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, Type *Ty,
84 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, 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 // By default, use the Source scheduling
129 setSchedulingPreference(Sched::RegPressure);
131 setSchedulingPreference(Sched::Source);
133 addRegisterClass(MVT::i1, &NVPTX::Int1RegsRegClass);
134 addRegisterClass(MVT::i16, &NVPTX::Int16RegsRegClass);
135 addRegisterClass(MVT::i32, &NVPTX::Int32RegsRegClass);
136 addRegisterClass(MVT::i64, &NVPTX::Int64RegsRegClass);
137 addRegisterClass(MVT::f32, &NVPTX::Float32RegsRegClass);
138 addRegisterClass(MVT::f64, &NVPTX::Float64RegsRegClass);
140 // Operations not directly supported by NVPTX.
141 setOperationAction(ISD::SELECT_CC, MVT::f32, Expand);
142 setOperationAction(ISD::SELECT_CC, MVT::f64, Expand);
143 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
144 setOperationAction(ISD::SELECT_CC, MVT::i8, Expand);
145 setOperationAction(ISD::SELECT_CC, MVT::i16, Expand);
146 setOperationAction(ISD::SELECT_CC, MVT::i32, Expand);
147 setOperationAction(ISD::SELECT_CC, MVT::i64, Expand);
148 setOperationAction(ISD::BR_CC, MVT::f32, Expand);
149 setOperationAction(ISD::BR_CC, MVT::f64, Expand);
150 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
151 setOperationAction(ISD::BR_CC, MVT::i8, Expand);
152 setOperationAction(ISD::BR_CC, MVT::i16, Expand);
153 setOperationAction(ISD::BR_CC, MVT::i32, Expand);
154 setOperationAction(ISD::BR_CC, MVT::i64, Expand);
155 // Some SIGN_EXTEND_INREG can be done using cvt instruction.
156 // For others we will expand to a SHL/SRA pair.
157 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i64, Legal);
158 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
159 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Legal);
160 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
161 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
163 setOperationAction(ISD::SHL_PARTS, MVT::i32 , Custom);
164 setOperationAction(ISD::SRA_PARTS, MVT::i32 , Custom);
165 setOperationAction(ISD::SRL_PARTS, MVT::i32 , Custom);
166 setOperationAction(ISD::SHL_PARTS, MVT::i64 , Custom);
167 setOperationAction(ISD::SRA_PARTS, MVT::i64 , Custom);
168 setOperationAction(ISD::SRL_PARTS, MVT::i64 , Custom);
170 if (STI.hasROT64()) {
171 setOperationAction(ISD::ROTL, MVT::i64, Legal);
172 setOperationAction(ISD::ROTR, MVT::i64, Legal);
174 setOperationAction(ISD::ROTL, MVT::i64, Expand);
175 setOperationAction(ISD::ROTR, MVT::i64, Expand);
177 if (STI.hasROT32()) {
178 setOperationAction(ISD::ROTL, MVT::i32, Legal);
179 setOperationAction(ISD::ROTR, MVT::i32, Legal);
181 setOperationAction(ISD::ROTL, MVT::i32, Expand);
182 setOperationAction(ISD::ROTR, MVT::i32, Expand);
185 setOperationAction(ISD::ROTL, MVT::i16, Expand);
186 setOperationAction(ISD::ROTR, MVT::i16, Expand);
187 setOperationAction(ISD::ROTL, MVT::i8, Expand);
188 setOperationAction(ISD::ROTR, MVT::i8, Expand);
189 setOperationAction(ISD::BSWAP, MVT::i16, Expand);
190 setOperationAction(ISD::BSWAP, MVT::i32, Expand);
191 setOperationAction(ISD::BSWAP, MVT::i64, Expand);
193 // Indirect branch is not supported.
194 // This also disables Jump Table creation.
195 setOperationAction(ISD::BR_JT, MVT::Other, Expand);
196 setOperationAction(ISD::BRIND, MVT::Other, Expand);
198 setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
199 setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
201 // We want to legalize constant related memmove and memcopy
203 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
205 // Turn FP extload into load/fextend
206 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
207 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
208 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
209 // Turn FP truncstore into trunc + store.
210 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
211 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
212 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
214 // PTX does not support load / store predicate registers
215 setOperationAction(ISD::LOAD, MVT::i1, Custom);
216 setOperationAction(ISD::STORE, MVT::i1, Custom);
218 for (MVT VT : MVT::integer_valuetypes()) {
219 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
220 setLoadExtAction(ISD::ZEXTLOAD, VT, MVT::i1, Promote);
221 setTruncStoreAction(VT, MVT::i1, Expand);
224 // This is legal in NVPTX
225 setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
226 setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
228 // TRAP can be lowered to PTX trap
229 setOperationAction(ISD::TRAP, MVT::Other, Legal);
231 setOperationAction(ISD::ADDC, MVT::i64, Expand);
232 setOperationAction(ISD::ADDE, MVT::i64, Expand);
234 // Register custom handling for vector loads/stores
235 for (MVT VT : MVT::vector_valuetypes()) {
236 if (IsPTXVectorType(VT)) {
237 setOperationAction(ISD::LOAD, VT, Custom);
238 setOperationAction(ISD::STORE, VT, Custom);
239 setOperationAction(ISD::INTRINSIC_W_CHAIN, VT, Custom);
243 // Custom handling for i8 intrinsics
244 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);
246 setOperationAction(ISD::CTLZ, MVT::i16, Legal);
247 setOperationAction(ISD::CTLZ, MVT::i32, Legal);
248 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
249 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Legal);
250 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Legal);
251 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Legal);
252 setOperationAction(ISD::CTTZ, MVT::i16, Expand);
253 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
254 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
255 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Expand);
256 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
257 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
258 setOperationAction(ISD::CTPOP, MVT::i16, Legal);
259 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
260 setOperationAction(ISD::CTPOP, MVT::i64, Legal);
262 // PTX does not directly support SELP of i1, so promote to i32 first
263 setOperationAction(ISD::SELECT, MVT::i1, Custom);
265 // We have some custom DAG combine patterns for these nodes
266 setTargetDAGCombine(ISD::ADD);
267 setTargetDAGCombine(ISD::AND);
268 setTargetDAGCombine(ISD::FADD);
269 setTargetDAGCombine(ISD::MUL);
270 setTargetDAGCombine(ISD::SHL);
272 // Now deduce the information based on the above mentioned
274 computeRegisterProperties(STI.getRegisterInfo());
277 const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
278 switch ((NVPTXISD::NodeType)Opcode) {
279 case NVPTXISD::FIRST_NUMBER:
282 return "NVPTXISD::CALL";
283 case NVPTXISD::RET_FLAG:
284 return "NVPTXISD::RET_FLAG";
285 case NVPTXISD::LOAD_PARAM:
286 return "NVPTXISD::LOAD_PARAM";
287 case NVPTXISD::Wrapper:
288 return "NVPTXISD::Wrapper";
289 case NVPTXISD::DeclareParam:
290 return "NVPTXISD::DeclareParam";
291 case NVPTXISD::DeclareScalarParam:
292 return "NVPTXISD::DeclareScalarParam";
293 case NVPTXISD::DeclareRet:
294 return "NVPTXISD::DeclareRet";
295 case NVPTXISD::DeclareScalarRet:
296 return "NVPTXISD::DeclareScalarRet";
297 case NVPTXISD::DeclareRetParam:
298 return "NVPTXISD::DeclareRetParam";
299 case NVPTXISD::PrintCall:
300 return "NVPTXISD::PrintCall";
301 case NVPTXISD::PrintCallUni:
302 return "NVPTXISD::PrintCallUni";
303 case NVPTXISD::LoadParam:
304 return "NVPTXISD::LoadParam";
305 case NVPTXISD::LoadParamV2:
306 return "NVPTXISD::LoadParamV2";
307 case NVPTXISD::LoadParamV4:
308 return "NVPTXISD::LoadParamV4";
309 case NVPTXISD::StoreParam:
310 return "NVPTXISD::StoreParam";
311 case NVPTXISD::StoreParamV2:
312 return "NVPTXISD::StoreParamV2";
313 case NVPTXISD::StoreParamV4:
314 return "NVPTXISD::StoreParamV4";
315 case NVPTXISD::StoreParamS32:
316 return "NVPTXISD::StoreParamS32";
317 case NVPTXISD::StoreParamU32:
318 return "NVPTXISD::StoreParamU32";
319 case NVPTXISD::CallArgBegin:
320 return "NVPTXISD::CallArgBegin";
321 case NVPTXISD::CallArg:
322 return "NVPTXISD::CallArg";
323 case NVPTXISD::LastCallArg:
324 return "NVPTXISD::LastCallArg";
325 case NVPTXISD::CallArgEnd:
326 return "NVPTXISD::CallArgEnd";
327 case NVPTXISD::CallVoid:
328 return "NVPTXISD::CallVoid";
329 case NVPTXISD::CallVal:
330 return "NVPTXISD::CallVal";
331 case NVPTXISD::CallSymbol:
332 return "NVPTXISD::CallSymbol";
333 case NVPTXISD::Prototype:
334 return "NVPTXISD::Prototype";
335 case NVPTXISD::MoveParam:
336 return "NVPTXISD::MoveParam";
337 case NVPTXISD::StoreRetval:
338 return "NVPTXISD::StoreRetval";
339 case NVPTXISD::StoreRetvalV2:
340 return "NVPTXISD::StoreRetvalV2";
341 case NVPTXISD::StoreRetvalV4:
342 return "NVPTXISD::StoreRetvalV4";
343 case NVPTXISD::PseudoUseParam:
344 return "NVPTXISD::PseudoUseParam";
345 case NVPTXISD::RETURN:
346 return "NVPTXISD::RETURN";
347 case NVPTXISD::CallSeqBegin:
348 return "NVPTXISD::CallSeqBegin";
349 case NVPTXISD::CallSeqEnd:
350 return "NVPTXISD::CallSeqEnd";
351 case NVPTXISD::CallPrototype:
352 return "NVPTXISD::CallPrototype";
353 case NVPTXISD::LoadV2:
354 return "NVPTXISD::LoadV2";
355 case NVPTXISD::LoadV4:
356 return "NVPTXISD::LoadV4";
357 case NVPTXISD::LDGV2:
358 return "NVPTXISD::LDGV2";
359 case NVPTXISD::LDGV4:
360 return "NVPTXISD::LDGV4";
361 case NVPTXISD::LDUV2:
362 return "NVPTXISD::LDUV2";
363 case NVPTXISD::LDUV4:
364 return "NVPTXISD::LDUV4";
365 case NVPTXISD::StoreV2:
366 return "NVPTXISD::StoreV2";
367 case NVPTXISD::StoreV4:
368 return "NVPTXISD::StoreV4";
369 case NVPTXISD::FUN_SHFL_CLAMP:
370 return "NVPTXISD::FUN_SHFL_CLAMP";
371 case NVPTXISD::FUN_SHFR_CLAMP:
372 return "NVPTXISD::FUN_SHFR_CLAMP";
374 return "NVPTXISD::IMAD";
375 case NVPTXISD::Dummy:
376 return "NVPTXISD::Dummy";
377 case NVPTXISD::MUL_WIDE_SIGNED:
378 return "NVPTXISD::MUL_WIDE_SIGNED";
379 case NVPTXISD::MUL_WIDE_UNSIGNED:
380 return "NVPTXISD::MUL_WIDE_UNSIGNED";
381 case NVPTXISD::Tex1DFloatS32: return "NVPTXISD::Tex1DFloatS32";
382 case NVPTXISD::Tex1DFloatFloat: return "NVPTXISD::Tex1DFloatFloat";
383 case NVPTXISD::Tex1DFloatFloatLevel:
384 return "NVPTXISD::Tex1DFloatFloatLevel";
385 case NVPTXISD::Tex1DFloatFloatGrad:
386 return "NVPTXISD::Tex1DFloatFloatGrad";
387 case NVPTXISD::Tex1DS32S32: return "NVPTXISD::Tex1DS32S32";
388 case NVPTXISD::Tex1DS32Float: return "NVPTXISD::Tex1DS32Float";
389 case NVPTXISD::Tex1DS32FloatLevel:
390 return "NVPTXISD::Tex1DS32FloatLevel";
391 case NVPTXISD::Tex1DS32FloatGrad:
392 return "NVPTXISD::Tex1DS32FloatGrad";
393 case NVPTXISD::Tex1DU32S32: return "NVPTXISD::Tex1DU32S32";
394 case NVPTXISD::Tex1DU32Float: return "NVPTXISD::Tex1DU32Float";
395 case NVPTXISD::Tex1DU32FloatLevel:
396 return "NVPTXISD::Tex1DU32FloatLevel";
397 case NVPTXISD::Tex1DU32FloatGrad:
398 return "NVPTXISD::Tex1DU32FloatGrad";
399 case NVPTXISD::Tex1DArrayFloatS32: return "NVPTXISD::Tex1DArrayFloatS32";
400 case NVPTXISD::Tex1DArrayFloatFloat: return "NVPTXISD::Tex1DArrayFloatFloat";
401 case NVPTXISD::Tex1DArrayFloatFloatLevel:
402 return "NVPTXISD::Tex1DArrayFloatFloatLevel";
403 case NVPTXISD::Tex1DArrayFloatFloatGrad:
404 return "NVPTXISD::Tex1DArrayFloatFloatGrad";
405 case NVPTXISD::Tex1DArrayS32S32: return "NVPTXISD::Tex1DArrayS32S32";
406 case NVPTXISD::Tex1DArrayS32Float: return "NVPTXISD::Tex1DArrayS32Float";
407 case NVPTXISD::Tex1DArrayS32FloatLevel:
408 return "NVPTXISD::Tex1DArrayS32FloatLevel";
409 case NVPTXISD::Tex1DArrayS32FloatGrad:
410 return "NVPTXISD::Tex1DArrayS32FloatGrad";
411 case NVPTXISD::Tex1DArrayU32S32: return "NVPTXISD::Tex1DArrayU32S32";
412 case NVPTXISD::Tex1DArrayU32Float: return "NVPTXISD::Tex1DArrayU32Float";
413 case NVPTXISD::Tex1DArrayU32FloatLevel:
414 return "NVPTXISD::Tex1DArrayU32FloatLevel";
415 case NVPTXISD::Tex1DArrayU32FloatGrad:
416 return "NVPTXISD::Tex1DArrayU32FloatGrad";
417 case NVPTXISD::Tex2DFloatS32: return "NVPTXISD::Tex2DFloatS32";
418 case NVPTXISD::Tex2DFloatFloat: return "NVPTXISD::Tex2DFloatFloat";
419 case NVPTXISD::Tex2DFloatFloatLevel:
420 return "NVPTXISD::Tex2DFloatFloatLevel";
421 case NVPTXISD::Tex2DFloatFloatGrad:
422 return "NVPTXISD::Tex2DFloatFloatGrad";
423 case NVPTXISD::Tex2DS32S32: return "NVPTXISD::Tex2DS32S32";
424 case NVPTXISD::Tex2DS32Float: return "NVPTXISD::Tex2DS32Float";
425 case NVPTXISD::Tex2DS32FloatLevel:
426 return "NVPTXISD::Tex2DS32FloatLevel";
427 case NVPTXISD::Tex2DS32FloatGrad:
428 return "NVPTXISD::Tex2DS32FloatGrad";
429 case NVPTXISD::Tex2DU32S32: return "NVPTXISD::Tex2DU32S32";
430 case NVPTXISD::Tex2DU32Float: return "NVPTXISD::Tex2DU32Float";
431 case NVPTXISD::Tex2DU32FloatLevel:
432 return "NVPTXISD::Tex2DU32FloatLevel";
433 case NVPTXISD::Tex2DU32FloatGrad:
434 return "NVPTXISD::Tex2DU32FloatGrad";
435 case NVPTXISD::Tex2DArrayFloatS32: return "NVPTXISD::Tex2DArrayFloatS32";
436 case NVPTXISD::Tex2DArrayFloatFloat: return "NVPTXISD::Tex2DArrayFloatFloat";
437 case NVPTXISD::Tex2DArrayFloatFloatLevel:
438 return "NVPTXISD::Tex2DArrayFloatFloatLevel";
439 case NVPTXISD::Tex2DArrayFloatFloatGrad:
440 return "NVPTXISD::Tex2DArrayFloatFloatGrad";
441 case NVPTXISD::Tex2DArrayS32S32: return "NVPTXISD::Tex2DArrayS32S32";
442 case NVPTXISD::Tex2DArrayS32Float: return "NVPTXISD::Tex2DArrayS32Float";
443 case NVPTXISD::Tex2DArrayS32FloatLevel:
444 return "NVPTXISD::Tex2DArrayS32FloatLevel";
445 case NVPTXISD::Tex2DArrayS32FloatGrad:
446 return "NVPTXISD::Tex2DArrayS32FloatGrad";
447 case NVPTXISD::Tex2DArrayU32S32: return "NVPTXISD::Tex2DArrayU32S32";
448 case NVPTXISD::Tex2DArrayU32Float: return "NVPTXISD::Tex2DArrayU32Float";
449 case NVPTXISD::Tex2DArrayU32FloatLevel:
450 return "NVPTXISD::Tex2DArrayU32FloatLevel";
451 case NVPTXISD::Tex2DArrayU32FloatGrad:
452 return "NVPTXISD::Tex2DArrayU32FloatGrad";
453 case NVPTXISD::Tex3DFloatS32: return "NVPTXISD::Tex3DFloatS32";
454 case NVPTXISD::Tex3DFloatFloat: return "NVPTXISD::Tex3DFloatFloat";
455 case NVPTXISD::Tex3DFloatFloatLevel:
456 return "NVPTXISD::Tex3DFloatFloatLevel";
457 case NVPTXISD::Tex3DFloatFloatGrad:
458 return "NVPTXISD::Tex3DFloatFloatGrad";
459 case NVPTXISD::Tex3DS32S32: return "NVPTXISD::Tex3DS32S32";
460 case NVPTXISD::Tex3DS32Float: return "NVPTXISD::Tex3DS32Float";
461 case NVPTXISD::Tex3DS32FloatLevel:
462 return "NVPTXISD::Tex3DS32FloatLevel";
463 case NVPTXISD::Tex3DS32FloatGrad:
464 return "NVPTXISD::Tex3DS32FloatGrad";
465 case NVPTXISD::Tex3DU32S32: return "NVPTXISD::Tex3DU32S32";
466 case NVPTXISD::Tex3DU32Float: return "NVPTXISD::Tex3DU32Float";
467 case NVPTXISD::Tex3DU32FloatLevel:
468 return "NVPTXISD::Tex3DU32FloatLevel";
469 case NVPTXISD::Tex3DU32FloatGrad:
470 return "NVPTXISD::Tex3DU32FloatGrad";
471 case NVPTXISD::TexCubeFloatFloat: return "NVPTXISD::TexCubeFloatFloat";
472 case NVPTXISD::TexCubeFloatFloatLevel:
473 return "NVPTXISD::TexCubeFloatFloatLevel";
474 case NVPTXISD::TexCubeS32Float: return "NVPTXISD::TexCubeS32Float";
475 case NVPTXISD::TexCubeS32FloatLevel:
476 return "NVPTXISD::TexCubeS32FloatLevel";
477 case NVPTXISD::TexCubeU32Float: return "NVPTXISD::TexCubeU32Float";
478 case NVPTXISD::TexCubeU32FloatLevel:
479 return "NVPTXISD::TexCubeU32FloatLevel";
480 case NVPTXISD::TexCubeArrayFloatFloat:
481 return "NVPTXISD::TexCubeArrayFloatFloat";
482 case NVPTXISD::TexCubeArrayFloatFloatLevel:
483 return "NVPTXISD::TexCubeArrayFloatFloatLevel";
484 case NVPTXISD::TexCubeArrayS32Float:
485 return "NVPTXISD::TexCubeArrayS32Float";
486 case NVPTXISD::TexCubeArrayS32FloatLevel:
487 return "NVPTXISD::TexCubeArrayS32FloatLevel";
488 case NVPTXISD::TexCubeArrayU32Float:
489 return "NVPTXISD::TexCubeArrayU32Float";
490 case NVPTXISD::TexCubeArrayU32FloatLevel:
491 return "NVPTXISD::TexCubeArrayU32FloatLevel";
492 case NVPTXISD::Tld4R2DFloatFloat:
493 return "NVPTXISD::Tld4R2DFloatFloat";
494 case NVPTXISD::Tld4G2DFloatFloat:
495 return "NVPTXISD::Tld4G2DFloatFloat";
496 case NVPTXISD::Tld4B2DFloatFloat:
497 return "NVPTXISD::Tld4B2DFloatFloat";
498 case NVPTXISD::Tld4A2DFloatFloat:
499 return "NVPTXISD::Tld4A2DFloatFloat";
500 case NVPTXISD::Tld4R2DS64Float:
501 return "NVPTXISD::Tld4R2DS64Float";
502 case NVPTXISD::Tld4G2DS64Float:
503 return "NVPTXISD::Tld4G2DS64Float";
504 case NVPTXISD::Tld4B2DS64Float:
505 return "NVPTXISD::Tld4B2DS64Float";
506 case NVPTXISD::Tld4A2DS64Float:
507 return "NVPTXISD::Tld4A2DS64Float";
508 case NVPTXISD::Tld4R2DU64Float:
509 return "NVPTXISD::Tld4R2DU64Float";
510 case NVPTXISD::Tld4G2DU64Float:
511 return "NVPTXISD::Tld4G2DU64Float";
512 case NVPTXISD::Tld4B2DU64Float:
513 return "NVPTXISD::Tld4B2DU64Float";
514 case NVPTXISD::Tld4A2DU64Float:
515 return "NVPTXISD::Tld4A2DU64Float";
517 case NVPTXISD::TexUnified1DFloatS32:
518 return "NVPTXISD::TexUnified1DFloatS32";
519 case NVPTXISD::TexUnified1DFloatFloat:
520 return "NVPTXISD::TexUnified1DFloatFloat";
521 case NVPTXISD::TexUnified1DFloatFloatLevel:
522 return "NVPTXISD::TexUnified1DFloatFloatLevel";
523 case NVPTXISD::TexUnified1DFloatFloatGrad:
524 return "NVPTXISD::TexUnified1DFloatFloatGrad";
525 case NVPTXISD::TexUnified1DS32S32:
526 return "NVPTXISD::TexUnified1DS32S32";
527 case NVPTXISD::TexUnified1DS32Float:
528 return "NVPTXISD::TexUnified1DS32Float";
529 case NVPTXISD::TexUnified1DS32FloatLevel:
530 return "NVPTXISD::TexUnified1DS32FloatLevel";
531 case NVPTXISD::TexUnified1DS32FloatGrad:
532 return "NVPTXISD::TexUnified1DS32FloatGrad";
