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 : TargetLowering(TM), nvTM(&TM),
111 nvptxSubtarget(TM.getSubtarget<NVPTXSubtarget>()) {
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 (nvptxSubtarget.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 (nvptxSubtarget.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::f16, Expand);
207 setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);
208 // Turn FP truncstore into trunc + store.
209 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
210 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
211 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
213 // PTX does not support load / store predicate registers
214 setOperationAction(ISD::LOAD, MVT::i1, Custom);
215 setOperationAction(ISD::STORE, MVT::i1, Custom);
217 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
218 setLoadExtAction(ISD::ZEXTLOAD, MVT::i1, Promote);
219 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
220 setTruncStoreAction(MVT::i32, MVT::i1, Expand);
221 setTruncStoreAction(MVT::i16, MVT::i1, Expand);
222 setTruncStoreAction(MVT::i8, 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 (int i = MVT::FIRST_VECTOR_VALUETYPE; i <= MVT::LAST_VECTOR_VALUETYPE;
237 MVT VT = (MVT::SimpleValueType) i;
238 if (IsPTXVectorType(VT)) {
239 setOperationAction(ISD::LOAD, VT, Custom);
240 setOperationAction(ISD::STORE, VT, Custom);
241 setOperationAction(ISD::INTRINSIC_W_CHAIN, VT, Custom);
245 // Custom handling for i8 intrinsics
246 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);
248 setOperationAction(ISD::CTLZ, MVT::i16, Legal);
249 setOperationAction(ISD::CTLZ, MVT::i32, Legal);
250 setOperationAction(ISD::CTLZ, MVT::i64, Legal);
251 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16, Legal);
252 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Legal);
253 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Legal);
254 setOperationAction(ISD::CTTZ, MVT::i16, Expand);
255 setOperationAction(ISD::CTTZ, MVT::i32, Expand);
256 setOperationAction(ISD::CTTZ, MVT::i64, Expand);
257 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16, Expand);
258 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
259 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
260 setOperationAction(ISD::CTPOP, MVT::i16, Legal);
261 setOperationAction(ISD::CTPOP, MVT::i32, Legal);
262 setOperationAction(ISD::CTPOP, MVT::i64, Legal);
264 // We have some custom DAG combine patterns for these nodes
265 setTargetDAGCombine(ISD::ADD);
266 setTargetDAGCombine(ISD::AND);
267 setTargetDAGCombine(ISD::FADD);
268 setTargetDAGCombine(ISD::MUL);
269 setTargetDAGCombine(ISD::SHL);
271 // Now deduce the information based on the above mentioned
273 computeRegisterProperties();
276 const char *NVPTXTargetLowering::getTargetNodeName(unsigned Opcode) const {
281 return "NVPTXISD::CALL";
282 case NVPTXISD::RET_FLAG:
283 return "NVPTXISD::RET_FLAG";
284 case NVPTXISD::Wrapper:
285 return "NVPTXISD::Wrapper";
286 case NVPTXISD::DeclareParam:
287 return "NVPTXISD::DeclareParam";
288 case NVPTXISD::DeclareScalarParam:
289 return "NVPTXISD::DeclareScalarParam";
290 case NVPTXISD::DeclareRet:
291 return "NVPTXISD::DeclareRet";
292 case NVPTXISD::DeclareRetParam:
293 return "NVPTXISD::DeclareRetParam";
294 case NVPTXISD::PrintCall:
295 return "NVPTXISD::PrintCall";
296 case NVPTXISD::LoadParam:
297 return "NVPTXISD::LoadParam";
298 case NVPTXISD::LoadParamV2:
299 return "NVPTXISD::LoadParamV2";
300 case NVPTXISD::LoadParamV4:
301 return "NVPTXISD::LoadParamV4";
302 case NVPTXISD::StoreParam:
303 return "NVPTXISD::StoreParam";
304 case NVPTXISD::StoreParamV2:
305 return "NVPTXISD::StoreParamV2";
306 case NVPTXISD::StoreParamV4:
307 return "NVPTXISD::StoreParamV4";
308 case NVPTXISD::StoreParamS32:
309 return "NVPTXISD::StoreParamS32";
310 case NVPTXISD::StoreParamU32:
311 return "NVPTXISD::StoreParamU32";
312 case NVPTXISD::CallArgBegin:
313 return "NVPTXISD::CallArgBegin";
314 case NVPTXISD::CallArg:
315 return "NVPTXISD::CallArg";
316 case NVPTXISD::LastCallArg:
317 return "NVPTXISD::LastCallArg";
318 case NVPTXISD::CallArgEnd:
319 return "NVPTXISD::CallArgEnd";
320 case NVPTXISD::CallVoid:
321 return "NVPTXISD::CallVoid";
322 case NVPTXISD::CallVal:
323 return "NVPTXISD::CallVal";
324 case NVPTXISD::CallSymbol:
325 return "NVPTXISD::CallSymbol";
326 case NVPTXISD::Prototype:
327 return "NVPTXISD::Prototype";
328 case NVPTXISD::MoveParam:
329 return "NVPTXISD::MoveParam";
330 case NVPTXISD::StoreRetval:
331 return "NVPTXISD::StoreRetval";
332 case NVPTXISD::StoreRetvalV2:
333 return "NVPTXISD::StoreRetvalV2";
334 case NVPTXISD::StoreRetvalV4:
335 return "NVPTXISD::StoreRetvalV4";
336 case NVPTXISD::PseudoUseParam:
337 return "NVPTXISD::PseudoUseParam";
338 case NVPTXISD::RETURN:
339 return "NVPTXISD::RETURN";
340 case NVPTXISD::CallSeqBegin:
341 return "NVPTXISD::CallSeqBegin";
342 case NVPTXISD::CallSeqEnd:
343 return "NVPTXISD::CallSeqEnd";
344 case NVPTXISD::CallPrototype:
345 return "NVPTXISD::CallPrototype";
346 case NVPTXISD::LoadV2:
347 return "NVPTXISD::LoadV2";
348 case NVPTXISD::LoadV4:
349 return "NVPTXISD::LoadV4";
350 case NVPTXISD::LDGV2:
351 return "NVPTXISD::LDGV2";
352 case NVPTXISD::LDGV4:
353 return "NVPTXISD::LDGV4";
354 case NVPTXISD::LDUV2:
355 return "NVPTXISD::LDUV2";
356 case NVPTXISD::LDUV4:
357 return "NVPTXISD::LDUV4";
358 case NVPTXISD::StoreV2:
359 return "NVPTXISD::StoreV2";
360 case NVPTXISD::StoreV4:
361 return "NVPTXISD::StoreV4";
362 case NVPTXISD::FUN_SHFL_CLAMP:
363 return "NVPTXISD::FUN_SHFL_CLAMP";
364 case NVPTXISD::FUN_SHFR_CLAMP:
365 return "NVPTXISD::FUN_SHFR_CLAMP";
367 return "NVPTXISD::IMAD";
368 case NVPTXISD::MUL_WIDE_SIGNED:
369 return "NVPTXISD::MUL_WIDE_SIGNED";
370 case NVPTXISD::MUL_WIDE_UNSIGNED:
371 return "NVPTXISD::MUL_WIDE_UNSIGNED";
372 case NVPTXISD::Tex1DFloatS32: return "NVPTXISD::Tex1DFloatS32";
373 case NVPTXISD::Tex1DFloatFloat: return "NVPTXISD::Tex1DFloatFloat";
374 case NVPTXISD::Tex1DFloatFloatLevel:
375 return "NVPTXISD::Tex1DFloatFloatLevel";
376 case NVPTXISD::Tex1DFloatFloatGrad:
377 return "NVPTXISD::Tex1DFloatFloatGrad";
378 case NVPTXISD::Tex1DS32S32: return "NVPTXISD::Tex1DS32S32";
379 case NVPTXISD::Tex1DS32Float: return "NVPTXISD::Tex1DS32Float";
380 case NVPTXISD::Tex1DS32FloatLevel:
381 return "NVPTXISD::Tex1DS32FloatLevel";
382 case NVPTXISD::Tex1DS32FloatGrad:
383 return "NVPTXISD::Tex1DS32FloatGrad";
384 case NVPTXISD::Tex1DU32S32: return "NVPTXISD::Tex1DU32S32";
385 case NVPTXISD::Tex1DU32Float: return "NVPTXISD::Tex1DU32Float";
386 case NVPTXISD::Tex1DU32FloatLevel:
387 return "NVPTXISD::Tex1DU32FloatLevel";
388 case NVPTXISD::Tex1DU32FloatGrad:
389 return "NVPTXISD::Tex1DU32FloatGrad";
390 case NVPTXISD::Tex1DArrayFloatS32: return "NVPTXISD::Tex1DArrayFloatS32";
391 case NVPTXISD::Tex1DArrayFloatFloat: return "NVPTXISD::Tex1DArrayFloatFloat";
392 case NVPTXISD::Tex1DArrayFloatFloatLevel:
393 return "NVPTXISD::Tex1DArrayFloatFloatLevel";
394 case NVPTXISD::Tex1DArrayFloatFloatGrad:
395 return "NVPTXISD::Tex1DArrayFloatFloatGrad";
396 case NVPTXISD::Tex1DArrayS32S32: return "NVPTXISD::Tex1DArrayS32S32";
397 case NVPTXISD::Tex1DArrayS32Float: return "NVPTXISD::Tex1DArrayS32Float";
398 case NVPTXISD::Tex1DArrayS32FloatLevel:
399 return "NVPTXISD::Tex1DArrayS32FloatLevel";
400 case NVPTXISD::Tex1DArrayS32FloatGrad:
401 return "NVPTXISD::Tex1DArrayS32FloatGrad";
402 case NVPTXISD::Tex1DArrayU32S32: return "NVPTXISD::Tex1DArrayU32S32";
403 case NVPTXISD::Tex1DArrayU32Float: return "NVPTXISD::Tex1DArrayU32Float";
404 case NVPTXISD::Tex1DArrayU32FloatLevel:
405 return "NVPTXISD::Tex1DArrayU32FloatLevel";
406 case NVPTXISD::Tex1DArrayU32FloatGrad:
407 return "NVPTXISD::Tex1DArrayU32FloatGrad";
408 case NVPTXISD::Tex2DFloatS32: return "NVPTXISD::Tex2DFloatS32";
409 case NVPTXISD::Tex2DFloatFloat: return "NVPTXISD::Tex2DFloatFloat";
410 case NVPTXISD::Tex2DFloatFloatLevel:
411 return "NVPTXISD::Tex2DFloatFloatLevel";
412 case NVPTXISD::Tex2DFloatFloatGrad:
413 return "NVPTXISD::Tex2DFloatFloatGrad";
414 case NVPTXISD::Tex2DS32S32: return "NVPTXISD::Tex2DS32S32";
415 case NVPTXISD::Tex2DS32Float: return "NVPTXISD::Tex2DS32Float";
416 case NVPTXISD::Tex2DS32FloatLevel:
417 return "NVPTXISD::Tex2DS32FloatLevel";
418 case NVPTXISD::Tex2DS32FloatGrad:
419 return "NVPTXISD::Tex2DS32FloatGrad";
420 case NVPTXISD::Tex2DU32S32: return "NVPTXISD::Tex2DU32S32";
421 case NVPTXISD::Tex2DU32Float: return "NVPTXISD::Tex2DU32Float";
422 case NVPTXISD::Tex2DU32FloatLevel:
423 return "NVPTXISD::Tex2DU32FloatLevel";
424 case NVPTXISD::Tex2DU32FloatGrad:
425 return "NVPTXISD::Tex2DU32FloatGrad";
426 case NVPTXISD::Tex2DArrayFloatS32: return "NVPTXISD::Tex2DArrayFloatS32";
427 case NVPTXISD::Tex2DArrayFloatFloat: return "NVPTXISD::Tex2DArrayFloatFloat";
428 case NVPTXISD::Tex2DArrayFloatFloatLevel:
429 return "NVPTXISD::Tex2DArrayFloatFloatLevel";
430 case NVPTXISD::Tex2DArrayFloatFloatGrad:
431 return "NVPTXISD::Tex2DArrayFloatFloatGrad";
432 case NVPTXISD::Tex2DArrayS32S32: return "NVPTXISD::Tex2DArrayS32S32";
433 case NVPTXISD::Tex2DArrayS32Float: return "NVPTXISD::Tex2DArrayS32Float";
434 case NVPTXISD::Tex2DArrayS32FloatLevel:
435 return "NVPTXISD::Tex2DArrayS32FloatLevel";
436 case NVPTXISD::Tex2DArrayS32FloatGrad:
437 return "NVPTXISD::Tex2DArrayS32FloatGrad";
438 case NVPTXISD::Tex2DArrayU32S32: return "NVPTXISD::Tex2DArrayU32S32";
439 case NVPTXISD::Tex2DArrayU32Float: return "NVPTXISD::Tex2DArrayU32Float";
440 case NVPTXISD::Tex2DArrayU32FloatLevel:
441 return "NVPTXISD::Tex2DArrayU32FloatLevel";
442 case NVPTXISD::Tex2DArrayU32FloatGrad:
443 return "NVPTXISD::Tex2DArrayU32FloatGrad";
444 case NVPTXISD::Tex3DFloatS32: return "NVPTXISD::Tex3DFloatS32";
445 case NVPTXISD::Tex3DFloatFloat: return "NVPTXISD::Tex3DFloatFloat";
446 case NVPTXISD::Tex3DFloatFloatLevel:
447 return "NVPTXISD::Tex3DFloatFloatLevel";
448 case NVPTXISD::Tex3DFloatFloatGrad:
449 return "NVPTXISD::Tex3DFloatFloatGrad";
450 case NVPTXISD::Tex3DS32S32: return "NVPTXISD::Tex3DS32S32";
451 case NVPTXISD::Tex3DS32Float: return "NVPTXISD::Tex3DS32Float";
452 case NVPTXISD::Tex3DS32FloatLevel:
453 return "NVPTXISD::Tex3DS32FloatLevel";
454 case NVPTXISD::Tex3DS32FloatGrad:
455 return "NVPTXISD::Tex3DS32FloatGrad";
456 case NVPTXISD::Tex3DU32S32: return "NVPTXISD::Tex3DU32S32";
457 case NVPTXISD::Tex3DU32Float: return "NVPTXISD::Tex3DU32Float";
458 case NVPTXISD::Tex3DU32FloatLevel:
459 return "NVPTXISD::Tex3DU32FloatLevel";
460 case NVPTXISD::Tex3DU32FloatGrad:
461 return "NVPTXISD::Tex3DU32FloatGrad";
462 case NVPTXISD::TexCubeFloatFloat: return "NVPTXISD::TexCubeFloatFloat";
463 case NVPTXISD::TexCubeFloatFloatLevel:
464 return "NVPTXISD::TexCubeFloatFloatLevel";
465 case NVPTXISD::TexCubeS32Float: return "NVPTXISD::TexCubeS32Float";
466 case NVPTXISD::TexCubeS32FloatLevel:
467 return "NVPTXISD::TexCubeS32FloatLevel";
468 case NVPTXISD::TexCubeU32Float: return "NVPTXISD::TexCubeU32Float";
469 case NVPTXISD::TexCubeU32FloatLevel:
470 return "NVPTXISD::TexCubeU32FloatLevel";
471 case NVPTXISD::TexCubeArrayFloatFloat:
472 return "NVPTXISD::TexCubeArrayFloatFloat";
473 case NVPTXISD::TexCubeArrayFloatFloatLevel:
474 return "NVPTXISD::TexCubeArrayFloatFloatLevel";
475 case NVPTXISD::TexCubeArrayS32Float:
476 return "NVPTXISD::TexCubeArrayS32Float";
477 case NVPTXISD::TexCubeArrayS32FloatLevel:
478 return "NVPTXISD::TexCubeArrayS32FloatLevel";
479 case NVPTXISD::TexCubeArrayU32Float:
480 return "NVPTXISD::TexCubeArrayU32Float";
481 case NVPTXISD::TexCubeArrayU32FloatLevel:
482 return "NVPTXISD::TexCubeArrayU32FloatLevel";
483 case NVPTXISD::Tld4R2DFloatFloat:
484 return "NVPTXISD::Tld4R2DFloatFloat";
485 case NVPTXISD::Tld4G2DFloatFloat:
486 return "NVPTXISD::Tld4G2DFloatFloat";
487 case NVPTXISD::Tld4B2DFloatFloat:
488 return "NVPTXISD::Tld4B2DFloatFloat";
489 case NVPTXISD::Tld4A2DFloatFloat:
490 return "NVPTXISD::Tld4A2DFloatFloat";
491 case NVPTXISD::Tld4R2DS64Float:
492 return "NVPTXISD::Tld4R2DS64Float";
493 case NVPTXISD::Tld4G2DS64Float:
494 return "NVPTXISD::Tld4G2DS64Float";
495 case NVPTXISD::Tld4B2DS64Float:
496 return "NVPTXISD::Tld4B2DS64Float";
497 case NVPTXISD::Tld4A2DS64Float:
498 return "NVPTXISD::Tld4A2DS64Float";
499 case NVPTXISD::Tld4R2DU64Float:
500 return "NVPTXISD::Tld4R2DU64Float";
501 case NVPTXISD::Tld4G2DU64Float:
502 return "NVPTXISD::Tld4G2DU64Float";
503 case NVPTXISD::Tld4B2DU64Float:
504 return "NVPTXISD::Tld4B2DU64Float";
505 case NVPTXISD::Tld4A2DU64Float:
506 return "NVPTXISD::Tld4A2DU64Float";
508 case NVPTXISD::TexUnified1DFloatS32:
509 return "NVPTXISD::TexUnified1DFloatS32";
510 case NVPTXISD::TexUnified1DFloatFloat:
511 return "NVPTXISD::TexUnified1DFloatFloat";
512 case NVPTXISD::TexUnified1DFloatFloatLevel:
513 return "NVPTXISD::TexUnified1DFloatFloatLevel";
514 case NVPTXISD::TexUnified1DFloatFloatGrad:
515 return "NVPTXISD::TexUnified1DFloatFloatGrad";
516 case NVPTXISD::TexUnified1DS32S32:
517 return "NVPTXISD::TexUnified1DS32S32";
518 case NVPTXISD::TexUnified1DS32Float:
519 return "NVPTXISD::TexUnified1DS32Float";
520 case NVPTXISD::TexUnified1DS32FloatLevel:
521 return "NVPTXISD::TexUnified1DS32FloatLevel";
522 case NVPTXISD::TexUnified1DS32FloatGrad:
523 return "NVPTXISD::TexUnified1DS32FloatGrad";
524 case NVPTXISD::TexUnified1DU32S32:
525 return "NVPTXISD::TexUnified1DU32S32";
526 case NVPTXISD::TexUnified1DU32Float:
527 return "NVPTXISD::TexUnified1DU32Float";
528 case NVPTXISD::TexUnified1DU32FloatLevel:
529 return "NVPTXISD::TexUnified1DU32FloatLevel";
530 case NVPTXISD::TexUnified1DU32FloatGrad:
531 return "NVPTXISD::TexUnified1DU32FloatGrad";
532 case NVPTXISD::TexUnified1DArrayFloatS32:
533 return "NVPTXISD::TexUnified1DArrayFloatS32";
534 case NVPTXISD::TexUnified1DArrayFloatFloat:
535 return "NVPTXISD::TexUnified1DArrayFloatFloat";
536 case NVPTXISD::TexUnified1DArrayFloatFloatLevel:
537 return "NVPTXISD::TexUnified1DArrayFloatFloatLevel";
538 case NVPTXISD::TexUnified1DArrayFloatFloatGrad:
539 return "NVPTXISD::TexUnified1DArrayFloatFloatGrad";
540 case NVPTXISD::TexUnified1DArrayS32S32:
541 return "NVPTXISD::TexUnified1DArrayS32S32";
542 case NVPTXISD::TexUnified1DArrayS32Float:
543 return "NVPTXISD::TexUnified1DArrayS32Float";
544 case NVPTXISD::TexUnified1DArrayS32FloatLevel:
545 return "NVPTXISD::TexUnified1DArrayS32FloatLevel";
546 case NVPTXISD::TexUnified1DArrayS32FloatGrad:
547 return "NVPTXISD::TexUnified1DArrayS32FloatGrad";
548 case NVPTXISD::TexUnified1DArrayU32S32:
549 return "NVPTXISD::TexUnified1DArrayU32S32";
550 case NVPTXISD::TexUnified1DArrayU32Float:
551 return "NVPTXISD::TexUnified1DArrayU32Float";
552 case NVPTXISD::TexUnified1DArrayU32FloatLevel:
553 return "NVPTXISD::TexUnified1DArrayU32FloatLevel";
554 case NVPTXISD::TexUnified1DArrayU32FloatGrad:
555 return "NVPTXISD::TexUnified1DArrayU32FloatGrad";
556 case NVPTXISD::TexUnified2DFloatS32:
557 return "NVPTXISD::TexUnified2DFloatS32";
558 case NVPTXISD::TexUnified2DFloatFloat:
559 return "NVPTXISD::TexUnified2DFloatFloat";
560 case NVPTXISD::TexUnified2DFloatFloatLevel:
561 return "NVPTXISD::TexUnified2DFloatFloatLevel";
562 case NVPTXISD::TexUnified2DFloatFloatGrad:
563 return "NVPTXISD::TexUnified2DFloatFloatGrad";
564 case NVPTXISD::TexUnified2DS32S32:
565 return "NVPTXISD::TexUnified2DS32S32";
566 case NVPTXISD::TexUnified2DS32Float:
567 return "NVPTXISD::TexUnified2DS32Float";
568 case NVPTXISD::TexUnified2DS32FloatLevel:
569 return "NVPTXISD::TexUnified2DS32FloatLevel";
570 case NVPTXISD::TexUnified2DS32FloatGrad:
571 return "NVPTXISD::TexUnified2DS32FloatGrad";
572 case NVPTXISD::TexUnified2DU32S32:
573 return "NVPTXISD::TexUnified2DU32S32";
574 case NVPTXISD::TexUnified2DU32Float:
575 return "NVPTXISD::TexUnified2DU32Float";
576 case NVPTXISD::TexUnified2DU32FloatLevel:
577 return "NVPTXISD::TexUnified2DU32FloatLevel";
578 case NVPTXISD::TexUnified2DU32FloatGrad:
579 return "NVPTXISD::TexUnified2DU32FloatGrad";
580 case NVPTXISD::TexUnified2DArrayFloatS32:
581 return "NVPTXISD::TexUnified2DArrayFloatS32";
582 case NVPTXISD::TexUnified2DArrayFloatFloat:
583 return "NVPTXISD::TexUnified2DArrayFloatFloat";
584 case NVPTXISD::TexUnified2DArrayFloatFloatLevel:
585 return "NVPTXISD::TexUnified2DArrayFloatFloatLevel";
586 case NVPTXISD::TexUnified2DArrayFloatFloatGrad:
587 return "NVPTXISD::TexUnified2DArrayFloatFloatGrad";
588 case NVPTXISD::TexUnified2DArrayS32S32:
589 return "NVPTXISD::TexUnified2DArrayS32S32";
590 case NVPTXISD::TexUnified2DArrayS32Float:
591 return "NVPTXISD::TexUnified2DArrayS32Float";
592 case NVPTXISD::TexUnified2DArrayS32FloatLevel:
593 return "NVPTXISD::TexUnified2DArrayS32FloatLevel";
594 case NVPTXISD::TexUnified2DArrayS32FloatGrad:
595 return "NVPTXISD::TexUnified2DArrayS32FloatGrad";
596 case NVPTXISD::TexUnified2DArrayU32S32:
597 return "NVPTXISD::TexUnified2DArrayU32S32";
598 case NVPTXISD::TexUnified2DArrayU32Float:
599 return "NVPTXISD::TexUnified2DArrayU32Float";
600 case NVPTXISD::TexUnified2DArrayU32FloatLevel:
601 return "NVPTXISD::TexUnified2DArrayU32FloatLevel";
602 case NVPTXISD::TexUnified2DArrayU32FloatGrad:
603 return "NVPTXISD::TexUnified2DArrayU32FloatGrad";
604 case NVPTXISD::TexUnified3DFloatS32:
605 return "NVPTXISD::TexUnified3DFloatS32";
606 case NVPTXISD::TexUnified3DFloatFloat:
607 return "NVPTXISD::TexUnified3DFloatFloat";
608 case NVPTXISD::TexUnified3DFloatFloatLevel:
609 return "NVPTXISD::TexUnified3DFloatFloatLevel";
610 case NVPTXISD::TexUnified3DFloatFloatGrad:
611 return "NVPTXISD::TexUnified3DFloatFloatGrad";
612 case NVPTXISD::TexUnified3DS32S32:
613 return "NVPTXISD::TexUnified3DS32S32";
614 case NVPTXISD::TexUnified3DS32Float:
615 return "NVPTXISD::TexUnified3DS32Float";
616 case NVPTXISD::TexUnified3DS32FloatLevel:
617 return "NVPTXISD::TexUnified3DS32FloatLevel";
618 case NVPTXISD::TexUnified3DS32FloatGrad:
619 return "NVPTXISD::TexUnified3DS32FloatGrad";
620 case NVPTXISD::TexUnified3DU32S32:
621 return "NVPTXISD::TexUnified3DU32S32";
622 case NVPTXISD::TexUnified3DU32Float:
623 return "NVPTXISD::TexUnified3DU32Float";
624 case NVPTXISD::TexUnified3DU32FloatLevel:
625 return "NVPTXISD::TexUnified3DU32FloatLevel";
626 case NVPTXISD::TexUnified3DU32FloatGrad:
