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"
31 #include "llvm/IR/DiagnosticInfo.h"
32 #include "llvm/IR/DiagnosticPrinter.h"
38 /// Diagnostic information for unimplemented or unsupported feature reporting.
39 class DiagnosticInfoUnsupported : public DiagnosticInfo {
41 const Twine &Description;
46 static int getKindID() {
48 KindID = llvm::getNextAvailablePluginDiagnosticKind();
53 DiagnosticInfoUnsupported(const Function &Fn, const Twine &Desc,
54 DiagnosticSeverity Severity = DS_Error)
55 : DiagnosticInfo(getKindID(), Severity),
59 const Function &getFunction() const { return Fn; }
60 const Twine &getDescription() const { return Description; }
62 void print(DiagnosticPrinter &DP) const override {
63 DP << "unsupported " << getDescription() << " in " << Fn.getName();
66 static bool classof(const DiagnosticInfo *DI) {
67 return DI->getKind() == getKindID();
71 int DiagnosticInfoUnsupported::KindID = 0;
75 static bool allocateStack(unsigned ValNo, MVT ValVT, MVT LocVT,
76 CCValAssign::LocInfo LocInfo,
77 ISD::ArgFlagsTy ArgFlags, CCState &State) {
78 unsigned Offset = State.AllocateStack(ValVT.getStoreSize(),
79 ArgFlags.getOrigAlign());
80 State.addLoc(CCValAssign::getMem(ValNo, ValVT, Offset, LocVT, LocInfo));
85 #include "AMDGPUGenCallingConv.inc"
87 AMDGPUTargetLowering::AMDGPUTargetLowering(TargetMachine &TM) :
88 TargetLowering(TM, new TargetLoweringObjectFileELF()) {
90 Subtarget = &TM.getSubtarget<AMDGPUSubtarget>();
92 // Initialize target lowering borrowed from AMDIL
95 // We need to custom lower some of the intrinsics
96 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
98 // Library functions. These default to Expand, but we have instructions
100 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
101 setOperationAction(ISD::FEXP2, MVT::f32, Legal);
102 setOperationAction(ISD::FPOW, MVT::f32, Legal);
103 setOperationAction(ISD::FLOG2, MVT::f32, Legal);
104 setOperationAction(ISD::FABS, MVT::f32, Legal);
105 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
106 setOperationAction(ISD::FRINT, MVT::f32, Legal);
107 setOperationAction(ISD::FROUND, MVT::f32, Legal);
108 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
110 // The hardware supports ROTR, but not ROTL
111 setOperationAction(ISD::ROTL, MVT::i32, Expand);
113 // Lower floating point store/load to integer store/load to reduce the number
114 // of patterns in tablegen.
115 setOperationAction(ISD::STORE, MVT::f32, Promote);
116 AddPromotedToType(ISD::STORE, MVT::f32, MVT::i32);
118 setOperationAction(ISD::STORE, MVT::v2f32, Promote);
119 AddPromotedToType(ISD::STORE, MVT::v2f32, MVT::v2i32);
121 setOperationAction(ISD::STORE, MVT::v4f32, Promote);
122 AddPromotedToType(ISD::STORE, MVT::v4f32, MVT::v4i32);
124 setOperationAction(ISD::STORE, MVT::v8f32, Promote);
125 AddPromotedToType(ISD::STORE, MVT::v8f32, MVT::v8i32);
127 setOperationAction(ISD::STORE, MVT::v16f32, Promote);
128 AddPromotedToType(ISD::STORE, MVT::v16f32, MVT::v16i32);
130 setOperationAction(ISD::STORE, MVT::f64, Promote);
131 AddPromotedToType(ISD::STORE, MVT::f64, MVT::i64);
133 // Custom lowering of vector stores is required for local address space
135 setOperationAction(ISD::STORE, MVT::v4i32, Custom);
136 // XXX: Native v2i32 local address space stores are possible, but not
137 // currently implemented.
