMP_DIM_MAX) - 1) |
1096 | ^ (GOMP_DIM_MASK (level) - 1)); |
1097 | } |
1098 | |
1099 | /* Finish off the current OpenACC loop ending at tail marker TAIL. |
1100 | Return the parent loop. */ |
1101 | |
1102 | static oacc_loop * |
1103 | finish_oacc_loop (oacc_loop *loop) |
1104 | { |
1105 | /* If the loop has been collapsed, don't partition it. */ |
1106 | if (loop->ifns.is_empty ()) |
1107 | loop->mask = loop->flags = 0; |
1108 | return loop->parent; |
1109 | } |
1110 | |
1111 | /* Free all OpenACC loop structures within LOOP (inclusive). */ |
1112 | |
1113 | static void |
1114 | free_oacc_loop (oacc_loop *loop) |
1115 | { |
1116 | if (loop->sibling) |
1117 | free_oacc_loop (loop: loop->sibling); |
1118 | if (loop->child) |
1119 | free_oacc_loop (loop: loop->child); |
1120 | |
1121 | loop->ifns.release (); |
1122 | free (ptr: loop); |
1123 | } |
1124 | |
1125 | /* Dump out the OpenACC loop head or tail beginning at FROM. */ |
1126 | |
1127 | static void |
1128 | dump_oacc_loop_part (FILE *file, gcall *from, int depth, |
1129 | const char *title, int level) |
1130 | { |
1131 | enum ifn_unique_kind kind |
1132 | = (enum ifn_unique_kind) TREE_INT_CST_LOW (gimple_call_arg (from, 0)); |
1133 | |
1134 | fprintf (stream: file, format: "%*s%s-%d:\n" , depth * 2, "" , title, level); |
1135 | for (gimple_stmt_iterator gsi = gsi_for_stmt (from);;) |
1136 | { |
1137 | gimple *stmt = gsi_stmt (i: gsi); |
1138 | |
1139 | if (gimple_call_internal_p (gs: stmt, fn: IFN_UNIQUE)) |
1140 | { |
1141 | enum ifn_unique_kind k |
1142 | = ((enum ifn_unique_kind) TREE_INT_CST_LOW |
1143 | (gimple_call_arg (stmt, 0))); |
1144 | |
1145 | if (k == kind && stmt != from) |
1146 | break; |
1147 | } |
1148 | print_gimple_stmt (file, stmt, depth * 2 + 2); |
1149 | |
1150 | gsi_next (i: &gsi); |
1151 | while (gsi_end_p (i: gsi)) |
1152 | gsi = gsi_start_bb (bb: single_succ (bb: gsi_bb (i: gsi))); |
1153 | } |
1154 | } |
1155 | |
1156 | /* Dump OpenACC loop LOOP, its children, and its siblings. */ |
1157 | |
1158 | static void |
1159 | dump_oacc_loop (FILE *file, oacc_loop *loop, int depth) |
1160 | { |
1161 | int ix; |
1162 | |
1163 | fprintf (stream: file, format: "%*sLoop %x(%x) %s:%u\n" , depth * 2, "" , |
1164 | loop->flags, loop->mask, |
1165 | LOCATION_FILE (loop->loc), LOCATION_LINE (loop->loc)); |
1166 | |
1167 | if (loop->marker) |
1168 | print_gimple_stmt (file, loop->marker, depth * 2); |
1169 | |
1170 | if (loop->routine) |
1171 | fprintf (stream: file, format: "%*sRoutine %s:%u:%s\n" , |
1172 | depth * 2, "" , DECL_SOURCE_FILE (loop->routine), |
1173 | DECL_SOURCE_LINE (loop->routine), |
1174 | IDENTIFIER_POINTER (DECL_NAME (loop->routine))); |
1175 | |
1176 | for (ix = GOMP_DIM_GANG; ix != GOMP_DIM_MAX; ix++) |
1177 | if (loop->heads[ix]) |
1178 | dump_oacc_loop_part (file, from: loop->heads[ix], depth, title: "Head" , level: ix); |
1179 | for (ix = GOMP_DIM_MAX; ix--;) |
1180 | if (loop->tails[ix]) |
1181 | dump_oacc_loop_part (file, from: loop->tails[ix], depth, title: "Tail" , level: ix); |
1182 | |
1183 | if (loop->child) |
1184 | dump_oacc_loop (file, loop: loop->child, depth: depth + 1); |
1185 | if (loop->sibling) |
1186 | dump_oacc_loop (file, loop: loop->sibling, depth); |
1187 | } |
1188 | |
1189 | void debug_oacc_loop (oacc_loop *); |
1190 | |
1191 | /* Dump loops to stderr. */ |
1192 | |
1193 | DEBUG_FUNCTION void |
1194 | debug_oacc_loop (oacc_loop *loop) |
1195 | { |
1196 | dump_oacc_loop (stderr, loop, depth: 0); |
1197 | } |
1198 | |
1199 | /* Provide diagnostics on OpenACC loop LOOP, its children, and its |
1200 | siblings. */ |
1201 | |
1202 | static void |
1203 | inform_oacc_loop (const oacc_loop *loop) |
1204 | { |
1205 | const char *gang |
1206 | = loop->mask & GOMP_DIM_MASK (GOMP_DIM_GANG) ? " gang" : "" ; |
1207 | const char *worker |
1208 | = loop->mask & GOMP_DIM_MASK (GOMP_DIM_WORKER) ? " worker" : "" ; |
1209 | const char *vector |
1210 | = loop->mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR) ? " vector" : "" ; |
1211 | const char *seq = loop->mask == 0 ? " seq" : "" ; |
1212 | const dump_user_location_t loc |
1213 | = dump_user_location_t::from_location_t (loc: loop->loc); |
1214 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc, |
1215 | "assigned OpenACC%s%s%s%s loop parallelism\n" , gang, worker, |
1216 | vector, seq); |
1217 | |
1218 | if (loop->child) |
1219 | inform_oacc_loop (loop: loop->child); |
1220 | if (loop->sibling) |
1221 | inform_oacc_loop (loop: loop->sibling); |
1222 | } |
1223 | |
1224 | /* DFS walk of basic blocks BB onwards, creating OpenACC loop |
1225 | structures as we go. By construction these loops are properly |
1226 | nested. */ |
1227 | |
1228 | static void |
1229 | oacc_loop_discover_walk (oacc_loop *loop, basic_block bb) |
1230 | { |
1231 | int marker = 0; |
1232 | int remaining = 0; |
1233 | |
1234 | if (bb->flags & BB_VISITED) |
1235 | return; |
1236 | |
1237 | follow: |
1238 | bb->flags |= BB_VISITED; |
1239 | |
1240 | /* Scan for loop markers. */ |
1241 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); |
1242 | gsi_next (i: &gsi)) |
1243 | { |
1244 | gimple *stmt = gsi_stmt (i: gsi); |
1245 | |
1246 | if (!is_gimple_call (gs: stmt)) |
1247 | continue; |
1248 | |
1249 | gcall *call = as_a <gcall *> (p: stmt); |
1250 | |
1251 | /* If this is a routine, make a dummy loop for it. */ |
1252 | if (tree decl = gimple_call_fndecl (gs: call)) |
1253 | if (tree attrs = oacc_get_fn_attrib (fn: decl)) |
1254 | { |
1255 | gcc_assert (!marker); |
1256 | new_oacc_loop_routine (parent: loop, call, decl, attrs); |
1257 | } |
1258 | |
1259 | if (!gimple_call_internal_p (gs: call)) |
1260 | continue; |
1261 | |
1262 | switch (gimple_call_internal_fn (gs: call)) |
1263 | { |
1264 | default: |
1265 | break; |
1266 | |
1267 | case IFN_GOACC_LOOP: |
1268 | case IFN_GOACC_TILE: |
1269 | /* Record the abstraction function, so we can manipulate it |
1270 | later. */ |
1271 | loop->ifns.safe_push (obj: call); |
1272 | break; |
1273 | |
1274 | case IFN_UNIQUE: |
1275 | enum ifn_unique_kind kind |
1276 | = (enum ifn_unique_kind) (TREE_INT_CST_LOW |
1277 | (gimple_call_arg (call, 0))); |
1278 | if (kind == IFN_UNIQUE_OACC_HEAD_MARK |
1279 | || kind == IFN_UNIQUE_OACC_TAIL_MARK) |
1280 | { |
1281 | if (gimple_call_num_args (gs: call) == 2) |
1282 | { |
1283 | gcc_assert (marker && !remaining); |
1284 | marker = 0; |
1285 | if (kind == IFN_UNIQUE_OACC_TAIL_MARK) |
1286 | loop = finish_oacc_loop (loop); |
1287 | else |
1288 | loop->head_end = call; |
1289 | } |
1290 | else |
1291 | { |
1292 | int count = TREE_INT_CST_LOW (gimple_call_arg (call, 2)); |
1293 | |
1294 | if (!marker) |
1295 | { |
1296 | if (kind == IFN_UNIQUE_OACC_HEAD_MARK) |
1297 | loop = new_oacc_loop (parent: loop, marker: call); |
1298 | remaining = count; |
1299 | } |
1300 | gcc_assert (count == remaining); |
1301 | if (remaining) |
1302 | { |
1303 | remaining--; |
1304 | if (kind == IFN_UNIQUE_OACC_HEAD_MARK) |
1305 | loop->heads[marker] = call; |
1306 | else |
1307 | loop->tails[remaining] = call; |
1308 | } |
1309 | marker++; |
1310 | } |
1311 | } |
1312 | } |
1313 | } |
1314 | if (remaining || marker) |
1315 | { |
1316 | bb = single_succ (bb); |
1317 | gcc_assert (single_pred_p (bb) && !(bb->flags & BB_VISITED)); |
1318 | goto follow; |
1319 | } |
1320 | |
1321 | /* Walk successor blocks. */ |
1322 | edge e; |
1323 | edge_iterator ei; |
1324 | |
1325 | FOR_EACH_EDGE (e, ei, bb->succs) |
1326 | oacc_loop_discover_walk (loop, bb: e->dest); |
1327 | } |
1328 | |
1329 | /* LOOP is the first sibling. Reverse the order in place and return |
1330 | the new first sibling. Recurse to child loops. */ |
1331 | |
1332 | static oacc_loop * |
1333 | oacc_loop_sibling_nreverse (oacc_loop *loop) |
1334 | { |
1335 | oacc_loop *last = NULL; |
1336 | do |
1337 | { |
1338 | if (loop->child) |
1339 | loop->child = oacc_loop_sibling_nreverse (loop: loop->child); |
1340 | |
1341 | oacc_loop *next = loop->sibling; |
1342 | loop->sibling = last; |
1343 | last = loop; |
1344 | loop = next; |
1345 | } |
1346 | while (loop); |
1347 | |
1348 | return last; |
1349 | } |
1350 | |
1351 | /* Discover the OpenACC loops marked up by HEAD and TAIL markers for |
