1 //===-- AMDGPUISelLowering.cpp - AMDGPU Common DAG lowering functions -----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 /// \brief This is the parent TargetLowering class for hardware code gen
14 //===----------------------------------------------------------------------===//
16 #include "AMDGPUISelLowering.h"
18 #include "AMDGPUFrameLowering.h"
19 #include "AMDGPURegisterInfo.h"
20 #include "AMDGPUSubtarget.h"
21 #include "AMDILIntrinsicInfo.h"
22 #include "R600MachineFunctionInfo.h"
23 #include "SIMachineFunctionInfo.h"
24 #include "llvm/Analysis/ValueTracking.h"
25 #include "llvm/CodeGen/CallingConvLower.h"
26 #include "llvm/CodeGen/MachineFunction.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/SelectionDAG.h"
29 #include "llvm/CodeGen/TargetLoweringObjectFileImpl.h"
30 #include "llvm/IR/DataLayout.h"
33 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
34 CCValAssign::LocInfo LocInfo,
35 ISD::ArgFlagsTy ArgFlags, CCState &State) {
36 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
37 ArgFlags.getOrigAlign());
38 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
43 #include "AMDGPUGenCallingConv.inc"
45 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
46 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
48 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
50 // Initialize target lowering borrowed from AMDIL
53 // We need to custom lower some of the intrinsics
54 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
56 // Library functions. These default to Expand, but we have instructions
58 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
59 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
60 setOperationAction(ISD::FPOW, MVT::f32, Legal);
61 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
62 setOperationAction(ISD::FABS, MVT::f32, Legal);
63 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
64 setOperationAction(ISD::FRINT, MVT::f32, Legal);
65 setOperationAction(ISD::FROUND, MVT::f32, Legal);
66 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
68 // The hardware supports ROTR, but not ROTL
69 setOperationAction(ISD::ROTL, MVT::i32, Expand);
71 // Lower floating point store/load to integer store/load to reduce the number
72 // of patterns in tablegen.
73 setOperationAction(ISD::STORE, MVT::f32, Promote);
74 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
76 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
77 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
79 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
80 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
82 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
83 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
85 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
86 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
88 setOperationAction(ISD::STORE, MVT::f64, Promote);
89 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
91 // Custom lowering of vector stores is required for local address space
93 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
94 // XXX: Native v2i32 local address space stores are possible, but not
95 // currently implemented.
96 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
98 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
99 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
100 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
102 // XXX: This can be change to Custom, once ExpandVectorStores can
103 // handle 64-bit stores.
104 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
106 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
107 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
108 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
111 setOperationAction(ISD::LOAD, MVT::f32, Promote);
112 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
114 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
115 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
117 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
118 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
120 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
121 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
123 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
124 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
126 setOperationAction(ISD::LOAD, MVT::f64, Promote);
127 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
129 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
130 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
131 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
132 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
133 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
134 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
135 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
136 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
137 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
138 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
140 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
141 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
142 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
143 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
144 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
145 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
146 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
147 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
148 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
149 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
150 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
151 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
153 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
155 setOperationAction(ISD::FNEG, MVT::v2f32, Expand);
156 setOperationAction(ISD::FNEG, MVT::v4f32, Expand);
158 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
160 setOperationAction(ISD::MUL, MVT::i64, Expand);
162 setOperationAction(ISD::UDIV, MVT::i32, Expand);
163 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
164 setOperationAction(ISD::UREM, MVT::i32, Expand);
165 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
166 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
168 static const MVT::SimpleValueType IntTypes[] = {
169 MVT::v2i32, MVT::v4i32
171 const size_t NumIntTypes = array_lengthof(IntTypes);
173 for (unsigned int x = 0; x < NumIntTypes; ++x) {
174 MVT::SimpleValueType VT = IntTypes[x];
175 //Expand the following operations for the current type by default
176 setOperationAction(ISD::ADD, VT, Expand);
177 setOperationAction(ISD::AND, VT, Expand);
178 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
179 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
180 setOperationAction(ISD::MUL, VT, Expand);
