]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - contrib/llvm/tools/clang/include/clang/Basic/BuiltinsNVPTX.def
Merge clang 7.0.1 and several follow-up changes
[FreeBSD/FreeBSD.git] / contrib / llvm / tools / clang / include / clang / Basic / BuiltinsNVPTX.def
1 //===--- BuiltinsPTX.def - PTX Builtin function database ----*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file defines the PTX-specific builtin function database.  Users of
11 // this file must define the BUILTIN macro to make use of this information.
12 //
13 //===----------------------------------------------------------------------===//
14
15 // The format of this database matches clang/Basic/Builtins.def.
16
17 #if defined(BUILTIN) && !defined(TARGET_BUILTIN)
18 #   define TARGET_BUILTIN(ID, TYPE, ATTRS, FEATURE) BUILTIN(ID, TYPE, ATTRS)
19 #endif
20
21 #pragma push_macro("SM_70")
22 #define SM_70 "sm_70|sm_71"
23 #pragma push_macro("SM_60")
24 #define SM_60 "sm_60|sm_61|sm_62|" SM_70
25
26 #pragma push_macro("PTX61")
27 #define PTX61 "ptx61"
28 #pragma push_macro("PTX60")
29 #define PTX60 "ptx60|" PTX61
30
31 #pragma push_macro("AND")
32 #define AND(a, b) a "," b
33
34 // Special Registers
35
36 BUILTIN(__nvvm_read_ptx_sreg_tid_x, "i", "nc")
37 BUILTIN(__nvvm_read_ptx_sreg_tid_y, "i", "nc")
38 BUILTIN(__nvvm_read_ptx_sreg_tid_z, "i", "nc")
39 BUILTIN(__nvvm_read_ptx_sreg_tid_w, "i", "nc")
40
41 BUILTIN(__nvvm_read_ptx_sreg_ntid_x, "i", "nc")
42 BUILTIN(__nvvm_read_ptx_sreg_ntid_y, "i", "nc")
43 BUILTIN(__nvvm_read_ptx_sreg_ntid_z, "i", "nc")
44 BUILTIN(__nvvm_read_ptx_sreg_ntid_w, "i", "nc")
45
46 BUILTIN(__nvvm_read_ptx_sreg_ctaid_x, "i", "nc")
47 BUILTIN(__nvvm_read_ptx_sreg_ctaid_y, "i", "nc")
48 BUILTIN(__nvvm_read_ptx_sreg_ctaid_z, "i", "nc")
49 BUILTIN(__nvvm_read_ptx_sreg_ctaid_w, "i", "nc")
50
51 BUILTIN(__nvvm_read_ptx_sreg_nctaid_x, "i", "nc")
52 BUILTIN(__nvvm_read_ptx_sreg_nctaid_y, "i", "nc")
53 BUILTIN(__nvvm_read_ptx_sreg_nctaid_z, "i", "nc")
54 BUILTIN(__nvvm_read_ptx_sreg_nctaid_w, "i", "nc")
55
56 BUILTIN(__nvvm_read_ptx_sreg_laneid, "i", "nc")
57 BUILTIN(__nvvm_read_ptx_sreg_warpid, "i", "nc")
58 BUILTIN(__nvvm_read_ptx_sreg_nwarpid, "i", "nc")
59
60 BUILTIN(__nvvm_read_ptx_sreg_smid, "i", "nc")
61 BUILTIN(__nvvm_read_ptx_sreg_nsmid, "i", "nc")
62 BUILTIN(__nvvm_read_ptx_sreg_gridid, "i", "nc")
63
64 BUILTIN(__nvvm_read_ptx_sreg_lanemask_eq, "i", "nc")
65 BUILTIN(__nvvm_read_ptx_sreg_lanemask_le, "i", "nc")
66 BUILTIN(__nvvm_read_ptx_sreg_lanemask_lt, "i", "nc")
67 BUILTIN(__nvvm_read_ptx_sreg_lanemask_ge, "i", "nc")
68 BUILTIN(__nvvm_read_ptx_sreg_lanemask_gt, "i", "nc")
69
70 BUILTIN(__nvvm_read_ptx_sreg_clock, "i", "n")
71 BUILTIN(__nvvm_read_ptx_sreg_clock64, "LLi", "n")
72
73 BUILTIN(__nvvm_read_ptx_sreg_pm0, "i", "n")
74 BUILTIN(__nvvm_read_ptx_sreg_pm1, "i", "n")
75 BUILTIN(__nvvm_read_ptx_sreg_pm2, "i", "n")
76 BUILTIN(__nvvm_read_ptx_sreg_pm3, "i", "n")
77
78 // MISC
79
80 BUILTIN(__nvvm_prmt, "UiUiUiUi", "")
81
82 // Min Max
83
84 BUILTIN(__nvvm_fmax_ftz_f, "fff",  "")
85 BUILTIN(__nvvm_fmax_f, "fff",  "")
86 BUILTIN(__nvvm_fmin_ftz_f, "fff",  "")
87 BUILTIN(__nvvm_fmin_f, "fff",  "")
88
89 BUILTIN(__nvvm_fmax_d, "ddd", "")
90 BUILTIN(__nvvm_fmin_d, "ddd", "")
91
92 // Multiplication
93
94 BUILTIN(__nvvm_mulhi_i, "iii", "")
95 BUILTIN(__nvvm_mulhi_ui, "UiUiUi", "")
96 BUILTIN(__nvvm_mulhi_ll, "LLiLLiLLi", "")
97 BUILTIN(__nvvm_mulhi_ull, "ULLiULLiULLi", "")
98
99 BUILTIN(__nvvm_mul_rn_ftz_f,  "fff", "")
100 BUILTIN(__nvvm_mul_rn_f,  "fff", "")
101 BUILTIN(__nvvm_mul_rz_ftz_f,  "fff", "")
102 BUILTIN(__nvvm_mul_rz_f,  "fff", "")
103 BUILTIN(__nvvm_mul_rm_ftz_f,  "fff", "")
