]>
Commit | Line | Data |
---|---|---|
1 | /* Target code for NVPTX. | |
2 | Copyright (C) 2014-2020 Free Software Foundation, Inc. | |
3 | Contributed by Bernd Schmidt <bernds@codesourcery.com> | |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it | |
8 | under the terms of the GNU General Public License as published | |
9 | by the Free Software Foundation; either version 3, or (at your | |
10 | option) any later version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT | |
13 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY | |
14 | or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public | |
15 | License for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | #define IN_TARGET_CODE 1 | |
22 | ||
23 | #include "config.h" | |
24 | #include <sstream> | |
25 | #include "system.h" | |
26 | #include "coretypes.h" | |
27 | #include "backend.h" | |
28 | #include "target.h" | |
29 | #include "rtl.h" | |
30 | #include "tree.h" | |
31 | #include "cfghooks.h" | |
32 | #include "df.h" | |
33 | #include "memmodel.h" | |
34 | #include "tm_p.h" | |
35 | #include "expmed.h" | |
36 | #include "optabs.h" | |
37 | #include "regs.h" | |
38 | #include "emit-rtl.h" | |
39 | #include "recog.h" | |
40 | #include "diagnostic.h" | |
41 | #include "alias.h" | |
42 | #include "insn-flags.h" | |
43 | #include "output.h" | |
44 | #include "insn-attr.h" | |
45 | #include "flags.h" | |
46 | #include "dojump.h" | |
47 | #include "explow.h" | |
48 | #include "calls.h" | |
49 | #include "varasm.h" | |
50 | #include "stmt.h" | |
51 | #include "expr.h" | |
52 | #include "tm-preds.h" | |
53 | #include "tm-constrs.h" | |
54 | #include "langhooks.h" | |
55 | #include "dbxout.h" | |
56 | #include "cfgrtl.h" | |
57 | #include "gimple.h" | |
58 | #include "stor-layout.h" | |
59 | #include "builtins.h" | |
60 | #include "omp-general.h" | |
61 | #include "omp-low.h" | |
62 | #include "omp-offload.h" | |
63 | #include "gomp-constants.h" | |
64 | #include "dumpfile.h" | |
65 | #include "internal-fn.h" | |
66 | #include "gimple-iterator.h" | |
67 | #include "stringpool.h" | |
68 | #include "attribs.h" | |
69 | #include "tree-vrp.h" | |
70 | #include "tree-ssa-operands.h" | |
71 | #include "tree-ssanames.h" | |
72 | #include "gimplify.h" | |
73 | #include "tree-phinodes.h" | |
74 | #include "cfgloop.h" | |
75 | #include "fold-const.h" | |
76 | #include "intl.h" | |
77 | ||
78 | /* This file should be included last. */ | |
79 | #include "target-def.h" | |
80 | ||
81 | #define WORKAROUND_PTXJIT_BUG 1 | |
82 | #define WORKAROUND_PTXJIT_BUG_2 1 | |
83 | #define WORKAROUND_PTXJIT_BUG_3 1 | |
84 | ||
85 | /* The PTX concept CTA (Concurrent Thread Array) maps on the CUDA concept thread | |
86 | block, which has had a maximum number of threads of 1024 since CUDA version | |
87 | 2.x. */ | |
88 | #define PTX_CTA_SIZE 1024 | |
89 | ||
90 | #define PTX_CTA_NUM_BARRIERS 16 | |
91 | #define PTX_WARP_SIZE 32 | |
92 | ||
93 | #define PTX_PER_CTA_BARRIER 0 | |
94 | #define PTX_NUM_PER_CTA_BARRIERS 1 | |
95 | #define PTX_FIRST_PER_WORKER_BARRIER (PTX_NUM_PER_CTA_BARRIERS) | |
96 | #define PTX_NUM_PER_WORKER_BARRIERS (PTX_CTA_NUM_BARRIERS - PTX_NUM_PER_CTA_BARRIERS) | |
97 | ||
98 | #define PTX_DEFAULT_VECTOR_LENGTH PTX_WARP_SIZE | |
99 | #define PTX_MAX_VECTOR_LENGTH PTX_CTA_SIZE | |
100 | #define PTX_WORKER_LENGTH 32 | |
101 | #define PTX_DEFAULT_RUNTIME_DIM 0 /* Defer to runtime. */ | |
102 | ||
103 | /* The various PTX memory areas an object might reside in. */ | |
104 | enum nvptx_data_area | |
105 | { | |
106 | DATA_AREA_GENERIC, | |
107 | DATA_AREA_GLOBAL, | |
108 | DATA_AREA_SHARED, | |
109 | DATA_AREA_LOCAL, | |
110 | DATA_AREA_CONST, | |
111 | DATA_AREA_PARAM, | |
112 | DATA_AREA_MAX | |
113 | }; | |
114 | ||
115 | /* We record the data area in the target symbol flags. */ | |
116 | #define SYMBOL_DATA_AREA(SYM) \ | |
117 | (nvptx_data_area)((SYMBOL_REF_FLAGS (SYM) >> SYMBOL_FLAG_MACH_DEP_SHIFT) \ | |
118 | & 7) | |
119 | #define SET_SYMBOL_DATA_AREA(SYM,AREA) \ | |
120 | (SYMBOL_REF_FLAGS (SYM) |= (AREA) << SYMBOL_FLAG_MACH_DEP_SHIFT) | |
121 | ||
122 | /* Record the function decls we've written, and the libfuncs and function | |
123 | decls corresponding to them. */ | |
124 | static std::stringstream func_decls; | |
125 | ||
126 | struct declared_libfunc_hasher : ggc_cache_ptr_hash<rtx_def> | |
127 | { | |
128 | static hashval_t hash (rtx x) { return htab_hash_pointer (x); } | |
129 | static bool equal (rtx a, rtx b) { return a == b; } | |
130 | }; | |
131 | ||
132 | static GTY((cache)) | |
133 | hash_table<declared_libfunc_hasher> *declared_libfuncs_htab; | |
134 | ||
135 | struct tree_hasher : ggc_cache_ptr_hash<tree_node> | |
136 | { | |
137 | static hashval_t hash (tree t) { return htab_hash_pointer (t); } | |
138 | static bool equal (tree a, tree b) { return a == b; } | |
139 | }; | |
140 | ||
141 | static GTY((cache)) hash_table<tree_hasher> *declared_fndecls_htab; | |
142 | static GTY((cache)) hash_table<tree_hasher> *needed_fndecls_htab; | |
143 | ||
144 | /* Buffer needed to broadcast across workers and vectors. This is | |
145 | used for both worker-neutering and worker broadcasting, and | |
146 | vector-neutering and boardcasting when vector_length > 32. It is | |
147 | shared by all functions emitted. The buffer is placed in shared | |
148 | memory. It'd be nice if PTX supported common blocks, because then | |
149 | this could be shared across TUs (taking the largest size). */ | |
150 | static unsigned oacc_bcast_size; | |
151 | static unsigned oacc_bcast_partition; | |
152 | static unsigned oacc_bcast_align; | |
153 | static GTY(()) rtx oacc_bcast_sym; | |
154 | ||
155 | /* Buffer needed for worker reductions. This has to be distinct from | |
156 | the worker broadcast array, as both may be live concurrently. */ | |
157 | static unsigned worker_red_size; | |
158 | static unsigned worker_red_align; | |
159 | static GTY(()) rtx worker_red_sym; | |
160 | ||
161 | /* Buffer needed for vector reductions, when vector_length > | |
162 | PTX_WARP_SIZE. This has to be distinct from the worker broadcast | |
163 | array, as both may be live concurrently. */ | |
164 | static unsigned vector_red_size; | |
165 | static unsigned vector_red_align; | |
166 | static unsigned vector_red_partition; | |
167 | static GTY(()) rtx vector_red_sym; | |
168 | ||
169 | /* Global lock variable, needed for 128bit worker & gang reductions. */ | |
170 | static GTY(()) tree global_lock_var; | |
171 | ||
172 | /* True if any function references __nvptx_stacks. */ | |
173 | static bool need_softstack_decl; | |
174 | ||
175 | /* True if any function references __nvptx_uni. */ | |
176 | static bool need_unisimt_decl; | |
177 | ||
178 | static int nvptx_mach_max_workers (); | |
179 | ||
180 | /* Allocate a new, cleared machine_function structure. */ | |
181 | ||
182 | static struct machine_function * | |
183 | nvptx_init_machine_status (void) | |
184 | { | |
185 | struct machine_function *p = ggc_cleared_alloc<machine_function> (); | |
186 | p->return_mode = VOIDmode; | |
187 | return p; | |
188 | } | |
189 | ||
190 | /* Issue a diagnostic when option OPTNAME is enabled (as indicated by OPTVAL) | |
191 | and -fopenacc is also enabled. */ | |
192 | ||
193 | static void | |
194 | diagnose_openacc_conflict (bool optval, const char *optname) | |
195 | { | |
196 | if (flag_openacc && optval) | |
197 | error ("option %s is not supported together with %<-fopenacc%>", optname); | |
198 | } | |
199 | ||
200 | /* Implement TARGET_OPTION_OVERRIDE. */ | |
201 | ||
202 | static void | |
203 | nvptx_option_override (void) | |
204 | { | |
205 | init_machine_status = nvptx_init_machine_status; | |
206 | ||
207 | /* Set toplevel_reorder, unless explicitly disabled. We need | |
208 | reordering so that we emit necessary assembler decls of | |
209 | undeclared variables. */ | |
210 | if (!global_options_set.x_flag_toplevel_reorder) | |
211 | flag_toplevel_reorder = 1; | |
212 | ||
213 | debug_nonbind_markers_p = 0; | |
214 | ||
215 | /* Set flag_no_common, unless explicitly disabled. We fake common | |
216 | using .weak, and that's not entirely accurate, so avoid it | |
217 | unless forced. */ | |
218 | if (!global_options_set.x_flag_no_common) | |
219 | flag_no_common = 1; | |
220 | ||
221 | /* The patch area requires nops, which we don't have. */ | |
222 | if (function_entry_patch_area_size > 0) | |
223 | sorry ("not generating patch area, nops not supported"); | |
224 | ||
225 | /* Assumes that it will see only hard registers. */ | |
226 | flag_var_tracking = 0; | |
227 | ||
228 | if (nvptx_optimize < 0) | |
229 | nvptx_optimize = optimize > 0; | |
230 | ||
231 | declared_fndecls_htab = hash_table<tree_hasher>::create_ggc (17); | |
232 | needed_fndecls_htab = hash_table<tree_hasher>::create_ggc (17); | |
233 | declared_libfuncs_htab | |
234 | = hash_table<declared_libfunc_hasher>::create_ggc (17); | |
235 | ||
236 | oacc_bcast_sym = gen_rtx_SYMBOL_REF (Pmode, "__oacc_bcast"); | |
237 | SET_SYMBOL_DATA_AREA (oacc_bcast_sym, DATA_AREA_SHARED); | |
238 | oacc_bcast_align = GET_MODE_ALIGNMENT (SImode) / BITS_PER_UNIT; | |
239 | oacc_bcast_partition = 0; | |
240 | ||
241 | worker_red_sym = gen_rtx_SYMBOL_REF (Pmode, "__worker_red"); | |
242 | SET_SYMBOL_DATA_AREA (worker_red_sym, DATA_AREA_SHARED); | |
243 | worker_red_align = GET_MODE_ALIGNMENT (SImode) / BITS_PER_UNIT; | |
244 | ||
245 | vector_red_sym = gen_rtx_SYMBOL_REF (Pmode, "__vector_red"); | |
246 | SET_SYMBOL_DATA_AREA (vector_red_sym, DATA_AREA_SHARED); | |
247 | vector_red_align = GET_MODE_ALIGNMENT (SImode) / BITS_PER_UNIT; | |
248 | vector_red_partition = 0; | |
249 | ||
250 | diagnose_openacc_conflict (TARGET_GOMP, "-mgomp"); | |
251 | diagnose_openacc_conflict (TARGET_SOFT_STACK, "-msoft-stack"); | |
252 | diagnose_openacc_conflict (TARGET_UNIFORM_SIMT, "-muniform-simt"); | |
253 | ||
254 | if (TARGET_GOMP) | |
255 | target_flags |= MASK_SOFT_STACK | MASK_UNIFORM_SIMT; | |
256 | } | |
257 | ||
258 | /* Return a ptx type for MODE. If PROMOTE, then use .u32 for QImode to | |
259 | deal with ptx ideosyncracies. */ | |
260 | ||
261 | const char * | |
262 | nvptx_ptx_type_from_mode (machine_mode mode, bool promote) | |
263 | { | |
264 | switch (mode) | |
265 | { | |
266 | case E_BLKmode: | |
267 | return ".b8"; | |
268 | case E_BImode: | |
269 | return ".pred"; | |
270 | case E_QImode: | |
271 | if (promote) | |
272 | return ".u32"; | |
273 | else | |
274 | return ".u8"; | |
275 | case E_HImode: | |
276 | return ".u16"; | |
277 | case E_SImode: | |
278 | return ".u32"; | |
279 | case E_DImode: | |
280 | return ".u64"; | |
281 | ||
282 | case E_SFmode: | |
283 | return ".f32"; | |
284 | case E_DFmode: | |
285 | return ".f64"; | |
286 | ||
287 | case E_V2SImode: | |
288 | return ".v2.u32"; | |
289 | case E_V2DImode: | |
290 | return ".v2.u64"; | |
291 | ||
292 | default: | |
293 | gcc_unreachable (); | |
294 | } | |
295 | } | |
296 | ||
297 | /* Encode the PTX data area that DECL (which might not actually be a | |
298 | _DECL) should reside in. */ | |
299 | ||
300 | static void | |
301 | nvptx_encode_section_info (tree decl, rtx rtl, int first) | |
302 | { | |
303 | default_encode_section_info (decl, rtl, first); | |
304 | if (first && MEM_P (rtl)) | |
305 | { | |
306 | nvptx_data_area area = DATA_AREA_GENERIC; | |
307 | ||
308 | if (TREE_CONSTANT (decl)) | |
309 | area = DATA_AREA_CONST; | |
310 | else if (TREE_CODE (decl) == VAR_DECL) | |
311 | { | |
312 | if (lookup_attribute ("shared", DECL_ATTRIBUTES (decl))) | |
313 | { | |
314 | area = DATA_AREA_SHARED; | |
315 | if (DECL_INITIAL (decl)) | |
316 | error ("static initialization of variable %q+D in %<.shared%>" | |
317 | " memory is not supported", decl); | |
318 | } | |
319 | else | |
320 | area = TREE_READONLY (decl) ? DATA_AREA_CONST : DATA_AREA_GLOBAL; | |
321 | } | |
322 | ||
323 | SET_SYMBOL_DATA_AREA (XEXP (rtl, 0), area); | |
324 | } | |
325 | } | |
326 | ||
327 | /* Return the PTX name of the data area in which SYM should be | |
328 | placed. The symbol must have already been processed by | |
329 | nvptx_encode_seciton_info, or equivalent. */ | |
330 | ||
331 | static const char * | |
332 | section_for_sym (rtx sym) | |
333 | { | |
334 | nvptx_data_area area = SYMBOL_DATA_AREA (sym); | |
335 | /* Same order as nvptx_data_area enum. */ | |
336 | static char const *const areas[] = | |
337 | {"", ".global", ".shared", ".local", ".const", ".param"}; | |
338 | ||
339 | return areas[area]; | |
340 | } | |
341 | ||
342 | /* Similarly for a decl. */ | |
343 | ||
344 | static const char * | |
345 | section_for_decl (const_tree decl) | |
346 | { | |
347 | return section_for_sym (XEXP (DECL_RTL (CONST_CAST (tree, decl)), 0)); | |
348 | } | |
349 | ||
350 | /* Check NAME for special function names and redirect them by returning a | |
351 | replacement. This applies to malloc, free and realloc, for which we | |
352 | want to use libgcc wrappers, and call, which triggers a bug in | |
353 | ptxas. We can't use TARGET_MANGLE_DECL_ASSEMBLER_NAME, as that's | |
354 | not active in an offload compiler -- the names are all set by the | |
355 | host-side compiler. */ | |
356 | ||
357 | static const char * | |
358 | nvptx_name_replacement (const char *name) | |
359 | { | |
360 | if (strcmp (name, "call") == 0) | |
361 | return "__nvptx_call"; | |
362 | if (strcmp (name, "malloc") == 0) | |
363 | return "__nvptx_malloc"; | |
364 | if (strcmp (name, "free") == 0) | |
365 | return "__nvptx_free"; | |
366 | if (strcmp (name, "realloc") == 0) | |
367 | return "__nvptx_realloc"; | |
368 | return name; | |
369 | } | |
370 | ||
371 | /* If MODE should be treated as two registers of an inner mode, return | |
372 | that inner mode. Otherwise return VOIDmode. */ | |
373 | ||
374 | static machine_mode | |
375 | maybe_split_mode (machine_mode mode) | |
376 | { | |
377 | if (COMPLEX_MODE_P (mode)) | |
378 | return GET_MODE_INNER (mode); | |
379 | ||
380 | if (mode == TImode) | |
381 | return DImode; | |
382 | ||
383 | return VOIDmode; | |
384 | } | |
385 | ||
386 | /* Return true if mode should be treated as two registers. */ | |
387 | ||
388 | static bool | |
389 | split_mode_p (machine_mode mode) | |
390 | { | |
391 | return maybe_split_mode (mode) != VOIDmode; | |
392 | } | |
393 | ||
394 | /* Output a register, subreg, or register pair (with optional | |
395 | enclosing braces). */ | |
396 | ||
397 | static void | |
398 | output_reg (FILE *file, unsigned regno, machine_mode inner_mode, | |
399 | int subreg_offset = -1) | |
400 | { | |
401 | if (inner_mode == VOIDmode) | |
402 | { | |
403 | if (HARD_REGISTER_NUM_P (regno)) | |
404 | fprintf (file, "%s", reg_names[regno]); | |
405 | else | |
406 | fprintf (file, "%%r%d", regno); | |
407 | } | |
408 | else if (subreg_offset >= 0) | |
409 | { | |
410 | output_reg (file, regno, VOIDmode); | |
411 | fprintf (file, "$%d", subreg_offset); | |
412 | } | |
413 | else | |
414 | { | |
415 | if (subreg_offset == -1) | |
416 | fprintf (file, "{"); | |
417 | output_reg (file, regno, inner_mode, GET_MODE_SIZE (inner_mode)); | |
418 | fprintf (file, ","); | |
419 | output_reg (file, regno, inner_mode, 0); | |
420 | if (subreg_offset == -1) | |
421 | fprintf (file, "}"); | |
422 | } | |
423 | } | |
424 | ||
425 | /* Emit forking instructions for MASK. */ | |
426 | ||
427 | static void | |
428 | nvptx_emit_forking (unsigned mask, bool is_call) | |
429 | { | |
430 | mask &= (GOMP_DIM_MASK (GOMP_DIM_WORKER) | |
431 | | GOMP_DIM_MASK (GOMP_DIM_VECTOR)); | |
432 | if (mask) | |
433 | { | |
434 | rtx op = GEN_INT (mask | (is_call << GOMP_DIM_MAX)); | |
435 | ||
436 | /* Emit fork at all levels. This helps form SESE regions, as | |
437 | it creates a block with a single successor before entering a | |
438 | partitooned region. That is a good candidate for the end of | |
439 | an SESE region. */ | |
440 | emit_insn (gen_nvptx_fork (op)); | |
441 | emit_insn (gen_nvptx_forked (op)); | |
442 | } | |
443 | } | |
444 | ||
445 | /* Emit joining instructions for MASK. */ | |
446 | ||
447 | static void | |
448 | nvptx_emit_joining (unsigned mask, bool is_call) | |
449 | { | |
450 | mask &= (GOMP_DIM_MASK (GOMP_DIM_WORKER) | |
451 | | GOMP_DIM_MASK (GOMP_DIM_VECTOR)); | |
452 | if (mask) | |
453 | { | |
454 | rtx op = GEN_INT (mask | (is_call << GOMP_DIM_MAX)); | |
455 | ||
456 | /* Emit joining for all non-call pars to ensure there's a single | |
457 | predecessor for the block the join insn ends up in. This is | |
458 | needed for skipping entire loops. */ | |
459 | emit_insn (gen_nvptx_joining (op)); | |
460 | emit_insn (gen_nvptx_join (op)); | |
461 | } | |
462 | } | |
463 | ||
464 | \f | |
465 | /* Determine whether MODE and TYPE (possibly NULL) should be passed or | |
466 | returned in memory. Integer and floating types supported by the | |
467 | machine are passed in registers, everything else is passed in | |
468 | memory. Complex types are split. */ | |
469 | ||
470 | static bool | |
471 | pass_in_memory (machine_mode mode, const_tree type, bool for_return) | |
472 | { | |
473 | if (type) | |
474 | { | |
475 | if (AGGREGATE_TYPE_P (type)) | |
476 | return true; | |
477 | if (TREE_CODE (type) == VECTOR_TYPE) | |
478 | return true; | |
479 | } | |
480 | ||
481 | if (!for_return && COMPLEX_MODE_P (mode)) | |
482 | /* Complex types are passed as two underlying args. */ | |
483 | mode = GET_MODE_INNER (mode); | |
484 | ||
485 | if (GET_MODE_CLASS (mode) != MODE_INT | |
486 | && GET_MODE_CLASS (mode) != MODE_FLOAT) | |
487 | return true; | |
488 | ||
489 | if (GET_MODE_SIZE (mode) > UNITS_PER_WORD) | |
490 | return true; | |
491 | ||
492 | return false; | |
493 | } | |
494 | ||
495 | /* A non-memory argument of mode MODE is being passed, determine the mode it | |
496 | should be promoted to. This is also used for determining return | |
497 | type promotion. */ | |
498 | ||
499 | static machine_mode | |
500 | promote_arg (machine_mode mode, bool prototyped) | |
501 | { | |
502 | if (!prototyped && mode == SFmode) | |
503 | /* K&R float promotion for unprototyped functions. */ | |
504 | mode = DFmode; | |
505 | else if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (SImode)) | |
506 | mode = SImode; | |
507 | ||
508 | return mode; | |
509 | } | |
510 | ||
511 | /* A non-memory return type of MODE is being returned. Determine the | |
512 | mode it should be promoted to. */ | |
513 | ||
514 | static machine_mode | |
515 | promote_return (machine_mode mode) | |
516 | { | |
517 | return promote_arg (mode, true); | |
518 | } | |
519 | ||
520 | /* Implement TARGET_FUNCTION_ARG. */ | |
521 | ||
522 | static rtx | |
523 | nvptx_function_arg (cumulative_args_t, const function_arg_info &arg) | |
524 | { | |
525 | if (arg.end_marker_p () || !arg.named) | |
526 | return NULL_RTX; | |
527 | ||
528 | return gen_reg_rtx (arg.mode); | |
529 | } | |
530 | ||
531 | /* Implement TARGET_FUNCTION_INCOMING_ARG. */ | |
532 | ||
533 | static rtx | |
534 | nvptx_function_incoming_arg (cumulative_args_t cum_v, | |
535 | const function_arg_info &arg) | |
536 | { | |
537 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
538 | ||
539 | if (arg.end_marker_p () || !arg.named) | |
540 | return NULL_RTX; | |
541 | ||
542 | /* No need to deal with split modes here, the only case that can | |
543 | happen is complex modes and those are dealt with by | |
544 | TARGET_SPLIT_COMPLEX_ARG. */ | |
545 | return gen_rtx_UNSPEC (arg.mode, | |
546 | gen_rtvec (1, GEN_INT (cum->count)), | |
547 | UNSPEC_ARG_REG); | |
548 | } | |
549 | ||
550 | /* Implement TARGET_FUNCTION_ARG_ADVANCE. */ | |
551 | ||
552 | static void | |
553 | nvptx_function_arg_advance (cumulative_args_t cum_v, const function_arg_info &) | |
554 | { | |
555 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
556 | ||
557 | cum->count++; | |
558 | } | |
559 | ||
560 | /* Implement TARGET_FUNCTION_ARG_BOUNDARY. | |
561 | ||
562 | For nvptx This is only used for varadic args. The type has already | |
563 | been promoted and/or converted to invisible reference. */ | |
564 | ||
565 | static unsigned | |
566 | nvptx_function_arg_boundary (machine_mode mode, const_tree ARG_UNUSED (type)) | |
567 | { | |
568 | return GET_MODE_ALIGNMENT (mode); | |
569 | } | |
570 | ||
571 | /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. | |
572 | ||
573 | For nvptx, we know how to handle functions declared as stdarg: by | |
574 | passing an extra pointer to the unnamed arguments. However, the | |
575 | Fortran frontend can produce a different situation, where a | |
576 | function pointer is declared with no arguments, but the actual | |
577 | function and calls to it take more arguments. In that case, we | |
578 | want to ensure the call matches the definition of the function. */ | |
579 | ||
580 | static bool | |
581 | nvptx_strict_argument_naming (cumulative_args_t cum_v) | |
582 | { | |
583 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
584 | ||
585 | return cum->fntype == NULL_TREE || stdarg_p (cum->fntype); | |
586 | } | |
587 | ||
588 | /* Implement TARGET_LIBCALL_VALUE. */ | |
589 | ||
590 | static rtx | |
591 | nvptx_libcall_value (machine_mode mode, const_rtx) | |
592 | { | |
593 | if (!cfun || !cfun->machine->doing_call) | |
594 | /* Pretend to return in a hard reg for early uses before pseudos can be | |
595 | generated. */ | |
596 | return gen_rtx_REG (mode, NVPTX_RETURN_REGNUM); | |
597 | ||
598 | return gen_reg_rtx (mode); | |
599 | } | |
600 | ||
601 | /* TARGET_FUNCTION_VALUE implementation. Returns an RTX representing the place | |
602 | where function FUNC returns or receives a value of data type TYPE. */ | |
603 | ||
604 | static rtx | |
605 | nvptx_function_value (const_tree type, const_tree ARG_UNUSED (func), | |
606 | bool outgoing) | |
607 | { | |
608 | machine_mode mode = promote_return (TYPE_MODE (type)); | |
609 | ||
610 | if (outgoing) | |
611 | { | |
612 | gcc_assert (cfun); | |
613 | cfun->machine->return_mode = mode; | |
614 | return gen_rtx_REG (mode, NVPTX_RETURN_REGNUM); | |
615 | } | |
616 | ||
617 | return nvptx_libcall_value (mode, NULL_RTX); | |
618 | } | |
619 | ||
620 | /* Implement TARGET_FUNCTION_VALUE_REGNO_P. */ | |
621 | ||
622 | static bool | |
623 | nvptx_function_value_regno_p (const unsigned int regno) | |
624 | { | |
625 | return regno == NVPTX_RETURN_REGNUM; | |
626 | } | |
627 | ||
628 | /* Types with a mode other than those supported by the machine are passed by | |
629 | reference in memory. */ | |
630 | ||
631 | static bool | |
632 | nvptx_pass_by_reference (cumulative_args_t, const function_arg_info &arg) | |
633 | { | |
634 | return pass_in_memory (arg.mode, arg.type, false); | |
635 | } | |
636 | ||
637 | /* Implement TARGET_RETURN_IN_MEMORY. */ | |
638 | ||
639 | static bool | |
640 | nvptx_return_in_memory (const_tree type, const_tree) | |
641 | { | |
642 | return pass_in_memory (TYPE_MODE (type), type, true); | |
643 | } | |
644 | ||
645 | /* Implement TARGET_PROMOTE_FUNCTION_MODE. */ | |
646 | ||
647 | static machine_mode | |
648 | nvptx_promote_function_mode (const_tree type, machine_mode mode, | |
649 | int *ARG_UNUSED (punsignedp), | |
650 | const_tree funtype, int for_return) | |
651 | { | |
652 | return promote_arg (mode, for_return || !type || TYPE_ARG_TYPES (funtype)); | |
653 | } | |
654 | ||
655 | /* Helper for write_arg. Emit a single PTX argument of MODE, either | |
656 | in a prototype, or as copy in a function prologue. ARGNO is the | |
657 | index of this argument in the PTX function. FOR_REG is negative, | |
658 | if we're emitting the PTX prototype. It is zero if we're copying | |
659 | to an argument register and it is greater than zero if we're | |
660 | copying to a specific hard register. */ | |
661 | ||
662 | static int | |
663 | write_arg_mode (std::stringstream &s, int for_reg, int argno, | |
664 | machine_mode mode) | |
665 | { | |
666 | const char *ptx_type = nvptx_ptx_type_from_mode (mode, false); | |
667 | ||
668 | if (for_reg < 0) | |
669 | { | |
670 | /* Writing PTX prototype. */ | |
671 | s << (argno ? ", " : " ("); | |
672 | s << ".param" << ptx_type << " %in_ar" << argno; | |
673 | } | |
674 | else | |
675 | { | |
676 | s << "\t.reg" << ptx_type << " "; | |
677 | if (for_reg) | |
678 | s << reg_names[for_reg]; | |
679 | else | |
680 | s << "%ar" << argno; | |
681 | s << ";\n"; | |
682 | if (argno >= 0) | |
683 | { | |
684 | s << "\tld.param" << ptx_type << " "; | |
685 | if (for_reg) | |
686 | s << reg_names[for_reg]; | |
687 | else | |
688 | s << "%ar" << argno; | |
689 | s << ", [%in_ar" << argno << "];\n"; | |
690 | } | |
691 | } | |
692 | return argno + 1; | |
693 | } | |
694 | ||
695 | /* Process function parameter TYPE to emit one or more PTX | |
696 | arguments. S, FOR_REG and ARGNO as for write_arg_mode. PROTOTYPED | |
697 | is true, if this is a prototyped function, rather than an old-style | |
698 | C declaration. Returns the next argument number to use. | |
699 | ||
700 | The promotion behavior here must match the regular GCC function | |
701 | parameter marshalling machinery. */ | |
702 | ||
703 | static int | |
704 | write_arg_type (std::stringstream &s, int for_reg, int argno, | |
705 | tree type, bool prototyped) | |
706 | { | |
707 | machine_mode mode = TYPE_MODE (type); | |
708 | ||
709 | if (mode == VOIDmode) | |
710 | return argno; | |
711 | ||
712 | if (pass_in_memory (mode, type, false)) | |
713 | mode = Pmode; | |
714 | else | |
715 | { | |
716 | bool split = TREE_CODE (type) == COMPLEX_TYPE; | |
717 | ||
718 | if (split) | |
719 | { | |
720 | /* Complex types are sent as two separate args. */ | |
721 | type = TREE_TYPE (type); | |
722 | mode = TYPE_MODE (type); | |
723 | prototyped = true; | |
724 | } | |
725 | ||
726 | mode = promote_arg (mode, prototyped); | |
727 | if (split) | |
728 | argno = write_arg_mode (s, for_reg, argno, mode); | |
729 | } | |
730 | ||
731 | return write_arg_mode (s, for_reg, argno, mode); | |
732 | } | |
733 | ||
734 | /* Emit a PTX return as a prototype or function prologue declaration | |
735 | for MODE. */ | |
736 | ||
737 | static void | |
738 | write_return_mode (std::stringstream &s, bool for_proto, machine_mode mode) | |
739 | { | |
740 | const char *ptx_type = nvptx_ptx_type_from_mode (mode, false); | |
741 | const char *pfx = "\t.reg"; | |
742 | const char *sfx = ";\n"; | |
743 | ||
744 | if (for_proto) | |
745 | pfx = "(.param", sfx = "_out) "; | |
746 | ||
747 | s << pfx << ptx_type << " " << reg_names[NVPTX_RETURN_REGNUM] << sfx; | |
748 | } | |
749 | ||
750 | /* Process a function return TYPE to emit a PTX return as a prototype | |
751 | or function prologue declaration. Returns true if return is via an | |
752 | additional pointer parameter. The promotion behavior here must | |
753 | match the regular GCC function return mashalling. */ | |
754 | ||
755 | static bool | |
756 | write_return_type (std::stringstream &s, bool for_proto, tree type) | |
757 | { | |
758 | machine_mode mode = TYPE_MODE (type); | |
759 | ||
760 | if (mode == VOIDmode) | |
761 | return false; | |
762 | ||
763 | bool return_in_mem = pass_in_memory (mode, type, true); | |
764 | ||
765 | if (return_in_mem) | |
766 | { | |
767 | if (for_proto) | |
768 | return return_in_mem; | |
769 | ||
770 | /* Named return values can cause us to return a pointer as well | |
771 | as expect an argument for the return location. This is | |
772 | optimization-level specific, so no caller can make use of | |
773 | this data, but more importantly for us, we must ensure it | |
774 | doesn't change the PTX prototype. */ | |
775 | mode = (machine_mode) cfun->machine->return_mode; | |
776 | ||
777 | if (mode == VOIDmode) | |
778 | return return_in_mem; | |
779 | ||
780 | /* Clear return_mode to inhibit copy of retval to non-existent | |
781 | retval parameter. */ | |
782 | cfun->machine->return_mode = VOIDmode; | |
783 | } | |
784 | else | |
785 | mode = promote_return (mode); | |
786 | ||
787 | write_return_mode (s, for_proto, mode); | |
788 | ||
789 | return return_in_mem; | |
790 | } | |
791 | ||
792 | /* Look for attributes in ATTRS that would indicate we must write a function | |
793 | as a .entry kernel rather than a .func. Return true if one is found. */ | |
794 | ||
795 | static bool | |
796 | write_as_kernel (tree attrs) | |
797 | { | |
798 | return (lookup_attribute ("kernel", attrs) != NULL_TREE | |
799 | || (lookup_attribute ("omp target entrypoint", attrs) != NULL_TREE | |
800 | && lookup_attribute ("oacc function", attrs) != NULL_TREE)); | |
801 | /* For OpenMP target regions, the corresponding kernel entry is emitted from | |
802 | write_omp_entry as a separate function. */ | |
803 | } | |
804 | ||
805 | /* Emit a linker marker for a function decl or defn. */ | |
806 | ||
807 | static void | |
808 | write_fn_marker (std::stringstream &s, bool is_defn, bool globalize, | |
809 | const char *name) | |
810 | { | |
811 | s << "\n// BEGIN"; | |
812 | if (globalize) | |
813 | s << " GLOBAL"; | |
814 | s << " FUNCTION " << (is_defn ? "DEF: " : "DECL: "); | |
815 | s << name << "\n"; | |
816 | } | |
817 | ||
818 | /* Emit a linker marker for a variable decl or defn. */ | |
819 | ||
820 | static void | |
821 | write_var_marker (FILE *file, bool is_defn, bool globalize, const char *name) | |
822 | { | |
823 | fprintf (file, "\n// BEGIN%s VAR %s: ", | |
824 | globalize ? " GLOBAL" : "", | |
825 | is_defn ? "DEF" : "DECL"); | |
826 | assemble_name_raw (file, name); | |
827 | fputs ("\n", file); | |
828 | } | |
829 | ||
830 | /* Write a .func or .kernel declaration or definition along with | |
831 | a helper comment for use by ld. S is the stream to write to, DECL | |
832 | the decl for the function with name NAME. For definitions, emit | |
833 | a declaration too. */ | |
834 | ||
835 | static const char * | |
836 | write_fn_proto (std::stringstream &s, bool is_defn, | |
837 | const char *name, const_tree decl) | |
838 | { | |
839 | if (is_defn) | |
840 | /* Emit a declaration. The PTX assembler gets upset without it. */ | |
841 | name = write_fn_proto (s, false, name, decl); | |
842 | else | |
843 | { | |
844 | /* Avoid repeating the name replacement. */ | |
845 | name = nvptx_name_replacement (name); | |
846 | if (name[0] == '*') | |
847 | name++; | |
848 | } | |
849 | ||
