]>
Commit | Line | Data |
---|---|---|
629b3d75 MJ |
1 | /* Bits of OpenMP and OpenACC handling that is specific to device offloading |
2 | and a lowering pass for OpenACC device directives. | |
3 | ||
8d9254fc | 4 | Copyright (C) 2005-2020 Free Software Foundation, Inc. |
629b3d75 MJ |
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" | |
0c6b03b5 | 36 | #include "langhooks.h" |
629b3d75 MJ |
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" | |
0c6b03b5 AM |
43 | #include "tree-nested.h" |
44 | #include "stor-layout.h" | |
629b3d75 MJ |
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" | |
324ff1a0 | 51 | #include "intl.h" |
314e6352 ML |
52 | #include "stringpool.h" |
53 | #include "attribs.h" | |
f64b12bd | 54 | #include "cfgloop.h" |
dc703151 | 55 | #include "context.h" |
629b3d75 MJ |
56 | |
57 | /* Describe the OpenACC looping structure of a function. The entire | |
58 | function is held in a 'NULL' loop. */ | |
59 | ||
60 | struct oacc_loop | |
61 | { | |
62 | oacc_loop *parent; /* Containing loop. */ | |
63 | ||
64 | oacc_loop *child; /* First inner loop. */ | |
65 | ||
66 | oacc_loop *sibling; /* Next loop within same parent. */ | |
67 | ||
68 | location_t loc; /* Location of the loop start. */ | |
69 | ||
70 | gcall *marker; /* Initial head marker. */ | |
71 | ||
01914336 MJ |
72 | gcall *heads[GOMP_DIM_MAX]; /* Head marker functions. */ |
73 | gcall *tails[GOMP_DIM_MAX]; /* Tail marker functions. */ | |
629b3d75 MJ |
74 | |
75 | tree routine; /* Pseudo-loop enclosing a routine. */ | |
76 | ||
77 | unsigned mask; /* Partitioning mask. */ | |
02889d23 | 78 | unsigned e_mask; /* Partitioning of element loops (when tiling). */ |
629b3d75 MJ |
79 | unsigned inner; /* Partitioning of inner loops. */ |
80 | unsigned flags; /* Partitioning flags. */ | |
02889d23 | 81 | vec<gcall *> ifns; /* Contained loop abstraction functions. */ |
629b3d75 MJ |
82 | tree chunk_size; /* Chunk size. */ |
83 | gcall *head_end; /* Final marker of head sequence. */ | |
84 | }; | |
85 | ||
86 | /* Holds offload tables with decls. */ | |
87 | vec<tree, va_gc> *offload_funcs, *offload_vars; | |
88 | ||
89 | /* Return level at which oacc routine may spawn a partitioned loop, or | |
90 | -1 if it is not a routine (i.e. is an offload fn). */ | |
91 | ||
4c187162 | 92 | int |
629b3d75 MJ |
93 | oacc_fn_attrib_level (tree attr) |
94 | { | |
95 | tree pos = TREE_VALUE (attr); | |
96 | ||
97 | if (!TREE_PURPOSE (pos)) | |
98 | return -1; | |
99 | ||
100 | int ix = 0; | |
101 | for (ix = 0; ix != GOMP_DIM_MAX; | |
102 | ix++, pos = TREE_CHAIN (pos)) | |
103 | if (!integer_zerop (TREE_PURPOSE (pos))) | |
104 | break; | |
105 | ||
106 | return ix; | |
107 | } | |
108 | ||
109 | /* Helper function for omp_finish_file routine. Takes decls from V_DECLS and | |
110 | adds their addresses and sizes to constructor-vector V_CTOR. */ | |
111 | ||
112 | static void | |
113 | add_decls_addresses_to_decl_constructor (vec<tree, va_gc> *v_decls, | |
114 | vec<constructor_elt, va_gc> *v_ctor) | |
115 | { | |
116 | unsigned len = vec_safe_length (v_decls); | |
117 | for (unsigned i = 0; i < len; i++) | |
118 | { | |
119 | tree it = (*v_decls)[i]; | |
120 | bool is_var = VAR_P (it); | |
121 | bool is_link_var | |
122 | = is_var | |
123 | #ifdef ACCEL_COMPILER | |
124 | && DECL_HAS_VALUE_EXPR_P (it) | |
125 | #endif | |
126 | && lookup_attribute ("omp declare target link", DECL_ATTRIBUTES (it)); | |
127 | ||
128 | tree size = NULL_TREE; | |
129 | if (is_var) | |
130 | size = fold_convert (const_ptr_type_node, DECL_SIZE_UNIT (it)); | |
131 | ||
132 | tree addr; | |
133 | if (!is_link_var) | |
134 | addr = build_fold_addr_expr (it); | |
135 | else | |
136 | { | |
137 | #ifdef ACCEL_COMPILER | |
138 | /* For "omp declare target link" vars add address of the pointer to | |
139 | the target table, instead of address of the var. */ | |
140 | tree value_expr = DECL_VALUE_EXPR (it); | |
141 | tree link_ptr_decl = TREE_OPERAND (value_expr, 0); | |
142 | varpool_node::finalize_decl (link_ptr_decl); | |
143 | addr = build_fold_addr_expr (link_ptr_decl); | |
144 | #else | |
145 | addr = build_fold_addr_expr (it); | |
146 | #endif | |
147 | ||
148 | /* Most significant bit of the size marks "omp declare target link" | |
149 | vars in host and target tables. */ | |
150 | unsigned HOST_WIDE_INT isize = tree_to_uhwi (size); | |
151 | isize |= 1ULL << (int_size_in_bytes (const_ptr_type_node) | |
152 | * BITS_PER_UNIT - 1); | |
153 | size = wide_int_to_tree (const_ptr_type_node, isize); | |
154 | } | |
155 | ||
156 | CONSTRUCTOR_APPEND_ELT (v_ctor, NULL_TREE, addr); | |
157 | if (is_var) | |
158 | CONSTRUCTOR_APPEND_ELT (v_ctor, NULL_TREE, size); | |
159 | } | |
160 | } | |
161 | ||
dc703151 JJ |
162 | /* Return true if DECL is a function for which its references should be |
163 | analyzed. */ | |
164 | ||
165 | static bool | |
166 | omp_declare_target_fn_p (tree decl) | |
167 | { | |
168 | return (TREE_CODE (decl) == FUNCTION_DECL | |
169 | && lookup_attribute ("omp declare target", DECL_ATTRIBUTES (decl)) | |
170 | && !lookup_attribute ("omp declare target host", | |
171 | DECL_ATTRIBUTES (decl)) | |
172 | && (!flag_openacc | |
173 | || oacc_get_fn_attrib (decl) == NULL_TREE)); | |
174 | } | |
175 | ||
176 | /* Return true if DECL Is a variable for which its initializer references | |
177 | should be analyzed. */ | |
178 | ||
179 | static bool | |
180 | omp_declare_target_var_p (tree decl) | |
181 | { | |
182 | return (VAR_P (decl) | |
183 | && lookup_attribute ("omp declare target", DECL_ATTRIBUTES (decl)) | |
184 | && !lookup_attribute ("omp declare target link", | |
185 | DECL_ATTRIBUTES (decl))); | |
186 | } | |
187 | ||
188 | /* Helper function for omp_discover_implicit_declare_target, called through | |
189 | walk_tree. Mark referenced FUNCTION_DECLs implicitly as | |
190 | declare target to. */ | |
191 | ||
192 | static tree | |
49ddde69 | 193 | omp_discover_declare_target_tgt_fn_r (tree *tp, int *walk_subtrees, void *data) |
dc703151 JJ |
194 | { |
195 | if (TREE_CODE (*tp) == FUNCTION_DECL | |
196 | && !omp_declare_target_fn_p (*tp) | |
197 | && !lookup_attribute ("omp declare target host", DECL_ATTRIBUTES (*tp))) | |
198 | { | |
199 | tree id = get_identifier ("omp declare target"); | |
200 | if (!DECL_EXTERNAL (*tp) && DECL_SAVED_TREE (*tp)) | |
201 | ((vec<tree> *) data)->safe_push (*tp); | |
202 | DECL_ATTRIBUTES (*tp) = tree_cons (id, NULL_TREE, DECL_ATTRIBUTES (*tp)); | |
203 | symtab_node *node = symtab_node::get (*tp); | |
204 | if (node != NULL) | |
205 | { | |
206 | node->offloadable = 1; | |
207 | if (ENABLE_OFFLOADING) | |
208 | g->have_offload = true; | |
209 | } | |
210 | } | |
211 | else if (TYPE_P (*tp)) | |
212 | *walk_subtrees = 0; | |
213 | /* else if (TREE_CODE (*tp) == OMP_TARGET) | |
214 | { | |
215 | if (tree dev = omp_find_clause (OMP_TARGET_CLAUSES (*tp))) | |
216 | if (OMP_DEVICE_ANCESTOR (dev)) | |
217 | *walk_subtrees = 0; | |
218 | } */ | |
219 | return NULL_TREE; | |
220 | } | |
221 | ||
49ddde69 JJ |
222 | /* Similarly, but ignore references outside of OMP_TARGET regions. */ |
223 | ||
224 | static tree | |
225 | omp_discover_declare_target_fn_r (tree *tp, int *walk_subtrees, void *data) | |
226 | { | |
227 | if (TREE_CODE (*tp) == OMP_TARGET) | |
228 | { | |
229 | /* And not OMP_DEVICE_ANCESTOR. */ | |
230 | walk_tree_without_duplicates (&OMP_TARGET_BODY (*tp), | |
231 | omp_discover_declare_target_tgt_fn_r, | |
232 | data); | |
233 | *walk_subtrees = 0; | |
234 | } | |
235 | else if (TYPE_P (*tp)) | |
236 | *walk_subtrees = 0; | |
237 | return NULL_TREE; | |
238 | } | |
239 | ||
dc703151 JJ |
240 | /* Helper function for omp_discover_implicit_declare_target, called through |
241 | walk_tree. Mark referenced FUNCTION_DECLs implicitly as | |
242 | declare target to. */ | |
243 | ||
244 | static tree | |
245 | omp_discover_declare_target_var_r (tree *tp, int *walk_subtrees, void *data) | |
246 | { | |
247 | if (TREE_CODE (*tp) == FUNCTION_DECL) | |
49ddde69 | 248 | return omp_discover_declare_target_tgt_fn_r (tp, walk_subtrees, data); |
dc703151 JJ |
249 | else if (VAR_P (*tp) |
250 | && is_global_var (*tp) | |
251 | && !omp_declare_target_var_p (*tp)) | |
252 | { | |
253 | tree id = get_identifier ("omp declare target"); | |
254 | if (lookup_attribute ("omp declare target link", DECL_ATTRIBUTES (*tp))) | |
255 | { | |
256 | error_at (DECL_SOURCE_LOCATION (*tp), | |
257 | "%qD specified both in declare target %<link%> and " | |
258 | "implicitly in %<to%> clauses", *tp); | |
259 | DECL_ATTRIBUTES (*tp) | |
260 | = remove_attribute ("omp declare target link", DECL_ATTRIBUTES (*tp)); | |
261 | } | |
262 | if (TREE_STATIC (*tp) && DECL_INITIAL (*tp)) | |
263 | ((vec<tree> *) data)->safe_push (*tp); | |
264 | DECL_ATTRIBUTES (*tp) = tree_cons (id, NULL_TREE, DECL_ATTRIBUTES (*tp)); | |
265 | symtab_node *node = symtab_node::get (*tp); | |
266 | if (node != NULL && !node->offloadable) | |
267 | { | |
268 | node->offloadable = 1; | |
269 | if (ENABLE_OFFLOADING) | |
270 | { | |
271 | g->have_offload = true; | |
272 | if (is_a <varpool_node *> (node)) | |
273 | vec_safe_push (offload_vars, node->decl); | |
274 | } | |
275 | } | |
276 | } | |
277 | else if (TYPE_P (*tp)) | |
278 | *walk_subtrees = 0; | |
279 | return NULL_TREE; | |
280 | } | |
281 | ||
282 | /* Perform the OpenMP implicit declare target to discovery. */ | |
283 | ||
284 | void | |
285 | omp_discover_implicit_declare_target (void) | |
286 | { | |
287 | cgraph_node *node; | |
288 | varpool_node *vnode; | |
289 | auto_vec<tree> worklist; | |
290 | ||
291 | FOR_EACH_DEFINED_FUNCTION (node) | |
49ddde69 JJ |
292 | if (DECL_SAVED_TREE (node->decl)) |
293 | { | |
294 | if (omp_declare_target_fn_p (node->decl)) | |
295 | worklist.safe_push (node->decl); | |
296 | else if (DECL_STRUCT_FUNCTION (node->decl) | |
297 | && DECL_STRUCT_FUNCTION (node->decl)->has_omp_target) | |
298 | worklist.safe_push (node->decl); | |
299 | } | |
dc703151 JJ |
300 | FOR_EACH_STATIC_INITIALIZER (vnode) |
301 | if (omp_declare_target_var_p (vnode->decl)) | |
302 | worklist.safe_push (vnode->decl); | |
303 | while (!worklist.is_empty ()) | |
304 | { | |
305 | tree decl = worklist.pop (); | |
49ddde69 JJ |
306 | if (VAR_P (decl)) |
307 | walk_tree_without_duplicates (&DECL_INITIAL (decl), | |
308 | omp_discover_declare_target_var_r, | |
