]>
Commit | Line | Data |
---|---|---|
738f2522 | 1 | /* Target code for NVPTX. |
818ab71a | 2 | Copyright (C) 2014-2016 Free Software Foundation, Inc. |
738f2522 BS |
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 | #include "config.h" | |
3a4d1cb1 | 22 | #include <sstream> |
738f2522 BS |
23 | #include "system.h" |
24 | #include "coretypes.h" | |
c7131fb2 | 25 | #include "backend.h" |
e11c4407 | 26 | #include "target.h" |
738f2522 | 27 | #include "rtl.h" |
e11c4407 AM |
28 | #include "tree.h" |
29 | #include "cfghooks.h" | |
c7131fb2 | 30 | #include "df.h" |
e11c4407 AM |
31 | #include "tm_p.h" |
32 | #include "expmed.h" | |
33 | #include "optabs.h" | |
34 | #include "regs.h" | |
35 | #include "emit-rtl.h" | |
36 | #include "recog.h" | |
37 | #include "diagnostic.h" | |
40e23961 | 38 | #include "alias.h" |
738f2522 BS |
39 | #include "insn-flags.h" |
40 | #include "output.h" | |
41 | #include "insn-attr.h" | |
36566b39 | 42 | #include "flags.h" |
36566b39 PK |
43 | #include "dojump.h" |
44 | #include "explow.h" | |
45 | #include "calls.h" | |
36566b39 PK |
46 | #include "varasm.h" |
47 | #include "stmt.h" | |
738f2522 | 48 | #include "expr.h" |
738f2522 BS |
49 | #include "tm-preds.h" |
50 | #include "tm-constrs.h" | |
738f2522 BS |
51 | #include "langhooks.h" |
52 | #include "dbxout.h" | |
738f2522 | 53 | #include "cfgrtl.h" |
d88cd9c4 | 54 | #include "gimple.h" |
738f2522 | 55 | #include "stor-layout.h" |
738f2522 | 56 | #include "builtins.h" |
3e32ee19 NS |
57 | #include "omp-low.h" |
58 | #include "gomp-constants.h" | |
d88cd9c4 | 59 | #include "dumpfile.h" |
f3552158 NS |
60 | #include "internal-fn.h" |
61 | #include "gimple-iterator.h" | |
62 | #include "stringpool.h" | |
63 | #include "tree-ssa-operands.h" | |
64 | #include "tree-ssanames.h" | |
65 | #include "gimplify.h" | |
66 | #include "tree-phinodes.h" | |
67 | #include "cfgloop.h" | |
68 | #include "fold-const.h" | |
738f2522 | 69 | |
994c5d85 | 70 | /* This file should be included last. */ |
d58627a0 RS |
71 | #include "target-def.h" |
72 | ||
59263259 NS |
73 | /* The kind of shuffe instruction. */ |
74 | enum nvptx_shuffle_kind | |
75 | { | |
76 | SHUFFLE_UP, | |
77 | SHUFFLE_DOWN, | |
78 | SHUFFLE_BFLY, | |
79 | SHUFFLE_IDX, | |
80 | SHUFFLE_MAX | |
81 | }; | |
d88cd9c4 | 82 | |
9a863523 NS |
83 | /* The various PTX memory areas an object might reside in. */ |
84 | enum nvptx_data_area | |
85 | { | |
86 | DATA_AREA_GENERIC, | |
87 | DATA_AREA_GLOBAL, | |
88 | DATA_AREA_SHARED, | |
89 | DATA_AREA_LOCAL, | |
90 | DATA_AREA_CONST, | |
91 | DATA_AREA_PARAM, | |
92 | DATA_AREA_MAX | |
93 | }; | |
94 | ||
95 | /* We record the data area in the target symbol flags. */ | |
96 | #define SYMBOL_DATA_AREA(SYM) \ | |
97 | (nvptx_data_area)((SYMBOL_REF_FLAGS (SYM) >> SYMBOL_FLAG_MACH_DEP_SHIFT) \ | |
98 | & 7) | |
99 | #define SET_SYMBOL_DATA_AREA(SYM,AREA) \ | |
100 | (SYMBOL_REF_FLAGS (SYM) |= (AREA) << SYMBOL_FLAG_MACH_DEP_SHIFT) | |
101 | ||
738f2522 BS |
102 | /* Record the function decls we've written, and the libfuncs and function |
103 | decls corresponding to them. */ | |
104 | static std::stringstream func_decls; | |
f3dba894 | 105 | |
6c907cff | 106 | struct declared_libfunc_hasher : ggc_cache_ptr_hash<rtx_def> |
f3dba894 TS |
107 | { |
108 | static hashval_t hash (rtx x) { return htab_hash_pointer (x); } | |
109 | static bool equal (rtx a, rtx b) { return a == b; } | |
110 | }; | |
111 | ||
112 | static GTY((cache)) | |
113 | hash_table<declared_libfunc_hasher> *declared_libfuncs_htab; | |
114 | ||
6c907cff | 115 | struct tree_hasher : ggc_cache_ptr_hash<tree_node> |
f3dba894 TS |
116 | { |
117 | static hashval_t hash (tree t) { return htab_hash_pointer (t); } | |
118 | static bool equal (tree a, tree b) { return a == b; } | |
119 | }; | |
120 | ||
121 | static GTY((cache)) hash_table<tree_hasher> *declared_fndecls_htab; | |
122 | static GTY((cache)) hash_table<tree_hasher> *needed_fndecls_htab; | |
738f2522 | 123 | |
f3552158 NS |
124 | /* Buffer needed to broadcast across workers. This is used for both |
125 | worker-neutering and worker broadcasting. It is shared by all | |
126 | functions emitted. The buffer is placed in shared memory. It'd be | |
127 | nice if PTX supported common blocks, because then this could be | |
128 | shared across TUs (taking the largest size). */ | |
d88cd9c4 NS |
129 | static unsigned worker_bcast_size; |
130 | static unsigned worker_bcast_align; | |
d88cd9c4 NS |
131 | static GTY(()) rtx worker_bcast_sym; |
132 | ||
f3552158 NS |
133 | /* Buffer needed for worker reductions. This has to be distinct from |
134 | the worker broadcast array, as both may be live concurrently. */ | |
135 | static unsigned worker_red_size; | |
136 | static unsigned worker_red_align; | |
f3552158 NS |
137 | static GTY(()) rtx worker_red_sym; |
138 | ||
33f47f42 NS |
139 | /* Global lock variable, needed for 128bit worker & gang reductions. */ |
140 | static GTY(()) tree global_lock_var; | |
141 | ||
738f2522 BS |
142 | /* Allocate a new, cleared machine_function structure. */ |
143 | ||
144 | static struct machine_function * | |
145 | nvptx_init_machine_status (void) | |
146 | { | |
147 | struct machine_function *p = ggc_cleared_alloc<machine_function> (); | |
44c068ae | 148 | p->return_mode = VOIDmode; |
738f2522 BS |
149 | return p; |
150 | } | |
151 | ||
152 | /* Implement TARGET_OPTION_OVERRIDE. */ | |
153 | ||
154 | static void | |
155 | nvptx_option_override (void) | |
156 | { | |
157 | init_machine_status = nvptx_init_machine_status; | |
158 | /* Gives us a predictable order, which we need especially for variables. */ | |
159 | flag_toplevel_reorder = 1; | |
160 | /* Assumes that it will see only hard registers. */ | |
161 | flag_var_tracking = 0; | |
1e5154e7 NS |
162 | |
163 | if (write_symbols == DBX_DEBUG) | |
164 | /* The stabs testcases want to know stabs isn't supported. */ | |
165 | sorry ("stabs debug format not supported"); | |
166 | ||
167 | /* Actually we don't have any debug format, but don't be | |
168 | unneccesarily noisy. */ | |
f324806d NS |
169 | write_symbols = NO_DEBUG; |
170 | debug_info_level = DINFO_LEVEL_NONE; | |
738f2522 | 171 | |
dba619f3 NS |
172 | if (nvptx_optimize < 0) |
173 | nvptx_optimize = optimize > 0; | |
174 | ||
f3dba894 TS |
175 | declared_fndecls_htab = hash_table<tree_hasher>::create_ggc (17); |
176 | needed_fndecls_htab = hash_table<tree_hasher>::create_ggc (17); | |
738f2522 | 177 | declared_libfuncs_htab |
f3dba894 | 178 | = hash_table<declared_libfunc_hasher>::create_ggc (17); |
d88cd9c4 | 179 | |
15ab6f00 | 180 | worker_bcast_sym = gen_rtx_SYMBOL_REF (Pmode, "__worker_bcast"); |
9a863523 | 181 | SET_SYMBOL_DATA_AREA (worker_bcast_sym, DATA_AREA_SHARED); |
d88cd9c4 | 182 | worker_bcast_align = GET_MODE_ALIGNMENT (SImode) / BITS_PER_UNIT; |
f3552158 | 183 | |
15ab6f00 | 184 | worker_red_sym = gen_rtx_SYMBOL_REF (Pmode, "__worker_red"); |
9a863523 | 185 | SET_SYMBOL_DATA_AREA (worker_red_sym, DATA_AREA_SHARED); |
f3552158 | 186 | worker_red_align = GET_MODE_ALIGNMENT (SImode) / BITS_PER_UNIT; |
738f2522 BS |
187 | } |
188 | ||
738f2522 BS |
189 | /* Return a ptx type for MODE. If PROMOTE, then use .u32 for QImode to |
190 | deal with ptx ideosyncracies. */ | |
191 | ||
192 | const char * | |
193 | nvptx_ptx_type_from_mode (machine_mode mode, bool promote) | |
194 | { | |
195 | switch (mode) | |
196 | { | |
197 | case BLKmode: | |
198 | return ".b8"; | |
199 | case BImode: | |
200 | return ".pred"; | |
201 | case QImode: | |
202 | if (promote) | |
203 | return ".u32"; | |
204 | else | |
205 | return ".u8"; | |
206 | case HImode: | |
207 | return ".u16"; | |
208 | case SImode: | |
209 | return ".u32"; | |
210 | case DImode: | |
211 | return ".u64"; | |
212 | ||
213 | case SFmode: | |
214 | return ".f32"; | |
215 | case DFmode: | |
216 | return ".f64"; | |
217 | ||
218 | default: | |
219 | gcc_unreachable (); | |
220 | } | |
221 | } | |
222 | ||
9a863523 NS |
223 | /* Encode the PTX data area that DECL (which might not actually be a |
224 | _DECL) should reside in. */ | |
7b8edc29 | 225 | |
9a863523 NS |
226 | static void |
227 | nvptx_encode_section_info (tree decl, rtx rtl, int first) | |
7b8edc29 | 228 | { |
9a863523 NS |
229 | default_encode_section_info (decl, rtl, first); |
230 | if (first && MEM_P (rtl)) | |
231 | { | |
232 | nvptx_data_area area = DATA_AREA_GENERIC; | |
7b8edc29 | 233 | |
9a863523 NS |
234 | if (TREE_CONSTANT (decl)) |
235 | area = DATA_AREA_CONST; | |
236 | else if (TREE_CODE (decl) == VAR_DECL) | |
237 | /* TODO: This would be a good place to check for a .shared or | |
238 | other section name. */ | |
239 | area = TREE_READONLY (decl) ? DATA_AREA_CONST : DATA_AREA_GLOBAL; | |
7b8edc29 | 240 | |
9a863523 NS |
241 | SET_SYMBOL_DATA_AREA (XEXP (rtl, 0), area); |
242 | } | |
243 | } | |
244 | ||
245 | /* Return the PTX name of the data area in which SYM should be | |
246 | placed. The symbol must have already been processed by | |
247 | nvptx_encode_seciton_info, or equivalent. */ | |
248 | ||
249 | static const char * | |
250 | section_for_sym (rtx sym) | |
251 | { | |
252 | nvptx_data_area area = SYMBOL_DATA_AREA (sym); | |
253 | /* Same order as nvptx_data_area enum. */ | |
254 | static char const *const areas[] = | |
255 | {"", ".global", ".shared", ".local", ".const", ".param"}; | |
256 | ||
257 | return areas[area]; | |
258 | } | |
259 | ||
260 | /* Similarly for a decl. */ | |
261 | ||
262 | static const char * | |
263 | section_for_decl (const_tree decl) | |
264 | { | |
265 | return section_for_sym (XEXP (DECL_RTL (CONST_CAST (tree, decl)), 0)); | |
7b8edc29 NS |
266 | } |
267 | ||
b699adcc NS |
268 | /* Check NAME for special function names and redirect them by returning a |
269 | replacement. This applies to malloc, free and realloc, for which we | |
270 | want to use libgcc wrappers, and call, which triggers a bug in ptxas. */ | |
271 | ||
272 | static const char * | |
273 | nvptx_name_replacement (const char *name) | |
274 | { | |
275 | if (strcmp (name, "call") == 0) | |
276 | return "__nvptx_call"; | |
277 | if (strcmp (name, "malloc") == 0) | |
278 | return "__nvptx_malloc"; | |
279 | if (strcmp (name, "free") == 0) | |
280 | return "__nvptx_free"; | |
281 | if (strcmp (name, "realloc") == 0) | |
282 | return "__nvptx_realloc"; | |
283 | return name; | |
284 | } | |
285 | ||
d7479262 NS |
286 | /* If MODE should be treated as two registers of an inner mode, return |
287 | that inner mode. Otherwise return VOIDmode. */ | |
738f2522 | 288 | |
d7479262 NS |
289 | static machine_mode |
290 | maybe_split_mode (machine_mode mode) | |
738f2522 | 291 | { |
738f2522 | 292 | if (COMPLEX_MODE_P (mode)) |
d7479262 | 293 | return GET_MODE_INNER (mode); |
738f2522 | 294 | |
738f2522 | 295 | if (mode == TImode) |
d7479262 NS |
296 | return DImode; |
297 | ||
298 | return VOIDmode; | |
738f2522 BS |
299 | } |
300 | ||
f313d112 NS |
301 | /* Output a register, subreg, or register pair (with optional |
302 | enclosing braces). */ | |
303 | ||
304 | static void | |
305 | output_reg (FILE *file, unsigned regno, machine_mode inner_mode, | |
306 | int subreg_offset = -1) | |
307 | { | |
308 | if (inner_mode == VOIDmode) | |
309 | { | |
310 | if (HARD_REGISTER_NUM_P (regno)) | |
311 | fprintf (file, "%s", reg_names[regno]); | |
312 | else | |
313 | fprintf (file, "%%r%d", regno); | |
314 | } | |
315 | else if (subreg_offset >= 0) | |
316 | { | |
317 | output_reg (file, regno, VOIDmode); | |
318 | fprintf (file, "$%d", subreg_offset); | |
319 | } | |
320 | else | |
321 | { | |
322 | if (subreg_offset == -1) | |
323 | fprintf (file, "{"); | |
324 | output_reg (file, regno, inner_mode, GET_MODE_SIZE (inner_mode)); | |
325 | fprintf (file, ","); | |
326 | output_reg (file, regno, inner_mode, 0); | |
327 | if (subreg_offset == -1) | |
328 | fprintf (file, "}"); | |
329 | } | |
330 | } | |
331 | ||
d88cd9c4 NS |
332 | /* Emit forking instructions for MASK. */ |
333 | ||
334 | static void | |
335 | nvptx_emit_forking (unsigned mask, bool is_call) | |
336 | { | |
337 | mask &= (GOMP_DIM_MASK (GOMP_DIM_WORKER) | |
338 | | GOMP_DIM_MASK (GOMP_DIM_VECTOR)); | |
339 | if (mask) | |
340 | { | |
341 | rtx op = GEN_INT (mask | (is_call << GOMP_DIM_MAX)); | |
342 | ||
343 | /* Emit fork at all levels. This helps form SESE regions, as | |
344 | it creates a block with a single successor before entering a | |
345 | partitooned region. That is a good candidate for the end of | |
346 | an SESE region. */ | |
347 | if (!is_call) | |
348 | emit_insn (gen_nvptx_fork (op)); | |
349 | emit_insn (gen_nvptx_forked (op)); | |
350 | } | |
351 | } | |
352 | ||
353 | /* Emit joining instructions for MASK. */ | |
354 | ||
355 | static void | |
356 | nvptx_emit_joining (unsigned mask, bool is_call) | |
357 | { | |
358 | mask &= (GOMP_DIM_MASK (GOMP_DIM_WORKER) | |
359 | | GOMP_DIM_MASK (GOMP_DIM_VECTOR)); | |
360 | if (mask) | |
361 | { | |
362 | rtx op = GEN_INT (mask | (is_call << GOMP_DIM_MAX)); | |
363 | ||
364 | /* Emit joining for all non-call pars to ensure there's a single | |
365 | predecessor for the block the join insn ends up in. This is | |
366 | needed for skipping entire loops. */ | |
367 | if (!is_call) | |
368 | emit_insn (gen_nvptx_joining (op)); | |
369 | emit_insn (gen_nvptx_join (op)); | |
370 | } | |
371 | } | |
372 | ||
738f2522 | 373 | \f |
44eba92d NS |
374 | /* Determine whether MODE and TYPE (possibly NULL) should be passed or |
375 | returned in memory. Integer and floating types supported by the | |
376 | machine are passed in registers, everything else is passed in | |
377 | memory. Complex types are split. */ | |
378 | ||
379 | static bool | |
380 | pass_in_memory (machine_mode mode, const_tree type, bool for_return) | |
381 | { | |
382 | if (type) | |
383 | { | |
384 | if (AGGREGATE_TYPE_P (type)) | |
385 | return true; | |
386 | if (TREE_CODE (type) == VECTOR_TYPE) | |
387 | return true; | |
388 | } | |
389 | ||
390 | if (!for_return && COMPLEX_MODE_P (mode)) | |
391 | /* Complex types are passed as two underlying args. */ | |
392 | mode = GET_MODE_INNER (mode); | |
393 | ||
394 | if (GET_MODE_CLASS (mode) != MODE_INT | |
395 | && GET_MODE_CLASS (mode) != MODE_FLOAT) | |
396 | return true; | |
397 | ||
398 | if (GET_MODE_SIZE (mode) > UNITS_PER_WORD) | |
399 | return true; | |
400 | ||
401 | return false; | |
402 | } | |
403 | ||
404 | /* A non-memory argument of mode MODE is being passed, determine the mode it | |
405 | should be promoted to. This is also used for determining return | |
406 | type promotion. */ | |
407 | ||
408 | static machine_mode | |
409 | promote_arg (machine_mode mode, bool prototyped) | |
410 | { | |
411 | if (!prototyped && mode == SFmode) | |
412 | /* K&R float promotion for unprototyped functions. */ | |
413 | mode = DFmode; | |
414 | else if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (SImode)) | |
415 | mode = SImode; | |
416 | ||
417 | return mode; | |
418 | } | |
419 | ||
420 | /* A non-memory return type of MODE is being returned. Determine the | |
421 | mode it should be promoted to. */ | |
422 | ||
423 | static machine_mode | |
424 | promote_return (machine_mode mode) | |
425 | { | |
426 | return promote_arg (mode, true); | |
427 | } | |
428 | ||
e74f7152 | 429 | /* Implement TARGET_FUNCTION_ARG. */ |
dc3d2aeb | 430 | |
e74f7152 | 431 | static rtx |
b49e35a9 | 432 | nvptx_function_arg (cumulative_args_t ARG_UNUSED (cum_v), machine_mode mode, |
e74f7152 NS |
433 | const_tree, bool named) |
434 | { | |
b49e35a9 | 435 | if (mode == VOIDmode || !named) |
e74f7152 | 436 | return NULL_RTX; |
738f2522 | 437 | |
b49e35a9 | 438 | return gen_reg_rtx (mode); |
e74f7152 NS |
439 | } |
440 | ||
441 | /* Implement TARGET_FUNCTION_INCOMING_ARG. */ | |
442 | ||
443 | static rtx | |
444 | nvptx_function_incoming_arg (cumulative_args_t cum_v, machine_mode mode, | |
445 | const_tree, bool named) | |
738f2522 | 446 | { |
e74f7152 | 447 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); |
dc3d2aeb | 448 | |
b49e35a9 | 449 | if (mode == VOIDmode || !named) |
e74f7152 | 450 | return NULL_RTX; |
738f2522 | 451 | |
e74f7152 NS |
452 | /* No need to deal with split modes here, the only case that can |
453 | happen is complex modes and those are dealt with by | |
454 | TARGET_SPLIT_COMPLEX_ARG. */ | |
455 | return gen_rtx_UNSPEC (mode, | |
456 | gen_rtvec (1, GEN_INT (cum->count)), | |
457 | UNSPEC_ARG_REG); | |
458 | } | |
459 | ||
460 | /* Implement TARGET_FUNCTION_ARG_ADVANCE. */ | |
461 | ||
462 | static void | |
463 | nvptx_function_arg_advance (cumulative_args_t cum_v, | |
464 | machine_mode ARG_UNUSED (mode), | |
465 | const_tree ARG_UNUSED (type), | |
466 | bool ARG_UNUSED (named)) | |
467 | { | |
468 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
b49e35a9 | 469 | |
e74f7152 NS |
470 | cum->count++; |
471 | } | |
472 | ||
473 | /* Handle the TARGET_STRICT_ARGUMENT_NAMING target hook. | |
474 | ||
475 | For nvptx, we know how to handle functions declared as stdarg: by | |
476 | passing an extra pointer to the unnamed arguments. However, the | |
477 | Fortran frontend can produce a different situation, where a | |
478 | function pointer is declared with no arguments, but the actual | |
479 | function and calls to it take more arguments. In that case, we | |
480 | want to ensure the call matches the definition of the function. */ | |
481 | ||
482 | static bool | |
483 | nvptx_strict_argument_naming (cumulative_args_t cum_v) | |
484 | { | |
485 | CUMULATIVE_ARGS *cum = get_cumulative_args (cum_v); | |
b49e35a9 | 486 | |
e74f7152 NS |
487 | return cum->fntype == NULL_TREE || stdarg_p (cum->fntype); |
488 | } | |
489 | ||
e74f7152 NS |
490 | /* Implement TARGET_LIBCALL_VALUE. */ |
491 | ||
492 | static rtx | |
493 | nvptx_libcall_value (machine_mode mode, const_rtx) | |
494 | { | |
44c068ae | 495 | if (!cfun->machine->doing_call) |
e74f7152 NS |
496 | /* Pretend to return in a hard reg for early uses before pseudos can be |
497 | generated. */ | |
498 | return gen_rtx_REG (mode, NVPTX_RETURN_REGNUM); | |
b49e35a9 | 499 | |
e74f7152 NS |
500 | return gen_reg_rtx (mode); |
501 | } | |
502 | ||
503 | /* TARGET_FUNCTION_VALUE implementation. Returns an RTX representing the place | |
504 | where function FUNC returns or receives a value of data type TYPE. */ | |
505 | ||
506 | static rtx | |
44eba92d | 507 | nvptx_function_value (const_tree type, const_tree ARG_UNUSED (func), |
e74f7152 NS |
508 | bool outgoing) |
509 | { | |
44eba92d NS |
510 | machine_mode mode = promote_return (TYPE_MODE (type)); |
511 | ||
e74f7152 | 512 | if (outgoing) |
5c036f3f | 513 | { |
44c068ae | 514 | cfun->machine->return_mode = mode; |
5c036f3f NS |
515 | return gen_rtx_REG (mode, NVPTX_RETURN_REGNUM); |
516 | } | |
b49e35a9 NS |
517 | |
518 | return nvptx_libcall_value (mode, NULL_RTX); | |
e74f7152 NS |
519 | } |
520 | ||
521 | /* Implement TARGET_FUNCTION_VALUE_REGNO_P. */ | |
522 | ||
523 | static bool | |
524 | nvptx_function_value_regno_p (const unsigned int regno) | |
525 | { | |
526 | return regno == NVPTX_RETURN_REGNUM; | |
527 | } | |
528 | ||
529 | /* Types with a mode other than those supported by the machine are passed by | |
530 | reference in memory. */ | |
531 | ||
532 | static bool | |
5563d5c0 NS |
533 | nvptx_pass_by_reference (cumulative_args_t ARG_UNUSED (cum), |
534 | machine_mode mode, const_tree type, | |
535 | bool ARG_UNUSED (named)) | |
e74f7152 | 536 | { |
44eba92d | 537 | return pass_in_memory (mode, type, false); |
e74f7152 NS |
538 | } |
539 | ||
540 | /* Implement TARGET_RETURN_IN_MEMORY. */ | |
541 | ||
542 | static bool | |
543 | nvptx_return_in_memory (const_tree type, const_tree) | |
544 | { | |
44eba92d | 545 | return pass_in_memory (TYPE_MODE (type), type, true); |
e74f7152 NS |
546 | } |
547 | ||
548 | /* Implement TARGET_PROMOTE_FUNCTION_MODE. */ | |
549 | ||
550 | static machine_mode | |
551 | nvptx_promote_function_mode (const_tree type, machine_mode mode, | |
44eba92d | 552 | int *ARG_UNUSED (punsignedp), |
e74f7152 NS |
553 | const_tree funtype, int for_return) |
554 | { | |
44eba92d | 555 | return promote_arg (mode, for_return || !type || TYPE_ARG_TYPES (funtype)); |
e74f7152 NS |
556 | } |
557 | ||
e74f7152 NS |
558 | /* Helper for write_arg. Emit a single PTX argument of MODE, either |
559 | in a prototype, or as copy in a function prologue. ARGNO is the | |
560 | index of this argument in the PTX function. FOR_REG is negative, | |
561 | if we're emitting the PTX prototype. It is zero if we're copying | |
562 | to an argument register and it is greater than zero if we're | |
563 | copying to a specific hard register. */ | |
564 | ||
565 | static int | |
1f065954 NS |
566 | write_arg_mode (std::stringstream &s, int for_reg, int argno, |
567 | machine_mode mode) | |
e74f7152 NS |
568 | { |
569 | const char *ptx_type = nvptx_ptx_type_from_mode (mode, false); | |
570 | ||
dc3d2aeb NS |
571 | if (for_reg < 0) |
572 | { | |
573 | /* Writing PTX prototype. */ | |
574 | s << (argno ? ", " : " ("); | |
e74f7152 | 575 | s << ".param" << ptx_type << " %in_ar" << argno; |
dc3d2aeb NS |
576 | } |
577 | else | |
578 | { | |
e74f7152 | 579 | s << "\t.reg" << ptx_type << " "; |
dc3d2aeb NS |
580 | if (for_reg) |
581 | s << reg_names[for_reg]; | |
582 | else | |
583 | s << "%ar" << argno; | |
584 | s << ";\n"; | |
5563d5c0 NS |
585 | if (argno >= 0) |
586 | { | |
587 | s << "\tld.param" << ptx_type << " "; | |
588 | if (for_reg) | |
589 | s << reg_names[for_reg]; | |
590 | else | |
591 | s << "%ar" << argno; | |
592 | s << ", [%in_ar" << argno << "];\n"; | |
593 | } | |
dc3d2aeb NS |
594 | } |
595 | return argno + 1; | |
738f2522 BS |
596 | } |
597 | ||
e74f7152 | 598 | /* Process function parameter TYPE to emit one or more PTX |
1f065954 | 599 | arguments. S, FOR_REG and ARGNO as for write_arg_mode. PROTOTYPED |
44eba92d NS |
600 | is true, if this is a prototyped function, rather than an old-style |
601 | C declaration. Returns the next argument number to use. | |
e74f7152 NS |
602 | |
603 | The promotion behaviour here must match the regular GCC function | |
604 | parameter marshalling machinery. */ | |
605 | ||
606 | static int | |
1f065954 NS |
607 | write_arg_type (std::stringstream &s, int for_reg, int argno, |
608 | tree type, bool prototyped) | |
e74f7152 NS |
609 | { |
610 | machine_mode mode = TYPE_MODE (type); | |
611 | ||
612 | if (mode == VOIDmode) | |
613 | return argno; | |
614 | ||
44eba92d | 615 | if (pass_in_memory (mode, type, false)) |
e74f7152 | 616 | mode = Pmode; |
44eba92d NS |
617 | else |
618 | { | |
619 | bool split = TREE_CODE (type) == COMPLEX_TYPE; | |
e74f7152 | 620 | |
44eba92d NS |
621 | if (split) |
622 | { | |
623 | /* Complex types are sent as two separate args. */ | |
624 | type = TREE_TYPE (type); | |
5563d5c0 | 625 | mode = TYPE_MODE (type); |
44eba92d NS |
626 | prototyped = true; |
627 | } | |
e74f7152 | 628 | |
44eba92d NS |
629 | mode = promote_arg (mode, prototyped); |
630 | if (split) | |