627 return "NVPTXISD::TexUnified3DU32FloatGrad";
628 case NVPTXISD::TexUnifiedCubeFloatFloat:
629 return "NVPTXISD::TexUnifiedCubeFloatFloat";
630 case NVPTXISD::TexUnifiedCubeFloatFloatLevel:
631 return "NVPTXISD::TexUnifiedCubeFloatFloatLevel";
632 case NVPTXISD::TexUnifiedCubeS32Float:
633 return "NVPTXISD::TexUnifiedCubeS32Float";
634 case NVPTXISD::TexUnifiedCubeS32FloatLevel:
635 return "NVPTXISD::TexUnifiedCubeS32FloatLevel";
636 case NVPTXISD::TexUnifiedCubeU32Float:
637 return "NVPTXISD::TexUnifiedCubeU32Float";
638 case NVPTXISD::TexUnifiedCubeU32FloatLevel:
639 return "NVPTXISD::TexUnifiedCubeU32FloatLevel";
640 case NVPTXISD::TexUnifiedCubeArrayFloatFloat:
641 return "NVPTXISD::TexUnifiedCubeArrayFloatFloat";
642 case NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel:
643 return "NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel";
644 case NVPTXISD::TexUnifiedCubeArrayS32Float:
645 return "NVPTXISD::TexUnifiedCubeArrayS32Float";
646 case NVPTXISD::TexUnifiedCubeArrayS32FloatLevel:
647 return "NVPTXISD::TexUnifiedCubeArrayS32FloatLevel";
648 case NVPTXISD::TexUnifiedCubeArrayU32Float:
649 return "NVPTXISD::TexUnifiedCubeArrayU32Float";
650 case NVPTXISD::TexUnifiedCubeArrayU32FloatLevel:
651 return "NVPTXISD::TexUnifiedCubeArrayU32FloatLevel";
652 case NVPTXISD::Tld4UnifiedR2DFloatFloat:
653 return "NVPTXISD::Tld4UnifiedR2DFloatFloat";
654 case NVPTXISD::Tld4UnifiedG2DFloatFloat:
655 return "NVPTXISD::Tld4UnifiedG2DFloatFloat";
656 case NVPTXISD::Tld4UnifiedB2DFloatFloat:
657 return "NVPTXISD::Tld4UnifiedB2DFloatFloat";
658 case NVPTXISD::Tld4UnifiedA2DFloatFloat:
659 return "NVPTXISD::Tld4UnifiedA2DFloatFloat";
660 case NVPTXISD::Tld4UnifiedR2DS64Float:
661 return "NVPTXISD::Tld4UnifiedR2DS64Float";
662 case NVPTXISD::Tld4UnifiedG2DS64Float:
663 return "NVPTXISD::Tld4UnifiedG2DS64Float";
664 case NVPTXISD::Tld4UnifiedB2DS64Float:
665 return "NVPTXISD::Tld4UnifiedB2DS64Float";
666 case NVPTXISD::Tld4UnifiedA2DS64Float:
667 return "NVPTXISD::Tld4UnifiedA2DS64Float";
668 case NVPTXISD::Tld4UnifiedR2DU64Float:
669 return "NVPTXISD::Tld4UnifiedR2DU64Float";
670 case NVPTXISD::Tld4UnifiedG2DU64Float:
671 return "NVPTXISD::Tld4UnifiedG2DU64Float";
672 case NVPTXISD::Tld4UnifiedB2DU64Float:
673 return "NVPTXISD::Tld4UnifiedB2DU64Float";
674 case NVPTXISD::Tld4UnifiedA2DU64Float:
675 return "NVPTXISD::Tld4UnifiedA2DU64Float";
677 case NVPTXISD::Suld1DI8Clamp: return "NVPTXISD::Suld1DI8Clamp";
678 case NVPTXISD::Suld1DI16Clamp: return "NVPTXISD::Suld1DI16Clamp";
679 case NVPTXISD::Suld1DI32Clamp: return "NVPTXISD::Suld1DI32Clamp";
680 case NVPTXISD::Suld1DI64Clamp: return "NVPTXISD::Suld1DI64Clamp";
681 case NVPTXISD::Suld1DV2I8Clamp: return "NVPTXISD::Suld1DV2I8Clamp";
682 case NVPTXISD::Suld1DV2I16Clamp: return "NVPTXISD::Suld1DV2I16Clamp";
683 case NVPTXISD::Suld1DV2I32Clamp: return "NVPTXISD::Suld1DV2I32Clamp";
684 case NVPTXISD::Suld1DV2I64Clamp: return "NVPTXISD::Suld1DV2I64Clamp";
685 case NVPTXISD::Suld1DV4I8Clamp: return "NVPTXISD::Suld1DV4I8Clamp";
686 case NVPTXISD::Suld1DV4I16Clamp: return "NVPTXISD::Suld1DV4I16Clamp";
687 case NVPTXISD::Suld1DV4I32Clamp: return "NVPTXISD::Suld1DV4I32Clamp";
689 case NVPTXISD::Suld1DArrayI8Clamp: return "NVPTXISD::Suld1DArrayI8Clamp";
690 case NVPTXISD::Suld1DArrayI16Clamp: return "NVPTXISD::Suld1DArrayI16Clamp";
691 case NVPTXISD::Suld1DArrayI32Clamp: return "NVPTXISD::Suld1DArrayI32Clamp";
692 case NVPTXISD::Suld1DArrayI64Clamp: return "NVPTXISD::Suld1DArrayI64Clamp";
693 case NVPTXISD::Suld1DArrayV2I8Clamp: return "NVPTXISD::Suld1DArrayV2I8Clamp";
694 case NVPTXISD::Suld1DArrayV2I16Clamp:return "NVPTXISD::Suld1DArrayV2I16Clamp";
695 case NVPTXISD::Suld1DArrayV2I32Clamp:return "NVPTXISD::Suld1DArrayV2I32Clamp";
696 case NVPTXISD::Suld1DArrayV2I64Clamp:return "NVPTXISD::Suld1DArrayV2I64Clamp";
697 case NVPTXISD::Suld1DArrayV4I8Clamp: return "NVPTXISD::Suld1DArrayV4I8Clamp";
698 case NVPTXISD::Suld1DArrayV4I16Clamp:return "NVPTXISD::Suld1DArrayV4I16Clamp";
699 case NVPTXISD::Suld1DArrayV4I32Clamp:return "NVPTXISD::Suld1DArrayV4I32Clamp";
701 case NVPTXISD::Suld2DI8Clamp: return "NVPTXISD::Suld2DI8Clamp";
702 case NVPTXISD::Suld2DI16Clamp: return "NVPTXISD::Suld2DI16Clamp";
703 case NVPTXISD::Suld2DI32Clamp: return "NVPTXISD::Suld2DI32Clamp";
704 case NVPTXISD::Suld2DI64Clamp: return "NVPTXISD::Suld2DI64Clamp";
705 case NVPTXISD::Suld2DV2I8Clamp: return "NVPTXISD::Suld2DV2I8Clamp";
706 case NVPTXISD::Suld2DV2I16Clamp: return "NVPTXISD::Suld2DV2I16Clamp";
707 case NVPTXISD::Suld2DV2I32Clamp: return "NVPTXISD::Suld2DV2I32Clamp";
708 case NVPTXISD::Suld2DV2I64Clamp: return "NVPTXISD::Suld2DV2I64Clamp";
709 case NVPTXISD::Suld2DV4I8Clamp: return "NVPTXISD::Suld2DV4I8Clamp";
710 case NVPTXISD::Suld2DV4I16Clamp: return "NVPTXISD::Suld2DV4I16Clamp";
711 case NVPTXISD::Suld2DV4I32Clamp: return "NVPTXISD::Suld2DV4I32Clamp";
713 case NVPTXISD::Suld2DArrayI8Clamp: return "NVPTXISD::Suld2DArrayI8Clamp";
714 case NVPTXISD::Suld2DArrayI16Clamp: return "NVPTXISD::Suld2DArrayI16Clamp";
715 case NVPTXISD::Suld2DArrayI32Clamp: return "NVPTXISD::Suld2DArrayI32Clamp";
716 case NVPTXISD::Suld2DArrayI64Clamp: return "NVPTXISD::Suld2DArrayI64Clamp";
717 case NVPTXISD::Suld2DArrayV2I8Clamp: return "NVPTXISD::Suld2DArrayV2I8Clamp";
718 case NVPTXISD::Suld2DArrayV2I16Clamp:return "NVPTXISD::Suld2DArrayV2I16Clamp";
719 case NVPTXISD::Suld2DArrayV2I32Clamp:return "NVPTXISD::Suld2DArrayV2I32Clamp";
720 case NVPTXISD::Suld2DArrayV2I64Clamp:return "NVPTXISD::Suld2DArrayV2I64Clamp";
721 case NVPTXISD::Suld2DArrayV4I8Clamp: return "NVPTXISD::Suld2DArrayV4I8Clamp";
722 case NVPTXISD::Suld2DArrayV4I16Clamp:return "NVPTXISD::Suld2DArrayV4I16Clamp";
723 case NVPTXISD::Suld2DArrayV4I32Clamp:return "NVPTXISD::Suld2DArrayV4I32Clamp";
725 case NVPTXISD::Suld3DI8Clamp: return "NVPTXISD::Suld3DI8Clamp";
726 case NVPTXISD::Suld3DI16Clamp: return "NVPTXISD::Suld3DI16Clamp";
727 case NVPTXISD::Suld3DI32Clamp: return "NVPTXISD::Suld3DI32Clamp";
728 case NVPTXISD::Suld3DI64Clamp: return "NVPTXISD::Suld3DI64Clamp";
729 case NVPTXISD::Suld3DV2I8Clamp: return "NVPTXISD::Suld3DV2I8Clamp";
730 case NVPTXISD::Suld3DV2I16Clamp: return "NVPTXISD::Suld3DV2I16Clamp";
731 case NVPTXISD::Suld3DV2I32Clamp: return "NVPTXISD::Suld3DV2I32Clamp";
732 case NVPTXISD::Suld3DV2I64Clamp: return "NVPTXISD::Suld3DV2I64Clamp";
733 case NVPTXISD::Suld3DV4I8Clamp: return "NVPTXISD::Suld3DV4I8Clamp";
734 case NVPTXISD::Suld3DV4I16Clamp: return "NVPTXISD::Suld3DV4I16Clamp";
735 case NVPTXISD::Suld3DV4I32Clamp: return "NVPTXISD::Suld3DV4I32Clamp";
737 case NVPTXISD::Suld1DI8Trap: return "NVPTXISD::Suld1DI8Trap";
738 case NVPTXISD::Suld1DI16Trap: return "NVPTXISD::Suld1DI16Trap";
739 case NVPTXISD::Suld1DI32Trap: return "NVPTXISD::Suld1DI32Trap";
740 case NVPTXISD::Suld1DI64Trap: return "NVPTXISD::Suld1DI64Trap";
741 case NVPTXISD::Suld1DV2I8Trap: return "NVPTXISD::Suld1DV2I8Trap";
742 case NVPTXISD::Suld1DV2I16Trap: return "NVPTXISD::Suld1DV2I16Trap";
743 case NVPTXISD::Suld1DV2I32Trap: return "NVPTXISD::Suld1DV2I32Trap";
744 case NVPTXISD::Suld1DV2I64Trap: return "NVPTXISD::Suld1DV2I64Trap";
745 case NVPTXISD::Suld1DV4I8Trap: return "NVPTXISD::Suld1DV4I8Trap";
746 case NVPTXISD::Suld1DV4I16Trap: return "NVPTXISD::Suld1DV4I16Trap";
747 case NVPTXISD::Suld1DV4I32Trap: return "NVPTXISD::Suld1DV4I32Trap";
749 case NVPTXISD::Suld1DArrayI8Trap: return "NVPTXISD::Suld1DArrayI8Trap";
750 case NVPTXISD::Suld1DArrayI16Trap: return "NVPTXISD::Suld1DArrayI16Trap";
751 case NVPTXISD::Suld1DArrayI32Trap: return "NVPTXISD::Suld1DArrayI32Trap";
752 case NVPTXISD::Suld1DArrayI64Trap: return "NVPTXISD::Suld1DArrayI64Trap";
753 case NVPTXISD::Suld1DArrayV2I8Trap: return "NVPTXISD::Suld1DArrayV2I8Trap";
754 case NVPTXISD::Suld1DArrayV2I16Trap: return "NVPTXISD::Suld1DArrayV2I16Trap";
755 case NVPTXISD::Suld1DArrayV2I32Trap: return "NVPTXISD::Suld1DArrayV2I32Trap";
756 case NVPTXISD::Suld1DArrayV2I64Trap: return "NVPTXISD::Suld1DArrayV2I64Trap";
757 case NVPTXISD::Suld1DArrayV4I8Trap: return "NVPTXISD::Suld1DArrayV4I8Trap";
758 case NVPTXISD::Suld1DArrayV4I16Trap: return "NVPTXISD::Suld1DArrayV4I16Trap";
759 case NVPTXISD::Suld1DArrayV4I32Trap: return "NVPTXISD::Suld1DArrayV4I32Trap";
761 case NVPTXISD::Suld2DI8Trap: return "NVPTXISD::Suld2DI8Trap";
762 case NVPTXISD::Suld2DI16Trap: return "NVPTXISD::Suld2DI16Trap";
763 case NVPTXISD::Suld2DI32Trap: return "NVPTXISD::Suld2DI32Trap";
764 case NVPTXISD::Suld2DI64Trap: return "NVPTXISD::Suld2DI64Trap";
765 case NVPTXISD::Suld2DV2I8Trap: return "NVPTXISD::Suld2DV2I8Trap";
766 case NVPTXISD::Suld2DV2I16Trap: return "NVPTXISD::Suld2DV2I16Trap";
767 case NVPTXISD::Suld2DV2I32Trap: return "NVPTXISD::Suld2DV2I32Trap";
768 case NVPTXISD::Suld2DV2I64Trap: return "NVPTXISD::Suld2DV2I64Trap";
769 case NVPTXISD::Suld2DV4I8Trap: return "NVPTXISD::Suld2DV4I8Trap";
770 case NVPTXISD::Suld2DV4I16Trap: return "NVPTXISD::Suld2DV4I16Trap";
771 case NVPTXISD::Suld2DV4I32Trap: return "NVPTXISD::Suld2DV4I32Trap";
773 case NVPTXISD::Suld2DArrayI8Trap: return "NVPTXISD::Suld2DArrayI8Trap";
774 case NVPTXISD::Suld2DArrayI16Trap: return "NVPTXISD::Suld2DArrayI16Trap";
775 case NVPTXISD::Suld2DArrayI32Trap: return "NVPTXISD::Suld2DArrayI32Trap";
776 case NVPTXISD::Suld2DArrayI64Trap: return "NVPTXISD::Suld2DArrayI64Trap";
777 case NVPTXISD::Suld2DArrayV2I8Trap: return "NVPTXISD::Suld2DArrayV2I8Trap";
778 case NVPTXISD::Suld2DArrayV2I16Trap: return "NVPTXISD::Suld2DArrayV2I16Trap";
779 case NVPTXISD::Suld2DArrayV2I32Trap: return "NVPTXISD::Suld2DArrayV2I32Trap";
780 case NVPTXISD::Suld2DArrayV2I64Trap: return "NVPTXISD::Suld2DArrayV2I64Trap";
781 case NVPTXISD::Suld2DArrayV4I8Trap: return "NVPTXISD::Suld2DArrayV4I8Trap";
782 case NVPTXISD::Suld2DArrayV4I16Trap: return "NVPTXISD::Suld2DArrayV4I16Trap";
783 case NVPTXISD::Suld2DArrayV4I32Trap: return "NVPTXISD::Suld2DArrayV4I32Trap";
785 case NVPTXISD::Suld3DI8Trap: return "NVPTXISD::Suld3DI8Trap";
786 case NVPTXISD::Suld3DI16Trap: return "NVPTXISD::Suld3DI16Trap";
787 case NVPTXISD::Suld3DI32Trap: return "NVPTXISD::Suld3DI32Trap";
788 case NVPTXISD::Suld3DI64Trap: return "NVPTXISD::Suld3DI64Trap";
789 case NVPTXISD::Suld3DV2I8Trap: return "NVPTXISD::Suld3DV2I8Trap";
790 case NVPTXISD::Suld3DV2I16Trap: return "NVPTXISD::Suld3DV2I16Trap";
791 case NVPTXISD::Suld3DV2I32Trap: return "NVPTXISD::Suld3DV2I32Trap";
792 case NVPTXISD::Suld3DV2I64Trap: return "NVPTXISD::Suld3DV2I64Trap";
793 case NVPTXISD::Suld3DV4I8Trap: return "NVPTXISD::Suld3DV4I8Trap";
794 case NVPTXISD::Suld3DV4I16Trap: return "NVPTXISD::Suld3DV4I16Trap";
795 case NVPTXISD::Suld3DV4I32Trap: return "NVPTXISD::Suld3DV4I32Trap";
797 case NVPTXISD::Suld1DI8Zero: return "NVPTXISD::Suld1DI8Zero";
798 case NVPTXISD::Suld1DI16Zero: return "NVPTXISD::Suld1DI16Zero";
799 case NVPTXISD::Suld1DI32Zero: return "NVPTXISD::Suld1DI32Zero";
800 case NVPTXISD::Suld1DI64Zero: return "NVPTXISD::Suld1DI64Zero";
801 case NVPTXISD::Suld1DV2I8Zero: return "NVPTXISD::Suld1DV2I8Zero";
802 case NVPTXISD::Suld1DV2I16Zero: return "NVPTXISD::Suld1DV2I16Zero";
803 case NVPTXISD::Suld1DV2I32Zero: return "NVPTXISD::Suld1DV2I32Zero";
804 case NVPTXISD::Suld1DV2I64Zero: return "NVPTXISD::Suld1DV2I64Zero";
805 case NVPTXISD::Suld1DV4I8Zero: return "NVPTXISD::Suld1DV4I8Zero";
806 case NVPTXISD::Suld1DV4I16Zero: return "NVPTXISD::Suld1DV4I16Zero";
807 case NVPTXISD::Suld1DV4I32Zero: return "NVPTXISD::Suld1DV4I32Zero";
809 case NVPTXISD::Suld1DArrayI8Zero: return "NVPTXISD::Suld1DArrayI8Zero";
810 case NVPTXISD::Suld1DArrayI16Zero: return "NVPTXISD::Suld1DArrayI16Zero";
811 case NVPTXISD::Suld1DArrayI32Zero: return "NVPTXISD::Suld1DArrayI32Zero";
812 case NVPTXISD::Suld1DArrayI64Zero: return "NVPTXISD::Suld1DArrayI64Zero";
813 case NVPTXISD::Suld1DArrayV2I8Zero: return "NVPTXISD::Suld1DArrayV2I8Zero";
814 case NVPTXISD::Suld1DArrayV2I16Zero: return "NVPTXISD::Suld1DArrayV2I16Zero";
815 case NVPTXISD::Suld1DArrayV2I32Zero: return "NVPTXISD::Suld1DArrayV2I32Zero";
816 case NVPTXISD::Suld1DArrayV2I64Zero: return "NVPTXISD::Suld1DArrayV2I64Zero";
817 case NVPTXISD::Suld1DArrayV4I8Zero: return "NVPTXISD::Suld1DArrayV4I8Zero";
818 case NVPTXISD::Suld1DArrayV4I16Zero: return "NVPTXISD::Suld1DArrayV4I16Zero";
819 case NVPTXISD::Suld1DArrayV4I32Zero: return "NVPTXISD::Suld1DArrayV4I32Zero";
821 case NVPTXISD::Suld2DI8Zero: return "NVPTXISD::Suld2DI8Zero";
822 case NVPTXISD::Suld2DI16Zero: return "NVPTXISD::Suld2DI16Zero";
823 case NVPTXISD::Suld2DI32Zero: return "NVPTXISD::Suld2DI32Zero";
824 case NVPTXISD::Suld2DI64Zero: return "NVPTXISD::Suld2DI64Zero";
825 case NVPTXISD::Suld2DV2I8Zero: return "NVPTXISD::Suld2DV2I8Zero";
826 case NVPTXISD::Suld2DV2I16Zero: return "NVPTXISD::Suld2DV2I16Zero";
827 case NVPTXISD::Suld2DV2I32Zero: return "NVPTXISD::Suld2DV2I32Zero";
828 case NVPTXISD::Suld2DV2I64Zero: return "NVPTXISD::Suld2DV2I64Zero";
829 case NVPTXISD::Suld2DV4I8Zero: return "NVPTXISD::Suld2DV4I8Zero";
830 case NVPTXISD::Suld2DV4I16Zero: return "NVPTXISD::Suld2DV4I16Zero";
831 case NVPTXISD::Suld2DV4I32Zero: return "NVPTXISD::Suld2DV4I32Zero";
833 case NVPTXISD::Suld2DArrayI8Zero: return "NVPTXISD::Suld2DArrayI8Zero";
834 case NVPTXISD::Suld2DArrayI16Zero: return "NVPTXISD::Suld2DArrayI16Zero";
835 case NVPTXISD::Suld2DArrayI32Zero: return "NVPTXISD::Suld2DArrayI32Zero";
836 case NVPTXISD::Suld2DArrayI64Zero: return "NVPTXISD::Suld2DArrayI64Zero";
837 case NVPTXISD::Suld2DArrayV2I8Zero: return "NVPTXISD::Suld2DArrayV2I8Zero";
838 case NVPTXISD::Suld2DArrayV2I16Zero: return "NVPTXISD::Suld2DArrayV2I16Zero";
839 case NVPTXISD::Suld2DArrayV2I32Zero: return "NVPTXISD::Suld2DArrayV2I32Zero";
840 case NVPTXISD::Suld2DArrayV2I64Zero: return "NVPTXISD::Suld2DArrayV2I64Zero";
841 case NVPTXISD::Suld2DArrayV4I8Zero: return "NVPTXISD::Suld2DArrayV4I8Zero";
842 case NVPTXISD::Suld2DArrayV4I16Zero: return "NVPTXISD::Suld2DArrayV4I16Zero";
843 case NVPTXISD::Suld2DArrayV4I32Zero: return "NVPTXISD::Suld2DArrayV4I32Zero";
845 case NVPTXISD::Suld3DI8Zero: return "NVPTXISD::Suld3DI8Zero";
846 case NVPTXISD::Suld3DI16Zero: return "NVPTXISD::Suld3DI16Zero";
847 case NVPTXISD::Suld3DI32Zero: return "NVPTXISD::Suld3DI32Zero";
848 case NVPTXISD::Suld3DI64Zero: return "NVPTXISD::Suld3DI64Zero";
849 case NVPTXISD::Suld3DV2I8Zero: return "NVPTXISD::Suld3DV2I8Zero";
850 case NVPTXISD::Suld3DV2I16Zero: return "NVPTXISD::Suld3DV2I16Zero";
851 case NVPTXISD::Suld3DV2I32Zero: return "NVPTXISD::Suld3DV2I32Zero";
852 case NVPTXISD::Suld3DV2I64Zero: return "NVPTXISD::Suld3DV2I64Zero";
853 case NVPTXISD::Suld3DV4I8Zero: return "NVPTXISD::Suld3DV4I8Zero";
854 case NVPTXISD::Suld3DV4I16Zero: return "NVPTXISD::Suld3DV4I16Zero";
855 case NVPTXISD::Suld3DV4I32Zero: return "NVPTXISD::Suld3DV4I32Zero";
859 TargetLoweringBase::LegalizeTypeAction
860 NVPTXTargetLowering::getPreferredVectorAction(EVT VT) const {
861 if (VT.getVectorNumElements() != 1 && VT.getScalarType() == MVT::i1)
862 return TypeSplitVector;
864 return TargetLoweringBase::getPreferredVectorAction(VT);
868 NVPTXTargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
870 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
871 Op = DAG.getTargetGlobalAddress(GV, dl, getPointerTy());
872 return DAG.getNode(NVPTXISD::Wrapper, dl, getPointerTy(), Op);
876 NVPTXTargetLowering::getPrototype(Type *retTy, const ArgListTy &Args,
877 const SmallVectorImpl<ISD::OutputArg> &Outs,
878 unsigned retAlignment,
879 const ImmutableCallSite *CS) const {
881 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
882 assert(isABI && "Non-ABI compilation is not supported");
887 O << "prototype_" << uniqueCallSite << " : .callprototype ";
889 if (retTy->getTypeID() == Type::VoidTyID) {
893 if (retTy->isFloatingPointTy() || retTy->isIntegerTy()) {
895 if (const IntegerType *ITy = dyn_cast<IntegerType>(retTy)) {
896 size = ITy->getBitWidth();
900 assert(retTy->isFloatingPointTy() &&
901 "Floating point type expected here");
902 size = retTy->getPrimitiveSizeInBits();
905 O << ".param .b" << size << " _";
906 } else if (isa<PointerType>(retTy)) {
907 O << ".param .b" << getPointerTy().getSizeInBits() << " _";
908 } else if ((retTy->getTypeID() == Type::StructTyID) ||
909 isa<VectorType>(retTy)) {
910 O << ".param .align "
913 << getDataLayout()->getTypeAllocSize(retTy) << "]";
915 llvm_unreachable("Unknown return type");
922 MVT thePointerTy = getPointerTy();
925 for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
926 Type *Ty = Args[i].Ty;
932 if (Outs[OIdx].Flags.isByVal() == false) {
933 if (Ty->isAggregateType() || Ty->isVectorTy()) {
935 const CallInst *CallI = cast<CallInst>(CS->getInstruction());
936 const DataLayout *TD = getDataLayout();
937 // +1 because index 0 is reserved for return type alignment
938 if (!llvm::getAlign(*CallI, i + 1, align))
939 align = TD->getABITypeAlignment(Ty);
940 unsigned sz = TD->getTypeAllocSize(Ty);
941 O << ".param .align " << align << " .b8 ";
943 O << "[" << sz << "]";
944 // update the index for Outs
945 SmallVector<EVT, 16> vtparts;
946 ComputeValueVTs(*this, Ty, vtparts);
947 if (unsigned len = vtparts.size())
951 // i8 types in IR will be i16 types in SDAG
952 assert((getValueType(Ty) == Outs[OIdx].VT ||
953 (getValueType(Ty) == MVT::i8 && Outs[OIdx].VT == MVT::i16)) &&
954 "type mismatch between callee prototype and arguments");
957 if (isa<IntegerType>(Ty)) {
958 sz = cast<IntegerType>(Ty)->getBitWidth();
961 } else if (isa<PointerType>(Ty))
962 sz = thePointerTy.getSizeInBits();
964 sz = Ty->getPrimitiveSizeInBits();
965 O << ".param .b" << sz << " ";
969 const PointerType *PTy = dyn_cast<PointerType>(Ty);
970 assert(PTy && "Param with byval attribute should be a pointer type");
971 Type *ETy = PTy->getElementType();
973 unsigned align = Outs[OIdx].Flags.getByValAlign();
974 unsigned sz = getDataLayout()->getTypeAllocSize(ETy);
975 O << ".param .align " << align << " .b8 ";
977 O << "[" << sz << "]";
984 NVPTXTargetLowering::getArgumentAlignment(SDValue Callee,
985 const ImmutableCallSite *CS,
987 unsigned Idx) const {
988 const DataLayout *TD = getDataLayout();
990 const Value *DirectCallee = CS->getCalledFunction();
993 // We don't have a direct function symbol, but that may be because of
994 // constant cast instructions in the call.