850 | write_fn_marker (s, is_defn, TREE_PUBLIC (decl), name); | |
851 | ||
852 | /* PTX declaration. */ | |
853 | if (DECL_EXTERNAL (decl)) | |
854 | s << ".extern "; | |
855 | else if (TREE_PUBLIC (decl)) | |
856 | s << (DECL_WEAK (decl) ? ".weak " : ".visible "); | |
857 | s << (write_as_kernel (DECL_ATTRIBUTES (decl)) ? ".entry " : ".func "); | |
858 | ||
859 | tree fntype = TREE_TYPE (decl); | |
860 | tree result_type = TREE_TYPE (fntype); | |
861 | ||
862 | /* atomic_compare_exchange_$n builtins have an exceptional calling | |
863 | convention. */ | |
864 | int not_atomic_weak_arg = -1; | |
865 | if (DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL) | |
866 | switch (DECL_FUNCTION_CODE (decl)) | |
867 | { | |
868 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_1: | |
869 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_2: | |
870 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_4: | |
871 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_8: | |
872 | case BUILT_IN_ATOMIC_COMPARE_EXCHANGE_16: | |
873 | /* These atomics skip the 'weak' parm in an actual library | |
874 | call. We must skip it in the prototype too. */ | |
875 | not_atomic_weak_arg = 3; | |
876 | break; | |
877 | ||
878 | default: | |
879 | break; | |
880 | } | |
881 | ||
882 | /* Declare the result. */ | |
883 | bool return_in_mem = write_return_type (s, true, result_type); | |
884 | ||
885 | s << name; | |
886 | ||
887 | int argno = 0; | |
888 | ||
889 | /* Emit argument list. */ | |
890 | if (return_in_mem) | |
891 | argno = write_arg_type (s, -1, argno, ptr_type_node, true); | |
892 | ||
893 | /* We get: | |
894 | NULL in TYPE_ARG_TYPES, for old-style functions | |
895 | NULL in DECL_ARGUMENTS, for builtin functions without another | |
896 | declaration. | |
897 | So we have to pick the best one we have. */ | |
898 | tree args = TYPE_ARG_TYPES (fntype); | |
899 | bool prototyped = true; | |
900 | if (!args) | |
901 | { | |
902 | args = DECL_ARGUMENTS (decl); | |
903 | prototyped = false; | |
904 | } | |
905 | ||
906 | for (; args; args = TREE_CHAIN (args), not_atomic_weak_arg--) | |
907 | { | |
908 | tree type = prototyped ? TREE_VALUE (args) : TREE_TYPE (args); | |
909 | ||
910 | if (not_atomic_weak_arg) | |
911 | argno = write_arg_type (s, -1, argno, type, prototyped); | |
912 | else | |
913 | gcc_assert (type == boolean_type_node); | |
914 | } | |
915 | ||
916 | if (stdarg_p (fntype)) | |
917 | argno = write_arg_type (s, -1, argno, ptr_type_node, true); | |
918 | ||
919 | if (DECL_STATIC_CHAIN (decl)) | |
920 | argno = write_arg_type (s, -1, argno, ptr_type_node, true); | |
921 | ||
922 | if (!argno && strcmp (name, "main") == 0) | |
923 | { | |
924 | argno = write_arg_type (s, -1, argno, integer_type_node, true); | |
925 | argno = write_arg_type (s, -1, argno, ptr_type_node, true); | |
926 | } | |
927 | ||
928 | if (argno) | |
929 | s << ")"; | |
930 | ||
931 | s << (is_defn ? "\n" : ";\n"); | |
932 | ||
933 | return name; | |
934 | } | |
935 | ||
936 | /* Construct a function declaration from a call insn. This can be | |
937 | necessary for two reasons - either we have an indirect call which | |
938 | requires a .callprototype declaration, or we have a libcall | |
939 | generated by emit_library_call for which no decl exists. */ | |
940 | ||
941 | static void | |
942 | write_fn_proto_from_insn (std::stringstream &s, const char *name, | |
943 | rtx result, rtx pat) | |
944 | { | |
945 | if (!name) | |
946 | { | |
947 | s << "\t.callprototype "; | |
948 | name = "_"; | |
949 | } | |
950 | else | |
951 | { | |
952 | name = nvptx_name_replacement (name); | |
953 | write_fn_marker (s, false, true, name); | |
954 | s << "\t.extern .func "; | |
955 | } | |
956 | ||
957 | if (result != NULL_RTX) | |
958 | write_return_mode (s, true, GET_MODE (result)); | |
959 | ||
960 | s << name; | |
961 | ||
962 | int arg_end = XVECLEN (pat, 0); | |
963 | for (int i = 1; i < arg_end; i++) | |
964 | { | |
965 | /* We don't have to deal with mode splitting & promotion here, | |
966 | as that was already done when generating the call | |
967 | sequence. */ | |
968 | machine_mode mode = GET_MODE (XEXP (XVECEXP (pat, 0, i), 0)); | |
969 | ||
970 | write_arg_mode (s, -1, i - 1, mode); | |
971 | } | |
972 | if (arg_end != 1) | |
973 | s << ")"; | |
974 | s << ";\n"; | |
975 | } | |
976 | ||
977 | /* DECL is an external FUNCTION_DECL, make sure its in the fndecl hash | |
978 | table and and write a ptx prototype. These are emitted at end of | |
979 | compilation. */ | |
980 | ||
981 | static void | |
982 | nvptx_record_fndecl (tree decl) | |
983 | { | |
984 | tree *slot = declared_fndecls_htab->find_slot (decl, INSERT); | |
985 | if (*slot == NULL) | |
986 | { | |
987 | *slot = decl; | |
988 | const char *name = get_fnname_from_decl (decl); | |
989 | write_fn_proto (func_decls, false, name, decl); | |
990 | } | |
991 | } | |
992 | ||
993 | /* Record a libcall or unprototyped external function. CALLEE is the | |
994 | SYMBOL_REF. Insert into the libfunc hash table and emit a ptx | |
995 | declaration for it. */ | |
996 | ||
997 | static void | |
998 | nvptx_record_libfunc (rtx callee, rtx retval, rtx pat) | |
999 | { | |
1000 | rtx *slot = declared_libfuncs_htab->find_slot (callee, INSERT); | |
1001 | if (*slot == NULL) | |
1002 | { | |
1003 | *slot = callee; | |
1004 | ||
1005 | const char *name = XSTR (callee, 0); | |
1006 | write_fn_proto_from_insn (func_decls, name, retval, pat); | |
1007 | } | |
1008 | } | |
1009 | ||
1010 | /* DECL is an external FUNCTION_DECL, that we're referencing. If it | |
1011 | is prototyped, record it now. Otherwise record it as needed at end | |
1012 | of compilation, when we might have more information about it. */ | |
1013 | ||
1014 | void | |
1015 | nvptx_record_needed_fndecl (tree decl) | |
1016 | { | |
1017 | if (TYPE_ARG_TYPES (TREE_TYPE (decl)) == NULL_TREE) | |
1018 | { | |
1019 | tree *slot = needed_fndecls_htab->find_slot (decl, INSERT); | |
1020 | if (*slot == NULL) | |
1021 | *slot = decl; | |
1022 | } | |
1023 | else | |
1024 | nvptx_record_fndecl (decl); | |
1025 | } | |
1026 | ||
1027 | /* SYM is a SYMBOL_REF. If it refers to an external function, record | |
1028 | it as needed. */ | |
1029 | ||
1030 | static void | |
1031 | nvptx_maybe_record_fnsym (rtx sym) | |
1032 | { | |
1033 | tree decl = SYMBOL_REF_DECL (sym); | |
1034 | ||
1035 | if (decl && TREE_CODE (decl) == FUNCTION_DECL && DECL_EXTERNAL (decl)) | |
1036 | nvptx_record_needed_fndecl (decl); | |
1037 | } | |
1038 | ||
1039 | /* Emit a local array to hold some part of a conventional stack frame | |
1040 | and initialize REGNO to point to it. If the size is zero, it'll | |
1041 | never be valid to dereference, so we can simply initialize to | |
1042 | zero. */ | |
1043 | ||
1044 | static void | |
1045 | init_frame (FILE *file, int regno, unsigned align, unsigned size) | |
1046 | { | |
1047 | if (size) | |
1048 | fprintf (file, "\t.local .align %d .b8 %s_ar[%u];\n", | |
1049 | align, reg_names[regno], size); | |
1050 | fprintf (file, "\t.reg.u%d %s;\n", | |
1051 | POINTER_SIZE, reg_names[regno]); | |
1052 | fprintf (file, (size ? "\tcvta.local.u%d %s, %s_ar;\n" | |
1053 | : "\tmov.u%d %s, 0;\n"), | |
1054 | POINTER_SIZE, reg_names[regno], reg_names[regno]); | |
1055 | } | |
1056 | ||
1057 | /* Emit soft stack frame setup sequence. */ | |
1058 | ||
1059 | static void | |
1060 | init_softstack_frame (FILE *file, unsigned alignment, HOST_WIDE_INT size) | |
1061 | { | |
1062 | /* Maintain 64-bit stack alignment. */ | |
1063 | unsigned keep_align = BIGGEST_ALIGNMENT / BITS_PER_UNIT; | |
1064 | size = ROUND_UP (size, keep_align); | |
1065 | int bits = POINTER_SIZE; | |
1066 | const char *reg_stack = reg_names[STACK_POINTER_REGNUM]; | |
1067 | const char *reg_frame = reg_names[FRAME_POINTER_REGNUM]; | |
1068 | const char *reg_sspslot = reg_names[SOFTSTACK_SLOT_REGNUM]; | |
1069 | const char *reg_sspprev = reg_names[SOFTSTACK_PREV_REGNUM]; | |
1070 | fprintf (file, "\t.reg.u%d %s;\n", bits, reg_stack); | |
1071 | fprintf (file, "\t.reg.u%d %s;\n", bits, reg_frame); | |
1072 | fprintf (file, "\t.reg.u%d %s;\n", bits, reg_sspslot); | |
1073 | fprintf (file, "\t.reg.u%d %s;\n", bits, reg_sspprev); | |
1074 | fprintf (file, "\t{\n"); | |
1075 | fprintf (file, "\t\t.reg.u32 %%fstmp0;\n"); | |
1076 | fprintf (file, "\t\t.reg.u%d %%fstmp1;\n", bits); | |
1077 | fprintf (file, "\t\t.reg.u%d %%fstmp2;\n", bits); | |
1078 | fprintf (file, "\t\tmov.u32 %%fstmp0, %%tid.y;\n"); | |
1079 | fprintf (file, "\t\tmul%s.u32 %%fstmp1, %%fstmp0, %d;\n", | |
1080 | bits == 64 ? ".wide" : ".lo", bits / 8); | |
1081 | fprintf (file, "\t\tmov.u%d %%fstmp2, __nvptx_stacks;\n", bits); | |
1082 | ||
1083 | /* Initialize %sspslot = &__nvptx_stacks[tid.y]. */ | |
1084 | fprintf (file, "\t\tadd.u%d %s, %%fstmp2, %%fstmp1;\n", bits, reg_sspslot); | |
1085 | ||
1086 | /* Initialize %sspprev = __nvptx_stacks[tid.y]. */ | |
1087 | fprintf (file, "\t\tld.shared.u%d %s, [%s];\n", | |
1088 | bits, reg_sspprev, reg_sspslot); | |
1089 | ||
1090 | /* Initialize %frame = %sspprev - size. */ | |
1091 | fprintf (file, "\t\tsub.u%d %s, %s, " HOST_WIDE_INT_PRINT_DEC ";\n", | |
1092 | bits, reg_frame, reg_sspprev, size); | |
1093 | ||
1094 | /* Apply alignment, if larger than 64. */ | |
1095 | if (alignment > keep_align) | |
1096 | fprintf (file, "\t\tand.b%d %s, %s, %d;\n", | |
1097 | bits, reg_frame, reg_frame, -alignment); | |
1098 | ||
1099 | size = crtl->outgoing_args_size; | |
1100 | gcc_assert (size % keep_align == 0); | |
1101 | ||
1102 | /* Initialize %stack. */ | |
1103 | fprintf (file, "\t\tsub.u%d %s, %s, " HOST_WIDE_INT_PRINT_DEC ";\n", | |
1104 | bits, reg_stack, reg_frame, size); | |
1105 | ||
1106 | if (!crtl->is_leaf) | |
1107 | fprintf (file, "\t\tst.shared.u%d [%s], %s;\n", | |
1108 | bits, reg_sspslot, reg_stack); | |
1109 | fprintf (file, "\t}\n"); | |
1110 | cfun->machine->has_softstack = true; | |
1111 | need_softstack_decl = true; | |
1112 | } | |
1113 | ||
1114 | /* Emit code to initialize the REGNO predicate register to indicate | |
1115 | whether we are not lane zero on the NAME axis. */ | |
1116 | ||
1117 | static void | |
1118 | nvptx_init_axis_predicate (FILE *file, int regno, const char *name) | |
1119 | { | |
1120 | fprintf (file, "\t{\n"); | |
1121 | fprintf (file, "\t\t.reg.u32\t%%%s;\n", name); | |
1122 | if (strcmp (name, "x") == 0 && cfun->machine->red_partition) | |
1123 | { | |
1124 | fprintf (file, "\t\t.reg.u64\t%%t_red;\n"); | |
1125 | fprintf (file, "\t\t.reg.u64\t%%y64;\n"); | |
1126 | } | |
1127 | fprintf (file, "\t\tmov.u32\t%%%s, %%tid.%s;\n", name, name); | |
1128 | fprintf (file, "\t\tsetp.ne.u32\t%%r%d, %%%s, 0;\n", regno, name); | |
1129 | if (strcmp (name, "x") == 0 && cfun->machine->red_partition) | |
1130 | { | |
1131 | fprintf (file, "\t\tcvt.u64.u32\t%%y64, %%tid.y;\n"); | |
1132 | fprintf (file, "\t\tcvta.shared.u64\t%%t_red, __vector_red;\n"); | |
1133 | fprintf (file, "\t\tmad.lo.u64\t%%r%d, %%y64, %d, %%t_red; " | |
1134 | "// vector reduction buffer\n", | |
1135 | REGNO (cfun->machine->red_partition), | |
1136 | vector_red_partition); | |
1137 | } | |
1138 | /* Verify vector_red_size. */ | |
1139 | gcc_assert (vector_red_partition * nvptx_mach_max_workers () | |
1140 | <= vector_red_size); | |
1141 | fprintf (file, "\t}\n"); | |
1142 | } | |
1143 | ||
1144 | /* Emit code to initialize OpenACC worker broadcast and synchronization | |
1145 | registers. */ | |
1146 | ||
1147 | static void | |
1148 | nvptx_init_oacc_workers (FILE *file) | |
1149 | { | |
1150 | fprintf (file, "\t{\n"); | |
1151 | fprintf (file, "\t\t.reg.u32\t%%tidy;\n"); | |
1152 | if (cfun->machine->bcast_partition) | |
1153 | { | |
1154 | fprintf (file, "\t\t.reg.u64\t%%t_bcast;\n"); | |
1155 | fprintf (file, "\t\t.reg.u64\t%%y64;\n"); | |
1156 | } | |
1157 | fprintf (file, "\t\tmov.u32\t\t%%tidy, %%tid.y;\n"); | |
1158 | if (cfun->machine->bcast_partition) | |
1159 | { | |
1160 | fprintf (file, "\t\tcvt.u64.u32\t%%y64, %%tidy;\n"); | |
1161 | fprintf (file, "\t\tadd.u64\t\t%%y64, %%y64, 1; // vector ID\n"); | |
1162 | fprintf (file, "\t\tcvta.shared.u64\t%%t_bcast, __oacc_bcast;\n"); | |
1163 | fprintf (file, "\t\tmad.lo.u64\t%%r%d, %%y64, %d, %%t_bcast; " | |
1164 | "// vector broadcast offset\n", | |
1165 | REGNO (cfun->machine->bcast_partition), | |
1166 | oacc_bcast_partition); | |
1167 | } | |
1168 | /* Verify oacc_bcast_size. */ | |
1169 | gcc_assert (oacc_bcast_partition * (nvptx_mach_max_workers () + 1) | |
1170 | <= oacc_bcast_size); | |
1171 | if (cfun->machine->sync_bar) | |
1172 | fprintf (file, "\t\tadd.u32\t\t%%r%d, %%tidy, 1; " | |
1173 | "// vector synchronization barrier\n", | |
1174 | REGNO (cfun->machine->sync_bar)); | |
1175 | fprintf (file, "\t}\n"); | |
1176 | } | |
1177 | ||
1178 | /* Emit code to initialize predicate and master lane index registers for | |
1179 | -muniform-simt code generation variant. */ | |
1180 | ||
1181 | static void | |
1182 | nvptx_init_unisimt_predicate (FILE *file) | |
1183 | { | |
1184 | cfun->machine->unisimt_location = gen_reg_rtx (Pmode); | |
1185 | int loc = REGNO (cfun->machine->unisimt_location); | |
1186 | int bits = POINTER_SIZE; | |
1187 | fprintf (file, "\t.reg.u%d %%r%d;\n", bits, loc); | |
1188 | fprintf (file, "\t{\n"); | |
1189 | fprintf (file, "\t\t.reg.u32 %%ustmp0;\n"); | |
1190 | fprintf (file, "\t\t.reg.u%d %%ustmp1;\n", bits); | |
1191 | fprintf (file, "\t\tmov.u32 %%ustmp0, %%tid.y;\n"); | |
1192 | fprintf (file, "\t\tmul%s.u32 %%ustmp1, %%ustmp0, 4;\n", | |
1193 | bits == 64 ? ".wide" : ".lo"); | |
1194 | fprintf (file, "\t\tmov.u%d %%r%d, __nvptx_uni;\n", bits, loc); | |
1195 | fprintf (file, "\t\tadd.u%d %%r%d, %%r%d, %%ustmp1;\n", bits, loc, loc); | |
1196 | if (cfun->machine->unisimt_predicate) | |
1197 | { | |
1198 | int master = REGNO (cfun->machine->unisimt_master); | |
1199 | int pred = REGNO (cfun->machine->unisimt_predicate); | |
1200 | fprintf (file, "\t\tld.shared.u32 %%r%d, [%%r%d];\n", master, loc); | |
1201 | fprintf (file, "\t\tmov.u32 %%ustmp0, %%laneid;\n"); | |
1202 | /* Compute 'master lane index' as 'laneid & __nvptx_uni[tid.y]'. */ | |
1203 | fprintf (file, "\t\tand.b32 %%r%d, %%r%d, %%ustmp0;\n", master, master); | |
1204 | /* Compute predicate as 'tid.x == master'. */ | |
1205 | fprintf (file, "\t\tsetp.eq.u32 %%r%d, %%r%d, %%ustmp0;\n", pred, master); | |
1206 | } | |
1207 | fprintf (file, "\t}\n"); | |
1208 | need_unisimt_decl = true; | |
1209 | } | |
1210 | ||
1211 | /* Emit kernel NAME for function ORIG outlined for an OpenMP 'target' region: | |
1212 | ||
1213 | extern void gomp_nvptx_main (void (*fn)(void*), void *fnarg); | |
1214 | void __attribute__((kernel)) NAME (void *arg, char *stack, size_t stacksize) | |
1215 | { | |
1216 | __nvptx_stacks[tid.y] = stack + stacksize * (ctaid.x * ntid.y + tid.y + 1); | |
1217 | __nvptx_uni[tid.y] = 0; | |
1218 | gomp_nvptx_main (ORIG, arg); | |
1219 | } | |
1220 | ORIG itself should not be emitted as a PTX .entry function. */ | |
1221 | ||
1222 | static void | |
1223 | write_omp_entry (FILE *file, const char *name, const char *orig) | |
1224 | { | |
1225 | static bool gomp_nvptx_main_declared; | |
1226 | if (!gomp_nvptx_main_declared) | |
1227 | { | |
1228 | gomp_nvptx_main_declared = true; | |
1229 | write_fn_marker (func_decls, false, true, "gomp_nvptx_main"); | |
1230 | func_decls << ".extern .func gomp_nvptx_main (.param.u" << POINTER_SIZE | |
1231 | << " %in_ar1, .param.u" << POINTER_SIZE << " %in_ar2);\n"; | |
1232 | } | |
1233 | /* PR79332. Single out this string; it confuses gcc.pot generation. */ | |
1234 | #define NTID_Y "%ntid.y" | |
1235 | #define ENTRY_TEMPLATE(PS, PS_BYTES, MAD_PS_32) "\ | |
1236 | (.param.u" PS " %arg, .param.u" PS " %stack, .param.u" PS " %sz)\n\ | |
1237 | {\n\ | |
1238 | .reg.u32 %r<3>;\n\ | |
1239 | .reg.u" PS " %R<4>;\n\ | |
1240 | mov.u32 %r0, %tid.y;\n\ | |
1241 | mov.u32 %r1, " NTID_Y ";\n\ | |
1242 | mov.u32 %r2, %ctaid.x;\n\ | |
1243 | cvt.u" PS ".u32 %R1, %r0;\n\ | |
1244 | " MAD_PS_32 " %R1, %r1, %r2, %R1;\n\ | |
1245 | mov.u" PS " %R0, __nvptx_stacks;\n\ | |
1246 | " MAD_PS_32 " %R0, %r0, " PS_BYTES ", %R0;\n\ | |
1247 | ld.param.u" PS " %R2, [%stack];\n\ | |
1248 | ld.param.u" PS " %R3, [%sz];\n\ | |
1249 | add.u" PS " %R2, %R2, %R3;\n\ | |
1250 | mad.lo.u" PS " %R2, %R1, %R3, %R2;\n\ | |
1251 | st.shared.u" PS " [%R0], %R2;\n\ | |
1252 | mov.u" PS " %R0, __nvptx_uni;\n\ | |
1253 | " MAD_PS_32 " %R0, %r0, 4, %R0;\n\ | |
1254 | mov.u32 %r0, 0;\n\ | |
1255 | st.shared.u32 [%R0], %r0;\n\ | |
1256 | mov.u" PS " %R0, \0;\n\ | |
1257 | ld.param.u" PS " %R1, [%arg];\n\ | |
1258 | {\n\ | |
1259 | .param.u" PS " %P<2>;\n\ | |
1260 | st.param.u" PS " [%P0], %R0;\n\ | |
1261 | st.param.u" PS " [%P1], %R1;\n\ | |
1262 | call.uni gomp_nvptx_main, (%P0, %P1);\n\ | |
1263 | }\n\ | |
1264 | ret.uni;\n\ | |
1265 | }\n" | |
1266 | static const char entry64[] = ENTRY_TEMPLATE ("64", "8", "mad.wide.u32"); | |
1267 | static const char entry32[] = ENTRY_TEMPLATE ("32", "4", "mad.lo.u32 "); | |
1268 | #undef ENTRY_TEMPLATE | |
1269 | #undef NTID_Y | |
1270 | const char *entry_1 = TARGET_ABI64 ? entry64 : entry32; | |
1271 | /* Position ENTRY_2 after the embedded nul using strlen of the prefix. */ | |
1272 | const char *entry_2 = entry_1 + strlen (entry64) + 1; | |
1273 | fprintf (file, ".visible .entry %s%s%s%s", name, entry_1, orig, entry_2); | |
1274 | need_softstack_decl = need_unisimt_decl = true; | |
1275 | } | |
1276 | ||
1277 | /* Implement ASM_DECLARE_FUNCTION_NAME. Writes the start of a ptx | |
1278 | function, including local var decls and copies from the arguments to | |
1279 | local regs. */ | |
1280 | ||
1281 | void | |
1282 | nvptx_declare_function_name (FILE *file, const char *name, const_tree decl) | |
1283 | { | |
1284 | tree fntype = TREE_TYPE (decl); | |
1285 | tree result_type = TREE_TYPE (fntype); | |
1286 | int argno = 0; | |
1287 | ||
1288 | if (lookup_attribute ("omp target entrypoint", DECL_ATTRIBUTES (decl)) | |
1289 | && !lookup_attribute ("oacc function", DECL_ATTRIBUTES (decl))) | |
1290 | { | |
1291 | char *buf = (char *) alloca (strlen (name) + sizeof ("$impl")); | |
1292 | sprintf (buf, "%s$impl", name); | |
1293 | write_omp_entry (file, name, buf); | |
1294 | name = buf; | |
1295 | } | |
1296 | /* We construct the initial part of the function into a string | |
1297 | stream, in order to share the prototype writing code. */ | |
1298 | std::stringstream s; | |
1299 | write_fn_proto (s, true, name, decl); | |
1300 | s << "{\n"; | |
1301 | ||
1302 | bool return_in_mem = write_return_type (s, false, result_type); | |
1303 | if (return_in_mem) | |
1304 | argno = write_arg_type (s, 0, argno, ptr_type_node, true); | |
1305 | ||
1306 | /* Declare and initialize incoming arguments. */ | |
1307 | tree args = TYPE_ARG_TYPES (fntype); | |
1308 | bool prototyped = true; | |
1309 | if (!args) | |
1310 | { | |
1311 | args = DECL_ARGUMENTS (decl); | |
1312 | prototyped = false; | |
1313 | } | |
1314 | ||
1315 | for (; args != NULL_TREE; args = TREE_CHAIN (args)) | |
1316 | { | |
1317 | tree type = prototyped ? TREE_VALUE (args) : TREE_TYPE (args); | |
1318 | ||
1319 | argno = write_arg_type (s, 0, argno, type, prototyped); | |
1320 | } | |
1321 | ||
1322 | if (stdarg_p (fntype)) | |
1323 | argno = write_arg_type (s, ARG_POINTER_REGNUM, argno, ptr_type_node, | |
1324 | true); | |
1325 | ||
1326 | if (DECL_STATIC_CHAIN (decl) || cfun->machine->has_chain) | |
1327 | write_arg_type (s, STATIC_CHAIN_REGNUM, | |
1328 | DECL_STATIC_CHAIN (decl) ? argno : -1, ptr_type_node, | |
1329 | true); | |
1330 | ||
1331 | fprintf (file, "%s", s.str().c_str()); | |
1332 | ||
1333 | /* Usually 'crtl->is_leaf' is computed during register allocator | |
1334 | initialization (which is not done on NVPTX) or for pressure-sensitive | |
1335 | optimizations. Initialize it here, except if already set. */ | |
1336 | if (!crtl->is_leaf) | |
1337 | crtl->is_leaf = leaf_function_p (); | |
1338 | ||
1339 | HOST_WIDE_INT sz = get_frame_size (); | |
1340 | bool need_frameptr = sz || cfun->machine->has_chain; | |
1341 | int alignment = crtl->stack_alignment_needed / BITS_PER_UNIT; | |
1342 | if (!TARGET_SOFT_STACK) | |
1343 | { | |
1344 | /* Declare a local var for outgoing varargs. */ | |
1345 | if (cfun->machine->has_varadic) | |
1346 | init_frame (file, STACK_POINTER_REGNUM, | |
1347 | UNITS_PER_WORD, crtl->outgoing_args_size); | |
1348 | ||
1349 | /* Declare a local variable for the frame. Force its size to be | |
1350 | DImode-compatible. */ | |
1351 | if (need_frameptr) | |
1352 | init_frame (file, FRAME_POINTER_REGNUM, alignment, | |
1353 | ROUND_UP (sz, GET_MODE_SIZE (DImode))); | |
1354 | } | |
1355 | else if (need_frameptr || cfun->machine->has_varadic || cfun->calls_alloca | |
1356 | || (cfun->machine->has_simtreg && !crtl->is_leaf)) | |
1357 | init_softstack_frame (file, alignment, sz); | |
1358 | ||
1359 | if (cfun->machine->has_simtreg) | |
1360 | { | |
1361 | unsigned HOST_WIDE_INT &simtsz = cfun->machine->simt_stack_size; | |
1362 | unsigned HOST_WIDE_INT &align = cfun->machine->simt_stack_align; | |
1363 | align = MAX (align, GET_MODE_SIZE (DImode)); | |
1364 | if (!crtl->is_leaf || cfun->calls_alloca) | |
1365 | simtsz = HOST_WIDE_INT_M1U; | |
1366 | if (simtsz == HOST_WIDE_INT_M1U) | |
1367 | simtsz = nvptx_softstack_size; | |
1368 | if (cfun->machine->has_softstack) | |
1369 | simtsz += POINTER_SIZE / 8; | |
1370 | simtsz = ROUND_UP (simtsz, GET_MODE_SIZE (DImode)); | |
1371 | if (align > GET_MODE_SIZE (DImode)) | |
1372 | simtsz += align - GET_MODE_SIZE (DImode); | |
1373 | if (simtsz) | |
1374 | fprintf (file, "\t.local.align 8 .b8 %%simtstack_ar[" | |
1375 | HOST_WIDE_INT_PRINT_DEC "];\n", simtsz); | |
1376 | } | |
1377 | ||
1378 | /* Restore the vector reduction partition register, if necessary. | |
1379 | FIXME: Find out when and why this is necessary, and fix it. */ | |
1380 | if (cfun->machine->red_partition) | |
1381 | regno_reg_rtx[REGNO (cfun->machine->red_partition)] | |
1382 | = cfun->machine->red_partition; | |
1383 | ||
1384 | /* Declare the pseudos we have as ptx registers. */ | |
1385 | int maxregs = max_reg_num (); | |
1386 | for (int i = LAST_VIRTUAL_REGISTER + 1; i < maxregs; i++) | |
1387 | { | |
1388 | if (regno_reg_rtx[i] != const0_rtx) | |
1389 | { | |
1390 | machine_mode mode = PSEUDO_REGNO_MODE (i); | |
1391 | machine_mode split = maybe_split_mode (mode); | |
1392 | ||
1393 | if (split_mode_p (mode)) | |
1394 | mode = split; | |
1395 | fprintf (file, "\t.reg%s ", nvptx_ptx_type_from_mode (mode, true)); | |
1396 | output_reg (file, i, split, -2); | |
1397 | fprintf (file, ";\n"); | |
1398 | } | |
1399 | } | |
1400 | ||
1401 | /* Emit axis predicates. */ | |
1402 | if (cfun->machine->axis_predicate[0]) | |
1403 | nvptx_init_axis_predicate (file, | |
1404 | REGNO (cfun->machine->axis_predicate[0]), "y"); | |
1405 | if (cfun->machine->axis_predicate[1]) | |
1406 | nvptx_init_axis_predicate (file, | |
1407 | REGNO (cfun->machine->axis_predicate[1]), "x"); | |
1408 | if (cfun->machine->unisimt_predicate | |
1409 | || (cfun->machine->has_simtreg && !crtl->is_leaf)) | |
1410 | nvptx_init_unisimt_predicate (file); | |
1411 | if (cfun->machine->bcast_partition || cfun->machine->sync_bar) | |
1412 | nvptx_init_oacc_workers (file); | |
1413 | } | |
1414 | ||
1415 | /* Output code for switching uniform-simt state. ENTERING indicates whether | |
1416 | we are entering or leaving non-uniform execution region. */ | |
1417 | ||
1418 | static void | |
1419 | nvptx_output_unisimt_switch (FILE *file, bool entering) | |
1420 | { | |
1421 | if (crtl->is_leaf && !cfun->machine->unisimt_predicate) | |
1422 | return; | |
1423 | fprintf (file, "\t{\n"); | |
1424 | fprintf (file, "\t\t.reg.u32 %%ustmp2;\n"); | |
1425 | fprintf (file, "\t\tmov.u32 %%ustmp2, %d;\n", entering ? -1 : 0); | |
1426 | if (!crtl->is_leaf) | |
1427 | { | |
1428 | int loc = REGNO (cfun->machine->unisimt_location); | |
1429 | fprintf (file, "\t\tst.shared.u32 [%%r%d], %%ustmp2;\n", loc); | |
1430 | } | |
1431 | if (cfun->machine->unisimt_predicate) | |
1432 | { | |
1433 | int master = REGNO (cfun->machine->unisimt_master); | |
1434 | int pred = REGNO (cfun->machine->unisimt_predicate); | |
1435 | fprintf (file, "\t\tmov.u32 %%ustmp2, %%laneid;\n"); | |
1436 | fprintf (file, "\t\tmov.u32 %%r%d, %s;\n", | |
1437 | master, entering ? "%ustmp2" : "0"); | |
1438 | fprintf (file, "\t\tsetp.eq.u32 %%r%d, %%r%d, %%ustmp2;\n", pred, master); | |
1439 | } | |
1440 | fprintf (file, "\t}\n"); | |
1441 | } | |
1442 | ||
1443 | /* Output code for allocating per-lane storage and switching soft-stack pointer. | |
1444 | ENTERING indicates whether we are entering or leaving non-uniform execution. | |
1445 | PTR is the register pointing to allocated storage, it is assigned to on | |
1446 | entering and used to restore state on leaving. SIZE and ALIGN are used only | |
1447 | on entering. */ | |
1448 | ||
1449 | static void | |
1450 | nvptx_output_softstack_switch (FILE *file, bool entering, | |
1451 | rtx ptr, rtx size, rtx align) | |
1452 | { | |
1453 | gcc_assert (REG_P (ptr) && !HARD_REGISTER_P (ptr)); | |
1454 | if (crtl->is_leaf && !cfun->machine->simt_stack_size) | |
1455 | return; | |
1456 | int bits = POINTER_SIZE, regno = REGNO (ptr); | |
1457 | fprintf (file, "\t{\n"); | |
1458 | if (entering) | |
1459 | { | |
1460 | fprintf (file, "\t\tcvta.local.u%d %%r%d, %%simtstack_ar + " | |
1461 | HOST_WIDE_INT_PRINT_DEC ";\n", bits, regno, | |
1462 | cfun->machine->simt_stack_size); | |
1463 | fprintf (file, "\t\tsub.u%d %%r%d, %%r%d, ", bits, regno, regno); | |
1464 | if (CONST_INT_P (size)) | |
1465 | fprintf (file, HOST_WIDE_INT_PRINT_DEC, | |
1466 | ROUND_UP (UINTVAL (size), GET_MODE_SIZE (DImode))); | |
1467 | else | |
1468 | output_reg (file, REGNO (size), VOIDmode); | |
1469 | fputs (";\n", file); | |
1470 | if (!CONST_INT_P (size) || UINTVAL (align) > GET_MODE_SIZE (DImode)) | |
1471 | fprintf (file, | |
1472 | "\t\tand.b%d %%r%d, %%r%d, -" HOST_WIDE_INT_PRINT_DEC ";\n", | |
1473 | bits, regno, regno, UINTVAL (align)); | |
1474 | } | |
1475 | if (cfun->machine->has_softstack) | |
1476 | { | |
1477 | const char *reg_stack = reg_names[STACK_POINTER_REGNUM]; | |
1478 | if (entering) | |
1479 | { | |
1480 | fprintf (file, "\t\tst.u%d [%%r%d + -%d], %s;\n", | |
1481 | bits, regno, bits / 8, reg_stack); | |
1482 | fprintf (file, "\t\tsub.u%d %s, %%r%d, %d;\n", | |
1483 | bits, reg_stack, regno, bits / 8); | |
1484 | } | |
1485 | else | |
1486 | { | |
1487 | fprintf (file, "\t\tld.u%d %s, [%%r%d + -%d];\n", | |
1488 | bits, reg_stack, regno, bits / 8); | |
1489 | } | |
1490 | nvptx_output_set_softstack (REGNO (stack_pointer_rtx)); | |
1491 | } | |
1492 | fprintf (file, "\t}\n"); | |
1493 | } | |
1494 | ||
1495 | /* Output code to enter non-uniform execution region. DEST is a register | |
1496 | to hold a per-lane allocation given by SIZE and ALIGN. */ | |
1497 | ||
1498 | const char * | |
1499 | nvptx_output_simt_enter (rtx dest, rtx size, rtx align) | |
1500 | { | |
1501 | nvptx_output_unisimt_switch (asm_out_file, true); | |
1502 | nvptx_output_softstack_switch (asm_out_file, true, dest, size, align); | |
1503 | return ""; | |
1504 | } | |
1505 | ||
1506 | /* Output code to leave non-uniform execution region. SRC is the register | |
1507 | holding per-lane storage previously allocated by omp_simt_enter insn. */ | |
1508 | ||
1509 | const char * | |
1510 | nvptx_output_simt_exit (rtx src) | |
1511 | { | |
1512 | nvptx_output_unisimt_switch (asm_out_file, false); | |
1513 | nvptx_output_softstack_switch (asm_out_file, false, src, NULL_RTX, NULL_RTX); | |
1514 | return ""; | |
1515 | } | |
1516 | ||
1517 | /* Output instruction that sets soft stack pointer in shared memory to the | |
1518 | value in register given by SRC_REGNO. */ | |
1519 | ||
1520 | const char * | |
1521 | nvptx_output_set_softstack (unsigned src_regno) | |
1522 | { | |
1523 | if (cfun->machine->has_softstack && !crtl->is_leaf) | |
1524 | { | |
1525 | fprintf (asm_out_file, "\tst.shared.u%d\t[%s], ", | |
1526 | POINTER_SIZE, reg_names[SOFTSTACK_SLOT_REGNUM]); | |
1527 | output_reg (asm_out_file, src_regno, VOIDmode); | |
1528 | fprintf (asm_out_file, ";\n"); | |
1529 | } | |
1530 | return ""; | |
1531 | } | |
1532 | /* Output a return instruction. Also copy the return value to its outgoing | |
1533 | location. */ | |
1534 | ||
1535 | const char * | |
1536 | nvptx_output_return (void) | |
1537 | { | |
1538 | machine_mode mode = (machine_mode)cfun->machine->return_mode; | |
1539 | ||
1540 | if (mode != VOIDmode) | |
1541 | fprintf (asm_out_file, "\tst.param%s\t[%s_out], %s;\n", | |
1542 | nvptx_ptx_type_from_mode (mode, false), | |
1543 | reg_names[NVPTX_RETURN_REGNUM], | |
1544 | reg_names[NVPTX_RETURN_REGNUM]); | |
1545 | ||
1546 | return "ret;"; | |
1547 | } | |
1548 | ||
1549 | /* Terminate a function by writing a closing brace to FILE. */ | |
1550 | ||
1551 | void | |
1552 | nvptx_function_end (FILE *file) | |
1553 | { | |
1554 | fprintf (file, "}\n"); | |
1555 | } | |
1556 | \f | |
1557 | /* Decide whether we can make a sibling call to a function. For ptx, we | |
1558 | can't. */ | |
1559 | ||
1560 | static bool | |
1561 | nvptx_function_ok_for_sibcall (tree, tree) | |
1562 | { | |
1563 | return false; | |
1564 | } | |
1565 | ||
1566 | /* Return Dynamic ReAlignment Pointer RTX. For PTX there isn't any. */ | |
1567 | ||
1568 | static rtx | |
1569 | nvptx_get_drap_rtx (void) | |
1570 | { | |
1571 | if (TARGET_SOFT_STACK && stack_realign_drap) | |
1572 | return arg_pointer_rtx; | |
1573 | return NULL_RTX; | |
1574 | } | |
1575 | ||
1576 | /* Implement the TARGET_CALL_ARGS hook. Record information about one | |
1577 | argument to the next call. */ | |
1578 | ||
1579 | static void | |
1580 | nvptx_call_args (rtx arg, tree fntype) | |
1581 | { | |
1582 | if (!cfun->machine->doing_call) | |
1583 | { | |
1584 | cfun->machine->doing_call = true; | |
1585 | cfun->machine->is_varadic = false; | |
1586 | cfun->machine->num_args = 0; | |
1587 | ||
1588 | if (fntype && stdarg_p (fntype)) | |
1589 | { | |
1590 | cfun->machine->is_varadic = true; | |
1591 | cfun->machine->has_varadic = true; | |
1592 | cfun->machine->num_args++; | |
