2 // The LLVM Compiler Infrastructure
4 // This file is distributed under the University of Illinois Open Source
5 // License. See LICENSE.TXT for details.
7 //===----------------------------------------------------------------------===//
9 // This file defines the interfaces that NVPTX uses to lower LLVM code into a
12 //===----------------------------------------------------------------------===//
14 #include "NVPTXISelLowering.h"
16 #include "NVPTXTargetMachine.h"
17 #include "NVPTXTargetObjectFile.h"
18 #include "NVPTXUtilities.h"
19 #include "llvm/CodeGen/Analysis.h"
20 #include "llvm/CodeGen/MachineFrameInfo.h"
21 #include "llvm/CodeGen/MachineFunction.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
25 #include "llvm/IR/CallSite.h"
26 #include "llvm/IR/DerivedTypes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GlobalValue.h"
29 #include "llvm/IR/IntrinsicInst.h"
30 #include "llvm/IR/Intrinsics.h"
31 #include "llvm/IR/Module.h"
32 #include "llvm/MC/MCSectionELF.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Support/Debug.h"
35 #include "llvm/Support/ErrorHandling.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/raw_ostream.h"
41 #define DEBUG_TYPE "nvptx-lower"
45 static unsigned int uniqueCallSite = 0;
47 static cl::opt<bool> sched4reg(
49 cl::desc("NVPTX Specific: schedule for register pressue"), cl::init(false));
51 static cl::opt<unsigned>
52 FMAContractLevelOpt("nvptx-fma-level", cl::ZeroOrMore, cl::Hidden,
53 cl::desc("NVPTX Specific: FMA contraction (0: don't do it"
54 " 1: do it 2: do it aggressively"),
57 static bool IsPTXVectorType(MVT VT) {
58 switch (VT.SimpleTy) {
77 /// ComputePTXValueVTs - For the given Type \p Ty, returns the set of primitive
78 /// EVTs that compose it. Unlike ComputeValueVTs, this will break apart vectors
79 /// into their primitive components.
80 /// NOTE: This is a band-aid for code that expects ComputeValueVTs to return the
81 /// same number of types as the Ins/Outs arrays in LowerFormalArguments,
82 /// LowerCall, and LowerReturn.
83 static void ComputePTXValueVTs(const TargetLowering &TLI, const DataLayout &DL,
84 Type *Ty, SmallVectorImpl<EVT> &ValueVTs,
85 SmallVectorImpl<uint64_t> *Offsets = nullptr,
86 uint64_t StartingOffset = 0) {
87 SmallVector<EVT, 16> TempVTs;
88 SmallVector<uint64_t, 16> TempOffsets;
90 ComputeValueVTs(TLI, DL, Ty, TempVTs, &TempOffsets, StartingOffset);
91 for (unsigned i = 0, e = TempVTs.size(); i != e; ++i) {
93 uint64_t Off = TempOffsets[i];
95 for (unsigned j = 0, je = VT.getVectorNumElements(); j != je; ++j) {
96 ValueVTs.push_back(VT.getVectorElementType());
98 Offsets->push_back(Off+j*VT.getVectorElementType().getStoreSize());
101 ValueVTs.push_back(VT);
103 Offsets->push_back(Off);
108 // NVPTXTargetLowering Constructor.
109 NVPTXTargetLowering::NVPTXTargetLowering(const NVPTXTargetMachine &TM,
110 const NVPTXSubtarget &STI)
111 : TargetLowering(TM), nvTM(&TM), STI(STI) {
113 // always lower memset, memcpy, and memmove intrinsics to load/store
114 // instructions, rather
115 // then generating calls to memset, mempcy or memmove.
116 MaxStoresPerMemset = (unsigned) 0xFFFFFFFF;
117 MaxStoresPerMemcpy = (unsigned) 0xFFFFFFFF;
118 MaxStoresPerMemmove = (unsigned) 0xFFFFFFFF;
120 setBooleanContents(ZeroOrNegativeOneBooleanContent);
121 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
123 // Jump is Expensive. Don't create extra control flow for 'and', 'or'
124 // condition branches.
125 setJumpIsExpensive(true);
127 // Wide divides are _very_ slow. Try to reduce the width of the divide if
129 addBypassSlowDiv(64, 32);
131 // By default, use the Source scheduling
133 setSchedulingPreference(Sched::RegPressure);
135 setSchedulingPreference(Sched::Source);
137 addRegisterClass(MVT::i1, &NVPTX::Int1RegsRegClass);
138 addRegisterClass(MVT::i16, &NVPTX::Int16RegsRegClass);
139 addRegisterClass(MVT::i32, &NVPTX::Int32RegsRegClass);
140 addRegisterClass(MVT::i64, &NVPTX::Int64RegsRegClass);
141 addRegisterClass(MVT::f32, &NVPTX::Float32RegsRegClass);
142 addRegisterClass(MVT::f64, &NVPTX::Float64RegsRegClass);
144 // Operations not directly supported by NVPTX.
145 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
146 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
147 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
148 setOperationAction(ISD::SELECT_CC, MVT::i8, Expand);
149 setOperationAction(ISD::SELECT_CC, MVT::i16, Expand);
150 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
151 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
152 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
153 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
154 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
155 setOperationAction(ISD::BR_CC, MVT::i8, Expand);
156 setOperationAction(ISD::BR_CC, MVT::i16, Expand);
157 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
158 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
159 // Some SIGN_EXTEND_INREG can be done using cvt instruction.
160 // For others we will expand to a SHL/SRA pair.
161 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i64, Legal);
162 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
163 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
164 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
165 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
167 setOperationAction(ISD::SHL_PARTS, MVT::i32 , Custom);
168 setOperationAction(ISD::SRA_PARTS, MVT::i32 , Custom);
169 setOperationAction(ISD::SRL_PARTS, MVT::i32 , Custom);
170 setOperationAction(ISD::SHL_PARTS, MVT::i64 , Custom);
171 setOperationAction(ISD::SRA_PARTS, MVT::i64 , Custom);
172 setOperationAction(ISD::SRL_PARTS, MVT::i64 , Custom);
174 if (STI.hasROT64()) {
175 setOperationAction(ISD::ROTL, MVT::i64, Legal);
176 setOperationAction(ISD::ROTR, MVT::i64, Legal);
178 setOperationAction(ISD::ROTL, MVT::i64, Expand);
179 setOperationAction(ISD::ROTR, MVT::i64, Expand);
181 if (STI.hasROT32()) {
182 setOperationAction(ISD::ROTL, MVT::i32, Legal);
183 setOperationAction(ISD::ROTR, MVT::i32, Legal);
185 setOperationAction(ISD::ROTL, MVT::i32, Expand);
186 setOperationAction(ISD::ROTR, MVT::i32, Expand);
189 setOperationAction(ISD::ROTL, MVT::i16, Expand);
190 setOperationAction(ISD::ROTR, MVT::i16, Expand);
191 setOperationAction(ISD::ROTL, MVT::i8, Expand);
192 setOperationAction(ISD::ROTR, MVT::i8, Expand);
193 setOperationAction(ISD::BSWAP, MVT::i16, Expand);
194 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
195 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
197 // Indirect branch is not supported.
198 // This also disables Jump Table creation.
199 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
200 setOperationAction(ISD::BRIND, MVT::Other, Expand);
202 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
203 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
205 // We want to legalize constant related memmove and memcopy
207 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
209 // Turn FP extload into load/fextend
210 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
211 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
212 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
213 setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, MVT::v2f16, Expand);
214 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Expand);
215 setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f32, Expand);
216 setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Expand);
217 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Expand);
218 setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f32, Expand);
219 // Turn FP truncstore into trunc + store.
220 // FIXME: vector types should also be expanded
221 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
222 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
223 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
225 // PTX does not support load / store predicate registers
226 setOperationAction(ISD::LOAD, MVT::i1, Custom);
227 setOperationAction(ISD::STORE, MVT::i1, Custom);
229 for (MVT VT : MVT::integer_valuetypes()) {
230 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
231 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
232 setTruncStoreAction(VT, MVT::i1, Expand);
235 // This is legal in NVPTX
236 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
237 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
239 // TRAP can be lowered to PTX trap
240 setOperationAction(ISD::TRAP, MVT::Other, Legal);
242 setOperationAction(ISD::ADDC, MVT::i64, Expand);
243 setOperationAction(ISD::ADDE, MVT::i64, Expand);
245 // Register custom handling for vector loads/stores
246 for (MVT VT : MVT::vector_valuetypes()) {
247 if (IsPTXVectorType(VT)) {
248 setOperationAction(ISD::LOAD, VT, Custom);
249 setOperationAction(ISD::STORE, VT, Custom);
250 setOperationAction(ISD::INTRINSIC_W_CHAIN, VT, Custom);
254 // Custom handling for i8 intrinsics
255 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);
257 setOperationAction(ISD::CTLZ, MVT::i16, Legal);
258 setOperationAction(ISD::CTLZ, MVT::i32, Legal);
259 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
260 setOperationAction(ISD::CTTZ, MVT::i16, Expand);
261 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
262 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
263 setOperationAction(ISD::CTPOP, MVT::i16, Legal);
264 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
265 setOperationAction(ISD::CTPOP, MVT::i64, Legal);
267 // PTX does not directly support SELP of i1, so promote to i32 first
268 setOperationAction(ISD::SELECT, MVT::i1, Custom);
270 // PTX cannot multiply two i64s in a single instruction.
271 setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);
272 setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
274 // We have some custom DAG combine patterns for these nodes
275 setTargetDAGCombine(ISD::ADD);
276 setTargetDAGCombine(ISD::AND);
277 setTargetDAGCombine(ISD::FADD);
278 setTargetDAGCombine(ISD::MUL);
279 setTargetDAGCombine(ISD::SHL);
280 setTargetDAGCombine(ISD::SELECT);
282 // Now deduce the information based on the above mentioned
284 computeRegisterProperties(STI.getRegisterInfo());
287 const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
288 switch ((NVPTXISD::NodeType)Opcode) {
289 case NVPTXISD::FIRST_NUMBER:
292 return "NVPTXISD::CALL";
293 case NVPTXISD::RET_FLAG:
294 return "NVPTXISD::RET_FLAG";
295 case NVPTXISD::LOAD_PARAM:
296 return "NVPTXISD::LOAD_PARAM";
297 case NVPTXISD::Wrapper:
298 return "NVPTXISD::Wrapper";
299 case NVPTXISD::DeclareParam:
300 return "NVPTXISD::DeclareParam";
301 case NVPTXISD::DeclareScalarParam:
302 return "NVPTXISD::DeclareScalarParam";
303 case NVPTXISD::DeclareRet:
304 return "NVPTXISD::DeclareRet";
305 case NVPTXISD::DeclareScalarRet:
306 return "NVPTXISD::DeclareScalarRet";
307 case NVPTXISD::DeclareRetParam:
308 return "NVPTXISD::DeclareRetParam";
309 case NVPTXISD::PrintCall:
310 return "NVPTXISD::PrintCall";
311 case NVPTXISD::PrintConvergentCall:
312 return "NVPTXISD::PrintConvergentCall";
313 case NVPTXISD::PrintCallUni:
314 return "NVPTXISD::PrintCallUni";
315 case NVPTXISD::PrintConvergentCallUni:
316 return "NVPTXISD::PrintConvergentCallUni";
317 case NVPTXISD::LoadParam:
318 return "NVPTXISD::LoadParam";
319 case NVPTXISD::LoadParamV2:
320 return "NVPTXISD::LoadParamV2";
321 case NVPTXISD::LoadParamV4:
322 return "NVPTXISD::LoadParamV4";
323 case NVPTXISD::StoreParam:
324 return "NVPTXISD::StoreParam";
325 case NVPTXISD::StoreParamV2:
326 return "NVPTXISD::StoreParamV2";
327 case NVPTXISD::StoreParamV4:
328 return "NVPTXISD::StoreParamV4";
329 case NVPTXISD::StoreParamS32:
330 return "NVPTXISD::StoreParamS32";
331 case NVPTXISD::StoreParamU32:
332 return "NVPTXISD::StoreParamU32";
333 case NVPTXISD::CallArgBegin:
334 return "NVPTXISD::CallArgBegin";
335 case NVPTXISD::CallArg:
336 return "NVPTXISD::CallArg";
337 case NVPTXISD::LastCallArg:
338 return "NVPTXISD::LastCallArg";
339 case NVPTXISD::CallArgEnd:
340 return "NVPTXISD::CallArgEnd";
341 case NVPTXISD::CallVoid:
342 return "NVPTXISD::CallVoid";
343 case NVPTXISD::CallVal:
344 return "NVPTXISD::CallVal";
345 case NVPTXISD::CallSymbol:
346 return "NVPTXISD::CallSymbol";
347 case NVPTXISD::Prototype:
348 return "NVPTXISD::Prototype";
349 case NVPTXISD::MoveParam:
350 return "NVPTXISD::MoveParam";
351 case NVPTXISD::StoreRetval:
352 return "NVPTXISD::StoreRetval";
353 case NVPTXISD::StoreRetvalV2:
354 return "NVPTXISD::StoreRetvalV2";
355 case NVPTXISD::StoreRetvalV4:
356 return "NVPTXISD::StoreRetvalV4";
357 case NVPTXISD::PseudoUseParam:
358 return "NVPTXISD::PseudoUseParam";
359 case NVPTXISD::RETURN:
360 return "NVPTXISD::RETURN";
361 case NVPTXISD::CallSeqBegin:
362 return "NVPTXISD::CallSeqBegin";
363 case NVPTXISD::CallSeqEnd:
364 return "NVPTXISD::CallSeqEnd";
365 case NVPTXISD::CallPrototype:
366 return "NVPTXISD::CallPrototype";
367 case NVPTXISD::LoadV2:
368 return "NVPTXISD::LoadV2";
369 case NVPTXISD::LoadV4:
370 return "NVPTXISD::LoadV4";
371 case NVPTXISD::LDGV2:
372 return "NVPTXISD::LDGV2";
373 case NVPTXISD::LDGV4:
374 return "NVPTXISD::LDGV4";
375 case NVPTXISD::LDUV2:
376 return "NVPTXISD::LDUV2";
377 case NVPTXISD::LDUV4:
378 return "NVPTXISD::LDUV4";
379 case NVPTXISD::StoreV2:
380 return "NVPTXISD::StoreV2";
381 case NVPTXISD::StoreV4:
382 return "NVPTXISD::StoreV4";
383 case NVPTXISD::FUN_SHFL_CLAMP:
384 return "NVPTXISD::FUN_SHFL_CLAMP";
385 case NVPTXISD::FUN_SHFR_CLAMP:
386 return "NVPTXISD::FUN_SHFR_CLAMP";
388 return "NVPTXISD::IMAD";
389 case NVPTXISD::Dummy:
390 return "NVPTXISD::Dummy";
391 case NVPTXISD::MUL_WIDE_SIGNED:
392 return "NVPTXISD::MUL_WIDE_SIGNED";
393 case NVPTXISD::MUL_WIDE_UNSIGNED:
394 return "NVPTXISD::MUL_WIDE_UNSIGNED";
395 case NVPTXISD::Tex1DFloatS32: return "NVPTXISD::Tex1DFloatS32";
396 case NVPTXISD::Tex1DFloatFloat: return "NVPTXISD::Tex1DFloatFloat";
397 case NVPTXISD::Tex1DFloatFloatLevel:
398 return "NVPTXISD::Tex1DFloatFloatLevel";
399 case NVPTXISD::Tex1DFloatFloatGrad:
400 return "NVPTXISD::Tex1DFloatFloatGrad";
401 case NVPTXISD::Tex1DS32S32: return "NVPTXISD::Tex1DS32S32";
402 case NVPTXISD::Tex1DS32Float: return "NVPTXISD::Tex1DS32Float";
403 case NVPTXISD::Tex1DS32FloatLevel:
404 return "NVPTXISD::Tex1DS32FloatLevel";
405 case NVPTXISD::Tex1DS32FloatGrad:
406 return "NVPTXISD::Tex1DS32FloatGrad";
407 case NVPTXISD::Tex1DU32S32: return "NVPTXISD::Tex1DU32S32";
408 case NVPTXISD::Tex1DU32Float: return "NVPTXISD::Tex1DU32Float";
409 case NVPTXISD::Tex1DU32FloatLevel:
410 return "NVPTXISD::Tex1DU32FloatLevel";
411 case NVPTXISD::Tex1DU32FloatGrad:
412 return "NVPTXISD::Tex1DU32FloatGrad";
413 case NVPTXISD::Tex1DArrayFloatS32: return "NVPTXISD::Tex1DArrayFloatS32";
414 case NVPTXISD::Tex1DArrayFloatFloat: return "NVPTXISD::Tex1DArrayFloatFloat";
415 case NVPTXISD::Tex1DArrayFloatFloatLevel:
416 return "NVPTXISD::Tex1DArrayFloatFloatLevel";
417 case NVPTXISD::Tex1DArrayFloatFloatGrad:
418 return "NVPTXISD::Tex1DArrayFloatFloatGrad";
419 case NVPTXISD::Tex1DArrayS32S32: return "NVPTXISD::Tex1DArrayS32S32";
420 case NVPTXISD::Tex1DArrayS32Float: return "NVPTXISD::Tex1DArrayS32Float";
421 case NVPTXISD::Tex1DArrayS32FloatLevel:
422 return "NVPTXISD::Tex1DArrayS32FloatLevel";
423 case NVPTXISD::Tex1DArrayS32FloatGrad:
424 return "NVPTXISD::Tex1DArrayS32FloatGrad";
425 case NVPTXISD::Tex1DArrayU32S32: return "NVPTXISD::Tex1DArrayU32S32";
426 case NVPTXISD::Tex1DArrayU32Float: return "NVPTXISD::Tex1DArrayU32Float";
427 case NVPTXISD::Tex1DArrayU32FloatLevel:
428 return "NVPTXISD::Tex1DArrayU32FloatLevel";
429 case NVPTXISD::Tex1DArrayU32FloatGrad:
430 return "NVPTXISD::Tex1DArrayU32FloatGrad";
431 case NVPTXISD::Tex2DFloatS32: return "NVPTXISD::Tex2DFloatS32";
432 case NVPTXISD::Tex2DFloatFloat: return "NVPTXISD::Tex2DFloatFloat";
433 case NVPTXISD::Tex2DFloatFloatLevel:
434 return "NVPTXISD::Tex2DFloatFloatLevel";
435 case NVPTXISD::Tex2DFloatFloatGrad:
436 return "NVPTXISD::Tex2DFloatFloatGrad";
437 case NVPTXISD::Tex2DS32S32: return "NVPTXISD::Tex2DS32S32";
438 case NVPTXISD::Tex2DS32Float: return "NVPTXISD::Tex2DS32Float";
439 case NVPTXISD::Tex2DS32FloatLevel:
440 return "NVPTXISD::Tex2DS32FloatLevel";
441 case NVPTXISD::Tex2DS32FloatGrad:
442 return "NVPTXISD::Tex2DS32FloatGrad";
443 case NVPTXISD::Tex2DU32S32: return "NVPTXISD::Tex2DU32S32";
444 case NVPTXISD::Tex2DU32Float: return "NVPTXISD::Tex2DU32Float";
445 case NVPTXISD::Tex2DU32FloatLevel:
446 return "NVPTXISD::Tex2DU32FloatLevel";
447 case NVPTXISD::Tex2DU32FloatGrad:
448 return "NVPTXISD::Tex2DU32FloatGrad";
449 case NVPTXISD::Tex2DArrayFloatS32: return "NVPTXISD::Tex2DArrayFloatS32";
450 case NVPTXISD::Tex2DArrayFloatFloat: return "NVPTXISD::Tex2DArrayFloatFloat";
451 case NVPTXISD::Tex2DArrayFloatFloatLevel:
452 return "NVPTXISD::Tex2DArrayFloatFloatLevel";
453 case NVPTXISD::Tex2DArrayFloatFloatGrad:
454 return "NVPTXISD::Tex2DArrayFloatFloatGrad";
455 case NVPTXISD::Tex2DArrayS32S32: return "NVPTXISD::Tex2DArrayS32S32";
456 case NVPTXISD::Tex2DArrayS32Float: return "NVPTXISD::Tex2DArrayS32Float";
457 case NVPTXISD::Tex2DArrayS32FloatLevel:
458 return "NVPTXISD::Tex2DArrayS32FloatLevel";
459 case NVPTXISD::Tex2DArrayS32FloatGrad:
460 return "NVPTXISD::Tex2DArrayS32FloatGrad";
461 case NVPTXISD::Tex2DArrayU32S32: return "NVPTXISD::Tex2DArrayU32S32";
462 case NVPTXISD::Tex2DArrayU32Float: return "NVPTXISD::Tex2DArrayU32Float";
463 case NVPTXISD::Tex2DArrayU32FloatLevel:
464 return "NVPTXISD::Tex2DArrayU32FloatLevel";
465 case NVPTXISD::Tex2DArrayU32FloatGrad:
466 return "NVPTXISD::Tex2DArrayU32FloatGrad";
467 case NVPTXISD::Tex3DFloatS32: return "NVPTXISD::Tex3DFloatS32";
468 case NVPTXISD::Tex3DFloatFloat: return "NVPTXISD::Tex3DFloatFloat";
469 case NVPTXISD::Tex3DFloatFloatLevel:
470 return "NVPTXISD::Tex3DFloatFloatLevel";
471 case NVPTXISD::Tex3DFloatFloatGrad:
472 return "NVPTXISD::Tex3DFloatFloatGrad";
473 case NVPTXISD::Tex3DS32S32: return "NVPTXISD::Tex3DS32S32";
474 case NVPTXISD::Tex3DS32Float: return "NVPTXISD::Tex3DS32Float";
