]>
Commit | Line | Data |
---|---|---|
56686608 | 1 | /* Implementation of commonly needed HSAIL related functions and methods. |
2 | Copyright (C) 2013-2016 Free Software Foundation, Inc. | |
3 | Contributed by Martin Jambor <mjambor@suse.cz> and | |
4 | Martin Liska <mliska@suse.cz>. | |
5 | ||
6 | This file is part of GCC. | |
7 | ||
8 | GCC is free software; you can redistribute it and/or modify | |
9 | it under the terms of the GNU General Public License as published by | |
10 | the Free Software Foundation; either version 3, or (at your option) | |
11 | any later version. | |
12 | ||
13 | GCC is distributed in the hope that it will be useful, | |
14 | but WITHOUT ANY WARRANTY; without even the implied warranty of | |
15 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
16 | GNU General Public License for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
19 | along with GCC; see the file COPYING3. If not see | |
20 | <http://www.gnu.org/licenses/>. */ | |
21 | ||
22 | #include "config.h" | |
23 | #include "system.h" | |
24 | #include "coretypes.h" | |
25 | #include "tm.h" | |
26 | #include "is-a.h" | |
27 | #include "hash-set.h" | |
28 | #include "hash-map.h" | |
29 | #include "vec.h" | |
30 | #include "tree.h" | |
31 | #include "dumpfile.h" | |
32 | #include "gimple-pretty-print.h" | |
33 | #include "diagnostic-core.h" | |
34 | #include "alloc-pool.h" | |
35 | #include "cgraph.h" | |
36 | #include "print-tree.h" | |
37 | #include "stringpool.h" | |
38 | #include "symbol-summary.h" | |
39 | #include "hsa.h" | |
40 | #include "internal-fn.h" | |
41 | #include "ctype.h" | |
42 | ||
43 | /* Structure containing intermediate HSA representation of the generated | |
44 | function. */ | |
45 | class hsa_function_representation *hsa_cfun; | |
46 | ||
47 | /* Element of the mapping vector between a host decl and an HSA kernel. */ | |
48 | ||
49 | struct GTY(()) hsa_decl_kernel_map_element | |
50 | { | |
51 | /* The decl of the host function. */ | |
52 | tree decl; | |
53 | /* Name of the HSA kernel in BRIG. */ | |
54 | char * GTY((skip)) name; | |
55 | /* Size of OMP data, if the kernel contains a kernel dispatch. */ | |
56 | unsigned omp_data_size; | |
57 | /* True if the function is gridified kernel. */ | |
58 | bool gridified_kernel_p; | |
59 | }; | |
60 | ||
61 | /* Mapping between decls and corresponding HSA kernels in this compilation | |
62 | unit. */ | |
63 | ||
64 | static GTY (()) vec<hsa_decl_kernel_map_element, va_gc> | |
65 | *hsa_decl_kernel_mapping; | |
66 | ||
67 | /* Mapping between decls and corresponding HSA kernels | |
68 | called by the function. */ | |
69 | hash_map <tree, vec <const char *> *> *hsa_decl_kernel_dependencies; | |
70 | ||
71 | /* Hash function to lookup a symbol for a decl. */ | |
72 | hash_table <hsa_noop_symbol_hasher> *hsa_global_variable_symbols; | |
73 | ||
74 | /* HSA summaries. */ | |
75 | hsa_summary_t *hsa_summaries = NULL; | |
76 | ||
77 | /* HSA number of threads. */ | |
78 | hsa_symbol *hsa_num_threads = NULL; | |
79 | ||
80 | /* HSA function that cannot be expanded to HSAIL. */ | |
81 | hash_set <tree> *hsa_failed_functions = NULL; | |
82 | ||
83 | /* True if compilation unit-wide data are already allocated and initialized. */ | |
84 | static bool compilation_unit_data_initialized; | |
85 | ||
86 | /* Return true if FNDECL represents an HSA-callable function. */ | |
87 | ||
88 | bool | |
89 | hsa_callable_function_p (tree fndecl) | |
90 | { | |
91 | return (lookup_attribute ("omp declare target", DECL_ATTRIBUTES (fndecl)) | |
92 | && !lookup_attribute ("oacc function", DECL_ATTRIBUTES (fndecl))); | |
93 | } | |
94 | ||
95 | /* Allocate HSA structures that are are used when dealing with different | |
96 | functions. */ | |
97 | ||
98 | void | |