533 case NVPTXISD::TexUnified1DU32S32:
534 return "NVPTXISD::TexUnified1DU32S32";
535 case NVPTXISD::TexUnified1DU32Float:
536 return "NVPTXISD::TexUnified1DU32Float";
537 case NVPTXISD::TexUnified1DU32FloatLevel:
538 return "NVPTXISD::TexUnified1DU32FloatLevel";
539 case NVPTXISD::TexUnified1DU32FloatGrad:
540 return "NVPTXISD::TexUnified1DU32FloatGrad";
541 case NVPTXISD::TexUnified1DArrayFloatS32:
542 return "NVPTXISD::TexUnified1DArrayFloatS32";
543 case NVPTXISD::TexUnified1DArrayFloatFloat:
544 return "NVPTXISD::TexUnified1DArrayFloatFloat";
545 case NVPTXISD::TexUnified1DArrayFloatFloatLevel:
546 return "NVPTXISD::TexUnified1DArrayFloatFloatLevel";
547 case NVPTXISD::TexUnified1DArrayFloatFloatGrad:
548 return "NVPTXISD::TexUnified1DArrayFloatFloatGrad";
549 case NVPTXISD::TexUnified1DArrayS32S32:
550 return "NVPTXISD::TexUnified1DArrayS32S32";
551 case NVPTXISD::TexUnified1DArrayS32Float:
552 return "NVPTXISD::TexUnified1DArrayS32Float";
553 case NVPTXISD::TexUnified1DArrayS32FloatLevel:
554 return "NVPTXISD::TexUnified1DArrayS32FloatLevel";
555 case NVPTXISD::TexUnified1DArrayS32FloatGrad:
556 return "NVPTXISD::TexUnified1DArrayS32FloatGrad";
557 case NVPTXISD::TexUnified1DArrayU32S32:
558 return "NVPTXISD::TexUnified1DArrayU32S32";
559 case NVPTXISD::TexUnified1DArrayU32Float:
560 return "NVPTXISD::TexUnified1DArrayU32Float";
561 case NVPTXISD::TexUnified1DArrayU32FloatLevel:
562 return "NVPTXISD::TexUnified1DArrayU32FloatLevel";
563 case NVPTXISD::TexUnified1DArrayU32FloatGrad:
564 return "NVPTXISD::TexUnified1DArrayU32FloatGrad";
565 case NVPTXISD::TexUnified2DFloatS32:
566 return "NVPTXISD::TexUnified2DFloatS32";
567 case NVPTXISD::TexUnified2DFloatFloat:
568 return "NVPTXISD::TexUnified2DFloatFloat";
569 case NVPTXISD::TexUnified2DFloatFloatLevel:
570 return "NVPTXISD::TexUnified2DFloatFloatLevel";
571 case NVPTXISD::TexUnified2DFloatFloatGrad:
572 return "NVPTXISD::TexUnified2DFloatFloatGrad";
573 case NVPTXISD::TexUnified2DS32S32:
574 return "NVPTXISD::TexUnified2DS32S32";
575 case NVPTXISD::TexUnified2DS32Float:
576 return "NVPTXISD::TexUnified2DS32Float";
577 case NVPTXISD::TexUnified2DS32FloatLevel:
578 return "NVPTXISD::TexUnified2DS32FloatLevel";
579 case NVPTXISD::TexUnified2DS32FloatGrad:
580 return "NVPTXISD::TexUnified2DS32FloatGrad";
581 case NVPTXISD::TexUnified2DU32S32:
582 return "NVPTXISD::TexUnified2DU32S32";
583 case NVPTXISD::TexUnified2DU32Float:
584 return "NVPTXISD::TexUnified2DU32Float";
585 case NVPTXISD::TexUnified2DU32FloatLevel:
586 return "NVPTXISD::TexUnified2DU32FloatLevel";
587 case NVPTXISD::TexUnified2DU32FloatGrad:
588 return "NVPTXISD::TexUnified2DU32FloatGrad";
589 case NVPTXISD::TexUnified2DArrayFloatS32:
590 return "NVPTXISD::TexUnified2DArrayFloatS32";
591 case NVPTXISD::TexUnified2DArrayFloatFloat:
592 return "NVPTXISD::TexUnified2DArrayFloatFloat";
593 case NVPTXISD::TexUnified2DArrayFloatFloatLevel:
594 return "NVPTXISD::TexUnified2DArrayFloatFloatLevel";
595 case NVPTXISD::TexUnified2DArrayFloatFloatGrad:
596 return "NVPTXISD::TexUnified2DArrayFloatFloatGrad";
597 case NVPTXISD::TexUnified2DArrayS32S32:
598 return "NVPTXISD::TexUnified2DArrayS32S32";
599 case NVPTXISD::TexUnified2DArrayS32Float:
600 return "NVPTXISD::TexUnified2DArrayS32Float";
601 case NVPTXISD::TexUnified2DArrayS32FloatLevel:
602 return "NVPTXISD::TexUnified2DArrayS32FloatLevel";
603 case NVPTXISD::TexUnified2DArrayS32FloatGrad:
604 return "NVPTXISD::TexUnified2DArrayS32FloatGrad";
605 case NVPTXISD::TexUnified2DArrayU32S32:
606 return "NVPTXISD::TexUnified2DArrayU32S32";
607 case NVPTXISD::TexUnified2DArrayU32Float:
608 return "NVPTXISD::TexUnified2DArrayU32Float";
609 case NVPTXISD::TexUnified2DArrayU32FloatLevel:
610 return "NVPTXISD::TexUnified2DArrayU32FloatLevel";
611 case NVPTXISD::TexUnified2DArrayU32FloatGrad:
612 return "NVPTXISD::TexUnified2DArrayU32FloatGrad";
613 case NVPTXISD::TexUnified3DFloatS32:
614 return "NVPTXISD::TexUnified3DFloatS32";
615 case NVPTXISD::TexUnified3DFloatFloat:
616 return "NVPTXISD::TexUnified3DFloatFloat";
617 case NVPTXISD::TexUnified3DFloatFloatLevel:
618 return "NVPTXISD::TexUnified3DFloatFloatLevel";
619 case NVPTXISD::TexUnified3DFloatFloatGrad:
620 return "NVPTXISD::TexUnified3DFloatFloatGrad";
621 case NVPTXISD::TexUnified3DS32S32:
622 return "NVPTXISD::TexUnified3DS32S32";
623 case NVPTXISD::TexUnified3DS32Float:
624 return "NVPTXISD::TexUnified3DS32Float";
625 case NVPTXISD::TexUnified3DS32FloatLevel:
626 return "NVPTXISD::TexUnified3DS32FloatLevel";
627 case NVPTXISD::TexUnified3DS32FloatGrad:
628 return "NVPTXISD::TexUnified3DS32FloatGrad";
629 case NVPTXISD::TexUnified3DU32S32:
630 return "NVPTXISD::TexUnified3DU32S32";
631 case NVPTXISD::TexUnified3DU32Float:
632 return "NVPTXISD::TexUnified3DU32Float";
633 case NVPTXISD::TexUnified3DU32FloatLevel:
634 return "NVPTXISD::TexUnified3DU32FloatLevel";
635 case NVPTXISD::TexUnified3DU32FloatGrad:
636 return "NVPTXISD::TexUnified3DU32FloatGrad";
637 case NVPTXISD::TexUnifiedCubeFloatFloat:
638 return "NVPTXISD::TexUnifiedCubeFloatFloat";
639 case NVPTXISD::TexUnifiedCubeFloatFloatLevel:
640 return "NVPTXISD::TexUnifiedCubeFloatFloatLevel";
641 case NVPTXISD::TexUnifiedCubeS32Float:
642 return "NVPTXISD::TexUnifiedCubeS32Float";
643 case NVPTXISD::TexUnifiedCubeS32FloatLevel:
644 return "NVPTXISD::TexUnifiedCubeS32FloatLevel";
645 case NVPTXISD::TexUnifiedCubeU32Float:
646 return "NVPTXISD::TexUnifiedCubeU32Float";
647 case NVPTXISD::TexUnifiedCubeU32FloatLevel:
648 return "NVPTXISD::TexUnifiedCubeU32FloatLevel";
649 case NVPTXISD::TexUnifiedCubeArrayFloatFloat:
650 return "NVPTXISD::TexUnifiedCubeArrayFloatFloat";
651 case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel:
652 return "NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel";
653 case NVPTXISD::TexUnifiedCubeArrayS32Float:
654 return "NVPTXISD::TexUnifiedCubeArrayS32Float";
655 case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel:
656 return "NVPTXISD::TexUnifiedCubeArrayS32FloatLevel";
657 case NVPTXISD::TexUnifiedCubeArrayU32Float:
658 return "NVPTXISD::TexUnifiedCubeArrayU32Float";
659 case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel:
660 return "NVPTXISD::TexUnifiedCubeArrayU32FloatLevel";
661 case NVPTXISD::Tld4UnifiedR2DFloatFloat:
662 return "NVPTXISD::Tld4UnifiedR2DFloatFloat";
663 case NVPTXISD::Tld4UnifiedG2DFloatFloat:
664 return "NVPTXISD::Tld4UnifiedG2DFloatFloat";
665 case NVPTXISD::Tld4UnifiedB2DFloatFloat:
666 return "NVPTXISD::Tld4UnifiedB2DFloatFloat";
667 case NVPTXISD::Tld4UnifiedA2DFloatFloat:
668 return "NVPTXISD::Tld4UnifiedA2DFloatFloat";
669 case NVPTXISD::Tld4UnifiedR2DS64Float:
670 return "NVPTXISD::Tld4UnifiedR2DS64Float";
671 case NVPTXISD::Tld4UnifiedG2DS64Float:
672 return "NVPTXISD::Tld4UnifiedG2DS64Float";
673 case NVPTXISD::Tld4UnifiedB2DS64Float:
674 return "NVPTXISD::Tld4UnifiedB2DS64Float";
675 case NVPTXISD::Tld4UnifiedA2DS64Float:
676 return "NVPTXISD::Tld4UnifiedA2DS64Float";
677 case NVPTXISD::Tld4UnifiedR2DU64Float:
678 return "NVPTXISD::Tld4UnifiedR2DU64Float";
679 case NVPTXISD::Tld4UnifiedG2DU64Float:
680 return "NVPTXISD::Tld4UnifiedG2DU64Float";
681 case NVPTXISD::Tld4UnifiedB2DU64Float:
682 return "NVPTXISD::Tld4UnifiedB2DU64Float";
683 case NVPTXISD::Tld4UnifiedA2DU64Float:
684 return "NVPTXISD::Tld4UnifiedA2DU64Float";
686 case NVPTXISD::Suld1DI8Clamp: return "NVPTXISD::Suld1DI8Clamp";
687 case NVPTXISD::Suld1DI16Clamp: return "NVPTXISD::Suld1DI16Clamp";
688 case NVPTXISD::Suld1DI32Clamp: return "NVPTXISD::Suld1DI32Clamp";
689 case NVPTXISD::Suld1DI64Clamp: return "NVPTXISD::Suld1DI64Clamp";
690 case NVPTXISD::Suld1DV2I8Clamp: return "NVPTXISD::Suld1DV2I8Clamp";
691 case NVPTXISD::Suld1DV2I16Clamp: return "NVPTXISD::Suld1DV2I16Clamp";
692 case NVPTXISD::Suld1DV2I32Clamp: return "NVPTXISD::Suld1DV2I32Clamp";
693 case NVPTXISD::Suld1DV2I64Clamp: return "NVPTXISD::Suld1DV2I64Clamp";
694 case NVPTXISD::Suld1DV4I8Clamp: return "NVPTXISD::Suld1DV4I8Clamp";
695 case NVPTXISD::Suld1DV4I16Clamp: return "NVPTXISD::Suld1DV4I16Clamp";
696 case NVPTXISD::Suld1DV4I32Clamp: return "NVPTXISD::Suld1DV4I32Clamp";
698 case NVPTXISD::Suld1DArrayI8Clamp: return "NVPTXISD::Suld1DArrayI8Clamp";
699 case NVPTXISD::Suld1DArrayI16Clamp: return "NVPTXISD::Suld1DArrayI16Clamp";
700 case NVPTXISD::Suld1DArrayI32Clamp: return "NVPTXISD::Suld1DArrayI32Clamp";
701 case NVPTXISD::Suld1DArrayI64Clamp: return "NVPTXISD::Suld1DArrayI64Clamp";
702 case NVPTXISD::Suld1DArrayV2I8Clamp: return "NVPTXISD::Suld1DArrayV2I8Clamp";
703 case NVPTXISD::Suld1DArrayV2I16Clamp:return "NVPTXISD::Suld1DArrayV2I16Clamp";
704 case NVPTXISD::Suld1DArrayV2I32Clamp:return "NVPTXISD::Suld1DArrayV2I32Clamp";
705 case NVPTXISD::Suld1DArrayV2I64Clamp:return "NVPTXISD::Suld1DArrayV2I64Clamp";
706 case NVPTXISD::Suld1DArrayV4I8Clamp: return "NVPTXISD::Suld1DArrayV4I8Clamp";
707 case NVPTXISD::Suld1DArrayV4I16Clamp:return "NVPTXISD::Suld1DArrayV4I16Clamp";
708 case NVPTXISD::Suld1DArrayV4I32Clamp:return "NVPTXISD::Suld1DArrayV4I32Clamp";
710 case NVPTXISD::Suld2DI8Clamp: return "NVPTXISD::Suld2DI8Clamp";
711 case NVPTXISD::Suld2DI16Clamp: return "NVPTXISD::Suld2DI16Clamp";
712 case NVPTXISD::Suld2DI32Clamp: return "NVPTXISD::Suld2DI32Clamp";
713 case NVPTXISD::Suld2DI64Clamp: return "NVPTXISD::Suld2DI64Clamp";
714 case NVPTXISD::Suld2DV2I8Clamp: return "NVPTXISD::Suld2DV2I8Clamp";
715 case NVPTXISD::Suld2DV2I16Clamp: return "NVPTXISD::Suld2DV2I16Clamp";
716 case NVPTXISD::Suld2DV2I32Clamp: return "NVPTXISD::Suld2DV2I32Clamp";
717 case NVPTXISD::Suld2DV2I64Clamp: return "NVPTXISD::Suld2DV2I64Clamp";
718 case NVPTXISD::Suld2DV4I8Clamp: return "NVPTXISD::Suld2DV4I8Clamp";
719 case NVPTXISD::Suld2DV4I16Clamp: return "NVPTXISD::Suld2DV4I16Clamp";
720 case NVPTXISD::Suld2DV4I32Clamp: return "NVPTXISD::Suld2DV4I32Clamp";
722 case NVPTXISD::Suld2DArrayI8Clamp: return "NVPTXISD::Suld2DArrayI8Clamp";
723 case NVPTXISD::Suld2DArrayI16Clamp: return "NVPTXISD::Suld2DArrayI16Clamp";
724 case NVPTXISD::Suld2DArrayI32Clamp: return "NVPTXISD::Suld2DArrayI32Clamp";
725 case NVPTXISD::Suld2DArrayI64Clamp: return "NVPTXISD::Suld2DArrayI64Clamp";
726 case NVPTXISD::Suld2DArrayV2I8Clamp: return "NVPTXISD::Suld2DArrayV2I8Clamp";
727 case NVPTXISD::Suld2DArrayV2I16Clamp:return "NVPTXISD::Suld2DArrayV2I16Clamp";
728 case NVPTXISD::Suld2DArrayV2I32Clamp:return "NVPTXISD::Suld2DArrayV2I32Clamp";
729 case NVPTXISD::Suld2DArrayV2I64Clamp:return "NVPTXISD::Suld2DArrayV2I64Clamp";
730 case NVPTXISD::Suld2DArrayV4I8Clamp: return "NVPTXISD::Suld2DArrayV4I8Clamp";
731 case NVPTXISD::Suld2DArrayV4I16Clamp:return "NVPTXISD::Suld2DArrayV4I16Clamp";
732 case NVPTXISD::Suld2DArrayV4I32Clamp:return "NVPTXISD::Suld2DArrayV4I32Clamp";
734 case NVPTXISD::Suld3DI8Clamp: return "NVPTXISD::Suld3DI8Clamp";
735 case NVPTXISD::Suld3DI16Clamp: return "NVPTXISD::Suld3DI16Clamp";
736 case NVPTXISD::Suld3DI32Clamp: return "NVPTXISD::Suld3DI32Clamp";
737 case NVPTXISD::Suld3DI64Clamp: return "NVPTXISD::Suld3DI64Clamp";
738 case NVPTXISD::Suld3DV2I8Clamp: return "NVPTXISD::Suld3DV2I8Clamp";
739 case NVPTXISD::Suld3DV2I16Clamp: return "NVPTXISD::Suld3DV2I16Clamp";
740 case NVPTXISD::Suld3DV2I32Clamp: return "NVPTXISD::Suld3DV2I32Clamp";
741 case NVPTXISD::Suld3DV2I64Clamp: return "NVPTXISD::Suld3DV2I64Clamp";
742 case NVPTXISD::Suld3DV4I8Clamp: return "NVPTXISD::Suld3DV4I8Clamp";
743 case NVPTXISD::Suld3DV4I16Clamp: return "NVPTXISD::Suld3DV4I16Clamp";
744 case NVPTXISD::Suld3DV4I32Clamp: return "NVPTXISD::Suld3DV4I32Clamp";
746 case NVPTXISD::Suld1DI8Trap: return "NVPTXISD::Suld1DI8Trap";
747 case NVPTXISD::Suld1DI16Trap: return "NVPTXISD::Suld1DI16Trap";
748 case NVPTXISD::Suld1DI32Trap: return "NVPTXISD::Suld1DI32Trap";
749 case NVPTXISD::Suld1DI64Trap: return "NVPTXISD::Suld1DI64Trap";
750 case NVPTXISD::Suld1DV2I8Trap: return "NVPTXISD::Suld1DV2I8Trap";
751 case NVPTXISD::Suld1DV2I16Trap: return "NVPTXISD::Suld1DV2I16Trap";
752 case NVPTXISD::Suld1DV2I32Trap: return "NVPTXISD::Suld1DV2I32Trap";
753 case NVPTXISD::Suld1DV2I64Trap: return "NVPTXISD::Suld1DV2I64Trap";
754 case NVPTXISD::Suld1DV4I8Trap: return "NVPTXISD::Suld1DV4I8Trap";
755 case NVPTXISD::Suld1DV4I16Trap: return "NVPTXISD::Suld1DV4I16Trap";
756 case NVPTXISD::Suld1DV4I32Trap: return "NVPTXISD::Suld1DV4I32Trap";
758 case NVPTXISD::Suld1DArrayI8Trap: return "NVPTXISD::Suld1DArrayI8Trap";
759 case NVPTXISD::Suld1DArrayI16Trap: return "NVPTXISD::Suld1DArrayI16Trap";
760 case NVPTXISD::Suld1DArrayI32Trap: return "NVPTXISD::Suld1DArrayI32Trap";
761 case NVPTXISD::Suld1DArrayI64Trap: return "NVPTXISD::Suld1DArrayI64Trap";
762 case NVPTXISD::Suld1DArrayV2I8Trap: return "NVPTXISD::Suld1DArrayV2I8Trap";
763 case NVPTXISD::Suld1DArrayV2I16Trap: return "NVPTXISD::Suld1DArrayV2I16Trap";
764 case NVPTXISD::Suld1DArrayV2I32Trap: return "NVPTXISD::Suld1DArrayV2I32Trap";
765 case NVPTXISD::Suld1DArrayV2I64Trap: return "NVPTXISD::Suld1DArrayV2I64Trap";
766 case NVPTXISD::Suld1DArrayV4I8Trap: return "NVPTXISD::Suld1DArrayV4I8Trap";
767 case NVPTXISD::Suld1DArrayV4I16Trap: return "NVPTXISD::Suld1DArrayV4I16Trap";
768 case NVPTXISD::Suld1DArrayV4I32Trap: return "NVPTXISD::Suld1DArrayV4I32Trap";
770 case NVPTXISD::Suld2DI8Trap: return "NVPTXISD::Suld2DI8Trap";
771 case NVPTXISD::Suld2DI16Trap: return "NVPTXISD::Suld2DI16Trap";
772 case NVPTXISD::Suld2DI32Trap: return "NVPTXISD::Suld2DI32Trap";
773 case NVPTXISD::Suld2DI64Trap: return "NVPTXISD::Suld2DI64Trap";
774 case NVPTXISD::Suld2DV2I8Trap: return "NVPTXISD::Suld2DV2I8Trap";
775 case NVPTXISD::Suld2DV2I16Trap: return "NVPTXISD::Suld2DV2I16Trap";
776 case NVPTXISD::Suld2DV2I32Trap: return "NVPTXISD::Suld2DV2I32Trap";
777 case NVPTXISD::Suld2DV2I64Trap: return "NVPTXISD::Suld2DV2I64Trap";
778 case NVPTXISD::Suld2DV4I8Trap: return "NVPTXISD::Suld2DV4I8Trap";
779 case NVPTXISD::Suld2DV4I16Trap: return "NVPTXISD::Suld2DV4I16Trap";
780 case NVPTXISD::Suld2DV4I32Trap: return "NVPTXISD::Suld2DV4I32Trap";
782 case NVPTXISD::Suld2DArrayI8Trap: return "NVPTXISD::Suld2DArrayI8Trap";
783 case NVPTXISD::Suld2DArrayI16Trap: return "NVPTXISD::Suld2DArrayI16Trap";
784 case NVPTXISD::Suld2DArrayI32Trap: return "NVPTXISD::Suld2DArrayI32Trap";
785 case NVPTXISD::Suld2DArrayI64Trap: return "NVPTXISD::Suld2DArrayI64Trap";
786 case NVPTXISD::Suld2DArrayV2I8Trap: return "NVPTXISD::Suld2DArrayV2I8Trap";
787 case NVPTXISD::Suld2DArrayV2I16Trap: return "NVPTXISD::Suld2DArrayV2I16Trap";
788 case NVPTXISD::Suld2DArrayV2I32Trap: return "NVPTXISD::Suld2DArrayV2I32Trap";
789 case NVPTXISD::Suld2DArrayV2I64Trap: return "NVPTXISD::Suld2DArrayV2I64Trap";
790 case NVPTXISD::Suld2DArrayV4I8Trap: return "NVPTXISD::Suld2DArrayV4I8Trap";
791 case NVPTXISD::Suld2DArrayV4I16Trap: return "NVPTXISD::Suld2DArrayV4I16Trap";
792 case NVPTXISD::Suld2DArrayV4I32Trap: return "NVPTXISD::Suld2DArrayV4I32Trap";
794 case NVPTXISD::Suld3DI8Trap: return "NVPTXISD::Suld3DI8Trap";
795 case NVPTXISD::Suld3DI16Trap: return "NVPTXISD::Suld3DI16Trap";
796 case NVPTXISD::Suld3DI32Trap: return "NVPTXISD::Suld3DI32Trap";
797 case NVPTXISD::Suld3DI64Trap: return "NVPTXISD::Suld3DI64Trap";
798 case NVPTXISD::Suld3DV2I8Trap: return "NVPTXISD::Suld3DV2I8Trap";
799 case NVPTXISD::Suld3DV2I16Trap: return "NVPTXISD::Suld3DV2I16Trap";
800 case NVPTXISD::Suld3DV2I32Trap: return "NVPTXISD::Suld3DV2I32Trap";
801 case NVPTXISD::Suld3DV2I64Trap: return "NVPTXISD::Suld3DV2I64Trap";
802 case NVPTXISD::Suld3DV4I8Trap: return "NVPTXISD::Suld3DV4I8Trap";
803 case NVPTXISD::Suld3DV4I16Trap: return "NVPTXISD::Suld3DV4I16Trap";
804 case NVPTXISD::Suld3DV4I32Trap: return "NVPTXISD::Suld3DV4I32Trap";
806 case NVPTXISD::Suld1DI8Zero: return "NVPTXISD::Suld1DI8Zero";
807 case NVPTXISD::Suld1DI16Zero: return "NVPTXISD::Suld1DI16Zero";
808 case NVPTXISD::Suld1DI32Zero: return "NVPTXISD::Suld1DI32Zero";
809 case NVPTXISD::Suld1DI64Zero: return "NVPTXISD::Suld1DI64Zero";
810 case NVPTXISD::Suld1DV2I8Zero: return "NVPTXISD::Suld1DV2I8Zero";
811 case NVPTXISD::Suld1DV2I16Zero: return "NVPTXISD::Suld1DV2I16Zero";
812 case NVPTXISD::Suld1DV2I32Zero: return "NVPTXISD::Suld1DV2I32Zero";
813 case NVPTXISD::Suld1DV2I64Zero: return "NVPTXISD::Suld1DV2I64Zero";
814 case NVPTXISD::Suld1DV4I8Zero: return "NVPTXISD::Suld1DV4I8Zero";
815 case NVPTXISD::Suld1DV4I16Zero: return "NVPTXISD::Suld1DV4I16Zero";
816 case NVPTXISD::Suld1DV4I32Zero: return "NVPTXISD::Suld1DV4I32Zero";
818 case NVPTXISD::Suld1DArrayI8Zero: return "NVPTXISD::Suld1DArrayI8Zero";
819 case NVPTXISD::Suld1DArrayI16Zero: return "NVPTXISD::Suld1DArrayI16Zero";
820 case NVPTXISD::Suld1DArrayI32Zero: return "NVPTXISD::Suld1DArrayI32Zero";
821 case NVPTXISD::Suld1DArrayI64Zero: return "NVPTXISD::Suld1DArrayI64Zero";
822 case NVPTXISD::Suld1DArrayV2I8Zero: return "NVPTXISD::Suld1DArrayV2I8Zero";
823 case NVPTXISD::Suld1DArrayV2I16Zero: return "NVPTXISD::Suld1DArrayV2I16Zero";
824 case NVPTXISD::Suld1DArrayV2I32Zero: return "NVPTXISD::Suld1DArrayV2I32Zero";
825 case NVPTXISD::Suld1DArrayV2I64Zero: return "NVPTXISD::Suld1DArrayV2I64Zero";
826 case NVPTXISD::Suld1DArrayV4I8Zero: return "NVPTXISD::Suld1DArrayV4I8Zero";
827 case NVPTXISD::Suld1DArrayV4I16Zero: return "NVPTXISD::Suld1DArrayV4I16Zero";
828 case NVPTXISD::Suld1DArrayV4I32Zero: return "NVPTXISD::Suld1DArrayV4I32Zero";
830 case NVPTXISD::Suld2DI8Zero: return "NVPTXISD::Suld2DI8Zero";
831 case NVPTXISD::Suld2DI16Zero: return "NVPTXISD::Suld2DI16Zero";