475 case NVPTXISD::Tex3DS32FloatLevel:
476 return "NVPTXISD::Tex3DS32FloatLevel";
477 case NVPTXISD::Tex3DS32FloatGrad:
478 return "NVPTXISD::Tex3DS32FloatGrad";
479 case NVPTXISD::Tex3DU32S32: return "NVPTXISD::Tex3DU32S32";
480 case NVPTXISD::Tex3DU32Float: return "NVPTXISD::Tex3DU32Float";
481 case NVPTXISD::Tex3DU32FloatLevel:
482 return "NVPTXISD::Tex3DU32FloatLevel";
483 case NVPTXISD::Tex3DU32FloatGrad:
484 return "NVPTXISD::Tex3DU32FloatGrad";
485 case NVPTXISD::TexCubeFloatFloat: return "NVPTXISD::TexCubeFloatFloat";
486 case NVPTXISD::TexCubeFloatFloatLevel:
487 return "NVPTXISD::TexCubeFloatFloatLevel";
488 case NVPTXISD::TexCubeS32Float: return "NVPTXISD::TexCubeS32Float";
489 case NVPTXISD::TexCubeS32FloatLevel:
490 return "NVPTXISD::TexCubeS32FloatLevel";
491 case NVPTXISD::TexCubeU32Float: return "NVPTXISD::TexCubeU32Float";
492 case NVPTXISD::TexCubeU32FloatLevel:
493 return "NVPTXISD::TexCubeU32FloatLevel";
494 case NVPTXISD::TexCubeArrayFloatFloat:
495 return "NVPTXISD::TexCubeArrayFloatFloat";
496 case NVPTXISD::TexCubeArrayFloatFloatLevel:
497 return "NVPTXISD::TexCubeArrayFloatFloatLevel";
498 case NVPTXISD::TexCubeArrayS32Float:
499 return "NVPTXISD::TexCubeArrayS32Float";
500 case NVPTXISD::TexCubeArrayS32FloatLevel:
501 return "NVPTXISD::TexCubeArrayS32FloatLevel";
502 case NVPTXISD::TexCubeArrayU32Float:
503 return "NVPTXISD::TexCubeArrayU32Float";
504 case NVPTXISD::TexCubeArrayU32FloatLevel:
505 return "NVPTXISD::TexCubeArrayU32FloatLevel";
506 case NVPTXISD::Tld4R2DFloatFloat:
507 return "NVPTXISD::Tld4R2DFloatFloat";
508 case NVPTXISD::Tld4G2DFloatFloat:
509 return "NVPTXISD::Tld4G2DFloatFloat";
510 case NVPTXISD::Tld4B2DFloatFloat:
511 return "NVPTXISD::Tld4B2DFloatFloat";
512 case NVPTXISD::Tld4A2DFloatFloat:
513 return "NVPTXISD::Tld4A2DFloatFloat";
514 case NVPTXISD::Tld4R2DS64Float:
515 return "NVPTXISD::Tld4R2DS64Float";
516 case NVPTXISD::Tld4G2DS64Float:
517 return "NVPTXISD::Tld4G2DS64Float";
518 case NVPTXISD::Tld4B2DS64Float:
519 return "NVPTXISD::Tld4B2DS64Float";
520 case NVPTXISD::Tld4A2DS64Float:
521 return "NVPTXISD::Tld4A2DS64Float";
522 case NVPTXISD::Tld4R2DU64Float:
523 return "NVPTXISD::Tld4R2DU64Float";
524 case NVPTXISD::Tld4G2DU64Float:
525 return "NVPTXISD::Tld4G2DU64Float";
526 case NVPTXISD::Tld4B2DU64Float:
527 return "NVPTXISD::Tld4B2DU64Float";
528 case NVPTXISD::Tld4A2DU64Float:
529 return "NVPTXISD::Tld4A2DU64Float";
531 case NVPTXISD::TexUnified1DFloatS32:
532 return "NVPTXISD::TexUnified1DFloatS32";
533 case NVPTXISD::TexUnified1DFloatFloat:
534 return "NVPTXISD::TexUnified1DFloatFloat";
535 case NVPTXISD::TexUnified1DFloatFloatLevel:
536 return "NVPTXISD::TexUnified1DFloatFloatLevel";
537 case NVPTXISD::TexUnified1DFloatFloatGrad:
538 return "NVPTXISD::TexUnified1DFloatFloatGrad";
539 case NVPTXISD::TexUnified1DS32S32:
540 return "NVPTXISD::TexUnified1DS32S32";
541 case NVPTXISD::TexUnified1DS32Float:
542 return "NVPTXISD::TexUnified1DS32Float";
543 case NVPTXISD::TexUnified1DS32FloatLevel:
544 return "NVPTXISD::TexUnified1DS32FloatLevel";
545 case NVPTXISD::TexUnified1DS32FloatGrad:
546 return "NVPTXISD::TexUnified1DS32FloatGrad";
547 case NVPTXISD::TexUnified1DU32S32:
548 return "NVPTXISD::TexUnified1DU32S32";
549 case NVPTXISD::TexUnified1DU32Float:
550 return "NVPTXISD::TexUnified1DU32Float";
551 case NVPTXISD::TexUnified1DU32FloatLevel:
552 return "NVPTXISD::TexUnified1DU32FloatLevel";
553 case NVPTXISD::TexUnified1DU32FloatGrad:
554 return "NVPTXISD::TexUnified1DU32FloatGrad";
555 case NVPTXISD::TexUnified1DArrayFloatS32:
556 return "NVPTXISD::TexUnified1DArrayFloatS32";
557 case NVPTXISD::TexUnified1DArrayFloatFloat:
558 return "NVPTXISD::TexUnified1DArrayFloatFloat";
559 case NVPTXISD::TexUnified1DArrayFloatFloatLevel:
560 return "NVPTXISD::TexUnified1DArrayFloatFloatLevel";
561 case NVPTXISD::TexUnified1DArrayFloatFloatGrad:
562 return "NVPTXISD::TexUnified1DArrayFloatFloatGrad";
563 case NVPTXISD::TexUnified1DArrayS32S32:
564 return "NVPTXISD::TexUnified1DArrayS32S32";
565 case NVPTXISD::TexUnified1DArrayS32Float:
566 return "NVPTXISD::TexUnified1DArrayS32Float";
567 case NVPTXISD::TexUnified1DArrayS32FloatLevel:
568 return "NVPTXISD::TexUnified1DArrayS32FloatLevel";
569 case NVPTXISD::TexUnified1DArrayS32FloatGrad:
570 return "NVPTXISD::TexUnified1DArrayS32FloatGrad";
571 case NVPTXISD::TexUnified1DArrayU32S32:
572 return "NVPTXISD::TexUnified1DArrayU32S32";
573 case NVPTXISD::TexUnified1DArrayU32Float:
574 return "NVPTXISD::TexUnified1DArrayU32Float";
575 case NVPTXISD::TexUnified1DArrayU32FloatLevel:
576 return "NVPTXISD::TexUnified1DArrayU32FloatLevel";
577 case NVPTXISD::TexUnified1DArrayU32FloatGrad:
578 return "NVPTXISD::TexUnified1DArrayU32FloatGrad";
579 case NVPTXISD::TexUnified2DFloatS32:
580 return "NVPTXISD::TexUnified2DFloatS32";
581 case NVPTXISD::TexUnified2DFloatFloat:
582 return "NVPTXISD::TexUnified2DFloatFloat";
583 case NVPTXISD::TexUnified2DFloatFloatLevel:
584 return "NVPTXISD::TexUnified2DFloatFloatLevel";
585 case NVPTXISD::TexUnified2DFloatFloatGrad:
586 return "NVPTXISD::TexUnified2DFloatFloatGrad";
587 case NVPTXISD::TexUnified2DS32S32:
588 return "NVPTXISD::TexUnified2DS32S32";
589 case NVPTXISD::TexUnified2DS32Float:
590 return "NVPTXISD::TexUnified2DS32Float";
591 case NVPTXISD::TexUnified2DS32FloatLevel:
592 return "NVPTXISD::TexUnified2DS32FloatLevel";
593 case NVPTXISD::TexUnified2DS32FloatGrad:
594 return "NVPTXISD::TexUnified2DS32FloatGrad";
595 case NVPTXISD::TexUnified2DU32S32:
596 return "NVPTXISD::TexUnified2DU32S32";
597 case NVPTXISD::TexUnified2DU32Float:
598 return "NVPTXISD::TexUnified2DU32Float";
599 case NVPTXISD::TexUnified2DU32FloatLevel:
600 return "NVPTXISD::TexUnified2DU32FloatLevel";
601 case NVPTXISD::TexUnified2DU32FloatGrad:
602 return "NVPTXISD::TexUnified2DU32FloatGrad";
603 case NVPTXISD::TexUnified2DArrayFloatS32:
604 return "NVPTXISD::TexUnified2DArrayFloatS32";
605 case NVPTXISD::TexUnified2DArrayFloatFloat:
606 return "NVPTXISD::TexUnified2DArrayFloatFloat";
607 case NVPTXISD::TexUnified2DArrayFloatFloatLevel:
608 return "NVPTXISD::TexUnified2DArrayFloatFloatLevel";
609 case NVPTXISD::TexUnified2DArrayFloatFloatGrad:
610 return "NVPTXISD::TexUnified2DArrayFloatFloatGrad";
611 case NVPTXISD::TexUnified2DArrayS32S32:
612 return "NVPTXISD::TexUnified2DArrayS32S32";
613 case NVPTXISD::TexUnified2DArrayS32Float:
614 return "NVPTXISD::TexUnified2DArrayS32Float";
615 case NVPTXISD::TexUnified2DArrayS32FloatLevel:
616 return "NVPTXISD::TexUnified2DArrayS32FloatLevel";
617 case NVPTXISD::TexUnified2DArrayS32FloatGrad:
618 return "NVPTXISD::TexUnified2DArrayS32FloatGrad";
619 case NVPTXISD::TexUnified2DArrayU32S32:
620 return "NVPTXISD::TexUnified2DArrayU32S32";
621 case NVPTXISD::TexUnified2DArrayU32Float:
622 return "NVPTXISD::TexUnified2DArrayU32Float";
623 case NVPTXISD::TexUnified2DArrayU32FloatLevel:
624 return "NVPTXISD::TexUnified2DArrayU32FloatLevel";
625 case NVPTXISD::TexUnified2DArrayU32FloatGrad:
626 return "NVPTXISD::TexUnified2DArrayU32FloatGrad";
627 case NVPTXISD::TexUnified3DFloatS32:
628 return "NVPTXISD::TexUnified3DFloatS32";
629 case NVPTXISD::TexUnified3DFloatFloat:
630 return "NVPTXISD::TexUnified3DFloatFloat";
631 case NVPTXISD::TexUnified3DFloatFloatLevel:
632 return "NVPTXISD::TexUnified3DFloatFloatLevel";
633 case NVPTXISD::TexUnified3DFloatFloatGrad:
634 return "NVPTXISD::TexUnified3DFloatFloatGrad";
635 case NVPTXISD::TexUnified3DS32S32:
636 return "NVPTXISD::TexUnified3DS32S32";
637 case NVPTXISD::TexUnified3DS32Float:
638 return "NVPTXISD::TexUnified3DS32Float";
639 case NVPTXISD::TexUnified3DS32FloatLevel:
640 return "NVPTXISD::TexUnified3DS32FloatLevel";
641 case NVPTXISD::TexUnified3DS32FloatGrad:
642 return "NVPTXISD::TexUnified3DS32FloatGrad";
643 case NVPTXISD::TexUnified3DU32S32:
644 return "NVPTXISD::TexUnified3DU32S32";
645 case NVPTXISD::TexUnified3DU32Float:
646 return "NVPTXISD::TexUnified3DU32Float";
647 case NVPTXISD::TexUnified3DU32FloatLevel:
648 return "NVPTXISD::TexUnified3DU32FloatLevel";
649 case NVPTXISD::TexUnified3DU32FloatGrad:
650 return "NVPTXISD::TexUnified3DU32FloatGrad";
651 case NVPTXISD::TexUnifiedCubeFloatFloat:
652 return "NVPTXISD::TexUnifiedCubeFloatFloat";
653 case NVPTXISD::TexUnifiedCubeFloatFloatLevel:
654 return "NVPTXISD::TexUnifiedCubeFloatFloatLevel";
655 case NVPTXISD::TexUnifiedCubeS32Float:
656 return "NVPTXISD::TexUnifiedCubeS32Float";
657 case NVPTXISD::TexUnifiedCubeS32FloatLevel:
658 return "NVPTXISD::TexUnifiedCubeS32FloatLevel";
659 case NVPTXISD::TexUnifiedCubeU32Float:
660 return "NVPTXISD::TexUnifiedCubeU32Float";
661 case NVPTXISD::TexUnifiedCubeU32FloatLevel:
662 return "NVPTXISD::TexUnifiedCubeU32FloatLevel";
663 case NVPTXISD::TexUnifiedCubeArrayFloatFloat:
664 return "NVPTXISD::TexUnifiedCubeArrayFloatFloat";
665 case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel:
666 return "NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel";
667 case NVPTXISD::TexUnifiedCubeArrayS32Float:
668 return "NVPTXISD::TexUnifiedCubeArrayS32Float";
669 case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel:
670 return "NVPTXISD::TexUnifiedCubeArrayS32FloatLevel";
671 case NVPTXISD::TexUnifiedCubeArrayU32Float:
672 return "NVPTXISD::TexUnifiedCubeArrayU32Float";
673 case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel:
674 return "NVPTXISD::TexUnifiedCubeArrayU32FloatLevel";
675 case NVPTXISD::Tld4UnifiedR2DFloatFloat:
676 return "NVPTXISD::Tld4UnifiedR2DFloatFloat";
677 case NVPTXISD::Tld4UnifiedG2DFloatFloat:
678 return "NVPTXISD::Tld4UnifiedG2DFloatFloat";
679 case NVPTXISD::Tld4UnifiedB2DFloatFloat:
680 return "NVPTXISD::Tld4UnifiedB2DFloatFloat";
681 case NVPTXISD::Tld4UnifiedA2DFloatFloat:
682 return "NVPTXISD::Tld4UnifiedA2DFloatFloat";
683 case NVPTXISD::Tld4UnifiedR2DS64Float:
684 return "NVPTXISD::Tld4UnifiedR2DS64Float";
685 case NVPTXISD::Tld4UnifiedG2DS64Float:
686 return "NVPTXISD::Tld4UnifiedG2DS64Float";
687 case NVPTXISD::Tld4UnifiedB2DS64Float:
688 return "NVPTXISD::Tld4UnifiedB2DS64Float";
689 case NVPTXISD::Tld4UnifiedA2DS64Float:
690 return "NVPTXISD::Tld4UnifiedA2DS64Float";
691 case NVPTXISD::Tld4UnifiedR2DU64Float:
692 return "NVPTXISD::Tld4UnifiedR2DU64Float";
693 case NVPTXISD::Tld4UnifiedG2DU64Float:
694 return "NVPTXISD::Tld4UnifiedG2DU64Float";
695 case NVPTXISD::Tld4UnifiedB2DU64Float:
696 return "NVPTXISD::Tld4UnifiedB2DU64Float";
697 case NVPTXISD::Tld4UnifiedA2DU64Float:
698 return "NVPTXISD::Tld4UnifiedA2DU64Float";
700 case NVPTXISD::Suld1DI8Clamp: return "NVPTXISD::Suld1DI8Clamp";
701 case NVPTXISD::Suld1DI16Clamp: return "NVPTXISD::Suld1DI16Clamp";
702 case NVPTXISD::Suld1DI32Clamp: return "NVPTXISD::Suld1DI32Clamp";
703 case NVPTXISD::Suld1DI64Clamp: return "NVPTXISD::Suld1DI64Clamp";
704 case NVPTXISD::Suld1DV2I8Clamp: return "NVPTXISD::Suld1DV2I8Clamp";
705 case NVPTXISD::Suld1DV2I16Clamp: return "NVPTXISD::Suld1DV2I16Clamp";
706 case NVPTXISD::Suld1DV2I32Clamp: return "NVPTXISD::Suld1DV2I32Clamp";
707 case NVPTXISD::Suld1DV2I64Clamp: return "NVPTXISD::Suld1DV2I64Clamp";
708 case NVPTXISD::Suld1DV4I8Clamp: return "NVPTXISD::Suld1DV4I8Clamp";
709 case NVPTXISD::Suld1DV4I16Clamp: return "NVPTXISD::Suld1DV4I16Clamp";
710 case NVPTXISD::Suld1DV4I32Clamp: return "NVPTXISD::Suld1DV4I32Clamp";
712 case NVPTXISD::Suld1DArrayI8Clamp: return "NVPTXISD::Suld1DArrayI8Clamp";
713 case NVPTXISD::Suld1DArrayI16Clamp: return "NVPTXISD::Suld1DArrayI16Clamp";
714 case NVPTXISD::Suld1DArrayI32Clamp: return "NVPTXISD::Suld1DArrayI32Clamp";
715 case NVPTXISD::Suld1DArrayI64Clamp: return "NVPTXISD::Suld1DArrayI64Clamp";
716 case NVPTXISD::Suld1DArrayV2I8Clamp: return "NVPTXISD::Suld1DArrayV2I8Clamp";
717 case NVPTXISD::Suld1DArrayV2I16Clamp:return "NVPTXISD::Suld1DArrayV2I16Clamp";
718 case NVPTXISD::Suld1DArrayV2I32Clamp:return "NVPTXISD::Suld1DArrayV2I32Clamp";
719 case NVPTXISD::Suld1DArrayV2I64Clamp:return "NVPTXISD::Suld1DArrayV2I64Clamp";
720 case NVPTXISD::Suld1DArrayV4I8Clamp: return "NVPTXISD::Suld1DArrayV4I8Clamp";
721 case NVPTXISD::Suld1DArrayV4I16Clamp:return "NVPTXISD::Suld1DArrayV4I16Clamp";
722 case NVPTXISD::Suld1DArrayV4I32Clamp:return "NVPTXISD::Suld1DArrayV4I32Clamp";
724 case NVPTXISD::Suld2DI8Clamp: return "NVPTXISD::Suld2DI8Clamp";
725 case NVPTXISD::Suld2DI16Clamp: return "NVPTXISD::Suld2DI16Clamp";
726 case NVPTXISD::Suld2DI32Clamp: return "NVPTXISD::Suld2DI32Clamp";
727 case NVPTXISD::Suld2DI64Clamp: return "NVPTXISD::Suld2DI64Clamp";
728 case NVPTXISD::Suld2DV2I8Clamp: return "NVPTXISD::Suld2DV2I8Clamp";
729 case NVPTXISD::Suld2DV2I16Clamp: return "NVPTXISD::Suld2DV2I16Clamp";
730 case NVPTXISD::Suld2DV2I32Clamp: return "NVPTXISD::Suld2DV2I32Clamp";
731 case NVPTXISD::Suld2DV2I64Clamp: return "NVPTXISD::Suld2DV2I64Clamp";
732 case NVPTXISD::Suld2DV4I8Clamp: return "NVPTXISD::Suld2DV4I8Clamp";
733 case NVPTXISD::Suld2DV4I16Clamp: return "NVPTXISD::Suld2DV4I16Clamp";
734 case NVPTXISD::Suld2DV4I32Clamp: return "NVPTXISD::Suld2DV4I32Clamp";
736 case NVPTXISD::Suld2DArrayI8Clamp: return "NVPTXISD::Suld2DArrayI8Clamp";
737 case NVPTXISD::Suld2DArrayI16Clamp: return "NVPTXISD::Suld2DArrayI16Clamp";
738 case NVPTXISD::Suld2DArrayI32Clamp: return "NVPTXISD::Suld2DArrayI32Clamp";
739 case NVPTXISD::Suld2DArrayI64Clamp: return "NVPTXISD::Suld2DArrayI64Clamp";
740 case NVPTXISD::Suld2DArrayV2I8Clamp: return "NVPTXISD::Suld2DArrayV2I8Clamp";
741 case NVPTXISD::Suld2DArrayV2I16Clamp:return "NVPTXISD::Suld2DArrayV2I16Clamp";
742 case NVPTXISD::Suld2DArrayV2I32Clamp:return "NVPTXISD::Suld2DArrayV2I32Clamp";
743 case NVPTXISD::Suld2DArrayV2I64Clamp:return "NVPTXISD::Suld2DArrayV2I64Clamp";
744 case NVPTXISD::Suld2DArrayV4I8Clamp: return "NVPTXISD::Suld2DArrayV4I8Clamp";
745 case NVPTXISD::Suld2DArrayV4I16Clamp:return "NVPTXISD::Suld2DArrayV4I16Clamp";
746 case NVPTXISD::Suld2DArrayV4I32Clamp:return "NVPTXISD::Suld2DArrayV4I32Clamp";
748 case NVPTXISD::Suld3DI8Clamp: return "NVPTXISD::Suld3DI8Clamp";
749 case NVPTXISD::Suld3DI16Clamp: return "NVPTXISD::Suld3DI16Clamp";
750 case NVPTXISD::Suld3DI32Clamp: return "NVPTXISD::Suld3DI32Clamp";
751 case NVPTXISD::Suld3DI64Clamp: return "NVPTXISD::Suld3DI64Clamp";
752 case NVPTXISD::Suld3DV2I8Clamp: return "NVPTXISD::Suld3DV2I8Clamp";
753 case NVPTXISD::Suld3DV2I16Clamp: return "NVPTXISD::Suld3DV2I16Clamp";
754 case NVPTXISD::Suld3DV2I32Clamp: return "NVPTXISD::Suld3DV2I32Clamp";
755 case NVPTXISD::Suld3DV2I64Clamp: return "NVPTXISD::Suld3DV2I64Clamp";
756 case NVPTXISD::Suld3DV4I8Clamp: return "NVPTXISD::Suld3DV4I8Clamp";
757 case NVPTXISD::Suld3DV4I16Clamp: return "NVPTXISD::Suld3DV4I16Clamp";
758 case NVPTXISD::Suld3DV4I32Clamp: return "NVPTXISD::Suld3DV4I32Clamp";
760 case NVPTXISD::Suld1DI8Trap: return "NVPTXISD::Suld1DI8Trap";
761 case NVPTXISD::Suld1DI16Trap: return "NVPTXISD::Suld1DI16Trap";
762 case NVPTXISD::Suld1DI32Trap: return "NVPTXISD::Suld1DI32Trap";
763 case NVPTXISD::Suld1DI64Trap: return "NVPTXISD::Suld1DI64Trap";
764 case NVPTXISD::Suld1DV2I8Trap: return "NVPTXISD::Suld1DV2I8Trap";
765 case NVPTXISD::Suld1DV2I16Trap: return "NVPTXISD::Suld1DV2I16Trap";
766 case NVPTXISD::Suld1DV2I32Trap: return "NVPTXISD::Suld1DV2I32Trap";
767 case NVPTXISD::Suld1DV2I64Trap: return "NVPTXISD::Suld1DV2I64Trap";
768 case NVPTXISD::Suld1DV4I8Trap: return "NVPTXISD::Suld1DV4I8Trap";
769 case NVPTXISD::Suld1DV4I16Trap: return "NVPTXISD::Suld1DV4I16Trap";
770 case NVPTXISD::Suld1DV4I32Trap: return "NVPTXISD::Suld1DV4I32Trap";
772 case NVPTXISD::Suld1DArrayI8Trap: return "NVPTXISD::Suld1DArrayI8Trap";
773 case NVPTXISD::Suld1DArrayI16Trap: return "NVPTXISD::Suld1DArrayI16Trap";
774 case NVPTXISD::Suld1DArrayI32Trap: return "NVPTXISD::Suld1DArrayI32Trap";
775 case NVPTXISD::Suld1DArrayI64Trap: return "NVPTXISD::Suld1DArrayI64Trap";
776 case NVPTXISD::Suld1DArrayV2I8Trap: return "NVPTXISD::Suld1DArrayV2I8Trap";
777 case NVPTXISD::Suld1DArrayV2I16Trap: return "NVPTXISD::Suld1DArrayV2I16Trap";
778 case NVPTXISD::Suld1DArrayV2I32Trap: return "NVPTXISD::Suld1DArrayV2I32Trap";
779 case NVPTXISD::Suld1DArrayV2I64Trap: return "NVPTXISD::Suld1DArrayV2I64Trap";
780 case NVPTXISD::Suld1DArrayV4I8Trap: return "NVPTXISD::Suld1DArrayV4I8Trap";
781 case NVPTXISD::Suld1DArrayV4I16Trap: return "NVPTXISD::Suld1DArrayV4I16Trap";
782 case NVPTXISD::Suld1DArrayV4I32Trap: return "NVPTXISD::Suld1DArrayV4I32Trap";
784 case NVPTXISD::Suld2DI8Trap: return "NVPTXISD::Suld2DI8Trap";
785 case NVPTXISD::Suld2DI16Trap: return "NVPTXISD::Suld2DI16Trap";
786 case NVPTXISD::Suld2DI32Trap: return "NVPTXISD::Suld2DI32Trap";
787 case NVPTXISD::Suld2DI64Trap: return "NVPTXISD::Suld2DI64Trap";
788 case NVPTXISD::Suld2DV2I8Trap: return "NVPTXISD::Suld2DV2I8Trap";
789 case NVPTXISD::Suld2DV2I16Trap: return "NVPTXISD::Suld2DV2I16Trap";
790 case NVPTXISD::Suld2DV2I32Trap: return "NVPTXISD::Suld2DV2I32Trap";
791 case NVPTXISD::Suld2DV2I64Trap: return "NVPTXISD::Suld2DV2I64Trap";
792 case NVPTXISD::Suld2DV4I8Trap: return "NVPTXISD::Suld2DV4I8Trap";
793 case NVPTXISD::Suld2DV4I16Trap: return "NVPTXISD::Suld2DV4I16Trap";
794 case NVPTXISD::Suld2DV4I32Trap: return "NVPTXISD::Suld2DV4I32Trap";
796 case NVPTXISD::Suld2DArrayI8Trap: return "NVPTXISD::Suld2DArrayI8Trap";
797 case NVPTXISD::Suld2DArrayI16Trap: return "NVPTXISD::Suld2DArrayI16Trap";
798 case NVPTXISD::Suld2DArrayI32Trap: return "NVPTXISD::Suld2DArrayI32Trap";
799 case NVPTXISD::Suld2DArrayI64Trap: return "NVPTXISD::Suld2DArrayI64Trap";
800 case NVPTXISD::Suld2DArrayV2I8Trap: return "NVPTXISD::Suld2DArrayV2I8Trap";
801 case NVPTXISD::Suld2DArrayV2I16Trap: return "NVPTXISD::Suld2DArrayV2I16Trap";
802 case NVPTXISD::Suld2DArrayV2I32Trap: return "NVPTXISD::Suld2DArrayV2I32Trap";
803 case NVPTXISD::Suld2DArrayV2I64Trap: return "NVPTXISD::Suld2DArrayV2I64Trap";
804 case NVPTXISD::Suld2DArrayV4I8Trap: return "NVPTXISD::Suld2DArrayV4I8Trap";
805 case NVPTXISD::Suld2DArrayV4I16Trap: return "NVPTXISD::Suld2DArrayV4I16Trap";
806 case NVPTXISD::Suld2DArrayV4I32Trap: return "NVPTXISD::Suld2DArrayV4I32Trap";
808 case NVPTXISD::Suld3DI8Trap: return "NVPTXISD::Suld3DI8Trap";
809 case NVPTXISD::Suld3DI16Trap: return "NVPTXISD::Suld3DI16Trap";