99 | hsa_init_compilation_unit_data (void) | |
100 | { | |
101 | if (compilation_unit_data_initialized) | |
102 | return; | |
103 | ||
104 | compilation_unit_data_initialized = true; | |
105 | ||
106 | hsa_global_variable_symbols = new hash_table <hsa_noop_symbol_hasher> (8); | |
107 | hsa_failed_functions = new hash_set <tree> (); | |
108 | hsa_emitted_internal_decls = new hash_table <hsa_internal_fn_hasher> (2); | |
109 | } | |
110 | ||
111 | /* Free data structures that are used when dealing with different | |
112 | functions. */ | |
113 | ||
114 | void | |
115 | hsa_deinit_compilation_unit_data (void) | |
116 | { | |
117 | gcc_assert (compilation_unit_data_initialized); | |
118 | ||
119 | delete hsa_failed_functions; | |
120 | delete hsa_emitted_internal_decls; | |
121 | ||
122 | for (hash_table <hsa_noop_symbol_hasher>::iterator it | |
123 | = hsa_global_variable_symbols->begin (); | |
124 | it != hsa_global_variable_symbols->end (); | |
125 | ++it) | |
126 | { | |
127 | hsa_symbol *sym = *it; | |
128 | delete sym; | |
129 | } | |
130 | ||
131 | delete hsa_global_variable_symbols; | |
132 | ||
133 | if (hsa_num_threads) | |
134 | { | |
135 | delete hsa_num_threads; | |
136 | hsa_num_threads = NULL; | |
137 | } | |
138 | ||
139 | compilation_unit_data_initialized = false; | |
140 | } | |
141 | ||
142 | /* Return true if we are generating large HSA machine model. */ | |
143 | ||
144 | bool | |
145 | hsa_machine_large_p (void) | |
146 | { | |
147 | /* FIXME: I suppose this is technically wrong but should work for me now. */ | |
148 | return (GET_MODE_BITSIZE (Pmode) == 64); | |
149 | } | |
150 | ||
151 | /* Return the HSA profile we are using. */ | |
152 | ||
153 | bool | |
154 | hsa_full_profile_p (void) | |
155 | { | |
156 | return true; | |
157 | } | |
158 | ||
159 | /* Return true if a register in operand number OPNUM of instruction | |
160 | is an output. False if it is an input. */ | |
161 | ||
162 | bool | |
163 | hsa_insn_basic::op_output_p (unsigned opnum) | |
164 | { | |
165 | switch (m_opcode) | |
166 | { | |
167 | case HSA_OPCODE_PHI: | |
168 | case BRIG_OPCODE_CBR: | |
169 | case BRIG_OPCODE_SBR: | |
170 | case BRIG_OPCODE_ST: | |
171 | case BRIG_OPCODE_SIGNALNORET: | |
172 | /* FIXME: There are probably missing cases here, double check. */ | |
173 | return false; | |
174 | case BRIG_OPCODE_EXPAND: | |
175 | /* Example: expand_v4_b32_b128 (dest0, dest1, dest2, dest3), src0. */ | |
176 | return opnum < operand_count () - 1; | |
177 | default: | |
178 | return opnum == 0; | |
179 | } | |
180 | } | |
181 | ||
182 | /* Return true if OPCODE is an floating-point bit instruction opcode. */ | |
183 | ||
184 | bool | |
185 | hsa_opcode_floating_bit_insn_p (BrigOpcode16_t opcode) | |
186 | { | |
187 | switch (opcode) | |
188 | { | |
189 | case BRIG_OPCODE_NEG: | |
190 | case BRIG_OPCODE_ABS: | |
191 | case BRIG_OPCODE_CLASS: | |
192 | case BRIG_OPCODE_COPYSIGN: | |
193 | return true; | |
194 | default: | |
195 | return false; | |
196 | } | |
197 | } | |
198 | ||
199 | /* Return the number of destination operands for this INSN. */ | |
200 | ||
201 | unsigned | |
202 | hsa_insn_basic::input_count () | |
203 | { | |
204 | switch (m_opcode) | |
205 | { | |
206 | default: | |
207 | return 1; | |
208 | ||
209 | case BRIG_OPCODE_NOP: | |
210 | return 0; | |
211 | ||
212 | case BRIG_OPCODE_EXPAND: | |
213 | return 2; | |
214 | ||
215 | case BRIG_OPCODE_LD: | |
216 | /* ld_v[234] not yet handled. */ | |
217 | return 1; | |
218 | ||
219 | case BRIG_OPCODE_ST: | |
220 | return 0; | |
221 | ||
222 | case BRIG_OPCODE_ATOMICNORET: | |
223 | return 0; | |
224 | ||
225 | case BRIG_OPCODE_SIGNAL: | |
226 | return 1; | |
227 | ||
228 | case BRIG_OPCODE_SIGNALNORET: | |
229 | return 0; | |
230 | ||
231 | case BRIG_OPCODE_MEMFENCE: | |