995 const Instruction *CalleeI = CS->getInstruction();
996 assert(CalleeI && "Call target is not a function or derived value?");
998 // With bitcast'd call targets, the instruction will be the call
999 if (isa<CallInst>(CalleeI)) {
1000 // Check if we have call alignment metadata
1001 if (llvm::getAlign(*cast<CallInst>(CalleeI), Idx, Align))
1004 const Value *CalleeV = cast<CallInst>(CalleeI)->getCalledValue();
1005 // Ignore any bitcast instructions
1006 while(isa<ConstantExpr>(CalleeV)) {
1007 const ConstantExpr *CE = cast<ConstantExpr>(CalleeV);
1010 // Look through the bitcast
1011 CalleeV = cast<ConstantExpr>(CalleeV)->getOperand(0);
1014 // We have now looked past all of the bitcasts. Do we finally have a
1016 if (isa<Function>(CalleeV))
1017 DirectCallee = CalleeV;
1021 // Check for function alignment information if we found that the
1022 // ultimate target is a Function
1024 if (llvm::getAlign(*cast<Function>(DirectCallee), Idx, Align))
1027 // Call is indirect or alignment information is not available, fall back to
1028 // the ABI type alignment
1029 return TD->getABITypeAlignment(Ty);
1032 SDValue NVPTXTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
1033 SmallVectorImpl<SDValue> &InVals) const {
1034 SelectionDAG &DAG = CLI.DAG;
1036 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
1037 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
1038 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
1039 SDValue Chain = CLI.Chain;
1040 SDValue Callee = CLI.Callee;
1041 bool &isTailCall = CLI.IsTailCall;
1042 ArgListTy &Args = CLI.getArgs();
1043 Type *retTy = CLI.RetTy;
1044 ImmutableCallSite *CS = CLI.CS;
1046 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
1047 assert(isABI && "Non-ABI compilation is not supported");
1050 const DataLayout *TD = getDataLayout();
1051 MachineFunction &MF = DAG.getMachineFunction();
1052 const Function *F = MF.getFunction();
1054 SDValue tempChain = Chain;
1056 DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(uniqueCallSite, true),
1058 SDValue InFlag = Chain.getValue(1);
1060 unsigned paramCount = 0;
1061 // Args.size() and Outs.size() need not match.
1062 // Outs.size() will be larger
1063 // * if there is an aggregate argument with multiple fields (each field
1064 // showing up separately in Outs)
1065 // * if there is a vector argument with more than typical vector-length
1066 // elements (generally if more than 4) where each vector element is
1067 // individually present in Outs.
1068 // So a different index should be used for indexing into Outs/OutVals.
1069 // See similar issue in LowerFormalArguments.
1071 // Declare the .params or .reg need to pass values
1073 for (unsigned i = 0, e = Args.size(); i != e; ++i, ++OIdx) {
1074 EVT VT = Outs[OIdx].VT;
1075 Type *Ty = Args[i].Ty;
1077 if (Outs[OIdx].Flags.isByVal() == false) {
1078 if (Ty->isAggregateType()) {
1080 SmallVector<EVT, 16> vtparts;
1081 SmallVector<uint64_t, 16> Offsets;
1082 ComputePTXValueVTs(*this, Ty, vtparts, &Offsets, 0);
1084 unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
1085 // declare .param .align <align> .b8 .param<n>[<size>];
1086 unsigned sz = TD->getTypeAllocSize(Ty);
1087 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1088 SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, MVT::i32),
1089 DAG.getConstant(paramCount, MVT::i32),
1090 DAG.getConstant(sz, MVT::i32), InFlag };
1091 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1093 InFlag = Chain.getValue(1);
1094 for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
1095 EVT elemtype = vtparts[j];
1096 unsigned ArgAlign = GreatestCommonDivisor64(align, Offsets[j]);
1097 if (elemtype.isInteger() && (sz < 8))
1099 SDValue StVal = OutVals[OIdx];
1100 if (elemtype.getSizeInBits() < 16) {
1101 StVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, StVal);
1103 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1104 SDValue CopyParamOps[] = { Chain,
1105 DAG.getConstant(paramCount, MVT::i32),
1106 DAG.getConstant(Offsets[j], MVT::i32),
1108 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
1109 CopyParamVTs, CopyParamOps,
1110 elemtype, MachinePointerInfo(),
1112 InFlag = Chain.getValue(1);
1115 if (vtparts.size() > 0)
1120 if (Ty->isVectorTy()) {
1121 EVT ObjectVT = getValueType(Ty);
1122 unsigned align = getArgumentAlignment(Callee, CS, Ty, paramCount + 1);
1123 // declare .param .align <align> .b8 .param<n>[<size>];
1124 unsigned sz = TD->getTypeAllocSize(Ty);
1125 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1126 SDValue DeclareParamOps[] = { Chain, DAG.getConstant(align, MVT::i32),
1127 DAG.getConstant(paramCount, MVT::i32),
1128 DAG.getConstant(sz, MVT::i32), InFlag };
1129 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1131 InFlag = Chain.getValue(1);
1132 unsigned NumElts = ObjectVT.getVectorNumElements();
1133 EVT EltVT = ObjectVT.getVectorElementType();
1135 bool NeedExtend = false;
1136 if (EltVT.getSizeInBits() < 16) {
1143 SDValue Elt = OutVals[OIdx++];
1145 Elt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt);
1147 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1148 SDValue CopyParamOps[] = { Chain,
1149 DAG.getConstant(paramCount, MVT::i32),
1150 DAG.getConstant(0, MVT::i32), Elt,
1152 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl,
1153 CopyParamVTs, CopyParamOps,
1154 MemVT, MachinePointerInfo());
1155 InFlag = Chain.getValue(1);
1156 } else if (NumElts == 2) {
1157 SDValue Elt0 = OutVals[OIdx++];
1158 SDValue Elt1 = OutVals[OIdx++];
1160 Elt0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt0);
1161 Elt1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Elt1);
1164 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1165 SDValue CopyParamOps[] = { Chain,
1166 DAG.getConstant(paramCount, MVT::i32),
1167 DAG.getConstant(0, MVT::i32), Elt0, Elt1,
1169 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParamV2, dl,
1170 CopyParamVTs, CopyParamOps,
1171 MemVT, MachinePointerInfo());
1172 InFlag = Chain.getValue(1);
1174 unsigned curOffset = 0;
1176 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
1178 // vector will be expanded to a power of 2 elements, so we know we can
1179 // always round up to the next multiple of 4 when creating the vector
1181 // e.g. 4 elem => 1 st.v4
1182 // 6 elem => 2 st.v4
1183 // 8 elem => 2 st.v4
1184 // 11 elem => 3 st.v4
1185 unsigned VecSize = 4;
1186 if (EltVT.getSizeInBits() == 64)
1189 // This is potentially only part of a vector, so assume all elements
1190 // are packed together.
1191 unsigned PerStoreOffset = MemVT.getStoreSizeInBits() / 8 * VecSize;
1193 for (unsigned i = 0; i < NumElts; i += VecSize) {
1196 SmallVector<SDValue, 8> Ops;
1197 Ops.push_back(Chain);
1198 Ops.push_back(DAG.getConstant(paramCount, MVT::i32));
1199 Ops.push_back(DAG.getConstant(curOffset, MVT::i32));
1201 unsigned Opc = NVPTXISD::StoreParamV2;
1203 StoreVal = OutVals[OIdx++];
1205 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1206 Ops.push_back(StoreVal);
1208 if (i + 1 < NumElts) {
1209 StoreVal = OutVals[OIdx++];
1212 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1214 StoreVal = DAG.getUNDEF(EltVT);
1216 Ops.push_back(StoreVal);
1219 Opc = NVPTXISD::StoreParamV4;
1220 if (i + 2 < NumElts) {
1221 StoreVal = OutVals[OIdx++];
1224 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1226 StoreVal = DAG.getUNDEF(EltVT);
1228 Ops.push_back(StoreVal);
1230 if (i + 3 < NumElts) {
1231 StoreVal = OutVals[OIdx++];
1234 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
1236 StoreVal = DAG.getUNDEF(EltVT);
1238 Ops.push_back(StoreVal);
1241 Ops.push_back(InFlag);
1243 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1244 Chain = DAG.getMemIntrinsicNode(Opc, dl, CopyParamVTs, Ops,
1245 MemVT, MachinePointerInfo());
1246 InFlag = Chain.getValue(1);
1247 curOffset += PerStoreOffset;
1255 // for ABI, declare .param .b<size> .param<n>;
1256 unsigned sz = VT.getSizeInBits();
1257 bool needExtend = false;
1258 if (VT.isInteger()) {
1264 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1265 SDValue DeclareParamOps[] = { Chain,
1266 DAG.getConstant(paramCount, MVT::i32),
1267 DAG.getConstant(sz, MVT::i32),
1268 DAG.getConstant(0, MVT::i32), InFlag };
1269 Chain = DAG.getNode(NVPTXISD::DeclareScalarParam, dl, DeclareParamVTs,
1271 InFlag = Chain.getValue(1);
1272 SDValue OutV = OutVals[OIdx];
1274 // zext/sext i1 to i16
1275 unsigned opc = ISD::ZERO_EXTEND;
1276 if (Outs[OIdx].Flags.isSExt())
1277 opc = ISD::SIGN_EXTEND;
1278 OutV = DAG.getNode(opc, dl, MVT::i16, OutV);
1280 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1281 SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32),
1282 DAG.getConstant(0, MVT::i32), OutV, InFlag };
1284 unsigned opcode = NVPTXISD::StoreParam;
1285 if (Outs[OIdx].Flags.isZExt())
1286 opcode = NVPTXISD::StoreParamU32;
1287 else if (Outs[OIdx].Flags.isSExt())
1288 opcode = NVPTXISD::StoreParamS32;
1289 Chain = DAG.getMemIntrinsicNode(opcode, dl, CopyParamVTs, CopyParamOps,
1290 VT, MachinePointerInfo());
1292 InFlag = Chain.getValue(1);
1297 SmallVector<EVT, 16> vtparts;
1298 SmallVector<uint64_t, 16> Offsets;
1299 const PointerType *PTy = dyn_cast<PointerType>(Args[i].Ty);
1300 assert(PTy && "Type of a byval parameter should be pointer");
1301 ComputePTXValueVTs(*this, PTy->getElementType(), vtparts, &Offsets, 0);
1303 // declare .param .align <align> .b8 .param<n>[<size>];
1304 unsigned sz = Outs[OIdx].Flags.getByValSize();
1305 SDVTList DeclareParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1306 unsigned ArgAlign = Outs[OIdx].Flags.getByValAlign();
1307 // The ByValAlign in the Outs[OIdx].Flags is alway set at this point,
1308 // so we don't need to worry about natural alignment or not.
1309 // See TargetLowering::LowerCallTo().
1310 SDValue DeclareParamOps[] = {
1311 Chain, DAG.getConstant(Outs[OIdx].Flags.getByValAlign(), MVT::i32),
1312 DAG.getConstant(paramCount, MVT::i32), DAG.getConstant(sz, MVT::i32),
1315 Chain = DAG.getNode(NVPTXISD::DeclareParam, dl, DeclareParamVTs,
1317 InFlag = Chain.getValue(1);
1318 for (unsigned j = 0, je = vtparts.size(); j != je; ++j) {
1319 EVT elemtype = vtparts[j];
1320 int curOffset = Offsets[j];
1321 unsigned PartAlign = GreatestCommonDivisor64(ArgAlign, curOffset);
1323 DAG.getNode(ISD::ADD, dl, getPointerTy(), OutVals[OIdx],
1324 DAG.getConstant(curOffset, getPointerTy()));
1325 SDValue theVal = DAG.getLoad(elemtype, dl, tempChain, srcAddr,
1326 MachinePointerInfo(), false, false, false,
1328 if (elemtype.getSizeInBits() < 16) {
1329 theVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, theVal);
1331 SDVTList CopyParamVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1332 SDValue CopyParamOps[] = { Chain, DAG.getConstant(paramCount, MVT::i32),
1333 DAG.getConstant(curOffset, MVT::i32), theVal,
1335 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreParam, dl, CopyParamVTs,
1336 CopyParamOps, elemtype,
1337 MachinePointerInfo());
1339 InFlag = Chain.getValue(1);
1344 GlobalAddressSDNode *Func = dyn_cast<GlobalAddressSDNode>(Callee.getNode());
1345 unsigned retAlignment = 0;
1348 if (Ins.size() > 0) {
1349 SmallVector<EVT, 16> resvtparts;
1350 ComputeValueVTs(*this, retTy, resvtparts);
1353 // .param .align 16 .b8 retval0[<size-in-bytes>], or
1354 // .param .b<size-in-bits> retval0
1355 unsigned resultsz = TD->getTypeAllocSizeInBits(retTy);
1356 // Emit ".param .b<size-in-bits> retval0" instead of byte arrays only for
1357 // these three types to match the logic in
1358 // NVPTXAsmPrinter::printReturnValStr and NVPTXTargetLowering::getPrototype.
1359 // Plus, this behavior is consistent with nvcc's.
1360 if (retTy->isFloatingPointTy() || retTy->isIntegerTy() ||
1361 retTy->isPointerTy()) {
1362 // Scalar needs to be at least 32bit wide
1365 SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1366 SDValue DeclareRetOps[] = { Chain, DAG.getConstant(1, MVT::i32),
1367 DAG.getConstant(resultsz, MVT::i32),
1368 DAG.getConstant(0, MVT::i32), InFlag };
1369 Chain = DAG.getNode(NVPTXISD::DeclareRet, dl, DeclareRetVTs,
1371 InFlag = Chain.getValue(1);
1373 retAlignment = getArgumentAlignment(Callee, CS, retTy, 0);
1374 SDVTList DeclareRetVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1375 SDValue DeclareRetOps[] = { Chain,
1376 DAG.getConstant(retAlignment, MVT::i32),
1377 DAG.getConstant(resultsz / 8, MVT::i32),
1378 DAG.getConstant(0, MVT::i32), InFlag };
1379 Chain = DAG.getNode(NVPTXISD::DeclareRetParam, dl, DeclareRetVTs,
1381 InFlag = Chain.getValue(1);
1386 // This is indirect function call case : PTX requires a prototype of the
1388 // proto_0 : .callprototype(.param .b32 _) _ (.param .b32 _);
1389 // to be emitted, and the label has to used as the last arg of call
1391 // The prototype is embedded in a string and put as the operand for a
1392 // CallPrototype SDNode which will print out to the value of the string.
1393 SDVTList ProtoVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1394 std::string Proto = getPrototype(retTy, Args, Outs, retAlignment, CS);
1395 const char *ProtoStr =
1396 nvTM->getManagedStrPool()->getManagedString(Proto.c_str())->c_str();
1397 SDValue ProtoOps[] = {
1398 Chain, DAG.getTargetExternalSymbol(ProtoStr, MVT::i32), InFlag,
1400 Chain = DAG.getNode(NVPTXISD::CallPrototype, dl, ProtoVTs, ProtoOps);
1401 InFlag = Chain.getValue(1);
1403 // Op to just print "call"
1404 SDVTList PrintCallVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1405 SDValue PrintCallOps[] = {
1406 Chain, DAG.getConstant((Ins.size() == 0) ? 0 : 1, MVT::i32), InFlag
1408 Chain = DAG.getNode(Func ? (NVPTXISD::PrintCallUni) : (NVPTXISD::PrintCall),
1409 dl, PrintCallVTs, PrintCallOps);
1410 InFlag = Chain.getValue(1);
1412 // Ops to print out the function name
1413 SDVTList CallVoidVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1414 SDValue CallVoidOps[] = { Chain, Callee, InFlag };
1415 Chain = DAG.getNode(NVPTXISD::CallVoid, dl, CallVoidVTs, CallVoidOps);
1416 InFlag = Chain.getValue(1);
1418 // Ops to print out the param list
1419 SDVTList CallArgBeginVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1420 SDValue CallArgBeginOps[] = { Chain, InFlag };
1421 Chain = DAG.getNode(NVPTXISD::CallArgBegin, dl, CallArgBeginVTs,
1423 InFlag = Chain.getValue(1);
1425 for (unsigned i = 0, e = paramCount; i != e; ++i) {
1428 opcode = NVPTXISD::LastCallArg;
1430 opcode = NVPTXISD::CallArg;
1431 SDVTList CallArgVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1432 SDValue CallArgOps[] = { Chain, DAG.getConstant(1, MVT::i32),
1433 DAG.getConstant(i, MVT::i32), InFlag };
1434 Chain = DAG.getNode(opcode, dl, CallArgVTs, CallArgOps);
1435 InFlag = Chain.getValue(1);
1437 SDVTList CallArgEndVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1438 SDValue CallArgEndOps[] = { Chain, DAG.getConstant(Func ? 1 : 0, MVT::i32),
1440 Chain = DAG.getNode(NVPTXISD::CallArgEnd, dl, CallArgEndVTs, CallArgEndOps);
1441 InFlag = Chain.getValue(1);
1444 SDVTList PrototypeVTs = DAG.getVTList(MVT::Other, MVT::Glue);
1445 SDValue PrototypeOps[] = { Chain, DAG.getConstant(uniqueCallSite, MVT::i32),
1447 Chain = DAG.getNode(NVPTXISD::Prototype, dl, PrototypeVTs, PrototypeOps);
1448 InFlag = Chain.getValue(1);
1451 // Generate loads from param memory/moves from registers for result
1452 if (Ins.size() > 0) {
1453 if (retTy && retTy->isVectorTy()) {
1454 EVT ObjectVT = getValueType(retTy);
1455 unsigned NumElts = ObjectVT.getVectorNumElements();
1456 EVT EltVT = ObjectVT.getVectorElementType();
1457 assert(nvTM->getSubtargetImpl()->getTargetLowering()->getNumRegisters(
1458 F->getContext(), ObjectVT) == NumElts &&
1459 "Vector was not scalarized");
1460 unsigned sz = EltVT.getSizeInBits();
1461 bool needTruncate = sz < 8 ? true : false;
1464 // Just a simple load
1465 SmallVector<EVT, 4> LoadRetVTs;
1466 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1467 // If loading i1/i8 result, generate
1471 LoadRetVTs.push_back(MVT::i16);
1473 LoadRetVTs.push_back(EltVT);
1474 LoadRetVTs.push_back(MVT::Other);
1475 LoadRetVTs.push_back(MVT::Glue);
1476 SmallVector<SDValue, 4> LoadRetOps;
1477 LoadRetOps.push_back(Chain);
1478 LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
1479 LoadRetOps.push_back(DAG.getConstant(0, MVT::i32));
1480 LoadRetOps.push_back(InFlag);
1481 SDValue retval = DAG.getMemIntrinsicNode(
1482 NVPTXISD::LoadParam, dl,
1483 DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
1484 Chain = retval.getValue(1);
1485 InFlag = retval.getValue(2);
1486 SDValue Ret0 = retval;
1488 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Ret0);
1489 InVals.push_back(Ret0);
1490 } else if (NumElts == 2) {
1492 SmallVector<EVT, 4> LoadRetVTs;
1493 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1494 // If loading i1/i8 result, generate
1498 LoadRetVTs.push_back(MVT::i16);
1499 LoadRetVTs.push_back(MVT::i16);
1501 LoadRetVTs.push_back(EltVT);
1502 LoadRetVTs.push_back(EltVT);
1504 LoadRetVTs.push_back(MVT::Other);
1505 LoadRetVTs.push_back(MVT::Glue);
1506 SmallVector<SDValue, 4> LoadRetOps;
1507 LoadRetOps.push_back(Chain);
1508 LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
1509 LoadRetOps.push_back(DAG.getConstant(0, MVT::i32));
1510 LoadRetOps.push_back(InFlag);
1511 SDValue retval = DAG.getMemIntrinsicNode(
1512 NVPTXISD::LoadParamV2, dl,
1513 DAG.getVTList(LoadRetVTs), LoadRetOps, EltVT, MachinePointerInfo());
1514 Chain = retval.getValue(2);
1515 InFlag = retval.getValue(3);
1516 SDValue Ret0 = retval.getValue(0);
1517 SDValue Ret1 = retval.getValue(1);
1519 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret0);
1520 InVals.push_back(Ret0);
1521 Ret1 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Ret1);
1522 InVals.push_back(Ret1);
1524 InVals.push_back(Ret0);
1525 InVals.push_back(Ret1);
1528 // Split into N LoadV4
1530 unsigned VecSize = 4;
1531 unsigned Opc = NVPTXISD::LoadParamV4;
1532 if (EltVT.getSizeInBits() == 64) {
1534 Opc = NVPTXISD::LoadParamV2;
1536 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
1537 for (unsigned i = 0; i < NumElts; i += VecSize) {
1538 SmallVector<EVT, 8> LoadRetVTs;
1539 if (EltVT == MVT::i1 || EltVT == MVT::i8) {
1540 // If loading i1/i8 result, generate
1544 for (unsigned j = 0; j < VecSize; ++j)
1545 LoadRetVTs.push_back(MVT::i16);
1547 for (unsigned j = 0; j < VecSize; ++j)
1548 LoadRetVTs.push_back(EltVT);
1550 LoadRetVTs.push_back(MVT::Other);
1551 LoadRetVTs.push_back(MVT::Glue);
1552 SmallVector<SDValue, 4> LoadRetOps;
1553 LoadRetOps.push_back(Chain);
1554 LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
1555 LoadRetOps.push_back(DAG.getConstant(Ofst, MVT::i32));
1556 LoadRetOps.push_back(InFlag);
1557 SDValue retval = DAG.getMemIntrinsicNode(
1558 Opc, dl, DAG.getVTList(LoadRetVTs),
1559 LoadRetOps, EltVT, MachinePointerInfo());
1561 Chain = retval.getValue(2);
1562 InFlag = retval.getValue(3);
1564 Chain = retval.getValue(4);
1565 InFlag = retval.getValue(5);
1568 for (unsigned j = 0; j < VecSize; ++j) {
1569 if (i + j >= NumElts)
1571 SDValue Elt = retval.getValue(j);
1573 Elt = DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
1574 InVals.push_back(Elt);
1576 Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
1580 SmallVector<EVT, 16> VTs;
1581 SmallVector<uint64_t, 16> Offsets;
1582 ComputePTXValueVTs(*this, retTy, VTs, &Offsets, 0);
1583 assert(VTs.size() == Ins.size() && "Bad value decomposition");
1584 unsigned RetAlign = getArgumentAlignment(Callee, CS, retTy, 0);
1585 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
1586 unsigned sz = VTs[i].getSizeInBits();
1587 unsigned AlignI = GreatestCommonDivisor64(RetAlign, Offsets[i]);
1588 bool needTruncate = sz < 8 ? true : false;
1589 if (VTs[i].isInteger() && (sz < 8))
1592 SmallVector<EVT, 4> LoadRetVTs;
1593 EVT TheLoadType = VTs[i];
1594 if (retTy->isIntegerTy() &&
1595 TD->getTypeAllocSizeInBits(retTy) < 32) {
1596 // This is for integer types only, and specifically not for
1598 LoadRetVTs.push_back(MVT::i32);
1599 TheLoadType = MVT::i32;
1600 } else if (sz < 16) {
1601 // If loading i1/i8 result, generate
1603 // trunc i16 to i1/i8
1604 LoadRetVTs.push_back(MVT::i16);
1606 LoadRetVTs.push_back(Ins[i].VT);
1607 LoadRetVTs.push_back(MVT::Other);
1608 LoadRetVTs.push_back(MVT::Glue);
1610 SmallVector<SDValue, 4> LoadRetOps;
1611 LoadRetOps.push_back(Chain);
1612 LoadRetOps.push_back(DAG.getConstant(1, MVT::i32));
1613 LoadRetOps.push_back(DAG.getConstant(Offsets[i], MVT::i32));
1614 LoadRetOps.push_back(InFlag);
1615 SDValue retval = DAG.getMemIntrinsicNode(
1616 NVPTXISD::LoadParam, dl,
1617 DAG.getVTList(LoadRetVTs), LoadRetOps,
1618 TheLoadType, MachinePointerInfo(), AlignI);
1619 Chain = retval.getValue(1);
1620 InFlag = retval.getValue(2);
1621 SDValue Ret0 = retval.getValue(0);
1623 Ret0 = DAG.getNode(ISD::TRUNCATE, dl, Ins[i].VT, Ret0);
1624 InVals.push_back(Ret0);
1629 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(uniqueCallSite, true),
1630 DAG.getIntPtrConstant(uniqueCallSite + 1, true),
1634 // set isTailCall to false for now, until we figure out how to express
1635 // tail call optimization in PTX
1640 // By default CONCAT_VECTORS is lowered by ExpandVectorBuildThroughStack()
1641 // (see LegalizeDAG.cpp). This is slow and uses local memory.