104 BUILTIN(__nvvm_mul_rm_f,  "fff", "")
105 BUILTIN(__nvvm_mul_rp_ftz_f,  "fff", "")
106 BUILTIN(__nvvm_mul_rp_f,  "fff", "")
107
108 BUILTIN(__nvvm_mul_rn_d,  "ddd", "")
109 BUILTIN(__nvvm_mul_rz_d,  "ddd", "")
110 BUILTIN(__nvvm_mul_rm_d,  "ddd", "")
111 BUILTIN(__nvvm_mul_rp_d,  "ddd", "")
112
113 BUILTIN(__nvvm_mul24_i,  "iii", "")
114 BUILTIN(__nvvm_mul24_ui,  "UiUiUi", "")
115
116 // Div
117
118 BUILTIN(__nvvm_div_approx_ftz_f,  "fff", "")
119 BUILTIN(__nvvm_div_approx_f,  "fff", "")
120
121 BUILTIN(__nvvm_div_rn_ftz_f,  "fff", "")
122 BUILTIN(__nvvm_div_rn_f,  "fff", "")
123 BUILTIN(__nvvm_div_rz_ftz_f,  "fff", "")
124 BUILTIN(__nvvm_div_rz_f,  "fff", "")
125 BUILTIN(__nvvm_div_rm_ftz_f,  "fff", "")
126 BUILTIN(__nvvm_div_rm_f,  "fff", "")
127 BUILTIN(__nvvm_div_rp_ftz_f,  "fff", "")
128 BUILTIN(__nvvm_div_rp_f,  "fff", "")
129
130 BUILTIN(__nvvm_div_rn_d,  "ddd", "")
131 BUILTIN(__nvvm_div_rz_d,  "ddd", "")
132 BUILTIN(__nvvm_div_rm_d,  "ddd", "")
133 BUILTIN(__nvvm_div_rp_d,  "ddd", "")
134
135 // Sad
136
137 BUILTIN(__nvvm_sad_i, "iiii", "")
138 BUILTIN(__nvvm_sad_ui, "UiUiUiUi", "")
139
140 // Floor, Ceil
141
142 BUILTIN(__nvvm_floor_ftz_f, "ff", "")
143 BUILTIN(__nvvm_floor_f, "ff", "")
144 BUILTIN(__nvvm_floor_d, "dd", "")
145
146 BUILTIN(__nvvm_ceil_ftz_f, "ff", "")
147 BUILTIN(__nvvm_ceil_f, "ff", "")
148 BUILTIN(__nvvm_ceil_d, "dd", "")
149
150 // Abs
151
152 BUILTIN(__nvvm_fabs_ftz_f, "ff", "")
153 BUILTIN(__nvvm_fabs_f, "ff", "")
154 BUILTIN(__nvvm_fabs_d, "dd", "")
155
156 // Round
157
158 BUILTIN(__nvvm_round_ftz_f, "ff", "")
159 BUILTIN(__nvvm_round_f, "ff", "")
160 BUILTIN(__nvvm_round_d, "dd", "")
161
162 // Trunc
163
164 BUILTIN(__nvvm_trunc_ftz_f, "ff", "")
165 BUILTIN(__nvvm_trunc_f, "ff", "")
166 BUILTIN(__nvvm_trunc_d, "dd", "")
167
168 // Saturate
169
170 BUILTIN(__nvvm_saturate_ftz_f, "ff", "")
171 BUILTIN(__nvvm_saturate_f, "ff", "")
172 BUILTIN(__nvvm_saturate_d, "dd", "")
173
174 // Exp2, Log2
175
176 BUILTIN(__nvvm_ex2_approx_ftz_f, "ff", "")
177 BUILTIN(__nvvm_ex2_approx_f, "ff", "")
178 BUILTIN(__nvvm_ex2_approx_d, "dd", "")
179
180 BUILTIN(__nvvm_lg2_approx_ftz_f, "ff", "")
181 BUILTIN(__nvvm_lg2_approx_f, "ff", "")
182 BUILTIN(__nvvm_lg2_approx_d, "dd", "")
183
184 // Sin, Cos
185
186 BUILTIN(__nvvm_sin_approx_ftz_f, "ff", "")
187 BUILTIN(__nvvm_sin_approx_f, "ff", "")
188
189 BUILTIN(__nvvm_cos_approx_ftz_f, "ff", "")
190 BUILTIN(__nvvm_cos_approx_f, "ff", "")
191
192 // Fma
193
194 BUILTIN(__nvvm_fma_rn_ftz_f, "ffff", "")
195 BUILTIN(__nvvm_fma_rn_f, "ffff", "")
196 BUILTIN(__nvvm_fma_rz_ftz_f, "ffff", "")
197 BUILTIN(__nvvm_fma_rz_f, "ffff", "")
198 BUILTIN(__nvvm_fma_rm_ftz_f, "ffff", "")
199 BUILTIN(__nvvm_fma_rm_f, "ffff", "")
200 BUILTIN(__nvvm_fma_rp_ftz_f, "ffff", "")
201 BUILTIN(__nvvm_fma_rp_f, "ffff", "")
202 BUILTIN(__nvvm_fma_rn_d, "dddd", "")
203 BUILTIN(__nvvm_fma_rz_d, "dddd", "")
204 BUILTIN(__nvvm_fma_rm_d, "dddd", "")
205 BUILTIN(__nvvm_fma_rp_d, "dddd", "")
206
207 // Rcp
208
209 BUILTIN(__nvvm_rcp_rn_ftz_f, "ff", "")
210 BUILTIN(__nvvm_rcp_rn_f, "ff", "")
211 BUILTIN(__nvvm_rcp_rz_ftz_f, "ff", "")
212 BUILTIN(__nvvm_rcp_rz_f, "ff", "")
213 BUILTIN(__nvvm_rcp_rm_ftz_f, "ff", "")
214 BUILTIN(__nvvm_rcp_rm_f, "ff", "")
215 BUILTIN(__nvvm_rcp_rp_ftz_f, "ff", "")
216 BUILTIN(__nvvm_rcp_rp_f, "ff", "")
217
218 BUILTIN(__nvvm_rcp_rn_d, "dd", "")
219 BUILTIN(__nvvm_rcp_rz_d, "dd", "")
220 BUILTIN(__nvvm_rcp_rm_d, "dd", "")
221 BUILTIN(__nvvm_rcp_rp_d, "dd", "")
222 BUILTIN(__nvvm_rcp_approx_ftz_d, "dd", "")
223
224 // Sqrt
225
226 BUILTIN(__nvvm_sqrt_rn_ftz_f, "ff", "")
227 BUILTIN(__nvvm_sqrt_rn_f, "ff", "")