1352 | the current function. */ |
1353 | |
1354 | static oacc_loop * |
1355 | oacc_loop_discovery () |
1356 | { |
1357 | /* Clear basic block flags, in particular BB_VISITED which we're going to use |
1358 | in the following. */ |
1359 | clear_bb_flags (); |
1360 | |
1361 | oacc_loop *top = new_oacc_loop_outer (decl: current_function_decl); |
1362 | oacc_loop_discover_walk (loop: top, ENTRY_BLOCK_PTR_FOR_FN (cfun)); |
1363 | |
1364 | /* The siblings were constructed in reverse order, reverse them so |
1365 | that diagnostics come out in an unsurprising order. */ |
1366 | top = oacc_loop_sibling_nreverse (loop: top); |
1367 | |
1368 | return top; |
1369 | } |
1370 | |
1371 | /* Transform the abstract internal function markers starting at FROM |
1372 | to be for partitioning level LEVEL. Stop when we meet another HEAD |
1373 | or TAIL marker. */ |
1374 | |
1375 | static void |
1376 | oacc_loop_xform_head_tail (gcall *from, int level) |
1377 | { |
1378 | enum ifn_unique_kind kind |
1379 | = (enum ifn_unique_kind) TREE_INT_CST_LOW (gimple_call_arg (from, 0)); |
1380 | tree replacement = build_int_cst (unsigned_type_node, level); |
1381 | |
1382 | for (gimple_stmt_iterator gsi = gsi_for_stmt (from);;) |
1383 | { |
1384 | gimple *stmt = gsi_stmt (i: gsi); |
1385 | |
1386 | if (gimple_call_internal_p (gs: stmt, fn: IFN_UNIQUE)) |
1387 | { |
1388 | enum ifn_unique_kind k |
1389 | = ((enum ifn_unique_kind) |
1390 | TREE_INT_CST_LOW (gimple_call_arg (stmt, 0))); |
1391 | |
1392 | if (k == IFN_UNIQUE_OACC_FORK |
1393 | || k == IFN_UNIQUE_OACC_JOIN |
1394 | || k == IFN_UNIQUE_OACC_PRIVATE) |
1395 | *gimple_call_arg_ptr (gs: stmt, index: 2) = replacement; |
1396 | else if (k == kind && stmt != from) |
1397 | break; |
1398 | } |
1399 | else if (gimple_call_internal_p (gs: stmt, fn: IFN_GOACC_REDUCTION)) |
1400 | *gimple_call_arg_ptr (gs: stmt, index: 3) = replacement; |
1401 | update_stmt (s: stmt); |
1402 | |
1403 | gsi_next (i: &gsi); |
1404 | while (gsi_end_p (i: gsi)) |
1405 | gsi = gsi_start_bb (bb: single_succ (bb: gsi_bb (i: gsi))); |
1406 | } |
1407 | } |
1408 | |
1409 | /* Process the discovered OpenACC loops, setting the correct |
1410 | partitioning level etc. */ |
1411 | |
1412 | static void |
1413 | oacc_loop_process (oacc_loop *loop, int fn_level) |
1414 | { |
1415 | if (loop->child) |
1416 | oacc_loop_process (loop: loop->child, fn_level); |
1417 | |
1418 | if (loop->mask && !loop->routine) |
1419 | { |
1420 | int ix; |
1421 | tree mask_arg = build_int_cst (unsigned_type_node, loop->mask); |
1422 | tree e_mask_arg = build_int_cst (unsigned_type_node, loop->e_mask); |
1423 | tree chunk_arg = loop->chunk_size; |
1424 | gcall *call; |
1425 | |
1426 | for (ix = 0; loop->ifns.iterate (ix, ptr: &call); ix++) |
1427 | { |
1428 | switch (gimple_call_internal_fn (gs: call)) |
1429 | { |
1430 | case IFN_GOACC_LOOP: |
1431 | { |
1432 | bool is_e = gimple_call_arg (gs: call, index: 5) == integer_minus_one_node; |
1433 | gimple_call_set_arg (gs: call, index: 5, arg: is_e ? e_mask_arg : mask_arg); |
1434 | if (!is_e) |
1435 | gimple_call_set_arg (gs: call, index: 4, arg: chunk_arg); |
1436 | } |
1437 | break; |
1438 | |
1439 | case IFN_GOACC_TILE: |
1440 | gimple_call_set_arg (gs: call, index: 3, arg: mask_arg); |
1441 | gimple_call_set_arg (gs: call, index: 4, arg: e_mask_arg); |
1442 | break; |
1443 | |
1444 | default: |
1445 | gcc_unreachable (); |
1446 | } |
1447 | update_stmt (s: call); |
1448 | } |
1449 | |
1450 | unsigned dim = GOMP_DIM_GANG; |
1451 | unsigned mask = loop->mask | loop->e_mask; |
1452 | for (ix = 0; ix != GOMP_DIM_MAX && mask; ix++) |
1453 | { |
1454 | while (!(GOMP_DIM_MASK (dim) & mask)) |
1455 | dim++; |
1456 | |
1457 | oacc_loop_xform_head_tail (from: loop->heads[ix], level: dim); |
1458 | oacc_loop_xform_head_tail (from: loop->tails[ix], level: dim); |
1459 | |
1460 | mask ^= GOMP_DIM_MASK (dim); |
1461 | } |
1462 | } |
1463 | |
1464 | if (loop->sibling) |
1465 | oacc_loop_process (loop: loop->sibling, fn_level); |
1466 | |
1467 | |
1468 | /* OpenACC 2.6, 2.9.11. "reduction clause" places a restriction such that |
1469 | "The 'reduction' clause may not be specified on an orphaned 'loop' |
1470 | construct with the 'gang' clause, or on an orphaned 'loop' construct that |
1471 | will generate gang parallelism in a procedure that is compiled with the |
1472 | 'routine gang' clause." */ |
1473 | if (fn_level == GOMP_DIM_GANG |
1474 | && (loop->mask & GOMP_DIM_MASK (GOMP_DIM_GANG)) |
1475 | && (loop->flags & OLF_REDUCTION)) |
1476 | error_at (loop->loc, |
1477 | "gang reduction on an orphan loop" ); |
1478 | } |
1479 | |
1480 | /* Walk the OpenACC loop heirarchy checking and assigning the |
1481 | programmer-specified partitionings. OUTER_MASK is the partitioning |
1482 | this loop is contained within. Return mask of partitioning |
1483 | encountered. If any auto loops are discovered, set GOMP_DIM_MAX |
1484 | bit. */ |
1485 | |
1486 | static unsigned |
1487 | oacc_loop_fixed_partitions (oacc_loop *loop, unsigned outer_mask) |
1488 | { |
1489 | unsigned this_mask = loop->mask; |
1490 | unsigned mask_all = 0; |
1491 | bool noisy = true; |
1492 | |
1493 | #ifdef ACCEL_COMPILER |
1494 | /* When device_type is supported, we want the device compiler to be |
1495 | noisy, if the loop parameters are device_type-specific. */ |
1496 | noisy = false; |
1497 | #endif |
1498 | |
1499 | if (!loop->routine) |
1500 | { |
1501 | bool auto_par = (loop->flags & OLF_AUTO) != 0; |
1502 | bool seq_par = (loop->flags & OLF_SEQ) != 0; |
1503 | bool tiling = (loop->flags & OLF_TILE) != 0; |
1504 | |
1505 | this_mask = ((loop->flags >> OLF_DIM_BASE) |
1506 | & (GOMP_DIM_MASK (GOMP_DIM_MAX) - 1)); |
1507 | |
1508 | /* Apply auto partitioning if this is a non-partitioned regular |
1509 | loop, or (no more than) single axis tiled loop. */ |
1510 | bool maybe_auto |
1511 | = !seq_par && this_mask == (tiling ? this_mask & -this_mask : 0); |
1512 | |
1513 | if ((this_mask != 0) + auto_par + seq_par > 1) |
1514 | { |
1515 | if (noisy) |
1516 | error_at (loop->loc, |
1517 | seq_par |
1518 | ? G_("%<seq%> overrides other OpenACC loop specifiers" ) |
1519 | : G_("%<auto%> conflicts with other OpenACC loop " |
1520 | "specifiers" )); |
1521 | maybe_auto = false; |
1522 | loop->flags &= ~OLF_AUTO; |
1523 | if (seq_par) |
1524 | { |
1525 | loop->flags |
1526 | &= ~((GOMP_DIM_MASK (GOMP_DIM_MAX) - 1) << OLF_DIM_BASE); |
1527 | this_mask = 0; |
1528 | } |
1529 | } |
1530 | |
1531 | if (maybe_auto && (loop->flags & OLF_INDEPENDENT)) |
1532 | { |
1533 | loop->flags |= OLF_AUTO; |
1534 | mask_all |= GOMP_DIM_MASK (GOMP_DIM_MAX); |
1535 | } |
1536 | } |
1537 | |
1538 | if (this_mask & outer_mask) |
1539 | { |
1540 | const oacc_loop *outer; |
1541 | for (outer = loop->parent; outer; outer = outer->parent) |
1542 | if ((outer->mask | outer->e_mask) & this_mask) |
1543 | break; |
1544 | |
1545 | if (noisy) |
1546 | { |
1547 | if (outer) |
1548 | { |
1549 | error_at (loop->loc, |
1550 | loop->routine |
1551 | ? G_("routine call uses same OpenACC parallelism" |
1552 | " as containing loop" ) |
1553 | : G_("inner loop uses same OpenACC parallelism" |
1554 | " as containing loop" )); |
1555 | inform (outer->loc, "containing loop here" ); |
1556 | } |
1557 | else |
1558 | error_at (loop->loc, |
1559 | loop->routine |
1560 | ? G_("routine call uses OpenACC parallelism disallowed" |
1561 | " by containing routine" ) |
1562 | : G_("loop uses OpenACC parallelism disallowed" |
1563 | " by containing routine" )); |
1564 | |
1565 | if (loop->routine) |
1566 | inform (DECL_SOURCE_LOCATION (loop->routine), |
1567 | "routine %qD declared here" , loop->routine); |
1568 | } |
1569 | this_mask &= ~outer_mask; |
1570 | } |
1571 | else |
1572 | { |
1573 | unsigned outermost = least_bit_hwi (x: this_mask); |
1574 | |
1575 | if (outermost && outermost <= outer_mask) |
1576 | { |
1577 | if (noisy) |
1578 | { |
1579 | error_at (loop->loc, |
1580 | "incorrectly nested OpenACC loop parallelism" ); |
1581 | |
1582 | const oacc_loop *outer; |
1583 | for (outer = loop->parent; |
1584 | outer->flags && outer->flags < outermost; |