1642 // We use extract/insert/build vector just as what LegalizeOp() does in llvm 2.5
1644 NVPTXTargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
1645 SDNode *Node = Op.getNode();
1647 SmallVector<SDValue, 8> Ops;
1648 unsigned NumOperands = Node->getNumOperands();
1649 for (unsigned i = 0; i < NumOperands; ++i) {
1650 SDValue SubOp = Node->getOperand(i);
1651 EVT VVT = SubOp.getNode()->getValueType(0);
1652 EVT EltVT = VVT.getVectorElementType();
1653 unsigned NumSubElem = VVT.getVectorNumElements();
1654 for (unsigned j = 0; j < NumSubElem; ++j) {
1655 Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, SubOp,
1656 DAG.getIntPtrConstant(j)));
1659 return DAG.getNode(ISD::BUILD_VECTOR, dl, Node->getValueType(0), Ops);
1662 /// LowerShiftRightParts - Lower SRL_PARTS, SRA_PARTS, which
1663 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1665 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1667 SDValue NVPTXTargetLowering::LowerShiftRightParts(SDValue Op,
1668 SelectionDAG &DAG) const {
1669 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1670 assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
1672 EVT VT = Op.getValueType();
1673 unsigned VTBits = VT.getSizeInBits();
1675 SDValue ShOpLo = Op.getOperand(0);
1676 SDValue ShOpHi = Op.getOperand(1);
1677 SDValue ShAmt = Op.getOperand(2);
1678 unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
1680 if (VTBits == 32 && nvptxSubtarget.getSmVersion() >= 35) {
1682 // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1683 // {dHi, dLo} = {aHi, aLo} >> Amt
1685 // dLo = shf.r.clamp aLo, aHi, Amt
1687 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1688 SDValue Lo = DAG.getNode(NVPTXISD::FUN_SHFR_CLAMP, dl, VT, ShOpLo, ShOpHi,
1691 SDValue Ops[2] = { Lo, Hi };
1692 return DAG.getMergeValues(Ops, dl);
1696 // {dHi, dLo} = {aHi, aLo} >> Amt
1697 // - if (Amt>=size) then
1698 // dLo = aHi >> (Amt-size)
1699 // dHi = aHi >> Amt (this is either all 0 or all 1)
1701 // dLo = (aLo >>logic Amt) | (aHi << (size-Amt))
1704 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1705 DAG.getConstant(VTBits, MVT::i32), ShAmt);
1706 SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
1707 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1708 DAG.getConstant(VTBits, MVT::i32));
1709 SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
1710 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
1711 SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
1713 SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
1714 DAG.getConstant(VTBits, MVT::i32), ISD::SETGE);
1715 SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
1716 SDValue Lo = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
1718 SDValue Ops[2] = { Lo, Hi };
1719 return DAG.getMergeValues(Ops, dl);
1723 /// LowerShiftLeftParts - Lower SHL_PARTS, which
1724 /// 1) returns two i32 values and take a 2 x i32 value to shift plus a shift
1726 /// 2) returns two i64 values and take a 2 x i64 value to shift plus a shift
1728 SDValue NVPTXTargetLowering::LowerShiftLeftParts(SDValue Op,
1729 SelectionDAG &DAG) const {
1730 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
1731 assert(Op.getOpcode() == ISD::SHL_PARTS);
1733 EVT VT = Op.getValueType();
1734 unsigned VTBits = VT.getSizeInBits();
1736 SDValue ShOpLo = Op.getOperand(0);
1737 SDValue ShOpHi = Op.getOperand(1);
1738 SDValue ShAmt = Op.getOperand(2);
1740 if (VTBits == 32 && nvptxSubtarget.getSmVersion() >= 35) {
1742 // For 32bit and sm35, we can use the funnel shift 'shf' instruction.
1743 // {dHi, dLo} = {aHi, aLo} << Amt
1744 // dHi = shf.l.clamp aLo, aHi, Amt
1747 SDValue Hi = DAG.getNode(NVPTXISD::FUN_SHFL_CLAMP, dl, VT, ShOpLo, ShOpHi,
1749 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1751 SDValue Ops[2] = { Lo, Hi };
1752 return DAG.getMergeValues(Ops, dl);
1756 // {dHi, dLo} = {aHi, aLo} << Amt
1757 // - if (Amt>=size) then
1758 // dLo = aLo << Amt (all 0)
1759 // dLo = aLo << (Amt-size)
1762 // dHi = (aHi << Amt) | (aLo >> (size-Amt))
1764 SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
1765 DAG.getConstant(VTBits, MVT::i32), ShAmt);
1766 SDValue Tmp1 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
1767 SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
1768 DAG.getConstant(VTBits, MVT::i32));
1769 SDValue Tmp2 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
1770 SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
1771 SDValue TrueVal = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
1773 SDValue Cmp = DAG.getSetCC(dl, MVT::i1, ShAmt,
1774 DAG.getConstant(VTBits, MVT::i32), ISD::SETGE);
1775 SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
1776 SDValue Hi = DAG.getNode(ISD::SELECT, dl, VT, Cmp, TrueVal, FalseVal);
1778 SDValue Ops[2] = { Lo, Hi };
1779 return DAG.getMergeValues(Ops, dl);
1784 NVPTXTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
1785 switch (Op.getOpcode()) {
1786 case ISD::RETURNADDR:
1788 case ISD::FRAMEADDR:
1790 case ISD::GlobalAddress:
1791 return LowerGlobalAddress(Op, DAG);
1792 case ISD::INTRINSIC_W_CHAIN:
1794 case ISD::BUILD_VECTOR:
1795 case ISD::EXTRACT_SUBVECTOR:
1797 case ISD::CONCAT_VECTORS:
1798 return LowerCONCAT_VECTORS(Op, DAG);
1800 return LowerSTORE(Op, DAG);
1802 return LowerLOAD(Op, DAG);
1803 case ISD::SHL_PARTS:
1804 return LowerShiftLeftParts(Op, DAG);
1805 case ISD::SRA_PARTS:
1806 case ISD::SRL_PARTS:
1807 return LowerShiftRightParts(Op, DAG);
1809 llvm_unreachable("Custom lowering not defined for operation");
1813 SDValue NVPTXTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
1814 if (Op.getValueType() == MVT::i1)
1815 return LowerLOADi1(Op, DAG);
1822 // v1 = ld i8* addr (-> i16)
1823 // v = trunc i16 to i1
1824 SDValue NVPTXTargetLowering::LowerLOADi1(SDValue Op, SelectionDAG &DAG) const {
1825 SDNode *Node = Op.getNode();
1826 LoadSDNode *LD = cast<LoadSDNode>(Node);
1828 assert(LD->getExtensionType() == ISD::NON_EXTLOAD);
1829 assert(Node->getValueType(0) == MVT::i1 &&
1830 "Custom lowering for i1 load only");
1832 DAG.getLoad(MVT::i16, dl, LD->getChain(), LD->getBasePtr(),
1833 LD->getPointerInfo(), LD->isVolatile(), LD->isNonTemporal(),
1834 LD->isInvariant(), LD->getAlignment());
1835 SDValue result = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, newLD);
1836 // The legalizer (the caller) is expecting two values from the legalized
1837 // load, so we build a MergeValues node for it. See ExpandUnalignedLoad()
1838 // in LegalizeDAG.cpp which also uses MergeValues.
1839 SDValue Ops[] = { result, LD->getChain() };
1840 return DAG.getMergeValues(Ops, dl);
1843 SDValue NVPTXTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
1844 EVT ValVT = Op.getOperand(1).getValueType();
1845 if (ValVT == MVT::i1)
1846 return LowerSTOREi1(Op, DAG);
1847 else if (ValVT.isVector())
1848 return LowerSTOREVector(Op, DAG);
1854 NVPTXTargetLowering::LowerSTOREVector(SDValue Op, SelectionDAG &DAG) const {
1855 SDNode *N = Op.getNode();
1856 SDValue Val = N->getOperand(1);
1858 EVT ValVT = Val.getValueType();
1860 if (ValVT.isVector()) {
1861 // We only handle "native" vector sizes for now, e.g. <4 x double> is not
1862 // legal. We can (and should) split that into 2 stores of <2 x double> here
1863 // but I'm leaving that as a TODO for now.
1864 if (!ValVT.isSimple())
1866 switch (ValVT.getSimpleVT().SimpleTy) {
1879 // This is a "native" vector type
1883 MemSDNode *MemSD = cast<MemSDNode>(N);
1884 const DataLayout *TD = getDataLayout();
1886 unsigned Align = MemSD->getAlignment();
1887 unsigned PrefAlign =
1888 TD->getPrefTypeAlignment(ValVT.getTypeForEVT(*DAG.getContext()));
1889 if (Align < PrefAlign) {
1890 // This store is not sufficiently aligned, so bail out and let this vector
1891 // store be scalarized. Note that we may still be able to emit smaller
1892 // vector stores. For example, if we are storing a <4 x float> with an
1893 // alignment of 8, this check will fail but the legalizer will try again
1894 // with 2 x <2 x float>, which will succeed with an alignment of 8.
1898 unsigned Opcode = 0;
1899 EVT EltVT = ValVT.getVectorElementType();
1900 unsigned NumElts = ValVT.getVectorNumElements();
1902 // Since StoreV2 is a target node, we cannot rely on DAG type legalization.
1903 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
1904 // stored type to i16 and propagate the "real" type as the memory type.
1905 bool NeedExt = false;
1906 if (EltVT.getSizeInBits() < 16)
1913 Opcode = NVPTXISD::StoreV2;
1916 Opcode = NVPTXISD::StoreV4;
1921 SmallVector<SDValue, 8> Ops;
1923 // First is the chain
1924 Ops.push_back(N->getOperand(0));
1926 // Then the split values
1927 for (unsigned i = 0; i < NumElts; ++i) {
1928 SDValue ExtVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, Val,
1929 DAG.getIntPtrConstant(i));
1931 ExtVal = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i16, ExtVal);
1932 Ops.push_back(ExtVal);
1935 // Then any remaining arguments
1936 for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i) {
1937 Ops.push_back(N->getOperand(i));
1940 SDValue NewSt = DAG.getMemIntrinsicNode(
1941 Opcode, DL, DAG.getVTList(MVT::Other), Ops,
1942 MemSD->getMemoryVT(), MemSD->getMemOperand());
1944 //return DCI.CombineTo(N, NewSt, true);
1953 // v1 = zxt v to i16
1955 SDValue NVPTXTargetLowering::LowerSTOREi1(SDValue Op, SelectionDAG &DAG) const {
1956 SDNode *Node = Op.getNode();
1958 StoreSDNode *ST = cast<StoreSDNode>(Node);
1959 SDValue Tmp1 = ST->getChain();
1960 SDValue Tmp2 = ST->getBasePtr();
1961 SDValue Tmp3 = ST->getValue();
1962 assert(Tmp3.getValueType() == MVT::i1 && "Custom lowering for i1 store only");
1963 unsigned Alignment = ST->getAlignment();
1964 bool isVolatile = ST->isVolatile();
1965 bool isNonTemporal = ST->isNonTemporal();
1966 Tmp3 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Tmp3);
1967 SDValue Result = DAG.getTruncStore(Tmp1, dl, Tmp3, Tmp2,
1968 ST->getPointerInfo(), MVT::i8, isNonTemporal,
1969 isVolatile, Alignment);
1973 SDValue NVPTXTargetLowering::getExtSymb(SelectionDAG &DAG, const char *inname,
1974 int idx, EVT v) const {
1975 std::string *name = nvTM->getManagedStrPool()->getManagedString(inname);
1976 std::stringstream suffix;
1978 *name += suffix.str();
1979 return DAG.getTargetExternalSymbol(name->c_str(), v);
1983 NVPTXTargetLowering::getParamSymbol(SelectionDAG &DAG, int idx, EVT v) const {
1984 std::string ParamSym;
1985 raw_string_ostream ParamStr(ParamSym);
1987 ParamStr << DAG.getMachineFunction().getName() << "_param_" << idx;
1990 std::string *SavedStr =
1991 nvTM->getManagedStrPool()->getManagedString(ParamSym.c_str());
1992 return DAG.getTargetExternalSymbol(SavedStr->c_str(), v);
1995 SDValue NVPTXTargetLowering::getParamHelpSymbol(SelectionDAG &DAG, int idx) {
1996 return getExtSymb(DAG, ".HLPPARAM", idx);
1999 // Check to see if the kernel argument is image*_t or sampler_t
2001 bool llvm::isImageOrSamplerVal(const Value *arg, const Module *context) {
2002 static const char *const specialTypes[] = { "struct._image2d_t",
2003 "struct._image3d_t",
2004 "struct._sampler_t" };
2006 const Type *Ty = arg->getType();
2007 const PointerType *PTy = dyn_cast<PointerType>(Ty);
2015 const StructType *STy = dyn_cast<StructType>(PTy->getElementType());
2016 const std::string TypeName = STy && !STy->isLiteral() ? STy->getName() : "";
2018 for (int i = 0, e = array_lengthof(specialTypes); i != e; ++i)
2019 if (TypeName == specialTypes[i])
2025 SDValue NVPTXTargetLowering::LowerFormalArguments(
2026 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2027 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2028 SmallVectorImpl<SDValue> &InVals) const {
2029 MachineFunction &MF = DAG.getMachineFunction();
2030 const DataLayout *TD = getDataLayout();
2032 const Function *F = MF.getFunction();
2033 const AttributeSet &PAL = F->getAttributes();
2034 const TargetLowering *TLI = DAG.getSubtarget().getTargetLowering();
2036 SDValue Root = DAG.getRoot();
2037 std::vector<SDValue> OutChains;
2039 bool isKernel = llvm::isKernelFunction(*F);
2040 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
2041 assert(isABI && "Non-ABI compilation is not supported");
2045 std::vector<Type *> argTypes;
2046 std::vector<const Argument *> theArgs;
2047 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
2049 theArgs.push_back(I);
2050 argTypes.push_back(I->getType());
2052 // argTypes.size() (or theArgs.size()) and Ins.size() need not match.
2053 // Ins.size() will be larger
2054 // * if there is an aggregate argument with multiple fields (each field
2055 // showing up separately in Ins)
2056 // * if there is a vector argument with more than typical vector-length
2057 // elements (generally if more than 4) where each vector element is
2058 // individually present in Ins.
2059 // So a different index should be used for indexing into Ins.
2060 // See similar issue in LowerCall.
2061 unsigned InsIdx = 0;
2064 for (unsigned i = 0, e = theArgs.size(); i != e; ++i, ++idx, ++InsIdx) {
2065 Type *Ty = argTypes[i];
2067 // If the kernel argument is image*_t or sampler_t, convert it to
2068 // a i32 constant holding the parameter position. This can later
2069 // matched in the AsmPrinter to output the correct mangled name.
2070 if (isImageOrSamplerVal(
2072 (theArgs[i]->getParent() ? theArgs[i]->getParent()->getParent()
2074 assert(isKernel && "Only kernels can have image/sampler params");
2075 InVals.push_back(DAG.getConstant(i + 1, MVT::i32));
2079 if (theArgs[i]->use_empty()) {
2081 if (Ty->isAggregateType()) {
2082 SmallVector<EVT, 16> vtparts;
2084 ComputePTXValueVTs(*this, Ty, vtparts);
2085 assert(vtparts.size() > 0 && "empty aggregate type not expected");
2086 for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2088 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2091 if (vtparts.size() > 0)
2095 if (Ty->isVectorTy()) {
2096 EVT ObjectVT = getValueType(Ty);
2097 unsigned NumRegs = TLI->getNumRegisters(F->getContext(), ObjectVT);
2098 for (unsigned parti = 0; parti < NumRegs; ++parti) {
2099 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2106 InVals.push_back(DAG.getNode(ISD::UNDEF, dl, Ins[InsIdx].VT));
2110 // In the following cases, assign a node order of "idx+1"
2111 // to newly created nodes. The SDNodes for params have to
2112 // appear in the same order as their order of appearance
2113 // in the original function. "idx+1" holds that order.
2114 if (PAL.hasAttribute(i + 1, Attribute::ByVal) == false) {
2115 if (Ty->isAggregateType()) {
2116 SmallVector<EVT, 16> vtparts;
2117 SmallVector<uint64_t, 16> offsets;
2119 // NOTE: Here, we lose the ability to issue vector loads for vectors
2120 // that are a part of a struct. This should be investigated in the
2122 ComputePTXValueVTs(*this, Ty, vtparts, &offsets, 0);
2123 assert(vtparts.size() > 0 && "empty aggregate type not expected");
2124 bool aggregateIsPacked = false;
2125 if (StructType *STy = llvm::dyn_cast<StructType>(Ty))
2126 aggregateIsPacked = STy->isPacked();
2128 SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2129 for (unsigned parti = 0, parte = vtparts.size(); parti != parte;
2131 EVT partVT = vtparts[parti];
2132 Value *srcValue = Constant::getNullValue(
2133 PointerType::get(partVT.getTypeForEVT(F->getContext()),
2134 llvm::ADDRESS_SPACE_PARAM));
2136 DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
2137 DAG.getConstant(offsets[parti], getPointerTy()));
2138 unsigned partAlign =
2139 aggregateIsPacked ? 1
2140 : TD->getABITypeAlignment(
2141 partVT.getTypeForEVT(F->getContext()));
2143 if (Ins[InsIdx].VT.getSizeInBits() > partVT.getSizeInBits()) {
2144 ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ?
2145 ISD::SEXTLOAD : ISD::ZEXTLOAD;
2146 p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, srcAddr,
2147 MachinePointerInfo(srcValue), partVT, false,
2148 false, false, partAlign);
2150 p = DAG.getLoad(partVT, dl, Root, srcAddr,
2151 MachinePointerInfo(srcValue), false, false, false,
2155 p.getNode()->setIROrder(idx + 1);
2156 InVals.push_back(p);
2159 if (vtparts.size() > 0)
2163 if (Ty->isVectorTy()) {
2164 EVT ObjectVT = getValueType(Ty);
2165 SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2166 unsigned NumElts = ObjectVT.getVectorNumElements();
2167 assert(TLI->getNumRegisters(F->getContext(), ObjectVT) == NumElts &&
2168 "Vector was not scalarized");
2169 EVT EltVT = ObjectVT.getVectorElementType();
2174 // We only have one element, so just directly load it
2175 Value *SrcValue = Constant::getNullValue(PointerType::get(
2176 EltVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2177 SDValue P = DAG.getLoad(
2178 EltVT, dl, Root, Arg, MachinePointerInfo(SrcValue), false,
2180 TD->getABITypeAlignment(EltVT.getTypeForEVT(F->getContext())));
2182 P.getNode()->setIROrder(idx + 1);
2184 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
2185 P = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, P);
2186 InVals.push_back(P);
2188 } else if (NumElts == 2) {
2190 // f32,f32 = load ...
2191 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, 2);
2192 Value *SrcValue = Constant::getNullValue(PointerType::get(
2193 VecVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2194 SDValue P = DAG.getLoad(
2195 VecVT, dl, Root, Arg, MachinePointerInfo(SrcValue), false,
2197 TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
2199 P.getNode()->setIROrder(idx + 1);
2201 SDValue Elt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2202 DAG.getIntPtrConstant(0));
2203 SDValue Elt1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2204 DAG.getIntPtrConstant(1));
2206 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits()) {
2207 Elt0 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt0);
2208 Elt1 = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt1);
2211 InVals.push_back(Elt0);
2212 InVals.push_back(Elt1);
2216 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and
2218 // vector will be expanded to a power of 2 elements, so we know we can
2219 // always round up to the next multiple of 4 when creating the vector
2221 // e.g. 4 elem => 1 ld.v4
2222 // 6 elem => 2 ld.v4
2223 // 8 elem => 2 ld.v4
2224 // 11 elem => 3 ld.v4
2225 unsigned VecSize = 4;
2226 if (EltVT.getSizeInBits() == 64) {
2229 EVT VecVT = EVT::getVectorVT(F->getContext(), EltVT, VecSize);
2231 for (unsigned i = 0; i < NumElts; i += VecSize) {
2232 Value *SrcValue = Constant::getNullValue(
2233 PointerType::get(VecVT.getTypeForEVT(F->getContext()),
2234 llvm::ADDRESS_SPACE_PARAM));
2236 DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg,
2237 DAG.getConstant(Ofst, getPointerTy()));
2238 SDValue P = DAG.getLoad(
2239 VecVT, dl, Root, SrcAddr, MachinePointerInfo(SrcValue), false,
2241 TD->getABITypeAlignment(VecVT.getTypeForEVT(F->getContext())));
2243 P.getNode()->setIROrder(idx + 1);
2245 for (unsigned j = 0; j < VecSize; ++j) {
2246 if (i + j >= NumElts)
2248 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, P,
2249 DAG.getIntPtrConstant(j));
2250 if (Ins[InsIdx].VT.getSizeInBits() > EltVT.getSizeInBits())
2251 Elt = DAG.getNode(ISD::ANY_EXTEND, dl, Ins[InsIdx].VT, Elt);
2252 InVals.push_back(Elt);
2254 Ofst += TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
2264 EVT ObjectVT = getValueType(Ty);
2265 // If ABI, load from the param symbol
2266 SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2267 Value *srcValue = Constant::getNullValue(PointerType::get(
2268 ObjectVT.getTypeForEVT(F->getContext()), llvm::ADDRESS_SPACE_PARAM));
2270 if (ObjectVT.getSizeInBits() < Ins[InsIdx].VT.getSizeInBits()) {
2271 ISD::LoadExtType ExtOp = Ins[InsIdx].Flags.isSExt() ?
2272 ISD::SEXTLOAD : ISD::ZEXTLOAD;
2273 p = DAG.getExtLoad(ExtOp, dl, Ins[InsIdx].VT, Root, Arg,
2274 MachinePointerInfo(srcValue), ObjectVT, false, false,
2276 TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
2278 p = DAG.getLoad(Ins[InsIdx].VT, dl, Root, Arg,
2279 MachinePointerInfo(srcValue), false, false, false,
2280 TD->getABITypeAlignment(ObjectVT.getTypeForEVT(F->getContext())));
2283 p.getNode()->setIROrder(idx + 1);
2284 InVals.push_back(p);
2288 // Param has ByVal attribute
2289 // Return MoveParam(param symbol).