1593 | } | |
1594 | } | |
1595 | ||
1596 | if (REG_P (arg) && arg != pc_rtx) | |
1597 | { | |
1598 | cfun->machine->num_args++; | |
1599 | cfun->machine->call_args = alloc_EXPR_LIST (VOIDmode, arg, | |
1600 | cfun->machine->call_args); | |
1601 | } | |
1602 | } | |
1603 | ||
1604 | /* Implement the corresponding END_CALL_ARGS hook. Clear and free the | |
1605 | information we recorded. */ | |
1606 | ||
1607 | static void | |
1608 | nvptx_end_call_args (void) | |
1609 | { | |
1610 | cfun->machine->doing_call = false; | |
1611 | free_EXPR_LIST_list (&cfun->machine->call_args); | |
1612 | } | |
1613 | ||
1614 | /* Emit the sequence for a call to ADDRESS, setting RETVAL. Keep | |
1615 | track of whether calls involving static chains or varargs were seen | |
1616 | in the current function. | |
1617 | For libcalls, maintain a hash table of decls we have seen, and | |
1618 | record a function decl for later when encountering a new one. */ | |
1619 | ||
1620 | void | |
1621 | nvptx_expand_call (rtx retval, rtx address) | |
1622 | { | |
1623 | rtx callee = XEXP (address, 0); | |
1624 | rtx varargs = NULL_RTX; | |
1625 | unsigned parallel = 0; | |
1626 | ||
1627 | if (!call_insn_operand (callee, Pmode)) | |
1628 | { | |
1629 | callee = force_reg (Pmode, callee); | |
1630 | address = change_address (address, QImode, callee); | |
1631 | } | |
1632 | ||
1633 | if (GET_CODE (callee) == SYMBOL_REF) | |
1634 | { | |
1635 | tree decl = SYMBOL_REF_DECL (callee); | |
1636 | if (decl != NULL_TREE) | |
1637 | { | |
1638 | if (DECL_STATIC_CHAIN (decl)) | |
1639 | cfun->machine->has_chain = true; | |
1640 | ||
1641 | tree attr = oacc_get_fn_attrib (decl); | |
1642 | if (attr) | |
1643 | { | |
1644 | tree dims = TREE_VALUE (attr); | |
1645 | ||
1646 | parallel = GOMP_DIM_MASK (GOMP_DIM_MAX) - 1; | |
1647 | for (int ix = 0; ix != GOMP_DIM_MAX; ix++) | |
1648 | { | |
1649 | if (TREE_PURPOSE (dims) | |
1650 | && !integer_zerop (TREE_PURPOSE (dims))) | |
1651 | break; | |
1652 | /* Not on this axis. */ | |
1653 | parallel ^= GOMP_DIM_MASK (ix); | |
1654 | dims = TREE_CHAIN (dims); | |
1655 | } | |
1656 | } | |
1657 | } | |
1658 | } | |
1659 | ||
1660 | unsigned nargs = cfun->machine->num_args; | |
1661 | if (cfun->machine->is_varadic) | |
1662 | { | |
1663 | varargs = gen_reg_rtx (Pmode); | |
1664 | emit_move_insn (varargs, stack_pointer_rtx); | |
1665 | } | |
1666 | ||
1667 | rtvec vec = rtvec_alloc (nargs + 1); | |
1668 | rtx pat = gen_rtx_PARALLEL (VOIDmode, vec); | |
1669 | int vec_pos = 0; | |
1670 | ||
1671 | rtx call = gen_rtx_CALL (VOIDmode, address, const0_rtx); | |
1672 | rtx tmp_retval = retval; | |
1673 | if (retval) | |
1674 | { | |
1675 | if (!nvptx_register_operand (retval, GET_MODE (retval))) | |
1676 | tmp_retval = gen_reg_rtx (GET_MODE (retval)); | |
1677 | call = gen_rtx_SET (tmp_retval, call); | |
1678 | } | |
1679 | XVECEXP (pat, 0, vec_pos++) = call; | |
1680 | ||
1681 | /* Construct the call insn, including a USE for each argument pseudo | |
1682 | register. These will be used when printing the insn. */ | |
1683 | for (rtx arg = cfun->machine->call_args; arg; arg = XEXP (arg, 1)) | |
1684 | XVECEXP (pat, 0, vec_pos++) = gen_rtx_USE (VOIDmode, XEXP (arg, 0)); | |
1685 | ||
1686 | if (varargs) | |
1687 | XVECEXP (pat, 0, vec_pos++) = gen_rtx_USE (VOIDmode, varargs); | |
1688 | ||
1689 | gcc_assert (vec_pos = XVECLEN (pat, 0)); | |
1690 | ||
1691 | nvptx_emit_forking (parallel, true); | |
1692 | emit_call_insn (pat); | |
1693 | nvptx_emit_joining (parallel, true); | |
1694 | ||
1695 | if (tmp_retval != retval) | |
1696 | emit_move_insn (retval, tmp_retval); | |
1697 | } | |
1698 | ||
1699 | /* Emit a comparison COMPARE, and return the new test to be used in the | |
1700 | jump. */ | |
1701 | ||
1702 | rtx | |
1703 | nvptx_expand_compare (rtx compare) | |
1704 | { | |
1705 | rtx pred = gen_reg_rtx (BImode); | |
1706 | rtx cmp = gen_rtx_fmt_ee (GET_CODE (compare), BImode, | |
1707 | XEXP (compare, 0), XEXP (compare, 1)); | |
1708 | emit_insn (gen_rtx_SET (pred, cmp)); | |
1709 | return gen_rtx_NE (BImode, pred, const0_rtx); | |
1710 | } | |
1711 | ||
1712 | /* Expand the oacc fork & join primitive into ptx-required unspecs. */ | |
1713 | ||
1714 | void | |
1715 | nvptx_expand_oacc_fork (unsigned mode) | |
1716 | { | |
1717 | nvptx_emit_forking (GOMP_DIM_MASK (mode), false); | |
1718 | } | |
1719 | ||
1720 | void | |
1721 | nvptx_expand_oacc_join (unsigned mode) | |
1722 | { | |
1723 | nvptx_emit_joining (GOMP_DIM_MASK (mode), false); | |
1724 | } | |
1725 | ||
1726 | /* Generate instruction(s) to unpack a 64 bit object into 2 32 bit | |
1727 | objects. */ | |
1728 | ||
1729 | static rtx | |
1730 | nvptx_gen_unpack (rtx dst0, rtx dst1, rtx src) | |
1731 | { | |
1732 | rtx res; | |
1733 | ||
1734 | switch (GET_MODE (src)) | |
1735 | { | |
1736 | case E_DImode: | |
1737 | res = gen_unpackdisi2 (dst0, dst1, src); | |
1738 | break; | |
1739 | case E_DFmode: | |
1740 | res = gen_unpackdfsi2 (dst0, dst1, src); | |
1741 | break; | |
1742 | default: gcc_unreachable (); | |
1743 | } | |
1744 | return res; | |
1745 | } | |
1746 | ||
1747 | /* Generate instruction(s) to pack 2 32 bit objects into a 64 bit | |
1748 | object. */ | |
1749 | ||
1750 | static rtx | |
1751 | nvptx_gen_pack (rtx dst, rtx src0, rtx src1) | |
1752 | { | |
1753 | rtx res; | |
1754 | ||
1755 | switch (GET_MODE (dst)) | |
1756 | { | |
1757 | case E_DImode: | |
1758 | res = gen_packsidi2 (dst, src0, src1); | |
1759 | break; | |
1760 | case E_DFmode: | |
1761 | res = gen_packsidf2 (dst, src0, src1); | |
1762 | break; | |
1763 | default: gcc_unreachable (); | |
1764 | } | |
1765 | return res; | |
1766 | } | |
1767 | ||
1768 | /* Generate an instruction or sequence to broadcast register REG | |
1769 | across the vectors of a single warp. */ | |
1770 | ||
1771 | rtx | |
1772 | nvptx_gen_shuffle (rtx dst, rtx src, rtx idx, nvptx_shuffle_kind kind) | |
1773 | { | |
1774 | rtx res; | |
1775 | ||
1776 | switch (GET_MODE (dst)) | |
1777 | { | |
1778 | case E_SImode: | |
1779 | res = gen_nvptx_shufflesi (dst, src, idx, GEN_INT (kind)); | |
1780 | break; | |
1781 | case E_SFmode: | |
1782 | res = gen_nvptx_shufflesf (dst, src, idx, GEN_INT (kind)); | |
1783 | break; | |
1784 | case E_DImode: | |
1785 | case E_DFmode: | |
1786 | { | |
1787 | rtx tmp0 = gen_reg_rtx (SImode); | |
1788 | rtx tmp1 = gen_reg_rtx (SImode); | |
1789 | ||
1790 | start_sequence (); | |
1791 | emit_insn (nvptx_gen_unpack (tmp0, tmp1, src)); | |
1792 | emit_insn (nvptx_gen_shuffle (tmp0, tmp0, idx, kind)); | |
1793 | emit_insn (nvptx_gen_shuffle (tmp1, tmp1, idx, kind)); | |
1794 | emit_insn (nvptx_gen_pack (dst, tmp0, tmp1)); | |
1795 | res = get_insns (); | |
1796 | end_sequence (); | |
1797 | } | |
1798 | break; | |
1799 | case E_BImode: | |
1800 | { | |
1801 | rtx tmp = gen_reg_rtx (SImode); | |
1802 | ||
1803 | start_sequence (); | |
1804 | emit_insn (gen_sel_truesi (tmp, src, GEN_INT (1), const0_rtx)); | |
1805 | emit_insn (nvptx_gen_shuffle (tmp, tmp, idx, kind)); | |
1806 | emit_insn (gen_rtx_SET (dst, gen_rtx_NE (BImode, tmp, const0_rtx))); | |
1807 | res = get_insns (); | |
1808 | end_sequence (); | |
1809 | } | |
1810 | break; | |
1811 | case E_QImode: | |
1812 | case E_HImode: | |
1813 | { | |
1814 | rtx tmp = gen_reg_rtx (SImode); | |
1815 | ||
1816 | start_sequence (); | |
1817 | emit_insn (gen_rtx_SET (tmp, gen_rtx_fmt_e (ZERO_EXTEND, SImode, src))); | |
1818 | emit_insn (nvptx_gen_shuffle (tmp, tmp, idx, kind)); | |
1819 | emit_insn (gen_rtx_SET (dst, gen_rtx_fmt_e (TRUNCATE, GET_MODE (dst), | |
1820 | tmp))); | |
1821 | res = get_insns (); | |
1822 | end_sequence (); | |
1823 | } | |
1824 | break; | |
1825 | ||
1826 | default: | |
1827 | gcc_unreachable (); | |
1828 | } | |
1829 | return res; | |
1830 | } | |
1831 | ||
1832 | /* Generate an instruction or sequence to broadcast register REG | |
1833 | across the vectors of a single warp. */ | |
1834 | ||
1835 | static rtx | |
1836 | nvptx_gen_warp_bcast (rtx reg) | |
1837 | { | |
1838 | return nvptx_gen_shuffle (reg, reg, const0_rtx, SHUFFLE_IDX); | |
1839 | } | |
1840 | ||
1841 | /* Structure used when generating a worker-level spill or fill. */ | |
1842 | ||
1843 | struct broadcast_data_t | |
1844 | { | |
1845 | rtx base; /* Register holding base addr of buffer. */ | |
1846 | rtx ptr; /* Iteration var, if needed. */ | |
1847 | unsigned offset; /* Offset into worker buffer. */ | |
1848 | }; | |
1849 | ||
1850 | /* Direction of the spill/fill and looping setup/teardown indicator. */ | |
1851 | ||
1852 | enum propagate_mask | |
1853 | { | |
1854 | PM_read = 1 << 0, | |
1855 | PM_write = 1 << 1, | |
1856 | PM_loop_begin = 1 << 2, | |
1857 | PM_loop_end = 1 << 3, | |
1858 | ||
1859 | PM_read_write = PM_read | PM_write | |
1860 | }; | |
1861 | ||
1862 | /* Generate instruction(s) to spill or fill register REG to/from the | |
1863 | worker broadcast array. PM indicates what is to be done, REP | |
1864 | how many loop iterations will be executed (0 for not a loop). */ | |
1865 | ||
1866 | static rtx | |
1867 | nvptx_gen_shared_bcast (rtx reg, propagate_mask pm, unsigned rep, | |
1868 | broadcast_data_t *data, bool vector) | |
1869 | { | |
1870 | rtx res; | |
1871 | machine_mode mode = GET_MODE (reg); | |
1872 | ||
1873 | switch (mode) | |
1874 | { | |
1875 | case E_BImode: | |
1876 | { | |
1877 | rtx tmp = gen_reg_rtx (SImode); | |
1878 | ||
1879 | start_sequence (); | |
1880 | if (pm & PM_read) | |
1881 | emit_insn (gen_sel_truesi (tmp, reg, GEN_INT (1), const0_rtx)); | |
1882 | emit_insn (nvptx_gen_shared_bcast (tmp, pm, rep, data, vector)); | |
1883 | if (pm & PM_write) | |
1884 | emit_insn (gen_rtx_SET (reg, gen_rtx_NE (BImode, tmp, const0_rtx))); | |
1885 | res = get_insns (); | |
1886 | end_sequence (); | |
1887 | } | |
1888 | break; | |
1889 | ||
1890 | default: | |
1891 | { | |
1892 | rtx addr = data->ptr; | |
1893 | ||
1894 | if (!addr) | |
1895 | { | |
1896 | unsigned align = GET_MODE_ALIGNMENT (mode) / BITS_PER_UNIT; | |
1897 | ||
1898 | oacc_bcast_align = MAX (oacc_bcast_align, align); | |
1899 | data->offset = ROUND_UP (data->offset, align); | |
1900 | addr = data->base; | |
1901 | gcc_assert (data->base != NULL); | |
1902 | if (data->offset) | |
1903 | addr = gen_rtx_PLUS (Pmode, addr, GEN_INT (data->offset)); | |
1904 | } | |
1905 | ||
1906 | addr = gen_rtx_MEM (mode, addr); | |
1907 | if (pm == PM_read) | |
1908 | res = gen_rtx_SET (addr, reg); | |
1909 | else if (pm == PM_write) | |
1910 | res = gen_rtx_SET (reg, addr); | |
1911 | else | |
1912 | gcc_unreachable (); | |
1913 | ||
1914 | if (data->ptr) | |
1915 | { | |
1916 | /* We're using a ptr, increment it. */ | |
1917 | start_sequence (); | |
1918 | ||
1919 | emit_insn (res); | |
1920 | emit_insn (gen_adddi3 (data->ptr, data->ptr, | |
1921 | GEN_INT (GET_MODE_SIZE (GET_MODE (reg))))); | |
1922 | res = get_insns (); | |
1923 | end_sequence (); | |
1924 | } | |
1925 | else | |
1926 | rep = 1; | |
1927 | data->offset += rep * GET_MODE_SIZE (GET_MODE (reg)); | |
1928 | } | |
1929 | break; | |
1930 | } | |
1931 | return res; | |
1932 | } | |
1933 | \f | |
1934 | /* Returns true if X is a valid address for use in a memory reference. */ | |
1935 | ||
1936 | static bool | |
1937 | nvptx_legitimate_address_p (machine_mode, rtx x, bool) | |
1938 | { | |
1939 | enum rtx_code code = GET_CODE (x); | |
1940 | ||
1941 | switch (code) | |
1942 | { | |
1943 | case REG: | |
1944 | return true; | |
1945 | ||
1946 | case PLUS: | |
1947 | if (REG_P (XEXP (x, 0)) && CONST_INT_P (XEXP (x, 1))) | |
1948 | return true; | |
1949 | return false; | |
1950 | ||
1951 | case CONST: | |
1952 | case SYMBOL_REF: | |
1953 | case LABEL_REF: | |
1954 | return true; | |
1955 | ||
1956 | default: | |
1957 | return false; | |
1958 | } | |
1959 | } | |
1960 | \f | |
1961 | /* Machinery to output constant initializers. When beginning an | |
1962 | initializer, we decide on a fragment size (which is visible in ptx | |
1963 | in the type used), and then all initializer data is buffered until | |
1964 | a fragment is filled and ready to be written out. */ | |
1965 | ||
1966 | static struct | |
1967 | { | |
1968 | unsigned HOST_WIDE_INT mask; /* Mask for storing fragment. */ | |
1969 | unsigned HOST_WIDE_INT val; /* Current fragment value. */ | |
1970 | unsigned HOST_WIDE_INT remaining; /* Remaining bytes to be written | |
1971 | out. */ | |
1972 | unsigned size; /* Fragment size to accumulate. */ | |
1973 | unsigned offset; /* Offset within current fragment. */ | |
1974 | bool started; /* Whether we've output any initializer. */ | |
1975 | } init_frag; | |
1976 | ||
1977 | /* The current fragment is full, write it out. SYM may provide a | |
1978 | symbolic reference we should output, in which case the fragment | |
1979 | value is the addend. */ | |
1980 | ||
1981 | static void | |
1982 | output_init_frag (rtx sym) | |
1983 | { | |
1984 | fprintf (asm_out_file, init_frag.started ? ", " : " = { "); | |
1985 | unsigned HOST_WIDE_INT val = init_frag.val; | |
1986 | ||
1987 | init_frag.started = true; | |
1988 | init_frag.val = 0; | |
1989 | init_frag.offset = 0; | |
1990 | init_frag.remaining--; | |
1991 | ||
1992 | if (sym) | |
1993 | { | |
1994 | bool function = (SYMBOL_REF_DECL (sym) | |
1995 | && (TREE_CODE (SYMBOL_REF_DECL (sym)) == FUNCTION_DECL)); | |
1996 | if (!function) | |
1997 | fprintf (asm_out_file, "generic("); | |
1998 | output_address (VOIDmode, sym); | |
1999 | if (!function) | |
2000 | fprintf (asm_out_file, ")"); | |
2001 | if (val) | |
2002 | fprintf (asm_out_file, " + "); | |
2003 | } | |
2004 | ||
2005 | if (!sym || val) | |
2006 | fprintf (asm_out_file, HOST_WIDE_INT_PRINT_DEC, val); | |
2007 | } | |
2008 | ||
2009 | /* Add value VAL of size SIZE to the data we're emitting, and keep | |
2010 | writing out chunks as they fill up. */ | |
2011 | ||
2012 | static void | |
2013 | nvptx_assemble_value (unsigned HOST_WIDE_INT val, unsigned size) | |
2014 | { | |
2015 | val &= ((unsigned HOST_WIDE_INT)2 << (size * BITS_PER_UNIT - 1)) - 1; | |
2016 | ||
2017 | for (unsigned part = 0; size; size -= part) | |
2018 | { | |
2019 | val >>= part * BITS_PER_UNIT; | |
2020 | part = init_frag.size - init_frag.offset; | |
2021 | part = MIN (part, size); | |
2022 | ||
2023 | unsigned HOST_WIDE_INT partial | |
2024 | = val << (init_frag.offset * BITS_PER_UNIT); | |
2025 | init_frag.val |= partial & init_frag.mask; | |
2026 | init_frag.offset += part; | |
2027 | ||
2028 | if (init_frag.offset == init_frag.size) | |
2029 | output_init_frag (NULL); | |
2030 | } | |
2031 | } | |
2032 | ||
2033 | /* Target hook for assembling integer object X of size SIZE. */ | |
2034 | ||
2035 | static bool | |
2036 | nvptx_assemble_integer (rtx x, unsigned int size, int ARG_UNUSED (aligned_p)) | |
2037 | { | |
2038 | HOST_WIDE_INT val = 0; | |
2039 | ||
2040 | switch (GET_CODE (x)) | |
2041 | { | |
2042 | default: | |
2043 | /* Let the generic machinery figure it out, usually for a | |
2044 | CONST_WIDE_INT. */ | |
2045 | return false; | |
2046 | ||
2047 | case CONST_INT: | |
2048 | nvptx_assemble_value (INTVAL (x), size); | |
2049 | break; | |
2050 | ||
2051 | case CONST: | |
2052 | x = XEXP (x, 0); | |
2053 | gcc_assert (GET_CODE (x) == PLUS); | |
2054 | val = INTVAL (XEXP (x, 1)); | |
2055 | x = XEXP (x, 0); | |
2056 | gcc_assert (GET_CODE (x) == SYMBOL_REF); | |
2057 | /* FALLTHROUGH */ | |
2058 | ||
2059 | case SYMBOL_REF: | |
2060 | gcc_assert (size == init_frag.size); | |
2061 | if (init_frag.offset) | |
2062 | sorry ("cannot emit unaligned pointers in ptx assembly"); | |
2063 | ||
2064 | nvptx_maybe_record_fnsym (x); | |
2065 | init_frag.val = val; | |
2066 | output_init_frag (x); | |
2067 | break; | |
2068 | } | |
2069 | ||
2070 | return true; | |
2071 | } | |
2072 | ||
2073 | /* Output SIZE zero bytes. We ignore the FILE argument since the | |
2074 | functions we're calling to perform the output just use | |
2075 | asm_out_file. */ | |
2076 | ||
2077 | void | |
2078 | nvptx_output_skip (FILE *, unsigned HOST_WIDE_INT size) | |
2079 | { | |
2080 | /* Finish the current fragment, if it's started. */ | |
2081 | if (init_frag.offset) | |
2082 | { | |
2083 | unsigned part = init_frag.size - init_frag.offset; | |
2084 | part = MIN (part, (unsigned)size); | |
2085 | size -= part; | |
2086 | nvptx_assemble_value (0, part); | |
2087 | } | |
2088 | ||
2089 | /* If this skip doesn't terminate the initializer, write as many | |
2090 | remaining pieces as possible directly. */ | |
2091 | if (size < init_frag.remaining * init_frag.size) | |
2092 | { | |
2093 | while (size >= init_frag.size) | |
2094 | { | |
2095 | size -= init_frag.size; | |
2096 | output_init_frag (NULL_RTX); | |
2097 | } | |
2098 | if (size) | |
2099 | nvptx_assemble_value (0, size); | |
2100 | } | |
2101 | } | |
2102 | ||
2103 | /* Output a string STR with length SIZE. As in nvptx_output_skip we | |
2104 | ignore the FILE arg. */ | |
2105 | ||
2106 | void | |
2107 | nvptx_output_ascii (FILE *, const char *str, unsigned HOST_WIDE_INT size) | |
2108 | { | |
2109 | for (unsigned HOST_WIDE_INT i = 0; i < size; i++) | |
2110 | nvptx_assemble_value (str[i], 1); | |
2111 | } | |
2112 | ||
2113 | /* Return true if TYPE is a record type where the last field is an array without | |
2114 | given dimension. */ | |
2115 | ||
2116 | static bool | |
2117 | flexible_array_member_type_p (const_tree type) | |
2118 | { | |
2119 | if (TREE_CODE (type) != RECORD_TYPE) | |
2120 | return false; | |
2121 | ||
2122 | const_tree last_field = NULL_TREE; | |
2123 | for (const_tree f = TYPE_FIELDS (type); f; f = TREE_CHAIN (f)) | |
2124 | last_field = f; | |
2125 | ||
2126 | if (!last_field) | |
2127 | return false; | |
2128 | ||
2129 | const_tree last_field_type = TREE_TYPE (last_field); | |
2130 | if (TREE_CODE (last_field_type) != ARRAY_TYPE) | |
2131 | return false; | |
2132 | ||
2133 | return (! TYPE_DOMAIN (last_field_type) | |
2134 | || ! TYPE_MAX_VALUE (TYPE_DOMAIN (last_field_type))); | |
2135 | } | |
2136 | ||
2137 | /* Emit a PTX variable decl and prepare for emission of its | |
2138 | initializer. NAME is the symbol name and SETION the PTX data | |
2139 | area. The type is TYPE, object size SIZE and alignment is ALIGN. | |
2140 | The caller has already emitted any indentation and linkage | |
2141 | specifier. It is responsible for any initializer, terminating ; | |
2142 | and newline. SIZE is in bytes, ALIGN is in bits -- confusingly | |
2143 | this is the opposite way round that PTX wants them! */ | |
2144 | ||
2145 | static void | |
2146 | nvptx_assemble_decl_begin (FILE *file, const char *name, const char *section, | |
2147 | const_tree type, HOST_WIDE_INT size, unsigned align, | |
2148 | bool undefined = false) | |
2149 | { | |
2150 | bool atype = (TREE_CODE (type) == ARRAY_TYPE) | |
2151 | && (TYPE_DOMAIN (type) == NULL_TREE); | |
2152 | ||
2153 | if (undefined && flexible_array_member_type_p (type)) | |
2154 | { | |
2155 | size = 0; | |
2156 | atype = true; | |
2157 | } | |
2158 | ||
2159 | while (TREE_CODE (type) == ARRAY_TYPE) | |
2160 | type = TREE_TYPE (type); | |
2161 | ||
2162 | if (TREE_CODE (type) == VECTOR_TYPE | |
2163 | || TREE_CODE (type) == COMPLEX_TYPE) | |
2164 | /* Neither vector nor complex types can contain the other. */ | |
2165 | type = TREE_TYPE (type); | |
2166 | ||
2167 | unsigned elt_size = int_size_in_bytes (type); | |
2168 | ||
2169 | /* Largest mode we're prepared to accept. For BLKmode types we | |
2170 | don't know if it'll contain pointer constants, so have to choose | |
2171 | pointer size, otherwise we can choose DImode. */ | |
2172 | machine_mode elt_mode = TYPE_MODE (type) == BLKmode ? Pmode : DImode; | |
2173 | ||
2174 | elt_size |= GET_MODE_SIZE (elt_mode); | |
2175 | elt_size &= -elt_size; /* Extract LSB set. */ | |
2176 | ||
2177 | init_frag.size = elt_size; | |
2178 | /* Avoid undefined shift behavior by using '2'. */ | |
2179 | init_frag.mask = ((unsigned HOST_WIDE_INT)2 | |
2180 | << (elt_size * BITS_PER_UNIT - 1)) - 1; | |
2181 | init_frag.val = 0; | |
2182 | init_frag.offset = 0; | |
2183 | init_frag.started = false; | |
2184 | /* Size might not be a multiple of elt size, if there's an | |
2185 | initialized trailing struct array with smaller type than | |
2186 | elt_size. */ | |
2187 | init_frag.remaining = (size + elt_size - 1) / elt_size; | |
2188 | ||
2189 | fprintf (file, "%s .align %d .u%d ", | |
2190 | section, align / BITS_PER_UNIT, | |
2191 | elt_size * BITS_PER_UNIT); | |
2192 | assemble_name (file, name); | |
2193 | ||
2194 | if (size) | |
2195 | /* We make everything an array, to simplify any initialization | |
2196 | emission. */ | |
2197 | fprintf (file, "[" HOST_WIDE_INT_PRINT_DEC "]", init_frag.remaining); | |
2198 | else if (atype) | |
2199 | fprintf (file, "[]"); | |
2200 | } | |
2201 | ||
2202 | /* Called when the initializer for a decl has been completely output through | |
2203 | combinations of the three functions above. */ | |
2204 | ||
2205 | static void | |
2206 | nvptx_assemble_decl_end (void) | |
2207 | { | |
2208 | if (init_frag.offset) | |
2209 | /* This can happen with a packed struct with trailing array member. */ | |
2210 | nvptx_assemble_value (0, init_frag.size - init_frag.offset); | |
2211 | fprintf (asm_out_file, init_frag.started ? " };\n" : ";\n"); | |
2212 | } | |
2213 | ||
2214 | /* Output an uninitialized common or file-scope variable. */ | |
2215 | ||
2216 | void | |
2217 | nvptx_output_aligned_decl (FILE *file, const char *name, | |
2218 | const_tree decl, HOST_WIDE_INT size, unsigned align) | |
2219 | { | |
2220 | write_var_marker (file, true, TREE_PUBLIC (decl), name); | |
2221 | ||
2222 | /* If this is public, it is common. The nearest thing we have to | |
2223 | common is weak. */ | |
2224 | fprintf (file, "\t%s", TREE_PUBLIC (decl) ? ".weak " : ""); | |
2225 | ||
2226 | nvptx_assemble_decl_begin (file, name, section_for_decl (decl), | |
2227 | TREE_TYPE (decl), size, align); | |
2228 | nvptx_assemble_decl_end (); | |
2229 | } | |
2230 | ||
2231 | /* Implement TARGET_ASM_DECLARE_CONSTANT_NAME. Begin the process of | |
2232 | writing a constant variable EXP with NAME and SIZE and its | |
2233 | initializer to FILE. */ | |
2234 | ||
2235 | static void | |
2236 | nvptx_asm_declare_constant_name (FILE *file, const char *name, | |
2237 | const_tree exp, HOST_WIDE_INT obj_size) | |
2238 | { | |
2239 | write_var_marker (file, true, false, name); | |
2240 | ||
2241 | fprintf (file, "\t"); | |
2242 | ||
2243 | tree type = TREE_TYPE (exp); | |
2244 | nvptx_assemble_decl_begin (file, name, ".const", type, obj_size, | |
2245 | TYPE_ALIGN (type)); | |
2246 | } | |
2247 | ||
2248 | /* Implement the ASM_DECLARE_OBJECT_NAME macro. Used to start writing | |
2249 | a variable DECL with NAME to FILE. */ | |
2250 | ||
2251 | void | |
2252 | nvptx_declare_object_name (FILE *file, const char *name, const_tree decl) | |
2253 | { | |
2254 | write_var_marker (file, true, TREE_PUBLIC (decl), name); | |
2255 | ||
2256 | fprintf (file, "\t%s", (!TREE_PUBLIC (decl) ? "" | |
2257 | : DECL_WEAK (decl) ? ".weak " : ".visible ")); | |
2258 | ||
2259 | tree type = TREE_TYPE (decl); | |
2260 | HOST_WIDE_INT obj_size = tree_to_shwi (DECL_SIZE_UNIT (decl)); | |
2261 | nvptx_assemble_decl_begin (file, name, section_for_decl (decl), | |
2262 | type, obj_size, DECL_ALIGN (decl)); | |
2263 | } | |
2264 | ||
2265 | /* Implement TARGET_ASM_GLOBALIZE_LABEL by doing nothing. */ | |
2266 | ||
2267 | static void | |
2268 | nvptx_globalize_label (FILE *, const char *) | |
2269 | { | |
2270 | } | |
2271 | ||
2272 | /* Implement TARGET_ASM_ASSEMBLE_UNDEFINED_DECL. Write an extern | |
2273 | declaration only for variable DECL with NAME to FILE. */ | |
2274 | ||
2275 | static void | |
2276 | nvptx_assemble_undefined_decl (FILE *file, const char *name, const_tree decl) | |
2277 | { | |
2278 | /* The middle end can place constant pool decls into the varpool as | |
2279 | undefined. Until that is fixed, catch the problem here. */ | |
2280 | if (DECL_IN_CONSTANT_POOL (decl)) | |
2281 | return; | |
2282 | ||
2283 | /* We support weak defintions, and hence have the right | |
2284 | ASM_WEAKEN_DECL definition. Diagnose the problem here. */ | |
2285 | if (DECL_WEAK (decl)) | |
2286 | error_at (DECL_SOURCE_LOCATION (decl), | |
2287 | "PTX does not support weak declarations" | |
2288 | " (only weak definitions)"); | |
2289 | write_var_marker (file, false, TREE_PUBLIC (decl), name); | |
2290 | ||
2291 | fprintf (file, "\t.extern "); | |
2292 | tree size = DECL_SIZE_UNIT (decl); | |
2293 | nvptx_assemble_decl_begin (file, name, section_for_decl (decl), | |
2294 | TREE_TYPE (decl), size ? tree_to_shwi (size) : 0, | |
2295 | DECL_ALIGN (decl), true); | |
2296 | nvptx_assemble_decl_end (); | |
2297 | } | |
2298 | ||
2299 | /* Output a pattern for a move instruction. */ | |
2300 | ||
2301 | const char * | |
2302 | nvptx_output_mov_insn (rtx dst, rtx src) | |
2303 | { | |
2304 | machine_mode dst_mode = GET_MODE (dst); | |
2305 | machine_mode dst_inner = (GET_CODE (dst) == SUBREG | |
2306 | ? GET_MODE (XEXP (dst, 0)) : dst_mode); | |
2307 | machine_mode src_inner = (GET_CODE (src) == SUBREG | |
2308 | ? GET_MODE (XEXP (src, 0)) : dst_mode); | |
2309 | ||
2310 | rtx sym = src; | |
2311 | if (GET_CODE (sym) == CONST) | |
2312 | sym = XEXP (XEXP (sym, 0), 0); | |
2313 | if (SYMBOL_REF_P (sym)) | |
2314 | { | |
2315 | if (SYMBOL_DATA_AREA (sym) != DATA_AREA_GENERIC) | |
2316 | return "%.\tcvta%D1%t0\t%0, %1;"; | |
2317 | nvptx_maybe_record_fnsym (sym); | |
2318 | } | |
2319 | ||
2320 | if (src_inner == dst_inner) | |
2321 | return "%.\tmov%t0\t%0, %1;"; | |
2322 | ||
2323 | if (CONSTANT_P (src)) | |
2324 | return (GET_MODE_CLASS (dst_inner) == MODE_INT | |
2325 | && GET_MODE_CLASS (src_inner) != MODE_FLOAT | |
2326 | ? "%.\tmov%t0\t%0, %1;" : "%.\tmov.b%T0\t%0, %1;"); | |
2327 | ||
2328 | if (GET_MODE_SIZE (dst_inner) == GET_MODE_SIZE (src_inner)) | |
2329 | { | |
2330 | if (GET_MODE_BITSIZE (dst_mode) == 128 | |
2331 | && GET_MODE_BITSIZE (GET_MODE (src)) == 128) | |
2332 | { | |
2333 | /* mov.b128 is not supported. */ | |
2334 | if (dst_inner == V2DImode && src_inner == TImode) | |
2335 | return "%.\tmov.u64\t%0.x, %L1;\n\t%.\tmov.u64\t%0.y, %H1;"; | |
2336 | else if (dst_inner == TImode && src_inner == V2DImode) | |
2337 | return "%.\tmov.u64\t%L0, %1.x;\n\t%.\tmov.u64\t%H0, %1.y;"; | |
2338 | ||
2339 | gcc_unreachable (); | |
2340 | } | |
2341 | return "%.\tmov.b%T0\t%0, %1;"; | |
2342 | } | |
2343 | ||
2344 | return "%.\tcvt%t0%t1\t%0, %1;"; | |
2345 | } | |
2346 | ||
2347 | static void nvptx_print_operand (FILE *, rtx, int); | |
2348 | ||
2349 | /* Output INSN, which is a call to CALLEE with result RESULT. For ptx, this | |
2350 | involves writing .param declarations and in/out copies into them. For | |
2351 | indirect calls, also write the .callprototype. */ | |
2352 | ||
2353 | const char * | |
2354 | nvptx_output_call_insn (rtx_insn *insn, rtx result, rtx callee) | |
2355 | { | |
2356 | char buf[16]; | |
2357 | static int labelno; | |
2358 | bool needs_tgt = register_operand (callee, Pmode); | |
2359 | rtx pat = PATTERN (insn); | |
2360 | if (GET_CODE (pat) == COND_EXEC) | |
2361 | pat = COND_EXEC_CODE (pat); | |
2362 | int arg_end = XVECLEN (pat, 0); | |
2363 | tree decl = NULL_TREE; | |
2364 | ||
2365 | fprintf (asm_out_file, "\t{\n"); | |
2366 | if (result != NULL) | |
2367 | fprintf (asm_out_file, "\t\t.param%s %s_in;\n", | |
2368 | nvptx_ptx_type_from_mode (GET_MODE (result), false), | |
2369 | reg_names[NVPTX_RETURN_REGNUM]); | |
2370 | ||
2371 | /* Ensure we have a ptx declaration in the output if necessary. */ | |
2372 | if (GET_CODE (callee) == SYMBOL_REF) | |
2373 | { | |
2374 | decl = SYMBOL_REF_DECL (callee); | |
2375 | if (!decl | |
2376 | || (DECL_EXTERNAL (decl) && !TYPE_ARG_TYPES (TREE_TYPE (decl)))) | |
2377 | nvptx_record_libfunc (callee, result, pat); | |
2378 | else if (DECL_EXTERNAL (decl)) | |
2379 | nvptx_record_fndecl (decl); | |
2380 | } | |
2381 | ||
2382 | if (needs_tgt) | |
2383 | { | |
2384 | ASM_GENERATE_INTERNAL_LABEL (buf, "LCT", labelno); | |
2385 | labelno++; | |
2386 | ASM_OUTPUT_LABEL (asm_out_file, buf); | |
2387 | std::stringstream s; | |
2388 | write_fn_proto_from_insn (s, NULL, result, pat); | |
2389 | fputs (s.str().c_str(), asm_out_file); | |
2390 | } | |
2391 | ||
2392 | for (int argno = 1; argno < arg_end; argno++) | |
2393 | { | |
2394 | rtx t = XEXP (XVECEXP (pat, 0, argno), 0); | |
2395 | machine_mode mode = GET_MODE (t); | |
2396 | const char *ptx_type = nvptx_ptx_type_from_mode (mode, false); | |
2397 | ||
2398 | /* Mode splitting has already been done. */ | |
2399 | fprintf (asm_out_file, "\t\t.param%s %%out_arg%d;\n" | |
2400 | "\t\tst.param%s [%%out_arg%d], ", | |
2401 | ptx_type, argno, ptx_type, argno); | |
2402 | output_reg (asm_out_file, REGNO (t), VOIDmode); | |
2403 | fprintf (asm_out_file, ";\n"); | |
2404 | } | |
2405 | ||
2406 | /* The '.' stands for the call's predicate, if any. */ | |
2407 | nvptx_print_operand (asm_out_file, NULL_RTX, '.'); | |
2408 | fprintf (asm_out_file, "\t\tcall "); | |
2409 | if (result != NULL_RTX) | |
2410 | fprintf (asm_out_file, "(%s_in), ", reg_names[NVPTX_RETURN_REGNUM]); | |
2411 | ||
2412 | if (decl) | |
2413 | { | |
2414 | const char *name = get_fnname_from_decl (decl); | |
2415 | name = nvptx_name_replacement (name); | |
2416 | assemble_name (asm_out_file, name); | |
2417 | } | |
2418 | else | |
2419 | output_address (VOIDmode, callee); | |
2420 | ||
2421 | const char *open = "("; | |
2422 | for (int argno = 1; argno < arg_end; argno++) | |
2423 | { | |
2424 | fprintf (asm_out_file, ", %s%%out_arg%d", open, argno); | |
2425 | open = ""; | |
2426 | } | |
2427 | if (decl && DECL_STATIC_CHAIN (decl)) | |
2428 | { | |
2429 | fprintf (asm_out_file, ", %s%s", open, reg_names [STATIC_CHAIN_REGNUM]); | |
2430 | open = ""; | |
2431 | } | |
2432 | if (!open[0]) | |
2433 | fprintf (asm_out_file, ")"); | |
2434 | ||
2435 | if (needs_tgt) | |
2436 | { | |
2437 | fprintf (asm_out_file, ", "); | |
2438 | assemble_name (asm_out_file, buf); | |
2439 | } | |
2440 | fprintf (asm_out_file, ";\n"); | |
2441 | ||
2442 | if (find_reg_note (insn, REG_NORETURN, NULL)) | |
2443 | { | |
