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::i16, Expand);
149 setTruncStoreAction(MVT::i64, MVT::i8, Expand);
150 setTruncStoreAction(MVT::i64, MVT::i1, Expand);
151 setTruncStoreAction(MVT::v2i64, MVT::v2i1, Expand);
152 setTruncStoreAction(MVT::v4i64, MVT::v4i1, Expand);
155 setOperationAction(ISD::LOAD, MVT::f32, Promote);
156 AddPromotedToType(ISD::LOAD, MVT::f32, MVT::i32);
158 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
159 AddPromotedToType(ISD::LOAD, MVT::v2f32, MVT::v2i32);
161 setOperationAction(ISD::LOAD, MVT::v4f32, Promote);
162 AddPromotedToType(ISD::LOAD, MVT::v4f32, MVT::v4i32);
164 setOperationAction(ISD::LOAD, MVT::v8f32, Promote);
165 AddPromotedToType(ISD::LOAD, MVT::v8f32, MVT::v8i32);
167 setOperationAction(ISD::LOAD, MVT::v16f32, Promote);
168 AddPromotedToType(ISD::LOAD, MVT::v16f32, MVT::v16i32);
170 setOperationAction(ISD::LOAD, MVT::f64, Promote);
171 AddPromotedToType(ISD::LOAD, MVT::f64, MVT::i64);
173 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
174 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Custom);
175 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
176 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
177 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2f32, Custom);
178 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v2i32, Custom);
179 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4f32, Custom);
180 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v4i32, Custom);
181 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8f32, Custom);
182 setOperationAction(ISD::EXTRACT_SUBVECTOR, MVT::v8i32, Custom);
184 setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Expand);
185 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Expand);
186 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i8, Expand);
187 setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Expand);
188 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Expand);
189 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i8, Expand);
190 setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Expand);
191 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Expand);
192 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i16, Expand);
193 setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Expand);
194 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Expand);
195 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, Expand);
197 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
199 setOperationAction(ISD::FNEG, MVT::v2f32, Expand);
200 setOperationAction(ISD::FNEG, MVT::v4f32, Expand);
202 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
204 setOperationAction(ISD::MUL, MVT::i64, Expand);
206 setOperationAction(ISD::UDIV, MVT::i32, Expand);
207 setOperationAction(ISD::UDIVREM, MVT::i32, Custom);
208 setOperationAction(ISD::UDIVREM, MVT::i64, Custom);
209 setOperationAction(ISD::UREM, MVT::i32, Expand);
210 setOperationAction(ISD::VSELECT, MVT::v2f32, Expand);
211 setOperationAction(ISD::VSELECT, MVT::v4f32, Expand);
213 static const MVT::SimpleValueType IntTypes[] = {
214 MVT::v2i32, MVT::v4i32
216 const size_t NumIntTypes = array_lengthof(IntTypes);
218 for (unsigned int x = 0; x < NumIntTypes; ++x) {
219 MVT::SimpleValueType VT = IntTypes[x];
220 //Expand the following operations for the current type by default
221 setOperationAction(ISD::ADD, VT, Expand);
222 setOperationAction(ISD::AND, VT, Expand);
223 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
224 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
225 setOperationAction(ISD::MUL, VT, Expand);
226 setOperationAction(ISD::OR, VT, Expand);
227 setOperationAction(ISD::SHL, VT, Expand);
228 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
229 setOperationAction(ISD::SRL, VT, Expand);
230 setOperationAction(ISD::SRA, VT, Expand);
231 setOperationAction(ISD::SUB, VT, Expand);
232 setOperationAction(ISD::UDIV, VT, Expand);
233 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
234 setOperationAction(ISD::UREM, VT, Expand);
235 setOperationAction(ISD::SELECT, VT, Expand);
236 setOperationAction(ISD::VSELECT, VT, Expand);
237 setOperationAction(ISD::XOR, VT, Expand);
240 static const MVT::SimpleValueType FloatTypes[] = {
241 MVT::v2f32, MVT::v4f32
243 const size_t NumFloatTypes = array_lengthof(FloatTypes);
245 for (unsigned int x = 0; x < NumFloatTypes; ++x) {
246 MVT::SimpleValueType VT = FloatTypes[x];
247 setOperationAction(ISD::FABS, VT, Expand);
248 setOperationAction(ISD::FADD, VT, Expand);
249 setOperationAction(ISD::FCOS, VT, Expand);
250 setOperationAction(ISD::FDIV, VT, Expand);
251 setOperationAction(ISD::FPOW, VT, Expand);
252 setOperationAction(ISD::FFLOOR, VT, Expand);
253 setOperationAction(ISD::FTRUNC, VT, Expand);
254 setOperationAction(ISD::FMUL, VT, Expand);
255 setOperationAction(ISD::FRINT, VT, Expand);
256 setOperationAction(ISD::FSQRT, VT, Expand);
257 setOperationAction(ISD::FSIN, VT, Expand);
258 setOperationAction(ISD::FSUB, VT, Expand);
259 setOperationAction(ISD::SELECT, VT, Expand);
262 setTargetDAGCombine(ISD::MUL);
265 //===----------------------------------------------------------------------===//
266 // Target Information
267 //===----------------------------------------------------------------------===//
269 MVT AMDGPUTargetLowering::getVectorIdxTy() const {
273 bool AMDGPUTargetLowering::isLoadBitCastBeneficial(EVT LoadTy,
275 if (LoadTy.getSizeInBits() != CastTy.getSizeInBits())
278 unsigned LScalarSize = LoadTy.getScalarType().getSizeInBits();
279 unsigned CastScalarSize = CastTy.getScalarType().getSizeInBits();
281 return ((LScalarSize <= CastScalarSize) ||
282 (CastScalarSize >= 32) ||
286 //===---------------------------------------------------------------------===//
288 //===---------------------------------------------------------------------===//
290 bool AMDGPUTargetLowering::isFAbsFree(EVT VT) const {
291 assert(VT.isFloatingPoint());
292 return VT == MVT::f32;
295 bool AMDGPUTargetLowering::isFNegFree(EVT VT) const {
296 assert(VT.isFloatingPoint());
297 return VT == MVT::f32;
300 bool AMDGPUTargetLowering::isTruncateFree(EVT Source, EVT Dest) const {
301 // Truncate is just accessing a subregister.
302 return Dest.bitsLT(Source) && (Dest.getSizeInBits() % 32 == 0);
305 bool AMDGPUTargetLowering::isTruncateFree(Type *Source, Type *Dest) const {
306 // Truncate is just accessing a subregister.