181 setOperationAction(ISD::OR, VT, Expand);
182 setOperationAction(ISD::SHL, VT, Expand);
183 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
184 setOperationAction(ISD::SRL, VT, Expand);
185 setOperationAction(ISD::SRA, VT, Expand);
186 setOperationAction(ISD::SUB, VT, Expand);
187 setOperationAction(ISD::UDIV, VT, Expand);
188 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
189 setOperationAction(ISD::UREM, VT, Expand);
190 setOperationAction(ISD::SELECT, VT, Expand);
191 setOperationAction(ISD::VSELECT, VT, Expand);
192 setOperationAction(ISD::XOR, VT, Expand);
195 static const MVT::SimpleValueType FloatTypes[] = {
196 MVT::v2f32, MVT::v4f32
198 const size_t NumFloatTypes = array_lengthof(FloatTypes);
200 for (unsigned int x = 0; x < NumFloatTypes; ++x) {
201 MVT::SimpleValueType VT = FloatTypes[x];
202 setOperationAction(ISD::FABS, VT, Expand);
203 setOperationAction(ISD::FADD, VT, Expand);
204 setOperationAction(ISD::FDIV, VT, Expand);
205 setOperationAction(ISD::FPOW, VT, Expand);
206 setOperationAction(ISD::FFLOOR, VT, Expand);
207 setOperationAction(ISD::FTRUNC, VT, Expand);
208 setOperationAction(ISD::FMUL, VT, Expand);
209 setOperationAction(ISD::FRINT, VT, Expand);
210 setOperationAction(ISD::FSQRT, VT, Expand);
211 setOperationAction(ISD::FSUB, VT, Expand);
212 setOperationAction(ISD::SELECT, VT, Expand);
215 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Custom);
216 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i1, Custom);
217 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i1, Custom);
219 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8, Custom);
220 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i8, Custom);
221 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i8, Custom);
223 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Custom);
224 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v2i16, Custom);
225 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::v4i16, Custom);
227 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Custom);
229 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::Other, Custom);
232 //===----------------------------------------------------------------------===//
233 // Target Information
234 //===----------------------------------------------------------------------===//
236 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
240 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
242 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
245 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
246 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
248 return ((LScalarSize <= CastScalarSize) ||
249 (CastScalarSize >= 32) ||
253 //===---------------------------------------------------------------------===//
255 //===---------------------------------------------------------------------===//
257 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
258 assert(VT.isFloatingPoint());
259 return VT == MVT::f32;
262 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
263 assert(VT.isFloatingPoint());
264 return VT == MVT::f32;
267 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
268 // Truncate is just accessing a subregister.
269 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
272 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
273 // Truncate is just accessing a subregister.
274 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
275 (Dest->getPrimitiveSizeInBits() % 32 == 0);
278 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
279 const DataLayout *DL = getDataLayout();
280 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
281 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
283 return SrcSize == 32 && DestSize == 64;
286 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
287 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
288 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
289 // this will enable reducing 64-bit operations the 32-bit, which is always
291 return Src == MVT::i32 && Dest == MVT::i64;
294 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
295 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
296 // limited number of native 64-bit operations. Shrinking an operation to fit
297 // in a single 32-bit register should always be helpful. As currently used,
298 // this is much less general than the name suggests, and is only used in
299 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
300 // not profitable, and may actually be harmful.
301 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
304 //===---------------------------------------------------------------------===//
305 // TargetLowering Callbacks
306 //===---------------------------------------------------------------------===//
308 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
309 const SmallVectorImpl<ISD::InputArg> &Ins) const {
311 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
314 SDValue AMDGPUTargetLowering::LowerReturn(
316 CallingConv::ID CallConv,
318 const SmallVectorImpl<ISD::OutputArg> &Outs,
319 const SmallVectorImpl<SDValue> &OutVals,
320 SDLoc DL, SelectionDAG &DAG) const {
321 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
324 //===---------------------------------------------------------------------===//
325 // Target specific lowering
326 //===---------------------------------------------------------------------===//
328 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
330 switch (Op.getOpcode()) {
332 Op.getNode()->dump();
333 llvm_unreachable("Custom lowering code for this"
334 "instruction is not implemented yet!");
336 // AMDIL DAG lowering
337 case ISD::SDIV: return LowerSDIV(Op, DAG);
338 case ISD::SREM: return LowerSREM(Op, DAG);
339 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
340 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
341 // AMDGPU DAG lowering
342 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
343 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
344 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
345 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
346 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
347 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
352 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
353 SmallVectorImpl<SDValue> &Results,
354 SelectionDAG &DAG) const {
355 switch (N->getOpcode()) {
356 case ISD::SIGN_EXTEND_INREG:
357 // Different parts of legalization seem to interpret which type of
358 // sign_extend_inreg is the one to check for custom lowering. The extended
359 // from type is what really matters, but some places check for custom
360 // lowering of the result type. This results in trying to use
361 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
362 // nothing here and let the illegal result integer be handled normally.