832 case NVPTXISD::Suld2DI32Zero: return "NVPTXISD::Suld2DI32Zero";
833 case NVPTXISD::Suld2DI64Zero: return "NVPTXISD::Suld2DI64Zero";
834 case NVPTXISD::Suld2DV2I8Zero: return "NVPTXISD::Suld2DV2I8Zero";
835 case NVPTXISD::Suld2DV2I16Zero: return "NVPTXISD::Suld2DV2I16Zero";
836 case NVPTXISD::Suld2DV2I32Zero: return "NVPTXISD::Suld2DV2I32Zero";
837 case NVPTXISD::Suld2DV2I64Zero: return "NVPTXISD::Suld2DV2I64Zero";
838 case NVPTXISD::Suld2DV4I8Zero: return "NVPTXISD::Suld2DV4I8Zero";
839 case NVPTXISD::Suld2DV4I16Zero: return "NVPTXISD::Suld2DV4I16Zero";
840 case NVPTXISD::Suld2DV4I32Zero: return "NVPTXISD::Suld2DV4I32Zero";
842 case NVPTXISD::Suld2DArrayI8Zero: return "NVPTXISD::Suld2DArrayI8Zero";
843 case NVPTXISD::Suld2DArrayI16Zero: return "NVPTXISD::Suld2DArrayI16Zero";
844 case NVPTXISD::Suld2DArrayI32Zero: return "NVPTXISD::Suld2DArrayI32Zero";
845 case NVPTXISD::Suld2DArrayI64Zero: return "NVPTXISD::Suld2DArrayI64Zero";
846 case NVPTXISD::Suld2DArrayV2I8Zero: return "NVPTXISD::Suld2DArrayV2I8Zero";
847 case NVPTXISD::Suld2DArrayV2I16Zero: return "NVPTXISD::Suld2DArrayV2I16Zero";
848 case NVPTXISD::Suld2DArrayV2I32Zero: return "NVPTXISD::Suld2DArrayV2I32Zero";
849 case NVPTXISD::Suld2DArrayV2I64Zero: return "NVPTXISD::Suld2DArrayV2I64Zero";
850 case NVPTXISD::Suld2DArrayV4I8Zero: return "NVPTXISD::Suld2DArrayV4I8Zero";
851 case NVPTXISD::Suld2DArrayV4I16Zero: return "NVPTXISD::Suld2DArrayV4I16Zero";
852 case NVPTXISD::Suld2DArrayV4I32Zero: return "NVPTXISD::Suld2DArrayV4I32Zero";
854 case NVPTXISD::Suld3DI8Zero: return "NVPTXISD::Suld3DI8Zero";
855 case NVPTXISD::Suld3DI16Zero: return "NVPTXISD::Suld3DI16Zero";
856 case NVPTXISD::Suld3DI32Zero: return "NVPTXISD::Suld3DI32Zero";
857 case NVPTXISD::Suld3DI64Zero: return "NVPTXISD::Suld3DI64Zero";
858 case NVPTXISD::Suld3DV2I8Zero: return "NVPTXISD::Suld3DV2I8Zero";
859 case NVPTXISD::Suld3DV2I16Zero: return "NVPTXISD::Suld3DV2I16Zero";
860 case NVPTXISD::Suld3DV2I32Zero: return "NVPTXISD::Suld3DV2I32Zero";
861 case NVPTXISD::Suld3DV2I64Zero: return "NVPTXISD::Suld3DV2I64Zero";
862 case NVPTXISD::Suld3DV4I8Zero: return "NVPTXISD::Suld3DV4I8Zero";
863 case NVPTXISD::Suld3DV4I16Zero: return "NVPTXISD::Suld3DV4I16Zero";
864 case NVPTXISD::Suld3DV4I32Zero: return "NVPTXISD::Suld3DV4I32Zero";
869 TargetLoweringBase::LegalizeTypeAction
870 NVPTXTargetLowering::getPreferredVectorAction(EVT VT) const {
871 if (VT.getVectorNumElements() != 1 && VT.getScalarType() == MVT::i1)
872 return TypeSplitVector;
874 return TargetLoweringBase::getPreferredVectorAction(VT);
878 NVPTXTargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
880 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
881 Op = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
882 return DAG.getNode(NVPTXISD::Wrapper, dl, getPointerTy(), Op);
886 NVPTXTargetLowering::getPrototype(Type *retTy, const ArgListTy &Args,
887 const SmallVectorImpl<ISD::OutputArg> &Outs,
888 unsigned retAlignment,
889 const ImmutableCallSite *CS) const {
891 bool isABI = (STI.getSmVersion() >= 20);
892 assert(isABI && "Non-ABI compilation is not supported");
897 O << "prototype_" << uniqueCallSite << " : .callprototype ";
899 if (retTy->getTypeID() == Type::VoidTyID) {
903 if (retTy->isFloatingPointTy() || retTy->isIntegerTy()) {
905 if (const IntegerType *ITy = dyn_cast<IntegerType>(retTy)) {
906 size = ITy->getBitWidth();
910 assert(retTy->isFloatingPointTy() &&
911 "Floating point type expected here");
912 size = retTy->getPrimitiveSizeInBits();
915 O << ".param .b" << size << " _";
916 } else if (isa<PointerType>(retTy)) {
917 O << ".param .b" << getPointerTy().getSizeInBits() << " _";
918 } else if ((retTy->getTypeID() == Type::StructTyID) ||
919 isa<VectorType>(retTy)) {
920 O << ".param .align "
923 << getDataLayout()->getTypeAllocSize(retTy) << "]";
925 llvm_unreachable("Unknown return type");
932 MVT thePointerTy = getPointerTy();
935 for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
936 Type *Ty = Args[i].Ty;
942 if (!Outs[OIdx].Flags.isByVal()) {
943 if (Ty->isAggregateType() || Ty->isVectorTy()) {
945 const CallInst *CallI = cast<CallInst>(CS->getInstruction());
946 const DataLayout *TD = getDataLayout();
947 // +1 because index 0 is reserved for return type alignment
948 if (!llvm::getAlign(*CallI, i + 1, align))
949 align = TD->getABITypeAlignment(Ty);
950 unsigned sz = TD->getTypeAllocSize(Ty);
951 O << ".param .align " << align << " .b8 ";
953 O << "[" << sz << "]";
954 // update the index for Outs
955 SmallVector<EVT, 16> vtparts;
956 ComputeValueVTs(*this, Ty, vtparts);
957 if (unsigned len = vtparts.size())
961 // i8 types in IR will be i16 types in SDAG
962 assert((getValueType(Ty) == Outs[OIdx].VT ||
963 (getValueType(Ty) == MVT::i8 && Outs[OIdx].VT == MVT::i16)) &&
964 "type mismatch between callee prototype and arguments");
967 if (isa<IntegerType>(Ty)) {
968 sz = cast<IntegerType>(Ty)->getBitWidth();
971 } else if (isa<PointerType>(Ty))
972 sz = thePointerTy.getSizeInBits();
974 sz = Ty->getPrimitiveSizeInBits();
975 O << ".param .b" << sz << " ";
979 const PointerType *PTy = dyn_cast<PointerType>(Ty);
980 assert(PTy && "Param with byval attribute should be a pointer type");
981 Type *ETy = PTy->getElementType();
983 unsigned align = Outs[OIdx].Flags.getByValAlign();
984 unsigned sz = getDataLayout()->getTypeAllocSize(ETy);
985 O << ".param .align " << align << " .b8 ";
987 O << "[" << sz << "]";
994 NVPTXTargetLowering::getArgumentAlignment(SDValue Callee,
995 const ImmutableCallSite *CS,
997 unsigned Idx) const {
998 const DataLayout *TD = getDataLayout();
1000 const Value *DirectCallee = CS->getCalledFunction();
1002 if (!DirectCallee) {
1003 // We don't have a direct function symbol, but that may be because of
1004 // constant cast instructions in the call.
1005 const Instruction *CalleeI = CS->getInstruction();
1006 assert(CalleeI && "Call target is not a function or derived value?");
1008 // With bitcast'd call targets, the instruction will be the call
1009 if (isa<CallInst>(CalleeI)) {
1010 // Check if we have call alignment metadata
1011 if (llvm::getAlign(*cast<CallInst>(CalleeI), Idx, Align))
1014 const Value *CalleeV = cast<CallInst>(CalleeI)->getCalledValue();
1015 // Ignore any bitcast instructions
1016 while(isa<ConstantExpr>(CalleeV)) {
1017 const ConstantExpr *CE = cast<ConstantExpr>(CalleeV);
1020 // Look through the bitcast
1021 CalleeV = cast<ConstantExpr>(CalleeV)->getOperand(0);
1024 // We have now looked past all of the bitcasts. Do we finally have a
1026 if (isa<Function>(CalleeV))
1027 DirectCallee = CalleeV;
1031 // Check for function alignment information if we found that the
1032 // ultimate target is a Function
1034 if (llvm::getAlign(*cast<Function>(DirectCallee), Idx, Align))
1037 // Call is indirect or alignment information is not available, fall back to
1038 // the ABI type alignment
1039 return TD->getABITypeAlignment(Ty);
1042 SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1043 SmallVectorImpl<SDValue> &InVals) const {
1044 SelectionDAG &DAG = CLI.DAG;
1046 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1047 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1048 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1049 SDValue Chain = CLI.Chain;
1050 SDValue Callee = CLI.Callee;
1051 bool &isTailCall = CLI.IsTailCall;
1052 ArgListTy &Args = CLI.getArgs();
1053 Type *retTy = CLI.RetTy;
1054 ImmutableCallSite *CS = CLI.CS;
1056 bool isABI = (STI.getSmVersion() >= 20);
1057 assert(isABI && "Non-ABI compilation is not supported");
1060 const DataLayout *TD = getDataLayout();
1061 MachineFunction &MF = DAG.getMachineFunction();
1062 const Function *F = MF.getFunction();
1064 SDValue tempChain = Chain;
1065 Chain = DAG.getCALLSEQ_START(Chain,
1066 DAG.getIntPtrConstant(uniqueCallSite, dl, true),
1068 SDValue InFlag = Chain.getValue(1);
1070 unsigned paramCount = 0;
1071 // Args.size() and Outs.size() need not match.
1072 // Outs.size() will be larger
1073 // * if there is an aggregate argument with multiple fields (each field
1074 // showing up separately in Outs)
1075 // * if there is a vector argument with more than typical vector-length
1076 // elements (generally if more than 4) where each vector element is
1077 // individually present in Outs.
1078 // So a different index should be used for indexing into Outs/OutVals.
1079 // See similar issue in LowerFormalArguments.
1081 // Declare the .params or .reg need to pass values
1083 for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
1084 EVT VT = Outs[OIdx].VT;
1085 Type *Ty = Args[i].Ty;
1087 if (!Outs[OIdx].Flags.isByVal()) {
1088 if (Ty->isAggregateType()) {
1090 SmallVector<EVT, 16> vtparts;
1091 SmallVector<uint64_t, 16> Offsets;
1092 ComputePTXValueVTs(*this, Ty, vtparts, &Offsets, 0);
1094 unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
1095 // declare .param .align <align> .b8 .param<n>[<size>];
1096 unsigned sz = TD->getTypeAllocSize(Ty);
1097 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1098 SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, dl,
1100 DAG.getConstant(paramCount, dl, MVT::i32),
1101 DAG.getConstant(sz, dl, MVT::i32),
1103 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1105 InFlag = Chain.getValue(1);
1106 for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
1107 EVT elemtype = vtparts[j];
1108 unsigned ArgAlign = GreatestCommonDivisor64(align, Offsets[j]);
1109 if (elemtype.isInteger() && (sz < 8))
1111 SDValue StVal = OutVals[OIdx];
1112 if (elemtype.getSizeInBits() < 16) {
1113 StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
1115 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1116 SDValue CopyParamOps[] = { Chain,
1117 DAG.getConstant(paramCount, dl, MVT::i32),
1118 DAG.getConstant(Offsets[j], dl, MVT::i32),
1120 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
1121 CopyParamVTs, CopyParamOps,
1122 elemtype, MachinePointerInfo(),
1124 InFlag = Chain.getValue(1);
1127 if (vtparts.size() > 0)
1132 if (Ty->isVectorTy()) {
1133 EVT ObjectVT = getValueType(Ty);
1134 unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
1135 // declare .param .align <align> .b8 .param<n>[<size>];
1136 unsigned sz = TD->getTypeAllocSize(Ty);
1137 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1138 SDValue DeclareParamOps[] = { Chain,
1139 DAG.getConstant(align, dl, MVT::i32),
1140 DAG.getConstant(paramCount, dl, MVT::i32),
1141 DAG.getConstant(sz, dl, MVT::i32),
1143 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1145 InFlag = Chain.getValue(1);
1146 unsigned NumElts = ObjectVT.getVectorNumElements();
1147 EVT EltVT = ObjectVT.getVectorElementType();
1149 bool NeedExtend = false;
1150 if (EltVT.getSizeInBits() < 16) {
1157 SDValue Elt = OutVals[OIdx++];
1159 Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt);
1161 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1162 SDValue CopyParamOps[] = { Chain,
1163 DAG.getConstant(paramCount, dl, MVT::i32),
1164 DAG.getConstant(0, dl, MVT::i32), Elt,
1166 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
1167 CopyParamVTs, CopyParamOps,
1168 MemVT, MachinePointerInfo());
1169 InFlag = Chain.getValue(1);
1170 } else if (NumElts == 2) {
1171 SDValue Elt0 = OutVals[OIdx++];
1172 SDValue Elt1 = OutVals[OIdx++];
1174 Elt0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt0);
1175 Elt1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt1);
1178 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1179 SDValue CopyParamOps[] = { Chain,
1180 DAG.getConstant(paramCount, dl, MVT::i32),
1181 DAG.getConstant(0, dl, MVT::i32), Elt0,
1183 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParamV2, dl,
1184 CopyParamVTs, CopyParamOps,
1185 MemVT, MachinePointerInfo());
1186 InFlag = Chain.getValue(1);
1188 unsigned curOffset = 0;
1190 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
1192 // vector will be expanded to a power of 2 elements, so we know we can
1193 // always round up to the next multiple of 4 when creating the vector
1195 // e.g. 4 elem => 1 st.v4
1196 // 6 elem => 2 st.v4
1197 // 8 elem => 2 st.v4
1198 // 11 elem => 3 st.v4
1199 unsigned VecSize = 4;
1200 if (EltVT.getSizeInBits() == 64)
1203 // This is potentially only part of a vector, so assume all elements
1204 // are packed together.
1205 unsigned PerStoreOffset = MemVT.getStoreSizeInBits() / 8 * VecSize;
1207 for (unsigned i = 0; i < NumElts; i += VecSize) {
1210 SmallVector<SDValue, 8> Ops;
1211 Ops.push_back(Chain);
1212 Ops.push_back(DAG.getConstant(paramCount, dl, MVT::i32));
1213 Ops.push_back(DAG.getConstant(curOffset, dl, MVT::i32));
1215 unsigned Opc = NVPTXISD::StoreParamV2;
1217 StoreVal = OutVals[OIdx++];
1219 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1220 Ops.push_back(StoreVal);
1222 if (i + 1 < NumElts) {
1223 StoreVal = OutVals[OIdx++];
1226 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1228 StoreVal = DAG.getUNDEF(EltVT);
1230 Ops.push_back(StoreVal);
1233 Opc = NVPTXISD::StoreParamV4;
1234 if (i + 2 < NumElts) {
1235 StoreVal = OutVals[OIdx++];
1238 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1240 StoreVal = DAG.getUNDEF(EltVT);
1242 Ops.push_back(StoreVal);
1244 if (i + 3 < NumElts) {
1245 StoreVal = OutVals[OIdx++];
1248 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1250 StoreVal = DAG.getUNDEF(EltVT);
1252 Ops.push_back(StoreVal);
1255 Ops.push_back(InFlag);
1257 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1258 Chain = DAG.getMemIntrinsicNode(Opc, dl, CopyParamVTs, Ops,
1259 MemVT, MachinePointerInfo());
1260 InFlag = Chain.getValue(1);
1261 curOffset += PerStoreOffset;
1269 // for ABI, declare .param .b<size> .param<n>;
1270 unsigned sz = VT.getSizeInBits();
1271 bool needExtend = false;
1272 if (VT.isInteger()) {
1278 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1279 SDValue DeclareParamOps[] = { Chain,
1280 DAG.getConstant(paramCount, dl, MVT::i32),
1281 DAG.getConstant(sz, dl, MVT::i32),
1282 DAG.getConstant(0, dl, MVT::i32), InFlag };
1283 Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
1285 InFlag = Chain.getValue(1);
1286 SDValue OutV = OutVals[OIdx];
1288 // zext/sext i1 to i16
1289 unsigned opc = ISD::ZERO_EXTEND;
1290 if (Outs[OIdx].Flags.isSExt())
1291 opc = ISD::SIGN_EXTEND;
1292 OutV = DAG.getNode(opc, dl, MVT::i16, OutV);
1294 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1295 SDValue CopyParamOps[] = { Chain,
1296 DAG.getConstant(paramCount, dl, MVT::i32),
1297 DAG.getConstant(0, dl, MVT::i32), OutV,
1300 unsigned opcode = NVPTXISD::StoreParam;
1301 if (Outs[OIdx].Flags.isZExt())
1302 opcode = NVPTXISD::StoreParamU32;
1303 else if (Outs[OIdx].Flags.isSExt())
1304 opcode = NVPTXISD::StoreParamS32;
1305 Chain = DAG.getMemIntrinsicNode(opcode, dl, CopyParamVTs, CopyParamOps,
1306 VT, MachinePointerInfo());
1308 InFlag = Chain.getValue(1);
1313 SmallVector<EVT, 16> vtparts;
1314 SmallVector<uint64_t, 16> Offsets;
1315 const PointerType *PTy = dyn_cast<PointerType>(Args[i].Ty);
1316 assert(PTy && "Type of a byval parameter should be pointer");
1317 ComputePTXValueVTs(*this, PTy->getElementType(), vtparts, &Offsets, 0);
1319 // declare .param .align <align> .b8 .param<n>[<size>];
1320 unsigned sz = Outs[OIdx].Flags.getByValSize();
1321 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1322 unsigned ArgAlign = Outs[OIdx].Flags.getByValAlign();
1323 // The ByValAlign in the Outs[OIdx].Flags is alway set at this point,
1324 // so we don't need to worry about natural alignment or not.
1325 // See TargetLowering::LowerCallTo().
1326 SDValue DeclareParamOps[] = {
1327 Chain, DAG.getConstant(Outs[OIdx].Flags.getByValAlign(), dl, MVT::i32),
1328 DAG.getConstant(paramCount, dl, MVT::i32),
1329 DAG.getConstant(sz, dl, MVT::i32), InFlag
1331 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1333 InFlag = Chain.getValue(1);
1334 for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
1335 EVT elemtype = vtparts[j];
1336 int curOffset = Offsets[j];
1337 unsigned PartAlign = GreatestCommonDivisor64(ArgAlign, curOffset);
1339 DAG.getNode(ISD::ADD, dl, getPointerTy(), OutVals[OIdx],
1340 DAG.getConstant(curOffset, dl, getPointerTy()));
1341 SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
1342 MachinePointerInfo(), false, false, false,
1344 if (elemtype.getSizeInBits() < 16) {
1345 theVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, theVal);
1347 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1348 SDValue CopyParamOps[] = { Chain,
1349 DAG.getConstant(paramCount, dl, MVT::i32),
1350 DAG.getConstant(curOffset, dl, MVT::i32),
1352 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, CopyParamVTs,
1353 CopyParamOps, elemtype,
1354 MachinePointerInfo());
1356 InFlag = Chain.getValue(1);
1361 GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
1362 unsigned retAlignment = 0;
1365 if (Ins.size() > 0) {
1366 SmallVector<EVT, 16> resvtparts;
1367 ComputeValueVTs(*this, retTy, resvtparts);
1370 // .param .align 16 .b8 retval0[<size-in-bytes>], or
1371 // .param .b<size-in-bits> retval0
1372 unsigned resultsz = TD->getTypeAllocSizeInBits(retTy);
1373 // Emit ".param .b<size-in-bits> retval0" instead of byte arrays only for
1374 // these three types to match the logic in
1375 // NVPTXAsmPrinter::printReturnValStr and NVPTXTargetLowering::getPrototype.