2444 | /* No return functions confuse the PTX JIT, as it doesn't realize | |
2445 | the flow control barrier they imply. It can seg fault if it | |
2446 | encounters what looks like an unexitable loop. Emit a trailing | |
2447 | trap and exit, which it does grok. */ | |
2448 | fprintf (asm_out_file, "\t\ttrap; // (noreturn)\n"); | |
2449 | fprintf (asm_out_file, "\t\texit; // (noreturn)\n"); | |
2450 | } | |
2451 | ||
2452 | if (result) | |
2453 | { | |
2454 | static char rval[sizeof ("\tld.param%%t0\t%%0, [%%%s_in];\n\t}") + 8]; | |
2455 | ||
2456 | if (!rval[0]) | |
2457 | /* We must escape the '%' that starts RETURN_REGNUM. */ | |
2458 | sprintf (rval, "\tld.param%%t0\t%%0, [%%%s_in];\n\t}", | |
2459 | reg_names[NVPTX_RETURN_REGNUM]); | |
2460 | return rval; | |
2461 | } | |
2462 | ||
2463 | return "}"; | |
2464 | } | |
2465 | ||
2466 | /* Implement TARGET_PRINT_OPERAND_PUNCT_VALID_P. */ | |
2467 | ||
2468 | static bool | |
2469 | nvptx_print_operand_punct_valid_p (unsigned char c) | |
2470 | { | |
2471 | return c == '.' || c== '#'; | |
2472 | } | |
2473 | ||
2474 | /* Subroutine of nvptx_print_operand; used to print a memory reference X to FILE. */ | |
2475 | ||
2476 | static void | |
2477 | nvptx_print_address_operand (FILE *file, rtx x, machine_mode) | |
2478 | { | |
2479 | rtx off; | |
2480 | if (GET_CODE (x) == CONST) | |
2481 | x = XEXP (x, 0); | |
2482 | switch (GET_CODE (x)) | |
2483 | { | |
2484 | case PLUS: | |
2485 | off = XEXP (x, 1); | |
2486 | output_address (VOIDmode, XEXP (x, 0)); | |
2487 | fprintf (file, "+"); | |
2488 | output_address (VOIDmode, off); | |
2489 | break; | |
2490 | ||
2491 | case SYMBOL_REF: | |
2492 | case LABEL_REF: | |
2493 | output_addr_const (file, x); | |
2494 | break; | |
2495 | ||
2496 | default: | |
2497 | gcc_assert (GET_CODE (x) != MEM); | |
2498 | nvptx_print_operand (file, x, 0); | |
2499 | break; | |
2500 | } | |
2501 | } | |
2502 | ||
2503 | /* Write assembly language output for the address ADDR to FILE. */ | |
2504 | ||
2505 | static void | |
2506 | nvptx_print_operand_address (FILE *file, machine_mode mode, rtx addr) | |
2507 | { | |
2508 | nvptx_print_address_operand (file, addr, mode); | |
2509 | } | |
2510 | ||
2511 | /* Print an operand, X, to FILE, with an optional modifier in CODE. | |
2512 | ||
2513 | Meaning of CODE: | |
2514 | . -- print the predicate for the instruction or an emptry string for an | |
2515 | unconditional one. | |
2516 | # -- print a rounding mode for the instruction | |
2517 | ||
2518 | A -- print a data area for a MEM | |
2519 | c -- print an opcode suffix for a comparison operator, including a type code | |
2520 | D -- print a data area for a MEM operand | |
2521 | S -- print a shuffle kind specified by CONST_INT | |
2522 | t -- print a type opcode suffix, promoting QImode to 32 bits | |
2523 | T -- print a type size in bits | |
2524 | u -- print a type opcode suffix without promotions. */ | |
2525 | ||
2526 | static void | |
2527 | nvptx_print_operand (FILE *file, rtx x, int code) | |
2528 | { | |
2529 | if (code == '.') | |
2530 | { | |
2531 | x = current_insn_predicate; | |
2532 | if (x) | |
2533 | { | |
2534 | fputs ("@", file); | |
2535 | if (GET_CODE (x) == EQ) | |
2536 | fputs ("!", file); | |
2537 | output_reg (file, REGNO (XEXP (x, 0)), VOIDmode); | |
2538 | } | |
2539 | return; | |
2540 | } | |
2541 | else if (code == '#') | |
2542 | { | |
2543 | fputs (".rn", file); | |
2544 | return; | |
2545 | } | |
2546 | ||
2547 | enum rtx_code x_code = GET_CODE (x); | |
2548 | machine_mode mode = GET_MODE (x); | |
2549 | ||
2550 | switch (code) | |
2551 | { | |
2552 | case 'A': | |
2553 | x = XEXP (x, 0); | |
2554 | /* FALLTHROUGH. */ | |
2555 | ||
2556 | case 'D': | |
2557 | if (GET_CODE (x) == CONST) | |
2558 | x = XEXP (x, 0); | |
2559 | if (GET_CODE (x) == PLUS) | |
2560 | x = XEXP (x, 0); | |
2561 | ||
2562 | if (GET_CODE (x) == SYMBOL_REF) | |
2563 | fputs (section_for_sym (x), file); | |
2564 | break; | |
2565 | ||
2566 | case 't': | |
2567 | case 'u': | |
2568 | if (x_code == SUBREG) | |
2569 | { | |
2570 | machine_mode inner_mode = GET_MODE (SUBREG_REG (x)); | |
2571 | if (VECTOR_MODE_P (inner_mode) | |
2572 | && (GET_MODE_SIZE (mode) | |
2573 | <= GET_MODE_SIZE (GET_MODE_INNER (inner_mode)))) | |
2574 | mode = GET_MODE_INNER (inner_mode); | |
2575 | else if (split_mode_p (inner_mode)) | |
2576 | mode = maybe_split_mode (inner_mode); | |
2577 | else | |
2578 | mode = inner_mode; | |
2579 | } | |
2580 | fprintf (file, "%s", nvptx_ptx_type_from_mode (mode, code == 't')); | |
2581 | break; | |
2582 | ||
2583 | case 'H': | |
2584 | case 'L': | |
2585 | { | |
2586 | rtx inner_x = SUBREG_REG (x); | |
2587 | machine_mode inner_mode = GET_MODE (inner_x); | |
2588 | machine_mode split = maybe_split_mode (inner_mode); | |
2589 | ||
2590 | output_reg (file, REGNO (inner_x), split, | |
2591 | (code == 'H' | |
2592 | ? GET_MODE_SIZE (inner_mode) / 2 | |
2593 | : 0)); | |
2594 | } | |
2595 | break; | |
2596 | ||
2597 | case 'S': | |
2598 | { | |
2599 | nvptx_shuffle_kind kind = (nvptx_shuffle_kind) UINTVAL (x); | |
2600 | /* Same order as nvptx_shuffle_kind. */ | |
2601 | static const char *const kinds[] = | |
2602 | {".up", ".down", ".bfly", ".idx"}; | |
2603 | fputs (kinds[kind], file); | |
2604 | } | |
2605 | break; | |
2606 | ||
2607 | case 'T': | |
2608 | fprintf (file, "%d", GET_MODE_BITSIZE (mode)); | |
2609 | break; | |
2610 | ||
2611 | case 'j': | |
2612 | fprintf (file, "@"); | |
2613 | goto common; | |
2614 | ||
2615 | case 'J': | |
2616 | fprintf (file, "@!"); | |
2617 | goto common; | |
2618 | ||
2619 | case 'c': | |
2620 | mode = GET_MODE (XEXP (x, 0)); | |
2621 | switch (x_code) | |
2622 | { | |
2623 | case EQ: | |
2624 | fputs (".eq", file); | |
2625 | break; | |
2626 | case NE: | |
2627 | if (FLOAT_MODE_P (mode)) | |
2628 | fputs (".neu", file); | |
2629 | else | |
2630 | fputs (".ne", file); | |
2631 | break; | |
2632 | case LE: | |
2633 | case LEU: | |
2634 | fputs (".le", file); | |
2635 | break; | |
2636 | case GE: | |
2637 | case GEU: | |
2638 | fputs (".ge", file); | |
2639 | break; | |
2640 | case LT: | |
2641 | case LTU: | |
2642 | fputs (".lt", file); | |
2643 | break; | |
2644 | case GT: | |
2645 | case GTU: | |
2646 | fputs (".gt", file); | |
2647 | break; | |
2648 | case LTGT: | |
2649 | fputs (".ne", file); | |
2650 | break; | |
2651 | case UNEQ: | |
2652 | fputs (".equ", file); | |
2653 | break; | |
2654 | case UNLE: | |
2655 | fputs (".leu", file); | |
2656 | break; | |
2657 | case UNGE: | |
2658 | fputs (".geu", file); | |
2659 | break; | |
2660 | case UNLT: | |
2661 | fputs (".ltu", file); | |
2662 | break; | |
2663 | case UNGT: | |
2664 | fputs (".gtu", file); | |
2665 | break; | |
2666 | case UNORDERED: | |
2667 | fputs (".nan", file); | |
2668 | break; | |
2669 | case ORDERED: | |
2670 | fputs (".num", file); | |
2671 | break; | |
2672 | default: | |
2673 | gcc_unreachable (); | |
2674 | } | |
2675 | if (FLOAT_MODE_P (mode) | |
2676 | || x_code == EQ || x_code == NE | |
2677 | || x_code == GEU || x_code == GTU | |
2678 | || x_code == LEU || x_code == LTU) | |
2679 | fputs (nvptx_ptx_type_from_mode (mode, true), file); | |
2680 | else | |
2681 | fprintf (file, ".s%d", GET_MODE_BITSIZE (mode)); | |
2682 | break; | |
2683 | default: | |
2684 | common: | |
2685 | switch (x_code) | |
2686 | { | |
2687 | case SUBREG: | |
2688 | { | |
2689 | rtx inner_x = SUBREG_REG (x); | |
2690 | machine_mode inner_mode = GET_MODE (inner_x); | |
2691 | machine_mode split = maybe_split_mode (inner_mode); | |
2692 | ||
2693 | if (VECTOR_MODE_P (inner_mode) | |
2694 | && (GET_MODE_SIZE (mode) | |
2695 | <= GET_MODE_SIZE (GET_MODE_INNER (inner_mode)))) | |
2696 | { | |
2697 | output_reg (file, REGNO (inner_x), VOIDmode); | |
2698 | fprintf (file, ".%s", SUBREG_BYTE (x) == 0 ? "x" : "y"); | |
2699 | } | |
2700 | else if (split_mode_p (inner_mode) | |
2701 | && (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (mode))) | |
2702 | output_reg (file, REGNO (inner_x), split); | |
2703 | else | |
2704 | output_reg (file, REGNO (inner_x), split, SUBREG_BYTE (x)); | |
2705 | } | |
2706 | break; | |
2707 | ||
2708 | case REG: | |
2709 | output_reg (file, REGNO (x), maybe_split_mode (mode)); | |
2710 | break; | |
2711 | ||
2712 | case MEM: | |
2713 | fputc ('[', file); | |
2714 | nvptx_print_address_operand (file, XEXP (x, 0), mode); | |
2715 | fputc (']', file); | |
2716 | break; | |
2717 | ||
2718 | case CONST_INT: | |
2719 | output_addr_const (file, x); | |
2720 | break; | |
2721 | ||
2722 | case CONST: | |
2723 | case SYMBOL_REF: | |
2724 | case LABEL_REF: | |
2725 | /* We could use output_addr_const, but that can print things like | |
2726 | "x-8", which breaks ptxas. Need to ensure it is output as | |
2727 | "x+-8". */ | |
2728 | nvptx_print_address_operand (file, x, VOIDmode); | |
2729 | break; | |
2730 | ||
2731 | case CONST_DOUBLE: | |
2732 | long vals[2]; | |
2733 | real_to_target (vals, CONST_DOUBLE_REAL_VALUE (x), mode); | |
2734 | vals[0] &= 0xffffffff; | |
2735 | vals[1] &= 0xffffffff; | |
2736 | if (mode == SFmode) | |
2737 | fprintf (file, "0f%08lx", vals[0]); | |
2738 | else | |
2739 | fprintf (file, "0d%08lx%08lx", vals[1], vals[0]); | |
2740 | break; | |
2741 | ||
2742 | case CONST_VECTOR: | |
2743 | { | |
2744 | unsigned n = CONST_VECTOR_NUNITS (x); | |
2745 | fprintf (file, "{ "); | |
2746 | for (unsigned i = 0; i < n; ++i) | |
2747 | { | |
2748 | if (i != 0) | |
2749 | fprintf (file, ", "); | |
2750 | ||
2751 | rtx elem = CONST_VECTOR_ELT (x, i); | |
2752 | output_addr_const (file, elem); | |
2753 | } | |
2754 | fprintf (file, " }"); | |
2755 | } | |
2756 | break; | |
2757 | ||
2758 | default: | |
2759 | output_addr_const (file, x); | |
2760 | } | |
2761 | } | |
2762 | } | |
2763 | \f | |
2764 | /* Record replacement regs used to deal with subreg operands. */ | |
2765 | struct reg_replace | |
2766 | { | |
2767 | rtx replacement[MAX_RECOG_OPERANDS]; | |
2768 | machine_mode mode; | |
2769 | int n_allocated; | |
2770 | int n_in_use; | |
2771 | }; | |
2772 | ||
2773 | /* Allocate or reuse a replacement in R and return the rtx. */ | |
2774 | ||
2775 | static rtx | |
2776 | get_replacement (struct reg_replace *r) | |
2777 | { | |
2778 | if (r->n_allocated == r->n_in_use) | |
2779 | r->replacement[r->n_allocated++] = gen_reg_rtx (r->mode); | |
2780 | return r->replacement[r->n_in_use++]; | |
2781 | } | |
2782 | ||
2783 | /* Clean up subreg operands. In ptx assembly, everything is typed, and | |
2784 | the presence of subregs would break the rules for most instructions. | |
2785 | Replace them with a suitable new register of the right size, plus | |
2786 | conversion copyin/copyout instructions. */ | |
2787 | ||
2788 | static void | |
2789 | nvptx_reorg_subreg (void) | |
2790 | { | |
2791 | struct reg_replace qiregs, hiregs, siregs, diregs; | |
2792 | rtx_insn *insn, *next; | |
2793 | ||
2794 | qiregs.n_allocated = 0; | |
2795 | hiregs.n_allocated = 0; | |
2796 | siregs.n_allocated = 0; | |
2797 | diregs.n_allocated = 0; | |
2798 | qiregs.mode = QImode; | |
2799 | hiregs.mode = HImode; | |
2800 | siregs.mode = SImode; | |
2801 | diregs.mode = DImode; | |
2802 | ||
2803 | for (insn = get_insns (); insn; insn = next) | |
2804 | { | |
2805 | next = NEXT_INSN (insn); | |
2806 | if (!NONDEBUG_INSN_P (insn) | |
2807 | || asm_noperands (PATTERN (insn)) >= 0 | |
2808 | || GET_CODE (PATTERN (insn)) == USE | |
2809 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
2810 | continue; | |
2811 | ||
2812 | qiregs.n_in_use = 0; | |
2813 | hiregs.n_in_use = 0; | |
2814 | siregs.n_in_use = 0; | |
2815 | diregs.n_in_use = 0; | |
2816 | extract_insn (insn); | |
2817 | enum attr_subregs_ok s_ok = get_attr_subregs_ok (insn); | |
2818 | ||
2819 | for (int i = 0; i < recog_data.n_operands; i++) | |
2820 | { | |
2821 | rtx op = recog_data.operand[i]; | |
2822 | if (GET_CODE (op) != SUBREG) | |
2823 | continue; | |
2824 | ||
2825 | rtx inner = SUBREG_REG (op); | |
2826 | ||
2827 | machine_mode outer_mode = GET_MODE (op); | |
2828 | machine_mode inner_mode = GET_MODE (inner); | |
2829 | gcc_assert (s_ok); | |
2830 | if (s_ok | |
2831 | && (GET_MODE_PRECISION (inner_mode) | |
2832 | >= GET_MODE_PRECISION (outer_mode))) | |
2833 | continue; | |
2834 | gcc_assert (SCALAR_INT_MODE_P (outer_mode)); | |
2835 | struct reg_replace *r = (outer_mode == QImode ? &qiregs | |
2836 | : outer_mode == HImode ? &hiregs | |
2837 | : outer_mode == SImode ? &siregs | |
2838 | : &diregs); | |
2839 | rtx new_reg = get_replacement (r); | |
2840 | ||
2841 | if (recog_data.operand_type[i] != OP_OUT) | |
2842 | { | |
2843 | enum rtx_code code; | |
2844 | if (GET_MODE_PRECISION (inner_mode) | |
2845 | < GET_MODE_PRECISION (outer_mode)) | |
2846 | code = ZERO_EXTEND; | |
2847 | else | |
2848 | code = TRUNCATE; | |
2849 | ||
2850 | rtx pat = gen_rtx_SET (new_reg, | |
2851 | gen_rtx_fmt_e (code, outer_mode, inner)); | |
2852 | emit_insn_before (pat, insn); | |
2853 | } | |
2854 | ||
2855 | if (recog_data.operand_type[i] != OP_IN) | |
2856 | { | |
2857 | enum rtx_code code; | |
2858 | if (GET_MODE_PRECISION (inner_mode) | |
2859 | < GET_MODE_PRECISION (outer_mode)) | |
2860 | code = TRUNCATE; | |
2861 | else | |
2862 | code = ZERO_EXTEND; | |
2863 | ||
2864 | rtx pat = gen_rtx_SET (inner, | |
2865 | gen_rtx_fmt_e (code, inner_mode, new_reg)); | |
2866 | emit_insn_after (pat, insn); | |
2867 | } | |
2868 | validate_change (insn, recog_data.operand_loc[i], new_reg, false); | |
2869 | } | |
2870 | } | |
2871 | } | |
2872 | ||
2873 | /* Return a SImode "master lane index" register for uniform-simt, allocating on | |
2874 | first use. */ | |
2875 | ||
2876 | static rtx | |
2877 | nvptx_get_unisimt_master () | |
2878 | { | |
2879 | rtx &master = cfun->machine->unisimt_master; | |
2880 | return master ? master : master = gen_reg_rtx (SImode); | |
2881 | } | |
2882 | ||
2883 | /* Return a BImode "predicate" register for uniform-simt, similar to above. */ | |
2884 | ||
2885 | static rtx | |
2886 | nvptx_get_unisimt_predicate () | |
2887 | { | |
2888 | rtx &pred = cfun->machine->unisimt_predicate; | |
2889 | return pred ? pred : pred = gen_reg_rtx (BImode); | |
2890 | } | |
2891 | ||
2892 | /* Return true if given call insn references one of the functions provided by | |
2893 | the CUDA runtime: malloc, free, vprintf. */ | |
2894 | ||
2895 | static bool | |
2896 | nvptx_call_insn_is_syscall_p (rtx_insn *insn) | |
2897 | { | |
2898 | rtx pat = PATTERN (insn); | |
2899 | gcc_checking_assert (GET_CODE (pat) == PARALLEL); | |
2900 | pat = XVECEXP (pat, 0, 0); | |
2901 | if (GET_CODE (pat) == SET) | |
2902 | pat = SET_SRC (pat); | |
2903 | gcc_checking_assert (GET_CODE (pat) == CALL | |
2904 | && GET_CODE (XEXP (pat, 0)) == MEM); | |
2905 | rtx addr = XEXP (XEXP (pat, 0), 0); | |
2906 | if (GET_CODE (addr) != SYMBOL_REF) | |
2907 | return false; | |
2908 | const char *name = XSTR (addr, 0); | |
2909 | /* Ordinary malloc/free are redirected to __nvptx_{malloc,free), so only the | |
2910 | references with forced assembler name refer to PTX syscalls. For vprintf, | |
2911 | accept both normal and forced-assembler-name references. */ | |
2912 | return (!strcmp (name, "vprintf") || !strcmp (name, "*vprintf") | |
2913 | || !strcmp (name, "*malloc") | |
2914 | || !strcmp (name, "*free")); | |
2915 | } | |
2916 | ||
2917 | /* If SET subexpression of INSN sets a register, emit a shuffle instruction to | |
2918 | propagate its value from lane MASTER to current lane. */ | |
2919 | ||
2920 | static void | |
2921 | nvptx_unisimt_handle_set (rtx set, rtx_insn *insn, rtx master) | |
2922 | { | |
2923 | rtx reg; | |
2924 | if (GET_CODE (set) == SET && REG_P (reg = SET_DEST (set))) | |
2925 | emit_insn_after (nvptx_gen_shuffle (reg, reg, master, SHUFFLE_IDX), insn); | |
2926 | } | |
2927 | ||
2928 | /* Adjust code for uniform-simt code generation variant by making atomics and | |
2929 | "syscalls" conditionally executed, and inserting shuffle-based propagation | |
2930 | for registers being set. */ | |
2931 | ||
2932 | static void | |
2933 | nvptx_reorg_uniform_simt () | |
2934 | { | |
2935 | rtx_insn *insn, *next; | |
2936 | ||
2937 | for (insn = get_insns (); insn; insn = next) | |
2938 | { | |
2939 | next = NEXT_INSN (insn); | |
2940 | if (!(CALL_P (insn) && nvptx_call_insn_is_syscall_p (insn)) | |
2941 | && !(NONJUMP_INSN_P (insn) | |
2942 | && GET_CODE (PATTERN (insn)) == PARALLEL | |
2943 | && get_attr_atomic (insn))) | |
2944 | continue; | |
2945 | rtx pat = PATTERN (insn); | |
2946 | rtx master = nvptx_get_unisimt_master (); | |
2947 | for (int i = 0; i < XVECLEN (pat, 0); i++) | |
2948 | nvptx_unisimt_handle_set (XVECEXP (pat, 0, i), insn, master); | |
2949 | rtx pred = nvptx_get_unisimt_predicate (); | |
2950 | pred = gen_rtx_NE (BImode, pred, const0_rtx); | |
2951 | pat = gen_rtx_COND_EXEC (VOIDmode, pred, pat); | |
2952 | validate_change (insn, &PATTERN (insn), pat, false); | |
2953 | } | |
2954 | } | |
2955 | ||
2956 | /* Offloading function attributes. */ | |
2957 | ||
2958 | struct offload_attrs | |
2959 | { | |
2960 | unsigned mask; | |
2961 | int num_gangs; | |
2962 | int num_workers; | |
2963 | int vector_length; | |
2964 | }; | |
2965 | ||
2966 | /* Define entries for cfun->machine->axis_dim. */ | |
2967 | ||
2968 | #define MACH_VECTOR_LENGTH 0 | |
2969 | #define MACH_MAX_WORKERS 1 | |
2970 | ||
2971 | static void populate_offload_attrs (offload_attrs *oa); | |
2972 | ||
2973 | static void | |
2974 | init_axis_dim (void) | |
2975 | { | |
2976 | offload_attrs oa; | |
2977 | int max_workers; | |
2978 | ||
2979 | populate_offload_attrs (&oa); | |
2980 | ||
2981 | if (oa.num_workers == 0) | |
2982 | max_workers = PTX_CTA_SIZE / oa.vector_length; | |
2983 | else | |
2984 | max_workers = oa.num_workers; | |
2985 | ||
2986 | cfun->machine->axis_dim[MACH_VECTOR_LENGTH] = oa.vector_length; | |
2987 | cfun->machine->axis_dim[MACH_MAX_WORKERS] = max_workers; | |
2988 | cfun->machine->axis_dim_init_p = true; | |
2989 | } | |
2990 | ||
2991 | static int ATTRIBUTE_UNUSED | |
2992 | nvptx_mach_max_workers () | |
2993 | { | |
2994 | if (!cfun->machine->axis_dim_init_p) | |
2995 | init_axis_dim (); | |
2996 | return cfun->machine->axis_dim[MACH_MAX_WORKERS]; | |
2997 | } | |
2998 | ||
2999 | static int ATTRIBUTE_UNUSED | |
3000 | nvptx_mach_vector_length () | |
3001 | { | |
3002 | if (!cfun->machine->axis_dim_init_p) | |
3003 | init_axis_dim (); | |
3004 | return cfun->machine->axis_dim[MACH_VECTOR_LENGTH]; | |
3005 | } | |
3006 | ||
3007 | /* Loop structure of the function. The entire function is described as | |
3008 | a NULL loop. */ | |
3009 | ||
3010 | struct parallel | |
3011 | { | |
3012 | /* Parent parallel. */ | |
3013 | parallel *parent; | |
3014 | ||
3015 | /* Next sibling parallel. */ | |
3016 | parallel *next; | |
3017 | ||
3018 | /* First child parallel. */ | |
3019 | parallel *inner; | |
3020 | ||
3021 | /* Partitioning mask of the parallel. */ | |
3022 | unsigned mask; | |
3023 | ||
3024 | /* Partitioning used within inner parallels. */ | |
3025 | unsigned inner_mask; | |
3026 | ||
3027 | /* Location of parallel forked and join. The forked is the first | |
3028 | block in the parallel and the join is the first block after of | |
3029 | the partition. */ | |
3030 | basic_block forked_block; | |
3031 | basic_block join_block; | |
3032 | ||
3033 | rtx_insn *forked_insn; | |
3034 | rtx_insn *join_insn; | |
3035 | ||
3036 | rtx_insn *fork_insn; | |
3037 | rtx_insn *joining_insn; | |
3038 | ||
3039 | /* Basic blocks in this parallel, but not in child parallels. The | |
3040 | FORKED and JOINING blocks are in the partition. The FORK and JOIN | |
3041 | blocks are not. */ | |
3042 | auto_vec<basic_block> blocks; | |
3043 | ||
3044 | public: | |
3045 | parallel (parallel *parent, unsigned mode); | |
3046 | ~parallel (); | |
3047 | }; | |
3048 | ||
3049 | /* Constructor links the new parallel into it's parent's chain of | |
3050 | children. */ | |
3051 | ||
3052 | parallel::parallel (parallel *parent_, unsigned mask_) | |
3053 | :parent (parent_), next (0), inner (0), mask (mask_), inner_mask (0) | |
3054 | { | |
3055 | forked_block = join_block = 0; | |
3056 | forked_insn = join_insn = 0; | |
3057 | fork_insn = joining_insn = 0; | |
3058 | ||
3059 | if (parent) | |
3060 | { | |
3061 | next = parent->inner; | |
3062 | parent->inner = this; | |
3063 | } | |
3064 | } | |
3065 | ||
3066 | parallel::~parallel () | |
3067 | { | |
3068 | delete inner; | |
3069 | delete next; | |
3070 | } | |
3071 | ||
3072 | /* Map of basic blocks to insns */ | |
3073 | typedef hash_map<basic_block, rtx_insn *> bb_insn_map_t; | |
3074 | ||
3075 | /* A tuple of an insn of interest and the BB in which it resides. */ | |
3076 | typedef std::pair<rtx_insn *, basic_block> insn_bb_t; | |
3077 | typedef auto_vec<insn_bb_t> insn_bb_vec_t; | |
3078 | ||
3079 | /* Split basic blocks such that each forked and join unspecs are at | |
3080 | the start of their basic blocks. Thus afterwards each block will | |
3081 | have a single partitioning mode. We also do the same for return | |
3082 | insns, as they are executed by every thread. Return the | |
3083 | partitioning mode of the function as a whole. Populate MAP with | |
3084 | head and tail blocks. We also clear the BB visited flag, which is | |
3085 | used when finding partitions. */ | |
3086 | ||
3087 | static void | |
3088 | nvptx_split_blocks (bb_insn_map_t *map) | |
3089 | { | |
3090 | insn_bb_vec_t worklist; | |
3091 | basic_block block; | |
3092 | rtx_insn *insn; | |
3093 | ||
3094 | /* Locate all the reorg instructions of interest. */ | |
3095 | FOR_ALL_BB_FN (block, cfun) | |
3096 | { | |
3097 | bool seen_insn = false; | |
3098 | ||
3099 | /* Clear visited flag, for use by parallel locator */ | |
3100 | block->flags &= ~BB_VISITED; | |
3101 | ||
3102 | FOR_BB_INSNS (block, insn) | |
3103 | { | |
3104 | if (!INSN_P (insn)) | |
3105 | continue; | |
3106 | switch (recog_memoized (insn)) | |
3107 | { | |
3108 | default: | |
3109 | seen_insn = true; | |
3110 | continue; | |
3111 | case CODE_FOR_nvptx_forked: | |
3112 | case CODE_FOR_nvptx_join: | |
3113 | break; | |
3114 | ||
3115 | case CODE_FOR_return: | |
3116 | /* We also need to split just before return insns, as | |
3117 | that insn needs executing by all threads, but the | |
3118 | block it is in probably does not. */ | |
3119 | break; | |
3120 | } | |
3121 | ||
3122 | if (seen_insn) | |
3123 | /* We've found an instruction that must be at the start of | |
3124 | a block, but isn't. Add it to the worklist. */ | |
3125 | worklist.safe_push (insn_bb_t (insn, block)); | |
3126 | else | |
3127 | /* It was already the first instruction. Just add it to | |
3128 | the map. */ | |
3129 | map->get_or_insert (block) = insn; | |
3130 | seen_insn = true; | |
3131 | } | |
3132 | } | |
3133 | ||
3134 | /* Split blocks on the worklist. */ | |
3135 | unsigned ix; | |
3136 | insn_bb_t *elt; | |
3137 | basic_block remap = 0; | |
3138 | for (ix = 0; worklist.iterate (ix, &elt); ix++) | |
3139 | { | |
3140 | if (remap != elt->second) | |
3141 | { | |
3142 | block = elt->second; | |
3143 | remap = block; | |
3144 | } | |
3145 | ||
3146 | /* Split block before insn. The insn is in the new block */ | |
3147 | edge e = split_block (block, PREV_INSN (elt->first)); | |
3148 | ||
3149 | block = e->dest; | |
3150 | map->get_or_insert (block) = elt->first; | |
3151 | } | |
3152 | } | |
3153 | ||
3154 | /* Return true if MASK contains parallelism that requires shared | |
3155 | memory to broadcast. */ | |
3156 | ||
3157 | static bool | |
3158 | nvptx_needs_shared_bcast (unsigned mask) | |
3159 | { | |
3160 | bool worker = mask & GOMP_DIM_MASK (GOMP_DIM_WORKER); | |
3161 | bool large_vector = (mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR)) | |
3162 | && nvptx_mach_vector_length () != PTX_WARP_SIZE; | |
3163 | ||
3164 | return worker || large_vector; | |
3165 | } | |
3166 | ||
3167 | /* BLOCK is a basic block containing a head or tail instruction. | |
3168 | Locate the associated prehead or pretail instruction, which must be | |
3169 | in the single predecessor block. */ | |
3170 | ||
3171 | static rtx_insn * | |
3172 | nvptx_discover_pre (basic_block block, int expected) | |
3173 | { | |
3174 | gcc_assert (block->preds->length () == 1); | |
3175 | basic_block pre_block = (*block->preds)[0]->src; | |
3176 | rtx_insn *pre_insn; | |
3177 | ||
3178 | for (pre_insn = BB_END (pre_block); !INSN_P (pre_insn); | |
3179 | pre_insn = PREV_INSN (pre_insn)) | |
3180 | gcc_assert (pre_insn != BB_HEAD (pre_block)); | |
3181 | ||
3182 | gcc_assert (recog_memoized (pre_insn) == expected); | |
3183 | return pre_insn; | |
3184 | } | |
3185 | ||
3186 | /* Dump this parallel and all its inner parallels. */ | |
3187 | ||
3188 | static void | |
3189 | nvptx_dump_pars (parallel *par, unsigned depth) | |
3190 | { | |
3191 | fprintf (dump_file, "%u: mask %d head=%d, tail=%d\n", | |
3192 | depth, par->mask, | |
3193 | par->forked_block ? par->forked_block->index : -1, | |
3194 | par->join_block ? par->join_block->index : -1); | |
3195 | ||
3196 | fprintf (dump_file, " blocks:"); | |
3197 | ||
3198 | basic_block block; | |
3199 | for (unsigned ix = 0; par->blocks.iterate (ix, &block); ix++) | |
3200 | fprintf (dump_file, " %d", block->index); | |
3201 | fprintf (dump_file, "\n"); | |
3202 | if (par->inner) | |
3203 | nvptx_dump_pars (par->inner, depth + 1); | |
3204 | ||
3205 | if (par->next) | |
3206 | nvptx_dump_pars (par->next, depth); | |
3207 | } | |
3208 | ||
3209 | /* If BLOCK contains a fork/join marker, process it to create or | |
3210 | terminate a loop structure. Add this block to the current loop, | |
3211 | and then walk successor blocks. */ | |
3212 | ||
3213 | static parallel * | |
3214 | nvptx_find_par (bb_insn_map_t *map, parallel *par, basic_block block) | |
3215 | { | |
3216 | if (block->flags & BB_VISITED) | |
3217 | return par; | |
3218 | block->flags |= BB_VISITED; | |
3219 | ||
3220 | if (rtx_insn **endp = map->get (block)) | |
3221 | { | |
3222 | rtx_insn *end = *endp; | |
3223 | ||
3224 | /* This is a block head or tail, or return instruction. */ | |
3225 | switch (recog_memoized (end)) | |
3226 | { | |
3227 | case CODE_FOR_return: | |
3228 | /* Return instructions are in their own block, and we | |
3229 | don't need to do anything more. */ | |
3230 | return par; | |
3231 | ||
3232 | case CODE_FOR_nvptx_forked: | |
3233 | /* Loop head, create a new inner loop and add it into | |
3234 | our parent's child list. */ | |
3235 | { | |
3236 | unsigned mask = UINTVAL (XVECEXP (PATTERN (end), 0, 0)); | |
3237 | ||
3238 | gcc_assert (mask); | |
3239 | par = new parallel (par, mask); | |
3240 | par->forked_block = block; | |
3241 | par->forked_insn = end; | |
3242 | if (nvptx_needs_shared_bcast (mask)) | |
3243 | par->fork_insn | |
3244 | = nvptx_discover_pre (block, CODE_FOR_nvptx_fork); | |
3245 | } | |
3246 | break; | |
3247 | ||
3248 | case CODE_FOR_nvptx_join: | |
3249 | /* A loop tail. Finish the current loop and return to | |
3250 | parent. */ | |
3251 | { | |
3252 | unsigned mask = UINTVAL (XVECEXP (PATTERN (end), 0, 0)); | |
3253 | ||
3254 | gcc_assert (par->mask == mask); | |
3255 | gcc_assert (par->join_block == NULL); | |
3256 | par->join_block = block; | |
3257 | par->join_insn = end; | |
3258 | if (nvptx_needs_shared_bcast (mask)) | |
3259 | par->joining_insn | |
3260 | = nvptx_discover_pre (block, CODE_FOR_nvptx_joining); | |
3261 | par = par->parent; | |
3262 | } | |
3263 | break; | |
3264 | ||
3265 | default: | |
3266 | gcc_unreachable (); | |
3267 | } | |
3268 | } | |
3269 | ||
3270 | if (par) | |
3271 | /* Add this block onto the current loop's list of blocks. */ | |
3272 | par->blocks.safe_push (block); | |
3273 | else | |
3274 | /* This must be the entry block. Create a NULL parallel. */ | |
3275 | par = new parallel (0, 0); | |
3276 | ||
3277 | /* Walk successor blocks. */ | |
3278 | edge e; | |
3279 | edge_iterator ei; | |
3280 | ||
3281 | FOR_EACH_EDGE (e, ei, block->succs) | |
3282 | nvptx_find_par (map, par, e->dest); | |
3283 | ||
3284 | return par; | |
3285 | } | |
3286 | ||
3287 | /* DFS walk the CFG looking for fork & join markers. Construct | |
3288 | loop structures as we go. MAP is a mapping of basic blocks | |
3289 | to head & tail markers, discovered when splitting blocks. This | |
3290 | speeds up the discovery. We rely on the BB visited flag having | |
3291 | been cleared when splitting blocks. */ | |
3292 | ||
3293 | static parallel * | |
3294 | nvptx_discover_pars (bb_insn_map_t *map) | |
3295 | { | |
3296 | basic_block block; | |
3297 | ||
3298 | /* Mark exit blocks as visited. */ | |
3299 | block = EXIT_BLOCK_PTR_FOR_FN (cfun); | |
3300 | block->flags |= BB_VISITED; | |
3301 | ||
3302 | /* And entry block as not. */ | |
3303 | block = ENTRY_BLOCK_PTR_FOR_FN (cfun); | |
3304 | block->flags &= ~BB_VISITED; | |
3305 | ||
3306 | parallel *par = nvptx_find_par (map, 0, block); | |
3307 | ||
3308 | if (dump_file) | |
3309 | { | |
3310 | fprintf (dump_file, "\nLoops\n"); | |
3311 | nvptx_dump_pars (par, 0); | |
3312 | fprintf (dump_file, "\n"); | |
3313 | } | |
3314 | ||
3315 | return par; | |
3316 | } | |
3317 | ||
3318 | /* Analyse a group of BBs within a partitioned region and create N | |
3319 | Single-Entry-Single-Exit regions. Some of those regions will be | |
3320 | trivial ones consisting of a single BB. The blocks of a | |
3321 | partitioned region might form a set of disjoint graphs -- because | |
3322 | the region encloses a differently partitoned sub region. | |
3323 | ||
3324 | We use the linear time algorithm described in 'Finding Regions Fast: | |
3325 | Single Entry Single Exit and control Regions in Linear Time' | |
3326 | Johnson, Pearson & Pingali. That algorithm deals with complete | |
3327 | CFGs, where a back edge is inserted from END to START, and thus the | |
3328 | problem becomes one of finding equivalent loops. | |
3329 | ||
3330 | In this case we have a partial CFG. We complete it by redirecting | |
3331 | any incoming edge to the graph to be from an arbitrary external BB, | |
3332 | and similarly redirecting any outgoing edge to be to that BB. | |
3333 | Thus we end up with a closed graph. | |
3334 | ||
3335 | The algorithm works by building a spanning tree of an undirected | |