309 | &worklist); | |
310 | else if (omp_declare_target_fn_p (decl)) | |
dc703151 | 311 | walk_tree_without_duplicates (&DECL_SAVED_TREE (decl), |
49ddde69 | 312 | omp_discover_declare_target_tgt_fn_r, |
dc703151 JJ |
313 | &worklist); |
314 | else | |
49ddde69 JJ |
315 | walk_tree_without_duplicates (&DECL_SAVED_TREE (decl), |
316 | omp_discover_declare_target_fn_r, | |
dc703151 JJ |
317 | &worklist); |
318 | } | |
319 | } | |
320 | ||
321 | ||
629b3d75 MJ |
322 | /* Create new symbols containing (address, size) pairs for global variables, |
323 | marked with "omp declare target" attribute, as well as addresses for the | |
324 | functions, which are outlined offloading regions. */ | |
325 | void | |
326 | omp_finish_file (void) | |
327 | { | |
328 | unsigned num_funcs = vec_safe_length (offload_funcs); | |
329 | unsigned num_vars = vec_safe_length (offload_vars); | |
330 | ||
331 | if (num_funcs == 0 && num_vars == 0) | |
332 | return; | |
333 | ||
334 | if (targetm_common.have_named_sections) | |
335 | { | |
336 | vec<constructor_elt, va_gc> *v_f, *v_v; | |
337 | vec_alloc (v_f, num_funcs); | |
338 | vec_alloc (v_v, num_vars * 2); | |
339 | ||
340 | add_decls_addresses_to_decl_constructor (offload_funcs, v_f); | |
341 | add_decls_addresses_to_decl_constructor (offload_vars, v_v); | |
342 | ||
343 | tree vars_decl_type = build_array_type_nelts (pointer_sized_int_node, | |
344 | num_vars * 2); | |
345 | tree funcs_decl_type = build_array_type_nelts (pointer_sized_int_node, | |
346 | num_funcs); | |
347 | SET_TYPE_ALIGN (vars_decl_type, TYPE_ALIGN (pointer_sized_int_node)); | |
348 | SET_TYPE_ALIGN (funcs_decl_type, TYPE_ALIGN (pointer_sized_int_node)); | |
349 | tree ctor_v = build_constructor (vars_decl_type, v_v); | |
350 | tree ctor_f = build_constructor (funcs_decl_type, v_f); | |
351 | TREE_CONSTANT (ctor_v) = TREE_CONSTANT (ctor_f) = 1; | |
352 | TREE_STATIC (ctor_v) = TREE_STATIC (ctor_f) = 1; | |
353 | tree funcs_decl = build_decl (UNKNOWN_LOCATION, VAR_DECL, | |
354 | get_identifier (".offload_func_table"), | |
355 | funcs_decl_type); | |
356 | tree vars_decl = build_decl (UNKNOWN_LOCATION, VAR_DECL, | |
357 | get_identifier (".offload_var_table"), | |
358 | vars_decl_type); | |
359 | TREE_STATIC (funcs_decl) = TREE_STATIC (vars_decl) = 1; | |
360 | /* Do not align tables more than TYPE_ALIGN (pointer_sized_int_node), | |
361 | otherwise a joint table in a binary will contain padding between | |
362 | tables from multiple object files. */ | |
363 | DECL_USER_ALIGN (funcs_decl) = DECL_USER_ALIGN (vars_decl) = 1; | |
364 | SET_DECL_ALIGN (funcs_decl, TYPE_ALIGN (funcs_decl_type)); | |
365 | SET_DECL_ALIGN (vars_decl, TYPE_ALIGN (vars_decl_type)); | |
366 | DECL_INITIAL (funcs_decl) = ctor_f; | |
367 | DECL_INITIAL (vars_decl) = ctor_v; | |
368 | set_decl_section_name (funcs_decl, OFFLOAD_FUNC_TABLE_SECTION_NAME); | |
369 | set_decl_section_name (vars_decl, OFFLOAD_VAR_TABLE_SECTION_NAME); | |
370 | ||
371 | varpool_node::finalize_decl (vars_decl); | |
372 | varpool_node::finalize_decl (funcs_decl); | |
373 | } | |
374 | else | |
375 | { | |
376 | for (unsigned i = 0; i < num_funcs; i++) | |
377 | { | |
378 | tree it = (*offload_funcs)[i]; | |
379 | targetm.record_offload_symbol (it); | |
380 | } | |
381 | for (unsigned i = 0; i < num_vars; i++) | |
382 | { | |
383 | tree it = (*offload_vars)[i]; | |
c2211a60 TB |
384 | #ifdef ACCEL_COMPILER |
385 | if (DECL_HAS_VALUE_EXPR_P (it) | |
386 | && lookup_attribute ("omp declare target link", | |
387 | DECL_ATTRIBUTES (it))) | |
388 | { | |
389 | tree value_expr = DECL_VALUE_EXPR (it); | |
390 | tree link_ptr_decl = TREE_OPERAND (value_expr, 0); | |
391 | targetm.record_offload_symbol (link_ptr_decl); | |
392 | varpool_node::finalize_decl (link_ptr_decl); | |
393 | } | |
394 | else | |
395 | #endif | |
396 | targetm.record_offload_symbol (it); | |
629b3d75 MJ |
397 | } |
398 | } | |
399 | } | |
400 | ||
02889d23 CLT |
401 | /* Call dim_pos (POS == true) or dim_size (POS == false) builtins for |
402 | axis DIM. Return a tmp var holding the result. */ | |
403 | ||
404 | static tree | |
405 | oacc_dim_call (bool pos, int dim, gimple_seq *seq) | |
406 | { | |
407 | tree arg = build_int_cst (unsigned_type_node, dim); | |
408 | tree size = create_tmp_var (integer_type_node); | |
409 | enum internal_fn fn = pos ? IFN_GOACC_DIM_POS : IFN_GOACC_DIM_SIZE; | |
410 | gimple *call = gimple_build_call_internal (fn, 1, arg); | |
411 | ||
412 | gimple_call_set_lhs (call, size); | |
413 | gimple_seq_add_stmt (seq, call); | |
414 | ||
415 | return size; | |
416 | } | |
417 | ||
629b3d75 MJ |
418 | /* Find the number of threads (POS = false), or thread number (POS = |
419 | true) for an OpenACC region partitioned as MASK. Setup code | |
420 | required for the calculation is added to SEQ. */ | |
421 | ||
422 | static tree | |
423 | oacc_thread_numbers (bool pos, int mask, gimple_seq *seq) | |
424 | { | |
425 | tree res = pos ? NULL_TREE : build_int_cst (unsigned_type_node, 1); | |
426 | unsigned ix; | |
427 | ||
428 | /* Start at gang level, and examine relevant dimension indices. */ | |
429 | for (ix = GOMP_DIM_GANG; ix != GOMP_DIM_MAX; ix++) | |
430 | if (GOMP_DIM_MASK (ix) & mask) | |
431 | { | |
629b3d75 MJ |
432 | if (res) |
433 | { | |
434 | /* We had an outer index, so scale that by the size of | |
435 | this dimension. */ | |
02889d23 | 436 | tree n = oacc_dim_call (false, ix, seq); |
629b3d75 MJ |
437 | res = fold_build2 (MULT_EXPR, integer_type_node, res, n); |
438 | } | |
439 | if (pos) | |
440 | { | |
441 | /* Determine index in this dimension. */ | |
02889d23 | 442 | tree id = oacc_dim_call (true, ix, seq); |
629b3d75 MJ |
443 | if (res) |
444 | res = fold_build2 (PLUS_EXPR, integer_type_node, res, id); | |
445 | else | |
446 | res = id; | |
447 | } | |
448 | } | |
449 | ||
450 | if (res == NULL_TREE) | |
451 | res = integer_zero_node; | |
452 | ||
453 | return res; | |
454 | } | |
455 | ||
456 | /* Transform IFN_GOACC_LOOP calls to actual code. See | |
457 | expand_oacc_for for where these are generated. At the vector | |
458 | level, we stride loops, such that each member of a warp will | |
459 | operate on adjacent iterations. At the worker and gang level, | |
460 | each gang/warp executes a set of contiguous iterations. Chunking | |
461 | can override this such that each iteration engine executes a | |
01914336 | 462 | contiguous chunk, and then moves on to stride to the next chunk. */ |
629b3d75 MJ |
463 | |
464 | static void | |
465 | oacc_xform_loop (gcall *call) | |
466 | { | |
467 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
468 | enum ifn_goacc_loop_kind code | |
469 | = (enum ifn_goacc_loop_kind) TREE_INT_CST_LOW (gimple_call_arg (call, 0)); | |
470 | tree dir = gimple_call_arg (call, 1); | |
471 | tree range = gimple_call_arg (call, 2); | |
472 | tree step = gimple_call_arg (call, 3); | |
473 | tree chunk_size = NULL_TREE; | |
474 | unsigned mask = (unsigned) TREE_INT_CST_LOW (gimple_call_arg (call, 5)); | |
475 | tree lhs = gimple_call_lhs (call); | |
c29c92c7 | 476 | tree type = NULL_TREE; |
629b3d75 MJ |
477 | tree diff_type = TREE_TYPE (range); |
478 | tree r = NULL_TREE; | |
479 | gimple_seq seq = NULL; | |
480 | bool chunking = false, striding = true; | |
481 | unsigned outer_mask = mask & (~mask + 1); // Outermost partitioning | |
482 | unsigned inner_mask = mask & ~outer_mask; // Inner partitioning (if any) | |
483 | ||
c29c92c7 FX |
484 | /* Skip lowering if return value of IFN_GOACC_LOOP call is not used. */ |
485 | if (!lhs) | |
486 | { | |
487 | gsi_replace_with_seq (&gsi, seq, true); | |
488 | return; | |
489 | } | |
490 | ||
491 | type = TREE_TYPE (lhs); | |
492 | ||
629b3d75 MJ |
493 | #ifdef ACCEL_COMPILER |
494 | chunk_size = gimple_call_arg (call, 4); | |
495 | if (integer_minus_onep (chunk_size) /* Force static allocation. */ | |
496 | || integer_zerop (chunk_size)) /* Default (also static). */ | |
497 | { | |
498 | /* If we're at the gang level, we want each to execute a | |
499 | contiguous run of iterations. Otherwise we want each element | |
500 | to stride. */ | |
501 | striding = !(outer_mask & GOMP_DIM_MASK (GOMP_DIM_GANG)); | |
502 | chunking = false; | |
503 | } | |
504 | else | |
505 | { | |
506 | /* Chunk of size 1 is striding. */ | |
507 | striding = integer_onep (chunk_size); | |
508 | chunking = !striding; | |
509 | } | |
510 | #endif | |
511 | ||
512 | /* striding=true, chunking=true | |
513 | -> invalid. | |
514 | striding=true, chunking=false | |
515 | -> chunks=1 | |
516 | striding=false,chunking=true | |
517 | -> chunks=ceil (range/(chunksize*threads*step)) | |
518 | striding=false,chunking=false | |
519 | -> chunk_size=ceil(range/(threads*step)),chunks=1 */ | |
520 | push_gimplify_context (true); | |
521 | ||
522 | switch (code) | |
523 | { | |
524 | default: gcc_unreachable (); | |
525 | ||
526 | case IFN_GOACC_LOOP_CHUNKS: | |
527 | if (!chunking) | |
528 | r = build_int_cst (type, 1); | |
529 | else | |
530 | { | |
531 | /* chunk_max | |
532 | = (range - dir) / (chunks * step * num_threads) + dir */ | |
533 | tree per = oacc_thread_numbers (false, mask, &seq); | |
534 | per = fold_convert (type, per); | |
535 | chunk_size = fold_convert (type, chunk_size); | |
536 | per = fold_build2 (MULT_EXPR, type, per, chunk_size); | |
537 | per = fold_build2 (MULT_EXPR, type, per, step); | |
538 | r = build2 (MINUS_EXPR, type, range, dir); | |
539 | r = build2 (PLUS_EXPR, type, r, per); | |
540 | r = build2 (TRUNC_DIV_EXPR, type, r, per); | |
541 | } | |
542 | break; | |
543 | ||
544 | case IFN_GOACC_LOOP_STEP: | |
545 | { | |
546 | /* If striding, step by the entire compute volume, otherwise | |
547 | step by the inner volume. */ | |
548 | unsigned volume = striding ? mask : inner_mask; | |
549 | ||
550 | r = oacc_thread_numbers (false, volume, &seq); | |
551 | r = build2 (MULT_EXPR, type, fold_convert (type, r), step); | |
552 | } | |
553 | break; | |
554 | ||
555 | case IFN_GOACC_LOOP_OFFSET: | |
f64b12bd CP |
556 | /* Enable vectorization on non-SIMT targets. */ |
557 | if (!targetm.simt.vf | |
558 | && outer_mask == GOMP_DIM_MASK (GOMP_DIM_VECTOR) | |
559 | /* If not -fno-tree-loop-vectorize, hint that we want to vectorize | |
560 | the loop. */ | |
561 | && (flag_tree_loop_vectorize | |
562 | || !global_options_set.x_flag_tree_loop_vectorize)) | |
563 | { | |