1f065954 | 631 | argno = write_arg_mode (s, for_reg, argno, mode); |
e74f7152 | 632 | } |
e74f7152 | 633 | |
1f065954 NS |
634 | return write_arg_mode (s, for_reg, argno, mode); |
635 | } | |
636 | ||
637 | /* Emit a PTX return as a prototype or function prologue declaration | |
638 | for MODE. */ | |
639 | ||
640 | static void | |
641 | write_return_mode (std::stringstream &s, bool for_proto, machine_mode mode) | |
642 | { | |
643 | const char *ptx_type = nvptx_ptx_type_from_mode (mode, false); | |
644 | const char *pfx = "\t.reg"; | |
645 | const char *sfx = ";\n"; | |
646 | ||
647 | if (for_proto) | |
648 | pfx = "(.param", sfx = "_out) "; | |
649 | ||
650 | s << pfx << ptx_type << " " << reg_names[NVPTX_RETURN_REGNUM] << sfx; | |
e74f7152 NS |
651 | } |
652 | ||
44eba92d | 653 | /* Process a function return TYPE to emit a PTX return as a prototype |
1f065954 NS |
654 | or function prologue declaration. Returns true if return is via an |
655 | additional pointer parameter. The promotion behaviour here must | |
656 | match the regular GCC function return mashalling. */ | |
44eba92d | 657 | |
0beb7c71 | 658 | static bool |
1f065954 | 659 | write_return_type (std::stringstream &s, bool for_proto, tree type) |
0beb7c71 NS |
660 | { |
661 | machine_mode mode = TYPE_MODE (type); | |
0beb7c71 | 662 | |
44eba92d NS |
663 | if (mode == VOIDmode) |
664 | return false; | |
665 | ||
666 | bool return_in_mem = pass_in_memory (mode, type, true); | |
667 | ||
668 | if (return_in_mem) | |
0beb7c71 | 669 | { |
44eba92d NS |
670 | if (for_proto) |
671 | return return_in_mem; | |
672 | ||
673 | /* Named return values can cause us to return a pointer as well | |
674 | as expect an argument for the return location. This is | |
675 | optimization-level specific, so no caller can make use of | |
676 | this data, but more importantly for us, we must ensure it | |
677 | doesn't change the PTX prototype. */ | |
44c068ae | 678 | mode = (machine_mode) cfun->machine->return_mode; |
5c036f3f | 679 | |
44eba92d NS |
680 | if (mode == VOIDmode) |
681 | return return_in_mem; | |
682 | ||
44c068ae | 683 | /* Clear return_mode to inhibit copy of retval to non-existent |
44eba92d | 684 | retval parameter. */ |
44c068ae | 685 | cfun->machine->return_mode = VOIDmode; |
0beb7c71 NS |
686 | } |
687 | else | |
44eba92d NS |
688 | mode = promote_return (mode); |
689 | ||
1f065954 | 690 | write_return_mode (s, for_proto, mode); |
0beb7c71 NS |
691 | |
692 | return return_in_mem; | |
693 | } | |
694 | ||
738f2522 BS |
695 | /* Look for attributes in ATTRS that would indicate we must write a function |
696 | as a .entry kernel rather than a .func. Return true if one is found. */ | |
697 | ||
698 | static bool | |
699 | write_as_kernel (tree attrs) | |
700 | { | |
701 | return (lookup_attribute ("kernel", attrs) != NULL_TREE | |
702 | || lookup_attribute ("omp target entrypoint", attrs) != NULL_TREE); | |
703 | } | |
704 | ||
69823d76 NS |
705 | /* Emit a linker marker for a function decl or defn. */ |
706 | ||
707 | static void | |
708 | write_fn_marker (std::stringstream &s, bool is_defn, bool globalize, | |
709 | const char *name) | |
710 | { | |
711 | s << "\n// BEGIN"; | |
712 | if (globalize) | |
713 | s << " GLOBAL"; | |
714 | s << " FUNCTION " << (is_defn ? "DEF: " : "DECL: "); | |
715 | s << name << "\n"; | |
716 | } | |
717 | ||
718 | /* Emit a linker marker for a variable decl or defn. */ | |
719 | ||
720 | static void | |
721 | write_var_marker (FILE *file, bool is_defn, bool globalize, const char *name) | |
722 | { | |
723 | fprintf (file, "\n// BEGIN%s VAR %s: ", | |
724 | globalize ? " GLOBAL" : "", | |
725 | is_defn ? "DEF" : "DECL"); | |
726 | assemble_name_raw (file, name); | |
727 | fputs ("\n", file); | |
728 | } | |
729 | ||
b699adcc NS |
730 | /* Write a .func or .kernel declaration or definition along with |
731 | a helper comment for use by ld. S is the stream to write to, DECL | |
732 | the decl for the function with name NAME. For definitions, emit | |
733 | a declaration too. */ | |
738f2522 | 734 | |
b699adcc NS |
735 | static const char * |
736 | write_fn_proto (std::stringstream &s, bool is_defn, | |
737 | const char *name, const_tree decl) | |
738f2522 | 738 | { |
b699adcc NS |
739 | if (is_defn) |
740 | /* Emit a declaration. The PTX assembler gets upset without it. */ | |
741 | name = write_fn_proto (s, false, name, decl); | |
742 | else | |
738f2522 | 743 | { |
b699adcc NS |
744 | /* Avoid repeating the name replacement. */ |
745 | name = nvptx_name_replacement (name); | |
746 | if (name[0] == '*') | |
747 | name++; | |
738f2522 BS |
748 | } |
749 | ||
69823d76 | 750 | write_fn_marker (s, is_defn, TREE_PUBLIC (decl), name); |
b699adcc NS |
751 | |
752 | /* PTX declaration. */ | |
738f2522 BS |
753 | if (DECL_EXTERNAL (decl)) |
754 | s << ".extern "; | |
755 | else if (TREE_PUBLIC (decl)) | |
0766660b | 756 | s << (DECL_WEAK (decl) ? ".weak " : ".visible "); |
b699adcc | 757 | s << (write_as_kernel (DECL_ATTRIBUTES (decl)) ? ".entry " : ".func "); |
738f2522 | 758 | |
b699adcc NS |
759 | tree fntype = TREE_TYPE (decl); |
760 | tree result_type = TREE_TYPE (fntype); | |
738f2522 BS |
761 | |
762 | /* Declare the result. */ | |
1f065954 | 763 | bool return_in_mem = write_return_type (s, true, result_type); |
738f2522 | 764 | |
b699adcc NS |
765 | s << name; |
766 | ||
dc3d2aeb | 767 | int argno = 0; |
b699adcc NS |
768 | |
769 | /* Emit argument list. */ | |
770 | if (return_in_mem) | |
1f065954 | 771 | argno = write_arg_type (s, -1, argno, ptr_type_node, true); |
df1bdded | 772 | |
b699adcc NS |
773 | /* We get: |
774 | NULL in TYPE_ARG_TYPES, for old-style functions | |
775 | NULL in DECL_ARGUMENTS, for builtin functions without another | |
776 | declaration. | |
777 | So we have to pick the best one we have. */ | |
778 | tree args = TYPE_ARG_TYPES (fntype); | |
dc3d2aeb NS |
779 | bool prototyped = true; |
780 | if (!args) | |
781 | { | |
782 | args = DECL_ARGUMENTS (decl); | |
783 | prototyped = false; | |
784 | } | |
738f2522 | 785 | |
b699adcc NS |
786 | for (; args; args = TREE_CHAIN (args)) |
787 | { | |
dc3d2aeb | 788 | tree type = prototyped ? TREE_VALUE (args) : TREE_TYPE (args); |
b699adcc | 789 | |
1f065954 | 790 | argno = write_arg_type (s, -1, argno, type, prototyped); |
738f2522 | 791 | } |
738f2522 | 792 | |
b699adcc | 793 | if (stdarg_p (fntype)) |
1f065954 | 794 | argno = write_arg_type (s, -1, argno, ptr_type_node, true); |
738f2522 | 795 | |
b699adcc | 796 | if (DECL_STATIC_CHAIN (decl)) |
1f065954 | 797 | argno = write_arg_type (s, -1, argno, ptr_type_node, true); |
b699adcc | 798 | |
dc3d2aeb | 799 | if (!argno && strcmp (name, "main") == 0) |
b699adcc | 800 | { |
1f065954 NS |
801 | argno = write_arg_type (s, -1, argno, integer_type_node, true); |
802 | argno = write_arg_type (s, -1, argno, ptr_type_node, true); | |
b699adcc NS |
803 | } |
804 | ||
dc3d2aeb | 805 | if (argno) |
b699adcc NS |
806 | s << ")"; |
807 | ||
808 | s << (is_defn ? "\n" : ";\n"); | |
809 | ||
810 | return name; | |
738f2522 BS |
811 | } |
812 | ||
00e52418 NS |
813 | /* Construct a function declaration from a call insn. This can be |
814 | necessary for two reasons - either we have an indirect call which | |
815 | requires a .callprototype declaration, or we have a libcall | |
816 | generated by emit_library_call for which no decl exists. */ | |
817 | ||
818 | static void | |
b699adcc NS |
819 | write_fn_proto_from_insn (std::stringstream &s, const char *name, |
820 | rtx result, rtx pat) | |
00e52418 NS |
821 | { |
822 | if (!name) | |
823 | { | |
824 | s << "\t.callprototype "; | |
825 | name = "_"; | |
826 | } | |
827 | else | |
828 | { | |
b699adcc | 829 | name = nvptx_name_replacement (name); |
69823d76 | 830 | write_fn_marker (s, false, true, name); |
00e52418 NS |
831 | s << "\t.extern .func "; |
832 | } | |
833 | ||
834 | if (result != NULL_RTX) | |
1f065954 | 835 | write_return_mode (s, true, GET_MODE (result)); |
00e52418 NS |
836 | |
837 | s << name; | |
838 | ||
00e52418 NS |
839 | int arg_end = XVECLEN (pat, 0); |
840 | for (int i = 1; i < arg_end; i++) | |
841 | { | |
1f065954 NS |
842 | /* We don't have to deal with mode splitting & promotion here, |
843 | as that was already done when generating the call | |
844 | sequence. */ | |
00e52418 NS |
845 | machine_mode mode = GET_MODE (XEXP (XVECEXP (pat, 0, i), 0)); |
846 | ||
1f065954 | 847 | write_arg_mode (s, -1, i - 1, mode); |
00e52418 NS |
848 | } |
849 | if (arg_end != 1) | |
850 | s << ")"; | |
851 | s << ";\n"; | |
852 | } | |
853 | ||
00e52418 NS |
854 | /* DECL is an external FUNCTION_DECL, make sure its in the fndecl hash |
855 | table and and write a ptx prototype. These are emitted at end of | |
856 | compilation. */ | |
738f2522 | 857 | |
00e52418 NS |
858 | static void |
859 | nvptx_record_fndecl (tree decl) | |
738f2522 | 860 | { |
f3dba894 | 861 | tree *slot = declared_fndecls_htab->find_slot (decl, INSERT); |
738f2522 BS |
862 | if (*slot == NULL) |
863 | { | |
864 | *slot = decl; | |
865 | const char *name = get_fnname_from_decl (decl); | |
b699adcc | 866 | write_fn_proto (func_decls, false, name, decl); |
738f2522 | 867 | } |
738f2522 BS |
868 | } |
869 | ||
00e52418 NS |
870 | /* Record a libcall or unprototyped external function. CALLEE is the |
871 | SYMBOL_REF. Insert into the libfunc hash table and emit a ptx | |
872 | declaration for it. */ | |
873 | ||
874 | static void | |
875 | nvptx_record_libfunc (rtx callee, rtx retval, rtx pat) | |
876 | { | |
877 | rtx *slot = declared_libfuncs_htab->find_slot (callee, INSERT); | |
878 | if (*slot == NULL) | |
879 | { | |
880 | *slot = callee; | |
881 | ||
882 | const char *name = XSTR (callee, 0); | |
b699adcc | 883 | write_fn_proto_from_insn (func_decls, name, retval, pat); |
00e52418 NS |
884 | } |
885 | } | |
886 | ||
887 | /* DECL is an external FUNCTION_DECL, that we're referencing. If it | |
888 | is prototyped, record it now. Otherwise record it as needed at end | |
889 | of compilation, when we might have more information about it. */ | |
738f2522 BS |
890 | |
891 | void | |
892 | nvptx_record_needed_fndecl (tree decl) | |
893 | { | |
00e52418 NS |
894 | if (TYPE_ARG_TYPES (TREE_TYPE (decl)) == NULL_TREE) |
895 | { | |
896 | tree *slot = needed_fndecls_htab->find_slot (decl, INSERT); | |
897 | if (*slot == NULL) | |
898 | *slot = decl; | |
899 | } | |
900 | else | |
901 | nvptx_record_fndecl (decl); | |
902 | } | |
738f2522 | 903 | |
00e52418 NS |
904 | /* SYM is a SYMBOL_REF. If it refers to an external function, record |
905 | it as needed. */ | |
906 | ||
907 | static void | |
908 | nvptx_maybe_record_fnsym (rtx sym) | |
909 | { | |
910 | tree decl = SYMBOL_REF_DECL (sym); | |
911 | ||
912 | if (decl && TREE_CODE (decl) == FUNCTION_DECL && DECL_EXTERNAL (decl)) | |
913 | nvptx_record_needed_fndecl (decl); | |
738f2522 BS |
914 | } |
915 | ||
1f065954 | 916 | /* Emit a local array to hold some part of a conventional stack frame |
5563d5c0 NS |
917 | and initialize REGNO to point to it. If the size is zero, it'll |
918 | never be valid to dereference, so we can simply initialize to | |
919 | zero. */ | |
1f065954 NS |
920 | |
921 | static void | |
922 | init_frame (FILE *file, int regno, unsigned align, unsigned size) | |
923 | { | |
5563d5c0 NS |
924 | if (size) |
925 | fprintf (file, "\t.local .align %d .b8 %s_ar[%u];\n", | |
926 | align, reg_names[regno], size); | |
927 | fprintf (file, "\t.reg.u%d %s;\n", | |
928 | POINTER_SIZE, reg_names[regno]); | |
929 | fprintf (file, (size ? "\tcvta.local.u%d %s, %s_ar;\n" | |
930 | : "\tmov.u%d %s, 0;\n"), | |
1f065954 NS |
931 | POINTER_SIZE, reg_names[regno], reg_names[regno]); |
932 | } | |
933 | ||
d88cd9c4 NS |
934 | /* Emit code to initialize the REGNO predicate register to indicate |
935 | whether we are not lane zero on the NAME axis. */ | |
936 | ||
937 | static void | |
938 | nvptx_init_axis_predicate (FILE *file, int regno, const char *name) | |
939 | { | |
940 | fprintf (file, "\t{\n"); | |
941 | fprintf (file, "\t\t.reg.u32\t%%%s;\n", name); | |
942 | fprintf (file, "\t\tmov.u32\t%%%s, %%tid.%s;\n", name, name); | |
943 | fprintf (file, "\t\tsetp.ne.u32\t%%r%d, %%%s, 0;\n", regno, name); | |
944 | fprintf (file, "\t}\n"); | |
945 | } | |
946 | ||
738f2522 BS |
947 | /* Implement ASM_DECLARE_FUNCTION_NAME. Writes the start of a ptx |
948 | function, including local var decls and copies from the arguments to | |
949 | local regs. */ | |
950 | ||
951 | void | |
952 | nvptx_declare_function_name (FILE *file, const char *name, const_tree decl) | |
953 | { | |
954 | tree fntype = TREE_TYPE (decl); | |
955 | tree result_type = TREE_TYPE (fntype); | |
dc3d2aeb | 956 | int argno = 0; |
738f2522 | 957 | |
dc3d2aeb NS |
958 | /* We construct the initial part of the function into a string |
959 | stream, in order to share the prototype writing code. */ | |
738f2522 | 960 | std::stringstream s; |
b699adcc | 961 | write_fn_proto (s, true, name, decl); |
dc3d2aeb | 962 | s << "{\n"; |
738f2522 | 963 | |
1f065954 | 964 | bool return_in_mem = write_return_type (s, false, result_type); |
738f2522 | 965 | if (return_in_mem) |
1f065954 | 966 | argno = write_arg_type (s, 0, argno, ptr_type_node, true); |
dc3d2aeb | 967 | |
5ab662d5 | 968 | /* Declare and initialize incoming arguments. */ |
dc3d2aeb NS |
969 | tree args = TYPE_ARG_TYPES (fntype); |
970 | bool prototyped = true; | |
971 | if (!args) | |
5ab662d5 | 972 | { |
dc3d2aeb NS |
973 | args = DECL_ARGUMENTS (decl); |
974 | prototyped = false; | |
5ab662d5 NS |
975 | } |
976 | ||
977 | for (; args != NULL_TREE; args = TREE_CHAIN (args)) | |
978 | { | |
979 | tree type = prototyped ? TREE_VALUE (args) : TREE_TYPE (args); | |
5ab662d5 | 980 | |
1f065954 | 981 | argno = write_arg_type (s, 0, argno, type, prototyped); |
dc3d2aeb | 982 | } |
5ab662d5 | 983 | |
dc3d2aeb | 984 | if (stdarg_p (fntype)) |
5563d5c0 | 985 | argno = write_arg_type (s, ARG_POINTER_REGNUM, argno, ptr_type_node, |
1f065954 | 986 | true); |
5ab662d5 | 987 | |
5563d5c0 NS |
988 | if (DECL_STATIC_CHAIN (decl) || cfun->machine->has_chain) |
989 | write_arg_type (s, STATIC_CHAIN_REGNUM, | |
990 | DECL_STATIC_CHAIN (decl) ? argno : -1, ptr_type_node, | |
991 | true); | |
992 | ||
dc3d2aeb | 993 | fprintf (file, "%s", s.str().c_str()); |
25662751 | 994 | |
44c068ae NS |
995 | /* Declare a local var for outgoing varargs. */ |
996 | if (cfun->machine->has_varadic) | |
997 | init_frame (file, STACK_POINTER_REGNUM, | |
998 | UNITS_PER_WORD, crtl->outgoing_args_size); | |
999 | ||
1000 | /* Declare a local variable for the frame. */ | |
1001 | HOST_WIDE_INT sz = get_frame_size (); | |
1002 | if (sz || cfun->machine->has_chain) | |
1003 | init_frame (file, FRAME_POINTER_REGNUM, | |
1004 | crtl->stack_alignment_needed / BITS_PER_UNIT, sz); | |
1005 | ||
738f2522 BS |
1006 | /* Declare the pseudos we have as ptx registers. */ |
1007 | int maxregs = max_reg_num (); | |
1008 | for (int i = LAST_VIRTUAL_REGISTER + 1; i < maxregs; i++) | |
1009 | { | |
1010 | if (regno_reg_rtx[i] != const0_rtx) | |
1011 | { | |
1012 | machine_mode mode = PSEUDO_REGNO_MODE (i); | |
d7479262 | 1013 | machine_mode split = maybe_split_mode (mode); |
f313d112 | 1014 | |
d7479262 | 1015 | if (split != VOIDmode) |
f313d112 NS |
1016 | mode = split; |
1017 | fprintf (file, "\t.reg%s ", nvptx_ptx_type_from_mode (mode, true)); | |
1018 | output_reg (file, i, split, -2); | |
1019 | fprintf (file, ";\n"); | |
738f2522 BS |
1020 | } |
1021 | } | |
1022 | ||
d88cd9c4 NS |
1023 | /* Emit axis predicates. */ |
1024 | if (cfun->machine->axis_predicate[0]) | |
1025 | nvptx_init_axis_predicate (file, | |
1026 | REGNO (cfun->machine->axis_predicate[0]), "y"); | |
1027 | if (cfun->machine->axis_predicate[1]) | |
1028 | nvptx_init_axis_predicate (file, | |
1029 | REGNO (cfun->machine->axis_predicate[1]), "x"); | |
738f2522 BS |
1030 | } |
1031 | ||
1032 | /* Output a return instruction. Also copy the return value to its outgoing | |
1033 | location. */ | |
1034 | ||
1035 | const char * | |
1036 | nvptx_output_return (void) | |
1037 | { | |
44c068ae | 1038 | machine_mode mode = (machine_mode)cfun->machine->return_mode; |
25662751 NS |
1039 | |
1040 | if (mode != VOIDmode) | |
1f065954 NS |
1041 | fprintf (asm_out_file, "\tst.param%s\t[%s_out], %s;\n", |
1042 | nvptx_ptx_type_from_mode (mode, false), | |
1043 | reg_names[NVPTX_RETURN_REGNUM], | |
1044 | reg_names[NVPTX_RETURN_REGNUM]); | |
738f2522 BS |
1045 | |
1046 | return "ret;"; | |
1047 | } | |
1048 | ||
738f2522 BS |
1049 | /* Terminate a function by writing a closing brace to FILE. */ |
1050 | ||
1051 | void | |
1052 | nvptx_function_end (FILE *file) | |
1053 | { | |
cf08c344 | 1054 | fprintf (file, "}\n"); |
738f2522 BS |
1055 | } |
1056 | \f | |
1057 | /* Decide whether we can make a sibling call to a function. For ptx, we | |
1058 | can't. */ | |
1059 | ||
1060 | static bool | |
1061 | nvptx_function_ok_for_sibcall (tree, tree) | |
1062 | { | |
1063 | return false; | |
1064 | } | |
1065 | ||
18c05628 NS |
1066 | /* Return Dynamic ReAlignment Pointer RTX. For PTX there isn't any. */ |
1067 | ||
1068 | static rtx | |
1069 | nvptx_get_drap_rtx (void) | |
1070 | { | |
1071 | return NULL_RTX; | |
1072 | } | |
1073 | ||
738f2522 BS |
1074 | /* Implement the TARGET_CALL_ARGS hook. Record information about one |
1075 | argument to the next call. */ | |
1076 | ||
1077 | static void | |
44c068ae | 1078 | nvptx_call_args (rtx arg, tree fntype) |
738f2522 | 1079 | { |
44c068ae | 1080 | if (!cfun->machine->doing_call) |
738f2522 | 1081 | { |
44c068ae NS |
1082 | cfun->machine->doing_call = true; |
1083 | cfun->machine->is_varadic = false; | |
1084 | cfun->machine->num_args = 0; | |
1085 | ||
1086 | if (fntype && stdarg_p (fntype)) | |
1087 | { | |
1088 | cfun->machine->is_varadic = true; | |
1089 | cfun->machine->has_varadic = true; | |
1090 | cfun->machine->num_args++; | |
1091 | } | |
738f2522 | 1092 | } |
738f2522 | 1093 | |
44c068ae NS |
1094 | if (REG_P (arg) && arg != pc_rtx) |
1095 | { | |
1096 | cfun->machine->num_args++; | |
1097 | cfun->machine->call_args = alloc_EXPR_LIST (VOIDmode, arg, | |
1098 | cfun->machine->call_args); | |
1099 | } | |
738f2522 BS |
1100 | } |
1101 | ||
1102 | /* Implement the corresponding END_CALL_ARGS hook. Clear and free the | |
1103 | information we recorded. */ | |
1104 | ||
1105 | static void | |
1106 | nvptx_end_call_args (void) | |
1107 | { | |
44c068ae | 1108 | cfun->machine->doing_call = false; |
738f2522 BS |
1109 | free_EXPR_LIST_list (&cfun->machine->call_args); |
1110 | } | |
1111 | ||
ecf6e535 BS |
1112 | /* Emit the sequence for a call to ADDRESS, setting RETVAL. Keep |
1113 | track of whether calls involving static chains or varargs were seen | |
1114 | in the current function. | |
1115 | For libcalls, maintain a hash table of decls we have seen, and | |
1116 | record a function decl for later when encountering a new one. */ | |
738f2522 BS |
1117 | |
1118 | void | |
1119 | nvptx_expand_call (rtx retval, rtx address) | |
1120 | { | |
738f2522 | 1121 | rtx callee = XEXP (address, 0); |
f324806d | 1122 | rtx varargs = NULL_RTX; |
d88cd9c4 | 1123 | unsigned parallel = 0; |
738f2522 | 1124 | |
738f2522 BS |
1125 | if (!call_insn_operand (callee, Pmode)) |
1126 | { | |
1127 | callee = force_reg (Pmode, callee); | |
1128 | address = change_address (address, QImode, callee); | |
1129 | } | |
1130 | ||
1131 | if (GET_CODE (callee) == SYMBOL_REF) | |
1132 | { | |
1133 | tree decl = SYMBOL_REF_DECL (callee); | |
1134 | if (decl != NULL_TREE) | |
1135 | { | |
738f2522 | 1136 | if (DECL_STATIC_CHAIN (decl)) |
44c068ae | 1137 | cfun->machine->has_chain = true; |
00e52418 | 1138 | |
d88cd9c4 NS |
1139 | tree attr = get_oacc_fn_attrib (decl); |
1140 | if (attr) | |
1141 | { | |
1142 | tree dims = TREE_VALUE (attr); | |
1143 | ||
1144 | parallel = GOMP_DIM_MASK (GOMP_DIM_MAX) - 1; | |
1145 | for (int ix = 0; ix != GOMP_DIM_MAX; ix++) | |
1146 | { | |
1147 | if (TREE_PURPOSE (dims) | |
1148 | && !integer_zerop (TREE_PURPOSE (dims))) | |
1149 | break; | |
1150 | /* Not on this axis. */ | |
1151 | parallel ^= GOMP_DIM_MASK (ix); | |
1152 | dims = TREE_CHAIN (dims); | |
1153 | } | |
1154 | } | |
738f2522 BS |
1155 | } |
1156 | } | |
c38f0d8c | 1157 | |
44c068ae NS |
1158 | unsigned nargs = cfun->machine->num_args; |
1159 | if (cfun->machine->is_varadic) | |
738f2522 | 1160 | { |
f324806d | 1161 | varargs = gen_reg_rtx (Pmode); |
863af9a4 | 1162 | emit_move_insn (varargs, stack_pointer_rtx); |
738f2522 BS |
1163 | } |
1164 | ||
44c068ae NS |
1165 | rtvec vec = rtvec_alloc (nargs + 1); |
1166 | rtx pat = gen_rtx_PARALLEL (VOIDmode, vec); | |
f324806d | 1167 | int vec_pos = 0; |
44c068ae NS |
1168 | |
1169 | rtx call = gen_rtx_CALL (VOIDmode, address, const0_rtx); | |
738f2522 | 1170 | rtx tmp_retval = retval; |
44c068ae | 1171 | if (retval) |
738f2522 BS |
1172 | { |
1173 | if (!nvptx_register_operand (retval, GET_MODE (retval))) | |
1174 | tmp_retval = gen_reg_rtx (GET_MODE (retval)); | |
44c068ae | 1175 | call = gen_rtx_SET (tmp_retval, call); |
738f2522 | 1176 | } |
44c068ae | 1177 | XVECEXP (pat, 0, vec_pos++) = call; |
f324806d NS |
1178 | |
1179 | /* Construct the call insn, including a USE for each argument pseudo | |
1180 | register. These will be used when printing the insn. */ | |
1181 | for (rtx arg = cfun->machine->call_args; arg; arg = XEXP (arg, 1)) | |
44c068ae | 1182 | XVECEXP (pat, 0, vec_pos++) = gen_rtx_USE (VOIDmode, XEXP (arg, 0)); |
f324806d NS |
1183 | |
1184 | if (varargs) | |
cf08c344 | 1185 | XVECEXP (pat, 0, vec_pos++) = gen_rtx_USE (VOIDmode, varargs); |
f324806d NS |
1186 | |
1187 | gcc_assert (vec_pos = XVECLEN (pat, 0)); | |
ecf6e535 | 1188 | |
d88cd9c4 | 1189 | nvptx_emit_forking (parallel, true); |
738f2522 | 1190 | emit_call_insn (pat); |
d88cd9c4 NS |
1191 | nvptx_emit_joining (parallel, true); |
1192 | ||
738f2522 BS |
1193 | if (tmp_retval != retval) |
1194 | emit_move_insn (retval, tmp_retval); | |
1195 | } | |
44eba92d | 1196 | |
738f2522 BS |
1197 | /* Emit a comparison COMPARE, and return the new test to be used in the |
1198 | jump. */ | |
1199 | ||
1200 | rtx | |
1201 | nvptx_expand_compare (rtx compare) | |
1202 | { | |
1203 | rtx pred = gen_reg_rtx (BImode); | |
1204 | rtx cmp = gen_rtx_fmt_ee (GET_CODE (compare), BImode, | |
1205 | XEXP (compare, 0), XEXP (compare, 1)); | |
f7df4a84 | 1206 | emit_insn (gen_rtx_SET (pred, cmp)); |
738f2522 BS |
1207 | return gen_rtx_NE (BImode, pred, const0_rtx); |
1208 | } | |
1209 | ||
d88cd9c4 NS |
1210 | /* Expand the oacc fork & join primitive into ptx-required unspecs. */ |
1211 | ||
1212 | void | |
1213 | nvptx_expand_oacc_fork (unsigned mode) | |
1214 | { | |
1215 | nvptx_emit_forking (GOMP_DIM_MASK (mode), false); | |
1216 | } | |
1217 | ||
1218 | void | |
1219 | nvptx_expand_oacc_join (unsigned mode) | |
1220 | { | |
1221 | nvptx_emit_joining (GOMP_DIM_MASK (mode), false); | |
1222 | } | |
1223 | ||