228 BUILTIN(__nvvm_sqrt_rz_ftz_f, "ff", "")
229 BUILTIN(__nvvm_sqrt_rz_f, "ff", "")
230 BUILTIN(__nvvm_sqrt_rm_ftz_f, "ff", "")
231 BUILTIN(__nvvm_sqrt_rm_f, "ff", "")
232 BUILTIN(__nvvm_sqrt_rp_ftz_f, "ff", "")
233 BUILTIN(__nvvm_sqrt_rp_f, "ff", "")
234 BUILTIN(__nvvm_sqrt_approx_ftz_f, "ff", "")
235 BUILTIN(__nvvm_sqrt_approx_f, "ff", "")
236
237 BUILTIN(__nvvm_sqrt_rn_d, "dd", "")
238 BUILTIN(__nvvm_sqrt_rz_d, "dd", "")
239 BUILTIN(__nvvm_sqrt_rm_d, "dd", "")
240 BUILTIN(__nvvm_sqrt_rp_d, "dd", "")
241
242 // Rsqrt
243
244 BUILTIN(__nvvm_rsqrt_approx_ftz_f, "ff", "")
245 BUILTIN(__nvvm_rsqrt_approx_f, "ff", "")
246 BUILTIN(__nvvm_rsqrt_approx_d, "dd", "")
247
248 // Add
249
250 BUILTIN(__nvvm_add_rn_ftz_f, "fff", "")
251 BUILTIN(__nvvm_add_rn_f, "fff", "")
252 BUILTIN(__nvvm_add_rz_ftz_f, "fff", "")
253 BUILTIN(__nvvm_add_rz_f, "fff", "")
254 BUILTIN(__nvvm_add_rm_ftz_f, "fff", "")
255 BUILTIN(__nvvm_add_rm_f, "fff", "")
256 BUILTIN(__nvvm_add_rp_ftz_f, "fff", "")
257 BUILTIN(__nvvm_add_rp_f, "fff", "")
258
259 BUILTIN(__nvvm_add_rn_d, "ddd", "")
260 BUILTIN(__nvvm_add_rz_d, "ddd", "")
261 BUILTIN(__nvvm_add_rm_d, "ddd", "")
262 BUILTIN(__nvvm_add_rp_d, "ddd", "")
263
264 // Convert
265
266 BUILTIN(__nvvm_d2f_rn_ftz, "fd", "")
267 BUILTIN(__nvvm_d2f_rn, "fd", "")
268 BUILTIN(__nvvm_d2f_rz_ftz, "fd", "")
269 BUILTIN(__nvvm_d2f_rz, "fd", "")
270 BUILTIN(__nvvm_d2f_rm_ftz, "fd", "")
271 BUILTIN(__nvvm_d2f_rm, "fd", "")
272 BUILTIN(__nvvm_d2f_rp_ftz, "fd", "")
273 BUILTIN(__nvvm_d2f_rp, "fd", "")
274
275 BUILTIN(__nvvm_d2i_rn, "id", "")
276 BUILTIN(__nvvm_d2i_rz, "id", "")
277 BUILTIN(__nvvm_d2i_rm, "id", "")
278 BUILTIN(__nvvm_d2i_rp, "id", "")
279
280 BUILTIN(__nvvm_d2ui_rn, "Uid", "")
281 BUILTIN(__nvvm_d2ui_rz, "Uid", "")
282 BUILTIN(__nvvm_d2ui_rm, "Uid", "")
283 BUILTIN(__nvvm_d2ui_rp, "Uid", "")
284
285 BUILTIN(__nvvm_i2d_rn, "di", "")
286 BUILTIN(__nvvm_i2d_rz, "di", "")
287 BUILTIN(__nvvm_i2d_rm, "di", "")
288 BUILTIN(__nvvm_i2d_rp, "di", "")
289
290 BUILTIN(__nvvm_ui2d_rn, "dUi", "")
291 BUILTIN(__nvvm_ui2d_rz, "dUi", "")
292 BUILTIN(__nvvm_ui2d_rm, "dUi", "")
293 BUILTIN(__nvvm_ui2d_rp, "dUi", "")
294
295 BUILTIN(__nvvm_f2i_rn_ftz, "if", "")
296 BUILTIN(__nvvm_f2i_rn, "if", "")
297 BUILTIN(__nvvm_f2i_rz_ftz, "if", "")
298 BUILTIN(__nvvm_f2i_rz, "if", "")
299 BUILTIN(__nvvm_f2i_rm_ftz, "if", "")
300 BUILTIN(__nvvm_f2i_rm, "if", "")
301 BUILTIN(__nvvm_f2i_rp_ftz, "if", "")
302 BUILTIN(__nvvm_f2i_rp, "if", "")
303
304 BUILTIN(__nvvm_f2ui_rn_ftz, "Uif", "")
305 BUILTIN(__nvvm_f2ui_rn, "Uif", "")
306 BUILTIN(__nvvm_f2ui_rz_ftz, "Uif", "")
307 BUILTIN(__nvvm_f2ui_rz, "Uif", "")
308 BUILTIN(__nvvm_f2ui_rm_ftz, "Uif", "")
309 BUILTIN(__nvvm_f2ui_rm, "Uif", "")
310 BUILTIN(__nvvm_f2ui_rp_ftz, "Uif", "")
311 BUILTIN(__nvvm_f2ui_rp, "Uif", "")
312
313 BUILTIN(__nvvm_i2f_rn, "fi", "")
314 BUILTIN(__nvvm_i2f_rz, "fi", "")
315 BUILTIN(__nvvm_i2f_rm, "fi", "")
316 BUILTIN(__nvvm_i2f_rp, "fi", "")
317
318 BUILTIN(__nvvm_ui2f_rn, "fUi", "")
319 BUILTIN(__nvvm_ui2f_rz, "fUi", "")
320 BUILTIN(__nvvm_ui2f_rm, "fUi", "")
321 BUILTIN(__nvvm_ui2f_rp, "fUi", "")
322
323 BUILTIN(__nvvm_lohi_i2d, "dii", "")
324
325 BUILTIN(__nvvm_d2i_lo, "id", "")
326 BUILTIN(__nvvm_d2i_hi, "id", "")
327
328 BUILTIN(__nvvm_f2ll_rn_ftz, "LLif", "")
329 BUILTIN(__nvvm_f2ll_rn, "LLif", "")
330 BUILTIN(__nvvm_f2ll_rz_ftz, "LLif", "")
331 BUILTIN(__nvvm_f2ll_rz, "LLif", "")
332 BUILTIN(__nvvm_f2ll_rm_ftz, "LLif", "")
333 BUILTIN(__nvvm_f2ll_rm, "LLif", "")
334 BUILTIN(__nvvm_f2ll_rp_ftz, "LLif", "")
335 BUILTIN(__nvvm_f2ll_rp, "LLif", "")
336
337 BUILTIN(__nvvm_f2ull_rn_ftz, "ULLif", "")
