1 //===-- NVPTXISelDAGToDAG.cpp - A dag to dag inst selector for NVPTX ------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // This file defines an instruction selector for the NVPTX target.
11 //===----------------------------------------------------------------------===//
13 #include "NVPTXISelDAGToDAG.h"
14 #include "MCTargetDesc/NVPTXBaseInfo.h"
15 #include "NVPTXUtilities.h"
16 #include "llvm/Analysis/ValueTracking.h"
17 #include "llvm/IR/GlobalValue.h"
18 #include "llvm/IR/Instructions.h"
19 #include "llvm/IR/IntrinsicsNVPTX.h"
20 #include "llvm/Support/AtomicOrdering.h"
21 #include "llvm/Support/CommandLine.h"
22 #include "llvm/Support/Debug.h"
23 #include "llvm/Support/ErrorHandling.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include "llvm/Target/TargetIntrinsicInfo.h"
29 #define DEBUG_TYPE "nvptx-isel"
31 /// createNVPTXISelDag - This pass converts a legalized DAG into a
32 /// NVPTX-specific DAG, ready for instruction scheduling.
33 FunctionPass *llvm::createNVPTXISelDag(NVPTXTargetMachine &TM,
34 llvm::CodeGenOpt::Level OptLevel) {
35 return new NVPTXDAGToDAGISel(TM, OptLevel);
38 NVPTXDAGToDAGISel::NVPTXDAGToDAGISel(NVPTXTargetMachine &tm,
39 CodeGenOpt::Level OptLevel)
40 : SelectionDAGISel(tm, OptLevel), TM(tm) {
41 doMulWide = (OptLevel > 0);
44 bool NVPTXDAGToDAGISel::runOnMachineFunction(MachineFunction &MF) {
45 Subtarget = &static_cast<const NVPTXSubtarget &>(MF.getSubtarget());
46 return SelectionDAGISel::runOnMachineFunction(MF);
49 int NVPTXDAGToDAGISel::getDivF32Level() const {
50 return Subtarget->getTargetLowering()->getDivF32Level();
53 bool NVPTXDAGToDAGISel::usePrecSqrtF32() const {
54 return Subtarget->getTargetLowering()->usePrecSqrtF32();
57 bool NVPTXDAGToDAGISel::useF32FTZ() const {
58 return Subtarget->getTargetLowering()->useF32FTZ(*MF);
61 bool NVPTXDAGToDAGISel::allowFMA() const {
62 const NVPTXTargetLowering *TL = Subtarget->getTargetLowering();
63 return TL->allowFMA(*MF, OptLevel);
66 bool NVPTXDAGToDAGISel::allowUnsafeFPMath() const {
67 const NVPTXTargetLowering *TL = Subtarget->getTargetLowering();
68 return TL->allowUnsafeFPMath(*MF);
71 bool NVPTXDAGToDAGISel::useShortPointers() const {
72 return TM.useShortPointers();
75 /// Select - Select instructions not customized! Used for
76 /// expanded, promoted and normal instructions.
77 void NVPTXDAGToDAGISel::Select(SDNode *N) {
79 if (N->isMachineOpcode()) {
81 return; // Already selected.
84 switch (N->getOpcode()) {
86 case ISD::ATOMIC_LOAD:
91 case ISD::ATOMIC_STORE:
95 case ISD::EXTRACT_VECTOR_ELT:
96 if (tryEXTRACT_VECTOR_ELEMENT(N))
99 case NVPTXISD::SETP_F16X2:
103 case NVPTXISD::LoadV2:
104 case NVPTXISD::LoadV4:
105 if (tryLoadVector(N))
108 case NVPTXISD::LDGV2:
109 case NVPTXISD::LDGV4:
110 case NVPTXISD::LDUV2:
111 case NVPTXISD::LDUV4:
115 case NVPTXISD::StoreV2:
116 case NVPTXISD::StoreV4:
117 if (tryStoreVector(N))
120 case NVPTXISD::LoadParam:
121 case NVPTXISD::LoadParamV2:
122 case NVPTXISD::LoadParamV4:
126 case NVPTXISD::StoreRetval:
127 case NVPTXISD::StoreRetvalV2:
128 case NVPTXISD::StoreRetvalV4:
129 if (tryStoreRetval(N))
132 case NVPTXISD::StoreParam:
133 case NVPTXISD::StoreParamV2:
134 case NVPTXISD::StoreParamV4:
135 case NVPTXISD::StoreParamS32:
136 case NVPTXISD::StoreParamU32:
137 if (tryStoreParam(N))
140 case ISD::INTRINSIC_WO_CHAIN:
141 if (tryIntrinsicNoChain(N))
144 case ISD::INTRINSIC_W_CHAIN:
145 if (tryIntrinsicChain(N))
148 case NVPTXISD::Tex1DFloatS32:
149 case NVPTXISD::Tex1DFloatFloat:
150 case NVPTXISD::Tex1DFloatFloatLevel:
151 case NVPTXISD::Tex1DFloatFloatGrad:
152 case NVPTXISD::Tex1DS32S32:
153 case NVPTXISD::Tex1DS32Float:
154 case NVPTXISD::Tex1DS32FloatLevel:
155 case NVPTXISD::Tex1DS32FloatGrad:
156 case NVPTXISD::Tex1DU32S32:
157 case NVPTXISD::Tex1DU32Float:
158 case NVPTXISD::Tex1DU32FloatLevel:
159 case NVPTXISD::Tex1DU32FloatGrad:
160 case NVPTXISD::Tex1DArrayFloatS32:
161 case NVPTXISD::Tex1DArrayFloatFloat:
162 case NVPTXISD::Tex1DArrayFloatFloatLevel:
163 case NVPTXISD::Tex1DArrayFloatFloatGrad:
164 case NVPTXISD::Tex1DArrayS32S32:
165 case NVPTXISD::Tex1DArrayS32Float:
166 case NVPTXISD::Tex1DArrayS32FloatLevel:
167 case NVPTXISD::Tex1DArrayS32FloatGrad:
168 case NVPTXISD::Tex1DArrayU32S32:
169 case NVPTXISD::Tex1DArrayU32Float:
170 case NVPTXISD::Tex1DArrayU32FloatLevel:
171 case NVPTXISD::Tex1DArrayU32FloatGrad:
172 case NVPTXISD::Tex2DFloatS32:
173 case NVPTXISD::Tex2DFloatFloat:
174 case NVPTXISD::Tex2DFloatFloatLevel:
175 case NVPTXISD::Tex2DFloatFloatGrad:
176 case NVPTXISD::Tex2DS32S32:
177 case NVPTXISD::Tex2DS32Float:
178 case NVPTXISD::Tex2DS32FloatLevel:
179 case NVPTXISD::Tex2DS32FloatGrad:
180 case NVPTXISD::Tex2DU32S32:
181 case NVPTXISD::Tex2DU32Float:
182 case NVPTXISD::Tex2DU32FloatLevel:
183 case NVPTXISD::Tex2DU32FloatGrad:
184 case NVPTXISD::Tex2DArrayFloatS32:
185 case NVPTXISD::Tex2DArrayFloatFloat:
186 case NVPTXISD::Tex2DArrayFloatFloatLevel:
187 case NVPTXISD::Tex2DArrayFloatFloatGrad:
188 case NVPTXISD::Tex2DArrayS32S32:
189 case NVPTXISD::Tex2DArrayS32Float:
190 case NVPTXISD::Tex2DArrayS32FloatLevel:
191 case NVPTXISD::Tex2DArrayS32FloatGrad:
192 case NVPTXISD::Tex2DArrayU32S32:
193 case NVPTXISD::Tex2DArrayU32Float:
194 case NVPTXISD::Tex2DArrayU32FloatLevel:
195 case NVPTXISD::Tex2DArrayU32FloatGrad:
196 case NVPTXISD::Tex3DFloatS32:
197 case NVPTXISD::Tex3DFloatFloat:
198 case NVPTXISD::Tex3DFloatFloatLevel:
199 case NVPTXISD::Tex3DFloatFloatGrad:
200 case NVPTXISD::Tex3DS32S32:
201 case NVPTXISD::Tex3DS32Float:
202 case NVPTXISD::Tex3DS32FloatLevel:
203 case NVPTXISD::Tex3DS32FloatGrad:
204 case NVPTXISD::Tex3DU32S32:
205 case NVPTXISD::Tex3DU32Float:
206 case NVPTXISD::Tex3DU32FloatLevel:
207 case NVPTXISD::Tex3DU32FloatGrad:
208 case NVPTXISD::TexCubeFloatFloat:
209 case NVPTXISD::TexCubeFloatFloatLevel:
210 case NVPTXISD::TexCubeS32Float:
211 case NVPTXISD::TexCubeS32FloatLevel:
212 case NVPTXISD::TexCubeU32Float:
213 case NVPTXISD::TexCubeU32FloatLevel:
214 case NVPTXISD::TexCubeArrayFloatFloat:
215 case NVPTXISD::TexCubeArrayFloatFloatLevel:
216 case NVPTXISD::TexCubeArrayS32Float:
217 case NVPTXISD::TexCubeArrayS32FloatLevel:
218 case NVPTXISD::TexCubeArrayU32Float:
219 case NVPTXISD::TexCubeArrayU32FloatLevel:
220 case NVPTXISD::Tld4R2DFloatFloat:
221 case NVPTXISD::Tld4G2DFloatFloat:
222 case NVPTXISD::Tld4B2DFloatFloat:
223 case NVPTXISD::Tld4A2DFloatFloat:
224 case NVPTXISD::Tld4R2DS64Float:
225 case NVPTXISD::Tld4G2DS64Float:
226 case NVPTXISD::Tld4B2DS64Float:
227 case NVPTXISD::Tld4A2DS64Float:
228 case NVPTXISD::Tld4R2DU64Float:
229 case NVPTXISD::Tld4G2DU64Float:
230 case NVPTXISD::Tld4B2DU64Float:
231 case NVPTXISD::Tld4A2DU64Float:
232 case NVPTXISD::TexUnified1DFloatS32:
233 case NVPTXISD::TexUnified1DFloatFloat:
234 case NVPTXISD::TexUnified1DFloatFloatLevel:
235 case NVPTXISD::TexUnified1DFloatFloatGrad:
236 case NVPTXISD::TexUnified1DS32S32:
237 case NVPTXISD::TexUnified1DS32Float:
238 case NVPTXISD::TexUnified1DS32FloatLevel:
239 case NVPTXISD::TexUnified1DS32FloatGrad:
240 case NVPTXISD::TexUnified1DU32S32:
241 case NVPTXISD::TexUnified1DU32Float:
242 case NVPTXISD::TexUnified1DU32FloatLevel:
243 case NVPTXISD::TexUnified1DU32FloatGrad:
244 case NVPTXISD::TexUnified1DArrayFloatS32:
245 case NVPTXISD::TexUnified1DArrayFloatFloat:
246 case NVPTXISD::TexUnified1DArrayFloatFloatLevel:
247 case NVPTXISD::TexUnified1DArrayFloatFloatGrad:
248 case NVPTXISD::TexUnified1DArrayS32S32:
249 case NVPTXISD::TexUnified1DArrayS32Float:
250 case NVPTXISD::TexUnified1DArrayS32FloatLevel:
251 case NVPTXISD::TexUnified1DArrayS32FloatGrad:
252 case NVPTXISD::TexUnified1DArrayU32S32:
253 case NVPTXISD::TexUnified1DArrayU32Float:
254 case NVPTXISD::TexUnified1DArrayU32FloatLevel:
255 case NVPTXISD::TexUnified1DArrayU32FloatGrad:
256 case NVPTXISD::TexUnified2DFloatS32:
257 case NVPTXISD::TexUnified2DFloatFloat:
258 case NVPTXISD::TexUnified2DFloatFloatLevel:
259 case NVPTXISD::TexUnified2DFloatFloatGrad:
260 case NVPTXISD::TexUnified2DS32S32:
261 case NVPTXISD::TexUnified2DS32Float:
262 case NVPTXISD::TexUnified2DS32FloatLevel:
263 case NVPTXISD::TexUnified2DS32FloatGrad:
264 case NVPTXISD::TexUnified2DU32S32:
265 case NVPTXISD::TexUnified2DU32Float:
266 case NVPTXISD::TexUnified2DU32FloatLevel:
267 case NVPTXISD::TexUnified2DU32FloatGrad:
268 case NVPTXISD::TexUnified2DArrayFloatS32:
269 case NVPTXISD::TexUnified2DArrayFloatFloat:
270 case NVPTXISD::TexUnified2DArrayFloatFloatLevel:
271 case NVPTXISD::TexUnified2DArrayFloatFloatGrad:
272 case NVPTXISD::TexUnified2DArrayS32S32:
273 case NVPTXISD::TexUnified2DArrayS32Float:
274 case NVPTXISD::TexUnified2DArrayS32FloatLevel:
275 case NVPTXISD::TexUnified2DArrayS32FloatGrad:
276 case NVPTXISD::TexUnified2DArrayU32S32:
277 case NVPTXISD::TexUnified2DArrayU32Float:
278 case NVPTXISD::TexUnified2DArrayU32FloatLevel:
279 case NVPTXISD::TexUnified2DArrayU32FloatGrad:
280 case NVPTXISD::TexUnified3DFloatS32:
281 case NVPTXISD::TexUnified3DFloatFloat:
282 case NVPTXISD::TexUnified3DFloatFloatLevel:
283 case NVPTXISD::TexUnified3DFloatFloatGrad:
284 case NVPTXISD::TexUnified3DS32S32:
285 case NVPTXISD::TexUnified3DS32Float:
286 case NVPTXISD::TexUnified3DS32FloatLevel:
287 case NVPTXISD::TexUnified3DS32FloatGrad:
288 case NVPTXISD::TexUnified3DU32S32:
289 case NVPTXISD::TexUnified3DU32Float:
290 case NVPTXISD::TexUnified3DU32FloatLevel:
291 case NVPTXISD::TexUnified3DU32FloatGrad:
292 case NVPTXISD::TexUnifiedCubeFloatFloat:
293 case NVPTXISD::TexUnifiedCubeFloatFloatLevel:
294 case NVPTXISD::TexUnifiedCubeS32Float:
295 case NVPTXISD::TexUnifiedCubeS32FloatLevel:
296 case NVPTXISD::TexUnifiedCubeU32Float:
297 case NVPTXISD::TexUnifiedCubeU32FloatLevel:
298 case NVPTXISD::TexUnifiedCubeArrayFloatFloat:
299 case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel:
300 case NVPTXISD::TexUnifiedCubeArrayS32Float:
301 case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel:
302 case NVPTXISD::TexUnifiedCubeArrayU32Float:
303 case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel:
304 case NVPTXISD::Tld4UnifiedR2DFloatFloat:
305 case NVPTXISD::Tld4UnifiedG2DFloatFloat:
306 case NVPTXISD::Tld4UnifiedB2DFloatFloat:
307 case NVPTXISD::Tld4UnifiedA2DFloatFloat:
308 case NVPTXISD::Tld4UnifiedR2DS64Float:
309 case NVPTXISD::Tld4UnifiedG2DS64Float:
310 case NVPTXISD::Tld4UnifiedB2DS64Float:
311 case NVPTXISD::Tld4UnifiedA2DS64Float:
312 case NVPTXISD::Tld4UnifiedR2DU64Float:
313 case NVPTXISD::Tld4UnifiedG2DU64Float:
314 case NVPTXISD::Tld4UnifiedB2DU64Float:
315 case NVPTXISD::Tld4UnifiedA2DU64Float:
316 if (tryTextureIntrinsic(N))
319 case NVPTXISD::Suld1DI8Clamp:
320 case NVPTXISD::Suld1DI16Clamp:
321 case NVPTXISD::Suld1DI32Clamp:
322 case NVPTXISD::Suld1DI64Clamp:
323 case NVPTXISD::Suld1DV2I8Clamp:
324 case NVPTXISD::Suld1DV2I16Clamp:
325 case NVPTXISD::Suld1DV2I32Clamp:
326 case NVPTXISD::Suld1DV2I64Clamp:
327 case NVPTXISD::Suld1DV4I8Clamp:
328 case NVPTXISD::Suld1DV4I16Clamp:
329 case NVPTXISD::Suld1DV4I32Clamp:
330 case NVPTXISD::Suld1DArrayI8Clamp:
331 case NVPTXISD::Suld1DArrayI16Clamp:
332 case NVPTXISD::Suld1DArrayI32Clamp:
333 case NVPTXISD::Suld1DArrayI64Clamp:
334 case NVPTXISD::Suld1DArrayV2I8Clamp:
335 case NVPTXISD::Suld1DArrayV2I16Clamp:
336 case NVPTXISD::Suld1DArrayV2I32Clamp:
337 case NVPTXISD::Suld1DArrayV2I64Clamp:
338 case NVPTXISD::Suld1DArrayV4I8Clamp:
339 case NVPTXISD::Suld1DArrayV4I16Clamp:
340 case NVPTXISD::Suld1DArrayV4I32Clamp:
341 case NVPTXISD::Suld2DI8Clamp:
342 case NVPTXISD::Suld2DI16Clamp:
343 case NVPTXISD::Suld2DI32Clamp:
344 case NVPTXISD::Suld2DI64Clamp:
345 case NVPTXISD::Suld2DV2I8Clamp:
346 case NVPTXISD::Suld2DV2I16Clamp:
347 case NVPTXISD::Suld2DV2I32Clamp:
348 case NVPTXISD::Suld2DV2I64Clamp:
349 case NVPTXISD::Suld2DV4I8Clamp:
350 case NVPTXISD::Suld2DV4I16Clamp:
351 case NVPTXISD::Suld2DV4I32Clamp:
352 case NVPTXISD::Suld2DArrayI8Clamp:
353 case NVPTXISD::Suld2DArrayI16Clamp:
354 case NVPTXISD::Suld2DArrayI32Clamp:
355 case NVPTXISD::Suld2DArrayI64Clamp:
356 case NVPTXISD::Suld2DArrayV2I8Clamp:
357 case NVPTXISD::Suld2DArrayV2I16Clamp:
358 case NVPTXISD::Suld2DArrayV2I32Clamp:
359 case NVPTXISD::Suld2DArrayV2I64Clamp:
360 case NVPTXISD::Suld2DArrayV4I8Clamp:
361 case NVPTXISD::Suld2DArrayV4I16Clamp:
362 case NVPTXISD::Suld2DArrayV4I32Clamp:
363 case NVPTXISD::Suld3DI8Clamp:
364 case NVPTXISD::Suld3DI16Clamp:
365 case NVPTXISD::Suld3DI32Clamp:
366 case NVPTXISD::Suld3DI64Clamp:
367 case NVPTXISD::Suld3DV2I8Clamp:
368 case NVPTXISD::Suld3DV2I16Clamp:
369 case NVPTXISD::Suld3DV2I32Clamp:
370 case NVPTXISD::Suld3DV2I64Clamp:
371 case NVPTXISD::Suld3DV4I8Clamp:
372 case NVPTXISD::Suld3DV4I16Clamp:
373 case NVPTXISD::Suld3DV4I32Clamp:
374 case NVPTXISD::Suld1DI8Trap:
375 case NVPTXISD::Suld1DI16Trap:
376 case NVPTXISD::Suld1DI32Trap:
377 case NVPTXISD::Suld1DI64Trap:
378 case NVPTXISD::Suld1DV2I8Trap:
379 case NVPTXISD::Suld1DV2I16Trap:
380 case NVPTXISD::Suld1DV2I32Trap:
381 case NVPTXISD::Suld1DV2I64Trap:
382 case NVPTXISD::Suld1DV4I8Trap:
383 case NVPTXISD::Suld1DV4I16Trap:
384 case NVPTXISD::Suld1DV4I32Trap:
385 case NVPTXISD::Suld1DArrayI8Trap:
386 case NVPTXISD::Suld1DArrayI16Trap:
387 case NVPTXISD::Suld1DArrayI32Trap:
388 case NVPTXISD::Suld1DArrayI64Trap:
389 case NVPTXISD::Suld1DArrayV2I8Trap:
390 case NVPTXISD::Suld1DArrayV2I16Trap:
391 case NVPTXISD::Suld1DArrayV2I32Trap:
392 case NVPTXISD::Suld1DArrayV2I64Trap:
393 case NVPTXISD::Suld1DArrayV4I8Trap:
394 case NVPTXISD::Suld1DArrayV4I16Trap:
395 case NVPTXISD::Suld1DArrayV4I32Trap:
396 case NVPTXISD::Suld2DI8Trap:
397 case NVPTXISD::Suld2DI16Trap:
398 case NVPTXISD::Suld2DI32Trap:
399 case NVPTXISD::Suld2DI64Trap:
400 case NVPTXISD::Suld2DV2I8Trap:
401 case NVPTXISD::Suld2DV2I16Trap:
402 case NVPTXISD::Suld2DV2I32Trap:
403 case NVPTXISD::Suld2DV2I64Trap:
404 case NVPTXISD::Suld2DV4I8Trap:
405 case NVPTXISD::Suld2DV4I16Trap:
406 case NVPTXISD::Suld2DV4I32Trap:
407 case NVPTXISD::Suld2DArrayI8Trap:
408 case NVPTXISD::Suld2DArrayI16Trap:
409 case NVPTXISD::Suld2DArrayI32Trap:
410 case NVPTXISD::Suld2DArrayI64Trap:
411 case NVPTXISD::Suld2DArrayV2I8Trap:
412 case NVPTXISD::Suld2DArrayV2I16Trap:
413 case NVPTXISD::Suld2DArrayV2I32Trap:
414 case NVPTXISD::Suld2DArrayV2I64Trap:
415 case NVPTXISD::Suld2DArrayV4I8Trap:
416 case NVPTXISD::Suld2DArrayV4I16Trap:
417 case NVPTXISD::Suld2DArrayV4I32Trap:
418 case NVPTXISD::Suld3DI8Trap:
419 case NVPTXISD::Suld3DI16Trap:
420 case NVPTXISD::Suld3DI32Trap:
421 case NVPTXISD::Suld3DI64Trap:
422 case NVPTXISD::Suld3DV2I8Trap:
423 case NVPTXISD::Suld3DV2I16Trap:
424 case NVPTXISD::Suld3DV2I32Trap:
425 case NVPTXISD::Suld3DV2I64Trap:
426 case NVPTXISD::Suld3DV4I8Trap:
427 case NVPTXISD::Suld3DV4I16Trap:
428 case NVPTXISD::Suld3DV4I32Trap:
429 case NVPTXISD::Suld1DI8Zero:
430 case NVPTXISD::Suld1DI16Zero:
431 case NVPTXISD::Suld1DI32Zero:
432 case NVPTXISD::Suld1DI64Zero:
433 case NVPTXISD::Suld1DV2I8Zero:
434 case NVPTXISD::Suld1DV2I16Zero:
435 case NVPTXISD::Suld1DV2I32Zero:
436 case NVPTXISD::Suld1DV2I64Zero:
437 case NVPTXISD::Suld1DV4I8Zero:
438 case NVPTXISD::Suld1DV4I16Zero:
439 case NVPTXISD::Suld1DV4I32Zero:
440 case NVPTXISD::Suld1DArrayI8Zero:
441 case NVPTXISD::Suld1DArrayI16Zero:
442 case NVPTXISD::Suld1DArrayI32Zero:
443 case NVPTXISD::Suld1DArrayI64Zero:
444 case NVPTXISD::Suld1DArrayV2I8Zero:
445 case NVPTXISD::Suld1DArrayV2I16Zero:
446 case NVPTXISD::Suld1DArrayV2I32Zero:
447 case NVPTXISD::Suld1DArrayV2I64Zero:
448 case NVPTXISD::Suld1DArrayV4I8Zero:
449 case NVPTXISD::Suld1DArrayV4I16Zero:
450 case NVPTXISD::Suld1DArrayV4I32Zero:
451 case NVPTXISD::Suld2DI8Zero:
452 case NVPTXISD::Suld2DI16Zero:
453 case NVPTXISD::Suld2DI32Zero:
454 case NVPTXISD::Suld2DI64Zero:
455 case NVPTXISD::Suld2DV2I8Zero:
456 case NVPTXISD::Suld2DV2I16Zero:
457 case NVPTXISD::Suld2DV2I32Zero:
458 case NVPTXISD::Suld2DV2I64Zero:
459 case NVPTXISD::Suld2DV4I8Zero:
460 case NVPTXISD::Suld2DV4I16Zero:
461 case NVPTXISD::Suld2DV4I32Zero:
462 case NVPTXISD::Suld2DArrayI8Zero:
463 case NVPTXISD::Suld2DArrayI16Zero:
464 case NVPTXISD::Suld2DArrayI32Zero:
465 case NVPTXISD::Suld2DArrayI64Zero:
466 case NVPTXISD::Suld2DArrayV2I8Zero:
467 case NVPTXISD::Suld2DArrayV2I16Zero:
468 case NVPTXISD::Suld2DArrayV2I32Zero:
469 case NVPTXISD::Suld2DArrayV2I64Zero:
470 case NVPTXISD::Suld2DArrayV4I8Zero:
471 case NVPTXISD::Suld2DArrayV4I16Zero:
472 case NVPTXISD::Suld2DArrayV4I32Zero:
473 case NVPTXISD::Suld3DI8Zero:
474 case NVPTXISD::Suld3DI16Zero:
475 case NVPTXISD::Suld3DI32Zero:
476 case NVPTXISD::Suld3DI64Zero:
477 case NVPTXISD::Suld3DV2I8Zero:
478 case NVPTXISD::Suld3DV2I16Zero:
479 case NVPTXISD::Suld3DV2I32Zero:
480 case NVPTXISD::Suld3DV2I64Zero:
481 case NVPTXISD::Suld3DV4I8Zero:
482 case NVPTXISD::Suld3DV4I16Zero:
483 case NVPTXISD::Suld3DV4I32Zero:
484 if (trySurfaceIntrinsic(N))
494 case ISD::ADDRSPACECAST:
495 SelectAddrSpaceCast(N);
497 case ISD::ConstantFP:
498 if (tryConstantFP16(N))
507 bool NVPTXDAGToDAGISel::tryIntrinsicChain(SDNode *N) {
508 unsigned IID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
512 case Intrinsic::nvvm_ldg_global_f:
513 case Intrinsic::nvvm_ldg_global_i:
514 case Intrinsic::nvvm_ldg_global_p:
515 case Intrinsic::nvvm_ldu_global_f:
516 case Intrinsic::nvvm_ldu_global_i:
517 case Intrinsic::nvvm_ldu_global_p:
522 // There's no way to specify FP16 immediates in .f16 ops, so we have to
523 // load them into an .f16 register first.