1376 // Plus, this behavior is consistent with nvcc's.
1377 if (retTy->isFloatingPointTy() || retTy->isIntegerTy() ||
1378 retTy->isPointerTy()) {
1379 // Scalar needs to be at least 32bit wide
1382 SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1383 SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, dl, MVT::i32),
1384 DAG.getConstant(resultsz, dl, MVT::i32),
1385 DAG.getConstant(0, dl, MVT::i32), InFlag };
1386 Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
1388 InFlag = Chain.getValue(1);
1390 retAlignment = getArgumentAlignment(Callee, CS, retTy, 0);
1391 SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1392 SDValue DeclareRetOps[] = { Chain,
1393 DAG.getConstant(retAlignment, dl, MVT::i32),
1394 DAG.getConstant(resultsz / 8, dl, MVT::i32),
1395 DAG.getConstant(0, dl, MVT::i32), InFlag };
1396 Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs,
1398 InFlag = Chain.getValue(1);
1403 // This is indirect function call case : PTX requires a prototype of the
1405 // proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _);
1406 // to be emitted, and the label has to used as the last arg of call
1408 // The prototype is embedded in a string and put as the operand for a
1409 // CallPrototype SDNode which will print out to the value of the string.
1410 SDVTList ProtoVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1411 std::string Proto = getPrototype(retTy, Args, Outs, retAlignment, CS);
1412 const char *ProtoStr =
1413 nvTM->getManagedStrPool()->getManagedString(Proto.c_str())->c_str();
1414 SDValue ProtoOps[] = {
1415 Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InFlag,
1417 Chain = DAG.getNode(NVPTXISD::CallPrototype, dl, ProtoVTs, ProtoOps);
1418 InFlag = Chain.getValue(1);
1420 // Op to just print "call"
1421 SDVTList PrintCallVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1422 SDValue PrintCallOps[] = {
1423 Chain, DAG.getConstant((Ins.size() == 0) ? 0 : 1, dl, MVT::i32), InFlag
1425 Chain = DAG.getNode(Func ? (NVPTXISD::PrintCallUni) : (NVPTXISD::PrintCall),
1426 dl, PrintCallVTs, PrintCallOps);
1427 InFlag = Chain.getValue(1);
1429 // Ops to print out the function name
1430 SDVTList CallVoidVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1431 SDValue CallVoidOps[] = { Chain, Callee, InFlag };
1432 Chain = DAG.getNode(NVPTXISD::CallVoid, dl, CallVoidVTs, CallVoidOps);
1433 InFlag = Chain.getValue(1);
1435 // Ops to print out the param list
1436 SDVTList CallArgBeginVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1437 SDValue CallArgBeginOps[] = { Chain, InFlag };
1438 Chain = DAG.getNode(NVPTXISD::CallArgBegin, dl, CallArgBeginVTs,
1440 InFlag = Chain.getValue(1);
1442 for (unsigned i = 0, e = paramCount; i != e; ++i) {
1445 opcode = NVPTXISD::LastCallArg;
1447 opcode = NVPTXISD::CallArg;
1448 SDVTList CallArgVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1449 SDValue CallArgOps[] = { Chain, DAG.getConstant(1, dl, MVT::i32),
1450 DAG.getConstant(i, dl, MVT::i32), InFlag };
1451 Chain = DAG.getNode(opcode, dl, CallArgVTs, CallArgOps);
1452 InFlag = Chain.getValue(1);
1454 SDVTList CallArgEndVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1455 SDValue CallArgEndOps[] = { Chain,
1456 DAG.getConstant(Func ? 1 : 0, dl, MVT::i32),
1458 Chain = DAG.getNode(NVPTXISD::CallArgEnd, dl, CallArgEndVTs, CallArgEndOps);
1459 InFlag = Chain.getValue(1);
1462 SDVTList PrototypeVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1463 SDValue PrototypeOps[] = { Chain,
1464 DAG.getConstant(uniqueCallSite, dl, MVT::i32),
1466 Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps);
1467 InFlag = Chain.getValue(1);
1470 // Generate loads from param memory/moves from registers for result
1471 if (Ins.size() > 0) {
1472 if (retTy && retTy->isVectorTy()) {
1473 EVT ObjectVT = getValueType(retTy);
1474 unsigned NumElts = ObjectVT.getVectorNumElements();
1475 EVT EltVT = ObjectVT.getVectorElementType();
1476 assert(STI.getTargetLowering()->getNumRegisters(F->getContext(),
1477 ObjectVT) == NumElts &&
1478 "Vector was not scalarized");
1479 unsigned sz = EltVT.getSizeInBits();
1480 bool needTruncate = sz < 8;
1483 // Just a simple load
1484 SmallVector<EVT, 4> LoadRetVTs;
1485 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1486 // If loading i1/i8 result, generate
1490 LoadRetVTs.push_back(MVT::i16);
1492 LoadRetVTs.push_back(EltVT);
1493 LoadRetVTs.push_back(MVT::Other);
1494 LoadRetVTs.push_back(MVT::Glue);
1495 SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
1496 DAG.getConstant(0, dl, MVT::i32), InFlag};
1497 SDValue retval = DAG.getMemIntrinsicNode(
1498 NVPTXISD::LoadParam, dl,
1499 DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
1500 Chain = retval.getValue(1);
1501 InFlag = retval.getValue(2);
1502 SDValue Ret0 = retval;
1504 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Ret0);
1505 InVals.push_back(Ret0);
1506 } else if (NumElts == 2) {
1508 SmallVector<EVT, 4> LoadRetVTs;
1509 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1510 // If loading i1/i8 result, generate
1514 LoadRetVTs.push_back(MVT::i16);
1515 LoadRetVTs.push_back(MVT::i16);
1517 LoadRetVTs.push_back(EltVT);
1518 LoadRetVTs.push_back(EltVT);
1520 LoadRetVTs.push_back(MVT::Other);
1521 LoadRetVTs.push_back(MVT::Glue);
1522 SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
1523 DAG.getConstant(0, dl, MVT::i32), InFlag};
1524 SDValue retval = DAG.getMemIntrinsicNode(
1525 NVPTXISD::LoadParamV2, dl,
1526 DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
1527 Chain = retval.getValue(2);
1528 InFlag = retval.getValue(3);
1529 SDValue Ret0 = retval.getValue(0);
1530 SDValue Ret1 = retval.getValue(1);
1532 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret0);
1533 InVals.push_back(Ret0);
1534 Ret1 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret1);
1535 InVals.push_back(Ret1);
1537 InVals.push_back(Ret0);
1538 InVals.push_back(Ret1);
1541 // Split into N LoadV4
1543 unsigned VecSize = 4;
1544 unsigned Opc = NVPTXISD::LoadParamV4;
1545 if (EltVT.getSizeInBits() == 64) {
1547 Opc = NVPTXISD::LoadParamV2;
1549 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
1550 for (unsigned i = 0; i < NumElts; i += VecSize) {
1551 SmallVector<EVT, 8> LoadRetVTs;
1552 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1553 // If loading i1/i8 result, generate
1557 for (unsigned j = 0; j < VecSize; ++j)
1558 LoadRetVTs.push_back(MVT::i16);
1560 for (unsigned j = 0; j < VecSize; ++j)
1561 LoadRetVTs.push_back(EltVT);
1563 LoadRetVTs.push_back(MVT::Other);
1564 LoadRetVTs.push_back(MVT::Glue);
1565 SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
1566 DAG.getConstant(Ofst, dl, MVT::i32), InFlag};
1567 SDValue retval = DAG.getMemIntrinsicNode(
1568 Opc, dl, DAG.getVTList(LoadRetVTs),
1569 LoadRetOps, EltVT, MachinePointerInfo());
1571 Chain = retval.getValue(2);
1572 InFlag = retval.getValue(3);
1574 Chain = retval.getValue(4);
1575 InFlag = retval.getValue(5);
1578 for (unsigned j = 0; j < VecSize; ++j) {
1579 if (i + j >= NumElts)
1581 SDValue Elt = retval.getValue(j);
1583 Elt = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
1584 InVals.push_back(Elt);
1586 Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
1590 SmallVector<EVT, 16> VTs;
1591 SmallVector<uint64_t, 16> Offsets;
1592 ComputePTXValueVTs(*this, retTy, VTs, &Offsets, 0);
1593 assert(VTs.size() == Ins.size() && "Bad value decomposition");
1594 unsigned RetAlign = getArgumentAlignment(Callee, CS, retTy, 0);
1595 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
1596 unsigned sz = VTs[i].getSizeInBits();
1597 unsigned AlignI = GreatestCommonDivisor64(RetAlign, Offsets[i]);
1598 bool needTruncate = sz < 8;
1599 if (VTs[i].isInteger() && (sz < 8))
1602 SmallVector<EVT, 4> LoadRetVTs;
1603 EVT TheLoadType = VTs[i];
1604 if (retTy->isIntegerTy() &&
1605 TD->getTypeAllocSizeInBits(retTy) < 32) {
1606 // This is for integer types only, and specifically not for
1608 LoadRetVTs.push_back(MVT::i32);
1609 TheLoadType = MVT::i32;
1610 } else if (sz < 16) {
1611 // If loading i1/i8 result, generate
1613 // trunc i16 to i1/i8
1614 LoadRetVTs.push_back(MVT::i16);
1616 LoadRetVTs.push_back(Ins[i].VT);
1617 LoadRetVTs.push_back(MVT::Other);
1618 LoadRetVTs.push_back(MVT::Glue);
1620 SDValue LoadRetOps[] = {Chain, DAG.getConstant(1, dl, MVT::i32),
1621 DAG.getConstant(Offsets[i], dl, MVT::i32),
1623 SDValue retval = DAG.getMemIntrinsicNode(
1624 NVPTXISD::LoadParam, dl,
1625 DAG.getVTList(LoadRetVTs), LoadRetOps,
1626 TheLoadType, MachinePointerInfo(), AlignI);
1627 Chain = retval.getValue(1);
1628 InFlag = retval.getValue(2);
1629 SDValue Ret0 = retval.getValue(0);
1631 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, Ins[i].VT, Ret0);
1632 InVals.push_back(Ret0);
1637 Chain = DAG.getCALLSEQ_END(Chain,
1638 DAG.getIntPtrConstant(uniqueCallSite, dl, true),
1639 DAG.getIntPtrConstant(uniqueCallSite + 1, dl,
1644 // set isTailCall to false for now, until we figure out how to express
1645 // tail call optimization in PTX
1650 // By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack()
1651 // (see LegalizeDAG.cpp). This is slow and uses local memory.
1652 // We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5
1654 NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
1655 SDNode *Node = Op.getNode();
1657 SmallVector<SDValue, 8> Ops;
1658 unsigned NumOperands = Node->getNumOperands();
1659 for (unsigned i = 0; i < NumOperands; ++i) {
1660 SDValue SubOp = Node->getOperand(i);
1661 EVT VVT = SubOp.getNode()->getValueType(0);
1662 EVT EltVT = VVT.getVectorElementType();
1663 unsigned NumSubElem = VVT.getVectorNumElements();
1664 for (unsigned j = 0; j < NumSubElem; ++j) {
1665 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp,
1666 DAG.getIntPtrConstant(j, dl)));
1669 return DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Ops);
1672 /// LowerShiftRightParts - Lower SRL_PARTS, SRA_PARTS, which
1673 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1675 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1677 SDValue NVPTXTargetLowering::LowerShiftRightParts(SDValue Op,
1678 SelectionDAG &DAG) const {
1679 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1680 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
1682 EVT VT = Op.getValueType();
1683 unsigned VTBits = VT.getSizeInBits();
1685 SDValue ShOpLo = Op.getOperand(0);
1686 SDValue ShOpHi = Op.getOperand(1);
1687 SDValue ShAmt = Op.getOperand(2);
1688 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
1690 if (VTBits == 32 && STI.getSmVersion() >= 35) {
1692 // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1693 // {dHi, dLo} = {aHi, aLo} >> Amt
1695 // dLo = shf.r.clamp aLo, aHi, Amt
1697 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1698 SDValue Lo = DAG.getNode(NVPTXISD::FUN_SHFR_CLAMP, dl, VT, ShOpLo, ShOpHi,
1701 SDValue Ops[2] = { Lo, Hi };
1702 return DAG.getMergeValues(Ops, dl);
1706 // {dHi, dLo} = {aHi, aLo} >> Amt
1707 // - if (Amt>=size) then
1708 // dLo = aHi >> (Amt-size)
1709 // dHi = aHi >> Amt (this is either all 0 or all 1)
1711 // dLo = (aLo >>logic Amt) | (aHi << (size-Amt))
1714 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1715 DAG.getConstant(VTBits, dl, MVT::i32),
1717 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
1718 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1719 DAG.getConstant(VTBits, dl, MVT::i32));
1720 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
1721 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
1722 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
1724 SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
1725 DAG.getConstant(VTBits, dl, MVT::i32),
1727 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1728 SDValue Lo = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
1730 SDValue Ops[2] = { Lo, Hi };
1731 return DAG.getMergeValues(Ops, dl);
1735 /// LowerShiftLeftParts - Lower SHL_PARTS, which
1736 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1738 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1740 SDValue NVPTXTargetLowering::LowerShiftLeftParts(SDValue Op,
1741 SelectionDAG &DAG) const {
1742 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1743 assert(Op.getOpcode() == ISD::SHL_PARTS);
1745 EVT VT = Op.getValueType();
1746 unsigned VTBits = VT.getSizeInBits();
1748 SDValue ShOpLo = Op.getOperand(0);
1749 SDValue ShOpHi = Op.getOperand(1);
1750 SDValue ShAmt = Op.getOperand(2);
1752 if (VTBits == 32 && STI.getSmVersion() >= 35) {
1754 // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1755 // {dHi, dLo} = {aHi, aLo} << Amt
1756 // dHi = shf.l.clamp aLo, aHi, Amt
1759 SDValue Hi = DAG.getNode(NVPTXISD::FUN_SHFL_CLAMP, dl, VT, ShOpLo, ShOpHi,
1761 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1763 SDValue Ops[2] = { Lo, Hi };
1764 return DAG.getMergeValues(Ops, dl);
1768 // {dHi, dLo} = {aHi, aLo} << Amt
1769 // - if (Amt>=size) then
1770 // dLo = aLo << Amt (all 0)
1771 // dLo = aLo << (Amt-size)
1774 // dHi = (aHi << Amt) | (aLo >> (size-Amt))
1776 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1777 DAG.getConstant(VTBits, dl, MVT::i32),
1779 SDValue Tmp1 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
1780 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1781 DAG.getConstant(VTBits, dl, MVT::i32));
1782 SDValue Tmp2 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
1783 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
1784 SDValue TrueVal = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
1786 SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
1787 DAG.getConstant(VTBits, dl, MVT::i32),
1789 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1790 SDValue Hi = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
1792 SDValue Ops[2] = { Lo, Hi };
1793 return DAG.getMergeValues(Ops, dl);
1798 NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
1799 switch (Op.getOpcode()) {
1800 case ISD::RETURNADDR:
1802 case ISD::FRAMEADDR:
1804 case ISD::GlobalAddress:
1805 return LowerGlobalAddress(Op, DAG);
1806 case ISD::INTRINSIC_W_CHAIN:
1808 case ISD::BUILD_VECTOR:
1809 case ISD::EXTRACT_SUBVECTOR:
1811 case ISD::CONCAT_VECTORS:
1812 return LowerCONCAT_VECTORS(Op, DAG);
1814 return LowerSTORE(Op, DAG);
1816 return LowerLOAD(Op, DAG);
1817 case ISD::SHL_PARTS:
1818 return LowerShiftLeftParts(Op, DAG);
1819 case ISD::SRA_PARTS:
1820 case ISD::SRL_PARTS:
1821 return LowerShiftRightParts(Op, DAG);
1823 return LowerSelect(Op, DAG);
1825 llvm_unreachable("Custom lowering not defined for operation");
1829 SDValue NVPTXTargetLowering::LowerSelect(SDValue Op, SelectionDAG &DAG) const {
1830 SDValue Op0 = Op->getOperand(0);
1831 SDValue Op1 = Op->getOperand(1);
1832 SDValue Op2 = Op->getOperand(2);
1833 SDLoc DL(Op.getNode());
1835 assert(Op.getValueType() == MVT::i1 && "Custom lowering enabled only for i1");
1837 Op1 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op1);
1838 Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op2);
1839 SDValue Select = DAG.getNode(ISD::SELECT, DL, MVT::i32, Op0, Op1, Op2);
1840 SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Select);
1845 SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1846 if (Op.getValueType() == MVT::i1)
1847 return LowerLOADi1(Op, DAG);
1854 // v1 = ld i8* addr (-> i16)
1855 // v = trunc i16 to i1
1856 SDValue NVPTXTargetLowering::LowerLOADi1(SDValue Op, SelectionDAG &DAG) const {
1857 SDNode *Node = Op.getNode();
1858 LoadSDNode *LD = cast<LoadSDNode>(Node);
1860 assert(LD->getExtensionType() == ISD::NON_EXTLOAD);
1861 assert(Node->getValueType(0) == MVT::i1 &&
1862 "Custom lowering for i1 load only");
1864 DAG.getLoad(MVT::i16, dl, LD->getChain(), LD->getBasePtr(),
1865 LD->getPointerInfo(), LD->isVolatile(), LD->isNonTemporal(),
1866 LD->isInvariant(), LD->getAlignment());
1867 SDValue result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, newLD);
1868 // The legalizer (the caller) is expecting two values from the legalized
1869 // load, so we build a MergeValues node for it. See ExpandUnalignedLoad()
1870 // in LegalizeDAG.cpp which also uses MergeValues.
1871 SDValue Ops[] = { result, LD->getChain() };
1872 return DAG.getMergeValues(Ops, dl);
1875 SDValue NVPTXTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1876 EVT ValVT = Op.getOperand(1).getValueType();
1877 if (ValVT == MVT::i1)
1878 return LowerSTOREi1(Op, DAG);
1879 else if (ValVT.isVector())
1880 return LowerSTOREVector(Op, DAG);
1886 NVPTXTargetLowering::LowerSTOREVector(SDValue Op, SelectionDAG &DAG) const {
1887 SDNode *N = Op.getNode();
1888 SDValue Val = N->getOperand(1);
1890 EVT ValVT = Val.getValueType();
1892 if (ValVT.isVector()) {
1893 // We only handle "native" vector sizes for now, e.g. <4 x double> is not
1894 // legal. We can (and should) split that into 2 stores of <2 x double> here
1895 // but I'm leaving that as a TODO for now.
1896 if (!ValVT.isSimple())
1898 switch (ValVT.getSimpleVT().SimpleTy) {
1911 // This is a "native" vector type
1915 MemSDNode *MemSD = cast<MemSDNode>(N);
1916 const DataLayout *TD = getDataLayout();
1918 unsigned Align = MemSD->getAlignment();
1919 unsigned PrefAlign =
1920 TD->getPrefTypeAlignment(ValVT.getTypeForEVT(*DAG.getContext()));
1921 if (Align < PrefAlign) {
1922 // This store is not sufficiently aligned, so bail out and let this vector
1923 // store be scalarized. Note that we may still be able to emit smaller
1924 // vector stores. For example, if we are storing a <4 x float> with an
1925 // alignment of 8, this check will fail but the legalizer will try again
1926 // with 2 x <2 x float>, which will succeed with an alignment of 8.
1930 unsigned Opcode = 0;
1931 EVT EltVT = ValVT.getVectorElementType();
1932 unsigned NumElts = ValVT.getVectorNumElements();
1934 // Since StoreV2 is a target node, we cannot rely on DAG type legalization.
1935 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
1936 // stored type to i16 and propagate the "real" type as the memory type.
1937 bool NeedExt = false;
1938 if (EltVT.getSizeInBits() < 16)
1945 Opcode = NVPTXISD::StoreV2;
1948 Opcode = NVPTXISD::StoreV4;
1953 SmallVector<SDValue, 8> Ops;
1955 // First is the chain
1956 Ops.push_back(N->getOperand(0));
1958 // Then the split values
1959 for (unsigned i = 0; i < NumElts; ++i) {
1960 SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Val,
1961 DAG.getIntPtrConstant(i, DL));
1963 ExtVal = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i16, ExtVal);
1964 Ops.push_back(ExtVal);
1967 // Then any remaining arguments
1968 Ops.append(N->op_begin() + 2, N->op_end());
1970 SDValue NewSt = DAG.getMemIntrinsicNode(
1971 Opcode, DL, DAG.getVTList(MVT::Other), Ops,
1972 MemSD->getMemoryVT(), MemSD->getMemOperand());
1974 //return DCI.CombineTo(N, NewSt, true);
1983 // v1 = zxt v to i16
1985 SDValue NVPTXTargetLowering::LowerSTOREi1(SDValue Op, SelectionDAG &DAG) const {
1986 SDNode *Node = Op.getNode();
1988 StoreSDNode *ST = cast<StoreSDNode>(Node);
1989 SDValue Tmp1 = ST->getChain();
1990 SDValue Tmp2 = ST->getBasePtr();
1991 SDValue Tmp3 = ST->getValue();
1992 assert(Tmp3.getValueType() == MVT::i1 && "Custom lowering for i1 store only");
1993 unsigned Alignment = ST->getAlignment();
1994 bool isVolatile = ST->isVolatile();
1995 bool isNonTemporal = ST->isNonTemporal();
1996 Tmp3 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Tmp3);
1997 SDValue Result = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2,
1998 ST->getPointerInfo(), MVT::i8, isNonTemporal,
1999 isVolatile, Alignment);
2003 SDValue NVPTXTargetLowering::getExtSymb(SelectionDAG &DAG, const char *inname,
2004 int idx, EVT v) const {
2005 std::string *name = nvTM->getManagedStrPool()->getManagedString(inname);
2006 std::stringstream suffix;
2008 *name += suffix.str();
2009 return DAG.getTargetExternalSymbol(name->c_str(), v);
2013 NVPTXTargetLowering::getParamSymbol(SelectionDAG &DAG, int idx, EVT v) const {
2014 std::string ParamSym;
2015 raw_string_ostream ParamStr(ParamSym);
2017 ParamStr << DAG.getMachineFunction().getName() << "_param_" << idx;
2020 std::string *SavedStr =
2021 nvTM->getManagedStrPool()->getManagedString(ParamSym.c_str());
2022 return DAG.getTargetExternalSymbol(SavedStr->c_str(), v);
2025 SDValue NVPTXTargetLowering::getParamHelpSymbol(SelectionDAG &DAG, int idx) {
2026 return getExtSymb(DAG, ".HLPPARAM", idx);
2029 // Check to see if the kernel argument is image*_t or sampler_t
2031 bool llvm::isImageOrSamplerVal(const Value *arg, const Module *context) {
2032 static const char *const specialTypes[] = { "struct._image2d_t",
2033 "struct._image3d_t",
2034 "struct._sampler_t" };
2036 const Type *Ty = arg->getType();
2037 const PointerType *PTy = dyn_cast<PointerType>(Ty);
2045 const StructType *STy = dyn_cast<StructType>(PTy->getElementType());
2046 const std::string TypeName = STy && !STy->isLiteral() ? STy->getName() : "";
2048 for (int i = 0, e = array_lengthof(specialTypes); i != e; ++i)
2049 if (TypeName == specialTypes[i])
2055 SDValue NVPTXTargetLowering::LowerFormalArguments(
2056 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2057 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2058 SmallVectorImpl<SDValue> &InVals) const {
2059 MachineFunction &MF = DAG.getMachineFunction();
2060 const DataLayout *TD = getDataLayout();
2062 const Function *F = MF.getFunction();
2063 const AttributeSet &PAL = F->getAttributes();
2064 const TargetLowering *TLI = STI.getTargetLowering();
2066 SDValue Root = DAG.getRoot();
2067 std::vector<SDValue> OutChains;
2069 bool isKernel = llvm::isKernelFunction(*F);
2070 bool isABI = (STI.getSmVersion() >= 20);
2071 assert(isABI && "Non-ABI compilation is not supported");
2075 std::vector<Type *> argTypes;
2076 std::vector<const Argument *> theArgs;
2077 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
2079 theArgs.push_back(I);
2080 argTypes.push_back(I->getType());
2082 // argTypes.size() (or theArgs.size()) and Ins.size() need not match.