1224 | /* Generate instruction(s) to unpack a 64 bit object into 2 32 bit | |
1225 | objects. */ | |
1226 | ||
1227 | static rtx | |
1228 | nvptx_gen_unpack (rtx dst0, rtx dst1, rtx src) | |
1229 | { | |
1230 | rtx res; | |
1231 | ||
1232 | switch (GET_MODE (src)) | |
1233 | { | |
1234 | case DImode: | |
1235 | res = gen_unpackdisi2 (dst0, dst1, src); | |
1236 | break; | |
1237 | case DFmode: | |
1238 | res = gen_unpackdfsi2 (dst0, dst1, src); | |
1239 | break; | |
1240 | default: gcc_unreachable (); | |
1241 | } | |
1242 | return res; | |
1243 | } | |
1244 | ||
1245 | /* Generate instruction(s) to pack 2 32 bit objects into a 64 bit | |
1246 | object. */ | |
1247 | ||
1248 | static rtx | |
1249 | nvptx_gen_pack (rtx dst, rtx src0, rtx src1) | |
1250 | { | |
1251 | rtx res; | |
1252 | ||
1253 | switch (GET_MODE (dst)) | |
1254 | { | |
1255 | case DImode: | |
1256 | res = gen_packsidi2 (dst, src0, src1); | |
1257 | break; | |
1258 | case DFmode: | |
1259 | res = gen_packsidf2 (dst, src0, src1); | |
1260 | break; | |
1261 | default: gcc_unreachable (); | |
1262 | } | |
1263 | return res; | |
1264 | } | |
1265 | ||
1266 | /* Generate an instruction or sequence to broadcast register REG | |
1267 | across the vectors of a single warp. */ | |
1268 | ||
1269 | static rtx | |
59263259 | 1270 | nvptx_gen_shuffle (rtx dst, rtx src, rtx idx, nvptx_shuffle_kind kind) |
d88cd9c4 NS |
1271 | { |
1272 | rtx res; | |
1273 | ||
1274 | switch (GET_MODE (dst)) | |
1275 | { | |
1276 | case SImode: | |
1277 | res = gen_nvptx_shufflesi (dst, src, idx, GEN_INT (kind)); | |
1278 | break; | |
1279 | case SFmode: | |
1280 | res = gen_nvptx_shufflesf (dst, src, idx, GEN_INT (kind)); | |
1281 | break; | |
1282 | case DImode: | |
1283 | case DFmode: | |
1284 | { | |
1285 | rtx tmp0 = gen_reg_rtx (SImode); | |
1286 | rtx tmp1 = gen_reg_rtx (SImode); | |
1287 | ||
1288 | start_sequence (); | |
1289 | emit_insn (nvptx_gen_unpack (tmp0, tmp1, src)); | |
1290 | emit_insn (nvptx_gen_shuffle (tmp0, tmp0, idx, kind)); | |
1291 | emit_insn (nvptx_gen_shuffle (tmp1, tmp1, idx, kind)); | |
1292 | emit_insn (nvptx_gen_pack (dst, tmp0, tmp1)); | |
1293 | res = get_insns (); | |
1294 | end_sequence (); | |
1295 | } | |
1296 | break; | |
1297 | case BImode: | |
1298 | { | |
1299 | rtx tmp = gen_reg_rtx (SImode); | |
1300 | ||
1301 | start_sequence (); | |
1302 | emit_insn (gen_sel_truesi (tmp, src, GEN_INT (1), const0_rtx)); | |
1303 | emit_insn (nvptx_gen_shuffle (tmp, tmp, idx, kind)); | |
1304 | emit_insn (gen_rtx_SET (dst, gen_rtx_NE (BImode, tmp, const0_rtx))); | |
1305 | res = get_insns (); | |
1306 | end_sequence (); | |
1307 | } | |
1308 | break; | |
1309 | ||
1310 | default: | |
1311 | gcc_unreachable (); | |
1312 | } | |
1313 | return res; | |
1314 | } | |
1315 | ||
1316 | /* Generate an instruction or sequence to broadcast register REG | |
1317 | across the vectors of a single warp. */ | |
1318 | ||
1319 | static rtx | |
1320 | nvptx_gen_vcast (rtx reg) | |
1321 | { | |
1322 | return nvptx_gen_shuffle (reg, reg, const0_rtx, SHUFFLE_IDX); | |
1323 | } | |
1324 | ||
1325 | /* Structure used when generating a worker-level spill or fill. */ | |
1326 | ||
1327 | struct wcast_data_t | |
1328 | { | |
1329 | rtx base; /* Register holding base addr of buffer. */ | |
1330 | rtx ptr; /* Iteration var, if needed. */ | |
1331 | unsigned offset; /* Offset into worker buffer. */ | |
1332 | }; | |
1333 | ||
1334 | /* Direction of the spill/fill and looping setup/teardown indicator. */ | |
1335 | ||
1336 | enum propagate_mask | |
1337 | { | |
1338 | PM_read = 1 << 0, | |
1339 | PM_write = 1 << 1, | |
1340 | PM_loop_begin = 1 << 2, | |
1341 | PM_loop_end = 1 << 3, | |
1342 | ||
1343 | PM_read_write = PM_read | PM_write | |
1344 | }; | |
1345 | ||
1346 | /* Generate instruction(s) to spill or fill register REG to/from the | |
1347 | worker broadcast array. PM indicates what is to be done, REP | |
1348 | how many loop iterations will be executed (0 for not a loop). */ | |
1349 | ||
1350 | static rtx | |
1351 | nvptx_gen_wcast (rtx reg, propagate_mask pm, unsigned rep, wcast_data_t *data) | |
1352 | { | |
1353 | rtx res; | |
1354 | machine_mode mode = GET_MODE (reg); | |
1355 | ||
1356 | switch (mode) | |
1357 | { | |
1358 | case BImode: | |
1359 | { | |
1360 | rtx tmp = gen_reg_rtx (SImode); | |
1361 | ||
1362 | start_sequence (); | |
1363 | if (pm & PM_read) | |
1364 | emit_insn (gen_sel_truesi (tmp, reg, GEN_INT (1), const0_rtx)); | |
1365 | emit_insn (nvptx_gen_wcast (tmp, pm, rep, data)); | |
1366 | if (pm & PM_write) | |
1367 | emit_insn (gen_rtx_SET (reg, gen_rtx_NE (BImode, tmp, const0_rtx))); | |
1368 | res = get_insns (); | |
1369 | end_sequence (); | |
1370 | } | |
1371 | break; | |
1372 | ||
1373 | default: | |
1374 | { | |
1375 | rtx addr = data->ptr; | |
1376 | ||
1377 | if (!addr) | |
1378 | { | |
1379 | unsigned align = GET_MODE_ALIGNMENT (mode) / BITS_PER_UNIT; | |
1380 | ||
1381 | if (align > worker_bcast_align) | |
1382 | worker_bcast_align = align; | |
1383 | data->offset = (data->offset + align - 1) & ~(align - 1); | |
1384 | addr = data->base; | |
1385 | if (data->offset) | |
1386 | addr = gen_rtx_PLUS (Pmode, addr, GEN_INT (data->offset)); | |
1387 | } | |
1388 | ||
1389 | addr = gen_rtx_MEM (mode, addr); | |
d88cd9c4 NS |
1390 | if (pm == PM_read) |
1391 | res = gen_rtx_SET (addr, reg); | |
1392 | else if (pm == PM_write) | |
1393 | res = gen_rtx_SET (reg, addr); | |
1394 | else | |
1395 | gcc_unreachable (); | |
1396 | ||
1397 | if (data->ptr) | |
1398 | { | |
1399 | /* We're using a ptr, increment it. */ | |
1400 | start_sequence (); | |
1401 | ||
1402 | emit_insn (res); | |
1403 | emit_insn (gen_adddi3 (data->ptr, data->ptr, | |
1404 | GEN_INT (GET_MODE_SIZE (GET_MODE (reg))))); | |
1405 | res = get_insns (); | |
1406 | end_sequence (); | |
1407 | } | |
1408 | else | |
1409 | rep = 1; | |
1410 | data->offset += rep * GET_MODE_SIZE (GET_MODE (reg)); | |
1411 | } | |
1412 | break; | |
1413 | } | |
1414 | return res; | |
1415 | } | |
738f2522 BS |
1416 | \f |
1417 | /* Returns true if X is a valid address for use in a memory reference. */ | |
1418 | ||
1419 | static bool | |
1420 | nvptx_legitimate_address_p (machine_mode, rtx x, bool) | |
1421 | { | |
1422 | enum rtx_code code = GET_CODE (x); | |
1423 | ||
1424 | switch (code) | |
1425 | { | |
1426 | case REG: | |
1427 | return true; | |
1428 | ||
1429 | case PLUS: | |
1430 | if (REG_P (XEXP (x, 0)) && CONST_INT_P (XEXP (x, 1))) | |
1431 | return true; | |
1432 | return false; | |
1433 | ||
1434 | case CONST: | |
1435 | case SYMBOL_REF: | |
1436 | case LABEL_REF: | |
1437 | return true; | |
1438 | ||
1439 | default: | |
1440 | return false; | |
1441 | } | |
1442 | } | |
738f2522 | 1443 | \f |
4d5438cd NS |
1444 | /* Machinery to output constant initializers. When beginning an |
1445 | initializer, we decide on a fragment size (which is visible in ptx | |
1446 | in the type used), and then all initializer data is buffered until | |
1447 | a fragment is filled and ready to be written out. */ | |
1448 | ||
1449 | static struct | |
1450 | { | |
1451 | unsigned HOST_WIDE_INT mask; /* Mask for storing fragment. */ | |
1452 | unsigned HOST_WIDE_INT val; /* Current fragment value. */ | |
1453 | unsigned HOST_WIDE_INT remaining; /* Remaining bytes to be written | |
1454 | out. */ | |
1455 | unsigned size; /* Fragment size to accumulate. */ | |
1456 | unsigned offset; /* Offset within current fragment. */ | |
1457 | bool started; /* Whether we've output any initializer. */ | |
1458 | } init_frag; | |
1459 | ||
1460 | /* The current fragment is full, write it out. SYM may provide a | |
1461 | symbolic reference we should output, in which case the fragment | |
1462 | value is the addend. */ | |
738f2522 BS |
1463 | |
1464 | static void | |
4d5438cd | 1465 | output_init_frag (rtx sym) |
738f2522 | 1466 | { |
4d5438cd NS |
1467 | fprintf (asm_out_file, init_frag.started ? ", " : " = { "); |
1468 | unsigned HOST_WIDE_INT val = init_frag.val; | |
738f2522 | 1469 | |
4d5438cd NS |
1470 | init_frag.started = true; |
1471 | init_frag.val = 0; | |
1472 | init_frag.offset = 0; | |
1473 | init_frag.remaining--; | |
1474 | ||
1475 | if (sym) | |
1476 | { | |
1477 | fprintf (asm_out_file, "generic("); | |
1478 | output_address (VOIDmode, sym); | |
1479 | fprintf (asm_out_file, val ? ") + " : ")"); | |
1480 | } | |
738f2522 | 1481 | |
4d5438cd NS |
1482 | if (!sym || val) |
1483 | fprintf (asm_out_file, HOST_WIDE_INT_PRINT_DEC, val); | |
738f2522 BS |
1484 | } |
1485 | ||
4d5438cd NS |
1486 | /* Add value VAL of size SIZE to the data we're emitting, and keep |
1487 | writing out chunks as they fill up. */ | |
738f2522 BS |
1488 | |
1489 | static void | |
4d5438cd | 1490 | nvptx_assemble_value (unsigned HOST_WIDE_INT val, unsigned size) |
738f2522 | 1491 | { |
4d5438cd NS |
1492 | val &= ((unsigned HOST_WIDE_INT)2 << (size * BITS_PER_UNIT - 1)) - 1; |
1493 | ||
1494 | for (unsigned part = 0; size; size -= part) | |
738f2522 | 1495 | { |
4d5438cd NS |
1496 | val >>= part * BITS_PER_UNIT; |
1497 | part = init_frag.size - init_frag.offset; | |
1498 | if (part > size) | |
1499 | part = size; | |
1500 | ||
1501 | unsigned HOST_WIDE_INT partial | |
1502 | = val << (init_frag.offset * BITS_PER_UNIT); | |
1503 | init_frag.val |= partial & init_frag.mask; | |
1504 | init_frag.offset += part; | |
1505 | ||
1506 | if (init_frag.offset == init_frag.size) | |
1507 | output_init_frag (NULL); | |
738f2522 BS |
1508 | } |
1509 | } | |
1510 | ||
1511 | /* Target hook for assembling integer object X of size SIZE. */ | |
1512 | ||
1513 | static bool | |
1514 | nvptx_assemble_integer (rtx x, unsigned int size, int ARG_UNUSED (aligned_p)) | |
1515 | { | |
00e52418 NS |
1516 | HOST_WIDE_INT val = 0; |
1517 | ||
1518 | switch (GET_CODE (x)) | |
738f2522 | 1519 | { |
00e52418 | 1520 | default: |
a9000e1e NS |
1521 | /* Let the generic machinery figure it out, usually for a |
1522 | CONST_WIDE_INT. */ | |
1523 | return false; | |
00e52418 NS |
1524 | |
1525 | case CONST_INT: | |
4d5438cd | 1526 | nvptx_assemble_value (INTVAL (x), size); |
00e52418 NS |
1527 | break; |
1528 | ||
1529 | case CONST: | |
1530 | x = XEXP (x, 0); | |
1531 | gcc_assert (GET_CODE (x) == PLUS); | |
1532 | val = INTVAL (XEXP (x, 1)); | |
1533 | x = XEXP (x, 0); | |
1534 | gcc_assert (GET_CODE (x) == SYMBOL_REF); | |
1535 | /* FALLTHROUGH */ | |
1536 | ||
1537 | case SYMBOL_REF: | |
4d5438cd NS |
1538 | gcc_assert (size == init_frag.size); |
1539 | if (init_frag.offset) | |
738f2522 | 1540 | sorry ("cannot emit unaligned pointers in ptx assembly"); |
738f2522 | 1541 | |
00e52418 | 1542 | nvptx_maybe_record_fnsym (x); |
4d5438cd NS |
1543 | init_frag.val = val; |
1544 | output_init_frag (x); | |
738f2522 | 1545 | break; |
738f2522 BS |
1546 | } |
1547 | ||
738f2522 BS |
1548 | return true; |
1549 | } | |
1550 | ||
1551 | /* Output SIZE zero bytes. We ignore the FILE argument since the | |
1552 | functions we're calling to perform the output just use | |
1553 | asm_out_file. */ | |
1554 | ||
1555 | void | |
1556 | nvptx_output_skip (FILE *, unsigned HOST_WIDE_INT size) | |
1557 | { | |
4d5438cd NS |
1558 | /* Finish the current fragment, if it's started. */ |
1559 | if (init_frag.offset) | |
738f2522 | 1560 | { |
4d5438cd NS |
1561 | unsigned part = init_frag.size - init_frag.offset; |
1562 | if (part > size) | |
1563 | part = (unsigned) size; | |
1564 | size -= part; | |
1565 | nvptx_assemble_value (0, part); | |
738f2522 BS |
1566 | } |
1567 | ||
4d5438cd NS |
1568 | /* If this skip doesn't terminate the initializer, write as many |
1569 | remaining pieces as possible directly. */ | |
1570 | if (size < init_frag.remaining * init_frag.size) | |
738f2522 | 1571 | { |
4d5438cd NS |
1572 | while (size >= init_frag.size) |
1573 | { | |
1574 | size -= init_frag.size; | |
1575 | output_init_frag (NULL_RTX); | |
1576 | } | |
1577 | if (size) | |
1578 | nvptx_assemble_value (0, size); | |
738f2522 | 1579 | } |
738f2522 BS |
1580 | } |
1581 | ||
1582 | /* Output a string STR with length SIZE. As in nvptx_output_skip we | |
1583 | ignore the FILE arg. */ | |
1584 | ||
1585 | void | |
1586 | nvptx_output_ascii (FILE *, const char *str, unsigned HOST_WIDE_INT size) | |
1587 | { | |
1588 | for (unsigned HOST_WIDE_INT i = 0; i < size; i++) | |
1589 | nvptx_assemble_value (str[i], 1); | |
1590 | } | |
1591 | ||
4ff3145a NS |
1592 | /* Emit a PTX variable decl and prepare for emission of its |
1593 | initializer. NAME is the symbol name and SETION the PTX data | |
1594 | area. The type is TYPE, object size SIZE and alignment is ALIGN. | |
1595 | The caller has already emitted any indentation and linkage | |
1596 | specifier. It is responsible for any initializer, terminating ; | |
1597 | and newline. SIZE is in bytes, ALIGN is in bits -- confusingly | |
1598 | this is the opposite way round that PTX wants them! */ | |
1599 | ||
1600 | static void | |
1601 | nvptx_assemble_decl_begin (FILE *file, const char *name, const char *section, | |
1602 | const_tree type, HOST_WIDE_INT size, unsigned align) | |
1603 | { | |
1604 | while (TREE_CODE (type) == ARRAY_TYPE) | |
1605 | type = TREE_TYPE (type); | |
1606 | ||
fc0efeea NS |
1607 | if (TREE_CODE (type) == VECTOR_TYPE |
1608 | || TREE_CODE (type) == COMPLEX_TYPE) | |
1609 | /* Neither vector nor complex types can contain the other. */ | |
1610 | type = TREE_TYPE (type); | |
1611 | ||
4ff3145a | 1612 | unsigned elt_size = int_size_in_bytes (type); |
fc0efeea NS |
1613 | |
1614 | /* Largest mode we're prepared to accept. For BLKmode types we | |
1615 | don't know if it'll contain pointer constants, so have to choose | |
1616 | pointer size, otherwise we can choose DImode. */ | |
1617 | machine_mode elt_mode = TYPE_MODE (type) == BLKmode ? Pmode : DImode; | |
1618 | ||
1619 | elt_size |= GET_MODE_SIZE (elt_mode); | |
1620 | elt_size &= -elt_size; /* Extract LSB set. */ | |
4ff3145a | 1621 | |
4d5438cd NS |
1622 | init_frag.size = elt_size; |
1623 | /* Avoid undefined shift behaviour by using '2'. */ | |
1624 | init_frag.mask = ((unsigned HOST_WIDE_INT)2 | |
1625 | << (elt_size * BITS_PER_UNIT - 1)) - 1; | |
1626 | init_frag.val = 0; | |
1627 | init_frag.offset = 0; | |
1628 | init_frag.started = false; | |
1629 | /* Size might not be a multiple of elt size, if there's an | |
1630 | initialized trailing struct array with smaller type than | |
1631 | elt_size. */ | |
1632 | init_frag.remaining = (size + elt_size - 1) / elt_size; | |
4ff3145a NS |
1633 | |
1634 | fprintf (file, "%s .align %d .u%d ", | |
1635 | section, align / BITS_PER_UNIT, | |
1636 | elt_size * BITS_PER_UNIT); | |
1637 | assemble_name (file, name); | |
1638 | ||
1639 | if (size) | |
1640 | /* We make everything an array, to simplify any initialization | |
1641 | emission. */ | |
4d5438cd | 1642 | fprintf (file, "[" HOST_WIDE_INT_PRINT_DEC "]", init_frag.remaining); |
4ff3145a NS |
1643 | } |
1644 | ||
738f2522 BS |
1645 | /* Called when the initializer for a decl has been completely output through |
1646 | combinations of the three functions above. */ | |
1647 | ||
1648 | static void | |
1649 | nvptx_assemble_decl_end (void) | |
1650 | { | |
4d5438cd NS |
1651 | if (init_frag.offset) |
1652 | /* This can happen with a packed struct with trailing array member. */ | |
1653 | nvptx_assemble_value (0, init_frag.size - init_frag.offset); | |
1654 | fprintf (asm_out_file, init_frag.started ? " };\n" : ";\n"); | |
738f2522 BS |
1655 | } |
1656 | ||
69823d76 NS |
1657 | /* Output an uninitialized common or file-scope variable. */ |
1658 | ||
1659 | void | |
1660 | nvptx_output_aligned_decl (FILE *file, const char *name, | |
1661 | const_tree decl, HOST_WIDE_INT size, unsigned align) | |
1662 | { | |
1663 | write_var_marker (file, true, TREE_PUBLIC (decl), name); | |
1664 | ||
1665 | /* If this is public, it is common. The nearest thing we have to | |
1666 | common is weak. */ | |
4ff3145a NS |
1667 | fprintf (file, "\t%s", TREE_PUBLIC (decl) ? ".weak " : ""); |
1668 | ||
1669 | nvptx_assemble_decl_begin (file, name, section_for_decl (decl), | |
1670 | TREE_TYPE (decl), size, align); | |
4d5438cd | 1671 | nvptx_assemble_decl_end (); |
69823d76 NS |
1672 | } |
1673 | ||
738f2522 BS |
1674 | /* Implement TARGET_ASM_DECLARE_CONSTANT_NAME. Begin the process of |
1675 | writing a constant variable EXP with NAME and SIZE and its | |
1676 | initializer to FILE. */ | |
1677 | ||
1678 | static void | |
1679 | nvptx_asm_declare_constant_name (FILE *file, const char *name, | |
4ff3145a | 1680 | const_tree exp, HOST_WIDE_INT obj_size) |
738f2522 | 1681 | { |
4ff3145a NS |
1682 | write_var_marker (file, true, false, name); |
1683 | ||
1684 | fprintf (file, "\t"); | |
1685 | ||
738f2522 | 1686 | tree type = TREE_TYPE (exp); |
4ff3145a NS |
1687 | nvptx_assemble_decl_begin (file, name, ".const", type, obj_size, |
1688 | TYPE_ALIGN (type)); | |
738f2522 BS |
1689 | } |
1690 | ||
1691 | /* Implement the ASM_DECLARE_OBJECT_NAME macro. Used to start writing | |
1692 | a variable DECL with NAME to FILE. */ | |
1693 | ||
1694 | void | |
1695 | nvptx_declare_object_name (FILE *file, const char *name, const_tree decl) | |
1696 | { | |
4ff3145a | 1697 | write_var_marker (file, true, TREE_PUBLIC (decl), name); |
9a863523 | 1698 | |
4ff3145a NS |
1699 | fprintf (file, "\t%s", (!TREE_PUBLIC (decl) ? "" |
1700 | : DECL_WEAK (decl) ? ".weak " : ".visible ")); | |
9a863523 | 1701 | |
4ff3145a NS |
1702 | tree type = TREE_TYPE (decl); |
1703 | HOST_WIDE_INT obj_size = tree_to_shwi (DECL_SIZE_UNIT (decl)); | |
1704 | nvptx_assemble_decl_begin (file, name, section_for_decl (decl), | |
1705 | type, obj_size, DECL_ALIGN (decl)); | |
738f2522 BS |
1706 | } |
1707 | ||
1708 | /* Implement TARGET_ASM_GLOBALIZE_LABEL by doing nothing. */ | |
1709 | ||
1710 | static void | |
1711 | nvptx_globalize_label (FILE *, const char *) | |
1712 | { | |
1713 | } | |
1714 | ||
1715 | /* Implement TARGET_ASM_ASSEMBLE_UNDEFINED_DECL. Write an extern | |
1716 | declaration only for variable DECL with NAME to FILE. */ | |
f313d112 | 1717 | |
738f2522 BS |
1718 | static void |
1719 | nvptx_assemble_undefined_decl (FILE *file, const char *name, const_tree decl) | |
1720 | { | |
0a0f74aa NS |
1721 | /* The middle end can place constant pool decls into the varpool as |
1722 | undefined. Until that is fixed, catch the problem here. */ | |
1723 | if (DECL_IN_CONSTANT_POOL (decl)) | |
1724 | return; | |
1725 | ||
69823d76 NS |
1726 | write_var_marker (file, false, TREE_PUBLIC (decl), name); |
1727 | ||
4ff3145a NS |
1728 | fprintf (file, "\t.extern "); |
1729 | tree size = DECL_SIZE_UNIT (decl); | |
1730 | nvptx_assemble_decl_begin (file, name, section_for_decl (decl), | |
1731 | TREE_TYPE (decl), size ? tree_to_shwi (size) : 0, | |
1732 | DECL_ALIGN (decl)); | |
1e5154e7 | 1733 | nvptx_assemble_decl_end (); |
738f2522 BS |
1734 | } |
1735 | ||
f313d112 NS |
1736 | /* Output a pattern for a move instruction. */ |
1737 | ||
1738 | const char * | |
1739 | nvptx_output_mov_insn (rtx dst, rtx src) | |
1740 | { | |
1741 | machine_mode dst_mode = GET_MODE (dst); | |
1742 | machine_mode dst_inner = (GET_CODE (dst) == SUBREG | |
1743 | ? GET_MODE (XEXP (dst, 0)) : dst_mode); | |
1744 | machine_mode src_inner = (GET_CODE (src) == SUBREG | |
1745 | ? GET_MODE (XEXP (src, 0)) : dst_mode); | |
1746 | ||
15113b03 NS |
1747 | rtx sym = src; |
1748 | if (GET_CODE (sym) == CONST) | |
1749 | sym = XEXP (XEXP (sym, 0), 0); | |
bd602b7f NS |
1750 | if (SYMBOL_REF_P (sym)) |
1751 | { | |
1752 | if (SYMBOL_DATA_AREA (sym) != DATA_AREA_GENERIC) | |
1753 | return "%.\tcvta%D1%t0\t%0, %1;"; | |
1754 | nvptx_maybe_record_fnsym (sym); | |
1755 | } | |
15113b03 | 1756 | |
f313d112 NS |
1757 | if (src_inner == dst_inner) |
1758 | return "%.\tmov%t0\t%0, %1;"; | |
1759 | ||
1760 | if (CONSTANT_P (src)) | |
1761 | return (GET_MODE_CLASS (dst_inner) == MODE_INT | |
1762 | && GET_MODE_CLASS (src_inner) != MODE_FLOAT | |
1763 | ? "%.\tmov%t0\t%0, %1;" : "%.\tmov.b%T0\t%0, %1;"); | |
1764 | ||
1765 | if (GET_MODE_SIZE (dst_inner) == GET_MODE_SIZE (src_inner)) | |
1766 | return "%.\tmov.b%T0\t%0, %1;"; | |
1767 | ||
1768 | return "%.\tcvt%t0%t1\t%0, %1;"; | |
1769 | } | |
1770 | ||
738f2522 | 1771 | /* Output INSN, which is a call to CALLEE with result RESULT. For ptx, this |
ecf6e535 BS |
1772 | involves writing .param declarations and in/out copies into them. For |
1773 | indirect calls, also write the .callprototype. */ | |
738f2522 BS |
1774 | |
1775 | const char * | |
1776 | nvptx_output_call_insn (rtx_insn *insn, rtx result, rtx callee) | |
1777 | { | |
863af9a4 | 1778 | char buf[16]; |
738f2522 BS |
1779 | static int labelno; |
1780 | bool needs_tgt = register_operand (callee, Pmode); | |
1781 | rtx pat = PATTERN (insn); | |
f324806d | 1782 | int arg_end = XVECLEN (pat, 0); |
738f2522 BS |
1783 | tree decl = NULL_TREE; |
1784 | ||
1785 | fprintf (asm_out_file, "\t{\n"); | |
1786 | if (result != NULL) | |
1f065954 NS |
1787 | fprintf (asm_out_file, "\t\t.param%s %s_in;\n", |
1788 | nvptx_ptx_type_from_mode (GET_MODE (result), false), | |
1789 | reg_names[NVPTX_RETURN_REGNUM]); | |
738f2522 | 1790 | |
ecf6e535 | 1791 | /* Ensure we have a ptx declaration in the output if necessary. */ |
738f2522 BS |
1792 | if (GET_CODE (callee) == SYMBOL_REF) |
1793 | { | |
1794 | decl = SYMBOL_REF_DECL (callee); | |
00e52418 NS |
1795 | if (!decl |
1796 | || (DECL_EXTERNAL (decl) && !TYPE_ARG_TYPES (TREE_TYPE (decl)))) | |
1797 | nvptx_record_libfunc (callee, result, pat); | |
1798 | else if (DECL_EXTERNAL (decl)) | |
738f2522 BS |
1799 | nvptx_record_fndecl (decl); |
1800 | } | |
1801 | ||
1802 | if (needs_tgt) | |
1803 | { | |
1804 | ASM_GENERATE_INTERNAL_LABEL (buf, "LCT", labelno); | |
1805 | labelno++; | |
1806 | ASM_OUTPUT_LABEL (asm_out_file, buf); | |
1807 | std::stringstream s; | |
b699adcc | 1808 | write_fn_proto_from_insn (s, NULL, result, pat); |
738f2522 BS |
1809 | fputs (s.str().c_str(), asm_out_file); |
1810 | } | |
1811 | ||
863af9a4 | 1812 | for (int argno = 1; argno < arg_end; argno++) |
738f2522 | 1813 | { |
863af9a4 | 1814 | rtx t = XEXP (XVECEXP (pat, 0, argno), 0); |
738f2522 | 1815 | machine_mode mode = GET_MODE (t); |
a02d84b6 | 1816 | const char *ptx_type = nvptx_ptx_type_from_mode (mode, false); |
738f2522 | 1817 | |
863af9a4 | 1818 | /* Mode splitting has already been done. */ |
a02d84b6 NS |
1819 | fprintf (asm_out_file, "\t\t.param%s %%out_arg%d;\n" |
1820 | "\t\tst.param%s [%%out_arg%d], ", | |
1821 | ptx_type, argno, ptx_type, argno); | |
1822 | output_reg (asm_out_file, REGNO (t), VOIDmode); | |
1823 | fprintf (asm_out_file, ";\n"); | |
738f2522 BS |
1824 | } |
1825 | ||
1826 | fprintf (asm_out_file, "\t\tcall "); | |
1827 | if (result != NULL_RTX) | |
1f065954 NS |
1828 | fprintf (asm_out_file, "(%s_in), ", reg_names[NVPTX_RETURN_REGNUM]); |
1829 | ||
738f2522 BS |
1830 | if (decl) |
1831 | { | |
1832 | const char *name = get_fnname_from_decl (decl); | |
1833 | name = nvptx_name_replacement (name); | |