1585 | outer = outer->parent) |
1586 | continue; |
1587 | inform (outer->loc, "containing loop here" ); |
1588 | } |
1589 | |
1590 | this_mask &= ~outermost; |
1591 | } |
1592 | } |
1593 | |
1594 | mask_all |= this_mask; |
1595 | |
1596 | if (loop->flags & OLF_TILE) |
1597 | { |
1598 | /* When tiling, vector goes to the element loop, and failing |
1599 | that we put worker there. The std doesn't contemplate |
1600 | specifying all three. We choose to put worker and vector on |
1601 | the element loops in that case. */ |
1602 | unsigned this_e_mask = this_mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR); |
1603 | if (!this_e_mask || this_mask & GOMP_DIM_MASK (GOMP_DIM_GANG)) |
1604 | this_e_mask |= this_mask & GOMP_DIM_MASK (GOMP_DIM_WORKER); |
1605 | |
1606 | loop->e_mask = this_e_mask; |
1607 | this_mask ^= this_e_mask; |
1608 | } |
1609 | |
1610 | loop->mask = this_mask; |
1611 | |
1612 | if (dump_file) |
1613 | fprintf (stream: dump_file, format: "Loop %s:%d user specified %d & %d\n" , |
1614 | LOCATION_FILE (loop->loc), LOCATION_LINE (loop->loc), |
1615 | loop->mask, loop->e_mask); |
1616 | |
1617 | if (loop->child) |
1618 | { |
1619 | unsigned tmp_mask = outer_mask | this_mask | loop->e_mask; |
1620 | loop->inner = oacc_loop_fixed_partitions (loop: loop->child, outer_mask: tmp_mask); |
1621 | mask_all |= loop->inner; |
1622 | } |
1623 | |
1624 | if (loop->sibling) |
1625 | mask_all |= oacc_loop_fixed_partitions (loop: loop->sibling, outer_mask); |
1626 | |
1627 | return mask_all; |
1628 | } |
1629 | |
1630 | /* Walk the OpenACC loop heirarchy to assign auto-partitioned loops. |
1631 | OUTER_MASK is the partitioning this loop is contained within. |
1632 | OUTER_ASSIGN is true if an outer loop is being auto-partitioned. |
1633 | Return the cumulative partitioning used by this loop, siblings and |
1634 | children. */ |
1635 | |
1636 | static unsigned |
1637 | oacc_loop_auto_partitions (oacc_loop *loop, unsigned outer_mask, |
1638 | bool outer_assign) |
1639 | { |
1640 | bool assign = (loop->flags & OLF_AUTO) && (loop->flags & OLF_INDEPENDENT); |
1641 | bool noisy = true; |
1642 | bool tiling = loop->flags & OLF_TILE; |
1643 | |
1644 | #ifdef ACCEL_COMPILER |
1645 | /* When device_type is supported, we want the device compiler to be |
1646 | noisy, if the loop parameters are device_type-specific. */ |
1647 | noisy = false; |
1648 | #endif |
1649 | |
1650 | if (assign && (!outer_assign || loop->inner)) |
1651 | { |
1652 | /* Allocate outermost and non-innermost loops at the outermost |
1653 | non-innermost available level. */ |
1654 | unsigned this_mask = GOMP_DIM_MASK (GOMP_DIM_GANG); |
1655 | |
1656 | /* Find the first outermost available partition. */ |
1657 | while (this_mask <= outer_mask) |
1658 | this_mask <<= 1; |
1659 | |
1660 | /* Grab two axes if tiling, and we've not assigned anything */ |
1661 | if (tiling && !(loop->mask | loop->e_mask)) |
1662 | this_mask |= this_mask << 1; |
1663 | |
1664 | /* Prohibit the innermost partitioning at the moment. */ |
1665 | this_mask &= GOMP_DIM_MASK (GOMP_DIM_MAX - 1) - 1; |
1666 | |
1667 | /* Don't use any dimension explicitly claimed by an inner loop. */ |
1668 | this_mask &= ~loop->inner; |
1669 | |
1670 | if (tiling && !loop->e_mask) |
1671 | { |
1672 | /* If we got two axes, allocate the inner one to the element |
1673 | loop. */ |
1674 | loop->e_mask = this_mask & (this_mask << 1); |
1675 | this_mask ^= loop->e_mask; |
1676 | } |
1677 | |
1678 | loop->mask |= this_mask; |
1679 | } |
1680 | |
1681 | if (loop->child) |
1682 | { |
1683 | unsigned tmp_mask = outer_mask | loop->mask | loop->e_mask; |
1684 | loop->inner = oacc_loop_auto_partitions (loop: loop->child, outer_mask: tmp_mask, |
1685 | outer_assign: outer_assign | assign); |
1686 | } |
1687 | |
1688 | if (assign && (!loop->mask || (tiling && !loop->e_mask) || !outer_assign)) |
1689 | { |
1690 | /* Allocate the loop at the innermost available level. Note |
1691 | that we do this even if we already assigned this loop the |
1692 | outermost available level above. That way we'll partition |
1693 | this along 2 axes, if they are available. */ |
1694 | unsigned this_mask = 0; |
1695 | |
1696 | /* Determine the outermost partitioning used within this loop. */ |
1697 | this_mask = loop->inner | GOMP_DIM_MASK (GOMP_DIM_MAX); |
1698 | this_mask = least_bit_hwi (x: this_mask); |
1699 | |
1700 | /* Pick the partitioning just inside that one. */ |
1701 | this_mask >>= 1; |
1702 | |
1703 | /* And avoid picking one use by an outer loop. */ |
1704 | this_mask &= ~outer_mask; |
1705 | |
1706 | /* If tiling and we failed completely above, grab the next one |
1707 | too. Making sure it doesn't hit an outer loop. */ |
1708 | if (tiling) |
1709 | { |
1710 | this_mask &= ~(loop->e_mask | loop->mask); |
1711 | unsigned tile_mask = ((this_mask >> 1) |
1712 | & ~(outer_mask | loop->e_mask | loop->mask)); |
1713 | |
1714 | if (tile_mask || loop->mask) |
1715 | { |
1716 | loop->e_mask |= this_mask; |
1717 | this_mask = tile_mask; |
1718 | } |
1719 | if (!loop->e_mask && noisy) |
1720 | warning_at (loop->loc, 0, |
1721 | "insufficient partitioning available" |
1722 | " to parallelize element loop" ); |
1723 | } |
1724 | |
1725 | loop->mask |= this_mask; |
1726 | if (!loop->mask && noisy) |
1727 | warning_at (loop->loc, 0, |
1728 | tiling |
1729 | ? G_("insufficient partitioning available" |
1730 | " to parallelize tile loop" ) |
1731 | : G_("insufficient partitioning available" |
1732 | " to parallelize loop" )); |
1733 | } |
1734 | |
1735 | if (assign && dump_file) |
1736 | fprintf (stream: dump_file, format: "Auto loop %s:%d assigned %d & %d\n" , |
1737 | LOCATION_FILE (loop->loc), LOCATION_LINE (loop->loc), |
1738 | loop->mask, loop->e_mask); |
1739 | |
1740 | unsigned inner_mask = 0; |
1741 | |
1742 | if (loop->sibling) |
1743 | inner_mask |= oacc_loop_auto_partitions (loop: loop->sibling, |
1744 | outer_mask, outer_assign); |
1745 | |
1746 | inner_mask |= loop->inner | loop->mask | loop->e_mask; |
1747 | |
1748 | return inner_mask; |
1749 | } |
1750 | |
1751 | /* Walk the OpenACC loop heirarchy to check and assign partitioning |
1752 | axes. Return mask of partitioning. */ |
1753 | |
1754 | static unsigned |
1755 | oacc_loop_partition (oacc_loop *loop, unsigned outer_mask) |
1756 | { |
1757 | unsigned mask_all = oacc_loop_fixed_partitions (loop, outer_mask); |
1758 | |
1759 | if (mask_all & GOMP_DIM_MASK (GOMP_DIM_MAX)) |
1760 | { |
1761 | mask_all ^= GOMP_DIM_MASK (GOMP_DIM_MAX); |
1762 | mask_all |= oacc_loop_auto_partitions (loop, outer_mask, outer_assign: false); |
1763 | } |
1764 | return mask_all; |
1765 | } |
1766 | |
1767 | /* Default fork/join early expander. Delete the function calls if |
1768 | there is no RTL expander. */ |
1769 | |
1770 | bool |
1771 | default_goacc_fork_join (gcall *ARG_UNUSED (call), |
1772 | const int *ARG_UNUSED (dims), bool is_fork) |
1773 | { |
1774 | if (is_fork) |
1775 | return targetm.have_oacc_fork (); |
1776 | else |
1777 | return targetm.have_oacc_join (); |
1778 | } |
1779 | |
1780 | /* Default goacc.reduction early expander. |
1781 | |
1782 | LHS-opt = IFN_REDUCTION (KIND, RES_PTR, VAR, LEVEL, OP, OFFSET) |
1783 | If RES_PTR is not integer-zerop: |
1784 | SETUP - emit 'LHS = *RES_PTR', LHS = NULL |
1785 | TEARDOWN - emit '*RES_PTR = VAR' |
1786 | If LHS is not NULL |
1787 | emit 'LHS = VAR' */ |
1788 | |
1789 | void |
1790 | default_goacc_reduction (gcall *call) |
1791 | { |
1792 | unsigned code = (unsigned)TREE_INT_CST_LOW (gimple_call_arg (call, 0)); |
1793 | gimple_stmt_iterator gsi = gsi_for_stmt (call); |
1794 | tree lhs = gimple_call_lhs (gs: call); |
1795 | tree var = gimple_call_arg (gs: call, index: 2); |
1796 | gimple_seq seq = NULL; |
1797 | |
1798 | if (code == IFN_GOACC_REDUCTION_SETUP |
1799 | || code == IFN_GOACC_REDUCTION_TEARDOWN) |
1800 | { |
1801 | /* Setup and Teardown need to copy from/to the receiver object, |
1802 | if there is one. */ |
1803 | tree ref_to_res = gimple_call_arg (gs: call, index: 1); |
1804 | |
1805 | if (!integer_zerop (ref_to_res)) |
1806 | { |
1807 | tree dst = build_simple_mem_ref (ref_to_res); |