524 bool NVPTXDAGToDAGISel::tryConstantFP16(SDNode *N) {
525 if (N->getValueType(0) != MVT::f16)
527 SDValue Val = CurDAG->getTargetConstantFP(
528 cast<ConstantFPSDNode>(N)->getValueAPF(), SDLoc(N), MVT::f16);
529 SDNode *LoadConstF16 =
530 CurDAG->getMachineNode(NVPTX::LOAD_CONST_F16, SDLoc(N), MVT::f16, Val);
531 ReplaceNode(N, LoadConstF16);
535 // Map ISD:CONDCODE value to appropriate CmpMode expected by
536 // NVPTXInstPrinter::printCmpMode()
537 static unsigned getPTXCmpMode(const CondCodeSDNode &CondCode, bool FTZ) {
538 using NVPTX::PTXCmpMode::CmpMode;
539 unsigned PTXCmpMode = [](ISD::CondCode CC) {
542 llvm_unreachable("Unexpected condition code.");
558 return CmpMode::NotANumber;
587 PTXCmpMode |= NVPTX::PTXCmpMode::FTZ_FLAG;
592 bool NVPTXDAGToDAGISel::SelectSETP_F16X2(SDNode *N) {
593 unsigned PTXCmpMode =
594 getPTXCmpMode(*cast<CondCodeSDNode>(N->getOperand(2)), useF32FTZ());
596 SDNode *SetP = CurDAG->getMachineNode(
597 NVPTX::SETP_f16x2rr, DL, MVT::i1, MVT::i1, N->getOperand(0),
598 N->getOperand(1), CurDAG->getTargetConstant(PTXCmpMode, DL, MVT::i32));
599 ReplaceNode(N, SetP);
603 // Find all instances of extract_vector_elt that use this v2f16 vector
604 // and coalesce them into a scattering move instruction.
605 bool NVPTXDAGToDAGISel::tryEXTRACT_VECTOR_ELEMENT(SDNode *N) {
606 SDValue Vector = N->getOperand(0);
608 // We only care about f16x2 as it's the only real vector type we
609 // need to deal with.
610 if (Vector.getSimpleValueType() != MVT::v2f16)
613 // Find and record all uses of this vector that extract element 0 or 1.
614 SmallVector<SDNode *, 4> E0, E1;
615 for (auto U : Vector.getNode()->uses()) {
616 if (U->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
618 if (U->getOperand(0) != Vector)
620 if (const ConstantSDNode *IdxConst =
621 dyn_cast<ConstantSDNode>(U->getOperand(1))) {
622 if (IdxConst->getZExtValue() == 0)
624 else if (IdxConst->getZExtValue() == 1)
627 llvm_unreachable("Invalid vector index.");
631 // There's no point scattering f16x2 if we only ever access one
633 if (E0.empty() || E1.empty())
636 unsigned Op = NVPTX::SplitF16x2;
637 // If the vector has been BITCAST'ed from i32, we can use original
638 // value directly and avoid register-to-register move.
639 SDValue Source = Vector;
640 if (Vector->getOpcode() == ISD::BITCAST) {
641 Op = NVPTX::SplitI32toF16x2;
642 Source = Vector->getOperand(0);
644 // Merge (f16 extractelt(V, 0), f16 extractelt(V,1))
645 // into f16,f16 SplitF16x2(V)
647 CurDAG->getMachineNode(Op, SDLoc(N), MVT::f16, MVT::f16, Source);
648 for (auto *Node : E0)
649 ReplaceUses(SDValue(Node, 0), SDValue(ScatterOp, 0));
650 for (auto *Node : E1)
651 ReplaceUses(SDValue(Node, 0), SDValue(ScatterOp, 1));
656 static unsigned int getCodeAddrSpace(MemSDNode *N) {
657 const Value *Src = N->getMemOperand()->getValue();
660 return NVPTX::PTXLdStInstCode::GENERIC;
662 if (auto *PT = dyn_cast<PointerType>(Src->getType())) {
663 switch (PT->getAddressSpace()) {
664 case llvm::ADDRESS_SPACE_LOCAL: return NVPTX::PTXLdStInstCode::LOCAL;
665 case llvm::ADDRESS_SPACE_GLOBAL: return NVPTX::PTXLdStInstCode::GLOBAL;
666 case llvm::ADDRESS_SPACE_SHARED: return NVPTX::PTXLdStInstCode::SHARED;
667 case llvm::ADDRESS_SPACE_GENERIC: return NVPTX::PTXLdStInstCode::GENERIC;
668 case llvm::ADDRESS_SPACE_PARAM: return NVPTX::PTXLdStInstCode::PARAM;
669 case llvm::ADDRESS_SPACE_CONST: return NVPTX::PTXLdStInstCode::CONSTANT;
673 return NVPTX::PTXLdStInstCode::GENERIC;
676 static bool canLowerToLDG(MemSDNode *N, const NVPTXSubtarget &Subtarget,
677 unsigned CodeAddrSpace, MachineFunction *F) {
678 // We use ldg (i.e. ld.global.nc) for invariant loads from the global address
681 // We have two ways of identifying invariant loads: Loads may be explicitly
682 // marked as invariant, or we may infer them to be invariant.
684 // We currently infer invariance for loads from
685 // - constant global variables, and
686 // - kernel function pointer params that are noalias (i.e. __restrict) and
689 // TODO: Perform a more powerful invariance analysis (ideally IPO, and ideally
690 // not during the SelectionDAG phase).
692 // TODO: Infer invariance only at -O2. We still want to use ldg at -O0 for
693 // explicitly invariant loads because these are how clang tells us to use ldg
694 // when the user uses a builtin.
695 if (!Subtarget.hasLDG() || CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL)
698 if (N->isInvariant())
701 bool IsKernelFn = isKernelFunction(F->getFunction());
703 // We use GetUnderlyingObjects() here instead of GetUnderlyingObject() mainly
704 // because the former looks through phi nodes while the latter does not. We
705 // need to look through phi nodes to handle pointer induction variables.
706 SmallVector<const Value *, 8> Objs;
707 GetUnderlyingObjects(N->getMemOperand()->getValue(),
708 Objs, F->getDataLayout());
710 return all_of(Objs, [&](const Value *V) {
711 if (auto *A = dyn_cast<const Argument>(V))
712 return IsKernelFn && A->onlyReadsMemory() && A->hasNoAliasAttr();
713 if (auto *GV = dyn_cast<const GlobalVariable>(V))
714 return GV->isConstant();
719 bool NVPTXDAGToDAGISel::tryIntrinsicNoChain(SDNode *N) {
720 unsigned IID = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
724 case Intrinsic::nvvm_texsurf_handle_internal:
725 SelectTexSurfHandle(N);
730 void NVPTXDAGToDAGISel::SelectTexSurfHandle(SDNode *N) {
731 // Op 0 is the intrinsic ID
732 SDValue Wrapper = N->getOperand(1);
733 SDValue GlobalVal = Wrapper.getOperand(0);
734 ReplaceNode(N, CurDAG->getMachineNode(NVPTX::texsurf_handles, SDLoc(N),
735 MVT::i64, GlobalVal));
738 void NVPTXDAGToDAGISel::SelectAddrSpaceCast(SDNode *N) {
739 SDValue Src = N->getOperand(0);
740 AddrSpaceCastSDNode *CastN = cast<AddrSpaceCastSDNode>(N);
741 unsigned SrcAddrSpace = CastN->getSrcAddressSpace();
742 unsigned DstAddrSpace = CastN->getDestAddressSpace();
743 assert(SrcAddrSpace != DstAddrSpace &&
744 "addrspacecast must be between different address spaces");
746 if (DstAddrSpace == ADDRESS_SPACE_GENERIC) {
747 // Specific to generic
749 switch (SrcAddrSpace) {
750 default: report_fatal_error("Bad address space in addrspacecast");
751 case ADDRESS_SPACE_GLOBAL:
752 Opc = TM.is64Bit() ? NVPTX::cvta_global_yes_64 : NVPTX::cvta_global_yes;
754 case ADDRESS_SPACE_SHARED:
755 Opc = TM.is64Bit() ? (useShortPointers() ? NVPTX::cvta_shared_yes_6432
756 : NVPTX::cvta_shared_yes_64)
757 : NVPTX::cvta_shared_yes;
759 case ADDRESS_SPACE_CONST:
760 Opc = TM.is64Bit() ? (useShortPointers() ? NVPTX::cvta_const_yes_6432
761 : NVPTX::cvta_const_yes_64)
762 : NVPTX::cvta_const_yes;
764 case ADDRESS_SPACE_LOCAL:
765 Opc = TM.is64Bit() ? (useShortPointers() ? NVPTX::cvta_local_yes_6432
766 : NVPTX::cvta_local_yes_64)
767 : NVPTX::cvta_local_yes;
770 ReplaceNode(N, CurDAG->getMachineNode(Opc, SDLoc(N), N->getValueType(0),
774 // Generic to specific
775 if (SrcAddrSpace != 0)
776 report_fatal_error("Cannot cast between two non-generic address spaces");
778 switch (DstAddrSpace) {
779 default: report_fatal_error("Bad address space in addrspacecast");
780 case ADDRESS_SPACE_GLOBAL:
781 Opc = TM.is64Bit() ? NVPTX::cvta_to_global_yes_64
782 : NVPTX::cvta_to_global_yes;
784 case ADDRESS_SPACE_SHARED:
785 Opc = TM.is64Bit() ? (useShortPointers() ? NVPTX::cvta_to_shared_yes_3264
786 : NVPTX::cvta_to_shared_yes_64)
787 : NVPTX::cvta_to_shared_yes;
789 case ADDRESS_SPACE_CONST:
790 Opc = TM.is64Bit() ? (useShortPointers() ? NVPTX::cvta_to_const_yes_3264
791 : NVPTX::cvta_to_const_yes_64)
792 : NVPTX::cvta_to_const_yes;
794 case ADDRESS_SPACE_LOCAL:
795 Opc = TM.is64Bit() ? (useShortPointers() ? NVPTX::cvta_to_local_yes_3264
796 : NVPTX::cvta_to_local_yes_64)
797 : NVPTX::cvta_to_local_yes;
799 case ADDRESS_SPACE_PARAM:
800 Opc = TM.is64Bit() ? NVPTX::nvvm_ptr_gen_to_param_64
801 : NVPTX::nvvm_ptr_gen_to_param;
804 ReplaceNode(N, CurDAG->getMachineNode(Opc, SDLoc(N), N->getValueType(0),
810 // Helper function template to reduce amount of boilerplate code for
812 static Optional<unsigned> pickOpcodeForVT(
813 MVT::SimpleValueType VT, unsigned Opcode_i8, unsigned Opcode_i16,
814 unsigned Opcode_i32, Optional<unsigned> Opcode_i64, unsigned Opcode_f16,
815 unsigned Opcode_f16x2, unsigned Opcode_f32, Optional<unsigned> Opcode_f64) {
839 bool NVPTXDAGToDAGISel::tryLoad(SDNode *N) {
841 MemSDNode *LD = cast<MemSDNode>(N);
842 assert(LD->readMem() && "Expected load");
843 LoadSDNode *PlainLoad = dyn_cast<LoadSDNode>(N);
844 EVT LoadedVT = LD->getMemoryVT();
845 SDNode *NVPTXLD = nullptr;
847 // do not support pre/post inc/dec
848 if (PlainLoad && PlainLoad->isIndexed())
851 if (!LoadedVT.isSimple())
854 AtomicOrdering Ordering = LD->getOrdering();
855 // In order to lower atomic loads with stronger guarantees we would need to
856 // use load.acquire or insert fences. However these features were only added
857 // with PTX ISA 6.0 / sm_70.
858 // TODO: Check if we can actually use the new instructions and implement them.
859 if (isStrongerThanMonotonic(Ordering))
862 // Address Space Setting
863 unsigned int CodeAddrSpace = getCodeAddrSpace(LD);
864 if (canLowerToLDG(LD, *Subtarget, CodeAddrSpace, MF)) {
868 unsigned int PointerSize =
869 CurDAG->getDataLayout().getPointerSizeInBits(LD->getAddressSpace());
872 // - .volatile is only available for .global and .shared
873 // - .volatile has the same memory synchronization semantics as .relaxed.sys
874 bool isVolatile = LD->isVolatile() || Ordering == AtomicOrdering::Monotonic;
875 if (CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL &&
876 CodeAddrSpace != NVPTX::PTXLdStInstCode::SHARED &&
877 CodeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC)
880 // Type Setting: fromType + fromTypeWidth
882 // Sign : ISD::SEXTLOAD
883 // Unsign : ISD::ZEXTLOAD, ISD::NON_EXTLOAD or ISD::EXTLOAD and the
885 // Float : ISD::NON_EXTLOAD or ISD::EXTLOAD and the type is float
886 MVT SimpleVT = LoadedVT.getSimpleVT();
887 MVT ScalarVT = SimpleVT.getScalarType();
888 // Read at least 8 bits (predicates are stored as 8-bit values)
889 unsigned fromTypeWidth = std::max(8U, (unsigned)ScalarVT.getSizeInBits());
890 unsigned int fromType;
893 unsigned vecType = NVPTX::PTXLdStInstCode::Scalar;
894 if (SimpleVT.isVector()) {
895 assert(LoadedVT == MVT::v2f16 && "Unexpected vector type");
896 // v2f16 is loaded using ld.b32
900 if (PlainLoad && (PlainLoad->getExtensionType() == ISD::SEXTLOAD))
901 fromType = NVPTX::PTXLdStInstCode::Signed;
902 else if (ScalarVT.isFloatingPoint())
903 // f16 uses .b16 as its storage type.