2290 // Ideally, the param symbol can be returned directly,
2291 // but when SDNode builder decides to use it in a CopyToReg(),
2292 // machine instruction fails because TargetExternalSymbol
2293 // (not lowered) is target dependent, and CopyToReg assumes
2294 // the source is lowered.
2295 EVT ObjectVT = getValueType(Ty);
2296 assert(ObjectVT == Ins[InsIdx].VT &&
2297 "Ins type did not match function type");
2298 SDValue Arg = getParamSymbol(DAG, idx, getPointerTy());
2299 SDValue p = DAG.getNode(NVPTXISD::MoveParam, dl, ObjectVT, Arg);
2301 p.getNode()->setIROrder(idx + 1);
2303 InVals.push_back(p);
2305 SDValue p2 = DAG.getNode(
2306 ISD::INTRINSIC_WO_CHAIN, dl, ObjectVT,
2307 DAG.getConstant(Intrinsic::nvvm_ptr_local_to_gen, MVT::i32), p);
2308 InVals.push_back(p2);
2312 // Clang will check explicit VarArg and issue error if any. However, Clang
2313 // will let code with
2314 // implicit var arg like f() pass. See bug 617733.
2315 // We treat this case as if the arg list is empty.
2316 // if (F.isVarArg()) {
2317 // assert(0 && "VarArg not supported yet!");
2320 if (!OutChains.empty())
2321 DAG.setRoot(DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains));
2328 NVPTXTargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
2330 const SmallVectorImpl<ISD::OutputArg> &Outs,
2331 const SmallVectorImpl<SDValue> &OutVals,
2332 SDLoc dl, SelectionDAG &DAG) const {
2333 MachineFunction &MF = DAG.getMachineFunction();
2334 const Function *F = MF.getFunction();
2335 Type *RetTy = F->getReturnType();
2336 const DataLayout *TD = getDataLayout();
2338 bool isABI = (nvptxSubtarget.getSmVersion() >= 20);
2339 assert(isABI && "Non-ABI compilation is not supported");
2343 if (VectorType *VTy = dyn_cast<VectorType>(RetTy)) {
2344 // If we have a vector type, the OutVals array will be the scalarized
2345 // components and we have combine them into 1 or more vector stores.
2346 unsigned NumElts = VTy->getNumElements();
2347 assert(NumElts == Outs.size() && "Bad scalarization of return value");
2349 // const_cast can be removed in later LLVM versions
2350 EVT EltVT = getValueType(RetTy).getVectorElementType();
2351 bool NeedExtend = false;
2352 if (EltVT.getSizeInBits() < 16)
2357 SDValue StoreVal = OutVals[0];
2358 // We only have one element, so just directly store it
2360 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal);
2361 SDValue Ops[] = { Chain, DAG.getConstant(0, MVT::i32), StoreVal };
2362 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
2363 DAG.getVTList(MVT::Other), Ops,
2364 EltVT, MachinePointerInfo());
2366 } else if (NumElts == 2) {
2368 SDValue StoreVal0 = OutVals[0];
2369 SDValue StoreVal1 = OutVals[1];
2372 StoreVal0 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal0);
2373 StoreVal1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, StoreVal1);
2376 SDValue Ops[] = { Chain, DAG.getConstant(0, MVT::i32), StoreVal0,
2378 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetvalV2, dl,
2379 DAG.getVTList(MVT::Other), Ops,
2380 EltVT, MachinePointerInfo());
2383 // We have at least 4 elements (<3 x Ty> expands to 4 elements) and the
2384 // vector will be expanded to a power of 2 elements, so we know we can
2385 // always round up to the next multiple of 4 when creating the vector
2387 // e.g. 4 elem => 1 st.v4
2388 // 6 elem => 2 st.v4
2389 // 8 elem => 2 st.v4
2390 // 11 elem => 3 st.v4
2392 unsigned VecSize = 4;
2393 if (OutVals[0].getValueType().getSizeInBits() == 64)
2396 unsigned Offset = 0;
2399 EVT::getVectorVT(F->getContext(), EltVT, VecSize);
2400 unsigned PerStoreOffset =
2401 TD->getTypeAllocSize(VecVT.getTypeForEVT(F->getContext()));
2403 for (unsigned i = 0; i < NumElts; i += VecSize) {
2406 SmallVector<SDValue, 8> Ops;
2407 Ops.push_back(Chain);
2408 Ops.push_back(DAG.getConstant(Offset, MVT::i32));
2409 unsigned Opc = NVPTXISD::StoreRetvalV2;
2410 EVT ExtendedVT = (NeedExtend) ? MVT::i16 : OutVals[0].getValueType();
2412 StoreVal = OutVals[i];
2414 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2415 Ops.push_back(StoreVal);
2417 if (i + 1 < NumElts) {
2418 StoreVal = OutVals[i + 1];
2420 StoreVal = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2422 StoreVal = DAG.getUNDEF(ExtendedVT);
2424 Ops.push_back(StoreVal);
2427 Opc = NVPTXISD::StoreRetvalV4;
2428 if (i + 2 < NumElts) {
2429 StoreVal = OutVals[i + 2];
2432 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2434 StoreVal = DAG.getUNDEF(ExtendedVT);
2436 Ops.push_back(StoreVal);
2438 if (i + 3 < NumElts) {
2439 StoreVal = OutVals[i + 3];
2442 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtendedVT, StoreVal);
2444 StoreVal = DAG.getUNDEF(ExtendedVT);
2446 Ops.push_back(StoreVal);
2449 // Chain = DAG.getNode(Opc, dl, MVT::Other, &Ops[0], Ops.size());
2451 DAG.getMemIntrinsicNode(Opc, dl, DAG.getVTList(MVT::Other), Ops,
2452 EltVT, MachinePointerInfo());
2453 Offset += PerStoreOffset;
2457 SmallVector<EVT, 16> ValVTs;
2458 SmallVector<uint64_t, 16> Offsets;
2459 ComputePTXValueVTs(*this, RetTy, ValVTs, &Offsets, 0);
2460 assert(ValVTs.size() == OutVals.size() && "Bad return value decomposition");
2462 for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
2463 SDValue theVal = OutVals[i];
2464 EVT TheValType = theVal.getValueType();
2465 unsigned numElems = 1;
2466 if (TheValType.isVector())
2467 numElems = TheValType.getVectorNumElements();
2468 for (unsigned j = 0, je = numElems; j != je; ++j) {
2469 SDValue TmpVal = theVal;
2470 if (TheValType.isVector())
2471 TmpVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
2472 TheValType.getVectorElementType(), TmpVal,
2473 DAG.getIntPtrConstant(j));
2474 EVT TheStoreType = ValVTs[i];
2475 if (RetTy->isIntegerTy() &&
2476 TD->getTypeAllocSizeInBits(RetTy) < 32) {
2477 // The following zero-extension is for integer types only, and
2478 // specifically not for aggregates.
2479 TmpVal = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, TmpVal);
2480 TheStoreType = MVT::i32;
2482 else if (TmpVal.getValueType().getSizeInBits() < 16)
2483 TmpVal = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i16, TmpVal);
2487 DAG.getConstant(Offsets[i], MVT::i32),
2489 Chain = DAG.getMemIntrinsicNode(NVPTXISD::StoreRetval, dl,
2490 DAG.getVTList(MVT::Other), Ops,
2492 MachinePointerInfo());
2497 return DAG.getNode(NVPTXISD::RET_FLAG, dl, MVT::Other, Chain);
2501 void NVPTXTargetLowering::LowerAsmOperandForConstraint(
2502 SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
2503 SelectionDAG &DAG) const {
2504 if (Constraint.length() > 1)
2507 TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
2510 // NVPTX suuport vector of legal types of any length in Intrinsics because the
2511 // NVPTX specific type legalizer
2512 // will legalize them to the PTX supported length.
2513 bool NVPTXTargetLowering::isTypeSupportedInIntrinsic(MVT VT) const {
2514 if (isTypeLegal(VT))
2516 if (VT.isVector()) {
2517 MVT eVT = VT.getVectorElementType();
2518 if (isTypeLegal(eVT))
2524 static unsigned getOpcForTextureInstr(unsigned Intrinsic) {
2525 switch (Intrinsic) {
2529 case Intrinsic::nvvm_tex_1d_v4f32_s32:
2530 return NVPTXISD::Tex1DFloatS32;
2531 case Intrinsic::nvvm_tex_1d_v4f32_f32:
2532 return NVPTXISD::Tex1DFloatFloat;
2533 case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
2534 return NVPTXISD::Tex1DFloatFloatLevel;
2535 case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
2536 return NVPTXISD::Tex1DFloatFloatGrad;
2537 case Intrinsic::nvvm_tex_1d_v4s32_s32:
2538 return NVPTXISD::Tex1DS32S32;
2539 case Intrinsic::nvvm_tex_1d_v4s32_f32:
2540 return NVPTXISD::Tex1DS32Float;
2541 case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
2542 return NVPTXISD::Tex1DS32FloatLevel;
2543 case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
2544 return NVPTXISD::Tex1DS32FloatGrad;
2545 case Intrinsic::nvvm_tex_1d_v4u32_s32:
2546 return NVPTXISD::Tex1DU32S32;
2547 case Intrinsic::nvvm_tex_1d_v4u32_f32:
2548 return NVPTXISD::Tex1DU32Float;
2549 case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
2550 return NVPTXISD::Tex1DU32FloatLevel;
2551 case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
2552 return NVPTXISD::Tex1DU32FloatGrad;
2554 case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
2555 return NVPTXISD::Tex1DArrayFloatS32;
2556 case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
2557 return NVPTXISD::Tex1DArrayFloatFloat;
2558 case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
2559 return NVPTXISD::Tex1DArrayFloatFloatLevel;
2560 case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
2561 return NVPTXISD::Tex1DArrayFloatFloatGrad;
2562 case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
2563 return NVPTXISD::Tex1DArrayS32S32;
2564 case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
2565 return NVPTXISD::Tex1DArrayS32Float;
2566 case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
2567 return NVPTXISD::Tex1DArrayS32FloatLevel;
2568 case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
2569 return NVPTXISD::Tex1DArrayS32FloatGrad;
2570 case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
2571 return NVPTXISD::Tex1DArrayU32S32;
2572 case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
2573 return NVPTXISD::Tex1DArrayU32Float;
2574 case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
2575 return NVPTXISD::Tex1DArrayU32FloatLevel;
2576 case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
2577 return NVPTXISD::Tex1DArrayU32FloatGrad;
2579 case Intrinsic::nvvm_tex_2d_v4f32_s32:
2580 return NVPTXISD::Tex2DFloatS32;
2581 case Intrinsic::nvvm_tex_2d_v4f32_f32:
2582 return NVPTXISD::Tex2DFloatFloat;
2583 case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
2584 return NVPTXISD::Tex2DFloatFloatLevel;
2585 case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
2586 return NVPTXISD::Tex2DFloatFloatGrad;
2587 case Intrinsic::nvvm_tex_2d_v4s32_s32:
2588 return NVPTXISD::Tex2DS32S32;
2589 case Intrinsic::nvvm_tex_2d_v4s32_f32:
2590 return NVPTXISD::Tex2DS32Float;
2591 case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
2592 return NVPTXISD::Tex2DS32FloatLevel;
2593 case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
2594 return NVPTXISD::Tex2DS32FloatGrad;
2595 case Intrinsic::nvvm_tex_2d_v4u32_s32:
2596 return NVPTXISD::Tex2DU32S32;
2597 case Intrinsic::nvvm_tex_2d_v4u32_f32:
2598 return NVPTXISD::Tex2DU32Float;
2599 case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
2600 return NVPTXISD::Tex2DU32FloatLevel;
2601 case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
2602 return NVPTXISD::Tex2DU32FloatGrad;
2604 case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
2605 return NVPTXISD::Tex2DArrayFloatS32;
2606 case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
2607 return NVPTXISD::Tex2DArrayFloatFloat;
2608 case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
2609 return NVPTXISD::Tex2DArrayFloatFloatLevel;
2610 case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
2611 return NVPTXISD::Tex2DArrayFloatFloatGrad;
2612 case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
2613 return NVPTXISD::Tex2DArrayS32S32;
2614 case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
2615 return NVPTXISD::Tex2DArrayS32Float;
2616 case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
2617 return NVPTXISD::Tex2DArrayS32FloatLevel;
2618 case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
2619 return NVPTXISD::Tex2DArrayS32FloatGrad;
2620 case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
2621 return NVPTXISD::Tex2DArrayU32S32;
2622 case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
2623 return NVPTXISD::Tex2DArrayU32Float;
2624 case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
2625 return NVPTXISD::Tex2DArrayU32FloatLevel;
2626 case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
2627 return NVPTXISD::Tex2DArrayU32FloatGrad;
2629 case Intrinsic::nvvm_tex_3d_v4f32_s32:
2630 return NVPTXISD::Tex3DFloatS32;
2631 case Intrinsic::nvvm_tex_3d_v4f32_f32:
2632 return NVPTXISD::Tex3DFloatFloat;
2633 case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
2634 return NVPTXISD::Tex3DFloatFloatLevel;
2635 case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
2636 return NVPTXISD::Tex3DFloatFloatGrad;
2637 case Intrinsic::nvvm_tex_3d_v4s32_s32:
2638 return NVPTXISD::Tex3DS32S32;
2639 case Intrinsic::nvvm_tex_3d_v4s32_f32:
2640 return NVPTXISD::Tex3DS32Float;
2641 case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
2642 return NVPTXISD::Tex3DS32FloatLevel;
2643 case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
2644 return NVPTXISD::Tex3DS32FloatGrad;
2645 case Intrinsic::nvvm_tex_3d_v4u32_s32:
2646 return NVPTXISD::Tex3DU32S32;
2647 case Intrinsic::nvvm_tex_3d_v4u32_f32:
2648 return NVPTXISD::Tex3DU32Float;
2649 case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
2650 return NVPTXISD::Tex3DU32FloatLevel;
2651 case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
2652 return NVPTXISD::Tex3DU32FloatGrad;
2654 case Intrinsic::nvvm_tex_cube_v4f32_f32:
2655 return NVPTXISD::TexCubeFloatFloat;
2656 case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
2657 return NVPTXISD::TexCubeFloatFloatLevel;
2658 case Intrinsic::nvvm_tex_cube_v4s32_f32:
2659 return NVPTXISD::TexCubeS32Float;
2660 case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
2661 return NVPTXISD::TexCubeS32FloatLevel;
2662 case Intrinsic::nvvm_tex_cube_v4u32_f32:
2663 return NVPTXISD::TexCubeU32Float;
2664 case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
2665 return NVPTXISD::TexCubeU32FloatLevel;
2667 case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
2668 return NVPTXISD::TexCubeArrayFloatFloat;
2669 case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
2670 return NVPTXISD::TexCubeArrayFloatFloatLevel;
2671 case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
2672 return NVPTXISD::TexCubeArrayS32Float;
2673 case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
2674 return NVPTXISD::TexCubeArrayS32FloatLevel;
2675 case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
2676 return NVPTXISD::TexCubeArrayU32Float;
2677 case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
2678 return NVPTXISD::TexCubeArrayU32FloatLevel;
2680 case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
2681 return NVPTXISD::Tld4R2DFloatFloat;
2682 case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
2683 return NVPTXISD::Tld4G2DFloatFloat;
2684 case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
2685 return NVPTXISD::Tld4B2DFloatFloat;
2686 case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
2687 return NVPTXISD::Tld4A2DFloatFloat;
2688 case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
2689 return NVPTXISD::Tld4R2DS64Float;
2690 case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
2691 return NVPTXISD::Tld4G2DS64Float;
2692 case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
2693 return NVPTXISD::Tld4B2DS64Float;
2694 case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
2695 return NVPTXISD::Tld4A2DS64Float;
2696 case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
2697 return NVPTXISD::Tld4R2DU64Float;
2698 case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
2699 return NVPTXISD::Tld4G2DU64Float;
2700 case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
2701 return NVPTXISD::Tld4B2DU64Float;
2702 case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
2703 return NVPTXISD::Tld4A2DU64Float;
2705 case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
2706 return NVPTXISD::TexUnified1DFloatS32;
2707 case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
2708 return NVPTXISD::TexUnified1DFloatFloat;
2709 case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
2710 return NVPTXISD::TexUnified1DFloatFloatLevel;
2711 case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
2712 return NVPTXISD::TexUnified1DFloatFloatGrad;
2713 case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
2714 return NVPTXISD::TexUnified1DS32S32;
2715 case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
2716 return NVPTXISD::TexUnified1DS32Float;
2717 case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
2718 return NVPTXISD::TexUnified1DS32FloatLevel;
2719 case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
2720 return NVPTXISD::TexUnified1DS32FloatGrad;
2721 case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
2722 return NVPTXISD::TexUnified1DU32S32;
2723 case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
2724 return NVPTXISD::TexUnified1DU32Float;
2725 case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
2726 return NVPTXISD::TexUnified1DU32FloatLevel;
2727 case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
2728 return NVPTXISD::TexUnified1DU32FloatGrad;
2730 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
2731 return NVPTXISD::TexUnified1DArrayFloatS32;
2732 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
2733 return NVPTXISD::TexUnified1DArrayFloatFloat;
2734 case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
2735 return NVPTXISD::TexUnified1DArrayFloatFloatLevel;
2736 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
2737 return NVPTXISD::TexUnified1DArrayFloatFloatGrad;
2738 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
2739 return NVPTXISD::TexUnified1DArrayS32S32;
2740 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
2741 return NVPTXISD::TexUnified1DArrayS32Float;
2742 case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
2743 return NVPTXISD::TexUnified1DArrayS32FloatLevel;
2744 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
2745 return NVPTXISD::TexUnified1DArrayS32FloatGrad;
2746 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
2747 return NVPTXISD::TexUnified1DArrayU32S32;
2748 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
2749 return NVPTXISD::TexUnified1DArrayU32Float;
2750 case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
2751 return NVPTXISD::TexUnified1DArrayU32FloatLevel;
2752 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
2753 return NVPTXISD::TexUnified1DArrayU32FloatGrad;
2755 case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
2756 return NVPTXISD::TexUnified2DFloatS32;
2757 case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
2758 return NVPTXISD::TexUnified2DFloatFloat;
2759 case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
2760 return NVPTXISD::TexUnified2DFloatFloatLevel;
2761 case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
2762 return NVPTXISD::TexUnified2DFloatFloatGrad;
2763 case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
2764 return NVPTXISD::TexUnified2DS32S32;
2765 case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
2766 return NVPTXISD::TexUnified2DS32Float;
2767 case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
2768 return NVPTXISD::TexUnified2DS32FloatLevel;
2769 case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
2770 return NVPTXISD::TexUnified2DS32FloatGrad;
2771 case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
2772 return NVPTXISD::TexUnified2DU32S32;
2773 case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
2774 return NVPTXISD::TexUnified2DU32Float;
2775 case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
2776 return NVPTXISD::TexUnified2DU32FloatLevel;
2777 case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
2778 return NVPTXISD::TexUnified2DU32FloatGrad;
2780 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
2781 return NVPTXISD::TexUnified2DArrayFloatS32;
2782 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
2783 return NVPTXISD::TexUnified2DArrayFloatFloat;
2784 case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
2785 return NVPTXISD::TexUnified2DArrayFloatFloatLevel;
2786 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
2787 return NVPTXISD::TexUnified2DArrayFloatFloatGrad;
2788 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
2789 return NVPTXISD::TexUnified2DArrayS32S32;
2790 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