138 setOperationAction(ISD::STORE, MVT::v2i32, Custom);
140 setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);
141 setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
142 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Custom);
144 // XXX: This can be change to Custom, once ExpandVectorStores can
145 // handle 64-bit stores.
146 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Expand);
148 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
149 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
150 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
153 setOperationAction(ISD::LOAD, MVT::f32, Promote);
154 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
156 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
157 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
159 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
160 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
162 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
163 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
165 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
166 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
168 setOperationAction(ISD::LOAD, MVT::f64, Promote);
169 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
171 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
172 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
173 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
174 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
175 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
176 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
177 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
178 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
179 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
180 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
182 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
183 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
184 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
185 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
186 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
187 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
188 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
189 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
190 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
191 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
192 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
193 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
195 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
197 setOperationAction(ISD::FNEG, MVT::v2f32, Expand);
198 setOperationAction(ISD::FNEG, MVT::v4f32, Expand);
200 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
202 setOperationAction(ISD::MUL, MVT::i64, Expand);
204 setOperationAction(ISD::UDIV, MVT::i32, Expand);
205 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
206 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
207 setOperationAction(ISD::UREM, MVT::i32, Expand);
208 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
209 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
211 static const MVT::SimpleValueType IntTypes[] = {
212 MVT::v2i32, MVT::v4i32
214 const size_t NumIntTypes = array_lengthof(IntTypes);
216 for (unsigned int x = 0; x < NumIntTypes; ++x) {
217 MVT::SimpleValueType VT = IntTypes[x];
218 //Expand the following operations for the current type by default
219 setOperationAction(ISD::ADD, VT, Expand);
220 setOperationAction(ISD::AND, VT, Expand);
221 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
222 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
223 setOperationAction(ISD::MUL, VT, Expand);
224 setOperationAction(ISD::OR, VT, Expand);
225 setOperationAction(ISD::SHL, VT, Expand);
226 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
227 setOperationAction(ISD::SRL, VT, Expand);
228 setOperationAction(ISD::SRA, VT, Expand);
229 setOperationAction(ISD::SUB, VT, Expand);
230 setOperationAction(ISD::UDIV, VT, Expand);
231 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
232 setOperationAction(ISD::UREM, VT, Expand);
233 setOperationAction(ISD::SELECT, VT, Expand);
234 setOperationAction(ISD::VSELECT, VT, Expand);
235 setOperationAction(ISD::XOR, VT, Expand);
238 static const MVT::SimpleValueType FloatTypes[] = {
239 MVT::v2f32, MVT::v4f32
241 const size_t NumFloatTypes = array_lengthof(FloatTypes);
243 for (unsigned int x = 0; x < NumFloatTypes; ++x) {
244 MVT::SimpleValueType VT = FloatTypes[x];
245 setOperationAction(ISD::FABS, VT, Expand);
246 setOperationAction(ISD::FADD, VT, Expand);
247 setOperationAction(ISD::FDIV, VT, Expand);
248 setOperationAction(ISD::FPOW, VT, Expand);
249 setOperationAction(ISD::FFLOOR, VT, Expand);
250 setOperationAction(ISD::FTRUNC, VT, Expand);
251 setOperationAction(ISD::FMUL, VT, Expand);
252 setOperationAction(ISD::FRINT, VT, Expand);
253 setOperationAction(ISD::FSQRT, VT, Expand);
254 setOperationAction(ISD::FSUB, VT, Expand);
255 setOperationAction(ISD::SELECT, VT, Expand);
258 setTargetDAGCombine(ISD::MUL);
261 //===----------------------------------------------------------------------===//
262 // Target Information
263 //===----------------------------------------------------------------------===//
265 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
269 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
271 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
274 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
275 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
277 return ((LScalarSize <= CastScalarSize) ||
278 (CastScalarSize >= 32) ||
282 //===---------------------------------------------------------------------===//
284 //===---------------------------------------------------------------------===//
286 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
287 assert(VT.isFloatingPoint());
288 return VT == MVT::f32;
291 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
292 assert(VT.isFloatingPoint());
293 return VT == MVT::f32;
296 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
297 // Truncate is just accessing a subregister.
