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));
431 SDValue Op = SDValue(N, 0);
433 EVT VT = Op.getValueType();
434 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
437 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT,
438 N->getOperand(0), DAG.getConstant(0, HalfVT));
439 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT,
440 N->getOperand(0), DAG.getConstant(1, HalfVT));
442 SDValue RHS = N->getOperand(1);
444 SDValue DIV = DAG.getConstant(0, VT);
445 SDValue REM = DAG.getConstant(0, VT);
447 const unsigned bitWidth = VT.getSizeInBits();
448 const unsigned halfBitWidth = HalfVT.getSizeInBits();
450 SDValue one = DAG.getConstant(1, HalfVT);
451 SDValue one_VT = DAG.getConstant(1, VT);
452 for (unsigned i = 0; i < bitWidth; ++i) {
453 SDValue POS = DAG.getConstant((bitWidth - i - 1) % halfBitWidth, HalfVT);
454 // Get Value of high bit
455 SDValue HBit = DAG.getNode(ISD::SRL, DL, HalfVT,
456 i < halfBitWidth ? LHS_Hi : LHS_Lo, POS);
457 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
458 HBit = DAG.getNode(ISD::ZERO_EXTEND, DL, VT, HBit);
460 // Add the high bit to shifted remainder
461 REM = DAG.getNode(ISD::SHL, DL, VT, REM, one);
462 REM = DAG.getNode(ISD::OR, DL, VT, REM, HBit);
465 SDValue ShDIV = DAG.getNode(ISD::SHL, DL, VT, DIV, one);
466 SDValue ShDIV_plus = DAG.getNode(ISD::OR, DL, VT, ShDIV, one_VT);
468 DIV = DAG.getSelectCC(DL, REM, RHS, ShDIV_plus, ShDIV, ISD::SETGE);
471 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
473 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
476 Results.push_back(DIV);
477 Results.push_back(REM);
485 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
486 const GlobalValue *GV,
487 const SDValue &InitPtr,
489 SelectionDAG &DAG) const {
490 const DataLayout *TD = getTargetMachine().getDataLayout();
492 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
493 EVT VT = EVT::getEVT(CI->getType());
494 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
495 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
496 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
497 TD->getPrefTypeAlignment(CI->getType()));
498 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
499 EVT VT = EVT::getEVT(CFP->getType());
500 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
501 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
502 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
503 TD->getPrefTypeAlignment(CFP->getType()));
504 } else if (Init->getType()->isAggregateType()) {
505 EVT PtrVT = InitPtr.getValueType();
506 unsigned NumElements = Init->getType()->getArrayNumElements();
507 SmallVector<SDValue, 8> Chains;
508 for (unsigned i = 0; i < NumElements; ++i) {
509 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
510 Init->getType()->getArrayElementType()), PtrVT);
511 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
512 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
513 GV, Ptr, Chain, DAG));
515 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
518 llvm_unreachable("Unhandled constant initializer");
522 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
524 SelectionDAG &DAG) const {
526 const DataLayout *TD = getTargetMachine().getDataLayout();
527 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
528 const GlobalValue *GV = G->getGlobal();
530 switch (G->getAddressSpace()) {
531 default: llvm_unreachable("Global Address lowering not implemented for this "
533 case AMDGPUAS::LOCAL_ADDRESS: {
534 // XXX: What does the value of G->getOffset() mean?
535 assert(G->getOffset() == 0 &&
536 "Do not know what to do with an non-zero offset");
539 if (MFI->LocalMemoryObjects.count(GV) == 0) {
540 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
541 Offset = MFI->LDSSize;
542 MFI->LocalMemoryObjects[GV] = Offset;
543 // XXX: Account for alignment?