1808 | tree src = var; |
1809 | |
1810 | if (code == IFN_GOACC_REDUCTION_SETUP) |
1811 | { |
1812 | src = dst; |
1813 | dst = lhs; |
1814 | lhs = NULL; |
1815 | } |
1816 | gimple_seq_add_stmt (&seq, gimple_build_assign (dst, src)); |
1817 | } |
1818 | } |
1819 | |
1820 | /* Copy VAR to LHS, if there is an LHS. */ |
1821 | if (lhs) |
1822 | gimple_seq_add_stmt (&seq, gimple_build_assign (lhs, var)); |
1823 | |
1824 | gsi_replace_with_seq (&gsi, seq, true); |
1825 | } |
1826 | |
1827 | struct var_decl_rewrite_info |
1828 | { |
1829 | gimple *stmt; |
1830 | hash_map<tree, tree> *adjusted_vars; |
1831 | bool avoid_pointer_conversion; |
1832 | bool modified; |
1833 | }; |
1834 | |
1835 | /* Helper function for execute_oacc_device_lower. Rewrite VAR_DECLs (by |
1836 | themselves or wrapped in various other nodes) according to ADJUSTED_VARS in |
1837 | the var_decl_rewrite_info pointed to via DATA. Used as part of coercing |
1838 | gang-private variables in OpenACC offload regions to reside in GPU shared |
1839 | memory. */ |
1840 | |
1841 | static tree |
1842 | oacc_rewrite_var_decl (tree *tp, int *walk_subtrees, void *data) |
1843 | { |
1844 | walk_stmt_info *wi = (walk_stmt_info *) data; |
1845 | var_decl_rewrite_info *info = (var_decl_rewrite_info *) wi->info; |
1846 | |
1847 | if (TREE_CODE (*tp) == ADDR_EXPR) |
1848 | { |
1849 | tree arg = TREE_OPERAND (*tp, 0); |
1850 | tree *new_arg = info->adjusted_vars->get (k: arg); |
1851 | |
1852 | if (new_arg) |
1853 | { |
1854 | if (info->avoid_pointer_conversion) |
1855 | { |
1856 | *tp = build_fold_addr_expr (*new_arg); |
1857 | info->modified = true; |
1858 | *walk_subtrees = 0; |
1859 | } |
1860 | else |
1861 | { |
1862 | gimple_stmt_iterator gsi = gsi_for_stmt (info->stmt); |
1863 | tree repl = build_fold_addr_expr (*new_arg); |
1864 | gimple *stmt1 |
1865 | = gimple_build_assign (make_ssa_name (TREE_TYPE (repl)), repl); |
1866 | tree conv = convert_to_pointer (TREE_TYPE (*tp), |
1867 | gimple_assign_lhs (gs: stmt1)); |
1868 | gimple *stmt2 |
1869 | = gimple_build_assign (make_ssa_name (TREE_TYPE (*tp)), conv); |
1870 | gsi_insert_before (&gsi, stmt1, GSI_SAME_STMT); |
1871 | gsi_insert_before (&gsi, stmt2, GSI_SAME_STMT); |
1872 | *tp = gimple_assign_lhs (gs: stmt2); |
1873 | info->modified = true; |
1874 | *walk_subtrees = 0; |
1875 | } |
1876 | } |
1877 | } |
1878 | else if (TREE_CODE (*tp) == COMPONENT_REF || TREE_CODE (*tp) == ARRAY_REF) |
1879 | { |
1880 | tree *base = &TREE_OPERAND (*tp, 0); |
1881 | |
1882 | while (TREE_CODE (*base) == COMPONENT_REF |
1883 | || TREE_CODE (*base) == ARRAY_REF) |
1884 | base = &TREE_OPERAND (*base, 0); |
1885 | |
1886 | if (TREE_CODE (*base) != VAR_DECL) |
1887 | return NULL; |
1888 | |
1889 | tree *new_decl = info->adjusted_vars->get (k: *base); |
1890 | if (!new_decl) |
1891 | return NULL; |
1892 | |
1893 | int base_quals = TYPE_QUALS (TREE_TYPE (*new_decl)); |
1894 | tree field = TREE_OPERAND (*tp, 1); |
1895 | |
1896 | /* Adjust the type of the field. */ |
1897 | int field_quals = TYPE_QUALS (TREE_TYPE (field)); |
1898 | if (TREE_CODE (field) == FIELD_DECL && field_quals != base_quals) |
1899 | { |
1900 | tree *field_type = &TREE_TYPE (field); |
1901 | while (TREE_CODE (*field_type) == ARRAY_TYPE) |
1902 | field_type = &TREE_TYPE (*field_type); |
1903 | field_quals |= base_quals; |
1904 | *field_type = build_qualified_type (*field_type, field_quals); |
1905 | } |
1906 | |
1907 | /* Adjust the type of the component ref itself. */ |
1908 | tree comp_type = TREE_TYPE (*tp); |
1909 | int comp_quals = TYPE_QUALS (comp_type); |
1910 | if (TREE_CODE (*tp) == COMPONENT_REF && comp_quals != base_quals) |
1911 | { |
1912 | comp_quals |= base_quals; |
1913 | TREE_TYPE (*tp) |
1914 | = build_qualified_type (comp_type, comp_quals); |
1915 | } |
1916 | |
1917 | *base = *new_decl; |
1918 | info->modified = true; |
1919 | } |
1920 | else if (VAR_P (*tp)) |
1921 | { |
1922 | tree *new_decl = info->adjusted_vars->get (k: *tp); |
1923 | if (new_decl) |
1924 | { |
1925 | *tp = *new_decl; |
1926 | info->modified = true; |
1927 | } |
1928 | } |
1929 | |
1930 | return NULL_TREE; |
1931 | } |
1932 | |
1933 | /* Return TRUE if CALL is a call to a builtin atomic/sync operation. */ |
1934 | |
1935 | static bool |
1936 | is_sync_builtin_call (gcall *call) |
1937 | { |
1938 | tree callee = gimple_call_fndecl (gs: call); |
1939 | |
1940 | if (callee != NULL_TREE |
1941 | && gimple_call_builtin_p (call, BUILT_IN_NORMAL)) |
1942 | switch (DECL_FUNCTION_CODE (decl: callee)) |
1943 | { |
1944 | #undef DEF_SYNC_BUILTIN |
1945 | #define DEF_SYNC_BUILTIN(ENUM, NAME, TYPE, ATTRS) case ENUM: |
1946 | #include "sync-builtins.def" |
1947 | #undef DEF_SYNC_BUILTIN |
1948 | return true; |
1949 | |
1950 | default: |
1951 | ; |
1952 | } |
1953 | |
1954 | return false; |
1955 | } |
1956 | |
1957 | /* Main entry point for oacc transformations which run on the device |
1958 | compiler after LTO, so we know what the target device is at this |
1959 | point (including the host fallback). */ |
1960 | |
1961 | static unsigned int |
1962 | execute_oacc_loop_designation () |
1963 | { |
1964 | tree attrs = oacc_get_fn_attrib (fn: current_function_decl); |
1965 | |
1966 | if (!attrs) |
1967 | /* Not an offloaded function. */ |
1968 | return 0; |
1969 | |
1970 | /* Parse the default dim argument exactly once. */ |
1971 | if ((const void *)flag_openacc_dims != &flag_openacc_dims) |
1972 | { |
1973 | oacc_parse_default_dims (flag_openacc_dims); |
1974 | flag_openacc_dims = (char *)&flag_openacc_dims; |
1975 | } |
1976 | |
1977 | bool is_oacc_parallel |
1978 | = (lookup_attribute (attr_name: "oacc parallel" , |
1979 | DECL_ATTRIBUTES (current_function_decl)) != NULL); |
1980 | bool is_oacc_kernels |
1981 | = (lookup_attribute (attr_name: "oacc kernels" , |
1982 | DECL_ATTRIBUTES (current_function_decl)) != NULL); |
1983 | bool is_oacc_serial |
1984 | = (lookup_attribute (attr_name: "oacc serial" , |
1985 | DECL_ATTRIBUTES (current_function_decl)) != NULL); |
1986 | bool is_oacc_parallel_kernels_parallelized |
1987 | = (lookup_attribute (attr_name: "oacc parallel_kernels_parallelized" , |
1988 | DECL_ATTRIBUTES (current_function_decl)) != NULL); |
1989 | bool is_oacc_parallel_kernels_gang_single |
1990 | = (lookup_attribute (attr_name: "oacc parallel_kernels_gang_single" , |
1991 | DECL_ATTRIBUTES (current_function_decl)) != NULL); |
1992 | int fn_level = oacc_fn_attrib_level (attr: attrs); |
1993 | bool is_oacc_routine = (fn_level >= 0); |
1994 | gcc_checking_assert (is_oacc_parallel |
1995 | + is_oacc_kernels |
1996 | + is_oacc_serial |
1997 | + is_oacc_parallel_kernels_parallelized |
1998 | + is_oacc_parallel_kernels_gang_single |
1999 | + is_oacc_routine |
2000 | == 1); |
2001 | |
2002 | bool is_oacc_kernels_parallelized |
2003 | = (lookup_attribute (attr_name: "oacc kernels parallelized" , |
2004 | DECL_ATTRIBUTES (current_function_decl)) != NULL); |
2005 | if (is_oacc_kernels_parallelized) |
2006 | gcc_checking_assert (is_oacc_kernels); |
2007 | |
2008 | if (dump_file) |
2009 | { |
2010 | if (is_oacc_parallel) |
2011 | fprintf (stream: dump_file, format: "Function is OpenACC parallel offload\n" ); |
2012 | else if (is_oacc_kernels) |
2013 | fprintf (stream: dump_file, format: "Function is %s OpenACC kernels offload\n" , |
2014 | (is_oacc_kernels_parallelized |
2015 | ? "parallelized" : "unparallelized" )); |
2016 | else if (is_oacc_serial) |
2017 | fprintf (stream: dump_file, format: "Function is OpenACC serial offload\n" ); |
2018 | else if (is_oacc_parallel_kernels_parallelized) |
2019 | fprintf (stream: dump_file, format: "Function is %s OpenACC kernels offload\n" , |
2020 | "parallel_kernels_parallelized" ); |
2021 | else if (is_oacc_parallel_kernels_gang_single) |
2022 | fprintf (stream: dump_file, format: "Function is %s OpenACC kernels offload\n" , |
2023 | "parallel_kernels_gang_single" ); |
2024 | else if (is_oacc_routine) |
2025 | fprintf (stream: dump_file, format: "Function is OpenACC routine level %d\n" , |
2026 | fn_level); |
2027 | else |
2028 | gcc_unreachable (); |
2029 | } |
2030 | |
2031 | /* This doesn't belong into 'pass_oacc_loop_designation' conceptually, but |