298 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
301 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
302 // Truncate is just accessing a subregister.
303 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
304 (Dest->getPrimitiveSizeInBits() % 32 == 0);
307 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
308 const DataLayout *DL = getDataLayout();
309 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
310 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
312 return SrcSize == 32 && DestSize == 64;
315 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
316 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
317 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
318 // this will enable reducing 64-bit operations the 32-bit, which is always
320 return Src == MVT::i32 && Dest == MVT::i64;
323 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
324 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
325 // limited number of native 64-bit operations. Shrinking an operation to fit
326 // in a single 32-bit register should always be helpful. As currently used,
327 // this is much less general than the name suggests, and is only used in
328 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
329 // not profitable, and may actually be harmful.
330 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
333 //===---------------------------------------------------------------------===//
334 // TargetLowering Callbacks
335 //===---------------------------------------------------------------------===//
337 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
338 const SmallVectorImpl<ISD::InputArg> &Ins) const {
340 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
343 SDValue AMDGPUTargetLowering::LowerReturn(
345 CallingConv::ID CallConv,
347 const SmallVectorImpl<ISD::OutputArg> &Outs,
348 const SmallVectorImpl<SDValue> &OutVals,
349 SDLoc DL, SelectionDAG &DAG) const {
350 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
353 //===---------------------------------------------------------------------===//
354 // Target specific lowering
355 //===---------------------------------------------------------------------===//
357 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
358 SmallVectorImpl<SDValue> &InVals) const {
359 SDValue Callee = CLI.Callee;
360 SelectionDAG &DAG = CLI.DAG;
362 const Function &Fn = *DAG.getMachineFunction().getFunction();
364 StringRef FuncName("<unknown>");
366 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
367 FuncName = G->getSymbol();
368 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
369 FuncName = G->getGlobal()->getName();
371 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
372 DAG.getContext()->diagnose(NoCalls);
376 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
378 switch (Op.getOpcode()) {
380 Op.getNode()->dump();
381 llvm_unreachable("Custom lowering code for this"
382 "instruction is not implemented yet!");
384 // AMDIL DAG lowering
385 case ISD::SDIV: return LowerSDIV(Op, DAG);
386 case ISD::SREM: return LowerSREM(Op, DAG);
387 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
388 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
389 // AMDGPU DAG lowering
390 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
391 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
392 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
393 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
394 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
395 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
400 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
401 SmallVectorImpl<SDValue> &Results,
402 SelectionDAG &DAG) const {
403 switch (N->getOpcode()) {
404 case ISD::SIGN_EXTEND_INREG:
405 // Different parts of legalization seem to interpret which type of
406 // sign_extend_inreg is the one to check for custom lowering. The extended
407 // from type is what really matters, but some places check for custom
408 // lowering of the result type. This results in trying to use
409 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
410 // nothing here and let the illegal result integer be handled normally.
413 SDValue Op = SDValue(N, 0);
415 EVT VT = Op.getValueType();
416 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
417 N->getOperand(0), N->getOperand(1));
418 Results.push_back(UDIVREM);
422 SDValue Op = SDValue(N, 0);
424 EVT VT = Op.getValueType();
425 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
426 N->getOperand(0), N->getOperand(1));
427 Results.push_back(UDIVREM.getValue(1));
435 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
436 const GlobalValue *GV,
437 const SDValue &InitPtr,
439 SelectionDAG &DAG) const {
440 const DataLayout *TD = getTargetMachine().getDataLayout();
442 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
443 EVT VT = EVT::getEVT(CI->getType());
444 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
445 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
446 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
447 TD->getPrefTypeAlignment(CI->getType()));
448 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
449 EVT VT = EVT::getEVT(CFP->getType());
450 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
451 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
452 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
453 TD->getPrefTypeAlignment(CFP->getType()));
454 } else if (Init->getType()->isAggregateType()) {
455 EVT PtrVT = InitPtr.getValueType();
456 unsigned NumElements = Init->getType()->getArrayNumElements();
457 SmallVector<SDValue, 8> Chains;
458 for (unsigned i = 0; i < NumElements; ++i) {
459 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
460 Init->getType()->getArrayElementType()), PtrVT);
461 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
462 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
463 GV, Ptr, Chain, DAG));
465 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
468 llvm_unreachable("Unhandled constant initializer");
472 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
474 SelectionDAG &DAG) const {
476 const DataLayout *TD = getTargetMachine().getDataLayout();
477 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
478 const GlobalValue *GV = G->getGlobal();
480 switch (G->getAddressSpace()) {
481 default: llvm_unreachable("Global Address lowering not implemented for this "
483 case AMDGPUAS::LOCAL_ADDRESS: {
484 // XXX: What does the value of G->getOffset() mean?