564 | basic_block bb = gsi_bb (gsi); | |
99b1c316 MS |
565 | class loop *parent = bb->loop_father; |
566 | class loop *body = parent->inner; | |
f64b12bd CP |
567 | |
568 | parent->force_vectorize = true; | |
569 | parent->safelen = INT_MAX; | |
570 | ||
571 | /* "Chunking loops" may have inner loops. */ | |
572 | if (parent->inner) | |
573 | { | |
574 | body->force_vectorize = true; | |
575 | body->safelen = INT_MAX; | |
576 | } | |
577 | ||
578 | cfun->has_force_vectorize_loops = true; | |
579 | } | |
629b3d75 MJ |
580 | if (striding) |
581 | { | |
582 | r = oacc_thread_numbers (true, mask, &seq); | |
583 | r = fold_convert (diff_type, r); | |
584 | } | |
585 | else | |
586 | { | |
587 | tree inner_size = oacc_thread_numbers (false, inner_mask, &seq); | |
588 | tree outer_size = oacc_thread_numbers (false, outer_mask, &seq); | |
589 | tree volume = fold_build2 (MULT_EXPR, TREE_TYPE (inner_size), | |
590 | inner_size, outer_size); | |
591 | ||
592 | volume = fold_convert (diff_type, volume); | |
593 | if (chunking) | |
594 | chunk_size = fold_convert (diff_type, chunk_size); | |
595 | else | |
596 | { | |
597 | tree per = fold_build2 (MULT_EXPR, diff_type, volume, step); | |
598 | ||
599 | chunk_size = build2 (MINUS_EXPR, diff_type, range, dir); | |
600 | chunk_size = build2 (PLUS_EXPR, diff_type, chunk_size, per); | |
601 | chunk_size = build2 (TRUNC_DIV_EXPR, diff_type, chunk_size, per); | |
602 | } | |
603 | ||
604 | tree span = build2 (MULT_EXPR, diff_type, chunk_size, | |
605 | fold_convert (diff_type, inner_size)); | |
606 | r = oacc_thread_numbers (true, outer_mask, &seq); | |
607 | r = fold_convert (diff_type, r); | |
608 | r = build2 (MULT_EXPR, diff_type, r, span); | |
609 | ||
610 | tree inner = oacc_thread_numbers (true, inner_mask, &seq); | |
611 | inner = fold_convert (diff_type, inner); | |
612 | r = fold_build2 (PLUS_EXPR, diff_type, r, inner); | |
613 | ||
614 | if (chunking) | |
615 | { | |
616 | tree chunk = fold_convert (diff_type, gimple_call_arg (call, 6)); | |
617 | tree per | |
618 | = fold_build2 (MULT_EXPR, diff_type, volume, chunk_size); | |
619 | per = build2 (MULT_EXPR, diff_type, per, chunk); | |
620 | ||
621 | r = build2 (PLUS_EXPR, diff_type, r, per); | |
622 | } | |
623 | } | |
624 | r = fold_build2 (MULT_EXPR, diff_type, r, step); | |
625 | if (type != diff_type) | |
626 | r = fold_convert (type, r); | |
627 | break; | |
628 | ||
629 | case IFN_GOACC_LOOP_BOUND: | |
630 | if (striding) | |
631 | r = range; | |
632 | else | |
633 | { | |
634 | tree inner_size = oacc_thread_numbers (false, inner_mask, &seq); | |
635 | tree outer_size = oacc_thread_numbers (false, outer_mask, &seq); | |
636 | tree volume = fold_build2 (MULT_EXPR, TREE_TYPE (inner_size), | |
637 | inner_size, outer_size); | |
638 | ||
639 | volume = fold_convert (diff_type, volume); | |
640 | if (chunking) | |
641 | chunk_size = fold_convert (diff_type, chunk_size); | |
642 | else | |
643 | { | |
644 | tree per = fold_build2 (MULT_EXPR, diff_type, volume, step); | |
645 | ||
646 | chunk_size = build2 (MINUS_EXPR, diff_type, range, dir); | |
647 | chunk_size = build2 (PLUS_EXPR, diff_type, chunk_size, per); | |
648 | chunk_size = build2 (TRUNC_DIV_EXPR, diff_type, chunk_size, per); | |
649 | } | |
650 | ||
651 | tree span = build2 (MULT_EXPR, diff_type, chunk_size, | |
652 | fold_convert (diff_type, inner_size)); | |
653 | ||
654 | r = fold_build2 (MULT_EXPR, diff_type, span, step); | |
655 | ||
656 | tree offset = gimple_call_arg (call, 6); | |
657 | r = build2 (PLUS_EXPR, diff_type, r, | |
658 | fold_convert (diff_type, offset)); | |
659 | r = build2 (integer_onep (dir) ? MIN_EXPR : MAX_EXPR, | |
660 | diff_type, r, range); | |
661 | } | |
662 | if (diff_type != type) | |
663 | r = fold_convert (type, r); | |
664 | break; | |
665 | } | |
666 | ||
667 | gimplify_assign (lhs, r, &seq); | |
668 | ||
669 | pop_gimplify_context (NULL); | |
670 | ||
671 | gsi_replace_with_seq (&gsi, seq, true); | |
672 | } | |
673 | ||
02889d23 CLT |
674 | /* Transform a GOACC_TILE call. Determines the element loop span for |
675 | the specified loop of the nest. This is 1 if we're not tiling. | |
676 | ||
677 | GOACC_TILE (collapse_count, loop_no, tile_arg, gwv_tile, gwv_element); */ | |
678 | ||
679 | static void | |
680 | oacc_xform_tile (gcall *call) | |
681 | { | |
682 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
683 | unsigned collapse = tree_to_uhwi (gimple_call_arg (call, 0)); | |
684 | /* Inner loops have higher loop_nos. */ | |
685 | unsigned loop_no = tree_to_uhwi (gimple_call_arg (call, 1)); | |
686 | tree tile_size = gimple_call_arg (call, 2); | |
687 | unsigned e_mask = tree_to_uhwi (gimple_call_arg (call, 4)); | |
688 | tree lhs = gimple_call_lhs (call); | |
689 | tree type = TREE_TYPE (lhs); | |
690 | gimple_seq seq = NULL; | |
691 | tree span = build_int_cst (type, 1); | |
692 | ||
693 | gcc_assert (!(e_mask | |
694 | & ~(GOMP_DIM_MASK (GOMP_DIM_VECTOR) | |
695 | | GOMP_DIM_MASK (GOMP_DIM_WORKER)))); | |
696 | push_gimplify_context (!seen_error ()); | |
697 | ||
698 | #ifndef ACCEL_COMPILER | |
699 | /* Partitioning disabled on host compilers. */ | |
700 | e_mask = 0; | |
701 | #endif | |
702 | if (!e_mask) | |
703 | /* Not paritioning. */ | |
704 | span = integer_one_node; | |
705 | else if (!integer_zerop (tile_size)) | |
706 | /* User explicitly specified size. */ | |
707 | span = tile_size; | |
708 | else | |
709 | { | |
710 | /* Pick a size based on the paritioning of the element loop and | |
711 | the number of loop nests. */ | |
712 | tree first_size = NULL_TREE; | |
713 | tree second_size = NULL_TREE; | |
714 | ||
715 | if (e_mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR)) | |
716 | first_size = oacc_dim_call (false, GOMP_DIM_VECTOR, &seq); | |
717 | if (e_mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)) | |
718 | second_size = oacc_dim_call (false, GOMP_DIM_WORKER, &seq); | |
719 | ||
720 | if (!first_size) | |
721 | { | |
722 | first_size = second_size; | |
723 | second_size = NULL_TREE; | |
724 | } | |
725 | ||
726 | if (loop_no + 1 == collapse) | |
727 | { | |
728 | span = first_size; | |
729 | if (!loop_no && second_size) | |
730 | span = fold_build2 (MULT_EXPR, TREE_TYPE (span), | |
731 | span, second_size); | |
732 | } | |
733 | else if (loop_no + 2 == collapse) | |
734 | span = second_size; | |
735 | else | |
736 | span = NULL_TREE; | |
737 | ||
738 | if (!span) | |
739 | /* There's no obvious element size for this loop. Options | |
740 | are 1, first_size or some non-unity constant (32 is my | |
741 | favourite). We should gather some statistics. */ | |
742 | span = first_size; | |
743 | } | |
744 | ||
745 | span = fold_convert (type, span); | |
746 | gimplify_assign (lhs, span, &seq); | |
747 | ||
748 | pop_gimplify_context (NULL); | |
749 | ||
750 | gsi_replace_with_seq (&gsi, seq, true); | |
751 | } | |
752 | ||
629b3d75 MJ |
753 | /* Default partitioned and minimum partitioned dimensions. */ |
754 | ||
755 | static int oacc_default_dims[GOMP_DIM_MAX]; | |
756 | static int oacc_min_dims[GOMP_DIM_MAX]; | |
757 | ||
b75e9c83 TV |
758 | int |
759 | oacc_get_default_dim (int dim) | |
760 | { | |
761 | gcc_assert (0 <= dim && dim < GOMP_DIM_MAX); | |
762 | return oacc_default_dims[dim]; | |
763 | } | |
764 | ||
6e373d13 TV |
765 | int |
766 | oacc_get_min_dim (int dim) | |
767 | { | |
768 | gcc_assert (0 <= dim && dim < GOMP_DIM_MAX); | |
769 | return oacc_min_dims[dim]; | |
770 | } | |
771 | ||
629b3d75 MJ |
772 | /* Parse the default dimension parameter. This is a set of |
773 | :-separated optional compute dimensions. Each specified dimension | |
774 | is a positive integer. When device type support is added, it is | |
775 | planned to be a comma separated list of such compute dimensions, | |
776 | with all but the first prefixed by the colon-terminated device | |
777 | type. */ | |
778 | ||
779 | static void | |
780 | oacc_parse_default_dims (const char *dims) | |
781 | { | |
782 | int ix; | |
783 | ||
784 | for (ix = GOMP_DIM_MAX; ix--;) | |
785 | { | |
786 | oacc_default_dims[ix] = -1; | |
787 | oacc_min_dims[ix] = 1; | |
788 | } | |
789 | ||
790 | #ifndef ACCEL_COMPILER | |
791 | /* Cannot be overridden on the host. */ | |
792 | dims = NULL; | |
793 | #endif | |
794 | if (dims) | |
795 | { | |
796 | const char *pos = dims; | |
797 | ||
798 | for (ix = 0; *pos && ix != GOMP_DIM_MAX; ix++) | |
799 | { | |
800 | if (ix) | |
801 | { | |
802 | if (*pos != ':') | |
803 | goto malformed; | |
804 | pos++; | |
805 | } | |
806 | ||
807 | if (*pos != ':') | |
808 | { | |
809 | long val; | |
810 | const char *eptr; | |
811 | ||
812 | errno = 0; | |
813 | val = strtol (pos, CONST_CAST (char **, &eptr), 10); | |
814 | if (errno || val <= 0 || (int) val != val) | |
815 | goto malformed; | |
816 | pos = eptr; | |
817 | oacc_default_dims[ix] = (int) val; | |
818 | } | |
819 | } | |
820 | if (*pos) | |
821 | { | |
822 | malformed: | |
823 | error_at (UNKNOWN_LOCATION, | |
904f3daa | 824 | "%<-fopenacc-dim%> operand is malformed at %qs", pos); |
629b3d75 MJ |
825 | } |
826 | } | |
827 | ||
828 | /* Allow the backend to validate the dimensions. */ | |
46dedae6 TV |
829 | targetm.goacc.validate_dims (NULL_TREE, oacc_default_dims, -1, 0); |
830 | targetm.goacc.validate_dims (NULL_TREE, oacc_min_dims, -2, 0); | |
629b3d75 MJ |
831 | } |
832 | ||
833 | /* Validate and update the dimensions for offloaded FN. ATTRS is the | |
834 | raw attribute. DIMS is an array of dimensions, which is filled in. | |
835 | LEVEL is the partitioning level of a routine, or -1 for an offload | |
01914336 | 836 | region itself. USED is the mask of partitioned execution in the |
629b3d75 MJ |
837 | function. */ |
838 | ||
839 | static void | |
840 | oacc_validate_dims (tree fn, tree attrs, int *dims, int level, unsigned used) | |
841 | { | |
842 | tree purpose[GOMP_DIM_MAX]; | |
843 | unsigned ix; | |
844 | tree pos = TREE_VALUE (attrs); | |
629b3d75 MJ |
845 | |
846 | /* Make sure the attribute creator attached the dimension | |
847 | information. */ | |
848 | gcc_assert (pos); | |
849 | ||
850 | for (ix = 0; ix != GOMP_DIM_MAX; ix++) | |
851 | { | |
852 | purpose[ix] = TREE_PURPOSE (pos); | |
853 | tree val = TREE_VALUE (pos); | |
854 | dims[ix] = val ? TREE_INT_CST_LOW (val) : -1; | |
855 | pos = TREE_CHAIN (pos); | |
856 | } | |
857 | ||