232 | return 0; | |
233 | ||
234 | case BRIG_OPCODE_RDIMAGE: | |
235 | case BRIG_OPCODE_LDIMAGE: | |
236 | case BRIG_OPCODE_STIMAGE: | |
237 | case BRIG_OPCODE_QUERYIMAGE: | |
238 | case BRIG_OPCODE_QUERYSAMPLER: | |
239 | sorry ("HSA image ops not handled"); | |
240 | return 0; | |
241 | ||
242 | case BRIG_OPCODE_CBR: | |
243 | case BRIG_OPCODE_BR: | |
244 | return 0; | |
245 | ||
246 | case BRIG_OPCODE_SBR: | |
247 | return 0; /* ??? */ | |
248 | ||
249 | case BRIG_OPCODE_WAVEBARRIER: | |
250 | return 0; /* ??? */ | |
251 | ||
252 | case BRIG_OPCODE_BARRIER: | |
253 | case BRIG_OPCODE_ARRIVEFBAR: | |
254 | case BRIG_OPCODE_INITFBAR: | |
255 | case BRIG_OPCODE_JOINFBAR: | |
256 | case BRIG_OPCODE_LEAVEFBAR: | |
257 | case BRIG_OPCODE_RELEASEFBAR: | |
258 | case BRIG_OPCODE_WAITFBAR: | |
259 | return 0; | |
260 | ||
261 | case BRIG_OPCODE_LDF: | |
262 | return 1; | |
263 | ||
264 | case BRIG_OPCODE_ACTIVELANECOUNT: | |
265 | case BRIG_OPCODE_ACTIVELANEID: | |
266 | case BRIG_OPCODE_ACTIVELANEMASK: | |
267 | case BRIG_OPCODE_ACTIVELANEPERMUTE: | |
268 | return 1; /* ??? */ | |
269 | ||
270 | case BRIG_OPCODE_CALL: | |
271 | case BRIG_OPCODE_SCALL: | |
272 | case BRIG_OPCODE_ICALL: | |
273 | return 0; | |
274 | ||
275 | case BRIG_OPCODE_RET: | |
276 | return 0; | |
277 | ||
278 | case BRIG_OPCODE_ALLOCA: | |
279 | return 1; | |
280 | ||
281 | case BRIG_OPCODE_CLEARDETECTEXCEPT: | |
282 | return 0; | |
283 | ||
284 | case BRIG_OPCODE_SETDETECTEXCEPT: | |
285 | return 0; | |
286 | ||
287 | case BRIG_OPCODE_PACKETCOMPLETIONSIG: | |
288 | case BRIG_OPCODE_PACKETID: | |
289 | case BRIG_OPCODE_CASQUEUEWRITEINDEX: | |
290 | case BRIG_OPCODE_LDQUEUEREADINDEX: | |
291 | case BRIG_OPCODE_LDQUEUEWRITEINDEX: | |
292 | case BRIG_OPCODE_STQUEUEREADINDEX: | |
293 | case BRIG_OPCODE_STQUEUEWRITEINDEX: | |
294 | return 1; /* ??? */ | |
295 | ||
296 | case BRIG_OPCODE_ADDQUEUEWRITEINDEX: | |
297 | return 1; | |
298 | ||
299 | case BRIG_OPCODE_DEBUGTRAP: | |
300 | return 0; | |
301 | ||
302 | case BRIG_OPCODE_GROUPBASEPTR: | |
303 | case BRIG_OPCODE_KERNARGBASEPTR: | |
304 | return 1; /* ??? */ | |
305 | ||
306 | case HSA_OPCODE_ARG_BLOCK: | |
307 | return 0; | |
308 | ||
309 | case BRIG_KIND_DIRECTIVE_COMMENT: | |
310 | return 0; | |
311 | } | |
312 | } | |
313 | ||
314 | /* Return the number of source operands for this INSN. */ | |
315 | ||
316 | unsigned | |
317 | hsa_insn_basic::num_used_ops () | |
318 | { | |
319 | gcc_checking_assert (input_count () <= operand_count ()); | |
320 | ||
321 | return operand_count () - input_count (); | |
322 | } | |
323 | ||
324 | /* Set alignment to VALUE. */ | |
325 | ||
326 | void | |
327 | hsa_insn_mem::set_align (BrigAlignment8_t value) | |
328 | { | |
329 | /* TODO: Perhaps remove this dump later on: */ | |
330 | if (dump_file && (dump_flags & TDF_DETAILS) && value < m_align) | |
331 | { | |
332 | fprintf (dump_file, "Decreasing alignment to %u in instruction ", value); | |
333 | dump_hsa_insn (dump_file, this); | |
334 | } | |
335 | m_align = value; | |
336 | } | |
337 | ||
338 | /* Return size of HSA type T in bits. */ | |
339 | ||
340 | unsigned | |
341 | hsa_type_bit_size (BrigType16_t t) | |
342 | { | |
343 | switch (t) | |
344 | { | |
345 | case BRIG_TYPE_B1: | |
346 | return 1; | |
347 | ||
348 | case BRIG_TYPE_U8: | |
349 | case BRIG_TYPE_S8: | |
350 | case BRIG_TYPE_B8: | |
351 | return 8; | |
352 | ||
353 | case BRIG_TYPE_U16: | |
354 | case BRIG_TYPE_S16: | |
355 | case BRIG_TYPE_B16: | |
356 | case BRIG_TYPE_F16: | |
357 | return 16; | |
358 | ||
359 | case BRIG_TYPE_U32: | |
360 | case BRIG_TYPE_S32: | |
361 | case BRIG_TYPE_B32: | |
362 | case BRIG_TYPE_F32: | |
363 | case BRIG_TYPE_U8X4: | |
364 | case BRIG_TYPE_U16X2: | |
365 | case BRIG_TYPE_S8X4: | |