370 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
371 const GlobalValue *GV,
372 const SDValue &InitPtr,
374 SelectionDAG &DAG) const {
375 const DataLayout *TD = getTargetMachine().getDataLayout();
377 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
378 EVT VT = EVT::getEVT(CI->getType());
379 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
380 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
381 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
382 TD->getPrefTypeAlignment(CI->getType()));
383 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
384 EVT VT = EVT::getEVT(CFP->getType());
385 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
386 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
387 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
388 TD->getPrefTypeAlignment(CFP->getType()));
389 } else if (Init->getType()->isAggregateType()) {
390 EVT PtrVT = InitPtr.getValueType();
391 unsigned NumElements = Init->getType()->getArrayNumElements();
392 SmallVector<SDValue, 8> Chains;
393 for (unsigned i = 0; i < NumElements; ++i) {
394 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
395 Init->getType()->getArrayElementType()), PtrVT);
396 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
397 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
398 GV, Ptr, Chain, DAG));
400 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, &Chains[0],
404 llvm_unreachable("Unhandled constant initializer");
408 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
410 SelectionDAG &DAG) const {
412 const DataLayout *TD = getTargetMachine().getDataLayout();
413 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
414 const GlobalValue *GV = G->getGlobal();
416 switch (G->getAddressSpace()) {
417 default: llvm_unreachable("Global Address lowering not implemented for this "
419 case AMDGPUAS::LOCAL_ADDRESS: {
420 // XXX: What does the value of G->getOffset() mean?
421 assert(G->getOffset() == 0 &&
422 "Do not know what to do with an non-zero offset");
425 if (MFI->LocalMemoryObjects.count(GV) == 0) {
426 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
427 Offset = MFI->LDSSize;
428 MFI->LocalMemoryObjects[GV] = Offset;
429 // XXX: Account for alignment?
430 MFI->LDSSize += Size;
432 Offset = MFI->LocalMemoryObjects[GV];
435 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
437 case AMDGPUAS::CONSTANT_ADDRESS: {
438 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
439 Type *EltType = GV->getType()->getElementType();
440 unsigned Size = TD->getTypeAllocSize(EltType);
441 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
443 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
444 const Constant *Init = Var->getInitializer();
445 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
446 SDValue InitPtr = DAG.getFrameIndex(FI,
447 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
448 SmallVector<SDNode*, 8> WorkList;
450 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
451 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
452 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
454 WorkList.push_back(*I);
456 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
457 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
458 E = WorkList.end(); I != E; ++I) {
459 SmallVector<SDValue, 8> Ops;
460 Ops.push_back(Chain);
461 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
462 Ops.push_back((*I)->getOperand(i));
464 DAG.UpdateNodeOperands(*I, &Ops[0], Ops.size());
466 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
467 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
472 void AMDGPUTargetLowering::ExtractVectorElements(SDValue Op, SelectionDAG &DAG,
473 SmallVectorImpl<SDValue> &Args,
475 unsigned Count) const {
476 EVT VT = Op.getValueType();
477 for (unsigned i = Start, e = Start + Count; i != e; ++i) {
478 Args.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SDLoc(Op),
479 VT.getVectorElementType(),
480 Op, DAG.getConstant(i, MVT::i32)));
484 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
485 SelectionDAG &DAG) const {
486 SmallVector<SDValue, 8> Args;
487 SDValue A = Op.getOperand(0);
488 SDValue B = Op.getOperand(1);
490 ExtractVectorElements(A, DAG, Args, 0,
491 A.getValueType().getVectorNumElements());
492 ExtractVectorElements(B, DAG, Args, 0,
493 B.getValueType().getVectorNumElements());
495 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(),
496 &Args[0], Args.size());
499 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
500 SelectionDAG &DAG) const {
502 SmallVector<SDValue, 8> Args;
503 EVT VT = Op.getValueType();
504 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
505 ExtractVectorElements(Op.getOperand(0), DAG, Args, Start,