904 fromType = ScalarVT.SimpleTy == MVT::f16 ? NVPTX::PTXLdStInstCode::Untyped
905 : NVPTX::PTXLdStInstCode::Float;
907 fromType = NVPTX::PTXLdStInstCode::Unsigned;
909 // Create the machine instruction DAG
910 SDValue Chain = N->getOperand(0);
911 SDValue N1 = N->getOperand(1);
913 SDValue Offset, Base;
914 Optional<unsigned> Opcode;
915 MVT::SimpleValueType TargetVT = LD->getSimpleValueType(0).SimpleTy;
917 if (SelectDirectAddr(N1, Addr)) {
918 Opcode = pickOpcodeForVT(
919 TargetVT, NVPTX::LD_i8_avar, NVPTX::LD_i16_avar, NVPTX::LD_i32_avar,
920 NVPTX::LD_i64_avar, NVPTX::LD_f16_avar, NVPTX::LD_f16x2_avar,
921 NVPTX::LD_f32_avar, NVPTX::LD_f64_avar);
924 SDValue Ops[] = { getI32Imm(isVolatile, dl), getI32Imm(CodeAddrSpace, dl),
925 getI32Imm(vecType, dl), getI32Imm(fromType, dl),
926 getI32Imm(fromTypeWidth, dl), Addr, Chain };
927 NVPTXLD = CurDAG->getMachineNode(Opcode.getValue(), dl, TargetVT,
929 } else if (PointerSize == 64 ? SelectADDRsi64(N1.getNode(), N1, Base, Offset)
930 : SelectADDRsi(N1.getNode(), N1, Base, Offset)) {
931 Opcode = pickOpcodeForVT(TargetVT, NVPTX::LD_i8_asi, NVPTX::LD_i16_asi,
932 NVPTX::LD_i32_asi, NVPTX::LD_i64_asi,
933 NVPTX::LD_f16_asi, NVPTX::LD_f16x2_asi,
934 NVPTX::LD_f32_asi, NVPTX::LD_f64_asi);
937 SDValue Ops[] = { getI32Imm(isVolatile, dl), getI32Imm(CodeAddrSpace, dl),
938 getI32Imm(vecType, dl), getI32Imm(fromType, dl),
939 getI32Imm(fromTypeWidth, dl), Base, Offset, Chain };
940 NVPTXLD = CurDAG->getMachineNode(Opcode.getValue(), dl, TargetVT,
942 } else if (PointerSize == 64 ? SelectADDRri64(N1.getNode(), N1, Base, Offset)
943 : SelectADDRri(N1.getNode(), N1, Base, Offset)) {
944 if (PointerSize == 64)
945 Opcode = pickOpcodeForVT(
946 TargetVT, NVPTX::LD_i8_ari_64, NVPTX::LD_i16_ari_64,
947 NVPTX::LD_i32_ari_64, NVPTX::LD_i64_ari_64, NVPTX::LD_f16_ari_64,
948 NVPTX::LD_f16x2_ari_64, NVPTX::LD_f32_ari_64, NVPTX::LD_f64_ari_64);
950 Opcode = pickOpcodeForVT(
951 TargetVT, NVPTX::LD_i8_ari, NVPTX::LD_i16_ari, NVPTX::LD_i32_ari,
952 NVPTX::LD_i64_ari, NVPTX::LD_f16_ari, NVPTX::LD_f16x2_ari,
953 NVPTX::LD_f32_ari, NVPTX::LD_f64_ari);
956 SDValue Ops[] = { getI32Imm(isVolatile, dl), getI32Imm(CodeAddrSpace, dl),
957 getI32Imm(vecType, dl), getI32Imm(fromType, dl),
958 getI32Imm(fromTypeWidth, dl), Base, Offset, Chain };
959 NVPTXLD = CurDAG->getMachineNode(Opcode.getValue(), dl, TargetVT,
962 if (PointerSize == 64)
963 Opcode = pickOpcodeForVT(
964 TargetVT, NVPTX::LD_i8_areg_64, NVPTX::LD_i16_areg_64,
965 NVPTX::LD_i32_areg_64, NVPTX::LD_i64_areg_64, NVPTX::LD_f16_areg_64,
966 NVPTX::LD_f16x2_areg_64, NVPTX::LD_f32_areg_64,
967 NVPTX::LD_f64_areg_64);
969 Opcode = pickOpcodeForVT(
970 TargetVT, NVPTX::LD_i8_areg, NVPTX::LD_i16_areg, NVPTX::LD_i32_areg,
971 NVPTX::LD_i64_areg, NVPTX::LD_f16_areg, NVPTX::LD_f16x2_areg,
972 NVPTX::LD_f32_areg, NVPTX::LD_f64_areg);
975 SDValue Ops[] = { getI32Imm(isVolatile, dl), getI32Imm(CodeAddrSpace, dl),
976 getI32Imm(vecType, dl), getI32Imm(fromType, dl),
977 getI32Imm(fromTypeWidth, dl), N1, Chain };
978 NVPTXLD = CurDAG->getMachineNode(Opcode.getValue(), dl, TargetVT,
985 MachineMemOperand *MemRef = cast<MemSDNode>(N)->getMemOperand();
986 CurDAG->setNodeMemRefs(cast<MachineSDNode>(NVPTXLD), {MemRef});
988 ReplaceNode(N, NVPTXLD);
992 bool NVPTXDAGToDAGISel::tryLoadVector(SDNode *N) {
994 SDValue Chain = N->getOperand(0);
995 SDValue Op1 = N->getOperand(1);
996 SDValue Addr, Offset, Base;
997 Optional<unsigned> Opcode;
1000 MemSDNode *MemSD = cast<MemSDNode>(N);
1001 EVT LoadedVT = MemSD->getMemoryVT();
1003 if (!LoadedVT.isSimple())
1006 // Address Space Setting
1007 unsigned int CodeAddrSpace = getCodeAddrSpace(MemSD);
1008 if (canLowerToLDG(MemSD, *Subtarget, CodeAddrSpace, MF)) {
1009 return tryLDGLDU(N);
1012 unsigned int PointerSize =
1013 CurDAG->getDataLayout().getPointerSizeInBits(MemSD->getAddressSpace());
1016 // - .volatile is only availalble for .global and .shared
1017 bool IsVolatile = MemSD->isVolatile();
1018 if (CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL &&
1019 CodeAddrSpace != NVPTX::PTXLdStInstCode::SHARED &&
1020 CodeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC)
1024 MVT SimpleVT = LoadedVT.getSimpleVT();
1026 // Type Setting: fromType + fromTypeWidth
1028 // Sign : ISD::SEXTLOAD
1029 // Unsign : ISD::ZEXTLOAD, ISD::NON_EXTLOAD or ISD::EXTLOAD and the
1031 // Float : ISD::NON_EXTLOAD or ISD::EXTLOAD and the type is float
1032 MVT ScalarVT = SimpleVT.getScalarType();
1033 // Read at least 8 bits (predicates are stored as 8-bit values)
1034 unsigned FromTypeWidth = std::max(8U, (unsigned)ScalarVT.getSizeInBits());
1035 unsigned int FromType;
1036 // The last operand holds the original LoadSDNode::getExtensionType() value
1037 unsigned ExtensionType = cast<ConstantSDNode>(
1038 N->getOperand(N->getNumOperands() - 1))->getZExtValue();
1039 if (ExtensionType == ISD::SEXTLOAD)
1040 FromType = NVPTX::PTXLdStInstCode::Signed;
1041 else if (ScalarVT.isFloatingPoint())
1042 FromType = ScalarVT.SimpleTy == MVT::f16 ? NVPTX::PTXLdStInstCode::Untyped
1043 : NVPTX::PTXLdStInstCode::Float;
1045 FromType = NVPTX::PTXLdStInstCode::Unsigned;
1049 switch (N->getOpcode()) {
1050 case NVPTXISD::LoadV2:
1051 VecType = NVPTX::PTXLdStInstCode::V2;
1053 case NVPTXISD::LoadV4:
1054 VecType = NVPTX::PTXLdStInstCode::V4;
1060 EVT EltVT = N->getValueType(0);
1062 // v8f16 is a special case. PTX doesn't have ld.v8.f16
1063 // instruction. Instead, we split the vector into v2f16 chunks and
1064 // load them with ld.v4.b32.
1065 if (EltVT == MVT::v2f16) {
1066 assert(N->getOpcode() == NVPTXISD::LoadV4 && "Unexpected load opcode.");
1068 FromType = NVPTX::PTXLdStInstCode::Untyped;
1072 if (SelectDirectAddr(Op1, Addr)) {
1073 switch (N->getOpcode()) {
1076 case NVPTXISD::LoadV2:
1077 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1078 NVPTX::LDV_i8_v2_avar, NVPTX::LDV_i16_v2_avar,
1079 NVPTX::LDV_i32_v2_avar, NVPTX::LDV_i64_v2_avar,
1080 NVPTX::LDV_f16_v2_avar, NVPTX::LDV_f16x2_v2_avar,
1081 NVPTX::LDV_f32_v2_avar, NVPTX::LDV_f64_v2_avar);
1083 case NVPTXISD::LoadV4:
1085 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v4_avar,
1086 NVPTX::LDV_i16_v4_avar, NVPTX::LDV_i32_v4_avar, None,
1087 NVPTX::LDV_f16_v4_avar, NVPTX::LDV_f16x2_v4_avar,
1088 NVPTX::LDV_f32_v4_avar, None);
1093 SDValue Ops[] = { getI32Imm(IsVolatile, DL), getI32Imm(CodeAddrSpace, DL),
1094 getI32Imm(VecType, DL), getI32Imm(FromType, DL),
1095 getI32Imm(FromTypeWidth, DL), Addr, Chain };
1096 LD = CurDAG->getMachineNode(Opcode.getValue(), DL, N->getVTList(), Ops);
1097 } else if (PointerSize == 64
1098 ? SelectADDRsi64(Op1.getNode(), Op1, Base, Offset)
1099 : SelectADDRsi(Op1.getNode(), Op1, Base, Offset)) {
1100 switch (N->getOpcode()) {
1103 case NVPTXISD::LoadV2:
1104 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1105 NVPTX::LDV_i8_v2_asi, NVPTX::LDV_i16_v2_asi,
1106 NVPTX::LDV_i32_v2_asi, NVPTX::LDV_i64_v2_asi,
1107 NVPTX::LDV_f16_v2_asi, NVPTX::LDV_f16x2_v2_asi,
1108 NVPTX::LDV_f32_v2_asi, NVPTX::LDV_f64_v2_asi);
1110 case NVPTXISD::LoadV4:
1112 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v4_asi,
1113 NVPTX::LDV_i16_v4_asi, NVPTX::LDV_i32_v4_asi, None,
1114 NVPTX::LDV_f16_v4_asi, NVPTX::LDV_f16x2_v4_asi,
1115 NVPTX::LDV_f32_v4_asi, None);
1120 SDValue Ops[] = { getI32Imm(IsVolatile, DL), getI32Imm(CodeAddrSpace, DL),
1121 getI32Imm(VecType, DL), getI32Imm(FromType, DL),
1122 getI32Imm(FromTypeWidth, DL), Base, Offset, Chain };
1123 LD = CurDAG->getMachineNode(Opcode.getValue(), DL, N->getVTList(), Ops);
1124 } else if (PointerSize == 64
1125 ? SelectADDRri64(Op1.getNode(), Op1, Base, Offset)
1126 : SelectADDRri(Op1.getNode(), Op1, Base, Offset)) {
1127 if (PointerSize == 64) {
1128 switch (N->getOpcode()) {
1131 case NVPTXISD::LoadV2:
1132 Opcode = pickOpcodeForVT(
1133 EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v2_ari_64,
1134 NVPTX::LDV_i16_v2_ari_64, NVPTX::LDV_i32_v2_ari_64,
1135 NVPTX::LDV_i64_v2_ari_64, NVPTX::LDV_f16_v2_ari_64,
1136 NVPTX::LDV_f16x2_v2_ari_64, NVPTX::LDV_f32_v2_ari_64,
1137 NVPTX::LDV_f64_v2_ari_64);
1139 case NVPTXISD::LoadV4:
1140 Opcode = pickOpcodeForVT(
1141 EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v4_ari_64,
1142 NVPTX::LDV_i16_v4_ari_64, NVPTX::LDV_i32_v4_ari_64, None,
1143 NVPTX::LDV_f16_v4_ari_64, NVPTX::LDV_f16x2_v4_ari_64,
1144 NVPTX::LDV_f32_v4_ari_64, None);
1148 switch (N->getOpcode()) {
1151 case NVPTXISD::LoadV2:
1152 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1153 NVPTX::LDV_i8_v2_ari, NVPTX::LDV_i16_v2_ari,
1154 NVPTX::LDV_i32_v2_ari, NVPTX::LDV_i64_v2_ari,
1155 NVPTX::LDV_f16_v2_ari, NVPTX::LDV_f16x2_v2_ari,
1156 NVPTX::LDV_f32_v2_ari, NVPTX::LDV_f64_v2_ari);
1158 case NVPTXISD::LoadV4:
1160 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v4_ari,
1161 NVPTX::LDV_i16_v4_ari, NVPTX::LDV_i32_v4_ari, None,
1162 NVPTX::LDV_f16_v4_ari, NVPTX::LDV_f16x2_v4_ari,
1163 NVPTX::LDV_f32_v4_ari, None);
1169 SDValue Ops[] = { getI32Imm(IsVolatile, DL), getI32Imm(CodeAddrSpace, DL),
1170 getI32Imm(VecType, DL), getI32Imm(FromType, DL),
1171 getI32Imm(FromTypeWidth, DL), Base, Offset, Chain };
1173 LD = CurDAG->getMachineNode(Opcode.getValue(), DL, N->getVTList(), Ops);
1175 if (PointerSize == 64) {
1176 switch (N->getOpcode()) {
1179 case NVPTXISD::LoadV2:
1180 Opcode = pickOpcodeForVT(
1181 EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v2_areg_64,
1182 NVPTX::LDV_i16_v2_areg_64, NVPTX::LDV_i32_v2_areg_64,
1183 NVPTX::LDV_i64_v2_areg_64, NVPTX::LDV_f16_v2_areg_64,
1184 NVPTX::LDV_f16x2_v2_areg_64, NVPTX::LDV_f32_v2_areg_64,
1185 NVPTX::LDV_f64_v2_areg_64);
1187 case NVPTXISD::LoadV4:
1188 Opcode = pickOpcodeForVT(
1189 EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v4_areg_64,
1190 NVPTX::LDV_i16_v4_areg_64, NVPTX::LDV_i32_v4_areg_64, None,
1191 NVPTX::LDV_f16_v4_areg_64, NVPTX::LDV_f16x2_v4_areg_64,
1192 NVPTX::LDV_f32_v4_areg_64, None);
1196 switch (N->getOpcode()) {
1199 case NVPTXISD::LoadV2:
1201 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v2_areg,
1202 NVPTX::LDV_i16_v2_areg, NVPTX::LDV_i32_v2_areg,
1203 NVPTX::LDV_i64_v2_areg, NVPTX::LDV_f16_v2_areg,
1204 NVPTX::LDV_f16x2_v2_areg, NVPTX::LDV_f32_v2_areg,
1205 NVPTX::LDV_f64_v2_areg);
1207 case NVPTXISD::LoadV4:
1208 Opcode = pickOpcodeForVT(
1209 EltVT.getSimpleVT().SimpleTy, NVPTX::LDV_i8_v4_areg,
1210 NVPTX::LDV_i16_v4_areg, NVPTX::LDV_i32_v4_areg, None,
1211 NVPTX::LDV_f16_v4_areg, NVPTX::LDV_f16x2_v4_areg,
1212 NVPTX::LDV_f32_v4_areg, None);
1218 SDValue Ops[] = { getI32Imm(IsVolatile, DL), getI32Imm(CodeAddrSpace, DL),
1219 getI32Imm(VecType, DL), getI32Imm(FromType, DL),
1220 getI32Imm(FromTypeWidth, DL), Op1, Chain };
1221 LD = CurDAG->getMachineNode(Opcode.getValue(), DL, N->getVTList(), Ops);
1224 MachineMemOperand *MemRef = cast<MemSDNode>(N)->getMemOperand();
1225 CurDAG->setNodeMemRefs(cast<MachineSDNode>(LD), {MemRef});
1231 bool NVPTXDAGToDAGISel::tryLDGLDU(SDNode *N) {
1233 SDValue Chain = N->getOperand(0);
1238 // If this is an LDG intrinsic, the address is the third operand. If its an
1239 // LDG/LDU SD node (from custom vector handling), then its the second operand
1240 if (N->getOpcode() == ISD::INTRINSIC_W_CHAIN) {
1241 Op1 = N->getOperand(2);
1242 Mem = cast<MemIntrinsicSDNode>(N);
1243 unsigned IID = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
1247 case Intrinsic::nvvm_ldg_global_f:
1248 case Intrinsic::nvvm_ldg_global_i:
1249 case Intrinsic::nvvm_ldg_global_p:
1252 case Intrinsic::nvvm_ldu_global_f:
1253 case Intrinsic::nvvm_ldu_global_i:
1254 case Intrinsic::nvvm_ldu_global_p:
1259 Op1 = N->getOperand(1);
1260 Mem = cast<MemSDNode>(N);
1263 Optional<unsigned> Opcode;
1266 SDValue Base, Offset, Addr;
1268 EVT EltVT = Mem->getMemoryVT();
1269 unsigned NumElts = 1;
1270 if (EltVT.isVector()) {
1271 NumElts = EltVT.getVectorNumElements();
1272 EltVT = EltVT.getVectorElementType();
1273 // vectors of f16 are loaded/stored as multiples of v2f16 elements.
1274 if (EltVT == MVT::f16 && N->getValueType(0) == MVT::v2f16) {
1275 assert(NumElts % 2 == 0 && "Vector must have even number of elements");
1281 // Build the "promoted" result VTList for the load. If we are really loading
1282 // i8s, then the return type will be promoted to i16 since we do not expose
1283 // 8-bit registers in NVPTX.
1284 EVT NodeVT = (EltVT == MVT::i8) ? MVT::i16 : EltVT;
1285 SmallVector<EVT, 5> InstVTs;
1286 for (unsigned i = 0; i != NumElts; ++i) {
1287 InstVTs.push_back(NodeVT);
1289 InstVTs.push_back(MVT::Other);
1290 SDVTList InstVTList = CurDAG->getVTList(InstVTs);
1292 if (SelectDirectAddr(Op1, Addr)) {
1293 switch (N->getOpcode()) {
1297 case ISD::INTRINSIC_W_CHAIN:
1299 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1300 NVPTX::INT_PTX_LDG_GLOBAL_i8avar,
1301 NVPTX::INT_PTX_LDG_GLOBAL_i16avar,
1302 NVPTX::INT_PTX_LDG_GLOBAL_i32avar,
1303 NVPTX::INT_PTX_LDG_GLOBAL_i64avar,
1304 NVPTX::INT_PTX_LDG_GLOBAL_f16avar,
1305 NVPTX::INT_PTX_LDG_GLOBAL_f16x2avar,
1306 NVPTX::INT_PTX_LDG_GLOBAL_f32avar,
1307 NVPTX::INT_PTX_LDG_GLOBAL_f64avar);
1309 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1310 NVPTX::INT_PTX_LDU_GLOBAL_i8avar,
1311 NVPTX::INT_PTX_LDU_GLOBAL_i16avar,
1312 NVPTX::INT_PTX_LDU_GLOBAL_i32avar,
1313 NVPTX::INT_PTX_LDU_GLOBAL_i64avar,
1314 NVPTX::INT_PTX_LDU_GLOBAL_f16avar,
1315 NVPTX::INT_PTX_LDU_GLOBAL_f16x2avar,
1316 NVPTX::INT_PTX_LDU_GLOBAL_f32avar,
1317 NVPTX::INT_PTX_LDU_GLOBAL_f64avar);
1319 case NVPTXISD::LoadV2:
1320 case NVPTXISD::LDGV2:
1321 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1322 NVPTX::INT_PTX_LDG_G_v2i8_ELE_avar,
1323 NVPTX::INT_PTX_LDG_G_v2i16_ELE_avar,
1324 NVPTX::INT_PTX_LDG_G_v2i32_ELE_avar,
1325 NVPTX::INT_PTX_LDG_G_v2i64_ELE_avar,
1326 NVPTX::INT_PTX_LDG_G_v2f16_ELE_avar,
1327 NVPTX::INT_PTX_LDG_G_v2f16x2_ELE_avar,
1328 NVPTX::INT_PTX_LDG_G_v2f32_ELE_avar,
1329 NVPTX::INT_PTX_LDG_G_v2f64_ELE_avar);
1331 case NVPTXISD::LDUV2:
1332 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1333 NVPTX::INT_PTX_LDU_G_v2i8_ELE_avar,
1334 NVPTX::INT_PTX_LDU_G_v2i16_ELE_avar,
1335 NVPTX::INT_PTX_LDU_G_v2i32_ELE_avar,
1336 NVPTX::INT_PTX_LDU_G_v2i64_ELE_avar,
1337 NVPTX::INT_PTX_LDU_G_v2f16_ELE_avar,
1338 NVPTX::INT_PTX_LDU_G_v2f16x2_ELE_avar,
1339 NVPTX::INT_PTX_LDU_G_v2f32_ELE_avar,
1340 NVPTX::INT_PTX_LDU_G_v2f64_ELE_avar);
1342 case NVPTXISD::LoadV4:
1343 case NVPTXISD::LDGV4:
1344 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1345 NVPTX::INT_PTX_LDG_G_v4i8_ELE_avar,
1346 NVPTX::INT_PTX_LDG_G_v4i16_ELE_avar,
1347 NVPTX::INT_PTX_LDG_G_v4i32_ELE_avar, None,