1834 | assemble_name (asm_out_file, name); | |
1835 | } | |
1836 | else | |
cc8ca59e | 1837 | output_address (VOIDmode, callee); |
738f2522 | 1838 | |
863af9a4 NS |
1839 | const char *open = "("; |
1840 | for (int argno = 1; argno < arg_end; argno++) | |
738f2522 | 1841 | { |
863af9a4 NS |
1842 | fprintf (asm_out_file, ", %s%%out_arg%d", open, argno); |
1843 | open = ""; | |
738f2522 | 1844 | } |
863af9a4 NS |
1845 | if (decl && DECL_STATIC_CHAIN (decl)) |
1846 | { | |
5563d5c0 | 1847 | fprintf (asm_out_file, ", %s%s", open, reg_names [STATIC_CHAIN_REGNUM]); |
863af9a4 NS |
1848 | open = ""; |
1849 | } | |
1850 | if (!open[0]) | |
1851 | fprintf (asm_out_file, ")"); | |
f324806d | 1852 | |
738f2522 BS |
1853 | if (needs_tgt) |
1854 | { | |
1855 | fprintf (asm_out_file, ", "); | |
1856 | assemble_name (asm_out_file, buf); | |
1857 | } | |
1858 | fprintf (asm_out_file, ";\n"); | |
738f2522 | 1859 | |
51baf85a NS |
1860 | if (find_reg_note (insn, REG_NORETURN, NULL)) |
1861 | /* No return functions confuse the PTX JIT, as it doesn't realize | |
1862 | the flow control barrier they imply. It can seg fault if it | |
1863 | encounters what looks like an unexitable loop. Emit a trailing | |
1864 | trap, which it does grok. */ | |
1865 | fprintf (asm_out_file, "\t\ttrap; // (noreturn)\n"); | |
1866 | ||
1f065954 NS |
1867 | if (result) |
1868 | { | |
1869 | static char rval[sizeof ("\tld.param%%t0\t%%0, [%%%s_in];\n\t}") + 8]; | |
1870 | ||
1871 | if (!rval[0]) | |
1872 | /* We must escape the '%' that starts RETURN_REGNUM. */ | |
1873 | sprintf (rval, "\tld.param%%t0\t%%0, [%%%s_in];\n\t}", | |
1874 | reg_names[NVPTX_RETURN_REGNUM]); | |
1875 | return rval; | |
1876 | } | |
1877 | ||
1878 | return "}"; | |
738f2522 BS |
1879 | } |
1880 | ||
1881 | /* Implement TARGET_PRINT_OPERAND_PUNCT_VALID_P. */ | |
1882 | ||
1883 | static bool | |
1884 | nvptx_print_operand_punct_valid_p (unsigned char c) | |
1885 | { | |
1886 | return c == '.' || c== '#'; | |
1887 | } | |
1888 | ||
1889 | static void nvptx_print_operand (FILE *, rtx, int); | |
1890 | ||
1891 | /* Subroutine of nvptx_print_operand; used to print a memory reference X to FILE. */ | |
1892 | ||
1893 | static void | |
1894 | nvptx_print_address_operand (FILE *file, rtx x, machine_mode) | |
1895 | { | |
1896 | rtx off; | |
1897 | if (GET_CODE (x) == CONST) | |
1898 | x = XEXP (x, 0); | |
1899 | switch (GET_CODE (x)) | |
1900 | { | |
1901 | case PLUS: | |
1902 | off = XEXP (x, 1); | |
cc8ca59e | 1903 | output_address (VOIDmode, XEXP (x, 0)); |
738f2522 | 1904 | fprintf (file, "+"); |
cc8ca59e | 1905 | output_address (VOIDmode, off); |
738f2522 BS |
1906 | break; |
1907 | ||
1908 | case SYMBOL_REF: | |
1909 | case LABEL_REF: | |
1910 | output_addr_const (file, x); | |
1911 | break; | |
1912 | ||
1913 | default: | |
1914 | gcc_assert (GET_CODE (x) != MEM); | |
1915 | nvptx_print_operand (file, x, 0); | |
1916 | break; | |
1917 | } | |
1918 | } | |
1919 | ||
1920 | /* Write assembly language output for the address ADDR to FILE. */ | |
1921 | ||
1922 | static void | |
cc8ca59e | 1923 | nvptx_print_operand_address (FILE *file, machine_mode mode, rtx addr) |
738f2522 | 1924 | { |
cc8ca59e | 1925 | nvptx_print_address_operand (file, addr, mode); |
738f2522 BS |
1926 | } |
1927 | ||
1928 | /* Print an operand, X, to FILE, with an optional modifier in CODE. | |
1929 | ||
1930 | Meaning of CODE: | |
1931 | . -- print the predicate for the instruction or an emptry string for an | |
1932 | unconditional one. | |
1933 | # -- print a rounding mode for the instruction | |
1934 | ||
9a863523 | 1935 | A -- print a data area for a MEM |
738f2522 | 1936 | c -- print an opcode suffix for a comparison operator, including a type code |
9a863523 | 1937 | D -- print a data area for a MEM operand |
d88cd9c4 | 1938 | S -- print a shuffle kind specified by CONST_INT |
738f2522 BS |
1939 | t -- print a type opcode suffix, promoting QImode to 32 bits |
1940 | T -- print a type size in bits | |
1941 | u -- print a type opcode suffix without promotions. */ | |
1942 | ||
1943 | static void | |
1944 | nvptx_print_operand (FILE *file, rtx x, int code) | |
1945 | { | |
738f2522 BS |
1946 | if (code == '.') |
1947 | { | |
1948 | x = current_insn_predicate; | |
1949 | if (x) | |
1950 | { | |
1951 | unsigned int regno = REGNO (XEXP (x, 0)); | |
1952 | fputs ("[", file); | |
1953 | if (GET_CODE (x) == EQ) | |
1954 | fputs ("!", file); | |
1955 | fputs (reg_names [regno], file); | |
1956 | fputs ("]", file); | |
1957 | } | |
1958 | return; | |
1959 | } | |
1960 | else if (code == '#') | |
1961 | { | |
1962 | fputs (".rn", file); | |
1963 | return; | |
1964 | } | |
1965 | ||
1966 | enum rtx_code x_code = GET_CODE (x); | |
f313d112 | 1967 | machine_mode mode = GET_MODE (x); |
738f2522 BS |
1968 | |
1969 | switch (code) | |
1970 | { | |
1971 | case 'A': | |
9a863523 NS |
1972 | x = XEXP (x, 0); |
1973 | /* FALLTHROUGH. */ | |
7b8edc29 | 1974 | |
9a863523 NS |
1975 | case 'D': |
1976 | if (GET_CODE (x) == CONST) | |
1977 | x = XEXP (x, 0); | |
1978 | if (GET_CODE (x) == PLUS) | |
1979 | x = XEXP (x, 0); | |
7b8edc29 | 1980 | |
9a863523 NS |
1981 | if (GET_CODE (x) == SYMBOL_REF) |
1982 | fputs (section_for_sym (x), file); | |
738f2522 BS |
1983 | break; |
1984 | ||
738f2522 | 1985 | case 't': |
738f2522 | 1986 | case 'u': |
f313d112 NS |
1987 | if (x_code == SUBREG) |
1988 | { | |
1989 | mode = GET_MODE (SUBREG_REG (x)); | |
1990 | if (mode == TImode) | |
1991 | mode = DImode; | |
1992 | else if (COMPLEX_MODE_P (mode)) | |
1993 | mode = GET_MODE_INNER (mode); | |
1994 | } | |
1995 | fprintf (file, "%s", nvptx_ptx_type_from_mode (mode, code == 't')); | |
738f2522 BS |
1996 | break; |
1997 | ||
d88cd9c4 NS |
1998 | case 'S': |
1999 | { | |
59263259 NS |
2000 | nvptx_shuffle_kind kind = (nvptx_shuffle_kind) UINTVAL (x); |
2001 | /* Same order as nvptx_shuffle_kind. */ | |
d88cd9c4 | 2002 | static const char *const kinds[] = |
59263259 NS |
2003 | {".up", ".down", ".bfly", ".idx"}; |
2004 | fputs (kinds[kind], file); | |
d88cd9c4 NS |
2005 | } |
2006 | break; | |
2007 | ||
738f2522 | 2008 | case 'T': |
f313d112 | 2009 | fprintf (file, "%d", GET_MODE_BITSIZE (mode)); |
738f2522 BS |
2010 | break; |
2011 | ||
2012 | case 'j': | |
2013 | fprintf (file, "@"); | |
2014 | goto common; | |
2015 | ||
2016 | case 'J': | |
2017 | fprintf (file, "@!"); | |
2018 | goto common; | |
2019 | ||
2020 | case 'c': | |
f313d112 | 2021 | mode = GET_MODE (XEXP (x, 0)); |
738f2522 BS |
2022 | switch (x_code) |
2023 | { | |
2024 | case EQ: | |
2025 | fputs (".eq", file); | |
2026 | break; | |
2027 | case NE: | |
f313d112 | 2028 | if (FLOAT_MODE_P (mode)) |
738f2522 BS |
2029 | fputs (".neu", file); |
2030 | else | |
2031 | fputs (".ne", file); | |
2032 | break; | |
2033 | case LE: | |
578fb225 | 2034 | case LEU: |
738f2522 BS |
2035 | fputs (".le", file); |
2036 | break; | |
2037 | case GE: | |
578fb225 | 2038 | case GEU: |
738f2522 BS |
2039 | fputs (".ge", file); |
2040 | break; | |
2041 | case LT: | |
578fb225 | 2042 | case LTU: |
738f2522 BS |
2043 | fputs (".lt", file); |
2044 | break; | |
2045 | case GT: | |
738f2522 | 2046 | case GTU: |
578fb225 | 2047 | fputs (".gt", file); |
738f2522 BS |
2048 | break; |
2049 | case LTGT: | |
2050 | fputs (".ne", file); | |
2051 | break; | |
2052 | case UNEQ: | |
2053 | fputs (".equ", file); | |
2054 | break; | |
2055 | case UNLE: | |
2056 | fputs (".leu", file); | |
2057 | break; | |
2058 | case UNGE: | |
2059 | fputs (".geu", file); | |
2060 | break; | |
2061 | case UNLT: | |
2062 | fputs (".ltu", file); | |
2063 | break; | |
2064 | case UNGT: | |
2065 | fputs (".gtu", file); | |
2066 | break; | |
2067 | case UNORDERED: | |
2068 | fputs (".nan", file); | |
2069 | break; | |
2070 | case ORDERED: | |
2071 | fputs (".num", file); | |
2072 | break; | |
2073 | default: | |
2074 | gcc_unreachable (); | |
2075 | } | |
f313d112 | 2076 | if (FLOAT_MODE_P (mode) |
738f2522 BS |
2077 | || x_code == EQ || x_code == NE |
2078 | || x_code == GEU || x_code == GTU | |
2079 | || x_code == LEU || x_code == LTU) | |
f313d112 | 2080 | fputs (nvptx_ptx_type_from_mode (mode, true), file); |
738f2522 | 2081 | else |
f313d112 | 2082 | fprintf (file, ".s%d", GET_MODE_BITSIZE (mode)); |
738f2522 BS |
2083 | break; |
2084 | default: | |
2085 | common: | |
2086 | switch (x_code) | |
2087 | { | |
2088 | case SUBREG: | |
f313d112 NS |
2089 | { |
2090 | rtx inner_x = SUBREG_REG (x); | |
2091 | machine_mode inner_mode = GET_MODE (inner_x); | |
2092 | machine_mode split = maybe_split_mode (inner_mode); | |
2093 | ||
2094 | if (split != VOIDmode | |
2095 | && (GET_MODE_SIZE (inner_mode) == GET_MODE_SIZE (mode))) | |
2096 | output_reg (file, REGNO (inner_x), split); | |
2097 | else | |
2098 | output_reg (file, REGNO (inner_x), split, SUBREG_BYTE (x)); | |
2099 | } | |
2100 | break; | |
738f2522 BS |
2101 | |
2102 | case REG: | |
f313d112 | 2103 | output_reg (file, REGNO (x), maybe_split_mode (mode)); |
738f2522 BS |
2104 | break; |
2105 | ||
2106 | case MEM: | |
2107 | fputc ('[', file); | |
f313d112 | 2108 | nvptx_print_address_operand (file, XEXP (x, 0), mode); |
738f2522 BS |
2109 | fputc (']', file); |
2110 | break; | |
2111 | ||
2112 | case CONST_INT: | |
2113 | output_addr_const (file, x); | |
2114 | break; | |
2115 | ||
2116 | case CONST: | |
2117 | case SYMBOL_REF: | |
2118 | case LABEL_REF: | |
2119 | /* We could use output_addr_const, but that can print things like | |
2120 | "x-8", which breaks ptxas. Need to ensure it is output as | |
2121 | "x+-8". */ | |
2122 | nvptx_print_address_operand (file, x, VOIDmode); | |
2123 | break; | |
2124 | ||
2125 | case CONST_DOUBLE: | |
2126 | long vals[2]; | |
f313d112 | 2127 | real_to_target (vals, CONST_DOUBLE_REAL_VALUE (x), mode); |
738f2522 BS |
2128 | vals[0] &= 0xffffffff; |
2129 | vals[1] &= 0xffffffff; | |
f313d112 | 2130 | if (mode == SFmode) |
738f2522 BS |
2131 | fprintf (file, "0f%08lx", vals[0]); |
2132 | else | |
2133 | fprintf (file, "0d%08lx%08lx", vals[1], vals[0]); | |
2134 | break; | |
2135 | ||
2136 | default: | |
2137 | output_addr_const (file, x); | |
2138 | } | |
2139 | } | |
2140 | } | |
2141 | \f | |
2142 | /* Record replacement regs used to deal with subreg operands. */ | |
2143 | struct reg_replace | |
2144 | { | |
2145 | rtx replacement[MAX_RECOG_OPERANDS]; | |
2146 | machine_mode mode; | |
2147 | int n_allocated; | |
2148 | int n_in_use; | |
2149 | }; | |
2150 | ||
2151 | /* Allocate or reuse a replacement in R and return the rtx. */ | |
2152 | ||
2153 | static rtx | |
2154 | get_replacement (struct reg_replace *r) | |
2155 | { | |
2156 | if (r->n_allocated == r->n_in_use) | |
2157 | r->replacement[r->n_allocated++] = gen_reg_rtx (r->mode); | |
2158 | return r->replacement[r->n_in_use++]; | |
2159 | } | |
2160 | ||
2161 | /* Clean up subreg operands. In ptx assembly, everything is typed, and | |
2162 | the presence of subregs would break the rules for most instructions. | |
2163 | Replace them with a suitable new register of the right size, plus | |
2164 | conversion copyin/copyout instructions. */ | |
2165 | ||
2166 | static void | |
517665b3 | 2167 | nvptx_reorg_subreg (void) |
738f2522 BS |
2168 | { |
2169 | struct reg_replace qiregs, hiregs, siregs, diregs; | |
2170 | rtx_insn *insn, *next; | |
2171 | ||
738f2522 BS |
2172 | qiregs.n_allocated = 0; |
2173 | hiregs.n_allocated = 0; | |
2174 | siregs.n_allocated = 0; | |
2175 | diregs.n_allocated = 0; | |
2176 | qiregs.mode = QImode; | |
2177 | hiregs.mode = HImode; | |
2178 | siregs.mode = SImode; | |
2179 | diregs.mode = DImode; | |
2180 | ||
2181 | for (insn = get_insns (); insn; insn = next) | |
2182 | { | |
2183 | next = NEXT_INSN (insn); | |
2184 | if (!NONDEBUG_INSN_P (insn) | |
1fe6befc | 2185 | || asm_noperands (PATTERN (insn)) >= 0 |
738f2522 BS |
2186 | || GET_CODE (PATTERN (insn)) == USE |
2187 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
2188 | continue; | |
f324806d | 2189 | |
738f2522 BS |
2190 | qiregs.n_in_use = 0; |
2191 | hiregs.n_in_use = 0; | |
2192 | siregs.n_in_use = 0; | |
2193 | diregs.n_in_use = 0; | |
2194 | extract_insn (insn); | |
2195 | enum attr_subregs_ok s_ok = get_attr_subregs_ok (insn); | |
f324806d | 2196 | |
738f2522 BS |
2197 | for (int i = 0; i < recog_data.n_operands; i++) |
2198 | { | |
2199 | rtx op = recog_data.operand[i]; | |
2200 | if (GET_CODE (op) != SUBREG) | |
2201 | continue; | |
2202 | ||
2203 | rtx inner = SUBREG_REG (op); | |
2204 | ||
2205 | machine_mode outer_mode = GET_MODE (op); | |
2206 | machine_mode inner_mode = GET_MODE (inner); | |
2207 | gcc_assert (s_ok); | |
2208 | if (s_ok | |
2209 | && (GET_MODE_PRECISION (inner_mode) | |
2210 | >= GET_MODE_PRECISION (outer_mode))) | |
2211 | continue; | |
2212 | gcc_assert (SCALAR_INT_MODE_P (outer_mode)); | |
2213 | struct reg_replace *r = (outer_mode == QImode ? &qiregs | |
2214 | : outer_mode == HImode ? &hiregs | |
2215 | : outer_mode == SImode ? &siregs | |
2216 | : &diregs); | |
2217 | rtx new_reg = get_replacement (r); | |
2218 | ||
2219 | if (recog_data.operand_type[i] != OP_OUT) | |
2220 | { | |
2221 | enum rtx_code code; | |
2222 | if (GET_MODE_PRECISION (inner_mode) | |
2223 | < GET_MODE_PRECISION (outer_mode)) | |
2224 | code = ZERO_EXTEND; | |
2225 | else | |
2226 | code = TRUNCATE; | |
2227 | ||
f7df4a84 | 2228 | rtx pat = gen_rtx_SET (new_reg, |
738f2522 BS |
2229 | gen_rtx_fmt_e (code, outer_mode, inner)); |
2230 | emit_insn_before (pat, insn); | |
2231 | } | |
2232 | ||
2233 | if (recog_data.operand_type[i] != OP_IN) | |
2234 | { | |
2235 | enum rtx_code code; | |
2236 | if (GET_MODE_PRECISION (inner_mode) | |
2237 | < GET_MODE_PRECISION (outer_mode)) | |
2238 | code = TRUNCATE; | |
2239 | else | |
2240 | code = ZERO_EXTEND; | |
2241 | ||
f7df4a84 | 2242 | rtx pat = gen_rtx_SET (inner, |
738f2522 BS |
2243 | gen_rtx_fmt_e (code, inner_mode, new_reg)); |
2244 | emit_insn_after (pat, insn); | |
2245 | } | |
2246 | validate_change (insn, recog_data.operand_loc[i], new_reg, false); | |
2247 | } | |
2248 | } | |
517665b3 | 2249 | } |
738f2522 | 2250 | |
d2d47a28 NS |
2251 | /* Loop structure of the function. The entire function is described as |
2252 | a NULL loop. */ | |
d88cd9c4 NS |
2253 | |
2254 | struct parallel | |
2255 | { | |
2256 | /* Parent parallel. */ | |
2257 | parallel *parent; | |
2258 | ||
2259 | /* Next sibling parallel. */ | |
2260 | parallel *next; | |
2261 | ||
2262 | /* First child parallel. */ | |
2263 | parallel *inner; | |
2264 | ||
2265 | /* Partitioning mask of the parallel. */ | |
2266 | unsigned mask; | |
2267 | ||
2268 | /* Partitioning used within inner parallels. */ | |
2269 | unsigned inner_mask; | |
2270 | ||
2271 | /* Location of parallel forked and join. The forked is the first | |
2272 | block in the parallel and the join is the first block after of | |
2273 | the partition. */ | |
2274 | basic_block forked_block; | |
2275 | basic_block join_block; | |
2276 | ||
2277 | rtx_insn *forked_insn; | |
2278 | rtx_insn *join_insn; | |
2279 | ||
2280 | rtx_insn *fork_insn; | |
2281 | rtx_insn *joining_insn; | |
2282 | ||
2283 | /* Basic blocks in this parallel, but not in child parallels. The | |
2284 | FORKED and JOINING blocks are in the partition. The FORK and JOIN | |
2285 | blocks are not. */ | |
2286 | auto_vec<basic_block> blocks; | |
2287 | ||
2288 | public: | |
2289 | parallel (parallel *parent, unsigned mode); | |
2290 | ~parallel (); | |
2291 | }; | |
2292 | ||
2293 | /* Constructor links the new parallel into it's parent's chain of | |
2294 | children. */ | |
2295 | ||
2296 | parallel::parallel (parallel *parent_, unsigned mask_) | |
2297 | :parent (parent_), next (0), inner (0), mask (mask_), inner_mask (0) | |
2298 | { | |
2299 | forked_block = join_block = 0; | |
2300 | forked_insn = join_insn = 0; | |
2301 | fork_insn = joining_insn = 0; | |
2302 | ||
2303 | if (parent) | |
2304 | { | |
2305 | next = parent->inner; | |
2306 | parent->inner = this; | |
2307 | } | |
2308 | } | |
2309 | ||
2310 | parallel::~parallel () | |
2311 | { | |
2312 | delete inner; | |
2313 | delete next; | |
2314 | } | |
2315 | ||
2316 | /* Map of basic blocks to insns */ | |
2317 | typedef hash_map<basic_block, rtx_insn *> bb_insn_map_t; | |
2318 | ||
2319 | /* A tuple of an insn of interest and the BB in which it resides. */ | |
2320 | typedef std::pair<rtx_insn *, basic_block> insn_bb_t; | |
2321 | typedef auto_vec<insn_bb_t> insn_bb_vec_t; | |
2322 | ||
2323 | /* Split basic blocks such that each forked and join unspecs are at | |
2324 | the start of their basic blocks. Thus afterwards each block will | |
2325 | have a single partitioning mode. We also do the same for return | |
2326 | insns, as they are executed by every thread. Return the | |
2327 | partitioning mode of the function as a whole. Populate MAP with | |
2328 | head and tail blocks. We also clear the BB visited flag, which is | |
2329 | used when finding partitions. */ | |
2330 | ||
2331 | static void | |
2332 | nvptx_split_blocks (bb_insn_map_t *map) | |
2333 | { | |
2334 | insn_bb_vec_t worklist; | |
2335 | basic_block block; | |
2336 | rtx_insn *insn; | |
2337 | ||
2338 | /* Locate all the reorg instructions of interest. */ | |
2339 | FOR_ALL_BB_FN (block, cfun) | |
2340 | { | |
2341 | bool seen_insn = false; | |
2342 | ||
2343 | /* Clear visited flag, for use by parallel locator */ | |
2344 | block->flags &= ~BB_VISITED; | |
2345 | ||
2346 | FOR_BB_INSNS (block, insn) | |
2347 | { | |
2348 | if (!INSN_P (insn)) | |
2349 | continue; | |
2350 | switch (recog_memoized (insn)) | |
2351 | { | |
2352 | default: | |
2353 | seen_insn = true; | |
2354 | continue; | |
2355 | case CODE_FOR_nvptx_forked: | |
2356 | case CODE_FOR_nvptx_join: | |
2357 | break; | |
2358 | ||
2359 | case CODE_FOR_return: | |
2360 | /* We also need to split just before return insns, as | |
2361 | that insn needs executing by all threads, but the | |
2362 | block it is in probably does not. */ | |
2363 | break; | |
2364 | } | |
2365 | ||
2366 | if (seen_insn) | |
2367 | /* We've found an instruction that must be at the start of | |
2368 | a block, but isn't. Add it to the worklist. */ | |
2369 | worklist.safe_push (insn_bb_t (insn, block)); | |
2370 | else | |
2371 | /* It was already the first instruction. Just add it to | |
2372 | the map. */ | |
2373 | map->get_or_insert (block) = insn; | |
2374 | seen_insn = true; | |
2375 | } | |
2376 | } | |
2377 | ||
2378 | /* Split blocks on the worklist. */ | |
2379 | unsigned ix; | |
2380 | insn_bb_t *elt; | |
2381 | basic_block remap = 0; | |
2382 | for (ix = 0; worklist.iterate (ix, &elt); ix++) | |
2383 | { | |
2384 | if (remap != elt->second) | |
2385 | { | |
2386 | block = elt->second; | |
2387 | remap = block; | |
2388 | } | |
2389 | ||
2390 | /* Split block before insn. The insn is in the new block */ | |
2391 | edge e = split_block (block, PREV_INSN (elt->first)); | |
2392 | ||
2393 | block = e->dest; | |
2394 | map->get_or_insert (block) = elt->first; | |
2395 | } | |
2396 | } | |
2397 | ||
2398 | /* BLOCK is a basic block containing a head or tail instruction. | |
2399 | Locate the associated prehead or pretail instruction, which must be | |
2400 | in the single predecessor block. */ | |
2401 | ||
2402 | static rtx_insn * | |
2403 | nvptx_discover_pre (basic_block block, int expected) | |
2404 | { | |
2405 | gcc_assert (block->preds->length () == 1); | |
2406 | basic_block pre_block = (*block->preds)[0]->src; | |
2407 | rtx_insn *pre_insn; | |
2408 | ||
2409 | for (pre_insn = BB_END (pre_block); !INSN_P (pre_insn); | |
2410 | pre_insn = PREV_INSN (pre_insn)) | |
2411 | gcc_assert (pre_insn != BB_HEAD (pre_block)); | |
2412 | ||
2413 | gcc_assert (recog_memoized (pre_insn) == expected); | |
2414 | return pre_insn; | |
2415 | } | |
2416 | ||
2417 | /* Dump this parallel and all its inner parallels. */ | |
2418 | ||
2419 | static void | |
2420 | nvptx_dump_pars (parallel *par, unsigned depth) | |
2421 | { | |
2422 | fprintf (dump_file, "%u: mask %d head=%d, tail=%d\n", | |
2423 | depth, par->mask, | |
2424 | par->forked_block ? par->forked_block->index : -1, | |
2425 | par->join_block ? par->join_block->index : -1); | |
2426 | ||
2427 | fprintf (dump_file, " blocks:"); | |
2428 | ||
2429 | basic_block block; | |
2430 | for (unsigned ix = 0; par->blocks.iterate (ix, &block); ix++) | |
2431 | fprintf (dump_file, " %d", block->index); | |
2432 | fprintf (dump_file, "\n"); | |
2433 | if (par->inner) | |
2434 | nvptx_dump_pars (par->inner, depth + 1); | |
2435 | ||
2436 | if (par->next) | |
2437 | nvptx_dump_pars (par->next, depth); | |
2438 | } | |
2439 | ||
2440 | /* If BLOCK contains a fork/join marker, process it to create or | |
2441 | terminate a loop structure. Add this block to the current loop, | |
2442 | and then walk successor blocks. */ | |
2443 | ||
2444 | static parallel * | |
2445 | nvptx_find_par (bb_insn_map_t *map, parallel *par, basic_block block) | |
2446 | { | |
2447 | if (block->flags & BB_VISITED) | |
2448 | return par; | |
2449 | block->flags |= BB_VISITED; | |
2450 | ||
2451 | if (rtx_insn **endp = map->get (block)) | |
2452 | { | |
2453 | rtx_insn *end = *endp; | |
2454 | ||
2455 | /* This is a block head or tail, or return instruction. */ | |
2456 | switch (recog_memoized (end)) | |
2457 | { | |
2458 | case CODE_FOR_return: | |
2459 | /* Return instructions are in their own block, and we | |
2460 | don't need to do anything more. */ | |
2461 | return par; | |
2462 | ||
2463 | case CODE_FOR_nvptx_forked: | |
2464 | /* Loop head, create a new inner loop and add it into | |
2465 | our parent's child list. */ | |
2466 | { | |
2467 | unsigned mask = UINTVAL (XVECEXP (PATTERN (end), 0, 0)); | |
2468 | ||
2469 | gcc_assert (mask); | |
2470 | par = new parallel (par, mask); | |
2471 | par->forked_block = block; | |
2472 | par->forked_insn = end; | |
2473 | if (!(mask & GOMP_DIM_MASK (GOMP_DIM_MAX)) | |
2474 | && (mask & GOMP_DIM_MASK (GOMP_DIM_WORKER))) | |
2475 | par->fork_insn | |
2476 | = nvptx_discover_pre (block, CODE_FOR_nvptx_fork); | |
2477 | } | |
2478 | break; | |
2479 | ||
2480 | case CODE_FOR_nvptx_join: | |
2481 | /* A loop tail. Finish the current loop and return to | |
2482 | parent. */ | |
2483 | { | |
2484 | unsigned mask = UINTVAL (XVECEXP (PATTERN (end), 0, 0)); | |
2485 | ||
2486 | gcc_assert (par->mask == mask); | |
2487 | par->join_block = block; | |
2488 | par->join_insn = end; | |
2489 | if (!(mask & GOMP_DIM_MASK (GOMP_DIM_MAX)) | |
2490 | && (mask & GOMP_DIM_MASK (GOMP_DIM_WORKER))) | |
2491 | par->joining_insn | |
2492 | = nvptx_discover_pre (block, CODE_FOR_nvptx_joining); | |
2493 | par = par->parent; | |
2494 | } | |
2495 | break; | |
2496 | ||
2497 | default: | |
2498 | gcc_unreachable (); | |
2499 | } | |
2500 | } | |
2501 | ||
2502 | if (par) | |
2503 | /* Add this block onto the current loop's list of blocks. */ | |
2504 | par->blocks.safe_push (block); | |
2505 | else | |
2506 | /* This must be the entry block. Create a NULL parallel. */ | |