366 | case BRIG_TYPE_S16X2: | |
367 | case BRIG_TYPE_F16X2: | |
368 | return 32; | |
369 | ||
370 | case BRIG_TYPE_U64: | |
371 | case BRIG_TYPE_S64: | |
372 | case BRIG_TYPE_F64: | |
373 | case BRIG_TYPE_B64: | |
374 | case BRIG_TYPE_U8X8: | |
375 | case BRIG_TYPE_U16X4: | |
376 | case BRIG_TYPE_U32X2: | |
377 | case BRIG_TYPE_S8X8: | |
378 | case BRIG_TYPE_S16X4: | |
379 | case BRIG_TYPE_S32X2: | |
380 | case BRIG_TYPE_F16X4: | |
381 | case BRIG_TYPE_F32X2: | |
382 | ||
383 | return 64; | |
384 | ||
385 | case BRIG_TYPE_B128: | |
386 | case BRIG_TYPE_U8X16: | |
387 | case BRIG_TYPE_U16X8: | |
388 | case BRIG_TYPE_U32X4: | |
389 | case BRIG_TYPE_U64X2: | |
390 | case BRIG_TYPE_S8X16: | |
391 | case BRIG_TYPE_S16X8: | |
392 | case BRIG_TYPE_S32X4: | |
393 | case BRIG_TYPE_S64X2: | |
394 | case BRIG_TYPE_F16X8: | |
395 | case BRIG_TYPE_F32X4: | |
396 | case BRIG_TYPE_F64X2: | |
397 | return 128; | |
398 | ||
399 | default: | |
400 | gcc_assert (hsa_seen_error ()); | |
401 | return t; | |
402 | } | |
403 | } | |
404 | ||
405 | /* Return BRIG bit-type with BITSIZE length. */ | |
406 | ||
407 | BrigType16_t | |
408 | hsa_bittype_for_bitsize (unsigned bitsize) | |
409 | { | |
410 | switch (bitsize) | |
411 | { | |
412 | case 1: | |
413 | return BRIG_TYPE_B1; | |
414 | case 8: | |
415 | return BRIG_TYPE_B8; | |
416 | case 16: | |
417 | return BRIG_TYPE_B16; | |
418 | case 32: | |
419 | return BRIG_TYPE_B32; | |
420 | case 64: | |
421 | return BRIG_TYPE_B64; | |
422 | case 128: | |
423 | return BRIG_TYPE_B128; | |
424 | default: | |
425 | gcc_unreachable (); | |
426 | } | |
427 | } | |
428 | ||
429 | /* Return BRIG unsigned int type with BITSIZE length. */ | |
430 | ||
431 | BrigType16_t | |
432 | hsa_uint_for_bitsize (unsigned bitsize) | |
433 | { | |
434 | switch (bitsize) | |
435 | { | |
436 | case 8: | |
437 | return BRIG_TYPE_U8; | |
438 | case 16: | |
439 | return BRIG_TYPE_U16; | |
440 | case 32: | |
441 | return BRIG_TYPE_U32; | |
442 | case 64: | |
443 | return BRIG_TYPE_U64; | |
444 | default: | |
445 | gcc_unreachable (); | |
446 | } | |
447 | } | |
448 | ||
449 | /* Return BRIG float type with BITSIZE length. */ | |
450 | ||
451 | BrigType16_t | |
452 | hsa_float_for_bitsize (unsigned bitsize) | |
453 | { | |
454 | switch (bitsize) | |
455 | { | |
456 | case 16: | |
457 | return BRIG_TYPE_F16; | |
458 | case 32: | |
459 | return BRIG_TYPE_F32; | |
460 | case 64: | |
461 | return BRIG_TYPE_F64; | |
462 | default: | |
463 | gcc_unreachable (); | |
464 | } | |
465 | } | |
466 | ||
467 | /* Return HSA bit-type with the same size as the type T. */ | |
468 | ||
469 | BrigType16_t | |
470 | hsa_bittype_for_type (BrigType16_t t) | |
471 | { | |
472 | return hsa_bittype_for_bitsize (hsa_type_bit_size (t)); | |
473 | } | |
474 | ||
0d1b26a7 | 475 | /* Return HSA unsigned integer type with the same size as the type T. */ |
476 | ||
477 | BrigType16_t | |
478 | hsa_unsigned_type_for_type (BrigType16_t t) | |
479 | { | |
480 | return hsa_uint_for_bitsize (hsa_type_bit_size (t)); | |
481 | } | |
482 | ||
bcd6e034 | 483 | /* Return true if TYPE is a packed HSA type. */ |
484 | ||
485 | bool | |
486 | hsa_type_packed_p (BrigType16_t type) | |
487 | { | |
488 | return (type & BRIG_TYPE_PACK_MASK) != BRIG_TYPE_PACK_NONE; | |
489 | } | |
490 | ||
56686608 | 491 | /* Return true if and only if TYPE is a floating point number type. */ |
492 | ||
493 | bool | |
494 | hsa_type_float_p (BrigType16_t type) | |
495 | { | |
496 | switch (type & BRIG_TYPE_BASE_MASK) | |
497 | { | |
498 | case BRIG_TYPE_F16: | |
499 | case BRIG_TYPE_F32: | |
500 | case BRIG_TYPE_F64: | |
501 | return true; | |
502 | default: | |
503 | return false; | |
504 | } | |
505 | } | |
506 | ||
507 | /* Return true if and only if TYPE is an integer number type. */ | |