2083 // Ins.size() will be larger
2084 // * if there is an aggregate argument with multiple fields (each field
2085 // showing up separately in Ins)
2086 // * if there is a vector argument with more than typical vector-length
2087 // elements (generally if more than 4) where each vector element is
2088 // individually present in Ins.
2089 // So a different index should be used for indexing into Ins.
2090 // See similar issue in LowerCall.
2091 unsigned InsIdx = 0;
2094 for (unsigned i = 0, e = theArgs.size(); i != e; ++i, ++idx, ++InsIdx) {
2095 Type *Ty = argTypes[i];
2097 // If the kernel argument is image*_t or sampler_t, convert it to
2098 // a i32 constant holding the parameter position. This can later
2099 // matched in the AsmPrinter to output the correct mangled name.
2100 if (isImageOrSamplerVal(
2102 (theArgs[i]->getParent() ? theArgs[i]->getParent()->getParent()
2104 assert(isKernel && "Only kernels can have image/sampler params");
2105 InVals.push_back(DAG.getConstant(i + 1, dl, MVT::i32));
2109 if (theArgs[i]->use_empty()) {
2111 if (Ty->isAggregateType()) {
2112 SmallVector<EVT, 16> vtparts;
2114 ComputePTXValueVTs(*this, Ty, vtparts);
2115 assert(vtparts.size() > 0 && "empty aggregate type not expected");
2116 for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2118 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2121 if (vtparts.size() > 0)
2125 if (Ty->isVectorTy()) {
2126 EVT ObjectVT = getValueType(Ty);
2127 unsigned NumRegs = TLI->getNumRegisters(F->getContext(), ObjectVT);
2128 for (unsigned parti = 0; parti < NumRegs; ++parti) {
2129 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2136 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2140 // In the following cases, assign a node order of "idx+1"
2141 // to newly created nodes. The SDNodes for params have to
2142 // appear in the same order as their order of appearance
2143 // in the original function. "idx+1" holds that order.
2144 if (!PAL.hasAttribute(i + 1, Attribute::ByVal)) {
2145 if (Ty->isAggregateType()) {
2146 SmallVector<EVT, 16> vtparts;
2147 SmallVector<uint64_t, 16> offsets;
2149 // NOTE: Here, we lose the ability to issue vector loads for vectors
2150 // that are a part of a struct. This should be investigated in the
2152 ComputePTXValueVTs(*this, Ty, vtparts, &offsets, 0);
2153 assert(vtparts.size() > 0 && "empty aggregate type not expected");
2154 bool aggregateIsPacked = false;
2155 if (StructType *STy = llvm::dyn_cast<StructType>(Ty))
2156 aggregateIsPacked = STy->isPacked();
2158 SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2159 for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2161 EVT partVT = vtparts[parti];
2162 Value *srcValue = Constant::getNullValue(
2163 PointerType::get(partVT.getTypeForEVT(F->getContext()),
2164 llvm::ADDRESS_SPACE_PARAM));
2166 DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
2167 DAG.getConstant(offsets[parti], dl, getPointerTy()));
2168 unsigned partAlign =
2169 aggregateIsPacked ? 1
2170 : TD->getABITypeAlignment(
2171 partVT.getTypeForEVT(F->getContext()));
2173 if (Ins[InsIdx].VT.getSizeInBits() > partVT.getSizeInBits()) {
2174 ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ?
2175 ISD::SEXTLOAD : ISD::ZEXTLOAD;
2176 p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, srcAddr,
2177 MachinePointerInfo(srcValue), partVT, false,
2178 false, false, partAlign);
2180 p = DAG.getLoad(partVT, dl, Root, srcAddr,
2181 MachinePointerInfo(srcValue), false, false, false,
2185 p.getNode()->setIROrder(idx + 1);
2186 InVals.push_back(p);
2189 if (vtparts.size() > 0)
2193 if (Ty->isVectorTy()) {
2194 EVT ObjectVT = getValueType(Ty);
2195 SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2196 unsigned NumElts = ObjectVT.getVectorNumElements();
2197 assert(TLI->getNumRegisters(F->getContext(), ObjectVT) == NumElts &&
2198 "Vector was not scalarized");
2199 EVT EltVT = ObjectVT.getVectorElementType();
2204 // We only have one element, so just directly load it
2205 Value *SrcValue = Constant::getNullValue(PointerType::get(
2206 EltVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2207 SDValue P = DAG.getLoad(
2208 EltVT, dl, Root, Arg, MachinePointerInfo(SrcValue), false,
2210 TD->getABITypeAlignment(EltVT.getTypeForEVT(F->getContext())));
2212 P.getNode()->setIROrder(idx + 1);
2214 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
2215 P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P);
2216 InVals.push_back(P);
2218 } else if (NumElts == 2) {
2220 // f32,f32 = load ...
2221 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2);
2222 Value *SrcValue = Constant::getNullValue(PointerType::get(
2223 VecVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2224 SDValue P = DAG.getLoad(
2225 VecVT, dl, Root, Arg, MachinePointerInfo(SrcValue), false,
2227 TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
2229 P.getNode()->setIROrder(idx + 1);
2231 SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2232 DAG.getIntPtrConstant(0, dl));
2233 SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2234 DAG.getIntPtrConstant(1, dl));
2236 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) {
2237 Elt0 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt0);
2238 Elt1 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt1);
2241 InVals.push_back(Elt0);
2242 InVals.push_back(Elt1);
2246 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
2248 // vector will be expanded to a power of 2 elements, so we know we can
2249 // always round up to the next multiple of 4 when creating the vector
2251 // e.g. 4 elem => 1 ld.v4
2252 // 6 elem => 2 ld.v4
2253 // 8 elem => 2 ld.v4
2254 // 11 elem => 3 ld.v4
2255 unsigned VecSize = 4;
2256 if (EltVT.getSizeInBits() == 64) {
2259 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
2261 for (unsigned i = 0; i < NumElts; i += VecSize) {
2262 Value *SrcValue = Constant::getNullValue(
2263 PointerType::get(VecVT.getTypeForEVT(F->getContext()),
2264 llvm::ADDRESS_SPACE_PARAM));
2266 DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
2267 DAG.getConstant(Ofst, dl, getPointerTy()));
2268 SDValue P = DAG.getLoad(
2269 VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false,
2271 TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
2273 P.getNode()->setIROrder(idx + 1);
2275 for (unsigned j = 0; j < VecSize; ++j) {
2276 if (i + j >= NumElts)
2278 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2279 DAG.getIntPtrConstant(j, dl));
2280 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
2281 Elt = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt);
2282 InVals.push_back(Elt);
2284 Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
2294 EVT ObjectVT = getValueType(Ty);
2295 // If ABI, load from the param symbol
2296 SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2297 Value *srcValue = Constant::getNullValue(PointerType::get(
2298 ObjectVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2300 if (ObjectVT.getSizeInBits() < Ins[InsIdx].VT.getSizeInBits()) {
2301 ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ?
2302 ISD::SEXTLOAD : ISD::ZEXTLOAD;
2303 p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, Arg,
2304 MachinePointerInfo(srcValue), ObjectVT, false, false,
2306 TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
2308 p = DAG.getLoad(Ins[InsIdx].VT, dl, Root, Arg,
2309 MachinePointerInfo(srcValue), false, false, false,
2310 TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
2313 p.getNode()->setIROrder(idx + 1);
2314 InVals.push_back(p);
2318 // Param has ByVal attribute
2319 // Return MoveParam(param symbol).
2320 // Ideally, the param symbol can be returned directly,
2321 // but when SDNode builder decides to use it in a CopyToReg(),
2322 // machine instruction fails because TargetExternalSymbol
2323 // (not lowered) is target dependent, and CopyToReg assumes
2324 // the source is lowered.
2325 EVT ObjectVT = getValueType(Ty);
2326 assert(ObjectVT == Ins[InsIdx].VT &&
2327 "Ins type did not match function type");
2328 SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2329 SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg);
2331 p.getNode()->setIROrder(idx + 1);
2333 InVals.push_back(p);
2335 SDValue p2 = DAG.getNode(
2336 ISD::INTRINSIC_WO_CHAIN, dl, ObjectVT,
2337 DAG.getConstant(Intrinsic::nvvm_ptr_local_to_gen, dl, MVT::i32), p);
2338 InVals.push_back(p2);
2342 // Clang will check explicit VarArg and issue error if any. However, Clang
2343 // will let code with
2344 // implicit var arg like f() pass. See bug 617733.
2345 // We treat this case as if the arg list is empty.
2346 // if (F.isVarArg()) {
2347 // assert(0 && "VarArg not supported yet!");
2350 if (!OutChains.empty())
2351 DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains));
2358 NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2360 const SmallVectorImpl<ISD::OutputArg> &Outs,
2361 const SmallVectorImpl<SDValue> &OutVals,
2362 SDLoc dl, SelectionDAG &DAG) const {
2363 MachineFunction &MF = DAG.getMachineFunction();
2364 const Function *F = MF.getFunction();
2365 Type *RetTy = F->getReturnType();
2366 const DataLayout *TD = getDataLayout();
2368 bool isABI = (STI.getSmVersion() >= 20);
2369 assert(isABI && "Non-ABI compilation is not supported");
2373 if (VectorType *VTy = dyn_cast<VectorType>(RetTy)) {
2374 // If we have a vector type, the OutVals array will be the scalarized
2375 // components and we have combine them into 1 or more vector stores.
2376 unsigned NumElts = VTy->getNumElements();
2377 assert(NumElts == Outs.size() && "Bad scalarization of return value");
2379 // const_cast can be removed in later LLVM versions
2380 EVT EltVT = getValueType(RetTy).getVectorElementType();
2381 bool NeedExtend = false;
2382 if (EltVT.getSizeInBits() < 16)
2387 SDValue StoreVal = OutVals[0];
2388 // We only have one element, so just directly store it
2390 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
2391 SDValue Ops[] = { Chain, DAG.getConstant(0, dl, MVT::i32), StoreVal };
2392 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
2393 DAG.getVTList(MVT::Other), Ops,
2394 EltVT, MachinePointerInfo());
2396 } else if (NumElts == 2) {
2398 SDValue StoreVal0 = OutVals[0];
2399 SDValue StoreVal1 = OutVals[1];
2402 StoreVal0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal0);
2403 StoreVal1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal1);
2406 SDValue Ops[] = { Chain, DAG.getConstant(0, dl, MVT::i32), StoreVal0,
2408 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetvalV2, dl,
2409 DAG.getVTList(MVT::Other), Ops,
2410 EltVT, MachinePointerInfo());
2413 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and the
2414 // vector will be expanded to a power of 2 elements, so we know we can
2415 // always round up to the next multiple of 4 when creating the vector
2417 // e.g. 4 elem => 1 st.v4
2418 // 6 elem => 2 st.v4
2419 // 8 elem => 2 st.v4
2420 // 11 elem => 3 st.v4
2422 unsigned VecSize = 4;
2423 if (OutVals[0].getValueType().getSizeInBits() == 64)
2426 unsigned Offset = 0;
2429 EVT::getVectorVT(F->getContext(), EltVT, VecSize);
2430 unsigned PerStoreOffset =
2431 TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
2433 for (unsigned i = 0; i < NumElts; i += VecSize) {
2436 SmallVector<SDValue, 8> Ops;
2437 Ops.push_back(Chain);
2438 Ops.push_back(DAG.getConstant(Offset, dl, MVT::i32));
2439 unsigned Opc = NVPTXISD::StoreRetvalV2;
2440 EVT ExtendedVT = (NeedExtend) ? MVT::i16 : OutVals[0].getValueType();
2442 StoreVal = OutVals[i];
2444 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2445 Ops.push_back(StoreVal);
2447 if (i + 1 < NumElts) {
2448 StoreVal = OutVals[i + 1];
2450 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2452 StoreVal = DAG.getUNDEF(ExtendedVT);
2454 Ops.push_back(StoreVal);
2457 Opc = NVPTXISD::StoreRetvalV4;
2458 if (i + 2 < NumElts) {
2459 StoreVal = OutVals[i + 2];
2462 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2464 StoreVal = DAG.getUNDEF(ExtendedVT);
2466 Ops.push_back(StoreVal);
2468 if (i + 3 < NumElts) {
2469 StoreVal = OutVals[i + 3];
2472 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2474 StoreVal = DAG.getUNDEF(ExtendedVT);
2476 Ops.push_back(StoreVal);
2479 // Chain = DAG.getNode(Opc, dl, MVT::Other, &Ops[0], Ops.size());
2481 DAG.getMemIntrinsicNode(Opc, dl, DAG.getVTList(MVT::Other), Ops,
2482 EltVT, MachinePointerInfo());
2483 Offset += PerStoreOffset;
2487 SmallVector<EVT, 16> ValVTs;
2488 SmallVector<uint64_t, 16> Offsets;
2489 ComputePTXValueVTs(*this, RetTy, ValVTs, &Offsets, 0);
2490 assert(ValVTs.size() == OutVals.size() && "Bad return value decomposition");
2492 for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
2493 SDValue theVal = OutVals[i];
2494 EVT TheValType = theVal.getValueType();
2495 unsigned numElems = 1;
2496 if (TheValType.isVector())
2497 numElems = TheValType.getVectorNumElements();
2498 for (unsigned j = 0, je = numElems; j != je; ++j) {
2499 SDValue TmpVal = theVal;
2500 if (TheValType.isVector())
2501 TmpVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
2502 TheValType.getVectorElementType(), TmpVal,
2503 DAG.getIntPtrConstant(j, dl));
2504 EVT TheStoreType = ValVTs[i];
2505 if (RetTy->isIntegerTy() &&
2506 TD->getTypeAllocSizeInBits(RetTy) < 32) {
2507 // The following zero-extension is for integer types only, and
2508 // specifically not for aggregates.
2509 TmpVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, TmpVal);
2510 TheStoreType = MVT::i32;
2512 else if (TmpVal.getValueType().getSizeInBits() < 16)
2513 TmpVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, TmpVal);
2517 DAG.getConstant(Offsets[i], dl, MVT::i32),
2519 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
2520 DAG.getVTList(MVT::Other), Ops,
2522 MachinePointerInfo());
2527 return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain);
2531 void NVPTXTargetLowering::LowerAsmOperandForConstraint(
2532 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
2533 SelectionDAG &DAG) const {
2534 if (Constraint.length() > 1)
2537 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2540 // NVPTX suuport vector of legal types of any length in Intrinsics because the
2541 // NVPTX specific type legalizer
2542 // will legalize them to the PTX supported length.