338 BUILTIN(__nvvm_f2ull_rn, "ULLif", "")
339 BUILTIN(__nvvm_f2ull_rz_ftz, "ULLif", "")
340 BUILTIN(__nvvm_f2ull_rz, "ULLif", "")
341 BUILTIN(__nvvm_f2ull_rm_ftz, "ULLif", "")
342 BUILTIN(__nvvm_f2ull_rm, "ULLif", "")
343 BUILTIN(__nvvm_f2ull_rp_ftz, "ULLif", "")
344 BUILTIN(__nvvm_f2ull_rp, "ULLif", "")
345
346 BUILTIN(__nvvm_d2ll_rn, "LLid", "")
347 BUILTIN(__nvvm_d2ll_rz, "LLid", "")
348 BUILTIN(__nvvm_d2ll_rm, "LLid", "")
349 BUILTIN(__nvvm_d2ll_rp, "LLid", "")
350
351 BUILTIN(__nvvm_d2ull_rn, "ULLid", "")
352 BUILTIN(__nvvm_d2ull_rz, "ULLid", "")
353 BUILTIN(__nvvm_d2ull_rm, "ULLid", "")
354 BUILTIN(__nvvm_d2ull_rp, "ULLid", "")
355
356 BUILTIN(__nvvm_ll2f_rn, "fLLi", "")
357 BUILTIN(__nvvm_ll2f_rz, "fLLi", "")
358 BUILTIN(__nvvm_ll2f_rm, "fLLi", "")
359 BUILTIN(__nvvm_ll2f_rp, "fLLi", "")
360
361 BUILTIN(__nvvm_ull2f_rn, "fULLi", "")
362 BUILTIN(__nvvm_ull2f_rz, "fULLi", "")
363 BUILTIN(__nvvm_ull2f_rm, "fULLi", "")
364 BUILTIN(__nvvm_ull2f_rp, "fULLi", "")
365
366 BUILTIN(__nvvm_ll2d_rn, "dLLi", "")
367 BUILTIN(__nvvm_ll2d_rz, "dLLi", "")
368 BUILTIN(__nvvm_ll2d_rm, "dLLi", "")
369 BUILTIN(__nvvm_ll2d_rp, "dLLi", "")
370
371 BUILTIN(__nvvm_ull2d_rn, "dULLi", "")
372 BUILTIN(__nvvm_ull2d_rz, "dULLi", "")
373 BUILTIN(__nvvm_ull2d_rm, "dULLi", "")
374 BUILTIN(__nvvm_ull2d_rp, "dULLi", "")
375
376 BUILTIN(__nvvm_f2h_rn_ftz, "Usf", "")
377 BUILTIN(__nvvm_f2h_rn, "Usf", "")
378
379 // Bitcast
380
381 BUILTIN(__nvvm_bitcast_f2i, "if", "")
382 BUILTIN(__nvvm_bitcast_i2f, "fi", "")
383
384 BUILTIN(__nvvm_bitcast_ll2d, "dLLi", "")
385 BUILTIN(__nvvm_bitcast_d2ll, "LLid", "")
386
387 // FNS
388 TARGET_BUILTIN(__nvvm_fns, "UiUiUii", "n", PTX60)
389
390 // Sync
391
392 BUILTIN(__syncthreads, "v", "")
393 BUILTIN(__nvvm_bar0_popc, "ii", "")
394 BUILTIN(__nvvm_bar0_and, "ii", "")
395 BUILTIN(__nvvm_bar0_or, "ii", "")
396 BUILTIN(__nvvm_bar_sync, "vi", "n")
397 TARGET_BUILTIN(__nvvm_bar_warp_sync, "vUi", "n", PTX60)
398 TARGET_BUILTIN(__nvvm_barrier_sync, "vUi", "n", PTX60)
399 TARGET_BUILTIN(__nvvm_barrier_sync_cnt, "vUiUi", "n", PTX60)
400
401 // Shuffle
402
403 BUILTIN(__nvvm_shfl_down_i32, "iiii", "")
404 BUILTIN(__nvvm_shfl_down_f32, "ffii", "")
405 BUILTIN(__nvvm_shfl_up_i32, "iiii", "")
406 BUILTIN(__nvvm_shfl_up_f32, "ffii", "")
407 BUILTIN(__nvvm_shfl_bfly_i32, "iiii", "")
408 BUILTIN(__nvvm_shfl_bfly_f32, "ffii", "")
409 BUILTIN(__nvvm_shfl_idx_i32, "iiii", "")
410 BUILTIN(__nvvm_shfl_idx_f32, "ffii", "")
411
412 TARGET_BUILTIN(__nvvm_shfl_sync_down_i32, "iUiiii", "", PTX60)
413 TARGET_BUILTIN(__nvvm_shfl_sync_down_f32, "fUifii", "", PTX60)
414 TARGET_BUILTIN(__nvvm_shfl_sync_up_i32, "iUiiii", "", PTX60)
415 TARGET_BUILTIN(__nvvm_shfl_sync_up_f32, "fUifii", "", PTX60)
416 TARGET_BUILTIN(__nvvm_shfl_sync_bfly_i32, "iUiiii", "", PTX60)
417 TARGET_BUILTIN(__nvvm_shfl_sync_bfly_f32, "fUifii", "", PTX60)
418 TARGET_BUILTIN(__nvvm_shfl_sync_idx_i32, "iUiiii", "", PTX60)
419 TARGET_BUILTIN(__nvvm_shfl_sync_idx_f32, "fUifii", "", PTX60)
420
421 // Vote
422 BUILTIN(__nvvm_vote_all, "bb", "")
423 BUILTIN(__nvvm_vote_any, "bb", "")
424 BUILTIN(__nvvm_vote_uni, "bb", "")
425 BUILTIN(__nvvm_vote_ballot, "Uib", "")
426
427 TARGET_BUILTIN(__nvvm_vote_all_sync, "bUib", "", PTX60)
428 TARGET_BUILTIN(__nvvm_vote_any_sync, "bUib", "", PTX60)
429 TARGET_BUILTIN(__nvvm_vote_uni_sync, "bUib", "", PTX60)
430 TARGET_BUILTIN(__nvvm_vote_ballot_sync, "UiUib", "", PTX60)
431
432 // Match
433 TARGET_BUILTIN(__nvvm_match_any_sync_i32, "UiUiUi", "", PTX60)
434 TARGET_BUILTIN(__nvvm_match_any_sync_i64, "WiUiWi", "", PTX60)
435 // These return a pair {value, predicate}, which requires custom lowering.