810 case NVPTXISD::Suld3DI32Trap: return "NVPTXISD::Suld3DI32Trap";
811 case NVPTXISD::Suld3DI64Trap: return "NVPTXISD::Suld3DI64Trap";
812 case NVPTXISD::Suld3DV2I8Trap: return "NVPTXISD::Suld3DV2I8Trap";
813 case NVPTXISD::Suld3DV2I16Trap: return "NVPTXISD::Suld3DV2I16Trap";
814 case NVPTXISD::Suld3DV2I32Trap: return "NVPTXISD::Suld3DV2I32Trap";
815 case NVPTXISD::Suld3DV2I64Trap: return "NVPTXISD::Suld3DV2I64Trap";
816 case NVPTXISD::Suld3DV4I8Trap: return "NVPTXISD::Suld3DV4I8Trap";
817 case NVPTXISD::Suld3DV4I16Trap: return "NVPTXISD::Suld3DV4I16Trap";
818 case NVPTXISD::Suld3DV4I32Trap: return "NVPTXISD::Suld3DV4I32Trap";
820 case NVPTXISD::Suld1DI8Zero: return "NVPTXISD::Suld1DI8Zero";
821 case NVPTXISD::Suld1DI16Zero: return "NVPTXISD::Suld1DI16Zero";
822 case NVPTXISD::Suld1DI32Zero: return "NVPTXISD::Suld1DI32Zero";
823 case NVPTXISD::Suld1DI64Zero: return "NVPTXISD::Suld1DI64Zero";
824 case NVPTXISD::Suld1DV2I8Zero: return "NVPTXISD::Suld1DV2I8Zero";
825 case NVPTXISD::Suld1DV2I16Zero: return "NVPTXISD::Suld1DV2I16Zero";
826 case NVPTXISD::Suld1DV2I32Zero: return "NVPTXISD::Suld1DV2I32Zero";
827 case NVPTXISD::Suld1DV2I64Zero: return "NVPTXISD::Suld1DV2I64Zero";
828 case NVPTXISD::Suld1DV4I8Zero: return "NVPTXISD::Suld1DV4I8Zero";
829 case NVPTXISD::Suld1DV4I16Zero: return "NVPTXISD::Suld1DV4I16Zero";
830 case NVPTXISD::Suld1DV4I32Zero: return "NVPTXISD::Suld1DV4I32Zero";
832 case NVPTXISD::Suld1DArrayI8Zero: return "NVPTXISD::Suld1DArrayI8Zero";
833 case NVPTXISD::Suld1DArrayI16Zero: return "NVPTXISD::Suld1DArrayI16Zero";
834 case NVPTXISD::Suld1DArrayI32Zero: return "NVPTXISD::Suld1DArrayI32Zero";
835 case NVPTXISD::Suld1DArrayI64Zero: return "NVPTXISD::Suld1DArrayI64Zero";
836 case NVPTXISD::Suld1DArrayV2I8Zero: return "NVPTXISD::Suld1DArrayV2I8Zero";
837 case NVPTXISD::Suld1DArrayV2I16Zero: return "NVPTXISD::Suld1DArrayV2I16Zero";
838 case NVPTXISD::Suld1DArrayV2I32Zero: return "NVPTXISD::Suld1DArrayV2I32Zero";
839 case NVPTXISD::Suld1DArrayV2I64Zero: return "NVPTXISD::Suld1DArrayV2I64Zero";
840 case NVPTXISD::Suld1DArrayV4I8Zero: return "NVPTXISD::Suld1DArrayV4I8Zero";
841 case NVPTXISD::Suld1DArrayV4I16Zero: return "NVPTXISD::Suld1DArrayV4I16Zero";
842 case NVPTXISD::Suld1DArrayV4I32Zero: return "NVPTXISD::Suld1DArrayV4I32Zero";
844 case NVPTXISD::Suld2DI8Zero: return "NVPTXISD::Suld2DI8Zero";
845 case NVPTXISD::Suld2DI16Zero: return "NVPTXISD::Suld2DI16Zero";
846 case NVPTXISD::Suld2DI32Zero: return "NVPTXISD::Suld2DI32Zero";
847 case NVPTXISD::Suld2DI64Zero: return "NVPTXISD::Suld2DI64Zero";
848 case NVPTXISD::Suld2DV2I8Zero: return "NVPTXISD::Suld2DV2I8Zero";
849 case NVPTXISD::Suld2DV2I16Zero: return "NVPTXISD::Suld2DV2I16Zero";
850 case NVPTXISD::Suld2DV2I32Zero: return "NVPTXISD::Suld2DV2I32Zero";
851 case NVPTXISD::Suld2DV2I64Zero: return "NVPTXISD::Suld2DV2I64Zero";
852 case NVPTXISD::Suld2DV4I8Zero: return "NVPTXISD::Suld2DV4I8Zero";
853 case NVPTXISD::Suld2DV4I16Zero: return "NVPTXISD::Suld2DV4I16Zero";
854 case NVPTXISD::Suld2DV4I32Zero: return "NVPTXISD::Suld2DV4I32Zero";
856 case NVPTXISD::Suld2DArrayI8Zero: return "NVPTXISD::Suld2DArrayI8Zero";
857 case NVPTXISD::Suld2DArrayI16Zero: return "NVPTXISD::Suld2DArrayI16Zero";
858 case NVPTXISD::Suld2DArrayI32Zero: return "NVPTXISD::Suld2DArrayI32Zero";
859 case NVPTXISD::Suld2DArrayI64Zero: return "NVPTXISD::Suld2DArrayI64Zero";
860 case NVPTXISD::Suld2DArrayV2I8Zero: return "NVPTXISD::Suld2DArrayV2I8Zero";
861 case NVPTXISD::Suld2DArrayV2I16Zero: return "NVPTXISD::Suld2DArrayV2I16Zero";
862 case NVPTXISD::Suld2DArrayV2I32Zero: return "NVPTXISD::Suld2DArrayV2I32Zero";
863 case NVPTXISD::Suld2DArrayV2I64Zero: return "NVPTXISD::Suld2DArrayV2I64Zero";
864 case NVPTXISD::Suld2DArrayV4I8Zero: return "NVPTXISD::Suld2DArrayV4I8Zero";
865 case NVPTXISD::Suld2DArrayV4I16Zero: return "NVPTXISD::Suld2DArrayV4I16Zero";
866 case NVPTXISD::Suld2DArrayV4I32Zero: return "NVPTXISD::Suld2DArrayV4I32Zero";
868 case NVPTXISD::Suld3DI8Zero: return "NVPTXISD::Suld3DI8Zero";
869 case NVPTXISD::Suld3DI16Zero: return "NVPTXISD::Suld3DI16Zero";
870 case NVPTXISD::Suld3DI32Zero: return "NVPTXISD::Suld3DI32Zero";
871 case NVPTXISD::Suld3DI64Zero: return "NVPTXISD::Suld3DI64Zero";
872 case NVPTXISD::Suld3DV2I8Zero: return "NVPTXISD::Suld3DV2I8Zero";
873 case NVPTXISD::Suld3DV2I16Zero: return "NVPTXISD::Suld3DV2I16Zero";
874 case NVPTXISD::Suld3DV2I32Zero: return "NVPTXISD::Suld3DV2I32Zero";
875 case NVPTXISD::Suld3DV2I64Zero: return "NVPTXISD::Suld3DV2I64Zero";
876 case NVPTXISD::Suld3DV4I8Zero: return "NVPTXISD::Suld3DV4I8Zero";
877 case NVPTXISD::Suld3DV4I16Zero: return "NVPTXISD::Suld3DV4I16Zero";
878 case NVPTXISD::Suld3DV4I32Zero: return "NVPTXISD::Suld3DV4I32Zero";
883 TargetLoweringBase::LegalizeTypeAction
884 NVPTXTargetLowering::getPreferredVectorAction(EVT VT) const {
885 if (VT.getVectorNumElements() != 1 && VT.getScalarType() == MVT::i1)
886 return TypeSplitVector;
888 return TargetLoweringBase::getPreferredVectorAction(VT);
892 NVPTXTargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
894 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
895 auto PtrVT = getPointerTy(DAG.getDataLayout());
896 Op = DAG.getTargetGlobalAddress(GV, dl, PtrVT);
897 return DAG.getNode(NVPTXISD::Wrapper, dl, PtrVT, Op);
900 std::string NVPTXTargetLowering::getPrototype(
901 const DataLayout &DL, Type *retTy, const ArgListTy &Args,
902 const SmallVectorImpl<ISD::OutputArg> &Outs, unsigned retAlignment,
903 const ImmutableCallSite *CS) const {
904 auto PtrVT = getPointerTy(DL);
906 bool isABI = (STI.getSmVersion() >= 20);
907 assert(isABI && "Non-ABI compilation is not supported");
912 O << "prototype_" << uniqueCallSite << " : .callprototype ";
914 if (retTy->getTypeID() == Type::VoidTyID) {
918 if (retTy->isFloatingPointTy() || retTy->isIntegerTy()) {
920 if (auto *ITy = dyn_cast<IntegerType>(retTy)) {
921 size = ITy->getBitWidth();
925 assert(retTy->isFloatingPointTy() &&
926 "Floating point type expected here");
927 size = retTy->getPrimitiveSizeInBits();
930 O << ".param .b" << size << " _";
931 } else if (isa<PointerType>(retTy)) {
932 O << ".param .b" << PtrVT.getSizeInBits() << " _";
933 } else if ((retTy->getTypeID() == Type::StructTyID) ||
934 isa<VectorType>(retTy)) {
935 auto &DL = CS->getCalledFunction()->getParent()->getDataLayout();
936 O << ".param .align " << retAlignment << " .b8 _["
937 << DL.getTypeAllocSize(retTy) << "]";
939 llvm_unreachable("Unknown return type");
948 for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
949 Type *Ty = Args[i].Ty;
955 if (!Outs[OIdx].Flags.isByVal()) {
956 if (Ty->isAggregateType() || Ty->isVectorTy()) {
958 const CallInst *CallI = cast<CallInst>(CS->getInstruction());
959 // +1 because index 0 is reserved for return type alignment
960 if (!llvm::getAlign(*CallI, i + 1, align))
961 align = DL.getABITypeAlignment(Ty);
962 unsigned sz = DL.getTypeAllocSize(Ty);
963 O << ".param .align " << align << " .b8 ";
965 O << "[" << sz << "]";
966 // update the index for Outs
967 SmallVector<EVT, 16> vtparts;
968 ComputeValueVTs(*this, DL, Ty, vtparts);
969 if (unsigned len = vtparts.size())
973 // i8 types in IR will be i16 types in SDAG
974 assert((getValueType(DL, Ty) == Outs[OIdx].VT ||
975 (getValueType(DL, Ty) == MVT::i8 && Outs[OIdx].VT == MVT::i16)) &&
976 "type mismatch between callee prototype and arguments");
979 if (isa<IntegerType>(Ty)) {
980 sz = cast<IntegerType>(Ty)->getBitWidth();
983 } else if (isa<PointerType>(Ty))
984 sz = PtrVT.getSizeInBits();
986 sz = Ty->getPrimitiveSizeInBits();
987 O << ".param .b" << sz << " ";
991 auto *PTy = dyn_cast<PointerType>(Ty);
992 assert(PTy && "Param with byval attribute should be a pointer type");
993 Type *ETy = PTy->getElementType();
995 unsigned align = Outs[OIdx].Flags.getByValAlign();
996 unsigned sz = DL.getTypeAllocSize(ETy);
997 O << ".param .align " << align << " .b8 ";
999 O << "[" << sz << "]";
1006 NVPTXTargetLowering::getArgumentAlignment(SDValue Callee,
1007 const ImmutableCallSite *CS,
1009 unsigned Idx) const {
1011 const Value *DirectCallee = CS->getCalledFunction();
1013 if (!DirectCallee) {
1014 // We don't have a direct function symbol, but that may be because of
1015 // constant cast instructions in the call.
1016 const Instruction *CalleeI = CS->getInstruction();
1017 assert(CalleeI && "Call target is not a function or derived value?");
1019 // With bitcast'd call targets, the instruction will be the call
1020 if (isa<CallInst>(CalleeI)) {
1021 // Check if we have call alignment metadata
1022 if (llvm::getAlign(*cast<CallInst>(CalleeI), Idx, Align))
1025 const Value *CalleeV = cast<CallInst>(CalleeI)->getCalledValue();
1026 // Ignore any bitcast instructions
1027 while(isa<ConstantExpr>(CalleeV)) {
1028 const ConstantExpr *CE = cast<ConstantExpr>(CalleeV);
1031 // Look through the bitcast
1032 CalleeV = cast<ConstantExpr>(CalleeV)->getOperand(0);
1035 // We have now looked past all of the bitcasts. Do we finally have a
1037 if (isa<Function>(CalleeV))
1038 DirectCallee = CalleeV;
1042 // Check for function alignment information if we found that the
1043 // ultimate target is a Function
1045 if (llvm::getAlign(*cast<Function>(DirectCallee), Idx, Align))
1048 // Call is indirect or alignment information is not available, fall back to
1049 // the ABI type alignment
1050 auto &DL = CS->getCaller()->getParent()->getDataLayout();
1051 return DL.getABITypeAlignment(Ty);
1054 SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1055 SmallVectorImpl<SDValue> &InVals) const {
1056 SelectionDAG &DAG = CLI.DAG;
1058 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1059 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1060 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1061 SDValue Chain = CLI.Chain;
1062 SDValue Callee = CLI.Callee;
1063 bool &isTailCall = CLI.IsTailCall;
1064 ArgListTy &Args = CLI.getArgs();
1065 Type *retTy = CLI.RetTy;
1066 ImmutableCallSite *CS = CLI.CS;
1068 bool isABI = (STI.getSmVersion() >= 20);
1069 assert(isABI && "Non-ABI compilation is not supported");
1072 MachineFunction &MF = DAG.getMachineFunction();
1073 const Function *F = MF.getFunction();
1074 auto &DL = MF.getDataLayout();
1076 SDValue tempChain = Chain;
1077 Chain = DAG.getCALLSEQ_START(Chain,
1078 DAG.getIntPtrConstant(uniqueCallSite, dl, true),
1080 SDValue InFlag = Chain.getValue(1);
1082 unsigned paramCount = 0;
1083 // Args.size() and Outs.size() need not match.
1084 // Outs.size() will be larger
1085 // * if there is an aggregate argument with multiple fields (each field
1086 // showing up separately in Outs)
1087 // * if there is a vector argument with more than typical vector-length
1088 // elements (generally if more than 4) where each vector element is
1089 // individually present in Outs.
1090 // So a different index should be used for indexing into Outs/OutVals.
1091 // See similar issue in LowerFormalArguments.
1093 // Declare the .params or .reg need to pass values
1095 for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
1096 EVT VT = Outs[OIdx].VT;
1097 Type *Ty = Args[i].Ty;
1099 if (!Outs[OIdx].Flags.isByVal()) {
1100 if (Ty->isAggregateType()) {
1102 SmallVector<EVT, 16> vtparts;
1103 SmallVector<uint64_t, 16> Offsets;
1104 ComputePTXValueVTs(*this, DAG.getDataLayout(), Ty, vtparts, &Offsets,
1107 unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
1108 // declare .param .align <align> .b8 .param<n>[<size>];
1109 unsigned sz = DL.getTypeAllocSize(Ty);
1110 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1111 SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, dl,
1113 DAG.getConstant(paramCount, dl, MVT::i32),
1114 DAG.getConstant(sz, dl, MVT::i32),
1116 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1118 InFlag = Chain.getValue(1);
1119 for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
1120 EVT elemtype = vtparts[j];
1121 unsigned ArgAlign = GreatestCommonDivisor64(align, Offsets[j]);
1122 if (elemtype.isInteger() && (sz < 8))
1124 SDValue StVal = OutVals[OIdx];
1125 if (elemtype.getSizeInBits() < 16) {
1126 StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
1128 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1129 SDValue CopyParamOps[] = { Chain,
1130 DAG.getConstant(paramCount, dl, MVT::i32),
1131 DAG.getConstant(Offsets[j], dl, MVT::i32),
1133 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
1134 CopyParamVTs, CopyParamOps,
1135 elemtype, MachinePointerInfo(),
1137 InFlag = Chain.getValue(1);
1140 if (vtparts.size() > 0)
1145 if (Ty->isVectorTy()) {
1146 EVT ObjectVT = getValueType(DL, Ty);
1147 unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
1148 // declare .param .align <align> .b8 .param<n>[<size>];
1149 unsigned sz = DL.getTypeAllocSize(Ty);
1150 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1151 SDValue DeclareParamOps[] = { Chain,
1152 DAG.getConstant(align, dl, MVT::i32),
1153 DAG.getConstant(paramCount, dl, MVT::i32),
1154 DAG.getConstant(sz, dl, MVT::i32),
1156 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1158 InFlag = Chain.getValue(1);
1159 unsigned NumElts = ObjectVT.getVectorNumElements();
1160 EVT EltVT = ObjectVT.getVectorElementType();
1162 bool NeedExtend = false;
1163 if (EltVT.getSizeInBits() < 16) {
1170 SDValue Elt = OutVals[OIdx++];
1172 Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt);
1174 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1175 SDValue CopyParamOps[] = { Chain,
1176 DAG.getConstant(paramCount, dl, MVT::i32),
1177 DAG.getConstant(0, dl, MVT::i32), Elt,
1179 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
1180 CopyParamVTs, CopyParamOps,
1181 MemVT, MachinePointerInfo());
1182 InFlag = Chain.getValue(1);
1183 } else if (NumElts == 2) {
1184 SDValue Elt0 = OutVals[OIdx++];
1185 SDValue Elt1 = OutVals[OIdx++];
1187 Elt0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt0);
1188 Elt1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt1);
1191 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1192 SDValue CopyParamOps[] = { Chain,
1193 DAG.getConstant(paramCount, dl, MVT::i32),
1194 DAG.getConstant(0, dl, MVT::i32), Elt0,
1196 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParamV2, dl,
1197 CopyParamVTs, CopyParamOps,
1198 MemVT, MachinePointerInfo());
1199 InFlag = Chain.getValue(1);
1201 unsigned curOffset = 0;
1203 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
1205 // vector will be expanded to a power of 2 elements, so we know we can
1206 // always round up to the next multiple of 4 when creating the vector
1208 // e.g. 4 elem => 1 st.v4
1209 // 6 elem => 2 st.v4
1210 // 8 elem => 2 st.v4
1211 // 11 elem => 3 st.v4
1212 unsigned VecSize = 4;
1213 if (EltVT.getSizeInBits() == 64)
1216 // This is potentially only part of a vector, so assume all elements
1217 // are packed together.
1218 unsigned PerStoreOffset = MemVT.getStoreSizeInBits() / 8 * VecSize;
1220 for (unsigned i = 0; i < NumElts; i += VecSize) {
1223 SmallVector<SDValue, 8> Ops;
1224 Ops.push_back(Chain);
1225 Ops.push_back(DAG.getConstant(paramCount, dl, MVT::i32));
1226 Ops.push_back(DAG.getConstant(curOffset, dl, MVT::i32));
1228 unsigned Opc = NVPTXISD::StoreParamV2;
1230 StoreVal = OutVals[OIdx++];
1232 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1233 Ops.push_back(StoreVal);
1235 if (i + 1 < NumElts) {
1236 StoreVal = OutVals[OIdx++];
1239 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1241 StoreVal = DAG.getUNDEF(EltVT);
1243 Ops.push_back(StoreVal);
1246 Opc = NVPTXISD::StoreParamV4;
1247 if (i + 2 < NumElts) {
1248 StoreVal = OutVals[OIdx++];
1251 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1253 StoreVal = DAG.getUNDEF(EltVT);
1255 Ops.push_back(StoreVal);
1257 if (i + 3 < NumElts) {
1258 StoreVal = OutVals[OIdx++];
1261 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1263 StoreVal = DAG.getUNDEF(EltVT);
1265 Ops.push_back(StoreVal);
1268 Ops.push_back(InFlag);
1270 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1271 Chain = DAG.getMemIntrinsicNode(Opc, dl, CopyParamVTs, Ops,
1272 MemVT, MachinePointerInfo());
1273 InFlag = Chain.getValue(1);
1274 curOffset += PerStoreOffset;
1282 // for ABI, declare .param .b<size> .param<n>;
1283 unsigned sz = VT.getSizeInBits();
1284 bool needExtend = false;
1285 if (VT.isInteger()) {
1291 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1292 SDValue DeclareParamOps[] = { Chain,
1293 DAG.getConstant(paramCount, dl, MVT::i32),
1294 DAG.getConstant(sz, dl, MVT::i32),
1295 DAG.getConstant(0, dl, MVT::i32), InFlag };
1296 Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
1298 InFlag = Chain.getValue(1);
1299 SDValue OutV = OutVals[OIdx];
1301 // zext/sext i1 to i16
1302 unsigned opc = ISD::ZERO_EXTEND;
1303 if (Outs[OIdx].Flags.isSExt())
1304 opc = ISD::SIGN_EXTEND;
1305 OutV = DAG.getNode(opc, dl, MVT::i16, OutV);
1307 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1308 SDValue CopyParamOps[] = { Chain,
1309 DAG.getConstant(paramCount, dl, MVT::i32),
1310 DAG.getConstant(0, dl, MVT::i32), OutV,
1313 unsigned opcode = NVPTXISD::StoreParam;
1314 if (Outs[OIdx].Flags.isZExt() && VT.getSizeInBits() < 32)
1315 opcode = NVPTXISD::StoreParamU32;
1316 else if (Outs[OIdx].Flags.isSExt() && VT.getSizeInBits() < 32)
1317 opcode = NVPTXISD::StoreParamS32;
1318 Chain = DAG.getMemIntrinsicNode(opcode, dl, CopyParamVTs, CopyParamOps,
1319 VT, MachinePointerInfo());
1321 InFlag = Chain.getValue(1);
1326 SmallVector<EVT, 16> vtparts;
1327 SmallVector<uint64_t, 16> Offsets;
1328 auto *PTy = dyn_cast<PointerType>(Args[i].Ty);
1329 assert(PTy && "Type of a byval parameter should be pointer");
1330 ComputePTXValueVTs(*this, DAG.getDataLayout(), PTy->getElementType(),
1331 vtparts, &Offsets, 0);
1333 // declare .param .align <align> .b8 .param<n>[<size>];
1334 unsigned sz = Outs[OIdx].Flags.getByValSize();
1335 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1336 unsigned ArgAlign = Outs[OIdx].Flags.getByValAlign();
1337 // The ByValAlign in the Outs[OIdx].Flags is alway set at this point,
1338 // so we don't need to worry about natural alignment or not.