508 | ||
509 | bool | |
510 | hsa_type_integer_p (BrigType16_t type) | |
511 | { | |
512 | switch (type & BRIG_TYPE_BASE_MASK) | |
513 | { | |
514 | case BRIG_TYPE_U8: | |
515 | case BRIG_TYPE_U16: | |
516 | case BRIG_TYPE_U32: | |
517 | case BRIG_TYPE_U64: | |
518 | case BRIG_TYPE_S8: | |
519 | case BRIG_TYPE_S16: | |
520 | case BRIG_TYPE_S32: | |
521 | case BRIG_TYPE_S64: | |
522 | return true; | |
523 | default: | |
524 | return false; | |
525 | } | |
526 | } | |
527 | ||
528 | /* Return true if and only if TYPE is an bit-type. */ | |
529 | ||
530 | bool | |
531 | hsa_btype_p (BrigType16_t type) | |
532 | { | |
533 | switch (type & BRIG_TYPE_BASE_MASK) | |
534 | { | |
535 | case BRIG_TYPE_B8: | |
536 | case BRIG_TYPE_B16: | |
537 | case BRIG_TYPE_B32: | |
538 | case BRIG_TYPE_B64: | |
539 | case BRIG_TYPE_B128: | |
540 | return true; | |
541 | default: | |
542 | return false; | |
543 | } | |
544 | } | |
545 | ||
546 | ||
547 | /* Return HSA alignment encoding alignment to N bits. */ | |
548 | ||
549 | BrigAlignment8_t | |
550 | hsa_alignment_encoding (unsigned n) | |
551 | { | |
552 | gcc_assert (n >= 8 && !(n & (n - 1))); | |
553 | if (n >= 256) | |
554 | return BRIG_ALIGNMENT_32; | |
555 | ||
556 | switch (n) | |
557 | { | |
558 | case 8: | |
559 | return BRIG_ALIGNMENT_1; | |
560 | case 16: | |
561 | return BRIG_ALIGNMENT_2; | |
562 | case 32: | |
563 | return BRIG_ALIGNMENT_4; | |
564 | case 64: | |
565 | return BRIG_ALIGNMENT_8; | |
566 | case 128: | |
567 | return BRIG_ALIGNMENT_16; | |
568 | default: | |
569 | gcc_unreachable (); | |
570 | } | |
571 | } | |
572 | ||
573 | /* Return natural alignment of HSA TYPE. */ | |
574 | ||
575 | BrigAlignment8_t | |
576 | hsa_natural_alignment (BrigType16_t type) | |
577 | { | |
578 | return hsa_alignment_encoding (hsa_type_bit_size (type & ~BRIG_TYPE_ARRAY)); | |
579 | } | |
580 | ||
581 | /* Call the correct destructor of a HSA instruction. */ | |
582 | ||
583 | void | |
584 | hsa_destroy_insn (hsa_insn_basic *insn) | |
585 | { | |
586 | if (hsa_insn_phi *phi = dyn_cast <hsa_insn_phi *> (insn)) | |
587 | phi->~hsa_insn_phi (); | |
588 | else if (hsa_insn_br *br = dyn_cast <hsa_insn_br *> (insn)) | |
589 | br->~hsa_insn_br (); | |
590 | else if (hsa_insn_cmp *cmp = dyn_cast <hsa_insn_cmp *> (insn)) | |
591 | cmp->~hsa_insn_cmp (); | |
592 | else if (hsa_insn_mem *mem = dyn_cast <hsa_insn_mem *> (insn)) | |
593 | mem->~hsa_insn_mem (); | |
594 | else if (hsa_insn_atomic *atomic = dyn_cast <hsa_insn_atomic *> (insn)) | |
595 | atomic->~hsa_insn_atomic (); | |
596 | else if (hsa_insn_seg *seg = dyn_cast <hsa_insn_seg *> (insn)) | |
597 | seg->~hsa_insn_seg (); | |
598 | else if (hsa_insn_call *call = dyn_cast <hsa_insn_call *> (insn)) | |
599 | call->~hsa_insn_call (); | |
600 | else if (hsa_insn_arg_block *block = dyn_cast <hsa_insn_arg_block *> (insn)) | |
601 | block->~hsa_insn_arg_block (); | |
602 | else if (hsa_insn_sbr *sbr = dyn_cast <hsa_insn_sbr *> (insn)) | |
603 | sbr->~hsa_insn_sbr (); | |
604 | else if (hsa_insn_comment *comment = dyn_cast <hsa_insn_comment *> (insn)) | |
605 | comment->~hsa_insn_comment (); | |
606 | else | |
607 | insn->~hsa_insn_basic (); | |
608 | } | |
609 | ||
610 | /* Call the correct destructor of a HSA operand. */ | |
611 | ||
612 | void | |
613 | hsa_destroy_operand (hsa_op_base *op) | |
614 | { | |
615 | if (hsa_op_code_list *list = dyn_cast <hsa_op_code_list *> (op)) | |
616 | list->~hsa_op_code_list (); | |
617 | else if (hsa_op_operand_list *list = dyn_cast <hsa_op_operand_list *> (op)) | |
618 | list->~hsa_op_operand_list (); | |
619 | else if (hsa_op_reg *reg = dyn_cast <hsa_op_reg *> (op)) | |
620 | reg->~hsa_op_reg (); | |
621 | else if (hsa_op_immed *immed = dyn_cast <hsa_op_immed *> (op)) | |
622 | immed->~hsa_op_immed (); | |