436 TARGET_BUILTIN(__nvvm_match_all_sync_i32p, "UiUiUii*", "", PTX60)
437 TARGET_BUILTIN(__nvvm_match_all_sync_i64p, "WiUiWii*", "", PTX60)
438
439 // Membar
440
441 BUILTIN(__nvvm_membar_cta, "v", "")
442 BUILTIN(__nvvm_membar_gl, "v", "")
443 BUILTIN(__nvvm_membar_sys, "v", "")
444
445 // Memcpy, Memset
446
447 BUILTIN(__nvvm_memcpy, "vUc*Uc*zi","")
448 BUILTIN(__nvvm_memset, "vUc*Uczi","")
449
450 // Image
451
452 BUILTIN(__builtin_ptx_read_image2Dfi_, "V4fiiii", "")
453 BUILTIN(__builtin_ptx_read_image2Dff_, "V4fiiff", "")
454 BUILTIN(__builtin_ptx_read_image2Dii_, "V4iiiii", "")
455 BUILTIN(__builtin_ptx_read_image2Dif_, "V4iiiff", "")
456
457 BUILTIN(__builtin_ptx_read_image3Dfi_, "V4fiiiiii", "")
458 BUILTIN(__builtin_ptx_read_image3Dff_, "V4fiiffff", "")
459 BUILTIN(__builtin_ptx_read_image3Dii_, "V4iiiiiii", "")
460 BUILTIN(__builtin_ptx_read_image3Dif_, "V4iiiffff", "")
461
462 BUILTIN(__builtin_ptx_write_image2Df_, "viiiffff", "")
463 BUILTIN(__builtin_ptx_write_image2Di_, "viiiiiii", "")
464 BUILTIN(__builtin_ptx_write_image2Dui_, "viiiUiUiUiUi", "")
465 BUILTIN(__builtin_ptx_get_image_depthi_, "ii", "")
466 BUILTIN(__builtin_ptx_get_image_heighti_, "ii", "")
467 BUILTIN(__builtin_ptx_get_image_widthi_, "ii", "")
468 BUILTIN(__builtin_ptx_get_image_channel_data_typei_, "ii", "")
469 BUILTIN(__builtin_ptx_get_image_channel_orderi_, "ii", "")
470
471 // Atomic
472 //
473 // We need the atom intrinsics because
474 // - they are used in converging analysis
475 // - they are used in address space analysis and optimization
476 // So it does not hurt to expose them as builtins.
477 //
478 BUILTIN(__nvvm_atom_add_gen_i, "iiD*i", "n")
479 TARGET_BUILTIN(__nvvm_atom_cta_add_gen_i, "iiD*i", "n", SM_60)
480 TARGET_BUILTIN(__nvvm_atom_sys_add_gen_i, "iiD*i", "n", SM_60)
481 BUILTIN(__nvvm_atom_add_gen_l, "LiLiD*Li", "n")
482 TARGET_BUILTIN(__nvvm_atom_cta_add_gen_l, "LiLiD*Li", "n", SM_60)
483 TARGET_BUILTIN(__nvvm_atom_sys_add_gen_l, "LiLiD*Li", "n", SM_60)
484 BUILTIN(__nvvm_atom_add_gen_ll, "LLiLLiD*LLi", "n")
485 TARGET_BUILTIN(__nvvm_atom_cta_add_gen_ll, "LLiLLiD*LLi", "n", SM_60)
486 TARGET_BUILTIN(__nvvm_atom_sys_add_gen_ll, "LLiLLiD*LLi", "n", SM_60)
487 BUILTIN(__nvvm_atom_add_gen_f, "ffD*f", "n")
488 TARGET_BUILTIN(__nvvm_atom_cta_add_gen_f, "ffD*f", "n", SM_60)
489 TARGET_BUILTIN(__nvvm_atom_sys_add_gen_f, "ffD*f", "n", SM_60)
490 TARGET_BUILTIN(__nvvm_atom_add_gen_d, "ddD*d", "n", SM_60)
491 TARGET_BUILTIN(__nvvm_atom_cta_add_gen_d, "ddD*d", "n", SM_60)
492 TARGET_BUILTIN(__nvvm_atom_sys_add_gen_d, "ddD*d", "n", SM_60)
493
494 BUILTIN(__nvvm_atom_sub_gen_i, "iiD*i", "n")
495 BUILTIN(__nvvm_atom_sub_gen_l, "LiLiD*Li", "n")
496 BUILTIN(__nvvm_atom_sub_gen_ll, "LLiLLiD*LLi", "n")
497
498 BUILTIN(__nvvm_atom_xchg_gen_i, "iiD*i", "n")
499 TARGET_BUILTIN(__nvvm_atom_cta_xchg_gen_i, "iiD*i", "n", SM_60)