2791 return NVPTXISD::TexUnified2DArrayS32Float;
2792 case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
2793 return NVPTXISD::TexUnified2DArrayS32FloatLevel;
2794 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
2795 return NVPTXISD::TexUnified2DArrayS32FloatGrad;
2796 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
2797 return NVPTXISD::TexUnified2DArrayU32S32;
2798 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
2799 return NVPTXISD::TexUnified2DArrayU32Float;
2800 case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
2801 return NVPTXISD::TexUnified2DArrayU32FloatLevel;
2802 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
2803 return NVPTXISD::TexUnified2DArrayU32FloatGrad;
2805 case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
2806 return NVPTXISD::TexUnified3DFloatS32;
2807 case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
2808 return NVPTXISD::TexUnified3DFloatFloat;
2809 case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
2810 return NVPTXISD::TexUnified3DFloatFloatLevel;
2811 case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
2812 return NVPTXISD::TexUnified3DFloatFloatGrad;
2813 case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
2814 return NVPTXISD::TexUnified3DS32S32;
2815 case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
2816 return NVPTXISD::TexUnified3DS32Float;
2817 case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
2818 return NVPTXISD::TexUnified3DS32FloatLevel;
2819 case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
2820 return NVPTXISD::TexUnified3DS32FloatGrad;
2821 case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
2822 return NVPTXISD::TexUnified3DU32S32;
2823 case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
2824 return NVPTXISD::TexUnified3DU32Float;
2825 case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
2826 return NVPTXISD::TexUnified3DU32FloatLevel;
2827 case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
2828 return NVPTXISD::TexUnified3DU32FloatGrad;
2830 case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
2831 return NVPTXISD::TexUnifiedCubeFloatFloat;
2832 case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
2833 return NVPTXISD::TexUnifiedCubeFloatFloatLevel;
2834 case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
2835 return NVPTXISD::TexUnifiedCubeS32Float;
2836 case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
2837 return NVPTXISD::TexUnifiedCubeS32FloatLevel;
2838 case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
2839 return NVPTXISD::TexUnifiedCubeU32Float;
2840 case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
2841 return NVPTXISD::TexUnifiedCubeU32FloatLevel;
2843 case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
2844 return NVPTXISD::TexUnifiedCubeArrayFloatFloat;
2845 case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
2846 return NVPTXISD::TexUnifiedCubeArrayFloatFloatLevel;
2847 case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
2848 return NVPTXISD::TexUnifiedCubeArrayS32Float;
2849 case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
2850 return NVPTXISD::TexUnifiedCubeArrayS32FloatLevel;
2851 case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
2852 return NVPTXISD::TexUnifiedCubeArrayU32Float;
2853 case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
2854 return NVPTXISD::TexUnifiedCubeArrayU32FloatLevel;
2856 case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
2857 return NVPTXISD::Tld4UnifiedR2DFloatFloat;
2858 case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
2859 return NVPTXISD::Tld4UnifiedG2DFloatFloat;
2860 case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
2861 return NVPTXISD::Tld4UnifiedB2DFloatFloat;
2862 case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32:
2863 return NVPTXISD::Tld4UnifiedA2DFloatFloat;
2864 case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
2865 return NVPTXISD::Tld4UnifiedR2DS64Float;
2866 case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
2867 return NVPTXISD::Tld4UnifiedG2DS64Float;
2868 case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
2869 return NVPTXISD::Tld4UnifiedB2DS64Float;
2870 case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
2871 return NVPTXISD::Tld4UnifiedA2DS64Float;
2872 case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
2873 return NVPTXISD::Tld4UnifiedR2DU64Float;
2874 case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
2875 return NVPTXISD::Tld4UnifiedG2DU64Float;
2876 case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
2877 return NVPTXISD::Tld4UnifiedB2DU64Float;
2878 case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32:
2879 return NVPTXISD::Tld4UnifiedA2DU64Float;
2883 static unsigned getOpcForSurfaceInstr(unsigned Intrinsic) {
2884 switch (Intrinsic) {
2887 case Intrinsic::nvvm_suld_1d_i8_clamp:
2888 return NVPTXISD::Suld1DI8Clamp;
2889 case Intrinsic::nvvm_suld_1d_i16_clamp:
2890 return NVPTXISD::Suld1DI16Clamp;
2891 case Intrinsic::nvvm_suld_1d_i32_clamp:
2892 return NVPTXISD::Suld1DI32Clamp;
2893 case Intrinsic::nvvm_suld_1d_i64_clamp:
2894 return NVPTXISD::Suld1DI64Clamp;
2895 case Intrinsic::nvvm_suld_1d_v2i8_clamp:
2896 return NVPTXISD::Suld1DV2I8Clamp;
2897 case Intrinsic::nvvm_suld_1d_v2i16_clamp:
2898 return NVPTXISD::Suld1DV2I16Clamp;
2899 case Intrinsic::nvvm_suld_1d_v2i32_clamp:
2900 return NVPTXISD::Suld1DV2I32Clamp;
2901 case Intrinsic::nvvm_suld_1d_v2i64_clamp:
2902 return NVPTXISD::Suld1DV2I64Clamp;
2903 case Intrinsic::nvvm_suld_1d_v4i8_clamp:
2904 return NVPTXISD::Suld1DV4I8Clamp;
2905 case Intrinsic::nvvm_suld_1d_v4i16_clamp:
2906 return NVPTXISD::Suld1DV4I16Clamp;
2907 case Intrinsic::nvvm_suld_1d_v4i32_clamp:
2908 return NVPTXISD::Suld1DV4I32Clamp;
2909 case Intrinsic::nvvm_suld_1d_array_i8_clamp:
2910 return NVPTXISD::Suld1DArrayI8Clamp;
2911 case Intrinsic::nvvm_suld_1d_array_i16_clamp:
2912 return NVPTXISD::Suld1DArrayI16Clamp;
2913 case Intrinsic::nvvm_suld_1d_array_i32_clamp:
2914 return NVPTXISD::Suld1DArrayI32Clamp;
2915 case Intrinsic::nvvm_suld_1d_array_i64_clamp:
2916 return NVPTXISD::Suld1DArrayI64Clamp;
2917 case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
2918 return NVPTXISD::Suld1DArrayV2I8Clamp;
2919 case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
2920 return NVPTXISD::Suld1DArrayV2I16Clamp;
2921 case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
2922 return NVPTXISD::Suld1DArrayV2I32Clamp;
2923 case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
2924 return NVPTXISD::Suld1DArrayV2I64Clamp;
2925 case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
2926 return NVPTXISD::Suld1DArrayV4I8Clamp;
2927 case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
2928 return NVPTXISD::Suld1DArrayV4I16Clamp;
2929 case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
2930 return NVPTXISD::Suld1DArrayV4I32Clamp;
2931 case Intrinsic::nvvm_suld_2d_i8_clamp:
2932 return NVPTXISD::Suld2DI8Clamp;
2933 case Intrinsic::nvvm_suld_2d_i16_clamp:
2934 return NVPTXISD::Suld2DI16Clamp;
2935 case Intrinsic::nvvm_suld_2d_i32_clamp:
2936 return NVPTXISD::Suld2DI32Clamp;
2937 case Intrinsic::nvvm_suld_2d_i64_clamp:
2938 return NVPTXISD::Suld2DI64Clamp;
2939 case Intrinsic::nvvm_suld_2d_v2i8_clamp:
2940 return NVPTXISD::Suld2DV2I8Clamp;
2941 case Intrinsic::nvvm_suld_2d_v2i16_clamp:
2942 return NVPTXISD::Suld2DV2I16Clamp;
2943 case Intrinsic::nvvm_suld_2d_v2i32_clamp:
2944 return NVPTXISD::Suld2DV2I32Clamp;
2945 case Intrinsic::nvvm_suld_2d_v2i64_clamp:
2946 return NVPTXISD::Suld2DV2I64Clamp;
2947 case Intrinsic::nvvm_suld_2d_v4i8_clamp:
2948 return NVPTXISD::Suld2DV4I8Clamp;
2949 case Intrinsic::nvvm_suld_2d_v4i16_clamp:
2950 return NVPTXISD::Suld2DV4I16Clamp;
2951 case Intrinsic::nvvm_suld_2d_v4i32_clamp:
2952 return NVPTXISD::Suld2DV4I32Clamp;
2953 case Intrinsic::nvvm_suld_2d_array_i8_clamp:
2954 return NVPTXISD::Suld2DArrayI8Clamp;
2955 case Intrinsic::nvvm_suld_2d_array_i16_clamp:
2956 return NVPTXISD::Suld2DArrayI16Clamp;
2957 case Intrinsic::nvvm_suld_2d_array_i32_clamp:
2958 return NVPTXISD::Suld2DArrayI32Clamp;
2959 case Intrinsic::nvvm_suld_2d_array_i64_clamp:
2960 return NVPTXISD::Suld2DArrayI64Clamp;
2961 case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
2962 return NVPTXISD::Suld2DArrayV2I8Clamp;
2963 case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
2964 return NVPTXISD::Suld2DArrayV2I16Clamp;
2965 case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
2966 return NVPTXISD::Suld2DArrayV2I32Clamp;
2967 case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
2968 return NVPTXISD::Suld2DArrayV2I64Clamp;
2969 case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
2970 return NVPTXISD::Suld2DArrayV4I8Clamp;
2971 case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
2972 return NVPTXISD::Suld2DArrayV4I16Clamp;
2973 case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
2974 return NVPTXISD::Suld2DArrayV4I32Clamp;
2975 case Intrinsic::nvvm_suld_3d_i8_clamp:
2976 return NVPTXISD::Suld3DI8Clamp;
2977 case Intrinsic::nvvm_suld_3d_i16_clamp:
2978 return NVPTXISD::Suld3DI16Clamp;
2979 case Intrinsic::nvvm_suld_3d_i32_clamp:
2980 return NVPTXISD::Suld3DI32Clamp;
2981 case Intrinsic::nvvm_suld_3d_i64_clamp:
2982 return NVPTXISD::Suld3DI64Clamp;
2983 case Intrinsic::nvvm_suld_3d_v2i8_clamp:
2984 return NVPTXISD::Suld3DV2I8Clamp;
2985 case Intrinsic::nvvm_suld_3d_v2i16_clamp:
2986 return NVPTXISD::Suld3DV2I16Clamp;
2987 case Intrinsic::nvvm_suld_3d_v2i32_clamp:
2988 return NVPTXISD::Suld3DV2I32Clamp;
2989 case Intrinsic::nvvm_suld_3d_v2i64_clamp:
2990 return NVPTXISD::Suld3DV2I64Clamp;
2991 case Intrinsic::nvvm_suld_3d_v4i8_clamp:
2992 return NVPTXISD::Suld3DV4I8Clamp;
2993 case Intrinsic::nvvm_suld_3d_v4i16_clamp:
2994 return NVPTXISD::Suld3DV4I16Clamp;
2995 case Intrinsic::nvvm_suld_3d_v4i32_clamp:
2996 return NVPTXISD::Suld3DV4I32Clamp;
2997 case Intrinsic::nvvm_suld_1d_i8_trap:
2998 return NVPTXISD::Suld1DI8Trap;
2999 case Intrinsic::nvvm_suld_1d_i16_trap:
3000 return NVPTXISD::Suld1DI16Trap;
3001 case Intrinsic::nvvm_suld_1d_i32_trap:
3002 return NVPTXISD::Suld1DI32Trap;
3003 case Intrinsic::nvvm_suld_1d_i64_trap:
3004 return NVPTXISD::Suld1DI64Trap;
3005 case Intrinsic::nvvm_suld_1d_v2i8_trap:
3006 return NVPTXISD::Suld1DV2I8Trap;
3007 case Intrinsic::nvvm_suld_1d_v2i16_trap:
3008 return NVPTXISD::Suld1DV2I16Trap;
3009 case Intrinsic::nvvm_suld_1d_v2i32_trap:
3010 return NVPTXISD::Suld1DV2I32Trap;
3011 case Intrinsic::nvvm_suld_1d_v2i64_trap:
3012 return NVPTXISD::Suld1DV2I64Trap;
3013 case Intrinsic::nvvm_suld_1d_v4i8_trap:
3014 return NVPTXISD::Suld1DV4I8Trap;
3015 case Intrinsic::nvvm_suld_1d_v4i16_trap:
3016 return NVPTXISD::Suld1DV4I16Trap;
3017 case Intrinsic::nvvm_suld_1d_v4i32_trap:
3018 return NVPTXISD::Suld1DV4I32Trap;
3019 case Intrinsic::nvvm_suld_1d_array_i8_trap:
3020 return NVPTXISD::Suld1DArrayI8Trap;
3021 case Intrinsic::nvvm_suld_1d_array_i16_trap:
3022 return NVPTXISD::Suld1DArrayI16Trap;
3023 case Intrinsic::nvvm_suld_1d_array_i32_trap:
3024 return NVPTXISD::Suld1DArrayI32Trap;
3025 case Intrinsic::nvvm_suld_1d_array_i64_trap:
3026 return NVPTXISD::Suld1DArrayI64Trap;
3027 case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3028 return NVPTXISD::Suld1DArrayV2I8Trap;
3029 case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3030 return NVPTXISD::Suld1DArrayV2I16Trap;
3031 case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3032 return NVPTXISD::Suld1DArrayV2I32Trap;
3033 case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3034 return NVPTXISD::Suld1DArrayV2I64Trap;
3035 case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3036 return NVPTXISD::Suld1DArrayV4I8Trap;
3037 case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3038 return NVPTXISD::Suld1DArrayV4I16Trap;
3039 case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3040 return NVPTXISD::Suld1DArrayV4I32Trap;
3041 case Intrinsic::nvvm_suld_2d_i8_trap:
3042 return NVPTXISD::Suld2DI8Trap;
3043 case Intrinsic::nvvm_suld_2d_i16_trap:
3044 return NVPTXISD::Suld2DI16Trap;
3045 case Intrinsic::nvvm_suld_2d_i32_trap:
3046 return NVPTXISD::Suld2DI32Trap;
3047 case Intrinsic::nvvm_suld_2d_i64_trap:
3048 return NVPTXISD::Suld2DI64Trap;
3049 case Intrinsic::nvvm_suld_2d_v2i8_trap:
3050 return NVPTXISD::Suld2DV2I8Trap;
3051 case Intrinsic::nvvm_suld_2d_v2i16_trap:
3052 return NVPTXISD::Suld2DV2I16Trap;
3053 case Intrinsic::nvvm_suld_2d_v2i32_trap:
3054 return NVPTXISD::Suld2DV2I32Trap;
3055 case Intrinsic::nvvm_suld_2d_v2i64_trap:
3056 return NVPTXISD::Suld2DV2I64Trap;
3057 case Intrinsic::nvvm_suld_2d_v4i8_trap:
3058 return NVPTXISD::Suld2DV4I8Trap;
3059 case Intrinsic::nvvm_suld_2d_v4i16_trap:
3060 return NVPTXISD::Suld2DV4I16Trap;
3061 case Intrinsic::nvvm_suld_2d_v4i32_trap:
3062 return NVPTXISD::Suld2DV4I32Trap;
3063 case Intrinsic::nvvm_suld_2d_array_i8_trap:
3064 return NVPTXISD::Suld2DArrayI8Trap;
3065 case Intrinsic::nvvm_suld_2d_array_i16_trap:
3066 return NVPTXISD::Suld2DArrayI16Trap;
3067 case Intrinsic::nvvm_suld_2d_array_i32_trap:
3068 return NVPTXISD::Suld2DArrayI32Trap;
3069 case Intrinsic::nvvm_suld_2d_array_i64_trap:
3070 return NVPTXISD::Suld2DArrayI64Trap;
3071 case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3072 return NVPTXISD::Suld2DArrayV2I8Trap;
3073 case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3074 return NVPTXISD::Suld2DArrayV2I16Trap;
3075 case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3076 return NVPTXISD::Suld2DArrayV2I32Trap;
3077 case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3078 return NVPTXISD::Suld2DArrayV2I64Trap;
3079 case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3080 return NVPTXISD::Suld2DArrayV4I8Trap;
3081 case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3082 return NVPTXISD::Suld2DArrayV4I16Trap;
3083 case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3084 return NVPTXISD::Suld2DArrayV4I32Trap;
3085 case Intrinsic::nvvm_suld_3d_i8_trap:
3086 return NVPTXISD::Suld3DI8Trap;
3087 case Intrinsic::nvvm_suld_3d_i16_trap:
3088 return NVPTXISD::Suld3DI16Trap;
3089 case Intrinsic::nvvm_suld_3d_i32_trap:
3090 return NVPTXISD::Suld3DI32Trap;
3091 case Intrinsic::nvvm_suld_3d_i64_trap:
3092 return NVPTXISD::Suld3DI64Trap;
3093 case Intrinsic::nvvm_suld_3d_v2i8_trap:
3094 return NVPTXISD::Suld3DV2I8Trap;
3095 case Intrinsic::nvvm_suld_3d_v2i16_trap:
3096 return NVPTXISD::Suld3DV2I16Trap;
3097 case Intrinsic::nvvm_suld_3d_v2i32_trap:
3098 return NVPTXISD::Suld3DV2I32Trap;
3099 case Intrinsic::nvvm_suld_3d_v2i64_trap:
3100 return NVPTXISD::Suld3DV2I64Trap;
3101 case Intrinsic::nvvm_suld_3d_v4i8_trap:
3102 return NVPTXISD::Suld3DV4I8Trap;
3103 case Intrinsic::nvvm_suld_3d_v4i16_trap:
3104 return NVPTXISD::Suld3DV4I16Trap;
3105 case Intrinsic::nvvm_suld_3d_v4i32_trap:
3106 return NVPTXISD::Suld3DV4I32Trap;
3107 case Intrinsic::nvvm_suld_1d_i8_zero:
3108 return NVPTXISD::Suld1DI8Zero;
3109 case Intrinsic::nvvm_suld_1d_i16_zero:
3110 return NVPTXISD::Suld1DI16Zero;
3111 case Intrinsic::nvvm_suld_1d_i32_zero:
3112 return NVPTXISD::Suld1DI32Zero;
3113 case Intrinsic::nvvm_suld_1d_i64_zero:
3114 return NVPTXISD::Suld1DI64Zero;
3115 case Intrinsic::nvvm_suld_1d_v2i8_zero:
3116 return NVPTXISD::Suld1DV2I8Zero;
3117 case Intrinsic::nvvm_suld_1d_v2i16_zero:
3118 return NVPTXISD::Suld1DV2I16Zero;
3119 case Intrinsic::nvvm_suld_1d_v2i32_zero:
3120 return NVPTXISD::Suld1DV2I32Zero;
3121 case Intrinsic::nvvm_suld_1d_v2i64_zero:
3122 return NVPTXISD::Suld1DV2I64Zero;
3123 case Intrinsic::nvvm_suld_1d_v4i8_zero:
3124 return NVPTXISD::Suld1DV4I8Zero;
3125 case Intrinsic::nvvm_suld_1d_v4i16_zero:
3126 return NVPTXISD::Suld1DV4I16Zero;
3127 case Intrinsic::nvvm_suld_1d_v4i32_zero:
3128 return NVPTXISD::Suld1DV4I32Zero;
3129 case Intrinsic::nvvm_suld_1d_array_i8_zero:
3130 return NVPTXISD::Suld1DArrayI8Zero;
3131 case Intrinsic::nvvm_suld_1d_array_i16_zero:
3132 return NVPTXISD::Suld1DArrayI16Zero;
3133 case Intrinsic::nvvm_suld_1d_array_i32_zero:
3134 return NVPTXISD::Suld1DArrayI32Zero;
3135 case Intrinsic::nvvm_suld_1d_array_i64_zero:
3136 return NVPTXISD::Suld1DArrayI64Zero;
3137 case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3138 return NVPTXISD::Suld1DArrayV2I8Zero;
3139 case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3140 return NVPTXISD::Suld1DArrayV2I16Zero;
3141 case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3142 return NVPTXISD::Suld1DArrayV2I32Zero;
3143 case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3144 return NVPTXISD::Suld1DArrayV2I64Zero;
3145 case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3146 return NVPTXISD::Suld1DArrayV4I8Zero;
3147 case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3148 return NVPTXISD::Suld1DArrayV4I16Zero;
3149 case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3150 return NVPTXISD::Suld1DArrayV4I32Zero;
3151 case Intrinsic::nvvm_suld_2d_i8_zero:
3152 return NVPTXISD::Suld2DI8Zero;
3153 case Intrinsic::nvvm_suld_2d_i16_zero:
3154 return NVPTXISD::Suld2DI16Zero;
3155 case Intrinsic::nvvm_suld_2d_i32_zero:
3156 return NVPTXISD::Suld2DI32Zero;
3157 case Intrinsic::nvvm_suld_2d_i64_zero:
3158 return NVPTXISD::Suld2DI64Zero;
3159 case Intrinsic::nvvm_suld_2d_v2i8_zero:
3160 return NVPTXISD::Suld2DV2I8Zero;
3161 case Intrinsic::nvvm_suld_2d_v2i16_zero:
3162 return NVPTXISD::Suld2DV2I16Zero;
3163 case Intrinsic::nvvm_suld_2d_v2i32_zero:
3164 return NVPTXISD::Suld2DV2I32Zero;
3165 case Intrinsic::nvvm_suld_2d_v2i64_zero:
3166 return NVPTXISD::Suld2DV2I64Zero;
3167 case Intrinsic::nvvm_suld_2d_v4i8_zero:
3168 return NVPTXISD::Suld2DV4I8Zero;
3169 case Intrinsic::nvvm_suld_2d_v4i16_zero:
3170 return NVPTXISD::Suld2DV4I16Zero;
3171 case Intrinsic::nvvm_suld_2d_v4i32_zero:
3172 return NVPTXISD::Suld2DV4I32Zero;
3173 case Intrinsic::nvvm_suld_2d_array_i8_zero:
3174 return NVPTXISD::Suld2DArrayI8Zero;
3175 case Intrinsic::nvvm_suld_2d_array_i16_zero:
3176 return NVPTXISD::Suld2DArrayI16Zero;
3177 case Intrinsic::nvvm_suld_2d_array_i32_zero:
3178 return NVPTXISD::Suld2DArrayI32Zero;
3179 case Intrinsic::nvvm_suld_2d_array_i64_zero:
3180 return NVPTXISD::Suld2DArrayI64Zero;
3181 case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3182 return NVPTXISD::Suld2DArrayV2I8Zero;
3183 case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3184 return NVPTXISD::Suld2DArrayV2I16Zero;
3185 case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3186 return NVPTXISD::Suld2DArrayV2I32Zero;
3187 case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3188 return NVPTXISD::Suld2DArrayV2I64Zero;
3189 case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3190 return NVPTXISD::Suld2DArrayV4I8Zero;
3191 case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3192 return NVPTXISD::Suld2DArrayV4I16Zero;
3193 case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3194 return NVPTXISD::Suld2DArrayV4I32Zero;
3195 case Intrinsic::nvvm_suld_3d_i8_zero:
3196 return NVPTXISD::Suld3DI8Zero;
3197 case Intrinsic::nvvm_suld_3d_i16_zero:
3198 return NVPTXISD::Suld3DI16Zero;
3199 case Intrinsic::nvvm_suld_3d_i32_zero:
3200 return NVPTXISD::Suld3DI32Zero;
3201 case Intrinsic::nvvm_suld_3d_i64_zero:
3202 return NVPTXISD::Suld3DI64Zero;
3203 case Intrinsic::nvvm_suld_3d_v2i8_zero:
3204 return NVPTXISD::Suld3DV2I8Zero;
3205 case Intrinsic::nvvm_suld_3d_v2i16_zero:
3206 return NVPTXISD::Suld3DV2I16Zero;
3207 case Intrinsic::nvvm_suld_3d_v2i32_zero:
3208 return NVPTXISD::Suld3DV2I32Zero;
3209 case Intrinsic::nvvm_suld_3d_v2i64_zero:
3210 return NVPTXISD::Suld3DV2I64Zero;
3211 case Intrinsic::nvvm_suld_3d_v4i8_zero:
3212 return NVPTXISD::Suld3DV4I8Zero;
3213 case Intrinsic::nvvm_suld_3d_v4i16_zero:
3214 return NVPTXISD::Suld3DV4I16Zero;
3215 case Intrinsic::nvvm_suld_3d_v4i32_zero:
3216 return NVPTXISD::Suld3DV4I32Zero;
3220 // llvm.ptx.memcpy.const and llvm.ptx.memmove.const need to be modeled as
3222 // because we need the information that is only available in the "Value" type
3224 // pointer. In particular, the address space information.