544 MFI->LDSSize += Size;
546 Offset = MFI->LocalMemoryObjects[GV];
549 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
551 case AMDGPUAS::CONSTANT_ADDRESS: {
552 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
553 Type *EltType = GV->getType()->getElementType();
554 unsigned Size = TD->getTypeAllocSize(EltType);
555 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
557 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
558 const Constant *Init = Var->getInitializer();
559 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
560 SDValue InitPtr = DAG.getFrameIndex(FI,
561 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
562 SmallVector<SDNode*, 8> WorkList;
564 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
565 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
566 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
568 WorkList.push_back(*I);
570 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
571 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
572 E = WorkList.end(); I != E; ++I) {
573 SmallVector<SDValue, 8> Ops;
574 Ops.push_back(Chain);
575 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
576 Ops.push_back((*I)->getOperand(i));
578 DAG.UpdateNodeOperands(*I, Ops);
580 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
581 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
586 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
587 SelectionDAG &DAG) const {
588 SmallVector<SDValue, 8> Args;
589 SDValue A = Op.getOperand(0);
590 SDValue B = Op.getOperand(1);
592 DAG.ExtractVectorElements(A, Args);
593 DAG.ExtractVectorElements(B, Args);
595 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
598 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
599 SelectionDAG &DAG) const {
601 SmallVector<SDValue, 8> Args;
602 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
603 EVT VT = Op.getValueType();
604 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
605 VT.getVectorNumElements());
607 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
610 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
611 SelectionDAG &DAG) const {
613 MachineFunction &MF = DAG.getMachineFunction();
614 const AMDGPUFrameLowering *TFL =
615 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
617 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
620 unsigned FrameIndex = FIN->getIndex();
621 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
622 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
626 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
627 SelectionDAG &DAG) const {
628 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
630 EVT VT = Op.getValueType();
632 switch (IntrinsicID) {
634 case AMDGPUIntrinsic::AMDIL_abs:
635 return LowerIntrinsicIABS(Op, DAG);
636 case AMDGPUIntrinsic::AMDIL_exp:
637 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
638 case AMDGPUIntrinsic::AMDGPU_lrp:
639 return LowerIntrinsicLRP(Op, DAG);
640 case AMDGPUIntrinsic::AMDIL_fraction:
641 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
642 case AMDGPUIntrinsic::AMDIL_max:
643 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
645 case AMDGPUIntrinsic::AMDGPU_imax:
646 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
648 case AMDGPUIntrinsic::AMDGPU_umax:
649 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
651 case AMDGPUIntrinsic::AMDIL_min:
652 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
654 case AMDGPUIntrinsic::AMDGPU_imin:
655 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
657 case AMDGPUIntrinsic::AMDGPU_umin:
658 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
661 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
662 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
667 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
668 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
673 case AMDGPUIntrinsic::AMDGPU_bfi:
674 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
679 case AMDGPUIntrinsic::AMDGPU_bfm:
680 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
684 case AMDGPUIntrinsic::AMDIL_round_nearest:
685 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
689 ///IABS(a) = SMAX(sub(0, a), a)
690 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
691 SelectionDAG &DAG) const {
694 EVT VT = Op.getValueType();
695 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
698 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
701 /// Linear Interpolation
702 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
703 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
704 SelectionDAG &DAG) const {
706 EVT VT = Op.getValueType();
707 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
708 DAG.getConstantFP(1.0f, MVT::f32),
710 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
712 return DAG.getNode(ISD::FADD, DL, VT,
713 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
717 /// \brief Generate Min/Max node
718 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
719 SelectionDAG &DAG) const {
721 EVT VT = Op.getValueType();
723 SDValue LHS = Op.getOperand(0);
724 SDValue RHS = Op.getOperand(1);
725 SDValue True = Op.getOperand(2);
726 SDValue False = Op.getOperand(3);
727 SDValue CC = Op.getOperand(4);
729 if (VT != MVT::f32 ||
730 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
734 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
748 llvm_unreachable("Operation should already be optimised!");
756 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
758 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
767 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
769 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
771 case ISD::SETCC_INVALID:
772 llvm_unreachable("Invalid setcc condcode!");
777 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
778 SelectionDAG &DAG) const {
779 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
780 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
781 EVT EltVT = Op.getValueType().getVectorElementType();
782 EVT PtrVT = Load->getBasePtr().getValueType();
783 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
784 SmallVector<SDValue, 8> Loads;
787 for (unsigned i = 0, e = NumElts; i != e; ++i) {
788 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
789 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
790 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
791 Load->getChain(), Ptr,
792 MachinePointerInfo(Load->getMemOperand()->getValue()),
793 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
794 Load->getAlignment()));
796 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
799 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
800 SelectionDAG &DAG) const {
801 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
802 EVT MemVT = Store->getMemoryVT();
803 unsigned MemBits = MemVT.getSizeInBits();
805 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
806 // truncating store into an i32 store.
807 // XXX: We could also handle optimize other vector bitwidths.