2543 bool NVPTXTargetLowering::isTypeSupportedInIntrinsic(MVT VT) const {
2544 if (isTypeLegal(VT))
2546 if (VT.isVector()) {
2547 MVT eVT = VT.getVectorElementType();
2548 if (isTypeLegal(eVT))
2554 static unsigned getOpcForTextureInstr(unsigned Intrinsic) {
2555 switch (Intrinsic) {
2559 case Intrinsic::nvvm_tex_1d_v4f32_s32:
2560 return NVPTXISD::Tex1DFloatS32;
2561 case Intrinsic::nvvm_tex_1d_v4f32_f32:
2562 return NVPTXISD::Tex1DFloatFloat;
2563 case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
2564 return NVPTXISD::Tex1DFloatFloatLevel;
2565 case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
2566 return NVPTXISD::Tex1DFloatFloatGrad;
2567 case Intrinsic::nvvm_tex_1d_v4s32_s32:
2568 return NVPTXISD::Tex1DS32S32;
2569 case Intrinsic::nvvm_tex_1d_v4s32_f32:
2570 return NVPTXISD::Tex1DS32Float;
2571 case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
2572 return NVPTXISD::Tex1DS32FloatLevel;
2573 case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
2574 return NVPTXISD::Tex1DS32FloatGrad;
2575 case Intrinsic::nvvm_tex_1d_v4u32_s32:
2576 return NVPTXISD::Tex1DU32S32;
2577 case Intrinsic::nvvm_tex_1d_v4u32_f32:
2578 return NVPTXISD::Tex1DU32Float;
2579 case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
2580 return NVPTXISD::Tex1DU32FloatLevel;
2581 case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
2582 return NVPTXISD::Tex1DU32FloatGrad;
2584 case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
2585 return NVPTXISD::Tex1DArrayFloatS32;
2586 case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
2587 return NVPTXISD::Tex1DArrayFloatFloat;
2588 case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
2589 return NVPTXISD::Tex1DArrayFloatFloatLevel;
2590 case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
2591 return NVPTXISD::Tex1DArrayFloatFloatGrad;
2592 case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
2593 return NVPTXISD::Tex1DArrayS32S32;
2594 case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
2595 return NVPTXISD::Tex1DArrayS32Float;
2596 case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
2597 return NVPTXISD::Tex1DArrayS32FloatLevel;
2598 case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
2599 return NVPTXISD::Tex1DArrayS32FloatGrad;
2600 case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
2601 return NVPTXISD::Tex1DArrayU32S32;
2602 case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
2603 return NVPTXISD::Tex1DArrayU32Float;
2604 case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
2605 return NVPTXISD::Tex1DArrayU32FloatLevel;
2606 case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
2607 return NVPTXISD::Tex1DArrayU32FloatGrad;
2609 case Intrinsic::nvvm_tex_2d_v4f32_s32:
2610 return NVPTXISD::Tex2DFloatS32;
2611 case Intrinsic::nvvm_tex_2d_v4f32_f32:
2612 return NVPTXISD::Tex2DFloatFloat;
2613 case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
2614 return NVPTXISD::Tex2DFloatFloatLevel;
2615 case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
2616 return NVPTXISD::Tex2DFloatFloatGrad;
2617 case Intrinsic::nvvm_tex_2d_v4s32_s32:
2618 return NVPTXISD::Tex2DS32S32;
2619 case Intrinsic::nvvm_tex_2d_v4s32_f32:
2620 return NVPTXISD::Tex2DS32Float;
2621 case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
2622 return NVPTXISD::Tex2DS32FloatLevel;
2623 case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
2624 return NVPTXISD::Tex2DS32FloatGrad;
2625 case Intrinsic::nvvm_tex_2d_v4u32_s32:
2626 return NVPTXISD::Tex2DU32S32;
2627 case Intrinsic::nvvm_tex_2d_v4u32_f32:
2628 return NVPTXISD::Tex2DU32Float;
2629 case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
2630 return NVPTXISD::Tex2DU32FloatLevel;
2631 case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
2632 return NVPTXISD::Tex2DU32FloatGrad;
2634 case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
2635 return NVPTXISD::Tex2DArrayFloatS32;
2636 case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
2637 return NVPTXISD::Tex2DArrayFloatFloat;
2638 case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
2639 return NVPTXISD::Tex2DArrayFloatFloatLevel;
2640 case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
2641 return NVPTXISD::Tex2DArrayFloatFloatGrad;
2642 case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
2643 return NVPTXISD::Tex2DArrayS32S32;
2644 case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
2645 return NVPTXISD::Tex2DArrayS32Float;
2646 case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
2647 return NVPTXISD::Tex2DArrayS32FloatLevel;
2648 case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
2649 return NVPTXISD::Tex2DArrayS32FloatGrad;
2650 case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
2651 return NVPTXISD::Tex2DArrayU32S32;
2652 case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
2653 return NVPTXISD::Tex2DArrayU32Float;
2654 case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
2655 return NVPTXISD::Tex2DArrayU32FloatLevel;
2656 case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
2657 return NVPTXISD::Tex2DArrayU32FloatGrad;
2659 case Intrinsic::nvvm_tex_3d_v4f32_s32:
2660 return NVPTXISD::Tex3DFloatS32;
2661 case Intrinsic::nvvm_tex_3d_v4f32_f32:
2662 return NVPTXISD::Tex3DFloatFloat;
2663 case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
2664 return NVPTXISD::Tex3DFloatFloatLevel;
2665 case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
2666 return NVPTXISD::Tex3DFloatFloatGrad;
2667 case Intrinsic::nvvm_tex_3d_v4s32_s32:
2668 return NVPTXISD::Tex3DS32S32;
2669 case Intrinsic::nvvm_tex_3d_v4s32_f32:
2670 return NVPTXISD::Tex3DS32Float;
2671 case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
2672 return NVPTXISD::Tex3DS32FloatLevel;
2673 case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
2674 return NVPTXISD::Tex3DS32FloatGrad;
2675 case Intrinsic::nvvm_tex_3d_v4u32_s32:
2676 return NVPTXISD::Tex3DU32S32;
2677 case Intrinsic::nvvm_tex_3d_v4u32_f32:
2678 return NVPTXISD::Tex3DU32Float;
2679 case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
2680 return NVPTXISD::Tex3DU32FloatLevel;
2681 case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
2682 return NVPTXISD::Tex3DU32FloatGrad;
2684 case Intrinsic::nvvm_tex_cube_v4f32_f32:
2685 return NVPTXISD::TexCubeFloatFloat;
2686 case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
2687 return NVPTXISD::TexCubeFloatFloatLevel;
2688 case Intrinsic::nvvm_tex_cube_v4s32_f32:
2689 return NVPTXISD::TexCubeS32Float;
2690 case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
2691 return NVPTXISD::TexCubeS32FloatLevel;
2692 case Intrinsic::nvvm_tex_cube_v4u32_f32:
2693 return NVPTXISD::TexCubeU32Float;
2694 case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
2695 return NVPTXISD::TexCubeU32FloatLevel;
2697 case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
2698 return NVPTXISD::TexCubeArrayFloatFloat;
2699 case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
2700 return NVPTXISD::TexCubeArrayFloatFloatLevel;
2701 case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
2702 return NVPTXISD::TexCubeArrayS32Float;
2703 case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
2704 return NVPTXISD::TexCubeArrayS32FloatLevel;
2705 case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
2706 return NVPTXISD::TexCubeArrayU32Float;
2707 case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
2708 return NVPTXISD::TexCubeArrayU32FloatLevel;
2710 case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
2711 return NVPTXISD::Tld4R2DFloatFloat;
2712 case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
2713 return NVPTXISD::Tld4G2DFloatFloat;
2714 case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
2715 return NVPTXISD::Tld4B2DFloatFloat;
2716 case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
2717 return NVPTXISD::Tld4A2DFloatFloat;
2718 case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
2719 return NVPTXISD::Tld4R2DS64Float;
2720 case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
2721 return NVPTXISD::Tld4G2DS64Float;
2722 case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
2723 return NVPTXISD::Tld4B2DS64Float;
2724 case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
2725 return NVPTXISD::Tld4A2DS64Float;
2726 case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
2727 return NVPTXISD::Tld4R2DU64Float;
2728 case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
2729 return NVPTXISD::Tld4G2DU64Float;
2730 case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
2731 return NVPTXISD::Tld4B2DU64Float;
2732 case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
2733 return NVPTXISD::Tld4A2DU64Float;
2735 case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
2736 return NVPTXISD::TexUnified1DFloatS32;
2737 case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
2738 return NVPTXISD::TexUnified1DFloatFloat;
2739 case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
2740 return NVPTXISD::TexUnified1DFloatFloatLevel;
2741 case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
2742 return NVPTXISD::TexUnified1DFloatFloatGrad;
2743 case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
2744 return NVPTXISD::TexUnified1DS32S32;
2745 case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
2746 return NVPTXISD::TexUnified1DS32Float;
2747 case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
2748 return NVPTXISD::TexUnified1DS32FloatLevel;
2749 case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
2750 return NVPTXISD::TexUnified1DS32FloatGrad;
2751 case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
2752 return NVPTXISD::TexUnified1DU32S32;
2753 case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
2754 return NVPTXISD::TexUnified1DU32Float;
2755 case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
2756 return NVPTXISD::TexUnified1DU32FloatLevel;
2757 case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
2758 return NVPTXISD::TexUnified1DU32FloatGrad;
2760 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
2761 return NVPTXISD::TexUnified1DArrayFloatS32;
2762 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
2763 return NVPTXISD::TexUnified1DArrayFloatFloat;
2764 case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
2765 return NVPTXISD::TexUnified1DArrayFloatFloatLevel;
2766 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
2767 return NVPTXISD::TexUnified1DArrayFloatFloatGrad;
2768 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
2769 return NVPTXISD::TexUnified1DArrayS32S32;
2770 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
2771 return NVPTXISD::TexUnified1DArrayS32Float;
2772 case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
2773 return NVPTXISD::TexUnified1DArrayS32FloatLevel;
2774 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
2775 return NVPTXISD::TexUnified1DArrayS32FloatGrad;
2776 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
2777 return NVPTXISD::TexUnified1DArrayU32S32;
2778 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
2779 return NVPTXISD::TexUnified1DArrayU32Float;
2780 case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
2781 return NVPTXISD::TexUnified1DArrayU32FloatLevel;
2782 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
2783 return NVPTXISD::TexUnified1DArrayU32FloatGrad;
2785 case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
2786 return NVPTXISD::TexUnified2DFloatS32;
2787 case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
2788 return NVPTXISD::TexUnified2DFloatFloat;
2789 case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
2790 return NVPTXISD::TexUnified2DFloatFloatLevel;
2791 case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
2792 return NVPTXISD::TexUnified2DFloatFloatGrad;
2793 case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
2794 return NVPTXISD::TexUnified2DS32S32;
2795 case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
2796 return NVPTXISD::TexUnified2DS32Float;
2797 case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
2798 return NVPTXISD::TexUnified2DS32FloatLevel;
2799 case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
2800 return NVPTXISD::TexUnified2DS32FloatGrad;
2801 case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
2802 return NVPTXISD::TexUnified2DU32S32;
2803 case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
2804 return NVPTXISD::TexUnified2DU32Float;
2805 case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
2806 return NVPTXISD::TexUnified2DU32FloatLevel;
2807 case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
2808 return NVPTXISD::TexUnified2DU32FloatGrad;
2810 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
2811 return NVPTXISD::TexUnified2DArrayFloatS32;
2812 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
2813 return NVPTXISD::TexUnified2DArrayFloatFloat;
2814 case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
2815 return NVPTXISD::TexUnified2DArrayFloatFloatLevel;
2816 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
2817 return NVPTXISD::TexUnified2DArrayFloatFloatGrad;
2818 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
2819 return NVPTXISD::TexUnified2DArrayS32S32;
2820 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
2821 return NVPTXISD::TexUnified2DArrayS32Float;
2822 case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
2823 return NVPTXISD::TexUnified2DArrayS32FloatLevel;
2824 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
2825 return NVPTXISD::TexUnified2DArrayS32FloatGrad;
2826 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
2827 return NVPTXISD::TexUnified2DArrayU32S32;
2828 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
2829 return NVPTXISD::TexUnified2DArrayU32Float;
2830 case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
2831 return NVPTXISD::TexUnified2DArrayU32FloatLevel;
2832 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
2833 return NVPTXISD::TexUnified2DArrayU32FloatGrad;
2835 case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
2836 return NVPTXISD::TexUnified3DFloatS32;
2837 case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
2838 return NVPTXISD::TexUnified3DFloatFloat;
2839 case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
2840 return NVPTXISD::TexUnified3DFloatFloatLevel;
2841 case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
2842 return NVPTXISD::TexUnified3DFloatFloatGrad;
2843 case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
2844 return NVPTXISD::TexUnified3DS32S32;
2845 case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
2846 return NVPTXISD::TexUnified3DS32Float;
2847 case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
2848 return NVPTXISD::TexUnified3DS32FloatLevel;
2849 case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
2850 return NVPTXISD::TexUnified3DS32FloatGrad;
2851 case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
2852 return NVPTXISD::TexUnified3DU32S32;
2853 case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
2854 return NVPTXISD::TexUnified3DU32Float;
2855 case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
2856 return NVPTXISD::TexUnified3DU32FloatLevel;
2857 case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
2858 return NVPTXISD::TexUnified3DU32FloatGrad;
2860 case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
2861 return NVPTXISD::TexUnifiedCubeFloatFloat;
2862 case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
2863 return NVPTXISD::TexUnifiedCubeFloatFloatLevel;
2864 case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
2865 return NVPTXISD::TexUnifiedCubeS32Float;
2866 case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
2867 return NVPTXISD::TexUnifiedCubeS32FloatLevel;
2868 case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
2869 return NVPTXISD::TexUnifiedCubeU32Float;
2870 case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
2871 return NVPTXISD::TexUnifiedCubeU32FloatLevel;
2873 case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
2874 return NVPTXISD::TexUnifiedCubeArrayFloatFloat;
2875 case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
2876 return NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel;
2877 case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
2878 return NVPTXISD::TexUnifiedCubeArrayS32Float;
2879 case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
2880 return NVPTXISD::TexUnifiedCubeArrayS32FloatLevel;
2881 case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
2882 return NVPTXISD::TexUnifiedCubeArrayU32Float;
2883 case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
2884 return NVPTXISD::TexUnifiedCubeArrayU32FloatLevel;
2886 case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
2887 return NVPTXISD::Tld4UnifiedR2DFloatFloat;
2888 case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
2889 return NVPTXISD::Tld4UnifiedG2DFloatFloat;
2890 case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
2891 return NVPTXISD::Tld4UnifiedB2DFloatFloat;
2892 case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32:
2893 return NVPTXISD::Tld4UnifiedA2DFloatFloat;
2894 case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
2895 return NVPTXISD::Tld4UnifiedR2DS64Float;
2896 case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
2897 return NVPTXISD::Tld4UnifiedG2DS64Float;
2898 case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
2899 return NVPTXISD::Tld4UnifiedB2DS64Float;
2900 case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
2901 return NVPTXISD::Tld4UnifiedA2DS64Float;
2902 case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
2903 return NVPTXISD::Tld4UnifiedR2DU64Float;
2904 case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
2905 return NVPTXISD::Tld4UnifiedG2DU64Float;
2906 case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
2907 return NVPTXISD::Tld4UnifiedB2DU64Float;
2908 case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32:
2909 return NVPTXISD::Tld4UnifiedA2DU64Float;
2913 static unsigned getOpcForSurfaceInstr(unsigned Intrinsic) {
2914 switch (Intrinsic) {
2917 case Intrinsic::nvvm_suld_1d_i8_clamp:
2918 return NVPTXISD::Suld1DI8Clamp;
2919 case Intrinsic::nvvm_suld_1d_i16_clamp:
2920 return NVPTXISD::Suld1DI16Clamp;
2921 case Intrinsic::nvvm_suld_1d_i32_clamp:
2922 return NVPTXISD::Suld1DI32Clamp;
2923 case Intrinsic::nvvm_suld_1d_i64_clamp:
2924 return NVPTXISD::Suld1DI64Clamp;
2925 case Intrinsic::nvvm_suld_1d_v2i8_clamp:
2926 return NVPTXISD::Suld1DV2I8Clamp;
2927 case Intrinsic::nvvm_suld_1d_v2i16_clamp:
2928 return NVPTXISD::Suld1DV2I16Clamp;
2929 case Intrinsic::nvvm_suld_1d_v2i32_clamp:
2930 return NVPTXISD::Suld1DV2I32Clamp;
2931 case Intrinsic::nvvm_suld_1d_v2i64_clamp:
2932 return NVPTXISD::Suld1DV2I64Clamp;
2933 case Intrinsic::nvvm_suld_1d_v4i8_clamp:
2934 return NVPTXISD::Suld1DV4I8Clamp;
2935 case Intrinsic::nvvm_suld_1d_v4i16_clamp:
2936 return NVPTXISD::Suld1DV4I16Clamp;
2937 case Intrinsic::nvvm_suld_1d_v4i32_clamp:
2938 return NVPTXISD::Suld1DV4I32Clamp;
2939 case Intrinsic::nvvm_suld_1d_array_i8_clamp:
2940 return NVPTXISD::Suld1DArrayI8Clamp;
2941 case Intrinsic::nvvm_suld_1d_array_i16_clamp:
2942 return NVPTXISD::Suld1DArrayI16Clamp;
2943 case Intrinsic::nvvm_suld_1d_array_i32_clamp:
2944 return NVPTXISD::Suld1DArrayI32Clamp;
2945 case Intrinsic::nvvm_suld_1d_array_i64_clamp:
2946 return NVPTXISD::Suld1DArrayI64Clamp;
2947 case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
2948 return NVPTXISD::Suld1DArrayV2I8Clamp;
2949 case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
2950 return NVPTXISD::Suld1DArrayV2I16Clamp;
2951 case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
2952 return NVPTXISD::Suld1DArrayV2I32Clamp;
2953 case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
2954 return NVPTXISD::Suld1DArrayV2I64Clamp;
2955 case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
2956 return NVPTXISD::Suld1DArrayV4I8Clamp;
2957 case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
2958 return NVPTXISD::Suld1DArrayV4I16Clamp;
2959 case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
2960 return NVPTXISD::Suld1DArrayV4I32Clamp;
2961 case Intrinsic::nvvm_suld_2d_i8_clamp:
2962 return NVPTXISD::Suld2DI8Clamp;
2963 case Intrinsic::nvvm_suld_2d_i16_clamp:
2964 return NVPTXISD::Suld2DI16Clamp;
2965 case Intrinsic::nvvm_suld_2d_i32_clamp:
2966 return NVPTXISD::Suld2DI32Clamp;
2967 case Intrinsic::nvvm_suld_2d_i64_clamp:
2968 return NVPTXISD::Suld2DI64Clamp;
2969 case Intrinsic::nvvm_suld_2d_v2i8_clamp:
2970 return NVPTXISD::Suld2DV2I8Clamp;
2971 case Intrinsic::nvvm_suld_2d_v2i16_clamp:
2972 return NVPTXISD::Suld2DV2I16Clamp;
2973 case Intrinsic::nvvm_suld_2d_v2i32_clamp:
2974 return NVPTXISD::Suld2DV2I32Clamp;
2975 case Intrinsic::nvvm_suld_2d_v2i64_clamp:
2976 return NVPTXISD::Suld2DV2I64Clamp;
2977 case Intrinsic::nvvm_suld_2d_v4i8_clamp:
2978 return NVPTXISD::Suld2DV4I8Clamp;
2979 case Intrinsic::nvvm_suld_2d_v4i16_clamp:
2980 return NVPTXISD::Suld2DV4I16Clamp;
2981 case Intrinsic::nvvm_suld_2d_v4i32_clamp:
2982 return NVPTXISD::Suld2DV4I32Clamp;
2983 case Intrinsic::nvvm_suld_2d_array_i8_clamp:
2984 return NVPTXISD::Suld2DArrayI8Clamp;
2985 case Intrinsic::nvvm_suld_2d_array_i16_clamp:
2986 return NVPTXISD::Suld2DArrayI16Clamp;
2987 case Intrinsic::nvvm_suld_2d_array_i32_clamp:
2988 return NVPTXISD::Suld2DArrayI32Clamp;
2989 case Intrinsic::nvvm_suld_2d_array_i64_clamp:
2990 return NVPTXISD::Suld2DArrayI64Clamp;
2991 case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
2992 return NVPTXISD::Suld2DArrayV2I8Clamp;
2993 case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
2994 return NVPTXISD::Suld2DArrayV2I16Clamp;
2995 case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
2996 return NVPTXISD::Suld2DArrayV2I32Clamp;
2997 case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
2998 return NVPTXISD::Suld2DArrayV2I64Clamp;
2999 case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
3000 return NVPTXISD::Suld2DArrayV4I8Clamp;
3001 case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
3002 return NVPTXISD::Suld2DArrayV4I16Clamp;
3003 case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
3004 return NVPTXISD::Suld2DArrayV4I32Clamp;
3005 case Intrinsic::nvvm_suld_3d_i8_clamp:
3006 return NVPTXISD::Suld3DI8Clamp;
3007 case Intrinsic::nvvm_suld_3d_i16_clamp:
3008 return NVPTXISD::Suld3DI16Clamp;
3009 case Intrinsic::nvvm_suld_3d_i32_clamp:
3010 return NVPTXISD::Suld3DI32Clamp;
3011 case Intrinsic::nvvm_suld_3d_i64_clamp:
3012 return NVPTXISD::Suld3DI64Clamp;
3013 case Intrinsic::nvvm_suld_3d_v2i8_clamp:
3014 return NVPTXISD::Suld3DV2I8Clamp;
3015 case Intrinsic::nvvm_suld_3d_v2i16_clamp:
3016 return NVPTXISD::Suld3DV2I16Clamp;
3017 case Intrinsic::nvvm_suld_3d_v2i32_clamp:
3018 return NVPTXISD::Suld3DV2I32Clamp;
3019 case Intrinsic::nvvm_suld_3d_v2i64_clamp:
3020 return NVPTXISD::Suld3DV2I64Clamp;
3021 case Intrinsic::nvvm_suld_3d_v4i8_clamp:
3022 return NVPTXISD::Suld3DV4I8Clamp;
3023 case Intrinsic::nvvm_suld_3d_v4i16_clamp:
3024 return NVPTXISD::Suld3DV4I16Clamp;
3025 case Intrinsic::nvvm_suld_3d_v4i32_clamp:
3026 return NVPTXISD::Suld3DV4I32Clamp;
3027 case Intrinsic::nvvm_suld_1d_i8_trap:
3028 return NVPTXISD::Suld1DI8Trap;
3029 case Intrinsic::nvvm_suld_1d_i16_trap:
3030 return NVPTXISD::Suld1DI16Trap;
3031 case Intrinsic::nvvm_suld_1d_i32_trap:
3032 return NVPTXISD::Suld1DI32Trap;
3033 case Intrinsic::nvvm_suld_1d_i64_trap:
3034 return NVPTXISD::Suld1DI64Trap;
3035 case Intrinsic::nvvm_suld_1d_v2i8_trap:
3036 return NVPTXISD::Suld1DV2I8Trap;
3037 case Intrinsic::nvvm_suld_1d_v2i16_trap:
3038 return NVPTXISD::Suld1DV2I16Trap;
3039 case Intrinsic::nvvm_suld_1d_v2i32_trap:
3040 return NVPTXISD::Suld1DV2I32Trap;
3041 case Intrinsic::nvvm_suld_1d_v2i64_trap:
3042 return NVPTXISD::Suld1DV2I64Trap;
3043 case Intrinsic::nvvm_suld_1d_v4i8_trap:
3044 return NVPTXISD::Suld1DV4I8Trap;
3045 case Intrinsic::nvvm_suld_1d_v4i16_trap:
3046 return NVPTXISD::Suld1DV4I16Trap;
3047 case Intrinsic::nvvm_suld_1d_v4i32_trap:
3048 return NVPTXISD::Suld1DV4I32Trap;
3049 case Intrinsic::nvvm_suld_1d_array_i8_trap:
3050 return NVPTXISD::Suld1DArrayI8Trap;
3051 case Intrinsic::nvvm_suld_1d_array_i16_trap:
3052 return NVPTXISD::Suld1DArrayI16Trap;
3053 case Intrinsic::nvvm_suld_1d_array_i32_trap:
3054 return NVPTXISD::Suld1DArrayI32Trap;
3055 case Intrinsic::nvvm_suld_1d_array_i64_trap:
3056 return NVPTXISD::Suld1DArrayI64Trap;
3057 case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3058 return NVPTXISD::Suld1DArrayV2I8Trap;
3059 case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3060 return NVPTXISD::Suld1DArrayV2I16Trap;
3061 case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3062 return NVPTXISD::Suld1DArrayV2I32Trap;
3063 case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3064 return NVPTXISD::Suld1DArrayV2I64Trap;
3065 case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3066 return NVPTXISD::Suld1DArrayV4I8Trap;
3067 case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3068 return NVPTXISD::Suld1DArrayV4I16Trap;
3069 case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3070 return NVPTXISD::Suld1DArrayV4I32Trap;
3071 case Intrinsic::nvvm_suld_2d_i8_trap:
3072 return NVPTXISD::Suld2DI8Trap;
3073 case Intrinsic::nvvm_suld_2d_i16_trap:
3074 return NVPTXISD::Suld2DI16Trap;
3075 case Intrinsic::nvvm_suld_2d_i32_trap:
3076 return NVPTXISD::Suld2DI32Trap;
3077 case Intrinsic::nvvm_suld_2d_i64_trap:
3078 return NVPTXISD::Suld2DI64Trap;
3079 case Intrinsic::nvvm_suld_2d_v2i8_trap:
3080 return NVPTXISD::Suld2DV2I8Trap;
3081 case Intrinsic::nvvm_suld_2d_v2i16_trap:
3082 return NVPTXISD::Suld2DV2I16Trap;
3083 case Intrinsic::nvvm_suld_2d_v2i32_trap:
3084 return NVPTXISD::Suld2DV2I32Trap;
3085 case Intrinsic::nvvm_suld_2d_v2i64_trap:
3086 return NVPTXISD::Suld2DV2I64Trap;
3087 case Intrinsic::nvvm_suld_2d_v4i8_trap:
3088 return NVPTXISD::Suld2DV4I8Trap;
3089 case Intrinsic::nvvm_suld_2d_v4i16_trap:
3090 return NVPTXISD::Suld2DV4I16Trap;
3091 case Intrinsic::nvvm_suld_2d_v4i32_trap:
3092 return NVPTXISD::Suld2DV4I32Trap;
3093 case Intrinsic::nvvm_suld_2d_array_i8_trap:
3094 return NVPTXISD::Suld2DArrayI8Trap;
3095 case Intrinsic::nvvm_suld_2d_array_i16_trap:
3096 return NVPTXISD::Suld2DArrayI16Trap;
3097 case Intrinsic::nvvm_suld_2d_array_i32_trap:
3098 return NVPTXISD::Suld2DArrayI32Trap;
3099 case Intrinsic::nvvm_suld_2d_array_i64_trap:
3100 return NVPTXISD::Suld2DArrayI64Trap;
3101 case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3102 return NVPTXISD::Suld2DArrayV2I8Trap;
3103 case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3104 return NVPTXISD::Suld2DArrayV2I16Trap;
3105 case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3106 return NVPTXISD::Suld2DArrayV2I32Trap;
3107 case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3108 return NVPTXISD::Suld2DArrayV2I64Trap;
3109 case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3110 return NVPTXISD::Suld2DArrayV4I8Trap;
3111 case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3112 return NVPTXISD::Suld2DArrayV4I16Trap;
3113 case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3114 return NVPTXISD::Suld2DArrayV4I32Trap;
3115 case Intrinsic::nvvm_suld_3d_i8_trap:
3116 return NVPTXISD::Suld3DI8Trap;
3117 case Intrinsic::nvvm_suld_3d_i16_trap:
3118 return NVPTXISD::Suld3DI16Trap;
3119 case Intrinsic::nvvm_suld_3d_i32_trap:
3120 return NVPTXISD::Suld3DI32Trap;
3121 case Intrinsic::nvvm_suld_3d_i64_trap:
3122 return NVPTXISD::Suld3DI64Trap;
3123 case Intrinsic::nvvm_suld_3d_v2i8_trap:
3124 return NVPTXISD::Suld3DV2I8Trap;
3125 case Intrinsic::nvvm_suld_3d_v2i16_trap:
3126 return NVPTXISD::Suld3DV2I16Trap;
3127 case Intrinsic::nvvm_suld_3d_v2i32_trap:
3128 return NVPTXISD::Suld3DV2I32Trap;
3129 case Intrinsic::nvvm_suld_3d_v2i64_trap:
3130 return NVPTXISD::Suld3DV2I64Trap;
3131 case Intrinsic::nvvm_suld_3d_v4i8_trap:
3132 return NVPTXISD::Suld3DV4I8Trap;
3133 case Intrinsic::nvvm_suld_3d_v4i16_trap:
3134 return NVPTXISD::Suld3DV4I16Trap;
3135 case Intrinsic::nvvm_suld_3d_v4i32_trap:
3136 return NVPTXISD::Suld3DV4I32Trap;
3137 case Intrinsic::nvvm_suld_1d_i8_zero:
3138 return NVPTXISD::Suld1DI8Zero;
3139 case Intrinsic::nvvm_suld_1d_i16_zero:
3140 return NVPTXISD::Suld1DI16Zero;
3141 case Intrinsic::nvvm_suld_1d_i32_zero:
3142 return NVPTXISD::Suld1DI32Zero;
3143 case Intrinsic::nvvm_suld_1d_i64_zero:
3144 return NVPTXISD::Suld1DI64Zero;
3145 case Intrinsic::nvvm_suld_1d_v2i8_zero:
3146 return NVPTXISD::Suld1DV2I8Zero;
3147 case Intrinsic::nvvm_suld_1d_v2i16_zero:
3148 return NVPTXISD::Suld1DV2I16Zero;
3149 case Intrinsic::nvvm_suld_1d_v2i32_zero:
3150 return NVPTXISD::Suld1DV2I32Zero;
3151 case Intrinsic::nvvm_suld_1d_v2i64_zero:
3152 return NVPTXISD::Suld1DV2I64Zero;
3153 case Intrinsic::nvvm_suld_1d_v4i8_zero:
3154 return NVPTXISD::Suld1DV4I8Zero;
3155 case Intrinsic::nvvm_suld_1d_v4i16_zero:
3156 return NVPTXISD::Suld1DV4I16Zero;
3157 case Intrinsic::nvvm_suld_1d_v4i32_zero:
3158 return NVPTXISD::Suld1DV4I32Zero;
3159 case Intrinsic::nvvm_suld_1d_array_i8_zero:
3160 return NVPTXISD::Suld1DArrayI8Zero;
3161 case Intrinsic::nvvm_suld_1d_array_i16_zero:
3162 return NVPTXISD::Suld1DArrayI16Zero;
3163 case Intrinsic::nvvm_suld_1d_array_i32_zero:
3164 return NVPTXISD::Suld1DArrayI32Zero;
3165 case Intrinsic::nvvm_suld_1d_array_i64_zero:
3166 return NVPTXISD::Suld1DArrayI64Zero;
3167 case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3168 return NVPTXISD::Suld1DArrayV2I8Zero;
3169 case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3170 return NVPTXISD::Suld1DArrayV2I16Zero;
3171 case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3172 return NVPTXISD::Suld1DArrayV2I32Zero;
3173 case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3174 return NVPTXISD::Suld1DArrayV2I64Zero;
3175 case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3176 return NVPTXISD::Suld1DArrayV4I8Zero;
3177 case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3178 return NVPTXISD::Suld1DArrayV4I16Zero;
3179 case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3180 return NVPTXISD::Suld1DArrayV4I32Zero;
3181 case Intrinsic::nvvm_suld_2d_i8_zero:
3182 return NVPTXISD::Suld2DI8Zero;
3183 case Intrinsic::nvvm_suld_2d_i16_zero:
3184 return NVPTXISD::Suld2DI16Zero;
3185 case Intrinsic::nvvm_suld_2d_i32_zero:
3186 return NVPTXISD::Suld2DI32Zero;
3187 case Intrinsic::nvvm_suld_2d_i64_zero:
3188 return NVPTXISD::Suld2DI64Zero;
3189 case Intrinsic::nvvm_suld_2d_v2i8_zero:
3190 return NVPTXISD::Suld2DV2I8Zero;
3191 case Intrinsic::nvvm_suld_2d_v2i16_zero:
3192 return NVPTXISD::Suld2DV2I16Zero;
3193 case Intrinsic::nvvm_suld_2d_v2i32_zero:
3194 return NVPTXISD::Suld2DV2I32Zero;
3195 case Intrinsic::nvvm_suld_2d_v2i64_zero:
3196 return NVPTXISD::Suld2DV2I64Zero;
3197 case Intrinsic::nvvm_suld_2d_v4i8_zero:
3198 return NVPTXISD::Suld2DV4I8Zero;
3199 case Intrinsic::nvvm_suld_2d_v4i16_zero:
3200 return NVPTXISD::Suld2DV4I16Zero;
3201 case Intrinsic::nvvm_suld_2d_v4i32_zero:
3202 return NVPTXISD::Suld2DV4I32Zero;
3203 case Intrinsic::nvvm_suld_2d_array_i8_zero:
3204 return NVPTXISD::Suld2DArrayI8Zero;
3205 case Intrinsic::nvvm_suld_2d_array_i16_zero:
3206 return NVPTXISD::Suld2DArrayI16Zero;
3207 case Intrinsic::nvvm_suld_2d_array_i32_zero:
3208 return NVPTXISD::Suld2DArrayI32Zero;
3209 case Intrinsic::nvvm_suld_2d_array_i64_zero:
3210 return NVPTXISD::Suld2DArrayI64Zero;
3211 case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3212 return NVPTXISD::Suld2DArrayV2I8Zero;
3213 case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3214 return NVPTXISD::Suld2DArrayV2I16Zero;
3215 case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3216 return NVPTXISD::Suld2DArrayV2I32Zero;
3217 case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3218 return NVPTXISD::Suld2DArrayV2I64Zero;
3219 case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3220 return NVPTXISD::Suld2DArrayV4I8Zero;
3221 case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3222 return NVPTXISD::Suld2DArrayV4I16Zero;
3223 case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3224 return NVPTXISD::Suld2DArrayV4I32Zero;
3225 case Intrinsic::nvvm_suld_3d_i8_zero:
3226 return NVPTXISD::Suld3DI8Zero;
3227 case Intrinsic::nvvm_suld_3d_i16_zero:
3228 return NVPTXISD::Suld3DI16Zero;
3229 case Intrinsic::nvvm_suld_3d_i32_zero:
3230 return NVPTXISD::Suld3DI32Zero;
3231 case Intrinsic::nvvm_suld_3d_i64_zero:
3232 return NVPTXISD::Suld3DI64Zero;
3233 case Intrinsic::nvvm_suld_3d_v2i8_zero:
3234 return NVPTXISD::Suld3DV2I8Zero;
3235 case Intrinsic::nvvm_suld_3d_v2i16_zero:
3236 return NVPTXISD::Suld3DV2I16Zero;
3237 case Intrinsic::nvvm_suld_3d_v2i32_zero:
3238 return NVPTXISD::Suld3DV2I32Zero;
3239 case Intrinsic::nvvm_suld_3d_v2i64_zero:
3240 return NVPTXISD::Suld3DV2I64Zero;
3241 case Intrinsic::nvvm_suld_3d_v4i8_zero:
3242 return NVPTXISD::Suld3DV4I8Zero;
3243 case Intrinsic::nvvm_suld_3d_v4i16_zero:
3244 return NVPTXISD::Suld3DV4I16Zero;
3245 case Intrinsic::nvvm_suld_3d_v4i32_zero:
3246 return NVPTXISD::Suld3DV4I32Zero;
3250 // llvm.ptx.memcpy.const and llvm.ptx.memmove.const need to be modeled as
3252 // because we need the information that is only available in the "Value" type
3254 // pointer. In particular, the address space information.