3225 bool NVPTXTargetLowering::getTgtMemIntrinsic(
3226 IntrinsicInfo &Info, const CallInst &I, unsigned Intrinsic) const {
3227 switch (Intrinsic) {
3231 case Intrinsic::nvvm_atomic_load_add_f32:
3232 Info.opc = ISD::INTRINSIC_W_CHAIN;
3233 Info.memVT = MVT::f32;
3234 Info.ptrVal = I.getArgOperand(0);
3237 Info.readMem = true;
3238 Info.writeMem = true;
3242 case Intrinsic::nvvm_atomic_load_inc_32:
3243 case Intrinsic::nvvm_atomic_load_dec_32:
3244 Info.opc = ISD::INTRINSIC_W_CHAIN;
3245 Info.memVT = MVT::i32;
3246 Info.ptrVal = I.getArgOperand(0);
3249 Info.readMem = true;
3250 Info.writeMem = true;
3254 case Intrinsic::nvvm_ldu_global_i:
3255 case Intrinsic::nvvm_ldu_global_f:
3256 case Intrinsic::nvvm_ldu_global_p: {
3258 Info.opc = ISD::INTRINSIC_W_CHAIN;
3259 if (Intrinsic == Intrinsic::nvvm_ldu_global_i)
3260 Info.memVT = getValueType(I.getType());
3261 else if(Intrinsic == Intrinsic::nvvm_ldu_global_p)
3262 Info.memVT = getPointerTy();
3264 Info.memVT = getValueType(I.getType());
3265 Info.ptrVal = I.getArgOperand(0);
3268 Info.readMem = true;
3269 Info.writeMem = false;
3270 Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
3274 case Intrinsic::nvvm_ldg_global_i:
3275 case Intrinsic::nvvm_ldg_global_f:
3276 case Intrinsic::nvvm_ldg_global_p: {
3278 Info.opc = ISD::INTRINSIC_W_CHAIN;
3279 if (Intrinsic == Intrinsic::nvvm_ldg_global_i)
3280 Info.memVT = getValueType(I.getType());
3281 else if(Intrinsic == Intrinsic::nvvm_ldg_global_p)
3282 Info.memVT = getPointerTy();
3284 Info.memVT = getValueType(I.getType());
3285 Info.ptrVal = I.getArgOperand(0);
3288 Info.readMem = true;
3289 Info.writeMem = false;
3290 Info.align = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
3295 case Intrinsic::nvvm_tex_1d_v4f32_s32:
3296 case Intrinsic::nvvm_tex_1d_v4f32_f32:
3297 case Intrinsic::nvvm_tex_1d_level_v4f32_f32:
3298 case Intrinsic::nvvm_tex_1d_grad_v4f32_f32:
3299 case Intrinsic::nvvm_tex_1d_array_v4f32_s32:
3300 case Intrinsic::nvvm_tex_1d_array_v4f32_f32:
3301 case Intrinsic::nvvm_tex_1d_array_level_v4f32_f32:
3302 case Intrinsic::nvvm_tex_1d_array_grad_v4f32_f32:
3303 case Intrinsic::nvvm_tex_2d_v4f32_s32:
3304 case Intrinsic::nvvm_tex_2d_v4f32_f32:
3305 case Intrinsic::nvvm_tex_2d_level_v4f32_f32:
3306 case Intrinsic::nvvm_tex_2d_grad_v4f32_f32:
3307 case Intrinsic::nvvm_tex_2d_array_v4f32_s32:
3308 case Intrinsic::nvvm_tex_2d_array_v4f32_f32:
3309 case Intrinsic::nvvm_tex_2d_array_level_v4f32_f32:
3310 case Intrinsic::nvvm_tex_2d_array_grad_v4f32_f32:
3311 case Intrinsic::nvvm_tex_3d_v4f32_s32:
3312 case Intrinsic::nvvm_tex_3d_v4f32_f32:
3313 case Intrinsic::nvvm_tex_3d_level_v4f32_f32:
3314 case Intrinsic::nvvm_tex_3d_grad_v4f32_f32:
3315 case Intrinsic::nvvm_tex_cube_v4f32_f32:
3316 case Intrinsic::nvvm_tex_cube_level_v4f32_f32:
3317 case Intrinsic::nvvm_tex_cube_array_v4f32_f32:
3318 case Intrinsic::nvvm_tex_cube_array_level_v4f32_f32:
3319 case Intrinsic::nvvm_tld4_r_2d_v4f32_f32:
3320 case Intrinsic::nvvm_tld4_g_2d_v4f32_f32:
3321 case Intrinsic::nvvm_tld4_b_2d_v4f32_f32:
3322 case Intrinsic::nvvm_tld4_a_2d_v4f32_f32:
3323 case Intrinsic::nvvm_tex_unified_1d_v4f32_s32:
3324 case Intrinsic::nvvm_tex_unified_1d_v4f32_f32:
3325 case Intrinsic::nvvm_tex_unified_1d_level_v4f32_f32:
3326 case Intrinsic::nvvm_tex_unified_1d_grad_v4f32_f32:
3327 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_s32:
3328 case Intrinsic::nvvm_tex_unified_1d_array_v4f32_f32:
3329 case Intrinsic::nvvm_tex_unified_1d_array_level_v4f32_f32:
3330 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4f32_f32:
3331 case Intrinsic::nvvm_tex_unified_2d_v4f32_s32:
3332 case Intrinsic::nvvm_tex_unified_2d_v4f32_f32:
3333 case Intrinsic::nvvm_tex_unified_2d_level_v4f32_f32:
3334 case Intrinsic::nvvm_tex_unified_2d_grad_v4f32_f32:
3335 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_s32:
3336 case Intrinsic::nvvm_tex_unified_2d_array_v4f32_f32:
3337 case Intrinsic::nvvm_tex_unified_2d_array_level_v4f32_f32:
3338 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4f32_f32:
3339 case Intrinsic::nvvm_tex_unified_3d_v4f32_s32:
3340 case Intrinsic::nvvm_tex_unified_3d_v4f32_f32:
3341 case Intrinsic::nvvm_tex_unified_3d_level_v4f32_f32:
3342 case Intrinsic::nvvm_tex_unified_3d_grad_v4f32_f32:
3343 case Intrinsic::nvvm_tex_unified_cube_v4f32_f32:
3344 case Intrinsic::nvvm_tex_unified_cube_level_v4f32_f32:
3345 case Intrinsic::nvvm_tex_unified_cube_array_v4f32_f32:
3346 case Intrinsic::nvvm_tex_unified_cube_array_level_v4f32_f32:
3347 case Intrinsic::nvvm_tld4_unified_r_2d_v4f32_f32:
3348 case Intrinsic::nvvm_tld4_unified_g_2d_v4f32_f32:
3349 case Intrinsic::nvvm_tld4_unified_b_2d_v4f32_f32:
3350 case Intrinsic::nvvm_tld4_unified_a_2d_v4f32_f32: {
3351 Info.opc = getOpcForTextureInstr(Intrinsic);
3352 Info.memVT = MVT::v4f32;
3353 Info.ptrVal = nullptr;
3356 Info.readMem = true;
3357 Info.writeMem = false;
3361 case Intrinsic::nvvm_tex_1d_v4s32_s32:
3362 case Intrinsic::nvvm_tex_1d_v4s32_f32:
3363 case Intrinsic::nvvm_tex_1d_level_v4s32_f32:
3364 case Intrinsic::nvvm_tex_1d_grad_v4s32_f32:
3365 case Intrinsic::nvvm_tex_1d_array_v4s32_s32:
3366 case Intrinsic::nvvm_tex_1d_array_v4s32_f32:
3367 case Intrinsic::nvvm_tex_1d_array_level_v4s32_f32:
3368 case Intrinsic::nvvm_tex_1d_array_grad_v4s32_f32:
3369 case Intrinsic::nvvm_tex_2d_v4s32_s32:
3370 case Intrinsic::nvvm_tex_2d_v4s32_f32:
3371 case Intrinsic::nvvm_tex_2d_level_v4s32_f32:
3372 case Intrinsic::nvvm_tex_2d_grad_v4s32_f32:
3373 case Intrinsic::nvvm_tex_2d_array_v4s32_s32:
3374 case Intrinsic::nvvm_tex_2d_array_v4s32_f32:
3375 case Intrinsic::nvvm_tex_2d_array_level_v4s32_f32:
3376 case Intrinsic::nvvm_tex_2d_array_grad_v4s32_f32:
3377 case Intrinsic::nvvm_tex_3d_v4s32_s32:
3378 case Intrinsic::nvvm_tex_3d_v4s32_f32:
3379 case Intrinsic::nvvm_tex_3d_level_v4s32_f32:
3380 case Intrinsic::nvvm_tex_3d_grad_v4s32_f32:
3381 case Intrinsic::nvvm_tex_cube_v4s32_f32:
3382 case Intrinsic::nvvm_tex_cube_level_v4s32_f32:
3383 case Intrinsic::nvvm_tex_cube_array_v4s32_f32:
3384 case Intrinsic::nvvm_tex_cube_array_level_v4s32_f32:
3385 case Intrinsic::nvvm_tex_cube_v4u32_f32:
3386 case Intrinsic::nvvm_tex_cube_level_v4u32_f32:
3387 case Intrinsic::nvvm_tex_cube_array_v4u32_f32:
3388 case Intrinsic::nvvm_tex_cube_array_level_v4u32_f32:
3389 case Intrinsic::nvvm_tex_1d_v4u32_s32:
3390 case Intrinsic::nvvm_tex_1d_v4u32_f32:
3391 case Intrinsic::nvvm_tex_1d_level_v4u32_f32:
3392 case Intrinsic::nvvm_tex_1d_grad_v4u32_f32:
3393 case Intrinsic::nvvm_tex_1d_array_v4u32_s32:
3394 case Intrinsic::nvvm_tex_1d_array_v4u32_f32:
3395 case Intrinsic::nvvm_tex_1d_array_level_v4u32_f32:
3396 case Intrinsic::nvvm_tex_1d_array_grad_v4u32_f32:
3397 case Intrinsic::nvvm_tex_2d_v4u32_s32:
3398 case Intrinsic::nvvm_tex_2d_v4u32_f32:
3399 case Intrinsic::nvvm_tex_2d_level_v4u32_f32:
3400 case Intrinsic::nvvm_tex_2d_grad_v4u32_f32:
3401 case Intrinsic::nvvm_tex_2d_array_v4u32_s32:
3402 case Intrinsic::nvvm_tex_2d_array_v4u32_f32:
3403 case Intrinsic::nvvm_tex_2d_array_level_v4u32_f32:
3404 case Intrinsic::nvvm_tex_2d_array_grad_v4u32_f32:
3405 case Intrinsic::nvvm_tex_3d_v4u32_s32:
3406 case Intrinsic::nvvm_tex_3d_v4u32_f32:
3407 case Intrinsic::nvvm_tex_3d_level_v4u32_f32:
3408 case Intrinsic::nvvm_tex_3d_grad_v4u32_f32:
3409 case Intrinsic::nvvm_tld4_r_2d_v4s32_f32:
3410 case Intrinsic::nvvm_tld4_g_2d_v4s32_f32:
3411 case Intrinsic::nvvm_tld4_b_2d_v4s32_f32:
3412 case Intrinsic::nvvm_tld4_a_2d_v4s32_f32:
3413 case Intrinsic::nvvm_tld4_r_2d_v4u32_f32:
3414 case Intrinsic::nvvm_tld4_g_2d_v4u32_f32:
3415 case Intrinsic::nvvm_tld4_b_2d_v4u32_f32:
3416 case Intrinsic::nvvm_tld4_a_2d_v4u32_f32:
3417 case Intrinsic::nvvm_tex_unified_1d_v4s32_s32:
3418 case Intrinsic::nvvm_tex_unified_1d_v4s32_f32:
3419 case Intrinsic::nvvm_tex_unified_1d_level_v4s32_f32:
3420 case Intrinsic::nvvm_tex_unified_1d_grad_v4s32_f32:
3421 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_s32:
3422 case Intrinsic::nvvm_tex_unified_1d_array_v4s32_f32:
3423 case Intrinsic::nvvm_tex_unified_1d_array_level_v4s32_f32:
3424 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4s32_f32:
3425 case Intrinsic::nvvm_tex_unified_2d_v4s32_s32:
3426 case Intrinsic::nvvm_tex_unified_2d_v4s32_f32:
3427 case Intrinsic::nvvm_tex_unified_2d_level_v4s32_f32:
3428 case Intrinsic::nvvm_tex_unified_2d_grad_v4s32_f32:
3429 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_s32:
3430 case Intrinsic::nvvm_tex_unified_2d_array_v4s32_f32:
3431 case Intrinsic::nvvm_tex_unified_2d_array_level_v4s32_f32:
3432 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4s32_f32:
3433 case Intrinsic::nvvm_tex_unified_3d_v4s32_s32:
3434 case Intrinsic::nvvm_tex_unified_3d_v4s32_f32:
3435 case Intrinsic::nvvm_tex_unified_3d_level_v4s32_f32:
3436 case Intrinsic::nvvm_tex_unified_3d_grad_v4s32_f32:
3437 case Intrinsic::nvvm_tex_unified_1d_v4u32_s32:
3438 case Intrinsic::nvvm_tex_unified_1d_v4u32_f32:
3439 case Intrinsic::nvvm_tex_unified_1d_level_v4u32_f32:
3440 case Intrinsic::nvvm_tex_unified_1d_grad_v4u32_f32:
3441 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_s32:
3442 case Intrinsic::nvvm_tex_unified_1d_array_v4u32_f32:
3443 case Intrinsic::nvvm_tex_unified_1d_array_level_v4u32_f32:
3444 case Intrinsic::nvvm_tex_unified_1d_array_grad_v4u32_f32:
3445 case Intrinsic::nvvm_tex_unified_2d_v4u32_s32:
3446 case Intrinsic::nvvm_tex_unified_2d_v4u32_f32:
3447 case Intrinsic::nvvm_tex_unified_2d_level_v4u32_f32:
3448 case Intrinsic::nvvm_tex_unified_2d_grad_v4u32_f32:
3449 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_s32:
3450 case Intrinsic::nvvm_tex_unified_2d_array_v4u32_f32:
3451 case Intrinsic::nvvm_tex_unified_2d_array_level_v4u32_f32:
3452 case Intrinsic::nvvm_tex_unified_2d_array_grad_v4u32_f32:
3453 case Intrinsic::nvvm_tex_unified_3d_v4u32_s32:
3454 case Intrinsic::nvvm_tex_unified_3d_v4u32_f32:
3455 case Intrinsic::nvvm_tex_unified_3d_level_v4u32_f32:
3456 case Intrinsic::nvvm_tex_unified_3d_grad_v4u32_f32:
3457 case Intrinsic::nvvm_tex_unified_cube_v4s32_f32:
3458 case Intrinsic::nvvm_tex_unified_cube_level_v4s32_f32:
3459 case Intrinsic::nvvm_tex_unified_cube_array_v4s32_f32:
3460 case Intrinsic::nvvm_tex_unified_cube_array_level_v4s32_f32:
3461 case Intrinsic::nvvm_tex_unified_cube_v4u32_f32:
3462 case Intrinsic::nvvm_tex_unified_cube_level_v4u32_f32:
3463 case Intrinsic::nvvm_tex_unified_cube_array_v4u32_f32:
3464 case Intrinsic::nvvm_tex_unified_cube_array_level_v4u32_f32:
3465 case Intrinsic::nvvm_tld4_unified_r_2d_v4s32_f32:
3466 case Intrinsic::nvvm_tld4_unified_g_2d_v4s32_f32:
3467 case Intrinsic::nvvm_tld4_unified_b_2d_v4s32_f32:
3468 case Intrinsic::nvvm_tld4_unified_a_2d_v4s32_f32:
3469 case Intrinsic::nvvm_tld4_unified_r_2d_v4u32_f32:
3470 case Intrinsic::nvvm_tld4_unified_g_2d_v4u32_f32:
3471 case Intrinsic::nvvm_tld4_unified_b_2d_v4u32_f32:
3472 case Intrinsic::nvvm_tld4_unified_a_2d_v4u32_f32: {
3473 Info.opc = getOpcForTextureInstr(Intrinsic);
3474 Info.memVT = MVT::v4i32;
3475 Info.ptrVal = nullptr;
3478 Info.readMem = true;
3479 Info.writeMem = false;
3483 case Intrinsic::nvvm_suld_1d_i8_clamp:
3484 case Intrinsic::nvvm_suld_1d_v2i8_clamp:
3485 case Intrinsic::nvvm_suld_1d_v4i8_clamp:
3486 case Intrinsic::nvvm_suld_1d_array_i8_clamp:
3487 case Intrinsic::nvvm_suld_1d_array_v2i8_clamp:
3488 case Intrinsic::nvvm_suld_1d_array_v4i8_clamp:
3489 case Intrinsic::nvvm_suld_2d_i8_clamp:
3490 case Intrinsic::nvvm_suld_2d_v2i8_clamp:
3491 case Intrinsic::nvvm_suld_2d_v4i8_clamp:
3492 case Intrinsic::nvvm_suld_2d_array_i8_clamp:
3493 case Intrinsic::nvvm_suld_2d_array_v2i8_clamp:
3494 case Intrinsic::nvvm_suld_2d_array_v4i8_clamp:
3495 case Intrinsic::nvvm_suld_3d_i8_clamp:
3496 case Intrinsic::nvvm_suld_3d_v2i8_clamp:
3497 case Intrinsic::nvvm_suld_3d_v4i8_clamp:
3498 case Intrinsic::nvvm_suld_1d_i8_trap:
3499 case Intrinsic::nvvm_suld_1d_v2i8_trap:
3500 case Intrinsic::nvvm_suld_1d_v4i8_trap:
3501 case Intrinsic::nvvm_suld_1d_array_i8_trap:
3502 case Intrinsic::nvvm_suld_1d_array_v2i8_trap:
3503 case Intrinsic::nvvm_suld_1d_array_v4i8_trap:
3504 case Intrinsic::nvvm_suld_2d_i8_trap:
3505 case Intrinsic::nvvm_suld_2d_v2i8_trap:
3506 case Intrinsic::nvvm_suld_2d_v4i8_trap:
3507 case Intrinsic::nvvm_suld_2d_array_i8_trap:
3508 case Intrinsic::nvvm_suld_2d_array_v2i8_trap:
3509 case Intrinsic::nvvm_suld_2d_array_v4i8_trap:
3510 case Intrinsic::nvvm_suld_3d_i8_trap:
3511 case Intrinsic::nvvm_suld_3d_v2i8_trap:
3512 case Intrinsic::nvvm_suld_3d_v4i8_trap:
3513 case Intrinsic::nvvm_suld_1d_i8_zero:
3514 case Intrinsic::nvvm_suld_1d_v2i8_zero:
3515 case Intrinsic::nvvm_suld_1d_v4i8_zero:
3516 case Intrinsic::nvvm_suld_1d_array_i8_zero:
3517 case Intrinsic::nvvm_suld_1d_array_v2i8_zero:
3518 case Intrinsic::nvvm_suld_1d_array_v4i8_zero:
3519 case Intrinsic::nvvm_suld_2d_i8_zero:
3520 case Intrinsic::nvvm_suld_2d_v2i8_zero:
3521 case Intrinsic::nvvm_suld_2d_v4i8_zero:
3522 case Intrinsic::nvvm_suld_2d_array_i8_zero:
3523 case Intrinsic::nvvm_suld_2d_array_v2i8_zero:
3524 case Intrinsic::nvvm_suld_2d_array_v4i8_zero:
3525 case Intrinsic::nvvm_suld_3d_i8_zero:
3526 case Intrinsic::nvvm_suld_3d_v2i8_zero:
3527 case Intrinsic::nvvm_suld_3d_v4i8_zero: {
3528 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3529 Info.memVT = MVT::i8;
3530 Info.ptrVal = nullptr;
3533 Info.readMem = true;
3534 Info.writeMem = false;
3538 case Intrinsic::nvvm_suld_1d_i16_clamp:
3539 case Intrinsic::nvvm_suld_1d_v2i16_clamp:
3540 case Intrinsic::nvvm_suld_1d_v4i16_clamp:
3541 case Intrinsic::nvvm_suld_1d_array_i16_clamp:
3542 case Intrinsic::nvvm_suld_1d_array_v2i16_clamp:
3543 case Intrinsic::nvvm_suld_1d_array_v4i16_clamp:
3544 case Intrinsic::nvvm_suld_2d_i16_clamp:
3545 case Intrinsic::nvvm_suld_2d_v2i16_clamp:
3546 case Intrinsic::nvvm_suld_2d_v4i16_clamp:
3547 case Intrinsic::nvvm_suld_2d_array_i16_clamp:
3548 case Intrinsic::nvvm_suld_2d_array_v2i16_clamp:
3549 case Intrinsic::nvvm_suld_2d_array_v4i16_clamp:
3550 case Intrinsic::nvvm_suld_3d_i16_clamp:
3551 case Intrinsic::nvvm_suld_3d_v2i16_clamp:
3552 case Intrinsic::nvvm_suld_3d_v4i16_clamp:
3553 case Intrinsic::nvvm_suld_1d_i16_trap:
3554 case Intrinsic::nvvm_suld_1d_v2i16_trap:
3555 case Intrinsic::nvvm_suld_1d_v4i16_trap:
3556 case Intrinsic::nvvm_suld_1d_array_i16_trap:
3557 case Intrinsic::nvvm_suld_1d_array_v2i16_trap:
3558 case Intrinsic::nvvm_suld_1d_array_v4i16_trap:
3559 case Intrinsic::nvvm_suld_2d_i16_trap:
3560 case Intrinsic::nvvm_suld_2d_v2i16_trap:
3561 case Intrinsic::nvvm_suld_2d_v4i16_trap:
3562 case Intrinsic::nvvm_suld_2d_array_i16_trap:
3563 case Intrinsic::nvvm_suld_2d_array_v2i16_trap:
3564 case Intrinsic::nvvm_suld_2d_array_v4i16_trap:
3565 case Intrinsic::nvvm_suld_3d_i16_trap:
3566 case Intrinsic::nvvm_suld_3d_v2i16_trap:
3567 case Intrinsic::nvvm_suld_3d_v4i16_trap:
3568 case Intrinsic::nvvm_suld_1d_i16_zero:
3569 case Intrinsic::nvvm_suld_1d_v2i16_zero:
3570 case Intrinsic::nvvm_suld_1d_v4i16_zero:
3571 case Intrinsic::nvvm_suld_1d_array_i16_zero:
3572 case Intrinsic::nvvm_suld_1d_array_v2i16_zero:
3573 case Intrinsic::nvvm_suld_1d_array_v4i16_zero:
3574 case Intrinsic::nvvm_suld_2d_i16_zero:
3575 case Intrinsic::nvvm_suld_2d_v2i16_zero:
3576 case Intrinsic::nvvm_suld_2d_v4i16_zero:
3577 case Intrinsic::nvvm_suld_2d_array_i16_zero:
3578 case Intrinsic::nvvm_suld_2d_array_v2i16_zero:
3579 case Intrinsic::nvvm_suld_2d_array_v4i16_zero:
3580 case Intrinsic::nvvm_suld_3d_i16_zero:
3581 case Intrinsic::nvvm_suld_3d_v2i16_zero:
3582 case Intrinsic::nvvm_suld_3d_v4i16_zero: {
3583 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3584 Info.memVT = MVT::i16;
3585 Info.ptrVal = nullptr;
3588 Info.readMem = true;
3589 Info.writeMem = false;
3593 case Intrinsic::nvvm_suld_1d_i32_clamp:
3594 case Intrinsic::nvvm_suld_1d_v2i32_clamp:
3595 case Intrinsic::nvvm_suld_1d_v4i32_clamp:
3596 case Intrinsic::nvvm_suld_1d_array_i32_clamp:
3597 case Intrinsic::nvvm_suld_1d_array_v2i32_clamp:
3598 case Intrinsic::nvvm_suld_1d_array_v4i32_clamp:
3599 case Intrinsic::nvvm_suld_2d_i32_clamp:
3600 case Intrinsic::nvvm_suld_2d_v2i32_clamp:
3601 case Intrinsic::nvvm_suld_2d_v4i32_clamp:
3602 case Intrinsic::nvvm_suld_2d_array_i32_clamp:
3603 case Intrinsic::nvvm_suld_2d_array_v2i32_clamp:
3604 case Intrinsic::nvvm_suld_2d_array_v4i32_clamp:
3605 case Intrinsic::nvvm_suld_3d_i32_clamp:
3606 case Intrinsic::nvvm_suld_3d_v2i32_clamp:
3607 case Intrinsic::nvvm_suld_3d_v4i32_clamp:
3608 case Intrinsic::nvvm_suld_1d_i32_trap:
3609 case Intrinsic::nvvm_suld_1d_v2i32_trap:
3610 case Intrinsic::nvvm_suld_1d_v4i32_trap:
3611 case Intrinsic::nvvm_suld_1d_array_i32_trap:
3612 case Intrinsic::nvvm_suld_1d_array_v2i32_trap:
3613 case Intrinsic::nvvm_suld_1d_array_v4i32_trap:
3614 case Intrinsic::nvvm_suld_2d_i32_trap:
3615 case Intrinsic::nvvm_suld_2d_v2i32_trap:
3616 case Intrinsic::nvvm_suld_2d_v4i32_trap:
3617 case Intrinsic::nvvm_suld_2d_array_i32_trap:
3618 case Intrinsic::nvvm_suld_2d_array_v2i32_trap:
3619 case Intrinsic::nvvm_suld_2d_array_v4i32_trap:
3620 case Intrinsic::nvvm_suld_3d_i32_trap:
3621 case Intrinsic::nvvm_suld_3d_v2i32_trap:
3622 case Intrinsic::nvvm_suld_3d_v4i32_trap:
3623 case Intrinsic::nvvm_suld_1d_i32_zero:
3624 case Intrinsic::nvvm_suld_1d_v2i32_zero:
3625 case Intrinsic::nvvm_suld_1d_v4i32_zero:
3626 case Intrinsic::nvvm_suld_1d_array_i32_zero:
3627 case Intrinsic::nvvm_suld_1d_array_v2i32_zero:
3628 case Intrinsic::nvvm_suld_1d_array_v4i32_zero:
3629 case Intrinsic::nvvm_suld_2d_i32_zero:
3630 case Intrinsic::nvvm_suld_2d_v2i32_zero:
3631 case Intrinsic::nvvm_suld_2d_v4i32_zero:
3632 case Intrinsic::nvvm_suld_2d_array_i32_zero:
3633 case Intrinsic::nvvm_suld_2d_array_v2i32_zero:
3634 case Intrinsic::nvvm_suld_2d_array_v4i32_zero:
3635 case Intrinsic::nvvm_suld_3d_i32_zero:
3636 case Intrinsic::nvvm_suld_3d_v2i32_zero:
3637 case Intrinsic::nvvm_suld_3d_v4i32_zero: {
3638 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3639 Info.memVT = MVT::i32;
3640 Info.ptrVal = nullptr;
3643 Info.readMem = true;
3644 Info.writeMem = false;
3648 case Intrinsic::nvvm_suld_1d_i64_clamp:
3649 case Intrinsic::nvvm_suld_1d_v2i64_clamp:
3650 case Intrinsic::nvvm_suld_1d_array_i64_clamp:
3651 case Intrinsic::nvvm_suld_1d_array_v2i64_clamp:
3652 case Intrinsic::nvvm_suld_2d_i64_clamp:
3653 case Intrinsic::nvvm_suld_2d_v2i64_clamp:
3654 case Intrinsic::nvvm_suld_2d_array_i64_clamp:
3655 case Intrinsic::nvvm_suld_2d_array_v2i64_clamp:
3656 case Intrinsic::nvvm_suld_3d_i64_clamp:
3657 case Intrinsic::nvvm_suld_3d_v2i64_clamp:
3658 case Intrinsic::nvvm_suld_1d_i64_trap:
3659 case Intrinsic::nvvm_suld_1d_v2i64_trap:
3660 case Intrinsic::nvvm_suld_1d_array_i64_trap:
3661 case Intrinsic::nvvm_suld_1d_array_v2i64_trap:
3662 case Intrinsic::nvvm_suld_2d_i64_trap:
3663 case Intrinsic::nvvm_suld_2d_v2i64_trap:
3664 case Intrinsic::nvvm_suld_2d_array_i64_trap:
3665 case Intrinsic::nvvm_suld_2d_array_v2i64_trap:
3666 case Intrinsic::nvvm_suld_3d_i64_trap:
3667 case Intrinsic::nvvm_suld_3d_v2i64_trap:
3668 case Intrinsic::nvvm_suld_1d_i64_zero:
3669 case Intrinsic::nvvm_suld_1d_v2i64_zero:
3670 case Intrinsic::nvvm_suld_1d_array_i64_zero:
3671 case Intrinsic::nvvm_suld_1d_array_v2i64_zero:
3672 case Intrinsic::nvvm_suld_2d_i64_zero:
3673 case Intrinsic::nvvm_suld_2d_v2i64_zero:
3674 case Intrinsic::nvvm_suld_2d_array_i64_zero:
3675 case Intrinsic::nvvm_suld_2d_array_v2i64_zero:
3676 case Intrinsic::nvvm_suld_3d_i64_zero:
3677 case Intrinsic::nvvm_suld_3d_v2i64_zero: {
3678 Info.opc = getOpcForSurfaceInstr(Intrinsic);
3679 Info.memVT = MVT::i64;
3680 Info.ptrVal = nullptr;
3683 Info.readMem = true;
3684 Info.writeMem = false;
3692 /// isLegalAddressingMode - Return true if the addressing mode represented
3693 /// by AM is legal for this target, for a load/store of the specified type.