808 if (!MemVT.isVector() || MemBits > 32) {
813 SDValue Value = Store->getValue();
814 EVT VT = Value.getValueType();
815 EVT ElemVT = VT.getVectorElementType();
816 SDValue Ptr = Store->getBasePtr();
817 EVT MemEltVT = MemVT.getVectorElementType();
818 unsigned MemEltBits = MemEltVT.getSizeInBits();
819 unsigned MemNumElements = MemVT.getVectorNumElements();
820 unsigned PackedSize = MemVT.getStoreSizeInBits();
821 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
823 assert(Value.getValueType().getScalarSizeInBits() >= 32);
826 for (unsigned i = 0; i < MemNumElements; ++i) {
827 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
828 DAG.getConstant(i, MVT::i32));
829 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
830 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
832 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
833 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
838 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
842 if (PackedSize < 32) {
843 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
844 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
845 Store->getMemOperand()->getPointerInfo(),
847 Store->isNonTemporal(), Store->isVolatile(),
848 Store->getAlignment());
851 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
852 Store->getMemOperand()->getPointerInfo(),
853 Store->isVolatile(), Store->isNonTemporal(),
854 Store->getAlignment());
857 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
858 SelectionDAG &DAG) const {
859 StoreSDNode *Store = cast<StoreSDNode>(Op);
860 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
861 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
862 EVT PtrVT = Store->getBasePtr().getValueType();
863 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
866 SmallVector<SDValue, 8> Chains;
868 for (unsigned i = 0, e = NumElts; i != e; ++i) {
869 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
870 Store->getValue(), DAG.getConstant(i, MVT::i32));
871 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
873 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
875 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
876 MachinePointerInfo(Store->getMemOperand()->getValue()),
877 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
878 Store->getAlignment()));
880 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
883 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
885 LoadSDNode *Load = cast<LoadSDNode>(Op);
886 ISD::LoadExtType ExtType = Load->getExtensionType();
887 EVT VT = Op.getValueType();
888 EVT MemVT = Load->getMemoryVT();
890 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
891 // We can do the extload to 32-bits, and then need to separately extend to
894 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
898 Load->getMemOperand());
899 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
902 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
903 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
904 // FIXME: Copied from PPC
905 // First, load into 32 bits, then truncate to 1 bit.
907 SDValue Chain = Load->getChain();
908 SDValue BasePtr = Load->getBasePtr();
909 MachineMemOperand *MMO = Load->getMemOperand();
911 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
912 BasePtr, MVT::i8, MMO);
913 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
916 // Lower loads constant address space global variable loads
917 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
919 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
921 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
922 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
923 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
924 DAG.getConstant(2, MVT::i32));
925 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
926 Load->getChain(), Ptr,
927 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
930 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
931 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
935 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
936 DAG.getConstant(2, MVT::i32));
937 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
938 Load->getChain(), Ptr,
939 DAG.getTargetConstant(0, MVT::i32),
941 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
943 DAG.getConstant(0x3, MVT::i32));
944 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
945 DAG.getConstant(3, MVT::i32));
947 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
949 EVT MemEltVT = MemVT.getScalarType();
950 if (ExtType == ISD::SEXTLOAD) {
951 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
952 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
955 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
958 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
960 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
961 if (Result.getNode()) {
965 StoreSDNode *Store = cast<StoreSDNode>(Op);
966 SDValue Chain = Store->getChain();
967 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
968 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
969 Store->getValue().getValueType().isVector()) {
970 return SplitVectorStore(Op, DAG);
973 EVT MemVT = Store->getMemoryVT();
974 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
975 MemVT.bitsLT(MVT::i32)) {
977 if (Store->getMemoryVT() == MVT::i8) {
979 } else if (Store->getMemoryVT() == MVT::i16) {
982 SDValue BasePtr = Store->getBasePtr();
983 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
984 DAG.getConstant(2, MVT::i32));
985 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
986 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
988 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
989 DAG.getConstant(0x3, MVT::i32));
991 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
992 DAG.getConstant(3, MVT::i32));
994 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
997 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
999 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1000 MaskedValue, ShiftAmt);
1002 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1004 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1005 DAG.getConstant(0xffffffff, MVT::i32));
1006 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1008 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1009 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1010 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1015 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1016 SelectionDAG &DAG) const {
1018 EVT VT = Op.getValueType();
1020 SDValue Num = Op.getOperand(0);
1021 SDValue Den = Op.getOperand(1);
1023 // RCP = URECIP(Den) = 2^32 / Den + e
1024 // e is rounding error.