3255 bool NVPTXTargetLowering::getTgtMemIntrinsic(
3256 IntrinsicInfo &Info, const CallInst &I, unsigned Intrinsic) const {
3257 switch (Intrinsic) {
3261 case Intrinsic::nvvm_atomic_load_add_f32:
3262 Info.opc = ISD::INTRINSIC_W_CHAIN;
3263 Info.memVT = MVT::f32;
3264 Info.ptrVal = I.getArgOperand(0);
3267 Info.readMem = true;
3268 Info.writeMem = true;
3272 case Intrinsic::nvvm_atomic_load_inc_32:
3273 case Intrinsic::nvvm_atomic_load_dec_32:
3274 Info.opc = ISD::INTRINSIC_W_CHAIN;
3275 Info.memVT = MVT::i32;
3276 Info.ptrVal = I.getArgOperand(0);
3279 Info.readMem = true;
3280 Info.writeMem = true;
3284 case Intrinsic::nvvm_ldu_global_i:
3285 case Intrinsic::nvvm_ldu_global_f:
3286 case Intrinsic::nvvm_ldu_global_p: {
3288 Info.opc = ISD::INTRINSIC_W_CHAIN;
3289 if (Intrinsic == Intrinsic::nvvm_ldu_global_i)
3290 Info.memVT = getValueType(I.getType());
3291 else if(Intrinsic == Intrinsic::nvvm_ldu_global_p)
3292 Info.memVT = getPointerTy();
3294 Info.memVT = getValueType(I.getType());
3295 Info.ptrVal = I.getArgOperand(0);
3298 Info.readMem = true;
3299 Info.writeMem = false;
3300 Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
3304 case Intrinsic::nvvm_ldg_global_i:
3305 case Intrinsic::nvvm_ldg_global_f:
3306 case Intrinsic::nvvm_ldg_global_p: {
3308 Info.opc = ISD::INTRINSIC_W_CHAIN;
3309 if (Intrinsic == Intrinsic::nvvm_ldg_global_i)
3310 Info.memVT = getValueType(I.getType());
3311 else if(Intrinsic == Intrinsic::nvvm_ldg_global_p)
3312 Info.memVT = getPointerTy();
3314 Info.memVT = getValueType(I.getType());
3315 Info.ptrVal = I.getArgOperand(0);
3318 Info.readMem = true;
3319 Info.writeMem = false;
3320 Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
3325 case Intrinsic::nvvm_tex_1d_v4f32_s32:
3326 case Intrinsic::nvvm_tex_1d_v4f32_f32:
3327 case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
3328 case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
3329 case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
3330 case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
3331 case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
3332 case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
3333 case Intrinsic::nvvm_tex_2d_v4f32_s32:
3334 case Intrinsic::nvvm_tex_2d_v4f32_f32:
3335 case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
3336 case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
3337 case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
3338 case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
3339 case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
3340 case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
3341 case Intrinsic::nvvm_tex_3d_v4f32_s32:
3342 case Intrinsic::nvvm_tex_3d_v4f32_f32:
3343 case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
3344 case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
3345 case Intrinsic::nvvm_tex_cube_v4f32_f32:
3346 case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
3347 case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
3348 case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
3349 case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
3350 case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
3351 case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
3352 case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
3353 case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
3354 case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
3355 case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
3356 case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
3357 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
3358 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
3359 case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
3360 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
3361 case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
3362 case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
3363 case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
3364 case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
3365 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
3366 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
3367 case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
3368 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
3369 case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
3370 case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
3371 case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
3372 case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
3373 case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
3374 case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
3375 case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
3376 case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
3377 case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
3378 case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
3379 case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
3380 case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32: {
3381 Info.opc = getOpcForTextureInstr(Intrinsic);
3382 Info.memVT = MVT::v4f32;
3383 Info.ptrVal = nullptr;
3386 Info.readMem = true;
3387 Info.writeMem = false;
3391 case Intrinsic::nvvm_tex_1d_v4s32_s32:
3392 case Intrinsic::nvvm_tex_1d_v4s32_f32:
3393 case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
3394 case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
3395 case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
3396 case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
3397 case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
3398 case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
3399 case Intrinsic::nvvm_tex_2d_v4s32_s32:
3400 case Intrinsic::nvvm_tex_2d_v4s32_f32:
3401 case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
3402 case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
3403 case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
3404 case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
3405 case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
3406 case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
3407 case Intrinsic::nvvm_tex_3d_v4s32_s32:
3408 case Intrinsic::nvvm_tex_3d_v4s32_f32:
3409 case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
3410 case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
3411 case Intrinsic::nvvm_tex_cube_v4s32_f32:
3412 case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
3413 case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
3414 case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
3415 case Intrinsic::nvvm_tex_cube_v4u32_f32:
3416 case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
3417 case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
3418 case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
3419 case Intrinsic::nvvm_tex_1d_v4u32_s32:
3420 case Intrinsic::nvvm_tex_1d_v4u32_f32:
3421 case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
3422 case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
3423 case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
3424 case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
3425 case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
3426 case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
3427 case Intrinsic::nvvm_tex_2d_v4u32_s32:
3428 case Intrinsic::nvvm_tex_2d_v4u32_f32:
3429 case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
3430 case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
3431 case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
3432 case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
3433 case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
3434 case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
3435 case Intrinsic::nvvm_tex_3d_v4u32_s32:
3436 case Intrinsic::nvvm_tex_3d_v4u32_f32:
3437 case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
3438 case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
3439 case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
3440 case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
3441 case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
3442 case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
3443 case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
3444 case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
3445 case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
3446 case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
3447 case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
3448 case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
3449 case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
3450 case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
3451 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
3452 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
3453 case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
3454 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
3455 case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
3456 case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
3457 case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
3458 case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
3459 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
3460 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
3461 case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
3462 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
3463 case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
3464 case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
3465 case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
3466 case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
3467 case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
3468 case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
3469 case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
3470 case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
3471 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
3472 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
3473 case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
3474 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
3475 case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
3476 case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
3477 case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
3478 case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
3479 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
3480 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
3481 case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
3482 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
3483 case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
3484 case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
3485 case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
3486 case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
3487 case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
3488 case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
3489 case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
3490 case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
3491 case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
3492 case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
3493 case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
3494 case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
3495 case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
3496 case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
3497 case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
3498 case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
3499 case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
3500 case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
3501 case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
3502 case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32: {
3503 Info.opc = getOpcForTextureInstr(Intrinsic);
3504 Info.memVT = MVT::v4i32;
3505 Info.ptrVal = nullptr;
3508 Info.readMem = true;
3509 Info.writeMem = false;
3513 case Intrinsic::nvvm_suld_1d_i8_clamp:
3514 case Intrinsic::nvvm_suld_1d_v2i8_clamp:
3515 case Intrinsic::nvvm_suld_1d_v4i8_clamp:
3516 case Intrinsic::nvvm_suld_1d_array_i8_clamp:
3517 case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
3518 case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
3519 case Intrinsic::nvvm_suld_2d_i8_clamp:
3520 case Intrinsic::nvvm_suld_2d_v2i8_clamp:
3521 case Intrinsic::nvvm_suld_2d_v4i8_clamp:
3522 case Intrinsic::nvvm_suld_2d_array_i8_clamp:
3523 case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
3524 case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
3525 case Intrinsic::nvvm_suld_3d_i8_clamp:
3526 case Intrinsic::nvvm_suld_3d_v2i8_clamp:
3527 case Intrinsic::nvvm_suld_3d_v4i8_clamp:
3528 case Intrinsic::nvvm_suld_1d_i8_trap:
3529 case Intrinsic::nvvm_suld_1d_v2i8_trap:
3530 case Intrinsic::nvvm_suld_1d_v4i8_trap:
3531 case Intrinsic::nvvm_suld_1d_array_i8_trap:
3532 case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3533 case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3534 case Intrinsic::nvvm_suld_2d_i8_trap:
3535 case Intrinsic::nvvm_suld_2d_v2i8_trap:
3536 case Intrinsic::nvvm_suld_2d_v4i8_trap:
3537 case Intrinsic::nvvm_suld_2d_array_i8_trap:
3538 case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3539 case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3540 case Intrinsic::nvvm_suld_3d_i8_trap:
3541 case Intrinsic::nvvm_suld_3d_v2i8_trap:
3542 case Intrinsic::nvvm_suld_3d_v4i8_trap:
3543 case Intrinsic::nvvm_suld_1d_i8_zero:
3544 case Intrinsic::nvvm_suld_1d_v2i8_zero:
3545 case Intrinsic::nvvm_suld_1d_v4i8_zero:
3546 case Intrinsic::nvvm_suld_1d_array_i8_zero:
3547 case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3548 case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3549 case Intrinsic::nvvm_suld_2d_i8_zero:
3550 case Intrinsic::nvvm_suld_2d_v2i8_zero:
3551 case Intrinsic::nvvm_suld_2d_v4i8_zero:
3552 case Intrinsic::nvvm_suld_2d_array_i8_zero:
3553 case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3554 case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3555 case Intrinsic::nvvm_suld_3d_i8_zero:
3556 case Intrinsic::nvvm_suld_3d_v2i8_zero:
3557 case Intrinsic::nvvm_suld_3d_v4i8_zero: {
3558 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3559 Info.memVT = MVT::i8;
3560 Info.ptrVal = nullptr;
3563 Info.readMem = true;
3564 Info.writeMem = false;
3568 case Intrinsic::nvvm_suld_1d_i16_clamp:
3569 case Intrinsic::nvvm_suld_1d_v2i16_clamp:
3570 case Intrinsic::nvvm_suld_1d_v4i16_clamp:
3571 case Intrinsic::nvvm_suld_1d_array_i16_clamp:
3572 case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
3573 case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
3574 case Intrinsic::nvvm_suld_2d_i16_clamp:
3575 case Intrinsic::nvvm_suld_2d_v2i16_clamp:
3576 case Intrinsic::nvvm_suld_2d_v4i16_clamp:
3577 case Intrinsic::nvvm_suld_2d_array_i16_clamp:
3578 case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
3579 case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
3580 case Intrinsic::nvvm_suld_3d_i16_clamp:
3581 case Intrinsic::nvvm_suld_3d_v2i16_clamp:
3582 case Intrinsic::nvvm_suld_3d_v4i16_clamp:
3583 case Intrinsic::nvvm_suld_1d_i16_trap:
3584 case Intrinsic::nvvm_suld_1d_v2i16_trap:
3585 case Intrinsic::nvvm_suld_1d_v4i16_trap:
3586 case Intrinsic::nvvm_suld_1d_array_i16_trap:
3587 case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3588 case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3589 case Intrinsic::nvvm_suld_2d_i16_trap:
3590 case Intrinsic::nvvm_suld_2d_v2i16_trap:
3591 case Intrinsic::nvvm_suld_2d_v4i16_trap:
3592 case Intrinsic::nvvm_suld_2d_array_i16_trap:
3593 case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3594 case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3595 case Intrinsic::nvvm_suld_3d_i16_trap:
3596 case Intrinsic::nvvm_suld_3d_v2i16_trap:
3597 case Intrinsic::nvvm_suld_3d_v4i16_trap:
3598 case Intrinsic::nvvm_suld_1d_i16_zero:
3599 case Intrinsic::nvvm_suld_1d_v2i16_zero:
3600 case Intrinsic::nvvm_suld_1d_v4i16_zero:
3601 case Intrinsic::nvvm_suld_1d_array_i16_zero:
3602 case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3603 case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3604 case Intrinsic::nvvm_suld_2d_i16_zero:
3605 case Intrinsic::nvvm_suld_2d_v2i16_zero:
3606 case Intrinsic::nvvm_suld_2d_v4i16_zero:
3607 case Intrinsic::nvvm_suld_2d_array_i16_zero:
3608 case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3609 case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3610 case Intrinsic::nvvm_suld_3d_i16_zero:
3611 case Intrinsic::nvvm_suld_3d_v2i16_zero:
3612 case Intrinsic::nvvm_suld_3d_v4i16_zero: {
3613 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3614 Info.memVT = MVT::i16;
3615 Info.ptrVal = nullptr;
3618 Info.readMem = true;
3619 Info.writeMem = false;
3623 case Intrinsic::nvvm_suld_1d_i32_clamp:
3624 case Intrinsic::nvvm_suld_1d_v2i32_clamp:
3625 case Intrinsic::nvvm_suld_1d_v4i32_clamp:
3626 case Intrinsic::nvvm_suld_1d_array_i32_clamp:
3627 case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
3628 case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
3629 case Intrinsic::nvvm_suld_2d_i32_clamp:
3630 case Intrinsic::nvvm_suld_2d_v2i32_clamp:
3631 case Intrinsic::nvvm_suld_2d_v4i32_clamp:
3632 case Intrinsic::nvvm_suld_2d_array_i32_clamp:
3633 case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
3634 case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
3635 case Intrinsic::nvvm_suld_3d_i32_clamp:
3636 case Intrinsic::nvvm_suld_3d_v2i32_clamp:
3637 case Intrinsic::nvvm_suld_3d_v4i32_clamp:
3638 case Intrinsic::nvvm_suld_1d_i32_trap:
3639 case Intrinsic::nvvm_suld_1d_v2i32_trap:
3640 case Intrinsic::nvvm_suld_1d_v4i32_trap:
3641 case Intrinsic::nvvm_suld_1d_array_i32_trap:
3642 case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3643 case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3644 case Intrinsic::nvvm_suld_2d_i32_trap:
3645 case Intrinsic::nvvm_suld_2d_v2i32_trap:
3646 case Intrinsic::nvvm_suld_2d_v4i32_trap:
3647 case Intrinsic::nvvm_suld_2d_array_i32_trap:
3648 case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3649 case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3650 case Intrinsic::nvvm_suld_3d_i32_trap:
3651 case Intrinsic::nvvm_suld_3d_v2i32_trap:
3652 case Intrinsic::nvvm_suld_3d_v4i32_trap:
3653 case Intrinsic::nvvm_suld_1d_i32_zero:
3654 case Intrinsic::nvvm_suld_1d_v2i32_zero:
3655 case Intrinsic::nvvm_suld_1d_v4i32_zero:
3656 case Intrinsic::nvvm_suld_1d_array_i32_zero:
3657 case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3658 case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3659 case Intrinsic::nvvm_suld_2d_i32_zero:
3660 case Intrinsic::nvvm_suld_2d_v2i32_zero:
3661 case Intrinsic::nvvm_suld_2d_v4i32_zero:
3662 case Intrinsic::nvvm_suld_2d_array_i32_zero:
3663 case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3664 case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3665 case Intrinsic::nvvm_suld_3d_i32_zero:
3666 case Intrinsic::nvvm_suld_3d_v2i32_zero:
3667 case Intrinsic::nvvm_suld_3d_v4i32_zero: {
3668 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3669 Info.memVT = MVT::i32;
3670 Info.ptrVal = nullptr;
3673 Info.readMem = true;
3674 Info.writeMem = false;
3678 case Intrinsic::nvvm_suld_1d_i64_clamp:
3679 case Intrinsic::nvvm_suld_1d_v2i64_clamp:
3680 case Intrinsic::nvvm_suld_1d_array_i64_clamp:
3681 case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
3682 case Intrinsic::nvvm_suld_2d_i64_clamp:
3683 case Intrinsic::nvvm_suld_2d_v2i64_clamp:
3684 case Intrinsic::nvvm_suld_2d_array_i64_clamp:
3685 case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
3686 case Intrinsic::nvvm_suld_3d_i64_clamp:
3687 case Intrinsic::nvvm_suld_3d_v2i64_clamp:
3688 case Intrinsic::nvvm_suld_1d_i64_trap:
3689 case Intrinsic::nvvm_suld_1d_v2i64_trap:
3690 case Intrinsic::nvvm_suld_1d_array_i64_trap:
3691 case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3692 case Intrinsic::nvvm_suld_2d_i64_trap:
3693 case Intrinsic::nvvm_suld_2d_v2i64_trap:
3694 case Intrinsic::nvvm_suld_2d_array_i64_trap:
3695 case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3696 case Intrinsic::nvvm_suld_3d_i64_trap:
3697 case Intrinsic::nvvm_suld_3d_v2i64_trap:
3698 case Intrinsic::nvvm_suld_1d_i64_zero:
3699 case Intrinsic::nvvm_suld_1d_v2i64_zero:
3700 case Intrinsic::nvvm_suld_1d_array_i64_zero:
3701 case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3702 case Intrinsic::nvvm_suld_2d_i64_zero:
3703 case Intrinsic::nvvm_suld_2d_v2i64_zero:
3704 case Intrinsic::nvvm_suld_2d_array_i64_zero:
3705 case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3706 case Intrinsic::nvvm_suld_3d_i64_zero:
3707 case Intrinsic::nvvm_suld_3d_v2i64_zero: {
3708 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3709 Info.memVT = MVT::i64;
3710 Info.ptrVal = nullptr;
3713 Info.readMem = true;
3714 Info.writeMem = false;
3722 /// isLegalAddressingMode - Return true if the addressing mode represented
3723 /// by AM is legal for this target, for a load/store of the specified type.