2507 | par = new parallel (0, 0); | |
2508 | ||
2509 | /* Walk successor blocks. */ | |
2510 | edge e; | |
2511 | edge_iterator ei; | |
2512 | ||
2513 | FOR_EACH_EDGE (e, ei, block->succs) | |
2514 | nvptx_find_par (map, par, e->dest); | |
2515 | ||
2516 | return par; | |
2517 | } | |
2518 | ||
2519 | /* DFS walk the CFG looking for fork & join markers. Construct | |
2520 | loop structures as we go. MAP is a mapping of basic blocks | |
2521 | to head & tail markers, discovered when splitting blocks. This | |
2522 | speeds up the discovery. We rely on the BB visited flag having | |
2523 | been cleared when splitting blocks. */ | |
2524 | ||
2525 | static parallel * | |
2526 | nvptx_discover_pars (bb_insn_map_t *map) | |
2527 | { | |
2528 | basic_block block; | |
2529 | ||
2530 | /* Mark exit blocks as visited. */ | |
2531 | block = EXIT_BLOCK_PTR_FOR_FN (cfun); | |
2532 | block->flags |= BB_VISITED; | |
2533 | ||
2534 | /* And entry block as not. */ | |
2535 | block = ENTRY_BLOCK_PTR_FOR_FN (cfun); | |
2536 | block->flags &= ~BB_VISITED; | |
2537 | ||
2538 | parallel *par = nvptx_find_par (map, 0, block); | |
2539 | ||
2540 | if (dump_file) | |
2541 | { | |
2542 | fprintf (dump_file, "\nLoops\n"); | |
2543 | nvptx_dump_pars (par, 0); | |
2544 | fprintf (dump_file, "\n"); | |
2545 | } | |
2546 | ||
2547 | return par; | |
2548 | } | |
2549 | ||
912442c2 NS |
2550 | /* Analyse a group of BBs within a partitioned region and create N |
2551 | Single-Entry-Single-Exit regions. Some of those regions will be | |
2552 | trivial ones consisting of a single BB. The blocks of a | |
2553 | partitioned region might form a set of disjoint graphs -- because | |
2554 | the region encloses a differently partitoned sub region. | |
2555 | ||
2556 | We use the linear time algorithm described in 'Finding Regions Fast: | |
2557 | Single Entry Single Exit and control Regions in Linear Time' | |
2558 | Johnson, Pearson & Pingali. That algorithm deals with complete | |
2559 | CFGs, where a back edge is inserted from END to START, and thus the | |
2560 | problem becomes one of finding equivalent loops. | |
2561 | ||
2562 | In this case we have a partial CFG. We complete it by redirecting | |
2563 | any incoming edge to the graph to be from an arbitrary external BB, | |
2564 | and similarly redirecting any outgoing edge to be to that BB. | |
2565 | Thus we end up with a closed graph. | |
2566 | ||
2567 | The algorithm works by building a spanning tree of an undirected | |
2568 | graph and keeping track of back edges from nodes further from the | |
2569 | root in the tree to nodes nearer to the root in the tree. In the | |
2570 | description below, the root is up and the tree grows downwards. | |
2571 | ||
2572 | We avoid having to deal with degenerate back-edges to the same | |
2573 | block, by splitting each BB into 3 -- one for input edges, one for | |
2574 | the node itself and one for the output edges. Such back edges are | |
2575 | referred to as 'Brackets'. Cycle equivalent nodes will have the | |
2576 | same set of brackets. | |
2577 | ||
2578 | Determining bracket equivalency is done by maintaining a list of | |
2579 | brackets in such a manner that the list length and final bracket | |
2580 | uniquely identify the set. | |
2581 | ||
2582 | We use coloring to mark all BBs with cycle equivalency with the | |
2583 | same color. This is the output of the 'Finding Regions Fast' | |
2584 | algorithm. Notice it doesn't actually find the set of nodes within | |
2585 | a particular region, just unorderd sets of nodes that are the | |
2586 | entries and exits of SESE regions. | |
2587 | ||
2588 | After determining cycle equivalency, we need to find the minimal | |
2589 | set of SESE regions. Do this with a DFS coloring walk of the | |
2590 | complete graph. We're either 'looking' or 'coloring'. When | |
2591 | looking, and we're in the subgraph, we start coloring the color of | |
2592 | the current node, and remember that node as the start of the | |
2593 | current color's SESE region. Every time we go to a new node, we | |
2594 | decrement the count of nodes with thet color. If it reaches zero, | |
2595 | we remember that node as the end of the current color's SESE region | |
2596 | and return to 'looking'. Otherwise we color the node the current | |
2597 | color. | |
2598 | ||
2599 | This way we end up with coloring the inside of non-trivial SESE | |
2600 | regions with the color of that region. */ | |
2601 | ||
2602 | /* A pair of BBs. We use this to represent SESE regions. */ | |
2603 | typedef std::pair<basic_block, basic_block> bb_pair_t; | |
2604 | typedef auto_vec<bb_pair_t> bb_pair_vec_t; | |
2605 | ||
2606 | /* A node in the undirected CFG. The discriminator SECOND indicates just | |
2607 | above or just below the BB idicated by FIRST. */ | |
2608 | typedef std::pair<basic_block, int> pseudo_node_t; | |
2609 | ||
2610 | /* A bracket indicates an edge towards the root of the spanning tree of the | |
2611 | undirected graph. Each bracket has a color, determined | |
2612 | from the currrent set of brackets. */ | |
2613 | struct bracket | |
2614 | { | |
2615 | pseudo_node_t back; /* Back target */ | |
2616 | ||
2617 | /* Current color and size of set. */ | |
2618 | unsigned color; | |
2619 | unsigned size; | |
2620 | ||
2621 | bracket (pseudo_node_t back_) | |
2622 | : back (back_), color (~0u), size (~0u) | |
2623 | { | |
2624 | } | |
2625 | ||
2626 | unsigned get_color (auto_vec<unsigned> &color_counts, unsigned length) | |
2627 | { | |
2628 | if (length != size) | |
2629 | { | |
2630 | size = length; | |
2631 | color = color_counts.length (); | |
2632 | color_counts.quick_push (0); | |
2633 | } | |
2634 | color_counts[color]++; | |
2635 | return color; | |
2636 | } | |
2637 | }; | |
2638 | ||
2639 | typedef auto_vec<bracket> bracket_vec_t; | |
2640 | ||
2641 | /* Basic block info for finding SESE regions. */ | |
2642 | ||
2643 | struct bb_sese | |
2644 | { | |
2645 | int node; /* Node number in spanning tree. */ | |
2646 | int parent; /* Parent node number. */ | |
2647 | ||
2648 | /* The algorithm splits each node A into Ai, A', Ao. The incoming | |
2649 | edges arrive at pseudo-node Ai and the outgoing edges leave at | |
2650 | pseudo-node Ao. We have to remember which way we arrived at a | |
2651 | particular node when generating the spanning tree. dir > 0 means | |
2652 | we arrived at Ai, dir < 0 means we arrived at Ao. */ | |
2653 | int dir; | |
2654 | ||
2655 | /* Lowest numbered pseudo-node reached via a backedge from thsis | |
2656 | node, or any descendant. */ | |
2657 | pseudo_node_t high; | |
2658 | ||
2659 | int color; /* Cycle-equivalence color */ | |
2660 | ||
2661 | /* Stack of brackets for this node. */ | |
2662 | bracket_vec_t brackets; | |
2663 | ||
2664 | bb_sese (unsigned node_, unsigned p, int dir_) | |
2665 | :node (node_), parent (p), dir (dir_) | |
2666 | { | |
2667 | } | |
2668 | ~bb_sese (); | |
2669 | ||
2670 | /* Push a bracket ending at BACK. */ | |
2671 | void push (const pseudo_node_t &back) | |
2672 | { | |
2673 | if (dump_file) | |
2674 | fprintf (dump_file, "Pushing backedge %d:%+d\n", | |
2675 | back.first ? back.first->index : 0, back.second); | |
2676 | brackets.safe_push (bracket (back)); | |
2677 | } | |
2678 | ||
2679 | void append (bb_sese *child); | |
2680 | void remove (const pseudo_node_t &); | |
2681 | ||
2682 | /* Set node's color. */ | |
2683 | void set_color (auto_vec<unsigned> &color_counts) | |
2684 | { | |
2685 | color = brackets.last ().get_color (color_counts, brackets.length ()); | |
2686 | } | |
2687 | }; | |
2688 | ||
2689 | bb_sese::~bb_sese () | |
2690 | { | |
2691 | } | |
2692 | ||
2693 | /* Destructively append CHILD's brackets. */ | |
2694 | ||
2695 | void | |
2696 | bb_sese::append (bb_sese *child) | |
2697 | { | |
2698 | if (int len = child->brackets.length ()) | |
2699 | { | |
2700 | int ix; | |
2701 | ||
2702 | if (dump_file) | |
2703 | { | |
2704 | for (ix = 0; ix < len; ix++) | |
2705 | { | |
2706 | const pseudo_node_t &pseudo = child->brackets[ix].back; | |
2707 | fprintf (dump_file, "Appending (%d)'s backedge %d:%+d\n", | |
2708 | child->node, pseudo.first ? pseudo.first->index : 0, | |
2709 | pseudo.second); | |
2710 | } | |
2711 | } | |
2712 | if (!brackets.length ()) | |
2713 | std::swap (brackets, child->brackets); | |
2714 | else | |
2715 | { | |
2716 | brackets.reserve (len); | |
2717 | for (ix = 0; ix < len; ix++) | |
2718 | brackets.quick_push (child->brackets[ix]); | |
2719 | } | |
2720 | } | |
2721 | } | |
2722 | ||
2723 | /* Remove brackets that terminate at PSEUDO. */ | |
2724 | ||
2725 | void | |
2726 | bb_sese::remove (const pseudo_node_t &pseudo) | |
2727 | { | |
2728 | unsigned removed = 0; | |
2729 | int len = brackets.length (); | |
2730 | ||
2731 | for (int ix = 0; ix < len; ix++) | |
2732 | { | |
2733 | if (brackets[ix].back == pseudo) | |
2734 | { | |
2735 | if (dump_file) | |
2736 | fprintf (dump_file, "Removing backedge %d:%+d\n", | |
2737 | pseudo.first ? pseudo.first->index : 0, pseudo.second); | |
2738 | removed++; | |
2739 | } | |
2740 | else if (removed) | |
2741 | brackets[ix-removed] = brackets[ix]; | |
2742 | } | |
2743 | while (removed--) | |
2744 | brackets.pop (); | |
2745 | } | |
2746 | ||
2747 | /* Accessors for BB's aux pointer. */ | |
2748 | #define BB_SET_SESE(B, S) ((B)->aux = (S)) | |
2749 | #define BB_GET_SESE(B) ((bb_sese *)(B)->aux) | |
2750 | ||
2751 | /* DFS walk creating SESE data structures. Only cover nodes with | |
2752 | BB_VISITED set. Append discovered blocks to LIST. We number in | |
2753 | increments of 3 so that the above and below pseudo nodes can be | |
2754 | implicitly numbered too. */ | |
2755 | ||
2756 | static int | |
2757 | nvptx_sese_number (int n, int p, int dir, basic_block b, | |
2758 | auto_vec<basic_block> *list) | |
2759 | { | |
2760 | if (BB_GET_SESE (b)) | |
2761 | return n; | |
2762 | ||
2763 | if (dump_file) | |
2764 | fprintf (dump_file, "Block %d(%d), parent (%d), orientation %+d\n", | |
2765 | b->index, n, p, dir); | |
2766 | ||
2767 | BB_SET_SESE (b, new bb_sese (n, p, dir)); | |
2768 | p = n; | |
2769 | ||
2770 | n += 3; | |
2771 | list->quick_push (b); | |
2772 | ||
2773 | /* First walk the nodes on the 'other side' of this node, then walk | |
2774 | the nodes on the same side. */ | |
2775 | for (unsigned ix = 2; ix; ix--) | |
2776 | { | |
2777 | vec<edge, va_gc> *edges = dir > 0 ? b->succs : b->preds; | |
2778 | size_t offset = (dir > 0 ? offsetof (edge_def, dest) | |
2779 | : offsetof (edge_def, src)); | |
2780 | edge e; | |
2781 | edge_iterator (ei); | |
2782 | ||
2783 | FOR_EACH_EDGE (e, ei, edges) | |
2784 | { | |
2785 | basic_block target = *(basic_block *)((char *)e + offset); | |
2786 | ||
2787 | if (target->flags & BB_VISITED) | |
2788 | n = nvptx_sese_number (n, p, dir, target, list); | |
2789 | } | |
2790 | dir = -dir; | |
2791 | } | |
2792 | return n; | |
2793 | } | |
2794 | ||
2795 | /* Process pseudo node above (DIR < 0) or below (DIR > 0) ME. | |
2796 | EDGES are the outgoing edges and OFFSET is the offset to the src | |
2797 | or dst block on the edges. */ | |
2798 | ||
2799 | static void | |
2800 | nvptx_sese_pseudo (basic_block me, bb_sese *sese, int depth, int dir, | |
2801 | vec<edge, va_gc> *edges, size_t offset) | |
2802 | { | |
2803 | edge e; | |
2804 | edge_iterator (ei); | |
2805 | int hi_back = depth; | |
2806 | pseudo_node_t node_back (0, depth); | |
2807 | int hi_child = depth; | |
2808 | pseudo_node_t node_child (0, depth); | |
2809 | basic_block child = NULL; | |
2810 | unsigned num_children = 0; | |
2811 | int usd = -dir * sese->dir; | |
2812 | ||
2813 | if (dump_file) | |
2814 | fprintf (dump_file, "\nProcessing %d(%d) %+d\n", | |
2815 | me->index, sese->node, dir); | |
2816 | ||
2817 | if (dir < 0) | |
2818 | { | |
2819 | /* This is the above pseudo-child. It has the BB itself as an | |
2820 | additional child node. */ | |
2821 | node_child = sese->high; | |
2822 | hi_child = node_child.second; | |
2823 | if (node_child.first) | |
2824 | hi_child += BB_GET_SESE (node_child.first)->node; | |
2825 | num_children++; | |
2826 | } | |
2827 | ||
2828 | /* Examine each edge. | |
2829 | - if it is a child (a) append its bracket list and (b) record | |
2830 | whether it is the child with the highest reaching bracket. | |
2831 | - if it is an edge to ancestor, record whether it's the highest | |
2832 | reaching backlink. */ | |
2833 | FOR_EACH_EDGE (e, ei, edges) | |
2834 | { | |
2835 | basic_block target = *(basic_block *)((char *)e + offset); | |
2836 | ||
2837 | if (bb_sese *t_sese = BB_GET_SESE (target)) | |
2838 | { | |
2839 | if (t_sese->parent == sese->node && !(t_sese->dir + usd)) | |
2840 | { | |
2841 | /* Child node. Append its bracket list. */ | |
2842 | num_children++; | |
2843 | sese->append (t_sese); | |
2844 | ||
2845 | /* Compare it's hi value. */ | |
2846 | int t_hi = t_sese->high.second; | |
2847 | ||
2848 | if (basic_block child_hi_block = t_sese->high.first) | |
2849 | t_hi += BB_GET_SESE (child_hi_block)->node; | |
2850 | ||
2851 | if (hi_child > t_hi) | |
2852 | { | |
2853 | hi_child = t_hi; | |
2854 | node_child = t_sese->high; | |
2855 | child = target; | |
2856 | } | |
2857 | } | |
2858 | else if (t_sese->node < sese->node + dir | |
2859 | && !(dir < 0 && sese->parent == t_sese->node)) | |
2860 | { | |
2861 | /* Non-parental ancestor node -- a backlink. */ | |
2862 | int d = usd * t_sese->dir; | |
2863 | int back = t_sese->node + d; | |
2864 | ||
2865 | if (hi_back > back) | |
2866 | { | |
2867 | hi_back = back; | |
2868 | node_back = pseudo_node_t (target, d); | |
2869 | } | |
2870 | } | |
2871 | } | |
2872 | else | |
2873 | { /* Fallen off graph, backlink to entry node. */ | |
2874 | hi_back = 0; | |
2875 | node_back = pseudo_node_t (0, 0); | |
2876 | } | |
2877 | } | |
2878 | ||
2879 | /* Remove any brackets that terminate at this pseudo node. */ | |
2880 | sese->remove (pseudo_node_t (me, dir)); | |
2881 | ||
2882 | /* Now push any backlinks from this pseudo node. */ | |
2883 | FOR_EACH_EDGE (e, ei, edges) | |
2884 | { | |
2885 | basic_block target = *(basic_block *)((char *)e + offset); | |
2886 | if (bb_sese *t_sese = BB_GET_SESE (target)) | |
2887 | { | |
2888 | if (t_sese->node < sese->node + dir | |
2889 | && !(dir < 0 && sese->parent == t_sese->node)) | |
2890 | /* Non-parental ancestor node - backedge from me. */ | |
2891 | sese->push (pseudo_node_t (target, usd * t_sese->dir)); | |
2892 | } | |
2893 | else | |
2894 | { | |
2895 | /* back edge to entry node */ | |
2896 | sese->push (pseudo_node_t (0, 0)); | |
2897 | } | |
2898 | } | |
2899 | ||
2900 | /* If this node leads directly or indirectly to a no-return region of | |
2901 | the graph, then fake a backedge to entry node. */ | |
2902 | if (!sese->brackets.length () || !edges || !edges->length ()) | |
2903 | { | |
2904 | hi_back = 0; | |
2905 | node_back = pseudo_node_t (0, 0); | |
2906 | sese->push (node_back); | |
2907 | } | |
2908 | ||
2909 | /* Record the highest reaching backedge from us or a descendant. */ | |
2910 | sese->high = hi_back < hi_child ? node_back : node_child; | |
2911 | ||
2912 | if (num_children > 1) | |
2913 | { | |
2914 | /* There is more than one child -- this is a Y shaped piece of | |
2915 | spanning tree. We have to insert a fake backedge from this | |
2916 | node to the highest ancestor reached by not-the-highest | |
2917 | reaching child. Note that there may be multiple children | |
2918 | with backedges to the same highest node. That's ok and we | |
2919 | insert the edge to that highest node. */ | |
2920 | hi_child = depth; | |
2921 | if (dir < 0 && child) | |
2922 | { | |
2923 | node_child = sese->high; | |
2924 | hi_child = node_child.second; | |
2925 | if (node_child.first) | |
2926 | hi_child += BB_GET_SESE (node_child.first)->node; | |
2927 | } | |
2928 | ||
2929 | FOR_EACH_EDGE (e, ei, edges) | |
2930 | { | |
2931 | basic_block target = *(basic_block *)((char *)e + offset); | |
2932 | ||
2933 | if (target == child) | |
2934 | /* Ignore the highest child. */ | |
2935 | continue; | |
2936 | ||
2937 | bb_sese *t_sese = BB_GET_SESE (target); | |
2938 | if (!t_sese) | |
2939 | continue; | |
2940 | if (t_sese->parent != sese->node) | |
2941 | /* Not a child. */ | |
2942 | continue; | |
2943 | ||
2944 | /* Compare its hi value. */ | |
2945 | int t_hi = t_sese->high.second; | |
2946 | ||
2947 | if (basic_block child_hi_block = t_sese->high.first) | |
2948 | t_hi += BB_GET_SESE (child_hi_block)->node; | |
2949 | ||
2950 | if (hi_child > t_hi) | |
2951 | { | |
2952 | hi_child = t_hi; | |
2953 | node_child = t_sese->high; | |
2954 | } | |
2955 | } | |
2956 | ||
2957 | sese->push (node_child); | |
2958 | } | |
2959 | } | |
2960 | ||
2961 | ||
2962 | /* DFS walk of BB graph. Color node BLOCK according to COLORING then | |
2963 | proceed to successors. Set SESE entry and exit nodes of | |
2964 | REGIONS. */ | |
2965 | ||
2966 | static void | |
2967 | nvptx_sese_color (auto_vec<unsigned> &color_counts, bb_pair_vec_t ®ions, | |
2968 | basic_block block, int coloring) | |
2969 | { | |
2970 | bb_sese *sese = BB_GET_SESE (block); | |
2971 | ||
2972 | if (block->flags & BB_VISITED) | |
2973 | { | |
2974 | /* If we've already encountered this block, either we must not | |
2975 | be coloring, or it must have been colored the current color. */ | |
2976 | gcc_assert (coloring < 0 || (sese && coloring == sese->color)); | |
2977 | return; | |
2978 | } | |
2979 | ||
2980 | block->flags |= BB_VISITED; | |
2981 | ||
2982 | if (sese) | |
2983 | { | |
2984 | if (coloring < 0) | |
2985 | { | |
2986 | /* Start coloring a region. */ | |
2987 | regions[sese->color].first = block; | |
2988 | coloring = sese->color; | |
2989 | } | |
2990 | ||
2991 | if (!--color_counts[sese->color] && sese->color == coloring) | |
2992 | { | |
2993 | /* Found final block of SESE region. */ | |
2994 | regions[sese->color].second = block; | |
2995 | coloring = -1; | |
2996 | } | |
2997 | else | |
2998 | /* Color the node, so we can assert on revisiting the node | |
2999 | that the graph is indeed SESE. */ | |
3000 | sese->color = coloring; | |
3001 | } | |
3002 | else | |
3003 | /* Fallen off the subgraph, we cannot be coloring. */ | |
3004 | gcc_assert (coloring < 0); | |
3005 | ||
3006 | /* Walk each successor block. */ | |
3007 | if (block->succs && block->succs->length ()) | |
3008 | { | |
3009 | edge e; | |
3010 | edge_iterator ei; | |
3011 | ||
3012 | FOR_EACH_EDGE (e, ei, block->succs) | |
3013 | nvptx_sese_color (color_counts, regions, e->dest, coloring); | |
3014 | } | |
3015 | else | |
3016 | gcc_assert (coloring < 0); | |
3017 | } | |
3018 | ||
3019 | /* Find minimal set of SESE regions covering BLOCKS. REGIONS might | |
3020 | end up with NULL entries in it. */ | |
3021 | ||
3022 | static void | |
3023 | nvptx_find_sese (auto_vec<basic_block> &blocks, bb_pair_vec_t ®ions) | |
3024 | { | |
3025 | basic_block block; | |
3026 | int ix; | |
3027 | ||
3028 | /* First clear each BB of the whole function. */ | |
3029 | FOR_EACH_BB_FN (block, cfun) | |
3030 | { | |
3031 | block->flags &= ~BB_VISITED; | |
3032 | BB_SET_SESE (block, 0); | |
3033 | } | |
3034 | block = EXIT_BLOCK_PTR_FOR_FN (cfun); | |
3035 | block->flags &= ~BB_VISITED; | |
3036 | BB_SET_SESE (block, 0); | |
3037 | block = ENTRY_BLOCK_PTR_FOR_FN (cfun); | |
3038 | block->flags &= ~BB_VISITED; | |
3039 | BB_SET_SESE (block, 0); | |
3040 | ||
3041 | /* Mark blocks in the function that are in this graph. */ | |
3042 | for (ix = 0; blocks.iterate (ix, &block); ix++) | |
3043 | block->flags |= BB_VISITED; | |
3044 | ||
3045 | /* Counts of nodes assigned to each color. There cannot be more | |
3046 | colors than blocks (and hopefully there will be fewer). */ | |
3047 | auto_vec<unsigned> color_counts; | |
3048 | color_counts.reserve (blocks.length ()); | |
3049 | ||
3050 | /* Worklist of nodes in the spanning tree. Again, there cannot be | |
3051 | more nodes in the tree than blocks (there will be fewer if the | |
3052 | CFG of blocks is disjoint). */ | |
3053 | auto_vec<basic_block> spanlist; | |
3054 | spanlist.reserve (blocks.length ()); | |
3055 | ||
3056 | /* Make sure every block has its cycle class determined. */ | |
3057 | for (ix = 0; blocks.iterate (ix, &block); ix++) | |
3058 | { | |
3059 | if (BB_GET_SESE (block)) | |
3060 | /* We already met this block in an earlier graph solve. */ | |
3061 | continue; | |
3062 | ||
3063 | if (dump_file) | |
3064 | fprintf (dump_file, "Searching graph starting at %d\n", block->index); | |
3065 | ||
3066 | /* Number the nodes reachable from block initial DFS order. */ | |
3067 | int depth = nvptx_sese_number (2, 0, +1, block, &spanlist); | |
3068 | ||
3069 | /* Now walk in reverse DFS order to find cycle equivalents. */ | |
3070 | while (spanlist.length ()) | |
3071 | { | |
3072 | block = spanlist.pop (); | |
3073 | bb_sese *sese = BB_GET_SESE (block); | |
3074 | ||
3075 | /* Do the pseudo node below. */ | |
3076 | nvptx_sese_pseudo (block, sese, depth, +1, | |
3077 | sese->dir > 0 ? block->succs : block->preds, | |
3078 | (sese->dir > 0 ? offsetof (edge_def, dest) | |
3079 | : offsetof (edge_def, src))); | |
3080 | sese->set_color (color_counts); | |
3081 | /* Do the pseudo node above. */ | |
3082 | nvptx_sese_pseudo (block, sese, depth, -1, | |
3083 | sese->dir < 0 ? block->succs : block->preds, | |
3084 | (sese->dir < 0 ? offsetof (edge_def, dest) | |
3085 | : offsetof (edge_def, src))); | |
3086 | } | |
3087 | if (dump_file) | |
3088 | fprintf (dump_file, "\n"); | |
3089 | } | |
3090 | ||
3091 | if (dump_file) | |
3092 | { | |
3093 | unsigned count; | |
3094 | const char *comma = ""; | |
3095 | ||
3096 | fprintf (dump_file, "Found %d cycle equivalents\n", | |
3097 | color_counts.length ()); | |
3098 | for (ix = 0; color_counts.iterate (ix, &count); ix++) | |
3099 | { | |
3100 | fprintf (dump_file, "%s%d[%d]={", comma, ix, count); | |
3101 | ||
3102 | comma = ""; | |
3103 | for (unsigned jx = 0; blocks.iterate (jx, &block); jx++) | |
3104 | if (BB_GET_SESE (block)->color == ix) | |
3105 | { | |
3106 | block->flags |= BB_VISITED; | |
3107 | fprintf (dump_file, "%s%d", comma, block->index); | |
3108 | comma=","; | |
3109 | } | |
3110 | fprintf (dump_file, "}"); | |
3111 | comma = ", "; | |
3112 | } | |
3113 | fprintf (dump_file, "\n"); | |
3114 | } | |
3115 | ||
3116 | /* Now we've colored every block in the subgraph. We now need to | |
3117 | determine the minimal set of SESE regions that cover that | |
3118 | subgraph. Do this with a DFS walk of the complete function. | |
3119 | During the walk we're either 'looking' or 'coloring'. When we | |
3120 | reach the last node of a particular color, we stop coloring and | |
3121 | return to looking. */ | |
3122 | ||
3123 | /* There cannot be more SESE regions than colors. */ | |
3124 | regions.reserve (color_counts.length ()); | |
3125 | for (ix = color_counts.length (); ix--;) | |
3126 | regions.quick_push (bb_pair_t (0, 0)); | |
3127 | ||