1348 NVPTX::INT_PTX_LDG_G_v4f16_ELE_avar,
1349 NVPTX::INT_PTX_LDG_G_v4f16x2_ELE_avar,
1350 NVPTX::INT_PTX_LDG_G_v4f32_ELE_avar, None);
1352 case NVPTXISD::LDUV4:
1353 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1354 NVPTX::INT_PTX_LDU_G_v4i8_ELE_avar,
1355 NVPTX::INT_PTX_LDU_G_v4i16_ELE_avar,
1356 NVPTX::INT_PTX_LDU_G_v4i32_ELE_avar, None,
1357 NVPTX::INT_PTX_LDU_G_v4f16_ELE_avar,
1358 NVPTX::INT_PTX_LDU_G_v4f16x2_ELE_avar,
1359 NVPTX::INT_PTX_LDU_G_v4f32_ELE_avar, None);
1364 SDValue Ops[] = { Addr, Chain };
1365 LD = CurDAG->getMachineNode(Opcode.getValue(), DL, InstVTList, Ops);
1366 } else if (TM.is64Bit() ? SelectADDRri64(Op1.getNode(), Op1, Base, Offset)
1367 : SelectADDRri(Op1.getNode(), Op1, Base, Offset)) {
1369 switch (N->getOpcode()) {
1373 case ISD::INTRINSIC_W_CHAIN:
1375 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1376 NVPTX::INT_PTX_LDG_GLOBAL_i8ari64,
1377 NVPTX::INT_PTX_LDG_GLOBAL_i16ari64,
1378 NVPTX::INT_PTX_LDG_GLOBAL_i32ari64,
1379 NVPTX::INT_PTX_LDG_GLOBAL_i64ari64,
1380 NVPTX::INT_PTX_LDG_GLOBAL_f16ari64,
1381 NVPTX::INT_PTX_LDG_GLOBAL_f16x2ari64,
1382 NVPTX::INT_PTX_LDG_GLOBAL_f32ari64,
1383 NVPTX::INT_PTX_LDG_GLOBAL_f64ari64);
1385 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1386 NVPTX::INT_PTX_LDU_GLOBAL_i8ari64,
1387 NVPTX::INT_PTX_LDU_GLOBAL_i16ari64,
1388 NVPTX::INT_PTX_LDU_GLOBAL_i32ari64,
1389 NVPTX::INT_PTX_LDU_GLOBAL_i64ari64,
1390 NVPTX::INT_PTX_LDU_GLOBAL_f16ari64,
1391 NVPTX::INT_PTX_LDU_GLOBAL_f16x2ari64,
1392 NVPTX::INT_PTX_LDU_GLOBAL_f32ari64,
1393 NVPTX::INT_PTX_LDU_GLOBAL_f64ari64);
1395 case NVPTXISD::LoadV2:
1396 case NVPTXISD::LDGV2:
1397 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1398 NVPTX::INT_PTX_LDG_G_v2i8_ELE_ari64,
1399 NVPTX::INT_PTX_LDG_G_v2i16_ELE_ari64,
1400 NVPTX::INT_PTX_LDG_G_v2i32_ELE_ari64,
1401 NVPTX::INT_PTX_LDG_G_v2i64_ELE_ari64,
1402 NVPTX::INT_PTX_LDG_G_v2f16_ELE_ari64,
1403 NVPTX::INT_PTX_LDG_G_v2f16x2_ELE_ari64,
1404 NVPTX::INT_PTX_LDG_G_v2f32_ELE_ari64,
1405 NVPTX::INT_PTX_LDG_G_v2f64_ELE_ari64);
1407 case NVPTXISD::LDUV2:
1408 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1409 NVPTX::INT_PTX_LDU_G_v2i8_ELE_ari64,
1410 NVPTX::INT_PTX_LDU_G_v2i16_ELE_ari64,
1411 NVPTX::INT_PTX_LDU_G_v2i32_ELE_ari64,
1412 NVPTX::INT_PTX_LDU_G_v2i64_ELE_ari64,
1413 NVPTX::INT_PTX_LDU_G_v2f16_ELE_ari64,
1414 NVPTX::INT_PTX_LDU_G_v2f16x2_ELE_ari64,
1415 NVPTX::INT_PTX_LDU_G_v2f32_ELE_ari64,
1416 NVPTX::INT_PTX_LDU_G_v2f64_ELE_ari64);
1418 case NVPTXISD::LoadV4:
1419 case NVPTXISD::LDGV4:
1420 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1421 NVPTX::INT_PTX_LDG_G_v4i8_ELE_ari64,
1422 NVPTX::INT_PTX_LDG_G_v4i16_ELE_ari64,
1423 NVPTX::INT_PTX_LDG_G_v4i32_ELE_ari64, None,
1424 NVPTX::INT_PTX_LDG_G_v4f16_ELE_ari64,
1425 NVPTX::INT_PTX_LDG_G_v4f16x2_ELE_ari64,
1426 NVPTX::INT_PTX_LDG_G_v4f32_ELE_ari64, None);
1428 case NVPTXISD::LDUV4:
1429 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1430 NVPTX::INT_PTX_LDU_G_v4i8_ELE_ari64,
1431 NVPTX::INT_PTX_LDU_G_v4i16_ELE_ari64,
1432 NVPTX::INT_PTX_LDU_G_v4i32_ELE_ari64, None,
1433 NVPTX::INT_PTX_LDU_G_v4f16_ELE_ari64,
1434 NVPTX::INT_PTX_LDU_G_v4f16x2_ELE_ari64,
1435 NVPTX::INT_PTX_LDU_G_v4f32_ELE_ari64, None);
1439 switch (N->getOpcode()) {
1443 case ISD::INTRINSIC_W_CHAIN:
1445 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1446 NVPTX::INT_PTX_LDG_GLOBAL_i8ari,
1447 NVPTX::INT_PTX_LDG_GLOBAL_i16ari,
1448 NVPTX::INT_PTX_LDG_GLOBAL_i32ari,
1449 NVPTX::INT_PTX_LDG_GLOBAL_i64ari,
1450 NVPTX::INT_PTX_LDG_GLOBAL_f16ari,
1451 NVPTX::INT_PTX_LDG_GLOBAL_f16x2ari,
1452 NVPTX::INT_PTX_LDG_GLOBAL_f32ari,
1453 NVPTX::INT_PTX_LDG_GLOBAL_f64ari);
1455 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1456 NVPTX::INT_PTX_LDU_GLOBAL_i8ari,
1457 NVPTX::INT_PTX_LDU_GLOBAL_i16ari,
1458 NVPTX::INT_PTX_LDU_GLOBAL_i32ari,
1459 NVPTX::INT_PTX_LDU_GLOBAL_i64ari,
1460 NVPTX::INT_PTX_LDU_GLOBAL_f16ari,
1461 NVPTX::INT_PTX_LDU_GLOBAL_f16x2ari,
1462 NVPTX::INT_PTX_LDU_GLOBAL_f32ari,
1463 NVPTX::INT_PTX_LDU_GLOBAL_f64ari);
1465 case NVPTXISD::LoadV2:
1466 case NVPTXISD::LDGV2:
1467 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1468 NVPTX::INT_PTX_LDG_G_v2i8_ELE_ari32,
1469 NVPTX::INT_PTX_LDG_G_v2i16_ELE_ari32,
1470 NVPTX::INT_PTX_LDG_G_v2i32_ELE_ari32,
1471 NVPTX::INT_PTX_LDG_G_v2i64_ELE_ari32,
1472 NVPTX::INT_PTX_LDG_G_v2f16_ELE_ari32,
1473 NVPTX::INT_PTX_LDG_G_v2f16x2_ELE_ari32,
1474 NVPTX::INT_PTX_LDG_G_v2f32_ELE_ari32,
1475 NVPTX::INT_PTX_LDG_G_v2f64_ELE_ari32);
1477 case NVPTXISD::LDUV2:
1478 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1479 NVPTX::INT_PTX_LDU_G_v2i8_ELE_ari32,
1480 NVPTX::INT_PTX_LDU_G_v2i16_ELE_ari32,
1481 NVPTX::INT_PTX_LDU_G_v2i32_ELE_ari32,
1482 NVPTX::INT_PTX_LDU_G_v2i64_ELE_ari32,
1483 NVPTX::INT_PTX_LDU_G_v2f16_ELE_ari32,
1484 NVPTX::INT_PTX_LDU_G_v2f16x2_ELE_ari32,
1485 NVPTX::INT_PTX_LDU_G_v2f32_ELE_ari32,
1486 NVPTX::INT_PTX_LDU_G_v2f64_ELE_ari32);
1488 case NVPTXISD::LoadV4:
1489 case NVPTXISD::LDGV4:
1490 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1491 NVPTX::INT_PTX_LDG_G_v4i8_ELE_ari32,
1492 NVPTX::INT_PTX_LDG_G_v4i16_ELE_ari32,
1493 NVPTX::INT_PTX_LDG_G_v4i32_ELE_ari32, None,
1494 NVPTX::INT_PTX_LDG_G_v4f16_ELE_ari32,
1495 NVPTX::INT_PTX_LDG_G_v4f16x2_ELE_ari32,
1496 NVPTX::INT_PTX_LDG_G_v4f32_ELE_ari32, None);
1498 case NVPTXISD::LDUV4:
1499 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1500 NVPTX::INT_PTX_LDU_G_v4i8_ELE_ari32,
1501 NVPTX::INT_PTX_LDU_G_v4i16_ELE_ari32,
1502 NVPTX::INT_PTX_LDU_G_v4i32_ELE_ari32, None,
1503 NVPTX::INT_PTX_LDU_G_v4f16_ELE_ari32,
1504 NVPTX::INT_PTX_LDU_G_v4f16x2_ELE_ari32,
1505 NVPTX::INT_PTX_LDU_G_v4f32_ELE_ari32, None);
1511 SDValue Ops[] = {Base, Offset, Chain};
1512 LD = CurDAG->getMachineNode(Opcode.getValue(), DL, InstVTList, Ops);
1515 switch (N->getOpcode()) {
1519 case ISD::INTRINSIC_W_CHAIN:
1521 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1522 NVPTX::INT_PTX_LDG_GLOBAL_i8areg64,
1523 NVPTX::INT_PTX_LDG_GLOBAL_i16areg64,
1524 NVPTX::INT_PTX_LDG_GLOBAL_i32areg64,
1525 NVPTX::INT_PTX_LDG_GLOBAL_i64areg64,
1526 NVPTX::INT_PTX_LDG_GLOBAL_f16areg64,
1527 NVPTX::INT_PTX_LDG_GLOBAL_f16x2areg64,
1528 NVPTX::INT_PTX_LDG_GLOBAL_f32areg64,
1529 NVPTX::INT_PTX_LDG_GLOBAL_f64areg64);
1531 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1532 NVPTX::INT_PTX_LDU_GLOBAL_i8areg64,
1533 NVPTX::INT_PTX_LDU_GLOBAL_i16areg64,
1534 NVPTX::INT_PTX_LDU_GLOBAL_i32areg64,
1535 NVPTX::INT_PTX_LDU_GLOBAL_i64areg64,
1536 NVPTX::INT_PTX_LDU_GLOBAL_f16areg64,
1537 NVPTX::INT_PTX_LDU_GLOBAL_f16x2areg64,
1538 NVPTX::INT_PTX_LDU_GLOBAL_f32areg64,
1539 NVPTX::INT_PTX_LDU_GLOBAL_f64areg64);
1541 case NVPTXISD::LoadV2:
1542 case NVPTXISD::LDGV2:
1543 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1544 NVPTX::INT_PTX_LDG_G_v2i8_ELE_areg64,
1545 NVPTX::INT_PTX_LDG_G_v2i16_ELE_areg64,
1546 NVPTX::INT_PTX_LDG_G_v2i32_ELE_areg64,
1547 NVPTX::INT_PTX_LDG_G_v2i64_ELE_areg64,
1548 NVPTX::INT_PTX_LDG_G_v2f16_ELE_areg64,
1549 NVPTX::INT_PTX_LDG_G_v2f16x2_ELE_areg64,
1550 NVPTX::INT_PTX_LDG_G_v2f32_ELE_areg64,
1551 NVPTX::INT_PTX_LDG_G_v2f64_ELE_areg64);
1553 case NVPTXISD::LDUV2:
1554 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1555 NVPTX::INT_PTX_LDU_G_v2i8_ELE_areg64,
1556 NVPTX::INT_PTX_LDU_G_v2i16_ELE_areg64,
1557 NVPTX::INT_PTX_LDU_G_v2i32_ELE_areg64,
1558 NVPTX::INT_PTX_LDU_G_v2i64_ELE_areg64,
1559 NVPTX::INT_PTX_LDU_G_v2f16_ELE_areg64,
1560 NVPTX::INT_PTX_LDU_G_v2f16x2_ELE_areg64,
1561 NVPTX::INT_PTX_LDU_G_v2f32_ELE_areg64,
1562 NVPTX::INT_PTX_LDU_G_v2f64_ELE_areg64);
1564 case NVPTXISD::LoadV4:
1565 case NVPTXISD::LDGV4:
1566 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1567 NVPTX::INT_PTX_LDG_G_v4i8_ELE_areg64,
1568 NVPTX::INT_PTX_LDG_G_v4i16_ELE_areg64,
1569 NVPTX::INT_PTX_LDG_G_v4i32_ELE_areg64, None,
1570 NVPTX::INT_PTX_LDG_G_v4f16_ELE_areg64,
1571 NVPTX::INT_PTX_LDG_G_v4f16x2_ELE_areg64,
1572 NVPTX::INT_PTX_LDG_G_v4f32_ELE_areg64, None);
1574 case NVPTXISD::LDUV4:
1575 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1576 NVPTX::INT_PTX_LDU_G_v4i8_ELE_areg64,
1577 NVPTX::INT_PTX_LDU_G_v4i16_ELE_areg64,
1578 NVPTX::INT_PTX_LDU_G_v4i32_ELE_areg64, None,
1579 NVPTX::INT_PTX_LDU_G_v4f16_ELE_areg64,
1580 NVPTX::INT_PTX_LDU_G_v4f16x2_ELE_areg64,
1581 NVPTX::INT_PTX_LDU_G_v4f32_ELE_areg64, None);
1585 switch (N->getOpcode()) {
1589 case ISD::INTRINSIC_W_CHAIN:
1591 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1592 NVPTX::INT_PTX_LDG_GLOBAL_i8areg,
1593 NVPTX::INT_PTX_LDG_GLOBAL_i16areg,
1594 NVPTX::INT_PTX_LDG_GLOBAL_i32areg,
1595 NVPTX::INT_PTX_LDG_GLOBAL_i64areg,
1596 NVPTX::INT_PTX_LDG_GLOBAL_f16areg,
1597 NVPTX::INT_PTX_LDG_GLOBAL_f16x2areg,
1598 NVPTX::INT_PTX_LDG_GLOBAL_f32areg,
1599 NVPTX::INT_PTX_LDG_GLOBAL_f64areg);
1601 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1602 NVPTX::INT_PTX_LDU_GLOBAL_i8areg,
1603 NVPTX::INT_PTX_LDU_GLOBAL_i16areg,
1604 NVPTX::INT_PTX_LDU_GLOBAL_i32areg,
1605 NVPTX::INT_PTX_LDU_GLOBAL_i64areg,
1606 NVPTX::INT_PTX_LDU_GLOBAL_f16areg,
1607 NVPTX::INT_PTX_LDU_GLOBAL_f16x2areg,
1608 NVPTX::INT_PTX_LDU_GLOBAL_f32areg,
1609 NVPTX::INT_PTX_LDU_GLOBAL_f64areg);
1611 case NVPTXISD::LoadV2:
1612 case NVPTXISD::LDGV2:
1613 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1614 NVPTX::INT_PTX_LDG_G_v2i8_ELE_areg32,
1615 NVPTX::INT_PTX_LDG_G_v2i16_ELE_areg32,
1616 NVPTX::INT_PTX_LDG_G_v2i32_ELE_areg32,
1617 NVPTX::INT_PTX_LDG_G_v2i64_ELE_areg32,
1618 NVPTX::INT_PTX_LDG_G_v2f16_ELE_areg32,
1619 NVPTX::INT_PTX_LDG_G_v2f16x2_ELE_areg32,
1620 NVPTX::INT_PTX_LDG_G_v2f32_ELE_areg32,
1621 NVPTX::INT_PTX_LDG_G_v2f64_ELE_areg32);
1623 case NVPTXISD::LDUV2:
1624 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1625 NVPTX::INT_PTX_LDU_G_v2i8_ELE_areg32,
1626 NVPTX::INT_PTX_LDU_G_v2i16_ELE_areg32,
1627 NVPTX::INT_PTX_LDU_G_v2i32_ELE_areg32,
1628 NVPTX::INT_PTX_LDU_G_v2i64_ELE_areg32,
1629 NVPTX::INT_PTX_LDU_G_v2f16_ELE_areg32,
1630 NVPTX::INT_PTX_LDU_G_v2f16x2_ELE_areg32,
1631 NVPTX::INT_PTX_LDU_G_v2f32_ELE_areg32,
1632 NVPTX::INT_PTX_LDU_G_v2f64_ELE_areg32);
1634 case NVPTXISD::LoadV4:
1635 case NVPTXISD::LDGV4:
1636 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1637 NVPTX::INT_PTX_LDG_G_v4i8_ELE_areg32,
1638 NVPTX::INT_PTX_LDG_G_v4i16_ELE_areg32,
1639 NVPTX::INT_PTX_LDG_G_v4i32_ELE_areg32, None,
1640 NVPTX::INT_PTX_LDG_G_v4f16_ELE_areg32,
1641 NVPTX::INT_PTX_LDG_G_v4f16x2_ELE_areg32,
1642 NVPTX::INT_PTX_LDG_G_v4f32_ELE_areg32, None);
1644 case NVPTXISD::LDUV4:
1645 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1646 NVPTX::INT_PTX_LDU_G_v4i8_ELE_areg32,
1647 NVPTX::INT_PTX_LDU_G_v4i16_ELE_areg32,
1648 NVPTX::INT_PTX_LDU_G_v4i32_ELE_areg32, None,
1649 NVPTX::INT_PTX_LDU_G_v4f16_ELE_areg32,
1650 NVPTX::INT_PTX_LDU_G_v4f16x2_ELE_areg32,
1651 NVPTX::INT_PTX_LDU_G_v4f32_ELE_areg32, None);
1657 SDValue Ops[] = { Op1, Chain };
1658 LD = CurDAG->getMachineNode(Opcode.getValue(), DL, InstVTList, Ops);
1661 MachineMemOperand *MemRef = Mem->getMemOperand();
1662 CurDAG->setNodeMemRefs(cast<MachineSDNode>(LD), {MemRef});
1664 // For automatic generation of LDG (through SelectLoad[Vector], not the
1665 // intrinsics), we may have an extending load like:
1667 // i32,ch = load<LD1[%data1(addrspace=1)], zext from i8> t0, t7, undef:i64
1669 // In this case, the matching logic above will select a load for the original
1670 // memory type (in this case, i8) and our types will not match (the node needs
1671 // to return an i32 in this case). Our LDG/LDU nodes do not support the
1672 // concept of sign-/zero-extension, so emulate it here by adding an explicit
1673 // CVT instruction. Ptxas should clean up any redundancies here.
1675 EVT OrigType = N->getValueType(0);
1676 LoadSDNode *LdNode = dyn_cast<LoadSDNode>(N);
1678 if (OrigType != EltVT && LdNode) {
1679 // We have an extending-load. The instruction we selected operates on the
1680 // smaller type, but the SDNode we are replacing has the larger type. We
1681 // need to emit a CVT to make the types match.
1682 bool IsSigned = LdNode->getExtensionType() == ISD::SEXTLOAD;
1683 unsigned CvtOpc = GetConvertOpcode(OrigType.getSimpleVT(),
1684 EltVT.getSimpleVT(), IsSigned);
1686 // For each output value, apply the manual sign/zero-extension and make sure
1687 // all users of the load go through that CVT.
1688 for (unsigned i = 0; i != NumElts; ++i) {
1690 SDValue OrigVal(N, i);
1693 CurDAG->getMachineNode(CvtOpc, DL, OrigType, Res,
1694 CurDAG->getTargetConstant(NVPTX::PTXCvtMode::NONE,
1696 ReplaceUses(OrigVal, SDValue(CvtNode, 0));
1704 bool NVPTXDAGToDAGISel::tryStore(SDNode *N) {
1706 MemSDNode *ST = cast<MemSDNode>(N);
1707 assert(ST->writeMem() && "Expected store");
1708 StoreSDNode *PlainStore = dyn_cast<StoreSDNode>(N);
1709 AtomicSDNode *AtomicStore = dyn_cast<AtomicSDNode>(N);
1710 assert((PlainStore || AtomicStore) && "Expected store");
1711 EVT StoreVT = ST->getMemoryVT();
1712 SDNode *NVPTXST = nullptr;
1714 // do not support pre/post inc/dec
1715 if (PlainStore && PlainStore->isIndexed())
1718 if (!StoreVT.isSimple())
1721 AtomicOrdering Ordering = ST->getOrdering();
1722 // In order to lower atomic loads with stronger guarantees we would need to
1723 // use store.release or insert fences. However these features were only added
1724 // with PTX ISA 6.0 / sm_70.
1725 // TODO: Check if we can actually use the new instructions and implement them.
1726 if (isStrongerThanMonotonic(Ordering))
1729 // Address Space Setting
1730 unsigned int CodeAddrSpace = getCodeAddrSpace(ST);
1731 unsigned int PointerSize =
1732 CurDAG->getDataLayout().getPointerSizeInBits(ST->getAddressSpace());
1735 // - .volatile is only available for .global and .shared
1736 // - .volatile has the same memory synchronization semantics as .relaxed.sys
1737 bool isVolatile = ST->isVolatile() || Ordering == AtomicOrdering::Monotonic;
1738 if (CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL &&
1739 CodeAddrSpace != NVPTX::PTXLdStInstCode::SHARED &&
1740 CodeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC)
1744 MVT SimpleVT = StoreVT.getSimpleVT();
1745 unsigned vecType = NVPTX::PTXLdStInstCode::Scalar;
1747 // Type Setting: toType + toTypeWidth
1748 // - for integer type, always use 'u'
1750 MVT ScalarVT = SimpleVT.getScalarType();
1751 unsigned toTypeWidth = ScalarVT.getSizeInBits();
1752 if (SimpleVT.isVector()) {
1753 assert(StoreVT == MVT::v2f16 && "Unexpected vector type");
1754 // v2f16 is stored using st.b32
1758 unsigned int toType;
1759 if (ScalarVT.isFloatingPoint())
1760 // f16 uses .b16 as its storage type.