1339 // See TargetLowering::LowerCallTo().
1340 SDValue DeclareParamOps[] = {
1341 Chain, DAG.getConstant(Outs[OIdx].Flags.getByValAlign(), dl, MVT::i32),
1342 DAG.getConstant(paramCount, dl, MVT::i32),
1343 DAG.getConstant(sz, dl, MVT::i32), InFlag
1345 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1347 InFlag = Chain.getValue(1);
1348 for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
1349 EVT elemtype = vtparts[j];
1350 int curOffset = Offsets[j];
1351 unsigned PartAlign = GreatestCommonDivisor64(ArgAlign, curOffset);
1352 auto PtrVT = getPointerTy(DAG.getDataLayout());
1353 SDValue srcAddr = DAG.getNode(ISD::ADD, dl, PtrVT, OutVals[OIdx],
1354 DAG.getConstant(curOffset, dl, PtrVT));
1355 SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
1356 MachinePointerInfo(), PartAlign);
1357 if (elemtype.getSizeInBits() < 16) {
1358 theVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, theVal);
1360 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1361 SDValue CopyParamOps[] = { Chain,
1362 DAG.getConstant(paramCount, dl, MVT::i32),
1363 DAG.getConstant(curOffset, dl, MVT::i32),
1365 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, CopyParamVTs,
1366 CopyParamOps, elemtype,
1367 MachinePointerInfo());
1369 InFlag = Chain.getValue(1);
1374 GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
1375 unsigned retAlignment = 0;
1378 if (Ins.size() > 0) {
1379 SmallVector<EVT, 16> resvtparts;
1380 ComputeValueVTs(*this, DL, retTy, resvtparts);
1383 // .param .align 16 .b8 retval0[<size-in-bytes>], or
1384 // .param .b<size-in-bits> retval0
1385 unsigned resultsz = DL.getTypeAllocSizeInBits(retTy);
1386 // Emit ".param .b<size-in-bits> retval0" instead of byte arrays only for
1387 // these three types to match the logic in
1388 // NVPTXAsmPrinter::printReturnValStr and NVPTXTargetLowering::getPrototype.
1389 // Plus, this behavior is consistent with nvcc's.
1390 if (retTy->isFloatingPointTy() || retTy->isIntegerTy() ||
1391 retTy->isPointerTy()) {
1392 // Scalar needs to be at least 32bit wide
1395 SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1396 SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, dl, MVT::i32),
1397 DAG.getConstant(resultsz, dl, MVT::i32),
1398 DAG.getConstant(0, dl, MVT::i32), InFlag };
1399 Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
1401 InFlag = Chain.getValue(1);
1403 retAlignment = getArgumentAlignment(Callee, CS, retTy, 0);
1404 SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1405 SDValue DeclareRetOps[] = { Chain,
1406 DAG.getConstant(retAlignment, dl, MVT::i32),
1407 DAG.getConstant(resultsz / 8, dl, MVT::i32),
1408 DAG.getConstant(0, dl, MVT::i32), InFlag };
1409 Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs,
1411 InFlag = Chain.getValue(1);
1416 // This is indirect function call case : PTX requires a prototype of the
1418 // proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _);
1419 // to be emitted, and the label has to used as the last arg of call
1421 // The prototype is embedded in a string and put as the operand for a
1422 // CallPrototype SDNode which will print out to the value of the string.
1423 SDVTList ProtoVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1425 getPrototype(DAG.getDataLayout(), retTy, Args, Outs, retAlignment, CS);
1426 const char *ProtoStr =
1427 nvTM->getManagedStrPool()->getManagedString(Proto.c_str())->c_str();
1428 SDValue ProtoOps[] = {
1429 Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InFlag,
1431 Chain = DAG.getNode(NVPTXISD::CallPrototype, dl, ProtoVTs, ProtoOps);
1432 InFlag = Chain.getValue(1);
1434 // Op to just print "call"
1435 SDVTList PrintCallVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1436 SDValue PrintCallOps[] = {
1437 Chain, DAG.getConstant((Ins.size() == 0) ? 0 : 1, dl, MVT::i32), InFlag
1439 // We model convergent calls as separate opcodes.
1440 unsigned Opcode = Func ? NVPTXISD::PrintCallUni : NVPTXISD::PrintCall;
1441 if (CLI.IsConvergent)
1442 Opcode = Opcode == NVPTXISD::PrintCallUni ? NVPTXISD::PrintConvergentCallUni
1443 : NVPTXISD::PrintConvergentCall;
1444 Chain = DAG.getNode(Opcode, dl, PrintCallVTs, PrintCallOps);
1445 InFlag = Chain.getValue(1);
1447 // Ops to print out the function name
1448 SDVTList CallVoidVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1449 SDValue CallVoidOps[] = { Chain, Callee, InFlag };
1450 Chain = DAG.getNode(NVPTXISD::CallVoid, dl, CallVoidVTs, CallVoidOps);
1451 InFlag = Chain.getValue(1);
1453 // Ops to print out the param list
1454 SDVTList CallArgBeginVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1455 SDValue CallArgBeginOps[] = { Chain, InFlag };
1456 Chain = DAG.getNode(NVPTXISD::CallArgBegin, dl, CallArgBeginVTs,
1458 InFlag = Chain.getValue(1);
1460 for (unsigned i = 0, e = paramCount; i != e; ++i) {
1463 opcode = NVPTXISD::LastCallArg;
1465 opcode = NVPTXISD::CallArg;
1466 SDVTList CallArgVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1467 SDValue CallArgOps[] = { Chain, DAG.getConstant(1, dl, MVT::i32),
1468 DAG.getConstant(i, dl, MVT::i32), InFlag };
1469 Chain = DAG.getNode(opcode, dl, CallArgVTs, CallArgOps);
1470 InFlag = Chain.getValue(1);
1472 SDVTList CallArgEndVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1473 SDValue CallArgEndOps[] = { Chain,
1474 DAG.getConstant(Func ? 1 : 0, dl, MVT::i32),
1476 Chain = DAG.getNode(NVPTXISD::CallArgEnd, dl, CallArgEndVTs, CallArgEndOps);
1477 InFlag = Chain.getValue(1);
1480 SDVTList PrototypeVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1481 SDValue PrototypeOps[] = { Chain,
1482 DAG.getConstant(uniqueCallSite, dl, MVT::i32),
1484 Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps);
1485 InFlag = Chain.getValue(1);
1488 // Generate loads from param memory/moves from registers for result
1489 if (Ins.size() > 0) {
1490 if (retTy && retTy->isVectorTy()) {
1491 EVT ObjectVT = getValueType(DL, retTy);
1492 unsigned NumElts = ObjectVT.getVectorNumElements();
1493 EVT EltVT = ObjectVT.getVectorElementType();
1494 assert(STI.getTargetLowering()->getNumRegisters(F->getContext(),
1495 ObjectVT) == NumElts &&
1496 "Vector was not scalarized");
1497 unsigned sz = EltVT.getSizeInBits();
1498 bool needTruncate = sz < 8;
1501 // Just a simple load
1502 SmallVector<EVT, 4> LoadRetVTs;
1503 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1504 // If loading i1/i8 result, generate
1508 LoadRetVTs.push_back(MVT::i16);
1510 LoadRetVTs.push_back(EltVT);
1511 LoadRetVTs.push_back(MVT::Other);
1512 LoadRetVTs.push_back(MVT::Glue);
1513 SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
1514 DAG.getConstant(0, dl, MVT::i32), InFlag};
1515 SDValue retval = DAG.getMemIntrinsicNode(
1516 NVPTXISD::LoadParam, dl,
1517 DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
1518 Chain = retval.getValue(1);
1519 InFlag = retval.getValue(2);
1520 SDValue Ret0 = retval;
1522 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Ret0);
1523 InVals.push_back(Ret0);
1524 } else if (NumElts == 2) {
1526 SmallVector<EVT, 4> LoadRetVTs;
1527 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1528 // If loading i1/i8 result, generate
1532 LoadRetVTs.push_back(MVT::i16);
1533 LoadRetVTs.push_back(MVT::i16);
1535 LoadRetVTs.push_back(EltVT);
1536 LoadRetVTs.push_back(EltVT);
1538 LoadRetVTs.push_back(MVT::Other);
1539 LoadRetVTs.push_back(MVT::Glue);
1540 SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
1541 DAG.getConstant(0, dl, MVT::i32), InFlag};
1542 SDValue retval = DAG.getMemIntrinsicNode(
1543 NVPTXISD::LoadParamV2, dl,
1544 DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
1545 Chain = retval.getValue(2);
1546 InFlag = retval.getValue(3);
1547 SDValue Ret0 = retval.getValue(0);
1548 SDValue Ret1 = retval.getValue(1);
1550 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret0);
1551 InVals.push_back(Ret0);
1552 Ret1 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret1);
1553 InVals.push_back(Ret1);
1555 InVals.push_back(Ret0);
1556 InVals.push_back(Ret1);
1559 // Split into N LoadV4
1561 unsigned VecSize = 4;
1562 unsigned Opc = NVPTXISD::LoadParamV4;
1563 if (EltVT.getSizeInBits() == 64) {
1565 Opc = NVPTXISD::LoadParamV2;
1567 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
1568 for (unsigned i = 0; i < NumElts; i += VecSize) {
1569 SmallVector<EVT, 8> LoadRetVTs;
1570 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1571 // If loading i1/i8 result, generate
1575 for (unsigned j = 0; j < VecSize; ++j)
1576 LoadRetVTs.push_back(MVT::i16);
1578 for (unsigned j = 0; j < VecSize; ++j)
1579 LoadRetVTs.push_back(EltVT);
1581 LoadRetVTs.push_back(MVT::Other);
1582 LoadRetVTs.push_back(MVT::Glue);
1583 SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
1584 DAG.getConstant(Ofst, dl, MVT::i32), InFlag};
1585 SDValue retval = DAG.getMemIntrinsicNode(
1586 Opc, dl, DAG.getVTList(LoadRetVTs),
1587 LoadRetOps, EltVT, MachinePointerInfo());
1589 Chain = retval.getValue(2);
1590 InFlag = retval.getValue(3);
1592 Chain = retval.getValue(4);
1593 InFlag = retval.getValue(5);
1596 for (unsigned j = 0; j < VecSize; ++j) {
1597 if (i + j >= NumElts)
1599 SDValue Elt = retval.getValue(j);
1601 Elt = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
1602 InVals.push_back(Elt);
1604 Ofst += DL.getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
1608 SmallVector<EVT, 16> VTs;
1609 SmallVector<uint64_t, 16> Offsets;
1610 ComputePTXValueVTs(*this, DAG.getDataLayout(), retTy, VTs, &Offsets, 0);
1611 assert(VTs.size() == Ins.size() && "Bad value decomposition");
1612 unsigned RetAlign = getArgumentAlignment(Callee, CS, retTy, 0);
1613 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
1614 unsigned sz = VTs[i].getSizeInBits();
1615 unsigned AlignI = GreatestCommonDivisor64(RetAlign, Offsets[i]);
1616 bool needTruncate = false;
1617 if (VTs[i].isInteger() && sz < 8) {
1619 needTruncate = true;
1622 SmallVector<EVT, 4> LoadRetVTs;
1623 EVT TheLoadType = VTs[i];
1624 if (retTy->isIntegerTy() && DL.getTypeAllocSizeInBits(retTy) < 32) {
1625 // This is for integer types only, and specifically not for
1627 LoadRetVTs.push_back(MVT::i32);
1628 TheLoadType = MVT::i32;
1629 needTruncate = true;
1630 } else if (sz < 16) {
1631 // If loading i1/i8 result, generate
1633 // trunc i16 to i1/i8
1635 // FIXME: Do we need to set needTruncate to true here, too? We could
1636 // not figure out what this branch is for in D17872, so we left it
1637 // alone. The comment above about loading i1/i8 may be wrong, as the
1638 // branch above seems to cover integers of size < 32.
1639 LoadRetVTs.push_back(MVT::i16);
1641 LoadRetVTs.push_back(Ins[i].VT);
1642 LoadRetVTs.push_back(MVT::Other);
1643 LoadRetVTs.push_back(MVT::Glue);
1645 SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
1646 DAG.getConstant(Offsets[i], dl, MVT::i32),
1648 SDValue retval = DAG.getMemIntrinsicNode(
1649 NVPTXISD::LoadParam, dl,
1650 DAG.getVTList(LoadRetVTs), LoadRetOps,
1651 TheLoadType, MachinePointerInfo(), AlignI);
1652 Chain = retval.getValue(1);
1653 InFlag = retval.getValue(2);
1654 SDValue Ret0 = retval.getValue(0);
1656 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, Ins[i].VT, Ret0);
1657 InVals.push_back(Ret0);
1662 Chain = DAG.getCALLSEQ_END(Chain,
1663 DAG.getIntPtrConstant(uniqueCallSite, dl, true),
1664 DAG.getIntPtrConstant(uniqueCallSite + 1, dl,
1669 // set isTailCall to false for now, until we figure out how to express
1670 // tail call optimization in PTX
1675 // By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack()
1676 // (see LegalizeDAG.cpp). This is slow and uses local memory.
1677 // We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5
1679 NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
1680 SDNode *Node = Op.getNode();
1682 SmallVector<SDValue, 8> Ops;
1683 unsigned NumOperands = Node->getNumOperands();
1684 for (unsigned i = 0; i < NumOperands; ++i) {
1685 SDValue SubOp = Node->getOperand(i);
1686 EVT VVT = SubOp.getNode()->getValueType(0);
1687 EVT EltVT = VVT.getVectorElementType();
1688 unsigned NumSubElem = VVT.getVectorNumElements();
1689 for (unsigned j = 0; j < NumSubElem; ++j) {
1690 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp,
1691 DAG.getIntPtrConstant(j, dl)));
1694 return DAG.getBuildVector(Node->getValueType(0), dl, Ops);
1697 /// LowerShiftRightParts - Lower SRL_PARTS, SRA_PARTS, which
1698 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1700 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1702 SDValue NVPTXTargetLowering::LowerShiftRightParts(SDValue Op,
1703 SelectionDAG &DAG) const {
1704 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1705 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
1707 EVT VT = Op.getValueType();
1708 unsigned VTBits = VT.getSizeInBits();
1710 SDValue ShOpLo = Op.getOperand(0);
1711 SDValue ShOpHi = Op.getOperand(1);
1712 SDValue ShAmt = Op.getOperand(2);
1713 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
1715 if (VTBits == 32 && STI.getSmVersion() >= 35) {
1717 // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1718 // {dHi, dLo} = {aHi, aLo} >> Amt
1720 // dLo = shf.r.clamp aLo, aHi, Amt
1722 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1723 SDValue Lo = DAG.getNode(NVPTXISD::FUN_SHFR_CLAMP, dl, VT, ShOpLo, ShOpHi,
1726 SDValue Ops[2] = { Lo, Hi };
1727 return DAG.getMergeValues(Ops, dl);
1731 // {dHi, dLo} = {aHi, aLo} >> Amt
1732 // - if (Amt>=size) then
1733 // dLo = aHi >> (Amt-size)
1734 // dHi = aHi >> Amt (this is either all 0 or all 1)
1736 // dLo = (aLo >>logic Amt) | (aHi << (size-Amt))
1739 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1740 DAG.getConstant(VTBits, dl, MVT::i32),
1742 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
1743 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1744 DAG.getConstant(VTBits, dl, MVT::i32));
1745 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
1746 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
1747 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
1749 SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
1750 DAG.getConstant(VTBits, dl, MVT::i32),
1752 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1753 SDValue Lo = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
1755 SDValue Ops[2] = { Lo, Hi };
1756 return DAG.getMergeValues(Ops, dl);
1760 /// LowerShiftLeftParts - Lower SHL_PARTS, which
1761 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1763 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1765 SDValue NVPTXTargetLowering::LowerShiftLeftParts(SDValue Op,
1766 SelectionDAG &DAG) const {
1767 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1768 assert(Op.getOpcode() == ISD::SHL_PARTS);
1770 EVT VT = Op.getValueType();
1771 unsigned VTBits = VT.getSizeInBits();
1773 SDValue ShOpLo = Op.getOperand(0);
1774 SDValue ShOpHi = Op.getOperand(1);
1775 SDValue ShAmt = Op.getOperand(2);
1777 if (VTBits == 32 && STI.getSmVersion() >= 35) {
1779 // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1780 // {dHi, dLo} = {aHi, aLo} << Amt
1781 // dHi = shf.l.clamp aLo, aHi, Amt
1784 SDValue Hi = DAG.getNode(NVPTXISD::FUN_SHFL_CLAMP, dl, VT, ShOpLo, ShOpHi,
1786 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1788 SDValue Ops[2] = { Lo, Hi };
1789 return DAG.getMergeValues(Ops, dl);
1793 // {dHi, dLo} = {aHi, aLo} << Amt
1794 // - if (Amt>=size) then
1795 // dLo = aLo << Amt (all 0)
1796 // dLo = aLo << (Amt-size)
1799 // dHi = (aHi << Amt) | (aLo >> (size-Amt))
1801 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1802 DAG.getConstant(VTBits, dl, MVT::i32),
1804 SDValue Tmp1 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
1805 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1806 DAG.getConstant(VTBits, dl, MVT::i32));
1807 SDValue Tmp2 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
1808 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
1809 SDValue TrueVal = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
1811 SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
1812 DAG.getConstant(VTBits, dl, MVT::i32),
1814 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1815 SDValue Hi = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
1817 SDValue Ops[2] = { Lo, Hi };
1818 return DAG.getMergeValues(Ops, dl);
1823 NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
1824 switch (Op.getOpcode()) {
1825 case ISD::RETURNADDR:
1827 case ISD::FRAMEADDR:
1829 case ISD::GlobalAddress:
1830 return LowerGlobalAddress(Op, DAG);
1831 case ISD::INTRINSIC_W_CHAIN:
1833 case ISD::BUILD_VECTOR:
1834 case ISD::EXTRACT_SUBVECTOR:
1836 case ISD::CONCAT_VECTORS:
1837 return LowerCONCAT_VECTORS(Op, DAG);
1839 return LowerSTORE(Op, DAG);
1841 return LowerLOAD(Op, DAG);
1842 case ISD::SHL_PARTS:
1843 return LowerShiftLeftParts(Op, DAG);
1844 case ISD::SRA_PARTS:
1845 case ISD::SRL_PARTS:
1846 return LowerShiftRightParts(Op, DAG);
1848 return LowerSelect(Op, DAG);
1850 llvm_unreachable("Custom lowering not defined for operation");
1854 SDValue NVPTXTargetLowering::LowerSelect(SDValue Op, SelectionDAG &DAG) const {
1855 SDValue Op0 = Op->getOperand(0);
1856 SDValue Op1 = Op->getOperand(1);
1857 SDValue Op2 = Op->getOperand(2);
1858 SDLoc DL(Op.getNode());
1860 assert(Op.getValueType() == MVT::i1 && "Custom lowering enabled only for i1");
1862 Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op1);
1863 Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op2);
1864 SDValue Select = DAG.getNode(ISD::SELECT, DL, MVT::i32, Op0, Op1, Op2);
1865 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Select);
1870 SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1871 if (Op.getValueType() == MVT::i1)
1872 return LowerLOADi1(Op, DAG);
1879 // v1 = ld i8* addr (-> i16)
1880 // v = trunc i16 to i1
1881 SDValue NVPTXTargetLowering::LowerLOADi1(SDValue Op, SelectionDAG &DAG) const {
1882 SDNode *Node = Op.getNode();
1883 LoadSDNode *LD = cast<LoadSDNode>(Node);
1885 assert(LD->getExtensionType() == ISD::NON_EXTLOAD);
1886 assert(Node->getValueType(0) == MVT::i1 &&
1887 "Custom lowering for i1 load only");
1888 SDValue newLD = DAG.getLoad(MVT::i16, dl, LD->getChain(), LD->getBasePtr(),
1889 LD->getPointerInfo(), LD->getAlignment(),
1890 LD->getMemOperand()->getFlags());
1891 SDValue result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, newLD);
1892 // The legalizer (the caller) is expecting two values from the legalized
1893 // load, so we build a MergeValues node for it. See ExpandUnalignedLoad()
1894 // in LegalizeDAG.cpp which also uses MergeValues.
1895 SDValue Ops[] = { result, LD->getChain() };
1896 return DAG.getMergeValues(Ops, dl);
1899 SDValue NVPTXTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1900 EVT ValVT = Op.getOperand(1).getValueType();
1901 if (ValVT == MVT::i1)
1902 return LowerSTOREi1(Op, DAG);
1903 else if (ValVT.isVector())
1904 return LowerSTOREVector(Op, DAG);
1910 NVPTXTargetLowering::LowerSTOREVector(SDValue Op, SelectionDAG &DAG) const {
1911 SDNode *N = Op.getNode();
1912 SDValue Val = N->getOperand(1);
1914 EVT ValVT = Val.getValueType();
1916 if (ValVT.isVector()) {
1917 // We only handle "native" vector sizes for now, e.g. <4 x double> is not
1918 // legal. We can (and should) split that into 2 stores of <2 x double> here
1919 // but I'm leaving that as a TODO for now.
1920 if (!ValVT.isSimple())
1922 switch (ValVT.getSimpleVT().SimpleTy) {
1935 // This is a "native" vector type
1939 MemSDNode *MemSD = cast<MemSDNode>(N);
1940 const DataLayout &TD = DAG.getDataLayout();
1942 unsigned Align = MemSD->getAlignment();
1943 unsigned PrefAlign =
1944 TD.getPrefTypeAlignment(ValVT.getTypeForEVT(*DAG.getContext()));
1945 if (Align < PrefAlign) {
1946 // This store is not sufficiently aligned, so bail out and let this vector
1947 // store be scalarized. Note that we may still be able to emit smaller
1948 // vector stores. For example, if we are storing a <4 x float> with an
1949 // alignment of 8, this check will fail but the legalizer will try again
1950 // with 2 x <2 x float>, which will succeed with an alignment of 8.
1954 unsigned Opcode = 0;
1955 EVT EltVT = ValVT.getVectorElementType();
1956 unsigned NumElts = ValVT.getVectorNumElements();
1958 // Since StoreV2 is a target node, we cannot rely on DAG type legalization.
1959 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
1960 // stored type to i16 and propagate the "real" type as the memory type.