46dedae6 | 858 | bool changed = targetm.goacc.validate_dims (fn, dims, level, used); |
629b3d75 MJ |
859 | |
860 | /* Default anything left to 1 or a partitioned default. */ | |
861 | for (ix = 0; ix != GOMP_DIM_MAX; ix++) | |
862 | if (dims[ix] < 0) | |
863 | { | |
864 | /* The OpenACC spec says 'If the [num_gangs] clause is not | |
865 | specified, an implementation-defined default will be used; | |
866 | the default may depend on the code within the construct.' | |
867 | (2.5.6). Thus an implementation is free to choose | |
868 | non-unity default for a parallel region that doesn't have | |
869 | any gang-partitioned loops. However, it appears that there | |
870 | is a sufficient body of user code that expects non-gang | |
871 | partitioned regions to not execute in gang-redundant mode. | |
872 | So we (a) don't warn about the non-portability and (b) pick | |
873 | the minimum permissible dimension size when there is no | |
874 | partitioned execution. Otherwise we pick the global | |
875 | default for the dimension, which the user can control. The | |
876 | same wording and logic applies to num_workers and | |
877 | vector_length, however the worker- or vector- single | |
878 | execution doesn't have the same impact as gang-redundant | |
879 | execution. (If the minimum gang-level partioning is not 1, | |
880 | the target is probably too confusing.) */ | |
881 | dims[ix] = (used & GOMP_DIM_MASK (ix) | |
882 | ? oacc_default_dims[ix] : oacc_min_dims[ix]); | |
883 | changed = true; | |
884 | } | |
885 | ||
886 | if (changed) | |
887 | { | |
888 | /* Replace the attribute with new values. */ | |
889 | pos = NULL_TREE; | |
890 | for (ix = GOMP_DIM_MAX; ix--;) | |
25651634 TS |
891 | pos = tree_cons (purpose[ix], |
892 | build_int_cst (integer_type_node, dims[ix]), pos); | |
629b3d75 MJ |
893 | oacc_replace_fn_attrib (fn, pos); |
894 | } | |
895 | } | |
896 | ||
897 | /* Create an empty OpenACC loop structure at LOC. */ | |
898 | ||
899 | static oacc_loop * | |
900 | new_oacc_loop_raw (oacc_loop *parent, location_t loc) | |
901 | { | |
902 | oacc_loop *loop = XCNEW (oacc_loop); | |
903 | ||
904 | loop->parent = parent; | |
629b3d75 MJ |
905 | |
906 | if (parent) | |
907 | { | |
908 | loop->sibling = parent->child; | |
909 | parent->child = loop; | |
910 | } | |
911 | ||
912 | loop->loc = loc; | |
629b3d75 MJ |
913 | return loop; |
914 | } | |
915 | ||
916 | /* Create an outermost, dummy OpenACC loop for offloaded function | |
917 | DECL. */ | |
918 | ||
919 | static oacc_loop * | |
920 | new_oacc_loop_outer (tree decl) | |
921 | { | |
922 | return new_oacc_loop_raw (NULL, DECL_SOURCE_LOCATION (decl)); | |
923 | } | |
924 | ||
925 | /* Start a new OpenACC loop structure beginning at head marker HEAD. | |
926 | Link into PARENT loop. Return the new loop. */ | |
927 | ||
928 | static oacc_loop * | |
929 | new_oacc_loop (oacc_loop *parent, gcall *marker) | |
930 | { | |
931 | oacc_loop *loop = new_oacc_loop_raw (parent, gimple_location (marker)); | |
932 | ||
933 | loop->marker = marker; | |
934 | ||
935 | /* TODO: This is where device_type flattening would occur for the loop | |
01914336 | 936 | flags. */ |
629b3d75 MJ |
937 | |
938 | loop->flags = TREE_INT_CST_LOW (gimple_call_arg (marker, 3)); | |
939 | ||
940 | tree chunk_size = integer_zero_node; | |
941 | if (loop->flags & OLF_GANG_STATIC) | |
942 | chunk_size = gimple_call_arg (marker, 4); | |
943 | loop->chunk_size = chunk_size; | |
944 | ||
945 | return loop; | |
946 | } | |
947 | ||
948 | /* Create a dummy loop encompassing a call to a openACC routine. | |
949 | Extract the routine's partitioning requirements. */ | |
950 | ||
951 | static void | |
952 | new_oacc_loop_routine (oacc_loop *parent, gcall *call, tree decl, tree attrs) | |
953 | { | |
954 | oacc_loop *loop = new_oacc_loop_raw (parent, gimple_location (call)); | |
955 | int level = oacc_fn_attrib_level (attrs); | |
956 | ||
957 | gcc_assert (level >= 0); | |
958 | ||
959 | loop->marker = call; | |
960 | loop->routine = decl; | |
961 | loop->mask = ((GOMP_DIM_MASK (GOMP_DIM_MAX) - 1) | |
962 | ^ (GOMP_DIM_MASK (level) - 1)); | |
963 | } | |
964 | ||
965 | /* Finish off the current OpenACC loop ending at tail marker TAIL. | |
966 | Return the parent loop. */ | |
967 | ||
968 | static oacc_loop * | |
969 | finish_oacc_loop (oacc_loop *loop) | |
970 | { | |
971 | /* If the loop has been collapsed, don't partition it. */ | |
02889d23 | 972 | if (loop->ifns.is_empty ()) |
629b3d75 MJ |
973 | loop->mask = loop->flags = 0; |
974 | return loop->parent; | |
975 | } | |
976 | ||
977 | /* Free all OpenACC loop structures within LOOP (inclusive). */ | |
978 | ||
979 | static void | |
980 | free_oacc_loop (oacc_loop *loop) | |
981 | { | |
982 | if (loop->sibling) | |
983 | free_oacc_loop (loop->sibling); | |
984 | if (loop->child) | |
985 | free_oacc_loop (loop->child); | |
986 | ||
622f6b64 | 987 | loop->ifns.release (); |
629b3d75 MJ |
988 | free (loop); |
989 | } | |
990 | ||
991 | /* Dump out the OpenACC loop head or tail beginning at FROM. */ | |
992 | ||
993 | static void | |
994 | dump_oacc_loop_part (FILE *file, gcall *from, int depth, | |
995 | const char *title, int level) | |
996 | { | |
997 | enum ifn_unique_kind kind | |
998 | = (enum ifn_unique_kind) TREE_INT_CST_LOW (gimple_call_arg (from, 0)); | |
999 | ||
1000 | fprintf (file, "%*s%s-%d:\n", depth * 2, "", title, level); | |
1001 | for (gimple_stmt_iterator gsi = gsi_for_stmt (from);;) | |
1002 | { | |
1003 | gimple *stmt = gsi_stmt (gsi); | |
1004 | ||
1005 | if (gimple_call_internal_p (stmt, IFN_UNIQUE)) | |
1006 | { | |
1007 | enum ifn_unique_kind k | |
1008 | = ((enum ifn_unique_kind) TREE_INT_CST_LOW | |
1009 | (gimple_call_arg (stmt, 0))); | |
1010 | ||
1011 | if (k == kind && stmt != from) | |
1012 | break; | |
1013 | } | |
ef6cb4c7 | 1014 | print_gimple_stmt (file, stmt, depth * 2 + 2); |
629b3d75 MJ |
1015 | |
1016 | gsi_next (&gsi); | |
1017 | while (gsi_end_p (gsi)) | |
1018 | gsi = gsi_start_bb (single_succ (gsi_bb (gsi))); | |
1019 | } | |
1020 | } | |
1021 | ||
5d390fd3 | 1022 | /* Dump OpenACC loop LOOP, its children, and its siblings. */ |
629b3d75 MJ |
1023 | |
1024 | static void | |
1025 | dump_oacc_loop (FILE *file, oacc_loop *loop, int depth) | |
1026 | { | |
1027 | int ix; | |
1028 | ||
1029 | fprintf (file, "%*sLoop %x(%x) %s:%u\n", depth * 2, "", | |
1030 | loop->flags, loop->mask, | |
1031 | LOCATION_FILE (loop->loc), LOCATION_LINE (loop->loc)); | |
1032 | ||
1033 | if (loop->marker) | |
ef6cb4c7 | 1034 | print_gimple_stmt (file, loop->marker, depth * 2); |
629b3d75 MJ |
1035 | |
1036 | if (loop->routine) | |
1037 | fprintf (file, "%*sRoutine %s:%u:%s\n", | |
1038 | depth * 2, "", DECL_SOURCE_FILE (loop->routine), | |
1039 | DECL_SOURCE_LINE (loop->routine), | |
1040 | IDENTIFIER_POINTER (DECL_NAME (loop->routine))); | |
1041 | ||
1042 | for (ix = GOMP_DIM_GANG; ix != GOMP_DIM_MAX; ix++) | |
1043 | if (loop->heads[ix]) | |
1044 | dump_oacc_loop_part (file, loop->heads[ix], depth, "Head", ix); | |
1045 | for (ix = GOMP_DIM_MAX; ix--;) | |
1046 | if (loop->tails[ix]) | |
1047 | dump_oacc_loop_part (file, loop->tails[ix], depth, "Tail", ix); | |
1048 | ||
1049 | if (loop->child) | |
1050 | dump_oacc_loop (file, loop->child, depth + 1); | |
1051 | if (loop->sibling) | |
1052 | dump_oacc_loop (file, loop->sibling, depth); | |
1053 | } | |
1054 | ||
1055 | void debug_oacc_loop (oacc_loop *); | |
1056 | ||
1057 | /* Dump loops to stderr. */ | |
1058 | ||
1059 | DEBUG_FUNCTION void | |
1060 | debug_oacc_loop (oacc_loop *loop) | |
1061 | { | |
1062 | dump_oacc_loop (stderr, loop, 0); | |
1063 | } | |
1064 | ||
5d390fd3 TS |
1065 | /* Provide diagnostics on OpenACC loop LOOP, its children, and its |
1066 | siblings. */ | |
1067 | ||
1068 | static void | |
1069 | inform_oacc_loop (const oacc_loop *loop) | |
1070 | { | |
1071 | const char *gang | |
1072 | = loop->mask & GOMP_DIM_MASK (GOMP_DIM_GANG) ? " gang" : ""; | |
1073 | const char *worker | |
1074 | = loop->mask & GOMP_DIM_MASK (GOMP_DIM_WORKER) ? " worker" : ""; | |
1075 | const char *vector | |
1076 | = loop->mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR) ? " vector" : ""; | |
1077 | const char *seq = loop->mask == 0 ? " seq" : ""; | |
1078 | const dump_user_location_t loc | |
1079 | = dump_user_location_t::from_location_t (loop->loc); | |
1080 | dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, loc, | |
1081 | "assigned OpenACC%s%s%s%s loop parallelism\n", gang, worker, | |
1082 | vector, seq); | |
1083 | ||
1084 | if (loop->child) | |
1085 | inform_oacc_loop (loop->child); | |
1086 | if (loop->sibling) | |
1087 | inform_oacc_loop (loop->sibling); | |
1088 | } | |
1089 | ||
629b3d75 MJ |
1090 | /* DFS walk of basic blocks BB onwards, creating OpenACC loop |
1091 | structures as we go. By construction these loops are properly | |
1092 | nested. */ | |
1093 | ||
1094 | static void | |
1095 | oacc_loop_discover_walk (oacc_loop *loop, basic_block bb) | |
1096 | { | |
1097 | int marker = 0; | |
1098 | int remaining = 0; | |
1099 | ||
1100 | if (bb->flags & BB_VISITED) | |
1101 | return; | |
1102 | ||
1103 | follow: | |
1104 | bb->flags |= BB_VISITED; | |
1105 | ||
1106 | /* Scan for loop markers. */ | |
1107 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi); | |
1108 | gsi_next (&gsi)) | |
1109 | { | |
1110 | gimple *stmt = gsi_stmt (gsi); | |
1111 | ||
1112 | if (!is_gimple_call (stmt)) | |
1113 | continue; | |
1114 | ||
1115 | gcall *call = as_a <gcall *> (stmt); | |
1116 | ||
1117 | /* If this is a routine, make a dummy loop for it. */ | |
1118 | if (tree decl = gimple_call_fndecl (call)) | |
1119 | if (tree attrs = oacc_get_fn_attrib (decl)) | |
1120 | { | |
1121 | gcc_assert (!marker); | |
1122 | new_oacc_loop_routine (loop, call, decl, attrs); | |
1123 | } | |
1124 | ||
1125 | if (!gimple_call_internal_p (call)) | |
1126 | continue; | |
1127 | ||
1128 | switch (gimple_call_internal_fn (call)) | |
1129 | { | |
1130 | default: | |
1131 | break; | |
1132 | ||
1133 | case IFN_GOACC_LOOP: | |
02889d23 CLT |
1134 | case IFN_GOACC_TILE: |
1135 | /* Record the abstraction function, so we can manipulate it | |
1136 | later. */ | |
1137 | loop->ifns.safe_push (call); | |
629b3d75 MJ |
1138 | break; |
1139 | ||
1140 | case IFN_UNIQUE: | |
1141 | enum ifn_unique_kind kind | |
1142 | = (enum ifn_unique_kind) (TREE_INT_CST_LOW | |