485 assert(G->getOffset() == 0 &&
486 "Do not know what to do with an non-zero offset");
489 if (MFI->LocalMemoryObjects.count(GV) == 0) {
490 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
491 Offset = MFI->LDSSize;
492 MFI->LocalMemoryObjects[GV] = Offset;
493 // XXX: Account for alignment?
494 MFI->LDSSize += Size;
496 Offset = MFI->LocalMemoryObjects[GV];
499 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
501 case AMDGPUAS::CONSTANT_ADDRESS: {
502 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
503 Type *EltType = GV->getType()->getElementType();
504 unsigned Size = TD->getTypeAllocSize(EltType);
505 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
507 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
508 const Constant *Init = Var->getInitializer();
509 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
510 SDValue InitPtr = DAG.getFrameIndex(FI,
511 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
512 SmallVector<SDNode*, 8> WorkList;
514 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
515 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
516 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
518 WorkList.push_back(*I);
520 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
521 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
522 E = WorkList.end(); I != E; ++I) {
523 SmallVector<SDValue, 8> Ops;
524 Ops.push_back(Chain);
525 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
526 Ops.push_back((*I)->getOperand(i));
528 DAG.UpdateNodeOperands(*I, Ops);
530 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
531 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
536 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
537 SelectionDAG &DAG) const {
538 SmallVector<SDValue, 8> Args;
539 SDValue A = Op.getOperand(0);
540 SDValue B = Op.getOperand(1);
542 DAG.ExtractVectorElements(A, Args);
543 DAG.ExtractVectorElements(B, Args);
545 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
548 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
549 SelectionDAG &DAG) const {
551 SmallVector<SDValue, 8> Args;
552 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
553 EVT VT = Op.getValueType();
554 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
555 VT.getVectorNumElements());
557 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
560 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
561 SelectionDAG &DAG) const {
563 MachineFunction &MF = DAG.getMachineFunction();
564 const AMDGPUFrameLowering *TFL =
565 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
567 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
570 unsigned FrameIndex = FIN->getIndex();
571 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
572 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
576 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
577 SelectionDAG &DAG) const {
578 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
580 EVT VT = Op.getValueType();
582 switch (IntrinsicID) {
584 case AMDGPUIntrinsic::AMDIL_abs:
585 return LowerIntrinsicIABS(Op, DAG);
586 case AMDGPUIntrinsic::AMDIL_exp:
587 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
588 case AMDGPUIntrinsic::AMDGPU_lrp:
589 return LowerIntrinsicLRP(Op, DAG);
590 case AMDGPUIntrinsic::AMDIL_fraction:
591 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
592 case AMDGPUIntrinsic::AMDIL_max:
593 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
595 case AMDGPUIntrinsic::AMDGPU_imax:
596 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
598 case AMDGPUIntrinsic::AMDGPU_umax:
599 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
601 case AMDGPUIntrinsic::AMDIL_min:
602 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
604 case AMDGPUIntrinsic::AMDGPU_imin:
605 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
607 case AMDGPUIntrinsic::AMDGPU_umin:
608 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
611 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
612 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
617 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
618 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
623 case AMDGPUIntrinsic::AMDGPU_bfi:
624 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
629 case AMDGPUIntrinsic::AMDGPU_bfm:
630 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
634 case AMDGPUIntrinsic::AMDIL_round_nearest:
635 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
639 ///IABS(a) = SMAX(sub(0, a), a)
640 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
641 SelectionDAG &DAG) const {
644 EVT VT = Op.getValueType();
645 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
648 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
651 /// Linear Interpolation
652 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
653 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
654 SelectionDAG &DAG) const {
656 EVT VT = Op.getValueType();
657 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
658 DAG.getConstantFP(1.0f, MVT::f32),
660 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
662 return DAG.getNode(ISD::FADD, DL, VT,
663 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