2032 | it's a convenient place, so... */ |
2033 | if (is_oacc_routine) |
2034 | { |
2035 | tree attr = lookup_attribute (attr_name: "omp declare target" , |
2036 | DECL_ATTRIBUTES (current_function_decl)); |
2037 | gcc_checking_assert (attr); |
2038 | tree clauses = TREE_VALUE (attr); |
2039 | gcc_checking_assert (clauses); |
2040 | |
2041 | /* Should this OpenACC routine be discarded? */ |
2042 | bool discard = false; |
2043 | |
2044 | tree clause_nohost = omp_find_clause (clauses, kind: OMP_CLAUSE_NOHOST); |
2045 | if (dump_file) |
2046 | fprintf (stream: dump_file, |
2047 | format: "OpenACC routine '%s' %s '%s' clause.\n" , |
2048 | lang_hooks.decl_printable_name (current_function_decl, 2), |
2049 | clause_nohost ? "has" : "doesn't have" , |
2050 | omp_clause_code_name[OMP_CLAUSE_NOHOST]); |
2051 | /* Host compiler, 'nohost' clause? */ |
2052 | #ifndef ACCEL_COMPILER |
2053 | if (clause_nohost) |
2054 | discard = true; |
2055 | #endif |
2056 | |
2057 | if (dump_file) |
2058 | fprintf (stream: dump_file, |
2059 | format: "OpenACC routine '%s' %sdiscarded.\n" , |
2060 | lang_hooks.decl_printable_name (current_function_decl, 2), |
2061 | discard ? "" : "not " ); |
2062 | if (discard) |
2063 | { |
2064 | TREE_ASM_WRITTEN (current_function_decl) = 1; |
2065 | return TODO_discard_function; |
2066 | } |
2067 | } |
2068 | |
2069 | /* Unparallelized OpenACC kernels constructs must get launched as 1 x 1 x 1 |
2070 | kernels, so remove the parallelism dimensions function attributes |
2071 | potentially set earlier on. */ |
2072 | if (is_oacc_kernels && !is_oacc_kernels_parallelized) |
2073 | { |
2074 | oacc_set_fn_attrib (fn: current_function_decl, NULL, NULL); |
2075 | attrs = oacc_get_fn_attrib (fn: current_function_decl); |
2076 | } |
2077 | |
2078 | /* Discover, partition and process the loops. */ |
2079 | oacc_loop *loops = oacc_loop_discovery (); |
2080 | |
2081 | unsigned outer_mask = 0; |
2082 | if (is_oacc_routine) |
2083 | outer_mask = GOMP_DIM_MASK (fn_level) - 1; |
2084 | unsigned used_mask = oacc_loop_partition (loop: loops, outer_mask); |
2085 | /* OpenACC kernels constructs are special: they currently don't use the |
2086 | generic oacc_loop infrastructure and attribute/dimension processing. */ |
2087 | if (is_oacc_kernels && is_oacc_kernels_parallelized) |
2088 | { |
2089 | /* Parallelized OpenACC kernels constructs use gang parallelism. See |
2090 | also tree-parloops.cc:create_parallel_loop. */ |
2091 | used_mask |= GOMP_DIM_MASK (GOMP_DIM_GANG); |
2092 | } |
2093 | |
2094 | int dims[GOMP_DIM_MAX]; |
2095 | oacc_validate_dims (fn: current_function_decl, attrs, dims, level: fn_level, used: used_mask); |
2096 | |
2097 | if (dump_file) |
2098 | { |
2099 | const char *comma = "Compute dimensions [" ; |
2100 | for (int ix = 0; ix != GOMP_DIM_MAX; ix++, comma = ", " ) |
2101 | fprintf (stream: dump_file, format: "%s%d" , comma, dims[ix]); |
2102 | fprintf (stream: dump_file, format: "]\n" ); |
2103 | } |
2104 | |
2105 | /* Verify that for OpenACC 'kernels' decomposed "gang-single" parts we launch |
2106 | a single gang only. */ |
2107 | if (is_oacc_parallel_kernels_gang_single) |
2108 | gcc_checking_assert (dims[GOMP_DIM_GANG] == 1); |
2109 | |
2110 | oacc_loop_process (loop: loops, fn_level); |
2111 | if (dump_file) |
2112 | { |
2113 | fprintf (stream: dump_file, format: "OpenACC loops\n" ); |
2114 | dump_oacc_loop (file: dump_file, loop: loops, depth: 0); |
2115 | fprintf (stream: dump_file, format: "\n" ); |
2116 | } |
2117 | if (dump_enabled_p ()) |
2118 | { |
2119 | oacc_loop *l = loops; |
2120 | /* OpenACC kernels constructs are special: they currently don't use the |
2121 | generic oacc_loop infrastructure. */ |
2122 | if (is_oacc_kernels) |
2123 | { |
2124 | /* Create a fake oacc_loop for diagnostic purposes. */ |
2125 | l = new_oacc_loop_raw (NULL, |
2126 | DECL_SOURCE_LOCATION (current_function_decl)); |
2127 | l->mask = used_mask; |
2128 | } |
2129 | else |
2130 | { |
2131 | /* Skip the outermost, dummy OpenACC loop */ |
2132 | l = l->child; |
2133 | } |
2134 | if (l) |
2135 | inform_oacc_loop (loop: l); |
2136 | if (is_oacc_kernels) |
2137 | free_oacc_loop (loop: l); |
2138 | } |
2139 | |
2140 | free_oacc_loop (loop: loops); |
2141 | |
2142 | return 0; |
2143 | } |
2144 | |
2145 | static unsigned int |
2146 | execute_oacc_device_lower () |
2147 | { |
2148 | tree attrs = oacc_get_fn_attrib (fn: current_function_decl); |
2149 | |
2150 | if (!attrs) |
2151 | /* Not an offloaded function. */ |
2152 | return 0; |
2153 | |
2154 | int dims[GOMP_DIM_MAX]; |
2155 | for (unsigned i = 0; i < GOMP_DIM_MAX; i++) |
2156 | dims[i] = oacc_get_fn_dim_size (fn: current_function_decl, axis: i); |
2157 | |
2158 | hash_map<tree, tree> adjusted_vars; |
2159 | |
2160 | /* Now lower internal loop functions to target-specific code |
2161 | sequences. */ |
2162 | basic_block bb; |
2163 | FOR_ALL_BB_FN (bb, cfun) |
2164 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi);) |
2165 | { |
2166 | gimple *stmt = gsi_stmt (i: gsi); |
2167 | if (!is_gimple_call (gs: stmt)) |
2168 | { |
2169 | gsi_next (i: &gsi); |
2170 | continue; |
2171 | } |
2172 | |
2173 | gcall *call = as_a <gcall *> (p: stmt); |
2174 | if (!gimple_call_internal_p (gs: call)) |
2175 | { |
2176 | gsi_next (i: &gsi); |
2177 | continue; |
2178 | } |
2179 | |
2180 | /* Rewind to allow rescan. */ |
2181 | gsi_prev (i: &gsi); |
2182 | bool rescan = false, remove = false; |
2183 | enum internal_fn ifn_code = gimple_call_internal_fn (gs: call); |
2184 | |
2185 | switch (ifn_code) |
2186 | { |
2187 | default: break; |
2188 | |
2189 | case IFN_GOACC_TILE: |
2190 | oacc_xform_tile (call); |
2191 | rescan = true; |
2192 | break; |
2193 | |
2194 | case IFN_GOACC_LOOP: |
2195 | oacc_xform_loop (call); |
2196 | rescan = true; |
2197 | break; |
2198 | |
2199 | case IFN_GOACC_REDUCTION: |
2200 | /* Mark the function for SSA renaming. */ |
2201 | mark_virtual_operands_for_renaming (cfun); |
2202 | |
2203 | /* If the level is -1, this ended up being an unused |
2204 | axis. Handle as a default. */ |
2205 | if (integer_minus_onep (gimple_call_arg (gs: call, index: 3))) |
2206 | default_goacc_reduction (call); |
2207 | else |
2208 | targetm.goacc.reduction (call); |
2209 | rescan = true; |
2210 | break; |
2211 | |
2212 | case IFN_UNIQUE: |
2213 | { |
2214 | enum ifn_unique_kind kind |
2215 | = ((enum ifn_unique_kind) |
2216 | TREE_INT_CST_LOW (gimple_call_arg (call, 0))); |
2217 | |
2218 | switch (kind) |
2219 | { |
2220 | default: |
2221 | break; |
2222 | |
2223 | case IFN_UNIQUE_OACC_FORK: |
2224 | case IFN_UNIQUE_OACC_JOIN: |
2225 | if (integer_minus_onep (gimple_call_arg (gs: call, index: 2))) |
2226 | remove = true; |
2227 | else if (!targetm.goacc.fork_join |
2228 | (call, dims, kind == IFN_UNIQUE_OACC_FORK)) |
2229 | remove = true; |
2230 | break; |
2231 | |
2232 | case IFN_UNIQUE_OACC_HEAD_MARK: |
2233 | case IFN_UNIQUE_OACC_TAIL_MARK: |
2234 | remove = true; |
2235 | break; |
2236 | |
2237 | case IFN_UNIQUE_OACC_PRIVATE: |
2238 | { |
2239 | dump_flags_t l_dump_flags |
2240 | = get_openacc_privatization_dump_flags (); |
2241 | |
2242 | location_t loc = gimple_location (g: stmt); |
2243 | if (LOCATION_LOCUS (loc) == UNKNOWN_LOCATION) |
2244 | loc = DECL_SOURCE_LOCATION (current_function_decl); |
2245 | const dump_user_location_t d_u_loc |
2246 | = dump_user_location_t::from_location_t (loc); |
2247 | |
2248 | HOST_WIDE_INT level |
2249 | = TREE_INT_CST_LOW (gimple_call_arg (call, 2)); |
2250 | gcc_checking_assert (level == -1 |
2251 | || (level >= 0 |
2252 | && level < GOMP_DIM_MAX)); |
2253 | for (unsigned i = 3; |
2254 | i < gimple_call_num_args (gs: call); |
2255 | i++) |
2256 | { |
2257 | static char const *const axes[] = |
2258 | /* Must be kept in sync with GOMP_DIM enumeration. */ |
2259 | { "gang" , "worker" , "vector" }; |
2260 | |
2261 | tree arg = gimple_call_arg (gs: call, index: i); |
2262 | gcc_checking_assert (TREE_CODE (arg) == ADDR_EXPR); |
2263 | tree decl = TREE_OPERAND (arg, 0); |
2264 | if (dump_enabled_p ()) |
2265 | /* PR100695 "Format decoder, quoting in 'dump_printf' etc." */ |
2266 | #if __GNUC__ >= 10 |