3336 | graph and keeping track of back edges from nodes further from the | |
3337 | root in the tree to nodes nearer to the root in the tree. In the | |
3338 | description below, the root is up and the tree grows downwards. | |
3339 | ||
3340 | We avoid having to deal with degenerate back-edges to the same | |
3341 | block, by splitting each BB into 3 -- one for input edges, one for | |
3342 | the node itself and one for the output edges. Such back edges are | |
3343 | referred to as 'Brackets'. Cycle equivalent nodes will have the | |
3344 | same set of brackets. | |
3345 | ||
3346 | Determining bracket equivalency is done by maintaining a list of | |
3347 | brackets in such a manner that the list length and final bracket | |
3348 | uniquely identify the set. | |
3349 | ||
3350 | We use coloring to mark all BBs with cycle equivalency with the | |
3351 | same color. This is the output of the 'Finding Regions Fast' | |
3352 | algorithm. Notice it doesn't actually find the set of nodes within | |
3353 | a particular region, just unorderd sets of nodes that are the | |
3354 | entries and exits of SESE regions. | |
3355 | ||
3356 | After determining cycle equivalency, we need to find the minimal | |
3357 | set of SESE regions. Do this with a DFS coloring walk of the | |
3358 | complete graph. We're either 'looking' or 'coloring'. When | |
3359 | looking, and we're in the subgraph, we start coloring the color of | |
3360 | the current node, and remember that node as the start of the | |
3361 | current color's SESE region. Every time we go to a new node, we | |
3362 | decrement the count of nodes with thet color. If it reaches zero, | |
3363 | we remember that node as the end of the current color's SESE region | |
3364 | and return to 'looking'. Otherwise we color the node the current | |
3365 | color. | |
3366 | ||
3367 | This way we end up with coloring the inside of non-trivial SESE | |
3368 | regions with the color of that region. */ | |
3369 | ||
3370 | /* A pair of BBs. We use this to represent SESE regions. */ | |
3371 | typedef std::pair<basic_block, basic_block> bb_pair_t; | |
3372 | typedef auto_vec<bb_pair_t> bb_pair_vec_t; | |
3373 | ||
3374 | /* A node in the undirected CFG. The discriminator SECOND indicates just | |
3375 | above or just below the BB idicated by FIRST. */ | |
3376 | typedef std::pair<basic_block, int> pseudo_node_t; | |
3377 | ||
3378 | /* A bracket indicates an edge towards the root of the spanning tree of the | |
3379 | undirected graph. Each bracket has a color, determined | |
3380 | from the currrent set of brackets. */ | |
3381 | struct bracket | |
3382 | { | |
3383 | pseudo_node_t back; /* Back target */ | |
3384 | ||
3385 | /* Current color and size of set. */ | |
3386 | unsigned color; | |
3387 | unsigned size; | |
3388 | ||
3389 | bracket (pseudo_node_t back_) | |
3390 | : back (back_), color (~0u), size (~0u) | |
3391 | { | |
3392 | } | |
3393 | ||
3394 | unsigned get_color (auto_vec<unsigned> &color_counts, unsigned length) | |
3395 | { | |
3396 | if (length != size) | |
3397 | { | |
3398 | size = length; | |
3399 | color = color_counts.length (); | |
3400 | color_counts.quick_push (0); | |
3401 | } | |
3402 | color_counts[color]++; | |
3403 | return color; | |
3404 | } | |
3405 | }; | |
3406 | ||
3407 | typedef auto_vec<bracket> bracket_vec_t; | |
3408 | ||
3409 | /* Basic block info for finding SESE regions. */ | |
3410 | ||
3411 | struct bb_sese | |
3412 | { | |
3413 | int node; /* Node number in spanning tree. */ | |
3414 | int parent; /* Parent node number. */ | |
3415 | ||
3416 | /* The algorithm splits each node A into Ai, A', Ao. The incoming | |
3417 | edges arrive at pseudo-node Ai and the outgoing edges leave at | |
3418 | pseudo-node Ao. We have to remember which way we arrived at a | |
3419 | particular node when generating the spanning tree. dir > 0 means | |
3420 | we arrived at Ai, dir < 0 means we arrived at Ao. */ | |
3421 | int dir; | |
3422 | ||
3423 | /* Lowest numbered pseudo-node reached via a backedge from thsis | |
3424 | node, or any descendant. */ | |
3425 | pseudo_node_t high; | |
3426 | ||
3427 | int color; /* Cycle-equivalence color */ | |
3428 | ||
3429 | /* Stack of brackets for this node. */ | |
3430 | bracket_vec_t brackets; | |
3431 | ||
3432 | bb_sese (unsigned node_, unsigned p, int dir_) | |
3433 | :node (node_), parent (p), dir (dir_) | |
3434 | { | |
3435 | } | |
3436 | ~bb_sese (); | |
3437 | ||
3438 | /* Push a bracket ending at BACK. */ | |
3439 | void push (const pseudo_node_t &back) | |
3440 | { | |
3441 | if (dump_file) | |
3442 | fprintf (dump_file, "Pushing backedge %d:%+d\n", | |
3443 | back.first ? back.first->index : 0, back.second); | |
3444 | brackets.safe_push (bracket (back)); | |
3445 | } | |
3446 | ||
3447 | void append (bb_sese *child); | |
3448 | void remove (const pseudo_node_t &); | |
3449 | ||
3450 | /* Set node's color. */ | |
3451 | void set_color (auto_vec<unsigned> &color_counts) | |
3452 | { | |
3453 | color = brackets.last ().get_color (color_counts, brackets.length ()); | |
3454 | } | |
3455 | }; | |
3456 | ||
3457 | bb_sese::~bb_sese () | |
3458 | { | |
3459 | } | |
3460 | ||
3461 | /* Destructively append CHILD's brackets. */ | |
3462 | ||
3463 | void | |
3464 | bb_sese::append (bb_sese *child) | |
3465 | { | |
3466 | if (int len = child->brackets.length ()) | |
3467 | { | |
3468 | int ix; | |
3469 | ||
3470 | if (dump_file) | |
3471 | { | |
3472 | for (ix = 0; ix < len; ix++) | |
3473 | { | |
3474 | const pseudo_node_t &pseudo = child->brackets[ix].back; | |
3475 | fprintf (dump_file, "Appending (%d)'s backedge %d:%+d\n", | |
3476 | child->node, pseudo.first ? pseudo.first->index : 0, | |
3477 | pseudo.second); | |
3478 | } | |
3479 | } | |
3480 | if (!brackets.length ()) | |
3481 | std::swap (brackets, child->brackets); | |
3482 | else | |
3483 | { | |
3484 | brackets.reserve (len); | |
3485 | for (ix = 0; ix < len; ix++) | |
3486 | brackets.quick_push (child->brackets[ix]); | |
3487 | } | |
3488 | } | |
3489 | } | |
3490 | ||
3491 | /* Remove brackets that terminate at PSEUDO. */ | |
3492 | ||
3493 | void | |
3494 | bb_sese::remove (const pseudo_node_t &pseudo) | |
3495 | { | |
3496 | unsigned removed = 0; | |
3497 | int len = brackets.length (); | |
3498 | ||
3499 | for (int ix = 0; ix < len; ix++) | |
3500 | { | |
3501 | if (brackets[ix].back == pseudo) | |
3502 | { | |
3503 | if (dump_file) | |
3504 | fprintf (dump_file, "Removing backedge %d:%+d\n", | |
3505 | pseudo.first ? pseudo.first->index : 0, pseudo.second); | |
3506 | removed++; | |
3507 | } | |
3508 | else if (removed) | |
3509 | brackets[ix-removed] = brackets[ix]; | |
3510 | } | |
3511 | while (removed--) | |
3512 | brackets.pop (); | |
3513 | } | |
3514 | ||
3515 | /* Accessors for BB's aux pointer. */ | |
3516 | #define BB_SET_SESE(B, S) ((B)->aux = (S)) | |
3517 | #define BB_GET_SESE(B) ((bb_sese *)(B)->aux) | |
3518 | ||
3519 | /* DFS walk creating SESE data structures. Only cover nodes with | |
3520 | BB_VISITED set. Append discovered blocks to LIST. We number in | |
3521 | increments of 3 so that the above and below pseudo nodes can be | |
3522 | implicitly numbered too. */ | |
3523 | ||
3524 | static int | |
3525 | nvptx_sese_number (int n, int p, int dir, basic_block b, | |
3526 | auto_vec<basic_block> *list) | |
3527 | { | |
3528 | if (BB_GET_SESE (b)) | |
3529 | return n; | |
3530 | ||
3531 | if (dump_file) | |
3532 | fprintf (dump_file, "Block %d(%d), parent (%d), orientation %+d\n", | |
3533 | b->index, n, p, dir); | |
3534 | ||
3535 | BB_SET_SESE (b, new bb_sese (n, p, dir)); | |
3536 | p = n; | |
3537 | ||
3538 | n += 3; | |
3539 | list->quick_push (b); | |
3540 | ||
3541 | /* First walk the nodes on the 'other side' of this node, then walk | |
3542 | the nodes on the same side. */ | |
3543 | for (unsigned ix = 2; ix; ix--) | |
3544 | { | |
3545 | vec<edge, va_gc> *edges = dir > 0 ? b->succs : b->preds; | |
3546 | size_t offset = (dir > 0 ? offsetof (edge_def, dest) | |
3547 | : offsetof (edge_def, src)); | |
3548 | edge e; | |
3549 | edge_iterator ei; | |
3550 | ||
3551 | FOR_EACH_EDGE (e, ei, edges) | |
3552 | { | |
3553 | basic_block target = *(basic_block *)((char *)e + offset); | |
3554 | ||
3555 | if (target->flags & BB_VISITED) | |
3556 | n = nvptx_sese_number (n, p, dir, target, list); | |
3557 | } | |
3558 | dir = -dir; | |
3559 | } | |
3560 | return n; | |
3561 | } | |
3562 | ||
3563 | /* Process pseudo node above (DIR < 0) or below (DIR > 0) ME. | |
3564 | EDGES are the outgoing edges and OFFSET is the offset to the src | |
3565 | or dst block on the edges. */ | |
3566 | ||
3567 | static void | |
3568 | nvptx_sese_pseudo (basic_block me, bb_sese *sese, int depth, int dir, | |
3569 | vec<edge, va_gc> *edges, size_t offset) | |
3570 | { | |
3571 | edge e; | |
3572 | edge_iterator ei; | |
3573 | int hi_back = depth; | |
3574 | pseudo_node_t node_back (0, depth); | |
3575 | int hi_child = depth; | |
3576 | pseudo_node_t node_child (0, depth); | |
3577 | basic_block child = NULL; | |
3578 | unsigned num_children = 0; | |
3579 | int usd = -dir * sese->dir; | |
3580 | ||
3581 | if (dump_file) | |
3582 | fprintf (dump_file, "\nProcessing %d(%d) %+d\n", | |
3583 | me->index, sese->node, dir); | |
3584 | ||
3585 | if (dir < 0) | |
3586 | { | |
3587 | /* This is the above pseudo-child. It has the BB itself as an | |
3588 | additional child node. */ | |
3589 | node_child = sese->high; | |
3590 | hi_child = node_child.second; | |
3591 | if (node_child.first) | |
3592 | hi_child += BB_GET_SESE (node_child.first)->node; | |
3593 | num_children++; | |
3594 | } | |
3595 | ||
3596 | /* Examine each edge. | |
3597 | - if it is a child (a) append its bracket list and (b) record | |
3598 | whether it is the child with the highest reaching bracket. | |
3599 | - if it is an edge to ancestor, record whether it's the highest | |
3600 | reaching backlink. */ | |
3601 | FOR_EACH_EDGE (e, ei, edges) | |
3602 | { | |
3603 | basic_block target = *(basic_block *)((char *)e + offset); | |
3604 | ||
3605 | if (bb_sese *t_sese = BB_GET_SESE (target)) | |
3606 | { | |
3607 | if (t_sese->parent == sese->node && !(t_sese->dir + usd)) | |
3608 | { | |
3609 | /* Child node. Append its bracket list. */ | |
3610 | num_children++; | |
3611 | sese->append (t_sese); | |
3612 | ||
3613 | /* Compare it's hi value. */ | |
3614 | int t_hi = t_sese->high.second; | |
3615 | ||
3616 | if (basic_block child_hi_block = t_sese->high.first) | |
3617 | t_hi += BB_GET_SESE (child_hi_block)->node; | |
3618 | ||
3619 | if (hi_child > t_hi) | |
3620 | { | |
3621 | hi_child = t_hi; | |
3622 | node_child = t_sese->high; | |
3623 | child = target; | |
3624 | } | |
3625 | } | |
3626 | else if (t_sese->node < sese->node + dir | |
3627 | && !(dir < 0 && sese->parent == t_sese->node)) | |
3628 | { | |
3629 | /* Non-parental ancestor node -- a backlink. */ | |
3630 | int d = usd * t_sese->dir; | |
3631 | int back = t_sese->node + d; | |
3632 | ||
3633 | if (hi_back > back) | |
3634 | { | |
3635 | hi_back = back; | |
3636 | node_back = pseudo_node_t (target, d); | |
3637 | } | |
3638 | } | |
3639 | } | |
3640 | else | |
3641 | { /* Fallen off graph, backlink to entry node. */ | |
3642 | hi_back = 0; | |
3643 | node_back = pseudo_node_t (0, 0); | |
3644 | } | |
3645 | } | |
3646 | ||
3647 | /* Remove any brackets that terminate at this pseudo node. */ | |
3648 | sese->remove (pseudo_node_t (me, dir)); | |
3649 | ||
3650 | /* Now push any backlinks from this pseudo node. */ | |
3651 | FOR_EACH_EDGE (e, ei, edges) | |
3652 | { | |
3653 | basic_block target = *(basic_block *)((char *)e + offset); | |
3654 | if (bb_sese *t_sese = BB_GET_SESE (target)) | |
3655 | { | |
3656 | if (t_sese->node < sese->node + dir | |
3657 | && !(dir < 0 && sese->parent == t_sese->node)) | |
3658 | /* Non-parental ancestor node - backedge from me. */ | |
3659 | sese->push (pseudo_node_t (target, usd * t_sese->dir)); | |
3660 | } | |
3661 | else | |
3662 | { | |
3663 | /* back edge to entry node */ | |
3664 | sese->push (pseudo_node_t (0, 0)); | |
3665 | } | |
3666 | } | |
3667 | ||
3668 | /* If this node leads directly or indirectly to a no-return region of | |
3669 | the graph, then fake a backedge to entry node. */ | |
3670 | if (!sese->brackets.length () || !edges || !edges->length ()) | |
3671 | { | |
3672 | hi_back = 0; | |
3673 | node_back = pseudo_node_t (0, 0); | |
3674 | sese->push (node_back); | |
3675 | } | |
3676 | ||
3677 | /* Record the highest reaching backedge from us or a descendant. */ | |
3678 | sese->high = hi_back < hi_child ? node_back : node_child; | |
3679 | ||
3680 | if (num_children > 1) | |
3681 | { | |
3682 | /* There is more than one child -- this is a Y shaped piece of | |
3683 | spanning tree. We have to insert a fake backedge from this | |
3684 | node to the highest ancestor reached by not-the-highest | |
3685 | reaching child. Note that there may be multiple children | |
3686 | with backedges to the same highest node. That's ok and we | |
3687 | insert the edge to that highest node. */ | |
3688 | hi_child = depth; | |
3689 | if (dir < 0 && child) | |
3690 | { | |
3691 | node_child = sese->high; | |
3692 | hi_child = node_child.second; | |
3693 | if (node_child.first) | |
3694 | hi_child += BB_GET_SESE (node_child.first)->node; | |
3695 | } | |
3696 | ||
3697 | FOR_EACH_EDGE (e, ei, edges) | |
3698 | { | |
3699 | basic_block target = *(basic_block *)((char *)e + offset); | |
3700 | ||
3701 | if (target == child) | |
3702 | /* Ignore the highest child. */ | |
3703 | continue; | |
3704 | ||
3705 | bb_sese *t_sese = BB_GET_SESE (target); | |
3706 | if (!t_sese) | |
3707 | continue; | |
3708 | if (t_sese->parent != sese->node) | |
3709 | /* Not a child. */ | |
3710 | continue; | |
3711 | ||
3712 | /* Compare its hi value. */ | |
3713 | int t_hi = t_sese->high.second; | |
3714 | ||
3715 | if (basic_block child_hi_block = t_sese->high.first) | |
3716 | t_hi += BB_GET_SESE (child_hi_block)->node; | |
3717 | ||
3718 | if (hi_child > t_hi) | |
3719 | { | |
3720 | hi_child = t_hi; | |
3721 | node_child = t_sese->high; | |
3722 | } | |
3723 | } | |
3724 | ||
3725 | sese->push (node_child); | |
3726 | } | |
3727 | } | |
3728 | ||
3729 | ||
3730 | /* DFS walk of BB graph. Color node BLOCK according to COLORING then | |
3731 | proceed to successors. Set SESE entry and exit nodes of | |
3732 | REGIONS. */ | |
3733 | ||
3734 | static void | |
3735 | nvptx_sese_color (auto_vec<unsigned> &color_counts, bb_pair_vec_t ®ions, | |
3736 | basic_block block, int coloring) | |
3737 | { | |
3738 | bb_sese *sese = BB_GET_SESE (block); | |
3739 | ||
3740 | if (block->flags & BB_VISITED) | |
3741 | { | |
3742 | /* If we've already encountered this block, either we must not | |
3743 | be coloring, or it must have been colored the current color. */ | |
3744 | gcc_assert (coloring < 0 || (sese && coloring == sese->color)); | |
3745 | return; | |
3746 | } | |
3747 | ||
3748 | block->flags |= BB_VISITED; | |
3749 | ||
3750 | if (sese) | |
3751 | { | |
3752 | if (coloring < 0) | |
3753 | { | |
3754 | /* Start coloring a region. */ | |
3755 | regions[sese->color].first = block; | |
3756 | coloring = sese->color; | |
3757 | } | |
3758 | ||
3759 | if (!--color_counts[sese->color] && sese->color == coloring) | |
3760 | { | |
3761 | /* Found final block of SESE region. */ | |
3762 | regions[sese->color].second = block; | |
3763 | coloring = -1; | |
3764 | } | |
3765 | else | |
3766 | /* Color the node, so we can assert on revisiting the node | |
3767 | that the graph is indeed SESE. */ | |
3768 | sese->color = coloring; | |
3769 | } | |
3770 | else | |
3771 | /* Fallen off the subgraph, we cannot be coloring. */ | |
3772 | gcc_assert (coloring < 0); | |
3773 | ||
3774 | /* Walk each successor block. */ | |
3775 | if (block->succs && block->succs->length ()) | |
3776 | { | |
3777 | edge e; | |
3778 | edge_iterator ei; | |
3779 | ||
3780 | FOR_EACH_EDGE (e, ei, block->succs) | |
3781 | nvptx_sese_color (color_counts, regions, e->dest, coloring); | |
3782 | } | |
3783 | else | |
3784 | gcc_assert (coloring < 0); | |
3785 | } | |
3786 | ||
3787 | /* Find minimal set of SESE regions covering BLOCKS. REGIONS might | |
3788 | end up with NULL entries in it. */ | |
3789 | ||
3790 | static void | |
3791 | nvptx_find_sese (auto_vec<basic_block> &blocks, bb_pair_vec_t ®ions) | |
3792 | { | |
3793 | basic_block block; | |
3794 | int ix; | |
3795 | ||
3796 | /* First clear each BB of the whole function. */ | |
3797 | FOR_ALL_BB_FN (block, cfun) | |
3798 | { | |
3799 | block->flags &= ~BB_VISITED; | |
3800 | BB_SET_SESE (block, 0); | |
3801 | } | |
3802 | ||
3803 | /* Mark blocks in the function that are in this graph. */ | |
3804 | for (ix = 0; blocks.iterate (ix, &block); ix++) | |
3805 | block->flags |= BB_VISITED; | |
3806 | ||
3807 | /* Counts of nodes assigned to each color. There cannot be more | |
3808 | colors than blocks (and hopefully there will be fewer). */ | |
3809 | auto_vec<unsigned> color_counts; | |
3810 | color_counts.reserve (blocks.length ()); | |
3811 | ||
3812 | /* Worklist of nodes in the spanning tree. Again, there cannot be | |
3813 | more nodes in the tree than blocks (there will be fewer if the | |
3814 | CFG of blocks is disjoint). */ | |
3815 | auto_vec<basic_block> spanlist; | |
3816 | spanlist.reserve (blocks.length ()); | |
3817 | ||
3818 | /* Make sure every block has its cycle class determined. */ | |
3819 | for (ix = 0; blocks.iterate (ix, &block); ix++) | |
3820 | { | |
3821 | if (BB_GET_SESE (block)) | |
3822 | /* We already met this block in an earlier graph solve. */ | |
3823 | continue; | |
3824 | ||
3825 | if (dump_file) | |
3826 | fprintf (dump_file, "Searching graph starting at %d\n", block->index); | |
3827 | ||
3828 | /* Number the nodes reachable from block initial DFS order. */ | |
3829 | int depth = nvptx_sese_number (2, 0, +1, block, &spanlist); | |
3830 | ||
3831 | /* Now walk in reverse DFS order to find cycle equivalents. */ | |
3832 | while (spanlist.length ()) | |
3833 | { | |
3834 | block = spanlist.pop (); | |
3835 | bb_sese *sese = BB_GET_SESE (block); | |
3836 | ||
3837 | /* Do the pseudo node below. */ | |
3838 | nvptx_sese_pseudo (block, sese, depth, +1, | |
3839 | sese->dir > 0 ? block->succs : block->preds, | |
3840 | (sese->dir > 0 ? offsetof (edge_def, dest) | |
3841 | : offsetof (edge_def, src))); | |
3842 | sese->set_color (color_counts); | |
3843 | /* Do the pseudo node above. */ | |
3844 | nvptx_sese_pseudo (block, sese, depth, -1, | |
3845 | sese->dir < 0 ? block->succs : block->preds, | |
3846 | (sese->dir < 0 ? offsetof (edge_def, dest) | |
3847 | : offsetof (edge_def, src))); | |
3848 | } | |
3849 | if (dump_file) | |
3850 | fprintf (dump_file, "\n"); | |
3851 | } | |
3852 | ||
3853 | if (dump_file) | |
3854 | { | |
3855 | unsigned count; | |
3856 | const char *comma = ""; | |
3857 | ||
3858 | fprintf (dump_file, "Found %d cycle equivalents\n", | |
3859 | color_counts.length ()); | |
3860 | for (ix = 0; color_counts.iterate (ix, &count); ix++) | |
3861 | { | |
3862 | fprintf (dump_file, "%s%d[%d]={", comma, ix, count); | |
3863 | ||
3864 | comma = ""; | |
3865 | for (unsigned jx = 0; blocks.iterate (jx, &block); jx++) | |
3866 | if (BB_GET_SESE (block)->color == ix) | |
3867 | { | |
3868 | block->flags |= BB_VISITED; | |
3869 | fprintf (dump_file, "%s%d", comma, block->index); | |
3870 | comma=","; | |
3871 | } | |
3872 | fprintf (dump_file, "}"); | |
3873 | comma = ", "; | |
3874 | } | |
3875 | fprintf (dump_file, "\n"); | |
3876 | } | |
3877 | ||
3878 | /* Now we've colored every block in the subgraph. We now need to | |
3879 | determine the minimal set of SESE regions that cover that | |
3880 | subgraph. Do this with a DFS walk of the complete function. | |
3881 | During the walk we're either 'looking' or 'coloring'. When we | |
3882 | reach the last node of a particular color, we stop coloring and | |
3883 | return to looking. */ | |
3884 | ||
3885 | /* There cannot be more SESE regions than colors. */ | |
3886 | regions.reserve (color_counts.length ()); | |
3887 | for (ix = color_counts.length (); ix--;) | |
3888 | regions.quick_push (bb_pair_t (0, 0)); | |
3889 | ||
3890 | for (ix = 0; blocks.iterate (ix, &block); ix++) | |
3891 | block->flags &= ~BB_VISITED; | |
3892 | ||
3893 | nvptx_sese_color (color_counts, regions, ENTRY_BLOCK_PTR_FOR_FN (cfun), -1); | |
3894 | ||
3895 | if (dump_file) | |
3896 | { | |
3897 | const char *comma = ""; | |
3898 | int len = regions.length (); | |
3899 | ||
3900 | fprintf (dump_file, "SESE regions:"); | |
3901 | for (ix = 0; ix != len; ix++) | |
3902 | { | |
3903 | basic_block from = regions[ix].first; | |
3904 | basic_block to = regions[ix].second; | |
3905 | ||
3906 | if (from) | |
3907 | { | |
3908 | fprintf (dump_file, "%s %d{%d", comma, ix, from->index); | |
3909 | if (to != from) | |
3910 | fprintf (dump_file, "->%d", to->index); | |
3911 | ||
3912 | int color = BB_GET_SESE (from)->color; | |
3913 | ||
3914 | /* Print the blocks within the region (excluding ends). */ | |
3915 | FOR_EACH_BB_FN (block, cfun) | |
3916 | { | |
3917 | bb_sese *sese = BB_GET_SESE (block); | |
3918 | ||
3919 | if (sese && sese->color == color | |
3920 | && block != from && block != to) | |
3921 | fprintf (dump_file, ".%d", block->index); | |
3922 | } | |
3923 | fprintf (dump_file, "}"); | |
3924 | } | |
3925 | comma = ","; | |
3926 | } | |
3927 | fprintf (dump_file, "\n\n"); | |
3928 | } | |
3929 | ||
3930 | for (ix = 0; blocks.iterate (ix, &block); ix++) | |
3931 | delete BB_GET_SESE (block); | |
3932 | } | |
3933 | ||
3934 | #undef BB_SET_SESE | |
3935 | #undef BB_GET_SESE | |
3936 | ||
3937 | /* Propagate live state at the start of a partitioned region. IS_CALL | |
3938 | indicates whether the propagation is for a (partitioned) call | |
3939 | instruction. BLOCK provides the live register information, and | |
3940 | might not contain INSN. Propagation is inserted just after INSN. RW | |
3941 | indicates whether we are reading and/or writing state. This | |
3942 | separation is needed for worker-level proppagation where we | |
3943 | essentially do a spill & fill. FN is the underlying worker | |
3944 | function to generate the propagation instructions for single | |
3945 | register. DATA is user data. | |
3946 | ||
3947 | Returns true if we didn't emit any instructions. | |
3948 | ||
3949 | We propagate the live register set for non-calls and the entire | |
3950 | frame for calls and non-calls. We could do better by (a) | |
3951 | propagating just the live set that is used within the partitioned | |
3952 | regions and (b) only propagating stack entries that are used. The | |
3953 | latter might be quite hard to determine. */ | |
3954 | ||
3955 | typedef rtx (*propagator_fn) (rtx, propagate_mask, unsigned, void *, bool); | |
3956 | ||
3957 | static bool | |
3958 | nvptx_propagate (bool is_call, basic_block block, rtx_insn *insn, | |
3959 | propagate_mask rw, propagator_fn fn, void *data, bool vector) | |
3960 | { | |
3961 | bitmap live = DF_LIVE_IN (block); | |
3962 | bitmap_iterator iterator; | |
3963 | unsigned ix; | |
3964 | bool empty = true; | |
3965 | ||
3966 | /* Copy the frame array. */ | |
3967 | HOST_WIDE_INT fs = get_frame_size (); | |
3968 | if (fs) | |
3969 | { | |
3970 | rtx tmp = gen_reg_rtx (DImode); | |
3971 | rtx idx = NULL_RTX; | |
3972 | rtx ptr = gen_reg_rtx (Pmode); | |
3973 | rtx pred = NULL_RTX; | |
3974 | rtx_code_label *label = NULL; | |
3975 | ||
3976 | empty = false; | |
3977 | /* The frame size might not be DImode compatible, but the frame | |
3978 | array's declaration will be. So it's ok to round up here. */ | |
3979 | fs = (fs + GET_MODE_SIZE (DImode) - 1) / GET_MODE_SIZE (DImode); | |
3980 | /* Detect single iteration loop. */ | |
3981 | if (fs == 1) | |
3982 | fs = 0; | |
3983 | ||
3984 | start_sequence (); | |
3985 | emit_insn (gen_rtx_SET (ptr, frame_pointer_rtx)); | |
3986 | if (fs) | |
3987 | { | |
3988 | idx = gen_reg_rtx (SImode); | |
3989 | pred = gen_reg_rtx (BImode); | |
3990 | label = gen_label_rtx (); | |
3991 | ||
3992 | emit_insn (gen_rtx_SET (idx, GEN_INT (fs))); | |
3993 | /* Allow worker function to initialize anything needed. */ | |
3994 | rtx init = fn (tmp, PM_loop_begin, fs, data, vector); | |
3995 | if (init) | |
3996 | emit_insn (init); | |
3997 | emit_label (label); | |
3998 | LABEL_NUSES (label)++; | |
3999 | emit_insn (gen_addsi3 (idx, idx, GEN_INT (-1))); | |
4000 | } | |
4001 | if (rw & PM_read) | |
4002 | emit_insn (gen_rtx_SET (tmp, gen_rtx_MEM (DImode, ptr))); | |
4003 | emit_insn (fn (tmp, rw, fs, data, vector)); | |
4004 | if (rw & PM_write) | |
4005 | emit_insn (gen_rtx_SET (gen_rtx_MEM (DImode, ptr), tmp)); | |
4006 | if (fs) | |
4007 | { | |
4008 | emit_insn (gen_rtx_SET (pred, gen_rtx_NE (BImode, idx, const0_rtx))); | |
4009 | emit_insn (gen_adddi3 (ptr, ptr, GEN_INT (GET_MODE_SIZE (DImode)))); | |
4010 | emit_insn (gen_br_true_uni (pred, label)); | |
4011 | rtx fini = fn (tmp, PM_loop_end, fs, data, vector); | |
4012 | if (fini) | |
4013 | emit_insn (fini); | |
4014 | emit_insn (gen_rtx_CLOBBER (GET_MODE (idx), idx)); | |
4015 | } | |
4016 | emit_insn (gen_rtx_CLOBBER (GET_MODE (tmp), tmp)); | |
4017 | emit_insn (gen_rtx_CLOBBER (GET_MODE (ptr), ptr)); | |
4018 | rtx cpy = get_insns (); | |
4019 | end_sequence (); | |
4020 | insn = emit_insn_after (cpy, insn); | |
4021 | } | |
4022 | ||
4023 | if (!is_call) | |
4024 | /* Copy live registers. */ | |
4025 | EXECUTE_IF_SET_IN_BITMAP (live, 0, ix, iterator) | |
4026 | { | |
4027 | rtx reg = regno_reg_rtx[ix]; | |
4028 | ||
4029 | if (REGNO (reg) >= FIRST_PSEUDO_REGISTER) | |
4030 | { | |
4031 | rtx bcast = fn (reg, rw, 0, data, vector); | |
4032 | ||
4033 | insn = emit_insn_after (bcast, insn); | |
4034 | empty = false; | |
4035 | } | |
4036 | } | |
4037 | return empty; | |
4038 | } | |
4039 | ||
4040 | /* Worker for nvptx_warp_propagate. */ | |
4041 | ||
4042 | static rtx | |
4043 | warp_prop_gen (rtx reg, propagate_mask pm, | |
4044 | unsigned ARG_UNUSED (count), void *ARG_UNUSED (data), | |
4045 | bool ARG_UNUSED (vector)) | |
4046 | { | |
4047 | if (!(pm & PM_read_write)) | |
4048 | return 0; | |
4049 | ||
4050 | return nvptx_gen_warp_bcast (reg); | |
4051 | } | |
4052 | ||
4053 | /* Propagate state that is live at start of BLOCK across the vectors | |
4054 | of a single warp. Propagation is inserted just after INSN. | |
4055 | IS_CALL and return as for nvptx_propagate. */ | |
4056 | ||
4057 | static bool | |
4058 | nvptx_warp_propagate (bool is_call, basic_block block, rtx_insn *insn) | |
4059 | { | |
4060 | return nvptx_propagate (is_call, block, insn, PM_read_write, | |
4061 | warp_prop_gen, 0, false); | |
4062 | } | |
4063 | ||
4064 | /* Worker for nvptx_shared_propagate. */ | |
4065 | ||
4066 | static rtx | |
4067 | shared_prop_gen (rtx reg, propagate_mask pm, unsigned rep, void *data_, | |
4068 | bool vector) | |
4069 | { | |
4070 | broadcast_data_t *data = (broadcast_data_t *)data_; | |
4071 | ||
4072 | if (pm & PM_loop_begin) | |
4073 | { | |
4074 | /* Starting a loop, initialize pointer. */ | |
4075 | unsigned align = GET_MODE_ALIGNMENT (GET_MODE (reg)) / BITS_PER_UNIT; | |
4076 | ||
4077 | oacc_bcast_align = MAX (oacc_bcast_align, align); | |
4078 | data->offset = ROUND_UP (data->offset, align); | |
4079 | ||
4080 | data->ptr = gen_reg_rtx (Pmode); | |
4081 | ||
4082 | return gen_adddi3 (data->ptr, data->base, GEN_INT (data->offset)); | |
4083 | } | |
4084 | else if (pm & PM_loop_end) | |
4085 | { | |
4086 | rtx clobber = gen_rtx_CLOBBER (GET_MODE (data->ptr), data->ptr); | |
4087 | data->ptr = NULL_RTX; | |
4088 | return clobber; | |
4089 | } | |
4090 | else | |
4091 | return nvptx_gen_shared_bcast (reg, pm, rep, data, vector); | |
4092 | } | |
4093 | ||
4094 | /* Spill or fill live state that is live at start of BLOCK. PRE_P | |
4095 | indicates if this is just before partitioned mode (do spill), or | |
4096 | just after it starts (do fill). Sequence is inserted just after | |
4097 | INSN. IS_CALL and return as for nvptx_propagate. */ | |
4098 | ||
4099 | static bool | |
4100 | nvptx_shared_propagate (bool pre_p, bool is_call, basic_block block, | |
4101 | rtx_insn *insn, bool vector) | |
4102 | { | |
4103 | broadcast_data_t data; | |
4104 | ||
4105 | data.base = gen_reg_rtx (Pmode); | |
4106 | data.offset = 0; | |
4107 | data.ptr = NULL_RTX; | |
4108 | ||
4109 | bool empty = nvptx_propagate (is_call, block, insn, | |
4110 | pre_p ? PM_read : PM_write, shared_prop_gen, | |
4111 | &data, vector); | |
4112 | gcc_assert (empty == !data.offset); | |
4113 | if (data.offset) | |
4114 | { | |
4115 | rtx bcast_sym = oacc_bcast_sym; | |
4116 | ||
4117 | /* Stuff was emitted, initialize the base pointer now. */ | |
4118 | if (vector && nvptx_mach_max_workers () > 1) | |
4119 | { | |
4120 | if (!cfun->machine->bcast_partition) | |
4121 | { | |
4122 | /* It would be nice to place this register in | |
4123 | DATA_AREA_SHARED. */ | |
4124 | cfun->machine->bcast_partition = gen_reg_rtx (DImode); | |
4125 | } | |
4126 | if (!cfun->machine->sync_bar) | |
4127 | cfun->machine->sync_bar = gen_reg_rtx (SImode); | |
4128 | ||
4129 | bcast_sym = cfun->machine->bcast_partition; | |
4130 | } | |
4131 | ||
4132 | rtx init = gen_rtx_SET (data.base, bcast_sym); | |
4133 | emit_insn_after (init, insn); | |
4134 | ||
4135 | unsigned int psize = ROUND_UP (data.offset, oacc_bcast_align); | |
4136 | unsigned int pnum = (nvptx_mach_vector_length () > PTX_WARP_SIZE | |
4137 | ? nvptx_mach_max_workers () + 1 | |
4138 | : 1); | |
4139 | ||
4140 | oacc_bcast_partition = MAX (oacc_bcast_partition, psize); | |
4141 | oacc_bcast_size = MAX (oacc_bcast_size, psize * pnum); | |
4142 | } | |
4143 | return empty; | |
4144 | } | |
4145 | ||
4146 | /* Emit a CTA-level synchronization barrier. LOCK is the barrier number, | |
4147 | which is an integer or a register. THREADS is the number of threads | |
4148 | controlled by the barrier. */ | |
4149 | ||
4150 | static rtx | |
4151 | nvptx_cta_sync (rtx lock, int threads) | |
4152 | { | |
4153 | return gen_nvptx_barsync (lock, GEN_INT (threads)); | |
4154 | } | |
4155 | ||
4156 | #if WORKAROUND_PTXJIT_BUG | |