667 /// \brief Generate Min/Max node
668 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
669 SelectionDAG &DAG) const {
671 EVT VT = Op.getValueType();
673 SDValue LHS = Op.getOperand(0);
674 SDValue RHS = Op.getOperand(1);
675 SDValue True = Op.getOperand(2);
676 SDValue False = Op.getOperand(3);
677 SDValue CC = Op.getOperand(4);
679 if (VT != MVT::f32 ||
680 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
684 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
698 llvm_unreachable("Operation should already be optimised!");
706 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
708 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
717 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
719 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
721 case ISD::SETCC_INVALID:
722 llvm_unreachable("Invalid setcc condcode!");
727 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
728 SelectionDAG &DAG) const {
729 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
730 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
731 EVT EltVT = Op.getValueType().getVectorElementType();
732 EVT PtrVT = Load->getBasePtr().getValueType();
733 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
734 SmallVector<SDValue, 8> Loads;
737 for (unsigned i = 0, e = NumElts; i != e; ++i) {
738 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
739 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
740 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
741 Load->getChain(), Ptr,
742 MachinePointerInfo(Load->getMemOperand()->getValue()),
743 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
744 Load->getAlignment()));
746 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
749 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
750 SelectionDAG &DAG) const {
751 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
752 EVT MemVT = Store->getMemoryVT();
753 unsigned MemBits = MemVT.getSizeInBits();
755 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
756 // truncating store into an i32 store.
757 // XXX: We could also handle optimize other vector bitwidths.
758 if (!MemVT.isVector() || MemBits > 32) {
763 SDValue Value = Store->getValue();
764 EVT VT = Value.getValueType();
765 EVT ElemVT = VT.getVectorElementType();
766 SDValue Ptr = Store->getBasePtr();
767 EVT MemEltVT = MemVT.getVectorElementType();
768 unsigned MemEltBits = MemEltVT.getSizeInBits();
769 unsigned MemNumElements = MemVT.getVectorNumElements();
770 unsigned PackedSize = MemVT.getStoreSizeInBits();
771 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
773 assert(Value.getValueType().getScalarSizeInBits() >= 32);
776 for (unsigned i = 0; i < MemNumElements; ++i) {
777 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
778 DAG.getConstant(i, MVT::i32));
779 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
780 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
782 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
783 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
788 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
792 if (PackedSize < 32) {
793 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
794 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
795 Store->getMemOperand()->getPointerInfo(),
797 Store->isNonTemporal(), Store->isVolatile(),
798 Store->getAlignment());
801 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
802 Store->getMemOperand()->getPointerInfo(),
803 Store->isVolatile(), Store->isNonTemporal(),
804 Store->getAlignment());
807 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
808 SelectionDAG &DAG) const {
809 StoreSDNode *Store = cast<StoreSDNode>(Op);
810 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
811 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
812 EVT PtrVT = Store->getBasePtr().getValueType();
813 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
816 SmallVector<SDValue, 8> Chains;
818 for (unsigned i = 0, e = NumElts; i != e; ++i) {
819 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
820 Store->getValue(), DAG.getConstant(i, MVT::i32));
821 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
823 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
825 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
826 MachinePointerInfo(Store->getMemOperand()->getValue()),
827 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
828 Store->getAlignment()));
830 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
833 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
835 LoadSDNode *Load = cast<LoadSDNode>(Op);
836 ISD::LoadExtType ExtType = Load->getExtensionType();
837 EVT VT = Op.getValueType();
838 EVT MemVT = Load->getMemoryVT();
840 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
841 // We can do the extload to 32-bits, and then need to separately extend to
844 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
848 Load->getMemOperand());
849 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
852 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
853 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
854 // FIXME: Copied from PPC
855 // First, load into 32 bits, then truncate to 1 bit.