1143 | (gimple_call_arg (call, 0))); | |
1144 | if (kind == IFN_UNIQUE_OACC_HEAD_MARK | |
1145 | || kind == IFN_UNIQUE_OACC_TAIL_MARK) | |
1146 | { | |
1147 | if (gimple_call_num_args (call) == 2) | |
1148 | { | |
1149 | gcc_assert (marker && !remaining); | |
1150 | marker = 0; | |
1151 | if (kind == IFN_UNIQUE_OACC_TAIL_MARK) | |
1152 | loop = finish_oacc_loop (loop); | |
1153 | else | |
1154 | loop->head_end = call; | |
1155 | } | |
1156 | else | |
1157 | { | |
1158 | int count = TREE_INT_CST_LOW (gimple_call_arg (call, 2)); | |
1159 | ||
1160 | if (!marker) | |
1161 | { | |
1162 | if (kind == IFN_UNIQUE_OACC_HEAD_MARK) | |
1163 | loop = new_oacc_loop (loop, call); | |
1164 | remaining = count; | |
1165 | } | |
1166 | gcc_assert (count == remaining); | |
1167 | if (remaining) | |
1168 | { | |
1169 | remaining--; | |
1170 | if (kind == IFN_UNIQUE_OACC_HEAD_MARK) | |
1171 | loop->heads[marker] = call; | |
1172 | else | |
1173 | loop->tails[remaining] = call; | |
1174 | } | |
1175 | marker++; | |
1176 | } | |
1177 | } | |
1178 | } | |
1179 | } | |
1180 | if (remaining || marker) | |
1181 | { | |
1182 | bb = single_succ (bb); | |
1183 | gcc_assert (single_pred_p (bb) && !(bb->flags & BB_VISITED)); | |
1184 | goto follow; | |
1185 | } | |
1186 | ||
1187 | /* Walk successor blocks. */ | |
1188 | edge e; | |
1189 | edge_iterator ei; | |
1190 | ||
1191 | FOR_EACH_EDGE (e, ei, bb->succs) | |
1192 | oacc_loop_discover_walk (loop, e->dest); | |
1193 | } | |
1194 | ||
1195 | /* LOOP is the first sibling. Reverse the order in place and return | |
1196 | the new first sibling. Recurse to child loops. */ | |
1197 | ||
1198 | static oacc_loop * | |
1199 | oacc_loop_sibling_nreverse (oacc_loop *loop) | |
1200 | { | |
1201 | oacc_loop *last = NULL; | |
1202 | do | |
1203 | { | |
1204 | if (loop->child) | |
01914336 | 1205 | loop->child = oacc_loop_sibling_nreverse (loop->child); |
629b3d75 MJ |
1206 | |
1207 | oacc_loop *next = loop->sibling; | |
1208 | loop->sibling = last; | |
1209 | last = loop; | |
1210 | loop = next; | |
1211 | } | |
1212 | while (loop); | |
1213 | ||
1214 | return last; | |
1215 | } | |
1216 | ||
1217 | /* Discover the OpenACC loops marked up by HEAD and TAIL markers for | |
1218 | the current function. */ | |
1219 | ||
1220 | static oacc_loop * | |
1221 | oacc_loop_discovery () | |
1222 | { | |
1223 | /* Clear basic block flags, in particular BB_VISITED which we're going to use | |
1224 | in the following. */ | |
1225 | clear_bb_flags (); | |
1226 | ||
1227 | oacc_loop *top = new_oacc_loop_outer (current_function_decl); | |
1228 | oacc_loop_discover_walk (top, ENTRY_BLOCK_PTR_FOR_FN (cfun)); | |
1229 | ||
1230 | /* The siblings were constructed in reverse order, reverse them so | |
1231 | that diagnostics come out in an unsurprising order. */ | |
1232 | top = oacc_loop_sibling_nreverse (top); | |
1233 | ||
1234 | return top; | |
1235 | } | |
1236 | ||
1237 | /* Transform the abstract internal function markers starting at FROM | |
1238 | to be for partitioning level LEVEL. Stop when we meet another HEAD | |
1239 | or TAIL marker. */ | |
1240 | ||
1241 | static void | |
1242 | oacc_loop_xform_head_tail (gcall *from, int level) | |
1243 | { | |
1244 | enum ifn_unique_kind kind | |
1245 | = (enum ifn_unique_kind) TREE_INT_CST_LOW (gimple_call_arg (from, 0)); | |
1246 | tree replacement = build_int_cst (unsigned_type_node, level); | |
1247 | ||
1248 | for (gimple_stmt_iterator gsi = gsi_for_stmt (from);;) | |
1249 | { | |
1250 | gimple *stmt = gsi_stmt (gsi); | |
1251 | ||
1252 | if (gimple_call_internal_p (stmt, IFN_UNIQUE)) | |
1253 | { | |
1254 | enum ifn_unique_kind k | |
1255 | = ((enum ifn_unique_kind) | |
1256 | TREE_INT_CST_LOW (gimple_call_arg (stmt, 0))); | |
1257 | ||
1258 | if (k == IFN_UNIQUE_OACC_FORK || k == IFN_UNIQUE_OACC_JOIN) | |
1259 | *gimple_call_arg_ptr (stmt, 2) = replacement; | |
1260 | else if (k == kind && stmt != from) | |
1261 | break; | |
1262 | } | |
1263 | else if (gimple_call_internal_p (stmt, IFN_GOACC_REDUCTION)) | |
1264 | *gimple_call_arg_ptr (stmt, 3) = replacement; | |
1265 | ||
1266 | gsi_next (&gsi); | |
1267 | while (gsi_end_p (gsi)) | |
1268 | gsi = gsi_start_bb (single_succ (gsi_bb (gsi))); | |
1269 | } | |
1270 | } | |
1271 | ||
629b3d75 MJ |
1272 | /* Process the discovered OpenACC loops, setting the correct |
1273 | partitioning level etc. */ | |
1274 | ||
1275 | static void | |
1276 | oacc_loop_process (oacc_loop *loop) | |
1277 | { | |
1278 | if (loop->child) | |
1279 | oacc_loop_process (loop->child); | |
1280 | ||
1281 | if (loop->mask && !loop->routine) | |
1282 | { | |
1283 | int ix; | |
02889d23 CLT |
1284 | tree mask_arg = build_int_cst (unsigned_type_node, loop->mask); |
1285 | tree e_mask_arg = build_int_cst (unsigned_type_node, loop->e_mask); | |
629b3d75 | 1286 | tree chunk_arg = loop->chunk_size; |
02889d23 CLT |
1287 | gcall *call; |
1288 | ||
1289 | for (ix = 0; loop->ifns.iterate (ix, &call); ix++) | |
1290 | switch (gimple_call_internal_fn (call)) | |
1291 | { | |
1292 | case IFN_GOACC_LOOP: | |
1293 | { | |
1294 | bool is_e = gimple_call_arg (call, 5) == integer_minus_one_node; | |
1295 | gimple_call_set_arg (call, 5, is_e ? e_mask_arg : mask_arg); | |
1296 | if (!is_e) | |
1297 | gimple_call_set_arg (call, 4, chunk_arg); | |
1298 | } | |
1299 | break; | |
1300 | ||
1301 | case IFN_GOACC_TILE: | |
1302 | gimple_call_set_arg (call, 3, mask_arg); | |
1303 | gimple_call_set_arg (call, 4, e_mask_arg); | |
1304 | break; | |
629b3d75 | 1305 | |
02889d23 CLT |
1306 | default: |
1307 | gcc_unreachable (); | |
1308 | } | |
629b3d75 | 1309 | |
02889d23 CLT |
1310 | unsigned dim = GOMP_DIM_GANG; |
1311 | unsigned mask = loop->mask | loop->e_mask; | |
629b3d75 MJ |
1312 | for (ix = 0; ix != GOMP_DIM_MAX && mask; ix++) |
1313 | { | |
1314 | while (!(GOMP_DIM_MASK (dim) & mask)) | |
1315 | dim++; | |
1316 | ||
1317 | oacc_loop_xform_head_tail (loop->heads[ix], dim); | |
1318 | oacc_loop_xform_head_tail (loop->tails[ix], dim); | |
1319 | ||
1320 | mask ^= GOMP_DIM_MASK (dim); | |
1321 | } | |
1322 | } | |
1323 | ||
1324 | if (loop->sibling) | |
1325 | oacc_loop_process (loop->sibling); | |
1326 | } | |
1327 | ||
1328 | /* Walk the OpenACC loop heirarchy checking and assigning the | |
1329 | programmer-specified partitionings. OUTER_MASK is the partitioning | |
1330 | this loop is contained within. Return mask of partitioning | |
1331 | encountered. If any auto loops are discovered, set GOMP_DIM_MAX | |
1332 | bit. */ | |
1333 | ||
1334 | static unsigned | |
1335 | oacc_loop_fixed_partitions (oacc_loop *loop, unsigned outer_mask) | |
1336 | { | |
1337 | unsigned this_mask = loop->mask; | |
1338 | unsigned mask_all = 0; | |
1339 | bool noisy = true; | |
1340 | ||
1341 | #ifdef ACCEL_COMPILER | |
1342 | /* When device_type is supported, we want the device compiler to be | |
1343 | noisy, if the loop parameters are device_type-specific. */ | |
1344 | noisy = false; | |
1345 | #endif | |
1346 | ||
1347 | if (!loop->routine) | |
1348 | { | |
1349 | bool auto_par = (loop->flags & OLF_AUTO) != 0; | |
1350 | bool seq_par = (loop->flags & OLF_SEQ) != 0; | |
02889d23 CLT |
1351 | bool tiling = (loop->flags & OLF_TILE) != 0; |
1352 | ||
629b3d75 MJ |
1353 | this_mask = ((loop->flags >> OLF_DIM_BASE) |
1354 | & (GOMP_DIM_MASK (GOMP_DIM_MAX) - 1)); | |
1355 | ||
02889d23 CLT |
1356 | /* Apply auto partitioning if this is a non-partitioned regular |
1357 | loop, or (no more than) single axis tiled loop. */ | |
1358 | bool maybe_auto | |
1359 | = !seq_par && this_mask == (tiling ? this_mask & -this_mask : 0); | |
1360 | ||
629b3d75 MJ |
1361 | if ((this_mask != 0) + auto_par + seq_par > 1) |
1362 | { | |
1363 | if (noisy) | |
1364 | error_at (loop->loc, | |
1365 | seq_par | |
324ff1a0 JJ |
1366 | ? G_("%<seq%> overrides other OpenACC loop specifiers") |
1367 | : G_("%<auto%> conflicts with other OpenACC loop " | |
1368 | "specifiers")); | |
02889d23 | 1369 | maybe_auto = false; |
629b3d75 MJ |
1370 | loop->flags &= ~OLF_AUTO; |
1371 | if (seq_par) | |
1372 | { | |
01914336 MJ |
1373 | loop->flags |
1374 | &= ~((GOMP_DIM_MASK (GOMP_DIM_MAX) - 1) << OLF_DIM_BASE); | |
629b3d75 MJ |
1375 | this_mask = 0; |
1376 | } | |
1377 | } | |
02889d23 CLT |
1378 | |
1379 | if (maybe_auto && (loop->flags & OLF_INDEPENDENT)) | |
1380 | { | |
1381 | loop->flags |= OLF_AUTO; | |
1382 | mask_all |= GOMP_DIM_MASK (GOMP_DIM_MAX); | |
1383 | } | |
629b3d75 MJ |
1384 | } |
1385 | ||
1386 | if (this_mask & outer_mask) | |
1387 | { | |
1388 | const oacc_loop *outer; | |
1389 | for (outer = loop->parent; outer; outer = outer->parent) | |
02889d23 | 1390 | if ((outer->mask | outer->e_mask) & this_mask) |
629b3d75 MJ |
1391 | break; |
1392 | ||
1393 | if (noisy) | |
1394 | { | |
1395 | if (outer) | |
1396 | { | |
1397 | error_at (loop->loc, | |
efebb49e DM |
1398 | loop->routine |
1399 | ? G_("routine call uses same OpenACC parallelism" | |
1400 | " as containing loop") | |
1401 | : G_("inner loop uses same OpenACC parallelism" | |
1402 | " as containing loop")); | |
629b3d75 MJ |
1403 | inform (outer->loc, "containing loop here"); |
1404 | } | |
1405 | else | |
1406 | error_at (loop->loc, | |
efebb49e DM |
1407 | loop->routine |
1408 | ? G_("routine call uses OpenACC parallelism disallowed" | |
1409 | " by containing routine") | |
1410 | : G_("loop uses OpenACC parallelism disallowed" | |
1411 | " by containing routine")); | |
629b3d75 MJ |
1412 | |
1413 | if (loop->routine) | |
1414 | inform (DECL_SOURCE_LOCATION (loop->routine), | |
1415 | "routine %qD declared here", loop->routine); | |
1416 | } | |
1417 | this_mask &= ~outer_mask; | |
1418 | } | |
1419 | else | |
1420 | { | |
1421 | unsigned outermost = least_bit_hwi (this_mask); | |
1422 | ||
1423 | if (outermost && outermost <= outer_mask) | |
1424 | { | |
1425 | if (noisy) | |
1426 | { | |
1427 | error_at (loop->loc, | |
1428 | "incorrectly nested OpenACC loop parallelism"); | |
1429 | ||
1430 | const oacc_loop *outer; | |
1431 | for (outer = loop->parent; | |
1432 | outer->flags && outer->flags < outermost; | |
1433 | outer = outer->parent) | |
1434 | continue; | |
1435 | inform (outer->loc, "containing loop here"); | |
1436 | } | |
1437 | ||
1438 | this_mask &= ~outermost; | |
1439 | } | |
1440 | } | |
1441 | ||
629b3d75 MJ |
1442 | mask_all |= this_mask; |
1443 | ||
02889d23 CLT |
1444 | if (loop->flags & OLF_TILE) |