2267 | # pragma GCC diagnostic push |
2268 | # pragma GCC diagnostic ignored "-Wformat" |
2269 | #endif |
2270 | dump_printf_loc (l_dump_flags, d_u_loc, |
2271 | "variable %<%T%> ought to be" |
2272 | " adjusted for OpenACC" |
2273 | " privatization level: %qs\n" , |
2274 | decl, |
2275 | (level == -1 |
2276 | ? "UNKNOWN" : axes[level])); |
2277 | #if __GNUC__ >= 10 |
2278 | # pragma GCC diagnostic pop |
2279 | #endif |
2280 | bool adjusted; |
2281 | if (level == -1) |
2282 | adjusted = false; |
2283 | else if (!targetm.goacc.adjust_private_decl) |
2284 | adjusted = false; |
2285 | else if (level == GOMP_DIM_VECTOR) |
2286 | { |
2287 | /* That's the default behavior. */ |
2288 | adjusted = true; |
2289 | } |
2290 | else |
2291 | { |
2292 | tree oldtype = TREE_TYPE (decl); |
2293 | tree newdecl |
2294 | = targetm.goacc.adjust_private_decl (loc, decl, |
2295 | level); |
2296 | adjusted = (TREE_TYPE (newdecl) != oldtype |
2297 | || newdecl != decl); |
2298 | if (adjusted) |
2299 | adjusted_vars.put (k: decl, v: newdecl); |
2300 | } |
2301 | if (adjusted |
2302 | && dump_enabled_p ()) |
2303 | /* PR100695 "Format decoder, quoting in 'dump_printf' etc." */ |
2304 | #if __GNUC__ >= 10 |
2305 | # pragma GCC diagnostic push |
2306 | # pragma GCC diagnostic ignored "-Wformat" |
2307 | #endif |
2308 | dump_printf_loc (l_dump_flags, d_u_loc, |
2309 | "variable %<%T%> adjusted for" |
2310 | " OpenACC privatization level:" |
2311 | " %qs\n" , |
2312 | decl, axes[level]); |
2313 | #if __GNUC__ >= 10 |
2314 | # pragma GCC diagnostic pop |
2315 | #endif |
2316 | } |
2317 | remove = true; |
2318 | } |
2319 | break; |
2320 | } |
2321 | break; |
2322 | } |
2323 | } |
2324 | |
2325 | if (gsi_end_p (i: gsi)) |
2326 | /* We rewound past the beginning of the BB. */ |
2327 | gsi = gsi_start_bb (bb); |
2328 | else |
2329 | /* Undo the rewind. */ |
2330 | gsi_next (i: &gsi); |
2331 | |
2332 | if (remove) |
2333 | { |
2334 | if (gimple_vdef (g: call)) |
2335 | replace_uses_by (gimple_vdef (g: call), gimple_vuse (g: call)); |
2336 | if (gimple_call_lhs (gs: call)) |
2337 | { |
2338 | /* Propagate the data dependency var. */ |
2339 | gimple *ass = gimple_build_assign (gimple_call_lhs (gs: call), |
2340 | gimple_call_arg (gs: call, index: 1)); |
2341 | gsi_replace (&gsi, ass, false); |
2342 | } |
2343 | else |
2344 | gsi_remove (&gsi, true); |
2345 | } |
2346 | else if (!rescan) |
2347 | /* If not rescanning, advance over the call. */ |
2348 | gsi_next (i: &gsi); |
2349 | } |
2350 | |
2351 | /* Regarding the OpenACC privatization level, we're currently only looking at |
2352 | making the gang-private level work. Regarding that, we have the following |
2353 | configurations: |
2354 | |
2355 | - GCN offloading: 'targetm.goacc.adjust_private_decl' does the work (in |
2356 | particular, change 'TREE_TYPE', etc.) and there is no |
2357 | 'targetm.goacc.expand_var_decl'. |
2358 | |
2359 | - nvptx offloading: 'targetm.goacc.adjust_private_decl' only sets a |
2360 | marker and then 'targetm.goacc.expand_var_decl' does the work. |
2361 | |
2362 | Eventually (in particular, for worker-private level?), both |
2363 | 'targetm.goacc.adjust_private_decl' and 'targetm.goacc.expand_var_decl' |
2364 | may need to do things, but that's currently not meant to be addressed, and |
2365 | thus not fully worked out and implemented, and thus untested. Hence, |
2366 | 'assert' what currently is implemented/tested, only. */ |
2367 | |
2368 | if (targetm.goacc.expand_var_decl) |
2369 | gcc_assert (adjusted_vars.is_empty ()); |
2370 | |
2371 | /* Make adjustments to gang-private local variables if required by the |
2372 | target, e.g. forcing them into a particular address space. Afterwards, |
2373 | ADDR_EXPR nodes which have adjusted variables as their argument need to |
2374 | be modified in one of two ways: |
2375 | |
2376 | 1. They can be recreated, making a pointer to the variable in the new |
2377 | address space, or |
2378 | |
2379 | 2. The address of the variable in the new address space can be taken, |
2380 | converted to the default (original) address space, and the result of |
2381 | that conversion subsituted in place of the original ADDR_EXPR node. |
2382 | |
2383 | Which of these is done depends on the gimple statement being processed. |
2384 | At present atomic operations and inline asms use (1), and everything else |
2385 | uses (2). At least on AMD GCN, there are atomic operations that work |
2386 | directly in the LDS address space. |
2387 | |
2388 | COMPONENT_REFS, ARRAY_REFS and plain VAR_DECLs are also rewritten to use |
2389 | the new decl, adjusting types of appropriate tree nodes as necessary. */ |
2390 | |
2391 | if (targetm.goacc.adjust_private_decl |
2392 | && !adjusted_vars.is_empty ()) |
2393 | { |
2394 | FOR_ALL_BB_FN (bb, cfun) |
2395 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); |
2396 | !gsi_end_p (i: gsi); |
2397 | gsi_next (i: &gsi)) |
2398 | { |
2399 | gimple *stmt = gsi_stmt (i: gsi); |
2400 | walk_stmt_info wi; |
2401 | var_decl_rewrite_info info; |
2402 | |
2403 | info.avoid_pointer_conversion |
2404 | = (is_gimple_call (gs: stmt) |
2405 | && is_sync_builtin_call (call: as_a <gcall *> (p: stmt))) |
2406 | || gimple_code (g: stmt) == GIMPLE_ASM; |
2407 | info.stmt = stmt; |
2408 | info.modified = false; |
2409 | info.adjusted_vars = &adjusted_vars; |
2410 | |
2411 | memset (s: &wi, c: 0, n: sizeof (wi)); |
2412 | wi.info = &info; |
2413 | |
2414 | walk_gimple_op (stmt, oacc_rewrite_var_decl, &wi); |
2415 | |
2416 | if (info.modified) |
2417 | update_stmt (s: stmt); |
2418 | } |
2419 | } |
2420 | |
2421 | return 0; |
2422 | } |
2423 | |
2424 | /* Default launch dimension validator. Force everything to 1. A |
2425 | backend that wants to provide larger dimensions must override this |
2426 | hook. */ |
2427 | |
2428 | bool |
2429 | default_goacc_validate_dims (tree ARG_UNUSED (decl), int *dims, |
2430 | int ARG_UNUSED (fn_level), |
2431 | unsigned ARG_UNUSED (used)) |
2432 | { |
2433 | bool changed = false; |
2434 | |
2435 | for (unsigned ix = 0; ix != GOMP_DIM_MAX; ix++) |
2436 | { |
2437 | if (dims[ix] != 1) |
2438 | { |
2439 | dims[ix] = 1; |
2440 | changed = true; |
2441 | } |
2442 | } |
2443 | |
2444 | return changed; |
2445 | } |
2446 | |
2447 | /* Default dimension bound is unknown on accelerator and 1 on host. */ |
2448 | |
2449 | int |
2450 | default_goacc_dim_limit (int ARG_UNUSED (axis)) |
2451 | { |
2452 | #ifdef ACCEL_COMPILER |
2453 | return 0; |
2454 | #else |
2455 | return 1; |
2456 | #endif |
2457 | } |
2458 | |
2459 | namespace { |
2460 | |
2461 | const pass_data pass_data_oacc_loop_designation = |
2462 | { |
2463 | .type: GIMPLE_PASS, /* type */ |
2464 | .name: "oaccloops" , /* name */ |
2465 | .optinfo_flags: OPTGROUP_OMP, /* optinfo_flags */ |
2466 | .tv_id: TV_NONE, /* tv_id */ |
2467 | PROP_cfg, /* properties_required */ |
2468 | .properties_provided: 0 /* Possibly PROP_gimple_eomp. */, /* properties_provided */ |
2469 | .properties_destroyed: 0, /* properties_destroyed */ |
2470 | .todo_flags_start: 0, /* todo_flags_start */ |
2471 | TODO_update_ssa | TODO_cleanup_cfg, /* todo_flags_finish */ |
2472 | }; |
2473 | |
2474 | class pass_oacc_loop_designation : public gimple_opt_pass |
2475 | { |
2476 | public: |
2477 | pass_oacc_loop_designation (gcc::context *ctxt) |
2478 | : gimple_opt_pass (pass_data_oacc_loop_designation, ctxt) |
2479 | {} |
2480 | |
2481 | /* opt_pass methods: */ |
2482 | bool gate (function *) final override { return flag_openacc; }; |
2483 | |
2484 | unsigned int execute (function *) final override |
2485 | { |
2486 | return execute_oacc_loop_designation (); |
2487 | } |
2488 | |
2489 | }; // class pass_oacc_loop_designation |
2490 | |
2491 | const pass_data pass_data_oacc_device_lower = |
2492 | { |
2493 | .type: GIMPLE_PASS, /* type */ |
2494 | .name: "oaccdevlow" , /* name */ |
2495 | .optinfo_flags: OPTGROUP_OMP, /* optinfo_flags */ |
2496 | .tv_id: TV_NONE, /* tv_id */ |
2497 | PROP_cfg, /* properties_required */ |
2498 | .properties_provided: 0 /* Possibly PROP_gimple_eomp. */, /* properties_provided */ |
2499 | .properties_destroyed: 0, /* properties_destroyed */ |