3128 | for (ix = 0; blocks.iterate (ix, &block); ix++) | |
3129 | block->flags &= ~BB_VISITED; | |
3130 | ||
3131 | nvptx_sese_color (color_counts, regions, ENTRY_BLOCK_PTR_FOR_FN (cfun), -1); | |
3132 | ||
3133 | if (dump_file) | |
3134 | { | |
3135 | const char *comma = ""; | |
3136 | int len = regions.length (); | |
3137 | ||
3138 | fprintf (dump_file, "SESE regions:"); | |
3139 | for (ix = 0; ix != len; ix++) | |
3140 | { | |
3141 | basic_block from = regions[ix].first; | |
3142 | basic_block to = regions[ix].second; | |
3143 | ||
3144 | if (from) | |
3145 | { | |
3146 | fprintf (dump_file, "%s %d{%d", comma, ix, from->index); | |
3147 | if (to != from) | |
3148 | fprintf (dump_file, "->%d", to->index); | |
3149 | ||
3150 | int color = BB_GET_SESE (from)->color; | |
3151 | ||
3152 | /* Print the blocks within the region (excluding ends). */ | |
3153 | FOR_EACH_BB_FN (block, cfun) | |
3154 | { | |
3155 | bb_sese *sese = BB_GET_SESE (block); | |
3156 | ||
3157 | if (sese && sese->color == color | |
3158 | && block != from && block != to) | |
3159 | fprintf (dump_file, ".%d", block->index); | |
3160 | } | |
3161 | fprintf (dump_file, "}"); | |
3162 | } | |
3163 | comma = ","; | |
3164 | } | |
3165 | fprintf (dump_file, "\n\n"); | |
3166 | } | |
3167 | ||
3168 | for (ix = 0; blocks.iterate (ix, &block); ix++) | |
3169 | delete BB_GET_SESE (block); | |
3170 | } | |
3171 | ||
3172 | #undef BB_SET_SESE | |
3173 | #undef BB_GET_SESE | |
3174 | ||
d88cd9c4 NS |
3175 | /* Propagate live state at the start of a partitioned region. BLOCK |
3176 | provides the live register information, and might not contain | |
3177 | INSN. Propagation is inserted just after INSN. RW indicates whether | |
3178 | we are reading and/or writing state. This | |
3179 | separation is needed for worker-level proppagation where we | |
3180 | essentially do a spill & fill. FN is the underlying worker | |
3181 | function to generate the propagation instructions for single | |
3182 | register. DATA is user data. | |
3183 | ||
3184 | We propagate the live register set and the entire frame. We could | |
3185 | do better by (a) propagating just the live set that is used within | |
3186 | the partitioned regions and (b) only propagating stack entries that | |
3187 | are used. The latter might be quite hard to determine. */ | |
3188 | ||
3189 | typedef rtx (*propagator_fn) (rtx, propagate_mask, unsigned, void *); | |
3190 | ||
3191 | static void | |
3192 | nvptx_propagate (basic_block block, rtx_insn *insn, propagate_mask rw, | |
3193 | propagator_fn fn, void *data) | |
3194 | { | |
3195 | bitmap live = DF_LIVE_IN (block); | |
3196 | bitmap_iterator iterator; | |
3197 | unsigned ix; | |
3198 | ||
3199 | /* Copy the frame array. */ | |
3200 | HOST_WIDE_INT fs = get_frame_size (); | |
3201 | if (fs) | |
3202 | { | |
3203 | rtx tmp = gen_reg_rtx (DImode); | |
3204 | rtx idx = NULL_RTX; | |
3205 | rtx ptr = gen_reg_rtx (Pmode); | |
3206 | rtx pred = NULL_RTX; | |
3207 | rtx_code_label *label = NULL; | |
3208 | ||
3209 | gcc_assert (!(fs & (GET_MODE_SIZE (DImode) - 1))); | |
3210 | fs /= GET_MODE_SIZE (DImode); | |
3211 | /* Detect single iteration loop. */ | |
3212 | if (fs == 1) | |
3213 | fs = 0; | |
3214 | ||
3215 | start_sequence (); | |
3216 | emit_insn (gen_rtx_SET (ptr, frame_pointer_rtx)); | |
3217 | if (fs) | |
3218 | { | |
3219 | idx = gen_reg_rtx (SImode); | |
3220 | pred = gen_reg_rtx (BImode); | |
3221 | label = gen_label_rtx (); | |
3222 | ||
3223 | emit_insn (gen_rtx_SET (idx, GEN_INT (fs))); | |
3224 | /* Allow worker function to initialize anything needed. */ | |
3225 | rtx init = fn (tmp, PM_loop_begin, fs, data); | |
3226 | if (init) | |
3227 | emit_insn (init); | |
3228 | emit_label (label); | |
3229 | LABEL_NUSES (label)++; | |
3230 | emit_insn (gen_addsi3 (idx, idx, GEN_INT (-1))); | |
3231 | } | |
3232 | if (rw & PM_read) | |
3233 | emit_insn (gen_rtx_SET (tmp, gen_rtx_MEM (DImode, ptr))); | |
3234 | emit_insn (fn (tmp, rw, fs, data)); | |
3235 | if (rw & PM_write) | |
3236 | emit_insn (gen_rtx_SET (gen_rtx_MEM (DImode, ptr), tmp)); | |
3237 | if (fs) | |
3238 | { | |
3239 | emit_insn (gen_rtx_SET (pred, gen_rtx_NE (BImode, idx, const0_rtx))); | |
3240 | emit_insn (gen_adddi3 (ptr, ptr, GEN_INT (GET_MODE_SIZE (DImode)))); | |
3241 | emit_insn (gen_br_true_uni (pred, label)); | |
3242 | rtx fini = fn (tmp, PM_loop_end, fs, data); | |
3243 | if (fini) | |
3244 | emit_insn (fini); | |
3245 | emit_insn (gen_rtx_CLOBBER (GET_MODE (idx), idx)); | |
3246 | } | |
3247 | emit_insn (gen_rtx_CLOBBER (GET_MODE (tmp), tmp)); | |
3248 | emit_insn (gen_rtx_CLOBBER (GET_MODE (ptr), ptr)); | |
3249 | rtx cpy = get_insns (); | |
3250 | end_sequence (); | |
3251 | insn = emit_insn_after (cpy, insn); | |
3252 | } | |
3253 | ||
3254 | /* Copy live registers. */ | |
3255 | EXECUTE_IF_SET_IN_BITMAP (live, 0, ix, iterator) | |
3256 | { | |
3257 | rtx reg = regno_reg_rtx[ix]; | |
3258 | ||
3259 | if (REGNO (reg) >= FIRST_PSEUDO_REGISTER) | |
3260 | { | |
3261 | rtx bcast = fn (reg, rw, 0, data); | |
3262 | ||
3263 | insn = emit_insn_after (bcast, insn); | |
3264 | } | |
3265 | } | |
3266 | } | |
3267 | ||
3268 | /* Worker for nvptx_vpropagate. */ | |
3269 | ||
3270 | static rtx | |
3271 | vprop_gen (rtx reg, propagate_mask pm, | |
3272 | unsigned ARG_UNUSED (count), void *ARG_UNUSED (data)) | |
3273 | { | |
3274 | if (!(pm & PM_read_write)) | |
3275 | return 0; | |
3276 | ||
3277 | return nvptx_gen_vcast (reg); | |
3278 | } | |
3279 | ||
3280 | /* Propagate state that is live at start of BLOCK across the vectors | |
3281 | of a single warp. Propagation is inserted just after INSN. */ | |
3282 | ||
3283 | static void | |
3284 | nvptx_vpropagate (basic_block block, rtx_insn *insn) | |
3285 | { | |
3286 | nvptx_propagate (block, insn, PM_read_write, vprop_gen, 0); | |
3287 | } | |
3288 | ||
3289 | /* Worker for nvptx_wpropagate. */ | |
3290 | ||
3291 | static rtx | |
3292 | wprop_gen (rtx reg, propagate_mask pm, unsigned rep, void *data_) | |
3293 | { | |
3294 | wcast_data_t *data = (wcast_data_t *)data_; | |
3295 | ||
3296 | if (pm & PM_loop_begin) | |
3297 | { | |
3298 | /* Starting a loop, initialize pointer. */ | |
3299 | unsigned align = GET_MODE_ALIGNMENT (GET_MODE (reg)) / BITS_PER_UNIT; | |
3300 | ||
3301 | if (align > worker_bcast_align) | |
3302 | worker_bcast_align = align; | |
3303 | data->offset = (data->offset + align - 1) & ~(align - 1); | |
3304 | ||
3305 | data->ptr = gen_reg_rtx (Pmode); | |
3306 | ||
3307 | return gen_adddi3 (data->ptr, data->base, GEN_INT (data->offset)); | |
3308 | } | |
3309 | else if (pm & PM_loop_end) | |
3310 | { | |
3311 | rtx clobber = gen_rtx_CLOBBER (GET_MODE (data->ptr), data->ptr); | |
3312 | data->ptr = NULL_RTX; | |
3313 | return clobber; | |
3314 | } | |
3315 | else | |
3316 | return nvptx_gen_wcast (reg, pm, rep, data); | |
3317 | } | |
3318 | ||
3319 | /* Spill or fill live state that is live at start of BLOCK. PRE_P | |
3320 | indicates if this is just before partitioned mode (do spill), or | |
3321 | just after it starts (do fill). Sequence is inserted just after | |
3322 | INSN. */ | |
3323 | ||
3324 | static void | |
3325 | nvptx_wpropagate (bool pre_p, basic_block block, rtx_insn *insn) | |
3326 | { | |
3327 | wcast_data_t data; | |
3328 | ||
3329 | data.base = gen_reg_rtx (Pmode); | |
3330 | data.offset = 0; | |
3331 | data.ptr = NULL_RTX; | |
3332 | ||
3333 | nvptx_propagate (block, insn, pre_p ? PM_read : PM_write, wprop_gen, &data); | |
3334 | if (data.offset) | |
3335 | { | |
3336 | /* Stuff was emitted, initialize the base pointer now. */ | |
bd602b7f | 3337 | rtx init = gen_rtx_SET (data.base, worker_bcast_sym); |
d88cd9c4 | 3338 | emit_insn_after (init, insn); |
15ab6f00 | 3339 | |
d88cd9c4 NS |
3340 | if (worker_bcast_size < data.offset) |
3341 | worker_bcast_size = data.offset; | |
3342 | } | |
3343 | } | |
3344 | ||
3345 | /* Emit a worker-level synchronization barrier. We use different | |
3346 | markers for before and after synchronizations. */ | |
3347 | ||
3348 | static rtx | |
3349 | nvptx_wsync (bool after) | |
3350 | { | |
3351 | return gen_nvptx_barsync (GEN_INT (after)); | |
3352 | } | |
3353 | ||
3354 | /* Single neutering according to MASK. FROM is the incoming block and | |
3355 | TO is the outgoing block. These may be the same block. Insert at | |
3356 | start of FROM: | |
3357 | ||
3358 | if (tid.<axis>) goto end. | |
3359 | ||
3360 | and insert before ending branch of TO (if there is such an insn): | |
3361 | ||
3362 | end: | |
3363 | <possibly-broadcast-cond> | |
3364 | <branch> | |
3365 | ||
3366 | We currently only use differnt FROM and TO when skipping an entire | |
3367 | loop. We could do more if we detected superblocks. */ | |
3368 | ||
3369 | static void | |
3370 | nvptx_single (unsigned mask, basic_block from, basic_block to) | |
3371 | { | |
3372 | rtx_insn *head = BB_HEAD (from); | |
3373 | rtx_insn *tail = BB_END (to); | |
3374 | unsigned skip_mask = mask; | |
3375 | ||
3376 | /* Find first insn of from block */ | |
3377 | while (head != BB_END (from) && !INSN_P (head)) | |
3378 | head = NEXT_INSN (head); | |
3379 | ||
3380 | /* Find last insn of to block */ | |
3381 | rtx_insn *limit = from == to ? head : BB_HEAD (to); | |
3382 | while (tail != limit && !INSN_P (tail) && !LABEL_P (tail)) | |
3383 | tail = PREV_INSN (tail); | |
3384 | ||
3385 | /* Detect if tail is a branch. */ | |
3386 | rtx tail_branch = NULL_RTX; | |
3387 | rtx cond_branch = NULL_RTX; | |
3388 | if (tail && INSN_P (tail)) | |
3389 | { | |
3390 | tail_branch = PATTERN (tail); | |
3391 | if (GET_CODE (tail_branch) != SET || SET_DEST (tail_branch) != pc_rtx) | |
3392 | tail_branch = NULL_RTX; | |
3393 | else | |
3394 | { | |
3395 | cond_branch = SET_SRC (tail_branch); | |
3396 | if (GET_CODE (cond_branch) != IF_THEN_ELSE) | |
3397 | cond_branch = NULL_RTX; | |
3398 | } | |
3399 | } | |
3400 | ||
3401 | if (tail == head) | |
3402 | { | |
3403 | /* If this is empty, do nothing. */ | |
3404 | if (!head || !INSN_P (head)) | |
3405 | return; | |
3406 | ||
3407 | /* If this is a dummy insn, do nothing. */ | |
3408 | switch (recog_memoized (head)) | |
3409 | { | |
3410 | default: | |
3411 | break; | |
3412 | case CODE_FOR_nvptx_fork: | |
3413 | case CODE_FOR_nvptx_forked: | |
3414 | case CODE_FOR_nvptx_joining: | |
3415 | case CODE_FOR_nvptx_join: | |
3416 | return; | |
3417 | } | |
3418 | ||
3419 | if (cond_branch) | |
3420 | { | |
3421 | /* If we're only doing vector single, there's no need to | |
3422 | emit skip code because we'll not insert anything. */ | |
3423 | if (!(mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR))) | |
3424 | skip_mask = 0; | |
3425 | } | |
3426 | else if (tail_branch) | |
3427 | /* Block with only unconditional branch. Nothing to do. */ | |
3428 | return; | |
3429 | } | |
3430 | ||
3431 | /* Insert the vector test inside the worker test. */ | |
3432 | unsigned mode; | |
3433 | rtx_insn *before = tail; | |
3434 | for (mode = GOMP_DIM_WORKER; mode <= GOMP_DIM_VECTOR; mode++) | |
3435 | if (GOMP_DIM_MASK (mode) & skip_mask) | |
3436 | { | |
3437 | rtx_code_label *label = gen_label_rtx (); | |
3438 | rtx pred = cfun->machine->axis_predicate[mode - GOMP_DIM_WORKER]; | |
3439 | ||
3440 | if (!pred) | |
3441 | { | |
3442 | pred = gen_reg_rtx (BImode); | |
3443 | cfun->machine->axis_predicate[mode - GOMP_DIM_WORKER] = pred; | |
3444 | } | |
3445 | ||
3446 | rtx br; | |
3447 | if (mode == GOMP_DIM_VECTOR) | |
3448 | br = gen_br_true (pred, label); | |
3449 | else | |
3450 | br = gen_br_true_uni (pred, label); | |
3451 | emit_insn_before (br, head); | |
3452 | ||
3453 | LABEL_NUSES (label)++; | |
3454 | if (tail_branch) | |
3455 | before = emit_label_before (label, before); | |
3456 | else | |
3457 | emit_label_after (label, tail); | |
3458 | } | |
3459 | ||
3460 | /* Now deal with propagating the branch condition. */ | |
3461 | if (cond_branch) | |
3462 | { | |
3463 | rtx pvar = XEXP (XEXP (cond_branch, 0), 0); | |
3464 | ||
3465 | if (GOMP_DIM_MASK (GOMP_DIM_VECTOR) == mask) | |
3466 | { | |
3467 | /* Vector mode only, do a shuffle. */ | |
3468 | emit_insn_before (nvptx_gen_vcast (pvar), tail); | |
3469 | } | |
3470 | else | |
3471 | { | |
3472 | /* Includes worker mode, do spill & fill. By construction | |
3473 | we should never have worker mode only. */ | |
3474 | wcast_data_t data; | |
3475 | ||
3476 | data.base = worker_bcast_sym; | |
3477 | data.ptr = 0; | |
3478 | ||
3479 | if (worker_bcast_size < GET_MODE_SIZE (SImode)) | |
3480 | worker_bcast_size = GET_MODE_SIZE (SImode); | |
3481 | ||
3482 | data.offset = 0; | |
3483 | emit_insn_before (nvptx_gen_wcast (pvar, PM_read, 0, &data), | |
3484 | before); | |
3485 | /* Barrier so other workers can see the write. */ | |
3486 | emit_insn_before (nvptx_wsync (false), tail); | |
3487 | data.offset = 0; | |
3488 | emit_insn_before (nvptx_gen_wcast (pvar, PM_write, 0, &data), tail); | |
3489 | /* This barrier is needed to avoid worker zero clobbering | |
3490 | the broadcast buffer before all the other workers have | |
3491 | had a chance to read this instance of it. */ | |
3492 | emit_insn_before (nvptx_wsync (true), tail); | |
3493 | } | |
3494 | ||
3495 | extract_insn (tail); | |
3496 | rtx unsp = gen_rtx_UNSPEC (BImode, gen_rtvec (1, pvar), | |
3497 | UNSPEC_BR_UNIFIED); | |
3498 | validate_change (tail, recog_data.operand_loc[0], unsp, false); | |
3499 | } | |
3500 | } | |
3501 | ||
3502 | /* PAR is a parallel that is being skipped in its entirety according to | |
3503 | MASK. Treat this as skipping a superblock starting at forked | |
3504 | and ending at joining. */ | |
3505 | ||
3506 | static void | |
3507 | nvptx_skip_par (unsigned mask, parallel *par) | |
3508 | { | |
3509 | basic_block tail = par->join_block; | |
3510 | gcc_assert (tail->preds->length () == 1); | |
3511 | ||
3512 | basic_block pre_tail = (*tail->preds)[0]->src; | |
3513 | gcc_assert (pre_tail->succs->length () == 1); | |
3514 | ||
3515 | nvptx_single (mask, par->forked_block, pre_tail); | |
3516 | } | |
3517 | ||
dba619f3 NS |
3518 | /* If PAR has a single inner parallel and PAR itself only contains |
3519 | empty entry and exit blocks, swallow the inner PAR. */ | |
3520 | ||
3521 | static void | |
3522 | nvptx_optimize_inner (parallel *par) | |
3523 | { | |
3524 | parallel *inner = par->inner; | |
3525 | ||
3526 | /* We mustn't be the outer dummy par. */ | |
3527 | if (!par->mask) | |
3528 | return; | |
3529 | ||
3530 | /* We must have a single inner par. */ | |
3531 | if (!inner || inner->next) | |
3532 | return; | |
3533 | ||
3534 | /* We must only contain 2 blocks ourselves -- the head and tail of | |
3535 | the inner par. */ | |
3536 | if (par->blocks.length () != 2) | |
3537 | return; | |
3538 | ||
3539 | /* We must be disjoint partitioning. As we only have vector and | |
3540 | worker partitioning, this is sufficient to guarantee the pars | |
3541 | have adjacent partitioning. */ | |
3542 | if ((par->mask & inner->mask) & (GOMP_DIM_MASK (GOMP_DIM_MAX) - 1)) | |
3543 | /* This indicates malformed code generation. */ | |
3544 | return; | |
3545 | ||
3546 | /* The outer forked insn should be immediately followed by the inner | |
3547 | fork insn. */ | |
3548 | rtx_insn *forked = par->forked_insn; | |
3549 | rtx_insn *fork = BB_END (par->forked_block); | |
3550 | ||
3551 | if (NEXT_INSN (forked) != fork) | |
3552 | return; | |
3553 | gcc_checking_assert (recog_memoized (fork) == CODE_FOR_nvptx_fork); | |
3554 | ||
3555 | /* The outer joining insn must immediately follow the inner join | |
3556 | insn. */ | |
3557 | rtx_insn *joining = par->joining_insn; | |
3558 | rtx_insn *join = inner->join_insn; | |
3559 | if (NEXT_INSN (join) != joining) | |
3560 | return; | |
3561 | ||
3562 | /* Preconditions met. Swallow the inner par. */ | |
3563 | if (dump_file) | |
3564 | fprintf (dump_file, "Merging loop %x [%d,%d] into %x [%d,%d]\n", | |
3565 | inner->mask, inner->forked_block->index, | |
3566 | inner->join_block->index, | |
3567 | par->mask, par->forked_block->index, par->join_block->index); | |
3568 | ||
3569 | par->mask |= inner->mask & (GOMP_DIM_MASK (GOMP_DIM_MAX) - 1); | |
3570 | ||
3571 | par->blocks.reserve (inner->blocks.length ()); | |
3572 | while (inner->blocks.length ()) | |
3573 | par->blocks.quick_push (inner->blocks.pop ()); | |
3574 | ||
3575 | par->inner = inner->inner; | |
3576 | inner->inner = NULL; | |
3577 | ||
3578 | delete inner; | |
3579 | } | |
3580 | ||
d88cd9c4 NS |
3581 | /* Process the parallel PAR and all its contained |
3582 | parallels. We do everything but the neutering. Return mask of | |
3583 | partitioned modes used within this parallel. */ | |
3584 | ||
3585 | static unsigned | |
3586 | nvptx_process_pars (parallel *par) | |
3587 | { | |
dba619f3 NS |
3588 | if (nvptx_optimize) |
3589 | nvptx_optimize_inner (par); | |
3590 | ||
d88cd9c4 NS |
3591 | unsigned inner_mask = par->mask; |
3592 | ||
3593 | /* Do the inner parallels first. */ | |
3594 | if (par->inner) | |
3595 | { | |
3596 | par->inner_mask = nvptx_process_pars (par->inner); | |
3597 | inner_mask |= par->inner_mask; | |
3598 | } | |
3599 | ||
3600 | if (par->mask & GOMP_DIM_MASK (GOMP_DIM_MAX)) | |
3601 | /* No propagation needed for a call. */; | |
5d306e55 | 3602 | else if (par->mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)) |
d88cd9c4 NS |
3603 | { |
3604 | nvptx_wpropagate (false, par->forked_block, par->forked_insn); | |
3605 | nvptx_wpropagate (true, par->forked_block, par->fork_insn); | |
3606 | /* Insert begin and end synchronizations. */ | |
3607 | emit_insn_after (nvptx_wsync (false), par->forked_insn); | |
3608 | emit_insn_before (nvptx_wsync (true), par->joining_insn); | |
3609 | } | |
3610 | else if (par->mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR)) | |
3611 | nvptx_vpropagate (par->forked_block, par->forked_insn); | |
3612 | ||
3613 | /* Now do siblings. */ | |
3614 | if (par->next) | |
3615 | inner_mask |= nvptx_process_pars (par->next); | |
3616 | return inner_mask; | |
3617 | } | |
3618 | ||
3619 | /* Neuter the parallel described by PAR. We recurse in depth-first | |
3620 | order. MODES are the partitioning of the execution and OUTER is | |
3621 | the partitioning of the parallels we are contained in. */ | |
3622 | ||
3623 | static void | |
3624 | nvptx_neuter_pars (parallel *par, unsigned modes, unsigned outer) | |
3625 | { | |
3626 | unsigned me = (par->mask | |
3627 | & (GOMP_DIM_MASK (GOMP_DIM_WORKER) | |
3628 | | GOMP_DIM_MASK (GOMP_DIM_VECTOR))); | |
3629 | unsigned skip_mask = 0, neuter_mask = 0; | |
3630 | ||
3631 | if (par->inner) | |
3632 | nvptx_neuter_pars (par->inner, modes, outer | me); | |
3633 | ||
3634 | for (unsigned mode = GOMP_DIM_WORKER; mode <= GOMP_DIM_VECTOR; mode++) | |
3635 | { | |
3636 | if ((outer | me) & GOMP_DIM_MASK (mode)) | |
3637 | {} /* Mode is partitioned: no neutering. */ | |
3638 | else if (!(modes & GOMP_DIM_MASK (mode))) | |
5d306e55 | 3639 | {} /* Mode is not used: nothing to do. */ |
d88cd9c4 NS |
3640 | else if (par->inner_mask & GOMP_DIM_MASK (mode) |
3641 | || !par->forked_insn) | |
3642 | /* Partitioned in inner parallels, or we're not a partitioned | |
3643 | at all: neuter individual blocks. */ | |
3644 | neuter_mask |= GOMP_DIM_MASK (mode); | |
3645 | else if (!par->parent || !par->parent->forked_insn | |
3646 | || par->parent->inner_mask & GOMP_DIM_MASK (mode)) | |
3647 | /* Parent isn't a parallel or contains this paralleling: skip | |
3648 | parallel at this level. */ | |
3649 | skip_mask |= GOMP_DIM_MASK (mode); | |
3650 | else | |
3651 | {} /* Parent will skip this parallel itself. */ | |
3652 | } | |
3653 | ||
3654 | if (neuter_mask) | |
3655 | { | |
912442c2 | 3656 | int ix, len; |
d88cd9c4 | 3657 | |
912442c2 NS |
3658 | if (nvptx_optimize) |
3659 | { | |
3660 | /* Neuter whole SESE regions. */ | |
3661 | bb_pair_vec_t regions; | |
3662 | ||
3663 | nvptx_find_sese (par->blocks, regions); | |
3664 | len = regions.length (); | |
3665 | for (ix = 0; ix != len; ix++) | |
3666 | { | |
3667 | basic_block from = regions[ix].first; | |
3668 | basic_block to = regions[ix].second; | |
3669 | ||
3670 | if (from) | |
3671 | nvptx_single (neuter_mask, from, to); | |
3672 | else | |
3673 | gcc_assert (!to); | |
3674 | } | |
3675 | } | |
3676 | else | |
d88cd9c4 | 3677 | { |
912442c2 NS |
3678 | /* Neuter each BB individually. */ |
3679 | len = par->blocks.length (); | |
3680 | for (ix = 0; ix != len; ix++) | |
3681 | { | |
3682 | basic_block block = par->blocks[ix]; | |
d88cd9c4 | 3683 | |
912442c2 NS |
3684 | nvptx_single (neuter_mask, block, block); |
3685 | } | |
d88cd9c4 NS |
3686 | } |
3687 | } | |
3688 | ||
3689 | if (skip_mask) | |
3690 | nvptx_skip_par (skip_mask, par); | |
3691 | ||
3692 | if (par->next) | |
3693 | nvptx_neuter_pars (par->next, modes, outer); | |
3694 | } | |
3695 | ||
517665b3 | 3696 | /* PTX-specific reorganization |
d88cd9c4 | 3697 | - Split blocks at fork and join instructions |
c38f0d8c NS |
3698 | - Compute live registers |
3699 | - Mark now-unused registers, so function begin doesn't declare | |
517665b3 | 3700 | unused registers. |
d88cd9c4 NS |
3701 | - Insert state propagation when entering partitioned mode |
3702 | - Insert neutering instructions when in single mode | |
c38f0d8c | 3703 | - Replace subregs with suitable sequences. |
517665b3 NS |
3704 | */ |
3705 | ||
3706 | static void | |
3707 | nvptx_reorg (void) | |
3708 | { | |
517665b3 NS |
3709 | /* We are freeing block_for_insn in the toplev to keep compatibility |
3710 | with old MDEP_REORGS that are not CFG based. Recompute it now. */ | |
3711 | compute_bb_for_insn (); | |
3712 | ||
3713 | thread_prologue_and_epilogue_insns (); | |
3714 | ||
d88cd9c4 NS |
3715 | /* Split blocks and record interesting unspecs. */ |
3716 | bb_insn_map_t bb_insn_map; | |
3717 | ||
3718 | nvptx_split_blocks (&bb_insn_map); | |
3719 | ||
c38f0d8c | 3720 | /* Compute live regs */ |
517665b3 NS |
3721 | df_clear_flags (DF_LR_RUN_DCE); |
3722 | df_set_flags (DF_NO_INSN_RESCAN | DF_NO_HARD_REGS); | |
d88cd9c4 NS |
3723 | df_live_add_problem (); |
3724 | df_live_set_all_dirty (); | |
517665b3 | 3725 | df_analyze (); |
738f2522 BS |
3726 | regstat_init_n_sets_and_refs (); |
3727 | ||
d88cd9c4 NS |
3728 | if (dump_file) |
3729 | df_dump (dump_file); | |
3730 | ||
517665b3 | 3731 | /* Mark unused regs as unused. */ |
d88cd9c4 | 3732 | int max_regs = max_reg_num (); |
44c068ae | 3733 | for (int i = LAST_VIRTUAL_REGISTER + 1; i < max_regs; i++) |
738f2522 BS |
3734 | if (REG_N_SETS (i) == 0 && REG_N_REFS (i) == 0) |
3735 | regno_reg_rtx[i] = const0_rtx; | |
517665b3 | 3736 | |
d88cd9c4 NS |
3737 | /* Determine launch dimensions of the function. If it is not an |
3738 | offloaded function (i.e. this is a regular compiler), the | |
3739 | function has no neutering. */ | |
3740 | tree attr = get_oacc_fn_attrib (current_function_decl); | |
3741 | if (attr) | |
3742 | { | |
3743 | /* If we determined this mask before RTL expansion, we could | |