1445 | { | |
1446 | /* When tiling, vector goes to the element loop, and failing | |
1447 | that we put worker there. The std doesn't contemplate | |
1448 | specifying all three. We choose to put worker and vector on | |
1449 | the element loops in that case. */ | |
1450 | unsigned this_e_mask = this_mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR); | |
1451 | if (!this_e_mask || this_mask & GOMP_DIM_MASK (GOMP_DIM_GANG)) | |
1452 | this_e_mask |= this_mask & GOMP_DIM_MASK (GOMP_DIM_WORKER); | |
1453 | ||
1454 | loop->e_mask = this_e_mask; | |
1455 | this_mask ^= this_e_mask; | |
1456 | } | |
1457 | ||
1458 | loop->mask = this_mask; | |
1459 | ||
1460 | if (dump_file) | |
1461 | fprintf (dump_file, "Loop %s:%d user specified %d & %d\n", | |
1462 | LOCATION_FILE (loop->loc), LOCATION_LINE (loop->loc), | |
1463 | loop->mask, loop->e_mask); | |
1464 | ||
629b3d75 MJ |
1465 | if (loop->child) |
1466 | { | |
02889d23 CLT |
1467 | unsigned tmp_mask = outer_mask | this_mask | loop->e_mask; |
1468 | loop->inner = oacc_loop_fixed_partitions (loop->child, tmp_mask); | |
629b3d75 MJ |
1469 | mask_all |= loop->inner; |
1470 | } | |
1471 | ||
1472 | if (loop->sibling) | |
1473 | mask_all |= oacc_loop_fixed_partitions (loop->sibling, outer_mask); | |
1474 | ||
1475 | return mask_all; | |
1476 | } | |
1477 | ||
1478 | /* Walk the OpenACC loop heirarchy to assign auto-partitioned loops. | |
1479 | OUTER_MASK is the partitioning this loop is contained within. | |
02889d23 | 1480 | OUTER_ASSIGN is true if an outer loop is being auto-partitioned. |
629b3d75 MJ |
1481 | Return the cumulative partitioning used by this loop, siblings and |
1482 | children. */ | |
1483 | ||
1484 | static unsigned | |
02889d23 CLT |
1485 | oacc_loop_auto_partitions (oacc_loop *loop, unsigned outer_mask, |
1486 | bool outer_assign) | |
629b3d75 MJ |
1487 | { |
1488 | bool assign = (loop->flags & OLF_AUTO) && (loop->flags & OLF_INDEPENDENT); | |
1489 | bool noisy = true; | |
02889d23 | 1490 | bool tiling = loop->flags & OLF_TILE; |
629b3d75 MJ |
1491 | |
1492 | #ifdef ACCEL_COMPILER | |
1493 | /* When device_type is supported, we want the device compiler to be | |
1494 | noisy, if the loop parameters are device_type-specific. */ | |
1495 | noisy = false; | |
1496 | #endif | |
1497 | ||
891ba5eb | 1498 | if (assign && (!outer_assign || loop->inner)) |
629b3d75 | 1499 | { |
02889d23 CLT |
1500 | /* Allocate outermost and non-innermost loops at the outermost |
1501 | non-innermost available level. */ | |
1502 | unsigned this_mask = GOMP_DIM_MASK (GOMP_DIM_GANG); | |
1503 | ||
1504 | /* Find the first outermost available partition. */ | |
1505 | while (this_mask <= outer_mask) | |
1506 | this_mask <<= 1; | |
1507 | ||
1508 | /* Grab two axes if tiling, and we've not assigned anything */ | |
1509 | if (tiling && !(loop->mask | loop->e_mask)) | |
1510 | this_mask |= this_mask << 1; | |
1511 | ||
1512 | /* Prohibit the innermost partitioning at the moment. */ | |
1513 | this_mask &= GOMP_DIM_MASK (GOMP_DIM_MAX - 1) - 1; | |
629b3d75 | 1514 | |
02889d23 CLT |
1515 | /* Don't use any dimension explicitly claimed by an inner loop. */ |
1516 | this_mask &= ~loop->inner; | |
1517 | ||
1518 | if (tiling && !loop->e_mask) | |
1519 | { | |
1520 | /* If we got two axes, allocate the inner one to the element | |
1521 | loop. */ | |
1522 | loop->e_mask = this_mask & (this_mask << 1); | |
1523 | this_mask ^= loop->e_mask; | |
1524 | } | |
1525 | ||
1526 | loop->mask |= this_mask; | |
629b3d75 MJ |
1527 | } |
1528 | ||
1529 | if (loop->child) | |
1530 | { | |
02889d23 CLT |
1531 | unsigned tmp_mask = outer_mask | loop->mask | loop->e_mask; |
1532 | loop->inner = oacc_loop_auto_partitions (loop->child, tmp_mask, | |
1533 | outer_assign | assign); | |
629b3d75 MJ |
1534 | } |
1535 | ||
02889d23 | 1536 | if (assign && (!loop->mask || (tiling && !loop->e_mask) || !outer_assign)) |
629b3d75 | 1537 | { |
02889d23 CLT |
1538 | /* Allocate the loop at the innermost available level. Note |
1539 | that we do this even if we already assigned this loop the | |
1540 | outermost available level above. That way we'll partition | |
1541 | this along 2 axes, if they are available. */ | |
629b3d75 MJ |
1542 | unsigned this_mask = 0; |
1543 | ||
01914336 | 1544 | /* Determine the outermost partitioning used within this loop. */ |
629b3d75 MJ |
1545 | this_mask = loop->inner | GOMP_DIM_MASK (GOMP_DIM_MAX); |
1546 | this_mask = least_bit_hwi (this_mask); | |
1547 | ||
1548 | /* Pick the partitioning just inside that one. */ | |
1549 | this_mask >>= 1; | |
1550 | ||
01914336 | 1551 | /* And avoid picking one use by an outer loop. */ |
629b3d75 MJ |
1552 | this_mask &= ~outer_mask; |
1553 | ||
02889d23 CLT |
1554 | /* If tiling and we failed completely above, grab the next one |
1555 | too. Making sure it doesn't hit an outer loop. */ | |
1556 | if (tiling) | |
1557 | { | |
1558 | this_mask &= ~(loop->e_mask | loop->mask); | |
1559 | unsigned tile_mask = ((this_mask >> 1) | |
1560 | & ~(outer_mask | loop->e_mask | loop->mask)); | |
1561 | ||
1562 | if (tile_mask || loop->mask) | |
1563 | { | |
1564 | loop->e_mask |= this_mask; | |
1565 | this_mask = tile_mask; | |
1566 | } | |
1567 | if (!loop->e_mask && noisy) | |
1568 | warning_at (loop->loc, 0, | |
1569 | "insufficient partitioning available" | |
1570 | " to parallelize element loop"); | |
1571 | } | |
629b3d75 | 1572 | |
02889d23 CLT |
1573 | loop->mask |= this_mask; |
1574 | if (!loop->mask && noisy) | |
1575 | warning_at (loop->loc, 0, | |
efebb49e DM |
1576 | tiling |
1577 | ? G_("insufficient partitioning available" | |
1578 | " to parallelize tile loop") | |
1579 | : G_("insufficient partitioning available" | |
1580 | " to parallelize loop")); | |
629b3d75 MJ |
1581 | } |
1582 | ||
1583 | if (assign && dump_file) | |
02889d23 | 1584 | fprintf (dump_file, "Auto loop %s:%d assigned %d & %d\n", |
629b3d75 | 1585 | LOCATION_FILE (loop->loc), LOCATION_LINE (loop->loc), |
02889d23 | 1586 | loop->mask, loop->e_mask); |
629b3d75 MJ |
1587 | |
1588 | unsigned inner_mask = 0; | |
1589 | ||
1590 | if (loop->sibling) | |
02889d23 CLT |
1591 | inner_mask |= oacc_loop_auto_partitions (loop->sibling, |
1592 | outer_mask, outer_assign); | |
629b3d75 | 1593 | |
02889d23 | 1594 | inner_mask |= loop->inner | loop->mask | loop->e_mask; |
629b3d75 MJ |
1595 | |
1596 | return inner_mask; | |
1597 | } | |
1598 | ||
1599 | /* Walk the OpenACC loop heirarchy to check and assign partitioning | |
1600 | axes. Return mask of partitioning. */ | |
1601 | ||
1602 | static unsigned | |
1603 | oacc_loop_partition (oacc_loop *loop, unsigned outer_mask) | |
1604 | { | |
1605 | unsigned mask_all = oacc_loop_fixed_partitions (loop, outer_mask); | |
1606 | ||
1607 | if (mask_all & GOMP_DIM_MASK (GOMP_DIM_MAX)) | |
1608 | { | |
1609 | mask_all ^= GOMP_DIM_MASK (GOMP_DIM_MAX); | |
02889d23 | 1610 | mask_all |= oacc_loop_auto_partitions (loop, outer_mask, false); |
629b3d75 MJ |
1611 | } |
1612 | return mask_all; | |
1613 | } | |
1614 | ||
1615 | /* Default fork/join early expander. Delete the function calls if | |
1616 | there is no RTL expander. */ | |
1617 | ||
1618 | bool | |
1619 | default_goacc_fork_join (gcall *ARG_UNUSED (call), | |
1620 | const int *ARG_UNUSED (dims), bool is_fork) | |
1621 | { | |
1622 | if (is_fork) | |
1623 | return targetm.have_oacc_fork (); | |
1624 | else | |
1625 | return targetm.have_oacc_join (); | |
1626 | } | |
1627 | ||
1628 | /* Default goacc.reduction early expander. | |
1629 | ||
1630 | LHS-opt = IFN_REDUCTION (KIND, RES_PTR, VAR, LEVEL, OP, OFFSET) | |
1631 | If RES_PTR is not integer-zerop: | |
1632 | SETUP - emit 'LHS = *RES_PTR', LHS = NULL | |
1633 | TEARDOWN - emit '*RES_PTR = VAR' | |
1634 | If LHS is not NULL | |
1635 | emit 'LHS = VAR' */ | |
1636 | ||
1637 | void | |
1638 | default_goacc_reduction (gcall *call) | |
1639 | { | |
1640 | unsigned code = (unsigned)TREE_INT_CST_LOW (gimple_call_arg (call, 0)); | |
1641 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
1642 | tree lhs = gimple_call_lhs (call); | |
1643 | tree var = gimple_call_arg (call, 2); | |
1644 | gimple_seq seq = NULL; | |
1645 | ||
1646 | if (code == IFN_GOACC_REDUCTION_SETUP | |
1647 | || code == IFN_GOACC_REDUCTION_TEARDOWN) | |
1648 | { | |
1649 | /* Setup and Teardown need to copy from/to the receiver object, | |
1650 | if there is one. */ | |
1651 | tree ref_to_res = gimple_call_arg (call, 1); | |
1652 | ||
1653 | if (!integer_zerop (ref_to_res)) | |
1654 | { | |
1655 | tree dst = build_simple_mem_ref (ref_to_res); | |
1656 | tree src = var; | |
1657 | ||
1658 | if (code == IFN_GOACC_REDUCTION_SETUP) | |
1659 | { | |
1660 | src = dst; | |
1661 | dst = lhs; | |
1662 | lhs = NULL; | |
1663 | } | |
1664 | gimple_seq_add_stmt (&seq, gimple_build_assign (dst, src)); | |
1665 | } | |
1666 | } | |
1667 | ||
1668 | /* Copy VAR to LHS, if there is an LHS. */ | |
1669 | if (lhs) | |
1670 | gimple_seq_add_stmt (&seq, gimple_build_assign (lhs, var)); | |
1671 | ||
1672 | gsi_replace_with_seq (&gsi, seq, true); | |
1673 | } | |
1674 | ||
1675 | /* Main entry point for oacc transformations which run on the device | |
1676 | compiler after LTO, so we know what the target device is at this | |
1677 | point (including the host fallback). */ | |
1678 | ||
1679 | static unsigned int | |
1680 | execute_oacc_device_lower () | |
1681 | { | |
1682 | tree attrs = oacc_get_fn_attrib (current_function_decl); | |
1683 | ||
1684 | if (!attrs) | |
1685 | /* Not an offloaded function. */ | |
1686 | return 0; | |
1687 | ||
1688 | /* Parse the default dim argument exactly once. */ | |
1689 | if ((const void *)flag_openacc_dims != &flag_openacc_dims) | |
1690 | { | |
1691 | oacc_parse_default_dims (flag_openacc_dims); | |
1692 | flag_openacc_dims = (char *)&flag_openacc_dims; | |
1693 | } | |
1694 | ||
b0f271ce TS |
1695 | bool is_oacc_kernels |
1696 | = (lookup_attribute ("oacc kernels", | |
1697 | DECL_ATTRIBUTES (current_function_decl)) != NULL); | |
1698 | bool is_oacc_kernels_parallelized | |
1699 | = (lookup_attribute ("oacc kernels parallelized", | |
1700 | DECL_ATTRIBUTES (current_function_decl)) != NULL); | |
1701 | ||
fd71a9a2 TS |
1702 | /* Unparallelized OpenACC kernels constructs must get launched as 1 x 1 x 1 |
1703 | kernels, so remove the parallelism dimensions function attributes | |