500 TARGET_BUILTIN(__nvvm_atom_sys_xchg_gen_i, "iiD*i", "n", SM_60)
501 BUILTIN(__nvvm_atom_xchg_gen_l, "LiLiD*Li", "n")
502 TARGET_BUILTIN(__nvvm_atom_cta_xchg_gen_l, "LiLiD*Li", "n", SM_60)
503 TARGET_BUILTIN(__nvvm_atom_sys_xchg_gen_l, "LiLiD*Li", "n", SM_60)
504 BUILTIN(__nvvm_atom_xchg_gen_ll, "LLiLLiD*LLi", "n")
505 TARGET_BUILTIN(__nvvm_atom_cta_xchg_gen_ll, "LLiLLiD*LLi", "n", SM_60)
506 TARGET_BUILTIN(__nvvm_atom_sys_xchg_gen_ll, "LLiLLiD*LLi", "n", SM_60)
507
508 BUILTIN(__nvvm_atom_max_gen_i, "iiD*i", "n")
509 TARGET_BUILTIN(__nvvm_atom_cta_max_gen_i, "iiD*i", "n", SM_60)
510 TARGET_BUILTIN(__nvvm_atom_sys_max_gen_i, "iiD*i", "n", SM_60)
511 BUILTIN(__nvvm_atom_max_gen_ui, "UiUiD*Ui", "n")
512 TARGET_BUILTIN(__nvvm_atom_cta_max_gen_ui, "UiUiD*Ui", "n", SM_60)
513 TARGET_BUILTIN(__nvvm_atom_sys_max_gen_ui, "UiUiD*Ui", "n", SM_60)
514 BUILTIN(__nvvm_atom_max_gen_l, "LiLiD*Li", "n")
515 TARGET_BUILTIN(__nvvm_atom_cta_max_gen_l, "LiLiD*Li", "n", SM_60)
516 TARGET_BUILTIN(__nvvm_atom_sys_max_gen_l, "LiLiD*Li", "n", SM_60)
517 BUILTIN(__nvvm_atom_max_gen_ul, "ULiULiD*ULi", "n")
518 TARGET_BUILTIN(__nvvm_atom_cta_max_gen_ul, "ULiULiD*ULi", "n", SM_60)
519 TARGET_BUILTIN(__nvvm_atom_sys_max_gen_ul, "ULiULiD*ULi", "n", SM_60)
520 BUILTIN(__nvvm_atom_max_gen_ll, "LLiLLiD*LLi", "n")
521 TARGET_BUILTIN(__nvvm_atom_cta_max_gen_ll, "LLiLLiD*LLi", "n", SM_60)
522 TARGET_BUILTIN(__nvvm_atom_sys_max_gen_ll, "LLiLLiD*LLi", "n", SM_60)
523 BUILTIN(__nvvm_atom_max_gen_ull, "ULLiULLiD*ULLi", "n")
524 TARGET_BUILTIN(__nvvm_atom_cta_max_gen_ull, "ULLiULLiD*ULLi", "n", SM_60)
525 TARGET_BUILTIN(__nvvm_atom_sys_max_gen_ull, "ULLiULLiD*ULLi", "n", SM_60)
526
527 BUILTIN(__nvvm_atom_min_gen_i, "iiD*i", "n")
528 TARGET_BUILTIN(__nvvm_atom_cta_min_gen_i, "iiD*i", "n", SM_60)
529 TARGET_BUILTIN(__nvvm_atom_sys_min_gen_i, "iiD*i", "n", SM_60)
530 BUILTIN(__nvvm_atom_min_gen_ui, "UiUiD*Ui", "n")
531 TARGET_BUILTIN(__nvvm_atom_cta_min_gen_ui, "UiUiD*Ui", "n", SM_60)
532 TARGET_BUILTIN(__nvvm_atom_sys_min_gen_ui, "UiUiD*Ui", "n", SM_60)
533 BUILTIN(__nvvm_atom_min_gen_l, "LiLiD*Li", "n")
534 TARGET_BUILTIN(__nvvm_atom_cta_min_gen_l, "LiLiD*Li", "n", SM_60)
535 TARGET_BUILTIN(__nvvm_atom_sys_min_gen_l, "LiLiD*Li", "n", SM_60)
536 BUILTIN(__nvvm_atom_min_gen_ul, "ULiULiD*ULi", "n")
537 TARGET_BUILTIN(__nvvm_atom_cta_min_gen_ul, "ULiULiD*ULi", "n", SM_60)
538 TARGET_BUILTIN(__nvvm_atom_sys_min_gen_ul, "ULiULiD*ULi", "n", SM_60)
539 BUILTIN(__nvvm_atom_min_gen_ll, "LLiLLiD*LLi", "n")
540 TARGET_BUILTIN(__nvvm_atom_cta_min_gen_ll, "LLiLLiD*LLi", "n", SM_60)
541 TARGET_BUILTIN(__nvvm_atom_sys_min_gen_ll, "LLiLLiD*LLi", "n", SM_60)
542 BUILTIN(__nvvm_atom_min_gen_ull, "ULLiULLiD*ULLi", "n")
543 TARGET_BUILTIN(__nvvm_atom_cta_min_gen_ull, "ULLiULLiD*ULLi", "n", SM_60)
544 TARGET_BUILTIN(__nvvm_atom_sys_min_gen_ull, "ULLiULLiD*ULLi", "n", SM_60)
545
546 BUILTIN(__nvvm_atom_inc_gen_ui, "UiUiD*Ui", "n")