1961 bool NeedExt = false;
1962 if (EltVT.getSizeInBits() < 16)
1969 Opcode = NVPTXISD::StoreV2;
1972 Opcode = NVPTXISD::StoreV4;
1977 SmallVector<SDValue, 8> Ops;
1979 // First is the chain
1980 Ops.push_back(N->getOperand(0));
1982 // Then the split values
1983 for (unsigned i = 0; i < NumElts; ++i) {
1984 SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Val,
1985 DAG.getIntPtrConstant(i, DL));
1987 ExtVal = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i16, ExtVal);
1988 Ops.push_back(ExtVal);
1991 // Then any remaining arguments
1992 Ops.append(N->op_begin() + 2, N->op_end());
1994 SDValue NewSt = DAG.getMemIntrinsicNode(
1995 Opcode, DL, DAG.getVTList(MVT::Other), Ops,
1996 MemSD->getMemoryVT(), MemSD->getMemOperand());
1998 //return DCI.CombineTo(N, NewSt, true);
2007 // v1 = zxt v to i16
2009 SDValue NVPTXTargetLowering::LowerSTOREi1(SDValue Op, SelectionDAG &DAG) const {
2010 SDNode *Node = Op.getNode();
2012 StoreSDNode *ST = cast<StoreSDNode>(Node);
2013 SDValue Tmp1 = ST->getChain();
2014 SDValue Tmp2 = ST->getBasePtr();
2015 SDValue Tmp3 = ST->getValue();
2016 assert(Tmp3.getValueType() == MVT::i1 && "Custom lowering for i1 store only");
2017 Tmp3 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Tmp3);
2019 DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2, ST->getPointerInfo(), MVT::i8,
2020 ST->getAlignment(), ST->getMemOperand()->getFlags());
2025 NVPTXTargetLowering::getParamSymbol(SelectionDAG &DAG, int idx, EVT v) const {
2026 std::string ParamSym;
2027 raw_string_ostream ParamStr(ParamSym);
2029 ParamStr << DAG.getMachineFunction().getName() << "_param_" << idx;
2032 std::string *SavedStr =
2033 nvTM->getManagedStrPool()->getManagedString(ParamSym.c_str());
2034 return DAG.getTargetExternalSymbol(SavedStr->c_str(), v);
2037 // Check to see if the kernel argument is image*_t or sampler_t
2039 static bool isImageOrSamplerVal(const Value *arg, const Module *context) {
2040 static const char *const specialTypes[] = { "struct._image2d_t",
2041 "struct._image3d_t",
2042 "struct._sampler_t" };
2044 Type *Ty = arg->getType();
2045 auto *PTy = dyn_cast<PointerType>(Ty);
2053 auto *STy = dyn_cast<StructType>(PTy->getElementType());
2054 if (!STy || STy->isLiteral())
2057 return std::find(std::begin(specialTypes), std::end(specialTypes),
2058 STy->getName()) != std::end(specialTypes);
2061 SDValue NVPTXTargetLowering::LowerFormalArguments(
2062 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2063 const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &dl,
2064 SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
2065 MachineFunction &MF = DAG.getMachineFunction();
2066 const DataLayout &DL = DAG.getDataLayout();
2067 auto PtrVT = getPointerTy(DAG.getDataLayout());
2069 const Function *F = MF.getFunction();
2070 const AttributeSet &PAL = F->getAttributes();
2071 const TargetLowering *TLI = STI.getTargetLowering();
2073 SDValue Root = DAG.getRoot();
2074 std::vector<SDValue> OutChains;
2076 bool isKernel = llvm::isKernelFunction(*F);
2077 bool isABI = (STI.getSmVersion() >= 20);
2078 assert(isABI && "Non-ABI compilation is not supported");
2082 std::vector<Type *> argTypes;
2083 std::vector<const Argument *> theArgs;
2084 for (const Argument &I : F->args()) {
2085 theArgs.push_back(&I);
2086 argTypes.push_back(I.getType());
2088 // argTypes.size() (or theArgs.size()) and Ins.size() need not match.
2089 // Ins.size() will be larger
2090 // * if there is an aggregate argument with multiple fields (each field
2091 // showing up separately in Ins)
2092 // * if there is a vector argument with more than typical vector-length
2093 // elements (generally if more than 4) where each vector element is
2094 // individually present in Ins.
2095 // So a different index should be used for indexing into Ins.
2096 // See similar issue in LowerCall.
2097 unsigned InsIdx = 0;
2100 for (unsigned i = 0, e = theArgs.size(); i != e; ++i, ++idx, ++InsIdx) {
2101 Type *Ty = argTypes[i];
2103 // If the kernel argument is image*_t or sampler_t, convert it to
2104 // a i32 constant holding the parameter position. This can later
2105 // matched in the AsmPrinter to output the correct mangled name.
2106 if (isImageOrSamplerVal(
2108 (theArgs[i]->getParent() ? theArgs[i]->getParent()->getParent()
2110 assert(isKernel && "Only kernels can have image/sampler params");
2111 InVals.push_back(DAG.getConstant(i + 1, dl, MVT::i32));
2115 if (theArgs[i]->use_empty()) {
2117 if (Ty->isAggregateType()) {
2118 SmallVector<EVT, 16> vtparts;
2120 ComputePTXValueVTs(*this, DAG.getDataLayout(), Ty, vtparts);
2121 assert(vtparts.size() > 0 && "empty aggregate type not expected");
2122 for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2124 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2127 if (vtparts.size() > 0)
2131 if (Ty->isVectorTy()) {
2132 EVT ObjectVT = getValueType(DL, Ty);
2133 unsigned NumRegs = TLI->getNumRegisters(F->getContext(), ObjectVT);
2134 for (unsigned parti = 0; parti < NumRegs; ++parti) {
2135 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2142 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2146 // In the following cases, assign a node order of "idx+1"
2147 // to newly created nodes. The SDNodes for params have to
2148 // appear in the same order as their order of appearance
2149 // in the original function. "idx+1" holds that order.
2150 if (!PAL.hasAttribute(i + 1, Attribute::ByVal)) {
2151 if (Ty->isAggregateType()) {
2152 SmallVector<EVT, 16> vtparts;
2153 SmallVector<uint64_t, 16> offsets;
2155 // NOTE: Here, we lose the ability to issue vector loads for vectors
2156 // that are a part of a struct. This should be investigated in the
2158 ComputePTXValueVTs(*this, DAG.getDataLayout(), Ty, vtparts, &offsets,
2160 assert(vtparts.size() > 0 && "empty aggregate type not expected");
2161 bool aggregateIsPacked = false;
2162 if (StructType *STy = llvm::dyn_cast<StructType>(Ty))
2163 aggregateIsPacked = STy->isPacked();
2165 SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
2166 for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2168 EVT partVT = vtparts[parti];
2169 Value *srcValue = Constant::getNullValue(
2170 PointerType::get(partVT.getTypeForEVT(F->getContext()),
2171 llvm::ADDRESS_SPACE_PARAM));
2173 DAG.getNode(ISD::ADD, dl, PtrVT, Arg,
2174 DAG.getConstant(offsets[parti], dl, PtrVT));
2175 unsigned partAlign = aggregateIsPacked
2177 : DL.getABITypeAlignment(
2178 partVT.getTypeForEVT(F->getContext()));
2180 if (Ins[InsIdx].VT.getSizeInBits() > partVT.getSizeInBits()) {
2181 ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ?
2182 ISD::SEXTLOAD : ISD::ZEXTLOAD;
2183 p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, srcAddr,
2184 MachinePointerInfo(srcValue), partVT, partAlign);
2186 p = DAG.getLoad(partVT, dl, Root, srcAddr,
2187 MachinePointerInfo(srcValue), partAlign);
2190 p.getNode()->setIROrder(idx + 1);
2191 InVals.push_back(p);
2194 if (vtparts.size() > 0)
2198 if (Ty->isVectorTy()) {
2199 EVT ObjectVT = getValueType(DL, Ty);
2200 SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
2201 unsigned NumElts = ObjectVT.getVectorNumElements();
2202 assert(TLI->getNumRegisters(F->getContext(), ObjectVT) == NumElts &&
2203 "Vector was not scalarized");
2204 EVT EltVT = ObjectVT.getVectorElementType();
2209 // We only have one element, so just directly load it
2210 Value *SrcValue = Constant::getNullValue(PointerType::get(
2211 EltVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2212 SDValue P = DAG.getLoad(
2213 EltVT, dl, Root, Arg, MachinePointerInfo(SrcValue),
2214 DL.getABITypeAlignment(EltVT.getTypeForEVT(F->getContext())),
2215 MachineMemOperand::MOInvariant);
2217 P.getNode()->setIROrder(idx + 1);
2219 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
2220 P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P);
2221 InVals.push_back(P);
2223 } else if (NumElts == 2) {
2225 // f32,f32 = load ...
2226 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2);
2227 Value *SrcValue = Constant::getNullValue(PointerType::get(
2228 VecVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2229 SDValue P = DAG.getLoad(
2230 VecVT, dl, Root, Arg, MachinePointerInfo(SrcValue),
2231 DL.getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())),
2232 MachineMemOperand::MOInvariant);
2234 P.getNode()->setIROrder(idx + 1);
2236 SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2237 DAG.getIntPtrConstant(0, dl));
2238 SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2239 DAG.getIntPtrConstant(1, dl));
2241 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) {
2242 Elt0 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt0);
2243 Elt1 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt1);
2246 InVals.push_back(Elt0);
2247 InVals.push_back(Elt1);
2251 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
2252 // the vector will be expanded to a power of 2 elements, so we know we
2253 // can always round up to the next multiple of 4 when creating the
2255 // e.g. 4 elem => 1 ld.v4
2256 // 6 elem => 2 ld.v4
2257 // 8 elem => 2 ld.v4
2258 // 11 elem => 3 ld.v4
2259 unsigned VecSize = 4;
2260 if (EltVT.getSizeInBits() == 64) {
2263 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
2265 for (unsigned i = 0; i < NumElts; i += VecSize) {
2266 Value *SrcValue = Constant::getNullValue(
2267 PointerType::get(VecVT.getTypeForEVT(F->getContext()),
2268 llvm::ADDRESS_SPACE_PARAM));
2269 SDValue SrcAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Arg,
2270 DAG.getConstant(Ofst, dl, PtrVT));
2271 SDValue P = DAG.getLoad(
2272 VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue),
2273 DL.getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())),
2274 MachineMemOperand::MOInvariant);
2276 P.getNode()->setIROrder(idx + 1);
2278 for (unsigned j = 0; j < VecSize; ++j) {
2279 if (i + j >= NumElts)
2281 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2282 DAG.getIntPtrConstant(j, dl));
2283 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
2284 Elt = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt);
2285 InVals.push_back(Elt);
2287 Ofst += DL.getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
2297 EVT ObjectVT = getValueType(DL, Ty);
2298 // If ABI, load from the param symbol
2299 SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
2300 Value *srcValue = Constant::getNullValue(PointerType::get(
2301 ObjectVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2303 if (ObjectVT.getSizeInBits() < Ins[InsIdx].VT.getSizeInBits()) {
2304 ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ?
2305 ISD::SEXTLOAD : ISD::ZEXTLOAD;
2307 ExtOp, dl, Ins[InsIdx].VT, Root, Arg, MachinePointerInfo(srcValue),
2309 DL.getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
2312 Ins[InsIdx].VT, dl, Root, Arg, MachinePointerInfo(srcValue),
2313 DL.getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
2316 p.getNode()->setIROrder(idx + 1);
2317 InVals.push_back(p);
2321 // Param has ByVal attribute
2322 // Return MoveParam(param symbol).
2323 // Ideally, the param symbol can be returned directly,
2324 // but when SDNode builder decides to use it in a CopyToReg(),
2325 // machine instruction fails because TargetExternalSymbol
2326 // (not lowered) is target dependent, and CopyToReg assumes
2327 // the source is lowered.
2328 EVT ObjectVT = getValueType(DL, Ty);
2329 assert(ObjectVT == Ins[InsIdx].VT &&
2330 "Ins type did not match function type");
2331 SDValue Arg = getParamSymbol(DAG, idx, PtrVT);
2332 SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg);
2334 p.getNode()->setIROrder(idx + 1);
2336 InVals.push_back(p);
2338 SDValue p2 = DAG.getNode(
2339 ISD::INTRINSIC_WO_CHAIN, dl, ObjectVT,
2340 DAG.getConstant(Intrinsic::nvvm_ptr_local_to_gen, dl, MVT::i32), p);
2341 InVals.push_back(p2);
2345 // Clang will check explicit VarArg and issue error if any. However, Clang
2346 // will let code with
2347 // implicit var arg like f() pass. See bug 617733.
2348 // We treat this case as if the arg list is empty.
2349 // if (F.isVarArg()) {
2350 // assert(0 && "VarArg not supported yet!");
2353 if (!OutChains.empty())
2354 DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains));
2360 NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2362 const SmallVectorImpl<ISD::OutputArg> &Outs,
2363 const SmallVectorImpl<SDValue> &OutVals,
2364 const SDLoc &dl, SelectionDAG &DAG) const {
2365 MachineFunction &MF = DAG.getMachineFunction();
2366 const Function *F = MF.getFunction();
2367 Type *RetTy = F->getReturnType();
2368 const DataLayout &TD = DAG.getDataLayout();
2370 bool isABI = (STI.getSmVersion() >= 20);
2371 assert(isABI && "Non-ABI compilation is not supported");
2375 if (VectorType *VTy = dyn_cast<VectorType>(RetTy)) {
2376 // If we have a vector type, the OutVals array will be the scalarized
2377 // components and we have combine them into 1 or more vector stores.
2378 unsigned NumElts = VTy->getNumElements();
2379 assert(NumElts == Outs.size() && "Bad scalarization of return value");
2381 // const_cast can be removed in later LLVM versions
2382 EVT EltVT = getValueType(TD, RetTy).getVectorElementType();
2383 bool NeedExtend = false;
2384 if (EltVT.getSizeInBits() < 16)
2389 SDValue StoreVal = OutVals[0];
2390 // We only have one element, so just directly store it
2392 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
2393 SDValue Ops[] = { Chain, DAG.getConstant(0, dl, MVT::i32), StoreVal };
2394 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
2395 DAG.getVTList(MVT::Other), Ops,
2396 EltVT, MachinePointerInfo());
2398 } else if (NumElts == 2) {
2400 SDValue StoreVal0 = OutVals[0];
2401 SDValue StoreVal1 = OutVals[1];
2404 StoreVal0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal0);
2405 StoreVal1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal1);
2408 SDValue Ops[] = { Chain, DAG.getConstant(0, dl, MVT::i32), StoreVal0,
2410 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetvalV2, dl,
2411 DAG.getVTList(MVT::Other), Ops,
2412 EltVT, MachinePointerInfo());
2415 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and the
2416 // vector will be expanded to a power of 2 elements, so we know we can
2417 // always round up to the next multiple of 4 when creating the vector
2419 // e.g. 4 elem => 1 st.v4
2420 // 6 elem => 2 st.v4
2421 // 8 elem => 2 st.v4
2422 // 11 elem => 3 st.v4
2424 unsigned VecSize = 4;
2425 if (OutVals[0].getValueType().getSizeInBits() == 64)
2428 unsigned Offset = 0;
2431 EVT::getVectorVT(F->getContext(), EltVT, VecSize);
2432 unsigned PerStoreOffset =
2433 TD.getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
2435 for (unsigned i = 0; i < NumElts; i += VecSize) {
2438 SmallVector<SDValue, 8> Ops;
2439 Ops.push_back(Chain);
2440 Ops.push_back(DAG.getConstant(Offset, dl, MVT::i32));
2441 unsigned Opc = NVPTXISD::StoreRetvalV2;
2442 EVT ExtendedVT = (NeedExtend) ? MVT::i16 : OutVals[0].getValueType();
2444 StoreVal = OutVals[i];
2446 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2447 Ops.push_back(StoreVal);
2449 if (i + 1 < NumElts) {
2450 StoreVal = OutVals[i + 1];
2452 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2454 StoreVal = DAG.getUNDEF(ExtendedVT);
2456 Ops.push_back(StoreVal);
2459 Opc = NVPTXISD::StoreRetvalV4;
2460 if (i + 2 < NumElts) {
2461 StoreVal = OutVals[i + 2];
2464 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2466 StoreVal = DAG.getUNDEF(ExtendedVT);
2468 Ops.push_back(StoreVal);
2470 if (i + 3 < NumElts) {
2471 StoreVal = OutVals[i + 3];
2474 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2476 StoreVal = DAG.getUNDEF(ExtendedVT);
2478 Ops.push_back(StoreVal);
2481 // Chain = DAG.getNode(Opc, dl, MVT::Other, &Ops[0], Ops.size());
2483 DAG.getMemIntrinsicNode(Opc, dl, DAG.getVTList(MVT::Other), Ops,
2484 EltVT, MachinePointerInfo());
2485 Offset += PerStoreOffset;
2489 SmallVector<EVT, 16> ValVTs;
2490 SmallVector<uint64_t, 16> Offsets;
2491 ComputePTXValueVTs(*this, DAG.getDataLayout(), RetTy, ValVTs, &Offsets, 0);
2492 assert(ValVTs.size() == OutVals.size() && "Bad return value decomposition");
2494 for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
2495 SDValue theVal = OutVals[i];
2496 EVT TheValType = theVal.getValueType();
2497 unsigned numElems = 1;
2498 if (TheValType.isVector())
2499 numElems = TheValType.getVectorNumElements();
2500 for (unsigned j = 0, je = numElems; j != je; ++j) {
2501 SDValue TmpVal = theVal;
2502 if (TheValType.isVector())
2503 TmpVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
2504 TheValType.getVectorElementType(), TmpVal,
2505 DAG.getIntPtrConstant(j, dl));
2506 EVT TheStoreType = ValVTs[i];
2507 if (RetTy->isIntegerTy() && TD.getTypeAllocSizeInBits(RetTy) < 32) {
2508 // The following zero-extension is for integer types only, and
2509 // specifically not for aggregates.
2510 TmpVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, TmpVal);
2511 TheStoreType = MVT::i32;
2513 else if (TmpVal.getValueType().getSizeInBits() < 16)
2514 TmpVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, TmpVal);
2518 DAG.getConstant(Offsets[i], dl, MVT::i32),
2520 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
2521 DAG.getVTList(MVT::Other), Ops,
2523 MachinePointerInfo());
2528 return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain);
2532 void NVPTXTargetLowering::LowerAsmOperandForConstraint(
2533 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
2534 SelectionDAG &DAG) const {
2535 if (Constraint.length() > 1)
2538 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2541 static unsigned getOpcForTextureInstr(unsigned Intrinsic) {
2542 switch (Intrinsic) {
2546 case Intrinsic::nvvm_tex_1d_v4f32_s32:
2547 return NVPTXISD::Tex1DFloatS32;
2548 case Intrinsic::nvvm_tex_1d_v4f32_f32:
2549 return NVPTXISD::Tex1DFloatFloat;
2550 case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
2551 return NVPTXISD::Tex1DFloatFloatLevel;
2552 case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
2553 return NVPTXISD::Tex1DFloatFloatGrad;
2554 case Intrinsic::nvvm_tex_1d_v4s32_s32:
2555 return NVPTXISD::Tex1DS32S32;
2556 case Intrinsic::nvvm_tex_1d_v4s32_f32:
2557 return NVPTXISD::Tex1DS32Float;
2558 case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
2559 return NVPTXISD::Tex1DS32FloatLevel;
2560 case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
2561 return NVPTXISD::Tex1DS32FloatGrad;
2562 case Intrinsic::nvvm_tex_1d_v4u32_s32:
2563 return NVPTXISD::Tex1DU32S32;
2564 case Intrinsic::nvvm_tex_1d_v4u32_f32:
2565 return NVPTXISD::Tex1DU32Float;
2566 case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
2567 return NVPTXISD::Tex1DU32FloatLevel;
2568 case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
2569 return NVPTXISD::Tex1DU32FloatGrad;
2571 case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
2572 return NVPTXISD::Tex1DArrayFloatS32;
2573 case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
2574 return NVPTXISD::Tex1DArrayFloatFloat;
2575 case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
2576 return NVPTXISD::Tex1DArrayFloatFloatLevel;
2577 case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
2578 return NVPTXISD::Tex1DArrayFloatFloatGrad;
2579 case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
2580 return NVPTXISD::Tex1DArrayS32S32;
2581 case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
2582 return NVPTXISD::Tex1DArrayS32Float;
2583 case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
2584 return NVPTXISD::Tex1DArrayS32FloatLevel;
2585 case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
2586 return NVPTXISD::Tex1DArrayS32FloatGrad;
2587 case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
2588 return NVPTXISD::Tex1DArrayU32S32;
2589 case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
2590 return NVPTXISD::Tex1DArrayU32Float;
2591 case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
2592 return NVPTXISD::Tex1DArrayU32FloatLevel;
2593 case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
2594 return NVPTXISD::Tex1DArrayU32FloatGrad;
2596 case Intrinsic::nvvm_tex_2d_v4f32_s32:
2597 return NVPTXISD::Tex2DFloatS32;
2598 case Intrinsic::nvvm_tex_2d_v4f32_f32:
2599 return NVPTXISD::Tex2DFloatFloat;
2600 case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
2601 return NVPTXISD::Tex2DFloatFloatLevel;
2602 case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
2603 return NVPTXISD::Tex2DFloatFloatGrad;
2604 case Intrinsic::nvvm_tex_2d_v4s32_s32:
2605 return NVPTXISD::Tex2DS32S32;
2606 case Intrinsic::nvvm_tex_2d_v4s32_f32:
2607 return NVPTXISD::Tex2DS32Float;
2608 case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
2609 return NVPTXISD::Tex2DS32FloatLevel;
2610 case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
2611 return NVPTXISD::Tex2DS32FloatGrad;
2612 case Intrinsic::nvvm_tex_2d_v4u32_s32:
2613 return NVPTXISD::Tex2DU32S32;
2614 case Intrinsic::nvvm_tex_2d_v4u32_f32:
2615 return NVPTXISD::Tex2DU32Float;
2616 case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
2617 return NVPTXISD::Tex2DU32FloatLevel;
2618 case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
2619 return NVPTXISD::Tex2DU32FloatGrad;
2621 case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
2622 return NVPTXISD::Tex2DArrayFloatS32;
2623 case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
2624 return NVPTXISD::Tex2DArrayFloatFloat;
2625 case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
2626 return NVPTXISD::Tex2DArrayFloatFloatLevel;
2627 case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
2628 return NVPTXISD::Tex2DArrayFloatFloatGrad;
2629 case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
2630 return NVPTXISD::Tex2DArrayS32S32;
2631 case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
2632 return NVPTXISD::Tex2DArrayS32Float;
2633 case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
2634 return NVPTXISD::Tex2DArrayS32FloatLevel;
2635 case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
2636 return NVPTXISD::Tex2DArrayS32FloatGrad;
2637 case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
2638 return NVPTXISD::Tex2DArrayU32S32;
2639 case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
2640 return NVPTXISD::Tex2DArrayU32Float;
2641 case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
2642 return NVPTXISD::Tex2DArrayU32FloatLevel;
2643 case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
2644 return NVPTXISD::Tex2DArrayU32FloatGrad;
2646 case Intrinsic::nvvm_tex_3d_v4f32_s32:
2647 return NVPTXISD::Tex3DFloatS32;
2648 case Intrinsic::nvvm_tex_3d_v4f32_f32:
2649 return NVPTXISD::Tex3DFloatFloat;
2650 case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
2651 return NVPTXISD::Tex3DFloatFloatLevel;
2652 case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
2653 return NVPTXISD::Tex3DFloatFloatGrad;
2654 case Intrinsic::nvvm_tex_3d_v4s32_s32:
2655 return NVPTXISD::Tex3DS32S32;
2656 case Intrinsic::nvvm_tex_3d_v4s32_f32:
2657 return NVPTXISD::Tex3DS32Float;
2658 case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
2659 return NVPTXISD::Tex3DS32FloatLevel;
2660 case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
2661 return NVPTXISD::Tex3DS32FloatGrad;
2662 case Intrinsic::nvvm_tex_3d_v4u32_s32:
2663 return NVPTXISD::Tex3DU32S32;
2664 case Intrinsic::nvvm_tex_3d_v4u32_f32:
2665 return NVPTXISD::Tex3DU32Float;
2666 case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
2667 return NVPTXISD::Tex3DU32FloatLevel;
2668 case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
2669 return NVPTXISD::Tex3DU32FloatGrad;
2671 case Intrinsic::nvvm_tex_cube_v4f32_f32:
2672 return NVPTXISD::TexCubeFloatFloat;
2673 case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
2674 return NVPTXISD::TexCubeFloatFloatLevel;
2675 case Intrinsic::nvvm_tex_cube_v4s32_f32:
2676 return NVPTXISD::TexCubeS32Float;
2677 case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
2678 return NVPTXISD::TexCubeS32FloatLevel;
2679 case Intrinsic::nvvm_tex_cube_v4u32_f32:
2680 return NVPTXISD::TexCubeU32Float;
2681 case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
2682 return NVPTXISD::TexCubeU32FloatLevel;
2684 case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
2685 return NVPTXISD::TexCubeArrayFloatFloat;
2686 case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
2687 return NVPTXISD::TexCubeArrayFloatFloatLevel;
2688 case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
2689 return NVPTXISD::TexCubeArrayS32Float;
2690 case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
2691 return NVPTXISD::TexCubeArrayS32FloatLevel;
2692 case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
2693 return NVPTXISD::TexCubeArrayU32Float;
2694 case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
2695 return NVPTXISD::TexCubeArrayU32FloatLevel;
2697 case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
2698 return NVPTXISD::Tld4R2DFloatFloat;
2699 case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
2700 return NVPTXISD::Tld4G2DFloatFloat;
2701 case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
2702 return NVPTXISD::Tld4B2DFloatFloat;
2703 case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
2704 return NVPTXISD::Tld4A2DFloatFloat;
2705 case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
2706 return NVPTXISD::Tld4R2DS64Float;