3744 | elide emission of some levels of forks and joins. */ | |
3745 | unsigned mask = 0; | |
3746 | tree dims = TREE_VALUE (attr); | |
3747 | unsigned ix; | |
3748 | ||
3749 | for (ix = 0; ix != GOMP_DIM_MAX; ix++, dims = TREE_CHAIN (dims)) | |
3750 | { | |
3751 | int size = TREE_INT_CST_LOW (TREE_VALUE (dims)); | |
3752 | tree allowed = TREE_PURPOSE (dims); | |
3753 | ||
3754 | if (size != 1 && !(allowed && integer_zerop (allowed))) | |
3755 | mask |= GOMP_DIM_MASK (ix); | |
3756 | } | |
3757 | /* If there is worker neutering, there must be vector | |
3758 | neutering. Otherwise the hardware will fail. */ | |
3759 | gcc_assert (!(mask & GOMP_DIM_MASK (GOMP_DIM_WORKER)) | |
3760 | || (mask & GOMP_DIM_MASK (GOMP_DIM_VECTOR))); | |
3761 | ||
3762 | /* Discover & process partitioned regions. */ | |
3763 | parallel *pars = nvptx_discover_pars (&bb_insn_map); | |
3764 | nvptx_process_pars (pars); | |
3765 | nvptx_neuter_pars (pars, mask, 0); | |
3766 | delete pars; | |
3767 | } | |
3768 | ||
517665b3 | 3769 | /* Replace subregs. */ |
c03b0416 | 3770 | nvptx_reorg_subreg (); |
517665b3 | 3771 | |
738f2522 | 3772 | regstat_free_n_sets_and_refs (); |
517665b3 NS |
3773 | |
3774 | df_finish_pass (true); | |
738f2522 BS |
3775 | } |
3776 | \f | |
3777 | /* Handle a "kernel" attribute; arguments as in | |
3778 | struct attribute_spec.handler. */ | |
3779 | ||
3780 | static tree | |
3781 | nvptx_handle_kernel_attribute (tree *node, tree name, tree ARG_UNUSED (args), | |
3782 | int ARG_UNUSED (flags), bool *no_add_attrs) | |
3783 | { | |
3784 | tree decl = *node; | |
3785 | ||
3786 | if (TREE_CODE (decl) != FUNCTION_DECL) | |
3787 | { | |
3788 | error ("%qE attribute only applies to functions", name); | |
3789 | *no_add_attrs = true; | |
3790 | } | |
b49e35a9 | 3791 | else if (!VOID_TYPE_P (TREE_TYPE (TREE_TYPE (decl)))) |
738f2522 BS |
3792 | { |
3793 | error ("%qE attribute requires a void return type", name); | |
3794 | *no_add_attrs = true; | |
3795 | } | |
3796 | ||
3797 | return NULL_TREE; | |
3798 | } | |
3799 | ||
3800 | /* Table of valid machine attributes. */ | |
3801 | static const struct attribute_spec nvptx_attribute_table[] = | |
3802 | { | |
3803 | /* { name, min_len, max_len, decl_req, type_req, fn_type_req, handler, | |
3804 | affects_type_identity } */ | |
3805 | { "kernel", 0, 0, true, false, false, nvptx_handle_kernel_attribute, false }, | |
3806 | { NULL, 0, 0, false, false, false, NULL, false } | |
3807 | }; | |
3808 | \f | |
3809 | /* Limit vector alignments to BIGGEST_ALIGNMENT. */ | |
3810 | ||
3811 | static HOST_WIDE_INT | |
3812 | nvptx_vector_alignment (const_tree type) | |
3813 | { | |
3814 | HOST_WIDE_INT align = tree_to_shwi (TYPE_SIZE (type)); | |
3815 | ||
3816 | return MIN (align, BIGGEST_ALIGNMENT); | |
3817 | } | |
d88cd9c4 NS |
3818 | |
3819 | /* Indicate that INSN cannot be duplicated. */ | |
3820 | ||
3821 | static bool | |
3822 | nvptx_cannot_copy_insn_p (rtx_insn *insn) | |
3823 | { | |
3824 | switch (recog_memoized (insn)) | |
3825 | { | |
3826 | case CODE_FOR_nvptx_shufflesi: | |
3827 | case CODE_FOR_nvptx_shufflesf: | |
3828 | case CODE_FOR_nvptx_barsync: | |
3829 | case CODE_FOR_nvptx_fork: | |
3830 | case CODE_FOR_nvptx_forked: | |
3831 | case CODE_FOR_nvptx_joining: | |
3832 | case CODE_FOR_nvptx_join: | |
3833 | return true; | |
3834 | default: | |
3835 | return false; | |
3836 | } | |
3837 | } | |
a794bd20 NS |
3838 | |
3839 | /* Section anchors do not work. Initialization for flag_section_anchor | |
3840 | probes the existence of the anchoring target hooks and prevents | |
3841 | anchoring if they don't exist. However, we may be being used with | |
3842 | a host-side compiler that does support anchoring, and hence see | |
3843 | the anchor flag set (as it's not recalculated). So provide an | |
3844 | implementation denying anchoring. */ | |
3845 | ||
3846 | static bool | |
3847 | nvptx_use_anchors_for_symbol_p (const_rtx ARG_UNUSED (a)) | |
3848 | { | |
3849 | return false; | |
3850 | } | |
738f2522 | 3851 | \f |
1f83528e TS |
3852 | /* Record a symbol for mkoffload to enter into the mapping table. */ |
3853 | ||
3854 | static void | |
3855 | nvptx_record_offload_symbol (tree decl) | |
3856 | { | |
3e32ee19 NS |
3857 | switch (TREE_CODE (decl)) |
3858 | { | |
3859 | case VAR_DECL: | |
3860 | fprintf (asm_out_file, "//:VAR_MAP \"%s\"\n", | |
3861 | IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl))); | |
3862 | break; | |
3863 | ||
3864 | case FUNCTION_DECL: | |
3865 | { | |
3866 | tree attr = get_oacc_fn_attrib (decl); | |
5d306e55 | 3867 | tree dims = TREE_VALUE (attr); |
3e32ee19 NS |
3868 | unsigned ix; |
3869 | ||
3e32ee19 NS |
3870 | fprintf (asm_out_file, "//:FUNC_MAP \"%s\"", |
3871 | IDENTIFIER_POINTER (DECL_ASSEMBLER_NAME (decl))); | |
3872 | ||
5d306e55 | 3873 | for (ix = 0; ix != GOMP_DIM_MAX; ix++, dims = TREE_CHAIN (dims)) |
3e32ee19 | 3874 | { |
5d306e55 | 3875 | int size = TREE_INT_CST_LOW (TREE_VALUE (dims)); |
3e32ee19 | 3876 | |
5d306e55 | 3877 | gcc_assert (!TREE_PURPOSE (dims)); |
3e32ee19 NS |
3878 | fprintf (asm_out_file, ", %#x", size); |
3879 | } | |
d2d47a28 | 3880 | |
3e32ee19 NS |
3881 | fprintf (asm_out_file, "\n"); |
3882 | } | |
3883 | break; | |
d2d47a28 | 3884 | |
3e32ee19 NS |
3885 | default: |
3886 | gcc_unreachable (); | |
3887 | } | |
1f83528e TS |
3888 | } |
3889 | ||
738f2522 BS |
3890 | /* Implement TARGET_ASM_FILE_START. Write the kinds of things ptxas expects |
3891 | at the start of a file. */ | |
3892 | ||
3893 | static void | |
3894 | nvptx_file_start (void) | |
3895 | { | |
3896 | fputs ("// BEGIN PREAMBLE\n", asm_out_file); | |
3897 | fputs ("\t.version\t3.1\n", asm_out_file); | |
3898 | fputs ("\t.target\tsm_30\n", asm_out_file); | |
3899 | fprintf (asm_out_file, "\t.address_size %d\n", GET_MODE_BITSIZE (Pmode)); | |
3900 | fputs ("// END PREAMBLE\n", asm_out_file); | |
3901 | } | |
3902 | ||
15ab6f00 NS |
3903 | /* Emit a declaration for a worker-level buffer in .shared memory. */ |
3904 | ||
3905 | static void | |
3906 | write_worker_buffer (FILE *file, rtx sym, unsigned align, unsigned size) | |
3907 | { | |
3908 | const char *name = XSTR (sym, 0); | |
3909 | ||
3910 | write_var_marker (file, true, false, name); | |
3911 | fprintf (file, ".shared .align %d .u8 %s[%d];\n", | |
3912 | align, name, size); | |
3913 | } | |
3914 | ||
ecf6e535 BS |
3915 | /* Write out the function declarations we've collected and declare storage |
3916 | for the broadcast buffer. */ | |
738f2522 BS |
3917 | |
3918 | static void | |
3919 | nvptx_file_end (void) | |
3920 | { | |
f3dba894 TS |
3921 | hash_table<tree_hasher>::iterator iter; |
3922 | tree decl; | |
3923 | FOR_EACH_HASH_TABLE_ELEMENT (*needed_fndecls_htab, decl, tree, iter) | |
00e52418 | 3924 | nvptx_record_fndecl (decl); |
738f2522 | 3925 | fputs (func_decls.str().c_str(), asm_out_file); |
d88cd9c4 NS |
3926 | |
3927 | if (worker_bcast_size) | |
15ab6f00 NS |
3928 | write_worker_buffer (asm_out_file, worker_bcast_sym, |
3929 | worker_bcast_align, worker_bcast_size); | |
f3552158 NS |
3930 | |
3931 | if (worker_red_size) | |
15ab6f00 NS |
3932 | write_worker_buffer (asm_out_file, worker_red_sym, |
3933 | worker_red_align, worker_red_size); | |
f3552158 NS |
3934 | } |
3935 | ||
3936 | /* Expander for the shuffle builtins. */ | |
3937 | ||
3938 | static rtx | |
3939 | nvptx_expand_shuffle (tree exp, rtx target, machine_mode mode, int ignore) | |
3940 | { | |
3941 | if (ignore) | |
3942 | return target; | |
3943 | ||
3944 | rtx src = expand_expr (CALL_EXPR_ARG (exp, 0), | |
3945 | NULL_RTX, mode, EXPAND_NORMAL); | |
3946 | if (!REG_P (src)) | |
3947 | src = copy_to_mode_reg (mode, src); | |
3948 | ||
3949 | rtx idx = expand_expr (CALL_EXPR_ARG (exp, 1), | |
3950 | NULL_RTX, SImode, EXPAND_NORMAL); | |
3951 | rtx op = expand_expr (CALL_EXPR_ARG (exp, 2), | |
3952 | NULL_RTX, SImode, EXPAND_NORMAL); | |
3953 | ||
3954 | if (!REG_P (idx) && GET_CODE (idx) != CONST_INT) | |
3955 | idx = copy_to_mode_reg (SImode, idx); | |
3956 | ||
59263259 NS |
3957 | rtx pat = nvptx_gen_shuffle (target, src, idx, |
3958 | (nvptx_shuffle_kind) INTVAL (op)); | |
f3552158 NS |
3959 | if (pat) |
3960 | emit_insn (pat); | |
3961 | ||
3962 | return target; | |
3963 | } | |
3964 | ||
3965 | /* Worker reduction address expander. */ | |
3966 | ||
3967 | static rtx | |
3968 | nvptx_expand_worker_addr (tree exp, rtx target, | |
3969 | machine_mode ARG_UNUSED (mode), int ignore) | |
3970 | { | |
3971 | if (ignore) | |
3972 | return target; | |
3973 | ||
3974 | unsigned align = TREE_INT_CST_LOW (CALL_EXPR_ARG (exp, 2)); | |
3975 | if (align > worker_red_align) | |
3976 | worker_red_align = align; | |
3977 | ||
3978 | unsigned offset = TREE_INT_CST_LOW (CALL_EXPR_ARG (exp, 0)); | |
3979 | unsigned size = TREE_INT_CST_LOW (CALL_EXPR_ARG (exp, 1)); | |
3980 | if (size + offset > worker_red_size) | |
3981 | worker_red_size = size + offset; | |
3982 | ||
9a863523 | 3983 | rtx addr = worker_red_sym; |
f3552158 | 3984 | if (offset) |
9a863523 NS |
3985 | { |
3986 | addr = gen_rtx_PLUS (Pmode, addr, GEN_INT (offset)); | |
3987 | addr = gen_rtx_CONST (Pmode, addr); | |
3988 | } | |
f3552158 | 3989 | |
9a863523 | 3990 | emit_move_insn (target, addr); |
f3552158 NS |
3991 | |
3992 | return target; | |
3993 | } | |
3994 | ||
3995 | /* Expand the CMP_SWAP PTX builtins. We have our own versions that do | |
3996 | not require taking the address of any object, other than the memory | |
3997 | cell being operated on. */ | |
3998 | ||
3999 | static rtx | |
4000 | nvptx_expand_cmp_swap (tree exp, rtx target, | |
4001 | machine_mode ARG_UNUSED (m), int ARG_UNUSED (ignore)) | |
4002 | { | |
4003 | machine_mode mode = TYPE_MODE (TREE_TYPE (exp)); | |
4004 | ||
4005 | if (!target) | |
4006 | target = gen_reg_rtx (mode); | |
4007 | ||
4008 | rtx mem = expand_expr (CALL_EXPR_ARG (exp, 0), | |
4009 | NULL_RTX, Pmode, EXPAND_NORMAL); | |
4010 | rtx cmp = expand_expr (CALL_EXPR_ARG (exp, 1), | |
4011 | NULL_RTX, mode, EXPAND_NORMAL); | |
4012 | rtx src = expand_expr (CALL_EXPR_ARG (exp, 2), | |
4013 | NULL_RTX, mode, EXPAND_NORMAL); | |
4014 | rtx pat; | |
4015 | ||
4016 | mem = gen_rtx_MEM (mode, mem); | |
4017 | if (!REG_P (cmp)) | |
4018 | cmp = copy_to_mode_reg (mode, cmp); | |
4019 | if (!REG_P (src)) | |
4020 | src = copy_to_mode_reg (mode, src); | |
4021 | ||
4022 | if (mode == SImode) | |
4023 | pat = gen_atomic_compare_and_swapsi_1 (target, mem, cmp, src, const0_rtx); | |
4024 | else | |
4025 | pat = gen_atomic_compare_and_swapdi_1 (target, mem, cmp, src, const0_rtx); | |
4026 | ||
4027 | emit_insn (pat); | |
4028 | ||
4029 | return target; | |
4030 | } | |
4031 | ||
4032 | ||
4033 | /* Codes for all the NVPTX builtins. */ | |
4034 | enum nvptx_builtins | |
4035 | { | |
4036 | NVPTX_BUILTIN_SHUFFLE, | |
4037 | NVPTX_BUILTIN_SHUFFLELL, | |
4038 | NVPTX_BUILTIN_WORKER_ADDR, | |
4039 | NVPTX_BUILTIN_CMP_SWAP, | |
4040 | NVPTX_BUILTIN_CMP_SWAPLL, | |
4041 | NVPTX_BUILTIN_MAX | |
4042 | }; | |
4043 | ||
4044 | static GTY(()) tree nvptx_builtin_decls[NVPTX_BUILTIN_MAX]; | |
4045 | ||
4046 | /* Return the NVPTX builtin for CODE. */ | |
4047 | ||
4048 | static tree | |
4049 | nvptx_builtin_decl (unsigned code, bool ARG_UNUSED (initialize_p)) | |
4050 | { | |
4051 | if (code >= NVPTX_BUILTIN_MAX) | |
4052 | return error_mark_node; | |
4053 | ||
4054 | return nvptx_builtin_decls[code]; | |
4055 | } | |
4056 | ||
4057 | /* Set up all builtin functions for this target. */ | |
4058 | ||
4059 | static void | |
4060 | nvptx_init_builtins (void) | |
4061 | { | |
4062 | #define DEF(ID, NAME, T) \ | |
4063 | (nvptx_builtin_decls[NVPTX_BUILTIN_ ## ID] \ | |
4064 | = add_builtin_function ("__builtin_nvptx_" NAME, \ | |
4065 | build_function_type_list T, \ | |
4066 | NVPTX_BUILTIN_ ## ID, BUILT_IN_MD, NULL, NULL)) | |
4067 | #define ST sizetype | |
4068 | #define UINT unsigned_type_node | |
4069 | #define LLUINT long_long_unsigned_type_node | |
4070 | #define PTRVOID ptr_type_node | |
4071 | ||
4072 | DEF (SHUFFLE, "shuffle", (UINT, UINT, UINT, UINT, NULL_TREE)); | |
4073 | DEF (SHUFFLELL, "shufflell", (LLUINT, LLUINT, UINT, UINT, NULL_TREE)); | |
4074 | DEF (WORKER_ADDR, "worker_addr", | |
4075 | (PTRVOID, ST, UINT, UINT, NULL_TREE)); | |
4076 | DEF (CMP_SWAP, "cmp_swap", (UINT, PTRVOID, UINT, UINT, NULL_TREE)); | |
4077 | DEF (CMP_SWAPLL, "cmp_swapll", (LLUINT, PTRVOID, LLUINT, LLUINT, NULL_TREE)); | |
4078 | ||
4079 | #undef DEF | |
4080 | #undef ST | |
4081 | #undef UINT | |
4082 | #undef LLUINT | |
4083 | #undef PTRVOID | |
4084 | } | |
4085 | ||
4086 | /* Expand an expression EXP that calls a built-in function, | |
4087 | with result going to TARGET if that's convenient | |
4088 | (and in mode MODE if that's convenient). | |
4089 | SUBTARGET may be used as the target for computing one of EXP's operands. | |
4090 | IGNORE is nonzero if the value is to be ignored. */ | |
4091 | ||
4092 | static rtx | |
4093 | nvptx_expand_builtin (tree exp, rtx target, rtx ARG_UNUSED (subtarget), | |
4094 | machine_mode mode, int ignore) | |
4095 | { | |
4096 | tree fndecl = TREE_OPERAND (CALL_EXPR_FN (exp), 0); | |
4097 | switch (DECL_FUNCTION_CODE (fndecl)) | |
4098 | { | |
4099 | case NVPTX_BUILTIN_SHUFFLE: | |
4100 | case NVPTX_BUILTIN_SHUFFLELL: | |
4101 | return nvptx_expand_shuffle (exp, target, mode, ignore); | |
4102 | ||
4103 | case NVPTX_BUILTIN_WORKER_ADDR: | |
4104 | return nvptx_expand_worker_addr (exp, target, mode, ignore); | |
4105 | ||
4106 | case NVPTX_BUILTIN_CMP_SWAP: | |
4107 | case NVPTX_BUILTIN_CMP_SWAPLL: | |
4108 | return nvptx_expand_cmp_swap (exp, target, mode, ignore); | |
4109 | ||
4110 | default: gcc_unreachable (); | |
4111 | } | |
738f2522 BS |
4112 | } |
4113 | \f | |
f3552158 NS |
4114 | /* Define dimension sizes for known hardware. */ |
4115 | #define PTX_VECTOR_LENGTH 32 | |
4116 | #define PTX_WORKER_LENGTH 32 | |
b6adbb9f | 4117 | #define PTX_GANG_DEFAULT 32 |
f3552158 | 4118 | |
94829f87 NS |
4119 | /* Validate compute dimensions of an OpenACC offload or routine, fill |
4120 | in non-unity defaults. FN_LEVEL indicates the level at which a | |
b6adbb9f NS |
4121 | routine might spawn a loop. It is negative for non-routines. If |
4122 | DECL is null, we are validating the default dimensions. */ | |
94829f87 NS |
4123 | |
4124 | static bool | |
5d306e55 | 4125 | nvptx_goacc_validate_dims (tree decl, int dims[], int fn_level) |
94829f87 NS |
4126 | { |
4127 | bool changed = false; | |
4128 | ||
ccc8282b | 4129 | /* The vector size must be 32, unless this is a SEQ routine. */ |
b6adbb9f NS |
4130 | if (fn_level <= GOMP_DIM_VECTOR && fn_level >= -1 |
4131 | && dims[GOMP_DIM_VECTOR] >= 0 | |
ccc8282b NS |
4132 | && dims[GOMP_DIM_VECTOR] != PTX_VECTOR_LENGTH) |
4133 | { | |
b6adbb9f NS |
4134 | if (fn_level < 0 && dims[GOMP_DIM_VECTOR] >= 0) |
4135 | warning_at (decl ? DECL_SOURCE_LOCATION (decl) : UNKNOWN_LOCATION, 0, | |
ccc8282b NS |
4136 | dims[GOMP_DIM_VECTOR] |
4137 | ? "using vector_length (%d), ignoring %d" | |
4138 | : "using vector_length (%d), ignoring runtime setting", | |
4139 | PTX_VECTOR_LENGTH, dims[GOMP_DIM_VECTOR]); | |
4140 | dims[GOMP_DIM_VECTOR] = PTX_VECTOR_LENGTH; | |
4141 | changed = true; | |
4142 | } | |
4143 | ||
4144 | /* Check the num workers is not too large. */ | |
4145 | if (dims[GOMP_DIM_WORKER] > PTX_WORKER_LENGTH) | |
4146 | { | |
b6adbb9f | 4147 | warning_at (decl ? DECL_SOURCE_LOCATION (decl) : UNKNOWN_LOCATION, 0, |
ccc8282b NS |
4148 | "using num_workers (%d), ignoring %d", |
4149 | PTX_WORKER_LENGTH, dims[GOMP_DIM_WORKER]); | |
4150 | dims[GOMP_DIM_WORKER] = PTX_WORKER_LENGTH; | |
4151 | changed = true; | |
4152 | } | |
94829f87 | 4153 | |
b6adbb9f NS |
4154 | if (!decl) |
4155 | { | |
4156 | dims[GOMP_DIM_VECTOR] = PTX_VECTOR_LENGTH; | |
4157 | if (dims[GOMP_DIM_WORKER] < 0) | |
4158 | dims[GOMP_DIM_WORKER] = PTX_WORKER_LENGTH; | |
4159 | if (dims[GOMP_DIM_GANG] < 0) | |
4160 | dims[GOMP_DIM_GANG] = PTX_GANG_DEFAULT; | |
4161 | changed = true; | |
4162 | } | |
4163 | ||
94829f87 NS |
4164 | return changed; |
4165 | } | |
d88cd9c4 | 4166 | |
bd751975 NS |
4167 | /* Return maximum dimension size, or zero for unbounded. */ |
4168 | ||
4169 | static int | |
4170 | nvptx_dim_limit (int axis) | |
4171 | { | |
4172 | switch (axis) | |
4173 | { | |
4174 | case GOMP_DIM_WORKER: | |
4175 | return PTX_WORKER_LENGTH; | |
4176 | ||
4177 | case GOMP_DIM_VECTOR: | |
4178 | return PTX_VECTOR_LENGTH; | |
4179 | ||
4180 | default: | |
4181 | break; | |
4182 | } | |
4183 | return 0; | |
4184 | } | |
4185 | ||
d88cd9c4 NS |
4186 | /* Determine whether fork & joins are needed. */ |
4187 | ||
4188 | static bool | |
4189 | nvptx_goacc_fork_join (gcall *call, const int dims[], | |
4190 | bool ARG_UNUSED (is_fork)) | |
4191 | { | |
4192 | tree arg = gimple_call_arg (call, 2); | |
4193 | unsigned axis = TREE_INT_CST_LOW (arg); | |
4194 | ||
4195 | /* We only care about worker and vector partitioning. */ | |
4196 | if (axis < GOMP_DIM_WORKER) | |
4197 | return false; | |
4198 | ||
4199 | /* If the size is 1, there's no partitioning. */ | |
4200 | if (dims[axis] == 1) | |
4201 | return false; | |
4202 | ||
4203 | return true; | |
4204 | } | |
4205 | ||
f3552158 NS |
4206 | /* Generate a PTX builtin function call that returns the address in |
4207 | the worker reduction buffer at OFFSET. TYPE is the type of the | |
4208 | data at that location. */ | |
4209 | ||
4210 | static tree | |
4211 | nvptx_get_worker_red_addr (tree type, tree offset) | |
4212 | { | |
4213 | machine_mode mode = TYPE_MODE (type); | |
4214 | tree fndecl = nvptx_builtin_decl (NVPTX_BUILTIN_WORKER_ADDR, true); | |
4215 | tree size = build_int_cst (unsigned_type_node, GET_MODE_SIZE (mode)); | |
4216 | tree align = build_int_cst (unsigned_type_node, | |
4217 | GET_MODE_ALIGNMENT (mode) / BITS_PER_UNIT); | |
4218 | tree call = build_call_expr (fndecl, 3, offset, size, align); | |
4219 | ||
4220 | return fold_convert (build_pointer_type (type), call); | |
4221 | } | |
4222 | ||
4223 | /* Emit a SHFL.DOWN using index SHFL of VAR into DEST_VAR. This function | |
4224 | will cast the variable if necessary. */ | |
4225 | ||
4226 | static void | |
4227 | nvptx_generate_vector_shuffle (location_t loc, | |
4228 | tree dest_var, tree var, unsigned shift, | |
4229 | gimple_seq *seq) | |
4230 | { | |
4231 | unsigned fn = NVPTX_BUILTIN_SHUFFLE; | |
4232 | tree_code code = NOP_EXPR; | |
dd3c1b14 NS |
4233 | tree arg_type = unsigned_type_node; |
4234 | tree var_type = TREE_TYPE (var); | |
4235 | tree dest_type = var_type; | |
f3552158 | 4236 | |
dd3c1b14 NS |
4237 | if (TREE_CODE (var_type) == COMPLEX_TYPE) |
4238 | var_type = TREE_TYPE (var_type); | |
4239 | ||
4240 | if (TREE_CODE (var_type) == REAL_TYPE) | |
f3552158 | 4241 | code = VIEW_CONVERT_EXPR; |
dd3c1b14 NS |
4242 | |
4243 | if (TYPE_SIZE (var_type) | |
4244 | == TYPE_SIZE (long_long_unsigned_type_node)) | |
f3552158 NS |
4245 | { |
4246 | fn = NVPTX_BUILTIN_SHUFFLELL; | |
dd3c1b14 | 4247 | arg_type = long_long_unsigned_type_node; |
f3552158 | 4248 | } |
dd3c1b14 | 4249 | |
f3552158 | 4250 | tree call = nvptx_builtin_decl (fn, true); |
dd3c1b14 NS |
4251 | tree bits = build_int_cst (unsigned_type_node, shift); |
4252 | tree kind = build_int_cst (unsigned_type_node, SHUFFLE_DOWN); | |
4253 | tree expr; | |
4254 | ||
4255 | if (var_type != dest_type) | |
4256 | { | |
4257 | /* Do real and imaginary parts separately. */ | |
4258 | tree real = fold_build1 (REALPART_EXPR, var_type, var); | |
4259 | real = fold_build1 (code, arg_type, real); | |
4260 | real = build_call_expr_loc (loc, call, 3, real, bits, kind); | |
4261 | real = fold_build1 (code, var_type, real); | |
f3552158 | 4262 | |
dd3c1b14 NS |
4263 | tree imag = fold_build1 (IMAGPART_EXPR, var_type, var); |
4264 | imag = fold_build1 (code, arg_type, imag); | |
4265 | imag = build_call_expr_loc (loc, call, 3, imag, bits, kind); | |
4266 | imag = fold_build1 (code, var_type, imag); | |
4267 | ||
4268 | expr = fold_build2 (COMPLEX_EXPR, dest_type, real, imag); | |
4269 | } | |
4270 | else | |
4271 | { | |
4272 | expr = fold_build1 (code, arg_type, var); | |
4273 | expr = build_call_expr_loc (loc, call, 3, expr, bits, kind); | |
4274 | expr = fold_build1 (code, dest_type, expr); | |
4275 | } | |
f3552158 | 4276 | |
dd3c1b14 | 4277 | gimplify_assign (dest_var, expr, seq); |
f3552158 NS |
4278 | } |
4279 | ||
33f47f42 NS |
4280 | /* Lazily generate the global lock var decl and return its address. */ |
4281 | ||
4282 | static tree | |
4283 | nvptx_global_lock_addr () | |
4284 | { | |
4285 | tree v = global_lock_var; | |
4286 | ||
4287 | if (!v) | |
4288 | { | |
4289 | tree name = get_identifier ("__reduction_lock"); | |
4290 | tree type = build_qualified_type (unsigned_type_node, | |
4291 | TYPE_QUAL_VOLATILE); | |
4292 | v = build_decl (BUILTINS_LOCATION, VAR_DECL, name, type); | |
4293 | global_lock_var = v; | |
4294 | DECL_ARTIFICIAL (v) = 1; | |
4295 | DECL_EXTERNAL (v) = 1; | |
4296 | TREE_STATIC (v) = 1; | |
4297 | TREE_PUBLIC (v) = 1; | |
4298 | TREE_USED (v) = 1; | |
4299 | mark_addressable (v); | |
4300 | mark_decl_referenced (v); | |
4301 | } | |
4302 | ||
4303 | return build_fold_addr_expr (v); | |
4304 | } | |
4305 | ||
4306 | /* Insert code to locklessly update *PTR with *PTR OP VAR just before | |
4307 | GSI. We use a lockless scheme for nearly all case, which looks | |
4308 | like: | |
4309 | actual = initval(OP); | |
4310 | do { | |
4311 | guess = actual; | |
4312 | write = guess OP myval; | |
4313 | actual = cmp&swap (ptr, guess, write) | |
4314 | } while (actual bit-different-to guess); | |
4315 | return write; | |
4316 | ||
4317 | This relies on a cmp&swap instruction, which is available for 32- | |
4318 | and 64-bit types. Larger types must use a locking scheme. */ | |
f3552158 NS |
4319 | |
4320 | static tree | |
4321 | nvptx_lockless_update (location_t loc, gimple_stmt_iterator *gsi, | |
4322 | tree ptr, tree var, tree_code op) | |
4323 | { | |
4324 | unsigned fn = NVPTX_BUILTIN_CMP_SWAP; | |
4325 | tree_code code = NOP_EXPR; | |
33f47f42 NS |
4326 | tree arg_type = unsigned_type_node; |
4327 | tree var_type = TREE_TYPE (var); | |
f3552158 | 4328 | |
33f47f42 NS |
4329 | if (TREE_CODE (var_type) == COMPLEX_TYPE |
4330 | || TREE_CODE (var_type) == REAL_TYPE) | |
f3552158 | 4331 | code = VIEW_CONVERT_EXPR; |
33f47f42 NS |
4332 | |
4333 | if (TYPE_SIZE (var_type) == TYPE_SIZE (long_long_unsigned_type_node)) | |
f3552158 | 4334 | { |
33f47f42 | 4335 | arg_type = long_long_unsigned_type_node; |
f3552158 | 4336 | fn = NVPTX_BUILTIN_CMP_SWAPLL; |
f3552158 NS |
4337 | } |
4338 | ||
33f47f42 NS |
4339 | tree swap_fn = nvptx_builtin_decl (fn, true); |
4340 | ||
f3552158 | 4341 | gimple_seq init_seq = NULL; |
33f47f42 NS |
4342 | tree init_var = make_ssa_name (arg_type); |
4343 | tree init_expr = omp_reduction_init_op (loc, op, var_type); | |
4344 | init_expr = fold_build1 (code, arg_type, init_expr); | |
f3552158 NS |