547 TARGET_BUILTIN(__nvvm_atom_cta_inc_gen_ui, "UiUiD*Ui", "n", SM_60)
548 TARGET_BUILTIN(__nvvm_atom_sys_inc_gen_ui, "UiUiD*Ui", "n", SM_60)
549 BUILTIN(__nvvm_atom_dec_gen_ui, "UiUiD*Ui", "n")
550 TARGET_BUILTIN(__nvvm_atom_cta_dec_gen_ui, "UiUiD*Ui", "n", SM_60)
551 TARGET_BUILTIN(__nvvm_atom_sys_dec_gen_ui, "UiUiD*Ui", "n", SM_60)
552
553 BUILTIN(__nvvm_atom_and_gen_i, "iiD*i", "n")
554 TARGET_BUILTIN(__nvvm_atom_cta_and_gen_i, "iiD*i", "n", SM_60)
555 TARGET_BUILTIN(__nvvm_atom_sys_and_gen_i, "iiD*i", "n", SM_60)
556 BUILTIN(__nvvm_atom_and_gen_l, "LiLiD*Li", "n")
557 TARGET_BUILTIN(__nvvm_atom_cta_and_gen_l, "LiLiD*Li", "n", SM_60)
558 TARGET_BUILTIN(__nvvm_atom_sys_and_gen_l, "LiLiD*Li", "n", SM_60)
559 BUILTIN(__nvvm_atom_and_gen_ll, "LLiLLiD*LLi", "n")
560 TARGET_BUILTIN(__nvvm_atom_cta_and_gen_ll, "LLiLLiD*LLi", "n", SM_60)
561 TARGET_BUILTIN(__nvvm_atom_sys_and_gen_ll, "LLiLLiD*LLi", "n", SM_60)
562
563 BUILTIN(__nvvm_atom_or_gen_i, "iiD*i", "n")
564 TARGET_BUILTIN(__nvvm_atom_cta_or_gen_i, "iiD*i", "n", SM_60)
565 TARGET_BUILTIN(__nvvm_atom_sys_or_gen_i, "iiD*i", "n", SM_60)
566 BUILTIN(__nvvm_atom_or_gen_l, "LiLiD*Li", "n")
567 TARGET_BUILTIN(__nvvm_atom_cta_or_gen_l, "LiLiD*Li", "n", SM_60)
568 TARGET_BUILTIN(__nvvm_atom_sys_or_gen_l, "LiLiD*Li", "n", SM_60)
569 BUILTIN(__nvvm_atom_or_gen_ll, "LLiLLiD*LLi", "n")
570 TARGET_BUILTIN(__nvvm_atom_cta_or_gen_ll, "LLiLLiD*LLi", "n", SM_60)
571 TARGET_BUILTIN(__nvvm_atom_sys_or_gen_ll, "LLiLLiD*LLi", "n", SM_60)
572
573 BUILTIN(__nvvm_atom_xor_gen_i, "iiD*i", "n")
574 TARGET_BUILTIN(__nvvm_atom_cta_xor_gen_i, "iiD*i", "n", SM_60)
575 TARGET_BUILTIN(__nvvm_atom_sys_xor_gen_i, "iiD*i", "n", SM_60)
576 BUILTIN(__nvvm_atom_xor_gen_l, "LiLiD*Li", "n")
577 TARGET_BUILTIN(__nvvm_atom_cta_xor_gen_l, "LiLiD*Li", "n", SM_60)
578 TARGET_BUILTIN(__nvvm_atom_sys_xor_gen_l, "LiLiD*Li", "n", SM_60)
579 BUILTIN(__nvvm_atom_xor_gen_ll, "LLiLLiD*LLi", "n")
580 TARGET_BUILTIN(__nvvm_atom_cta_xor_gen_ll, "LLiLLiD*LLi", "n", SM_60)
581 TARGET_BUILTIN(__nvvm_atom_sys_xor_gen_ll, "LLiLLiD*LLi", "n", SM_60)
582
583 BUILTIN(__nvvm_atom_cas_gen_i, "iiD*ii", "n")
584 TARGET_BUILTIN(__nvvm_atom_cta_cas_gen_i, "iiD*ii", "n", SM_60)
585 TARGET_BUILTIN(__nvvm_atom_sys_cas_gen_i, "iiD*ii", "n", SM_60)
586 BUILTIN(__nvvm_atom_cas_gen_l, "LiLiD*LiLi", "n")
587 TARGET_BUILTIN(__nvvm_atom_cta_cas_gen_l, "LiLiD*LiLi", "n", SM_60)
588 TARGET_BUILTIN(__nvvm_atom_sys_cas_gen_l, "LiLiD*LiLi", "n", SM_60)
589 BUILTIN(__nvvm_atom_cas_gen_ll, "LLiLLiD*LLiLLi", "n")
590 TARGET_BUILTIN(__nvvm_atom_cta_cas_gen_ll, "LLiLLiD*LLiLLi", "n", SM_60)
591 TARGET_BUILTIN(__nvvm_atom_sys_cas_gen_ll, "LLiLLiD*LLiLLi", "n", SM_60)
592
593 // Compiler Error Warn
594 BUILTIN(__nvvm_compiler_error, "vcC*4", "n")
595 BUILTIN(__nvvm_compiler_warn, "vcC*4", "n")
596
597 // __ldg.  This is not implemented as a builtin by nvcc.