2707 case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
2708 return NVPTXISD::Tld4G2DS64Float;
2709 case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
2710 return NVPTXISD::Tld4B2DS64Float;
2711 case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
2712 return NVPTXISD::Tld4A2DS64Float;
2713 case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
2714 return NVPTXISD::Tld4R2DU64Float;
2715 case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
2716 return NVPTXISD::Tld4G2DU64Float;
2717 case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
2718 return NVPTXISD::Tld4B2DU64Float;
2719 case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
2720 return NVPTXISD::Tld4A2DU64Float;
2722 case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
2723 return NVPTXISD::TexUnified1DFloatS32;
2724 case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
2725 return NVPTXISD::TexUnified1DFloatFloat;
2726 case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
2727 return NVPTXISD::TexUnified1DFloatFloatLevel;
2728 case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
2729 return NVPTXISD::TexUnified1DFloatFloatGrad;
2730 case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
2731 return NVPTXISD::TexUnified1DS32S32;
2732 case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
2733 return NVPTXISD::TexUnified1DS32Float;
2734 case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
2735 return NVPTXISD::TexUnified1DS32FloatLevel;
2736 case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
2737 return NVPTXISD::TexUnified1DS32FloatGrad;
2738 case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
2739 return NVPTXISD::TexUnified1DU32S32;
2740 case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
2741 return NVPTXISD::TexUnified1DU32Float;
2742 case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
2743 return NVPTXISD::TexUnified1DU32FloatLevel;
2744 case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
2745 return NVPTXISD::TexUnified1DU32FloatGrad;
2747 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
2748 return NVPTXISD::TexUnified1DArrayFloatS32;
2749 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
2750 return NVPTXISD::TexUnified1DArrayFloatFloat;
2751 case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
2752 return NVPTXISD::TexUnified1DArrayFloatFloatLevel;
2753 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
2754 return NVPTXISD::TexUnified1DArrayFloatFloatGrad;
2755 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
2756 return NVPTXISD::TexUnified1DArrayS32S32;
2757 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
2758 return NVPTXISD::TexUnified1DArrayS32Float;
2759 case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
2760 return NVPTXISD::TexUnified1DArrayS32FloatLevel;
2761 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
2762 return NVPTXISD::TexUnified1DArrayS32FloatGrad;
2763 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
2764 return NVPTXISD::TexUnified1DArrayU32S32;
2765 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
2766 return NVPTXISD::TexUnified1DArrayU32Float;
2767 case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
2768 return NVPTXISD::TexUnified1DArrayU32FloatLevel;
2769 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
2770 return NVPTXISD::TexUnified1DArrayU32FloatGrad;
2772 case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
2773 return NVPTXISD::TexUnified2DFloatS32;
2774 case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
2775 return NVPTXISD::TexUnified2DFloatFloat;
2776 case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
2777 return NVPTXISD::TexUnified2DFloatFloatLevel;
2778 case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
2779 return NVPTXISD::TexUnified2DFloatFloatGrad;
2780 case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
2781 return NVPTXISD::TexUnified2DS32S32;
2782 case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
2783 return NVPTXISD::TexUnified2DS32Float;
2784 case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
2785 return NVPTXISD::TexUnified2DS32FloatLevel;
2786 case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
2787 return NVPTXISD::TexUnified2DS32FloatGrad;
2788 case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
2789 return NVPTXISD::TexUnified2DU32S32;
2790 case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
2791 return NVPTXISD::TexUnified2DU32Float;
2792 case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
2793 return NVPTXISD::TexUnified2DU32FloatLevel;
2794 case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
2795 return NVPTXISD::TexUnified2DU32FloatGrad;
2797 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
2798 return NVPTXISD::TexUnified2DArrayFloatS32;
2799 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
2800 return NVPTXISD::TexUnified2DArrayFloatFloat;
2801 case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
2802 return NVPTXISD::TexUnified2DArrayFloatFloatLevel;
2803 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
2804 return NVPTXISD::TexUnified2DArrayFloatFloatGrad;
2805 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
2806 return NVPTXISD::TexUnified2DArrayS32S32;
2807 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
2808 return NVPTXISD::TexUnified2DArrayS32Float;
2809 case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
2810 return NVPTXISD::TexUnified2DArrayS32FloatLevel;
2811 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
2812 return NVPTXISD::TexUnified2DArrayS32FloatGrad;
2813 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
2814 return NVPTXISD::TexUnified2DArrayU32S32;
2815 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
2816 return NVPTXISD::TexUnified2DArrayU32Float;
2817 case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
2818 return NVPTXISD::TexUnified2DArrayU32FloatLevel;
2819 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
2820 return NVPTXISD::TexUnified2DArrayU32FloatGrad;
2822 case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
2823 return NVPTXISD::TexUnified3DFloatS32;
2824 case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
2825 return NVPTXISD::TexUnified3DFloatFloat;
2826 case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
2827 return NVPTXISD::TexUnified3DFloatFloatLevel;
2828 case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
2829 return NVPTXISD::TexUnified3DFloatFloatGrad;
2830 case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
2831 return NVPTXISD::TexUnified3DS32S32;
2832 case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
2833 return NVPTXISD::TexUnified3DS32Float;
2834 case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
2835 return NVPTXISD::TexUnified3DS32FloatLevel;
2836 case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
2837 return NVPTXISD::TexUnified3DS32FloatGrad;
2838 case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
2839 return NVPTXISD::TexUnified3DU32S32;
2840 case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
2841 return NVPTXISD::TexUnified3DU32Float;
2842 case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
2843 return NVPTXISD::TexUnified3DU32FloatLevel;
2844 case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
2845 return NVPTXISD::TexUnified3DU32FloatGrad;
2847 case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
2848 return NVPTXISD::TexUnifiedCubeFloatFloat;
2849 case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
2850 return NVPTXISD::TexUnifiedCubeFloatFloatLevel;
2851 case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
2852 return NVPTXISD::TexUnifiedCubeS32Float;
2853 case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
2854 return NVPTXISD::TexUnifiedCubeS32FloatLevel;
2855 case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
2856 return NVPTXISD::TexUnifiedCubeU32Float;
2857 case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
2858 return NVPTXISD::TexUnifiedCubeU32FloatLevel;
2860 case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
2861 return NVPTXISD::TexUnifiedCubeArrayFloatFloat;
2862 case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
2863 return NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel;
2864 case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
2865 return NVPTXISD::TexUnifiedCubeArrayS32Float;
2866 case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
2867 return NVPTXISD::TexUnifiedCubeArrayS32FloatLevel;
2868 case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
2869 return NVPTXISD::TexUnifiedCubeArrayU32Float;
2870 case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
2871 return NVPTXISD::TexUnifiedCubeArrayU32FloatLevel;
2873 case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
2874 return NVPTXISD::Tld4UnifiedR2DFloatFloat;
2875 case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
2876 return NVPTXISD::Tld4UnifiedG2DFloatFloat;
2877 case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
2878 return NVPTXISD::Tld4UnifiedB2DFloatFloat;
2879 case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32:
2880 return NVPTXISD::Tld4UnifiedA2DFloatFloat;
2881 case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
2882 return NVPTXISD::Tld4UnifiedR2DS64Float;
2883 case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
2884 return NVPTXISD::Tld4UnifiedG2DS64Float;
2885 case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
2886 return NVPTXISD::Tld4UnifiedB2DS64Float;
2887 case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
2888 return NVPTXISD::Tld4UnifiedA2DS64Float;
2889 case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
2890 return NVPTXISD::Tld4UnifiedR2DU64Float;
2891 case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
2892 return NVPTXISD::Tld4UnifiedG2DU64Float;
2893 case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
2894 return NVPTXISD::Tld4UnifiedB2DU64Float;
2895 case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32:
2896 return NVPTXISD::Tld4UnifiedA2DU64Float;
2900 static unsigned getOpcForSurfaceInstr(unsigned Intrinsic) {
2901 switch (Intrinsic) {
2904 case Intrinsic::nvvm_suld_1d_i8_clamp:
2905 return NVPTXISD::Suld1DI8Clamp;
2906 case Intrinsic::nvvm_suld_1d_i16_clamp:
2907 return NVPTXISD::Suld1DI16Clamp;
2908 case Intrinsic::nvvm_suld_1d_i32_clamp:
2909 return NVPTXISD::Suld1DI32Clamp;
2910 case Intrinsic::nvvm_suld_1d_i64_clamp:
2911 return NVPTXISD::Suld1DI64Clamp;
2912 case Intrinsic::nvvm_suld_1d_v2i8_clamp:
2913 return NVPTXISD::Suld1DV2I8Clamp;
2914 case Intrinsic::nvvm_suld_1d_v2i16_clamp:
2915 return NVPTXISD::Suld1DV2I16Clamp;
2916 case Intrinsic::nvvm_suld_1d_v2i32_clamp:
2917 return NVPTXISD::Suld1DV2I32Clamp;
2918 case Intrinsic::nvvm_suld_1d_v2i64_clamp:
2919 return NVPTXISD::Suld1DV2I64Clamp;
2920 case Intrinsic::nvvm_suld_1d_v4i8_clamp:
2921 return NVPTXISD::Suld1DV4I8Clamp;
2922 case Intrinsic::nvvm_suld_1d_v4i16_clamp:
2923 return NVPTXISD::Suld1DV4I16Clamp;
2924 case Intrinsic::nvvm_suld_1d_v4i32_clamp:
2925 return NVPTXISD::Suld1DV4I32Clamp;
2926 case Intrinsic::nvvm_suld_1d_array_i8_clamp:
2927 return NVPTXISD::Suld1DArrayI8Clamp;
2928 case Intrinsic::nvvm_suld_1d_array_i16_clamp:
2929 return NVPTXISD::Suld1DArrayI16Clamp;
2930 case Intrinsic::nvvm_suld_1d_array_i32_clamp:
2931 return NVPTXISD::Suld1DArrayI32Clamp;
2932 case Intrinsic::nvvm_suld_1d_array_i64_clamp:
2933 return NVPTXISD::Suld1DArrayI64Clamp;
2934 case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
2935 return NVPTXISD::Suld1DArrayV2I8Clamp;
2936 case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
2937 return NVPTXISD::Suld1DArrayV2I16Clamp;
2938 case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
2939 return NVPTXISD::Suld1DArrayV2I32Clamp;
2940 case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
2941 return NVPTXISD::Suld1DArrayV2I64Clamp;
2942 case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
2943 return NVPTXISD::Suld1DArrayV4I8Clamp;
2944 case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
2945 return NVPTXISD::Suld1DArrayV4I16Clamp;
2946 case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
2947 return NVPTXISD::Suld1DArrayV4I32Clamp;
2948 case Intrinsic::nvvm_suld_2d_i8_clamp:
2949 return NVPTXISD::Suld2DI8Clamp;
2950 case Intrinsic::nvvm_suld_2d_i16_clamp:
2951 return NVPTXISD::Suld2DI16Clamp;
2952 case Intrinsic::nvvm_suld_2d_i32_clamp:
2953 return NVPTXISD::Suld2DI32Clamp;
2954 case Intrinsic::nvvm_suld_2d_i64_clamp:
2955 return NVPTXISD::Suld2DI64Clamp;
2956 case Intrinsic::nvvm_suld_2d_v2i8_clamp:
2957 return NVPTXISD::Suld2DV2I8Clamp;
2958 case Intrinsic::nvvm_suld_2d_v2i16_clamp:
2959 return NVPTXISD::Suld2DV2I16Clamp;
2960 case Intrinsic::nvvm_suld_2d_v2i32_clamp:
2961 return NVPTXISD::Suld2DV2I32Clamp;
2962 case Intrinsic::nvvm_suld_2d_v2i64_clamp:
2963 return NVPTXISD::Suld2DV2I64Clamp;
2964 case Intrinsic::nvvm_suld_2d_v4i8_clamp:
2965 return NVPTXISD::Suld2DV4I8Clamp;
2966 case Intrinsic::nvvm_suld_2d_v4i16_clamp:
2967 return NVPTXISD::Suld2DV4I16Clamp;
2968 case Intrinsic::nvvm_suld_2d_v4i32_clamp:
2969 return NVPTXISD::Suld2DV4I32Clamp;
2970 case Intrinsic::nvvm_suld_2d_array_i8_clamp:
2971 return NVPTXISD::Suld2DArrayI8Clamp;
2972 case Intrinsic::nvvm_suld_2d_array_i16_clamp:
2973 return NVPTXISD::Suld2DArrayI16Clamp;
2974 case Intrinsic::nvvm_suld_2d_array_i32_clamp:
2975 return NVPTXISD::Suld2DArrayI32Clamp;
2976 case Intrinsic::nvvm_suld_2d_array_i64_clamp:
2977 return NVPTXISD::Suld2DArrayI64Clamp;
2978 case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
2979 return NVPTXISD::Suld2DArrayV2I8Clamp;
2980 case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
2981 return NVPTXISD::Suld2DArrayV2I16Clamp;
2982 case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
2983 return NVPTXISD::Suld2DArrayV2I32Clamp;
2984 case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
2985 return NVPTXISD::Suld2DArrayV2I64Clamp;
2986 case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
2987 return NVPTXISD::Suld2DArrayV4I8Clamp;
2988 case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
2989 return NVPTXISD::Suld2DArrayV4I16Clamp;
2990 case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
2991 return NVPTXISD::Suld2DArrayV4I32Clamp;
2992 case Intrinsic::nvvm_suld_3d_i8_clamp:
2993 return NVPTXISD::Suld3DI8Clamp;
2994 case Intrinsic::nvvm_suld_3d_i16_clamp:
2995 return NVPTXISD::Suld3DI16Clamp;
2996 case Intrinsic::nvvm_suld_3d_i32_clamp:
2997 return NVPTXISD::Suld3DI32Clamp;
2998 case Intrinsic::nvvm_suld_3d_i64_clamp:
2999 return NVPTXISD::Suld3DI64Clamp;
3000 case Intrinsic::nvvm_suld_3d_v2i8_clamp:
3001 return NVPTXISD::Suld3DV2I8Clamp;
3002 case Intrinsic::nvvm_suld_3d_v2i16_clamp:
3003 return NVPTXISD::Suld3DV2I16Clamp;
3004 case Intrinsic::nvvm_suld_3d_v2i32_clamp:
3005 return NVPTXISD::Suld3DV2I32Clamp;
3006 case Intrinsic::nvvm_suld_3d_v2i64_clamp:
3007 return NVPTXISD::Suld3DV2I64Clamp;
3008 case Intrinsic::nvvm_suld_3d_v4i8_clamp:
3009 return NVPTXISD::Suld3DV4I8Clamp;
3010 case Intrinsic::nvvm_suld_3d_v4i16_clamp:
3011 return NVPTXISD::Suld3DV4I16Clamp;
3012 case Intrinsic::nvvm_suld_3d_v4i32_clamp:
3013 return NVPTXISD::Suld3DV4I32Clamp;
3014 case Intrinsic::nvvm_suld_1d_i8_trap:
3015 return NVPTXISD::Suld1DI8Trap;
3016 case Intrinsic::nvvm_suld_1d_i16_trap:
3017 return NVPTXISD::Suld1DI16Trap;
3018 case Intrinsic::nvvm_suld_1d_i32_trap:
3019 return NVPTXISD::Suld1DI32Trap;
3020 case Intrinsic::nvvm_suld_1d_i64_trap:
3021 return NVPTXISD::Suld1DI64Trap;
3022 case Intrinsic::nvvm_suld_1d_v2i8_trap:
3023 return NVPTXISD::Suld1DV2I8Trap;
3024 case Intrinsic::nvvm_suld_1d_v2i16_trap:
3025 return NVPTXISD::Suld1DV2I16Trap;
3026 case Intrinsic::nvvm_suld_1d_v2i32_trap:
3027 return NVPTXISD::Suld1DV2I32Trap;
3028 case Intrinsic::nvvm_suld_1d_v2i64_trap:
3029 return NVPTXISD::Suld1DV2I64Trap;
3030 case Intrinsic::nvvm_suld_1d_v4i8_trap:
3031 return NVPTXISD::Suld1DV4I8Trap;
3032 case Intrinsic::nvvm_suld_1d_v4i16_trap:
3033 return NVPTXISD::Suld1DV4I16Trap;
3034 case Intrinsic::nvvm_suld_1d_v4i32_trap:
3035 return NVPTXISD::Suld1DV4I32Trap;
3036 case Intrinsic::nvvm_suld_1d_array_i8_trap:
3037 return NVPTXISD::Suld1DArrayI8Trap;
3038 case Intrinsic::nvvm_suld_1d_array_i16_trap:
3039 return NVPTXISD::Suld1DArrayI16Trap;
3040 case Intrinsic::nvvm_suld_1d_array_i32_trap:
3041 return NVPTXISD::Suld1DArrayI32Trap;
3042 case Intrinsic::nvvm_suld_1d_array_i64_trap:
3043 return NVPTXISD::Suld1DArrayI64Trap;
3044 case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3045 return NVPTXISD::Suld1DArrayV2I8Trap;
3046 case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3047 return NVPTXISD::Suld1DArrayV2I16Trap;
3048 case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3049 return NVPTXISD::Suld1DArrayV2I32Trap;
3050 case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3051 return NVPTXISD::Suld1DArrayV2I64Trap;
3052 case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3053 return NVPTXISD::Suld1DArrayV4I8Trap;
3054 case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3055 return NVPTXISD::Suld1DArrayV4I16Trap;
3056 case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3057 return NVPTXISD::Suld1DArrayV4I32Trap;
3058 case Intrinsic::nvvm_suld_2d_i8_trap:
3059 return NVPTXISD::Suld2DI8Trap;
3060 case Intrinsic::nvvm_suld_2d_i16_trap:
3061 return NVPTXISD::Suld2DI16Trap;
3062 case Intrinsic::nvvm_suld_2d_i32_trap:
3063 return NVPTXISD::Suld2DI32Trap;
3064 case Intrinsic::nvvm_suld_2d_i64_trap:
3065 return NVPTXISD::Suld2DI64Trap;
3066 case Intrinsic::nvvm_suld_2d_v2i8_trap:
3067 return NVPTXISD::Suld2DV2I8Trap;
3068 case Intrinsic::nvvm_suld_2d_v2i16_trap:
3069 return NVPTXISD::Suld2DV2I16Trap;
3070 case Intrinsic::nvvm_suld_2d_v2i32_trap:
3071 return NVPTXISD::Suld2DV2I32Trap;
3072 case Intrinsic::nvvm_suld_2d_v2i64_trap:
3073 return NVPTXISD::Suld2DV2I64Trap;
3074 case Intrinsic::nvvm_suld_2d_v4i8_trap:
3075 return NVPTXISD::Suld2DV4I8Trap;
3076 case Intrinsic::nvvm_suld_2d_v4i16_trap:
3077 return NVPTXISD::Suld2DV4I16Trap;
3078 case Intrinsic::nvvm_suld_2d_v4i32_trap:
3079 return NVPTXISD::Suld2DV4I32Trap;
3080 case Intrinsic::nvvm_suld_2d_array_i8_trap:
3081 return NVPTXISD::Suld2DArrayI8Trap;
3082 case Intrinsic::nvvm_suld_2d_array_i16_trap:
3083 return NVPTXISD::Suld2DArrayI16Trap;
3084 case Intrinsic::nvvm_suld_2d_array_i32_trap:
3085 return NVPTXISD::Suld2DArrayI32Trap;
3086 case Intrinsic::nvvm_suld_2d_array_i64_trap:
3087 return NVPTXISD::Suld2DArrayI64Trap;
3088 case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3089 return NVPTXISD::Suld2DArrayV2I8Trap;
3090 case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3091 return NVPTXISD::Suld2DArrayV2I16Trap;
3092 case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3093 return NVPTXISD::Suld2DArrayV2I32Trap;
3094 case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3095 return NVPTXISD::Suld2DArrayV2I64Trap;
3096 case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3097 return NVPTXISD::Suld2DArrayV4I8Trap;
3098 case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3099 return NVPTXISD::Suld2DArrayV4I16Trap;
3100 case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3101 return NVPTXISD::Suld2DArrayV4I32Trap;
3102 case Intrinsic::nvvm_suld_3d_i8_trap:
3103 return NVPTXISD::Suld3DI8Trap;
3104 case Intrinsic::nvvm_suld_3d_i16_trap:
3105 return NVPTXISD::Suld3DI16Trap;
3106 case Intrinsic::nvvm_suld_3d_i32_trap:
3107 return NVPTXISD::Suld3DI32Trap;
3108 case Intrinsic::nvvm_suld_3d_i64_trap:
3109 return NVPTXISD::Suld3DI64Trap;
3110 case Intrinsic::nvvm_suld_3d_v2i8_trap:
3111 return NVPTXISD::Suld3DV2I8Trap;
3112 case Intrinsic::nvvm_suld_3d_v2i16_trap:
3113 return NVPTXISD::Suld3DV2I16Trap;
3114 case Intrinsic::nvvm_suld_3d_v2i32_trap:
3115 return NVPTXISD::Suld3DV2I32Trap;
3116 case Intrinsic::nvvm_suld_3d_v2i64_trap:
3117 return NVPTXISD::Suld3DV2I64Trap;
3118 case Intrinsic::nvvm_suld_3d_v4i8_trap:
3119 return NVPTXISD::Suld3DV4I8Trap;
3120 case Intrinsic::nvvm_suld_3d_v4i16_trap:
3121 return NVPTXISD::Suld3DV4I16Trap;
3122 case Intrinsic::nvvm_suld_3d_v4i32_trap:
3123 return NVPTXISD::Suld3DV4I32Trap;
3124 case Intrinsic::nvvm_suld_1d_i8_zero:
3125 return NVPTXISD::Suld1DI8Zero;
3126 case Intrinsic::nvvm_suld_1d_i16_zero:
3127 return NVPTXISD::Suld1DI16Zero;
3128 case Intrinsic::nvvm_suld_1d_i32_zero:
3129 return NVPTXISD::Suld1DI32Zero;
3130 case Intrinsic::nvvm_suld_1d_i64_zero:
3131 return NVPTXISD::Suld1DI64Zero;
3132 case Intrinsic::nvvm_suld_1d_v2i8_zero:
3133 return NVPTXISD::Suld1DV2I8Zero;
3134 case Intrinsic::nvvm_suld_1d_v2i16_zero:
3135 return NVPTXISD::Suld1DV2I16Zero;
3136 case Intrinsic::nvvm_suld_1d_v2i32_zero:
3137 return NVPTXISD::Suld1DV2I32Zero;
3138 case Intrinsic::nvvm_suld_1d_v2i64_zero:
3139 return NVPTXISD::Suld1DV2I64Zero;
3140 case Intrinsic::nvvm_suld_1d_v4i8_zero:
3141 return NVPTXISD::Suld1DV4I8Zero;
3142 case Intrinsic::nvvm_suld_1d_v4i16_zero:
3143 return NVPTXISD::Suld1DV4I16Zero;
3144 case Intrinsic::nvvm_suld_1d_v4i32_zero:
3145 return NVPTXISD::Suld1DV4I32Zero;
3146 case Intrinsic::nvvm_suld_1d_array_i8_zero:
3147 return NVPTXISD::Suld1DArrayI8Zero;
3148 case Intrinsic::nvvm_suld_1d_array_i16_zero:
3149 return NVPTXISD::Suld1DArrayI16Zero;
3150 case Intrinsic::nvvm_suld_1d_array_i32_zero:
3151 return NVPTXISD::Suld1DArrayI32Zero;
3152 case Intrinsic::nvvm_suld_1d_array_i64_zero:
3153 return NVPTXISD::Suld1DArrayI64Zero;
3154 case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3155 return NVPTXISD::Suld1DArrayV2I8Zero;
3156 case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3157 return NVPTXISD::Suld1DArrayV2I16Zero;
3158 case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3159 return NVPTXISD::Suld1DArrayV2I32Zero;
3160 case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3161 return NVPTXISD::Suld1DArrayV2I64Zero;
3162 case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3163 return NVPTXISD::Suld1DArrayV4I8Zero;
3164 case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3165 return NVPTXISD::Suld1DArrayV4I16Zero;
3166 case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3167 return NVPTXISD::Suld1DArrayV4I32Zero;
3168 case Intrinsic::nvvm_suld_2d_i8_zero:
3169 return NVPTXISD::Suld2DI8Zero;
3170 case Intrinsic::nvvm_suld_2d_i16_zero:
3171 return NVPTXISD::Suld2DI16Zero;
3172 case Intrinsic::nvvm_suld_2d_i32_zero:
3173 return NVPTXISD::Suld2DI32Zero;
3174 case Intrinsic::nvvm_suld_2d_i64_zero:
3175 return NVPTXISD::Suld2DI64Zero;
3176 case Intrinsic::nvvm_suld_2d_v2i8_zero:
3177 return NVPTXISD::Suld2DV2I8Zero;
3178 case Intrinsic::nvvm_suld_2d_v2i16_zero:
3179 return NVPTXISD::Suld2DV2I16Zero;
3180 case Intrinsic::nvvm_suld_2d_v2i32_zero:
3181 return NVPTXISD::Suld2DV2I32Zero;
3182 case Intrinsic::nvvm_suld_2d_v2i64_zero:
3183 return NVPTXISD::Suld2DV2I64Zero;
3184 case Intrinsic::nvvm_suld_2d_v4i8_zero:
3185 return NVPTXISD::Suld2DV4I8Zero;
3186 case Intrinsic::nvvm_suld_2d_v4i16_zero:
3187 return NVPTXISD::Suld2DV4I16Zero;
3188 case Intrinsic::nvvm_suld_2d_v4i32_zero:
3189 return NVPTXISD::Suld2DV4I32Zero;
3190 case Intrinsic::nvvm_suld_2d_array_i8_zero:
3191 return NVPTXISD::Suld2DArrayI8Zero;
3192 case Intrinsic::nvvm_suld_2d_array_i16_zero:
3193 return NVPTXISD::Suld2DArrayI16Zero;
3194 case Intrinsic::nvvm_suld_2d_array_i32_zero:
3195 return NVPTXISD::Suld2DArrayI32Zero;
3196 case Intrinsic::nvvm_suld_2d_array_i64_zero:
3197 return NVPTXISD::Suld2DArrayI64Zero;
3198 case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3199 return NVPTXISD::Suld2DArrayV2I8Zero;
3200 case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3201 return NVPTXISD::Suld2DArrayV2I16Zero;
3202 case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3203 return NVPTXISD::Suld2DArrayV2I32Zero;
3204 case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3205 return NVPTXISD::Suld2DArrayV2I64Zero;
3206 case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3207 return NVPTXISD::Suld2DArrayV4I8Zero;
3208 case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3209 return NVPTXISD::Suld2DArrayV4I16Zero;
3210 case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3211 return NVPTXISD::Suld2DArrayV4I32Zero;
3212 case Intrinsic::nvvm_suld_3d_i8_zero:
3213 return NVPTXISD::Suld3DI8Zero;
3214 case Intrinsic::nvvm_suld_3d_i16_zero:
3215 return NVPTXISD::Suld3DI16Zero;
3216 case Intrinsic::nvvm_suld_3d_i32_zero:
3217 return NVPTXISD::Suld3DI32Zero;
3218 case Intrinsic::nvvm_suld_3d_i64_zero:
3219 return NVPTXISD::Suld3DI64Zero;
3220 case Intrinsic::nvvm_suld_3d_v2i8_zero:
3221 return NVPTXISD::Suld3DV2I8Zero;
3222 case Intrinsic::nvvm_suld_3d_v2i16_zero:
3223 return NVPTXISD::Suld3DV2I16Zero;
3224 case Intrinsic::nvvm_suld_3d_v2i32_zero:
3225 return NVPTXISD::Suld3DV2I32Zero;
3226 case Intrinsic::nvvm_suld_3d_v2i64_zero:
3227 return NVPTXISD::Suld3DV2I64Zero;
3228 case Intrinsic::nvvm_suld_3d_v4i8_zero:
3229 return NVPTXISD::Suld3DV4I8Zero;
3230 case Intrinsic::nvvm_suld_3d_v4i16_zero:
3231 return NVPTXISD::Suld3DV4I16Zero;
3232 case Intrinsic::nvvm_suld_3d_v4i32_zero:
3233 return NVPTXISD::Suld3DV4I32Zero;
3237 // llvm.ptx.memcpy.const and llvm.ptx.memmove.const need to be modeled as
3239 // because we need the information that is only available in the "Value" type
3241 // pointer. In particular, the address space information.