623 | else | |
624 | op->~hsa_op_base (); | |
625 | } | |
626 | ||
627 | /* Create a mapping between the original function DECL and kernel name NAME. */ | |
628 | ||
629 | void | |
630 | hsa_add_kern_decl_mapping (tree decl, char *name, unsigned omp_data_size, | |
631 | bool gridified_kernel_p) | |
632 | { | |
633 | hsa_decl_kernel_map_element dkm; | |
634 | dkm.decl = decl; | |
635 | dkm.name = name; | |
636 | dkm.omp_data_size = omp_data_size; | |
637 | dkm.gridified_kernel_p = gridified_kernel_p; | |
638 | vec_safe_push (hsa_decl_kernel_mapping, dkm); | |
639 | } | |
640 | ||
641 | /* Return the number of kernel decl name mappings. */ | |
642 | ||
643 | unsigned | |
644 | hsa_get_number_decl_kernel_mappings (void) | |
645 | { | |
646 | return vec_safe_length (hsa_decl_kernel_mapping); | |
647 | } | |
648 | ||
649 | /* Return the decl in the Ith kernel decl name mapping. */ | |
650 | ||
651 | tree | |
652 | hsa_get_decl_kernel_mapping_decl (unsigned i) | |
653 | { | |
654 | return (*hsa_decl_kernel_mapping)[i].decl; | |
655 | } | |
656 | ||
657 | /* Return the name in the Ith kernel decl name mapping. */ | |
658 | ||
659 | char * | |
660 | hsa_get_decl_kernel_mapping_name (unsigned i) | |
661 | { | |
662 | return (*hsa_decl_kernel_mapping)[i].name; | |
663 | } | |
664 | ||
665 | /* Return maximum OMP size for kernel decl name mapping. */ | |
666 | ||
667 | unsigned | |
668 | hsa_get_decl_kernel_mapping_omp_size (unsigned i) | |
669 | { | |
670 | return (*hsa_decl_kernel_mapping)[i].omp_data_size; | |
671 | } | |
672 | ||
673 | /* Return if the function is gridified kernel in decl name mapping. */ | |
674 | ||
675 | bool | |
676 | hsa_get_decl_kernel_mapping_gridified (unsigned i) | |
677 | { | |
678 | return (*hsa_decl_kernel_mapping)[i].gridified_kernel_p; | |
679 | } | |
680 | ||
681 | /* Free the mapping between original decls and kernel names. */ | |
682 | ||
683 | void | |
684 | hsa_free_decl_kernel_mapping (void) | |
685 | { | |
686 | if (hsa_decl_kernel_mapping == NULL) | |
687 | return; | |
688 | ||
689 | for (unsigned i = 0; i < hsa_decl_kernel_mapping->length (); ++i) | |
690 | free ((*hsa_decl_kernel_mapping)[i].name); | |
691 | ggc_free (hsa_decl_kernel_mapping); | |
692 | } | |
693 | ||
694 | /* Add new kernel dependency. */ | |
695 | ||
696 | void | |
697 | hsa_add_kernel_dependency (tree caller, const char *called_function) | |
698 | { | |
699 | if (hsa_decl_kernel_dependencies == NULL) | |
700 | hsa_decl_kernel_dependencies = new hash_map<tree, vec<const char *> *> (); | |
701 | ||
702 | vec <const char *> *s = NULL; | |
703 | vec <const char *> **slot = hsa_decl_kernel_dependencies->get (caller); | |
704 | if (slot == NULL) | |
705 | { | |
706 | s = new vec <const char *> (); | |
707 | hsa_decl_kernel_dependencies->put (caller, s); | |
708 | } | |
709 | else | |
710 | s = *slot; | |
711 | ||
712 | s->safe_push (called_function); | |
713 | } | |
714 | ||
e2f980b8 | 715 | /* Expansion to HSA needs a few gc roots to hold types, constructors etc. In |
716 | order to minimize the number of GTY roots, we'll root them all in the | |
717 | following array. The individual elements should only be accessed by the | |
718 | very simple getters (of a pointer-to-tree) below. */ | |
719 | ||
720 | static GTY(()) tree hsa_tree_gt_roots[3]; | |
721 | ||
722 | tree * | |
723 | hsa_get_ctor_statements (void) | |
724 | { | |
725 | return &hsa_tree_gt_roots[0]; | |
726 | } | |
727 | ||
728 | tree * | |
729 | hsa_get_dtor_statements (void) | |
730 | { | |
731 | return &hsa_tree_gt_roots[1]; | |
732 | } | |
733 | ||
734 | tree * | |
735 | hsa_get_kernel_dispatch_type (void) | |
736 | { | |
737 | return &hsa_tree_gt_roots[2]; | |
738 | } | |
739 | ||
56686608 | 740 | /* Modify the name P in-place so that it is a valid HSA identifier. */ |