2500 | .todo_flags_start: 0, /* todo_flags_start */ |
2501 | TODO_update_ssa | TODO_cleanup_cfg, /* todo_flags_finish */ |
2502 | }; |
2503 | |
2504 | class pass_oacc_device_lower : public gimple_opt_pass |
2505 | { |
2506 | public: |
2507 | pass_oacc_device_lower (gcc::context *ctxt) |
2508 | : gimple_opt_pass (pass_data_oacc_device_lower, ctxt) |
2509 | {} |
2510 | |
2511 | /* opt_pass methods: */ |
2512 | bool gate (function *) final override { return flag_openacc; }; |
2513 | |
2514 | unsigned int execute (function *) final override |
2515 | { |
2516 | return execute_oacc_device_lower (); |
2517 | } |
2518 | |
2519 | }; // class pass_oacc_device_lower |
2520 | |
2521 | } // anon namespace |
2522 | |
2523 | gimple_opt_pass * |
2524 | make_pass_oacc_loop_designation (gcc::context *ctxt) |
2525 | { |
2526 | return new pass_oacc_loop_designation (ctxt); |
2527 | } |
2528 | |
2529 | gimple_opt_pass * |
2530 | make_pass_oacc_device_lower (gcc::context *ctxt) |
2531 | { |
2532 | return new pass_oacc_device_lower (ctxt); |
2533 | } |
2534 | |
2535 | |
2536 | /* Rewrite GOMP_SIMT_ENTER_ALLOC call given by GSI and remove the preceding |
2537 | GOMP_SIMT_ENTER call identifying the privatized variables, which are |
2538 | turned to structure fields and receive a DECL_VALUE_EXPR accordingly. |
2539 | Set *REGIMPLIFY to true, except if no privatized variables were seen. */ |
2540 | |
2541 | static void |
2542 | ompdevlow_adjust_simt_enter (gimple_stmt_iterator *gsi, bool *regimplify) |
2543 | { |
2544 | gimple *alloc_stmt = gsi_stmt (i: *gsi); |
2545 | tree simtrec = gimple_call_lhs (gs: alloc_stmt); |
2546 | tree simduid = gimple_call_arg (gs: alloc_stmt, index: 0); |
2547 | gimple *enter_stmt = SSA_NAME_DEF_STMT (simduid); |
2548 | gcc_assert (gimple_call_internal_p (enter_stmt, IFN_GOMP_SIMT_ENTER)); |
2549 | tree rectype = lang_hooks.types.make_type (RECORD_TYPE); |
2550 | TYPE_ARTIFICIAL (rectype) = TYPE_NAMELESS (rectype) = 1; |
2551 | TREE_ADDRESSABLE (rectype) = 1; |
2552 | TREE_TYPE (simtrec) = build_pointer_type (rectype); |
2553 | for (unsigned i = 1; i < gimple_call_num_args (gs: enter_stmt); i++) |
2554 | { |
2555 | tree *argp = gimple_call_arg_ptr (gs: enter_stmt, index: i); |
2556 | if (*argp == null_pointer_node) |
2557 | continue; |
2558 | gcc_assert (TREE_CODE (*argp) == ADDR_EXPR |
2559 | && VAR_P (TREE_OPERAND (*argp, 0))); |
2560 | tree var = TREE_OPERAND (*argp, 0); |
2561 | |
2562 | tree field = build_decl (DECL_SOURCE_LOCATION (var), FIELD_DECL, |
2563 | DECL_NAME (var), TREE_TYPE (var)); |
2564 | SET_DECL_ALIGN (field, DECL_ALIGN (var)); |
2565 | DECL_USER_ALIGN (field) = DECL_USER_ALIGN (var); |
2566 | TREE_THIS_VOLATILE (field) = TREE_THIS_VOLATILE (var); |
2567 | |
2568 | insert_field_into_struct (rectype, field); |
2569 | |
2570 | tree t = build_simple_mem_ref (simtrec); |
2571 | t = build3 (COMPONENT_REF, TREE_TYPE (var), t, field, NULL); |
2572 | TREE_THIS_VOLATILE (t) = TREE_THIS_VOLATILE (var); |
2573 | SET_DECL_VALUE_EXPR (var, t); |
2574 | DECL_HAS_VALUE_EXPR_P (var) = 1; |
2575 | *regimplify = true; |
2576 | } |
2577 | layout_type (rectype); |
2578 | tree size = TYPE_SIZE_UNIT (rectype); |
2579 | tree align = build_int_cst (TREE_TYPE (size), TYPE_ALIGN_UNIT (rectype)); |
2580 | |
2581 | alloc_stmt |
2582 | = gimple_build_call_internal (IFN_GOMP_SIMT_ENTER_ALLOC, 2, size, align); |
2583 | gimple_call_set_lhs (gs: alloc_stmt, lhs: simtrec); |
2584 | gsi_replace (gsi, alloc_stmt, false); |
2585 | gimple_stmt_iterator enter_gsi = gsi_for_stmt (enter_stmt); |
2586 | enter_stmt = gimple_build_assign (simduid, gimple_call_arg (gs: enter_stmt, index: 0)); |
2587 | gsi_replace (&enter_gsi, enter_stmt, false); |
2588 | |
2589 | use_operand_p use; |
2590 | gimple *exit_stmt; |
2591 | if (single_imm_use (var: simtrec, use_p: &use, stmt: &exit_stmt)) |
2592 | { |
2593 | gcc_assert (gimple_call_internal_p (exit_stmt, IFN_GOMP_SIMT_EXIT)); |
2594 | gimple_stmt_iterator exit_gsi = gsi_for_stmt (exit_stmt); |
2595 | tree clobber = build_clobber (rectype); |
2596 | exit_stmt = gimple_build_assign (build_simple_mem_ref (simtrec), clobber); |
2597 | gsi_insert_before (&exit_gsi, exit_stmt, GSI_SAME_STMT); |
2598 | } |
2599 | else |
2600 | gcc_checking_assert (has_zero_uses (simtrec)); |
2601 | } |
2602 | |
2603 | /* Callback for walk_gimple_stmt used to scan for SIMT-privatized variables. */ |
2604 | |
2605 | static tree |
2606 | find_simtpriv_var_op (tree *tp, int *walk_subtrees, void *) |
2607 | { |
2608 | tree t = *tp; |
2609 | |
2610 | if (VAR_P (t) |
2611 | && DECL_HAS_VALUE_EXPR_P (t) |
2612 | && lookup_attribute (attr_name: "omp simt private" , DECL_ATTRIBUTES (t))) |
2613 | { |
2614 | *walk_subtrees = 0; |
2615 | return t; |
2616 | } |
2617 | return NULL_TREE; |
2618 | } |
2619 | |
2620 | /* Cleanup uses of SIMT placeholder internal functions: on non-SIMT targets, |
2621 | VF is 1 and LANE is 0; on SIMT targets, VF is folded to a constant, and |
2622 | LANE is kept to be expanded to RTL later on. Also cleanup all other SIMT |
2623 | internal functions on non-SIMT targets, and likewise some SIMD internal |
2624 | functions on SIMT targets. */ |
2625 | |
2626 | static unsigned int |
2627 | execute_omp_device_lower () |
2628 | { |
2629 | int vf = targetm.simt.vf ? targetm.simt.vf () : 1; |
2630 | bool regimplify = false; |
2631 | basic_block bb; |
2632 | gimple_stmt_iterator gsi; |
2633 | bool calls_declare_variant_alt |
2634 | = cgraph_node::get (cfun->decl)->calls_declare_variant_alt; |
2635 | #ifdef ACCEL_COMPILER |
2636 | bool omp_redirect_indirect_calls = vec_safe_length (offload_ind_funcs) > 0; |
2637 | tree map_ptr_fn |
2638 | = builtin_decl_explicit (BUILT_IN_GOMP_TARGET_MAP_INDIRECT_PTR); |
2639 | #endif |
2640 | FOR_EACH_BB_FN (bb, cfun) |
2641 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
2642 | { |
2643 | gimple *stmt = gsi_stmt (i: gsi); |
2644 | if (!is_gimple_call (gs: stmt)) |
2645 | continue; |
2646 | if (!gimple_call_internal_p (gs: stmt)) |
2647 | { |
2648 | if (calls_declare_variant_alt) |
2649 | if (tree fndecl = gimple_call_fndecl (gs: stmt)) |
2650 | { |
2651 | tree new_fndecl = omp_resolve_declare_variant (fndecl); |
2652 | if (new_fndecl != fndecl) |
2653 | { |
2654 | gimple_call_set_fndecl (gs: stmt, decl: new_fndecl); |
2655 | update_stmt (s: stmt); |
2656 | } |
2657 | } |
2658 | #ifdef ACCEL_COMPILER |
2659 | if (omp_redirect_indirect_calls |
2660 | && gimple_call_fndecl (stmt) == NULL_TREE) |
2661 | { |
2662 | gcall *orig_call = dyn_cast <gcall *> (stmt); |
2663 | tree call_fn = gimple_call_fn (stmt); |
2664 | tree fn_ty = TREE_TYPE (call_fn); |
2665 | |
2666 | if (TREE_CODE (call_fn) == OBJ_TYPE_REF) |
2667 | { |
2668 | tree obj_ref = create_tmp_reg (TREE_TYPE (call_fn), |
2669 | ".ind_fn_objref" ); |
2670 | gimple *gassign = gimple_build_assign (obj_ref, call_fn); |
2671 | gsi_insert_before (&gsi, gassign, GSI_SAME_STMT); |
2672 | call_fn = obj_ref; |
2673 | } |
2674 | tree mapped_fn = create_tmp_reg (fn_ty, ".ind_fn" ); |
2675 | gimple *gcall = |
2676 | gimple_build_call (map_ptr_fn, 1, call_fn); |
2677 | gimple_set_location (gcall, gimple_location (stmt)); |
2678 | gimple_call_set_lhs (gcall, mapped_fn); |
2679 | gsi_insert_before (&gsi, gcall, GSI_SAME_STMT); |
2680 | |
2681 | gimple_call_set_fn (orig_call, mapped_fn); |
2682 | update_stmt (orig_call); |
2683 | } |
2684 | #endif |
2685 | continue; |
2686 | } |
2687 | tree lhs = gimple_call_lhs (gs: stmt), rhs = NULL_TREE; |
2688 | tree type = lhs ? TREE_TYPE (lhs) : integer_type_node; |
2689 | switch (gimple_call_internal_fn (gs: stmt)) |
2690 | { |
2691 | case IFN_GOMP_TARGET_REV: |
2692 | { |
2693 | #ifndef ACCEL_COMPILER |
2694 | gimple_stmt_iterator gsi2 = gsi; |
2695 | gsi_next (i: &gsi2); |
2696 | gcc_assert (!gsi_end_p (gsi2)); |
2697 | gcc_assert (gimple_call_builtin_p (gsi_stmt (gsi2), |
2698 | BUILT_IN_GOMP_TARGET)); |
2699 | tree old_decl |
2700 | = TREE_OPERAND (gimple_call_arg (gsi_stmt (gsi2), 1), 0); |
2701 | tree new_decl = gimple_call_arg (gs: gsi_stmt (i: gsi), index: 0); |
2702 | gimple_call_set_arg (gs: gsi_stmt (i: gsi2), index: 1, arg: new_decl); |