506 VT.getVectorNumElements());
508 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(),
509 &Args[0], Args.size());
512 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
513 SelectionDAG &DAG) const {
515 MachineFunction &MF = DAG.getMachineFunction();
516 const AMDGPUFrameLowering *TFL =
517 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
519 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
522 unsigned FrameIndex = FIN->getIndex();
523 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
524 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
528 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
529 SelectionDAG &DAG) const {
530 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
532 EVT VT = Op.getValueType();
534 switch (IntrinsicID) {
536 case AMDGPUIntrinsic::AMDIL_abs:
537 return LowerIntrinsicIABS(Op, DAG);
538 case AMDGPUIntrinsic::AMDIL_exp:
539 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
540 case AMDGPUIntrinsic::AMDGPU_lrp:
541 return LowerIntrinsicLRP(Op, DAG);
542 case AMDGPUIntrinsic::AMDIL_fraction:
543 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
544 case AMDGPUIntrinsic::AMDIL_max:
545 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
547 case AMDGPUIntrinsic::AMDGPU_imax:
548 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
550 case AMDGPUIntrinsic::AMDGPU_umax:
551 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
553 case AMDGPUIntrinsic::AMDIL_min:
554 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
556 case AMDGPUIntrinsic::AMDGPU_imin:
557 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
559 case AMDGPUIntrinsic::AMDGPU_umin:
560 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
562 case AMDGPUIntrinsic::AMDIL_round_nearest:
563 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
567 ///IABS(a) = SMAX(sub(0, a), a)
568 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
569 SelectionDAG &DAG) const {
572 EVT VT = Op.getValueType();
573 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
576 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
579 /// Linear Interpolation
580 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
581 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
582 SelectionDAG &DAG) const {
584 EVT VT = Op.getValueType();
585 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
586 DAG.getConstantFP(1.0f, MVT::f32),
588 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
590 return DAG.getNode(ISD::FADD, DL, VT,
591 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
595 /// \brief Generate Min/Max node
596 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
597 SelectionDAG &DAG) const {
599 EVT VT = Op.getValueType();
601 SDValue LHS = Op.getOperand(0);
602 SDValue RHS = Op.getOperand(1);
603 SDValue True = Op.getOperand(2);
604 SDValue False = Op.getOperand(3);
605 SDValue CC = Op.getOperand(4);
607 if (VT != MVT::f32 ||
608 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
612 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
626 llvm_unreachable("Operation should already be optimised!");
634 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
636 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
645 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
647 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
649 case ISD::SETCC_INVALID:
650 llvm_unreachable("Invalid setcc condcode!");
655 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
656 SelectionDAG &DAG) const {
657 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
658 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
659 EVT EltVT = Op.getValueType().getVectorElementType();
660 EVT PtrVT = Load->getBasePtr().getValueType();
661 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
662 SmallVector<SDValue, 8> Loads;
665 for (unsigned i = 0, e = NumElts; i != e; ++i) {
666 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
667 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
668 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
669 Load->getChain(), Ptr,
670 MachinePointerInfo(Load->getMemOperand()->getValue()),
671 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
672 Load->getAlignment()));
674 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(),
675 Loads.data(), Loads.size());
678 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
679 SelectionDAG &DAG) const {
680 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
681 EVT MemVT = Store->getMemoryVT();
682 unsigned MemBits = MemVT.getSizeInBits();
684 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
685 // truncating store into an i32 store.
686 // XXX: We could also handle optimize other vector bitwidths.