741 | ||
742 | void | |
743 | hsa_sanitize_name (char *p) | |
744 | { | |
745 | for (; *p; p++) | |
746 | if (*p == '.' || *p == '-') | |
747 | *p = '_'; | |
748 | } | |
749 | ||
750 | /* Clone the name P, set trailing ampersand and sanitize the name. */ | |
751 | ||
752 | char * | |
753 | hsa_brig_function_name (const char *p) | |
754 | { | |
755 | unsigned len = strlen (p); | |
756 | char *buf = XNEWVEC (char, len + 2); | |
757 | ||
758 | buf[0] = '&'; | |
759 | buf[len + 1] = '\0'; | |
760 | memcpy (buf + 1, p, len); | |
761 | ||
762 | hsa_sanitize_name (buf); | |
763 | return buf; | |
764 | } | |
765 | ||
766 | /* Return declaration name if exists. */ | |
767 | ||
768 | const char * | |
769 | hsa_get_declaration_name (tree decl) | |
770 | { | |
771 | if (!DECL_NAME (decl)) | |
772 | { | |
773 | char buf[64]; | |
774 | snprintf (buf, 64, "__hsa_anonymous_%i", DECL_UID (decl)); | |
775 | const char *ggc_str = ggc_strdup (buf); | |
776 | return ggc_str; | |
777 | } | |
778 | ||
779 | tree name_tree; | |
780 | if (TREE_CODE (decl) == FUNCTION_DECL | |
781 | || (TREE_CODE (decl) == VAR_DECL && is_global_var (decl))) | |
782 | name_tree = DECL_ASSEMBLER_NAME (decl); | |
783 | else | |
784 | name_tree = DECL_NAME (decl); | |
785 | ||
786 | const char *name = IDENTIFIER_POINTER (name_tree); | |
787 | /* User-defined assembly names have prepended asterisk symbol. */ | |
788 | if (name[0] == '*') | |
789 | name++; | |
790 | ||
791 | return name; | |
792 | } | |
793 | ||
794 | void | |
795 | hsa_summary_t::link_functions (cgraph_node *gpu, cgraph_node *host, | |
796 | hsa_function_kind kind, bool gridified_kernel_p) | |
797 | { | |
798 | hsa_function_summary *gpu_summary = get (gpu); | |
799 | hsa_function_summary *host_summary = get (host); | |
800 | ||
801 | gpu_summary->m_kind = kind; | |
802 | host_summary->m_kind = kind; | |
803 | ||
804 | gpu_summary->m_gpu_implementation_p = true; | |
805 | host_summary->m_gpu_implementation_p = false; | |
806 | ||
807 | gpu_summary->m_gridified_kernel_p = gridified_kernel_p; | |
808 | host_summary->m_gridified_kernel_p = gridified_kernel_p; | |
809 | ||
810 | gpu_summary->m_binded_function = host; | |
811 | host_summary->m_binded_function = gpu; | |
812 | ||
813 | tree gdecl = gpu->decl; | |
814 | DECL_ATTRIBUTES (gdecl) | |
815 | = tree_cons (get_identifier ("flatten"), NULL_TREE, | |
816 | DECL_ATTRIBUTES (gdecl)); | |
817 | ||
818 | tree fn_opts = DECL_FUNCTION_SPECIFIC_OPTIMIZATION (gdecl); | |
819 | if (fn_opts == NULL_TREE) | |
820 | fn_opts = optimization_default_node; | |
821 | fn_opts = copy_node (fn_opts); | |
822 | TREE_OPTIMIZATION (fn_opts)->x_flag_tree_loop_vectorize = false; | |
823 | TREE_OPTIMIZATION (fn_opts)->x_flag_tree_slp_vectorize = false; | |
824 | DECL_FUNCTION_SPECIFIC_OPTIMIZATION (gdecl) = fn_opts; | |
874e154a | 825 | |
826 | /* Create reference between a kernel and a corresponding host implementation | |
827 | to quarantee LTO streaming to a same LTRANS. */ | |
828 | if (kind == HSA_KERNEL) | |
829 | gpu->create_reference (host, IPA_REF_ADDR); | |
56686608 | 830 | } |
831 | ||
832 | /* Add a HOST function to HSA summaries. */ | |
833 | ||
834 | void | |
835 | hsa_register_kernel (cgraph_node *host) | |
836 | { | |
837 | if (hsa_summaries == NULL) | |
838 | hsa_summaries = new hsa_summary_t (symtab); | |
839 | hsa_function_summary *s = hsa_summaries->get (host); | |
840 | s->m_kind = HSA_KERNEL; | |
841 | } | |
842 | ||
843 | /* Add a pair of functions to HSA summaries. GPU is an HSA implementation of | |
844 | a HOST function. */ | |
845 | ||
846 | void | |
847 | hsa_register_kernel (cgraph_node *gpu, cgraph_node *host) | |
848 | { | |
849 | if (hsa_summaries == NULL) | |
850 | hsa_summaries = new hsa_summary_t (symtab); | |