307 return Dest->getPrimitiveSizeInBits() < Source->getPrimitiveSizeInBits() &&
308 (Dest->getPrimitiveSizeInBits() % 32 == 0);
311 bool AMDGPUTargetLowering::isZExtFree(Type *Src, Type *Dest) const {
312 const DataLayout *DL = getDataLayout();
313 unsigned SrcSize = DL->getTypeSizeInBits(Src->getScalarType());
314 unsigned DestSize = DL->getTypeSizeInBits(Dest->getScalarType());
316 return SrcSize == 32 && DestSize == 64;
319 bool AMDGPUTargetLowering::isZExtFree(EVT Src, EVT Dest) const {
320 // Any register load of a 64-bit value really requires 2 32-bit moves. For all
321 // practical purposes, the extra mov 0 to load a 64-bit is free. As used,
322 // this will enable reducing 64-bit operations the 32-bit, which is always
324 return Src == MVT::i32 && Dest == MVT::i64;
327 bool AMDGPUTargetLowering::isNarrowingProfitable(EVT SrcVT, EVT DestVT) const {
328 // There aren't really 64-bit registers, but pairs of 32-bit ones and only a
329 // limited number of native 64-bit operations. Shrinking an operation to fit
330 // in a single 32-bit register should always be helpful. As currently used,
331 // this is much less general than the name suggests, and is only used in
332 // places trying to reduce the sizes of loads. Shrinking loads to < 32-bits is
333 // not profitable, and may actually be harmful.
334 return SrcVT.getSizeInBits() > 32 && DestVT.getSizeInBits() == 32;
337 //===---------------------------------------------------------------------===//
338 // TargetLowering Callbacks
339 //===---------------------------------------------------------------------===//
341 void AMDGPUTargetLowering::AnalyzeFormalArguments(CCState &State,
342 const SmallVectorImpl<ISD::InputArg> &Ins) const {
344 State.AnalyzeFormalArguments(Ins, CC_AMDGPU);
347 SDValue AMDGPUTargetLowering::LowerReturn(
349 CallingConv::ID CallConv,
351 const SmallVectorImpl<ISD::OutputArg> &Outs,
352 const SmallVectorImpl<SDValue> &OutVals,
353 SDLoc DL, SelectionDAG &DAG) const {
354 return DAG.getNode(AMDGPUISD::RET_FLAG, DL, MVT::Other, Chain);
357 //===---------------------------------------------------------------------===//
358 // Target specific lowering
359 //===---------------------------------------------------------------------===//
361 SDValue AMDGPUTargetLowering::LowerCall(CallLoweringInfo &CLI,
362 SmallVectorImpl<SDValue> &InVals) const {
363 SDValue Callee = CLI.Callee;
364 SelectionDAG &DAG = CLI.DAG;
366 const Function &Fn = *DAG.getMachineFunction().getFunction();
368 StringRef FuncName("<unknown>");
370 if (const ExternalSymbolSDNode *G = dyn_cast<ExternalSymbolSDNode>(Callee))
371 FuncName = G->getSymbol();
372 else if (const GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
373 FuncName = G->getGlobal()->getName();
375 DiagnosticInfoUnsupported NoCalls(Fn, "call to function " + FuncName);
376 DAG.getContext()->diagnose(NoCalls);
380 SDValue AMDGPUTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG)
382 switch (Op.getOpcode()) {
384 Op.getNode()->dump();
385 llvm_unreachable("Custom lowering code for this"
386 "instruction is not implemented yet!");
388 // AMDIL DAG lowering
389 case ISD::SDIV: return LowerSDIV(Op, DAG);
390 case ISD::SREM: return LowerSREM(Op, DAG);
391 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op, DAG);
392 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
393 // AMDGPU DAG lowering
394 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
395 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
396 case ISD::FrameIndex: return LowerFrameIndex(Op, DAG);
397 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
398 case ISD::UDIVREM: return LowerUDIVREM(Op, DAG);
399 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
404 void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
405 SmallVectorImpl<SDValue> &Results,
406 SelectionDAG &DAG) const {
407 switch (N->getOpcode()) {
408 case ISD::SIGN_EXTEND_INREG:
409 // Different parts of legalization seem to interpret which type of
410 // sign_extend_inreg is the one to check for custom lowering. The extended
411 // from type is what really matters, but some places check for custom
412 // lowering of the result type. This results in trying to use
413 // ReplaceNodeResults to sext_in_reg to an illegal type, so we'll just do
414 // nothing here and let the illegal result integer be handled normally.
417 SDValue Op = SDValue(N, 0);
419 EVT VT = Op.getValueType();
420 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
421 N->getOperand(0), N->getOperand(1));
422 Results.push_back(UDIVREM);
426 SDValue Op = SDValue(N, 0);
428 EVT VT = Op.getValueType();
429 SDValue UDIVREM = DAG.getNode(ISD::UDIVREM, DL, DAG.getVTList(VT, VT),
430 N->getOperand(0), N->getOperand(1));
431 Results.push_back(UDIVREM.getValue(1));
435 SDValue Op = SDValue(N, 0);
437 EVT VT = Op.getValueType();
438 EVT HalfVT = VT.getHalfSizedIntegerVT(*DAG.getContext());
440 SDValue one = DAG.getConstant(1, HalfVT);
441 SDValue zero = DAG.getConstant(0, HalfVT);
444 SDValue LHS = N->getOperand(0);
445 SDValue LHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, zero);
446 SDValue LHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, LHS, one);
448 SDValue RHS = N->getOperand(1);
449 SDValue RHS_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, zero);
450 SDValue RHS_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, RHS, one);
452 // Get Speculative values
453 SDValue DIV_Part = DAG.getNode(ISD::UDIV, DL, HalfVT, LHS_Hi, RHS_Lo);
454 SDValue REM_Part = DAG.getNode(ISD::UREM, DL, HalfVT, LHS_Hi, RHS_Lo);
456 SDValue REM_Hi = zero;
457 SDValue REM_Lo = DAG.getSelectCC(DL, RHS_Hi, zero, REM_Part, LHS_Hi, ISD::SETEQ);