1025 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1027 // RCP_LO = umulo(RCP, Den) */
1028 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1030 // RCP_HI = mulhu (RCP, Den) */
1031 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1033 // NEG_RCP_LO = -RCP_LO
1034 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1037 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1038 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1041 // Calculate the rounding error from the URECIP instruction
1042 // E = mulhu(ABS_RCP_LO, RCP)
1043 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1045 // RCP_A_E = RCP + E
1046 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1048 // RCP_S_E = RCP - E
1049 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1051 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1052 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1055 // Quotient = mulhu(Tmp0, Num)
1056 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1058 // Num_S_Remainder = Quotient * Den
1059 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1061 // Remainder = Num - Num_S_Remainder
1062 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1064 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1065 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1066 DAG.getConstant(-1, VT),
1067 DAG.getConstant(0, VT),
1069 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1070 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1072 DAG.getConstant(-1, VT),
1073 DAG.getConstant(0, VT),
1075 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1076 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1079 // Calculate Division result:
1081 // Quotient_A_One = Quotient + 1
1082 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1083 DAG.getConstant(1, VT));
1085 // Quotient_S_One = Quotient - 1
1086 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1087 DAG.getConstant(1, VT));
1089 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1090 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1091 Quotient, Quotient_A_One, ISD::SETEQ);
1093 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1094 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1095 Quotient_S_One, Div, ISD::SETEQ);
1097 // Calculate Rem result:
1099 // Remainder_S_Den = Remainder - Den
1100 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1102 // Remainder_A_Den = Remainder + Den
1103 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1105 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1106 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1107 Remainder, Remainder_S_Den, ISD::SETEQ);
1109 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1110 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1111 Remainder_A_Den, Rem, ISD::SETEQ);
1116 return DAG.getMergeValues(Ops, DL);
1119 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1120 SelectionDAG &DAG) const {
1121 SDValue S0 = Op.getOperand(0);
1123 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1126 // f32 uint_to_fp i64
1127 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1128 DAG.getConstant(0, MVT::i32));
1129 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1130 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1131 DAG.getConstant(1, MVT::i32));
1132 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1133 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1134 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1135 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1139 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1141 SelectionDAG &DAG) const {
1142 MVT VT = Op.getSimpleValueType();
1144 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1145 // Shift left by 'Shift' bits.
1146 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1147 // Signed shift Right by 'Shift' bits.
1148 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1151 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1152 SelectionDAG &DAG) const {
1153 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1154 MVT VT = Op.getSimpleValueType();
1155 MVT ScalarVT = VT.getScalarType();
1160 SDValue Src = Op.getOperand(0);
1163 // TODO: Don't scalarize on Evergreen?
1164 unsigned NElts = VT.getVectorNumElements();
1165 SmallVector<SDValue, 8> Args;
1166 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1168 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1169 for (unsigned I = 0; I < NElts; ++I)
1170 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1172 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1175 //===----------------------------------------------------------------------===//
1176 // Custom DAG optimizations
1177 //===----------------------------------------------------------------------===//
1179 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1180 APInt KnownZero, KnownOne;
1181 EVT VT = Op.getValueType();
1182 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
1184 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1187 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1188 EVT VT = Op.getValueType();
1190 // In order for this to be a signed 24-bit value, bit 23, must
1192 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1193 // as unsigned 24-bit values.
1194 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1197 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1199 SelectionDAG &DAG = DCI.DAG;
1200 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1201 EVT VT = Op.getValueType();
1203 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1204 APInt KnownZero, KnownOne;
1205 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1206 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1207 DCI.CommitTargetLoweringOpt(TLO);
1210 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1211 DAGCombinerInfo &DCI) const {
1212 SelectionDAG &DAG = DCI.DAG;
1215 switch(N->getOpcode()) {
1218 EVT VT = N->getValueType(0);
1219 SDValue N0 = N->getOperand(0);
1220 SDValue N1 = N->getOperand(1);
1223 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1224 if (VT.isVector() || VT.getSizeInBits() > 32)
1227 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1228 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1229 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1230 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1231 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1232 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1233 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1234 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1239 // We need to use sext even for MUL_U24, because MUL_U24 is used
1240 // for signed multiply of 8 and 16-bit types.
1241 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1245 case AMDGPUISD::MUL_I24:
1246 case AMDGPUISD::MUL_U24: {
1247 SDValue N0 = N->getOperand(0);
1248 SDValue N1 = N->getOperand(1);
1249 simplifyI24(N0, DCI);
1250 simplifyI24(N1, DCI);
1257 //===----------------------------------------------------------------------===//
1259 //===----------------------------------------------------------------------===//
1261 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1264 const SmallVectorImpl<ISD::InputArg> &Ins,
1265 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1267 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1268 if (Ins[i].ArgVT == Ins[i].VT) {
1269 OrigIns.push_back(Ins[i]);
1274 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1275 // Vector has been split into scalars.