2703 | update_stmt (s: gsi_stmt (i: gsi2)); |
2704 | new_decl = TREE_OPERAND (new_decl, 0); |
2705 | unsigned i; |
2706 | unsigned num_funcs = vec_safe_length (v: offload_funcs); |
2707 | for (i = 0; i < num_funcs; i++) |
2708 | { |
2709 | if ((*offload_funcs)[i] == old_decl) |
2710 | { |
2711 | (*offload_funcs)[i] = new_decl; |
2712 | break; |
2713 | } |
2714 | else if ((*offload_funcs)[i] == new_decl) |
2715 | break; /* This can happen due to inlining. */ |
2716 | } |
2717 | gcc_assert (i < num_funcs); |
2718 | #else |
2719 | tree old_decl = TREE_OPERAND (gimple_call_arg (gsi_stmt (gsi), 0), |
2720 | 0); |
2721 | #endif |
2722 | /* FIXME: Find a way to actually prevent outputting the empty-body |
2723 | old_decl as debug symbol + function in the assembly file. */ |
2724 | cgraph_node *node = cgraph_node::get (decl: old_decl); |
2725 | node->address_taken = false; |
2726 | node->need_lto_streaming = false; |
2727 | node->offloadable = false; |
2728 | |
2729 | unlink_stmt_vdef (stmt); |
2730 | } |
2731 | break; |
2732 | case IFN_GOMP_USE_SIMT: |
2733 | rhs = vf == 1 ? integer_zero_node : integer_one_node; |
2734 | break; |
2735 | case IFN_GOMP_SIMT_ENTER: |
2736 | rhs = vf == 1 ? gimple_call_arg (gs: stmt, index: 0) : NULL_TREE; |
2737 | goto simtreg_enter_exit; |
2738 | case IFN_GOMP_SIMT_ENTER_ALLOC: |
2739 | if (vf != 1) |
2740 | ompdevlow_adjust_simt_enter (gsi: &gsi, regimplify: ®implify); |
2741 | rhs = vf == 1 ? null_pointer_node : NULL_TREE; |
2742 | goto simtreg_enter_exit; |
2743 | case IFN_GOMP_SIMT_EXIT: |
2744 | simtreg_enter_exit: |
2745 | if (vf != 1) |
2746 | continue; |
2747 | unlink_stmt_vdef (stmt); |
2748 | break; |
2749 | case IFN_GOMP_SIMT_LANE: |
2750 | case IFN_GOMP_SIMT_LAST_LANE: |
2751 | rhs = vf == 1 ? build_zero_cst (type) : NULL_TREE; |
2752 | break; |
2753 | case IFN_GOMP_SIMT_VF: |
2754 | rhs = build_int_cst (type, vf); |
2755 | break; |
2756 | case IFN_GOMP_SIMT_ORDERED_PRED: |
2757 | rhs = vf == 1 ? integer_zero_node : NULL_TREE; |
2758 | if (rhs || !lhs) |
2759 | unlink_stmt_vdef (stmt); |
2760 | break; |
2761 | case IFN_GOMP_SIMT_VOTE_ANY: |
2762 | case IFN_GOMP_SIMT_XCHG_BFLY: |
2763 | case IFN_GOMP_SIMT_XCHG_IDX: |
2764 | rhs = vf == 1 ? gimple_call_arg (gs: stmt, index: 0) : NULL_TREE; |
2765 | break; |
2766 | case IFN_GOMP_SIMD_LANE: |
2767 | case IFN_GOMP_SIMD_LAST_LANE: |
2768 | rhs = vf != 1 ? build_zero_cst (type) : NULL_TREE; |
2769 | break; |
2770 | case IFN_GOMP_SIMD_VF: |
2771 | rhs = vf != 1 ? build_one_cst (type) : NULL_TREE; |
2772 | break; |
2773 | default: |
2774 | continue; |
2775 | } |
2776 | if (lhs && !rhs) |
2777 | continue; |
2778 | stmt = lhs ? gimple_build_assign (lhs, rhs) : gimple_build_nop (); |
2779 | gsi_replace (&gsi, stmt, false); |
2780 | } |
2781 | if (regimplify) |
2782 | FOR_EACH_BB_REVERSE_FN (bb, cfun) |
2783 | for (gsi = gsi_last_bb (bb); !gsi_end_p (i: gsi); gsi_prev (i: &gsi)) |
2784 | if (walk_gimple_stmt (&gsi, NULL, find_simtpriv_var_op, NULL)) |
2785 | { |
2786 | if (gimple_clobber_p (s: gsi_stmt (i: gsi))) |
2787 | gsi_remove (&gsi, true); |
2788 | else |
2789 | gimple_regimplify_operands (gsi_stmt (i: gsi), &gsi); |
2790 | } |
2791 | if (vf != 1) |
2792 | cfun->has_force_vectorize_loops = false; |
2793 | return 0; |
2794 | } |
2795 | |
2796 | namespace { |
2797 | |
2798 | const pass_data pass_data_omp_device_lower = |
2799 | { |
2800 | .type: GIMPLE_PASS, /* type */ |
2801 | .name: "ompdevlow" , /* name */ |
2802 | .optinfo_flags: OPTGROUP_OMP, /* optinfo_flags */ |
2803 | .tv_id: TV_NONE, /* tv_id */ |
2804 | PROP_cfg, /* properties_required */ |
2805 | PROP_gimple_lomp_dev, /* properties_provided */ |
2806 | .properties_destroyed: 0, /* properties_destroyed */ |
2807 | .todo_flags_start: 0, /* todo_flags_start */ |
2808 | TODO_update_ssa, /* todo_flags_finish */ |
2809 | }; |
2810 | |
2811 | class pass_omp_device_lower : public gimple_opt_pass |
2812 | { |
2813 | public: |
2814 | pass_omp_device_lower (gcc::context *ctxt) |
2815 | : gimple_opt_pass (pass_data_omp_device_lower, ctxt) |
2816 | {} |
2817 | |
2818 | /* opt_pass methods: */ |
2819 | bool gate (function *fun) final override |
2820 | { |
2821 | #ifdef ACCEL_COMPILER |
2822 | bool offload_ind_funcs_p = vec_safe_length (offload_ind_funcs) > 0; |
2823 | #else |
2824 | bool offload_ind_funcs_p = false; |
2825 | #endif |
2826 | return (!(fun->curr_properties & PROP_gimple_lomp_dev) |
2827 | || (flag_openmp |
2828 | && (cgraph_node::get (decl: fun->decl)->calls_declare_variant_alt |
2829 | || offload_ind_funcs_p))); |
2830 | } |
2831 | unsigned int execute (function *) final override |
2832 | { |
2833 | return execute_omp_device_lower (); |
2834 | } |
2835 | |
2836 | }; // class pass_expand_omp_ssa |
2837 | |
2838 | } // anon namespace |
2839 | |
2840 | gimple_opt_pass * |
2841 | make_pass_omp_device_lower (gcc::context *ctxt) |
2842 | { |
2843 | return new pass_omp_device_lower (ctxt); |
2844 | } |
2845 | |
2846 | /* "omp declare target link" handling pass. */ |
2847 | |
2848 | namespace { |
2849 | |
2850 | const pass_data pass_data_omp_target_link = |
2851 | { |
2852 | .type: GIMPLE_PASS, /* type */ |
2853 | .name: "omptargetlink" , /* name */ |
2854 | .optinfo_flags: OPTGROUP_OMP, /* optinfo_flags */ |
2855 | .tv_id: TV_NONE, /* tv_id */ |
2856 | PROP_ssa, /* properties_required */ |
2857 | .properties_provided: 0, /* properties_provided */ |
2858 | .properties_destroyed: 0, /* properties_destroyed */ |
2859 | .todo_flags_start: 0, /* todo_flags_start */ |
2860 | TODO_update_ssa, /* todo_flags_finish */ |
2861 | }; |
2862 | |
2863 | class pass_omp_target_link : public gimple_opt_pass |
2864 | { |
2865 | public: |
2866 | pass_omp_target_link (gcc::context *ctxt) |
2867 | : gimple_opt_pass (pass_data_omp_target_link, ctxt) |
2868 | {} |
2869 | |
2870 | /* opt_pass methods: */ |
2871 | bool gate (function *fun) final override |
2872 | { |
2873 | #ifdef ACCEL_COMPILER |
2874 | return offloading_function_p (fun->decl); |
2875 | #else |
2876 | (void) fun; |
2877 | return false; |
2878 | #endif |
2879 | } |
2880 | |
2881 | unsigned execute (function *) final override; |
2882 | }; |
2883 | |
2884 | /* Callback for walk_gimple_stmt used to scan for link var operands. */ |
2885 | |
2886 | static tree |
2887 | find_link_var_op (tree *tp, int *walk_subtrees, void *) |
2888 | { |
2889 | tree t = *tp; |
2890 | |
2891 | if (VAR_P (t) |
2892 | && DECL_HAS_VALUE_EXPR_P (t) |
2893 | && is_global_var (t) |
2894 | && lookup_attribute (attr_name: "omp declare target link" , DECL_ATTRIBUTES (t))) |
2895 | { |
2896 | *walk_subtrees = 0; |
2897 | return t; |
2898 | } |
2899 | |
2900 | return NULL_TREE; |
2901 | } |
2902 | |
2903 | unsigned |
2904 | pass_omp_target_link::execute (function *fun) |
2905 | { |
2906 | basic_block bb; |
2907 | FOR_EACH_BB_FN (bb, fun) |
2908 | { |
2909 | gimple_stmt_iterator gsi; |
2910 | for (gsi = gsi_start_bb (bb); !gsi_end_p (i: gsi); gsi_next (i: &gsi)) |
2911 | { |
2912 | if (gimple_call_builtin_p (gsi_stmt (i: gsi), BUILT_IN_GOMP_TARGET)) |
2913 | { |
2914 | tree dev = gimple_call_arg (gs: gsi_stmt (i: gsi), index: 0); |
2915 | tree fn = gimple_call_arg (gs: gsi_stmt (i: gsi), index: 1); |
2916 | if (POINTER_TYPE_P (TREE_TYPE (fn))) |
2917 | fn = TREE_OPERAND (fn, 0); |
2918 | if (TREE_CODE (dev) == INTEGER_CST |
2919 | && wi::to_wide (t: dev) == GOMP_DEVICE_HOST_FALLBACK |
2920 | && lookup_attribute (attr_name: "omp target device_ancestor_nohost" , |
2921 | DECL_ATTRIBUTES (fn)) != NULL_TREE) |
2922 | continue; /* ancestor:1 */ |
2923 | /* Nullify the second argument of __builtin_GOMP_target_ext. */ |
2924 | gimple_call_set_arg (gs: gsi_stmt (i: gsi), index: 1, null_pointer_node); |
2925 | update_stmt (s: gsi_stmt (i: gsi)); |
2926 | } |
2927 | if (walk_gimple_stmt (&gsi, NULL, find_link_var_op, NULL)) |
2928 | gimple_regimplify_operands (gsi_stmt (i: gsi), &gsi); |
2929 | } |
2930 | } |
2931 | |
2932 | return 0; |
2933 | } |
2934 | |
2935 | } // anon namespace |
2936 | |
2937 | gimple_opt_pass * |
2938 | make_pass_omp_target_link (gcc::context *ctxt) |
2939 | { |
2940 | return new pass_omp_target_link (ctxt); |
2941 | } |
2942 | |