459 SDValue DIV_Hi = DAG.getSelectCC(DL, RHS_Hi, zero, DIV_Part, zero, ISD::SETEQ);
460 SDValue DIV_Lo = zero;
462 const unsigned halfBitWidth = HalfVT.getSizeInBits();
464 for (unsigned i = 0; i < halfBitWidth; ++i) {
465 SDValue POS = DAG.getConstant(halfBitWidth - i - 1, HalfVT);
466 // Get Value of high bit
468 if (halfBitWidth == 32 && Subtarget->hasBFE()) {
469 HBit = DAG.getNode(AMDGPUISD::BFE_U32, DL, HalfVT, LHS_Lo, POS, one);
471 HBit = DAG.getNode(ISD::SRL, DL, HalfVT, LHS_Lo, POS);
472 HBit = DAG.getNode(ISD::AND, DL, HalfVT, HBit, one);
475 SDValue Carry = DAG.getNode(ISD::SRL, DL, HalfVT, REM_Lo,
476 DAG.getConstant(halfBitWidth - 1, HalfVT));
477 REM_Hi = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Hi, one);
478 REM_Hi = DAG.getNode(ISD::OR, DL, HalfVT, REM_Hi, Carry);
480 REM_Lo = DAG.getNode(ISD::SHL, DL, HalfVT, REM_Lo, one);
481 REM_Lo = DAG.getNode(ISD::OR, DL, HalfVT, REM_Lo, HBit);
484 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
486 SDValue BIT = DAG.getConstant(1 << (halfBitWidth - i - 1), HalfVT);
487 SDValue realBIT = DAG.getSelectCC(DL, REM, RHS, BIT, zero, ISD::SETGE);
489 DIV_Lo = DAG.getNode(ISD::OR, DL, HalfVT, DIV_Lo, realBIT);
493 SDValue REM_sub = DAG.getNode(ISD::SUB, DL, VT, REM, RHS);
495 REM = DAG.getSelectCC(DL, REM, RHS, REM_sub, REM, ISD::SETGE);
496 REM_Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, zero);
497 REM_Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, HalfVT, REM, one);
500 SDValue REM = DAG.getNode(ISD::BUILD_PAIR, DL, VT, REM_Lo, REM_Hi);
501 SDValue DIV = DAG.getNode(ISD::BUILD_PAIR, DL, VT, DIV_Lo, DIV_Hi);
502 Results.push_back(DIV);
503 Results.push_back(REM);
511 SDValue AMDGPUTargetLowering::LowerConstantInitializer(const Constant* Init,
512 const GlobalValue *GV,
513 const SDValue &InitPtr,
515 SelectionDAG &DAG) const {
516 const DataLayout *TD = getTargetMachine().getDataLayout();
518 if (const ConstantInt *CI = dyn_cast<ConstantInt>(Init)) {
519 EVT VT = EVT::getEVT(CI->getType());
520 PointerType *PtrTy = PointerType::get(CI->getType(), 0);
521 return DAG.getStore(Chain, DL, DAG.getConstant(*CI, VT), InitPtr,
522 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
523 TD->getPrefTypeAlignment(CI->getType()));
524 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(Init)) {
525 EVT VT = EVT::getEVT(CFP->getType());
526 PointerType *PtrTy = PointerType::get(CFP->getType(), 0);
527 return DAG.getStore(Chain, DL, DAG.getConstantFP(*CFP, VT), InitPtr,
528 MachinePointerInfo(UndefValue::get(PtrTy)), false, false,
529 TD->getPrefTypeAlignment(CFP->getType()));
530 } else if (Init->getType()->isAggregateType()) {
531 EVT PtrVT = InitPtr.getValueType();
532 unsigned NumElements = Init->getType()->getArrayNumElements();
533 SmallVector<SDValue, 8> Chains;
534 for (unsigned i = 0; i < NumElements; ++i) {
535 SDValue Offset = DAG.getConstant(i * TD->getTypeAllocSize(
536 Init->getType()->getArrayElementType()), PtrVT);
537 SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, InitPtr, Offset);
538 Chains.push_back(LowerConstantInitializer(Init->getAggregateElement(i),
539 GV, Ptr, Chain, DAG));
541 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
544 llvm_unreachable("Unhandled constant initializer");
548 SDValue AMDGPUTargetLowering::LowerGlobalAddress(AMDGPUMachineFunction* MFI,
550 SelectionDAG &DAG) const {
552 const DataLayout *TD = getTargetMachine().getDataLayout();
553 GlobalAddressSDNode *G = cast<GlobalAddressSDNode>(Op);
554 const GlobalValue *GV = G->getGlobal();
556 switch (G->getAddressSpace()) {
557 default: llvm_unreachable("Global Address lowering not implemented for this "
559 case AMDGPUAS::LOCAL_ADDRESS: {
560 // XXX: What does the value of G->getOffset() mean?
561 assert(G->getOffset() == 0 &&
562 "Do not know what to do with an non-zero offset");
565 if (MFI->LocalMemoryObjects.count(GV) == 0) {
566 uint64_t Size = TD->getTypeAllocSize(GV->getType()->getElementType());
567 Offset = MFI->LDSSize;
568 MFI->LocalMemoryObjects[GV] = Offset;
569 // XXX: Account for alignment?
570 MFI->LDSSize += Size;
572 Offset = MFI->LocalMemoryObjects[GV];
575 return DAG.getConstant(Offset, getPointerTy(G->getAddressSpace()));
577 case AMDGPUAS::CONSTANT_ADDRESS: {
578 MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
579 Type *EltType = GV->getType()->getElementType();
580 unsigned Size = TD->getTypeAllocSize(EltType);
581 unsigned Alignment = TD->getPrefTypeAlignment(EltType);
583 const GlobalVariable *Var = dyn_cast<GlobalVariable>(GV);
584 const Constant *Init = Var->getInitializer();
585 int FI = FrameInfo->CreateStackObject(Size, Alignment, false);
586 SDValue InitPtr = DAG.getFrameIndex(FI,
587 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
588 SmallVector<SDNode*, 8> WorkList;
590 for (SDNode::use_iterator I = DAG.getEntryNode()->use_begin(),
591 E = DAG.getEntryNode()->use_end(); I != E; ++I) {
592 if (I->getOpcode() != AMDGPUISD::REGISTER_LOAD && I->getOpcode() != ISD::LOAD)
594 WorkList.push_back(*I);
596 SDValue Chain = LowerConstantInitializer(Init, GV, InitPtr, DAG.getEntryNode(), DAG);
597 for (SmallVector<SDNode*, 8>::iterator I = WorkList.begin(),
598 E = WorkList.end(); I != E; ++I) {
599 SmallVector<SDValue, 8> Ops;
600 Ops.push_back(Chain);
601 for (unsigned i = 1; i < (*I)->getNumOperands(); ++i) {
602 Ops.push_back((*I)->getOperand(i));
604 DAG.UpdateNodeOperands(*I, Ops);
606 return DAG.getZExtOrTrunc(InitPtr, SDLoc(Op),
607 getPointerTy(AMDGPUAS::CONSTANT_ADDRESS));
612 SDValue AMDGPUTargetLowering::LowerCONCAT_VECTORS(SDValue Op,
613 SelectionDAG &DAG) const {
614 SmallVector<SDValue, 8> Args;
615 SDValue A = Op.getOperand(0);
616 SDValue B = Op.getOperand(1);