1704 | potentially set earlier on. */ | |
1705 | if (is_oacc_kernels && !is_oacc_kernels_parallelized) | |
1706 | { | |
1707 | oacc_set_fn_attrib (current_function_decl, NULL, NULL); | |
1708 | attrs = oacc_get_fn_attrib (current_function_decl); | |
1709 | } | |
1710 | ||
629b3d75 MJ |
1711 | /* Discover, partition and process the loops. */ |
1712 | oacc_loop *loops = oacc_loop_discovery (); | |
1713 | int fn_level = oacc_fn_attrib_level (attrs); | |
1714 | ||
1715 | if (dump_file) | |
25651634 TS |
1716 | { |
1717 | if (fn_level >= 0) | |
1718 | fprintf (dump_file, "Function is OpenACC routine level %d\n", | |
1719 | fn_level); | |
b0f271ce TS |
1720 | else if (is_oacc_kernels) |
1721 | fprintf (dump_file, "Function is %s OpenACC kernels offload\n", | |
1722 | (is_oacc_kernels_parallelized | |
1723 | ? "parallelized" : "unparallelized")); | |
25651634 TS |
1724 | else |
1725 | fprintf (dump_file, "Function is OpenACC parallel offload\n"); | |
1726 | } | |
629b3d75 MJ |
1727 | |
1728 | unsigned outer_mask = fn_level >= 0 ? GOMP_DIM_MASK (fn_level) - 1 : 0; | |
1729 | unsigned used_mask = oacc_loop_partition (loops, outer_mask); | |
b0f271ce TS |
1730 | /* OpenACC kernels constructs are special: they currently don't use the |
1731 | generic oacc_loop infrastructure and attribute/dimension processing. */ | |
1732 | if (is_oacc_kernels && is_oacc_kernels_parallelized) | |
1733 | { | |
1734 | /* Parallelized OpenACC kernels constructs use gang parallelism. See | |
1735 | also tree-parloops.c:create_parallel_loop. */ | |
1736 | used_mask |= GOMP_DIM_MASK (GOMP_DIM_GANG); | |
1737 | } | |
629b3d75 | 1738 | |
b0f271ce | 1739 | int dims[GOMP_DIM_MAX]; |
629b3d75 MJ |
1740 | oacc_validate_dims (current_function_decl, attrs, dims, fn_level, used_mask); |
1741 | ||
1742 | if (dump_file) | |
1743 | { | |
1744 | const char *comma = "Compute dimensions ["; | |
1745 | for (int ix = 0; ix != GOMP_DIM_MAX; ix++, comma = ", ") | |
1746 | fprintf (dump_file, "%s%d", comma, dims[ix]); | |
1747 | fprintf (dump_file, "]\n"); | |
1748 | } | |
1749 | ||
1750 | oacc_loop_process (loops); | |
1751 | if (dump_file) | |
1752 | { | |
1753 | fprintf (dump_file, "OpenACC loops\n"); | |
1754 | dump_oacc_loop (dump_file, loops, 0); | |
1755 | fprintf (dump_file, "\n"); | |
1756 | } | |
5d390fd3 TS |
1757 | if (dump_enabled_p ()) |
1758 | { | |
1759 | oacc_loop *l = loops; | |
1760 | /* OpenACC kernels constructs are special: they currently don't use the | |
1761 | generic oacc_loop infrastructure. */ | |
1762 | if (is_oacc_kernels) | |
1763 | { | |
1764 | /* Create a fake oacc_loop for diagnostic purposes. */ | |
1765 | l = new_oacc_loop_raw (NULL, | |
1766 | DECL_SOURCE_LOCATION (current_function_decl)); | |
1767 | l->mask = used_mask; | |
1768 | } | |
1769 | else | |
1770 | { | |
1771 | /* Skip the outermost, dummy OpenACC loop */ | |
1772 | l = l->child; | |
1773 | } | |
1774 | if (l) | |
1775 | inform_oacc_loop (l); | |
1776 | if (is_oacc_kernels) | |
1777 | free_oacc_loop (l); | |
1778 | } | |
629b3d75 MJ |
1779 | |
1780 | /* Offloaded targets may introduce new basic blocks, which require | |
1781 | dominance information to update SSA. */ | |
1782 | calculate_dominance_info (CDI_DOMINATORS); | |
1783 | ||
1784 | /* Now lower internal loop functions to target-specific code | |
1785 | sequences. */ | |
1786 | basic_block bb; | |
1787 | FOR_ALL_BB_FN (bb, cfun) | |
1788 | for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);) | |
1789 | { | |
1790 | gimple *stmt = gsi_stmt (gsi); | |
1791 | if (!is_gimple_call (stmt)) | |
1792 | { | |
1793 | gsi_next (&gsi); | |
1794 | continue; | |
1795 | } | |
1796 | ||
1797 | gcall *call = as_a <gcall *> (stmt); | |
1798 | if (!gimple_call_internal_p (call)) | |
1799 | { | |
1800 | gsi_next (&gsi); | |
1801 | continue; | |
1802 | } | |
1803 | ||
1804 | /* Rewind to allow rescan. */ | |
1805 | gsi_prev (&gsi); | |
1806 | bool rescan = false, remove = false; | |
1807 | enum internal_fn ifn_code = gimple_call_internal_fn (call); | |
1808 | ||
1809 | switch (ifn_code) | |
1810 | { | |
1811 | default: break; | |
1812 | ||
02889d23 CLT |
1813 | case IFN_GOACC_TILE: |
1814 | oacc_xform_tile (call); | |
1815 | rescan = true; | |
1816 | break; | |
1817 | ||
629b3d75 MJ |
1818 | case IFN_GOACC_LOOP: |
1819 | oacc_xform_loop (call); | |
1820 | rescan = true; | |
1821 | break; | |
1822 | ||
1823 | case IFN_GOACC_REDUCTION: | |
1824 | /* Mark the function for SSA renaming. */ | |
1825 | mark_virtual_operands_for_renaming (cfun); | |
1826 | ||
1827 | /* If the level is -1, this ended up being an unused | |
1828 | axis. Handle as a default. */ | |
1829 | if (integer_minus_onep (gimple_call_arg (call, 3))) | |
1830 | default_goacc_reduction (call); | |
1831 | else | |
1832 | targetm.goacc.reduction (call); | |
1833 | rescan = true; | |
1834 | break; | |
1835 | ||
1836 | case IFN_UNIQUE: | |
1837 | { | |
1838 | enum ifn_unique_kind kind | |
1839 | = ((enum ifn_unique_kind) | |
1840 | TREE_INT_CST_LOW (gimple_call_arg (call, 0))); | |
1841 | ||
1842 | switch (kind) | |
1843 | { | |
1844 | default: | |
02889d23 | 1845 | break; |
629b3d75 MJ |
1846 | |
1847 | case IFN_UNIQUE_OACC_FORK: | |
1848 | case IFN_UNIQUE_OACC_JOIN: | |
1849 | if (integer_minus_onep (gimple_call_arg (call, 2))) | |
1850 | remove = true; | |
1851 | else if (!targetm.goacc.fork_join | |
1852 | (call, dims, kind == IFN_UNIQUE_OACC_FORK)) | |
1853 | remove = true; | |
1854 | break; | |
1855 | ||
1856 | case IFN_UNIQUE_OACC_HEAD_MARK: | |
1857 | case IFN_UNIQUE_OACC_TAIL_MARK: | |
1858 | remove = true; | |
1859 | break; | |
1860 | } | |
1861 | break; | |
1862 | } | |
1863 | } | |
1864 | ||
1865 | if (gsi_end_p (gsi)) | |
1866 | /* We rewound past the beginning of the BB. */ | |
1867 | gsi = gsi_start_bb (bb); | |
1868 | else | |
1869 | /* Undo the rewind. */ | |
1870 | gsi_next (&gsi); | |
1871 | ||
1872 | if (remove) | |
1873 | { | |
1874 | if (gimple_vdef (call)) | |
1875 | replace_uses_by (gimple_vdef (call), gimple_vuse (call)); | |
1876 | if (gimple_call_lhs (call)) | |
1877 | { | |
1878 | /* Propagate the data dependency var. */ | |
1879 | gimple *ass = gimple_build_assign (gimple_call_lhs (call), | |
1880 | gimple_call_arg (call, 1)); | |
1881 | gsi_replace (&gsi, ass, false); | |
1882 | } | |
1883 | else | |
1884 | gsi_remove (&gsi, true); | |
1885 | } | |
1886 | else if (!rescan) | |
1887 | /* If not rescanning, advance over the call. */ | |
1888 | gsi_next (&gsi); | |
1889 | } | |
1890 | ||
1891 | free_oacc_loop (loops); | |
1892 | ||
1893 | return 0; | |
1894 | } | |
1895 | ||
1896 | /* Default launch dimension validator. Force everything to 1. A | |
1897 | backend that wants to provide larger dimensions must override this | |
1898 | hook. */ | |
1899 | ||
1900 | bool | |
1901 | default_goacc_validate_dims (tree ARG_UNUSED (decl), int *dims, | |
46dedae6 TV |
1902 | int ARG_UNUSED (fn_level), |
1903 | unsigned ARG_UNUSED (used)) | |
629b3d75 MJ |
1904 | { |
1905 | bool changed = false; | |
1906 | ||
1907 | for (unsigned ix = 0; ix != GOMP_DIM_MAX; ix++) | |
1908 | { | |
1909 | if (dims[ix] != 1) | |
1910 | { | |
1911 | dims[ix] = 1; | |
1912 | changed = true; | |
1913 | } | |
1914 | } | |
1915 | ||
1916 | return changed; | |
1917 | } | |
1918 | ||
01914336 | 1919 | /* Default dimension bound is unknown on accelerator and 1 on host. */ |
629b3d75 MJ |
1920 | |
1921 | int | |
1922 | default_goacc_dim_limit (int ARG_UNUSED (axis)) | |
1923 | { | |
1924 | #ifdef ACCEL_COMPILER | |
1925 | return 0; | |
1926 | #else | |
1927 | return 1; | |
1928 | #endif | |
1929 | } | |
1930 | ||
1931 | namespace { | |
1932 | ||
1933 | const pass_data pass_data_oacc_device_lower = | |
1934 | { | |
1935 | GIMPLE_PASS, /* type */ | |
1936 | "oaccdevlow", /* name */ | |
fd2b8c8b | 1937 | OPTGROUP_OMP, /* optinfo_flags */ |
629b3d75 MJ |
1938 | TV_NONE, /* tv_id */ |
1939 | PROP_cfg, /* properties_required */ | |
1940 | 0 /* Possibly PROP_gimple_eomp. */, /* properties_provided */ | |
1941 | 0, /* properties_destroyed */ | |
1942 | 0, /* todo_flags_start */ | |
1943 | TODO_update_ssa | TODO_cleanup_cfg, /* todo_flags_finish */ | |
1944 | }; | |
1945 | ||
1946 | class pass_oacc_device_lower : public gimple_opt_pass | |
1947 | { | |
1948 | public: | |
1949 | pass_oacc_device_lower (gcc::context *ctxt) | |
1950 | : gimple_opt_pass (pass_data_oacc_device_lower, ctxt) | |
1951 | {} | |
1952 | ||
1953 | /* opt_pass methods: */ | |
1954 | virtual bool gate (function *) { return flag_openacc; }; | |
1955 | ||
1956 | virtual unsigned int execute (function *) | |
1957 | { | |
1958 | return execute_oacc_device_lower (); | |
1959 | } | |
1960 | ||
1961 | }; // class pass_oacc_device_lower | |
1962 | ||
1963 | } // anon namespace | |
1964 | ||
1965 | gimple_opt_pass * | |
1966 | make_pass_oacc_device_lower (gcc::context *ctxt) | |
1967 | { | |
1968 | return new pass_oacc_device_lower (ctxt); | |
1969 | } | |
1970 | ||
0c6b03b5 AM |
1971 | \f |
1972 | /* Rewrite GOMP_SIMT_ENTER_ALLOC call given by GSI and remove the preceding | |
1973 | GOMP_SIMT_ENTER call identifying the privatized variables, which are | |
1974 | turned to structure fields and receive a DECL_VALUE_EXPR accordingly. | |
1975 | Set *REGIMPLIFY to true, except if no privatized variables were seen. */ | |
1976 | ||
1977 | static void | |
1978 | ompdevlow_adjust_simt_enter (gimple_stmt_iterator *gsi, bool *regimplify) | |
1979 | { | |
1980 | gimple *alloc_stmt = gsi_stmt (*gsi); | |
1981 | tree simtrec = gimple_call_lhs (alloc_stmt); | |
1982 | tree simduid = gimple_call_arg (alloc_stmt, 0); | |
1983 | gimple *enter_stmt = SSA_NAME_DEF_STMT (simduid); | |
1984 | gcc_assert (gimple_call_internal_p (enter_stmt, IFN_GOMP_SIMT_ENTER)); | |
1985 | tree rectype = lang_hooks.types.make_type (RECORD_TYPE); | |
1986 | TYPE_ARTIFICIAL (rectype) = TYPE_NAMELESS (rectype) = 1; | |
1987 | TREE_ADDRESSABLE (rectype) = 1; | |
1988 | TREE_TYPE (simtrec) = build_pointer_type (rectype); | |
1989 | for (unsigned i = 1; i < gimple_call_num_args (enter_stmt); i++) | |
1990 | { | |
1991 | tree *argp = gimple_call_arg_ptr (enter_stmt, i); | |
1992 | if (*argp == null_pointer_node) | |
1993 | continue; | |
1994 | gcc_assert (TREE_CODE (*argp) == ADDR_EXPR | |
1995 | && VAR_P (TREE_OPERAND (*argp, 0))); | |
1996 | tree var = TREE_OPERAND (*argp, 0); | |
1997 | ||
1998 | tree field = build_decl (DECL_SOURCE_LOCATION (var), FIELD_DECL, | |