687 if (!MemVT.isVector() || MemBits > 32) {
692 const SDValue &Value = Store->getValue();
693 EVT VT = Value.getValueType();
694 const SDValue &Ptr = Store->getBasePtr();
695 EVT MemEltVT = MemVT.getVectorElementType();
696 unsigned MemEltBits = MemEltVT.getSizeInBits();
697 unsigned MemNumElements = MemVT.getVectorNumElements();
698 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), MemVT.getSizeInBits());
699 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, PackedVT);
702 for (unsigned i = 0; i < MemNumElements; ++i) {
703 EVT ElemVT = VT.getVectorElementType();
704 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
705 DAG.getConstant(i, MVT::i32));
706 Elt = DAG.getZExtOrTrunc(Elt, DL, PackedVT);
707 Elt = DAG.getNode(ISD::AND, DL, PackedVT, Elt, Mask);
708 SDValue Shift = DAG.getConstant(MemEltBits * i, PackedVT);
709 Elt = DAG.getNode(ISD::SHL, DL, PackedVT, Elt, Shift);
713 PackedValue = DAG.getNode(ISD::OR, DL, PackedVT, PackedValue, Elt);
716 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
717 MachinePointerInfo(Store->getMemOperand()->getValue()),
718 Store->isVolatile(), Store->isNonTemporal(),
719 Store->getAlignment());
722 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
723 SelectionDAG &DAG) const {
724 StoreSDNode *Store = cast<StoreSDNode>(Op);
725 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
726 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
727 EVT PtrVT = Store->getBasePtr().getValueType();
728 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
731 SmallVector<SDValue, 8> Chains;
733 for (unsigned i = 0, e = NumElts; i != e; ++i) {
734 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
735 Store->getValue(), DAG.getConstant(i, MVT::i32));
736 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
738 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
740 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
741 MachinePointerInfo(Store->getMemOperand()->getValue()),
742 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
743 Store->getAlignment()));
745 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, &Chains[0], NumElts);
748 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
750 LoadSDNode *Load = cast<LoadSDNode>(Op);
751 ISD::LoadExtType ExtType = Load->getExtensionType();
752 EVT VT = Op.getValueType();
753 EVT MemVT = Load->getMemoryVT();
755 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
756 // We can do the extload to 32-bits, and then need to separately extend to
759 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
763 Load->getMemOperand());
764 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
767 // Lower loads constant address space global variable loads
768 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
769 isa<GlobalVariable>(GetUnderlyingObject(Load->getPointerInfo().V))) {
771 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
772 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
773 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
774 DAG.getConstant(2, MVT::i32));
775 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
776 Load->getChain(), Ptr,
777 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
780 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
781 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
785 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
786 DAG.getConstant(2, MVT::i32));
787 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
788 Load->getChain(), Ptr,
789 DAG.getTargetConstant(0, MVT::i32),
791 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
793 DAG.getConstant(0x3, MVT::i32));
794 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
795 DAG.getConstant(3, MVT::i32));
797 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
799 EVT MemEltVT = MemVT.getScalarType();
800 if (ExtType == ISD::SEXTLOAD) {
801 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
802 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
805 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
808 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
810 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
811 if (Result.getNode()) {
815 StoreSDNode *Store = cast<StoreSDNode>(Op);
816 SDValue Chain = Store->getChain();
817 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
818 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
819 Store->getValue().getValueType().isVector()) {
820 return SplitVectorStore(Op, DAG);
823 EVT MemVT = Store->getMemoryVT();
824 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
825 MemVT.bitsLT(MVT::i32)) {
827 if (Store->getMemoryVT() == MVT::i8) {
829 } else if (Store->getMemoryVT() == MVT::i16) {
832 SDValue BasePtr = Store->getBasePtr();
833 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
834 DAG.getConstant(2, MVT::i32));
835 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
836 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
838 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
839 DAG.getConstant(0x3, MVT::i32));
841 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
842 DAG.getConstant(3, MVT::i32));
844 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
847 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
849 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
850 MaskedValue, ShiftAmt);
852 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
854 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
855 DAG.getConstant(0xffffffff, MVT::i32));
856 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
858 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
859 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
860 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
865 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
866 SelectionDAG &DAG) const {
868 EVT VT = Op.getValueType();
870 SDValue Num = Op.getOperand(0);
871 SDValue Den = Op.getOperand(1);
873 SmallVector<SDValue, 8> Results;
875 // RCP = URECIP(Den) = 2^32 / Den + e
876 // e is rounding error.