618 DAG.ExtractVectorElements(A, Args);
619 DAG.ExtractVectorElements(B, Args);
621 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
624 SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
625 SelectionDAG &DAG) const {
627 SmallVector<SDValue, 8> Args;
628 unsigned Start = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
629 EVT VT = Op.getValueType();
630 DAG.ExtractVectorElements(Op.getOperand(0), Args, Start,
631 VT.getVectorNumElements());
633 return DAG.getNode(ISD::BUILD_VECTOR, SDLoc(Op), Op.getValueType(), Args);
636 SDValue AMDGPUTargetLowering::LowerFrameIndex(SDValue Op,
637 SelectionDAG &DAG) const {
639 MachineFunction &MF = DAG.getMachineFunction();
640 const AMDGPUFrameLowering *TFL =
641 static_cast<const AMDGPUFrameLowering*>(getTargetMachine().getFrameLowering());
643 FrameIndexSDNode *FIN = dyn_cast<FrameIndexSDNode>(Op);
646 unsigned FrameIndex = FIN->getIndex();
647 unsigned Offset = TFL->getFrameIndexOffset(MF, FrameIndex);
648 return DAG.getConstant(Offset * 4 * TFL->getStackWidth(MF),
652 SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
653 SelectionDAG &DAG) const {
654 unsigned IntrinsicID = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
656 EVT VT = Op.getValueType();
658 switch (IntrinsicID) {
660 case AMDGPUIntrinsic::AMDIL_abs:
661 return LowerIntrinsicIABS(Op, DAG);
662 case AMDGPUIntrinsic::AMDIL_exp:
663 return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
664 case AMDGPUIntrinsic::AMDGPU_lrp:
665 return LowerIntrinsicLRP(Op, DAG);
666 case AMDGPUIntrinsic::AMDIL_fraction:
667 return DAG.getNode(AMDGPUISD::FRACT, DL, VT, Op.getOperand(1));
668 case AMDGPUIntrinsic::AMDIL_max:
669 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, Op.getOperand(1),
671 case AMDGPUIntrinsic::AMDGPU_imax:
672 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
674 case AMDGPUIntrinsic::AMDGPU_umax:
675 return DAG.getNode(AMDGPUISD::UMAX, DL, VT, Op.getOperand(1),
677 case AMDGPUIntrinsic::AMDIL_min:
678 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, Op.getOperand(1),
680 case AMDGPUIntrinsic::AMDGPU_imin:
681 return DAG.getNode(AMDGPUISD::SMIN, DL, VT, Op.getOperand(1),
683 case AMDGPUIntrinsic::AMDGPU_umin:
684 return DAG.getNode(AMDGPUISD::UMIN, DL, VT, Op.getOperand(1),
687 case AMDGPUIntrinsic::AMDGPU_bfe_i32:
688 return DAG.getNode(AMDGPUISD::BFE_I32, DL, VT,
693 case AMDGPUIntrinsic::AMDGPU_bfe_u32:
694 return DAG.getNode(AMDGPUISD::BFE_U32, DL, VT,
699 case AMDGPUIntrinsic::AMDGPU_bfi:
700 return DAG.getNode(AMDGPUISD::BFI, DL, VT,
705 case AMDGPUIntrinsic::AMDGPU_bfm:
706 return DAG.getNode(AMDGPUISD::BFM, DL, VT,
710 case AMDGPUIntrinsic::AMDIL_round_nearest:
711 return DAG.getNode(ISD::FRINT, DL, VT, Op.getOperand(1));
715 ///IABS(a) = SMAX(sub(0, a), a)
716 SDValue AMDGPUTargetLowering::LowerIntrinsicIABS(SDValue Op,
717 SelectionDAG &DAG) const {
720 EVT VT = Op.getValueType();
721 SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
724 return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Neg, Op.getOperand(1));
727 /// Linear Interpolation
728 /// LRP(a, b, c) = muladd(a, b, (1 - a) * c)
729 SDValue AMDGPUTargetLowering::LowerIntrinsicLRP(SDValue Op,
730 SelectionDAG &DAG) const {
732 EVT VT = Op.getValueType();
733 SDValue OneSubA = DAG.getNode(ISD::FSUB, DL, VT,
734 DAG.getConstantFP(1.0f, MVT::f32),
736 SDValue OneSubAC = DAG.getNode(ISD::FMUL, DL, VT, OneSubA,
738 return DAG.getNode(ISD::FADD, DL, VT,
739 DAG.getNode(ISD::FMUL, DL, VT, Op.getOperand(1), Op.getOperand(2)),
743 /// \brief Generate Min/Max node
744 SDValue AMDGPUTargetLowering::LowerMinMax(SDValue Op,
745 SelectionDAG &DAG) const {
747 EVT VT = Op.getValueType();
749 SDValue LHS = Op.getOperand(0);
750 SDValue RHS = Op.getOperand(1);
751 SDValue True = Op.getOperand(2);
752 SDValue False = Op.getOperand(3);
753 SDValue CC = Op.getOperand(4);
755 if (VT != MVT::f32 ||
756 !((LHS == True && RHS == False) || (LHS == False && RHS == True))) {
760 ISD::CondCode CCOpcode = cast<CondCodeSDNode>(CC)->get();
774 llvm_unreachable("Operation should already be optimised!");
782 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
784 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
793 return DAG.getNode(AMDGPUISD::FMAX, DL, VT, LHS, RHS);
795 return DAG.getNode(AMDGPUISD::FMIN, DL, VT, LHS, RHS);
797 case ISD::SETCC_INVALID:
798 llvm_unreachable("Invalid setcc condcode!");
803 SDValue AMDGPUTargetLowering::SplitVectorLoad(const SDValue &Op,
804 SelectionDAG &DAG) const {
805 LoadSDNode *Load = dyn_cast<LoadSDNode>(Op);
806 EVT MemEltVT = Load->getMemoryVT().getVectorElementType();
807 EVT EltVT = Op.getValueType().getVectorElementType();
808 EVT PtrVT = Load->getBasePtr().getValueType();
809 unsigned NumElts = Load->getMemoryVT().getVectorNumElements();
810 SmallVector<SDValue, 8> Loads;
813 for (unsigned i = 0, e = NumElts; i != e; ++i) {
814 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT, Load->getBasePtr(),
815 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8), PtrVT));
816 Loads.push_back(DAG.getExtLoad(Load->getExtensionType(), SL, EltVT,
817 Load->getChain(), Ptr,
818 MachinePointerInfo(Load->getMemOperand()->getValue()),
819 MemEltVT, Load->isVolatile(), Load->isNonTemporal(),
820 Load->getAlignment()));
822 return DAG.getNode(ISD::BUILD_VECTOR, SL, Op.getValueType(), Loads);
825 SDValue AMDGPUTargetLowering::MergeVectorStore(const SDValue &Op,
826 SelectionDAG &DAG) const {
827 StoreSDNode *Store = dyn_cast<StoreSDNode>(Op);
828 EVT MemVT = Store->getMemoryVT();
829 unsigned MemBits = MemVT.getSizeInBits();
831 // Byte stores are really expensive, so if possible, try to pack 32-bit vector
832 // truncating store into an i32 store.