3242 bool NVPTXTargetLowering::getTgtMemIntrinsic(
3243 IntrinsicInfo &Info, const CallInst &I, unsigned Intrinsic) const {
3244 switch (Intrinsic) {
3248 case Intrinsic::nvvm_atomic_load_add_f32:
3249 Info.opc = ISD::INTRINSIC_W_CHAIN;
3250 Info.memVT = MVT::f32;
3251 Info.ptrVal = I.getArgOperand(0);
3254 Info.readMem = true;
3255 Info.writeMem = true;
3259 case Intrinsic::nvvm_atomic_load_inc_32:
3260 case Intrinsic::nvvm_atomic_load_dec_32:
3261 Info.opc = ISD::INTRINSIC_W_CHAIN;
3262 Info.memVT = MVT::i32;
3263 Info.ptrVal = I.getArgOperand(0);
3266 Info.readMem = true;
3267 Info.writeMem = true;
3271 case Intrinsic::nvvm_ldu_global_i:
3272 case Intrinsic::nvvm_ldu_global_f:
3273 case Intrinsic::nvvm_ldu_global_p: {
3274 auto &DL = I.getModule()->getDataLayout();
3275 Info.opc = ISD::INTRINSIC_W_CHAIN;
3276 if (Intrinsic == Intrinsic::nvvm_ldu_global_i)
3277 Info.memVT = getValueType(DL, I.getType());
3278 else if(Intrinsic == Intrinsic::nvvm_ldu_global_p)
3279 Info.memVT = getPointerTy(DL);
3281 Info.memVT = getValueType(DL, I.getType());
3282 Info.ptrVal = I.getArgOperand(0);
3285 Info.readMem = true;
3286 Info.writeMem = false;
3287 Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
3291 case Intrinsic::nvvm_ldg_global_i:
3292 case Intrinsic::nvvm_ldg_global_f:
3293 case Intrinsic::nvvm_ldg_global_p: {
3294 auto &DL = I.getModule()->getDataLayout();
3296 Info.opc = ISD::INTRINSIC_W_CHAIN;
3297 if (Intrinsic == Intrinsic::nvvm_ldg_global_i)
3298 Info.memVT = getValueType(DL, I.getType());
3299 else if(Intrinsic == Intrinsic::nvvm_ldg_global_p)
3300 Info.memVT = getPointerTy(DL);
3302 Info.memVT = getValueType(DL, I.getType());
3303 Info.ptrVal = I.getArgOperand(0);
3306 Info.readMem = true;
3307 Info.writeMem = false;
3308 Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
3313 case Intrinsic::nvvm_tex_1d_v4f32_s32:
3314 case Intrinsic::nvvm_tex_1d_v4f32_f32:
3315 case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
3316 case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
3317 case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
3318 case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
3319 case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
3320 case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
3321 case Intrinsic::nvvm_tex_2d_v4f32_s32:
3322 case Intrinsic::nvvm_tex_2d_v4f32_f32:
3323 case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
3324 case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
3325 case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
3326 case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
3327 case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
3328 case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
3329 case Intrinsic::nvvm_tex_3d_v4f32_s32:
3330 case Intrinsic::nvvm_tex_3d_v4f32_f32:
3331 case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
3332 case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
3333 case Intrinsic::nvvm_tex_cube_v4f32_f32:
3334 case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
3335 case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
3336 case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
3337 case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
3338 case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
3339 case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
3340 case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
3341 case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
3342 case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
3343 case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
3344 case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
3345 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
3346 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
3347 case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
3348 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
3349 case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
3350 case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
3351 case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
3352 case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
3353 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
3354 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
3355 case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
3356 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
3357 case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
3358 case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
3359 case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
3360 case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
3361 case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
3362 case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
3363 case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
3364 case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
3365 case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
3366 case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
3367 case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
3368 case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32: {
3369 Info.opc = getOpcForTextureInstr(Intrinsic);
3370 Info.memVT = MVT::v4f32;
3371 Info.ptrVal = nullptr;
3374 Info.readMem = true;
3375 Info.writeMem = false;
3379 case Intrinsic::nvvm_tex_1d_v4s32_s32:
3380 case Intrinsic::nvvm_tex_1d_v4s32_f32:
3381 case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
3382 case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
3383 case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
3384 case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
3385 case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
3386 case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
3387 case Intrinsic::nvvm_tex_2d_v4s32_s32:
3388 case Intrinsic::nvvm_tex_2d_v4s32_f32:
3389 case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
3390 case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
3391 case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
3392 case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
3393 case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
3394 case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
3395 case Intrinsic::nvvm_tex_3d_v4s32_s32:
3396 case Intrinsic::nvvm_tex_3d_v4s32_f32:
3397 case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
3398 case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
3399 case Intrinsic::nvvm_tex_cube_v4s32_f32:
3400 case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
3401 case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
3402 case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
3403 case Intrinsic::nvvm_tex_cube_v4u32_f32:
3404 case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
3405 case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
3406 case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
3407 case Intrinsic::nvvm_tex_1d_v4u32_s32:
3408 case Intrinsic::nvvm_tex_1d_v4u32_f32:
3409 case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
3410 case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
3411 case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
3412 case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
3413 case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
3414 case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
3415 case Intrinsic::nvvm_tex_2d_v4u32_s32:
3416 case Intrinsic::nvvm_tex_2d_v4u32_f32:
3417 case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
3418 case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
3419 case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
3420 case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
3421 case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
3422 case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
3423 case Intrinsic::nvvm_tex_3d_v4u32_s32:
3424 case Intrinsic::nvvm_tex_3d_v4u32_f32:
3425 case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
3426 case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
3427 case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
3428 case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
3429 case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
3430 case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
3431 case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
3432 case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
3433 case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
3434 case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
3435 case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
3436 case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
3437 case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
3438 case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
3439 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
3440 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
3441 case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
3442 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
3443 case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
3444 case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
3445 case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
3446 case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
3447 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
3448 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
3449 case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
3450 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
3451 case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
3452 case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
3453 case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
3454 case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
3455 case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
3456 case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
3457 case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
3458 case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
3459 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
3460 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
3461 case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
3462 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
3463 case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
3464 case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
3465 case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
3466 case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
3467 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
3468 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
3469 case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
3470 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
3471 case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
3472 case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
3473 case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
3474 case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
3475 case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
3476 case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
3477 case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
3478 case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
3479 case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
3480 case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
3481 case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
3482 case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
3483 case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
3484 case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
3485 case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
3486 case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
3487 case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
3488 case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
3489 case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
3490 case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32: {
3491 Info.opc = getOpcForTextureInstr(Intrinsic);
3492 Info.memVT = MVT::v4i32;
3493 Info.ptrVal = nullptr;
3496 Info.readMem = true;
3497 Info.writeMem = false;
3501 case Intrinsic::nvvm_suld_1d_i8_clamp:
3502 case Intrinsic::nvvm_suld_1d_v2i8_clamp:
3503 case Intrinsic::nvvm_suld_1d_v4i8_clamp:
3504 case Intrinsic::nvvm_suld_1d_array_i8_clamp:
3505 case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
3506 case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
3507 case Intrinsic::nvvm_suld_2d_i8_clamp:
3508 case Intrinsic::nvvm_suld_2d_v2i8_clamp:
3509 case Intrinsic::nvvm_suld_2d_v4i8_clamp:
3510 case Intrinsic::nvvm_suld_2d_array_i8_clamp:
3511 case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
3512 case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
3513 case Intrinsic::nvvm_suld_3d_i8_clamp:
3514 case Intrinsic::nvvm_suld_3d_v2i8_clamp:
3515 case Intrinsic::nvvm_suld_3d_v4i8_clamp:
3516 case Intrinsic::nvvm_suld_1d_i8_trap:
3517 case Intrinsic::nvvm_suld_1d_v2i8_trap:
3518 case Intrinsic::nvvm_suld_1d_v4i8_trap:
3519 case Intrinsic::nvvm_suld_1d_array_i8_trap:
3520 case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3521 case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3522 case Intrinsic::nvvm_suld_2d_i8_trap:
3523 case Intrinsic::nvvm_suld_2d_v2i8_trap:
3524 case Intrinsic::nvvm_suld_2d_v4i8_trap:
3525 case Intrinsic::nvvm_suld_2d_array_i8_trap:
3526 case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3527 case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3528 case Intrinsic::nvvm_suld_3d_i8_trap:
3529 case Intrinsic::nvvm_suld_3d_v2i8_trap:
3530 case Intrinsic::nvvm_suld_3d_v4i8_trap:
3531 case Intrinsic::nvvm_suld_1d_i8_zero:
3532 case Intrinsic::nvvm_suld_1d_v2i8_zero:
3533 case Intrinsic::nvvm_suld_1d_v4i8_zero:
3534 case Intrinsic::nvvm_suld_1d_array_i8_zero:
3535 case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3536 case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3537 case Intrinsic::nvvm_suld_2d_i8_zero:
3538 case Intrinsic::nvvm_suld_2d_v2i8_zero:
3539 case Intrinsic::nvvm_suld_2d_v4i8_zero:
3540 case Intrinsic::nvvm_suld_2d_array_i8_zero:
3541 case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3542 case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3543 case Intrinsic::nvvm_suld_3d_i8_zero:
3544 case Intrinsic::nvvm_suld_3d_v2i8_zero:
3545 case Intrinsic::nvvm_suld_3d_v4i8_zero: {
3546 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3547 Info.memVT = MVT::i8;
3548 Info.ptrVal = nullptr;
3551 Info.readMem = true;
3552 Info.writeMem = false;
3556 case Intrinsic::nvvm_suld_1d_i16_clamp:
3557 case Intrinsic::nvvm_suld_1d_v2i16_clamp:
3558 case Intrinsic::nvvm_suld_1d_v4i16_clamp:
3559 case Intrinsic::nvvm_suld_1d_array_i16_clamp:
3560 case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
3561 case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
3562 case Intrinsic::nvvm_suld_2d_i16_clamp:
3563 case Intrinsic::nvvm_suld_2d_v2i16_clamp:
3564 case Intrinsic::nvvm_suld_2d_v4i16_clamp:
3565 case Intrinsic::nvvm_suld_2d_array_i16_clamp:
3566 case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
3567 case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
3568 case Intrinsic::nvvm_suld_3d_i16_clamp:
3569 case Intrinsic::nvvm_suld_3d_v2i16_clamp:
3570 case Intrinsic::nvvm_suld_3d_v4i16_clamp:
3571 case Intrinsic::nvvm_suld_1d_i16_trap:
3572 case Intrinsic::nvvm_suld_1d_v2i16_trap:
3573 case Intrinsic::nvvm_suld_1d_v4i16_trap:
3574 case Intrinsic::nvvm_suld_1d_array_i16_trap:
3575 case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3576 case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3577 case Intrinsic::nvvm_suld_2d_i16_trap:
3578 case Intrinsic::nvvm_suld_2d_v2i16_trap:
3579 case Intrinsic::nvvm_suld_2d_v4i16_trap:
3580 case Intrinsic::nvvm_suld_2d_array_i16_trap:
3581 case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3582 case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3583 case Intrinsic::nvvm_suld_3d_i16_trap:
3584 case Intrinsic::nvvm_suld_3d_v2i16_trap:
3585 case Intrinsic::nvvm_suld_3d_v4i16_trap:
3586 case Intrinsic::nvvm_suld_1d_i16_zero:
3587 case Intrinsic::nvvm_suld_1d_v2i16_zero:
3588 case Intrinsic::nvvm_suld_1d_v4i16_zero:
3589 case Intrinsic::nvvm_suld_1d_array_i16_zero:
3590 case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3591 case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3592 case Intrinsic::nvvm_suld_2d_i16_zero:
3593 case Intrinsic::nvvm_suld_2d_v2i16_zero:
3594 case Intrinsic::nvvm_suld_2d_v4i16_zero:
3595 case Intrinsic::nvvm_suld_2d_array_i16_zero:
3596 case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3597 case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3598 case Intrinsic::nvvm_suld_3d_i16_zero:
3599 case Intrinsic::nvvm_suld_3d_v2i16_zero:
3600 case Intrinsic::nvvm_suld_3d_v4i16_zero: {
3601 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3602 Info.memVT = MVT::i16;
3603 Info.ptrVal = nullptr;
3606 Info.readMem = true;
3607 Info.writeMem = false;
3611 case Intrinsic::nvvm_suld_1d_i32_clamp:
3612 case Intrinsic::nvvm_suld_1d_v2i32_clamp:
3613 case Intrinsic::nvvm_suld_1d_v4i32_clamp:
3614 case Intrinsic::nvvm_suld_1d_array_i32_clamp:
3615 case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
3616 case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
3617 case Intrinsic::nvvm_suld_2d_i32_clamp:
3618 case Intrinsic::nvvm_suld_2d_v2i32_clamp:
3619 case Intrinsic::nvvm_suld_2d_v4i32_clamp:
3620 case Intrinsic::nvvm_suld_2d_array_i32_clamp:
3621 case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
3622 case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
3623 case Intrinsic::nvvm_suld_3d_i32_clamp:
3624 case Intrinsic::nvvm_suld_3d_v2i32_clamp:
3625 case Intrinsic::nvvm_suld_3d_v4i32_clamp:
3626 case Intrinsic::nvvm_suld_1d_i32_trap:
3627 case Intrinsic::nvvm_suld_1d_v2i32_trap:
3628 case Intrinsic::nvvm_suld_1d_v4i32_trap:
3629 case Intrinsic::nvvm_suld_1d_array_i32_trap:
3630 case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3631 case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3632 case Intrinsic::nvvm_suld_2d_i32_trap:
3633 case Intrinsic::nvvm_suld_2d_v2i32_trap:
3634 case Intrinsic::nvvm_suld_2d_v4i32_trap:
3635 case Intrinsic::nvvm_suld_2d_array_i32_trap:
3636 case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3637 case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3638 case Intrinsic::nvvm_suld_3d_i32_trap:
3639 case Intrinsic::nvvm_suld_3d_v2i32_trap:
3640 case Intrinsic::nvvm_suld_3d_v4i32_trap:
3641 case Intrinsic::nvvm_suld_1d_i32_zero:
3642 case Intrinsic::nvvm_suld_1d_v2i32_zero:
3643 case Intrinsic::nvvm_suld_1d_v4i32_zero:
3644 case Intrinsic::nvvm_suld_1d_array_i32_zero:
3645 case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3646 case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3647 case Intrinsic::nvvm_suld_2d_i32_zero:
3648 case Intrinsic::nvvm_suld_2d_v2i32_zero:
3649 case Intrinsic::nvvm_suld_2d_v4i32_zero:
3650 case Intrinsic::nvvm_suld_2d_array_i32_zero:
3651 case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3652 case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3653 case Intrinsic::nvvm_suld_3d_i32_zero:
3654 case Intrinsic::nvvm_suld_3d_v2i32_zero:
3655 case Intrinsic::nvvm_suld_3d_v4i32_zero: {
3656 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3657 Info.memVT = MVT::i32;
3658 Info.ptrVal = nullptr;
3661 Info.readMem = true;
3662 Info.writeMem = false;
3666 case Intrinsic::nvvm_suld_1d_i64_clamp:
3667 case Intrinsic::nvvm_suld_1d_v2i64_clamp:
3668 case Intrinsic::nvvm_suld_1d_array_i64_clamp:
3669 case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
3670 case Intrinsic::nvvm_suld_2d_i64_clamp:
3671 case Intrinsic::nvvm_suld_2d_v2i64_clamp:
3672 case Intrinsic::nvvm_suld_2d_array_i64_clamp:
3673 case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
3674 case Intrinsic::nvvm_suld_3d_i64_clamp:
3675 case Intrinsic::nvvm_suld_3d_v2i64_clamp:
3676 case Intrinsic::nvvm_suld_1d_i64_trap:
3677 case Intrinsic::nvvm_suld_1d_v2i64_trap:
3678 case Intrinsic::nvvm_suld_1d_array_i64_trap:
3679 case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3680 case Intrinsic::nvvm_suld_2d_i64_trap:
3681 case Intrinsic::nvvm_suld_2d_v2i64_trap:
3682 case Intrinsic::nvvm_suld_2d_array_i64_trap:
3683 case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3684 case Intrinsic::nvvm_suld_3d_i64_trap:
3685 case Intrinsic::nvvm_suld_3d_v2i64_trap:
3686 case Intrinsic::nvvm_suld_1d_i64_zero:
3687 case Intrinsic::nvvm_suld_1d_v2i64_zero:
3688 case Intrinsic::nvvm_suld_1d_array_i64_zero:
3689 case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3690 case Intrinsic::nvvm_suld_2d_i64_zero:
3691 case Intrinsic::nvvm_suld_2d_v2i64_zero:
3692 case Intrinsic::nvvm_suld_2d_array_i64_zero:
3693 case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3694 case Intrinsic::nvvm_suld_3d_i64_zero:
3695 case Intrinsic::nvvm_suld_3d_v2i64_zero: {
3696 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3697 Info.memVT = MVT::i64;
3698 Info.ptrVal = nullptr;
3701 Info.readMem = true;
3702 Info.writeMem = false;
3710 /// isLegalAddressingMode - Return true if the addressing mode represented
3711 /// by AM is legal for this target, for a load/store of the specified type.
3712 /// Used to guide target specific optimizations, like loop strength reduction
3713 /// (LoopStrengthReduce.cpp) and memory optimization for address mode
3714 /// (CodeGenPrepare.cpp)
3715 bool NVPTXTargetLowering::isLegalAddressingMode(const DataLayout &DL,
3716 const AddrMode &AM, Type *Ty,
3717 unsigned AS) const {
3719 // AddrMode - This represents an addressing mode of:
3720 // BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
3722 // The legal address modes are
3729 return !AM.BaseOffs && !AM.HasBaseReg && !AM.Scale;
3733 case 0: // "r", "r+i" or "i" is allowed
3736 if (AM.HasBaseReg) // "r+r+i" or "r+r" is not allowed.
3738 // Otherwise we have r+i.
3741 // No scale > 1 is allowed
3747 //===----------------------------------------------------------------------===//
3748 // NVPTX Inline Assembly Support
3749 //===----------------------------------------------------------------------===//
3751 /// getConstraintType - Given a constraint letter, return the type of
3752 /// constraint it is for this target.