1999 | DECL_NAME (var), TREE_TYPE (var)); | |
2000 | SET_DECL_ALIGN (field, DECL_ALIGN (var)); | |
2001 | DECL_USER_ALIGN (field) = DECL_USER_ALIGN (var); | |
2002 | TREE_THIS_VOLATILE (field) = TREE_THIS_VOLATILE (var); | |
2003 | ||
2004 | insert_field_into_struct (rectype, field); | |
2005 | ||
2006 | tree t = build_simple_mem_ref (simtrec); | |
2007 | t = build3 (COMPONENT_REF, TREE_TYPE (var), t, field, NULL); | |
2008 | TREE_THIS_VOLATILE (t) = TREE_THIS_VOLATILE (var); | |
2009 | SET_DECL_VALUE_EXPR (var, t); | |
2010 | DECL_HAS_VALUE_EXPR_P (var) = 1; | |
2011 | *regimplify = true; | |
2012 | } | |
2013 | layout_type (rectype); | |
2014 | tree size = TYPE_SIZE_UNIT (rectype); | |
2015 | tree align = build_int_cst (TREE_TYPE (size), TYPE_ALIGN_UNIT (rectype)); | |
2016 | ||
2017 | alloc_stmt | |
2018 | = gimple_build_call_internal (IFN_GOMP_SIMT_ENTER_ALLOC, 2, size, align); | |
2019 | gimple_call_set_lhs (alloc_stmt, simtrec); | |
2020 | gsi_replace (gsi, alloc_stmt, false); | |
2021 | gimple_stmt_iterator enter_gsi = gsi_for_stmt (enter_stmt); | |
2022 | enter_stmt = gimple_build_assign (simduid, gimple_call_arg (enter_stmt, 0)); | |
2023 | gsi_replace (&enter_gsi, enter_stmt, false); | |
2024 | ||
2025 | use_operand_p use; | |
2026 | gimple *exit_stmt; | |
2027 | if (single_imm_use (simtrec, &use, &exit_stmt)) | |
2028 | { | |
2029 | gcc_assert (gimple_call_internal_p (exit_stmt, IFN_GOMP_SIMT_EXIT)); | |
2030 | gimple_stmt_iterator exit_gsi = gsi_for_stmt (exit_stmt); | |
25b45c7c | 2031 | tree clobber = build_clobber (rectype); |
0c6b03b5 AM |
2032 | exit_stmt = gimple_build_assign (build_simple_mem_ref (simtrec), clobber); |
2033 | gsi_insert_before (&exit_gsi, exit_stmt, GSI_SAME_STMT); | |
2034 | } | |
2035 | else | |
2036 | gcc_checking_assert (has_zero_uses (simtrec)); | |
2037 | } | |
2038 | ||
2039 | /* Callback for walk_gimple_stmt used to scan for SIMT-privatized variables. */ | |
2040 | ||
2041 | static tree | |
2042 | find_simtpriv_var_op (tree *tp, int *walk_subtrees, void *) | |
2043 | { | |
2044 | tree t = *tp; | |
2045 | ||
2046 | if (VAR_P (t) | |
2047 | && DECL_HAS_VALUE_EXPR_P (t) | |
2048 | && lookup_attribute ("omp simt private", DECL_ATTRIBUTES (t))) | |
2049 | { | |
2050 | *walk_subtrees = 0; | |
2051 | return t; | |
2052 | } | |
2053 | return NULL_TREE; | |
2054 | } | |
2055 | ||
629b3d75 MJ |
2056 | /* Cleanup uses of SIMT placeholder internal functions: on non-SIMT targets, |
2057 | VF is 1 and LANE is 0; on SIMT targets, VF is folded to a constant, and | |
2058 | LANE is kept to be expanded to RTL later on. Also cleanup all other SIMT | |
2059 | internal functions on non-SIMT targets, and likewise some SIMD internal | |
2060 | functions on SIMT targets. */ | |
2061 | ||
2062 | static unsigned int | |
2063 | execute_omp_device_lower () | |
2064 | { | |
2065 | int vf = targetm.simt.vf ? targetm.simt.vf () : 1; | |
0c6b03b5 | 2066 | bool regimplify = false; |
629b3d75 MJ |
2067 | basic_block bb; |
2068 | gimple_stmt_iterator gsi; | |
7a50e708 JJ |
2069 | bool calls_declare_variant_alt |
2070 | = cgraph_node::get (cfun->decl)->calls_declare_variant_alt; | |
629b3d75 MJ |
2071 | FOR_EACH_BB_FN (bb, cfun) |
2072 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2073 | { | |
2074 | gimple *stmt = gsi_stmt (gsi); | |
7a50e708 | 2075 | if (!is_gimple_call (stmt)) |
629b3d75 | 2076 | continue; |
7a50e708 JJ |
2077 | if (!gimple_call_internal_p (stmt)) |
2078 | { | |
2079 | if (calls_declare_variant_alt) | |
2080 | if (tree fndecl = gimple_call_fndecl (stmt)) | |
2081 | { | |
2082 | tree new_fndecl = omp_resolve_declare_variant (fndecl); | |
2083 | if (new_fndecl != fndecl) | |
2084 | { | |
2085 | gimple_call_set_fndecl (stmt, new_fndecl); | |
2086 | update_stmt (stmt); | |
2087 | } | |
2088 | } | |
2089 | continue; | |
2090 | } | |
629b3d75 MJ |
2091 | tree lhs = gimple_call_lhs (stmt), rhs = NULL_TREE; |
2092 | tree type = lhs ? TREE_TYPE (lhs) : integer_type_node; | |
2093 | switch (gimple_call_internal_fn (stmt)) | |
2094 | { | |
2095 | case IFN_GOMP_USE_SIMT: | |
2096 | rhs = vf == 1 ? integer_zero_node : integer_one_node; | |
2097 | break; | |
0c6b03b5 AM |
2098 | case IFN_GOMP_SIMT_ENTER: |
2099 | rhs = vf == 1 ? gimple_call_arg (stmt, 0) : NULL_TREE; | |
2100 | goto simtreg_enter_exit; | |
2101 | case IFN_GOMP_SIMT_ENTER_ALLOC: | |
2102 | if (vf != 1) | |
2103 | ompdevlow_adjust_simt_enter (&gsi, ®implify); | |
2104 | rhs = vf == 1 ? null_pointer_node : NULL_TREE; | |
2105 | goto simtreg_enter_exit; | |
2106 | case IFN_GOMP_SIMT_EXIT: | |
2107 | simtreg_enter_exit: | |
2108 | if (vf != 1) | |
2109 | continue; | |
2110 | unlink_stmt_vdef (stmt); | |
2111 | break; | |
629b3d75 MJ |
2112 | case IFN_GOMP_SIMT_LANE: |
2113 | case IFN_GOMP_SIMT_LAST_LANE: | |
2114 | rhs = vf == 1 ? build_zero_cst (type) : NULL_TREE; | |
2115 | break; | |
2116 | case IFN_GOMP_SIMT_VF: | |
2117 | rhs = build_int_cst (type, vf); | |
2118 | break; | |
2119 | case IFN_GOMP_SIMT_ORDERED_PRED: | |
2120 | rhs = vf == 1 ? integer_zero_node : NULL_TREE; | |
2121 | if (rhs || !lhs) | |
2122 | unlink_stmt_vdef (stmt); | |
2123 | break; | |
2124 | case IFN_GOMP_SIMT_VOTE_ANY: | |
2125 | case IFN_GOMP_SIMT_XCHG_BFLY: | |
2126 | case IFN_GOMP_SIMT_XCHG_IDX: | |
2127 | rhs = vf == 1 ? gimple_call_arg (stmt, 0) : NULL_TREE; | |
2128 | break; | |
2129 | case IFN_GOMP_SIMD_LANE: | |
2130 | case IFN_GOMP_SIMD_LAST_LANE: | |
2131 | rhs = vf != 1 ? build_zero_cst (type) : NULL_TREE; | |
2132 | break; | |
2133 | case IFN_GOMP_SIMD_VF: | |
2134 | rhs = vf != 1 ? build_one_cst (type) : NULL_TREE; | |
2135 | break; | |
2136 | default: | |
2137 | continue; | |
2138 | } | |
2139 | if (lhs && !rhs) | |
2140 | continue; | |
2141 | stmt = lhs ? gimple_build_assign (lhs, rhs) : gimple_build_nop (); | |
2142 | gsi_replace (&gsi, stmt, false); | |
2143 | } | |
0c6b03b5 AM |
2144 | if (regimplify) |
2145 | FOR_EACH_BB_REVERSE_FN (bb, cfun) | |
2146 | for (gsi = gsi_last_bb (bb); !gsi_end_p (gsi); gsi_prev (&gsi)) | |
2147 | if (walk_gimple_stmt (&gsi, NULL, find_simtpriv_var_op, NULL)) | |
2148 | { | |
2149 | if (gimple_clobber_p (gsi_stmt (gsi))) | |
2150 | gsi_remove (&gsi, true); | |
2151 | else | |
2152 | gimple_regimplify_operands (gsi_stmt (gsi), &gsi); | |
2153 | } | |
629b3d75 MJ |
2154 | if (vf != 1) |
2155 | cfun->has_force_vectorize_loops = false; | |
2156 | return 0; | |
2157 | } | |
2158 | ||
2159 | namespace { | |
2160 | ||
2161 | const pass_data pass_data_omp_device_lower = | |
2162 | { | |
2163 | GIMPLE_PASS, /* type */ | |
2164 | "ompdevlow", /* name */ | |
fd2b8c8b | 2165 | OPTGROUP_OMP, /* optinfo_flags */ |
629b3d75 MJ |
2166 | TV_NONE, /* tv_id */ |
2167 | PROP_cfg, /* properties_required */ | |
2168 | PROP_gimple_lomp_dev, /* properties_provided */ | |
2169 | 0, /* properties_destroyed */ | |
2170 | 0, /* todo_flags_start */ | |
2171 | TODO_update_ssa, /* todo_flags_finish */ | |
2172 | }; | |
2173 | ||
2174 | class pass_omp_device_lower : public gimple_opt_pass | |
2175 | { | |
2176 | public: | |
2177 | pass_omp_device_lower (gcc::context *ctxt) | |
2178 | : gimple_opt_pass (pass_data_omp_device_lower, ctxt) | |
2179 | {} | |
2180 | ||
2181 | /* opt_pass methods: */ | |
4cea8675 | 2182 | virtual bool gate (function *fun) |
629b3d75 | 2183 | { |
7a50e708 JJ |
2184 | return (!(fun->curr_properties & PROP_gimple_lomp_dev) |
2185 | || (flag_openmp | |
2186 | && cgraph_node::get (fun->decl)->calls_declare_variant_alt)); | |
629b3d75 MJ |
2187 | } |
2188 | virtual unsigned int execute (function *) | |
2189 | { | |
2190 | return execute_omp_device_lower (); | |
2191 | } | |
2192 | ||
2193 | }; // class pass_expand_omp_ssa | |
2194 | ||
2195 | } // anon namespace | |
2196 | ||
2197 | gimple_opt_pass * | |
2198 | make_pass_omp_device_lower (gcc::context *ctxt) | |
2199 | { | |
2200 | return new pass_omp_device_lower (ctxt); | |
2201 | } | |
2202 | ||
2203 | /* "omp declare target link" handling pass. */ | |
2204 | ||
2205 | namespace { | |
2206 | ||
2207 | const pass_data pass_data_omp_target_link = | |
2208 | { | |
2209 | GIMPLE_PASS, /* type */ | |
2210 | "omptargetlink", /* name */ | |
fd2b8c8b | 2211 | OPTGROUP_OMP, /* optinfo_flags */ |
629b3d75 MJ |
2212 | TV_NONE, /* tv_id */ |
2213 | PROP_ssa, /* properties_required */ | |
2214 | 0, /* properties_provided */ | |
2215 | 0, /* properties_destroyed */ | |
2216 | 0, /* todo_flags_start */ | |
2217 | TODO_update_ssa, /* todo_flags_finish */ | |
2218 | }; | |
2219 | ||
2220 | class pass_omp_target_link : public gimple_opt_pass | |
2221 | { | |
2222 | public: | |
2223 | pass_omp_target_link (gcc::context *ctxt) | |
2224 | : gimple_opt_pass (pass_data_omp_target_link, ctxt) | |
2225 | {} | |
2226 | ||
2227 | /* opt_pass methods: */ | |
2228 | virtual bool gate (function *fun) | |
2229 | { | |
2230 | #ifdef ACCEL_COMPILER | |
46dbeb40 | 2231 | return offloading_function_p (fun->decl); |
629b3d75 MJ |
2232 | #else |
2233 | (void) fun; | |
2234 | return false; | |
2235 | #endif | |
2236 | } | |
2237 | ||
2238 | virtual unsigned execute (function *); | |
2239 | }; | |
2240 | ||
2241 | /* Callback for walk_gimple_stmt used to scan for link var operands. */ | |
2242 | ||
2243 | static tree | |
2244 | find_link_var_op (tree *tp, int *walk_subtrees, void *) | |
2245 | { | |
2246 | tree t = *tp; | |
2247 | ||
56f71478 JJ |
2248 | if (VAR_P (t) |
2249 | && DECL_HAS_VALUE_EXPR_P (t) | |
2250 | && is_global_var (t) | |
629b3d75 MJ |
2251 | && lookup_attribute ("omp declare target link", DECL_ATTRIBUTES (t))) |
2252 | { | |
2253 | *walk_subtrees = 0; | |
2254 | return t; | |
2255 | } | |
2256 | ||
2257 | return NULL_TREE; | |
2258 | } | |
2259 | ||
2260 | unsigned | |
2261 | pass_omp_target_link::execute (function *fun) | |
2262 | { | |
2263 | basic_block bb; | |
2264 | FOR_EACH_BB_FN (bb, fun) | |
2265 | { | |
2266 | gimple_stmt_iterator gsi; | |
2267 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
2268 | if (walk_gimple_stmt (&gsi, NULL, find_link_var_op, NULL)) | |
2269 | gimple_regimplify_operands (gsi_stmt (gsi), &gsi); | |
2270 | } | |
2271 | ||
2272 | return 0; | |
2273 | } | |
2274 | ||
2275 | } // anon namespace | |
2276 | ||
2277 | gimple_opt_pass * | |
2278 | make_pass_omp_target_link (gcc::context *ctxt) | |
2279 | { | |
2280 | return new pass_omp_target_link (ctxt); | |
2281 | } |