833 // XXX: We could also handle optimize other vector bitwidths.
834 if (!MemVT.isVector() || MemBits > 32) {
839 SDValue Value = Store->getValue();
840 EVT VT = Value.getValueType();
841 EVT ElemVT = VT.getVectorElementType();
842 SDValue Ptr = Store->getBasePtr();
843 EVT MemEltVT = MemVT.getVectorElementType();
844 unsigned MemEltBits = MemEltVT.getSizeInBits();
845 unsigned MemNumElements = MemVT.getVectorNumElements();
846 unsigned PackedSize = MemVT.getStoreSizeInBits();
847 SDValue Mask = DAG.getConstant((1 << MemEltBits) - 1, MVT::i32);
849 assert(Value.getValueType().getScalarSizeInBits() >= 32);
852 for (unsigned i = 0; i < MemNumElements; ++i) {
853 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ElemVT, Value,
854 DAG.getConstant(i, MVT::i32));
855 Elt = DAG.getZExtOrTrunc(Elt, DL, MVT::i32);
856 Elt = DAG.getNode(ISD::AND, DL, MVT::i32, Elt, Mask); // getZeroExtendInReg
858 SDValue Shift = DAG.getConstant(MemEltBits * i, MVT::i32);
859 Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt, Shift);
864 PackedValue = DAG.getNode(ISD::OR, DL, MVT::i32, PackedValue, Elt);
868 if (PackedSize < 32) {
869 EVT PackedVT = EVT::getIntegerVT(*DAG.getContext(), PackedSize);
870 return DAG.getTruncStore(Store->getChain(), DL, PackedValue, Ptr,
871 Store->getMemOperand()->getPointerInfo(),
873 Store->isNonTemporal(), Store->isVolatile(),
874 Store->getAlignment());
877 return DAG.getStore(Store->getChain(), DL, PackedValue, Ptr,
878 Store->getMemOperand()->getPointerInfo(),
879 Store->isVolatile(), Store->isNonTemporal(),
880 Store->getAlignment());
883 SDValue AMDGPUTargetLowering::SplitVectorStore(SDValue Op,
884 SelectionDAG &DAG) const {
885 StoreSDNode *Store = cast<StoreSDNode>(Op);
886 EVT MemEltVT = Store->getMemoryVT().getVectorElementType();
887 EVT EltVT = Store->getValue().getValueType().getVectorElementType();
888 EVT PtrVT = Store->getBasePtr().getValueType();
889 unsigned NumElts = Store->getMemoryVT().getVectorNumElements();
892 SmallVector<SDValue, 8> Chains;
894 for (unsigned i = 0, e = NumElts; i != e; ++i) {
895 SDValue Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, SL, EltVT,
896 Store->getValue(), DAG.getConstant(i, MVT::i32));
897 SDValue Ptr = DAG.getNode(ISD::ADD, SL, PtrVT,
899 DAG.getConstant(i * (MemEltVT.getSizeInBits() / 8),
901 Chains.push_back(DAG.getTruncStore(Store->getChain(), SL, Val, Ptr,
902 MachinePointerInfo(Store->getMemOperand()->getValue()),
903 MemEltVT, Store->isVolatile(), Store->isNonTemporal(),
904 Store->getAlignment()));
906 return DAG.getNode(ISD::TokenFactor, SL, MVT::Other, Chains);
909 SDValue AMDGPUTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
911 LoadSDNode *Load = cast<LoadSDNode>(Op);
912 ISD::LoadExtType ExtType = Load->getExtensionType();
913 EVT VT = Op.getValueType();
914 EVT MemVT = Load->getMemoryVT();
916 if (ExtType != ISD::NON_EXTLOAD && !VT.isVector() && VT.getSizeInBits() > 32) {
917 // We can do the extload to 32-bits, and then need to separately extend to
920 SDValue ExtLoad32 = DAG.getExtLoad(ExtType, DL, MVT::i32,
924 Load->getMemOperand());
925 return DAG.getNode(ISD::getExtForLoadExtType(ExtType), DL, VT, ExtLoad32);
928 if (ExtType == ISD::NON_EXTLOAD && VT.getSizeInBits() < 32) {
929 assert(VT == MVT::i1 && "Only i1 non-extloads expected");
930 // FIXME: Copied from PPC
931 // First, load into 32 bits, then truncate to 1 bit.