851 | hsa_summaries->link_functions (gpu, host, HSA_KERNEL, true); | |
852 | } | |
853 | ||
854 | /* Return true if expansion of the current HSA function has already failed. */ | |
855 | ||
856 | bool | |
857 | hsa_seen_error (void) | |
858 | { | |
859 | return hsa_cfun->m_seen_error; | |
860 | } | |
861 | ||
862 | /* Mark current HSA function as failed. */ | |
863 | ||
864 | void | |
865 | hsa_fail_cfun (void) | |
866 | { | |
867 | hsa_failed_functions->add (hsa_cfun->m_decl); | |
868 | hsa_cfun->m_seen_error = true; | |
869 | } | |
870 | ||
871 | char * | |
872 | hsa_internal_fn::name () | |
873 | { | |
874 | char *name = xstrdup (internal_fn_name (m_fn)); | |
875 | for (char *ptr = name; *ptr; ptr++) | |
876 | *ptr = TOLOWER (*ptr); | |
877 | ||
878 | const char *suffix = NULL; | |
879 | if (m_type_bit_size == 32) | |
880 | suffix = "f"; | |
881 | ||
882 | if (suffix) | |
883 | { | |
884 | char *name2 = concat (name, suffix, NULL); | |
885 | free (name); | |
886 | name = name2; | |
887 | } | |
888 | ||
889 | hsa_sanitize_name (name); | |
890 | return name; | |
891 | } | |
892 | ||
893 | unsigned | |
894 | hsa_internal_fn::get_arity () | |
895 | { | |
896 | switch (m_fn) | |
897 | { | |
898 | case IFN_ACOS: | |
899 | case IFN_ASIN: | |
900 | case IFN_ATAN: | |
901 | case IFN_COS: | |
902 | case IFN_EXP: | |
903 | case IFN_EXP10: | |
904 | case IFN_EXP2: | |
905 | case IFN_EXPM1: | |
906 | case IFN_LOG: | |
907 | case IFN_LOG10: | |
908 | case IFN_LOG1P: | |
909 | case IFN_LOG2: | |
910 | case IFN_LOGB: | |
911 | case IFN_SIGNIFICAND: | |
912 | case IFN_SIN: | |
913 | case IFN_SQRT: | |
914 | case IFN_TAN: | |
915 | case IFN_CEIL: | |
916 | case IFN_FLOOR: | |
917 | case IFN_NEARBYINT: | |
918 | case IFN_RINT: | |
919 | case IFN_ROUND: | |
920 | case IFN_TRUNC: | |
921 | return 1; | |
922 | case IFN_ATAN2: | |
923 | case IFN_COPYSIGN: | |
924 | case IFN_FMOD: | |
925 | case IFN_POW: | |
926 | case IFN_REMAINDER: | |
927 | case IFN_SCALB: | |
928 | case IFN_LDEXP: | |
929 | return 2; | |
930 | break; | |
931 | case IFN_CLRSB: | |
932 | case IFN_CLZ: | |
933 | case IFN_CTZ: | |
934 | case IFN_FFS: | |
935 | case IFN_PARITY: | |
936 | case IFN_POPCOUNT: | |
937 | default: | |
938 | /* As we produce sorry message for unknown internal functions, | |
939 | reaching this label is definitely a bug. */ | |
940 | gcc_unreachable (); | |
941 | } | |
942 | } | |
943 | ||
944 | BrigType16_t | |
945 | hsa_internal_fn::get_argument_type (int n) | |
946 | { | |
947 | switch (m_fn) | |
948 | { | |
949 | case IFN_ACOS: | |
950 | case IFN_ASIN: | |
951 | case IFN_ATAN: | |
952 | case IFN_COS: | |
953 | case IFN_EXP: | |
954 | case IFN_EXP10: | |
955 | case IFN_EXP2: | |
956 | case IFN_EXPM1: | |
957 | case IFN_LOG: | |
958 | case IFN_LOG10: | |
959 | case IFN_LOG1P: | |
960 | case IFN_LOG2: | |
961 | case IFN_LOGB: | |
962 | case IFN_SIGNIFICAND: | |
963 | case IFN_SIN: | |
964 | case IFN_SQRT: | |
965 | case IFN_TAN: | |
966 | case IFN_CEIL: | |
967 | case IFN_FLOOR: | |
968 | case IFN_NEARBYINT: | |
969 | case IFN_RINT: | |
970 | case IFN_ROUND: | |
971 | case IFN_TRUNC: | |
972 | case IFN_ATAN2: | |
973 | case IFN_COPYSIGN: | |
974 | case IFN_FMOD: | |
975 | case IFN_POW: | |
976 | case IFN_REMAINDER: | |
977 | case IFN_SCALB: | |
978 | return hsa_float_for_bitsize (m_type_bit_size); | |
979 | case IFN_LDEXP: | |
980 | { | |
981 | if (n == -1 || n == 0) | |
982 | return hsa_float_for_bitsize (m_type_bit_size); | |
983 | else | |
984 | return BRIG_TYPE_S32; | |
985 | } | |
986 | default: | |
987 | /* As we produce sorry message for unknown internal functions, | |
988 | reaching this label is definitely a bug. */ | |
989 | gcc_unreachable (); | |
990 | } | |
991 | } | |
992 | ||
993 | #include "gt-hsa.h" |