4157 | /* Return first real insn in BB, or return NULL_RTX if BB does not contain | |
4158 | real insns. */ | |
4159 | ||
4160 | static rtx_insn * | |
4161 | bb_first_real_insn (basic_block bb) | |
4162 | { | |
4163 | rtx_insn *insn; | |
4164 | ||
4165 | /* Find first insn of from block. */ | |
4166 | FOR_BB_INSNS (bb, insn) | |
4167 | if (INSN_P (insn)) | |
4168 | return insn; | |
4169 | ||
4170 | return 0; | |
4171 | } | |
4172 | #endif | |
4173 | ||
4174 | /* Return true if INSN needs neutering. */ | |
4175 | ||
4176 | static bool | |
4177 | needs_neutering_p (rtx_insn *insn) | |
4178 | { | |
4179 | if (!INSN_P (insn)) | |
4180 | return false; | |
4181 | ||
4182 | switch (recog_memoized (insn)) | |
4183 | { | |
4184 | case CODE_FOR_nvptx_fork: | |
4185 | case CODE_FOR_nvptx_forked: | |
4186 | case CODE_FOR_nvptx_joining: | |
4187 | case CODE_FOR_nvptx_join: | |
4188 | case CODE_FOR_nvptx_barsync: | |
4189 | return false; | |
4190 | default: | |
4191 | return true; | |
4192 | } | |
4193 | } | |
4194 | ||
4195 | /* Verify position of VECTOR_{JUMP,LABEL} and WORKER_{JUMP,LABEL} in FROM. */ | |
4196 | ||
4197 | static bool | |
4198 | verify_neutering_jumps (basic_block from, | |
4199 | rtx_insn *vector_jump, rtx_insn *worker_jump, | |
4200 | rtx_insn *vector_label, rtx_insn *worker_label) | |
4201 | { | |
4202 | basic_block bb = from; | |
4203 | rtx_insn *insn = BB_HEAD (bb); | |
4204 | bool seen_worker_jump = false; | |
4205 | bool seen_vector_jump = false; | |
4206 | bool seen_worker_label = false; | |
4207 | bool seen_vector_label = false; | |
4208 | bool worker_neutered = false; | |
4209 | bool vector_neutered = false; | |
4210 | while (true) | |
4211 | { | |
4212 | if (insn == worker_jump) | |
4213 | { | |
4214 | seen_worker_jump = true; | |
4215 | worker_neutered = true; | |
4216 | gcc_assert (!vector_neutered); | |
4217 | } | |
4218 | else if (insn == vector_jump) | |
4219 | { | |
4220 | seen_vector_jump = true; | |
4221 | vector_neutered = true; | |
4222 | } | |
4223 | else if (insn == worker_label) | |
4224 | { | |
4225 | seen_worker_label = true; | |
4226 | gcc_assert (worker_neutered); | |
4227 | worker_neutered = false; | |
4228 | } | |
4229 | else if (insn == vector_label) | |
4230 | { | |
4231 | seen_vector_label = true; | |
4232 | gcc_assert (vector_neutered); | |
4233 | vector_neutered = false; | |
4234 | } | |
4235 | else if (INSN_P (insn)) | |
4236 | switch (recog_memoized (insn)) | |
4237 | { | |
4238 | case CODE_FOR_nvptx_barsync: | |
4239 | gcc_assert (!vector_neutered && !worker_neutered); | |
4240 | break; | |
4241 | default: | |
4242 | break; | |
4243 | } | |
4244 | ||
4245 | if (insn != BB_END (bb)) | |
4246 | insn = NEXT_INSN (insn); | |
4247 | else if (JUMP_P (insn) && single_succ_p (bb) | |
4248 | && !seen_vector_jump && !seen_worker_jump) | |
4249 | { | |
4250 | bb = single_succ (bb); | |
4251 | insn = BB_HEAD (bb); | |
4252 | } | |
4253 | else | |
4254 | break; | |
4255 | } | |
4256 | ||
4257 | gcc_assert (!(vector_jump && !seen_vector_jump)); | |
4258 | gcc_assert (!(worker_jump && !seen_worker_jump)); | |
4259 | ||
4260 | if (seen_vector_label || seen_worker_label) | |
4261 | { | |
4262 | gcc_assert (!(vector_label && !seen_vector_label)); | |
4263 | gcc_assert (!(worker_label && !seen_worker_label)); | |
4264 | ||
4265 | return true; | |
4266 | } | |
4267 | ||
4268 | return false; | |
4269 | } | |
4270 | ||
4271 | /* Verify position of VECTOR_LABEL and WORKER_LABEL in TO. */ | |
4272 | ||
4273 | static void | |
4274 | verify_neutering_labels (basic_block to, rtx_insn *vector_label, | |
4275 | rtx_insn *worker_label) | |
4276 | { | |
4277 | basic_block bb = to; | |
4278 | rtx_insn *insn = BB_END (bb); | |
4279 | bool seen_worker_label = false; | |
4280 | bool seen_vector_label = false; | |
4281 | while (true) | |
4282 | { | |
4283 | if (insn == worker_label) | |
4284 | { | |
4285 | seen_worker_label = true; | |
4286 | gcc_assert (!seen_vector_label); | |
4287 | } | |
4288 | else if (insn == vector_label) | |
4289 | seen_vector_label = true; | |
4290 | else if (INSN_P (insn)) | |
4291 | switch (recog_memoized (insn)) | |
4292 | { | |
4293 | case CODE_FOR_nvptx_barsync: | |
4294 | gcc_assert (!seen_vector_label && !seen_worker_label); | |
4295 | break; | |
4296 | } | |
4297 | ||
4298 | if (insn != BB_HEAD (bb)) | |
4299 | insn = PREV_INSN (insn); | |
4300 | else | |
4301 | break; | |
4302 | } | |
4303 | ||
4304 | gcc_assert (!(vector_label && !seen_vector_label)); | |
4305 | gcc_assert (!(worker_label && !seen_worker_label)); | |
4306 | } | |
4307 | ||
4308 | /* Single neutering according to MASK. FROM is the incoming block and | |
4309 | TO is the outgoing block. These may be the same block. Insert at | |
4310 | start of FROM: | |
4311 | ||
4312 | if (tid.<axis>) goto end. | |
4313 | ||
4314 | and insert before ending branch of TO (if there is such an insn): | |
4315 | ||
4316 | end: | |
4317 | <possibly-broadcast-cond> | |
4318 | <branch> | |
4319 | ||
4320 | We currently only use differnt FROM and TO when skipping an entire | |
4321 | loop. We could do more if we detected superblocks. */ | |
4322 | ||
4323 | static void | |
4324 | nvptx_single (unsigned mask, basic_block from, basic_block to) | |
4325 | { | |
4326 | rtx_insn *head = BB_HEAD (from); | |
4327 | rtx_insn *tail = BB_END (to); | |
4328 | unsigned skip_mask = mask; | |
4329 | ||
4330 | while (true) | |
4331 | { | |
4332 | /* Find first insn of from block. */ | |
4333 | while (head != BB_END (from) && !needs_neutering_p (head)) | |
4334 | head = NEXT_INSN (head); | |
4335 | ||
4336 | if (from == to) | |
4337 | break; | |
4338 | ||
4339 | if (!(JUMP_P (head) && single_succ_p (from))) | |
4340 | break; | |
4341 | ||
4342 | basic_block jump_target = single_succ (from); | |
4343 | if (!single_pred_p (jump_target)) | |
4344 | break; | |
4345 | ||
4346 | from = jump_target; | |
4347 | head = BB_HEAD (from); | |
4348 | } | |
4349 | ||
4350 | /* Find last insn of to block */ | |
4351 | rtx_insn *limit = from == to ? head : BB_HEAD (to); | |
4352 | while (tail != limit && !INSN_P (tail) && !LABEL_P (tail)) | |
4353 | tail = PREV_INSN (tail); | |
4354 | ||
4355 | /* Detect if tail is a branch. */ | |
4356 | rtx tail_branch = NULL_RTX; | |
4357 | rtx cond_branch = NULL_RTX; | |
4358 | if (tail && INSN_P (tail)) | |
4359 | { | |
4360 | tail_branch = PATTERN (tail); | |
4361 | if (GET_CODE (tail_branch) != SET || SET_DEST (tail_branch) != pc_rtx) | |
4362 | tail_branch = NULL_RTX; | |
4363 | else | |
4364 | { | |
4365 | cond_branch = SET_SRC (tail_branch); | |
4366 | if (GET_CODE (cond_branch) != IF_THEN_ELSE) | |
4367 | cond_branch = NULL_RTX; | |
4368 | } | |
4369 | } | |
4370 | ||
4371 | if (tail == head) | |
4372 | { | |
4373 | /* If this is empty, do nothing. */ | |
4374 | if (!head || !needs_neutering_p (head)) | |
4375 | return; | |
4376 | ||
4377 | if (cond_branch) | |
4378 | { | |
4379 | /* If we're only doing vector single, there's no need to | |
4380 | emit skip code because we'll not insert anything. */ | |
4381 | if (!(mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR))) | |
4382 | skip_mask = 0; | |
4383 | } | |
4384 | else if (tail_branch) | |
4385 | /* Block with only unconditional branch. Nothing to do. */ | |
4386 | return; | |
4387 | } | |
4388 | ||
4389 | /* Insert the vector test inside the worker test. */ | |
4390 | unsigned mode; | |
4391 | rtx_insn *before = tail; | |
4392 | rtx_insn *neuter_start = NULL; | |
4393 | rtx_insn *worker_label = NULL, *vector_label = NULL; | |
4394 | rtx_insn *worker_jump = NULL, *vector_jump = NULL; | |
4395 | for (mode = GOMP_DIM_WORKER; mode <= GOMP_DIM_VECTOR; mode++) | |
4396 | if (GOMP_DIM_MASK (mode) & skip_mask) | |
4397 | { | |
4398 | rtx_code_label *label = gen_label_rtx (); | |
4399 | rtx pred = cfun->machine->axis_predicate[mode - GOMP_DIM_WORKER]; | |
4400 | rtx_insn **mode_jump | |
4401 | = mode == GOMP_DIM_VECTOR ? &vector_jump : &worker_jump; | |
4402 | rtx_insn **mode_label | |
4403 | = mode == GOMP_DIM_VECTOR ? &vector_label : &worker_label; | |
4404 | ||
4405 | if (!pred) | |
4406 | { | |
4407 | pred = gen_reg_rtx (BImode); | |
4408 | cfun->machine->axis_predicate[mode - GOMP_DIM_WORKER] = pred; | |
4409 | } | |
4410 | ||
4411 | rtx br; | |
4412 | if (mode == GOMP_DIM_VECTOR) | |
4413 | br = gen_br_true (pred, label); | |
4414 | else | |
4415 | br = gen_br_true_uni (pred, label); | |
4416 | if (neuter_start) | |
4417 | neuter_start = emit_insn_after (br, neuter_start); | |
4418 | else | |
4419 | neuter_start = emit_insn_before (br, head); | |
4420 | *mode_jump = neuter_start; | |
4421 | ||
4422 | LABEL_NUSES (label)++; | |
4423 | rtx_insn *label_insn; | |
4424 | if (tail_branch) | |
4425 | { | |
4426 | label_insn = emit_label_before (label, before); | |
4427 | before = label_insn; | |
4428 | } | |
4429 | else | |
4430 | { | |
4431 | label_insn = emit_label_after (label, tail); | |
4432 | if ((mode == GOMP_DIM_VECTOR || mode == GOMP_DIM_WORKER) | |
4433 | && CALL_P (tail) && find_reg_note (tail, REG_NORETURN, NULL)) | |
4434 | emit_insn_after (gen_exit (), label_insn); | |
4435 | } | |
4436 | ||
4437 | *mode_label = label_insn; | |
4438 | } | |
4439 | ||
4440 | /* Now deal with propagating the branch condition. */ | |
4441 | if (cond_branch) | |
4442 | { | |
4443 | rtx pvar = XEXP (XEXP (cond_branch, 0), 0); | |
4444 | ||
4445 | if (GOMP_DIM_MASK (GOMP_DIM_VECTOR) == mask | |
4446 | && nvptx_mach_vector_length () == PTX_WARP_SIZE) | |
4447 | { | |
4448 | /* Vector mode only, do a shuffle. */ | |
4449 | #if WORKAROUND_PTXJIT_BUG | |
4450 | /* The branch condition %rcond is propagated like this: | |
4451 | ||
4452 | { | |
4453 | .reg .u32 %x; | |
4454 | mov.u32 %x,%tid.x; | |
4455 | setp.ne.u32 %rnotvzero,%x,0; | |
4456 | } | |
4457 | ||
4458 | @%rnotvzero bra Lskip; | |
4459 | setp.<op>.<type> %rcond,op1,op2; | |
4460 | Lskip: | |
4461 | selp.u32 %rcondu32,1,0,%rcond; | |
4462 | shfl.idx.b32 %rcondu32,%rcondu32,0,31; | |
4463 | setp.ne.u32 %rcond,%rcondu32,0; | |
4464 | ||
4465 | There seems to be a bug in the ptx JIT compiler (observed at driver | |
4466 | version 381.22, at -O1 and higher for sm_61), that drops the shfl | |
4467 | unless %rcond is initialized to something before 'bra Lskip'. The | |
4468 | bug is not observed with ptxas from cuda 8.0.61. | |
4469 | ||
4470 | It is true that the code is non-trivial: at Lskip, %rcond is | |
4471 | uninitialized in threads 1-31, and after the selp the same holds | |
4472 | for %rcondu32. But shfl propagates the defined value in thread 0 | |
4473 | to threads 1-31, so after the shfl %rcondu32 is defined in threads | |
4474 | 0-31, and after the setp.ne %rcond is defined in threads 0-31. | |
4475 | ||
4476 | There is nothing in the PTX spec to suggest that this is wrong, or | |
4477 | to explain why the extra initialization is needed. So, we classify | |
4478 | it as a JIT bug, and the extra initialization as workaround: | |
4479 | ||
4480 | { | |
4481 | .reg .u32 %x; | |
4482 | mov.u32 %x,%tid.x; | |
4483 | setp.ne.u32 %rnotvzero,%x,0; | |
4484 | } | |
4485 | ||
4486 | +.reg .pred %rcond2; | |
4487 | +setp.eq.u32 %rcond2, 1, 0; | |
4488 | ||
4489 | @%rnotvzero bra Lskip; | |
4490 | setp.<op>.<type> %rcond,op1,op2; | |
4491 | +mov.pred %rcond2, %rcond; | |
4492 | Lskip: | |
4493 | +mov.pred %rcond, %rcond2; | |
4494 | selp.u32 %rcondu32,1,0,%rcond; | |
4495 | shfl.idx.b32 %rcondu32,%rcondu32,0,31; | |
4496 | setp.ne.u32 %rcond,%rcondu32,0; | |
4497 | */ | |
4498 | rtx_insn *label = PREV_INSN (tail); | |
4499 | gcc_assert (label && LABEL_P (label)); | |
4500 | rtx tmp = gen_reg_rtx (BImode); | |
4501 | emit_insn_before (gen_movbi (tmp, const0_rtx), | |
4502 | bb_first_real_insn (from)); | |
4503 | emit_insn_before (gen_rtx_SET (tmp, pvar), label); | |
4504 | emit_insn_before (gen_rtx_SET (pvar, tmp), tail); | |
4505 | #endif | |
4506 | emit_insn_before (nvptx_gen_warp_bcast (pvar), tail); | |
4507 | } | |
4508 | else | |
4509 | { | |
4510 | /* Includes worker mode, do spill & fill. By construction | |
4511 | we should never have worker mode only. */ | |
4512 | broadcast_data_t data; | |
4513 | unsigned size = GET_MODE_SIZE (SImode); | |
4514 | bool vector = (GOMP_DIM_MASK (GOMP_DIM_VECTOR) == mask) != 0; | |
4515 | bool worker = (GOMP_DIM_MASK (GOMP_DIM_WORKER) == mask) != 0; | |
4516 | rtx barrier = GEN_INT (0); | |
4517 | int threads = 0; | |
4518 | ||
4519 | data.base = oacc_bcast_sym; | |
4520 | data.ptr = 0; | |
4521 | ||
4522 | bool use_partitioning_p = (vector && !worker | |
4523 | && nvptx_mach_max_workers () > 1 | |
4524 | && cfun->machine->bcast_partition); | |
4525 | if (use_partitioning_p) | |
4526 | { | |
4527 | data.base = cfun->machine->bcast_partition; | |
4528 | barrier = cfun->machine->sync_bar; | |
4529 | threads = nvptx_mach_vector_length (); | |
4530 | } | |
4531 | gcc_assert (data.base != NULL); | |
4532 | gcc_assert (barrier); | |
4533 | ||
4534 | unsigned int psize = ROUND_UP (size, oacc_bcast_align); | |
4535 | unsigned int pnum = (nvptx_mach_vector_length () > PTX_WARP_SIZE | |
4536 | ? nvptx_mach_max_workers () + 1 | |
4537 | : 1); | |
4538 | ||
4539 | oacc_bcast_partition = MAX (oacc_bcast_partition, psize); | |
4540 | oacc_bcast_size = MAX (oacc_bcast_size, psize * pnum); | |
4541 | ||
4542 | data.offset = 0; | |
4543 | emit_insn_before (nvptx_gen_shared_bcast (pvar, PM_read, 0, &data, | |
4544 | vector), | |
4545 | before); | |
4546 | ||
4547 | /* Barrier so other workers can see the write. */ | |
4548 | emit_insn_before (nvptx_cta_sync (barrier, threads), tail); | |
4549 | data.offset = 0; | |
4550 | emit_insn_before (nvptx_gen_shared_bcast (pvar, PM_write, 0, &data, | |
4551 | vector), | |
4552 | tail); | |
4553 | /* This barrier is needed to avoid worker zero clobbering | |
4554 | the broadcast buffer before all the other workers have | |
4555 | had a chance to read this instance of it. */ | |
4556 | emit_insn_before (nvptx_cta_sync (barrier, threads), tail); | |
4557 | } | |
4558 | ||
4559 | extract_insn (tail); | |
4560 | rtx unsp = gen_rtx_UNSPEC (BImode, gen_rtvec (1, pvar), | |
4561 | UNSPEC_BR_UNIFIED); | |
4562 | validate_change (tail, recog_data.operand_loc[0], unsp, false); | |
4563 | } | |
4564 | ||
4565 | bool seen_label = verify_neutering_jumps (from, vector_jump, worker_jump, | |
4566 | vector_label, worker_label); | |
4567 | if (!seen_label) | |
4568 | verify_neutering_labels (to, vector_label, worker_label); | |
4569 | } | |
4570 | ||
4571 | /* PAR is a parallel that is being skipped in its entirety according to | |
4572 | MASK. Treat this as skipping a superblock starting at forked | |
4573 | and ending at joining. */ | |
4574 | ||
4575 | static void | |
4576 | nvptx_skip_par (unsigned mask, parallel *par) | |
4577 | { | |
4578 | basic_block tail = par->join_block; | |
4579 | gcc_assert (tail->preds->length () == 1); | |
4580 | ||
4581 | basic_block pre_tail = (*tail->preds)[0]->src; | |
4582 | gcc_assert (pre_tail->succs->length () == 1); | |
4583 | ||
4584 | nvptx_single (mask, par->forked_block, pre_tail); | |
4585 | } | |
4586 | ||
4587 | /* If PAR has a single inner parallel and PAR itself only contains | |
4588 | empty entry and exit blocks, swallow the inner PAR. */ | |
4589 | ||
4590 | static void | |
4591 | nvptx_optimize_inner (parallel *par) | |
4592 | { | |
4593 | parallel *inner = par->inner; | |
4594 | ||
4595 | /* We mustn't be the outer dummy par. */ | |
4596 | if (!par->mask) | |
4597 | return; | |
4598 | ||
4599 | /* We must have a single inner par. */ | |
4600 | if (!inner || inner->next) | |
4601 | return; | |
4602 | ||
4603 | /* We must only contain 2 blocks ourselves -- the head and tail of | |
4604 | the inner par. */ | |
4605 | if (par->blocks.length () != 2) | |
4606 | return; | |
4607 | ||
4608 | /* We must be disjoint partitioning. As we only have vector and | |
4609 | worker partitioning, this is sufficient to guarantee the pars | |
4610 | have adjacent partitioning. */ | |
4611 | if ((par->mask & inner->mask) & (GOMP_DIM_MASK (GOMP_DIM_MAX) - 1)) | |
4612 | /* This indicates malformed code generation. */ | |
4613 | return; | |
4614 | ||
4615 | /* The outer forked insn should be immediately followed by the inner | |
4616 | fork insn. */ | |
4617 | rtx_insn *forked = par->forked_insn; | |
4618 | rtx_insn *fork = BB_END (par->forked_block); | |
4619 | ||
4620 | if (NEXT_INSN (forked) != fork) | |
4621 | return; | |
4622 | gcc_checking_assert (recog_memoized (fork) == CODE_FOR_nvptx_fork); | |
4623 | ||
4624 | /* The outer joining insn must immediately follow the inner join | |
4625 | insn. */ | |
4626 | rtx_insn *joining = par->joining_insn; | |
4627 | rtx_insn *join = inner->join_insn; | |
4628 | if (NEXT_INSN (join) != joining) | |
4629 | return; | |
4630 | ||
4631 | /* Preconditions met. Swallow the inner par. */ | |
4632 | if (dump_file) | |
4633 | fprintf (dump_file, "Merging loop %x [%d,%d] into %x [%d,%d]\n", | |
4634 | inner->mask, inner->forked_block->index, | |
4635 | inner->join_block->index, | |
4636 | par->mask, par->forked_block->index, par->join_block->index); | |
4637 | ||
4638 | par->mask |= inner->mask & (GOMP_DIM_MASK (GOMP_DIM_MAX) - 1); | |
4639 | ||
4640 | par->blocks.reserve (inner->blocks.length ()); | |
4641 | while (inner->blocks.length ()) | |
4642 | par->blocks.quick_push (inner->blocks.pop ()); | |
4643 | ||
4644 | par->inner = inner->inner; | |
4645 | inner->inner = NULL; | |
4646 | ||
4647 | delete inner; | |
4648 | } | |
4649 | ||
4650 | /* Process the parallel PAR and all its contained | |
4651 | parallels. We do everything but the neutering. Return mask of | |
4652 | partitioned modes used within this parallel. */ | |
4653 | ||
4654 | static unsigned | |
4655 | nvptx_process_pars (parallel *par) | |
4656 | { | |
4657 | if (nvptx_optimize) | |
4658 | nvptx_optimize_inner (par); | |
4659 | ||
4660 | unsigned inner_mask = par->mask; | |
4661 | ||
4662 | /* Do the inner parallels first. */ | |
4663 | if (par->inner) | |
4664 | { | |
4665 | par->inner_mask = nvptx_process_pars (par->inner); | |
4666 | inner_mask |= par->inner_mask; | |
4667 | } | |
4668 | ||
4669 | bool is_call = (par->mask & GOMP_DIM_MASK (GOMP_DIM_MAX)) != 0; | |
4670 | bool worker = (par->mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)); | |
4671 | bool large_vector = ((par->mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR)) | |
4672 | && nvptx_mach_vector_length () > PTX_WARP_SIZE); | |
4673 | ||
4674 | if (worker || large_vector) | |
4675 | { | |
4676 | nvptx_shared_propagate (false, is_call, par->forked_block, | |
4677 | par->forked_insn, !worker); | |
4678 | bool no_prop_p | |
4679 | = nvptx_shared_propagate (true, is_call, par->forked_block, | |
4680 | par->fork_insn, !worker); | |
4681 | bool empty_loop_p | |
4682 | = !is_call && (NEXT_INSN (par->forked_insn) | |
4683 | && NEXT_INSN (par->forked_insn) == par->joining_insn); | |
4684 | rtx barrier = GEN_INT (0); | |
4685 | int threads = 0; | |
4686 | ||
4687 | if (!worker && cfun->machine->sync_bar) | |
4688 | { | |
4689 | barrier = cfun->machine->sync_bar; | |
4690 | threads = nvptx_mach_vector_length (); | |
4691 | } | |
4692 | ||
4693 | if (no_prop_p && empty_loop_p) | |
4694 | ; | |
4695 | else if (no_prop_p && is_call) | |
4696 | ; | |
4697 | else | |
4698 | { | |
4699 | /* Insert begin and end synchronizations. */ | |
4700 | emit_insn_before (nvptx_cta_sync (barrier, threads), | |
4701 | par->forked_insn); | |
4702 | emit_insn_before (nvptx_cta_sync (barrier, threads), par->join_insn); | |
4703 | } | |
4704 | } | |
4705 | else if (par->mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR)) | |
4706 | nvptx_warp_propagate (is_call, par->forked_block, par->forked_insn); | |
4707 | ||
4708 | /* Now do siblings. */ | |
4709 | if (par->next) | |
4710 | inner_mask |= nvptx_process_pars (par->next); | |
4711 | return inner_mask; | |
4712 | } | |
4713 | ||
4714 | /* Neuter the parallel described by PAR. We recurse in depth-first | |
4715 | order. MODES are the partitioning of the execution and OUTER is | |
4716 | the partitioning of the parallels we are contained in. */ | |
4717 | ||
4718 | static void | |
4719 | nvptx_neuter_pars (parallel *par, unsigned modes, unsigned outer) | |
4720 | { | |
4721 | unsigned me = (par->mask | |
4722 | & (GOMP_DIM_MASK (GOMP_DIM_WORKER) | |
4723 | | GOMP_DIM_MASK (GOMP_DIM_VECTOR))); | |
4724 | unsigned skip_mask = 0, neuter_mask = 0; | |
4725 | ||
4726 | if (par->inner) | |
4727 | nvptx_neuter_pars (par->inner, modes, outer | me); | |
4728 | ||
4729 | for (unsigned mode = GOMP_DIM_WORKER; mode <= GOMP_DIM_VECTOR; mode++) | |
4730 | { | |
4731 | if ((outer | me) & GOMP_DIM_MASK (mode)) | |
4732 | {} /* Mode is partitioned: no neutering. */ | |
4733 | else if (!(modes & GOMP_DIM_MASK (mode))) | |
4734 | {} /* Mode is not used: nothing to do. */ | |
4735 | else if (par->inner_mask & GOMP_DIM_MASK (mode) | |
4736 | || !par->forked_insn) | |
4737 | /* Partitioned in inner parallels, or we're not a partitioned | |
4738 | at all: neuter individual blocks. */ | |
4739 | neuter_mask |= GOMP_DIM_MASK (mode); | |
4740 | else if (!par->parent || !par->parent->forked_insn | |
4741 | || par->parent->inner_mask & GOMP_DIM_MASK (mode)) | |
4742 | /* Parent isn't a parallel or contains this paralleling: skip | |
4743 | parallel at this level. */ | |
4744 | skip_mask |= GOMP_DIM_MASK (mode); | |
4745 | else | |
4746 | {} /* Parent will skip this parallel itself. */ | |
4747 | } | |
4748 | ||
4749 | if (neuter_mask) | |
4750 | { | |
4751 | int ix, len; | |
4752 | ||
4753 | if (nvptx_optimize) | |
4754 | { | |
4755 | /* Neuter whole SESE regions. */ | |
4756 | bb_pair_vec_t regions; | |
4757 | ||
4758 | nvptx_find_sese (par->blocks, regions); | |
4759 | len = regions.length (); | |
4760 | for (ix = 0; ix != len; ix++) | |
4761 | { | |
4762 | basic_block from = regions[ix].first; | |
4763 | basic_block to = regions[ix].second; | |
4764 | ||
4765 | if (from) | |
4766 | nvptx_single (neuter_mask, from, to); | |
4767 | else | |
4768 | gcc_assert (!to); | |
4769 | } | |
4770 | } | |
4771 | else | |
4772 | { | |
4773 | /* Neuter each BB individually. */ | |
4774 | len = par->blocks.length (); | |
4775 | for (ix = 0; ix != len; ix++) | |
4776 | { | |
4777 | basic_block block = par->blocks[ix]; | |
4778 | ||
4779 | nvptx_single (neuter_mask, block, block); | |
4780 | } | |
4781 | } | |
4782 | } | |
4783 | ||
4784 | if (skip_mask) | |
4785 | nvptx_skip_par (skip_mask, par); | |
4786 | ||
4787 | if (par->next) | |
4788 | nvptx_neuter_pars (par->next, modes, outer); | |
4789 | } | |
4790 | ||
4791 | static void | |
4792 | populate_offload_attrs (offload_attrs *oa) | |
4793 | { | |
4794 | tree attr = oacc_get_fn_attrib (current_function_decl); | |
4795 | tree dims = TREE_VALUE (attr); | |
4796 | unsigned ix; | |
4797 | ||
4798 | oa->mask = 0; | |
4799 | ||
4800 | for (ix = 0; ix != GOMP_DIM_MAX; ix++, dims = TREE_CHAIN (dims)) | |
4801 | { | |
4802 | tree t = TREE_VALUE (dims); | |
4803 | int size = (t == NULL_TREE) ? -1 : TREE_INT_CST_LOW (t); | |
4804 | tree allowed = TREE_PURPOSE (dims); | |
4805 | ||
4806 | if (size != 1 && !(allowed && integer_zerop (allowed))) | |
4807 | oa->mask |= GOMP_DIM_MASK (ix); | |
4808 | ||
4809 | switch (ix) | |
4810 | { | |
4811 | case GOMP_DIM_GANG: | |
4812 | oa->num_gangs = size; | |
4813 | break; | |
4814 | ||
4815 | case GOMP_DIM_WORKER: | |
4816 | oa->num_workers = size; | |
4817 | break; | |
4818 | ||
4819 | case GOMP_DIM_VECTOR: | |
4820 | oa->vector_length = size; | |
4821 | break; | |
4822 | } | |
4823 | } | |
4824 | } | |
4825 | ||
4826 | #if WORKAROUND_PTXJIT_BUG_2 | |
4827 | /* Variant of pc_set that only requires JUMP_P (INSN) if STRICT. This variant | |
4828 | is needed in the nvptx target because the branches generated for | |
4829 | parititioning are NONJUMP_INSN_P, not JUMP_P. */ | |
4830 | ||
4831 | static rtx | |
4832 | nvptx_pc_set (const rtx_insn *insn, bool strict = true) | |
4833 | { | |
4834 | rtx pat; | |
4835 | if ((strict && !JUMP_P (insn)) | |
4836 | || (!strict && !INSN_P (insn))) | |
4837 | return NULL_RTX; | |
4838 | pat = PATTERN (insn); | |
4839 | ||
4840 | /* The set is allowed to appear either as the insn pattern or | |
4841 | the first set in a PARALLEL. */ | |
4842 | if (GET_CODE (pat) == PARALLEL) | |
4843 | pat = XVECEXP (pat, 0, 0); | |
4844 | if (GET_CODE (pat) == SET && GET_CODE (SET_DEST (pat)) == PC) | |
4845 | return pat; | |
4846 | ||
4847 | return NULL_RTX; | |
4848 | } | |
4849 | ||
4850 | /* Variant of condjump_label that only requires JUMP_P (INSN) if STRICT. */ | |
4851 | ||
4852 | static rtx | |
4853 | nvptx_condjump_label (const rtx_insn *insn, bool strict = true) | |
4854 | { | |
4855 | rtx x = nvptx_pc_set (insn, strict); | |
4856 | ||
4857 | if (!x) | |
4858 | return NULL_RTX; | |
4859 | x = SET_SRC (x); | |
4860 | if (GET_CODE (x) == LABEL_REF) | |
4861 | return x; | |
4862 | if (GET_CODE (x) != IF_THEN_ELSE) | |
4863 | return NULL_RTX; | |
4864 | if (XEXP (x, 2) == pc_rtx && GET_CODE (XEXP (x, 1)) == LABEL_REF) | |
4865 | return XEXP (x, 1); | |
4866 | if (XEXP (x, 1) == pc_rtx && GET_CODE (XEXP (x, 2)) == LABEL_REF) | |
4867 | return XEXP (x, 2); | |
4868 | return NULL_RTX; | |
4869 | } | |
4870 | ||
4871 | /* Insert a dummy ptx insn when encountering a branch to a label with no ptx | |
4872 | insn inbetween the branch and the label. This works around a JIT bug | |
4873 | observed at driver version 384.111, at -O0 for sm_50. */ | |
4874 | ||
4875 | static void | |
4876 | prevent_branch_around_nothing (void) | |
4877 | { | |
4878 | rtx_insn *seen_label = NULL; | |
4879 | for (rtx_insn *insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
4880 | { | |
4881 | if (INSN_P (insn) && condjump_p (insn)) | |
4882 | { | |
4883 | seen_label = label_ref_label (nvptx_condjump_label (insn, false)); | |
4884 | continue; | |
4885 | } | |
4886 | ||
4887 | if (seen_label == NULL) | |
4888 | continue; | |
4889 | ||
4890 | if (NOTE_P (insn) || DEBUG_INSN_P (insn)) | |
4891 | continue; | |
4892 | ||
4893 | if (INSN_P (insn)) | |
4894 | switch (recog_memoized (insn)) | |
4895 | { | |
4896 | case CODE_FOR_nvptx_fork: | |
4897 | case CODE_FOR_nvptx_forked: | |
4898 | case CODE_FOR_nvptx_joining: | |
4899 | case CODE_FOR_nvptx_join: | |
4900 | continue; | |
4901 | default: | |
4902 | seen_label = NULL; | |
4903 | continue; | |
4904 | } | |
4905 | ||
4906 | if (LABEL_P (insn) && insn == seen_label) | |
4907 | emit_insn_before (gen_fake_nop (), insn); | |
4908 | ||
4909 | seen_label = NULL; | |
4910 | } | |
4911 | } | |
4912 | #endif | |
4913 | ||
4914 | #ifdef WORKAROUND_PTXJIT_BUG_3 | |
4915 | /* Insert two membar.cta insns inbetween two subsequent bar.sync insns. This | |
4916 | works around a hang observed at driver version 390.48 for sm_50. */ | |
4917 | ||
4918 | static void | |
4919 | workaround_barsyncs (void) | |
4920 | { | |
4921 | bool seen_barsync = false; | |
4922 | for (rtx_insn *insn = get_insns (); insn; insn = NEXT_INSN (insn)) | |
4923 | { | |
4924 | if (INSN_P (insn) && recog_memoized (insn) == CODE_FOR_nvptx_barsync) | |
4925 | { | |
4926 | if (seen_barsync) | |
4927 | { | |
4928 | emit_insn_before (gen_nvptx_membar_cta (), insn); | |
4929 | emit_insn_before (gen_nvptx_membar_cta (), insn); | |
4930 | } | |
4931 | ||
4932 | seen_barsync = true; | |
4933 | continue; | |
4934 | } | |
4935 | ||
4936 | if (!seen_barsync) | |
4937 | continue; | |
4938 | ||
4939 | if (NOTE_P (insn) || DEBUG_INSN_P (insn)) | |
4940 | continue; | |
4941 | else if (INSN_P (insn)) | |
4942 | switch (recog_memoized (insn)) | |
4943 | { | |
4944 | case CODE_FOR_nvptx_fork: | |
4945 | case CODE_FOR_nvptx_forked: | |
4946 | case CODE_FOR_nvptx_joining: | |
4947 | case CODE_FOR_nvptx_join: | |
4948 | continue; | |
4949 | default: | |
4950 | break; | |
4951 | } | |
4952 | ||
4953 | seen_barsync = false; | |
4954 | } | |
4955 | } | |
4956 | #endif | |
4957 | ||
4958 | /* PTX-specific reorganization | |
4959 | - Split blocks at fork and join instructions | |
4960 | - Compute live registers | |
4961 | - Mark now-unused registers, so function begin doesn't declare | |
4962 | unused registers. | |
4963 | - Insert state propagation when entering partitioned mode | |
4964 | - Insert neutering instructions when in single mode | |
4965 | - Replace subregs with suitable sequences. | |
4966 | */ | |
4967 | ||
4968 | static void | |
4969 | nvptx_reorg (void) | |
4970 | { | |
4971 | /* We are freeing block_for_insn in the toplev to keep compatibility | |
4972 | with old MDEP_REORGS that are not CFG based. Recompute it now. */ | |
4973 | compute_bb_for_insn (); | |
4974 | ||
4975 | thread_prologue_and_epilogue_insns (); | |
4976 | ||
4977 | /* Split blocks and record interesting unspecs. */ | |
4978 | bb_insn_map_t bb_insn_map; | |
4979 | ||
4980 | nvptx_split_blocks (&bb_insn_map); | |
4981 | ||
4982 | /* Compute live regs */ | |
4983 | df_clear_flags (DF_LR_RUN_DCE); | |
4984 | df_set_flags (DF_NO_INSN_RESCAN | DF_NO_HARD_REGS); | |
4985 | df_live_add_problem (); | |
4986 | df_live_set_all_dirty (); | |
4987 | df_analyze (); | |
4988 | regstat_init_n_sets_and_refs (); | |
4989 | ||
4990 | if (dump_file) | |
4991 | df_dump (dump_file); | |
4992 | ||
4993 | /* Mark unused regs as unused. */ | |
4994 | int max_regs = max_reg_num (); | |
4995 | for (int i = LAST_VIRTUAL_REGISTER + 1; i < max_regs; i++) | |
4996 | if (REG_N_SETS (i) == 0 && REG_N_REFS (i) == 0) | |
4997 | regno_reg_rtx[i] = const0_rtx; | |
4998 | ||
4999 | /* Determine launch dimensions of the function. If it is not an | |
5000 | offloaded function (i.e. this is a regular compiler), the | |
5001 | function has no neutering. */ | |
5002 | tree attr = oacc_get_fn_attrib (current_function_decl); | |
5003 | if (attr) | |
5004 | { | |
5005 | /* If we determined this mask before RTL expansion, we could | |
5006 | elide emission of some levels of forks and joins. */ | |
5007 | offload_attrs oa; | |
5008 | ||
5009 | populate_offload_attrs (&oa); | |
5010 | ||
5011 | /* If there is worker neutering, there must be vector | |
5012 | neutering. Otherwise the hardware will fail. */ | |
5013 | gcc_assert (!(oa.mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)) | |