933 SDValue Chain = Load->getChain();
934 SDValue BasePtr = Load->getBasePtr();
935 MachineMemOperand *MMO = Load->getMemOperand();
937 SDValue NewLD = DAG.getExtLoad(ISD::EXTLOAD, DL, MVT::i32, Chain,
938 BasePtr, MVT::i8, MMO);
939 return DAG.getNode(ISD::TRUNCATE, DL, VT, NewLD);
942 // Lower loads constant address space global variable loads
943 if (Load->getAddressSpace() == AMDGPUAS::CONSTANT_ADDRESS &&
945 GetUnderlyingObject(Load->getMemOperand()->getValue()))) {
947 SDValue Ptr = DAG.getZExtOrTrunc(Load->getBasePtr(), DL,
948 getPointerTy(AMDGPUAS::PRIVATE_ADDRESS));
949 Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Ptr,
950 DAG.getConstant(2, MVT::i32));
951 return DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
952 Load->getChain(), Ptr,
953 DAG.getTargetConstant(0, MVT::i32), Op.getOperand(2));
956 if (Load->getAddressSpace() != AMDGPUAS::PRIVATE_ADDRESS ||
957 ExtType == ISD::NON_EXTLOAD || Load->getMemoryVT().bitsGE(MVT::i32))
961 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, Load->getBasePtr(),
962 DAG.getConstant(2, MVT::i32));
963 SDValue Ret = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, Op.getValueType(),
964 Load->getChain(), Ptr,
965 DAG.getTargetConstant(0, MVT::i32),
967 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32,
969 DAG.getConstant(0x3, MVT::i32));
970 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
971 DAG.getConstant(3, MVT::i32));
973 Ret = DAG.getNode(ISD::SRL, DL, MVT::i32, Ret, ShiftAmt);
975 EVT MemEltVT = MemVT.getScalarType();
976 if (ExtType == ISD::SEXTLOAD) {
977 SDValue MemEltVTNode = DAG.getValueType(MemEltVT);
978 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i32, Ret, MemEltVTNode);
981 return DAG.getZeroExtendInReg(Ret, DL, MemEltVT);
984 SDValue AMDGPUTargetLowering::LowerSTORE(SDValue Op, SelectionDAG &DAG) const {
986 SDValue Result = AMDGPUTargetLowering::MergeVectorStore(Op, DAG);
987 if (Result.getNode()) {
991 StoreSDNode *Store = cast<StoreSDNode>(Op);
992 SDValue Chain = Store->getChain();
993 if ((Store->getAddressSpace() == AMDGPUAS::LOCAL_ADDRESS ||
994 Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS) &&
995 Store->getValue().getValueType().isVector()) {
996 return SplitVectorStore(Op, DAG);
999 EVT MemVT = Store->getMemoryVT();
1000 if (Store->getAddressSpace() == AMDGPUAS::PRIVATE_ADDRESS &&
1001 MemVT.bitsLT(MVT::i32)) {
1003 if (Store->getMemoryVT() == MVT::i8) {
1005 } else if (Store->getMemoryVT() == MVT::i16) {
1008 SDValue BasePtr = Store->getBasePtr();
1009 SDValue Ptr = DAG.getNode(ISD::SRL, DL, MVT::i32, BasePtr,
1010 DAG.getConstant(2, MVT::i32));
1011 SDValue Dst = DAG.getNode(AMDGPUISD::REGISTER_LOAD, DL, MVT::i32,
1012 Chain, Ptr, DAG.getTargetConstant(0, MVT::i32));
1014 SDValue ByteIdx = DAG.getNode(ISD::AND, DL, MVT::i32, BasePtr,
1015 DAG.getConstant(0x3, MVT::i32));
1017 SDValue ShiftAmt = DAG.getNode(ISD::SHL, DL, MVT::i32, ByteIdx,
1018 DAG.getConstant(3, MVT::i32));
1020 SDValue SExtValue = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i32,
1023 SDValue MaskedValue = DAG.getZeroExtendInReg(SExtValue, DL, MemVT);
1025 SDValue ShiftedValue = DAG.getNode(ISD::SHL, DL, MVT::i32,
1026 MaskedValue, ShiftAmt);
1028 SDValue DstMask = DAG.getNode(ISD::SHL, DL, MVT::i32, DAG.getConstant(Mask, MVT::i32),
1030 DstMask = DAG.getNode(ISD::XOR, DL, MVT::i32, DstMask,
1031 DAG.getConstant(0xffffffff, MVT::i32));
1032 Dst = DAG.getNode(ISD::AND, DL, MVT::i32, Dst, DstMask);
1034 SDValue Value = DAG.getNode(ISD::OR, DL, MVT::i32, Dst, ShiftedValue);
1035 return DAG.getNode(AMDGPUISD::REGISTER_STORE, DL, MVT::Other,
1036 Chain, Value, Ptr, DAG.getTargetConstant(0, MVT::i32));
1041 SDValue AMDGPUTargetLowering::LowerUDIVREM(SDValue Op,
1042 SelectionDAG &DAG) const {
1044 EVT VT = Op.getValueType();
1046 SDValue Num = Op.getOperand(0);
1047 SDValue Den = Op.getOperand(1);
1049 // RCP = URECIP(Den) = 2^32 / Den + e
1050 // e is rounding error.
1051 SDValue RCP = DAG.getNode(AMDGPUISD::URECIP, DL, VT, Den);
1053 // RCP_LO = umulo(RCP, Den) */
1054 SDValue RCP_LO = DAG.getNode(ISD::UMULO, DL, VT, RCP, Den);
1056 // RCP_HI = mulhu (RCP, Den) */
1057 SDValue RCP_HI = DAG.getNode(ISD::MULHU, DL, VT, RCP, Den);
1059 // NEG_RCP_LO = -RCP_LO
1060 SDValue NEG_RCP_LO = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, VT),
1063 // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
1064 SDValue ABS_RCP_LO = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1067 // Calculate the rounding error from the URECIP instruction
1068 // E = mulhu(ABS_RCP_LO, RCP)
1069 SDValue E = DAG.getNode(ISD::MULHU, DL, VT, ABS_RCP_LO, RCP);
1071 // RCP_A_E = RCP + E
1072 SDValue RCP_A_E = DAG.getNode(ISD::ADD, DL, VT, RCP, E);
1074 // RCP_S_E = RCP - E
1075 SDValue RCP_S_E = DAG.getNode(ISD::SUB, DL, VT, RCP, E);
1077 // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
1078 SDValue Tmp0 = DAG.getSelectCC(DL, RCP_HI, DAG.getConstant(0, VT),
1081 // Quotient = mulhu(Tmp0, Num)
1082 SDValue Quotient = DAG.getNode(ISD::MULHU, DL, VT, Tmp0, Num);
1084 // Num_S_Remainder = Quotient * Den
1085 SDValue Num_S_Remainder = DAG.getNode(ISD::UMULO, DL, VT, Quotient, Den);
1087 // Remainder = Num - Num_S_Remainder
1088 SDValue Remainder = DAG.getNode(ISD::SUB, DL, VT, Num, Num_S_Remainder);
1090 // Remainder_GE_Den = (Remainder >= Den ? -1 : 0)
1091 SDValue Remainder_GE_Den = DAG.getSelectCC(DL, Remainder, Den,
1092 DAG.getConstant(-1, VT),
1093 DAG.getConstant(0, VT),
1095 // Remainder_GE_Zero = (Num >= Num_S_Remainder ? -1 : 0)
1096 SDValue Remainder_GE_Zero = DAG.getSelectCC(DL, Num,
1098 DAG.getConstant(-1, VT),
1099 DAG.getConstant(0, VT),