3694 /// Used to guide target specific optimizations, like loop strength reduction
3695 /// (LoopStrengthReduce.cpp) and memory optimization for address mode
3696 /// (CodeGenPrepare.cpp)
3697 bool NVPTXTargetLowering::isLegalAddressingMode(const AddrMode &AM,
3700 // AddrMode - This represents an addressing mode of:
3701 // BaseGV + BaseOffs + BaseReg + Scale*ScaleReg
3703 // The legal address modes are
3710 if (AM.BaseOffs || AM.HasBaseReg || AM.Scale)
3716 case 0: // "r", "r+i" or "i" is allowed
3719 if (AM.HasBaseReg) // "r+r+i" or "r+r" is not allowed.
3721 // Otherwise we have r+i.
3724 // No scale > 1 is allowed
3730 //===----------------------------------------------------------------------===//
3731 // NVPTX Inline Assembly Support
3732 //===----------------------------------------------------------------------===//
3734 /// getConstraintType - Given a constraint letter, return the type of
3735 /// constraint it is for this target.
3736 NVPTXTargetLowering::ConstraintType
3737 NVPTXTargetLowering::getConstraintType(const std::string &Constraint) const {
3738 if (Constraint.size() == 1) {
3739 switch (Constraint[0]) {
3751 return C_RegisterClass;
3754 return TargetLowering::getConstraintType(Constraint);
3757 std::pair<unsigned, const TargetRegisterClass *>
3758 NVPTXTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
3760 if (Constraint.size() == 1) {
3761 switch (Constraint[0]) {
3763 return std::make_pair(0U, &NVPTX::Int1RegsRegClass);
3765 return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
3767 return std::make_pair(0U, &NVPTX::Int16RegsRegClass);
3769 return std::make_pair(0U, &NVPTX::Int32RegsRegClass);
3772 return std::make_pair(0U, &NVPTX::Int64RegsRegClass);
3774 return std::make_pair(0U, &NVPTX::Float32RegsRegClass);
3776 return std::make_pair(0U, &NVPTX::Float64RegsRegClass);
3779 return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
3782 /// getFunctionAlignment - Return the Log2 alignment of this function.
3783 unsigned NVPTXTargetLowering::getFunctionAlignment(const Function *) const {
3787 //===----------------------------------------------------------------------===//
3788 // NVPTX DAG Combining
3789 //===----------------------------------------------------------------------===//
3791 bool NVPTXTargetLowering::allowFMA(MachineFunction &MF,
3792 CodeGenOpt::Level OptLevel) const {
3793 const Function *F = MF.getFunction();
3794 const TargetOptions &TO = MF.getTarget().Options;
3796 // Always honor command-line argument
3797 if (FMAContractLevelOpt.getNumOccurrences() > 0) {
3798 return FMAContractLevelOpt > 0;
3799 } else if (OptLevel == 0) {
3800 // Do not contract if we're not optimizing the code
3802 } else if (TO.AllowFPOpFusion == FPOpFusion::Fast || TO.UnsafeFPMath) {
3803 // Honor TargetOptions flags that explicitly say fusion is okay
3805 } else if (F->hasFnAttribute("unsafe-fp-math")) {
3806 // Check for unsafe-fp-math=true coming from Clang
3807 Attribute Attr = F->getFnAttribute("unsafe-fp-math");
3808 StringRef Val = Attr.getValueAsString();
3813 // We did not have a clear indication that fusion is allowed, so assume not
3817 /// PerformADDCombineWithOperands - Try DAG combinations for an ADD with
3818 /// operands N0 and N1. This is a helper for PerformADDCombine that is
3819 /// called with the default operands, and if that fails, with commuted
3821 static SDValue PerformADDCombineWithOperands(SDNode *N, SDValue N0, SDValue N1,
3822 TargetLowering::DAGCombinerInfo &DCI,
3823 const NVPTXSubtarget &Subtarget,
3824 CodeGenOpt::Level OptLevel) {
3825 SelectionDAG &DAG = DCI.DAG;
3826 // Skip non-integer, non-scalar case
3827 EVT VT=N0.getValueType();
3831 // fold (add (mul a, b), c) -> (mad a, b, c)
3833 if (N0.getOpcode() == ISD::MUL) {
3834 assert (VT.isInteger());
3836 // Since integer multiply-add costs the same as integer multiply
3837 // but is more costly than integer add, do the fusion only when
3838 // the mul is only used in the add.
3839 if (OptLevel==CodeGenOpt::None || VT != MVT::i32 ||
3840 !N0.getNode()->hasOneUse())
3844 return DAG.getNode(NVPTXISD::IMAD, SDLoc(N), VT,
3845 N0.getOperand(0), N0.getOperand(1), N1);
3847 else if (N0.getOpcode() == ISD::FMUL) {
3848 if (VT == MVT::f32 || VT == MVT::f64) {
3849 const auto *TLI = static_cast<const NVPTXTargetLowering *>(
3850 &DAG.getTargetLoweringInfo());
3851 if (!TLI->allowFMA(DAG.getMachineFunction(), OptLevel))
3854 // For floating point:
3855 // Do the fusion only when the mul has less than 5 uses and all
3857 // The heuristic is that if a use is not an add, then that use
3858 // cannot be fused into fma, therefore mul is still needed anyway.
3859 // If there are more than 4 uses, even if they are all add, fusing
3860 // them will increase register pressue.
3863 int nonAddCount = 0;
3864 for (SDNode::use_iterator UI = N0.getNode()->use_begin(),
3865 UE = N0.getNode()->use_end();
3869 if (User->getOpcode() != ISD::FADD)
3875 int orderNo = N->getIROrder();
3876 int orderNo2 = N0.getNode()->getIROrder();
3877 // simple heuristics here for considering potential register
3878 // pressure, the logics here is that the differnce are used
3879 // to measure the distance between def and use, the longer distance
3880 // more likely cause register pressure.
3881 if (orderNo - orderNo2 < 500)
3884 // Now, check if at least one of the FMUL's operands is live beyond the node N,
3885 // which guarantees that the FMA will not increase register pressure at node N.
3886 bool opIsLive = false;
3887 const SDNode *left = N0.getOperand(0).getNode();
3888 const SDNode *right = N0.getOperand(1).getNode();
3890 if (dyn_cast<ConstantSDNode>(left) || dyn_cast<ConstantSDNode>(right))
3894 for (SDNode::use_iterator UI = left->use_begin(), UE = left->use_end(); UI != UE; ++UI) {
3896 int orderNo3 = User->getIROrder();
3897 if (orderNo3 > orderNo) {
3904 for (SDNode::use_iterator UI = right->use_begin(), UE = right->use_end(); UI != UE; ++UI) {
3906 int orderNo3 = User->getIROrder();
3907 if (orderNo3 > orderNo) {
3917 return DAG.getNode(ISD::FMA, SDLoc(N), VT,
3918 N0.getOperand(0), N0.getOperand(1), N1);
3925 /// PerformADDCombine - Target-specific dag combine xforms for ISD::ADD.
3927 static SDValue PerformADDCombine(SDNode *N,
3928 TargetLowering::DAGCombinerInfo &DCI,
3929 const NVPTXSubtarget &Subtarget,
3930 CodeGenOpt::Level OptLevel) {
3931 SDValue N0 = N->getOperand(0);
3932 SDValue N1 = N->getOperand(1);
3934 // First try with the default operand order.
3935 SDValue Result = PerformADDCombineWithOperands(N, N0, N1, DCI, Subtarget,
3937 if (Result.getNode())
3940 // If that didn't work, try again with the operands commuted.
3941 return PerformADDCombineWithOperands(N, N1, N0, DCI, Subtarget, OptLevel);
3944 static SDValue PerformANDCombine(SDNode *N,
3945 TargetLowering::DAGCombinerInfo &DCI) {
3946 // The type legalizer turns a vector load of i8 values into a zextload to i16
3947 // registers, optionally ANY_EXTENDs it (if target type is integer),
3948 // and ANDs off the high 8 bits. Since we turn this load into a
3949 // target-specific DAG node, the DAG combiner fails to eliminate these AND
3950 // nodes. Do that here.
3951 SDValue Val = N->getOperand(0);
3952 SDValue Mask = N->getOperand(1);
3954 if (isa<ConstantSDNode>(Val)) {
3955 std::swap(Val, Mask);
3959 // Generally, we will see zextload -> IMOV16rr -> ANY_EXTEND -> and
3960 if (Val.getOpcode() == ISD::ANY_EXTEND) {
3962 Val = Val->getOperand(0);
3965 if (Val->isMachineOpcode() && Val->getMachineOpcode() == NVPTX::IMOV16rr) {
3966 Val = Val->getOperand(0);
3969 if (Val->getOpcode() == NVPTXISD::LoadV2 ||
3970 Val->getOpcode() == NVPTXISD::LoadV4) {
3971 ConstantSDNode *MaskCnst = dyn_cast<ConstantSDNode>(Mask);
3973 // Not an AND with a constant
3977 uint64_t MaskVal = MaskCnst->getZExtValue();
3978 if (MaskVal != 0xff) {
3979 // Not an AND that chops off top 8 bits
3983 MemSDNode *Mem = dyn_cast<MemSDNode>(Val);
3985 // Not a MemSDNode?!?
3989 EVT MemVT = Mem->getMemoryVT();
3990 if (MemVT != MVT::v2i8 && MemVT != MVT::v4i8) {
3991 // We only handle the i8 case
3996 cast<ConstantSDNode>(Val->getOperand(Val->getNumOperands()-1))->
3998 if (ExtType == ISD::SEXTLOAD) {
3999 // If for some reason the load is a sextload, the and is needed to zero
4000 // out the high 8 bits
4005 if (AExt.getNode() != 0) {
4006 // Re-insert the ext as a zext.
4007 Val = DCI.DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N),
4008 AExt.getValueType(), Val);
4012 // If we get here, the AND is unnecessary. Just replace it with the load
4013 DCI.CombineTo(N, Val, AddTo);
4019 enum OperandSignedness {
4025 /// IsMulWideOperandDemotable - Checks if the provided DAG node is an operand
4026 /// that can be demoted to \p OptSize bits without loss of information. The
4027 /// signedness of the operand, if determinable, is placed in \p S.
4028 static bool IsMulWideOperandDemotable(SDValue Op,
4030 OperandSignedness &S) {
4033 if (Op.getOpcode() == ISD::SIGN_EXTEND ||
4034 Op.getOpcode() == ISD::SIGN_EXTEND_INREG) {
4035 EVT OrigVT = Op.getOperand(0).getValueType();
4036 if (OrigVT.getSizeInBits() <= OptSize) {
4040 } else if (Op.getOpcode() == ISD::ZERO_EXTEND) {
4041 EVT OrigVT = Op.getOperand(0).getValueType();
4042 if (OrigVT.getSizeInBits() <= OptSize) {
4051 /// AreMulWideOperandsDemotable - Checks if the given LHS and RHS operands can
4052 /// be demoted to \p OptSize bits without loss of information. If the operands
4053 /// contain a constant, it should appear as the RHS operand. The signedness of
4054 /// the operands is placed in \p IsSigned.
4055 static bool AreMulWideOperandsDemotable(SDValue LHS, SDValue RHS,
4059 OperandSignedness LHSSign;
4061 // The LHS operand must be a demotable op
4062 if (!IsMulWideOperandDemotable(LHS, OptSize, LHSSign))
4065 // We should have been able to determine the signedness from the LHS
4066 if (LHSSign == Unknown)
4069 IsSigned = (LHSSign == Signed);
4071 // The RHS can be a demotable op or a constant
4072 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(RHS)) {
4073 APInt Val = CI->getAPIntValue();
4074 if (LHSSign == Unsigned) {
4075 if (Val.isIntN(OptSize)) {
4080 if (Val.isSignedIntN(OptSize)) {
4086 OperandSignedness RHSSign;
4087 if (!IsMulWideOperandDemotable(RHS, OptSize, RHSSign))
4090 if (LHSSign != RHSSign)
4097 /// TryMULWIDECombine - Attempt to replace a multiply of M bits with a multiply
4098 /// of M/2 bits that produces an M-bit result (i.e. mul.wide). This transform
4099 /// works on both multiply DAG nodes and SHL DAG nodes with a constant shift
4101 static SDValue TryMULWIDECombine(SDNode *N,
4102 TargetLowering::DAGCombinerInfo &DCI) {
4103 EVT MulType = N->getValueType(0);
4104 if (MulType != MVT::i32 && MulType != MVT::i64) {
4108 unsigned OptSize = MulType.getSizeInBits() >> 1;
4109 SDValue LHS = N->getOperand(0);
4110 SDValue RHS = N->getOperand(1);
4112 // Canonicalize the multiply so the constant (if any) is on the right
4113 if (N->getOpcode() == ISD::MUL) {
4114 if (isa<ConstantSDNode>(LHS)) {
4115 std::swap(LHS, RHS);
4119 // If we have a SHL, determine the actual multiply amount
4120 if (N->getOpcode() == ISD::SHL) {
4121 ConstantSDNode *ShlRHS = dyn_cast<ConstantSDNode>(RHS);
4126 APInt ShiftAmt = ShlRHS->getAPIntValue();
4127 unsigned BitWidth = MulType.getSizeInBits();
4128 if (ShiftAmt.sge(0) && ShiftAmt.slt(BitWidth)) {
4129 APInt MulVal = APInt(BitWidth, 1) << ShiftAmt;
4130 RHS = DCI.DAG.getConstant(MulVal, MulType);
4137 // Verify that our operands are demotable
4138 if (!AreMulWideOperandsDemotable(LHS, RHS, OptSize, Signed)) {
4143 if (MulType == MVT::i32) {
4144 DemotedVT = MVT::i16;
4146 DemotedVT = MVT::i32;
4149 // Truncate the operands to the correct size. Note that these are just for
4150 // type consistency and will (likely) be eliminated in later phases.
4152 DCI.DAG.getNode(ISD::TRUNCATE, SDLoc(N), DemotedVT, LHS);
4154 DCI.DAG.getNode(ISD::TRUNCATE, SDLoc(N), DemotedVT, RHS);
4158 Opc = NVPTXISD::MUL_WIDE_SIGNED;
4160 Opc = NVPTXISD::MUL_WIDE_UNSIGNED;
4163 return DCI.DAG.getNode(Opc, SDLoc(N), MulType, TruncLHS, TruncRHS);
4166 /// PerformMULCombine - Runs PTX-specific DAG combine patterns on MUL nodes.
4167 static SDValue PerformMULCombine(SDNode *N,
4168 TargetLowering::DAGCombinerInfo &DCI,
4169 CodeGenOpt::Level OptLevel) {
4171 // Try mul.wide combining at OptLevel > 0
4172 SDValue Ret = TryMULWIDECombine(N, DCI);
4180 /// PerformSHLCombine - Runs PTX-specific DAG combine patterns on SHL nodes.
4181 static SDValue PerformSHLCombine(SDNode *N,
4182 TargetLowering::DAGCombinerInfo &DCI,
4183 CodeGenOpt::Level OptLevel) {
4185 // Try mul.wide combining at OptLevel > 0
4186 SDValue Ret = TryMULWIDECombine(N, DCI);
4194 SDValue NVPTXTargetLowering::PerformDAGCombine(SDNode *N,
4195 DAGCombinerInfo &DCI) const {
4196 CodeGenOpt::Level OptLevel = getTargetMachine().getOptLevel();
4197 switch (N->getOpcode()) {
4201 return PerformADDCombine(N, DCI, nvptxSubtarget, OptLevel);
4203 return PerformMULCombine(N, DCI, OptLevel);
4205 return PerformSHLCombine(N, DCI, OptLevel);
4207 return PerformANDCombine(N, DCI);
4212 /// ReplaceVectorLoad - Convert vector loads into multi-output scalar loads.
4213 static void ReplaceLoadVector(SDNode *N, SelectionDAG &DAG,
4214 const DataLayout *TD,
4215 SmallVectorImpl<SDValue> &Results) {
4216 EVT ResVT = N->getValueType(0);
4219 assert(ResVT.isVector() && "Vector load must have vector type");
4221 // We only handle "native" vector sizes for now, e.g. <4 x double> is not
4222 // legal. We can (and should) split that into 2 loads of <2 x double> here
4223 // but I'm leaving that as a TODO for now.
4224 assert(ResVT.isSimple() && "Can only handle simple types");
4225 switch (ResVT.getSimpleVT().SimpleTy) {
4238 // This is a "native" vector type
4242 LoadSDNode *LD = cast<LoadSDNode>(N);
4244 unsigned Align = LD->getAlignment();
4245 unsigned PrefAlign =
4246 TD->getPrefTypeAlignment(ResVT.getTypeForEVT(*DAG.getContext()));
4247 if (Align < PrefAlign) {
4248 // This load is not sufficiently aligned, so bail out and let this vector
4249 // load be scalarized. Note that we may still be able to emit smaller
4250 // vector loads. For example, if we are loading a <4 x float> with an
4251 // alignment of 8, this check will fail but the legalizer will try again
4252 // with 2 x <2 x float>, which will succeed with an alignment of 8.
4256 EVT EltVT = ResVT.getVectorElementType();
4257 unsigned NumElts = ResVT.getVectorNumElements();
4259 // Since LoadV2 is a target node, we cannot rely on DAG type legalization.
4260 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
4261 // loaded type to i16 and propagate the "real" type as the memory type.
4262 bool NeedTrunc = false;
4263 if (EltVT.getSizeInBits() < 16) {
4268 unsigned Opcode = 0;
4275 Opcode = NVPTXISD::LoadV2;
4276 LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4279 Opcode = NVPTXISD::LoadV4;
4280 EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4281 LdResVTs = DAG.getVTList(ListVTs);
4286 SmallVector<SDValue, 8> OtherOps;
4288 // Copy regular operands
4289 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4290 OtherOps.push_back(N->getOperand(i));
4292 // The select routine does not have access to the LoadSDNode instance, so
4293 // pass along the extension information
4294 OtherOps.push_back(DAG.getIntPtrConstant(LD->getExtensionType()));
4296 SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4298 LD->getMemOperand());
4300 SmallVector<SDValue, 4> ScalarRes;
4302 for (unsigned i = 0; i < NumElts; ++i) {
4303 SDValue Res = NewLD.getValue(i);
4305 Res = DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4306 ScalarRes.push_back(Res);
4309 SDValue LoadChain = NewLD.getValue(NumElts);
4311 SDValue BuildVec = DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes);
4313 Results.push_back(BuildVec);
4314 Results.push_back(LoadChain);
4317 static void ReplaceINTRINSIC_W_CHAIN(SDNode *N, SelectionDAG &DAG,
4318 SmallVectorImpl<SDValue> &Results) {
4319 SDValue Chain = N->getOperand(0);
4320 SDValue Intrin = N->getOperand(1);
4323 // Get the intrinsic ID
4324 unsigned IntrinNo = cast<ConstantSDNode>(Intrin.getNode())->getZExtValue();
4328 case Intrinsic::nvvm_ldg_global_i:
4329 case Intrinsic::nvvm_ldg_global_f:
4330 case Intrinsic::nvvm_ldg_global_p:
4331 case Intrinsic::nvvm_ldu_global_i:
4332 case Intrinsic::nvvm_ldu_global_f:
4333 case Intrinsic::nvvm_ldu_global_p: {
4334 EVT ResVT = N->getValueType(0);
4336 if (ResVT.isVector()) {
4339 unsigned NumElts = ResVT.getVectorNumElements();
4340 EVT EltVT = ResVT.getVectorElementType();
4342 // Since LDU/LDG are target nodes, we cannot rely on DAG type
4344 // Therefore, we must ensure the type is legal. For i1 and i8, we set the
4345 // loaded type to i16 and propagate the "real" type as the memory type.
4346 bool NeedTrunc = false;
4347 if (EltVT.getSizeInBits() < 16) {
4352 unsigned Opcode = 0;
4362 case Intrinsic::nvvm_ldg_global_i:
4363 case Intrinsic::nvvm_ldg_global_f:
4364 case Intrinsic::nvvm_ldg_global_p:
4365 Opcode = NVPTXISD::LDGV2;
4367 case Intrinsic::nvvm_ldu_global_i:
4368 case Intrinsic::nvvm_ldu_global_f:
4369 case Intrinsic::nvvm_ldu_global_p:
4370 Opcode = NVPTXISD::LDUV2;
4373 LdResVTs = DAG.getVTList(EltVT, EltVT, MVT::Other);
4379 case Intrinsic::nvvm_ldg_global_i:
4380 case Intrinsic::nvvm_ldg_global_f:
4381 case Intrinsic::nvvm_ldg_global_p:
4382 Opcode = NVPTXISD::LDGV4;
4384 case Intrinsic::nvvm_ldu_global_i:
4385 case Intrinsic::nvvm_ldu_global_f:
4386 case Intrinsic::nvvm_ldu_global_p:
4387 Opcode = NVPTXISD::LDUV4;
4390 EVT ListVTs[] = { EltVT, EltVT, EltVT, EltVT, MVT::Other };
4391 LdResVTs = DAG.getVTList(ListVTs);
4396 SmallVector<SDValue, 8> OtherOps;
4398 // Copy regular operands
4400 OtherOps.push_back(Chain); // Chain
4401 // Skip operand 1 (intrinsic ID)
4403 for (unsigned i = 2, e = N->getNumOperands(); i != e; ++i)
4404 OtherOps.push_back(N->getOperand(i));
4406 MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4408 SDValue NewLD = DAG.getMemIntrinsicNode(Opcode, DL, LdResVTs, OtherOps,
4409 MemSD->getMemoryVT(),
4410 MemSD->getMemOperand());
4412 SmallVector<SDValue, 4> ScalarRes;
4414 for (unsigned i = 0; i < NumElts; ++i) {
4415 SDValue Res = NewLD.getValue(i);
4418 DAG.getNode(ISD::TRUNCATE, DL, ResVT.getVectorElementType(), Res);
4419 ScalarRes.push_back(Res);
4422 SDValue LoadChain = NewLD.getValue(NumElts);
4425 DAG.getNode(ISD::BUILD_VECTOR, DL, ResVT, ScalarRes);
4427 Results.push_back(BuildVec);
4428 Results.push_back(LoadChain);
4431 assert(ResVT.isSimple() && ResVT.getSimpleVT().SimpleTy == MVT::i8 &&
4432 "Custom handling of non-i8 ldu/ldg?");
4434 // Just copy all operands as-is
4435 SmallVector<SDValue, 4> Ops;
4436 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
4437 Ops.push_back(N->getOperand(i));
4439 // Force output to i16
4440 SDVTList LdResVTs = DAG.getVTList(MVT::i16, MVT::Other);
4442 MemIntrinsicSDNode *MemSD = cast<MemIntrinsicSDNode>(N);
4444 // We make sure the memory type is i8, which will be used during isel
4445 // to select the proper instruction.
4447 DAG.getMemIntrinsicNode(ISD::INTRINSIC_W_CHAIN, DL, LdResVTs, Ops,
4448 MVT::i8, MemSD->getMemOperand());
4450 Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
4451 NewLD.getValue(0)));
4452 Results.push_back(NewLD.getValue(1));
4458 void NVPTXTargetLowering::ReplaceNodeResults(
4459 SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
4460 switch (N->getOpcode()) {
4462 report_fatal_error("Unhandled custom legalization");
4464 ReplaceLoadVector(N, DAG, getDataLayout(), Results);
4466 case ISD::INTRINSIC_W_CHAIN:
4467 ReplaceINTRINSIC_W_CHAIN(N, DAG, Results);
4472 // Pin NVPTXSection's and NVPTXTargetObjectFile's vtables to this file.
4473 void NVPTXSection::anchor() {}
4475 NVPTXTargetObjectFile::~NVPTXTargetObjectFile() {
4479 delete ReadOnlySection;
4481 delete StaticCtorSection;
4482 delete StaticDtorSection;
4484 delete EHFrameSection;
4485 delete DwarfAbbrevSection;
4486 delete DwarfInfoSection;
4487 delete DwarfLineSection;
4488 delete DwarfFrameSection;
4489 delete DwarfPubTypesSection;
4490 delete DwarfDebugInlineSection;
4491 delete DwarfStrSection;
4492 delete DwarfLocSection;
4493 delete DwarfARangesSection;
4494 delete DwarfRangesSection;
4495 delete DwarfMacroInfoSection;
4499 NVPTXTargetObjectFile::SelectSectionForGlobal(const GlobalValue *GV,
4500 SectionKind Kind, Mangler &Mang,
4501 const TargetMachine &TM) const {
4502 return getDataSection();