5014 | || (oa.mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR))); | |
5015 | ||
5016 | /* Discover & process partitioned regions. */ | |
5017 | parallel *pars = nvptx_discover_pars (&bb_insn_map); | |
5018 | nvptx_process_pars (pars); | |
5019 | nvptx_neuter_pars (pars, oa.mask, 0); | |
5020 | delete pars; | |
5021 | } | |
5022 | ||
5023 | /* Replace subregs. */ | |
5024 | nvptx_reorg_subreg (); | |
5025 | ||
5026 | if (TARGET_UNIFORM_SIMT) | |
5027 | nvptx_reorg_uniform_simt (); | |
5028 | ||
5029 | #if WORKAROUND_PTXJIT_BUG_2 | |
5030 | prevent_branch_around_nothing (); | |
5031 | #endif | |
5032 | ||
5033 | #ifdef WORKAROUND_PTXJIT_BUG_3 | |
5034 | workaround_barsyncs (); | |
5035 | #endif | |
5036 | ||
5037 | regstat_free_n_sets_and_refs (); | |
5038 | ||
5039 | df_finish_pass (true); | |
5040 | } | |
5041 | \f | |
5042 | /* Handle a "kernel" attribute; arguments as in | |
5043 | struct attribute_spec.handler. */ | |
5044 | ||
5045 | static tree | |
5046 | nvptx_handle_kernel_attribute (tree *node, tree name, tree ARG_UNUSED (args), | |
5047 | int ARG_UNUSED (flags), bool *no_add_attrs) | |
5048 | { | |
5049 | tree decl = *node; | |
5050 | ||
5051 | if (TREE_CODE (decl) != FUNCTION_DECL) | |
5052 | { | |
5053 | error ("%qE attribute only applies to functions", name); | |
5054 | *no_add_attrs = true; | |
5055 | } | |
5056 | else if (!VOID_TYPE_P (TREE_TYPE (TREE_TYPE (decl)))) | |
5057 | { | |
5058 | error ("%qE attribute requires a void return type", name); | |
5059 | *no_add_attrs = true; | |
5060 | } | |
5061 | ||
5062 | return NULL_TREE; | |
5063 | } | |
5064 | ||
5065 | /* Handle a "shared" attribute; arguments as in | |
5066 | struct attribute_spec.handler. */ | |
5067 | ||
5068 | static tree | |
5069 | nvptx_handle_shared_attribute (tree *node, tree name, tree ARG_UNUSED (args), | |
5070 | int ARG_UNUSED (flags), bool *no_add_attrs) | |
5071 | { | |
5072 | tree decl = *node; | |
5073 | ||
5074 | if (TREE_CODE (decl) != VAR_DECL) | |
5075 | { | |
5076 | error ("%qE attribute only applies to variables", name); | |
5077 | *no_add_attrs = true; | |
5078 | } | |
5079 | else if (!(TREE_PUBLIC (decl) || TREE_STATIC (decl))) | |
5080 | { | |
5081 | error ("%qE attribute not allowed with auto storage class", name); | |
5082 | *no_add_attrs = true; | |
5083 | } | |
5084 | ||
5085 | return NULL_TREE; | |
5086 | } | |
5087 | ||
5088 | /* Table of valid machine attributes. */ | |
5089 | static const struct attribute_spec nvptx_attribute_table[] = | |
5090 | { | |
5091 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, | |
5092 | affects_type_identity, handler, exclude } */ | |
5093 | { "kernel", 0, 0, true, false, false, false, nvptx_handle_kernel_attribute, | |
5094 | NULL }, | |
5095 | { "shared", 0, 0, true, false, false, false, nvptx_handle_shared_attribute, | |
5096 | NULL }, | |
5097 | { NULL, 0, 0, false, false, false, false, NULL, NULL } | |
5098 | }; | |
5099 | \f | |
5100 | /* Limit vector alignments to BIGGEST_ALIGNMENT. */ | |
5101 | ||
5102 | static HOST_WIDE_INT | |
5103 | nvptx_vector_alignment (const_tree type) | |
5104 | { | |
5105 | HOST_WIDE_INT align = tree_to_shwi (TYPE_SIZE (type)); | |
5106 | ||
5107 | return MIN (align, BIGGEST_ALIGNMENT); | |
5108 | } | |
5109 | ||
5110 | /* Indicate that INSN cannot be duplicated. */ | |
5111 | ||
5112 | static bool | |
5113 | nvptx_cannot_copy_insn_p (rtx_insn *insn) | |
5114 | { | |
5115 | switch (recog_memoized (insn)) | |
5116 | { | |
5117 | case CODE_FOR_nvptx_shufflesi: | |
5118 | case CODE_FOR_nvptx_shufflesf: | |
5119 | case CODE_FOR_nvptx_barsync: | |
5120 | case CODE_FOR_nvptx_fork: | |
5121 | case CODE_FOR_nvptx_forked: | |
5122 | case CODE_FOR_nvptx_joining: | |
5123 | case CODE_FOR_nvptx_join: | |
5124 | return true; | |
5125 | default: | |
5126 | return false; | |
5127 | } | |
5128 | } | |
5129 | ||
5130 | /* Section anchors do not work. Initialization for flag_section_anchor | |
5131 | probes the existence of the anchoring target hooks and prevents | |
5132 | anchoring if they don't exist. However, we may be being used with | |
5133 | a host-side compiler that does support anchoring, and hence see | |
5134 | the anchor flag set (as it's not recalculated). So provide an | |
5135 | implementation denying anchoring. */ | |
5136 | ||
5137 | static bool | |
5138 | nvptx_use_anchors_for_symbol_p (const_rtx ARG_UNUSED (a)) | |
5139 | { | |
5140 | return false; | |
5141 | } | |
5142 | \f | |
5143 | /* Record a symbol for mkoffload to enter into the mapping table. */ | |
5144 | ||
5145 | static void | |
5146 | nvptx_record_offload_symbol (tree decl) | |
5147 | { | |
5148 | switch (TREE_CODE (decl)) | |
5149 | { | |
5150 | case VAR_DECL: | |
5151 | fprintf (asm_out_file, "//:VAR_MAP \"%s\"\n", | |
5152 | IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl))); | |
5153 | break; | |
5154 | ||
5155 | case FUNCTION_DECL: | |
5156 | { | |
5157 | tree attr = oacc_get_fn_attrib (decl); | |
5158 | /* OpenMP offloading does not set this attribute. */ | |
5159 | tree dims = attr ? TREE_VALUE (attr) : NULL_TREE; | |
5160 | ||
5161 | fprintf (asm_out_file, "//:FUNC_MAP \"%s\"", | |
5162 | IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl))); | |
5163 | ||
5164 | for (; dims; dims = TREE_CHAIN (dims)) | |
5165 | { | |
5166 | int size = TREE_INT_CST_LOW (TREE_VALUE (dims)); | |
5167 | ||
5168 | gcc_assert (!TREE_PURPOSE (dims)); | |
5169 | fprintf (asm_out_file, ", %#x", size); | |
5170 | } | |
5171 | ||
5172 | fprintf (asm_out_file, "\n"); | |
5173 | } | |
5174 | break; | |
5175 | ||
5176 | default: | |
5177 | gcc_unreachable (); | |
5178 | } | |
5179 | } | |
5180 | ||
5181 | /* Implement TARGET_ASM_FILE_START. Write the kinds of things ptxas expects | |
5182 | at the start of a file. */ | |
5183 | ||
5184 | static void | |
5185 | nvptx_file_start (void) | |
5186 | { | |
5187 | fputs ("// BEGIN PREAMBLE\n", asm_out_file); | |
5188 | fputs ("\t.version\t3.1\n", asm_out_file); | |
5189 | if (TARGET_SM35) | |
5190 | fputs ("\t.target\tsm_35\n", asm_out_file); | |
5191 | else | |
5192 | fputs ("\t.target\tsm_30\n", asm_out_file); | |
5193 | fprintf (asm_out_file, "\t.address_size %d\n", GET_MODE_BITSIZE (Pmode)); | |
5194 | fputs ("// END PREAMBLE\n", asm_out_file); | |
5195 | } | |
5196 | ||
5197 | /* Emit a declaration for a worker and vector-level buffer in .shared | |
5198 | memory. */ | |
5199 | ||
5200 | static void | |
5201 | write_shared_buffer (FILE *file, rtx sym, unsigned align, unsigned size) | |
5202 | { | |
5203 | const char *name = XSTR (sym, 0); | |
5204 | ||
5205 | write_var_marker (file, true, false, name); | |
5206 | fprintf (file, ".shared .align %d .u8 %s[%d];\n", | |
5207 | align, name, size); | |
5208 | } | |
5209 | ||
5210 | /* Write out the function declarations we've collected and declare storage | |
5211 | for the broadcast buffer. */ | |
5212 | ||
5213 | static void | |
5214 | nvptx_file_end (void) | |
5215 | { | |
5216 | hash_table<tree_hasher>::iterator iter; | |
5217 | tree decl; | |
5218 | FOR_EACH_HASH_TABLE_ELEMENT (*needed_fndecls_htab, decl, tree, iter) | |
5219 | nvptx_record_fndecl (decl); | |
5220 | fputs (func_decls.str().c_str(), asm_out_file); | |
5221 | ||
5222 | if (oacc_bcast_size) | |
5223 | write_shared_buffer (asm_out_file, oacc_bcast_sym, | |
5224 | oacc_bcast_align, oacc_bcast_size); | |
5225 | ||
5226 | if (worker_red_size) | |
5227 | write_shared_buffer (asm_out_file, worker_red_sym, | |
5228 | worker_red_align, worker_red_size); | |
5229 | ||
5230 | if (vector_red_size) | |
5231 | write_shared_buffer (asm_out_file, vector_red_sym, | |
5232 | vector_red_align, vector_red_size); | |
5233 | ||
5234 | if (need_softstack_decl) | |
5235 | { | |
5236 | write_var_marker (asm_out_file, false, true, "__nvptx_stacks"); | |
5237 | /* 32 is the maximum number of warps in a block. Even though it's an | |
5238 | external declaration, emit the array size explicitly; otherwise, it | |
5239 | may fail at PTX JIT time if the definition is later in link order. */ | |
5240 | fprintf (asm_out_file, ".extern .shared .u%d __nvptx_stacks[32];\n", | |
5241 | POINTER_SIZE); | |
5242 | } | |
5243 | if (need_unisimt_decl) | |
5244 | { | |
5245 | write_var_marker (asm_out_file, false, true, "__nvptx_uni"); | |
5246 | fprintf (asm_out_file, ".extern .shared .u32 __nvptx_uni[32];\n"); | |
5247 | } | |
5248 | } | |
5249 | ||
5250 | /* Expander for the shuffle builtins. */ | |
5251 | ||
5252 | static rtx | |
5253 | nvptx_expand_shuffle (tree exp, rtx target, machine_mode mode, int ignore) | |
5254 | { | |
5255 | if (ignore) | |
5256 | return target; | |
5257 | ||
5258 | rtx src = expand_expr (CALL_EXPR_ARG (exp, 0), | |
5259 | NULL_RTX, mode, EXPAND_NORMAL); | |
5260 | if (!REG_P (src)) | |
5261 | src = copy_to_mode_reg (mode, src); | |
5262 | ||
5263 | rtx idx = expand_expr (CALL_EXPR_ARG (exp, 1), | |
5264 | NULL_RTX, SImode, EXPAND_NORMAL); | |
5265 | rtx op = expand_expr (CALL_EXPR_ARG (exp, 2), | |
5266 | NULL_RTX, SImode, EXPAND_NORMAL); | |
5267 | ||
5268 | if (!REG_P (idx) && GET_CODE (idx) != CONST_INT) | |
5269 | idx = copy_to_mode_reg (SImode, idx); | |
5270 | ||
5271 | rtx pat = nvptx_gen_shuffle (target, src, idx, | |
5272 | (nvptx_shuffle_kind) INTVAL (op)); | |
5273 | if (pat) | |
5274 | emit_insn (pat); | |
5275 | ||
5276 | return target; | |
5277 | } | |
5278 | ||
5279 | const char * | |
5280 | nvptx_output_red_partition (rtx dst, rtx offset) | |
5281 | { | |
5282 | const char *zero_offset = "\t\tmov.u64\t%%r%d, %%r%d; // vred buffer\n"; | |
5283 | const char *with_offset = "\t\tadd.u64\t%%r%d, %%r%d, %d; // vred buffer\n"; | |
5284 | ||
5285 | if (offset == const0_rtx) | |
5286 | fprintf (asm_out_file, zero_offset, REGNO (dst), | |
5287 | REGNO (cfun->machine->red_partition)); | |
5288 | else | |
5289 | fprintf (asm_out_file, with_offset, REGNO (dst), | |
5290 | REGNO (cfun->machine->red_partition), UINTVAL (offset)); | |
5291 | ||
5292 | return ""; | |
5293 | } | |
5294 | ||
5295 | /* Shared-memory reduction address expander. */ | |
5296 | ||
5297 | static rtx | |
5298 | nvptx_expand_shared_addr (tree exp, rtx target, | |
5299 | machine_mode ARG_UNUSED (mode), int ignore, | |
5300 | int vector) | |
5301 | { | |
5302 | if (ignore) | |
5303 | return target; | |
5304 | ||
5305 | unsigned align = TREE_INT_CST_LOW (CALL_EXPR_ARG (exp, 2)); | |
5306 | unsigned offset = TREE_INT_CST_LOW (CALL_EXPR_ARG (exp, 0)); | |
5307 | unsigned size = TREE_INT_CST_LOW (CALL_EXPR_ARG (exp, 1)); | |
5308 | rtx addr = worker_red_sym; | |
5309 | ||
5310 | if (vector) | |
5311 | { | |
5312 | offload_attrs oa; | |
5313 | ||
5314 | populate_offload_attrs (&oa); | |
5315 | ||
5316 | unsigned int psize = ROUND_UP (size + offset, align); | |
5317 | unsigned int pnum = nvptx_mach_max_workers (); | |
5318 | vector_red_partition = MAX (vector_red_partition, psize); | |
5319 | vector_red_size = MAX (vector_red_size, psize * pnum); | |
5320 | vector_red_align = MAX (vector_red_align, align); | |
5321 | ||
5322 | if (cfun->machine->red_partition == NULL) | |
5323 | cfun->machine->red_partition = gen_reg_rtx (Pmode); | |
5324 | ||
5325 | addr = gen_reg_rtx (Pmode); | |
5326 | emit_insn (gen_nvptx_red_partition (addr, GEN_INT (offset))); | |
5327 | } | |
5328 | else | |
5329 | { | |
5330 | worker_red_align = MAX (worker_red_align, align); | |
5331 | worker_red_size = MAX (worker_red_size, size + offset); | |
5332 | ||
5333 | if (offset) | |
5334 | { | |
5335 | addr = gen_rtx_PLUS (Pmode, addr, GEN_INT (offset)); | |
5336 | addr = gen_rtx_CONST (Pmode, addr); | |
5337 | } | |
5338 | } | |
5339 | ||
5340 | emit_move_insn (target, addr); | |
5341 | return target; | |
5342 | } | |
5343 | ||
5344 | /* Expand the CMP_SWAP PTX builtins. We have our own versions that do | |
5345 | not require taking the address of any object, other than the memory | |
5346 | cell being operated on. */ | |
5347 | ||
5348 | static rtx | |
5349 | nvptx_expand_cmp_swap (tree exp, rtx target, | |
5350 | machine_mode ARG_UNUSED (m), int ARG_UNUSED (ignore)) | |
5351 | { | |
5352 | machine_mode mode = TYPE_MODE (TREE_TYPE (exp)); | |
5353 | ||
5354 | if (!target) | |
5355 | target = gen_reg_rtx (mode); | |
5356 | ||
5357 | rtx mem = expand_expr (CALL_EXPR_ARG (exp, 0), | |
5358 | NULL_RTX, Pmode, EXPAND_NORMAL); | |
5359 | rtx cmp = expand_expr (CALL_EXPR_ARG (exp, 1), | |
5360 | NULL_RTX, mode, EXPAND_NORMAL); | |
5361 | rtx src = expand_expr (CALL_EXPR_ARG (exp, 2), | |
5362 | NULL_RTX, mode, EXPAND_NORMAL); | |
5363 | rtx pat; | |
5364 | ||
5365 | mem = gen_rtx_MEM (mode, mem); | |
5366 | if (!REG_P (cmp)) | |
5367 | cmp = copy_to_mode_reg (mode, cmp); | |
5368 | if (!REG_P (src)) | |
5369 | src = copy_to_mode_reg (mode, src); | |
5370 | ||
5371 | if (mode == SImode) | |
5372 | pat = gen_atomic_compare_and_swapsi_1 (target, mem, cmp, src, const0_rtx); | |
5373 | else | |
5374 | pat = gen_atomic_compare_and_swapdi_1 (target, mem, cmp, src, const0_rtx); | |
5375 | ||
5376 | emit_insn (pat); | |
5377 | ||
5378 | return target; | |
5379 | } | |
5380 | ||
5381 | ||
5382 | /* Codes for all the NVPTX builtins. */ | |
5383 | enum nvptx_builtins | |
5384 | { | |
5385 | NVPTX_BUILTIN_SHUFFLE, | |
5386 | NVPTX_BUILTIN_SHUFFLELL, | |
5387 | NVPTX_BUILTIN_WORKER_ADDR, | |
5388 | NVPTX_BUILTIN_VECTOR_ADDR, | |
5389 | NVPTX_BUILTIN_CMP_SWAP, | |
5390 | NVPTX_BUILTIN_CMP_SWAPLL, | |
5391 | NVPTX_BUILTIN_MAX | |
5392 | }; | |
5393 | ||
5394 | static GTY(()) tree nvptx_builtin_decls[NVPTX_BUILTIN_MAX]; | |
5395 | ||
5396 | /* Return the NVPTX builtin for CODE. */ | |
5397 | ||
5398 | static tree | |
5399 | nvptx_builtin_decl (unsigned code, bool ARG_UNUSED (initialize_p)) | |
5400 | { | |
5401 | if (code >= NVPTX_BUILTIN_MAX) | |
5402 | return error_mark_node; | |
5403 | ||
5404 | return nvptx_builtin_decls[code]; | |
5405 | } | |
5406 | ||
5407 | /* Set up all builtin functions for this target. */ | |
5408 | ||
5409 | static void | |
5410 | nvptx_init_builtins (void) | |
5411 | { | |
5412 | #define DEF(ID, NAME, T) \ | |
5413 | (nvptx_builtin_decls[NVPTX_BUILTIN_ ## ID] \ | |
5414 | = add_builtin_function ("__builtin_nvptx_" NAME, \ | |
5415 | build_function_type_list T, \ | |
5416 | NVPTX_BUILTIN_ ## ID, BUILT_IN_MD, NULL, NULL)) | |
5417 | #define ST sizetype | |
5418 | #define UINT unsigned_type_node | |
5419 | #define LLUINT long_long_unsigned_type_node | |
5420 | #define PTRVOID ptr_type_node | |
5421 | ||
5422 | DEF (SHUFFLE, "shuffle", (UINT, UINT, UINT, UINT, NULL_TREE)); | |
5423 | DEF (SHUFFLELL, "shufflell", (LLUINT, LLUINT, UINT, UINT, NULL_TREE)); | |
5424 | DEF (WORKER_ADDR, "worker_addr", | |
5425 | (PTRVOID, ST, UINT, UINT, NULL_TREE)); | |
5426 | DEF (VECTOR_ADDR, "vector_addr", | |
5427 | (PTRVOID, ST, UINT, UINT, NULL_TREE)); | |
5428 | DEF (CMP_SWAP, "cmp_swap", (UINT, PTRVOID, UINT, UINT, NULL_TREE)); | |
5429 | DEF (CMP_SWAPLL, "cmp_swapll", (LLUINT, PTRVOID, LLUINT, LLUINT, NULL_TREE)); | |
5430 | ||
5431 | #undef DEF | |
5432 | #undef ST | |
5433 | #undef UINT | |
5434 | #undef LLUINT | |
5435 | #undef PTRVOID | |
5436 | } | |
5437 | ||
5438 | /* Expand an expression EXP that calls a built-in function, | |
5439 | with result going to TARGET if that's convenient | |
5440 | (and in mode MODE if that's convenient). | |
5441 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
5442 | IGNORE is nonzero if the value is to be ignored. */ | |
5443 | ||
5444 | static rtx | |
5445 | nvptx_expand_builtin (tree exp, rtx target, rtx ARG_UNUSED (subtarget), | |
5446 | machine_mode mode, int ignore) | |
5447 | { | |
5448 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
5449 | switch (DECL_MD_FUNCTION_CODE (fndecl)) | |
5450 | { | |
5451 | case NVPTX_BUILTIN_SHUFFLE: | |
5452 | case NVPTX_BUILTIN_SHUFFLELL: | |
5453 | return nvptx_expand_shuffle (exp, target, mode, ignore); | |
5454 | ||
5455 | case NVPTX_BUILTIN_WORKER_ADDR: | |
5456 | return nvptx_expand_shared_addr (exp, target, mode, ignore, false); | |
5457 | ||
5458 | case NVPTX_BUILTIN_VECTOR_ADDR: | |
5459 | return nvptx_expand_shared_addr (exp, target, mode, ignore, true); | |
5460 | ||
5461 | case NVPTX_BUILTIN_CMP_SWAP: | |
5462 | case NVPTX_BUILTIN_CMP_SWAPLL: | |
5463 | return nvptx_expand_cmp_swap (exp, target, mode, ignore); | |
5464 | ||
5465 | default: gcc_unreachable (); | |
5466 | } | |
5467 | } | |
5468 | ||
5469 | /* Implement TARGET_SIMT_VF target hook: number of threads in a warp. */ | |
5470 | ||
5471 | static int | |
5472 | nvptx_simt_vf () | |
5473 | { | |
5474 | return PTX_WARP_SIZE; | |
5475 | } | |
5476 | ||
5477 | /* Return 1 if TRAIT NAME is present in the OpenMP context's | |
5478 | device trait set, return 0 if not present in any OpenMP context in the | |
5479 | whole translation unit, or -1 if not present in the current OpenMP context | |
5480 | but might be present in another OpenMP context in the same TU. */ | |
5481 | ||
5482 | int | |
5483 | nvptx_omp_device_kind_arch_isa (enum omp_device_kind_arch_isa trait, | |
5484 | const char *name) | |
5485 | { | |
5486 | switch (trait) | |
5487 | { | |
5488 | case omp_device_kind: | |
5489 | return strcmp (name, "gpu") == 0; | |
5490 | case omp_device_arch: | |
5491 | return strcmp (name, "nvptx") == 0; | |
5492 | case omp_device_isa: | |
5493 | if (strcmp (name, "sm_30") == 0) | |
5494 | return !TARGET_SM35; | |
5495 | if (strcmp (name, "sm_35") == 0) | |
5496 | return TARGET_SM35; | |
5497 | return 0; | |
5498 | default: | |
5499 | gcc_unreachable (); | |
5500 | } | |
5501 | } | |
5502 | ||
5503 | static bool | |
5504 | nvptx_welformed_vector_length_p (int l) | |
5505 | { | |
5506 | gcc_assert (l > 0); | |
5507 | return l % PTX_WARP_SIZE == 0; | |
5508 | } | |
5509 | ||
5510 | static void | |
5511 | nvptx_apply_dim_limits (int dims[]) | |
5512 | { | |
5513 | /* Check that the vector_length is not too large. */ | |
5514 | if (dims[GOMP_DIM_VECTOR] > PTX_MAX_VECTOR_LENGTH) | |
5515 | dims[GOMP_DIM_VECTOR] = PTX_MAX_VECTOR_LENGTH; | |
5516 | ||
5517 | /* Check that the number of workers is not too large. */ | |
5518 | if (dims[GOMP_DIM_WORKER] > PTX_WORKER_LENGTH) | |
5519 | dims[GOMP_DIM_WORKER] = PTX_WORKER_LENGTH; | |
5520 | ||
5521 | /* Ensure that num_worker * vector_length <= cta size. */ | |
5522 | if (dims[GOMP_DIM_WORKER] > 0 && dims[GOMP_DIM_VECTOR] > 0 | |
5523 | && dims[GOMP_DIM_WORKER] * dims[GOMP_DIM_VECTOR] > PTX_CTA_SIZE) | |
5524 | dims[GOMP_DIM_VECTOR] = PTX_WARP_SIZE; | |
5525 | ||
5526 | /* If we need a per-worker barrier ... . */ | |
5527 | if (dims[GOMP_DIM_WORKER] > 0 && dims[GOMP_DIM_VECTOR] > 0 | |
5528 | && dims[GOMP_DIM_VECTOR] > PTX_WARP_SIZE) | |
5529 | /* Don't use more barriers than available. */ | |
5530 | dims[GOMP_DIM_WORKER] = MIN (dims[GOMP_DIM_WORKER], | |
5531 | PTX_NUM_PER_WORKER_BARRIERS); | |
5532 | } | |
5533 | ||
5534 | /* Return true if FNDECL contains calls to vector-partitionable routines. */ | |
5535 | ||
5536 | static bool | |
5537 | has_vector_partitionable_routine_calls_p (tree fndecl) | |
5538 | { | |
5539 | if (!fndecl) | |
5540 | return false; | |
5541 | ||
5542 | basic_block bb; | |
5543 | FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (fndecl)) | |
5544 | for (gimple_stmt_iterator i = gsi_start_bb (bb); !gsi_end_p (i); | |
5545 | gsi_next_nondebug (&i)) | |
5546 | { | |
5547 | gimple *stmt = gsi_stmt (i); | |
5548 | if (gimple_code (stmt) != GIMPLE_CALL) | |
5549 | continue; | |
5550 | ||
5551 | tree callee = gimple_call_fndecl (stmt); | |
5552 | if (!callee) | |
5553 | continue; | |
5554 | ||
5555 | tree attrs = oacc_get_fn_attrib (callee); | |
5556 | if (attrs == NULL_TREE) | |
5557 | return false; | |
5558 | ||
5559 | int partition_level = oacc_fn_attrib_level (attrs); | |
5560 | bool seq_routine_p = partition_level == GOMP_DIM_MAX; | |
5561 | if (!seq_routine_p) | |
5562 | return true; | |
5563 | } | |
5564 | ||
5565 | return false; | |
5566 | } | |
5567 | ||
5568 | /* As nvptx_goacc_validate_dims, but does not return bool to indicate whether | |
5569 | DIMS has changed. */ | |
5570 | ||
5571 | static void | |
5572 | nvptx_goacc_validate_dims_1 (tree decl, int dims[], int fn_level, unsigned used) | |
5573 | { | |
5574 | bool oacc_default_dims_p = false; | |
5575 | bool oacc_min_dims_p = false; | |
5576 | bool offload_region_p = false; | |
5577 | bool routine_p = false; | |
5578 | bool routine_seq_p = false; | |
5579 | int default_vector_length = -1; | |
5580 | ||
5581 | if (decl == NULL_TREE) | |
5582 | { | |
5583 | if (fn_level == -1) | |
5584 | oacc_default_dims_p = true; | |
5585 | else if (fn_level == -2) | |
5586 | oacc_min_dims_p = true; | |
5587 | else | |
5588 | gcc_unreachable (); | |
5589 | } | |
5590 | else if (fn_level == -1) | |
5591 | offload_region_p = true; | |
5592 | else if (0 <= fn_level && fn_level <= GOMP_DIM_MAX) | |
5593 | { | |
5594 | routine_p = true; | |
5595 | routine_seq_p = fn_level == GOMP_DIM_MAX; | |
5596 | } | |
5597 | else | |
5598 | gcc_unreachable (); | |
5599 | ||
5600 | if (oacc_min_dims_p) | |
5601 | { | |
5602 | gcc_assert (dims[GOMP_DIM_VECTOR] == 1); | |
5603 | gcc_assert (dims[GOMP_DIM_WORKER] == 1); | |
5604 | gcc_assert (dims[GOMP_DIM_GANG] == 1); | |
5605 | ||
5606 | dims[GOMP_DIM_VECTOR] = PTX_WARP_SIZE; | |
5607 | return; | |
5608 | } | |
5609 | ||
5610 | if (routine_p) | |
5611 | { | |
5612 | if (!routine_seq_p) | |
5613 | dims[GOMP_DIM_VECTOR] = PTX_WARP_SIZE; | |
5614 | ||
5615 | return; | |
5616 | } | |
5617 | ||
5618 | if (oacc_default_dims_p) | |
5619 | { | |
5620 | /* -1 : not set | |
5621 | 0 : set at runtime, f.i. -fopenacc-dims=- | |
5622 | >= 1: set at compile time, f.i. -fopenacc-dims=1. */ | |
5623 | gcc_assert (dims[GOMP_DIM_VECTOR] >= -1); | |
5624 | gcc_assert (dims[GOMP_DIM_WORKER] >= -1); | |
5625 | gcc_assert (dims[GOMP_DIM_GANG] >= -1); | |
5626 | ||
5627 | /* But -fopenacc-dims=- is not yet supported on trunk. */ | |
5628 | gcc_assert (dims[GOMP_DIM_VECTOR] != 0); | |
5629 | gcc_assert (dims[GOMP_DIM_WORKER] != 0); | |
5630 | gcc_assert (dims[GOMP_DIM_GANG] != 0); | |
5631 | } | |
5632 | ||
5633 | if (offload_region_p) | |
5634 | { | |
5635 | /* -1 : not set | |
5636 | 0 : set using variable, f.i. num_gangs (n) | |
5637 | >= 1: set using constant, f.i. num_gangs (1). */ | |
5638 | gcc_assert (dims[GOMP_DIM_VECTOR] >= -1); | |
5639 | gcc_assert (dims[GOMP_DIM_WORKER] >= -1); | |
5640 | gcc_assert (dims[GOMP_DIM_GANG] >= -1); | |
5641 | } | |
5642 | ||
5643 | if (offload_region_p) | |
5644 | default_vector_length = oacc_get_default_dim (GOMP_DIM_VECTOR); | |
5645 | else | |
5646 | /* oacc_default_dims_p. */ | |
5647 | default_vector_length = PTX_DEFAULT_VECTOR_LENGTH; | |
5648 | ||
5649 | int old_dims[GOMP_DIM_MAX]; | |
5650 | unsigned int i; | |
5651 | for (i = 0; i < GOMP_DIM_MAX; ++i) | |
5652 | old_dims[i] = dims[i]; | |
5653 | ||
5654 | const char *vector_reason = NULL; | |
5655 | if (offload_region_p && has_vector_partitionable_routine_calls_p (decl)) | |
5656 | { | |
5657 | default_vector_length = PTX_WARP_SIZE; | |
5658 | ||
5659 | if (dims[GOMP_DIM_VECTOR] > PTX_WARP_SIZE) | |
5660 | { | |
5661 | vector_reason = G_("using vector_length (%d) due to call to" | |
5662 | " vector-partitionable routine, ignoring %d"); | |
5663 | dims[GOMP_DIM_VECTOR] = PTX_WARP_SIZE; | |
5664 | } | |
5665 | } | |
5666 | ||
5667 | if (dims[GOMP_DIM_VECTOR] == 0) | |
5668 | { | |
5669 | vector_reason = G_("using vector_length (%d), ignoring runtime setting"); | |
5670 | dims[GOMP_DIM_VECTOR] = default_vector_length; | |
5671 | } | |
5672 | ||
5673 | if (dims[GOMP_DIM_VECTOR] > 0 | |
5674 | && !nvptx_welformed_vector_length_p (dims[GOMP_DIM_VECTOR])) | |
5675 | dims[GOMP_DIM_VECTOR] = default_vector_length; | |
5676 | ||
5677 | nvptx_apply_dim_limits (dims); | |
5678 | ||
5679 | if (dims[GOMP_DIM_VECTOR] != old_dims[GOMP_DIM_VECTOR]) | |
5680 | warning_at (decl ? DECL_SOURCE_LOCATION (decl) : UNKNOWN_LOCATION, 0, | |
5681 | vector_reason != NULL | |
5682 | ? vector_reason | |
5683 | : G_("using vector_length (%d), ignoring %d"), | |
5684 | dims[GOMP_DIM_VECTOR], old_dims[GOMP_DIM_VECTOR]); | |
5685 | ||
5686 | if (dims[GOMP_DIM_WORKER] != old_dims[GOMP_DIM_WORKER]) | |
5687 | warning_at (decl ? DECL_SOURCE_LOCATION (decl) : UNKNOWN_LOCATION, 0, | |
5688 | G_("using num_workers (%d), ignoring %d"), | |
5689 | dims[GOMP_DIM_WORKER], old_dims[GOMP_DIM_WORKER]); | |
5690 | ||
5691 | if (oacc_default_dims_p) | |
5692 | { | |
5693 | if (dims[GOMP_DIM_VECTOR] < 0) | |
5694 | dims[GOMP_DIM_VECTOR] = default_vector_length; | |
5695 | if (dims[GOMP_DIM_WORKER] < 0) | |
5696 | dims[GOMP_DIM_WORKER] = PTX_DEFAULT_RUNTIME_DIM; | |
5697 | if (dims[GOMP_DIM_GANG] < 0) | |
5698 | dims[GOMP_DIM_GANG] = PTX_DEFAULT_RUNTIME_DIM; | |
5699 | nvptx_apply_dim_limits (dims); | |
5700 | } | |
5701 | ||
5702 | if (offload_region_p) | |
5703 | { | |
5704 | for (i = 0; i < GOMP_DIM_MAX; i++) | |
5705 | { | |
5706 | if (!(dims[i] < 0)) | |
5707 | continue; | |
5708 | ||
5709 | if ((used & GOMP_DIM_MASK (i)) == 0) | |
5710 | /* Function oacc_validate_dims will apply the minimal dimension. */ | |
5711 | continue; | |
5712 | ||
5713 | dims[i] = (i == GOMP_DIM_VECTOR | |
5714 | ? default_vector_length | |
5715 | : oacc_get_default_dim (i)); | |
5716 | } | |
5717 | ||
5718 | nvptx_apply_dim_limits (dims); | |
5719 | } | |
5720 | } | |
5721 | ||
5722 | /* Validate compute dimensions of an OpenACC offload or routine, fill | |
5723 | in non-unity defaults. FN_LEVEL indicates the level at which a | |
5724 | routine might spawn a loop. It is negative for non-routines. If | |
5725 | DECL is null, we are validating the default dimensions. */ | |
5726 | ||
5727 | static bool | |
5728 | nvptx_goacc_validate_dims (tree decl, int dims[], int fn_level, unsigned used) | |
5729 | { | |
5730 | int old_dims[GOMP_DIM_MAX]; | |
5731 | unsigned int i; | |
5732 | ||
5733 | for (i = 0; i < GOMP_DIM_MAX; ++i) | |
5734 | old_dims[i] = dims[i]; | |
5735 | ||
5736 | nvptx_goacc_validate_dims_1 (decl, dims, fn_level, used); | |
5737 | ||
5738 | gcc_assert (dims[GOMP_DIM_VECTOR] != 0); | |
5739 | if (dims[GOMP_DIM_WORKER] > 0 && dims[GOMP_DIM_VECTOR] > 0) | |
5740 | gcc_assert (dims[GOMP_DIM_WORKER] * dims[GOMP_DIM_VECTOR] <= PTX_CTA_SIZE); | |
5741 | ||
5742 | for (i = 0; i < GOMP_DIM_MAX; ++i) | |
5743 | if (old_dims[i] != dims[i]) | |
5744 | return true; | |
5745 | ||
5746 | return false; | |
5747 | } | |
5748 | ||
5749 | /* Return maximum dimension size, or zero for unbounded. */ | |
5750 | ||
5751 | static int | |
5752 | nvptx_dim_limit (int axis) | |
5753 | { | |
5754 | switch (axis) | |
5755 | { | |
5756 | case GOMP_DIM_VECTOR: | |
5757 | return PTX_MAX_VECTOR_LENGTH; | |
5758 | ||
5759 | default: | |
5760 | break; | |
5761 | } | |
5762 | return 0; | |
5763 | } | |
5764 | ||
5765 | /* Determine whether fork & joins are needed. */ | |
5766 | ||
5767 | static bool | |
5768 | nvptx_goacc_fork_join (gcall *call, const int dims[], | |
5769 | bool ARG_UNUSED (is_fork)) | |
5770 | { | |
5771 | tree arg = gimple_call_arg (call, 2); | |
5772 | unsigned axis = TREE_INT_CST_LOW (arg); | |
5773 | ||
5774 | /* We only care about worker and vector partitioning. */ | |
5775 | if (axis < GOMP_DIM_WORKER) | |
5776 | return false; | |
5777 | ||
5778 | /* If the size is 1, there's no partitioning. */ | |
5779 | if (dims[axis] == 1) | |
5780 | return false; | |
5781 | ||
5782 | return true; | |
5783 | } | |
5784 | ||
5785 | /* Generate a PTX builtin function call that returns the address in | |
5786 | the worker reduction buffer at OFFSET. TYPE is the type of the | |
5787 | data at that location. */ | |
5788 | ||
5789 | static tree | |
5790 | nvptx_get_shared_red_addr (tree type, tree offset, bool vector) | |
5791 | { | |
5792 | enum nvptx_builtins addr_dim = NVPTX_BUILTIN_WORKER_ADDR; | |
5793 | if (vector) | |
5794 | addr_dim = NVPTX_BUILTIN_VECTOR_ADDR; | |
5795 | machine_mode mode = TYPE_MODE (type); | |
5796 | tree fndecl = nvptx_builtin_decl (addr_dim, true); | |
5797 | tree size = build_int_cst (unsigned_type_node, GET_MODE_SIZE (mode)); | |
5798 | tree align = build_int_cst (unsigned_type_node, | |
5799 | GET_MODE_ALIGNMENT (mode) / BITS_PER_UNIT); | |
5800 | tree call = build_call_expr (fndecl, 3, offset, size, align); | |
5801 | ||
5802 | return fold_convert (build_pointer_type (type), call); | |
5803 | } | |
5804 | ||
5805 | /* Emit a SHFL.DOWN using index SHFL of VAR into DEST_VAR. This function | |
5806 | will cast the variable if necessary. */ | |
5807 | ||
5808 | static void | |
5809 | nvptx_generate_vector_shuffle (location_t loc, | |
5810 | tree dest_var, tree var, unsigned shift, | |
5811 | gimple_seq *seq) | |
5812 | { | |
5813 | unsigned fn = NVPTX_BUILTIN_SHUFFLE; | |
5814 | tree_code code = NOP_EXPR; | |
5815 | tree arg_type = unsigned_type_node; | |
5816 | tree var_type = TREE_TYPE (var); | |
5817 | tree dest_type = var_type; | |
5818 | ||
5819 | if (TREE_CODE (var_type) == COMPLEX_TYPE) | |
5820 | var_type = TREE_TYPE (var_type); | |
5821 | ||
5822 | if (TREE_CODE (var_type) == REAL_TYPE) | |
5823 | code = VIEW_CONVERT_EXPR; | |
5824 | ||
5825 | if (TYPE_SIZE (var_type) | |
5826 | == TYPE_SIZE (long_long_unsigned_type_node)) | |
5827 | { | |
5828 | fn = NVPTX_BUILTIN_SHUFFLELL; | |
5829 | arg_type = long_long_unsigned_type_node; | |
5830 | } | |
5831 | ||
5832 | tree call = nvptx_builtin_decl (fn, true); | |
5833 | tree bits = build_int_cst (unsigned_type_node, shift); | |
5834 | tree kind = build_int_cst (unsigned_type_node, SHUFFLE_DOWN); | |
5835 | tree expr; | |
5836 | ||
5837 | if (var_type != dest_type) | |
5838 | { | |
5839 | /* Do real and imaginary parts separately. */ | |
5840 | tree real = fold_build1 (REALPART_EXPR, var_type, var); | |
5841 | real = fold_build1 (code, arg_type, real); | |
5842 | real = build_call_expr_loc (loc, call, 3, real, bits, kind); | |
5843 | real = fold_build1 (code, var_type, real); | |
5844 | ||
5845 | tree imag = fold_build1 (IMAGPART_EXPR, var_type, var); | |
5846 | imag = fold_build1 (code, arg_type, imag); | |
5847 | imag = build_call_expr_loc (loc, call, 3, imag, bits, kind); | |
5848 | imag = fold_build1 (code, var_type, imag); | |
5849 | ||
5850 | expr = fold_build2 (COMPLEX_EXPR, dest_type, real, imag); | |
5851 | } | |
5852 | else | |
5853 | { | |
5854 | expr = fold_build1 (code, arg_type, var); | |