857 SDValue Chain = Load->getChain();
858 SDValue BasePtr = Load->getBasePtr();
859 MachineMemOperand *MMO = Load->getMemOperand();
861 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
862 BasePtr, MVT::i8, MMO);
863 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
866 // Lower loads constant address space global variable loads
867 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
869 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
871 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
872 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
873 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
874 DAG.getConstant(2, MVT::i32));
875 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
876 Load->getChain(), Ptr,
877 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
880 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
881 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
885 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
886 DAG.getConstant(2, MVT::i32));
887 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
888 Load->getChain(), Ptr,
889 DAG.getTargetConstant(0, MVT::i32),
891 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
893 DAG.getConstant(0x3, MVT::i32));
894 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
895 DAG.getConstant(3, MVT::i32));
897 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
899 EVT MemEltVT = MemVT.getScalarType();
900 if (ExtType == ISD::SEXTLOAD) {
901 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
902 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
905 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
908 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
910 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
911 if (Result.getNode()) {
915 StoreSDNode *Store = cast<StoreSDNode>(Op);
916 SDValue Chain = Store->getChain();
917 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
918 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
919 Store->getValue().getValueType().isVector()) {
920 return SplitVectorStore(Op, DAG);
923 EVT MemVT = Store->getMemoryVT();
924 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
925 MemVT.bitsLT(MVT::i32)) {
927 if (Store->getMemoryVT() == MVT::i8) {
929 } else if (Store->getMemoryVT() == MVT::i16) {
932 SDValue BasePtr = Store->getBasePtr();
933 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
934 DAG.getConstant(2, MVT::i32));
935 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
936 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
938 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
939 DAG.getConstant(0x3, MVT::i32));
941 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
942 DAG.getConstant(3, MVT::i32));
944 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
947 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
949 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
950 MaskedValue, ShiftAmt);
952 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
954 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
955 DAG.getConstant(0xffffffff, MVT::i32));
956 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
958 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
959 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
960 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
965 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
966 SelectionDAG &DAG) const {
968 EVT VT = Op.getValueType();
970 SDValue Num = Op.getOperand(0);
971 SDValue Den = Op.getOperand(1);
973 // RCP = URECIP(Den) = 2^32 / Den + e
974 // e is rounding error.
975 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
977 // RCP_LO = umulo(RCP, Den) */
978 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
980 // RCP_HI = mulhu (RCP, Den) */
981 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
983 // NEG_RCP_LO = -RCP_LO
984 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
987 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
988 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
991 // Calculate the rounding error from the URECIP instruction
992 // E = mulhu(ABS_RCP_LO, RCP)
993 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
996 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
999 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1001 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1002 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1005 // Quotient = mulhu(Tmp0, Num)
1006 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1008 // Num_S_Remainder = Quotient * Den
1009 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1011 // Remainder = Num - Num_S_Remainder
1012 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1014 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1015 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1016 DAG.getConstant(-1, VT),
1017 DAG.getConstant(0, VT),
1019 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1020 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1022 DAG.getConstant(-1, VT),
1023 DAG.getConstant(0, VT),
1025 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1026 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1029 // Calculate Division result:
1031 // Quotient_A_One = Quotient + 1
1032 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1033 DAG.getConstant(1, VT));
1035 // Quotient_S_One = Quotient - 1
1036 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1037 DAG.getConstant(1, VT));
1039 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1040 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1041 Quotient, Quotient_A_One, ISD::SETEQ);
1043 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1044 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1045 Quotient_S_One, Div, ISD::SETEQ);
1047 // Calculate Rem result:
1049 // Remainder_S_Den = Remainder - Den
1050 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1052 // Remainder_A_Den = Remainder + Den
1053 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1055 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1056 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1057 Remainder, Remainder_S_Den, ISD::SETEQ);
1059 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1060 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1061 Remainder_A_Den, Rem, ISD::SETEQ);
1066 return DAG.getMergeValues(Ops, DL);
1069 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1070 SelectionDAG &DAG) const {
1071 SDValue S0 = Op.getOperand(0);
1073 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1076 // f32 uint_to_fp i64
1077 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1078 DAG.getConstant(0, MVT::i32));
1079 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1080 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1081 DAG.getConstant(1, MVT::i32));
1082 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1083 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1084 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1085 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1089 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1091 SelectionDAG &DAG) const {
1092 MVT VT = Op.getSimpleValueType();
1094 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1095 // Shift left by 'Shift' bits.