3724 /// Used to guide target specific optimizations, like loop strength reduction
3725 /// (LoopStrengthReduce.cpp) and memory optimization for address mode
3726 /// (CodeGenPrepare.cpp)
3727 bool NVPTXTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3730 // AddrMode - This represents an addressing mode of:
3731 // BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
3733 // The legal address modes are
3740 if (AM.BaseOffs || AM.HasBaseReg || AM.Scale)
3746 case 0: // "r", "r+i" or "i" is allowed
3749 if (AM.HasBaseReg) // "r+r+i" or "r+r" is not allowed.
3751 // Otherwise we have r+i.
3754 // No scale > 1 is allowed
3760 //===----------------------------------------------------------------------===//
3761 // NVPTX Inline Assembly Support
3762 //===----------------------------------------------------------------------===//
3764 /// getConstraintType - Given a constraint letter, return the type of
3765 /// constraint it is for this target.
3766 NVPTXTargetLowering::ConstraintType
3767 NVPTXTargetLowering::getConstraintType(const std::string &Constraint) const {
3768 if (Constraint.size() == 1) {
3769 switch (Constraint[0]) {
3781 return C_RegisterClass;
3784 return TargetLowering::getConstraintType(Constraint);
3787 std::pair<unsigned, const TargetRegisterClass *>
3788 NVPTXTargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
3789 const std::string &Constraint,
3791 if (Constraint.size() == 1) {
3792 switch (Constraint[0]) {
3794 return std::make_pair(0U, &NVPTX::Int1RegsRegClass);
3796 return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
3798 return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
3800 return std::make_pair(0U, &NVPTX::Int32RegsRegClass);
3803 return std::make_pair(0U, &NVPTX::Int64RegsRegClass);
3805 return std::make_pair(0U, &NVPTX::Float32RegsRegClass);
3807 return std::make_pair(0U, &NVPTX::Float64RegsRegClass);
3810 return TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
3813 /// getFunctionAlignment - Return the Log2 alignment of this function.
3814 unsigned NVPTXTargetLowering::getFunctionAlignment(const Function *) const {
3818 //===----------------------------------------------------------------------===//
3819 // NVPTX DAG Combining
3820 //===----------------------------------------------------------------------===//
3822 bool NVPTXTargetLowering::allowFMA(MachineFunction &MF,
3823 CodeGenOpt::Level OptLevel) const {
3824 const Function *F = MF.getFunction();
3825 const TargetOptions &TO = MF.getTarget().Options;
3827 // Always honor command-line argument
3828 if (FMAContractLevelOpt.getNumOccurrences() > 0) {
3829 return FMAContractLevelOpt > 0;
3830 } else if (OptLevel == 0) {
3831 // Do not contract if we're not optimizing the code
3833 } else if (TO.AllowFPOpFusion == FPOpFusion::Fast || TO.UnsafeFPMath) {
3834 // Honor TargetOptions flags that explicitly say fusion is okay
3836 } else if (F->hasFnAttribute("unsafe-fp-math")) {
3837 // Check for unsafe-fp-math=true coming from Clang
3838 Attribute Attr = F->getFnAttribute("unsafe-fp-math");
3839 StringRef Val = Attr.getValueAsString();
3844 // We did not have a clear indication that fusion is allowed, so assume not
3848 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
3849 /// operands N0 and N1. This is a helper for PerformADDCombine that is
3850 /// called with the default operands, and if that fails, with commuted
3852 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
3853 TargetLowering::DAGCombinerInfo &DCI,
3854 const NVPTXSubtarget &Subtarget,
3855 CodeGenOpt::Level OptLevel) {
3856 SelectionDAG &DAG = DCI.DAG;
3857 // Skip non-integer, non-scalar case
3858 EVT VT=N0.getValueType();
3862 // fold (add (mul a, b), c) -> (mad a, b, c)
3864 if (N0.getOpcode() == ISD::MUL) {
3865 assert (VT.isInteger());
3867 // Since integer multiply-add costs the same as integer multiply
3868 // but is more costly than integer add, do the fusion only when
3869 // the mul is only used in the add.
3870 if (OptLevel==CodeGenOpt::None || VT != MVT::i32 ||
3871 !N0.getNode()->hasOneUse())
3875 return DAG.getNode(NVPTXISD::IMAD, SDLoc(N), VT,
3876 N0.getOperand(0), N0.getOperand(1), N1);
3878 else if (N0.getOpcode() == ISD::FMUL) {
3879 if (VT == MVT::f32 || VT == MVT::f64) {
3880 const auto *TLI = static_cast<const NVPTXTargetLowering *>(
3881 &DAG.getTargetLoweringInfo());
3882 if (!TLI->allowFMA(DAG.getMachineFunction(), OptLevel))
3885 // For floating point:
3886 // Do the fusion only when the mul has less than 5 uses and all
3888 // The heuristic is that if a use is not an add, then that use
3889 // cannot be fused into fma, therefore mul is still needed anyway.
3890 // If there are more than 4 uses, even if they are all add, fusing
3891 // them will increase register pressue.
3894 int nonAddCount = 0;
3895 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
3896 UE = N0.getNode()->use_end();
3900 if (User->getOpcode() != ISD::FADD)
3906 int orderNo = N->getIROrder();
3907 int orderNo2 = N0.getNode()->getIROrder();
3908 // simple heuristics here for considering potential register
3909 // pressure, the logics here is that the differnce are used
3910 // to measure the distance between def and use, the longer distance
3911 // more likely cause register pressure.
3912 if (orderNo - orderNo2 < 500)
3915 // Now, check if at least one of the FMUL's operands is live beyond the node N,
3916 // which guarantees that the FMA will not increase register pressure at node N.
3917 bool opIsLive = false;
3918 const SDNode *left = N0.getOperand(0).getNode();
3919 const SDNode *right = N0.getOperand(1).getNode();
3921 if (isa<ConstantSDNode>(left) || isa<ConstantSDNode>(right))
3925 for (SDNode::use_iterator UI = left->use_begin(), UE = left->use_end(); UI != UE; ++UI) {
3927 int orderNo3 = User->getIROrder();
3928 if (orderNo3 > orderNo) {
3935 for (SDNode::use_iterator UI = right->use_begin(), UE = right->use_end(); UI != UE; ++UI) {
3937 int orderNo3 = User->getIROrder();
3938 if (orderNo3 > orderNo) {
3948 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
3949 N0.getOperand(0), N0.getOperand(1), N1);
3956 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3958 static SDValue PerformADDCombine(SDNode *N,
3959 TargetLowering::DAGCombinerInfo &DCI,
3960 const NVPTXSubtarget &Subtarget,
3961 CodeGenOpt::Level OptLevel) {
3962 SDValue N0 = N->getOperand(0);
3963 SDValue N1 = N->getOperand(1);
3965 // First try with the default operand order.
3966 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget,
3968 if (Result.getNode())
3971 // If that didn't work, try again with the operands commuted.
3972 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget, OptLevel);
3975 static SDValue PerformANDCombine(SDNode *N,
3976 TargetLowering::DAGCombinerInfo &DCI) {
3977 // The type legalizer turns a vector load of i8 values into a zextload to i16
3978 // registers, optionally ANY_EXTENDs it (if target type is integer),
3979 // and ANDs off the high 8 bits. Since we turn this load into a
3980 // target-specific DAG node, the DAG combiner fails to eliminate these AND
3981 // nodes. Do that here.
3982 SDValue Val = N->getOperand(0);
3983 SDValue Mask = N->getOperand(1);
3985 if (isa<ConstantSDNode>(Val)) {
3986 std::swap(Val, Mask);
3990 // Generally, we will see zextload -> IMOV16rr -> ANY_EXTEND -> and
3991 if (Val.getOpcode() == ISD::ANY_EXTEND) {
3993 Val = Val->getOperand(0);
3996 if (Val->isMachineOpcode() && Val->getMachineOpcode() == NVPTX::IMOV16rr) {
3997 Val = Val->getOperand(0);
4000 if (Val->getOpcode() == NVPTXISD::LoadV2 ||
4001 Val->getOpcode() == NVPTXISD::LoadV4) {
4002 ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(Mask);
4004 // Not an AND with a constant
4008 uint64_t MaskVal = MaskCnst->getZExtValue();
4009 if (MaskVal != 0xff) {
4010 // Not an AND that chops off top 8 bits
4014 MemSDNode *Mem = dyn_cast<MemSDNode>(Val);
4016 // Not a MemSDNode?!?
4020 EVT MemVT = Mem->getMemoryVT();
4021 if (MemVT != MVT::v2i8 && MemVT != MVT::v4i8) {
4022 // We only handle the i8 case
4027 cast<ConstantSDNode>(Val->getOperand(Val->getNumOperands()-1))->
4029 if (ExtType == ISD::SEXTLOAD) {
4030 // If for some reason the load is a sextload, the and is needed to zero
4031 // out the high 8 bits
4036 if (AExt.getNode() != 0) {
4037 // Re-insert the ext as a zext.
4038 Val = DCI.DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
4039 AExt.getValueType(), Val);
4043 // If we get here, the AND is unnecessary. Just replace it with the load
4044 DCI.CombineTo(N, Val, AddTo);
4050 enum OperandSignedness {
4056 /// IsMulWideOperandDemotable - Checks if the provided DAG node is an operand
4057 /// that can be demoted to \p OptSize bits without loss of information. The
4058 /// signedness of the operand, if determinable, is placed in \p S.
4059 static bool IsMulWideOperandDemotable(SDValue Op,
4061 OperandSignedness &S) {
4064 if (Op.getOpcode() == ISD::SIGN_EXTEND ||
4065 Op.getOpcode() == ISD::SIGN_EXTEND_INREG) {
4066 EVT OrigVT = Op.getOperand(0).getValueType();
4067 if (OrigVT.getSizeInBits() <= OptSize) {
4071 } else if (Op.getOpcode() == ISD::ZERO_EXTEND) {
4072 EVT OrigVT = Op.getOperand(0).getValueType();
4073 if (OrigVT.getSizeInBits() <= OptSize) {
4082 /// AreMulWideOperandsDemotable - Checks if the given LHS and RHS operands can
4083 /// be demoted to \p OptSize bits without loss of information. If the operands
4084 /// contain a constant, it should appear as the RHS operand. The signedness of
4085 /// the operands is placed in \p IsSigned.
4086 static bool AreMulWideOperandsDemotable(SDValue LHS, SDValue RHS,
4090 OperandSignedness LHSSign;
4092 // The LHS operand must be a demotable op
4093 if (!IsMulWideOperandDemotable(LHS, OptSize, LHSSign))
4096 // We should have been able to determine the signedness from the LHS
4097 if (LHSSign == Unknown)
4100 IsSigned = (LHSSign == Signed);
4102 // The RHS can be a demotable op or a constant
4103 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(RHS)) {
4104 APInt Val = CI->getAPIntValue();
4105 if (LHSSign == Unsigned) {
4106 if (Val.isIntN(OptSize)) {
4111 if (Val.isSignedIntN(OptSize)) {
4117 OperandSignedness RHSSign;
4118 if (!IsMulWideOperandDemotable(RHS, OptSize, RHSSign))
4121 if (LHSSign != RHSSign)
4128 /// TryMULWIDECombine - Attempt to replace a multiply of M bits with a multiply
4129 /// of M/2 bits that produces an M-bit result (i.e. mul.wide). This transform
4130 /// works on both multiply DAG nodes and SHL DAG nodes with a constant shift
4132 static SDValue TryMULWIDECombine(SDNode *N,
4133 TargetLowering::DAGCombinerInfo &DCI) {
4134 EVT MulType = N->getValueType(0);
4135 if (MulType != MVT::i32 && MulType != MVT::i64) {
4140 unsigned OptSize = MulType.getSizeInBits() >> 1;
4141 SDValue LHS = N->getOperand(0);
4142 SDValue RHS = N->getOperand(1);
4144 // Canonicalize the multiply so the constant (if any) is on the right
4145 if (N->getOpcode() == ISD::MUL) {
4146 if (isa<ConstantSDNode>(LHS)) {
4147 std::swap(LHS, RHS);
4151 // If we have a SHL, determine the actual multiply amount
4152 if (N->getOpcode() == ISD::SHL) {
4153 ConstantSDNode *ShlRHS = dyn_cast<ConstantSDNode>(RHS);
4158 APInt ShiftAmt = ShlRHS->getAPIntValue();
4159 unsigned BitWidth = MulType.getSizeInBits();
4160 if (ShiftAmt.sge(0) && ShiftAmt.slt(BitWidth)) {
4161 APInt MulVal = APInt(BitWidth, 1) << ShiftAmt;
4162 RHS = DCI.DAG.getConstant(MulVal, DL, MulType);
4169 // Verify that our operands are demotable
4170 if (!AreMulWideOperandsDemotable(LHS, RHS, OptSize, Signed)) {
4175 if (MulType == MVT::i32) {
4176 DemotedVT = MVT::i16;
4178 DemotedVT = MVT::i32;
4181 // Truncate the operands to the correct size. Note that these are just for
4182 // type consistency and will (likely) be eliminated in later phases.
4184 DCI.DAG.getNode(ISD::TRUNCATE, DL, DemotedVT, LHS);
4186 DCI.DAG.getNode(ISD::TRUNCATE, DL, DemotedVT, RHS);
4190 Opc = NVPTXISD::MUL_WIDE_SIGNED;
4192 Opc = NVPTXISD::MUL_WIDE_UNSIGNED;
4195 return DCI.DAG.getNode(Opc, DL, MulType, TruncLHS, TruncRHS);
4198 /// PerformMULCombine - Runs PTX-specific DAG combine patterns on MUL nodes.
4199 static SDValue PerformMULCombine(SDNode *N,
4200 TargetLowering::DAGCombinerInfo &DCI,
4201 CodeGenOpt::Level OptLevel) {
4203 // Try mul.wide combining at OptLevel > 0
4204 SDValue Ret = TryMULWIDECombine(N, DCI);
4212 /// PerformSHLCombine - Runs PTX-specific DAG combine patterns on SHL nodes.
4213 static SDValue PerformSHLCombine(SDNode *N,
4214 TargetLowering::DAGCombinerInfo &DCI,
4215 CodeGenOpt::Level OptLevel) {
4217 // Try mul.wide combining at OptLevel > 0
4218 SDValue Ret = TryMULWIDECombine(N, DCI);
4226 SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
4227 DAGCombinerInfo &DCI) const {
4228 CodeGenOpt::Level OptLevel = getTargetMachine().getOptLevel();
4229 switch (N->getOpcode()) {
4233 return PerformADDCombine(N, DCI, STI, OptLevel);
4235 return PerformMULCombine(N, DCI, OptLevel);
4237 return PerformSHLCombine(N, DCI, OptLevel);
4239 return PerformANDCombine(N, DCI);
4244 /// ReplaceVectorLoad - Convert vector loads into multi-output scalar loads.
4245 static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG,
4246 const DataLayout *TD,
4247 SmallVectorImpl<SDValue> &Results) {
4248 EVT ResVT = N->getValueType(0);
4251 assert(ResVT.isVector() && "Vector load must have vector type");
4253 // We only handle "native" vector sizes for now, e.g. <4 x double> is not
4254 // legal. We can (and should) split that into 2 loads of <2 x double> here
4255 // but I'm leaving that as a TODO for now.
4256 assert(ResVT.isSimple() && "Can only handle simple types");
4257 switch (ResVT.getSimpleVT().SimpleTy) {
4270 // This is a "native" vector type
4274 LoadSDNode *LD = cast<LoadSDNode>(N);
4276 unsigned Align = LD->getAlignment();
4277 unsigned PrefAlign =
4278 TD->getPrefTypeAlignment(ResVT.getTypeForEVT(*DAG.getContext()));
4279 if (Align < PrefAlign) {
4280 // This load is not sufficiently aligned, so bail out and let this vector
4281 // load be scalarized. Note that we may still be able to emit smaller
4282 // vector loads. For example, if we are loading a <4 x float> with an
4283 // alignment of 8, this check will fail but the legalizer will try again
4284 // with 2 x <2 x float>, which will succeed with an alignment of 8.
4288 EVT EltVT = ResVT.getVectorElementType();
4289 unsigned NumElts = ResVT.getVectorNumElements();
4291 // Since LoadV2 is a target node, we cannot rely on DAG type legalization.
4292 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
4293 // loaded type to i16 and propagate the "real" type as the memory type.
4294 bool NeedTrunc = false;
4295 if (EltVT.getSizeInBits() < 16) {
4300 unsigned Opcode = 0;
4307 Opcode = NVPTXISD::LoadV2;
4308 LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4311 Opcode = NVPTXISD::LoadV4;
4312 EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4313 LdResVTs = DAG.getVTList(ListVTs);
4318 // Copy regular operands
4319 SmallVector<SDValue, 8> OtherOps(N->op_begin(), N->op_end());
4321 // The select routine does not have access to the LoadSDNode instance, so
4322 // pass along the extension information
4323 OtherOps.push_back(DAG.getIntPtrConstant(LD->getExtensionType(), DL));
4325 SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4327 LD->getMemOperand());
4329 SmallVector<SDValue, 4> ScalarRes;
4331 for (unsigned i = 0; i < NumElts; ++i) {
4332 SDValue Res = NewLD.getValue(i);
4334 Res = DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4335 ScalarRes.push_back(Res);
4338 SDValue LoadChain = NewLD.getValue(NumElts);
4340 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes);
4342 Results.push_back(BuildVec);
4343 Results.push_back(LoadChain);
4346 static void ReplaceINTRINSIC_W_CHAIN(SDNode *N, SelectionDAG &DAG,
4347 SmallVectorImpl<SDValue> &Results) {
4348 SDValue Chain = N->getOperand(0);
4349 SDValue Intrin = N->getOperand(1);
4352 // Get the intrinsic ID
4353 unsigned IntrinNo = cast<ConstantSDNode>(Intrin.getNode())->getZExtValue();
4357 case Intrinsic::nvvm_ldg_global_i:
4358 case Intrinsic::nvvm_ldg_global_f:
4359 case Intrinsic::nvvm_ldg_global_p:
4360 case Intrinsic::nvvm_ldu_global_i:
4361 case Intrinsic::nvvm_ldu_global_f:
4362 case Intrinsic::nvvm_ldu_global_p: {
4363 EVT ResVT = N->getValueType(0);
4365 if (ResVT.isVector()) {
4368 unsigned NumElts = ResVT.getVectorNumElements();
4369 EVT EltVT = ResVT.getVectorElementType();
4371 // Since LDU/LDG are target nodes, we cannot rely on DAG type
4373 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
4374 // loaded type to i16 and propagate the "real" type as the memory type.
4375 bool NeedTrunc = false;
4376 if (EltVT.getSizeInBits() < 16) {
4381 unsigned Opcode = 0;
4391 case Intrinsic::nvvm_ldg_global_i:
4392 case Intrinsic::nvvm_ldg_global_f:
4393 case Intrinsic::nvvm_ldg_global_p:
4394 Opcode = NVPTXISD::LDGV2;
4396 case Intrinsic::nvvm_ldu_global_i:
4397 case Intrinsic::nvvm_ldu_global_f:
4398 case Intrinsic::nvvm_ldu_global_p:
4399 Opcode = NVPTXISD::LDUV2;
4402 LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4408 case Intrinsic::nvvm_ldg_global_i:
4409 case Intrinsic::nvvm_ldg_global_f:
4410 case Intrinsic::nvvm_ldg_global_p:
4411 Opcode = NVPTXISD::LDGV4;
4413 case Intrinsic::nvvm_ldu_global_i:
4414 case Intrinsic::nvvm_ldu_global_f:
4415 case Intrinsic::nvvm_ldu_global_p:
4416 Opcode = NVPTXISD::LDUV4;
4419 EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4420 LdResVTs = DAG.getVTList(ListVTs);
4425 SmallVector<SDValue, 8> OtherOps;
4427 // Copy regular operands
4429 OtherOps.push_back(Chain); // Chain
4430 // Skip operand 1 (intrinsic ID)
4432 OtherOps.append(N->op_begin() + 2, N->op_end());
4434 MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4436 SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4437 MemSD->getMemoryVT(),
4438 MemSD->getMemOperand());
4440 SmallVector<SDValue, 4> ScalarRes;
4442 for (unsigned i = 0; i < NumElts; ++i) {
4443 SDValue Res = NewLD.getValue(i);
4446 DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4447 ScalarRes.push_back(Res);
4450 SDValue LoadChain = NewLD.getValue(NumElts);
4453 DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes);
4455 Results.push_back(BuildVec);
4456 Results.push_back(LoadChain);
4459 assert(ResVT.isSimple() && ResVT.getSimpleVT().SimpleTy == MVT::i8 &&
4460 "Custom handling of non-i8 ldu/ldg?");
4462 // Just copy all operands as-is
4463 SmallVector<SDValue, 4> Ops(N->op_begin(), N->op_end());
4465 // Force output to i16
4466 SDVTList LdResVTs = DAG.getVTList(MVT::i16, MVT::Other);
4468 MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4470 // We make sure the memory type is i8, which will be used during isel
4471 // to select the proper instruction.
4473 DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, LdResVTs, Ops,
4474 MVT::i8, MemSD->getMemOperand());
4476 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
4477 NewLD.getValue(0)));
4478 Results.push_back(NewLD.getValue(1));
4484 void NVPTXTargetLowering::ReplaceNodeResults(
4485 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
4486 switch (N->getOpcode()) {
4488 report_fatal_error("Unhandled custom legalization");
4490 ReplaceLoadVector(N, DAG, getDataLayout(), Results);
4492 case ISD::INTRINSIC_W_CHAIN:
4493 ReplaceINTRINSIC_W_CHAIN(N, DAG, Results);
4498 // Pin NVPTXSection's and NVPTXTargetObjectFile's vtables to this file.
4499 void NVPTXSection::anchor() {}
4501 NVPTXTargetObjectFile::~NVPTXTargetObjectFile() {
4505 delete ReadOnlySection;
4507 delete StaticCtorSection;
4508 delete StaticDtorSection;
4510 delete EHFrameSection;
4511 delete DwarfAbbrevSection;
4512 delete DwarfInfoSection;
4513 delete DwarfLineSection;
4514 delete DwarfFrameSection;
4515 delete DwarfPubTypesSection;
4516 delete DwarfDebugInlineSection;
4517 delete DwarfStrSection;
4518 delete DwarfLocSection;
4519 delete DwarfARangesSection;
4520 delete DwarfRangesSection;
4524 NVPTXTargetObjectFile::SelectSectionForGlobal(const GlobalValue *GV,
4525 SectionKind Kind, Mangler &Mang,
4526 const TargetMachine &TM) const {
4527 return getDataSection();