1761 toType = ScalarVT.SimpleTy == MVT::f16 ? NVPTX::PTXLdStInstCode::Untyped
1762 : NVPTX::PTXLdStInstCode::Float;
1764 toType = NVPTX::PTXLdStInstCode::Unsigned;
1766 // Create the machine instruction DAG
1767 SDValue Chain = ST->getChain();
1768 SDValue Value = PlainStore ? PlainStore->getValue() : AtomicStore->getVal();
1769 SDValue BasePtr = ST->getBasePtr();
1771 SDValue Offset, Base;
1772 Optional<unsigned> Opcode;
1773 MVT::SimpleValueType SourceVT =
1774 Value.getNode()->getSimpleValueType(0).SimpleTy;
1776 if (SelectDirectAddr(BasePtr, Addr)) {
1777 Opcode = pickOpcodeForVT(SourceVT, NVPTX::ST_i8_avar, NVPTX::ST_i16_avar,
1778 NVPTX::ST_i32_avar, NVPTX::ST_i64_avar,
1779 NVPTX::ST_f16_avar, NVPTX::ST_f16x2_avar,
1780 NVPTX::ST_f32_avar, NVPTX::ST_f64_avar);
1783 SDValue Ops[] = {Value,
1784 getI32Imm(isVolatile, dl),
1785 getI32Imm(CodeAddrSpace, dl),
1786 getI32Imm(vecType, dl),
1787 getI32Imm(toType, dl),
1788 getI32Imm(toTypeWidth, dl),
1791 NVPTXST = CurDAG->getMachineNode(Opcode.getValue(), dl, MVT::Other, Ops);
1792 } else if (PointerSize == 64
1793 ? SelectADDRsi64(BasePtr.getNode(), BasePtr, Base, Offset)
1794 : SelectADDRsi(BasePtr.getNode(), BasePtr, Base, Offset)) {
1795 Opcode = pickOpcodeForVT(SourceVT, NVPTX::ST_i8_asi, NVPTX::ST_i16_asi,
1796 NVPTX::ST_i32_asi, NVPTX::ST_i64_asi,
1797 NVPTX::ST_f16_asi, NVPTX::ST_f16x2_asi,
1798 NVPTX::ST_f32_asi, NVPTX::ST_f64_asi);
1801 SDValue Ops[] = {Value,
1802 getI32Imm(isVolatile, dl),
1803 getI32Imm(CodeAddrSpace, dl),
1804 getI32Imm(vecType, dl),
1805 getI32Imm(toType, dl),
1806 getI32Imm(toTypeWidth, dl),
1810 NVPTXST = CurDAG->getMachineNode(Opcode.getValue(), dl, MVT::Other, Ops);
1811 } else if (PointerSize == 64
1812 ? SelectADDRri64(BasePtr.getNode(), BasePtr, Base, Offset)
1813 : SelectADDRri(BasePtr.getNode(), BasePtr, Base, Offset)) {
1814 if (PointerSize == 64)
1815 Opcode = pickOpcodeForVT(
1816 SourceVT, NVPTX::ST_i8_ari_64, NVPTX::ST_i16_ari_64,
1817 NVPTX::ST_i32_ari_64, NVPTX::ST_i64_ari_64, NVPTX::ST_f16_ari_64,
1818 NVPTX::ST_f16x2_ari_64, NVPTX::ST_f32_ari_64, NVPTX::ST_f64_ari_64);
1820 Opcode = pickOpcodeForVT(SourceVT, NVPTX::ST_i8_ari, NVPTX::ST_i16_ari,
1821 NVPTX::ST_i32_ari, NVPTX::ST_i64_ari,
1822 NVPTX::ST_f16_ari, NVPTX::ST_f16x2_ari,
1823 NVPTX::ST_f32_ari, NVPTX::ST_f64_ari);
1827 SDValue Ops[] = {Value,
1828 getI32Imm(isVolatile, dl),
1829 getI32Imm(CodeAddrSpace, dl),
1830 getI32Imm(vecType, dl),
1831 getI32Imm(toType, dl),
1832 getI32Imm(toTypeWidth, dl),
1836 NVPTXST = CurDAG->getMachineNode(Opcode.getValue(), dl, MVT::Other, Ops);
1838 if (PointerSize == 64)
1840 pickOpcodeForVT(SourceVT, NVPTX::ST_i8_areg_64, NVPTX::ST_i16_areg_64,
1841 NVPTX::ST_i32_areg_64, NVPTX::ST_i64_areg_64,
1842 NVPTX::ST_f16_areg_64, NVPTX::ST_f16x2_areg_64,
1843 NVPTX::ST_f32_areg_64, NVPTX::ST_f64_areg_64);
1845 Opcode = pickOpcodeForVT(SourceVT, NVPTX::ST_i8_areg, NVPTX::ST_i16_areg,
1846 NVPTX::ST_i32_areg, NVPTX::ST_i64_areg,
1847 NVPTX::ST_f16_areg, NVPTX::ST_f16x2_areg,
1848 NVPTX::ST_f32_areg, NVPTX::ST_f64_areg);
1851 SDValue Ops[] = {Value,
1852 getI32Imm(isVolatile, dl),
1853 getI32Imm(CodeAddrSpace, dl),
1854 getI32Imm(vecType, dl),
1855 getI32Imm(toType, dl),
1856 getI32Imm(toTypeWidth, dl),
1859 NVPTXST = CurDAG->getMachineNode(Opcode.getValue(), dl, MVT::Other, Ops);
1865 MachineMemOperand *MemRef = cast<MemSDNode>(N)->getMemOperand();
1866 CurDAG->setNodeMemRefs(cast<MachineSDNode>(NVPTXST), {MemRef});
1867 ReplaceNode(N, NVPTXST);
1871 bool NVPTXDAGToDAGISel::tryStoreVector(SDNode *N) {
1872 SDValue Chain = N->getOperand(0);
1873 SDValue Op1 = N->getOperand(1);
1874 SDValue Addr, Offset, Base;
1875 Optional<unsigned> Opcode;
1878 EVT EltVT = Op1.getValueType();
1879 MemSDNode *MemSD = cast<MemSDNode>(N);
1880 EVT StoreVT = MemSD->getMemoryVT();
1882 // Address Space Setting
1883 unsigned CodeAddrSpace = getCodeAddrSpace(MemSD);
1884 if (CodeAddrSpace == NVPTX::PTXLdStInstCode::CONSTANT) {
1885 report_fatal_error("Cannot store to pointer that points to constant "
1888 unsigned int PointerSize =
1889 CurDAG->getDataLayout().getPointerSizeInBits(MemSD->getAddressSpace());
1892 // - .volatile is only availalble for .global and .shared
1893 bool IsVolatile = MemSD->isVolatile();
1894 if (CodeAddrSpace != NVPTX::PTXLdStInstCode::GLOBAL &&
1895 CodeAddrSpace != NVPTX::PTXLdStInstCode::SHARED &&
1896 CodeAddrSpace != NVPTX::PTXLdStInstCode::GENERIC)
1899 // Type Setting: toType + toTypeWidth
1900 // - for integer type, always use 'u'
1901 assert(StoreVT.isSimple() && "Store value is not simple");
1902 MVT ScalarVT = StoreVT.getSimpleVT().getScalarType();
1903 unsigned ToTypeWidth = ScalarVT.getSizeInBits();
1905 if (ScalarVT.isFloatingPoint())
1906 ToType = ScalarVT.SimpleTy == MVT::f16 ? NVPTX::PTXLdStInstCode::Untyped
1907 : NVPTX::PTXLdStInstCode::Float;
1909 ToType = NVPTX::PTXLdStInstCode::Unsigned;
1911 SmallVector<SDValue, 12> StOps;
1915 switch (N->getOpcode()) {
1916 case NVPTXISD::StoreV2:
1917 VecType = NVPTX::PTXLdStInstCode::V2;
1918 StOps.push_back(N->getOperand(1));
1919 StOps.push_back(N->getOperand(2));
1920 N2 = N->getOperand(3);
1922 case NVPTXISD::StoreV4:
1923 VecType = NVPTX::PTXLdStInstCode::V4;
1924 StOps.push_back(N->getOperand(1));
1925 StOps.push_back(N->getOperand(2));
1926 StOps.push_back(N->getOperand(3));
1927 StOps.push_back(N->getOperand(4));
1928 N2 = N->getOperand(5);
1934 // v8f16 is a special case. PTX doesn't have st.v8.f16
1935 // instruction. Instead, we split the vector into v2f16 chunks and
1936 // store them with st.v4.b32.
1937 if (EltVT == MVT::v2f16) {
1938 assert(N->getOpcode() == NVPTXISD::StoreV4 && "Unexpected load opcode.");
1940 ToType = NVPTX::PTXLdStInstCode::Untyped;
1944 StOps.push_back(getI32Imm(IsVolatile, DL));
1945 StOps.push_back(getI32Imm(CodeAddrSpace, DL));
1946 StOps.push_back(getI32Imm(VecType, DL));
1947 StOps.push_back(getI32Imm(ToType, DL));
1948 StOps.push_back(getI32Imm(ToTypeWidth, DL));
1950 if (SelectDirectAddr(N2, Addr)) {
1951 switch (N->getOpcode()) {
1954 case NVPTXISD::StoreV2:
1955 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1956 NVPTX::STV_i8_v2_avar, NVPTX::STV_i16_v2_avar,
1957 NVPTX::STV_i32_v2_avar, NVPTX::STV_i64_v2_avar,
1958 NVPTX::STV_f16_v2_avar, NVPTX::STV_f16x2_v2_avar,
1959 NVPTX::STV_f32_v2_avar, NVPTX::STV_f64_v2_avar);
1961 case NVPTXISD::StoreV4:
1963 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v4_avar,
1964 NVPTX::STV_i16_v4_avar, NVPTX::STV_i32_v4_avar, None,
1965 NVPTX::STV_f16_v4_avar, NVPTX::STV_f16x2_v4_avar,
1966 NVPTX::STV_f32_v4_avar, None);
1969 StOps.push_back(Addr);
1970 } else if (PointerSize == 64 ? SelectADDRsi64(N2.getNode(), N2, Base, Offset)
1971 : SelectADDRsi(N2.getNode(), N2, Base, Offset)) {
1972 switch (N->getOpcode()) {
1975 case NVPTXISD::StoreV2:
1976 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
1977 NVPTX::STV_i8_v2_asi, NVPTX::STV_i16_v2_asi,
1978 NVPTX::STV_i32_v2_asi, NVPTX::STV_i64_v2_asi,
1979 NVPTX::STV_f16_v2_asi, NVPTX::STV_f16x2_v2_asi,
1980 NVPTX::STV_f32_v2_asi, NVPTX::STV_f64_v2_asi);
1982 case NVPTXISD::StoreV4:
1984 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v4_asi,
1985 NVPTX::STV_i16_v4_asi, NVPTX::STV_i32_v4_asi, None,
1986 NVPTX::STV_f16_v4_asi, NVPTX::STV_f16x2_v4_asi,
1987 NVPTX::STV_f32_v4_asi, None);
1990 StOps.push_back(Base);
1991 StOps.push_back(Offset);
1992 } else if (PointerSize == 64 ? SelectADDRri64(N2.getNode(), N2, Base, Offset)
1993 : SelectADDRri(N2.getNode(), N2, Base, Offset)) {
1994 if (PointerSize == 64) {
1995 switch (N->getOpcode()) {
1998 case NVPTXISD::StoreV2:
1999 Opcode = pickOpcodeForVT(
2000 EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v2_ari_64,
2001 NVPTX::STV_i16_v2_ari_64, NVPTX::STV_i32_v2_ari_64,
2002 NVPTX::STV_i64_v2_ari_64, NVPTX::STV_f16_v2_ari_64,
2003 NVPTX::STV_f16x2_v2_ari_64, NVPTX::STV_f32_v2_ari_64,
2004 NVPTX::STV_f64_v2_ari_64);
2006 case NVPTXISD::StoreV4:
2007 Opcode = pickOpcodeForVT(
2008 EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v4_ari_64,
2009 NVPTX::STV_i16_v4_ari_64, NVPTX::STV_i32_v4_ari_64, None,
2010 NVPTX::STV_f16_v4_ari_64, NVPTX::STV_f16x2_v4_ari_64,
2011 NVPTX::STV_f32_v4_ari_64, None);
2015 switch (N->getOpcode()) {
2018 case NVPTXISD::StoreV2:
2019 Opcode = pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy,
2020 NVPTX::STV_i8_v2_ari, NVPTX::STV_i16_v2_ari,
2021 NVPTX::STV_i32_v2_ari, NVPTX::STV_i64_v2_ari,
2022 NVPTX::STV_f16_v2_ari, NVPTX::STV_f16x2_v2_ari,
2023 NVPTX::STV_f32_v2_ari, NVPTX::STV_f64_v2_ari);
2025 case NVPTXISD::StoreV4:
2027 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v4_ari,
2028 NVPTX::STV_i16_v4_ari, NVPTX::STV_i32_v4_ari, None,
2029 NVPTX::STV_f16_v4_ari, NVPTX::STV_f16x2_v4_ari,
2030 NVPTX::STV_f32_v4_ari, None);
2034 StOps.push_back(Base);
2035 StOps.push_back(Offset);
2037 if (PointerSize == 64) {
2038 switch (N->getOpcode()) {
2041 case NVPTXISD::StoreV2:
2042 Opcode = pickOpcodeForVT(
2043 EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v2_areg_64,
2044 NVPTX::STV_i16_v2_areg_64, NVPTX::STV_i32_v2_areg_64,
2045 NVPTX::STV_i64_v2_areg_64, NVPTX::STV_f16_v2_areg_64,
2046 NVPTX::STV_f16x2_v2_areg_64, NVPTX::STV_f32_v2_areg_64,
2047 NVPTX::STV_f64_v2_areg_64);
2049 case NVPTXISD::StoreV4:
2050 Opcode = pickOpcodeForVT(
2051 EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v4_areg_64,
2052 NVPTX::STV_i16_v4_areg_64, NVPTX::STV_i32_v4_areg_64, None,
2053 NVPTX::STV_f16_v4_areg_64, NVPTX::STV_f16x2_v4_areg_64,
2054 NVPTX::STV_f32_v4_areg_64, None);
2058 switch (N->getOpcode()) {
2061 case NVPTXISD::StoreV2:
2063 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v2_areg,
2064 NVPTX::STV_i16_v2_areg, NVPTX::STV_i32_v2_areg,
2065 NVPTX::STV_i64_v2_areg, NVPTX::STV_f16_v2_areg,
2066 NVPTX::STV_f16x2_v2_areg, NVPTX::STV_f32_v2_areg,
2067 NVPTX::STV_f64_v2_areg);
2069 case NVPTXISD::StoreV4:
2071 pickOpcodeForVT(EltVT.getSimpleVT().SimpleTy, NVPTX::STV_i8_v4_areg,
2072 NVPTX::STV_i16_v4_areg, NVPTX::STV_i32_v4_areg, None,
2073 NVPTX::STV_f16_v4_areg, NVPTX::STV_f16x2_v4_areg,
2074 NVPTX::STV_f32_v4_areg, None);
2078 StOps.push_back(N2);
2084 StOps.push_back(Chain);
2086 ST = CurDAG->getMachineNode(Opcode.getValue(), DL, MVT::Other, StOps);
2088 MachineMemOperand *MemRef = cast<MemSDNode>(N)->getMemOperand();
2089 CurDAG->setNodeMemRefs(cast<MachineSDNode>(ST), {MemRef});
2095 bool NVPTXDAGToDAGISel::tryLoadParam(SDNode *Node) {
2096 SDValue Chain = Node->getOperand(0);
2097 SDValue Offset = Node->getOperand(2);
2098 SDValue Flag = Node->getOperand(3);
2100 MemSDNode *Mem = cast<MemSDNode>(Node);
2103 switch (Node->getOpcode()) {
2106 case NVPTXISD::LoadParam:
2109 case NVPTXISD::LoadParamV2:
2112 case NVPTXISD::LoadParamV4:
2117 EVT EltVT = Node->getValueType(0);
2118 EVT MemVT = Mem->getMemoryVT();
2120 Optional<unsigned> Opcode;
2126 Opcode = pickOpcodeForVT(MemVT.getSimpleVT().SimpleTy,
2127 NVPTX::LoadParamMemI8, NVPTX::LoadParamMemI16,
2128 NVPTX::LoadParamMemI32, NVPTX::LoadParamMemI64,
2129 NVPTX::LoadParamMemF16, NVPTX::LoadParamMemF16x2,
2130 NVPTX::LoadParamMemF32, NVPTX::LoadParamMemF64);
2134 pickOpcodeForVT(MemVT.getSimpleVT().SimpleTy, NVPTX::LoadParamMemV2I8,
2135 NVPTX::LoadParamMemV2I16, NVPTX::LoadParamMemV2I32,
2136 NVPTX::LoadParamMemV2I64, NVPTX::LoadParamMemV2F16,
2137 NVPTX::LoadParamMemV2F16x2, NVPTX::LoadParamMemV2F32,
2138 NVPTX::LoadParamMemV2F64);
2141 Opcode = pickOpcodeForVT(
2142 MemVT.getSimpleVT().SimpleTy, NVPTX::LoadParamMemV4I8,
2143 NVPTX::LoadParamMemV4I16, NVPTX::LoadParamMemV4I32, None,
2144 NVPTX::LoadParamMemV4F16, NVPTX::LoadParamMemV4F16x2,
2145 NVPTX::LoadParamMemV4F32, None);
2153 VTs = CurDAG->getVTList(EltVT, MVT::Other, MVT::Glue);
2154 } else if (VecSize == 2) {
2155 VTs = CurDAG->getVTList(EltVT, EltVT, MVT::Other, MVT::Glue);
2157 EVT EVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other, MVT::Glue };
2158 VTs = CurDAG->getVTList(EVTs);
2161 unsigned OffsetVal = cast<ConstantSDNode>(Offset)->getZExtValue();
2163 SmallVector<SDValue, 2> Ops;
2164 Ops.push_back(CurDAG->getTargetConstant(OffsetVal, DL, MVT::i32));
2165 Ops.push_back(Chain);
2166 Ops.push_back(Flag);
2168 ReplaceNode(Node, CurDAG->getMachineNode(Opcode.getValue(), DL, VTs, Ops));
2172 bool NVPTXDAGToDAGISel::tryStoreRetval(SDNode *N) {
2174 SDValue Chain = N->getOperand(0);
2175 SDValue Offset = N->getOperand(1);
2176 unsigned OffsetVal = cast<ConstantSDNode>(Offset)->getZExtValue();
2177 MemSDNode *Mem = cast<MemSDNode>(N);
2179 // How many elements do we have?
2180 unsigned NumElts = 1;
2181 switch (N->getOpcode()) {
2184 case NVPTXISD::StoreRetval:
2187 case NVPTXISD::StoreRetvalV2:
2190 case NVPTXISD::StoreRetvalV4:
2195 // Build vector of operands
2196 SmallVector<SDValue, 6> Ops;
2197 for (unsigned i = 0; i < NumElts; ++i)
2198 Ops.push_back(N->getOperand(i + 2));
2199 Ops.push_back(CurDAG->getTargetConstant(OffsetVal, DL, MVT::i32));
2200 Ops.push_back(Chain);
2202 // Determine target opcode
2203 // If we have an i1, use an 8-bit store. The lowering code in
2204 // NVPTXISelLowering will have already emitted an upcast.
2205 Optional<unsigned> Opcode = 0;
2210 Opcode = pickOpcodeForVT(Mem->getMemoryVT().getSimpleVT().SimpleTy,
2211 NVPTX::StoreRetvalI8, NVPTX::StoreRetvalI16,
2212 NVPTX::StoreRetvalI32, NVPTX::StoreRetvalI64,
2213 NVPTX::StoreRetvalF16, NVPTX::StoreRetvalF16x2,
2214 NVPTX::StoreRetvalF32, NVPTX::StoreRetvalF64);
2217 Opcode = pickOpcodeForVT(Mem->getMemoryVT().getSimpleVT().SimpleTy,
2218 NVPTX::StoreRetvalV2I8, NVPTX::StoreRetvalV2I16,
2219 NVPTX::StoreRetvalV2I32, NVPTX::StoreRetvalV2I64,
2220 NVPTX::StoreRetvalV2F16, NVPTX::StoreRetvalV2F16x2,
2221 NVPTX::StoreRetvalV2F32, NVPTX::StoreRetvalV2F64);
2224 Opcode = pickOpcodeForVT(Mem->getMemoryVT().getSimpleVT().SimpleTy,
2225 NVPTX::StoreRetvalV4I8, NVPTX::StoreRetvalV4I16,
2226 NVPTX::StoreRetvalV4I32, None,
2227 NVPTX::StoreRetvalV4F16, NVPTX::StoreRetvalV4F16x2,
2228 NVPTX::StoreRetvalV4F32, None);
2234 SDNode *Ret = CurDAG->getMachineNode(Opcode.getValue(), DL, MVT::Other, Ops);
2235 MachineMemOperand *MemRef = cast<MemSDNode>(N)->getMemOperand();
2236 CurDAG->setNodeMemRefs(cast<MachineSDNode>(Ret), {MemRef});
2238 ReplaceNode(N, Ret);
2242 bool NVPTXDAGToDAGISel::tryStoreParam(SDNode *N) {
2244 SDValue Chain = N->getOperand(0);
2245 SDValue Param = N->getOperand(1);
2246 unsigned ParamVal = cast<ConstantSDNode>(Param)->getZExtValue();
2247 SDValue Offset = N->getOperand(2);
2248 unsigned OffsetVal = cast<ConstantSDNode>(Offset)->getZExtValue();
2249 MemSDNode *Mem = cast<MemSDNode>(N);
2250 SDValue Flag = N->getOperand(N->getNumOperands() - 1);
2252 // How many elements do we have?
2253 unsigned NumElts = 1;
2254 switch (N->getOpcode()) {
2257 case NVPTXISD::StoreParamU32:
2258 case NVPTXISD::StoreParamS32:
2259 case NVPTXISD::StoreParam:
2262 case NVPTXISD::StoreParamV2:
2265 case NVPTXISD::StoreParamV4:
2270 // Build vector of operands
2271 SmallVector<SDValue, 8> Ops;
2272 for (unsigned i = 0; i < NumElts; ++i)
2273 Ops.push_back(N->getOperand(i + 3));
2274 Ops.push_back(CurDAG->getTargetConstant(ParamVal, DL, MVT::i32));
2275 Ops.push_back(CurDAG->getTargetConstant(OffsetVal, DL, MVT::i32));
2276 Ops.push_back(Chain);
2277 Ops.push_back(Flag);
2279 // Determine target opcode
2280 // If we have an i1, use an 8-bit store. The lowering code in
2281 // NVPTXISelLowering will have already emitted an upcast.