1101 // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
1102 SDValue Tmp1 = DAG.getNode(ISD::AND, DL, VT, Remainder_GE_Den,
1105 // Calculate Division result:
1107 // Quotient_A_One = Quotient + 1
1108 SDValue Quotient_A_One = DAG.getNode(ISD::ADD, DL, VT, Quotient,
1109 DAG.getConstant(1, VT));
1111 // Quotient_S_One = Quotient - 1
1112 SDValue Quotient_S_One = DAG.getNode(ISD::SUB, DL, VT, Quotient,
1113 DAG.getConstant(1, VT));
1115 // Div = (Tmp1 == 0 ? Quotient : Quotient_A_One)
1116 SDValue Div = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1117 Quotient, Quotient_A_One, ISD::SETEQ);
1119 // Div = (Remainder_GE_Zero == 0 ? Quotient_S_One : Div)
1120 Div = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1121 Quotient_S_One, Div, ISD::SETEQ);
1123 // Calculate Rem result:
1125 // Remainder_S_Den = Remainder - Den
1126 SDValue Remainder_S_Den = DAG.getNode(ISD::SUB, DL, VT, Remainder, Den);
1128 // Remainder_A_Den = Remainder + Den
1129 SDValue Remainder_A_Den = DAG.getNode(ISD::ADD, DL, VT, Remainder, Den);
1131 // Rem = (Tmp1 == 0 ? Remainder : Remainder_S_Den)
1132 SDValue Rem = DAG.getSelectCC(DL, Tmp1, DAG.getConstant(0, VT),
1133 Remainder, Remainder_S_Den, ISD::SETEQ);
1135 // Rem = (Remainder_GE_Zero == 0 ? Remainder_A_Den : Rem)
1136 Rem = DAG.getSelectCC(DL, Remainder_GE_Zero, DAG.getConstant(0, VT),
1137 Remainder_A_Den, Rem, ISD::SETEQ);
1142 return DAG.getMergeValues(Ops, DL);
1145 SDValue AMDGPUTargetLowering::LowerUINT_TO_FP(SDValue Op,
1146 SelectionDAG &DAG) const {
1147 SDValue S0 = Op.getOperand(0);
1149 if (Op.getValueType() != MVT::f32 || S0.getValueType() != MVT::i64)
1152 // f32 uint_to_fp i64
1153 SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1154 DAG.getConstant(0, MVT::i32));
1155 SDValue FloatLo = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Lo);
1156 SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DL, MVT::i32, S0,
1157 DAG.getConstant(1, MVT::i32));
1158 SDValue FloatHi = DAG.getNode(ISD::UINT_TO_FP, DL, MVT::f32, Hi);
1159 FloatHi = DAG.getNode(ISD::FMUL, DL, MVT::f32, FloatHi,
1160 DAG.getConstantFP(4294967296.0f, MVT::f32)); // 2^32
1161 return DAG.getNode(ISD::FADD, DL, MVT::f32, FloatLo, FloatHi);
1165 SDValue AMDGPUTargetLowering::ExpandSIGN_EXTEND_INREG(SDValue Op,
1167 SelectionDAG &DAG) const {
1168 MVT VT = Op.getSimpleValueType();
1170 SDValue Shift = DAG.getConstant(BitsDiff, VT);
1171 // Shift left by 'Shift' bits.
1172 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, Op.getOperand(0), Shift);
1173 // Signed shift Right by 'Shift' bits.
1174 return DAG.getNode(ISD::SRA, DL, VT, Shl, Shift);
1177 SDValue AMDGPUTargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op,
1178 SelectionDAG &DAG) const {
1179 EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
1180 MVT VT = Op.getSimpleValueType();
1181 MVT ScalarVT = VT.getScalarType();
1186 SDValue Src = Op.getOperand(0);
1189 // TODO: Don't scalarize on Evergreen?
1190 unsigned NElts = VT.getVectorNumElements();
1191 SmallVector<SDValue, 8> Args;
1192 DAG.ExtractVectorElements(Src, Args, 0, NElts);
1194 SDValue VTOp = DAG.getValueType(ExtraVT.getScalarType());
1195 for (unsigned I = 0; I < NElts; ++I)
1196 Args[I] = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ScalarVT, Args[I], VTOp);
1198 return DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Args);
1201 //===----------------------------------------------------------------------===//
1202 // Custom DAG optimizations
1203 //===----------------------------------------------------------------------===//
1205 static bool isU24(SDValue Op, SelectionDAG &DAG) {
1206 APInt KnownZero, KnownOne;
1207 EVT VT = Op.getValueType();
1208 DAG.ComputeMaskedBits(Op, KnownZero, KnownOne);
1210 return (VT.getSizeInBits() - KnownZero.countLeadingOnes()) <= 24;
1213 static bool isI24(SDValue Op, SelectionDAG &DAG) {
1214 EVT VT = Op.getValueType();
1216 // In order for this to be a signed 24-bit value, bit 23, must
1218 return VT.getSizeInBits() >= 24 && // Types less than 24-bit should be treated
1219 // as unsigned 24-bit values.
1220 (VT.getSizeInBits() - DAG.ComputeNumSignBits(Op)) < 24;
1223 static void simplifyI24(SDValue Op, TargetLowering::DAGCombinerInfo &DCI) {
1225 SelectionDAG &DAG = DCI.DAG;
1226 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
1227 EVT VT = Op.getValueType();
1229 APInt Demanded = APInt::getLowBitsSet(VT.getSizeInBits(), 24);
1230 APInt KnownZero, KnownOne;
1231 TargetLowering::TargetLoweringOpt TLO(DAG, true, true);
1232 if (TLI.SimplifyDemandedBits(Op, Demanded, KnownZero, KnownOne, TLO))
1233 DCI.CommitTargetLoweringOpt(TLO);
1236 SDValue AMDGPUTargetLowering::PerformDAGCombine(SDNode *N,
1237 DAGCombinerInfo &DCI) const {
1238 SelectionDAG &DAG = DCI.DAG;
1241 switch(N->getOpcode()) {
1244 EVT VT = N->getValueType(0);
1245 SDValue N0 = N->getOperand(0);
1246 SDValue N1 = N->getOperand(1);
1249 // FIXME: Add support for 24-bit multiply with 64-bit output on SI.
1250 if (VT.isVector() || VT.getSizeInBits() > 32)
1253 if (Subtarget->hasMulU24() && isU24(N0, DAG) && isU24(N1, DAG)) {
1254 N0 = DAG.getZExtOrTrunc(N0, DL, MVT::i32);
1255 N1 = DAG.getZExtOrTrunc(N1, DL, MVT::i32);
1256 Mul = DAG.getNode(AMDGPUISD::MUL_U24, DL, MVT::i32, N0, N1);
1257 } else if (Subtarget->hasMulI24() && isI24(N0, DAG) && isI24(N1, DAG)) {
1258 N0 = DAG.getSExtOrTrunc(N0, DL, MVT::i32);
1259 N1 = DAG.getSExtOrTrunc(N1, DL, MVT::i32);
1260 Mul = DAG.getNode(AMDGPUISD::MUL_I24, DL, MVT::i32, N0, N1);
1265 // We need to use sext even for MUL_U24, because MUL_U24 is used
1266 // for signed multiply of 8 and 16-bit types.