877 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
879 // RCP_LO = umulo(RCP, Den) */
880 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
882 // RCP_HI = mulhu (RCP, Den) */
883 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
885 // NEG_RCP_LO = -RCP_LO
886 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
889 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
890 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
893 // Calculate the rounding error from the URECIP instruction
894 // E = mulhu(ABS_RCP_LO, RCP)
895 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
898 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
901 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
903 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
904 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
907 // Quotient = mulhu(Tmp0, Num)
908 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
910 // Num_S_Remainder = Quotient * Den
911 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
913 // Remainder = Num - Num_S_Remainder
914 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
916 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
917 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
918 DAG.getConstant(-1, VT),
919 DAG.getConstant(0, VT),
921 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
922 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
924 DAG.getConstant(-1, VT),
925 DAG.getConstant(0, VT),
927 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
928 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
931 // Calculate Division result:
933 // Quotient_A_One = Quotient + 1
934 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
935 DAG.getConstant(1, VT));
937 // Quotient_S_One = Quotient - 1
938 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
939 DAG.getConstant(1, VT));
941 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
942 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
943 Quotient, Quotient_A_One, ISD::SETEQ);
945 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
946 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
947 Quotient_S_One, Div, ISD::SETEQ);
949 // Calculate Rem result:
951 // Remainder_S_Den = Remainder - Den
952 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
954 // Remainder_A_Den = Remainder + Den
955 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
957 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
958 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
959 Remainder, Remainder_S_Den, ISD::SETEQ);
961 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
962 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
963 Remainder_A_Den, Rem, ISD::SETEQ);
967 return DAG.getMergeValues(Ops, 2, DL);
970 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
971 SelectionDAG &DAG) const {
972 SDValue S0 = Op.getOperand(0);
974 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
977 // f32 uint_to_fp i64
978 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
979 DAG.getConstant(0, MVT::i32));
980 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
981 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
982 DAG.getConstant(1, MVT::i32));
983 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
984 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
985 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
986 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
990 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
992 SelectionDAG &DAG) const {
993 MVT VT = Op.getSimpleValueType();
995 SDValue Shift = DAG.getConstant(BitsDiff, VT);
996 // Shift left by 'Shift' bits.
997 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
998 // Signed shift Right by 'Shift' bits.
999 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1002 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1003 SelectionDAG &DAG) const {
1004 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1005 MVT VT = Op.getSimpleValueType();
1006 MVT ScalarVT = VT.getScalarType();
1008 unsigned SrcBits = ExtraVT.getScalarType().getSizeInBits();
1009 unsigned DestBits = ScalarVT.getSizeInBits();
1010 unsigned BitsDiff = DestBits - SrcBits;
1012 if (!Subtarget->hasBFE())
1013 return ExpandSIGN_EXTEND_INREG(Op, BitsDiff, DAG);
1015 SDValue Src = Op.getOperand(0);
1016 if (VT.isVector()) {
1018 // Need to scalarize this, and revisit each of the scalars later.
1019 // TODO: Don't scalarize on Evergreen?
1020 unsigned NElts = VT.getVectorNumElements();
1021 SmallVector<SDValue, 8> Args;
1022 ExtractVectorElements(Src, DAG, Args, 0, NElts);
1024 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1025 for (unsigned I = 0; I < NElts; ++I)
1026 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1028 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args.data(), Args.size());
1031 if (SrcBits == 32) {
1034 // If the source is 32-bits, this is really half of a 2-register pair, and
1035 // we need to discard the unused half of the pair.
1036 SDValue TruncSrc = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Src);
1037 return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, TruncSrc);
1040 unsigned NElts = VT.isVector() ? VT.getVectorNumElements() : 1;
1042 // TODO: Match 64-bit BFE. SI has a 64-bit BFE, but it's scalar only so it
1043 // might not be worth the effort, and will need to expand to shifts when
1044 // fixing SGPR copies.
1045 if (SrcBits < 32 && DestBits <= 32) {
1047 MVT ExtVT = (NElts == 1) ? MVT::i32 : MVT::getVectorVT(MVT::i32, NElts);
1050 Src = DAG.getNode(ISD::ZERO_EXTEND, DL, ExtVT, Src);
1052 // FIXME: This should use TargetConstant, but that hits assertions for
1054 SDValue Ext = DAG.getNode(AMDGPUISD::BFE_I32, DL, ExtVT,
1055 Op.getOperand(0), // Operand
1056 DAG.getConstant(0, ExtVT), // Offset
1057 DAG.getConstant(SrcBits, ExtVT)); // Width
1059 // Truncate to the original type if necessary.