1096 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1097 // Signed shift Right by 'Shift' bits.
1098 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1101 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1102 SelectionDAG &DAG) const {
1103 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1104 MVT VT = Op.getSimpleValueType();
1105 MVT ScalarVT = VT.getScalarType();
1110 SDValue Src = Op.getOperand(0);
1113 // TODO: Don't scalarize on Evergreen?
1114 unsigned NElts = VT.getVectorNumElements();
1115 SmallVector<SDValue, 8> Args;
1116 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1118 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1119 for (unsigned I = 0; I < NElts; ++I)
1120 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1122 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1125 //===----------------------------------------------------------------------===//
1126 // Custom DAG optimizations
1127 //===----------------------------------------------------------------------===//
1129 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1130 APInt KnownZero, KnownOne;
1131 EVT VT = Op.getValueType();
1132 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
1134 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1137 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1138 EVT VT = Op.getValueType();
1140 // In order for this to be a signed 24-bit value, bit 23, must
1142 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1143 // as unsigned 24-bit values.
1144 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1147 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1149 SelectionDAG &DAG = DCI.DAG;
1150 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1151 EVT VT = Op.getValueType();
1153 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1154 APInt KnownZero, KnownOne;
1155 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1156 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1157 DCI.CommitTargetLoweringOpt(TLO);
1160 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1161 DAGCombinerInfo &DCI) const {
1162 SelectionDAG &DAG = DCI.DAG;
1165 switch(N->getOpcode()) {
1168 EVT VT = N->getValueType(0);
1169 SDValue N0 = N->getOperand(0);
1170 SDValue N1 = N->getOperand(1);
1173 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1174 if (VT.isVector() || VT.getSizeInBits() > 32)
1177 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1178 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1179 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1180 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1181 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1182 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1183 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1184 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1189 // We need to use sext even for MUL_U24, because MUL_U24 is used
1190 // for signed multiply of 8 and 16-bit types.
1191 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1195 case AMDGPUISD::MUL_I24:
1196 case AMDGPUISD::MUL_U24: {
1197 SDValue N0 = N->getOperand(0);
1198 SDValue N1 = N->getOperand(1);
1199 simplifyI24(N0, DCI);
1200 simplifyI24(N1, DCI);
1207 //===----------------------------------------------------------------------===//
1209 //===----------------------------------------------------------------------===//
1211 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1214 const SmallVectorImpl<ISD::InputArg> &Ins,
1215 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1217 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1218 if (Ins[i].ArgVT == Ins[i].VT) {
1219 OrigIns.push_back(Ins[i]);
1224 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1225 // Vector has been split into scalars.
1226 VT = Ins[i].ArgVT.getVectorElementType();
1227 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1228 Ins[i].ArgVT.getVectorElementType() !=
1229 Ins[i].VT.getVectorElementType()) {
1230 // Vector elements have been promoted
1233 // Vector has been spilt into smaller vectors.