1267 SDValue Reg = DAG.getSExtOrTrunc(Mul, DL, VT);
1271 case AMDGPUISD::MUL_I24:
1272 case AMDGPUISD::MUL_U24: {
1273 SDValue N0 = N->getOperand(0);
1274 SDValue N1 = N->getOperand(1);
1275 simplifyI24(N0, DCI);
1276 simplifyI24(N1, DCI);
1283 //===----------------------------------------------------------------------===//
1285 //===----------------------------------------------------------------------===//
1287 void AMDGPUTargetLowering::getOriginalFunctionArgs(
1290 const SmallVectorImpl<ISD::InputArg> &Ins,
1291 SmallVectorImpl<ISD::InputArg> &OrigIns) const {
1293 for (unsigned i = 0, e = Ins.size(); i < e; ++i) {
1294 if (Ins[i].ArgVT == Ins[i].VT) {
1295 OrigIns.push_back(Ins[i]);
1300 if (Ins[i].ArgVT.isVector() && !Ins[i].VT.isVector()) {
1301 // Vector has been split into scalars.
1302 VT = Ins[i].ArgVT.getVectorElementType();
1303 } else if (Ins[i].VT.isVector() && Ins[i].ArgVT.isVector() &&
1304 Ins[i].ArgVT.getVectorElementType() !=
1305 Ins[i].VT.getVectorElementType()) {
1306 // Vector elements have been promoted
1309 // Vector has been spilt into smaller vectors.
1313 ISD::InputArg Arg(Ins[i].Flags, VT, VT, Ins[i].Used,
1314 Ins[i].OrigArgIndex, Ins[i].PartOffset);
1315 OrigIns.push_back(Arg);
1319 bool AMDGPUTargetLowering::isHWTrueValue(SDValue Op) const {
1320 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1321 return CFP->isExactlyValue(1.0);
1323 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1324 return C->isAllOnesValue();
1329 bool AMDGPUTargetLowering::isHWFalseValue(SDValue Op) const {
1330 if (ConstantFPSDNode * CFP = dyn_cast<ConstantFPSDNode>(Op)) {
1331 return CFP->getValueAPF().isZero();
1333 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
1334 return C->isNullValue();
1339 SDValue AMDGPUTargetLowering::CreateLiveInRegister(SelectionDAG &DAG,
1340 const TargetRegisterClass *RC,
1341 unsigned Reg, EVT VT) const {
1342 MachineFunction &MF = DAG.getMachineFunction();
1343 MachineRegisterInfo &MRI = MF.getRegInfo();
1344 unsigned VirtualRegister;
1345 if (!MRI.isLiveIn(Reg)) {
1346 VirtualRegister = MRI.createVirtualRegister(RC);
1347 MRI.addLiveIn(Reg, VirtualRegister);
1349 VirtualRegister = MRI.getLiveInVirtReg(Reg);
1351 return DAG.getRegister(VirtualRegister, VT);
1354 #define NODE_NAME_CASE(node) case AMDGPUISD::node: return #node;
1356 const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
1358 default: return nullptr;
1360 NODE_NAME_CASE(CALL);
1361 NODE_NAME_CASE(UMUL);
1362 NODE_NAME_CASE(DIV_INF);
1363 NODE_NAME_CASE(RET_FLAG);
1364 NODE_NAME_CASE(BRANCH_COND);
1367 NODE_NAME_CASE(DWORDADDR)
1368 NODE_NAME_CASE(FRACT)
1369 NODE_NAME_CASE(FMAX)
1370 NODE_NAME_CASE(SMAX)
1371 NODE_NAME_CASE(UMAX)
1372 NODE_NAME_CASE(FMIN)
1373 NODE_NAME_CASE(SMIN)
1374 NODE_NAME_CASE(UMIN)
1375 NODE_NAME_CASE(BFE_U32)
1376 NODE_NAME_CASE(BFE_I32)
1379 NODE_NAME_CASE(MUL_U24)
1380 NODE_NAME_CASE(MUL_I24)
1381 NODE_NAME_CASE(URECIP)
1382 NODE_NAME_CASE(DOT4)
1383 NODE_NAME_CASE(EXPORT)
1384 NODE_NAME_CASE(CONST_ADDRESS)
1385 NODE_NAME_CASE(REGISTER_LOAD)
1386 NODE_NAME_CASE(REGISTER_STORE)
1387 NODE_NAME_CASE(LOAD_CONSTANT)
1388 NODE_NAME_CASE(LOAD_INPUT)
1389 NODE_NAME_CASE(SAMPLE)
1390 NODE_NAME_CASE(SAMPLEB)
1391 NODE_NAME_CASE(SAMPLED)
1392 NODE_NAME_CASE(SAMPLEL)
1393 NODE_NAME_CASE(STORE_MSKOR)
1394 NODE_NAME_CASE(TBUFFER_STORE_FORMAT)
1398 static void computeMaskedBitsForMinMax(const SDValue Op0,
1402 const SelectionDAG &DAG,
1404 APInt Op0Zero, Op0One;
1405 APInt Op1Zero, Op1One;
1406 DAG.ComputeMaskedBits(Op0, Op0Zero, Op0One, Depth);
1407 DAG.ComputeMaskedBits(Op1, Op1Zero, Op1One, Depth);
1409 KnownZero = Op0Zero & Op1Zero;
1410 KnownOne = Op0One & Op1One;
1413 void AMDGPUTargetLowering::computeMaskedBitsForTargetNode(
1417 const SelectionDAG &DAG,
1418 unsigned Depth) const {
1420 KnownZero = KnownOne = APInt(KnownOne.getBitWidth(), 0); // Don't know anything.
1421 unsigned Opc = Op.getOpcode();
1423 case ISD::INTRINSIC_WO_CHAIN: {
1424 // FIXME: The intrinsic should just use the node.
1425 switch (cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue()) {
1426 case AMDGPUIntrinsic::AMDGPU_imax:
1427 case AMDGPUIntrinsic::AMDGPU_umax:
1428 case AMDGPUIntrinsic::AMDGPU_imin:
1429 case AMDGPUIntrinsic::AMDGPU_umin:
1430 computeMaskedBitsForMinMax(Op.getOperand(1), Op.getOperand(2),
1431 KnownZero, KnownOne, DAG, Depth);
1439 case AMDGPUISD::SMAX:
1440 case AMDGPUISD::UMAX:
1441 case AMDGPUISD::SMIN:
1442 case AMDGPUISD::UMIN:
1443 computeMaskedBitsForMinMax(Op.getOperand(0), Op.getOperand(1),
1444 KnownZero, KnownOne, DAG, Depth);