1276 VT = Ins[i].ArgVT.getVectorElementType();
1277 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1278 Ins[i].ArgVT.getVectorElementType() !=
1279 Ins[i].VT.getVectorElementType()) {
1280 // Vector elements have been promoted
1283 // Vector has been spilt into smaller vectors.
1287 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1288 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1289 OrigIns.push_back(Arg);
1293 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1294 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1295 return CFP->isExactlyValue(1.0);
1297 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1298 return C->isAllOnesValue();
1303 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1304 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1305 return CFP->getValueAPF().isZero();
1307 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1308 return C->isNullValue();
1313 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1314 const TargetRegisterClass *RC,
1315 unsigned Reg, EVT VT) const {
1316 MachineFunction &MF = DAG.getMachineFunction();
1317 MachineRegisterInfo &MRI = MF.getRegInfo();
1318 unsigned VirtualRegister;
1319 if (!MRI.isLiveIn(Reg)) {
1320 VirtualRegister = MRI.createVirtualRegister(RC);
1321 MRI.addLiveIn(Reg, VirtualRegister);
1323 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1325 return DAG.getRegister(VirtualRegister, VT);
1328 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1330 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1332 default: return nullptr;
1334 NODE_NAME_CASE(CALL);
1335 NODE_NAME_CASE(UMUL);
1336 NODE_NAME_CASE(DIV_INF);
1337 NODE_NAME_CASE(RET_FLAG);
1338 NODE_NAME_CASE(BRANCH_COND);
1341 NODE_NAME_CASE(DWORDADDR)
1342 NODE_NAME_CASE(FRACT)
1343 NODE_NAME_CASE(FMAX)
1344 NODE_NAME_CASE(SMAX)
1345 NODE_NAME_CASE(UMAX)
1346 NODE_NAME_CASE(FMIN)
1347 NODE_NAME_CASE(SMIN)
1348 NODE_NAME_CASE(UMIN)
1349 NODE_NAME_CASE(BFE_U32)
1350 NODE_NAME_CASE(BFE_I32)
1353 NODE_NAME_CASE(MUL_U24)
1354 NODE_NAME_CASE(MUL_I24)
1355 NODE_NAME_CASE(URECIP)
1356 NODE_NAME_CASE(DOT4)
1357 NODE_NAME_CASE(EXPORT)
1358 NODE_NAME_CASE(CONST_ADDRESS)
1359 NODE_NAME_CASE(REGISTER_LOAD)
1360 NODE_NAME_CASE(REGISTER_STORE)
1361 NODE_NAME_CASE(LOAD_CONSTANT)
1362 NODE_NAME_CASE(LOAD_INPUT)
1363 NODE_NAME_CASE(SAMPLE)
1364 NODE_NAME_CASE(SAMPLEB)
1365 NODE_NAME_CASE(SAMPLED)
1366 NODE_NAME_CASE(SAMPLEL)
1367 NODE_NAME_CASE(STORE_MSKOR)
1368 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1372 static void computeMaskedBitsForMinMax(const SDValue Op0,
1376 const SelectionDAG &DAG,
1378 APInt Op0Zero, Op0One;
1379 APInt Op1Zero, Op1One;
1380 DAG.ComputeMaskedBits(Op0, Op0Zero, Op0One, Depth);
1381 DAG.ComputeMaskedBits(Op1, Op1Zero, Op1One, Depth);
1383 KnownZero = Op0Zero & Op1Zero;
1384 KnownOne = Op0One & Op1One;
1387 void AMDGPUTargetLowering::computeMaskedBitsForTargetNode(
1391 const SelectionDAG &DAG,
1392 unsigned Depth) const {
1394 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1395 unsigned Opc = Op.getOpcode();
1397 case ISD::INTRINSIC_WO_CHAIN: {
1398 // FIXME: The intrinsic should just use the node.
1399 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1400 case AMDGPUIntrinsic::AMDGPU_imax:
1401 case AMDGPUIntrinsic::AMDGPU_umax:
1402 case AMDGPUIntrinsic::AMDGPU_imin:
1403 case AMDGPUIntrinsic::AMDGPU_umin:
1404 computeMaskedBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1405 KnownZero, KnownOne, DAG, Depth);
1413 case AMDGPUISD::SMAX:
1414 case AMDGPUISD::UMAX:
1415 case AMDGPUISD::SMIN:
1416 case AMDGPUISD::UMIN:
1417 computeMaskedBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1418 KnownZero, KnownOne, DAG, Depth);