4345 | gimplify_assign (init_var, init_expr, &init_seq); |
4346 | gimple *init_end = gimple_seq_last (init_seq); | |
4347 | ||
4348 | gsi_insert_seq_before (gsi, init_seq, GSI_SAME_STMT); | |
4349 | ||
f3552158 NS |
4350 | /* Split the block just after the init stmts. */ |
4351 | basic_block pre_bb = gsi_bb (*gsi); | |
4352 | edge pre_edge = split_block (pre_bb, init_end); | |
4353 | basic_block loop_bb = pre_edge->dest; | |
4354 | pre_bb = pre_edge->src; | |
4355 | /* Reset the iterator. */ | |
4356 | *gsi = gsi_for_stmt (gsi_stmt (*gsi)); | |
4357 | ||
33f47f42 NS |
4358 | tree expect_var = make_ssa_name (arg_type); |
4359 | tree actual_var = make_ssa_name (arg_type); | |
4360 | tree write_var = make_ssa_name (arg_type); | |
4361 | ||
4362 | /* Build and insert the reduction calculation. */ | |
4363 | gimple_seq red_seq = NULL; | |
4364 | tree write_expr = fold_build1 (code, var_type, expect_var); | |
4365 | write_expr = fold_build2 (op, var_type, write_expr, var); | |
4366 | write_expr = fold_build1 (code, arg_type, write_expr); | |
4367 | gimplify_assign (write_var, write_expr, &red_seq); | |
4368 | ||
4369 | gsi_insert_seq_before (gsi, red_seq, GSI_SAME_STMT); | |
4370 | ||
4371 | /* Build & insert the cmp&swap sequence. */ | |
4372 | gimple_seq latch_seq = NULL; | |
4373 | tree swap_expr = build_call_expr_loc (loc, swap_fn, 3, | |
4374 | ptr, expect_var, write_var); | |
4375 | gimplify_assign (actual_var, swap_expr, &latch_seq); | |
4376 | ||
4377 | gcond *cond = gimple_build_cond (EQ_EXPR, actual_var, expect_var, | |
4378 | NULL_TREE, NULL_TREE); | |
4379 | gimple_seq_add_stmt (&latch_seq, cond); | |
4380 | ||
4381 | gimple *latch_end = gimple_seq_last (latch_seq); | |
4382 | gsi_insert_seq_before (gsi, latch_seq, GSI_SAME_STMT); | |
f3552158 | 4383 | |
33f47f42 NS |
4384 | /* Split the block just after the latch stmts. */ |
4385 | edge post_edge = split_block (loop_bb, latch_end); | |
f3552158 NS |
4386 | basic_block post_bb = post_edge->dest; |
4387 | loop_bb = post_edge->src; | |
4388 | *gsi = gsi_for_stmt (gsi_stmt (*gsi)); | |
4389 | ||
4390 | post_edge->flags ^= EDGE_TRUE_VALUE | EDGE_FALLTHRU; | |
4391 | edge loop_edge = make_edge (loop_bb, loop_bb, EDGE_FALSE_VALUE); | |
4392 | set_immediate_dominator (CDI_DOMINATORS, loop_bb, pre_bb); | |
4393 | set_immediate_dominator (CDI_DOMINATORS, post_bb, loop_bb); | |
4394 | ||
4395 | gphi *phi = create_phi_node (expect_var, loop_bb); | |
4396 | add_phi_arg (phi, init_var, pre_edge, loc); | |
4397 | add_phi_arg (phi, actual_var, loop_edge, loc); | |
4398 | ||
4399 | loop *loop = alloc_loop (); | |
4400 | loop->header = loop_bb; | |
4401 | loop->latch = loop_bb; | |
4402 | add_loop (loop, loop_bb->loop_father); | |
4403 | ||
33f47f42 NS |
4404 | return fold_build1 (code, var_type, write_var); |
4405 | } | |
4406 | ||
4407 | /* Insert code to lockfully update *PTR with *PTR OP VAR just before | |
4408 | GSI. This is necessary for types larger than 64 bits, where there | |
4409 | is no cmp&swap instruction to implement a lockless scheme. We use | |
4410 | a lock variable in global memory. | |
4411 | ||
4412 | while (cmp&swap (&lock_var, 0, 1)) | |
4413 | continue; | |
4414 | T accum = *ptr; | |
4415 | accum = accum OP var; | |
4416 | *ptr = accum; | |
4417 | cmp&swap (&lock_var, 1, 0); | |
4418 | return accum; | |
4419 | ||
4420 | A lock in global memory is necessary to force execution engine | |
4421 | descheduling and avoid resource starvation that can occur if the | |
4422 | lock is in .shared memory. */ | |
4423 | ||
4424 | static tree | |
4425 | nvptx_lockfull_update (location_t loc, gimple_stmt_iterator *gsi, | |
4426 | tree ptr, tree var, tree_code op) | |
4427 | { | |
4428 | tree var_type = TREE_TYPE (var); | |
4429 | tree swap_fn = nvptx_builtin_decl (NVPTX_BUILTIN_CMP_SWAP, true); | |
4430 | tree uns_unlocked = build_int_cst (unsigned_type_node, 0); | |
4431 | tree uns_locked = build_int_cst (unsigned_type_node, 1); | |
4432 | ||
4433 | /* Split the block just before the gsi. Insert a gimple nop to make | |
4434 | this easier. */ | |
4435 | gimple *nop = gimple_build_nop (); | |
4436 | gsi_insert_before (gsi, nop, GSI_SAME_STMT); | |
4437 | basic_block entry_bb = gsi_bb (*gsi); | |
4438 | edge entry_edge = split_block (entry_bb, nop); | |
4439 | basic_block lock_bb = entry_edge->dest; | |
4440 | /* Reset the iterator. */ | |
4441 | *gsi = gsi_for_stmt (gsi_stmt (*gsi)); | |
4442 | ||
4443 | /* Build and insert the locking sequence. */ | |
4444 | gimple_seq lock_seq = NULL; | |
4445 | tree lock_var = make_ssa_name (unsigned_type_node); | |
4446 | tree lock_expr = nvptx_global_lock_addr (); | |
4447 | lock_expr = build_call_expr_loc (loc, swap_fn, 3, lock_expr, | |
4448 | uns_unlocked, uns_locked); | |
4449 | gimplify_assign (lock_var, lock_expr, &lock_seq); | |
4450 | gcond *cond = gimple_build_cond (EQ_EXPR, lock_var, uns_unlocked, | |
4451 | NULL_TREE, NULL_TREE); | |
4452 | gimple_seq_add_stmt (&lock_seq, cond); | |
4453 | gimple *lock_end = gimple_seq_last (lock_seq); | |
4454 | gsi_insert_seq_before (gsi, lock_seq, GSI_SAME_STMT); | |
4455 | ||
4456 | /* Split the block just after the lock sequence. */ | |
4457 | edge locked_edge = split_block (lock_bb, lock_end); | |
4458 | basic_block update_bb = locked_edge->dest; | |
4459 | lock_bb = locked_edge->src; | |
4460 | *gsi = gsi_for_stmt (gsi_stmt (*gsi)); | |
4461 | ||
4462 | /* Create the lock loop ... */ | |
4463 | locked_edge->flags ^= EDGE_TRUE_VALUE | EDGE_FALLTHRU; | |
4464 | make_edge (lock_bb, lock_bb, EDGE_FALSE_VALUE); | |
4465 | set_immediate_dominator (CDI_DOMINATORS, lock_bb, entry_bb); | |
4466 | set_immediate_dominator (CDI_DOMINATORS, update_bb, lock_bb); | |
4467 | ||
4468 | /* ... and the loop structure. */ | |
4469 | loop *lock_loop = alloc_loop (); | |
4470 | lock_loop->header = lock_bb; | |
4471 | lock_loop->latch = lock_bb; | |
4472 | lock_loop->nb_iterations_estimate = 1; | |
4473 | lock_loop->any_estimate = true; | |
4474 | add_loop (lock_loop, entry_bb->loop_father); | |
4475 | ||
4476 | /* Build and insert the reduction calculation. */ | |
4477 | gimple_seq red_seq = NULL; | |
4478 | tree acc_in = make_ssa_name (var_type); | |
4479 | tree ref_in = build_simple_mem_ref (ptr); | |
4480 | TREE_THIS_VOLATILE (ref_in) = 1; | |
4481 | gimplify_assign (acc_in, ref_in, &red_seq); | |
4482 | ||
4483 | tree acc_out = make_ssa_name (var_type); | |
4484 | tree update_expr = fold_build2 (op, var_type, ref_in, var); | |
4485 | gimplify_assign (acc_out, update_expr, &red_seq); | |
4486 | ||
4487 | tree ref_out = build_simple_mem_ref (ptr); | |
4488 | TREE_THIS_VOLATILE (ref_out) = 1; | |
4489 | gimplify_assign (ref_out, acc_out, &red_seq); | |
4490 | ||
4491 | gsi_insert_seq_before (gsi, red_seq, GSI_SAME_STMT); | |
4492 | ||
4493 | /* Build & insert the unlock sequence. */ | |
4494 | gimple_seq unlock_seq = NULL; | |
4495 | tree unlock_expr = nvptx_global_lock_addr (); | |
4496 | unlock_expr = build_call_expr_loc (loc, swap_fn, 3, unlock_expr, | |
4497 | uns_locked, uns_unlocked); | |
4498 | gimplify_and_add (unlock_expr, &unlock_seq); | |
4499 | gsi_insert_seq_before (gsi, unlock_seq, GSI_SAME_STMT); | |
4500 | ||
4501 | return acc_out; | |
4502 | } | |
4503 | ||
4504 | /* Emit a sequence to update a reduction accumlator at *PTR with the | |
4505 | value held in VAR using operator OP. Return the updated value. | |
4506 | ||
4507 | TODO: optimize for atomic ops and indepedent complex ops. */ | |
4508 | ||
4509 | static tree | |
4510 | nvptx_reduction_update (location_t loc, gimple_stmt_iterator *gsi, | |
4511 | tree ptr, tree var, tree_code op) | |
4512 | { | |
4513 | tree type = TREE_TYPE (var); | |
4514 | tree size = TYPE_SIZE (type); | |
4515 | ||
4516 | if (size == TYPE_SIZE (unsigned_type_node) | |
4517 | || size == TYPE_SIZE (long_long_unsigned_type_node)) | |
4518 | return nvptx_lockless_update (loc, gsi, ptr, var, op); | |
4519 | else | |
4520 | return nvptx_lockfull_update (loc, gsi, ptr, var, op); | |
f3552158 NS |
4521 | } |
4522 | ||
4523 | /* NVPTX implementation of GOACC_REDUCTION_SETUP. */ | |
4524 | ||
4525 | static void | |
4526 | nvptx_goacc_reduction_setup (gcall *call) | |
4527 | { | |
4528 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
4529 | tree lhs = gimple_call_lhs (call); | |
4530 | tree var = gimple_call_arg (call, 2); | |
4531 | int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); | |
4532 | gimple_seq seq = NULL; | |
4533 | ||
4534 | push_gimplify_context (true); | |
4535 | ||
4536 | if (level != GOMP_DIM_GANG) | |
4537 | { | |
4538 | /* Copy the receiver object. */ | |
4539 | tree ref_to_res = gimple_call_arg (call, 1); | |
4540 | ||
4541 | if (!integer_zerop (ref_to_res)) | |
4542 | var = build_simple_mem_ref (ref_to_res); | |
4543 | } | |
4544 | ||
4545 | if (level == GOMP_DIM_WORKER) | |
4546 | { | |
4547 | /* Store incoming value to worker reduction buffer. */ | |
4548 | tree offset = gimple_call_arg (call, 5); | |
4549 | tree call = nvptx_get_worker_red_addr (TREE_TYPE (var), offset); | |
4550 | tree ptr = make_ssa_name (TREE_TYPE (call)); | |
4551 | ||
4552 | gimplify_assign (ptr, call, &seq); | |
4553 | tree ref = build_simple_mem_ref (ptr); | |
4554 | TREE_THIS_VOLATILE (ref) = 1; | |
4555 | gimplify_assign (ref, var, &seq); | |
4556 | } | |
4557 | ||
4558 | if (lhs) | |
4559 | gimplify_assign (lhs, var, &seq); | |
4560 | ||
4561 | pop_gimplify_context (NULL); | |
4562 | gsi_replace_with_seq (&gsi, seq, true); | |
4563 | } | |
4564 | ||
4565 | /* NVPTX implementation of GOACC_REDUCTION_INIT. */ | |
4566 | ||
4567 | static void | |
4568 | nvptx_goacc_reduction_init (gcall *call) | |
4569 | { | |
4570 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
4571 | tree lhs = gimple_call_lhs (call); | |
4572 | tree var = gimple_call_arg (call, 2); | |
4573 | int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); | |
4574 | enum tree_code rcode | |
4575 | = (enum tree_code)TREE_INT_CST_LOW (gimple_call_arg (call, 4)); | |
4576 | tree init = omp_reduction_init_op (gimple_location (call), rcode, | |
4577 | TREE_TYPE (var)); | |
4578 | gimple_seq seq = NULL; | |
4579 | ||
4580 | push_gimplify_context (true); | |
4581 | ||
4582 | if (level == GOMP_DIM_VECTOR) | |
4583 | { | |
4584 | /* Initialize vector-non-zeroes to INIT_VAL (OP). */ | |
4585 | tree tid = make_ssa_name (integer_type_node); | |
4586 | tree dim_vector = gimple_call_arg (call, 3); | |
4587 | gimple *tid_call = gimple_build_call_internal (IFN_GOACC_DIM_POS, 1, | |
4588 | dim_vector); | |
4589 | gimple *cond_stmt = gimple_build_cond (NE_EXPR, tid, integer_zero_node, | |
4590 | NULL_TREE, NULL_TREE); | |
4591 | ||
4592 | gimple_call_set_lhs (tid_call, tid); | |
4593 | gimple_seq_add_stmt (&seq, tid_call); | |
4594 | gimple_seq_add_stmt (&seq, cond_stmt); | |
4595 | ||
4596 | /* Split the block just after the call. */ | |
4597 | edge init_edge = split_block (gsi_bb (gsi), call); | |
4598 | basic_block init_bb = init_edge->dest; | |
4599 | basic_block call_bb = init_edge->src; | |
4600 | ||
4601 | /* Fixup flags from call_bb to init_bb. */ | |
4602 | init_edge->flags ^= EDGE_FALLTHRU | EDGE_TRUE_VALUE; | |
4603 | ||
4604 | /* Set the initialization stmts. */ | |
4605 | gimple_seq init_seq = NULL; | |
4606 | tree init_var = make_ssa_name (TREE_TYPE (var)); | |
4607 | gimplify_assign (init_var, init, &init_seq); | |
4608 | gsi = gsi_start_bb (init_bb); | |
4609 | gsi_insert_seq_before (&gsi, init_seq, GSI_SAME_STMT); | |
4610 | ||
4611 | /* Split block just after the init stmt. */ | |
4612 | gsi_prev (&gsi); | |
4613 | edge inited_edge = split_block (gsi_bb (gsi), gsi_stmt (gsi)); | |
4614 | basic_block dst_bb = inited_edge->dest; | |
4615 | ||
4616 | /* Create false edge from call_bb to dst_bb. */ | |
4617 | edge nop_edge = make_edge (call_bb, dst_bb, EDGE_FALSE_VALUE); | |
4618 | ||
4619 | /* Create phi node in dst block. */ | |
4620 | gphi *phi = create_phi_node (lhs, dst_bb); | |
4621 | add_phi_arg (phi, init_var, inited_edge, gimple_location (call)); | |
4622 | add_phi_arg (phi, var, nop_edge, gimple_location (call)); | |
4623 | ||
4624 | /* Reset dominator of dst bb. */ | |
4625 | set_immediate_dominator (CDI_DOMINATORS, dst_bb, call_bb); | |
4626 | ||
4627 | /* Reset the gsi. */ | |
4628 | gsi = gsi_for_stmt (call); | |
4629 | } | |
4630 | else | |
4631 | { | |
4632 | if (level == GOMP_DIM_GANG) | |
4633 | { | |
4634 | /* If there's no receiver object, propagate the incoming VAR. */ | |
4635 | tree ref_to_res = gimple_call_arg (call, 1); | |
4636 | if (integer_zerop (ref_to_res)) | |
4637 | init = var; | |
4638 | } | |
4639 | ||
4640 | gimplify_assign (lhs, init, &seq); | |
4641 | } | |
4642 | ||
4643 | pop_gimplify_context (NULL); | |
4644 | gsi_replace_with_seq (&gsi, seq, true); | |
4645 | } | |
4646 | ||
4647 | /* NVPTX implementation of GOACC_REDUCTION_FINI. */ | |
4648 | ||
4649 | static void | |
4650 | nvptx_goacc_reduction_fini (gcall *call) | |
4651 | { | |
4652 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
4653 | tree lhs = gimple_call_lhs (call); | |
4654 | tree ref_to_res = gimple_call_arg (call, 1); | |
4655 | tree var = gimple_call_arg (call, 2); | |
4656 | int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); | |
4657 | enum tree_code op | |
4658 | = (enum tree_code)TREE_INT_CST_LOW (gimple_call_arg (call, 4)); | |
4659 | gimple_seq seq = NULL; | |
4660 | tree r = NULL_TREE;; | |
4661 | ||
4662 | push_gimplify_context (true); | |
4663 | ||
4664 | if (level == GOMP_DIM_VECTOR) | |
4665 | { | |
4666 | /* Emit binary shuffle tree. TODO. Emit this as an actual loop, | |
4667 | but that requires a method of emitting a unified jump at the | |
4668 | gimple level. */ | |
4669 | for (int shfl = PTX_VECTOR_LENGTH / 2; shfl > 0; shfl = shfl >> 1) | |
4670 | { | |
4671 | tree other_var = make_ssa_name (TREE_TYPE (var)); | |
4672 | nvptx_generate_vector_shuffle (gimple_location (call), | |
4673 | other_var, var, shfl, &seq); | |
4674 | ||
4675 | r = make_ssa_name (TREE_TYPE (var)); | |
4676 | gimplify_assign (r, fold_build2 (op, TREE_TYPE (var), | |
4677 | var, other_var), &seq); | |
4678 | var = r; | |
4679 | } | |
4680 | } | |
4681 | else | |
4682 | { | |
4683 | tree accum = NULL_TREE; | |
4684 | ||
4685 | if (level == GOMP_DIM_WORKER) | |
4686 | { | |
4687 | /* Get reduction buffer address. */ | |
4688 | tree offset = gimple_call_arg (call, 5); | |
4689 | tree call = nvptx_get_worker_red_addr (TREE_TYPE (var), offset); | |
4690 | tree ptr = make_ssa_name (TREE_TYPE (call)); | |
4691 | ||
4692 | gimplify_assign (ptr, call, &seq); | |
4693 | accum = ptr; | |
4694 | } | |
4695 | else if (integer_zerop (ref_to_res)) | |
4696 | r = var; | |
4697 | else | |
4698 | accum = ref_to_res; | |
4699 | ||
4700 | if (accum) | |
4701 | { | |
33f47f42 | 4702 | /* UPDATE the accumulator. */ |
f3552158 NS |
4703 | gsi_insert_seq_before (&gsi, seq, GSI_SAME_STMT); |
4704 | seq = NULL; | |
33f47f42 NS |
4705 | r = nvptx_reduction_update (gimple_location (call), &gsi, |
4706 | accum, var, op); | |
f3552158 NS |
4707 | } |
4708 | } | |
4709 | ||
4710 | if (lhs) | |
4711 | gimplify_assign (lhs, r, &seq); | |
4712 | pop_gimplify_context (NULL); | |
4713 | ||
4714 | gsi_replace_with_seq (&gsi, seq, true); | |
4715 | } | |
4716 | ||
4717 | /* NVPTX implementation of GOACC_REDUCTION_TEARDOWN. */ | |
4718 | ||
4719 | static void | |
4720 | nvptx_goacc_reduction_teardown (gcall *call) | |
4721 | { | |
4722 | gimple_stmt_iterator gsi = gsi_for_stmt (call); | |
4723 | tree lhs = gimple_call_lhs (call); | |
4724 | tree var = gimple_call_arg (call, 2); | |
4725 | int level = TREE_INT_CST_LOW (gimple_call_arg (call, 3)); | |
4726 | gimple_seq seq = NULL; | |
4727 | ||
4728 | push_gimplify_context (true); | |
4729 | if (level == GOMP_DIM_WORKER) | |
4730 | { | |
4731 | /* Read the worker reduction buffer. */ | |
4732 | tree offset = gimple_call_arg (call, 5); | |
4733 | tree call = nvptx_get_worker_red_addr(TREE_TYPE (var), offset); | |
4734 | tree ptr = make_ssa_name (TREE_TYPE (call)); | |
4735 | ||
4736 | gimplify_assign (ptr, call, &seq); | |
4737 | var = build_simple_mem_ref (ptr); | |
4738 | TREE_THIS_VOLATILE (var) = 1; | |
4739 | } | |
4740 | ||
4741 | if (level != GOMP_DIM_GANG) | |
4742 | { | |
4743 | /* Write to the receiver object. */ | |
4744 | tree ref_to_res = gimple_call_arg (call, 1); | |
4745 | ||
4746 | if (!integer_zerop (ref_to_res)) | |
4747 | gimplify_assign (build_simple_mem_ref (ref_to_res), var, &seq); | |
4748 | } | |
4749 | ||
4750 | if (lhs) | |
4751 | gimplify_assign (lhs, var, &seq); | |
4752 | ||
4753 | pop_gimplify_context (NULL); | |
4754 | ||
4755 | gsi_replace_with_seq (&gsi, seq, true); | |
4756 | } | |
4757 | ||
4758 | /* NVPTX reduction expander. */ | |
4759 | ||
5563d5c0 | 4760 | static void |
f3552158 NS |
4761 | nvptx_goacc_reduction (gcall *call) |
4762 | { | |
4763 | unsigned code = (unsigned)TREE_INT_CST_LOW (gimple_call_arg (call, 0)); | |
4764 | ||
4765 | switch (code) | |
4766 | { | |
4767 | case IFN_GOACC_REDUCTION_SETUP: | |
4768 | nvptx_goacc_reduction_setup (call); | |
4769 | break; | |
4770 | ||
4771 | case IFN_GOACC_REDUCTION_INIT: | |
4772 | nvptx_goacc_reduction_init (call); | |
4773 | break; | |
4774 | ||
4775 | case IFN_GOACC_REDUCTION_FINI: | |
4776 | nvptx_goacc_reduction_fini (call); | |
4777 | break; | |
4778 | ||
4779 | case IFN_GOACC_REDUCTION_TEARDOWN: | |
4780 | nvptx_goacc_reduction_teardown (call); | |
4781 | break; | |
4782 | ||
4783 | default: | |
4784 | gcc_unreachable (); | |
4785 | } | |
4786 | } | |
4787 | ||
738f2522 BS |
4788 | #undef TARGET_OPTION_OVERRIDE |
4789 | #define TARGET_OPTION_OVERRIDE nvptx_option_override | |
4790 | ||
4791 | #undef TARGET_ATTRIBUTE_TABLE | |
4792 | #define TARGET_ATTRIBUTE_TABLE nvptx_attribute_table | |
4793 | ||
4794 | #undef TARGET_LEGITIMATE_ADDRESS_P | |
4795 | #define TARGET_LEGITIMATE_ADDRESS_P nvptx_legitimate_address_p | |
4796 | ||
4797 | #undef TARGET_PROMOTE_FUNCTION_MODE | |
4798 | #define TARGET_PROMOTE_FUNCTION_MODE nvptx_promote_function_mode | |
4799 | ||
4800 | #undef TARGET_FUNCTION_ARG | |
4801 | #define TARGET_FUNCTION_ARG nvptx_function_arg | |
4802 | #undef TARGET_FUNCTION_INCOMING_ARG | |
4803 | #define TARGET_FUNCTION_INCOMING_ARG nvptx_function_incoming_arg | |
4804 | #undef TARGET_FUNCTION_ARG_ADVANCE | |
4805 | #define TARGET_FUNCTION_ARG_ADVANCE nvptx_function_arg_advance | |
738f2522 BS |
4806 | #undef TARGET_PASS_BY_REFERENCE |
4807 | #define TARGET_PASS_BY_REFERENCE nvptx_pass_by_reference | |
4808 | #undef TARGET_FUNCTION_VALUE_REGNO_P | |
4809 | #define TARGET_FUNCTION_VALUE_REGNO_P nvptx_function_value_regno_p | |
4810 | #undef TARGET_FUNCTION_VALUE | |
4811 | #define TARGET_FUNCTION_VALUE nvptx_function_value | |
4812 | #undef TARGET_LIBCALL_VALUE | |
4813 | #define TARGET_LIBCALL_VALUE nvptx_libcall_value | |
4814 | #undef TARGET_FUNCTION_OK_FOR_SIBCALL | |
4815 | #define TARGET_FUNCTION_OK_FOR_SIBCALL nvptx_function_ok_for_sibcall | |
18c05628 NS |
4816 | #undef TARGET_GET_DRAP_RTX |
4817 | #define TARGET_GET_DRAP_RTX nvptx_get_drap_rtx | |
738f2522 BS |
4818 | #undef TARGET_SPLIT_COMPLEX_ARG |
4819 | #define TARGET_SPLIT_COMPLEX_ARG hook_bool_const_tree_true | |
4820 | #undef TARGET_RETURN_IN_MEMORY | |
4821 | #define TARGET_RETURN_IN_MEMORY nvptx_return_in_memory | |
4822 | #undef TARGET_OMIT_STRUCT_RETURN_REG | |
4823 | #define TARGET_OMIT_STRUCT_RETURN_REG true | |
4824 | #undef TARGET_STRICT_ARGUMENT_NAMING | |
4825 | #define TARGET_STRICT_ARGUMENT_NAMING nvptx_strict_argument_naming | |
738f2522 BS |
4826 | #undef TARGET_CALL_ARGS |
4827 | #define TARGET_CALL_ARGS nvptx_call_args | |
4828 | #undef TARGET_END_CALL_ARGS | |
4829 | #define TARGET_END_CALL_ARGS nvptx_end_call_args | |
4830 | ||
4831 | #undef TARGET_ASM_FILE_START | |
4832 | #define TARGET_ASM_FILE_START nvptx_file_start | |
4833 | #undef TARGET_ASM_FILE_END | |
4834 | #define TARGET_ASM_FILE_END nvptx_file_end | |
4835 | #undef TARGET_ASM_GLOBALIZE_LABEL | |
4836 | #define TARGET_ASM_GLOBALIZE_LABEL nvptx_globalize_label | |
4837 | #undef TARGET_ASM_ASSEMBLE_UNDEFINED_DECL | |
4838 | #define TARGET_ASM_ASSEMBLE_UNDEFINED_DECL nvptx_assemble_undefined_decl | |
4839 | #undef TARGET_PRINT_OPERAND | |
4840 | #define TARGET_PRINT_OPERAND nvptx_print_operand | |
4841 | #undef TARGET_PRINT_OPERAND_ADDRESS | |
4842 | #define TARGET_PRINT_OPERAND_ADDRESS nvptx_print_operand_address | |
4843 | #undef TARGET_PRINT_OPERAND_PUNCT_VALID_P | |
4844 | #define TARGET_PRINT_OPERAND_PUNCT_VALID_P nvptx_print_operand_punct_valid_p | |
4845 | #undef TARGET_ASM_INTEGER | |
4846 | #define TARGET_ASM_INTEGER nvptx_assemble_integer | |
4847 | #undef TARGET_ASM_DECL_END | |
4848 | #define TARGET_ASM_DECL_END nvptx_assemble_decl_end | |
4849 | #undef TARGET_ASM_DECLARE_CONSTANT_NAME | |
4850 | #define TARGET_ASM_DECLARE_CONSTANT_NAME nvptx_asm_declare_constant_name | |
4851 | #undef TARGET_USE_BLOCKS_FOR_CONSTANT_P | |
4852 | #define TARGET_USE_BLOCKS_FOR_CONSTANT_P hook_bool_mode_const_rtx_true | |
4853 | #undef TARGET_ASM_NEED_VAR_DECL_BEFORE_USE | |
4854 | #define TARGET_ASM_NEED_VAR_DECL_BEFORE_USE true | |
4855 | ||
4856 | #undef TARGET_MACHINE_DEPENDENT_REORG | |
4857 | #define TARGET_MACHINE_DEPENDENT_REORG nvptx_reorg | |
4858 | #undef TARGET_NO_REGISTER_ALLOCATION | |
4859 | #define TARGET_NO_REGISTER_ALLOCATION true | |
4860 | ||
9a863523 NS |
4861 | #undef TARGET_ENCODE_SECTION_INFO |
4862 | #define TARGET_ENCODE_SECTION_INFO nvptx_encode_section_info | |
1f83528e TS |
4863 | #undef TARGET_RECORD_OFFLOAD_SYMBOL |
4864 | #define TARGET_RECORD_OFFLOAD_SYMBOL nvptx_record_offload_symbol | |
4865 | ||
738f2522 BS |
4866 | #undef TARGET_VECTOR_ALIGNMENT |
4867 | #define TARGET_VECTOR_ALIGNMENT nvptx_vector_alignment | |
4868 | ||
d88cd9c4 NS |
4869 | #undef TARGET_CANNOT_COPY_INSN_P |
4870 | #define TARGET_CANNOT_COPY_INSN_P nvptx_cannot_copy_insn_p | |
4871 | ||
a794bd20 NS |
4872 | #undef TARGET_USE_ANCHORS_FOR_SYMBOL_P |
4873 | #define TARGET_USE_ANCHORS_FOR_SYMBOL_P nvptx_use_anchors_for_symbol_p | |
4874 | ||
f3552158 NS |
4875 | #undef TARGET_INIT_BUILTINS |
4876 | #define TARGET_INIT_BUILTINS nvptx_init_builtins | |
4877 | #undef TARGET_EXPAND_BUILTIN | |
4878 | #define TARGET_EXPAND_BUILTIN nvptx_expand_builtin | |
4879 | #undef TARGET_BUILTIN_DECL | |
4880 | #define TARGET_BUILTIN_DECL nvptx_builtin_decl | |
4881 | ||
94829f87 NS |
4882 | #undef TARGET_GOACC_VALIDATE_DIMS |
4883 | #define TARGET_GOACC_VALIDATE_DIMS nvptx_goacc_validate_dims | |
4884 | ||
bd751975 NS |
4885 | #undef TARGET_GOACC_DIM_LIMIT |
4886 | #define TARGET_GOACC_DIM_LIMIT nvptx_dim_limit | |
4887 | ||
d88cd9c4 NS |
4888 | #undef TARGET_GOACC_FORK_JOIN |
4889 | #define TARGET_GOACC_FORK_JOIN nvptx_goacc_fork_join | |
4890 | ||
f3552158 NS |
4891 | #undef TARGET_GOACC_REDUCTION |
4892 | #define TARGET_GOACC_REDUCTION nvptx_goacc_reduction | |
4893 | ||
738f2522 BS |
4894 | struct gcc_target targetm = TARGET_INITIALIZER; |
4895 | ||
4896 | #include "gt-nvptx.h" |