1060 if (ScalarVT == MVT::i32)
1062 return DAG.getNode(ISD::TRUNCATE, DL, VT, Ext);
1065 // For small types, extend to 32-bits first.
1068 MVT ExtVT = (NElts == 1) ? MVT::i32 : MVT::getVectorVT(MVT::i32, NElts);
1070 SDValue TruncSrc = DAG.getNode(ISD::TRUNCATE, DL, ExtVT, Src);
1071 SDValue Ext32 = DAG.getNode(AMDGPUISD::BFE_I32,
1074 TruncSrc, // Operand
1075 DAG.getConstant(0, ExtVT), // Offset
1076 DAG.getConstant(SrcBits, ExtVT)); // Width
1078 return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Ext32);
1081 // For everything else, use the standard bitshift expansion.
1082 return ExpandSIGN_EXTEND_INREG(Op, BitsDiff, DAG);
1085 //===----------------------------------------------------------------------===//
1087 //===----------------------------------------------------------------------===//
1089 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1092 const SmallVectorImpl<ISD::InputArg> &Ins,
1093 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1095 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1096 if (Ins[i].ArgVT == Ins[i].VT) {
1097 OrigIns.push_back(Ins[i]);
1102 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1103 // Vector has been split into scalars.
1104 VT = Ins[i].ArgVT.getVectorElementType();
1105 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1106 Ins[i].ArgVT.getVectorElementType() !=
1107 Ins[i].VT.getVectorElementType()) {
1108 // Vector elements have been promoted
1111 // Vector has been spilt into smaller vectors.
1115 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1116 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1117 OrigIns.push_back(Arg);
1121 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1122 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1123 return CFP->isExactlyValue(1.0);
1125 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1126 return C->isAllOnesValue();
1131 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1132 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1133 return CFP->getValueAPF().isZero();
1135 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1136 return C->isNullValue();
1141 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1142 const TargetRegisterClass *RC,
1143 unsigned Reg, EVT VT) const {
1144 MachineFunction &MF = DAG.getMachineFunction();
1145 MachineRegisterInfo &MRI = MF.getRegInfo();
1146 unsigned VirtualRegister;
1147 if (!MRI.isLiveIn(Reg)) {
1148 VirtualRegister = MRI.createVirtualRegister(RC);
1149 MRI.addLiveIn(Reg, VirtualRegister);
1151 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1153 return DAG.getRegister(VirtualRegister, VT);
1156 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1158 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1162 NODE_NAME_CASE(CALL);
1163 NODE_NAME_CASE(UMUL);
1164 NODE_NAME_CASE(DIV_INF);
1165 NODE_NAME_CASE(RET_FLAG);
1166 NODE_NAME_CASE(BRANCH_COND);
1169 NODE_NAME_CASE(DWORDADDR)
1170 NODE_NAME_CASE(FRACT)
1171 NODE_NAME_CASE(FMAX)
1172 NODE_NAME_CASE(SMAX)
1173 NODE_NAME_CASE(UMAX)
1174 NODE_NAME_CASE(FMIN)
1175 NODE_NAME_CASE(SMIN)
1176 NODE_NAME_CASE(UMIN)
1177 NODE_NAME_CASE(BFE_U32)
1178 NODE_NAME_CASE(BFE_I32)
1181 NODE_NAME_CASE(URECIP)
1182 NODE_NAME_CASE(DOT4)
1183 NODE_NAME_CASE(EXPORT)
1184 NODE_NAME_CASE(CONST_ADDRESS)
1185 NODE_NAME_CASE(REGISTER_LOAD)
1186 NODE_NAME_CASE(REGISTER_STORE)
1187 NODE_NAME_CASE(LOAD_CONSTANT)
1188 NODE_NAME_CASE(LOAD_INPUT)
1189 NODE_NAME_CASE(SAMPLE)
1190 NODE_NAME_CASE(SAMPLEB)
1191 NODE_NAME_CASE(SAMPLED)
1192 NODE_NAME_CASE(SAMPLEL)
1193 NODE_NAME_CASE(STORE_MSKOR)
1194 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1198 void AMDGPUTargetLowering::computeMaskedBitsForTargetNode(
1202 const SelectionDAG &DAG,
1203 unsigned Depth) const {
1204 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.