3753 NVPTXTargetLowering::ConstraintType
3754 NVPTXTargetLowering::getConstraintType(StringRef Constraint) const {
3755 if (Constraint.size() == 1) {
3756 switch (Constraint[0]) {
3768 return C_RegisterClass;
3771 return TargetLowering::getConstraintType(Constraint);
3774 std::pair<unsigned, const TargetRegisterClass *>
3775 NVPTXTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
3776 StringRef Constraint,
3778 if (Constraint.size() == 1) {
3779 switch (Constraint[0]) {
3781 return std::make_pair(0U, &NVPTX::Int1RegsRegClass);
3783 return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
3785 return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
3787 return std::make_pair(0U, &NVPTX::Int32RegsRegClass);
3790 return std::make_pair(0U, &NVPTX::Int64RegsRegClass);
3792 return std::make_pair(0U, &NVPTX::Float32RegsRegClass);
3794 return std::make_pair(0U, &NVPTX::Float64RegsRegClass);
3797 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
3800 //===----------------------------------------------------------------------===//
3801 // NVPTX DAG Combining
3802 //===----------------------------------------------------------------------===//
3804 bool NVPTXTargetLowering::allowFMA(MachineFunction &MF,
3805 CodeGenOpt::Level OptLevel) const {
3806 const Function *F = MF.getFunction();
3807 const TargetOptions &TO = MF.getTarget().Options;
3809 // Always honor command-line argument
3810 if (FMAContractLevelOpt.getNumOccurrences() > 0) {
3811 return FMAContractLevelOpt > 0;
3812 } else if (OptLevel == 0) {
3813 // Do not contract if we're not optimizing the code
3815 } else if (TO.AllowFPOpFusion == FPOpFusion::Fast || TO.UnsafeFPMath) {
3816 // Honor TargetOptions flags that explicitly say fusion is okay
3818 } else if (F->hasFnAttribute("unsafe-fp-math")) {
3819 // Check for unsafe-fp-math=true coming from Clang
3820 Attribute Attr = F->getFnAttribute("unsafe-fp-math");
3821 StringRef Val = Attr.getValueAsString();
3826 // We did not have a clear indication that fusion is allowed, so assume not
3830 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
3831 /// operands N0 and N1. This is a helper for PerformADDCombine that is
3832 /// called with the default operands, and if that fails, with commuted
3834 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
3835 TargetLowering::DAGCombinerInfo &DCI,
3836 const NVPTXSubtarget &Subtarget,
3837 CodeGenOpt::Level OptLevel) {
3838 SelectionDAG &DAG = DCI.DAG;
3839 // Skip non-integer, non-scalar case
3840 EVT VT=N0.getValueType();
3844 // fold (add (mul a, b), c) -> (mad a, b, c)
3846 if (N0.getOpcode() == ISD::MUL) {
3847 assert (VT.isInteger());
3849 // Since integer multiply-add costs the same as integer multiply
3850 // but is more costly than integer add, do the fusion only when
3851 // the mul is only used in the add.
3852 if (OptLevel==CodeGenOpt::None || VT != MVT::i32 ||
3853 !N0.getNode()->hasOneUse())
3857 return DAG.getNode(NVPTXISD::IMAD, SDLoc(N), VT,
3858 N0.getOperand(0), N0.getOperand(1), N1);
3860 else if (N0.getOpcode() == ISD::FMUL) {
3861 if (VT == MVT::f32 || VT == MVT::f64) {
3862 const auto *TLI = static_cast<const NVPTXTargetLowering *>(
3863 &DAG.getTargetLoweringInfo());
3864 if (!TLI->allowFMA(DAG.getMachineFunction(), OptLevel))
3867 // For floating point:
3868 // Do the fusion only when the mul has less than 5 uses and all
3870 // The heuristic is that if a use is not an add, then that use
3871 // cannot be fused into fma, therefore mul is still needed anyway.
3872 // If there are more than 4 uses, even if they are all add, fusing
3873 // them will increase register pressue.
3876 int nonAddCount = 0;
3877 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
3878 UE = N0.getNode()->use_end();
3882 if (User->getOpcode() != ISD::FADD)
3888 int orderNo = N->getIROrder();
3889 int orderNo2 = N0.getNode()->getIROrder();
3890 // simple heuristics here for considering potential register
3891 // pressure, the logics here is that the differnce are used
3892 // to measure the distance between def and use, the longer distance
3893 // more likely cause register pressure.
3894 if (orderNo - orderNo2 < 500)
3897 // Now, check if at least one of the FMUL's operands is live beyond the node N,
3898 // which guarantees that the FMA will not increase register pressure at node N.
3899 bool opIsLive = false;
3900 const SDNode *left = N0.getOperand(0).getNode();
3901 const SDNode *right = N0.getOperand(1).getNode();
3903 if (isa<ConstantSDNode>(left) || isa<ConstantSDNode>(right))
3907 for (SDNode::use_iterator UI = left->use_begin(), UE = left->use_end(); UI != UE; ++UI) {
3909 int orderNo3 = User->getIROrder();
3910 if (orderNo3 > orderNo) {
3917 for (SDNode::use_iterator UI = right->use_begin(), UE = right->use_end(); UI != UE; ++UI) {
3919 int orderNo3 = User->getIROrder();
3920 if (orderNo3 > orderNo) {
3930 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
3931 N0.getOperand(0), N0.getOperand(1), N1);
3938 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3940 static SDValue PerformADDCombine(SDNode *N,
3941 TargetLowering::DAGCombinerInfo &DCI,
3942 const NVPTXSubtarget &Subtarget,
3943 CodeGenOpt::Level OptLevel) {
3944 SDValue N0 = N->getOperand(0);
3945 SDValue N1 = N->getOperand(1);
3947 // First try with the default operand order.
3948 if (SDValue Result =
3949 PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget, OptLevel))
3952 // If that didn't work, try again with the operands commuted.
3953 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget, OptLevel);
3956 static SDValue PerformANDCombine(SDNode *N,
3957 TargetLowering::DAGCombinerInfo &DCI) {
3958 // The type legalizer turns a vector load of i8 values into a zextload to i16
3959 // registers, optionally ANY_EXTENDs it (if target type is integer),
3960 // and ANDs off the high 8 bits. Since we turn this load into a
3961 // target-specific DAG node, the DAG combiner fails to eliminate these AND
3962 // nodes. Do that here.
3963 SDValue Val = N->getOperand(0);
3964 SDValue Mask = N->getOperand(1);
3966 if (isa<ConstantSDNode>(Val)) {
3967 std::swap(Val, Mask);
3971 // Generally, we will see zextload -> IMOV16rr -> ANY_EXTEND -> and
3972 if (Val.getOpcode() == ISD::ANY_EXTEND) {
3974 Val = Val->getOperand(0);
3977 if (Val->isMachineOpcode() && Val->getMachineOpcode() == NVPTX::IMOV16rr) {
3978 Val = Val->getOperand(0);
3981 if (Val->getOpcode() == NVPTXISD::LoadV2 ||
3982 Val->getOpcode() == NVPTXISD::LoadV4) {
3983 ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(Mask);
3985 // Not an AND with a constant
3989 uint64_t MaskVal = MaskCnst->getZExtValue();
3990 if (MaskVal != 0xff) {
3991 // Not an AND that chops off top 8 bits
3995 MemSDNode *Mem = dyn_cast<MemSDNode>(Val);
3997 // Not a MemSDNode?!?
4001 EVT MemVT = Mem->getMemoryVT();
4002 if (MemVT != MVT::v2i8 && MemVT != MVT::v4i8) {
4003 // We only handle the i8 case
4008 cast<ConstantSDNode>(Val->getOperand(Val->getNumOperands()-1))->
4010 if (ExtType == ISD::SEXTLOAD) {
4011 // If for some reason the load is a sextload, the and is needed to zero
4012 // out the high 8 bits
4017 if (AExt.getNode() != 0) {
4018 // Re-insert the ext as a zext.
4019 Val = DCI.DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
4020 AExt.getValueType(), Val);
4024 // If we get here, the AND is unnecessary. Just replace it with the load
4025 DCI.CombineTo(N, Val, AddTo);
4031 static SDValue PerformSELECTCombine(SDNode *N,
4032 TargetLowering::DAGCombinerInfo &DCI) {
4033 // Currently this detects patterns for integer min and max and
4034 // lowers them to PTX-specific intrinsics that enable hardware
4037 const SDValue Cond = N->getOperand(0);
4038 if (Cond.getOpcode() != ISD::SETCC) return SDValue();
4040 const SDValue LHS = Cond.getOperand(0);
4041 const SDValue RHS = Cond.getOperand(1);
4042 const SDValue True = N->getOperand(1);
4043 const SDValue False = N->getOperand(2);
4044 if (!(LHS == True && RHS == False) && !(LHS == False && RHS == True))
4047 const EVT VT = N->getValueType(0);
4048 if (VT != MVT::i32 && VT != MVT::i64) return SDValue();
4050 const ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
4051 SDValue Larger; // The larger of LHS and RHS when condition is true.
4070 const bool IsMax = (Larger == True);
4071 const bool IsSigned = ISD::isSignedIntSetCC(CC);
4073 unsigned IntrinsicId;
4074 if (VT == MVT::i32) {
4076 IntrinsicId = IsMax ? Intrinsic::nvvm_max_i : Intrinsic::nvvm_min_i;
4078 IntrinsicId = IsMax ? Intrinsic::nvvm_max_ui : Intrinsic::nvvm_min_ui;
4080 assert(VT == MVT::i64);
4082 IntrinsicId = IsMax ? Intrinsic::nvvm_max_ll : Intrinsic::nvvm_min_ll;
4084 IntrinsicId = IsMax ? Intrinsic::nvvm_max_ull : Intrinsic::nvvm_min_ull;
4088 return DCI.DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
4089 DCI.DAG.getConstant(IntrinsicId, DL, VT), LHS, RHS);
4092 enum OperandSignedness {
4098 /// IsMulWideOperandDemotable - Checks if the provided DAG node is an operand
4099 /// that can be demoted to \p OptSize bits without loss of information. The
4100 /// signedness of the operand, if determinable, is placed in \p S.
4101 static bool IsMulWideOperandDemotable(SDValue Op,
4103 OperandSignedness &S) {
4106 if (Op.getOpcode() == ISD::SIGN_EXTEND ||
4107 Op.getOpcode() == ISD::SIGN_EXTEND_INREG) {
4108 EVT OrigVT = Op.getOperand(0).getValueType();
4109 if (OrigVT.getSizeInBits() <= OptSize) {
4113 } else if (Op.getOpcode() == ISD::ZERO_EXTEND) {
4114 EVT OrigVT = Op.getOperand(0).getValueType();
4115 if (OrigVT.getSizeInBits() <= OptSize) {
4124 /// AreMulWideOperandsDemotable - Checks if the given LHS and RHS operands can
4125 /// be demoted to \p OptSize bits without loss of information. If the operands
4126 /// contain a constant, it should appear as the RHS operand. The signedness of
4127 /// the operands is placed in \p IsSigned.
4128 static bool AreMulWideOperandsDemotable(SDValue LHS, SDValue RHS,
4132 OperandSignedness LHSSign;
4134 // The LHS operand must be a demotable op
4135 if (!IsMulWideOperandDemotable(LHS, OptSize, LHSSign))
4138 // We should have been able to determine the signedness from the LHS
4139 if (LHSSign == Unknown)
4142 IsSigned = (LHSSign == Signed);
4144 // The RHS can be a demotable op or a constant
4145 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(RHS)) {
4146 const APInt &Val = CI->getAPIntValue();
4147 if (LHSSign == Unsigned) {
4148 return Val.isIntN(OptSize);
4150 return Val.isSignedIntN(OptSize);
4153 OperandSignedness RHSSign;
4154 if (!IsMulWideOperandDemotable(RHS, OptSize, RHSSign))
4157 return LHSSign == RHSSign;
4161 /// TryMULWIDECombine - Attempt to replace a multiply of M bits with a multiply
4162 /// of M/2 bits that produces an M-bit result (i.e. mul.wide). This transform
4163 /// works on both multiply DAG nodes and SHL DAG nodes with a constant shift
4165 static SDValue TryMULWIDECombine(SDNode *N,
4166 TargetLowering::DAGCombinerInfo &DCI) {
4167 EVT MulType = N->getValueType(0);
4168 if (MulType != MVT::i32 && MulType != MVT::i64) {
4173 unsigned OptSize = MulType.getSizeInBits() >> 1;
4174 SDValue LHS = N->getOperand(0);
4175 SDValue RHS = N->getOperand(1);
4177 // Canonicalize the multiply so the constant (if any) is on the right
4178 if (N->getOpcode() == ISD::MUL) {
4179 if (isa<ConstantSDNode>(LHS)) {
4180 std::swap(LHS, RHS);
4184 // If we have a SHL, determine the actual multiply amount
4185 if (N->getOpcode() == ISD::SHL) {
4186 ConstantSDNode *ShlRHS = dyn_cast<ConstantSDNode>(RHS);
4191 APInt ShiftAmt = ShlRHS->getAPIntValue();
4192 unsigned BitWidth = MulType.getSizeInBits();
4193 if (ShiftAmt.sge(0) && ShiftAmt.slt(BitWidth)) {
4194 APInt MulVal = APInt(BitWidth, 1) << ShiftAmt;
4195 RHS = DCI.DAG.getConstant(MulVal, DL, MulType);
4202 // Verify that our operands are demotable
4203 if (!AreMulWideOperandsDemotable(LHS, RHS, OptSize, Signed)) {
4208 if (MulType == MVT::i32) {
4209 DemotedVT = MVT::i16;
4211 DemotedVT = MVT::i32;
4214 // Truncate the operands to the correct size. Note that these are just for
4215 // type consistency and will (likely) be eliminated in later phases.
4217 DCI.DAG.getNode(ISD::TRUNCATE, DL, DemotedVT, LHS);
4219 DCI.DAG.getNode(ISD::TRUNCATE, DL, DemotedVT, RHS);
4223 Opc = NVPTXISD::MUL_WIDE_SIGNED;
4225 Opc = NVPTXISD::MUL_WIDE_UNSIGNED;
4228 return DCI.DAG.getNode(Opc, DL, MulType, TruncLHS, TruncRHS);
4231 /// PerformMULCombine - Runs PTX-specific DAG combine patterns on MUL nodes.
4232 static SDValue PerformMULCombine(SDNode *N,
4233 TargetLowering::DAGCombinerInfo &DCI,
4234 CodeGenOpt::Level OptLevel) {
4236 // Try mul.wide combining at OptLevel > 0
4237 if (SDValue Ret = TryMULWIDECombine(N, DCI))
4244 /// PerformSHLCombine - Runs PTX-specific DAG combine patterns on SHL nodes.
4245 static SDValue PerformSHLCombine(SDNode *N,
4246 TargetLowering::DAGCombinerInfo &DCI,
4247 CodeGenOpt::Level OptLevel) {
4249 // Try mul.wide combining at OptLevel > 0
4250 if (SDValue Ret = TryMULWIDECombine(N, DCI))
4257 SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
4258 DAGCombinerInfo &DCI) const {
4259 CodeGenOpt::Level OptLevel = getTargetMachine().getOptLevel();
4260 switch (N->getOpcode()) {
4264 return PerformADDCombine(N, DCI, STI, OptLevel);
4266 return PerformMULCombine(N, DCI, OptLevel);
4268 return PerformSHLCombine(N, DCI, OptLevel);
4270 return PerformANDCombine(N, DCI);
4272 return PerformSELECTCombine(N, DCI);
4277 /// ReplaceVectorLoad - Convert vector loads into multi-output scalar loads.
4278 static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG,
4279 SmallVectorImpl<SDValue> &Results) {
4280 EVT ResVT = N->getValueType(0);
4283 assert(ResVT.isVector() && "Vector load must have vector type");
4285 // We only handle "native" vector sizes for now, e.g. <4 x double> is not
4286 // legal. We can (and should) split that into 2 loads of <2 x double> here
4287 // but I'm leaving that as a TODO for now.
4288 assert(ResVT.isSimple() && "Can only handle simple types");
4289 switch (ResVT.getSimpleVT().SimpleTy) {
4302 // This is a "native" vector type
4306 LoadSDNode *LD = cast<LoadSDNode>(N);
4308 unsigned Align = LD->getAlignment();
4309 auto &TD = DAG.getDataLayout();
4310 unsigned PrefAlign =
4311 TD.getPrefTypeAlignment(ResVT.getTypeForEVT(*DAG.getContext()));
4312 if (Align < PrefAlign) {
4313 // This load is not sufficiently aligned, so bail out and let this vector
4314 // load be scalarized. Note that we may still be able to emit smaller
4315 // vector loads. For example, if we are loading a <4 x float> with an
4316 // alignment of 8, this check will fail but the legalizer will try again
4317 // with 2 x <2 x float>, which will succeed with an alignment of 8.
4321 EVT EltVT = ResVT.getVectorElementType();
4322 unsigned NumElts = ResVT.getVectorNumElements();
4324 // Since LoadV2 is a target node, we cannot rely on DAG type legalization.
4325 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
4326 // loaded type to i16 and propagate the "real" type as the memory type.
4327 bool NeedTrunc = false;
4328 if (EltVT.getSizeInBits() < 16) {
4333 unsigned Opcode = 0;
4340 Opcode = NVPTXISD::LoadV2;
4341 LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4344 Opcode = NVPTXISD::LoadV4;
4345 EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4346 LdResVTs = DAG.getVTList(ListVTs);
4351 // Copy regular operands
4352 SmallVector<SDValue, 8> OtherOps(N->op_begin(), N->op_end());
4354 // The select routine does not have access to the LoadSDNode instance, so
4355 // pass along the extension information
4356 OtherOps.push_back(DAG.getIntPtrConstant(LD->getExtensionType(), DL));
4358 SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4360 LD->getMemOperand());
4362 SmallVector<SDValue, 4> ScalarRes;
4364 for (unsigned i = 0; i < NumElts; ++i) {
4365 SDValue Res = NewLD.getValue(i);
4367 Res = DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4368 ScalarRes.push_back(Res);
4371 SDValue LoadChain = NewLD.getValue(NumElts);
4373 SDValue BuildVec = DAG.getBuildVector(ResVT, DL, ScalarRes);
4375 Results.push_back(BuildVec);
4376 Results.push_back(LoadChain);
4379 static void ReplaceINTRINSIC_W_CHAIN(SDNode *N, SelectionDAG &DAG,
4380 SmallVectorImpl<SDValue> &Results) {
4381 SDValue Chain = N->getOperand(0);
4382 SDValue Intrin = N->getOperand(1);
4385 // Get the intrinsic ID
4386 unsigned IntrinNo = cast<ConstantSDNode>(Intrin.getNode())->getZExtValue();
4390 case Intrinsic::nvvm_ldg_global_i:
4391 case Intrinsic::nvvm_ldg_global_f:
4392 case Intrinsic::nvvm_ldg_global_p:
4393 case Intrinsic::nvvm_ldu_global_i:
4394 case Intrinsic::nvvm_ldu_global_f:
4395 case Intrinsic::nvvm_ldu_global_p: {
4396 EVT ResVT = N->getValueType(0);
4398 if (ResVT.isVector()) {
4401 unsigned NumElts = ResVT.getVectorNumElements();
4402 EVT EltVT = ResVT.getVectorElementType();
4404 // Since LDU/LDG are target nodes, we cannot rely on DAG type
4406 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
4407 // loaded type to i16 and propagate the "real" type as the memory type.
4408 bool NeedTrunc = false;
4409 if (EltVT.getSizeInBits() < 16) {
4414 unsigned Opcode = 0;
4424 case Intrinsic::nvvm_ldg_global_i:
4425 case Intrinsic::nvvm_ldg_global_f:
4426 case Intrinsic::nvvm_ldg_global_p:
4427 Opcode = NVPTXISD::LDGV2;
4429 case Intrinsic::nvvm_ldu_global_i:
4430 case Intrinsic::nvvm_ldu_global_f:
4431 case Intrinsic::nvvm_ldu_global_p:
4432 Opcode = NVPTXISD::LDUV2;
4435 LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4441 case Intrinsic::nvvm_ldg_global_i:
4442 case Intrinsic::nvvm_ldg_global_f:
4443 case Intrinsic::nvvm_ldg_global_p:
4444 Opcode = NVPTXISD::LDGV4;
4446 case Intrinsic::nvvm_ldu_global_i:
4447 case Intrinsic::nvvm_ldu_global_f:
4448 case Intrinsic::nvvm_ldu_global_p:
4449 Opcode = NVPTXISD::LDUV4;
4452 EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4453 LdResVTs = DAG.getVTList(ListVTs);
4458 SmallVector<SDValue, 8> OtherOps;
4460 // Copy regular operands
4462 OtherOps.push_back(Chain); // Chain
4463 // Skip operand 1 (intrinsic ID)
4465 OtherOps.append(N->op_begin() + 2, N->op_end());
4467 MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4469 SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4470 MemSD->getMemoryVT(),
4471 MemSD->getMemOperand());
4473 SmallVector<SDValue, 4> ScalarRes;
4475 for (unsigned i = 0; i < NumElts; ++i) {
4476 SDValue Res = NewLD.getValue(i);
4479 DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4480 ScalarRes.push_back(Res);
4483 SDValue LoadChain = NewLD.getValue(NumElts);
4486 DAG.getBuildVector(ResVT, DL, ScalarRes);
4488 Results.push_back(BuildVec);
4489 Results.push_back(LoadChain);
4492 assert(ResVT.isSimple() && ResVT.getSimpleVT().SimpleTy == MVT::i8 &&
4493 "Custom handling of non-i8 ldu/ldg?");
4495 // Just copy all operands as-is
4496 SmallVector<SDValue, 4> Ops(N->op_begin(), N->op_end());
4498 // Force output to i16
4499 SDVTList LdResVTs = DAG.getVTList(MVT::i16, MVT::Other);
4501 MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4503 // We make sure the memory type is i8, which will be used during isel
4504 // to select the proper instruction.
4506 DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, LdResVTs, Ops,
4507 MVT::i8, MemSD->getMemOperand());
4509 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
4510 NewLD.getValue(0)));
4511 Results.push_back(NewLD.getValue(1));
4517 void NVPTXTargetLowering::ReplaceNodeResults(
4518 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
4519 switch (N->getOpcode()) {
4521 report_fatal_error("Unhandled custom legalization");
4523 ReplaceLoadVector(N, DAG, Results);
4525 case ISD::INTRINSIC_W_CHAIN:
4526 ReplaceINTRINSIC_W_CHAIN(N, DAG, Results);
4531 // Pin NVPTXSection's and NVPTXTargetObjectFile's vtables to this file.
4532 void NVPTXSection::anchor() {}
4534 NVPTXTargetObjectFile::~NVPTXTargetObjectFile() {
4535 delete static_cast<NVPTXSection *>(TextSection);
4536 delete static_cast<NVPTXSection *>(DataSection);
4537 delete static_cast<NVPTXSection *>(BSSSection);
4538 delete static_cast<NVPTXSection *>(ReadOnlySection);
4540 delete static_cast<NVPTXSection *>(StaticCtorSection);
4541 delete static_cast<NVPTXSection *>(StaticDtorSection);
4542 delete static_cast<NVPTXSection *>(LSDASection);
4543 delete static_cast<NVPTXSection *>(EHFrameSection);
4544 delete static_cast<NVPTXSection *>(DwarfAbbrevSection);
4545 delete static_cast<NVPTXSection *>(DwarfInfoSection);
4546 delete static_cast<NVPTXSection *>(DwarfLineSection);
4547 delete static_cast<NVPTXSection *>(DwarfFrameSection);
4548 delete static_cast<NVPTXSection *>(DwarfPubTypesSection);
4549 delete static_cast<const NVPTXSection *>(DwarfDebugInlineSection);
4550 delete static_cast<NVPTXSection *>(DwarfStrSection);
4551 delete static_cast<NVPTXSection *>(DwarfLocSection);
4552 delete static_cast<NVPTXSection *>(DwarfARangesSection);
4553 delete static_cast<NVPTXSection *>(DwarfRangesSection);
4554 delete static_cast<NVPTXSection *>(DwarfMacinfoSection);
4558 NVPTXTargetObjectFile::SelectSectionForGlobal(const GlobalValue *GV,
4559 SectionKind Kind, Mangler &Mang,
4560 const TargetMachine &TM) const {
4561 return getDataSection();