598 BUILTIN(__nvvm_ldg_c, "ccC*", "")
599 BUILTIN(__nvvm_ldg_s, "ssC*", "")
600 BUILTIN(__nvvm_ldg_i, "iiC*", "")
601 BUILTIN(__nvvm_ldg_l, "LiLiC*", "")
602 BUILTIN(__nvvm_ldg_ll, "LLiLLiC*", "")
603
604 BUILTIN(__nvvm_ldg_uc, "UcUcC*", "")
605 BUILTIN(__nvvm_ldg_us, "UsUsC*", "")
606 BUILTIN(__nvvm_ldg_ui, "UiUiC*", "")
607 BUILTIN(__nvvm_ldg_ul, "ULiULiC*", "")
608 BUILTIN(__nvvm_ldg_ull, "ULLiULLiC*", "")
609
610 BUILTIN(__nvvm_ldg_f, "ffC*", "")
611 BUILTIN(__nvvm_ldg_d, "ddC*", "")
612
613 BUILTIN(__nvvm_ldg_c2, "E2cE2cC*", "")
614 BUILTIN(__nvvm_ldg_c4, "E4cE4cC*", "")
615 BUILTIN(__nvvm_ldg_s2, "E2sE2sC*", "")
616 BUILTIN(__nvvm_ldg_s4, "E4sE4sC*", "")
617 BUILTIN(__nvvm_ldg_i2, "E2iE2iC*", "")
618 BUILTIN(__nvvm_ldg_i4, "E4iE4iC*", "")
619 BUILTIN(__nvvm_ldg_ll2, "E2LLiE2LLiC*", "")
620
621 BUILTIN(__nvvm_ldg_uc2, "E2UcE2UcC*", "")
622 BUILTIN(__nvvm_ldg_uc4, "E4UcE4UcC*", "")
623 BUILTIN(__nvvm_ldg_us2, "E2UsE2UsC*", "")
624 BUILTIN(__nvvm_ldg_us4, "E4UsE4UsC*", "")
625 BUILTIN(__nvvm_ldg_ui2, "E2UiE2UiC*", "")
626 BUILTIN(__nvvm_ldg_ui4, "E4UiE4UiC*", "")
627 BUILTIN(__nvvm_ldg_ull2, "E2ULLiE2ULLiC*", "")
628
629 BUILTIN(__nvvm_ldg_f2, "E2fE2fC*", "")
630 BUILTIN(__nvvm_ldg_f4, "E4fE4fC*", "")
631 BUILTIN(__nvvm_ldg_d2, "E2dE2dC*", "")
632
633 // Builtins to support WMMA instructions on sm_70
634 TARGET_BUILTIN(__hmma_m16n16k16_ld_a, "vi*iC*UiIi", "", AND(SM_70,PTX60))
635 TARGET_BUILTIN(__hmma_m16n16k16_ld_b, "vi*iC*UiIi", "", AND(SM_70,PTX60))
636 TARGET_BUILTIN(__hmma_m16n16k16_ld_c_f16, "vi*iC*UiIi", "", AND(SM_70,PTX60))
637 TARGET_BUILTIN(__hmma_m16n16k16_ld_c_f32, "vf*fC*UiIi", "", AND(SM_70,PTX60))
638 TARGET_BUILTIN(__hmma_m16n16k16_st_c_f16, "vi*i*UiIi", "", AND(SM_70,PTX60))
639 TARGET_BUILTIN(__hmma_m16n16k16_st_c_f32, "vf*f*UiIi", "", AND(SM_70,PTX60))
640
641 TARGET_BUILTIN(__hmma_m32n8k16_ld_a, "vi*iC*UiIi", "", AND(SM_70,PTX61))
642 TARGET_BUILTIN(__hmma_m32n8k16_ld_b, "vi*iC*UiIi", "", AND(SM_70,PTX61))
643 TARGET_BUILTIN(__hmma_m32n8k16_ld_c_f16, "vi*iC*UiIi", "", AND(SM_70,PTX61))
644 TARGET_BUILTIN(__hmma_m32n8k16_ld_c_f32, "vf*fC*UiIi", "", AND(SM_70,PTX61))
645 TARGET_BUILTIN(__hmma_m32n8k16_st_c_f16, "vi*i*UiIi", "", AND(SM_70,PTX61))
646 TARGET_BUILTIN(__hmma_m32n8k16_st_c_f32, "vf*f*UiIi", "", AND(SM_70,PTX61))
647
648 TARGET_BUILTIN(__hmma_m8n32k16_ld_a, "vi*iC*UiIi", "", AND(SM_70,PTX61))
649 TARGET_BUILTIN(__hmma_m8n32k16_ld_b, "vi*iC*UiIi", "", AND(SM_70,PTX61))
650 TARGET_BUILTIN(__hmma_m8n32k16_ld_c_f16, "vi*iC*UiIi", "", AND(SM_70,PTX61))
651 TARGET_BUILTIN(__hmma_m8n32k16_ld_c_f32, "vf*fC*UiIi", "", AND(SM_70,PTX61))
652 TARGET_BUILTIN(__hmma_m8n32k16_st_c_f16, "vi*i*UiIi", "", AND(SM_70,PTX61))
653 TARGET_BUILTIN(__hmma_m8n32k16_st_c_f32, "vf*f*UiIi", "", AND(SM_70,PTX61))
654
655 TARGET_BUILTIN(__hmma_m16n16k16_mma_f16f16, "vi*iC*iC*iC*IiIi", "", AND(SM_70,PTX60))
656 TARGET_BUILTIN(__hmma_m16n16k16_mma_f32f16, "vf*iC*iC*iC*IiIi", "", AND(SM_70,PTX60))
657 TARGET_BUILTIN(__hmma_m16n16k16_mma_f32f32, "vf*iC*iC*fC*IiIi", "", AND(SM_70,PTX60))
658 TARGET_BUILTIN(__hmma_m16n16k16_mma_f16f32, "vi*iC*iC*fC*IiIi", "", AND(SM_70,PTX60))
659
660 TARGET_BUILTIN(__hmma_m32n8k16_mma_f16f16, "vi*iC*iC*iC*IiIi", "", AND(SM_70,PTX61))
661 TARGET_BUILTIN(__hmma_m32n8k16_mma_f32f16, "vf*iC*iC*iC*IiIi", "", AND(SM_70,PTX61))
662 TARGET_BUILTIN(__hmma_m32n8k16_mma_f32f32, "vf*iC*iC*fC*IiIi", "", AND(SM_70,PTX61))
663 TARGET_BUILTIN(__hmma_m32n8k16_mma_f16f32, "vi*iC*iC*fC*IiIi", "", AND(SM_70,PTX61))
664
665 TARGET_BUILTIN(__hmma_m8n32k16_mma_f16f16, "vi*iC*iC*iC*IiIi", "", AND(SM_70,PTX61))
666 TARGET_BUILTIN(__hmma_m8n32k16_mma_f32f16, "vf*iC*iC*iC*IiIi", "", AND(SM_70,PTX61))
667 TARGET_BUILTIN(__hmma_m8n32k16_mma_f32f32, "vf*iC*iC*fC*IiIi", "", AND(SM_70,PTX61))
668 TARGET_BUILTIN(__hmma_m8n32k16_mma_f16f32, "vi*iC*iC*fC*IiIi", "", AND(SM_70,PTX61))
669
670 #undef BUILTIN
671 #undef TARGET_BUILTIN
672 #pragma pop_macro("AND")
673 #pragma pop_macro("SM_60")
674 #pragma pop_macro("SM_70")
675 #pragma pop_macro("PTX60")
676 #pragma pop_macro("PTX61")