1237 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1238 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1239 OrigIns.push_back(Arg);
1243 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1244 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1245 return CFP->isExactlyValue(1.0);
1247 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1248 return C->isAllOnesValue();
1253 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1254 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1255 return CFP->getValueAPF().isZero();
1257 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1258 return C->isNullValue();
1263 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1264 const TargetRegisterClass *RC,
1265 unsigned Reg, EVT VT) const {
1266 MachineFunction &MF = DAG.getMachineFunction();
1267 MachineRegisterInfo &MRI = MF.getRegInfo();
1268 unsigned VirtualRegister;
1269 if (!MRI.isLiveIn(Reg)) {
1270 VirtualRegister = MRI.createVirtualRegister(RC);
1271 MRI.addLiveIn(Reg, VirtualRegister);
1273 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1275 return DAG.getRegister(VirtualRegister, VT);
1278 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1280 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1282 default: return nullptr;
1284 NODE_NAME_CASE(CALL);
1285 NODE_NAME_CASE(UMUL);
1286 NODE_NAME_CASE(DIV_INF);
1287 NODE_NAME_CASE(RET_FLAG);
1288 NODE_NAME_CASE(BRANCH_COND);
1291 NODE_NAME_CASE(DWORDADDR)
1292 NODE_NAME_CASE(FRACT)
1293 NODE_NAME_CASE(FMAX)
1294 NODE_NAME_CASE(SMAX)
1295 NODE_NAME_CASE(UMAX)
1296 NODE_NAME_CASE(FMIN)
1297 NODE_NAME_CASE(SMIN)
1298 NODE_NAME_CASE(UMIN)
1299 NODE_NAME_CASE(BFE_U32)
1300 NODE_NAME_CASE(BFE_I32)
1303 NODE_NAME_CASE(MUL_U24)
1304 NODE_NAME_CASE(MUL_I24)
1305 NODE_NAME_CASE(URECIP)
1306 NODE_NAME_CASE(DOT4)
1307 NODE_NAME_CASE(EXPORT)
1308 NODE_NAME_CASE(CONST_ADDRESS)
1309 NODE_NAME_CASE(REGISTER_LOAD)
1310 NODE_NAME_CASE(REGISTER_STORE)
1311 NODE_NAME_CASE(LOAD_CONSTANT)
1312 NODE_NAME_CASE(LOAD_INPUT)
1313 NODE_NAME_CASE(SAMPLE)
1314 NODE_NAME_CASE(SAMPLEB)
1315 NODE_NAME_CASE(SAMPLED)
1316 NODE_NAME_CASE(SAMPLEL)
1317 NODE_NAME_CASE(STORE_MSKOR)
1318 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1322 static void computeMaskedBitsForMinMax(const SDValue Op0,
1326 const SelectionDAG &DAG,
1328 APInt Op0Zero, Op0One;
1329 APInt Op1Zero, Op1One;
1330 DAG.ComputeMaskedBits(Op0, Op0Zero, Op0One, Depth);
1331 DAG.ComputeMaskedBits(Op1, Op1Zero, Op1One, Depth);
1333 KnownZero = Op0Zero & Op1Zero;
1334 KnownOne = Op0One & Op1One;
1337 void AMDGPUTargetLowering::computeMaskedBitsForTargetNode(
1341 const SelectionDAG &DAG,
1342 unsigned Depth) const {
1344 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1345 unsigned Opc = Op.getOpcode();
1347 case ISD::INTRINSIC_WO_CHAIN: {
1348 // FIXME: The intrinsic should just use the node.
1349 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1350 case AMDGPUIntrinsic::AMDGPU_imax:
1351 case AMDGPUIntrinsic::AMDGPU_umax:
1352 case AMDGPUIntrinsic::AMDGPU_imin:
1353 case AMDGPUIntrinsic::AMDGPU_umin:
1354 computeMaskedBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1355 KnownZero, KnownOne, DAG, Depth);
1363 case AMDGPUISD::SMAX:
1364 case AMDGPUISD::UMAX:
1365 case AMDGPUISD::SMIN:
1366 case AMDGPUISD::UMIN:
1367 computeMaskedBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1368 KnownZero, KnownOne, DAG, Depth);