2282 Optional<unsigned> Opcode = 0;
2283 switch (N->getOpcode()) {
2289 Opcode = pickOpcodeForVT(Mem->getMemoryVT().getSimpleVT().SimpleTy,
2290 NVPTX::StoreParamI8, NVPTX::StoreParamI16,
2291 NVPTX::StoreParamI32, NVPTX::StoreParamI64,
2292 NVPTX::StoreParamF16, NVPTX::StoreParamF16x2,
2293 NVPTX::StoreParamF32, NVPTX::StoreParamF64);
2296 Opcode = pickOpcodeForVT(Mem->getMemoryVT().getSimpleVT().SimpleTy,
2297 NVPTX::StoreParamV2I8, NVPTX::StoreParamV2I16,
2298 NVPTX::StoreParamV2I32, NVPTX::StoreParamV2I64,
2299 NVPTX::StoreParamV2F16, NVPTX::StoreParamV2F16x2,
2300 NVPTX::StoreParamV2F32, NVPTX::StoreParamV2F64);
2303 Opcode = pickOpcodeForVT(Mem->getMemoryVT().getSimpleVT().SimpleTy,
2304 NVPTX::StoreParamV4I8, NVPTX::StoreParamV4I16,
2305 NVPTX::StoreParamV4I32, None,
2306 NVPTX::StoreParamV4F16, NVPTX::StoreParamV4F16x2,
2307 NVPTX::StoreParamV4F32, None);
2313 // Special case: if we have a sign-extend/zero-extend node, insert the
2314 // conversion instruction first, and use that as the value operand to
2315 // the selected StoreParam node.
2316 case NVPTXISD::StoreParamU32: {
2317 Opcode = NVPTX::StoreParamI32;
2318 SDValue CvtNone = CurDAG->getTargetConstant(NVPTX::PTXCvtMode::NONE, DL,
2320 SDNode *Cvt = CurDAG->getMachineNode(NVPTX::CVT_u32_u16, DL,
2321 MVT::i32, Ops[0], CvtNone);
2322 Ops[0] = SDValue(Cvt, 0);
2325 case NVPTXISD::StoreParamS32: {
2326 Opcode = NVPTX::StoreParamI32;
2327 SDValue CvtNone = CurDAG->getTargetConstant(NVPTX::PTXCvtMode::NONE, DL,
2329 SDNode *Cvt = CurDAG->getMachineNode(NVPTX::CVT_s32_s16, DL,
2330 MVT::i32, Ops[0], CvtNone);
2331 Ops[0] = SDValue(Cvt, 0);
2336 SDVTList RetVTs = CurDAG->getVTList(MVT::Other, MVT::Glue);
2338 CurDAG->getMachineNode(Opcode.getValue(), DL, RetVTs, Ops);
2339 MachineMemOperand *MemRef = cast<MemSDNode>(N)->getMemOperand();
2340 CurDAG->setNodeMemRefs(cast<MachineSDNode>(Ret), {MemRef});
2342 ReplaceNode(N, Ret);
2346 bool NVPTXDAGToDAGISel::tryTextureIntrinsic(SDNode *N) {
2349 switch (N->getOpcode()) {
2350 default: return false;
2351 case NVPTXISD::Tex1DFloatS32:
2352 Opc = NVPTX::TEX_1D_F32_S32;
2354 case NVPTXISD::Tex1DFloatFloat:
2355 Opc = NVPTX::TEX_1D_F32_F32;
2357 case NVPTXISD::Tex1DFloatFloatLevel:
2358 Opc = NVPTX::TEX_1D_F32_F32_LEVEL;
2360 case NVPTXISD::Tex1DFloatFloatGrad:
2361 Opc = NVPTX::TEX_1D_F32_F32_GRAD;
2363 case NVPTXISD::Tex1DS32S32:
2364 Opc = NVPTX::TEX_1D_S32_S32;
2366 case NVPTXISD::Tex1DS32Float:
2367 Opc = NVPTX::TEX_1D_S32_F32;
2369 case NVPTXISD::Tex1DS32FloatLevel:
2370 Opc = NVPTX::TEX_1D_S32_F32_LEVEL;
2372 case NVPTXISD::Tex1DS32FloatGrad:
2373 Opc = NVPTX::TEX_1D_S32_F32_GRAD;
2375 case NVPTXISD::Tex1DU32S32:
2376 Opc = NVPTX::TEX_1D_U32_S32;
2378 case NVPTXISD::Tex1DU32Float:
2379 Opc = NVPTX::TEX_1D_U32_F32;
2381 case NVPTXISD::Tex1DU32FloatLevel:
2382 Opc = NVPTX::TEX_1D_U32_F32_LEVEL;
2384 case NVPTXISD::Tex1DU32FloatGrad:
2385 Opc = NVPTX::TEX_1D_U32_F32_GRAD;
2387 case NVPTXISD::Tex1DArrayFloatS32:
2388 Opc = NVPTX::TEX_1D_ARRAY_F32_S32;
2390 case NVPTXISD::Tex1DArrayFloatFloat:
2391 Opc = NVPTX::TEX_1D_ARRAY_F32_F32;
2393 case NVPTXISD::Tex1DArrayFloatFloatLevel:
2394 Opc = NVPTX::TEX_1D_ARRAY_F32_F32_LEVEL;
2396 case NVPTXISD::Tex1DArrayFloatFloatGrad:
2397 Opc = NVPTX::TEX_1D_ARRAY_F32_F32_GRAD;
2399 case NVPTXISD::Tex1DArrayS32S32:
2400 Opc = NVPTX::TEX_1D_ARRAY_S32_S32;
2402 case NVPTXISD::Tex1DArrayS32Float:
2403 Opc = NVPTX::TEX_1D_ARRAY_S32_F32;
2405 case NVPTXISD::Tex1DArrayS32FloatLevel:
2406 Opc = NVPTX::TEX_1D_ARRAY_S32_F32_LEVEL;
2408 case NVPTXISD::Tex1DArrayS32FloatGrad:
2409 Opc = NVPTX::TEX_1D_ARRAY_S32_F32_GRAD;
2411 case NVPTXISD::Tex1DArrayU32S32:
2412 Opc = NVPTX::TEX_1D_ARRAY_U32_S32;
2414 case NVPTXISD::Tex1DArrayU32Float:
2415 Opc = NVPTX::TEX_1D_ARRAY_U32_F32;
2417 case NVPTXISD::Tex1DArrayU32FloatLevel:
2418 Opc = NVPTX::TEX_1D_ARRAY_U32_F32_LEVEL;
2420 case NVPTXISD::Tex1DArrayU32FloatGrad:
2421 Opc = NVPTX::TEX_1D_ARRAY_U32_F32_GRAD;
2423 case NVPTXISD::Tex2DFloatS32:
2424 Opc = NVPTX::TEX_2D_F32_S32;
2426 case NVPTXISD::Tex2DFloatFloat:
2427 Opc = NVPTX::TEX_2D_F32_F32;
2429 case NVPTXISD::Tex2DFloatFloatLevel:
2430 Opc = NVPTX::TEX_2D_F32_F32_LEVEL;
2432 case NVPTXISD::Tex2DFloatFloatGrad:
2433 Opc = NVPTX::TEX_2D_F32_F32_GRAD;
2435 case NVPTXISD::Tex2DS32S32:
2436 Opc = NVPTX::TEX_2D_S32_S32;
2438 case NVPTXISD::Tex2DS32Float:
2439 Opc = NVPTX::TEX_2D_S32_F32;
2441 case NVPTXISD::Tex2DS32FloatLevel:
2442 Opc = NVPTX::TEX_2D_S32_F32_LEVEL;
2444 case NVPTXISD::Tex2DS32FloatGrad:
2445 Opc = NVPTX::TEX_2D_S32_F32_GRAD;
2447 case NVPTXISD::Tex2DU32S32:
2448 Opc = NVPTX::TEX_2D_U32_S32;
2450 case NVPTXISD::Tex2DU32Float:
2451 Opc = NVPTX::TEX_2D_U32_F32;
2453 case NVPTXISD::Tex2DU32FloatLevel:
2454 Opc = NVPTX::TEX_2D_U32_F32_LEVEL;
2456 case NVPTXISD::Tex2DU32FloatGrad:
2457 Opc = NVPTX::TEX_2D_U32_F32_GRAD;
2459 case NVPTXISD::Tex2DArrayFloatS32:
2460 Opc = NVPTX::TEX_2D_ARRAY_F32_S32;
2462 case NVPTXISD::Tex2DArrayFloatFloat:
2463 Opc = NVPTX::TEX_2D_ARRAY_F32_F32;
2465 case NVPTXISD::Tex2DArrayFloatFloatLevel:
2466 Opc = NVPTX::TEX_2D_ARRAY_F32_F32_LEVEL;
2468 case NVPTXISD::Tex2DArrayFloatFloatGrad:
2469 Opc = NVPTX::TEX_2D_ARRAY_F32_F32_GRAD;
2471 case NVPTXISD::Tex2DArrayS32S32:
2472 Opc = NVPTX::TEX_2D_ARRAY_S32_S32;
2474 case NVPTXISD::Tex2DArrayS32Float:
2475 Opc = NVPTX::TEX_2D_ARRAY_S32_F32;
2477 case NVPTXISD::Tex2DArrayS32FloatLevel:
2478 Opc = NVPTX::TEX_2D_ARRAY_S32_F32_LEVEL;
2480 case NVPTXISD::Tex2DArrayS32FloatGrad:
2481 Opc = NVPTX::TEX_2D_ARRAY_S32_F32_GRAD;
2483 case NVPTXISD::Tex2DArrayU32S32:
2484 Opc = NVPTX::TEX_2D_ARRAY_U32_S32;
2486 case NVPTXISD::Tex2DArrayU32Float:
2487 Opc = NVPTX::TEX_2D_ARRAY_U32_F32;
2489 case NVPTXISD::Tex2DArrayU32FloatLevel:
2490 Opc = NVPTX::TEX_2D_ARRAY_U32_F32_LEVEL;
2492 case NVPTXISD::Tex2DArrayU32FloatGrad:
2493 Opc = NVPTX::TEX_2D_ARRAY_U32_F32_GRAD;
2495 case NVPTXISD::Tex3DFloatS32:
2496 Opc = NVPTX::TEX_3D_F32_S32;
2498 case NVPTXISD::Tex3DFloatFloat:
2499 Opc = NVPTX::TEX_3D_F32_F32;
2501 case NVPTXISD::Tex3DFloatFloatLevel:
2502 Opc = NVPTX::TEX_3D_F32_F32_LEVEL;
2504 case NVPTXISD::Tex3DFloatFloatGrad:
2505 Opc = NVPTX::TEX_3D_F32_F32_GRAD;
2507 case NVPTXISD::Tex3DS32S32:
2508 Opc = NVPTX::TEX_3D_S32_S32;
2510 case NVPTXISD::Tex3DS32Float:
2511 Opc = NVPTX::TEX_3D_S32_F32;
2513 case NVPTXISD::Tex3DS32FloatLevel:
2514 Opc = NVPTX::TEX_3D_S32_F32_LEVEL;
2516 case NVPTXISD::Tex3DS32FloatGrad:
2517 Opc = NVPTX::TEX_3D_S32_F32_GRAD;
2519 case NVPTXISD::Tex3DU32S32:
2520 Opc = NVPTX::TEX_3D_U32_S32;
2522 case NVPTXISD::Tex3DU32Float:
2523 Opc = NVPTX::TEX_3D_U32_F32;
2525 case NVPTXISD::Tex3DU32FloatLevel:
2526 Opc = NVPTX::TEX_3D_U32_F32_LEVEL;
2528 case NVPTXISD::Tex3DU32FloatGrad:
2529 Opc = NVPTX::TEX_3D_U32_F32_GRAD;
2531 case NVPTXISD::TexCubeFloatFloat:
2532 Opc = NVPTX::TEX_CUBE_F32_F32;
2534 case NVPTXISD::TexCubeFloatFloatLevel:
2535 Opc = NVPTX::TEX_CUBE_F32_F32_LEVEL;
2537 case NVPTXISD::TexCubeS32Float:
2538 Opc = NVPTX::TEX_CUBE_S32_F32;
2540 case NVPTXISD::TexCubeS32FloatLevel:
2541 Opc = NVPTX::TEX_CUBE_S32_F32_LEVEL;
2543 case NVPTXISD::TexCubeU32Float:
2544 Opc = NVPTX::TEX_CUBE_U32_F32;
2546 case NVPTXISD::TexCubeU32FloatLevel:
2547 Opc = NVPTX::TEX_CUBE_U32_F32_LEVEL;
2549 case NVPTXISD::TexCubeArrayFloatFloat:
2550 Opc = NVPTX::TEX_CUBE_ARRAY_F32_F32;
2552 case NVPTXISD::TexCubeArrayFloatFloatLevel:
2553 Opc = NVPTX::TEX_CUBE_ARRAY_F32_F32_LEVEL;
2555 case NVPTXISD::TexCubeArrayS32Float:
2556 Opc = NVPTX::TEX_CUBE_ARRAY_S32_F32;
2558 case NVPTXISD::TexCubeArrayS32FloatLevel:
2559 Opc = NVPTX::TEX_CUBE_ARRAY_S32_F32_LEVEL;
2561 case NVPTXISD::TexCubeArrayU32Float:
2562 Opc = NVPTX::TEX_CUBE_ARRAY_U32_F32;
2564 case NVPTXISD::TexCubeArrayU32FloatLevel:
2565 Opc = NVPTX::TEX_CUBE_ARRAY_U32_F32_LEVEL;
2567 case NVPTXISD::Tld4R2DFloatFloat:
2568 Opc = NVPTX::TLD4_R_2D_F32_F32;
2570 case NVPTXISD::Tld4G2DFloatFloat:
2571 Opc = NVPTX::TLD4_G_2D_F32_F32;
2573 case NVPTXISD::Tld4B2DFloatFloat:
2574 Opc = NVPTX::TLD4_B_2D_F32_F32;
2576 case NVPTXISD::Tld4A2DFloatFloat:
2577 Opc = NVPTX::TLD4_A_2D_F32_F32;
2579 case NVPTXISD::Tld4R2DS64Float:
2580 Opc = NVPTX::TLD4_R_2D_S32_F32;
2582 case NVPTXISD::Tld4G2DS64Float:
2583 Opc = NVPTX::TLD4_G_2D_S32_F32;
2585 case NVPTXISD::Tld4B2DS64Float:
2586 Opc = NVPTX::TLD4_B_2D_S32_F32;
2588 case NVPTXISD::Tld4A2DS64Float:
2589 Opc = NVPTX::TLD4_A_2D_S32_F32;
2591 case NVPTXISD::Tld4R2DU64Float:
2592 Opc = NVPTX::TLD4_R_2D_U32_F32;
2594 case NVPTXISD::Tld4G2DU64Float:
2595 Opc = NVPTX::TLD4_G_2D_U32_F32;
2597 case NVPTXISD::Tld4B2DU64Float:
2598 Opc = NVPTX::TLD4_B_2D_U32_F32;
2600 case NVPTXISD::Tld4A2DU64Float:
2601 Opc = NVPTX::TLD4_A_2D_U32_F32;
2603 case NVPTXISD::TexUnified1DFloatS32:
2604 Opc = NVPTX::TEX_UNIFIED_1D_F32_S32;
2606 case NVPTXISD::TexUnified1DFloatFloat:
2607 Opc = NVPTX::TEX_UNIFIED_1D_F32_F32;
2609 case NVPTXISD::TexUnified1DFloatFloatLevel:
2610 Opc = NVPTX::TEX_UNIFIED_1D_F32_F32_LEVEL;
2612 case NVPTXISD::TexUnified1DFloatFloatGrad:
2613 Opc = NVPTX::TEX_UNIFIED_1D_F32_F32_GRAD;
2615 case NVPTXISD::TexUnified1DS32S32:
2616 Opc = NVPTX::TEX_UNIFIED_1D_S32_S32;
2618 case NVPTXISD::TexUnified1DS32Float:
2619 Opc = NVPTX::TEX_UNIFIED_1D_S32_F32;
2621 case NVPTXISD::TexUnified1DS32FloatLevel:
2622 Opc = NVPTX::TEX_UNIFIED_1D_S32_F32_LEVEL;
2624 case NVPTXISD::TexUnified1DS32FloatGrad:
2625 Opc = NVPTX::TEX_UNIFIED_1D_S32_F32_GRAD;
2627 case NVPTXISD::TexUnified1DU32S32:
2628 Opc = NVPTX::TEX_UNIFIED_1D_U32_S32;
2630 case NVPTXISD::TexUnified1DU32Float:
2631 Opc = NVPTX::TEX_UNIFIED_1D_U32_F32;
2633 case NVPTXISD::TexUnified1DU32FloatLevel:
2634 Opc = NVPTX::TEX_UNIFIED_1D_U32_F32_LEVEL;
2636 case NVPTXISD::TexUnified1DU32FloatGrad:
2637 Opc = NVPTX::TEX_UNIFIED_1D_U32_F32_GRAD;
2639 case NVPTXISD::TexUnified1DArrayFloatS32:
2640 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_F32_S32;
2642 case NVPTXISD::TexUnified1DArrayFloatFloat:
2643 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_F32_F32;
2645 case NVPTXISD::TexUnified1DArrayFloatFloatLevel:
2646 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_F32_F32_LEVEL;
2648 case NVPTXISD::TexUnified1DArrayFloatFloatGrad:
2649 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_F32_F32_GRAD;
2651 case NVPTXISD::TexUnified1DArrayS32S32:
2652 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_S32_S32;
2654 case NVPTXISD::TexUnified1DArrayS32Float:
2655 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_S32_F32;
2657 case NVPTXISD::TexUnified1DArrayS32FloatLevel:
2658 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_S32_F32_LEVEL;
2660 case NVPTXISD::TexUnified1DArrayS32FloatGrad:
2661 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_S32_F32_GRAD;
2663 case NVPTXISD::TexUnified1DArrayU32S32:
2664 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_U32_S32;
2666 case NVPTXISD::TexUnified1DArrayU32Float:
2667 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_U32_F32;
2669 case NVPTXISD::TexUnified1DArrayU32FloatLevel:
2670 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_U32_F32_LEVEL;
2672 case NVPTXISD::TexUnified1DArrayU32FloatGrad:
2673 Opc = NVPTX::TEX_UNIFIED_1D_ARRAY_U32_F32_GRAD;
2675 case NVPTXISD::TexUnified2DFloatS32:
2676 Opc = NVPTX::TEX_UNIFIED_2D_F32_S32;
2678 case NVPTXISD::TexUnified2DFloatFloat:
2679 Opc = NVPTX::TEX_UNIFIED_2D_F32_F32;
2681 case NVPTXISD::TexUnified2DFloatFloatLevel:
2682 Opc = NVPTX::TEX_UNIFIED_2D_F32_F32_LEVEL;
2684 case NVPTXISD::TexUnified2DFloatFloatGrad:
2685 Opc = NVPTX::TEX_UNIFIED_2D_F32_F32_GRAD;
2687 case NVPTXISD::TexUnified2DS32S32:
2688 Opc = NVPTX::TEX_UNIFIED_2D_S32_S32;
2690 case NVPTXISD::TexUnified2DS32Float:
2691 Opc = NVPTX::TEX_UNIFIED_2D_S32_F32;
2693 case NVPTXISD::TexUnified2DS32FloatLevel:
2694 Opc = NVPTX::TEX_UNIFIED_2D_S32_F32_LEVEL;
2696 case NVPTXISD::TexUnified2DS32FloatGrad:
2697 Opc = NVPTX::TEX_UNIFIED_2D_S32_F32_GRAD;
2699 case NVPTXISD::TexUnified2DU32S32:
2700 Opc = NVPTX::TEX_UNIFIED_2D_U32_S32;
2702 case NVPTXISD::TexUnified2DU32Float:
2703 Opc = NVPTX::TEX_UNIFIED_2D_U32_F32;
2705 case NVPTXISD::TexUnified2DU32FloatLevel:
2706 Opc = NVPTX::TEX_UNIFIED_2D_U32_F32_LEVEL;
2708 case NVPTXISD::TexUnified2DU32FloatGrad:
2709 Opc = NVPTX::TEX_UNIFIED_2D_U32_F32_GRAD;
2711 case NVPTXISD::TexUnified2DArrayFloatS32:
2712 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_F32_S32;
2714 case NVPTXISD::TexUnified2DArrayFloatFloat:
2715 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_F32_F32;
2717 case NVPTXISD::TexUnified2DArrayFloatFloatLevel:
2718 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_F32_F32_LEVEL;
2720 case NVPTXISD::TexUnified2DArrayFloatFloatGrad:
2721 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_F32_F32_GRAD;
2723 case NVPTXISD::TexUnified2DArrayS32S32:
2724 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_S32_S32;
2726 case NVPTXISD::TexUnified2DArrayS32Float:
2727 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_S32_F32;
2729 case NVPTXISD::TexUnified2DArrayS32FloatLevel:
2730 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_S32_F32_LEVEL;
2732 case NVPTXISD::TexUnified2DArrayS32FloatGrad:
2733 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_S32_F32_GRAD;
2735 case NVPTXISD::TexUnified2DArrayU32S32:
2736 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_U32_S32;
2738 case NVPTXISD::TexUnified2DArrayU32Float:
2739 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_U32_F32;
2741 case NVPTXISD::TexUnified2DArrayU32FloatLevel:
2742 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_U32_F32_LEVEL;
2744 case NVPTXISD::TexUnified2DArrayU32FloatGrad:
2745 Opc = NVPTX::TEX_UNIFIED_2D_ARRAY_U32_F32_GRAD;
2747 case NVPTXISD::TexUnified3DFloatS32:
2748 Opc = NVPTX::TEX_UNIFIED_3D_F32_S32;
2750 case NVPTXISD::TexUnified3DFloatFloat:
2751 Opc = NVPTX::TEX_UNIFIED_3D_F32_F32;
2753 case NVPTXISD::TexUnified3DFloatFloatLevel:
2754 Opc = NVPTX::TEX_UNIFIED_3D_F32_F32_LEVEL;
2756 case NVPTXISD::TexUnified3DFloatFloatGrad:
2757 Opc = NVPTX::TEX_UNIFIED_3D_F32_F32_GRAD;
2759 case NVPTXISD::TexUnified3DS32S32:
2760 Opc = NVPTX::TEX_UNIFIED_3D_S32_S32;
2762 case NVPTXISD::TexUnified3DS32Float:
2763 Opc = NVPTX::TEX_UNIFIED_3D_S32_F32;
2765 case NVPTXISD::TexUnified3DS32FloatLevel:
2766 Opc = NVPTX::TEX_UNIFIED_3D_S32_F32_LEVEL;
2768 case NVPTXISD::TexUnified3DS32FloatGrad:
2769 Opc = NVPTX::TEX_UNIFIED_3D_S32_F32_GRAD;
2771 case NVPTXISD::TexUnified3DU32S32:
2772 Opc = NVPTX::TEX_UNIFIED_3D_U32_S32;
2774 case NVPTXISD::TexUnified3DU32Float:
2775 Opc = NVPTX::TEX_UNIFIED_3D_U32_F32;
2777 case NVPTXISD::TexUnified3DU32FloatLevel:
2778 Opc = NVPTX::TEX_UNIFIED_3D_U32_F32_LEVEL;
2780 case NVPTXISD::TexUnified3DU32FloatGrad:
2781 Opc = NVPTX::TEX_UNIFIED_3D_U32_F32_GRAD;
2783 case NVPTXISD::TexUnifiedCubeFloatFloat:
2784 Opc = NVPTX::TEX_UNIFIED_CUBE_F32_F32;
2786 case NVPTXISD::TexUnifiedCubeFloatFloatLevel:
2787 Opc = NVPTX::TEX_UNIFIED_CUBE_F32_F32_LEVEL;
2789 case NVPTXISD::TexUnifiedCubeS32Float:
2790 Opc = NVPTX::TEX_UNIFIED_CUBE_S32_F32;
2792 case NVPTXISD::TexUnifiedCubeS32FloatLevel:
2793 Opc = NVPTX::TEX_UNIFIED_CUBE_S32_F32_LEVEL;
2795 case NVPTXISD::TexUnifiedCubeU32Float:
2796 Opc = NVPTX::TEX_UNIFIED_CUBE_U32_F32;
2798 case NVPTXISD::TexUnifiedCubeU32FloatLevel:
2799 Opc = NVPTX::TEX_UNIFIED_CUBE_U32_F32_LEVEL;
2801 case NVPTXISD::TexUnifiedCubeArrayFloatFloat:
2802 Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_F32_F32;
2804 case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel:
2805 Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_F32_F32_LEVEL;
2807 case NVPTXISD::TexUnifiedCubeArrayS32Float:
2808 Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_S32_F32;
2810 case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel:
2811 Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_S32_F32_LEVEL;
2813 case NVPTXISD::TexUnifiedCubeArrayU32Float:
2814 Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_U32_F32;
2816 case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel:
2817 Opc = NVPTX::TEX_UNIFIED_CUBE_ARRAY_U32_F32_LEVEL;
2819 case NVPTXISD::Tld4UnifiedR2DFloatFloat:
2820 Opc = NVPTX::TLD4_UNIFIED_R_2D_F32_F32;
2822 case NVPTXISD::Tld4UnifiedG2DFloatFloat:
2823 Opc = NVPTX::TLD4_UNIFIED_G_2D_F32_F32;
2825 case NVPTXISD::Tld4UnifiedB2DFloatFloat:
2826 Opc = NVPTX::TLD4_UNIFIED_B_2D_F32_F32;
2828 case NVPTXISD::Tld4UnifiedA2DFloatFloat:
2829 Opc = NVPTX::TLD4_UNIFIED_A_2D_F32_F32;
2831 case NVPTXISD::Tld4UnifiedR2DS64Float:
2832 Opc = NVPTX::TLD4_UNIFIED_R_2D_S32_F32;
2834 case NVPTXISD::Tld4UnifiedG2DS64Float:
2835 Opc = NVPTX::TLD4_UNIFIED_G_2D_S32_F32;
2837 case NVPTXISD::Tld4UnifiedB2DS64Float:
2838 Opc = NVPTX::TLD4_UNIFIED_B_2D_S32_F32;
2840 case NVPTXISD::Tld4UnifiedA2DS64Float:
2841 Opc = NVPTX::TLD4_UNIFIED_A_2D_S32_F32;
2843 case NVPTXISD::Tld4UnifiedR2DU64Float:
2844 Opc = NVPTX::TLD4_UNIFIED_R_2D_U32_F32;
2846 case NVPTXISD::Tld4UnifiedG2DU64Float:
2847 Opc = NVPTX::TLD4_UNIFIED_G_2D_U32_F32;
2849 case NVPTXISD::Tld4UnifiedB2DU64Float:
2850 Opc = NVPTX::TLD4_UNIFIED_B_2D_U32_F32;
2852 case NVPTXISD::Tld4UnifiedA2DU64Float:
2853 Opc = NVPTX::TLD4_UNIFIED_A_2D_U32_F32;
2857 // Copy over operands
2858 SmallVector<SDValue, 8> Ops(N->op_begin() + 1, N->op_end());
2859 Ops.push_back(N->getOperand(0)); // Move chain to the back.
2861 ReplaceNode(N, CurDAG->getMachineNode(Opc, SDLoc(N), N->getVTList(), Ops));
2865 bool NVPTXDAGToDAGISel::trySurfaceIntrinsic(SDNode *N) {
2867 switch (N->getOpcode()) {
2868 default: return false;
2869 case NVPTXISD::Suld1DI8Clamp:
2870 Opc = NVPTX::SULD_1D_I8_CLAMP;
2872 case NVPTXISD::Suld1DI16Clamp:
2873 Opc = NVPTX::SULD_1D_I16_CLAMP;
2875 case NVPTXISD::Suld1DI32Clamp:
2876 Opc = NVPTX::SULD_1D_I32_CLAMP;
2878 case NVPTXISD::Suld1DI64Clamp:
2879 Opc = NVPTX::SULD_1D_I64_CLAMP;
2881 case NVPTXISD::Suld1DV2I8Clamp:
2882 Opc = NVPTX::SULD_1D_V2I8_CLAMP;
2884 case NVPTXISD::Suld1DV2I16Clamp:
2885 Opc = NVPTX::SULD_1D_V2I16_CLAMP;
2887 case NVPTXISD::Suld1DV2I32Clamp:
2888 Opc = NVPTX::SULD_1D_V2I32_CLAMP;
2890 case NVPTXISD::Suld1DV2I64Clamp:
2891 Opc = NVPTX::SULD_1D_V2I64_CLAMP;
2893 case NVPTXISD::Suld1DV4I8Clamp:
2894 Opc = NVPTX::SULD_1D_V4I8_CLAMP;
2896 case NVPTXISD::Suld1DV4I16Clamp:
2897 Opc = NVPTX::SULD_1D_V4I16_CLAMP;
2899 case NVPTXISD::Suld1DV4I32Clamp:
2900 Opc = NVPTX::SULD_1D_V4I32_CLAMP;
2902 case NVPTXISD::Suld1DArrayI8Clamp:
2903 Opc = NVPTX::SULD_1D_ARRAY_I8_CLAMP;
2905 case NVPTXISD::Suld1DArrayI16Clamp:
2906 Opc = NVPTX::SULD_1D_ARRAY_I16_CLAMP;
2908 case NVPTXISD::Suld1DArrayI32Clamp:
2909 Opc = NVPTX::SULD_1D_ARRAY_I32_CLAMP;
2911 case NVPTXISD::Suld1DArrayI64Clamp:
2912 Opc = NVPTX::SULD_1D_ARRAY_I64_CLAMP;
2914 case NVPTXISD::Suld1DArrayV2I8Clamp:
2915 Opc = NVPTX::SULD_1D_ARRAY_V2I8_CLAMP;
2917 case NVPTXISD::Suld1DArrayV2I16Clamp:
2918 Opc = NVPTX::SULD_1D_ARRAY_V2I16_CLAMP;
2920 case NVPTXISD::Suld1DArrayV2I32Clamp:
2921 Opc = NVPTX::SULD_1D_ARRAY_V2I32_CLAMP;
2923 case NVPTXISD::Suld1DArrayV2I64Clamp:
2924 Opc = NVPTX::SULD_1D_ARRAY_V2I64_CLAMP;
2926 case NVPTXISD::Suld1DArrayV4I8Clamp:
2927 Opc = NVPTX::SULD_1D_ARRAY_V4I8_CLAMP;
2929 case NVPTXISD::Suld1DArrayV4I16Clamp:
2930 Opc = NVPTX::SULD_1D_ARRAY_V4I16_CLAMP;
2932 case NVPTXISD::Suld1DArrayV4I32Clamp:
2933 Opc = NVPTX::SULD_1D_ARRAY_V4I32_CLAMP;
2935 case NVPTXISD::Suld2DI8Clamp:
2936 Opc = NVPTX::SULD_2D_I8_CLAMP;
2938 case NVPTXISD::Suld2DI16Clamp:
2939 Opc = NVPTX::SULD_2D_I16_CLAMP;
2941 case NVPTXISD::Suld2DI32Clamp:
2942 Opc = NVPTX::SULD_2D_I32_CLAMP;
2944 case NVPTXISD::Suld2DI64Clamp:
2945 Opc = NVPTX::SULD_2D_I64_CLAMP;
2947 case NVPTXISD::Suld2DV2I8Clamp:
2948 Opc = NVPTX::SULD_2D_V2I8_CLAMP;
2950 case NVPTXISD::Suld2DV2I16Clamp:
2951 Opc = NVPTX::SULD_2D_V2I16_CLAMP;
2953 case NVPTXISD::Suld2DV2I32Clamp:
2954 Opc = NVPTX::SULD_2D_V2I32_CLAMP;
2956 case NVPTXISD::Suld2DV2I64Clamp:
2957 Opc = NVPTX::SULD_2D_V2I64_CLAMP;
2959 case NVPTXISD::Suld2DV4I8Clamp:
2960 Opc = NVPTX::SULD_2D_V4I8_CLAMP;
2962 case NVPTXISD::Suld2DV4I16Clamp:
2963 Opc = NVPTX::SULD_2D_V4I16_CLAMP;
2965 case NVPTXISD::Suld2DV4I32Clamp:
2966 Opc = NVPTX::SULD_2D_V4I32_CLAMP;
2968 case NVPTXISD::Suld2DArrayI8Clamp:
2969 Opc = NVPTX::SULD_2D_ARRAY_I8_CLAMP;
2971 case NVPTXISD::Suld2DArrayI16Clamp:
2972 Opc = NVPTX::SULD_2D_ARRAY_I16_CLAMP;
2974 case NVPTXISD::Suld2DArrayI32Clamp:
2975 Opc = NVPTX::SULD_2D_ARRAY_I32_CLAMP;
2977 case NVPTXISD::Suld2DArrayI64Clamp:
2978 Opc = NVPTX::SULD_2D_ARRAY_I64_CLAMP;
2980 case NVPTXISD::Suld2DArrayV2I8Clamp:
2981 Opc = NVPTX::SULD_2D_ARRAY_V2I8_CLAMP;
2983 case NVPTXISD::Suld2DArrayV2I16Clamp:
2984 Opc = NVPTX::SULD_2D_ARRAY_V2I16_CLAMP;
2986 case NVPTXISD::Suld2DArrayV2I32Clamp:
2987 Opc = NVPTX::SULD_2D_ARRAY_V2I32_CLAMP;
2989 case NVPTXISD::Suld2DArrayV2I64Clamp:
2990 Opc = NVPTX::SULD_2D_ARRAY_V2I64_CLAMP;
2992 case NVPTXISD::Suld2DArrayV4I8Clamp:
2993 Opc = NVPTX::SULD_2D_ARRAY_V4I8_CLAMP;
2995 case NVPTXISD::Suld2DArrayV4I16Clamp:
2996 Opc = NVPTX::SULD_2D_ARRAY_V4I16_CLAMP;
2998 case NVPTXISD::Suld2DArrayV4I32Clamp:
2999 Opc = NVPTX::SULD_2D_ARRAY_V4I32_CLAMP;
3001 case NVPTXISD::Suld3DI8Clamp:
3002 Opc = NVPTX::SULD_3D_I8_CLAMP;
3004 case NVPTXISD::Suld3DI16Clamp:
3005 Opc = NVPTX::SULD_3D_I16_CLAMP;
3007 case NVPTXISD::Suld3DI32Clamp:
3008 Opc = NVPTX::SULD_3D_I32_CLAMP;
3010 case NVPTXISD::Suld3DI64Clamp:
3011 Opc = NVPTX::SULD_3D_I64_CLAMP;
3013 case NVPTXISD::Suld3DV2I8Clamp:
3014 Opc = NVPTX::SULD_3D_V2I8_CLAMP;
3016 case NVPTXISD::Suld3DV2I16Clamp:
3017 Opc = NVPTX::SULD_3D_V2I16_CLAMP;
3019 case NVPTXISD::Suld3DV2I32Clamp:
3020 Opc = NVPTX::SULD_3D_V2I32_CLAMP;
3022 case NVPTXISD::Suld3DV2I64Clamp:
3023 Opc = NVPTX::SULD_3D_V2I64_CLAMP;
3025 case NVPTXISD::Suld3DV4I8Clamp:
3026 Opc = NVPTX::SULD_3D_V4I8_CLAMP;
3028 case NVPTXISD::Suld3DV4I16Clamp:
3029 Opc = NVPTX::SULD_3D_V4I16_CLAMP;
3031 case NVPTXISD::Suld3DV4I32Clamp:
3032 Opc = NVPTX::SULD_3D_V4I32_CLAMP;
3034 case NVPTXISD::Suld1DI8Trap:
3035 Opc = NVPTX::SULD_1D_I8_TRAP;
3037 case NVPTXISD::Suld1DI16Trap:
3038 Opc = NVPTX::SULD_1D_I16_TRAP;
3040 case NVPTXISD::Suld1DI32Trap:
3041 Opc = NVPTX::SULD_1D_I32_TRAP;
3043 case NVPTXISD::Suld1DI64Trap:
3044 Opc = NVPTX::SULD_1D_I64_TRAP;
3046 case NVPTXISD::Suld1DV2I8Trap:
3047 Opc = NVPTX::SULD_1D_V2I8_TRAP;
3049 case NVPTXISD::Suld1DV2I16Trap:
3050 Opc = NVPTX::SULD_1D_V2I16_TRAP;
3052 case NVPTXISD::Suld1DV2I32Trap:
3053 Opc = NVPTX::SULD_1D_V2I32_TRAP;
3055 case NVPTXISD::Suld1DV2I64Trap:
3056 Opc = NVPTX::SULD_1D_V2I64_TRAP;
3058 case NVPTXISD::Suld1DV4I8Trap:
3059 Opc = NVPTX::SULD_1D_V4I8_TRAP;
3061 case NVPTXISD::Suld1DV4I16Trap:
3062 Opc = NVPTX::SULD_1D_V4I16_TRAP;
3064 case NVPTXISD::Suld1DV4I32Trap:
3065 Opc = NVPTX::SULD_1D_V4I32_TRAP;
3067 case NVPTXISD::Suld1DArrayI8Trap:
3068 Opc = NVPTX::SULD_1D_ARRAY_I8_TRAP;
3070 case NVPTXISD::Suld1DArrayI16Trap:
3071 Opc = NVPTX::SULD_1D_ARRAY_I16_TRAP;
3073 case NVPTXISD::Suld1DArrayI32Trap:
3074 Opc = NVPTX::SULD_1D_ARRAY_I32_TRAP;
3076 case NVPTXISD::Suld1DArrayI64Trap:
3077 Opc = NVPTX::SULD_1D_ARRAY_I64_TRAP;
3079 case NVPTXISD::Suld1DArrayV2I8Trap:
3080 Opc = NVPTX::SULD_1D_ARRAY_V2I8_TRAP;
3082 case NVPTXISD::Suld1DArrayV2I16Trap:
3083 Opc = NVPTX::SULD_1D_ARRAY_V2I16_TRAP;
3085 case NVPTXISD::Suld1DArrayV2I32Trap:
3086 Opc = NVPTX::SULD_1D_ARRAY_V2I32_TRAP;
3088 case NVPTXISD::Suld1DArrayV2I64Trap:
3089 Opc = NVPTX::SULD_1D_ARRAY_V2I64_TRAP;
3091 case NVPTXISD::Suld1DArrayV4I8Trap:
3092 Opc = NVPTX::SULD_1D_ARRAY_V4I8_TRAP;
3094 case NVPTXISD::Suld1DArrayV4I16Trap:
3095 Opc = NVPTX::SULD_1D_ARRAY_V4I16_TRAP;
3097 case NVPTXISD::Suld1DArrayV4I32Trap:
3098 Opc = NVPTX::SULD_1D_ARRAY_V4I32_TRAP;
3100 case NVPTXISD::Suld2DI8Trap:
3101 Opc = NVPTX::SULD_2D_I8_TRAP;
3103 case NVPTXISD::Suld2DI16Trap:
3104 Opc = NVPTX::SULD_2D_I16_TRAP;
3106 case NVPTXISD::Suld2DI32Trap:
3107 Opc = NVPTX::SULD_2D_I32_TRAP;
3109 case NVPTXISD::Suld2DI64Trap:
3110 Opc = NVPTX::SULD_2D_I64_TRAP;
3112 case NVPTXISD::Suld2DV2I8Trap:
3113 Opc = NVPTX::SULD_2D_V2I8_TRAP;
3115 case NVPTXISD::Suld2DV2I16Trap:
3116 Opc = NVPTX::SULD_2D_V2I16_TRAP;
3118 case NVPTXISD::Suld2DV2I32Trap:
3119 Opc = NVPTX::SULD_2D_V2I32_TRAP;
3121 case NVPTXISD::Suld2DV2I64Trap:
3122 Opc = NVPTX::SULD_2D_V2I64_TRAP;
3124 case NVPTXISD::Suld2DV4I8Trap:
3125 Opc = NVPTX::SULD_2D_V4I8_TRAP;
3127 case NVPTXISD::Suld2DV4I16Trap:
3128 Opc = NVPTX::SULD_2D_V4I16_TRAP;
3130 case NVPTXISD::Suld2DV4I32Trap:
3131 Opc = NVPTX::SULD_2D_V4I32_TRAP;
3133 case NVPTXISD::Suld2DArrayI8Trap:
3134 Opc = NVPTX::SULD_2D_ARRAY_I8_TRAP;
3136 case NVPTXISD::Suld2DArrayI16Trap:
3137 Opc = NVPTX::SULD_2D_ARRAY_I16_TRAP;
3139 case NVPTXISD::Suld2DArrayI32Trap:
3140 Opc = NVPTX::SULD_2D_ARRAY_I32_TRAP;
3142 case NVPTXISD::Suld2DArrayI64Trap:
3143 Opc = NVPTX::SULD_2D_ARRAY_I64_TRAP;
3145 case NVPTXISD::Suld2DArrayV2I8Trap:
3146 Opc = NVPTX::SULD_2D_ARRAY_V2I8_TRAP;
3148 case NVPTXISD::Suld2DArrayV2I16Trap:
3149 Opc = NVPTX::SULD_2D_ARRAY_V2I16_TRAP;
3151 case NVPTXISD::Suld2DArrayV2I32Trap:
3152 Opc = NVPTX::SULD_2D_ARRAY_V2I32_TRAP;
3154 case NVPTXISD::Suld2DArrayV2I64Trap:
3155 Opc = NVPTX::SULD_2D_ARRAY_V2I64_TRAP;
3157 case NVPTXISD::Suld2DArrayV4I8Trap:
3158 Opc = NVPTX::SULD_2D_ARRAY_V4I8_TRAP;
3160 case NVPTXISD::Suld2DArrayV4I16Trap:
3161 Opc = NVPTX::SULD_2D_ARRAY_V4I16_TRAP;
3163 case NVPTXISD::Suld2DArrayV4I32Trap:
3164 Opc = NVPTX::SULD_2D_ARRAY_V4I32_TRAP;
3166 case NVPTXISD::Suld3DI8Trap:
3167 Opc = NVPTX::SULD_3D_I8_TRAP;
3169 case NVPTXISD::Suld3DI16Trap:
3170 Opc = NVPTX::SULD_3D_I16_TRAP;
3172 case NVPTXISD::Suld3DI32Trap:
3173 Opc = NVPTX::SULD_3D_I32_TRAP;
3175 case NVPTXISD::Suld3DI64Trap:
3176 Opc = NVPTX::SULD_3D_I64_TRAP;
3178 case NVPTXISD::Suld3DV2I8Trap:
3179 Opc = NVPTX::SULD_3D_V2I8_TRAP;
3181 case NVPTXISD::Suld3DV2I16Trap:
3182 Opc = NVPTX::SULD_3D_V2I16_TRAP;
3184 case NVPTXISD::Suld3DV2I32Trap:
3185 Opc = NVPTX::SULD_3D_V2I32_TRAP;
3187 case NVPTXISD::Suld3DV2I64Trap:
3188 Opc = NVPTX::SULD_3D_V2I64_TRAP;
3190 case NVPTXISD::Suld3DV4I8Trap:
3191 Opc = NVPTX::SULD_3D_V4I8_TRAP;
3193 case NVPTXISD::Suld3DV4I16Trap:
3194 Opc = NVPTX::SULD_3D_V4I16_TRAP;
3196 case NVPTXISD::Suld3DV4I32Trap:
3197 Opc = NVPTX::SULD_3D_V4I32_TRAP;
3199 case NVPTXISD::Suld1DI8Zero:
3200 Opc = NVPTX::SULD_1D_I8_ZERO;
3202 case NVPTXISD::Suld1DI16Zero:
3203 Opc = NVPTX::SULD_1D_I16_ZERO;
3205 case NVPTXISD::Suld1DI32Zero:
3206 Opc = NVPTX::SULD_1D_I32_ZERO;
3208 case NVPTXISD::Suld1DI64Zero:
3209 Opc = NVPTX::SULD_1D_I64_ZERO;
3211 case NVPTXISD::Suld1DV2I8Zero:
3212 Opc = NVPTX::SULD_1D_V2I8_ZERO;
3214 case NVPTXISD::Suld1DV2I16Zero:
3215 Opc = NVPTX::SULD_1D_V2I16_ZERO;
3217 case NVPTXISD::Suld1DV2I32Zero:
3218 Opc = NVPTX::SULD_1D_V2I32_ZERO;
3220 case NVPTXISD::Suld1DV2I64Zero:
3221 Opc = NVPTX::SULD_1D_V2I64_ZERO;
3223 case NVPTXISD::Suld1DV4I8Zero:
3224 Opc = NVPTX::SULD_1D_V4I8_ZERO;
3226 case NVPTXISD::Suld1DV4I16Zero:
3227 Opc = NVPTX::SULD_1D_V4I16_ZERO;
3229 case NVPTXISD::Suld1DV4I32Zero:
3230 Opc = NVPTX::SULD_1D_V4I32_ZERO;
3232 case NVPTXISD::Suld1DArrayI8Zero:
3233 Opc = NVPTX::SULD_1D_ARRAY_I8_ZERO;
3235 case NVPTXISD::Suld1DArrayI16Zero:
3236 Opc = NVPTX::SULD_1D_ARRAY_I16_ZERO;
3238 case NVPTXISD::Suld1DArrayI32Zero:
3239 Opc = NVPTX::SULD_1D_ARRAY_I32_ZERO;
3241 case NVPTXISD::Suld1DArrayI64Zero:
3242 Opc = NVPTX::SULD_1D_ARRAY_I64_ZERO;
3244 case NVPTXISD::Suld1DArrayV2I8Zero:
3245 Opc = NVPTX::SULD_1D_ARRAY_V2I8_ZERO;
3247 case NVPTXISD::Suld1DArrayV2I16Zero:
3248 Opc = NVPTX::SULD_1D_ARRAY_V2I16_ZERO;
3250 case NVPTXISD::Suld1DArrayV2I32Zero:
3251 Opc = NVPTX::SULD_1D_ARRAY_V2I32_ZERO;
3253 case NVPTXISD::Suld1DArrayV2I64Zero:
3254 Opc = NVPTX::SULD_1D_ARRAY_V2I64_ZERO;
3256 case NVPTXISD::Suld1DArrayV4I8Zero:
3257 Opc = NVPTX::SULD_1D_ARRAY_V4I8_ZERO;
3259 case NVPTXISD::Suld1DArrayV4I16Zero:
3260 Opc = NVPTX::SULD_1D_ARRAY_V4I16_ZERO;
3262 case NVPTXISD::Suld1DArrayV4I32Zero:
3263 Opc = NVPTX::SULD_1D_ARRAY_V4I32_ZERO;
3265 case NVPTXISD::Suld2DI8Zero:
3266 Opc = NVPTX::SULD_2D_I8_ZERO;
3268 case NVPTXISD::Suld2DI16Zero:
3269 Opc = NVPTX::SULD_2D_I16_ZERO;
3271 case NVPTXISD::Suld2DI32Zero:
3272 Opc = NVPTX::SULD_2D_I32_ZERO;
3274 case NVPTXISD::Suld2DI64Zero:
3275 Opc = NVPTX::SULD_2D_I64_ZERO;
3277 case NVPTXISD::Suld2DV2I8Zero:
3278 Opc = NVPTX::SULD_2D_V2I8_ZERO;
3280 case NVPTXISD::Suld2DV2I16Zero:
3281 Opc = NVPTX::SULD_2D_V2I16_ZERO;
3283 case NVPTXISD::Suld2DV2I32Zero:
3284 Opc = NVPTX::SULD_2D_V2I32_ZERO;
3286 case NVPTXISD::Suld2DV2I64Zero:
3287 Opc = NVPTX::SULD_2D_V2I64_ZERO;
3289 case NVPTXISD::Suld2DV4I8Zero:
3290 Opc = NVPTX::SULD_2D_V4I8_ZERO;
3292 case NVPTXISD::Suld2DV4I16Zero:
3293 Opc = NVPTX::SULD_2D_V4I16_ZERO;
3295 case NVPTXISD::Suld2DV4I32Zero:
3296 Opc = NVPTX::SULD_2D_V4I32_ZERO;
3298 case NVPTXISD::Suld2DArrayI8Zero:
3299 Opc = NVPTX::SULD_2D_ARRAY_I8_ZERO;
3301 case NVPTXISD::Suld2DArrayI16Zero:
3302 Opc = NVPTX::SULD_2D_ARRAY_I16_ZERO;
3304 case NVPTXISD::Suld2DArrayI32Zero:
3305 Opc = NVPTX::SULD_2D_ARRAY_I32_ZERO;
3307 case NVPTXISD::Suld2DArrayI64Zero:
3308 Opc = NVPTX::SULD_2D_ARRAY_I64_ZERO;
3310 case NVPTXISD::Suld2DArrayV2I8Zero:
3311 Opc = NVPTX::SULD_2D_ARRAY_V2I8_ZERO;
3313 case NVPTXISD::Suld2DArrayV2I16Zero:
3314 Opc = NVPTX::SULD_2D_ARRAY_V2I16_ZERO;
3316 case NVPTXISD::Suld2DArrayV2I32Zero:
3317 Opc = NVPTX::SULD_2D_ARRAY_V2I32_ZERO;
3319 case NVPTXISD::Suld2DArrayV2I64Zero:
3320 Opc = NVPTX::SULD_2D_ARRAY_V2I64_ZERO;
3322 case NVPTXISD::Suld2DArrayV4I8Zero:
3323 Opc = NVPTX::SULD_2D_ARRAY_V4I8_ZERO;
3325 case NVPTXISD::Suld2DArrayV4I16Zero:
3326 Opc = NVPTX::SULD_2D_ARRAY_V4I16_ZERO;
3328 case NVPTXISD::Suld2DArrayV4I32Zero:
3329 Opc = NVPTX::SULD_2D_ARRAY_V4I32_ZERO;
3331 case NVPTXISD::Suld3DI8Zero:
3332 Opc = NVPTX::SULD_3D_I8_ZERO;
3334 case NVPTXISD::Suld3DI16Zero:
3335 Opc = NVPTX::SULD_3D_I16_ZERO;
3337 case NVPTXISD::Suld3DI32Zero:
3338 Opc = NVPTX::SULD_3D_I32_ZERO;
3340 case NVPTXISD::Suld3DI64Zero:
3341 Opc = NVPTX::SULD_3D_I64_ZERO;
3343 case NVPTXISD::Suld3DV2I8Zero:
3344 Opc = NVPTX::SULD_3D_V2I8_ZERO;
3346 case NVPTXISD::Suld3DV2I16Zero:
3347 Opc = NVPTX::SULD_3D_V2I16_ZERO;
3349 case NVPTXISD::Suld3DV2I32Zero:
3350 Opc = NVPTX::SULD_3D_V2I32_ZERO;
3352 case NVPTXISD::Suld3DV2I64Zero:
3353 Opc = NVPTX::SULD_3D_V2I64_ZERO;
3355 case NVPTXISD::Suld3DV4I8Zero:
3356 Opc = NVPTX::SULD_3D_V4I8_ZERO;
3358 case NVPTXISD::Suld3DV4I16Zero:
3359 Opc = NVPTX::SULD_3D_V4I16_ZERO;
3361 case NVPTXISD::Suld3DV4I32Zero:
3362 Opc = NVPTX::SULD_3D_V4I32_ZERO;
3366 // Copy over operands
3367 SmallVector<SDValue, 8> Ops(N->op_begin() + 1, N->op_end());
3368 Ops.push_back(N->getOperand(0)); // Move chain to the back.
3370 ReplaceNode(N, CurDAG->getMachineNode(Opc, SDLoc(N), N->getVTList(), Ops));
3375 /// SelectBFE - Look for instruction sequences that can be made more efficient
3376 /// by using the 'bfe' (bit-field extract) PTX instruction
3377 bool NVPTXDAGToDAGISel::tryBFE(SDNode *N) {
3379 SDValue LHS = N->getOperand(0);
3380 SDValue RHS = N->getOperand(1);
3384 bool IsSigned = false;
3386 if (N->getOpcode() == ISD::AND) {
3387 // Canonicalize the operands
3388 // We want 'and %val, %mask'
3389 if (isa<ConstantSDNode>(LHS) && !isa<ConstantSDNode>(RHS)) {
3390 std::swap(LHS, RHS);
3393 ConstantSDNode *Mask = dyn_cast<ConstantSDNode>(RHS);
3395 // We need a constant mask on the RHS of the AND
3399 // Extract the mask bits
3400 uint64_t MaskVal = Mask->getZExtValue();
3401 if (!isMask_64(MaskVal)) {
3402 // We *could* handle shifted masks here, but doing so would require an
3403 // 'and' operation to fix up the low-order bits so we would trade
3404 // shr+and for bfe+and, which has the same throughput
3408 // How many bits are in our mask?
3409 uint64_t NumBits = countTrailingOnes(MaskVal);
3410 Len = CurDAG->getTargetConstant(NumBits, DL, MVT::i32);
3412 if (LHS.getOpcode() == ISD::SRL || LHS.getOpcode() == ISD::SRA) {
3413 // We have a 'srl/and' pair, extract the effective start bit and length
3414 Val = LHS.getNode()->getOperand(0);
3415 Start = LHS.getNode()->getOperand(1);
3416 ConstantSDNode *StartConst = dyn_cast<ConstantSDNode>(Start);
3418 uint64_t StartVal = StartConst->getZExtValue();
3419 // How many "good" bits do we have left? "good" is defined here as bits
3420 // that exist in the original value, not shifted in.
3421 uint64_t GoodBits = Start.getValueSizeInBits() - StartVal;
3422 if (NumBits > GoodBits) {
3423 // Do not handle the case where bits have been shifted in. In theory
3424 // we could handle this, but the cost is likely higher than just
3425 // emitting the srl/and pair.
3428 Start = CurDAG->getTargetConstant(StartVal, DL, MVT::i32);
3430 // Do not handle the case where the shift amount (can be zero if no srl
3431 // was found) is not constant. We could handle this case, but it would
3432 // require run-time logic that would be more expensive than just
3433 // emitting the srl/and pair.
3437 // Do not handle the case where the LHS of the and is not a shift. While
3438 // it would be trivial to handle this case, it would just transform
3439 // 'and' -> 'bfe', but 'and' has higher-throughput.
3442 } else if (N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) {
3443 if (LHS->getOpcode() == ISD::AND) {
3444 ConstantSDNode *ShiftCnst = dyn_cast<ConstantSDNode>(RHS);
3446 // Shift amount must be constant
3450 uint64_t ShiftAmt = ShiftCnst->getZExtValue();
3452 SDValue AndLHS = LHS->getOperand(0);
3453 SDValue AndRHS = LHS->getOperand(1);
3455 // Canonicalize the AND to have the mask on the RHS
3456 if (isa<ConstantSDNode>(AndLHS)) {
3457 std::swap(AndLHS, AndRHS);
3460 ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(AndRHS);
3462 // Mask must be constant
3466 uint64_t MaskVal = MaskCnst->getZExtValue();
3469 if (isMask_64(MaskVal)) {
3471 // The number of bits in the result bitfield will be the number of
3472 // trailing ones (the AND) minus the number of bits we shift off
3473 NumBits = countTrailingOnes(MaskVal) - ShiftAmt;
3474 } else if (isShiftedMask_64(MaskVal)) {
3475 NumZeros = countTrailingZeros(MaskVal);
3476 unsigned NumOnes = countTrailingOnes(MaskVal >> NumZeros);
3477 // The number of bits in the result bitfield will be the number of
3478 // trailing zeros plus the number of set bits in the mask minus the
3479 // number of bits we shift off
3480 NumBits = NumZeros + NumOnes - ShiftAmt;
3482 // This is not a mask we can handle
3486 if (ShiftAmt < NumZeros) {
3487 // Handling this case would require extra logic that would make this
3488 // transformation non-profitable
3493 Start = CurDAG->getTargetConstant(ShiftAmt, DL, MVT::i32);
3494 Len = CurDAG->getTargetConstant(NumBits, DL, MVT::i32);
3495 } else if (LHS->getOpcode() == ISD::SHL) {
3496 // Here, we have a pattern like:
3498 // (sra (shl val, NN), MM)
3500 // (srl (shl val, NN), MM)
3502 // If MM >= NN, we can efficiently optimize this with bfe
3503 Val = LHS->getOperand(0);
3505 SDValue ShlRHS = LHS->getOperand(1);
3506 ConstantSDNode *ShlCnst = dyn_cast<ConstantSDNode>(ShlRHS);
3508 // Shift amount must be constant
3511 uint64_t InnerShiftAmt = ShlCnst->getZExtValue();
3513 SDValue ShrRHS = RHS;
3514 ConstantSDNode *ShrCnst = dyn_cast<ConstantSDNode>(ShrRHS);
3516 // Shift amount must be constant
3519 uint64_t OuterShiftAmt = ShrCnst->getZExtValue();
3521 // To avoid extra codegen and be profitable, we need Outer >= Inner
3522 if (OuterShiftAmt < InnerShiftAmt) {
3526 // If the outer shift is more than the type size, we have no bitfield to
3527 // extract (since we also check that the inner shift is <= the outer shift
3528 // then this also implies that the inner shift is < the type size)
3529 if (OuterShiftAmt >= Val.getValueSizeInBits()) {
3533 Start = CurDAG->getTargetConstant(OuterShiftAmt - InnerShiftAmt, DL,
3535 Len = CurDAG->getTargetConstant(Val.getValueSizeInBits() - OuterShiftAmt,
3538 if (N->getOpcode() == ISD::SRA) {
3539 // If we have a arithmetic right shift, we need to use the signed bfe
3554 // For the BFE operations we form here from "and" and "srl", always use the
3555 // unsigned variants.
3556 if (Val.getValueType() == MVT::i32) {
3558 Opc = NVPTX::BFE_S32rii;
3560 Opc = NVPTX::BFE_U32rii;
3562 } else if (Val.getValueType() == MVT::i64) {
3564 Opc = NVPTX::BFE_S64rii;
3566 Opc = NVPTX::BFE_U64rii;
3569 // We cannot handle this type
3577 ReplaceNode(N, CurDAG->getMachineNode(Opc, DL, N->getVTList(), Ops));
3581 // SelectDirectAddr - Match a direct address for DAG.
3582 // A direct address could be a globaladdress or externalsymbol.
3583 bool NVPTXDAGToDAGISel::SelectDirectAddr(SDValue N, SDValue &Address) {
3584 // Return true if TGA or ES.
3585 if (N.getOpcode() == ISD::TargetGlobalAddress ||
3586 N.getOpcode() == ISD::TargetExternalSymbol) {
3590 if (N.getOpcode() == NVPTXISD::Wrapper) {
3591 Address = N.getOperand(0);
3594 // addrspacecast(MoveParam(arg_symbol) to addrspace(PARAM)) -> arg_symbol
3595 if (AddrSpaceCastSDNode *CastN = dyn_cast<AddrSpaceCastSDNode>(N)) {
3596 if (CastN->getSrcAddressSpace() == ADDRESS_SPACE_GENERIC &&
3597 CastN->getDestAddressSpace() == ADDRESS_SPACE_PARAM &&
3598 CastN->getOperand(0).getOpcode() == NVPTXISD::MoveParam)
3599 return SelectDirectAddr(CastN->getOperand(0).getOperand(0), Address);
3605 bool NVPTXDAGToDAGISel::SelectADDRsi_imp(
3606 SDNode *OpNode, SDValue Addr, SDValue &Base, SDValue &Offset, MVT mvt) {
3607 if (Addr.getOpcode() == ISD::ADD) {
3608 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1))) {
3609 SDValue base = Addr.getOperand(0);
3610 if (SelectDirectAddr(base, Base)) {
3611 Offset = CurDAG->getTargetConstant(CN->getZExtValue(), SDLoc(OpNode),
3621 bool NVPTXDAGToDAGISel::SelectADDRsi(SDNode *OpNode, SDValue Addr,
3622 SDValue &Base, SDValue &Offset) {
3623 return SelectADDRsi_imp(OpNode, Addr, Base, Offset, MVT::i32);
3627 bool NVPTXDAGToDAGISel::SelectADDRsi64(SDNode *OpNode, SDValue Addr,
3628 SDValue &Base, SDValue &Offset) {
3629 return SelectADDRsi_imp(OpNode, Addr, Base, Offset, MVT::i64);
3633 bool NVPTXDAGToDAGISel::SelectADDRri_imp(
3634 SDNode *OpNode, SDValue Addr, SDValue &Base, SDValue &Offset, MVT mvt) {
3635 if (FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Addr)) {
3636 Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), mvt);
3637 Offset = CurDAG->getTargetConstant(0, SDLoc(OpNode), mvt);
3640 if (Addr.getOpcode() == ISD::TargetExternalSymbol ||
3641 Addr.getOpcode() == ISD::TargetGlobalAddress)
3642 return false; // direct calls.
3644 if (Addr.getOpcode() == ISD::ADD) {
3645 if (SelectDirectAddr(Addr.getOperand(0), Addr)) {
3648 if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(Addr.getOperand(1))) {
3649 if (FrameIndexSDNode *FIN =
3650 dyn_cast<FrameIndexSDNode>(Addr.getOperand(0)))
3651 // Constant offset from frame ref.
3652 Base = CurDAG->getTargetFrameIndex(FIN->getIndex(), mvt);
3654 Base = Addr.getOperand(0);
3655 Offset = CurDAG->getTargetConstant(CN->getZExtValue(), SDLoc(OpNode),
3664 bool NVPTXDAGToDAGISel::SelectADDRri(SDNode *OpNode, SDValue Addr,
3665 SDValue &Base, SDValue &Offset) {
3666 return SelectADDRri_imp(OpNode, Addr, Base, Offset, MVT::i32);
3670 bool NVPTXDAGToDAGISel::SelectADDRri64(SDNode *OpNode, SDValue Addr,
3671 SDValue &Base, SDValue &Offset) {
3672 return SelectADDRri_imp(OpNode, Addr, Base, Offset, MVT::i64);
3675 bool NVPTXDAGToDAGISel::ChkMemSDNodeAddressSpace(SDNode *N,
3676 unsigned int spN) const {
3677 const Value *Src = nullptr;
3678 if (MemSDNode *mN = dyn_cast<MemSDNode>(N)) {
3679 if (spN == 0 && mN->getMemOperand()->getPseudoValue())
3681 Src = mN->getMemOperand()->getValue();
3685 if (auto *PT = dyn_cast<PointerType>(Src->getType()))
3686 return (PT->getAddressSpace() == spN);
3690 /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
3691 /// inline asm expressions.
3692 bool NVPTXDAGToDAGISel::SelectInlineAsmMemoryOperand(
3693 const SDValue &Op, unsigned ConstraintID, std::vector<SDValue> &OutOps) {
3695 switch (ConstraintID) {
3698 case InlineAsm::Constraint_m: // memory
3699 if (SelectDirectAddr(Op, Op0)) {
3700 OutOps.push_back(Op0);
3701 OutOps.push_back(CurDAG->getTargetConstant(0, SDLoc(Op), MVT::i32));
3704 if (SelectADDRri(Op.getNode(), Op, Op0, Op1)) {
3705 OutOps.push_back(Op0);
3706 OutOps.push_back(Op1);
3714 /// GetConvertOpcode - Returns the CVT_ instruction opcode that implements a
3715 /// conversion from \p SrcTy to \p DestTy.
3716 unsigned NVPTXDAGToDAGISel::GetConvertOpcode(MVT DestTy, MVT SrcTy,
3718 switch (SrcTy.SimpleTy) {
3720 llvm_unreachable("Unhandled source type");
3722 switch (DestTy.SimpleTy) {
3724 llvm_unreachable("Unhandled dest type");
3726 return IsSigned ? NVPTX::CVT_s16_s8 : NVPTX::CVT_u16_u8;
3728 return IsSigned ? NVPTX::CVT_s32_s8 : NVPTX::CVT_u32_u8;
3730 return IsSigned ? NVPTX::CVT_s64_s8 : NVPTX::CVT_u64_u8;
3733 switch (DestTy.SimpleTy) {
3735 llvm_unreachable("Unhandled dest type");
3737 return IsSigned ? NVPTX::CVT_s8_s16 : NVPTX::CVT_u8_u16;
3739 return IsSigned ? NVPTX::CVT_s32_s16 : NVPTX::CVT_u32_u16;
3741 return IsSigned ? NVPTX::CVT_s64_s16 : NVPTX::CVT_u64_u16;
3744 switch (DestTy.SimpleTy) {
3746 llvm_unreachable("Unhandled dest type");
3748 return IsSigned ? NVPTX::CVT_s8_s32 : NVPTX::CVT_u8_u32;
3750 return IsSigned ? NVPTX::CVT_s16_s32 : NVPTX::CVT_u16_u32;
3752 return IsSigned ? NVPTX::CVT_s64_s32 : NVPTX::CVT_u64_u32;
3755 switch (DestTy.SimpleTy) {
3757 llvm_unreachable("Unhandled dest type");
3759 return IsSigned ? NVPTX::CVT_s8_s64 : NVPTX::CVT_u8_u64;
3761 return IsSigned ? NVPTX::CVT_s16_s64 : NVPTX::CVT_u16_u64;
3763 return IsSigned ? NVPTX::CVT_s32_s64 : NVPTX::CVT_u32_u64;