5855 | expr = build_call_expr_loc (loc, call, 3, expr, bits, kind); | |
5856 | expr = fold_build1 (code, dest_type, expr); | |
5857 | } | |
5858 | ||
5859 | gimplify_assign (dest_var, expr, seq); | |
5860 | } | |
5861 | ||
5862 | /* Lazily generate the global lock var decl and return its address. */ | |
5863 | ||
5864 | static tree | |
5865 | nvptx_global_lock_addr () | |
5866 | { | |
5867 | tree v = global_lock_var; | |
5868 | ||
5869 | if (!v) | |
5870 | { | |
5871 | tree name = get_identifier ("__reduction_lock"); | |
5872 | tree type = build_qualified_type (unsigned_type_node, | |
5873 | TYPE_QUAL_VOLATILE); | |
5874 | v = build_decl (BUILTINS_LOCATION, VAR_DECL, name, type); | |
5875 | global_lock_var = v; | |
5876 | DECL_ARTIFICIAL (v) = 1; | |
5877 | DECL_EXTERNAL (v) = 1; | |
5878 | TREE_STATIC (v) = 1; | |
5879 | TREE_PUBLIC (v) = 1; | |
5880 | TREE_USED (v) = 1; | |
5881 | mark_addressable (v); | |
5882 | mark_decl_referenced (v); | |
5883 | } | |
5884 | ||
5885 | return build_fold_addr_expr (v); | |
5886 | } | |
5887 | ||
5888 | /* Insert code to locklessly update *PTR with *PTR OP VAR just before | |
5889 | GSI. We use a lockless scheme for nearly all case, which looks | |
5890 | like: | |
5891 | actual = initval(OP); | |
5892 | do { | |
5893 | guess = actual; | |
5894 | write = guess OP myval; | |
5895 | actual = cmp&swap (ptr, guess, write) | |
5896 | } while (actual bit-different-to guess); | |
5897 | return write; | |
5898 | ||
5899 | This relies on a cmp&swap instruction, which is available for 32- | |
5900 | and 64-bit types. Larger types must use a locking scheme. */ | |
5901 | ||
5902 | static tree | |
5903 | nvptx_lockless_update (location_t loc, gimple_stmt_iterator *gsi, | |
5904 | tree ptr, tree var, tree_code op) | |
5905 | { | |
5906 | unsigned fn = NVPTX_BUILTIN_CMP_SWAP; | |
5907 | tree_code code = NOP_EXPR; | |
5908 | tree arg_type = unsigned_type_node; | |
5909 | tree var_type = TREE_TYPE (var); | |
5910 | ||
5911 | if (TREE_CODE (var_type) == COMPLEX_TYPE | |
5912 | || TREE_CODE (var_type) == REAL_TYPE) | |
5913 | code = VIEW_CONVERT_EXPR; | |
5914 | ||
5915 | if (TYPE_SIZE (var_type) == TYPE_SIZE (long_long_unsigned_type_node)) | |
5916 | { | |
5917 | arg_type = long_long_unsigned_type_node; | |
5918 | fn = NVPTX_BUILTIN_CMP_SWAPLL; | |
5919 | } | |
5920 | ||
5921 | tree swap_fn = nvptx_builtin_decl (fn, true); | |
5922 | ||
5923 | gimple_seq init_seq = NULL; | |
5924 | tree init_var = make_ssa_name (arg_type); | |
5925 | tree init_expr = omp_reduction_init_op (loc, op, var_type); | |
5926 | init_expr = fold_build1 (code, arg_type, init_expr); | |
5927 | gimplify_assign (init_var, init_expr, &init_seq); | |
5928 | gimple *init_end = gimple_seq_last (init_seq); | |
5929 | ||
5930 | gsi_insert_seq_before (gsi, init_seq, GSI_SAME_STMT); | |
5931 | ||
5932 | /* Split the block just after the init stmts. */ | |
5933 | basic_block pre_bb = gsi_bb (*gsi); | |
5934 | edge pre_edge = split_block (pre_bb, init_end); | |
5935 | basic_block loop_bb = pre_edge->dest; | |
5936 | pre_bb = pre_edge->src; | |
5937 | /* Reset the iterator. */ | |
5938 | *gsi = gsi_for_stmt (gsi_stmt (*gsi)); | |
5939 | ||
5940 | tree expect_var = make_ssa_name (arg_type); | |
5941 | tree actual_var = make_ssa_name (arg_type); | |
5942 | tree write_var = make_ssa_name (arg_type); | |
5943 | ||
5944 | /* Build and insert the reduction calculation. */ | |
5945 | gimple_seq red_seq = NULL; | |
5946 | tree write_expr = fold_build1 (code, var_type, expect_var); | |
5947 | write_expr = fold_build2 (op, var_type, write_expr, var); | |
5948 | write_expr = fold_build1 (code, arg_type, write_expr); | |
5949 | gimplify_assign (write_var, write_expr, &red_seq); | |
5950 | ||
5951 | gsi_insert_seq_before (gsi, red_seq, GSI_SAME_STMT); | |
5952 | ||
5953 | /* Build & insert the cmp&swap sequence. */ | |
5954 | gimple_seq latch_seq = NULL; | |
5955 | tree swap_expr = build_call_expr_loc (loc, swap_fn, 3, | |
5956 | ptr, expect_var, write_var); | |
5957 | gimplify_assign (actual_var, swap_expr, &latch_seq); | |
5958 | ||
5959 | gcond *cond = gimple_build_cond (EQ_EXPR, actual_var, expect_var, | |
5960 | NULL_TREE, NULL_TREE); | |
5961 | gimple_seq_add_stmt (&latch_seq, cond); | |
5962 | ||
5963 | gimple *latch_end = gimple_seq_last (latch_seq); | |
5964 | gsi_insert_seq_before (gsi, latch_seq, GSI_SAME_STMT); | |
5965 | ||
5966 | /* Split the block just after the latch stmts. */ | |
5967 | edge post_edge = split_block (loop_bb, latch_end); | |
5968 | basic_block post_bb = post_edge->dest; | |
5969 | loop_bb = post_edge->src; | |
5970 | *gsi = gsi_for_stmt (gsi_stmt (*gsi)); | |
5971 | ||
5972 | post_edge->flags ^= EDGE_TRUE_VALUE | EDGE_FALLTHRU; | |
5973 | post_edge->probability = profile_probability::even (); | |
5974 | edge loop_edge = make_edge (loop_bb, loop_bb, EDGE_FALSE_VALUE); | |
5975 | loop_edge->probability = profile_probability::even (); | |
5976 | set_immediate_dominator (CDI_DOMINATORS, loop_bb, pre_bb); | |
5977 | set_immediate_dominator (CDI_DOMINATORS, post_bb, loop_bb); | |
5978 | ||
5979 | gphi *phi = create_phi_node (expect_var, loop_bb); | |
5980 | add_phi_arg (phi, init_var, pre_edge, loc); | |
5981 | add_phi_arg (phi, actual_var, loop_edge, loc); | |
5982 | ||
5983 | loop *loop = alloc_loop (); | |
5984 | loop->header = loop_bb; | |
5985 | loop->latch = loop_bb; | |
5986 | add_loop (loop, loop_bb->loop_father); | |
5987 | ||
5988 | return fold_build1 (code, var_type, write_var); | |
5989 | } | |
5990 | ||
5991 | /* Insert code to lockfully update *PTR with *PTR OP VAR just before | |
5992 | GSI. This is necessary for types larger than 64 bits, where there | |
5993 | is no cmp&swap instruction to implement a lockless scheme. We use | |
5994 | a lock variable in global memory. | |
5995 | ||
5996 | while (cmp&swap (&lock_var, 0, 1)) | |
5997 | continue; | |
5998 | T accum = *ptr; | |
5999 | accum = accum OP var; | |
6000 | *ptr = accum; | |
6001 | cmp&swap (&lock_var, 1, 0); | |
6002 | return accum; | |
6003 | ||
6004 | A lock in global memory is necessary to force execution engine | |
6005 | descheduling and avoid resource starvation that can occur if the | |
6006 | lock is in .shared memory. */ | |
6007 | ||
6008 | static tree | |
6009 | nvptx_lockfull_update (location_t loc, gimple_stmt_iterator *gsi, | |
6010 | tree ptr, tree var, tree_code op) | |
6011 | { | |
6012 | tree var_type = TREE_TYPE (var); | |
6013 | tree swap_fn = nvptx_builtin_decl (NVPTX_BUILTIN_CMP_SWAP, true); | |
6014 | tree uns_unlocked = build_int_cst (unsigned_type_node, 0); | |
6015 | tree uns_locked = build_int_cst (unsigned_type_node, 1); | |
6016 | ||
6017 | /* Split the block just before the gsi. Insert a gimple nop to make | |
6018 | this easier. */ | |
6019 | gimple *nop = gimple_build_nop (); | |
6020 | gsi_insert_before (gsi, nop, GSI_SAME_STMT); | |
6021 | basic_block entry_bb = gsi_bb (*gsi); | |
6022 | edge entry_edge = split_block (entry_bb, nop); | |
6023 | basic_block lock_bb = entry_edge->dest; | |
6024 | /* Reset the iterator. */ | |
6025 | *gsi = gsi_for_stmt (gsi_stmt (*gsi)); | |
6026 | ||
6027 | /* Build and insert the locking sequence. */ | |
6028 | gimple_seq lock_seq = NULL; | |
6029 | tree lock_var = make_ssa_name (unsigned_type_node); | |
6030 | tree lock_expr = nvptx_global_lock_addr (); | |
6031 | lock_expr = build_call_expr_loc (loc, swap_fn, 3, lock_expr, | |
6032 | uns_unlocked, uns_locked); | |
6033 | gimplify_assign (lock_var, lock_expr, &lock_seq); | |
6034 | gcond *cond = gimple_build_cond (EQ_EXPR, lock_var, uns_unlocked, | |
6035 | NULL_TREE, NULL_TREE); | |
6036 | gimple_seq_add_stmt (&lock_seq, cond); | |
6037 | gimple *lock_end = gimple_seq_last (lock_seq); | |
6038 | gsi_insert_seq_before (gsi, lock_seq, GSI_SAME_STMT); | |
6039 | ||
6040 | /* Split the block just after the lock sequence. */ | |
6041 | edge locked_edge = split_block (lock_bb, lock_end); | |
6042 | basic_block update_bb = locked_edge->dest; | |
6043 | lock_bb = locked_edge->src; | |
6044 | *gsi = gsi_for_stmt (gsi_stmt (*gsi)); | |
6045 | ||
6046 | /* Create the lock loop ... */ | |
6047 | locked_edge->flags ^= EDGE_TRUE_VALUE | EDGE_FALLTHRU; | |
6048 | locked_edge->probability = profile_probability::even (); | |
6049 | edge loop_edge = make_edge (lock_bb, lock_bb, EDGE_FALSE_VALUE); | |
6050 | loop_edge->probability = profile_probability::even (); | |
6051 | set_immediate_dominator (CDI_DOMINATORS, lock_bb, entry_bb); | |
6052 | set_immediate_dominator (CDI_DOMINATORS, update_bb, lock_bb); | |
6053 | ||
6054 | /* ... and the loop structure. */ | |
6055 | loop *lock_loop = alloc_loop (); | |
6056 | lock_loop->header = lock_bb; | |
6057 | lock_loop->latch = lock_bb; | |
6058 | lock_loop->nb_iterations_estimate = 1; | |
6059 | lock_loop->any_estimate = true; | |
6060 | add_loop (lock_loop, entry_bb->loop_father); | |
6061 | ||
6062 | /* Build and insert the reduction calculation. */ | |
6063 | gimple_seq red_seq = NULL; | |
6064 | tree acc_in = make_ssa_name (var_type); | |
6065 | tree ref_in = build_simple_mem_ref (ptr); | |
6066 | TREE_THIS_VOLATILE (ref_in) = 1; | |
6067 | gimplify_assign (acc_in, ref_in, &red_seq); | |
6068 | ||
6069 | tree acc_out = make_ssa_name (var_type); | |
6070 | tree update_expr = fold_build2 (op, var_type, ref_in, var); | |
6071 | gimplify_assign (acc_out, update_expr, &red_seq); | |
6072 | ||
6073 | tree ref_out = build_simple_mem_ref (ptr); | |
6074 | TREE_THIS_VOLATILE (ref_out) = 1; | |
6075 | gimplify_assign (ref_out, acc_out, &red_seq); | |
6076 | ||
6077 | gsi_insert_seq_before (gsi, red_seq, GSI_SAME_STMT); | |
6078 | ||
6079 | /* Build & insert the unlock sequence. */ | |
6080 | gimple_seq unlock_seq = NULL; | |
6081 | tree unlock_expr = nvptx_global_lock_addr (); | |
6082 | unlock_expr = build_call_expr_loc (loc, swap_fn, 3, unlock_expr, | |
6083 | uns_locked, uns_unlocked); | |
6084 | gimplify_and_add (unlock_expr, &unlock_seq); | |
6085 | gsi_insert_seq_before (gsi, unlock_seq, GSI_SAME_STMT); | |
6086 | ||
6087 | return acc_out; | |
6088 | } | |
6089 | ||
6090 | /* Emit a sequence to update a reduction accumlator at *PTR with the | |
6091 | value held in VAR using operator OP. Return the updated value. | |
6092 | ||
6093 | TODO: optimize for atomic ops and indepedent complex ops. */ | |
6094 | ||
6095 | static tree | |
6096 | nvptx_reduction_update (location_t loc, gimple_stmt_iterator *gsi, | |
6097 | tree ptr, tree var, tree_code op) | |
6098 | { | |
6099 | tree type = TREE_TYPE (var); | |
6100 | tree size = TYPE_SIZE (type); | |
6101 | ||
6102 | if (size == TYPE_SIZE (unsigned_type_node) | |
6103 | || size == TYPE_SIZE (long_long_unsigned_type_node)) | |
6104 | return nvptx_lockless_update (loc, gsi, ptr, var, op); | |
6105 | else | |
6106 | return nvptx_lockfull_update (loc, gsi, ptr, var, op); | |
6107 | } | |
6108 | ||
6109 | /* NVPTX implementation of GOACC_REDUCTION_SETUP. */ | |
6110 | ||
6111 | static void | |
6112 | nvptx_goacc_reduction_setup (gcall *call, offload_attrs *oa) | |
6113 | { | |
6114 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
6115 | tree lhs = gimple_call_lhs (call); | |
6116 | tree var = gimple_call_arg (call, 2); | |
6117 | int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); | |
6118 | gimple_seq seq = NULL; | |
6119 | ||
6120 | push_gimplify_context (true); | |
6121 | ||
6122 | if (level != GOMP_DIM_GANG) | |
6123 | { | |
6124 | /* Copy the receiver object. */ | |
6125 | tree ref_to_res = gimple_call_arg (call, 1); | |
6126 | ||
6127 | if (!integer_zerop (ref_to_res)) | |
6128 | var = build_simple_mem_ref (ref_to_res); | |
6129 | } | |
6130 | ||
6131 | if (level == GOMP_DIM_WORKER | |
6132 | || (level == GOMP_DIM_VECTOR && oa->vector_length > PTX_WARP_SIZE)) | |
6133 | { | |
6134 | /* Store incoming value to worker reduction buffer. */ | |
6135 | tree offset = gimple_call_arg (call, 5); | |
6136 | tree call = nvptx_get_shared_red_addr (TREE_TYPE (var), offset, | |
6137 | level == GOMP_DIM_VECTOR); | |
6138 | tree ptr = make_ssa_name (TREE_TYPE (call)); | |
6139 | ||
6140 | gimplify_assign (ptr, call, &seq); | |
6141 | tree ref = build_simple_mem_ref (ptr); | |
6142 | TREE_THIS_VOLATILE (ref) = 1; | |
6143 | gimplify_assign (ref, var, &seq); | |
6144 | } | |
6145 | ||
6146 | if (lhs) | |
6147 | gimplify_assign (lhs, var, &seq); | |
6148 | ||
6149 | pop_gimplify_context (NULL); | |
6150 | gsi_replace_with_seq (&gsi, seq, true); | |
6151 | } | |
6152 | ||
6153 | /* NVPTX implementation of GOACC_REDUCTION_INIT. */ | |
6154 | ||
6155 | static void | |
6156 | nvptx_goacc_reduction_init (gcall *call, offload_attrs *oa) | |
6157 | { | |
6158 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
6159 | tree lhs = gimple_call_lhs (call); | |
6160 | tree var = gimple_call_arg (call, 2); | |
6161 | int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); | |
6162 | enum tree_code rcode | |
6163 | = (enum tree_code)TREE_INT_CST_LOW (gimple_call_arg (call, 4)); | |
6164 | tree init = omp_reduction_init_op (gimple_location (call), rcode, | |
6165 | TREE_TYPE (var)); | |
6166 | gimple_seq seq = NULL; | |
6167 | ||
6168 | push_gimplify_context (true); | |
6169 | ||
6170 | if (level == GOMP_DIM_VECTOR && oa->vector_length == PTX_WARP_SIZE) | |
6171 | { | |
6172 | /* Initialize vector-non-zeroes to INIT_VAL (OP). */ | |
6173 | tree tid = make_ssa_name (integer_type_node); | |
6174 | tree dim_vector = gimple_call_arg (call, 3); | |
6175 | gimple *tid_call = gimple_build_call_internal (IFN_GOACC_DIM_POS, 1, | |
6176 | dim_vector); | |
6177 | gimple *cond_stmt = gimple_build_cond (NE_EXPR, tid, integer_zero_node, | |
6178 | NULL_TREE, NULL_TREE); | |
6179 | ||
6180 | gimple_call_set_lhs (tid_call, tid); | |
6181 | gimple_seq_add_stmt (&seq, tid_call); | |
6182 | gimple_seq_add_stmt (&seq, cond_stmt); | |
6183 | ||
6184 | /* Split the block just after the call. */ | |
6185 | edge init_edge = split_block (gsi_bb (gsi), call); | |
6186 | basic_block init_bb = init_edge->dest; | |
6187 | basic_block call_bb = init_edge->src; | |
6188 | ||
6189 | /* Fixup flags from call_bb to init_bb. */ | |
6190 | init_edge->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; | |
6191 | init_edge->probability = profile_probability::even (); | |
6192 | ||
6193 | /* Set the initialization stmts. */ | |
6194 | gimple_seq init_seq = NULL; | |
6195 | tree init_var = make_ssa_name (TREE_TYPE (var)); | |
6196 | gimplify_assign (init_var, init, &init_seq); | |
6197 | gsi = gsi_start_bb (init_bb); | |
6198 | gsi_insert_seq_before (&gsi, init_seq, GSI_SAME_STMT); | |
6199 | ||
6200 | /* Split block just after the init stmt. */ | |
6201 | gsi_prev (&gsi); | |
6202 | edge inited_edge = split_block (gsi_bb (gsi), gsi_stmt (gsi)); | |
6203 | basic_block dst_bb = inited_edge->dest; | |
6204 | ||
6205 | /* Create false edge from call_bb to dst_bb. */ | |
6206 | edge nop_edge = make_edge (call_bb, dst_bb, EDGE_FALSE_VALUE); | |
6207 | nop_edge->probability = profile_probability::even (); | |
6208 | ||
6209 | /* Create phi node in dst block. */ | |
6210 | gphi *phi = create_phi_node (lhs, dst_bb); | |
6211 | add_phi_arg (phi, init_var, inited_edge, gimple_location (call)); | |
6212 | add_phi_arg (phi, var, nop_edge, gimple_location (call)); | |
6213 | ||
6214 | /* Reset dominator of dst bb. */ | |
6215 | set_immediate_dominator (CDI_DOMINATORS, dst_bb, call_bb); | |
6216 | ||
6217 | /* Reset the gsi. */ | |
6218 | gsi = gsi_for_stmt (call); | |
6219 | } | |
6220 | else | |
6221 | { | |
6222 | if (level == GOMP_DIM_GANG) | |
6223 | { | |
6224 | /* If there's no receiver object, propagate the incoming VAR. */ | |
6225 | tree ref_to_res = gimple_call_arg (call, 1); | |
6226 | if (integer_zerop (ref_to_res)) | |
6227 | init = var; | |
6228 | } | |
6229 | ||
6230 | if (lhs != NULL_TREE) | |
6231 | gimplify_assign (lhs, init, &seq); | |
6232 | } | |
6233 | ||
6234 | pop_gimplify_context (NULL); | |
6235 | gsi_replace_with_seq (&gsi, seq, true); | |
6236 | } | |
6237 | ||
6238 | /* NVPTX implementation of GOACC_REDUCTION_FINI. */ | |
6239 | ||
6240 | static void | |
6241 | nvptx_goacc_reduction_fini (gcall *call, offload_attrs *oa) | |
6242 | { | |
6243 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
6244 | tree lhs = gimple_call_lhs (call); | |
6245 | tree ref_to_res = gimple_call_arg (call, 1); | |
6246 | tree var = gimple_call_arg (call, 2); | |
6247 | int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); | |
6248 | enum tree_code op | |
6249 | = (enum tree_code)TREE_INT_CST_LOW (gimple_call_arg (call, 4)); | |
6250 | gimple_seq seq = NULL; | |
6251 | tree r = NULL_TREE;; | |
6252 | ||
6253 | push_gimplify_context (true); | |
6254 | ||
6255 | if (level == GOMP_DIM_VECTOR && oa->vector_length == PTX_WARP_SIZE) | |
6256 | { | |
6257 | /* Emit binary shuffle tree. TODO. Emit this as an actual loop, | |
6258 | but that requires a method of emitting a unified jump at the | |
6259 | gimple level. */ | |
6260 | for (int shfl = PTX_WARP_SIZE / 2; shfl > 0; shfl = shfl >> 1) | |
6261 | { | |
6262 | tree other_var = make_ssa_name (TREE_TYPE (var)); | |
6263 | nvptx_generate_vector_shuffle (gimple_location (call), | |
6264 | other_var, var, shfl, &seq); | |
6265 | ||
6266 | r = make_ssa_name (TREE_TYPE (var)); | |
6267 | gimplify_assign (r, fold_build2 (op, TREE_TYPE (var), | |
6268 | var, other_var), &seq); | |
6269 | var = r; | |
6270 | } | |
6271 | } | |
6272 | else | |
6273 | { | |
6274 | tree accum = NULL_TREE; | |
6275 | ||
6276 | if (level == GOMP_DIM_WORKER || level == GOMP_DIM_VECTOR) | |
6277 | { | |
6278 | /* Get reduction buffer address. */ | |
6279 | tree offset = gimple_call_arg (call, 5); | |
6280 | tree call = nvptx_get_shared_red_addr (TREE_TYPE (var), offset, | |
6281 | level == GOMP_DIM_VECTOR); | |
6282 | tree ptr = make_ssa_name (TREE_TYPE (call)); | |
6283 | ||
6284 | gimplify_assign (ptr, call, &seq); | |
6285 | accum = ptr; | |
6286 | } | |
6287 | else if (integer_zerop (ref_to_res)) | |
6288 | r = var; | |
6289 | else | |
6290 | accum = ref_to_res; | |
6291 | ||
6292 | if (accum) | |
6293 | { | |
6294 | /* UPDATE the accumulator. */ | |
6295 | gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); | |
6296 | seq = NULL; | |
6297 | r = nvptx_reduction_update (gimple_location (call), &gsi, | |
6298 | accum, var, op); | |
6299 | } | |
6300 | } | |
6301 | ||
6302 | if (lhs) | |
6303 | gimplify_assign (lhs, r, &seq); | |
6304 | pop_gimplify_context (NULL); | |
6305 | ||
6306 | gsi_replace_with_seq (&gsi, seq, true); | |
6307 | } | |
6308 | ||
6309 | /* NVPTX implementation of GOACC_REDUCTION_TEARDOWN. */ | |
6310 | ||
6311 | static void | |
6312 | nvptx_goacc_reduction_teardown (gcall *call, offload_attrs *oa) | |
6313 | { | |
6314 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
6315 | tree lhs = gimple_call_lhs (call); | |
6316 | tree var = gimple_call_arg (call, 2); | |
6317 | int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); | |
6318 | gimple_seq seq = NULL; | |
6319 | ||
6320 | push_gimplify_context (true); | |
6321 | if (level == GOMP_DIM_WORKER | |
6322 | || (level == GOMP_DIM_VECTOR && oa->vector_length > PTX_WARP_SIZE)) | |
6323 | { | |
6324 | /* Read the worker reduction buffer. */ | |
6325 | tree offset = gimple_call_arg (call, 5); | |
6326 | tree call = nvptx_get_shared_red_addr (TREE_TYPE (var), offset, | |
6327 | level == GOMP_DIM_VECTOR); | |
6328 | tree ptr = make_ssa_name (TREE_TYPE (call)); | |
6329 | ||
6330 | gimplify_assign (ptr, call, &seq); | |
6331 | var = build_simple_mem_ref (ptr); | |
6332 | TREE_THIS_VOLATILE (var) = 1; | |
6333 | } | |
6334 | ||
6335 | if (level != GOMP_DIM_GANG) | |
6336 | { | |
6337 | /* Write to the receiver object. */ | |
6338 | tree ref_to_res = gimple_call_arg (call, 1); | |
6339 | ||
6340 | if (!integer_zerop (ref_to_res)) | |
6341 | gimplify_assign (build_simple_mem_ref (ref_to_res), var, &seq); | |
6342 | } | |
6343 | ||
6344 | if (lhs) | |
6345 | gimplify_assign (lhs, var, &seq); | |
6346 | ||
6347 | pop_gimplify_context (NULL); | |
6348 | ||
6349 | gsi_replace_with_seq (&gsi, seq, true); | |
6350 | } | |
6351 | ||
6352 | /* NVPTX reduction expander. */ | |
6353 | ||
6354 | static void | |
6355 | nvptx_goacc_reduction (gcall *call) | |
6356 | { | |
6357 | unsigned code = (unsigned)TREE_INT_CST_LOW (gimple_call_arg (call, 0)); | |
6358 | offload_attrs oa; | |
6359 | ||
6360 | populate_offload_attrs (&oa); | |
6361 | ||
6362 | switch (code) | |
6363 | { | |
6364 | case IFN_GOACC_REDUCTION_SETUP: | |
6365 | nvptx_goacc_reduction_setup (call, &oa); | |
6366 | break; | |
6367 | ||
6368 | case IFN_GOACC_REDUCTION_INIT: | |
6369 | nvptx_goacc_reduction_init (call, &oa); | |
6370 | break; | |
6371 | ||
6372 | case IFN_GOACC_REDUCTION_FINI: | |
6373 | nvptx_goacc_reduction_fini (call, &oa); | |
6374 | break; | |
6375 | ||
6376 | case IFN_GOACC_REDUCTION_TEARDOWN: | |
6377 | nvptx_goacc_reduction_teardown (call, &oa); | |
6378 | break; | |
6379 | ||
6380 | default: | |
6381 | gcc_unreachable (); | |
6382 | } | |
6383 | } | |
6384 | ||
6385 | static bool | |
6386 | nvptx_cannot_force_const_mem (machine_mode mode ATTRIBUTE_UNUSED, | |
6387 | rtx x ATTRIBUTE_UNUSED) | |
6388 | { | |
6389 | return true; | |
6390 | } | |
6391 | ||
6392 | static bool | |
6393 | nvptx_vector_mode_supported (machine_mode mode) | |
6394 | { | |
6395 | return (mode == V2SImode | |
6396 | || mode == V2DImode); | |
6397 | } | |
6398 | ||
6399 | /* Return the preferred mode for vectorizing scalar MODE. */ | |
6400 | ||
6401 | static machine_mode | |
6402 | nvptx_preferred_simd_mode (scalar_mode mode) | |
6403 | { | |
6404 | switch (mode) | |
6405 | { | |
6406 | case E_DImode: | |
6407 | return V2DImode; | |
6408 | case E_SImode: | |
6409 | return V2SImode; | |
6410 | ||
6411 | default: | |
6412 | return default_preferred_simd_mode (mode); | |
6413 | } | |
6414 | } | |
6415 | ||
6416 | unsigned int | |
6417 | nvptx_data_alignment (const_tree type, unsigned int basic_align) | |
6418 | { | |
6419 | if (TREE_CODE (type) == INTEGER_TYPE) | |
6420 | { | |
6421 | unsigned HOST_WIDE_INT size = tree_to_uhwi (TYPE_SIZE_UNIT (type)); | |
6422 | if (size == GET_MODE_SIZE (TImode)) | |
6423 | return GET_MODE_BITSIZE (maybe_split_mode (TImode)); | |
6424 | } | |
6425 | ||
6426 | return basic_align; | |
6427 | } | |
6428 | ||
6429 | /* Implement TARGET_MODES_TIEABLE_P. */ | |
6430 | ||
6431 | static bool | |
6432 | nvptx_modes_tieable_p (machine_mode, machine_mode) | |
6433 | { | |
6434 | return false; | |
6435 | } | |
6436 | ||
6437 | /* Implement TARGET_HARD_REGNO_NREGS. */ | |
6438 | ||
6439 | static unsigned int | |
6440 | nvptx_hard_regno_nregs (unsigned int, machine_mode) | |
6441 | { | |
6442 | return 1; | |
6443 | } | |
6444 | ||
6445 | /* Implement TARGET_CAN_CHANGE_MODE_CLASS. */ | |
6446 | ||
6447 | static bool | |
6448 | nvptx_can_change_mode_class (machine_mode, machine_mode, reg_class_t) | |
6449 | { | |
6450 | return false; | |
6451 | } | |
6452 | ||
6453 | static GTY(()) tree nvptx_previous_fndecl; | |
6454 | ||
6455 | static void | |
6456 | nvptx_set_current_function (tree fndecl) | |
6457 | { | |
6458 | if (!fndecl || fndecl == nvptx_previous_fndecl) | |
6459 | return; | |
6460 | ||
6461 | nvptx_previous_fndecl = fndecl; | |
6462 | vector_red_partition = 0; | |
6463 | oacc_bcast_partition = 0; | |
6464 | } | |
6465 | ||
6466 | #undef TARGET_OPTION_OVERRIDE | |
6467 | #define TARGET_OPTION_OVERRIDE nvptx_option_override | |
6468 | ||
6469 | #undef TARGET_ATTRIBUTE_TABLE | |
6470 | #define TARGET_ATTRIBUTE_TABLE nvptx_attribute_table | |
6471 | ||
6472 | #undef TARGET_LRA_P | |
6473 | #define TARGET_LRA_P hook_bool_void_false | |
6474 | ||
6475 | #undef TARGET_LEGITIMATE_ADDRESS_P | |
6476 | #define TARGET_LEGITIMATE_ADDRESS_P nvptx_legitimate_address_p | |
6477 | ||
6478 | #undef TARGET_PROMOTE_FUNCTION_MODE | |
6479 | #define TARGET_PROMOTE_FUNCTION_MODE nvptx_promote_function_mode | |
6480 | ||
6481 | #undef TARGET_FUNCTION_ARG | |
6482 | #define TARGET_FUNCTION_ARG nvptx_function_arg | |
6483 | #undef TARGET_FUNCTION_INCOMING_ARG | |
6484 | #define TARGET_FUNCTION_INCOMING_ARG nvptx_function_incoming_arg | |
6485 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
6486 | #define TARGET_FUNCTION_ARG_ADVANCE nvptx_function_arg_advance | |
6487 | #undef TARGET_FUNCTION_ARG_BOUNDARY | |
6488 | #define TARGET_FUNCTION_ARG_BOUNDARY nvptx_function_arg_boundary | |
6489 | #undef TARGET_PASS_BY_REFERENCE | |
6490 | #define TARGET_PASS_BY_REFERENCE nvptx_pass_by_reference | |
6491 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
6492 | #define TARGET_FUNCTION_VALUE_REGNO_P nvptx_function_value_regno_p | |
6493 | #undef TARGET_FUNCTION_VALUE | |
6494 | #define TARGET_FUNCTION_VALUE nvptx_function_value | |
6495 | #undef TARGET_LIBCALL_VALUE | |
6496 | #define TARGET_LIBCALL_VALUE nvptx_libcall_value | |
6497 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
6498 | #define TARGET_FUNCTION_OK_FOR_SIBCALL nvptx_function_ok_for_sibcall | |
6499 | #undef TARGET_GET_DRAP_RTX | |
6500 | #define TARGET_GET_DRAP_RTX nvptx_get_drap_rtx | |
6501 | #undef TARGET_SPLIT_COMPLEX_ARG | |
6502 | #define TARGET_SPLIT_COMPLEX_ARG hook_bool_const_tree_true | |
6503 | #undef TARGET_RETURN_IN_MEMORY | |
6504 | #define TARGET_RETURN_IN_MEMORY nvptx_return_in_memory | |
6505 | #undef TARGET_OMIT_STRUCT_RETURN_REG | |
6506 | #define TARGET_OMIT_STRUCT_RETURN_REG true | |
6507 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
6508 | #define TARGET_STRICT_ARGUMENT_NAMING nvptx_strict_argument_naming | |
6509 | #undef TARGET_CALL_ARGS | |
6510 | #define TARGET_CALL_ARGS nvptx_call_args | |
6511 | #undef TARGET_END_CALL_ARGS | |
6512 | #define TARGET_END_CALL_ARGS nvptx_end_call_args | |
6513 | ||
6514 | #undef TARGET_ASM_FILE_START | |
6515 | #define TARGET_ASM_FILE_START nvptx_file_start | |
6516 | #undef TARGET_ASM_FILE_END | |
6517 | #define TARGET_ASM_FILE_END nvptx_file_end | |
6518 | #undef TARGET_ASM_GLOBALIZE_LABEL | |
6519 | #define TARGET_ASM_GLOBALIZE_LABEL nvptx_globalize_label | |
6520 | #undef TARGET_ASM_ASSEMBLE_UNDEFINED_DECL | |
6521 | #define TARGET_ASM_ASSEMBLE_UNDEFINED_DECL nvptx_assemble_undefined_decl | |
6522 | #undef TARGET_PRINT_OPERAND | |
6523 | #define TARGET_PRINT_OPERAND nvptx_print_operand | |
6524 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
6525 | #define TARGET_PRINT_OPERAND_ADDRESS nvptx_print_operand_address | |
6526 | #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P | |
6527 | #define TARGET_PRINT_OPERAND_PUNCT_VALID_P nvptx_print_operand_punct_valid_p | |
6528 | #undef TARGET_ASM_INTEGER | |
6529 | #define TARGET_ASM_INTEGER nvptx_assemble_integer | |
6530 | #undef TARGET_ASM_DECL_END | |
6531 | #define TARGET_ASM_DECL_END nvptx_assemble_decl_end | |
6532 | #undef TARGET_ASM_DECLARE_CONSTANT_NAME | |
6533 | #define TARGET_ASM_DECLARE_CONSTANT_NAME nvptx_asm_declare_constant_name | |
6534 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
6535 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P hook_bool_mode_const_rtx_true | |
6536 | #undef TARGET_ASM_NEED_VAR_DECL_BEFORE_USE | |
6537 | #define TARGET_ASM_NEED_VAR_DECL_BEFORE_USE true | |
6538 | ||
6539 | #undef TARGET_MACHINE_DEPENDENT_REORG | |
6540 | #define TARGET_MACHINE_DEPENDENT_REORG nvptx_reorg | |
6541 | #undef TARGET_NO_REGISTER_ALLOCATION | |
6542 | #define TARGET_NO_REGISTER_ALLOCATION true | |
6543 | ||
6544 | #undef TARGET_ENCODE_SECTION_INFO | |
6545 | #define TARGET_ENCODE_SECTION_INFO nvptx_encode_section_info | |
6546 | #undef TARGET_RECORD_OFFLOAD_SYMBOL | |
6547 | #define TARGET_RECORD_OFFLOAD_SYMBOL nvptx_record_offload_symbol | |
6548 | ||
6549 | #undef TARGET_VECTOR_ALIGNMENT | |
6550 | #define TARGET_VECTOR_ALIGNMENT nvptx_vector_alignment | |
6551 | ||
6552 | #undef TARGET_CANNOT_COPY_INSN_P | |
6553 | #define TARGET_CANNOT_COPY_INSN_P nvptx_cannot_copy_insn_p | |
6554 | ||
6555 | #undef TARGET_USE_ANCHORS_FOR_SYMBOL_P | |
6556 | #define TARGET_USE_ANCHORS_FOR_SYMBOL_P nvptx_use_anchors_for_symbol_p | |
6557 | ||
6558 | #undef TARGET_INIT_BUILTINS | |
6559 | #define TARGET_INIT_BUILTINS nvptx_init_builtins | |
6560 | #undef TARGET_EXPAND_BUILTIN | |
6561 | #define TARGET_EXPAND_BUILTIN nvptx_expand_builtin | |
6562 | #undef TARGET_BUILTIN_DECL | |
6563 | #define TARGET_BUILTIN_DECL nvptx_builtin_decl | |
6564 | ||
6565 | #undef TARGET_SIMT_VF | |
6566 | #define TARGET_SIMT_VF nvptx_simt_vf | |
6567 | ||
6568 | #undef TARGET_OMP_DEVICE_KIND_ARCH_ISA | |
6569 | #define TARGET_OMP_DEVICE_KIND_ARCH_ISA nvptx_omp_device_kind_arch_isa | |
6570 | ||
6571 | #undef TARGET_GOACC_VALIDATE_DIMS | |
6572 | #define TARGET_GOACC_VALIDATE_DIMS nvptx_goacc_validate_dims | |
6573 | ||
6574 | #undef TARGET_GOACC_DIM_LIMIT | |
6575 | #define TARGET_GOACC_DIM_LIMIT nvptx_dim_limit | |
6576 | ||
6577 | #undef TARGET_GOACC_FORK_JOIN | |
6578 | #define TARGET_GOACC_FORK_JOIN nvptx_goacc_fork_join | |
6579 | ||
6580 | #undef TARGET_GOACC_REDUCTION | |
6581 | #define TARGET_GOACC_REDUCTION nvptx_goacc_reduction | |
6582 | ||
6583 | #undef TARGET_CANNOT_FORCE_CONST_MEM | |
6584 | #define TARGET_CANNOT_FORCE_CONST_MEM nvptx_cannot_force_const_mem | |
6585 | ||
6586 | #undef TARGET_VECTOR_MODE_SUPPORTED_P | |
6587 | #define TARGET_VECTOR_MODE_SUPPORTED_P nvptx_vector_mode_supported | |
6588 | ||
6589 | #undef TARGET_VECTORIZE_PREFERRED_SIMD_MODE | |
6590 | #define TARGET_VECTORIZE_PREFERRED_SIMD_MODE \ | |
6591 | nvptx_preferred_simd_mode | |
6592 | ||
6593 | #undef TARGET_MODES_TIEABLE_P | |
6594 | #define TARGET_MODES_TIEABLE_P nvptx_modes_tieable_p | |
6595 | ||
6596 | #undef TARGET_HARD_REGNO_NREGS | |
6597 | #define TARGET_HARD_REGNO_NREGS nvptx_hard_regno_nregs | |
6598 | ||
6599 | #undef TARGET_CAN_CHANGE_MODE_CLASS | |
6600 | #define TARGET_CAN_CHANGE_MODE_CLASS nvptx_can_change_mode_class | |
6601 | ||
6602 | #undef TARGET_HAVE_SPECULATION_SAFE_VALUE | |
6603 | #define TARGET_HAVE_SPECULATION_SAFE_VALUE speculation_safe_value_not_needed | |
6604 | ||
6605 | #undef TARGET_SET_CURRENT_FUNCTION | |
6606 | #define TARGET_SET_CURRENT_FUNCTION nvptx_set_current_function | |
6607 | ||
6608 | struct gcc_target targetm = TARGET_INITIALIZER; | |
6609 | ||
6610 | #include "gt-nvptx.h" |